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  *  version: DPDK.L.1.2.3-3
  34  */
  35
  36 #include <errno.h>
  37 #include <stdarg.h>
  38 #include <stdlib.h>
  39 #include <stdio.h>
  40 #include <stdint.h>
  41 #include <inttypes.h>
  42 #include <string.h>
  43 #include <stdarg.h>
  44 #include <sys/mman.h>
  45 #include <sys/types.h>
  46 #include <sys/stat.h>
  47 #include <sys/queue.h>
  48 #include <fcntl.h>
  49 #include <unistd.h>
  50 #include <limits.h>
  51 #include <errno.h>
  52 #include <sys/ioctl.h>
  53
  54 #include <rte_log.h>
  55 #include <rte_memory.h>
  56 #include <rte_memzone.h>
  57 #include <rte_launch.h>
  58 #include <rte_tailq.h>
  59 #include <rte_eal.h>
  60 #include <rte_per_lcore.h>
  61 #include <rte_lcore.h>
  62 #include <rte_common.h>
  63 #include <rte_string_fns.h>
  64
  65 #include "eal_private.h"
  66 #include "eal_internal_cfg.h"
  67 #include "eal_fs_paths.h"
  68 #include "eal_hugepages.h"
  69
  70 /**
  71  * @file
  72  * Huge page mapping under linux
  73  *
  74  * To reserve a big contiguous amount of memory, we use the hugepage
  75  * feature of linux. For that, we need to have hugetlbfs mounted. This
  76  * code will create many files in this directory (one per page) and
  77  * map them in virtual memory. For each page, we will retrieve its
  78  * physical address and remap it in order to have a virtual contiguous
  79  * zone as well as a physical contiguous zone.
  80  */
  81
  82
  83 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
  84
  85 /*
  86  * Check whether address-space layout randomization is enabled in
  87  * the kernel. This is important for multi-process as it can prevent
  88  * two processes mapping data to the same virtual address
  89  * Returns:
  90  *    0 - address space randomization disabled
  91  *    1/2 - address space randomization enabled
  92  *    negative error code on error
  93  */
  94 static int
  95 aslr_enabled(void)
  96 {
  97         char c;
  98         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
  99         if (fd < 0)
 100                 return -errno;
 101         retval = read(fd, &c, 1);
 102         close(fd);
 103         if (retval < 0)
 104                 return -errno;
 105         if (retval == 0)
 106                 return -EIO;
 107         switch (c) {
 108                 case '0' : return 0;
 109                 case '1' : return 1;
 110                 case '2' : return 2;
 111                 default: return -EINVAL;
 112         }
 113 }
 114
 115 /*
 116  * Try to mmap *size bytes in /dev/zero. If it is succesful, return the
 117  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 118  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 119  * 0. In this case, return NULL. Note: this function returns an address
 120  * which is a multiple of hugepage size.
 121  */
 122 static void *
 123 get_virtual_area(uint64_t *size, uint64_t hugepage_sz)
 124 {
 125         void *addr;
 126         int fd;
 127         long aligned_addr;
 128
 129         RTE_LOG(INFO, EAL, "Ask a virtual area of 0x%"PRIx64" bytes\n", *size);
 130
 131         fd = open("/dev/zero", O_RDONLY);
 132         if (fd < 0){
 133                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 134                 return NULL;
 135         }
 136         do {
 137                 addr = mmap(NULL, (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
 138                 if (addr == MAP_FAILED)
 139                         *size -= hugepage_sz;
 140         } while (addr == MAP_FAILED && *size > 0);
 141
 142         if (addr == MAP_FAILED) {
 143                 close(fd);
 144                 RTE_LOG(INFO, EAL, "Cannot get a virtual area\n");
 145                 return NULL;
 146         }
 147
 148         munmap(addr, (*size) + hugepage_sz);
 149         close(fd);
 150
 151         /* align addr to a huge page size boundary */
 152         aligned_addr = (long)addr;
 153         aligned_addr += (hugepage_sz - 1);
 154         aligned_addr &= (~(hugepage_sz - 1));
 155         addr = (void *)(aligned_addr);
 156
 157         RTE_LOG(INFO, EAL, "Virtual area found at %p (size = 0x%"PRIx64")\n",
 158                 addr, *size);
 159
 160         return addr;
 161 }
 162
 163 /*
 164  * Mmap all hugepages of hugepage table: it first open a file in
 165  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 166  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 167  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 168  * map continguous physical blocks in contiguous virtual blocks.
