On 01.09.2023 15:30, Stefano Garzarella wrote:
> On Sun, Aug 27, 2023 at 11:54:36AM +0300, Arseniy Krasnov wrote:
>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>> flag is set and zerocopy transmission is possible (enabled in socket
>> options and transport allows zerocopy), then non-linear skb will be
>> created and filled with the pages of user's buffer. Pages of user's
>> buffer are locked in memory by 'get_user_pages()'. Second thing that
>> this patch does is replace type of skb owning: instead of calling
>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>> of socket, so to decrease this field correctly proper skb destructor is
>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>>
>> Signed-off-by: Arseniy Krasnov <[email protected]>
>> ---
>> Changelog:
>> v5(big patchset) -> v1:
>> * Refactorings of 'if' conditions.
>> * Remove extra blank line.
>> * Remove 'frag_off' field unneeded init.
>> * Add function 'virtio_transport_fill_skb()' which fills both linear
>> and non-linear skb with provided data.
>> v1 -> v2:
>> * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>> v2 -> v3:
>> * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>> provided 'iov_iter' with data could be sent in a zerocopy mode.
>> If this callback is not set in transport - transport allows to send
>> any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>> then zerocopy is allowed. Reason of this callback is that in case of
>> G2H transmission we insert whole skb to the tx virtio queue and such
>> skb must fit to the size of the virtio queue to be sent in a single
>> iteration (may be tx logic in 'virtio_transport.c' could be reworked
>> as in vhost to support partial send of current skb). This callback
>> will be enabled only for G2H path. For details pls see comment
>> 'Check that tx queue...' below.
>> v3 -> v4:
>> * 'msgzerocopy_check_iov' moved from 'struct vsock_transport' to
>> 'struct virtio_transport' as it is virtio specific callback and
>> never needed in other transports.
>> v4 -> v5:
>> * 'msgzerocopy_check_iov' renamed to 'can_msgzerocopy' and now it
>> uses number of buffers to send as input argument. I think there is
>> no need to pass iov to this callback (at least today, it is used only
>> by guest side of virtio transport), because the only thing that this
>> callback does is comparison of number of buffers to be inserted to
>> the tx queue and size of this queue.
>> * Remove any checks for type of current 'iov_iter' with payload (is it
>> 'iovec' or 'ubuf'). These checks left from the earlier versions where I
>> didn't use already implemented kernel API which handles every type of
>> 'iov_iter'.
>> v5 -> v6:
>> * Refactor 'virtio_transport_fill_skb()'.
>> * Add 'WARN_ON_ONCE()' and comment on invalid combination of destination
>> socket and payload in 'virtio_transport_alloc_skb()'.
>>
>> include/linux/virtio_vsock.h | 5 +
>> net/vmw_vsock/virtio_transport.c | 33 ++++
>> net/vmw_vsock/virtio_transport_common.c | 250 ++++++++++++++++++------
>> 3 files changed, 231 insertions(+), 57 deletions(-)
>>
>> diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
>> index a91fbdf233e4..56501cd9843f 100644
>> --- a/include/linux/virtio_vsock.h
>> +++ b/include/linux/virtio_vsock.h
>> @@ -160,6 +160,11 @@ struct virtio_transport {
>>
>> /* Takes ownership of the packet */
>> int (*send_pkt)(struct sk_buff *skb);
>> +
>> + /* Used in MSG_ZEROCOPY mode. Checks that provided data
>> + * could be transmitted with zerocopy mode.
>> + */
>> + bool (*can_msgzerocopy)(int bufs_num);
>> };
>>
>> ssize_t
>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>> index 8636477cf088..4ce44916e585 100644
>> --- a/net/vmw_vsock/virtio_transport.c
>> +++ b/net/vmw_vsock/virtio_transport.c
>> @@ -453,6 +453,38 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>> queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>> }
>>
>> +static bool virtio_transport_can_msgzerocopy(int bufs_num)
>> +{
>> + struct virtio_vsock *vsock;
>> + bool res = false;
>> +
>> + rcu_read_lock();
>> +
>> + vsock = rcu_dereference(the_virtio_vsock);
>> + if (vsock) {
>> + struct virtqueue *vq = vsock->vqs[VSOCK_VQ_TX];
>> +
>> + /* Check that tx queue is large enough to keep whole
>> + * data to send. This is needed, because when there is
>> + * not enough free space in the queue, current skb to
>> + * send will be reinserted to the head of tx list of
>> + * the socket to retry transmission later, so if skb
>> + * is bigger than whole queue, it will be reinserted
>> + * again and again, thus blocking other skbs to be sent.
