Fix 'unexpected data beyond EOF' on replica restart

On restart, a replica can fail with an error like 'unexpected data
beyond EOF in block 200 of relation T/D/R'. These are the steps to
reproduce it:

- A relation has a size of 400 blocks.
  - Blocks 201 to 400 are empty.
  - Block 200 has two rows.
  - Blocks 100 to 199 are empty.
- A restartpoint is done
- Vacuum truncates the relation to 200 blocks
- A FPW deletes a row in block 200
- A checkpoint is done
- A FPW deletes the last row in block 200
- Vacuum truncates the relation to 100 blocks
- The replica restarts

When the replica restarts:

- The relation on disk starts at 100 blocks, because all the
  truncations were applied before restart.
- The first truncate to 200 blocks is replayed. It silently fails, but
  it will still (incorrectly!) update the cache size to 200 blocks
- The first FPW on block 200 is applied. XLogReadBufferForRead relies
  on the cached size and incorrectly assumes that the page already
  exists in the file, and thus won't extend the relation.
- The online checkpoint record is replayed, calling smgrdestroyall
  which causes the cached size to be discarded
- The second FPW on block 200 is applied. This time, the detected size
  is 100 blocks, an extend is attempted. However, the block 200 is
  already present in the buffer cache due to the first FPW. This
  triggers the 'unexpected data beyond EOF'.

To fix, update the cached size in SmgrRelation with the current size
rather than the requested new size, when the requested new size is
greater.

Author: Anthonin Bonnefoy <anthonin.bonnefoy@datadoghq.com>
Discussion: https://www.postgresql.org/message-id/CAO6_Xqrv-snNJNhbj1KjQmWiWHX3nYGDgAc=vxaZP3qc4g1Siw@mail.gmail.com
Backpatch-through: 14

src/backend/storage/smgr/md.c

@@ -1282,6 +1282,9 @@ mdnblocks(SMgrRelation reln, ForkNumber forknum)
  * functions for this relation or handled interrupts in between.  This makes
  * sure we have opened all active segments, so that truncate loop will get
  * them all!
+ *
+ * If nblocks > curnblk, the request is ignored when we are InRecovery,
+ * otherwise, an error is raised.
  */
 void
 mdtruncate(SMgrRelation reln, ForkNumber forknum,

src/backend/storage/smgr/smgr.c

@@ -910,8 +910,17 @@ smgrtruncate(SMgrRelation reln, ForkNumber *forknum, int nforks,
 		 * backends to invalidate their copies of smgr_cached_nblocks, and
 		 * these ones too at the next command boundary. But ensure they aren't
 		 * outright wrong until then.
+		 *
+		 * We can have nblocks > old_nblocks when a relation was truncated
+		 * multiple times, a replica applied all the truncations, and later
+		 * restarts from a restartpoint located before the truncations. The
+		 * relation on disk will be the size of the last truncate. When
+		 * replaying the first truncate, we will have nblocks > current size.
+		 * In such cases, smgr_truncate does nothing, so set the cached size
+		 * to the old size rather than the requested size.
 		 */
-		reln->smgr_cached_nblocks[forknum[i]] = nblocks[i];
+		reln->smgr_cached_nblocks[forknum[i]] =
+			nblocks[i] > old_nblocks[i] ? old_nblocks[i] : nblocks[i];
 	}
 }