xref: /freebsd/sys/net/mp_ring.c (revision 273c26a3c3bea87a241d6879abd4f991db180bf0)
1 /*-
2  * Copyright (c) 2014 Chelsio Communications, Inc.
3  * All rights reserved.
4  * Written by: Navdeep Parhar <np@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include <sys/types.h>
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/counter.h>
35 #include <sys/lock.h>
36 #include <sys/mutex.h>
37 #include <sys/malloc.h>
38 #include <machine/cpu.h>
39 
40 #if defined(__powerpc__) || defined(__mips__)
41 #define NO_64BIT_ATOMICS
42 #endif
43 
44 #if defined(__i386__)
45 #define atomic_cmpset_acq_64 atomic_cmpset_64
46 #define atomic_cmpset_rel_64 atomic_cmpset_64
47 #endif
48 
49 #include <net/mp_ring.h>
50 
51 union ring_state {
52 	struct {
53 		uint16_t pidx_head;
54 		uint16_t pidx_tail;
55 		uint16_t cidx;
56 		uint16_t flags;
57 	};
58 	uint64_t state;
59 };
60 
61 enum {
62 	IDLE = 0,	/* consumer ran to completion, nothing more to do. */
63 	BUSY,		/* consumer is running already, or will be shortly. */
64 	STALLED,	/* consumer stopped due to lack of resources. */
65 	ABDICATED,	/* consumer stopped even though there was work to be
66 			   done because it wants another thread to take over. */
67 };
68 
69 static inline uint16_t
70 space_available(struct ifmp_ring *r, union ring_state s)
71 {
72 	uint16_t x = r->size - 1;
73 
74 	if (s.cidx == s.pidx_head)
75 		return (x);
76 	else if (s.cidx > s.pidx_head)
77 		return (s.cidx - s.pidx_head - 1);
78 	else
79 		return (x - s.pidx_head + s.cidx);
80 }
81 
82 static inline uint16_t
83 increment_idx(struct ifmp_ring *r, uint16_t idx, uint16_t n)
84 {
85 	int x = r->size - idx;
86 
87 	MPASS(x > 0);
88 	return (x > n ? idx + n : n - x);
89 }
90 
91 /* Consumer is about to update the ring's state to s */
92 static inline uint16_t
93 state_to_flags(union ring_state s, int abdicate)
94 {
95 
96 	if (s.cidx == s.pidx_tail)
97 		return (IDLE);
98 	else if (abdicate && s.pidx_tail != s.pidx_head)
99 		return (ABDICATED);
100 
101 	return (BUSY);
102 }
103 
104 #ifdef NO_64BIT_ATOMICS
105 static void
106 drain_ring_locked(struct ifmp_ring *r, union ring_state os, uint16_t prev, int budget)
107 {
108 	union ring_state ns;
109 	int n, pending, total;
110 	uint16_t cidx = os.cidx;
111 	uint16_t pidx = os.pidx_tail;
112 
113 	MPASS(os.flags == BUSY);
114 	MPASS(cidx != pidx);
115 
116 	if (prev == IDLE)
117 		counter_u64_add(r->starts, 1);
118 	pending = 0;
119 	total = 0;
120 
121 	while (cidx != pidx) {
122 
123 		/* Items from cidx to pidx are available for consumption. */
124 		n = r->drain(r, cidx, pidx);
125 		if (n == 0) {
126 			os.state = ns.state = r->state;
127 			ns.cidx = cidx;
128 			ns.flags = STALLED;
129 			r->state = ns.state;
130 			if (prev != STALLED)
131 				counter_u64_add(r->stalls, 1);
132 			else if (total > 0) {
133 				counter_u64_add(r->restarts, 1);
134 				counter_u64_add(r->stalls, 1);
135 			}
136 			break;
137 		}
138 		cidx = increment_idx(r, cidx, n);
139 		pending += n;
140 		total += n;
141 
142 		/*
143 		 * We update the cidx only if we've caught up with the pidx, the
144 		 * real cidx is getting too far ahead of the one visible to
145 		 * everyone else, or we have exceeded our budget.
146 		 */
147 		if (cidx != pidx && pending < 64 && total < budget)
148 			continue;
149 
150 		os.state = ns.state = r->state;
151 		ns.cidx = cidx;
152 		ns.flags = state_to_flags(ns, total >= budget);
153 		r->state = ns.state;
154 
155 		if (ns.flags == ABDICATED)
156 			counter_u64_add(r->abdications, 1);
157 		if (ns.flags != BUSY) {
158 			/* Wrong loop exit if we're going to stall. */
159 			MPASS(ns.flags != STALLED);
160 			if (prev == STALLED) {
161 				MPASS(total > 0);
162 				counter_u64_add(r->restarts, 1);
163 			}
164 			break;
165 		}
166 
167 		/*
168 		 * The acquire style atomic above guarantees visibility of items
169 		 * associated with any pidx change that we notice here.
170 		 */
171 		pidx = ns.pidx_tail;
172 		pending = 0;
173 	}
174 }
175 #else
176 /*
177  * Caller passes in a state, with a guarantee that there is work to do and that
178  * all items up to the pidx_tail in the state are visible.
179  */
180 static void
181 drain_ring_lockless(struct ifmp_ring *r, union ring_state os, uint16_t prev, int budget)
182 {
183 	union ring_state ns;
184 	int n, pending, total;
185 	uint16_t cidx = os.cidx;
186 	uint16_t pidx = os.pidx_tail;
187 
188 	MPASS(os.flags == BUSY);
189 	MPASS(cidx != pidx);
190 
191 	if (prev == IDLE)
192 		counter_u64_add(r->starts, 1);
193 	pending = 0;
194 	total = 0;
195 
196 	while (cidx != pidx) {
197 
198 		/* Items from cidx to pidx are available for consumption. */
199 		n = r->drain(r, cidx, pidx);
200 		if (n == 0) {
201 			critical_enter();
202 			do {
203 				os.state = ns.state = r->state;
204 				ns.cidx = cidx;
205 				ns.flags = STALLED;
206 			} while (atomic_cmpset_64(&r->state, os.state,
207 			    ns.state) == 0);
208 			critical_exit();
209 			if (prev != STALLED)
210 				counter_u64_add(r->stalls, 1);
211 			else if (total > 0) {
212 				counter_u64_add(r->restarts, 1);
213 				counter_u64_add(r->stalls, 1);
214 			}
215 			break;
216 		}
217 		cidx = increment_idx(r, cidx, n);
218 		pending += n;
219 		total += n;
220 
221 		/*
222 		 * We update the cidx only if we've caught up with the pidx, the
223 		 * real cidx is getting too far ahead of the one visible to
224 		 * everyone else, or we have exceeded our budget.
225 		 */
226 		if (cidx != pidx && pending < 64 && total < budget)
227 			continue;
228 		critical_enter();
229 		do {
230 			os.state = ns.state = r->state;
231 			ns.cidx = cidx;
232 			ns.flags = state_to_flags(ns, total >= budget);
233 		} while (atomic_cmpset_acq_64(&r->state, os.state, ns.state) == 0);
234 		critical_exit();
235 
236 		if (ns.flags == ABDICATED)
237 			counter_u64_add(r->abdications, 1);
238 		if (ns.flags != BUSY) {
239 			/* Wrong loop exit if we're going to stall. */
240 			MPASS(ns.flags != STALLED);
241 			if (prev == STALLED) {
242 				MPASS(total > 0);
243 				counter_u64_add(r->restarts, 1);
244 			}
245 			break;
246 		}
247 
248 		/*
249 		 * The acquire style atomic above guarantees visibility of items
250 		 * associated with any pidx change that we notice here.
251 		 */
252 		pidx = ns.pidx_tail;
253 		pending = 0;
254 	}
255 }
256 #endif
257 
258 int
259 ifmp_ring_alloc(struct ifmp_ring **pr, int size, void *cookie, mp_ring_drain_t drain,
260     mp_ring_can_drain_t can_drain, struct malloc_type *mt, int flags)
261 {
262 	struct ifmp_ring *r;
263 
264 	/* All idx are 16b so size can be 65536 at most */
265 	if (pr == NULL || size < 2 || size > 65536 || drain == NULL ||
266 	    can_drain == NULL)
267 		return (EINVAL);
268 	*pr = NULL;
269 	flags &= M_NOWAIT | M_WAITOK;
270 	MPASS(flags != 0);
271 
272 	r = malloc(__offsetof(struct ifmp_ring, items[size]), mt, flags | M_ZERO);
273 	if (r == NULL)
274 		return (ENOMEM);
275 	r->size = size;
276 	r->cookie = cookie;
277 	r->mt = mt;
278 	r->drain = drain;
279 	r->can_drain = can_drain;
280 	r->enqueues = counter_u64_alloc(flags);
281 	r->drops = counter_u64_alloc(flags);
282 	r->starts = counter_u64_alloc(flags);
283 	r->stalls = counter_u64_alloc(flags);
284 	r->restarts = counter_u64_alloc(flags);
285 	r->abdications = counter_u64_alloc(flags);
286 	if (r->enqueues == NULL || r->drops == NULL || r->starts == NULL ||
287 	    r->stalls == NULL || r->restarts == NULL ||
288 	    r->abdications == NULL) {
289 		ifmp_ring_free(r);
290 		return (ENOMEM);
291 	}
292 
293 	*pr = r;
294 #ifdef NO_64BIT_ATOMICS
295 	mtx_init(&r->lock, "mp_ring lock", NULL, MTX_DEF);
296 #endif
297 	return (0);
298 }
299 
300 void
301 ifmp_ring_free(struct ifmp_ring *r)
302 {
303 
304 	if (r == NULL)
305 		return;
306 
307 	if (r->enqueues != NULL)
308 		counter_u64_free(r->enqueues);
309 	if (r->drops != NULL)
310 		counter_u64_free(r->drops);
311 	if (r->starts != NULL)
312 		counter_u64_free(r->starts);
313 	if (r->stalls != NULL)
314 		counter_u64_free(r->stalls);
315 	if (r->restarts != NULL)
316 		counter_u64_free(r->restarts);
317 	if (r->abdications != NULL)
318 		counter_u64_free(r->abdications);
319 
320 	free(r, r->mt);
321 }
322 
323 /*
324  * Enqueue n items and maybe drain the ring for some time.
325  *
326  * Returns an errno.
327  */
328 #ifdef NO_64BIT_ATOMICS
329 int
330 ifmp_ring_enqueue(struct ifmp_ring *r, void **items, int n, int budget)
331 {
332 	union ring_state os, ns;
333 	uint16_t pidx_start, pidx_stop;
334 	int i;
335 
336 	MPASS(items != NULL);
337 	MPASS(n > 0);
338 
339 	mtx_lock(&r->lock);
340 	/*
341 	 * Reserve room for the new items.  Our reservation, if successful, is
342 	 * from 'pidx_start' to 'pidx_stop'.
343 	 */
344 	os.state = r->state;
345 	if (n >= space_available(r, os)) {
346 		counter_u64_add(r->drops, n);
347 		MPASS(os.flags != IDLE);
348 		if (os.flags == STALLED)
349 			ifmp_ring_check_drainage(r, 0);
350 		return (ENOBUFS);
351 	}
352 	ns.state = os.state;
353 	ns.pidx_head = increment_idx(r, os.pidx_head, n);
354 	r->state = ns.state;
355 	pidx_start = os.pidx_head;
356 	pidx_stop = ns.pidx_head;
357 
358 	/*
359 	 * Wait for other producers who got in ahead of us to enqueue their
360 	 * items, one producer at a time.  It is our turn when the ring's
361 	 * pidx_tail reaches the beginning of our reservation (pidx_start).
362 	 */
363 	while (ns.pidx_tail != pidx_start) {
364 		cpu_spinwait();
365 		ns.state = r->state;
366 	}
367 
368 	/* Now it is our turn to fill up the area we reserved earlier. */
369 	i = pidx_start;
370 	do {
371 		r->items[i] = *items++;
372 		if (__predict_false(++i == r->size))
373 			i = 0;
374 	} while (i != pidx_stop);
375 
376 	/*
377 	 * Update the ring's pidx_tail.  The release style atomic guarantees
378 	 * that the items are visible to any thread that sees the updated pidx.
379 	 */
380 	os.state = ns.state = r->state;
381 	ns.pidx_tail = pidx_stop;
382 	ns.flags = BUSY;
383 	r->state = ns.state;
384 	counter_u64_add(r->enqueues, n);
385 
386 	/*
387 	 * Turn into a consumer if some other thread isn't active as a consumer
388 	 * already.
389 	 */
390 	if (os.flags != BUSY)
391 		drain_ring_locked(r, ns, os.flags, budget);
392 
393 	mtx_unlock(&r->lock);
394 	return (0);
395 }
396 
397 #else
398 int
399 ifmp_ring_enqueue(struct ifmp_ring *r, void **items, int n, int budget)
400 {
401 	union ring_state os, ns;
402 	uint16_t pidx_start, pidx_stop;
403 	int i;
404 
405 	MPASS(items != NULL);
406 	MPASS(n > 0);
407 
408 	/*
409 	 * Reserve room for the new items.  Our reservation, if successful, is
410 	 * from 'pidx_start' to 'pidx_stop'.
411 	 */
412 	for (;;) {
413 		os.state = r->state;
414 		if (n >= space_available(r, os)) {
415 			counter_u64_add(r->drops, n);
416 			MPASS(os.flags != IDLE);
417 			if (os.flags == STALLED)
418 				ifmp_ring_check_drainage(r, 0);
419 			return (ENOBUFS);
420 		}
421 		ns.state = os.state;
422 		ns.pidx_head = increment_idx(r, os.pidx_head, n);
423 		critical_enter();
424 		if (atomic_cmpset_64(&r->state, os.state, ns.state))
425 			break;
426 		critical_exit();
427 		cpu_spinwait();
428 	}
429 	pidx_start = os.pidx_head;
430 	pidx_stop = ns.pidx_head;
431 
432 	/*
433 	 * Wait for other producers who got in ahead of us to enqueue their
434 	 * items, one producer at a time.  It is our turn when the ring's
435 	 * pidx_tail reaches the beginning of our reservation (pidx_start).
436 	 */
437 	while (ns.pidx_tail != pidx_start) {
438 		cpu_spinwait();
439 		ns.state = r->state;
440 	}
441 
442 	/* Now it is our turn to fill up the area we reserved earlier. */
443 	i = pidx_start;
444 	do {
445 		r->items[i] = *items++;
446 		if (__predict_false(++i == r->size))
447 			i = 0;
448 	} while (i != pidx_stop);
449 
450 	/*
451 	 * Update the ring's pidx_tail.  The release style atomic guarantees
452 	 * that the items are visible to any thread that sees the updated pidx.
453 	 */
454 	do {
455 		os.state = ns.state = r->state;
456 		ns.pidx_tail = pidx_stop;
457 		ns.flags = BUSY;
458 	} while (atomic_cmpset_rel_64(&r->state, os.state, ns.state) == 0);
459 	critical_exit();
460 	counter_u64_add(r->enqueues, n);
461 
462 	/*
463 	 * Turn into a consumer if some other thread isn't active as a consumer
464 	 * already.
465 	 */
466 	if (os.flags != BUSY)
467 		drain_ring_lockless(r, ns, os.flags, budget);
468 
469 	return (0);
470 }
471 #endif
472 
473 void
474 ifmp_ring_check_drainage(struct ifmp_ring *r, int budget)
475 {
476 	union ring_state os, ns;
477 
478 	os.state = r->state;
479 	if (os.flags != STALLED || os.pidx_head != os.pidx_tail || r->can_drain(r) == 0)
480 		return;
481 
482 	MPASS(os.cidx != os.pidx_tail);	/* implied by STALLED */
483 	ns.state = os.state;
484 	ns.flags = BUSY;
485 
486 
487 #ifdef NO_64BIT_ATOMICS
488 	mtx_lock(&r->lock);
489 	if (r->state != os.state) {
490 		mtx_unlock(&r->lock);
491 		return;
492 	}
493 	r->state = ns.state;
494 	drain_ring_locked(r, ns, os.flags, budget);
495 	mtx_unlock(&r->lock);
496 #else
497 	/*
498 	 * The acquire style atomic guarantees visibility of items associated
499 	 * with the pidx that we read here.
500 	 */
501 	if (!atomic_cmpset_acq_64(&r->state, os.state, ns.state))
502 		return;
503 
504 
505 	drain_ring_lockless(r, ns, os.flags, budget);
506 #endif
507 }
508 
509 void
510 ifmp_ring_reset_stats(struct ifmp_ring *r)
511 {
512 
513 	counter_u64_zero(r->enqueues);
514 	counter_u64_zero(r->drops);
515 	counter_u64_zero(r->starts);
516 	counter_u64_zero(r->stalls);
517 	counter_u64_zero(r->restarts);
518 	counter_u64_zero(r->abdications);
519 }
520 
521 int
522 ifmp_ring_is_idle(struct ifmp_ring *r)
523 {
524 	union ring_state s;
525 
526 	s.state = r->state;
527 	if (s.pidx_head == s.pidx_tail && s.pidx_tail == s.cidx &&
528 	    s.flags == IDLE)
529 		return (1);
530 
531 	return (0);
532 }
533 
534 int
535 ifmp_ring_is_stalled(struct ifmp_ring *r)
536 {
537 	union ring_state s;
538 
539 	s.state = r->state;
540 	if (s.pidx_head == s.pidx_tail && s.flags == STALLED)
541 		return (1);
542 
543 	return (0);
544 }
545