On Mon, Aug 23, 2010 at 07:17:20PM +0300, Avi Kivity wrote:
> On 07/19/2010 06:30 PM, Gleb Natapov wrote:
> >If guest access swapped out memory do not swap it in from vcpu thread
> >context. Setup slow work to do swapping and send async page fault to
> >a guest.
> >
> >Allow async page fault injection only when guest is in user mode since
> >otherwise guest may be in non-sleepable context and will not be able to
> >reschedule.
> >
> >
> >
> > struct kvm_arch {
> >@@ -444,6 +446,8 @@ struct kvm_vcpu_stat {
> > u32 hypercalls;
> > u32 irq_injections;
> > u32 nmi_injections;
> >+ u32 apf_not_present;
> >+ u32 apf_present;
> > };
>
> Please don't add more stats, instead add tracepoints which can be
> converted to stats by userspace.
>
> Would be good to have both guest and host tracepoints.
>
I do have host tracepoints for all events. I still prefer to have also
kvm stats since they are so much easier to use right now. We can delete
them later when we replace kvm_stat with perf.
> >@@ -2345,6 +2346,21 @@ static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
> > error_code& PFERR_WRITE_MASK, gfn);
> > }
> >
> >+int kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn)
> >+{
> >+ struct kvm_arch_async_pf arch;
> >+ arch.token = (vcpu->arch.async_pf_id++<< 12) | vcpu->vcpu_id;
> >+ return kvm_setup_async_pf(vcpu, gva, gfn,&arch);
> >+}
>
> Ok. so token is globally unique. We're limited to 4096 vcpus, I
> guess that's fine. Wraparound at 1M faults/vcpu, what's the impact?
> failure if we have 1M faulting processes on one vcpu?
1M faulting processes on one vcpu simultaneously. And we limit number of
outstanding apfs anyway.
>
> I guess that's fine too.
>
> >+
> >+static bool can_do_async_pf(struct kvm_vcpu *vcpu)
> >+{
> >+ if (!vcpu->arch.apf_data || kvm_event_needs_reinjection(vcpu))
> >+ return false;
> >+
> >+ return !!kvm_x86_ops->get_cpl(vcpu);
>
> !! !needed, bool autoconverts. But > 0 is more readable.
OK.
>
> >+}
> >+
> > static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
> > u32 error_code)
> > {
> >@@ -2353,6 +2369,7 @@ static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
> > int level;
> > gfn_t gfn = gpa>> PAGE_SHIFT;
> > unsigned long mmu_seq;
> >+ bool async;
> >
> > ASSERT(vcpu);
> > ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
> >@@ -2367,7 +2384,23 @@ static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
> >
> > mmu_seq = vcpu->kvm->mmu_notifier_seq;
> > smp_rmb();
> >- pfn = gfn_to_pfn(vcpu->kvm, gfn);
> >+
> >+ if (can_do_async_pf(vcpu)) {
> >+ pfn = gfn_to_pfn_async(vcpu->kvm, gfn,&async);
> >+ trace_kvm_try_async_get_page(async, pfn);
> >+ } else {
> >+do_sync:
> >+ async = false;
> >+ pfn = gfn_to_pfn(vcpu->kvm, gfn);
> >+ }
> >+
> >+ if (async) {
> >+ if (!kvm_arch_setup_async_pf(vcpu, gpa, gfn))
> >+ goto do_sync;
> >+ return 0;
> >+ }
> >+
>
> This goto is pretty ugly. How about:
>
> async = false;
> if (can_do_async_pf(&async)) {
> }
> if (async && !setup())
> async = false;
> if (async)
> ...
>
> or something.
>
Will try to re-factor.
> >@@ -459,7 +460,21 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr,
> >
> > mmu_seq = vcpu->kvm->mmu_notifier_seq;
> > smp_rmb();
> >- pfn = gfn_to_pfn(vcpu->kvm, walker.gfn);
> >+
> >+ if (can_do_async_pf(vcpu)) {
> >+ pfn = gfn_to_pfn_async(vcpu->kvm, walker.gfn,&async);
> >+ trace_kvm_try_async_get_page(async, pfn);
> >+ } else {
> >+do_sync:
> >+ async = false;
> >+ pfn = gfn_to_pfn(vcpu->kvm, walker.gfn);
> >+ }
> >+
> >+ if (async) {
> >+ if (!kvm_arch_setup_async_pf(vcpu, addr, walker.gfn))
> >+ goto do_sync;
> >+ return 0;
> >+ }
> >
>
> This repetition is ugly too.
>
Yeah. May be it can be moved to separate function and this will help to
get rid of goto as well.
