|
|
|
|
/* Copyright (C) 2007-2024 Open Information Security Foundation
|
|
|
|
|
*
|
|
|
|
|
* You can copy, redistribute or modify this Program under the terms of
|
|
|
|
|
* the GNU General Public License version 2 as published by the Free
|
|
|
|
|
* Software Foundation.
|
|
|
|
|
*
|
|
|
|
|
* This program is distributed in the hope that it will be useful,
|
|
|
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
|
|
|
* GNU General Public License for more details.
|
|
|
|
|
*
|
|
|
|
|
* You should have received a copy of the GNU General Public License
|
|
|
|
|
* version 2 along with this program; if not, write to the Free Software
|
|
|
|
|
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
|
|
|
|
|
* 02110-1301, USA.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* \file
|
|
|
|
|
*
|
|
|
|
|
* \author Anoop Saldanha <anoopsaldanha@gmail.com>
|
|
|
|
|
* \author Victor Julien <victor@inliniac.net>
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
#include "suricata-common.h"
|
|
|
|
|
#include "conf.h"
|
|
|
|
|
#include "threadvars.h"
|
|
|
|
|
#include "tm-threads.h"
|
|
|
|
|
#include "runmodes.h"
|
|
|
|
|
|
|
|
|
|
#include "util-time.h"
|
|
|
|
|
|
|
|
|
|
#include "flow.h"
|
|
|
|
|
#include "flow-queue.h"
|
|
|
|
|
#include "flow-hash.h"
|
|
|
|
|
#include "flow-util.h"
|
|
|
|
|
#include "flow-private.h"
|
|
|
|
|
#include "flow-timeout.h"
|
|
|
|
|
#include "flow-manager.h"
|
|
|
|
|
#include "flow-storage.h"
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
#include "flow-spare-pool.h"
|
|
|
|
|
#include "flow-callbacks.h"
|
|
|
|
|
|
|
|
|
|
#include "stream-tcp.h"
|
|
|
|
|
#include "stream-tcp-cache.h"
|
|
|
|
|
|
|
|
|
|
#include "util-device-private.h"
|
|
|
|
|
|
|
|
|
|
#include "util-debug.h"
|
|
|
|
|
|
|
|
|
|
#include "threads.h"
|
|
|
|
|
#include "detect-engine-threshold.h"
|
|
|
|
|
|
|
|
|
|
#include "host-timeout.h"
|
|
|
|
|
#include "defrag-hash.h"
|
|
|
|
|
#include "defrag-timeout.h"
|
|
|
|
|
#include "ippair-timeout.h"
|
|
|
|
|
#include "rust.h"
|
|
|
|
|
#include "app-layer-htp-range.h"
|
|
|
|
|
|
|
|
|
|
#include "output-flow.h"
|
|
|
|
|
|
|
|
|
|
#include "runmode-unix-socket.h"
|
|
|
|
|
|
|
|
|
|
/** queue to pass flows to cleanup/log thread(s) */
|
|
|
|
|
FlowQueue flow_recycle_q;
|
|
|
|
|
|
|
|
|
|
/* multi flow manager support */
|
|
|
|
|
static uint32_t flowmgr_number = 1;
|
|
|
|
|
/* atomic counter for flow managers, to assign instance id */
|
|
|
|
|
SC_ATOMIC_DECLARE(uint32_t, flowmgr_cnt);
|
|
|
|
|
|
|
|
|
|
/* multi flow recycler support */
|
|
|
|
|
static uint32_t flowrec_number = 1;
|
|
|
|
|
/* atomic counter for flow recyclers, to assign instance id */
|
|
|
|
|
SC_ATOMIC_DECLARE(uint32_t, flowrec_cnt);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
SC_ATOMIC_DECLARE(uint32_t, flowrec_busy);
|
|
|
|
|
SC_ATOMIC_EXTERN(unsigned int, flow_flags);
|
|
|
|
|
|
|
|
|
|
static SCCtrlCondT flow_manager_ctrl_cond = PTHREAD_COND_INITIALIZER;
|
|
|
|
|
static SCCtrlMutex flow_manager_ctrl_mutex = PTHREAD_MUTEX_INITIALIZER;
|
|
|
|
|
static SCCtrlCondT flow_recycler_ctrl_cond = PTHREAD_COND_INITIALIZER;
|
|
|
|
|
static SCCtrlMutex flow_recycler_ctrl_mutex = PTHREAD_MUTEX_INITIALIZER;
|
|
|
|
|
|
|
|
|
|
void FlowWakeupFlowManagerThread(void)
|
|
|
|
|
{
|
|
|
|
|
SCCtrlMutexLock(&flow_manager_ctrl_mutex);
|
|
|
|
|
SCCtrlCondSignal(&flow_manager_ctrl_cond);
|
|
|
|
|
SCCtrlMutexUnlock(&flow_manager_ctrl_mutex);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void FlowWakeupFlowRecyclerThread(void)
|
|
|
|
|
{
|
|
|
|
|
SCCtrlMutexLock(&flow_recycler_ctrl_mutex);
|
|
|
|
|
SCCtrlCondSignal(&flow_recycler_ctrl_cond);
|
|
|
|
|
SCCtrlMutexUnlock(&flow_recycler_ctrl_mutex);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void FlowTimeoutsInit(void)
|
|
|
|
|
{
|
|
|
|
|
SC_ATOMIC_SET(flow_timeouts, flow_timeouts_normal);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void FlowTimeoutsEmergency(void)
|
|
|
|
|
{
|
|
|
|
|
SC_ATOMIC_SET(flow_timeouts, flow_timeouts_emerg);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
typedef struct FlowTimeoutCounters_ {
|
|
|
|
|
uint32_t rows_checked;
|
|
|
|
|
uint32_t rows_skipped;
|
|
|
|
|
uint32_t rows_empty;
|
|
|
|
|
uint32_t rows_maxlen;
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
uint32_t flows_checked;
|
|
|
|
|
uint32_t flows_notimeout;
|
|
|
|
|
uint32_t flows_timeout;
|
|
|
|
|
uint32_t flows_removed;
|
|
|
|
|
uint32_t flows_aside;
|
|
|
|
|
uint32_t flows_aside_needs_work;
|
|
|
|
|
|
|
|
|
|
uint32_t bypassed_count;
|
|
|
|
|
uint64_t bypassed_pkts;
|
|
|
|
|
uint64_t bypassed_bytes;
|
|
|
|
|
} FlowTimeoutCounters;
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* \brief Used to disable flow manager thread(s).
|
|
|
|
|
*
|
|
|
|
|
* \todo Kinda hackish since it uses the tv name to identify flow manager
|
|
|
|
|
* thread. We need an all weather identification scheme.
|
|
|
|
|
*/
|
|
|
|
|
void FlowDisableFlowManagerThread(void)
|
|
|
|
|
{
|
|
|
|
|
SCMutexLock(&tv_root_lock);
|
|
|
|
|
/* flow manager thread(s) is/are a part of mgmt threads */
|
|
|
|
|
for (ThreadVars *tv = tv_root[TVT_MGMT]; tv != NULL; tv = tv->next) {
|
|
|
|
|
if (strncasecmp(tv->name, thread_name_flow_mgr,
|
|
|
|
|
strlen(thread_name_flow_mgr)) == 0)
|
|
|
|
|
{
|
|
|
|
|
TmThreadsSetFlag(tv, THV_KILL);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
SCMutexUnlock(&tv_root_lock);
|
|
|
|
|
|
|
|
|
|
struct timeval start_ts;
|
|
|
|
|
struct timeval cur_ts;
|
|
|
|
|
gettimeofday(&start_ts, NULL);
|
|
|
|
|
|
|
|
|
|
again:
|
|
|
|
|
gettimeofday(&cur_ts, NULL);
|
|
|
|
|
if ((cur_ts.tv_sec - start_ts.tv_sec) > 60) {
|
|
|
|
|
FatalError("unable to get all flow manager "
|
|
|
|
|
"threads to shutdown in time");
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
SCMutexLock(&tv_root_lock);
|
|
|
|
|
for (ThreadVars *tv = tv_root[TVT_MGMT]; tv != NULL; tv = tv->next) {
|
|
|
|
|
if (strncasecmp(tv->name, thread_name_flow_mgr,
|
|
|
|
|
strlen(thread_name_flow_mgr)) == 0)
|
|
|
|
|
{
|
|
|
|
|
if (!TmThreadsCheckFlag(tv, THV_RUNNING_DONE)) {
|
|
|
|
|
SCMutexUnlock(&tv_root_lock);
|
|
|
|
|
/* sleep outside lock */
|
|
|
|
|
SleepMsec(1);
|
|
|
|
|
goto again;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
SCMutexUnlock(&tv_root_lock);
|
|
|
|
|
|
|
|
|
|
/* reset count, so we can kill and respawn (unix socket) */
|
|
|
|
|
SC_ATOMIC_SET(flowmgr_cnt, 0);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/** \internal
|
|
|
|
|
* \brief check if a flow is timed out
|
|
|
|
|
*
|
|
|
|
|
* Takes lastts, adds the timeout policy to it, compared to current time `ts`.
|
|
|
|
|
* In case of emergency mode, timeout_policy is ignored and the emerg table
|
|
|
|
|
* is used.
