Date: Mon, 5 May 2008 11:56:14 -0400 From: John Baldwin <jhb@freebsd.org> To: Kostik Belousov <kostikbel@gmail.com> Cc: freebsd-arch@freebsd.org Subject: Re: Per-open file private data for the cdevs Message-ID: <200805051156.14706.jhb@freebsd.org> In-Reply-To: <20080505142051.GS18958@deviant.kiev.zoral.com.ua> References: <20080504171002.GN18958@deviant.kiev.zoral.com.ua> <200805050939.42425.jhb@freebsd.org> <20080505142051.GS18958@deviant.kiev.zoral.com.ua>
next in thread | previous in thread | raw e-mail | index | archive | help
On Monday 05 May 2008 10:20:51 am Kostik Belousov wrote: > On Mon, May 05, 2008 at 09:39:42AM -0400, John Baldwin wrote: > > On Sunday 04 May 2008 01:10:02 pm Kostik Belousov wrote: > > > Since the review for the clone-at-open patch (fdclone) posted some time ago > > > mostly says that it would be better to implement per-file private data > > > instead, I produced the patch along this line, > > > > > > The patch does not change the cdevsw ABI, instead, three new functions > > > nt devfs_get_cdevpriv(void **datap); > > > int devfs_set_cdevpriv(void *priv, cdevpriv_dtr_t dtr); > > > void devfs_clear_cdevpriv(void); > > > are provided for manipulation of the per-file private data. > > > > > > devfs_set_cdevpriv assigns the priv as private data for the file descriptor > > > which is used to initiate currently performed driver operation. dtr > > > is the function that will be called when either the last refernce to > > > the file goes away or devfs_clear_cdevpriv is called. > > > > > > devfs_get_cdevpriv is the obvious accessor. > > > > > > devfs_clear_cdevpriv allows to clear the private data for the still > > > open file. > > > > > > The synchronization of the cdev data and file private data is left > > > to the driver code, I did not found any generic helper mechanism that > > > could be useful there. > > > > > > Patch: > > > http://people.freebsd.org/~kib/misc/fdpriv.1.patch > > > > > > Dumb driver that shows the basic usage of the proposed KPI: > > > http://people.freebsd.org/~kib/misc/fpclone.c > > > > > > Previous version of the patch was tested by Peter Holm. > > > > I like this very much and will use it instead of devfs cloning in ipmi(4) so > > we can use ipmievd when it is committed. IWBN if there was a more automated > > way of handling the unload race, for example if destroy_dev() could somehow > > clear all the per-open fd data. That may not be easily feasible, however. > > (In theory each cdev could have a list of "attached" 'struct file' objects > > that it updates in VOP_CLOSE() and for a destroy dev it detaches all the > > private data after marking the cdev with a bad/null cdevsw, however, that > > would bloat 'struct file' with at least one more pointer (if not two).) > > Probably, I shall clarify that the dtr is called when the last reference > to the struct file going away, unless the priv data is already cleared > by the call to the devfs_clear_cdevpriv. > > The problem with VOP_CLOSE() is that some actions (like forced unmount > or revoke) may cause less calls to the devfs_close then the files > pointing to the cdev. Your proposal basically means that we need, > besides the normal pointers file->vnode->cdev, have the reverse path > cdev->file. I think this is ugly and better be handled by the fdrop(). The disconnect as I see it is that right now destroy_dev() "orphans" devices in that the files still exist but the backing cdev's now have dead operations. However, now you can have a partially orphaned cdev in that not all of the cdev state is torn down in destroy_dev(). In that sense it would be nice to have the behavior described above, and yes you would have to have the cdev's keep track of 'file's as I suggested. Could you handle the revoke(2) and forced umount cases either via VOP_REVOKE() or by adding a VOP_INACTIVE() to devfs? -- John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200805051156.14706.jhb>