kernel-ark/include/net/red.h
Hannes Frederic Sowa 809fa972fd reciprocal_divide: update/correction of the algorithm
Jakub Zawadzki noticed that some divisions by reciprocal_divide()
were not correct [1][2], which he could also show with BPF code
after divisions are transformed into reciprocal_value() for runtime
invariance which can be passed to reciprocal_divide() later on;
reverse in BPF dump ended up with a different, off-by-one K in
some situations.

This has been fixed by Eric Dumazet in commit aee636c480
("bpf: do not use reciprocal divide"). This follow-up patch
improves reciprocal_value() and reciprocal_divide() to work in
all cases by using Granlund and Montgomery method, so that also
future use is safe and without any non-obvious side-effects.
Known problems with the old implementation were that division by 1
always returned 0 and some off-by-ones when the dividend and divisor
where very large. This seemed to not be problematic with its
current users, as far as we can tell. Eric Dumazet checked for
the slab usage, we cannot surely say so in the case of flex_array.
Still, in order to fix that, we propose an extension from the
original implementation from commit 6a2d7a955d resp. [3][4],
by using the algorithm proposed in "Division by Invariant Integers
Using Multiplication" [5], Torbjörn Granlund and Peter L.
Montgomery, that is, pseudocode for q = n/d where q, n, d is in
u32 universe:

1) Initialization:

  int l = ceil(log_2 d)
  uword m' = floor((1<<32)*((1<<l)-d)/d)+1
  int sh_1 = min(l,1)
  int sh_2 = max(l-1,0)

2) For q = n/d, all uword:

  uword t = (n*m')>>32
  q = (t+((n-t)>>sh_1))>>sh_2

The assembler implementation from Agner Fog [6] also helped a lot
while implementing. We have tested the implementation on x86_64,
ppc64, i686, s390x; on x86_64/haswell we're still half the latency
compared to normal divide.

Joint work with Daniel Borkmann.

  [1] http://www.wireshark.org/~darkjames/reciprocal-buggy.c
  [2] http://www.wireshark.org/~darkjames/set-and-dump-filter-k-bug.c
  [3] https://gmplib.org/~tege/division-paper.pdf
  [4] http://homepage.cs.uiowa.edu/~jones/bcd/divide.html
  [5] http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.1.2556
  [6] http://www.agner.org/optimize/asmlib.zip

Reported-by: Jakub Zawadzki <darkjames-ws@darkjames.pl>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Austin S Hemmelgarn <ahferroin7@gmail.com>
Cc: linux-kernel@vger.kernel.org
Cc: Jesse Gross <jesse@nicira.com>
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Cc: Stephen Hemminger <stephen@networkplumber.org>
Cc: Matt Mackall <mpm@selenic.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Andy Gospodarek <andy@greyhouse.net>
Cc: Veaceslav Falico <vfalico@redhat.com>
Cc: Jay Vosburgh <fubar@us.ibm.com>
Cc: Jakub Zawadzki <darkjames-ws@darkjames.pl>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-01-21 23:17:20 -08:00

