Keep up with a small amount of churn in the upstream repository.
Upstream now represents the version of CAKE as found in the linux kernel
from 4.19 onwards but with some compatibility stubs to allow building on
<4.19.
After a diversion related to an experimental ECN implementation which
has now been reverted, the important and relevant changes for us are:
8a8946b sch_cake: avoid possible divide by zero in cake_enqueue()
183b320 RFC 8622 diffserv3, 4 & 8 LE PHB support
6ff4561 sch_cake: Make sure we can write the IP header before changing DSCP bits
9fba602 sch_cake: Use tc_skb_protocol for getting packet protocol
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
CAKE made it to kernel 4.19 and since OpenWrt now at kernel 4.19 we can
drop the out of tree cake package in base repository.
Add kmod-sched-cake to netsupport so package dependencies are still met.
Similarly CAKE is retained as an optional qdisc module to avoid base
scheduler package size implications.
Backport upstream patches from k5.1 to address some small bugs and
support fwmark usage.
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
No target is using kernel 3.18 anymore, remove all the generic
support for kernel 3.18.
The removed packages are depending on kernel 3.18 only and are not used on
any recent kernel.
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
Update the out of tree build of cake. Applicable patches are also in
net-next.
057c738 Fix fwmark_shft assignment (again)
ca6c162 Add support for storing mark back into conntrack
7ed9b6c Fix off-by-one error when setting fwmark_shft.
a4a243a sch_cake: Interpret fwmark parameter as a bitmask
29d707e Simplify logic in cake_select_tin()
8acaaee Permit use of connmarks as tin classifiers
348f186 Make the dual modes fairer
99a7297 compat: Don't lock root qdisc when dumping stats on old kernels
Signed-off-by: Ryan Mounce <ryan@mounce.com.au>
331ac70 Correctly update parent qlen when splitting GSO packets
581967c Makefile: Hook into Kbuild/Kconfig infrastructure
The parent qlen change is relevant if using cake as a leaf qdisc,
the makefile is a no-op.
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
Revert "Add workaround for wrong skb->mac_len values after splitting GSO"
Remove our local patch which did the same thing.
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
Kernels 4.14.73 & 4.9.140 include the gso fixup fix, so cake
doesn't need to do it. Let's not waste cpu cycles by doing it in
cake which could be really important on cpu constrained devices.
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
Bump to latest upstream cake:
Add workaround for wrong skb->mac_len values after splitting GSO
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
Expand filter flow mapping to include hosts as well
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
(cherry picked from commit d14ffdc307d36bd9abe908b46ff7baece54c9551)
Follow upstream kernel patch that restores always splitting gso packets
by default whilst making the option configurable from (tc) userspace.
No ABI/API change
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
Bump to the latest cake recipe.
This backports tc class support to kernel 4.9 and other than conditional
kernel compilation pre-processor macros represents the cake that has
gone upstream into kernel 4.19. Loud cheer!
Fun may be had by changing cake tin classification for packets on
ingress. e.g.
tc filter add dev ifb4eth0 parent 800b: protocol ip u32 match \
ip dport 6981 0xffff action skbedit priority 800b:1
Where 800b: represents the filter handle for the ifb obtained by 'tc
qdisc' and the 1 from 800b:1 represents the cake tin number. So the
above example puts all incoming packets destined for port 6981 into the
BULK (lowest priority) tin.
f39ab9a Obey tin_order for tc filter classifiers
1e2473f Clean up after latest backport.
82531d0 Reorder includes to fix out of tree compilation
52cbc00 Code style cleanup
6cdb496 Fix argument order for NL_SET_ERR_MSG_ATTR()
cab17b6 Remove duplicate call to qdisc_watchdog_init()
71c7991 Merge branch 'backport-classful'
32aa7fb Fix compilation on Linux 4.9
9f8fe7a Fix compilation on Linux 4.14
ceab7a3 Rework filter classification
aad5436 Fixed version of class stats
be1c549 Add cake-specific class stats
483399d Use tin_order for class dumps
80dc129 Add class dumping
0c8e6c1 Fix dropping when using filters
c220493 Add the minimum class ops
5ed54d2 Start implementing tc filter/class support
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
Fixes a potential infinite loop bug when in unlimited (ie not using
built in shaper) mode.
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
This bumps to the latest & possibly greatest cake, sadly it's still
inedible but from an SQM point of view quite tasty :-)
Main tweaks since previous bump, improved ack_filter, some extra stats,
support for 64bit netlink parameters (higher rates/byte counters)
0520a6c Fix NAT option handling
8da93e1 Make sure we always call qdisc_watchdog_init() in cake_init()
f65daf6 Fix mismatched parenthesis
51d4ab3 Change flag handling to be safe even when mixing with non-eligible ACKs
f2ea091 ack_filter: protect DCTCP with stricter filtering of ECE marks
28b4560 ACK filter: Handle wrapping sequence numbers and DSACKs
73f62d9 Use the right PAD attribute for options
5969c14 Use 32 for tin backlog
e289f31 Move all the u64 netlink attributes together
36180a0 Check ACK seqno before parsing SACKs
91bbc01 Merge branch 'mine' into cobalt
58c55ec Rework SACK check to compare the ranges of two SACKs
9a5d593 ack_filter: Add proper handling of SACKs
eca95d4 ack_filter: short-circuit TCP flag check
d50a246 compat: backport some ktime functions
7b7ad11 compat: define tcpopt_fastopen for pre-4.1 kernels
ca54cdb Fix ktime compare
9d7dcc0 ack filter: Parse TCP options and only drop safe ones
b119882 Return EOPNOTSUPP on NAT option if conntrack is not available
842d7f0 Don't try to pad stats with tin_stats padding
bd46dc2 Use 64-bit divide helper
8e41bf0 Make sure we never drop SACKs when filtering ACKs
66e5d60 Avoid comparing ktime_t to scalar values
7fab017 Actually commit the ktime_t changes
fca6d13 Switch to ktime_t and get rid of cobalt.h
6f7e5af Can't use do_div with 64-bit divisors
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
Following changes as part of the kernel
upstreaming attempts. And fix a slight fsck up
when calculating overheads for GSO packets.
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
No functional change. Code tidy ups.
735eaf2 Make sure we don't reallocate q->tins (we didn't anyway but his
really makes sure)
6c5ad6e Get rid of __GFP_NOWARN flag for memory allocation
2a37333 Don't need the wrapper for kvfree, and no need to check before calling it
2b1c631 Whitespace fix
7fe6e28 compat tidyup (for older kernel versions <4.4)
93b805c pedant tidy up superfluous semicolons on switch statements
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
The sender domain has a DMARC Reject/Quarantine policy which disallows
sending mailing list messages using the original "From" header.
To mitigate this problem, the original message has been wrapped
automatically by the mailing list software.
Cake is bearing fruits of kernel upstreaming efforts.
diffserv-llt dropped. DSCP mapping paper died and no one using it.
ack-filter re-written & simplified
tc userspace & cake kmod netlink interface usage changed in non
backwards compatible way, thus this once requires tc & cake to be
in-step. Change due to upstream requirements.
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
Relevant changes:
0afc1be Fixes for kernel 4.16
d2d6780 Reinitialise overhead compensation stats when reconfiguring.
a3bab9d Export overhead compensation stats to userspace.
9cd2fa8 Split tin stats to its own structure to decrease size of tc_cake_xstats
71c7b44 Gather more statistics about packet length transformations.
0517357 Rework overhead compensation to use dynamic transport header offset instead of (inaccurate) static one.
c1a0c8e Refactor length handling code to better centralise overhead calculations
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
More important bug fix:
402f05c Use full-rate mtu_time in all tins. Fixes an issue where some
cake tins experienced excessive latency since 49776da (dynamically
adjust target)
Minor bug fixes:
31277c2 Avoid unsigned comparison against zero. Fix compiler warning,
no known impact.
8cf5278 ack_filter: fix TCP flag check. A very contrived case may have
lead to dropping a SYN packet that should not be dropped.
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
Fix overhead accounting error introduced by f33c4d6 refactor
cake_advance_shaper and ack_filter
Symptoms were links running under rate.
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
There has been recent significant activity with the cake qdisc of late
but in the cobalt branch. Some of that effort is related to upstreaming
to kernel & iproute2 mainline but we're not quite there yet. Relevant
feature changes:
ingress mode: Instead of only counting packets that make it past the
shaper, include packets we've decided to drop as well, since they did
arrive with us on the link and took link capacity.
