Bug 595499

+++ This bug was initially created as a clone of Bug #459901 +++

I've found a race condition between AIO read/write and setresuid(). The
race results in AIO in Samba sometimes not completing, leading to stuck clients.

You can grab a little standalone test program that demonstrates the bug here:

  http://samba.org/~tridge/junkcode/aio_uid.c

Here is the description of the bug from that code:

  The race condition is in setresuid(), which in glibc tries to be
  smart about threads and change the euid of threads when the euid of
  the main program changes. The problem is that this makes setresuid()
  non-atomic, which means that if an IO completes during the complex
  series of system calls that setresuid() becomes, then the thread
  completing the IO may get -1/EPERM back from the rt_sigqueueinfo()
  call that it uses to notify its parent of the completing IO. In that
  case two things happen:

    1) the signal is never delivered, so the caller never is told that
    the IO has completed

    2) if the caller polls for completion using aio_error() then it
    will see a -1/EPERM result, rather than the real result of the IO

  The simplest fix in existing code that mixes uid changing with AIO
  (such as Samba) is to not use setresuid() and use setreuid()
  instead, which in glibc doesn't try to play any games with the euid
  of threads. That does mean that you will need to manually gain root
  privileges before calling aio_read() or aio_write() to ensure that
  the thread has permission to send signals to the main thread

If you strace the above program then the bug should be fairly clear.

I've reproduced this bug on RHEL 5.2, but I expect it will be on all
current Linux distros. I've also reproduced it on Ubuntu Hardy.

I've added a workaround to the Samba 3.2 tree, but it will take a while
before this workaround gets to end users. I imagine there are probably
other applications that are affected.

Cheers, Tridge