This patch implements RFC 5827 early retransmit (ER) for TCP.
It reduces DUPACK threshold (dupthresh) if outstanding packets are
less than 4 to recover losses by fast recovery instead of timeout.
While the algorithm is simple, small but frequent network reordering
makes this feature dangerous: the connection repeatedly enter
false recovery and degrade performance. Therefore we implement
a mitigation suggested in the appendix of the RFC that delays
entering fast recovery by a small interval, i.e., RTT/4. Currently
ER is conservative and is disabled for the rest of the connection
after the first reordering event. A large scale web server
experiment on the performance impact of ER is summarized in
section 6 of the paper "Proportional Rate Reduction for TCP”,
IMC 2011. http://conferences.sigcomm.org/imc/2011/docs/p155.pdf
Note that Linux has a similar feature called THIN_DUPACK. The
differences are THIN_DUPACK do not mitigate reorderings and is only
used after slow start. Currently ER is disabled if THIN_DUPACK is
enabled. I would be happy to merge THIN_DUPACK feature with ER if
people think it's a good idea.
ER is enabled by sysctl_tcp_early_retrans:
0: Disables ER
1: Reduce dupthresh to packets_out - 1 when outstanding packets < 4.
2: (Default) reduce dupthresh like mode 1. In addition, delay
entering fast recovery by RTT/4.
Note: mode 2 is implemented in the third part of this patch series.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This a prepartion patch that refactors the code to enter recovery
into a new function tcp_enter_recovery(). It's needed to implement
the delayed fast retransmit in ER.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP coalesce can check if skb to be merged has its skb->head mapped to a
page fragment, instead of a kmalloc() area.
We had to disable coalescing in this case, for performance reasons.
We 'upgrade' skb->head as a fragment in itself.
This reduces number of cache misses when user makes its copies, since a
less sk_buff are fetched.
This makes receive and ofo queues shorter and thus reduce cache line
misses in TCP stack.
This is a followup of patch "net: allow skb->head to be a page fragment"
Tested with tg3 nic, with GRO on or off. We can see "TCPRcvCoalesce"
counter being incremented.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Matt Carlson <mcarlson@broadcom.com>
Cc: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This clarifies code intention, as suggested by David.
Suggested-by: David Miller <davem@davemloft.net>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix merge between commit 3adadc08cc ("net ax25: Reorder ax25_exit to
remove races") and commit 0ca7a4c87d ("net ax25: Simplify and
cleanup the ax25 sysctl handling")
The former moved around the sysctl register/unregister calls, the
later simply removed them.
With help from Stephen Rothwell.
Signed-off-by: David S. Miller <davem@davemloft.net>
commit c8628155ec (tcp: reduce out_of_order memory use) took care of
coalescing tcp segments provided by legacy devices (linear skbs)
We extend this idea to fragged skbs, as their truesize can be heavy.
ixgbe for example uses 256+1024+PAGE_SIZE/2 = 3328 bytes per segment.
Use this coalescing strategy for receive queue too.
This contributes to reduce number of tcp collapses, at minimal cost, and
reduces memory overhead and packets drops.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is just the preparation patch, which makes the needed for
TCP repair code ready for use.
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_grow_window() has to grow rcv_ssthresh up to window_clamp, allowing
sender to increase its window.
tcp_grow_window() still assumes a tcp frame is under MSS, but its no
longer true with LRO/GRO.
This patch fixes one of the performance issue we noticed with GRO on.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use of "unsigned int" is preferred to bare "unsigned" in net tree.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_enter_quickack_mode() already calls tcp_incr_quickack() and sets
icsk->icsk_ack.ato to TCP_ATO_MIN. This patch removes the duplication.
Signed-off-by: Vijay Subramanian <subramanian.vijay@gmail.com>
Reviewed-by: Flavio Leitner <fbl@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Updates some comments to track RFC6298
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: H.K. Jerry Chu <hkchu@google.com>
Cc: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pull networking fixes from David Miller:
1) Fix bluetooth userland regression reported by Keith Packard, from
Gustavo Padovan.
2) Revert ath9k PS idle change, from Sujith Manoharan.
3) Correct default TCP memory limits (again), from Eric Dumazet.
4) Fix tcp_rcv_rtt_update() accidental use of unscaled RTT, from Neal
Cardwell.
5) We made a facility for layers like wireless to say how much tailroom
they need in the SKB for link layer stuff such as wireless
encryption etc., but TCP works hard to fill every SKB out to the end
defeating this specification.
This leads to every TCP packet getting reallocated by the wireless
code in order to have the right amount of tailroom available.
Fix TCP to only fill SKBs out to the real amount of data area it
asked for during the allocation, this way it won't eat into the
slack added for the device's tailroom needs.
Reported by Marc Merlin and fixed by Eric Dumazet.
6) Leaks, endian bugs, and new device IDs in bluetooth from Santosh
Nayak, João Paulo Rechi Vita, Cho, Yu-Chen, Andrei Emeltchenko,
AceLan Kao, and Andrei Emeltchenko.
7) OOPS on tty_close fix in bluetooth's hci_ldisc from Johan Hovold.
8) netfilter erroneously scales TCP window twice, fix from Changli Gao.
9) Memleak fix in wext-core from Julia Lawall.
10) Consistently handle invalid TCP packets in ipv4 vs. ipv6 conntrack,
from Jozsef Kadlecsik.
11) Validate IP header length properly in netfilter conntrack's
ipv4_get_l4proto().
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net: (39 commits)
NFC: Fix the LLCP Tx fragmentation loop
rtlwifi: Add missing DMA buffer unmapping for PCI drivers
rtlwifi: Preallocate USB read buffers and eliminate kalloc in read routine
tcp: avoid order-1 allocations on wifi and tx path
net: allow pskb_expand_head() to get maximum tailroom
bridge: Do not send queries on multicast group leaves
MAINTAINERS: Mark NATSEMI driver as orphan'd.
tcp: fix tcp_rcv_rtt_update() use of an unscaled RTT sample
tcp: restore correct limit
Revert "ath9k: fix going to full-sleep on PS idle"
rt2x00: Fix rfkill_polling register function.
bcma: fix build error on MIPS; implicit pcibios_enable_device
netfilter: nf_conntrack: fix incorrect logic in nf_conntrack_init_net
netfilter: nf_ct_ipv4: packets with wrong ihl are invalid
netfilter: nf_ct_ipv4: handle invalid IPv4 and IPv6 packets consistently
net/wireless/wext-core.c: add missing kfree
rtlwifi: Fix oops on rate-control failure
mac80211: Convert WARN_ON to WARN_ON_ONCE
rtlwifi: rtl8192de: Fix firmware initialization
nl80211: ensure interface is up in various APIs
...
1/ regression fix for Xen as it now trips over a broken assumption
about the dma address size on 32-bit builds
2/ new quirk for netdma to ignore dma channels that cannot meet
netdma alignment requirements
3/ fixes for two long standing issues in ioatdma (ring size overflow)
and iop-adma (potential stack corruption)
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.12 (GNU/Linux)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=F3IR
-----END PGP SIGNATURE-----
Merge tag 'dmaengine-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/djbw/dmaengine
Pull dmaengine fixes from Dan Williams:
1/ regression fix for Xen as it now trips over a broken assumption
about the dma address size on 32-bit builds
2/ new quirk for netdma to ignore dma channels that cannot meet
netdma alignment requirements
3/ fixes for two long standing issues in ioatdma (ring size overflow)
and iop-adma (potential stack corruption)
* tag 'dmaengine-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/djbw/dmaengine:
netdma: adding alignment check for NETDMA ops
ioatdma: DMA copy alignment needed to address IOAT DMA silicon errata
ioat: ring size variables need to be 32bit to avoid overflow
iop-adma: Corrected array overflow in RAID6 Xscale(R) test.
ioat: fix size of 'completion' for Xen
Fix a code path in tcp_rcv_rtt_update() that was comparing scaled and
unscaled RTT samples.
