Count the number of packets that a TCP connection marks lost.
Congestion control modules can use this loss rate information for more
intelligent decisions about how fast to send.
Specifically, this is used in TCP BBR policer detection. BBR uses a
high packet loss rate as one signal in its policer detection and
policer bandwidth estimation algorithm.
The BBR policer detection algorithm cannot simply track retransmits,
because a retransmit can be (and often is) an indicator of packets
lost long, long ago. This is particularly true in a long CA_Loss
period that repairs the initial massive losses when a policer kicks
in.
Signed-off-by: Van Jacobson <vanj@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Refactor the TCP min_rtt code to reuse the new win_minmax library in
lib/win_minmax.c to simplify the TCP code.
This is a pure refactor: the functionality is exactly the same. We
just moved the windowed min code to make TCP easier to read and
maintain, and to allow other parts of the kernel to use the windowed
min/max filter code.
Signed-off-by: Van Jacobson <vanj@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Over the years, TCP BDP has increased by several orders of magnitude,
and some people are considering to reach the 2 Gbytes limit.
Even with current window scale limit of 14, ~1 Gbytes maps to ~740,000
MSS.
In presence of packet losses (or reorders), TCP stores incoming packets
into an out of order queue, and number of skbs sitting there waiting for
the missing packets to be received can be in the 10^5 range.
Most packets are appended to the tail of this queue, and when
packets can finally be transferred to receive queue, we scan the queue
from its head.
However, in presence of heavy losses, we might have to find an arbitrary
point in this queue, involving a linear scan for every incoming packet,
throwing away cpu caches.
This patch converts it to a RB tree, to get bounded latencies.
Yaogong wrote a preliminary patch about 2 years ago.
Eric did the rebase, added ofo_last_skb cache, polishing and tests.
Tested with network dropping between 1 and 10 % packets, with good
success (about 30 % increase of throughput in stress tests)
Next step would be to also use an RB tree for the write queue at sender
side ;)
Signed-off-by: Yaogong Wang <wygivan@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Acked-By: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Per RFC4898, they count segments sent/received
containing a positive length data segment (that includes
retransmission segments carrying data). Unlike
tcpi_segs_out/in, tcpi_data_segs_out/in excludes segments
carrying no data (e.g. pure ack).
The patch also updates the segs_in in tcp_fastopen_add_skb()
so that segs_in >= data_segs_in property is kept.
Together with retransmission data, tcpi_data_segs_out
gives a better signal on the rxmit rate.
v6: Rebase on the latest net-next
v5: Eric pointed out that checking skb->len is still needed in
tcp_fastopen_add_skb() because skb can carry a FIN without data.
Hence, instead of open coding segs_in and data_segs_in, tcp_segs_in()
helper is used. Comment is added to the fastopen case to explain why
segs_in has to be reset and tcp_segs_in() has to be called before
__skb_pull().
v4: Add comment to the changes in tcp_fastopen_add_skb()
and also add remark on this case in the commit message.
v3: Add const modifier to the skb parameter in tcp_segs_in()
v2: Rework based on recent fix by Eric:
commit a9d99ce28e ("tcp: fix tcpi_segs_in after connection establishment")
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Cc: Chris Rapier <rapier@psc.edu>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Marcelo Ricardo Leitner <mleitner@redhat.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_hdrlen is wasteful if you already have a pointer to struct tcphdr.
This splits the size calculation into a helper function that can be
used if a struct tcphdr is already available.
Signed-off-by: Craig Gallek <kraig@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch changes the accounting of how many packets are
newly acked or sacked when the sender receives an ACK.
The current approach basically computes
newly_acked_sacked = (prior_packets - prior_sacked) -
(tp->packets_out - tp->sacked_out)
where prior_packets and prior_sacked out are snapshot
at the beginning of the ACK processing.
The new approach tracks the delivery information via a new
TCP state variable "delivered" which monotically increases
as new packets are delivered in order or out-of-order.
The reason for this change is that the current approach is
brittle that produces negative or inaccurate estimate.
1) For non-SACK connections, an ACK that advances the SND.UNA
could reset the DUPACK counters (tp->sacked_out) in
tcp_process_loss() or tcp_fastretrans_alert(). This inflates
the inflight suddenly and causes under-estimate or even
negative estimate. Here is a real example:
before after (processing ACK)
packets_out 75 73
sacked_out 23 0
ca state Loss Open
The old approach computes (75-23) - (73 - 0) = -21 delivered
while the new approach computes 1 delivered since it
considers the 2nd-24th packets are delivered OOO.
2) MSS change would re-count packets_out and sacked_out so
the estimate is in-accurate and can even become negative.
E.g., the inflight is doubled when MSS is halved.
3) Spurious retransmission signaled by DSACK is not accounted
The new approach is simpler and more robust. For SACK connections,
tp->delivered increments as packets are being acked or sacked in
SACK and ACK processing.
For non-sack connections, it's done in tcp_remove_reno_sacks() and
tcp_add_reno_sack(). When an ACK advances the SND.UNA, tp->delivered
is incremented by the number of packets ACKed (less the current
number of DUPACKs received plus one packet hole). Upon receiving
a DUPACK, tp->delivered is incremented assuming one out-of-order
packet is delivered.
Upon receiving a DSACK, tp->delivered is incremtened assuming one
retransmission is delivered in tcp_sacktag_write_queue().
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
For the reasons explained in commit ce1050089c ("tcp/dccp: fix
ireq->pktopts race"), we need to make sure we do not access
req->saved_syn unless we own the request sock.
This fixes races for listeners using TCP_SAVE_SYN option.
Fixes: e994b2f0fb ("tcp: do not lock listener to process SYN packets")
Fixes: 079096f103 ("tcp/dccp: install syn_recv requests into ehash table")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Ying Cai <ycai@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Allowing an application to set whatever limit for
the list of recently RST fastopen sessions [1] is not wise,
as it open ways to deplete kernel memory.
Cap the user provided limit by somaxconn sysctl,
like listen() backlog.
[1] https://tools.ietf.org/html/rfc7413#section-5.1
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch is the first half of the RACK loss recovery.
RACK loss recovery uses the notion of time instead
of packet sequence (FACK) or counts (dupthresh). It's inspired by the
previous FACK heuristic in tcp_mark_lost_retrans(): when a limited
transmit (new data packet) is sacked, then current retransmitted
sequence below the newly sacked sequence must been lost,
since at least one round trip time has elapsed.
But it has several limitations:
1) can't detect tail drops since it depends on limited transmit
2) is disabled upon reordering (assumes no reordering)
3) only enabled in fast recovery ut not timeout recovery
RACK (Recently ACK) addresses these limitations with the notion
of time instead: a packet P1 is lost if a later packet P2 is s/acked,
as at least one round trip has passed.
Since RACK cares about the time sequence instead of the data sequence
of packets, it can detect tail drops when later retransmission is
s/acked while FACK or dupthresh can't. For reordering RACK uses a
dynamically adjusted reordering window ("reo_wnd") to reduce false
positives on ever (small) degree of reordering.
This patch implements tcp_advanced_rack() which tracks the
most recent transmission time among the packets that have been
delivered (ACKed or SACKed) in tp->rack.mstamp. This timestamp
is the key to determine which packet has been lost.
Consider an example that the sender sends six packets:
T1: P1 (lost)
T2: P2
T3: P3
T4: P4
T100: sack of P2. rack.mstamp = T2
T101: retransmit P1
T102: sack of P2,P3,P4. rack.mstamp = T4
T205: ACK of P4 since the hole is repaired. rack.mstamp = T101
We need to be careful about spurious retransmission because it may
falsely advance tp->rack.mstamp by an RTT or an RTO, causing RACK
to falsely mark all packets lost, just like a spurious timeout.
We identify spurious retransmission by the ACK's TS echo value.
If TS option is not applicable but the retransmission is acknowledged
less than min-RTT ago, it is likely to be spurious. We refrain from
using the transmission time of these spurious retransmissions.
The second half is implemented in the next patch that marks packet
lost using RACK timestamp.
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>
Remove the existing lost retransmit detection because RACK subsumes
it completely. This also stops the overloading the ack_seq field of
the skb control block.
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>
Kathleen Nichols' algorithm for tracking the minimum RTT of a
data stream over some measurement window. It uses constant space
and constant time per update. Yet it almost always delivers
the same minimum as an implementation that has to keep all
the data in the window. The measurement window is tunable via
sysctl.net.ipv4.tcp_min_rtt_wlen with a default value of 5 minutes.
The algorithm keeps track of the best, 2nd best & 3rd best min
values, maintaining an invariant that the measurement time of
the n'th best >= n-1'th best. It also makes sure that the three
values are widely separated in the time window since that bounds
the worse case error when that data is monotonically increasing
over the window.
Upon getting a new min, we can forget everything earlier because
it has no value - the new min is less than everything else in the
window by definition and it's the most recent. So we restart fresh
on every new min and overwrites the 2nd & 3rd choices. The same
property holds for the 2nd & 3rd best.
Therefore we have to maintain two invariants to maximize the
information in the samples, one on values (1st.v <= 2nd.v <=
3rd.v) and the other on times (now-win <=1st.t <= 2nd.t <= 3rd.t <=
now). These invariants determine the structure of the code
The RTT input to the windowed filter is the minimum RTT measured
from ACK or SACK, or as the last resort from TCP timestamps.
The accessor tcp_min_rtt() returns the minimum RTT seen in the
window. ~0U indicates it is not available. The minimum is 1usec
even if the true RTT is below that.
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>
Reducing tcp_timewait_sock from 280 bytes to 272 bytes
allows SLAB to pack 15 objects per page instead of 14 (on x86)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While auditing TCP stack for upcoming 'lockless' listener changes,
I found I had to change fastopen_init_queue() to properly init the object
before publishing it.
Otherwise an other cpu could try to lock the spinlock before it gets
properly initialized.
Instead of adding appropriate barriers, just remove dynamic memory
allocations :
- Structure is 28 bytes on 64bit arches. Using additional 8 bytes
for holding a pointer seems overkill.
- Two listeners can share same cache line and performance would suffer.
If we really want to save few bytes, we would instead dynamically allocate
whole struct request_sock_queue in the future.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently SYN/ACK RTT is measured in jiffies. For LAN the SYN/ACK
RTT is often measured as 0ms or sometimes 1ms, which would affect
RTT estimation and min RTT samping used by some congestion control.
This patch improves SYN/ACK RTT to be usec resolution if platform
supports it. While the timestamping of SYN/ACK is done in request
sock, the RTT measurement is carefully arranged to avoid storing
another u64 timestamp in tcp_sock.
For regular handshake w/o SYNACK retransmission, the RTT is sampled
right after the child socket is created and right before the request
sock is released (tcp_check_req() in tcp_minisocks.c)
For Fast Open the child socket is already created when SYN/ACK was
sent, the RTT is sampled in tcp_rcv_state_process() after processing
the final ACK an right before the request socket is released.
If the SYN/ACK was retransmistted or SYN-cookie was used, we rely
on TCP timestamps to measure the RTT. The sample is taken at the
same place in tcp_rcv_state_process() after the timestamp values
are validated in tcp_validate_incoming(). Note that we do not store
TS echo value in request_sock for SYN-cookies, because the value
is already stored in tp->rx_opt used by tcp_ack_update_rtt().
One side benefit is that the RTT measurement now happens before
initializing congestion control (of the passive side). Therefore
the congestion control can use the SYN/ACK RTT.
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>
In commit b73c3d0e4f ("net: Save TX flow hash in sock and set in skbuf
on xmit"), Tom provided a l4 hash to most outgoing TCP packets.
We'd like to provide one as well for SYNACK packets, so that all packets
of a given flow share same txhash, to later enable bonding driver to
also use skb->hash to perform slave selection.
Note that a SYNACK retransmit shuffles the tx hash, as Tom did
in commit 265f94ff54 ("net: Recompute sk_txhash on negative routing
advice") for established sockets.
This has nice effect making TCP flows resilient to some kind of black
holes, even at connection establish phase.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Tom Herbert <tom@herbertland.com>
Cc: Mahesh Bandewar <maheshb@google.com>
Acked-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
drivers/net/ethernet/cadence/macb.c
drivers/net/phy/phy.c
include/linux/skbuff.h
net/ipv4/tcp.c
net/switchdev/switchdev.c
Switchdev was a case of RTNH_H_{EXTERNAL --> OFFLOAD}
renaming overlapping with net-next changes of various
sorts.
phy.c was a case of two changes, one adding a local
variable to a function whilst the second was removing
one.
tcp.c overlapped a deadlock fix with the addition of new tcp_info
statistic values.
macb.c involved the addition of two zyncq device entries.
skbuff.h involved adding back ipv4_daddr to nf_bridge_info
whilst net-next changes put two other existing members of
that struct into a union.
Signed-off-by: David S. Miller <davem@davemloft.net>
Taking socket spinlock in tcp_get_info() can deadlock, as
inet_diag_dump_icsk() holds the &hashinfo->ehash_locks[i],
while packet processing can use the reverse locking order.
We could avoid this locking for TCP_LISTEN states, but lockdep would
certainly get confused as all TCP sockets share same lockdep classes.
[ 523.722504] ======================================================
[ 523.728706] [ INFO: possible circular locking dependency detected ]
[ 523.734990] 4.1.0-dbg-DEV #1676 Not tainted
[ 523.739202] -------------------------------------------------------
[ 523.745474] ss/18032 is trying to acquire lock:
[ 523.750002] (slock-AF_INET){+.-...}, at: [<ffffffff81669d44>] tcp_get_info+0x2c4/0x360
[ 523.758129]
[ 523.758129] but task is already holding lock:
[ 523.763968] (&(&hashinfo->ehash_locks[i])->rlock){+.-...}, at: [<ffffffff816bcb75>] inet_diag_dump_icsk+0x1d5/0x6c0
[ 523.774661]
[ 523.774661] which lock already depends on the new lock.
[ 523.774661]
[ 523.782850]
[ 523.782850] the existing dependency chain (in reverse order) is:
[ 523.790326]
-> #1 (&(&hashinfo->ehash_locks[i])->rlock){+.-...}:
[ 523.796599] [<ffffffff811126bb>] lock_acquire+0xbb/0x270
[ 523.802565] [<ffffffff816f5868>] _raw_spin_lock+0x38/0x50
[ 523.808628] [<ffffffff81665af8>] __inet_hash_nolisten+0x78/0x110
[ 523.815273] [<ffffffff816819db>] tcp_v4_syn_recv_sock+0x24b/0x350
[ 523.822067] [<ffffffff81684d41>] tcp_check_req+0x3c1/0x500
[ 523.828199] [<ffffffff81682d09>] tcp_v4_do_rcv+0x239/0x3d0
[ 523.834331] [<ffffffff816842fe>] tcp_v4_rcv+0xa8e/0xc10
[ 523.840202] [<ffffffff81658fa3>] ip_local_deliver_finish+0x133/0x3e0
[ 523.847214] [<ffffffff81659a9a>] ip_local_deliver+0xaa/0xc0
[ 523.853440] [<ffffffff816593b8>] ip_rcv_finish+0x168/0x5c0
[ 523.859624] [<ffffffff81659db7>] ip_rcv+0x307/0x420
Lets use u64_sync infrastructure instead. As a bonus, 64bit
arches get optimized, as these are nop for them.
Fixes: 0df48c26d8 ("tcp: add tcpi_bytes_acked to tcp_info")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch tracks the total number of inbound and outbound segments on a
TCP socket. One may use this number to have an idea on connection
quality when compared against the retransmissions.
RFC4898 named these : tcpEStatsPerfSegsIn and tcpEStatsPerfSegsOut
These are a 32bit field each and can be fetched both from TCP_INFO
getsockopt() if one has a handle on a TCP socket, or from inet_diag
netlink facility (iproute2/ss patch will follow)
Note that tp->segs_out was placed near tp->snd_nxt for good data
locality and minimal performance impact, while tp->segs_in was placed
near tp->bytes_received for the same reason.
Join work with Eric Dumazet.
Note that received SYN are accounted on the listener, but sent SYNACK
are not accounted.
Signed-off-by: Marcelo Ricardo Leitner <mleitner@redhat.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch allows a server application to get the TCP SYN headers for
its passive connections. This is useful if the server is doing
fingerprinting of clients based on SYN packet contents.
Two socket options are added: TCP_SAVE_SYN and TCP_SAVED_SYN.
The first is used on a socket to enable saving the SYN headers
for child connections. This can be set before or after the listen()
call.
The latter is used to retrieve the SYN headers for passive connections,
if the parent listener has enabled TCP_SAVE_SYN.
TCP_SAVED_SYN is read once, it frees the saved SYN headers.
The data returned in TCP_SAVED_SYN are network (IPv4/IPv6) and TCP
headers.
Original patch was written by Tom Herbert, I changed it to not hold
a full skb (and associated dst and conntracking reference).
We have used such patch for about 3 years at Google.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Tested-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch tracks total number of payload bytes received on a TCP socket.
This is the sum of all changes done to tp->rcv_nxt
RFC4898 named this : tcpEStatsAppHCThruOctetsReceived
This is a 64bit field, and can be fetched both from TCP_INFO
getsockopt() if one has a handle on a TCP socket, or from inet_diag
netlink facility (iproute2/ss patch will follow)
Note that tp->bytes_received was placed near tp->rcv_nxt for
best data locality and minimal performance impact.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Matt Mathis <mattmathis@google.com>
Cc: Eric Salo <salo@google.com>
Cc: Martin Lau <kafai@fb.com>
Cc: Chris Rapier <rapier@psc.edu>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch tracks total number of bytes acked for a TCP socket.
This is the sum of all changes done to tp->snd_una, and allows
for precise tracking of delivered data.
RFC4898 named this : tcpEStatsAppHCThruOctetsAcked
This is a 64bit field, and can be fetched both from TCP_INFO
getsockopt() if one has a handle on a TCP socket, or from inet_diag
netlink facility (iproute2/ss patch will follow)
Note that tp->bytes_acked was placed near tp->snd_una for
best data locality and minimal performance impact.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Cc: Matt Mathis <mattmathis@google.com>
Cc: Eric Salo <salo@google.com>
Cc: Martin Lau <kafai@fb.com>
Cc: Chris Rapier <rapier@psc.edu>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fast Open has been using an experimental option with a magic number
(RFC6994). This patch makes the client by default use the RFC7413
option (34) to get and send Fast Open cookies. This patch makes
the client solicit cookies from a given server first with the
RFC7413 option. If that fails to elicit a cookie, then it tries
the RFC6994 experimental option. If that also fails, it uses the
RFC7413 option on all subsequent connect attempts. If the server
returns a Fast Open cookie then the client caches the form of the
option that successfully elicited a cookie, and uses that form on
later connects when it presents that cookie.
