cachepc-linux

Fork of AMDESE/linux with modifications for CachePC side-channel attack
git clone https://git.sinitax.com/sinitax/cachepc-linux
Log | Files | Refs | README | LICENSE | sfeed.txt

red.h (11654B)


      1/* SPDX-License-Identifier: GPL-2.0 */
      2#ifndef __NET_SCHED_RED_H
      3#define __NET_SCHED_RED_H
      4
      5#include <linux/types.h>
      6#include <linux/bug.h>
      7#include <net/pkt_sched.h>
      8#include <net/inet_ecn.h>
      9#include <net/dsfield.h>
     10#include <linux/reciprocal_div.h>
     11
     12/*	Random Early Detection (RED) algorithm.
     13	=======================================
     14
     15	Source: Sally Floyd and Van Jacobson, "Random Early Detection Gateways
     16	for Congestion Avoidance", 1993, IEEE/ACM Transactions on Networking.
     17
     18	This file codes a "divisionless" version of RED algorithm
     19	as written down in Fig.17 of the paper.
     20
     21	Short description.
     22	------------------
     23
     24	When a new packet arrives we calculate the average queue length:
     25
     26	avg = (1-W)*avg + W*current_queue_len,
     27
     28	W is the filter time constant (chosen as 2^(-Wlog)), it controls
     29	the inertia of the algorithm. To allow larger bursts, W should be
     30	decreased.
     31
     32	if (avg > th_max) -> packet marked (dropped).
     33	if (avg < th_min) -> packet passes.
     34	if (th_min < avg < th_max) we calculate probability:
     35
     36	Pb = max_P * (avg - th_min)/(th_max-th_min)
     37
     38	and mark (drop) packet with this probability.
     39	Pb changes from 0 (at avg==th_min) to max_P (avg==th_max).
     40	max_P should be small (not 1), usually 0.01..0.02 is good value.
     41
     42	max_P is chosen as a number, so that max_P/(th_max-th_min)
     43	is a negative power of two in order arithmetics to contain
     44	only shifts.
     45
     46
     47	Parameters, settable by user:
     48	-----------------------------
     49
     50	qth_min		- bytes (should be < qth_max/2)
     51	qth_max		- bytes (should be at least 2*qth_min and less limit)
     52	Wlog	       	- bits (<32) log(1/W).
     53	Plog	       	- bits (<32)
     54
     55	Plog is related to max_P by formula:
     56
     57	max_P = (qth_max-qth_min)/2^Plog;
     58
     59	F.e. if qth_max=128K and qth_min=32K, then Plog=22
     60	corresponds to max_P=0.02
     61
     62	Scell_log
     63	Stab
     64
     65	Lookup table for log((1-W)^(t/t_ave).
     66
     67
     68	NOTES:
     69
     70	Upper bound on W.
     71	-----------------
     72
     73	If you want to allow bursts of L packets of size S,
     74	you should choose W:
     75
     76	L + 1 - th_min/S < (1-(1-W)^L)/W
     77
     78	th_min/S = 32         th_min/S = 4
     79
     80	log(W)	L
     81	-1	33
     82	-2	35
     83	-3	39
     84	-4	46
     85	-5	57
     86	-6	75
     87	-7	101
     88	-8	135
     89	-9	190
     90	etc.
     91 */
     92
     93/*
     94 * Adaptative RED : An Algorithm for Increasing the Robustness of RED's AQM
     95 * (Sally FLoyd, Ramakrishna Gummadi, and Scott Shenker) August 2001
     96 *
     97 * Every 500 ms:
     98 *  if (avg > target and max_p <= 0.5)
     99 *   increase max_p : max_p += alpha;
    100 *  else if (avg < target and max_p >= 0.01)
    101 *   decrease max_p : max_p *= beta;
    102 *
    103 * target :[qth_min + 0.4*(qth_min - qth_max),
    104 *          qth_min + 0.6*(qth_min - qth_max)].
