| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
We observed high 99 and 99.9% latencies when doing RPCs with DCTCP. The
problem is triggered when the last packet of a request arrives CE
marked. The reply will carry the ECE mark causing TCP to shrink its cwnd
to 1 (because there are no packets in flight). When the 1st packet of
the next request arrives, the ACK was sometimes delayed even though it
is CWR marked, adding up to 40ms to the RPC latency.
This patch insures that CWR marked data packets arriving will be acked
immediately.
Packetdrill script to reproduce the problem:
0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
0.000 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
0.000 setsockopt(3, SOL_TCP, TCP_CONGESTION, "dctcp", 5) = 0
0.000 bind(3, ..., ...) = 0
0.000 listen(3, 1) = 0
0.100 < [ect0] SEW 0:0(0) win 32792 <mss 1000,sackOK,nop,nop,nop,wscale 7>
0.100 > SE. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 8>
0.110 < [ect0] . 1:1(0) ack 1 win 257
0.200 accept(3, ..., ...) = 4
0.200 < [ect0] . 1:1001(1000) ack 1 win 257
0.200 > [ect01] . 1:1(0) ack 1001
0.200 write(4, ..., 1) = 1
0.200 > [ect01] P. 1:2(1) ack 1001
0.200 < [ect0] . 1001:2001(1000) ack 2 win 257
0.200 write(4, ..., 1) = 1
0.200 > [ect01] P. 2:3(1) ack 2001
0.200 < [ect0] . 2001:3001(1000) ack 3 win 257
0.200 < [ect0] . 3001:4001(1000) ack 3 win 257
0.200 > [ect01] . 3:3(0) ack 4001
0.210 < [ce] P. 4001:4501(500) ack 3 win 257
+0.001 read(4, ..., 4500) = 4500
+0 write(4, ..., 1) = 1
+0 > [ect01] PE. 3:4(1) ack 4501
+0.010 < [ect0] W. 4501:5501(1000) ack 4 win 257
// Previously the ACK sequence below would be 4501, causing a long RTO
+0.040~+0.045 > [ect01] . 4:4(0) ack 5501 // delayed ack
+0.311 < [ect0] . 5501:6501(1000) ack 4 win 257 // More data
+0 > [ect01] . 4:4(0) ack 6501 // now acks everything
+0.500 < F. 9501:9501(0) ack 4 win 257
Modified based on comments by Neal Cardwell <ncardwell@google.com>
Signed-off-by: Lawrence Brakmo <brakmo@fb.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
In case skb in out_or_order_queue is the result of
multiple skbs coalescing, we would like to get a proper gso_segs
counter tracking, so that future tcp_drop() can report an accurate
number.
I chose to not implement this tracking for skbs in receive queue,
since they are not dropped, unless socket is disconnected.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
In order to be able to give better diagnostics and detect
malicious traffic, we need to have better sk->sk_drops tracking.
Fixes: 9f5afeae5152 ("tcp: use an RB tree for ooo receive queue")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
In case an attacker feeds tiny packets completely out of order,
tcp_collapse_ofo_queue() might scan the whole rb-tree, performing
expensive copies, but not changing socket memory usage at all.
1) Do not attempt to collapse tiny skbs.
2) Add logic to exit early when too many tiny skbs are detected.
We prefer not doing aggressive collapsing (which copies packets)
for pathological flows, and revert to tcp_prune_ofo_queue() which
will be less expensive.
In the future, we might add the possibility of terminating flows
that are proven to be malicious.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Right after a TCP flow is created, receiving tiny out of order
packets allways hit the condition :
if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
tcp_clamp_window(sk);
tcp_clamp_window() increases sk_rcvbuf to match sk_rmem_alloc
(guarded by tcp_rmem[2])
Calling tcp_collapse_ofo_queue() in this case is not useful,
and offers a O(N^2) surface attack to malicious peers.
