[PATCH net v1] net: stmmac: TSO: Fix unbalanced DMA map/unmap for non-paged SKB data

[Jon Hunter]

On 04/12/2024 18:18, Thierry Reding wrote:
> On Wed, Dec 04, 2024 at 05:45:43PM +0000, Russell King (Oracle) wrote:
>> On Wed, Dec 04, 2024 at 05:02:19PM +0000, Jon Hunter wrote:
>>> Hi Russell,
>>>
>>> On 04/12/2024 16:39, Russell King (Oracle) wrote:
>>>> On Wed, Dec 04, 2024 at 04:58:34PM +0100, Thierry Reding wrote:
>>>>> This doesn't match the location from earlier, but at least there's
>>>>> something afoot here that needs fixing. I suppose this could simply be
>>>>> hiding any subsequent errors, so once this is fixed we might see other
>>>>> similar issues.
>>>>
>>>> Well, having a quick look at this, the first thing which stands out is:
>>>>
>>>> In stmmac_tx_clean(), we have:
>>>>
>>>>                   if (likely(tx_q->tx_skbuff_dma[entry].buf &&
>>>>                              tx_q->tx_skbuff_dma[entry].buf_type != STMMAC_TXBUF_T
>>>> _XDP_TX)) {
>>>>                           if (tx_q->tx_skbuff_dma[entry].map_as_page)
>>>>                                   dma_unmap_page(priv->device,
>>>>                                                  tx_q->tx_skbuff_dma[entry].buf,
>>>>                                                  tx_q->tx_skbuff_dma[entry].len,
>>>>                                                  DMA_TO_DEVICE);
>>>>                           else
>>>>                                   dma_unmap_single(priv->device,
>>>>                                                    tx_q->tx_skbuff_dma[entry].buf,
>>>>                                                    tx_q->tx_skbuff_dma[entry].len,
>>>>                                                    DMA_TO_DEVICE);
>>>>                           tx_q->tx_skbuff_dma[entry].buf = 0;
>>>>                           tx_q->tx_skbuff_dma[entry].len = 0;
>>>>                           tx_q->tx_skbuff_dma[entry].map_as_page = false;
>>>>                   }
>>>>
>>>> So, tx_skbuff_dma[entry].buf is expected to point appropriately to the
>>>> DMA region.
>>>>
>>>> Now if we look at stmmac_tso_xmit():
>>>>
>>>>           des = dma_map_single(priv->device, skb->data, skb_headlen(skb),
>>>>                                DMA_TO_DEVICE);
>>>>           if (dma_mapping_error(priv->device, des))
>>>>                   goto dma_map_err;
>>>>
>>>>           if (priv->dma_cap.addr64 <= 32) {
>>>> ...
>>>>           } else {
>>>> ...
>>>>                   des += proto_hdr_len;
>>>> ...
>>>> 	}
>>>>
>>>>           tx_q->tx_skbuff_dma[tx_q->cur_tx].buf = des;
>>>>           tx_q->tx_skbuff_dma[tx_q->cur_tx].len = skb_headlen(skb);
>>>>           tx_q->tx_skbuff_dma[tx_q->cur_tx].map_as_page = false;
>>>>           tx_q->tx_skbuff_dma[tx_q->cur_tx].buf_type = STMMAC_TXBUF_T_SKB;
>>>>
>>>> This will result in stmmac_tx_clean() calling dma_unmap_single() using
>>>> "des" and "skb_headlen(skb)" as the buffer start and length.
>>>>
>>>> One of the requirements of the DMA mapping API is that the DMA handle
>>>> returned by the map operation will be passed into the unmap function.
>>>> Not something that was offset. The length will also be the same.
>>>>
>>>> We can clearly see above that there is a case where the DMA handle has
>>>> been offset by proto_hdr_len, and when this is so, the value that is
>>>> passed into the unmap operation no longer matches this requirement.
>>>>
>>>> So, a question to the reporter - what is the value of
>>>> priv->dma_cap.addr64 in your failing case? You should see the value
>>>> in the "Using %d/%d bits DMA host/device width" kernel message.
>>>
>>> It is ...
>>>
>>>   dwc-eth-dwmac 2490000.ethernet: Using 40/40 bits DMA host/device width
>>
>> So yes, "des" is being offset, which will upset the unmap operation.
>> Please try the following patch, thanks:
>>
>> diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
>> index 9b262cdad60b..c81ea8cdfe6e 100644
>> --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
>> +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
>> @@ -4192,8 +4192,8 @@ static netdev_tx_t stmmac_tso_xmit(struct sk_buff *skb, struct net_device *dev)
>>   	struct stmmac_txq_stats *txq_stats;
>>   	struct stmmac_tx_queue *tx_q;
>>   	u32 pay_len, mss, queue;
>> +	dma_addr_t tso_des, des;
>>   	u8 proto_hdr_len, hdr;
>> -	dma_addr_t des;
>>   	bool set_ic;
>>   	int i;
>>   
>> @@ -4289,14 +4289,15 @@ static netdev_tx_t stmmac_tso_xmit(struct sk_buff *skb, struct net_device *dev)
>>   
>>   		/* If needed take extra descriptors to fill the remaining payload */
>>   		tmp_pay_len = pay_len - TSO_MAX_BUFF_SIZE;
>> +		tso_des = des;
>>   	} else {
>>   		stmmac_set_desc_addr(priv, first, des);
>>   		tmp_pay_len = pay_len;
>> -		des += proto_hdr_len;
>> +		tso_des = des + proto_hdr_len;
>>   		pay_len = 0;
>>   	}
>>   
>> -	stmmac_tso_allocator(priv, des, tmp_pay_len, (nfrags == 0), queue);
>> +	stmmac_tso_allocator(priv, tso_des, tmp_pay_len, (nfrags == 0), queue);
>>   
>>   	/* In case two or more DMA transmit descriptors are allocated for this
>>   	 * non-paged SKB data, the DMA buffer address should be saved to
> 
> I see, that makes sense. Looks like this has been broken for a few years
> (since commit 34c15202896d ("net: stmmac: Fix the problem of tso_xmit"))
> and Furong's patch ended up exposing it.
> 
> Anyway, this seems to fix it for me. I can usually trigger the issue
> within one or two iperf runs, with your patch I haven't seen it break
> after a dozen or so runs.
> 
> It may be good to have Jon's test results as well, but looks good so
> far.


I have been running tests on my side and so far so good too. I have not 
seen any more mapping failure cases.

Russell, if you are planning to send a fix for this, please add my ...

Tested-by: Jon Hunter <jonathanh at nvidia.com>

Thanks!
Jon

-- 
nvpublic