lib/librte_eal/linuxapp/eal/eal_memory.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2012 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  *
  33  */
  34
  35 #include <errno.h>
  36 #include <stdarg.h>
  37 #include <stdlib.h>
  38 #include <stdio.h>
  39 #include <stdint.h>
  40 #include <inttypes.h>
  41 #include <string.h>
  42 #include <stdarg.h>
  43 #include <sys/mman.h>
  44 #include <sys/types.h>
  45 #include <sys/stat.h>
  46 #include <sys/queue.h>
  47 #include <fcntl.h>
  48 #include <unistd.h>
  49 #include <limits.h>
  50 #include <errno.h>
  51 #include <sys/ioctl.h>
  52
  53 #include <rte_log.h>
  54 #include <rte_memory.h>
  55 #include <rte_memzone.h>
  56 #include <rte_launch.h>
  57 #include <rte_tailq.h>
  58 #include <rte_eal.h>
  59 #include <rte_per_lcore.h>
  60 #include <rte_lcore.h>
  61 #include <rte_common.h>
  62 #include <rte_string_fns.h>
  63
  64 #include "eal_private.h"
  65 #include "eal_internal_cfg.h"
  66 #include "eal_fs_paths.h"
  67 #include "eal_hugepages.h"
  68
  69 /**
  70  * @file
  71  * Huge page mapping under linux
  72  *
  73  * To reserve a big contiguous amount of memory, we use the hugepage
  74  * feature of linux. For that, we need to have hugetlbfs mounted. This
  75  * code will create many files in this directory (one per page) and
  76  * map them in virtual memory. For each page, we will retrieve its
  77  * physical address and remap it in order to have a virtual contiguous
  78  * zone as well as a physical contiguous zone.
  79  */
  80
  81
  82 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
  83
  84 /*
  85  * Check whether address-space layout randomization is enabled in
  86  * the kernel. This is important for multi-process as it can prevent
  87  * two processes mapping data to the same virtual address
  88  * Returns:
  89  *    0 - address space randomization disabled
  90  *    1/2 - address space randomization enabled
  91  *    negative error code on error
  92  */
  93 static int
  94 aslr_enabled(void)
  95 {
  96         char c;
  97         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
  98         if (fd < 0)
  99                 return -errno;
 100         retval = read(fd, &c, 1);
 101         close(fd);
 102         if (retval < 0)
 103                 return -errno;
 104         if (retval == 0)
 105                 return -EIO;
 106         switch (c) {
 107                 case '0' : return 0;
 108                 case '1' : return 1;
 109                 case '2' : return 2;
 110                 default: return -EINVAL;
 111         }
 112 }
 113
 114 /*
 115  * Try to mmap *size bytes in /dev/zero. If it is succesful, return the
 116  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 117  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 118  * 0. In this case, return NULL. Note: this function returns an address
 119  * which is a multiple of hugepage size.
 120  */
 121 static void *
 122 get_virtual_area(uint64_t *size, uint64_t hugepage_sz)
 123 {
 124         void *addr;
 125         int fd;
 126         long aligned_addr;
 127
 128         RTE_LOG(INFO, EAL, "Ask a virtual area of 0x%"PRIx64" bytes\n", *size);
 129
 130         fd = open("/dev/zero", O_RDONLY);
 131         if (fd < 0){
 132                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 133                 return NULL;
 134         }
 135         do {
 136                 addr = mmap(NULL, (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
 137                 if (addr == MAP_FAILED)
 138                         *size -= hugepage_sz;
 139         } while (addr == MAP_FAILED && *size > 0);
 140
 141         if (addr == MAP_FAILED) {
 142                 close(fd);
 143                 RTE_LOG(INFO, EAL, "Cannot get a virtual area\n");
 144                 return NULL;
 145         }
 146
 147         munmap(addr, (*size) + hugepage_sz);
 148         close(fd);
 149
 150         /* align addr to a huge page size boundary */
 151         aligned_addr = (long)addr;
 152         aligned_addr += (hugepage_sz - 1);
 153         aligned_addr &= (~(hugepage_sz - 1));
 154         addr = (void *)(aligned_addr);
 155
 156         RTE_LOG(INFO, EAL, "Virtual area found at %p (size = 0x%"PRIx64")\n",
 157                 addr, *size);
 158
 159         return addr;
 160 }
 161
 162 /*
 163  * Mmap all hugepages of hugepage table: it first open a file in
 164  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 165  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 166  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 167  * map continguous physical blocks in contiguous virtual blocks.
