At present, if the last frag of paged data in a skb has < PAGE_SIZE
data, we compute the recv_skip_hint as being equal to the size of that
frag and the entire next skb.
Instead, just return the runt frag size as the hint.
recv_skip_hint is used by the application to skip over
bytes that can not be mmaped, so returning a too big
chunk is pessimistic and forces more bytes to be copied.
Signed-off-by: Arjun Roy <arjunroy@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
FASTOPEN setsockopt() or sendmsg() may switch the SMC socket to fallback
mode. Once fallback mode is active, the native TCP socket functions are
called. Nevertheless there is a small race window, when FASTOPEN
setsockopt/sendmsg runs in parallel to a connect(), and switch the
socket into fallback mode before connect() takes the sock lock.
Make sure the SMC-specific connect setup is omitted in this case.
This way a syzbot-reported refcount problem is fixed, triggered by
different threads running non-blocking connect() and FASTOPEN_KEY
setsockopt.
Reported-by: syzbot+96d3f9ff6a86d37e44c8@syzkaller.appspotmail.com
Fixes: 6d6dd528d5 ("net/smc: fix refcount non-blocking connect() -part 2")
Signed-off-by: Ursula Braun <ubraun@linux.ibm.com>
Signed-off-by: Karsten Graul <kgraul@linux.ibm.com>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
This adds rx_bpdu, tx_bpdu, rx_tcn, tx_tcn, transition_blk,
transition_fwd xstats counters to the bridge ports copied over via
netlink, providing useful information for STP.
Signed-off-by: Vivien Didelot <vivien.didelot@gmail.com>
Acked-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
At the time commit ce5ec44099 ("tcp: ensure epoll edge trigger
wakeup when write queue is empty") was added to the kernel,
we still had a single write queue, combining rtx and write queues.
Once we moved the rtx queue into a separate rb-tree, testing
if sk_write_queue is empty has been suboptimal.
Indeed, if we have packets in the rtx queue, we probably want
to delay the EPOLLOUT generation at the time incoming packets
will free them, making room, but more importantly avoiding
flooding application with EPOLLOUT events.
Solution is to use tcp_rtx_and_write_queues_empty() helper.
Fixes: 75c119afe1 ("tcp: implement rb-tree based retransmit queue")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Jason Baron <jbaron@akamai.com>
Cc: Neal Cardwell <ncardwell@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Due to how tcp_sendmsg() is implemented, we can have an empty
skb at the tail of the write queue.
Most [1] tcp_write_queue_empty() callers want to know if there is
anything to send (payload and/or FIN)
Instead of checking if the sk_write_queue is empty, we need
to test if tp->write_seq == tp->snd_nxt
[1] tcp_send_fin() was the only caller that expected to
see if an skb was in the write queue, I have changed the code
to reuse the tcp_write_queue_tail() result.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Backport of commit fdfc5c8594 ("tcp: remove empty skb from
write queue in error cases") in linux-4.14 stable triggered
various bugs. One of them has been fixed in commit ba2ddb43f270
("tcp: Don't dequeue SYN/FIN-segments from write-queue"), but
we still have crashes in some occasions.
Root-cause is that when tcp_sendmsg() has allocated a fresh
skb and could not append a fragment before being blocked
in sk_stream_wait_memory(), tcp_write_xmit() might be called
and decide to send this fresh and empty skb.
Sending an empty packet is not only silly, it might have caused
many issues we had in the past with tp->packets_out being
out of sync.
Fixes: c65f7f00c5 ("[TCP]: Simplify SKB data portion allocation with NETIF_F_SG.")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Christoph Paasch <cpaasch@apple.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Cc: Jason Baron <jbaron@akamai.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Michal Kubecek and Firo Yang did a very nice analysis of crashes
happening in __inet_lookup_established().
Since a TCP socket can go from TCP_ESTABLISH to TCP_LISTEN
(via a close()/socket()/listen() cycle) without a RCU grace period,
I should not have changed listeners linkage in their hash table.
They must use the nulls protocol (Documentation/RCU/rculist_nulls.txt),
so that a lookup can detect a socket in a hash list was moved in
another one.
