After blamed commit, TFO client-ack-dropped-then-recovery-ms-timestamps
packetdrill test failed.
David Morley and Neal Cardwell started investigating and Neal pointed
that we had :
tcp_conn_request()
tcp_try_fastopen()
-> tcp_fastopen_create_child
-> child = inet_csk(sk)->icsk_af_ops->syn_recv_sock()
-> tcp_create_openreq_child()
-> copy req_usec_ts from req:
newtp->tcp_usec_ts = treq->req_usec_ts;
// now the new TFO server socket always does usec TS, no matter
// what the route options are...
send_synack()
-> tcp_make_synack()
// disable tcp_rsk(req)->req_usec_ts if route option is not present:
if (tcp_rsk(req)->req_usec_ts < 0)
tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst);
tcp_conn_request() has the initial dst, we can initialize
tcp_rsk(req)->req_usec_ts there instead of later in send_synack();
This means tcp_rsk(req)->req_usec_ts can be a boolean.
Many thanks to David an Neal for their help.
Fixes: 614e8316aa ("tcp: add support for usec resolution in TCP TS values")
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202310302216.f79d78bc-oliver.sang@intel.com
Suggested-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: David Morley <morleyd@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Similarly to RST segments, wire SYN-ACKs to TCP-AO.
tcp_rsk_used_ao() is handy here to check if the request socket used AO
and needs a signature on the outgoing segments.
Co-developed-by: Francesco Ruggeri <fruggeri@arista.com>
Signed-off-by: Francesco Ruggeri <fruggeri@arista.com>
Co-developed-by: Salam Noureddine <noureddine@arista.com>
Signed-off-by: Salam Noureddine <noureddine@arista.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Acked-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now when the new request socket is created from the listening socket,
it's recorded what MKT was used by the peer. tcp_rsk_used_ao() is
a new helper for checking if TCP-AO option was used to create the
request socket.
tcp_ao_copy_all_matching() will copy all keys that match the peer on the
request socket, as well as preparing them for the usage (creating
traffic keys).
Co-developed-by: Francesco Ruggeri <fruggeri@arista.com>
Signed-off-by: Francesco Ruggeri <fruggeri@arista.com>
Co-developed-by: Salam Noureddine <noureddine@arista.com>
Signed-off-by: Salam Noureddine <noureddine@arista.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Acked-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add support for sockets in time-wait state.
ao_info as well as all keys are inherited on transition to time-wait
socket. The lifetime of ao_info is now protected by ref counter, so
that tcp_ao_destroy_sock() will destruct it only when the last user is
gone.
Co-developed-by: Francesco Ruggeri <fruggeri@arista.com>
Signed-off-by: Francesco Ruggeri <fruggeri@arista.com>
Co-developed-by: Salam Noureddine <noureddine@arista.com>
Signed-off-by: Salam Noureddine <noureddine@arista.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Acked-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Using precalculated traffic keys, sign TCP segments as prescribed by
RFC5925. Per RFC, TCP header options are included in sign calculation:
"The TCP header, by default including options, and where the TCP
checksum and TCP-AO MAC fields are set to zero, all in network-
byte order." (5.1.3)
tcp_ao_hash_header() has exclude_options parameter to optionally exclude
TCP header from hash calculation, as described in RFC5925 (9.1), this is
needed for interaction with middleboxes that may change "some TCP
options". This is wired up to AO key flags and setsockopt() later.
Similarly to TCP-MD5 hash TCP segment fragments.
From this moment a user can start sending TCP-AO signed segments with
one of crypto ahash algorithms from supported by Linux kernel. It can
have a user-specified MAC length, to either save TCP option header space
or provide higher protection using a longer signature.
The inbound segments are not yet verified, TCP-AO option is ignored and
they are accepted.
Co-developed-by: Francesco Ruggeri <fruggeri@arista.com>
Signed-off-by: Francesco Ruggeri <fruggeri@arista.com>
Co-developed-by: Salam Noureddine <noureddine@arista.com>
Signed-off-by: Salam Noureddine <noureddine@arista.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Acked-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add traffic key calculation the way it's described in RFC5926.
Wire it up to tcp_finish_connect() and cache the new keys straight away
on already established TCP connections.
