This reverts commit 0de2a5c4b824da2205658ebebb99a55c43cdf60f.
I forgot that a TCP socket could receive messages in its error queue.
sock_queue_err_skb() can be called without socket lock being held,
and changes sk->sk_rmem_alloc.
The fact that skbs in error queue are limited by sk->sk_rcvbuf
means that error messages can be dropped if socket receive
queues are full, which is an orthogonal issue.
In future kernels, we could use a separate sk->sk_error_mem_alloc
counter specifically for the error queue.
Fixes: 0de2a5c4b824 ("tcp: avoid atomic operations on sk->sk_rmem_alloc")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250331075946.31960-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
TCP uses generic skb_set_owner_r() and sock_rfree()
for received packets, with socket lock being owned.
Switch to private versions, avoiding two atomic operations
per packet.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250320121604.3342831-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
As a followup of my presentation in Zagreb for netdev 0x19:
icsk_clean_acked is only used by TCP when/if CONFIG_TLS_DEVICE
is enabled from tcp_ack().
Rename it to tcp_clean_acked, move it to tcp_sock structure
in the tcp_sock_read_rx for better cache locality in TCP
fast path.
Define this field only when CONFIG_TLS_DEVICE is enabled
saving 8 bytes on configs not using it.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Sabrina Dubroca <sd@queasysnail.net>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250317085313.2023214-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
tcp_in_quickack_mode() is called from input path for small packets.
It calls __sk_dst_get() which reads sk->sk_dst_cache which has been
put in sock_read_tx group (for good reasons).
Then dst_metric(dst, RTAX_QUICKACK) also needs extra cache line misses.
Cache RTAX_QUICKACK in icsk->icsk_ack.dst_quick_ack to no longer pull
these cache lines for the cases a delayed ACK is scheduled.
After this patch TCP receive path does not longer access sock_read_tx
group.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250312083907.1931644-1-edumazet@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Create helpers for TCP ECN modes. No functional changes.
Signed-off-by: Ilpo Järvinen <ij@kernel.org>
Signed-off-by: Chia-Yu Chang <chia-yu.chang@nokia-bell-labs.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Rename tcp_ecn_check_ce to tcp_data_ecn_check as it is
called only for data segments, not for ACKs (with AccECN,
also ACKs may get ECN bits).
The extra "layer" in tcp_ecn_check_ce() function just
checks for ECN being enabled, that can be moved into
tcp_ecn_field_check rather than having the __ variant.
No functional changes.
Signed-off-by: Ilpo Järvinen <ij@kernel.org>
Signed-off-by: Chia-Yu Chang <chia-yu.chang@nokia-bell-labs.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Whenever timestamp advances, it declares progress which
can be used by the other parts of the stack to decide that
the ACK is the most recent one seen so far.
AccECN will use this flag when deciding whether to use the
ACK to update AccECN state or not.
Signed-off-by: Ilpo Järvinen <ij@kernel.org>
Signed-off-by: Chia-Yu Chang <chia-yu.chang@nokia-bell-labs.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
- Move tcp_count_delivered() earlier and split tcp_count_delivered_ce()
out of it
- Move tcp_in_ack_event() later
- While at it, remove the inline from tcp_in_ack_event() and let
the compiler to decide
Accurate ECN's heuristics does not know if there is going
to be ACE field based CE counter increase or not until after
rtx queue has been processed. Only then the number of ACKed
bytes/pkts is available. As CE or not affects presence of
FLAG_ECE, that information for tcp_in_ack_event is not yet
available in the old location of the call to tcp_in_ack_event().
Signed-off-by: Ilpo Järvinen <ij@kernel.org>
Signed-off-by: Chia-Yu Chang <chia-yu.chang@nokia-bell-labs.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We want to add new drop reasons for packets dropped in 3WHS in the
following patches.
tcp_rcv_state_process() has to set reason to TCP_FASTOPEN,
because tcp_check_req() will conditionally overwrite the drop_reason.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250301201424.2046477-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Yong-Hao Zou mentioned that linux was not strict as other OS in 3WHS,
for flows using TCP TS option (RFC 7323)
As hinted by an old comment in tcp_check_req(),
we can check the TSEcr value in the incoming packet corresponds
to one of the SYNACK TSval values we have sent.
In this patch, I record the oldest and most recent values
that SYNACK packets have used.
Send a challenge ACK if we receive a TSEcr outside
of this range, and increase a new SNMP counter.
nstat -az | grep TSEcrRejected
TcpExtTSEcrRejected 0 0.0
Due to TCP fastopen implementation, do not apply yet these checks
for fastopen flows.
v2: No longer use req->num_timeout, but treq->snt_tsval_first
to detect when first SYNACK is prepared. This means
we make sure to not send an initial zero TSval.
Make sure MPTCP and TCP selftests are passing.
Change MIB name to TcpExtTSEcrRejected
v1: https://lore.kernel.org/netdev/CADVnQykD8i4ArpSZaPKaoNxLJ2if2ts9m4As+=Jvdkrgx1qMHw@mail.gmail.com/T/
Reported-by: Yong-Hao Zou <yonghaoz1994@gmail.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Matthieu Baerts (NGI0) <matttbe@kernel.org>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250225171048.3105061-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
-----BEGIN PGP SIGNATURE-----
iHUEABYKAB0WIQQ6NaUOruQGUkvPdG4raS+Z+3y5EwUCZ7ffOQAKCRAraS+Z+3y5
EzVHAP9h/QkeYoOZW9gul08I8vFiZsFe/lbOSLJWxeVfxb9JhgD/cMqby3qAxQK6
lsdNQ9jYG2232Wym89ag7fvTBK15Wg4=
=gkN2
-----END PGP SIGNATURE-----
Merge tag 'for-netdev' of https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next
Martin KaFai Lau says:
====================
pull-request: bpf-next 2025-02-20
We've added 19 non-merge commits during the last 8 day(s) which contain
a total of 35 files changed, 1126 insertions(+), 53 deletions(-).
