Re: [PATCH net-next v3 5/6] net: devmem: Implement TX path

[Pavel Begunkov <asml.silence@gmail.com>]

On 2/17/25 23:26, Mina Almasry wrote:
> On Thu, Feb 13, 2025 at 5:17 AM Pavel Begunkov <asml.silence@gmail.com> wrote:
...
>>>>> It's asserting that sizeof(ubuf_info_msgzc) <= sizeof(skb->cb), and
>>>>> I'm guessing increasing skb->cb size is not really the way to go.
>>>>>
>>>>> What I may be able to do here is stash the binding somewhere in
>>>>> ubuf_info_msgzc via union with fields we don't need for devmem, and/or
>>>>
>>>> It doesn't need to account the memory against the user, and you
>>>> actually don't want that because dmabuf should take care of that.
>>>> So, it should be fine to reuse ->mmp.
>>>>
>>>> It's also not a real sk_buff, so maybe maintainers wouldn't mind
>>>> reusing some more space out of it, if that would even be needed.
>>>>
>>>
>>> netmem skb are real sk_buff, with the modification that frags are not
>>
>> We were discussing ubuf_info allocation, take a look at
>> msg_zerocopy_alloc(), it has nothing to do with netmems and all that.
>>
> 
> Yes. My response was regarding the suggestion that we can use space in
> devmem skbs however we want though.

Well, at least I didn't suggest that, assuming "devmem skbs" are skbs
filled with devmem frags. I think the confusion here is thinking
that skb->cb you mentioned above is about "devmem skbs", while it's
special skbs without data used only to piggy back ubuf allocation.
Functionally speaking, it'd be perfectly fine to get rid of the
warning and allocate it with kmalloc().

...
>>> But MSG_ZEROCOPY doesn't set msg->msg_ubuf. And not setting
>>> msg->msg_ubuf fails to trigger msg->sg_from_iter altogether.
>>>
>>> And also currently sg_from_iter isn't set up to take in a ubuf_info.
>>> We'd need that if we stash the binding in the ubuf_info.
>>
>> https://github.com/isilence/linux.git sg-iter-ops
>>
>> I have old patches for all of that, they even rebased cleanly. That
>> should do it for you, and I need to send then regardless of devmem.
>>
>>
> 
> These patches help a bit, but do not make any meaningful dent in
> addressing the concern I have in the earlier emails.
> 
> The concern is that we're piggybacking devmem TX on MSG_ZEROCOPY, and
> currently the MSG_ZEROCOPY code carefully avoids any code paths
> setting msg->[sg_from_iter|msg_ubuf].

Fwiw, with that branch you don't need ->msg_ubuf at all, just pass
it as an argument from tcp_sendmsg_locked() as usual, and
->sg_from_iter is gone from there as well.

> If we want devmem to reuse both the MSG_ZEROCOPY mechanisms and the
> msg->[sg_from_iter|ubuf_info] mechanism, I have to dissect the
> MSG_ZEROCOPY code carefully so that it works with and without
> setting msg->[ubuf_info|msg->sg_from_iter]. Having gone through this
> rabbit hole so far I see that it complicates the implementation and
> adds more checks to the fast MSG_ZEROCOPY paths.

If you've already done, maybe you can post it as a draft? At least
it'll be obvious why you say it's more complicated.

> The complication could be worth it if there was some upside, but I
> don't see one tbh. Passing the binding down to
> zerocopy_fill_skb_from_devmem seems like a better approach to my eye
> so far

The upside is that 1) you currently you add overhead to common
path (incl copy), 2) passing it down through all the function also
have overhead to the zerocopy and MSG_ZEROCOPY path, which I'd
assume is comparable to those extra checks you have. 3) tcp would
need to know about devmem tcp and its bindings, while it all could
be in one spot under the MSG_ZEROCOPY check. 4) When you'd want
another protocol to support that, instead of a simple

ubuf = get_devmem_ubuf();

You'd need to plumb binding passing through the stack there as
well.

5) And keeping it in one place makes it easier to keep around.

I just don't see why it'd be complicated, but maybe I miss
something, which is why a draft prototype would explain it
better than any words.

> I'm afraid I'm going to table this for now. If there is overwhelming
> consensus that msg->sg_from_iter is the right approach here I will
> revisit, but it seems to me to complicate code without a significant
> upside.

-- 
Pavel Begunkov