The idea is to gradually obsolete the use of experimental options as
the servers and clients upgrade, while keeping the interoperability
meanwhile.
Signed-off-by: Daniel Lee <Longinus00@gmail.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>
Fast Open has been using the experimental option with a magic number
(RFC6994) to request and grant Fast Open cookies. This patch enables
the server to support the official IANA option 34 in RFC7413 in
addition.
The change has passed all existing Fast Open tests with both
old and new options at Google.
Signed-off-by: Daniel Lee <Longinus00@gmail.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>
The listener field in struct tcp_request_sock is a pointer
back to the listener. We now have req->rsk_listener, so TCP
only needs one boolean and not a full pointer.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TSO relies on ability to defer sending a small amount of packets.
Heuristic is to wait for future ACKS in hope to send more packets at once.
Current algorithm uses a per socket tso_deferred field as a pseudo timer.
This pseudo timer relies on future ACK, but there is no guarantee
we receive them in time.
Fix would be to use a real timer, but cost of such timer is probably too
expensive for typical cases.
This patch changes the logic to test the time of last transmit,
because we should not add bursts of more than 1ms for any given flow.
We've used this patch for about two years at Google, before FQ/pacing
as it would reduce a fair amount of bursts.
Signed-off-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>
Ensure that in state FIN_WAIT2 or TIME_WAIT, where the connection is
represented by a tcp_timewait_sock, we rate limit dupacks in response
to incoming packets (a) with TCP timestamps that fail PAWS checks, or
(b) with sequence numbers that are out of the acceptable window.
We do not send a dupack in response to out-of-window packets if it has
been less than sysctl_tcp_invalid_ratelimit (default 500ms) since we
last sent a dupack in response to an out-of-window packet.
Reported-by: Avery Fay <avery@mixpanel.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>
Signed-off-by: David S. Miller <davem@davemloft.net>
Ensure that in state ESTABLISHED, where the connection is represented
by a tcp_sock, we rate limit dupacks in response to incoming packets
(a) with TCP timestamps that fail PAWS checks, or (b) with sequence
numbers or ACK numbers that are out of the acceptable window.
We do not send a dupack in response to out-of-window packets if it has
been less than sysctl_tcp_invalid_ratelimit (default 500ms) since we
last sent a dupack in response to an out-of-window packet.
There is already a similar (although global) rate-limiting mechanism
for "challenge ACKs". When deciding whether to send a challence ACK,
we first consult the new per-connection rate limit, and then the
global rate limit.
Reported-by: Avery Fay <avery@mixpanel.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>
Signed-off-by: David S. Miller <davem@davemloft.net>
In the SYN_RECV state, where the TCP connection is represented by
tcp_request_sock, we now rate-limit SYNACKs in response to a client's
retransmitted SYNs: we do not send a SYNACK in response to client SYN
if it has been less than sysctl_tcp_invalid_ratelimit (default 500ms)
since we last sent a SYNACK in response to a client's retransmitted
SYN.
This allows the vast majority of legitimate client connections to
proceed unimpeded, even for the most aggressive platforms, iOS and
MacOS, which actually retransmit SYNs 1-second intervals for several
times in a row. They use SYN RTO timeouts following the progression:
1,1,1,1,1,2,4,8,16,32.
Reported-by: Avery Fay <avery@mixpanel.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>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 95bd09eb27 ("tcp: TSO packets automatic sizing") tried to
control TSO size, but did this at the wrong place (sendmsg() time)
At sendmsg() time, we might have a pessimistic view of flow rate,
and we end up building very small skbs (with 2 MSS per skb).
This is bad because :
- It sends small TSO packets even in Slow Start where rate quickly
increases.
- It tends to make socket write queue very big, increasing tcp_ack()
processing time, but also increasing memory needs, not necessarily
accounted for, as fast clones overhead is currently ignored.
- Lower GRO efficiency and more ACK packets.
Servers with a lot of small lived connections suffer from this.
Lets instead fill skbs as much as possible (64KB of payload), but split
them at xmit time, when we have a precise idea of the flow rate.
skb split is actually quite efficient.
Patch looks bigger than necessary, because TCP Small Queue decision now
has to take place after the eventual split.
As Neal suggested, introduce a new tcp_tso_autosize() helper, so that
tcp_tso_should_defer() can be synchronized on same goal.
Rename tp->xmit_size_goal_segs to tp->gso_segs, as this variable
contains number of mss that we can put in GSO packet, and is not
related to the autosizing goal anymore.
Tested:
40 ms rtt link
nstat >/dev/null
netperf -H remote -l -2000000 -- -s 1000000
nstat | egrep "IpInReceives|IpOutRequests|TcpOutSegs|IpExtOutOctets"
Before patch :
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/s
87380 2000000 2000000 0.36 44.22
IpInReceives 600 0.0
IpOutRequests 599 0.0
TcpOutSegs 1397 0.0
IpExtOutOctets 2033249 0.0
After patch :
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/sec
87380 2000000 2000000 0.36 44.27
IpInReceives 221 0.0
IpOutRequests 232 0.0
TcpOutSegs 1397 0.0
IpExtOutOctets 2013953 0.0
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While testing upcoming Yaogong patch (converting out of order queue
into an RB tree), I hit the max reordering level of linux TCP stack.
Reordering level was limited to 127 for no good reason, and some
network setups [1] can easily reach this limit and get limited
throughput.
Allow a new max limit of 300, and add a sysctl to allow admins to even
allow bigger (or lower) values if needed.
[1] Aggregation of links, per packet load balancing, fabrics not doing
deep packet inspections, alternative TCP congestion modules...
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Yaogong Wang <wygivan@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pull networking updates from David Miller:
"Most notable changes in here:
1) By far the biggest accomplishment, thanks to a large range of
contributors, is the addition of multi-send for transmit. This is
the result of discussions back in Chicago, and the hard work of
several individuals.
Now, when the ->ndo_start_xmit() method of a driver sees
skb->xmit_more as true, it can choose to defer the doorbell
telling the driver to start processing the new TX queue entires.
skb->xmit_more means that the generic networking is guaranteed to
call the driver immediately with another SKB to send.
There is logic added to the qdisc layer to dequeue multiple
packets at a time, and the handling mis-predicted offloads in
software is now done with no locks held.
Finally, pktgen is extended to have a "burst" parameter that can
be used to test a multi-send implementation.
Several drivers have xmit_more support: i40e, igb, ixgbe, mlx4,
virtio_net
Adding support is almost trivial, so export more drivers to
support this optimization soon.
I want to thank, in no particular or implied order, Jesper
Dangaard Brouer, Eric Dumazet, Alexander Duyck, Tom Herbert, Jamal
Hadi Salim, John Fastabend, Florian Westphal, Daniel Borkmann,
David Tat, Hannes Frederic Sowa, and Rusty Russell.
2) PTP and timestamping support in bnx2x, from Michal Kalderon.
3) Allow adjusting the rx_copybreak threshold for a driver via
ethtool, and add rx_copybreak support to enic driver. From
Govindarajulu Varadarajan.
4) Significant enhancements to the generic PHY layer and the bcm7xxx
driver in particular (EEE support, auto power down, etc.) from
Florian Fainelli.
5) Allow raw buffers to be used for flow dissection, allowing drivers
to determine the optimal "linear pull" size for devices that DMA
into pools of pages. The objective is to get exactly the
necessary amount of headers into the linear SKB area pre-pulled,
but no more. The new interface drivers use is eth_get_headlen().
From WANG Cong, with driver conversions (several had their own
by-hand duplicated implementations) by Alexander Duyck and Eric
Dumazet.
6) Support checksumming more smoothly and efficiently for
encapsulations, and add "foo over UDP" facility. From Tom
Herbert.
7) Add Broadcom SF2 switch driver to DSA layer, from Florian
Fainelli.
8) eBPF now can load programs via a system call and has an extensive
testsuite. Alexei Starovoitov and Daniel Borkmann.
9) Major overhaul of the packet scheduler to use RCU in several major
areas such as the classifiers and rate estimators. From John
Fastabend.
10) Add driver for Intel FM10000 Ethernet Switch, from Alexander
Duyck.
11) Rearrange TCP_SKB_CB() to reduce cache line misses, from Eric
Dumazet.
12) Add Datacenter TCP congestion control algorithm support, From
Florian Westphal.
13) Reorganize sk_buff so that __copy_skb_header() is significantly
faster. From Eric Dumazet"
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next: (1558 commits)
netlabel: directly return netlbl_unlabel_genl_init()
net: add netdev_txq_bql_{enqueue, complete}_prefetchw() helpers
net: description of dma_cookie cause make xmldocs warning
cxgb4: clean up a type issue
cxgb4: potential shift wrapping bug
i40e: skb->xmit_more support
net: fs_enet: Add NAPI TX
net: fs_enet: Remove non NAPI RX
r8169:add support for RTL8168EP
net_sched: copy exts->type in tcf_exts_change()
wimax: convert printk to pr_foo()
af_unix: remove 0 assignment on static
ipv6: Do not warn for informational ICMP messages, regardless of type.
Update Intel Ethernet Driver maintainers list
bridge: Save frag_max_size between PRE_ROUTING and POST_ROUTING
tipc: fix bug in multicast congestion handling
net: better IFF_XMIT_DST_RELEASE support
net/mlx4_en: remove NETDEV_TX_BUSY
3c59x: fix bad split of cpu_to_le32(pci_map_single())
net: bcmgenet: fix Tx ring priority programming
...
1/ Step down as dmaengine maintainer see commit 08223d80df "dmaengine
maintainer update"
2/ Removal of net_dma, as it has been marked 'broken' since 3.13 (commit
7787380336 "net_dma: mark broken"), without reports of performance
regression.
3/ Miscellaneous fixes
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1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=BBBv
-----END PGP SIGNATURE-----
Merge tag 'dmaengine-3.17' of git://git.kernel.org/pub/scm/linux/kernel/git/djbw/dmaengine
Pull dmaengine updates from Dan Williams:
"Even though this has fixes marked for -stable, given the size and the
needed conflict resolutions this is 3.18-rc1/merge-window material.
These patches have been languishing in my tree for a long while. The
fact that I do not have the time to do proper/prompt maintenance of
this tree is a primary factor in the decision to step down as
dmaengine maintainer. That and the fact that the bulk of drivers/dma/
activity is going through Vinod these days.
The net_dma removal has not been in -next. It has developed simple
conflicts against mainline and net-next (for-3.18).
Continuing thanks to Vinod for staying on top of drivers/dma/.
Summary:
1/ Step down as dmaengine maintainer see commit 08223d80df
"dmaengine maintainer update"
2/ Removal of net_dma, as it has been marked 'broken' since 3.13
(commit 7787380336 "net_dma: mark broken"), without reports of
performance regression.
3/ Miscellaneous fixes"
* tag 'dmaengine-3.17' of git://git.kernel.org/pub/scm/linux/kernel/git/djbw/dmaengine:
net: make tcp_cleanup_rbuf private
net_dma: revert 'copied_early'
net_dma: simple removal
dmaengine maintainer update
dmatest: prevent memory leakage on error path in thread
ioat: Use time_before_jiffies()
dmaengine: fix xor sources continuation
dma: mv_xor: Rename __mv_xor_slot_cleanup() to mv_xor_slot_cleanup()
dma: mv_xor: Remove all callers of mv_xor_slot_cleanup()
dma: mv_xor: Remove unneeded mv_xor_clean_completed_slots() call
ioat: Use pci_enable_msix_exact() instead of pci_enable_msix()
drivers: dma: Include appropriate header file in dca.c
drivers: dma: Mark functions as static in dma_v3.c
dma: mv_xor: Add DMA API error checks
ioat/dca: Use dev_is_pci() to check whether it is pci device
Per commit "77873803363c net_dma: mark broken" net_dma is no longer used
and there is no plan to fix it.
This is the mechanical removal of bits in CONFIG_NET_DMA ifdef guards.
Reverting the remainder of the net_dma induced changes is deferred to
subsequent patches.
