2 * Copyright (c) 2014 Chelsio Communications, Inc.
4 * Written by: Navdeep Parhar <np@FreeBSD.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
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.
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
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
31 #include <sys/types.h>
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/counter.h>
36 #include <sys/mutex.h>
37 #include <sys/malloc.h>
38 #include <machine/cpu.h>
39 #include <net/mp_ring.h>
52 IDLE = 0, /* consumer ran to completion, nothing more to do. */
53 BUSY, /* consumer is running already, or will be shortly. */
54 STALLED, /* consumer stopped due to lack of resources. */
55 ABDICATED, /* consumer stopped even though there was work to be
56 done because it wants another thread to take over. */
59 static inline uint16_t
60 space_available(struct ifmp_ring *r, union ring_state s)
62 uint16_t x = r->size - 1;
64 if (s.cidx == s.pidx_head)
66 else if (s.cidx > s.pidx_head)
67 return (s.cidx - s.pidx_head - 1);
69 return (x - s.pidx_head + s.cidx);
72 static inline uint16_t
73 increment_idx(struct ifmp_ring *r, uint16_t idx, uint16_t n)
75 int x = r->size - idx;
78 return (x > n ? idx + n : n - x);
81 /* Consumer is about to update the ring's state to s */
82 static inline uint16_t
83 state_to_flags(union ring_state s, int abdicate)
86 if (s.cidx == s.pidx_tail)
88 else if (abdicate && s.pidx_tail != s.pidx_head)
94 #ifdef MP_RING_NO_64BIT_ATOMICS
96 drain_ring_locked(struct ifmp_ring *r, union ring_state os, uint16_t prev, int budget)
99 int n, pending, total;
100 uint16_t cidx = os.cidx;
101 uint16_t pidx = os.pidx_tail;
103 MPASS(os.flags == BUSY);
107 counter_u64_add(r->starts, 1);
111 while (cidx != pidx) {
112 /* Items from cidx to pidx are available for consumption. */
113 n = r->drain(r, cidx, pidx);
115 os.state = ns.state = r->state;
120 counter_u64_add(r->stalls, 1);
121 else if (total > 0) {
122 counter_u64_add(r->restarts, 1);
123 counter_u64_add(r->stalls, 1);
127 cidx = increment_idx(r, cidx, n);
132 * We update the cidx only if we've caught up with the pidx, the
133 * real cidx is getting too far ahead of the one visible to
134 * everyone else, or we have exceeded our budget.
136 if (cidx != pidx && pending < 64 && total < budget)
139 os.state = ns.state = r->state;
141 ns.flags = state_to_flags(ns, total >= budget);
144 if (ns.flags == ABDICATED)
145 counter_u64_add(r->abdications, 1);
146 if (ns.flags != BUSY) {
147 /* Wrong loop exit if we're going to stall. */
148 MPASS(ns.flags != STALLED);
149 if (prev == STALLED) {
151 counter_u64_add(r->restarts, 1);
157 * The acquire style atomic above guarantees visibility of items
158 * associated with any pidx change that we notice here.
166 * Caller passes in a state, with a guarantee that there is work to do and that
167 * all items up to the pidx_tail in the state are visible.
170 drain_ring_lockless(struct ifmp_ring *r, union ring_state os, uint16_t prev, int budget)
173 int n, pending, total;
174 uint16_t cidx = os.cidx;
175 uint16_t pidx = os.pidx_tail;
177 MPASS(os.flags == BUSY);
181 counter_u64_add(r->starts, 1);
185 while (cidx != pidx) {
186 /* Items from cidx to pidx are available for consumption. */
187 n = r->drain(r, cidx, pidx);
195 } while (atomic_fcmpset_64(&r->state, &os.state,
199 counter_u64_add(r->stalls, 1);
200 else if (total > 0) {
201 counter_u64_add(r->restarts, 1);
202 counter_u64_add(r->stalls, 1);
206 cidx = increment_idx(r, cidx, n);
211 * We update the cidx only if we've caught up with the pidx, the
212 * real cidx is getting too far ahead of the one visible to
213 * everyone else, or we have exceeded our budget.
