lib/librte_eal/freebsd/eal/eal_memory.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2014 Intel Corporation
   3  */
   4 #include <sys/mman.h>
   5 #include <unistd.h>
   6 #include <sys/types.h>
   7 #include <sys/sysctl.h>
   8 #include <inttypes.h>
   9 #include <errno.h>
  10 #include <string.h>
  11 #include <fcntl.h>
  12
  13 #include <rte_eal.h>
  14 #include <rte_eal_memconfig.h>
  15 #include <rte_errno.h>
  16 #include <rte_log.h>
  17 #include <rte_string_fns.h>
  18 #include "eal_private.h"
  19 #include "eal_internal_cfg.h"
  20 #include "eal_filesystem.h"
  21 #include "eal_memcfg.h"
  22
  23 #define EAL_PAGE_SIZE (sysconf(_SC_PAGESIZE))
  24
  25 /*
  26  * Get physical address of any mapped virtual address in the current process.
  27  */
  28 phys_addr_t
  29 rte_mem_virt2phy(const void *virtaddr)
  30 {
  31         /* XXX not implemented. This function is only used by
  32          * rte_mempool_virt2iova() when hugepages are disabled. */
  33         (void)virtaddr;
  34         return RTE_BAD_IOVA;
  35 }
  36 rte_iova_t
  37 rte_mem_virt2iova(const void *virtaddr)
  38 {
  39         return rte_mem_virt2phy(virtaddr);
  40 }
  41
  42 int
  43 rte_eal_hugepage_init(void)
  44 {
  45         struct rte_mem_config *mcfg;
  46         uint64_t total_mem = 0;
  47         void *addr;
  48         unsigned int i, j, seg_idx = 0;
  49
  50         /* get pointer to global configuration */
  51         mcfg = rte_eal_get_configuration()->mem_config;
  52
  53         /* for debug purposes, hugetlbfs can be disabled */
  54         if (internal_config.no_hugetlbfs) {
  55                 struct rte_memseg_list *msl;
  56                 struct rte_fbarray *arr;
  57                 struct rte_memseg *ms;
  58                 uint64_t page_sz;
  59                 int n_segs, cur_seg;
  60
  61                 /* create a memseg list */
  62                 msl = &mcfg->memsegs[0];
  63
  64                 page_sz = RTE_PGSIZE_4K;
  65                 n_segs = internal_config.memory / page_sz;
  66
  67                 if (rte_fbarray_init(&msl->memseg_arr, "nohugemem", n_segs,
  68                                 sizeof(struct rte_memseg))) {
  69                         RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
  70                         return -1;
  71                 }
  72
  73                 addr = mmap(NULL, internal_config.memory,
  74                                 PROT_READ | PROT_WRITE,
  75                                 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
  76                 if (addr == MAP_FAILED) {
  77                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
  78                                         strerror(errno));
  79                         return -1;
  80                 }
  81                 msl->base_va = addr;
  82                 msl->page_sz = page_sz;
  83                 msl->len = internal_config.memory;
  84                 msl->socket_id = 0;
  85
  86                 /* populate memsegs. each memseg is 1 page long */
  87                 for (cur_seg = 0; cur_seg < n_segs; cur_seg++) {
  88                         arr = &msl->memseg_arr;
  89
  90                         ms = rte_fbarray_get(arr, cur_seg);
  91                         if (rte_eal_iova_mode() == RTE_IOVA_VA)
  92                                 ms->iova = (uintptr_t)addr;
  93                         else
  94                                 ms->iova = RTE_BAD_IOVA;
  95                         ms->addr = addr;
  96                         ms->hugepage_sz = page_sz;
  97                         ms->len = page_sz;
  98                         ms->socket_id = 0;
  99
 100                         rte_fbarray_set_used(arr, cur_seg);
 101
 102                         addr = RTE_PTR_ADD(addr, page_sz);
 103                 }
 104                 return 0;
 105         }
 106
 107         /* map all hugepages and sort them */
 108         for (i = 0; i < internal_config.num_hugepage_sizes; i ++){
 109                 struct hugepage_info *hpi;
 110                 rte_iova_t prev_end = 0;
 111                 int prev_ms_idx = -1;
 112                 uint64_t page_sz, mem_needed;
 113                 unsigned int n_pages, max_pages;
 114
 115                 hpi = &internal_config.hugepage_info[i];
 116                 page_sz = hpi->hugepage_sz;
 117                 max_pages = hpi->num_pages[0];
 118                 mem_needed = RTE_ALIGN_CEIL(internal_config.memory - total_mem,
 119                                 page_sz);
 120
 121                 n_pages = RTE_MIN(mem_needed / page_sz, max_pages);
 122
 123                 for (j = 0; j < n_pages; j++) {
 124                         struct rte_memseg_list *msl;
 125                         struct rte_fbarray *arr;
 126                         struct rte_memseg *seg;
 127                         int msl_idx, ms_idx;
 128                         rte_iova_t physaddr;
 129                         int error;
 130                         size_t sysctl_size = sizeof(physaddr);
 131                         char physaddr_str[64];
 132                         bool is_adjacent;
 133
 134                         /* first, check if this segment is IOVA-adjacent to
 135                          * the previous one.
 136                          */
 137                         snprintf(physaddr_str, sizeof(physaddr_str),
 138                                         "hw.contigmem.physaddr.%d", j);
 139                         error = sysctlbyname(physaddr_str, &physaddr,
 140                                         &sysctl_size, NULL, 0);
 141                         if (error < 0) {
 142                                 RTE_LOG(ERR, EAL, "Failed to get physical addr for buffer %u "
 143                                                 "from %s\n", j, hpi->hugedir);
 144                                 return -1;
 145                         }
 146
 147                         is_adjacent = prev_end != 0 && physaddr == prev_end;
 148                         prev_end = physaddr + hpi->hugepage_sz;
 149
 150                         for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS;
 151                                         msl_idx++) {
 152                                 bool empty, need_hole;
 153                                 msl = &mcfg->memsegs[msl_idx];
 154                                 arr = &msl->memseg_arr;
 155
 156                                 if (msl->page_sz != page_sz)
 157                                         continue;
 158
 159                                 empty = arr->count == 0;
 160
 161                                 /* we need a hole if this isn't an empty memseg
 162                                  * list, and if previous segment was not
 163                                  * adjacent to current one.
 164                                  */
 165                                 need_hole = !empty && !is_adjacent;
 166
 167                                 /* we need 1, plus hole if not adjacent */
 168                                 ms_idx = rte_fbarray_find_next_n_free(arr,
 169                                                 0, 1 + (need_hole ? 1 : 0));
 170
 171                                 /* memseg list is full? */
 172                                 if (ms_idx < 0)
 173                                         continue;
 174
 175                                 if (need_hole && prev_ms_idx == ms_idx - 1)
 176                                         ms_idx++;
 177                                 prev_ms_idx = ms_idx;
 178
 179                                 break;
 180                         }
 181                         if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
 182                                 RTE_LOG(ERR, EAL, "Could not find space for memseg. Please increase %s and/or %s in configuration.\n",
 183                                         RTE_STR(CONFIG_RTE_MAX_MEMSEG_PER_TYPE),
 184                                         RTE_STR(CONFIG_RTE_MAX_MEM_PER_TYPE));
 185                                 return -1;
 186                         }
 187                         arr = &msl->memseg_arr;
 188                         seg = rte_fbarray_get(arr, ms_idx);
 189
 190                         addr = RTE_PTR_ADD(msl->base_va,
 191                                         (size_t)msl->page_sz * ms_idx);
 192
 193                         /* address is already mapped in memseg list, so using
 194                          * MAP_FIXED here is safe.
