Use tcf lock to protect csum action struct private data from concurrent
modification in init and dump. Use rcu swap operation to reassign params
pointer under protection of tcf lock. (old params value is not used by
init, so there is no need of standalone rcu dereference step)
Remove rtnl assertion that is no longer necessary.
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use tcf spinlock to protect bpf action private data from concurrent
modification during dump and init. Remove rtnl lock assertion that is no
longer necessary.
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently the refcnt is never decremented in case the value is not 1.
Fix it by adding decrement in case the refcnt is not 1.
Reported-by: Vlad Buslov <vladbu@mellanox.com>
Fixes: f71e0ca4db ("net: sched: Avoid implicit chain 0 creation")
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
fl_reoffload implementation sets following members of struct
tc_cls_flower_offload incorrectly:
- masked key instead of mask
- key instead of masked key
Fix fl_reoffload to provide correct data to offload callback.
Fixes: 31533cba43 ("net: sched: cls_flower: implement offload tcf_proto_op")
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Allow matching on options in Geneve tunnel headers.
This makes use of existing tunnel metadata support.
The options can be described in the form
CLASS:TYPE:DATA/CLASS_MASK:TYPE_MASK:DATA_MASK, where CLASS is
represented as a 16bit hexadecimal value, TYPE as an 8bit
hexadecimal value and DATA as a variable length hexadecimal value.
e.g.
# ip link add name geneve0 type geneve dstport 0 external
# tc qdisc add dev geneve0 ingress
# tc filter add dev geneve0 protocol ip parent ffff: \
flower \
enc_src_ip 10.0.99.192 \
enc_dst_ip 10.0.99.193 \
enc_key_id 11 \
geneve_opts 0102:80:1122334421314151/ffff:ff:ffffffffffffffff \
ip_proto udp \
action mirred egress redirect dev eth1
This patch adds support for matching Geneve options in the order
supplied by the user. This leads to an efficient implementation in
the software datapath (and in our opinion hardware datapaths that
offload this feature). It is also compatible with Geneve options
matching provided by the Open vSwitch kernel datapath which is
relevant here as the Flower classifier may be used as a mechanism
to program flows into hardware as a form of Open vSwitch datapath
offload (sometimes referred to as OVS-TC). The netlink
Kernel/Userspace API may be extended, for example by adding a flag,
if other matching options are desired, for example matching given
options in any order. This would require an implementation in the
TC software datapath. And be done in a way that drivers that
facilitate offload of the Flower classifier can reject or accept
such flows based on hardware datapath capabilities.
This approach was discussed and agreed on at Netconf 2017 in Seoul.
Signed-off-by: Simon Horman <simon.horman@netronome.com>
Signed-off-by: Pieter Jansen van Vuuren <pieter.jansenvanvuuren@netronome.com>
Acked-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We forgot to set the error code on this path, so we return NULL instead
of an error pointer. In the current code kzalloc() won't fail for small
allocations so this doesn't really affect runtime.
Fixes: b95ec7eb3b ("net: sched: cls_flower: implement chain templates")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
User was able to perform filter flush on chain 0 even if it didn't have
any filters in it. With the patch that avoided implicit chain 0
creation, this changed. So in case user wants filter flush on chain
which does not exist, just return success. There's no reason for non-0
chains to behave differently than chain 0, so do the same for them.
Reported-by: Ido Schimmel <idosch@mellanox.com>
Fixes: f71e0ca4db ("net: sched: Avoid implicit chain 0 creation")
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
These are no longer used outside of cls_api.c so make them static.
Move tcf_chain_flush() to avoid fwd declaration of tcf_chain_put().
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
v1->v2:
- new patch
Signed-off-by: David S. Miller <davem@davemloft.net>
Chains that only have action references serve as placeholders.
Until a non-action reference is created, user should not be aware
of the chain. Also he should not receive any notifications about it.
So send notifications for the new chain only in case the chain gets
the first non-action reference. Symmetrically to that, when
the last non-action reference is dropped, send the notification about
deleted chain.