Better not attempt anything before full queue capacity is reached,
forcing attacker to spend lots of resource and allow us to more
easily detect the abuse.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Juha-Matti Tilli reported that malicious peers could inject tiny
packets in out_of_order_queue, forcing very expensive calls
to tcp_collapse_ofo_queue() and tcp_prune_ofo_queue() for
every incoming packet. out_of_order_queue rb-tree can contain
thousands of nodes, iterating over all of them is not nice.
Before linux-4.9, we would have pruned all packets in ofo_queue
in one go, every XXXX packets. XXXX depends on sk_rcvbuf and skbs
truesize, but is about 7000 packets with tcp_rmem[2] default of 6 MB.
Since we plan to increase tcp_rmem[2] in the future to cope with
modern BDP, can not revert to the old behavior, without great pain.
Strategy taken in this patch is to purge ~12.5 % of the queue capacity.
Fixes: 36a6503fedda ("tcp: refine tcp_prune_ofo_queue() to not drop all packets")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Juha-Matti Tilli <juha-matti.tilli@iki.fi>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Per DCTCP RFC8257 (Section 3.2) the ACK reflecting the CE status change
has to be sent immediately so the sender can respond quickly:
""" When receiving packets, the CE codepoint MUST be processed as follows:
1. If the CE codepoint is set and DCTCP.CE is false, set DCTCP.CE to
true and send an immediate ACK.
2. If the CE codepoint is not set and DCTCP.CE is true, set DCTCP.CE
to false and send an immediate ACK.
"""
Previously DCTCP implementation may continue to delay the ACK. This
patch fixes that to implement the RFC by forcing an immediate ACK.
Tested with this packetdrill script provided by Larry Brakmo
0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
0.000 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
0.000 setsockopt(3, SOL_TCP, TCP_CONGESTION, "dctcp", 5) = 0
0.000 bind(3, ..., ...) = 0
0.000 listen(3, 1) = 0
0.100 < [ect0] SEW 0:0(0) win 32792 <mss 1000,sackOK,nop,nop,nop,wscale 7>
0.100 > SE. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 8>
0.110 < [ect0] . 1:1(0) ack 1 win 257
0.200 accept(3, ..., ...) = 4
+0 setsockopt(4, SOL_SOCKET, SO_DEBUG, [1], 4) = 0
0.200 < [ect0] . 1:1001(1000) ack 1 win 257
0.200 > [ect01] . 1:1(0) ack 1001
0.200 write(4, ..., 1) = 1
0.200 > [ect01] P. 1:2(1) ack 1001
0.200 < [ect0] . 1001:2001(1000) ack 2 win 257
+0.005 < [ce] . 2001:3001(1000) ack 2 win 257
+0.000 > [ect01] . 2:2(0) ack 2001
// Previously the ACK below would be delayed by 40ms
+0.000 > [ect01] E. 2:2(0) ack 3001
+0.500 < F. 9501:9501(0) ack 4 win 257
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
If SACK is not enabled and the first cumulative ACK after the RTO
retransmission covers more than the retransmitted skb, a spurious
FRTO undo will trigger (assuming FRTO is enabled for that RTO).
The reason is that any non-retransmitted segment acknowledged will
set FLAG_ORIG_SACK_ACKED in tcp_clean_rtx_queue even if there is
no indication that it would have been delivered for real (the
scoreboard is not kept with TCPCB_SACKED_ACKED bits in the non-SACK
case so the check for that bit won't help like it does with SACK).
Having FLAG_ORIG_SACK_ACKED set results in the spurious FRTO undo
in tcp_process_loss.
We need to use more strict condition for non-SACK case and check
that none of the cumulatively ACKed segments were retransmitted
to prove that progress is due to original transmissions. Only then
keep FLAG_ORIG_SACK_ACKED set, allowing FRTO undo to proceed in
non-SACK case.
