Consider a bi-directional TCP connection:
1. we use all IOBs for tx queue
2. we advertize zero recv window because we have no free IOBs
3. if the peer tcp does the same thing,
both sides advertize zero window and can not drain the tx queue.
For a similar stall to happen, the peer doesn't need to be
a naive tcp implementation like nuttx. A naive application blocking
on send() without draining its read buffer is enough.
(Probably such an application should be fixed to drain rx even
when tx is full. However, it's another story.)
This commit avoids the situation by prevent tx from grabbing
the all IOBs in the first place. (assuming CONFIG_IOB_THROTTLE > 0)
Gregory Nutt is the copyright holder for those files and he has submitted the
SGA as a result we can migrate the licenses to Apache.
Signed-off-by: Alin Jerpelea <alin.jerpelea@sony.com>
My recent changes to buffered tcp send broke this. [1]
One of my local apps using non-blocking tcp is working
again with this fix.
[1]
```
commit 837e1a72a4
Author: YAMAMOTO Takashi <yamamoto@midokura.com>
Date: Mon Mar 15 16:19:42 2021 +0900
tcp_send_buffered.c: improve tcp write buffering
```
Summary:
- Based on the discussion (PR#2772), let me revert the commit
Impact:
- None
Testing:
- N/A
This reverts commit ec8bf5c8c1.
Suggested-by: YAMAMOTO Takashi <yamamoto@midokura.com>
Signed-off-by: Masayuki Ishikawa <Masayuki.Ishikawa@jp.sony.com>
Summary:
- This commit adds DEBUGASSERT() to check the IOB size
Impact:
- None
Testing:
- Tested with sabre-6quad:netnsh with QEMU
Signed-off-by: Masayuki Ishikawa <Masayuki.Ishikawa@jp.sony.com>
RFC2001: TCP Slow Start, Congestion Avoidance, Fast Retransmit,
and Fast Recovery Algorithms
...
3. Fast Retransmit
Modifications to the congestion avoidance algorithm were proposed in
1990 [3]. Before describing the change, realize that TCP may
generate an immediate acknowledgment (a duplicate ACK) when an out-
of-order segment is received (Section 4.2.2.21 of [1], with a note
that one reason for doing so was for the experimental fast-
retransmit algorithm). This duplicate ACK should not be delayed.
The purpose of this duplicate ACK is to let the other end know that a
segment was received out of order, and to tell it what sequence
number is expected.
Since TCP does not know whether a duplicate ACK is caused by a lost
segment or just a reordering of segments, it waits for a small number
of duplicate ACKs to be received. It is assumed that if there is
just a reordering of the segments, there will be only one or two
duplicate ACKs before the reordered segment is processed, which will
then generate a new ACK. If three or more duplicate ACKs are
received in a row, it is a strong indication that a segment has been
lost. TCP then performs a retransmission of what appears to be the
missing segment, without waiting for a retransmission timer to
expire.
Change-Id: Ie2cbcecab507c3d831f74390a6a85e0c5c8e0652
Signed-off-by: chao.an <anchao@xiaomi.com>
remove the connection assertion since the instance will be invalid
if the network device has been taken down.
net/netdev/netdev_ioctl.c:
1847 void netdev_ifdown(FAR struct net_driver_s *dev)
1848 {
...
1871 /* Notify clients that the network has been taken down */
1872
1873 devif_dev_event(dev, NULL, NETDEV_DOWN);
...
1883 }
Change-Id: I492b97b5ebe035ea67bbdd7ed635cb13d085e89c
Signed-off-by: chao.an <anchao@xiaomi.com>
In the current net stack implementation, there is no mechanism
for notifying the loss of the wireless connection, if the network
is disconnected then application sends data packets through tcp,
the tcp_timer will keep retrying fetch the ack for awhile, the
connection status will not be able to be switched timely.
Change-Id: I84d1121527edafc6ee6ad56ba164838694e7e11c
Signed-off-by: chao.an <anchao@xiaomi.com>
MSG_DONTWAIT (since Linux 2.2)
Enables nonblocking operation; if the operation would block, the
call fails with the error EAGAIN or EWOULDBLOCK. This provides
similar behavior to setting the O_NONBLOCK flag (via the fcntl(2)
F_SETFL operation), but differs in that MSG_DONTWAIT is a per-call
option, whereas O_NONBLOCK is a setting on the open file description
(see open(2)), which will affect all threads in the calling process
and as well as other processes that hold file descriptors referring
to the same open file description.
1.Consolidate absolute to relative timeout conversion into one place(_net_timedwait)
2.Drive the wait timeout logic by net_timedwait instead of devif_timer
This patch help us remove devif_timer(period tick) to save the power in the future.
