lib/librte_eal/linux/eal/eal_memory.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 <errno.h>
   7 #include <fcntl.h>
   8 #include <stdarg.h>
   9 #include <stdbool.h>
  10 #include <stdlib.h>
  11 #include <stdio.h>
  12 #include <stdint.h>
  13 #include <inttypes.h>
  14 #include <string.h>
  15 #include <sys/mman.h>
  16 #include <sys/types.h>
  17 #include <sys/stat.h>
  18 #include <sys/queue.h>
  19 #include <sys/file.h>
  20 #include <sys/resource.h>
  21 #include <unistd.h>
  22 #include <limits.h>
  23 #include <sys/ioctl.h>
  24 #include <sys/time.h>
  25 #include <signal.h>
  26 #include <setjmp.h>
  27 #ifdef F_ADD_SEALS /* if file sealing is supported, so is memfd */
  28 #include <linux/memfd.h>
  29 #define MEMFD_SUPPORTED
  30 #endif
  31 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
  32 #include <numa.h>
  33 #include <numaif.h>
  34 #endif
  35
  36 #include <rte_errno.h>
  37 #include <rte_log.h>
  38 #include <rte_memory.h>
  39 #include <rte_launch.h>
  40 #include <rte_eal.h>
  41 #include <rte_per_lcore.h>
  42 #include <rte_lcore.h>
  43 #include <rte_common.h>
  44 #include <rte_string_fns.h>
  45
  46 #include "eal_private.h"
  47 #include "eal_memalloc.h"
  48 #include "eal_memcfg.h"
  49 #include "eal_internal_cfg.h"
  50 #include "eal_filesystem.h"
  51 #include "eal_hugepages.h"
  52 #include "eal_options.h"
  53
  54 #define PFN_MASK_SIZE   8
  55
  56 /**
  57  * @file
  58  * Huge page mapping under linux
  59  *
  60  * To reserve a big contiguous amount of memory, we use the hugepage
  61  * feature of linux. For that, we need to have hugetlbfs mounted. This
  62  * code will create many files in this directory (one per page) and
  63  * map them in virtual memory. For each page, we will retrieve its
  64  * physical address and remap it in order to have a virtual contiguous
  65  * zone as well as a physical contiguous zone.
  66  */
  67
  68 static int phys_addrs_available = -1;
  69
  70 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
  71
  72 uint64_t eal_get_baseaddr(void)
  73 {
  74         /*
  75          * Linux kernel uses a really high address as starting address for
  76          * serving mmaps calls. If there exists addressing limitations and IOVA
  77          * mode is VA, this starting address is likely too high for those
  78          * devices. However, it is possible to use a lower address in the
  79          * process virtual address space as with 64 bits there is a lot of
  80          * available space.
  81          *
  82          * Current known limitations are 39 or 40 bits. Setting the starting
  83          * address at 4GB implies there are 508GB or 1020GB for mapping the
  84          * available hugepages. This is likely enough for most systems, although
  85          * a device with addressing limitations should call
  86          * rte_mem_check_dma_mask for ensuring all memory is within supported
  87          * range.
  88          */
  89         return 0x100000000ULL;
  90 }
  91
  92 /*
  93  * Get physical address of any mapped virtual address in the current process.
  94  */
  95 phys_addr_t
  96 rte_mem_virt2phy(const void *virtaddr)
  97 {
  98         int fd, retval;
  99         uint64_t page, physaddr;
 100         unsigned long virt_pfn;
 101         int page_size;
 102         off_t offset;
 103
 104         if (phys_addrs_available == 0)
 105                 return RTE_BAD_IOVA;
 106
 107         /* standard page size */
 108         page_size = getpagesize();
 109
 110         fd = open("/proc/self/pagemap", O_RDONLY);
 111         if (fd < 0) {
 112                 RTE_LOG(INFO, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 113                         __func__, strerror(errno));
 114                 return RTE_BAD_IOVA;
 115         }
 116
 117         virt_pfn = (unsigned long)virtaddr / page_size;
 118         offset = sizeof(uint64_t) * virt_pfn;
 119         if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 120                 RTE_LOG(INFO, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 121                                 __func__, strerror(errno));
 122                 close(fd);
 123                 return RTE_BAD_IOVA;
 124         }
 125
 126         retval = read(fd, &page, PFN_MASK_SIZE);
 127         close(fd);
 128         if (retval < 0) {
 129                 RTE_LOG(INFO, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 130                                 __func__, strerror(errno));
 131                 return RTE_BAD_IOVA;
 132         } else if (retval != PFN_MASK_SIZE) {
 133                 RTE_LOG(INFO, EAL, "%s(): read %d bytes from /proc/self/pagemap "
 134                                 "but expected %d:\n",
 135                                 __func__, retval, PFN_MASK_SIZE);
 136                 return RTE_BAD_IOVA;
 137         }
 138
 139         /*
 140          * the pfn (page frame number) are bits 0-54 (see
 141          * pagemap.txt in linux Documentation)
 142          */
 143         if ((page & 0x7fffffffffffffULL) == 0)
 144                 return RTE_BAD_IOVA;
 145
 146         physaddr = ((page & 0x7fffffffffffffULL) * page_size)
 147                 + ((unsigned long)virtaddr % page_size);
 148
 149         return physaddr;
 150 }
 151
 152 rte_iova_t
 153 rte_mem_virt2iova(const void *virtaddr)
 154 {
 155         if (rte_eal_iova_mode() == RTE_IOVA_VA)
 156                 return (uintptr_t)virtaddr;
 157         return rte_mem_virt2phy(virtaddr);
 158 }
 159
 160 /*
 161  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 162  * it by browsing the /proc/self/pagemap special file.
 163  */
 164 static int
 165 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 166 {
 167         unsigned int i;
 168         phys_addr_t addr;
 169
 170         for (i = 0; i < hpi->num_pages[0]; i++) {
 171                 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
 172                 if (addr == RTE_BAD_PHYS_ADDR)
 173                         return -1;
 174                 hugepg_tbl[i].physaddr = addr;
 175         }
 176         return 0;
 177 }
 178
 179 /*
 180  * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
 181  */
 182 static int
 183 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 184 {
 185         unsigned int i;
 186         static phys_addr_t addr;
 187
 188         for (i = 0; i < hpi->num_pages[0]; i++) {
 189                 hugepg_tbl[i].physaddr = addr;
 190                 addr += hugepg_tbl[i].size;
 191         }
 192         return 0;
 193 }
 194
 195 /*
 196  * Check whether address-space layout randomization is enabled in
 197  * the kernel. This is important for multi-process as it can prevent
 198  * two processes mapping data to the same virtual address
 199  * Returns:
 200  *    0 - address space randomization disabled
 201  *    1/2 - address space randomization enabled
 202  *    negative error code on error
 203  */
 204 static int
 205 aslr_enabled(void)
 206 {
 207         char c;
 208         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 209         if (fd < 0)
 210                 return -errno;
 211         retval = read(fd, &c, 1);
 212         close(fd);
 213         if (retval < 0)
 214                 return -errno;
 215         if (retval == 0)
 216                 return -EIO;
 217         switch (c) {
 218                 case '0' : return 0;
 219                 case '1' : return 1;
 220                 case '2' : return 2;
 221                 default: return -EINVAL;
 222         }
 223 }
 224
 225 static sigjmp_buf huge_jmpenv;
 226
 227 static void huge_sigbus_handler(int signo __rte_unused)
 228 {
 229         siglongjmp(huge_jmpenv, 1);
 230 }
 231
 232 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
 233  * non-static local variable in the stack frame calling sigsetjmp might be
 234  * clobbered by a call to longjmp.
 235  */
 236 static int huge_wrap_sigsetjmp(void)
 237 {
 238         return sigsetjmp(huge_jmpenv, 1);
 239 }
 240
 241 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 242 /* Callback for numa library. */
 243 void numa_error(char *where)
 244 {
 245         RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
 246 }
 247 #endif
 248
 249 /*
 250  * Mmap all hugepages of hugepage table: it first open a file in
 251  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 252  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 253  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 254  * map contiguous physical blocks in contiguous virtual blocks.
 255  */
 256 static unsigned
 257 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
 258                   uint64_t *essential_memory __rte_unused)
 259 {
 260         int fd;
 261         unsigned i;
 262         void *virtaddr;
 263 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 264         int node_id = -1;
 265         int essential_prev = 0;
 266         int oldpolicy;
 267         struct bitmask *oldmask = NULL;
 268         bool have_numa = true;
 269         unsigned long maxnode = 0;
 270
 271         /* Check if kernel supports NUMA. */
 272         if (numa_available() != 0) {
 273                 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
 274                 have_numa = false;
 275         }
 276
 277         if (have_numa) {
 278                 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
 279                 oldmask = numa_allocate_nodemask();
 280                 if (get_mempolicy(&oldpolicy, oldmask->maskp,
 281                                   oldmask->size + 1, 0, 0) < 0) {
 282                         RTE_LOG(ERR, EAL,
 283                                 "Failed to get current mempolicy: %s. "
 284                                 "Assuming MPOL_DEFAULT.\n", strerror(errno));
 285                         oldpolicy = MPOL_DEFAULT;
 286                 }
 287                 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 288                         if (internal_config.socket_mem[i])
 289                                 maxnode = i + 1;
 290         }
 291 #endif
 292
 293         for (i = 0; i < hpi->num_pages[0]; i++) {
 294                 struct hugepage_file *hf = &hugepg_tbl[i];
 295                 uint64_t hugepage_sz = hpi->hugepage_sz;
 296
 297 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 298                 if (maxnode) {
 299                         unsigned int j;
 300
 301                         for (j = 0; j < maxnode; j++)
 302                                 if (essential_memory[j])
 303                                         break;
 304
 305                         if (j == maxnode) {
 306                                 node_id = (node_id + 1) % maxnode;
 307                                 while (!internal_config.socket_mem[node_id]) {
 308                                         node_id++;
 309                                         node_id %= maxnode;
 310                                 }
 311                                 essential_prev = 0;
 312                         } else {
 313                                 node_id = j;
 314                                 essential_prev = essential_memory[j];
 315
 316                                 if (essential_memory[j] < hugepage_sz)
 317                                         essential_memory[j] = 0;
 318                                 else
 319                                         essential_memory[j] -= hugepage_sz;
 320                         }
 321
 322                         RTE_LOG(DEBUG, EAL,
 323                                 "Setting policy MPOL_PREFERRED for socket %d\n",
 324                                 node_id);
 325                         numa_set_preferred(node_id);
 326                 }
 327 #endif
 328
 329                 hf->file_id = i;
 330                 hf->size = hugepage_sz;
 331                 eal_get_hugefile_path(hf->filepath, sizeof(hf->filepath),
 332                                 hpi->hugedir, hf->file_id);
 333                 hf->filepath[sizeof(hf->filepath) - 1] = '\0';
 334
 335                 /* try to create hugepage file */
 336                 fd = open(hf->filepath, O_CREAT | O_RDWR, 0600);
 337                 if (fd < 0) {
 338                         RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
 339                                         strerror(errno));
 340                         goto out;
 341                 }
 342
 343                 /* map the segment, and populate page tables,
 344                  * the kernel fills this segment with zeros. we don't care where
 345                  * this gets mapped - we already have contiguous memory areas
 346                  * ready for us to map into.
