drivers/mempool/bucket/rte_mempool_bucket.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  *
   3  * Copyright (c) 2017-2018 Solarflare Communications Inc.
   4  * All rights reserved.
   5  *
   6  * This software was jointly developed between OKTET Labs (under contract
   7  * for Solarflare) and Solarflare Communications, Inc.
   8  */
   9
  10 #include <stdbool.h>
  11 #include <stdio.h>
  12 #include <string.h>
  13
  14 #include <rte_errno.h>
  15 #include <rte_ring.h>
  16 #include <rte_mempool.h>
  17 #include <rte_malloc.h>
  18
  19 /*
  20  * The general idea of the bucket mempool driver is as follows.
  21  * We keep track of physically contiguous groups (buckets) of objects
  22  * of a certain size. Every such a group has a counter that is
  23  * incremented every time an object from that group is enqueued.
  24  * Until the bucket is full, no objects from it are eligible for allocation.
  25  * If a request is made to dequeue a multiply of bucket size, it is
  26  * satisfied by returning the whole buckets, instead of separate objects.
  27  */
  28
  29
  30 struct bucket_header {
  31         unsigned int lcore_id;
  32         uint8_t fill_cnt;
  33 };
  34
  35 struct bucket_stack {
  36         unsigned int top;
  37         unsigned int limit;
  38         void *objects[];
  39 };
  40
  41 struct bucket_data {
  42         unsigned int header_size;
  43         unsigned int total_elt_size;
  44         unsigned int obj_per_bucket;
  45         uintptr_t bucket_page_mask;
  46         struct rte_ring *shared_bucket_ring;
  47         struct bucket_stack *buckets[RTE_MAX_LCORE];
  48         /*
  49          * Multi-producer single-consumer ring to hold objects that are
  50          * returned to the mempool at a different lcore than initially
  51          * dequeued
  52          */
  53         struct rte_ring *adoption_buffer_rings[RTE_MAX_LCORE];
  54         struct rte_ring *shared_orphan_ring;
  55         struct rte_mempool *pool;
  56         unsigned int bucket_mem_size;
  57 };
  58
  59 static struct bucket_stack *
  60 bucket_stack_create(const struct rte_mempool *mp, unsigned int n_elts)
  61 {
  62         struct bucket_stack *stack;
  63
  64         stack = rte_zmalloc_socket("bucket_stack",
  65                                    sizeof(struct bucket_stack) +
  66                                    n_elts * sizeof(void *),
  67                                    RTE_CACHE_LINE_SIZE,
  68                                    mp->socket_id);
  69         if (stack == NULL)
  70                 return NULL;
  71         stack->limit = n_elts;
  72         stack->top = 0;
  73
  74         return stack;
  75 }
  76
  77 static void
  78 bucket_stack_push(struct bucket_stack *stack, void *obj)
  79 {
  80         RTE_ASSERT(stack->top < stack->limit);
  81         stack->objects[stack->top++] = obj;
  82 }
  83
  84 static void *
  85 bucket_stack_pop_unsafe(struct bucket_stack *stack)
  86 {
  87         RTE_ASSERT(stack->top > 0);
  88         return stack->objects[--stack->top];
  89 }
  90
  91 static void *
  92 bucket_stack_pop(struct bucket_stack *stack)
  93 {
  94         if (stack->top == 0)
  95                 return NULL;
  96         return bucket_stack_pop_unsafe(stack);
  97 }
  98
  99 static int
 100 bucket_enqueue_single(struct bucket_data *bd, void *obj)
 101 {
 102         int rc = 0;
 103         uintptr_t addr = (uintptr_t)obj;
 104         struct bucket_header *hdr;
 105         unsigned int lcore_id = rte_lcore_id();
 106
 107         addr &= bd->bucket_page_mask;
 108         hdr = (struct bucket_header *)addr;
 109
 110         if (likely(hdr->lcore_id == lcore_id)) {
 111                 if (hdr->fill_cnt < bd->obj_per_bucket - 1) {
 112                         hdr->fill_cnt++;
 113                 } else {
 114                         hdr->fill_cnt = 0;
 115                         /* Stack is big enough to put all buckets */
 116                         bucket_stack_push(bd->buckets[lcore_id], hdr);
 117                 }
 118         } else if (hdr->lcore_id != LCORE_ID_ANY) {
 119                 struct rte_ring *adopt_ring =
 120                         bd->adoption_buffer_rings[hdr->lcore_id];
 121
 122                 rc = rte_ring_enqueue(adopt_ring, obj);
 123                 /* Ring is big enough to put all objects */
 124                 RTE_ASSERT(rc == 0);
