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 /*   BSD LICENSE
  35  *
  36  *   Copyright(c) 2013 6WIND.
  37  *
  38  *   Redistribution and use in source and binary forms, with or without
  39  *   modification, are permitted provided that the following conditions
  40  *   are met:
  41  *
  42  *     * Redistributions of source code must retain the above copyright
  43  *       notice, this list of conditions and the following disclaimer.
  44  *     * Redistributions in binary form must reproduce the above copyright
  45  *       notice, this list of conditions and the following disclaimer in
  46  *       the documentation and/or other materials provided with the
  47  *       distribution.
  48  *     * Neither the name of 6WIND S.A. nor the names of its
  49  *       contributors may be used to endorse or promote products derived
  50  *       from this software without specific prior written permission.
  51  *
  52  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  53  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  54  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  55  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  56  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  57  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  58  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  59  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  60  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  61  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  62  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  63  */
  64
  65 #include <errno.h>
  66 #include <stdarg.h>
  67 #include <stdlib.h>
  68 #include <stdio.h>
  69 #include <stdint.h>
  70 #include <inttypes.h>
  71 #include <string.h>
  72 #include <stdarg.h>
  73 #include <sys/mman.h>
  74 #include <sys/types.h>
  75 #include <sys/stat.h>
  76 #include <sys/queue.h>
  77 #include <sys/file.h>
  78 #include <unistd.h>
  79 #include <limits.h>
  80 #include <errno.h>
  81 #include <sys/ioctl.h>
  82 #include <sys/time.h>
  83 #include <sys/resource.h>
  84
  85 #include <rte_log.h>
  86 #include <rte_memory.h>
  87 #include <rte_memzone.h>
  88 #include <rte_launch.h>
  89 #include <rte_tailq.h>
  90 #include <rte_eal.h>
  91 #include <rte_eal_memconfig.h>
  92 #include <rte_per_lcore.h>
  93 #include <rte_lcore.h>
  94 #include <rte_common.h>
  95 #include <rte_string_fns.h>
  96
  97 #include "eal_private.h"
  98 #include "eal_internal_cfg.h"
  99 #include "eal_filesystem.h"
 100 #include "eal_hugepages.h"
 101
 102 /**
 103  * @file
 104  * Huge page mapping under linux
 105  *
 106  * To reserve a big contiguous amount of memory, we use the hugepage
 107  * feature of linux. For that, we need to have hugetlbfs mounted. This
 108  * code will create many files in this directory (one per page) and
 109  * map them in virtual memory. For each page, we will retrieve its
 110  * physical address and remap it in order to have a virtual contiguous
 111  * zone as well as a physical contiguous zone.
 112  */
 113
 114
 115 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
 116
 117 /*
 118  * Check whether address-space layout randomization is enabled in
 119  * the kernel. This is important for multi-process as it can prevent
 120  * two processes mapping data to the same virtual address
 121  * Returns:
 122  *    0 - address space randomization disabled
 123  *    1/2 - address space randomization enabled
 124  *    negative error code on error
 125  */
 126 static int
 127 aslr_enabled(void)
 128 {
 129         char c;
 130         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 131         if (fd < 0)
 132                 return -errno;
 133         retval = read(fd, &c, 1);
 134         close(fd);
 135         if (retval < 0)
 136                 return -errno;
 137         if (retval == 0)
 138                 return -EIO;
 139         switch (c) {
 140                 case '0' : return 0;
 141                 case '1' : return 1;
 142                 case '2' : return 2;
 143                 default: return -EINVAL;
 144         }
 145 }
 146
 147 /*
 148  * Increase limit for open files for current process
 149  */
 150 static int
 151 increase_open_file_limit(void)
 152 {
 153         struct rlimit limit;
 154
 155         /* read current limits */
 156         if (getrlimit(RLIMIT_NOFILE, &limit) != 0) {
 157                 RTE_LOG(ERR, EAL, "Error reading resource limit: %s\n",
 158                                 strerror(errno));
 159                 return -1;
 160         }
 161
 162         /* check if current soft limit matches the hard limit */
 163         if (limit.rlim_cur < limit.rlim_max) {
 164                 /* set soft limit to match hard limit */
 165                 limit.rlim_cur = limit.rlim_max;
 166         }
 167         else {
 168                 /* we can't increase the soft limit so now we try to increase
 169                  * soft and hard limit. this might fail when run as non-root.
