lib/librte_eal/common/eal_common_dynmem.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2014 Intel Corporation.
   3  * Copyright(c) 2013 6WIND S.A.
   4  */
   5
   6 #include <inttypes.h>
   7 #include <string.h>
   8
   9 #include <rte_log.h>
  10 #include <rte_string_fns.h>
  11
  12 #include "eal_internal_cfg.h"
  13 #include "eal_memalloc.h"
  14 #include "eal_memcfg.h"
  15 #include "eal_private.h"
  16
  17 /** @file Functions common to EALs that support dynamic memory allocation. */
  18
  19 int
  20 eal_dynmem_memseg_lists_init(void)
  21 {
  22         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
  23         struct memtype {
  24                 uint64_t page_sz;
  25                 int socket_id;
  26         } *memtypes = NULL;
  27         int i, hpi_idx, msl_idx, ret = -1; /* fail unless told to succeed */
  28         struct rte_memseg_list *msl;
  29         uint64_t max_mem, max_mem_per_type;
  30         unsigned int max_seglists_per_type;
  31         unsigned int n_memtypes, cur_type;
  32
  33         /* no-huge does not need this at all */
  34         if (internal_config.no_hugetlbfs)
  35                 return 0;
  36
  37         /*
  38          * figuring out amount of memory we're going to have is a long and very
  39          * involved process. the basic element we're operating with is a memory
  40          * type, defined as a combination of NUMA node ID and page size (so that
  41          * e.g. 2 sockets with 2 page sizes yield 4 memory types in total).
  42          *
  43          * deciding amount of memory going towards each memory type is a
  44          * balancing act between maximum segments per type, maximum memory per
  45          * type, and number of detected NUMA nodes. the goal is to make sure
  46          * each memory type gets at least one memseg list.
  47          *
  48          * the total amount of memory is limited by RTE_MAX_MEM_MB value.
  49          *
  50          * the total amount of memory per type is limited by either
  51          * RTE_MAX_MEM_MB_PER_TYPE, or by RTE_MAX_MEM_MB divided by the number
  52          * of detected NUMA nodes. additionally, maximum number of segments per
  53          * type is also limited by RTE_MAX_MEMSEG_PER_TYPE. this is because for
  54          * smaller page sizes, it can take hundreds of thousands of segments to
  55          * reach the above specified per-type memory limits.
  56          *
  57          * additionally, each type may have multiple memseg lists associated
  58          * with it, each limited by either RTE_MAX_MEM_MB_PER_LIST for bigger
  59          * page sizes, or RTE_MAX_MEMSEG_PER_LIST segments for smaller ones.
  60          *
  61          * the number of memseg lists per type is decided based on the above
  62          * limits, and also taking number of detected NUMA nodes, to make sure
  63          * that we don't run out of memseg lists before we populate all NUMA
  64          * nodes with memory.
  65          *
  66          * we do this in three stages. first, we collect the number of types.
  67          * then, we figure out memory constraints and populate the list of
  68          * would-be memseg lists. then, we go ahead and allocate the memseg
  69          * lists.
