Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S934118AbdC3QoJ (ORCPT ); Thu, 30 Mar 2017 12:44:09 -0400 Received: from zeniv.linux.org.uk ([195.92.253.2]:50158 "EHLO ZenIV.linux.org.uk" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932919AbdC3QoG (ORCPT ); Thu, 30 Mar 2017 12:44:06 -0400 Date: Thu, 30 Mar 2017 17:43:42 +0100 From: Al Viro To: Russell King - ARM Linux Cc: linux-arch@vger.kernel.org, linux-kernel@vger.kernel.org, Linus Torvalds , Richard Henderson , Will Deacon , Haavard Skinnemoen , Vineet Gupta , Steven Miao , Jesper Nilsson , Mark Salter , Yoshinori Sato , Richard Kuo , Tony Luck , Geert Uytterhoeven , James Hogan , Michal Simek , David Howells , Ley Foon Tan , Jonas Bonn , Helge Deller , Martin Schwidefsky , Ralf Baechle , Benjamin Herrenschmidt , Chen Liqin , "David S. Miller" , Chris Metcalf , Richard Weinberger , Guan Xuetao , Thomas Gleixner , Chris Zankel Subject: Re: [RFC][CFT][PATCHSET v1] uaccess unification Message-ID: <20170330164342.GR29622@ZenIV.linux.org.uk> References: <20170329055706.GH29622@ZenIV.linux.org.uk> <20170330162241.GG7909@n2100.armlinux.org.uk> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20170330162241.GG7909@n2100.armlinux.org.uk> User-Agent: Mutt/1.7.1 (2016-10-04) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 4728 Lines: 121 On Thu, Mar 30, 2017 at 05:22:41PM +0100, Russell King - ARM Linux wrote: > On Wed, Mar 29, 2017 at 06:57:06AM +0100, Al Viro wrote: > > Comments, review, testing, replacement patches, etc. are very welcome. > > I've given this a spin, and it appears to work (in that the box boots). > > Kernel size wise: > > text data bss dec hex filename > 8020229 3014220 10243276 21277725 144ac1d vmlinux.orig > 8034741 3014388 10243276 21292405 144e575 vmlinux.uaccess > 7976719 3014324 10243276 21234319 144028f vmlinux.noinline > > Performance using hdparm -T (cached reads) to evaluate against a SSD > gives me the following results: > > * original: > Timing cached reads: 580 MB in 2.00 seconds = 289.64 MB/sec > Timing cached reads: 580 MB in 2.00 seconds = 290.06 MB/sec > Timing cached reads: 580 MB in 2.00 seconds = 289.65 MB/sec > Timing cached reads: 582 MB in 2.00 seconds = 290.82 MB/sec > Timing cached reads: 578 MB in 2.00 seconds = 289.07 MB/sec > > Average = 289.85MB/s > > * uaccess: > Timing cached reads: 578 MB in 2.00 seconds = 288.36 MB/sec > Timing cached reads: 534 MB in 2.00 seconds = 266.68 MB/sec > Timing cached reads: 534 MB in 2.00 seconds = 267.07 MB/sec > Timing cached reads: 552 MB in 2.00 seconds = 275.45 MB/sec > Timing cached reads: 532 MB in 2.00 seconds = 266.08 MB/sec > > Average = 272.73 MB/sec > > * noinline: > Timing cached reads: 548 MB in 2.00 seconds = 274.16 MB/sec > Timing cached reads: 574 MB in 2.00 seconds = 287.19 MB/sec > Timing cached reads: 574 MB in 2.00 seconds = 286.47 MB/sec > Timing cached reads: 572 MB in 2.00 seconds = 286.20 MB/sec > Timing cached reads: 578 MB in 2.00 seconds = 288.86 MB/sec > > Average = 284.58 MB/sec > > I've run the test twice, and there's definitely a reproducable drop in > performance for some reason when switching between current and Al's > uaccess patches, which is partly recovered by switching to the out of > line versions. > > The only difference that I can identify that could explain this are > the extra might_fault() checks in Al's version but which are missing > from the ARM version. How would the following affect things? diff --git a/lib/iov_iter.c b/lib/iov_iter.c index e68604ae3ced..d24d338f0682 100644 --- a/lib/iov_iter.c +++ b/lib/iov_iter.c @@ -184,7 +184,7 @@ static size_t copy_page_to_iter_iovec(struct page *page, size_t offset, size_t b kaddr = kmap(page); from = kaddr + offset; - left = __copy_to_user(buf, from, copy); + left = __copy_to_user_inatomic(buf, from, copy); copy -= left; skip += copy; from += copy; @@ -193,7 +193,7 @@ static size_t copy_page_to_iter_iovec(struct page *page, size_t offset, size_t b iov++; buf = iov->iov_base; copy = min(bytes, iov->iov_len); - left = __copy_to_user(buf, from, copy); + left = __copy_to_user_inatomic(buf, from, copy); copy -= left; skip = copy; from += copy; @@ -267,7 +267,7 @@ static size_t copy_page_from_iter_iovec(struct page *page, size_t offset, size_t kaddr = kmap(page); to = kaddr + offset; - left = __copy_from_user(to, buf, copy); + left = __copy_from_user_inatomic(to, buf, copy); copy -= left; skip += copy; to += copy; @@ -276,7 +276,7 @@ static size_t copy_page_from_iter_iovec(struct page *page, size_t offset, size_t iov++; buf = iov->iov_base; copy = min(bytes, iov->iov_len); - left = __copy_from_user(to, buf, copy); + left = __copy_from_user_inatomic(to, buf, copy); copy -= left; skip = copy; to += copy; @@ -541,7 +541,7 @@ size_t copy_to_iter(const void *addr, size_t bytes, struct iov_iter *i) if (unlikely(i->type & ITER_PIPE)) return copy_pipe_to_iter(addr, bytes, i); iterate_and_advance(i, bytes, v, - __copy_to_user(v.iov_base, (from += v.iov_len) - v.iov_len, + __copy_to_user_inatomic(v.iov_base, (from += v.iov_len) - v.iov_len, v.iov_len), memcpy_to_page(v.bv_page, v.bv_offset, (from += v.bv_len) - v.bv_len, v.bv_len), @@ -560,7 +560,7 @@ size_t copy_from_iter(void *addr, size_t bytes, struct iov_iter *i) return 0; } iterate_and_advance(i, bytes, v, - __copy_from_user((to += v.iov_len) - v.iov_len, v.iov_base, + __copy_from_user_inatomic((to += v.iov_len) - v.iov_len, v.iov_base, v.iov_len), memcpy_from_page((to += v.bv_len) - v.bv_len, v.bv_page, v.bv_offset, v.bv_len), @@ -582,7 +582,7 @@ bool copy_from_iter_full(void *addr, size_t bytes, struct iov_iter *i) return false; iterate_all_kinds(i, bytes, v, ({ - if (__copy_from_user((to += v.iov_len) - v.iov_len, + if (__copy_from_user_inatomic((to += v.iov_len) - v.iov_len, v.iov_base, v.iov_len)) return false; 0;}),