lib/librte_mbuf/rte_mbuf_dyn.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2019 6WIND S.A.
   3  */
   4
   5 #include <sys/queue.h>
   6 #include <stdint.h>
   7 #include <limits.h>
   8
   9 #include <rte_common.h>
  10 #include <rte_eal.h>
  11 #include <rte_eal_memconfig.h>
  12 #include <rte_tailq.h>
  13 #include <rte_errno.h>
  14 #include <rte_malloc.h>
  15 #include <rte_string_fns.h>
  16 #include <rte_bitops.h>
  17 #include <rte_mbuf.h>
  18 #include <rte_mbuf_dyn.h>
  19
  20 #define RTE_MBUF_DYN_MZNAME "rte_mbuf_dyn"
  21
  22 struct mbuf_dynfield_elt {
  23         struct rte_mbuf_dynfield params;
  24         size_t offset;
  25 };
  26 TAILQ_HEAD(mbuf_dynfield_list, rte_tailq_entry);
  27
  28 static struct rte_tailq_elem mbuf_dynfield_tailq = {
  29         .name = "RTE_MBUF_DYNFIELD",
  30 };
  31 EAL_REGISTER_TAILQ(mbuf_dynfield_tailq);
  32
  33 struct mbuf_dynflag_elt {
  34         struct rte_mbuf_dynflag params;
  35         unsigned int bitnum;
  36 };
  37 TAILQ_HEAD(mbuf_dynflag_list, rte_tailq_entry);
  38
  39 static struct rte_tailq_elem mbuf_dynflag_tailq = {
  40         .name = "RTE_MBUF_DYNFLAG",
  41 };
  42 EAL_REGISTER_TAILQ(mbuf_dynflag_tailq);
  43
  44 struct mbuf_dyn_shm {
  45         /**
  46          * For each mbuf byte, free_space[i] != 0 if space is free.
  47          * The value is the size of the biggest aligned element that
  48          * can fit in the zone.
  49          */
  50         uint8_t free_space[sizeof(struct rte_mbuf)];
  51         /** Bitfield of available flags. */
  52         uint64_t free_flags;
  53 };
  54 static struct mbuf_dyn_shm *shm;
  55
  56 /* Set the value of free_space[] according to the size and alignment of
  57  * the free areas. This helps to select the best place when reserving a
  58  * dynamic field. Assume tailq is locked.
  59  */
  60 static void
  61 process_score(void)
  62 {
  63         size_t off, align, size, i;
  64
  65         /* first, erase previous info */
  66         for (i = 0; i < sizeof(struct rte_mbuf); i++) {
  67                 if (shm->free_space[i])
  68                         shm->free_space[i] = 1;
  69         }
  70
  71         off = 0;
  72         while (off < sizeof(struct rte_mbuf)) {
  73                 /* get the size of the free zone */
  74                 for (size = 0; (off + size) < sizeof(struct rte_mbuf) &&
  75                              shm->free_space[off + size]; size++)
  76                         ;
  77                 if (size == 0) {
  78                         off++;
  79                         continue;
  80                 }
  81
  82                 /* get the alignment of biggest object that can fit in
  83                  * the zone at this offset.
  84                  */
  85                 for (align = 1;
  86                      (off % (align << 1)) == 0 && (align << 1) <= size;
  87                      align <<= 1)
  88                         ;
  89
  90                 /* save it in free_space[] */
  91                 for (i = off; i < off + align; i++)
  92                         shm->free_space[i] = RTE_MAX(align, shm->free_space[i]);
  93
  94                 off += align;
  95         }
  96 }
  97
  98 /* Mark the area occupied by a mbuf field as available in the shm. */
  99 #define mark_free(field)                                                \
 100         memset(&shm->free_space[offsetof(struct rte_mbuf, field)],      \
 101                 1, sizeof(((struct rte_mbuf *)0)->field))
 102
 103 /* Allocate and initialize the shared memory. Assume tailq is locked */
 104 static int
 105 init_shared_mem(void)
 106 {
 107         const struct rte_memzone *mz;
 108         uint64_t mask;
 109
 110         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
 111                 mz = rte_memzone_reserve_aligned(RTE_MBUF_DYN_MZNAME,
 112                                                 sizeof(struct mbuf_dyn_shm),
 113                                                 SOCKET_ID_ANY, 0,
 114                                                 RTE_CACHE_LINE_SIZE);
 115         } else {
 116                 mz = rte_memzone_lookup(RTE_MBUF_DYN_MZNAME);
 117         }
 118         if (mz == NULL)
 119                 return -1;
 120
 121         shm = mz->addr;
 122
 123         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
 124                 /* init free_space, keep it sync'd with
 125                  * rte_mbuf_dynfield_copy().
