On 2023/3/23 19:44, Jan Kara wrote:
On Thu 16-03-23 19:28:32, Baokun Li wrote:
Above issue may happen as follows:
cpu1 cpu2
______________________________|_____________________________
mount -o dioread_lock
ext4_writepages
ext4_do_writepages
*if (ext4_should_dioread_nolock(inode))*
// rsv_blocks is not assigned here
mount -o remount,dioread_nolock
ext4_journal_start_with_reserve
__ext4_journal_start
__ext4_journal_start_sb
jbd2__journal_start
*if (rsv_blocks)*
// h_rsv_handle is not initialized here
mpage_map_and_submit_extent
mpage_map_one_extent
dioread_nolock = ext4_should_dioread_nolock(inode)
if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN))
mpd->io_submit.io_end->handle = handle->h_rsv_handle
ext4_set_io_unwritten_flag
io_end->flag |= EXT4_IO_END_UNWRITTEN
// now io_end->handle is NULL but has EXT4_IO_END_UNWRITTEN flag
scsi_finish_command
scsi_io_completion
scsi_io_completion_action
scsi_end_request
blk_update_request
req_bio_endio
bio_endio
bio->bi_end_io > ext4_end_bio
ext4_put_io_end_defer
ext4_add_complete_io
// trigger WARN_ON(!io_end->handle && sbi->s_journal);
The immediate cause of this problem is that ext4_should_dioread_nolock()
function returns inconsistent values in the ext4_do_writepages() and
mpage_map_one_extent(). There are four conditions in this function that
can be changed at mount time to cause this problem. These four conditions
can be divided into two categories:
(1) journal_data and EXT4_EXTENTS_FL, which can be changed by ioctl
(2) DELALLOC and DIOREAD_NOLOCK, which can be changed by remount
The two in the first category have been fixed by commit c8585c6fcaf2
("ext4: fix races between changing inode journal mode and ext4_writepages")
and commit cb85f4d23f79 ("ext4: fix race between writepages and enabling
EXT4_EXTENTS_FL") respectively.
Two cases in the other category have not yet been fixed, and the above
issue is caused by this situation. We refer to the fix for the first
category, When DELALLOC or DIOREAD_NOLOCK is detected to be changed
during remount, we hold the s_writepages_rwsem lock to avoid racing with
ext4_writepages to trigger the problem.
Moreover, we add an EXT4_MOUNT_SHOULD_DIOREAD_NOLOCK macro to ensure that
the mount options used by ext4_should_dioread_nolock() and __ext4_remount()
are always consistent.
Fixes: 6b523df4fb5a ("ext4: use transaction reservation for extent conversion in ext4_end_io")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Baokun Li <libaokun1@xxxxxxxxxx>
Nice catch! One comment below:
Thank you very much for your review!
---
fs/ext4/ext4.h | 3 ++-
fs/ext4/ext4_jbd2.h | 9 +++++----
fs/ext4/super.c | 14 ++++++++++++++
3 files changed, 21 insertions(+), 5 deletions(-)
diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
index 08b29c289da4..f60967fa648f 100644
--- a/fs/ext4/ext4.h
+++ b/fs/ext4/ext4.h
@@ -1703,7 +1703,8 @@ struct ext4_sb_info {
/*
* Barrier between writepages ops and changing any inode's JOURNAL_DATA
- * or EXTENTS flag.
+ * or EXTENTS flag or between changing SHOULD_DIOREAD_NOLOCK flag on
+ * remount and writepages ops.
