Created attachment 80949 [details]
Demo test prog.

Running the attached test2.c, valgrind typically fails as follows:

$ gcc -O0 -g test2.c -lpthread && /var/tmp/valgrind/bin/valgrind --vgdb-error=1 ./a.out
==23755== Memcheck, a memory error detector
==23755== Copyright (C) 2002-2012, and GNU GPL'd, by Julian Seward et al.
==23755== Using Valgrind-3.9.0.SVN and LibVEX; rerun with -h for copyright info
==23755== Command: ./a.out
==23755== 
==23755== 
==23755== TO DEBUG THIS PROCESS USING GDB: start GDB like this
==23755==   /path/to/gdb ./a.out
==23755== and then give GDB the following command
==23755==   target remote | /var/tmp/valgrind/lib/valgrind/../../bin/vgdb --pid=23755
==23755== --pid is optional if only one valgrind process is running
==23755== 
==23755== Invalid write of size 8
==23755==    at 0x4006B7: __yell (test2.c:16)
==23755==    by 0x40076C: main (test2.c:30)
==23755==  Address 0xffeffeed0 is on thread 1's stack
==23755== 
==23755== (action on error) vgdb me ... 

(gdb) print &buf
$6 = (char (*)[256]) 0xffeffeed0

(gdb) monitor get_vbits 0xffeffeed0 256
________ ________ ________ ________ ________ ________ ________ ________
________ ________ ________ ________ ________ ________ ________ ________
________ ________ ________ ________ ________ ________ ________ ________
________ ________ ________ ________ ________ ________ ________ ________
________ ________ ________ ________ ffffffff ffffffff ffffffff ffffffff
ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff
ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff
ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff
Address 0xFFEFFEED0 len 256 has 144 bytes unaddressable

It seems to depend on:
 - Some (small) number of threads being spawned.
 - A > page-sized alloca().
 - Reasonably sized memset on top.
 - It's always the main thread which suffers.

Seen with valgrind 3.7, 3.8.1 and yesterday's SVN. 

Got one comment from email:

--snip--
From: 	John Reiser <jreiser@...>
To: 	valgrind-users@lists.sourceforge.net
Subject: 	Re: [Valgrind-users] threads vs main and invalid stack writes
Date: 	Thu, 04 Jul 2013 11:07:30 -0700

>     alloca(4096);
>     __yell();

> 
> (gdb) monitor get_vbits 0xffeffeed0 256
> ________ ________ ________ ________ ________ ________ ________ ________
> ________ ________ ________ ________ ________ ________ ________ ________
> ________ ________ ________ ________ ________ ________ ________ ________
> ________ ________ ________ ________ ________ ________ ________ ________
> ________ ________ ________ ________ ffffffff ffffffff ffffffff ffffffff
> ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff
> ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff
> ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff
> Address 0xFFEFFEED0 len 256 has 144 bytes unaddressable
> 
> Any ideas? It seems to depend on:
> 
>  - Some (small) number of threads being spawned.
>  - A > page-sized alloca().
>  - Reasonably sized memset on top.
>  - It's always the main thread which suffers.

Thank you for the small, reproducible testcase!

It's a bug in valgrind, so please file a bug report, and include the testcase.
See the "Bug Reports" entry in the left column of the main page
http://www.valgrind.org/ .

Meanwhile the trivial workaround is to memset every result of alloca.
[..]
--snip--