This mode is more suitable for shaping the ingress of a link
(e.g. from ISP) rather than the more normal egress.
ptm mode: Minor optimisation in packet overhead calculation.
dual-src/dsthost/triple-isolate: Optimise only calculating src or dst
host hashes only if required.
ack-filter/ack-filter-aggressive: Filter excessive TCP ACKS. Useful in
highly assymetric links (downstream v upstream capacity) where the
majority of upstream link capacity is occupied with ACKS for downstream
traffic.
A separate iproute2 patch to teach it about Cake's new features will
follow.
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
dfb2f6c pkt_sched: make compile again
5ab7026 sch_cake: make compile again
6f28803 codel5: make more checkpatch compliant
bd426aa Fix build error on 4.12
e4a3628 Whitespace tidy up
Signed-off-by: Fushan Wen <qydwhotmail@gmail.com>
This allows packages to use kernel make options without the forced
-C $(LINUX_DIR). It also makes it more clear that it to be called from
kernel module packages directly.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
wash, mpu & some memory optimisation have now made it to the official
cake repository.
Point LEDE to the official repository.
Signed-off-by: Kevin Darbyshire-Bryant <kevin@darbyshire-bryant.me.uk>
The last two parameters passed between user space tc and kernel space
sched-cake were transposed due to a merge mistake in a parameter header
file.
As such, using a packet overhead figure was likely to set cake to wash
packet DSCP values. Similarly, the DSCP wash flag was used as an offset
to the displayed packet overhead value.
Signed-off-by: Kevin Darbyshire-Bryant <kevin@darbyshire-bryant.me.uk>
Follow upstream cake:
diffserv3: a simple 3 tin classifier more in line with sqm-scripts
defaults for other qdisc algorithms.
defaults: default qdisc options are now diffserv3 and triple-isolate.
Signed-off-by: Kevin Darbyshire-Bryant <kevin@darbyshire-bryant.me.uk>
Track upstream changes, incl changes in packet overhead accounting
(automatically taking care of linux' packet sizing knowledge),
improvements to triple isolated DRR handling (new flow dominance),
statistics tweak & allow more packet drops in stressed conditions.
Under tests this has significantly improved latency control under
'many flows to one' scenarious as is typical of bittorrent and MS
Windows update.
I also restored 'DSCP washing' functionality in my repo which follows
upstream closely (like a hawk!) with tc keywords 'wash/nowash'. This
allows cake to limit/control packets in bands determined by a packet's
DSCP but to clear those DSCP bits on qdisc egress. This functionality
was originally removed as part of an attempt to push cake into the
kernel, which hasn't actually happened as yet.
A matching commit is required to iproute2/tc to support the new overhead
handling, keyword changes as well as the 'wash/nowash' tweak.
Signed-off-by: Kevin Darbyshire-Bryant <kevin@darbyshire-bryant.me.uk>
* Change git packages to xz
* Update mirror checksums in packages where they are used
* Change a few source tarballs to xz if available upstream
* Remove unused lines in packages we're touching, requested by jow- and blogic
* We're relying more on xz-utils so add official mirror as primary source, master site as secondary.
* Add SHA256 checksums to multiple git tarball packages
Signed-off-by: Daniel Engberg <daniel.engberg.lists@pyret.net>
COBALT is a hybrid codel/blue algo combining best elements of both.
Exhibits improved behaviour in presence of abuse from unresponsive flows
handled by 'blue', whereas responsive flows are still handled by codel.
Signed-off-by: Kevin Darbyshire-Bryant <kevin@darbyshire-bryant.me.uk>
In company networks everything except the http and https protocol is
often causes problems, because the network administrators try to block
everything else. To make it easier to use LEDE in company networks use
the https/http protocol for git access when possible.
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
Add 'cake' qdisc kernel module package.
V2 - KDB Small update to base on latest cake tc changes (wash option
deprecated)
V3 - KDB Move kmod-sched-cake package to kernel as is kernel related
V4 - KDB Split into individual patches, kmod & tc
Signed-off-by: Hannu Nyman <hannu.nyman@iki.fi>
Acked-by: Kevin Darbyshire-Bryant <kevin@darbyshire-bryant.me.uk>
