dm: do not replace bioset for request based dm

This patch fixes a regression introduced in v3.8, which causes oops
like this when dm-multipath is used:

general protection fault: 0000 [#1] SMP
RIP: 0010:[<ffffffff810fe754>]  [<ffffffff810fe754>] mempool_free+0x24/0xb0
Call Trace:
  <IRQ>
  [<ffffffff81187417>] bio_put+0x97/0xc0
  [<ffffffffa02247a5>] end_clone_bio+0x35/0x90 [dm_mod]
  [<ffffffff81185efd>] bio_endio+0x1d/0x30
  [<ffffffff811f03a3>] req_bio_endio.isra.51+0xa3/0xe0
  [<ffffffff811f2f68>] blk_update_request+0x118/0x520
  [<ffffffff811f3397>] blk_update_bidi_request+0x27/0xa0
  [<ffffffff811f343c>] blk_end_bidi_request+0x2c/0x80
  [<ffffffff811f34d0>] blk_end_request+0x10/0x20
  [<ffffffffa000b32b>] scsi_io_completion+0xfb/0x6c0 [scsi_mod]
  [<ffffffffa000107d>] scsi_finish_command+0xbd/0x120 [scsi_mod]
  [<ffffffffa000b12f>] scsi_softirq_done+0x13f/0x160 [scsi_mod]
  [<ffffffff811f9fd0>] blk_done_softirq+0x80/0xa0
  [<ffffffff81044551>] __do_softirq+0xf1/0x250
  [<ffffffff8142ee8c>] call_softirq+0x1c/0x30
  [<ffffffff8100420d>] do_softirq+0x8d/0xc0
  [<ffffffff81044885>] irq_exit+0xd5/0xe0
  [<ffffffff8142f3e3>] do_IRQ+0x63/0xe0
  [<ffffffff814257af>] common_interrupt+0x6f/0x6f
  <EOI>
  [<ffffffffa021737c>] srp_queuecommand+0x8c/0xcb0 [ib_srp]
  [<ffffffffa0002f18>] scsi_dispatch_cmd+0x148/0x310 [scsi_mod]
  [<ffffffffa000a38e>] scsi_request_fn+0x31e/0x520 [scsi_mod]
  [<ffffffff811f1e57>] __blk_run_queue+0x37/0x50
  [<ffffffff811f1f69>] blk_delay_work+0x29/0x40
  [<ffffffff81059003>] process_one_work+0x1c3/0x5c0
  [<ffffffff8105b22e>] worker_thread+0x15e/0x440
  [<ffffffff8106164b>] kthread+0xdb/0xe0
  [<ffffffff8142db9c>] ret_from_fork+0x7c/0xb0

The regression was introduced by the change
c0820cf "dm: introduce per_bio_data", where dm started to replace
bioset during table replacement.
For bio-based dm, it is good because clone bios do not exist during the
table replacement.
For request-based dm, however, (not-yet-mapped) clone bios may stay in
request queue and survive during the table replacement.
So freeing the old bioset could cause the oops in bio_put().

Since the size of front_pad may change only with bio-based dm,
it is not necessary to replace bioset for request-based dm.

Reported-by: Bart Van Assche <[email protected]>
Tested-by: Bart Van Assche <[email protected]>
Signed-off-by: Jun'ichi Nomura <[email protected]>
Acked-by: Mikulas Patocka <[email protected]>
Acked-by: Mike Snitzer <[email protected]>
Cc: <[email protected]>
Signed-off-by: Alasdair G Kergon <[email protected]>

Author: nomuranec

drivers/md/dm.c

@@ -1947,15 +1947,27 @@ static void __bind_mempools(struct mapped_device *md, struct dm_table *t)
 {
 	struct dm_md_mempools *p = dm_table_get_md_mempools(t);
 
-	if (md->io_pool && (md->tio_pool || dm_table_get_type(t) == DM_TYPE_BIO_BASED) && md->bs) {
-		/*
-		 * The md already has necessary mempools. Reload just the
-		 * bioset because front_pad may have changed because
-		 * a different table was loaded.
-		 */
-		bioset_free(md->bs);
-		md->bs = p->bs;
-		p->bs = NULL;
+	if (md->io_pool && md->bs) {
+		/* The md already has necessary mempools. */
+		if (dm_table_get_type(t) == DM_TYPE_BIO_BASED) {
+			/*
+			 * Reload bioset because front_pad may have changed
+			 * because a different table was loaded.
+			 */
+			bioset_free(md->bs);
+			md->bs = p->bs;
+			p->bs = NULL;
+		} else if (dm_table_get_type(t) == DM_TYPE_REQUEST_BASED) {
+			BUG_ON(!md->tio_pool);
+			/*
+			 * There's no need to reload with request-based dm
+			 * because the size of front_pad doesn't change.
+			 * Note for future: If you are to reload bioset,
+			 * prep-ed requests in the queue may refer
+			 * to bio from the old bioset, so you must walk
+			 * through the queue to unprep.
+			 */
+		}
 		goto out;
 	}