The main changes are:
1) Add TCP_RTO_MAX_MS support to bpf_set/getsockopt, from Jason Xing
2) Add network TX timestamping support to BPF sock_ops, from Jason Xing
3) Add TX metadata Launch Time support, from Song Yoong Siang
* tag 'for-netdev' of https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next:
igc: Add launch time support to XDP ZC
igc: Refactor empty frame insertion for launch time support
net: stmmac: Add launch time support to XDP ZC
selftests/bpf: Add launch time request to xdp_hw_metadata
xsk: Add launch time hardware offload support to XDP Tx metadata
selftests/bpf: Add simple bpf tests in the tx path for timestamping feature
bpf: Support selective sampling for bpf timestamping
bpf: Add BPF_SOCK_OPS_TSTAMP_SENDMSG_CB callback
bpf: Add BPF_SOCK_OPS_TSTAMP_ACK_CB callback
bpf: Add BPF_SOCK_OPS_TSTAMP_SND_HW_CB callback
bpf: Add BPF_SOCK_OPS_TSTAMP_SND_SW_CB callback
bpf: Add BPF_SOCK_OPS_TSTAMP_SCHED_CB callback
net-timestamp: Prepare for isolating two modes of SO_TIMESTAMPING
bpf: Disable unsafe helpers in TX timestamping callbacks
bpf: Prevent unsafe access to the sock fields in the BPF timestamping callback
bpf: Prepare the sock_ops ctx and call bpf prog for TX timestamping
bpf: Add networking timestamping support to bpf_get/setsockopt()
selftests/bpf: Add rto max for bpf_setsockopt test
bpf: Support TCP_RTO_MAX_MS for bpf_setsockopt
====================
Link: https://patch.msgid.link/20250221022104.386462-1-martin.lau@linux.dev
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The subsequent patch will implement BPF TX timestamping. It will
call the sockops BPF program without holding the sock lock.
This breaks the current assumption that all sock ops programs will
hold the sock lock. The sock's fields of the uapi's bpf_sock_ops
requires this assumption.
To address this, a new "u8 is_locked_tcp_sock;" field is added. This
patch sets it in the current sock_ops callbacks. The "is_fullsock"
test is then replaced by the "is_locked_tcp_sock" test during
sock_ops_convert_ctx_access().
The new TX timestamping callbacks added in the subsequent patch will
not have this set. This will prevent unsafe access from the new
timestamping callbacks.
Potentially, we could allow read-only access. However, this would
require identifying which callback is read-safe-only and also requires
additional BPF instruction rewrites in the covert_ctx. Since the BPF
program can always read everything from a socket (e.g., by using
bpf_core_cast), this patch keeps it simple and disables all read
and write access to any socket fields through the bpf_sock_ops
UAPI from the new TX timestamping callback.
Moreover, note that some of the fields in bpf_sock_ops are specific
to tcp_sock, and sock_ops currently only supports tcp_sock. In
the future, UDP timestamping will be added, which will also break
this assumption. The same idea used in this patch will be reused.
Considering that the current sock_ops only supports tcp_sock, the
variable is named is_locked_"tcp"_sock.
Signed-off-by: Jason Xing <kerneljasonxing@gmail.com>
Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>
Link: https://patch.msgid.link/20250220072940.99994-4-kerneljasonxing@gmail.com
Xiumei reported hitting the WARN in xfrm6_tunnel_net_exit while
running tests that boil down to:
- create a pair of netns
- run a basic TCP test over ipcomp6
- delete the pair of netns
The xfrm_state found on spi_byaddr was not deleted at the time we
delete the netns, because we still have a reference on it. This
lingering reference comes from a secpath (which holds a ref on the
xfrm_state), which is still attached to an skb. This skb is not
leaked, it ends up on sk_receive_queue and then gets defer-free'd by
skb_attempt_defer_free.
The problem happens when we defer freeing an skb (push it on one CPU's
defer_list), and don't flush that list before the netns is deleted. In
that case, we still have a reference on the xfrm_state that we don't
expect at this point.
We already drop the skb's dst in the TCP receive path when it's no
longer needed, so let's also drop the secpath. At this point,
tcp_filter has already called into the LSM hooks that may require the
secpath, so it should not be needed anymore. However, in some of those
places, the MPTCP extension has just been attached to the skb, so we
cannot simply drop all extensions.
Fixes: 68822bdf76f1 ("net: generalize skb freeing deferral to per-cpu lists")
Reported-by: Xiumei Mu <xmu@redhat.com>
Signed-off-by: Sabrina Dubroca <sd@queasysnail.net>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/5055ba8f8f72bdcb602faa299faca73c280b7735.1739743613.git.sd@queasysnail.net
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Since commit under Fixes we set the window clamp in accordance
to newly measured rcvbuf scaling_ratio. If the scaling_ratio
decreased significantly we may put ourselves in a situation
where windows become smaller than rcvq_space, preventing
tcp_rcv_space_adjust() from increasing rcvbuf.
The significant decrease of scaling_ratio is far more likely
since commit 697a6c8cec03 ("tcp: increase the default TCP scaling ratio"),
which increased the "default" scaling ratio from ~30% to 50%.
Hitting the bad condition depends a lot on TCP tuning, and
drivers at play. One of Meta's workloads hits it reliably
under following conditions:
- default rcvbuf of 125k
- sender MTU 1500, receiver MTU 5000
- driver settles on scaling_ratio of 78 for the config above.
Initial rcvq_space gets calculated as TCP_INIT_CWND * tp->advmss
(10 * 5k = 50k). Once we find out the true scaling ratio and
MSS we clamp the windows to 38k. Triggering the condition also
depends on the message sequence of this workload. I can't repro
the problem with simple iperf or TCP_RR-style tests.
Fixes: a2cbb1603943 ("tcp: Update window clamping condition")
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Link: https://patch.msgid.link/20250217232905.3162187-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add a lightweight tracepoint to monitor TCP congestion window
adjustments via tcp_cwnd_reduction(). This tracepoint enables tracking
of:
- TCP window size fluctuations
- Active socket behavior
- Congestion window reduction events
Meta has been using BPF programs to monitor this function for years.
Adding a proper tracepoint provides a stable API for all users who need
to monitor TCP congestion window behavior.
Use DECLARE_TRACE instead of TRACE_EVENT to avoid creating trace event
infrastructure and exporting to tracefs, keeping the implementation
minimal. (Thanks Steven Rostedt)
Given that this patch creates a rawtracepoint, you could hook into it
using regular tooling, like bpftrace, using regular rawtracepoint
infrastructure, such as:
rawtracepoint:tcp_cwnd_reduction_tp {
....
}
Signed-off-by: Breno Leitao <leitao@debian.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250214-cwnd_tracepoint-v2-1-ef8d15162d95@debian.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Use EXPORT_IPV6_MOD[_GPL]() for symbols that don't need
to be exported unless CONFIG_IPV6=m
tcp_hashinfo and tcp_openreq_init_rwin() are no longer
used from any module anyway.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Mateusz Polchlopek <mateusz.polchlopek@intel.com>
Link: https://patch.msgid.link/20250212132418.1524422-4-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently, TCP stack uses a constant (120 seconds)
to limit the RTO value exponential growth.