The intent in the code was to only use the 'm' measurement if it was a
new minimum. However, since 'm' had not yet been shifted left 3 bits
but 'new_sample' had, this comparison would nearly always succeed,
leading us to erroneously set our receive-side RTT estimate to the 'm'
sample when that sample could be nearly 8x too high to use.
The overall effect is to often cause the receive-side RTT estimate to
be significantly too large (up to 40% too large for brief periods in
my tests).
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is the fallout from adding memcpy alignment workaround for certain
IOATDMA hardware. NetDMA will only use DMA engine that can handle byte align
ops.
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
With increasing receive window sizes, but speed of light not improved
that much, out of order queue can contain a huge number of skbs, waiting
to be moved to receive_queue when missing packets can fill the holes.
Some devices happen to use fat skbs (truesize of 4096 + sizeof(struct
sk_buff)) to store regular (MTU <= 1500) frames. This makes highly
probable sk_rmem_alloc hits sk_rcvbuf limit, which can be 4Mbytes in
many cases.
When limit is hit, tcp stack calls tcp_collapse_ofo_queue(), a true
latency killer and cpu cache blower.
Doing the coalescing attempt each time we add a frame in ofo queue
permits to keep memory use tight and in many cases avoid the
tcp_collapse() thing later.
Tested on various wireless setups (b43, ath9k, ...) known to use big skb
truesize, this patch removed the "packets collapsed in receive queue due
to low socket buffer" I had before.
This also reduced average memory used by tcp sockets.
With help from Neal Cardwell.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: H.K. Jerry Chu <hkchu@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Split tcp_data_queue() in two parts for better readability.
tcp_data_queue_ofo() is responsible for queueing incoming skb into out
of order queue.
Change code layout so that the skb_set_owner_r() is performed only if
skb is not dropped.
This is a preliminary patch before "reduce out_of_order memory use"
following patch.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: H.K. Jerry Chu <hkchu@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add #define pr_fmt(fmt) as appropriate.
Add "IPv4: ", "TCP: ", and "IPsec: " to appropriate files.
Standardize on "UDPLite: " for appropriate uses.
Some prefixes were previously "UDPLITE: " and "UDP-Lite: ".
Add KBUILD_MODNAME ": " to icmp and gre.
Remove embedded prefixes as appropriate.
Add missing "\n" to pr_info in gre.c.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use a more current kernel messaging style.
Convert a printk block to print_hex_dump.
Coalesce formats, align arguments.
Use %s, __func__ instead of embedding function names.
Some messages that were prefixed with <foo>_close are
now prefixed with <foo>_fini. Some ah4 and esp messages
are now not prefixed with "ip ".
The intent of this patch is to later add something like
#define pr_fmt(fmt) "IPv4: " fmt.
to standardize the output messages.
Text size is trivially reduced. (x86-32 allyesconfig)
$ size net/ipv4/built-in.o*
text data bss dec hex filename
887888 31558 249696 1169142 11d6f6 net/ipv4/built-in.o.new
887934 31558 249800 1169292 11d78c net/ipv4/built-in.o.old
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit fixes tcp_shift_skb_data() so that it does not shift
SACKed data below snd_una.
This fixes an issue whose symptoms exactly match reports showing
tp->sacked_out going negative since 3.3.0-rc4 (see "WARNING: at
net/ipv4/tcp_input.c:3418" thread on netdev).
Since 2008 (832d11c5cd)
tcp_shift_skb_data() had been shifting SACKed ranges that were below
snd_una. It checked that the *end* of the skb it was about to shift
from was above snd_una, but did not check that the end of the actual
shifted range was above snd_una; this commit adds that check.
Shifting SACKed ranges below snd_una is problematic because for such
ranges tcp_sacktag_one() short-circuits: it does not declare anything
as SACKed and does not increase sacked_out.