Since we added code in commit d296ba60d8 ("soreuseport: Resolve
merge conflict for v4/v6 ordering fix"), we have to add
hlist_nulls_add_tail_rcu() helper.
Fixes: 3b24d854cb ("tcp/dccp: do not touch listener sk_refcnt under synflood")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Michal Kubecek <mkubecek@suse.cz>
Reported-by: Firo Yang <firo.yang@suse.com>
Reviewed-by: Michal Kubecek <mkubecek@suse.cz>
Link: https://lore.kernel.org/netdev/20191120083919.GH27852@unicorn.suse.cz/
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
In commit 4b1373de73 ("net: ipv6: addr: perform strict checks also for
doit handlers") we add strict check for inet6_rtm_getaddr(). But we did
the invalid header values check before checking if NETLINK_F_STRICT_CHK
is set. This may break backwards compatibility if user already set the
ifm->ifa_prefixlen, ifm->ifa_flags, ifm->ifa_scope in their netlink code.
I didn't move the nlmsg_len check because I thought it's a valid check.
Reported-by: Jianlin Shi <jishi@redhat.com>
Fixes: 4b1373de73 ("net: ipv6: addr: perform strict checks also for doit handlers")
Signed-off-by: Hangbin Liu <liuhangbin@gmail.com>
Reviewed-by: David Ahern <dsahern@gmail.com>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Redirecting a packet from ingress to egress by using bpf_redirect
breaks if the egress interface has an fq qdisc installed. This is the same
problem as fixed in 'commit 8203e2d844 ("net: clear skb->tstamp in forwarding paths")
Clear skb->tstamp when redirecting into the egress path.
Fixes: 80b14dee2b ("net: Add a new socket option for a future transmit time.")
Fixes: fb420d5d91 ("tcp/fq: move back to CLOCK_MONOTONIC")
Signed-off-by: Lorenz Bauer <lmb@cloudflare.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/bpf/20191213180817.2510-1-lmb@cloudflare.com
Pull io_uring fixes from Jens Axboe:
- A tweak to IOSQE_IO_LINK (also marked for stable) to allow links that
don't sever if the result is < 0.
This is mostly for linked timeouts, where if we ask for a pure
timeout we always get -ETIME. This makes links useless for that case,
hence allow a case where it works.
- Five minor optimizations to fix and improve cases that regressed
since v5.4.
- An SQTHREAD locking fix.
- A sendmsg/recvmsg iov assignment fix.
- Net fix where read_iter/write_iter don't honor IOCB_NOWAIT, and
subsequently ensuring that works for io_uring.
- Fix a case where for an invalid opcode we might return -EBADF instead
of -EINVAL, if the ->fd of that sqe was set to an invalid fd value.
* tag 'io_uring-5.5-20191212' of git://git.kernel.dk/linux-block:
io_uring: ensure we return -EINVAL on unknown opcode
io_uring: add sockets to list of files that support non-blocking issue
net: make socket read/write_iter() honor IOCB_NOWAIT
io_uring: only hash regular files for async work execution
io_uring: run next sqe inline if possible
io_uring: don't dynamically allocate poll data
io_uring: deferred send/recvmsg should assign iov
io_uring: sqthread should grab ctx->uring_lock for submissions
io-wq: briefly spin for new work after finishing work
io-wq: remove worker->wait waitqueue
io_uring: allow unbreakable links
This patch adds a new transmit path for hardware that supports 802.11
encapsulation offloading. In those cases 802.3 frames get passed
directly to the driver allowing the hardware to handle the encapsulation.
Some features such as monitor mode and TKIP would break when encapsulation
offloading is enabled. If any of these get enabled, the code will alwyas
fallback to the normal sw encapsulation data path.
The patch defines a secondary netdev_ops struct that the device gets
assigned if 802.11 encap support is available and enabled. The driver
needs to enable the support on a per vif basis if it finds that all
pre-reqs are meet.