 169  */
 170 static int
 171 map_all_hugepages(struct hugepage *hugepg_tbl,
 172                 struct hugepage_info *hpi, int orig)
 173 {
 174         int fd;
 175         unsigned i;
 176         void *virtaddr;
 177         void *vma_addr = NULL;
 178         uint64_t vma_len = 0;
 179
 180         for (i = 0; i < hpi->num_pages; i++) {
 181                 uint64_t hugepage_sz = hpi->hugepage_sz;
 182
 183                 if (orig) {
 184                         hugepg_tbl[i].file_id = i;
 185                         hugepg_tbl[i].size = hugepage_sz;
 186                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 187                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 188                                         hugepg_tbl[i].file_id);
 189                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 190                 }
 191 #ifndef RTE_ARCH_X86_64
 192                 /* for 32-bit systems, don't remap 1G pages, just reuse original
 193                  * map address as final map address.
 194                  */
 195                 else if (hugepage_sz == RTE_PGSIZE_1G){
 196                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 197                         hugepg_tbl[i].orig_va = NULL;
 198                         continue;
 199                 }
 200 #endif
 201                 else if (vma_len == 0) {
 202                         unsigned j, num_pages;
 203
 204                         /* reserve a virtual area for next contiguous
 205                          * physical block: count the number of
 206                          * contiguous physical pages. */
 207                         for (j = i+1; j < hpi->num_pages ; j++) {
 208                                 if (hugepg_tbl[j].physaddr !=
 209                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 210                                         break;
 211                         }
 212                         num_pages = j - i;
 213                         vma_len = num_pages * hugepage_sz;
 214
 215                         /* get the biggest virtual memory area up to
 216                          * vma_len. If it fails, vma_addr is NULL, so
 217                          * let the kernel provide the address. */
 218                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 219                         if (vma_addr == NULL)
 220                                 vma_len = hugepage_sz;
 221                 }
 222
 223                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
 224                 if (fd < 0) {
 225                         RTE_LOG(ERR, EAL, "%s(): open failed: %s", __func__,
 226                                         strerror(errno));
 227                         return -1;
 228                 }
 229
 230                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 231                                 MAP_SHARED, fd, 0);
 232                 if (virtaddr == MAP_FAILED) {
 233                         RTE_LOG(ERR, EAL, "%s(): mmap failed: %s", __func__,
 234                                         strerror(errno));
 235                         close(fd);
 236                         return -1;
 237                 }
 238                 if (orig) {
 239                         hugepg_tbl[i].orig_va = virtaddr;
 240                         memset(virtaddr, 0, hugepage_sz);
 241                 }
 242                 else {
 243                         hugepg_tbl[i].final_va = virtaddr;
 244                 }
 245
 246                 vma_addr = (char *)vma_addr + hugepage_sz;
 247                 vma_len -= hugepage_sz;
 248                 close(fd);
 249         }
 250         return 0;
 251 }
 252
 253 /* Unmap all hugepages from original mapping. */
 254 static int
 255 unmap_all_hugepages_orig(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 256 {
 257         unsigned i;
 258         for (i = 0; i < hpi->num_pages; i++) {
 259                 if (hugepg_tbl[i].orig_va) {
 260                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 261                         hugepg_tbl[i].orig_va = NULL;
 262                 }
 263         }
 264         return 0;
 265 }
 266
 267 /*
 268  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 269  * it by browsing the /proc/self/pagemap special file.
 270  */
 271 static int
 272 find_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 273 {
 274         int fd;
 275         unsigned i;
 276         uint64_t page;
 277         unsigned long virt_pfn;
 278         int page_size;
 279
 280         /* standard page size */
 281         page_size = getpagesize();
 282
 283         fd = open("/proc/self/pagemap", O_RDONLY);
 284         if (fd < 0) {
 285                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s",
 286                         __func__, strerror(errno));
 287                 return -1;
 288         }
 289
 290         for (i = 0; i < hpi->num_pages; i++) {
 291                 off_t offset;
 292                 virt_pfn = (unsigned long)hugepg_tbl[i].orig_va /
 293                         page_size;
 294                 offset = sizeof(uint64_t) * virt_pfn;
 295                 if (lseek(fd, offset, SEEK_SET) != offset){
 296                         RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s",
 297                                         __func__, strerror(errno));
 298                         close(fd);
 299                         return -1;
 300                 }
 301                 if (read(fd, &page, sizeof(uint64_t)) < 0) {
 302                         RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s",
 303                                         __func__, strerror(errno));
 304                         close(fd);
 305                         return -1;
 306                 }
 307
 308                 /*
 309                  * the pfn (page frame number) are bits 0-54 (see
 310                  * pagemap.txt in linux Documentation)
 311                  */
 312                 hugepg_tbl[i].physaddr = ((page & 0x7fffffffffffffULL) * page_size);
 313         }
 314         close(fd);
 315         return 0;
 316 }
 317
 318 /*
 319  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 320  * page.
 321  */
 322 static int
 323 find_numasocket(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 324 {
 325         int socket_id;
 326         char *end, *nodestr;
 327         unsigned i, hp_count = 0;
 328         uint64_t virt_addr;
 329         char buf[BUFSIZ];
 330         char hugedir_str[PATH_MAX];
 331         FILE *f;
 332
 333         f = fopen("/proc/self/numa_maps", "r");
 334         if (f == NULL) {
 335                 RTE_LOG(INFO, EAL, "cannot open /proc/self/numa_maps,"
 336                                 "consider that all memory is in socket_id 0");
 337                 return 0;
 338         }
 339
 340         rte_snprintf(hugedir_str, sizeof(hugedir_str),
 341                         "%s/", hpi->hugedir);
 342
 343         /* parse numa map */
 344         while (fgets(buf, sizeof(buf), f) != NULL) {
 345
 346                 /* ignore non huge page */
 347                 if (strstr(buf, " huge ") == NULL &&
 348                                 strstr(buf, hugedir_str) == NULL)
 349                         continue;
 350
 351                 /* get zone addr */
 352                 virt_addr = strtoull(buf, &end, 16);
 353                 if (virt_addr == 0 || end == buf) {
 354                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 355                         goto error;
 356                 }
 357
 358                 /* get node id (socket id) */
 359                 nodestr = strstr(buf, " N");
 360                 if (nodestr == NULL) {
 361                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 362                         goto error;
 363                 }
 364                 nodestr += 2;
 365                 end = strstr(nodestr, "=");
 366                 if (end == NULL) {
 367                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 368                         goto error;
 369                 }
 370                 end[0] = '\0';
 371                 end = NULL;
 372
 373                 socket_id = strtoul(nodestr, &end, 0);
 374                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 375                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 376                         goto error;
 377                 }
 378
 379                 /* if we find this page in our mappings, set socket_id */
 380                 for (i = 0; i < hpi->num_pages; i++) {
 381                         void *va = (void *)(unsigned long)virt_addr;
 382                         if (hugepg_tbl[i].orig_va == va) {
 383                                 hugepg_tbl[i].socket_id = socket_id;
 384                                 hp_count++;
 385                         }
 386                 }
 387         }
 388         if (hp_count < hpi->num_pages)
 389                 goto error;
 390         fclose(f);
 391         return 0;
 392
 393 error:
 394         fclose(f);
 395         return -1;
 396 }
 397
 398 /*
 399  * Sort the hugepg_tbl by physical address (lower addresses first). We
 400  * use a slow algorithm, but we won't have millions of pages, and this
 401  * is only done at init time.