|
|
|
|
|
*
|
|
|
|
|
* \param f flow
|
|
|
|
|
* \param ts timestamp - realtime or a minimum of active threads in offline mode
|
|
|
|
|
* \param next_ts tracking the next timeout ts, so FM can skip the row until that time
|
|
|
|
|
* \param emerg bool to indicate if emergency timeout settings should be used
|
|
|
|
|
*
|
|
|
|
|
* \retval false not timed out
|
|
|
|
|
* \retval true timed out
|
|
|
|
|
*/
|
|
|
|
|
static bool FlowManagerFlowTimeout(Flow *f, SCTime_t ts, uint32_t *next_ts, const bool emerg)
|
|
|
|
|
{
|
|
|
|
|
SCTime_t timesout_at;
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
if (emerg) {
|
|
|
|
|
extern FlowProtoTimeout flow_timeouts_emerg[FLOW_PROTO_MAX];
|
|
|
|
|
timesout_at = SCTIME_ADD_SECS(f->lastts,
|
|
|
|
|
FlowGetFlowTimeoutDirect(flow_timeouts_emerg, f->flow_state, f->protomap));
|
|
|
|
|
} else {
|
|
|
|
|
timesout_at = SCTIME_ADD_SECS(f->lastts, f->timeout_policy);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
}
|
|
|
|
|
/* update next_ts if needed */
|
|
|
|
|
if (*next_ts == 0 || (uint32_t)SCTIME_SECS(timesout_at) < *next_ts)
|
|
|
|
|
*next_ts = (uint32_t)SCTIME_SECS(timesout_at);
|
|
|
|
|
|
|
|
|
|
/* if time is live, we just use the `ts` */
|
|
|
|
|
if (TimeModeIsLive() || f->thread_id[0] == 0) {
|
|
|
|
|
/* do the timeout check */
|
|
|
|
|
if (SCTIME_CMP_LT(ts, timesout_at)) {
|
|
|
|
|
return false;
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
/* offline: take last ts from "owning" thread */
|
|
|
|
|
SCTime_t checkts = TmThreadsGetThreadTime(f->thread_id[0]);
|
|
|
|
|
/* do the timeout check */
|
|
|
|
|
if (SCTIME_CMP_LT(checkts, timesout_at)) {
|
|
|
|
|
return false;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return true;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
#ifdef CAPTURE_OFFLOAD
|
|
|
|
|
/** \internal
|
|
|
|
|
* \brief check timeout of captured bypassed flow by querying capture method
|
|
|
|
|
*
|
|
|
|
|
* \param f Flow
|
|
|
|
|
* \param ts timestamp
|
|
|
|
|
* \param counters Flow timeout counters
|
|
|
|
|
*
|
|
|
|
|
* \retval false not timeout
|
|
|
|
|
* \retval true timeout (or not capture bypassed)
|
|
|
|
|
*/
|
|
|
|
|
static inline bool FlowBypassedTimeout(Flow *f, SCTime_t ts, FlowTimeoutCounters *counters)
|
|
|
|
|
{
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
if (f->flow_state != FLOW_STATE_CAPTURE_BYPASSED) {
|
|
|
|
|
return true;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
FlowBypassInfo *fc = FlowGetStorageById(f, GetFlowBypassInfoID());
|
|
|
|
|
if (fc && fc->BypassUpdate) {
|
|
|
|
|
/* flow will be possibly updated */
|
|
|
|
|
uint64_t pkts_tosrc = fc->tosrcpktcnt;
|
|
|
|
|
uint64_t bytes_tosrc = fc->tosrcbytecnt;
|
|
|
|
|
uint64_t pkts_todst = fc->todstpktcnt;
|
|
|
|
|
uint64_t bytes_todst = fc->todstbytecnt;
|
|
|
|
|
bool update = fc->BypassUpdate(f, fc->bypass_data, SCTIME_SECS(ts));
|
|
|
|
|
if (update) {
|
|
|
|
|
SCLogDebug("Updated flow: %" PRIu64 "", FlowGetId(f));
|
|
|
|
|
pkts_tosrc = fc->tosrcpktcnt - pkts_tosrc;
|
|
|
|
|
bytes_tosrc = fc->tosrcbytecnt - bytes_tosrc;
|
|
|
|
|
pkts_todst = fc->todstpktcnt - pkts_todst;
|
|
|
|
|
bytes_todst = fc->todstbytecnt - bytes_todst;
|
|
|
|
|
if (f->livedev) {
|
|
|
|
|
SC_ATOMIC_ADD(f->livedev->bypassed,
|
|
|
|
|
pkts_tosrc + pkts_todst);
|
|
|
|
|
}
|
|
|
|
|
counters->bypassed_pkts += pkts_tosrc + pkts_todst;
|
|
|
|
|
counters->bypassed_bytes += bytes_tosrc + bytes_todst;
|
|
|
|
|
return false;
|
|
|
|
|
}
|
|
|
|
|
SCLogDebug("No new packet, dead flow %" PRIu64 "", FlowGetId(f));
|
|
|
|
|
if (f->livedev) {
|
|
|
|
|
if (FLOW_IS_IPV4(f)) {
|
|
|
|
|
LiveDevSubBypassStats(f->livedev, 1, AF_INET);
|
|
|
|
|
} else if (FLOW_IS_IPV6(f)) {
|
|
|
|
|
LiveDevSubBypassStats(f->livedev, 1, AF_INET6);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
counters->bypassed_count++;
|
|
|
|
|
}
|
|
|
|
|
return true;
|
|
|
|
|
}
|
|
|
|
|
#endif /* CAPTURE_OFFLOAD */
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
typedef struct FlowManagerTimeoutThread {
|
|
|
|
|
/* used to temporarily store flows that have timed out and are
|
|
|
|
|
* removed from the hash to reduce locking contention */
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
FlowQueuePrivate aside_queue;
|
|
|
|
|
} FlowManagerTimeoutThread;
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* \internal
|
|
|
|
|
*
|
|
|
|
|
* \brief Process the temporary Aside Queue
|
|
|
|
|
* This means that as long as a flow f is not waiting on detection
|
|
|
|
|
* engine to finish dealing with it, f will be put in the recycle
|
|
|
|
|
* queue for further processing later on.
|
|
|
|
|
*
|
|
|
|
|
* \param td FM Timeout Thread instance
|
|
|
|
|
* \param counters Flow Timeout counters to be updated
|
|
|
|
|
*
|
|
|
|
|
* \retval Number of flows that were recycled
|
|
|
|
|
*/
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
static uint32_t ProcessAsideQueue(FlowManagerTimeoutThread *td, FlowTimeoutCounters *counters)
|
|
|
|
|
{
|
|
|
|
|
FlowQueuePrivate recycle = { NULL, NULL, 0 };
|
|
|
|
|
counters->flows_aside += td->aside_queue.len;
|
|
|
|
|
|
|
|
|
|
uint32_t cnt = 0;
|
|
|
|
|
Flow *f;
|
|
|
|
|
while ((f = FlowQueuePrivateGetFromTop(&td->aside_queue)) != NULL) {
|
|
|
|
|
/* flow is still locked */
|
|
|
|
|
|
|
|
|
|
if (f->proto == IPPROTO_TCP &&
|
|
|
|
|
!(f->flags & (FLOW_TIMEOUT_REASSEMBLY_DONE | FLOW_ACTION_DROP)) &&
|
|
|
|
|
!FlowIsBypassed(f) && FlowNeedsReassembly(f)) {
|
|
|
|
|
/* Send the flow to its thread */
|
|
|
|
|
FlowSendToLocalThread(f);
|
|
|
|
|
FLOWLOCK_UNLOCK(f);
|
|
|
|
|
/* flow ownership is already passed to the worker thread */
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
|
|
|
|
|
counters->flows_aside_needs_work++;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
FLOWLOCK_UNLOCK(f);
|
|
|
|
|
|
|
|
|
|
FlowQueuePrivateAppendFlow(&recycle, f);
|
|
|
|
|
if (recycle.len == 100) {
|
|
|
|
|
FlowQueueAppendPrivate(&flow_recycle_q, &recycle);
|
|
|
|
|
FlowWakeupFlowRecyclerThread();
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
}
|
|
|
|
|
cnt++;
|
|
|
|
|
}
|
|
|
|
|
if (recycle.len) {
|
|
|
|
|
FlowQueueAppendPrivate(&flow_recycle_q, &recycle);
|
|
|
|
|
FlowWakeupFlowRecyclerThread();
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
}
|
|
|
|
|
return cnt;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* \internal
|
|
|
|
|
*
|
|
|
|
|
* \brief check all flows in a hash row for timing out
|
|
|
|
|
*
|
|
|
|
|
* \param f last flow in the hash row
|
|
|
|
|
* \param ts timestamp
|
|
|
|
|
* \param emergency bool indicating emergency mode
|
|
|
|
|
* \param counters ptr to FlowTimeoutCounters structure
|
|
|
|
|
*/
|
|
|
|
|
static void FlowManagerHashRowTimeout(FlowManagerTimeoutThread *td, Flow *f, SCTime_t ts,
|
|
|
|
|
int emergency, FlowTimeoutCounters *counters, uint32_t *next_ts)
|
|
|
|
|
{
|
|
|
|
|
uint32_t checked = 0;
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
Flow *prev_f = NULL;
|
|
|
|
|
|
|
|
|
|
do {
|
|
|
|
|
checked++;
|
|
|
|
|
|
|
|
|
|
FLOWLOCK_WRLOCK(f);
|
|
|
|
|
|
|
|
|
|
/* check flow timeout based on lastts and state. Both can be
|
|
|
|
|
* accessed w/o Flow lock as we do have the hash row lock (so flow
|
|
|
|
|
* can't disappear) and flow_state is atomic. lastts can only
|
|
|
|
|
* be modified when we have both the flow and hash row lock */
|
|
|
|
|
|
|
|
|
|
/* timeout logic goes here */
|
|
|
|
|
if (!FlowManagerFlowTimeout(f, ts, next_ts, emergency)) {
|
|
|
|
|
FLOWLOCK_UNLOCK(f);
|
|
|
|
|
counters->flows_notimeout++;
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
prev_f = f;
|
|
|
|
|
f = f->next;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
Flow *next_flow = f->next;
|
|
|
|
|
|
|
|
|
|
#ifdef CAPTURE_OFFLOAD
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
/* never prune a flow that is used by a packet we
|
|
|
|
|
* are currently processing in one of the threads */
|
|
|
|
|
if (!FlowBypassedTimeout(f, ts, counters)) {
|
|
|
|
|
FLOWLOCK_UNLOCK(f);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
prev_f = f;
|
|
|
|
|
f = f->next;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
f->flow_end_flags |= FLOW_END_FLAG_TIMEOUT;
|
|
|
|
|
|
|
|
|
|
counters->flows_timeout++;
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
RemoveFromHash(f, prev_f);
|
|
|
|
|
|
|
|
|
|
FlowQueuePrivateAppendFlow(&td->aside_queue, f);
|
|
|
|
|
/* flow is still locked in the queue */
|
|
|
|
|
|
|
|
|
|
f = next_flow;
|
|
|
|
|
} while (f != NULL);
|
|
|
|
|
|
|
|
|
|
counters->flows_checked += checked;
|
|
|
|
|
if (checked > counters->rows_maxlen)
|
|
|
|
|
counters->rows_maxlen = checked;
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* \internal
|
|
|
|
|
*
|
|
|
|
|
* \brief Clear evicted list from Flow Manager.
