When tcp sends a TSO packet, adding a PSH flag on it
reduces the sojourn time of GRO packet in GRO receivers.
This is particularly the case under pressure, since RX queues
receive packets for many concurrent flows.
A sender can give a hint to GRO engines when it is
appropriate to flush a super-packet, especially when pacing
is in the picture, since next packet is probably delayed by
one ms.
Having less packets in GRO engine reduces chance
of LRU eviction or inflated RTT, and reduces GRO cost.
We found recently that we must not set the PSH flag on
individual full-size MSS segments [1] :
Under pressure (CWR state), we better let the packet sit
for a small delay (depending on NAPI logic) so that the
ACK packet is delayed, and thus next packet we send is
also delayed a bit. Eventually the bottleneck queue can
be drained. DCTCP flows with CWND=1 have demonstrated
the issue.
This patch allows to slowdown the aggregate traffic without
involving high resolution timers on senders and/or
receivers.
It has been used at Google for about four years,
and has been discussed at various networking conferences.
[1] segments smaller than MSS already have PSH flag set
by tcp_sendmsg() / tcp_mark_push(), unless MSG_MORE
has been requested by the user.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Tariq Toukan <tariqt@mellanox.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>
TCP associates tx timestamp requests with a byte in the bytestream.
If merging skbs in tcp_mtu_probe, migrate the tstamp request.
Similar to MSG_EOR, do not allow moving a timestamp from any segment
in the probe but the last. This to avoid merging multiple timestamps.
Tested with the packetdrill script at
https://github.com/wdebruij/packetdrill/commits/mtu_probe-1
Link: http://patchwork.ozlabs.org/patch/1143278/#2232897
Fixes: 4ed2d765df ("net-timestamp: TCP timestamping")
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
sk_validate_xmit_skb() and drivers depend on the sk member of
struct sk_buff to identify segments requiring encryption.
Any operation which removes or does not preserve the original TLS
socket such as skb_orphan() or skb_clone() will cause clear text
leaks.
Make the TCP socket underlying an offloaded TLS connection
mark all skbs as decrypted, if TLS TX is in offload mode.
Then in sk_validate_xmit_skb() catch skbs which have no socket
(or a socket with no validation) and decrypted flag set.
Note that CONFIG_SOCK_VALIDATE_XMIT, CONFIG_TLS_DEVICE and
sk->sk_validate_xmit_skb are slightly interchangeable right now,
they all imply TLS offload. The new checks are guarded by
CONFIG_TLS_DEVICE because that's the option guarding the
sk_buff->decrypted member.
Second, smaller issue with orphaning is that it breaks
the guarantee that packets will be delivered to device
queues in-order. All TLS offload drivers depend on that
scheduling property. This means skb_orphan_partial()'s
trick of preserving partial socket references will cause
issues in the drivers. We need a full orphan, and as a
result netem delay/throttling will cause all TLS offload
skbs to be dropped.
Reusing the sk_buff->decrypted flag also protects from
leaking clear text when incoming, decrypted skb is redirected
(e.g. by TC).
See commit 0608c69c9a ("bpf: sk_msg, sock{map|hash} redirect
through ULP") for justification why the internal flag is safe.
The only location which could leak the flag in is tcp_bpf_sendmsg(),
which is taken care of by clearing the previously unused bit.
v2:
- remove superfluous decrypted mark copy (Willem);
- remove the stale doc entry (Boris);
- rely entirely on EOR marking to prevent coalescing (Boris);
- use an internal sendpages flag instead of marking the socket
(Boris).
v3 (Willem):
- reorganize the can_skb_orphan_partial() condition;
- fix the flag leak-in through tcp_bpf_sendmsg.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use accessor functions for skb fragment's page_offset instead
of direct references, in preparation for bvec conversion.
Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some applications set tiny SO_SNDBUF values and expect
TCP to just work. Recent patches to address CVE-2019-11478
broke them in case of losses, since retransmits might
be prevented.
We should allow these flows to make progress.
This patch allows the first and last skb in retransmit queue
to be split even if memory limits are hit.
It also adds the some room due to the fact that tcp_sendmsg()
and tcp_sendpage() might overshoot sk_wmem_queued by about one full
TSO skb (64KB size). Note this allowance was already present
in stable backports for kernels < 4.15
Note for < 4.15 backports :
tcp_rtx_queue_tail() will probably look like :
static inline struct sk_buff *tcp_rtx_queue_tail(const struct sock *sk)
{
struct sk_buff *skb = tcp_send_head(sk);
return skb ? tcp_write_queue_prev(sk, skb) : tcp_write_queue_tail(sk);
}
Fixes: f070ef2ac6 ("tcp: tcp_fragment() should apply sane memory limits")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Andrew Prout <aprout@ll.mit.edu>
Tested-by: Andrew Prout <aprout@ll.mit.edu>
Tested-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Tested-by: Michal Kubecek <mkubecek@suse.cz>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Christoph Paasch <cpaasch@apple.com>
Cc: Jonathan Looney <jtl@netflix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_fragment() might be called for skbs in the write queue.
Memory limits might have been exceeded because tcp_sendmsg() only
checks limits at full skb (64KB) boundaries.
Therefore, we need to make sure tcp_fragment() wont punish applications
that might have setup very low SO_SNDBUF values.
Fixes: f070ef2ac6 ("tcp: tcp_fragment() should apply sane memory limits")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Christoph Paasch <cpaasch@apple.com>
Tested-by: Christoph Paasch <cpaasch@apple.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some TCP peers announce a very small MSS option in their SYN and/or
SYN/ACK messages.
This forces the stack to send packets with a very high network/cpu
overhead.
Linux has enforced a minimal value of 48. Since this value includes
the size of TCP options, and that the options can consume up to 40
bytes, this means that each segment can include only 8 bytes of payload.
In some cases, it can be useful to increase the minimal value
to a saner value.
We still let the default to 48 (TCP_MIN_SND_MSS), for compatibility
reasons.
Note that TCP_MAXSEG socket option enforces a minimal value
of (TCP_MIN_MSS). David Miller increased this minimal value
in commit c39508d6f1 ("tcp: Make TCP_MAXSEG minimum more correct.")
from 64 to 88.
We might in the future merge TCP_MIN_SND_MSS and TCP_MIN_MSS.
CVE-2019-11479 -- tcp mss hardcoded to 48
Signed-off-by: Eric Dumazet <edumazet@google.com>
Suggested-by: Jonathan Looney <jtl@netflix.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Tyler Hicks <tyhicks@canonical.com>
Cc: Bruce Curtis <brucec@netflix.com>
Cc: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Jonathan Looney reported that a malicious peer can force a sender
to fragment its retransmit queue into tiny skbs, inflating memory
usage and/or overflow 32bit counters.