 195                          */
 196                         addr = mmap(addr, page_sz, PROT_READ|PROT_WRITE,
 197                                         MAP_SHARED | MAP_FIXED,
 198                                         hpi->lock_descriptor,
 199                                         j * EAL_PAGE_SIZE);
 200                         if (addr == MAP_FAILED) {
 201                                 RTE_LOG(ERR, EAL, "Failed to mmap buffer %u from %s\n",
 202                                                 j, hpi->hugedir);
 203                                 return -1;
 204                         }
 205
 206                         seg->addr = addr;
 207                         seg->iova = physaddr;
 208                         seg->hugepage_sz = page_sz;
 209                         seg->len = page_sz;
 210                         seg->nchannel = mcfg->nchannel;
 211                         seg->nrank = mcfg->nrank;
 212                         seg->socket_id = 0;
 213
 214                         rte_fbarray_set_used(arr, ms_idx);
 215
 216                         RTE_LOG(INFO, EAL, "Mapped memory segment %u @ %p: physaddr:0x%"
 217                                         PRIx64", len %zu\n",
 218                                         seg_idx++, addr, physaddr, page_sz);
 219
 220                         total_mem += seg->len;
 221                 }
 222                 if (total_mem >= internal_config.memory)
 223                         break;
 224         }
 225         if (total_mem < internal_config.memory) {
 226                 RTE_LOG(ERR, EAL, "Couldn't reserve requested memory, "
 227                                 "requested: %" PRIu64 "M "
 228                                 "available: %" PRIu64 "M\n",
 229                                 internal_config.memory >> 20, total_mem >> 20);
 230                 return -1;
 231         }
 232         return 0;
 233 }
 234
 235 struct attach_walk_args {
 236         int fd_hugepage;
 237         int seg_idx;
 238 };
 239 static int
 240 attach_segment(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
 241                 void *arg)
 242 {
 243         struct attach_walk_args *wa = arg;
 244         void *addr;
 245
 246         if (msl->external)
 247                 return 0;
 248
 249         addr = mmap(ms->addr, ms->len, PROT_READ | PROT_WRITE,
 250                         MAP_SHARED | MAP_FIXED, wa->fd_hugepage,
 251                         wa->seg_idx * EAL_PAGE_SIZE);
 252         if (addr == MAP_FAILED || addr != ms->addr)
 253                 return -1;
 254         wa->seg_idx++;
 255
 256         return 0;
 257 }
 258
 259 int
 260 rte_eal_hugepage_attach(void)
 261 {
 262         const struct hugepage_info *hpi;
 263         int fd_hugepage = -1;
 264         unsigned int i;
 265
 266         hpi = &internal_config.hugepage_info[0];
 267
 268         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 269                 const struct hugepage_info *cur_hpi = &hpi[i];
 270                 struct attach_walk_args wa;
 271
 272                 memset(&wa, 0, sizeof(wa));
 273
 274                 /* Obtain a file descriptor for contiguous memory */
 275                 fd_hugepage = open(cur_hpi->hugedir, O_RDWR);
 276                 if (fd_hugepage < 0) {
 277                         RTE_LOG(ERR, EAL, "Could not open %s\n",
 278                                         cur_hpi->hugedir);
 279                         goto error;
 280                 }
 281                 wa.fd_hugepage = fd_hugepage;
 282                 wa.seg_idx = 0;
 283
 284                 /* Map the contiguous memory into each memory segment */
 285                 if (rte_memseg_walk(attach_segment, &wa) < 0) {
 286                         RTE_LOG(ERR, EAL, "Failed to mmap buffer %u from %s\n",
 287                                 wa.seg_idx, cur_hpi->hugedir);
 288                         goto error;
 289                 }
 290
 291                 close(fd_hugepage);
 292                 fd_hugepage = -1;
 293         }
 294
 295         /* hugepage_info is no longer required */
 296         return 0;
 297
 298 error:
 299         if (fd_hugepage >= 0)
 300                 close(fd_hugepage);
 301         return -1;
 302 }
 303
 304 int
 305 rte_eal_using_phys_addrs(void)
 306 {
 307         return 0;
 308 }
 309
 310 static uint64_t
 311 get_mem_amount(uint64_t page_sz, uint64_t max_mem)
 312 {
 313         uint64_t area_sz, max_pages;
 314
 315         /* limit to RTE_MAX_MEMSEG_PER_LIST pages or RTE_MAX_MEM_MB_PER_LIST */
 316         max_pages = RTE_MAX_MEMSEG_PER_LIST;
 317         max_mem = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20, max_mem);
 318
 319         area_sz = RTE_MIN(page_sz * max_pages, max_mem);
 320
 321         /* make sure the list isn't smaller than the page size */
 322         area_sz = RTE_MAX(area_sz, page_sz);
 323
 324         return RTE_ALIGN(area_sz, page_sz);
 325 }
 326
 327 #define MEMSEG_LIST_FMT "memseg-%" PRIu64 "k-%i-%i"
 328 static int
 329 alloc_memseg_list(struct rte_memseg_list *msl, uint64_t page_sz,
 330                 int n_segs, int socket_id, int type_msl_idx)
 331 {
 332         char name[RTE_FBARRAY_NAME_LEN];
 333
 334         snprintf(name, sizeof(name), MEMSEG_LIST_FMT, page_sz >> 10, socket_id,
 335                  type_msl_idx);
 336         if (rte_fbarray_init(&msl->memseg_arr, name, n_segs,
 337                         sizeof(struct rte_memseg))) {
 338                 RTE_LOG(ERR, EAL, "Cannot allocate memseg list: %s\n",
 339                         rte_strerror(rte_errno));
 340                 return -1;
 341         }
 342
 343         msl->page_sz = page_sz;
 344         msl->socket_id = socket_id;
 345         msl->base_va = NULL;
 346
 347         RTE_LOG(DEBUG, EAL, "Memseg list allocated: 0x%zxkB at socket %i\n",
 348                         (size_t)page_sz >> 10, socket_id);
 349
 350         return 0;
 351 }
 352
 353 static int
 354 alloc_va_space(struct rte_memseg_list *msl)
 355 {
 356         uint64_t page_sz;
 357         size_t mem_sz;
 358         void *addr;
 359         int flags = 0;
 360
 361 #ifdef RTE_ARCH_PPC_64
 362         flags |= MAP_HUGETLB;
 363 #endif
 364
 365         page_sz = msl->page_sz;
 366         mem_sz = page_sz * msl->memseg_arr.len;
 367
 368         addr = eal_get_virtual_area(msl->base_va, &mem_sz, page_sz, 0, flags);
 369         if (addr == NULL) {
 370                 if (rte_errno == EADDRNOTAVAIL)
 371                         RTE_LOG(ERR, EAL, "Could not mmap %llu bytes at [%p] - please use '--base-virtaddr' option\n",
 372                                 (unsigned long long)mem_sz, msl->base_va);
 373                 else
 374                         RTE_LOG(ERR, EAL, "Cannot reserve memory\n");
 375                 return -1;
 376         }
 377         msl->base_va = addr;
 378         msl->len = mem_sz;
 379
 380         return 0;
 381 }
 382
 383
 384 static int
 385 memseg_primary_init(void)
 386 {
 387         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 388         int hpi_idx, msl_idx = 0;
 389         struct rte_memseg_list *msl;
 390         uint64_t max_mem, total_mem;
 391
 392         /* no-huge does not need this at all */
 393         if (internal_config.no_hugetlbfs)
 394                 return 0;
 395
 396         /* FreeBSD has an issue where core dump will dump the entire memory
 397          * contents, including anonymous zero-page memory. Therefore, while we
 398          * will be limiting total amount of memory to RTE_MAX_MEM_MB, we will
 399          * also be further limiting total memory amount to whatever memory is
 400          * available to us through contigmem driver (plus spacing blocks).
 401          *
 402          * so, at each stage, we will be checking how much memory we are
 403          * preallocating, and adjust all the values accordingly.