Signed-off-by: Vasanthakumar Thiagarajan <vthiagar@qti.qualcomm.com>
Signed-off-by: John Crispin <john@phrozen.org>
Link: https://lore.kernel.org/r/20191125100438.16539-1-john@phrozen.org
[reword comments, remove SUPPORTS_80211_ENCAP HW flag, minor cleanups]
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
The AQL statistics are displayed in the 'airtime' station debugfs file, but
that file was only shown if a driver has enabled the airtime fairness
feature flag. Since AQL can be enabled without airtime fairness, let's
expose the airtime file whenever TXQs are enabled, so the AQL data can be
read.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://lore.kernel.org/r/20191210143417.142964-1-toke@redhat.com
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Instead of just having an airtime flag in debugfs, turn AQL into a proper
NL80211_EXT_FEATURE, so drivers can turn it on when they are ready, and so
we also expose the presence of the feature to userspace.
This also has the effect of flipping the default, so drivers have to opt in
to using AQL instead of getting it by default with TXQs. To keep
functionality the same as pre-patch, we set this feature for ath10k (which
is where it is needed the most).
While we're at it, split out the debugfs interface so AQL gets its own
per-station debugfs file instead of using the 'airtime' file.
[Johannes:]
This effectively disables AQL for iwlwifi, where it fixes a number of
issues:
* TSO in iwlwifi is causing underflows and associated warnings in AQL
* HE (802.11ax) rates aren't reported properly so at HE rates, AQL could
never have a valid estimate (it'd use 6 Mbps instead of up to 2400!)
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://lore.kernel.org/r/20191212111437.224294-1-toke@redhat.com
Fixes: 3ace10f5b5 ("mac80211: Implement Airtime-based Queue Limit (AQL)")
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Fix overwriting of the qos_ctrl.tid field for encrypted frames injected on
a monitor interface. While qos_ctrl.tid is not encrypted, it's used as an
input into the encryption algorithm so it's protected, and thus cannot be
modified after encryption. For injected frames, the encryption may already
have been done in userspace, so we cannot change any fields.
Before passing the frame to the driver, the qos_ctrl.tid field is updated
from skb->priority. Prior to dbd50a851c skb->priority was updated in
ieee80211_select_queue_80211(), but this function is no longer always
called.
Update skb->priority in ieee80211_monitor_start_xmit() so that the value
is stored, and when later code 'modifies' the TID it really sets it to
the same value as before, preserving the encryption.
Fixes: dbd50a851c ("mac80211: only allocate one queue when using iTXQs")
Signed-off-by: Fredrik Olofsson <fredrik.olofsson@anyfinetworks.com>
Link: https://lore.kernel.org/r/20191119133451.14711-1-fredrik.olofsson@anyfinetworks.com
[rewrite commit message based on our discussion]
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Unlike e.g. netdev features, the ethtool ioctl interface requires link mode
table to be in sync between kernel and userspace for userspace to be able
to display and set all link modes supported by kernel. The way arbitrary
length bitsets are implemented in netlink interface, this will be no longer
needed.
To allow userspace to access all link modes running kernel supports, add
table of ethernet link mode names and make it available as a string set to
userspace GET_STRSET requests. Add build time check to make sure names
are defined for all modes declared in enum ethtool_link_mode_bit_indices.
Once the string set is available, make it also accessible via ioctl.
Signed-off-by: Michal Kubecek <mkubecek@suse.cz>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Reviewed-by: Jiri Pirko <jiri@mellanox.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Introduce file net/ethtool/common.c for code shared by ioctl and netlink
ethtool interface. Move name tables of features, RSS hash functions,
tunables and PHY tunables into this file.
Signed-off-by: Michal Kubecek <mkubecek@suse.cz>
Reviewed-by: Jiri Pirko <jiri@mellanox.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The ethtool netlink interface is going to be split into multiple files so
that it will be more convenient to put all of them in a separate directory
net/ethtool. Start by moving current ethtool.c with ioctl interface into
this directory and renaming it to ioctl.c.
Signed-off-by: Michal Kubecek <mkubecek@suse.cz>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Function nl80211_validate_nested() is not specific to nl80211, it's
a counterpart to nla_validate_nested_deprecated() with strict validation.
For consistency with other validation and parse functions, rename it to
nla_validate_nested().
