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.
Previously each 'TmSlot' had it's own packet queue that was passed
to the registered SlotFunc as an argument. This was used mostly for
tunnel packets by the decoders and by defrag.
This patch removes that in favor of a single queue in the ThreadVars:
decode_pq. This is the non-locked version of the queue as this is
only a temporary store for handling packets within a thread.
This patch removes the PacketQueue pointer argument from the API.
The new queue can be accessed directly through the ThreadVars
pointer.
Shrink ThreadVars by removing the queue handler names that are only
used at shutdown. Since this is not performance critical, we can use
the id's to look up the queue handler.
Use a single packet queue per thread for flow timeout packet
injection. The per slot queue was unused except for this use
case. Having a single queue makes the logic and implementation
simpler.
In case of 'autofp', the per thread packet queue will actually
use the threads input queue. For workers/single a dedicated
queue will be set up.
Rename TmThreadsSlotHandlePostPQs to TmThreadsHandleInjectedPackets
to reflect the changed logic.
Set flags by default:
-Wmissing-prototypes
-Wmissing-declarations
-Wstrict-prototypes
-Wwrite-strings
-Wcast-align
-Wbad-function-cast
-Wformat-security
-Wno-format-nonliteral
-Wmissing-format-attribute
-funsigned-char
Fix minor compiler warnings for these new flags on gcc and clang.
In preparation of the introduction of more general purpose prefilter
engines, rename PatternMatcherQueue to PrefilterRuleStore. The new
engines will fill this structure a similar way to the current mpm
prefilters.
Thread restarts never worked well and the rest of the engine was
never really expecting errors to lead to thread restarts. Either
and error is recoverable in the thread, or not at all.
So this patch removes the functionality completely.
Capture methods that are non blocking will still not generate packets
that go through the system if there is no traffic. Some maintenance
tasks, like rule reloads rely on packets to complete.
This patch introduces a new thread flag, THV_CAPTURE_INJECT_PKT, that
instructs the capture thread to create a fake packet.
The capture implementations can call the TmThreadsCaptureInjectPacket
utility function either with the packet they already got from the pool
or without a packet. In this case the util func will get it's own
packet.
Implementations for pcap, AF_PACKET and PF_RING.
Update pktacq loop to process flow timeouts in a running engine.
Add a new step to the shutdown phase of packet acquisition loop
threads (pktacqloop).
The shutdown code lets the pktacqloop break out of it's packet
acquisition loop. The thread then enters a flow timeout loop, where
it processes packets from it's tv->stream_pq queue until it's
empty _and_ the KILL flag is set.
Make sure receive threads are done before moving on to flow hash
cleanup (recycle all). Without this the flow recycler could start
it's unconditional hash clean up while detect threads are still
running on the flows.
Update unix socket to match live modes.
It is possible to have a non-contiguous CPU set, which was not being
handled correctly on the TILE architecture.
Added a "rank" field in the ThreadVar to store the worker's rank separately
from the cpu for this case.
On Tile, replace pthread_mutex_locks with queued spin locks (ticket
locks) for dataplane processing code. This is safe when running on
dataplane cores with one thread per core. The condition variables are
no-ops when the thread is spinning anyway.
For control plane threads, unix-manager, stats-logs, thread startup,
use pthread_mutex_locks. For these locks replaced SCMutex with SCCtrlMutex
and SCCond with SCCtrlCond.
Tilera's GCC supports the GCC __sync_ intrinsics.
Increase the size of some atomic variables for better performance on
Tile. The Tile-Gx architecture has native support for 32-bit and
64-bit atomic operations, but not 8-bit and 16-bit, which are emulated
using 32-bit atomics, so changing some 16-bit and 8-bit atomic into
ints improves performance.
Increasing the size of the atomic variables modified in this change
does not increase the total size of the structures in which they
reside because of existing padding requirements. The one case that
would increase the size of the structure (Flow_) was confitionalized
to only change the size on Tile.
This patch features the implementation of affinity related
changes in tm-threads. In place code has been used but some
refactoring has been done to avoid code duplication.
Signed-off-by: Eric Leblond <eric@regit.org>
This small patch add inclusion of util-affinity.h in the
files that will have to use affinity related features.
Signed-off-by: Eric Leblond <eric@regit.org>
Philippe Antoine
Victor Julien
Sascha Steinbiss
Zachary Rasmor
Ken Steele
Anoop Saldanha
Anoop Saldanha
Eric Leblond
Pablo Rincon