x86: kvm: hyperv: simplify SynIC message delivery
authorRoman Kagan <rkagan@virtuozzo.com>
Mon, 10 Dec 2018 18:47:26 +0000 (18:47 +0000)
committerPaolo Bonzini <pbonzini@redhat.com>
Fri, 14 Dec 2018 16:59:51 +0000 (17:59 +0100)
SynIC message delivery is somewhat overengineered: it pretends to follow
the ordering rules when grabbing the message slot, using atomic
operations and all that, but does it incorrectly and unnecessarily.

The correct order would be to first set .msg_pending, then atomically
replace .message_type if it was zero, and then clear .msg_pending if
the previous step was successful.  But this all is done in vcpu context
so the whole update looks atomic to the guest (it's assumed to only
access the message page from this cpu), and therefore can be done in
whatever order is most convenient (and is also the reason why the
incorrect order didn't trigger any bugs so far).

While at this, also switch to kvm_vcpu_{read,write}_guest_page, and drop
the no longer needed synic_clear_sint_msg_pending.

Signed-off-by: Roman Kagan <rkagan@virtuozzo.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
arch/x86/kvm/hyperv.c

index 4e80080f277a66abdc7350af0bc2bc279457b926..17d04d2c6d4fe57a9de34d247115e771ee3e53d6 100644 (file)
@@ -158,32 +158,6 @@ static struct kvm_vcpu_hv_synic *synic_get(struct kvm *kvm, u32 vpidx)
        return (synic->active) ? synic : NULL;
 }
 
-static void synic_clear_sint_msg_pending(struct kvm_vcpu_hv_synic *synic,
-                                       u32 sint)
-{
-       struct kvm_vcpu *vcpu = synic_to_vcpu(synic);
-       struct page *page;
-       gpa_t gpa;
-       struct hv_message *msg;
-       struct hv_message_page *msg_page;
-
-       gpa = synic->msg_page & PAGE_MASK;
-       page = kvm_vcpu_gfn_to_page(vcpu, gpa >> PAGE_SHIFT);
-       if (is_error_page(page)) {
-               vcpu_err(vcpu, "Hyper-V SynIC can't get msg page, gpa 0x%llx\n",
-                        gpa);
-               return;
-       }
-       msg_page = kmap_atomic(page);
-
-       msg = &msg_page->sint_message[sint];
-       msg->header.message_flags.msg_pending = 0;
-
-       kunmap_atomic(msg_page);
-       kvm_release_page_dirty(page);
-       kvm_vcpu_mark_page_dirty(vcpu, gpa >> PAGE_SHIFT);
-}
-
 static void kvm_hv_notify_acked_sint(struct kvm_vcpu *vcpu, u32 sint)
 {
        struct kvm *kvm = vcpu->kvm;
@@ -194,9 +168,6 @@ static void kvm_hv_notify_acked_sint(struct kvm_vcpu *vcpu, u32 sint)
 
        trace_kvm_hv_notify_acked_sint(vcpu->vcpu_id, sint);
 
-       if (synic->msg_page & HV_SYNIC_SIMP_ENABLE)
-               synic_clear_sint_msg_pending(synic, sint);
-
        /* Try to deliver pending Hyper-V SynIC timers messages */
        stimers_pending = 0;
        for (idx = 0; idx < ARRAY_SIZE(hv_vcpu->stimer); idx++) {
@@ -589,41 +560,54 @@ static int synic_deliver_msg(struct kvm_vcpu_hv_synic *synic, u32 sint,
                             struct hv_message *src_msg)
 {
        struct kvm_vcpu *vcpu = synic_to_vcpu(synic);
-       struct page *page;
-       gpa_t gpa;
-       struct hv_message *dst_msg;
+       int msg_off = offsetof(struct hv_message_page, sint_message[sint]);
+       gfn_t msg_page_gfn;
+       struct hv_message_header hv_hdr;
        int r;
-       struct hv_message_page *msg_page;
 
        if (!(synic->msg_page & HV_SYNIC_SIMP_ENABLE))
                return -ENOENT;
 
-       gpa = synic->msg_page & PAGE_MASK;
-       page = kvm_vcpu_gfn_to_page(vcpu, gpa >> PAGE_SHIFT);
-       if (is_error_page(page))
-               return -EFAULT;
+       msg_page_gfn = synic->msg_page >> PAGE_SHIFT;
 
-       msg_page = kmap_atomic(page);
-       dst_msg = &msg_page->sint_message[sint];
-       if (sync_cmpxchg(&dst_msg->header.message_type, HVMSG_NONE,
-                        src_msg->header.message_type) != HVMSG_NONE) {
-               dst_msg->header.message_flags.msg_pending = 1;
-               r = -EAGAIN;
-       } else {
-               memcpy(&dst_msg->u.payload, &src_msg->u.payload,
-                      src_msg->header.payload_size);
-               dst_msg->header.message_type = src_msg->header.message_type;
-               dst_msg->header.payload_size = src_msg->header.payload_size;
-               r = synic_set_irq(synic, sint);
-               if (r >= 1)
-                       r = 0;
-               else if (r == 0)
-                       r = -EFAULT;
+       /*
+        * Strictly following the spec-mandated ordering would assume setting
+        * .msg_pending before checking .message_type.  However, this function
+        * is only called in vcpu context so the entire update is atomic from
+        * guest POV and thus the exact order here doesn't matter.
+        */
+       r = kvm_vcpu_read_guest_page(vcpu, msg_page_gfn, &hv_hdr.message_type,
+                                    msg_off + offsetof(struct hv_message,
+                                                       header.message_type),
+                                    sizeof(hv_hdr.message_type));
+       if (r < 0)
+               return r;
+
+       if (hv_hdr.message_type != HVMSG_NONE) {
+               hv_hdr.message_flags.msg_pending = 1;
+               r = kvm_vcpu_write_guest_page(vcpu, msg_page_gfn,
+                                             &hv_hdr.message_flags,
+                                             msg_off +
+                                             offsetof(struct hv_message,
+                                                      header.message_flags),
+                                             sizeof(hv_hdr.message_flags));
+               if (r < 0)
+                       return r;
+               return -EAGAIN;
        }
-       kunmap_atomic(msg_page);
-       kvm_release_page_dirty(page);
-       kvm_vcpu_mark_page_dirty(vcpu, gpa >> PAGE_SHIFT);
-       return r;
+
+       r = kvm_vcpu_write_guest_page(vcpu, msg_page_gfn, src_msg, msg_off,
+                                     sizeof(src_msg->header) +
+                                     src_msg->header.payload_size);
+       if (r < 0)
+               return r;
+
+       r = synic_set_irq(synic, sint);
+       if (r < 0)
+               return r;
+       if (r == 0)
+               return -EFAULT;
+       return 0;
 }
 
 static int stimer_send_msg(struct kvm_vcpu_hv_stimer *stimer)