[Bug libgcj/27170] New: Deadlock in garbage collector

[Bug libgcj/27170] New: Deadlock in garbage collector

[green at redhat dot com]

I believe I've identified a deadlock that is periodically reproducible
with Eclipse running on FC5.  I'll attach the stack trace to this bug
report.  See threads 4 and 13.

1. Thread 4 aquires the GC lock and starts a GC
2. Thread 13 dlopens a .so, acquiring a glibc maintained lock
3. Thread 13 tries to acquire the GC lock while registering classes
4. Thread 4 tries to acquire the glibc lock in dl_iterate_phdr

...and we're deadlocked.


-- 
           Summary: Deadlock in garbage collector
           Product: gcc
           Version: 4.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: libgcj
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: green at redhat dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27170

[Bug libgcj/27170] Deadlock in garbage collector

[green at redhat dot com]

------- Comment #1 from green at redhat dot com  2006-04-14 18:58 -------
Created an attachment (id=11273)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=11273&action=view)
stack traces from gdb.  


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27170

[Bug libgcj/27170] Deadlock in garbage collector

[Hans dot Boehm at hp dot com]

------- Comment #2 from Hans dot Boehm at hp dot com  2006-04-14 20:51 -------
Based on the thread 13 stack trace, it looks to me like we're calling dlopen
directly, when we should somehow be arranging to call GC_dlopen.  GC_dlopen is
included in the GC to avoid this sort of deadlock.

(The way GC_dlopen does this is by waiting for, and then disabling, GC during
dlopen.  That's not a great solution, since it can cause unneeded heap growth
in the presence of frequent dlopens.  It should perhaps conditionally trigger a
GC before disabling it.  But in any case, calling GC_dlopen is much better than
generating a deadlock.)


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27170

[Bug libgcj/27170] Deadlock in garbage collector

[green at redhat dot com]

------- Comment #3 from green at redhat dot com  2006-04-14 21:01 -------
(In reply to comment #2)
> Based on the thread 13 stack trace, it looks to me like we're calling dlopen
> directly, when we should somehow be arranging to call GC_dlopen.  GC_dlopen is
> included in the GC to avoid this sort of deadlock.

Thanks Hans.  I'll try making this change to see what happens.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27170

[Bug libgcj/27170] Deadlock in garbage collector

[green at redhat dot com]

------- Comment #4 from green at redhat dot com  2006-04-17 14:43 -------
Created an attachment (id=11283)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=11283&action=view)
Sample patch


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27170

[Bug libgcj/27170] Deadlock in garbage collector

[green at redhat dot com]

------- Comment #5 from green at redhat dot com  2006-04-17 14:44 -------
I've uploaded a patch that I've applied to the FC5 gcc RPM.  I can't reproduce
the problem with this patch.


-- 

green at redhat dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |green at redhat dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27170

[Bug libgcj/27170] Deadlock in garbage collector

[tromey at gcc dot gnu dot org]

------- Comment #6 from tromey at gcc dot gnu dot org  2006-04-17 16:00 -------
We could either use <gc.h> here (see ltdl.c) or we could
use libltdl instead of directly using dlopen...


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27170

[Bug libgcj/27170] Deadlock in garbage collector

[bryce at gcc dot gnu dot org]

------- Comment #7 from bryce at gcc dot gnu dot org  2006-04-20 23:48 -------
Subject: Bug 27170

Author: bryce
Date: Thu Apr 20 23:47:56 2006
New Revision: 113123

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=113123
Log:
        PR libgcj/27170
        * gnu/gcj/runtime/natSharedLibLoader.cc: Include gc.h to override
        dlopen(). From Anthony Green.

Modified:
    trunk/libjava/ChangeLog
    trunk/libjava/gnu/gcj/runtime/natSharedLibLoader.cc


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27170

[Bug libgcj/27170] Deadlock in garbage collector

[bryce at gcc dot gnu dot org]

------- Comment #8 from bryce at gcc dot gnu dot org  2006-04-20 23:52 -------
Subject: Bug 27170

Author: bryce
Date: Thu Apr 20 23:52:14 2006
New Revision: 113124

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=113124
Log:
        PR libgcj/27170
        * gnu/gcj/runtime/natSharedLibLoader.cc: Include gc.h to override
        dlopen(). From Anthony Green.

Modified:
    branches/gcc-4_1-branch/libjava/ChangeLog
    branches/gcc-4_1-branch/libjava/gnu/gcj/runtime/natSharedLibLoader.cc


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27170

[Bug libgcj/27170] Deadlock in garbage collector

[mckinlay at redhat dot com]

------- Comment #9 from mckinlay at redhat dot com  2006-04-20 23:55 -------
I have checked in the fix to the trunk and gcc-4_1-branch.


-- 

mckinlay at redhat dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |FIXED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27170