"David S. Miller" wrote:
>
> David Dillow writes:
> > F.e. anything that calls rtmsg_ifinfo() almost always calls it without
> > the dev_base_lock held. This includes the dev_ioctl(), dev_ifsioc(),
> > dev_change_flags(), dev_open() call chain, among others. And anything
> > that notfies using the netdev_chain notifier list.
>
> Holding the rtnl_sem semaphore makes holding the dev_base_lock()
> unnecessary in most cases.
>
> Actually, let's go over the cases in detail:
>
> net/core/rtnetlink.c:rtnetlink_fill_ifinfo()
>
> Called from:
>
> 1) rtnetlink_dump_ifinfo, holds dev_base_lock
> 2) rtmsg_ifinfo, all callers paths hold rtnl_sem
>
> net/core/dev.c:sprintf_stats()
>
> Caller dev_get_info() holds dev_base_lock
>
> And that's it.
I agree that that looks like all of the cases. My parse tree was larger,
but I also wasn't taking the rntl_{lock,sem} into account. But, I'm
still getting multiple callers into get_stats() at the same time, unless
my detect code is buggy:
static atomic_t mylock = ATOMIC_INIT(1);
int printed = 0;
while(!atomic_dec_and_test(&mylock))
{
atomic_inc(&mylock);
if(!printed) {
printk("%s: reentered get_stats\n", dev->name);
printed = 1;
}
}
... rest of get_stats ...
atomic_inc(&mylock);
Maybe I'm missing something, but there seems to be a problem somewhere.
I'll keep looking.
Thanks,
David Dillow
dillowd@y12.doe.gov
-
: send the line "unsubscribe linux-net" in
the body of a message to majordomo@vger.kernel.org
