Created attachment 353967 [details]
Test program to reproduce the bug.

Description of problem:
Reads scheduled with libaio sometimes fail when fork()ing simultaneously despite success is being reported.

Version-Release number of selected component (if applicable):
Checked on Centos 5.2.
uname -a
Linux w-50 2.6.18-128.2.1.el5 #1 SMP Tue Jul 14 06:36:37 EDT 2009 x86_64 x86_64 x86_64 GNU/Linux
libaio: 0.3.106
and on
Red Hat Enterprise Linux Server release 5.1 (Tikanga)
uname -a:
Linux beta12 2.6.18-92.1.6.1 #1 SMP Mon Aug 25 12:16:35 EDT 2008 x86_64 x86_64 x86_64 GNU/Linux
libaio: 0.3.106

How reproducible:
Interleave fork()'s with submitting reads through libaio. There is a test program attached, which reproduces the bug.

Its input should be a 4Gb file filled with zeros. It launches 2 new threads, which do:
1:
keeps calling fork() and wait()

2:
allocates a buffer, fills it with 0xff and submits reads from the test file with a callback set on completion (loops until the end of file)

3:
calls io_queue_run() in a loop to launch callbacks; callbacks check if (in case of no error reported) the buffer contains zeros and exit()s if the condition is not met (with a "corruption found" message")

Steps to Reproduce:
1. gcc -o aiotest aiotest.c -lpthread -laio
2. dd if=/dev/zero of=test bs=1048576 count=4096 oflag=direct
3. ./aiotest test
  
Actual results:
Terminates with a message "corruption found" (after a nondeterministic amount of time, usually a second).

Expected results:
Having read the whole 4Gb it should fall into an infinite loop forking and waiting.

Additional info:
When the loop with fork()'s is replaced with an empty loop, everything works correctly.