[PATCH 16/18] cpus: clear exit_request in qemu_wait_io_event

Paolo Bonzini posted 18 patches 1 week ago
Maintainers: Richard Henderson <richard.henderson@linaro.org>, Paolo Bonzini <pbonzini@redhat.com>, "Philippe Mathieu-Daudé" <philmd@linaro.org>, Cameron Esfahani <dirty@apple.com>, Roman Bolshakov <rbolshakov@ddn.com>, Phil Dennis-Jordan <phil@philjordan.eu>, Mads Ynddal <mads@ynddal.dk>, Riku Voipio <riku.voipio@iki.fi>, Warner Losh <imp@bsdimp.com>, Kyle Evans <kevans@freebsd.org>, Eduardo Habkost <eduardo@habkost.net>, Marcel Apfelbaum <marcel.apfelbaum@gmail.com>, Yanan Wang <wangyanan55@huawei.com>, Zhao Liu <zhao1.liu@intel.com>, Nicholas Piggin <npiggin@gmail.com>, Chinmay Rath <rathc@linux.ibm.com>, Harsh Prateek Bora <harshpb@linux.ibm.com>, Laurent Vivier <laurent@vivier.eu>, Brian Cain <brian.cain@oss.qualcomm.com>, "Alex Bennée" <alex.bennee@linaro.org>, Peter Maydell <peter.maydell@linaro.org>, Michael Rolnik <mrolnik@gmail.com>, Marcelo Tosatti <mtosatti@redhat.com>, Reinoud Zandijk <reinoud@netbsd.org>, Sunil Muthuswamy <sunilmut@microsoft.com>, Stafford Horne <shorne@gmail.com>, Yoshinori Sato <yoshinori.sato@nifty.com>, David Hildenbrand <david@redhat.com>, Ilya Leoshkevich <iii@linux.ibm.com>, Thomas Huth <thuth@redhat.com>, Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>, Artyom Tarasenko <atar4qemu@gmail.com>
[PATCH 16/18] cpus: clear exit_request in qemu_wait_io_event
Posted by Paolo Bonzini 1 week ago
Make the code common to all accelerators: after seeing cpu->exit_request
set to true, accelerator code needs to reach qemu_wait_io_event_common().

So for the common cases where they use qemu_wait_io_event(), go ahead and
clear it in there.  Note that the cheap qatomic_set() is enough because
at this point the thread has taken the BQL; qatomic_set_mb() is not needed.
In particular, this is the ordering of the communication between
I/O and vCPU threads is always the same.

In the I/O thread:

(a) store other memory locations that will be checked if cpu->exit_request
    or cpu->interrupt_request is 1 (for example cpu->stop or cpu->work_list
    for cpu->exit_request)

(b) cpu_exit(): store-release cpu->exit_request, or
(b) cpu_interrupt(): store-release cpu->interrupt_request

>>> at this point, cpu->halt_cond is broadcast and the BQL released

(c) do the accelerator-specific kick (e.g. write icount_decr for TCG,
    pthread_kill for KVM, etc.)

In the vCPU thread instead the opposite order is respected:

(c) the accelerator's execution loop exits thanks to the kick

(b) then the inner execution loop checks cpu->interrupt_request
    and cpu->exit_request.  If needed cpu->interrupt_request is
    converted into cpu->exit_request when work is needed outside
    the execution loop.

(a) then the other memory locations are checked.  Some may need to
    be read under the BQL, but the vCPU thread may also take other
    locks (e.g. for queued work items) or none at all.

qatomic_set_mb() would only be needed if the halt sleep was done
outside the BQL (though in that case, cpu->exit_request probably
would be replaced by a QemuEvent or something like that).

