Work towards making `suricata-common.h` only introduce system headers
and other things that are independent of complex internal Suricata
data structures.
Update files to compile after this.
Remove special DPDK handling for strlcpy and strlcat, as this caused
many compilation failures w/o including DPDK headers for all files.
Remove packet macros from decode.h and move them into their own file,
turn them into functions and rename them to match our function naming
policy.
Adds a framework for setting exception policies. These would be called
when the engine reaches some kind of exception condition, like hitting
a memcap or some traffic processing error.
The policy gives control over what should happen next: drop the packet,
drop the packet and flow, bypass, etc.
Implements the policy for:
stream: If stream session or reassembly memcaps are hit call the
memcap policy on the packet and flow.
flow: Apply policy when memcap is reached and no flow could be
freed up.
defrag: Apply policy when no tracker could be picked up.
app-layer: Apply ppolicy if a parser reaches an error state.
All options default to 'ignore', which means the default behavior
is unchanged.
Adds commandline options: add simulation options for exceptions. These
are only exposed if compiled with `--enable-debug`.
Ticket: #5214.
Ticket: #5215.
Ticket: #5216.
Ticket: #5218.
Ticket: #5194.
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.
Kernel time is not available (and/or costly) on NIC such as
Netronome so we update the logic to detect dead flows based on a
lack of update of packets counters. This way, the XDP filter will
be usable by network card.
This patch also updates the ebpf code to support per CPU and
regular mapping. Netronome is not supporting it and the structure
is using atomic for counter so the cost of simultaneous update
is really low.
This patch also updates the xdp_filter to be able to select if the
flow table is per CPU on shared. Second option will be used for
hardward offload. To deactivate the per cpu hash, you need to set
USE_PERCPU_HASH to 0.
This patch also adds an new option to af-packet named no-percpu-hash
If this option is set to yes then the Flow bypassed manager thread
will use one CPU instead of the number of cores. By doing that
we are able to handle the case where USE_PERCPU_HASH is unset (so
hardware offload for Netronome).
This patch also remove aligment indications in the eBPF filter. This
was not really needed and it seems it is causing problem with
some recent version of LLVM toolchain.
Until now the flow manager would walk the entire flow hash table on an
interval. It would thus touch all flows, leading to a lot of memory
and cache pressure. In scenario's where the number of tracked flows run
into the hundreds on thousands, and the memory used can run into many
hundreds of megabytes or even gigabytes, this would lead to serious
performance degradation.
This patch introduces a new approach. A timestamp per flow bucket
(hash row) is maintained by the flow manager. It holds the timestamp
of the earliest possible timeout of a flow in the list. The hash walk
skips rows with timestamps beyond the current time.
As the timestamp depends on the flows in the hash row's list, and on
the 'state' of each flow in the list, any addition of a flow or
changing of a flow's state invalidates the timestamp. The flow manager
then has to walk the list again to set a new timestamp.
A utility function FlowUpdateState is introduced to change Flow states,
taking care of the bucket timestamp invalidation while at it.
Empty flow buckets use a special value so that we don't have to take
the flow bucket lock to find out the bucket is empty.
This patch also adds more performance counters:
flow_mgr.flows_checked | Total | 929
flow_mgr.flows_notimeout | Total | 391
flow_mgr.flows_timeout | Total | 538
flow_mgr.flows_removed | Total | 277
flow_mgr.flows_timeout_inuse | Total | 261
flow_mgr.rows_checked | Total | 1000000
flow_mgr.rows_skipped | Total | 998835
flow_mgr.rows_empty | Total | 290
flow_mgr.rows_maxlen | Total | 2
flow_mgr.flows_checked: number of flows checked for timeout in the
last pass
flow_mgr.flows_notimeout: number of flows out of flow_mgr.flows_checked
that didn't time out
flow_mgr.flows_timeout: number of out of flow_mgr.flows_checked that
did reach the time out
flow_mgr.flows_removed: number of flows out of flow_mgr.flows_timeout
that were really removed
flow_mgr.flows_timeout_inuse: number of flows out of flow_mgr.flows_timeout
that were still in use or needed work
flow_mgr.rows_checked: hash table rows checked
flow_mgr.rows_skipped: hash table rows skipped because non of the flows
would time out anyway
The counters below are only relating to rows that were not skipped.
flow_mgr.rows_empty: empty hash rows
flow_mgr.rows_maxlen: max number of flows per hash row. Best to keep low,
so increase hash-size if needed.
flow_mgr.rows_busy: row skipped because it was locked by another thread
Update Flow lookup functions to get a flow reference during lookup.
This reference is set under the FlowBucket lock.
This paves the way to not getting a flow lock during lookups.
In case of autofp (or more general, when flow and stream engine
run in different threads) the flow engine should not trigger a flow
reuse as this can lead to race conditions between the flow and the
stream engine.
In such cases, the flow engine can be far ahead of the stream engine
as packets are in a queue between the threads.
Observed:
Flow engine tags packet 10 as start of new flow. Flow is tagged as
'reused'.
Stream engine evaluates packet 5 which belongs to the old flow. It
rejects the flow as it's tagged 'reused'. Attaches packet 5 to the
new flow which is wrong.
Solution:
This patch connects the flow engines handling of reuse cases to
the runmode. It hooks into the RunmodeSetFlowStreamAsync() call to
notify the flow engine that it shouldn't handle the reuse.
Most flows are marked for clean up by the flow manager, which then
passes them to the recycler. The recycler logs and cleans up. However,
under resource stress conditions, the packet threads can recycle
existing flow directly. So here the recycler has no role to play, as
the flow is immediately used.
For this reason, the packet threads need to be able to invoke the
flow logger directly.
The flow logging thread ctx will stored in the DecodeThreadVars
stucture. Therefore, this patch makes the DecodeThreadVars an argument
to FlowHandlePacket.
By moving FlowReference() out of FlowGetFlowFromHash() and into the one
function that calls it, all the flow functions take const Packet * instead
of Packet *.
Major redesign of the flow engine. Remove the flow queues that turned
out to be major choke points when using many threads. Flow manager now
walks the hash table directly. Simplify the way we get a new flow in
case of emergency.
Eric Leblond
Victor Julien
Victor Julien
Eric Leblond
Ken Steele
William Metcalf
Pablo Rincon