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 int32_t
 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         int status;
 541
 542         internal_lpm = container_of(lpm, struct __rte_lpm, lpm);
 543         if (internal_lpm->v == NULL) {
 544                 /* Set tbl8 group invalid*/
 545                 __atomic_store(&lpm->tbl8[tbl8_group_start], &zero_tbl8_entry,
 546                                 __ATOMIC_RELAXED);
 547         } else if (internal_lpm->rcu_mode == RTE_LPM_QSBR_MODE_SYNC) {
 548                 /* Wait for quiescent state change. */
 549                 rte_rcu_qsbr_synchronize(internal_lpm->v,
 550                         RTE_QSBR_THRID_INVALID);
 551                 /* Set tbl8 group invalid*/
 552                 __atomic_store(&lpm->tbl8[tbl8_group_start], &zero_tbl8_entry,
 553                                 __ATOMIC_RELAXED);
 554         } else if (internal_lpm->rcu_mode == RTE_LPM_QSBR_MODE_DQ) {
 555                 /* Push into QSBR defer queue. */
 556                 status = rte_rcu_qsbr_dq_enqueue(internal_lpm->dq,
 557                                 (void *)&tbl8_group_start);
 558                 if (status == 1) {
 559                         RTE_LOG(ERR, LPM, "Failed to push QSBR FIFO\n");
 560                         return -rte_errno;
 561                 }
 562         }
 563
 564         return 0;
 565 }
 566
 567 static __rte_noinline int32_t
 568 add_depth_small(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
 569                 uint32_t next_hop)
 570 {
 571 #define group_idx next_hop
 572         uint32_t tbl24_index, tbl24_range, tbl8_index, tbl8_group_end, i, j;
 573
 574         /* Calculate the index into Table24. */
 575         tbl24_index = ip >> 8;
 576         tbl24_range = depth_to_range(depth);
 577
 578         for (i = tbl24_index; i < (tbl24_index + tbl24_range); i++) {
 579                 /*
 580                  * For invalid OR valid and non-extended tbl 24 entries set
 581                  * entry.
 582                  */
 583                 if (!lpm->tbl24[i].valid || (lpm->tbl24[i].valid_group == 0 &&
 584                                 lpm->tbl24[i].depth <= depth)) {
 585
 586                         struct rte_lpm_tbl_entry new_tbl24_entry = {
 587                                 .next_hop = next_hop,
 588                                 .valid = VALID,
 589                                 .valid_group = 0,
 590                                 .depth = depth,
 591                         };
 592
 593                         /* Setting tbl24 entry in one go to avoid race
 594                          * conditions
 595                          */
 596                         __atomic_store(&lpm->tbl24[i], &new_tbl24_entry,
 597                                         __ATOMIC_RELEASE);
 598
 599                         continue;
 600                 }
 601
 602                 if (lpm->tbl24[i].valid_group == 1) {
 603                         /* If tbl24 entry is valid and extended calculate the
 604                          *  index into tbl8.
 605                          */
 606                         tbl8_index = lpm->tbl24[i].group_idx *
 607                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 608                         tbl8_group_end = tbl8_index +
 609                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 610
 611                         for (j = tbl8_index; j < tbl8_group_end; j++) {
 612                                 if (!lpm->tbl8[j].valid ||
 613                                                 lpm->tbl8[j].depth <= depth) {
 614                                         struct rte_lpm_tbl_entry
 615                                                 new_tbl8_entry = {
 616                                                 .valid = VALID,
 617                                                 .valid_group = VALID,
 618                                                 .depth = depth,
 619                                                 .next_hop = next_hop,
 620                                         };
 621
 622                                         /*
 623                                          * Setting tbl8 entry in one go to avoid
 624                                          * race conditions
 625                                          */
 626                                         __atomic_store(&lpm->tbl8[j],
 627                                                 &new_tbl8_entry,
 628                                                 __ATOMIC_RELAXED);
 629
 630                                         continue;
 631                                 }
 632                         }
 633                 }
 634         }
 635 #undef group_idx
 636         return 0;
 637 }
 638
 639 static __rte_noinline int32_t
 640 add_depth_big(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth,
 641                 uint32_t next_hop)
 642 {
 643 #define group_idx next_hop
 644         uint32_t tbl24_index;
 645         int32_t tbl8_group_index, tbl8_group_start, tbl8_group_end, tbl8_index,
 646                 tbl8_range, i;
 647
 648         tbl24_index = (ip_masked >> 8);
 649         tbl8_range = depth_to_range(depth);
