1 // SPDX-License-Identifier: GPL-2.0-only OR BSD-2-Clause
2 /* Copyright (C) 2024 Nokia
3 *
4 * Author: Koen De Schepper <koen.de_schepper@nokia-bell-labs.com>
5 * Author: Olga Albisser <olga@albisser.org>
6 * Author: Henrik Steen <henrist@henrist.net>
7 * Author: Olivier Tilmans <olivier.tilmans@nokia.com>
8 * Author: Chia-Yu Chang <chia-yu.chang@nokia-bell-labs.com>
9 *
10 * DualPI Improved with a Square (dualpi2):
11 * - Supports congestion controls that comply with the Prague requirements
12 * in RFC9331 (e.g. TCP-Prague)
13 * - Supports coupled dual-queue with PI2 as defined in RFC9332
14 * - Supports ECN L4S-identifier (IP.ECN==0b*1)
15 *
16 * note: Although DCTCP and BBRv3 can use shallow-threshold ECN marks,
17 * they do not meet the 'Prague L4S Requirements' listed in RFC 9331
18 * Section 4, so they can only be used with DualPI2 in a datacenter
19 * context.
20 *
21 * References:
22 * - RFC9332: https://datatracker.ietf.org/doc/html/rfc9332
23 * - De Schepper, Koen, et al. "PI 2: A linearized AQM for both classic and
24 * scalable TCP." in proc. ACM CoNEXT'16, 2016.
25 */
26
27 #include <linux/errno.h>
28 #include <linux/hrtimer.h>
29 #include <linux/if_vlan.h>
30 #include <linux/kernel.h>
31 #include <linux/limits.h>
32 #include <linux/module.h>
33 #include <linux/skbuff.h>
34 #include <linux/types.h>
35
36 #include <net/gso.h>
37 #include <net/inet_ecn.h>
38 #include <net/pkt_cls.h>
39 #include <net/pkt_sched.h>
40
41 /* 32b enable to support flows with windows up to ~8.6 * 1e9 packets
42 * i.e., twice the maximal snd_cwnd.
43 * MAX_PROB must be consistent with the RNG in dualpi2_roll().
44 */
45 #define MAX_PROB U32_MAX
46
47 /* alpha/beta values exchanged over netlink are in units of 256ns */
48 #define ALPHA_BETA_SHIFT 8
49
50 /* Scaled values of alpha/beta must fit in 32b to avoid overflow in later
51 * computations. Consequently (see and dualpi2_scale_alpha_beta()), their
52 * netlink-provided values can use at most 31b, i.e. be at most (2^23)-1
53 * (~4MHz) as those are given in 1/256th. This enable to tune alpha/beta to
54 * control flows whose maximal RTTs can be in usec up to few secs.
55 */
56 #define ALPHA_BETA_MAX ((1U << 31) - 1)
57
58 /* Internal alpha/beta are in units of 64ns.
59 * This enables to use all alpha/beta values in the allowed range without loss
60 * of precision due to rounding when scaling them internally, e.g.,
61 * scale_alpha_beta(1) will not round down to 0.
62 */
63 #define ALPHA_BETA_GRANULARITY 6
64
65 #define ALPHA_BETA_SCALING (ALPHA_BETA_SHIFT - ALPHA_BETA_GRANULARITY)
66
67 /* We express the weights (wc, wl) in %, i.e., wc + wl = 100 */
68 #define MAX_WC 100
69
70 struct dualpi2_sched_data {
71 struct Qdisc *l_queue; /* The L4S Low latency queue (L-queue) */
72 struct Qdisc *sch; /* The Classic queue (C-queue) */
73
74 /* Registered tc filters */
75 struct tcf_proto __rcu *tcf_filters;
76 struct tcf_block *tcf_block;
77
78 /* PI2 parameters */
79 u64 pi2_target; /* Target delay in nanoseconds */
80 u32 pi2_tupdate; /* Timer frequency in nanoseconds */
81 u32 pi2_prob; /* Base PI probability */
82 u32 pi2_alpha; /* Gain factor for the integral rate response */
83 u32 pi2_beta; /* Gain factor for the proportional response */
84 struct hrtimer pi2_timer; /* prob update timer */
85
86 /* Step AQM (L-queue only) parameters */
87 u32 step_thresh; /* Step threshold */
88 bool step_in_packets; /* Step thresh in packets (1) or time (0) */
89
90 /* C-queue starvation protection */
91 s32 c_protection_credit; /* Credit (sign indicates which queue) */
92 s32 c_protection_init; /* Reset value of the credit */
93 u8 c_protection_wc; /* C-queue weight (between 0 and MAX_WC) */
94 u8 c_protection_wl; /* L-queue weight (MAX_WC - wc) */
95
96 /* General dualQ parameters */
97 u32 memory_limit; /* Memory limit of both queues */
98 u8 coupling_factor;/* Coupling factor (k) between both queues */
99 u8 ecn_mask; /* Mask to match packets into L-queue */
100 u32 min_qlen_step; /* Minimum queue length to apply step thresh */
101 bool drop_early; /* Drop at enqueue (1) instead of dequeue (0) */
102 bool drop_overload; /* Drop (1) on overload, or overflow (0) */
103 bool split_gso; /* Split aggregated skb (1) or leave as is (0) */
104
105 /* Statistics */
106 u64 c_head_ts; /* Enqueue timestamp of the C-queue head */
107 u64 l_head_ts; /* Enqueue timestamp of the L-queue head */
108 u64 last_qdelay; /* Q delay val at the last probability update */
109 u32 packets_in_c; /* Enqueue packet counter of the C-queue */
110 u32 packets_in_l; /* Enqueue packet counter of the L-queue */
111 u32 maxq; /* Maximum queue size of the C-queue */
112 u32 ecn_mark; /* ECN mark pkt counter due to PI probability */
113 u32 step_marks; /* ECN mark pkt counter due to step AQM */
114 u32 memory_used; /* Memory used of both queues */
115 u32 max_memory_used;/* Maximum used memory */
116
117 /* Deferred drop statistics */
118 u32 deferred_drops_cnt; /* Packets dropped */
119 u32 deferred_drops_len; /* Bytes dropped */
120 };
121
122 struct dualpi2_skb_cb {
123 u64 ts; /* Timestamp at enqueue */
124 u8 apply_step:1, /* Can we apply the step threshold */
125 classified:2, /* Packet classification results */
126 ect:2; /* Packet ECT codepoint */
127 };
128
129 enum dualpi2_classification_results {
130 DUALPI2_C_CLASSIC = 0, /* C-queue */
131 DUALPI2_C_L4S = 1, /* L-queue (scale mark/classic drop) */
132 DUALPI2_C_LLLL = 2, /* L-queue (no drops/marks) */