Change-Id: I534748a5d767ca6da8a7843c3c2f993ed9ea77d4
Signed-off-by: Xiang Xiao <xiaoxiang@xiaomi.com>
Author: Gregory Nutt <gnutt@nuttx.org>
Run all .h and .c files modified in last PR through nxstyle.
Author: Xiang Xiao <xiaoxiang@xiaomi.com>
Net cleanup (#17)
* Fix the semaphore usage issue found in tcp/udp
1. The count semaphore need disable priority inheritance
2. Loop again if net_lockedwait return -EINTR
3. Call nxsem_trywait to avoid the race condition
4. Call nxsem_post instead of sem_post
* Put the work notifier into free list to avoid the heap fragment in the long run. Since the allocation strategy is encapsulated internally, we can even refine the implementation later.
* Network stack shouldn't allocate memory in the poll implementation to avoid the heap fragment in the long run, other modification include:
1. Select MM_IOB automatically since ICMP[v6] socket can't work without the read ahead buffer
2. Remove the net lock since xxx_callback_free already do the same thing
3. TCP/UDP poll should work even the read ahead buffer isn't enabled at all
* Add NET_ prefix for UDP_NOTIFIER and TCP_NOTIFIER option to align with other UDP/TCP option convention
* Remove the unused _SF_[IDLE|ACCEPT|SEND|RECV|MASK] flags since there are code to set/clear these flags, but nobody check them.
Squashed commit of the following:
net/tmp: Rename the unacked field of the tcp connection structure to tx_unacked. Too confusing with the implementation of delayed RX ACKs.
net/tcp: Initial implementation of TCP delayed ACKs.
net/tcp: Add delayed ACK configuration selection. Rename tcp_ack() to tcp_synack(). It may or may not send a ACK. It will always send SYN or SYN/ACK.
1. For buffered tcp/udp case, if CONFIG_NET_ARP_SEND/CONFIG_NET_ARP_IPIN/CONFIG_NET_ICMPv6_NEIGHBOR isn't enabled and the table doesn't contain ip<->ethaddr mapping yet, the logic will skip the realtransmission and then arp/neighbor can't steal the final buffer to generate arp/icmpv6 packet.
2.for all other case, the tcp layer or user program should already contain the retransmit logic, the check is redundancy and may generate many duplicated packets if arp/icmpv6 response is too slow because the cursor stop forward. If user still concern about the very first packet lost, he could fix the issue by enabling CONFIG_NET_ARP_SEND/CONFIG_NET_ICMPv6_NEIGHBOR at begin.
Squashed commit of the following:
net/: Fix some naming inconsistencies, Fix final compilation issies.
net/inet/inet_close(): Now that we have logic to drain the buffered TX data, we can implement a proper lingering close.
net/inet,tcp,udp: Add functions to wait for write buffers to drain.
net/udp: Add support for notification when the UDP write buffer becomes empty.
net/tcp: Add support for notification when the TCP write buffer becomes empty.
sched/wqueue/kwork_notifier.c: Redesign some data structures. struct works_s must appear at the beginning of the notifier entry structure. That is because it contains the work queue indices. This solves a harfault issue.
net/tcp/tcp_netpoll.c: tcp_iob_work() needs to free the allocated argument when it is finished.
net/tcp/tcp_send_buffered.c: Extend psock_tcp_cansend() so that it also requires that at least on IOB is also avaialble.
mm/iob: iob_navail() was returning the number of free IOB chain queue entries, not the number of free IOBs. Completely misnamed.
net/tcp/tcp_netpoll.c: Add logic to receive notifications when IOBs are freed (Needs CONFIG_NET_TCP_WRITE_BUFFERS and CONFIG_IOB_NOTIFIER). At present, does nothing because the logic in in psock_tcp_cansend() does not check for the availability of IOBs. That will change.
If there is no active connection (e.g. it is waiting in accept), then
the connection object, which doesn't yet exist, should not be cleaned
up when the socket is closed.
During a write, if there is no more buffer space for the user data,
return the amount that was written instead of waiting until there
is free space. If nothing has been written yet, then block as before.
This solves a deadlock that occurs if the user data is too large to
fit in the available buffer: the write thread will block before any
data is added to the write queue, leaving no possibility that more
buffers will free up when they are ACKed (since they have not yet been
sent). The write thread will then block forever and hold all of the
buffers.
When a poll requesting POLLOUT happens, the poll should return
immediately if a write will not block. This change adds that, as
opposed to the old behaviour of blocking until a timer from the
Ethernet driver eventually triggers the poll to complete.
This is only implemented for buffered TCP. Unbuffered TCP should
behave as before.
