On Tue 26-11-19 18:15:25, Ritesh Harjani wrote: > On 11/26/19 4:21 PM, Ritesh Harjani wrote: > > On 11/20/19 8:02 PM, Jan Kara wrote: > > > On Wed 20-11-19 10:30:24, Ritesh Harjani wrote: > > > > We were using shared locking only in case of dioread_nolock > > > > mount option in case of DIO overwrites. This mount condition > > > > is not needed anymore with current code, since:- > > > > > > > > 1. No race between buffered writes & DIO overwrites. > > > > Since buffIO writes takes exclusive locks & DIO overwrites > > > > will take share locking. Also DIO path will make sure > > > > to flush and wait for any dirty page cache data. > > > > > > > > 2. No race between buffered reads & DIO overwrites, since there > > > > is no block allocation that is possible with DIO overwrites. > > > > So no stale data exposure should happen. Same is the case > > > > between DIO reads & DIO overwrites. > > > > > > > > 3. Also other paths like truncate is protected, > > > > since we wait there for any DIO in flight to be over. > > > > > > > > 4. In case of buffIO writes followed by DIO reads: > > > > Since here also we take exclusive locks in ext4_write_begin/end(). > > > > There is no risk of exposing any stale data in this case. > > > > Since after ext4_write_end, iomap_dio_rw() will wait to flush & > > > > wait for any dirty page cache data. > > > > > > > > Signed-off-by: Ritesh Harjani <riteshh@xxxxxxxxxxxxx> > > > > > > There's one more case to consider here as I mentioned in [1]. There > > > can be > > > > Yes, I should have mentioned about this in cover letter and about my > > thoughts on that. > > I was of the opinion that since the race is already existing > > and it may not be caused due to this patch, so we should handle that in > > incremental fashion and as a separate patch series after this one. > > Let me know your thoughts on above. > > > > Also, I wanted to have some more discussions on this race before > > making the changes. > > But nevertheless, it's the right time to discuss those changes here. > > > > > mmap write instantiating dirty page and then someone starting writeback > > > against that page while DIO read is running still theoretically > > > leading to > > > stale data exposure. Now this patch does not have influence on that race > > > but: > > > > Yes, agreed. > > > > > > > > 1) We need to close the race mentioned above. Maybe we could do that by > > > proactively allocating unwritten blocks for a page being faulted > > > when there > > > is direct IO running against the file - the one who fills holes through > > > mmap write while direct IO is running on the file deserves to suffer the > > > performance penalty... > > > > I was giving this a thought. So even if we try to penalize mmap > > write as you mentioned above, what I am not sure about it, is that, how > > can we reliably detect that the DIO is in progress? > > > > Say even if we try to check for atomic_read(&inode->i_dio_count) in mmap > > ext4_page_mkwrite path, it cannot be reliable unless there is some sort > > of a lock protection, no? > > Because after the check the DIO can still snoop in, right? > > IIRC, we had some discussion around this at [1] last time. > IIUC, you were mentioning to always using unwritten extents > as ->get_block in ext4_page_mkwrite. > And in ext4_writepages(), we replace 'ext4_should_dioread_nolock()' check > with 'is there any DIO in flight'. Yes. > It was discussed to do that check reliably we should have all pages > locked for writeback. But how does that ensure that DIO is not currently in > flight? One such case could be:- > It may still happen that the check to filemap_write_and_wait_range() > from DIO (iomap_dio_rw) got completed and before calling inode_dio_begin() a > context switch happens. > And in parallel we got the page fault which somehow also resulted into > writeback of pages calling ext4_writepages(). Here when we checked for > 'is DIO in progress' after making sure all the writeback pages are locked, > we still say may miss the reliable check if the context switch back to the > DIO process happens right. Am I missing anything? > > 1. Is there any lock guarantee which I am overlooking here? > > 2. Do you think we should use some other lock to provide the guarantee > between page_mkwrite & DIO read? See my previous reply. > 3. What if we always go via unwritten blocks in ext4_writepages too? > hmm, but I am not sure if should really do this today as there are some > known xfstests failures for blocksize < pagesize with dioread_nolock path > right. > Although those problems I have mostly observed with 1K blocksize & 4K > pagesize on x86 platform. Well, the problems with dioread_nolock writeback need to be fixed. That's for sure but orthogonal to this discussion. The reason why we don't want to always use unwritten extents for writeback in ext4_writepages() is because it is slower (we have to convert unwritten extents to written ones after IO completion and that does add up, although it may be worthwhile to get fresh performance numbers). Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
