lkml.org 
[lkml]   [2021]   [Jun]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [MASSMAIL KLMS] Re: [PATCH v11 11/18] virtio/vsock: dequeue callback for SOCK_SEQPACKET
On Fri, Jun 18, 2021 at 07:08:30PM +0300, Arseny Krasnov wrote:
>
>On 18.06.2021 18:55, Stefano Garzarella wrote:
>> On Fri, Jun 18, 2021 at 06:04:37PM +0300, Arseny Krasnov wrote:
>>> On 18.06.2021 16:44, Stefano Garzarella wrote:
>>>> Hi Arseny,
>>>> the series looks great, I have just a question below about
>>>> seqpacket_dequeue.
>>>>
>>>> I also sent a couple a simple fixes, it would be great if you can review
>>>> them:
>>>> https://lore.kernel.org/netdev/20210618133526.300347-1-sgarzare@redhat.com/
>>>>
>>>>
>>>> On Fri, Jun 11, 2021 at 02:12:38PM +0300, Arseny Krasnov wrote:
>>>>> Callback fetches RW packets from rx queue of socket until whole record
>>>>> is copied(if user's buffer is full, user is not woken up). This is done
>>>>> to not stall sender, because if we wake up user and it leaves syscall,
>>>>> nobody will send credit update for rest of record, and sender will wait
>>>>> for next enter of read syscall at receiver's side. So if user buffer is
>>>>> full, we just send credit update and drop data.
>>>>>
>>>>> Signed-off-by: Arseny Krasnov <arseny.krasnov@kaspersky.com>
>>>>> ---
>>>>> v10 -> v11:
>>>>> 1) 'msg_count' field added to count current number of EORs.
>>>>> 2) 'msg_ready' argument removed from callback.
>>>>> 3) If 'memcpy_to_msg()' failed during copy loop, there will be
>>>>> no next attempts to copy data, rest of record will be freed.
>>>>>
>>>>> include/linux/virtio_vsock.h | 5 ++
>>>>> net/vmw_vsock/virtio_transport_common.c | 84 +++++++++++++++++++++++++
>>>>> 2 files changed, 89 insertions(+)
>>>>>
>>>>> diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
>>>>> index dc636b727179..1d9a302cb91d 100644
>>>>> --- a/include/linux/virtio_vsock.h
>>>>> +++ b/include/linux/virtio_vsock.h
>>>>> @@ -36,6 +36,7 @@ struct virtio_vsock_sock {
>>>>> u32 rx_bytes;
>>>>> u32 buf_alloc;
>>>>> struct list_head rx_queue;
>>>>> + u32 msg_count;
>>>>> };
>>>>>
>>>>> struct virtio_vsock_pkt {
>>>>> @@ -80,6 +81,10 @@ virtio_transport_dgram_dequeue(struct vsock_sock *vsk,
>>>>> struct msghdr *msg,
>>>>> size_t len, int flags);
>>>>>
>>>>> +ssize_t
>>>>> +virtio_transport_seqpacket_dequeue(struct vsock_sock *vsk,
>>>>> + struct msghdr *msg,
>>>>> + int flags);
>>>>> s64 virtio_transport_stream_has_data(struct vsock_sock *vsk);
>>>>> s64 virtio_transport_stream_has_space(struct vsock_sock *vsk);
>>>>>
>>>>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
>>>>> index ad0d34d41444..1e1df19ec164 100644
>>>>> --- a/net/vmw_vsock/virtio_transport_common.c
>>>>> +++ b/net/vmw_vsock/virtio_transport_common.c
>>>>> @@ -393,6 +393,78 @@ virtio_transport_stream_do_dequeue(struct vsock_sock *vsk,
>>>>> return err;
>>>>> }
>>>>>
>>>>> +static int virtio_transport_seqpacket_do_dequeue(struct vsock_sock *vsk,
>>>>> + struct msghdr *msg,
>>>>> + int flags)
>>>>> +{
>>>>> + struct virtio_vsock_sock *vvs = vsk->trans;
>>>>> + struct virtio_vsock_pkt *pkt;
>>>>> + int dequeued_len = 0;
>>>>> + size_t user_buf_len = msg_data_left(msg);
