[lttng-dev] segfault when exiting process with background thread

Thu Sep 22 22:48:06 UTC 2016

Yes, tracepoint_unregister_lib is the function I was referring. If the function is okay to call with background thread still writing trace, I am definitely fine. I just want to double check it, if there could be a potential bug. Thanks.

________________________________
From: Mathieu Desnoyers <mathieu.desnoyers at efficios.com>
Sent: Wednesday, September 21, 2016 2:05 PM
To: Jeffrey Chen
Cc: Jeffrey Chen; lttng-dev
Subject: Re: segfault when exiting process with background thread

I don't see anything relevant that changed since commit:

commit 664ccf245fcb343e1ea10e145bab3749423e9d0f
Author: Mathieu Desnoyers <mathieu.desnoyers at efficios.com>
Date:   Thu Mar 10 09:32:59 2016 -0500

    Add tracepoint_disable_destructors()

    Calling this function from an instrumented program allows disabling
    tracepoint destructors. This allows threads to continue calling
    tracepoint code even after the tracepoint destructors have run. This is
    needed for applications that exit without joining all their threads.

    Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers at efficios.com>
    CC: Jeffrey Chen <cpthk at live.com>

Can you point it out ? And let me know what is segfaulting, with a backtrace.

I see that tracepoint_dlopen_ptr->tracepoint_unregister_lib(__start___tracepoints_ptrs)
is still invoked from __tracepoints__ptrs_destroy() even if the disable destructors is set,
but this only unregisters the library from the list of available events. I don't see how it would
trigger a crash of threads still actively performing tracing.

Thanks,

Mathieu

----- On Sep 21, 2016, at 4:00 PM, Jeffrey Chen <cpthk at live.com> wrote:

It seems that some changes went in recently that caused this workaround to not work again. Originally, the __tracepoints__disable_destructors variable should be checked around the beginning of the destructor functions (__tracepoints__destroy and __tracepoints__ptrs_destroy), and return the function when the variable is true. It is not in the beginning any more. So, the segfault problem could happen. Is it possible to fix it again at your convenience? Thanks a lot.

________________________________
From: lttng-dev <lttng-dev-bounces at lists.lttng.org> on behalf of Mathieu Desnoyers <mathieu.desnoyers at efficios.com>
Sent: Tuesday, March 15, 2016 12:00 PM
To: Jeffrey Chen
Cc: lttng-dev
Subject: Re: [lttng-dev] segfault when exiting process with background thread

Commit merged into master, thanks for your feedback!

Mathieu

----- On Mar 15, 2016, at 2:48 PM, Jeffrey Chen <cpthk at hotmail.com> wrote:

Hi Mathieu:

Will this fix be checked-in? Thanks.

________________________________
From: Mathieu Desnoyers <mathieu.desnoyers at efficios.com>
Sent: Thursday, March 10, 2016 3:27 PM
To: Jeffrey Chen
Cc: Jonathan Rajotte Julien; lttng-dev
Subject: Re: segfault when exiting process with background thread

----- On Mar 10, 2016, at 5:59 PM, Jeffrey Chen <cpthk at hotmail.com> wrote:

Thanks, Mathieu.

Yes, the fix works as a workaround. This is also the same workaround we have now.

It seems the problem is due to a race condition, that it checked whether "tracepoint_dlopen.rcu_read_lock_sym_bp" is not NULL. If not NULL, it calls to that. But, at this moment, memset cleared the memory.

This fix is okay with us, but I am not certain if it would break other component, since I do not understand LTTng enough to tell. Thanks.

In addition to memset clearing that memory, dlclose() of the lttng-ust-tracepoint
shared object will clear the memory containing the called functions. Therefore,
we need to skip both steps of the destructor if we want threads to survive after
this destructor execution.

Thanks,

Mathieu

________________________________
From: Mathieu Desnoyers <mathieu.desnoyers at efficios.com>
Sent: Thursday, March 10, 2016 10:52 AM
To: Jeffrey Chen
Cc: Jonathan Rajotte Julien; lttng-dev
Subject: Re: segfault when exiting process with background thread
Hi Jeffrey,

I CC'd your @live.com address on the RFC patch. Please let me know if the approach
is OK with you (and try it out).

See:
https://lists.lttng.org/pipermail/lttng-dev/2016-March/025608.html

Thanks,

Mathieu

----- On Mar 10, 2016, at 1:35 PM, Jeffrey Chen <cpthk at hotmail.com> wrote:

Thanks Mathieu.