Signed-off-by: Michal Kubecek <mkubecek@suse.cz>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Reviewed-by: Johannes Berg <johannes@sipsolutions.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Permanent hardware address of a network device was traditionally provided
via ethtool ioctl interface but as Jiri Pirko pointed out in a review of
ethtool netlink interface, rtnetlink is much more suitable for it so let's
add it to the RTM_NEWLINK message.
Add IFLA_PERM_ADDRESS attribute to RTM_NEWLINK messages unless the
permanent address is all zeros (i.e. device driver did not fill it). As
permanent address is not modifiable, reject userspace requests containing
IFLA_PERM_ADDRESS attribute.
Note: we already provide permanent hardware address for bond slaves;
unfortunately we cannot drop that attribute for backward compatibility
reasons.
v5 -> v6: only add the attribute if permanent address is not zero
Signed-off-by: Michal Kubecek <mkubecek@suse.cz>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Unix sockets like a block box. You never know what is stored there:
there may be a file descriptor holding a mount or a block device,
or there may be whole universes with namespaces, sockets with receive
queues full of sockets etc.
The patch adds a little debug and accounts number of files (not recursive),
which is in receive queue of a unix socket. Sometimes this is useful
to determine, that socket should be investigated or which task should
be killed to put reference counter on a resourse.
v2: Pass correct argument to lockdep
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds .show_fdinfo to socket_file_ops, so protocols will be able
to print their specific data in fdinfo.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We can remove the loopback handling from virtio_transport,
because now the vsock core is able to handle local communication
using the new vsock_loopback device.
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now that we have a transport that can handle the local communication,
we can use it when it is loaded.
A socket will use the local transport (loopback) when the remote
CID is:
- equal to VMADDR_CID_LOCAL
- or equal to transport_g2h->get_local_cid(), if transport_g2h
is loaded (this allows us to keep the same behavior implemented
by virtio and vmci transports)
- or equal to VMADDR_CID_HOST, if transport_g2h is not loaded
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds a new vsock_loopback transport to handle local
communication.
This transport is based on the loopback implementation of
virtio_transport, so it uses the virtio_transport_common APIs
to interface with the vsock core.
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch allows to register a transport able to handle
local communication (loopback).
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The VMADDR_CID_RESERVED (1) was used by VMCI, but now it is not
used anymore, so we can reuse it for local communication
(loopback) adding the new well-know CID: VMADDR_CID_LOCAL.
Cc: Jorgen Hansen <jhansen@vmware.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Jorgen Hansen <jhansen@vmware.com>
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We can remove virtio header includes, because virtio_transport_common
doesn't use virtio API, but provides common functions to interface
virtio/vhost transports with the af_vsock core, and to handle
the protocol.
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In the function 'tipc_disc_rcv()', the 'msg_peer_net_hash()' is called
to read the header data field but after the message skb has been freed,
that might result in a garbage value...
This commit fixes it by defining a new local variable to store the data
first, just like the other header fields' handling.
Fixes: f73b12812a ("tipc: improve throughput between nodes in netns")
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: Tuong Lien <tuong.t.lien@dektech.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a user message is sent, TIPC will check if the socket has faced a
congestion at link layer. If that happens, it will make a sleep to wait
for the congestion to disappear. This leaves a gap for other users to
take over the socket (e.g. multi threads) since the socket is released
as well. Also, in case of connectionless (e.g. SOCK_RDM), user is free
to send messages to various destinations (e.g. via 'sendto()'), then
the socket's preformatted header has to be updated correspondingly
prior to the actual payload message building.
Unfortunately, the latter action is done before the first action which
causes a condition issue that the destination of a certain message can
be modified incorrectly in the middle, leading to wrong destination
when that message is built. Consequently, when the message is sent to
the link layer, it gets stuck there forever because the peer node will
simply reject it. After a number of retransmission attempts, the link
is eventually taken down and the retransmission failure is reported.
This commit fixes the problem by rearranging the order of actions to
prevent the race condition from occurring, so the message building is
'atomic' and its header will not be modified by anyone.
