From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S932132AbWJEPgL (ORCPT ); Thu, 5 Oct 2006 11:36:11 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S932133AbWJEPgK (ORCPT ); Thu, 5 Oct 2006 11:36:10 -0400 Received: from mailhub.sw.ru ([195.214.233.200]:31257 "EHLO relay.sw.ru") by vger.kernel.org with ESMTP id S932132AbWJEPgJ (ORCPT ); Thu, 5 Oct 2006 11:36:09 -0400 To: linux-kernel@vger.kernel.org CC: linux-mm@kvack.org Subject: D-cache aliasing issue in __block_prepare_write From: Monakhov Dmitriy User-Agent: Gnus/5.1006 (Gnus v5.10.6) Emacs/21.4 (gnu/linux) Date: Thu, 05 Oct 2006 19:16:46 +0400 Message-ID: <87ejtmn675.fsf@sw.ru> MIME-Version: 1.0 Content-Type: multipart/mixed; boundary="=-=-=" Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org --=-=-= It's seems I've found D-cache aliasing issue in fs/buffer.c 1902 static int __block_prepare_write(struct inode *inode, struct page *page, 1903 unsigned from, unsigned to, get_block_t *get_block) ...... 1951 kaddr = kmap_atomic(page, KM_USER0); 1952 if (block_end > to) 1953 memset(kaddr+to, 0, 1954 block_end-to); 1955 if (block_start < from) 1956 memset(kaddr+block_start, 1957 0, from-block_start); 1958 flush_dcache_page(page); ##### We call flush_dcache_page() due to page was changed ##### and user space mapping potentially exist. 1959 kunmap_atomic(kaddr, KM_USER0); ...... 2008 clear_buffer_new(bh); 2009 kaddr = kmap_atomic(page, KM_USER0); 2010 memset(kaddr+block_start, 0, bh->b_size); 2011 kunmap_atomic(kaddr, KM_USER0); ###### Here we have absolutely identical situation. ###### D-cache have to be flushed here too. ###### It seems it is just forgotten here. 2012 set_buffer_uptodate(bh); 2013 mark_buffer_dirty(bh); 2014 } 2015 next_bh: 2016 block_start = block_end; 2017 bh = bh->b_this_page; 2018 } while (bh != head); 2019 return err; 2020 } nobh_commit_write() has analogical issue 2515 kaddr = kmap_atomic(page, KM_USER0); 2516 memset(kaddr, 0, PAGE_CACHE_SIZE); ###### flush_dcache_page() have to called here ###### It seems it is just forgotten here too. 2517 kunmap_atomic(kaddr, KM_USER0); 2518 SetPageUptodate(page); 2519 set_page_dirty(page); x86 does not have cache aliasing problems, the problem could show up only on marginal archs, ia64 is the most frequently used. Following is the patch against 2.6.18 fix this issue: --=-=-= Content-Disposition: inline; filename=diff-buffer-flush-dcache-page diff --git a/fs/buffer.c b/fs/buffer.c index 71649ef..b2652aa 100644 --- a/fs/buffer.c +++ b/fs/buffer.c @@ -2008,6 +2008,7 @@ static int __block_prepare_write(struct clear_buffer_new(bh); kaddr = kmap_atomic(page, KM_USER0); memset(kaddr+block_start, 0, bh->b_size); + flush_dcache_page(page); kunmap_atomic(kaddr, KM_USER0); set_buffer_uptodate(bh); mark_buffer_dirty(bh); @@ -2514,6 +2515,7 @@ failed: */ kaddr = kmap_atomic(page, KM_USER0); memset(kaddr, 0, PAGE_CACHE_SIZE); + flush_dcache_page(page); kunmap_atomic(kaddr, KM_USER0); SetPageUptodate(page); set_page_dirty(page); --=-=-=--