 168  */
 169 static int
 170 map_all_hugepages(struct hugepage *hugepg_tbl,
 171                 struct hugepage_info *hpi, int orig)
 172 {
 173         int fd;
 174         unsigned i;
 175         void *virtaddr;
 176         void *vma_addr = NULL;
 177         uint64_t vma_len = 0;
 178
 179         for (i = 0; i < hpi->num_pages; i++) {
 180                 uint64_t hugepage_sz = hpi->hugepage_sz;
 181
 182                 if (orig) {
 183                         hugepg_tbl[i].file_id = i;
 184                         hugepg_tbl[i].size = hugepage_sz;
 185                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 186                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 187                                         hugepg_tbl[i].file_id);
 188                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 189                 }
 190 #ifndef RTE_ARCH_X86_64
 191                 /* for 32-bit systems, don't remap 1G pages, just reuse original
 192                  * map address as final map address.
 193                  */
 194                 else if (hugepage_sz == RTE_PGSIZE_1G){
 195                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 196                         hugepg_tbl[i].orig_va = NULL;
 197                         continue;
 198                 }
 199 #endif
 200                 else if (vma_len == 0) {
 201                         unsigned j, num_pages;
 202
 203                         /* reserve a virtual area for next contiguous
 204                          * physical block: count the number of
 205                          * contiguous physical pages. */
 206                         for (j = i+1; j < hpi->num_pages ; j++) {
 207                                 if (hugepg_tbl[j].physaddr !=
 208                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 209                                         break;
 210                         }
 211                         num_pages = j - i;
 212                         vma_len = num_pages * hugepage_sz;
 213
 214                         /* get the biggest virtual memory area up to
 215                          * vma_len. If it fails, vma_addr is NULL, so
 216                          * let the kernel provide the address. */
 217                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 218                         if (vma_addr == NULL)
 219                                 vma_len = hugepage_sz;
 220                 }
 221
 222                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
 223                 if (fd < 0) {
 224                         RTE_LOG(ERR, EAL, "%s(): open failed: %s", __func__,
 225                                         strerror(errno));
 226                         return -1;
 227                 }
 228
 229                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 230                                 MAP_SHARED, fd, 0);
 231                 if (virtaddr == MAP_FAILED) {
 232                         RTE_LOG(ERR, EAL, "%s(): mmap failed: %s", __func__,
 233                                         strerror(errno));
 234                         close(fd);
 235                         return -1;
 236                 }
 237                 if (orig) {
 238                         hugepg_tbl[i].orig_va = virtaddr;
 239                         memset(virtaddr, 0, hugepage_sz);
 240                 }
 241                 else {
 242                         hugepg_tbl[i].final_va = virtaddr;
 243                 }
 244
 245                 vma_addr = (char *)vma_addr + hugepage_sz;
 246                 vma_len -= hugepage_sz;
 247                 close(fd);
 248         }
 249         return 0;
 250 }
 251
 252 /* Unmap all hugepages from original mapping. */
 253 static int
 254 unmap_all_hugepages_orig(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 255 {
 256         unsigned i;
 257         for (i = 0; i < hpi->num_pages; i++) {
 258                 if (hugepg_tbl[i].orig_va) {
 259                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 260                         hugepg_tbl[i].orig_va = NULL;
 261                 }
 262         }
 263         return 0;
 264 }
 265
 266 /*
 267  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 268  * it by browsing the /proc/self/pagemap special file.
 269  */
 270 static int
 271 find_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 272 {
 273         int fd;
 274         unsigned i;
 275         uint64_t page;
 276         unsigned long virt_pfn;
 277         int page_size;
 278
 279         /* standard page size */
 280         page_size = getpagesize();
 281
 282         fd = open("/proc/self/pagemap", O_RDONLY);
 283         if (fd < 0) {
 284                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s",
 285                         __func__, strerror(errno));
 286                 return -1;
 287         }
 288
 289         for (i = 0; i < hpi->num_pages; i++) {
 290                 off_t offset;
 291                 virt_pfn = (unsigned long)hugepg_tbl[i].orig_va /
 292                         page_size;
 293                 offset = sizeof(uint64_t) * virt_pfn;
 294                 if (lseek(fd, offset, SEEK_SET) != offset){
 295                         RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s",
 296                                         __func__, strerror(errno));
 297                         close(fd);
 298                         return -1;
 299                 }
 300                 if (read(fd, &page, sizeof(uint64_t)) < 0) {
 301                         RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s",
 302                                         __func__, strerror(errno));
 303                         close(fd);
 304                         return -1;
 305                 }
 306
 307                 /*
 308                  * the pfn (page frame number) are bits 0-54 (see
 309                  * pagemap.txt in linux Documentation)
 310                  */
 311                 hugepg_tbl[i].physaddr = ((page & 0x7fffffffffffffULL) * page_size);
 312         }
 313         close(fd);
 314         return 0;
 315 }
 316
 317 /*
 318  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 319  * page.
 320  */
 321 static int
 322 find_numasocket(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 323 {
 324         int socket_id;
 325         char *end, *nodestr;
 326         unsigned i, hp_count = 0;
 327         uint64_t virt_addr;
 328         char buf[BUFSIZ];
 329         char hugedir_str[PATH_MAX];
 330         FILE *f;
 331
 332         f = fopen("/proc/self/numa_maps", "r");
 333         if (f == NULL) {
 334                 RTE_LOG(INFO, EAL, "cannot open /proc/self/numa_maps,"
 335                                 "consider that all memory is in socket_id 0");
 336                 return 0;
 337         }
 338
 339         rte_snprintf(hugedir_str, sizeof(hugedir_str),
 340                         "%s/", hpi->hugedir);
 341
 342         /* parse numa map */
 343         while (fgets(buf, sizeof(buf), f) != NULL) {
 344
 345                 /* ignore non huge page */
 346                 if (strstr(buf, " huge ") == NULL &&
 347                                 strstr(buf, hugedir_str) == NULL)
 348                         continue;
 349
 350                 /* get zone addr */
 351                 virt_addr = strtoull(buf, &end, 16);
 352                 if (virt_addr == 0 || end == buf) {
 353                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 354                         goto error;
 355                 }
 356
 357                 /* get node id (socket id) */
 358                 nodestr = strstr(buf, " N");
 359                 if (nodestr == NULL) {
 360                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 361                         goto error;
 362                 }
 363                 nodestr += 2;
 364                 end = strstr(nodestr, "=");
 365                 if (end == NULL) {
 366                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 367                         goto error;
 368                 }
 369                 end[0] = '\0';
 370                 end = NULL;
 371
 372                 socket_id = strtoul(nodestr, &end, 0);
 373                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 374                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 375                         goto error;
 376                 }
 377
 378                 /* if we find this page in our mappings, set socket_id */
 379                 for (i = 0; i < hpi->num_pages; i++) {
 380                         void *va = (void *)(unsigned long)virt_addr;
 381                         if (hugepg_tbl[i].orig_va == va) {
 382                                 hugepg_tbl[i].socket_id = socket_id;
 383                                 hp_count++;
 384                         }
 385                 }
 386         }
 387         if (hp_count < hpi->num_pages)
 388                 goto error;
 389         fclose(f);
 390         return 0;
 391
 392 error:
 393         fclose(f);
 394         return -1;
 395 }
 396
 397 /*
 398  * Sort the hugepg_tbl by physical address (lower addresses first). We
 399  * use a slow algorithm, but we won't have millions of pages, and this
 400  * is only done at init time.
 401  */
 402 static int
 403 sort_by_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 404 {
 405         unsigned i, j;
 406         int smallest_idx;
 407         uint64_t smallest_addr;
 408         struct hugepage tmp;
 409
 410         for (i = 0; i < hpi->num_pages; i++) {
 411                 smallest_addr = 0;
 412                 smallest_idx = -1;
 413
 414                 /*
 415                  * browse all entries starting at 'i', and find the
 416                  * entry with the smallest addr
 417                  */
 418                 for (j=i; j<hpi->num_pages; j++) {
 419
 420                         if (smallest_addr == 0 ||
 421                             hugepg_tbl[j].physaddr < smallest_addr) {
 422                                 smallest_addr = hugepg_tbl[j].physaddr;
 423                                 smallest_idx = j;
 424                         }
 425                 }
 426
 427                 /* should not happen */
 428                 if (smallest_idx == -1) {
 429                         RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
 430                         return -1;
 431                 }
 432
 433                 /* swap the 2 entries in the table */
 434                 memcpy(&tmp, &hugepg_tbl[smallest_idx], sizeof(struct hugepage));
 435                 memcpy(&hugepg_tbl[smallest_idx], &hugepg_tbl[i],
 436                                 sizeof(struct hugepage));
 437                 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage));
 438         }
 439         return 0;
 440 }
 441
 442 /*
 443  * Uses mmap to create a shared memory area for storage of data
 444  *Used in this file to store the hugepage file map on disk
 445  */
 446 static void *
 447 create_shared_memory(const char *filename, const size_t mem_size)
 448 {
 449         void *retval;
 450         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 451         if (fd < 0)
 452                 return NULL;
 453         if (ftruncate(fd, mem_size) < 0) {
 454                 close(fd);
 455                 return NULL;
 456         }
 457         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 458         close(fd);
 459         return retval;
 460 }
 461
 462 /*
 463  * This function takes in the list of hugepage sizes and the
 464  * number of pages thereof, and calculates the best number of
 465  * pages of each size to fulfill the request for <memory> ram
 466  */
 467 static int
 468 calc_num_pages(uint64_t memory,
 469                 struct hugepage_info *hp_info,
 470                 struct hugepage_info *hp_used,
 471                 unsigned num_hp_info)
 472 {
 473         unsigned i = 0;
 474         int total_num_pages = 0;
 475         if (num_hp_info == 0)
 476                 return -1;
 477
 478         for (i = 0; i < num_hp_info; i++){
 479                 hp_used[i].hugepage_sz = hp_info[i].hugepage_sz;
 480                 hp_used[i].hugedir = hp_info[i].hugedir;
 481                 hp_used[i].num_pages = RTE_MIN(memory / hp_info[i].hugepage_sz,
 482                                 hp_info[i].num_pages);
 483
 484                 memory -= hp_used[i].num_pages * hp_used[i].hugepage_sz;
 485                 total_num_pages += hp_used[i].num_pages;
 486
 487                 /* check if we have met all memory requests */
 488                 if (memory == 0)
 489                         break;
 490                 /* check if we have any more pages left at this size, if so
 491                  * move on to next size */
 492                 if (hp_used[i].num_pages == hp_info[i].num_pages)
 493                         continue;
 494                 /* At this point we know that there are more pages available that are
 495                  * bigger than the memory we want, so lets see if we can get enough
 496                  * from other page sizes.
 497                  */
 498                 unsigned j;
 499                 uint64_t remaining_mem = 0;
 500                 for (j = i+1; j < num_hp_info; j++)
 501                         remaining_mem += hp_info[j].hugepage_sz * hp_info[j].num_pages;
 502
 503                 /* is there enough other memory, if not allocate another page and quit*/
 504                 if (remaining_mem < memory){
 505                         memory -= hp_info[i].hugepage_sz;
 506                         hp_used[i].num_pages++;
 507                         total_num_pages++;
 508                         break; /* we are done */
 509                 }
 510         }
 511         return total_num_pages;
 512 }
 513
 514 /*
 515  * Prepare physical memory mapping: fill configuration structure with
 516  * these infos, return 0 on success.