 170                  */
 171                 limit.rlim_cur *= 2;
 172                 limit.rlim_max *= 2;
 173         }
 174
 175         /* set current resource limit */
 176         if (setrlimit(RLIMIT_NOFILE, &limit) != 0) {
 177                 RTE_LOG(ERR, EAL, "Error increasing open files limit: %s\n",
 178                                 strerror(errno));
 179                 return -1;
 180         }
 181
 182         return 0;
 183 }
 184
 185 /*
 186  * Try to mmap *size bytes in /dev/zero. If it is succesful, return the
 187  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 188  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 189  * 0. In this case, return NULL. Note: this function returns an address
 190  * which is a multiple of hugepage size.
 191  */
 192 static void *
 193 get_virtual_area(uint64_t *size, uint64_t hugepage_sz)
 194 {
 195         void *addr;
 196         int fd;
 197         long aligned_addr;
 198
 199         RTE_LOG(INFO, EAL, "Ask a virtual area of 0x%"PRIx64" bytes\n", *size);
 200
 201         fd = open("/dev/zero", O_RDONLY);
 202         if (fd < 0){
 203                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 204                 return NULL;
 205         }
 206         do {
 207                 addr = mmap(NULL, (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
 208                 if (addr == MAP_FAILED)
 209                         *size -= hugepage_sz;
 210         } while (addr == MAP_FAILED && *size > 0);
 211
 212         if (addr == MAP_FAILED) {
 213                 close(fd);
 214                 RTE_LOG(INFO, EAL, "Cannot get a virtual area\n");
 215                 return NULL;
 216         }
 217
 218         munmap(addr, (*size) + hugepage_sz);
 219         close(fd);
 220
 221         /* align addr to a huge page size boundary */
 222         aligned_addr = (long)addr;
 223         aligned_addr += (hugepage_sz - 1);
 224         aligned_addr &= (~(hugepage_sz - 1));
 225         addr = (void *)(aligned_addr);
 226
 227         RTE_LOG(INFO, EAL, "Virtual area found at %p (size = 0x%"PRIx64")\n",
 228                 addr, *size);
 229
 230         return addr;
 231 }
 232
 233 /*
 234  * Mmap all hugepages of hugepage table: it first open a file in
 235  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 236  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 237  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 238  * map continguous physical blocks in contiguous virtual blocks.
 239  */
 240 static int
 241 map_all_hugepages(struct hugepage *hugepg_tbl,
 242                 struct hugepage_info *hpi, int orig)
 243 {
 244         int fd;
 245         unsigned i;
 246         void *virtaddr;
 247         void *vma_addr = NULL;
 248         uint64_t vma_len = 0;
 249
 250         for (i = 0; i < hpi->num_pages[0]; i++) {
 251                 uint64_t hugepage_sz = hpi->hugepage_sz;
 252
 253                 if (orig) {
 254                         hugepg_tbl[i].file_id = i;
 255                         hugepg_tbl[i].size = hugepage_sz;
 256                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 257                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 258                                         hugepg_tbl[i].file_id);
 259                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 260                 }
 261 #ifndef RTE_ARCH_X86_64
 262                 /* for 32-bit systems, don't remap 1G pages, just reuse original
 263                  * map address as final map address.
 264                  */
 265                 else if (hugepage_sz == RTE_PGSIZE_1G){
 266                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 267                         hugepg_tbl[i].orig_va = NULL;
 268                         continue;
 269                 }
 270 #endif
 271                 else if (vma_len == 0) {
 272                         unsigned j, num_pages;
 273
 274                         /* reserve a virtual area for next contiguous
 275                          * physical block: count the number of
 276                          * contiguous physical pages. */
 277                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 278                                 if (hugepg_tbl[j].physaddr !=
 279                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 280                                         break;
 281                         }
 282                         num_pages = j - i;
 283                         vma_len = num_pages * hugepage_sz;
 284
 285                         /* get the biggest virtual memory area up to
 286                          * vma_len. If it fails, vma_addr is NULL, so
 287                          * let the kernel provide the address. */
 288                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 289                         if (vma_addr == NULL)
 290                                 vma_len = hugepage_sz;
 291                 }
 292
 293                 /* try to create hugepage file */
 294                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
 295                 if (fd < 0) {
 296                         RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
 297                                         strerror(errno));
 298                         return -1;
 299                 }
 300
 301                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 302                                 MAP_SHARED, fd, 0);
 303                 if (virtaddr == MAP_FAILED) {
 304                         RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__,
 305                                         strerror(errno));
 306                         close(fd);
 307                         return -1;
 308                 }
 309
 310                 if (orig) {
 311                         hugepg_tbl[i].orig_va = virtaddr;
 312                         memset(virtaddr, 0, hugepage_sz);
 313                 }
 314                 else {
 315                         hugepg_tbl[i].final_va = virtaddr;
 316                 }
 317
 318                 /* close the file descriptor, files will be locked later */
 319                 close(fd);
 320
 321                 vma_addr = (char *)vma_addr + hugepage_sz;
 322                 vma_len -= hugepage_sz;
 323         }
 324         return 0;
 325 }
 326
 327 /* Unmap all hugepages from original mapping. */
 328 static int
 329 unmap_all_hugepages_orig(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 330 {
 331         unsigned i;
 332         for (i = 0; i < hpi->num_pages[0]; i++) {
 333                 if (hugepg_tbl[i].orig_va) {
 334                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 335                         hugepg_tbl[i].orig_va = NULL;
 336                 }
 337         }
 338         return 0;
 339 }
 340
 341 /*
 342  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 343  * it by browsing the /proc/self/pagemap special file.