TCP allows an application to queue up to sk_sndbuf bytes,
so we need to give some allowance for non malicious splitting
of retransmit queue.
A new SNMP counter is added to monitor how many times TCP
did not allow to split an skb if the allowance was exceeded.
Note that this counter might increase in the case applications
use SO_SNDBUF socket option to lower sk_sndbuf.
CVE-2019-11478 : tcp_fragment, prevent fragmenting a packet when the
socket is already using more than half the allowed space
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Jonathan Looney <jtl@netflix.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Reviewed-by: Tyler Hicks <tyhicks@canonical.com>
Cc: Bruce Curtis <brucec@netflix.com>
Cc: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Jonathan Looney reported that TCP can trigger the following crash
in tcp_shifted_skb() :
BUG_ON(tcp_skb_pcount(skb) < pcount);
This can happen if the remote peer has advertized the smallest
MSS that linux TCP accepts : 48
An skb can hold 17 fragments, and each fragment can hold 32KB
on x86, or 64KB on PowerPC.
This means that the 16bit witdh of TCP_SKB_CB(skb)->tcp_gso_segs
can overflow.
Note that tcp_sendmsg() builds skbs with less than 64KB
of payload, so this problem needs SACK to be enabled.
SACK blocks allow TCP to coalesce multiple skbs in the retransmit
queue, thus filling the 17 fragments to maximal capacity.
CVE-2019-11477 -- u16 overflow of TCP_SKB_CB(skb)->tcp_gso_segs
Fixes: 832d11c5cd ("tcp: Try to restore large SKBs while SACK processing")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Jonathan Looney <jtl@netflix.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Tyler Hicks <tyhicks@canonical.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Bruce Curtis <brucec@netflix.com>
Cc: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Adding delays to TCP flows is crucial for studying behavior
of TCP stacks, including congestion control modules.
Linux offers netem module, but it has unpractical constraints :
- Need root access to change qdisc
- Hard to setup on egress if combined with non trivial qdisc like FQ
- Single delay for all flows.
EDT (Earliest Departure Time) adoption in TCP stack allows us
to enable a per socket delay at a very small cost.
Networking tools can now establish thousands of flows, each of them
with a different delay, simulating real world conditions.
This requires FQ packet scheduler or a EDT-enabled NIC.
This patchs adds TCP_TX_DELAY socket option, to set a delay in
usec units.
unsigned int tx_delay = 10000; /* 10 msec */
setsockopt(fd, SOL_TCP, TCP_TX_DELAY, &tx_delay, sizeof(tx_delay));
Note that FQ packet scheduler limits might need some tweaking :
man tc-fq
PARAMETERS
limit
Hard limit on the real queue size. When this limit is
reached, new packets are dropped. If the value is lowered,
packets are dropped so that the new limit is met. Default
is 10000 packets.
flow_limit
Hard limit on the maximum number of packets queued per
flow. Default value is 100.
Use of TCP_TX_DELAY option will increase number of skbs in FQ qdisc,
so packets would be dropped if any of the previous limit is hit.
Use of a jump label makes this support runtime-free, for hosts
never using the option.
Also note that TSQ (TCP Small Queues) limits are slightly changed
with this patch : we need to account that skbs artificially delayed
wont stop us providind more skbs to feed the pipe (netem uses
skb_orphan_partial() for this purpose, but FQ can not use this trick)
Because of that, using big delays might very well trigger
old bugs in TSO auto defer logic and/or sndbuf limited detection.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add SPDX license identifiers to all files which:
- Have no license information of any form
- Have EXPORT_.*_SYMBOL_GPL inside which was used in the
initial scan/conversion to ignore the file
These files fall under the project license, GPL v2 only. The resulting SPDX
license identifier is:
GPL-2.0-only
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Detecting spurious SYNACK timeout using timestamp option requires
recording the exact SYNACK skb timestamp. Previously the SYNACK
sent timestamp was stamped slightly earlier before the skb
was transmitted. This patch uses the SYNACK skb transmission
timestamp directly.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The non-null check on tskb is always false because it is in an else
path of a check on tskb and hence tskb is null in this code block.
This is check is therefore redundant and can be removed as well
as the label coalesc.
if (tsbk) {
...
} else {
...
if (unlikely(!skb)) {
if (tskb) /* can never be true, redundant code */
goto coalesc;
return;
}
}
Addresses-Coverity: ("Logically dead code")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Reviewed-by: Mukesh Ojha <mojha@codeaurora.org>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_clock_ns() (aka ktime_get_ns()) is using monotonic clock,
so the checks we had in tcp_mstamp_refresh() are no longer
relevant.
This patch removes cpu stall (when the cache line is not hot)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tso_fragment() is only called for packets still in write queue.
Remove the tcp_queue parameter to make this more obvious,
even if the comment clearly states this.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We prefer static_branch_unlikely() over static_key_false() these days.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Three conflicts, one of which, for marvell10g.c is non-trivial and
requires some follow-up from Heiner or someone else.
The issue is that Heiner converted the marvell10g driver over to
use the generic c45 code as much as possible.
However, in 'net' a bug fix appeared which makes sure that a new
local mask (MDIO_AN_10GBT_CTRL_ADV_NBT_MASK) with value 0x01e0
is cleared.
Signed-off-by: David S. Miller <davem@davemloft.net>
In order to be more confident about an on-going interactive session, we
increment pingpong count by 1 for every interactive transaction and we
adjust TCP_PINGPONG_THRESH to 3.
This means, we only consider a session in pingpong mode after we see 3
interactive transactions, and start to activate delayed acks in quick
ack mode.
And in order to not over-count the credits, we only increase pingpong
count for the first packet sent in response for the previous received
packet.
This is mainly to prevent delaying the ack immediately after some
handshake protocol but no real interactive traffic pattern afterwards.
Signed-off-by: Wei Wang <weiwan@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>
Instead of using pingpong as a single bit information, we refactor the
code to treat it as a counter. When interactive session is detected,
we set pingpong count to TCP_PINGPONG_THRESH. And when pingpong count
is >= TCP_PINGPONG_THRESH, we consider the session in pingpong mode.
This patch is a pure refactor and sets foundation for the next patch.
This patch itself does not change any pingpong logic.
Signed-off-by: Wei Wang <weiwan@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>
Accept MSG_ZEROCOPY in all the TCP states that allow sendmsg. Remove
the explicit check for ESTABLISHED and CLOSE_WAIT states.
