[PATCH v21 05/20] nvme-tcp: Add DDP offload control path

[Aurelien Aptel]

Max Gurtovoy <mgurtovoy at nvidia.com> writes:
>> @@ -739,6 +937,9 @@ static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
>>   	size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
>>   	int ret;
>>   
>> +	if (test_bit(NVME_TCP_Q_OFF_DDP, &queue->flags))
>> +		nvme_tcp_resync_response(queue, skb, *offset);
>
> lets try to optimize the fast path with:
>
> if (IS_ENABLED(CONFIG_ULP_DDP) && test_bit(NVME_TCP_Q_OFF_DDP, 
> &queue->flags))
>      nvme_tcp_resync_response(queue, skb, *offset);
>

For this one, when ULP_DDP is disabled, I do see 1 extra mov instruction
but no branching... I think it's negligible personally.

$ gdb drivers/nvme/host/nvme-tcp.ko
(gdb) disass /s nvme_tcp_recv_skb
...
1088    static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
1089                    unsigned int *offset, size_t *len)
1090    {
1091            struct nvme_tcp_hdr *hdr;
1092            char *pdu = queue->pdu;
   0x00000000000046a6 <+118>:   mov    %rsi,-0x70(%rbp)

880             return  (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
   0x00000000000046aa <+122>:   test   %ebx,%ebx
   0x00000000000046ac <+124>:   je     0x4975 <nvme_tcp_recv_skb+837>

1093            size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
   0x00000000000046b2 <+130>:   cmp    %r14,%rbx

1100                    &pdu[queue->pdu_offset], rcv_len);
   0x00000000000046b5 <+133>:   movslq 0x19c(%r12),%rdx

1099            ret = skb_copy_bits(skb, *offset,
   0x00000000000046bd <+141>:   mov    -0x58(%rbp),%rdi

1093            size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
   0x00000000000046c1 <+145>:   cmova  %r14,%rbx

./arch/x86/include/asm/bitops.h:
205             return ((1UL << (nr & (BITS_PER_LONG-1)))
   0x00000000000046c5 <+149>:   mov    0x1d8(%r12),%rax

Extra mov of queue->flags offset here ^^^^^^^^

  (gdb) p &((struct nvme_tcp_queue *)0)->flags
  $1 = (unsigned long *) 0x1d8