That patch series introduced new virtual clock type for use in external
subsystems. It breaks desired behavior in non-record/replay usage
scenarios due to a small change to existing behavior. Processing of
virtual timers belonging to new clock type is kicked off to the main
loop, which makes these timers asynchronous with vCPU thread and,
in icount mode, with whole guest execution. This breaks expected
determinism in non-record/replay icount mode of emulation where these
"external subsystems" are isolated from the host (i.e. they are
external only to guest core, not to the entire emulation environment).
Example for slirp ("user" backend for network device):
User runs qemu in icount mode with rtc clock=vm without any external
communication interfaces but with "-netdev user,restrict=on". It expects
deterministic execution, because network services are emulated inside
qemu and isolated from host. There are no reasons to get reply from DHCP
server with different delay or something like that.
The next patches revert reimplements the same changes in a better way.
This reverts commit 87f4fe7653.
This reverts commit 775a412bf8.
This reverts commit 9888091404.
Signed-off-by: Artem Pisarenko <artem.k.pisarenko@gmail.com>
Message-Id: <18b1e7c8f155fe26976f91be06bde98eef6f8751.1539764043.git.artem.k.pisarenko@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Slirp and VNC modules use virtual clock for processing some events that
are related to the guest execution speed.
But virtual clock-related events are consideres to be deterministic and
are recorded/replayed by icount mechanism. But slirp and VNC lie outside
the recorded guest core (which includes CPU and peripherals).
Therefore slirp and VNC are external for the guest, but should work at
guest speed.
This patch introduces new virtual clock which can be used for external
subsystems for running timers that are synchronized with the guest.
Signed-off-by: Pavel Dovgalyuk <Pavel.Dovgaluk@ispras.ru>
Message-Id: <20180912082002.3228.82417.stgit@pasha-VirtualBox>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Ciro Santilli reported that commit a5ed352596
breaks the execution replay. It happens due to the probing the clock
for the new instances of iothread.
However, this probing was made in replay mode for the timer lists that
are empty.
This patch removes clock probing in replay mode.
It is an artifact of the old version with another thread model.
Signed-off-by: Pavel Dovgalyuk <Pavel.Dovgaluk@ispras.ru>
Message-Id: <20180725121526.12867.17866.stgit@pasha-VirtualBox>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This patch adds saving/restoring of the host clock field 'last'.
It is used in host clock calculation and therefore clock may
become incorrect when using restored vmstate.
Signed-off-by: Pavel Dovgalyuk <pavel.dovgaluk@ispras.ru>
Acked-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20180227095226.1060.50975.stgit@pasha-VirtualBox>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Pavel Dovgalyuk <Pavel.Dovgaluk@ispras.ru>
icount has become much slower after tcg_cpu_exec has stopped
using the BQL. There is also a latent bug that is masked by
the slowness.
The slowness happens because every occurrence of a QEMU_CLOCK_VIRTUAL
timer now has to wake up the I/O thread and wait for it. The rendez-vous
is mediated by the BQL QemuMutex:
- handle_icount_deadline wakes up the I/O thread with BQL taken
- the I/O thread wakes up and waits on the BQL
- the VCPU thread releases the BQL a little later
- the I/O thread raises an interrupt, which calls qemu_cpu_kick
- the VCPU thread notices the interrupt, takes the BQL to
process it and waits on it
All this back and forth is extremely expensive, causing a 6 to 8-fold
slowdown when icount is turned on.
One may think that the issue is that the VCPU thread is too dependent
on the BQL, but then the latent bug comes in. I first tried removing
the BQL completely from the x86 cpu_exec, only to see everything break.
The only way to fix it (and make everything slow again) was to add a dummy
BQL lock/unlock pair.
This is because in -icount mode you really have to process the events
before the CPU restarts executing the next instruction. Therefore, this
series moves the processing of QEMU_CLOCK_VIRTUAL timers straight in
the vCPU thread when running in icount mode.
The required changes include:
- make the timer notification callback wake up TCG's single vCPU thread
when run from another thread. By using async_run_on_cpu, the callback
can override all_cpu_threads_idle() when the CPU is halted.
- move handle_icount_deadline after qemu_tcg_wait_io_event, so that
the timer notification callback is invoked after the dummy work item
wakes up the vCPU thread
- make handle_icount_deadline run the timers instead of just waking the
I/O thread.
- stop processing the timers in the main loop
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
There is no change for now, because the callback just invokes
qemu_notify_event.
Reviewed-by: Edgar E. Iglesias <edgar.iglesias@xilinx.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This dependency is the wrong way, and we will need util/qemu-timer.h from
sysemu/cpus.h in the next patch.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Edgar E. Iglesias <edgar.iglesias@xilinx.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
If the first timer is exactly at the current value of the clock, the
deadline is met and the timer should fire. This fixes itself on the next
iteration of the loop without icount; with icount, however, execution
of instructions will stop exactly at the deadline and won't proceed.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Edgar E. Iglesias <edgar.iglesias@xilinx.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Similarly to allocation, do it from an inline function. This allows
tests to only use the headers for allocation/free of timer.
Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
AioContext is fairly self contained, the only dependency is QEMUTimer but
that in turn doesn't need anything else. So move them out of block-obj-y
to avoid introducing a dependency from io/ to block-obj-y.
main-loop and its dependency iohandler also need to be moved, because
later in this series io/ will call iohandler_get_aio_context.
[Changed copyright "the QEMU team" to "other QEMU contributors" as
suggested by Daniel Berrange and agreed by Paolo.
--Stefan]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Message-id: 20170213135235.12274-2-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>