 344  */
 345 static int
 346 find_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 347 {
 348         int fd;
 349         unsigned i;
 350         uint64_t page;
 351         unsigned long virt_pfn;
 352         int page_size;
 353
 354         /* standard page size */
 355         page_size = getpagesize();
 356
 357         fd = open("/proc/self/pagemap", O_RDONLY);
 358         if (fd < 0) {
 359                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 360                         __func__, strerror(errno));
 361                 return -1;
 362         }
 363
 364         for (i = 0; i < hpi->num_pages[0]; i++) {
 365                 off_t offset;
 366                 virt_pfn = (unsigned long)hugepg_tbl[i].orig_va /
 367                         page_size;
 368                 offset = sizeof(uint64_t) * virt_pfn;
 369                 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 370                         RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 371                                         __func__, strerror(errno));
 372                         close(fd);
 373                         return -1;
 374                 }
 375                 if (read(fd, &page, sizeof(uint64_t)) < 0) {
 376                         RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 377                                         __func__, strerror(errno));
 378                         close(fd);
 379                         return -1;
 380                 }
 381
 382                 /*
 383                  * the pfn (page frame number) are bits 0-54 (see
 384                  * pagemap.txt in linux Documentation)
 385                  */
 386                 hugepg_tbl[i].physaddr = ((page & 0x7fffffffffffffULL) * page_size);
 387         }
 388         close(fd);
 389         return 0;
 390 }
 391
 392 /*
 393  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 394  * page.
 395  */
 396 static int
 397 find_numasocket(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 398 {
 399         int socket_id;
 400         char *end, *nodestr;
 401         unsigned i, hp_count = 0;
 402         uint64_t virt_addr;
 403         char buf[BUFSIZ];
 404         char hugedir_str[PATH_MAX];
 405         FILE *f;
 406
 407         f = fopen("/proc/self/numa_maps", "r");
 408         if (f == NULL) {
 409                 RTE_LOG(INFO, EAL, "cannot open /proc/self/numa_maps,"
 410                                 " consider that all memory is in socket_id 0\n");
 411                 return 0;
 412         }
 413
 414         rte_snprintf(hugedir_str, sizeof(hugedir_str),
 415                         "%s/", hpi->hugedir);
 416
 417         /* parse numa map */
 418         while (fgets(buf, sizeof(buf), f) != NULL) {
 419
 420                 /* ignore non huge page */
 421                 if (strstr(buf, " huge ") == NULL &&
 422                                 strstr(buf, hugedir_str) == NULL)
 423                         continue;
 424
 425                 /* get zone addr */
 426                 virt_addr = strtoull(buf, &end, 16);
 427                 if (virt_addr == 0 || end == buf) {
 428                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 429                         goto error;
 430                 }
 431
 432                 /* get node id (socket id) */
 433                 nodestr = strstr(buf, " N");
 434                 if (nodestr == NULL) {
 435                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 436                         goto error;
 437                 }
 438                 nodestr += 2;
 439                 end = strstr(nodestr, "=");
 440                 if (end == NULL) {
 441                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 442                         goto error;
 443                 }
 444                 end[0] = '\0';
 445                 end = NULL;
 446
 447                 socket_id = strtoul(nodestr, &end, 0);
 448                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 449                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 450                         goto error;
 451                 }
 452
 453                 /* if we find this page in our mappings, set socket_id */
 454                 for (i = 0; i < hpi->num_pages[0]; i++) {
 455                         void *va = (void *)(unsigned long)virt_addr;
 456                         if (hugepg_tbl[i].orig_va == va) {
 457                                 hugepg_tbl[i].socket_id = socket_id;
 458                                 hp_count++;
 459                         }
 460                 }
 461         }
 462
 463         if (hp_count < hpi->num_pages[0])
 464                 goto error;
 465
 466         fclose(f);
 467         return 0;
 468
 469 error:
 470         fclose(f);
 471         return -1;
 472 }
 473
 474 /*
 475  * Sort the hugepg_tbl by physical address (lower addresses first). We
 476  * use a slow algorithm, but we won't have millions of pages, and this
 477  * is only done at init time.