 125         } else if (hdr->fill_cnt < bd->obj_per_bucket - 1) {
 126                 hdr->fill_cnt++;
 127         } else {
 128                 hdr->fill_cnt = 0;
 129                 rc = rte_ring_enqueue(bd->shared_bucket_ring, hdr);
 130                 /* Ring is big enough to put all buckets */
 131                 RTE_ASSERT(rc == 0);
 132         }
 133
 134         return rc;
 135 }
 136
 137 static int
 138 bucket_enqueue(struct rte_mempool *mp, void * const *obj_table,
 139                unsigned int n)
 140 {
 141         struct bucket_data *bd = mp->pool_data;
 142         unsigned int i;
 143         int rc = 0;
 144
 145         for (i = 0; i < n; i++) {
 146                 rc = bucket_enqueue_single(bd, obj_table[i]);
 147                 RTE_ASSERT(rc == 0);
 148         }
 149         return rc;
 150 }
 151
 152 static void **
 153 bucket_fill_obj_table(const struct bucket_data *bd, void **pstart,
 154                       void **obj_table, unsigned int n)
 155 {
 156         unsigned int i;
 157         uint8_t *objptr = *pstart;
 158
 159         for (objptr += bd->header_size, i = 0; i < n;
 160              i++, objptr += bd->total_elt_size)
 161                 *obj_table++ = objptr;
 162         *pstart = objptr;
 163         return obj_table;
 164 }
 165
 166 static int
 167 bucket_dequeue_orphans(struct bucket_data *bd, void **obj_table,
 168                        unsigned int n_orphans)
 169 {
 170         unsigned int i;
 171         int rc;
 172         uint8_t *objptr;
 173
 174         rc = rte_ring_dequeue_bulk(bd->shared_orphan_ring, obj_table,
 175                                    n_orphans, NULL);
 176         if (unlikely(rc != (int)n_orphans)) {
 177                 struct bucket_header *hdr;
 178
 179                 objptr = bucket_stack_pop(bd->buckets[rte_lcore_id()]);
 180                 hdr = (struct bucket_header *)objptr;
 181
 182                 if (objptr == NULL) {
 183                         rc = rte_ring_dequeue(bd->shared_bucket_ring,
 184                                               (void **)&objptr);
 185                         if (rc != 0) {
 186                                 rte_errno = ENOBUFS;
 187                                 return -rte_errno;
 188                         }
 189                         hdr = (struct bucket_header *)objptr;
 190                         hdr->lcore_id = rte_lcore_id();
 191                 }
 192                 hdr->fill_cnt = 0;
 193                 bucket_fill_obj_table(bd, (void **)&objptr, obj_table,
 194                                       n_orphans);
 195                 for (i = n_orphans; i < bd->obj_per_bucket; i++,
 196                              objptr += bd->total_elt_size) {
 197                         rc = rte_ring_enqueue(bd->shared_orphan_ring,
 198                                               objptr);
 199                         if (rc != 0) {
 200                                 RTE_ASSERT(0);
 201                                 rte_errno = -rc;
 202                                 return rc;
 203                         }
 204                 }
 205         }
 206
 207         return 0;
 208 }
 209
 210 static int
 211 bucket_dequeue_buckets(struct bucket_data *bd, void **obj_table,
 212                        unsigned int n_buckets)
 213 {
 214         struct bucket_stack *cur_stack = bd->buckets[rte_lcore_id()];
 215         unsigned int n_buckets_from_stack = RTE_MIN(n_buckets, cur_stack->top);
 216         void **obj_table_base = obj_table;
 217
 218         n_buckets -= n_buckets_from_stack;
 219         while (n_buckets_from_stack-- > 0) {
 220                 void *obj = bucket_stack_pop_unsafe(cur_stack);
 221
 222                 obj_table = bucket_fill_obj_table(bd, &obj, obj_table,
 223                                                   bd->obj_per_bucket);
 224         }
 225         while (n_buckets-- > 0) {
 226                 struct bucket_header *hdr;
 227
 228                 if (unlikely(rte_ring_dequeue(bd->shared_bucket_ring,
 229                                               (void **)&hdr) != 0)) {
 230                         /*
 231                          * Return the already-dequeued buffers
 232                          * back to the mempool
 233                          */
 234                         bucket_enqueue(bd->pool, obj_table_base,