|
|
|
|
|
* All the evicted flows are removed from the Flow bucket and added
|
|
|
|
|
* to the temporary Aside Queue.
|
|
|
|
|
*
|
|
|
|
|
* \param td FM timeout thread instance
|
|
|
|
|
* \param f head of the evicted list
|
|
|
|
|
*/
|
|
|
|
|
static void FlowManagerHashRowClearEvictedList(FlowManagerTimeoutThread *td, Flow *f)
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
{
|
|
|
|
|
do {
|
|
|
|
|
FLOWLOCK_WRLOCK(f);
|
|
|
|
|
Flow *next_flow = f->next;
|
|
|
|
|
f->next = NULL;
|
|
|
|
|
f->fb = NULL;
|
|
|
|
|
|
|
|
|
|
FlowQueuePrivateAppendFlow(&td->aside_queue, f);
|
|
|
|
|
/* flow is still locked in the queue */
|
|
|
|
|
|
|
|
|
|
f = next_flow;
|
|
|
|
|
} while (f != NULL);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* \brief time out flows from the hash
|
|
|
|
|
*
|
|
|
|
|
* \param ts timestamp
|
|
|
|
|
* \param hash_min min hash index to consider
|
|
|
|
|
* \param hash_max max hash index to consider
|
|
|
|
|
* \param counters ptr to FlowTimeoutCounters structure
|
|
|
|
|
*
|
|
|
|
|
* \retval cnt number of timed out flow
|
|
|
|
|
*/
|
|
|
|
|
static uint32_t FlowTimeoutHash(FlowManagerTimeoutThread *td, SCTime_t ts, const uint32_t hash_min,
|
|
|
|
|
const uint32_t hash_max, FlowTimeoutCounters *counters)
|
|
|
|
|
{
|
|
|
|
|
uint32_t cnt = 0;
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
const int emergency = ((SC_ATOMIC_GET(flow_flags) & FLOW_EMERGENCY));
|
|
|
|
|
const uint32_t rows_checked = hash_max - hash_min;
|
|
|
|
|
uint32_t rows_skipped = 0;
|
|
|
|
|
uint32_t rows_empty = 0;
|
|
|
|
|
|
|
|
|
|
#if __WORDSIZE==64
|
|
|
|
|
#define BITS 64
|
|
|
|
|
#define TYPE uint64_t
|
|
|
|
|
#else
|
|
|
|
|
#define BITS 32
|
|
|
|
|
#define TYPE uint32_t
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
const uint32_t ts_secs = (uint32_t)SCTIME_SECS(ts);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
for (uint32_t idx = hash_min; idx < hash_max; idx+=BITS) {
|
|
|
|
|
TYPE check_bits = 0;
|
|
|
|
|
const uint32_t check = MIN(BITS, (hash_max - idx));
|
|
|
|
|
for (uint32_t i = 0; i < check; i++) {
|
|
|
|
|
FlowBucket *fb = &flow_hash[idx+i];
|
|
|
|
|
check_bits |= (TYPE)(SC_ATOMIC_LOAD_EXPLICIT(
|
|
|
|
|
fb->next_ts, SC_ATOMIC_MEMORY_ORDER_RELAXED) <= ts_secs)
|
|
|
|
|
<< (TYPE)i;
|
|
|
|
|
}
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
if (check_bits == 0)
|
|
|
|
|
continue;
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
for (uint32_t i = 0; i < check; i++) {
|
|
|
|
|
FlowBucket *fb = &flow_hash[idx+i];
|
|
|
|
|
if ((check_bits & ((TYPE)1 << (TYPE)i)) != 0 && SC_ATOMIC_GET(fb->next_ts) <= ts_secs) {
|
|
|
|
|
FBLOCK_LOCK(fb);
|
|
|
|
|
Flow *evicted = NULL;
|
|
|
|
|
if (fb->evicted != NULL || fb->head != NULL) {
|
|
|
|
|
if (fb->evicted != NULL) {
|
|
|
|
|
/* transfer out of bucket so we can do additional work outside
|
|
|
|
|
* of the bucket lock */
|
|
|
|
|
evicted = fb->evicted;
|
|
|
|
|
fb->evicted = NULL;
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
}
|
|
|
|
|
if (fb->head != NULL) {
|
|
|
|
|
uint32_t next_ts = 0;
|
|
|
|
|
FlowManagerHashRowTimeout(td, fb->head, ts, emergency, counters, &next_ts);
|
|
|
|
|
|
|
|
|
|
if (SC_ATOMIC_GET(fb->next_ts) != next_ts)
|
|
|
|
|
SC_ATOMIC_SET(fb->next_ts, next_ts);
|
|
|
|
|
}
|
|
|
|
|
if (fb->evicted == NULL && fb->head == NULL) {
|
|
|
|
|
/* row is empty */
|
|
|
|
|
SC_ATOMIC_SET(fb->next_ts, UINT_MAX);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
SC_ATOMIC_SET(fb->next_ts, UINT_MAX);
|
|
|
|
|
rows_empty++;
|
|
|
|
|
}
|
|
|
|
|
FBLOCK_UNLOCK(fb);
|
|
|
|
|
/* processed evicted list */
|
|
|
|
|
if (evicted) {
|
|
|
|
|
FlowManagerHashRowClearEvictedList(td, evicted);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
rows_skipped++;
|
|
|
|
|
}
|
|
|
|
|
}
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
if (td->aside_queue.len) {
|
|
|
|
|
cnt += ProcessAsideQueue(td, counters);
|
|
|
|
|
}
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
}
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
counters->rows_checked += rows_checked;
|
|
|
|
|
counters->rows_skipped += rows_skipped;
|
|
|
|
|
counters->rows_empty += rows_empty;
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
if (td->aside_queue.len) {
|
|
|
|
|
cnt += ProcessAsideQueue(td, counters);
|
|
|
|
|
}
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
counters->flows_removed += cnt;
|
|
|
|
|
/* coverity[missing_unlock : FALSE] */
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
return cnt;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/** \internal
|
|
|
|
|
*
|
|
|
|
|
* \brief handle timeout for a slice of hash rows
|
|
|
|
|
* If we wrap around we call FlowTimeoutHash twice
|
|
|
|
|
* \param td FM timeout thread
|
|
|
|
|
* \param ts timeout timestamp
|
|
|
|
|
* \param hash_min lower bound of the row slice
|
|
|
|
|
* \param hash_max upper bound of the row slice
|
|
|
|
|
* \param counters Flow timeout counters to be passed
|
|
|
|
|
* \param rows number of rows for this worker unit
|
|
|
|
|
* \param pos absolute position of the beginning of row slice in the hash table
|
|
|
|
|
* \param instance instance id of this FM
|
|
|
|
|
*
|
|
|
|
|
* \retval number of successfully timed out flows
|
|
|
|
|
*/
|
|
|
|
|
static uint32_t FlowTimeoutHashInChunks(FlowManagerTimeoutThread *td, SCTime_t ts,
|
|
|
|
|
const uint32_t hash_min, const uint32_t hash_max, FlowTimeoutCounters *counters,
|
|
|
|
|
const uint32_t rows, uint32_t *pos, const uint32_t instance)
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
{
|
|
|
|
|
uint32_t start = 0;
|
|
|
|
|
uint32_t end = 0;
|
|
|
|
|
uint32_t cnt = 0;
|
|
|
|
|
uint32_t rows_left = rows;
|
|
|
|
|
|
|
|
|
|
again:
|
|
|
|
|
start = (*pos);
|
|
|
|
|
if (start >= hash_max) {
|
|
|
|
|
start = hash_min;
|
|
|
|
|
}
|
|
|
|
|
end = start + rows_left;
|
|
|
|
|
if (end > hash_max) {
|
|
|
|
|
end = hash_max;
|
|
|
|
|
}
|
|
|
|
|
*pos = (end == hash_max) ? hash_min : end;
|
|
|
|
|
rows_left = rows_left - (end - start);
|
|
|
|
|
|
|
|
|
|
SCLogDebug("instance %u: %u:%u (hash_min %u, hash_max %u *pos %u)", instance, start, end,
|
|
|
|
|
hash_min, hash_max, *pos);
|
|
|
|
|
|
|
|
|
|
cnt += FlowTimeoutHash(td, ts, start, end, counters);
|
|
|
|
|
if (rows_left) {
|
|
|
|
|
goto again;
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
}
|
|
|
|
|
return cnt;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* \internal
|
|
|
|
|
*
|
|
|
|
|
* \brief move all flows out of a hash row
|
|
|
|
|
*
|
|
|
|
|
* \param f last flow in the hash row
|
|
|
|
|
* \param recycle_q Flow recycle queue
|
|
|
|
|
* \param mode emergency or not
|
|
|
|
|
*
|
|
|
|
|
* \retval cnt number of flows removed from the hash and added to the recycle queue
|
|
|
|
|
*/
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
static uint32_t FlowManagerHashRowCleanup(Flow *f, FlowQueuePrivate *recycle_q, const int mode)
|
|
|
|
|
{
|
|
|
|
|
uint32_t cnt = 0;
|
