Got a suggestion from mod_pagespeed project: call _exit instead of exit. 
I'll test it out to see if the segmentation fault would be gone.

Alex

On Monday, May 12, 2014 6:28:28 PM UTC-7, Graham Dumpleton wrote:
>
>
> On 13/05/2014, at 11:13 AM, Alex Wu <[email protected] <javascript:>> 
> wrote:
>
> My question is that if mod_wsgi should wipe out all meomry inherited from 
> parent once it forks?
>
>
> It can't. It relies on it being a fork (and not a fork/exec) to inherit 
> everything.
>
> I am not clear if a module inherits an C++ object from parent, does it 
> trigger a destructor call?
>
>
> Most likely it would.
>
> What I don't know is if you unload a module does that by pass execution of 
> finaliser sections.
>
> I would imagine it cannot by pass them else memory from the heap would not 
> be released otherwise, if referenced by global C++ objects, and you would 
> get a potential memory leak.
>
> This may not matter on process shutdown, but would during an Apache 
> restart as Apache will unload and reload modules when that occurs.
>
> So although in the Apache parent it does appear to unload modules on 
> process shutdown:
>
>     /*
>      * Register a cleanup in the config apr_pool_t (normally pconf). When
>      * we do a restart (or shutdown) this cleanup will cause the
>      * shared object to be unloaded.
>      */
>     apr_pool_cleanup_register(cmd->pool, modi, unload_module, 
> apr_pool_cleanup_null);
>
>
> int main(…) {
>     …
>
>     destroy_and_exit_process(process, 0);
>
>     return 0; /* Termination 'ok' */
> }
>
> static void destroy_and_exit_process(process_rec *process,
>                                      int process_exit_value)
> {
>     /*
>      * Sleep for TASK_SWITCH_SLEEP micro seconds to cause a task switch on
>      * OS layer and thus give possibly started piped loggers a chance to
>      * process their input. Otherwise it is possible that they get killed
>      * by us before they can do so. In this case maybe valueable log 
> messages
>      * might get lost.
>      */
>     apr_sleep(TASK_SWITCH_SLEEP);
>     apr_pool_destroy(process->pool); /* and destroy all descendent pools */
>     apr_terminate();
>     exit(process_exit_value);
> }
>
> doing that may not help and may just trigger it at that point instead.
>
> I will though need to look into whether I should introduce something 
> similar just prior to calling exit() in the daemon processes.
>
> I would have to be very careful about what pools I destroy though. Or 
> perhaps work out how just to trigger cleanup routines on selected pools.
>
> Graham
>
> Alex
>
> On Monday, May 12, 2014 5:42:27 PM UTC-7, Graham Dumpleton wrote:
>>
>> Okay. So this isn't an atexit() callback but global C++ object 
>> destructors kicking in from the automatic execution of finaliser sections 
>> on the object files.
>>
>> Same issue applies though in part. It looks like the page speed module 
>> could be making some assumption that certain data will always be 
>> initialised by the time the process is terminated, but possibly because 
>> Apache module child init handlers are not called for the page speed module 
>> in the mod_wsgi daemon processes, then that data isn't initialised and as a 
>> result it crashes.
>>
>> When this happens though it is usual to see a NULL pointer dereference or 
>> low memory access due to relative reference to NULL pointer. I can't see an 
>> obvious case of that, but is hard to tell what the module is doing.
>>
>> Another problem with this thought is that since the page speed module 
>> doesn't get to do anything at all in the mod_wsgi daemon mode process, then 
>> can't see how this issue wouldn't also arise in the Apache parent process 
>> unless the fact that the module might be unloaded from memory by Apache 
>> first before shutdown (can't remember) might mean that global C++ 
>> destructors aren't called in that case.
>>
>> Now one could argue that if this is happening that the page speed module 
>> is being sloppy, but at the same time, under normal circumstances an Apache 
>> module would never need to contend with possibility that something like the 
>> Apache child init handler might not be called in a child process. That is 
>> an oddity caused by mod_wsgi daemon mode.
>>
>> Anyway, all can do right now is confirm whether it is the page speed 
>> module by disabling that module temporarily.
