sys/net/mp_ring.c

   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 }