|
|
|
|
|
|
|
|
|
do {
|
|
|
|
|
FLOWLOCK_WRLOCK(f);
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
Flow *next_flow = f->next;
|
|
|
|
|
|
|
|
|
|
/* remove from the hash */
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
if (mode == 0) {
|
|
|
|
|
RemoveFromHash(f, NULL);
|
|
|
|
|
} else {
|
|
|
|
|
FlowBucket *fb = f->fb;
|
|
|
|
|
fb->evicted = f->next;
|
|
|
|
|
f->next = NULL;
|
|
|
|
|
f->fb = NULL;
|
|
|
|
|
}
|
|
|
|
|
f->flow_end_flags |= FLOW_END_FLAG_SHUTDOWN;
|
|
|
|
|
|
|
|
|
|
/* no one is referring to this flow, removed from hash
|
|
|
|
|
* so we can unlock it and move it to the recycle queue. */
|
|
|
|
|
FLOWLOCK_UNLOCK(f);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
FlowQueuePrivateAppendFlow(recycle_q, f);
|
|
|
|
|
|
|
|
|
|
cnt++;
|
|
|
|
|
|
|
|
|
|
f = next_flow;
|
|
|
|
|
} while (f != NULL);
|
|
|
|
|
|
|
|
|
|
return cnt;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
#define RECYCLE_MAX_QUEUE_ITEMS 25
|
|
|
|
|
/**
|
|
|
|
|
* \brief remove all flows from the hash
|
|
|
|
|
*
|
|
|
|
|
* \retval cnt number of removes out flows
|
|
|
|
|
*/
|
|
|
|
|
static uint32_t FlowCleanupHash(void)
|
|
|
|
|
{
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
FlowQueuePrivate local_queue = { NULL, NULL, 0 };
|
|
|
|
|
uint32_t cnt = 0;
|
|
|
|
|
|
|
|
|
|
for (uint32_t idx = 0; idx < flow_config.hash_size; idx++) {
|
|
|
|
|
FlowBucket *fb = &flow_hash[idx];
|
|
|
|
|
|
|
|
|
|
FBLOCK_LOCK(fb);
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
if (fb->head != NULL) {
|
|
|
|
|
/* we have a flow, or more than one */
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
cnt += FlowManagerHashRowCleanup(fb->head, &local_queue, 0);
|
|
|
|
|
}
|
|
|
|
|
if (fb->evicted != NULL) {
|
|
|
|
|
/* we have a flow, or more than one */
|
|
|
|
|
cnt += FlowManagerHashRowCleanup(fb->evicted, &local_queue, 1);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
FBLOCK_UNLOCK(fb);
|
|
|
|
|
if (local_queue.len >= RECYCLE_MAX_QUEUE_ITEMS) {
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
FlowQueueAppendPrivate(&flow_recycle_q, &local_queue);
|
|
|
|
|
FlowWakeupFlowRecyclerThread();
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
DEBUG_VALIDATE_BUG_ON(local_queue.len >= RECYCLE_MAX_QUEUE_ITEMS);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
FlowQueueAppendPrivate(&flow_recycle_q, &local_queue);
|
|
|
|
|
FlowWakeupFlowRecyclerThread();
|
|
|
|
|
|
|
|
|
|
return cnt;
|
|
|
|
|
}
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
typedef struct FlowCounters_ {
|
|
|
|
|
StatsCounterId flow_mgr_full_pass;
|
|
|
|
|
StatsCounterId flow_mgr_rows_sec;
|
|
|
|
|
|
|
|
|
|
StatsCounterId flow_mgr_spare;
|
|
|
|
|
StatsCounterId flow_emerg_mode_enter;
|
|
|
|
|
StatsCounterId flow_emerg_mode_over;
|
|
|
|
|
|
|
|
|
|
StatsCounterId flow_mgr_flows_checked;
|
|
|
|
|
StatsCounterId flow_mgr_flows_notimeout;
|
|
|
|
|
StatsCounterId flow_mgr_flows_timeout;
|
|
|
|
|
StatsCounterId flow_mgr_flows_aside;
|
|
|
|
|
StatsCounterId flow_mgr_flows_aside_needs_work;
|
|
|
|
|
|
|
|
|
|
StatsCounterMaxId flow_mgr_rows_maxlen;
|
|
|
|
|
|
|
|
|
|
StatsCounterId flow_bypassed_cnt_clo;
|
|
|
|
|
StatsCounterId flow_bypassed_pkts;
|
|
|
|
|
StatsCounterId flow_bypassed_bytes;
|
|
|
|
|
|
|
|
|
|
StatsCounterId memcap_pressure;
|
|
|
|
|
StatsCounterMaxId memcap_pressure_max;
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
} FlowCounters;
|
|
|
|
|
|
|
|
|
|
typedef struct FlowManagerThreadData_ {
|
|
|
|
|
uint32_t instance;
|
|
|
|
|
uint32_t min;
|
|
|
|
|
uint32_t max;
|
|
|
|
|
|
|
|
|
|
FlowCounters cnt;
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
FlowManagerTimeoutThread timeout;
|
|
|
|
|
StatsCounterId counter_defrag_timeout;
|
|
|
|
|
StatsCounterId counter_defrag_memuse;
|
|
|
|
|
} FlowManagerThreadData;
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
static void FlowCountersInit(ThreadVars *t, FlowCounters *fc)
|
|
|
|
|
{
|
|
|
|
|
fc->flow_mgr_full_pass = StatsRegisterCounter("flow.mgr.full_hash_pass", t);
|
|
|
|
|
fc->flow_mgr_rows_sec = StatsRegisterCounter("flow.mgr.rows_per_sec", t);
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
fc->flow_mgr_spare = StatsRegisterCounter("flow.spare", t);
|
|
|
|
|
fc->flow_emerg_mode_enter = StatsRegisterCounter("flow.emerg_mode_entered", t);
|
|
|
|
|
fc->flow_emerg_mode_over = StatsRegisterCounter("flow.emerg_mode_over", t);
|
|
|
|
|
|
|
|
|
|
fc->flow_mgr_rows_maxlen = StatsRegisterMaxCounter("flow.mgr.rows_maxlen", t);
|
|
|
|
|
fc->flow_mgr_flows_checked = StatsRegisterCounter("flow.mgr.flows_checked", t);
|
|
|
|
|
fc->flow_mgr_flows_notimeout = StatsRegisterCounter("flow.mgr.flows_notimeout", t);
|
|
|
|
|
fc->flow_mgr_flows_timeout = StatsRegisterCounter("flow.mgr.flows_timeout", t);
|
|
|
|
|
fc->flow_mgr_flows_aside = StatsRegisterCounter("flow.mgr.flows_evicted", t);
|
|
|
|
|
fc->flow_mgr_flows_aside_needs_work = StatsRegisterCounter("flow.mgr.flows_evicted_needs_work", t);
|
|
|
|
|
|
|
|
|
|
fc->flow_bypassed_cnt_clo = StatsRegisterCounter("flow_bypassed.closed", t);
|
|
|
|
|
fc->flow_bypassed_pkts = StatsRegisterCounter("flow_bypassed.pkts", t);
|
|
|
|
|
fc->flow_bypassed_bytes = StatsRegisterCounter("flow_bypassed.bytes", t);
|
|
|
|
|
|
|
|
|
|
fc->memcap_pressure = StatsRegisterCounter("memcap.pressure", t);
|
|
|
|
|
fc->memcap_pressure_max = StatsRegisterMaxCounter("memcap.pressure_max", t);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void FlowCountersUpdate(
|
|
|
|
|
ThreadVars *th_v, const FlowManagerThreadData *ftd, const FlowTimeoutCounters *counters)
|
|
|
|
|
{
|
|
|
|
|
StatsCounterAddI64(
|
|
|
|
|
&th_v->stats, ftd->cnt.flow_mgr_flows_checked, (int64_t)counters->flows_checked);
|
|
|
|
|
StatsCounterAddI64(
|
|
|
|
|
&th_v->stats, ftd->cnt.flow_mgr_flows_notimeout, (int64_t)counters->flows_notimeout);
|
|
|
|
|
|
|
|
|
|
StatsCounterAddI64(
|
|
|
|
|
&th_v->stats, ftd->cnt.flow_mgr_flows_timeout, (int64_t)counters->flows_timeout);
|
|
|
|
|
StatsCounterAddI64(&th_v->stats, ftd->cnt.flow_mgr_flows_aside, (int64_t)counters->flows_aside);
|
|
|
|
|
StatsCounterAddI64(&th_v->stats, ftd->cnt.flow_mgr_flows_aside_needs_work,
|
|
|
|
|
(int64_t)counters->flows_aside_needs_work);
|
|
|
|
|
|
|
|
|
|
StatsCounterAddI64(
|
|
|
|
|
&th_v->stats, ftd->cnt.flow_bypassed_cnt_clo, (int64_t)counters->bypassed_count);
|
|
|
|
|
StatsCounterAddI64(&th_v->stats, ftd->cnt.flow_bypassed_pkts, (int64_t)counters->bypassed_pkts);
|
|
|
|
|
StatsCounterAddI64(
|
|
|
|
|
&th_v->stats, ftd->cnt.flow_bypassed_bytes, (int64_t)counters->bypassed_bytes);
|
|
|
|
|
|
|
|
|
|
StatsCounterMaxUpdateI64(
|
|
|
|
|
&th_v->stats, ftd->cnt.flow_mgr_rows_maxlen, (int64_t)counters->rows_maxlen);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static TmEcode FlowManagerThreadInit(ThreadVars *t, const void *initdata, void **data)
|
|
|
|
|
{
|
|
|
|
|
FlowManagerThreadData *ftd = SCCalloc(1, sizeof(FlowManagerThreadData));