Co-developed-by: Francesco Ruggeri <fruggeri@arista.com>
Signed-off-by: Francesco Ruggeri <fruggeri@arista.com>
Co-developed-by: Salam Noureddine <noureddine@arista.com>
Signed-off-by: Salam Noureddine <noureddine@arista.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Acked-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Be as conservative as possible: if there is TCP-MD5 key for a given peer
regardless of L3 interface - don't allow setting TCP-AO key for the same
peer. According to RFC5925, TCP-AO is supposed to replace TCP-MD5 and
there can't be any switch between both on any connected tuple.
Later it can be relaxed, if there's a use, but in the beginning restrict
any intersection.
Note: it's still should be possible to set both TCP-MD5 and TCP-AO keys
on a listening socket for *different* peers.
Co-developed-by: Francesco Ruggeri <fruggeri@arista.com>
Signed-off-by: Francesco Ruggeri <fruggeri@arista.com>
Co-developed-by: Salam Noureddine <noureddine@arista.com>
Signed-off-by: Salam Noureddine <noureddine@arista.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Acked-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
RTAX_FEATURE_ALLFRAG was added before the first git commit:
https://www.mail-archive.com/bk-commits-head@vger.kernel.org/msg03399.html
The feature would send packets to the fragmentation path if a box
receives a PMTU value with less than 1280 byte. However, since commit
9d289715eb ("ipv6: stop sending PTB packets for MTU < 1280"), such
message would be simply discarded. The feature flag is neither supported
in iproute2 utility. In theory one can still manipulate it with direct
netlink message, but it is not ideal because it was based on obsoleted
guidance of RFC-2460 (replaced by RFC-8200).
The feature would always test false at the moment, so remove related
code or mark them as unused.
Signed-off-by: Yan Zhai <yan@cloudflare.com>
Reviewed-by: Florian Westphal <fw@strlen.de>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/d78e44dcd9968a252143ffe78460446476a472a1.1698156966.git.yan@cloudflare.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Back in 2015, Van Jacobson suggested to use usec resolution in TCP TS values.
This has been implemented in our private kernels.
Goals were :
1) better observability of delays in networking stacks.
2) better disambiguation of events based on TSval/ecr values.
3) building block for congestion control modules needing usec resolution.
Back then we implemented a schem based on private SYN options
to negotiate the feature.
For upstream submission, we chose to use a route attribute,
because this feature is probably going to be used in private
networks [1] [2].
ip route add 10/8 ... features tcp_usec_ts
Note that RFC 7323 recommends a
"timestamp clock frequency in the range 1 ms to 1 sec per tick.",
but also mentions
"the maximum acceptable clock frequency is one tick every 59 ns."
[1] Unfortunately RFC 7323 5.5 (Outdated Timestamps) suggests
to invalidate TS.Recent values after a flow was idle for more
than 24 days. This is the part making usec_ts a problem
for peers following this recommendation for long living
idle flows.
[2] Attempts to standardize usec ts went nowhere:
https://www.ietf.org/proceedings/97/slides/slides-97-tcpm-tcp-options-for-low-latency-00.pdfhttps://datatracker.ietf.org/doc/draft-wang-tcpm-low-latency-opt/
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This helper returns a TSval from a TCP socket.
It currently calls tcp_time_stamp_ms() but will soon
be able to return a usec based TSval, depending
on an upcoming tp->tcp_usec_ts field.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This helper returns a 32bit TCP TSval from skb->tstamp.
As we are going to support usec or ms units soon, rename it
to tcp_skb_timestamp_ts() and add a boolean to select the unit.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit 75eefc6c59 ("tcp: tsq: add a shortcut in tcp_small_queue_check()")
we allowed to send an skb regardless of TSQ limits being hit if rtx queue
was empty or had a single skb, in order to better fill the pipe
when/if TX completions were slow.
Then later, commit 75c119afe1 ("tcp: implement rb-tree based
retransmit queue") accidentally removed the special case for
one skb in rtx queue.
Stefan Wahren reported a regression in single TCP flow throughput
using a 100Mbit fec link, starting from commit 65466904b0 ("tcp: adjust
TSO packet sizes based on min_rtt"). This last commit only made the
regression more visible, because it locked the TCP flow on a particular
behavior where TSQ prevented two skbs being pushed downstream,
adding silences on the wire between each TSO packet.
Many thanks to Stefan for his invaluable help !
Fixes: 75c119afe1 ("tcp: implement rb-tree based retransmit queue")
Link: https://lore.kernel.org/netdev/7f31ddc8-9971-495e-a1f6-819df542e0af@gmx.net/
Reported-by: Stefan Wahren <wahrenst@gmx.net>
Tested-by: Stefan Wahren <wahrenst@gmx.net>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Link: https://lore.kernel.org/r/20231017124526.4060202-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We discovered from packet traces of slow loss recovery on kernels with
the default HZ=250 setting (and min_rtt < 1ms) that after reordering,
when receiving a SACKed sequence range, the RACK reordering timer was
firing after about 16ms rather than the desired value of roughly
min_rtt/4 + 2ms. The problem is largely due to the RACK reorder timer
calculation adding in TCP_TIMEOUT_MIN, which is 2 jiffies. On kernels
with HZ=250, this is 2*4ms = 8ms. The TLP timer calculation has the
exact same issue.
This commit fixes the TLP transmit timer and RACK reordering timer
floor calculation to more closely match the intended 2ms floor even on
kernels with HZ=250. It does this by adding in a new
TCP_TIMEOUT_MIN_US floor of 2000 us and then converting to jiffies,
instead of the current approach of converting to jiffies and then
adding th TCP_TIMEOUT_MIN value of 2 jiffies.
Our testing has verified that on kernels with HZ=1000, as expected,
this does not produce significant changes in behavior, but on kernels
with the default HZ=250 the latency improvement can be large. For
example, our tests show that for HZ=250 kernels at low RTTs this fix
roughly halves the latency for the RACK reorder timer: instead of
mostly firing at 16ms it mostly fires at 8ms.
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Fixes: bb4d991a28 ("tcp: adjust tail loss probe timeout")
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20231015174700.2206872-1-ncardwell.sw@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
TCP pingpong threshold is 1 by default. But some applications, like SQL DB
may prefer a higher pingpong threshold to activate delayed acks in quick
ack mode for better performance.
The pingpong threshold and related code were changed to 3 in the year
2019 in:
commit 4a41f453be ("tcp: change pingpong threshold to 3")
And reverted to 1 in the year 2022 in:
commit 4d8f24eeed ("Revert "tcp: change pingpong threshold to 3"")
There is no single value that fits all applications.
Add net.ipv4.tcp_pingpong_thresh sysctl tunable, so it can be tuned for
optimal performance based on the application needs.
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://lore.kernel.org/r/1697056244-21888-1-git-send-email-haiyangz@microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Cross-merge networking fixes after downstream PR.
No conflicts.
Adjacent changes:
kernel/bpf/verifier.c
829955981c ("bpf: Fix verifier log for async callback return values")
a923819fb2 ("bpf: Treat first argument as return value for bpf_throw")
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
tcp_stream_alloc_skb() initializes the skb to use tcp_tsorted_anchor
which is a union with the destructor. We need to clean that
TCP-iness up before freeing.
Fixes: 736013292e ("tcp: let tcp_mtu_probe() build headless packets")
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20231010173651.3990234-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This commit fixes quick-ack counting so that it only considers that a
quick-ack has been provided if we are sending an ACK that newly
acknowledges data.
The code was erroneously using the number of data segments in outgoing
skbs when deciding how many quick-ack credits to remove. This logic
does not make sense, and could cause poor performance in
request-response workloads, like RPC traffic, where requests or
responses can be multi-segment skbs.
When a TCP connection decides to send N quick-acks, that is to
accelerate the cwnd growth of the congestion control module
controlling the remote endpoint of the TCP connection. That quick-ack
decision is purely about the incoming data and outgoing ACKs. It has
nothing to do with the outgoing data or the size of outgoing data.
And in particular, an ACK only serves the intended purpose of allowing
the remote congestion control to grow the congestion window quickly if
the ACK is ACKing or SACKing new data.
The fix is simple: only count packets as serving the goal of the
quickack mechanism if they are ACKing/SACKing new data. We can tell
whether this is the case by checking inet_csk_ack_scheduled(), since
we schedule an ACK exactly when we are ACKing/SACKing new data.
Fixes: fc6415bcb0 ("[TCP]: Fix quick-ack decrementing with TSO.")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Yuchung Cheng <ycheng@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20231001151239.1866845-1-ncardwell.sw@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This field can be read or written without socket lock being held.
Add annotations to avoid load-store tearing.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
SO_MAX_PACING_RATE setsockopt() does not need to hold
the socket lock, because sk->sk_pacing_rate readers
can run fine if the value is changed by other threads,
after adding READ_ONCE() accessors.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While BPF allows to set icsk->->icsk_delack_max
and/or icsk->icsk_rto_min, we have an ip route
attribute (RTAX_RTO_MIN) to be able to tune rto_min,
but nothing to consequently adjust max delayed ack,
which vary from 40ms to 200 ms (TCP_DELACK_{MIN|MAX}).