 235                                        obj_table - obj_table_base);
 236                         rte_errno = ENOBUFS;
 237                         return -rte_errno;
 238                 }
 239                 hdr->lcore_id = rte_lcore_id();
 240                 obj_table = bucket_fill_obj_table(bd, (void **)&hdr,
 241                                                   obj_table,
 242                                                   bd->obj_per_bucket);
 243         }
 244
 245         return 0;
 246 }
 247
 248 static int
 249 bucket_adopt_orphans(struct bucket_data *bd)
 250 {
 251         int rc = 0;
 252         struct rte_ring *adopt_ring =
 253                 bd->adoption_buffer_rings[rte_lcore_id()];
 254
 255         if (unlikely(!rte_ring_empty(adopt_ring))) {
 256                 void *orphan;
 257
 258                 while (rte_ring_sc_dequeue(adopt_ring, &orphan) == 0) {
 259                         rc = bucket_enqueue_single(bd, orphan);
 260                         RTE_ASSERT(rc == 0);
 261                 }
 262         }
 263         return rc;
 264 }
 265
 266 static int
 267 bucket_dequeue(struct rte_mempool *mp, void **obj_table, unsigned int n)
 268 {
 269         struct bucket_data *bd = mp->pool_data;
 270         unsigned int n_buckets = n / bd->obj_per_bucket;
 271         unsigned int n_orphans = n - n_buckets * bd->obj_per_bucket;
 272         int rc = 0;
 273
 274         bucket_adopt_orphans(bd);
 275
 276         if (unlikely(n_orphans > 0)) {
 277                 rc = bucket_dequeue_orphans(bd, obj_table +
 278                                             (n_buckets * bd->obj_per_bucket),
 279                                             n_orphans);
 280                 if (rc != 0)
 281                         return rc;
 282         }
 283
 284         if (likely(n_buckets > 0)) {
 285                 rc = bucket_dequeue_buckets(bd, obj_table, n_buckets);
 286                 if (unlikely(rc != 0) && n_orphans > 0) {
 287                         rte_ring_enqueue_bulk(bd->shared_orphan_ring,
 288                                               obj_table + (n_buckets *
 289                                                            bd->obj_per_bucket),
 290                                               n_orphans, NULL);
 291                 }
 292         }
 293
 294         return rc;
 295 }
 296
 297 static void
 298 count_underfilled_buckets(struct rte_mempool *mp,
 299                           void *opaque,
 300                           struct rte_mempool_memhdr *memhdr,
 301                           __rte_unused unsigned int mem_idx)
 302 {
 303         unsigned int *pcount = opaque;
 304         const struct bucket_data *bd = mp->pool_data;
 305         unsigned int bucket_page_sz =
 306                 (unsigned int)(~bd->bucket_page_mask + 1);
 307         uintptr_t align;
 308         uint8_t *iter;
 309
 310         align = (uintptr_t)RTE_PTR_ALIGN_CEIL(memhdr->addr, bucket_page_sz) -
 311                 (uintptr_t)memhdr->addr;
 312
 313         for (iter = (uint8_t *)memhdr->addr + align;
 314              iter < (uint8_t *)memhdr->addr + memhdr->len;
 315              iter += bucket_page_sz) {
 316                 struct bucket_header *hdr = (struct bucket_header *)iter;
 317
 318                 *pcount += hdr->fill_cnt;
 319         }
 320 }
 321
 322 static unsigned int
 323 bucket_get_count(const struct rte_mempool *mp)
 324 {
 325         const struct bucket_data *bd = mp->pool_data;
 326         unsigned int count =
 327                 bd->obj_per_bucket * rte_ring_count(bd->shared_bucket_ring) +
 328                 rte_ring_count(bd->shared_orphan_ring);
 329         unsigned int i;
 330
 331         for (i = 0; i < RTE_MAX_LCORE; i++) {
 332                 if (!rte_lcore_is_enabled(i))
 333                         continue;
 334                 count += bd->obj_per_bucket * bd->buckets[i]->top +
 335                         rte_ring_count(bd->adoption_buffer_rings[i]);
 336         }
 337
 338         rte_mempool_mem_iter((struct rte_mempool *)(uintptr_t)mp,
 339                              count_underfilled_buckets, &count);
 340
 341         return count;
 342 }
 343
 344 static int
 345 bucket_alloc(struct rte_mempool *mp)
 346 {
 347         int rg_flags = 0;
 348         int rc = 0;
 349         char rg_name[RTE_RING_NAMESIZE];
 350         struct bucket_data *bd;