|
|
|
|
|
if (ftd == NULL)
|
|
|
|
|
return TM_ECODE_FAILED;
|
|
|
|
|
|
|
|
|
|
ftd->instance = SC_ATOMIC_ADD(flowmgr_cnt, 1);
|
|
|
|
|
SCLogDebug("flow manager instance %u", ftd->instance);
|
|
|
|
|
|
|
|
|
|
/* set the min and max value used for hash row walking
|
|
|
|
|
* each thread has it's own section of the flow hash */
|
|
|
|
|
uint32_t range = flow_config.hash_size / flowmgr_number;
|
|
|
|
|
|
|
|
|
|
ftd->min = ftd->instance * range;
|
|
|
|
|
ftd->max = (ftd->instance + 1) * range;
|
|
|
|
|
|
|
|
|
|
/* last flow-manager takes on hash_size % flowmgr_number extra rows */
|
|
|
|
|
if ((ftd->instance + 1) == flowmgr_number) {
|
|
|
|
|
ftd->max = flow_config.hash_size;
|
|
|
|
|
}
|
|
|
|
|
BUG_ON(ftd->min > flow_config.hash_size || ftd->max > flow_config.hash_size);
|
|
|
|
|
|
|
|
|
|
SCLogDebug("instance %u hash range %u %u", ftd->instance, ftd->min, ftd->max);
|
|
|
|
|
|
|
|
|
|
/* pass thread data back to caller */
|
|
|
|
|
*data = ftd;
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
FlowCountersInit(t, &ftd->cnt);
|
|
|
|
|
ftd->counter_defrag_timeout = StatsRegisterCounter("defrag.mgr.tracker_timeout", t);
|
|
|
|
|
ftd->counter_defrag_memuse = StatsRegisterCounter("defrag.memuse", t);
|
|
|
|
|
|
|
|
|
|
PacketPoolInit();
|
|
|
|
|
return TM_ECODE_OK;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static TmEcode FlowManagerThreadDeinit(ThreadVars *t, void *data)
|
|
|
|
|
{
|
|
|
|
|
StreamTcpThreadCacheCleanup();
|
|
|
|
|
PacketPoolDestroy();
|
|
|
|
|
SCFree(data);
|
|
|
|
|
return TM_ECODE_OK;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/** \internal
|
|
|
|
|
* \brief calculate number of rows to scan and how much time to sleep
|
|
|
|
|
* based on the busy score `mp` (0 idle, 100 max busy).
|
|
|
|
|
*
|
|
|
|
|
* We try to to make sure we scan the hash once a second. The number size
|
|
|
|
|
* of the slice of the hash scanned is determined by our busy score 'mp'.
|
|
|
|
|
* We sleep for the remainder of the second after processing the slice,
|
|
|
|
|
* or at least an approximation of it.
|
|
|
|
|
* A minimum busy score of 10 is assumed to avoid a longer than 10 second
|
|
|
|
|
* full hash pass. This is to avoid burstiness in scanning when there is
|
|
|
|
|
* a rapid increase of the busy score, which could lead to the flow manager
|
|
|
|
|
* suddenly scanning a much larger slice of the hash leading to a burst
|
|
|
|
|
* in scan/eviction work.
|
|
|
|
|
*
|
|
|
|
|
* \param rows number of rows for the work unit
|
|
|
|
|
* \param mp current memcap pressure value
|
|
|
|
|
* \param emergency emergency mode is set or not
|
|
|
|
|
* \param wu_sleep holds value of sleep time per worker unit
|
|
|
|
|
* \param wu_rows holds value of calculated rows to be processed per second
|
|
|
|
|
* \param rows_sec same as wu_rows, only used for counter updates
|
|
|
|
|
*/
|
|
|
|
|
static void GetWorkUnitSizing(const uint32_t rows, const uint32_t mp, const bool emergency,
|
|
|
|
|
uint64_t *wu_sleep, uint32_t *wu_rows, uint32_t *rows_sec)
|
|
|
|
|
{
|
|
|
|
|
if (emergency) {
|
|
|
|
|
*wu_rows = rows;
|
|
|
|
|
*wu_sleep = 250;
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
/* minimum busy score is 10 */
|
|
|
|
|
const uint32_t emp = MAX(mp, 10);
|
|
|
|
|
const uint32_t rows_per_sec = (uint32_t)((float)rows * (float)((float)emp / (float)100));
|
|
|
|
|
/* calc how much time we estimate the work will take, in ms. We assume
|
|
|
|
|
* each row takes an average of 1usec. Maxing out at 1sec. */
|
|
|
|
|
const uint32_t work_per_unit = MIN(rows_per_sec / 1000, 1000);
|
|
|
|
|
/* calc how much time we need to sleep to get to the per second cadence
|
|
|
|
|
* but sleeping for at least 250ms. */
|
|
|
|
|
const uint32_t sleep_per_unit = MAX(250, 1000 - work_per_unit);
|
|
|
|
|
SCLogDebug("mp %u emp %u rows %u rows_sec %u sleep %ums", mp, emp, rows, rows_per_sec,
|
|
|
|
|
sleep_per_unit);
|
|
|
|
|
|
|
|
|
|
*wu_sleep = sleep_per_unit;
|
|
|
|
|
*wu_rows = rows_per_sec;
|
|
|
|
|
*rows_sec = rows_per_sec;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/** \brief Thread that manages the flow table and times out flows.
|
|
|
|
|
*
|
|
|
|
|
* \param td ThreadVars cast to void ptr
|
|
|
|
|
*
|
|
|
|
|
* Keeps an eye on the spare list, alloc flows if needed...
|
|
|
|
|
*/
|
|
|
|
|
static TmEcode FlowManager(ThreadVars *th_v, void *thread_data)
|
|
|
|
|
{
|
|
|
|
|
FlowManagerThreadData *ftd = thread_data;
|
|
|
|
|
const uint32_t rows = ftd->max - ftd->min;
|
|
|
|
|
const bool time_is_live = TimeModeIsLive();
|
|
|
|
|
|
|
|
|
|
uint32_t emerg_over_cnt = 0;
|
|
|
|
|
uint64_t next_run_ms = 0;
|
|
|
|
|
uint32_t pos = ftd->min;
|
|
|
|
|
uint32_t rows_sec = 0;
|
|
|
|
|
uint32_t rows_per_wu = 0;
|
|
|
|
|
uint64_t sleep_per_wu = 0;
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
bool prev_emerg = false;
|
|
|
|
|
uint32_t other_last_sec = 0; /**< last sec stamp when defrag etc ran */
|
|
|
|
|
|
|
|
|
|
uint32_t mp = MemcapsGetPressure() * 100;
|
|
|
|
|
if (ftd->instance == 0) {
|
|
|
|
|
StatsSetUI64(th_v, ftd->cnt.memcap_pressure, mp);
|
|
|
|
|
StatsCounterMaxUpdateI64(&th_v->stats, ftd->cnt.memcap_pressure_max, (int64_t)mp);
|
|
|
|
|
}
|
|
|
|
|
GetWorkUnitSizing(rows, mp, false, &sleep_per_wu, &rows_per_wu, &rows_sec);
|
|
|
|
|
StatsSetUI64(th_v, ftd->cnt.flow_mgr_rows_sec, rows_sec);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
|
|
|
|
|
TmThreadsSetFlag(th_v, THV_RUNNING);
|
|
|
|
|
/* don't start our activities until time is setup */
|
|
|
|
|
while (!TimeModeIsReady()) {
|
|
|
|
|
if (suricata_ctl_flags != 0)
|
|
|
|
|
return TM_ECODE_OK;
|
|
|
|
|
SleepUsec(10);
|
|
|
|
|
}
|
|
|
|
|
bool run = TmThreadsWaitForUnpause(th_v);
|
|
|
|
|
|
|
|
|
|
while (run) {
|
|
|
|
|
bool emerg = ((SC_ATOMIC_GET(flow_flags) & FLOW_EMERGENCY) != 0);
|
|
|
|
|
|
|
|
|
|
/* Get the time: real time in live mode, or a min() of the
|
|
|
|
|
* "active" threads in offline mode. See TmThreadsGetMinimalTimestamp */
|
|
|
|
|
SCTime_t ts = TimeGet();
|
|
|
|
|
|
|
|
|
|
SCLogDebug("ts %" PRIdMAX "", (intmax_t)SCTIME_SECS(ts));
|
|
|
|
|
uint64_t ts_ms = SCTIME_MSECS(ts);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
const bool emerge_p = (emerg && !prev_emerg);
|
|
|
|
|
if (emerge_p) {
|
|
|
|
|
next_run_ms = 0;
|
|
|
|
|
prev_emerg = true;
|
|
|
|
|
SCLogNotice("Flow emergency mode entered...");
|
|
|
|
|
StatsIncr(th_v, ftd->cnt.flow_emerg_mode_enter);
|
|
|
|
|
}
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
if (ts_ms >= next_run_ms) {
|
|
|
|
|
if (ftd->instance == 0) {
|
|
|
|
|
const uint32_t sq_len = FlowSpareGetPoolSize();
|
|
|
|
|
const uint32_t spare_perc = sq_len * 100 / MAX(flow_config.prealloc, 1);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
/* see if we still have enough spare flows */
|
|
|
|
|
if (spare_perc < 90 || spare_perc > 110) {
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
