Ojaswin Mujoo <ojaswin@xxxxxxxxxxxxx> writes:
> On Sun, Mar 09, 2025 at 12:11:22AM +0530, Ritesh Harjani wrote:
>> Ojaswin Mujoo <ojaswin@xxxxxxxxxxxxx> writes:
>>
>> > On Sat, Mar 08, 2025 at 06:56:23PM +0530, Ritesh Harjani wrote:
>> >> Ojaswin Mujoo <ojaswin@xxxxxxxxxxxxx> writes:
>> >>
>> >> > On Sat, Mar 08, 2025 at 03:25:04PM +0530, Ritesh Harjani (IBM) wrote:
>> >> >> Ojaswin Mujoo <ojaswin@xxxxxxxxxxxxx> writes:
>> >> >>
>> >> >> > Presently we always BUG_ON if trying to start a transaction on a journal marked
>> >> >> > with JBD2_UNMOUNT, since this should never happen. However, while ltp running
>> >> >> > stress tests, it was observed that in case of some error handling paths, it is
>> >> >> > possible for update_super_work to start a transaction after the journal is
>> >> >> > destroyed eg:
>> >> >> >
>> >> >> > (umount)
>> >> >> > ext4_kill_sb
>> >> >> > kill_block_super
>> >> >> > generic_shutdown_super
>> >> >> > sync_filesystem /* commits all txns */
>> >> >> > evict_inodes
>> >> >> > /* might start a new txn */
>> >> >> > ext4_put_super
>> >> >> > flush_work(&sbi->s_sb_upd_work) /* flush the workqueue */
>> >> >> > jbd2_journal_destroy
>> >> >> > journal_kill_thread
>> >> >> > journal->j_flags |= JBD2_UNMOUNT;
>> >> >> > jbd2_journal_commit_transaction
>> >> >> > jbd2_journal_get_descriptor_buffer
>> >> >> > jbd2_journal_bmap
>> >> >> > ext4_journal_bmap
>> >> >> > ext4_map_blocks
>> >> >> > ...
>> >> >> > ext4_inode_error
>> >> >> > ext4_handle_error
>> >> >> > schedule_work(&sbi->s_sb_upd_work)
>> >> >> >
>> >> >> > /* work queue kicks in */
>> >> >> > update_super_work
>> >> >> > jbd2_journal_start
>> >> >> > start_this_handle
>> >> >> > BUG_ON(journal->j_flags &
>> >> >> > JBD2_UNMOUNT)
>> >> >> >
>> >> >> > Hence, introduce a new sbi flag s_journal_destroying to indicate journal is
>> >> >> > destroying only do a journaled (and deferred) update of sb if this flag is not
>> >> >> > set. Otherwise, just fallback to an un-journaled commit.
>> >> >> >
>> >> >> > We set sbi->s_journal_destroying = true only after all the FS updates are done
>> >> >> > during ext4_put_super() (except a running transaction that will get commited
>> >> >> > during jbd2_journal_destroy()). After this point, it is safe to commit the sb
>> >> >> > outside the journal as it won't race with a journaled update (refer
>> >> >> > 2d01ddc86606).
>> >> >> >
>> >> >> > Also, we don't need a similar check in ext4_grp_locked_error since it is only
>> >> >> > called from mballoc and AFAICT it would be always valid to schedule work here.