>> + * Each page of the user provided buffer will be added
>> + * as a single buffer to the tx virtqueue, so compare
>> + * number of pages against maximum capacity of the queue.
>> + * +1 means buffer for the packet header.
>> + */
>> + if (bufs_num + 1 <= vq->num_max)
>
> I'd move the + 1 in the caller.
>
>> + res = true;
>> + }
>> +
>> + rcu_read_unlock();
>> +
>> + return res;
>> +}
>> +
>> static bool virtio_transport_seqpacket_allow(u32 remote_cid);
>>
>> static struct virtio_transport virtio_transport = {
>> @@ -502,6 +534,7 @@ static struct virtio_transport virtio_transport = {
>> },
>>
>> .send_pkt = virtio_transport_send_pkt,
>> + .can_msgzerocopy = virtio_transport_can_msgzerocopy,
>> };
>>
>> static bool virtio_transport_seqpacket_allow(u32 remote_cid)
>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
>> index 3a48e48a99ac..7f42ad131a69 100644
>> --- a/net/vmw_vsock/virtio_transport_common.c
>> +++ b/net/vmw_vsock/virtio_transport_common.c
>> @@ -37,27 +37,99 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
>> return container_of(t, struct virtio_transport, transport);
>> }
>>
>> -/* Returns a new packet on success, otherwise returns NULL.
>> - *
>> - * If NULL is returned, errp is set to a negative errno.
>> - */
>> -static struct sk_buff *
>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
>> - size_t len,
>> - u32 src_cid,
>> - u32 src_port,
>> - u32 dst_cid,
>> - u32 dst_port)
>> -{
>> - const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
>> - struct virtio_vsock_hdr *hdr;
>> - struct sk_buff *skb;
>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
>> + size_t max_to_send)
> ^
> I'd call it `pkt_len`, `max_to_send` is confusing IMHO. I didn't
> initially if it was the number of buffers or bytes.
>
>> +{
>> + const struct virtio_transport *t_ops;
>> + struct iov_iter *iov_iter;
>> +
>> + if (!info->msg)
>> + return false;
>> +
>> + iov_iter = &info->msg->msg_iter;
>> +
>> + if (iov_iter->iov_offset)
>> + return false;
>> +
>> + /* We can't send whole iov. */
>> + if (iov_iter->count > max_to_send)
>> + return false;
>> +
>> + /* Check that transport can send data in zerocopy mode. */
>> + t_ops = virtio_transport_get_ops(info->vsk);
>> +
>> + if (t_ops->can_msgzerocopy) {
>
> So if `can_msgzerocopy` is not implemented, we always return true after
> this point. Should we mention it in the .can_msgzerocopy documentation?
Ops, this is my mistake, I must return 'false' in this case. Seems I didn't
catch this problem with my tests, because there was no test case where
zerocopy will fallback to copy!
I'll fix it and add new test!
>
> Can we also mention in the commit description why this is need only for
> virtio_tranport and not for vhost and loopback?
