Aggregation effects are extremely common with wifi, cellular, and cable
modem link technologies, ACK decimation in middleboxes, and LRO and GRO
in receiving hosts. The aggregation can happen in either direction,
data or ACKs, but in either case the aggregation effect is visible
to the sender in the ACK stream.
Previously BBR's sending was often limited by cwnd under severe ACK
aggregation/decimation because BBR sized the cwnd at 2*BDP. If packets
were acked in bursts after long delays (e.g. one ACK acking 5*BDP after
5*RTT), BBR's sending was halted after sending 2*BDP over 2*RTT, leaving
the bottleneck idle for potentially long periods. Note that loss-based
congestion control does not have this issue because when facing
aggregation it continues increasing cwnd after bursts of ACKs, growing
cwnd until the buffer is full.
To achieve good throughput in the presence of aggregation effects, this
algorithm allows the BBR sender to put extra data in flight to keep the
bottleneck utilized during silences in the ACK stream that it has evidence
to suggest were caused by aggregation.
A summary of the algorithm: when a burst of packets are acked by a
stretched ACK or a burst of ACKs or both, BBR first estimates the expected
amount of data that should have been acked, based on its estimated
bandwidth. Then the surplus ("extra_acked") is recorded in a windowed-max
filter to estimate the recent level of observed ACK aggregation. Then cwnd
is increased by the ACK aggregation estimate. The larger cwnd avoids BBR
being cwnd-limited in the face of ACK silences that recent history suggests
were caused by aggregation. As a sanity check, the ACK aggregation degree
is upper-bounded by the cwnd (at the time of measurement) and a global max
of BW * 100ms. The algorithm is further described by the following
presentation:
https://datatracker.ietf.org/meeting/101/materials/slides-101-iccrg-an-update-on-bbr-work-at-google-00
In our internal testing, we observed a significant increase in BBR
throughput (measured using netperf), in a basic wifi setup.
- Host1 (sender on ethernet) -> AP -> Host2 (receiver on wifi)
- 2.4 GHz -> BBR before: ~73 Mbps; BBR after: ~102 Mbps; CUBIC: ~100 Mbps
- 5.0 GHz -> BBR before: ~362 Mbps; BBR after: ~593 Mbps; CUBIC: ~601 Mbps
Also, this code is running globally on YouTube TCP connections and produced
significant bandwidth increases for YouTube traffic.
This is based on Ian Swett's max_ack_height_ algorithm from the
QUIC BBR implementation.
Signed-off-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As part of the effort to reduce the code duplication between _THIS_IP_
and current_text_addr(), let's consolidate callers of
current_text_addr() to use _THIS_IP_.
Signed-off-by: Nick Desaulniers <ndesaulniers@google.com>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a new flag to indicate a one-time immediate ACK. This flag is
occasionaly set under specific TCP protocol states in addition to
the more common quickack mechanism for interactive application.
In several cases in the TCP code we want to force an immediate ACK
but do not want to call tcp_enter_quickack_mode() because we do
not want to forget the icsk_ack.pingpong or icsk_ack.ato state.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Wei Wang <weiwan@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
INET_CSK_DEBUG is always set and only is used for 2 pr_debug calls.
EXPORT_SYMBOL(inet_csk_timer_bug_msg) is only used by these 2
pr_debug calls and is also unnecessary as the exported string can
be used directly by these calls.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Called when a TCP segment is acknowledged.
Could be used by application protocols who hold additional
metadata associated with the stream data.
This is required by TLS device offload to release
metadata associated with acknowledged TLS records.
Signed-off-by: Ilya Lesokhin <ilyal@mellanox.com>
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: Aviad Yehezkel <aviadye@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Ran simple script to find/remove trailing whitespace and blank lines
at EOF because that kind of stuff git whines about and editors leave
behind.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is the mindless scripted replacement of kernel use of POLL*
variables as described by Al, done by this script:
for V in IN OUT PRI ERR RDNORM RDBAND WRNORM WRBAND HUP RDHUP NVAL MSG; do
L=`git grep -l -w POLL$V | grep -v '^t' | grep -v /um/ | grep -v '^sa' | grep -v '/poll.h$'|grep -v '^D'`
for f in $L; do sed -i "-es/^\([^\"]*\)\(\<POLL$V\>\)/\\1E\\2/" $f; done
done
with de-mangling cleanups yet to come.
