Fix: ring buffer: handle concurrent update in nested buffer wrap around check
With stress-test loads that trigger sub-buffer switch very frequently
(small 4kB sub-buffers, frequent flush), we currently observe this kind
of warnings once every few minutes:
[65335.896208] ring buffer relay-overwrite-mmap, cpu 5: records were lost. Caused by:
[65335.896208] [ 0 buffer full, 1 nest buffer wrap-around, 0 event too big ]
It appears that the check for nested buffer wrap-around does not take
into account that a concurrent execution contexts (either nested for
per-cpu buffers, or from another CPU or nested for global buffers) can
update the commit_count value concurrently.
What we really want to do with this check is to ensure that if we enter
a sub-buffer that had an unbalanced reserve/commit count, assuming there
is no hope that this gets rebalanced promptly, we detect this and drop
the current event. However, in the case where the commit counter has
been concurrently updated by another reserve or a switch, we want to
retry the entire reserve operation.
One way to detect this is to sample the reserve offset twice, around the
commit counter read, along with the appropriate memory barriers.
Therefore, we can detect if the mismatch between reserve and commit
counter is actually caused by a concurrent update, which necessarily has
updated the reserve counter.
lib_ring_buffer_write(), lib_ring_buffer_memset() and
lib_ring_buffer_copy_from_user_inatomic() could be passed a length of 0.
This typically has no side-effect as far as writing into the buffers is
concerned, except for one detail: in overwrite mode, there is a check to
make sure the sub-buffer can be written into. This check is performed
even if length is 0. In the case where this would fall exactly at the
end of a sub-buffer, the check would fail, because the offset would fall
exactly at the beginning of the next sub-buffer.
Fix: ring buffer: RING_BUFFER_FLUSH ioctl buffer corruption
lib_ring_buffer_switch_slow() clearly states:
* Note, however, that as a v_cmpxchg is used for some atomic
* operations, this function must be called from the CPU which owns the
* buffer for a ACTIVE flush.
But unfortunately, the RING_BUFFER_FLUSH ioctl does not follow these
important directives. Therefore, whenever the consumer daemon or session
daemon explicitly triggers a "flush" on a buffer, it can race with data
being written to the buffer, leading to corruption of the reserve/commit
counters, and therefore corruption of data in the buffer. It triggers
these warnings for overwrite mode buffers:
[65356.890016] WARNING: at
/home/compudj/git/lttng-modules/wrapper/ringbuffer/../../lib/ringbuffer/../../wrapper/ringbuffer/../../lib/ringbuffer/backend.h:110 lttng_event_write+0x118/0x140 [lttng_ring_buffer_client_mmap_overwrite]()
Which indicates that we are trying to write into a sub-buffer for which
we don't have exclusive access. It also causes the following warnings to
show up:
[65335.896208] ring buffer relay-overwrite-mmap, cpu 5: records were lost. Caused by:
[65335.896208] [ 0 buffer full, 80910 nest buffer wrap-around, 0 event too big ]
Which is caused by corrupted commit counter.
Fix this by sending an IPI to the CPU owning the flushed buffer for
per-cpu synchronization. For global synchronization, no IPI is needed,
since we allow writes from remote CPUs.
88ec3fa...
by
Samuel Martin <email address hidden>
Fix build and load against linux-2.6.33.x
* lttng-event.h declared but did not implement
lttng_add_perf_counter_to_ctx on kernel >=2.6.33, the implementation
was in lttng-context-perf-counters.c, which was only included for
kernel >=2.6.34. This prevented the module from being loaded.
* on kernel 2.6.33.x, lttng-context-perf-counters.c complains about
implicit declaration for {get,put}_online_cpus and
{,un}register_cpu_notifier; so fix header inclusion.
Signed-off-by: Samuel Martin <email address hidden>
Signed-off-by: Mathieu Desnoyers <email address hidden>