>>
>> Will then need to work out what to do and perhaps raise issue with page 
>> speed module authors if that is where it is arising and see if they want to 
>> say not their problem since mod_wsgi does weird stuff. :-)
>>
>> Graham
>>
>> On 13/05/2014, at 9:51 AM, Alex Wu <[email protected]> wrote:
>>
>> Here is one example:
>>
>> warning: Can't read pathname for load map: Input/output error.
>> [Thread debugging using libthread_db enabled]
>> Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
>> Core was generated by `(wsgi:dataplane)                  -D Dataplane -D 
>> pagespeed -D fwd_proxy -D DAT'.
>> Program terminated with signal 11, Segmentation fault.
>> #0  0x00007fb46da18e4a in ?? () from /usr/lib/libpython2.7.so.1.0
>> (gdb) info threads
>>   Id   Target Id         Frame 
>>   5    Thread 0x7fb458fe9700 (LWP 25847) 0x00007fb4777566e0 in 
>> sigprocmask () from /lib/x86_64-linux-gnu/libc.so.6
>>   4    Thread 0x7fb4785bb740 (LWP 22886) 0x00007fb46cadd678 in (anonymous 
>> namespace)::scribble (ptr=0x7fb478f13a38, size=34008, 
>> scribble_word=-559038737)
>>     at pagespeed/kernel/base/mem_debug.cc:81
>>   3    Thread 0x7fb469870700 (LWP 22895) 0x00007fb477814a93 in epoll_wait 
>> () from /lib/x86_64-linux-gnu/libc.so.6
>>   2    Thread 0x7fb46aa76700 (LWP 22893) 0x00007fb47780d763 in select () 
>> from /lib/x86_64-linux-gnu/libc.so.6
>> * 1    Thread 0x7fb46a071700 (LWP 22894) 0x00007fb46da18e4a in ?? () from 
>> /usr/lib/libpython2.7.so.1.0
>> (gdb) thread 4
>> [Switching to thread 4 (Thread 0x7fb4785bb740 (LWP 22886))]
>> #0  0x00007fb46cadd678 in (anonymous namespace)::scribble 
>> (ptr=0x7fb478f13a38, size=34008, scribble_word=-559038737) at 
>> pagespeed/kernel/base/mem_debug.cc:81
>> 81    pagespeed/kernel/base/mem_debug.cc: No such file or directory.
>> (gdb) bt
>> #0  0x00007fb46cadd678 in (anonymous namespace)::scribble 
>> (ptr=0x7fb478f13a38, size=34008, scribble_word=-559038737) at 
>> pagespeed/kernel/base/mem_debug.cc:81
>> #1  0x00007fb46cadd827 in (anonymous namespace)::debug_free 
>> (ptr=0x7fb478f13a38) at pagespeed/kernel/base/mem_debug.cc:100
>> #2  0x00007fb46cadd9f9 in operator delete[] (ptr=0x7fb478f13a38) at 
>> pagespeed/kernel/base/mem_debug.cc:142
>> #3  0x00007fb46ce2256e in re2::Prog::~Prog (this=0x7fb478c260e8, 
>> __in_chrg=<optimized out>) at third_party/re2/src/re2/prog.cc:123
>> #4  0x00007fb46cdf5402 in re2::RE2::~RE2 (this=0x7fb478ff3dd8, 
>> __in_chrg=<optimized out>) at third_party/re2/src/re2/re2.cc:272
>> #5  0x00007fb46d1033af in 
>> pagespeed::js::JsTokenizerPatterns::~JsTokenizerPatterns 
>> (this=0x7fb478ff3dd8, __in_chrg=<optimized out>)
>>     at pagespeed/kernel/js/js_tokenizer.cc:1096
>> #6  0x00007fb46cf9f00c in 
>> base::DefaultDeleter<pagespeed::js::JsTokenizerPatterns>::operator() 
>> (this=0x7fb46d6a6fe8, ptr=0x7fb478ff3dd8)
>>     at third_party/chromium/src/base/memory/scoped_ptr.h:137
>> #7  0x00007fb46cf9efc2 in 
>> base::internal::scoped_ptr_impl<pagespeed::js::JsTokenizerPatterns, 
>> base::DefaultDeleter<pagespeed::js::JsTokenizerPatterns> >::~scoped_ptr_impl
>>     (this=0x7fb46d6a6fe8, __in_chrg=<optimized out>) at 
>> third_party/chromium/src/base/memory/scoped_ptr.h:220
>> #8  0x00007fb46cf9ef6c in scoped_ptr<pagespeed::js::JsTokenizerPatterns, 