This makes RTAX_RTO_MIN of almost no practical use,
unless customers are in big trouble.
Modern days datacenter communications want to set
rto_min to ~5 ms, and the max delayed ack one jiffie
smaller to avoid spurious retransmits.
After this patch, an "rto_min 5" route attribute will
effectively lower max delayed ack timers to 4 ms.
Note in the following ss output, "rto:6 ... ato:4"
$ ss -temoi dst XXXXXX
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
ESTAB 0 0 [2002:a05:6608:295::]:52950 [2002:a05:6608:297::]:41597
ino:255134 sk:1001 <->
skmem:(r0,rb1707063,t872,tb262144,f0,w0,o0,bl0,d0) ts sack
cubic wscale:8,8 rto:6 rtt:0.02/0.002 ato:4 mss:4096 pmtu:4500
rcvmss:536 advmss:4096 cwnd:10 bytes_sent:54823160 bytes_acked:54823121
bytes_received:54823120 segs_out:1370582 segs_in:1370580
data_segs_out:1370579 data_segs_in:1370578 send 16.4Gbps
pacing_rate 32.6Gbps delivery_rate 1.72Gbps delivered:1370579
busy:26920ms unacked:1 rcv_rtt:34.615 rcv_space:65920
rcv_ssthresh:65535 minrtt:0.015 snd_wnd:65536
While we could argue this patch fixes a bug with RTAX_RTO_MIN,
I do not add a Fixes: tag, so that we can soak it a bit before
asking backports to stable branches.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This idea came after a particular workload requested
the quickack attribute set on routes, and a performance
drop was noticed for large bulk transfers.
For high throughput flows, it is best to use one cpu
running the user thread issuing socket system calls,
and a separate cpu to process incoming packets from BH context.
(With TSO/GRO, bottleneck is usually the 'user' cpu)
Problem is the user thread can spend a lot of time while holding
the socket lock, forcing BH handler to queue most of incoming
packets in the socket backlog.
Whenever the user thread releases the socket lock, it must first
process all accumulated packets in the backlog, potentially
adding latency spikes. Due to flood mitigation, having too many
packets in the backlog increases chance of unexpected drops.
Backlog processing unfortunately shifts a fair amount of cpu cycles
from the BH cpu to the 'user' cpu, thus reducing max throughput.
This patch takes advantage of the backlog processing,
and the fact that ACK are mostly cumulative.
The idea is to detect we are in the backlog processing
and defer all eligible ACK into a single one,
sent from tcp_release_cb().
This saves cpu cycles on both sides, and network resources.
Performance of a single TCP flow on a 200Gbit NIC:
- Throughput is increased by 20% (100Gbit -> 120Gbit).
- Number of generated ACK per second shrinks from 240,000 to 40,000.
- Number of backlog drops per second shrinks from 230 to 0.
Benchmark context:
- Regular netperf TCP_STREAM (no zerocopy)
- Intel(R) Xeon(R) Platinum 8481C (Saphire Rapids)
- MAX_SKB_FRAGS = 17 (~60KB per GRO packet)
This feature is guarded by a new sysctl, and enabled by default:
/proc/sys/net/ipv4/tcp_backlog_ack_defer
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Dave Taht <dave.taht@gmail.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
This partially reverts c3f9b01849 ("tcp: tcp_release_cb()
should release socket ownership").
prequeue has been removed by Florian in commit e7942d0633
("tcp: remove prequeue support")
__tcp_checksum_complete_user() being gone, we no longer
have to release socket ownership in tcp_release_cb().
This is a prereq for third patch in the series
("net: call prot->release_cb() when processing backlog").
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Every time sk->sk_forward_alloc is read locklessly,
add a READ_ONCE().
Add sk_forward_alloc_add() helper to centralize updates,
to reduce number of WRITE_ONCE().
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Enabling BIG TCP on a low end platform apparently increased
chances of getting flows locked on one busy TX queue.
A similar problem was handled in commit 9b462d02d6
("tcp: TCP Small Queues and strange attractors"),
but the strategy worked for either bulk flows,
or 'large enough' RPC. BIG TCP changed how large
RPC needed to be to enable the work around:
If RPC fits in a single skb, TSQ never triggers.
