lib/librte_eal/linuxapp/eal/eal_memory.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2013 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
  84 #include <rte_log.h>
  85 #include <rte_memory.h>
  86 #include <rte_memzone.h>
  87 #include <rte_launch.h>
  88 #include <rte_tailq.h>
  89 #include <rte_eal.h>
  90 #include <rte_eal_memconfig.h>
  91 #include <rte_per_lcore.h>
  92 #include <rte_lcore.h>
  93 #include <rte_common.h>
  94 #include <rte_string_fns.h>
  95
  96 #include "eal_private.h"
  97 #include "eal_internal_cfg.h"
  98 #include "eal_filesystem.h"
  99 #include "eal_hugepages.h"
 100
 101 /**
 102  * @file
 103  * Huge page mapping under linux
 104  *
 105  * To reserve a big contiguous amount of memory, we use the hugepage
 106  * feature of linux. For that, we need to have hugetlbfs mounted. This
 107  * code will create many files in this directory (one per page) and
 108  * map them in virtual memory. For each page, we will retrieve its
 109  * physical address and remap it in order to have a virtual contiguous
 110  * zone as well as a physical contiguous zone.
 111  */
 112
 113
 114 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
 115
 116 /*
 117  * Check whether address-space layout randomization is enabled in
 118  * the kernel. This is important for multi-process as it can prevent
 119  * two processes mapping data to the same virtual address
 120  * Returns:
 121  *    0 - address space randomization disabled
 122  *    1/2 - address space randomization enabled
 123  *    negative error code on error
 124  */
 125 static int
 126 aslr_enabled(void)
 127 {
 128         char c;
 129         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 130         if (fd < 0)
 131                 return -errno;
 132         retval = read(fd, &c, 1);
 133         close(fd);
 134         if (retval < 0)
 135                 return -errno;
 136         if (retval == 0)
 137                 return -EIO;
 138         switch (c) {
 139                 case '0' : return 0;
 140                 case '1' : return 1;
 141                 case '2' : return 2;
 142                 default: return -EINVAL;
 143         }
 144 }
 145
 146 /*
 147  * Try to mmap *size bytes in /dev/zero. If it is succesful, return the
 148  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 149  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 150  * 0. In this case, return NULL. Note: this function returns an address
 151  * which is a multiple of hugepage size.
 152  */
 153 static void *
 154 get_virtual_area(size_t *size, size_t hugepage_sz)
 155 {
 156         void *addr;
 157         int fd;
 158         long aligned_addr;
 159
 160         RTE_LOG(INFO, EAL, "Ask a virtual area of 0x%zu bytes\n", *size);
 161
 162         fd = open("/dev/zero", O_RDONLY);
 163         if (fd < 0){
 164                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 165                 return NULL;
 166         }
 167         do {
 168                 addr = mmap(NULL, (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
 169                 if (addr == MAP_FAILED)
 170                         *size -= hugepage_sz;
 171         } while (addr == MAP_FAILED && *size > 0);
 172
 173         if (addr == MAP_FAILED) {
 174                 close(fd);
 175                 RTE_LOG(INFO, EAL, "Cannot get a virtual area\n");
 176                 return NULL;
 177         }
 178
 179         munmap(addr, (*size) + hugepage_sz);
 180         close(fd);
 181
 182         /* align addr to a huge page size boundary */
 183         aligned_addr = (long)addr;
 184         aligned_addr += (hugepage_sz - 1);
 185         aligned_addr &= (~(hugepage_sz - 1));
 186         addr = (void *)(aligned_addr);
 187
 188         RTE_LOG(INFO, EAL, "Virtual area found at %p (size = 0x%zx)\n",
 189                 addr, *size);
 190
 191         return addr;
 192 }
 193
 194 /*
 195  * Mmap all hugepages of hugepage table: it first open a file in
 196  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 197  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 198  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 199  * map continguous physical blocks in contiguous virtual blocks.
 200  */
 201 static int
 202 map_all_hugepages(struct hugepage *hugepg_tbl,
 203                 struct hugepage_info *hpi, int orig)
 204 {
 205         int fd;
 206         unsigned i;
 207         void *virtaddr;
 208         void *vma_addr = NULL;
 209         size_t vma_len = 0;
 210
 211         for (i = 0; i < hpi->num_pages[0]; i++) {
 212                 size_t hugepage_sz = hpi->hugepage_sz;
 213
 214                 if (orig) {
 215                         hugepg_tbl[i].file_id = i;
 216                         hugepg_tbl[i].size = hugepage_sz;
 217                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 218                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 219                                         hugepg_tbl[i].file_id);
 220                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 221                 }
 222 #ifndef RTE_ARCH_X86_64
 223                 /* for 32-bit systems, don't remap 1G pages, just reuse original
 224                  * map address as final map address.