>> base::DefaultDeleter<pagespeed::js::JsTokenizerPatterns> >::~scoped_ptr 
>> (this=0x7fb46d6a6fe8, 
>>     __in_chrg=<optimized out>) at 
>> third_party/chromium/src/base/memory/scoped_ptr.h:310
>> #9  0x00007fb46cf9ef33 in net_instaweb::ProcessContext::~ProcessContext 
>> (this=0x7fb46d6a6fe8, __in_chrg=<optimized out>) at 
>> net/instaweb/rewriter/process_context.cc:54
>> #10 0x00007fb46cad3969 in net_instaweb::(anonymous 
>> namespace)::ApacheProcessContext::~ApacheProcessContext 
>> (this=0x7fb46d6a6fe0, __in_chrg=<optimized out>)
>>     at net/instaweb/apache/mod_instaweb.cc:313
>> #11 0x00007fb47775b901 in ?? () from /lib/x86_64-linux-gnu/libc.so.6
>> #12 0x00007fb47775b985 in exit () from /lib/x86_64-linux-gnu/libc.so.6
>> #13 0x00007fb46dddfd96 in wsgi_start_process (p=<optimized out>, 
>> daemon=<optimized out>) at mod_wsgi.c:11969
>> #14 0x00007fb46dde1344 in wsgi_start_daemons (p=0x7fb478bac138) at 
>> mod_wsgi.c:12166
>> #15 wsgi_hook_init (pconf=0x7fb478bac138, ptemp=<optimized out>, 
>> plog=<optimized out>, s=<optimized out>) at mod_wsgi.c:13737
>> #16 0x00007fb478633113 in ap_run_post_config (pconf=0x7fb478bac138, 
>> plog=0x7fb478bd9378, ptemp=0x7fb478bd7348, s=0x7fb478bd5538) at config.c:106
>> #17 0x00007fb478608993 in main (argc=15, argv=0x7fff2ee8cfd8) at 
>> main.c:765
>> On Monday, May 12, 2014 4:07:35 PM UTC-7, Graham Dumpleton wrote:
>>>
>>> Can you point out to me where in the Apache 2.4 code base it calls 
>>> atexit() to register anything on process shutdown?
>>>
>>> Neither Apache nor the underlying APR/APU libraries that it uses rely on 
>>> atexit() to have anything triggered on process shutdown that I know of and 
>>> I cannot find anything in the code I have handy for those which uses 
>>> atexit() in such a generic way.
>>>
>>> Normally Apache relies on cleanup actions attached to deletion of memory 
>>> pools and not atexit(). Thus it requires orderly Apache process shutdown 
>>> and for memory pools to be destroyed for actions to be performed on process 
>>> shutdown. The destruction of memory pools is not triggered via atexit().
>>>
>>> Do you also have a more extensive stack trace that that one line so I 
>>> can see in what actual code the crash occurs? That may give me more clues.
>>>
>>> Graham
>>>
>>> On 13/05/2014, at 8:58 AM, Alex Wu <[email protected]> wrote:
>>>
>>> we do not specifically add hook to atexit. It is called/triggered by 
>>> apache frame work when a module is written within the apache 2.4 frame 
>>> work. Also, mod_pagespeed used scoped point on their server context, it 
>>> triggers auto clean once exit is called and library is unloaded. 
>>>
>>> Alex
>>>
>>>
>>>
>>> On Monday, May 12, 2014 3:40:26 PM UTC-7, Graham Dumpleton wrote:
>>>>
>>>> If your own Apache modules are using atexit() to perform cleanup on 
>>>> process exit, rather than Apache's own mechanisms for performing cleanup 
>>>> actions when the pool the module uses is cleaned up, then the atexit() 
>>>> callback will have to take into consideration that under mod_wsgi when 
>>>> using daemon mode, that the Apache module child init handler will not be 
>>>> called in the daemon process for your Apache module. Thus the callback 
>>>> should check whether global data pointers are in fact non NULL before 
>>>> trying to do things with them.
