]> www.infradead.org Git - users/dwmw2/qemu.git/commit
coroutine-win32.c: Add noinline attribute to work around gcc bug
authorPeter Maydell <peter.maydell@linaro.org>
Mon, 23 Jun 2014 14:55:03 +0000 (15:55 +0100)
committerMichael Roth <mdroth@linux.vnet.ibm.com>
Wed, 16 Jul 2014 00:28:01 +0000 (19:28 -0500)
commit8a93721d04a55b3f23d1594287fe812da01c0d31
tree6a68141380092025b09613c39b3fd65637f2593c
parentb47506f55cf4fb01d04e3c76c77ca09b75cf75c6
coroutine-win32.c: Add noinline attribute to work around gcc bug

A gcc codegen bug in x86_64-w64-mingw32-gcc (GCC) 4.6.3 means that
non-debug builds of QEMU for Windows tend to assert when using
coroutines. Work around this by marking qemu_coroutine_switch
as noinline.

If we allow gcc to inline qemu_coroutine_switch into
coroutine_trampoline, then it hoists the code to get the
address of the TLS variable "current" out of the while() loop.
This is an invalid transformation because the SwitchToFiber()
call may be called when running thread A but return in thread B,
and so we might be in a different thread context each time
round the loop. This can happen quite often.  Typically.
a coroutine is started when a VCPU thread does bdrv_aio_readv:

     VCPU thread

     main VCPU thread coroutine      I/O coroutine
        bdrv_aio_readv ----->
                                     start I/O operation
                                       thread_pool_submit_co
                       <------------ yields
        back to emulation

Then I/O finishes and the thread-pool.c event notifier triggers in
the I/O thread.  event_notifier_ready calls thread_pool_co_cb, and
the I/O coroutine now restarts *in another thread*:

     iothread

     main iothread coroutine         I/O coroutine (formerly in VCPU thread)
        event_notifier_ready
          thread_pool_co_cb ----->   current = I/O coroutine;
                                     call AIO callback

But on Win32, because of the bug, the "current" being set here the
current coroutine of the VCPU thread, not the iothread.

noinline is a good-enough workaround, and quite unlikely to break in
the future.

(Thanks to Paolo Bonzini for assistance in diagnosing the problem
and providing the detailed example/ascii art quoted above.)

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Message-id: 1403535303-14939-1-git-send-email-peter.maydell@linaro.org
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Richard Henderson <rth@twiddle.net>
(cherry picked from commit ff4873cb8c81db89668d8b56e19e57b852edb5f5)
Signed-off-by: Michael Roth <mdroth@linux.vnet.ibm.com>
coroutine-win32.c