Re: [PATCH v9 4/4] xfs: replace mrlock_t with rw_semaphores

[Pavel Reichl <preichl@redhat.com>]

On 10/7/20 5:25 PM, Darrick J. Wong wrote:
> On Wed, Oct 07, 2020 at 09:17:13AM -0500, Eric Sandeen wrote:
>> On 10/6/20 8:21 PM, Darrick J. Wong wrote:
>>> On Tue, Oct 06, 2020 at 09:15:41PM +0200, Pavel Reichl wrote:
>>>> Remove mrlock_t as it does not provide any extra value over
>>>> rw_semaphores. Make i_lock and i_mmaplock native rw_semaphores and
>>>> replace mr*() functions with native rwsem calls.
>>>>
>>>> Release the lock in xfs_btree_split() just before the work-queue
>>>> executing xfs_btree_split_worker() is scheduled and make
>>>> xfs_btree_split_worker() to acquire the lock as a first thing and
>>>> release it just before returning from the function. This it done so the
>>>> ownership of the lock is transfered between kernel threads and thus
>>>> lockdep won't complain about lock being held by a different kernel
>>>> thread.
>>>>
>>>> Signed-off-by: Pavel Reichl <preichl@redhat.com>
>>>> ---
>>>>  fs/xfs/libxfs/xfs_btree.c | 14 +++++++
>>>>  fs/xfs/mrlock.h           | 78 ---------------------------------------
>>>>  fs/xfs/xfs_inode.c        | 36 ++++++++++--------
>>>>  fs/xfs/xfs_inode.h        |  4 +-
>>>>  fs/xfs/xfs_iops.c         |  4 +-
>>>>  fs/xfs/xfs_linux.h        |  2 +-
>>>>  fs/xfs/xfs_super.c        |  6 +--
>>>>  7 files changed, 41 insertions(+), 103 deletions(-)
>>>>  delete mode 100644 fs/xfs/mrlock.h
>>>>
>>>> diff --git a/fs/xfs/libxfs/xfs_btree.c b/fs/xfs/libxfs/xfs_btree.c
>>>> index 2d25bab68764..1d1bb8423688 100644
>>>> --- a/fs/xfs/libxfs/xfs_btree.c
>>>> +++ b/fs/xfs/libxfs/xfs_btree.c
>>>> @@ -2816,6 +2816,7 @@ xfs_btree_split_worker(
>>>>  	unsigned long		pflags;
>>>>  	unsigned long		new_pflags = PF_MEMALLOC_NOFS;
>>>>  
>>>> +	rwsem_acquire(&args->cur->bc_ino.ip->i_lock.dep_map, 0, 0, _RET_IP_);
>>> These calls also need a comment explaining just what they're doing.
>>>
>>>>  	/*
>>>>  	 * we are in a transaction context here, but may also be doing work
>>>>  	 * in kswapd context, and hence we may need to inherit that state
>>>> @@ -2832,6 +2833,7 @@ xfs_btree_split_worker(
>>>>  	complete(args->done);
>>>>  
>>>>  	current_restore_flags_nested(&pflags, new_pflags);
>>>> +	rwsem_release(&args->cur->bc_ino.ip->i_lock.dep_map, _THIS_IP_);
>>> Note that as soon as you call complete(), xfs_btree_split can wake up
>>> and return, which means that *args could now point to reclaimed stack
>>> space.  This leads to crashes and memory corruption in generic/562 on
>>> a 1k block filesystem (though in principle this can happen anywhere):
>>
>>
>> What's the right way out of this; store *ip when we enter the function
>> and use that to get to the map, rather than args i guess?
> 
> Er, no, because the worker could also get preempted right after
> complete() and take so long to get rescheduled that the the inode have
> been reclaimed.  Think about it -- the original thread is waiting on the
> completion that it passed to the worker through $args, and therefore the
> worker cannot touch any of the resources it was accessing through $args
> after calling complete()....

Hi,

thanks for the comments, however for some reason I cannot reproduce the same memory corruption you are getting.

Do you think that moving the 'rwsem_release()' right before the 'complete()' should fix the problem?

Something like:


+       /*
+        * Update lockdep's lock ownership information to point to
+        * this thread as the thread that scheduled this worker is waiting
+        * for it's completion.
+        */
        rwsem_acquire(&args->cur->bc_ino.ip->i_lock.dep_map, 0, 0, _RET_IP_);
        /*
         * we are in a transaction context here, but may also be doing work
@@ -2830,10 +2835,15 @@ xfs_btree_split_worker(
 
        args->result = __xfs_btree_split(args->cur, args->level, args->ptrp,
                                         args->key, args->curp, args->stat);
+       /*
+        * Update lockdep's lock ownership information to reflect that we will
+        * be transferring the ilock from this worker back to the scheduling
+        * thread.
+        */
+       rwsem_release(&args->cur->bc_ino.ip->i_lock.dep_map, _THIS_IP_);
        complete(args->done);
 
        current_restore_flags_nested(&pflags, new_pflags);
-       rwsem_release(&args->cur->bc_ino.ip->i_lock.dep_map, _THIS_IP_);



> 
> --D
> 
>> Thanks,
>> -Eric
>