Re: Why lock_kernel() in drivers/pci/proc.c (2.6.0-test9)??

[Andrew Morton <akpm@osdl.org>]

Peter Chubb <peterc@gelato.unsw.edu.au> wrote:
>
> 
> Hi,
> 	What is the BKL protecting in
> drivers/pci/proc.c:proc_bus_pci_lseek() ?

nothing much.

> It looks useless to me, as file->f_pos is changed outside the lock
> anyway, and all the other variables inside the locked region are
> effectively constants for the purpose of this code.
> 
> So unless something subtle's going on, I suggest this:
> 

We normally use i_sem to prevent parallel lseeks against the same fd from
scribbling on f_pos.  Which seems fairly pointless since parallel reads
aren't using i_sem.

But this would be a more conventional debklification.  Does it work OK?


 drivers/pci/proc.c |   10 ++++++----
 1 files changed, 6 insertions(+), 4 deletions(-)

diff -puN drivers/pci/proc.c~proc_bus_pci_lseek-remove-lock_kernel drivers/pci/proc.c
--- 25/drivers/pci/proc.c~proc_bus_pci_lseek-remove-lock_kernel	2003-11-18 18:47:20.000000000 -0800
+++ 25-akpm/drivers/pci/proc.c	2003-11-18 18:48:16.000000000 -0800
@@ -25,7 +25,7 @@ proc_bus_pci_lseek(struct file *file, lo
 {
 	loff_t new = -1;
 
-	lock_kernel();
+	down(&file->f_dentry->d_inode->i_sem);
 	switch (whence) {
 	case 0:
 		new = off;
@@ -37,10 +37,12 @@ proc_bus_pci_lseek(struct file *file, lo
 		new = PCI_CFG_SPACE_SIZE + off;
 		break;
 	}
-	unlock_kernel();
 	if (new < 0 || new > PCI_CFG_SPACE_SIZE)
-		return -EINVAL;
-	return (file->f_pos = new);
+		new = -EINVAL;
+	else
+		file->f_pos = new;
+	up(&file->f_dentry->d_inode->i_sem);
+	return new;
 }
 
 static ssize_t

_