Marked for stable due to Roman's report of a memory leak in
dma_pin_iovec_pages():
https://lkml.org/lkml/2014/9/3/177
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Vinod Koul <vinod.koul@intel.com>
Cc: David Whipple <whipple@securedatainnovations.ch>
Cc: Alexander Duyck <alexander.h.duyck@intel.com>
Cc: <stable@vger.kernel.org>
Reported-by: Roman Gushchin <klamm@yandex-team.ru>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Upon timeout, undo (via both timestamps/Eifel and DSACKs) was
disabled if any retransmits were still in flight. The concern was
perhaps that spurious retransmission sent in a previous recovery
episode may trigger DSACKs to falsely undo the current recovery.
However, this inadvertently misses undo opportunities (using either
TCP timestamps or DSACKs) when timeout occurs during a loss episode,
i.e. recurring timeouts or timeout during fast recovery. In these
cases some retransmissions will be in flight but we should allow
undo. Furthermore, we should only reset undo_marker and undo_retrans
upon timeout if we are starting a new recovery episode. Finally,
when we do reset our undo state, we now do so in a manner similar
to tcp_enter_recovery(), so that we require a DSACK for each of
the outstsanding retransmissions. This will achieve the original
goal by requiring that we receive the same number of DSACKs as
retransmissions.
This patch increases the undo events by 50% on Google servers.
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>
Move the specific IPv4/IPv6 intializations to a new method in
tcp_request_sock_ops in preparation for unifying tcp_v4_conn_request
and tcp_v6_conn_request.
Signed-off-by: Octavian Purdila <octavian.purdila@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Experience with the recent e114a710aa ("tcp: fix cwnd limited
checking to improve congestion control") has shown that there are
common cases where that commit can cause cwnd to be much larger than
necessary. This leads to TSO autosizing cooking skbs that are too
large, among other things.
The main problems seemed to be:
(1) That commit attempted to predict the future behavior of the
connection by looking at the write queue (if TSO or TSQ limit
sending). That prediction sometimes overestimated future outstanding
packets.
(2) That commit always allowed cwnd to grow to twice the number of
outstanding packets (even in congestion avoidance, where this is not
needed).
This commit improves both of these, by:
(1) Switching to a measurement-based approach where we explicitly
track the largest number of packets in flight during the past window
("max_packets_out"), and remember whether we were cwnd-limited at the
moment we finished sending that flight.
(2) Only allowing cwnd to grow to twice the number of outstanding
packets ("max_packets_out") in slow start. In congestion avoidance
mode we now only allow cwnd to grow if it was fully utilized.
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>
Consolidate various cookie checking and generation code to simplify
the fast open processing. The main goal is to reduce code duplication
in tcp_v4_conn_request() for IPv6 support.
Removes two experimental sysctl flags TFO_SERVER_ALWAYS and
TFO_SERVER_COOKIE_NOT_CHKD used primarily for developmental debugging
purposes.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Daniel Lee <longinus00@gmail.com>
Signed-off-by: Jerry Chu <hkchu@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>
Yuchung discovered tcp_is_cwnd_limited() was returning false in
slow start phase even if the application filled the socket write queue.
All congestion modules take into account tcp_is_cwnd_limited()
before increasing cwnd, so this behavior limits slow start from
probing the bandwidth at full speed.
The problem is that even if write queue is full (aka we are _not_
application limited), cwnd can be under utilized if TSO should auto
defer or TCP Small queues decided to hold packets.
So the in_flight can be kept to smaller value, and we can get to the
point tcp_is_cwnd_limited() returns false.
With TCP Small Queues and FQ/pacing, this issue is more visible.
We fix this by having tcp_cwnd_validate(), which is supposed to track
such things, take into account unsent_segs, the number of segs that we
are not sending at the moment due to TSO or TSQ, but intend to send
real soon. Then when we are cwnd-limited, remember this fact while we
are processing the window of ACKs that comes back.
For example, suppose we have a brand new connection with cwnd=10; we
are in slow start, and we send a flight of 9 packets. By the time we
have received ACKs for all 9 packets we want our cwnd to be 18.
We implement this by setting tp->lsnd_pending to 9, and
considering ourselves to be cwnd-limited while cwnd is less than
twice tp->lsnd_pending (2*9 -> 18).
This makes tcp_is_cwnd_limited() more understandable, by removing
the GSO/TSO kludge, that tried to work around the issue.
Note the in_flight parameter can be removed in a followup cleanup
patch.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Upcoming congestion controls for TCP require usec resolution for RTT
estimations. Millisecond resolution is simply not enough these days.
FQ/pacing in DC environments also require this change for finer control
and removal of bimodal behavior due to the current hack in
tcp_update_pacing_rate() for 'small rtt'
TCP_CONG_RTT_STAMP is no longer needed.
As Julian Anastasov pointed out, we need to keep user compatibility :
tcp_metrics used to export RTT and RTTVAR in msec resolution,
so we added RTT_US and RTTVAR_US. An iproute2 patch is needed
to use the new attributes if provided by the kernel.
In this example ss command displays a srtt of 32 usecs (10Gbit link)
lpk51:~# ./ss -i dst lpk52
Netid State Recv-Q Send-Q Local Address:Port Peer
Address:Port
tcp ESTAB 0 1 10.246.11.51:42959
10.246.11.52:64614
cubic wscale:6,6 rto:201 rtt:0.032/0.001 ato:40 mss:1448
cwnd:10 send
3620.0Mbps pacing_rate 7240.0Mbps unacked:1 rcv_rtt:993 rcv_space:29559
Updated iproute2 ip command displays :
lpk51:~# ./ip tcp_metrics | grep 10.246.11.52
10.246.11.52 age 561.914sec cwnd 10 rtt 274us rttvar 213us source
10.246.11.51
Old binary displays :
lpk51:~# ip tcp_metrics | grep 10.246.11.52
10.246.11.52 age 561.914sec cwnd 10 rtt 250us rttvar 125us source
10.246.11.51
With help from Julian Anastasov, Stephen Hemminger and Yuchung Cheng
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Cc: Stephen Hemminger <stephen@networkplumber.org>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Larry Brakmo <brakmo@google.com>
Cc: Julian Anastasov <ja@ssi.bg>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP out_of_order_queue lock is not used, as queue manipulation
happens with socket lock held and we therefore use the lockless
skb queue routines (as __skb_queue_head())
We can use __skb_queue_head_init() instead of skb_queue_head_init()
to make this more consistent.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove declaration, 4 defines and confusing comment that are no longer used
since 1a2c6181c4 ("tcp: Remove TCPCT").