 517  *  1. map N huge pages in separate files in hugetlbfs
 518  *  2. find associated physical addr
 519  *  3. find associated NUMA socket ID
 520  *  4. sort all huge pages by physical address
 521  *  5. remap these N huge pages in the correct order
 522  *  6. unmap the first mapping
 523  *  7. fill memsegs in configuration with contiguous zones
 524  */
 525 static int
 526 rte_eal_hugepage_init(void)
 527 {
 528         struct rte_mem_config *mcfg;
 529         struct hugepage *hugepage;
 530         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
 531         int i, j, new_memseg;
 532         int nrpages;
 533         void *addr;
 534
 535         memset(used_hp, 0, sizeof(used_hp));
 536
 537         /* get pointer to global configuration */
 538         mcfg = rte_eal_get_configuration()->mem_config;
 539
 540         /* for debug purposes, hugetlbfs can be disabled */
 541         if (internal_config.no_hugetlbfs) {
 542                 addr = malloc(internal_config.memory);
 543                 mcfg->memseg[0].phys_addr = (unsigned long)addr;
 544                 mcfg->memseg[0].addr = addr;
 545                 mcfg->memseg[0].len = internal_config.memory;
 546                 mcfg->memseg[0].socket_id = 0;
 547                 return 0;
 548         }
 549
 550         nrpages = calc_num_pages(internal_config.memory,
 551                         &internal_config.hugepage_info[0], &used_hp[0],
 552                         internal_config.num_hugepage_sizes);
 553         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i++)
 554                 RTE_LOG(INFO, EAL, "Requesting %u pages of size %"PRIu64"\n",
 555                                 used_hp[i].num_pages, used_hp[i].hugepage_sz);
 556
 557         hugepage = create_shared_memory(eal_hugepage_info_path(),
 558                         nrpages * sizeof(struct hugepage));
 559         if (hugepage == NULL)
 560                 return -1;
 561         memset(hugepage, 0, nrpages * sizeof(struct hugepage));
 562
 563         unsigned hp_offset = 0; /* where we start the current page size entries */
 564         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
 565                 struct hugepage_info *hpi = &used_hp[i];
 566                 if (hpi->num_pages == 0)
 567                         continue;
 568
 569                 if (map_all_hugepages(&hugepage[hp_offset], hpi, 1) < 0){
 570                         RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
 571                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 572                         goto fail;
 573                 }
 574
 575                 if (find_physaddr(&hugepage[hp_offset], hpi) < 0){
 576                         RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
 577                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 578                         goto fail;
 579                 }
 580
 581                 if (find_numasocket(&hugepage[hp_offset], hpi) < 0){
 582                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
 583                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 584                         goto fail;
 585                 }
 586
 587                 if (sort_by_physaddr(&hugepage[hp_offset], hpi) < 0)
 588                         goto fail;
 589
 590                 if (map_all_hugepages(&hugepage[hp_offset], hpi, 0) < 0){
 591                         RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
 592                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 593                         goto fail;
 594                 }
 595
 596                 if (unmap_all_hugepages_orig(&hugepage[hp_offset], hpi) < 0)
 597                         goto fail;
 598
 599                 /* we have processed a num of hugepages of this size, so inc offset */
 600                 hp_offset += hpi->num_pages;
 601         }
 602
 603         memset(mcfg->memseg, 0, sizeof(mcfg->memseg));
 604         j = -1;
 605         for (i = 0; i < nrpages; i++) {
 606                 new_memseg = 0;
 607
 608                 /* if this is a new section, create a new memseg */
 609                 if (i == 0)
 610                         new_memseg = 1;
 611                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
 612                         new_memseg = 1;
 613                 else if (hugepage[i].size != hugepage[i-1].size)
 614                         new_memseg = 1;
 615                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
 616                          hugepage[i].size)
 617                         new_memseg = 1;
 618                 else if (((unsigned long)hugepage[i].final_va -
 619                      (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
 620                         new_memseg = 1;
 621
 622                 if (new_memseg) {
 623                         j += 1;
 624                         if (j == RTE_MAX_MEMSEG)
 625                                 break;
 626
 627                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
 628                         mcfg->memseg[j].addr = hugepage[i].final_va;
 629                         mcfg->memseg[j].len = hugepage[i].size;
 630                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
 631                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
 632                 }
 633                 /* continuation of previous memseg */
 634                 else {
 635                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
 636                 }
 637                 hugepage[i].memseg_id = j;
 638         }
 639
 640         return 0;
 641
 642
 643  fail:
 644         return -1;
 645 }
 646
 647 /*
 648  * uses fstat to report the size of a file on disk
 649  */
 650 static off_t
 651 getFileSize(int fd)
 652 {
 653         struct stat st;
 654         if (fstat(fd, &st) < 0)
 655                 return 0;
 656         return st.st_size;
 657 }
 658
 659 /*
 660  * This creates the memory mappings in the secondary process to match that of