 347                  */
 348                 virtaddr = mmap(NULL, hugepage_sz, PROT_READ | PROT_WRITE,
 349                                 MAP_SHARED | MAP_POPULATE, fd, 0);
 350                 if (virtaddr == MAP_FAILED) {
 351                         RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
 352                                         strerror(errno));
 353                         close(fd);
 354                         goto out;
 355                 }
 356
 357                 hf->orig_va = virtaddr;
 358
 359                 /* In linux, hugetlb limitations, like cgroup, are
 360                  * enforced at fault time instead of mmap(), even
 361                  * with the option of MAP_POPULATE. Kernel will send
 362                  * a SIGBUS signal. To avoid to be killed, save stack
 363                  * environment here, if SIGBUS happens, we can jump
 364                  * back here.
 365                  */
 366                 if (huge_wrap_sigsetjmp()) {
 367                         RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
 368                                 "hugepages of size %u MB\n",
 369                                 (unsigned int)(hugepage_sz / 0x100000));
 370                         munmap(virtaddr, hugepage_sz);
 371                         close(fd);
 372                         unlink(hugepg_tbl[i].filepath);
 373 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 374                         if (maxnode)
 375                                 essential_memory[node_id] =
 376                                         essential_prev;
 377 #endif
 378                         goto out;
 379                 }
 380                 *(int *)virtaddr = 0;
 381
 382                 /* set shared lock on the file. */
 383                 if (flock(fd, LOCK_SH) < 0) {
 384                         RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
 385                                 __func__, strerror(errno));
 386                         close(fd);
 387                         goto out;
 388                 }
 389
 390                 close(fd);
 391         }
 392
 393 out:
 394 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 395         if (maxnode) {
 396                 RTE_LOG(DEBUG, EAL,
 397                         "Restoring previous memory policy: %d\n", oldpolicy);
 398                 if (oldpolicy == MPOL_DEFAULT) {
 399                         numa_set_localalloc();
 400                 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
 401                                          oldmask->size + 1) < 0) {
 402                         RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
 403                                 strerror(errno));
 404                         numa_set_localalloc();
 405                 }
 406         }
 407         if (oldmask != NULL)
 408                 numa_free_cpumask(oldmask);
 409 #endif
 410         return i;
 411 }
 412
 413 /*
 414  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 415  * page.
 416  */
 417 static int
 418 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 419 {
 420         int socket_id;
 421         char *end, *nodestr;
 422         unsigned i, hp_count = 0;
 423         uint64_t virt_addr;
 424         char buf[BUFSIZ];
 425         char hugedir_str[PATH_MAX];
 426         FILE *f;
 427
 428         f = fopen("/proc/self/numa_maps", "r");
 429         if (f == NULL) {
 430                 RTE_LOG(NOTICE, EAL, "NUMA support not available"
 431                         " consider that all memory is in socket_id 0\n");
 432                 return 0;
 433         }
 434
 435         snprintf(hugedir_str, sizeof(hugedir_str),
 436                         "%s/%s", hpi->hugedir, eal_get_hugefile_prefix());
 437
 438         /* parse numa map */
 439         while (fgets(buf, sizeof(buf), f) != NULL) {
 440
 441                 /* ignore non huge page */
 442                 if (strstr(buf, " huge ") == NULL &&
 443                                 strstr(buf, hugedir_str) == NULL)
 444                         continue;
 445
 446                 /* get zone addr */
 447                 virt_addr = strtoull(buf, &end, 16);
 448                 if (virt_addr == 0 || end == buf) {
 449                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 450                         goto error;
 451                 }
 452
 453                 /* get node id (socket id) */
 454                 nodestr = strstr(buf, " N");
 455                 if (nodestr == NULL) {
 456                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 457                         goto error;
 458                 }
 459                 nodestr += 2;
 460                 end = strstr(nodestr, "=");
 461                 if (end == NULL) {
 462                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 463                         goto error;
 464                 }
 465                 end[0] = '\0';
 466                 end = NULL;
 467
 468                 socket_id = strtoul(nodestr, &end, 0);
 469                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 470                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 471                         goto error;
 472                 }
 473
 474                 /* if we find this page in our mappings, set socket_id */
 475                 for (i = 0; i < hpi->num_pages[0]; i++) {
 476                         void *va = (void *)(unsigned long)virt_addr;
 477                         if (hugepg_tbl[i].orig_va == va) {
 478                                 hugepg_tbl[i].socket_id = socket_id;
 479                                 hp_count++;
 480 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 481                                 RTE_LOG(DEBUG, EAL,
 482                                         "Hugepage %s is on socket %d\n",
 483                                         hugepg_tbl[i].filepath, socket_id);
 484 #endif
 485                         }
 486                 }
 487         }
 488
 489         if (hp_count < hpi->num_pages[0])
 490                 goto error;
 491
 492         fclose(f);
 493         return 0;
 494
 495 error:
 496         fclose(f);
 497         return -1;
 498 }
 499
 500 static int
 501 cmp_physaddr(const void *a, const void *b)
 502 {
 503 #ifndef RTE_ARCH_PPC_64
 504         const struct hugepage_file *p1 = a;
 505         const struct hugepage_file *p2 = b;
 506 #else
 507         /* PowerPC needs memory sorted in reverse order from x86 */
 508         const struct hugepage_file *p1 = b;
 509         const struct hugepage_file *p2 = a;
 510 #endif
 511         if (p1->physaddr < p2->physaddr)
 512                 return -1;
 513         else if (p1->physaddr > p2->physaddr)
 514                 return 1;
 515         else
 516                 return 0;
 517 }
 518
 519 /*
 520  * Uses mmap to create a shared memory area for storage of data
 521  * Used in this file to store the hugepage file map on disk
 522  */
 523 static void *
 524 create_shared_memory(const char *filename, const size_t mem_size)
 525 {
 526         void *retval;
 527         int fd;
 528
 529         /* if no shared files mode is used, create anonymous memory instead */
 530         if (internal_config.no_shconf) {
 531                 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE,
 532                                 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 533                 if (retval == MAP_FAILED)
 534                         return NULL;
 535                 return retval;
 536         }
 537
 538         fd = open(filename, O_CREAT | O_RDWR, 0600);
 539         if (fd < 0)
 540                 return NULL;
 541         if (ftruncate(fd, mem_size) < 0) {
 542                 close(fd);
 543                 return NULL;
 544         }
 545         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 546         close(fd);
 547         if (retval == MAP_FAILED)
 548                 return NULL;
 549         return retval;
 550 }
 551
 552 /*
 553  * this copies *active* hugepages from one hugepage table to another.
 554  * destination is typically the shared memory.
 555  */
 556 static int
 557 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
 558                 const struct hugepage_file * src, int src_size)
 559 {
 560         int src_pos, dst_pos = 0;
 561
 562         for (src_pos = 0; src_pos < src_size; src_pos++) {
 563                 if (src[src_pos].orig_va != NULL) {
 564                         /* error on overflow attempt */
 565                         if (dst_pos == dest_size)
 566                                 return -1;
 567                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
 568                         dst_pos++;
 569                 }
 570         }
 571         return 0;
 572 }
 573
 574 static int
 575 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
 576                 unsigned num_hp_info)
 577 {
 578         unsigned socket, size;
 579         int page, nrpages = 0;
 580
 581         /* get total number of hugepages */
 582         for (size = 0; size < num_hp_info; size++)
 583                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 584                         nrpages +=
 585                         internal_config.hugepage_info[size].num_pages[socket];
 586
 587         for (page = 0; page < nrpages; page++) {
 588                 struct hugepage_file *hp = &hugepg_tbl[page];
 589
 590                 if (hp->orig_va != NULL && unlink(hp->filepath)) {
 591                         RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
 592                                 __func__, hp->filepath, strerror(errno));
 593                 }
 594         }
 595         return 0;
 596 }
 597
 598 /*
 599  * unmaps hugepages that are not going to be used. since we originally allocate
 600  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 601  */
 602 static int
 603 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
 604                 struct hugepage_info *hpi,
 605                 unsigned num_hp_info)
 606 {
 607         unsigned socket, size;
 608         int page, nrpages = 0;
 609
 610         /* get total number of hugepages */
 611         for (size = 0; size < num_hp_info; size++)
 612                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 613                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 614
 615         for (size = 0; size < num_hp_info; size++) {
 616                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 617                         unsigned pages_found = 0;
 618
 619                         /* traverse until we have unmapped all the unused pages */
 620                         for (page = 0; page < nrpages; page++) {
 621                                 struct hugepage_file *hp = &hugepg_tbl[page];
 622
 623                                 /* find a page that matches the criteria */
 624                                 if ((hp->size == hpi[size].hugepage_sz) &&
 625                                                 (hp->socket_id == (int) socket)) {
 626
 627                                         /* if we skipped enough pages, unmap the rest */
 628                                         if (pages_found == hpi[size].num_pages[socket]) {
 629                                                 uint64_t unmap_len;
 630
 631                                                 unmap_len = hp->size;
 632
 633                                                 /* get start addr and len of the remaining segment */
 634                                                 munmap(hp->orig_va,
 635                                                         (size_t)unmap_len);
 636
 637                                                 hp->orig_va = NULL;
 638                                                 if (unlink(hp->filepath) == -1) {
 639                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 640                                                                         __func__, hp->filepath, strerror(errno));
 641                                                         return -1;
 642                                                 }
 643                                         } else {
 644                                                 /* lock the page and skip */
 645                                                 pages_found++;
 646                                         }
 647
 648                                 } /* match page */
 649                         } /* foreach page */
 650                 } /* foreach socket */
 651         } /* foreach pagesize */
 652
 653         return 0;
 654 }
 655
 656 static int
 657 remap_segment(struct hugepage_file *hugepages, int seg_start, int seg_end)
 658 {
 659         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 660         struct rte_memseg_list *msl;
 661         struct rte_fbarray *arr;
 662         int cur_page, seg_len;
 663         unsigned int msl_idx;
 664         int ms_idx;
 665         uint64_t page_sz;
 666         size_t memseg_len;
 667         int socket_id;
 668
 669         page_sz = hugepages[seg_start].size;
 670         socket_id = hugepages[seg_start].socket_id;
 671         seg_len = seg_end - seg_start;
 672
 673         RTE_LOG(DEBUG, EAL, "Attempting to map %" PRIu64 "M on socket %i\n",
 674                         (seg_len * page_sz) >> 20ULL, socket_id);
 675
 676         /* find free space in memseg lists */
 677         for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
 678                 bool empty;
 679                 msl = &mcfg->memsegs[msl_idx];
 680                 arr = &msl->memseg_arr;
 681
 682                 if (msl->page_sz != page_sz)
 683                         continue;
 684                 if (msl->socket_id != socket_id)
 685                         continue;
 686
 687                 /* leave space for a hole if array is not empty */
 688                 empty = arr->count == 0;
 689                 ms_idx = rte_fbarray_find_next_n_free(arr, 0,
 690                                 seg_len + (empty ? 0 : 1));
 691
 692                 /* memseg list is full? */
 693                 if (ms_idx < 0)
 694                         continue;
 695
 696                 /* leave some space between memsegs, they are not IOVA
 697                  * contiguous, so they shouldn't be VA contiguous either.
 698                  */
 699                 if (!empty)
 700                         ms_idx++;
 701                 break;
 702         }
 703         if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
 704                 RTE_LOG(ERR, EAL, "Could not find space for memseg. Please increase %s and/or %s in configuration.\n",
 705                                 RTE_STR(CONFIG_RTE_MAX_MEMSEG_PER_TYPE),
 706                                 RTE_STR(CONFIG_RTE_MAX_MEM_PER_TYPE));
 707                 return -1;
 708         }
 709
 710 #ifdef RTE_ARCH_PPC_64
 711         /* for PPC64 we go through the list backwards */
 712         for (cur_page = seg_end - 1; cur_page >= seg_start;
 713                         cur_page--, ms_idx++) {
 714 #else
 715         for (cur_page = seg_start; cur_page < seg_end; cur_page++, ms_idx++) {
 716 #endif
 717                 struct hugepage_file *hfile = &hugepages[cur_page];
 718                 struct rte_memseg *ms = rte_fbarray_get(arr, ms_idx);
 719                 void *addr;
 720                 int fd;
 721
 722                 fd = open(hfile->filepath, O_RDWR);
 723                 if (fd < 0) {
 724                         RTE_LOG(ERR, EAL, "Could not open '%s': %s\n",
 725                                         hfile->filepath, strerror(errno));
 726                         return -1;
 727                 }
 728                 /* set shared lock on the file. */
 729                 if (flock(fd, LOCK_SH) < 0) {
 730                         RTE_LOG(DEBUG, EAL, "Could not lock '%s': %s\n",
 731                                         hfile->filepath, strerror(errno));
 732                         close(fd);
 733                         return -1;
 734                 }
 735                 memseg_len = (size_t)page_sz;
 736                 addr = RTE_PTR_ADD(msl->base_va, ms_idx * memseg_len);
 737
 738                 /* we know this address is already mmapped by memseg list, so
 739                  * using MAP_FIXED here is safe
 740                  */
 741                 addr = mmap(addr, page_sz, PROT_READ | PROT_WRITE,
 742                                 MAP_SHARED | MAP_POPULATE | MAP_FIXED, fd, 0);
 743                 if (addr == MAP_FAILED) {
 744                         RTE_LOG(ERR, EAL, "Couldn't remap '%s': %s\n",
 745                                         hfile->filepath, strerror(errno));
 746                         close(fd);
 747                         return -1;
 748                 }
 749
 750                 /* we have a new address, so unmap previous one */
 751 #ifndef RTE_ARCH_64
 752                 /* in 32-bit legacy mode, we have already unmapped the page */
 753                 if (!internal_config.legacy_mem)
 754                         munmap(hfile->orig_va, page_sz);
 755 #else
 756                 munmap(hfile->orig_va, page_sz);
 757 #endif
 758
 759                 hfile->orig_va = NULL;
 760                 hfile->final_va = addr;
 761
 762                 /* rewrite physical addresses in IOVA as VA mode */
 763                 if (rte_eal_iova_mode() == RTE_IOVA_VA)
 764                         hfile->physaddr = (uintptr_t)addr;
 765
 766                 /* set up memseg data */
 767                 ms->addr = addr;
 768                 ms->hugepage_sz = page_sz;
 769                 ms->len = memseg_len;
 770                 ms->iova = hfile->physaddr;
 771                 ms->socket_id = hfile->socket_id;
 772                 ms->nchannel = rte_memory_get_nchannel();
 773                 ms->nrank = rte_memory_get_nrank();
 774
 775                 rte_fbarray_set_used(arr, ms_idx);
 776
 777                 /* store segment fd internally */
 778                 if (eal_memalloc_set_seg_fd(msl_idx, ms_idx, fd) < 0)
 779                         RTE_LOG(ERR, EAL, "Could not store segment fd: %s\n",
 780                                 rte_strerror(rte_errno));
 781         }
 782         RTE_LOG(DEBUG, EAL, "Allocated %" PRIu64 "M on socket %i\n",
 783                         (seg_len * page_sz) >> 20, socket_id);
 784         return 0;
 785 }
 786
 787 static uint64_t
 788 get_mem_amount(uint64_t page_sz, uint64_t max_mem)
 789 {
 790         uint64_t area_sz, max_pages;
 791
 792         /* limit to RTE_MAX_MEMSEG_PER_LIST pages or RTE_MAX_MEM_MB_PER_LIST */
 793         max_pages = RTE_MAX_MEMSEG_PER_LIST;
 794         max_mem = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20, max_mem);
 795
 796         area_sz = RTE_MIN(page_sz * max_pages, max_mem);
 797
 798         /* make sure the list isn't smaller than the page size */
 799         area_sz = RTE_MAX(area_sz, page_sz);
 800
 801         return RTE_ALIGN(area_sz, page_sz);
 802 }
 803
 804 static int
 805 free_memseg_list(struct rte_memseg_list *msl)
 806 {
 807         if (rte_fbarray_destroy(&msl->memseg_arr)) {
 808                 RTE_LOG(ERR, EAL, "Cannot destroy memseg list\n");
 809                 return -1;
 810         }
 811         memset(msl, 0, sizeof(*msl));
 812         return 0;
 813 }
 814
 815 #define MEMSEG_LIST_FMT "memseg-%" PRIu64 "k-%i-%i"
 816 static int
 817 alloc_memseg_list(struct rte_memseg_list *msl, uint64_t page_sz,
 818                 int n_segs, int socket_id, int type_msl_idx)
 819 {
 820         char name[RTE_FBARRAY_NAME_LEN];
 821
 822         snprintf(name, sizeof(name), MEMSEG_LIST_FMT, page_sz >> 10, socket_id,
 823                  type_msl_idx);
 824         if (rte_fbarray_init(&msl->memseg_arr, name, n_segs,
 825                         sizeof(struct rte_memseg))) {
 826                 RTE_LOG(ERR, EAL, "Cannot allocate memseg list: %s\n",
 827                         rte_strerror(rte_errno));
 828                 return -1;
 829         }
 830
 831         msl->page_sz = page_sz;
 832         msl->socket_id = socket_id;
 833         msl->base_va = NULL;
 834         msl->heap = 1; /* mark it as a heap segment */
 835
 836         RTE_LOG(DEBUG, EAL, "Memseg list allocated: 0x%zxkB at socket %i\n",
 837                         (size_t)page_sz >> 10, socket_id);
 838
 839         return 0;
 840 }
 841
 842 static int
 843 alloc_va_space(struct rte_memseg_list *msl)
 844 {
 845         uint64_t page_sz;
 846         size_t mem_sz;
 847         void *addr;
 848         int flags = 0;
 849
 850         page_sz = msl->page_sz;
 851         mem_sz = page_sz * msl->memseg_arr.len;
 852
 853         addr = eal_get_virtual_area(msl->base_va, &mem_sz, page_sz, 0, flags);
 854         if (addr == NULL) {
 855                 if (rte_errno == EADDRNOTAVAIL)
 856                         RTE_LOG(ERR, EAL, "Could not mmap %llu bytes at [%p] - "
 857                                 "please use '--" OPT_BASE_VIRTADDR "' option\n",
 858                                 (unsigned long long)mem_sz, msl->base_va);
 859                 else
 860                         RTE_LOG(ERR, EAL, "Cannot reserve memory\n");
 861                 return -1;
 862         }
 863         msl->base_va = addr;
 864         msl->len = mem_sz;
 865
 866         return 0;
 867 }
 868
 869 /*
 870  * Our VA space is not preallocated yet, so preallocate it here. We need to know
 871  * how many segments there are in order to map all pages into one address space,
 872  * and leave appropriate holes between segments so that rte_malloc does not
 873  * concatenate them into one big segment.
 874  *
 875  * we also need to unmap original pages to free up address space.
 876  */
 877 static int __rte_unused
 878 prealloc_segments(struct hugepage_file *hugepages, int n_pages)
 879 {
 880         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 881         int cur_page, seg_start_page, end_seg, new_memseg;
 882         unsigned int hpi_idx, socket, i;
 883         int n_contig_segs, n_segs;
 884         int msl_idx;
 885
 886         /* before we preallocate segments, we need to free up our VA space.
 887          * we're not removing files, and we already have information about
 888          * PA-contiguousness, so it is safe to unmap everything.
 889          */
 890         for (cur_page = 0; cur_page < n_pages; cur_page++) {
 891                 struct hugepage_file *hpi = &hugepages[cur_page];
 892                 munmap(hpi->orig_va, hpi->size);
 893                 hpi->orig_va = NULL;
 894         }
 895
 896         /* we cannot know how many page sizes and sockets we have discovered, so
 897          * loop over all of them
 898          */
 899         for (hpi_idx = 0; hpi_idx < internal_config.num_hugepage_sizes;
 900                         hpi_idx++) {
 901                 uint64_t page_sz =
 902                         internal_config.hugepage_info[hpi_idx].hugepage_sz;
 903
 904                 for (i = 0; i < rte_socket_count(); i++) {
 905                         struct rte_memseg_list *msl;
 906
 907                         socket = rte_socket_id_by_idx(i);
 908                         n_contig_segs = 0;
 909                         n_segs = 0;
 910                         seg_start_page = -1;
 911
 912                         for (cur_page = 0; cur_page < n_pages; cur_page++) {
 913                                 struct hugepage_file *prev, *cur;
 914                                 int prev_seg_start_page = -1;
 915
 916                                 cur = &hugepages[cur_page];
 917                                 prev = cur_page == 0 ? NULL :
 918                                                 &hugepages[cur_page - 1];
 919
 920                                 new_memseg = 0;
 921                                 end_seg = 0;
 922
 923                                 if (cur->size == 0)
 924                                         end_seg = 1;
 925                                 else if (cur->socket_id != (int) socket)
 926                                         end_seg = 1;
 927                                 else if (cur->size != page_sz)
 928                                         end_seg = 1;
 929                                 else if (cur_page == 0)
 930                                         new_memseg = 1;
 931 #ifdef RTE_ARCH_PPC_64
 932                                 /* On PPC64 architecture, the mmap always start
 933                                  * from higher address to lower address. Here,
 934                                  * physical addresses are in descending order.
 935                                  */
 936                                 else if ((prev->physaddr - cur->physaddr) !=
 937                                                 cur->size)
 938                                         new_memseg = 1;
 939 #else
 940                                 else if ((cur->physaddr - prev->physaddr) !=
 941                                                 cur->size)
 942                                         new_memseg = 1;
 943 #endif
 944                                 if (new_memseg) {
 945                                         /* if we're already inside a segment,
 946                                          * new segment means end of current one
 947                                          */
 948                                         if (seg_start_page != -1) {
 949                                                 end_seg = 1;
 950                                                 prev_seg_start_page =
 951                                                                 seg_start_page;
 952                                         }
 953                                         seg_start_page = cur_page;
 954                                 }
 955
 956                                 if (end_seg) {
 957                                         if (prev_seg_start_page != -1) {
 958                                                 /* we've found a new segment */
 959                                                 n_contig_segs++;
 960                                                 n_segs += cur_page -
 961                                                         prev_seg_start_page;
 962                                         } else if (seg_start_page != -1) {
 963                                                 /* we didn't find new segment,
 964                                                  * but did end current one
 965                                                  */
 966                                                 n_contig_segs++;
 967                                                 n_segs += cur_page -
 968                                                                 seg_start_page;
 969                                                 seg_start_page = -1;
 970                                                 continue;
 971                                         } else {
 972                                                 /* we're skipping this page */
 973                                                 continue;
 974                                         }
 975                                 }
 976                                 /* segment continues */
 977                         }
 978                         /* check if we missed last segment */
 979                         if (seg_start_page != -1) {
 980                                 n_contig_segs++;
 981                                 n_segs += cur_page - seg_start_page;
 982                         }
 983
 984                         /* if no segments were found, do not preallocate */
 985                         if (n_segs == 0)
 986                                 continue;
 987
 988                         /* we now have total number of pages that we will
 989                          * allocate for this segment list. add separator pages
 990                          * to the total count, and preallocate VA space.