>>>>
>>>> Can you confirm you are using atexit() callbacks in C code with your 
>>>> Apache modules and explain at what point you are registering the callback 
>>>> with atexit()?
>>>>
>>>> Is there a specific reason you are using atexit() callbacks rather than 
>>>> doing the normal thing of in the Apache module child init handler 
>>>> registering a cleanup callback on the memory pool given to the Apache 
>>>> module on child init and relying on that being triggered by Apache when 
>>>> shutting things down?
>>>>
>>>> Graham
>>>>
>>>> On 13/05/2014, at 8:23 AM, Alex Wu <[email protected]> wrote:
>>>>
>>>> some are our own, one is mod_pagespeed. We use python 2.7.3 with apache 
>>>> 2.4.7 in MPM mode. The segmentation fault is cleanup routine of each 
>>>> modules other than mod_wsgi after exit call.
>>>>
>>>> Alex
>>>>
>>>>
>>>> On Monday, May 12, 2014 1:50:35 PM UTC-7, Graham Dumpleton wrote:
>>>>>
>>>>> On 13/05/2014, at 4:40 AM, Alex Wu <[email protected]> wrote: 
>>>>>
>>>>> > We have observed various segmentation fault caused by exit call from 
>>>>> mod_wsgi 3.5: 
>>>>> > 
>>>>> > #20 0x00007f9490a94d96 in wsgi_start_process (p=<optimized out>, 
>>>>> daemon=<optimized out>) at mod_wsgi.c:11969 
>>>>> > 
>>>>> > The exit call triggers cleanup from other modules, that cleanup 
>>>>> caused segmentation fault, 
>>>>>
>>>>> What version of Apache and Python are you using? 
>>>>>
>>>>> What other non standard Apache modules are you using? For example, is 
>>>>> PHP being used in the same Apache instance? 
>>>>>
>>>>> Graham 
>>>>>
>>>>>
>>>>>
>>>> -- 
>>>> You received this message because you are subscribed to the Google 
>>>> Groups "modwsgi" group.
>>>> To unsubscribe from this group and stop receiving emails from it, send 
>>>> an email to [email protected].
>>>> To post to this group, send email to [email protected].
>>>> Visit this group at http://groups.google.com/group/modwsgi.
>>>> For more options, visit https://groups.google.com/d/optout.
>>>>
>>>>
>>>>
>>> -- 
>>> You received this message because you are subscribed to the Google 
>>> Groups "modwsgi" group.
>>> To unsubscribe from this group and stop receiving emails from it, send 
>>> an email to [email protected].
>>> To post to this group, send email to [email protected].
>>> Visit this group at http://groups.google.com/group/modwsgi.
>>> For more options, visit https://groups.google.com/d/optout.
>>>
>>>
>>>
>> -- 
>> You received this message because you are subscribed to the Google Groups 
>> "modwsgi" group.
>> To unsubscribe from this group and stop receiving emails from it, send an 
>> email to [email protected].
>> To post to this group, send email to [email protected].
>> Visit this group at http://groups.google.com/group/modwsgi.
>> For more options, visit https://groups.google.com/d/optout.
>>
>>
>>
> -- 
> You received this message because you are subscribed to the Google Groups 
> "modwsgi" group.
> To unsubscribe from this group and stop receiving emails from it, send an 
> email to [email protected] <javascript:>.
> To post to this group, send email to [email protected] <javascript:>
> .
> Visit this group at http://groups.google.com/group/modwsgi.
> For more options, visit https://groups.google.com/d/optout.
>
>
>

-- 
You received this message because you are subscribed to the Google Groups 
"modwsgi" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
To post to this group, send email to [email protected].
Visit this group at http://groups.google.com/group/modwsgi.
For more options, visit https://groups.google.com/d/optout.

Reply via email to