>> >> >> >
>> >> >> > Fixes: 2d01ddc86606 ("ext4: save error info to sb through journal if available")
>> >> >> > Reported-by: Mahesh Kumar <maheshkumar657g@xxxxxxxxx>
>> >> >> > Suggested-by: Jan Kara <jack@xxxxxxx>
>> >> >> > Signed-off-by: Ojaswin Mujoo <ojaswin@xxxxxxxxxxxxx>
>> >> >> > ---
>> >> >> > fs/ext4/ext4.h | 2 ++
>> >> >> > fs/ext4/ext4_jbd2.h | 8 ++++++++
>> >> >> > fs/ext4/super.c | 4 +++-
>> >> >> > 3 files changed, 13 insertions(+), 1 deletion(-)
>> >> >> >
>> >> >> > diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
>> >> >> > index 2b7d781bfcad..d48e93bd5690 100644
>> >> >> > --- a/fs/ext4/ext4.h
>> >> >> > +++ b/fs/ext4/ext4.h
>> >> >> > @@ -1728,6 +1728,8 @@ struct ext4_sb_info {
>> >> >> > */
>> >> >> > struct work_struct s_sb_upd_work;
>> >> >> >
>> >> >> > + bool s_journal_destorying;
>> >> >> > +
>> >> >> > /* Atomic write unit values in bytes */
>> >> >> > unsigned int s_awu_min;
>> >> >> > unsigned int s_awu_max;
>> >> >> > diff --git a/fs/ext4/ext4_jbd2.h b/fs/ext4/ext4_jbd2.h
>> >> >> > index 9b3c9df02a39..6bd3ca84410d 100644
>> >> >> > --- a/fs/ext4/ext4_jbd2.h
>> >> >> > +++ b/fs/ext4/ext4_jbd2.h
>> >> >> > @@ -437,6 +437,14 @@ static inline int ext4_journal_destroy(struct ext4_sb_info *sbi, journal_t *jour
>> >> >> > {
>> >> >> > int err = 0;
>> >> >> >
>> >> >> > + /*
>> >> >> > + * At this point all pending FS updates should be done except a possible
>> >> >> > + * running transaction (which will commit in jbd2_journal_destroy). It
>> >> >> > + * is now safe for any new errors to directly commit superblock rather
>> >> >> > + * than going via journal.
>> >> >> > + */
>> >> >> > + sbi->s_journal_destorying = true;
>> >> >>
>> >> >> This is not correct right. I think what we decided to set this flag
>> >> >> before we flush the workqueue. So that we don't schedule any new
>> >> >> work after this flag has been set. At least that is what I understood.
>> >> >>
>> >> >> [1]: https://lore.kernel.org/all/87eczc6rlt.fsf@xxxxxxxxx/
>> >> >>
>> >> >> -ritesh
>> >> >
>> >> > Hey Ritesh,
>> >> >
>> >> > Yes that is not correct, I missed that in my patch however we realised
>> >> > that adding it before flush_work() also has issues [1]. More
>> >> > specifically:
>> >>
>> >> Ohk. right.
>> >>
>> >> >
>> >> > **kjournald2**
>> >> > jbd2_journal_commit_transaction()
>> >> > ...
>> >> > ext4_handle_error()
>> >> > /* s_journal_destorying is not set */
>> >> > if (journal && !s_journal_destorying)
>> >>
>> >> Then maybe we should not schedule another work to update the superblock
>> >> via journalling, it the error itself occurred while were trying to
>> >> commit the journal txn?
>> >>
>> >>
>> >> -ritesh
>> >
>> > Hmm, ideally yes that should not happen, but how can we achieve that?
>> > For example with the trace we saw:
>> >
>> > **kjournald2**
>> > jbd2_journal_commit_transaction()
>> > jbd2_journal_get_descriptor_buffer
>> > jbd2_journal_bmap
>> > ext4_journal_bmap
>> > ext4_map_blocks
>> > ...
>> > ext4_inode_error
>> > ext4_handle_error
>> > schedule_work(&sbi->s_sb_upd_work)
>> >
>> > How do we tell ext4_handle_error that it is in the context of a
>> > committing txn.
>>
>> So even if we identify that the current
>> jbd2_journal_commit_transaction() is coming from kjournald2(), that is
>> sufficient right? Because the only other place where we call
>> jbd2_journal_commit_transaction() is jbd2_journal_destroy() and that
>> happens after we can set few things from ext4_put_super() and flush work
>> is completed, correct?
>>
>>
>> > We can't pass down an argument all the way down
>> > cause that is not feasible. An sb level flag will also not work
>> > I think. Any thoughts on this?
>>
>> I was thinking if we should have a per task flag? Something like
>> PF_KJOURNALD? (Similar to how we have PF_KSWAPD or PF_KCOMPACTD)? This
>> can help us identify if we are a kjournald2() kthread.
>>
>> That will help prevent scheduling another work item to start a new
>> transaction in case an error occurs while committing the currently
>> running transaction. Correct?
>
> Yes, I like this approach. I think this will also help us avoid the
> extra checks in ext4_journal_destroy() since the journal will no longer
> schecule work for updating the sb. Hence we can be sure that after the
> final flush_work() noone will try to schedule more work or start a new
> transaction.
>
> I'll try to spin up a poc and test it. Does seem like we are out of flags in
> task struct though.
>
> Regards,
> ojaswin
>
>>
>> Now I don't know if we have any free bit available in current->flags. If
>> not shall we use current->journal_info pointer to have 0th bit as a
>> flag? Basically override current->journal_info to also store a flag. We
>> can create a wrapper to get the journal_info from current by masking
>> this flag bit and use it to dereference journal_info.
>
> Hmm so journal_info will be holding a kernel address of the handle. Is
> it possible to have it share a flag as well? I thought the address would
> utilize the full 64bits?
What I meant here was -
In general I assume we should be able to play some tricks with a pointer
which at least should be 4 byte aligned. So we can save some flag bits in
the lower bits of a pointer. For e.g. check struct address_space &
PAGE_MAPPING_FLAGS. There maybe few other examples too.
-ritesh
>
> Regards,
> ojaswin
>
>>
>> But before going down that road, it's better to know what others think?
>>
>> -ritesh
>>
>>
>> >
>> > regards,
>> > ojaswin
>> >
>> >>
>> >>
>> >> > ext4_put_super()
>> >> > sbi->s_journal_destorying = true;
>> >> > flush_work(&sbi->s_sb_upd_work)
>> >> > schedule_work()
>> >> > jbd2_journal_destroy()
>> >> > journal->j_flags |= JBD2_UNMOUNT;
>> >> >
>> >> > jbd2_journal_start()
>> >> > start_this_handle()
>> >> > BUG_ON(JBD2_UNMOUNT)
>> >> >
>> >> > So the right thing to do seems to be that we need to force a journal
>> >> > commit before the final flush as well. [1] Has more info on this and
>> >> > some followup discussion as well.
>> >> >
>> >> > [1] https://lore.kernel.org/all/cover.1741270780.git.ojaswin@xxxxxxxxxxxxx/T/#mc8046d47b357665bdbd2878c91e51eb660f94b3e
>> >> >
>> >> > Regards,
>> >> > ojaswin
>> >> >>
>> >> >>
>> >> >> > +
>> >> >> > err = jbd2_journal_destroy(journal);
>> >> >> > sbi->s_journal = NULL;
>> >> >> >
>> >> >> > diff --git a/fs/ext4/super.c b/fs/ext4/super.c
>> >> >> > index 8ad664d47806..31552cf0519a 100644
>> >> >> > --- a/fs/ext4/super.c
>> >> >> > +++ b/fs/ext4/super.c
>> >> >> > @@ -706,7 +706,7 @@ static void ext4_handle_error(struct super_block *sb, bool force_ro, int error,
>> >> >> > * constraints, it may not be safe to do it right here so we
>> >> >> > * defer superblock flushing to a workqueue.
>> >> >> > */
>> >> >> > - if (continue_fs && journal)
>> >> >> > + if (continue_fs && journal && !EXT4_SB(sb)->s_journal_destorying)
>> >> >> > schedule_work(&EXT4_SB(sb)->s_sb_upd_work);
>> >> >> > else
>> >> >> > ext4_commit_super(sb);
>> >> >> > @@ -5311,6 +5311,8 @@ static int __ext4_fill_super(struct fs_context *fc, struct super_block *sb)
>> >> >> > spin_lock_init(&sbi->s_error_lock);
>> >> >> > INIT_WORK(&sbi->s_sb_upd_work, update_super_work);
>> >> >> >
>> >> >> > + sbi->s_journal_destorying = false;
>> >> >> > +
>> >> >> > err = ext4_group_desc_init(sb, es, logical_sb_block, &first_not_zeroed);
>> >> >> > if (err)
>> >> >> > goto failed_mount3;
>> >> >> > --
>> >> >> > 2.48.1
