[ath9k-devel] DMA issues with ar9280 cards

[Bernhard Schmidt <bschmidt@techwires.net>]

On Wednesday, March 02, 2011 15:51:37 Mohammed Shafi wrote:
> On Thu, Feb 24, 2011 at 1:36 AM, Bernhard Schmidt
> <bschmidt@techwires.net> wrote:
> > On Wednesday 23 February 2011 18:46:49 Luis R. Rodriguez wrote:
> >> On Wed, Feb 23, 2011 at 12:55:35AM -0800, Bernhard Schmidt wrote:
> >> > % scan done, setting up the BSS
> >> > [  296.060536] ath: Failed to stop TX DMA in 100 msec after killing last frame
> >> > [  296.075985] ath: Failed to stop TX DMA in 100 msec after killing last frame
> >> > [  296.083032] ath: Failed to stop TX DMA!
> >> > [  296.099187] ath: DMA failed to stop in 10 ms AR_CR=0x00000024 AR_DIAG_SW=0x42000020
> >> > [  296.106884] ath: Could not stop RX, we could be confusing the DMA engine when we start RX up
> >> > [  296.127247] ath: DMA failed to stop in 10 ms AR_CR=0x00000024 AR_DIAG_SW=0x42000020
> >> > [  297.163998] ath: Failed to stop TX DMA in 100 msec after killing last frame
> >> > [  297.182977] ath: DMA failed to stop in 10 ms AR_CR=0x00000024 AR_DIAG_SW=0x42000020
> >> > [  297.190677] ath: Could not stop RX, we could be confusing the DMA engine when we start RX up
> >> > [  297.211004] ath: DMA failed to stop in 10 ms AR_CR=0x00000024 AR_DIAG_SW=0x42000020
> >> > [  298.247659] ath: Failed to stop TX DMA in 100 msec after killing last frame
> >> > % ..
> >>
> >> If its so easy to reproduce we can likely fix this for good!
> >
> > Actually, haven't found a way to *not* reproduce it. :)
> 
> the obvious way of increasing the timeout for stopping the DMA might help ?
> in mac.c
> 
> 148 #define ATH9K_TX_STOP_DMA_TIMEOUT       4000    /* usec */
> increase it to 10000 usec
> 
> and
> 781#define AH_RX_STOP_DMA_TIMEOUT 10000   /* usec */
> increase it to 20,000 usec

Did try to play with various values, no differences whatsoever,
obviously. Even tried polling a whole second. The messages do indicate
that there is an issue. It's not that the reg isn't polled for long
enough, but instead this is an indicator for something wrong going on.
What I'm wondering though is, that the queue resets which eventually
happen do not have any effect, only after a "full" reset (e.g.
ifconfig down/up) the queues will be usable again.

On a site note, after falling back to other chips (9380, 9160), I see
the same issue there too. Not as often, or as immediate as with 9280,
but they do occur (even on different hardware). Having a hostapd
instance idle long enough is able to trigger it eventually.

-- 
Bernhard