Root cause for the problem is a busy TX queue,
with delayed TX completions.
This patch changes how we set skb->ooo_okay to detect
the case TX completion was not done, but incoming ACK
already was processed and emptied rtx queue.
Update the comment to explain the tricky details.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20230817182353.2523746-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
For now, skb will be dropped when no memory, which makes client keep
retrans util timeout and it's not friendly to the users.
In this patch, we reply an ACK with zero-window in this case to update
the snd_wnd of the sender to 0. Therefore, the sender won't timeout the
connection and will probe the zero-window with the retransmits.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP socket saves the minimum required header length in tcp_header_len
of struct tcp_sock, and later the value is used in __tcp_fast_path_on()
to generate a part of TCP header in tcp_sock(sk)->pred_flags.
In tcp_rcv_established(), if the incoming packet has the same pattern
with pred_flags, we enter the fast path and skip full option parsing.
The MD5 option is parsed in tcp_v[46]_rcv(), so we need not parse it
again later in tcp_rcv_established() unless other options exist. We
add TCPOLEN_MD5SIG_ALIGNED to tcp_header_len in two paths to avoid the
slow path.
For passive open connections with MD5, we add TCPOLEN_MD5SIG_ALIGNED
to tcp_header_len in tcp_create_openreq_child() after 3WHS.
On the other hand, we do it in tcp_connect_init() for active open
connections. However, the value is overwritten while processing
SYN+ACK or crossed SYN in tcp_rcv_synsent_state_process().
These two cases will have the wrong value in pred_flags and never go
into the fast path.
We could update tcp_header_len in tcp_rcv_synsent_state_process(), but
a test with slightly modified netperf which uses MD5 for each flow shows
that the slow path is actually a bit faster than the fast path.
On c5.4xlarge EC2 instance (16 vCPU, 32 GiB mem)
$ for i in {1..10}; do
./super_netperf $(nproc) -H localhost -l 10 -- -m 256 -M 256;
done
Avg of 10
* 36e68eadd3 : 10.376 Gbps
* all fast path : 10.374 Gbps (patch v2, See Link)
* all slow path : 10.394 Gbps
The header prediction is not worth adding complexity for MD5, so let's
disable it for MD5.
Link: https://lore.kernel.org/netdev/20230803042214.38309-1-kuniyu@amazon.com/
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20230803224552.69398-2-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Under certain circumstances, the tcp receive buffer memory limit
set by autotuning (sk_rcvbuf) is increased due to incoming data
packets as a result of the window not closing when it should be.
This can result in the receive buffer growing all the way up to
tcp_rmem[2], even for tcp sessions with a low BDP.
To reproduce: Connect a TCP session with the receiver doing
nothing and the sender sending small packets (an infinite loop
of socket send() with 4 bytes of payload with a sleep of 1 ms
in between each send()). This will cause the tcp receive buffer
to grow all the way up to tcp_rmem[2].
As a result, a host can have individual tcp sessions with receive
buffers of size tcp_rmem[2], and the host itself can reach tcp_mem
limits, causing the host to go into tcp memory pressure mode.
The fundamental issue is the relationship between the granularity
of the window scaling factor and the number of byte ACKed back
to the sender. This problem has previously been identified in
RFC 7323, appendix F [1].
The Linux kernel currently adheres to never shrinking the window.
In addition to the overallocation of memory mentioned above, the
current behavior is functionally incorrect, because once tcp_rmem[2]
is reached when no remediations remain (i.e. tcp collapse fails to
free up any more memory and there are no packets to prune from the
out-of-order queue), the receiver will drop in-window packets
resulting in retransmissions and an eventual timeout of the tcp
session. A receive buffer full condition should instead result
in a zero window and an indefinite wait.
In practice, this problem is largely hidden for most flows. It
is not applicable to mice flows. Elephant flows can send data
fast enough to "overrun" the sk_rcvbuf limit (in a single ACK),
triggering a zero window.
But this problem does show up for other types of flows. Examples
are websockets and other type of flows that send small amounts of
data spaced apart slightly in time. In these cases, we directly
encounter the problem described in [1].
RFC 7323, section 2.4 [2], says there are instances when a retracted
window can be offered, and that TCP implementations MUST ensure
that they handle a shrinking window, as specified in RFC 1122,
section 4.2.2.16 [3]. All prior RFCs on the topic of tcp window
management have made clear that sender must accept a shrunk window
from the receiver, including RFC 793 [4] and RFC 1323 [5].