NOTE! On almost all architectures, the EPOLL* constants have the same
values as the POLL* constants do. But they keyword here is "almost".
For various bad reasons they aren't the same, and epoll() doesn't
actually work quite correctly in some cases due to this on Sparc et al.
The next patch from Al will sort out the final differences, and we
should be all done.
Scripted-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull networking updates from David Miller:
1) Significantly shrink the core networking routing structures. Result
of http://vger.kernel.org/~davem/seoul2017_netdev_keynote.pdf
2) Add netdevsim driver for testing various offloads, from Jakub
Kicinski.
3) Support cross-chip FDB operations in DSA, from Vivien Didelot.
4) Add a 2nd listener hash table for TCP, similar to what was done for
UDP. From Martin KaFai Lau.
5) Add eBPF based queue selection to tun, from Jason Wang.
6) Lockless qdisc support, from John Fastabend.
7) SCTP stream interleave support, from Xin Long.
8) Smoother TCP receive autotuning, from Eric Dumazet.
9) Lots of erspan tunneling enhancements, from William Tu.
10) Add true function call support to BPF, from Alexei Starovoitov.
11) Add explicit support for GRO HW offloading, from Michael Chan.
12) Support extack generation in more netlink subsystems. From Alexander
Aring, Quentin Monnet, and Jakub Kicinski.
13) Add 1000BaseX, flow control, and EEE support to mvneta driver. From
Russell King.
14) Add flow table abstraction to netfilter, from Pablo Neira Ayuso.
15) Many improvements and simplifications to the NFP driver bpf JIT,
from Jakub Kicinski.
16) Support for ipv6 non-equal cost multipath routing, from Ido
Schimmel.
17) Add resource abstration to devlink, from Arkadi Sharshevsky.
18) Packet scheduler classifier shared filter block support, from Jiri
Pirko.
19) Avoid locking in act_csum, from Davide Caratti.
20) devinet_ioctl() simplifications from Al viro.
21) More TCP bpf improvements from Lawrence Brakmo.
22) Add support for onlink ipv6 route flag, similar to ipv4, from David
Ahern.
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next: (1925 commits)
tls: Add support for encryption using async offload accelerator
ip6mr: fix stale iterator
net/sched: kconfig: Remove blank help texts
openvswitch: meter: Use 64-bit arithmetic instead of 32-bit
tcp_nv: fix potential integer overflow in tcpnv_acked
r8169: fix RTL8168EP take too long to complete driver initialization.
qmi_wwan: Add support for Quectel EP06
rtnetlink: enable IFLA_IF_NETNSID for RTM_NEWLINK
ipmr: Fix ptrdiff_t print formatting
ibmvnic: Wait for device response when changing MAC
qlcnic: fix deadlock bug
tcp: release sk_frag.page in tcp_disconnect
ipv4: Get the address of interface correctly.
net_sched: gen_estimator: fix lockdep splat
net: macb: Handle HRESP error
net/mlx5e: IPoIB, Fix copy-paste bug in flow steering refactoring
ipv6: addrconf: break critical section in addrconf_verify_rtnl()
ipv6: change route cache aging logic
i40e/i40evf: Update DESC_NEEDED value to reflect larger value
bnxt_en: cleanup DIM work on device shutdown
...
The current listener hashtable is hashed by port only.
When a process is listening at many IP addresses with the same port (e.g.
[IP1]:443, [IP2]:443... [IPN]:443), the inet[6]_lookup_listener()
performance is degraded to a link list. It is prone to syn attack.
UDP had a similar issue and a second hashtable was added to resolve it.
This patch adds a second hashtable for the listener's sockets.
The second hashtable is hashed by port and address.
It cannot reuse the existing skc_portaddr_node which is shared
with skc_bind_node. TCP listener needs to use skc_bind_node.
Instead, this patch adds a hlist_node 'icsk_listen_portaddr_node' to
the inet_connection_sock which the listener (like TCP) also belongs to.
The new portaddr hashtable may need two lookup (First by IP:PORT.