Reported-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Acked-by: Cong Wang <xiyou.wangcong@gmail.com>
v1->v2:
- made __tcf_chain_{get,put}() static as suggested by Cong
Signed-off-by: David S. Miller <davem@davemloft.net>
As mentioned by Cong and Jakub during the review process, it is a bit
odd to sometimes (act flow) create a new chain which would be
immediately a "zombie". So just rename it to "held_by_acts_only".
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Suggested-by: Cong Wang <xiyou.wangcong@gmail.com>
Suggested-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Acked-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When mirred is invoked from the ingress path, and it wants to redirect
the processed packet, it can now use the TC_ACT_REINSERT action,
filling the tcf_result accordingly, and avoiding a per packet
skb_clone().
Overall this gives a ~10% improvement in forwarding performance for the
TC S/W data path and TC S/W performances are now comparable to the
kernel openvswitch datapath.
v1 -> v2: use ACT_MIRRED instead of ACT_REDIRECT
v2 -> v3: updated after action rename, fixed typo into the commit
message
v3 -> v4: updated again after action rename, added more comments to
the code (JiriP), skip the optimization if the control action
need to touch the tcf_result (Paolo)
v4 -> v5: fix sparse warning (kbuild bot)
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Each lockless action currently does its own RCU locking in ->act().
This allows using plain RCU accessor, even if the context
is really RCU BH.
This change drops the per action RCU lock, replace the accessors
with the _bh variant, cleans up a bit the surrounding code and
documents the RCU status in the relevant header.
No functional nor performance change is intended.
The goal of this patch is clarifying that the RCU critical section
used by the tc actions extends up to the classifier's caller.
v1 -> v2:
- preserve rcu lock in act_bpf: it's needed by eBPF helpers,
as pointed out by Daniel
v3 -> v4:
- fixed some typos in the commit message (JiriP)
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, when initializing an action, the user-space can specify
and use arbitrary values for the tcfa_action field. If the value
is unknown by the kernel, is implicitly threaded as TC_ACT_UNSPEC.
This change explicitly checks for unknown values at action creation
time, and explicitly convert them to TC_ACT_UNSPEC. No functional
changes are introduced, but this will allow introducing tcfa_action
values not exposed to user-space in a later patch.
Note: we can't use the above to hide TC_ACT_REDIRECT from user-space,
as the latter is already part of uAPI.
v3 -> v4:
- use an helper to check for action validity (JiriP)
- emit an extack for invalid actions (JiriP)
v4 -> v5:
- keep messages on a single line, drop net_warn (Marcelo)
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Replace calls to kmalloc followed by a memcpy with a direct call to
kmemdup.
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Replace calls to kmalloc followed by a memcpy with a direct call to
kmemdup.
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
In case a chain is empty and not explicitly created by a user,
such chain should not exist. The only exception is if there is
an action "goto chain" pointing to it. In that case, don't show the
chain in the dump. Track the chain references held by actions and
use them to find out if a chain should or should not be shown
in chain dump.
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Once user manually deletes the chain using "chain del", the chain cannot
be marked as explicitly created anymore.
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Fixes: 32a4f5ecd7 ("net: sched: introduce chain object to uapi")
Signed-off-by: David S. Miller <davem@davemloft.net>
Code at line 1850 is unreachable. Fix this by removing the break
statement above it, so the code for case RTM_GETCHAIN can be
properly executed.
Addresses-Coverity-ID: 1472050 ("Structurally dead code")
Fixes: 32a4f5ecd7 ("net: sched: introduce chain object to uapi")
Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This will allow to install a child qdisc under cbs. The main use case
is to install ETF (Earliest TxTime First) qdisc under cbs, so there's
another level of control for time-sensitive traffic.
Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This fixes the following sparse warning:
net/sched/cls_flower.c:1356:36: warning: incorrect type in argument 3 (different base types)
net/sched/cls_flower.c:1356:36: expected unsigned short [unsigned] [usertype] value
net/sched/cls_flower.c:1356:36: got restricted __be16 [usertype] vlan_tpid
Signed-off-by: Jianbo Liu <jianbol@mellanox.com>
Reported-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Skbprio (SKB Priority Queue) is a queueing discipline that prioritizes packets
according to their skb->priority field. Under congestion, already-enqueued lower
priority packets will be dropped to make space available for higher priority
packets. Skbprio was conceived as a solution for denial-of-service defenses that
need to route packets with different priorities as a means to overcome DoS
attacks.
v5
*Do not reference qdisc_dev(sch)->tx_queue_len for setting limit. Instead set
default sch->limit to 64.
v4
*Drop Documentation/networking/sch_skbprio.txt doc file to move it to tc man
page for Skbprio, in iproute2.
v3
*Drop max_limit parameter in struct skbprio_sched_data and instead use
sch->limit.
*Reference qdisc_dev(sch)->tx_queue_len only once, during initialisation for
qdisc (previously being referenced every time qdisc changes).
*Move qdisc's detailed description from in-code to Documentation/networking.
*When qdisc is saturated, enqueue incoming packet first before dequeueing
lowest priority packet in queue - improves usage of call stack registers.
*Introduce and use overlimit stat to keep track of number of dropped packets.
v2
*Use skb->priority field rather than DS field. Rename queueing discipline as
SKB Priority Queue (previously Gatekeeper Priority Queue).
*Queueing discipline is made classful to expose Skbprio's internal priority
queues.
Signed-off-by: Nishanth Devarajan <ndev2021@gmail.com>
Reviewed-by: Sachin Paryani <sachin.paryani@gmail.com>
Reviewed-by: Cody Doucette <doucette@bu.edu>
Reviewed-by: Michel Machado <michel@digirati.com.br>
Acked-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove trailing whitespace and blank lines at EOF
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Introduce a couple of flower offload commands in order to propagate
template creation/destruction events down to device drivers.
Drivers may use this information to prepare HW in an optimal way
for future filter insertions.
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use the previously introduced template extension and implement
callback to create, destroy and dump chain template. The existing
parsing and dumping functions are re-used. Also, check if newly added
filters fit the template if it is set.
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This function is going to be used for templates as well, so we need to
pass the pointer separately.
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Push key/mask dumping from fl_dump() into a separate function
fl_dump_key(), that will be reused for template dumping.
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Allow user to set a template for newly created chains. Template lock
down the chain for particular classifier type/options combinations.
The classifier needs to support templates, otherwise kernel would
reply with error.
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Allow user to create, destroy, get and dump chain objects. Do that by
extending rtnl commands by the chain-specific ones. User will now be
able to explicitly create or destroy chains (so far this was done only
automatically according the filter/act needs and refcounting). Also, the
user will receive notification about any chain creation or destuction.
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, chain 0 is implicitly created during block creation. However
that does not align with chain object exposure, creation and destruction
api introduced later on. So make the chain 0 behave the same way as any
other chain and only create it when it is needed. Since chain 0 is
somehow special as the qdiscs need to hold pointer to the first chain
tp, this requires to move the chain head change callback infra to the
block structure.
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Push all bits that take care of ops lookup, including module loading
outside tcf_proto_create() function, into tcf_proto_lookup_ops()
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This line makes up what macro PTR_ERR_OR_ZERO already does. So,
make use of PTR_ERR_OR_ZERO rather than an open-code version.
This code was detected with the help of Coccinelle.
Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Allow users to set rules matching on ipv4 tos and ttl or
ipv6 traffic-class and hoplimit of tunnel headers.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Allow user-space to provide tos and ttl to be set for the tunnel headers.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fixes the following sparse warnings:
net/sched/cls_api.c:1101:43: warning: Using plain integer as NULL pointer
net/sched/cls_api.c:1492:75: warning: Using plain integer as NULL pointer
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In diffserv mode, CAKE stores tins in a different order internally than
the logical order exposed to userspace. The order remapping was missing
in the handling of 'tc filter' priority mappings through skb->priority,
resulting in bulk and best effort mappings being reversed relative to
how they are displayed.