Fixes: 365ad353c2 ("tipc: reduce risk of user starvation during link congestion")
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: Tuong Lien <tuong.t.lien@dektech.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit c55c8edafa ("tipc: smooth change between replicast and
broadcast"), we allow instant switching between replicast and broadcast
by sending a dummy 'SYN' packet on the last used link to synchronize
packets on the links. The 'SYN' message is an object of link congestion
also, so if that happens, a 'SOCK_WAKEUP' will be scheduled to be sent
back to the socket...
However, in that commit, we simply use the same socket 'cong_link_cnt'
counter for both the 'SYN' & normal payload message sending. Therefore,
if both the replicast & broadcast links are congested, the counter will
be not updated correctly but overwritten by the latter congestion.
Later on, when the 'SOCK_WAKEUP' messages are processed, the counter is
reduced one by one and eventually overflowed. Consequently, further
activities on the socket will only wait for the false congestion signal
to disappear but never been met.
Because sending the 'SYN' message is vital for the mechanism, it should
be done anyway. This commit fixes the issue by marking the message with
an error code e.g. 'TIPC_ERR_NO_PORT', so its sending should not face a
link congestion, there is no need to touch the socket 'cong_link_cnt'
either. In addition, in the event of any error (e.g. -ENOBUFS), we will
purge the entire payload message queue and make a return immediately.
Fixes: c55c8edafa ("tipc: smooth change between replicast and broadcast")
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: Tuong Lien <tuong.t.lien@dektech.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
The current rbtree for service ranges in the name table is built based
on the 'lower' & 'upper' range values resulting in a flaw in the rbtree
searching. Some issues have been observed in case of range overlapping:
Case #1: unable to withdraw a name entry:
After some name services are bound, all of them are withdrawn by user
but one remains in the name table forever. This corrupts the table and
that service becomes dummy i.e. no real port.
E.g.
/
{22, 22}
/
/
---> {10, 50}
/ \
/ \
{10, 30} {20, 60}
The node {10, 30} cannot be removed since the rbtree searching stops at
the node's ancestor i.e. {10, 50}, so starting from it will never reach
the finding node.
Case #2: failed to send data in some cases:
E.g. Two service ranges: {20, 60}, {10, 50} are bound. The rbtree for
this service will be one of the two cases below depending on the order
of the bindings:
{20, 60} {10, 50} <--
/ \ / \
/ \ / \
{10, 50} NIL <-- NIL {20, 60}
(a) (b)
Now, try to send some data to service {30}, there will be two results:
(a): Failed, no route to host.
(b): Ok.
The reason is that the rbtree searching will stop at the pointing node
as shown above.
Case #3: Same as case #2b above but if the data sending's scope is
local and the {10, 50} is published by a peer node, then it will result
in 'no route to host' even though the other {20, 60} is for example on
the local node which should be able to get the data.
The issues are actually due to the way we built the rbtree. This commit
fixes it by introducing an additional field to each node - named 'max',
which is the largest 'upper' of that node subtree. The 'max' value for
each subtrees will be propagated correctly whenever a node is inserted/
removed or the tree is rebalanced by the augmented rbtree callbacks.
By this way, we can change the rbtree searching appoarch to solve the
issues above. Another benefit from this is that we can now improve the
searching for a next range matching e.g. in case of multicast, so get
rid of the unneeded looping over all nodes in the tree.
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: Tuong Lien <tuong.t.lien@dektech.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a TCP socket is created, sk->sk_state is initialized twice as
TCP_CLOSE in sock_init_data() and tcp_init_sock(). The tcp_init_sock() is
always called after the sock_init_data(), so it is not necessary to update
sk->sk_state in the tcp_init_sock().
Before v2.1.8, the code of the two functions was in the inet_create(). In
the patch of v2.1.8, the tcp_v4/v6_init_sock() were added and the code of
initialization of sk->state was duplicated.
Signed-off-by: Kuniyuki Iwashima <kuni1840@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We introduce a simple variable window congestion control for links.
The algorithm is inspired by the Reno algorithm, covering both 'slow
start', 'congestion avoidance', and 'fast recovery' modes.
- We introduce hard lower and upper window limits per link, still
different and configurable per bearer type.
- We introduce a 'slow start theshold' variable, initially set to
the maximum window size.