 991                          */
 992                         n_segs += n_contig_segs - 1;
 993
 994                         /* now, preallocate VA space for these segments */
 995
 996                         /* first, find suitable memseg list for this */
 997                         for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS;
 998                                         msl_idx++) {
 999                                 msl = &mcfg->memsegs[msl_idx];
1000
1001                                 if (msl->base_va != NULL)
1002                                         continue;
1003                                 break;
1004                         }
1005                         if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
1006                                 RTE_LOG(ERR, EAL, "Not enough space in memseg lists, please increase %s\n",
1007                                         RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
1008                                 return -1;
1009                         }
1010
1011                         /* now, allocate fbarray itself */
1012                         if (alloc_memseg_list(msl, page_sz, n_segs, socket,
1013                                                 msl_idx) < 0)
1014                                 return -1;
1015
1016                         /* finally, allocate VA space */
1017                         if (alloc_va_space(msl) < 0)
1018                                 return -1;
1019                 }
1020         }
1021         return 0;
1022 }
1023
1024 /*
1025  * We cannot reallocate memseg lists on the fly because PPC64 stores pages
1026  * backwards, therefore we have to process the entire memseg first before
1027  * remapping it into memseg list VA space.
1028  */
1029 static int
1030 remap_needed_hugepages(struct hugepage_file *hugepages, int n_pages)
1031 {
1032         int cur_page, seg_start_page, new_memseg, ret;
1033
1034         seg_start_page = 0;
1035         for (cur_page = 0; cur_page < n_pages; cur_page++) {
1036                 struct hugepage_file *prev, *cur;
1037
1038                 new_memseg = 0;
1039
1040                 cur = &hugepages[cur_page];
1041                 prev = cur_page == 0 ? NULL : &hugepages[cur_page - 1];
1042
1043                 /* if size is zero, no more pages left */
1044                 if (cur->size == 0)
1045                         break;
1046
1047                 if (cur_page == 0)
1048                         new_memseg = 1;
1049                 else if (cur->socket_id != prev->socket_id)
1050                         new_memseg = 1;
1051                 else if (cur->size != prev->size)
1052                         new_memseg = 1;
1053 #ifdef RTE_ARCH_PPC_64
1054                 /* On PPC64 architecture, the mmap always start from higher
1055                  * address to lower address. Here, physical addresses are in
1056                  * descending order.
1057                  */
1058                 else if ((prev->physaddr - cur->physaddr) != cur->size)
1059                         new_memseg = 1;
1060 #else
1061                 else if ((cur->physaddr - prev->physaddr) != cur->size)
1062                         new_memseg = 1;
1063 #endif
1064
1065                 if (new_memseg) {
1066                         /* if this isn't the first time, remap segment */
1067                         if (cur_page != 0) {
1068                                 ret = remap_segment(hugepages, seg_start_page,
1069                                                 cur_page);
1070                                 if (ret != 0)
1071                                         return -1;
1072                         }
1073                         /* remember where we started */
1074                         seg_start_page = cur_page;
1075                 }
1076                 /* continuation of previous memseg */
1077         }
1078         /* we were stopped, but we didn't remap the last segment, do it now */
1079         if (cur_page != 0) {
1080                 ret = remap_segment(hugepages, seg_start_page,
1081                                 cur_page);
1082                 if (ret != 0)
1083                         return -1;
1084         }
1085         return 0;
1086 }
1087
1088 __rte_unused /* function is unused on 32-bit builds */
1089 static inline uint64_t
1090 get_socket_mem_size(int socket)
1091 {
1092         uint64_t size = 0;
1093         unsigned i;
1094
1095         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
1096                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
1097                 size += hpi->hugepage_sz * hpi->num_pages[socket];
1098         }
1099
1100         return size;
1101 }
1102
1103 /*
1104  * This function is a NUMA-aware equivalent of calc_num_pages.
1105  * It takes in the list of hugepage sizes and the
1106  * number of pages thereof, and calculates the best number of
1107  * pages of each size to fulfill the request for <memory> ram
1108  */
1109 static int
1110 calc_num_pages_per_socket(uint64_t * memory,
1111                 struct hugepage_info *hp_info,
1112                 struct hugepage_info *hp_used,
1113                 unsigned num_hp_info)
1114 {
1115         unsigned socket, j, i = 0;
1116         unsigned requested, available;
1117         int total_num_pages = 0;
1118         uint64_t remaining_mem, cur_mem;
1119         uint64_t total_mem = internal_config.memory;
1120
1121         if (num_hp_info == 0)
1122                 return -1;
1123
1124         /* if specific memory amounts per socket weren't requested */
1125         if (internal_config.force_sockets == 0) {
1126                 size_t total_size;
1127 #ifdef RTE_ARCH_64
1128                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
1129                 size_t default_size;
1130                 unsigned lcore_id;
1131
1132                 /* Compute number of cores per socket */
1133                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
1134                 RTE_LCORE_FOREACH(lcore_id) {
1135                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
1136                 }
1137
1138                 /*
1139                  * Automatically spread requested memory amongst detected sockets according
1140                  * to number of cores from cpu mask present on each socket
1141                  */
1142                 total_size = internal_config.memory;
1143                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
1144
1145                         /* Set memory amount per socket */
1146                         default_size = (internal_config.memory * cpu_per_socket[socket])
1147                                         / rte_lcore_count();
1148
1149                         /* Limit to maximum available memory on socket */
1150                         default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
1151
1152                         /* Update sizes */
1153                         memory[socket] = default_size;
1154                         total_size -= default_size;
1155                 }
1156
1157                 /*
1158                  * If some memory is remaining, try to allocate it by getting all
1159                  * available memory from sockets, one after the other
1160                  */
1161                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
1162                         /* take whatever is available */
1163                         default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
1164                                                total_size);
1165
1166                         /* Update sizes */
1167                         memory[socket] += default_size;
1168                         total_size -= default_size;
1169                 }
1170 #else
1171                 /* in 32-bit mode, allocate all of the memory only on master
1172                  * lcore socket
1173                  */
1174                 total_size = internal_config.memory;
1175                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0;
1176                                 socket++) {
1177                         struct rte_config *cfg = rte_eal_get_configuration();
1178                         unsigned int master_lcore_socket;
1179
1180                         master_lcore_socket =
1181                                 rte_lcore_to_socket_id(cfg->master_lcore);
1182
1183                         if (master_lcore_socket != socket)
1184                                 continue;
1185
1186                         /* Update sizes */
1187                         memory[socket] = total_size;
1188                         break;
1189                 }
1190 #endif
1191         }
1192
1193         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
1194                 /* skips if the memory on specific socket wasn't requested */
1195                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
1196                         strlcpy(hp_used[i].hugedir, hp_info[i].hugedir,
1197                                 sizeof(hp_used[i].hugedir));
1198                         hp_used[i].num_pages[socket] = RTE_MIN(
1199                                         memory[socket] / hp_info[i].hugepage_sz,
1200                                         hp_info[i].num_pages[socket]);
1201
1202                         cur_mem = hp_used[i].num_pages[socket] *
1203                                         hp_used[i].hugepage_sz;
1204
1205                         memory[socket] -= cur_mem;
1206                         total_mem -= cur_mem;
1207
1208                         total_num_pages += hp_used[i].num_pages[socket];
1209
1210                         /* check if we have met all memory requests */
1211                         if (memory[socket] == 0)
1212                                 break;
1213
1214                         /* check if we have any more pages left at this size, if so
1215                          * move on to next size */
1216                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
1217                                 continue;
1218                         /* At this point we know that there are more pages available that are
1219                          * bigger than the memory we want, so lets see if we can get enough
1220                          * from other page sizes.
1221                          */
1222                         remaining_mem = 0;
1223                         for (j = i+1; j < num_hp_info; j++)
1224                                 remaining_mem += hp_info[j].hugepage_sz *
1225                                 hp_info[j].num_pages[socket];
1226
1227                         /* is there enough other memory, if not allocate another page and quit */
1228                         if (remaining_mem < memory[socket]){
1229                                 cur_mem = RTE_MIN(memory[socket],
1230                                                 hp_info[i].hugepage_sz);
1231                                 memory[socket] -= cur_mem;
1232                                 total_mem -= cur_mem;
1233                                 hp_used[i].num_pages[socket]++;
1234                                 total_num_pages++;
1235                                 break; /* we are done with this socket*/
1236                         }
1237                 }
1238                 /* if we didn't satisfy all memory requirements per socket */
1239                 if (memory[socket] > 0 &&
1240                                 internal_config.socket_mem[socket] != 0) {
1241                         /* to prevent icc errors */
1242                         requested = (unsigned) (internal_config.socket_mem[socket] /
1243                                         0x100000);
1244                         available = requested -
1245                                         ((unsigned) (memory[socket] / 0x100000));
1246                         RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
1247                                         "Requested: %uMB, available: %uMB\n", socket,
1248                                         requested, available);
1249                         return -1;
1250                 }
1251         }
1252
1253         /* if we didn't satisfy total memory requirements */
1254         if (total_mem > 0) {
1255                 requested = (unsigned) (internal_config.memory / 0x100000);
1256                 available = requested - (unsigned) (total_mem / 0x100000);
1257                 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
1258                                 " available: %uMB\n", requested, available);
1259                 return -1;
1260         }
1261         return total_num_pages;
1262 }
1263
1264 static inline size_t
1265 eal_get_hugepage_mem_size(void)
1266 {
1267         uint64_t size = 0;
1268         unsigned i, j;
1269
1270         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
1271                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
1272                 if (strnlen(hpi->hugedir, sizeof(hpi->hugedir)) != 0) {
1273                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1274                                 size += hpi->hugepage_sz * hpi->num_pages[j];
1275                         }
1276                 }
1277         }
1278
1279         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
1280 }
1281
1282 static struct sigaction huge_action_old;
1283 static int huge_need_recover;
1284
1285 static void
1286 huge_register_sigbus(void)
1287 {
1288         sigset_t mask;
1289         struct sigaction action;
1290
1291         sigemptyset(&mask);
1292         sigaddset(&mask, SIGBUS);
1293         action.sa_flags = 0;
1294         action.sa_mask = mask;
1295         action.sa_handler = huge_sigbus_handler;
1296
1297         huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
1298 }
1299
1300 static void
1301 huge_recover_sigbus(void)
1302 {
1303         if (huge_need_recover) {
1304                 sigaction(SIGBUS, &huge_action_old, NULL);
1305                 huge_need_recover = 0;
1306         }
1307 }
1308
1309 /*
1310  * Prepare physical memory mapping: fill configuration structure with
1311  * these infos, return 0 on success.