(FLAG_ORIG_SACK_ACKED is planned to be renamed to FLAG_ORIG_PROGRESS
to better indicate its purpose but to keep this change minimal, it
will be done in another patch).
Besides burstiness and congestion control violations, this problem
can result in RTO loop: When the loss recovery is prematurely
undoed, only new data will be transmitted (if available) and
the next retransmission can occur only after a new RTO which in case
of multiple losses (that are not for consecutive packets) requires
one RTO per loss to recover.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Tested-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Larry Brakmo proposal ( https://patchwork.ozlabs.org/patch/935233/
tcp: force cwnd at least 2 in tcp_cwnd_reduction) made us rethink
about our recent patch removing ~16 quick acks after ECN events.
tcp_enter_quickack_mode(sk, 1) makes sure one immediate ack is sent,
but in the case the sender cwnd was lowered to 1, we do not want
to have a delayed ack for the next packet we will receive.
Fixes: 522040ea5fdd ("tcp: do not aggressively quick ack after ECN events")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Neal Cardwell <ncardwell@google.com>
Cc: Lawrence Brakmo <brakmo@fb.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Refactor tcp_ecn_check_ce and __tcp_ecn_check_ce to accept struct sock*
instead of tcp_sock* to clean up type casts. This is a pure refactor
patch.
Signed-off-by: Yousuk Seung <ysseung@google.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>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This is additional to the
commit ea1627c20c34 ("tcp: minor optimizations around tcp_hdr() usage").
At this point, skb->data is same with tcp_hdr() as tcp header has not
been pulled yet. So use the less expensive one to get the tcp header.
Remove the third parameter of tcp_rcv_established() and put it into
the function body.
Furthermore, the local variables are listed as a reverse christmas tree :)
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
ECN signals currently forces TCP to enter quickack mode for
up to 16 (TCP_MAX_QUICKACKS) following incoming packets.
We believe this is not needed, and only sending one immediate ack
for the current packet should be enough.
This should reduce the extra load noticed in DCTCP environments,
after congestion events.
This is part 2 of our effort to reduce pure ACK packets.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
We want to add finer control of the number of ACK packets sent after
ECN events.
This patch is not changing current behavior, it only enables following
change.
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 per netns sysctl allows for TCP SACK compression fine-tuning.
This limits number of SACK that can be compressed.
Using 0 disables SACK compression.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This per netns sysctl allows for TCP SACK compression fine-tuning.
Its default value is 1,000,000, or 1 ms to meet TSO autosizing period.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
When TCP receives an out-of-order packet, it immediately sends
a SACK packet, generating network load but also forcing the
receiver to send 1-MSS pathological packets, increasing its
RTX queue length/depth, and thus processing time.
Wifi networks suffer from this aggressive behavior, but generally
speaking, all these SACK packets add fuel to the fire when networks
are under congestion.
This patch adds a high resolution timer and tp->compressed_ack counter.
Instead of sending a SACK, we program this timer with a small delay,
based on RTT and capped to 1 ms :
delay = min ( 5 % of RTT, 1 ms)
If subsequent SACKs need to be sent while the timer has not yet
expired, we simply increment tp->compressed_ack.
When timer expires, a SACK is sent with the latest information.
Whenever an ACK is sent (if data is sent, or if in-order
data is received) timer is canceled.
Note that tcp_sack_new_ofo_skb() is able to force a SACK to be sent
if the sack blocks need to be shuffled, even if the timer has not
expired.
A new SNMP counter is added in the following patch.
Two other patches add sysctls to allow changing the 1,000,000 and 44
values that this commit hard-coded.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
As explained in commit 9f9843a751d0 ("tcp: properly handle stretch
acks in slow start"), TCP stacks have to consider how many packets
are acknowledged in one single ACK, because of GRO, but also
because of ACK compression or losses.