Second by INADDR_ANY:PORT if the IP:PORT is a not found). Hence,
it implements a similar cut off as UDP such that it will only consult the
new portaddr hashtable if the current port-only hashtable has >10
sk in the link-list.
lhash2 and lhash2_mask are added to 'struct inet_hashinfo'. I take
this chance to plug a 4 bytes hole. It is done by first moving
the existing bind_bucket_cachep up and then add the new
(int lhash2_mask, *lhash2) after the existing bhash_size.
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In preparation for unconditionally passing the struct timer_list pointer to
all timer callbacks, switch to using the new timer_setup() and from_timer()
to pass the timer pointer explicitly.
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Cc: Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: dccp@vger.kernel.org
Signed-off-by: Kees Cook <keescook@chromium.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add the infrustructure for attaching Upper Layer Protocols (ULPs) over TCP
sockets. Based on a similar infrastructure in tcp_cong. The idea is that any
ULP can add its own logic by changing the TCP proto_ops structure to its own
methods.
Example usage:
setsockopt(sock, SOL_TCP, TCP_ULP, "tls", sizeof("tls"));
modules will call:
tcp_register_ulp(&tcp_tls_ulp_ops);
to register/unregister their ulp, with an init function and name.
A list of registered ulps will be returned by tcp_get_available_ulp, which is
hooked up to /proc. Example:
$ cat /proc/sys/net/ipv4/tcp_available_ulp
tls
There is currently no functionality to remove or chain ULPs, but
it should be possible to add these in the future if needed.
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Lockdep issues a circular dependency warning when AFS issues an operation
through AF_RXRPC from a context in which the VFS/VM holds the mmap_sem.
The theory lockdep comes up with is as follows:
(1) If the pagefault handler decides it needs to read pages from AFS, it
calls AFS with mmap_sem held and AFS begins an AF_RXRPC call, but
creating a call requires the socket lock:
mmap_sem must be taken before sk_lock-AF_RXRPC
(2) afs_open_socket() opens an AF_RXRPC socket and binds it. rxrpc_bind()
binds the underlying UDP socket whilst holding its socket lock.
inet_bind() takes its own socket lock:
sk_lock-AF_RXRPC must be taken before sk_lock-AF_INET
(3) Reading from a TCP socket into a userspace buffer might cause a fault
and thus cause the kernel to take the mmap_sem, but the TCP socket is
locked whilst doing this:
sk_lock-AF_INET must be taken before mmap_sem
However, lockdep's theory is wrong in this instance because it deals only
with lock classes and not individual locks. The AF_INET lock in (2) isn't
really equivalent to the AF_INET lock in (3) as the former deals with a
socket entirely internal to the kernel that never sees userspace. This is
a limitation in the design of lockdep.
Fix the general case by:
(1) Double up all the locking keys used in sockets so that one set are
used if the socket is created by userspace and the other set is used
if the socket is created by the kernel.
(2) Store the kern parameter passed to sk_alloc() in a variable in the
sock struct (sk_kern_sock). This informs sock_lock_init(),
sock_init_data() and sk_clone_lock() as to the lock keys to be used.
Note that the child created by sk_clone_lock() inherits the parent's
kern setting.
(3) Add a 'kern' parameter to ->accept() that is analogous to the one
passed in to ->create() that distinguishes whether kernel_accept() or
sys_accept4() was the caller and can be passed to sk_alloc().
Note that a lot of accept functions merely dequeue an already
allocated socket. I haven't touched these as the new socket already
exists before we get the parameter.
Note also that there are a couple of places where I've made the accepted
socket unconditionally kernel-based:
irda_accept()
rds_rcp_accept_one()
tcp_accept_from_sock()
because they follow a sock_create_kern() and accept off of that.
Whilst creating this, I noticed that lustre and ocfs don't create sockets
through sock_create_kern() and thus they aren't marked as for-kernel,
though they appear to be internal. I wonder if these should do that so
that they use the new set of lock keys.
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The only difference between inet6_csk_bind_conflict and inet_csk_bind_conflict
is how they check the rcv_saddr, so delete this call back and simply
change inet_csk_bind_conflict to call inet_rcv_saddr_equal.
