link: close tracefs event if perf event cannot be opened

[nashuiliang]

When getTraceEventID() or openTracepointPerfEvent() returns an error,
tracefs event (e.g. /sys/kernel/debug/tracing/events/), which has
been created by createTraceFSProbeEvent(), is not cleaned up.

Signed-off-by: Chuang Wang nashuiliang@gmail.com

[mmat11]

I remember pointing this out in the initial Kprobe implementation, here's the discussion: #265 (comment). I have no strong opinion but prefer cleaning it up /cc @ti-mo

[nashuiliang]

Thanks, I read the discussion: #265 (comment).

I should briefly describe my scene:

When livepatch and kprobe are using the same kernel function X, tracefs event (e.g. /sys/kernel/debug/tracing/events/), which has been created by createTraceFSProbeEvent(), cannot be enabled by syscall perf_event_open.

Therefore, I should add closeTraceFSProbeEvent() on failed openTracepointPerfEvent().
Hmm, I saw getTraceEventID() as well, and added it by the way, although I thought it might not return an error.

[ti-mo]

@nashuiliang Thanks for the patch! I've force-pushed an alternative solution to your branch that makes sure this is not forgotten in the future.

Is there a way this could be tested? This seems fairly tricky to do.

[nashuiliang]

@ti-mo Wow, cool!

The test method is more complicated. It can be triggered manually, and it is difficult to write test code at the same time.

[lmb]

I should briefly describe my scene:

When livepatch and kprobe are using the same kernel function X, tracefs event (e.g. /sys/kernel/debug/tracing/events/), which has been created by createTraceFSProbeEvent(), cannot be enabled by syscall perf_event_open.

Interesting! Out of curiosity, which errno do you get from perf_event_open?

[nashuiliang]

Error: opening tracepoint perf event: device or resource busy.
Errno: -EBUSY