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 <machine/cpu.h>
38
39 #include "t4_mp_ring.h"
40
41 #if defined(__i386__)
42 #define atomic_cmpset_acq_64 atomic_cmpset_64
43 #define atomic_cmpset_rel_64 atomic_cmpset_64
44 #endif
45
46 union ring_state {
47 struct {
48 uint16_t pidx_head;
49 uint16_t pidx_tail;
50 uint16_t cidx;
51 uint16_t flags;
52 };
53 uint64_t state;
54 };
55
56 enum {
57 IDLE = 0, /* consumer ran to completion, nothing more to do. */
58 BUSY, /* consumer is running already, or will be shortly. */
59 STALLED, /* consumer stopped due to lack of resources. */
60 ABDICATED, /* consumer stopped even though there was work to be
61 done because it wants another thread to take over. */
62 };
63
64 static inline uint16_t
space_available(struct mp_ring * r,union ring_state s)65 space_available(struct mp_ring *r, union ring_state s)
66 {
67 uint16_t x = r->size - 1;
68
69 if (s.cidx == s.pidx_head)
70 return (x);
71 else if (s.cidx > s.pidx_head)
72 return (s.cidx - s.pidx_head - 1);
73 else
74 return (x - s.pidx_head + s.cidx);
75 }
76
77 static inline uint16_t
increment_idx(struct mp_ring * r,uint16_t idx,uint16_t n)78 increment_idx(struct mp_ring *r, uint16_t idx, uint16_t n)
79 {
80 int x = r->size - idx;
81
82 MPASS(x > 0);
83 return (x > n ? idx + n : n - x);
84 }
85
86 /* Consumer is about to update the ring's state to s */
87 static inline uint16_t
state_to_flags(union ring_state s,int abdicate)88 state_to_flags(union ring_state s, int abdicate)
89 {
90
91 if (s.cidx == s.pidx_tail)
92 return (IDLE);
93 else if (abdicate && s.pidx_tail != s.pidx_head)
94 return (ABDICATED);
95
96 return (BUSY);
97 }
98
99 /*
100 * Caller passes in a state, with a guarantee that there is work to do and that
101 * all items up to the pidx_tail in the state are visible.
102 */
103 static void
drain_ring(struct mp_ring * r,union ring_state os,uint16_t prev,int budget)104 drain_ring(struct mp_ring *r, union ring_state os, uint16_t prev, int budget)
105 {
106 union ring_state ns;
107 int n, pending, total;
108 uint16_t cidx = os.cidx;
109 uint16_t pidx = os.pidx_tail;
110
111 MPASS(os.flags == BUSY);
112 MPASS(cidx != pidx);
113
114 if (prev == IDLE)
115 counter_u64_add(r->starts, 1);
116 pending = 0;
117 total = 0;
118
119 while (cidx != pidx) {
120
121 /* Items from cidx to pidx are available for consumption. */
122 n = r->drain(r, cidx, pidx);
123 if (n == 0) {
124 critical_enter();
125 do {
126 os.state = ns.state = r->state;
127 ns.cidx = cidx;
128 ns.flags = STALLED;
129 } while (atomic_cmpset_64(&r->state, os.state,
130 ns.state) == 0);
131 critical_exit();
132 if (prev != STALLED)
133 counter_u64_add(r->stalls, 1);
134 else if (total > 0) {
135 counter_u64_add(r->restarts, 1);
136 counter_u64_add(r->stalls, 1);
137 }
138 break;
139 }
140 cidx = increment_idx(r, cidx, n);
141 pending += n;
142 total += n;
143
144 /*
145 * We update the cidx only if we've caught up with the pidx, the
146 * real cidx is getting too far ahead of the one visible to
147 * everyone else, or we have exceeded our budget.
148 */
149 if (cidx != pidx && pending < 64 && total < budget)
150 continue;
151 critical_enter();
152 do {
153 os.state = ns.state = r->state;
154 ns.cidx = cidx;
155 ns.flags = state_to_flags(ns, total >= budget);
156 } while (atomic_cmpset_acq_64(&r->state, os.state, ns.state) == 0);
157 critical_exit();
158
159 if (ns.flags == ABDICATED)
160 counter_u64_add(r->abdications, 1);
161 if (ns.flags != BUSY) {
162 /* Wrong loop exit if we're going to stall. */
163 MPASS(ns.flags != STALLED);
164 if (prev == STALLED) {
165 MPASS(total > 0);
166 counter_u64_add(r->restarts, 1);
167 }
168 break;
169 }
170
171 /*
172 * The acquire style atomic above guarantees visibility of items
173 * associated with any pidx change that we notice here.