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch makes RACK install a reordering timer when it suspects
some packets might be lost, but wants to delay the decision
a little bit to accomodate reordering.
It does not create a new timer but instead repurposes the existing
RTO timer, because both are meant to retransmit packets.
Specifically it arms a timer ICSK_TIME_REO_TIMEOUT when
the RACK timing check fails. The wait time is set to
RACK.RTT + RACK.reo_wnd - (NOW - Packet.xmit_time) + fudge
This translates to expecting a packet (Packet) should take
(RACK.RTT + RACK.reo_wnd + fudge) to deliver after it was sent.
When there are multiple packets that need a timer, we use one timer
with the maximum timeout. Therefore the timer conservatively uses
the maximum window to expire N packets by one timeout, instead of
N timeouts to expire N packets sent at different times.
The fudge factor is 2 jiffies to ensure when the timer fires, all
the suspected packets would exceed the deadline and be marked lost
by tcp_rack_detect_loss(). It has to be at least 1 jiffy because the
clock may tick between calling icsk_reset_xmit_timer(timeout) and
actually hang the timer. The next jiffy is to lower-bound the timeout
to 2 jiffies when reo_wnd is < 1ms.
When the reordering timer fires (tcp_rack_reo_timeout): If we aren't
in Recovery we'll enter fast recovery and force fast retransmit.
This is very similar to the early retransmit (RFC5827) except RACK
is not constrained to only enter recovery for small outstanding
flights.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
A user may call listen with binding an explicit port with the intent
that the kernel will assign an available port to the socket. In this
case inet_csk_get_port does a port scan. For such sockets, the user may
also set soreuseport with the intent a creating more sockets for the
port that is selected. The problem is that the initial socket being
opened could inadvertently choose an existing and unreleated port
number that was already created with soreuseport.
This patch adds a boolean parameter to inet_bind_conflict that indicates
rather soreuseport is allowed for the check (in addition to
sk->sk_reuseport). In calls to inet_bind_conflict from inet_csk_get_port
the argument is set to true if an explicit port is being looked up (snum
argument is nonzero), and is false if port scan is done.
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Per listen(fd, backlog) rules, there is really no point accepting a SYN,
sending a SYNACK, and dropping the following ACK packet if accept queue
is full, because application is not draining accept queue fast enough.
This behavior is fooling TCP clients that believe they established a
flow, while there is nothing at server side. They might then send about
10 MSS (if using IW10) that will be dropped anyway while server is under
stress.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Ilya reported following lockdep splat:
kernel: =========================
kernel: [ BUG: held lock freed! ]
kernel: 4.5.0-rc1-ceph-00026-g5e0a311 #1 Not tainted
kernel: -------------------------
kernel: swapper/5/0 is freeing memory
ffff880035c9d200-ffff880035c9dbff, with a lock still held there!
kernel: (&(&queue->rskq_lock)->rlock){+.-...}, at:
[<ffffffff816f6a88>] inet_csk_reqsk_queue_add+0x28/0xa0
kernel: 4 locks held by swapper/5/0:
kernel: #0: (rcu_read_lock){......}, at: [<ffffffff8169ef6b>]
netif_receive_skb_internal+0x4b/0x1f0
kernel: #1: (rcu_read_lock){......}, at: [<ffffffff816e977f>]
ip_local_deliver_finish+0x3f/0x380
kernel: #2: (slock-AF_INET){+.-...}, at: [<ffffffff81685ffb>]
sk_clone_lock+0x19b/0x440
kernel: #3: (&(&queue->rskq_lock)->rlock){+.-...}, at:
[<ffffffff816f6a88>] inet_csk_reqsk_queue_add+0x28/0xa0
To properly fix this issue, inet_csk_reqsk_queue_add() needs
to return to its callers if the child as been queued
into accept queue.
We also need to make sure listener is still there before
calling sk->sk_data_ready(), by holding a reference on it,
since the reference carried by the child can disappear as
soon as the child is put on accept queue.
Reported-by: Ilya Dryomov <idryomov@gmail.com>
Fixes: ebb516af60 ("tcp/dccp: fix race at listener dismantle phase")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Multiple cpus can process duplicates of incoming ACK messages
matching a SYN_RECV request socket. This is a rare event under
normal operations, but definitely can happen.