We plan to add SACK compression in the following patch, we
must therefore not call tcp_enter_quickack_mode()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
An RTO event indicates the head has not been acked for a long time
after its last (re)transmission. But the other packets are not
necessarily lost if they have been only sent recently (for example
due to application limit). This patch would prohibit marking packets
sent within an RTT to be lost on RTO event, using similar logic in
TCP RACK detection.
Normally the head (SND.UNA) would be marked lost since RTO should
fire strictly after the head was sent. An exception is when the
most recent RACK RTT measurement is larger than the (previous)
RTO. To address this exception the head is always marked lost.
Congestion control interaction: since we may not mark every packet
lost, the congestion window may be more than 1 (inflight plus 1).
But only one packet will be retransmitted after RTO, since
tcp_retransmit_timer() calls tcp_retransmit_skb(...,segs=1). The
connection still performs slow start from one packet (with Cubic
congestion control).
This commit was tested in an A/B test with Google web servers,
and showed a reduction of 2% in (spurious) retransmits post
timeout (SlowStartRetrans), and correspondingly reduced DSACKs
(DSACKIgnoredOld) by 7%.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Create and export a new helper tcp_rack_skb_timeout and move tcp_is_rack
to prepare the final RTO change.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Previously when TCP times out, it first updates cwnd and ssthresh,
marks packets lost, and then updates congestion state again. This
was fine because everything not yet delivered is marked lost,
so the inflight is always 0 and cwnd can be safely set to 1 to
retransmit one packet on timeout.
But the inflight may not always be 0 on timeout if TCP changes to
mark packets lost based on packet sent time. Therefore we must
first mark the packet lost, then set the cwnd based on the
(updated) inflight.
This is not a pure refactor. Congestion control may potentially
break if it uses (not yet updated) inflight to compute ssthresh.
Fortunately all existing congestion control modules does not do that.
Also it changes the inflight when CA_LOSS_EVENT is called, and only
westwood processes such an event but does not use inflight.
This change has two other minor side benefits:
1) consistent with Fast Recovery s.t. the inflight is updated
first before tcp_enter_recovery flips state to CA_Recovery.
2) avoid intertwining loss marking with state update, making the
code more readable.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Refactor using a new helper, tcp_timeout_mark_loss(), that marks packets
lost upon RTO.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The previous approach for the lost and retransmit bits was to
wipe the slate clean: zero all the lost and retransmit bits,
correspondingly zero the lost_out and retrans_out counters, and
then add back the lost bits (and correspondingly increment lost_out).
The new approach is to treat this very much like marking packets
lost in fast recovery. We don’t wipe the slate clean. We just say
that for all packets that were not yet marked sacked or lost, we now
mark them as lost in exactly the same way we do for fast recovery.
This fixes the lost retransmit accounting at RTO time and greatly
simplifies the RTO code by sharing much of the logic with Fast
Recovery.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This is a rewrite of NewReno loss recovery implementation that is
simpler and standalone for readability and better performance by
using less states.
Note that NewReno refers to RFC6582 as a modification to the fast
recovery algorithm. It is used only if the connection does not
support SACK in Linux. It should not to be confused with the Reno
(AIMD) congestion control.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This patch disables RFC6675 loss detection and make sysctl
net.ipv4.tcp_recovery = 1 controls a binary choice between RACK
(1) or RFC6675 (0).
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Called when a TCP segment is acknowledged.
Could be used by application protocols who hold additional
metadata associated with the stream data.
This is required by TLS device offload to release
metadata associated with acknowledged TLS records.
Signed-off-by: Ilya Lesokhin <ilyal@mellanox.com>
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: Aviad Yehezkel <aviadye@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
|
|
This reverts commit <c6849a3ac17e> ("net: init sk_cookie for inet socket")
Per discussion with Eric, when update sock_net(sk)->cookie_gen, the
whole cache cache line will be invalidated, as this cache line is shared
with all cpus, that may cause great performace hit.
Bellow is the data form Eric.
"Performance is reduced from ~5 Mpps to ~3.8 Mpps with 16 RX queues on
my host" when running synflood test.