    105 * alpha : min(0.01, max_p / 4)
    106 * beta : 0.9
    107 * max_P is a Q0.32 fixed point number (with 32 bits mantissa)
    108 * max_P between 0.01 and 0.5 (1% - 50%) [ Its no longer a negative power of two ]
    109 */
    110#define RED_ONE_PERCENT ((u32)DIV_ROUND_CLOSEST(1ULL<<32, 100))
    111
    112#define MAX_P_MIN (1 * RED_ONE_PERCENT)
    113#define MAX_P_MAX (50 * RED_ONE_PERCENT)
    114#define MAX_P_ALPHA(val) min(MAX_P_MIN, val / 4)
    115
    116#define RED_STAB_SIZE	256
    117#define RED_STAB_MASK	(RED_STAB_SIZE - 1)
    118
    119struct red_stats {
    120	u32		prob_drop;	/* Early probability drops */
    121	u32		prob_mark;	/* Early probability marks */
    122	u32		forced_drop;	/* Forced drops, qavg > max_thresh */
    123	u32		forced_mark;	/* Forced marks, qavg > max_thresh */
    124	u32		pdrop;          /* Drops due to queue limits */
    125	u32		other;          /* Drops due to drop() calls */
    126};
    127
    128struct red_parms {
    129	/* Parameters */
    130	u32		qth_min;	/* Min avg length threshold: Wlog scaled */
    131	u32		qth_max;	/* Max avg length threshold: Wlog scaled */
    132	u32		Scell_max;
    133	u32		max_P;		/* probability, [0 .. 1.0] 32 scaled */
    134	/* reciprocal_value(max_P / qth_delta) */
    135	struct reciprocal_value	max_P_reciprocal;
    136	u32		qth_delta;	/* max_th - min_th */
    137	u32		target_min;	/* min_th + 0.4*(max_th - min_th) */
    138	u32		target_max;	/* min_th + 0.6*(max_th - min_th) */
    139	u8		Scell_log;
    140	u8		Wlog;		/* log(W)		*/
    141	u8		Plog;		/* random number bits	*/
    142	u8		Stab[RED_STAB_SIZE];
    143};
    144
    145struct red_vars {
    146	/* Variables */
    147	int		qcount;		/* Number of packets since last random
    148					   number generation */
    149	u32		qR;		/* Cached random number */
    150
    151	unsigned long	qavg;		/* Average queue length: Wlog scaled */
    152	ktime_t		qidlestart;	/* Start of current idle period */
    153};
    154
    155static inline u32 red_maxp(u8 Plog)
    156{
    157	return Plog < 32 ? (~0U >> Plog) : ~0U;
    158}
    159
    160static inline void red_set_vars(struct red_vars *v)
    161{
    162	/* Reset average queue length, the value is strictly bound
    163	 * to the parameters below, reseting hurts a bit but leaving
    164	 * it might result in an unreasonable qavg for a while. --TGR
    165	 */
    166	v->qavg		= 0;
    167
    168	v->qcount	= -1;
    169}
    170
    171static inline bool red_check_params(u32 qth_min, u32 qth_max, u8 Wlog,
    172				    u8 Scell_log, u8 *stab)
    173{
    174	if (fls(qth_min) + Wlog >= 32)
    175		return false;
    176	if (fls(qth_max) + Wlog >= 32)
    177		return false;
    178	if (Scell_log >= 32)
    179		return false;
    180	if (qth_max < qth_min)
    181		return false;
    182	if (stab) {
    183		int i;
    184
    185		for (i = 0; i < RED_STAB_SIZE; i++)
    186			if (stab[i] >= 32)
    187				return false;
    188	}
    189	return true;
    190}
    191
    192static inline int red_get_flags(unsigned char qopt_flags,
    193				unsigned char historic_mask,
    194				struct nlattr *flags_attr,
    195				unsigned char supported_mask,
    196				struct nla_bitfield32 *p_flags,
    197				unsigned char *p_userbits,
    198				struct netlink_ext_ack *extack)
    199{
    200	struct nla_bitfield32 flags;
    201
    202	if (qopt_flags && flags_attr) {
    203		NL_SET_ERR_MSG_MOD(extack, "flags should be passed either through qopt, or through a dedicated attribute");
    204		return -EINVAL;
    205	}
    206
    207	if (flags_attr) {
    208		flags = nla_get_bitfield32(flags_attr);
    209	} else {
    210		flags.selector = historic_mask;
    211		flags.value = qopt_flags & historic_mask;
    212	}
    213
    214	*p_flags = flags;
    215	*p_userbits = qopt_flags & ~historic_mask;
    216	return 0;
    217}
    218
    219static inline int red_validate_flags(unsigned char flags,