Only one must win the race, otherwise corruption would occur.
To fix this without adding new atomic ops, we use logic in
inet_ehash_nolisten() to detect the request was present in the same
ehash bucket where we try to insert the new child.
If request socket was not found, we have to undo the child creation.
This actually removes a spin_lock()/spin_unlock() pair in
reqsk_queue_unlink() for the fast path.
Fixes: e994b2f0fb ("tcp: do not lock listener to process SYN packets")
Fixes: 079096f103 ("tcp/dccp: install syn_recv requests into ehash table")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Under stress, a close() on a listener can trigger the
WARN_ON(sk->sk_ack_backlog) in inet_csk_listen_stop()
We need to test if listener is still active before queueing
a child in inet_csk_reqsk_queue_add()
Create a common inet_child_forget() helper, and use it
from inet_csk_reqsk_queue_add() and inet_csk_listen_stop()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Let's reduce the confusion about inet_csk_reqsk_queue_drop() :
In many cases we also need to release reference on request socket,
so add a helper to do this, reducing code size and complexity.
Fixes: 4bdc3d6614 ("tcp/dccp: fix behavior of stale SYN_RECV request sockets")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
At listen() time, there is a small window where listener is visible with
a zero backlog, triggering a spurious "Possible SYN flooding on port"
message.
Nothing prevents us from setting the correct backlog.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This control variable was set at first listen(fd, backlog)
call, but not updated if application tried to increase or decrease
backlog. It made sense at the time listener had a non resizeable
hash table.
Also rounding to powers of two was not very friendly.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In this patch, we insert request sockets into TCP/DCCP
regular ehash table (where ESTABLISHED and TIMEWAIT sockets
are) instead of using the per listener hash table.
ACK packets find SYN_RECV pseudo sockets without having
to find and lock the listener.
In nominal conditions, this halves pressure on listener lock.
Note that this will allow for SO_REUSEPORT refinements,
so that we can select a listener using cpu/numa affinities instead
of the prior 'consistent hash', since only SYN packets will
apply this selection logic.
We will shrink listen_sock in the following patch to ease
code review.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ying Cai <ycai@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We'll soon no longer hold listener socket lock, these
functions do not modify the socket in any way.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is used by TCP listener core, and listener socket shall
not be modified by inet_csk_route_req().
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Linux 3.17 and earlier are explicitly engineered so that if the app
doesn't specifically request a CC module on a listener before the SYN
arrives, then the child gets the system default CC when the connection
is established. See tcp_init_congestion_control() in 3.17 or earlier,
which says "if no choice made yet assign the current value set as
default". The change ("net: tcp: assign tcp cong_ops when tcp sk is
created") altered these semantics, so that children got their parent
listener's congestion control even if the system default had changed
after the listener was created.
This commit returns to those original semantics from 3.17 and earlier,
since they are the original semantics from 2007 in 4d4d3d1e8 ("[TCP]:
Congestion control initialization."), and some Linux congestion
control workflows depend on that.
In summary, if a listener socket specifically sets TCP_CONGESTION to
"x", or the route locks the CC module to "x", then the child gets
"x". Otherwise the child gets current system default from
net.ipv4.tcp_congestion_control. That's the behavior in 3.17 and
earlier, and this commit restores that.
Fixes: 55d8694fa8 ("net: tcp: assign tcp cong_ops when tcp sk is created")
Cc: Florian Westphal <fw@strlen.de>
Cc: Daniel Borkmann <dborkman@redhat.com>
Cc: Glenn Judd <glenn.judd@morganstanley.com>
Cc: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_illinois and upcoming tcp_cdg require 64bit alignment of
icsk_ca_priv
x86 does not care, but other architectures might.
Fixes: 05cbc0db03 ("ipv4: Create probe timer for tcp PMTU as per RFC4821")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Fan Du <fan.du@intel.com>
Acked-by: Fan Du <fan.du@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
[ 3897.923145] BUG: unable to handle kernel NULL pointer dereference at
0000000000000080
[ 3897.931025] IP: [<ffffffffa9f27686>] reqsk_timer_handler+0x1a6/0x243
There is a race when reqsk_timer_handler() and tcp_check_req() call
inet_csk_reqsk_queue_unlink() on the same req at the same time.
