On Mon, Mar 27, 2023 at 12:01:08PM -0600, Jens Axboe wrote: > On 3/24/23 10:46 PM, Al Viro wrote: > > On Fri, Mar 24, 2023 at 02:44:41PM -0600, Jens Axboe wrote: > >> Hi, > >> > >> We've been doing a few conversions of ITER_IOVEC to ITER_UBUF in select > >> spots, as the latter is cheaper to iterate and hence saves some cycles. > >> I recently experimented [1] with io_uring converting single segment READV > >> and WRITEV into non-vectored variants, as we can save some cycles through > >> that as well. > >> > >> But there's really no reason why we can't just do this further down, > >> enabling it for everyone. It's quite common to use vectored reads or > >> writes even with a single segment, unfortunately, even for cases where > >> there's no specific reason to do so. From a bit of non-scientific > >> testing on a vm on my laptop, I see about 60% of the import_iovec() > >> calls being for a single segment. > >> > >> I initially was worried that we'd have callers assuming an ITER_IOVEC > >> iter after a call import_iovec() or import_single_range(), but an audit > >> of the kernel code actually looks sane in that regard. Of the ones that > >> do call it, I ran the ltp test cases and they all still pass. > > > > Which tree was that audit on? Mainline? Some branch in block.git? > > It was just master in -git. But looks like I did miss two spots, I've > updated the series here and will send out a v2: > > https://git.kernel.dk/cgit/linux-block/log/?h=iter-ubuf Just to make sure - head's at 4d0ba2f0250d? One obvious comment (just about the problems you've dealt with in that branch; I'll go over that tree and look for other sources of trouble, will post tonight): all 3 callers of iov_iter_iovec() in there are accompanied by the identical chunks that deal with ITER_UBUF case; it would make more sense to teach iov_iter_iovec() to handle that. loop_rw_iter() would turn into if (!iov_iter_is_bvec(iter)) { iovec = iov_iter_iovec(iter); } else { iovec.iov_base = u64_to_user_ptr(rw->addr); iovec.iov_len = rw->len; } and process_madvise() and do_loop_readv_writev() patches simply go away. Again, I'm _not_ saying there's no other problems left, just that these are better dealt with that way. Something like static inline struct iovec iov_iter_iovec(const struct iov_iter *iter) { if (WARN_ON(!iter->user_backed)) return (struct iovec) { .iov_base = NULL, .iov_len = 0 }; else if (iov_iter_is_ubuf(iter)) return (struct iovec) { .iov_base = iter->ubuf + iter->iov_offset, .iov_len = iter->count }; else return (struct iovec) { .iov_base = iter->iov->iov_base + iter->iov_offset, .iov_len = min(iter->count, iter->iov->iov_len - iter->iov_offset), }; } and no need to duplicate that logics in all callers. Or get rid of those elses, seeing that each alternative is a plain return - matter of taste...