This requires correctly handling zerocopy state (uarg, sk_zckey) in
all paths reachable from other TCP states. Such as the EPIPE case
in sk_stream_wait_connect, which a sendmsg() in incorrect state will
now hit. Most paths are already safe.
Only extension needed is for TCP Fastopen active open. This can build
an skb with data in tcp_send_syn_data. Pass the uarg along with other
fastopen state, so that this skb also generates a zerocopy
notification on release.
Tested with active and passive tcp fastopen packetdrill scripts at
1747eef03d
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously when the sender fails to send (original) data packet or
window probes due to congestion in the local host (e.g. throttling
in qdisc), it'll retry within an RTO or two up to 500ms.
In low-RTT networks such as data-centers, RTO is often far below
the default minimum 200ms. Then local host congestion could trigger
a retry storm pouring gas to the fire. Worse yet, the probe counter
(icsk_probes_out) is not properly updated so the aggressive retry
may exceed the system limit (15 rounds) until the packet finally
slips through.
On such rare events, it's wise to retry more conservatively
(500ms) and update the stats properly to reflect these incidents
and follow the system limit. Note that this is consistent with
the behaviors when a keep-alive probe or RTO retry is dropped
due to local congestion.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously TCP socket's retrans_stamp is not set if the
retransmission has failed to send. As a result if a socket is
experiencing local issues to retransmit packets, determining when
to abort a socket is complicated w/o knowning the starting time of
the recovery since retrans_stamp may remain zero.
This complication causes sub-optimal behavior that TCP may use the
latest, instead of the first, retransmission time to compute the
elapsed time of a stalling connection due to local issues. Then TCP
may disrecard TCP retries settings and keep retrying until it finally
succeed: not a good idea when the local host is already strained.
The simple fix is to always timestamp the start of a recovery.
It's worth noting that retrans_stamp is also used to compare echo
timestamp values to detect spurious recovery. This patch does
not break that because retrans_stamp is still later than when the
original packet was sent.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously TCP skbs are not always timestamped if the transmission
failed due to memory or other local issues. This makes deciding
when to abort a socket tricky and complicated because the first
unacknowledged skb's timestamp may be 0 on TCP timeout.
The straight-forward fix is to always timestamp skb on every
transmission attempt. Also every skb retransmission needs to be
flagged properly to avoid RTT under-estimation. This can happen
upon receiving an ACK for the original packet and the a previous
(spurious) retransmission has failed.
It's worth noting that this reverts to the old time-stamping
style before commit 8c72c65b42 ("tcp: update skb->skb_mstamp more
carefully") which addresses a problem in computing the elapsed time
of a stalled window-probing socket. The problem will be addressed
differently in the next patches with a simpler approach.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit f9bfe4e6a9 ("tcp: lack of available data can also cause
TSO defer") we moved the test in tcp_tso_should_defer() for packets
with a FIN flag, and we mentioned that the same would be done
later for EOR flag.
Both flags should be handled at the same time, after all other
heuristics have been considered. They both mean that no more bytes
can be added to this skb by an application.
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>
Several conflicts, seemingly all over the place.
I used Stephen Rothwell's sample resolutions for many of these, if not
just to double check my own work, so definitely the credit largely
goes to him.
The NFP conflict consisted of a bug fix (moving operations
past the rhashtable operation) while chaning the initial
argument in the function call in the moved code.
The net/dsa/master.c conflict had to do with a bug fix intermixing of
making dsa_master_set_mtu() static with the fixing of the tagging
attribute location.
cls_flower had a conflict because the dup reject fix from Or
overlapped with the addition of port range classifiction.
__set_phy_supported()'s conflict was relatively easy to resolve
because Andrew fixed it in both trees, so it was just a matter
of taking the net-next copy. Or at least I think it was :-)
Joe Stringer's fix to the handling of netns id 0 in bpf_sk_lookup()
intermixed with changes on how the sdif and caller_net are calculated
in these code paths in net-next.
The remaining BPF conflicts were largely about the addition of the
__bpf_md_ptr stuff in 'net' overlapping with adjustments and additions
to the relevant data structure where the MD pointer macros are used.
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_tso_should_defer() can return true in three different cases :
1) We are cwnd-limited
2) We are rwnd-limited
3) We are application limited.
Neal pointed out that my recent fix went too far, since
it assumed that if we were not in 1) case, we must be rwnd-limited
Fix this by properly populating the is_cwnd_limited and
is_rwnd_limited booleans.
After this change, we can finally move the silly check for FIN
flag only for the application-limited case.
The same move for EOR bit will be handled in net-next,
since commit 1c09f7d073 ("tcp: do not try to defer skbs
with eor mark (MSG_EOR)") is scheduled for linux-4.21
Tested by running 200 concurrent netperf -t TCP_RR -- -r 60000,100
and checking none of them was rwnd_limited in the chrono_stat
output from "ss -ti" command.
Fixes: 41727549de ("tcp: Do not underestimate rwnd_limited")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Suggested-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP loss probe timer may fire when the retranmission queue is empty but
has a non-zero tp->packets_out counter. tcp_send_loss_probe will call
tcp_rearm_rto which triggers NULL pointer reference by fetching the
retranmission queue head in its sub-routines.
Add a more detailed warning to help catch the root cause of the inflight
accounting inconsistency.
Reported-by: Rafael Tinoco <rafael.tinoco@linaro.org>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If available rwnd is too small, tcp_tso_should_defer()
can decide it is worth waiting before splitting a TSO packet.
This really means we are rwnd limited.
Fixes: 5615f88614 ("tcp: instrument how long TCP is limited by receive window")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously the SNMP counter LINUX_MIB_TCPRETRANSFAIL is not counting
the TSO/GSO properly on failed retransmission. This patch fixes that.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Most linux hosts never setup TCP MD5 keys. We can avoid a
cache line miss (accessing tp->md5ig_info) on RX and TX
using a jump label.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We can remove the loop and conditional branches
and compute wscale efficiently thanks to ilog2()
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>
Jean-Louis reported a TCP regression and bisected to recent SACK
compression.
After a loss episode (receiver not able to keep up and dropping
packets because its backlog is full), linux TCP stack is sending
a single SACK (DUPACK).
Sender waits a full RTO timer before recovering losses.
While RFC 6675 says in section 5, "Algorithm Details",
(2) If DupAcks < DupThresh but IsLost (HighACK + 1) returns true --
indicating at least three segments have arrived above the current
cumulative acknowledgment point, which is taken to indicate loss
-- go to step (4).
...
(4) Invoke fast retransmit and enter loss recovery as follows:
there are old TCP stacks not implementing this strategy, and
still counting the dupacks before starting fast retransmit.
While these stacks probably perform poorly when receivers implement
LRO/GRO, we should be a little more gentle to them.
This patch makes sure we do not enable SACK compression unless
3 dupacks have been sent since last rcv_nxt update.
Ideally we should even rearm the timer to send one or two
more DUPACK if no more packets are coming, but that will
be work aiming for linux-4.21.
Many thanks to Jean-Louis for bisecting the issue, providing
packet captures and testing this patch.
Fixes: 5d9f4262b7 ("tcp: add SACK compression")
Reported-by: Jean-Louis Dupond <jean-louis@dupond.be>
Tested-by: Jean-Louis Dupond <jean-louis@dupond.be>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
FQ pacing guarantees that paced packets queued by one flow do not
add head-of-line blocking for other flows.
After TCP GSO conversion, increasing limit_output_bytes to 1 MB is safe,
since this maps to 16 skbs at most in qdisc or device queues.
(or slightly more if some drivers lower {gso_max_segs|size})
We still can queue at most 1 ms worth of traffic (this can be scaled
by wifi drivers if they need to)
Tested:
# ethtool -c eth0 | egrep "tx-usecs:|tx-frames:" # 40 Gbit mlx4 NIC
tx-usecs: 16
tx-frames: 16
# tc qdisc replace dev eth0 root fq
# for f in {1..10};do netperf -P0 -H lpaa24,6 -o THROUGHPUT;done
Before patch:
27711
26118
27107
27377
27712
27388
27340
27117
27278
27509
After patch:
37434
36949
36658
36998
37711
37291
37605
36659
36544
37349
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_tso_should_defer() first heuristic is to not defer
if last send is "old enough".
Its current implementation uses jiffies and its low granularity.
TSO autodefer performance should not rely on kernel HZ :/
After EDT conversion, we have state variables in nanoseconds that
can allow us to properly implement the heuristic.
This patch increases TSO chunk sizes on medium rate flows,
especially when receivers do not use GRO or similar aggregation.
It also reduces bursts for HZ=100 or HZ=250 kernels, making TCP
behavior more uniform.
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>
tcp_tso_should_defer() last step tries to check if the probable
next ACK packet is coming in less than half rtt.
Problem is that the head->tstamp might be in the future,
so we need to use signed arithmetics to avoid overflows.
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>
Applications using MSG_EOR are giving a strong hint to TCP stack :
Subsequent sendmsg() can not append more bytes to skbs having
the EOR mark.
Do not try to TSO defer suchs skbs, there is really no hope.
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>
With EDT model, SRTT no longer is inflated by pacing delays.
This means that RTO and some other xmit timers might be setup
incorrectly. This is particularly visible with either :
- Very small enforced pacing rates (SO_MAX_PACING_RATE)
- Reduced rto (from the default 200 ms)
This can lead to TCP flows aborts in the worst case,
or spurious retransmits in other cases.
For example, this session gets far more throughput
than the requested 80kbit :
$ netperf -H 127.0.0.2 -l 100 -- -q 10000
MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 127.0.0.2 () port 0 AF_INET
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/sec
540000 262144 262144 104.00 2.66
With the fix :
$ netperf -H 127.0.0.2 -l 100 -- -q 10000
MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 127.0.0.2 () port 0 AF_INET
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/sec
540000 262144 262144 104.00 0.12
EDT allows for better control of rtx timers, since TCP has
a better idea of the earliest departure time of each skb
in the rtx queue. We only have to eventually add to the
timer the difference of the EDT time with current time.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Andrey reported the following warning triggered while running CRIU tests:
tcp_clean_rtx_queue()
...
last_ackt = tcp_skb_timestamp_us(skb);
WARN_ON_ONCE(last_ackt == 0);
This is caused by 5f6188a800 ("tcp: do not change tcp_wstamp_ns
in tcp_mstamp_refresh"), as we end up having skbs in retransmit queue
with a zero skb->skb_mstamp_ns field.
We could fix this bug in different ways, like making sure
tp->tcp_wstamp_ns is not zero at socket creation, but as Neal pointed
out, we also do not want that pacing status of a repaired socket
could push tp->tcp_wstamp_ns far ahead in the future.
So we prefer changing tcp_write_xmit() to not call tcp_update_skb_after_send()
and instead do what is requested by TCP_REPAIR logic.
Fixes: 5f6188a800 ("tcp: do not change tcp_wstamp_ns in tcp_mstamp_refresh")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Andrey Vagin <avagin@openvz.org>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When TCP implements its own pacing (when no fq packet scheduler is used),
it is arming high resolution timer after a packet is sent.
But in many cases (like TCP_RR kind of workloads), this high resolution
timer expires before the application attempts to write the following
packet. This overhead also happens when the flow is ACK clocked and
cwnd limited instead of being limited by the pacing rate.
This leads to extra overhead (high number of IRQ)
Now tcp_wstamp_ns is reserved for the pacing timer only
(after commit "tcp: do not change tcp_wstamp_ns in tcp_mstamp_refresh"),
we can setup the timer only when a packet is about to be sent,
and if tcp_wstamp_ns is in the future.
This leads to a ~10% performance increase in TCP_RR workloads.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit fefa569a9d ("net_sched: sch_fq: account for schedule/timers
drifts") we added a mitigation for scheduling jitter in fq packet scheduler.
This patch does the same in TCP stack, now it is using EDT model.
Note that this mitigation is valid for both external (fq packet scheduler)
or internal TCP pacing.
This uses the same strategy than the above commit, allowing
a time credit of half the packet currently sent.
Consider following case :
An skb is sent, after an idle period of 300 usec.
The air-time (skb->len/pacing_rate) is 500 usec
Instead of setting the pacing timer to now+500 usec,
it will use now+min(500/2, 300) -> now+250usec
This is like having a token bucket with a depth of half
an skb.
Tested:
tc qdisc replace dev eth0 root pfifo_fast
Before
netperf -P0 -H remote -- -q 1000000000 # 8000Mbit
540000 262144 262144 10.00 7710.43
After :
netperf -P0 -H remote -- -q 1000000000 # 8000 Mbit
540000 262144 262144 10.00 7999.75 # Much closer to 8000Mbit target
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
sk_pacing_rate has beed introduced as a u32 field in 2013,
effectively limiting per flow pacing to 34Gbit.
We believe it is time to allow TCP to pace high speed flows
on 64bit hosts, as we now can reach 100Gbit on one TCP flow.
