[PATCH 10/18] nvme-tcp: fixup send workflow for kTLS

[Sagi Grimberg]

On 3/22/23 12:08, Hannes Reinecke wrote:
> On 3/22/23 10:31, Sagi Grimberg wrote:
>>
>>
>> On 3/21/23 14:43, Hannes Reinecke wrote:
>>> kTLS does not support MSG_EOR flag for sendmsg(), and the ->sendpage()
>>> call really doesn't bring any benefit as data has to be copied
>>> anyway.
>>> So use sock_no_sendpage() or sendmsg() instead, and ensure that the
>>> MSG_EOR flag is blanked out for kTLS.
>>>
>>> Signed-off-by: Hannes Reinecke <hare at suse.de>
>>> ---
>>>   drivers/nvme/host/tcp.c | 33 +++++++++++++++++++++------------
>>>   1 file changed, 21 insertions(+), 12 deletions(-)
>>>
>>> diff --git a/drivers/nvme/host/tcp.c b/drivers/nvme/host/tcp.c
>>> index bbff1f52a167..007d457cacf9 100644
>>> --- a/drivers/nvme/host/tcp.c
>>> +++ b/drivers/nvme/host/tcp.c
>>> @@ -1034,13 +1034,19 @@ static int nvme_tcp_try_send_data(struct 
>>> nvme_tcp_request *req)
>>>           bool last = nvme_tcp_pdu_last_send(req, len);
>>>           int req_data_sent = req->data_sent;
>>>           int ret, flags = MSG_DONTWAIT;
>>> +        bool do_sendpage = sendpage_ok(page);
>>> -        if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
>>> +        if (!last || queue->data_digest || nvme_tcp_queue_more(queue))
>>> +            flags |= MSG_MORE;
>>> +        else if (!test_bit(NVME_TCP_Q_TLS, &queue->flags))
>>>               flags |= MSG_EOR;
>>> -        else
>>> -            flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
>>
>> I think its time to move the flags setting to a helper.
>>
>>> -        if (sendpage_ok(page)) {
>>> +        if (test_bit(NVME_TCP_Q_TLS, &queue->flags))
>>> +            do_sendpage = false;
>>> +
>>> +        if (do_sendpage) {
>>
>> The do_sendpage looks redundant to me.
>>
>>> +            if (flags & MSG_MORE)
>>> +                flags |= MSG_SENDPAGE_NOTLAST;
>>>               ret = kernel_sendpage(queue->sock, page, offset, len,
>>>                       flags);
>>
>> I think that the SENDPAGE_NOLAST should be set together with MSG_MORE
>> regardless.
>>
>>>           } else {
>>> @@ -1088,19 +1094,22 @@ static int nvme_tcp_try_send_cmd_pdu(struct 
>>> nvme_tcp_request *req)
>>>       bool inline_data = nvme_tcp_has_inline_data(req);
>>>       u8 hdgst = nvme_tcp_hdgst_len(queue);
>>>       int len = sizeof(*pdu) + hdgst - req->offset;
>>> -    int flags = MSG_DONTWAIT;
>>> +    struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
>>> +    struct kvec iov = {
>>> +        .iov_base = (u8 *)req->pdu + req->offset,
>>> +        .iov_len = len,
>>> +    };
>>>       int ret;
>>>       if (inline_data || nvme_tcp_queue_more(queue))
>>> -        flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
>>> -    else
>>> -        flags |= MSG_EOR;
>>> +        msg.msg_flags |= MSG_MORE;
>>> +    else if (!test_bit(NVME_TCP_Q_TLS, &queue->flags))
>>> +        msg.msg_flags |= MSG_EOR;
>>>       if (queue->hdr_digest && !req->offset)
>>>           nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
>>> -    ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
>>> -            offset_in_page(pdu) + req->offset, len,  flags);
>>> +    ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
>>
>> I'd prefer to do kernel_sednpage/sock_no_sendpage similar to how we do
>> it for data and data pdu.
>>
>>>       if (unlikely(ret <= 0))
>>>           return ret;
>>> @@ -1131,7 +1140,7 @@ static int nvme_tcp_try_send_data_pdu(struct 
>>> nvme_tcp_request *req)
>>>       if (queue->hdr_digest && !req->offset)
>>>           nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
>>> -    if (!req->h2cdata_left)
>>> +    if (!test_bit(NVME_TCP_Q_TLS, &queue->flags) && !req->h2cdata_left)
>>>           ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
>>>                   offset_in_page(pdu) + req->offset, len,
>>>                   MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
>>
>> Something is unclear to me. Is kernel_sendpage unsupported with tls? (I
>> think it is). I understand the motivation to add more checks in the code
>> for kernel_sendpage vs. sock_no_sendpage given that it should be
>> perfectly fine to use either.
>>
>> Did you see any regressions with using kernel_sendpage? If so, isn't
>> that a bug in the tls code?
> 
> The actual issue with the tls code is the 'MSG_EOR' handling.
> Problem is that tls is using MSG_EOR internally, and bails out on 
> unknown MSG_ settings:

That is fine, lets separate MSG_EOR for TLS, and any change from
kernel_sendpage to kernel_sendmsg to a different patch (or eliminate
it altogether if it is unneeded).

> 
> int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
> {
> [ .. ]
>          if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
>                                 MSG_CMSG_COMPAT))
>                  return -EOPNOTSUPP;
> 
> I would _vastly_ prefer to blank out unsupported flags (like MSG_EOR) 
> from the TLS code, because to all intents and purposes MSG_EOR is just 
> the opposite of MSG_MORE.

Not exactly. But ok.

> Or drop MSG_EOR usage from the nvme tcp code.

Possible to do. But MSG_EOR hints the network stack that no other
payload is expected, so it can send it down the wire asap and avoid
any batching heuristics.

> But then I'm not _that_ into the networking code to make a judgement here.
> And as we're using sendmsg() already I had been switching to use it for 
> ktls, too (as I know that the sendmsg() flow worked).
> But in the end I guess we could use sendpage going forward.

I'd prefer not to change that at this point if it is supported properly.