215 if (cidx != pidx && pending < 64 && total < budget)
222 ns.flags = state_to_flags(ns, total >= budget);
223 } while (atomic_fcmpset_acq_64(&r->state, &os.state,
227 if (ns.flags == ABDICATED)
228 counter_u64_add(r->abdications, 1);
229 if (ns.flags != BUSY) {
230 /* Wrong loop exit if we're going to stall. */
231 MPASS(ns.flags != STALLED);
232 if (prev == STALLED) {
234 counter_u64_add(r->restarts, 1);
240 * The acquire style atomic above guarantees visibility of items
241 * associated with any pidx change that we notice here.
250 ifmp_ring_alloc(struct ifmp_ring **pr, int size, void *cookie, mp_ring_drain_t drain,
251 mp_ring_can_drain_t can_drain, struct malloc_type *mt, int flags)
255 /* All idx are 16b so size can be 65536 at most */
256 if (pr == NULL || size < 2 || size > 65536 || drain == NULL ||
260 flags &= M_NOWAIT | M_WAITOK;
263 r = malloc(__offsetof(struct ifmp_ring, items[size]), mt, flags | M_ZERO);
270 r->can_drain = can_drain;
271 r->enqueues = counter_u64_alloc(flags);
272 r->drops = counter_u64_alloc(flags);
273 r->starts = counter_u64_alloc(flags);
274 r->stalls = counter_u64_alloc(flags);
275 r->restarts = counter_u64_alloc(flags);
276 r->abdications = counter_u64_alloc(flags);
277 if (r->enqueues == NULL || r->drops == NULL || r->starts == NULL ||
278 r->stalls == NULL || r->restarts == NULL ||
279 r->abdications == NULL) {
285 #ifdef MP_RING_NO_64BIT_ATOMICS
286 mtx_init(&r->lock, "mp_ring lock", NULL, MTX_DEF);
292 ifmp_ring_free(struct ifmp_ring *r)
298 if (r->enqueues != NULL)
299 counter_u64_free(r->enqueues);
300 if (r->drops != NULL)
301 counter_u64_free(r->drops);
302 if (r->starts != NULL)
303 counter_u64_free(r->starts);
304 if (r->stalls != NULL)
305 counter_u64_free(r->stalls);
306 if (r->restarts != NULL)
307 counter_u64_free(r->restarts);
308 if (r->abdications != NULL)
309 counter_u64_free(r->abdications);
315 * Enqueue n items and maybe drain the ring for some time.
319 #ifdef MP_RING_NO_64BIT_ATOMICS
321 ifmp_ring_enqueue(struct ifmp_ring *r, void **items, int n, int budget, int abdicate)
323 union ring_state os, ns;
324 uint16_t pidx_start, pidx_stop;
327 MPASS(items != NULL);
332 * Reserve room for the new items. Our reservation, if successful, is
333 * from 'pidx_start' to 'pidx_stop'.
336 if (n >= space_available(r, os)) {
337 counter_u64_add(r->drops, n);
338 MPASS(os.flags != IDLE);
339 mtx_unlock(&r->lock);
340 if (os.flags == STALLED)
341 ifmp_ring_check_drainage(r, 0);
345 ns.pidx_head = increment_idx(r, os.pidx_head, n);
347 pidx_start = os.pidx_head;
348 pidx_stop = ns.pidx_head;
351 * Wait for other producers who got in ahead of us to enqueue their
352 * items, one producer at a time. It is our turn when the ring's
353 * pidx_tail reaches the beginning of our reservation (pidx_start).
355 while (ns.pidx_tail != pidx_start) {
360 /* Now it is our turn to fill up the area we reserved earlier. */
363 r->items[i] = *items++;
364 if (__predict_false(++i == r->size))
366 } while (i != pidx_stop);
369 * Update the ring's pidx_tail. The release style atomic guarantees
370 * that the items are visible to any thread that sees the updated pidx.