This patch adds no cost for 32bit kernels.
The tcpi_pacing_rate and tcpi_max_pacing_rate were already
exported as 64bit, so iproute2/ss command require no changes.
Unfortunately the SO_MAX_PACING_RATE socket option will stay
32bit and we will need to add a new option to let applications
control high pacing rates.
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 1787144 10.246.9.76:49992 10.246.9.77:36741
timer:(on,003ms,0) ino:91863 sk:2 <->
skmem:(r0,rb540000,t66440,tb2363904,f605944,w1822984,o0,bl0,d0)
ts sack bbr wscale:8,8 rto:201 rtt:0.057/0.006 mss:1448
rcvmss:536 advmss:1448
cwnd:138 ssthresh:178 bytes_acked:256699822585 segs_out:177279177
segs_in:3916318 data_segs_out:177279175
bbr:(bw:31276.8Mbps,mrtt:0,pacing_gain:1.25,cwnd_gain:2)
send 28045.5Mbps lastrcv:73333
pacing_rate 38705.0Mbps delivery_rate 22997.6Mbps
busy:73333ms unacked:135 retrans:0/157 rcv_space:14480
notsent:2085120 minrtt:0.013
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In EDT design, I made the mistake of using tcp_wstamp_ns
to store the last tcp_clock_ns() sample and to store the
pacing virtual timer.
This causes major regressions at high speed flows.
Introduce tcp_clock_cache to store last tcp_clock_ns().
This is needed because some arches have slow high-resolution
kernel time service.
tcp_wstamp_ns is only updated when a packet is sent.
Note that we can remove tcp_mstamp in the future since
tcp_mstamp is essentially tcp_clock_cache/1000, so the
apparent socket size increase is temporary.
Fixes: 9799ccb0e9 ("tcp: add tcp_wstamp_ns socket field")
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>
Previously TCP initial receive buffer is ~87KB by default and
the initial receive window is ~29KB (20 MSS). This patch changes
the two numbers to 128KB and ~64KB (rounding down to the multiples
of MSS) respectively. The patch also simplifies the calculations s.t.
the two numbers are directly controlled by sysctl tcp_rmem[1]:
1) Initial receiver buffer budget (sk_rcvbuf): while this should
be configured via sysctl tcp_rmem[1], previously tcp_fixup_rcvbuf()
always override and set a larger size when a new connection
establishes.
2) Initial receive window in SYN: previously it is set to 20
packets if MSS <= 1460. The number 20 was based on the initial
congestion window of 10: the receiver needs twice amount to
avoid being limited by the receive window upon out-of-order
delivery in the first window burst. But since this only
applies if the receiving MSS <= 1460, connection using large MTU
(e.g. to utilize receiver zero-copy) may be limited by the
receive window.
With this patch TCP memory configuration is more straight-forward and
more properly sized to modern high-speed networks by default. Several
popular stacks have been announcing 64KB rwin in SYNs as well.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Wei Wang <weiwan@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now TCP keeps track of tcp_wstamp_ns, recording the earliest
departure time of next packet, we can remove duplicate code
from tcp_internal_pacing()
This removes one ktime_get_tai_ns() call, and a divide.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP keeps track of tcp_wstamp_ns by itself, meaning sch_fq
no longer has to do it.
Thanks to this model, TCP can get more accurate RTT samples,
since pacing no longer inflates them.
This has the nice effect of removing some delays caused by FQ
quantum mechanism, causing inflated max/P99 latencies.
Also we might relax TCP Small Queue tight limits in the future,
since this new model allow TCP to build bigger batches, since
sch_fq (or a device with earliest departure time offload) ensure
these packets will be delivered on time.
Note that other protocols are not converted (they will probably
never be) so sch_fq has still support for SO_MAX_PACING_RATE
Tested:
Test showing FQ pacing quantum artifact for low-rate flows,
adding unexpected throttles for RPC flows, inflating max and P99 latencies.
The parameters chosen here are to show what happens typically when
a TCP flow has a reduced pacing rate (this can be caused by a reduced
cwin after few losses, or/and rtt above few ms)
MIBS="MIN_LATENCY,MEAN_LATENCY,MAX_LATENCY,P99_LATENCY,STDDEV_LATENCY"
Before :
$ netperf -H 10.246.7.133 -t TCP_RR -Cc -T6,6 -- -q 2000000 -r 100,100 -o $MIBS
MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.246.7.133 () port 0 AF_INET : first burst 0 : cpu bind
Minimum Latency Microseconds,Mean Latency Microseconds,Maximum Latency Microseconds,99th Percentile Latency Microseconds,Stddev Latency Microseconds
19,82.78,5279,3825,482.02
After :
$ netperf -H 10.246.7.133 -t TCP_RR -Cc -T6,6 -- -q 2000000 -r 100,100 -o $MIBS
MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.246.7.133 () port 0 AF_INET : first burst 0 : cpu bind
Minimum Latency Microseconds,Mean Latency Microseconds,Maximum Latency Microseconds,99th Percentile Latency Microseconds,Stddev Latency Microseconds
20,49.94,128,63,3.18
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Next patch will use tcp_wstamp_ns to feed internal
TCP pacing timer, so switch to CLOCK_TAI to share same base.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Switch internal TCP skb->skb_mstamp to skb->skb_mstamp_ns,
from usec units to nsec units.
Do not clear skb->tstamp before entering IP stacks in TX,
so that qdisc or devices can implement pacing based on the
earliest departure time instead of socket sk->sk_pacing_rate
Packets are fed with tcp_wstamp_ns, and following patch
will update tcp_wstamp_ns when both TCP and sch_fq switch to
the earliest departure time mechanism.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP will soon provide earliest departure time on TX skbs.
It needs to track this in a new variable.
tcp_mstamp_refresh() needs to update this variable, and
became too big to stay an inline.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are few places where TCP reads skb->skb_mstamp expecting
a value in usec unit.
skb->tstamp (aka skb->skb_mstamp) will soon store CLOCK_TAI nsec value.
Add tcp_skb_timestamp_us() to provide proper conversion when needed.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fixes gcc '-Wunused-but-set-variable' warning:
net/ipv4/tcp_output.c: In function 'tcp_collapse_retrans':
net/ipv4/tcp_output.c:2700:6: warning:
variable 'skb_size' set but not used [-Wunused-but-set-variable]
int skb_size, next_skb_size;
^
Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Introduce a new TCP stat to record the number of bytes retransmitted
(RFC4898 tcpEStatsPerfOctetsRetrans) and expose it in both tcp_info
(TCP_INFO) and opt_stats (SOF_TIMESTAMPING_OPT_STATS).
Signed-off-by: Wei Wang <weiwan@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@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>
Introduce a new TCP stat to record the number of bytes sent
(RFC4898 tcpEStatsPerfHCDataOctetsOut) and expose it in both tcp_info
(TCP_INFO) and opt_stats (SOF_TIMESTAMPING_OPT_STATS).
Signed-off-by: Wei Wang <weiwan@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@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>
Currently when a DCTCP receiver delays an ACK and receive a
data packet with a different CE mark from the previous one's, it
sends two immediate ACKs acking previous and latest sequences
respectly (for ECN accounting).
Previously sending the first ACK may mark off the delayed ACK timer
(tcp_event_ack_sent). This may subsequently prevent sending the
second ACK to acknowledge the latest sequence (tcp_ack_snd_check).
The culprit is that tcp_send_ack() assumes it always acknowleges
the latest sequence, which is not true for the first special ACK.
The fix is to not make the assumption in tcp_send_ack and check the
actual ack sequence before cancelling the delayed ACK. Further it's
safer to pass the ack sequence number as a local variable into
tcp_send_ack routine, instead of intercepting tp->rcv_nxt to avoid
future bugs like this.
Reported-by: Neal Cardwell <ncardwell@google.com>
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>
Refactor and create helpers to send the special ACK in DCTCP.
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>
After fixing the way DCTCP tracking delayed ACKs, the delayed-ACK
related callbacks are no longer needed
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Lawrence Brakmo <brakmo@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit makes BBR use only the MSS (without any headers) to
calculate pacing rates when internal TCP-layer pacing is used.
This is necessary to achieve the correct pacing behavior in this case,
since tcp_internal_pacing() uses only the payload length to calculate
pacing delays.
Signed-off-by: Kevin Yang <yyd@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
S390 bpf_jit.S is removed in net-next and had changes in 'net',
since that code isn't used any more take the removal.
TLS data structures split the TX and RX components in 'net-next',
put the new struct members from the bug fix in 'net' into the RX
part.
The 'net-next' tree had some reworking of how the ERSPAN code works in
the GRE tunneling code, overlapping with a one-line headroom
calculation fix in 'net'.
Overlapping changes in __sock_map_ctx_update_elem(), keep the bits
that read the prog members via READ_ONCE() into local variables
before using them.
Signed-off-by: David S. Miller <davem@davemloft.net>
This counter tracks number of ACK packets that the host has not sent,
thanks to ACK compression.
Sample output :
$ nstat -n;sleep 1;nstat|egrep "IpInReceives|IpOutRequests|TcpInSegs|TcpOutSegs|TcpExtTCPAckCompressed"
IpInReceives 123250 0.0
IpOutRequests 3684 0.0
TcpInSegs 123251 0.0
TcpOutSegs 3684 0.0
TcpExtTCPAckCompressed 119252 0.0
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>
linux-4.16 got support for softirq based hrtimers.
TCP can switch its pacing hrtimer to this variant, since this
avoids going through a tasklet and some atomic operations.
pacing timer logic looks like other (jiffies based) tcp timers.
v2: use hrtimer_try_to_cancel() in tcp_clear_xmit_timers()
to correctly release reference on socket if needed.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In tcp_select_initial_window(), we only set rcv_wnd to
tcp_default_init_rwnd() if current mss > (1 << wscale). Otherwise,
rcv_wnd is kept at the full receive space of the socket which is a
value way larger than tcp_default_init_rwnd().
With larger initial rcv_wnd value, receive buffer autotuning logic
takes longer to kick in and increase the receive buffer.
In a TCP throughput test where receiver has rmem[2] set to 125MB
(wscale is 11), we see the connection gets recvbuf limited at the
beginning of the connection and gets less throughput overall.
Signed-off-by: Wei Wang <weiwan@google.com>
Acked-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>
RETPOLINE made calls to tp->af_specific->md5_lookup() quite expensive,
given they have no result.
We can omit the calls for sockets that have no md5 keys.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is second part of dealing with suboptimal device gso parameters.
In first patch (350c9f484b "tcp_bbr: better deal with suboptimal GSO")
we dealt with devices having low gso_max_segs
Some devices lower gso_max_size from 64KB to 16 KB (r8152 is an example)
In order to probe an optimal cwnd, we want BBR being not sensitive
to whatever GSO constraint a device can have.
This patch removes tso_segs_goal() CC callback in favor of
min_tso_segs() for CC wanting to override sysctl_tcp_min_tso_segs
Next patch will remove bbr->tso_segs_goal since it does not have
to be persistent.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
BBR uses tcp_tso_autosize() in an attempt to probe what would be the
burst sizes and to adjust cwnd in bbr_target_cwnd() with following
gold formula :
/* Allow enough full-sized skbs in flight to utilize end systems. */
cwnd += 3 * bbr->tso_segs_goal;
But GSO can be lacking or be constrained to very small
units (ip link set dev ... gso_max_segs 2)
What we really want is to have enough packets in flight so that both
GSO and GRO are efficient.
So in the case GSO is off or downgraded, we still want to have the same
number of packets in flight as if GSO/TSO was fully operational, so
that GRO can hopefully be working efficiently.
To fix this issue, we make tcp_tso_autosize() unaware of
sk->sk_gso_max_segs
Only tcp_tso_segs() has to enforce the gso_max_segs limit.
Tested:
ethtool -K eth0 tso off gso off
tc qd replace dev eth0 root pfifo_fast
Before patch:
for f in {1..5}; do ./super_netperf 1 -H lpaa24 -- -K bbr; done
691 (ss -temoi shows cwnd is stuck around 6 )
667
651
631
517
After patch :
# for f in {1..5}; do ./super_netperf 1 -H lpaa24 -- -K bbr; done
1733 (ss -temoi shows cwnd is around 386 )
1778
1746
1781
1718
Fixes: 0f8782ea14 ("tcp_bbr: add BBR congestion control")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since all skbs in write/rtx queues have CHECKSUM_PARTIAL,
we can remove dead code.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We no longer have skbs with skb->ip_summed == CHECKSUM_NONE
in TCP write queues.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Avoid SKB coalescing if eor bit is set in one of the relevant
SKBs.
Fixes: c134ecb878 ("tcp: Make use of MSG_EOR in tcp_sendmsg")
Signed-off-by: Ilya Lesokhin <ilyal@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Adds support for calling sock_ops BPF program when there is a
retransmission. Three arguments are used; one for the sequence number,
another for the number of segments retransmitted, and the last one for
the return value of tcp_transmit_skb (0 => success).
