On 12/23/21 4:39 PM, Linus Torvalds wrote: > On Thu, Dec 23, 2021 at 1:11 PM Jens Axboe <axboe@xxxxxxxxx> wrote: >> >> Single fix for not clearing kiocb->ki_pos back to 0 for a stream, >> destined for stable as well. > > I don't think any of this is right. > > You can't use f_pos without using fdget_pos() to actually get the lock > to it. > > Which you don't do in io_uring. Well, current position doesn't really make much sense to begin with in an async API, but there are various use cases that want to use it for sync IO. We do have the file at that point, but it's most certainly not sychronized across completions (or serialized with other fdget_pos() users). We could hold ->f_pos_lock over the duration of the IO however, that would probably be saner... As completions are always done from the task itself as well, that should work. I'll take a look at that for 5.17. -- Jens Axboe