 650
 651         if (!lpm->tbl24[tbl24_index].valid) {
 652                 /* Search for a free tbl8 group. */
 653                 tbl8_group_index = tbl8_alloc(lpm);
 654
 655                 /* Check tbl8 allocation was successful. */
 656                 if (tbl8_group_index < 0) {
 657                         return tbl8_group_index;
 658                 }
 659
 660                 /* Find index into tbl8 and range. */
 661                 tbl8_index = (tbl8_group_index *
 662                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES) +
 663                                 (ip_masked & 0xFF);
 664
 665                 /* Set tbl8 entry. */
 666                 for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
 667                         struct rte_lpm_tbl_entry new_tbl8_entry = {
 668                                 .valid = VALID,
 669                                 .depth = depth,
 670                                 .valid_group = lpm->tbl8[i].valid_group,
 671                                 .next_hop = next_hop,
 672                         };
 673                         __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
 674                                         __ATOMIC_RELAXED);
 675                 }
 676
 677                 /*
 678                  * Update tbl24 entry to point to new tbl8 entry. Note: The
 679                  * ext_flag and tbl8_index need to be updated simultaneously,
 680                  * so assign whole structure in one go
 681                  */
 682
 683                 struct rte_lpm_tbl_entry new_tbl24_entry = {
 684                         .group_idx = tbl8_group_index,
 685                         .valid = VALID,
 686                         .valid_group = 1,
 687                         .depth = 0,
 688                 };
 689
 690                 /* The tbl24 entry must be written only after the
 691                  * tbl8 entries are written.
 692                  */
 693                 __atomic_store(&lpm->tbl24[tbl24_index], &new_tbl24_entry,
 694                                 __ATOMIC_RELEASE);
 695
 696         } /* If valid entry but not extended calculate the index into Table8. */
 697         else if (lpm->tbl24[tbl24_index].valid_group == 0) {
 698                 /* Search for free tbl8 group. */
 699                 tbl8_group_index = tbl8_alloc(lpm);
 700
 701                 if (tbl8_group_index < 0) {
 702                         return tbl8_group_index;
 703                 }
 704
 705                 tbl8_group_start = tbl8_group_index *
 706                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 707                 tbl8_group_end = tbl8_group_start +
 708                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 709
 710                 /* Populate new tbl8 with tbl24 value. */
 711                 for (i = tbl8_group_start; i < tbl8_group_end; i++) {
 712                         struct rte_lpm_tbl_entry new_tbl8_entry = {
 713                                 .valid = VALID,
 714                                 .depth = lpm->tbl24[tbl24_index].depth,
 715                                 .valid_group = lpm->tbl8[i].valid_group,
 716                                 .next_hop = lpm->tbl24[tbl24_index].next_hop,
 717                         };
 718                         __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
 719                                         __ATOMIC_RELAXED);
 720                 }
 721
 722                 tbl8_index = tbl8_group_start + (ip_masked & 0xFF);
 723
 724                 /* Insert new rule into the tbl8 entry. */
 725                 for (i = tbl8_index; i < tbl8_index + tbl8_range; i++) {
 726                         struct rte_lpm_tbl_entry new_tbl8_entry = {
 727                                 .valid = VALID,
 728                                 .depth = depth,
 729                                 .valid_group = lpm->tbl8[i].valid_group,
 730                                 .next_hop = next_hop,
 731                         };
 732                         __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
 733                                         __ATOMIC_RELAXED);
 734                 }
 735
 736                 /*
 737                  * Update tbl24 entry to point to new tbl8 entry. Note: The
 738                  * ext_flag and tbl8_index need to be updated simultaneously,
 739                  * so assign whole structure in one go.
 740                  */
 741
 742                 struct rte_lpm_tbl_entry new_tbl24_entry = {
 743                                 .group_idx = tbl8_group_index,
 744                                 .valid = VALID,
 745                                 .valid_group = 1,
 746                                 .depth = 0,
 747                 };
 748
 749                 /* The tbl24 entry must be written only after the
 750                  * tbl8 entries are written.
 751                  */
 752                 __atomic_store(&lpm->tbl24[tbl24_index], &new_tbl24_entry,
 753                                 __ATOMIC_RELEASE);
 754
 755         } else { /*
 756                 * If it is valid, extended entry calculate the index into tbl8.