 351         unsigned int i;
 352         unsigned int bucket_header_size;
 353
 354         bd = rte_zmalloc_socket("bucket_pool", sizeof(*bd),
 355                                 RTE_CACHE_LINE_SIZE, mp->socket_id);
 356         if (bd == NULL) {
 357                 rc = -ENOMEM;
 358                 goto no_mem_for_data;
 359         }
 360         bd->pool = mp;
 361         if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
 362                 bucket_header_size = sizeof(struct bucket_header);
 363         else
 364                 bucket_header_size = RTE_CACHE_LINE_SIZE;
 365         RTE_BUILD_BUG_ON(sizeof(struct bucket_header) > RTE_CACHE_LINE_SIZE);
 366         bd->header_size = mp->header_size + bucket_header_size;
 367         bd->total_elt_size = mp->header_size + mp->elt_size + mp->trailer_size;
 368         bd->bucket_mem_size = RTE_DRIVER_MEMPOOL_BUCKET_SIZE_KB * 1024;
 369         bd->obj_per_bucket = (bd->bucket_mem_size - bucket_header_size) /
 370                 bd->total_elt_size;
 371         bd->bucket_page_mask = ~(rte_align64pow2(bd->bucket_mem_size) - 1);
 372
 373         if (mp->flags & MEMPOOL_F_SP_PUT)
 374                 rg_flags |= RING_F_SP_ENQ;
 375         if (mp->flags & MEMPOOL_F_SC_GET)
 376                 rg_flags |= RING_F_SC_DEQ;
 377
 378         for (i = 0; i < RTE_MAX_LCORE; i++) {
 379                 if (!rte_lcore_is_enabled(i))
 380                         continue;
 381                 bd->buckets[i] =
 382                         bucket_stack_create(mp, mp->size / bd->obj_per_bucket);
 383                 if (bd->buckets[i] == NULL) {
 384                         rc = -ENOMEM;
 385                         goto no_mem_for_stacks;
 386                 }
 387                 rc = snprintf(rg_name, sizeof(rg_name),
 388                               RTE_MEMPOOL_MZ_FORMAT ".a%u", mp->name, i);
 389                 if (rc < 0 || rc >= (int)sizeof(rg_name)) {
 390                         rc = -ENAMETOOLONG;
 391                         goto no_mem_for_stacks;
 392                 }
 393                 bd->adoption_buffer_rings[i] =
 394                         rte_ring_create(rg_name, rte_align32pow2(mp->size + 1),
 395                                         mp->socket_id,
 396                                         rg_flags | RING_F_SC_DEQ);
 397                 if (bd->adoption_buffer_rings[i] == NULL) {
 398                         rc = -rte_errno;
 399                         goto no_mem_for_stacks;
 400                 }
 401         }
 402
 403         rc = snprintf(rg_name, sizeof(rg_name),
 404                       RTE_MEMPOOL_MZ_FORMAT ".0", mp->name);
 405         if (rc < 0 || rc >= (int)sizeof(rg_name)) {
 406                 rc = -ENAMETOOLONG;
 407                 goto invalid_shared_orphan_ring;
 408         }
 409         bd->shared_orphan_ring =
 410                 rte_ring_create(rg_name, rte_align32pow2(mp->size + 1),
 411                                 mp->socket_id, rg_flags);
 412         if (bd->shared_orphan_ring == NULL) {
 413                 rc = -rte_errno;
 414                 goto cannot_create_shared_orphan_ring;
 415         }
 416
 417         rc = snprintf(rg_name, sizeof(rg_name),
 418                        RTE_MEMPOOL_MZ_FORMAT ".1", mp->name);
 419         if (rc < 0 || rc >= (int)sizeof(rg_name)) {
 420                 rc = -ENAMETOOLONG;
 421                 goto invalid_shared_bucket_ring;
 422         }
 423         bd->shared_bucket_ring =
 424                 rte_ring_create(rg_name,
 425                                 rte_align32pow2((mp->size + 1) /
 426                                                 bd->obj_per_bucket),
 427                                 mp->socket_id, rg_flags);
 428         if (bd->shared_bucket_ring == NULL) {
 429                 rc = -rte_errno;
 430                 goto cannot_create_shared_bucket_ring;
 431         }
 432
 433         mp->pool_data = bd;
 434
 435         return 0;
 436
 437 cannot_create_shared_bucket_ring:
 438 invalid_shared_bucket_ring:
 439         rte_ring_free(bd->shared_orphan_ring);
 440 cannot_create_shared_orphan_ring:
 441 invalid_shared_orphan_ring:
 442 no_mem_for_stacks:
 443         for (i = 0; i < RTE_MAX_LCORE; i++) {
 444                 rte_free(bd->buckets[i]);
 445                 rte_ring_free(bd->adoption_buffer_rings[i]);
 446         }
 447         rte_free(bd);
