On Tue, Feb 11, 2020 at 08:38:31AM -0800, Ira Weiny wrote: > On Tue, Feb 11, 2020 at 04:34:01PM +1100, Dave Chinner wrote: > > On Sat, Feb 08, 2020 at 11:34:38AM -0800, ira.weiny@xxxxxxxxx wrote: > > > From: Ira Weiny <ira.weiny@xxxxxxxxx> > > > > > > The IS_DAX() check in io_is_direct() causes a race between changing the > > > DAX state and creating the iocb flags. > > > > > > Remove the check because DAX now emulates the page cache API and > > > therefore it does not matter if the file state is DAX or not when the > > > iocb flags are created. > > > > This statement is ... weird. > > > > DAX doesn't "emulate" the page cache API at all > > ah... yea emulate is a bad word here. > > > - it has it's own > > read/write methods that filesystems call based on the iomap > > infrastructure (dax_iomap_rw()). i.e. there are 3 different IO paths > > through the filesystems: the DAX IO path, the direct IO path, and > > the buffered IO path. > > > > Indeed, it seems like this works a bit by luck: Ext4 and XFS always > > check IS_DAX(inode) in the read/write_iter methods before checking > > for IOCB_DIRECT, and hence the IOCB_DIRECT flag is ignored by the > > filesystems. i.e. when we got rid of the O_DIRECT paths from DAX, we > > forgot to clean up io_is_direct() and it's only due to the ordering > > of checks that we went down the DAX path correctly.... > > > > That said, the code change is good, but the commit message needs a > > rewrite. > > How about? > > <commit msg> > fs: Remove unneeded IS_DAX() check > > The IS_DAX() check in io_is_direct() causes a race between changing the > DAX state and creating the iocb flags. This is irrelevant - the check is simply wrong and has been since ~2016 when we moved DAX to use the iomap infrastructure... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx
