Date: Tue, 2 Dec 2008 18:28:36 -0500 From: John Baldwin <jhb@freebsd.org> To: Kostik Belousov <kostikbel@gmail.com> Cc: scottl@freebsd.org, current@freebsd.org Subject: Re: [PATCH] Make udf(4) MPSAFE and use shared lookups Message-ID: <200812021828.36857.jhb@freebsd.org> In-Reply-To: <20081122115028.GB6408@deviant.kiev.zoral.com.ua> References: <200811201627.58289.jhb@freebsd.org> <200811211452.02545.jhb@freebsd.org> <20081122115028.GB6408@deviant.kiev.zoral.com.ua>
next in thread | previous in thread | raw e-mail | index | archive | help
On Saturday 22 November 2008 06:50:28 am Kostik Belousov wrote: > udf_vget() does insmntque() before vnode is fully initialized, allowing > other threads to find the vnode on the mount list. This is typical for > !MPSAFE fs, and it seems corresponding call was not marked XXX for udf. It does the same as ufs. ufs only partially initializes the i-node (as much as both cd9660 and udf do) and then exclusive locks the vnode before insmntque(). They then finish initializing the i-node (bread() the d-node, for example) and finally drop the vnode lock. > udf_lookup for ISDOTDOT case unlocks dvp before vget'ing "..", allowing > the same race on forced unmount as ufs (I will finally commit ufs patch > today). The race happens for !MPSAFE code too, but it is easier to > execute without Giant. Every fs is going to need this workaround it seems. Would be nice if there was an easier way to avoid cut and pasting this code N times. Perhaps we could make lookup() check VI_DOOMED instead? I had changed it do that at one point, but then someone pointed me at the deadfs stuff and said that was sufficient. -- John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200812021828.36857.jhb>