>
>> + int pages_in_iov = iov_iter_npages(iov_iter, MAX_SKB_FRAGS);
>> + int pages_to_send = min(pages_in_iov, MAX_SKB_FRAGS);
>> +
>> + return t_ops->can_msgzerocopy(pages_to_send);
>> + }
>> +
>> + return true;
>> +}
>> +
>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
>> + struct sk_buff *skb,
>> + struct msghdr *msg,
>> + bool zerocopy)
>> +{
>> + struct ubuf_info *uarg;
>> +
>> + if (msg->msg_ubuf) {
>> + uarg = msg->msg_ubuf;
>> + net_zcopy_get(uarg);
>> + } else {
>> + struct iov_iter *iter = &msg->msg_iter;
>> + struct ubuf_info_msgzc *uarg_zc;
>> +
>> + uarg = msg_zerocopy_realloc(sk_vsock(vsk),
>> + iter->count,
>> + NULL);
>> + if (!uarg)
>> + return -1;
>> +
>> + uarg_zc = uarg_to_msgzc(uarg);
>> + uarg_zc->zerocopy = zerocopy ? 1 : 0;
>> + }
>> +
>> + skb_zcopy_init(skb, uarg);
>> +
>> + return 0;
>> +}
>> +
>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
>> + struct virtio_vsock_pkt_info *info,
>> + size_t len,
>> + bool zcopy)
>> +{
>> void *payload;
>> int err;
>>
>> - skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>> - if (!skb)
>> - return NULL;
>> + if (zcopy)
>> + return __zerocopy_sg_from_iter(info->msg, NULL, skb,
>> + &info->msg->msg_iter,
>> + len);
>> +
>> + payload = skb_put(skb, len);
>> + err = memcpy_from_msg(payload, info->msg, len);
>> + if (err)
>> + return -1;
>> +
>> + if (msg_data_left(info->msg))
>> + return 0;
>> +
>> + return 0;
>> +}
>> +
>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
>> + struct virtio_vsock_pkt_info *info,
>> + u32 src_cid,
>> + u32 src_port,
>> + u32 dst_cid,
>> + u32 dst_port,
>> + size_t len)
>
> I'd move `len` after `info` like in virtio_transport_alloc_skb().
> And maybe we can use the same name on both (`len` or `payload_len`).
>
>> +{
>> + struct virtio_vsock_hdr *hdr;
>>
>> hdr = virtio_vsock_hdr(skb);
>> hdr->type = cpu_to_le16(info->type);
>> @@ -68,42 +140,6 @@ virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
>> hdr->dst_port = cpu_to_le32(dst_port);
>> hdr->flags = cpu_to_le32(info->flags);
>> hdr->len = cpu_to_le32(len);
>> -
>> - if (info->msg && len > 0) {
>> - payload = skb_put(skb, len);
>> - err = memcpy_from_msg(payload, info->msg, len);
>> - if (err)
>> - goto out;
>> -
>> - if (msg_data_left(info->msg) == 0 &&
>> - info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>> - hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>> -
>> - if (info->msg->msg_flags & MSG_EOR)
>> - hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>> - }
>> - }
>> -
>> - if (info->reply)
>> - virtio_vsock_skb_set_reply(skb);
>> -
>> - trace_virtio_transport_alloc_pkt(src_cid, src_port,
>> - dst_cid, dst_port,
>> - len,
>> - info->type,
>> - info->op,
>> - info->flags);
>> -
>> - if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
>> - WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
>> - goto out;
>> - }
>> -
>> - return skb;
>> -
>> -out:
>> - kfree_skb(skb);
>> - return NULL;
>> }
>>
>> static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
>> @@ -214,6 +250,78 @@ static u16 virtio_transport_get_type(struct sock *sk)
>> return VIRTIO_VSOCK_TYPE_SEQPACKET;
>> }
>>
>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
>> + struct virtio_vsock_pkt_info *info,
>> + size_t payload_len,
>> + bool zcopy,
>> + u32 src_cid,
>> + u32 src_port,
>> + u32 dst_cid,
>> + u32 dst_port)
>> +{
>> + struct sk_buff *skb;
>> + size_t skb_len;
>> +
>> + skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
>> +
>> + if (!zcopy)
>> + skb_len += payload_len;
>> +
>> + skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>> + if (!skb)
>> + return NULL;
>> +
>> + virtio_transport_init_hdr(skb, info, src_cid, src_port,
>> + dst_cid, dst_port,
>> + payload_len);
>> +
>> + /* If 'vsk' != NULL then payload is always present, so we
>> + * will never call '__zerocopy_sg_from_iter()' below without
>> + * setting skb owner in 'skb_set_owner_w()'. The only case
>> + * when 'vsk' == NULL is VIRTIO_VSOCK_OP_RST control message
>> + * without payload.