1312  *  1. map N huge pages in separate files in hugetlbfs
1313  *  2. find associated physical addr
1314  *  3. find associated NUMA socket ID
1315  *  4. sort all huge pages by physical address
1316  *  5. remap these N huge pages in the correct order
1317  *  6. unmap the first mapping
1318  *  7. fill memsegs in configuration with contiguous zones
1319  */
1320 static int
1321 eal_legacy_hugepage_init(void)
1322 {
1323         struct rte_mem_config *mcfg;
1324         struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1325         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1326         struct rte_fbarray *arr;
1327         struct rte_memseg *ms;
1328
1329         uint64_t memory[RTE_MAX_NUMA_NODES];
1330
1331         unsigned hp_offset;
1332         int i, j;
1333         int nr_hugefiles, nr_hugepages = 0;
1334         void *addr;
1335
1336         memset(used_hp, 0, sizeof(used_hp));
1337
1338         /* get pointer to global configuration */
1339         mcfg = rte_eal_get_configuration()->mem_config;
1340
1341         /* hugetlbfs can be disabled */
1342         if (internal_config.no_hugetlbfs) {
1343                 void *prealloc_addr;
1344                 size_t mem_sz;
1345                 struct rte_memseg_list *msl;
1346                 int n_segs, cur_seg, fd, flags;
1347 #ifdef MEMFD_SUPPORTED
1348                 int memfd;
1349 #endif
1350                 uint64_t page_sz;
1351
1352                 /* nohuge mode is legacy mode */
1353                 internal_config.legacy_mem = 1;
1354
1355                 /* nohuge mode is single-file segments mode */
1356                 internal_config.single_file_segments = 1;
1357
1358                 /* create a memseg list */
1359                 msl = &mcfg->memsegs[0];
1360
1361                 page_sz = RTE_PGSIZE_4K;
1362                 n_segs = internal_config.memory / page_sz;
1363
1364                 if (rte_fbarray_init(&msl->memseg_arr, "nohugemem", n_segs,
1365                                         sizeof(struct rte_memseg))) {
1366                         RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
1367                         return -1;
1368                 }
1369
1370                 /* set up parameters for anonymous mmap */
1371                 fd = -1;
1372                 flags = MAP_PRIVATE | MAP_ANONYMOUS;
1373
1374 #ifdef MEMFD_SUPPORTED
1375                 /* create a memfd and store it in the segment fd table */
1376                 memfd = memfd_create("nohuge", 0);
1377                 if (memfd < 0) {
1378                         RTE_LOG(DEBUG, EAL, "Cannot create memfd: %s\n",
1379                                         strerror(errno));
1380                         RTE_LOG(DEBUG, EAL, "Falling back to anonymous map\n");
1381                 } else {
1382                         /* we got an fd - now resize it */
1383                         if (ftruncate(memfd, internal_config.memory) < 0) {
1384                                 RTE_LOG(ERR, EAL, "Cannot resize memfd: %s\n",
1385                                                 strerror(errno));
1386                                 RTE_LOG(ERR, EAL, "Falling back to anonymous map\n");
1387                                 close(memfd);
1388                         } else {
1389                                 /* creating memfd-backed file was successful.
1390                                  * we want changes to memfd to be visible to
1391                                  * other processes (such as vhost backend), so
1392                                  * map it as shared memory.
1393                                  */
1394                                 RTE_LOG(DEBUG, EAL, "Using memfd for anonymous memory\n");
1395                                 fd = memfd;
1396                                 flags = MAP_SHARED;
1397                         }
1398                 }
1399 #endif
1400                 /* preallocate address space for the memory, so that it can be
1401                  * fit into the DMA mask.
1402                  */
1403                 mem_sz = internal_config.memory;
1404                 prealloc_addr = eal_get_virtual_area(
1405                                 NULL, &mem_sz, page_sz, 0, 0);
1406                 if (prealloc_addr == NULL) {
1407                         RTE_LOG(ERR, EAL,
1408                                         "%s: reserving memory area failed: "
1409                                         "%s\n",
1410                                         __func__, strerror(errno));
1411                         return -1;
1412                 }
1413                 addr = mmap(prealloc_addr, mem_sz, PROT_READ | PROT_WRITE,
1414                                 flags | MAP_FIXED, fd, 0);
1415                 if (addr == MAP_FAILED || addr != prealloc_addr) {
1416                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1417                                         strerror(errno));
1418                         munmap(prealloc_addr, mem_sz);
1419                         return -1;
1420                 }
1421                 msl->base_va = addr;
1422                 msl->page_sz = page_sz;
1423                 msl->socket_id = 0;
1424                 msl->len = mem_sz;
1425                 msl->heap = 1;
1426
1427                 /* we're in single-file segments mode, so only the segment list
1428                  * fd needs to be set up.
1429                  */
1430                 if (fd != -1) {
1431                         if (eal_memalloc_set_seg_list_fd(0, fd) < 0) {
1432                                 RTE_LOG(ERR, EAL, "Cannot set up segment list fd\n");
1433                                 /* not a serious error, proceed */
1434                         }
1435                 }
1436
1437                 /* populate memsegs. each memseg is one page long */
1438                 for (cur_seg = 0; cur_seg < n_segs; cur_seg++) {
1439                         arr = &msl->memseg_arr;
1440
1441                         ms = rte_fbarray_get(arr, cur_seg);
1442                         if (rte_eal_iova_mode() == RTE_IOVA_VA)
1443                                 ms->iova = (uintptr_t)addr;
1444                         else
1445                                 ms->iova = RTE_BAD_IOVA;
1446                         ms->addr = addr;
1447                         ms->hugepage_sz = page_sz;
1448                         ms->socket_id = 0;
1449                         ms->len = page_sz;
1450
1451                         rte_fbarray_set_used(arr, cur_seg);
1452
1453                         addr = RTE_PTR_ADD(addr, (size_t)page_sz);
1454                 }
1455                 if (mcfg->dma_maskbits &&
1456                     rte_mem_check_dma_mask_thread_unsafe(mcfg->dma_maskbits)) {
1457                         RTE_LOG(ERR, EAL,
1458                                 "%s(): couldn't allocate memory due to IOVA exceeding limits of current DMA mask.\n",
1459                                 __func__);
1460                         if (rte_eal_iova_mode() == RTE_IOVA_VA &&
1461                             rte_eal_using_phys_addrs())
1462                                 RTE_LOG(ERR, EAL,
1463                                         "%s(): Please try initializing EAL with --iova-mode=pa parameter.\n",
1464                                         __func__);
1465                         goto fail;
1466                 }
1467                 return 0;
1468         }
1469
1470         /* calculate total number of hugepages available. at this point we haven't
1471          * yet started sorting them so they all are on socket 0 */
1472         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1473                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1474                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1475
1476                 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1477         }
1478
1479         /*
1480          * allocate a memory area for hugepage table.
1481          * this isn't shared memory yet. due to the fact that we need some
1482          * processing done on these pages, shared memory will be created
1483          * at a later stage.
1484          */
1485         tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1486         if (tmp_hp == NULL)
1487                 goto fail;
1488
1489         memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1490
1491         hp_offset = 0; /* where we start the current page size entries */
1492
1493         huge_register_sigbus();
1494
1495         /* make a copy of socket_mem, needed for balanced allocation. */
1496         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1497                 memory[i] = internal_config.socket_mem[i];
1498
1499         /* map all hugepages and sort them */
1500         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1501                 unsigned pages_old, pages_new;
1502                 struct hugepage_info *hpi;
1503
1504                 /*
1505                  * we don't yet mark hugepages as used at this stage, so
1506                  * we just map all hugepages available to the system
1507                  * all hugepages are still located on socket 0
1508                  */
1509                 hpi = &internal_config.hugepage_info[i];
1510
1511                 if (hpi->num_pages[0] == 0)
1512                         continue;
1513
1514                 /* map all hugepages available */
1515                 pages_old = hpi->num_pages[0];
1516                 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, memory);
1517                 if (pages_new < pages_old) {
1518                         RTE_LOG(DEBUG, EAL,
1519                                 "%d not %d hugepages of size %u MB allocated\n",
1520                                 pages_new, pages_old,
1521                                 (unsigned)(hpi->hugepage_sz / 0x100000));
1522
1523                         int pages = pages_old - pages_new;
1524
1525                         nr_hugepages -= pages;
1526                         hpi->num_pages[0] = pages_new;
1527                         if (pages_new == 0)
1528                                 continue;
1529                 }
1530
1531                 if (rte_eal_using_phys_addrs() &&
1532                                 rte_eal_iova_mode() != RTE_IOVA_VA) {
1533                         /* find physical addresses for each hugepage */
1534                         if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1535                                 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1536                                         "for %u MB pages\n",
1537                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1538                                 goto fail;
1539                         }
1540                 } else {
1541                         /* set physical addresses for each hugepage */
1542                         if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1543                                 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1544                                         "for %u MB pages\n",
1545                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1546                                 goto fail;
1547                         }
1548                 }
1549
1550                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1551                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1552                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1553                         goto fail;
1554                 }
1555
1556                 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1557                       sizeof(struct hugepage_file), cmp_physaddr);
1558
1559                 /* we have processed a num of hugepages of this size, so inc offset */
1560                 hp_offset += hpi->num_pages[0];
1561         }
1562
1563         huge_recover_sigbus();
1564
1565         if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1566                 internal_config.memory = eal_get_hugepage_mem_size();
1567
1568         nr_hugefiles = nr_hugepages;
1569
1570
1571         /* clean out the numbers of pages */
1572         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1573                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1574                         internal_config.hugepage_info[i].num_pages[j] = 0;
1575
1576         /* get hugepages for each socket */
1577         for (i = 0; i < nr_hugefiles; i++) {
1578                 int socket = tmp_hp[i].socket_id;
1579
1580                 /* find a hugepage info with right size and increment num_pages */
1581                 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1582                                 (int)internal_config.num_hugepage_sizes);
1583                 for (j = 0; j < nb_hpsizes; j++) {
1584                         if (tmp_hp[i].size ==
1585                                         internal_config.hugepage_info[j].hugepage_sz) {
1586                                 internal_config.hugepage_info[j].num_pages[socket]++;
1587                         }
1588                 }
1589         }
1590
1591         /* make a copy of socket_mem, needed for number of pages calculation */
1592         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1593                 memory[i] = internal_config.socket_mem[i];
1594
1595         /* calculate final number of pages */
1596         nr_hugepages = calc_num_pages_per_socket(memory,
1597                         internal_config.hugepage_info, used_hp,
1598                         internal_config.num_hugepage_sizes);
1599
1600         /* error if not enough memory available */
1601         if (nr_hugepages < 0)
1602                 goto fail;
1603
1604         /* reporting in! */
1605         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1606                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1607                         if (used_hp[i].num_pages[j] > 0) {
1608                                 RTE_LOG(DEBUG, EAL,
1609                                         "Requesting %u pages of size %uMB"
1610                                         " from socket %i\n",
1611                                         used_hp[i].num_pages[j],
1612                                         (unsigned)
1613                                         (used_hp[i].hugepage_sz / 0x100000),
1614                                         j);
1615                         }
1616                 }
1617         }
1618
1619         /* create shared memory */
1620         hugepage = create_shared_memory(eal_hugepage_data_path(),
1621                         nr_hugefiles * sizeof(struct hugepage_file));
1622
1623         if (hugepage == NULL) {
1624                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1625                 goto fail;
1626         }
1627         memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1628
1629         /*
1630          * unmap pages that we won't need (looks at used_hp).