Before commit fa76ce7328 ("inet: get rid of central tcp/dccp listener
timer"), listener spinlock was held and race could not happen.
To solve this bug, we change reqsk_queue_unlink() to not assume req
must be found, and we return a status, to conditionally release a
refcount on the request sock.
This also means tcp_check_req() in non fastopen case might or not
consume req refcount, so tcp_v6_hnd_req() & tcp_v4_hnd_req() have
to properly handle this.
(Same remark for dccp_check_req() and its callers)
inet_csk_reqsk_queue_drop() is now too big to be inlined, as it is
called 4 times in tcp and 3 times in dccp.
Fixes: fa76ce7328 ("inet: get rid of central tcp/dccp listener timer")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When request sock are put in ehash table, the whole notion
of having a previous request to update dl_next is pointless.
Also, following patch will get rid of big purge timer,
so we want to delete a request sock without holding listener lock.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While testing last patch series, I found req sock refcounting was wrong.
We must set skc_refcnt to 1 for all request socks added in hashes,
but also on request sockets created by FastOpen or syncookies.
It is tricky because we need to defer this initialization so that
future RCU lookups do not try to take a refcount on a not yet
fully initialized request socket.
Also get rid of ireq_refcnt alias.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Fixes: 13854e5a60 ("inet: add proper refcounting to request sock")
Signed-off-by: David S. Miller <davem@davemloft.net>
reqsk_put() is the generic function that should be used
to release a refcount (and automatically call reqsk_free())
reqsk_free() might be called if refcount is known to be 0
or undefined.
refcnt is set to one in inet_csk_reqsk_queue_add()
As request socks are not yet in global ehash table,
I added temporary debugging checks in reqsk_put() and reqsk_free()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As per RFC4821 7.3. Selecting Probe Size, a probe timer should
be armed once probing has converged. Once this timer expired,
probing again to take advantage of any path PMTU change. The
recommended probing interval is 10 minutes per RFC1981. Probing
interval could be sysctled by sysctl_tcp_probe_interval.
Eric Dumazet suggested to implement pseudo timer based on 32bits
jiffies tcp_time_stamp instead of using classic timer for such
rare event.
Signed-off-by: Fan Du <fan.du@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds necessary infrastructure to the congestion control
framework for later per route congestion control support.
For a per route congestion control possibility, our aim is to store
a unique u32 key identifier into dst metrics, which can then be
mapped into a tcp_congestion_ops struct. We argue that having a
RTAX key entry is the most simple, generic and easy way to manage,
and also keeps the memory footprint of dst entries lower on 64 bit
than with storing a pointer directly, for example. Having a unique
key id also allows for decoupling actual TCP congestion control
module management from the FIB layer, i.e. we don't have to care
about expensive module refcounting inside the FIB at this point.
We first thought of using an IDR store for the realization, which
takes over dynamic assignment of unused key space and also performs
the key to pointer mapping in RCU. While doing so, we stumbled upon
the issue that due to the nature of dynamic key distribution, it
just so happens, arguably in very rare occasions, that excessive
module loads and unloads can lead to a possible reuse of previously
used key space. Thus, previously stale keys in the dst metric are
now being reassigned to a different congestion control algorithm,
which might lead to unexpected behaviour. One way to resolve this
would have been to walk FIBs on the actually rare occasion of a
module unload and reset the metric keys for each FIB in each netns,
but that's just very costly.
Therefore, we argue a better solution is to reuse the unique
congestion control algorithm name member and map that into u32 key
space through jhash. For that, we split the flags attribute (as it
currently uses 2 bits only anyway) into two u32 attributes, flags
and key, so that we can keep the cacheline boundary of 2 cachelines
on x86_64 and cache the precalculated key at registration time for
the fast path. On average we might expect 2 - 4 modules being loaded
worst case perhaps 15, so a key collision possibility is extremely
low, and guaranteed collision-free on LE/BE for all in-tree modules.
Overall this results in much simpler code, and all without the
overhead of an IDR. Due to the deterministic nature, modules can
now be unloaded, the congestion control algorithm for a specific
but unloaded key will fall back to the default one, and on module
reload time it will switch back to the expected algorithm
transparently.