    220				     struct netlink_ext_ack *extack)
    221{
    222	if ((flags & TC_RED_NODROP) && !(flags & TC_RED_ECN)) {
    223		NL_SET_ERR_MSG_MOD(extack, "nodrop mode is only meaningful with ECN");
    224		return -EINVAL;
    225	}
    226
    227	return 0;
    228}
    229
    230static inline void red_set_parms(struct red_parms *p,
    231				 u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog,
    232				 u8 Scell_log, u8 *stab, u32 max_P)
    233{
    234	int delta = qth_max - qth_min;
    235	u32 max_p_delta;
    236
    237	p->qth_min	= qth_min << Wlog;
    238	p->qth_max	= qth_max << Wlog;
    239	p->Wlog		= Wlog;
    240	p->Plog		= Plog;
    241	if (delta <= 0)
    242		delta = 1;
    243	p->qth_delta	= delta;
    244	if (!max_P) {
    245		max_P = red_maxp(Plog);
    246		max_P *= delta; /* max_P = (qth_max - qth_min)/2^Plog */
    247	}
    248	p->max_P = max_P;
    249	max_p_delta = max_P / delta;
    250	max_p_delta = max(max_p_delta, 1U);
    251	p->max_P_reciprocal  = reciprocal_value(max_p_delta);
    252
    253	/* RED Adaptative target :
    254	 * [min_th + 0.4*(min_th - max_th),
    255	 *  min_th + 0.6*(min_th - max_th)].
    256	 */
    257	delta /= 5;
    258	p->target_min = qth_min + 2*delta;
    259	p->target_max = qth_min + 3*delta;
    260
    261	p->Scell_log	= Scell_log;
    262	p->Scell_max	= (255 << Scell_log);
    263
    264	if (stab)
    265		memcpy(p->Stab, stab, sizeof(p->Stab));
    266}
    267
    268static inline int red_is_idling(const struct red_vars *v)
    269{
    270	return v->qidlestart != 0;
    271}
    272
    273static inline void red_start_of_idle_period(struct red_vars *v)
    274{
    275	v->qidlestart = ktime_get();
    276}
    277
    278static inline void red_end_of_idle_period(struct red_vars *v)
    279{
    280	v->qidlestart = 0;
    281}
    282
    283static inline void red_restart(struct red_vars *v)
    284{
    285	red_end_of_idle_period(v);
    286	v->qavg = 0;
    287	v->qcount = -1;
    288}
    289
    290static inline unsigned long red_calc_qavg_from_idle_time(const struct red_parms *p,
    291							 const struct red_vars *v)
    292{
    293	s64 delta = ktime_us_delta(ktime_get(), v->qidlestart);
    294	long us_idle = min_t(s64, delta, p->Scell_max);
    295	int  shift;
    296
    297	/*
    298	 * The problem: ideally, average length queue recalculation should
    299	 * be done over constant clock intervals. This is too expensive, so
    300	 * that the calculation is driven by outgoing packets.
    301	 * When the queue is idle we have to model this clock by hand.
    302	 *
    303	 * SF+VJ proposed to "generate":
    304	 *
    305	 *	m = idletime / (average_pkt_size / bandwidth)
    306	 *
    307	 * dummy packets as a burst after idle time, i.e.
    308	 *
    309	 * 	v->qavg *= (1-W)^m
    310	 *
    311	 * This is an apparently overcomplicated solution (f.e. we have to
    312	 * precompute a table to make this calculation in reasonable time)
    313	 * I believe that a simpler model may be used here,
    314	 * but it is field for experiments.
    315	 */
    316
    317	shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK];
    318
    319	if (shift)
    320		return v->qavg >> shift;
    321	else {
    322		/* Approximate initial part of exponent with linear function:
    323		 *
    324		 * 	(1-W)^m ~= 1-mW + ...
    325		 *
    326		 * Seems, it is the best solution to
    327		 * problem of too coarse exponent tabulation.
    328		 */
    329		us_idle = (v->qavg * (u64)us_idle) >> p->Scell_log;
    330
    331		if (us_idle < (v->qavg >> 1))
    332			return v->qavg - us_idle;
    333		else
    334			return v->qavg >> 1;
    335	}
    336}
    337
    338static inline unsigned long red_calc_qavg_no_idle_time(const struct red_parms *p,
    339						       const struct red_vars *v,
    340						       unsigned int backlog)
    341{
    342	/*
    343	 * NOTE: v->qavg is fixed point number with point at Wlog.