Before the fixes in commits cc9a672ee5
and daef52bab1, shifting SACKed ranges
below snd_una happened to work because tcp_shifted_skb() was always
(incorrectly) passing in to tcp_sacktag_one() an skb whose end_seq
tcp_shift_skb_data() had already guaranteed was beyond snd_una. Hence
tcp_sacktag_one() never short-circuited and always increased
tp->sacked_out in this case.
After those two fixes, my testing has verified that shifting SACKed
ranges below snd_una could cause tp->sacked_out to go negative with
the following sequence of events:
(1) tcp_shift_skb_data() sees an skb whose end_seq is beyond snd_una,
then shifts a prefix of that skb that is below snd_una
(2) tcp_shifted_skb() increments the packet count of the
already-SACKed prev sk_buff
(3) tcp_sacktag_one() sees the end of the new SACKed range is below
snd_una, so it short-circuits and doesn't increase tp->sacked_out
(5) tcp_clean_rtx_queue() sees the SACKed skb has been ACKed,
decrements tp->sacked_out by this "inflated" pcount that was
missing a matching increase in tp->sacked_out, and hence
tp->sacked_out underflows to a u32 like 0xFFFFFFFF, which casted
to s32 is negative.
(6) this leads to the warnings seen in the recent "WARNING: at
net/ipv4/tcp_input.c:3418" thread on the netdev list; e.g.:
tcp_input.c:3418 WARN_ON((int)tp->sacked_out < 0);
More generally, I think this bug can be tickled in some cases where
two or more ACKs from the receiver are lost and then a DSACK arrives
that is immediately above an existing SACKed skb in the write queue.
This fix changes tcp_shift_skb_data() to abort this sequence at step
(1) in the scenario above by noticing that the bytes are below snd_una
and not shifting them.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In tcp_mark_head_lost() we should not attempt to fragment a SACKed skb
to mark the first portion as lost. This is for two primary reasons:
(1) tcp_shifted_skb() coalesces adjacent regions of SACKed skbs. When
doing this, it preserves the sum of their packet counts in order to
reflect the real-world dynamics on the wire. But given that skbs can
have remainders that do not align to MSS boundaries, this packet count
preservation means that for SACKed skbs there is not necessarily a
direct linear relationship between tcp_skb_pcount(skb) and
skb->len. Thus tcp_mark_head_lost()'s previous attempts to fragment
off and mark as lost a prefix of length (packets - oldcnt)*mss from
SACKed skbs were leading to occasional failures of the WARN_ON(len >
skb->len) in tcp_fragment() (which used to be a BUG_ON(); see the
recent "crash in tcp_fragment" thread on netdev).
(2) there is no real point in fragmenting off part of a SACKed skb and
calling tcp_skb_mark_lost() on it, since tcp_skb_mark_lost() is a NOP
for SACKed skbs.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When tcp_shifted_skb() shifts bytes from the skb that is currently
pointed to by 'highest_sack' then the increment of
TCP_SKB_CB(skb)->seq implicitly advances tcp_highest_sack_seq(). This
implicit advancement, combined with the recent fix to pass the correct
SACKed range into tcp_sacktag_one(), caused tcp_sacktag_one() to think
that the newly SACKed range was before the tcp_highest_sack_seq(),
leading to a call to tcp_update_reordering() with a degree of
reordering matching the size of the newly SACKed range (typically just
1 packet, which is a NOP, but potentially larger).
This commit fixes this by simply calling tcp_sacktag_one() before the
TCP_SKB_CB(skb)->seq advancement that can advance our notion of the
highest SACKed sequence.
Correspondingly, we can simplify the code a little now that
tcp_shifted_skb() should update the lost_cnt_hint in all cases where
skb == tp->lost_skb_hint.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit ensures that lost_cnt_hint is correctly updated in
tcp_shifted_skb() for FACK TCP senders. The lost_cnt_hint adjustment
in tcp_sacktag_one() only applies to non-FACK senders, so FACK senders
need their own adjustment.
