On Wed 23-01-13 14:11:41, Zheng Liu wrote: > On Wed, Jan 23, 2013 at 12:14:32AM +0100, Jan Kara wrote: > > On Wed 23-01-13 00:00:17, Zheng Liu wrote: > > > On Tue, Jan 22, 2013 at 04:22:43PM +0100, Jan Kara wrote: > > > > On Tue 22-01-13 22:22:21, Zheng Liu wrote: > > > > > On Tue, Jan 22, 2013 at 02:44:00PM +0100, Jan Kara wrote: > > > > > > On Tue 22-01-13 15:11:24, Dmitry Monakhov wrote: > > > > > > > On Fri, 18 Jan 2013 13:00:37 +0100, Jan Kara <jack@xxxxxxx> wrote: > > > > > > > > When using indirect blocks there is no possibility to have any unwritten > > > > > > > > extents. So wait for them in ext4_ind_direct_IO() is just bogus. > > > > > > > But as soon as i remember indirect implementation may also be used by > > > > > > > extents based inodes 3074: ext4_ext_direct_IO > > > > > > > /* Use the old path for reads and writes beyond i_size. */ > > > > > > > if (rw != WRITE || final_size > inode->i_size) > > > > > > > return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs); > > > > > > > > > > > > > > Am I missing ? > > > > > > Ah, that's a catch. Thanks for pointing that out! So my patch is wrong > > > > > > and that code path needs some cleaning and commenting. In particular I'm > > > > > > afraid using dioread_nolock for inodes with indirect map causes data > > > > > > exposure bugs when unlocked DIO read races with DIO write because such > > > > > > inodes don't support uninitialized extents. > > > > > > > > > > Sorry, but I am still confused. dioread_nolock is only for extent-based > > > > > file. So when a file system without extent feature, dioread_nolock > > > > > couldn't be enabled. It seems that we don't need to worry about > > > > > exposing stale data here. > > > > Well, you can have fs with extent feature enabled but still with inodes > > > > using indirect map. But as Dmitry pointed out, ext4_should_dioread_nolock() > > > > handles that correctly. So there's not a bug I was suspecting. > > > > > > Yep, the patch itself is fine. But that would be great if a comment is > > > added here. > > No, the patch is wrong. The code before the patch is correct. We can get > > to that code for extent based inode which has unwritten conversions pending > > and we need to wait for those as otherwise we could return 0s in places > > where we acknowledged successful write just a while ago. Or am I missing > > something? > > Ah, I see. I guess that the problem is that the dio read races with buffered > write. > > dio read buffered write > ->ext4_file_write > ->ext4_da_write_begin > ->ext4_da_write_end > [buffered write has finished, but the data > and metadata has not been flushed] > ->generic_file_aio_read > ->filemap_write_and_wait_range > ->do_writepages > ->ext4_da_writepages > ->filemap_fdatawait_range > ->wait_on_page_writeback > ->ext4_end_bio > ->end_page_writeback > [unwritten extent has not been > converted] > ->ext4_ind_direct_IO > [here we need to flush unwritten io] Yes, exactly. > So it seems that this patch could be applied after reworking unwritten extent > conversion. Yes. When PageWriteback is cleared after extent conversion, this waiting can go away and everything should be fine. > FWIW, after applied this patch, the latency of dio read could be reduced > dramatically. So that would be great if this patch can be applied when it > doesn't break something. Sure, I'll have that in mind. Honza -- Jan Kara <jack@xxxxxxx> SUSE Labs, CR -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