  70          */
  71
  72         /* create space for mem types */
  73         n_memtypes = internal_config.num_hugepage_sizes * rte_socket_count();
  74         memtypes = calloc(n_memtypes, sizeof(*memtypes));
  75         if (memtypes == NULL) {
  76                 RTE_LOG(ERR, EAL, "Cannot allocate space for memory types\n");
  77                 return -1;
  78         }
  79
  80         /* populate mem types */
  81         cur_type = 0;
  82         for (hpi_idx = 0; hpi_idx < (int) internal_config.num_hugepage_sizes;
  83                         hpi_idx++) {
  84                 struct hugepage_info *hpi;
  85                 uint64_t hugepage_sz;
  86
  87                 hpi = &internal_config.hugepage_info[hpi_idx];
  88                 hugepage_sz = hpi->hugepage_sz;
  89
  90                 for (i = 0; i < (int) rte_socket_count(); i++, cur_type++) {
  91                         int socket_id = rte_socket_id_by_idx(i);
  92
  93 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
  94                         /* we can still sort pages by socket in legacy mode */
  95                         if (!internal_config.legacy_mem && socket_id > 0)
  96                                 break;
  97 #endif
  98                         memtypes[cur_type].page_sz = hugepage_sz;
  99                         memtypes[cur_type].socket_id = socket_id;
 100
 101                         RTE_LOG(DEBUG, EAL, "Detected memory type: "
 102                                 "socket_id:%u hugepage_sz:%" PRIu64 "\n",
 103                                 socket_id, hugepage_sz);
 104                 }
 105         }
 106         /* number of memtypes could have been lower due to no NUMA support */
 107         n_memtypes = cur_type;
 108
 109         /* set up limits for types */
 110         max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
 111         max_mem_per_type = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_TYPE << 20,
 112                         max_mem / n_memtypes);
 113         /*
 114          * limit maximum number of segment lists per type to ensure there's
 115          * space for memseg lists for all NUMA nodes with all page sizes
 116          */
 117         max_seglists_per_type = RTE_MAX_MEMSEG_LISTS / n_memtypes;
 118
 119         if (max_seglists_per_type == 0) {
 120                 RTE_LOG(ERR, EAL, "Cannot accommodate all memory types, please increase %s\n",
 121                         RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
 122                 goto out;
 123         }
 124
 125         /* go through all mem types and create segment lists */
 126         msl_idx = 0;
 127         for (cur_type = 0; cur_type < n_memtypes; cur_type++) {
 128                 unsigned int cur_seglist, n_seglists, n_segs;
 129                 unsigned int max_segs_per_type, max_segs_per_list;
 130                 struct memtype *type = &memtypes[cur_type];
 131                 uint64_t max_mem_per_list, pagesz;
 132                 int socket_id;
 133
 134                 pagesz = type->page_sz;
 135                 socket_id = type->socket_id;
 136
 137                 /*
 138                  * we need to create segment lists for this type. we must take
 139                  * into account the following things:
 140                  *
 141                  * 1. total amount of memory we can use for this memory type
 142                  * 2. total amount of memory per memseg list allowed
 143                  * 3. number of segments needed to fit the amount of memory
 144                  * 4. number of segments allowed per type
 145                  * 5. number of segments allowed per memseg list
 146                  * 6. number of memseg lists we are allowed to take up
 147                  */
 148
 149                 /* calculate how much segments we will need in total */
 150                 max_segs_per_type = max_mem_per_type / pagesz;
 151                 /* limit number of segments to maximum allowed per type */
 152                 max_segs_per_type = RTE_MIN(max_segs_per_type,
 153                                 (unsigned int)RTE_MAX_MEMSEG_PER_TYPE);
 154                 /* limit number of segments to maximum allowed per list */
 155                 max_segs_per_list = RTE_MIN(max_segs_per_type,
 156                                 (unsigned int)RTE_MAX_MEMSEG_PER_LIST);
 157
 158                 /* calculate how much memory we can have per segment list */
 159                 max_mem_per_list = RTE_MIN(max_segs_per_list * pagesz,
 160                                 (uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20);
 161
 162                 /* calculate how many segments each segment list will have */
 163                 n_segs = RTE_MIN(max_segs_per_list, max_mem_per_list / pagesz);
 164
 165                 /* calculate how many segment lists we can have */
 166                 n_seglists = RTE_MIN(max_segs_per_type / n_segs,
 167                                 max_mem_per_type / max_mem_per_list);
 168
 169                 /* limit number of segment lists according to our maximum */
 170                 n_seglists = RTE_MIN(n_seglists, max_seglists_per_type);
 171
 172                 RTE_LOG(DEBUG, EAL, "Creating %i segment lists: "
 173                                 "n_segs:%i socket_id:%i hugepage_sz:%" PRIu64 "\n",
 174                         n_seglists, n_segs, socket_id, pagesz);
 175
 176                 /* create all segment lists */
 177                 for (cur_seglist = 0; cur_seglist < n_seglists; cur_seglist++) {
 178                         if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
 179                                 RTE_LOG(ERR, EAL,
 180                                         "No more space in memseg lists, please increase %s\n",
 181                                         RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
 182                                 goto out;
 183                         }
 184                         msl = &mcfg->memsegs[msl_idx++];
 185
 186                         if (eal_memseg_list_init(msl, pagesz, n_segs,
 187                                         socket_id, cur_seglist, true))
 188                                 goto out;
 189
 190                         if (eal_memseg_list_alloc(msl, 0)) {
 191                                 RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list\n");
 192                                 goto out;
 193                         }
 194                 }
 195         }
 196         /* we're successful */
 197         ret = 0;
 198 out:
 199         free(memtypes);
 200         return ret;
 201 }
 202
 203 static int __rte_unused
 204 hugepage_count_walk(const struct rte_memseg_list *msl, void *arg)
 205 {
 206         struct hugepage_info *hpi = arg;
 207
 208         if (msl->page_sz != hpi->hugepage_sz)
 209                 return 0;
 210
 211         hpi->num_pages[msl->socket_id] += msl->memseg_arr.len;
 212         return 0;
 213 }
 214
 215 static int
 216 limits_callback(int socket_id, size_t cur_limit, size_t new_len)
 217 {
 218         RTE_SET_USED(socket_id);
 219         RTE_SET_USED(cur_limit);
 220         RTE_SET_USED(new_len);
 221         return -1;
 222 }
 223
 224 int
 225 eal_dynmem_hugepage_init(void)
 226 {
 227         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
 228         uint64_t memory[RTE_MAX_NUMA_NODES];
 229         int hp_sz_idx, socket_id;
 230
 231         memset(used_hp, 0, sizeof(used_hp));
 232
 233         for (hp_sz_idx = 0;
 234                         hp_sz_idx < (int) internal_config.num_hugepage_sizes;
 235                         hp_sz_idx++) {
 236 #ifndef RTE_ARCH_64
 237                 struct hugepage_info dummy;
 238                 unsigned int i;
 239 #endif
 240                 /* also initialize used_hp hugepage sizes in used_hp */
 241                 struct hugepage_info *hpi;
 242                 hpi = &internal_config.hugepage_info[hp_sz_idx];
 243                 used_hp[hp_sz_idx].hugepage_sz = hpi->hugepage_sz;
 244
 245 #ifndef RTE_ARCH_64
 246                 /* for 32-bit, limit number of pages on socket to whatever we've
 247                  * preallocated, as we cannot allocate more.
 248                  */
 249                 memset(&dummy, 0, sizeof(dummy));
 250                 dummy.hugepage_sz = hpi->hugepage_sz;
 251                 if (rte_memseg_list_walk(hugepage_count_walk, &dummy) < 0)
 252                         return -1;
 253
 254                 for (i = 0; i < RTE_DIM(dummy.num_pages); i++) {
 255                         hpi->num_pages[i] = RTE_MIN(hpi->num_pages[i],
 256                                         dummy.num_pages[i]);
 257                 }
 258 #endif
 259         }
 260
 261         /* make a copy of socket_mem, needed for balanced allocation. */
 262         for (hp_sz_idx = 0; hp_sz_idx < RTE_MAX_NUMA_NODES; hp_sz_idx++)
 263                 memory[hp_sz_idx] = internal_config.socket_mem[hp_sz_idx];
 264
 265         /* calculate final number of pages */
 266         if (eal_dynmem_calc_num_pages_per_socket(memory,
 267                         internal_config.hugepage_info, used_hp,
 268                         internal_config.num_hugepage_sizes) < 0)
 269                 return -1;
 270
 271         for (hp_sz_idx = 0;
 272                         hp_sz_idx < (int)internal_config.num_hugepage_sizes;
 273                         hp_sz_idx++) {
 274                 for (socket_id = 0; socket_id < RTE_MAX_NUMA_NODES;
 275                                 socket_id++) {
 276                         struct rte_memseg **pages;
 277                         struct hugepage_info *hpi = &used_hp[hp_sz_idx];
 278                         unsigned int num_pages = hpi->num_pages[socket_id];
 279                         unsigned int num_pages_alloc;
 280
 281                         if (num_pages == 0)
 282                                 continue;
 283
 284                         RTE_LOG(DEBUG, EAL,
 285                                 "Allocating %u pages of size %" PRIu64 "M "
 286                                 "on socket %i\n",
 287                                 num_pages, hpi->hugepage_sz >> 20, socket_id);
 288
 289                         /* we may not be able to allocate all pages in one go,
 290                          * because we break up our memory map into multiple
 291                          * memseg lists. therefore, try allocating multiple
 292                          * times and see if we can get the desired number of
 293                          * pages from multiple allocations.
 294                          */
 295
 296                         num_pages_alloc = 0;
 297                         do {
 298                                 int i, cur_pages, needed;
 299
 300                                 needed = num_pages - num_pages_alloc;
 301
 302                                 pages = malloc(sizeof(*pages) * needed);
 303
 304                                 /* do not request exact number of pages */
 305                                 cur_pages = eal_memalloc_alloc_seg_bulk(pages,
 306                                                 needed, hpi->hugepage_sz,
 307                                                 socket_id, false);
 308                                 if (cur_pages <= 0) {
 309                                         free(pages);
 310                                         return -1;
 311                                 }
 312
 313                                 /* mark preallocated pages as unfreeable */
 314                                 for (i = 0; i < cur_pages; i++) {
 315                                         struct rte_memseg *ms = pages[i];
 316                                         ms->flags |=
 317                                                 RTE_MEMSEG_FLAG_DO_NOT_FREE;
 318                                 }
 319                                 free(pages);
 320
 321                                 num_pages_alloc += cur_pages;
 322                         } while (num_pages_alloc != num_pages);
 323                 }
 324         }
 325
 326         /* if socket limits were specified, set them */
 327         if (internal_config.force_socket_limits) {
 328                 unsigned int i;
 329                 for (i = 0; i < RTE_MAX_NUMA_NODES; i++) {
 330                         uint64_t limit = internal_config.socket_limit[i];
 331                         if (limit == 0)
 332                                 continue;
 333                         if (rte_mem_alloc_validator_register("socket-limit",
 334                                         limits_callback, i, limit))
 335                                 RTE_LOG(ERR, EAL, "Failed to register socket limits validator callback\n");
 336                 }
 337         }
 338         return 0;
 339 }
 340
 341 __rte_unused /* function is unused on 32-bit builds */
 342 static inline uint64_t
 343 get_socket_mem_size(int socket)
 344 {
 345         uint64_t size = 0;
 346         unsigned int i;
 347
 348         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 349                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 350                 size += hpi->hugepage_sz * hpi->num_pages[socket];
 351         }
 352
 353         return size;
 354 }
 355
 356 int
 357 eal_dynmem_calc_num_pages_per_socket(
 358         uint64_t *memory, struct hugepage_info *hp_info,
 359         struct hugepage_info *hp_used, unsigned int num_hp_info)
 360 {
 361         unsigned int socket, j, i = 0;
 362         unsigned int requested, available;
 363         int total_num_pages = 0;
 364         uint64_t remaining_mem, cur_mem;
 365         uint64_t total_mem = internal_config.memory;
 366
 367         if (num_hp_info == 0)
 368                 return -1;
 369
 370         /* if specific memory amounts per socket weren't requested */
 371         if (internal_config.force_sockets == 0) {
 372                 size_t total_size;
 373 #ifdef RTE_ARCH_64
 374                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
 375                 size_t default_size;
 376                 unsigned int lcore_id;
 377
 378                 /* Compute number of cores per socket */
 379                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
 380                 RTE_LCORE_FOREACH(lcore_id) {
 381                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
 382                 }
 383
 384                 /*
 385                  * Automatically spread requested memory amongst detected
 386                  * sockets according to number of cores from CPU mask present
 387                  * on each socket.
 388                  */
 389                 total_size = internal_config.memory;
 390                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0;
 391                                 socket++) {
 392
 393                         /* Set memory amount per socket */
 394                         default_size = internal_config.memory *
 395                                 cpu_per_socket[socket] / rte_lcore_count();
 396
 397                         /* Limit to maximum available memory on socket */
 398                         default_size = RTE_MIN(
 399                                 default_size, get_socket_mem_size(socket));
 400
 401                         /* Update sizes */
 402                         memory[socket] = default_size;
 403                         total_size -= default_size;
 404                 }
 405
 406                 /*
 407                  * If some memory is remaining, try to allocate it by getting
 408                  * all available memory from sockets, one after the other.