*/
struct percpu_rw_semaphore s_writepages_rwsem;
struct dax_device *s_daxdev;
diff --git a/fs/ext4/ext4_jbd2.h b/fs/ext4/ext4_jbd2.h
index 0c77697d5e90..d82bfcdd56e5 100644
--- a/fs/ext4/ext4_jbd2.h
+++ b/fs/ext4/ext4_jbd2.h
@@ -488,6 +488,9 @@ static inline int ext4_free_data_revoke_credits(struct inode *inode, int blocks)
return blocks + 2*(EXT4_SB(inode->i_sb)->s_cluster_ratio - 1);
}
+/* delalloc is a temporary fix to prevent generic/422 test failures*/
+#define EXT4_MOUNT_SHOULD_DIOREAD_NOLOCK (EXT4_MOUNT_DIOREAD_NOLOCK | \
+ EXT4_MOUNT_DELALLOC)
/*
* This function controls whether or not we should try to go down the
* dioread_nolock code paths, which makes it safe to avoid taking
@@ -499,7 +502,8 @@ static inline int ext4_free_data_revoke_credits(struct inode *inode, int blocks)
*/
static inline int ext4_should_dioread_nolock(struct inode *inode)
{
- if (!test_opt(inode->i_sb, DIOREAD_NOLOCK))
+ if (test_opt(inode->i_sb, SHOULD_DIOREAD_NOLOCK) !=
+ EXT4_MOUNT_SHOULD_DIOREAD_NOLOCK)
return 0;
if (!S_ISREG(inode->i_mode))
return 0;
@@ -507,9 +511,6 @@ static inline int ext4_should_dioread_nolock(struct inode *inode)
return 0;
if (ext4_should_journal_data(inode))
return 0;
- /* temporary fix to prevent generic/422 test failures */
- if (!test_opt(inode->i_sb, DELALLOC))
- return 0;
return 1;
}
Is there a need for this SHOULD_DIOREAD_NOLOCK? When called from writeback
we will be protected by s_writepages_rwsem anyway. When called from other
places, we either decide to do dioread_nolock or don't but the situation
can change at any instant so I don't see how unifying this check would
help. And the new SHOULD_DIOREAD_NOLOCK somewhat obfuscates what's going
on.
We're thinking that the mount-related flags in
ext4_should_dioread_nolock() might
be modified, such as DELALLOC being removed because generic/422 test
failures
were fixed in some other way, resulting in some unnecessary locking
during remount,
or for whatever reason a mount-related flag was added to
ext4_should_dioread_nolock(),
and we didn't make a synchronization change in __ext4_remount() that
would cause
the problem to recur. So we added this flag to this function (instead
of in ext4.h), so that
when we change the mount option in ext4_should_dioread_nolock(), we
directly change
this flag, and we don't have to consider making synchronization changes
in __ext4_remount().
We have checked where this function is called and there are two types of
calls to this function:
1. One category is ext4_do_writepages() and mpage_map_one_extent(),
which are protected
by s_writepages_rwsem, the location of the problem;
2. The other type is in ext4_page_mkwrite(),
ext4_convert_inline_data_to_extent(),ext4_write_begin()
to determine whether to get the block using
ext4_get_block_unwritten() or ext4_get_ block().
1) If we just started fetching written blocks, it looks like there
is no problem;
2) If we start getting unwritten blocks, when DIOREAD_NOLOCK is
cleared by remount,
we will convert the blocks to written in ext4_map_blocks(). The
data=ordered mode
ensures that we don't see stale data.
diff --git a/fs/ext4/super.c b/fs/ext4/super.c
index fefcd42f34ea..bdf6b288aeff 100644
--- a/fs/ext4/super.c
+++ b/fs/ext4/super.c
@@ -6403,8 +6403,22 @@ static int __ext4_remount(struct fs_context *fc, struct super_block *sb)
}
+ /* Get the flag we really need to set/clear. */
+ ctx->mask_s_mount_opt &= sbi->s_mount_opt;
+ ctx->vals_s_mount_opt &= ~sbi->s_mount_opt;
+
+ /*
+ * If EXT4_MOUNT_SHOULD_DIOREAD_NOLOCK change on remount, we need
+ * to hold s_writepages_rwsem to avoid racing with writepages ops.
+ */
+ if (ctx_changed_mount_opt(ctx, EXT4_MOUNT_SHOULD_DIOREAD_NOLOCK))
+ percpu_down_write(&sbi->s_writepages_rwsem);
+
Honestly, I'd be inclined to grab s_writepages_rwsem unconditionally during
remount. Remount is not a fast path operation and waiting for writepages
isn't too bad. Also it's easier for testing :).
Honza
Initially I was concerned that adding this lock to remount would cause
users to
wait too long when remounting, so I added the lock only when the **actual**
mount flags being modified **contained** DIOREAD_NOLOCK or DELALLOC.
But now that I think about it, if there is no better way to do it, it
seems to be
acceptable to add the lock unconditionally, so that the code looks cleaner.
Thanks again!
Any comments are welcome!
--
With Best Regards,
Baokun Li
.
