diff mbox series

jbd2: can b_transaction be NULL in refile_buffer ?

Message ID 20200611083417.4akdykeubd7kfuuh@work
State New
Headers show
Series jbd2: can b_transaction be NULL in refile_buffer ? | expand

Commit Message

Lukas Czerner June 11, 2020, 8:34 a.m. UTC 
Hi,

I am tracking a rare and very hard to reproduce bug that ends up hittng

J_ASSERT_JH(jh, jh->b_transaction == NULL)

in __journal_remove_journal_head(). In fact we can get there with
b_next_transaction set and b_jlist == BJ_Forget so it's clear that we
should not have dropped the last JH reference at that point.

Most of the time that I've seen we get there from
__jbd2_journal_remove_checkpoint() called from
jbd2_journal_commit_transaction().

The locking in and around grabbing and putting the journal head
reference (b_jcount) looks solid as well as the use of j_list_lock. But
I have noticed a problem in logic of __jbd2_journal_refile_buffer().

The idea is that b_next_transaction will inherit the reference from
b_transaction so that we do not need to grab a new reference of
journal_head. However this will only be true if b_transaction is set.
But if it is indeed NULL, then we will do

WRITE_ONCE(jh->b_transaction, jh->b_next_transaction);

and __jbd2_journal_file_buffer() will not grab the jh reference. AFAICT
the b_next_transaction is not holding it's own jh reference. This will
result in b_transaction _not_ holding it's own jh reference and we will
be able to drop the last jh reference at unexpected places - hence we can
hit the asserts in __journal_remove_journal_head().

However I am not really sure if it is indeed possible to get into
__jbd2_journal_refile_buffer() with b_transaction == NULL and
b_next_transaction set. Jan do you have any idea if that's possible and
what would be the circumstances to lead us there ?


Regardless I still think this is a bug in the logic and we should either
make sure that b_transaction is _not_ NULL in
__jbd2_journal_refile_buffer(), or let __jbd2_journal_file_buffer() grab
the jh reference if b_transaction was indeen NULL. How about something
like the following untested patch ?

Thanks!
-Lukas

Comments

Jan Kara June 11, 2020, 10:37 a.m. UTC | #1 
Hi!

On Thu 11-06-20 10:34:17, Lukas Czerner wrote:
> I am tracking a rare and very hard to reproduce bug that ends up hittng
> 
> J_ASSERT_JH(jh, jh->b_transaction == NULL)
> 
> in __journal_remove_journal_head(). In fact we can get there with
> b_next_transaction set and b_jlist == BJ_Forget so it's clear that we
> should not have dropped the last JH reference at that point.
> 
> Most of the time that I've seen we get there from
> __jbd2_journal_remove_checkpoint() called from
> jbd2_journal_commit_transaction().
> 
> The locking in and around grabbing and putting the journal head
> reference (b_jcount) looks solid as well as the use of j_list_lock. But
> I have noticed a problem in logic of __jbd2_journal_refile_buffer().

Yeah, the trouble with refcounting bugs is that if *any* of the users
releases a reference it should not, we will (much later) hit the problem you
describe.

> The idea is that b_next_transaction will inherit the reference from
> b_transaction so that we do not need to grab a new reference of
> journal_head. However this will only be true if b_transaction is set.
>
> But if it is indeed NULL, then we will do
> 
> WRITE_ONCE(jh->b_transaction, jh->b_next_transaction);
> 
> and __jbd2_journal_file_buffer() will not grab the jh reference. AFAICT
> the b_next_transaction is not holding it's own jh reference. This will
> result in b_transaction _not_ holding it's own jh reference and we will
> be able to drop the last jh reference at unexpected places - hence we can
> hit the asserts in __journal_remove_journal_head().
>
> However I am not really sure if it is indeed possible to get into
> __jbd2_journal_refile_buffer() with b_transaction == NULL and
> b_next_transaction set. Jan do you have any idea if that's possible and
> what would be the circumstances to lead us there ?

__jbd2_journal_refile_buffer() should always be called with b_transaction
!= NULL and as I've checked (all three) callers, that indeed seems to be
the case. Feel free to add assert along those lines to
__jbd2_journal_refile_buffer() to see whether it triggers...

> Regardless I still think this is a bug in the logic and we should either
> make sure that b_transaction is _not_ NULL in
> __jbd2_journal_refile_buffer(), or let __jbd2_journal_file_buffer() grab
> the jh reference if b_transaction was indeen NULL. How about something
> like the following untested patch ?

I'd rather got for the assert. It makes things simpler, also the "meaning"
of __jbd2_journal_refile_buffer() is "jh is done in its current
transaction, deal with it" and that doesn't have a great meaning if
b_transaction is NULL.

And when you're adding asserts, then adding one in
__jbd2_journal_unfile_buffer() checking b_transaction != NULL and
b_next_transaction == NULL would be good as well because lot of callers
assume this. I've checked the code and I didn't find any problematic one
but that code is complex enough that I could have missed something.