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 accel/kvm/kvm-all.c          | 2 --
 accel/tcg/cpu-exec.c         | 1 -
 accel/tcg/tcg-accel-ops-rr.c | 9 +++++++--
 accel/tcg/tcg-accel-ops.c    | 2 --
 accel/tcg/user-exec.c        | 1 +
 system/cpus.c                | 1 +
 target/i386/nvmm/nvmm-all.c  | 2 --
 target/i386/whpx/whpx-all.c  | 2 --
 8 files changed, 9 insertions(+), 11 deletions(-)

diff --git a/accel/kvm/kvm-all.c b/accel/kvm/kvm-all.c
index e4167d94b4f..d13156bee87 100644
--- a/accel/kvm/kvm-all.c
+++ b/accel/kvm/kvm-all.c
@@ -3155,7 +3155,6 @@ int kvm_cpu_exec(CPUState *cpu)
     trace_kvm_cpu_exec();
 
     if (kvm_arch_process_async_events(cpu)) {
-        qatomic_set(&cpu->exit_request, 0);
         return EXCP_HLT;
     }
 
@@ -3345,7 +3344,6 @@ int kvm_cpu_exec(CPUState *cpu)
         vm_stop(RUN_STATE_INTERNAL_ERROR);
     }
 
-    qatomic_set(&cpu->exit_request, 0);
     return ret;
 }
 
diff --git a/accel/tcg/cpu-exec.c b/accel/tcg/cpu-exec.c
index 3ae545e888f..ad94f96b252 100644
--- a/accel/tcg/cpu-exec.c
+++ b/accel/tcg/cpu-exec.c
@@ -872,7 +872,6 @@ static inline bool cpu_handle_interrupt(CPUState *cpu,
      * The corresponding store-release is in cpu_exit.
      */
     if (unlikely(qatomic_load_acquire(&cpu->exit_request)) || icount_exit_request(cpu)) {
-        qatomic_set(&cpu->exit_request, 0);
         if (cpu->exception_index == -1) {
             cpu->exception_index = EXCP_INTERRUPT;
         }
diff --git a/accel/tcg/tcg-accel-ops-rr.c b/accel/tcg/tcg-accel-ops-rr.c
index 610292d3bac..e9d291dc391 100644
--- a/accel/tcg/tcg-accel-ops-rr.c
+++ b/accel/tcg/tcg-accel-ops-rr.c
@@ -286,8 +286,13 @@ static void *rr_cpu_thread_fn(void *arg)
         /* Does not need a memory barrier because a spurious wakeup is okay.  */
         qatomic_set(&rr_current_cpu, NULL);
 
-        if (cpu && qatomic_read(&cpu->exit_request)) {
-            qatomic_set_mb(&cpu->exit_request, 0);
+        if (cpu) {
+            /*
+             * This could even reset exit_request for all CPUs, but in practice
+             * races between CPU exits and changes to "cpu" are so rare that
+             * there's no advantage in doing so.
+             */
+            qatomic_set(&cpu->exit_request, false);
         }
 
         if (icount_enabled() && all_cpu_threads_idle()) {
diff --git a/accel/tcg/tcg-accel-ops.c b/accel/tcg/tcg-accel-ops.c
index 1f662a9c745..3bd98005042 100644
--- a/accel/tcg/tcg-accel-ops.c
+++ b/accel/tcg/tcg-accel-ops.c
@@ -82,8 +82,6 @@ int tcg_cpu_exec(CPUState *cpu)
     ret = cpu_exec(cpu);
     cpu_exec_end(cpu);
 
-    qatomic_set_mb(&cpu->exit_request, 0);
-
     return ret;
 }
 
diff --git a/accel/tcg/user-exec.c b/accel/tcg/user-exec.c
index 81906d2e033..8f4f049b924 100644
--- a/accel/tcg/user-exec.c
+++ b/accel/tcg/user-exec.c
@@ -54,6 +54,7 @@ void qemu_cpu_kick(CPUState *cpu)
 
 void qemu_wait_io_event(CPUState *cpu)
 {
+    qatomic_set(&cpu->exit_request, false);
     process_queued_cpu_work(cpu);
 }
 
diff --git a/system/cpus.c b/system/cpus.c
index bb13942cbb7..f989d9938b6 100644
--- a/system/cpus.c
+++ b/system/cpus.c
@@ -463,6 +463,7 @@ void qemu_wait_io_event(CPUState *cpu)
 {
     bool slept = false;
 
