Re: [PATCH] sky2: receive dma mapping error handling

[Michael Breuer <mbreuer@majjas.com>]

On 1/31/2010 5:18 PM, Jarek Poplawski wrote:
> On Sun, Jan 31, 2010 at 04:58:42PM -0500, Michael Breuer wrote:
>    
>> On 1/31/2010 1:50 PM, Michael Breuer wrote:
>>      
>>> On 1/30/2010 11:55 PM, Michael Breuer wrote:
>>>        
>>>> On 01/30/2010 07:34 PM, Jarek Poplawski wrote:
>>>>          
>>>>> Could you try the patch below to show maybe some other users of
>>>>> dma-debug entries?
>>>>>
>>>>> Jarek P.
>>>>> ---
>>>>>
>>>>>            
>>>> With the default # entries&  dma_debug_driver=sky2:
>>>>
>>>> 6:00 is eth0&  4:00 is eth1.
>>>>
>>>> Jan 30 23:53:14 mail kernel: DMA-API: 0000:06:00.0: entries: 31961
>>>> ...
>>>>
>>>>          
>>> I put a printk as a third else case in sky2_tx_unmap. Looks like
>>> the issue is that a large number (perhaps all) calls to
>>> sky2_tx_unmap have re->flags set to neither TX_MAP_SINGLE or
>>> TX_MAP_PAGE. Thus the elements are never being unmapped.
>>>
>>> I suspect that the system collapses when using DMAR sooner than if
>>> not using DMAR. Probably some hardware limitation on the number of
>>> mapped elements that is less than the software limitation. I don't
>>> see at present how a ring element can ever get to this code
>>> without re->flags being set to one or the other.
>>>
>>>
>>>        
>> Put some more debugging code in... re->flags is always NULL upon
>> entry to sky2_tx_unmap.
>>
>>      
> Yes, good point! Could you try if this patch can fix it. (not compiled)
>
> Thanks,
> Jarek P.
> ---
>
>   drivers/net/sky2.c |   10 +++++++---
>   1 files changed, 7 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/net/sky2.c b/drivers/net/sky2.c
> index d760650..3437917 100644
> --- a/drivers/net/sky2.c
> +++ b/drivers/net/sky2.c
> @@ -1025,9 +1025,10 @@ static void sky2_prefetch_init(struct sky2_hw *hw, u32 qaddr,
>   static inline struct sky2_tx_le *get_tx_le(struct sky2_port *sky2, u16 *slot)
>   {
>   	struct sky2_tx_le *le = sky2->tx_le + *slot;
> -	struct tx_ring_info *re = sky2->tx_ring + *slot;
> +	struct tx_ring_info *re;
>
>   	*slot = RING_NEXT(*slot, sky2->tx_ring_size);
> +	re = sky2->tx_ring + *slot;
>   	re->flags = 0;
>   	re->skb = NULL;
>   	le->ctrl = 0;
> @@ -1036,13 +1037,16 @@ static inline struct sky2_tx_le *get_tx_le(struct sky2_port *sky2, u16 *slot)
>
>   static void tx_init(struct sky2_port *sky2)
>   {
> -	struct sky2_tx_le *le;
> +	struct sky2_tx_le *le = sky2->tx_le;
> +	struct tx_ring_info *re = sky2->tx_ring;
>
>   	sky2->tx_prod = sky2->tx_cons = 0;
>   	sky2->tx_tcpsum = 0;
>   	sky2->tx_last_mss = 0;
>
> -	le = get_tx_le(sky2,&sky2->tx_prod);
> +	re->flags = 0;
> +	re->skb = NULL;
> +	le->ctrl = 0;
>   	le->addr = 0;
>   	le->opcode = OP_ADDR64 | HW_OWNER;
>   	sky2->tx_last_upper = 0;
>    
Ok- solves the dma-debug issue - i.e., elements are now being unmapped.

Will leave up and hit with traffic unless a crash occurs. If I hit 
something unrelated I'll backport to 2.6.32.7 and try that for a while. 
I do think it's plausible that the dma errors after (during) load were 
due to hardware limitations on the number of mapped entries (haven't 
researched what that limit was). I would also assume that the sw map 
would also have failed eventually.

I'd suggest that regardless of whether this patch solves my crash that 
it ought to be backported as it seems unlikely that any machine would be 
able to survive for long without the tx entries being unmapped.