 126                  */
 127                 memset(shm, 0, sizeof(*shm));
 128                 mark_free(dynfield1);
 129
 130                 /* init free_flags */
 131                 for (mask = PKT_FIRST_FREE; mask <= PKT_LAST_FREE; mask <<= 1)
 132                         shm->free_flags |= mask;
 133
 134                 process_score();
 135         }
 136
 137         return 0;
 138 }
 139
 140 /* check if this offset can be used */
 141 static int
 142 check_offset(size_t offset, size_t size, size_t align)
 143 {
 144         size_t i;
 145
 146         if ((offset & (align - 1)) != 0)
 147                 return -1;
 148         if (offset + size > sizeof(struct rte_mbuf))
 149                 return -1;
 150
 151         for (i = 0; i < size; i++) {
 152                 if (!shm->free_space[i + offset])
 153                         return -1;
 154         }
 155
 156         return 0;
 157 }
 158
 159 /* assume tailq is locked */
 160 static struct mbuf_dynfield_elt *
 161 __mbuf_dynfield_lookup(const char *name)
 162 {
 163         struct mbuf_dynfield_list *mbuf_dynfield_list;
 164         struct mbuf_dynfield_elt *mbuf_dynfield;
 165         struct rte_tailq_entry *te;
 166
 167         mbuf_dynfield_list = RTE_TAILQ_CAST(
 168                 mbuf_dynfield_tailq.head, mbuf_dynfield_list);
 169
 170         TAILQ_FOREACH(te, mbuf_dynfield_list, next) {
 171                 mbuf_dynfield = (struct mbuf_dynfield_elt *)te->data;
 172                 if (strcmp(name, mbuf_dynfield->params.name) == 0)
 173                         break;
 174         }
 175
 176         if (te == NULL || mbuf_dynfield == NULL) {
 177                 rte_errno = ENOENT;
 178                 return NULL;
 179         }
 180
 181         return mbuf_dynfield;
 182 }
 183
 184 int
 185 rte_mbuf_dynfield_lookup(const char *name, struct rte_mbuf_dynfield *params)
 186 {
 187         struct mbuf_dynfield_elt *mbuf_dynfield;
 188
 189         rte_mcfg_tailq_read_lock();
 190         if (shm == NULL && init_shared_mem() < 0)
 191                 mbuf_dynfield = NULL;
 192         else
 193                 mbuf_dynfield = __mbuf_dynfield_lookup(name);
 194         rte_mcfg_tailq_read_unlock();
 195
 196         if (mbuf_dynfield == NULL)
 197                 return -1;
 198
 199         if (params != NULL)
 200                 memcpy(params, &mbuf_dynfield->params, sizeof(*params));
 201
 202         return mbuf_dynfield->offset;
 203 }
 204
 205 static int mbuf_dynfield_cmp(const struct rte_mbuf_dynfield *params1,
 206                 const struct rte_mbuf_dynfield *params2)
 207 {
 208         if (strcmp(params1->name, params2->name))
 209                 return -1;
 210         if (params1->size != params2->size)
 211                 return -1;
 212         if (params1->align != params2->align)
 213                 return -1;
 214         if (params1->flags != params2->flags)
 215                 return -1;
 216         return 0;
 217 }
 218
 219 /* assume tailq is locked */
 220 static int
 221 __rte_mbuf_dynfield_register_offset(const struct rte_mbuf_dynfield *params,
 222                                 size_t req)
 223 {
 224         struct mbuf_dynfield_list *mbuf_dynfield_list;
 225         struct mbuf_dynfield_elt *mbuf_dynfield = NULL;
 226         struct rte_tailq_entry *te = NULL;
 227         unsigned int best_zone = UINT_MAX;
 228         size_t i, offset;
 229         int ret;
 230
 231         if (shm == NULL && init_shared_mem() < 0)
 232                 return -1;
 233
 234         mbuf_dynfield = __mbuf_dynfield_lookup(params->name);
