Skip to contentSkip to conversation

dm: also check for conflicting table type in dm_table_set_type

PATCH In review Mike Snitzer 2 participants 2 messages May 20, 2010 – May 21, 2010 Device mapper

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 dm: also check for conflicting table type in dm_table_set_type

Factorize live table type compatibility check into dm_table_type_matches_live_table(). Catching a conflicting table type early, during table load, avoids failing during resume -- which always leaves the DM device suspended.

So also check type during dm_table_set_type() -- which happens as part of a table load. But preserve the existing check during resume, in dm_swap_table(), because a racing table load could stage an inactive table that conflicts with the table type that dm_swap_table() is about to make live:

table_load() for bio-based do_resume() for rq-based
-----------------------------------------------------------------
dm_table_type_matches_live_table(t)

                                         down_write(_hash_lock)
                                         new_map = hc->new_map
                                         up_write(_hash_lock)
down_write(_hash_lock)
hc->new_map = t
up_write(_hash_lock)
                                         dm_swap_table(md, new_map)
                                         dm_table_type_matches_live_table(t)
                                         __bind(md, new_map)
---
 drivers/md/dm-table.c |   24 +++++++++++++++++++++++-
 drivers/md/dm.c       |    6 +-----
 drivers/md/dm.h       |    1 +
 3 files changed, 25 insertions(+), 6 deletions(-)
drivers/md/dm-table.c +23 -1
Index: linux-2.6/drivers/md/dm-table.c
===================================================================
@@ -803,7 +803,7 @@ int dm_table_set_type(struct dm_table *t
 	if (bio_based) {
 		/* We must use this table as bio-based */
 		t->type = DM_TYPE_BIO_BASED;
-		return 0;
+		goto finish;
 	}
 
 	BUG_ON(!request_based); /* No targets in this table */
@@ -831,6 +831,10 @@ int dm_table_set_type(struct dm_table *t
 
 	t->type = DM_TYPE_REQUEST_BASED;
 
+finish:
+	if (!dm_table_type_matches_live_table(t))
+		return -EINVAL;
+
 	return 0;
 }
 
@@ -839,6 +843,24 @@ unsigned dm_table_get_type(struct dm_tab
 	return t->type;
 }
 
+int dm_table_type_matches_live_table(struct dm_table *t)
+{
+	int r = 1;
+	struct dm_table *live_table = NULL;
+
+	live_table = dm_get_live_table(t->md);
+	if (!live_table)
+		return 1; /* no live table, return success */
+
+	if (dm_table_get_type(live_table) != dm_table_get_type(t)) {
+		DMWARN("can't change the device type after a table is bound");
+		r = 0;
+	}
+	dm_table_put(live_table);
+
+	return r;
+}
+
 bool dm_table_request_based(struct dm_table *t)
 {
 	return dm_table_get_type(t) == DM_TYPE_REQUEST_BASED;
drivers/md/dm.c +1 -5
Index: linux-2.6/drivers/md/dm.c
===================================================================
@@ -2403,12 +2403,8 @@ struct dm_table *dm_swap_table(struct ma
 		goto out;
 	}
 
-	/* cannot change the device type, once a table is bound */
-	if (md->map &&
-	    (dm_table_get_type(md->map) != dm_table_get_type(table))) {
-		DMWARN("can't change the device type after a table is bound");
+	if (!dm_table_type_matches_live_table(table))
 		goto out;
-	}
 
 	map = __bind(md, table, &limits);
 
drivers/md/dm.h +1 -0
Index: linux-2.6/drivers/md/dm.h
===================================================================
@@ -61,6 +61,7 @@ int dm_table_any_congested(struct dm_tab
 int dm_table_any_busy_target(struct dm_table *t);
 int dm_table_set_type(struct dm_table *t);
 unsigned dm_table_get_type(struct dm_table *t);
+int dm_table_type_matches_live_table(struct dm_table *t);
 bool dm_table_request_based(struct dm_table *t);
 int dm_table_alloc_md_mempools(struct dm_table *t);
 void dm_table_free_md_mempools(struct dm_table *t);
#1 lore raw

Hi Mike,

Quoted 10 lines · Mike Snitzer

Factorize live table type compatibility check into dm_table_type_matches_live_table(). Catching a conflicting table type early, during table load, avoids failing during resume -- which always leaves the DM device suspended.

So also check type during dm_table_set_type() -- which happens as part of a table load. But preserve the existing check during resume, in dm_swap_table(), because a racing table load could stage an inactive table that conflicts with the table type that dm_swap_table() is about to make live:

As you noted, the additional check in dm_table_set_type() is racy and has no guarantee to detect problematic table loading in some cases. Although I'm not sure such a duplicated racy check is really needed, it will detect most cases and users may be happy with that.

So I have no objection, but at least, please put some comments that it's racy in the code below.

drivers/md/dm-table.c int dm_table_set_type(struct dm_table *t · line 837

Quoted 4 lines
drivers/md/dm-table.c +4 -0
+finish:
+	if (!dm_table_type_matches_live_table(t))
+		return -EINVAL;
+

Thanks,
Kiyoshi Ueda