xref: /openbmc/linux/include/net/fq_impl.h (revision f7af616c)
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 	bool oom;
204 
205 	lockdep_assert_held(&fq->lock);
206 
207 	flow = fq_flow_classify(fq, tin, idx, skb);
208 
209 	if (!flow->backlog) {
210 		if (flow != &tin->default_flow)
211 			__set_bit(idx, fq->flows_bitmap);
212 		else if (list_empty(&tin->tin_list))
213 			list_add(&tin->tin_list, &fq->tin_backlog);
214 	}
215 
216 	flow->tin = tin;
217 	flow->backlog += skb->len;
218 	tin->backlog_bytes += skb->len;
219 	tin->backlog_packets++;
220 	fq->memory_usage += skb->truesize;
221 	fq->backlog++;
222 
223 	if (list_empty(&flow->flowchain)) {
224 		flow->deficit = fq->quantum;
225 		list_add_tail(&flow->flowchain,
226 			      &tin->new_flows);
227 	}
228 
229 	__skb_queue_tail(&flow->queue, skb);
230 	oom = (fq->memory_usage > fq->memory_limit);
231 	while (fq->backlog > fq->limit || oom) {
232 		flow = fq_find_fattest_flow(fq);
233 		if (!flow)
234 			return;
235 
236 		if (!fq_flow_drop(fq, flow, free_func))
237 			return;
238 
239 		flow->tin->overlimit++;
240 		fq->overlimit++;
241 		if (oom) {
242 			fq->overmemory++;
243 			oom = (fq->memory_usage > fq->memory_limit);
244 		}
245 	}
246 }
247 
248 static void fq_flow_filter(struct fq *fq,
249 			   struct fq_flow *flow,
250 			   fq_skb_filter_t filter_func,
251 			   void *filter_data,
252 			   fq_skb_free_t free_func)
253 {
254 	struct fq_tin *tin = flow->tin;
255 	struct sk_buff *skb, *tmp;
256 
257 	lockdep_assert_held(&fq->lock);
258 
259 	skb_queue_walk_safe(&flow->queue, skb, tmp) {
260 		if (!filter_func(fq, tin, flow, skb, filter_data))
261 			continue;
262 
263 		__skb_unlink(skb, &flow->queue);
264 		fq_adjust_removal(fq, flow, skb);
265 		free_func(fq, tin, flow, skb);
266 	}
267 }
268 
269 static void fq_tin_filter(struct fq *fq,
270 			  struct fq_tin *tin,
271 			  fq_skb_filter_t filter_func,
272 			  void *filter_data,
273 			  fq_skb_free_t free_func)
274 {
275 	struct fq_flow *flow;
276 
277 	lockdep_assert_held(&fq->lock);
278 
279 	list_for_each_entry(flow, &tin->new_flows, flowchain)
280 		fq_flow_filter(fq, flow, filter_func, filter_data, free_func);
281 	list_for_each_entry(flow, &tin->old_flows, flowchain)
282 		fq_flow_filter(fq, flow, filter_func, filter_data, free_func);
283 }
284 
285 static void fq_flow_reset(struct fq *fq,
286 			  struct fq_flow *flow,
287 			  fq_skb_free_t free_func)
288 {
289 	struct fq_tin *tin = flow->tin;
290 	struct sk_buff *skb;
291 
292 	while ((skb = fq_flow_dequeue(fq, flow)))
293 		free_func(fq, tin, flow, skb);
294 
295 	if (!list_empty(&flow->flowchain)) {
296 		list_del_init(&flow->flowchain);
297 		if (list_empty(&tin->new_flows) &&
298 		    list_empty(&tin->old_flows))
299 			list_del_init(&tin->tin_list);
300 	}
301 
302 	flow->tin = NULL;
303 
304 	WARN_ON_ONCE(flow->backlog);
305 }
306 
307 static void fq_tin_reset(struct fq *fq,
308 			 struct fq_tin *tin,
309 			 fq_skb_free_t free_func)
310 {
311 	struct list_head *head;
312 	struct fq_flow *flow;
313 
314 	for (;;) {
315 		head = &tin->new_flows;
316 		if (list_empty(head)) {
317 			head = &tin->old_flows;
318 			if (list_empty(head))
319 				break;
320 		}
321 
322 		flow = list_first_entry(head, struct fq_flow, flowchain);
323 		fq_flow_reset(fq, flow, free_func);
324 	}
325 
326 	WARN_ON_ONCE(!list_empty(&tin->tin_list));
327 	WARN_ON_ONCE(tin->backlog_bytes);
328 	WARN_ON_ONCE(tin->backlog_packets);
329 }
330 
331 static void fq_flow_init(struct fq_flow *flow)
332 {
333 	INIT_LIST_HEAD(&flow->flowchain);
334 	__skb_queue_head_init(&flow->queue);
335 }
336 
337 static void fq_tin_init(struct fq_tin *tin)
338 {
339 	INIT_LIST_HEAD(&tin->new_flows);
340 	INIT_LIST_HEAD(&tin->old_flows);
341 	INIT_LIST_HEAD(&tin->tin_list);
342 	fq_flow_init(&tin->default_flow);
343 }
344 
345 static int fq_init(struct fq *fq, int flows_cnt)
346 {
347 	int i;
348 
349 	memset(fq, 0, sizeof(fq[0]));
350 	spin_lock_init(&fq->lock);
351 	INIT_LIST_HEAD(&fq->tin_backlog);
352 	fq->flows_cnt = max_t(u32, flows_cnt, 1);
353 	fq->quantum = 300;
354 	fq->limit = 8192;
355 	fq->memory_limit = 16 << 20; /* 16 MBytes */
356 
357 	fq->flows = kvcalloc(fq->flows_cnt, sizeof(fq->flows[0]), GFP_KERNEL);
358 	if (!fq->flows)
359 		return -ENOMEM;
360 
361 	fq->flows_bitmap = kcalloc(BITS_TO_LONGS(fq->flows_cnt), sizeof(long),
362 				   GFP_KERNEL);
363 	if (!fq->flows_bitmap) {
364 		kvfree(fq->flows);
365 		fq->flows = NULL;
366 		return -ENOMEM;
367 	}
368 
369 	for (i = 0; i < fq->flows_cnt; i++)
370 		fq_flow_init(&fq->flows[i]);
371 
372 	return 0;
373 }
374 
375 static void fq_reset(struct fq *fq,
376 		     fq_skb_free_t free_func)
377 {
378 	int i;
379 
380 	for (i = 0; i < fq->flows_cnt; i++)
381 		fq_flow_reset(fq, &fq->flows[i], free_func);
382 
383 	kvfree(fq->flows);
384 	fq->flows = NULL;
385 
386 	kfree(fq->flows_bitmap);
387 	fq->flows_bitmap = NULL;
388 }
389 
390 #endif
391