[Bug go/56172] New: net FAILs on Solaris

[Bug go/56172] New: net FAILs on Solaris

[ro at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56172

             Bug #: 56172
           Summary: net FAILs on Solaris
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: go
        AssignedTo: ian@airs.com
        ReportedBy: ro@gcc.gnu.org
              Host: *-*-solaris2.*
            Target: *-*-solaris2.*
             Build: *-*-solaris2.*


The libgo net test currently FAILs on Solaris.  With -test.v=true I find

=== RUN TestAddFDReturnsError
panic: runtime error: invalid memory address or nil pointer dereference
[signal 0xb code=0x1 addr=0xe]

goroutine 19 [running]:
created by runtime_MHeap_Scavenger
        /vol/gcc/src/hg/trunk/local/libgo/runtime/mheap.c:408

goroutine 1 [chan receive]:
[...]

Running the test with -test.run=TestAddFDReturnsError under gdb reveals:

Program received signal SIGSEGV, Segmentation fault.
[Switching to Thread 3 (LWP 3)]
0xff23db2c in sem_invalid () from /lib/libc.so.1
(gdb) where
#0  0xff23db2c in sem_invalid () from /lib/libc.so.1
#1  0xff23dbf0 in sem_post () from /lib/libc.so.1
#2  0xfdfa1a5c in runtime_semawakeup (mp=0xddc0cf38)
    at /vol/gcc/src/hg/trunk/local/libgo/runtime/thread-sema.c:125
#3  0xfdfa8f48 in kickoff ()
    at /vol/gcc/src/hg/trunk/local/libgo/runtime/proc.c:369
#4  0xff1cb730 in __makecontext_v2 () from /lib/libc.so.1

  sem_invalid () does

    if (sem->sem_magic != SEMA_MAGIC) {
        errno = EINVAL;
        return (-1);
    }

I haven't yet debugged how this happens.

  Rainer

[Bug go/56172] net FAILs on Solaris

[ian at airs dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56172

Ian Lance Taylor <ian at airs dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-02-02
     Ever Confirmed|0                           |1

--- Comment #1 from Ian Lance Taylor <ian at airs dot com> 2013-02-02 16:08:32 UTC ---
The error suggests that runtime_semawakeup was called with an m whose waitsema
field is NULL.  There should be no way that kickoff can call runtime_semawakeup
directly, but it could be happening via a tailcall at the end of the function
that some goroutine was started to run.  Possible tailcalls could come from
runtime_unlock or runtime_notewakeup.  The stack trace shows that the goroutine
was started at mheap.c:408, and indeed that line starts a goroutine that runs 
forcegchelper, and indeed forcegchelper ends with a tailcall to
runtime_notewakeup.  runtime_notewakeup will look at the note's waitm field. 
In this case the waitm field must be not NULL, and it must point to an m whose
waitsema field is NULL.

Unfortunately I don't see any way that could happen.  The only m that could be
in the note is the one running runtime_MHeap_Scavenger.  And that could only be
in the note if it calls runtime_notesleep before forcegchelper gets around to
calling runtime_notewakeup.  And runtime_notesleep initializes the waitsema
field before it puts the m in the note.

Since you can recreate the bug, I guess the next step is to check the
mp->waitsema field in the runtime_semawakeup frame.  My prediction is that it
is 0.  But I'm not sure where to go from there since I don't see how it could
be 0.

And by the way I bet this is hiding something else, since the scavenger should
only run if nothing much else is happening, so once we fix this problem I
expect that the test will time out for some reason.

[Bug go/56172] net FAILs on Solaris

[ro at CeBiTec dot Uni-Bielefeld.DE]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56172

--- Comment #2 from ro at CeBiTec dot Uni-Bielefeld.DE <ro at CeBiTec dot Uni-Bielefeld.DE> 2013-02-04 13:00:59 UTC ---
Thanks for the analysis.

> Since you can recreate the bug, I guess the next step is to check the
> mp->waitsema field in the runtime_semawakeup frame.  My prediction is that it
> is 0.  But I'm not sure where to go from there since I don't see how it could
> be 0.

In my case, mp->waitsema is 8.

In case you want to have a look yourself (which will be way faster), you
can find the build in
apoc:/var/gcc/regression/trunk/11-gcc-gas/build/sparc-sun-solaris2.11/libgo/net-check/test.

    Rainer

[Bug go/56172] net FAILs on Solaris

[ian at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56172

--- Comment #3 from ian at gcc dot gnu.org <ian at gcc dot gnu.org> 2013-02-05 23:59:28 UTC ---
Author: ian
Date: Tue Feb  5 23:59:24 2013
New Revision: 195774

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=195774
Log:
    PR go/56172
runtime: Fix argument passed to forcegchelper.

Modified:
    trunk/libgo/runtime/mheap.c

[Bug go/56172] net FAILs on Solaris

[ian at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56172

--- Comment #4 from ian at gcc dot gnu.org <ian at gcc dot gnu.org> 2013-02-06 22:40:26 UTC ---
Author: ian
Date: Wed Feb  6 22:40:18 2013
New Revision: 195823

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=195823
Log:
    PR go/56172
net: Fixes for select based pollster.

Make Close work properly, mainly for testing.  Restart the
select if a descriptor is closed.

Modified:
    trunk/libgo/go/net/fd_bsd.go
    trunk/libgo/go/net/fd_linux.go
    trunk/libgo/go/net/fd_select.go
    trunk/libgo/go/net/fd_unix.go

[Bug go/56172] net FAILs on Solaris

[ian at airs dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56172

Ian Lance Taylor <ian at airs dot com> changed:

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

--- Comment #5 from Ian Lance Taylor <ian at airs dot com> 2013-02-06 22:41:54 UTC ---
The patch to mheap.c should fix this segmentation violation.

As expected, there was another underlying problem, which should be fixed by the
patch to the net package.

[Bug go/56172] net FAILs on Solaris

[ian at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56172

--- Comment #6 from ian at gcc dot gnu.org <ian at gcc dot gnu.org> 2013-02-07 17:04:37 UTC ---
Author: ian
Date: Thu Feb  7 17:04:24 2013
New Revision: 195855

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=195855
Log:
    PR go/56172
net: Skip TestMulticastListener on Solaris

>From Rainer Orth.

Modified:
    trunk/libgo/go/net/multicast_posix_test.go