This applies the spirit of 1e5289e121 -
except now that the sequence range passed into tcp_sacktag_one() is
correct we need only have a special case adjustment for FACK.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix the newly-SACKed range to be the range of newly-shifted bytes.
Previously - since 832d11c5cd -
tcp_shifted_skb() incorrectly called tcp_sacktag_one() with the start
and end sequence numbers of the skb it passes in set to the range just
beyond the range that is newly-SACKed.
This commit also removes a special-case adjustment to lost_cnt_hint in
tcp_shifted_skb() since the pre-existing adjustment of lost_cnt_hint
in tcp_sacktag_one() now properly handles this things now that the
correct start sequence number is passed in.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit allows callers of tcp_sacktag_one() to pass in sequence
ranges that do not align with skb boundaries, as tcp_shifted_skb()
needs to do in an upcoming fix in this patch series.
In fact, now tcp_sacktag_one() does not need to depend on an input skb
at all, which makes its semantics and dependencies more clear.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Correctly implement a loss detection heuristic: New sequences (above
high_seq) sent during the fast recovery are deemed lost when higher
sequences are SACKed.
Current code does not catch these losses, because tcp_mark_head_lost()
does not check packets beyond high_seq. The fix is straight-forward by
checking packets until the highest sacked packet. In addition, all the
FLAG_DATA_LOST logic are in-effective and redundant and can be removed.
Update the loss heuristic comments. The algorithm above is documented
as heuristic B, but it is redundant too because heuristic A already
covers B.
Note that this change only marks some forward-retransmitted packets LOST.
It does NOT forbid TCP performing further CWR on new losses. A potential
follow-up patch under preparation is to perform another CWR on "new"
losses such as
1) sequence above high_seq is lost (by resetting high_seq to snd_nxt)
2) retransmission is lost.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
to record the state of SACK/FACK and DSACK for better readability and maintenance.
Signed-off-by: Vijay Subramanian <subramanian.vijay@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch replaces all uses of struct sock fields' memory_pressure,
memory_allocated, sockets_allocated, and sysctl_mem to acessor
macros. Those macros can either receive a socket argument, or a mem_cgroup
argument, depending on the context they live in.
Since we're only doing a macro wrapping here, no performance impact at all is
expected in the case where we don't have cgroups disabled.
Signed-off-by: Glauber Costa <glommer@parallels.com>
Reviewed-by: Hiroyouki Kamezawa <kamezawa.hiroyu@jp.fujitsu.com>
CC: David S. Miller <davem@davemloft.net>
CC: Eric W. Biederman <ebiederm@xmission.com>
CC: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Instead of testing defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Denys Fedoryshchenko reported that SYN+FIN attacks were bringing his
linux machines to their limits.
Dont call conn_request() if the TCP flags includes SYN flag
Reported-by: Denys Fedoryshchenko <denys@visp.net.lb>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The problem: Senders were overriding cwnd values picked during an undo
by calling tcp_moderate_cwnd() in tcp_try_to_open().
The fix: Don't moderate cwnd in tcp_try_to_open() if we're in
TCP_CA_Open, since doing so is generally unnecessary and specifically
would override a DSACK-based undo of a cwnd reduction made in fast
recovery.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously, SACK-enabled connections hung around in TCP_CA_Disorder
state while snd_una==high_seq, just waiting to accumulate DSACKs and
hopefully undo a cwnd reduction. This could and did lead to the
following unfortunate scenario: if some incoming ACKs advance snd_una
beyond high_seq then we were setting undo_marker to 0 and moving to
TCP_CA_Open, so if (due to reordering in the ACK return path) we
shortly thereafter received a DSACK then we were no longer able to
undo the cwnd reduction.
The change: Simplify the congestion avoidance state machine by
removing the behavior where SACK-enabled connections hung around in
the TCP_CA_Disorder state just waiting for DSACKs. Instead, when
snd_una advances to high_seq or beyond we typically move to
TCP_CA_Open immediately and allow an undo in either TCP_CA_Open or
TCP_CA_Disorder if we later receive enough DSACKs.