 757                 */
 758                 tbl8_group_index = lpm->tbl24[tbl24_index].group_idx;
 759                 tbl8_group_start = tbl8_group_index *
 760                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 761                 tbl8_index = tbl8_group_start + (ip_masked & 0xFF);
 762
 763                 for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
 764
 765                         if (!lpm->tbl8[i].valid ||
 766                                         lpm->tbl8[i].depth <= depth) {
 767                                 struct rte_lpm_tbl_entry new_tbl8_entry = {
 768                                         .valid = VALID,
 769                                         .depth = depth,
 770                                         .next_hop = next_hop,
 771                                         .valid_group = lpm->tbl8[i].valid_group,
 772                                 };
 773
 774                                 /*
 775                                  * Setting tbl8 entry in one go to avoid race
 776                                  * condition
 777                                  */
 778                                 __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
 779                                                 __ATOMIC_RELAXED);
 780
 781                                 continue;
 782                         }
 783                 }
 784         }
 785 #undef group_idx
 786         return 0;
 787 }
 788
 789 /*
 790  * Add a route
 791  */
 792 int
 793 rte_lpm_add(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
 794                 uint32_t next_hop)
 795 {
 796         int32_t rule_index, status = 0;
 797         uint32_t ip_masked;
 798
 799         /* Check user arguments. */
 800         if ((lpm == NULL) || (depth < 1) || (depth > RTE_LPM_MAX_DEPTH))
 801                 return -EINVAL;
 802
 803         ip_masked = ip & depth_to_mask(depth);
 804
 805         /* Add the rule to the rule table. */
 806         rule_index = rule_add(lpm, ip_masked, depth, next_hop);
 807
 808         /* Skip table entries update if The rule is the same as
 809          * the rule in the rules table.
 810          */
 811         if (rule_index == -EEXIST)
 812                 return 0;
 813
 814         /* If the is no space available for new rule return error. */
 815         if (rule_index < 0) {
 816                 return rule_index;
 817         }
 818
 819         if (depth <= MAX_DEPTH_TBL24) {
 820                 status = add_depth_small(lpm, ip_masked, depth, next_hop);
 821         } else { /* If depth > RTE_LPM_MAX_DEPTH_TBL24 */
 822                 status = add_depth_big(lpm, ip_masked, depth, next_hop);
 823
 824                 /*
 825                  * If add fails due to exhaustion of tbl8 extensions delete
 826                  * rule that was added to rule table.
 827                  */
 828                 if (status < 0) {
 829                         rule_delete(lpm, rule_index, depth);
 830
 831                         return status;
 832                 }
 833         }
 834
 835         return 0;
 836 }
 837
 838 /*
 839  * Look for a rule in the high-level rules table
 840  */
 841 int
 842 rte_lpm_is_rule_present(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
 843 uint32_t *next_hop)
 844 {
 845         uint32_t ip_masked;
 846         int32_t rule_index;
 847
 848         /* Check user arguments. */
 849         if ((lpm == NULL) ||
 850                 (next_hop == NULL) ||
 851                 (depth < 1) || (depth > RTE_LPM_MAX_DEPTH))
 852                 return -EINVAL;
 853
 854         /* Look for the rule using rule_find. */
 855         ip_masked = ip & depth_to_mask(depth);
 856         rule_index = rule_find(lpm, ip_masked, depth);
 857
 858         if (rule_index >= 0) {
 859                 *next_hop = lpm->rules_tbl[rule_index].next_hop;
 860                 return 1;
 861         }
 862
 863         /* If rule is not found return 0. */
 864         return 0;
 865 }
 866
 867 static int32_t
 868 find_previous_rule(struct rte_lpm *lpm, uint32_t ip, uint8_t depth,
 869                 uint8_t *sub_rule_depth)
 870 {
 871         int32_t rule_index;
 872         uint32_t ip_masked;
 873         uint8_t prev_depth;
 874
 875         for (prev_depth = (uint8_t)(depth - 1); prev_depth > 0; prev_depth--) {
 876                 ip_masked = ip & depth_to_mask(prev_depth);
 877
 878                 rule_index = rule_find(lpm, ip_masked, prev_depth);
 879
 880                 if (rule_index >= 0) {
 881                         *sub_rule_depth = prev_depth;
 882                         return rule_index;
 883                 }
 884         }
 885
 886         return -1;
 887 }
 888
 889 static int32_t
 890 delete_depth_small(struct rte_lpm *lpm, uint32_t ip_masked,
 891         uint8_t depth, int32_t sub_rule_index, uint8_t sub_rule_depth)
 892 {
 893 #define group_idx next_hop
 894         uint32_t tbl24_range, tbl24_index, tbl8_group_index, tbl8_index, i, j;
 895
 896         /* Calculate the range and index into Table24. */
 897         tbl24_range = depth_to_range(depth);
 898         tbl24_index = (ip_masked >> 8);
 899         struct rte_lpm_tbl_entry zero_tbl24_entry = {0};
 900
 901         /*
 902          * Firstly check the sub_rule_index. A -1 indicates no replacement rule
 903          * and a positive number indicates a sub_rule_index.
