CHROMIUM: net: sched: Adaptive Rate Liming Qdisc

Adaptive Rate Limiting Qdisc (ARL) is designed for home routers to
eliminate bufferbloat at upstream CPE (Cable/DSL) modem. It prevents
a standing queue from forming at the upstream node by limit egress
throughput to match the available upstream bandwidth. ARL does not
use a statically preconfigured limit for its rate limiting. Instead,
it passively measures the latency of the TCP connctions on its egress
link in realtime and throttles the rate when the minimuim latency of
all connections is seen to increase, which indicates the presence of
excessive queue at the upstream device.

ARL can be applied as root qdisc for WAN interface to prevent upstream
bufferbloat at the CPE modem. Queue is then managed locally at the
router, by applying another qdisc such as fq_codel as child qdisc.

This is a port from chromeos-3.18 to chromeos-4.14-gw of CL:1121778
(ported from commit cee41c3dd7f50e0605368402c1625863ba278216)
Note changes introduced as part of Linux-3.18 -> 4.14 porting...
  * Removed usage of nla_put_u64() [deprecated by Linux
    commit 50225243 ("netlink: kill nla_put_u64()")], replaced with
    nla_put_u64_64bit() [introduced by Linux
    commit 73520786 ("libnl: add nla_put_u64_64bit() helper")] that
    forces 64bit alignment of attribute. This also required definition
    of TCA_ARL_PAD attribute used for alignment pad.
  * Removed usage of out-of-line struct ewma [deprecated by Linux
    commit f4e774f5 ("average: remove out-of-line implementation")]
    replaced with static / inline macro [introduced by Linux commit
    commit 2377799c ("average: provide macro to create static
    EWMA")] and adapted to change in 'factor' parameter to 'precision'
    [introduced by Linux commit eb1e011a ("average: change to
    declare precision, not factor")]. The old 'weight' parameter of 8
    remains unchanged, just renamed 'weight_rcp'.  The old 'factor'
    parameter of 8 translates to 3 bits 'precision'.
  * Adapt to changes in kernel nanosecond timer APIs, ktime (and
    ktime_t) are now simple signed 64bit value rather than a union
    type [Linux commit 2456e855 ("ktime: Get rid of the union")].
  * The 'zone' parameter passed into nf_conntrack_find_get() is no
    longer an integer identifier
    [Linux commit 308ac914 ("netfilter: nf_conntrack: push zone
    object into functions")], a pointer to the 'nf_ct_zone_dflt' global
    const object is now required.
  * Removed all references to NF_CT_ASSERT() macro. The nf_conn
    'timeout' field checked was never actually referenced within the
    ARL code, the 'timeout.data' pointer was eliminated by conntrack
    change that switched from using kernel timer to simple timestamp
    [Linux commit f330a7fd ("netfilter: conntrack: get rid of
    conntrack timer")], and the macro has now been totally deprecated
    [Linux commit 9efdb14f ("net: Remove CONFIG_NETFILTER_DEBUG and
    _ASSERT() macros.")].
  * Remove nf_ct_is_untracked() check, the UNTRACKED state was
    deprecated [Linux commit cc41c84b ("netfilter: kill the fake
    untracked conntrack objects")] and then the API was completely
    removed [Linux commit ab8bc7ed ("netfilter: remove
    nf_ct_is_untracked")].
  * Adapt to changes in Qdisc_class_ops:
    - graft(): Replaced function body with call to qdisc_replace(),
        this appears to be what all the supported Qdisc have migrated
	to.
    - get()/put(): APIs were deprecated
      [Linux commit 143976ce ("net_sched: remove tc class reference
      counting")], new find() API is introduced and basically identical
      to old get().
  * Adapt to changes in Qdisc_ops:
    - enqueue(): Accepts new packets 'to_free' list, and use
        qdisc_drop() rather than qdisc_reshape_fail() when packet
	cannot be forwarded.  gso_segment() helper function now also
	utilizes packets 'to_free' list and qdisc_tree_reduce_backlog()
	rather than qdisc_tree_decrease_qlen() to update queue length
	accounting.
    - dequeue(): Removed call to qdisc_unthrottled(), generic throttled
      management was deprecated [Linux commit 45f50bed ("net_sched:
      remove generic throttled management")].
    - drop(): Deprecated [Linux commit a09ceb0e ("sched: remove
        qdisc->drop")].
    - change(): Added call to qdisc_hash_add() after default FIFO child
        Qdisc is created and when adjusting queue length accounting
	utilize qdisc_tree_reduce_backlog() rather than
	qdisc_tree_decrease_qlen(). This is based on coding practices
	observed in other supported Qdisc (e.g. sch_tbf, sch_prio).
Note changes introduced as part of 64bit porting...
  * Added explicit cast (unsigned long) for first parameter of clamp()
    when 'lo'/'hi' are 'UL' constants. clamp() strongly enforces type
    equality at compile time. On 64bit architecture the first parameter
    (2 * minmax_get() / 1000L) (i.e. int * u32 * long) results in
    'signed long' which doesn't match the 'unsigned long' second
    parameter type.

BUG=b:138685581
TEST=Latency test with heavy upstream traffic.

Cq-Depend: chromium:1727770, chromium:1730603
Change-Id: I727e4737ba45b8136e70643f159534cf7d06739e
Signed-off-by: Danny J. Mitzel <mitzel@google.com>
Reviewed-on: https://chromium-review.googlesource.com/1731583


Tested-by: Danny Mitzel <mitzel@google.com>
Commit-Ready: Danny Mitzel <mitzel@google.com>
Legacy-Commit-Queue: Commit Bot <commit-bot@chromium.org>
Reviewed-by: Kan Yan <kyan@chromium.org>