Other patches in this series will provide other changes that are
necessary to fully fix this problem.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The bug: When the ACK field is below snd_una (which can happen when
ACKs are reordered), senders ignored DSACKs (preventing undo) and did
not call tcp_fastretrans_alert, so they did not increment
prr_delivered to reflect newly-SACKed sequence ranges, and did not
call tcp_xmit_retransmit_queue, thus passing up chances to send out
more retransmitted and new packets based on any newly-SACKed packets.
The change: When the ACK field is below snd_una (the "old_ack" goto
label), call tcp_fastretrans_alert to allow undo based on any
newly-arrived DSACKs and try to send out more packets based on
newly-SACKed packets.
Other patches in this series will provide other changes that are
necessary to fully fix this problem.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
The bug: Senders ignored DSACKs after recovery when there were no
outstanding packets (a common scenario for HTTP servers).
The change: when there are no outstanding packets (the "no_queue" goto
label), call tcp_fastretrans_alert() in order to use DSACKs to undo
congestion window reductions.
Other patches in this series will provide other changes that are
necessary to fully fix this problem.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Allow callers to decide whether an ACK is a duplicate ACK. This is a
prerequisite to allowing fastretrans_alert to be called from new
contexts, such as the no_queue and old_ack code paths, from which we
have extra info that tells us whether an ACK is a dupack.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Adding const qualifiers to pointers can ease code review, and spot some
bugs. It might allow compiler to optimize code further.
For example, is it legal to temporary write a null cksum into tcphdr
in tcp_md5_hash_header() ? I am afraid a sniffer could catch the
temporary null value...
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_fin() only needs socket pointer, we can remove skb and th params.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since commit 356f039822 (TCP: increase default initial receive
window.), we allow sender to send 10 (TCP_DEFAULT_INIT_RCVWND) segments.
Change tcp_fixup_rcvbuf() to reflect this change, even if no real change
is expected, since sysctl_tcp_rmem[1] = 87380 and this value
is bigger than tcp_fixup_rcvbuf() computed rcvmem (~23720)
Note: Since commit 356f039822 limited default window to maximum of
10*1460 and 2*MSS, we use same heuristic in this patch.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Initial cwnd being 10 (TCP_INIT_CWND) instead of 3, change
tcp_fixup_sndbuf() to get more than 16384 bytes (sysctl_tcp_wmem[1]) in
initial sk_sndbuf
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
skb truesize currently accounts for sk_buff struct and part of skb head.
kmalloc() roundings are also ignored.
Considering that skb_shared_info is larger than sk_buff, its time to
take it into account for better memory accounting.
This patch introduces SKB_TRUESIZE(X) macro to centralize various
assumptions into a single place.
At skb alloc phase, we put skb_shared_info struct at the exact end of
skb head, to allow a better use of memory (lowering number of
reallocations), since kmalloc() gives us power-of-two memory blocks.
Unless SLUB/SLUB debug is active, both skb->head and skb_shared_info are
aligned to cache lines, as before.
Note: This patch might trigger performance regressions because of
misconfigured protocol stacks, hitting per socket or global memory
limits that were previously not reached. But its a necessary step for a
more accurate memory accounting.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Andi Kleen <ak@linux.intel.com>
CC: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
lost_skb_hint is used by tcp_mark_head_lost() to mark the first unhandled skb.
lost_cnt_hint is the number of packets or sacked packets before the lost_skb_hint;
When shifting a skb that is before the lost_skb_hint, if tcp_is_fack() is ture,
the skb has already been counted in the lost_cnt_hint; if tcp_is_fack() is false,
tcp_sacktag_one() will increase the lost_cnt_hint. So tcp_shifted_skb() does not
need to adjust the lost_cnt_hint by itself. When shifting a skb that is equal to
lost_skb_hint, the shifted packets will not be counted by tcp_mark_head_lost().
