The flow timeout mechanism called both from the flow manager at run time
and at shutdown creates pseudo packets. For this it has it's own packet
pool, which can be depleted if the timeout logic is faster than the packet
processing threads. In this case the flow timeout would enter a wait loop.
The problem however, is that this wait loop would happen while keeping a
flow locked. This could lead to a race condition when the packet thread(s)
are waiting for the lock that the flow manager has.
This patch introduces a new packet pool call 'PacketPoolWaitForN', meant
to make sure that the thread's packet pool has at least N available
packets. The flow timeout paths use this to make sure enough packets are
available *before* grabbing the flow lock. If there aren't enough packets
available yet, the wait happens before the lock as well.
This still means the wait can happen while the flow hash row is locked, so
we do make sure some more packets are available when entering that. But
perhaps in the future we need a more precise logic there as well.
Don't kill flow manager and recyclers before the rest of the threads. The
packet threads may still have packets from their pools. As the flow threads
would destroy their pools the packets would be lost.
This patch doesn't kill the threads, it just pulls them out of their run
loop and into a wait loop. The packet pools won't be cleared until all
threads are killed.
Wait for flow management threads to close before moving on to the
next steps in the shutdown process.
Don't destroy flow force reassembly packet pool too early. Worker
threads may still want to return packets to it.
This replaces the tcp.reused_ssn counter. The flow engine now
enforces the TCP flow reuse logic.
The counter is incremented only when the flow is timed out, so
after the "tcp closed" timeout expired for a flow.
Until this point, the flow manager would check for timed out flows
by walking the flow hash, locking first the hash row and then each
individual flow to get it's state and timestamp. To not be too
intrusive trylocks were used so that a busy flow wouldn't cause the
flow manager to wait for a long time while holding the hash row lock.
Building on the changes in handling of the flow state and lastts
fields, this patch changes the flow managers behavior.
It can now get a flows state atomically and the lastts can be safely
read while holding just the flow hash row lock. This allows the flow
manager to do the basic time out check much more cheaply:
1. it doesn't have to wait for getting a lock
2. it doesn't interupt the packet path
As a consequence the trylock is now also gone. A flow that returns
'true' on timeout is pretty much certainly not going to be busy so
we can safely lock it unconditionally. This also means the flow
manager now walks the entire row unconditionally and is guaranteed
to inspect each flow in the row.
To make sure the functions called before the flow lock don't
accidentally change the flow (which would require a lock) the args
to these flows are changed to const pointers.
A flow has 3 states: NEW, ESTABLISHED and CLOSED.
For all protocols except TCP, a flow is in state NEW as long as just one
side of the conversation has been seen. When both sides have been
observed the state is moved to ESTABLISHED.
TCP has a different logic, controlled by the stream engine. Here the TCP
state is leading.
Until now, when parts of the engine needed to know the flow state, it
would invoke a per protocol callback 'GetProtoState'. For TCP this would
return the state based on the TcpSession.
This patch changes this logic. It introduces an atomic variable in the
flow 'flow_state'. It defaults to NEW and is set to ESTABLISHED for non-
TCP protocols when we've seen both sides of the conversation.
For TCP, the state is updated from the TCP engine directly.
The goal is to allow for access to the state without holding the Flow's
main mutex lock. This will later allow the Flow Manager(s) to evaluate
the Flow w/o interupting it.
Use new management API to run the flow recycler.
Make number of threads configurable:
flow:
memcap: 64mb
hash-size: 65536
prealloc: 10000
emergency-recovery: 30
managers: 2
recyclers: 2
This sets up 2 flow recyclers.
Use new management API to run the flow manager.
Support multiple flow managers, where each of them works with it's
own part of the flow hash.
Make number of threads configurable:
flow:
memcap: 64mb
hash-size: 65536
prealloc: 10000
emergency-recovery: 30
managers: 2
This sets up 2 flow managers.
Handle misc tasks only in instance 1: Handle defrag hash timeout
handing, host hash timeout handling and flow spare queue updating
only from the first instance.
The flow end flags field is filled by the flow manager or the flow
hash (in case of forced timeout of a flow) to record the timeout
conditions in the flow:
- emergency mode
- state
- reason (timed out or forced)
Add logging to the flow logger.
Thread was killed by the generic TmThreadKillThreads instead of
the FlowKillFlowRecyclerThread. The latter wakes the thread up, so
that shutdown is quite a bit faster.
Flow timeout code keeps track of thread module running detect, and
fails (hard) if it doesn't find it.
This changeset retrieves the global g_detect_disabled and passes
it to the timeout handling code during setup.
Flow timeout code worked by luck when checking if a flow still needed
reassembly for app layer inspection or logging. It would check for a
part of raw reassembly (smsg list) to determine if detection was
needed. In this case it would also process app layer cleanup,
including logging.
Introduced AppLayerTransactionGetActive which returns the lowest tx_id
in a direction that still needs some work.
FlowForceReassemblyNeedReassmbly now uses it to determine if the
applayer still needs work.
Converted FlowForceReassemblyForHash to use the checking function
FlowForceReassemblyNeedReassmbly as well, so that checking if a flow
needs work is now unified.
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.
Victor Julien
Eric Leblond
Ken Steele
Anoop Saldanha
Ludovico Cavedon
Last G