xref: /netbsd-src/sys/net/pktqueue.c (revision 4d342c046e3288fb5a1edcd33cfec48c41c80664)
1 /*	$NetBSD: pktqueue.c,v 1.12 2020/09/11 14:29:00 riastradh Exp $	*/
2 
3 /*-
4  * Copyright (c) 2014 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Mindaugas Rasiukevicius.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  * The packet queue (pktqueue) interface is a lockless IP input queue
34  * which also abstracts and handles network ISR scheduling.  It provides
35  * a mechanism to enable receiver-side packet steering (RPS).
36  */
37 
38 #include <sys/cdefs.h>
39 __KERNEL_RCSID(0, "$NetBSD: pktqueue.c,v 1.12 2020/09/11 14:29:00 riastradh Exp $");
40 
41 #include <sys/param.h>
42 #include <sys/types.h>
43 
44 #include <sys/atomic.h>
45 #include <sys/cpu.h>
46 #include <sys/pcq.h>
47 #include <sys/intr.h>
48 #include <sys/mbuf.h>
49 #include <sys/proc.h>
50 #include <sys/percpu.h>
51 #include <sys/xcall.h>
52 
53 #include <net/pktqueue.h>
54 
55 /*
56  * WARNING: update this if struct pktqueue changes.
57  */
58 #define	PKTQ_CLPAD	\
59     MAX(COHERENCY_UNIT, COHERENCY_UNIT - sizeof(kmutex_t) - sizeof(u_int))
60 
61 struct pktqueue {
62 	/*
63 	 * The lock used for a barrier mechanism.  The barrier counter,
64 	 * as well as the drop counter, are managed atomically though.
65 	 * Ensure this group is in a separate cache line.
66 	 */
67 	kmutex_t	pq_lock;
68 	volatile u_int	pq_barrier;
69 	uint8_t		_pad[PKTQ_CLPAD];
70 
71 	/* The size of the queue, counters and the interrupt handler. */
72 	u_int		pq_maxlen;
73 	percpu_t *	pq_counters;
74 	void *		pq_sih;
75 
76 	/* Finally, per-CPU queues. */
77 	struct percpu *	pq_pcq;	/* struct pcq * */
78 };
79 
80 /* The counters of the packet queue. */
81 #define	PQCNT_ENQUEUE	0
82 #define	PQCNT_DEQUEUE	1
83 #define	PQCNT_DROP	2
84 #define	PQCNT_NCOUNTERS	3
85 
86 typedef struct {
87 	uint64_t	count[PQCNT_NCOUNTERS];
88 } pktq_counters_t;
89 
90 /* Special marker value used by pktq_barrier() mechanism. */
91 #define	PKTQ_MARKER	((void *)(~0ULL))
92 
93 static void
94 pktq_init_cpu(void *vqp, void *vpq, struct cpu_info *ci)
95 {
96 	struct pcq **qp = vqp;
97 	struct pktqueue *pq = vpq;
98 
99 	*qp = pcq_create(pq->pq_maxlen, KM_SLEEP);
100 }
101 
102 static void
103 pktq_fini_cpu(void *vqp, void *vpq, struct cpu_info *ci)
104 {
105 	struct pcq **qp = vqp, *q = *qp;
106 
107 	KASSERT(pcq_peek(q) == NULL);
108 	pcq_destroy(q);
109 	*qp = NULL;		/* paranoia */
110 }
111 
112 static struct pcq *
113 pktq_pcq(struct pktqueue *pq, struct cpu_info *ci)
114 {
115 	struct pcq **qp, *q;
116 
117 	/*
118 	 * As long as preemption is disabled, the xcall to swap percpu
119 	 * buffers can't complete, so it is safe to read the pointer.
120 	 */
121 	KASSERT(kpreempt_disabled());
122 
123 	qp = percpu_getptr_remote(pq->pq_pcq, ci);
124 	q = *qp;
125 
126 	return q;
127 }
128 
129 pktqueue_t *
130 pktq_create(size_t maxlen, void (*intrh)(void *), void *sc)
131 {
132 	const u_int sflags = SOFTINT_NET | SOFTINT_MPSAFE | SOFTINT_RCPU;
133 	pktqueue_t *pq;
134 	percpu_t *pc;
135 	void *sih;
136 
137 	pc = percpu_alloc(sizeof(pktq_counters_t));
138 	if ((sih = softint_establish(sflags, intrh, sc)) == NULL) {
139 		percpu_free(pc, sizeof(pktq_counters_t));
140 		return NULL;
141 	}
142 
143 	pq = kmem_zalloc(sizeof(*pq), KM_SLEEP);
144 	mutex_init(&pq->pq_lock, MUTEX_DEFAULT, IPL_NONE);
145 	pq->pq_maxlen = maxlen;
146 	pq->pq_counters = pc;
147 	pq->pq_sih = sih;
148 	pq->pq_pcq = percpu_create(sizeof(struct pcq *),
149 	    pktq_init_cpu, pktq_fini_cpu, pq);
150 
151 	return pq;
152 }
153 
154 void
155 pktq_destroy(pktqueue_t *pq)
156 {
157 
158 	percpu_free(pq->pq_pcq, sizeof(struct pcq *));
159 	percpu_free(pq->pq_counters, sizeof(pktq_counters_t));
160 	softint_disestablish(pq->pq_sih);
161 	mutex_destroy(&pq->pq_lock);
162 	kmem_free(pq, sizeof(*pq));
163 }
164 
165 /*
166  * - pktq_inc_counter: increment the counter given an ID.
167  * - pktq_collect_counts: handler to sum up the counts from each CPU.
168  * - pktq_getcount: return the effective count given an ID.
169  */
170 
171 static inline void
172 pktq_inc_count(pktqueue_t *pq, u_int i)
173 {
174 	percpu_t *pc = pq->pq_counters;
175 	pktq_counters_t *c;
176 
177 	c = percpu_getref(pc);
178 	c->count[i]++;
179 	percpu_putref(pc);
180 }
181 
182 static void
183 pktq_collect_counts(void *mem, void *arg, struct cpu_info *ci)
184 {
185 	const pktq_counters_t *c = mem;
186 	pktq_counters_t *sum = arg;
187 
188 	int s = splnet();
189 
190 	for (u_int i = 0; i < PQCNT_NCOUNTERS; i++) {
191 		sum->count[i] += c->count[i];
192 	}
193 
194 	splx(s);
195 }
196 
197 uint64_t
198 pktq_get_count(pktqueue_t *pq, pktq_count_t c)
199 {
200 	pktq_counters_t sum;
201 
202 	if (c != PKTQ_MAXLEN) {
203 		memset(&sum, 0, sizeof(sum));
204 		percpu_foreach_xcall(pq->pq_counters,
205 		    XC_HIGHPRI_IPL(IPL_SOFTNET), pktq_collect_counts, &sum);
206 	}
207 	switch (c) {
208 	case PKTQ_NITEMS:
209 		return sum.count[PQCNT_ENQUEUE] - sum.count[PQCNT_DEQUEUE];
210 	case PKTQ_DROPS:
211 		return sum.count[PQCNT_DROP];
212 	case PKTQ_MAXLEN:
213 		return pq->pq_maxlen;
214 	}
215 	return 0;
216 }
217 
218 uint32_t
219 pktq_rps_hash(const struct mbuf *m __unused)
220 {
221 	/*
222 	 * XXX: No distribution yet; the softnet_lock contention
223 	 * XXX: must be eliminated first.
224 	 */
225 	return 0;
226 }
227 
228 /*
229  * pktq_enqueue: inject the packet into the end of the queue.
230  *
231  * => Must be called from the interrupt or with the preemption disabled.
232  * => Consumes the packet and returns true on success.
233  * => Returns false on failure; caller is responsible to free the packet.
234  */
235 bool
236 pktq_enqueue(pktqueue_t *pq, struct mbuf *m, const u_int hash __unused)
237 {
238 #if defined(_RUMPKERNEL) || defined(_RUMP_NATIVE_ABI)
239 	struct cpu_info *ci = curcpu();
240 #else
241 	struct cpu_info *ci = cpu_lookup(hash % ncpu);
242 #endif
243 
244 	KASSERT(kpreempt_disabled());
245 
246 	if (__predict_false(!pcq_put(pktq_pcq(pq, ci), m))) {
247 		pktq_inc_count(pq, PQCNT_DROP);
248 		return false;
249 	}
250 	softint_schedule_cpu(pq->pq_sih, ci);
251 	pktq_inc_count(pq, PQCNT_ENQUEUE);
252 	return true;
253 }
254 
255 /*
256  * pktq_dequeue: take a packet from the queue.
257  *
258  * => Must be called with preemption disabled.
259  * => Must ensure there are not concurrent dequeue calls.
260  */
261 struct mbuf *
262 pktq_dequeue(pktqueue_t *pq)
263 {
264 	struct cpu_info *ci = curcpu();
265 	struct mbuf *m;
266 
267 	KASSERT(kpreempt_disabled());
268 
269 	m = pcq_get(pktq_pcq(pq, ci));
270 	if (__predict_false(m == PKTQ_MARKER)) {
271 		/* Note the marker entry. */
272 		atomic_inc_uint(&pq->pq_barrier);
273 		return NULL;
274 	}
275 	if (__predict_true(m != NULL)) {
276 		pktq_inc_count(pq, PQCNT_DEQUEUE);
277 	}
278 	return m;
279 }
280 
281 /*
282  * pktq_barrier: waits for a grace period when all packets enqueued at
283  * the moment of calling this routine will be processed.  This is used
284  * to ensure that e.g. packets referencing some interface were drained.
285  */
286 void
287 pktq_barrier(pktqueue_t *pq)
288 {
289 	CPU_INFO_ITERATOR cii;
290 	struct cpu_info *ci;
291 	u_int pending = 0;
292 
293 	mutex_enter(&pq->pq_lock);
294 	KASSERT(pq->pq_barrier == 0);
295 
296 	for (CPU_INFO_FOREACH(cii, ci)) {
297 		struct pcq *q;
298 
299 		kpreempt_disable();
300 		q = pktq_pcq(pq, ci);
301 		kpreempt_enable();
302 
303 		/* If the queue is empty - nothing to do. */
304 		if (pcq_peek(q) == NULL) {
305 			continue;
306 		}
307 		/* Otherwise, put the marker and entry. */
308 		while (!pcq_put(q, PKTQ_MARKER)) {
309 			kpause("pktqsync", false, 1, NULL);
310 		}
311 		kpreempt_disable();
312 		softint_schedule_cpu(pq->pq_sih, ci);
313 		kpreempt_enable();
314 		pending++;
315 	}
316 
317 	/* Wait for each queue to process the markers. */
318 	while (pq->pq_barrier != pending) {
319 		kpause("pktqsync", false, 1, NULL);
320 	}
321 	pq->pq_barrier = 0;
322 	mutex_exit(&pq->pq_lock);
323 }
324 
325 /*
326  * pktq_flush: free mbufs in all queues.
327  *
328  * => The caller must ensure there are no concurrent writers or flush calls.
329  */
330 void
331 pktq_flush(pktqueue_t *pq)
332 {
333 	CPU_INFO_ITERATOR cii;
334 	struct cpu_info *ci;
335 	struct mbuf *m;
336 
337 	for (CPU_INFO_FOREACH(cii, ci)) {
338 		struct pcq *q;
339 
340 		kpreempt_disable();
341 		q = pktq_pcq(pq, ci);
342 		kpreempt_enable();
343 
344 		/*
345 		 * XXX This can't be right -- if the softint is running
346 		 * then pcq_get isn't safe here.
347 		 */
348 		while ((m = pcq_get(q)) != NULL) {
349 			pktq_inc_count(pq, PQCNT_DEQUEUE);
350 			m_freem(m);
351 		}
352 	}
353 }
354 
355 static void
356 pktq_set_maxlen_cpu(void *vpq, void *vqs)
357 {
358 	struct pktqueue *pq = vpq;
359 	struct pcq **qp, *q, **qs = vqs;
360 	unsigned i = cpu_index(curcpu());
361 	int s;
362 
363 	s = splnet();
364 	qp = percpu_getref(pq->pq_pcq);
365 	q = *qp;
366 	*qp = qs[i];
367 	qs[i] = q;
368 	percpu_putref(pq->pq_pcq);
369 	splx(s);
370 }
371 
372 /*
373  * pktq_set_maxlen: create per-CPU queues using a new size and replace
374  * the existing queues without losing any packets.
375  *
376  * XXX ncpu must remain stable throughout.
377  */
378 int
379 pktq_set_maxlen(pktqueue_t *pq, size_t maxlen)
380 {
381 	const u_int slotbytes = ncpu * sizeof(pcq_t *);
382 	pcq_t **qs;
383 
384 	if (!maxlen || maxlen > PCQ_MAXLEN)
385 		return EINVAL;
386 	if (pq->pq_maxlen == maxlen)
387 		return 0;
388 
389 	/* First, allocate the new queues. */
390 	qs = kmem_zalloc(slotbytes, KM_SLEEP);
391 	for (u_int i = 0; i < ncpu; i++) {
392 		qs[i] = pcq_create(maxlen, KM_SLEEP);
393 	}
394 
395 	/*
396 	 * Issue an xcall to replace the queue pointers on each CPU.
397 	 * This implies all the necessary memory barriers.
398 	 */
399 	mutex_enter(&pq->pq_lock);
400 	xc_wait(xc_broadcast(XC_HIGHPRI, pktq_set_maxlen_cpu, pq, qs));
401 	pq->pq_maxlen = maxlen;
402 	mutex_exit(&pq->pq_lock);
403 
404 	/*
405 	 * At this point, the new packets are flowing into the new
406 	 * queues.  However, the old queues may have some packets
407 	 * present which are no longer being processed.  We are going
408 	 * to re-enqueue them.  This may change the order of packet
409 	 * arrival, but it is not considered an issue.
410 	 *
411 	 * There may be in-flight interrupts calling pktq_dequeue()
412 	 * which reference the old queues.  Issue a barrier to ensure
413 	 * that we are going to be the only pcq_get() callers on the
414 	 * old queues.
415 	 */
416 	pktq_barrier(pq);
417 
418 	for (u_int i = 0; i < ncpu; i++) {
419 		struct pcq *q;
420 		struct mbuf *m;
421 
422 		kpreempt_disable();
423 		q = pktq_pcq(pq, cpu_lookup(i));
424 		kpreempt_enable();
425 
426 		while ((m = pcq_get(qs[i])) != NULL) {
427 			while (!pcq_put(q, m)) {
428 				kpause("pktqrenq", false, 1, NULL);
429 			}
430 		}
431 		pcq_destroy(qs[i]);
432 	}
433 
434 	/* Well, that was fun. */
435 	kmem_free(qs, slotbytes);
436 	return 0;
437 }
438 
439 int
440 sysctl_pktq_maxlen(SYSCTLFN_ARGS, pktqueue_t *pq)
441 {
442 	u_int nmaxlen = pktq_get_count(pq, PKTQ_MAXLEN);
443 	struct sysctlnode node = *rnode;
444 	int error;
445 
446 	node.sysctl_data = &nmaxlen;
447 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
448 	if (error || newp == NULL)
449 		return error;
450 	return pktq_set_maxlen(pq, nmaxlen);
451 }
452 
453 int
454 sysctl_pktq_count(SYSCTLFN_ARGS, pktqueue_t *pq, u_int count_id)
455 {
456 	uint64_t count = pktq_get_count(pq, count_id);
457 	struct sysctlnode node = *rnode;
458 
459 	node.sysctl_data = &count;
460 	return sysctl_lookup(SYSCTLFN_CALL(&node));
461 }
462