On Thu, Dec 15, 2022 at 04:04:53AM -0500, Michael S. Tsirkin wrote:

On Thu, Dec 15, 2022 at 09:47:57AM +0100, Stefano Garzarella wrote:

quoted

On Thu, Dec 15, 2022 at 04:36:44AM +0000, Bobby Eshleman wrote:

quoted

This commit changes virtio/vsock to use sk_buff instead of
virtio_vsock_pkt. Beyond better conforming to other net code, using
sk_buff allows vsock to use sk_buff-dependent features in the future
(such as sockmap) and improves throughput.

This patch introduces the following performance changes:

Tool/Config: uperf w/ 64 threads, SOCK_STREAM
Test Runs: 5, mean of results
Before: commit 95ec6bce2a0b ("Merge branch 'net-ipa-more-endpoints'")

Test: 64KB, g2h
Before: 21.63 Gb/s
After: 25.59 Gb/s (+18%)

Test: 16B, g2h
Before: 11.86 Mb/s
After: 17.41 Mb/s (+46%)

Test: 64KB, h2g
Before: 2.15 Gb/s
After: 3.6 Gb/s (+67%)

Test: 16B, h2g
Before: 14.38 Mb/s
After: 18.43 Mb/s (+28%)

Signed-off-by: Bobby Eshleman <redacted>
---

Testing: passed vsock_test h2g and g2h
Note2: net-next is closed, but sending out now in case more comments
roll in, as discussed here:
https://lore.kernel.org/all/Y34SoH1nFTXXLWbK@bullseye/ (local)

Changes in v8:
- vhost/vsock: remove unused enum
- vhost/vsock: use spin_lock_bh() instead of spin_lock()
- vsock/loopback: use __skb_dequeue instead of skb_dequeue

Changes in v7:
- use skb_queue_empty() instead of skb_queue_empty_lockless()

Changes in v6:
- use skb->cb instead of skb->_skb_refdst
- use lock-free __skb_queue_purge for rx_queue when rx_lock held

Changes in v5:
- last_skb instead of skb: last_hdr->len = cpu_to_le32(last_skb->len)

Changes in v4:
- vdso/bits.h -> linux/bits.h
- add virtio_vsock_alloc_skb() helper
- virtio/vsock: rename buf_len -> total_len
- update last_hdr->len
- fix build_skb() for vsockmon (tested)
- add queue helpers
- use spin_{unlock/lock}_bh() instead of spin_lock()/spin_unlock()
- note: I only ran a few g2h tests to check that this change
 had no perf impact. The above data is still from patch
 v3.

Changes in v3:
- fix seqpacket bug
- use zero in vhost_add_used(..., 0) device doesn't write to buffer
- use xmas tree style declarations
- vsock_hdr() -> virtio_vsock_hdr() and other include file style fixes
- no skb merging
- save space by not using vsock_metadata
- use _skb_refdst instead of skb buffer space for flags
- use skb_pull() to keep track of read bytes instead of using an an
 extra variable 'off' in the skb buffer space
- remove unnecessary sk_allocation assignment
- do not zero hdr needlessly
- introduce virtio_transport_skb_len() because skb->len changes now
- use spin_lock() directly on queue lock instead of sk_buff_head helpers
 which use spin_lock_irqsave() (e.g., skb_dequeue)
- do not reduce buffer size to be page size divisible
- Note: the biggest performance change came from loosening the spinlock
 variation and not reducing the buffer size.

Changes in v2:
- Use alloc_skb() directly instead of sock_alloc_send_pskb() to minimize
 uAPI changes.
- Do not marshal errors to -ENOMEM for non-virtio implementations.
- No longer a part of the original series
- Some code cleanup and refactoring
- Include performance stats

drivers/vhost/vsock.c                   | 213 +++++-------
include/linux/virtio_vsock.h            | 126 +++++--
net/vmw_vsock/virtio_transport.c        | 149 +++------
net/vmw_vsock/virtio_transport_common.c | 422 +++++++++++++-----------
net/vmw_vsock/vsock_loopback.c          |  51 +--
5 files changed, 495 insertions(+), 466 deletions(-)

diff --git a/drivers/vhost/vsock.c b/drivers/vhost/vsock.c
index 5703775af129..ee0a00432cb1 100644
--- a/drivers/vhost/vsock.c
+++ b/drivers/vhost/vsock.c

@@ -51,8 +51,7 @@ struct vhost_vsock {

	struct hlist_node hash;

	struct vhost_work send_pkt_work;
-	spinlock_t send_pkt_list_lock;
-	struct list_head send_pkt_list;	/* host->guest pending packets */
+	struct sk_buff_head send_pkt_queue; /* host->guest pending packets */

	atomic_t queued_replies;

@@ -108,40 +107,33 @@ vhost_transport_do_send_pkt(struct vhost_vsock

*vsock,
	vhost_disable_notify(&vsock->dev, vq);

	do {
-		struct virtio_vsock_pkt *pkt;
+		struct virtio_vsock_hdr *hdr;
+		size_t iov_len, payload_len;
		struct iov_iter iov_iter;
+		u32 flags_to_restore = 0;
+		struct sk_buff *skb;
		unsigned out, in;
		size_t nbytes;
-		size_t iov_len, payload_len;
		int head;
-		u32 flags_to_restore = 0;

-		spin_lock_bh(&vsock->send_pkt_list_lock);
-		if (list_empty(&vsock->send_pkt_list)) {
-			spin_unlock_bh(&vsock->send_pkt_list_lock);
+		spin_lock_bh(&vsock->send_pkt_queue.lock);
+		skb = __skb_dequeue(&vsock->send_pkt_queue);
+		spin_unlock_bh(&vsock->send_pkt_queue.lock);

Sorry for coming late, but I just commented in Paolo's reply.
Here I think we can directly use the new virtio_vsock_skb_dequeue().

Yea, pretty late.
And using that will prevent me from merging this since it's not in my tree yet.
We can do a cleanup patch on top.

Yep, sure!
Functionally nothing changes, so it's fine with a patch on top.

So, for this patch:

Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>

Thanks,
Stefano