FlowSparePoolUpdate(sq_len);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* try to time out flows */
|
|
|
|
|
// clang-format off
|
|
|
|
|
FlowTimeoutCounters counters = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, };
|
|
|
|
|
// clang-format on
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
|
|
|
|
|
if (emerg) {
|
|
|
|
|
/* in emergency mode, do a full pass of the hash table */
|
|
|
|
|
FlowTimeoutHash(&ftd->timeout, ts, ftd->min, ftd->max, &counters);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
StatsIncr(th_v, ftd->cnt.flow_mgr_full_pass);
|
|
|
|
|
} else {
|
|
|
|
|
SCLogDebug("hash %u:%u slice starting at %u with %u rows", ftd->min, ftd->max, pos,
|
|
|
|
|
rows_per_wu);
|
|
|
|
|
|
|
|
|
|
const uint32_t ppos = pos;
|
|
|
|
|
FlowTimeoutHashInChunks(&ftd->timeout, ts, ftd->min, ftd->max, &counters,
|
|
|
|
|
rows_per_wu, &pos, ftd->instance);
|
|
|
|
|
if (ppos > pos) {
|
|
|
|
|
StatsIncr(th_v, ftd->cnt.flow_mgr_full_pass);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
const uint32_t spare_pool_len = FlowSpareGetPoolSize();
|
|
|
|
|
StatsSetUI64(th_v, ftd->cnt.flow_mgr_spare, (uint64_t)spare_pool_len);
|
|
|
|
|
|
|
|
|
|
FlowCountersUpdate(th_v, ftd, &counters);
|
|
|
|
|
|
|
|
|
|
if (emerg) {
|
|
|
|
|
SCLogDebug("flow_sparse_q.len = %" PRIu32 " prealloc: %" PRIu32
|
|
|
|
|
"flow_spare_q status: %" PRIu32 "%% flows at the queue",
|
|
|
|
|
spare_pool_len, flow_config.prealloc,
|
|
|
|
|
spare_pool_len * 100 / MAX(flow_config.prealloc, 1));
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
|
|
|
|
|
/* only if we have pruned this "emergency_recovery" percentage
|
|
|
|
|
* of flows, we will unset the emergency bit */
|
|
|
|
|
if ((spare_pool_len * 100 / MAX(flow_config.prealloc, 1)) >
|
|
|
|
|
flow_config.emergency_recovery) {
|
|
|
|
|
emerg_over_cnt++;
|
|
|
|
|
} else {
|
|
|
|
|
emerg_over_cnt = 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (emerg_over_cnt >= 30) {
|
|
|
|
|
SC_ATOMIC_AND(flow_flags, ~FLOW_EMERGENCY);
|
|
|
|
|
FlowTimeoutsReset();
|
|
|
|
|
|
|
|
|
|
emerg = false;
|
|
|
|
|
prev_emerg = false;
|
|
|
|
|
emerg_over_cnt = 0;
|
|
|
|
|
SCLogNotice("Flow emergency mode over, back to normal... unsetting"
|
|
|
|
|
" FLOW_EMERGENCY bit (ts.tv_sec: %" PRIuMAX ", "
|
|
|
|
|
"ts.tv_usec:%" PRIuMAX ") flow_spare_q status(): %" PRIu32
|
|
|
|
|
"%% flows at the queue",
|
|
|
|
|
(uintmax_t)SCTIME_SECS(ts), (uintmax_t)SCTIME_USECS(ts),
|
|
|
|
|
spare_pool_len * 100 / MAX(flow_config.prealloc, 1));
|
|
|
|
|
|
|
|
|
|
StatsIncr(th_v, ftd->cnt.flow_emerg_mode_over);
|
|
|
|
|
}
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* update work units */
|
|
|
|
|
const uint32_t pmp = mp;
|
|
|
|
|
mp = MemcapsGetPressure() * 100;
|
|
|
|
|
if (ftd->instance == 0) {
|
|
|
|
|
StatsSetUI64(th_v, ftd->cnt.memcap_pressure, mp);
|
|
|
|
|
StatsCounterMaxUpdateI64(&th_v->stats, ftd->cnt.memcap_pressure_max, (int64_t)mp);
|
|
|
|
|
}
|
|
|
|
|
GetWorkUnitSizing(rows, mp, emerg, &sleep_per_wu, &rows_per_wu, &rows_sec);
|
|
|
|
|
if (pmp != mp) {
|
|
|
|
|
StatsSetUI64(th_v, ftd->cnt.flow_mgr_rows_sec, rows_sec);
|
|
|
|
|
}
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
|
|
|
|
|
next_run_ms = ts_ms + sleep_per_wu;
|
|
|
|
|
}
|
|
|
|
|
if (other_last_sec == 0 || other_last_sec < (uint32_t)SCTIME_SECS(ts)) {
|
|
|
|
|
if (ftd->instance == 0) {
|
|
|
|
|
StatsSetUI64(th_v, ftd->counter_defrag_memuse, DefragTrackerGetMemcap());
|
|
|
|
|
uint32_t defrag_cnt = DefragTimeoutHash(ts);
|
|
|
|
|
if (defrag_cnt) {
|
|
|
|
|
StatsCounterAddI64(
|
|
|
|
|
&th_v->stats, ftd->counter_defrag_timeout, (int64_t)defrag_cnt);
|
|
|
|
|
}
|
|
|
|
|
HostTimeoutHash(ts);
|
|
|
|
|
IPPairTimeoutHash(ts);
|
|
|
|
|
HttpRangeContainersTimeoutHash(ts);
|
|
|
|
|
ThresholdsExpire(ts);
|
|
|
|
|
other_last_sec = (uint32_t)SCTIME_SECS(ts);
|
|
|
|
|
}
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (TmThreadsCheckFlag(th_v, THV_KILL)) {
|
|
|
|
|
StatsSyncCounters(&th_v->stats);
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (emerg || !time_is_live) {
|
|
|
|
|
SleepUsec(250);
|
|
|
|
|
} else {
|
|
|
|
|
struct timeval cond_tv;
|
|
|
|
|
gettimeofday(&cond_tv, NULL);
|
|
|
|
|
struct timeval add_tv;
|
|
|
|
|
add_tv.tv_sec = sleep_per_wu / 1000;
|
|
|
|
|
add_tv.tv_usec = (sleep_per_wu % 1000) * 1000;
|
|
|
|
|
timeradd(&cond_tv, &add_tv, &cond_tv);
|
|
|
|
|
|
|
|
|
|
struct timespec cond_time = FROM_TIMEVAL(cond_tv);
|
|
|
|
|
SCCtrlMutexLock(&flow_manager_ctrl_mutex);
|
|
|
|
|
while (1) {
|
|
|
|
|
if (SC_ATOMIC_GET(flow_flags) & FLOW_EMERGENCY) {
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
int rc = SCCtrlCondTimedwait(
|
|
|
|
|
&flow_manager_ctrl_cond, &flow_manager_ctrl_mutex, &cond_time);
|
|
|
|
|
if (rc == ETIMEDOUT || rc < 0) {
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
SCCtrlMutexUnlock(&flow_manager_ctrl_mutex);
|
|
|
|
|
}
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
|
|
|
|
|
SCLogDebug("woke up... %s", SC_ATOMIC_GET(flow_flags) & FLOW_EMERGENCY ? "emergency":"");
|
|
|
|
|
|
|
|
|
|
StatsSyncCountersIfSignalled(&th_v->stats);
|
|
|
|
|
}
|
|
|
|
|
return TM_ECODE_OK;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/** \brief spawn the flow manager thread */
|
|
|
|
|
void FlowManagerThreadSpawn(void)
|
|
|
|
|
{
|
|
|
|
|
intmax_t setting = 1;
|
|
|
|
|
(void)SCConfGetInt("flow.managers", &setting);
|
|
|
|
|
|
|
|
|
|
if (setting < 1 || setting > 1024) {
|
|
|
|
|
FatalError("invalid flow.managers setting %" PRIdMAX, setting);
|
|
|
|
|
}
|
|
|
|
|
flowmgr_number = (uint32_t)setting;
|
|
|
|
|
|
|
|
|
|
SCLogConfig("using %u flow manager threads", flowmgr_number);
|
|
|
|
|
StatsRegisterGlobalCounter("flow.memuse", FlowGetMemuse);
|
|
|
|
|
|
|
|
|
|
for (uint32_t u = 0; u < flowmgr_number; u++) {
|
|
|
|
|
char name[TM_THREAD_NAME_MAX];
|
|
|
|
|
snprintf(name, sizeof(name), "%s#%02u", thread_name_flow_mgr, u+1);
|
|
|
|
|
|
|
|
|
|
ThreadVars *tv_flowmgr = TmThreadCreateMgmtThreadByName(name,
|
|
|
|
|
"FlowManager", 0);
|
|
|
|
|
BUG_ON(tv_flowmgr == NULL);
|
|
|
|
|
|
|
|
|
|
if (tv_flowmgr == NULL) {
|
|
|
|
|
FatalError("flow manager thread creation failed");
|
|
|
|
|
}
|
|
|
|
|
if (TmThreadSpawn(tv_flowmgr) != TM_ECODE_OK) {
|
|
|
|
|
FatalError("flow manager thread spawn failed");
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
typedef struct FlowRecyclerThreadData_ {
|
|
|
|
|
void *output_thread_data;
|
|
|
|
|
|
|
|
|
|
StatsCounterId counter_flows;
|
|
|
|
|
StatsCounterAvgId counter_queue_avg;
|
|
|
|
|
StatsCounterMaxId counter_queue_max;
|
|
|
|
|
|
|
|
|
|
StatsCounterId counter_flow_active;
|
|
|
|
|
StatsCounterId counter_tcp_active_sessions;
|
|
|
|
|
FlowEndCounters fec;
|
|
|
|
|
} FlowRecyclerThreadData;
|
|
|
|
|
|
|
|
|
|
static TmEcode FlowRecyclerThreadInit(ThreadVars *t, const void *initdata, void **data)
|
|
|
|
|
{
|
|
|
|
|
FlowRecyclerThreadData *ftd = SCCalloc(1, sizeof(FlowRecyclerThreadData));
|
|
|
|
|