    344	 * The formula below is equvalent to floating point
    345	 * version:
    346	 *
    347	 * 	qavg = qavg*(1-W) + backlog*W;
    348	 *
    349	 * --ANK (980924)
    350	 */
    351	return v->qavg + (backlog - (v->qavg >> p->Wlog));
    352}
    353
    354static inline unsigned long red_calc_qavg(const struct red_parms *p,
    355					  const struct red_vars *v,
    356					  unsigned int backlog)
    357{
    358	if (!red_is_idling(v))
    359		return red_calc_qavg_no_idle_time(p, v, backlog);
    360	else
    361		return red_calc_qavg_from_idle_time(p, v);
    362}
    363
    364
    365static inline u32 red_random(const struct red_parms *p)
    366{
    367	return reciprocal_divide(prandom_u32(), p->max_P_reciprocal);
    368}
    369
    370static inline int red_mark_probability(const struct red_parms *p,
    371				       const struct red_vars *v,
    372				       unsigned long qavg)
    373{
    374	/* The formula used below causes questions.
    375
    376	   OK. qR is random number in the interval
    377		(0..1/max_P)*(qth_max-qth_min)
    378	   i.e. 0..(2^Plog). If we used floating point
    379	   arithmetics, it would be: (2^Plog)*rnd_num,
    380	   where rnd_num is less 1.
    381
    382	   Taking into account, that qavg have fixed
    383	   point at Wlog, two lines
    384	   below have the following floating point equivalent:
    385
    386	   max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount
    387
    388	   Any questions? --ANK (980924)
    389	 */
    390	return !(((qavg - p->qth_min) >> p->Wlog) * v->qcount < v->qR);
    391}
    392
    393enum {
    394	RED_BELOW_MIN_THRESH,
    395	RED_BETWEEN_TRESH,
    396	RED_ABOVE_MAX_TRESH,
    397};
    398
    399static inline int red_cmp_thresh(const struct red_parms *p, unsigned long qavg)
    400{
    401	if (qavg < p->qth_min)
    402		return RED_BELOW_MIN_THRESH;
    403	else if (qavg >= p->qth_max)
    404		return RED_ABOVE_MAX_TRESH;
    405	else
    406		return RED_BETWEEN_TRESH;
    407}
    408
    409enum {
    410	RED_DONT_MARK,
    411	RED_PROB_MARK,
    412	RED_HARD_MARK,
    413};
    414
    415static inline int red_action(const struct red_parms *p,
    416			     struct red_vars *v,
    417			     unsigned long qavg)
    418{
    419	switch (red_cmp_thresh(p, qavg)) {
    420		case RED_BELOW_MIN_THRESH:
    421			v->qcount = -1;
    422			return RED_DONT_MARK;
    423
    424		case RED_BETWEEN_TRESH:
    425			if (++v->qcount) {
    426				if (red_mark_probability(p, v, qavg)) {
    427					v->qcount = 0;
    428					v->qR = red_random(p);
    429					return RED_PROB_MARK;
    430				}
    431			} else
    432				v->qR = red_random(p);
    433
    434			return RED_DONT_MARK;
    435
    436		case RED_ABOVE_MAX_TRESH:
    437			v->qcount = -1;
    438			return RED_HARD_MARK;
    439	}
    440
    441	BUG();
    442	return RED_DONT_MARK;
    443}
    444
    445static inline void red_adaptative_algo(struct red_parms *p, struct red_vars *v)
    446{
    447	unsigned long qavg;
    448	u32 max_p_delta;
    449
    450	qavg = v->qavg;
    451	if (red_is_idling(v))
    452		qavg = red_calc_qavg_from_idle_time(p, v);
    453
    454	/* v->qavg is fixed point number with point at Wlog */
    455	qavg >>= p->Wlog;
    456
    457	if (qavg > p->target_max && p->max_P <= MAX_P_MAX)
    458		p->max_P += MAX_P_ALPHA(p->max_P); /* maxp = maxp + alpha */
    459	else if (qavg < p->target_min && p->max_P >= MAX_P_MIN)
    460		p->max_P = (p->max_P/10)*9; /* maxp = maxp * Beta */
    461
    462	max_p_delta = DIV_ROUND_CLOSEST(p->max_P, p->qth_delta);
    463	max_p_delta = max(max_p_delta, 1U);
    464	p->max_P_reciprocal = reciprocal_value(max_p_delta);
    465}
    466#endif