Some applications want to set a lower value.
Add TCP_RTO_MAX_MS socket option to set a value (in ms)
between 1 and 120 seconds.
It is discouraged to change the socket rto max on a live
socket, as it might lead to unexpected disconnects.
Following patch is adding a netns sysctl to control the
default value at socket creation time.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
We want to factorize calls to inet_csk_reset_xmit_timer(),
to ease TCP_RTO_MAX change.
Current users want to add tcp_pacing_delay(sk)
to the timeout.
Remaining calls to inet_csk_reset_xmit_timer()
do not add the pacing delay. Following patch
will convert them, passing false for @pace_delay.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
All callers use TCP_RTO_MAX, we can factorize this constant,
becoming a variable soon.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
inet_csk_delete_keepalive_timer() and inet_csk_reset_keepalive_timer()
are only used from core TCP, there is no need to export them.
Replace their prefix by tcp.
Move them to net/ipv4/tcp_timer.c and make tcp_delete_keepalive_timer()
static.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Joe Damato <jdamato@fastly.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250206094605.2694118-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
These two functions are not called from modules.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Joe Damato <jdamato@fastly.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250206093436.2609008-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Prior patch in the series added TCP_RFC7323_PAWS_ACK drop reason.
This patch adds the corresponding SNMP counter, for folks
using nstat instead of tracing for TCP diagnostics.
nstat -az | grep PAWSOldAck
Suggested-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Tested-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250113135558.3180360-4-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
XPS can cause reorders because of the relaxed OOO
conditions for pure ACK packets.
For hosts not using RFS, what can happpen is that ACK
packets are sent on behalf of the cpu processing NIC
interrupts, selecting TX queue A for ACK packet P1.
Then a subsequent sendmsg() can run on another cpu.
TX queue selection uses the socket hash and can choose
another queue B for packets P2 (with payload).
If queue A is more congested than queue B,
the ACK packet P1 could be sent on the wire after
P2.
A linux receiver when processing P1 (after P2) currently increments
LINUX_MIB_PAWSESTABREJECTED (TcpExtPAWSEstab)
and use TCP_RFC7323_PAWS drop reason.
It might also send a DUPACK if not rate limited.
In order to better understand this pattern, this
patch adds a new drop_reason : TCP_RFC7323_PAWS_ACK.
For old ACKS like these, we no longer increment
LINUX_MIB_PAWSESTABREJECTED and no longer sends a DUPACK,
keeping credit for other more interesting DUPACK.
perf record -e skb:kfree_skb -a
perf script
...
swapper 0 [148] 27475.438637: skb:kfree_skb: ... location=tcp_validate_incoming+0x4f0 reason: TCP_RFC7323_PAWS_ACK
swapper 0 [208] 27475.438706: skb:kfree_skb: ... location=tcp_validate_incoming+0x4f0 reason: TCP_RFC7323_PAWS_ACK
swapper 0 [208] 27475.438908: skb:kfree_skb: ... location=tcp_validate_incoming+0x4f0 reason: TCP_RFC7323_PAWS_ACK
swapper 0 [148] 27475.439010: skb:kfree_skb: ... location=tcp_validate_incoming+0x4f0 reason: TCP_RFC7323_PAWS_ACK
swapper 0 [148] 27475.439214: skb:kfree_skb: ... location=tcp_validate_incoming+0x4f0 reason: TCP_RFC7323_PAWS_ACK
swapper 0 [208] 27475.439286: skb:kfree_skb: ... location=tcp_validate_incoming+0x4f0 reason: TCP_RFC7323_PAWS_ACK
...
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Link: https://patch.msgid.link/20250113135558.3180360-3-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Following patch is adding a new drop_reason to tcp_validate_incoming().
Change tcp_disordered_ack() to not return a boolean anymore,
but a drop reason.
Change its name to tcp_disordered_ack_check()
Refactor tcp_validate_incoming() to ease the code
review of the following patch, and reduce indentation
level.
This patch is a refactor, with no functional change.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Link: https://patch.msgid.link/20250113135558.3180360-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We previously hooked the tcp_drop_reason() function using BPF to monitor
TCP drop reasons. However, after upgrading our compiler from GCC 9 to GCC
11, tcp_drop_reason() is now inlined, preventing us from hooking into it.
To address this, it would be beneficial to make noinline explicitly for
tracing.
Link: https://lore.kernel.org/netdev/CANn89iJuShCmidCi_ZkYABtmscwbVjhuDta1MS5LxV_4H9tKOA@mail.gmail.com/
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
Cc: Menglong Dong <menglong8.dong@gmail.com>
Link: https://patch.msgid.link/20241024093742.87681-3-laoar.shao@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Fix tcp_rcv_synrecv_state_fastopen() to not zero retrans_stamp
if retransmits are outstanding.
tcp_fastopen_synack_timer() sets retrans_stamp, so typically we'll
need to zero retrans_stamp here to prevent spurious
retransmits_timed_out(). The logic to zero retrans_stamp is from this
2019 commit:
commit cd736d8b67fb ("tcp: fix retrans timestamp on passive Fast Open")
However, in the corner case where the ACK of our TFO SYNACK carried
some SACK blocks that caused us to enter TCP_CA_Recovery then that
non-zero retrans_stamp corresponds to the active fast recovery, and we
need to leave retrans_stamp with its current non-zero value, for
correct ETIMEDOUT and undo behavior.
Fixes: cd736d8b67fb ("tcp: fix retrans timestamp on passive Fast Open")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20241001200517.2756803-4-ncardwell.sw@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Fix tcp_enter_recovery() so that if there are no retransmits out then
we zero retrans_stamp when entering fast recovery. This is necessary
to fix two buggy behaviors.
Currently a non-zero retrans_stamp value can persist across multiple
back-to-back loss recovery episodes. This is because we generally only
clears retrans_stamp if we are completely done with loss recoveries,
and get to tcp_try_to_open() and find !tcp_any_retrans_done(sk). This
behavior causes two bugs:
(1) When a loss recovery episode (CA_Loss or CA_Recovery) is followed
immediately by a new CA_Recovery, the retrans_stamp value can persist
and can be a time before this new CA_Recovery episode starts. That
means that timestamp-based undo will be using the wrong retrans_stamp
(a value that is too old) when comparing incoming TS ecr values to
retrans_stamp to see if the current fast recovery episode can be
undone.