 402  */
 403 static int
 404 sort_by_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 405 {
 406         unsigned i, j;
 407         int smallest_idx;
 408         uint64_t smallest_addr;
 409         struct hugepage tmp;
 410
 411         for (i = 0; i < hpi->num_pages; i++) {
 412                 smallest_addr = 0;
 413                 smallest_idx = -1;
 414
 415                 /*
 416                  * browse all entries starting at 'i', and find the
 417                  * entry with the smallest addr
 418                  */
 419                 for (j=i; j<hpi->num_pages; j++) {
 420
 421                         if (smallest_addr == 0 ||
 422                             hugepg_tbl[j].physaddr < smallest_addr) {
 423                                 smallest_addr = hugepg_tbl[j].physaddr;
 424                                 smallest_idx = j;
 425                         }
 426                 }
 427
 428                 /* should not happen */
 429                 if (smallest_idx == -1) {
 430                         RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
 431                         return -1;
 432                 }
 433
 434                 /* swap the 2 entries in the table */
 435                 memcpy(&tmp, &hugepg_tbl[smallest_idx], sizeof(struct hugepage));
 436                 memcpy(&hugepg_tbl[smallest_idx], &hugepg_tbl[i],
 437                                 sizeof(struct hugepage));
 438                 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage));
 439         }
 440         return 0;
 441 }
 442
 443 /*
 444  * Uses mmap to create a shared memory area for storage of data
 445  *Used in this file to store the hugepage file map on disk
 446  */
 447 static void *
 448 create_shared_memory(const char *filename, const size_t mem_size)
 449 {
 450         void *retval;
 451         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 452         if (fd < 0)
 453                 return NULL;
 454         if (ftruncate(fd, mem_size) < 0) {
 455                 close(fd);
 456                 return NULL;
 457         }
 458         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 459         close(fd);
 460         return retval;
 461 }
 462
 463 /*
 464  * This function takes in the list of hugepage sizes and the
 465  * number of pages thereof, and calculates the best number of
 466  * pages of each size to fulfill the request for <memory> ram
 467  */
 468 static int
 469 calc_num_pages(uint64_t memory,
 470                 struct hugepage_info *hp_info,
 471                 struct hugepage_info *hp_used,
 472                 unsigned num_hp_info)
 473 {
 474         unsigned i = 0;
 475         int total_num_pages = 0;
 476         if (num_hp_info == 0)
 477                 return -1;
 478
 479         for (i = 0; i < num_hp_info; i++){
 480                 hp_used[i].hugepage_sz = hp_info[i].hugepage_sz;
 481                 hp_used[i].hugedir = hp_info[i].hugedir;
 482                 hp_used[i].num_pages = RTE_MIN(memory / hp_info[i].hugepage_sz,
 483                                 hp_info[i].num_pages);
 484
 485                 memory -= hp_used[i].num_pages * hp_used[i].hugepage_sz;
 486                 total_num_pages += hp_used[i].num_pages;
 487
 488                 /* check if we have met all memory requests */
 489                 if (memory == 0)
 490                         break;
 491                 /* check if we have any more pages left at this size, if so
 492                  * move on to next size */
 493                 if (hp_used[i].num_pages == hp_info[i].num_pages)
 494                         continue;
 495                 /* At this point we know that there are more pages available that are
 496                  * bigger than the memory we want, so lets see if we can get enough
 497                  * from other page sizes.
 498                  */
 499                 unsigned j;
 500                 uint64_t remaining_mem = 0;
 501                 for (j = i+1; j < num_hp_info; j++)
 502                         remaining_mem += hp_info[j].hugepage_sz * hp_info[j].num_pages;
 503
 504                 /* is there enough other memory, if not allocate another page and quit*/
 505                 if (remaining_mem < memory){
 506                         memory -= hp_info[i].hugepage_sz;
 507                         hp_used[i].num_pages++;
 508                         total_num_pages++;
 509                         break; /* we are done */
 510                 }
 511         }
 512         return total_num_pages;
 513 }
 514
 515 /*
 516  * Prepare physical memory mapping: fill configuration structure with
 517  * these infos, return 0 on success.
 518  *  1. map N huge pages in separate files in hugetlbfs
 519  *  2. find associated physical addr
 520  *  3. find associated NUMA socket ID
 521  *  4. sort all huge pages by physical address
 522  *  5. remap these N huge pages in the correct order
 523  *  6. unmap the first mapping
 524  *  7. fill memsegs in configuration with contiguous zones
 525  */
 526 static int
 527 rte_eal_hugepage_init(void)
 528 {
 529         struct rte_mem_config *mcfg;
 530         struct hugepage *hugepage;
 531         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
 532         int i, j, new_memseg;
 533         int nrpages;
 534         void *addr;
 535
 536         memset(used_hp, 0, sizeof(used_hp));
 537
 538         /* get pointer to global configuration */
 539         mcfg = rte_eal_get_configuration()->mem_config;
 540
 541         /* for debug purposes, hugetlbfs can be disabled */
 542         if (internal_config.no_hugetlbfs) {
 543                 addr = malloc(internal_config.memory);
 544                 mcfg->memseg[0].phys_addr = (unsigned long)addr;
 545                 mcfg->memseg[0].addr = addr;
 546                 mcfg->memseg[0].len = internal_config.memory;
 547                 mcfg->memseg[0].socket_id = 0;
 548                 return 0;
 549         }
 550
 551         nrpages = calc_num_pages(internal_config.memory,
 552                         &internal_config.hugepage_info[0], &used_hp[0],
 553                         internal_config.num_hugepage_sizes);
 554         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i++)
 555                 RTE_LOG(INFO, EAL, "Requesting %u pages of size %"PRIu64"\n",
 556                                 used_hp[i].num_pages, used_hp[i].hugepage_sz);
 557
 558         hugepage = create_shared_memory(eal_hugepage_info_path(),
 559                         nrpages * sizeof(struct hugepage));
 560         if (hugepage == NULL)
 561                 return -1;
 562         memset(hugepage, 0, nrpages * sizeof(struct hugepage));
 563
 564         unsigned hp_offset = 0; /* where we start the current page size entries */
 565         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
 566                 struct hugepage_info *hpi = &used_hp[i];
 567                 if (hpi->num_pages == 0)
 568                         continue;
 569
 570                 if (map_all_hugepages(&hugepage[hp_offset], hpi, 1) < 0){
 571                         RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
 572                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 573                         goto fail;
 574                 }
 575
 576                 if (find_physaddr(&hugepage[hp_offset], hpi) < 0){
 577                         RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
 578                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 579                         goto fail;
 580                 }
 581
 582                 if (find_numasocket(&hugepage[hp_offset], hpi) < 0){
 583                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
 584                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 585                         goto fail;
 586                 }
 587
 588                 if (sort_by_physaddr(&hugepage[hp_offset], hpi) < 0)
 589                         goto fail;
 590
 591                 if (map_all_hugepages(&hugepage[hp_offset], hpi, 0) < 0){
 592                         RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
 593                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 594                         goto fail;
 595                 }
 596
 597                 if (unmap_all_hugepages_orig(&hugepage[hp_offset], hpi) < 0)
 598                         goto fail;
 599
 600                 /* we have processed a num of hugepages of this size, so inc offset */
 601                 hp_offset += hpi->num_pages;
 602         }
 603
 604         memset(mcfg->memseg, 0, sizeof(mcfg->memseg));
 605         j = -1;
 606         for (i = 0; i < nrpages; i++) {
 607                 new_memseg = 0;
 608
 609                 /* if this is a new section, create a new memseg */
 610                 if (i == 0)
 611                         new_memseg = 1;
 612                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
 613                         new_memseg = 1;
 614                 else if (hugepage[i].size != hugepage[i-1].size)
 615                         new_memseg = 1;
 616                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
 617                          hugepage[i].size)
 618                         new_memseg = 1;
 619                 else if (((unsigned long)hugepage[i].final_va -
 620                      (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
 621                         new_memseg = 1;
 622
 623                 if (new_memseg) {
 624                         j += 1;
 625                         if (j == RTE_MAX_MEMSEG)
 626                                 break;
 627
 628                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
 629                         mcfg->memseg[j].addr = hugepage[i].final_va;
 630                         mcfg->memseg[j].len = hugepage[i].size;
 631                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
 632                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
 633                 }
 634                 /* continuation of previous memseg */
 635                 else {
 636                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
 637                 }
 638                 hugepage[i].memseg_id = j;
 639         }
 640
 641         return 0;
 642
 643
 644  fail:
 645         return -1;
 646 }
 647
 648 /*
 649  * uses fstat to report the size of a file on disk
 650  */
 651 static off_t
 652 getFileSize(int fd)
 653 {
 654         struct stat st;
 655         if (fstat(fd, &st) < 0)
 656                 return 0;
 657         return st.st_size;
 658 }
 659
 660 /*
 661  * This creates the memory mappings in the secondary process to match that of
