lib/librte_lpm/rte_lpm.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2014 Intel Corporation
   3  * Copyright(c) 2020 Arm Limited
   4  */
   5
   6 #include <string.h>
   7 #include <stdint.h>
   8 #include <errno.h>
   9 #include <stdarg.h>
  10 #include <stdio.h>
  11 #include <sys/queue.h>
  12
  13 #include <rte_log.h>
  14 #include <rte_branch_prediction.h>
  15 #include <rte_common.h>
  16 #include <rte_memory.h>        /* for definition of RTE_CACHE_LINE_SIZE */
  17 #include <rte_malloc.h>
  18 #include <rte_eal.h>
  19 #include <rte_eal_memconfig.h>
  20 #include <rte_per_lcore.h>
  21 #include <rte_string_fns.h>
  22 #include <rte_errno.h>
  23 #include <rte_rwlock.h>
  24 #include <rte_spinlock.h>
  25 #include <rte_tailq.h>
  26
  27 #include "rte_lpm.h"
  28
  29 TAILQ_HEAD(rte_lpm_list, rte_tailq_entry);
  30
  31 static struct rte_tailq_elem rte_lpm_tailq = {
  32         .name = "RTE_LPM",
  33 };
  34 EAL_REGISTER_TAILQ(rte_lpm_tailq)
  35
  36 #define MAX_DEPTH_TBL24 24
  37
  38 enum valid_flag {
  39         INVALID = 0,
  40         VALID
  41 };
  42
  43 /** @internal LPM structure. */
  44 struct __rte_lpm {
  45         /* LPM metadata. */
  46         struct rte_lpm lpm;
  47
  48         /* RCU config. */
  49         struct rte_rcu_qsbr *v;         /* RCU QSBR variable. */
  50         enum rte_lpm_qsbr_mode rcu_mode;/* Blocking, defer queue. */
  51         struct rte_rcu_qsbr_dq *dq;     /* RCU QSBR defer queue. */
  52 };
  53
  54 /* Macro to enable/disable run-time checks. */
  55 #if defined(RTE_LIBRTE_LPM_DEBUG)
  56 #include <rte_debug.h>
  57 #define VERIFY_DEPTH(depth) do {                                \
  58         if ((depth == 0) || (depth > RTE_LPM_MAX_DEPTH))        \
  59                 rte_panic("LPM: Invalid depth (%u) at line %d", \
  60                                 (unsigned)(depth), __LINE__);   \
  61 } while (0)
  62 #else
  63 #define VERIFY_DEPTH(depth)
  64 #endif
  65
  66 /*
  67  * Converts a given depth value to its corresponding mask value.
  68  *
  69  * depth  (IN)          : range = 1 - 32
  70  * mask   (OUT)         : 32bit mask
  71  */
  72 static uint32_t __attribute__((pure))
  73 depth_to_mask(uint8_t depth)
  74 {
  75         VERIFY_DEPTH(depth);
  76
  77         /* To calculate a mask start with a 1 on the left hand side and right
  78          * shift while populating the left hand side with 1's
  79          */
  80         return (int)0x80000000 >> (depth - 1);
  81 }
  82
  83 /*
  84  * Converts given depth value to its corresponding range value.
  85  */
  86 static uint32_t __attribute__((pure))
  87 depth_to_range(uint8_t depth)
  88 {
  89         VERIFY_DEPTH(depth);
  90
  91         /*
  92          * Calculate tbl24 range. (Note: 2^depth = 1 << depth)
  93          */
  94         if (depth <= MAX_DEPTH_TBL24)
  95                 return 1 << (MAX_DEPTH_TBL24 - depth);
  96
  97         /* Else if depth is greater than 24 */
  98         return 1 << (RTE_LPM_MAX_DEPTH - depth);
  99 }
 100
 101 /*
 102  * Find an existing lpm table and return a pointer to it.
 103  */
 104 struct rte_lpm *
 105 rte_lpm_find_existing(const char *name)
 106 {
 107         struct rte_lpm *l = NULL;
 108         struct rte_tailq_entry *te;
 109         struct rte_lpm_list *lpm_list;
 110
 111         lpm_list = RTE_TAILQ_CAST(rte_lpm_tailq.head, rte_lpm_list);
 112
 113         rte_mcfg_tailq_read_lock();
 114         TAILQ_FOREACH(te, lpm_list, next) {
 115                 l = te->data;
 116                 if (strncmp(name, l->name, RTE_LPM_NAMESIZE) == 0)
 117                         break;
 118         }
 119         rte_mcfg_tailq_read_unlock();
 120
 121         if (te == NULL) {
 122                 rte_errno = ENOENT;
 123                 return NULL;
 124         }
 125
 126         return l;
 127 }
 128
 129 /*
 130  * Allocates memory for LPM object
 131  */
 132 struct rte_lpm *
 133 rte_lpm_create(const char *name, int socket_id,
 134                 const struct rte_lpm_config *config)
 135 {
 136         char mem_name[RTE_LPM_NAMESIZE];
 137         struct __rte_lpm *internal_lpm;
 138         struct rte_lpm *lpm = NULL;
 139         struct rte_tailq_entry *te;
 140         uint32_t mem_size, rules_size, tbl8s_size;
 141         struct rte_lpm_list *lpm_list;
 142
 143         lpm_list = RTE_TAILQ_CAST(rte_lpm_tailq.head, rte_lpm_list);
 144
 145         RTE_BUILD_BUG_ON(sizeof(struct rte_lpm_tbl_entry) != 4);
 146
 147         /* Check user arguments. */
 148         if ((name == NULL) || (socket_id < -1) || (config->max_rules == 0)
 149                         || config->number_tbl8s > RTE_LPM_MAX_TBL8_NUM_GROUPS) {
 150                 rte_errno = EINVAL;
 151                 return NULL;
 152         }
 153
 154         snprintf(mem_name, sizeof(mem_name), "LPM_%s", name);
 155
 156         rte_mcfg_tailq_write_lock();
 157
 158         /* guarantee there's no existing */
 159         TAILQ_FOREACH(te, lpm_list, next) {
 160                 lpm = te->data;
 161                 if (strncmp(name, lpm->name, RTE_LPM_NAMESIZE) == 0)
 162                         break;
 163         }
 164
 165         if (te != NULL) {
 166                 lpm = NULL;
 167                 rte_errno = EEXIST;
 168                 goto exit;
 169         }
 170
 171         /* Determine the amount of memory to allocate. */
 172         mem_size = sizeof(*internal_lpm);
 173         rules_size = sizeof(struct rte_lpm_rule) * config->max_rules;
 174         tbl8s_size = sizeof(struct rte_lpm_tbl_entry) *
 175                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES * config->number_tbl8s;
 176
 177         /* allocate tailq entry */
 178         te = rte_zmalloc("LPM_TAILQ_ENTRY", sizeof(*te), 0);
 179         if (te == NULL) {
 180                 RTE_LOG(ERR, LPM, "Failed to allocate tailq entry\n");
 181                 rte_errno = ENOMEM;
 182                 goto exit;
 183         }
 184
 185         /* Allocate memory to store the LPM data structures. */
 186         internal_lpm = rte_zmalloc_socket(mem_name, mem_size,
 187                         RTE_CACHE_LINE_SIZE, socket_id);
 188         if (internal_lpm == NULL) {
 189                 RTE_LOG(ERR, LPM, "LPM memory allocation failed\n");
 190                 rte_free(te);
 191                 rte_errno = ENOMEM;
 192                 goto exit;
 193         }
 194
 195         lpm = &internal_lpm->lpm;
 196         lpm->rules_tbl = rte_zmalloc_socket(NULL,
 197                         (size_t)rules_size, RTE_CACHE_LINE_SIZE, socket_id);
 198
 199         if (lpm->rules_tbl == NULL) {
 200                 RTE_LOG(ERR, LPM, "LPM rules_tbl memory allocation failed\n");
 201                 rte_free(internal_lpm);
 202                 internal_lpm = NULL;
 203                 lpm = NULL;
 204                 rte_free(te);
 205                 rte_errno = ENOMEM;
 206                 goto exit;
 207         }
 208
 209         lpm->tbl8 = rte_zmalloc_socket(NULL,
 210                         (size_t)tbl8s_size, RTE_CACHE_LINE_SIZE, socket_id);
 211
 212         if (lpm->tbl8 == NULL) {
 213                 RTE_LOG(ERR, LPM, "LPM tbl8 memory allocation failed\n");
 214                 rte_free(lpm->rules_tbl);
 215                 rte_free(internal_lpm);
 216                 internal_lpm = NULL;
 217                 lpm = NULL;
 218                 rte_free(te);
 219                 rte_errno = ENOMEM;
 220                 goto exit;
 221         }
 222
 223         /* Save user arguments. */
 224         lpm->max_rules = config->max_rules;
 225         lpm->number_tbl8s = config->number_tbl8s;
 226         strlcpy(lpm->name, name, sizeof(lpm->name));
 227
 228         te->data = lpm;
 229
 230         TAILQ_INSERT_TAIL(lpm_list, te, next);
 231
 232 exit:
 233         rte_mcfg_tailq_write_unlock();
 234
 235         return lpm;
 236 }
 237
 238 /*
 239  * Deallocates memory for given LPM table.
 240  */
 241 void
 242 rte_lpm_free(struct rte_lpm *lpm)
 243 {
 244         struct __rte_lpm *internal_lpm;
 245         struct rte_lpm_list *lpm_list;
 246         struct rte_tailq_entry *te;
 247
 248         /* Check user arguments. */
 249         if (lpm == NULL)
 250                 return;
 251
 252         lpm_list = RTE_TAILQ_CAST(rte_lpm_tailq.head, rte_lpm_list);
 253
 254         rte_mcfg_tailq_write_lock();
 255
 256         /* find our tailq entry */
 257         TAILQ_FOREACH(te, lpm_list, next) {
 258                 if (te->data == (void *) lpm)
 259                         break;
 260         }
 261         if (te != NULL)
 262                 TAILQ_REMOVE(lpm_list, te, next);
 263
 264         rte_mcfg_tailq_write_unlock();
 265
 266         internal_lpm = container_of(lpm, struct __rte_lpm, lpm);
 267         if (internal_lpm->dq != NULL)
 268                 rte_rcu_qsbr_dq_delete(internal_lpm->dq);
 269         rte_free(lpm->tbl8);
 270         rte_free(lpm->rules_tbl);
 271         rte_free(lpm);
 272         rte_free(te);
 273 }
 274
 275 static void
 276 __lpm_rcu_qsbr_free_resource(void *p, void *data, unsigned int n)
 277 {
 278         struct rte_lpm_tbl_entry *tbl8 = ((struct rte_lpm *)p)->tbl8;
 279         struct rte_lpm_tbl_entry zero_tbl8_entry = {0};
 280         uint32_t tbl8_group_index = *(uint32_t *)data;
 281
 282         RTE_SET_USED(n);
 283         /* Set tbl8 group invalid */
 284         __atomic_store(&tbl8[tbl8_group_index], &zero_tbl8_entry,
 285                 __ATOMIC_RELAXED);
 286 }
 287
 288 /* Associate QSBR variable with an LPM object.
 289  */
 290 int
 291 rte_lpm_rcu_qsbr_add(struct rte_lpm *lpm, struct rte_lpm_rcu_config *cfg,
 292         struct rte_rcu_qsbr_dq **dq)
 293 {
 294         struct rte_rcu_qsbr_dq_parameters params = {0};
 295         char rcu_dq_name[RTE_RCU_QSBR_DQ_NAMESIZE];
 296         struct __rte_lpm *internal_lpm;
 297
 298         if (lpm == NULL || cfg == NULL) {
 299                 rte_errno = EINVAL;
 300                 return 1;
 301         }
 302
 303         internal_lpm = container_of(lpm, struct __rte_lpm, lpm);
 304         if (internal_lpm->v != NULL) {
 305                 rte_errno = EEXIST;
 306                 return 1;
 307         }
 308
 309         if (cfg->mode == RTE_LPM_QSBR_MODE_SYNC) {
 310                 /* No other things to do. */
 311         } else if (cfg->mode == RTE_LPM_QSBR_MODE_DQ) {
 312                 /* Init QSBR defer queue. */
 313                 snprintf(rcu_dq_name, sizeof(rcu_dq_name),
 314                                 "LPM_RCU_%s", lpm->name);
 315                 params.name = rcu_dq_name;
 316                 params.size = cfg->dq_size;
 317                 if (params.size == 0)
 318                         params.size = lpm->number_tbl8s;
 319                 params.trigger_reclaim_limit = cfg->reclaim_thd;
 320                 params.max_reclaim_size = cfg->reclaim_max;
 321                 if (params.max_reclaim_size == 0)
 322                         params.max_reclaim_size = RTE_LPM_RCU_DQ_RECLAIM_MAX;
 323                 params.esize = sizeof(uint32_t);        /* tbl8 group index */
 324                 params.free_fn = __lpm_rcu_qsbr_free_resource;
 325                 params.p = lpm;
 326                 params.v = cfg->v;
 327                 internal_lpm->dq = rte_rcu_qsbr_dq_create(&params);
 328                 if (internal_lpm->dq == NULL) {
 329                         RTE_LOG(ERR, LPM, "LPM defer queue creation failed\n");
 330                         return 1;
 331                 }
 332                 if (dq != NULL)
 333                         *dq = internal_lpm->dq;
 334         } else {
 335                 rte_errno = EINVAL;
 336                 return 1;
 337         }
 338         internal_lpm->rcu_mode = cfg->mode;
 339         internal_lpm->v = cfg->v;
 340
 341         return 0;
 342 }
 343
 344 /*
 345  * Adds a rule to the rule table.
 346  *
 347  * NOTE: The rule table is split into 32 groups. Each group contains rules that
 348  * apply to a specific prefix depth (i.e. group 1 contains rules that apply to
 349  * prefixes with a depth of 1 etc.). In the following code (depth - 1) is used
 350  * to refer to depth 1 because even though the depth range is 1 - 32, depths
 351  * are stored in the rule table from 0 - 31.
 352  * NOTE: Valid range for depth parameter is 1 .. 32 inclusive.
 353  */
 354 static int32_t
 355 rule_add(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth,
 356         uint32_t next_hop)
 357 {
 358         uint32_t rule_gindex, rule_index, last_rule;
 359         int i;
 360
 361         VERIFY_DEPTH(depth);
 362
 363         /* Scan through rule group to see if rule already exists. */
 364         if (lpm->rule_info[depth - 1].used_rules > 0) {
 365
 366                 /* rule_gindex stands for rule group index. */
 367                 rule_gindex = lpm->rule_info[depth - 1].first_rule;
 368                 /* Initialise rule_index to point to start of rule group. */
 369                 rule_index = rule_gindex;
 370                 /* Last rule = Last used rule in this rule group. */
 371                 last_rule = rule_gindex + lpm->rule_info[depth - 1].used_rules;
 372
 373                 for (; rule_index < last_rule; rule_index++) {
 374
 375                         /* If rule already exists update next hop and return. */
 376                         if (lpm->rules_tbl[rule_index].ip == ip_masked) {
 377
 378                                 if (lpm->rules_tbl[rule_index].next_hop
 379                                                 == next_hop)
 380                                         return -EEXIST;
 381                                 lpm->rules_tbl[rule_index].next_hop = next_hop;
 382
 383                                 return rule_index;
 384                         }
 385                 }
 386
 387                 if (rule_index == lpm->max_rules)
 388                         return -ENOSPC;
 389         } else {
 390                 /* Calculate the position in which the rule will be stored. */
 391                 rule_index = 0;
 392
 393                 for (i = depth - 1; i > 0; i--) {
 394                         if (lpm->rule_info[i - 1].used_rules > 0) {
 395                                 rule_index = lpm->rule_info[i - 1].first_rule
 396                                                 + lpm->rule_info[i - 1].used_rules;
 397                                 break;
 398                         }
 399                 }
 400                 if (rule_index == lpm->max_rules)
 401                         return -ENOSPC;
 402
 403                 lpm->rule_info[depth - 1].first_rule = rule_index;
 404         }
 405
 406         /* Make room for the new rule in the array. */
 407         for (i = RTE_LPM_MAX_DEPTH; i > depth; i--) {
 408                 if (lpm->rule_info[i - 1].first_rule
 409                                 + lpm->rule_info[i - 1].used_rules == lpm->max_rules)
 410                         return -ENOSPC;
 411
 412                 if (lpm->rule_info[i - 1].used_rules > 0) {
 413                         lpm->rules_tbl[lpm->rule_info[i - 1].first_rule
 414                                 + lpm->rule_info[i - 1].used_rules]
 415                                         = lpm->rules_tbl[lpm->rule_info[i - 1].first_rule];
 416                         lpm->rule_info[i - 1].first_rule++;
 417                 }
 418         }
 419
 420         /* Add the new rule. */
 421         lpm->rules_tbl[rule_index].ip = ip_masked;
 422         lpm->rules_tbl[rule_index].next_hop = next_hop;
 423
 424         /* Increment the used rules counter for this rule group. */
 425         lpm->rule_info[depth - 1].used_rules++;
 426
 427         return rule_index;
 428 }
 429
 430 /*
 431  * Delete a rule from the rule table.
 432  * NOTE: Valid range for depth parameter is 1 .. 32 inclusive.
 433  */
 434 static void
 435 rule_delete(struct rte_lpm *lpm, int32_t rule_index, uint8_t depth)
 436 {
 437         int i;
 438
 439         VERIFY_DEPTH(depth);
 440
 441         lpm->rules_tbl[rule_index] =
 442                         lpm->rules_tbl[lpm->rule_info[depth - 1].first_rule
 443                         + lpm->rule_info[depth - 1].used_rules - 1];
 444
 445         for (i = depth; i < RTE_LPM_MAX_DEPTH; i++) {
 446                 if (lpm->rule_info[i].used_rules > 0) {
 447                         lpm->rules_tbl[lpm->rule_info[i].first_rule - 1] =
 448                                         lpm->rules_tbl[lpm->rule_info[i].first_rule
 449                                                 + lpm->rule_info[i].used_rules - 1];
 450                         lpm->rule_info[i].first_rule--;
 451                 }
 452         }
 453
 454         lpm->rule_info[depth - 1].used_rules--;
 455 }
 456
 457 /*
 458  * Finds a rule in rule table.
 459  * NOTE: Valid range for depth parameter is 1 .. 32 inclusive.
 460  */
 461 static int32_t
 462 rule_find(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth)
 463 {
 464         uint32_t rule_gindex, last_rule, rule_index;
 465
 466         VERIFY_DEPTH(depth);
 467
 468         rule_gindex = lpm->rule_info[depth - 1].first_rule;
 469         last_rule = rule_gindex + lpm->rule_info[depth - 1].used_rules;
 470
 471         /* Scan used rules at given depth to find rule. */
 472         for (rule_index = rule_gindex; rule_index < last_rule; rule_index++) {
 473                 /* If rule is found return the rule index. */
 474                 if (lpm->rules_tbl[rule_index].ip == ip_masked)
 475                         return rule_index;
 476         }
 477
 478         /* If rule is not found return -EINVAL. */
 479         return -EINVAL;
 480 }
 481
 482 /*
 483  * Find, clean and allocate a tbl8.
 484  */
 485 static int32_t
 486 _tbl8_alloc(struct rte_lpm *lpm)
 487 {
 488         uint32_t group_idx; /* tbl8 group index. */
 489         struct rte_lpm_tbl_entry *tbl8_entry;
 490
 491         /* Scan through tbl8 to find a free (i.e. INVALID) tbl8 group. */
 492         for (group_idx = 0; group_idx < lpm->number_tbl8s; group_idx++) {
 493                 tbl8_entry = &lpm->tbl8[group_idx *
 494                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES];
 495                 /* If a free tbl8 group is found clean it and set as VALID. */
 496                 if (!tbl8_entry->valid_group) {
 497                         struct rte_lpm_tbl_entry new_tbl8_entry = {
 498                                 .next_hop = 0,
 499                                 .valid = INVALID,
 500                                 .depth = 0,
 501                                 .valid_group = VALID,
 502                         };
 503
 504                         memset(&tbl8_entry[0], 0,
 505                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES *
 506                                         sizeof(tbl8_entry[0]));
 507
 508                         __atomic_store(tbl8_entry, &new_tbl8_entry,
 509                                         __ATOMIC_RELAXED);
 510
 511                         /* Return group index for allocated tbl8 group. */
 512                         return group_idx;
 513                 }
 514         }
 515
 516         /* If there are no tbl8 groups free then return error. */
 517         return -ENOSPC;
 518 }
 519
 520 static int32_t
 521 tbl8_alloc(struct rte_lpm *lpm)
 522 {
 523         int32_t group_idx; /* tbl8 group index. */
 524         struct __rte_lpm *internal_lpm;
 525
 526         internal_lpm = container_of(lpm, struct __rte_lpm, lpm);
 527         group_idx = _tbl8_alloc(lpm);
 528         if (group_idx == -ENOSPC && internal_lpm->dq != NULL) {
 529                 /* If there are no tbl8 groups try to reclaim one. */
 530                 if (rte_rcu_qsbr_dq_reclaim(internal_lpm->dq, 1,
 531                                 NULL, NULL, NULL) == 0)
 532                         group_idx = _tbl8_alloc(lpm);
 533         }
 534
 535         return group_idx;
 536 }
 537
 538 static void
 539 tbl8_free(struct rte_lpm *lpm, uint32_t tbl8_group_start)
 540 {
 541         struct rte_lpm_tbl_entry zero_tbl8_entry = {0};
 542         struct __rte_lpm *internal_lpm;
 543
 544         internal_lpm = container_of(lpm, struct __rte_lpm, lpm);
 545         if (internal_lpm->v == NULL) {
 546                 /* Set tbl8 group invalid*/
 547                 __atomic_store(&lpm->tbl8[tbl8_group_start], &zero_tbl8_entry,
 548                                 __ATOMIC_RELAXED);
 549         } else if (internal_lpm->rcu_mode == RTE_LPM_QSBR_MODE_SYNC) {
 550                 /* Wait for quiescent state change. */
 551                 rte_rcu_qsbr_synchronize(internal_lpm->v,
 552                         RTE_QSBR_THRID_INVALID);
 553                 /* Set tbl8 group invalid*/
 554                 __atomic_store(&lpm->tbl8[tbl8_group_start], &zero_tbl8_entry,
 555                                 __ATOMIC_RELAXED);
 556         } else if (internal_lpm->rcu_mode == RTE_LPM_QSBR_MODE_DQ) {
 557                 /* Push into QSBR defer queue. */
 558                 rte_rcu_qsbr_dq_enqueue(internal_lpm->dq,
 559                                 (void *)&tbl8_group_start);
 560         }
 561 }
 562
 563 static __rte_noinline int32_t
 564 add_depth_small(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
 565                 uint32_t next_hop)
 566 {
 567 #define group_idx next_hop
 568         uint32_t tbl24_index, tbl24_range, tbl8_index, tbl8_group_end, i, j;
 569
 570         /* Calculate the index into Table24. */
 571         tbl24_index = ip >> 8;
 572         tbl24_range = depth_to_range(depth);
 573
 574         for (i = tbl24_index; i < (tbl24_index + tbl24_range); i++) {
 575                 /*
 576                  * For invalid OR valid and non-extended tbl 24 entries set
 577                  * entry.
 578                  */
 579                 if (!lpm->tbl24[i].valid || (lpm->tbl24[i].valid_group == 0 &&
 580                                 lpm->tbl24[i].depth <= depth)) {
 581
 582                         struct rte_lpm_tbl_entry new_tbl24_entry = {
 583                                 .next_hop = next_hop,
 584                                 .valid = VALID,
 585                                 .valid_group = 0,
 586                                 .depth = depth,
 587                         };
 588
 589                         /* Setting tbl24 entry in one go to avoid race
 590                          * conditions
 591                          */
 592                         __atomic_store(&lpm->tbl24[i], &new_tbl24_entry,
 593                                         __ATOMIC_RELEASE);
 594
 595                         continue;
 596                 }
 597
 598                 if (lpm->tbl24[i].valid_group == 1) {
 599                         /* If tbl24 entry is valid and extended calculate the
 600                          *  index into tbl8.
 601                          */
 602                         tbl8_index = lpm->tbl24[i].group_idx *
 603                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 604                         tbl8_group_end = tbl8_index +
 605                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 606
 607                         for (j = tbl8_index; j < tbl8_group_end; j++) {
 608                                 if (!lpm->tbl8[j].valid ||
 609                                                 lpm->tbl8[j].depth <= depth) {
 610                                         struct rte_lpm_tbl_entry
 611                                                 new_tbl8_entry = {
 612                                                 .valid = VALID,
 613                                                 .valid_group = VALID,
 614                                                 .depth = depth,
 615                                                 .next_hop = next_hop,
 616                                         };
 617
 618                                         /*
 619                                          * Setting tbl8 entry in one go to avoid
 620                                          * race conditions
 621                                          */
 622                                         __atomic_store(&lpm->tbl8[j],
 623                                                 &new_tbl8_entry,
 624                                                 __ATOMIC_RELAXED);
 625
 626                                         continue;
 627                                 }
 628                         }
 629                 }
 630         }
 631 #undef group_idx
 632         return 0;
 633 }
 634
 635 static __rte_noinline int32_t
 636 add_depth_big(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth,
 637                 uint32_t next_hop)
 638 {
 639 #define group_idx next_hop
 640         uint32_t tbl24_index;
 641         int32_t tbl8_group_index, tbl8_group_start, tbl8_group_end, tbl8_index,
 642                 tbl8_range, i;
 643
 644         tbl24_index = (ip_masked >> 8);
 645         tbl8_range = depth_to_range(depth);
 646
 647         if (!lpm->tbl24[tbl24_index].valid) {
 648                 /* Search for a free tbl8 group. */
 649                 tbl8_group_index = tbl8_alloc(lpm);
 650
 651                 /* Check tbl8 allocation was successful. */
 652                 if (tbl8_group_index < 0) {
 653                         return tbl8_group_index;
 654                 }
 655
 656                 /* Find index into tbl8 and range. */
 657                 tbl8_index = (tbl8_group_index *
 658                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES) +
 659                                 (ip_masked & 0xFF);
 660
 661                 /* Set tbl8 entry. */
 662                 for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
 663                         struct rte_lpm_tbl_entry new_tbl8_entry = {
 664                                 .valid = VALID,
 665                                 .depth = depth,
 666                                 .valid_group = lpm->tbl8[i].valid_group,
 667                                 .next_hop = next_hop,
 668                         };
 669                         __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
 670                                         __ATOMIC_RELAXED);
 671                 }
 672
 673                 /*
 674                  * Update tbl24 entry to point to new tbl8 entry. Note: The
 675                  * ext_flag and tbl8_index need to be updated simultaneously,
 676                  * so assign whole structure in one go
 677                  */
 678
 679                 struct rte_lpm_tbl_entry new_tbl24_entry = {
 680                         .group_idx = tbl8_group_index,
 681                         .valid = VALID,
 682                         .valid_group = 1,
 683                         .depth = 0,
 684                 };
 685
 686                 /* The tbl24 entry must be written only after the
 687                  * tbl8 entries are written.
 688                  */
 689                 __atomic_store(&lpm->tbl24[tbl24_index], &new_tbl24_entry,
 690                                 __ATOMIC_RELEASE);
 691
 692         } /* If valid entry but not extended calculate the index into Table8. */
 693         else if (lpm->tbl24[tbl24_index].valid_group == 0) {
 694                 /* Search for free tbl8 group. */
 695                 tbl8_group_index = tbl8_alloc(lpm);
 696
 697                 if (tbl8_group_index < 0) {
 698                         return tbl8_group_index;
 699                 }
 700
 701                 tbl8_group_start = tbl8_group_index *
 702                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 703                 tbl8_group_end = tbl8_group_start +
 704                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 705
 706                 /* Populate new tbl8 with tbl24 value. */
 707                 for (i = tbl8_group_start; i < tbl8_group_end; i++) {
 708                         struct rte_lpm_tbl_entry new_tbl8_entry = {
 709                                 .valid = VALID,
 710                                 .depth = lpm->tbl24[tbl24_index].depth,
 711                                 .valid_group = lpm->tbl8[i].valid_group,
 712                                 .next_hop = lpm->tbl24[tbl24_index].next_hop,
 713                         };
 714                         __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
 715                                         __ATOMIC_RELAXED);
 716                 }
 717
 718                 tbl8_index = tbl8_group_start + (ip_masked & 0xFF);
 719
 720                 /* Insert new rule into the tbl8 entry. */
 721                 for (i = tbl8_index; i < tbl8_index + tbl8_range; i++) {
 722                         struct rte_lpm_tbl_entry new_tbl8_entry = {
 723                                 .valid = VALID,
 724                                 .depth = depth,
 725                                 .valid_group = lpm->tbl8[i].valid_group,
 726                                 .next_hop = next_hop,
 727                         };
 728                         __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
 729                                         __ATOMIC_RELAXED);
 730                 }
 731
 732                 /*
 733                  * Update tbl24 entry to point to new tbl8 entry. Note: The
 734                  * ext_flag and tbl8_index need to be updated simultaneously,
 735                  * so assign whole structure in one go.
 736                  */
 737
 738                 struct rte_lpm_tbl_entry new_tbl24_entry = {
 739                                 .group_idx = tbl8_group_index,
 740                                 .valid = VALID,
 741                                 .valid_group = 1,
 742                                 .depth = 0,
 743                 };
 744
 745                 /* The tbl24 entry must be written only after the
 746                  * tbl8 entries are written.
 747                  */
 748                 __atomic_store(&lpm->tbl24[tbl24_index], &new_tbl24_entry,
 749                                 __ATOMIC_RELEASE);
 750
 751         } else { /*
 752                 * If it is valid, extended entry calculate the index into tbl8.
 753                 */
 754                 tbl8_group_index = lpm->tbl24[tbl24_index].group_idx;
 755                 tbl8_group_start = tbl8_group_index *
 756                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 757                 tbl8_index = tbl8_group_start + (ip_masked & 0xFF);
 758
 759                 for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
 760
 761                         if (!lpm->tbl8[i].valid ||
 762                                         lpm->tbl8[i].depth <= depth) {
 763                                 struct rte_lpm_tbl_entry new_tbl8_entry = {
 764                                         .valid = VALID,
 765                                         .depth = depth,
 766                                         .next_hop = next_hop,
 767                                         .valid_group = lpm->tbl8[i].valid_group,
 768                                 };
 769
 770                                 /*
 771                                  * Setting tbl8 entry in one go to avoid race
 772                                  * condition
 773                                  */
 774                                 __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
 775                                                 __ATOMIC_RELAXED);
 776
 777                                 continue;
 778                         }
 779                 }
 780         }
 781 #undef group_idx
 782         return 0;
 783 }
 784
 785 /*
 786  * Add a route
 787  */
 788 int
 789 rte_lpm_add(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
 790                 uint32_t next_hop)
 791 {
 792         int32_t rule_index, status = 0;
 793         uint32_t ip_masked;
 794
 795         /* Check user arguments. */
 796         if ((lpm == NULL) || (depth < 1) || (depth > RTE_LPM_MAX_DEPTH))
 797                 return -EINVAL;
 798
 799         ip_masked = ip & depth_to_mask(depth);
 800
 801         /* Add the rule to the rule table. */
 802         rule_index = rule_add(lpm, ip_masked, depth, next_hop);
 803
 804         /* Skip table entries update if The rule is the same as
 805          * the rule in the rules table.
 806          */
 807         if (rule_index == -EEXIST)
 808                 return 0;
 809
 810         /* If the is no space available for new rule return error. */
 811         if (rule_index < 0) {
 812                 return rule_index;
 813         }
 814
 815         if (depth <= MAX_DEPTH_TBL24) {
 816                 status = add_depth_small(lpm, ip_masked, depth, next_hop);
 817         } else { /* If depth > RTE_LPM_MAX_DEPTH_TBL24 */
 818                 status = add_depth_big(lpm, ip_masked, depth, next_hop);
 819
 820                 /*
 821                  * If add fails due to exhaustion of tbl8 extensions delete
 822                  * rule that was added to rule table.
 823                  */
 824                 if (status < 0) {
 825                         rule_delete(lpm, rule_index, depth);
 826
 827                         return status;
 828                 }
 829         }
 830
 831         return 0;
 832 }
 833
 834 /*
 835  * Look for a rule in the high-level rules table
 836  */
 837 int
 838 rte_lpm_is_rule_present(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
 839 uint32_t *next_hop)
 840 {
 841         uint32_t ip_masked;
 842         int32_t rule_index;
 843
 844         /* Check user arguments. */
 845         if ((lpm == NULL) ||
 846                 (next_hop == NULL) ||
 847                 (depth < 1) || (depth > RTE_LPM_MAX_DEPTH))
 848                 return -EINVAL;
 849
 850         /* Look for the rule using rule_find. */
 851         ip_masked = ip & depth_to_mask(depth);
 852         rule_index = rule_find(lpm, ip_masked, depth);
 853
 854         if (rule_index >= 0) {
 855                 *next_hop = lpm->rules_tbl[rule_index].next_hop;
 856                 return 1;
 857         }
 858
 859         /* If rule is not found return 0. */
 860         return 0;
 861 }
 862
 863 static int32_t
 864 find_previous_rule(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
 865                 uint8_t *sub_rule_depth)
 866 {
 867         int32_t rule_index;
 868         uint32_t ip_masked;
 869         uint8_t prev_depth;
 870
 871         for (prev_depth = (uint8_t)(depth - 1); prev_depth > 0; prev_depth--) {
 872                 ip_masked = ip & depth_to_mask(prev_depth);
 873
 874                 rule_index = rule_find(lpm, ip_masked, prev_depth);
 875
 876                 if (rule_index >= 0) {
 877                         *sub_rule_depth = prev_depth;
 878                         return rule_index;
 879                 }
 880         }
 881
 882         return -1;
 883 }
 884
 885 static int32_t
 886 delete_depth_small(struct rte_lpm *lpm, uint32_t ip_masked,
 887         uint8_t depth, int32_t sub_rule_index, uint8_t sub_rule_depth)
 888 {
 889 #define group_idx next_hop
 890         uint32_t tbl24_range, tbl24_index, tbl8_group_index, tbl8_index, i, j;
 891
 892         /* Calculate the range and index into Table24. */
 893         tbl24_range = depth_to_range(depth);
 894         tbl24_index = (ip_masked >> 8);
 895         struct rte_lpm_tbl_entry zero_tbl24_entry = {0};
 896
 897         /*
 898          * Firstly check the sub_rule_index. A -1 indicates no replacement rule
 899          * and a positive number indicates a sub_rule_index.
 900          */
 901         if (sub_rule_index < 0) {
 902                 /*
 903                  * If no replacement rule exists then invalidate entries
 904                  * associated with this rule.
 905                  */
 906                 for (i = tbl24_index; i < (tbl24_index + tbl24_range); i++) {
 907
 908                         if (lpm->tbl24[i].valid_group == 0 &&
 909                                         lpm->tbl24[i].depth <= depth) {
 910                                 __atomic_store(&lpm->tbl24[i],
 911                                         &zero_tbl24_entry, __ATOMIC_RELEASE);
 912                         } else if (lpm->tbl24[i].valid_group == 1) {
 913                                 /*
 914                                  * If TBL24 entry is extended, then there has
 915                                  * to be a rule with depth >= 25 in the
 916                                  * associated TBL8 group.
 917                                  */
 918
 919                                 tbl8_group_index = lpm->tbl24[i].group_idx;
 920                                 tbl8_index = tbl8_group_index *
 921                                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 922
 923                                 for (j = tbl8_index; j < (tbl8_index +
 924                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES); j++) {
 925
 926                                         if (lpm->tbl8[j].depth <= depth)
 927                                                 lpm->tbl8[j].valid = INVALID;
 928                                 }
 929                         }
 930                 }
 931         } else {
 932                 /*
 933                  * If a replacement rule exists then modify entries
 934                  * associated with this rule.
 935                  */
 936
 937                 struct rte_lpm_tbl_entry new_tbl24_entry = {
 938                         .next_hop = lpm->rules_tbl[sub_rule_index].next_hop,
 939                         .valid = VALID,
 940                         .valid_group = 0,
 941                         .depth = sub_rule_depth,
 942                 };
 943
 944                 struct rte_lpm_tbl_entry new_tbl8_entry = {
 945                         .valid = VALID,
 946                         .valid_group = VALID,
 947                         .depth = sub_rule_depth,
 948                         .next_hop = lpm->rules_tbl
 949                         [sub_rule_index].next_hop,
 950                 };
 951
 952                 for (i = tbl24_index; i < (tbl24_index + tbl24_range); i++) {
 953
 954                         if (lpm->tbl24[i].valid_group == 0 &&
 955                                         lpm->tbl24[i].depth <= depth) {
 956                                 __atomic_store(&lpm->tbl24[i], &new_tbl24_entry,
 957                                                 __ATOMIC_RELEASE);
 958                         } else  if (lpm->tbl24[i].valid_group == 1) {
 959                                 /*
 960                                  * If TBL24 entry is extended, then there has
 961                                  * to be a rule with depth >= 25 in the
 962                                  * associated TBL8 group.
 963                                  */
 964
 965                                 tbl8_group_index = lpm->tbl24[i].group_idx;
 966                                 tbl8_index = tbl8_group_index *
 967                                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 968
 969                                 for (j = tbl8_index; j < (tbl8_index +
 970                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES); j++) {
 971
 972                                         if (lpm->tbl8[j].depth <= depth)
 973                                                 __atomic_store(&lpm->tbl8[j],
 974                                                         &new_tbl8_entry,
 975                                                         __ATOMIC_RELAXED);
 976                                 }
 977                         }
 978                 }
 979         }
 980 #undef group_idx
 981         return 0;
 982 }
 983
 984 /*
 985  * Checks if table 8 group can be recycled.
 986  *
 987  * Return of -EEXIST means tbl8 is in use and thus can not be recycled.
 988  * Return of -EINVAL means tbl8 is empty and thus can be recycled
 989  * Return of value > -1 means tbl8 is in use but has all the same values and
 990  * thus can be recycled
 991  */
 992 static int32_t
 993 tbl8_recycle_check(struct rte_lpm_tbl_entry *tbl8,
 994                 uint32_t tbl8_group_start)
 995 {
 996         uint32_t tbl8_group_end, i;
 997         tbl8_group_end = tbl8_group_start + RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 998
 999         /*
1000          * Check the first entry of the given tbl8. If it is invalid we know
1001          * this tbl8 does not contain any rule with a depth < RTE_LPM_MAX_DEPTH
1002          *  (As they would affect all entries in a tbl8) and thus this table
1003          *  can not be recycled.
1004          */
1005         if (tbl8[tbl8_group_start].valid) {
1006                 /*
1007                  * If first entry is valid check if the depth is less than 24
1008                  * and if so check the rest of the entries to verify that they
1009                  * are all of this depth.
1010                  */
1011                 if (tbl8[tbl8_group_start].depth <= MAX_DEPTH_TBL24) {
1012                         for (i = (tbl8_group_start + 1); i < tbl8_group_end;
1013                                         i++) {
1014
1015                                 if (tbl8[i].depth !=
1016                                                 tbl8[tbl8_group_start].depth) {
1017
1018                                         return -EEXIST;
1019                                 }
1020                         }
1021                         /* If all entries are the same return the tb8 index */
1022                         return tbl8_group_start;
1023                 }
1024
1025                 return -EEXIST;
1026         }
1027         /*
1028          * If the first entry is invalid check if the rest of the entries in
1029          * the tbl8 are invalid.
1030          */
1031         for (i = (tbl8_group_start + 1); i < tbl8_group_end; i++) {
1032                 if (tbl8[i].valid)
1033                         return -EEXIST;
1034         }
1035         /* If no valid entries are found then return -EINVAL. */
1036         return -EINVAL;
1037 }
1038
1039 static int32_t
1040 delete_depth_big(struct rte_lpm *lpm, uint32_t ip_masked,
1041         uint8_t depth, int32_t sub_rule_index, uint8_t sub_rule_depth)
1042 {
1043 #define group_idx next_hop
1044         uint32_t tbl24_index, tbl8_group_index, tbl8_group_start, tbl8_index,
1045                         tbl8_range, i;
1046         int32_t tbl8_recycle_index;
1047
1048         /*
1049          * Calculate the index into tbl24 and range. Note: All depths larger
1050          * than MAX_DEPTH_TBL24 are associated with only one tbl24 entry.