 448 no_mem_for_data:
 449         rte_errno = -rc;
 450         return rc;
 451 }
 452
 453 static void
 454 bucket_free(struct rte_mempool *mp)
 455 {
 456         unsigned int i;
 457         struct bucket_data *bd = mp->pool_data;
 458
 459         if (bd == NULL)
 460                 return;
 461
 462         for (i = 0; i < RTE_MAX_LCORE; i++) {
 463                 rte_free(bd->buckets[i]);
 464                 rte_ring_free(bd->adoption_buffer_rings[i]);
 465         }
 466
 467         rte_ring_free(bd->shared_orphan_ring);
 468         rte_ring_free(bd->shared_bucket_ring);
 469
 470         rte_free(bd);
 471 }
 472
 473 static ssize_t
 474 bucket_calc_mem_size(const struct rte_mempool *mp, uint32_t obj_num,
 475                      __rte_unused uint32_t pg_shift, size_t *min_total_elt_size,
 476                      size_t *align)
 477 {
 478         struct bucket_data *bd = mp->pool_data;
 479         unsigned int bucket_page_sz;
 480
 481         if (bd == NULL)
 482                 return -EINVAL;
 483
 484         bucket_page_sz = rte_align32pow2(bd->bucket_mem_size);
 485         *align = bucket_page_sz;
 486         *min_total_elt_size = bucket_page_sz;
 487         /*
 488          * Each bucket occupies its own block aligned to
 489          * bucket_page_sz, so the required amount of memory is
 490          * a multiple of bucket_page_sz.
 491          * We also need extra space for a bucket header
 492          */
 493         return ((obj_num + bd->obj_per_bucket - 1) /
 494                 bd->obj_per_bucket) * bucket_page_sz;
 495 }
 496
 497 static int
 498 bucket_populate(struct rte_mempool *mp, unsigned int max_objs,
 499                 void *vaddr, rte_iova_t iova, size_t len,
 500                 rte_mempool_populate_obj_cb_t *obj_cb, void *obj_cb_arg)
 501 {
 502         struct bucket_data *bd = mp->pool_data;
 503         unsigned int bucket_page_sz;
 504         unsigned int bucket_header_sz;
 505         unsigned int n_objs;
 506         uintptr_t align;
 507         uint8_t *iter;
 508         int rc;
 509
 510         if (bd == NULL)
 511                 return -EINVAL;
 512
 513         bucket_page_sz = rte_align32pow2(bd->bucket_mem_size);
 514         align = RTE_PTR_ALIGN_CEIL((uintptr_t)vaddr, bucket_page_sz) -
 515                 (uintptr_t)vaddr;
 516
 517         bucket_header_sz = bd->header_size - mp->header_size;
 518         if (iova != RTE_BAD_IOVA)
 519                 iova += align + bucket_header_sz;
 520
 521         for (iter = (uint8_t *)vaddr + align, n_objs = 0;
 522              iter < (uint8_t *)vaddr + len && n_objs < max_objs;
 523              iter += bucket_page_sz) {
 524                 struct bucket_header *hdr = (struct bucket_header *)iter;
 525                 unsigned int chunk_len = bd->bucket_mem_size;
 526
 527                 if ((size_t)(iter - (uint8_t *)vaddr) + chunk_len > len)
 528                         chunk_len = len - (iter - (uint8_t *)vaddr);
 529                 if (chunk_len <= bucket_header_sz)
 530                         break;
 531                 chunk_len -= bucket_header_sz;
 532
 533                 hdr->fill_cnt = 0;
 534                 hdr->lcore_id = LCORE_ID_ANY;
 535                 rc = rte_mempool_op_populate_default(mp,
 536                                                      RTE_MIN(bd->obj_per_bucket,
 537                                                              max_objs - n_objs),
 538                                                      iter + bucket_header_sz,
 539                                                      iova, chunk_len,
 540                                                      obj_cb, obj_cb_arg);
 541                 if (rc < 0)
 542                         return rc;
 543                 n_objs += rc;
 544                 if (iova != RTE_BAD_IOVA)
 545                         iova += bucket_page_sz;
 546         }
 547
 548         return n_objs;
 549 }
 550
 551 static const struct rte_mempool_ops ops_bucket = {
 552         .name = "bucket",
 553         .alloc = bucket_alloc,
 554         .free = bucket_free,
 555         .enqueue = bucket_enqueue,
 556         .dequeue = bucket_dequeue,
 557         .get_count = bucket_get_count,
 558         .calc_mem_size = bucket_calc_mem_size,
 559         .populate = bucket_populate,
 560 };
 561
 562
 563 MEMPOOL_REGISTER_OPS(ops_bucket);