]> git.ipfire.org Git - thirdparty/qemu.git/commit
replay: wake up vCPU when replaying
authorPavel Dovgalyuk <Pavel.Dovgaluk@ispras.ru>
Wed, 12 Sep 2018 08:19:45 +0000 (11:19 +0300)
committerPaolo Bonzini <pbonzini@redhat.com>
Tue, 2 Oct 2018 17:08:51 +0000 (19:08 +0200)
commit0c08185f8fe1eb20edec1a2bf32b4d219cc023f0
treedab130711a48942f7179666888316612a1e725c0
parent0a7fa00a13f0852ec6fa83ab987a5ee7978d9867
replay: wake up vCPU when replaying

In record/replay icount mode vCPU thread and iothread synchronize
the execution using the checkpoints.
vCPU thread processes the virtual timers and iothread processes all others.
When iothread wants to wake up sleeping vCPU thread, it sends dummy queued
work. Therefore it could be the following sequence of the events in
record mode:
 - IO: sending dummy work
 - IO: processing timers
 - CPU: wakeup
 - CPU: clearing dummy work
 - CPU: processing virtual timers

But due to the races in replay mode the sequence may change:
 - IO: sending dummy work
 - CPU: wakeup
 - CPU: clearing dummy work
 - CPU: sleeping again because nothing to do
 - IO: Processing timers
 - CPU: zzzz

In this case vCPU will not wake up, because dummy work is not to be set up
again.

This patch tries to wake up the vCPU when it sleeps and the icount warp
checkpoint isn't met. It means that vCPU has something to do, because
there are no other reasons of non-matching warp checkpoint.

Signed-off-by: Pavel Dovgalyuk <Pavel.Dovgaluk@ispras.ru>
--

v5: improve checking that vCPU is still sleeping
Message-Id: <20180912081945.3228.19776.stgit@pasha-VirtualBox>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
cpus.c
include/sysemu/replay.h
replay/replay.c