KVM: PPC: Book3S HV: Drop locks around call to kvmppc_pin_guest_page

At the moment we call kvmppc_pin_guest_page() in kvmppc_update_vpa()
with two spinlocks held: the vcore lock and the vcpu->vpa_update_lock.
This is not good, since kvmppc_pin_guest_page() calls down_read() and
get_user_pages_fast(), both of which can sleep.  This bug was introduced
in 2e25aa5f ("KVM: PPC: Book3S HV: Make virtual processor area
registration more robust").

This arranges to drop those spinlocks before calling
kvmppc_pin_guest_page() and re-take them afterwards.  Dropping the
vcore lock in kvmppc_run_core() means we have to set the vcore_state
field to VCORE_RUNNING before we drop the lock, so that other vcpus
won't try to run this vcore.

Signed-off-by: Paul Mackerras <paulus@samba.org>
Acked-by: Alexander Graf <agraf@suse.de>
Signed-off-by: Avi Kivity <avi@redhat.com>

1 files changed, 66 insertions, 30 deletions

diff --git a/arch/powerpc/kvm/book3s_hv.c b/arch/powerpc/kvm/book3s_hv.c
index c6af1d623839..3abe1b86e583 100644
--- a/arch/powerpc/kvm/book3s_hv.c
+++ b/arch/powerpc/kvm/book3s_hv.c
@@ -268,24 +268,45 @@ static unsigned long do_h_register_vpa(struct kvm_vcpu *vcpu,
         return err;
 }
 
-static void kvmppc_update_vpa(struct kvm *kvm, struct kvmppc_vpa *vpap)
+static void kvmppc_update_vpa(struct kvm_vcpu *vcpu, struct kvmppc_vpa *vpap)
 {
+        struct kvm *kvm = vcpu->kvm;
         void *va;
         unsigned long nb;
+        unsigned long gpa;
 
-        vpap->update_pending = 0;
-        va = NULL;
-        if (vpap->next_gpa) {
-                va = kvmppc_pin_guest_page(kvm, vpap->next_gpa, &nb);
-                if (nb < vpap->len) {
-                        /*
-                         * If it's now too short, it must be that userspace
-                         * has changed the mappings underlying guest memory,
-                         * so unregister the region.
-                         */
+        /*
+         * We need to pin the page pointed to by vpap->next_gpa,
+         * but we can't call kvmppc_pin_guest_page under the lock
+         * as it does get_user_pages() and down_read().  So we
+         * have to drop the lock, pin the page, then get the lock
+         * again and check that a new area didn't get registered
+         * in the meantime.
+         */
+        for (;;) {
+                gpa = vpap->next_gpa;
+                spin_unlock(&vcpu->arch.vpa_update_lock);
+                va = NULL;
+                nb = 0;
+                if (gpa)
+                        va = kvmppc_pin_guest_page(kvm, vpap->next_gpa, &nb);
+                spin_lock(&vcpu->arch.vpa_update_lock);
+                if (gpa == vpap->next_gpa)
+                        break;
+                /* sigh... unpin that one and try again */
+                if (va)
                         kvmppc_unpin_guest_page(kvm, va);
-                        va = NULL;
-                }
+        }
+
+        vpap->update_pending = 0;
+        if (va && nb < vpap->len) {
+                /*
+                 * If it's now too short, it must be that userspace
+                 * has changed the mappings underlying guest memory,
+                 * so unregister the region.
+                 */
+                kvmppc_unpin_guest_page(kvm, va);
+                va = NULL;
         }
         if (vpap->pinned_addr)
                 kvmppc_unpin_guest_page(kvm, vpap->pinned_addr);
@@ -296,20 +317,18 @@ static void kvmppc_update_vpa(struct kvm *kvm, struct kvmppc_vpa *vpap)
 