if (ftd == NULL)
|
|
|
|
|
return TM_ECODE_FAILED;
|
|
|
|
|
if (OutputFlowLogThreadInit(t, &ftd->output_thread_data) != TM_ECODE_OK) {
|
|
|
|
|
SCLogError("initializing flow log API for thread failed");
|
|
|
|
|
SCFree(ftd);
|
|
|
|
|
return TM_ECODE_FAILED;
|
|
|
|
|
}
|
|
|
|
|
SCLogDebug("output_thread_data %p", ftd->output_thread_data);
|
|
|
|
|
|
|
|
|
|
ftd->counter_flows = StatsRegisterCounter("flow.recycler.recycled", t);
|
|
|
|
|
ftd->counter_queue_avg = StatsRegisterAvgCounter("flow.recycler.queue_avg", t);
|
|
|
|
|
ftd->counter_queue_max = StatsRegisterMaxCounter("flow.recycler.queue_max", t);
|
|
|
|
|
|
|
|
|
|
ftd->counter_flow_active = StatsRegisterCounter("flow.active", t);
|
|
|
|
|
ftd->counter_tcp_active_sessions = StatsRegisterCounter("tcp.active_sessions", t);
|
|
|
|
|
|
|
|
|
|
FlowEndCountersRegister(t, &ftd->fec);
|
|
|
|
|
|
|
|
|
|
*data = ftd;
|
|
|
|
|
return TM_ECODE_OK;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static TmEcode FlowRecyclerThreadDeinit(ThreadVars *t, void *data)
|
|
|
|
|
{
|
|
|
|
|
StreamTcpThreadCacheCleanup();
|
|
|
|
|
|
|
|
|
|
FlowRecyclerThreadData *ftd = (FlowRecyclerThreadData *)data;
|
|
|
|
|
if (ftd->output_thread_data != NULL)
|
|
|
|
|
OutputFlowLogThreadDeinit(t, ftd->output_thread_data);
|
|
|
|
|
|
|
|
|
|
SCFree(data);
|
|
|
|
|
return TM_ECODE_OK;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void Recycler(ThreadVars *tv, FlowRecyclerThreadData *ftd, Flow *f)
|
|
|
|
|
{
|
|
|
|
|
FLOWLOCK_WRLOCK(f);
|
|
|
|
|
|
|
|
|
|
(void)OutputFlowLog(tv, ftd->output_thread_data, f);
|
|
|
|
|
|
|
|
|
|
FlowEndCountersUpdate(tv, &ftd->fec, f);
|
|
|
|
|
if (f->proto == IPPROTO_TCP && f->protoctx != NULL) {
|
|
|
|
|
StatsDecr(tv, ftd->counter_tcp_active_sessions);
|
|
|
|
|
}
|
|
|
|
|
StatsDecr(tv, ftd->counter_flow_active);
|
|
|
|
|
SCFlowRunFinishCallbacks(tv, f);
|
|
|
|
|
FlowClearMemory(f, f->protomap);
|
|
|
|
|
FLOWLOCK_UNLOCK(f);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/** \brief Thread that manages timed out flows.
|
|
|
|
|
*
|
|
|
|
|
* \param td ThreadVars cast to void ptr
|
|
|
|
|
*/
|
|
|
|
|
static TmEcode FlowRecycler(ThreadVars *th_v, void *thread_data)
|
|
|
|
|
{
|
|
|
|
|
FlowRecyclerThreadData *ftd = (FlowRecyclerThreadData *)thread_data;
|
|
|
|
|
BUG_ON(ftd == NULL);
|
|
|
|
|
const bool time_is_live = TimeModeIsLive();
|
|
|
|
|
uint64_t recycled_cnt = 0;
|
|
|
|
|
FlowQueuePrivate ret_queue = { NULL, NULL, 0 };
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
|
|
|
|
|
TmThreadsSetFlag(th_v, THV_RUNNING);
|
|
|
|
|
bool run = TmThreadsWaitForUnpause(th_v);
|
|
|
|
|
|
|
|
|
|
while (run) {
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
SC_ATOMIC_ADD(flowrec_busy,1);
|
|
|
|
|
FlowQueuePrivate list = FlowQueueExtractPrivate(&flow_recycle_q);
|
|
|
|
|
|
|
|
|
|
StatsCounterAvgAddI64(&th_v->stats, ftd->counter_queue_avg, (int64_t)list.len);
|
|
|
|
|
StatsCounterMaxUpdateI64(&th_v->stats, ftd->counter_queue_max, (int64_t)list.len);
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
const int bail = (TmThreadsCheckFlag(th_v, THV_KILL));
|
|
|
|
|
|
|
|
|
|
/* Get the time */
|
|
|
|
|
SCLogDebug("ts %" PRIdMAX "", (intmax_t)SCTIME_SECS(TimeGet()));
|
|
|
|
|
|
|
|
|
|
int64_t cnt = 0;
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
Flow *f;
|
|
|
|
|
while ((f = FlowQueuePrivateGetFromTop(&list)) != NULL) {
|
|
|
|
|
Recycler(th_v, ftd, f);
|
|
|
|
|
cnt++;
|
|
|
|
|
|
|
|
|
|
/* for every full sized block, add it to the spare pool */
|
|
|
|
|
FlowQueuePrivateAppendFlow(&ret_queue, f);
|
|
|
|
|
if (ret_queue.len == FLOW_SPARE_POOL_BLOCK_SIZE) {
|
|
|
|
|
FlowSparePoolReturnFlows(&ret_queue);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
if (ret_queue.len > 0) {
|
|
|
|
|
FlowSparePoolReturnFlows(&ret_queue);
|
|
|
|
|
}
|
|
|
|
|
if (cnt > 0) {
|
|
|
|
|
recycled_cnt += cnt;
|
|
|
|
|
StatsCounterAddI64(&th_v->stats, ftd->counter_flows, cnt);
|
|
|
|
|
}
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
SC_ATOMIC_SUB(flowrec_busy,1);
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
if (bail) {
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
const bool emerg = (SC_ATOMIC_GET(flow_flags) & FLOW_EMERGENCY);
|
|
|
|
|
if (emerg || !time_is_live) {
|
|
|
|
|
SleepUsec(250);
|
|
|
|
|
} else {
|
|
|
|
|
struct timeval cond_tv;
|
|
|
|
|
gettimeofday(&cond_tv, NULL);
|
|
|
|
|
cond_tv.tv_sec += 1;
|
|
|
|
|
struct timespec cond_time = FROM_TIMEVAL(cond_tv);
|
|
|
|
|
SCCtrlMutexLock(&flow_recycler_ctrl_mutex);
|
|
|
|
|
while (1) {
|
|
|
|
|
if (SC_ATOMIC_GET(flow_flags) & FLOW_EMERGENCY) {
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
if (SC_ATOMIC_GET(flow_recycle_q.non_empty)) {
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
int rc = SCCtrlCondTimedwait(
|
|
|
|
|
&flow_recycler_ctrl_cond, &flow_recycler_ctrl_mutex, &cond_time);
|
|
|
|
|
if (rc == ETIMEDOUT || rc < 0) {
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
SCCtrlMutexUnlock(&flow_recycler_ctrl_mutex);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
SCLogDebug("woke up...");
|
|
|
|
|
|
|
|
|
|
StatsSyncCountersIfSignalled(&th_v->stats);
|
|
|
|
|
}
|
|
|
|
|
StatsSyncCounters(&th_v->stats);
|
|
|
|
|
SCLogPerf("%"PRIu64" flows processed", recycled_cnt);
|
|
|
|
|
return TM_ECODE_OK;
|
|
|
|
|
}
|
|
|
|
|
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
static bool FlowRecyclerReadyToShutdown(void)
|
|
|
|
|
{
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
if (SC_ATOMIC_GET(flowrec_busy) != 0) {
|
|
|
|
|
return false;
|
|
|
|
|
}
|
|
|
|
|
uint32_t len = 0;
|
|
|
|
|
FQLOCK_LOCK(&flow_recycle_q);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
len = flow_recycle_q.qlen;
|
|
|
|
|
FQLOCK_UNLOCK(&flow_recycle_q);
|
|
|
|
|
|
|
|
|
|
return ((len == 0));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/** \brief spawn the flow recycler thread */
|
|
|
|
|
void FlowRecyclerThreadSpawn(void)
|
|
|
|
|
{
|
|
|
|
|
intmax_t setting = 1;
|
|
|
|
|
(void)SCConfGetInt("flow.recyclers", &setting);
|
|
|
|
|
|
|
|
|
|
if (setting < 1 || setting > 1024) {
|
|
|
|
|
FatalError("invalid flow.recyclers setting %" PRIdMAX, setting);
|
|
|
|
|
}
|
|
|
|
|
flowrec_number = (uint32_t)setting;
|
|
|
|
|
|
|
|
|
|
SCLogConfig("using %u flow recycler threads", flowrec_number);
|
|
|
|
|
|
|
|
|
|
for (uint32_t u = 0; u < flowrec_number; u++) {
|
|
|
|
|
char name[TM_THREAD_NAME_MAX];
|
|
|
|
|
snprintf(name, sizeof(name), "%s#%02u", thread_name_flow_rec, u+1);
|
|
|
|
|
|
|
|
|
|
ThreadVars *tv_flowrec = TmThreadCreateMgmtThreadByName(name,
|
|
|
|
|
"FlowRecycler", 0);
|
|
|
|
|
|
|
|
|
|
if (tv_flowrec == NULL) {
|
|
|
|
|
FatalError("flow recycler thread creation failed");
|
|
|
|
|
}
|
|
|
|
|
if (TmThreadSpawn(tv_flowrec) != TM_ECODE_OK) {
|
|
|
|
|
FatalError("flow recycler thread spawn failed");
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* \brief Used to disable flow recycler thread(s).