 235         if (mbuf_dynfield != NULL) {
 236                 if (req != SIZE_MAX && req != mbuf_dynfield->offset) {
 237                         rte_errno = EEXIST;
 238                         return -1;
 239                 }
 240                 if (mbuf_dynfield_cmp(params, &mbuf_dynfield->params) < 0) {
 241                         rte_errno = EEXIST;
 242                         return -1;
 243                 }
 244                 return mbuf_dynfield->offset;
 245         }
 246
 247         if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
 248                 rte_errno = EPERM;
 249                 return -1;
 250         }
 251
 252         if (req == SIZE_MAX) {
 253                 /* Find the best place to put this field: we search the
 254                  * lowest value of shm->free_space[offset]: the zones
 255                  * containing room for larger fields are kept for later.
 256                  */
 257                 for (offset = 0;
 258                      offset < sizeof(struct rte_mbuf);
 259                      offset++) {
 260                         if (check_offset(offset, params->size,
 261                                                 params->align) == 0 &&
 262                                         shm->free_space[offset] < best_zone) {
 263                                 best_zone = shm->free_space[offset];
 264                                 req = offset;
 265                         }
 266                 }
 267                 if (req == SIZE_MAX) {
 268                         rte_errno = ENOENT;
 269                         return -1;
 270                 }
 271         } else {
 272                 if (check_offset(req, params->size, params->align) < 0) {
 273                         rte_errno = EBUSY;
 274                         return -1;
 275                 }
 276         }
 277
 278         offset = req;
 279         mbuf_dynfield_list = RTE_TAILQ_CAST(
 280                 mbuf_dynfield_tailq.head, mbuf_dynfield_list);
 281
 282         te = rte_zmalloc("MBUF_DYNFIELD_TAILQ_ENTRY", sizeof(*te), 0);
 283         if (te == NULL) {
 284                 rte_errno = ENOMEM;
 285                 return -1;
 286         }
 287
 288         mbuf_dynfield = rte_zmalloc("mbuf_dynfield", sizeof(*mbuf_dynfield), 0);
 289         if (mbuf_dynfield == NULL) {
 290                 rte_free(te);
 291                 rte_errno = ENOMEM;
 292                 return -1;
 293         }
 294
 295         ret = strlcpy(mbuf_dynfield->params.name, params->name,
 296                 sizeof(mbuf_dynfield->params.name));
 297         if (ret < 0 || ret >= (int)sizeof(mbuf_dynfield->params.name)) {
 298                 rte_errno = ENAMETOOLONG;
 299                 rte_free(mbuf_dynfield);
 300                 rte_free(te);
 301                 return -1;
 302         }
 303         memcpy(&mbuf_dynfield->params, params, sizeof(mbuf_dynfield->params));
 304         mbuf_dynfield->offset = offset;
 305         te->data = mbuf_dynfield;
 306
 307         TAILQ_INSERT_TAIL(mbuf_dynfield_list, te, next);
 308
 309         for (i = offset; i < offset + params->size; i++)
 310                 shm->free_space[i] = 0;
 311         process_score();
 312
 313         RTE_LOG(DEBUG, MBUF, "Registered dynamic field %s (sz=%zu, al=%zu, fl=0x%x) -> %zd\n",
 314                 params->name, params->size, params->align, params->flags,
 315                 offset);
 316
 317         return offset;
 318 }
 319
 320 int
 321 rte_mbuf_dynfield_register_offset(const struct rte_mbuf_dynfield *params,
 322                                 size_t req)
 323 {
 324         int ret;
 325
 326         if (params->size >= sizeof(struct rte_mbuf)) {
 327                 rte_errno = EINVAL;
 328                 return -1;
 329         }
 330         if (!rte_is_power_of_2(params->align)) {
 331                 rte_errno = EINVAL;
 332                 return -1;
 333         }
 334         if (params->flags != 0) {
 335                 rte_errno = EINVAL;
 336                 return -1;
 337         }
 338
 339         rte_mcfg_tailq_write_lock();
 340         ret = __rte_mbuf_dynfield_register_offset(params, req);
 341         rte_mcfg_tailq_write_unlock();
 342
 343         return ret;
 344 }
 345
 346 int
 347 rte_mbuf_dynfield_register(const struct rte_mbuf_dynfield *params)
 348 {
 349         return rte_mbuf_dynfield_register_offset(params, SIZE_MAX);
 350 }
 351
 352 /* assume tailq is locked */
 353 static struct mbuf_dynflag_elt *
 354 __mbuf_dynflag_lookup(const char *name)
 355 {
 356         struct mbuf_dynflag_list *mbuf_dynflag_list;
 357         struct mbuf_dynflag_elt *mbuf_dynflag;
 358         struct rte_tailq_entry *te;
 359
 360         mbuf_dynflag_list = RTE_TAILQ_CAST(
 361                 mbuf_dynflag_tailq.head, mbuf_dynflag_list);
 362
 363         TAILQ_FOREACH(te, mbuf_dynflag_list, next) {
 364                 mbuf_dynflag = (struct mbuf_dynflag_elt *)te->data;
 365                 if (strncmp(name, mbuf_dynflag->params.name,
 366                                 RTE_MBUF_DYN_NAMESIZE) == 0)
 367                         break;
 368         }
 369
 370         if (te == NULL) {
 371                 rte_errno = ENOENT;
 372                 return NULL;
 373         }
 374
 375         return mbuf_dynflag;
 376 }
 377
 378 int
 379 rte_mbuf_dynflag_lookup(const char *name,
 380                         struct rte_mbuf_dynflag *params)
 381 {
 382         struct mbuf_dynflag_elt *mbuf_dynflag;
 383
 384         rte_mcfg_tailq_read_lock();
 385         if (shm == NULL && init_shared_mem() < 0)
 386                 mbuf_dynflag = NULL;
 387         else
 388                 mbuf_dynflag = __mbuf_dynflag_lookup(name);
 389         rte_mcfg_tailq_read_unlock();
 390
 391         if (mbuf_dynflag == NULL)
 392                 return -1;
 393
 394         if (params != NULL)
 395                 memcpy(params, &mbuf_dynflag->params, sizeof(*params));
 396
 397         return mbuf_dynflag->bitnum;
 398 }
 399
 400 static int mbuf_dynflag_cmp(const struct rte_mbuf_dynflag *params1,
 401                 const struct rte_mbuf_dynflag *params2)
 402 {
 403         if (strcmp(params1->name, params2->name))
 404                 return -1;
 405         if (params1->flags != params2->flags)
 406                 return -1;
 407         return 0;
 408 }
 409
 410 /* assume tailq is locked */
 411 static int
 412 __rte_mbuf_dynflag_register_bitnum(const struct rte_mbuf_dynflag *params,
 413                                 unsigned int req)
 414 {
 415         struct mbuf_dynflag_list *mbuf_dynflag_list;
 416         struct mbuf_dynflag_elt *mbuf_dynflag = NULL;
 417         struct rte_tailq_entry *te = NULL;
 418         unsigned int bitnum;
 419         int ret;
 420
 421         if (shm == NULL && init_shared_mem() < 0)
 422                 return -1;
 423
 424         mbuf_dynflag = __mbuf_dynflag_lookup(params->name);
 425         if (mbuf_dynflag != NULL) {
 426                 if (req != UINT_MAX && req != mbuf_dynflag->bitnum) {
 427                         rte_errno = EEXIST;
 428                         return -1;
 429                 }
 430                 if (mbuf_dynflag_cmp(params, &mbuf_dynflag->params) < 0) {
