On Mon, 18 Aug 2003 12:00:46 -0700
Jean Tourrilhes <jt@xxxxxxxxxxxxxxxxxx> wrote:
> On Mon, Aug 18, 2003 at 11:39:36AM -0700, Stephen Hemminger wrote:
> > The locking in hashbin_delete is a problem since unregister_netdevice
> > can't be called with locks held in 2.6.0-test3.
> >
> > Replace it with simpler list macros.
> > Insertion/deletion protected with RTNL semaphore, and read
> > uses RCU.
> >
> > Don't have client hardware to test, but load/unload works on SMP.
>
> Why don't you try the much simpler version :
> ----------------------------------------------
> --- irlan_common.h1.c Mon Aug 18 11:57:39 2003
> +++ irlan_common.c Mon Aug 18 11:58:25 2003
> @@ -125,7 +125,7 @@ int __init irlan_init(void)
>
> IRDA_DEBUG(0, "%s()\n", __FUNCTION__ );
> /* Allocate master structure */
> - irlan = hashbin_new(HB_LOCK); /* protect from /proc */
> + irlan = hashbin_new(HB_NOLOCK); /* network layer will protect us */
> if (irlan == NULL) {
> printk(KERN_WARNING "IrLAN: Can't allocate hashbin!\n");
> return -ENOMEM;
> ----------------------------------------------
> We would still need the RCU, but the overall patch would be
> much simpler and safer IMHO. But it's up to you ;-)
>
> Have fun...
>
> Jean
It really is a personal choice; there is no "right and wrong", but
I felt more secure using the same locking and list model as other network
drivers. You probably feel more secure using the hashbin stuff that the
rest of the IRDA stack does.
Also, RCU is well tested and if something comes up, it would get fixed
in list.h. I wouldn't trust hashbin's with RCU without explicit testing
and review by Paul or Dipankar.
|