This patch implements the functionality to shrink the tcp window
when necessary to keep the right edge within the memory limit by
autotuning (sk_rcvbuf). This new functionality is enabled with
the new sysctl: net.ipv4.tcp_shrink_window
Additional information can be found at:
https://blog.cloudflare.com/unbounded-memory-usage-by-tcp-for-receive-buffers-and-how-we-fixed-it/
[1] https://www.rfc-editor.org/rfc/rfc7323#appendix-F
[2] https://www.rfc-editor.org/rfc/rfc7323#section-2.4
[3] https://www.rfc-editor.org/rfc/rfc1122#page-91
[4] https://www.rfc-editor.org/rfc/rfc793
[5] https://www.rfc-editor.org/rfc/rfc1323
Signed-off-by: Mike Freemon <mfreemon@cloudflare.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now all tcp_stream_alloc_skb() callers pass @size == 0, we can
remove this parameter.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now all skbs in write queue do not contain any payload in skb->head,
we can remove some dead code.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_send_syn_data() is the last component in TCP transmit
path to put payload in skb->head.
Switch it to use page frags, so that we can remove dead
code later.
This allows to put more payload than previous implementation.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_mtu_probe() is still copying payload from skbs in the write queue,
using skb_copy_bits(), ignoring potential errors.
Modern TCP stack wants to only deal with payload found in page frags,
as this is a prereq for TCPDirect (host stack might not have access
to the payload)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20230607214113.1992947-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We can change tcp_sk() to propagate its argument const qualifier,
thanks to container_of_const().
We have two places where a const sock pointer has to be upgraded
to a write one. We have been using const qualifier for lockless
listeners to clearly identify points where writes could happen.
Add tcp_sk_rw() helper to better document these.
tcp_inbound_md5_hash(), __tcp_grow_window(), tcp_reset_check()
and tcp_rack_reo_wnd() get an additional const qualififer
for their @tp local variables.
smc_check_reset_syn_req() also needs a similar change.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Simon Horman <simon.horman@corigine.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_poll() reads sk->sk_err without socket lock held/owned.
We should used READ_ONCE() here, and update writers
to use WRITE_ONCE().
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_rtx_synack() now could be called in process context as explained in
0a375c8224 ("tcp: tcp_rtx_synack() can be called from process
context").
tcp_rtx_synack() might call tcp_make_synack(), which will touch per-CPU
variables with preemption enabled. This causes the following BUG:
BUG: using __this_cpu_add() in preemptible [00000000] code: ThriftIO1/5464
caller is tcp_make_synack+0x841/0xac0
Call Trace:
<TASK>
dump_stack_lvl+0x10d/0x1a0
check_preemption_disabled+0x104/0x110
tcp_make_synack+0x841/0xac0
tcp_v6_send_synack+0x5c/0x450
tcp_rtx_synack+0xeb/0x1f0
inet_rtx_syn_ack+0x34/0x60
tcp_check_req+0x3af/0x9e0
tcp_rcv_state_process+0x59b/0x2030
tcp_v6_do_rcv+0x5f5/0x700
release_sock+0x3a/0xf0
tcp_sendmsg+0x33/0x40
____sys_sendmsg+0x2f2/0x490
__sys_sendmsg+0x184/0x230
do_syscall_64+0x3d/0x90
Avoid calling __TCP_INC_STATS() with will touch per-cpu variables. Use
TCP_INC_STATS() which is safe to be called from context switch.
Fixes: 8336886f78 ("tcp: TCP Fast Open Server - support TFO listeners")
Signed-off-by: Breno Leitao <leitao@debian.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20230308190745.780221-1-leitao@debian.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
To do that, separate two scenarios:
- where it's the first MD5 key on the system, which means that enabling
of the static key may need to sleep;
- copying of an existing key from a listening socket to the request
socket upon receiving a signed TCP segment, where static key was
already enabled (when the key was added to the listening socket).
Now the life-time of the static branch for TCP-MD5 is until:
- last tcp_md5sig_info is destroyed
- last socket in time-wait state with MD5 key is closed.
Which means that after all sockets with TCP-MD5 keys are gone, the
system gets back the performance of disabled md5-key static branch.
While at here, provide static_key_fast_inc() helper that does ref
counter increment in atomic fashion (without grabbing cpus_read_lock()
on CONFIG_JUMP_LABEL=y). This is needed to add a new user for
a static_key when the caller controls the lifetime of another user.
