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

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <c42aabf5-45bc-4ce4-acbf-82f544ce8dc9@nvidia.com>
Date: Wed, 20 Dec 2023 13:30:25 +0200
From: Max Gurtovoy <mgurtovoy@...dia.com>
To: Aurelien Aptel <aaptel@...dia.com>, linux-nvme@...ts.infradead.org,
 netdev@...r.kernel.org, sagi@...mberg.me, hch@....de, kbusch@...nel.org,
 axboe@...com, chaitanyak@...dia.com, davem@...emloft.net, kuba@...nel.org
Cc: Boris Pismenny <borisp@...dia.com>, aurelien.aptel@...il.com,
 smalin@...dia.com, malin1024@...il.com, ogerlitz@...dia.com,
 yorayz@...dia.com, galshalom@...dia.com, brauner@...nel.org
Subject: Re: [PATCH v21 05/20] nvme-tcp: Add DDP offload control path



On 18/12/2023 22:00, Aurelien Aptel wrote:
> Max Gurtovoy <mgurtovoy@...dia.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

Ok we can keep it as is.

Sagi,
any comments on the NVMf patches or on others before we send next version ?
we would like to make it merge for the 6_8 window..