| Message ID | 20180703085250.28380.33220.stgit@pasha-VirtualBox (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
There are some situations when this patch still doesn't help.
I think this happens due to the race condition in qemu_tcg_rr_wait_io_event
static void qemu_tcg_rr_wait_io_event(CPUState *cpu)
{
while (all_cpu_threads_idle()) {
stop_tcg_kick_timer();
qemu_cond_wait(cpu->halt_cond, &qemu_global_mutex);
}
start_tcg_kick_timer();
qemu_wait_io_event_common(cpu);
}
all_cpu_threads_idle() returns true when there is no queued work.
But between this call and qemu_cond_wait() iothread may add queued work
and the vCPU thread will sleep infinitely.
Does anyone have an idea how to fix this?
Pavel Dovgalyuk
> -----Original Message-----
> From: Pavel Dovgalyuk [mailto:Pavel.Dovgaluk@ispras.ru]
> Sent: Tuesday, July 03, 2018 11:53 AM
> To: qemu-devel@nongnu.org
> Cc: alex.bennee@linaro.org; pbonzini@redhat.com; maria.klimushenkova@ispras.ru;
> dovgaluk@ispras.ru; pavel.dovgaluk@ispras.ru
> Subject: [PATCH] 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>
> ---
> cpus.c | 15 ++++++++++-----
> 1 file changed, 10 insertions(+), 5 deletions(-)
>
> diff --git a/cpus.c b/cpus.c
> index 181ce33..bad6a33 100644
> --- a/cpus.c
> +++ b/cpus.c
> @@ -539,11 +539,6 @@ void qemu_start_warp_timer(void)
> return;
> }
>
> - /* warp clock deterministically in record/replay mode */
> - if (!replay_checkpoint(CHECKPOINT_CLOCK_WARP_START)) {
> - return;
> - }
> -
> if (!all_cpu_threads_idle()) {
> return;
> }
> @@ -553,6 +548,16 @@ void qemu_start_warp_timer(void)
> return;
> }
>
> + /* warp clock deterministically in record/replay mode */
> + if (!replay_checkpoint(CHECKPOINT_CLOCK_WARP_START)) {
> + /* vCPU is sleeping and warp can't be started.
> + It is probably a race condition: notification sent
> + to vCPU was processed in advance and vCPU went to sleep.
> + Therefore we have to wake it up for doing someting. */
> + qemu_clock_notify(QEMU_CLOCK_VIRTUAL);
> + return;
> + }
> +
> /* We want to use the earliest deadline from ALL vm_clocks */
> clock = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL_RT);
> deadline = qemu_clock_deadline_ns_all(QEMU_CLOCK_VIRTUAL);
On 09/07/2018 13:24, Pavel Dovgalyuk wrote: > static void qemu_tcg_rr_wait_io_event(CPUState *cpu) > { > while (all_cpu_threads_idle()) { > stop_tcg_kick_timer(); > qemu_cond_wait(cpu->halt_cond, &qemu_global_mutex); > } > > start_tcg_kick_timer(); > > qemu_wait_io_event_common(cpu); > } > > all_cpu_threads_idle() returns true when there is no queued work. > But between this call and qemu_cond_wait() iothread may add queued work > and the vCPU thread will sleep infinitely. Maybe queue_work_on_cpu is called outside BQL? Paolo
> From: Paolo Bonzini [mailto:pbonzini@redhat.com] > On 09/07/2018 13:24, Pavel Dovgalyuk wrote: > > static void qemu_tcg_rr_wait_io_event(CPUState *cpu) > > { > > while (all_cpu_threads_idle()) { > > stop_tcg_kick_timer(); > > qemu_cond_wait(cpu->halt_cond, &qemu_global_mutex); > > } > > > > start_tcg_kick_timer(); > > > > qemu_wait_io_event_common(cpu); > > } > > > > all_cpu_threads_idle() returns true when there is no queued work. > > But between this call and qemu_cond_wait() iothread may add queued work > > and the vCPU thread will sleep infinitely. > > Maybe queue_work_on_cpu is called outside BQL? I don't remember now. However, rr series includes the better version of that patch. Pavel Dovgalyuk
On 11/09/2018 09:37, Pavel Dovgalyuk wrote: >> From: Paolo Bonzini [mailto:pbonzini@redhat.com] >> On 09/07/2018 13:24, Pavel Dovgalyuk wrote: >>> static void qemu_tcg_rr_wait_io_event(CPUState *cpu) >>> { >>> while (all_cpu_threads_idle()) { >>> stop_tcg_kick_timer(); >>> qemu_cond_wait(cpu->halt_cond, &qemu_global_mutex); >>> } >>> >>> start_tcg_kick_timer(); >>> >>> qemu_wait_io_event_common(cpu); >>> } >>> >>> all_cpu_threads_idle() returns true when there is no queued work. >>> But between this call and qemu_cond_wait() iothread may add queued work >>> and the vCPU thread will sleep infinitely. >> >> Maybe queue_work_on_cpu is called outside BQL? > > I don't remember now. > However, rr series includes the better version of that patch. Good, thanks. Paolo
diff --git a/cpus.c b/cpus.c index 181ce33..bad6a33 100644 --- a/cpus.c +++ b/cpus.c @@ -539,11 +539,6 @@ void qemu_start_warp_timer(void) return; } - /* warp clock deterministically in record/replay mode */ - if (!replay_checkpoint(CHECKPOINT_CLOCK_WARP_START)) { - return; - } - if (!all_cpu_threads_idle()) { return; } @@ -553,6 +548,16 @@ void qemu_start_warp_timer(void) return; } + /* warp clock deterministically in record/replay mode */ + if (!replay_checkpoint(CHECKPOINT_CLOCK_WARP_START)) { + /* vCPU is sleeping and warp can't be started. + It is probably a race condition: notification sent + to vCPU was processed in advance and vCPU went to sleep. + Therefore we have to wake it up for doing someting. */ + qemu_clock_notify(QEMU_CLOCK_VIRTUAL); + return; + } + /* We want to use the earliest deadline from ALL vm_clocks */ clock = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL_RT); deadline = qemu_clock_deadline_ns_all(QEMU_CLOCK_VIRTUAL);
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> --- cpus.c | 15 ++++++++++----- 1 file changed, 10 insertions(+), 5 deletions(-)