 478  */
 479 static int
 480 sort_by_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 481 {
 482         unsigned i, j;
 483         int smallest_idx;
 484         uint64_t smallest_addr;
 485         struct hugepage tmp;
 486
 487         for (i = 0; i < hpi->num_pages[0]; i++) {
 488                 smallest_addr = 0;
 489                 smallest_idx = -1;
 490
 491                 /*
 492                  * browse all entries starting at 'i', and find the
 493                  * entry with the smallest addr
 494                  */
 495                 for (j=i; j< hpi->num_pages[0]; j++) {
 496
 497                         if (smallest_addr == 0 ||
 498                             hugepg_tbl[j].physaddr < smallest_addr) {
 499                                 smallest_addr = hugepg_tbl[j].physaddr;
 500                                 smallest_idx = j;
 501                         }
 502                 }
 503
 504                 /* should not happen */
 505                 if (smallest_idx == -1) {
 506                         RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
 507                         return -1;
 508                 }
 509
 510                 /* swap the 2 entries in the table */
 511                 memcpy(&tmp, &hugepg_tbl[smallest_idx], sizeof(struct hugepage));
 512                 memcpy(&hugepg_tbl[smallest_idx], &hugepg_tbl[i],
 513                                 sizeof(struct hugepage));
 514                 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage));
 515         }
 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 = open(filename, O_CREAT | O_RDWR, 0666);
 528         if (fd < 0)
 529                 return NULL;
 530         if (ftruncate(fd, mem_size) < 0) {
 531                 close(fd);
 532                 return NULL;
 533         }
 534         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 535         close(fd);
 536         return retval;
 537 }
 538
 539 /*
 540  * this copies *active* hugepages from one hugepage table to another.
 541  * destination is typically the shared memory.
 542  */
 543 static int
 544 copy_hugepages_to_shared_mem(struct hugepage * dst, int dest_size,
 545                 const struct hugepage * src, int src_size)
 546 {
 547         int src_pos, dst_pos = 0;
 548
 549         for (src_pos = 0; src_pos < src_size; src_pos++) {
 550                 if (src[src_pos].final_va != NULL) {
 551                         /* error on overflow attempt */
 552                         if (dst_pos == dest_size)
 553                                 return -1;
 554                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage));
 555                         dst_pos++;
 556                 }
 557         }
 558         return 0;
 559 }
 560
 561 /*
 562  * unmaps hugepages that are not going to be used. since we originally allocate
 563  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 564  */
 565 static int
 566 unmap_unneeded_hugepages(struct hugepage *hugepg_tbl,
 567                 struct hugepage_info *hpi,
 568                 unsigned num_hp_info)
 569 {
 570         unsigned socket, size;
 571         int page, nrpages = 0;
 572         int fd;
 573
 574         /* get total number of hugepages */
 575         for (size = 0; size < num_hp_info; size++)
 576                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 577                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 578
 579         for (size = 0; size < num_hp_info; size++) {
 580                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 581                         unsigned pages_found = 0;
 582                         /* traverse until we have unmapped all the unused pages */
 583                         for (page = 0; page < nrpages; page++) {
 584                                 struct hugepage *hp = &hugepg_tbl[page];
 585
 586                                 /* find a page that matches the criteria */
 587                                 if ((hp->size == hpi[size].hugepage_sz) &&
 588                                                 (hp->socket_id == (int) socket)) {
 589
 590                                         /* if we skipped enough pages, unmap the rest */
 591                                         if (pages_found == hpi[size].num_pages[socket]) {
 592                                                 munmap(hp->final_va, hp->size);
 593                                                 hp->final_va = NULL;
 594                                         }
 595                                         /* lock the page and skip */
 596                                         else {
 597                                                 /* try and open the hugepage file */
 598                                                 while ((fd = open(hp->filepath, O_CREAT | O_RDWR, 0755)) < 0) {
 599                                                         /* if we can't open due to resource limits */
 600                                                         if (errno == EMFILE) {
 601                                                                 RTE_LOG(INFO, EAL, "Increasing open file limit\n");
 602
 603                                                                 /* if we manage to increase resource limit, try again */
 604                                                                 if (increase_open_file_limit() == 0)
 605                                                                         continue;
 606                                                         }
 607                                                         else
 608                                                                 RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
 609                                                                                 strerror(errno));
 610                                                         return -1;
 611                                                 }
 612                                                 /* try and lock the hugepage */
 613                                                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 614                                                         RTE_LOG(ERR, EAL, "Locking hugepage file failed!\n");
 615                                                         close(fd);
 616                                                         return -1;
 617                                                 }
 618                                                 hp->page_lock = fd;
 619                                                 pages_found++;
 620                                         }
 621                                 } /* match page */
 622                         } /* foreach page */
 623                 } /* foreach socket */
 624         } /* foreach pagesize */
 625
 626         return 0;
 627 }
 628
 629 static inline uint64_t
 630 get_socket_mem_size(int socket)
 631 {
 632         uint64_t size = 0;
 633         unsigned i;
 634
 635         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 636                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 637                 if (hpi->hugedir != NULL)
 638                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 639         }
 640
 641         return (size);
 642 }
 643
 644 /*
 645  * This function is a NUMA-aware equivalent of calc_num_pages.
 646  * It takes in the list of hugepage sizes and the
 647  * number of pages thereof, and calculates the best number of
 648  * pages of each size to fulfill the request for <memory> ram
 649  */
 650 static int
 651 calc_num_pages_per_socket(uint64_t * memory,
 652                 struct hugepage_info *hp_info,
 653                 struct hugepage_info *hp_used,
 654                 unsigned num_hp_info)
 655 {
 656         unsigned socket, j, i = 0;
 657         unsigned requested, available;
 658         int total_num_pages = 0;
 659         uint64_t remaining_mem, cur_mem;
 660         uint64_t total_mem = internal_config.memory;
 661
 662         if (num_hp_info == 0)
 663                 return -1;
 664
 665         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 666                 /* if specific memory amounts per socket weren't requested */
 667                 if (internal_config.force_sockets == 0) {
 668                         /* take whatever is available */
 669                         memory[socket] = RTE_MIN(get_socket_mem_size(socket),
 670                                         total_mem);
 671                 }
 672                 /* skips if the memory on specific socket wasn't requested */
 673                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 674                         hp_used[i].hugedir = hp_info[i].hugedir;
 675                         hp_used[i].num_pages[socket] = RTE_MIN(
 676                                         memory[socket] / hp_info[i].hugepage_sz,
 677                                         hp_info[i].num_pages[socket]);
 678
 679                         cur_mem = hp_used[i].num_pages[socket] *
 680                                         hp_used[i].hugepage_sz;
 681
 682                         memory[socket] -= cur_mem;
 683                         total_mem -= cur_mem;
 684
 685                         total_num_pages += hp_used[i].num_pages[socket];
 686
 687                         /* check if we have met all memory requests */
 688                         if (memory[socket] == 0)
 689                                 break;
 690
 691                         /* check if we have any more pages left at this size, if so
 692                          * move on to next size */
 693                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
 694                                 continue;
 695                         /* At this point we know that there are more pages available that are
 696                          * bigger than the memory we want, so lets see if we can get enough
 697                          * from other page sizes.
 698                          */
 699                         remaining_mem = 0;
 700                         for (j = i+1; j < num_hp_info; j++)
 701                                 remaining_mem += hp_info[j].hugepage_sz *
 702                                 hp_info[j].num_pages[socket];
 703
 704                         /* is there enough other memory, if not allocate another page and quit */
 705                         if (remaining_mem < memory[socket]){
 706                                 cur_mem = RTE_MIN(memory[socket],
 707                                                 hp_info[i].hugepage_sz);
 708                                 memory[socket] -= cur_mem;
 709                                 total_mem -= cur_mem;
 710                                 hp_used[i].num_pages[socket]++;
 711                                 total_num_pages++;
 712                                 break; /* we are done with this socket*/
 713                         }
 714                 }
 715                 /* if we didn't satisfy all memory requirements per socket */
 716                 if (memory[socket] > 0) {
 717                         /* to prevent icc errors */
 718                         requested = (unsigned) (internal_config.socket_mem[socket] /
 719                                         0x100000);
 720                         available = requested -
 721                                         ((unsigned) (memory[socket] / 0x100000));
 722                         RTE_LOG(INFO, EAL, "Not enough memory available on socket %u! "
 723                                         "Requested: %uMB, available: %uMB\n", socket,
 724                                         requested, available);
 725                         return -1;
 726                 }
 727         }
 728
 729         /* if we didn't satisfy total memory requirements */
 730         if (total_mem > 0) {
 731                 requested = (unsigned) (internal_config.memory / 0x100000);
 732                 available = requested - (unsigned) (total_mem / 0x100000);
 733                 RTE_LOG(INFO, EAL, "Not enough memory available! Requested: %uMB,"
 734                                 " available: %uMB\n", requested, available);
 735                 return -1;
 736         }
 737         return total_num_pages;
 738 }
 739
 740 /*
 741  * Prepare physical memory mapping: fill configuration structure with
 742  * these infos, return 0 on success.
