ioreq: handle pending emulation racing with ioreq server destruction

When an emulation request is initiated in hvm_send_ioreq() the guest vcpu is
blocked on an event channel until that request is completed. If, however,
the emulator is killed whilst that emulation is pending then the ioreq
server may be destroyed. Thus when the vcpu is awoken the code in
handle_hvm_io_completion() will find no pending request to wait for, but will
leave the internal vcpu io_req.state set to IOREQ_READY and the vcpu shutdown
deferall flag in place (because hvm_io_assist() will never be called). The
emulation request is then completed anyway. This means that any subsequent call
to hvmemul_do_io() will find an unexpected value in io_req.state and will
return X86EMUL_UNHANDLEABLE, which in some cases will result in continuous
re-tries.

This patch fixes the issue by moving the setting of io_req.state and clearing
of shutdown deferral (as will as MSI-X write completion) out of hvm_io_assist()
and directly into handle_hvm_io_completion().

Reported-by: Marek Marczykowski-Górecki <marmarek@invisiblethingslab.com>
Signed-off-by: Paul Durrant <pdurrant@amazon.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Release-acked-by: Paul Durrant <paul@xen.org>

diff --git a/xen/arch/x86/hvm/ioreq.c b/xen/arch/x86/hvm/ioreq.c
index c55c4bc4bca5ece57d8bca87cb68494f7bd02318..724007016d4a76278996ed4cd63fb02d25decbd5 100644 (file)
--- a/xen/arch/x86/hvm/ioreq.c
+++ b/xen/arch/x86/hvm/ioreq.c
@@ -109,15 +109,7 @@ static void hvm_io_assist(struct hvm_ioreq_vcpu *sv, uint64_t data)
     ioreq_t *ioreq = &v->arch.hvm.hvm_io.io_req;
 
     if ( hvm_ioreq_needs_completion(ioreq) )
-    {
-        ioreq->state = STATE_IORESP_READY;
         ioreq->data = data;
-    }
-    else
-        ioreq->state = STATE_IOREQ_NONE;
-
-    msix_write_completion(v);
-    vcpu_end_shutdown_deferral(v);
 
     sv->pending = false;
 }
@@ -209,6 +201,12 @@ bool handle_hvm_io_completion(struct vcpu *v)
         }
     }
 
+    vio->io_req.state = hvm_ioreq_needs_completion(&vio->io_req) ?
+        STATE_IORESP_READY : STATE_IOREQ_NONE;
+
+    msix_write_completion(v);
+    vcpu_end_shutdown_deferral(v);
+
     io_completion = vio->io_completion;
     vio->io_completion = HVMIO_no_completion;