407 lines
10 KiB
C

#ifndef __NET_SCHED_RED_H
#define __NET_SCHED_RED_H
#include <linux/types.h>
#include <linux/bug.h>
#include <net/pkt_sched.h>
#include <net/inet_ecn.h>
#include <net/dsfield.h>
#include <linux/reciprocal_div.h>
/* Random Early Detection (RED) algorithm.
=======================================
Source: Sally Floyd and Van Jacobson, "Random Early Detection Gateways
for Congestion Avoidance", 1993, IEEE/ACM Transactions on Networking.
This file codes a "divisionless" version of RED algorithm
as written down in Fig.17 of the paper.
Short description.
------------------
When a new packet arrives we calculate the average queue length:
avg = (1-W)*avg + W*current_queue_len,
W is the filter time constant (chosen as 2^(-Wlog)), it controls
the inertia of the algorithm. To allow larger bursts, W should be
decreased.
if (avg > th_max) -> packet marked (dropped).
if (avg < th_min) -> packet passes.
if (th_min < avg < th_max) we calculate probability:
Pb = max_P * (avg - th_min)/(th_max-th_min)
and mark (drop) packet with this probability.
Pb changes from 0 (at avg==th_min) to max_P (avg==th_max).
max_P should be small (not 1), usually 0.01..0.02 is good value.
max_P is chosen as a number, so that max_P/(th_max-th_min)
is a negative power of two in order arithmetics to contain
only shifts.
Parameters, settable by user:
-----------------------------
qth_min - bytes (should be < qth_max/2)
qth_max - bytes (should be at least 2*qth_min and less limit)
Wlog - bits (<32) log(1/W).
Plog - bits (<32)
Plog is related to max_P by formula:
max_P = (qth_max-qth_min)/2^Plog;
F.e. if qth_max=128K and qth_min=32K, then Plog=22
corresponds to max_P=0.02
Scell_log
Stab
Lookup table for log((1-W)^(t/t_ave).
NOTES:
Upper bound on W.
-----------------
If you want to allow bursts of L packets of size S,
you should choose W:
L + 1 - th_min/S < (1-(1-W)^L)/W
th_min/S = 32 th_min/S = 4
log(W) L
-1 33
-2 35
-3 39
-4 46
-5 57
-6 75
-7 101
-8 135
-9 190
etc.
*/
/*
* Adaptative RED : An Algorithm for Increasing the Robustness of RED's AQM
* (Sally FLoyd, Ramakrishna Gummadi, and Scott Shenker) August 2001
*
* Every 500 ms:
* if (avg > target and max_p <= 0.5)
* increase max_p : max_p += alpha;
* else if (avg < target and max_p >= 0.01)
* decrease max_p : max_p *= beta;
*
* target :[qth_min + 0.4*(qth_min - qth_max),
* qth_min + 0.6*(qth_min - qth_max)].
* alpha : min(0.01, max_p / 4)
* beta : 0.9
* max_P is a Q0.32 fixed point number (with 32 bits mantissa)
* max_P between 0.01 and 0.5 (1% - 50%) [ Its no longer a negative power of two ]
*/
#define RED_ONE_PERCENT ((u32)DIV_ROUND_CLOSEST(1ULL<<32, 100))
#define MAX_P_MIN (1 * RED_ONE_PERCENT)
#define MAX_P_MAX (50 * RED_ONE_PERCENT)
#define MAX_P_ALPHA(val) min(MAX_P_MIN, val / 4)
#define RED_STAB_SIZE 256
#define RED_STAB_MASK (RED_STAB_SIZE - 1)
struct red_stats {
u32 prob_drop; /* Early probability drops */
u32 prob_mark; /* Early probability marks */
u32 forced_drop; /* Forced drops, qavg > max_thresh */
u32 forced_mark; /* Forced marks, qavg > max_thresh */
u32 pdrop; /* Drops due to queue limits */
u32 other; /* Drops due to drop() calls */
};
struct red_parms {
/* Parameters */
u32 qth_min; /* Min avg length threshold: Wlog scaled */
u32 qth_max; /* Max avg length threshold: Wlog scaled */
u32 Scell_max;
u32 max_P; /* probability, [0 .. 1.0] 32 scaled */
/* reciprocal_value(max_P / qth_delta) */
struct reciprocal_value max_P_reciprocal;
u32 qth_delta; /* max_th - min_th */
u32 target_min; /* min_th + 0.4*(max_th - min_th) */
u32 target_max; /* min_th + 0.6*(max_th - min_th) */
u8 Scell_log;
u8 Wlog; /* log(W) */
u8 Plog; /* random number bits */
u8 Stab[RED_STAB_SIZE];
};
struct red_vars {
/* Variables */
int qcount; /* Number of packets since last random
number generation */
u32 qR; /* Cached random number */
unsigned long qavg; /* Average queue length: Wlog scaled */
ktime_t qidlestart; /* Start of current idle period */
};
static inline u32 red_maxp(u8 Plog)
{
return Plog < 32 ? (~0U >> Plog) : ~0U;
}
static inline void red_set_vars(struct red_vars *v)
{
/* Reset average queue length, the value is strictly bound
* to the parameters below, reseting hurts a bit but leaving
* it might result in an unreasonable qavg for a while. --TGR
*/
v->qavg = 0;
v->qcount = -1;
}
static inline void red_set_parms(struct red_parms *p,
u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog,
u8 Scell_log, u8 *stab, u32 max_P)
{
int delta = qth_max - qth_min;
u32 max_p_delta;
p->qth_min = qth_min << Wlog;
p->qth_max = qth_max << Wlog;
p->Wlog = Wlog;
p->Plog = Plog;
if (delta < 0)
delta = 1;
p->qth_delta = delta;
if (!max_P) {
max_P = red_maxp(Plog);
max_P *= delta; /* max_P = (qth_max - qth_min)/2^Plog */
}
p->max_P = max_P;
max_p_delta = max_P / delta;
max_p_delta = max(max_p_delta, 1U);
