mm/migrate: fix sleep in atomic for large folios and buffer heads

The large folio + buffer head noref migration scenarios are
being naughty and blocking while holding a spinlock.

As a consequence of the pagecache lookup path taking the
folio lock this serializes against migration paths, so
they can wait for each other. For the private_lock
atomic case, a new BH_Migrate flag is introduced which
enables the lookup to bail.

This allows the critical region of the private_lock on
the migration path to be reduced to the way it was before
ebdf4de ("mm: migrate: fix reference  check race
between __find_get_block() and migration"), that is covering
the count checks.

The scope is always noref migration.

Reported-by: kernel test robot <[email protected]>
Reported-by: [email protected]
Closes: https://lore.kernel.org/oe-lkp/[email protected]
Fixes: 3c20917 ("block/bdev: enable large folio support for large logical block sizes")
Reviewed-by: Jan Kara <[email protected]>
Co-developed-by: Luis Chamberlain <[email protected]>
Signed-off-by: Davidlohr Bueso <[email protected]>
Link: https://kdevops.org/ext4/v6.15-rc2.html # [0]
Link: https://lore.kernel.org/all/[email protected]/ # [1]
Link: https://lore.kernel.org/[email protected]
Tested-by: [email protected] # [0] [1]
Reviewed-by: Luis Chamberlain <[email protected]>
Signed-off-by: Christian Brauner <[email protected]>

Author: Davidlohr Bueso

fs/buffer.c

@@ -207,6 +207,15 @@ __find_get_block_slow(struct block_device *bdev, sector_t block, bool atomic)
 	head = folio_buffers(folio);
 	if (!head)
 		goto out_unlock;
+	/*
+	 * Upon a noref migration, the folio lock serializes here;
+	 * otherwise bail.
+	 */
+	if (test_bit_acquire(BH_Migrate, &head->b_state)) {
+		WARN_ON(!atomic);
+		goto out_unlock;
+	}
+
 	bh = head;
 	do {
 		if (!buffer_mapped(bh))
@@ -1390,7 +1399,8 @@ lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
 /*
  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
  * it in the LRU and mark it as accessed.  If it is not present then return
- * NULL
+ * NULL. Atomic context callers may also return NULL if the buffer is being
+ * migrated; similarly the page is not marked accessed either.
  */
 static struct buffer_head *
 find_get_block_common(struct block_device *bdev, sector_t block,

fs/ext4/ialloc.c

@@ -691,7 +691,8 @@ static int recently_deleted(struct super_block *sb, ext4_group_t group, int ino)
 	if (!bh || !buffer_uptodate(bh))
 		/*
 		 * If the block is not in the buffer cache, then it
-		 * must have been written out.
+		 * must have been written out, or, most unlikely, is
+		 * being migrated - false failure should be OK here.
 		 */
 		goto out;
 

include/linux/buffer_head.h

@@ -34,6 +34,7 @@ enum bh_state_bits {
 	BH_Meta,	/* Buffer contains metadata */
 	BH_Prio,	/* Buffer should be submitted with REQ_PRIO */
 	BH_Defer_Completion, /* Defer AIO completion to workqueue */
+	BH_Migrate,     /* Buffer is being migrated (norefs) */
 
 	BH_PrivateStart,/* not a state bit, but the first bit available
 			 * for private allocation by other entities

mm/migrate.c

@@ -845,9 +845,11 @@ static int __buffer_migrate_folio(struct address_space *mapping,
 		return -EAGAIN;
 
 	if (check_refs) {
-		bool busy;
+		bool busy, migrating;
 		bool invalidated = false;
 
+		migrating = test_and_set_bit_lock(BH_Migrate, &head->b_state);
+		VM_WARN_ON_ONCE(migrating);
 recheck_buffers:
 		busy = false;
 		spin_lock(&mapping->i_private_lock);
@@ -859,12 +861,12 @@ static int __buffer_migrate_folio(struct address_space *mapping,
 			}
 			bh = bh->b_this_page;
 		} while (bh != head);
+		spin_unlock(&mapping->i_private_lock);
 		if (busy) {
 			if (invalidated) {
 				rc = -EAGAIN;
 				goto unlock_buffers;
 			}
-			spin_unlock(&mapping->i_private_lock);
 			invalidate_bh_lrus();
 			invalidated = true;
 			goto recheck_buffers;
@@ -883,7 +885,7 @@ static int __buffer_migrate_folio(struct address_space *mapping,
 
 unlock_buffers:
 	if (check_refs)
-		spin_unlock(&mapping->i_private_lock);
+		clear_bit_unlock(BH_Migrate, &head->b_state);
 	bh = head;
 	do {
 		unlock_buffer(bh);