(2) If there is a roughly minutes-long sequence of back-to-back fast
recovery episodes, one after another (e.g. in a shallow-buffered or
policed bottleneck), where each fast recovery successfully makes
forward progress and recovers one window of sequence space (but leaves
at least one retransmit in flight at the end of the recovery),
followed by several RTOs, then the ETIMEDOUT check may be using the
wrong retrans_stamp (a value set at the start of the first fast
recovery in the sequence). This can cause a very premature ETIMEDOUT,
killing the connection prematurely.
This commit changes the code to zero retrans_stamp when entering fast
recovery, when this is known to be safe (no retransmits are out in the
network). That ensures that when starting a fast recovery episode, and
it is safe to do so, retrans_stamp is set when we send the fast
retransmit packet. That addresses both bug (1) and bug (2) by ensuring
that (if no retransmits are out when we start a fast recovery) we use
the initial fast retransmit of this fast recovery as the time value
for undo and ETIMEDOUT calculations.
This makes intuitive sense, since the start of a new fast recovery
episode (in a scenario where no lost packets are out in the network)
means that the connection has made forward progress since the last RTO
or fast recovery, and we should thus "restart the clock" used for both
undo and ETIMEDOUT logic.
Note that if when we start fast recovery there *are* retransmits out
in the network, there can still be undesirable (1)/(2) issues. For
example, after this patch we can still have the (1) and (2) problems
in cases like this:
+ round 1: sender sends flight 1
+ round 2: sender receives SACKs and enters fast recovery 1,
retransmits some packets in flight 1 and then sends some new data as
flight 2
+ round 3: sender receives some SACKs for flight 2, notes losses, and
retransmits some packets to fill the holes in flight 2
+ fast recovery has some lost retransmits in flight 1 and continues
for one or more rounds sending retransmits for flight 1 and flight 2
+ fast recovery 1 completes when snd_una reaches high_seq at end of
flight 1
+ there are still holes in the SACK scoreboard in flight 2, so we
enter fast recovery 2, but some retransmits in the flight 2 sequence
range are still in flight (retrans_out > 0), so we can't execute the
new retrans_stamp=0 added here to clear retrans_stamp
It's not yet clear how to fix these remaining (1)/(2) issues in an
efficient way without breaking undo behavior, given that retrans_stamp
is currently used for undo and ETIMEDOUT. Perhaps the optimal (but
expensive) strategy would be to set retrans_stamp to the timestamp of
the earliest outstanding retransmit when entering fast recovery. But
at least this commit makes things better.
Note that this does not change the semantics of retrans_stamp; it
simply makes retrans_stamp accurate in some cases where it was not
before:
(1) Some loss recovery, followed by an immediate entry into a fast
recovery, where there are no retransmits out when entering the fast
recovery.
(2) When a TFO server has a SYNACK retransmit that sets retrans_stamp,
and then the ACK that completes the 3-way handshake has SACK blocks
that trigger a fast recovery. In this case when entering fast recovery
we want to zero out the retrans_stamp from the TFO SYNACK retransmit,
and set the retrans_stamp based on the timestamp of the fast recovery.
We introduce a tcp_retrans_stamp_cleanup() helper, because this
two-line sequence already appears in 3 places and is about to appear
in 2 more as a result of this bug fix patch series. Once this bug fix
patches series in the net branch makes it into the net-next branch
we'll update the 3 other call sites to use the new helper.
This is a long-standing issue. The Fixes tag below is chosen to be the
oldest commit at which the patch will apply cleanly, which is from
Linux v3.5 in 2012.
Fixes: 1fbc340514fc ("tcp: early retransmit: tcp_enter_recovery()")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20241001200517.2756803-3-ncardwell.sw@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Fix the TCP loss recovery undo logic in tcp_packet_delayed() so that
it can trigger undo even if TSQ prevents a fast recovery episode from
reaching tcp_retransmit_skb().
Geumhwan Yu <geumhwan.yu@samsung.com> recently reported that after
this commit from 2019:
commit bc9f38c8328e ("tcp: avoid unconditional congestion window undo
on SYN retransmit")
...and before this fix we could have buggy scenarios like the
following:
+ Due to reordering, a TCP connection receives some SACKs and enters a
spurious fast recovery.
+ TSQ prevents all invocations of tcp_retransmit_skb(), because many
skbs are queued in lower layers of the sending machine's network
stack; thus tp->retrans_stamp remains 0.
+ The connection receives a TCP timestamp ECR value echoing a
timestamp before the fast recovery, indicating that the fast
recovery was spurious.
+ The connection fails to undo the spurious fast recovery because
tp->retrans_stamp is 0, and thus tcp_packet_delayed() returns false,
due to the new logic in the 2019 commit: commit bc9f38c8328e ("tcp:
avoid unconditional congestion window undo on SYN retransmit")
This fix tweaks the logic to be more similar to the
tcp_packet_delayed() logic before bc9f38c8328e, except that we take
care not to be fooled by the FLAG_SYN_ACKED code path zeroing out
tp->retrans_stamp (the bug noted and fixed by Yuchung in
bc9f38c8328e).
Note that this returns the high-level behavior of tcp_packet_delayed()
to again match the comment for the function, which says: "Nothing was
retransmitted or returned timestamp is less than timestamp of the
first retransmission." Note that this comment is in the original
2005-04-16 Linux git commit, so this is evidently long-standing
behavior.
Fixes: bc9f38c8328e ("tcp: avoid unconditional congestion window undo on SYN retransmit")
Reported-by: Geumhwan Yu <geumhwan.yu@samsung.com>
Diagnosed-by: Geumhwan Yu <geumhwan.yu@samsung.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20241001200517.2756803-2-ncardwell.sw@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
asm/unaligned.h is always an include of asm-generic/unaligned.h;
might as well move that thing to linux/unaligned.h and include
that - there's nothing arch-specific in that header.
auto-generated by the following:
for i in `git grep -l -w asm/unaligned.h`; do
sed -i -e "s/asm\/unaligned.h/linux\/unaligned.h/" $i
done
for i in `git grep -l -w asm-generic/unaligned.h`; do
sed -i -e "s/asm-generic\/unaligned.h/linux\/unaligned.h/" $i
done
git mv include/asm-generic/unaligned.h include/linux/unaligned.h
git mv tools/include/asm-generic/unaligned.h tools/include/linux/unaligned.h
sed -i -e "/unaligned.h/d" include/asm-generic/Kbuild
sed -i -e "s/__ASM_GENERIC/__LINUX/" include/linux/unaligned.h tools/include/linux/unaligned.h
For device memory TCP, we expect the skb headers to be available in host
memory for access, and we expect the skb frags to be in device memory
and unaccessible to the host. We expect there to be no mixing and
matching of device memory frags (unaccessible) with host memory frags
(accessible) in the same skb.
