On Fri 26-03-21 09:18:45, Alexander Lochmann wrote:
> On 11.02.21 10:30, Jan Kara wrote:
> >> diff --git a/include/linux/jbd2.h b/include/linux/jbd2.h
> >> index 99d3cd051ac3..18f77d9b1745 100644
> >> --- a/include/linux/jbd2.h
> >> +++ b/include/linux/jbd2.h
> >> @@ -594,18 +594,18 @@ struct transaction_s
> >> */
> >> unsigned long t_log_start;
> >>
> >> - /* Number of buffers on the t_buffers list [j_list_lock] */
> >> + /* Number of buffers on the t_buffers list [j_list_lock, no lock for quick racy checks] */
> >> int t_nr_buffers;
> >
> > So this case is actually somewhat different now that I audited the uses.
> > There are two types of users - commit code (fs/jbd2/commit.c) and others.
> > Other users properly use j_list_lock to access t_nr_buffers. Commit code
> > does not use any locks because committing transaction is fully in
> > ownership of the jbd2 thread and all other users need to check & wait for
> > commit to be finished before doing anything with the transaction's buffers.
>
> I'm still trying understand how thinks work:
> Accesses to transaction_t might occur from different contexts. Thus,
> locks are necessary. If it comes to the commit phase, every other
> context has to wait until jbd2 thread has done its work. Therefore, jbd2
> thread does not need any locks to access a transaction_t (or just parts
> of it?) during commit phase.
> Is that correct?
Yes, that is correct.
> If so: I was thinking whether it make sense to ignore all memory
> accesses to a transaction_t (or parts of it) that happen in the commit
> phase. They deliberately ignore the locking policy, and would confuse
> our approach.
>
> Is the commit phase performed by jbd2_journal_commit_transaction()?
> We would add this function to our blacklist for transaction_t.
Yes, commit phase is implemented by jbd2_journal_commit_transaction() and
the functions it calls.
Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR