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 {
 293         struct rte_rcu_qsbr_dq_parameters params = {0};
 294         char rcu_dq_name[RTE_RCU_QSBR_DQ_NAMESIZE];
 295         struct __rte_lpm *internal_lpm;
 296
 297         if (lpm == NULL || cfg == NULL) {
 298                 rte_errno = EINVAL;
 299                 return 1;
 300         }
 301
 302         internal_lpm = container_of(lpm, struct __rte_lpm, lpm);
 303         if (internal_lpm->v != NULL) {
 304                 rte_errno = EEXIST;
 305                 return 1;
 306         }
 307
 308         if (cfg->mode == RTE_LPM_QSBR_MODE_SYNC) {
 309                 /* No other things to do. */
 310         } else if (cfg->mode == RTE_LPM_QSBR_MODE_DQ) {
 311                 /* Init QSBR defer queue. */
 312                 snprintf(rcu_dq_name, sizeof(rcu_dq_name),
 313                                 "LPM_RCU_%s", lpm->name);
 314                 params.name = rcu_dq_name;
 315                 params.size = cfg->dq_size;
 316                 if (params.size == 0)
 317                         params.size = lpm->number_tbl8s;
 318                 params.trigger_reclaim_limit = cfg->reclaim_thd;
 319                 params.max_reclaim_size = cfg->reclaim_max;
 320                 if (params.max_reclaim_size == 0)
 321                         params.max_reclaim_size = RTE_LPM_RCU_DQ_RECLAIM_MAX;
 322                 params.esize = sizeof(uint32_t);        /* tbl8 group index */
 323                 params.free_fn = __lpm_rcu_qsbr_free_resource;
 324                 params.p = lpm;
 325                 params.v = cfg->v;
 326                 internal_lpm->dq = rte_rcu_qsbr_dq_create(&params);
 327                 if (internal_lpm->dq == NULL) {
 328                         RTE_LOG(ERR, LPM, "LPM defer queue creation failed\n");
 329                         return 1;
 330                 }
 331         } else {
 332                 rte_errno = EINVAL;
 333                 return 1;
 334         }
 335         internal_lpm->rcu_mode = cfg->mode;
 336         internal_lpm->v = cfg->v;
 337
 338         return 0;
 339 }
 340
 341 /*
 342  * Adds a rule to the rule table.
 343  *
 344  * NOTE: The rule table is split into 32 groups. Each group contains rules that
 345  * apply to a specific prefix depth (i.e. group 1 contains rules that apply to
 346  * prefixes with a depth of 1 etc.). In the following code (depth - 1) is used
 347  * to refer to depth 1 because even though the depth range is 1 - 32, depths
 348  * are stored in the rule table from 0 - 31.
 349  * NOTE: Valid range for depth parameter is 1 .. 32 inclusive.
 350  */
 351 static int32_t
 352 rule_add(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth,
 353         uint32_t next_hop)
 354 {
 355         uint32_t rule_gindex, rule_index, last_rule;
 356         int i;
 357
 358         VERIFY_DEPTH(depth);
 359
 360         /* Scan through rule group to see if rule already exists. */
 361         if (lpm->rule_info[depth - 1].used_rules > 0) {
 362
 363                 /* rule_gindex stands for rule group index. */
 364                 rule_gindex = lpm->rule_info[depth - 1].first_rule;
 365                 /* Initialise rule_index to point to start of rule group. */
 366                 rule_index = rule_gindex;
 367                 /* Last rule = Last used rule in this rule group. */
 368                 last_rule = rule_gindex + lpm->rule_info[depth - 1].used_rules;
 369
 370                 for (; rule_index < last_rule; rule_index++) {
 371
 372                         /* If rule already exists update next hop and return. */
 373                         if (lpm->rules_tbl[rule_index].ip == ip_masked) {
 374
 375                                 if (lpm->rules_tbl[rule_index].next_hop
 376                                                 == next_hop)
 377                                         return -EEXIST;
 378                                 lpm->rules_tbl[rule_index].next_hop = next_hop;
 379
 380                                 return rule_index;
 381                         }
 382                 }
 383
 384                 if (rule_index == lpm->max_rules)
 385                         return -ENOSPC;
 386         } else {
 387                 /* Calculate the position in which the rule will be stored. */
 388                 rule_index = 0;
 389
 390                 for (i = depth - 1; i > 0; i--) {
 391                         if (lpm->rule_info[i - 1].used_rules > 0) {
 392                                 rule_index = lpm->rule_info[i - 1].first_rule
 393                                                 + lpm->rule_info[i - 1].used_rules;
 394                                 break;
 395                         }
 396                 }
 397                 if (rule_index == lpm->max_rules)
 398                         return -ENOSPC;
 399
 400                 lpm->rule_info[depth - 1].first_rule = rule_index;
 401         }
 402
 403         /* Make room for the new rule in the array. */
 404         for (i = RTE_LPM_MAX_DEPTH; i > depth; i--) {
 405                 if (lpm->rule_info[i - 1].first_rule
 406                                 + lpm->rule_info[i - 1].used_rules == lpm->max_rules)
 407                         return -ENOSPC;
 408
 409                 if (lpm->rule_info[i - 1].used_rules > 0) {
 410                         lpm->rules_tbl[lpm->rule_info[i - 1].first_rule
 411                                 + lpm->rule_info[i - 1].used_rules]
 412                                         = lpm->rules_tbl[lpm->rule_info[i - 1].first_rule];
 413                         lpm->rule_info[i - 1].first_rule++;
 414                 }
 415         }
 416
 417         /* Add the new rule. */
 418         lpm->rules_tbl[rule_index].ip = ip_masked;
 419         lpm->rules_tbl[rule_index].next_hop = next_hop;
 420
 421         /* Increment the used rules counter for this rule group. */
 422         lpm->rule_info[depth - 1].used_rules++;
 423
 424         return rule_index;
 425 }
 426
 427 /*
 428  * Delete a rule from the rule table.
 429  * NOTE: Valid range for depth parameter is 1 .. 32 inclusive.
 430  */
 431 static void
 432 rule_delete(struct rte_lpm *lpm, int32_t rule_index, uint8_t depth)
 433 {
 434         int i;
 435
 436         VERIFY_DEPTH(depth);
 437
 438         lpm->rules_tbl[rule_index] =
 439                         lpm->rules_tbl[lpm->rule_info[depth - 1].first_rule
 440                         + lpm->rule_info[depth - 1].used_rules - 1];
 441
 442         for (i = depth; i < RTE_LPM_MAX_DEPTH; i++) {
 443                 if (lpm->rule_info[i].used_rules > 0) {
 444                         lpm->rules_tbl[lpm->rule_info[i].first_rule - 1] =
 445                                         lpm->rules_tbl[lpm->rule_info[i].first_rule
 446                                                 + lpm->rule_info[i].used_rules - 1];
 447                         lpm->rule_info[i].first_rule--;
 448                 }
 449         }
 450
 451         lpm->rule_info[depth - 1].used_rules--;
 452 }
 453
 454 /*
 455  * Finds a rule in rule table.
 456  * NOTE: Valid range for depth parameter is 1 .. 32 inclusive.
 457  */
 458 static int32_t
 459 rule_find(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth)
 460 {
 461         uint32_t rule_gindex, last_rule, rule_index;
 462
 463         VERIFY_DEPTH(depth);
 464
 465         rule_gindex = lpm->rule_info[depth - 1].first_rule;
 466         last_rule = rule_gindex + lpm->rule_info[depth - 1].used_rules;
 467
 468         /* Scan used rules at given depth to find rule. */
 469         for (rule_index = rule_gindex; rule_index < last_rule; rule_index++) {
 470                 /* If rule is found return the rule index. */
 471                 if (lpm->rules_tbl[rule_index].ip == ip_masked)
 472                         return rule_index;
 473         }
 474
 475         /* If rule is not found return -EINVAL. */
 476         return -EINVAL;
 477 }
 478
 479 /*
 480  * Find, clean and allocate a tbl8.
 481  */
 482 static int32_t
 483 _tbl8_alloc(struct rte_lpm *lpm)
 484 {
 485         uint32_t group_idx; /* tbl8 group index. */
 486         struct rte_lpm_tbl_entry *tbl8_entry;
 487
 488         /* Scan through tbl8 to find a free (i.e. INVALID) tbl8 group. */
 489         for (group_idx = 0; group_idx < lpm->number_tbl8s; group_idx++) {
 490                 tbl8_entry = &lpm->tbl8[group_idx *
 491                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES];
 492                 /* If a free tbl8 group is found clean it and set as VALID. */
 493                 if (!tbl8_entry->valid_group) {
 494                         struct rte_lpm_tbl_entry new_tbl8_entry = {
 495                                 .next_hop = 0,
 496                                 .valid = INVALID,
 497                                 .depth = 0,
 498                                 .valid_group = VALID,
 499                         };
 500
 501                         memset(&tbl8_entry[0], 0,
 502                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES *
 503                                         sizeof(tbl8_entry[0]));
 504
 505                         __atomic_store(tbl8_entry, &new_tbl8_entry,
 506                                         __ATOMIC_RELAXED);
 507
 508                         /* Return group index for allocated tbl8 group. */
 509                         return group_idx;
 510                 }
 511         }
 512
 513         /* If there are no tbl8 groups free then return error. */
 514         return -ENOSPC;
 515 }
 516
 517 static int32_t
 518 tbl8_alloc(struct rte_lpm *lpm)
 519 {
 520         int32_t group_idx; /* tbl8 group index. */
 521         struct __rte_lpm *internal_lpm;
 522
 523         internal_lpm = container_of(lpm, struct __rte_lpm, lpm);
 524         group_idx = _tbl8_alloc(lpm);
 525         if (group_idx == -ENOSPC && internal_lpm->dq != NULL) {
 526                 /* If there are no tbl8 groups try to reclaim one. */
 527                 if (rte_rcu_qsbr_dq_reclaim(internal_lpm->dq, 1,
 528                                 NULL, NULL, NULL) == 0)
 529                         group_idx = _tbl8_alloc(lpm);
 530         }
 531
 532         return group_idx;
 533 }
 534
 535 static void
 536 tbl8_free(struct rte_lpm *lpm, uint32_t tbl8_group_start)
 537 {
 538         struct rte_lpm_tbl_entry zero_tbl8_entry = {0};
 539         struct __rte_lpm *internal_lpm;
 540
 541         internal_lpm = container_of(lpm, struct __rte_lpm, lpm);
 542         if (internal_lpm->v == NULL) {
 543                 /* Set tbl8 group invalid*/
 544                 __atomic_store(&lpm->tbl8[tbl8_group_start], &zero_tbl8_entry,
 545                                 __ATOMIC_RELAXED);
 546         } else if (internal_lpm->rcu_mode == RTE_LPM_QSBR_MODE_SYNC) {
 547                 /* Wait for quiescent state change. */
 548                 rte_rcu_qsbr_synchronize(internal_lpm->v,
 549                         RTE_QSBR_THRID_INVALID);
 550                 /* Set tbl8 group invalid*/
 551                 __atomic_store(&lpm->tbl8[tbl8_group_start], &zero_tbl8_entry,
 552                                 __ATOMIC_RELAXED);
 553         } else if (internal_lpm->rcu_mode == RTE_LPM_QSBR_MODE_DQ) {
 554                 /* Push into QSBR defer queue. */
 555                 rte_rcu_qsbr_dq_enqueue(internal_lpm->dq,
 556                                 (void *)&tbl8_group_start);
 557         }
 558 }
 559
 560 static __rte_noinline int32_t
 561 add_depth_small(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
 562                 uint32_t next_hop)
 563 {
 564 #define group_idx next_hop
 565         uint32_t tbl24_index, tbl24_range, tbl8_index, tbl8_group_end, i, j;
 566
 567         /* Calculate the index into Table24. */
 568         tbl24_index = ip >> 8;
 569         tbl24_range = depth_to_range(depth);
 570
 571         for (i = tbl24_index; i < (tbl24_index + tbl24_range); i++) {
 572                 /*
 573                  * For invalid OR valid and non-extended tbl 24 entries set
 574                  * entry.
 575                  */
 576                 if (!lpm->tbl24[i].valid || (lpm->tbl24[i].valid_group == 0 &&
 577                                 lpm->tbl24[i].depth <= depth)) {
 578
 579                         struct rte_lpm_tbl_entry new_tbl24_entry = {
 580                                 .next_hop = next_hop,
 581                                 .valid = VALID,
 582                                 .valid_group = 0,
 583                                 .depth = depth,
 584                         };
 585
 586                         /* Setting tbl24 entry in one go to avoid race
 587                          * conditions
 588                          */
 589                         __atomic_store(&lpm->tbl24[i], &new_tbl24_entry,
 590                                         __ATOMIC_RELEASE);
 591
 592                         continue;
 593                 }
 594
 595                 if (lpm->tbl24[i].valid_group == 1) {
 596                         /* If tbl24 entry is valid and extended calculate the
 597                          *  index into tbl8.
 598                          */
 599                         tbl8_index = lpm->tbl24[i].group_idx *
 600                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 601                         tbl8_group_end = tbl8_index +
 602                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 603
 604                         for (j = tbl8_index; j < tbl8_group_end; j++) {
 605                                 if (!lpm->tbl8[j].valid ||
 606                                                 lpm->tbl8[j].depth <= depth) {
 607                                         struct rte_lpm_tbl_entry
 608                                                 new_tbl8_entry = {
 609                                                 .valid = VALID,
 610                                                 .valid_group = VALID,
 611                                                 .depth = depth,
 612                                                 .next_hop = next_hop,
 613                                         };
 614
 615                                         /*
 616                                          * Setting tbl8 entry in one go to avoid
 617                                          * race conditions
 618                                          */
 619                                         __atomic_store(&lpm->tbl8[j],
 620                                                 &new_tbl8_entry,
 621                                                 __ATOMIC_RELAXED);
 622
 623                                         continue;
 624                                 }
 625                         }
 626                 }
 627         }
 628 #undef group_idx
 629         return 0;
 630 }
 631
 632 static __rte_noinline int32_t
 633 add_depth_big(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth,
 634                 uint32_t next_hop)
 635 {
 636 #define group_idx next_hop
 637         uint32_t tbl24_index;
 638         int32_t tbl8_group_index, tbl8_group_start, tbl8_group_end, tbl8_index,
 639                 tbl8_range, i;
 640
 641         tbl24_index = (ip_masked >> 8);
 642         tbl8_range = depth_to_range(depth);
 643
 644         if (!lpm->tbl24[tbl24_index].valid) {
 645                 /* Search for a free tbl8 group. */
 646                 tbl8_group_index = tbl8_alloc(lpm);
 647
 648                 /* Check tbl8 allocation was successful. */
 649                 if (tbl8_group_index < 0) {
 650                         return tbl8_group_index;
 651                 }
 652
 653                 /* Find index into tbl8 and range. */
 654                 tbl8_index = (tbl8_group_index *
 655                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES) +
 656                                 (ip_masked & 0xFF);
 657
 658                 /* Set tbl8 entry. */
 659                 for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
 660                         struct rte_lpm_tbl_entry new_tbl8_entry = {
 661                                 .valid = VALID,
 662                                 .depth = depth,
 663                                 .valid_group = lpm->tbl8[i].valid_group,
 664                                 .next_hop = next_hop,
 665                         };
 666                         __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
 667                                         __ATOMIC_RELAXED);
 668                 }
 669
 670                 /*
 671                  * Update tbl24 entry to point to new tbl8 entry. Note: The
 672                  * ext_flag and tbl8_index need to be updated simultaneously,
 673                  * so assign whole structure in one go
 674                  */
 675
 676                 struct rte_lpm_tbl_entry new_tbl24_entry = {
 677                         .group_idx = tbl8_group_index,
 678                         .valid = VALID,
 679                         .valid_group = 1,
 680                         .depth = 0,
 681                 };
 682
 683                 /* The tbl24 entry must be written only after the
 684                  * tbl8 entries are written.
 685                  */
 686                 __atomic_store(&lpm->tbl24[tbl24_index], &new_tbl24_entry,
 687                                 __ATOMIC_RELEASE);
 688
 689         } /* If valid entry but not extended calculate the index into Table8. */
 690         else if (lpm->tbl24[tbl24_index].valid_group == 0) {
 691                 /* Search for free tbl8 group. */
 692                 tbl8_group_index = tbl8_alloc(lpm);
 693
 694                 if (tbl8_group_index < 0) {
 695                         return tbl8_group_index;
 696                 }
 697
 698                 tbl8_group_start = tbl8_group_index *
 699                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 700                 tbl8_group_end = tbl8_group_start +
 701                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 702
 703                 /* Populate new tbl8 with tbl24 value. */
 704                 for (i = tbl8_group_start; i < tbl8_group_end; i++) {
 705                         struct rte_lpm_tbl_entry new_tbl8_entry = {
 706                                 .valid = VALID,
 707                                 .depth = lpm->tbl24[tbl24_index].depth,
 708                                 .valid_group = lpm->tbl8[i].valid_group,
 709                                 .next_hop = lpm->tbl24[tbl24_index].next_hop,
 710                         };
 711                         __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
 712                                         __ATOMIC_RELAXED);
 713                 }
 714
 715                 tbl8_index = tbl8_group_start + (ip_masked & 0xFF);
 716
 717                 /* Insert new rule into the tbl8 entry. */
 718                 for (i = tbl8_index; i < tbl8_index + tbl8_range; i++) {
 719                         struct rte_lpm_tbl_entry new_tbl8_entry = {
 720                                 .valid = VALID,
 721                                 .depth = depth,
 722                                 .valid_group = lpm->tbl8[i].valid_group,
 723                                 .next_hop = next_hop,
 724                         };
 725                         __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
 726                                         __ATOMIC_RELAXED);
 727                 }
 728
 729                 /*
 730                  * Update tbl24 entry to point to new tbl8 entry. Note: The
 731                  * ext_flag and tbl8_index need to be updated simultaneously,
 732                  * so assign whole structure in one go.
 733                  */
 734
 735                 struct rte_lpm_tbl_entry new_tbl24_entry = {
 736                                 .group_idx = tbl8_group_index,
 737                                 .valid = VALID,
 738                                 .valid_group = 1,
 739                                 .depth = 0,
 740                 };
 741
 742                 /* The tbl24 entry must be written only after the
 743                  * tbl8 entries are written.
 744                  */
 745                 __atomic_store(&lpm->tbl24[tbl24_index], &new_tbl24_entry,
 746                                 __ATOMIC_RELEASE);
 747
 748         } else { /*
 749                 * If it is valid, extended entry calculate the index into tbl8.
 750                 */
 751                 tbl8_group_index = lpm->tbl24[tbl24_index].group_idx;
 752                 tbl8_group_start = tbl8_group_index *
 753                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 754                 tbl8_index = tbl8_group_start + (ip_masked & 0xFF);
 755
 756                 for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
 757
 758                         if (!lpm->tbl8[i].valid ||
 759                                         lpm->tbl8[i].depth <= depth) {
 760                                 struct rte_lpm_tbl_entry new_tbl8_entry = {
 761                                         .valid = VALID,
 762                                         .depth = depth,
 763                                         .next_hop = next_hop,
 764                                         .valid_group = lpm->tbl8[i].valid_group,
 765                                 };
 766
 767                                 /*
 768                                  * Setting tbl8 entry in one go to avoid race
 769                                  * condition
 770                                  */
 771                                 __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
 772                                                 __ATOMIC_RELAXED);
 773
 774                                 continue;
 775                         }
 776                 }
 777         }
 778 #undef group_idx
 779         return 0;
 780 }
 781
 782 /*
 783  * Add a route
 784  */
 785 int
 786 rte_lpm_add(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
 787                 uint32_t next_hop)
 788 {
 789         int32_t rule_index, status = 0;
 790         uint32_t ip_masked;
 791
 792         /* Check user arguments. */
 793         if ((lpm == NULL) || (depth < 1) || (depth > RTE_LPM_MAX_DEPTH))
 794                 return -EINVAL;
 795
 796         ip_masked = ip & depth_to_mask(depth);
 797
 798         /* Add the rule to the rule table. */
 799         rule_index = rule_add(lpm, ip_masked, depth, next_hop);
 800
 801         /* Skip table entries update if The rule is the same as
 802          * the rule in the rules table.
 803          */
 804         if (rule_index == -EEXIST)
 805                 return 0;
 806
 807         /* If the is no space available for new rule return error. */
 808         if (rule_index < 0) {
 809                 return rule_index;
 810         }
 811
 812         if (depth <= MAX_DEPTH_TBL24) {
 813                 status = add_depth_small(lpm, ip_masked, depth, next_hop);
 814         } else { /* If depth > RTE_LPM_MAX_DEPTH_TBL24 */
 815                 status = add_depth_big(lpm, ip_masked, depth, next_hop);
 816
 817                 /*
 818                  * If add fails due to exhaustion of tbl8 extensions delete
 819                  * rule that was added to rule table.
 820                  */
 821                 if (status < 0) {
 822                         rule_delete(lpm, rule_index, depth);
 823
 824                         return status;
 825                 }
 826         }
 827
 828         return 0;
 829 }
 830
 831 /*
 832  * Look for a rule in the high-level rules table
 833  */
 834 int
 835 rte_lpm_is_rule_present(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
 836 uint32_t *next_hop)
 837 {
 838         uint32_t ip_masked;
 839         int32_t rule_index;
 840
 841         /* Check user arguments. */
 842         if ((lpm == NULL) ||
 843                 (next_hop == NULL) ||
 844                 (depth < 1) || (depth > RTE_LPM_MAX_DEPTH))
 845                 return -EINVAL;
 846
 847         /* Look for the rule using rule_find. */
 848         ip_masked = ip & depth_to_mask(depth);
 849         rule_index = rule_find(lpm, ip_masked, depth);
 850
 851         if (rule_index >= 0) {
 852                 *next_hop = lpm->rules_tbl[rule_index].next_hop;
 853                 return 1;
 854         }
 855
 856         /* If rule is not found return 0. */
 857         return 0;
 858 }
 859
 860 static int32_t
 861 find_previous_rule(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
 862                 uint8_t *sub_rule_depth)
 863 {
 864         int32_t rule_index;
 865         uint32_t ip_masked;
 866         uint8_t prev_depth;
 867
 868         for (prev_depth = (uint8_t)(depth - 1); prev_depth > 0; prev_depth--) {
 869                 ip_masked = ip & depth_to_mask(prev_depth);
 870
 871                 rule_index = rule_find(lpm, ip_masked, prev_depth);
 872
 873                 if (rule_index >= 0) {
 874                         *sub_rule_depth = prev_depth;
 875                         return rule_index;
 876                 }
 877         }
 878
 879         return -1;
 880 }
 881
 882 static int32_t
 883 delete_depth_small(struct rte_lpm *lpm, uint32_t ip_masked,
 884         uint8_t depth, int32_t sub_rule_index, uint8_t sub_rule_depth)
 885 {
 886 #define group_idx next_hop
 887         uint32_t tbl24_range, tbl24_index, tbl8_group_index, tbl8_index, i, j;
 888
 889         /* Calculate the range and index into Table24. */
 890         tbl24_range = depth_to_range(depth);
 891         tbl24_index = (ip_masked >> 8);
 892         struct rte_lpm_tbl_entry zero_tbl24_entry = {0};
 893
 894         /*
 895          * Firstly check the sub_rule_index. A -1 indicates no replacement rule
 896          * and a positive number indicates a sub_rule_index.
 897          */
 898         if (sub_rule_index < 0) {
 899                 /*
 900                  * If no replacement rule exists then invalidate entries
 901                  * associated with this rule.
 902                  */
 903                 for (i = tbl24_index; i < (tbl24_index + tbl24_range); i++) {
 904
 905                         if (lpm->tbl24[i].valid_group == 0 &&
 906                                         lpm->tbl24[i].depth <= depth) {
 907                                 __atomic_store(&lpm->tbl24[i],
 908                                         &zero_tbl24_entry, __ATOMIC_RELEASE);
 909                         } else if (lpm->tbl24[i].valid_group == 1) {
 910                                 /*
 911                                  * If TBL24 entry is extended, then there has
 912                                  * to be a rule with depth >= 25 in the
 913                                  * associated TBL8 group.
 914                                  */
 915
 916                                 tbl8_group_index = lpm->tbl24[i].group_idx;
 917                                 tbl8_index = tbl8_group_index *
 918                                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 919
 920                                 for (j = tbl8_index; j < (tbl8_index +
 921                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES); j++) {
 922
 923                                         if (lpm->tbl8[j].depth <= depth)
 924                                                 lpm->tbl8[j].valid = INVALID;
 925                                 }
 926                         }
 927                 }
 928         } else {
 929                 /*
 930                  * If a replacement rule exists then modify entries
 931                  * associated with this rule.
 932                  */
 933
 934                 struct rte_lpm_tbl_entry new_tbl24_entry = {
 935                         .next_hop = lpm->rules_tbl[sub_rule_index].next_hop,
 936                         .valid = VALID,
 937                         .valid_group = 0,
 938                         .depth = sub_rule_depth,
 939                 };
 940
 941                 struct rte_lpm_tbl_entry new_tbl8_entry = {
 942                         .valid = VALID,
 943                         .valid_group = VALID,
 944                         .depth = sub_rule_depth,
 945                         .next_hop = lpm->rules_tbl
 946                         [sub_rule_index].next_hop,
 947                 };
 948
 949                 for (i = tbl24_index; i < (tbl24_index + tbl24_range); i++) {
 950
 951                         if (lpm->tbl24[i].valid_group == 0 &&
 952                                         lpm->tbl24[i].depth <= depth) {
 953                                 __atomic_store(&lpm->tbl24[i], &new_tbl24_entry,
 954                                                 __ATOMIC_RELEASE);
 955                         } else  if (lpm->tbl24[i].valid_group == 1) {
 956                                 /*
 957                                  * If TBL24 entry is extended, then there has
 958                                  * to be a rule with depth >= 25 in the
 959                                  * associated TBL8 group.
 960                                  */
 961
 962                                 tbl8_group_index = lpm->tbl24[i].group_idx;
 963                                 tbl8_index = tbl8_group_index *
 964                                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 965
 966                                 for (j = tbl8_index; j < (tbl8_index +
 967                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES); j++) {
 968
 969                                         if (lpm->tbl8[j].depth <= depth)
 970                                                 __atomic_store(&lpm->tbl8[j],
 971                                                         &new_tbl8_entry,
 972                                                         __ATOMIC_RELAXED);
 973                                 }
 974                         }
 975                 }
 976         }
 977 #undef group_idx
 978         return 0;
 979 }
 980
 981 /*
 982  * Checks if table 8 group can be recycled.
 983  *
 984  * Return of -EEXIST means tbl8 is in use and thus can not be recycled.
 985  * Return of -EINVAL means tbl8 is empty and thus can be recycled
 986  * Return of value > -1 means tbl8 is in use but has all the same values and
 987  * thus can be recycled
 988  */
 989 static int32_t
 990 tbl8_recycle_check(struct rte_lpm_tbl_entry *tbl8,
 991                 uint32_t tbl8_group_start)
 992 {
 993         uint32_t tbl8_group_end, i;
 994         tbl8_group_end = tbl8_group_start + RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 995
 996         /*
 997          * Check the first entry of the given tbl8. If it is invalid we know
 998          * this tbl8 does not contain any rule with a depth < RTE_LPM_MAX_DEPTH
 999          *  (As they would affect all entries in a tbl8) and thus this table
1000          *  can not be recycled.
1001          */
1002         if (tbl8[tbl8_group_start].valid) {
1003                 /*
1004                  * If first entry is valid check if the depth is less than 24
1005                  * and if so check the rest of the entries to verify that they
1006                  * are all of this depth.
1007                  */
1008                 if (tbl8[tbl8_group_start].depth <= MAX_DEPTH_TBL24) {
1009                         for (i = (tbl8_group_start + 1); i < tbl8_group_end;
1010                                         i++) {
1011
1012                                 if (tbl8[i].depth !=
1013                                                 tbl8[tbl8_group_start].depth) {
1014
1015                                         return -EEXIST;
1016                                 }
1017                         }
1018                         /* If all entries are the same return the tb8 index */
1019                         return tbl8_group_start;
1020                 }
1021
1022                 return -EEXIST;
1023         }
1024         /*
1025          * If the first entry is invalid check if the rest of the entries in
1026          * the tbl8 are invalid.
1027          */
1028         for (i = (tbl8_group_start + 1); i < tbl8_group_end; i++) {
1029                 if (tbl8[i].valid)
1030                         return -EEXIST;
1031         }
1032         /* If no valid entries are found then return -EINVAL. */
1033         return -EINVAL;
1034 }
1035
1036 static int32_t
1037 delete_depth_big(struct rte_lpm *lpm, uint32_t ip_masked,
1038         uint8_t depth, int32_t sub_rule_index, uint8_t sub_rule_depth)
1039 {
1040 #define group_idx next_hop
1041         uint32_t tbl24_index, tbl8_group_index, tbl8_group_start, tbl8_index,
1042                         tbl8_range, i;
1043         int32_t tbl8_recycle_index;
1044
1045         /*
1046          * Calculate the index into tbl24 and range. Note: All depths larger
1047          * than MAX_DEPTH_TBL24 are associated with only one tbl24 entry.
1048          */
1049         tbl24_index = ip_masked >> 8;
1050
1051         /* Calculate the index into tbl8 and range. */
1052         tbl8_group_index = lpm->tbl24[tbl24_index].group_idx;
1053         tbl8_group_start = tbl8_group_index * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
1054         tbl8_index = tbl8_group_start + (ip_masked & 0xFF);
1055         tbl8_range = depth_to_range(depth);
1056
1057         if (sub_rule_index < 0) {
1058                 /*
1059                  * Loop through the range of entries on tbl8 for which the
1060                  * rule_to_delete must be removed or modified.
1061                  */
1062                 for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
1063                         if (lpm->tbl8[i].depth <= depth)
1064                                 lpm->tbl8[i].valid = INVALID;
1065                 }
1066         } else {
1067                 /* Set new tbl8 entry. */
1068                 struct rte_lpm_tbl_entry new_tbl8_entry = {
1069                         .valid = VALID,
1070                         .depth = sub_rule_depth,
1071                         .valid_group = lpm->tbl8[tbl8_group_start].valid_group,
1072                         .next_hop = lpm->rules_tbl[sub_rule_index].next_hop,
1073                 };
1074
1075                 /*
1076                  * Loop through the range of entries on tbl8 for which the
1077                  * rule_to_delete must be modified.
1078                  */
1079                 for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
1080                         if (lpm->tbl8[i].depth <= depth)
1081                                 __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
1082                                                 __ATOMIC_RELAXED);
1083                 }
1084         }
1085
1086         /*
1087          * Check if there are any valid entries in this tbl8 group. If all
1088          * tbl8 entries are invalid we can free the tbl8 and invalidate the
1089          * associated tbl24 entry.
1090          */
1091
1092         tbl8_recycle_index = tbl8_recycle_check(lpm->tbl8, tbl8_group_start);
1093
1094         if (tbl8_recycle_index == -EINVAL) {
1095                 /* Set tbl24 before freeing tbl8 to avoid race condition.
1096                  * Prevent the free of the tbl8 group from hoisting.
1097                  */
1098                 lpm->tbl24[tbl24_index].valid = 0;
1099                 __atomic_thread_fence(__ATOMIC_RELEASE);
1100                 tbl8_free(lpm, tbl8_group_start);
1101         } else if (tbl8_recycle_index > -1) {
1102                 /* Update tbl24 entry. */
1103                 struct rte_lpm_tbl_entry new_tbl24_entry = {
1104                         .next_hop = lpm->tbl8[tbl8_recycle_index].next_hop,
1105                         .valid = VALID,
1106                         .valid_group = 0,
1107                         .depth = lpm->tbl8[tbl8_recycle_index].depth,
1108                 };
1109
1110                 /* Set tbl24 before freeing tbl8 to avoid race condition.
1111                  * Prevent the free of the tbl8 group from hoisting.
1112                  */
1113                 __atomic_store(&lpm->tbl24[tbl24_index], &new_tbl24_entry,
1114                                 __ATOMIC_RELAXED);
1115                 __atomic_thread_fence(__ATOMIC_RELEASE);
1116                 tbl8_free(lpm, tbl8_group_start);
1117         }
1118 #undef group_idx
1119         return 0;
1120 }
1121
1122 /*
1123  * Deletes a rule
1124  */
1125 int
1126 rte_lpm_delete(struct rte_lpm *lpm, uint32_t ip, uint8_t depth)
1127 {
1128         int32_t rule_to_delete_index, sub_rule_index;
1129         uint32_t ip_masked;
1130         uint8_t sub_rule_depth;
1131         /*
1132          * Check input arguments. Note: IP must be a positive integer of 32
1133          * bits in length therefore it need not be checked.
1134          */
1135         if ((lpm == NULL) || (depth < 1) || (depth > RTE_LPM_MAX_DEPTH)) {
1136                 return -EINVAL;
1137         }
1138
1139         ip_masked = ip & depth_to_mask(depth);
1140
1141         /*
1142          * Find the index of the input rule, that needs to be deleted, in the
1143          * rule table.
1144          */
1145         rule_to_delete_index = rule_find(lpm, ip_masked, depth);
1146
1147         /*
1148          * Check if rule_to_delete_index was found. If no rule was found the
1149          * function rule_find returns -EINVAL.
1150          */
1151         if (rule_to_delete_index < 0)
1152                 return -EINVAL;
1153
1154         /* Delete the rule from the rule table. */
1155         rule_delete(lpm, rule_to_delete_index, depth);
1156
1157         /*
1158          * Find rule to replace the rule_to_delete. If there is no rule to
1159          * replace the rule_to_delete we return -1 and invalidate the table
1160          * entries associated with this rule.
1161          */
1162         sub_rule_depth = 0;
1163         sub_rule_index = find_previous_rule(lpm, ip, depth, &sub_rule_depth);
1164
1165         /*
1166          * If the input depth value is less than 25 use function
1167          * delete_depth_small otherwise use delete_depth_big.
1168          */
1169         if (depth <= MAX_DEPTH_TBL24) {
1170                 return delete_depth_small(lpm, ip_masked, depth,
1171                                 sub_rule_index, sub_rule_depth);
1172         } else { /* If depth > MAX_DEPTH_TBL24 */
1173                 return delete_depth_big(lpm, ip_masked, depth, sub_rule_index,
1174                                 sub_rule_depth);
1175         }
1176 }
1177
1178 /*
1179  * Delete all rules from the LPM table.
1180  */
1181 void
1182 rte_lpm_delete_all(struct rte_lpm *lpm)
1183 {
1184         /* Zero rule information. */
1185         memset(lpm->rule_info, 0, sizeof(lpm->rule_info));
1186
1187         /* Zero tbl24. */
1188         memset(lpm->tbl24, 0, sizeof(lpm->tbl24));
1189
1190         /* Zero tbl8. */
1191         memset(lpm->tbl8, 0, sizeof(lpm->tbl8[0])
1192                         * RTE_LPM_TBL8_GROUP_NUM_ENTRIES * lpm->number_tbl8s);
1193
1194         /* Delete all rules form the rules table. */
1195         memset(lpm->rules_tbl, 0, sizeof(lpm->rules_tbl[0]) * lpm->max_rules);
1196 }