p->max_P_reciprocal = reciprocal_value(max_p_delta);
/* RED Adaptative target :
* [min_th + 0.4*(min_th - max_th),
* min_th + 0.6*(min_th - max_th)].
*/
delta /= 5;
p->target_min = qth_min + 2*delta;
p->target_max = qth_min + 3*delta;
p->Scell_log = Scell_log;
p->Scell_max = (255 << Scell_log);
if (stab)
memcpy(p->Stab, stab, sizeof(p->Stab));
}
static inline int red_is_idling(const struct red_vars *v)
{
return v->qidlestart.tv64 != 0;
}
static inline void red_start_of_idle_period(struct red_vars *v)
{
v->qidlestart = ktime_get();
}
static inline void red_end_of_idle_period(struct red_vars *v)
{
v->qidlestart.tv64 = 0;
}
static inline void red_restart(struct red_vars *v)
{
red_end_of_idle_period(v);
v->qavg = 0;
v->qcount = -1;
}
static inline unsigned long red_calc_qavg_from_idle_time(const struct red_parms *p,
const struct red_vars *v)
{
s64 delta = ktime_us_delta(ktime_get(), v->qidlestart);
long us_idle = min_t(s64, delta, p->Scell_max);
int shift;
/*
* The problem: ideally, average length queue recalcultion should
* be done over constant clock intervals. This is too expensive, so
* that the calculation is driven by outgoing packets.
* When the queue is idle we have to model this clock by hand.
*
* SF+VJ proposed to "generate":
*
* m = idletime / (average_pkt_size / bandwidth)
*
* dummy packets as a burst after idle time, i.e.
*
* v->qavg *= (1-W)^m
*
* This is an apparently overcomplicated solution (f.e. we have to
* precompute a table to make this calculation in reasonable time)
* I believe that a simpler model may be used here,
* but it is field for experiments.
*/
shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK];
if (shift)
return v->qavg >> shift;
else {
/* Approximate initial part of exponent with linear function:
*
* (1-W)^m ~= 1-mW + ...
*
* Seems, it is the best solution to
* problem of too coarse exponent tabulation.
*/
us_idle = (v->qavg * (u64)us_idle) >> p->Scell_log;
if (us_idle < (v->qavg >> 1))
return v->qavg - us_idle;
else
return v->qavg >> 1;
}
}
static inline unsigned long red_calc_qavg_no_idle_time(const struct red_parms *p,
const struct red_vars *v,
unsigned int backlog)
{
/*
* NOTE: v->qavg is fixed point number with point at Wlog.
* The formula below is equvalent to floating point
* version:
*
* qavg = qavg*(1-W) + backlog*W;
*
* --ANK (980924)
*/
return v->qavg + (backlog - (v->qavg >> p->Wlog));
}
static inline unsigned long red_calc_qavg(const struct red_parms *p,
const struct red_vars *v,
unsigned int backlog)
{
if (!red_is_idling(v))
return red_calc_qavg_no_idle_time(p, v, backlog);
else
return red_calc_qavg_from_idle_time(p, v);
}
static inline u32 red_random(const struct red_parms *p)
{
return reciprocal_divide(prandom_u32(), p->max_P_reciprocal);
}
static inline int red_mark_probability(const struct red_parms *p,
const struct red_vars *v,
unsigned long qavg)
{
/* The formula used below causes questions.
OK. qR is random number in the interval
(0..1/max_P)*(qth_max-qth_min)
i.e. 0..(2^Plog). If we used floating point
arithmetics, it would be: (2^Plog)*rnd_num,
where rnd_num is less 1.
Taking into account, that qavg have fixed
point at Wlog, two lines
below have the following floating point equivalent:
max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount
Any questions? --ANK (980924)
*/
return !(((qavg - p->qth_min) >> p->Wlog) * v->qcount < v->qR);
}
enum {
RED_BELOW_MIN_THRESH,
RED_BETWEEN_TRESH,
RED_ABOVE_MAX_TRESH,
};
static inline int red_cmp_thresh(const struct red_parms *p, unsigned long qavg)
{
if (qavg < p->qth_min)
return RED_BELOW_MIN_THRESH;
else if (qavg >= p->qth_max)
return RED_ABOVE_MAX_TRESH;
else
return RED_BETWEEN_TRESH;
}
enum {
RED_DONT_MARK,
RED_PROB_MARK,
RED_HARD_MARK,
};
static inline int red_action(const struct red_parms *p,
struct red_vars *v,
unsigned long qavg)
{
switch (red_cmp_thresh(p, qavg)) {
case RED_BELOW_MIN_THRESH:
v->qcount = -1;
return RED_DONT_MARK;
case RED_BETWEEN_TRESH:
if (++v->qcount) {
if (red_mark_probability(p, v, qavg)) {
v->qcount = 0;
v->qR = red_random(p);
return RED_PROB_MARK;
}
} else
v->qR = red_random(p);
return RED_DONT_MARK;
case RED_ABOVE_MAX_TRESH:
v->qcount = -1;
return RED_HARD_MARK;
}
BUG();
return RED_DONT_MARK;
}
static inline void red_adaptative_algo(struct red_parms *p, struct red_vars *v)
{
unsigned long qavg;
u32 max_p_delta;
qavg = v->qavg;
if (red_is_idling(v))
qavg = red_calc_qavg_from_idle_time(p, v);
/* v->qavg is fixed point number with point at Wlog */
qavg >>= p->Wlog;
if (qavg > p->target_max && p->max_P <= MAX_P_MAX)
p->max_P += MAX_P_ALPHA(p->max_P); /* maxp = maxp + alpha */
else if (qavg < p->target_min && p->max_P >= MAX_P_MIN)
p->max_P = (p->max_P/10)*9; /* maxp = maxp * Beta */
max_p_delta = DIV_ROUND_CLOSEST(p->max_P, p->qth_delta);
max_p_delta = max(max_p_delta, 1U);
p->max_P_reciprocal = reciprocal_value(max_p_delta);
}
#endif