On 9/10/21 8:42 AM, Al Viro wrote: > On Fri, Sep 10, 2021 at 07:57:49AM -0600, Jens Axboe wrote: > >> It was just a quick hack, might very well be too eager to go through >> those motions. But pondering this instead of sleeping, we don't need to >> copy all of iov_iter in order to restore the state, and we can use the >> same advance after restoring. So something like this may be more >> palatable. Caveat - again untested, and I haven't tested the performance >> impact of this at all. > > You actually can cut it down even more - nr_segs + iov remains constant > all along, so you could get away with just 3 words here... I would be Mmm, the iov pointer remains constant? Maybe I'm missing your point, but the various advance functions are quite happy to increment iter->iov or iter->bvec, so we need to restore them. From a quick look, looks like iter->nr_segs is modified for advancing too. What am I missing? > surprised if extra memory traffic had shown up - it's well within the > noise from register spills, (un)inlining, etc. We are talking about > 3 (or 4, with your variant) extra words on one stack frame (and that'd > be further offset by removal of ->truncated); I'd still like to see the > profiling data, but concerns about extra memory traffic due to that > are, IMO, misplaced. See other email that was just sent out, it is measurable but pretty minimal. But that's also down to about 1/3rd of copying the whole thing blindly, so definitely a better case. -- Jens Axboe