Joint work with Florian Westphal.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
icsk_rto is a 32bit field, and icsk_backoff can reach 15 by default,
or more if some sysctl (eg tcp_retries2) are changed.
Better use 64bit to perform icsk_rto << icsk_backoff operations
As Joe Perches suggested, add a helper for this.
Yuchung spotted the tcp_v4_err() case.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Make sure we use the correct address-family-specific function for
handling MTU reductions from within tcp_release_cb().
Previously AF_INET6 sockets were incorrectly always using the IPv6
code path when sometimes they were handling IPv4 traffic and thus had
an IPv4 dst.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Diagnosed-by: Willem de Bruijn <willemb@google.com>
Fixes: 563d34d057 ("tcp: dont drop MTU reduction indications")
Reviewed-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
ip_queue_xmit() assumes the skb it has to transmit is attached to an
inet socket. Commit 31c70d5956 ("l2tp: keep original skb ownership")
changed l2tp to not change skb ownership and thus broke this assumption.
One fix is to add a new 'struct sock *sk' parameter to ip_queue_xmit(),
so that we do not assume skb->sk points to the socket used by l2tp
tunnel.
Fixes: 31c70d5956 ("l2tp: keep original skb ownership")
Reported-by: Zhan Jianyu <nasa4836@gmail.com>
Tested-by: Zhan Jianyu <nasa4836@gmail.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are a mix of function prototypes with and without extern
in the kernel sources. Standardize on not using extern for
function prototypes.
Function prototypes don't need to be written with extern.
extern is assumed by the compiler. Its use is as unnecessary as
using auto to declare automatic/local variables in a block.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch series implement the Tail loss probe (TLP) algorithm described
in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The
first patch implements the basic algorithm.
TLP's goal is to reduce tail latency of short transactions. It achieves
this by converting retransmission timeouts (RTOs) occuring due
to tail losses (losses at end of transactions) into fast recovery.
TLP transmits one packet in two round-trips when a connection is in
Open state and isn't receiving any ACKs. The transmitted packet, aka
loss probe, can be either new or a retransmission. When there is tail
loss, the ACK from a loss probe triggers FACK/early-retransmit based
fast recovery, thus avoiding a costly RTO. In the absence of loss,
there is no change in the connection state.
PTO stands for probe timeout. It is a timer event indicating
that an ACK is overdue and triggers a loss probe packet. The PTO value
is set to max(2*SRTT, 10ms) and is adjusted to account for delayed
ACK timer when there is only one oustanding packet.
TLP Algorithm
On transmission of new data in Open state:
-> packets_out > 1: schedule PTO in max(2*SRTT, 10ms).
-> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms)
-> PTO = min(PTO, RTO)
Conditions for scheduling PTO:
-> Connection is in Open state.
-> Connection is either cwnd limited or no new data to send.
-> Number of probes per tail loss episode is limited to one.
-> Connection is SACK enabled.
When PTO fires:
new_segment_exists:
-> transmit new segment.
-> packets_out++. cwnd remains same.
no_new_packet:
-> retransmit the last segment.
Its ACK triggers FACK or early retransmit based recovery.
ACK path:
-> rearm RTO at start of ACK processing.
-> reschedule PTO if need be.
In addition, the patch includes a small variation to the Early Retransmit
(ER) algorithm, such that ER and TLP together can in principle recover any
N-degree of tail loss through fast recovery. TLP is controlled by the same
sysctl as ER, tcp_early_retrans sysctl.
tcp_early_retrans==0; disables TLP and ER.
==1; enables RFC5827 ER.
==2; delayed ER.
==3; TLP and delayed ER. [DEFAULT]
==4; TLP only.
The TLP patch series have been extensively tested on Google Web servers.
It is most effective for short Web trasactions, where it reduced RTOs by 15%
and improved HTTP response time (average by 6%, 99th percentile by 10%).
The transmitted probes account for <0.5% of the overall transmissions.
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If in either of the above functions inet_csk_route_child_sock() or
__inet_inherit_port() fails, the newsk will not be freed:
unreferenced object 0xffff88022e8a92c0 (size 1592):
comm "softirq", pid 0, jiffies 4294946244 (age 726.160s)
hex dump (first 32 bytes):
0a 01 01 01 0a 01 01 02 00 00 00 00 a7 cc 16 00 ................