> >
> >+static int apf_put_user(struct kvm_vcpu *vcpu, u32 val)
> >+{
> >+ if (unlikely(vcpu->arch.apf_memslot_ver !=
> >+ vcpu->kvm->memslot_version)) {
> >+ u64 gpa = vcpu->arch.apf_msr_val& ~0x3f;
> >+ unsigned long addr;
> >+ int offset = offset_in_page(gpa);
> >+
> >+ addr = gfn_to_hva(vcpu->kvm, gpa>> PAGE_SHIFT);
> >+ vcpu->arch.apf_data = (u32 __user *)(addr + offset);
> >+ if (kvm_is_error_hva(addr)) {
> >+ vcpu->arch.apf_data = NULL;
> >+ return -EFAULT;
> >+ }
> >+ }
> >+
> >+ return put_user(val, vcpu->arch.apf_data);
> >+}
>
> This nice cache needs to be outside apf to reduce complexity for
> reviewers and since it is useful for others.
>
> Would be good to have memslot-cached kvm_put_guest() and kvm_get_guest().
Will look into it.
>
> >diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> >index 292514c..f56e8ac 100644
> >--- a/virt/kvm/kvm_main.c
> >+++ b/virt/kvm/kvm_main.c
> >@@ -78,6 +78,11 @@ static atomic_t hardware_enable_failed;
> > struct kmem_cache *kvm_vcpu_cache;
> > EXPORT_SYMBOL_GPL(kvm_vcpu_cache);
> >
> >+#ifdef CONFIG_KVM_ASYNC_PF
> >+#define ASYNC_PF_PER_VCPU 100
> >+static struct kmem_cache *async_pf_cache;
> >+#endif
>
> All those #ifdefs can be eliminated with virt/kvm/apf.[ch].
>
OK.
> >+
> > static __read_mostly struct preempt_ops kvm_preempt_ops;
> >
> > struct dentry *kvm_debugfs_dir;
> >@@ -186,6 +191,11 @@ int kvm_vcpu_init(struct kvm_vcpu *vcpu, struct kvm *kvm, unsigned id)
> > vcpu->kvm = kvm;
> > vcpu->vcpu_id = id;
> > init_waitqueue_head(&vcpu->wq);
> >+#ifdef CONFIG_KVM_ASYNC_PF
> >+ INIT_LIST_HEAD(&vcpu->async_pf_done);
> >+ INIT_LIST_HEAD(&vcpu->async_pf_queue);
> >+ spin_lock_init(&vcpu->async_pf_lock);
> >+#endif
>
> kvm_apf_init() etc.
>
> >+ struct kvm_arch_async_pf *arch)
> >+{
> >+ struct kvm_async_pf *work;
> >+
> >+ if (vcpu->async_pf_queued>= ASYNC_PF_PER_VCPU)
> >+ return 0;
>
> 100 == too high. At 16 vcpus, this allows 1600 kernel threads to
> wait for I/O.
Number of kernel threads are limited by other means. Slow work subsystem
has its own knobs to tune that. Here we limit how much slow work items
can be queued per vcpu.
>
> Would have been best if we could ask for a page to be paged in
> asynchronously.
>
You mean to have core kernel facility for that? I agree it would be
nice, but much harder.
> >+
> >+ /* setup slow work */
> >+
> >+ /* do alloc nowait since if we are going to sleep anyway we
> >+ may as well sleep faulting in page */
> >+ work = kmem_cache_zalloc(async_pf_cache, GFP_NOWAIT);
> >+ if (!work)
> >+ return 0;
> >+
> >+ atomic_set(&work->used, 1);
> >+ work->page = NULL;
> >+ work->vcpu = vcpu;
> >+ work->gva = gva;
> >+ work->addr = gfn_to_hva(vcpu->kvm, gfn);
> >+ work->arch = *arch;
> >+ work->mm = current->mm;
> >+ atomic_inc(&work->mm->mm_count);
> >+ kvm_get_kvm(work->vcpu->kvm);
> >+
> >+ /* this can't really happen otherwise gfn_to_pfn_async
> >+ would succeed */
> >+ if (unlikely(kvm_is_error_hva(work->addr)))
> >+ goto retry_sync;
> >+
> >+ slow_work_init(&work->work,&async_pf_ops);
> >+ if (slow_work_enqueue(&work->work) != 0)
> >+ goto retry_sync;
> >+
> >+ vcpu->async_pf_work = work;
> >+ list_add_tail(&work->queue,&vcpu->async_pf_queue);
> >+ vcpu->async_pf_queued++;
> >+ return 1;
> >+retry_sync:
> >+ kvm_put_kvm(work->vcpu->kvm);
> >+ mmdrop(work->mm);
> >+ kmem_cache_free(async_pf_cache, work);
> >+ return 0;
> >+}
> >+
> >+
>
> --
> error compiling committee.c: too many arguments to function
--
Gleb.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>