 431                         rte_errno = EEXIST;
 432                         return -1;
 433                 }
 434                 return mbuf_dynflag->bitnum;
 435         }
 436
 437         if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
 438                 rte_errno = EPERM;
 439                 return -1;
 440         }
 441
 442         if (req == UINT_MAX) {
 443                 if (shm->free_flags == 0) {
 444                         rte_errno = ENOENT;
 445                         return -1;
 446                 }
 447                 bitnum = rte_bsf64(shm->free_flags);
 448         } else {
 449                 if ((shm->free_flags & (1ULL << req)) == 0) {
 450                         rte_errno = EBUSY;
 451                         return -1;
 452                 }
 453                 bitnum = req;
 454         }
 455
 456         mbuf_dynflag_list = RTE_TAILQ_CAST(
 457                 mbuf_dynflag_tailq.head, mbuf_dynflag_list);
 458
 459         te = rte_zmalloc("MBUF_DYNFLAG_TAILQ_ENTRY", sizeof(*te), 0);
 460         if (te == NULL) {
 461                 rte_errno = ENOMEM;
 462                 return -1;
 463         }
 464
 465         mbuf_dynflag = rte_zmalloc("mbuf_dynflag", sizeof(*mbuf_dynflag), 0);
 466         if (mbuf_dynflag == NULL) {
 467                 rte_free(te);
 468                 rte_errno = ENOMEM;
 469                 return -1;
 470         }
 471
 472         ret = strlcpy(mbuf_dynflag->params.name, params->name,
 473                 sizeof(mbuf_dynflag->params.name));
 474         if (ret < 0 || ret >= (int)sizeof(mbuf_dynflag->params.name)) {
 475                 rte_free(mbuf_dynflag);
 476                 rte_free(te);
 477                 rte_errno = ENAMETOOLONG;
 478                 return -1;
 479         }
 480         mbuf_dynflag->bitnum = bitnum;
 481         te->data = mbuf_dynflag;
 482
 483         TAILQ_INSERT_TAIL(mbuf_dynflag_list, te, next);
 484
 485         shm->free_flags &= ~(1ULL << bitnum);
 486
 487         RTE_LOG(DEBUG, MBUF, "Registered dynamic flag %s (fl=0x%x) -> %u\n",
 488                 params->name, params->flags, bitnum);
 489
 490         return bitnum;
 491 }
 492
 493 int
 494 rte_mbuf_dynflag_register_bitnum(const struct rte_mbuf_dynflag *params,
 495                                 unsigned int req)
 496 {
 497         int ret;
 498
 499         if (req >= RTE_SIZEOF_FIELD(struct rte_mbuf, ol_flags) * CHAR_BIT &&
 500                         req != UINT_MAX) {
 501                 rte_errno = EINVAL;
 502                 return -1;
 503         }
 504
 505         rte_mcfg_tailq_write_lock();
 506         ret = __rte_mbuf_dynflag_register_bitnum(params, req);
 507         rte_mcfg_tailq_write_unlock();
 508
 509         return ret;
 510 }
 511
 512 int
 513 rte_mbuf_dynflag_register(const struct rte_mbuf_dynflag *params)
 514 {
 515         return rte_mbuf_dynflag_register_bitnum(params, UINT_MAX);
 516 }
 517
 518 void rte_mbuf_dyn_dump(FILE *out)
 519 {
 520         struct mbuf_dynfield_list *mbuf_dynfield_list;
 521         struct mbuf_dynfield_elt *dynfield;
 522         struct mbuf_dynflag_list *mbuf_dynflag_list;
 523         struct mbuf_dynflag_elt *dynflag;
 524         struct rte_tailq_entry *te;
 525         size_t i;
 526
 527         rte_mcfg_tailq_write_lock();
 528         init_shared_mem();
 529         fprintf(out, "Reserved fields:\n");
 530         mbuf_dynfield_list = RTE_TAILQ_CAST(
 531                 mbuf_dynfield_tailq.head, mbuf_dynfield_list);
 532         TAILQ_FOREACH(te, mbuf_dynfield_list, next) {