Add a skb->devmem flag which indicates whether the frags in this skb
are device memory frags or not.
__skb_fill_netmem_desc() now checks frags added to skbs for net_iov,
and marks the skb as skb->devmem accordingly.
Add checks through the network stack to avoid accessing the frags of
devmem skbs and avoid coalescing devmem skbs with non devmem skbs.
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Kaiyuan Zhang <kaiyuanz@google.com>
Signed-off-by: Mina Almasry <almasrymina@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Link: https://patch.msgid.link/20240910171458.219195-9-almasrymina@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This patch is based on the discussions between Neal Cardwell and
Eric Dumazet in the link
https://lore.kernel.org/netdev/20240726204105.1466841-1-quic_subashab@quicinc.com/
It was correctly pointed out that tp->window_clamp would not be
updated in cases where net.ipv4.tcp_moderate_rcvbuf=0 or if
(copied <= tp->rcvq_space.space). While it is expected for most
setups to leave the sysctl enabled, the latter condition may
not end up hitting depending on the TCP receive queue size and
the pattern of arriving data.
The updated check should be hit only on initial MSS update from
TCP_MIN_MSS to measured MSS value and subsequently if there was
an update to a larger value.
Fixes: 05f76b2d634e ("tcp: Adjust clamping window for applications specifying SO_RCVBUF")
Signed-off-by: Sean Tranchetti <quic_stranche@quicinc.com>
Signed-off-by: Subash Abhinov Kasiviswanathan <quic_subashab@quicinc.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tp->scaling_ratio is not updated based on skb->len/skb->truesize once
SO_RCVBUF is set leading to the maximum window scaling to be 25% of
rcvbuf after
commit dfa2f0483360 ("tcp: get rid of sysctl_tcp_adv_win_scale")
and 50% of rcvbuf after
commit 697a6c8cec03 ("tcp: increase the default TCP scaling ratio").
50% tries to emulate the behavior of older kernels using
sysctl_tcp_adv_win_scale with default value.
Systems which were using a different values of sysctl_tcp_adv_win_scale
in older kernels ended up seeing reduced download speeds in certain
cases as covered in https://lists.openwall.net/netdev/2024/05/15/13
While the sysctl scheme is no longer acceptable, the value of 50% is
a bit conservative when the skb->len/skb->truesize ratio is later
determined to be ~0.66.
Applications not specifying SO_RCVBUF update the window scaling and
the receiver buffer every time data is copied to userspace. This
computation is now used for applications setting SO_RCVBUF to update
the maximum window scaling while ensuring that the receive buffer
is within the application specified limit.
Fixes: dfa2f0483360 ("tcp: get rid of sysctl_tcp_adv_win_scale")
Signed-off-by: Sean Tranchetti <quic_stranche@quicinc.com>
Signed-off-by: Subash Abhinov Kasiviswanathan <quic_subashab@quicinc.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The 'Fixes' commit recently changed the behaviour of TCP by skipping the
processing of the 3rd ACK when a sk->sk_socket is set. The goal was to
skip tcp_ack_snd_check() in tcp_rcv_state_process() not to send an
unnecessary ACK in case of simultaneous connect(). Unfortunately, that
had an impact on TFO and MPTCP.
I started to look at the impact on MPTCP, because the MPTCP CI found
some issues with the MPTCP Packetdrill tests [1]. Then Paolo Abeni
suggested me to look at the impact on TFO with "plain" TCP.
For MPTCP, when receiving the 3rd ACK of a request adding a new path
(MP_JOIN), sk->sk_socket will be set, and point to the MPTCP sock that
has been created when the MPTCP connection got established before with
the first path. The newly added 'goto' will then skip the processing of
the segment text (step 7) and not go through tcp_data_queue() where the
MPTCP options are validated, and some actions are triggered, e.g.
sending the MPJ 4th ACK [2] as demonstrated by the new errors when
running a packetdrill test [3] establishing a second subflow.
This doesn't fully break MPTCP, mainly the 4th MPJ ACK that will be
delayed. Still, we don't want to have this behaviour as it delays the
switch to the fully established mode, and invalid MPTCP options in this
3rd ACK will not be caught any more. This modification also affects the
MPTCP + TFO feature as well, and being the reason why the selftests
started to be unstable the last few days [4].
For TFO, the existing 'basic-cookie-not-reqd' test [5] was no longer
passing: if the 3rd ACK contains data, and the connection is accept()ed
before receiving them, these data would no longer be processed, and thus
not ACKed.
One last thing about MPTCP, in case of simultaneous connect(), a
fallback to TCP will be done, which seems fine:
`../common/defaults.sh`
0 socket(..., SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_MPTCP) = 3
+0 connect(3, ..., ...) = -1 EINPROGRESS (Operation now in progress)
+0 > S 0:0(0) <mss 1460, sackOK, TS val 100 ecr 0, nop, wscale 8, mpcapable v1 flags[flag_h] nokey>
+0 < S 0:0(0) win 1000 <mss 1460, sackOK, TS val 407 ecr 0, nop, wscale 8, mpcapable v1 flags[flag_h] nokey>
+0 > S. 0:0(0) ack 1 <mss 1460, sackOK, TS val 330 ecr 0, nop, wscale 8, mpcapable v1 flags[flag_h] nokey>
+0 < S. 0:0(0) ack 1 win 65535 <mss 1460, sackOK, TS val 700 ecr 100, nop, wscale 8, mpcapable v1 flags[flag_h] key[skey=2]>
+0 > . 1:1(0) ack 1 <nop, nop, TS val 845707014 ecr 700, nop, nop, sack 0:1>
Simultaneous SYN-data crossing is also not supported by TFO, see [6].
Kuniyuki Iwashima suggested to restrict the processing to SYN+ACK only:
that's a more generic solution than the one initially proposed, and
also enough to fix the issues described above.