Fix this by adding the missing mapping when reading skb->priority.
Fixes: 83f8fd69af ("sch_cake: Add DiffServ handling")
Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Extend struct tcf_walker with additional 'cookie' field. It is intended to
be used by classifier walk implementations to continue iteration directly
from particular filter, instead of iterating 'skip' number of times.
Change flower walk implementation to save filter handle in 'cookie'. Each
time flower walk is called, it looks up filter with saved handle directly
with idr, instead of iterating over filter linked list 'skip' number of
times. This change improves complexity of dumping flower classifier from
quadratic to linearithmic. (assuming idr lookup has logarithmic complexity)
Reviewed-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Reported-by: Simon Horman <simon.horman@netronome.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
use RCU instead of spin_{,un}lock_bh, to protect concurrent read/write on
act_skbedit configuration. This reduces the effects of contention in the
data path, in case multiple readers are present.
Signed-off-by: Davide Caratti <dcaratti@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
use per-CPU counters, instead of sharing a single set of stats with all
cores: this removes the need of spinlocks when stats are read/updated.
Signed-off-by: Davide Caratti <dcaratti@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix action attribute size calculation function to take rcu read lock and
access act_cookie pointer with rcu dereference.
Fixes: eec94fdb04 ("net: sched: use rcu for action cookie update")
Reported-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Free params if tcf_idr_check_alloc() returned error.
Fixes: 0190c1d452 ("net: sched: atomically check-allocate action")
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Zahari issued tc vlan command without setting vlan_ethtype, which will
crash kernel. To avoid this, we must check tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]
is not null before use it.
Also we don't need to dump vlan_ethtype or cvlan_ethtype in this case.
Fixes: d64efd0926 ('net/sched: flower: Add supprt for matching on QinQ vlan headers')
Signed-off-by: Jianbo Liu <jianbol@mellanox.com>
Reported-by: Zahari Doychev <zahari.doychev@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
At lower bandwidths, the transmission time of a single GSO segment can add
an unacceptable amount of latency due to HOL blocking. Furthermore, with a
software shaper, any tuning mechanism employed by the kernel to control the
maximum size of GSO segments is thrown off by the artificial limit on
bandwidth. For this reason, we split GSO segments into their individual
packets iff the shaper is active and configured to a bandwidth <= 1 Gbps.
Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit adds configurable overhead compensation support to the rate
shaper. With this feature, userspace can configure the actual bottleneck
link overhead and encapsulation mode used, which will be used by the shaper
to calculate the precise duration of each packet on the wire.
This feature is needed because CAKE is often deployed one or two hops
upstream of the actual bottleneck (which can be, e.g., inside a DSL or
cable modem). In this case, the link layer characteristics and overhead
reported by the kernel does not match the actual bottleneck. Being able to
set the actual values in use makes it possible to configure the shaper rate
much closer to the actual bottleneck rate (our experience shows it is
possible to get with 0.1% of the actual physical bottleneck rate), thus
keeping latency low without sacrificing bandwidth.
The overhead compensation has three tunables: A fixed per-packet overhead
size (which, if set, will be accounted from the IP packet header), a
minimum packet size (MPU) and a framing mode supporting either ATM or PTM
framing. We include a set of common keywords in TC to help users configure
the right parameters. If no overhead value is set, the value reported by
the kernel is used.
Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds support for DiffServ-based priority queueing to CAKE. If the
shaper is in use, each priority tier gets its own virtual clock, which
limits that tier's rate to a fraction of the overall shaped rate, to
discourage trying to game the priority mechanism.
CAKE defaults to a simple, three-tier mode that interprets most code points
as "best effort", but places CS1 traffic into a low-priority "bulk" tier
which is assigned 1/16 of the total rate, and a few code points indicating
latency-sensitive or control traffic (specifically TOS4, VA, EF, CS6, CS7)
into a "latency sensitive" high-priority tier, which is assigned 1/4 rate.