 533                 dynfield = (struct mbuf_dynfield_elt *)te->data;
 534                 fprintf(out, "  name=%s offset=%zd size=%zd align=%zd flags=%x\n",
 535                         dynfield->params.name, dynfield->offset,
 536                         dynfield->params.size, dynfield->params.align,
 537                         dynfield->params.flags);
 538         }
 539         fprintf(out, "Reserved flags:\n");
 540         mbuf_dynflag_list = RTE_TAILQ_CAST(
 541                 mbuf_dynflag_tailq.head, mbuf_dynflag_list);
 542         TAILQ_FOREACH(te, mbuf_dynflag_list, next) {
 543                 dynflag = (struct mbuf_dynflag_elt *)te->data;
 544                 fprintf(out, "  name=%s bitnum=%u flags=%x\n",
 545                         dynflag->params.name, dynflag->bitnum,
 546                         dynflag->params.flags);
 547         }
 548         fprintf(out, "Free space in mbuf (0 = occupied, value = free zone alignment):\n");
 549         for (i = 0; i < sizeof(struct rte_mbuf); i++) {
 550                 if ((i % 8) == 0)
 551                         fprintf(out, "  %4.4zx: ", i);
 552                 fprintf(out, "%2.2x%s", shm->free_space[i],
 553                         (i % 8 != 7) ? " " : "\n");
 554         }
 555         fprintf(out, "Free bit in mbuf->ol_flags (0 = occupied, 1 = free):\n");
 556         for (i = 0; i < sizeof(uint64_t) * CHAR_BIT; i++) {
 557                 if ((i % 8) == 0)
 558                         fprintf(out, "  %4.4zx: ", i);
 559                 fprintf(out, "%1.1x%s", (shm->free_flags & (1ULL << i)) ? 1 : 0,
 560                         (i % 8 != 7) ? " " : "\n");
 561         }
 562
 563         rte_mcfg_tailq_write_unlock();
 564 }
 565
 566 static int
 567 rte_mbuf_dyn_timestamp_register(int *field_offset, uint64_t *flag,
 568                 const char *direction, const char *flag_name)
 569 {
 570         static const struct rte_mbuf_dynfield field_desc = {
 571                 .name = RTE_MBUF_DYNFIELD_TIMESTAMP_NAME,
 572                 .size = sizeof(rte_mbuf_timestamp_t),
 573                 .align = __alignof__(rte_mbuf_timestamp_t),
 574         };
 575         struct rte_mbuf_dynflag flag_desc = {};
 576         int offset;
 577
 578         offset = rte_mbuf_dynfield_register(&field_desc);
 579         if (offset < 0) {
 580                 RTE_LOG(ERR, MBUF,
 581                         "Failed to register mbuf field for timestamp\n");
 582                 return -1;
 583         }
 584         if (field_offset != NULL)
 585                 *field_offset = offset;
 586
 587         strlcpy(flag_desc.name, flag_name, sizeof(flag_desc.name));
 588         offset = rte_mbuf_dynflag_register(&flag_desc);
 589         if (offset < 0) {
 590                 RTE_LOG(ERR, MBUF,
 591                         "Failed to register mbuf flag for %s timestamp\n",
 592                         direction);
 593                 return -1;
 594         }
 595         if (flag != NULL)
 596                 *flag = RTE_BIT64(offset);
 597
 598         return 0;
 599 }
 600
 601 int
 602 rte_mbuf_dyn_rx_timestamp_register(int *field_offset, uint64_t *rx_flag)
 603 {
 604         return rte_mbuf_dyn_timestamp_register(field_offset, rx_flag,
 605                         "Rx", RTE_MBUF_DYNFLAG_RX_TIMESTAMP_NAME);
 606 }
 607
 608 int
 609 rte_mbuf_dyn_tx_timestamp_register(int *field_offset, uint64_t *tx_flag)
 610 {
 611         return rte_mbuf_dyn_timestamp_register(field_offset, tx_flag,
 612                         "Tx", RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME);
 613 }