xref: /freebsd/sys/dev/cxgbe/t4_mp_ring.c (revision e0c4386e7e71d93b0edc0c8fa156263fc4a8b0b6)
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/types.h>
29 #include <sys/param.h>
30 #include <sys/systm.h>
31 #include <sys/counter.h>
32 #include <sys/lock.h>
33 #include <sys/malloc.h>
34 #include <sys/mutex.h>
35 #include <sys/sysctl.h>
36 #include <machine/cpu.h>
37 
38 #include "t4_mp_ring.h"
39 
40 #if defined(__i386__)
41 #define atomic_cmpset_acq_64 atomic_cmpset_64
42 #define atomic_cmpset_rel_64 atomic_cmpset_64
43 #endif
44 
45 /*
46  * mp_ring handles multiple threads (producers) enqueueing data to a tx queue.
47  * The thread that is writing the hardware descriptors is the consumer and it
48  * runs with the consumer lock held.  A producer becomes the consumer if there
49  * isn't one already.  The consumer runs with the flags sets to BUSY and
50  * consumes everything (IDLE or COALESCING) or gets STALLED.  If it is running
51  * over its budget it sets flags to TOO_BUSY.  A producer that observes a
52  * TOO_BUSY consumer will become the new consumer by setting flags to
53  * TAKING_OVER.  The original consumer stops and sets the flags back to BUSY for
54  * the new consumer.
55  *
56  * COALESCING is the same as IDLE except there are items being held in the hope
57  * that they can be coalesced with items that follow.  The driver must arrange
58  * for a tx update or some other event that transmits all the held items in a
59  * timely manner if nothing else is enqueued.
60  */
61 
62 union ring_state {
63 	struct {
64 		uint16_t pidx_head;
65 		uint16_t pidx_tail;
66 		uint16_t cidx;
67 		uint16_t flags;
68 	};
69 	uint64_t state;
70 };
71 
72 enum {
73 	IDLE = 0,	/* tx is all caught up, nothing to do. */
74 	COALESCING,	/* IDLE, but tx frames are being held for coalescing */
75 	BUSY,		/* consumer is running already, or will be shortly. */
76 	TOO_BUSY,	/* consumer is running and is beyond its budget */
77 	TAKING_OVER,	/* new consumer taking over from a TOO_BUSY consumer */
78 	STALLED,	/* consumer stopped due to lack of resources. */
79 };
80 
81 enum {
82 	C_FAST = 0,
83 	C_2,
84 	C_3,
85 	C_TAKEOVER,
86 };
87 
88 static inline uint16_t
89 space_available(struct mp_ring *r, union ring_state s)
90 {
91 	uint16_t x = r->size - 1;
92 
93 	if (s.cidx == s.pidx_head)
94 		return (x);
95 	else if (s.cidx > s.pidx_head)
96 		return (s.cidx - s.pidx_head - 1);
97 	else
98 		return (x - s.pidx_head + s.cidx);
99 }
100 
101 static inline uint16_t
102 increment_idx(struct mp_ring *r, uint16_t idx, uint16_t n)
103 {
104 	int x = r->size - idx;
105 
106 	MPASS(x > 0);
107 	return (x > n ? idx + n : n - x);
108 }
109 
110 /*
111  * Consumer.  Called with the consumer lock held and a guarantee that there is
112  * work to do.
113  */
114 static void
115 drain_ring(struct mp_ring *r, int budget)
116 {
117 	union ring_state os, ns;
118 	int n, pending, total;
119 	uint16_t cidx;
120 	uint16_t pidx;
121 	bool coalescing;
122 
123 	mtx_assert(r->cons_lock, MA_OWNED);
124 
125 	os.state = atomic_load_acq_64(&r->state);
126 	MPASS(os.flags == BUSY);
127 
128 	cidx = os.cidx;
129 	pidx = os.pidx_tail;
130 	MPASS(cidx != pidx);
131 
132 	pending = 0;
133 	total = 0;
134 
135 	while (cidx != pidx) {
136 
137 		/* Items from cidx to pidx are available for consumption. */
138 		n = r->drain(r, cidx, pidx, &coalescing);
139 		if (n == 0) {
140 			critical_enter();
141 			os.state = atomic_load_64(&r->state);
142 			do {
143 				ns.state = os.state;
144 				ns.cidx = cidx;
145 
146 				MPASS(os.flags == BUSY ||
147 				    os.flags == TOO_BUSY ||
148 				    os.flags == TAKING_OVER);
149 
150 				if (os.flags == TAKING_OVER)
151 					ns.flags = BUSY;
152 				else
153 					ns.flags = STALLED;
154 			} while (atomic_fcmpset_64(&r->state, &os.state,
155 			    ns.state) == 0);
156 			critical_exit();
157 			if (os.flags == TAKING_OVER)
158 				counter_u64_add(r->abdications, 1);
159 			else if (ns.flags == STALLED)
160 				counter_u64_add(r->stalls, 1);
161 			break;
162 		}
163 		cidx = increment_idx(r, cidx, n);
164 		pending += n;
165 		total += n;
166 		counter_u64_add(r->consumed, n);
167 
168 		os.state = atomic_load_64(&r->state);
169 		do {
170 			MPASS(os.flags == BUSY || os.flags == TOO_BUSY ||
171 			    os.flags == TAKING_OVER);
172 
173 			ns.state = os.state;
174 			ns.cidx = cidx;
175 			if (__predict_false(os.flags == TAKING_OVER)) {
176 				MPASS(total >= budget);
177 				ns.flags = BUSY;
178 				continue;
179 			}
180 			if (cidx == os.pidx_tail) {
181 				ns.flags = coalescing ? COALESCING : IDLE;
182 				continue;
183 			}
184 			if (total >= budget) {
185 				ns.flags = TOO_BUSY;
186 				continue;
187 			}
188 			MPASS(os.flags == BUSY);
189 			if (pending < 32)
190 				break;
191 		} while (atomic_fcmpset_acq_64(&r->state, &os.state, ns.state) == 0);
192 
193 		if (__predict_false(os.flags == TAKING_OVER)) {
194 			MPASS(ns.flags == BUSY);
195 			counter_u64_add(r->abdications, 1);
196 			break;
197 		}
198 
199 		if (ns.flags == IDLE || ns.flags == COALESCING) {
200 			MPASS(ns.pidx_tail == cidx);
201 			if (ns.pidx_head != ns.pidx_tail)
202 				counter_u64_add(r->cons_idle2, 1);
203 			else
204 				counter_u64_add(r->cons_idle, 1);
205 			break;
206 		}
207 
208 		/*
209 		 * The acquire style atomic above guarantees visibility of items
210 		 * associated with any pidx change that we notice here.
211 		 */
212 		pidx = ns.pidx_tail;
213 		pending = 0;
214 	}
215 
216 #ifdef INVARIANTS
217 	if (os.flags == TAKING_OVER)
218 		MPASS(ns.flags == BUSY);
219 	else {
220 		MPASS(ns.flags == IDLE || ns.flags == COALESCING ||
221 		    ns.flags == STALLED);
222 	}
223 #endif
224 }
225 
226 static void
227 drain_txpkts(struct mp_ring *r, union ring_state os, int budget)
228 {
229 	union ring_state ns;
230 	uint16_t cidx = os.cidx;
231 	uint16_t pidx = os.pidx_tail;
232 	bool coalescing;
233 
234 	mtx_assert(r->cons_lock, MA_OWNED);
235 	MPASS(os.flags == BUSY);
236 	MPASS(cidx == pidx);
237 
238 	r->drain(r, cidx, pidx, &coalescing);
239 	MPASS(coalescing == false);
240 	critical_enter();
241 	os.state = atomic_load_64(&r->state);
242 	do {
243 		ns.state = os.state;
244 		MPASS(os.flags == BUSY);
245 		MPASS(os.cidx == cidx);
246 		if (ns.cidx == ns.pidx_tail)
247 			ns.flags = IDLE;
248 		else
249 			ns.flags = BUSY;
250 	} while (atomic_fcmpset_acq_64(&r->state, &os.state, ns.state) == 0);
251 	critical_exit();
252 
253 	if (ns.flags == BUSY)
254 		drain_ring(r, budget);
255 }
256 
257 int
258 mp_ring_alloc(struct mp_ring **pr, int size, void *cookie, ring_drain_t drain,
259     ring_can_drain_t can_drain, struct malloc_type *mt, struct mtx *lck,
260     int flags)
261 {
262 	struct mp_ring *r;
263 	int i;
264 
265 	/* All idx are 16b so size can be 65536 at most */
266 	if (pr == NULL || size < 2 || size > 65536 || drain == NULL ||
267 	    can_drain == NULL)
268 		return (EINVAL);
269 	*pr = NULL;
270 	flags &= M_NOWAIT | M_WAITOK;
271 	MPASS(flags != 0);
272 
273 	r = malloc(__offsetof(struct mp_ring, items[size]), mt, flags | M_ZERO);
274 	if (r == NULL)
275 		return (ENOMEM);
276 	r->size = size;
277 	r->cookie = cookie;
278 	r->mt = mt;
279 	r->drain = drain;
280 	r->can_drain = can_drain;
281 	r->cons_lock = lck;
282 	if ((r->dropped = counter_u64_alloc(flags)) == NULL)
283 		goto failed;
284 	for (i = 0; i < nitems(r->consumer); i++) {
285 		if ((r->consumer[i] = counter_u64_alloc(flags)) == NULL)
286 			goto failed;
287 	}
288 	if ((r->not_consumer = counter_u64_alloc(flags)) == NULL)
289 		goto failed;
290 	if ((r->abdications = counter_u64_alloc(flags)) == NULL)
291 		goto failed;
292 	if ((r->stalls = counter_u64_alloc(flags)) == NULL)
293 		goto failed;
294 	if ((r->consumed = counter_u64_alloc(flags)) == NULL)
295 		goto failed;
296 	if ((r->cons_idle = counter_u64_alloc(flags)) == NULL)
297 		goto failed;
298 	if ((r->cons_idle2 = counter_u64_alloc(flags)) == NULL)
299 		goto failed;
300 	*pr = r;
301 	return (0);
302 failed:
303 	mp_ring_free(r);
304 	return (ENOMEM);
305 }
306 
307 void
308 
309 mp_ring_free(struct mp_ring *r)
310 {
311 	int i;
312 
313 	if (r == NULL)
314 		return;
315 
316 	if (r->dropped != NULL)
317 		counter_u64_free(r->dropped);
318 	for (i = 0; i < nitems(r->consumer); i++) {
319 		if (r->consumer[i] != NULL)
320 			counter_u64_free(r->consumer[i]);
321 	}
322 	if (r->not_consumer != NULL)
323 		counter_u64_free(r->not_consumer);
324 	if (r->abdications != NULL)
325 		counter_u64_free(r->abdications);
326 	if (r->stalls != NULL)
327 		counter_u64_free(r->stalls);
328 	if (r->consumed != NULL)
329 		counter_u64_free(r->consumed);
330 	if (r->cons_idle != NULL)
331 		counter_u64_free(r->cons_idle);
332 	if (r->cons_idle2 != NULL)
333 		counter_u64_free(r->cons_idle2);
334 
335 	free(r, r->mt);
336 }
337 
338 /*
339  * Enqueue n items and maybe drain the ring for some time.
340  *
341  * Returns an errno.
342  */
343 int
344 mp_ring_enqueue(struct mp_ring *r, void **items, int n, int budget)
345 {
346 	union ring_state os, ns;
347 	uint16_t pidx_start, pidx_stop;
348 	int i, nospc, cons;
349 	bool consumer;
350 
351 	MPASS(items != NULL);
352 	MPASS(n > 0);
353 
354 	/*
355 	 * Reserve room for the new items.  Our reservation, if successful, is
356 	 * from 'pidx_start' to 'pidx_stop'.
357 	 */
358 	nospc = 0;
359 	os.state = atomic_load_64(&r->state);
360 	for (;;) {
361 		for (;;) {
362 			if (__predict_true(space_available(r, os) >= n))
363 				break;
364 
365 			/* Not enough room in the ring. */
366 
367 			MPASS(os.flags != IDLE);
368 			MPASS(os.flags != COALESCING);
369 			if (__predict_false(++nospc > 100)) {
370 				counter_u64_add(r->dropped, n);
371 				return (ENOBUFS);
372 			}
373 			if (os.flags == STALLED)
374 				mp_ring_check_drainage(r, 64);
375 			else
376 				cpu_spinwait();
377 			os.state = atomic_load_64(&r->state);
378 		}
379 
380 		/* There is room in the ring. */
381 
382 		cons = -1;
383 		ns.state = os.state;
384 		ns.pidx_head = increment_idx(r, os.pidx_head, n);
385 		if (os.flags == IDLE || os.flags == COALESCING) {
386 			MPASS(os.pidx_tail == os.cidx);
387 			if (os.pidx_head == os.pidx_tail) {
388 				cons = C_FAST;
389 				ns.pidx_tail = increment_idx(r, os.pidx_tail, n);
390 			} else
391 				cons = C_2;
392 			ns.flags = BUSY;
393 		} else if (os.flags == TOO_BUSY) {
394 			cons = C_TAKEOVER;
395 			ns.flags = TAKING_OVER;
396 		}
397 		critical_enter();
398 		if (atomic_fcmpset_64(&r->state, &os.state, ns.state))
399 			break;
400 		critical_exit();
401 		cpu_spinwait();
402 	};
403 
404 	pidx_start = os.pidx_head;
405 	pidx_stop = ns.pidx_head;
406 
407 	if (cons == C_FAST) {
408 		i = pidx_start;
409 		do {
410 			r->items[i] = *items++;
411 			if (__predict_false(++i == r->size))
412 				i = 0;
413 		} while (i != pidx_stop);
414 		critical_exit();
415 		counter_u64_add(r->consumer[C_FAST], 1);
416 		mtx_lock(r->cons_lock);
417 		drain_ring(r, budget);
418 		mtx_unlock(r->cons_lock);
419 		return (0);
420 	}
421 
422 	/*
423 	 * Wait for other producers who got in ahead of us to enqueue their
424 	 * items, one producer at a time.  It is our turn when the ring's
425 	 * pidx_tail reaches the beginning of our reservation (pidx_start).
426 	 */
427 	while (ns.pidx_tail != pidx_start) {
428 		cpu_spinwait();
429 		ns.state = atomic_load_64(&r->state);
430 	}
431 
432 	/* Now it is our turn to fill up the area we reserved earlier. */
433 	i = pidx_start;
434 	do {
435 		r->items[i] = *items++;
436 		if (__predict_false(++i == r->size))
437 			i = 0;
438 	} while (i != pidx_stop);
439 
440 	/*
441 	 * Update the ring's pidx_tail.  The release style atomic guarantees
442 	 * that the items are visible to any thread that sees the updated pidx.
443 	 */
444 	os.state = atomic_load_64(&r->state);
445 	do {
446 		consumer = false;
447 		ns.state = os.state;
448 		ns.pidx_tail = pidx_stop;
449 		if (os.flags == IDLE || os.flags == COALESCING ||
450 		    (os.flags == STALLED && r->can_drain(r))) {
451 			MPASS(cons == -1);
452 			consumer = true;
453 			ns.flags = BUSY;
454 		}
455 	} while (atomic_fcmpset_rel_64(&r->state, &os.state, ns.state) == 0);
456 	critical_exit();
457 
458 	if (cons == -1) {
459 		if (consumer)
460 			cons = C_3;
461 		else {
462 			counter_u64_add(r->not_consumer, 1);
463 			return (0);
464 		}
465 	}
466 	MPASS(cons > C_FAST && cons < nitems(r->consumer));
467 	counter_u64_add(r->consumer[cons], 1);
468 	mtx_lock(r->cons_lock);
469 	drain_ring(r, budget);
470 	mtx_unlock(r->cons_lock);
471 
472 	return (0);
473 }
474 
475 void
476 mp_ring_check_drainage(struct mp_ring *r, int budget)
477 {
478 	union ring_state os, ns;
479 
480 	os.state = atomic_load_64(&r->state);
481 	if (os.flags == STALLED && r->can_drain(r)) {
482 		MPASS(os.cidx != os.pidx_tail);	/* implied by STALLED */
483 		ns.state = os.state;
484 		ns.flags = BUSY;
485 		if (atomic_cmpset_acq_64(&r->state, os.state, ns.state)) {
486 			mtx_lock(r->cons_lock);
487 			drain_ring(r, budget);
488 			mtx_unlock(r->cons_lock);
489 		}
490 	} else if (os.flags == COALESCING) {
491 		MPASS(os.cidx == os.pidx_tail);
492 		ns.state = os.state;
493 		ns.flags = BUSY;
494 		if (atomic_cmpset_acq_64(&r->state, os.state, ns.state)) {
495 			mtx_lock(r->cons_lock);
496 			drain_txpkts(r, ns, budget);
497 			mtx_unlock(r->cons_lock);
498 		}
499 	}
500 }
501 
502 void
503 mp_ring_reset_stats(struct mp_ring *r)
504 {
505 	int i;
506 
507 	counter_u64_zero(r->dropped);
508 	for (i = 0; i < nitems(r->consumer); i++)
509 		counter_u64_zero(r->consumer[i]);
510 	counter_u64_zero(r->not_consumer);
511 	counter_u64_zero(r->abdications);
512 	counter_u64_zero(r->stalls);
513 	counter_u64_zero(r->consumed);
514 	counter_u64_zero(r->cons_idle);
515 	counter_u64_zero(r->cons_idle2);
516 }
517 
518 bool
519 mp_ring_is_idle(struct mp_ring *r)
520 {
521 	union ring_state s;
522 
523 	s.state = atomic_load_64(&r->state);
524 	if (s.pidx_head == s.pidx_tail && s.pidx_tail == s.cidx &&
525 	    s.flags == IDLE)
526 		return (true);
527 
528 	return (false);
529 }
530 
531 void
532 mp_ring_sysctls(struct mp_ring *r, struct sysctl_ctx_list *ctx,
533     struct sysctl_oid_list *children)
534 {
535 	struct sysctl_oid *oid;
536 
537 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "mp_ring", CTLFLAG_RD |
538 	    CTLFLAG_MPSAFE, NULL, "mp_ring statistics");
539 	children = SYSCTL_CHILDREN(oid);
540 
541 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "state", CTLFLAG_RD,
542 	    __DEVOLATILE(uint64_t *, &r->state), 0, "ring state");
543 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "dropped", CTLFLAG_RD,
544 	    &r->dropped, "# of items dropped");
545 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "consumed",
546 	    CTLFLAG_RD, &r->consumed, "# of items consumed");
547 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "fast_consumer",
548 	    CTLFLAG_RD, &r->consumer[C_FAST],
549 	    "# of times producer became consumer (fast)");
550 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "consumer2",
551 	    CTLFLAG_RD, &r->consumer[C_2],
552 	    "# of times producer became consumer (2)");
553 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "consumer3",
554 	    CTLFLAG_RD, &r->consumer[C_3],
555 	    "# of times producer became consumer (3)");
556 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "takeovers",
557 	    CTLFLAG_RD, &r->consumer[C_TAKEOVER],
558 	    "# of times producer took over from another consumer.");
559 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "not_consumer",
560 	    CTLFLAG_RD, &r->not_consumer,
561 	    "# of times producer did not become consumer");
562 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "abdications",
563 	    CTLFLAG_RD, &r->abdications, "# of consumer abdications");
564 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "stalls",
565 	    CTLFLAG_RD, &r->stalls, "# of consumer stalls");
566 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "cons_idle",
567 	    CTLFLAG_RD, &r->cons_idle,
568 	    "# of times consumer ran fully to completion");
569 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "cons_idle2",
570 	    CTLFLAG_RD, &r->cons_idle2,
571 	    "# of times consumer idled when another enqueue was in progress");
572 }
573