When there are many threads and/or the packet pool (max-pending-packets) is
small, a potential dead lock exists between the packet pool return pool
logic and the capture threads. The autofp workers together can have all the
packets in their return pools, while the capture thread(s) are waiting at an
empty pool. A race between the worker threads and the capture thread, where
the latter signals the former, is lost by the capture thread. Now everyone
is waiting.
To avoid this scenario, this patch makes the previously hardcoded 'return
pool' threshold dynamic based on the number of threads and the packet pool
size.
It sets the threshold to the max pending packets value, divided by the number
of lister threads. The max value hasn't changed. Normally, in the autofp
runmode these are the stream/detect/log worker threads.
The max_pending_return_packets value needs to stay below the packet pool size
of the 'producers' (normally pkt capture threads but also flow timeout
injection) to avoid the deadlock.
As it's quite impossible at this time to learn how many threads will be
created before starting the runmodes, and thus spawning the threads and
already initializing the packet pools, this code sets a global variable
after runmode setup, but before the threads are 'unpaused'.
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.
PacketPoolWait in autofp can wait for considerable time. Until now
it was essentially spinning, keeping the CPU 100% busy.
This patch introduces a condition to wait in such cases.
Atomically flag pool that consumer is waiting, so that we can sync
the pending pool right away instead of waiting for the
MAX_PENDING_RETURN_PACKETS limit.
Better handle the autofp case where one thread allocates the majority
of the packets and other threads free those packets.
Add a list of locally pending packets. The first packet freed goes on the
pending list, then subsequent freed packets for the same Packet Pool are
added to this list until it hits a fixed number of packets, then the
entire list of packets is pushed onto the pool's return stack. If a freed
packet is not for the pending pool, it is freed immediately to its pool's
return stack, as before.
For the autofp case, since there is only one Packet Pool doing all the
allocation, every other thread will keep a list of pending packets for
that pool.
For the worker run mode, most packets are allocated and freed locally. For
the case where packets are being returned to a remote pool, a pending list
will be kept for one of those other threads, all others are returned as before.
Which remote pool for which to keep a pending list is changed each time the
pending list is returned. Since the return pending pool is cleared when it is
freed, then next packet to be freed chooses the new pending pool.
Using a stack for free Packet storage causes recently freed Packets to be
reused quickly, while there is more likelihood of the data still being in
cache.
The new structure has a per-thread private stack for allocating Packets
which does not need any locking. Since Packets can be freed by any thread,
there is a second stack (return stack) for freeing packets by other threads.
The return stack is protected by a mutex. Packets are moved from the return
stack to the private stack when the private stack is empty.
Returning packets back to their "home" stack keeps the stacks from getting out
of balance.
The PacketPoolInit() function is now called by each thread that will be
allocating packets. Each thread allocates max_pending_packets, which is a
change from before, where that was the total number of packets across all
threads.
This commit allows handling Packets allocated by different methods.
The ReleaseData function pointer in the Packet structure is replaced
with ReleasePacket function pointer, which is then always called to
release the memory associated with a Packet.
Currently, the only usage of ReleaseData is in AF Packet. Previously
ReleaseData was only called when it was not NULL. To implement the
same functionality as before in AF Packet, a new function is defined
in AF Packet to first call the AFP specific ReleaseData function and
then releases the Packet structure.
Three new general functions are defined for releasing packets in the
default case:
1) PacketFree() - To release a packet alloced with SCMalloc()
2) PacketPoolReturnPacket() - For packets allocated from the Packet Pool.
Calls RECYCLE_PACKET(p)
3) PacketFreeOrRelease() - Calls PacketFree() or PacketPoolReturnPacket()
based on the PKT_ALLOC flag.
Having these functions removes the need to check the PKT_ALLOC flag
when releasing a packet in most cases, since the ReleasePacket
function encodes how the Packet was allocated. The PKT_ALLOC flag is
still set and is needed when AF Packet releases a packet, since it
replaces the ReleasePacket function pointer with its own function and
then calls PacketFreeOfRelease(), which uses the PKT_ALLOC flag.
Victor Julien
Ken Steele
William Metcalf
Anoop Saldanha