xen: io: Fix race between sending an I/O and domain shutdown

Xen provides hypercalls to shutdown (SCHEDOP_shutdown{,_code}) and
resume a domain (XEN_DOMCTL_resumedomain). They can be used for checkpoint
where the expectation is the domain should continue as nothing happened
afterwards.

hvmemul_do_io() and handle_pio() will act differently if the return
code of hvm_send_ioreq() (resp. hvmemul_do_pio_buffer()) is X86EMUL_RETRY.

In this case, the I/O state will be reset to STATE_IOREQ_NONE (i.e
no I/O is pending) and/or the PC will not be advanced.

If the shutdown request happens right after the I/O was sent to the
IOREQ, then emulation code will end up to re-execute the instruction
and therefore forward again the same I/O (at least when reading IO port).

This would be problem if the access has a side-effect. A dumb example,
is a device implementing a counter which is incremented by one for every
access. When running shutdown/resume in a loop, the value read by the
OS may not be the old value + 1.

Add an extra boolean in the structure hvm_vcpu_io to indicate whether
the I/O was suspended. This is then used in place of checking the domain
is shutting down in hvmemul_do_io() and handle_pio() as they should
act on suspend (i.e. vcpu_start_shutdown_deferral() returns false) rather
than shutdown.

Signed-off-by: Julien Grall <jgrall@amazon.com>
Reviewed-by: Paul Durrant <paul@xen.org>
master commit: b7e0d89788
master date: 2022-05-06 17:16:22 +0100

xen/arch/arm/ioreq.c

@@ -80,9 +80,10 @@ enum io_state try_fwd_ioserv(struct cpu_user_regs *regs,
         return IO_ABORT;
 
     vio->req = p;
+    vio->suspended = false;
 
     rc = ioreq_send(s, &p, 0);
-    if ( rc != IO_RETRY || v->domain->is_shutting_down )
+    if ( rc != IO_RETRY || vio->suspended )
         vio->req.state = STATE_IOREQ_NONE;
     else if ( !ioreq_needs_completion(&vio->req) )
         rc = IO_HANDLED;

xen/arch/x86/hvm/emulate.c

@@ -239,6 +239,7 @@ static int hvmemul_do_io(
     ASSERT(p.count);
 
     vio->req = p;
+    vio->suspended = false;
 
     rc = hvm_io_intercept(&p);
 
@@ -334,7 +335,7 @@ static int hvmemul_do_io(
         else
         {
             rc = ioreq_send(s, &p, 0);
-            if ( rc != X86EMUL_RETRY || currd->is_shutting_down )
+            if ( rc != X86EMUL_RETRY || vio->suspended )
                 vio->req.state = STATE_IOREQ_NONE;
             else if ( !ioreq_needs_completion(&vio->req) )
                 rc = X86EMUL_OKAY;

xen/arch/x86/hvm/io.c

@@ -138,10 +138,11 @@ bool handle_pio(uint16_t port, unsigned int size, int dir)
 
     case X86EMUL_RETRY:
         /*
-         * We should not advance RIP/EIP if the domain is shutting down or
-         * if X86EMUL_RETRY has been returned by an internal handler.
+         * We should not advance RIP/EIP if the vio was suspended (e.g.
+         * because the domain is shutting down) or if X86EMUL_RETRY has
+         * been returned by an internal handler.
          */
-        if ( curr->domain->is_shutting_down || !vcpu_ioreq_pending(curr) )
+        if ( vio->suspended || !vcpu_ioreq_pending(curr) )
             return false;
         break;
 

xen/common/ioreq.c

@@ -1256,6 +1256,7 @@ int ioreq_send(struct ioreq_server *s, ioreq_t *proto_p,
     struct vcpu *curr = current;
     struct domain *d = curr->domain;
     struct ioreq_vcpu *sv;
+    struct vcpu_io *vio = &curr->io;
 
     ASSERT(s);
 
@@ -1263,7 +1264,10 @@ int ioreq_send(struct ioreq_server *s, ioreq_t *proto_p,
         return ioreq_send_buffered(s, proto_p);
 
     if ( unlikely(!vcpu_start_shutdown_deferral(curr)) )
+    {
+        vio->suspended = true;
         return IOREQ_STATUS_RETRY;
+    }
 
     list_for_each_entry ( sv,
                           &s->ioreq_vcpu_list,

xen/include/xen/sched.h

@@ -159,6 +159,11 @@ enum vio_completion {
 struct vcpu_io {
     /* I/O request in flight to device model. */
     enum vio_completion  completion;
+    /*
+     * Indicate whether the I/O was not handled because the domain
+     * is about to be paused.
+     */
+    bool                 suspended;
     ioreq_t              req;
 };