From patchwork Tue Jan 24 21:23:27 2017 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Jeff Layton X-Patchwork-Id: 9535909 Return-Path: Received: from mail.wl.linuxfoundation.org (pdx-wl-mail.web.codeaurora.org [172.30.200.125]) by pdx-korg-patchwork.web.codeaurora.org (Postfix) with ESMTP id 137EC601D3 for ; Tue, 24 Jan 2017 21:23:44 +0000 (UTC) Received: from mail.wl.linuxfoundation.org (localhost [127.0.0.1]) by mail.wl.linuxfoundation.org (Postfix) with ESMTP id 0692526E97 for ; Tue, 24 Jan 2017 21:23:44 +0000 (UTC) Received: by mail.wl.linuxfoundation.org (Postfix, from userid 486) id EE3BD27813; Tue, 24 Jan 2017 21:23:43 +0000 (UTC) X-Spam-Checker-Version: SpamAssassin 3.3.1 (2010-03-16) on pdx-wl-mail.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-6.9 required=2.0 tests=BAYES_00,RCVD_IN_DNSWL_HI autolearn=unavailable version=3.3.1 Received: from lists.sourceforge.net (lists.sourceforge.net [216.34.181.88]) (using TLSv1 with cipher DHE-RSA-AES256-SHA (256/256 bits)) (No client certificate requested) by mail.wl.linuxfoundation.org (Postfix) with ESMTPS id 64E2426E97 for ; Tue, 24 Jan 2017 21:23:43 +0000 (UTC) Received: from localhost ([127.0.0.1] helo=sfs-ml-2.v29.ch3.sourceforge.com) by sfs-ml-2.v29.ch3.sourceforge.com with esmtp (Exim 4.76) (envelope-from ) id 1cW8Z2-0001pp-PO; Tue, 24 Jan 2017 21:23:40 +0000 Received: from sog-mx-1.v43.ch3.sourceforge.com ([172.29.43.191] helo=mx.sourceforge.net) by sfs-ml-2.v29.ch3.sourceforge.com with esmtp (Exim 4.76) (envelope-from ) id 1cW8Z1-0001pj-Ir for v9fs-developer@lists.sourceforge.net; Tue, 24 Jan 2017 21:23:39 +0000 Received-SPF: pass (sog-mx-1.v43.ch3.sourceforge.com: domain of redhat.com designates 209.132.183.28 as permitted sender) client-ip=209.132.183.28; envelope-from=jlayton@redhat.com; helo=mx1.redhat.com; Received: from mx1.redhat.com ([209.132.183.28]) by sog-mx-1.v43.ch3.sourceforge.com with esmtps (TLSv1:AES256-SHA:256) (Exim 4.76) id 1cW8Yz-00032K-Jx for v9fs-developer@lists.sourceforge.net; Tue, 24 Jan 2017 21:23:39 +0000 Received: from smtp.corp.redhat.com (int-mx16.intmail.prod.int.phx2.redhat.com [10.5.11.28]) (using TLSv1.2 with cipher AECDH-AES256-SHA (256/256 bits)) (No client certificate requested) by mx1.redhat.com (Postfix) with ESMTPS id 479A480480; Tue, 24 Jan 2017 21:23:32 +0000 (UTC) Received: from tleilax.poochiereds.net (ovpn-116-60.rdu2.redhat.com [10.10.116.60]) by smtp.corp.redhat.com (Postfix) with ESMTP id 91F96BDF59; Tue, 24 Jan 2017 21:23:31 +0000 (UTC) From: Jeff Layton To: viro@zeniv.linux.org.uk Date: Tue, 24 Jan 2017 16:23:27 -0500 Message-Id: <20170124212327.14517-1-jlayton@redhat.com> X-Scanned-By: MIMEDefang 2.74 on 10.5.11.28 X-Greylist: Sender IP whitelisted, not delayed by milter-greylist-4.5.16 (mx1.redhat.com [10.5.110.28]); Tue, 24 Jan 2017 21:23:32 +0000 (UTC) X-Headers-End: 1cW8Yz-00032K-Jx Cc: linux-nfs@vger.kernel.org, linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org, v9fs-developer@lists.sourceforge.net, ceph-devel@vger.kernel.org, lustre-devel@lists.lustre.org Subject: [V9fs-developer] [PATCH] iov_iter: allow iov_iter_get_pages_alloc to allocate more pages per call X-BeenThere: v9fs-developer@lists.sourceforge.net X-Mailman-Version: 2.1.9 Precedence: list List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , MIME-Version: 1.0 Errors-To: v9fs-developer-bounces@lists.sourceforge.net X-Virus-Scanned: ClamAV using ClamSMTP Currently, iov_iter_get_pages_alloc will only ever operate on the first vector that iterate_all_kinds hands back. Many of the callers however would like to have as long a set of pages as possible, to allow for fewer, but larger I/Os. When the previous vector ends on a page boundary and the current one begins on one, we can continue to add more pages. Change the function to first scan the iov_iter to see how long an array of pages we could create from the current position. Then, allocate an array that large (or up to the maxsize), and fill that many pages. This patch would allow us to rip out a chunk of code in ceph that tries to do the same thing, but doesn't handle it right. For NFS, this allows the client to do larger I/Os when userland passes down an array of page-aligned iovecs in an O_DIRECT request. I imagine this will also make splice writes into an O_DIRECT