1631          * also, sets final_va to NULL on pages that were unmapped.
1632          */
1633         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1634                         internal_config.num_hugepage_sizes) < 0) {
1635                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1636                 goto fail;
1637         }
1638
1639         /*
1640          * copy stuff from malloc'd hugepage* to the actual shared memory.
1641          * this procedure only copies those hugepages that have orig_va
1642          * not NULL. has overflow protection.
1643          */
1644         if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1645                         tmp_hp, nr_hugefiles) < 0) {
1646                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1647                 goto fail;
1648         }
1649
1650 #ifndef RTE_ARCH_64
1651         /* for legacy 32-bit mode, we did not preallocate VA space, so do it */
1652         if (internal_config.legacy_mem &&
1653                         prealloc_segments(hugepage, nr_hugefiles)) {
1654                 RTE_LOG(ERR, EAL, "Could not preallocate VA space for hugepages\n");
1655                 goto fail;
1656         }
1657 #endif
1658
1659         /* remap all pages we do need into memseg list VA space, so that those
1660          * pages become first-class citizens in DPDK memory subsystem
1661          */
1662         if (remap_needed_hugepages(hugepage, nr_hugefiles)) {
1663                 RTE_LOG(ERR, EAL, "Couldn't remap hugepage files into memseg lists\n");
1664                 goto fail;
1665         }
1666
1667         /* free the hugepage backing files */
1668         if (internal_config.hugepage_unlink &&
1669                 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1670                 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1671                 goto fail;
1672         }
1673
1674         /* free the temporary hugepage table */
1675         free(tmp_hp);
1676         tmp_hp = NULL;
1677
1678         munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1679         hugepage = NULL;
1680
1681         /* we're not going to allocate more pages, so release VA space for
1682          * unused memseg lists
1683          */
1684         for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
1685                 struct rte_memseg_list *msl = &mcfg->memsegs[i];
1686                 size_t mem_sz;
1687
1688                 /* skip inactive lists */
1689                 if (msl->base_va == NULL)
1690                         continue;
1691                 /* skip lists where there is at least one page allocated */
1692                 if (msl->memseg_arr.count > 0)
1693                         continue;
1694                 /* this is an unused list, deallocate it */
1695                 mem_sz = msl->len;
1696                 munmap(msl->base_va, mem_sz);
1697                 msl->base_va = NULL;
1698                 msl->heap = 0;
1699
1700                 /* destroy backing fbarray */
1701                 rte_fbarray_destroy(&msl->memseg_arr);
1702         }
1703
1704         if (mcfg->dma_maskbits &&
1705             rte_mem_check_dma_mask_thread_unsafe(mcfg->dma_maskbits)) {
1706                 RTE_LOG(ERR, EAL,
1707                         "%s(): couldn't allocate memory due to IOVA exceeding limits of current DMA mask.\n",
1708                         __func__);
1709                 goto fail;
1710         }
1711
1712         return 0;
1713
1714 fail:
1715         huge_recover_sigbus();
1716         free(tmp_hp);
1717         if (hugepage != NULL)
1718                 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1719
1720         return -1;
1721 }
1722
1723 static int __rte_unused
1724 hugepage_count_walk(const struct rte_memseg_list *msl, void *arg)
1725 {
1726         struct hugepage_info *hpi = arg;
1727
1728         if (msl->page_sz != hpi->hugepage_sz)
1729                 return 0;
1730
1731         hpi->num_pages[msl->socket_id] += msl->memseg_arr.len;
1732         return 0;
1733 }
1734
1735 static int
1736 limits_callback(int socket_id, size_t cur_limit, size_t new_len)
1737 {
1738         RTE_SET_USED(socket_id);
1739         RTE_SET_USED(cur_limit);
1740         RTE_SET_USED(new_len);
1741         return -1;
1742 }
1743
1744 static int
1745 eal_hugepage_init(void)
1746 {
1747         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1748         uint64_t memory[RTE_MAX_NUMA_NODES];
1749         int hp_sz_idx, socket_id;
1750
1751         memset(used_hp, 0, sizeof(used_hp));
1752
1753         for (hp_sz_idx = 0;
1754                         hp_sz_idx < (int) internal_config.num_hugepage_sizes;
1755                         hp_sz_idx++) {
1756 #ifndef RTE_ARCH_64
1757                 struct hugepage_info dummy;
1758                 unsigned int i;
1759 #endif
1760                 /* also initialize used_hp hugepage sizes in used_hp */
1761                 struct hugepage_info *hpi;
1762                 hpi = &internal_config.hugepage_info[hp_sz_idx];
1763                 used_hp[hp_sz_idx].hugepage_sz = hpi->hugepage_sz;
1764
1765 #ifndef RTE_ARCH_64
1766                 /* for 32-bit, limit number of pages on socket to whatever we've
1767                  * preallocated, as we cannot allocate more.
1768                  */
1769                 memset(&dummy, 0, sizeof(dummy));
1770                 dummy.hugepage_sz = hpi->hugepage_sz;
1771                 if (rte_memseg_list_walk(hugepage_count_walk, &dummy) < 0)
1772                         return -1;
1773
1774                 for (i = 0; i < RTE_DIM(dummy.num_pages); i++) {
1775                         hpi->num_pages[i] = RTE_MIN(hpi->num_pages[i],
1776                                         dummy.num_pages[i]);
1777                 }
1778 #endif
1779         }
1780
1781         /* make a copy of socket_mem, needed for balanced allocation. */
1782         for (hp_sz_idx = 0; hp_sz_idx < RTE_MAX_NUMA_NODES; hp_sz_idx++)
1783                 memory[hp_sz_idx] = internal_config.socket_mem[hp_sz_idx];
1784
1785         /* calculate final number of pages */
1786         if (calc_num_pages_per_socket(memory,
1787                         internal_config.hugepage_info, used_hp,
1788                         internal_config.num_hugepage_sizes) < 0)
1789                 return -1;
1790
1791         for (hp_sz_idx = 0;
1792                         hp_sz_idx < (int)internal_config.num_hugepage_sizes;
1793                         hp_sz_idx++) {
1794                 for (socket_id = 0; socket_id < RTE_MAX_NUMA_NODES;
1795                                 socket_id++) {
1796                         struct rte_memseg **pages;
1797                         struct hugepage_info *hpi = &used_hp[hp_sz_idx];
1798                         unsigned int num_pages = hpi->num_pages[socket_id];
1799                         unsigned int num_pages_alloc;
1800
1801                         if (num_pages == 0)
1802                                 continue;
1803
1804                         RTE_LOG(DEBUG, EAL, "Allocating %u pages of size %" PRIu64 "M on socket %i\n",
1805                                 num_pages, hpi->hugepage_sz >> 20, socket_id);
1806
1807                         /* we may not be able to allocate all pages in one go,
1808                          * because we break up our memory map into multiple
1809                          * memseg lists. therefore, try allocating multiple
1810                          * times and see if we can get the desired number of
1811                          * pages from multiple allocations.
1812                          */
1813
1814                         num_pages_alloc = 0;
1815                         do {
1816                                 int i, cur_pages, needed;
1817
1818                                 needed = num_pages - num_pages_alloc;
1819
1820                                 pages = malloc(sizeof(*pages) * needed);
1821
1822                                 /* do not request exact number of pages */
1823                                 cur_pages = eal_memalloc_alloc_seg_bulk(pages,
1824                                                 needed, hpi->hugepage_sz,
1825                                                 socket_id, false);
1826                                 if (cur_pages <= 0) {
1827                                         free(pages);
1828                                         return -1;
1829                                 }
1830
1831                                 /* mark preallocated pages as unfreeable */
1832                                 for (i = 0; i < cur_pages; i++) {
1833                                         struct rte_memseg *ms = pages[i];
1834                                         ms->flags |= RTE_MEMSEG_FLAG_DO_NOT_FREE;
1835                                 }
1836                                 free(pages);
1837
1838                                 num_pages_alloc += cur_pages;
1839                         } while (num_pages_alloc != num_pages);
1840                 }
1841         }
1842         /* if socket limits were specified, set them */
1843         if (internal_config.force_socket_limits) {
1844                 unsigned int i;
1845                 for (i = 0; i < RTE_MAX_NUMA_NODES; i++) {
1846                         uint64_t limit = internal_config.socket_limit[i];
1847                         if (limit == 0)
1848                                 continue;
1849                         if (rte_mem_alloc_validator_register("socket-limit",
1850                                         limits_callback, i, limit))
1851                                 RTE_LOG(ERR, EAL, "Failed to register socket limits validator callback\n");
1852                 }
1853         }
1854         return 0;
1855 }
1856
1857 /*
1858  * uses fstat to report the size of a file on disk
1859  */
1860 static off_t
1861 getFileSize(int fd)
1862 {
1863         struct stat st;
1864         if (fstat(fd, &st) < 0)
1865                 return 0;
1866         return st.st_size;
1867 }
1868
1869 /*
1870  * This creates the memory mappings in the secondary process to match that of
1871  * the server process. It goes through each memory segment in the DPDK runtime
1872  * configuration and finds the hugepages which form that segment, mapping them
1873  * in order to form a contiguous block in the virtual memory space
1874  */
1875 static int
1876 eal_legacy_hugepage_attach(void)
1877 {
1878         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1879         struct hugepage_file *hp = NULL;
1880         unsigned int num_hp = 0;
1881         unsigned int i = 0;
1882         unsigned int cur_seg;
1883         off_t size = 0;
1884         int fd, fd_hugepage = -1;
1885
1886         if (aslr_enabled() > 0) {
1887                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1888                                 "(ASLR) is enabled in the kernel.\n");
1889                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1890                                 "into secondary processes\n");
1891         }
1892
1893         fd_hugepage = open(eal_hugepage_data_path(), O_RDONLY);
1894         if (fd_hugepage < 0) {
1895                 RTE_LOG(ERR, EAL, "Could not open %s\n",
1896                                 eal_hugepage_data_path());
1897                 goto error;
1898         }
1899
1900         size = getFileSize(fd_hugepage);
1901         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1902         if (hp == MAP_FAILED) {
1903                 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1904                                 eal_hugepage_data_path());
1905                 goto error;
1906         }
1907
1908         num_hp = size / sizeof(struct hugepage_file);
1909         RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1910
1911         /* map all segments into memory to make sure we get the addrs. the
1912          * segments themselves are already in memseg list (which is shared and
1913          * has its VA space already preallocated), so we just need to map
1914          * everything into correct addresses.
