Here is an excerpt from a stack trace I got while profiling with OProfile:

#0  sem_wait () from /lib64/libpthread.so.0
#1  thread_suspendAll () at core/thread.d:2471
#2  gc.gcx.Gcx.fullcollect() (this=...) at gc/gcx.d:2427
#3  gc.gcx.Gcx.bigAlloc() (this=..., size=16401, poolPtr=0x7fc3d4bfe3c8, alloc_size=0x7fc3d4bfe418) at gc/gcx.d:2099
#4  gc.gcx.GC.mallocNoSync (alloc_size=0x7fc3d4bfe418, bits=10, size=16401, this=...) gc/gcx.d:503
#5  gc.gcx.GC.malloc() (this=..., size=16401, bits=10, alloc_size=0x7fc3d4bfe418) gc/gcx.d:421
#6  gc.gc.gc_qalloc (ba=10, sz=<optimized out>) gc/gc.d:203
#7  gc_qalloc (sz=<optimized out>, ba=10) gc/gc.d:198
#8  _d_newarrayT (ti=..., length=4096) rt/lifetime.d:807
#9  sequencer.algorithm.gzip.HuffmanTree.__T6__ctorTG32hZ.__ctor() (this=..., bitLengths=...) sequencer/algorithm/gzip.d:444

Two more threads are alive, but waiting on a condition variable (i.e.: in pthread_cond_wait(), but from my own and not from druntime code. Is there some obvious way I could have dead-locked the GC ? Or is there a bug ?

This was compiled with GDC using DMD FE 2.062.

-- 
Marco

One more note: I get this consistently during profiling, but
not without.
I don't count kernel involvement out either, since OProfile is
a kernel based profiler and there could be a quirk in its
interaction with semaphores.

-- 
Marco

On Jun 13, 2013, at 2:22 AM, Marco Leise <Marco.Leise@gmx.de> wrote:

> Here is an excerpt from a stack trace I got while profiling with OProfile:
> 
> #0  sem_wait () from /lib64/libpthread.so.0
> #1  thread_suspendAll () at core/thread.d:2471
> #2  gc.gcx.Gcx.fullcollect() (this=...) at gc/gcx.d:2427
> #3  gc.gcx.Gcx.bigAlloc() (this=..., size=16401, poolPtr=0x7fc3d4bfe3c8, alloc_size=0x7fc3d4bfe418) at gc/gcx.d:2099
> #4  gc.gcx.GC.mallocNoSync (alloc_size=0x7fc3d4bfe418, bits=10, size=16401, this=...) gc/gcx.d:503
> #5  gc.gcx.GC.malloc() (this=..., size=16401, bits=10, alloc_size=0x7fc3d4bfe418) gc/gcx.d:421
> #6  gc.gc.gc_qalloc (ba=10, sz=<optimized out>) gc/gc.d:203
> #7  gc_qalloc (sz=<optimized out>, ba=10) gc/gc.d:198
> #8  _d_newarrayT (ti=..., length=4096) rt/lifetime.d:807
> #9  sequencer.algorithm.gzip.HuffmanTree.__T6__ctorTG32hZ.__ctor() (this=..., bitLengths=...) sequencer/algorithm/gzip.d:444
> 
> Two more threads are alive, but waiting on a condition variable (i.e.: in pthread_cond_wait(), but from my own and not from druntime code. Is there some obvious way I could have dead-locked the GC ? Or is there a bug ?

I assume you're running on Linux, which uses signals (SIGUSR1, specifically) to suspend threads for a collection.  So I imagine what's happening is that your thread is trying to suspend all the other threads so it can collect, and those threads are ignoring the signal for some reason.  I would expect pthread_cond_wait to be interrupted if a signal arrives though.  Have you overridden the signal handler for SIGUSR1?