Later on, Eric Dumazet mentioned that an ACK should still be sent in
reaction to the second SYN+ACK that is received: not sending a DUPACK
here seems wrong and could hurt:
0 socket(..., SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 3
+0 connect(3, ..., ...) = -1 EINPROGRESS (Operation now in progress)
+0 > S 0:0(0) <mss 1460, sackOK, TS val 1000 ecr 0,nop,wscale 8>
+0 < S 0:0(0) win 1000 <mss 1000, sackOK, nop, nop>
+0 > S. 0:0(0) ack 1 <mss 1460, sackOK, TS val 3308134035 ecr 0,nop,wscale 8>
+0 < S. 0:0(0) ack 1 win 1000 <mss 1000, sackOK, nop, nop>
+0 > . 1:1(0) ack 1 <nop, nop, sack 0:1> // <== Here
So in this version, the 'goto consume' is dropped, to always send an ACK
when switching from TCP_SYN_RECV to TCP_ESTABLISHED. This ACK will be
seen as a DUPACK -- with DSACK if SACK has been negotiated -- in case of
simultaneous SYN crossing: that's what is expected here.
Link: https://github.com/multipath-tcp/mptcp_net-next/actions/runs/9936227696 [1]
Link: https://datatracker.ietf.org/doc/html/rfc8684#fig_tokens [2]
Link: https://github.com/multipath-tcp/packetdrill/blob/mptcp-net-next/gtests/net/mptcp/syscalls/accept.pkt#L28 [3]
Link: https://netdev.bots.linux.dev/contest.html?executor=vmksft-mptcp-dbg&test=mptcp-connect-sh [4]
Link: https://github.com/google/packetdrill/blob/master/gtests/net/tcp/fastopen/server/basic-cookie-not-reqd.pkt#L21 [5]
Link: https://github.com/google/packetdrill/blob/master/gtests/net/tcp/fastopen/client/simultaneous-fast-open.pkt [6]
Fixes: 23e89e8ee7be ("tcp: Don't drop SYN+ACK for simultaneous connect().")
Suggested-by: Paolo Abeni <pabeni@redhat.com>
Suggested-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Matthieu Baerts (NGI0) <matttbe@kernel.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20240724-upstream-net-next-20240716-tcp-3rd-ack-consume-sk_socket-v3-1-d48339764ce9@kernel.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
RFC 9293 states that in the case of simultaneous connect(), the connection
gets established when SYN+ACK is received. [0]
TCP Peer A TCP Peer B
1. CLOSED CLOSED
2. SYN-SENT --> <SEQ=100><CTL=SYN> ...
3. SYN-RECEIVED <-- <SEQ=300><CTL=SYN> <-- SYN-SENT
4. ... <SEQ=100><CTL=SYN> --> SYN-RECEIVED
5. SYN-RECEIVED --> <SEQ=100><ACK=301><CTL=SYN,ACK> ...
6. ESTABLISHED <-- <SEQ=300><ACK=101><CTL=SYN,ACK> <-- SYN-RECEIVED
7. ... <SEQ=100><ACK=301><CTL=SYN,ACK> --> ESTABLISHED
However, since commit 0c24604b68fc ("tcp: implement RFC 5961 4.2"), such a
SYN+ACK is dropped in tcp_validate_incoming() and responded with Challenge
ACK.
For example, the write() syscall in the following packetdrill script fails
with -EAGAIN, and wrong SNMP stats get incremented.
0 socket(..., SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 3
+0 connect(3, ..., ...) = -1 EINPROGRESS (Operation now in progress)
+0 > S 0:0(0) <mss 1460,sackOK,TS val 1000 ecr 0,nop,wscale 8>
+0 < S 0:0(0) win 1000 <mss 1000>
+0 > S. 0:0(0) ack 1 <mss 1460,sackOK,TS val 3308134035 ecr 0,nop,wscale 8>
+0 < S. 0:0(0) ack 1 win 1000
+0 write(3, ..., 100) = 100
+0 > P. 1:101(100) ack 1
--
# packetdrill cross-synack.pkt
cross-synack.pkt:13: runtime error in write call: Expected result 100 but got -1 with errno 11 (Resource temporarily unavailable)
# nstat
...
TcpExtTCPChallengeACK 1 0.0
TcpExtTCPSYNChallenge 1 0.0
The problem is that bpf_skops_established() is triggered by the Challenge
ACK instead of SYN+ACK. This causes the bpf prog to miss the chance to
check if the peer supports a TCP option that is expected to be exchanged
in SYN and SYN+ACK.
Let's accept a bare SYN+ACK for active-open TCP_SYN_RECV sockets to avoid
such a situation.
Note that tcp_ack_snd_check() in tcp_rcv_state_process() is skipped not to
send an unnecessary ACK, but this could be a bit risky for net.git, so this
targets for net-next.
Link: https://www.rfc-editor.org/rfc/rfc9293.html#section-3.5-7 [0]
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20240710171246.87533-2-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Cross-merge networking fixes after downstream PR.
Conflicts:
net/sched/act_ct.c
26488172b029 ("net/sched: Fix UAF when resolving a clash")
3abbd7ed8b76 ("act_ct: prepare for stolen verdict coming from conntrack and nat engine")
No adjacent changes.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Loss recovery undo_retrans bookkeeping had a long-standing bug where a
DSACK from a spurious TLP retransmit packet could cause an erroneous
undo of a fast recovery or RTO recovery that repaired a single
really-lost packet (in a sequence range outside that of the TLP
retransmit). Basically, because the loss recovery state machine didn't
account for the fact that it sent a TLP retransmit, the DSACK for the
TLP retransmit could erroneously be implicitly be interpreted as
corresponding to the normal fast recovery or RTO recovery retransmit
that plugged a real hole, thus resulting in an improper undo.
For example, consider the following buggy scenario where there is a
real packet loss but the congestion control response is improperly
undone because of this bug:
+ send packets P1, P2, P3, P4
+ P1 is really lost
+ send TLP retransmit of P4
+ receive SACK for original P2, P3, P4
+ enter fast recovery, fast-retransmit P1, increment undo_retrans to 1
+ receive DSACK for TLP P4, decrement undo_retrans to 0, undo (bug!)
+ receive cumulative ACK for P1-P4 (fast retransmit plugged real hole)
The fix: when we initialize undo machinery in tcp_init_undo(), if
there is a TLP retransmit in flight, then increment tp->undo_retrans
so that we make sure that we receive a DSACK corresponding to the TLP
retransmit, as well as DSACKs for all later normal retransmits, before
triggering a loss recovery undo. Note that we also have to move the
line that clears tp->tlp_high_seq for RTO recovery, so that upon RTO
we remember the tp->tlp_high_seq value until tcp_init_undo() and clear
it only afterward.