The other supported DiffServ modes are a 4-tier mode matching the 802.11e
precedence rules, as well as two 8-tier modes, one of which implements
strict precedence of the eight priority levels.
This commit also adds an optional DiffServ 'wash' mode, which will zero out
the DSCP fields of any packet passing through CAKE. While this can
technically be done with other mechanisms in the kernel, having the feature
available in CAKE significantly decreases configuration complexity; and the
implementation cost is low on top of the other DiffServ-handling code.
Filters and applications can set the skb->priority field to override the
DSCP-based classification into tiers. If TC_H_MAJ(skb->priority) matches
CAKE's qdisc handle, the minor number will be interpreted as a priority
tier if it is less than or equal to the number of configured priority
tiers.
Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
When CAKE is deployed on a gateway that also performs NAT (which is a
common deployment mode), the host fairness mechanism cannot distinguish
internal hosts from each other, and so fails to work correctly.
To fix this, we add an optional NAT awareness mode, which will query the
kernel conntrack mechanism to obtain the pre-NAT addresses for each packet
and use that in the flow and host hashing.
When the shaper is enabled and the host is already performing NAT, the cost
of this lookup is negligible. However, in unlimited mode with no NAT being
performed, there is a significant CPU cost at higher bandwidths. For this
reason, the feature is turned off by default.
Cc: netfilter-devel@vger.kernel.org
Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
The ACK filter is an optional feature of CAKE which is designed to improve
performance on links with very asymmetrical rate limits. On such links
(which are unfortunately quite prevalent, especially for DSL and cable
subscribers), the downstream throughput can be limited by the number of
ACKs capable of being transmitted in the *upstream* direction.
Filtering ACKs can, in general, have adverse effects on TCP performance
because it interferes with ACK clocking (especially in slow start), and it
reduces the flow's resiliency to ACKs being dropped further along the path.
To alleviate these drawbacks, the ACK filter in CAKE tries its best to
always keep enough ACKs queued to ensure forward progress in the TCP flow
being filtered. It does this by only filtering redundant ACKs. In its
default 'conservative' mode, the filter will always keep at least two
redundant ACKs in the queue, while in 'aggressive' mode, it will filter
down to a single ACK.
The ACK filter works by inspecting the per-flow queue on every packet
enqueue. Starting at the head of the queue, the filter looks for another
eligible packet to drop (so the ACK being dropped is always closer to the
head of the queue than the packet being enqueued). An ACK is eligible only
if it ACKs *fewer* bytes than the new packet being enqueued, including any
SACK options. This prevents duplicate ACKs from being filtered, to avoid
interfering with retransmission logic. In addition, we check TCP header
options and only drop those that are known to not interfere with sender
state. In particular, packets with unknown option codes are never dropped.
In aggressive mode, an eligible packet is always dropped, while in
conservative mode, at least two ACKs are kept in the queue. Only pure ACKs
(with no data segments) are considered eligible for dropping, but when an
ACK with data segments is enqueued, this can cause another pure ACK to
become eligible for dropping.
The approach described above ensures that this ACK filter avoids most of
the drawbacks of a naive filtering mechanism that only keeps flow state but
does not inspect the queue. This is the rationale for including the ACK
filter in CAKE itself rather than as separate module (as the TC filter, for
instance).
Our performance evaluation has shown that on a 30/1 Mbps link with a
bidirectional traffic test (RRUL), turning on the ACK filter on the
upstream link improves downstream throughput by ~20% (both modes) and
upstream throughput by ~12% in conservative mode and ~40% in aggressive
mode, at the cost of ~5ms of inter-flow latency due to the increased
congestion.
In *really* pathological cases, the effect can be a lot more; for instance,
the ACK filter increases the achievable downstream throughput on a link
with 100 Kbps in the upstream direction by an order of magnitude (from ~2.5
Mbps to ~25 Mbps).
Finally, even though we consider the ACK filter to be safer than most, we
do not recommend turning it on everywhere: on more symmetrical link
bandwidths the effect is negligible at best.
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