133 __DUALPI2_C_MAX /* Keep last*/
134 };
135
dualpi2_skb_cb(struct sk_buff * skb)136 static struct dualpi2_skb_cb *dualpi2_skb_cb(struct sk_buff *skb)
137 {
138 qdisc_cb_private_validate(skb, sizeof(struct dualpi2_skb_cb));
139 return (struct dualpi2_skb_cb *)qdisc_skb_cb(skb)->data;
140 }
141
dualpi2_sojourn_time(struct sk_buff * skb,u64 reference)142 static u64 dualpi2_sojourn_time(struct sk_buff *skb, u64 reference)
143 {
144 return reference - dualpi2_skb_cb(skb)->ts;
145 }
146
head_enqueue_time(struct Qdisc * q)147 static u64 head_enqueue_time(struct Qdisc *q)
148 {
149 struct sk_buff *skb = qdisc_peek_head(q);
150
151 return skb ? dualpi2_skb_cb(skb)->ts : 0;
152 }
153
dualpi2_scale_alpha_beta(u32 param)154 static u32 dualpi2_scale_alpha_beta(u32 param)
155 {
156 u64 tmp = ((u64)param * MAX_PROB >> ALPHA_BETA_SCALING);
157
158 do_div(tmp, NSEC_PER_SEC);
159 return tmp;
160 }
161
dualpi2_unscale_alpha_beta(u32 param)162 static u32 dualpi2_unscale_alpha_beta(u32 param)
163 {
164 u64 tmp = ((u64)param * NSEC_PER_SEC << ALPHA_BETA_SCALING);
165
166 do_div(tmp, MAX_PROB);
167 return tmp;
168 }
169
next_pi2_timeout(struct dualpi2_sched_data * q)170 static ktime_t next_pi2_timeout(struct dualpi2_sched_data *q)
171 {
172 return ktime_add_ns(ktime_get_ns(), q->pi2_tupdate);
173 }
174
skb_is_l4s(struct sk_buff * skb)175 static bool skb_is_l4s(struct sk_buff *skb)
176 {
177 return dualpi2_skb_cb(skb)->classified == DUALPI2_C_L4S;
178 }
179
skb_in_l_queue(struct sk_buff * skb)180 static bool skb_in_l_queue(struct sk_buff *skb)
181 {
182 return dualpi2_skb_cb(skb)->classified != DUALPI2_C_CLASSIC;
183 }
184
skb_apply_step(struct sk_buff * skb,struct dualpi2_sched_data * q)185 static bool skb_apply_step(struct sk_buff *skb, struct dualpi2_sched_data *q)
186 {
187 return skb_is_l4s(skb) && qdisc_qlen(q->l_queue) >= q->min_qlen_step;
188 }
189
dualpi2_mark(struct dualpi2_sched_data * q,struct sk_buff * skb)190 static bool dualpi2_mark(struct dualpi2_sched_data *q, struct sk_buff *skb)
191 {
192 if (INET_ECN_set_ce(skb)) {
193 WRITE_ONCE(q->ecn_mark, q->ecn_mark + 1);
194 return true;
195 }
196 return false;
197 }
198
dualpi2_reset_c_protection(struct dualpi2_sched_data * q)199 static void dualpi2_reset_c_protection(struct dualpi2_sched_data *q)
200 {
201 WRITE_ONCE(q->c_protection_credit, q->c_protection_init);
202 }
203
204 /* This computes the initial credit value and WRR weight for the L queue (wl)
205 * from the weight of the C queue (wc).
206 * If wl > wc, the scheduler will start with the L queue when reset.
207 */
dualpi2_calculate_c_protection(struct Qdisc * sch,struct dualpi2_sched_data * q,u32 wc)208 static void dualpi2_calculate_c_protection(struct Qdisc *sch,
209 struct dualpi2_sched_data *q, u32 wc)
210 {
211 q->c_protection_wc = wc;
212 q->c_protection_wl = MAX_WC - wc;
213 q->c_protection_init = (s32)psched_mtu(qdisc_dev(sch)) *
214 ((int)q->c_protection_wc - (int)q->c_protection_wl);
215 dualpi2_reset_c_protection(q);
216 }
217
dualpi2_roll(u32 prob)218 static bool dualpi2_roll(u32 prob)
219 {
220 return get_random_u32() <= prob;
221 }
222
223 /* Packets in the C-queue are subject to a marking probability pC, which is the
224 * square of the internal PI probability (i.e., have an overall lower mark/drop
225 * probability). If the qdisc is overloaded, ignore ECT values and only drop.
226 *
227 * Note that this marking scheme is also applied to L4S packets during overload.
228 * Return true if packet dropping is required in C queue
229 */
dualpi2_classic_marking(struct dualpi2_sched_data * q,struct sk_buff * skb,u32 prob,bool overload)230 static bool dualpi2_classic_marking(struct dualpi2_sched_data *q,
231 struct sk_buff *skb, u32 prob,
232 bool overload)
233 {
234 if (dualpi2_roll(prob) && dualpi2_roll(prob)) {
235 if (overload || dualpi2_skb_cb(skb)->ect == INET_ECN_NOT_ECT)
236 return true;
237 dualpi2_mark(q, skb);
238 }
239 return false;
240 }
241
242 /* Packets in the L-queue are subject to a marking probability pL given by the
243 * internal PI probability scaled by the coupling factor.
244 *
245 * On overload (i.e., @local_l_prob is >= 100%):
246 * - if the qdisc is configured to trade losses to preserve latency (i.e.,
247 * @q->drop_overload), apply classic drops first before marking.
248 * - otherwise, preserve the "no loss" property of ECN at the cost of queueing
249 * delay, eventually resulting in taildrop behavior once sch->limit is
250 * reached.
251 * Return true if packet dropping is required in L queue
252 */
dualpi2_scalable_marking(struct dualpi2_sched_data * q,struct sk_buff * skb,u64 local_l_prob,u32 prob,bool overload)253 static bool dualpi2_scalable_marking(struct dualpi2_sched_data *q,
254 struct sk_buff *skb,
255 u64 local_l_prob, u32 prob,
256 bool overload)
257 {
258 if (overload) {
259 /* Apply classic drop */
260 if (!q->drop_overload ||
261 !(dualpi2_roll(prob) && dualpi2_roll(prob)))
262 goto mark;
263 return true;
264 }
265
266 /* We can safely cut the upper 32b as overload==false */
267 if (dualpi2_roll(local_l_prob)) {
268 /* Non-ECT packets could have classified as L4S by filters. */
269 if (dualpi2_skb_cb(skb)->ect == INET_ECN_NOT_ECT)
270 return true;
271 mark:
272 dualpi2_mark(q, skb);
273 }
274 return false;
275 }
276
277 /* Decide whether a given packet must be dropped (or marked if ECT), according
278 * to the PI2 probability.
279 *
280 * Never mark/drop if we have a standing queue of less than 2 MTUs.
281 */
must_drop(struct Qdisc * sch,struct dualpi2_sched_data * q,struct sk_buff * skb)282 static bool must_drop(struct Qdisc *sch, struct dualpi2_sched_data *q,
283 struct sk_buff *skb)
284 {
285 u64 local_l_prob;
286 bool overload;
287 u32 prob;
288
289 if (sch->qstats.backlog < 2 * psched_mtu(qdisc_dev(sch)))
290 return false;
291
292 prob = READ_ONCE(q->pi2_prob);
293 local_l_prob = (u64)prob * q->coupling_factor;
294 overload = local_l_prob > MAX_PROB;
295
296 switch (dualpi2_skb_cb(skb)->classified) {
297 case DUALPI2_C_CLASSIC:
298 return dualpi2_classic_marking(q, skb, prob, overload);
299 case DUALPI2_C_L4S:
300 return dualpi2_scalable_marking(q, skb, local_l_prob, prob,
301 overload);
302 default: /* DUALPI2_C_LLLL */
303 return false;
304 }
305 }
306
dualpi2_read_ect(struct sk_buff * skb)307 static void dualpi2_read_ect(struct sk_buff *skb)
308 {
309 struct dualpi2_skb_cb *cb = dualpi2_skb_cb(skb);
310 int wlen = skb_network_offset(skb);
311
312 switch (skb_protocol(skb, true)) {
313 case htons(ETH_P_IP):
314 wlen += sizeof(struct iphdr);
315 if (!pskb_may_pull(skb, wlen) ||
316 skb_try_make_writable(skb, wlen))
317 goto not_ecn;
318
319 cb->ect = ipv4_get_dsfield(ip_hdr(skb)) & INET_ECN_MASK;
320 break;
321 case htons(ETH_P_IPV6):
322 wlen += sizeof(struct ipv6hdr);
323 if (!pskb_may_pull(skb, wlen) ||
324 skb_try_make_writable(skb, wlen))
325 goto not_ecn;
326
327 cb->ect = ipv6_get_dsfield(ipv6_hdr(skb)) & INET_ECN_MASK;
328 break;
329 default:
330 goto not_ecn;
331 }
332 return;
333
334 not_ecn:
335 /* Non pullable/writable packets can only be dropped hence are
336 * classified as not ECT.
337 */
338 cb->ect = INET_ECN_NOT_ECT;
339 }
340
dualpi2_skb_classify(struct dualpi2_sched_data * q,struct sk_buff * skb)341 static int dualpi2_skb_classify(struct dualpi2_sched_data *q,
342 struct sk_buff *skb)
343 {
344 struct dualpi2_skb_cb *cb = dualpi2_skb_cb(skb);
345 struct tcf_result res;
346 struct tcf_proto *fl;
347 int result;
348
349 cb->classified = DUALPI2_C_CLASSIC;
350
351 dualpi2_read_ect(skb);
352 if (cb->ect & q->ecn_mask) {
353 cb->classified = DUALPI2_C_L4S;
354 return NET_XMIT_SUCCESS;
355 }
356
357 if (TC_H_MAJ(skb->priority) == q->sch->handle &&
358 TC_H_MIN(skb->priority) < __DUALPI2_C_MAX) {
359 cb->classified = TC_H_MIN(skb->priority);
360 return NET_XMIT_SUCCESS;
361 }
362
363 fl = rcu_dereference_bh(q->tcf_filters);
364 if (!fl)
365 return NET_XMIT_SUCCESS;
366
367 result = tcf_classify_qdisc(skb, fl, &res, false);
368 if (result >= 0) {
369 #ifdef CONFIG_NET_CLS_ACT
370 switch (result) {
371 case TC_ACT_STOLEN:
372 case TC_ACT_QUEUED:
373 case TC_ACT_TRAP:
374 return NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
375 case TC_ACT_SHOT:
376 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
377 }
378 #endif
379 cb->classified = TC_H_MIN(res.classid) < __DUALPI2_C_MAX ?
380 TC_H_MIN(res.classid) : DUALPI2_C_CLASSIC;
381 }
382 return NET_XMIT_SUCCESS;
383 }
384
dualpi2_enqueue_skb(struct sk_buff * skb,struct Qdisc * sch,struct sk_buff ** to_free)385 static int dualpi2_enqueue_skb(struct sk_buff *skb, struct Qdisc *sch,
386 struct sk_buff **to_free)
387 {
388 struct dualpi2_sched_data *q = qdisc_priv(sch);
389 struct dualpi2_skb_cb *cb;
390
391 if (unlikely(qdisc_qlen(sch) >= sch->limit) ||
392 unlikely((u64)q->memory_used + skb->truesize > q->memory_limit)) {
393 qdisc_qstats_overlimit(sch);
394 if (skb_in_l_queue(skb))
395 qdisc_qstats_overlimit(q->l_queue);
396 return qdisc_drop_reason(skb, sch, to_free, QDISC_DROP_OVERLIMIT);
397 }
398
399 if (q->drop_early && must_drop(sch, q, skb)) {
400 qdisc_drop_reason(skb, sch, to_free, QDISC_DROP_CONGESTED);
401 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
402 }
403
404 cb = dualpi2_skb_cb(skb);
405 cb->ts = ktime_get_ns();
406 WRITE_ONCE(q->memory_used, q->memory_used + skb->truesize);
407 if (q->memory_used > q->max_memory_used)
408 WRITE_ONCE(q->max_memory_used, q->memory_used);
409
410 if (qdisc_qlen(sch) > q->maxq)
411 WRITE_ONCE(q->maxq, qdisc_qlen(sch));
412
413 if (skb_in_l_queue(skb)) {
414 /* Apply step thresh if skb is L4S && L-queue len >= min_qlen */
415 dualpi2_skb_cb(skb)->apply_step = skb_apply_step(skb, q);
416
417 /* Keep the overall qdisc stats consistent */
418 qdisc_qlen_inc(sch);
419 qdisc_qstats_backlog_inc(sch, skb);
420 WRITE_ONCE(q->packets_in_l, q->packets_in_l + 1);
421 if (!q->l_head_ts)
422 WRITE_ONCE(q->l_head_ts, cb->ts);
423 return qdisc_enqueue_tail(skb, q->l_queue);
424 }
425 WRITE_ONCE(q->packets_in_c, q->packets_in_c + 1);
426 if (!q->c_head_ts)
427 WRITE_ONCE(q->c_head_ts, cb->ts);
428 return qdisc_enqueue_tail(skb, sch);
429 }
430
431 /* By default, dualpi2 will split GSO skbs into independent skbs and enqueue
432 * each of those individually. This yields the following benefits, at the
433 * expense of CPU usage:
434 * - Finer-grained AQM actions as the sub-packets of a burst no longer share the
435 * same fate (e.g., the random mark/drop probability is applied individually)
436 * - Improved precision of the starvation protection/WRR scheduler at dequeue,
437 * as the size of the dequeued packets will be smaller.
438 */
dualpi2_qdisc_enqueue(struct sk_buff * skb,struct Qdisc * sch,struct sk_buff ** to_free)439 static int dualpi2_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
440 struct sk_buff **to_free)
441 {
442 struct dualpi2_sched_data *q = qdisc_priv(sch);
443 int err;
444
445 err = dualpi2_skb_classify(q, skb);
446 if (err != NET_XMIT_SUCCESS) {
447 if (err & __NET_XMIT_BYPASS)
448 qdisc_qstats_drop(sch);
449 __qdisc_drop(skb, to_free);
450 return err;
451 }
452
453 if (q->split_gso && skb_is_gso(skb)) {
454 netdev_features_t features;
455 struct sk_buff *nskb, *next;
456 int cnt, byte_len, orig_len;
457 int err;
458
459 features = netif_skb_features(skb);
460 nskb = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
461 if (IS_ERR_OR_NULL(nskb))
462 return qdisc_drop(skb, sch, to_free);
463
464 cnt = 0;
465 byte_len = 0;
466 orig_len = qdisc_pkt_len(skb);
467 skb_list_walk_safe(nskb, nskb, next) {
468 skb_mark_not_on_list(nskb);
469
470 /* Iterate through GSO fragments of an skb:
471 * (1) Set pkt_len from the single GSO fragments
472 * (2) Copy classified and ect values of an skb
473 * (3) Enqueue fragment & set ts in dualpi2_enqueue_skb
474 */
475 qdisc_skb_cb(nskb)->pkt_len = nskb->len;
476 qdisc_skb_cb(nskb)->pkt_segs = 1;
477 dualpi2_skb_cb(nskb)->classified =
478 dualpi2_skb_cb(skb)->classified;
479 dualpi2_skb_cb(nskb)->ect = dualpi2_skb_cb(skb)->ect;
480 err = dualpi2_enqueue_skb(nskb, sch, to_free);
481
482 if (err == NET_XMIT_SUCCESS) {
483 /* Compute the backlog adjustment that needs
484 * to be propagated in the qdisc tree to reflect
485 * all new skbs successfully enqueued.
486 */
487 ++cnt;
488 byte_len += nskb->len;
489 }
490 }
491 if (cnt > 0) {
492 /* The caller will add the original skb stats to its
493 * backlog, compensate this if any nskb is enqueued.
494 */
495 qdisc_tree_reduce_backlog(sch, 1 - cnt,
496 orig_len - byte_len);
497 }
498 consume_skb(skb);
499 return cnt > 0 ? NET_XMIT_SUCCESS : err;
500 }
501 return dualpi2_enqueue_skb(skb, sch, to_free);
502 }
503
504 /* Select the queue from which the next packet can be dequeued, ensuring that
505 * neither queue can starve the other with a WRR scheduler.
506 *
507 * The sign of the WRR credit determines the next queue, while the size of
508 * the dequeued packet determines the magnitude of the WRR credit change. If
509 * either queue is empty, the WRR credit is kept unchanged.
510 *
511 * As the dequeued packet can be dropped later, the caller has to perform the
512 * qdisc_bstats_update() calls.
513 */
dequeue_packet(struct Qdisc * sch,struct dualpi2_sched_data * q,int * credit_change,u64 now)514 static struct sk_buff *dequeue_packet(struct Qdisc *sch,
515 struct dualpi2_sched_data *q,
516 int *credit_change,
517 u64 now)
518 {
519 struct sk_buff *skb = NULL;
520 int c_len;
521
522 *credit_change = 0;
523 c_len = qdisc_qlen(sch) - qdisc_qlen(q->l_queue);
524 if (qdisc_qlen(q->l_queue) && (!c_len || q->c_protection_credit <= 0)) {
525 skb = __qdisc_dequeue_head(&q->l_queue->q);
526 WRITE_ONCE(q->l_head_ts, head_enqueue_time(q->l_queue));
527 if (c_len)
528 *credit_change = q->c_protection_wc;
529 qdisc_qstats_backlog_dec(q->l_queue, skb);
530
531 /* Keep the global queue size consistent */
532 qdisc_qlen_dec(sch);
533 } else if (c_len) {
534 skb = __qdisc_dequeue_head(&sch->q);
535 WRITE_ONCE(q->c_head_ts, head_enqueue_time(sch));
536 if (qdisc_qlen(q->l_queue))
537 *credit_change = ~((s32)q->c_protection_wl) + 1;
538 } else {
539 dualpi2_reset_c_protection(q);
540 return NULL;
541 }
542 WRITE_ONCE(q->memory_used, q->memory_used - skb->truesize);
543 *credit_change *= qdisc_pkt_len(skb);
544 qdisc_qstats_backlog_dec(sch, skb);
545 return skb;
546 }
547
do_step_aqm(struct dualpi2_sched_data * q,struct sk_buff * skb,u64 now)548 static int do_step_aqm(struct dualpi2_sched_data *q, struct sk_buff *skb,
549 u64 now)
550 {
551 u64 qdelay = 0;
552
553 if (q->step_in_packets)
554 qdelay = qdisc_qlen(q->l_queue);
555 else
556 qdelay = dualpi2_sojourn_time(skb, now);
557
558 if (dualpi2_skb_cb(skb)->apply_step && qdelay > q->step_thresh) {
559 if (!dualpi2_skb_cb(skb)->ect) {
560 /* Drop this non-ECT packet */
561 return 1;
562 }
563
564 if (dualpi2_mark(q, skb))
565 WRITE_ONCE(q->step_marks, q->step_marks + 1);
566 }
567 qdisc_bstats_update(q->l_queue, skb);
568 return 0;
569 }
570
drop_and_retry(struct dualpi2_sched_data * q,struct sk_buff * skb,struct Qdisc * sch,enum qdisc_drop_reason reason)571 static void drop_and_retry(struct dualpi2_sched_data *q, struct sk_buff *skb,
572 struct Qdisc *sch, enum qdisc_drop_reason reason)
573 {
574 ++q->deferred_drops_cnt;
575 q->deferred_drops_len += qdisc_pkt_len(skb);
576 qdisc_dequeue_drop(sch, skb, reason);
577 qdisc_qstats_drop(sch);
578 }
579
__dualpi2_qdisc_dequeue(struct Qdisc * sch)580 static struct sk_buff *__dualpi2_qdisc_dequeue(struct Qdisc *sch)
581 {
582 struct dualpi2_sched_data *q = qdisc_priv(sch);
583 struct sk_buff *skb;
584 int credit_change;
585 u64 now;
586
587 now = ktime_get_ns();
588
589 while ((skb = dequeue_packet(sch, q, &credit_change, now))) {
590 if (!q->drop_early && must_drop(sch, q, skb)) {
591 drop_and_retry(q, skb, sch, QDISC_DROP_CONGESTED);
592 continue;
593 }
594
595 if (skb_in_l_queue(skb) && do_step_aqm(q, skb, now)) {
596 qdisc_qstats_drop(q->l_queue);
597 drop_and_retry(q, skb, sch, QDISC_DROP_L4S_STEP_NON_ECN);
598 continue;
599 }
600
601 WRITE_ONCE(q->c_protection_credit,
602 q->c_protection_credit + credit_change);
603 qdisc_bstats_update(sch, skb);
604 break;
605 }
606
607 return skb;
608 }
609
dualpi2_dequeue_drop(struct Qdisc * sch)610 static void dualpi2_dequeue_drop(struct Qdisc *sch)
611 {
612 struct dualpi2_sched_data *q = qdisc_priv(sch);
613
614 if (q->deferred_drops_cnt) {
615 qdisc_tree_reduce_backlog(sch, q->deferred_drops_cnt,
616 q->deferred_drops_len);
617 q->deferred_drops_cnt = 0;
618 q->deferred_drops_len = 0;
619 }
620 }
621
dualpi2_qdisc_dequeue(struct Qdisc * sch)622 static struct sk_buff *dualpi2_qdisc_dequeue(struct Qdisc *sch)
623 {
624 struct sk_buff *skb;
625
626 skb = __dualpi2_qdisc_dequeue(sch);
627
628 dualpi2_dequeue_drop(sch);
629
630 return skb;
631 }
632
dualpi2_peek(struct Qdisc * sch)633 static struct sk_buff *dualpi2_peek(struct Qdisc *sch)
634 {
635 struct sk_buff *skb = skb_peek(&sch->gso_skb);
636
637 if (!skb) {
638 skb = __dualpi2_qdisc_dequeue(sch);
639
640 if (skb) {
641 __skb_queue_head(&sch->gso_skb, skb);
642 /* it's still part of the queue */
643 qdisc_qstats_backlog_inc(sch, skb);
644 sch->q.qlen++;
645 }
646
647 dualpi2_dequeue_drop(sch);
648 }
649
650 return skb;
651 }
652
__scale_delta(u64 diff)653 static s64 __scale_delta(u64 diff)
654 {
655 do_div(diff, 1 << ALPHA_BETA_GRANULARITY);
656 return diff;
657 }
658
get_queue_delays(struct dualpi2_sched_data * q,u64 * qdelay_c,u64 * qdelay_l)659 static void get_queue_delays(struct dualpi2_sched_data *q, u64 *qdelay_c,
660 u64 *qdelay_l)
661 {
662 u64 now, qc, ql;
663
664 now = ktime_get_ns();
665 qc = READ_ONCE(q->c_head_ts);
666 ql = READ_ONCE(q->l_head_ts);
667
668 *qdelay_c = qc ? now - qc : 0;
669 *qdelay_l = ql ? now - ql : 0;
670 }
671
calculate_probability(struct Qdisc * sch)672 static u32 calculate_probability(struct Qdisc *sch)
673 {
674 struct dualpi2_sched_data *q = qdisc_priv(sch);
675 u32 new_prob;
676 u64 qdelay_c;
677 u64 qdelay_l;
678 u64 qdelay;
679 s64 delta;
680
681 get_queue_delays(q, &qdelay_c, &qdelay_l);
682 qdelay = max(qdelay_l, qdelay_c);
683
684 /* Alpha and beta take at most 32b, i.e, the delay difference would
685 * overflow for queuing delay differences > ~4.2sec.
686 */
687 delta = ((s64)qdelay - (s64)q->pi2_target) * q->pi2_alpha;
688 delta += ((s64)qdelay - (s64)q->last_qdelay) * q->pi2_beta;
689 q->last_qdelay = qdelay;
690
691 /* Bound new_prob between 0 and MAX_PROB */
692 if (delta > 0) {
693 new_prob = __scale_delta(delta) + q->pi2_prob;
694 if (new_prob < q->pi2_prob)
695 new_prob = MAX_PROB;
696 } else {
697 new_prob = q->pi2_prob - __scale_delta(~delta + 1);
698 if (new_prob > q->pi2_prob)
699 new_prob = 0;
700 }
701
702 /* If we do not drop on overload, ensure we cap the L4S probability to
703 * 100% to keep window fairness when overflowing.
704 */
705 if (!q->drop_overload)
706 return min_t(u32, new_prob, MAX_PROB / q->coupling_factor);
707 return new_prob;
708 }
709
get_memory_limit(struct Qdisc * sch,u32 limit)710 static u32 get_memory_limit(struct Qdisc *sch, u32 limit)
711 {
712 /* Apply rule of thumb, i.e., doubling the packet length,
713 * to further include per packet overhead in memory_limit.
714 */
715 u64 memlim = mul_u32_u32(limit, 2 * psched_mtu(qdisc_dev(sch)));
716
717 if (upper_32_bits(memlim))
718 return U32_MAX;
719 else
720 return lower_32_bits(memlim);
721 }
722
convert_us_to_nsec(u32 us)723 static u32 convert_us_to_nsec(u32 us)
724 {
725 u64 ns = mul_u32_u32(us, NSEC_PER_USEC);
726
727 if (upper_32_bits(ns))
728 return U32_MAX;
729
730 return lower_32_bits(ns);
731 }
732
convert_ns_to_usec(u64 ns)733 static u32 convert_ns_to_usec(u64 ns)
734 {
735 do_div(ns, NSEC_PER_USEC);
736 if (upper_32_bits(ns))
737 return U32_MAX;
738
739 return lower_32_bits(ns);
740 }
741
dualpi2_timer(struct hrtimer * timer)742 static enum hrtimer_restart dualpi2_timer(struct hrtimer *timer)
743 {
744 struct dualpi2_sched_data *q = timer_container_of(q, timer, pi2_timer);
745 struct Qdisc *sch = q->sch;
746 spinlock_t *root_lock; /* to lock qdisc for probability calculations */
747
748 rcu_read_lock();
749 root_lock = qdisc_lock(qdisc_root_sleeping(sch));
750 spin_lock(root_lock);
751
752 WRITE_ONCE(q->pi2_prob, calculate_probability(sch));
753 hrtimer_set_expires(&q->pi2_timer, next_pi2_timeout(q));
754
755 spin_unlock(root_lock);
756 rcu_read_unlock();
757 return HRTIMER_RESTART;
758 }
759
760 static struct netlink_range_validation dualpi2_alpha_beta_range = {
761 .min = 1,
762 .max = ALPHA_BETA_MAX,
763 };
764
765 static const struct nla_policy dualpi2_policy[TCA_DUALPI2_MAX + 1] = {
766 [TCA_DUALPI2_LIMIT] = NLA_POLICY_MIN(NLA_U32, 1),
767 [TCA_DUALPI2_MEMORY_LIMIT] = NLA_POLICY_MIN(NLA_U32, 1),
768 [TCA_DUALPI2_TARGET] = { .type = NLA_U32 },
769 [TCA_DUALPI2_TUPDATE] = NLA_POLICY_MIN(NLA_U32, 1),
770 [TCA_DUALPI2_ALPHA] =
771 NLA_POLICY_FULL_RANGE(NLA_U32, &dualpi2_alpha_beta_range),
772 [TCA_DUALPI2_BETA] =
773 NLA_POLICY_FULL_RANGE(NLA_U32, &dualpi2_alpha_beta_range),
774 [TCA_DUALPI2_STEP_THRESH_PKTS] = { .type = NLA_U32 },
775 [TCA_DUALPI2_STEP_THRESH_US] = { .type = NLA_U32 },
776 [TCA_DUALPI2_MIN_QLEN_STEP] = { .type = NLA_U32 },
777 [TCA_DUALPI2_COUPLING] = NLA_POLICY_MIN(NLA_U8, 1),
778 [TCA_DUALPI2_DROP_OVERLOAD] =
779 NLA_POLICY_MAX(NLA_U8, TCA_DUALPI2_DROP_OVERLOAD_MAX),
780 [TCA_DUALPI2_DROP_EARLY] =
781 NLA_POLICY_MAX(NLA_U8, TCA_DUALPI2_DROP_EARLY_MAX),
782 [TCA_DUALPI2_C_PROTECTION] =
783 NLA_POLICY_RANGE(NLA_U8, 0, MAX_WC),
784 [TCA_DUALPI2_ECN_MASK] =
785 NLA_POLICY_RANGE(NLA_U8, TC_DUALPI2_ECN_MASK_L4S_ECT,
786 TCA_DUALPI2_ECN_MASK_MAX),
787 [TCA_DUALPI2_SPLIT_GSO] =
788 NLA_POLICY_MAX(NLA_U8, TCA_DUALPI2_SPLIT_GSO_MAX),
789 };
790
dualpi2_change(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)791 static int dualpi2_change(struct Qdisc *sch, struct nlattr *opt,
792 struct netlink_ext_ack *extack)
793 {
794 struct nlattr *tb[TCA_DUALPI2_MAX + 1];
795 struct dualpi2_sched_data *q;
796 int old_backlog;
797 int old_qlen;
798 int err;
799
800 if (!opt || !nla_len(opt)) {
801 NL_SET_ERR_MSG_MOD(extack, "Dualpi2 options are required");
802 return -EINVAL;
803 }
804 err = nla_parse_nested(tb, TCA_DUALPI2_MAX, opt, dualpi2_policy,
805 extack);
806 if (err < 0)
807 return err;
808 if (tb[TCA_DUALPI2_STEP_THRESH_PKTS] && tb[TCA_DUALPI2_STEP_THRESH_US]) {
809 NL_SET_ERR_MSG_MOD(extack, "multiple step thresh attributes");
810 return -EINVAL;
811 }
812
813 q = qdisc_priv(sch);
814 sch_tree_lock(sch);
815
816 if (tb[TCA_DUALPI2_LIMIT]) {
817 u32 limit = nla_get_u32(tb[TCA_DUALPI2_LIMIT]);
818
819 WRITE_ONCE(sch->limit, limit);
820 WRITE_ONCE(q->memory_limit, get_memory_limit(sch, limit));
821 }
822
823 if (tb[TCA_DUALPI2_MEMORY_LIMIT])
824 WRITE_ONCE(q->memory_limit,
825 nla_get_u32(tb[TCA_DUALPI2_MEMORY_LIMIT]));
826
827 if (tb[TCA_DUALPI2_TARGET]) {
828 u64 target = nla_get_u32(tb[TCA_DUALPI2_TARGET]);
829
830 WRITE_ONCE(q->pi2_target, target * NSEC_PER_USEC);
831 }
832
833 if (tb[TCA_DUALPI2_TUPDATE]) {
834 u64 tupdate = nla_get_u32(tb[TCA_DUALPI2_TUPDATE]);
835
836 WRITE_ONCE(q->pi2_tupdate, convert_us_to_nsec(tupdate));
837 }
838
839 if (tb[TCA_DUALPI2_ALPHA]) {
840 u32 alpha = nla_get_u32(tb[TCA_DUALPI2_ALPHA]);
841
842 WRITE_ONCE(q->pi2_alpha, dualpi2_scale_alpha_beta(alpha));
843 }
844
845 if (tb[TCA_DUALPI2_BETA]) {
846 u32 beta = nla_get_u32(tb[TCA_DUALPI2_BETA]);
847
848 WRITE_ONCE(q->pi2_beta, dualpi2_scale_alpha_beta(beta));
849 }
850
851 if (tb[TCA_DUALPI2_STEP_THRESH_PKTS]) {
852 u32 step_th = nla_get_u32(tb[TCA_DUALPI2_STEP_THRESH_PKTS]);
853
854 WRITE_ONCE(q->step_in_packets, true);
855 WRITE_ONCE(q->step_thresh, step_th);
856 } else if (tb[TCA_DUALPI2_STEP_THRESH_US]) {
857 u32 step_th = nla_get_u32(tb[TCA_DUALPI2_STEP_THRESH_US]);
858
859 WRITE_ONCE(q->step_in_packets, false);
860 WRITE_ONCE(q->step_thresh, convert_us_to_nsec(step_th));
861 }
862
863 if (tb[TCA_DUALPI2_MIN_QLEN_STEP])
864 WRITE_ONCE(q->min_qlen_step,
865 nla_get_u32(tb[TCA_DUALPI2_MIN_QLEN_STEP]));
866
867 if (tb[TCA_DUALPI2_COUPLING]) {
868 u8 coupling = nla_get_u8(tb[TCA_DUALPI2_COUPLING]);
869
870 WRITE_ONCE(q->coupling_factor, coupling);
871 }
872
873 if (tb[TCA_DUALPI2_DROP_OVERLOAD]) {
874 u8 drop_overload = nla_get_u8(tb[TCA_DUALPI2_DROP_OVERLOAD]);
875
876 WRITE_ONCE(q->drop_overload, (bool)drop_overload);
877 }
878
879 if (tb[TCA_DUALPI2_DROP_EARLY]) {
880 u8 drop_early = nla_get_u8(tb[TCA_DUALPI2_DROP_EARLY]);
881
882 WRITE_ONCE(q->drop_early, (bool)drop_early);
883 }
884
885 if (tb[TCA_DUALPI2_C_PROTECTION]) {
886 u8 wc = nla_get_u8(tb[TCA_DUALPI2_C_PROTECTION]);
887
888 dualpi2_calculate_c_protection(sch, q, wc);
889 }
890
891 if (tb[TCA_DUALPI2_ECN_MASK]) {
892 u8 ecn_mask = nla_get_u8(tb[TCA_DUALPI2_ECN_MASK]);
893
894 WRITE_ONCE(q->ecn_mask, ecn_mask);
895 }
896
897 if (tb[TCA_DUALPI2_SPLIT_GSO]) {
898 u8 split_gso = nla_get_u8(tb[TCA_DUALPI2_SPLIT_GSO]);
899
900 WRITE_ONCE(q->split_gso, (bool)split_gso);
901 }
902
903 old_qlen = qdisc_qlen(sch);
904 old_backlog = sch->qstats.backlog;
905 while (qdisc_qlen(sch) > sch->limit ||
906 q->memory_used > q->memory_limit) {
907 struct sk_buff *skb = NULL;
908
909 if (qdisc_qlen(sch) > qdisc_qlen(q->l_queue)) {
910 skb = qdisc_dequeue_internal(sch, true);
911 if (unlikely(!skb)) {
912 WARN_ON_ONCE(1);
913 break;
914 }
915 WRITE_ONCE(q->memory_used, q->memory_used - skb->truesize);
916 rtnl_qdisc_drop(skb, sch);
917 } else if (qdisc_qlen(q->l_queue)) {
918 skb = qdisc_dequeue_internal(q->l_queue, true);
919 if (unlikely(!skb)) {
920 WARN_ON_ONCE(1);
921 break;
922 }
923 /* L-queue packets are counted in both sch and
924 * l_queue on enqueue; qdisc_dequeue_internal()
925 * handled l_queue, so we further account for sch.
926 */
927 qdisc_qlen_dec(sch);
928 qdisc_qstats_backlog_dec(sch, skb);
929 WRITE_ONCE(q->memory_used, q->memory_used - skb->truesize);
930 rtnl_qdisc_drop(skb, q->l_queue);
931 qdisc_qstats_drop(sch);
932 } else {
933 WARN_ON_ONCE(1);
934 break;
935 }
936 }
937 qdisc_tree_reduce_backlog(sch, old_qlen - qdisc_qlen(sch),
938 old_backlog - sch->qstats.backlog);
939
940 sch_tree_unlock(sch);
941 return 0;
942 }
943
944 /* Default alpha/beta values give a 10dB stability margin with max_rtt=100ms. */
dualpi2_reset_default(struct Qdisc * sch)945 static void dualpi2_reset_default(struct Qdisc *sch)
946 {
947 struct dualpi2_sched_data *q = qdisc_priv(sch);
948
949 q->sch->limit = 10000; /* Max 125ms at 1Gbps */
950 q->memory_limit = get_memory_limit(sch, q->sch->limit);
951
952 q->pi2_target = 15 * NSEC_PER_MSEC;
953 q->pi2_tupdate = 16 * NSEC_PER_MSEC;
954 q->pi2_alpha = dualpi2_scale_alpha_beta(41); /* ~0.16 Hz * 256 */
955 q->pi2_beta = dualpi2_scale_alpha_beta(819); /* ~3.20 Hz * 256 */
956
957 q->step_thresh = 1 * NSEC_PER_MSEC;
958 q->step_in_packets = false;
959
960 dualpi2_calculate_c_protection(q->sch, q, 10); /* wc=10%, wl=90% */
961
962 q->ecn_mask = TC_DUALPI2_ECN_MASK_L4S_ECT; /* INET_ECN_ECT_1 */
963 q->min_qlen_step = 0; /* Always apply step mark in L-queue */
964 q->coupling_factor = 2; /* window fairness for equal RTTs */
965 q->drop_overload = TC_DUALPI2_DROP_OVERLOAD_DROP; /* Drop overload */
966 q->drop_early = TC_DUALPI2_DROP_EARLY_DROP_DEQUEUE; /* Drop dequeue */
967 q->split_gso = TC_DUALPI2_SPLIT_GSO_SPLIT_GSO; /* Split GSO */
968 }
969
dualpi2_init(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)970 static int dualpi2_init(struct Qdisc *sch, struct nlattr *opt,
971 struct netlink_ext_ack *extack)
972 {
973 struct dualpi2_sched_data *q = qdisc_priv(sch);
974 int err;
975
976 sch->flags |= TCQ_F_DEQUEUE_DROPS;
977 hrtimer_setup(&q->pi2_timer, dualpi2_timer, CLOCK_MONOTONIC,
978 HRTIMER_MODE_ABS_PINNED_SOFT);
979
980 q->l_queue = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
981 TC_H_MAKE(sch->handle, 1), extack);
982 if (!q->l_queue)
983 return -ENOMEM;
984
985 err = tcf_block_get(&q->tcf_block, &q->tcf_filters, sch, extack);
986 if (err)
987 return err;
988
989 q->sch = sch;
990 dualpi2_reset_default(sch);
991
992 if (opt && nla_len(opt)) {
993 err = dualpi2_change(sch, opt, extack);
994
995 if (err)
996 return err;
997 }
998
999 hrtimer_start(&q->pi2_timer, next_pi2_timeout(q),
1000 HRTIMER_MODE_ABS_PINNED_SOFT);
1001 return 0;
1002 }
1003
dualpi2_dump(struct Qdisc * sch,struct sk_buff * skb)1004 static int dualpi2_dump(struct Qdisc *sch, struct sk_buff *skb)
1005 {
1006 struct dualpi2_sched_data *q = qdisc_priv(sch);
1007 struct nlattr *opts;
1008 bool step_in_pkts;
1009 u32 step_th;
1010
1011 step_in_pkts = READ_ONCE(q->step_in_packets);
1012 step_th = READ_ONCE(q->step_thresh);
1013
1014 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
1015 if (!opts)
1016 goto nla_put_failure;
1017
1018 if (step_in_pkts &&
1019 (nla_put_u32(skb, TCA_DUALPI2_LIMIT, READ_ONCE(sch->limit)) ||
1020 nla_put_u32(skb, TCA_DUALPI2_MEMORY_LIMIT,
1021 READ_ONCE(q->memory_limit)) ||
1022 nla_put_u32(skb, TCA_DUALPI2_TARGET,
1023 convert_ns_to_usec(READ_ONCE(q->pi2_target))) ||
1024 nla_put_u32(skb, TCA_DUALPI2_TUPDATE,
1025 convert_ns_to_usec(READ_ONCE(q->pi2_tupdate))) ||
1026 nla_put_u32(skb, TCA_DUALPI2_ALPHA,
1027 dualpi2_unscale_alpha_beta(READ_ONCE(q->pi2_alpha))) ||
1028 nla_put_u32(skb, TCA_DUALPI2_BETA,
1029 dualpi2_unscale_alpha_beta(READ_ONCE(q->pi2_beta))) ||
1030 nla_put_u32(skb, TCA_DUALPI2_STEP_THRESH_PKTS, step_th) ||
1031 nla_put_u32(skb, TCA_DUALPI2_MIN_QLEN_STEP,
1032 READ_ONCE(q->min_qlen_step)) ||
1033 nla_put_u8(skb, TCA_DUALPI2_COUPLING,
1034 READ_ONCE(q->coupling_factor)) ||
1035 nla_put_u8(skb, TCA_DUALPI2_DROP_OVERLOAD,
1036 READ_ONCE(q->drop_overload)) ||
1037 nla_put_u8(skb, TCA_DUALPI2_DROP_EARLY,
1038 READ_ONCE(q->drop_early)) ||
1039 nla_put_u8(skb, TCA_DUALPI2_C_PROTECTION,
1040 READ_ONCE(q->c_protection_wc)) ||
1041 nla_put_u8(skb, TCA_DUALPI2_ECN_MASK, READ_ONCE(q->ecn_mask)) ||
1042 nla_put_u8(skb, TCA_DUALPI2_SPLIT_GSO, READ_ONCE(q->split_gso))))
1043 goto nla_put_failure;
1044
1045 if (!step_in_pkts &&
1046 (nla_put_u32(skb, TCA_DUALPI2_LIMIT, READ_ONCE(sch->limit)) ||
1047 nla_put_u32(skb, TCA_DUALPI2_MEMORY_LIMIT,
1048 READ_ONCE(q->memory_limit)) ||
1049 nla_put_u32(skb, TCA_DUALPI2_TARGET,
1050 convert_ns_to_usec(READ_ONCE(q->pi2_target))) ||
1051 nla_put_u32(skb, TCA_DUALPI2_TUPDATE,
1052 convert_ns_to_usec(READ_ONCE(q->pi2_tupdate))) ||
1053 nla_put_u32(skb, TCA_DUALPI2_ALPHA,
1054 dualpi2_unscale_alpha_beta(READ_ONCE(q->pi2_alpha))) ||
1055 nla_put_u32(skb, TCA_DUALPI2_BETA,
1056 dualpi2_unscale_alpha_beta(READ_ONCE(q->pi2_beta))) ||
1057 nla_put_u32(skb, TCA_DUALPI2_STEP_THRESH_US,
1058 convert_ns_to_usec(step_th)) ||
1059 nla_put_u32(skb, TCA_DUALPI2_MIN_QLEN_STEP,
1060 READ_ONCE(q->min_qlen_step)) ||
1061 nla_put_u8(skb, TCA_DUALPI2_COUPLING,
1062 READ_ONCE(q->coupling_factor)) ||
1063 nla_put_u8(skb, TCA_DUALPI2_DROP_OVERLOAD,
1064 READ_ONCE(q->drop_overload)) ||
1065 nla_put_u8(skb, TCA_DUALPI2_DROP_EARLY,
1066 READ_ONCE(q->drop_early)) ||
1067 nla_put_u8(skb, TCA_DUALPI2_C_PROTECTION,
1068 READ_ONCE(q->c_protection_wc)) ||
1069 nla_put_u8(skb, TCA_DUALPI2_ECN_MASK, READ_ONCE(q->ecn_mask)) ||
1070 nla_put_u8(skb, TCA_DUALPI2_SPLIT_GSO, READ_ONCE(q->split_gso))))
1071 goto nla_put_failure;
1072
1073 return nla_nest_end(skb, opts);
1074
1075 nla_put_failure:
1076 nla_nest_cancel(skb, opts);
1077 return -1;
1078 }
1079
dualpi2_dump_stats(struct Qdisc * sch,struct gnet_dump * d)1080 static int dualpi2_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1081 {
1082 struct dualpi2_sched_data *q = qdisc_priv(sch);
1083 struct tc_dualpi2_xstats st = {
1084 .prob = READ_ONCE(q->pi2_prob),
1085 .packets_in_c = READ_ONCE(q->packets_in_c),
1086 .packets_in_l = READ_ONCE(q->packets_in_l),
1087 .maxq = READ_ONCE(q->maxq),
1088 .ecn_mark = READ_ONCE(q->ecn_mark),
1089 .credit = READ_ONCE(q->c_protection_credit),
1090 .step_marks = READ_ONCE(q->step_marks),
1091 .memory_used = READ_ONCE(q->memory_used),
1092 .max_memory_used = READ_ONCE(q->max_memory_used),
1093 .memory_limit = READ_ONCE(q->memory_limit),
1094 };
1095 u64 qc, ql;
1096
1097 get_queue_delays(q, &qc, &ql);
1098 st.delay_l = convert_ns_to_usec(ql);
1099 st.delay_c = convert_ns_to_usec(qc);
1100 return gnet_stats_copy_app(d, &st, sizeof(st));
1101 }
1102
1103 /* Reset both L-queue and C-queue, internal packet counters, PI probability,
1104 * C-queue protection credit, and timestamps, while preserving current
1105 * configuration of DUALPI2.
