[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [RFC v3 08/13] exec.c: introduce a simple rendezvous su
From: |
Alex Bennée |
Subject: |
Re: [Qemu-devel] [RFC v3 08/13] exec.c: introduce a simple rendezvous support |
Date: |
Fri, 17 Jul 2015 14:45:47 +0100 |
Alvise Rigo <address@hidden> writes:
> When a vCPU is about to set a memory page as exclusive, it needs to wait
> that all the running vCPUs finish to execute the current TB and to be aware
> of the exact moment when that happens. For this, add a simple rendezvous
> mechanism that will be used in softmmu_llsc_template.h to implement the
> ldlink operation.
>
> Suggested-by: Jani Kokkonen <address@hidden>
> Suggested-by: Claudio Fontana <address@hidden>
> Signed-off-by: Alvise Rigo <address@hidden>
> ---
> cpus.c | 5 +++++
> exec.c | 45 +++++++++++++++++++++++++++++++++++++++++++++
> include/qom/cpu.h | 16 ++++++++++++++++
> 3 files changed, 66 insertions(+)
>
> diff --git a/cpus.c b/cpus.c
> index aee445a..f4d938e 100644
> --- a/cpus.c
> +++ b/cpus.c
> @@ -1423,6 +1423,11 @@ static int tcg_cpu_exec(CPUArchState *env)
> qemu_mutex_unlock_iothread();
> ret = cpu_exec(env);
> cpu->tcg_executing = 0;
> +
> + if (unlikely(cpu->pending_rdv)) {
> + cpu_exit_do_rendezvous(cpu);
> + }
> +
I'll ignore this stuff for now as I assume we can all use the async work
patch of Fred's?
> qemu_mutex_lock_iothread();
> #ifdef CONFIG_PROFILER
> tcg_time += profile_getclock() - ti;
> diff --git a/exec.c b/exec.c
> index 964e922..51958ed 100644
> --- a/exec.c
> +++ b/exec.c
> @@ -746,6 +746,51 @@ void cpu_breakpoint_remove_all(CPUState *cpu, int mask)
> }
> }
>
> +/* Rendezvous implementation.
> + * The corresponding definitions are in include/qom/cpu.h. */
> +CpuExitRendezvous cpu_exit_rendezvous;
> +inline void cpu_exit_init_rendezvous(void)
> +{
> + CpuExitRendezvous *rdv = &cpu_exit_rendezvous;
> +
> + rdv->attendees = 0;
> +}
> +
> +inline void cpu_exit_rendezvous_add_attendee(CPUState *cpu)
> +{
> + CpuExitRendezvous *rdv = &cpu_exit_rendezvous;
> +
> + if (!cpu->pending_rdv) {
> + cpu->pending_rdv = 1;
> + atomic_inc(&rdv->attendees);
> + }
> +}
> +
> +void cpu_exit_do_rendezvous(CPUState *cpu)
> +{
> + CpuExitRendezvous *rdv = &cpu_exit_rendezvous;
> +
> + atomic_dec(&rdv->attendees);
> +
> + cpu->pending_rdv = 0;
> +}
> +
> +void cpu_exit_rendezvous_wait_others(CPUState *cpu)
> +{
> + CpuExitRendezvous *rdv = &cpu_exit_rendezvous;
> +
> + while (rdv->attendees) {
> + g_usleep(TCG_RDV_POLLING_PERIOD);
> + }
> +}
> +
> +void cpu_exit_rendezvous_release(void)
> +{
> + CpuExitRendezvous *rdv = &cpu_exit_rendezvous;
> +
> + rdv->attendees = 0;
> +}
> +
> /* enable or disable single step mode. EXCP_DEBUG is returned by the
> CPU loop after each instruction */
> void cpu_single_step(CPUState *cpu, int enabled)
> diff --git a/include/qom/cpu.h b/include/qom/cpu.h
> index 8f3fe56..8d121b3 100644
> --- a/include/qom/cpu.h
> +++ b/include/qom/cpu.h
> @@ -201,6 +201,20 @@ typedef struct CPUWatchpoint {
> QTAILQ_ENTRY(CPUWatchpoint) entry;
> } CPUWatchpoint;
>
> +/* Rendezvous support */
> +#define TCG_RDV_POLLING_PERIOD 10
> +typedef struct CpuExitRendezvous {
> + volatile int attendees;
> + QemuMutex lock;
> +} CpuExitRendezvous;
> +
> +extern CpuExitRendezvous cpu_exit_rendezvous;
> +void cpu_exit_init_rendezvous(void);
> +void cpu_exit_rendezvous_add_attendee(CPUState *cpu);
> +void cpu_exit_do_rendezvous(CPUState *cpu);
> +void cpu_exit_rendezvous_wait_others(CPUState *cpu);
> +void cpu_exit_rendezvous_release(void);
> +
> struct KVMState;
> struct kvm_run;
>
> @@ -291,6 +305,8 @@ struct CPUState {
>
> void *opaque;
>
> + volatile int pending_rdv;
> +
I will however echo the "hmmm" on Fred's patch about the optimistic use
of volatile here. As Peter mentioned it is a red flag and I would prefer
explicit memory consistency behaviour used and documented.
> /* In order to avoid passing too many arguments to the MMIO helpers,
> * we store some rarely used information in the CPU context.
> */
--
Alex Bennée
- Re: [Qemu-devel] [RFC v3 03/13] softmmu: Add helpers for a new slow-path, (continued)
- [Qemu-devel] [RFC v3 06/13] target-i386: translate: implement qemu_ldlink and qemu_stcond ops, Alvise Rigo, 2015/07/10
- [Qemu-devel] [RFC v3 04/13] tcg-op: create new TCG qemu_ldlink and qemu_stcond instructions, Alvise Rigo, 2015/07/10
- [Qemu-devel] [RFC v3 07/13] ram_addr.h: Make exclusive bitmap accessors atomic, Alvise Rigo, 2015/07/10
- [Qemu-devel] [RFC v3 08/13] exec.c: introduce a simple rendezvous support, Alvise Rigo, 2015/07/10
- Re: [Qemu-devel] [RFC v3 08/13] exec.c: introduce a simple rendezvous support,
Alex Bennée <=
- [Qemu-devel] [RFC v3 05/13] target-arm: translate: implement qemu_ldlink and qemu_stcond ops, Alvise Rigo, 2015/07/10
- [Qemu-devel] [RFC v3 09/13] cpus.c: introduce simple callback support, Alvise Rigo, 2015/07/10