1 /* SPDX-License-Identifier: GPL-2.0-only */
2 /*
3 * Copyright (c) 2016 Qualcomm Atheros, Inc
4 *
5 * Based on net/sched/sch_fq_codel.c
6 */
7 #ifndef __NET_SCHED_FQ_IMPL_H
8 #define __NET_SCHED_FQ_IMPL_H
9
10 #include <net/fq.h>
11
12 /* functions that are embedded into includer */
13
14
15 static void
16 __fq_adjust_removal(struct fq *fq, struct fq_flow *flow, unsigned int packets,
17 unsigned int bytes, unsigned int truesize)
18 {
19 struct fq_tin *tin = flow->tin;
20 int idx;
21
22 tin->backlog_bytes -= bytes;
23 tin->backlog_packets -= packets;
24 flow->backlog -= bytes;
25 fq->backlog -= packets;
26 fq->memory_usage -= truesize;
27
28 if (flow->backlog)
29 return;
30
31 if (flow == &tin->default_flow) {
32 list_del_init(&tin->tin_list);
33 return;
34 }
35
36 idx = flow - fq->flows;
37 __clear_bit(idx, fq->flows_bitmap);
38 }
39
40 static void fq_adjust_removal(struct fq *fq,
41 struct fq_flow *flow,
42 struct sk_buff *skb)
43 {
44 __fq_adjust_removal(fq, flow, 1, skb->len, skb->truesize);
45 }
46
47 static struct sk_buff *fq_flow_dequeue(struct fq *fq,
48 struct fq_flow *flow)
49 {
50 struct sk_buff *skb;
51
52 lockdep_assert_held(&fq->lock);
53
54 skb = __skb_dequeue(&flow->queue);
55 if (!skb)
56 return NULL;
57
58 fq_adjust_removal(fq, flow, skb);
59
60 return skb;
61 }
62
63 static int fq_flow_drop(struct fq *fq, struct fq_flow *flow,
64 fq_skb_free_t free_func)
65 {
66 unsigned int packets = 0, bytes = 0, truesize = 0;
67 struct fq_tin *tin = flow->tin;
68 struct sk_buff *skb;
69 int pending;
70
71 lockdep_assert_held(&fq->lock);
72
73 pending = min_t(int, 32, skb_queue_len(&flow->queue) / 2);
74 do {
75 skb = __skb_dequeue(&flow->queue);
76 if (!skb)
77 break;
78
79 packets++;
80 bytes += skb->len;
81 truesize += skb->truesize;
82 free_func(fq, tin, flow, skb);
83 } while (packets < pending);
84
85 __fq_adjust_removal(fq, flow, packets, bytes, truesize);
86
87 return packets;
88 }
89
90 static struct sk_buff *fq_tin_dequeue(struct fq *fq,
91 struct fq_tin *tin,
92 fq_tin_dequeue_t dequeue_func)
93 {
94 struct fq_flow *flow;
95 struct list_head *head;
96 struct sk_buff *skb;
97
98 lockdep_assert_held(&fq->lock);
99
100 begin:
101 head = &tin->new_flows;
102 if (list_empty(head)) {
103 head = &tin->old_flows;
104 if (list_empty(head))
105 return NULL;
106 }
107
108 flow = list_first_entry(head, struct fq_flow, flowchain);
109
110 if (flow->deficit <= 0) {
111 flow->deficit += fq->quantum;
112 list_move_tail(&flow->flowchain,
113 &tin->old_flows);
114 goto begin;
115 }
116
117 skb = dequeue_func(fq, tin, flow);
118 if (!skb) {
119 /* force a pass through old_flows to prevent starvation */
120 if ((head == &tin->new_flows) &&
121 !list_empty(&tin->old_flows)) {
122 list_move_tail(&flow->flowchain, &tin->old_flows);
123 } else {
124 list_del_init(&flow->flowchain);
125 flow->tin = NULL;
126 }
127 goto begin;
128 }
129
130 flow->deficit -= skb->len;
131 tin->tx_bytes += skb->len;
132 tin->tx_packets++;
133
134 return skb;
135 }
136
137 static u32 fq_flow_idx(struct fq *fq, struct sk_buff *skb)
138 {
139 u32 hash = skb_get_hash(skb);
140
141 return reciprocal_scale(hash, fq->flows_cnt);
142 }
143
144 static struct fq_flow *fq_flow_classify(struct fq *fq,
145 struct fq_tin *tin, u32 idx,
146 struct sk_buff *skb)
147 {
148 struct fq_flow *flow;
149
150 lockdep_assert_held(&fq->lock);
151
152 flow = &fq->flows[idx];
153 if (flow->tin && flow->tin != tin) {
154 flow = &tin->default_flow;
155 tin->collisions++;
156 fq->collisions++;
157 }
158
159 if (!flow->tin)
160 tin->flows++;
161
162 return flow;
163 }
164
165 static struct fq_flow *fq_find_fattest_flow(struct fq *fq)
166 {
167 struct fq_tin *tin;
168 struct fq_flow *flow = NULL;
169 u32 len = 0;
170 int i;
171
172 for_each_set_bit(i, fq->flows_bitmap, fq->flows_cnt) {
173 struct fq_flow *cur = &fq->flows[i];
174 unsigned int cur_len;
175
176 cur_len = cur->backlog;
177 if (cur_len <= len)
178 continue;
179
180 flow = cur;
181 len = cur_len;
182 }
183
184 list_for_each_entry(tin, &fq->tin_backlog, tin_list) {
185 unsigned int cur_len = tin->default_flow.backlog;
186
187 if (cur_len <= len)
188 continue;
189
190 flow = &tin->default_flow;
191 len = cur_len;
192 }
193
194 return flow;
195 }
196
197 static void fq_tin_enqueue(struct fq *fq,
198 struct fq_tin *tin, u32 idx,
199 struct sk_buff *skb,
200 fq_skb_free_t free_func)
201 {
202 struct fq_flow *flow;
203 struct sk_buff *next;
204 bool oom;
205
206 lockdep_assert_held(&fq->lock);
207
208 flow = fq_flow_classify(fq, tin, idx, skb);
209
210 if (!flow->backlog) {
211 if (flow != &tin->default_flow)
212 __set_bit(idx, fq->flows_bitmap);
213 else if (list_empty(&tin->tin_list))
214 list_add(&tin->tin_list, &fq->tin_backlog);
215 }
216
217 flow->tin = tin;
218 skb_list_walk_safe(skb, skb, next) {
219 skb_mark_not_on_list(skb);
220 flow->backlog += skb->len;
221 tin->backlog_bytes += skb->len;
222 tin->backlog_packets++;
223 fq->memory_usage += skb->truesize;
224 fq->backlog++;
225 __skb_queue_tail(&flow->queue, skb);
226 }
227
228 if (list_empty(&flow->flowchain)) {
229 flow->deficit = fq->quantum;
230 list_add_tail(&flow->flowchain,
231 &tin->new_flows);
232 }
233
234 oom = (fq->memory_usage > fq->memory_limit);
235 while (fq->backlog > fq->limit || oom) {
236 flow = fq_find_fattest_flow(fq);
237 if (!flow)
238 return;
239
240 if (!fq_flow_drop(fq, flow, free_func))
241 return;
242
243 flow->tin->overlimit++;
244 fq->overlimit++;
245 if (oom) {
246 fq->overmemory++;
247 oom = (fq->memory_usage > fq->memory_limit);
248 }
249 }
250 }
251
252 static void fq_flow_filter(struct fq *fq,
253 struct fq_flow *flow,
254 fq_skb_filter_t filter_func,
255 void *filter_data,
256 fq_skb_free_t free_func)
257 {
258 struct fq_tin *tin = flow->tin;
259 struct sk_buff *skb, *tmp;
260
261 lockdep_assert_held(&fq->lock);
262
263 skb_queue_walk_safe(&flow->queue, skb, tmp) {
264 if (!filter_func(fq, tin, flow, skb, filter_data))
265 continue;
266
267 __skb_unlink(skb, &flow->queue);
268 fq_adjust_removal(fq, flow, skb);
269 free_func(fq, tin, flow, skb);
270 }
271 }
272
273 static void fq_tin_filter(struct fq *fq,
274 struct fq_tin *tin,
275 fq_skb_filter_t filter_func,
276 void *filter_data,
277 fq_skb_free_t free_func)
278 {
279 struct fq_flow *flow;
280
281 lockdep_assert_held(&fq->lock);
282
283 list_for_each_entry(flow, &tin->new_flows, flowchain)
284 fq_flow_filter(fq, flow, filter_func, filter_data, free_func);
285 list_for_each_entry(flow, &tin->old_flows, flowchain)
286 fq_flow_filter(fq, flow, filter_func, filter_data, free_func);
287 }
288
289 static void fq_flow_reset(struct fq *fq,
290 struct fq_flow *flow,
291 fq_skb_free_t free_func)
292 {
293 struct fq_tin *tin = flow->tin;
294 struct sk_buff *skb;
295
296 while ((skb = fq_flow_dequeue(fq, flow)))
297 free_func(fq, tin, flow, skb);
298
299 if (!list_empty(&flow->flowchain)) {
300 list_del_init(&flow->flowchain);
301 if (list_empty(&tin->new_flows) &&
302 list_empty(&tin->old_flows))
303 list_del_init(&tin->tin_list);
304 }
305
306 flow->tin = NULL;
307
308 WARN_ON_ONCE(flow->backlog);
309 }
310
311 static void fq_tin_reset(struct fq *fq,
312 struct fq_tin *tin,
313 fq_skb_free_t free_func)
314 {
315 struct list_head *head;
316 struct fq_flow *flow;
317
318 for (;;) {
319 head = &tin->new_flows;
320 if (list_empty(head)) {
321 head = &tin->old_flows;
322 if (list_empty(head))
323 break;
324 }
325
326 flow = list_first_entry(head, struct fq_flow, flowchain);
327 fq_flow_reset(fq, flow, free_func);
328 }
329
330 WARN_ON_ONCE(!list_empty(&tin->tin_list));
331 WARN_ON_ONCE(tin->backlog_bytes);
332 WARN_ON_ONCE(tin->backlog_packets);
333 }
334
335 static void fq_flow_init(struct fq_flow *flow)
336 {
337 INIT_LIST_HEAD(&flow->flowchain);
338 __skb_queue_head_init(&flow->queue);
339 }
340
341 static void fq_tin_init(struct fq_tin *tin)
342 {
343 INIT_LIST_HEAD(&tin->new_flows);
344 INIT_LIST_HEAD(&tin->old_flows);
345 INIT_LIST_HEAD(&tin->tin_list);
346 fq_flow_init(&tin->default_flow);
347 }
348
349 static int fq_init(struct fq *fq, int flows_cnt)
350 {
351 int i;
352
353 memset(fq, 0, sizeof(fq[0]));
354 spin_lock_init(&fq->lock);
355 INIT_LIST_HEAD(&fq->tin_backlog);
356 fq->flows_cnt = max_t(u32, flows_cnt, 1);
357 fq->quantum = 300;
358 fq->limit = 8192;
359 fq->memory_limit = 16 << 20; /* 16 MBytes */
360
361 fq->flows = kvcalloc(fq->flows_cnt, sizeof(fq->flows[0]), GFP_KERNEL);
362 if (!fq->flows)
363 return -ENOMEM;
364
365 fq->flows_bitmap = bitmap_zalloc(fq->flows_cnt, GFP_KERNEL);
366 if (!fq->flows_bitmap) {
367 kvfree(fq->flows);
368 fq->flows = NULL;
369 return -ENOMEM;
370 }
371
372 for (i = 0; i < fq->flows_cnt; i++)
373 fq_flow_init(&fq->flows[i]);
374
375 return 0;
376 }
377
378 static void fq_reset(struct fq *fq,
379 fq_skb_free_t free_func)
380 {
381 int i;
382
383 for (i = 0; i < fq->flows_cnt; i++)
384 fq_flow_reset(fq, &fq->flows[i], free_func);
385
386 kvfree(fq->flows);
387 fq->flows = NULL;
388
389 bitmap_free(fq->flows_bitmap);
390 fq->flows_bitmap = NULL;
391 }
392
393 #endif
394
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.