 662  * the server process. It goes through each memory segment in the DPDK runtime
 663  * configuration and finds the hugepages which form that segment, mapping them
 664  * in order to form a contiguous block in the virtual memory space
 665  */
 666 static int
 667 rte_eal_hugepage_attach(void)
 668 {
 669         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 670         const struct hugepage *hp = NULL;
 671         unsigned num_hp = 0;
 672         unsigned i, s = 0; /* s used to track the segment number */
 673         off_t size;
 674         int fd, fd_zero = -1, fd_hugepage = -1;
 675
 676         if (aslr_enabled() > 0) {
 677                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
 678                                 "(ASLR) is enabled in the kernel.\n");
 679                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
 680                                 "into secondary processes\n");
 681         }
 682
 683         fd_zero = open("/dev/zero", O_RDONLY);
 684         if (fd_zero < 0) {
 685                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
 686                 goto error;
 687         }
 688         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
 689         if (fd_hugepage < 0) {
 690                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
 691                 goto error;
 692         }
 693
 694         size = getFileSize(fd_hugepage);
 695         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
 696         if (hp == NULL) {
 697                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
 698                 goto error;
 699         }
 700
 701         num_hp = size / sizeof(struct hugepage);
 702         RTE_LOG(DEBUG, EAL, "Analysing %u hugepages\n", num_hp);
 703
 704         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
 705                 void *addr, *base_addr;
 706                 uintptr_t offset = 0;
 707
 708                 /* fdzero is mmapped to get a contiguous block of virtual addresses
 709                  * get a block of free memory of the appropriate size -
 710                  * use mmap to attempt to get an identical address as server.
 711                  */
 712                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
 713                                 PROT_READ, MAP_PRIVATE, fd_zero, 0);
 714                 if (base_addr == MAP_FAILED || base_addr != mcfg->memseg[s].addr) {
 715                         RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
 716                                 "in /dev/zero to requested address [%p]\n",
 717                                 (unsigned long long)mcfg->memseg[s].len,
 718                                 mcfg->memseg[s].addr);
 719                         if (aslr_enabled() > 0)
 720                                 RTE_LOG(ERR, EAL, "It is recommended to disable ASLR in the kernel "
 721                                                 "and retry running both primary and secondary processes\n");
 722                         goto error;
 723                 }
 724                 /* free memory so we can map the hugepages into the space */
 725                 munmap(base_addr, mcfg->memseg[s].len);
 726
 727                 /* find the hugepages for this segment and map them
 728                  * we don't need to worry about order, as the server sorted the
 729                  * entries before it did the second mmap of them */
 730                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
 731                         if (hp[i].memseg_id == (int)s){
 732                                 fd = open(hp[i].filepath, O_RDWR);
 733                                 if (fd < 0) {
 734                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
 735                                                 hp[i].filepath);
 736                                         goto error;
 737                                 }
 738                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
 739                                                 hp[i].size, PROT_READ | PROT_WRITE,
 740                                                 MAP_SHARED | MAP_FIXED, fd, 0);
 741                                 close(fd); /* close file both on success and on failure */
 742                                 if (addr == MAP_FAILED) {
 743                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
 744                                                 hp[i].filepath);
 745                                         goto error;
 746                                 }
 747                                 offset+=hp[i].size;
 748                         }
 749                 }
 750                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
 751                                 (unsigned long long)mcfg->memseg[s].len);
 752                 s++;
 753         }
 754         close(fd_zero);
 755         close(fd_hugepage);
 756         return 0;
 757
 758 error:
 759         if (fd_zero >= 0)
 760                 close(fd_zero);
 761         if (fd_hugepage >= 0)
 762                 close(fd_hugepage);
 763         return -1;
 764 }
 765
 766 static int
 767 rte_eal_memdevice_init(void)
 768 {
 769         struct rte_config *config;
 770
 771         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
 772                 return 0;
 773
 774         config = rte_eal_get_configuration();
 775         config->mem_config->nchannel = internal_config.force_nchannel;
 776         config->mem_config->nrank = internal_config.force_nrank;
 777
 778         return 0;
 779 }
 780
 781
 782 /* init memory subsystem */
 783 int
 784 rte_eal_memory_init(void)
 785 {
 786         const int retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
 787                         rte_eal_hugepage_init() :
 788                         rte_eal_hugepage_attach();
 789         if (retval < 0)
 790                 return -1;
 791
 792         if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)
 793                 return -1;
 794
 795         return 0;
 796 }