xref: /freebsd/sys/dev/cxgbe/t4_mp_ring.c (revision eda14cbc264d6969b02f2b1994cef11148e914f1)
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/malloc.h>
37 #include <sys/mutex.h>
38 #include <sys/sysctl.h>
39 #include <machine/cpu.h>
40 
41 #include "t4_mp_ring.h"
42 
43 #if defined(__i386__)
44 #define atomic_cmpset_acq_64 atomic_cmpset_64
45 #define atomic_cmpset_rel_64 atomic_cmpset_64
46 #endif
47 
48 /*
49  * mp_ring handles multiple threads (producers) enqueueing data to a tx queue.
50  * The thread that is writing the hardware descriptors is the consumer and it
51  * runs with the consumer lock held.  A producer becomes the consumer if there
52  * isn't one already.  The consumer runs with the flags sets to BUSY and
53  * consumes everything (IDLE or COALESCING) or gets STALLED.  If it is running
54  * over its budget it sets flags to TOO_BUSY.  A producer that observes a
55  * TOO_BUSY consumer will become the new consumer by setting flags to
56  * TAKING_OVER.  The original consumer stops and sets the flags back to BUSY for
57  * the new consumer.
58  *
59  * COALESCING is the same as IDLE except there are items being held in the hope
60  * that they can be coalesced with items that follow.  The driver must arrange
61  * for a tx update or some other event that transmits all the held items in a
62  * timely manner if nothing else is enqueued.
63  */
64 
65 union ring_state {
66 	struct {
67 		uint16_t pidx_head;
68 		uint16_t pidx_tail;
69 		uint16_t cidx;
70 		uint16_t flags;
71 	};
72 	uint64_t state;
73 };
74 
75 enum {
76 	IDLE = 0,	/* tx is all caught up, nothing to do. */
77 	COALESCING,	/* IDLE, but tx frames are being held for coalescing */
78 	BUSY,		/* consumer is running already, or will be shortly. */
79 	TOO_BUSY,	/* consumer is running and is beyond its budget */
80 	TAKING_OVER,	/* new consumer taking over from a TOO_BUSY consumer */
81 	STALLED,	/* consumer stopped due to lack of resources. */
82 };
83 
84 enum {
85 	C_FAST = 0,
86 	C_2,
87 	C_3,
88 	C_TAKEOVER,
89 };
90 
91 static inline uint16_t
92 space_available(struct mp_ring *r, union ring_state s)
93 {
94 	uint16_t x = r->size - 1;
95 
96 	if (s.cidx == s.pidx_head)
97 		return (x);
98 	else if (s.cidx > s.pidx_head)
99 		return (s.cidx - s.pidx_head - 1);
100 	else
101 		return (x - s.pidx_head + s.cidx);
102 }
103 
104 static inline uint16_t
105 increment_idx(struct mp_ring *r, uint16_t idx, uint16_t n)
106 {
107 	int x = r->size - idx;
108 
109 	MPASS(x > 0);
110 	return (x > n ? idx + n : n - x);
111 }
112 
113 /*
114  * Consumer.  Called with the consumer lock held and a guarantee that there is
115  * work to do.
116  */
117 static void
118 drain_ring(struct mp_ring *r, int budget)
119 {
120 	union ring_state os, ns;
121 	int n, pending, total;
122 	uint16_t cidx;
123 	uint16_t pidx;
124 	bool coalescing;
125 
126 	mtx_assert(r->cons_lock, MA_OWNED);
127 
128 	os.state = atomic_load_acq_64(&r->state);
129 	MPASS(os.flags == BUSY);
130 
131 	cidx = os.cidx;
132 	pidx = os.pidx_tail;
133 	MPASS(cidx != pidx);
134 
135 	pending = 0;
136 	total = 0;
137 
138 	while (cidx != pidx) {
139 
140 		/* Items from cidx to pidx are available for consumption. */
141 		n = r->drain(r, cidx, pidx, &coalescing);
142 		if (n == 0) {
143 			critical_enter();
144 			os.state = atomic_load_64(&r->state);
145 			do {
146 				ns.state = os.state;
147 				ns.cidx = cidx;
148 
149 				MPASS(os.flags == BUSY ||
150 				    os.flags == TOO_BUSY ||
151 				    os.flags == TAKING_OVER);
152 
153 				if (os.flags == TAKING_OVER)
154 					ns.flags = BUSY;
155 				else
156 					ns.flags = STALLED;
157 			} while (atomic_fcmpset_64(&r->state, &os.state,
158 			    ns.state) == 0);
159 			critical_exit();
160 			if (os.flags == TAKING_OVER)
161 				counter_u64_add(r->abdications, 1);
162 			else if (ns.flags == STALLED)
163 				counter_u64_add(r->stalls, 1);
164 			break;
165 		}
166 		cidx = increment_idx(r, cidx, n);
167 		pending += n;
168 		total += n;
169 		counter_u64_add(r->consumed, n);
170 
171 		os.state = atomic_load_64(&r->state);
172 		do {
173 			MPASS(os.flags == BUSY || os.flags == TOO_BUSY ||
174 			    os.flags == TAKING_OVER);
175 
176 			ns.state = os.state;
177 			ns.cidx = cidx;
178 			if (__predict_false(os.flags == TAKING_OVER)) {
179 				MPASS(total >= budget);
180 				ns.flags = BUSY;
181 				continue;
182 			}
183 			if (cidx == os.pidx_tail) {
184 				ns.flags = coalescing ? COALESCING : IDLE;
185 				continue;
186 			}
187 			if (total >= budget) {
188 				ns.flags = TOO_BUSY;
189 				continue;
190 			}
191 			MPASS(os.flags == BUSY);
192 			if (pending < 32)
193 				break;
194 		} while (atomic_fcmpset_acq_64(&r->state, &os.state, ns.state) == 0);
195 
196 		if (__predict_false(os.flags == TAKING_OVER)) {
197 			MPASS(ns.flags == BUSY);
198 			counter_u64_add(r->abdications, 1);
199 			break;
200 		}
201 
202 		if (ns.flags == IDLE || ns.flags == COALESCING) {
203 			MPASS(ns.pidx_tail == cidx);
204 			if (ns.pidx_head != ns.pidx_tail)
205 				counter_u64_add(r->cons_idle2, 1);
206 			else
207 				counter_u64_add(r->cons_idle, 1);
208 			break;
209 		}
210 
211 		/*
212 		 * The acquire style atomic above guarantees visibility of items
213 		 * associated with any pidx change that we notice here.
214 		 */
215 		pidx = ns.pidx_tail;
216 		pending = 0;
217 	}
218 
219 #ifdef INVARIANTS
220 	if (os.flags == TAKING_OVER)
221 		MPASS(ns.flags == BUSY);
222 	else {
223 		MPASS(ns.flags == IDLE || ns.flags == COALESCING ||
224 		    ns.flags == STALLED);
225 	}
226 #endif
227 }
228 
229 static void
230 drain_txpkts(struct mp_ring *r, union ring_state os, int budget)
231 {
232 	union ring_state ns;
233 	uint16_t cidx = os.cidx;
234 	uint16_t pidx = os.pidx_tail;
235 	bool coalescing;
236 
237 	mtx_assert(r->cons_lock, MA_OWNED);
238 	MPASS(os.flags == BUSY);
239 	MPASS(cidx == pidx);
240 
241 	r->drain(r, cidx, pidx, &coalescing);