+    qatomic_set(&cpu->exit_request, false);
     while (cpu_thread_is_idle(cpu)) {
         if (!slept) {
             slept = true;
diff --git a/target/i386/nvmm/nvmm-all.c b/target/i386/nvmm/nvmm-all.c
index 7e36c42fbb4..ed424251673 100644
--- a/target/i386/nvmm/nvmm-all.c
+++ b/target/i386/nvmm/nvmm-all.c
@@ -817,8 +817,6 @@ nvmm_vcpu_loop(CPUState *cpu)
     cpu_exec_end(cpu);
     bql_lock();
 
-    qatomic_set(&cpu->exit_request, false);
-
     return ret < 0;
 }
 
diff --git a/target/i386/whpx/whpx-all.c b/target/i386/whpx/whpx-all.c
index 00fb7e23100..2a85168ed51 100644
--- a/target/i386/whpx/whpx-all.c
+++ b/target/i386/whpx/whpx-all.c
@@ -2050,8 +2050,6 @@ static int whpx_vcpu_run(CPUState *cpu)
         whpx_last_vcpu_stopping(cpu);
     }
 
-    qatomic_set(&cpu->exit_request, false);
-
     return ret < 0;
 }
 
-- 
2.51.0
Re: [PATCH 16/18] cpus: clear exit_request in qemu_wait_io_event
Posted by Igor Mammedov 4 days, 14 hours ago
On Fri, 29 Aug 2025 17:31:13 +0200
Paolo Bonzini <pbonzini@redhat.com> wrote:

> Make the code common to all accelerators: after seeing cpu->exit_request
> set to true, accelerator code needs to reach qemu_wait_io_event_common().
> 
> So for the common cases where they use qemu_wait_io_event(), go ahead and
> clear it in there.  Note that the cheap qatomic_set() is enough because
> at this point the thread has taken the BQL; qatomic_set_mb() is not needed.
> In particular, this is the ordering of the communication between
> I/O and vCPU threads is always the same.
> 
> In the I/O thread:
> 
> (a) store other memory locations that will be checked if cpu->exit_request
>     or cpu->interrupt_request is 1 (for example cpu->stop or cpu->work_list
>     for cpu->exit_request)
> 
> (b) cpu_exit(): store-release cpu->exit_request, or
> (b) cpu_interrupt(): store-release cpu->interrupt_request
> 
> >>> at this point, cpu->halt_cond is broadcast and the BQL released  
> 
> (c) do the accelerator-specific kick (e.g. write icount_decr for TCG,
>     pthread_kill for KVM, etc.)
> 
> In the vCPU thread instead the opposite order is respected:
> 
> (c) the accelerator's execution loop exits thanks to the kick
> 
> (b) then the inner execution loop checks cpu->interrupt_request
>     and cpu->exit_request.  If needed cpu->interrupt_request is
>     converted into cpu->exit_request when work is needed outside
>     the execution loop.
> 
> (a) then the other memory locations are checked.  Some may need to
>     be read under the BQL, but the vCPU thread may also take other
>     locks (e.g. for queued work items) or none at all.
> 
> qatomic_set_mb() would only be needed if the halt sleep was done
> outside the BQL (though in that case, cpu->exit_request probably
> would be replaced by a QemuEvent or something like that).
> 
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Reviewed-by: Igor Mammedov <imammedo@redhat.com>