02 00 03 01 00 00 00 00 00 00 00 00 00 00 00 00 ................
backtrace:
[<ffffffff8153d190>] kmemleak_alloc+0x21/0x3e
[<ffffffff810ab3e7>] kmem_cache_alloc+0xb5/0xc5
[<ffffffff8149b65b>] sk_prot_alloc.isra.53+0x2b/0xcd
[<ffffffff8149b784>] sk_clone_lock+0x16/0x21e
[<ffffffff814d711a>] inet_csk_clone_lock+0x10/0x7b
[<ffffffff814ebbc3>] tcp_create_openreq_child+0x21/0x481
[<ffffffff814e8fa5>] tcp_v4_syn_recv_sock+0x3a/0x23b
[<ffffffff814ec5ba>] tcp_check_req+0x29f/0x416
[<ffffffff814e8e10>] tcp_v4_do_rcv+0x161/0x2bc
[<ffffffff814eb917>] tcp_v4_rcv+0x6c9/0x701
[<ffffffff814cea9f>] ip_local_deliver_finish+0x70/0xc4
[<ffffffff814cec20>] ip_local_deliver+0x4e/0x7f
[<ffffffff814ce9f8>] ip_rcv_finish+0x1fc/0x233
[<ffffffff814cee68>] ip_rcv+0x217/0x267
[<ffffffff814a7bbe>] __netif_receive_skb+0x49e/0x553
[<ffffffff814a7cc3>] netif_receive_skb+0x50/0x82
This happens, because sk_clone_lock initializes sk_refcnt to 2, and thus
a single sock_put() is not enough to free the memory. Additionally, things
like xfrm, memcg, cookie_values,... may have been initialized.
We have to free them properly.
This is fixed by forcing a call to tcp_done(), ending up in
inet_csk_destroy_sock, doing the final sock_put(). tcp_done() is necessary,
because it ends up doing all the cleanup on xfrm, memcg, cookie_values,
xfrm,...
Before calling tcp_done, we have to set the socket to SOCK_DEAD, to
force it entering inet_csk_destroy_sock. To avoid the warning in
inet_csk_destroy_sock, inet_num has to be set to 0.
As inet_csk_destroy_sock does a dec on orphan_count, we first have to
increase it.
Calling tcp_done() allows us to remove the calls to
tcp_clear_xmit_timer() and tcp_cleanup_congestion_control().
A similar approach is taken for dccp by calling dccp_done().
This is in the kernel since 093d282321 (tproxy: fix hash locking issue
when using port redirection in __inet_inherit_port()), thus since
version >= 2.6.37.
Signed-off-by: Christoph Paasch <christoph.paasch@uclouvain.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
IPv6 needs a cookie in dst_check() call.
We need to add rx_dst_cookie and provide a family independent
sk_rx_dst_set(sk, skb) method to properly support IPv6 TCP early demux.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This abstracts away the call to dst_ops->update_pmtu() so that we can
transparently handle the fact that, in the future, the dst itself can
be invalidated by the PMTU update (when we have non-host routes cached
in sockets).
So we try to rebuild the socket cached route after the method
invocation if necessary.
This isn't used by SCTP because it needs to cache dsts per-transport,
and thus will need it's own local version of this helper.
Signed-off-by: David S. Miller <davem@davemloft.net>
Don't cache output dst for syncookies, as this adds pressure on IP route
cache and rcu subsystem for no gain.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Hans Schillstrom <hans.schillstrom@ericsson.com>
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The get_peer method TCP uses is full of special cases that make no
sense accommodating, and it also gets in the way of doing more
reasonable things here.
First of all, if the socket doesn't have a usable cached route, there
is no sense in trying to optimize timewait recycling.
Likewise for the case where we have IP options, such as SRR enabled,
that make the IP header destination address (and thus the destination
address of the route key) differ from that of the connection's
destination address.
Just return a NULL peer in these cases, and thus we're also able to
get rid of the clumsy inetpeer release logic.
Signed-off-by: David S. Miller <davem@davemloft.net>
Priyaranjan Jha