On Fri, May 11, 2018 at 09:32:28PM +0100, Al Viro wrote: > On Fri, May 11, 2018 at 01:30:01PM -0700, Omar Sandoval wrote: > > On Fri, May 11, 2018 at 09:05:38PM +0100, Al Viro wrote: > > > On Thu, May 10, 2018 at 11:30:10PM -0700, Omar Sandoval wrote: > > > > do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode, > > > > struct block_device *bdev, struct iov_iter *iter, > > > > get_block_t get_block, dio_iodone_t end_io, > > > > - dio_submit_t submit_io, int flags) > > > > + dio_submit_t submit_io, int flags, void *private) > > > > > > Oh, dear... That's what, 9 arguments? I agree that the hack in question > > > is obscene, but so is this ;-/ > > > > So looking at these one by one, obviously needed: > > > > - iocb > > - inode > > - iter > > > > bdev is almost always inode->i_sb->s_bdev, except for Btrfs :( > > > > These could _maybe_ go in struct kiocb: > > > > - flags could maybe be folded into ki_flags > > - private could maybe go in iocb->private, but I haven't yet read > > through to figure out if we're already using iocb->private for direct > > I/O Modifying kiocb isn't going to pan out, it's constructed way up in the stack so that'd be a mess. > > That leaves the callbacks, get_block, end_io, and submit_io. Perhaps we > > can add those to inode_operations? > > Or, perhaps, btrfs shouldn't be using the common helper? The question > is not where to stash the bits and pieces - it's how unreadable the callers > are and how much boilerplate/hidden information is involved... I need to call through to do_blockdev_direct_IO() eventually, I'm sure no one wants me to reimplement the 200 lines in there :) so I'd be happy to add a separate helper that only Btrfs uses, but if we're going to call do_blockdev_direct_IO() eventually then we still need the 9 arguments in some form. Am I misunderstanding?