372 os.state = ns.state = r->state;
373 ns.pidx_tail = pidx_stop;
375 if (os.flags == IDLE)
376 ns.flags = ABDICATED;
380 counter_u64_add(r->enqueues, n);
384 * Turn into a consumer if some other thread isn't active as a consumer
387 if (os.flags != BUSY)
388 drain_ring_locked(r, ns, os.flags, budget);
391 mtx_unlock(&r->lock);
396 ifmp_ring_enqueue(struct ifmp_ring *r, void **items, int n, int budget, int abdicate)
398 union ring_state os, ns;
399 uint16_t pidx_start, pidx_stop;
402 MPASS(items != NULL);
406 * Reserve room for the new items. Our reservation, if successful, is
407 * from 'pidx_start' to 'pidx_stop'.
411 if (n >= space_available(r, os)) {
412 counter_u64_add(r->drops, n);
413 MPASS(os.flags != IDLE);
414 if (os.flags == STALLED)
415 ifmp_ring_check_drainage(r, 0);
419 ns.pidx_head = increment_idx(r, os.pidx_head, n);
421 if (atomic_fcmpset_64(&r->state, &os.state, ns.state))
426 pidx_start = os.pidx_head;
427 pidx_stop = ns.pidx_head;
430 * Wait for other producers who got in ahead of us to enqueue their
431 * items, one producer at a time. It is our turn when the ring's
432 * pidx_tail reaches the beginning of our reservation (pidx_start).
434 while (ns.pidx_tail != pidx_start) {
439 /* Now it is our turn to fill up the area we reserved earlier. */
442 r->items[i] = *items++;
443 if (__predict_false(++i == r->size))
445 } while (i != pidx_stop);
448 * Update the ring's pidx_tail. The release style atomic guarantees
449 * that the items are visible to any thread that sees the updated pidx.
454 ns.pidx_tail = pidx_stop;
456 if (os.flags == IDLE)
457 ns.flags = ABDICATED;
460 } while (atomic_fcmpset_rel_64(&r->state, &os.state, ns.state) == 0);
462 counter_u64_add(r->enqueues, n);
466 * Turn into a consumer if some other thread isn't active as a consumer
469 if (os.flags != BUSY)
470 drain_ring_lockless(r, ns, os.flags, budget);
478 ifmp_ring_check_drainage(struct ifmp_ring *r, int budget)
480 union ring_state os, ns;
483 if ((os.flags != STALLED && os.flags != ABDICATED) || // Only continue in STALLED and ABDICATED
484 os.pidx_head != os.pidx_tail || // Require work to be available
485 (os.flags != ABDICATED && r->can_drain(r) == 0)) // Can either drain, or everyone left
488 MPASS(os.cidx != os.pidx_tail); /* implied by STALLED */
492 #ifdef MP_RING_NO_64BIT_ATOMICS
494 if (r->state != os.state) {
495 mtx_unlock(&r->lock);
499 drain_ring_locked(r, ns, os.flags, budget);
500 mtx_unlock(&r->lock);
503 * The acquire style atomic guarantees visibility of items associated
504 * with the pidx that we read here.
506 if (!atomic_cmpset_acq_64(&r->state, os.state, ns.state))
509 drain_ring_lockless(r, ns, os.flags, budget);
514 ifmp_ring_reset_stats(struct ifmp_ring *r)
517 counter_u64_zero(r->enqueues);
518 counter_u64_zero(r->drops);
519 counter_u64_zero(r->starts);
520 counter_u64_zero(r->stalls);
521 counter_u64_zero(r->restarts);
522 counter_u64_zero(r->abdications);
526 ifmp_ring_is_idle(struct ifmp_ring *r)
531 if (s.pidx_head == s.pidx_tail && s.pidx_tail == s.cidx &&
539 ifmp_ring_is_stalled(struct ifmp_ring *r)
544 if (s.pidx_head == s.pidx_tail && s.flags == STALLED)