Signed-off-by: Dmitry Popov <dp@highloadlab.com>
Acked-by: Christoph Paasch <christoph.paasch@uclouvain.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
Idea of this patch is to add optional limitation of number of
unsent bytes in TCP sockets, to reduce usage of kernel memory.
TCP receiver might announce a big window, and TCP sender autotuning
might allow a large amount of bytes in write queue, but this has little
performance impact if a large part of this buffering is wasted :
Write queue needs to be large only to deal with large BDP, not
necessarily to cope with scheduling delays (incoming ACKS make room
for the application to queue more bytes)
For most workloads, using a value of 128 KB or less is OK to give
applications enough time to react to POLLOUT events in time
(or being awaken in a blocking sendmsg())
This patch adds two ways to set the limit :
1) Per socket option TCP_NOTSENT_LOWAT
2) A sysctl (/proc/sys/net/ipv4/tcp_notsent_lowat) for sockets
not using TCP_NOTSENT_LOWAT socket option (or setting a zero value)
Default value being UINT_MAX (0xFFFFFFFF), meaning this has no effect.
This changes poll()/select()/epoll() to report POLLOUT
only if number of unsent bytes is below tp->nosent_lowat
Note this might increase number of sendmsg()/sendfile() calls
when using non blocking sockets,
and increase number of context switches for blocking sockets.
Note this is not related to SO_SNDLOWAT (as SO_SNDLOWAT is
defined as :
Specify the minimum number of bytes in the buffer until
the socket layer will pass the data to the protocol)
Tested:
netperf sessions, and watching /proc/net/protocols "memory" column for TCP
With 200 concurrent netperf -t TCP_STREAM sessions, amount of kernel memory
used by TCP buffers shrinks by ~55 % (20567 pages instead of 45458)
lpq83:~# echo -1 >/proc/sys/net/ipv4/tcp_notsent_lowat
lpq83:~# (super_netperf 200 -t TCP_STREAM -H remote -l 90 &); sleep 60 ; grep TCP /proc/net/protocols
TCPv6 1880 2 45458 no 208 yes ipv6 y y y y y y y y y y y y y n y y y y y
TCP 1696 508 45458 no 208 yes kernel y y y y y y y y y y y y y n y y y y y
lpq83:~# echo 131072 >/proc/sys/net/ipv4/tcp_notsent_lowat
lpq83:~# (super_netperf 200 -t TCP_STREAM -H remote -l 90 &); sleep 60 ; grep TCP /proc/net/protocols
TCPv6 1880 2 20567 no 208 yes ipv6 y y y y y y y y y y y y y n y y y y y
TCP 1696 508 20567 no 208 yes kernel y y y y y y y y y y y y y n y y y y y
Using 128KB has no bad effect on the throughput or cpu usage
of a single flow, although there is an increase of context switches.
A bonus is that we hold socket lock for a shorter amount
of time and should improve latencies of ACK processing.
lpq83:~# echo -1 >/proc/sys/net/ipv4/tcp_notsent_lowat
lpq83:~# perf stat -e context-switches ./netperf -H 7.7.7.84 -t omni -l 20 -c -i10,3
OMNI Send TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 7.7.7.84 () port 0 AF_INET : +/-2.500% @ 99% conf.
Local Remote Local Elapsed Throughput Throughput Local Local Remote Remote Local Remote Service
Send Socket Recv Socket Send Time Units CPU CPU CPU CPU Service Service Demand
Size Size Size (sec) Util Util Util Util Demand Demand Units
Final Final % Method % Method
1651584 6291456 16384 20.00 17447.90 10^6bits/s 3.13 S -1.00 U 0.353 -1.000 usec/KB
Performance counter stats for './netperf -H 7.7.7.84 -t omni -l 20 -c -i10,3':
412,514 context-switches
200.034645535 seconds time elapsed
lpq83:~# echo 131072 >/proc/sys/net/ipv4/tcp_notsent_lowat
lpq83:~# perf stat -e context-switches ./netperf -H 7.7.7.84 -t omni -l 20 -c -i10,3
OMNI Send TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 7.7.7.84 () port 0 AF_INET : +/-2.500% @ 99% conf.
Local Remote Local Elapsed Throughput Throughput Local Local Remote Remote Local Remote Service
Send Socket Recv Socket Send Time Units CPU CPU CPU CPU Service Service Demand
Size Size Size (sec) Util Util Util Util Demand Demand Units
Final Final % Method % Method
1593240 6291456 16384 20.00 17321.16 10^6bits/s 3.35 S -1.00 U 0.381 -1.000 usec/KB
Performance counter stats for './netperf -H 7.7.7.84 -t omni -l 20 -c -i10,3':
2,675,818 context-switches
200.029651391 seconds time elapsed
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-By: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_timeout_skb() was intended to trigger fast recovery on timeout,
unfortunately in reality it often causes spurious retransmission
storms during fast recovery. The particular sign is a fast retransmit
over the highest sacked sequence (SND.FACK).
Currently the RTO timer re-arming (as in RFC6298) offers a nice cushion
to avoid spurious timeout: when SND.UNA advances the sender re-arms
RTO and extends the timeout by icsk_rto. The sender does not offset
the time elapsed since the packet at SND.UNA was sent.
But if the next (DUP)ACK arrives later than ~RTTVAR and triggers
tcp_fastretrans_alert(), then tcp_timeout_skb() will mark any packet
sent before the icsk_rto interval lost, including one that's above the
highest sacked sequence. Most likely a large part of scorebard will be
marked.
If most packets are not lost then the subsequent DUPACKs with new SACK
blocks will cause the sender to continue to retransmit packets beyond
SND.FACK spuriously. Even if only one packet is lost the sender may
falsely retransmit almost the entire window.
The situation becomes common in the world of bufferbloat: the RTT
continues to grow as the queue builds up but RTTVAR remains small and
close to the minimum 200ms. If a data packet is lost and the DUPACK
triggered by the next data packet is slightly delayed, then a spurious
retransmission storm forms.
As the original comment on tcp_timeout_skb() suggests: the usefulness
of this feature is questionable. It also wastes cycles walking the
sack scoreboard and is actually harmful because of false recovery.