>> + */
>> + WARN_ON_ONCE(!(vsk && (info->msg && payload_len)) && zcopy);
>> +
>> + /* Set owner here, because '__zerocopy_sg_from_iter()' uses
>> + * owner of skb without check to update 'sk_wmem_alloc'.
>> + */
>> + if (vsk)
>> + skb_set_owner_w(skb, sk_vsock(vsk));
>> +
>> + if (info->msg && payload_len > 0) {
>> + int err;
>> +
>> + err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
>> + if (err)
>> + goto out;
>> +
>> + if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>> + struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
>> +
>> + hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>> +
>> + if (info->msg->msg_flags & MSG_EOR)
>> + hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>> + }
>> + }
>> +
>> + if (info->reply)
>> + virtio_vsock_skb_set_reply(skb);
>> +
>> + trace_virtio_transport_alloc_pkt(src_cid, src_port,
>> + dst_cid, dst_port,
>> + payload_len,
>> + info->type,
>> + info->op,
>> + info->flags);
>> +
>> + return skb;
>> +out:
>> + kfree_skb(skb);
>> + return NULL;
>> +}
>> +
>> /* This function can only be used on connecting/connected sockets,
>> * since a socket assigned to a transport is required.
>> *
>> @@ -222,10 +330,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
>> static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>> struct virtio_vsock_pkt_info *info)
>> {
>> + u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
>> u32 src_cid, src_port, dst_cid, dst_port;
>> const struct virtio_transport *t_ops;
>> struct virtio_vsock_sock *vvs;
>> u32 pkt_len = info->pkt_len;
>> + bool can_zcopy = false;
>> u32 rest_len;
>> int ret;
>>
>> @@ -254,15 +364,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>> if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
>> return pkt_len;
>>
>> + if (info->msg) {
>> + /* If zerocopy is not enabled by 'setsockopt()', we behave as
>> + * there is no MSG_ZEROCOPY flag set.
>> + */
>> + if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
>> + info->msg->msg_flags &= ~MSG_ZEROCOPY;
>> +
>> + if (info->msg->msg_flags & MSG_ZEROCOPY)
>> + can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
>> +
>> + if (can_zcopy)
>> + max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
>> + (MAX_SKB_FRAGS * PAGE_SIZE));
>> + }
>> +
>> rest_len = pkt_len;
>>
>> do {
>> struct sk_buff *skb;
>> size_t skb_len;
>>
>> - skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
>> + skb_len = min(max_skb_len, rest_len);
>>
>> - skb = virtio_transport_alloc_skb(info, skb_len,
>> + skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
>> src_cid, src_port,
>> dst_cid, dst_port);
>> if (!skb) {
>> @@ -270,6 +395,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>> break;
>> }
>>
>> + /* This is last skb to send this portion of data. */
>
> Sorry I didn't get it :-(
>
> Can you elaborate this a bit more?
I mean that we iterate over user's buffer here, allocating skb on each
iteration. And for last skb for this buffer we initialize completion
for user (we need to allocate one completion for one syscall).
Thanks for review, I'll fix all other comments and resend patchset when
'net-next' will be opened again.
Thanks, Arseniy
>
> Thanks,
> Stefano
>
>> + if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
>> + skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
>> + if (virtio_transport_init_zcopy_skb(vsk, skb,
>> + info->msg,
>> + can_zcopy)) {
>> + ret = -ENOMEM;
>> + break;
>> + }
>> + }
>> +
>> virtio_transport_inc_tx_pkt(vvs, skb);
>>
>> ret = t_ops->send_pkt(skb);
>> @@ -985,7 +1121,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
>> if (!t)
>> return -ENOTCONN;
>>
>> - reply = virtio_transport_alloc_skb(&info, 0,
>> + reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
>> le64_to_cpu(hdr->dst_cid),
>> le32_to_cpu(hdr->dst_port),
>> le64_to_cpu(hdr->src_cid),
>> --
>> 2.25.1
>>
>