Also note that the bug dates back to the original 2013 TLP
implementation, commit 6ba8a3b19e76 ("tcp: Tail loss probe (TLP)").
However, this patch will only compile and work correctly with kernels
that have tp->tlp_retrans, which was added only in v5.8 in 2020 in
commit 76be93fc0702 ("tcp: allow at most one TLP probe per flight").
So we associate this fix with that later commit.
Fixes: 76be93fc0702 ("tcp: allow at most one TLP probe per flight")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Kevin Yang <yyd@google.com>
Link: https://patch.msgid.link/20240703171246.1739561-1-ncardwell.sw@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When we process segments with TCP AO, we don't check it in
tcp_parse_options(). Thus, opt_rx->saw_unknown is set to 1,
which unconditionally triggers the BPF TCP option parser.
Let's avoid the unnecessary BPF invocation.
Fixes: 0a3a809089eb ("net/tcp: Verify inbound TCP-AO signed segments")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Acked-by: Dmitry Safonov <0x7f454c46@gmail.com>
Link: https://patch.msgid.link/20240703033508.6321-1-kuniyu@amazon.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
In some production workloads we noticed that connections could
sometimes close extremely prematurely with ETIMEDOUT after
transmitting only 1 TLP and RTO retransmission (when we would normally
expect roughly tcp_retries2 = TCP_RETR2 = 15 RTOs before a connection
closes with ETIMEDOUT).
From tracing we determined that these workloads can suffer from a
scenario where in fast recovery, after some retransmits, a DSACK undo
can happen at a point where the scoreboard is totally clear (we have
retrans_out == sacked_out == lost_out == 0). In such cases, calling
tcp_try_keep_open() means that we do not execute any code path that
clears tp->retrans_stamp to 0. That means that tp->retrans_stamp can
remain erroneously set to the start time of the undone fast recovery,
even after the fast recovery is undone. If minutes or hours elapse,
and then a TLP/RTO/RTO sequence occurs, then the start_ts value in
retransmits_timed_out() (which is from tp->retrans_stamp) will be
erroneously ancient (left over from the fast recovery undone via
DSACKs). Thus this ancient tp->retrans_stamp value can cause the
connection to die very prematurely with ETIMEDOUT via
tcp_write_err().
The fix: we change DSACK undo in fast recovery (TCP_CA_Recovery) to
call tcp_try_to_open() instead of tcp_try_keep_open(). This ensures
that if no retransmits are in flight at the time of DSACK undo in fast
recovery then we properly zero retrans_stamp. Note that calling
tcp_try_to_open() is more consistent with other loss recovery
behavior, since normal fast recovery (CA_Recovery) and RTO recovery
(CA_Loss) both normally end when tp->snd_una meets or exceeds
tp->high_seq and then in tcp_fastretrans_alert() the "default" switch
case executes tcp_try_to_open(). Also note that by inspection this
change to call tcp_try_to_open() implies at least one other nice bug
fix, where now an ECE-marked DSACK that causes an undo will properly
invoke tcp_enter_cwr() rather than ignoring the ECE mark.
Fixes: c7d9d6a185a7 ("tcp: undo on DSACK during recovery")
Signed-off-by: Neal Cardwell <ncardwell@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>
Cross-merge networking fixes after downstream PR.
No conflicts.
Adjacent changes:
e3f02f32a050 ("ionic: fix kernel panic due to multi-buffer handling")
d9c04209990b ("ionic: Mark error paths in the data path as unlikely")
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Testing determined that the recent commit 9e046bb111f1 ("tcp: clear
tp->retrans_stamp in tcp_rcv_fastopen_synack()") has a race, and does
not always ensure retrans_stamp is 0 after a TFO payload retransmit.
If transmit completion for the SYN+data skb happens after the client
TCP stack receives the SYNACK (which sometimes happens), then
retrans_stamp can erroneously remain non-zero for the lifetime of the
connection, causing a premature ETIMEDOUT later.
Testing and tracing showed that the buggy scenario is the following
somewhat tricky sequence:
+ Client attempts a TFO handshake. tcp_send_syn_data() sends SYN + TFO
cookie + data in a single packet in the syn_data skb. It hands the
syn_data skb to tcp_transmit_skb(), which makes a clone. Crucially,
it then reuses the same original (non-clone) syn_data skb,
transforming it by advancing the seq by one byte and removing the
FIN bit, and enques the resulting payload-only skb in the
sk->tcp_rtx_queue.
+ Client sets retrans_stamp to the start time of the three-way
handshake.
+ Cookie mismatches or server has TFO disabled, and server only ACKs
SYN.
+ tcp_ack() sees SYN is acked, tcp_clean_rtx_queue() clears
retrans_stamp.
+ Since the client SYN was acked but not the payload, the TFO failure
code path in tcp_rcv_fastopen_synack() tries to retransmit the
payload skb. However, in some cases the transmit completion for the
clone of the syn_data (which had SYN + TFO cookie + data) hasn't
happened. In those cases, skb_still_in_host_queue() returns true
for the retransmitted TFO payload, because the clone of the syn_data
skb has not had its tx completetion.
+ Because skb_still_in_host_queue() finds skb_fclone_busy() is true,
it sets the TSQ_THROTTLED bit and the retransmit does not happen in
the tcp_rcv_fastopen_synack() call chain.
+ The tcp_rcv_fastopen_synack() code next implicitly assumes the
retransmit process is finished, and sets retrans_stamp to 0 to clear
it, but this is later overwritten (see below).
+ Later, upon tx completion, tcp_tsq_write() calls
tcp_xmit_retransmit_queue(), which puts the retransmit in flight and
sets retrans_stamp to a non-zero value.
+ The client receives an ACK for the retransmitted TFO payload data.
+ Since we're in CA_Open and there are no dupacks/SACKs/DSACKs/ECN to
make tcp_ack_is_dubious() true and make us call
tcp_fastretrans_alert() and reach a code path that clears
retrans_stamp, retrans_stamp stays nonzero.
+ Later, if there is a TLP, RTO, RTO sequence, then the connection
will suffer an early ETIMEDOUT due to the erroneously ancient
retrans_stamp.
