Re: [Qemu-devel] [RFC PATCH] ui/console: ensure graphic updates don't race with TCG vCPUs

[Laurent Desnogues]

On Wed, Mar 22, 2017 at 5:28 PM, Alex Bennée <address@hidden> wrote:
>
> Laurent Desnogues <address@hidden> writes:
>
>> Hi Alex,
>>
>> this patch breaks:
>>
>>   http://wiki.qemu.org/download/arm-test-0.2.tar.gz
>>
>> qemu-system-arm -kernel zImage.integrator -initrd arm_root.img
>> -append "console=ttyAMA0" -machine integratorcp -serial stdio -icount
>> 0
>> Uncompressing 
>> Linux..........................................................................
>> done, booting the kernel.
>> qemu-system-arm: /work/qemu/qemu/memory.c:920:
>> memory_region_transaction_commit: Assertion
>> `qemu_mutex_iothread_locked()' failed.
>
> Doh - too many levels of BQL. So wi->exclusive tasks drop the iothread
> so they don't deadlock as any other threads try and wake up and get to
> cpu_exec_end. Of course the update was originally under BQL so the
> deferred work should be as well. We have to make a minor tweak to avoid
> triggering RCU clean-up outside the execution context though. Try this:
>
> ui/console: ensure do_safe_dpy_refresh holds BQL
>
> I missed the fact that when an exclusive work item runs it drops the
> BQL to ensure all no vCPUs are stuck waiting for it, hence causing a
> deadlock. However the actual helper needs to take the BQL especially
> as we'll be messing with device emulation bits during the update which
> all assume BQL is held.
>
> We make a minor cpu_reloading_memory_map which must try and unlock the
> RCU if we are actually outside the running context.
>
> Reported-by: Laurent Desnogues <address@hidden>
> CC: Gerd Hoffmann <address@hidden>
> CC: Paolo Bonzini <address@hidden>
> Signed-off-by: Alex Bennée <address@hidden>

It works fine on the image I reported and another AArch64 one.

Tested-by: Laurent Desnogues <address@hidden>

Thanks for the quick fix,

Laurent

> 2 files changed, 3 insertions(+), 1 deletion(-)
> cpu-exec-common.c | 2 +-
> ui/console.c      | 2 ++
>
> modified   cpu-exec-common.c
> @@ -35,7 +35,7 @@ void cpu_loop_exit_noexc(CPUState *cpu)
>  #if defined(CONFIG_SOFTMMU)
>  void cpu_reloading_memory_map(void)
>  {
> -    if (qemu_in_vcpu_thread()) {
> +    if (qemu_in_vcpu_thread() && current_cpu->running) {
>          /* The guest can in theory prolong the RCU critical section as long
>           * as it feels like. The major problem with this is that because it
>           * can do multiple reconfigurations of the memory map within the
> modified   ui/console.c
> @@ -1586,7 +1586,9 @@ bool dpy_gfx_check_format(QemuConsole *con,
>  static void do_safe_dpy_refresh(CPUState *cpu, run_on_cpu_data opaque)
>  {
>      DisplayChangeListener *dcl = opaque.host_ptr;
> +    qemu_mutex_lock_iothread();
>      dcl->ops->dpy_refresh(dcl);
> +    qemu_mutex_unlock_iothread();
>  }
>
>  static void dpy_refresh(DisplayState *s)