Re: [f2fs-dev] [PATCH] Revert "f2fs: give a warning only for readonly partition"

From: Chao Yu
Date: Tue Mar 30 2021 - 23:18:43 EST


On 2021/3/31 9:57, Jaegeuk Kim wrote:
On 03/27, Chao Yu wrote:
On 2021/3/27 9:52, Chao Yu wrote:
On 2021/3/27 1:30, Jaegeuk Kim wrote:
On 03/26, Chao Yu wrote:
On 2021/3/26 9:19, Jaegeuk Kim wrote:
On 03/26, Chao Yu wrote:
On 2021/3/25 9:59, Chao Yu wrote:
On 2021/3/25 6:44, Jaegeuk Kim wrote:
On 03/24, Chao Yu wrote:
On 2021/3/24 12:22, Jaegeuk Kim wrote:
On 03/24, Chao Yu wrote:
On 2021/3/24 2:39, Jaegeuk Kim wrote:
On 03/23, Chao Yu wrote:
This reverts commit 938a184265d75ea474f1c6fe1da96a5196163789.

Because that commit fails generic/050 testcase which expect failure
during mount a recoverable readonly partition.

I think we need to change generic/050, since f2fs can recover this partition,

Well, not sure we can change that testcase, since it restricts all generic
filesystems behavior. At least, ext4's behavior makes sense to me:

journal_dev_ro = bdev_read_only(journal->j_dev);
really_read_only = bdev_read_only(sb->s_bdev) | journal_dev_ro;

if (journal_dev_ro && !sb_rdonly(sb)) {
ext4_msg(sb, KERN_ERR,
"journal device read-only, try mounting with '-o ro'");
err = -EROFS;
goto err_out;
}

if (ext4_has_feature_journal_needs_recovery(sb)) {
if (sb_rdonly(sb)) {
ext4_msg(sb, KERN_INFO, "INFO: recovery "
"required on readonly filesystem");
if (really_read_only) {
ext4_msg(sb, KERN_ERR, "write access "
"unavailable, cannot proceed "
"(try mounting with noload)");
err = -EROFS;
goto err_out;
}
ext4_msg(sb, KERN_INFO, "write access will "
"be enabled during recovery");
}
}

even though using it as readonly. And, valid checkpoint can allow for user to
read all the data without problem.

if (f2fs_hw_is_readonly(sbi)) {

Since device is readonly now, all write to the device will fail, checkpoint can
not persist recovered data, after page cache is expired, user can see stale data.

My point is, after mount with ro, there'll be no data write which preserves the
current status. So, in the next time, we can recover fsync'ed data later, if
user succeeds to mount as rw. Another point is, with the current checkpoint, we
should not have any corrupted metadata. So, why not giving a chance to show what
data remained to user? I think this can be doable only with CoW filesystems.

I guess we're talking about the different things...

Let me declare two different readonly status:

1. filesystem readonly: file system is mount with ro mount option, and
app from userspace can not modify any thing of filesystem, but filesystem
itself can modify data on device since device may be writable.

2. device readonly: device is set to readonly status via 'blockdev --setro'
command, and then filesystem should never issue any write IO to the device.

So, what I mean is, *when device is readonly*, rather than f2fs mountpoint
is readonly (f2fs_hw_is_readonly() returns true as below code, instead of
f2fs_readonly() returns true), in this condition, we should not issue any
write IO to device anyway, because, AFAIK, write IO will fail due to
bio_check_ro() check.

In that case, mount(2) will try readonly, no?

Yes, if device is readonly, mount (2) can not mount/remount device to rw
mountpoint.

Any other concern about this patch?

Indeed we're talking about different things. :)

This case is mount(ro) with device(ro) having some data to recover.
My point is why not giving a chance to mount(ro) to show the current data
covered by a valid checkpoint. This doesn't change anything in the disk,
Got your idea.

IMO, it has potential issue in above condition:

Since device is readonly now, all write to the device will fail, checkpoint can
not persist recovered data, after page cache is expired, user can see stale data.

e.g.

Recovery writes one inode and then triggers a checkpoint, all writes fail

I'm confused. Currently we don't trigger the roll-forward recovery.

Oh, my miss, sorry. :-P

My point is in this condition we can return error and try to notice user to
mount with disable_roll_forward or norecovery option, then at least user can
know he should not expect last fsynced data in newly mounted image.

Or we can use f2fs_recover_fsync_data() to check whether there is fsynced data,
if there is no such data, then let mount() succeed.