1915          */
1916         for (i = 0; i < num_hp; i++) {
1917                 struct hugepage_file *hf = &hp[i];
1918                 size_t map_sz = hf->size;
1919                 void *map_addr = hf->final_va;
1920                 int msl_idx, ms_idx;
1921                 struct rte_memseg_list *msl;
1922                 struct rte_memseg *ms;
1923
1924                 /* if size is zero, no more pages left */
1925                 if (map_sz == 0)
1926                         break;
1927
1928                 fd = open(hf->filepath, O_RDWR);
1929                 if (fd < 0) {
1930                         RTE_LOG(ERR, EAL, "Could not open %s: %s\n",
1931                                 hf->filepath, strerror(errno));
1932                         goto error;
1933                 }
1934
1935                 map_addr = mmap(map_addr, map_sz, PROT_READ | PROT_WRITE,
1936                                 MAP_SHARED | MAP_FIXED, fd, 0);
1937                 if (map_addr == MAP_FAILED) {
1938                         RTE_LOG(ERR, EAL, "Could not map %s: %s\n",
1939                                 hf->filepath, strerror(errno));
1940                         goto fd_error;
1941                 }
1942
1943                 /* set shared lock on the file. */
1944                 if (flock(fd, LOCK_SH) < 0) {
1945                         RTE_LOG(DEBUG, EAL, "%s(): Locking file failed: %s\n",
1946                                 __func__, strerror(errno));
1947                         goto mmap_error;
1948                 }
1949
1950                 /* find segment data */
1951                 msl = rte_mem_virt2memseg_list(map_addr);
1952                 if (msl == NULL) {
1953                         RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg list\n",
1954                                 __func__);
1955                         goto mmap_error;
1956                 }
1957                 ms = rte_mem_virt2memseg(map_addr, msl);
1958                 if (ms == NULL) {
1959                         RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg\n",
1960                                 __func__);
1961                         goto mmap_error;
1962                 }
1963
1964                 msl_idx = msl - mcfg->memsegs;
1965                 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
1966                 if (ms_idx < 0) {
1967                         RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg idx\n",
1968                                 __func__);
1969                         goto mmap_error;
1970                 }
1971
1972                 /* store segment fd internally */
1973                 if (eal_memalloc_set_seg_fd(msl_idx, ms_idx, fd) < 0)
1974                         RTE_LOG(ERR, EAL, "Could not store segment fd: %s\n",
1975                                 rte_strerror(rte_errno));
1976         }
1977         /* unmap the hugepage config file, since we are done using it */
1978         munmap(hp, size);
1979         close(fd_hugepage);
1980         return 0;
1981
1982 mmap_error:
1983         munmap(hp[i].final_va, hp[i].size);
1984 fd_error:
1985         close(fd);
1986 error:
1987         /* unwind mmap's done so far */
1988         for (cur_seg = 0; cur_seg < i; cur_seg++)
1989                 munmap(hp[cur_seg].final_va, hp[cur_seg].size);
1990
1991         if (hp != NULL && hp != MAP_FAILED)
1992                 munmap(hp, size);
1993         if (fd_hugepage >= 0)
1994                 close(fd_hugepage);
1995         return -1;
1996 }
1997
1998 static int
1999 eal_hugepage_attach(void)
2000 {
2001         if (eal_memalloc_sync_with_primary()) {
2002                 RTE_LOG(ERR, EAL, "Could not map memory from primary process\n");
2003                 if (aslr_enabled() > 0)
2004                         RTE_LOG(ERR, EAL, "It is recommended to disable ASLR in the kernel and retry running both primary and secondary processes\n");
2005                 return -1;
2006         }
2007         return 0;
2008 }
2009
2010 int
2011 rte_eal_hugepage_init(void)
2012 {
2013         return internal_config.legacy_mem ?
2014                         eal_legacy_hugepage_init() :
2015                         eal_hugepage_init();
2016 }
2017
2018 int
2019 rte_eal_hugepage_attach(void)
2020 {
2021         return internal_config.legacy_mem ?
2022                         eal_legacy_hugepage_attach() :
2023                         eal_hugepage_attach();
2024 }
2025
2026 int
2027 rte_eal_using_phys_addrs(void)
2028 {
2029         if (phys_addrs_available == -1) {
2030                 uint64_t tmp = 0;
2031
2032                 if (rte_eal_has_hugepages() != 0 &&
2033                     rte_mem_virt2phy(&tmp) != RTE_BAD_PHYS_ADDR)
2034                         phys_addrs_available = 1;
2035                 else
2036                         phys_addrs_available = 0;
2037         }
2038         return phys_addrs_available;
2039 }
2040
2041 static int __rte_unused
2042 memseg_primary_init_32(void)
2043 {
2044         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
2045         int active_sockets, hpi_idx, msl_idx = 0;
2046         unsigned int socket_id, i;
2047         struct rte_memseg_list *msl;
2048         uint64_t extra_mem_per_socket, total_extra_mem, total_requested_mem;
2049         uint64_t max_mem;
2050
2051         /* no-huge does not need this at all */
2052         if (internal_config.no_hugetlbfs)
2053                 return 0;
2054
2055         /* this is a giant hack, but desperate times call for desperate
2056          * measures. in legacy 32-bit mode, we cannot preallocate VA space,
2057          * because having upwards of 2 gigabytes of VA space already mapped will
2058          * interfere with our ability to map and sort hugepages.
2059          *
2060          * therefore, in legacy 32-bit mode, we will be initializing memseg
2061          * lists much later - in eal_memory.c, right after we unmap all the
2062          * unneeded pages. this will not affect secondary processes, as those
2063          * should be able to mmap the space without (too many) problems.
2064          */
2065         if (internal_config.legacy_mem)
2066                 return 0;
2067
2068         /* 32-bit mode is a very special case. we cannot know in advance where
2069          * the user will want to allocate their memory, so we have to do some
2070          * heuristics.
2071          */
2072         active_sockets = 0;
2073         total_requested_mem = 0;
2074         if (internal_config.force_sockets)
2075                 for (i = 0; i < rte_socket_count(); i++) {
2076                         uint64_t mem;
2077
2078                         socket_id = rte_socket_id_by_idx(i);
2079                         mem = internal_config.socket_mem[socket_id];
2080
2081                         if (mem == 0)
2082                                 continue;
2083
2084                         active_sockets++;
2085                         total_requested_mem += mem;
2086                 }
2087         else
2088                 total_requested_mem = internal_config.memory;
2089
2090         max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
2091         if (total_requested_mem > max_mem) {
2092                 RTE_LOG(ERR, EAL, "Invalid parameters: 32-bit process can at most use %uM of memory\n",
2093                                 (unsigned int)(max_mem >> 20));
2094                 return -1;
2095         }
2096         total_extra_mem = max_mem - total_requested_mem;
2097         extra_mem_per_socket = active_sockets == 0 ? total_extra_mem :
2098                         total_extra_mem / active_sockets;
2099
2100         /* the allocation logic is a little bit convoluted, but here's how it
2101          * works, in a nutshell:
2102          *  - if user hasn't specified on which sockets to allocate memory via
2103          *    --socket-mem, we allocate all of our memory on master core socket.
2104          *  - if user has specified sockets to allocate memory on, there may be
2105          *    some "unused" memory left (e.g. if user has specified --socket-mem
2106          *    such that not all memory adds up to 2 gigabytes), so add it to all
2107          *    sockets that are in use equally.
2108          *
2109          * page sizes are sorted by size in descending order, so we can safely
2110          * assume that we dispense with bigger page sizes first.
2111          */
2112
2113         /* create memseg lists */
2114         for (i = 0; i < rte_socket_count(); i++) {
2115                 int hp_sizes = (int) internal_config.num_hugepage_sizes;
2116                 uint64_t max_socket_mem, cur_socket_mem;
2117                 unsigned int master_lcore_socket;
2118                 struct rte_config *cfg = rte_eal_get_configuration();
2119                 bool skip;
2120
2121                 socket_id = rte_socket_id_by_idx(i);
2122
2123 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
2124                 /* we can still sort pages by socket in legacy mode */
2125                 if (!internal_config.legacy_mem && socket_id > 0)
2126                         break;
2127 #endif
2128
2129                 /* if we didn't specifically request memory on this socket */
2130                 skip = active_sockets != 0 &&
2131                                 internal_config.socket_mem[socket_id] == 0;
2132                 /* ...or if we didn't specifically request memory on *any*
2133                  * socket, and this is not master lcore
2134                  */
2135                 master_lcore_socket = rte_lcore_to_socket_id(cfg->master_lcore);
2136                 skip |= active_sockets == 0 && socket_id != master_lcore_socket;
2137
2138                 if (skip) {
2139                         RTE_LOG(DEBUG, EAL, "Will not preallocate memory on socket %u\n",
2140                                         socket_id);
2141                         continue;
2142                 }
2143
2144                 /* max amount of memory on this socket */
2145                 max_socket_mem = (active_sockets != 0 ?
2146                                         internal_config.socket_mem[socket_id] :
2147                                         internal_config.memory) +
2148                                         extra_mem_per_socket;
2149                 cur_socket_mem = 0;
2150
2151                 for (hpi_idx = 0; hpi_idx < hp_sizes; hpi_idx++) {
2152                         uint64_t max_pagesz_mem, cur_pagesz_mem = 0;
2153                         uint64_t hugepage_sz;
2154                         struct hugepage_info *hpi;
2155                         int type_msl_idx, max_segs, total_segs = 0;
2156
2157                         hpi = &internal_config.hugepage_info[hpi_idx];
2158                         hugepage_sz = hpi->hugepage_sz;
2159
2160                         /* check if pages are actually available */
2161                         if (hpi->num_pages[socket_id] == 0)
2162                                 continue;
2163
2164                         max_segs = RTE_MAX_MEMSEG_PER_TYPE;
2165                         max_pagesz_mem = max_socket_mem - cur_socket_mem;
2166
2167                         /* make it multiple of page size */
2168                         max_pagesz_mem = RTE_ALIGN_FLOOR(max_pagesz_mem,
2169                                         hugepage_sz);
2170
2171                         RTE_LOG(DEBUG, EAL, "Attempting to preallocate "
2172                                         "%" PRIu64 "M on socket %i\n",
2173                                         max_pagesz_mem >> 20, socket_id);
2174
2175                         type_msl_idx = 0;
2176                         while (cur_pagesz_mem < max_pagesz_mem &&
2177                                         total_segs < max_segs) {
2178                                 uint64_t cur_mem;
2179                                 unsigned int n_segs;
2180
2181                                 if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
2182                                         RTE_LOG(ERR, EAL,
2183                                                 "No more space in memseg lists, please increase %s\n",
2184                                                 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2185                                         return -1;
2186                                 }
2187
2188                                 msl = &mcfg->memsegs[msl_idx];
2189
2190                                 cur_mem = get_mem_amount(hugepage_sz,
2191                                                 max_pagesz_mem);
2192                                 n_segs = cur_mem / hugepage_sz;
2193
2194                                 if (alloc_memseg_list(msl, hugepage_sz, n_segs,
2195                                                 socket_id, type_msl_idx)) {
2196                                         /* failing to allocate a memseg list is
2197                                          * a serious error.
2198                                          */
2199                                         RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
2200                                         return -1;
2201                                 }
2202
2203                                 if (alloc_va_space(msl)) {
2204                                         /* if we couldn't allocate VA space, we
2205                                          * can try with smaller page sizes.
