From: Christoph Hellwig <hch@infradead.org> To: Trond Myklebust <trondmy@primarydata.com> Cc: "linux-nfs@vger.kernel.org" <linux-nfs@vger.kernel.org>, linux-fsdevel@vger.kernel.org, linux-mm@kvack.org Subject: Re: [PATCH v2 10/12] NFS: Do not serialise O_DIRECT reads and writes Date: Thu, 23 Jun 2016 04:00:11 -0700 [thread overview] Message-ID: <20160623110011.GA6247@infradead.org> (raw) In-Reply-To: <97494C37-23D3-44FA-A9B8-1887E17429D9@primarydata.com> On Wed, Jun 22, 2016 at 05:24:50PM +0000, Trond Myklebust wrote: > If we?re going to worry about write atomicity in the buffered I/O case, > then we really should also make sure that O_DIRECT writes are atomic > w.r.t. page cache updates too. With this locking model, a buffered > read() can race with the O_DIRECT write() and get a mixture of old > data and new. The difference between buffered I/O and direct I/O is that the former is covered by standards, and the latter is a Linux extension with very lose semantics. But I'm perfectly fine with removing the buffered reader shared lock for now - for the purposes of direct I/O synchronization it's not nessecary. Yes. > > + if (mapping->nrpages) { > > + inode_lock(inode); > > This is unnecessary now that we have a rw_semaphore. You don?t need to > take an exclusive lock in order to serialise w.r.t. new writes, and by > doing so you end up serialising all reads if there happens to be pages > in the page cache. This is true whether or not those pages are dirty. Traditionally we needed the exclusive lock around invalidate_inode_pages2 and unmap_mapping_range, and from a quick look that's what the existing callers all have. I don't actually see that requirement documented anywhere, though. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@kvack.org. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>
WARNING: multiple messages have this Message-ID (diff)
From: Christoph Hellwig <hch@infradead.org> To: Trond Myklebust <trondmy@primarydata.com> Cc: "linux-nfs@vger.kernel.org" <linux-nfs@vger.kernel.org>, linux-fsdevel@vger.kernel.org, linux-mm@kvack.org Subject: Re: [PATCH v2 10/12] NFS: Do not serialise O_DIRECT reads and writes Date: Thu, 23 Jun 2016 04:00:11 -0700 [thread overview] Message-ID: <20160623110011.GA6247@infradead.org> (raw) In-Reply-To: <97494C37-23D3-44FA-A9B8-1887E17429D9@primarydata.com> On Wed, Jun 22, 2016 at 05:24:50PM +0000, Trond Myklebust wrote: > If we?re going to worry about write atomicity in the buffered I/O case, > then we really should also make sure that O_DIRECT writes are atomic > w.r.t. page cache updates too. With this locking model, a buffered > read() can race with the O_DIRECT write() and get a mixture of old > data and new. The difference between buffered I/O and direct I/O is that the former is covered by standards, and the latter is a Linux extension with very lose semantics. But I'm perfectly fine with removing the buffered reader shared lock for now - for the purposes of direct I/O synchronization it's not nessecary. Yes. > > + if (mapping->nrpages) { > > + inode_lock(inode); > > This is unnecessary now that we have a rw_semaphore. You don?t need to > take an exclusive lock in order to serialise w.r.t. new writes, and by > doing so you end up serialising all reads if there happens to be pages > in the page cache. This is true whether or not those pages are dirty. Traditionally we needed the exclusive lock around invalidate_inode_pages2 and unmap_mapping_range, and from a quick look that's what the existing callers all have. I don't actually see that requirement documented anywhere, though.
next prev parent reply other threads:[~2016-06-23 11:00 UTC|newest] Thread overview: 29+ messages / expand[flat|nested] mbox.gz Atom feed top 2016-06-21 21:34 [PATCH v2 01/12] NFS: Don't flush caches for a getattr that races with writeback Trond Myklebust 2016-06-21 21:34 ` [PATCH v2 02/12] NFS: Cache access checks more aggressively Trond Myklebust 2016-06-21 21:34 ` [PATCH v2 03/12] NFS: Cache aggressively when file is open for writing Trond Myklebust 2016-06-21 21:34 ` [PATCH v2 04/12] NFS: Kill NFS_INO_NFS_INO_FLUSHING: it is a performance killer Trond Myklebust 2016-06-21 21:34 ` [PATCH v2 05/12] NFS: writepage of a single page should not be synchronous Trond Myklebust 2016-06-21 21:34 ` [PATCH v2 06/12] NFS: Don't hold the inode lock across fsync() Trond Myklebust 2016-06-21 21:34 ` [PATCH v2 07/12] NFS: Don't call COMMIT in ->releasepage() Trond Myklebust 2016-06-21 21:34 ` [PATCH v2 08/12] NFS: Fix O_DIRECT verifier problems Trond Myklebust 2016-06-21 21:34 ` [PATCH v2 09/12] NFS: Ensure we reset the write verifier 'committed' value on resend Trond Myklebust 2016-06-21 21:34 ` [PATCH v2 10/12] NFS: Do not serialise O_DIRECT reads and writes Trond Myklebust 2016-06-21 21:34 ` [PATCH v2 11/12] NFS: Remove inode->i_dio_count from the NFS O_DIRECT code Trond Myklebust 2016-06-21 21:34 ` [PATCH v2 12/12] NFS: Clean up nfs_direct_complete() Trond Myklebust 2016-06-22 16:43 ` Christoph Hellwig 2016-06-22 16:42 ` [PATCH v2 11/12] NFS: Remove inode->i_dio_count from the NFS O_DIRECT code Christoph Hellwig 2016-06-22 16:58 ` Trond Myklebust 2016-06-23 10:19 ` Christoph Hellwig 2016-06-22 17:58 ` Anna Schumaker 2016-06-22 18:06 ` Trond Myklebust 2016-06-22 18:08 ` Anna Schumaker 2016-06-22 18:51 ` Anna Schumaker 2016-06-22 19:42 ` Trond Myklebust 2016-06-22 16:47 ` [PATCH v2 10/12] NFS: Do not serialise O_DIRECT reads and writes Christoph Hellwig 2016-06-22 17:24 ` Trond Myklebust 2016-06-23 11:00 ` Christoph Hellwig [this message] 2016-06-23 11:00 ` Christoph Hellwig 2016-06-21 22:25 ` [PATCH v2 03/12] NFS: Cache aggressively when file is open for writing Oleg Drokin 2016-06-22 13:06 ` Trond Myklebust 2016-06-22 16:19 ` Oleg Drokin 2016-06-22 15:48 ` [PATCH v2 01/12] NFS: Don't flush caches for a getattr that races with writeback Christoph Hellwig
Reply instructions: You may reply publicly to this message via plain-text email using any one of the following methods: * Save the following mbox file, import it into your mail client, and reply-to-all from there: mbox Avoid top-posting and favor interleaved quoting: https://en.wikipedia.org/wiki/Posting_style#Interleaved_style * Reply using the --to, --cc, and --in-reply-to switches of git-send-email(1): git send-email \ --in-reply-to=20160623110011.GA6247@infradead.org \ --to=hch@infradead.org \ --cc=linux-fsdevel@vger.kernel.org \ --cc=linux-mm@kvack.org \ --cc=linux-nfs@vger.kernel.org \ --cc=trondmy@primarydata.com \ /path/to/YOUR_REPLY https://kernel.org/pub/software/scm/git/docs/git-send-email.html * If your mail client supports setting the In-Reply-To header via mailto: links, try the mailto: linkBe sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes, see mirroring instructions on how to clone and mirror all data and code used by this external index.