On Wed, Feb 12, 2025 at 7:52 AM Pavel Begunkov asml.silence@gmail.com wrote:
On 2/10/25 21:09, Mina Almasry wrote:
On Wed, Feb 5, 2025 at 4:20 AM Pavel Begunkov asml.silence@gmail.com wrote:
On 2/3/25 22:39, Mina Almasry wrote: ...
diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h index bb2b751d274a..3ff8f568c382 100644 --- a/include/linux/skbuff.h +++ b/include/linux/skbuff.h @@ -1711,9 +1711,12 @@ struct ubuf_info *msg_zerocopy_realloc(struct sock *sk, size_t size,
...
int zerocopy_fill_skb_from_iter(struct sk_buff *skb, struct iov_iter *from, size_t length); @@ -1721,12 +1724,14 @@ int zerocopy_fill_skb_from_iter(struct sk_buff *skb, static inline int skb_zerocopy_iter_dgram(struct sk_buff *skb, struct msghdr *msg, int len) {
return __zerocopy_sg_from_iter(msg, skb->sk, skb, &msg->msg_iter, len);
return __zerocopy_sg_from_iter(msg, skb->sk, skb, &msg->msg_iter, len,NULL);Instead of propagating it all the way down and carving a new path, why not reuse the existing infra? You already hook into where ubuf is allocated, you can stash the binding in there. And
It looks like it's not possible to increase the side of ubuf_info at all, otherwise the BUILD_BUG_ON in msg_zerocopy_alloc() fires.
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 readable, only in the case that the netmem is unreadable. I would not approve of considering netmem/devmem skbs "not real skbs", and start messing with the semantics of skb fields for devmem skbs, and having to start adding skb_is_devmem() checks through all code in the skb handlers that touch the fields being overwritten in the devmem case. No, I don't think we can re-use random fields in the skb for devmem.
stashing the binding in ubuf_info_ops (very hacky). Neither approach seems ideal, but the former may work and may be cleaner.
I'll take a deeper look here. I had looked before and concluded that we're piggybacking devmem TX on MSG_ZEROCOPY path, because we need almost all of the functionality there (no copying, send complete notifications, etc), with one minor change in the skb filling. I had concluded that if MSG_ZEROCOPY was never updated to use the existing infra, then it's appropriate for devmem TX piggybacking on top of it
MSG_ZEROCOPY does use the common infra, i.e. passing ubuf_info, but doesn't need ->sg_from_iter as zerocopy_fill_skb_from_iter() and it's what was there first.
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.
All in all I think I wanna prototype an msg->sg_from_iter approach and make a judgement call on whether it's cleaner than just passing the binding through a couple of helpers just as I'm doing here. My feeling is that the implementation in this patch may be cleaner than refactoring the entire msg_ubuf/sg_from_iter flows so we can sort of use it for MSG_ZEROCOPY with devmem when it currently doesn't use it.
to follow that. I would not want to get into a refactor of MSG_ZEROCOPY for no real reason.
But I'll take a deeper look here and see if I can make something slightly cleaner work.
zerocopy_fill_skb_from_devmem can implement ->sg_from_iter, see __zerocopy_sg_from_iter().
...
diff --git a/net/core/datagram.c b/net/core/datagram.c index f0693707aece..c989606ff58d 100644 --- a/net/core/datagram.c +++ b/net/core/datagram.c @@ -63,6 +63,8 @@ +static int +zerocopy_fill_skb_from_devmem(struct sk_buff *skb, struct iov_iter *from,
int length,struct net_devmem_dmabuf_binding *binding)+{
int i = skb_shinfo(skb)->nr_frags;size_t virt_addr, size, off;struct net_iov *niov;while (length && iov_iter_count(from)) {if (i == MAX_SKB_FRAGS)return -EMSGSIZE;virt_addr = (size_t)iter_iov_addr(from);Unless I missed it somewhere it needs to check that the iter is iovec based.
How do we end up here with an iterator that is not iovec based? Is the user able to trigger that somehow and I missed it?
Hopefully not, but for example io_uring passes bvecs for a number of requests that can end up in tcp_sendmsg_locked(). Those probably would work with the current patch, but check the order of some of the checks it will break. And once io_uring starts passing bvecs for normal send[msg] requests, it'd definitely be possible. And there are other in kernel users apart from send(2) path, so who knows.
The api allows it and therefore should be checked, it's better to avoid quite possible latent bugs.
Sounds good.