174 */
175 pidx = ns.pidx_tail;
176 pending = 0;
177 }
178 }
179
180 int
mp_ring_alloc(struct mp_ring ** pr,int size,void * cookie,ring_drain_t drain,ring_can_drain_t can_drain,struct malloc_type * mt,int flags)181 mp_ring_alloc(struct mp_ring **pr, int size, void *cookie, ring_drain_t drain,
182 ring_can_drain_t can_drain, struct malloc_type *mt, int flags)
183 {
184 struct mp_ring *r;
185
186 /* All idx are 16b so size can be 65536 at most */
187 if (pr == NULL || size < 2 || size > 65536 || drain == NULL ||
188 can_drain == NULL)
189 return (EINVAL);
190 *pr = NULL;
191 flags &= M_NOWAIT | M_WAITOK;
192 MPASS(flags != 0);
193
194 r = malloc(__offsetof(struct mp_ring, items[size]), mt, flags | M_ZERO);
195 if (r == NULL)
196 return (ENOMEM);
197 r->size = size;
198 r->cookie = cookie;
199 r->mt = mt;
200 r->drain = drain;
201 r->can_drain = can_drain;
202 r->enqueues = counter_u64_alloc(flags);
203 r->drops = counter_u64_alloc(flags);
204 r->starts = counter_u64_alloc(flags);
205 r->stalls = counter_u64_alloc(flags);
206 r->restarts = counter_u64_alloc(flags);
207 r->abdications = counter_u64_alloc(flags);
208 if (r->enqueues == NULL || r->drops == NULL || r->starts == NULL ||
209 r->stalls == NULL || r->restarts == NULL ||
210 r->abdications == NULL) {
211 mp_ring_free(r);
212 return (ENOMEM);
213 }
214
215 *pr = r;
216 return (0);
217 }
218
219 void
220
mp_ring_free(struct mp_ring * r)221 mp_ring_free(struct mp_ring *r)
222 {
223
224 if (r == NULL)
225 return;
226
227 if (r->enqueues != NULL)
228 counter_u64_free(r->enqueues);
229 if (r->drops != NULL)
230 counter_u64_free(r->drops);
231 if (r->starts != NULL)
232 counter_u64_free(r->starts);
233 if (r->stalls != NULL)
234 counter_u64_free(r->stalls);
235 if (r->restarts != NULL)
236 counter_u64_free(r->restarts);
237 if (r->abdications != NULL)
238 counter_u64_free(r->abdications);
239
240 free(r, r->mt);
241 }
242
243 /*
244 * Enqueue n items and maybe drain the ring for some time.
245 *
246 * Returns an errno.
247 */
248 int
mp_ring_enqueue(struct mp_ring * r,void ** items,int n,int budget)249 mp_ring_enqueue(struct mp_ring *r, void **items, int n, int budget)
250 {
251 union ring_state os, ns;
252 uint16_t pidx_start, pidx_stop;
253 int i;
254
255 MPASS(items != NULL);
256 MPASS(n > 0);
257
258 /*
259 * Reserve room for the new items. Our reservation, if successful, is
260 * from 'pidx_start' to 'pidx_stop'.
261 */
262 for (;;) {
263 os.state = r->state;
264 if (n >= space_available(r, os)) {
265 counter_u64_add(r->drops, n);
266 MPASS(os.flags != IDLE);
267 if (os.flags == STALLED)
268 mp_ring_check_drainage(r, 0);
269 return (ENOBUFS);
270 }
271 ns.state = os.state;
272 ns.pidx_head = increment_idx(r, os.pidx_head, n);
273 critical_enter();
274 if (atomic_cmpset_64(&r->state, os.state, ns.state))
275 break;
276 critical_exit();
277 cpu_spinwait();
278 }
279 pidx_start = os.pidx_head;
280 pidx_stop = ns.pidx_head;
281
282 /*
283 * Wait for other producers who got in ahead of us to enqueue their
284 * items, one producer at a time. It is our turn when the ring's
285 * pidx_tail reaches the begining of our reservation (pidx_start).
286 */
287 while (ns.pidx_tail != pidx_start) {
288 cpu_spinwait();
289 ns.state = r->state;
290 }
291
292 /* Now it is our turn to fill up the area we reserved earlier. */
293 i = pidx_start;
294 do {
295 r->items[i] = *items++;
296 if (__predict_false(++i == r->size))
297 i = 0;
298 } while (i != pidx_stop);
299
300 /*
301 * Update the ring's pidx_tail. The release style atomic guarantees
302 * that the items are visible to any thread that sees the updated pidx.
303 */
304 do {
305 os.state = ns.state = r->state;
306 ns.pidx_tail = pidx_stop;
307 ns.flags = BUSY;
308 } while (atomic_cmpset_rel_64(&r->state, os.state, ns.state) == 0);
309 critical_exit();
310 counter_u64_add(r->enqueues, n);
311
312 /*
313 * Turn into a consumer if some other thread isn't active as a consumer
314 * already.
315 */
316 if (os.flags != BUSY)
317 drain_ring(r, ns, os.flags, budget);
318
319 return (0);
320 }
321
322 void
mp_ring_check_drainage(struct mp_ring * r,int budget)323 mp_ring_check_drainage(struct mp_ring *r, int budget)
324 {
325 union ring_state os, ns;
326
327 os.state = r->state;
328 if (os.flags != STALLED || os.pidx_head != os.pidx_tail ||
329 r->can_drain(r) == 0)
330 return;
331
332 MPASS(os.cidx != os.pidx_tail); /* implied by STALLED */
333 ns.state = os.state;
334 ns.flags = BUSY;
335
336 /*
337 * The acquire style atomic guarantees visibility of items associated
338 * with the pidx that we read here.
339 */
340 if (!atomic_cmpset_acq_64(&r->state, os.state, ns.state))
341 return;
342
343 drain_ring(r, ns, os.flags, budget);
344 }
345
346 void
mp_ring_reset_stats(struct mp_ring * r)347 mp_ring_reset_stats(struct mp_ring *r)
348 {
349
350 counter_u64_zero(r->enqueues);
351 counter_u64_zero(r->drops);
352 counter_u64_zero(r->starts);
353 counter_u64_zero(r->stalls);
354 counter_u64_zero(r->restarts);
355 counter_u64_zero(r->abdications);
356 }
357
358 int
mp_ring_is_idle(struct mp_ring * r)359 mp_ring_is_idle(struct mp_ring *r)
360 {
361 union ring_state s;
362
363 s.state = r->state;
364 if (s.pidx_head == s.pidx_tail && s.pidx_tail == s.cidx &&
365 s.flags == IDLE)
366 return (1);
367
368 return (0);
369 }
370