>>>>> + bool copy_failed = false;
>>>>> + bool msg_ready = false;
>>>>> +
>>>>> + spin_lock_bh(&vvs->rx_lock);
>>>>> +
>>>>> + if (vvs->msg_count == 0) {
>>>>> + spin_unlock_bh(&vvs->rx_lock);
>>>>> + return 0;
>>>>> + }
>>>>> +
>>>>> + while (!msg_ready) {
>>>>> + pkt = list_first_entry(&vvs->rx_queue, struct virtio_vsock_pkt, list);
>>>>> +
>>>>> + if (!copy_failed) {
>>>>> + size_t pkt_len;
>>>>> + size_t bytes_to_copy;
>>>>> +
>>>>> + pkt_len = (size_t)le32_to_cpu(pkt->hdr.len);
>>>>> + bytes_to_copy = min(user_buf_len, pkt_len);
>>>>> +
>>>>> + if (bytes_to_copy) {
>>>>> + int err;
>>>>> +
>>>>> + /* sk_lock is held by caller so no one else can dequeue.
>>>>> + * Unlock rx_lock since memcpy_to_msg() may sleep.
>>>>> + */
>>>>> + spin_unlock_bh(&vvs->rx_lock);
>>>>> +
>>>>> + err = memcpy_to_msg(msg, pkt->buf, bytes_to_copy);
>>>>> + if (err) {
>>>>> + /* Copy of message failed, set flag to skip
>>>>> + * copy path for rest of fragments. Rest of
>>>>> + * fragments will be freed without copy.
>>>>> + */
>>>>> + copy_failed = true;
>>>>> + dequeued_len = err;
>>>> If we fail to copy the message we will discard the entire packet.
>>>> Is it acceptable for the user point of view, or we should leave the
>>>> packet in the queue and the user can retry, maybe with a different
>>>> buffer?
>>>>
>>>> Then we can remove the packets only when we successfully copied all the
>>>> fragments.
>>>>
>>>> I'm not sure make sense, maybe better to check also other
>>>> implementations :-)
>>>>
>>>> Thanks,
>>>> Stefano
>>> Understand, i'll check it on weekend, anyway I think it is
>>> not critical for implementation.
>> Yep, I agree.
>>
>>>
>>> I have another question: may be it is useful to research for
>>> approach where packets are not queued until whole message
>>> is received, but copied to user's buffer thus freeing memory.
>>> (like previous implementation, of course with solution of problem
>>> where part of message still in queue, while reader was woken
>>> by timeout or signal).
>>>
>>> I think it is better, because  in current version, sender may set
>>> 'peer_alloc_buf' to  for example 1MB, so at receiver we get
>>> 1MB of 'kmalloc()' memory allocated, while having user's buffer
>>> to copy data there or drop it(if user's buffer is full). This way
>>> won't change spec(e.g. no message id or SEQ_BEGIN will be added).
>>>
>>> What do You think?
>> Yep, I see your point and it would be great, but I think the main issues
>> to fix is how to handle a signal while we are waiting other fragments
>> since the other peer can take unspecified time to send them.
>
>What about transport callback, something like 'seqpacket_drain()' or
>
>'seqpacket_drop_curr()' - when we got signal or timeout, notify transport
>
>to drop current message. In virtio case this will set special flag in transport,
>
>so on next dequeue, this flag is checked and if it is set - we drop all packets
>
>until EOR found. Then we can copy untouched new record.
>

But in this way, we will lose the entire message.

Is it acceptable for seqpacket?

Stefano

\
 
 \ /
  Last update: 2021-06-18 18:27    [W:0.105 / U:1.916 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site