file more efficient. Signed-off-by: Jeff Layton --- lib/iov_iter.c | 140 +++++++++++++++++++++++++++++++++++++++++++++------------ 1 file changed, 111 insertions(+), 29 deletions(-) diff --git a/lib/iov_iter.c b/lib/iov_iter.c index e68604ae3ced..956d17767a3e 100644 --- a/lib/iov_iter.c +++ b/lib/iov_iter.c @@ -883,6 +883,58 @@ unsigned long iov_iter_gap_alignment(const struct iov_iter *i) } EXPORT_SYMBOL(iov_iter_gap_alignment); +/** + * iov_iter_pvec_size - find length of page aligned iovecs in iov_iter + * @i: iov_iter to in which to find the size + * + * Some filesystems can stitch together multiple iovecs into a single + * page vector when both the previous tail and current base are page + * aligned. This function discovers the length that can fit in a single + * pagevec and returns it. + */ +static size_t iov_iter_pvec_size(const struct iov_iter *i) +{ + size_t size = i->count; + size_t pv_size = 0; + bool contig = false, first = true; + + if (!size) + return 0; + + /* Pipes are naturally aligned for this */ + if (unlikely(i->type & ITER_PIPE)) + return size; + + /* + * An iov can be page vectored when the current base and previous + * tail are both page aligned. Note that we don't require that the + * initial base in the first iovec also be page aligned. + */ + iterate_all_kinds(i, size, v, + ({ + if (first || (contig && PAGE_ALIGNED(v.iov_base))) { + pv_size += v.iov_len; + first = false; + contig = PAGE_ALIGNED(v.iov_base + v.iov_len); + }; 0; + }), + ({ + if (first || (contig && v.bv_offset == 0)) { + pv_size += v.bv_len; + first = false; + contig = PAGE_ALIGNED(v.bv_offset + v.bv_len); + } + }), + ({ + if (first || (contig && PAGE_ALIGNED(v.iov_base))) { + pv_size += v.iov_len; + first = false; + contig = PAGE_ALIGNED(v.iov_base + v.iov_len); + } + })) + return pv_size; +} + static inline size_t __pipe_get_pages(struct iov_iter *i, size_t maxsize, struct page **pages, @@ -1006,47 +1058,77 @@ static ssize_t pipe_get_pages_alloc(struct iov_iter *i, } ssize_t iov_iter_get_pages_alloc(struct iov_iter *i, - struct page ***pages, size_t maxsize, - size_t *start) + struct page ***ppages, size_t maxsize, + size_t *pstart) { - struct page **p; - - if (maxsize > i->count) - maxsize = i->count; + struct page **p, **pc; + size_t start = 0; + ssize_t len = 0; + int npages, res = 0; + bool first = true; if (unlikely(i->type & ITER_PIPE)) - return pipe_get_pages_alloc(i, pages, maxsize, start); + return pipe_get_pages_alloc(i, ppages, maxsize, pstart); + + maxsize = min(iov_iter_pvec_size(i), maxsize); + npages = DIV_ROUND_UP(maxsize, PAGE_SIZE); + p = get_pages_array(npages); + if (!p) + return -ENOMEM; + + pc = p; iterate_all_kinds(i, maxsize, v, ({ unsigned long addr = (unsigned long)v.iov_base; - size_t len = v.iov_len + (*start = addr & (PAGE_SIZE - 1)); + size_t slen = v.iov_len; int n; - int res; - addr &= ~(PAGE_SIZE - 1); - n = DIV_ROUND_UP(len, PAGE_SIZE); - p = get_pages_array(n); - if (!p) - return -ENOMEM; - res = get_user_pages_fast(addr, n, (i->type & WRITE) != WRITE, p); - if (unlikely(res < 0)) { - kvfree(p); - return res; + if (first) { + start = addr & (PAGE_SIZE - 1); + slen += start; + first = false; } - *pages = p; - return (res == n ? len : res * PAGE_SIZE) - *start; + + n = DIV_ROUND_UP(slen, PAGE_SIZE); + if ((pc + n) > (p + npages)) { + /* Did something change the iov array?!? */ + res = -EFAULT; + goto out; + } + addr &= ~(PAGE_SIZE - 1); + res = get_user_pages_fast(addr, n, (i->type & WRITE) != WRITE, pc); + if (unlikely(res < 0)) + goto out; + len += (res == n ? slen : res * PAGE_SIZE) - start; + pc += res; 0;}),({ - /* can't be more than PAGE_SIZE */ - *start = v.bv_offset; - *pages = p = get_pages_array(1); - if (!p) - return -ENOMEM; - get_page(*p = v.bv_page); - return v.bv_len; + /* bio_vecs are limited to a single page each */ + if (first) { + start = v.bv_offset; + first = false; + } + get_page(*pc = v.bv_page); + len += v.bv_len; + ++pc; + BUG_ON(pc > p + npages); }),({ - return -EFAULT; + /* FIXME: should we handle this case? */ + res = -EFAULT; + goto out; }) ) - return 0; +out: + if (unlikely(res < 0)) { + struct page **i; + + for (i = p; i < pc; i++) + put_page(*i); + kvfree(p); + return res; + } + + *ppages = p; + *pstart = start; + return len; } EXPORT_SYMBOL(iov_iter_get_pages_alloc);