 743  *  1. map N huge pages in separate files in hugetlbfs
 744  *  2. find associated physical addr
 745  *  3. find associated NUMA socket ID
 746  *  4. sort all huge pages by physical address
 747  *  5. remap these N huge pages in the correct order
 748  *  6. unmap the first mapping
 749  *  7. fill memsegs in configuration with contiguous zones
 750  */
 751 static int
 752 rte_eal_hugepage_init(void)
 753 {
 754         struct rte_mem_config *mcfg;
 755         struct hugepage *hugepage, *tmp_hp = NULL;
 756         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
 757
 758         uint64_t memory[RTE_MAX_NUMA_NODES];
 759
 760         unsigned hp_offset;
 761         int i, j, new_memseg;
 762         int nrpages, total_pages = 0;
 763         void *addr;
 764
 765         memset(used_hp, 0, sizeof(used_hp));
 766
 767         /* get pointer to global configuration */
 768         mcfg = rte_eal_get_configuration()->mem_config;
 769
 770         /* for debug purposes, hugetlbfs can be disabled */
 771         if (internal_config.no_hugetlbfs) {
 772                 addr = malloc(internal_config.memory);
 773                 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
 774                 mcfg->memseg[0].addr = addr;
 775                 mcfg->memseg[0].len = internal_config.memory;
 776                 mcfg->memseg[0].socket_id = 0;
 777                 return 0;
 778         }
 779
 780
 781         /* calculate total number of hugepages available. at this point we haven't
 782          * yet started sorting them so they all are on socket 0 */
 783         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
 784                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
 785                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
 786
 787                 total_pages += internal_config.hugepage_info[i].num_pages[0];
 788         }
 789
 790         /*
 791          * allocate a memory area for hugepage table.
 792          * this isn't shared memory yet. due to the fact that we need some
 793          * processing done on these pages, shared memory will be created
 794          * at a later stage.
 795          */
 796         tmp_hp = malloc(total_pages * sizeof(struct hugepage));
 797         if (tmp_hp == NULL)
 798                 goto fail;
 799
 800         memset(tmp_hp, 0, total_pages * sizeof(struct hugepage));
 801
 802         hp_offset = 0; /* where we start the current page size entries */
 803
 804         /* map all hugepages and sort them */
 805         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
 806                 struct hugepage_info *hpi;
 807
 808                 /*
 809                  * we don't yet mark hugepages as used at this stage, so
 810                  * we just map all hugepages available to the system
 811                  * all hugepages are still located on socket 0
 812                  */
 813                 hpi = &internal_config.hugepage_info[i];
 814
 815                 if (hpi->num_pages == 0)
 816                         continue;
 817
 818                 /* map all hugepages available */
 819                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 1) < 0){
 820                         RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
 821                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 822                         goto fail;
 823                 }
 824
 825                 /* find physical addresses and sockets for each hugepage */
 826                 if (find_physaddr(&tmp_hp[hp_offset], hpi) < 0){
 827                         RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
 828                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 829                         goto fail;
 830                 }
 831
 832                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
 833                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
 834                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 835                         goto fail;
 836                 }
 837
 838                 if (sort_by_physaddr(&tmp_hp[hp_offset], hpi) < 0)
 839                         goto fail;
 840
 841                 /* remap all hugepages */
 842                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) < 0){
 843                         RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
 844                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 845                         goto fail;
 846                 }
 847
 848                 /* unmap original mappings */
 849                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
 850                         goto fail;
 851
 852                 /* we have processed a num of hugepages of this size, so inc offset */
 853                 hp_offset += hpi->num_pages[0];
 854         }
 855
 856         /* clean out the numbers of pages */
 857         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
 858                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
 859                         internal_config.hugepage_info[i].num_pages[j] = 0;
 860
 861         /* get hugepages for each socket */
 862         for (i = 0; i < total_pages; i++) {
 863                 int socket = tmp_hp[i].socket_id;
 864
 865                 /* find a hugepage info with right size and increment num_pages */
 866                 for (j = 0; j < (int) internal_config.num_hugepage_sizes; j++) {
 867                         if (tmp_hp[i].size ==
 868                                         internal_config.hugepage_info[j].hugepage_sz) {
 869                                 internal_config.hugepage_info[j].num_pages[socket]++;
 870                         }
 871                 }
 872         }
 873
