Avoid improbable PANIC during heap_update, redux.

Commit 34f581c39 intended to ensure that RelationGetBufferForTuple
would acquire a visibility-map page pin in case the otherBuffer's
all-visible bit had become set since we last had lock on that page.
But I missed a case: when we're extending the relation, VM concerns
were dealt with only in the relatively-less-likely case that we
fail to conditionally lock the otherBuffer.  I think I'd believed
that we couldn't need to worry about it if the conditional lock
succeeds, which is true for the target buffer; but the otherBuffer
was unlocked for awhile so its bit might be set anyway.  So we need
to do the GetVisibilityMapPins dance, and then also recheck the
page's free space, in both cases.

Per report from Jaime Casanova.  Back-patch to v12 as the previous
patch was (although there's still no evidence that the bug is
reachable pre-v14).

Discussion: https://postgr.es/m/[email protected]

diff --git a/src/backend/access/heap/hio.c b/src/backend/access/heap/hio.c
index ae2e2ce37a57ffa35c8d0beb97d6326dda5a4ea8..b0ece666293bdddf718a4cfb1796f66fb4ec19c3 100644 (file)
--- a/src/backend/access/heap/hio.c
+++ b/src/backend/access/heap/hio.c
@@ -678,29 +678,34 @@ loop:
            LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
            LockBuffer(otherBuffer, BUFFER_LOCK_EXCLUSIVE);
            LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
+       }
 
-           /*
-            * Because the buffers were unlocked for a while, it's possible,
-            * although unlikely, that an all-visible flag became set or that
-            * somebody used up the available space in the new page.  We can
-            * use GetVisibilityMapPins to deal with the first case.  In the
-            * second case, just retry from start.
-            */
-           GetVisibilityMapPins(relation, otherBuffer, buffer,
-                                otherBlock, targetBlock, vmbuffer_other,
-                                vmbuffer);
+       /*
+        * Because the buffers were unlocked for a while, it's possible,
+        * although unlikely, that an all-visible flag became set or that
+        * somebody used up the available space in the new page.  We can use
+        * GetVisibilityMapPins to deal with the first case.  In the second
+        * case, just retry from start.
+        */
+       GetVisibilityMapPins(relation, otherBuffer, buffer,
+                            otherBlock, targetBlock, vmbuffer_other,
+                            vmbuffer);
 
-           if (len > PageGetHeapFreeSpace(page))
-           {
-               LockBuffer(otherBuffer, BUFFER_LOCK_UNLOCK);
-               UnlockReleaseBuffer(buffer);
+       /*
+        * Note that we have to check the available space even if our
+        * conditional lock succeeded, because GetVisibilityMapPins might've
+        * transiently released lock on the target buffer to acquire a VM pin
+        * for the otherBuffer.
+        */
+       if (len > PageGetHeapFreeSpace(page))
+       {
+           LockBuffer(otherBuffer, BUFFER_LOCK_UNLOCK);
+           UnlockReleaseBuffer(buffer);
 
-               goto loop;
-           }
+           goto loop;
        }
    }
-
-   if (len > PageGetHeapFreeSpace(page))
+   else if (len > PageGetHeapFreeSpace(page))
    {
        /* We should not get here given the test at the top */
        elog(PANIC, "tuple is too big: size %zu", len);