Am Mon, 17 Jun 2013 10:46:19 -0700
schrieb Sean Kelly <sean@invisibleduck.org>:

> On Jun 13, 2013, at 2:22 AM, Marco Leise <Marco.Leise@gmx.de> wrote:
> 
> > Here is an excerpt from a stack trace I got while profiling with OProfile:
> > 
> > #0  sem_wait () from /lib64/libpthread.so.0
> > #1  thread_suspendAll () at core/thread.d:2471
> > #2  gc.gcx.Gcx.fullcollect() (this=...) at gc/gcx.d:2427
> > #3  gc.gcx.Gcx.bigAlloc() (this=..., size=16401, poolPtr=0x7fc3d4bfe3c8, alloc_size=0x7fc3d4bfe418) at gc/gcx.d:2099
> > #4  gc.gcx.GC.mallocNoSync (alloc_size=0x7fc3d4bfe418, bits=10, size=16401, this=...) gc/gcx.d:503
> > #5  gc.gcx.GC.malloc() (this=..., size=16401, bits=10, alloc_size=0x7fc3d4bfe418) gc/gcx.d:421
> > #6  gc.gc.gc_qalloc (ba=10, sz=<optimized out>) gc/gc.d:203
> > #7  gc_qalloc (sz=<optimized out>, ba=10) gc/gc.d:198
> > #8  _d_newarrayT (ti=..., length=4096) rt/lifetime.d:807
> > #9  sequencer.algorithm.gzip.HuffmanTree.__T6__ctorTG32hZ.__ctor() (this=..., bitLengths=...) sequencer/algorithm/gzip.d:444
> > 
> > Two more threads are alive, but waiting on a condition variable (i.e.: in pthread_cond_wait(), but from my own and not from druntime code. Is there some obvious way I could have dead-locked the GC ? Or is there a bug ?
> 
> I assume you're running on Linux, which uses signals (SIGUSR1, specifically) to suspend threads for a collection.  So I imagine what's happening is that your thread is trying to suspend all the other threads so it can collect, and those threads are ignoring the signal for some reason.  I would expect pthread_cond_wait to be interrupted if a signal arrives though.  Have you overridden the signal handler for SIGUSR1?

No, I have not overridden the signal handler. I'm aware of the fact that signals make pthread_cond_wait() return early and put them in a while loop as one would expect, that is all.

-- 
Marco

On Jun 18, 2013, at 7:01 AM, Marco Leise <Marco.Leise@gmx.de> wrote:

> Am Mon, 17 Jun 2013 10:46:19 -0700
> schrieb Sean Kelly <sean@invisibleduck.org>:
> 
>> On Jun 13, 2013, at 2:22 AM, Marco Leise <Marco.Leise@gmx.de> wrote:
>> 
>>> Here is an excerpt from a stack trace I got while profiling with OProfile:
>>> 
>>> #0  sem_wait () from /lib64/libpthread.so.0
>>> #1  thread_suspendAll () at core/thread.d:2471
>>> #2  gc.gcx.Gcx.fullcollect() (this=...) at gc/gcx.d:2427
>>> #3  gc.gcx.Gcx.bigAlloc() (this=..., size=16401, poolPtr=0x7fc3d4bfe3c8, alloc_size=0x7fc3d4bfe418) at gc/gcx.d:2099
>>> #4  gc.gcx.GC.mallocNoSync (alloc_size=0x7fc3d4bfe418, bits=10, size=16401, this=...) gc/gcx.d:503
>>> #5  gc.gcx.GC.malloc() (this=..., size=16401, bits=10, alloc_size=0x7fc3d4bfe418) gc/gcx.d:421
>>> #6  gc.gc.gc_qalloc (ba=10, sz=<optimized out>) gc/gc.d:203
>>> #7  gc_qalloc (sz=<optimized out>, ba=10) gc/gc.d:198
>>> #8  _d_newarrayT (ti=..., length=4096) rt/lifetime.d:807
>>> #9  sequencer.algorithm.gzip.HuffmanTree.__T6__ctorTG32hZ.__ctor() (this=..., bitLengths=...) sequencer/algorithm/gzip.d:444
>>> 
>>> Two more threads are alive, but waiting on a condition variable (i.e.: in pthread_cond_wait(), but from my own and not from druntime code. Is there some obvious way I could have dead-locked the GC ? Or is there a bug ?
>> 
>> I assume you're running on Linux, which uses signals (SIGUSR1, specifically) to suspend threads for a collection.  So I imagine what's happening is that your thread is trying to suspend all the other threads so it can collect, and those threads are ignoring the signal for some reason.  I would expect pthread_cond_wait to be interrupted if a signal arrives though.  Have you overridden the signal handler for SIGUSR1?
> 
> No, I have not overridden the signal handler. I'm aware of the fact that signals make pthread_cond_wait() return early and put them in a while loop as one would expect, that is all.