|
|
|
|
|
*
|
|
|
|
|
* \note this should only be called when the flow manager is already gone
|
|
|
|
|
*
|
|
|
|
|
* \todo Kinda hackish since it uses the tv name to identify flow recycler
|
|
|
|
|
* thread. We need an all weather identification scheme.
|
|
|
|
|
*/
|
|
|
|
|
void FlowDisableFlowRecyclerThread(void)
|
|
|
|
|
{
|
|
|
|
|
/* move all flows still in the hash to the recycler queue */
|
|
|
|
|
#ifndef DEBUG
|
|
|
|
|
(void)FlowCleanupHash();
|
|
|
|
|
#else
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
uint32_t flows = FlowCleanupHash();
|
|
|
|
|
SCLogDebug("flows to progress: %u", flows);
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
/* make sure all flows are processed */
|
|
|
|
|
do {
|
|
|
|
|
FlowWakeupFlowRecyclerThread();
|
|
|
|
|
SleepUsec(10);
|
|
|
|
|
} while (!FlowRecyclerReadyToShutdown());
|
|
|
|
|
|
|
|
|
|
SCMutexLock(&tv_root_lock);
|
|
|
|
|
/* flow recycler thread(s) is/are a part of mgmt threads */
|
|
|
|
|
for (ThreadVars *tv = tv_root[TVT_MGMT]; tv != NULL; tv = tv->next) {
|
|
|
|
|
if (strncasecmp(tv->name, thread_name_flow_rec,
|
|
|
|
|
strlen(thread_name_flow_rec)) == 0)
|
|
|
|
|
{
|
|
|
|
|
TmThreadsSetFlag(tv, THV_KILL);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
SCMutexUnlock(&tv_root_lock);
|
|
|
|
|
|
|
|
|
|
struct timeval start_ts;
|
|
|
|
|
struct timeval cur_ts;
|
|
|
|
|
gettimeofday(&start_ts, NULL);
|
|
|
|
|
|
|
|
|
|
again:
|
|
|
|
|
gettimeofday(&cur_ts, NULL);
|
|
|
|
|
if ((cur_ts.tv_sec - start_ts.tv_sec) > 60) {
|
|
|
|
|
FatalError("unable to get all flow recycler "
|
|
|
|
|
"threads to shutdown in time");
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
SCMutexLock(&tv_root_lock);
|
|
|
|
|
for (ThreadVars *tv = tv_root[TVT_MGMT]; tv != NULL; tv = tv->next) {
|
|
|
|
|
if (strncasecmp(tv->name, thread_name_flow_rec,
|
|
|
|
|
strlen(thread_name_flow_rec)) == 0)
|
|
|
|
|
{
|
|
|
|
|
if (!TmThreadsCheckFlag(tv, THV_RUNNING_DONE)) {
|
|
|
|
|
SCMutexUnlock(&tv_root_lock);
|
|
|
|
|
FlowWakeupFlowRecyclerThread();
|
|
|
|
|
/* sleep outside lock */
|
|
|
|
|
SleepMsec(1);
|
|
|
|
|
goto again;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
SCMutexUnlock(&tv_root_lock);
|
|
|
|
|
|
|
|
|
|
/* reset count, so we can kill and respawn (unix socket) */
|
|
|
|
|
SC_ATOMIC_SET(flowrec_cnt, 0);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void TmModuleFlowManagerRegister (void)
|
|
|
|
|
{
|
|
|
|
|
tmm_modules[TMM_FLOWMANAGER].name = "FlowManager";
|
|
|
|
|
tmm_modules[TMM_FLOWMANAGER].ThreadInit = FlowManagerThreadInit;
|
|
|
|
|
tmm_modules[TMM_FLOWMANAGER].ThreadDeinit = FlowManagerThreadDeinit;
|
|
|
|
|
tmm_modules[TMM_FLOWMANAGER].Management = FlowManager;
|
|
|
|
|
tmm_modules[TMM_FLOWMANAGER].cap_flags = 0;
|
|
|
|
|
tmm_modules[TMM_FLOWMANAGER].flags = TM_FLAG_MANAGEMENT_TM;
|
|
|
|
|
SCLogDebug("%s registered", tmm_modules[TMM_FLOWMANAGER].name);
|
|
|
|
|
|
|
|
|
|
SC_ATOMIC_INIT(flowmgr_cnt);
|
|
|
|
|
SC_ATOMIC_INITPTR(flow_timeouts);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void TmModuleFlowRecyclerRegister (void)
|
|
|
|
|
{
|
|
|
|
|
tmm_modules[TMM_FLOWRECYCLER].name = "FlowRecycler";
|
|
|
|
|
tmm_modules[TMM_FLOWRECYCLER].ThreadInit = FlowRecyclerThreadInit;
|
|
|
|
|
tmm_modules[TMM_FLOWRECYCLER].ThreadDeinit = FlowRecyclerThreadDeinit;
|
|
|
|
|
tmm_modules[TMM_FLOWRECYCLER].Management = FlowRecycler;
|
|
|
|
|
tmm_modules[TMM_FLOWRECYCLER].cap_flags = 0;
|
|
|
|
|
tmm_modules[TMM_FLOWRECYCLER].flags = TM_FLAG_MANAGEMENT_TM;
|
|
|
|
|
SCLogDebug("%s registered", tmm_modules[TMM_FLOWRECYCLER].name);
|
|
|
|
|
|
|
|
|
|
SC_ATOMIC_INIT(flowrec_cnt);
|
flow: redesign of flow timeout handling
Goals:
- reduce locking
- take advantage of 'hot' caches
- better locality
Locking reduction
New flow spare pool. The global pool is implmented as a list of blocks,
where each block has a 100 spare flows. Worker threads fetch a block at
a time, storing the block in the local thread storage.
Flow Recycler now returns flows to the pool is blocks as well.
Flow Recycler fetches all flows to be processed in one step instead of
one at a time.
Cache 'hot'ness
Worker threads now check the timeout of flows they evaluate during lookup.
The worker will have to read the flow into cache anyway, so the added
overhead of checking the timeout value is minimal. When a flow is considered
timed out, one of 2 things happens:
- if the flow is 'owned' by the thread it is handled locally. Handling means
checking if the flow needs 'timeout' work.
- otherwise, the flow is added to a special 'evicted' list in the flow
bucket where it will be picked up by the flow manager.
Flow Manager timing
By default the flow manager now tries to do passes of the flow hash in
smaller steps, where the goal is to do full pass in 8 x the lowest timeout
value it has to enforce. So if the lowest timeout value is 30s, a full pass
will take 4 minutes. The goal here is to reduce locking overhead and not
get in the way of the workers.
In emergency mode each pass is full, and lower timeouts are used.
Timing of the flow manager is also no longer relying on pthread condition
variables, as these generally cause waking up much quicker than the desired
timout. Instead a simple (u)sleep loop is used.
Both changes reduce the number of hash passes a lot.
Emergency behavior
In emergency mode there a number of changes to the workers. In this scenario
the flow memcap is fully used up and it is unavoidable that some flows won't
be tracked.
1. flow spare pool fetches are reduced to once a second. This avoids locking
overhead, while the chance of success was very low.
2. getting an active flow directly from the hash skips flows that had very
recent activity to avoid the scenario where all flows get only into the
NEW state before getting reused. Rather allow some to have a chance of
completing.
3. TCP packets that are not SYN packets will not get a used flow, unless
stream.midstream is enabled. The goal here is again to avoid evicting
active flows unnecessarily.
Better Localily
Flow Manager injects flows into the worker threads now, instead of one or
two packets. Advantage of this is that the worker threads can get packets
from their local packet pools, avoiding constant overhead of packets returning
to 'foreign' pools.
Counters
A lot of flow counters have been added and some have been renamed.
Overall the worker threads increment 'flow.wrk.*' counters, while the flow
manager increments 'flow.mgr.*'.
Additionally, none of the counters are snapshots anymore, they all increment
over time. The flow.memuse and flow.spare counters are exceptions.
Misc
FlowQueue has been split into a FlowQueuePrivate (unlocked) and FlowQueue.
Flow no longer has 'prev' pointers and used a unified 'next' pointer for
both hash and queue use.
7 years ago
|
|
|
SC_ATOMIC_INIT(flowrec_busy);
|
|
|
|
|
}
|