So tcp_shifted_skb() should adjust the lost_cnt_hint even tcp_is_fack(tp) is true.
Signed-off-by: Zheng Yan <zheng.z.yan@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Rename struct tcp_skb_cb "flags" to "tcp_flags" to ease code review and
maintenance.
Its content is a combination of FIN/SYN/RST/PSH/ACK/URG/ECE/CWR flags
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
struct tcp_skb_cb contains a "flags" field containing either tcp flags
or IP dsfield depending on context (input or output path)
Introduce ip_dsfield to make the difference clear and ease maintenance.
If later we want to save space, we can union flags/ip_dsfield
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While playing with a new ADSL box at home, I discovered that ECN
blackhole can trigger suboptimal quickack mode on linux : We send one
ACK for each incoming data frame, without any delay and eventual
piggyback.
This is because TCP_ECN_check_ce() considers that if no ECT is seen on a
segment, this is because this segment was a retransmit.
Refine this heuristic and apply it only if we seen ECT in a previous
segment, to detect ECN blackhole at IP level.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Jamal Hadi Salim <jhs@mojatatu.com>
CC: Jerry Chu <hkchu@google.com>
CC: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
CC: Jim Gettys <jg@freedesktop.org>
CC: Dave Taht <dave.taht@gmail.com>
Acked-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
D-SACK is allowed to reside below snd_una. But the corresponding check
in tcp_is_sackblock_valid() is the exact opposite. It looks like a typo.
Signed-off-by: Zheng Yan <zheng.z.yan@intel.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch implements Proportional Rate Reduction (PRR) for TCP.
PRR is an algorithm that determines TCP's sending rate in fast
recovery. PRR avoids excessive window reductions and aims for
the actual congestion window size at the end of recovery to be as
close as possible to the window determined by the congestion control
algorithm. PRR also improves accuracy of the amount of data sent
during loss recovery.
The patch implements the recommended flavor of PRR called PRR-SSRB
(Proportional rate reduction with slow start reduction bound) and
replaces the existing rate halving algorithm. PRR improves upon the
existing Linux fast recovery under a number of conditions including:
1) burst losses where the losses implicitly reduce the amount of
outstanding data (pipe) below the ssthresh value selected by the
congestion control algorithm and,
2) losses near the end of short flows where application runs out of
data to send.
As an example, with the existing rate halving implementation a single
loss event can cause a connection carrying short Web transactions to
go into the slow start mode after the recovery. This is because during
recovery Linux pulls the congestion window down to packets_in_flight+1
on every ACK. A short Web response often runs out of new data to send
and its pipe reduces to zero by the end of recovery when all its packets
are drained from the network. Subsequent HTTP responses using the same
connection will have to slow start to raise cwnd to ssthresh. PRR on
the other hand aims for the cwnd to be as close as possible to ssthresh
by the end of recovery.
A description of PRR and a discussion of its performance can be found at
the following links:
- IETF Draft:
http://tools.ietf.org/html/draft-mathis-tcpm-proportional-rate-reduction-01
- IETF Slides:
http://www.ietf.org/proceedings/80/slides/tcpm-6.pdfhttp://tools.ietf.org/agenda/81/slides/tcpm-2.pdf
- Paper to appear in Internet Measurements Conference (IMC) 2011:
Improving TCP Loss Recovery
Nandita Dukkipati, Matt Mathis, Yuchung Cheng
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch lowers the default initRTO from 3secs to 1sec per
RFC2988bis. It falls back to 3secs if the SYN or SYN-ACK packet
has been retransmitted, AND the TCP timestamp option is not on.
It also adds support to take RTT sample during 3WHS on the passive
open side, just like its active open counterpart, and uses it, if
valid, to seed the initRTO for the data transmission phase.
The patch also resets ssthresh to its initial default at the
beginning of the data transmission phase, and reduces cwnd to 1 if
there has been MORE THAN ONE retransmission during 3WHS per RFC5681.
Signed-off-by: H.K. Jerry Chu <hkchu@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