 409                  */
 410                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0;
 411                                 socket++) {
 412                         /* take whatever is available */
 413                         default_size = RTE_MIN(
 414                                 get_socket_mem_size(socket) - memory[socket],
 415                                 total_size);
 416
 417                         /* Update sizes */
 418                         memory[socket] += default_size;
 419                         total_size -= default_size;
 420                 }
 421 #else
 422                 /* in 32-bit mode, allocate all of the memory only on master
 423                  * lcore socket
 424                  */
 425                 total_size = internal_config.memory;
 426                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0;
 427                                 socket++) {
 428                         struct rte_config *cfg = rte_eal_get_configuration();
 429                         unsigned int master_lcore_socket;
 430
 431                         master_lcore_socket =
 432                                 rte_lcore_to_socket_id(cfg->master_lcore);
 433
 434                         if (master_lcore_socket != socket)
 435                                 continue;
 436
 437                         /* Update sizes */
 438                         memory[socket] = total_size;
 439                         break;
 440                 }
 441 #endif
 442         }
 443
 444         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0;
 445                         socket++) {
 446                 /* skips if the memory on specific socket wasn't requested */
 447                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++) {
 448                         rte_strscpy(hp_used[i].hugedir, hp_info[i].hugedir,
 449                                 sizeof(hp_used[i].hugedir));
 450                         hp_used[i].num_pages[socket] = RTE_MIN(
 451                                         memory[socket] / hp_info[i].hugepage_sz,
 452                                         hp_info[i].num_pages[socket]);
 453
 454                         cur_mem = hp_used[i].num_pages[socket] *
 455                                         hp_used[i].hugepage_sz;
 456
 457                         memory[socket] -= cur_mem;
 458                         total_mem -= cur_mem;
 459
 460                         total_num_pages += hp_used[i].num_pages[socket];
 461
 462                         /* check if we have met all memory requests */
 463                         if (memory[socket] == 0)
 464                                 break;
 465
 466                         /* Check if we have any more pages left at this size,
 467                          * if so, move on to next size.
 468                          */
 469                         if (hp_used[i].num_pages[socket] ==
 470                                         hp_info[i].num_pages[socket])
 471                                 continue;
 472                         /* At this point we know that there are more pages
 473                          * available that are bigger than the memory we want,
 474                          * so lets see if we can get enough from other page
 475                          * sizes.
 476                          */
 477                         remaining_mem = 0;
 478                         for (j = i+1; j < num_hp_info; j++)
 479                                 remaining_mem += hp_info[j].hugepage_sz *
 480                                 hp_info[j].num_pages[socket];
 481
 482                         /* Is there enough other memory?
 483                          * If not, allocate another page and quit.
 484                          */
 485                         if (remaining_mem < memory[socket]) {
 486                                 cur_mem = RTE_MIN(
 487                                         memory[socket], hp_info[i].hugepage_sz);
 488                                 memory[socket] -= cur_mem;
 489                                 total_mem -= cur_mem;
 490                                 hp_used[i].num_pages[socket]++;
 491                                 total_num_pages++;
 492                                 break; /* we are done with this socket*/
 493                         }
 494                 }
 495
 496                 /* if we didn't satisfy all memory requirements per socket */
 497                 if (memory[socket] > 0 &&
 498                                 internal_config.socket_mem[socket] != 0) {
 499                         /* to prevent icc errors */
 500                         requested = (unsigned int)(
 501                                 internal_config.socket_mem[socket] / 0x100000);
 502                         available = requested -
 503                                 ((unsigned int)(memory[socket] / 0x100000));
 504                         RTE_LOG(ERR, EAL, "Not enough memory available on "
 505                                 "socket %u! Requested: %uMB, available: %uMB\n",
 506                                 socket, requested, available);
 507                         return -1;
 508                 }
 509         }
 510
 511         /* if we didn't satisfy total memory requirements */
 512         if (total_mem > 0) {
 513                 requested = (unsigned int)(internal_config.memory / 0x100000);
 514                 available = requested - (unsigned int)(total_mem / 0x100000);
 515                 RTE_LOG(ERR, EAL, "Not enough memory available! "
 516                         "Requested: %uMB, available: %uMB\n",
 517                         requested, available);
 518                 return -1;
 519         }
 520         return total_num_pages;
 521 }