 904          */
 905         if (sub_rule_index < 0) {
 906                 /*
 907                  * If no replacement rule exists then invalidate entries
 908                  * associated with this rule.
 909                  */
 910                 for (i = tbl24_index; i < (tbl24_index + tbl24_range); i++) {
 911
 912                         if (lpm->tbl24[i].valid_group == 0 &&
 913                                         lpm->tbl24[i].depth <= depth) {
 914                                 __atomic_store(&lpm->tbl24[i],
 915                                         &zero_tbl24_entry, __ATOMIC_RELEASE);
 916                         } else if (lpm->tbl24[i].valid_group == 1) {
 917                                 /*
 918                                  * If TBL24 entry is extended, then there has
 919                                  * to be a rule with depth >= 25 in the
 920                                  * associated TBL8 group.
 921                                  */
 922
 923                                 tbl8_group_index = lpm->tbl24[i].group_idx;
 924                                 tbl8_index = tbl8_group_index *
 925                                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 926
 927                                 for (j = tbl8_index; j < (tbl8_index +
 928                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES); j++) {
 929
 930                                         if (lpm->tbl8[j].depth <= depth)
 931                                                 lpm->tbl8[j].valid = INVALID;
 932                                 }
 933                         }
 934                 }
 935         } else {
 936                 /*
 937                  * If a replacement rule exists then modify entries
 938                  * associated with this rule.
 939                  */
 940
 941                 struct rte_lpm_tbl_entry new_tbl24_entry = {
 942                         .next_hop = lpm->rules_tbl[sub_rule_index].next_hop,
 943                         .valid = VALID,
 944                         .valid_group = 0,
 945                         .depth = sub_rule_depth,
 946                 };
 947
 948                 struct rte_lpm_tbl_entry new_tbl8_entry = {
 949                         .valid = VALID,
 950                         .valid_group = VALID,
 951                         .depth = sub_rule_depth,
 952                         .next_hop = lpm->rules_tbl
 953                         [sub_rule_index].next_hop,
 954                 };
 955
 956                 for (i = tbl24_index; i < (tbl24_index + tbl24_range); i++) {
 957
 958                         if (lpm->tbl24[i].valid_group == 0 &&
 959                                         lpm->tbl24[i].depth <= depth) {
 960                                 __atomic_store(&lpm->tbl24[i], &new_tbl24_entry,
 961                                                 __ATOMIC_RELEASE);
 962                         } else  if (lpm->tbl24[i].valid_group == 1) {
 963                                 /*
 964                                  * If TBL24 entry is extended, then there has
 965                                  * to be a rule with depth >= 25 in the
 966                                  * associated TBL8 group.
 967                                  */
 968
 969                                 tbl8_group_index = lpm->tbl24[i].group_idx;
 970                                 tbl8_index = tbl8_group_index *
 971                                                 RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 972
 973                                 for (j = tbl8_index; j < (tbl8_index +
 974                                         RTE_LPM_TBL8_GROUP_NUM_ENTRIES); j++) {
 975
 976                                         if (lpm->tbl8[j].depth <= depth)
 977                                                 __atomic_store(&lpm->tbl8[j],
 978                                                         &new_tbl8_entry,
 979                                                         __ATOMIC_RELAXED);
 980                                 }
 981                         }
 982                 }
 983         }
 984 #undef group_idx
 985         return 0;
 986 }
 987
 988 /*
 989  * Checks if table 8 group can be recycled.
 990  *
 991  * Return of -EEXIST means tbl8 is in use and thus can not be recycled.
 992  * Return of -EINVAL means tbl8 is empty and thus can be recycled
 993  * Return of value > -1 means tbl8 is in use but has all the same values and
 994  * thus can be recycled
 995  */
 996 static int32_t
 997 tbl8_recycle_check(struct rte_lpm_tbl_entry *tbl8,
 998                 uint32_t tbl8_group_start)
 999 {
1000         uint32_t tbl8_group_end, i;
1001         tbl8_group_end = tbl8_group_start + RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
1002
1003         /*
1004          * Check the first entry of the given tbl8. If it is invalid we know
1005          * this tbl8 does not contain any rule with a depth < RTE_LPM_MAX_DEPTH
1006          *  (As they would affect all entries in a tbl8) and thus this table
1007          *  can not be recycled.
1008          */
1009         if (tbl8[tbl8_group_start].valid) {
1010                 /*
1011                  * If first entry is valid check if the depth is less than 24
1012                  * and if so check the rest of the entries to verify that they
1013                  * are all of this depth.
1014                  */
1015                 if (tbl8[tbl8_group_start].depth <= MAX_DEPTH_TBL24) {
1016                         for (i = (tbl8_group_start + 1); i < tbl8_group_end;
1017                                         i++) {
1018
1019                                 if (tbl8[i].depth !=
1020                                                 tbl8[tbl8_group_start].depth) {
1021
1022                                         return -EEXIST;
1023                                 }
1024                         }
1025                         /* If all entries are the same return the tb8 index */
1026                         return tbl8_group_start;
1027                 }
1028
1029                 return -EEXIST;
1030         }
1031         /*
1032          * If the first entry is invalid check if the rest of the entries in
1033          * the tbl8 are invalid.
1034          */
1035         for (i = (tbl8_group_start + 1); i < tbl8_group_end; i++) {
1036                 if (tbl8[i].valid)
1037                         return -EEXIST;
1038         }
1039         /* If no valid entries are found then return -EINVAL. */
1040         return -EINVAL;
1041 }
1042
1043 static int32_t
1044 delete_depth_big(struct rte_lpm *lpm, uint32_t ip_masked,
1045         uint8_t depth, int32_t sub_rule_index, uint8_t sub_rule_depth)
1046 {
1047 #define group_idx next_hop
1048         uint32_t tbl24_index, tbl8_group_index, tbl8_group_start, tbl8_index,
1049                         tbl8_range, i;
1050         int32_t tbl8_recycle_index, status = 0;
1051
1052         /*
1053          * Calculate the index into tbl24 and range. Note: All depths larger
1054          * than MAX_DEPTH_TBL24 are associated with only one tbl24 entry.
1055          */
1056         tbl24_index = ip_masked >> 8;
1057
1058         /* Calculate the index into tbl8 and range. */
1059         tbl8_group_index = lpm->tbl24[tbl24_index].group_idx;
1060         tbl8_group_start = tbl8_group_index * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
1061         tbl8_index = tbl8_group_start + (ip_masked & 0xFF);
1062         tbl8_range = depth_to_range(depth);
1063
1064         if (sub_rule_index < 0) {
1065                 /*
1066                  * Loop through the range of entries on tbl8 for which the
1067                  * rule_to_delete must be removed or modified.
1068                  */
1069                 for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
1070                         if (lpm->tbl8[i].depth <= depth)
1071                                 lpm->tbl8[i].valid = INVALID;
1072                 }
1073         } else {
1074                 /* Set new tbl8 entry. */
1075                 struct rte_lpm_tbl_entry new_tbl8_entry = {
1076                         .valid = VALID,
1077                         .depth = sub_rule_depth,
1078                         .valid_group = lpm->tbl8[tbl8_group_start].valid_group,
1079                         .next_hop = lpm->rules_tbl[sub_rule_index].next_hop,
1080                 };
1081
1082                 /*
1083                  * Loop through the range of entries on tbl8 for which the
1084                  * rule_to_delete must be modified.
1085                  */
1086                 for (i = tbl8_index; i < (tbl8_index + tbl8_range); i++) {
1087                         if (lpm->tbl8[i].depth <= depth)
1088                                 __atomic_store(&lpm->tbl8[i], &new_tbl8_entry,
1089                                                 __ATOMIC_RELAXED);
1090                 }
1091         }
1092
1093         /*
1094          * Check if there are any valid entries in this tbl8 group. If all
1095          * tbl8 entries are invalid we can free the tbl8 and invalidate the
1096          * associated tbl24 entry.
1097          */
1098
1099         tbl8_recycle_index = tbl8_recycle_check(lpm->tbl8, tbl8_group_start);
1100
1101         if (tbl8_recycle_index == -EINVAL) {
1102                 /* Set tbl24 before freeing tbl8 to avoid race condition.