Signed-off-by: Dmitry Safonov <dima@arista.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Use try_cmpxchg() (instead of cmpxchg()) in a more readable way.
oval = smp_load_acquire(&sk->sk_tsq_flags);
do {
...
} while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval));
Reduce indentation level.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20221110190239.3531280-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
try_cmpxchg() is slighly more efficient (at least on x86),
and smp_load_acquire(&sk->sk_tsq_flags) could avoid a KCSAN report.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20221110174829.3403442-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This commit fixes a bug in the tracking of max_packets_out and
is_cwnd_limited. This bug can cause the connection to fail to remember
that is_cwnd_limited is true, causing the connection to fail to grow
cwnd when it should, causing throughput to be lower than it should be.
The following event sequence is an example that triggers the bug:
(a) The connection is cwnd_limited, but packets_out is not at its
peak due to TSO deferral deciding not to send another skb yet.
In such cases the connection can advance max_packets_seq and set
tp->is_cwnd_limited to true and max_packets_out to a small
number.
(b) Then later in the round trip the connection is pacing-limited (not
cwnd-limited), and packets_out is larger. In such cases the
connection would raise max_packets_out to a bigger number but
(unexpectedly) flip tp->is_cwnd_limited from true to false.
This commit fixes that bug.
One straightforward fix would be to separately track (a) the next
window after max_packets_out reaches a maximum, and (b) the next
window after tp->is_cwnd_limited is set to true. But this would
require consuming an extra u32 sequence number.
Instead, to save space we track only the most important
information. Specifically, we track the strongest available signal of
the degree to which the cwnd is fully utilized:
(1) If the connection is cwnd-limited then we remember that fact for
the current window.
(2) If the connection not cwnd-limited then we track the maximum
number of outstanding packets in the current window.
In particular, note that the new logic cannot trigger the buggy
(a)/(b) sequence above because with the new logic a condition where
tp->packets_out > tp->max_packets_out can only trigger an update of
tp->is_cwnd_limited if tp->is_cwnd_limited is false.
This first showed up in a testing of a BBRv2 dev branch, but this
buggy behavior highlighted a general issue with the
tcp_cwnd_validate() logic that can cause cwnd to fail to increase at
the proper rate for any TCP congestion control, including Reno or
CUBIC.
Fixes: ca8a226343 ("tcp: make cwnd-limited checks measurement-based, and gentler")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Kevin(Yudong) Yang <yyd@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While reading sysctl_[rw]mem_(max|default), they can be changed
concurrently. Thus, we need to add READ_ONCE() to its readers.
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While reading these sysctl variables, they can be changed concurrently.
Thus, we need to add READ_ONCE() to their readers.
- .sysctl_rmem
- .sysctl_rwmem
- .sysctl_rmem_offset
- .sysctl_wmem_offset
- sysctl_tcp_rmem[1, 2]
- sysctl_tcp_wmem[1, 2]
- sysctl_decnet_rmem[1]
- sysctl_decnet_wmem[1]
- sysctl_tipc_rmem[1]
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This reverts commit 4a41f453be.
This to-be-reverted commit was meant to apply a stricter rule for the
stack to enter pingpong mode. However, the condition used to check for
interactive session "before(tp->lsndtime, icsk->icsk_ack.lrcvtime)" is
jiffy based and might be too coarse, which delays the stack entering
pingpong mode.
We revert this patch so that we no longer use the above condition to
determine interactive session, and also reduce pingpong threshold to 1.
Fixes: 4a41f453be ("tcp: change pingpong threshold to 3")
Reported-by: LemmyHuang <hlm3280@163.com>
Suggested-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20220721204404.388396-1-weiwan@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
While reading sysctl_tcp_tso_rtt_log, it can be changed concurrently.
Thus, we need to add READ_ONCE() to its reader.
Fixes: 65466904b0 ("tcp: adjust TSO packet sizes based on min_rtt")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While reading sysctl_tcp_min_tso_segs, it can be changed concurrently.
Thus, we need to add READ_ONCE() to its reader.
Fixes: 95bd09eb27 ("tcp: TSO packets automatic sizing")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While reading sysctl_tcp_limit_output_bytes, it can be changed
concurrently. Thus, we need to add READ_ONCE() to its reader.
Fixes: 46d3ceabd8 ("tcp: TCP Small Queues")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