1051          */
1052         tbl24_index = ip_masked >> 8;
1053
1054         /* Calculate the index into tbl8 and range. */
1055         tbl8_group_index = lpm->tbl24[tbl24_index].group_idx;
1056         tbl8_group_start = tbl8_group_index * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
1057         tbl8_index = tbl8_group_start + (ip_masked & 0xFF);
1058         tbl8_range = depth_to_range(depth);
1059
1060         if (sub_rule_index < 0) {
1061                 /*
1062                  * Loop through the range of entries on tbl8 for which the
1063                  * rule_to_delete must be removed or modified.
1064                  */
1065                 for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
1066                         if (lpm->tbl8[i].depth <= depth)
1067                                 lpm->tbl8[i].valid = INVALID;
1068                 }
1069         } else {
1070                 /* Set new tbl8 entry. */
1071                 struct rte_lpm_tbl_entry new_tbl8_entry = {
1072                         .valid = VALID,
1073                         .depth = sub_rule_depth,
1074                         .valid_group = lpm->tbl8[tbl8_group_start].valid_group,
1075                         .next_hop = lpm->rules_tbl[sub_rule_index].next_hop,
1076                 };
1077
1078                 /*
1079                  * Loop through the range of entries on tbl8 for which the
1080                  * rule_to_delete must be modified.
1081                  */
1082                 for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
1083                         if (lpm->tbl8[i].depth <= depth)
1084                                 __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
1085                                                 __ATOMIC_RELAXED);
1086                 }
1087         }
1088
1089         /*
1090          * Check if there are any valid entries in this tbl8 group. If all
1091          * tbl8 entries are invalid we can free the tbl8 and invalidate the
1092          * associated tbl24 entry.
1093          */
1094
1095         tbl8_recycle_index = tbl8_recycle_check(lpm->tbl8, tbl8_group_start);
1096
1097         if (tbl8_recycle_index == -EINVAL) {
1098                 /* Set tbl24 before freeing tbl8 to avoid race condition.
1099                  * Prevent the free of the tbl8 group from hoisting.
1100                  */
1101                 lpm->tbl24[tbl24_index].valid = 0;
1102                 __atomic_thread_fence(__ATOMIC_RELEASE);
1103                 tbl8_free(lpm, tbl8_group_start);
1104         } else if (tbl8_recycle_index > -1) {
1105                 /* Update tbl24 entry. */
1106                 struct rte_lpm_tbl_entry new_tbl24_entry = {
1107                         .next_hop = lpm->tbl8[tbl8_recycle_index].next_hop,
1108                         .valid = VALID,
1109                         .valid_group = 0,
1110                         .depth = lpm->tbl8[tbl8_recycle_index].depth,
1111                 };
1112
1113                 /* Set tbl24 before freeing tbl8 to avoid race condition.
1114                  * Prevent the free of the tbl8 group from hoisting.
1115                  */
1116                 __atomic_store(&lpm->tbl24[tbl24_index], &new_tbl24_entry,
1117                                 __ATOMIC_RELAXED);
1118                 __atomic_thread_fence(__ATOMIC_RELEASE);
1119                 tbl8_free(lpm, tbl8_group_start);
1120         }
1121 #undef group_idx
1122         return 0;
1123 }
1124
1125 /*
1126  * Deletes a rule
1127  */
1128 int
1129 rte_lpm_delete(struct rte_lpm *lpm, uint32_t ip, uint8_t depth)
1130 {
1131         int32_t rule_to_delete_index, sub_rule_index;
1132         uint32_t ip_masked;
1133         uint8_t sub_rule_depth;
1134         /*
1135          * Check input arguments. Note: IP must be a positive integer of 32
1136          * bits in length therefore it need not be checked.
1137          */
1138         if ((lpm == NULL) || (depth < 1) || (depth > RTE_LPM_MAX_DEPTH)) {
1139                 return -EINVAL;
1140         }
1141
1142         ip_masked = ip & depth_to_mask(depth);
1143
1144         /*
1145          * Find the index of the input rule, that needs to be deleted, in the
1146          * rule table.
1147          */
1148         rule_to_delete_index = rule_find(lpm, ip_masked, depth);
1149
1150         /*
1151          * Check if rule_to_delete_index was found. If no rule was found the
1152          * function rule_find returns -EINVAL.
1153          */
1154         if (rule_to_delete_index < 0)
1155                 return -EINVAL;
1156
1157         /* Delete the rule from the rule table. */
1158         rule_delete(lpm, rule_to_delete_index, depth);
1159
1160         /*
1161          * Find rule to replace the rule_to_delete. If there is no rule to
1162          * replace the rule_to_delete we return -1 and invalidate the table
1163          * entries associated with this rule.
1164          */
1165         sub_rule_depth = 0;
1166         sub_rule_index = find_previous_rule(lpm, ip, depth, &sub_rule_depth);
1167
1168         /*
1169          * If the input depth value is less than 25 use function
1170          * delete_depth_small otherwise use delete_depth_big.
1171          */
1172         if (depth <= MAX_DEPTH_TBL24) {
1173                 return delete_depth_small(lpm, ip_masked, depth,
1174                                 sub_rule_index, sub_rule_depth);
1175         } else { /* If depth > MAX_DEPTH_TBL24 */
1176                 return delete_depth_big(lpm, ip_masked, depth, sub_rule_index,
1177                                 sub_rule_depth);
1178         }
1179 }
1180
1181 /*
1182  * Delete all rules from the LPM table.
1183  */
1184 void
1185 rte_lpm_delete_all(struct rte_lpm *lpm)
1186 {
1187         /* Zero rule information. */
1188         memset(lpm->rule_info, 0, sizeof(lpm->rule_info));
1189
1190         /* Zero tbl24. */
1191         memset(lpm->tbl24, 0, sizeof(lpm->tbl24));
1192
1193         /* Zero tbl8. */
1194         memset(lpm->tbl8, 0, sizeof(lpm->tbl8[0])
1195                         * RTE_LPM_TBL8_GROUP_NUM_ENTRIES * lpm->number_tbl8s);
1196
1197         /* Delete all rules form the rules table. */
1198         memset(lpm->rules_tbl, 0, sizeof(lpm->rules_tbl[0]) * lpm->max_rules);
1199 }