x86/mm: Don't re-set PGT_pinned on a partially de-validated page

When unpinning pagetables, if an operation is interrupted,
relinquish_memory() re-sets PGT_pinned so that the un-pin will
pickedup again when the hypercall restarts.

This is appropriate when put_page_and_type_preemptible() returns
-EINTR, which indicates that the page is back in its initial state
(i.e., completely validated).  However, for -ERESTART, this leads to a
state where a page has both PGT_pinned and PGT_partial set.

This happens to work at the moment, although it's not really a
"canonical" state; but in subsequent patches, where we need to make a
distinction in handling between PGT_validated and PGT_partial pages,
this causes issues.

Move to a "canonical" state by:
- Only re-setting PGT_pinned on -EINTR
- Re-dropping the refcount held by PGT_pinned on -ERESTART

In the latter case, the PGT_partial bit will be cleared further down
with the rest of the other PGT_partial pages.

While here, clean up some trainling whitespace.

This is part of XSA-299.

Reported-by: George Dunlap <george.dunlap@citrix.com>
Signed-off-by: George Dunlap <george.dunlap@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>

xen/arch/x86/domain.c

@@ -120,7 +120,7 @@ void play_dead(void)
      * this case, heap corruption or #PF can occur (when heap debugging is
      * enabled). For example, even printk() can involve tasklet scheduling,
      * which touches per-cpu vars.
-     * 
+     *
      * Consider very carefully when adding code to *dead_idle. Most hypervisor
      * subsystems are unsafe to call.
      */
@@ -1972,9 +1972,34 @@ static int relinquish_memory(
             break;
         case -ERESTART:
         case -EINTR:
+            /*
+             * -EINTR means PGT_validated has been re-set; re-set
+             * PGT_pinned again so that it gets picked up next time
+             * around.
+             *
+             * -ERESTART, OTOH, means PGT_partial is set instead.  Put
+             * it back on the list, but don't set PGT_pinned; the
+             * section below will finish off de-validation.  But we do
+             * need to drop the general ref associated with
+             * PGT_pinned, since put_page_and_type_preemptible()
+             * didn't do it.
+             *
+             * NB we can do an ASSERT for PGT_validated, since we
+             * "own" the type ref; but theoretically, the PGT_partial
+             * could be cleared by someone else.
+             */
+            if ( ret == -EINTR )
+            {
+                ASSERT(page->u.inuse.type_info & PGT_validated);
+                set_bit(_PGT_pinned, &page->u.inuse.type_info);
+            }
+            else
+                put_page(page);
+
             ret = -ERESTART;
+
+            /* Put the page back on the list and drop the ref we grabbed above */
             page_list_add(page, list);
-            set_bit(_PGT_pinned, &page->u.inuse.type_info);
             put_page(page);
             goto out;
         default:
@@ -2225,7 +2250,7 @@ void vcpu_kick(struct vcpu *v)
      * pending flag. These values may fluctuate (after all, we hold no
      * locks) but the key insight is that each change will cause
      * evtchn_upcall_pending to be polled.
-     * 
+     *
      * NB2. We save the running flag across the unblock to avoid a needless
      * IPI for domains that we IPI'd to unblock.
      */