- We let a link start at the minimum congestion window, i.e. in slow
start mode, and then let is grow rapidly (+1 per rceived ACK) until
it reaches the slow start threshold and enters congestion avoidance
mode.
- In congestion avoidance mode we increment the congestion window for
each window-size number of acked packets, up to a possible maximum
equal to the configured maximum window.
- For each non-duplicate NACK received, we drop back to fast recovery
mode, by setting the both the slow start threshold to and the
congestion window to (current_congestion_window / 2).
- If the timeout handler finds that the transmit queue has not moved
since the previous timeout, it drops the link back to slow start
and forces a probe containing the last sent sequence number to the
sent to the peer, so that this can discover the stale situation.
This change does in reality have effect only on unicast ethernet
transport, as we have seen that there is no room whatsoever for
increasing the window max size for the UDP bearer.
For now, we also choose to keep the limits for the broadcast link
unchanged and equal.
This algorithm seems to give a 50-100% throughput improvement for
messages larger than MTU.
Suggested-by: Xin Long <lucien.xin@gmail.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When we increase the link tranmsit window we often observe the following
scenario:
1) A STATE message bypasses a sequence of traffic packets and arrives
far ahead of those to the receiver. STATE messages contain a
'peers_nxt_snt' field to indicate which was the last packet sent
from the peer. This mechanism is intended as a last resort for the
receiver to detect missing packets, e.g., during very low traffic
when there is no packet flow to help early loss detection.
3) The receiving link compares the 'peer_nxt_snt' field to its own
'rcv_nxt', finds that there is a gap, and immediately sends a
NACK message back to the peer.
4) When this NACKs arrives at the sender, all the requested
retransmissions are performed, since it is a first-time request.
Just like in the scenario described in the previous commit this leads
to many redundant retransmissions, with decreased throughput as a
consequence.
We fix this by adding two more conditions before we send a NACK in
this sitution. First, the deferred queue must be empty, so we cannot
assume that the potential packet loss has already been detected by
other means. Second, we check the 'peers_snd_nxt' field only in probe/
probe_reply messages, thus turning this into a true mechanism of last
resort as it was really meant to be.
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When we increase the link send window we sometimes observe the
following scenario:
1) A packet #N arrives out of order far ahead of a sequence of older
packets which are still under way. The packet is added to the
deferred queue.
2) The missing packets arrive in sequence, and for each 16th of them
an ACK is sent back to the receiver, as it should be.
3) When building those ACK messages, it is checked if there is a gap
between the link's 'rcv_nxt' and the first packet in the deferred
queue. This is always the case until packet number #N-1 arrives, and
a 'gap' indicator is added, effectively turning them into NACK
messages.
4) When those NACKs arrive at the sender, all the requested
retransmissions are done, since it is a first-time request.
This sometimes leads to a huge amount of redundant retransmissions,
causing a drop in max throughput. This problem gets worse when we
in a later commit introduce variable window congestion control,
since it drops the link back to 'fast recovery' much more often
than necessary.
We now fix this by not sending any 'gap' indicator in regular ACK
messages. We already have a mechanism for sending explicit NACKs
in place, and this is sufficient to keep up the packet flow.
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The socket read/write helpers only look at the file O_NONBLOCK. not
the iocb IOCB_NOWAIT flag. This breaks users like preadv2/pwritev2
and io_uring that rely on not having the file itself marked nonblocking,
but rather the iocb itself.
Cc: netdev@vger.kernel.org
Acked-by: David Miller <davem@davemloft.net>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This patch introduces a sysctl knob "net.ipv4.tcp_no_ssthresh_metrics_save"
that disables TCP ssthresh metrics cache by default. Other parts of TCP
metrics cache, e.g. rtt, cwnd, remain unchanged.
As modern networks becoming more and more dynamic, TCP metrics cache
today often causes more harm than benefits. For example, the same IP
address is often shared by different subscribers behind NAT in residential
networks. Even if the IP address is not shared by different users,
caching the slow-start threshold of a previous short flow using loss-based
congestion control (e.g. cubic) often causes the future longer flows of
the same network path to exit slow-start prematurely with abysmal
throughput.
