jbd2,ext4: add a shrinker to release checkpointed buffers

Current metadata buffer release logic in bdev_try_to_free_page() have
a lot of use-after-free issues when umount filesystem concurrently, and
it is difficult to fix directly because ext4 is the only user of
s_op->bdev_try_to_free_page callback and we may have to add more special
refcount or lock that is only used by ext4 into the common vfs layer,
which is unacceptable.

One better solution is remove the bdev_try_to_free_page callback, but
the real problem is we cannot easily release journal_head on the
checkpointed buffer, so try_to_free_buffers() cannot release buffers and
page under memory pressure, which is more likely to trigger
out-of-memory. So we cannot remove the callback directly before we find
another way to release journal_head.

This patch introduce a shrinker to free journal_head on the checkpointed
transaction. After the journal_head got freed, try_to_free_buffers()
could free buffer properly.

Signed-off-by: Zhang Yi <yi.zhang@huawei.com>
Suggested-by: Jan Kara <jack@suse.cz>
Reviewed-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20210610112440.3438139-6-yi.zhang@huawei.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

1 files changed, 26 insertions, 0 deletions

diff --git a/include/linux/jbd2.h b/include/linux/jbd2.h
index d5db408ae064..6cc035321562 100644
--- a/include/linux/jbd2.h
+++ b/include/linux/jbd2.h
@@ -910,6 +910,29 @@ struct journal_s
 	struct buffer_head	*j_chkpt_bhs[JBD2_NR_BATCH];
 
 	/**
+	 * @j_shrinker:
+	 *
+	 * Journal head shrinker, reclaim buffer's journal head which
+	 * has been written back.
+	 */
+	struct shrinker		j_shrinker;
+
+	/**
+	 * @j_jh_shrink_count:
+	 *
+	 * Number of journal buffers on the checkpoint list. [j_list_lock]
+	 */
+	struct percpu_counter	j_jh_shrink_count;
+
+	/**
+	 * @j_shrink_transaction:
+	 *
+	 * Record next transaction will shrink on the checkpoint list.
+	 * [j_list_lock]
+	 */
+	transaction_t		*j_shrink_transaction;
+
+	/**
 	 * @j_head:
 	 *
 	 * Journal head: identifies the first unused block in the journal.
@@ -1422,6 +1445,7 @@ extern void jbd2_journal_commit_transaction(journal_t *);
 
 /* Checkpoint list management */
 void __jbd2_journal_clean_checkpoint_list(journal_t *journal, bool destroy);
+unsigned long jbd2_journal_shrink_checkpoint_list(journal_t *journal, unsigned long *nr_to_scan);
 int __jbd2_journal_remove_checkpoint(struct journal_head *);
 void jbd2_journal_destroy_checkpoint(journal_t *journal);
 void __jbd2_journal_insert_checkpoint(struct journal_head *, transaction_t *);
@@ -1532,6 +1556,8 @@ extern int	   jbd2_journal_set_features
 		   (journal_t *, unsigned long, unsigned long, unsigned long);
 extern void	   jbd2_journal_clear_features
 		   (journal_t *, unsigned long, unsigned long, unsigned long);
+extern int	   jbd2_journal_register_shrinker(journal_t *journal);
+extern void	   jbd2_journal_unregister_shrinker(journal_t *journal);
 extern int	   jbd2_journal_load       (journal_t *journal);
 extern int	   jbd2_journal_destroy    (journal_t *);
 extern int	   jbd2_journal_recover    (journal_t *journal);