Have to revert it to prevent from cache line false sharing.
Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
With sk_cookie we can identify a socket, that is very helpful for
traceing and statistic, i.e. tcp tracepiont and ebpf.
So we'd better init it by default for inet socket.
When using it, we just need call atomic64_read(&sk->sk_cookie).
Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
tcp_rcv_space_adjust is called every time data is copied to user space,
introducing a tcp tracepoint for which could show us when the packet is
copied to user.
When a tcp packet arrives, tcp_rcv_established() will be called and with
the existed tracepoint tcp_probe we could get the time when this packet
arrives.
Then this packet will be copied to user, and tcp_rcv_space_adjust will
be called and with this new introduced tracepoint we could get the time
when this packet is copied to user.
With these two tracepoints, we could figure out whether the user program
processes this packet immediately or there's latency.
Hence in the printk message, sk_cookie is printed as a key to relate
tcp_rcv_space_adjust with tcp_probe.
Maybe we could export sockfd in this new tracepoint as well, then we
could relate this new tracepoint with epoll/read/recv* tracepoints, and
finally that could show us the whole lifespan of this packet. But we
could also implement that with pid as these functions are executed in
process context.
Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The old code reads the "opsize" variable from out-of-bounds memory (first
byte behind the segment) if a broken TCP segment ends directly after an
opcode that is neither EOL nor NOP.
The result of the read isn't used for anything, so the worst thing that
could theoretically happen is a pagefault; and since the physmap is usually
mostly contiguous, even that seems pretty unlikely.
The following C reproducer triggers the uninitialized read - however, you
can't actually see anything happen unless you put something like a
pr_warn() in tcp_parse_md5sig_option() to print the opsize.
====================================
#define _GNU_SOURCE
#include <arpa/inet.h>
#include <stdlib.h>
#include <errno.h>
#include <stdarg.h>
#include <net/if.h>
#include <linux/if.h>
#include <linux/ip.h>
#include <linux/tcp.h>
#include <linux/in.h>
#include <linux/if_tun.h>
#include <err.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <string.h>
#include <stdio.h>
#include <unistd.h>
#include <sys/ioctl.h>
#include <assert.h>
void systemf(const char *command, ...) {
char *full_command;
va_list ap;
va_start(ap, command);
if (vasprintf(&full_command, command, ap) == -1)
err(1, "vasprintf");
va_end(ap);
printf("systemf: <<<%s>>>\n", full_command);
system(full_command);
}
char *devname;
int tun_alloc(char *name) {
int fd = open("/dev/net/tun", O_RDWR);
if (fd == -1)
err(1, "open tun dev");
static struct ifreq req = { .ifr_flags = IFF_TUN|IFF_NO_PI };
strcpy(req.ifr_name, name);
if (ioctl(fd, TUNSETIFF, &req))
err(1, "TUNSETIFF");
devname = req.ifr_name;
printf("device name: %s\n", devname);
return fd;
}
#define IPADDR(a,b,c,d) (((a)<<0)+((b)<<8)+((c)<<16)+((d)<<24))
void sum_accumulate(unsigned int *sum, void *data, int len) {
assert((len&2)==0);
for (int i=0; i<len/2; i++) {
*sum += ntohs(((unsigned short *)data)[i]);
}
}
unsigned short sum_final(unsigned int sum) {
sum = (sum >> 16) + (sum & 0xffff);
sum = (sum >> 16) + (sum & 0xffff);
return htons(~sum);
}
void fix_ip_sum(struct iphdr *ip) {
unsigned int sum = 0;
sum_accumulate(&sum, ip, sizeof(*ip));