Caching ssthresh is very risky and can lead to terrible performance.
Therefore it makes sense to make disabling ssthresh caching by
default and opt-in for specific networks by the administrators.
This practice also has worked well for several years of deployment with
CUBIC congestion control at Google.
Acked-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Kevin(Yudong) Yang <yyd@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 312434617c ("sctp: cache netns in sctp_ep_common") set netns
in asoc and ep base since they're created, and it will never change.
It's a better way to get netns from asoc and ep base, comparing to
calling sock_net().
This patch is to replace them.
v1->v2:
- no change.
Suggested-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Xin Long <lucien.xin@gmail.com>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Acked-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is softlockup when using TPACKET_V3:
...
NMI watchdog: BUG: soft lockup - CPU#2 stuck for 60010ms!
(__irq_svc) from [<c0558a0c>] (_raw_spin_unlock_irqrestore+0x44/0x54)
(_raw_spin_unlock_irqrestore) from [<c027b7e8>] (mod_timer+0x210/0x25c)
(mod_timer) from [<c0549c30>]
(prb_retire_rx_blk_timer_expired+0x68/0x11c)
(prb_retire_rx_blk_timer_expired) from [<c027a7ac>]
(call_timer_fn+0x90/0x17c)
(call_timer_fn) from [<c027ab6c>] (run_timer_softirq+0x2d4/0x2fc)
(run_timer_softirq) from [<c021eaf4>] (__do_softirq+0x218/0x318)
(__do_softirq) from [<c021eea0>] (irq_exit+0x88/0xac)
(irq_exit) from [<c0240130>] (msa_irq_exit+0x11c/0x1d4)
(msa_irq_exit) from [<c0209cf0>] (handle_IPI+0x650/0x7f4)
(handle_IPI) from [<c02015bc>] (gic_handle_irq+0x108/0x118)
(gic_handle_irq) from [<c0558ee4>] (__irq_usr+0x44/0x5c)
...
If __ethtool_get_link_ksettings() is failed in
prb_calc_retire_blk_tmo(), msec and tmo will be zero, so tov_in_jiffies
is zero and the timer expire for retire_blk_timer is turn to
mod_timer(&pkc->retire_blk_timer, jiffies + 0),
which will trigger cpu usage of softirq is 100%.
Fixes: f6fb8f100b ("af-packet: TPACKET_V3 flexible buffer implementation.")
Tested-by: Xiao Jiangfeng <xiaojiangfeng@huawei.com>
Signed-off-by: Mao Wenan <maowenan@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pablo Neira Ayuso says:
====================
Netfilter fixes for net
The following patchset contains Netfilter fixes for net:
1) Wait for rcu grace period after releasing netns in ctnetlink,
from Florian Westphal.
2) Incorrect command type in flowtable offload ndo invocation,
from wenxu.
3) Incorrect callback type in flowtable offload flow tuple
updates, also from wenxu.
4) Fix compile warning on flowtable offload infrastructure due to
possible reference to uninitialized variable, from Nathan Chancellor.
5) Do not inline nf_ct_resolve_clash(), this is called from slow
path / stress situations. From Florian Westphal.
6) Missing IPv6 flow selector description in flowtable offload.
7) Missing check for NETDEV_UNREGISTER in nf_tables offload
infrastructure, from wenxu.
8) Update NAT selftest to use randomized netns names, from
Florian Westphal.
9) Restore nfqueue bridge support, from Marco Oliverio.
10) Compilation warning in SCTP_CHUNKMAP_*() on xt_sctp header.
From Phil Sutter.
11) Fix bogus lookup/get match for non-anonymous rbtree sets.
12) Missing netlink validation for NFT_SET_ELEM_INTERVAL_END
elements.
13) Missing netlink validation for NFT_DATA_VALUE after
nft_data_init().
14) If rule specifies no actions, offload infrastructure returns
EOPNOTSUPP.
15) Module refcount leak in object updates.
16) Missing sanitization for ARP traffic from br_netfilter, from
Eric Dumazet.
17) Compilation breakage on big-endian due to incorrect memcpy()
size in the flowtable offload infrastructure.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