Hrm... Can you trap this in a debugger and post the stack traces of all threads?  That stack above is a thread waiting for others to say they're suspended so it can collect.

Am Tue, 18 Jun 2013 19:12:06 -0700
schrieb Sean Kelly <sean@invisibleduck.org>:

> On Jun 18, 2013, at 7:01 AM, Marco Leise <Marco.Leise@gmx.de> wrote:
> 
> > Am Mon, 17 Jun 2013 10:46:19 -0700
> > schrieb Sean Kelly <sean@invisibleduck.org>:
> > 
> >> On Jun 13, 2013, at 2:22 AM, Marco Leise <Marco.Leise@gmx.de> wrote:
> >> 
> >>> Here is an excerpt from a stack trace I got while profiling with OProfile:
> >>> 
> >>> #0  sem_wait () from /lib64/libpthread.so.0
> >>> #1  thread_suspendAll () at core/thread.d:2471
> >>> #2  gc.gcx.Gcx.fullcollect() (this=...) at gc/gcx.d:2427
> >>> #3  gc.gcx.Gcx.bigAlloc() (this=..., size=16401, poolPtr=0x7fc3d4bfe3c8, alloc_size=0x7fc3d4bfe418) at gc/gcx.d:2099
> >>> #4  gc.gcx.GC.mallocNoSync (alloc_size=0x7fc3d4bfe418, bits=10, size=16401, this=...) gc/gcx.d:503
> >>> #5  gc.gcx.GC.malloc() (this=..., size=16401, bits=10, alloc_size=0x7fc3d4bfe418) gc/gcx.d:421
> >>> #6  gc.gc.gc_qalloc (ba=10, sz=<optimized out>) gc/gc.d:203
> >>> #7  gc_qalloc (sz=<optimized out>, ba=10) gc/gc.d:198
> >>> #8  _d_newarrayT (ti=..., length=4096) rt/lifetime.d:807
> >>> #9  sequencer.algorithm.gzip.HuffmanTree.__T6__ctorTG32hZ.__ctor() (this=..., bitLengths=...) sequencer/algorithm/gzip.d:444
> >>> 
> >>> Two more threads are alive, but waiting on a condition variable (i.e.: in pthread_cond_wait(), but from my own and not from druntime code. Is there some obvious way I could have dead-locked the GC ? Or is there a bug ?
> >> 
> >> I assume you're running on Linux, which uses signals (SIGUSR1, specifically) to suspend threads for a collection.  So I imagine what's happening is that your thread is trying to suspend all the other threads so it can collect, and those threads are ignoring the signal for some reason.  I would expect pthread_cond_wait to be interrupted if a signal arrives though.  Have you overridden the signal handler for SIGUSR1?
> > 
> > No, I have not overridden the signal handler. I'm aware of the fact that signals make pthread_cond_wait() return early and put them in a while loop as one would expect, that is all.
> 
> Hrm... Can you trap this in a debugger and post the stack traces of all threads?  That stack above is a thread waiting for others to say they're suspended so it can collect.

I could do that (with a little work setting the scenario up again), but it wont help. As I said, the other two threads were paused in pthread_cond_wait() in my own code. There was nothing special about their stack trace.

-- 
Marco