ip->check = sum_final(sum);
}
void fix_tcp_sum(struct iphdr *ip, struct tcphdr *tcp) {
unsigned int sum = 0;
struct {
unsigned int saddr;
unsigned int daddr;
unsigned char pad;
unsigned char proto_num;
unsigned short tcp_len;
} fakehdr = {
.saddr = ip->saddr,
.daddr = ip->daddr,
.proto_num = ip->protocol,
.tcp_len = htons(ntohs(ip->tot_len) - ip->ihl*4)
};
sum_accumulate(&sum, &fakehdr, sizeof(fakehdr));
sum_accumulate(&sum, tcp, tcp->doff*4);
tcp->check = sum_final(sum);
}
int main(void) {
int tun_fd = tun_alloc("inject_dev%d");
systemf("ip link set %s up", devname);
systemf("ip addr add 192.168.42.1/24 dev %s", devname);
struct {
struct iphdr ip;
struct tcphdr tcp;
unsigned char tcp_opts[20];
} __attribute__((packed)) syn_packet = {
.ip = {
.ihl = sizeof(struct iphdr)/4,
.version = 4,
.tot_len = htons(sizeof(syn_packet)),
.ttl = 30,
.protocol = IPPROTO_TCP,
/* FIXUP check */
.saddr = IPADDR(192,168,42,2),
.daddr = IPADDR(192,168,42,1)
},
.tcp = {
.source = htons(1),
.dest = htons(1337),
.seq = 0x12345678,
.doff = (sizeof(syn_packet.tcp)+sizeof(syn_packet.tcp_opts))/4,
.syn = 1,
.window = htons(64),
.check = 0 /*FIXUP*/
},
.tcp_opts = {
/* INVALID: trailing MD5SIG opcode after NOPs */
1, 1, 1, 1, 1,
1, 1, 1, 1, 1,
1, 1, 1, 1, 1,
1, 1, 1, 1, 19
}
};
fix_ip_sum(&syn_packet.ip);
fix_tcp_sum(&syn_packet.ip, &syn_packet.tcp);
while (1) {
int write_res = write(tun_fd, &syn_packet, sizeof(syn_packet));
if (write_res != sizeof(syn_packet))
err(1, "packet write failed");
}
}
====================================
Fixes: cfb6eeb4c860 ("[TCP]: MD5 Signature Option (RFC2385) support.")
Signed-off-by: Jann Horn <jannh@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Export data delivered and delivered with CE marks to
1) SNMP TCPDelivered and TCPDeliveredCE
2) getsockopt(TCP_INFO)
3) Timestamping API SOF_TIMESTAMPING_OPT_STATS
Note that for SCM_TSTAMP_ACK, the delivery info in
SOF_TIMESTAMPING_OPT_STATS is reported before the info
was fully updated on the ACK.
These stats help application monitor TCP delivery and ECN status
on per host, per connection, even per message level.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Introduce a new delivered_ce stat in tcp socket to estimate
number of packets being marked with CE bits. The estimation is
done via ACKs with ECE bit. Depending on the actual receiver
behavior, the estimation could have biases.
Since the TCP sender can't really see the CE bit in the data path,
so the sender is technically counting packets marked delivered with
the "ECE / ECN-Echo" flag set.
With RFC3168 ECN, because the ECE bit is sticky, this count can
drastically overestimate the nummber of CE-marked data packets
With DCTCP-style ECN this should be reasonably precise unless there
is loss in the ACK path, in which case it's not precise.
With AccECN proposal this can be made still more precise, even in
the case some degree of ACK loss.
However this is sender's best estimate of CE information.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Add new helper tcp_newly_delivered() to prepare the ECN accounting change.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
the tcp_sock:delivered has inconsistent accounting for SYN and FIN.
1. it counts pure FIN
2. it counts pure SYN
3. it counts SYN-data twice
4. it does not count SYN-ACK
For congestion control perspective it does not matter much as C.C. only
cares about the difference not the aboslute value. But the next patch
would export this field to user-space so it's better to report the absolute
value w/o these caveats.