 404          */
 405
 406         max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
 407         total_mem = 0;
 408
 409         /* create memseg lists */
 410         for (hpi_idx = 0; hpi_idx < (int) internal_config.num_hugepage_sizes;
 411                         hpi_idx++) {
 412                 uint64_t max_type_mem, total_type_mem = 0;
 413                 uint64_t avail_mem;
 414                 int type_msl_idx, max_segs, avail_segs, total_segs = 0;
 415                 struct hugepage_info *hpi;
 416                 uint64_t hugepage_sz;
 417
 418                 hpi = &internal_config.hugepage_info[hpi_idx];
 419                 hugepage_sz = hpi->hugepage_sz;
 420
 421                 /* no NUMA support on FreeBSD */
 422
 423                 /* check if we've already exceeded total memory amount */
 424                 if (total_mem >= max_mem)
 425                         break;
 426
 427                 /* first, calculate theoretical limits according to config */
 428                 max_type_mem = RTE_MIN(max_mem - total_mem,
 429                         (uint64_t)RTE_MAX_MEM_MB_PER_TYPE << 20);
 430                 max_segs = RTE_MAX_MEMSEG_PER_TYPE;
 431
 432                 /* now, limit all of that to whatever will actually be
 433                  * available to us, because without dynamic allocation support,
 434                  * all of that extra memory will be sitting there being useless
 435                  * and slowing down core dumps in case of a crash.
 436                  *
 437                  * we need (N*2)-1 segments because we cannot guarantee that
 438                  * each segment will be IOVA-contiguous with the previous one,
 439                  * so we will allocate more and put spaces inbetween segments
 440                  * that are non-contiguous.
 441                  */
 442                 avail_segs = (hpi->num_pages[0] * 2) - 1;
 443                 avail_mem = avail_segs * hugepage_sz;
 444
 445                 max_type_mem = RTE_MIN(avail_mem, max_type_mem);
 446                 max_segs = RTE_MIN(avail_segs, max_segs);
 447
 448                 type_msl_idx = 0;
 449                 while (total_type_mem < max_type_mem &&
 450                                 total_segs < max_segs) {
 451                         uint64_t cur_max_mem, cur_mem;
 452                         unsigned int n_segs;
 453
 454                         if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
 455                                 RTE_LOG(ERR, EAL,
 456                                         "No more space in memseg lists, please increase %s\n",
 457                                         RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
 458                                 return -1;
 459                         }
 460
 461                         msl = &mcfg->memsegs[msl_idx++];
 462
 463                         cur_max_mem = max_type_mem - total_type_mem;
 464
 465                         cur_mem = get_mem_amount(hugepage_sz,
 466                                         cur_max_mem);
 467                         n_segs = cur_mem / hugepage_sz;
 468
 469                         if (alloc_memseg_list(msl, hugepage_sz, n_segs,
 470                                         0, type_msl_idx))
 471                                 return -1;
 472
 473                         total_segs += msl->memseg_arr.len;
 474                         total_type_mem = total_segs * hugepage_sz;
 475                         type_msl_idx++;
 476
 477                         if (alloc_va_space(msl)) {
 478                                 RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list\n");
 479                                 return -1;
 480                         }
 481                 }
 482                 total_mem += total_type_mem;
 483         }
 484         return 0;
 485 }
 486
 487 static int
 488 memseg_secondary_init(void)
 489 {
 490         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 491         int msl_idx = 0;
 492         struct rte_memseg_list *msl;
 493
 494         for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
 495
 496                 msl = &mcfg->memsegs[msl_idx];
 497
 498                 /* skip empty memseg lists */
 499                 if (msl->memseg_arr.len == 0)
 500                         continue;
 501
 502                 if (rte_fbarray_attach(&msl->memseg_arr)) {
 503                         RTE_LOG(ERR, EAL, "Cannot attach to primary process memseg lists\n");
 504                         return -1;
 505                 }
 506
 507                 /* preallocate VA space */
 508                 if (alloc_va_space(msl)) {
 509                         RTE_LOG(ERR, EAL, "Cannot preallocate VA space for hugepage memory\n");
 510                         return -1;
 511                 }
 512         }
 513
 514         return 0;
 515 }
 516
 517 int
 518 rte_eal_memseg_init(void)
 519 {
 520         return rte_eal_process_type() == RTE_PROC_PRIMARY ?
 521                         memseg_primary_init() :
 522                         memseg_secondary_init();
 523 }