Messages in this thread | | | From | Arnd Bergmann <> | Subject | Re: [PATCH] sony_pi: Remove the BKL from sonypi_misc_open | Date | Wed, 21 Oct 2009 10:29:21 +0200 |
| |
On Wednesday 21 October 2009, John Kacur wrote: > From 11e6a3b690413c3926e6db1c53a53410b5214c3d Mon Sep 17 00:00:00 2001 > From: John Kacur <jkacur@redhat.com> > Date: Wed, 21 Oct 2009 01:51:41 +0200 > Subject: [PATCH] sonypi: Use non-BKL version of llseek. > > The default version of llseek uses the BKL. > We have removed the use of the BKL in open and the ioctl. > Now lets remove the last use of the BKL by explictly calling the > generic unlocked llseek, under the sonypi_device.lock mutex > > Signed-off-by: John Kacur <jkacur@redhat.com>
Well, for this specific case, the driver does not actually need to seek, because the read function never looks at the position and there is no write function. For annotation purposes, IMHO we should have a simple '.llseek = no_llseek' in there.
In other files, the best solution may be to just point to generic_file_llseek and make sure we hold the i_mutex when accessing the file->f_pos explicitly.
Interestingly, atomicity of updates to f_pos is currently maintained through file_pos_write(), which does not hold any lock but assumes that a store to an loff_t is atomic. It is not atomic in general, so concurrent lseek and read/write operations seem to have undefined behaviour no matter what kind of locking we have in the llseek function.
Arnd <><
| |