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_errno.h>
  15 #include <rte_log.h>
  16 #include <rte_string_fns.h>
  17
  18 #include "eal_private.h"
  19 #include "eal_internal_cfg.h"
  20 #include "eal_filesystem.h"
  21 #include "eal_memcfg.h"
  22 #include "eal_options.h"
  23
  24 #define EAL_PAGE_SIZE (sysconf(_SC_PAGESIZE))
  25
  26 uint64_t eal_get_baseaddr(void)
  27 {
  28         /*
  29          * FreeBSD may allocate something in the space we will be mapping things
  30          * before we get a chance to do that, so use a base address that's far
  31          * away from where malloc() et al usually map things.
  32          */
  33         return 0x1000000000ULL;
  34 }
  35
  36 /*
  37  * Get physical address of any mapped virtual address in the current process.
  38  */
  39 phys_addr_t
  40 rte_mem_virt2phy(const void *virtaddr)
  41 {
  42         /* XXX not implemented. This function is only used by
  43          * rte_mempool_virt2iova() when hugepages are disabled. */
  44         (void)virtaddr;
  45         return RTE_BAD_IOVA;
  46 }
  47 rte_iova_t
  48 rte_mem_virt2iova(const void *virtaddr)
  49 {
  50         return rte_mem_virt2phy(virtaddr);
  51 }
  52
  53 int
  54 rte_eal_hugepage_init(void)
  55 {
  56         struct rte_mem_config *mcfg;
  57         uint64_t total_mem = 0;
  58         void *addr;
  59         unsigned int i, j, seg_idx = 0;
  60
  61         /* get pointer to global configuration */
  62         mcfg = rte_eal_get_configuration()->mem_config;
  63
  64         /* for debug purposes, hugetlbfs can be disabled */
  65         if (internal_config.no_hugetlbfs) {
  66                 struct rte_memseg_list *msl;
  67                 struct rte_fbarray *arr;
  68                 struct rte_memseg *ms;
  69                 uint64_t page_sz;
  70                 int n_segs, cur_seg;
  71
  72                 /* create a memseg list */
  73                 msl = &mcfg->memsegs[0];
  74
  75                 page_sz = RTE_PGSIZE_4K;
  76                 n_segs = internal_config.memory / page_sz;
  77
  78                 if (rte_fbarray_init(&msl->memseg_arr, "nohugemem", n_segs,
  79                                 sizeof(struct rte_memseg))) {
  80                         RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
  81                         return -1;
  82                 }
  83
  84                 addr = mmap(NULL, internal_config.memory,
  85                                 PROT_READ | PROT_WRITE,
  86                                 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
  87                 if (addr == MAP_FAILED) {
  88                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
  89                                         strerror(errno));
  90                         return -1;
  91                 }
  92                 msl->base_va = addr;
  93                 msl->page_sz = page_sz;
  94                 msl->len = internal_config.memory;
  95                 msl->socket_id = 0;
  96
  97                 /* populate memsegs. each memseg is 1 page long */
  98                 for (cur_seg = 0; cur_seg < n_segs; cur_seg++) {
  99                         arr = &msl->memseg_arr;
 100
 101                         ms = rte_fbarray_get(arr, cur_seg);
 102                         if (rte_eal_iova_mode() == RTE_IOVA_VA)
 103                                 ms->iova = (uintptr_t)addr;
 104                         else
 105                                 ms->iova = RTE_BAD_IOVA;
 106                         ms->addr = addr;
 107                         ms->hugepage_sz = page_sz;
 108                         ms->len = page_sz;
 109                         ms->socket_id = 0;
 110
 111                         rte_fbarray_set_used(arr, cur_seg);
 112
 113                         addr = RTE_PTR_ADD(addr, page_sz);
 114                 }
 115                 return 0;
 116         }
 117
 118         /* map all hugepages and sort them */
 119         for (i = 0; i < internal_config.num_hugepage_sizes; i ++){
 120                 struct hugepage_info *hpi;
 121                 rte_iova_t prev_end = 0;
 122                 int prev_ms_idx = -1;
 123                 uint64_t page_sz, mem_needed;
 124                 unsigned int n_pages, max_pages;
 125
 126                 hpi = &internal_config.hugepage_info[i];
 127                 page_sz = hpi->hugepage_sz;
 128                 max_pages = hpi->num_pages[0];
 129                 mem_needed = RTE_ALIGN_CEIL(internal_config.memory - total_mem,
 130                                 page_sz);
 131
 132                 n_pages = RTE_MIN(mem_needed / page_sz, max_pages);
 133
 134                 for (j = 0; j < n_pages; j++) {
 135                         struct rte_memseg_list *msl;
 136                         struct rte_fbarray *arr;
 137                         struct rte_memseg *seg;
 138                         int msl_idx, ms_idx;
 139                         rte_iova_t physaddr;
 140                         int error;
 141                         size_t sysctl_size = sizeof(physaddr);
 142                         char physaddr_str[64];
 143                         bool is_adjacent;
 144
 145                         /* first, check if this segment is IOVA-adjacent to
 146                          * the previous one.
