From mboxrd@z Thu Jan 1 00:00:00 1970 From: Miao Xie Subject: Re: [RFC PATCH 1/2] vfs: add an interface to do some preparation for umount Date: Thu, 22 Mar 2012 13:07:19 +0800 Message-ID: <4F6AB387.5010800@cn.fujitsu.com> References: <4F6A98CB.1070308@cn.fujitsu.com> <20120322034109.GU6589@ZenIV.linux.org.uk> Reply-To: miaox@cn.fujitsu.com Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Cc: Chris Mason , Linux Fsdevel , Linux Btrfs To: Al Viro Return-path: In-Reply-To: <20120322034109.GU6589@ZenIV.linux.org.uk> List-ID: On Thu, 22 Mar 2012 03:41:09 +0000, Al Viro wrote: > On Thu, Mar 22, 2012 at 11:13:15AM +0800, Miao Xie wrote: >> Some file systems can do some works in the background by kthreads, we'd >> better stop those works before we umount the file system, or it is easy to >> introduce some problems. So we add an interface that is used to do some >> preparation for umount. > > NAK. First of all, fs might be in active use _after_ umount. man 8 umount, > see umount -l there. Moreover, the same superblock may very well be mounted > more than once, so umount of an individual mountpoint would better not do > anything nasty to users of other ones. > > IOW, this is completely misguided - you are dealing with whatever problem it > is at least two layers above the right one. > > 1) vfsmount may be detached from mount trees but remain in active use. > 2) there may be many vfsmounts over given struct super_block. > > Doing things earlier than the final mntput() runs afoul of (1); doing > them before the final deactivate_locked_super() runs afoul of (2). > > What the hell is that thread doing that needs ->s_umount for serialization > and why is it doing that? AFAIK, acquiring ->s_umount is to avoid the filesystem going from r/w to r/o. This is patch is slapdash, I will find another way to deal with the bug we found. Thanks for your review. Miao