> ---
>  accel/kvm/kvm-all.c          | 2 --
>  accel/tcg/cpu-exec.c         | 1 -
>  accel/tcg/tcg-accel-ops-rr.c | 9 +++++++--
>  accel/tcg/tcg-accel-ops.c    | 2 --
>  accel/tcg/user-exec.c        | 1 +
>  system/cpus.c                | 1 +
>  target/i386/nvmm/nvmm-all.c  | 2 --
>  target/i386/whpx/whpx-all.c  | 2 --
>  8 files changed, 9 insertions(+), 11 deletions(-)
> 
> diff --git a/accel/kvm/kvm-all.c b/accel/kvm/kvm-all.c
> index e4167d94b4f..d13156bee87 100644
> --- a/accel/kvm/kvm-all.c
> +++ b/accel/kvm/kvm-all.c
> @@ -3155,7 +3155,6 @@ int kvm_cpu_exec(CPUState *cpu)
>      trace_kvm_cpu_exec();
>  
>      if (kvm_arch_process_async_events(cpu)) {
> -        qatomic_set(&cpu->exit_request, 0);
>          return EXCP_HLT;
>      }
>  
> @@ -3345,7 +3344,6 @@ int kvm_cpu_exec(CPUState *cpu)
>          vm_stop(RUN_STATE_INTERNAL_ERROR);
>      }
>  
> -    qatomic_set(&cpu->exit_request, 0);
>      return ret;
>  }
>  
> diff --git a/accel/tcg/cpu-exec.c b/accel/tcg/cpu-exec.c
> index 3ae545e888f..ad94f96b252 100644
> --- a/accel/tcg/cpu-exec.c
> +++ b/accel/tcg/cpu-exec.c
> @@ -872,7 +872,6 @@ static inline bool cpu_handle_interrupt(CPUState *cpu,
>       * The corresponding store-release is in cpu_exit.
>       */
>      if (unlikely(qatomic_load_acquire(&cpu->exit_request)) || icount_exit_request(cpu)) {
> -        qatomic_set(&cpu->exit_request, 0);
>          if (cpu->exception_index == -1) {
>              cpu->exception_index = EXCP_INTERRUPT;
>          }
> diff --git a/accel/tcg/tcg-accel-ops-rr.c b/accel/tcg/tcg-accel-ops-rr.c
> index 610292d3bac..e9d291dc391 100644
> --- a/accel/tcg/tcg-accel-ops-rr.c
> +++ b/accel/tcg/tcg-accel-ops-rr.c
> @@ -286,8 +286,13 @@ static void *rr_cpu_thread_fn(void *arg)
>          /* Does not need a memory barrier because a spurious wakeup is okay.  */
>          qatomic_set(&rr_current_cpu, NULL);
>  
> -        if (cpu && qatomic_read(&cpu->exit_request)) {
> -            qatomic_set_mb(&cpu->exit_request, 0);
> +        if (cpu) {
> +            /*
> +             * This could even reset exit_request for all CPUs, but in practice
> +             * races between CPU exits and changes to "cpu" are so rare that
> +             * there's no advantage in doing so.
> +             */
> +            qatomic_set(&cpu->exit_request, false);
>          }
>  
>          if (icount_enabled() && all_cpu_threads_idle()) {
> diff --git a/accel/tcg/tcg-accel-ops.c b/accel/tcg/tcg-accel-ops.c
> index 1f662a9c745..3bd98005042 100644
> --- a/accel/tcg/tcg-accel-ops.c
> +++ b/accel/tcg/tcg-accel-ops.c
> @@ -82,8 +82,6 @@ int tcg_cpu_exec(CPUState *cpu)
>      ret = cpu_exec(cpu);
>      cpu_exec_end(cpu);
>  
> -    qatomic_set_mb(&cpu->exit_request, 0);
> -
>      return ret;
>  }
>  
> diff --git a/accel/tcg/user-exec.c b/accel/tcg/user-exec.c
> index 81906d2e033..8f4f049b924 100644
> --- a/accel/tcg/user-exec.c
> +++ b/accel/tcg/user-exec.c
> @@ -54,6 +54,7 @@ void qemu_cpu_kick(CPUState *cpu)
>  
>  void qemu_wait_io_event(CPUState *cpu)
>  {
> +    qatomic_set(&cpu->exit_request, false);
>      process_queued_cpu_work(cpu);
>  }
>  
> diff --git a/system/cpus.c b/system/cpus.c
> index bb13942cbb7..f989d9938b6 100644
> --- a/system/cpus.c
> +++ b/system/cpus.c
> @@ -463,6 +463,7 @@ void qemu_wait_io_event(CPUState *cpu)
>  {
>      bool slept = false;
>  
> +    qatomic_set(&cpu->exit_request, false);
>      while (cpu_thread_is_idle(cpu)) {
>          if (!slept) {
>              slept = true;
> diff --git a/target/i386/nvmm/nvmm-all.c b/target/i386/nvmm/nvmm-all.c
> index 7e36c42fbb4..ed424251673 100644
> --- a/target/i386/nvmm/nvmm-all.c
> +++ b/target/i386/nvmm/nvmm-all.c
> @@ -817,8 +817,6 @@ nvmm_vcpu_loop(CPUState *cpu)
>      cpu_exec_end(cpu);
>      bql_lock();
>  
> -    qatomic_set(&cpu->exit_request, false);
> -
>      return ret < 0;
>  }
>  
> diff --git a/target/i386/whpx/whpx-all.c b/target/i386/whpx/whpx-all.c
> index 00fb7e23100..2a85168ed51 100644
> --- a/target/i386/whpx/whpx-all.c
> +++ b/target/i386/whpx/whpx-all.c
> @@ -2050,8 +2050,6 @@ static int whpx_vcpu_run(CPUState *cpu)
>          whpx_last_vcpu_stopping(cpu);
>      }
>  
> -    qatomic_set(&cpu->exit_request, false);
> -
>      return ret < 0;
>  }
>
Re: [PATCH 16/18] cpus: clear exit_request in qemu_wait_io_event
Posted by Richard Henderson 1 week ago
On 8/30/25 01:31, Paolo Bonzini wrote:
> Make the code common to all accelerators: after seeing cpu->exit_request
> set to true, accelerator code needs to reach qemu_wait_io_event_common().
> 
> So for the common cases where they use qemu_wait_io_event(), go ahead and
> clear it in there.  Note that the cheap qatomic_set() is enough because
> at this point the thread has taken the BQL; qatomic_set_mb() is not needed.
> In particular, this is the ordering of the communication between
> I/O and vCPU threads is always the same.
> 
> In the I/O thread:
> 
> (a) store other memory locations that will be checked if cpu->exit_request
>      or cpu->interrupt_request is 1 (for example cpu->stop or cpu->work_list
>      for cpu->exit_request)
> 
> (b) cpu_exit(): store-release cpu->exit_request, or
> (b) cpu_interrupt(): store-release cpu->interrupt_request

