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

[Eric Dumazet]

On 06/16/2018 12:14 AM, Mathieu Malaterre wrote:
> Hi Eric,
> 
> On Fri, Jun 15, 2018 at 9:14 PM Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote:
>>
>>
>>
>> On 06/15/2018 11:56 AM, Mathieu Malaterre wrote:
>>> This reverts commit 88078d98d1bb085d72af8437707279e203524fa5.
>>>
>>> It causes regressions for people using chips driven by the sungem
>>> driver. Suspicion is that the skb->csum value isn't being adjusted
>>> properly.
>>>
>>> Symptoms as seen on G4+sungem are:
>>>
>>> [   34.023281] eth0: hw csum failure
>>> [   34.023438] CPU: 0 PID: 0 Comm: swapper Not tainted 4.17.0+ #2
>>> [   34.023618] Call Trace:
>>> [   34.023707] [dffedbd0] [c069ddac] __skb_checksum_complete+0xf0/0x108 (unreliable)
>>> [   34.023948] [dffedbf0] [c0777a70] tcp_v4_rcv+0x604/0xe00
>>> [   34.024118] [dffedc70] [c0731624] ip_local_deliver_finish+0xa8/0x3c4
>>> [   34.024315] [dffedcb0] [c0732430] ip_local_deliver+0xf0/0x154
>>> [   34.024493] [dffedcf0] [c07328dc] ip_rcv+0x448/0x774
>>> [   34.024653] [dffedd50] [c06aeae0] __netif_receive_skb_core+0x5e8/0x1184
>>> [   34.024857] [dffedde0] [c06bba20] napi_gro_receive+0x160/0x22c
>>> [   34.025044] [dffede10] [e14b2590] gem_poll+0x7fc/0x1ac0 [sungem]
>>> [   34.025228] [dffedee0] [c06bacf0] net_rx_action+0x34c/0x618
>>> [   34.025402] [dffedf60] [c07fd27c] __do_softirq+0x16c/0x5f0
>>> [   34.025575] [dffedfd0] [c0064c7c] irq_exit+0x110/0x1a8
>>> [   34.025738] [dffedff0] [c0016170] call_do_irq+0x24/0x3c
>>> [   34.025903] [c0cf7e80] [c0009a84] do_IRQ+0x98/0x1a0
>>> [   34.026055] [c0cf7eb0] [c001b474] ret_from_except+0x0/0x14
>>> [   34.026225] --- interrupt: 501 at arch_cpu_idle+0x30/0x78
>>>                    LR = arch_cpu_idle+0x30/0x78
>>> [   34.026510] [c0cf7f70] [c0cf6000] 0xc0cf6000 (unreliable)
>>> [   34.026682] [c0cf7f80] [c00a3868] do_idle+0xc4/0x158
>>> [   34.026835] [c0cf7fb0] [c00a3ab0] cpu_startup_entry+0x20/0x28
>>> [   34.027013] [c0cf7fc0] [c0998820] start_kernel+0x47c/0x490
>>> [   34.027181] [c0cf7ff0] [00003444] 0x3444
>>>
>>> See commit 7ce5a27f2ef8 ("Revert "net: Handle CHECKSUM_COMPLETE more
>>> adequately in pskb_trim_rcsum()."") for previous reference.
>>
>> This fix seems to hide a bug in csum functions on this architecture.
> 
> That's odd since it seems to only affect g4+sungem user. None of the
> ppc64 seems to be having it. And some ppc32 users are not even seeing
> it.
> 
>> Or a bug on this NIC when receiving a small packet (less than 60 bytes).
>> Maybe the padding bytes are not included in NIC provided csum, and not 0.
> 
> Ok in that case the bug is located in
> ./drivers/net/ethernet/sun/sungem.c that seems more likely. I'll try
> to understand that code, then.


I would try something like :

Basically do not bother using CHECKSUM_COMPLETE for small frames that might have been padded.

Since we need to bring one cache line in eth_type_trans() and further header processing,
computing the checksum in software will be almost free anyway.

diff --git a/drivers/net/ethernet/sun/sungem.c b/drivers/net/ethernet/sun/sungem.c
index 7a16d40a72d13cf1d522e8a3a396c826fe76f9b9..071039f211a8a33153e888bd4014314ba5e686a4 100644
--- a/drivers/net/ethernet/sun/sungem.c
+++ b/drivers/net/ethernet/sun/sungem.c
@@ -855,9 +855,11 @@ static int gem_rx(struct gem *gp, int work_to_do)
                        skb = copy_skb;
                }
 
-               csum = (__force __sum16)htons((status & RXDCTRL_TCPCSUM) ^ 0xffff);
-               skb->csum = csum_unfold(csum);
-               skb->ip_summed = CHECKSUM_COMPLETE;
+               if (len > ETH_ZLEN) {
+                       csum = (__force __sum16)htons((status & RXDCTRL_TCPCSUM) ^ 0xffff);
+                       skb->csum = csum_unfold(csum);
+                       skb->ip_summed = CHECKSUM_COMPLETE;
+               }
                skb->protocol = eth_type_trans(skb, gp->dev);
 
                napi_gro_receive(&gp->napi, skb);