Something like this, maybe:

---
fs/f2fs/super.c | 17 +++++++++++++----
1 file changed, 13 insertions(+), 4 deletions(-)

diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
index 954b1fe97d67..5e1a1caf412d 100644
--- a/fs/f2fs/super.c
+++ b/fs/f2fs/super.c
@@ -3966,10 +3966,19 @@ static int f2fs_fill_super(struct super_block *sb, void *data, int silent)
* previous checkpoint was not done by clean system shutdown.
*/
if (f2fs_hw_is_readonly(sbi)) {
- if (!is_set_ckpt_flags(sbi, CP_UMOUNT_FLAG))
- f2fs_err(sbi, "Need to recover fsync data, but write access unavailable");
- else
- f2fs_info(sbi, "write access unavailable, skipping recovery");
+ if (!is_set_ckpt_flags(sbi, CP_UMOUNT_FLAG)) {
+ err = f2fs_recover_fsync_data(sbi, true);

Can we do like this?

if (err > 0) {
err = -EROFS;
f2fs_err(sbi, "Need to recover fsync data, but "
"write access unavailable, please try "
"mount w/ disable_roll_forward or norecovery");
}
if (err < 0)
goto free_meta;
}
f2fs_info(sbi, "write access unavailable, skipping recovery");
goto reset_checkpoint;

More clear, revised in v2.

Thanks,


+ if (!err)
+ goto reset_checkpoint;
+ else if (err < 0)
+ goto free_meta;
+ err = -EROFS;
+ f2fs_err(sbi, "Need to recover fsync data, but "
+ "write access unavailable, please try "
+ "mount w/ disable_roll_forward or norecovery");
+ goto free_meta;
+ }
+ f2fs_info(sbi, "write access unavailable, skipping recovery");
goto reset_checkpoint;
}

--
2.29.2

Thanks,


Thanks,


due to device is readonly, once inode cache is reclaimed by vm, user will see
old inode when reloading it, or even see corrupted fs if partial meta inode's
cache is expired.

Thoughts?

Thanks,

and in the next time, it allows mount(rw|ro) with device(rw) to recover
the data seamlessly.


Thanks,


Thanks,


# blockdev --setro /dev/vdb
# mount -t f2fs /dev/vdb /mnt/test/
mount: /mnt/test: WARNING: source write-protected, mounted read-only.


if (f2fs_hw_is_readonly(sbi)) {
- if (!is_set_ckpt_flags(sbi, CP_UMOUNT_FLAG)) {
- err = -EROFS;
+ if (!is_set_ckpt_flags(sbi, CP_UMOUNT_FLAG))
f2fs_err(sbi, "Need to recover fsync data, but write access unavailable");
- goto free_meta;
- }
- f2fs_info(sbi, "write access unavailable, skipping recovery");
+ else
+ f2fs_info(sbi, "write access unavailable, skipping recovery");
goto reset_checkpoint;
}

For the case of filesystem is readonly and device is writable, it's fine
to do recovery in order to let user to see fsynced data.

Thanks,



Am I missing something?

Thanks,



Fixes: 938a184265d7 ("f2fs: give a warning only for readonly partition")
Signed-off-by: Chao Yu <yuchao0@xxxxxxxxxx>
---
fs/f2fs/super.c | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
index b48281642e98..2b78ee11f093 100644
--- a/fs/f2fs/super.c
+++ b/fs/f2fs/super.c
@@ -3952,10 +3952,12 @@ static int f2fs_fill_super(struct super_block *sb, void *data, int silent)
* previous checkpoint was not done by clean system shutdown.
*/
if (f2fs_hw_is_readonly(sbi)) {
- if (!is_set_ckpt_flags(sbi, CP_UMOUNT_FLAG))
+ if (!is_set_ckpt_flags(sbi, CP_UMOUNT_FLAG)) {
+ err = -EROFS;
f2fs_err(sbi, "Need to recover fsync data, but write access unavailable");
- else
- f2fs_info(sbi, "write access unavailable, skipping recovery");
+ goto free_meta;
+ }
+ f2fs_info(sbi, "write access unavailable, skipping recovery");
goto reset_checkpoint;
}
--
2.29.2
.

.

.



_______________________________________________
Linux-f2fs-devel mailing list
Linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel
.

.

.



_______________________________________________
Linux-f2fs-devel mailing list
Linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel
.

.