 874         /* make a copy of socket_mem, needed for number of pages calculation */
 875         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 876                 memory[i] = internal_config.socket_mem[i];
 877
 878         /* calculate final number of pages */
 879         nrpages = calc_num_pages_per_socket(memory,
 880                         internal_config.hugepage_info, used_hp,
 881                         internal_config.num_hugepage_sizes);
 882
 883         /* error if not enough memory available */
 884         if (nrpages < 0)
 885                 goto fail;
 886
 887         /* reporting in! */
 888         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
 889                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 890                         if (used_hp[i].num_pages[j] > 0) {
 891                                 RTE_LOG(INFO, EAL,
 892                                                 "Requesting %u pages of size %uMB"
 893                                                 " from socket %i\n",
 894                                                 used_hp[i].num_pages[j],
 895                                                 (unsigned)
 896                                                         (used_hp[i].hugepage_sz / 0x100000),
 897                                                 j);
 898                         }
 899                 }
 900         }
 901
 902         /* create shared memory */
 903         hugepage = create_shared_memory(eal_hugepage_info_path(),
 904                                         nrpages * sizeof(struct hugepage));
 905
 906         if (hugepage == NULL) {
 907                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
 908                 goto fail;
 909         }
 910
 911         /*
 912          * unmap pages that we won't need (looks at used_hp).
 913          * also, sets final_va to NULL on pages that were unmapped.
 914          */
 915         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
 916                         internal_config.num_hugepage_sizes) < 0) {
 917                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
 918                 goto fail;
 919         }
 920
 921         /*
 922          * copy stuff from malloc'd hugepage* to the actual shared memory.
 923          * this procedure only copies those hugepages that have final_va
 924          * not NULL. has overflow protection.
 925          */
 926         if (copy_hugepages_to_shared_mem(hugepage, nrpages,
 927                         tmp_hp, total_pages) < 0) {
 928                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
 929                 goto fail;
 930         }
 931
 932         /* free the temporary hugepage table */
 933         free(tmp_hp);
 934         tmp_hp = NULL;
 935
 936         memset(mcfg->memseg, 0, sizeof(mcfg->memseg));
 937         j = -1;
 938         for (i = 0; i < nrpages; i++) {
 939                 new_memseg = 0;
 940
 941                 /* if this is a new section, create a new memseg */
 942                 if (i == 0)
 943                         new_memseg = 1;
 944                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
 945                         new_memseg = 1;
 946                 else if (hugepage[i].size != hugepage[i-1].size)
 947                         new_memseg = 1;
 948                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
 949                     hugepage[i].size)
 950                         new_memseg = 1;
 951                 else if (((unsigned long)hugepage[i].final_va -
 952                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
 953                         new_memseg = 1;
 954
 955                 if (new_memseg) {
 956                         j += 1;
 957                         if (j == RTE_MAX_MEMSEG)
 958                                 break;
 959
 960                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
 961                         mcfg->memseg[j].addr = hugepage[i].final_va;
 962                         mcfg->memseg[j].len = hugepage[i].size;
 963                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
 964                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
 965                 }
 966                 /* continuation of previous memseg */
 967                 else {
 968                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
 969                 }
 970                 hugepage[i].memseg_id = j;
 971         }
 972
 973         if (i < nrpages) {
 974                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
 975                         "from %d requested\n"
 976                         "Current %s=%d is not enough\n"
 977                         "Please either increase it or request less amount "
 978                         "of memory.\n",
 979                         i, nrpages, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
 980                         RTE_MAX_MEMSEG);
 981                 return (-ENOMEM);
 982         }
 983
 984
 985         return 0;
 986
 987
 988 fail:
 989         if (tmp_hp)
 990                 free(tmp_hp);
 991         return -1;
 992 }
 993
 994 /*
 995  * uses fstat to report the size of a file on disk
 996  */
 997 static off_t
 998 getFileSize(int fd)
 999 {
1000         struct stat st;
1001         if (fstat(fd, &st) < 0)
1002                 return 0;
1003         return st.st_size;
1004 }
1005
1006 /*
1007  * This creates the memory mappings in the secondary process to match that of
1008  * the server process. It goes through each memory segment in the DPDK runtime
1009  * configuration and finds the hugepages which form that segment, mapping them
1010  * in order to form a contiguous block in the virtual memory space
1011  */
1012 static int
1013 rte_eal_hugepage_attach(void)
1014 {