242 	MPASS(coalescing == false);
243 	critical_enter();
244 	os.state = atomic_load_64(&r->state);
245 	do {
246 		ns.state = os.state;
247 		MPASS(os.flags == BUSY);
248 		MPASS(os.cidx == cidx);
249 		if (ns.cidx == ns.pidx_tail)
250 			ns.flags = IDLE;
251 		else
252 			ns.flags = BUSY;
253 	} while (atomic_fcmpset_acq_64(&r->state, &os.state, ns.state) == 0);
254 	critical_exit();
255 
256 	if (ns.flags == BUSY)
257 		drain_ring(r, budget);
258 }
259 
260 int
261 mp_ring_alloc(struct mp_ring **pr, int size, void *cookie, ring_drain_t drain,
262     ring_can_drain_t can_drain, struct malloc_type *mt, struct mtx *lck,
263     int flags)
264 {
265 	struct mp_ring *r;
266 	int i;
267 
268 	/* All idx are 16b so size can be 65536 at most */
269 	if (pr == NULL || size < 2 || size > 65536 || drain == NULL ||
270 	    can_drain == NULL)
271 		return (EINVAL);
272 	*pr = NULL;
273 	flags &= M_NOWAIT | M_WAITOK;
274 	MPASS(flags != 0);
275 
276 	r = malloc(__offsetof(struct mp_ring, items[size]), mt, flags | M_ZERO);
277 	if (r == NULL)
278 		return (ENOMEM);
279 	r->size = size;
280 	r->cookie = cookie;
281 	r->mt = mt;
282 	r->drain = drain;
283 	r->can_drain = can_drain;
284 	r->cons_lock = lck;
285 	if ((r->dropped = counter_u64_alloc(flags)) == NULL)
286 		goto failed;
287 	for (i = 0; i < nitems(r->consumer); i++) {
288 		if ((r->consumer[i] = counter_u64_alloc(flags)) == NULL)
289 			goto failed;
290 	}
291 	if ((r->not_consumer = counter_u64_alloc(flags)) == NULL)
292 		goto failed;
293 	if ((r->abdications = counter_u64_alloc(flags)) == NULL)
294 		goto failed;
295 	if ((r->stalls = counter_u64_alloc(flags)) == NULL)
296 		goto failed;
297 	if ((r->consumed = counter_u64_alloc(flags)) == NULL)
298 		goto failed;
299 	if ((r->cons_idle = counter_u64_alloc(flags)) == NULL)
300 		goto failed;
301 	if ((r->cons_idle2 = counter_u64_alloc(flags)) == NULL)
302 		goto failed;
303 	*pr = r;
304 	return (0);
305 failed:
306 	mp_ring_free(r);
307 	return (ENOMEM);
308 }
309 
310 void
311 
312 mp_ring_free(struct mp_ring *r)
313 {
314 	int i;
315 
316 	if (r == NULL)
317 		return;
318 
319 	if (r->dropped != NULL)
320 		counter_u64_free(r->dropped);
321 	for (i = 0; i < nitems(r->consumer); i++) {
322 		if (r->consumer[i] != NULL)
323 			counter_u64_free(r->consumer[i]);
324 	}
325 	if (r->not_consumer != NULL)
326 		counter_u64_free(r->not_consumer);
327 	if (r->abdications != NULL)
328 		counter_u64_free(r->abdications);
329 	if (r->stalls != NULL)
330 		counter_u64_free(r->stalls);
331 	if (r->consumed != NULL)
332 		counter_u64_free(r->consumed);
333 	if (r->cons_idle != NULL)
334 		counter_u64_free(r->cons_idle);
335 	if (r->cons_idle2 != NULL)
336 		counter_u64_free(r->cons_idle2);
337 
338 	free(r, r->mt);
339 }
340 
341 /*
342  * Enqueue n items and maybe drain the ring for some time.
343  *
344  * Returns an errno.
345  */
346 int
347 mp_ring_enqueue(struct mp_ring *r, void **items, int n, int budget)
348 {
349 	union ring_state os, ns;
350 	uint16_t pidx_start, pidx_stop;
351 	int i, nospc, cons;
352 	bool consumer;
353 
354 	MPASS(items != NULL);
355 	MPASS(n > 0);
356 
357 	/*
358 	 * Reserve room for the new items.  Our reservation, if successful, is
359 	 * from 'pidx_start' to 'pidx_stop'.