Is there a plan at LTTng side to fix this issue? If so, we could wait for the fix. If not, we will have to workaround the problem for now (probably by applying your fix ourselves). Thanks.

________________________________
From: Mathieu Desnoyers <mathieu.desnoyers at efficios.com>
Sent: Tuesday, March 8, 2016 6:09 PM
To: Jonathan Rajotte Julien
Cc: lttng-dev
Subject: Re: [lttng-dev] segfault when exiting process with background thread
The attached patch works around the issue. The problem is
the tracepoint destructors that run in parallel with the thread
that still exists.

If we remove this, we leak the liblttng-ust-tracepoint.so.0 shared
object on dlclose of the instrumented code. Not sure if we should
care though...

Thoughts ?

Thanks,

Mathieu

----- On Mar 8, 2016, at 7:35 PM, Mathieu Desnoyers <mathieu.desnoyers at efficios.com> wrote:
Reproduced it on master (ust and tools).

Repro with:

lttng create
lttng enable-event -u -a
lttng start

then run hello a few times, it eventually segfaults in the spawned thread.

Looking into it.

Thanks,

Mathieu

----- On Mar 8, 2016, at 7:12 PM, Jonathan Rajotte Julien <Jonathan.rajotte-julien at efficios.com> wrote:
Hi Jeffrey,

Could you provide the version for lttng-ust, lttng-tools, lttng-modules (if installed)?

Were you able to reproduce the issue on master (git) ?

What are the exact order of command to reproduce the problem ?

Thanks

On 2016-03-08 06:43 PM, Jeffrey Chen wrote:

Hi LTTng community:

We are using LTTng for our production environment. We have been noticing segfault problem when our process exit. We think the problem is due to the background thread continue to write traces, while LTTng cleanup its memory. We do not know how to fix this issue without changing LTTng code base. I have written a very simple app that could repro the problem. Most of my codes are copy pasted from the LTTng doc sample. Is there any fix that the LTTng side could do?

hello.c

#include <stdio.h>
#include "hello-tp.h"
#include <pthread.h>

void* doSomeThing(void *arg)
{
    int x;
    for (x = 0; x < 100000; ++x) {
        tracepoint(hello_world, my_first_tracepoint, x, "test");
    }
}

int main(int argc, char *argv[])
{
    int x;

    getchar();

    pthread_t inc_x_thread;

    if(pthread_create(&inc_x_thread, NULL, doSomeThing, NULL)) {

        fprintf(stderr, "Error creating thread\n");
        return 1;
    }

    tracepoint(hello_world, my_first_tracepoint, 23, "hi there!");

    return 0;
}

hello-tp.c

#define TRACEPOINT_CREATE_PROBES
#define TRACEPOINT_DEFINE

#include "hello-tp.h"

hello-tp.h

#undef TRACEPOINT_PROVIDER
#define TRACEPOINT_PROVIDER hello_world

#undef TRACEPOINT_INCLUDE
#define TRACEPOINT_INCLUDE "./hello-tp.h"

#if !defined(_HELLO_TP_H) || defined(TRACEPOINT_HEADER_MULTI_READ)
#define _HELLO_TP_H

#include <lttng/tracepoint.h>

TRACEPOINT_EVENT(
    hello_world,
    my_first_tracepoint,
    TP_ARGS(
        int, my_integer_arg,
        char*, my_string_arg
    ),
    TP_FIELDS(
        ctf_string(my_string_field, my_string_arg)
        ctf_integer(int, my_integer_field, my_integer_arg)
    )
)

#endif /* _HELLO_TP_H */

#include <lttng/tracepoint-event.h>

Compile

gcc -c -I. hello-tp.c
gcc -c hello.c
gcc -o hello hello.o hello-tp.o -llttng-ust -ldl -lpthread

_______________________________________________
lttng-dev mailing list
lttng-dev at lists.lttng.org<mailto:lttng-dev at lists.lttng.org>http://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev

--
Jonathan R. Julien
Efficios

_______________________________________________
lttng-dev mailing list
lttng-dev at lists.lttng.org
http://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev

--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

_______________________________________________
lttng-dev mailing list
lttng-dev at lists.lttng.org
http://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev

--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.lttng.org/pipermail/lttng-dev/attachments/20160922/136bb2b6/attachment-0001.html>