It's time to remove this.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch implements F-RTO (foward RTO recovery):
When the first retransmission after timeout is acknowledged, F-RTO
sends new data instead of old data. If the next ACK acknowledges
some never-retransmitted data, then the timeout was spurious and the
congestion state is reverted. Otherwise if the next ACK selectively
acknowledges the new data, then the timeout was genuine and the
loss recovery continues. This idea applies to recurring timeouts
as well. While F-RTO sends different data during timeout recovery,
it does not (and should not) change the congestion control.
The implementaion follows the three steps of SACK enhanced algorithm
(section 3) in RFC5682. Step 1 is in tcp_enter_loss(). Step 2 and
3 are in tcp_process_loss(). The basic version is not supported
because SACK enhanced version also works for non-SACK connections.
The new implementation is functionally in parity with the old F-RTO
implementation except the one case where it increases undo events:
In addition to the RFC algorithm, a spurious timeout may be detected
without sending data in step 2, as long as the SACK confirms not
all the original data are dropped. When this happens, the sender
will undo the cwnd and perhaps enter fast recovery instead. This
additional check increases the F-RTO undo events by 5x compared
to the prior implementation on Google Web servers, since the sender
often does not have new data to send for HTTP.
Note F-RTO may detect spurious timeout before Eifel with timestamps
does so.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The patch series refactor the F-RTO feature (RFC4138/5682).
This is to simplify the loss recovery processing. Existing F-RTO
was developed during the experimental stage (RFC4138) and has
many experimental features. It takes a separate code path from
the traditional timeout processing by overloading CA_Disorder
instead of using CA_Loss state. This complicates CA_Disorder state
handling because it's also used for handling dubious ACKs and undos.
While the algorithm in the RFC does not change the congestion control,
the implementation intercepts congestion control in various places
(e.g., frto_cwnd in tcp_ack()).
The new code implements newer F-RTO RFC5682 using CA_Loss processing
path. F-RTO becomes a small extension in the timeout processing
and interfaces with congestion control and Eifel undo modules.
It lets congestion control (module) determines how many to send
independently. F-RTO only chooses what to send in order to detect
spurious retranmission. If timeout is found spurious it invokes
existing Eifel undo algorithms like DSACK or TCP timestamp based
detection.
The first patch removes all F-RTO code except the sysctl_tcp_frto is
left for the new implementation. Since CA_EVENT_FRTO is removed, TCP
westwood now computes ssthresh on regular timeout CA_EVENT_LOSS event.
Signed-off-by: Yuchung Cheng <ycheng@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>
TCPCT uses option-number 253, reserved for experimental use and should
not be used in production environments.
Further, TCPCT does not fully implement RFC 6013.
As a nice side-effect, removing TCPCT increases TCP's performance for
very short flows:
Doing an apache-benchmark with -c 100 -n 100000, sending HTTP-requests
for files of 1KB size.
before this patch:
average (among 7 runs) of 20845.5 Requests/Second
after:
average (among 7 runs) of 21403.6 Requests/Second
Signed-off-by: Christoph Paasch <christoph.paasch@uclouvain.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is the second of the TLP patch series; it augments the basic TLP
algorithm with a loss detection scheme.
This patch implements a mechanism for loss detection when a Tail
loss probe retransmission plugs a hole thereby masking packet loss
from the sender. The loss detection algorithm relies on counting
TLP dupacks as outlined in Sec. 3 of:
http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01
The basic idea is: Sender keeps track of TLP "episode" upon
retransmission of a TLP packet. An episode ends when the sender receives
an ACK above the SND.NXT (tracked by tlp_high_seq) at the time of the
episode. We want to make sure that before the episode ends the sender
receives a "TLP dupack", indicating that the TLP retransmission was
unnecessary, so there was no loss/hole that needed plugging. If the
sender gets no TLP dupack before the end of the episode, then it reduces
ssthresh and the congestion window, because the TLP packet arriving at
the receiver probably plugged a hole.
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch series implement the Tail loss probe (TLP) algorithm described
in http://tools.ietf.org/html/draft-dukkipati-tcpm-tcp-loss-probe-01. The
first patch implements the basic algorithm.
TLP's goal is to reduce tail latency of short transactions. It achieves
this by converting retransmission timeouts (RTOs) occuring due
to tail losses (losses at end of transactions) into fast recovery.
TLP transmits one packet in two round-trips when a connection is in
Open state and isn't receiving any ACKs. The transmitted packet, aka
loss probe, can be either new or a retransmission. When there is tail
loss, the ACK from a loss probe triggers FACK/early-retransmit based
fast recovery, thus avoiding a costly RTO. In the absence of loss,
there is no change in the connection state.
PTO stands for probe timeout. It is a timer event indicating
that an ACK is overdue and triggers a loss probe packet. The PTO value
is set to max(2*SRTT, 10ms) and is adjusted to account for delayed
ACK timer when there is only one oustanding packet.
TLP Algorithm
On transmission of new data in Open state:
-> packets_out > 1: schedule PTO in max(2*SRTT, 10ms).
-> packets_out == 1: schedule PTO in max(2*RTT, 1.5*RTT + 200ms)
-> PTO = min(PTO, RTO)
Conditions for scheduling PTO:
-> Connection is in Open state.
-> Connection is either cwnd limited or no new data to send.
-> Number of probes per tail loss episode is limited to one.
-> Connection is SACK enabled.
When PTO fires:
new_segment_exists:
-> transmit new segment.
-> packets_out++. cwnd remains same.
no_new_packet:
-> retransmit the last segment.
Its ACK triggers FACK or early retransmit based recovery.
ACK path:
-> rearm RTO at start of ACK processing.
-> reschedule PTO if need be.
In addition, the patch includes a small variation to the Early Retransmit
(ER) algorithm, such that ER and TLP together can in principle recover any
N-degree of tail loss through fast recovery. TLP is controlled by the same
sysctl as ER, tcp_early_retrans sysctl.
tcp_early_retrans==0; disables TLP and ER.
==1; enables RFC5827 ER.
==2; delayed ER.
==3; TLP and delayed ER. [DEFAULT]
==4; TLP only.
The TLP patch series have been extensively tested on Google Web servers.
It is most effective for short Web trasactions, where it reduced RTOs by 15%
and improved HTTP response time (average by 6%, 99th percentile by 10%).
The transmitted probes account for <0.5% of the overall transmissions.
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>