On Sun, Aug 27, 2023 at 08:41:22PM +0100, Al Viro wrote: > That part is somewhat fishy - there's a case where you return a positive value > and advance ->ki_pos by more than that amount. I really wonder if all callers > of ->write_iter() are OK with that. Consider e.g. this: This should not exist in the latest version merged by Jens. Can you check if you still see issues in the version in the block tree or linux-next. > Suppose ->write_iter() ends up doing returning a positive value smaller than > the increment of kiocb.ki_pos. What do we get? ret is positive, so > kiocb.ki_pos gets copied into *ppos, which is ksys_write's pos and there > we copy it into file->f_pos. > > Is it really OK to have write() return 4096 and advance the file position > by 16K? AFAICS, userland wouldn't get any indication of something > odd going on - just a short write to a regular file, with followup write > of remaining 12K getting quietly written in the range 16K..28K. > > I don't remember what POSIX says about that, but it would qualify as > nasty surprise for any userland program - sure, one can check fsync() > results before closing the sucker and see if everything looks fine, > but the way it's usually discussed could easily lead to assumption that > (synchronous) O_DIRECT writes would not be affected by anything of that > sort. ki_pos should always be updated by the write return value. Everything else is a bug.