Does not include syn-ack retransmissions.
New op: BPF_SOCK_OPS_RETRANS_CB.
Signed-off-by: Lawrence Brakmo <brakmo@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Adds support for passing up to 4 arguments to sock_ops bpf functions. It
reusues the reply union, so the bpf_sock_ops structures are not
increased in size.
Signed-off-by: Lawrence Brakmo <brakmo@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
The two conditions triggering BUG_ON() are somewhat unrelated:
the tcp_skb_pcount() check is meant to catch TSO flaws, the
second one checks sanity of congestion window bookkeeping.
Split them into two separate BUG_ON() assertions on two lines,
so that we know which one actually triggers, when they do.
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch enables tail loss probe in cwnd reduction (CWR) state
to detect potential losses. Prior to this patch, since the sender
uses PRR to determine the cwnd in CWR state, the combination of
CWR+PRR plus tcp_tso_should_defer() could cause unnecessary stalls
upon losses: PRR makes cwnd so gentle that tcp_tso_should_defer()
defers sending wait for more ACKs. The ACKs may not come due to
packet losses.
Disallowing TLP when there is unused cwnd had the primary effect
of disallowing TLP when there is TSO deferral, Nagle deferral,
or we hit the rwin limit. Because basically every application
write() or incoming ACK will cause us to run tcp_write_xmit()
to see if we can send more, and then if we sent something we call
tcp_schedule_loss_probe() to see if we should schedule a TLP. At
that point, there are a few common reasons why some cwnd budget
could still be unused:
(a) rwin limit
(b) nagle check
(c) TSO deferral
(d) TSQ
For (d), after the next packet tx completion the TSQ mechanism
will allow us to send more packets, so we don't really need a
TLP (in practice it shouldn't matter whether we schedule one
or not). But for (a), (b), (c) the sender won't send any more
packets until it gets another ACK. But if the whole flight was
lost, or all the ACKs were lost, then we won't get any more ACKs,
and ideally we should schedule and send a TLP to get more feedback.
In particular for a long time we have wanted some kind of timer for
TSO deferral, and at least this would give us some kind of timer
Reported-by: Steve Ibanez <sibanez@stanford.edu>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Reviewed-by: Nandita Dukkipati <nanditad@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix the TLP scheduling logic so that when scheduling a TLP probe, we
ensure that the estimated time at which an RTO would fire accounts for
the fact that ACKs indicating forward progress should push back RTO
times.
After the following fix:
df92c8394e ("tcp: fix xmit timer to only be reset if data ACKed/SACKed")
we had an unintentional behavior change in the following kind of
scenario: suppose the RTT variance has been very low recently. Then
suppose we send out a flight of N packets and our RTT is 100ms:
t=0: send a flight of N packets
t=100ms: receive an ACK for N-1 packets
The response before df92c8394e that was:
-> schedule a TLP for now + RTO_interval
The response after df92c8394e is:
-> schedule a TLP for t=0 + RTO_interval
Since RTO_interval = srtt + RTT_variance, this means that we have
scheduled a TLP timer at a point in the future that only accounts for
RTT_variance. If the RTT_variance term is small, this means that the
timer fires soon.
Before df92c8394e this would not happen, because in that code, when
we receive an ACK for a prefix of flight, we did:
1) Near the top of tcp_ack(), switch from TLP timer to RTO
at write_queue_head->paket_tx_time + RTO_interval:
if (icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
tcp_rearm_rto(sk);
2) In tcp_clean_rtx_queue(), update the RTO to now + RTO_interval:
if (flag & FLAG_ACKED) {
tcp_rearm_rto(sk);
3) In tcp_ack() after tcp_fastretrans_alert() switch from RTO
to TLP at now + RTO_interval:
if (icsk->icsk_pending == ICSK_TIME_RETRANS)
tcp_schedule_loss_probe(sk);
In df92c8394e we removed that 3-phase dance, and instead directly
set the TLP timer once: we set the TLP timer in cases like this to
write_queue_head->packet_tx_time + RTO_interval. So if the RTT
variance is small, then this means that this is setting the TLP timer
to fire quite soon. This means if the ACK for the tail of the flight
takes longer than an RTT to arrive (often due to delayed ACKs), then
the TLP timer fires too quickly.
Fixes: df92c8394e ("tcp: fix xmit timer to only be reset if data ACKed/SACKed")
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>
Pull networking updates from David Miller:
"Highlights:
1) Maintain the TCP retransmit queue using an rbtree, with 1GB
windows at 100Gb this really has become necessary. From Eric
Dumazet.
2) Multi-program support for cgroup+bpf, from Alexei Starovoitov.
3) Perform broadcast flooding in hardware in mv88e6xxx, from Andrew
Lunn.
4) Add meter action support to openvswitch, from Andy Zhou.
5) Add a data meta pointer for BPF accessible packets, from Daniel
Borkmann.
6) Namespace-ify almost all TCP sysctl knobs, from Eric Dumazet.
7) Turn on Broadcom Tags in b53 driver, from Florian Fainelli.
8) More work to move the RTNL mutex down, from Florian Westphal.
9) Add 'bpftool' utility, to help with bpf program introspection.
From Jakub Kicinski.
10) Add new 'cpumap' type for XDP_REDIRECT action, from Jesper
Dangaard Brouer.
11) Support 'blocks' of transformations in the packet scheduler which
can span multiple network devices, from Jiri Pirko.
12) TC flower offload support in cxgb4, from Kumar Sanghvi.
13) Priority based stream scheduler for SCTP, from Marcelo Ricardo
Leitner.
14) Thunderbolt networking driver, from Amir Levy and Mika Westerberg.
15) Add RED qdisc offloadability, and use it in mlxsw driver. From
Nogah Frankel.
16) eBPF based device controller for cgroup v2, from Roman Gushchin.
17) Add some fundamental tracepoints for TCP, from Song Liu.
18) Remove garbage collection from ipv6 route layer, this is a
significant accomplishment. From Wei Wang.
19) Add multicast route offload support to mlxsw, from Yotam Gigi"
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next: (2177 commits)
tcp: highest_sack fix
geneve: fix fill_info when link down
bpf: fix lockdep splat
net: cdc_ncm: GetNtbFormat endian fix
openvswitch: meter: fix NULL pointer dereference in ovs_meter_cmd_reply_start
netem: remove unnecessary 64 bit modulus
netem: use 64 bit divide by rate
tcp: Namespace-ify sysctl_tcp_default_congestion_control
net: Protect iterations over net::fib_notifier_ops in fib_seq_sum()
ipv6: set all.accept_dad to 0 by default
uapi: fix linux/tls.h userspace compilation error
usbnet: ipheth: prevent TX queue timeouts when device not ready
vhost_net: conditionally enable tx polling
uapi: fix linux/rxrpc.h userspace compilation errors
net: stmmac: fix LPI transitioning for dwmac4
atm: horizon: Fix irq release error
net-sysfs: trigger netlink notification on ifalias change via sysfs
openvswitch: Using kfree_rcu() to simplify the code
openvswitch: Make local function ovs_nsh_key_attr_size() static
openvswitch: Fix return value check in ovs_meter_cmd_features()
...