This patch counts SYN, SYN-ACK, or SYN-data delivery once always in
the "delivered" field.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
SO_RCVLOWAT is properly handled in tcp_poll(), so that POLLIN is only
generated when enough bytes are available in receive queue, after
David change (commit c7004482e8dc "tcp: Respect SO_RCVLOWAT in tcp_poll().")
But TCP still calls sk->sk_data_ready() for each chunk added in receive
queue, meaning thread is awaken, and goes back to sleep shortly after.
Tested:
tcp_mmap test program, receiving 32768 MB of data with SO_RCVLOWAT set to 512KB
-> Should get ~2 wakeups (c-switches) per MB, regardless of how many
(tiny or big) packets were received.
High speed (mostly full size GRO packets)
received 32768 MB (100 % mmap'ed) in 8.03112 s, 34.2266 Gbit,
cpu usage user:0.037 sys:1.404, 43.9758 usec per MB, 65497 c-switches
received 32768 MB (99.9954 % mmap'ed) in 7.98453 s, 34.4263 Gbit,
cpu usage user:0.03 sys:1.422, 44.3115 usec per MB, 65485 c-switches
Low speed (sender is ratelimited and sends 1-MSS at a time, so GRO is not helping)
received 22474.5 MB (100 % mmap'ed) in 6015.35 s, 0.0313414 Gbit,
cpu usage user:0.05 sys:1.586, 72.7952 usec per MB, 44950 c-switches
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
We should not delay acks if there are not enough bytes
in receive queue to satisfy SO_RCVLOWAT.
Since [E]POLLIN event is not going to be generated, there is little
hope for a delayed ack to be useful.
In fact, delaying ACK prevents sender from completing
the transfer.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Minor conflicts in drivers/net/ethernet/mellanox/mlx5/core/en_rep.c,
we had some overlapping changes:
1) In 'net' MLX5E_PARAMS_LOG_{SQ,RQ}_SIZE -->
MLX5E_REP_PARAMS_LOG_{SQ,RQ}_SIZE
2) In 'net-next' params->log_rq_size is renamed to be
params->log_rq_mtu_frames.
3) In 'net-next' params->hard_mtu is added.
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Currently, the SMC experimental TCP option in a SYN packet is lost on
the server side when SYN Cookies are active. However, the corresponding
SYNACK sent back to the client contains the SMC option. This causes an
inconsistent view of the SMC capabilities on the client and server.
This patch disables the SMC option in the SYNACK when SYN Cookies are
active to avoid this issue.
Fixes: 60e2a7780793b ("tcp: TCP experimental option for SMC")
Signed-off-by: Hans Wippel <hwippel@linux.vnet.ibm.com>
Signed-off-by: Ursula Braun <ubraun@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
All of the conflicts were cases of overlapping changes.
In net/core/devlink.c, we have to make care that the
resouce size_params have become a struct member rather
than a pointer to such an object.
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
When the connection is reset, there is no point in
keeping the packets on the write queue until the connection
is closed.
RFC 793 (page 70) and RFC 793-bis (page 64) both suggest
purging the write queue upon RST:
https://tools.ietf.org/html/draft-ietf-tcpm-rfc793bis-07
Moreover, this is essential for a correct MSG_ZEROCOPY
implementation, because userspace cannot call close(fd)
before receiving zerocopy signals even when the connection
is reset.
Fixes: f214f915e7db ("tcp: enable MSG_ZEROCOPY")
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This reverts commit 89fe18e44f7ee5ab1c90d0dff5835acee7751427.
While the patch could detect more spurious timeouts, it could cause
poor TCP performance on broken middle-boxes that modifies TCP packets
(e.g. receive window, SACK options). Since the performance gain is
much smaller compared to the potential loss. The best solution is
to fully revert the change.
Fixes: 89fe18e44f7e ("tcp: extend F-RTO to catch more spurious timeouts")
Reported-by: Teodor Milkov <tm@del.bg>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This reverts commit cc663f4d4c97b7297fb45135ab23cfd508b35a77. While fixing
some broken middle-boxes that modifies receive window fields, it does not
address middle-boxes that strip off SACK options. The best solution is
to fully revert this patch and the root F-RTO enhancement.