360 	 */
361 	nospc = 0;
362 	os.state = atomic_load_64(&r->state);
363 	for (;;) {
364 		for (;;) {
365 			if (__predict_true(space_available(r, os) >= n))
366 				break;
367 
368 			/* Not enough room in the ring. */
369 
370 			MPASS(os.flags != IDLE);
371 			MPASS(os.flags != COALESCING);
372 			if (__predict_false(++nospc > 100)) {
373 				counter_u64_add(r->dropped, n);
374 				return (ENOBUFS);
375 			}
376 			if (os.flags == STALLED)
377 				mp_ring_check_drainage(r, 64);
378 			else
379 				cpu_spinwait();
380 			os.state = atomic_load_64(&r->state);
381 		}
382 
383 		/* There is room in the ring. */
384 
385 		cons = -1;
386 		ns.state = os.state;
387 		ns.pidx_head = increment_idx(r, os.pidx_head, n);
388 		if (os.flags == IDLE || os.flags == COALESCING) {
389 			MPASS(os.pidx_tail == os.cidx);
390 			if (os.pidx_head == os.pidx_tail) {
391 				cons = C_FAST;
392 				ns.pidx_tail = increment_idx(r, os.pidx_tail, n);
393 			} else
394 				cons = C_2;
395 			ns.flags = BUSY;
396 		} else if (os.flags == TOO_BUSY) {
397 			cons = C_TAKEOVER;
398 			ns.flags = TAKING_OVER;
399 		}
400 		critical_enter();
401 		if (atomic_fcmpset_64(&r->state, &os.state, ns.state))
402 			break;
403 		critical_exit();
404 		cpu_spinwait();
405 	};
406 
407 	pidx_start = os.pidx_head;
408 	pidx_stop = ns.pidx_head;
409 
410 	if (cons == C_FAST) {
411 		i = pidx_start;
412 		do {
413 			r->items[i] = *items++;
414 			if (__predict_false(++i == r->size))
415 				i = 0;
416 		} while (i != pidx_stop);
417 		critical_exit();
418 		counter_u64_add(r->consumer[C_FAST], 1);
419 		mtx_lock(r->cons_lock);
420 		drain_ring(r, budget);
421 		mtx_unlock(r->cons_lock);
422 		return (0);
423 	}
424 
425 	/*
426 	 * Wait for other producers who got in ahead of us to enqueue their
427 	 * items, one producer at a time.  It is our turn when the ring's
428 	 * pidx_tail reaches the beginning of our reservation (pidx_start).
429 	 */
430 	while (ns.pidx_tail != pidx_start) {
431 		cpu_spinwait();
432 		ns.state = atomic_load_64(&r->state);
433 	}
434 
435 	/* Now it is our turn to fill up the area we reserved earlier. */
436 	i = pidx_start;
437 	do {
438 		r->items[i] = *items++;
439 		if (__predict_false(++i == r->size))
440 			i = 0;
441 	} while (i != pidx_stop);
442 
443 	/*
444 	 * Update the ring's pidx_tail.  The release style atomic guarantees
445 	 * that the items are visible to any thread that sees the updated pidx.
446 	 */
447 	os.state = atomic_load_64(&r->state);
448 	do {
449 		consumer = false;
450 		ns.state = os.state;
451 		ns.pidx_tail = pidx_stop;
452 		if (os.flags == IDLE || os.flags == COALESCING ||
453 		    (os.flags == STALLED && r->can_drain(r))) {
454 			MPASS(cons == -1);
455 			consumer = true;
456 			ns.flags = BUSY;
457 		}
458 	} while (atomic_fcmpset_rel_64(&r->state, &os.state, ns.state) == 0);
459 	critical_exit();
460 
461 	if (cons == -1) {
462 		if (consumer)
463 			cons = C_3;
464 		else {
465 			counter_u64_add(r->not_consumer, 1);
466 			return (0);
467 		}
468 	}
469 	MPASS(cons > C_FAST && cons < nitems(r->consumer));
470 	counter_u64_add(r->consumer[cons], 1);
471 	mtx_lock(r->cons_lock);
472 	drain_ring(r, budget);
473 	mtx_unlock(r->cons_lock);
474 
475 	return (0);
476 }
477 
478 void
479 mp_ring_check_drainage(struct mp_ring *r, int budget)
480 {
481 	union ring_state os, ns;
482 
483 	os.state = atomic_load_64(&r->state);
484 	if (os.flags == STALLED && r->can_drain(r)) {
485 		MPASS(os.cidx != os.pidx_tail);	/* implied by STALLED */
486 		ns.state = os.state;
487 		ns.flags = BUSY;
488 		if (atomic_cmpset_acq_64(&r->state, os.state, ns.state)) {
489 			mtx_lock(r->cons_lock);
490 			drain_ring(r, budget);
491 			mtx_unlock(r->cons_lock);
492 		}
493 	} else if (os.flags == COALESCING) {
494 		MPASS(os.cidx == os.pidx_tail);
495 		ns.state = os.state;
496 		ns.flags = BUSY;
497 		if (atomic_cmpset_acq_64(&r->state, os.state, ns.state)) {
498 			mtx_lock(r->cons_lock);
499 			drain_txpkts(r, ns, budget);
500 			mtx_unlock(r->cons_lock);
501 		}
502 	}
503 }
504 
505 void
506 mp_ring_reset_stats(struct mp_ring *r)
507 {
508 	int i;
509 
510 	counter_u64_zero(r->dropped);
511 	for (i = 0; i < nitems(r->consumer); i++)
512 		counter_u64_zero(r->consumer[i]);
513 	counter_u64_zero(r->not_consumer);
514 	counter_u64_zero(r->abdications);
515 	counter_u64_zero(r->stalls);
516 	counter_u64_zero(r->consumed);
517 	counter_u64_zero(r->cons_idle);
518 	counter_u64_zero(r->cons_idle2);
519 }
520 
521 bool
522 mp_ring_is_idle(struct mp_ring *r)
523 {
524 	union ring_state s;
525 
526 	s.state = atomic_load_64(&r->state);
527 	if (s.pidx_head == s.pidx_tail && s.pidx_tail == s.cidx &&
528 	    s.flags == IDLE)
529 		return (true);
530 
531 	return (false);
532 }
533 
534 void
535 mp_ring_sysctls(struct mp_ring *r, struct sysctl_ctx_list *ctx,
536     struct sysctl_oid_list *children)
537 {
538 	struct sysctl_oid *oid;
539 
540 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "mp_ring", CTLFLAG_RD |
541 	    CTLFLAG_MPSAFE, NULL, "mp_ring statistics");
542 	children = SYSCTL_CHILDREN(oid);
543 
544 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "state", CTLFLAG_RD,
545 	    __DEVOLATILE(uint64_t *, &r->state), 0, "ring state");
546 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "dropped", CTLFLAG_RD,
547 	    &r->dropped, "# of items dropped");
548 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "consumed",
549 	    CTLFLAG_RD, &r->consumed, "# of items consumed");
550 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "fast_consumer",
551 	    CTLFLAG_RD, &r->consumer[C_FAST],
552 	    "# of times producer became consumer (fast)");
553 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "consumer2",
554 	    CTLFLAG_RD, &r->consumer[C_2],
555 	    "# of times producer became consumer (2)");
556 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "consumer3",
557 	    CTLFLAG_RD, &r->consumer[C_3],
558 	    "# of times producer became consumer (3)");
559 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "takeovers",
560 	    CTLFLAG_RD, &r->consumer[C_TAKEOVER],
561 	    "# of times producer took over from another consumer.");
562 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "not_consumer",
563 	    CTLFLAG_RD, &r->not_consumer,
564 	    "# of times producer did not become consumer");
565 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "abdications",
566 	    CTLFLAG_RD, &r->abdications, "# of consumer abdications");
567 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "stalls",
568 	    CTLFLAG_RD, &r->stalls, "# of consumer stalls");
569 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "cons_idle",
570 	    CTLFLAG_RD, &r->cons_idle,
571 	    "# of times consumer ran fully to completion");
572 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "cons_idle2",
573 	    CTLFLAG_RD, &r->cons_idle2,
574 	    "# of times consumer idled when another enqueue was in progress");
575 }
576