io_uring/notif: implement notification stacking

The network stack allows only one ubuf_info per skb, and unlike
MSG_ZEROCOPY, each io_uring zerocopy send will carry a separate
ubuf_info. That means that send requests can't reuse a previosly
allocated skb and need to get one more or more of new ones. That's fine
for large sends, but otherwise it would spam the stack with lots of skbs
carrying just a little data each.

To help with that implement linking notification (i.e. an io_uring wrapper
around ubuf_info) into a list. Each is refcounted by skbs and the stack
as usual. additionally all non head entries keep a reference to the
head, which they put down when their refcount hits 0. When the head have
no more users, it'll efficiently put all notifications in a batch.

As mentioned previously about ->io_link_skb, the callback implementation
always allows to bind to an skb without a ubuf_info.

Reviewed-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/bf1e7f9b72f9ecc99999fdc0d2cded5eea87fd0b.1713369317.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

io_uring/notif.c

@@ -9,10 +9,17 @@
 #include "notif.h"
 #include "rsrc.h"
 
+static const struct ubuf_info_ops io_ubuf_ops;
+
 static void io_notif_tw_complete(struct io_kiocb *notif, struct io_tw_state *ts)
 {
 	struct io_notif_data *nd = io_notif_to_data(notif);
 
+	do {
+		notif = cmd_to_io_kiocb(nd);
+
+		lockdep_assert(refcount_read(&nd->uarg.refcnt) == 0);
+
 		if (unlikely(nd->zc_report) && (nd->zc_copied || !nd->zc_used))
 			notif->cqe.res |= IORING_NOTIF_USAGE_ZC_COPIED;
 
@@ -20,7 +27,10 @@ static void io_notif_tw_complete(struct io_kiocb *notif, struct io_tw_state *ts)
 			__io_unaccount_mem(notif->ctx->user, nd->account_pages);
 			nd->account_pages = 0;
 		}
+
+		nd = nd->next;
 		io_req_task_complete(notif, ts);
+	} while (nd);
 }
 
 void io_tx_ubuf_complete(struct sk_buff *skb, struct ubuf_info *uarg,
@@ -39,12 +49,56 @@ void io_tx_ubuf_complete(struct sk_buff *skb, struct ubuf_info *uarg,
 	if (!refcount_dec_and_test(&uarg->refcnt))
 		return;
 
+	if (nd->head != nd) {
+		io_tx_ubuf_complete(skb, &nd->head->uarg, success);
+		return;
+	}
 	notif->io_task_work.func = io_notif_tw_complete;
 	__io_req_task_work_add(notif, IOU_F_TWQ_LAZY_WAKE);
 }
 
+static int io_link_skb(struct sk_buff *skb, struct ubuf_info *uarg)
+{
+	struct io_notif_data *nd, *prev_nd;
+	struct io_kiocb *prev_notif, *notif;
+	struct ubuf_info *prev_uarg = skb_zcopy(skb);
+
+	nd = container_of(uarg, struct io_notif_data, uarg);
+	notif = cmd_to_io_kiocb(nd);
+
+	if (!prev_uarg) {
+		net_zcopy_get(&nd->uarg);
+		skb_zcopy_init(skb, &nd->uarg);
+		return 0;
+	}
+	/* handle it separately as we can't link a notif to itself */
+	if (unlikely(prev_uarg == &nd->uarg))
+		return 0;
+	/* we can't join two links together, just request a fresh skb */
+	if (unlikely(nd->head != nd || nd->next))
+		return -EEXIST;
+	/* don't mix zc providers */
+	if (unlikely(prev_uarg->ops != &io_ubuf_ops))
+		return -EEXIST;
+
+	prev_nd = container_of(prev_uarg, struct io_notif_data, uarg);
+	prev_notif = cmd_to_io_kiocb(nd);
+
+	/* make sure all noifications can be finished in the same task_work */
+	if (unlikely(notif->ctx != prev_notif->ctx ||
+		     notif->task != prev_notif->task))
+		return -EEXIST;
+
+	nd->head = prev_nd->head;
+	nd->next = prev_nd->next;
+	prev_nd->next = nd;
+	net_zcopy_get(&nd->head->uarg);
+	return 0;
+}
+
 static const struct ubuf_info_ops io_ubuf_ops = {
 	.complete = io_tx_ubuf_complete,
+	.link_skb = io_link_skb,
 };
 
 struct io_kiocb *io_alloc_notif(struct io_ring_ctx *ctx)
@@ -65,6 +119,9 @@ struct io_kiocb *io_alloc_notif(struct io_ring_ctx *ctx)
 	nd = io_notif_to_data(notif);
 	nd->zc_report = false;
 	nd->account_pages = 0;
+	nd->next = NULL;
+	nd->head = nd;
+
 	nd->uarg.flags = IO_NOTIF_UBUF_FLAGS;
 	nd->uarg.ops = &io_ubuf_ops;
 	refcount_set(&nd->uarg.refcnt, 1);

io_uring/notif.h

@@ -14,6 +14,9 @@ struct io_notif_data {
 	struct file		*file;
 	struct ubuf_info	uarg;
 
+	struct io_notif_data	*next;
+	struct io_notif_data	*head;
+
 	unsigned		account_pages;
 	bool			zc_report;
 	bool			zc_used;