On Thu, Sep 09, 2021 at 09:22:30PM -0600, Jens Axboe wrote: > On 9/9/21 9:11 PM, Al Viro wrote: > > On Thu, Sep 09, 2021 at 09:05:13PM -0600, Jens Axboe wrote: > >> On 9/9/21 8:57 PM, Al Viro wrote: > >>> On Thu, Sep 09, 2021 at 03:19:56PM -0600, Jens Axboe wrote: > >>> > >>>> Not sure how we'd do that, outside of stupid tricks like copy the > >>>> iov_iter before we pass it down. But that's obviously not going to be > >>>> very efficient. Hence we're left with having some way to reset/reexpand, > >>>> even in the presence of someone having done truncate on it. > >>> > >>> "Obviously" why, exactly? It's not that large a structure; it's not > >>> the optimal variant, but I'd like to see profiling data before assuming > >>> that it'll cause noticable slowdowns. > >> > >> It's 48 bytes, and we have to do it upfront. That means we'd be doing it > >> for _all_ requests, not just when we need to retry. As an example, current > >> benchmarks are at ~4M read requests per core. That'd add ~200MB/sec of > >> memory traffic just doing this copy. > > > > Umm... How much of that will be handled by cache? > > Depends? And what if the iovec itself has been modified in the middle? > We'd need to copy that whole thing too. It's just not workable as a > solution. Huh? Why the hell would we need to copy iovecs themselves? They are never modified by ->read_iter()/->write_iter(). That's the whole fucking point of iov_iter - the iovec itself is made constant, with all movable parts taken to iov_iter. Again, we should never, ever modify the iovec (or bvec, etc.) array in ->read_iter()/->write_iter()/->sendmsg()/etc. instances. If you see such behaviour anywhere, report it immediately. Any such is a blatant bug.