1103                  * Prevent the free of the tbl8 group from hoisting.
1104                  */
1105                 lpm->tbl24[tbl24_index].valid = 0;
1106                 __atomic_thread_fence(__ATOMIC_RELEASE);
1107                 status = tbl8_free(lpm, tbl8_group_start);
1108         } else if (tbl8_recycle_index > -1) {
1109                 /* Update tbl24 entry. */
1110                 struct rte_lpm_tbl_entry new_tbl24_entry = {
1111                         .next_hop = lpm->tbl8[tbl8_recycle_index].next_hop,
1112                         .valid = VALID,
1113                         .valid_group = 0,
1114                         .depth = lpm->tbl8[tbl8_recycle_index].depth,
1115                 };
1116
1117                 /* Set tbl24 before freeing tbl8 to avoid race condition.
1118                  * Prevent the free of the tbl8 group from hoisting.
1119                  */
1120                 __atomic_store(&lpm->tbl24[tbl24_index], &new_tbl24_entry,
1121                                 __ATOMIC_RELAXED);
1122                 __atomic_thread_fence(__ATOMIC_RELEASE);
1123                 status = tbl8_free(lpm, tbl8_group_start);
1124         }
1125 #undef group_idx
1126         return status;
1127 }
1128
1129 /*
1130  * Deletes a rule
1131  */
1132 int
1133 rte_lpm_delete(struct rte_lpm *lpm, uint32_t ip, uint8_t depth)
1134 {
1135         int32_t rule_to_delete_index, sub_rule_index;
1136         uint32_t ip_masked;
1137         uint8_t sub_rule_depth;
1138         /*
1139          * Check input arguments. Note: IP must be a positive integer of 32
1140          * bits in length therefore it need not be checked.
1141          */
1142         if ((lpm == NULL) || (depth < 1) || (depth > RTE_LPM_MAX_DEPTH)) {
1143                 return -EINVAL;
1144         }
1145
1146         ip_masked = ip & depth_to_mask(depth);
1147
1148         /*
1149          * Find the index of the input rule, that needs to be deleted, in the
1150          * rule table.
1151          */
1152         rule_to_delete_index = rule_find(lpm, ip_masked, depth);
1153
1154         /*
1155          * Check if rule_to_delete_index was found. If no rule was found the
1156          * function rule_find returns -EINVAL.
1157          */
1158         if (rule_to_delete_index < 0)
1159                 return -EINVAL;
1160
1161         /* Delete the rule from the rule table. */
1162         rule_delete(lpm, rule_to_delete_index, depth);
1163
1164         /*
1165          * Find rule to replace the rule_to_delete. If there is no rule to
1166          * replace the rule_to_delete we return -1 and invalidate the table
1167          * entries associated with this rule.
1168          */
1169         sub_rule_depth = 0;
1170         sub_rule_index = find_previous_rule(lpm, ip, depth, &sub_rule_depth);
1171
1172         /*
1173          * If the input depth value is less than 25 use function
1174          * delete_depth_small otherwise use delete_depth_big.
1175          */
1176         if (depth <= MAX_DEPTH_TBL24) {
1177                 return delete_depth_small(lpm, ip_masked, depth,
1178                                 sub_rule_index, sub_rule_depth);
1179         } else { /* If depth > MAX_DEPTH_TBL24 */
1180                 return delete_depth_big(lpm, ip_masked, depth, sub_rule_index,
1181                                 sub_rule_depth);
1182         }
1183 }
1184
1185 /*
1186  * Delete all rules from the LPM table.
1187  */
1188 void
1189 rte_lpm_delete_all(struct rte_lpm *lpm)
1190 {
1191         /* Zero rule information. */
1192         memset(lpm->rule_info, 0, sizeof(lpm->rule_info));
1193
1194         /* Zero tbl24. */
1195         memset(lpm->tbl24, 0, sizeof(lpm->tbl24));
1196
1197         /* Zero tbl8. */
1198         memset(lpm->tbl8, 0, sizeof(lpm->tbl8[0])
1199                         * RTE_LPM_TBL8_GROUP_NUM_ENTRIES * lpm->number_tbl8s);
1200
1201         /* Delete all rules form the rules table. */
1202         memset(lpm->rules_tbl, 0, sizeof(lpm->rules_tbl[0]) * lpm->max_rules);
1203 }