2206                                          */
2207                                         RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list, retrying with different page size\n");
2208                                         /* deallocate memseg list */
2209                                         if (free_memseg_list(msl))
2210                                                 return -1;
2211                                         break;
2212                                 }
2213
2214                                 total_segs += msl->memseg_arr.len;
2215                                 cur_pagesz_mem = total_segs * hugepage_sz;
2216                                 type_msl_idx++;
2217                                 msl_idx++;
2218                         }
2219                         cur_socket_mem += cur_pagesz_mem;
2220                 }
2221                 if (cur_socket_mem == 0) {
2222                         RTE_LOG(ERR, EAL, "Cannot allocate VA space on socket %u\n",
2223                                 socket_id);
2224                         return -1;
2225                 }
2226         }
2227
2228         return 0;
2229 }
2230
2231 static int __rte_unused
2232 memseg_primary_init(void)
2233 {
2234         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
2235         struct memtype {
2236                 uint64_t page_sz;
2237                 int socket_id;
2238         } *memtypes = NULL;
2239         int i, hpi_idx, msl_idx, ret = -1; /* fail unless told to succeed */
2240         struct rte_memseg_list *msl;
2241         uint64_t max_mem, max_mem_per_type;
2242         unsigned int max_seglists_per_type;
2243         unsigned int n_memtypes, cur_type;
2244
2245         /* no-huge does not need this at all */
2246         if (internal_config.no_hugetlbfs)
2247                 return 0;
2248
2249         /*
2250          * figuring out amount of memory we're going to have is a long and very
2251          * involved process. the basic element we're operating with is a memory
2252          * type, defined as a combination of NUMA node ID and page size (so that
2253          * e.g. 2 sockets with 2 page sizes yield 4 memory types in total).
2254          *
2255          * deciding amount of memory going towards each memory type is a
2256          * balancing act between maximum segments per type, maximum memory per
2257          * type, and number of detected NUMA nodes. the goal is to make sure
2258          * each memory type gets at least one memseg list.
2259          *
2260          * the total amount of memory is limited by RTE_MAX_MEM_MB value.
2261          *
2262          * the total amount of memory per type is limited by either
2263          * RTE_MAX_MEM_MB_PER_TYPE, or by RTE_MAX_MEM_MB divided by the number
2264          * of detected NUMA nodes. additionally, maximum number of segments per
2265          * type is also limited by RTE_MAX_MEMSEG_PER_TYPE. this is because for
2266          * smaller page sizes, it can take hundreds of thousands of segments to
2267          * reach the above specified per-type memory limits.
2268          *
2269          * additionally, each type may have multiple memseg lists associated
2270          * with it, each limited by either RTE_MAX_MEM_MB_PER_LIST for bigger
2271          * page sizes, or RTE_MAX_MEMSEG_PER_LIST segments for smaller ones.
2272          *
2273          * the number of memseg lists per type is decided based on the above
2274          * limits, and also taking number of detected NUMA nodes, to make sure
2275          * that we don't run out of memseg lists before we populate all NUMA
2276          * nodes with memory.
2277          *
2278          * we do this in three stages. first, we collect the number of types.
2279          * then, we figure out memory constraints and populate the list of
2280          * would-be memseg lists. then, we go ahead and allocate the memseg
2281          * lists.
2282          */
2283
2284         /* create space for mem types */
2285         n_memtypes = internal_config.num_hugepage_sizes * rte_socket_count();
2286         memtypes = calloc(n_memtypes, sizeof(*memtypes));
2287         if (memtypes == NULL) {
2288                 RTE_LOG(ERR, EAL, "Cannot allocate space for memory types\n");
2289                 return -1;
2290         }
2291
2292         /* populate mem types */
2293         cur_type = 0;
2294         for (hpi_idx = 0; hpi_idx < (int) internal_config.num_hugepage_sizes;
2295                         hpi_idx++) {
2296                 struct hugepage_info *hpi;
2297                 uint64_t hugepage_sz;
2298
2299                 hpi = &internal_config.hugepage_info[hpi_idx];
2300                 hugepage_sz = hpi->hugepage_sz;
2301
2302                 for (i = 0; i < (int) rte_socket_count(); i++, cur_type++) {
2303                         int socket_id = rte_socket_id_by_idx(i);
2304
2305 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
2306                         /* we can still sort pages by socket in legacy mode */
2307                         if (!internal_config.legacy_mem && socket_id > 0)
2308                                 break;
2309 #endif
2310                         memtypes[cur_type].page_sz = hugepage_sz;
2311                         memtypes[cur_type].socket_id = socket_id;
2312
2313                         RTE_LOG(DEBUG, EAL, "Detected memory type: "
2314                                 "socket_id:%u hugepage_sz:%" PRIu64 "\n",
2315                                 socket_id, hugepage_sz);
2316                 }
2317         }
2318         /* number of memtypes could have been lower due to no NUMA support */
2319         n_memtypes = cur_type;
2320
2321         /* set up limits for types */
2322         max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
2323         max_mem_per_type = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_TYPE << 20,
2324                         max_mem / n_memtypes);
2325         /*
2326          * limit maximum number of segment lists per type to ensure there's
2327          * space for memseg lists for all NUMA nodes with all page sizes
2328          */
2329         max_seglists_per_type = RTE_MAX_MEMSEG_LISTS / n_memtypes;
2330
2331         if (max_seglists_per_type == 0) {
2332                 RTE_LOG(ERR, EAL, "Cannot accommodate all memory types, please increase %s\n",
2333                         RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2334                 goto out;
2335         }
2336
2337         /* go through all mem types and create segment lists */
2338         msl_idx = 0;
2339         for (cur_type = 0; cur_type < n_memtypes; cur_type++) {
2340                 unsigned int cur_seglist, n_seglists, n_segs;
2341                 unsigned int max_segs_per_type, max_segs_per_list;
2342                 struct memtype *type = &memtypes[cur_type];
2343                 uint64_t max_mem_per_list, pagesz;
2344                 int socket_id;
2345
2346                 pagesz = type->page_sz;
2347                 socket_id = type->socket_id;
2348
2349                 /*
2350                  * we need to create segment lists for this type. we must take
2351                  * into account the following things:
2352                  *
2353                  * 1. total amount of memory we can use for this memory type
2354                  * 2. total amount of memory per memseg list allowed
2355                  * 3. number of segments needed to fit the amount of memory
2356                  * 4. number of segments allowed per type
2357                  * 5. number of segments allowed per memseg list
2358                  * 6. number of memseg lists we are allowed to take up
2359                  */
2360
2361                 /* calculate how much segments we will need in total */
2362                 max_segs_per_type = max_mem_per_type / pagesz;
2363                 /* limit number of segments to maximum allowed per type */
2364                 max_segs_per_type = RTE_MIN(max_segs_per_type,
2365                                 (unsigned int)RTE_MAX_MEMSEG_PER_TYPE);
2366                 /* limit number of segments to maximum allowed per list */
2367                 max_segs_per_list = RTE_MIN(max_segs_per_type,
2368                                 (unsigned int)RTE_MAX_MEMSEG_PER_LIST);
2369
2370                 /* calculate how much memory we can have per segment list */
2371                 max_mem_per_list = RTE_MIN(max_segs_per_list * pagesz,
2372                                 (uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20);
2373
2374                 /* calculate how many segments each segment list will have */
2375                 n_segs = RTE_MIN(max_segs_per_list, max_mem_per_list / pagesz);
2376
2377                 /* calculate how many segment lists we can have */
2378                 n_seglists = RTE_MIN(max_segs_per_type / n_segs,
2379                                 max_mem_per_type / max_mem_per_list);
2380
2381                 /* limit number of segment lists according to our maximum */
2382                 n_seglists = RTE_MIN(n_seglists, max_seglists_per_type);
2383
2384                 RTE_LOG(DEBUG, EAL, "Creating %i segment lists: "
2385                                 "n_segs:%i socket_id:%i hugepage_sz:%" PRIu64 "\n",
2386                         n_seglists, n_segs, socket_id, pagesz);
2387
2388                 /* create all segment lists */
2389                 for (cur_seglist = 0; cur_seglist < n_seglists; cur_seglist++) {
2390                         if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
2391                                 RTE_LOG(ERR, EAL,
2392                                         "No more space in memseg lists, please increase %s\n",
2393                                         RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2394                                 goto out;
2395                         }
2396                         msl = &mcfg->memsegs[msl_idx++];
2397
2398                         if (alloc_memseg_list(msl, pagesz, n_segs,
2399                                         socket_id, cur_seglist))
2400                                 goto out;
2401
2402                         if (alloc_va_space(msl)) {
2403                                 RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list\n");
2404                                 goto out;
2405                         }
2406                 }
2407         }
2408         /* we're successful */
2409         ret = 0;
2410 out:
2411         free(memtypes);
2412         return ret;
2413 }
2414
2415 static int
2416 memseg_secondary_init(void)
2417 {
2418         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
2419         int msl_idx = 0;
2420         struct rte_memseg_list *msl;
2421
2422         for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
2423
2424                 msl = &mcfg->memsegs[msl_idx];
2425
2426                 /* skip empty memseg lists */
2427                 if (msl->memseg_arr.len == 0)
2428                         continue;
2429
2430                 if (rte_fbarray_attach(&msl->memseg_arr)) {
2431                         RTE_LOG(ERR, EAL, "Cannot attach to primary process memseg lists\n");
2432                         return -1;
2433                 }
2434
2435                 /* preallocate VA space */
2436                 if (alloc_va_space(msl)) {
2437                         RTE_LOG(ERR, EAL, "Cannot preallocate VA space for hugepage memory\n");
2438                         return -1;
2439                 }
2440         }
2441
2442         return 0;
2443 }
2444
2445 int
2446 rte_eal_memseg_init(void)
2447 {
2448         /* increase rlimit to maximum */
2449         struct rlimit lim;
2450
2451         if (getrlimit(RLIMIT_NOFILE, &lim) == 0) {
2452                 /* set limit to maximum */
2453                 lim.rlim_cur = lim.rlim_max;
2454
2455                 if (setrlimit(RLIMIT_NOFILE, &lim) < 0) {
2456                         RTE_LOG(DEBUG, EAL, "Setting maximum number of open files failed: %s\n",
2457                                         strerror(errno));
2458                 } else {
2459                         RTE_LOG(DEBUG, EAL, "Setting maximum number of open files to %"
2460                                         PRIu64 "\n",
2461                                         (uint64_t)lim.rlim_cur);
2462                 }
2463         } else {
2464                 RTE_LOG(ERR, EAL, "Cannot get current resource limits\n");
2465         }
2466 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
2467         if (!internal_config.legacy_mem && rte_socket_count() > 1) {
2468                 RTE_LOG(WARNING, EAL, "DPDK is running on a NUMA system, but is compiled without NUMA support.\n");
2469                 RTE_LOG(WARNING, EAL, "This will have adverse consequences for performance and usability.\n");
2470                 RTE_LOG(WARNING, EAL, "Please use --"OPT_LEGACY_MEM" option, or recompile with NUMA support.\n");
2471         }
2472 #endif
2473
2474         return rte_eal_process_type() == RTE_PROC_PRIMARY ?
2475 #ifndef RTE_ARCH_64
2476                         memseg_primary_init_32() :
2477 #else
2478                         memseg_primary_init() :
2479 #endif
2480                         memseg_secondary_init();
2481 }