This define is used to remove reference to capture bypass in case
no capture method implementing this is active.
This patch also introduces CAPTURE_OFFLOAD_MANAGER that is defined
if we need the flow bypass manager code.
If vlan.use-for-tracking is disabled, set the vlan_id fields to 0 when
hashing or comparing flows. This is done using a bitmask as suggested by
Victor Julien in IRC, in order to avoid adding more branches to this
code.
Currently, suricata does not fill in vlan_id fields if
vlan.use-for-tracking is disabled and instead leaves them at the default
0 value, so this commit makes no functional change. This change is in
preparation for future commits where the vlan_ids will be always filled
in.
Related to https://redmine.openinfosecfoundation.org/issues/3076
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.
On MinGW the result of ntohl needs to be casted to uint32_t and
the result of ntohs to uint16_t. To avoid doing this everywhere
add SCNtohl and SCNtohs macros.
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
Instead of a single big FlowProto array containing timeouts separately
for normal and emergency cases, plus the 'Free' pointer for the
protoctx, split up these arrays.
An array made of FlowProtoTimeout for just the normal timeouts and an
mirror of that for emergency timeouts are used through a pointer that
will be set at init and by swapped by the emergency logic. It's swapped
back when the emergency is over.
The free funcs are moved to their own array.
This simplifies the timeout lookup code and shrinks the data that is
commonly used.
Now that the flow lookup is done in the worker threads the flow
queue handlers running after the capture thread(s) no longer have
access to the flow. This limits the options of how flow balancing
can be done.
This patch removes all code that is now useless. The only 2 methods
that still make sense are 'hash' and 'ippair'.
Instead of handling the packet update during flow lookup, handle
it in the stream/detect threads. This lowers the load of the
capture thread(s) in autofp mode.
The decoders now set a flag in the packet if the packet needs a
flow lookup. Then the workers will take care of this. The decoders
also already calculate the raw flow hash value. This is so that
this value can be used in flow balancing in autofp.
Because the flow lookup/creation is now done in the worker threads,
the flow balancing can no longer use the flow. It's not yet
available. Autofp load balancing uses raw hash values instead.
In the same line, move UDP AppLayer out of the DecodeUDP module,
and also into the stream/detect threads.
Handle TCP session reuse inside the flow engine itself. If a looked up
flow matches the packet, but is a TCP stream starter, check if the
ssn needs to be reused. If that is the case handle it within the
lookup function. Simplies the locking and removes potential race
conditions.
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.
This adds a counter indicating how many times
the flow max memcap has been reached
Since there is no always a reference to FlowManagerThreadData,
the counter is put in DecodeThreadVars.
Currently when there is no counter increase in one call of FlowGetNew
because we don't have tv or dtv at the time of the call.
The following is a snippet of the generated EVE entry:
"flow":{"memcap":0,"spare":10000,"emerg_mode_entered":0,"emerg_mode_over":0,"tcp_reuse":0,"memuse":7085248}
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.
For the autofp case, handling TCP reuse in the flow engine didn't work.
The problem is the mismatch between the moment the flow engine looks at
packets and the stream, and the moment the stream engine runs. Flow engine
is invoked in the packet capture thread(s), while the stream engine runs
as part of the stream/detect thread(s). Because of the queues between
those threads the flow manager may already inspect a new SYN while the
stream engine still has to process the previous session.
Moving the flow engine to the stream/detect thread(s) wasn't an option
as the 'autofp' load balancing depends on the flow already being
available in the packet.
The solution here is to add a check for this condition to the stream
engine. At this point the TCP state is up to date. If a TCP reuse case
is encountered, this is the global logic:
- detach packet for old flow
- get a new flow and attach it to the packet
- flag the old flow that it is now obsolete
Additional logic makes sure that the packets already in the queue
between the flow thread(s) and the stream thread are reassigned the
new flow.
Some special handling:
Apply previous 'reuse' before checking for a new reuse. Otherwise we're
tagging the wrong flow in some cases (multiple reuses in the same tuple).
When in a flow/ssn reuse condition, properly remove the packet from
the flow.
Don't 'reuse' if packet is a SYN retransmission.
The old flow is timed out normally by the flow manager.
Until now, TCP session reuse was handled in the TCP stream engine.
If the state was TCP_CLOSED, a new SYN packet was received and a few
other conditions were met, the flow was 'reset' and reused for the
'new' TCP session.
There are a number of problems with this approach:
- it breaks the normal flow lifecycle wrt timeout, detection, logging
- new TCP sessions could come in on different threads due to mismatches
in timeouts between suricata and flow balancing hw/nic/drivers
- cleanup code was often causing problems
- it complicated locking because of the possible thread mismatch
This patch implements a different solution, where a new TCP session also
gets a new flow. To do this 2 main things needed to be done:
1. the flow engine needed to be aware of when the TCP reuse case was
happening
2. the flow engine needs to be able to 'skip' the old flow once it was
replaced by a new one
To handle (1), a new function TcpSessionPacketSsnReuse() is introduced
to check for the TCP reuse conditions. It's called from 'FlowCompare()'
for TCP packets / TCP flows that are candidates for reuse. FlowCompare
returns FALSE for the 'old' flow in the case of TCP reuse.
This in turn will lead to the flow engine not finding a flow for the TCP
SYN packet, resulting in the creation of a new flow.
To handle (2), FlowCompare flags the 'old' flow. This flag causes future
FlowCompare calls to always return FALSE on it. In other words, the flow
can't be found anymore. It can only be accessed by:
1. existing packets with a reference to it
2. flow timeout handling as this logic gets the flows from walking the
hash directly
3. flow timeout pseudo packets, as they are set up by (2)
The old flow will time out normally, as governed by the "tcp closed"
flow timeout setting. At timeout, the normal detection, logging and
cleanup code will process it.
The flagging of a flow making it 'unfindable' in the flow hash is a bit
of a hack. The reason for this approach over for example putting the
old flow into a forced timeout queue where it could be timed out, is
that such a queue could easily become a contention point. The TCP
session reuse case can easily be created by an attacker. In case of
multiple packet handlers, this could lead to contention on such a flow
timeout queue.
In the lastts timeval struct field in the flow the timestamp of the
last packet to update is recorded. This allows for tracking the timeout
of the flow. So far, this value was updated under the flow lock and also
read under the flow lock.
This patch moves the updating of this field to the FlowGetFlowFromHash
function, where it updated at the point where both the Flow and the
Flow Hash Row lock are held. This guarantees that the field is only
updated when both locks are held.
This makes reading the field safe when either lock is held, which is the
purpose of this patch.
The flow manager, while holding the flow hash row lock, can now safely
read the lastts value. This allows it to do the flow timeout check
without actually locking the flow.
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.
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.
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 *.
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.
clang was issuing some warnings related to unused return in function.
This patch adds some needed error treatment and ignore the rest of the
warnings by adding a cast to void.
Victor Julien
Eric Leblond
Max Fillinger
Jason Ish
Giuseppe Longo
Ken Steele
Anoop Saldanha