I had many reports that TSQ logic breaks wifi aggregation.
Current logic is to allow up to 1 ms of bytes to be queued into qdisc
and drivers queues.
But Wifi aggregation needs a bigger budget to allow bigger rates to
be discovered by various TCP Congestion Controls algorithms.
This patch adds an extra socket field, allowing wifi drivers to select
another log scale to derive TCP Small Queue credit from current pacing
rate.
Initial value is 10, meaning that this patch does not change current
behavior.
We expect wifi drivers to set this field to smaller values (tests have
been done with values from 6 to 9)
They would have to use following template :
if (skb->sk && skb->sk->sk_pacing_shift != MY_PACING_SHIFT)
skb->sk->sk_pacing_shift = MY_PACING_SHIFT;
Ref: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1670041
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Johannes Berg <johannes.berg@intel.com>
Cc: Toke Høiland-Jørgensen <toke@toke.dk>
Cc: Kir Kolyshkin <kir@openvz.org>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Replace the reordering distance measurement in packet unit with
sequence based approach. Previously it trackes the number of "packets"
toward the forward ACK (i.e. highest sacked sequence)in a state
variable "fackets_out".
Precisely measuring reordering degree on packet distance has not much
benefit, as the degree constantly changes by factors like path, load,
and congestion window. It is also complicated and prone to arcane bugs.
This patch replaces with sequence-based approach that's much simpler.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@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>
FACK loss detection has been disabled by default and the
successor RACK subsumed FACK and can handle reordering better.
This patch removes FACK to simplify TCP loss recovery.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@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>
Note that when a new netns is created, it inherits its
sysctl_tcp_rmem and sysctl_tcp_wmem from initial netns.
This change is needed so that we can refine TCP rcvbuf autotuning,
to take RTT into consideration.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Wei Wang <weiwan@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_init_nondata_skb() is fed with freshly allocated skbs.
They already have a cleared csum field, no need to clear it again.
This is based on Neal review on commit 3b11775033 ("tcp: do not mangle
skb->cb[] in tcp_make_synack()"), noticing I did not clear skb->csum.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Files removed in 'net-next' had their license header updated
in 'net'. We take the remove from 'net-next'.
Signed-off-by: David S. Miller <davem@davemloft.net>
While stress testing MTU probing, we had crashes in list_del() that we root-caused
to the fact that tcp_fragment() is unconditionally inserting the freshly allocated
skb into tsorted_sent_queue list.
But this list is supposed to contain skbs that were sent.
This was mostly harmless until MTU probing was enabled.
Fortunately we can use the tcp_queue enum added later (but in same linux version)
for rtx-rb-tree to fix the bug.
Fixes: e2080072ed ("tcp: new list for sent but unacked skbs for RACK recovery")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Priyaranjan Jha <priyarjha@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>
Christoph Paasch sent a patch to address the following issue :
tcp_make_synack() is leaving some TCP private info in skb->cb[],
then send the packet by other means than tcp_transmit_skb()
tcp_transmit_skb() makes sure to clear skb->cb[] to not confuse
IPv4/IPV6 stacks, but we have no such cleanup for SYNACK.
tcp_make_synack() should not use tcp_init_nondata_skb() :
tcp_init_nondata_skb() really should be limited to skbs put in write/rtx
queues (the ones that are only sent via tcp_transmit_skb())
This patch fixes the issue and should even save few cpu cycles ;)
Fixes: 971f10eca1 ("tcp: better TCP_SKB_CB layout to reduce cache line misses")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Christoph Paasch <cpaasch@apple.com>
Reviewed-by: Christoph Paasch <cpaasch@apple.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This tracepoint can be used to trace synack retransmits. It maintains
pointer to struct request_sock.
We cannot simply reuse trace_tcp_retransmit_skb() here, because the
sk here is the LISTEN socket. The IP addresses and ports should be
extracted from struct request_sock.
Note that, like many other tracepoints, this patch uses IS_ENABLED
in TP_fast_assign macro, which triggers sparse warning like:
./include/trace/events/tcp.h:274:1: error: directive in argument list
./include/trace/events/tcp.h:281:1: error: directive in argument list
However, there is no good solution to avoid these warnings. To the
best of our knowledge, these warnings are harmless.
Signed-off-by: Song Liu <songliubraving@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Smooth Cong Wang's bug fix into 'net-next'. Basically put
the bulk of the tcf_block_put() logic from 'net' into
tcf_block_put_ext(), but after the offload unbind.
Signed-off-by: David S. Miller <davem@davemloft.net>
Based on SNMP values provided by Roman, Yuchung made the observation
that some crashes in tcp_sacktag_walk() might be caused by MTU probing.
Looking at tcp_mtu_probe(), I found that when a new skb was placed
in front of the write queue, we were not updating tcp highest sack.
If one skb is freed because all its content was copied to the new skb
(for MTU probing), then tp->highest_sack could point to a now freed skb.
Bad things would then happen, including infinite loops.
This patch renames tcp_highest_sack_combine() and uses it
from tcp_mtu_probe() to fix the bug.
Note that I also removed one test against tp->sacked_out,
since we want to replace tp->highest_sack regardless of whatever
condition, since keeping a stale pointer to freed skb is a recipe
for disaster.
Fixes: a47e5a988a ("[TCP]: Convert highest_sack to sk_buff to allow direct access")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Alexei Starovoitov <alexei.starovoitov@gmail.com>
Reported-by: Roman Gushchin <guro@fb.com>
Reported-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Several conflicts here.
NFP driver bug fix adding nfp_netdev_is_nfp_repr() check to
nfp_fl_output() needed some adjustments because the code block is in
an else block now.
Parallel additions to net/pkt_cls.h and net/sch_generic.h
A bug fix in __tcp_retransmit_skb() conflicted with some of
the rbtree changes in net-next.
The tc action RCU callback fixes in 'net' had some overlap with some
of the recent tcf_block reworking.
Signed-off-by: David S. Miller <davem@davemloft.net>
Eric Dumazet