[Devel] Re: [PATCH] incorrect direct io error handling

[Dmitriy Monakhov]

Sorry for long delay (russian holidays are very hard time :) )

David Chinner <dgc at sgi.com> writes:
> On Tue, Dec 19, 2006 at 09:07:12AM +0300, Dmitriy Monakhov wrote:
>> David Chinner <dgc at sgi.com> writes:
>> > On Mon, Dec 18, 2006 at 04:22:44PM +0300, Dmitriy Monakhov wrote:
>> >> diff --git a/mm/filemap.c b/mm/filemap.c
>> >> index 8332c77..7c571dd 100644
>> >> --- a/mm/filemap.c
>> >> +++ b/mm/filemap.c
>
> <snip stuff>
>
>> > You comment in the first hunk that i_mutex may not be held here,
>> > but there's no comment in __generic_file_aio_write_nolock() that the
>> > i_mutex must be held for !S_ISBLK devices.
>> Any one may call directly call generic_file_direct_write() with i_mutex not held. 
>
> Only block devices based on the implementation (i.e. buffered I/O is
> done here).  but one can't call vmtruncate without the i_mutex held,
> so if a filesystem is calling generic_file_direct_write() it won't
> be able to use __generic_file_aio_write_nolock() without the i_mutex
> held (because it can right now if it doesn't need the buffered I/O
> fallback path), then 
>
>> >
>> >> @@ -2341,6 +2353,13 @@ ssize_t generic_file_aio_write_nolock(st
>> >>  	ssize_t ret;
>> >>  
>> >>  	BUG_ON(iocb->ki_pos != pos);
>> >> +	/*
>> >> +	 *  generic_file_buffered_write() may be called inside 
>> >> +	 *  __generic_file_aio_write_nolock() even in case of
>> >> +	 *  O_DIRECT for non S_ISBLK files. So i_mutex must be held.
>> >> +	 */
>> >> +	if (!S_ISBLK(inode->i_mode))
>> >> +		BUG_ON(!mutex_is_locked(&inode->i_mutex));
>> >>  
>> >>  	ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
>> >>  			&iocb->ki_pos);
>> >
>> > I note that you comment here in generic_file_aio_write_nolock(),
>> > but it's not immediately obvious that this is refering to the
>> > vmtruncate() call in __generic_file_aio_write_nolock().
>> This is not about vmtruncate(). __generic_file_aio_write_nolock() may 
>> call generic_file_buffered_write() even in case of O_DIRECT for !S_ISBLK, and 
>
> No, the need for i_mutex is currently dependent on doing direct I/O
> and the return value from generic_file_buffered_write().
> A filesystem that doesn't fall back to buffered I/O (e.g. XFS) can currently
> use generic_file_aio_write_nolock() without needing to hold i_mutex.
> use generic_file_aio_write_nolock() without needing to hold i_mutex.
But it doesn't use it. XFS implement it's own write method with it's own locking
rules and explicitly call generic_file_direct_write() in case of O_DIRECT.
BTW XFS correctly handling ENOSPC in case of O_DIRECT (fs corruption not happend
after error occur).

>
> Your change prevents that by introducing a vmtruncate() before the
> generic_file_buffered_write() return value check, which means that a
> filesystem now _must_ hold the i_mutex when calling
> generic_file_aio_write_nolock() even when it doesn't do buffered I/O
> through this path.
Yes it's so. But it is just explicitly document the fact that every fs call 
generic_file_aio_write_nolock() with i_mutex held (where is no any fs that
invoke it without i_mutex). As i understand Andrew Morton think so too: 
http://lkml.org/lkml/2006/12/12/67
<snip>
  I guess we can make that a rule (document it, add
  BUG_ON(!mutex_is_locked(..)) if it isn't a blockdev) if needs be.  After
  really checking that this matches reality for all callers.
<snip>

>
>> generic_file_buffered_write() has documented locking rules (i_mutex held).
>> IMHO it is important to explicitly document this . And after we realize
>> that i_mutex always held, vmtruncate() may be safely called.
>
> I don't think changing the locking semantics of
> generic_file_aio_write_nolock() to require a lock for all
> filesystem-based users is a good way to fix a filesystem specific
> direct I/O problem which can be easily fixed in filesystem specific
> code - i.e. call vmtruncate() in ext3_file_write() on failure....
Where are more than 10 filesystems where we have to fix it then.
And fix is almost the same for all fs, so we have to do many copy/paste work
IMHO fix it inside generic_file_aio_write_nolock is realy straightforward way.
What do you think?
>
> Cheers,
>
> Dave.
> -- 
> Dave Chinner
> Principal Engineer
> SGI Australian Software Group