Commit | Line | Data |
---|---|---|
09c434b8 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
10239edf TL |
2 | /* net/sched/sch_hhf.c Heavy-Hitter Filter (HHF) |
3 | * | |
4 | * Copyright (C) 2013 Terry Lam <vtlam@google.com> | |
5 | * Copyright (C) 2013 Nandita Dukkipati <nanditad@google.com> | |
6 | */ | |
7 | ||
10239edf TL |
8 | #include <linux/jiffies.h> |
9 | #include <linux/module.h> | |
10 | #include <linux/skbuff.h> | |
11 | #include <linux/vmalloc.h> | |
55667441 | 12 | #include <linux/siphash.h> |
10239edf TL |
13 | #include <net/pkt_sched.h> |
14 | #include <net/sock.h> | |
15 | ||
16 | /* Heavy-Hitter Filter (HHF) | |
17 | * | |
18 | * Principles : | |
19 | * Flows are classified into two buckets: non-heavy-hitter and heavy-hitter | |
20 | * buckets. Initially, a new flow starts as non-heavy-hitter. Once classified | |
21 | * as heavy-hitter, it is immediately switched to the heavy-hitter bucket. | |
22 | * The buckets are dequeued by a Weighted Deficit Round Robin (WDRR) scheduler, | |
23 | * in which the heavy-hitter bucket is served with less weight. | |
24 | * In other words, non-heavy-hitters (e.g., short bursts of critical traffic) | |
25 | * are isolated from heavy-hitters (e.g., persistent bulk traffic) and also have | |
26 | * higher share of bandwidth. | |
27 | * | |
28 | * To capture heavy-hitters, we use the "multi-stage filter" algorithm in the | |
29 | * following paper: | |
30 | * [EV02] C. Estan and G. Varghese, "New Directions in Traffic Measurement and | |
31 | * Accounting", in ACM SIGCOMM, 2002. | |
32 | * | |
33 | * Conceptually, a multi-stage filter comprises k independent hash functions | |
34 | * and k counter arrays. Packets are indexed into k counter arrays by k hash | |
35 | * functions, respectively. The counters are then increased by the packet sizes. | |
36 | * Therefore, | |
37 | * - For a heavy-hitter flow: *all* of its k array counters must be large. | |
38 | * - For a non-heavy-hitter flow: some of its k array counters can be large | |
39 | * due to hash collision with other small flows; however, with high | |
40 | * probability, not *all* k counters are large. | |
41 | * | |
42 | * By the design of the multi-stage filter algorithm, the false negative rate | |
43 | * (heavy-hitters getting away uncaptured) is zero. However, the algorithm is | |
44 | * susceptible to false positives (non-heavy-hitters mistakenly classified as | |
45 | * heavy-hitters). | |
46 | * Therefore, we also implement the following optimizations to reduce false | |
47 | * positives by avoiding unnecessary increment of the counter values: | |
48 | * - Optimization O1: once a heavy-hitter is identified, its bytes are not | |
49 | * accounted in the array counters. This technique is called "shielding" | |
50 | * in Section 3.3.1 of [EV02]. | |
51 | * - Optimization O2: conservative update of counters | |
52 | * (Section 3.3.2 of [EV02]), | |
53 | * New counter value = max {old counter value, | |
54 | * smallest counter value + packet bytes} | |
55 | * | |
56 | * Finally, we refresh the counters periodically since otherwise the counter | |
57 | * values will keep accumulating. | |
58 | * | |
59 | * Once a flow is classified as heavy-hitter, we also save its per-flow state | |
60 | * in an exact-matching flow table so that its subsequent packets can be | |
61 | * dispatched to the heavy-hitter bucket accordingly. | |
62 | * | |
63 | * | |
64 | * At a high level, this qdisc works as follows: | |
65 | * Given a packet p: | |
66 | * - If the flow-id of p (e.g., TCP 5-tuple) is already in the exact-matching | |
67 | * heavy-hitter flow table, denoted table T, then send p to the heavy-hitter | |
68 | * bucket. | |
69 | * - Otherwise, forward p to the multi-stage filter, denoted filter F | |
70 | * + If F decides that p belongs to a non-heavy-hitter flow, then send p | |
71 | * to the non-heavy-hitter bucket. | |
72 | * + Otherwise, if F decides that p belongs to a new heavy-hitter flow, | |
73 | * then set up a new flow entry for the flow-id of p in the table T and | |
74 | * send p to the heavy-hitter bucket. | |
75 | * | |
76 | * In this implementation: | |
77 | * - T is a fixed-size hash-table with 1024 entries. Hash collision is | |
78 | * resolved by linked-list chaining. | |
79 | * - F has four counter arrays, each array containing 1024 32-bit counters. | |
80 | * That means 4 * 1024 * 32 bits = 16KB of memory. | |
81 | * - Since each array in F contains 1024 counters, 10 bits are sufficient to | |
82 | * index into each array. | |
83 | * Hence, instead of having four hash functions, we chop the 32-bit | |
84 | * skb-hash into three 10-bit chunks, and the remaining 10-bit chunk is | |
85 | * computed as XOR sum of those three chunks. | |
86 | * - We need to clear the counter arrays periodically; however, directly | |
87 | * memsetting 16KB of memory can lead to cache eviction and unwanted delay. | |
88 | * So by representing each counter by a valid bit, we only need to reset | |
89 | * 4K of 1 bit (i.e. 512 bytes) instead of 16KB of memory. | |
90 | * - The Deficit Round Robin engine is taken from fq_codel implementation | |
91 | * (net/sched/sch_fq_codel.c). Note that wdrr_bucket corresponds to | |
92 | * fq_codel_flow in fq_codel implementation. | |
93 | * | |
94 | */ | |
95 | ||
96 | /* Non-configurable parameters */ | |
97 | #define HH_FLOWS_CNT 1024 /* number of entries in exact-matching table T */ | |
98 | #define HHF_ARRAYS_CNT 4 /* number of arrays in multi-stage filter F */ | |
99 | #define HHF_ARRAYS_LEN 1024 /* number of counters in each array of F */ | |
100 | #define HHF_BIT_MASK_LEN 10 /* masking 10 bits */ | |
101 | #define HHF_BIT_MASK 0x3FF /* bitmask of 10 bits */ | |
102 | ||
103 | #define WDRR_BUCKET_CNT 2 /* two buckets for Weighted DRR */ | |
104 | enum wdrr_bucket_idx { | |
105 | WDRR_BUCKET_FOR_HH = 0, /* bucket id for heavy-hitters */ | |
106 | WDRR_BUCKET_FOR_NON_HH = 1 /* bucket id for non-heavy-hitters */ | |
107 | }; | |
108 | ||
109 | #define hhf_time_before(a, b) \ | |
110 | (typecheck(u32, a) && typecheck(u32, b) && ((s32)((a) - (b)) < 0)) | |
111 | ||
112 | /* Heavy-hitter per-flow state */ | |
113 | struct hh_flow_state { | |
114 | u32 hash_id; /* hash of flow-id (e.g. TCP 5-tuple) */ | |
115 | u32 hit_timestamp; /* last time heavy-hitter was seen */ | |
116 | struct list_head flowchain; /* chaining under hash collision */ | |
117 | }; | |
118 | ||
119 | /* Weighted Deficit Round Robin (WDRR) scheduler */ | |
120 | struct wdrr_bucket { | |
121 | struct sk_buff *head; | |
122 | struct sk_buff *tail; | |
123 | struct list_head bucketchain; | |
124 | int deficit; | |
125 | }; | |
126 | ||
127 | struct hhf_sched_data { | |
128 | struct wdrr_bucket buckets[WDRR_BUCKET_CNT]; | |
55667441 | 129 | siphash_key_t perturbation; /* hash perturbation */ |
10239edf TL |
130 | u32 quantum; /* psched_mtu(qdisc_dev(sch)); */ |
131 | u32 drop_overlimit; /* number of times max qdisc packet | |
132 | * limit was hit | |
133 | */ | |
134 | struct list_head *hh_flows; /* table T (currently active HHs) */ | |
135 | u32 hh_flows_limit; /* max active HH allocs */ | |
136 | u32 hh_flows_overlimit; /* num of disallowed HH allocs */ | |
137 | u32 hh_flows_total_cnt; /* total admitted HHs */ | |
138 | u32 hh_flows_current_cnt; /* total current HHs */ | |
139 | u32 *hhf_arrays[HHF_ARRAYS_CNT]; /* HH filter F */ | |
140 | u32 hhf_arrays_reset_timestamp; /* last time hhf_arrays | |
141 | * was reset | |
142 | */ | |
143 | unsigned long *hhf_valid_bits[HHF_ARRAYS_CNT]; /* shadow valid bits | |
144 | * of hhf_arrays | |
145 | */ | |
146 | /* Similar to the "new_flows" vs. "old_flows" concept in fq_codel DRR */ | |
147 | struct list_head new_buckets; /* list of new buckets */ | |
148 | struct list_head old_buckets; /* list of old buckets */ | |
149 | ||
150 | /* Configurable HHF parameters */ | |
151 | u32 hhf_reset_timeout; /* interval to reset counter | |
152 | * arrays in filter F | |
153 | * (default 40ms) | |
154 | */ | |
155 | u32 hhf_admit_bytes; /* counter thresh to classify as | |
156 | * HH (default 128KB). | |
157 | * With these default values, | |
158 | * 128KB / 40ms = 25 Mbps | |
159 | * i.e., we expect to capture HHs | |
160 | * sending > 25 Mbps. | |
161 | */ | |
162 | u32 hhf_evict_timeout; /* aging threshold to evict idle | |
163 | * HHs out of table T. This should | |
164 | * be large enough to avoid | |
165 | * reordering during HH eviction. | |
166 | * (default 1s) | |
167 | */ | |
168 | u32 hhf_non_hh_weight; /* WDRR weight for non-HHs | |
169 | * (default 2, | |
170 | * i.e., non-HH : HH = 2 : 1) | |
171 | */ | |
172 | }; | |
173 | ||
174 | static u32 hhf_time_stamp(void) | |
175 | { | |
176 | return jiffies; | |
177 | } | |
178 | ||
10239edf TL |
179 | /* Looks up a heavy-hitter flow in a chaining list of table T. */ |
180 | static struct hh_flow_state *seek_list(const u32 hash, | |
181 | struct list_head *head, | |
182 | struct hhf_sched_data *q) | |
183 | { | |
184 | struct hh_flow_state *flow, *next; | |
185 | u32 now = hhf_time_stamp(); | |
186 | ||
187 | if (list_empty(head)) | |
188 | return NULL; | |
189 | ||
190 | list_for_each_entry_safe(flow, next, head, flowchain) { | |
191 | u32 prev = flow->hit_timestamp + q->hhf_evict_timeout; | |
192 | ||
193 | if (hhf_time_before(prev, now)) { | |
194 | /* Delete expired heavy-hitters, but preserve one entry | |
195 | * to avoid kzalloc() when next time this slot is hit. | |
196 | */ | |
197 | if (list_is_last(&flow->flowchain, head)) | |
198 | return NULL; | |
199 | list_del(&flow->flowchain); | |
200 | kfree(flow); | |
201 | q->hh_flows_current_cnt--; | |
202 | } else if (flow->hash_id == hash) { | |
203 | return flow; | |
204 | } | |
205 | } | |
206 | return NULL; | |
207 | } | |
208 | ||
209 | /* Returns a flow state entry for a new heavy-hitter. Either reuses an expired | |
210 | * entry or dynamically alloc a new entry. | |
211 | */ | |
212 | static struct hh_flow_state *alloc_new_hh(struct list_head *head, | |
213 | struct hhf_sched_data *q) | |
214 | { | |
215 | struct hh_flow_state *flow; | |
216 | u32 now = hhf_time_stamp(); | |
217 | ||
218 | if (!list_empty(head)) { | |
219 | /* Find an expired heavy-hitter flow entry. */ | |
220 | list_for_each_entry(flow, head, flowchain) { | |
221 | u32 prev = flow->hit_timestamp + q->hhf_evict_timeout; | |
222 | ||
223 | if (hhf_time_before(prev, now)) | |
224 | return flow; | |
225 | } | |
226 | } | |
227 | ||
228 | if (q->hh_flows_current_cnt >= q->hh_flows_limit) { | |
229 | q->hh_flows_overlimit++; | |
230 | return NULL; | |
231 | } | |
232 | /* Create new entry. */ | |
233 | flow = kzalloc(sizeof(struct hh_flow_state), GFP_ATOMIC); | |
234 | if (!flow) | |
235 | return NULL; | |
236 | ||
237 | q->hh_flows_current_cnt++; | |
238 | INIT_LIST_HEAD(&flow->flowchain); | |
239 | list_add_tail(&flow->flowchain, head); | |
240 | ||
241 | return flow; | |
242 | } | |
243 | ||
244 | /* Assigns packets to WDRR buckets. Implements a multi-stage filter to | |
245 | * classify heavy-hitters. | |
246 | */ | |
247 | static enum wdrr_bucket_idx hhf_classify(struct sk_buff *skb, struct Qdisc *sch) | |
248 | { | |
249 | struct hhf_sched_data *q = qdisc_priv(sch); | |
250 | u32 tmp_hash, hash; | |
251 | u32 xorsum, filter_pos[HHF_ARRAYS_CNT], flow_pos; | |
252 | struct hh_flow_state *flow; | |
253 | u32 pkt_len, min_hhf_val; | |
254 | int i; | |
255 | u32 prev; | |
256 | u32 now = hhf_time_stamp(); | |
257 | ||
258 | /* Reset the HHF counter arrays if this is the right time. */ | |
259 | prev = q->hhf_arrays_reset_timestamp + q->hhf_reset_timeout; | |
260 | if (hhf_time_before(prev, now)) { | |
261 | for (i = 0; i < HHF_ARRAYS_CNT; i++) | |
262 | bitmap_zero(q->hhf_valid_bits[i], HHF_ARRAYS_LEN); | |
263 | q->hhf_arrays_reset_timestamp = now; | |
264 | } | |
265 | ||
266 | /* Get hashed flow-id of the skb. */ | |
55667441 | 267 | hash = skb_get_hash_perturb(skb, &q->perturbation); |
10239edf TL |
268 | |
269 | /* Check if this packet belongs to an already established HH flow. */ | |
270 | flow_pos = hash & HHF_BIT_MASK; | |
271 | flow = seek_list(hash, &q->hh_flows[flow_pos], q); | |
272 | if (flow) { /* found its HH flow */ | |
273 | flow->hit_timestamp = now; | |
274 | return WDRR_BUCKET_FOR_HH; | |
275 | } | |
276 | ||
277 | /* Now pass the packet through the multi-stage filter. */ | |
278 | tmp_hash = hash; | |
279 | xorsum = 0; | |
280 | for (i = 0; i < HHF_ARRAYS_CNT - 1; i++) { | |
281 | /* Split the skb_hash into three 10-bit chunks. */ | |
282 | filter_pos[i] = tmp_hash & HHF_BIT_MASK; | |
283 | xorsum ^= filter_pos[i]; | |
284 | tmp_hash >>= HHF_BIT_MASK_LEN; | |
285 | } | |
286 | /* The last chunk is computed as XOR sum of other chunks. */ | |
287 | filter_pos[HHF_ARRAYS_CNT - 1] = xorsum ^ tmp_hash; | |
288 | ||
289 | pkt_len = qdisc_pkt_len(skb); | |
290 | min_hhf_val = ~0U; | |
291 | for (i = 0; i < HHF_ARRAYS_CNT; i++) { | |
292 | u32 val; | |
293 | ||
294 | if (!test_bit(filter_pos[i], q->hhf_valid_bits[i])) { | |
295 | q->hhf_arrays[i][filter_pos[i]] = 0; | |
296 | __set_bit(filter_pos[i], q->hhf_valid_bits[i]); | |
297 | } | |
298 | ||
299 | val = q->hhf_arrays[i][filter_pos[i]] + pkt_len; | |
300 | if (min_hhf_val > val) | |
301 | min_hhf_val = val; | |
302 | } | |
303 | ||
304 | /* Found a new HH iff all counter values > HH admit threshold. */ | |
305 | if (min_hhf_val > q->hhf_admit_bytes) { | |
306 | /* Just captured a new heavy-hitter. */ | |
307 | flow = alloc_new_hh(&q->hh_flows[flow_pos], q); | |
308 | if (!flow) /* memory alloc problem */ | |
309 | return WDRR_BUCKET_FOR_NON_HH; | |
310 | flow->hash_id = hash; | |
311 | flow->hit_timestamp = now; | |
312 | q->hh_flows_total_cnt++; | |
313 | ||
314 | /* By returning without updating counters in q->hhf_arrays, | |
315 | * we implicitly implement "shielding" (see Optimization O1). | |
316 | */ | |
317 | return WDRR_BUCKET_FOR_HH; | |
318 | } | |
319 | ||
320 | /* Conservative update of HHF arrays (see Optimization O2). */ | |
321 | for (i = 0; i < HHF_ARRAYS_CNT; i++) { | |
322 | if (q->hhf_arrays[i][filter_pos[i]] < min_hhf_val) | |
323 | q->hhf_arrays[i][filter_pos[i]] = min_hhf_val; | |
324 | } | |
325 | return WDRR_BUCKET_FOR_NON_HH; | |
326 | } | |
327 | ||
328 | /* Removes one skb from head of bucket. */ | |
329 | static struct sk_buff *dequeue_head(struct wdrr_bucket *bucket) | |
330 | { | |
331 | struct sk_buff *skb = bucket->head; | |
332 | ||
333 | bucket->head = skb->next; | |
a8305bff | 334 | skb_mark_not_on_list(skb); |
10239edf TL |
335 | return skb; |
336 | } | |
337 | ||
338 | /* Tail-adds skb to bucket. */ | |
339 | static void bucket_add(struct wdrr_bucket *bucket, struct sk_buff *skb) | |
340 | { | |
341 | if (bucket->head == NULL) | |
342 | bucket->head = skb; | |
343 | else | |
344 | bucket->tail->next = skb; | |
345 | bucket->tail = skb; | |
346 | skb->next = NULL; | |
347 | } | |
348 | ||
520ac30f | 349 | static unsigned int hhf_drop(struct Qdisc *sch, struct sk_buff **to_free) |
10239edf TL |
350 | { |
351 | struct hhf_sched_data *q = qdisc_priv(sch); | |
352 | struct wdrr_bucket *bucket; | |
353 | ||
354 | /* Always try to drop from heavy-hitters first. */ | |
355 | bucket = &q->buckets[WDRR_BUCKET_FOR_HH]; | |
356 | if (!bucket->head) | |
357 | bucket = &q->buckets[WDRR_BUCKET_FOR_NON_HH]; | |
358 | ||
359 | if (bucket->head) { | |
360 | struct sk_buff *skb = dequeue_head(bucket); | |
361 | ||
362 | sch->q.qlen--; | |
25331d6c | 363 | qdisc_qstats_backlog_dec(sch, skb); |
520ac30f | 364 | qdisc_drop(skb, sch, to_free); |
10239edf TL |
365 | } |
366 | ||
367 | /* Return id of the bucket from which the packet was dropped. */ | |
368 | return bucket - q->buckets; | |
369 | } | |
370 | ||
520ac30f ED |
371 | static int hhf_enqueue(struct sk_buff *skb, struct Qdisc *sch, |
372 | struct sk_buff **to_free) | |
10239edf TL |
373 | { |
374 | struct hhf_sched_data *q = qdisc_priv(sch); | |
375 | enum wdrr_bucket_idx idx; | |
376 | struct wdrr_bucket *bucket; | |
2ccccf5f | 377 | unsigned int prev_backlog; |
10239edf TL |
378 | |
379 | idx = hhf_classify(skb, sch); | |
380 | ||
381 | bucket = &q->buckets[idx]; | |
382 | bucket_add(bucket, skb); | |
25331d6c | 383 | qdisc_qstats_backlog_inc(sch, skb); |
10239edf TL |
384 | |
385 | if (list_empty(&bucket->bucketchain)) { | |
386 | unsigned int weight; | |
387 | ||
388 | /* The logic of new_buckets vs. old_buckets is the same as | |
389 | * new_flows vs. old_flows in the implementation of fq_codel, | |
390 | * i.e., short bursts of non-HHs should have strict priority. | |
391 | */ | |
392 | if (idx == WDRR_BUCKET_FOR_HH) { | |
393 | /* Always move heavy-hitters to old bucket. */ | |
394 | weight = 1; | |
395 | list_add_tail(&bucket->bucketchain, &q->old_buckets); | |
396 | } else { | |
397 | weight = q->hhf_non_hh_weight; | |
398 | list_add_tail(&bucket->bucketchain, &q->new_buckets); | |
399 | } | |
400 | bucket->deficit = weight * q->quantum; | |
401 | } | |
b2ce49e7 | 402 | if (++sch->q.qlen <= sch->limit) |
10239edf TL |
403 | return NET_XMIT_SUCCESS; |
404 | ||
2ccccf5f | 405 | prev_backlog = sch->qstats.backlog; |
10239edf TL |
406 | q->drop_overlimit++; |
407 | /* Return Congestion Notification only if we dropped a packet from this | |
408 | * bucket. | |
409 | */ | |
520ac30f | 410 | if (hhf_drop(sch, to_free) == idx) |
10239edf TL |
411 | return NET_XMIT_CN; |
412 | ||
413 | /* As we dropped a packet, better let upper stack know this. */ | |
2ccccf5f | 414 | qdisc_tree_reduce_backlog(sch, 1, prev_backlog - sch->qstats.backlog); |
10239edf TL |
415 | return NET_XMIT_SUCCESS; |
416 | } | |
417 | ||
418 | static struct sk_buff *hhf_dequeue(struct Qdisc *sch) | |
419 | { | |
420 | struct hhf_sched_data *q = qdisc_priv(sch); | |
421 | struct sk_buff *skb = NULL; | |
422 | struct wdrr_bucket *bucket; | |
423 | struct list_head *head; | |
424 | ||
425 | begin: | |
426 | head = &q->new_buckets; | |
427 | if (list_empty(head)) { | |
428 | head = &q->old_buckets; | |
429 | if (list_empty(head)) | |
430 | return NULL; | |
431 | } | |
432 | bucket = list_first_entry(head, struct wdrr_bucket, bucketchain); | |
433 | ||
434 | if (bucket->deficit <= 0) { | |
435 | int weight = (bucket - q->buckets == WDRR_BUCKET_FOR_HH) ? | |
436 | 1 : q->hhf_non_hh_weight; | |
437 | ||
438 | bucket->deficit += weight * q->quantum; | |
439 | list_move_tail(&bucket->bucketchain, &q->old_buckets); | |
440 | goto begin; | |
441 | } | |
442 | ||
443 | if (bucket->head) { | |
444 | skb = dequeue_head(bucket); | |
445 | sch->q.qlen--; | |
25331d6c | 446 | qdisc_qstats_backlog_dec(sch, skb); |
10239edf TL |
447 | } |
448 | ||
449 | if (!skb) { | |
450 | /* Force a pass through old_buckets to prevent starvation. */ | |
451 | if ((head == &q->new_buckets) && !list_empty(&q->old_buckets)) | |
452 | list_move_tail(&bucket->bucketchain, &q->old_buckets); | |
453 | else | |
454 | list_del_init(&bucket->bucketchain); | |
455 | goto begin; | |
456 | } | |
457 | qdisc_bstats_update(sch, skb); | |
458 | bucket->deficit -= qdisc_pkt_len(skb); | |
459 | ||
460 | return skb; | |
461 | } | |
462 | ||
463 | static void hhf_reset(struct Qdisc *sch) | |
464 | { | |
465 | struct sk_buff *skb; | |
466 | ||
467 | while ((skb = hhf_dequeue(sch)) != NULL) | |
e7e424cd | 468 | rtnl_kfree_skbs(skb, skb); |
10239edf TL |
469 | } |
470 | ||
10239edf TL |
471 | static void hhf_destroy(struct Qdisc *sch) |
472 | { | |
473 | int i; | |
474 | struct hhf_sched_data *q = qdisc_priv(sch); | |
475 | ||
476 | for (i = 0; i < HHF_ARRAYS_CNT; i++) { | |
752ade68 MH |
477 | kvfree(q->hhf_arrays[i]); |
478 | kvfree(q->hhf_valid_bits[i]); | |
10239edf TL |
479 | } |
480 | ||
32db864d NA |
481 | if (!q->hh_flows) |
482 | return; | |
483 | ||
10239edf TL |
484 | for (i = 0; i < HH_FLOWS_CNT; i++) { |
485 | struct hh_flow_state *flow, *next; | |
486 | struct list_head *head = &q->hh_flows[i]; | |
487 | ||
488 | if (list_empty(head)) | |
489 | continue; | |
490 | list_for_each_entry_safe(flow, next, head, flowchain) { | |
491 | list_del(&flow->flowchain); | |
492 | kfree(flow); | |
493 | } | |
494 | } | |
752ade68 | 495 | kvfree(q->hh_flows); |
10239edf TL |
496 | } |
497 | ||
498 | static const struct nla_policy hhf_policy[TCA_HHF_MAX + 1] = { | |
499 | [TCA_HHF_BACKLOG_LIMIT] = { .type = NLA_U32 }, | |
500 | [TCA_HHF_QUANTUM] = { .type = NLA_U32 }, | |
501 | [TCA_HHF_HH_FLOWS_LIMIT] = { .type = NLA_U32 }, | |
502 | [TCA_HHF_RESET_TIMEOUT] = { .type = NLA_U32 }, | |
503 | [TCA_HHF_ADMIT_BYTES] = { .type = NLA_U32 }, | |
504 | [TCA_HHF_EVICT_TIMEOUT] = { .type = NLA_U32 }, | |
505 | [TCA_HHF_NON_HH_WEIGHT] = { .type = NLA_U32 }, | |
506 | }; | |
507 | ||
2030721c AA |
508 | static int hhf_change(struct Qdisc *sch, struct nlattr *opt, |
509 | struct netlink_ext_ack *extack) | |
10239edf TL |
510 | { |
511 | struct hhf_sched_data *q = qdisc_priv(sch); | |
512 | struct nlattr *tb[TCA_HHF_MAX + 1]; | |
2ccccf5f | 513 | unsigned int qlen, prev_backlog; |
10239edf TL |
514 | int err; |
515 | u64 non_hh_quantum; | |
516 | u32 new_quantum = q->quantum; | |
517 | u32 new_hhf_non_hh_weight = q->hhf_non_hh_weight; | |
518 | ||
519 | if (!opt) | |
520 | return -EINVAL; | |
521 | ||
8cb08174 JB |
522 | err = nla_parse_nested_deprecated(tb, TCA_HHF_MAX, opt, hhf_policy, |
523 | NULL); | |
10239edf TL |
524 | if (err < 0) |
525 | return err; | |
526 | ||
10239edf TL |
527 | if (tb[TCA_HHF_QUANTUM]) |
528 | new_quantum = nla_get_u32(tb[TCA_HHF_QUANTUM]); | |
529 | ||
530 | if (tb[TCA_HHF_NON_HH_WEIGHT]) | |
531 | new_hhf_non_hh_weight = nla_get_u32(tb[TCA_HHF_NON_HH_WEIGHT]); | |
532 | ||
533 | non_hh_quantum = (u64)new_quantum * new_hhf_non_hh_weight; | |
d4d6ec6d | 534 | if (non_hh_quantum == 0 || non_hh_quantum > INT_MAX) |
10239edf | 535 | return -EINVAL; |
f6a082fe JF |
536 | |
537 | sch_tree_lock(sch); | |
538 | ||
539 | if (tb[TCA_HHF_BACKLOG_LIMIT]) | |
540 | sch->limit = nla_get_u32(tb[TCA_HHF_BACKLOG_LIMIT]); | |
541 | ||
10239edf TL |
542 | q->quantum = new_quantum; |
543 | q->hhf_non_hh_weight = new_hhf_non_hh_weight; | |
544 | ||
545 | if (tb[TCA_HHF_HH_FLOWS_LIMIT]) | |
546 | q->hh_flows_limit = nla_get_u32(tb[TCA_HHF_HH_FLOWS_LIMIT]); | |
547 | ||
548 | if (tb[TCA_HHF_RESET_TIMEOUT]) { | |
6c76a07a | 549 | u32 us = nla_get_u32(tb[TCA_HHF_RESET_TIMEOUT]); |
10239edf | 550 | |
6c76a07a | 551 | q->hhf_reset_timeout = usecs_to_jiffies(us); |
10239edf TL |
552 | } |
553 | ||
554 | if (tb[TCA_HHF_ADMIT_BYTES]) | |
555 | q->hhf_admit_bytes = nla_get_u32(tb[TCA_HHF_ADMIT_BYTES]); | |
556 | ||
557 | if (tb[TCA_HHF_EVICT_TIMEOUT]) { | |
6c76a07a | 558 | u32 us = nla_get_u32(tb[TCA_HHF_EVICT_TIMEOUT]); |
10239edf | 559 | |
6c76a07a | 560 | q->hhf_evict_timeout = usecs_to_jiffies(us); |
10239edf TL |
561 | } |
562 | ||
563 | qlen = sch->q.qlen; | |
2ccccf5f | 564 | prev_backlog = sch->qstats.backlog; |
10239edf TL |
565 | while (sch->q.qlen > sch->limit) { |
566 | struct sk_buff *skb = hhf_dequeue(sch); | |
567 | ||
e7e424cd | 568 | rtnl_kfree_skbs(skb, skb); |
10239edf | 569 | } |
2ccccf5f WC |
570 | qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, |
571 | prev_backlog - sch->qstats.backlog); | |
10239edf TL |
572 | |
573 | sch_tree_unlock(sch); | |
574 | return 0; | |
575 | } | |
576 | ||
e63d7dfd AA |
577 | static int hhf_init(struct Qdisc *sch, struct nlattr *opt, |
578 | struct netlink_ext_ack *extack) | |
10239edf TL |
579 | { |
580 | struct hhf_sched_data *q = qdisc_priv(sch); | |
581 | int i; | |
582 | ||
583 | sch->limit = 1000; | |
584 | q->quantum = psched_mtu(qdisc_dev(sch)); | |
55667441 | 585 | get_random_bytes(&q->perturbation, sizeof(q->perturbation)); |
10239edf TL |
586 | INIT_LIST_HEAD(&q->new_buckets); |
587 | INIT_LIST_HEAD(&q->old_buckets); | |
588 | ||
589 | /* Configurable HHF parameters */ | |
590 | q->hhf_reset_timeout = HZ / 25; /* 40 ms */ | |
591 | q->hhf_admit_bytes = 131072; /* 128 KB */ | |
592 | q->hhf_evict_timeout = HZ; /* 1 sec */ | |
593 | q->hhf_non_hh_weight = 2; | |
594 | ||
595 | if (opt) { | |
2030721c | 596 | int err = hhf_change(sch, opt, extack); |
10239edf TL |
597 | |
598 | if (err) | |
599 | return err; | |
600 | } | |
601 | ||
602 | if (!q->hh_flows) { | |
603 | /* Initialize heavy-hitter flow table. */ | |
778e1cdd KC |
604 | q->hh_flows = kvcalloc(HH_FLOWS_CNT, sizeof(struct list_head), |
605 | GFP_KERNEL); | |
10239edf TL |
606 | if (!q->hh_flows) |
607 | return -ENOMEM; | |
608 | for (i = 0; i < HH_FLOWS_CNT; i++) | |
609 | INIT_LIST_HEAD(&q->hh_flows[i]); | |
610 | ||
611 | /* Cap max active HHs at twice len of hh_flows table. */ | |
612 | q->hh_flows_limit = 2 * HH_FLOWS_CNT; | |
613 | q->hh_flows_overlimit = 0; | |
614 | q->hh_flows_total_cnt = 0; | |
615 | q->hh_flows_current_cnt = 0; | |
616 | ||
617 | /* Initialize heavy-hitter filter arrays. */ | |
618 | for (i = 0; i < HHF_ARRAYS_CNT; i++) { | |
778e1cdd KC |
619 | q->hhf_arrays[i] = kvcalloc(HHF_ARRAYS_LEN, |
620 | sizeof(u32), | |
621 | GFP_KERNEL); | |
10239edf | 622 | if (!q->hhf_arrays[i]) { |
87b60cfa ED |
623 | /* Note: hhf_destroy() will be called |
624 | * by our caller. | |
625 | */ | |
10239edf TL |
626 | return -ENOMEM; |
627 | } | |
628 | } | |
629 | q->hhf_arrays_reset_timestamp = hhf_time_stamp(); | |
630 | ||
631 | /* Initialize valid bits of heavy-hitter filter arrays. */ | |
632 | for (i = 0; i < HHF_ARRAYS_CNT; i++) { | |
752ade68 MH |
633 | q->hhf_valid_bits[i] = kvzalloc(HHF_ARRAYS_LEN / |
634 | BITS_PER_BYTE, GFP_KERNEL); | |
10239edf | 635 | if (!q->hhf_valid_bits[i]) { |
87b60cfa ED |
636 | /* Note: hhf_destroy() will be called |
637 | * by our caller. | |
638 | */ | |
10239edf TL |
639 | return -ENOMEM; |
640 | } | |
641 | } | |
642 | ||
643 | /* Initialize Weighted DRR buckets. */ | |
644 | for (i = 0; i < WDRR_BUCKET_CNT; i++) { | |
645 | struct wdrr_bucket *bucket = q->buckets + i; | |
646 | ||
647 | INIT_LIST_HEAD(&bucket->bucketchain); | |
648 | } | |
649 | } | |
650 | ||
651 | return 0; | |
652 | } | |
653 | ||
654 | static int hhf_dump(struct Qdisc *sch, struct sk_buff *skb) | |
655 | { | |
656 | struct hhf_sched_data *q = qdisc_priv(sch); | |
657 | struct nlattr *opts; | |
658 | ||
ae0be8de | 659 | opts = nla_nest_start_noflag(skb, TCA_OPTIONS); |
10239edf TL |
660 | if (opts == NULL) |
661 | goto nla_put_failure; | |
662 | ||
663 | if (nla_put_u32(skb, TCA_HHF_BACKLOG_LIMIT, sch->limit) || | |
664 | nla_put_u32(skb, TCA_HHF_QUANTUM, q->quantum) || | |
665 | nla_put_u32(skb, TCA_HHF_HH_FLOWS_LIMIT, q->hh_flows_limit) || | |
666 | nla_put_u32(skb, TCA_HHF_RESET_TIMEOUT, | |
6c76a07a | 667 | jiffies_to_usecs(q->hhf_reset_timeout)) || |
10239edf TL |
668 | nla_put_u32(skb, TCA_HHF_ADMIT_BYTES, q->hhf_admit_bytes) || |
669 | nla_put_u32(skb, TCA_HHF_EVICT_TIMEOUT, | |
6c76a07a | 670 | jiffies_to_usecs(q->hhf_evict_timeout)) || |
10239edf TL |
671 | nla_put_u32(skb, TCA_HHF_NON_HH_WEIGHT, q->hhf_non_hh_weight)) |
672 | goto nla_put_failure; | |
673 | ||
d59b7d80 | 674 | return nla_nest_end(skb, opts); |
10239edf TL |
675 | |
676 | nla_put_failure: | |
677 | return -1; | |
678 | } | |
679 | ||
680 | static int hhf_dump_stats(struct Qdisc *sch, struct gnet_dump *d) | |
681 | { | |
682 | struct hhf_sched_data *q = qdisc_priv(sch); | |
683 | struct tc_hhf_xstats st = { | |
684 | .drop_overlimit = q->drop_overlimit, | |
685 | .hh_overlimit = q->hh_flows_overlimit, | |
686 | .hh_tot_count = q->hh_flows_total_cnt, | |
687 | .hh_cur_count = q->hh_flows_current_cnt, | |
688 | }; | |
689 | ||
690 | return gnet_stats_copy_app(d, &st, sizeof(st)); | |
691 | } | |
692 | ||
c49fa257 | 693 | static struct Qdisc_ops hhf_qdisc_ops __read_mostly = { |
10239edf TL |
694 | .id = "hhf", |
695 | .priv_size = sizeof(struct hhf_sched_data), | |
696 | ||
697 | .enqueue = hhf_enqueue, | |
698 | .dequeue = hhf_dequeue, | |
699 | .peek = qdisc_peek_dequeued, | |
10239edf TL |
700 | .init = hhf_init, |
701 | .reset = hhf_reset, | |
702 | .destroy = hhf_destroy, | |
703 | .change = hhf_change, | |
704 | .dump = hhf_dump, | |
705 | .dump_stats = hhf_dump_stats, | |
706 | .owner = THIS_MODULE, | |
707 | }; | |
10239edf TL |
708 | |
709 | static int __init hhf_module_init(void) | |
710 | { | |
711 | return register_qdisc(&hhf_qdisc_ops); | |
712 | } | |
713 | ||
714 | static void __exit hhf_module_exit(void) | |
715 | { | |
716 | unregister_qdisc(&hhf_qdisc_ops); | |
717 | } | |
718 | ||
719 | module_init(hhf_module_init) | |
720 | module_exit(hhf_module_exit) | |
721 | MODULE_AUTHOR("Terry Lam"); | |
722 | MODULE_AUTHOR("Nandita Dukkipati"); | |
723 | MODULE_LICENSE("GPL"); |