From: Li Zefan Subject: Re: [RFC][PATCH] vfs: always protect diretory file->fpos with inode mutex Date: Wed, 20 Feb 2013 09:49:36 +0800 Message-ID: <51242BB0.3060103@huawei.com> References: <5122D3E0.6070800@huawei.com> <20130219091931.GB21945@quack.suse.cz> <51236652.1050608@huawei.com> <20130219125913.GD21945@quack.suse.cz> Mime-Version: 1.0 Content-Type: text/plain; charset="ISO-8859-1" Content-Transfer-Encoding: 7bit Cc: , LKML , Ext4 Developers List , "Theodore Ts'o" , Andrew Morton , , Wuqixuan , Al Viro , To: Jan Kara Return-path: In-Reply-To: <20130219125913.GD21945@quack.suse.cz> Sender: linux-kernel-owner@vger.kernel.org List-Id: linux-ext4.vger.kernel.org On 2013/2/19 20:59, Jan Kara wrote: > On Tue 19-02-13 19:47:30, Li Zefan wrote: >> On 2013/2/19 17:19, Jan Kara wrote: >>> On Tue 19-02-13 09:22:40, Li Zefan wrote: >>>> There's a long long-standing bug...As long as I don't know when it dates >>>> from. >>>> >>>> I've written and attached a simple program to reproduce this bug, and it can >>>> immediately trigger the bug in my box. It uses two threads, one keeps calling >>>> read(), and the other calling readdir(), both on the same directory fd. >>> So the fact that read() or even write() to fd opened O_RDONLY has *any* >>> effect on f_pos looks really unexpected to me. I think we really should >>> have there: >>> if (ret >= 0) >>> file_pos_write(...); >> >> I thought about this. The problem is then we have to check every fop->write() >> to see if any of them can return -errno with file->f_pos changed and fix them, >> though it's do-able. > But returning error and advancing f_pos would be a bug - specification > says write() returns the number of bytes written or -1 and f_pos should be > advanced by the number of bytes written. > Oh, I had an illusion that vfs saves f_pos and calls write() and restore f_pos if write() fails.