 static void kvmppc_update_vpas(struct kvm_vcpu *vcpu)
 {
-        struct kvm *kvm = vcpu->kvm;
-
         spin_lock(&vcpu->arch.vpa_update_lock);
         if (vcpu->arch.vpa.update_pending) {
-                kvmppc_update_vpa(kvm, &vcpu->arch.vpa);
+                kvmppc_update_vpa(vcpu, &vcpu->arch.vpa);
                 init_vpa(vcpu, vcpu->arch.vpa.pinned_addr);
         }
         if (vcpu->arch.dtl.update_pending) {
-                kvmppc_update_vpa(kvm, &vcpu->arch.dtl);
+                kvmppc_update_vpa(vcpu, &vcpu->arch.dtl);
                 vcpu->arch.dtl_ptr = vcpu->arch.dtl.pinned_addr;
                 vcpu->arch.dtl_index = 0;
         }
         if (vcpu->arch.slb_shadow.update_pending)
-                kvmppc_update_vpa(kvm, &vcpu->arch.slb_shadow);
+                kvmppc_update_vpa(vcpu, &vcpu->arch.slb_shadow);
         spin_unlock(&vcpu->arch.vpa_update_lock);
 }
 
@@ -800,12 +819,39 @@ static int kvmppc_run_core(struct kvmppc_vcore *vc)
         struct kvm_vcpu *vcpu, *vcpu0, *vnext;
         long ret;
         u64 now;
-        int ptid, i;
+        int ptid, i, need_vpa_update;
 
         /* don't start if any threads have a signal pending */
-        list_for_each_entry(vcpu, &vc->runnable_threads, arch.run_list)
+        need_vpa_update = 0;
+        list_for_each_entry(vcpu, &vc->runnable_threads, arch.run_list) {
                 if (signal_pending(vcpu->arch.run_task))
                         return 0;
+                need_vpa_update |= vcpu->arch.vpa.update_pending |
+                        vcpu->arch.slb_shadow.update_pending |
+                        vcpu->arch.dtl.update_pending;
+        }
+
+        /*
+         * Initialize *vc, in particular vc->vcore_state, so we can
+         * drop the vcore lock if necessary.
+         */
+        vc->n_woken = 0;
+        vc->nap_count = 0;
+        vc->entry_exit_count = 0;
+        vc->vcore_state = VCORE_RUNNING;
+        vc->in_guest = 0;
+        vc->napping_threads = 0;
+
+        /*
+         * Updating any of the vpas requires calling kvmppc_pin_guest_page,
+         * which can't be called with any spinlocks held.
+         */
+        if (need_vpa_update) {
+                spin_unlock(&vc->lock);
+                list_for_each_entry(vcpu, &vc->runnable_threads, arch.run_list)
+                        kvmppc_update_vpas(vcpu);
+                spin_lock(&vc->lock);
+        }
 
         /*
          * Make sure we are running on thread 0, and that
@@ -838,20 +884,10 @@ static int kvmppc_run_core(struct kvmppc_vcore *vc)
                 if (vcpu->arch.ceded)
                         vcpu->arch.ptid = ptid++;
 
-        vc->n_woken = 0;
-        vc->nap_count = 0;
-        vc->entry_exit_count = 0;
-        vc->vcore_state = VCORE_RUNNING;
         vc->stolen_tb += mftb() - vc->preempt_tb;
-        vc->in_guest = 0;
         vc->pcpu = smp_processor_id();
-        vc->napping_threads = 0;
         list_for_each_entry(vcpu, &vc->runnable_threads, arch.run_list) {
                 kvmppc_start_thread(vcpu);
-                if (vcpu->arch.vpa.update_pending ||
-                    vcpu->arch.slb_shadow.update_pending ||
-                    vcpu->arch.dtl.update_pending)
-                        kvmppc_update_vpas(vcpu);
                 kvmppc_create_dtl_entry(vcpu, vc);
         }
         /* Grab any remaining hw threads so they can't go into the kernel */