Fixes: cc663f4d4c97 ("tcp: restrict F-RTO to work-around broken middle-boxes")
Reported-by: Teodor Milkov <tm@del.bg>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
After previous commit, sk_can_gso() is always true.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
배석진 reported that in some situations, packets for a given 5-tuple
end up being processed by different CPUS.
This involves RPS, and fragmentation.
배석진 is seeing packet drops when a SYN_RECV request socket is
moved into ESTABLISH state. Other states are protected by socket lock.
This is caused by a CPU losing the race, and simply not caring enough.
Since this seems to occur frequently, we can do better and perform
a second lookup.
Note that all needed memory barriers are already in the existing code,
thanks to the spin_lock()/spin_unlock() pair in inet_ehash_insert()
and reqsk_put(). The second lookup must find the new socket,
unless it has already been accepted and closed by another cpu.
Note that the fragmentation could be avoided in the first place by
use of a correct TCP MSS option in the SYN{ACK} packet, but this
does not mean we can not be more robust.
Many thanks to 배석진 for a very detailed analysis.
Reported-by: 배석진 <soukjin.bae@samsung.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This is the mindless scripted replacement of kernel use of POLL*
variables as described by Al, done by this script:
for V in IN OUT PRI ERR RDNORM RDBAND WRNORM WRBAND HUP RDHUP NVAL MSG; do
L=`git grep -l -w POLL$V | grep -v '^t' | grep -v /um/ | grep -v '^sa' | grep -v '/poll.h$'|grep -v '^D'`
for f in $L; do sed -i "-es/^\([^\"]*\)\(\<POLL$V\>\)/\\1E\\2/" $f; done
done
with de-mangling cleanups yet to come.
NOTE! On almost all architectures, the EPOLL* constants have the same
values as the POLL* constants do. But they keyword here is "almost".
For various bad reasons they aren't the same, and epoll() doesn't
actually work quite correctly in some cases due to this on Sparc et al.
The next patch from Al will sort out the final differences, and we
should be all done.
Scripted-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
A persistent connection may send tiny amount of data (e.g. health-check)
for a long period of time. BBR's windowed min RTT filter may only see
RTT samples from delayed ACKs causing BBR to grossly over-estimate
the path delay depending how much the ACK was delayed at the receiver.
This patch skips RTT samples that are likely coming from delayed ACKs. Note
that it is possible the sender never obtains a valid measure to set the
min RTT. In this case BBR will continue to set cwnd to initial window
which seems fine because the connection is thin stream.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This patch avoids having TCP sender or congestion control
overestimate the min RTT by orders of magnitude. This happens when
all the samples in the windowed filter are one-packet transfer
like small request and health-check like chit-chat, which is farily
common for applications using persistent connections. This patch
tries to conservatively labels and skip RTT samples obtained from
this type of workload.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This adds an event to trace TCP stat variables with
slightly intrusive trace-event. This uses ftrace/perf
event log buffer to trace those state, no needs to
prepare own ring-buffer, nor custom user apps.
User can use ftrace to trace this event as below;
# cd /sys/kernel/debug/tracing
# echo 1 > events/tcp/tcp_probe/enable
(run workloads)
# cat trace
Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Three sets of overlapping changes, two in the packet scheduler
and one in the meson-gxl PHY driver.
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
When ms timestamp is used, current logic uses 1us in
tcp_rcv_rtt_update() when the real rcv_rtt is within 1 - 999us.
This could cause rcv_rtt underestimation.
Fix it by always using a min value of 1ms if ms timestamp is used.
Fixes: 645f4c6f2ebd ("tcp: switch rcv_rtt_est and rcvq_space to high resolution timestamps")
Signed-off-by: Wei Wang <weiwan@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