								Honza
Lukas Czerner June 11, 2020, 10:50 a.m. UTC | #2 
On Thu, Jun 11, 2020 at 12:37:09PM +0200, Jan Kara wrote:
> Hi!
> 
> On Thu 11-06-20 10:34:17, Lukas Czerner wrote:
> > I am tracking a rare and very hard to reproduce bug that ends up hittng
> > 
> > J_ASSERT_JH(jh, jh->b_transaction == NULL)
> > 
> > in __journal_remove_journal_head(). In fact we can get there with
> > b_next_transaction set and b_jlist == BJ_Forget so it's clear that we
> > should not have dropped the last JH reference at that point.
> > 
> > Most of the time that I've seen we get there from
> > __jbd2_journal_remove_checkpoint() called from
> > jbd2_journal_commit_transaction().
> > 
> > The locking in and around grabbing and putting the journal head
> > reference (b_jcount) looks solid as well as the use of j_list_lock. But
> > I have noticed a problem in logic of __jbd2_journal_refile_buffer().
> 
> Yeah, the trouble with refcounting bugs is that if *any* of the users
> releases a reference it should not, we will (much later) hit the problem you
> describe.
> 
> > The idea is that b_next_transaction will inherit the reference from
> > b_transaction so that we do not need to grab a new reference of
> > journal_head. However this will only be true if b_transaction is set.
> >
> > But if it is indeed NULL, then we will do
> > 
> > WRITE_ONCE(jh->b_transaction, jh->b_next_transaction);
> > 
> > and __jbd2_journal_file_buffer() will not grab the jh reference. AFAICT
> > the b_next_transaction is not holding it's own jh reference. This will
> > result in b_transaction _not_ holding it's own jh reference and we will
> > be able to drop the last jh reference at unexpected places - hence we can
> > hit the asserts in __journal_remove_journal_head().
> >
> > However I am not really sure if it is indeed possible to get into
> > __jbd2_journal_refile_buffer() with b_transaction == NULL and
> > b_next_transaction set. Jan do you have any idea if that's possible and
> > what would be the circumstances to lead us there ?
> 
> __jbd2_journal_refile_buffer() should always be called with b_transaction
> != NULL and as I've checked (all three) callers, that indeed seems to be
> the case. Feel free to add assert along those lines to
> __jbd2_journal_refile_buffer() to see whether it triggers...
> 
> > Regardless I still think this is a bug in the logic and we should either
> > make sure that b_transaction is _not_ NULL in
> > __jbd2_journal_refile_buffer(), or let __jbd2_journal_file_buffer() grab
> > the jh reference if b_transaction was indeen NULL. How about something
> > like the following untested patch ?
> 
> I'd rather got for the assert. It makes things simpler, also the "meaning"
> of __jbd2_journal_refile_buffer() is "jh is done in its current
> transaction, deal with it" and that doesn't have a great meaning if
> b_transaction is NULL.
> 
> And when you're adding asserts, then adding one in
> __jbd2_journal_unfile_buffer() checking b_transaction != NULL and
> b_next_transaction == NULL would be good as well because lot of callers
> assume this. I've checked the code and I didn't find any problematic one
> but that code is complex enough that I could have missed something.

Ok, thanks. I'll prepare a patch.

Regards,
-Lukas

> 
> 								Honza
> -- 
> Jan Kara <jack@suse.com>
> SUSE Labs, CR
>
diff mbox series

Patch

diff --git a/fs/jbd2/transaction.c b/fs/jbd2/transaction.c
index e91aad3637a2..55e5cb6b4bb5 100644
--- a/fs/jbd2/transaction.c
+++ b/fs/jbd2/transaction.c
@@ -2498,7 +2498,7 @@  void __jbd2_journal_file_buffer(struct journal_head *jh,
                __jbd2_journal_temp_unlink_buffer(jh);
        else
                jbd2_journal_grab_journal_head(bh);
-       jh->b_transaction = transaction;
+       WRITE_ONCE(jh->b_transaction, transaction);

        switch (jlist) {
        case BJ_None:
@@ -2577,15 +2577,14 @@  bool __jbd2_journal_refile_buffer(struct journal_head *jh)
         * our jh reference and thus __jbd2_journal_file_buffer() must not
         * take a new one.
         */
-       WRITE_ONCE(jh->b_transaction, jh->b_next_transaction);
-       WRITE_ONCE(jh->b_next_transaction, NULL);
        if (buffer_freed(bh))
                jlist = BJ_Forget;
        else if (jh->b_modified)
                jlist = BJ_Metadata;
        else
                jlist = BJ_Reserved;
-       __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
+       __jbd2_journal_file_buffer(jh, jh->b_next_transaction, jlist);
+       WRITE_ONCE(jh->b_next_transaction, NULL);
        J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);

        if (was_dirty)