On 7/9/20 7:32 AM, Matthew Wilcox wrote: > On Thu, Jul 09, 2020 at 02:26:11PM +0100, Christoph Hellwig wrote: >> On Thu, Jul 09, 2020 at 12:10:36PM +0100, Matthew Wilcox wrote: >>> On Thu, Jul 09, 2020 at 11:17:05AM +0100, Christoph Hellwig wrote: >>>> I really don't like this series at all. If saves a single pointer >>>> but introduces a complicated machinery that just doesn't follow any >>>> natural flow. And there doesn't seem to be any good reason for it to >>>> start with. >>> >>> Jens doesn't want the kiocb to grow beyond a single cacheline, and we >>> want the ability to set the loff_t in userspace for an appending write, >>> so the plan was to replace the ki_complete member in kiocb with an >>> loff_t __user *ki_posp. >>> >>> I don't think it's worth worrying about growing kiocb, personally, >>> but this seemed like the easiest way to make room for a new pointer. >> >> The user offset pointer has absolutely no business in the the kiocb >> itself - it is a io_uring concept which needs to go into the io_kiocb, >> which has 14 bytes left in the last cache line in my build. It would >> fit in very well there right next to the result and user pointer. > > I agree. Jens doesn't. Stop putting words in my mouth, especially when they are totally untrue. I was opposed to growing struct io_rw in io_uring, which is where the extra append variable belonds, beyond a cacheline. You mentioned you could probably shave some bits out of struct kiocb, which is how this completion handling business came about. If kiocb was shrunk, then io_rw has room for the needed variable. At no point have I said that whatever we need to shove in there for io_uring should be in the kiocb, that would not make any sense. I'm just opposed to growing the per-op data field in io_kiocb beyond a cacheline. And that's especially true for something like append writes, which I don't consider super interesting. -- Jens Axboe