The fix: this commit refactors the code to have
tcp_rcv_fastopen_synack() retransmit by reusing the relevant parts of
tcp_simple_retransmit() that enter CA_Loss (without changing cwnd) and
call tcp_xmit_retransmit_queue(). We have tcp_simple_retransmit() and
tcp_rcv_fastopen_synack() share code in this way because in both cases
we get a packet indicating non-congestion loss (MTU reduction or TFO
failure) and thus in both cases we want to retransmit as many packets
as cwnd allows, without reducing cwnd. And given that retransmits will
set retrans_stamp to a non-zero value (and may do so in a later
calling context due to TSQ), we also want to enter CA_Loss so that we
track when all retransmitted packets are ACked and clear retrans_stamp
when that happens (to ensure later recurring RTOs are using the
correct retrans_stamp and don't declare ETIMEDOUT prematurely).
Fixes: 9e046bb111f1 ("tcp: clear tp->retrans_stamp in tcp_rcv_fastopen_synack()")
Fixes: a7abf3cd76e1 ("tcp: consider using standard rtx logic in tcp_rcv_fastopen_synack()")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Link: https://patch.msgid.link/20240624144323.2371403-1-ncardwell.sw@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When bonding is configured in BOND_MODE_BROADCAST mode, if two identical
SYN packets are received at the same time and processed on different CPUs,
it can potentially create the same sk (sock) but two different reqsk
(request_sock) in tcp_conn_request().
These two different reqsk will respond with two SYNACK packets, and since
the generation of the seq (ISN) incorporates a timestamp, the final two
SYNACK packets will have different seq values.
The consequence is that when the Client receives and replies with an ACK
to the earlier SYNACK packet, we will reset(RST) it.
========================================================================
This behavior is consistently reproducible in my local setup,
which comprises:
| NETA1 ------ NETB1 |
PC_A --- bond --- | | --- bond --- PC_B
| NETA2 ------ NETB2 |
- PC_A is the Server and has two network cards, NETA1 and NETA2. I have
bonded these two cards using BOND_MODE_BROADCAST mode and configured
them to be handled by different CPU.
- PC_B is the Client, also equipped with two network cards, NETB1 and
NETB2, which are also bonded and configured in BOND_MODE_BROADCAST mode.
If the client attempts a TCP connection to the server, it might encounter
a failure. Capturing packets from the server side reveals:
10.10.10.10.45182 > localhost: Flags [S], seq 320236027,
10.10.10.10.45182 > localhost: Flags [S], seq 320236027,
localhost > 10.10.10.10.45182: Flags [S.], seq 2967855116,
localhost > 10.10.10.10.45182: Flags [S.], seq 2967855123, <==
10.10.10.10.45182 > localhost: Flags [.], ack 4294967290,
10.10.10.10.45182 > localhost: Flags [.], ack 4294967290,
localhost > 10.10.10.10.45182: Flags [R], seq 2967855117, <==
localhost > 10.10.10.10.45182: Flags [R], seq 2967855117,
Two SYNACKs with different seq numbers are sent by localhost,
resulting in an anomaly.
========================================================================
The attempted solution is as follows:
Add a return value to inet_csk_reqsk_queue_hash_add() to confirm if the
ehash insertion is successful (Up to now, the reason for unsuccessful
insertion is that a reqsk for the same connection has already been
inserted). If the insertion fails, release the reqsk.
Due to the refcnt, Kuniyuki suggests also adding a return value check
for the DCCP module; if ehash insertion fails, indicating a successful
insertion of the same connection, simply release the reqsk as well.
Simultaneously, In the reqsk_queue_hash_req(), the start of the
req->rsk_timer is adjusted to be after successful insertion.
Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Signed-off-by: luoxuanqiang <luoxuanqiang@kylinos.cn>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240621013929.1386815-1-luoxuanqiang@kylinos.cn
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Cross-merge networking fixes after downstream PR.
Conflicts:
drivers/net/ethernet/broadcom/bnxt/bnxt.c
1e7962114c10 ("bnxt_en: Restore PTP tx_avail count in case of skb_pad() error")
165f87691a89 ("bnxt_en: add timestamping statistics support")
No adjacent changes.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Replace kfree_skb_reason with sk_skb_reason_drop and pass the receiving
socket to the tracepoint.
Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/r/202406011539.jhwBd7DX-lkp@intel.com/
Signed-off-by: Yan Zhai <yan@cloudflare.com>
Acked-by: Jesper Dangaard Brouer <hawk@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some applications were reporting ETIMEDOUT errors on apparently
good looking flows, according to packet dumps.
We were able to root cause the issue to an accidental setting
of tp->retrans_stamp in the following scenario:
- client sends TFO SYN with data.
- server has TFO disabled, ACKs only SYN but not payload.
- client receives SYNACK covering only SYN.
- tcp_ack() eats SYN and sets tp->retrans_stamp to 0.
- tcp_rcv_fastopen_synack() calls tcp_xmit_retransmit_queue()
to retransmit TFO payload w/o SYN, sets tp->retrans_stamp to "now",
but we are not in any loss recovery state.
- TFO payload is ACKed.
- we are not in any loss recovery state, and don't see any dupacks,
so we don't get to any code path that clears tp->retrans_stamp.
- tp->retrans_stamp stays non-zero for the lifetime of the connection.
- after first RTO, tcp_clamp_rto_to_user_timeout() clamps second RTO
to 1 jiffy due to bogus tp->retrans_stamp.
- on clamped RTO with non-zero icsk_retransmits, retransmits_timed_out()
sets start_ts from tp->retrans_stamp from TFO payload retransmit
hours/days ago, and computes bogus long elapsed time for loss recovery,
and suffers ETIMEDOUT early.
Fixes: a7abf3cd76e1 ("tcp: consider using standard rtx logic in tcp_rcv_fastopen_synack()")
CC: stable@vger.kernel.org
Co-developed-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Co-developed-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240614130615.396837-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Instead of forcing userspace to parse dmesg (that's what currently is
happening, at least in codebase of my current company), provide a better
way, that can be enabled/disabled in runtime.
Currently, there are already tcp events, add hashing related ones there,
too. Rasdaemon currently exercises net_dev_xmit_timeout,
devlink_health_report, but it'll be trivial to teach it to deal with
failed hashes. Otherwise, BGP may trace/log them itself. Especially
exciting for possible investigations is key rotation (RNext_key
requests).
Suggested-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Dmitry Safonov <0x7f454c46@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
inet_reqsk_alloc() does not belong to tcp_input.c,
move it to inet_connection_sock.c instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
