On Thu, Feb 19, 2015 at 04:42:41PM +0100, Jan Kara wrote: > On Thu 19-02-15 08:55:23, Dave Chinner wrote: > > On Wed, Feb 18, 2015 at 11:40:09AM +0100, Jan Kara wrote: > > > On Tue 17-02-15 08:37:45, Matthew Wilcox wrote: > > > > On Tue, Feb 17, 2015 at 09:52:00AM +0100, Jan Kara wrote: > > > > > > > This got added to fix a problem that Dave Chinner pointed out. We need > > > > > > > the allocated extent to either be zeroed (as ext2 does), or marked as > > > > > > > unwritten (ext4, XFS) so that a racing read/page fault doesn't return > > > > > > > uninitialized data. If it's marked as unwritten, we need to convert it > > > > > > > to a written extent after we've initialised the contents. We use the > > > > > > > b_end_io() callback to do this, and it's called from the DAX code, not in > > > > > > > softirq context. > > > > > > OK, I see. But I didn't find where ->b_end_io gets called from dax code > > > > > > (specifically I don't see it anywhere in dax_do_IO() or dax_io()). Can you > > > > > > point me please? > > > > > > > > For faults, we call it in dax_insert_mapping(), the very last thing > > > > before returning in the fault path. The normal I/O path gets to use > > > > the dio_iodone_t for the same purpose. > > > I see. I didn't think of races with reads (hum, I actually wonder whether > > > we don't have this data exposure problem for ext4 for mmapped write into > > > a hole vs direct read as well). So I guess we do need those unwritten > > > extent dances after all (or we would need to have a page covering hole when > > > writing to it via mmap but I guess unwritten extent dances are somewhat > > > more standard). > > > > Right, that was the reason for doing it that way - it leveraged all > > the existing methods we have for avoiding data exposure races in > > XFS. but it's also not just for races - it's for ensuring that if we > > crash between the allocation and the write to the persistent store > > we don't expose the underlying contents when the system next comes > > up. > Well, ext3/4 handles the crash situation differently - we make sure we > flush data to allocated blocks before committing a transaction that > allocates them. That works perfectly for crashes but doesn't avoid the > race with DIO. I was talking about direct IO, not buffered IO. DAX is modeled on the direct IO stack, not buffered IO. I did go and look at the ext4 IO completion path, and I can see where ext4_end_io_dio() triggers a commit outside of doing unwritten extent conversion. Can you clue me in - IO completion in ext4 is a maze of twisty passages... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