1015         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1016         const struct hugepage *hp = NULL;
1017         unsigned num_hp = 0;
1018         unsigned i, s = 0; /* s used to track the segment number */
1019         off_t size;
1020         int fd, fd_zero = -1, fd_hugepage = -1;
1021
1022         if (aslr_enabled() > 0) {
1023                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1024                                 "(ASLR) is enabled in the kernel.\n");
1025                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1026                                 "into secondary processes\n");
1027         }
1028
1029         fd_zero = open("/dev/zero", O_RDONLY);
1030         if (fd_zero < 0) {
1031                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1032                 goto error;
1033         }
1034         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1035         if (fd_hugepage < 0) {
1036                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1037                 goto error;
1038         }
1039
1040         size = getFileSize(fd_hugepage);
1041         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1042         if (hp == NULL) {
1043                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1044                 goto error;
1045         }
1046
1047         num_hp = size / sizeof(struct hugepage);
1048         RTE_LOG(DEBUG, EAL, "Analysing %u hugepages\n", num_hp);
1049
1050         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1051                 void *addr, *base_addr;
1052                 uintptr_t offset = 0;
1053
1054                 /* fdzero is mmapped to get a contiguous block of virtual addresses
1055                  * get a block of free memory of the appropriate size -
1056                  * use mmap to attempt to get an identical address as server.
1057                  */
1058                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1059                                 PROT_READ, MAP_PRIVATE, fd_zero, 0);
1060                 if (base_addr == MAP_FAILED || base_addr != mcfg->memseg[s].addr) {
1061                         RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1062                                 "in /dev/zero to requested address [%p]\n",
1063                                 (unsigned long long)mcfg->memseg[s].len,
1064                                 mcfg->memseg[s].addr);
1065                         if (aslr_enabled() > 0)
1066                                 RTE_LOG(ERR, EAL, "It is recommended to disable ASLR in the kernel "
1067                                                 "and retry running both primary and secondary processes\n");
1068                         goto error;
1069                 }
1070                 /* free memory so we can map the hugepages into the space */
1071                 munmap(base_addr, mcfg->memseg[s].len);
1072
1073                 /* find the hugepages for this segment and map them
1074                  * we don't need to worry about order, as the server sorted the
1075                  * entries before it did the second mmap of them */
1076                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1077                         if (hp[i].memseg_id == (int)s){
1078                                 fd = open(hp[i].filepath, O_RDWR);
1079                                 if (fd < 0) {
1080                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1081                                                 hp[i].filepath);
1082                                         goto error;
1083                                 }
1084                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1085                                                 hp[i].size, PROT_READ | PROT_WRITE,
1086                                                 MAP_SHARED | MAP_FIXED, fd, 0);
1087                                 close(fd); /* close file both on success and on failure */
1088                                 if (addr == MAP_FAILED) {
1089                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1090                                                 hp[i].filepath);
1091                                         goto error;
1092                                 }
1093                                 offset+=hp[i].size;
1094                         }
1095                 }
1096                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1097                                 (unsigned long long)mcfg->memseg[s].len);
1098                 s++;
1099         }
1100         close(fd_zero);
1101         close(fd_hugepage);
1102         return 0;
1103
1104 error:
1105         if (fd_zero >= 0)
1106                 close(fd_zero);
1107         if (fd_hugepage >= 0)
1108                 close(fd_hugepage);
1109         return -1;
1110 }
1111
1112 static int
1113 rte_eal_memdevice_init(void)
1114 {
1115         struct rte_config *config;
1116
1117         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1118                 return 0;
1119
1120         config = rte_eal_get_configuration();
1121         config->mem_config->nchannel = internal_config.force_nchannel;
1122         config->mem_config->nrank = internal_config.force_nrank;
1123
1124         return 0;
1125 }
1126
1127
1128 /* init memory subsystem */
1129 int
1130 rte_eal_memory_init(void)
1131 {
1132         RTE_LOG(INFO, EAL, "Setting up hugepage memory...\n");
1133         const int retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
1134                         rte_eal_hugepage_init() :
1135                         rte_eal_hugepage_attach();
1136         if (retval < 0)
1137                 return -1;
1138
1139         if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)
1140                 return -1;
1141
1142         return 0;
1143 }