 661  * the server process. It goes through each memory segment in the DPDK runtime
 662  * configuration and finds the hugepages which form that segment, mapping them
 663  * in order to form a contiguous block in the virtual memory space
 664  */
 665 static int
 666 rte_eal_hugepage_attach(void)
 667 {
 668         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 669         const struct hugepage *hp = NULL;
 670         unsigned num_hp = 0;
 671         unsigned i, s = 0; /* s used to track the segment number */
 672         off_t size;
 673         int fd, fd_zero = -1, fd_hugepage = -1;
 674
 675         if (aslr_enabled() > 0) {
 676                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
 677                                 "(ASLR) is enabled in the kernel.\n");
 678                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
 679                                 "into secondary processes\n");
 680         }
 681
 682         fd_zero = open("/dev/zero", O_RDONLY);
 683         if (fd_zero < 0) {
 684                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
 685                 goto error;
 686         }
 687         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
 688         if (fd_hugepage < 0) {
 689                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
 690                 goto error;
 691         }
 692
 693         size = getFileSize(fd_hugepage);
 694         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
 695         if (hp == NULL) {
 696                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
 697                 goto error;
 698         }
 699
 700         num_hp = size / sizeof(struct hugepage);
 701         RTE_LOG(DEBUG, EAL, "Analysing %u hugepages\n", num_hp);
 702
 703         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
 704                 void *addr, *base_addr;
 705                 uintptr_t offset = 0;
 706
 707                 /* fdzero is mmapped to get a contiguous block of virtual addresses
 708                  * get a block of free memory of the appropriate size -
 709                  * use mmap to attempt to get an identical address as server.
 710                  */
 711                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
 712                                 PROT_READ, MAP_PRIVATE, fd_zero, 0);
 713                 if (base_addr == MAP_FAILED || base_addr != mcfg->memseg[s].addr) {
 714                         RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
 715                                 "in /dev/zero to requested address [%p]\n",
 716                                 (unsigned long long)mcfg->memseg[s].len,
 717                                 mcfg->memseg[s].addr);
 718                         if (aslr_enabled() > 0)
 719                                 RTE_LOG(ERR, EAL, "It is recommended to disable ASLR in the kernel "
 720                                                 "and retry running both primary and secondary processes\n");
 721                         goto error;
 722                 }
 723                 /* free memory so we can map the hugepages into the space */
 724                 munmap(base_addr, mcfg->memseg[s].len);
 725
 726                 /* find the hugepages for this segment and map them
 727                  * we don't need to worry about order, as the server sorted the
 728                  * entries before it did the second mmap of them */
 729                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
 730                         if (hp[i].memseg_id == (int)s){
 731                                 fd = open(hp[i].filepath, O_RDWR);
 732                                 if (fd < 0) {
 733                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
 734                                                 hp[i].filepath);
 735                                         goto error;
 736                                 }
 737                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
 738                                                 hp[i].size, PROT_READ | PROT_WRITE,
 739                                                 MAP_SHARED | MAP_FIXED, fd, 0);
 740                                 close(fd); /* close file both on success and on failure */
 741                                 if (addr == MAP_FAILED) {
 742                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
 743                                                 hp[i].filepath);
 744                                         goto error;
 745                                 }
 746                                 offset+=hp[i].size;
 747                         }
 748                 }
 749                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
 750                                 (unsigned long long)mcfg->memseg[s].len);
 751                 s++;
 752         }
 753         close(fd_zero);
 754         close(fd_hugepage);
 755         return 0;
 756
 757 error:
 758         if (fd_zero >= 0)
 759                 close(fd_zero);
 760         if (fd_hugepage >= 0)
 761                 close(fd_hugepage);
 762         return -1;
 763 }
 764
 765 static int
 766 rte_eal_memdevice_init(void)
 767 {
 768         struct rte_config *config;
 769
 770         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
 771                 return 0;
 772
 773         config = rte_eal_get_configuration();
 774         config->mem_config->nchannel = internal_config.force_nchannel;
 775         config->mem_config->nrank = internal_config.force_nrank;
 776
 777         return 0;
 778 }
 779
 780
 781 /* init memory subsystem */
 782 int
 783 rte_eal_memory_init(void)
 784 {
 785         const int retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
 786                         rte_eal_hugepage_init() :
 787                         rte_eal_hugepage_attach();
 788         if (retval < 0)
 789                 return -1;
 790
 791         if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)
 792                 return -1;
 793
 794         return 0;
 795 }