1106 */
dualpi2_reset(struct Qdisc * sch)1107 static void dualpi2_reset(struct Qdisc *sch)
1108 {
1109 struct dualpi2_sched_data *q = qdisc_priv(sch);
1110
1111 qdisc_reset_queue(sch);
1112 qdisc_reset_queue(q->l_queue);
1113 WRITE_ONCE(q->c_head_ts, 0);
1114 WRITE_ONCE(q->l_head_ts, 0);
1115 WRITE_ONCE(q->pi2_prob, 0);
1116 WRITE_ONCE(q->packets_in_c, 0);
1117 WRITE_ONCE(q->packets_in_l, 0);
1118 WRITE_ONCE(q->maxq, 0);
1119 WRITE_ONCE(q->ecn_mark, 0);
1120 WRITE_ONCE(q->step_marks, 0);
1121 WRITE_ONCE(q->memory_used, 0);
1122 WRITE_ONCE(q->max_memory_used, 0);
1123 dualpi2_reset_c_protection(q);
1124 }
1125
dualpi2_destroy(struct Qdisc * sch)1126 static void dualpi2_destroy(struct Qdisc *sch)
1127 {
1128 struct dualpi2_sched_data *q = qdisc_priv(sch);
1129
1130 q->pi2_tupdate = 0;
1131 hrtimer_cancel(&q->pi2_timer);
1132 if (q->l_queue)
1133 qdisc_put(q->l_queue);
1134 tcf_block_put(q->tcf_block);
1135 }
1136
dualpi2_leaf(struct Qdisc * sch,unsigned long arg)1137 static struct Qdisc *dualpi2_leaf(struct Qdisc *sch, unsigned long arg)
1138 {
1139 return NULL;
1140 }
1141
dualpi2_find(struct Qdisc * sch,u32 classid)1142 static unsigned long dualpi2_find(struct Qdisc *sch, u32 classid)
1143 {
1144 return 0;
1145 }
1146
dualpi2_bind(struct Qdisc * sch,unsigned long parent,u32 classid)1147 static unsigned long dualpi2_bind(struct Qdisc *sch, unsigned long parent,
1148 u32 classid)
1149 {
1150 return 0;
1151 }
1152
dualpi2_unbind(struct Qdisc * q,unsigned long cl)1153 static void dualpi2_unbind(struct Qdisc *q, unsigned long cl)
1154 {
1155 }
1156
dualpi2_tcf_block(struct Qdisc * sch,unsigned long cl,struct netlink_ext_ack * extack)1157 static struct tcf_block *dualpi2_tcf_block(struct Qdisc *sch, unsigned long cl,
1158 struct netlink_ext_ack *extack)
1159 {
1160 struct dualpi2_sched_data *q = qdisc_priv(sch);
1161
1162 if (cl)
1163 return NULL;
1164 return q->tcf_block;
1165 }
1166
dualpi2_walk(struct Qdisc * sch,struct qdisc_walker * arg)1167 static void dualpi2_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1168 {
1169 unsigned int i;
1170
1171 if (arg->stop)
1172 return;
1173
1174 /* We statically define only 2 queues */
1175 for (i = 0; i < 2; i++) {
1176 if (arg->count < arg->skip) {
1177 arg->count++;
1178 continue;
1179 }
1180 if (arg->fn(sch, i + 1, arg) < 0) {
1181 arg->stop = 1;
1182 break;
1183 }
1184 arg->count++;
1185 }
1186 }
1187
1188 /* Minimal class support to handle tc filters */
1189 static const struct Qdisc_class_ops dualpi2_class_ops = {
1190 .leaf = dualpi2_leaf,
1191 .find = dualpi2_find,
1192 .tcf_block = dualpi2_tcf_block,
1193 .bind_tcf = dualpi2_bind,
1194 .unbind_tcf = dualpi2_unbind,
1195 .walk = dualpi2_walk,
1196 };
1197
1198 static struct Qdisc_ops dualpi2_qdisc_ops __read_mostly = {
1199 .id = "dualpi2",
1200 .cl_ops = &dualpi2_class_ops,
1201 .priv_size = sizeof(struct dualpi2_sched_data),
1202 .enqueue = dualpi2_qdisc_enqueue,
1203 .dequeue = dualpi2_qdisc_dequeue,
1204 .peek = dualpi2_peek,
1205 .init = dualpi2_init,
1206 .destroy = dualpi2_destroy,
1207 .reset = dualpi2_reset,
1208 .change = dualpi2_change,
1209 .dump = dualpi2_dump,
1210 .dump_stats = dualpi2_dump_stats,
1211 .owner = THIS_MODULE,
1212 };
1213 MODULE_ALIAS_NET_SCH("dualpi2");
1214
dualpi2_module_init(void)1215 static int __init dualpi2_module_init(void)
1216 {
1217 return register_qdisc(&dualpi2_qdisc_ops);
1218 }
1219
dualpi2_module_exit(void)1220 static void __exit dualpi2_module_exit(void)
1221 {
1222 unregister_qdisc(&dualpi2_qdisc_ops);
1223 }
1224
1225 module_init(dualpi2_module_init);
1226 module_exit(dualpi2_module_exit);
1227
1228 MODULE_DESCRIPTION("Dual Queue with Proportional Integral controller Improved with a Square (dualpi2) scheduler");
1229 MODULE_AUTHOR("Koen De Schepper <koen.de_schepper@nokia-bell-labs.com>");
1230 MODULE_AUTHOR("Chia-Yu Chang <chia-yu.chang@nokia-bell-labs.com>");
1231 MODULE_AUTHOR("Olga Albisser <olga@albisser.org>");
1232 MODULE_AUTHOR("Henrik Steen <henrist@henrist.net>");
1233 MODULE_AUTHOR("Olivier Tilmans <olivier.tilmans@nokia.com>");
1234
1235 MODULE_LICENSE("Dual BSD/GPL");
1236 MODULE_VERSION("1.0");
1237