Re: [PATCH] Revert "net: pskb_trim_rcsum() and CHECKSUM_COMPLETE are friends"

[Eric Dumazet]

On 06/19/2018 03:32 PM, Andreas Schwab wrote:
> On Jun 19 2018, Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote:
> 
>> diff --git a/drivers/net/ethernet/sun/sungem.c b/drivers/net/ethernet/sun/sungem.c
>> index 7a16d40a72d13cf1d522e8a3a396c826fe76f9b9..672d6748ab44f0890e92d5ca55d6ff6834c20dc9 100644
>> --- a/drivers/net/ethernet/sun/sungem.c
>> +++ b/drivers/net/ethernet/sun/sungem.c
>> @@ -60,8 +60,7 @@
>>  #include <linux/sungem_phy.h>
>>  #include "sungem.h"
>>  
>> -/* Stripping FCS is causing problems, disabled for now */
>> -#undef STRIP_FCS
>> +#define STRIP_FCS
>>  
>>  #define DEFAULT_MSG    (NETIF_MSG_DRV          | \
>>                          NETIF_MSG_PROBE        | \
>> @@ -435,7 +434,7 @@ static int gem_rxmac_reset(struct gem *gp)
>>         writel(desc_dma & 0xffffffff, gp->regs + RXDMA_DBLOW);
>>         writel(RX_RING_SIZE - 4, gp->regs + RXDMA_KICK);
>>         val = (RXDMA_CFG_BASE | (RX_OFFSET << 10) |
>> -              ((14 / 2) << 13) | RXDMA_CFG_FTHRESH_128);
>> +              (ETH_HLEN << 13) | RXDMA_CFG_FTHRESH_128);
>>         writel(val, gp->regs + RXDMA_CFG);
>>         if (readl(gp->regs + GREG_BIFCFG) & GREG_BIFCFG_M66EN)
>>                 writel(((5 & RXDMA_BLANK_IPKTS) |
>> @@ -857,6 +856,14 @@ static int gem_rx(struct gem *gp, int work_to_do)
>>  
>>                 csum = (__force __sum16)htons((status & RXDCTRL_TCPCSUM) ^ 0xffff);
>>                 skb->csum = csum_unfold(csum);
>> +               {
>> +               __wsum rsum = csum_partial(skb->data + ETH_HLEN, len - ETH_HLEN, 0);
>> +               if (csum != csum_fold(rsum) && net_ratelimit())
>> +                       pr_err("sungem wrong csum : %x/%x, len %u bytes\n",
>> +                               csum, csum_fold(rsum), len);
>> +                       print_hex_dump(KERN_ERR, "raw data: ", DUMP_PREFIX_OFFSET,
>> +                                           16, 1, skb->data, len, true);
>> +               }
>>                 skb->ip_summed = CHECKSUM_COMPLETE;
>>                 skb->protocol = eth_type_trans(skb, gp->dev);
>>  
>> @@ -1761,7 +1768,7 @@ static void gem_init_dma(struct gem *gp)
>>         writel(0, gp->regs + TXDMA_KICK);
>>  
>>         val = (RXDMA_CFG_BASE | (RX_OFFSET << 10) |
>> -              ((14 / 2) << 13) | RXDMA_CFG_FTHRESH_128);
>> +              (ETH_HLEN << 13) | RXDMA_CFG_FTHRESH_128);
>>         writel(val, gp->regs + RXDMA_CFG);
>>  
>>         writel(desc_dma >> 32, gp->regs + RXDMA_DBHI);
> 
> With that patch I still get the wrong csum messages, but no longer the
> hw csum failure messages (tested on a PowerMac G5).
> 
> [  662.659767] sungem: sungem wrong csum : 8359/7ca6, len 86 bytes, c0000001fee9cc02
> [  662.659775] raw data: 00000000: 00 0d 93 43 81 62 d4 3d 7e 4c 48 b7 86 dd 61 01  ...C.b.=~LH...a.
> [  662.659778] raw data: 00000010: 1c 1e 00 20 06 40 20 01 0a 62 17 11 88 01 00 00  ... .@ ..b......
> [  662.659780] raw data: 00000020: 00 00 00 00 0a 38 20 01 0a 62 17 11 88 01 00 00  .....8 ..b......
> [  662.659783] raw data: 00000030: 00 00 00 00 00 07 9a 18 00 16 c1 9a 7e ea ea 44  ............~..D
> [  662.659785] raw data: 00000040: fb 4a 80 10 05 93 44 08 00 00 01 01 08 0a 59 68  .J....D.......Yh
> [  662.659788] raw data: 00000050: ba e2 0e bb ac ae                                ......
> 
> Andreas.
> 

Note that 8359 and 7ca6 are the same really (a missing ~ to invert csum_partial())

So the bug was that :

1) Driver programmed a wrong start offset for the csum  (7 bytes instead of 14 bytes to skip Ethernet Header)

2) FCS was not stripped.

Basically CHECKSUM_COMPLETE support never worked, but this was hidden by the fact that linux stack
had to throw away this CHECKSUM_COMPLETE because the FCS had to be removed.

Thanks !