Mm.  This is the reason we want the seq_cst of the qatomic_or.
Perhaps comments in patch 7 should be expanded to document this?

Anyway for this patch,
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>


r~
Re: [PATCH 16/18] cpus: clear exit_request in qemu_wait_io_event
Posted by Paolo Bonzini 1 week ago
On Sat, Aug 30, 2025 at 12:11 AM Richard Henderson
<richard.henderson@linaro.org> wrote:
> > (b) cpu_exit(): store-release cpu->exit_request, or
> > (b) cpu_interrupt(): store-release cpu->interrupt_request
>
> Mm.  This is the reason we want the seq_cst of the qatomic_or.
> Perhaps comments in patch 7 should be expanded to document this?

Not entirely.  There are cases in which a relaxed RMW atomic would
also work, for example this one in the CPU thread:

    if (ctl_has_irq(env)) {
        cpu_set_interrupt(cs, CPU_INTERRUPT_VIRQ);
    }

and there are cases in which store-release is needed but atomicity
isn't (the ones under the BQL). It's putting the two requirements
together that requires patch 7.

And to be honest I am not sure there are other cases than the one
above, where cpu_set_interrupt()/cpu_reset_interrupt() is called
outside the BQL. But it's really a matter of time. Sooner or later
someone will move the APIC or GIC under its own lock, therefore we
might as well keep things clean already. "Written only under lock X"
is handy but I prefer to use it with moderation - like it's okay if
it's a primitive but not if there are dozens of calls across the code
base.

Thanks for the speedy review!

Paolo