On Wed, Sep 15, 2010 at 04:57:15PM +0200, Attilio Rao wrote: > 2010/9/15 Kostik Belousov <kostik...@gmail.com>: > > On Wed, Sep 15, 2010 at 02:24:21PM +0000, Attilio Rao wrote: > >> Author: attilio > >> Date: Wed Sep 15 14:24:21 2010 > >> New Revision: 212661 > >> URL: http://svn.freebsd.org/changeset/base/212661 > >> > >> Log: > >> Fix bogus busying mechanism from cdevsw callbacks: > >> - D_TRACKCLOSE may be used there as d_close() are expected to match up > >> d_open() calls > > VFS is not very good at properly calling VOP_CLOSE(). As example, premature > > vnode reclaim due to devfs unmount would cause VOP_CLOSE() to be called > > only once despite the number of opens being > 1. > > Yes. > That makes implementing a similar semantic in drivers very difficult > and not very well fixable, in particular within the d_* callbacks. > I'm seriously wondering if we might just make a shortcut just for > supporting such a feature (busying the device on real devfs entry > activity) as several of them may be needing and may be probabilly > needing to be 100%. > Luckilly, it seems that such paths are not experienced very frequently.
I am not sure what do you mean by "busying the device on real devfs entry activity". Is the operation made by dev_refthread() and friends enough ? Device cannot be destroyed by destroy_dev() until all threads leave the cdevsw methods.
pgpeUG62tg5Fj.pgp
Description: PGP signature
