Re: memory clobber in rx path, maybe related to ath9k.

["Luis R. Rodriguez" <lrodriguez@atheros.com>]

On Fri, Oct 15, 2010 at 04:33:50PM -0700, Ben Greear wrote:
> On 10/15/2010 04:21 PM, Luis R. Rodriguez wrote:
> > Ben, please give this patch a shot. I addresses three races on the PCU:
> >
> >    * When we were stopping the CPU for non-EDMA cards we never locked against
> >      anything starting the PCU again
> >
> >    * ath9k_hw_startpcureceive() was being called without locking
> >
> >    * Although we lock on the rxbuf lock for contention against starting/stopping
> >      the PCU, we also need to lock on the driver in locations where we start/stop
> >      the PCU within the same location otherwise we end up in inconsistant states
> >      and the hardware may end up proessing an incorrect buffer for DMA. To
> >      protect against this we use a new PCU lock on the main part of the driver to
> >      ensure each start/stop/reset operation is done atomically.
> >
> > And fixes one issue as a side effect:
> >
> >    * No more packet loss on ping flood when you have one STA associated :)
> >
> > The only issue I see with this is I eventually run out of memory and my box
> > becomes useless, unless I am mistaking that for some other issue.
> >
> > Please give this a shot and if it cures your woes I'll split it up into
> > 3 separate patches, or maybe just two, one for the first two and one for
> > the last issue.
> 
> Sounds good, but this lockdep splat happens almost immediately upon starting
> my app:
> 
> =======================================================
> [ INFO: possible circular locking dependency detected ]
> 2.6.36-rc8-wl+ #32
> -------------------------------------------------------
> swapper/0 is trying to acquire lock:
>   (&(&sc->rx.pcu_lock)->rlock){+.-...}, at: [<fa16e5c7>] ath9k_tasklet+0x7e/0x140 [ath9k]
> 
> but task is already holding lock:
>   (&(&sc->rx.rxflushlock)->rlock){+.-...}, at: [<fa16e5b9>] ath9k_tasklet+0x70/0x140 [ath9k]
> 
> which lock already depends on the new lock.
> 
> 
> the existing dependency chain (in reverse order) is:
> 
> -> #1 (&(&sc->rx.rxflushlock)->rlock){+.-...}:
>         [<c0457639>] lock_acquire+0x5a/0x78
>         [<c075f6ed>] _raw_spin_lock_bh+0x20/0x2f
>         [<fa170513>] ath_flushrecv+0x14/0x61 [ath9k]

Ah we just need to nuke the flush lock, one second. Also remove my
skb_copy() otherwise you will really run out of memory quickly,
unless you really want to test with it :)

  Luis