 147                          */
 148                         snprintf(physaddr_str, sizeof(physaddr_str),
 149                                         "hw.contigmem.physaddr.%d", j);
 150                         error = sysctlbyname(physaddr_str, &physaddr,
 151                                         &sysctl_size, NULL, 0);
 152                         if (error < 0) {
 153                                 RTE_LOG(ERR, EAL, "Failed to get physical addr for buffer %u "
 154                                                 "from %s\n", j, hpi->hugedir);
 155                                 return -1;
 156                         }
 157
 158                         is_adjacent = prev_end != 0 && physaddr == prev_end;
 159                         prev_end = physaddr + hpi->hugepage_sz;
 160
 161                         for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS;
 162                                         msl_idx++) {
 163                                 bool empty, need_hole;
 164                                 msl = &mcfg->memsegs[msl_idx];
 165                                 arr = &msl->memseg_arr;
 166
 167                                 if (msl->page_sz != page_sz)
 168                                         continue;
 169
 170                                 empty = arr->count == 0;
 171
 172                                 /* we need a hole if this isn't an empty memseg
 173                                  * list, and if previous segment was not
 174                                  * adjacent to current one.
 175                                  */
 176                                 need_hole = !empty && !is_adjacent;
 177
 178                                 /* we need 1, plus hole if not adjacent */
 179                                 ms_idx = rte_fbarray_find_next_n_free(arr,
 180                                                 0, 1 + (need_hole ? 1 : 0));
 181
 182                                 /* memseg list is full? */
 183                                 if (ms_idx < 0)
 184                                         continue;
 185
 186                                 if (need_hole && prev_ms_idx == ms_idx - 1)
 187                                         ms_idx++;
 188                                 prev_ms_idx = ms_idx;
 189
 190                                 break;
 191                         }
 192                         if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
 193                                 RTE_LOG(ERR, EAL, "Could not find space for memseg. Please increase %s and/or %s in configuration.\n",
 194                                         RTE_STR(CONFIG_RTE_MAX_MEMSEG_PER_TYPE),
 195                                         RTE_STR(CONFIG_RTE_MAX_MEM_PER_TYPE));
 196                                 return -1;
 197                         }
 198                         arr = &msl->memseg_arr;
 199                         seg = rte_fbarray_get(arr, ms_idx);
 200
 201                         addr = RTE_PTR_ADD(msl->base_va,
 202                                         (size_t)msl->page_sz * ms_idx);
 203
 204                         /* address is already mapped in memseg list, so using
 205                          * MAP_FIXED here is safe.
 206                          */
 207                         addr = mmap(addr, page_sz, PROT_READ|PROT_WRITE,
 208                                         MAP_SHARED | MAP_FIXED,
 209                                         hpi->lock_descriptor,
 210                                         j * EAL_PAGE_SIZE);
 211                         if (addr == MAP_FAILED) {
 212                                 RTE_LOG(ERR, EAL, "Failed to mmap buffer %u from %s\n",
 213                                                 j, hpi->hugedir);
 214                                 return -1;
 215                         }
 216
 217                         seg->addr = addr;
 218                         seg->iova = physaddr;
 219                         seg->hugepage_sz = page_sz;
 220                         seg->len = page_sz;
 221                         seg->nchannel = mcfg->nchannel;
 222                         seg->nrank = mcfg->nrank;
 223                         seg->socket_id = 0;
 224
 225                         rte_fbarray_set_used(arr, ms_idx);
 226
 227                         RTE_LOG(INFO, EAL, "Mapped memory segment %u @ %p: physaddr:0x%"
 228                                         PRIx64", len %zu\n",
 229                                         seg_idx++, addr, physaddr, page_sz);
 230
 231                         total_mem += seg->len;
 232                 }
 233                 if (total_mem >= internal_config.memory)
 234                         break;
 235         }
 236         if (total_mem < internal_config.memory) {
 237                 RTE_LOG(ERR, EAL, "Couldn't reserve requested memory, "
 238                                 "requested: %" PRIu64 "M "
 239                                 "available: %" PRIu64 "M\n",
 240                                 internal_config.memory >> 20, total_mem >> 20);
 241                 return -1;
 242         }
 243         return 0;
 244 }
 245
 246 struct attach_walk_args {
 247         int fd_hugepage;
 248         int seg_idx;
 249 };
 250 static int
 251 attach_segment(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
 252                 void *arg)
 253 {
 254         struct attach_walk_args *wa = arg;
 255         void *addr;
 256
 257         if (msl->external)
 258                 return 0;
 259
 260         addr = mmap(ms->addr, ms->len, PROT_READ | PROT_WRITE,
 261                         MAP_SHARED | MAP_FIXED, wa->fd_hugepage,
 262                         wa->seg_idx * EAL_PAGE_SIZE);
 263         if (addr == MAP_FAILED || addr != ms->addr)
 264                 return -1;
 265         wa->seg_idx++;
 266
 267         return 0;
 268 }
 269
 270 int
 271 rte_eal_hugepage_attach(void)
 272 {
 273         const struct hugepage_info *hpi;
 274         int fd_hugepage = -1;
 275         unsigned int i;
 276
 277         hpi = &internal_config.hugepage_info[0];
 278
 279         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 280                 const struct hugepage_info *cur_hpi = &hpi[i];
 281                 struct attach_walk_args wa;
 282
 283                 memset(&wa, 0, sizeof(wa));
 284
 285                 /* Obtain a file descriptor for contiguous memory */
 286                 fd_hugepage = open(cur_hpi->hugedir, O_RDWR);
 287                 if (fd_hugepage < 0) {
 288                         RTE_LOG(ERR, EAL, "Could not open %s\n",
 289                                         cur_hpi->hugedir);
 290                         goto error;
 291                 }
 292                 wa.fd_hugepage = fd_hugepage;
 293                 wa.seg_idx = 0;
 294
 295                 /* Map the contiguous memory into each memory segment */
 296                 if (rte_memseg_walk(attach_segment, &wa) < 0) {
 297                         RTE_LOG(ERR, EAL, "Failed to mmap buffer %u from %s\n",
 298                                 wa.seg_idx, cur_hpi->hugedir);
 299                         goto error;
 300                 }
 301
 302                 close(fd_hugepage);
 303                 fd_hugepage = -1;
 304         }
 305
 306         /* hugepage_info is no longer required */
 307         return 0;
 308
 309 error:
 310         if (fd_hugepage >= 0)
 311                 close(fd_hugepage);
 312         return -1;
 313 }
 314
 315 int
 316 rte_eal_using_phys_addrs(void)
 317 {
 318         return 0;
 319 }
 320
 321 static uint64_t
 322 get_mem_amount(uint64_t page_sz, uint64_t max_mem)
 323 {
 324         uint64_t area_sz, max_pages;
 325
 326         /* limit to RTE_MAX_MEMSEG_PER_LIST pages or RTE_MAX_MEM_MB_PER_LIST */
 327         max_pages = RTE_MAX_MEMSEG_PER_LIST;
 328         max_mem = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20, max_mem);
 329
 330         area_sz = RTE_MIN(page_sz * max_pages, max_mem);
 331
 332         /* make sure the list isn't smaller than the page size */
 333         area_sz = RTE_MAX(area_sz, page_sz);
 334
 335         return RTE_ALIGN(area_sz, page_sz);
 336 }
 337
 338 #define MEMSEG_LIST_FMT "memseg-%" PRIu64 "k-%i-%i"
 339 static int
 340 alloc_memseg_list(struct rte_memseg_list *msl, uint64_t page_sz,
 341                 int n_segs, int socket_id, int type_msl_idx)
 342 {
 343         char name[RTE_FBARRAY_NAME_LEN];
 344
 345         snprintf(name, sizeof(name), MEMSEG_LIST_FMT, page_sz >> 10, socket_id,
 346                  type_msl_idx);
 347         if (rte_fbarray_init(&msl->memseg_arr, name, n_segs,
 348                         sizeof(struct rte_memseg))) {
 349                 RTE_LOG(ERR, EAL, "Cannot allocate memseg list: %s\n",
 350                         rte_strerror(rte_errno));
 351                 return -1;
 352         }
 353
 354         msl->page_sz = page_sz;
 355         msl->socket_id = socket_id;
 356         msl->base_va = NULL;
 357
 358         RTE_LOG(DEBUG, EAL, "Memseg list allocated: 0x%zxkB at socket %i\n",
 359                         (size_t)page_sz >> 10, socket_id);
 360
 361         return 0;
 362 }
 363
 364 static int
 365 alloc_va_space(struct rte_memseg_list *msl)
 366 {
 367         uint64_t page_sz;
 368         size_t mem_sz;
 369         void *addr;
 370         int flags = 0;
 371
 372 #ifdef RTE_ARCH_PPC_64
 373         flags |= MAP_HUGETLB;
 374 #endif
 375
 376         page_sz = msl->page_sz;
 377         mem_sz = page_sz * msl->memseg_arr.len;
 378
 379         addr = eal_get_virtual_area(msl->base_va, &mem_sz, page_sz, 0, flags);
 380         if (addr == NULL) {
 381                 if (rte_errno == EADDRNOTAVAIL)
 382                         RTE_LOG(ERR, EAL, "Could not mmap %llu bytes at [%p] - "
 383                                 "please use '--" OPT_BASE_VIRTADDR "' option\n",
 384                                 (unsigned long long)mem_sz, msl->base_va);
 385                 else
 386                         RTE_LOG(ERR, EAL, "Cannot reserve memory\n");
 387                 return -1;
 388         }
 389         msl->base_va = addr;
 390         msl->len = mem_sz;
 391
 392         return 0;
 393 }
 394
 395
 396 static int
 397 memseg_primary_init(void)
 398 {
 399         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 400         int hpi_idx, msl_idx = 0;
 401         struct rte_memseg_list *msl;
 402         uint64_t max_mem, total_mem;
 403
 404         /* no-huge does not need this at all */
 405         if (internal_config.no_hugetlbfs)
 406                 return 0;
 407
 408         /* FreeBSD has an issue where core dump will dump the entire memory
 409          * contents, including anonymous zero-page memory. Therefore, while we
 410          * will be limiting total amount of memory to RTE_MAX_MEM_MB, we will
 411          * also be further limiting total memory amount to whatever memory is
 412          * available to us through contigmem driver (plus spacing blocks).
 413          *
 414          * so, at each stage, we will be checking how much memory we are
 415          * preallocating, and adjust all the values accordingly.
 416          */
 417
 418         max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
 419         total_mem = 0;
 420
 421         /* create memseg lists */
 422         for (hpi_idx = 0; hpi_idx < (int) internal_config.num_hugepage_sizes;
 423                         hpi_idx++) {
 424                 uint64_t max_type_mem, total_type_mem = 0;
 425                 uint64_t avail_mem;
 426                 int type_msl_idx, max_segs, avail_segs, total_segs = 0;
 427                 struct hugepage_info *hpi;
 428                 uint64_t hugepage_sz;
 429
 430                 hpi = &internal_config.hugepage_info[hpi_idx];
 431                 hugepage_sz = hpi->hugepage_sz;
 432
 433                 /* no NUMA support on FreeBSD */
 434
 435                 /* check if we've already exceeded total memory amount */
 436                 if (total_mem >= max_mem)
 437                         break;
 438
 439                 /* first, calculate theoretical limits according to config */
 440                 max_type_mem = RTE_MIN(max_mem - total_mem,
 441                         (uint64_t)RTE_MAX_MEM_MB_PER_TYPE << 20);
 442                 max_segs = RTE_MAX_MEMSEG_PER_TYPE;
 443
 444                 /* now, limit all of that to whatever will actually be
 445                  * available to us, because without dynamic allocation support,
 446                  * all of that extra memory will be sitting there being useless
 447                  * and slowing down core dumps in case of a crash.
 448                  *
 449                  * we need (N*2)-1 segments because we cannot guarantee that
 450                  * each segment will be IOVA-contiguous with the previous one,
 451                  * so we will allocate more and put spaces inbetween segments
 452                  * that are non-contiguous.
 453                  */
 454                 avail_segs = (hpi->num_pages[0] * 2) - 1;
 455                 avail_mem = avail_segs * hugepage_sz;
 456
 457                 max_type_mem = RTE_MIN(avail_mem, max_type_mem);
 458                 max_segs = RTE_MIN(avail_segs, max_segs);
 459
 460                 type_msl_idx = 0;
 461                 while (total_type_mem < max_type_mem &&
 462                                 total_segs < max_segs) {
 463                         uint64_t cur_max_mem, cur_mem;
 464                         unsigned int n_segs;
 465
 466                         if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
 467                                 RTE_LOG(ERR, EAL,
 468                                         "No more space in memseg lists, please increase %s\n",
 469                                         RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
 470                                 return -1;
 471                         }
 472
 473                         msl = &mcfg->memsegs[msl_idx++];
 474
 475                         cur_max_mem = max_type_mem - total_type_mem;
 476
 477                         cur_mem = get_mem_amount(hugepage_sz,
 478                                         cur_max_mem);
 479                         n_segs = cur_mem / hugepage_sz;
 480
 481                         if (alloc_memseg_list(msl, hugepage_sz, n_segs,
 482                                         0, type_msl_idx))
 483                                 return -1;
 484
 485                         total_segs += msl->memseg_arr.len;
 486                         total_type_mem = total_segs * hugepage_sz;
 487                         type_msl_idx++;
 488
 489                         if (alloc_va_space(msl)) {
 490                                 RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list\n");
 491                                 return -1;
 492                         }
 493                 }
 494                 total_mem += total_type_mem;
 495         }
 496         return 0;
 497 }
 498
 499 static int
 500 memseg_secondary_init(void)
 501 {
 502         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 503         int msl_idx = 0;
 504         struct rte_memseg_list *msl;
 505
 506         for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
 507
 508                 msl = &mcfg->memsegs[msl_idx];
 509
 510                 /* skip empty memseg lists */
 511                 if (msl->memseg_arr.len == 0)
 512                         continue;
 513
 514                 if (rte_fbarray_attach(&msl->memseg_arr)) {
 515                         RTE_LOG(ERR, EAL, "Cannot attach to primary process memseg lists\n");
 516                         return -1;
 517                 }
 518
 519                 /* preallocate VA space */
 520                 if (alloc_va_space(msl)) {
 521                         RTE_LOG(ERR, EAL, "Cannot preallocate VA space for hugepage memory\n");
 522                         return -1;
 523                 }
 524         }
 525
 526         return 0;
 527 }
 528
 529 int
 530 rte_eal_memseg_init(void)
 531 {
 532         return rte_eal_process_type() == RTE_PROC_PRIMARY ?
 533                         memseg_primary_init() :
 534                         memseg_secondary_init();
 535 }