 225                  */
 226                 else if (hugepage_sz == RTE_PGSIZE_1G){
 227                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 228                         hugepg_tbl[i].orig_va = NULL;
 229                         continue;
 230                 }
 231 #endif
 232                 else if (vma_len == 0) {
 233                         unsigned j, num_pages;
 234
 235                         /* reserve a virtual area for next contiguous
 236                          * physical block: count the number of
 237                          * contiguous physical pages. */
 238                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 239                                 if (hugepg_tbl[j].physaddr !=
 240                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 241                                         break;
 242                         }
 243                         num_pages = j - i;
 244                         vma_len = num_pages * hugepage_sz;
 245
 246                         /* get the biggest virtual memory area up to
 247                          * vma_len. If it fails, vma_addr is NULL, so
 248                          * let the kernel provide the address. */
 249                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 250                         if (vma_addr == NULL)
 251                                 vma_len = hugepage_sz;
 252                 }
 253
 254                 /* try to create hugepage file */
 255                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
 256                 if (fd < 0) {
 257                         RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
 258                                         strerror(errno));
 259                         return -1;
 260                 }
 261
 262                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 263                                 MAP_SHARED, fd, 0);
 264                 if (virtaddr == MAP_FAILED) {
 265                         RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__,
 266                                         strerror(errno));
 267                         close(fd);
 268                         return -1;
 269                 }
 270
 271                 if (orig) {
 272                         hugepg_tbl[i].orig_va = virtaddr;
 273                         memset(virtaddr, 0, hugepage_sz);
 274                 }
 275                 else {
 276                         hugepg_tbl[i].final_va = virtaddr;
 277                 }
 278
 279                 /* set shared flock on the file. */
 280                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 281                         RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
 282                                 __func__, strerror(errno));
 283                         close(fd);
 284                         return -1;
 285                 }
 286
 287                 close(fd);
 288
 289                 vma_addr = (char *)vma_addr + hugepage_sz;
 290                 vma_len -= hugepage_sz;
 291         }
 292         return 0;
 293 }
 294
 295 /* Unmap all hugepages from original mapping. */
 296 static int
 297 unmap_all_hugepages_orig(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 298 {
 299         unsigned i;
 300         for (i = 0; i < hpi->num_pages[0]; i++) {
 301                 if (hugepg_tbl[i].orig_va) {
 302                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 303                         hugepg_tbl[i].orig_va = NULL;
 304                 }
 305         }
 306         return 0;
 307 }
 308
 309 /*
 310  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 311  * it by browsing the /proc/self/pagemap special file.
 312  */
 313 static int
 314 find_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 315 {
 316         int fd;
 317         unsigned i;
 318         uint64_t page;
 319         unsigned long virt_pfn;
 320         int page_size;
 321
 322         /* standard page size */
 323         page_size = getpagesize();
 324
 325         fd = open("/proc/self/pagemap", O_RDONLY);
 326         if (fd < 0) {
 327                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 328                         __func__, strerror(errno));
 329                 return -1;
 330         }
 331
 332         for (i = 0; i < hpi->num_pages[0]; i++) {
 333                 off_t offset;
 334                 virt_pfn = (unsigned long)hugepg_tbl[i].orig_va /
 335                         page_size;
 336                 offset = sizeof(uint64_t) * virt_pfn;
 337                 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 338                         RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 339                                         __func__, strerror(errno));
 340                         close(fd);
 341                         return -1;
 342                 }
 343                 if (read(fd, &page, sizeof(uint64_t)) < 0) {
 344                         RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 345                                         __func__, strerror(errno));
 346                         close(fd);
 347                         return -1;
 348                 }
 349
 350                 /*
 351                  * the pfn (page frame number) are bits 0-54 (see
 352                  * pagemap.txt in linux Documentation)
 353                  */
 354                 hugepg_tbl[i].physaddr = ((page & 0x7fffffffffffffULL) * page_size);
 355         }
 356         close(fd);
 357         return 0;
 358 }
 359
 360 /*
 361  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 362  * page.
 363  */
 364 static int
 365 find_numasocket(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 366 {
 367         int socket_id;
 368         char *end, *nodestr;
 369         unsigned i, hp_count = 0;
 370         uint64_t virt_addr;
 371         char buf[BUFSIZ];
 372         char hugedir_str[PATH_MAX];
 373         FILE *f;
 374
 375         f = fopen("/proc/self/numa_maps", "r");
 376         if (f == NULL) {
 377                 RTE_LOG(INFO, EAL, "cannot open /proc/self/numa_maps,"
 378                                 " consider that all memory is in socket_id 0\n");
 379                 return 0;
 380         }
 381
 382         rte_snprintf(hugedir_str, sizeof(hugedir_str),
 383                         "%s/", hpi->hugedir);
 384
 385         /* parse numa map */
 386         while (fgets(buf, sizeof(buf), f) != NULL) {
 387
 388                 /* ignore non huge page */
 389                 if (strstr(buf, " huge ") == NULL &&
 390                                 strstr(buf, hugedir_str) == NULL)
 391                         continue;
 392
 393                 /* get zone addr */
 394                 virt_addr = strtoull(buf, &end, 16);
 395                 if (virt_addr == 0 || end == buf) {
 396                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 397                         goto error;
 398                 }
 399
 400                 /* get node id (socket id) */
 401                 nodestr = strstr(buf, " N");
 402                 if (nodestr == NULL) {
 403                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 404                         goto error;
 405                 }
 406                 nodestr += 2;
 407                 end = strstr(nodestr, "=");
 408                 if (end == NULL) {
 409                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 410                         goto error;
 411                 }
 412                 end[0] = '\0';
 413                 end = NULL;
 414
 415                 socket_id = strtoul(nodestr, &end, 0);
 416                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 417                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 418                         goto error;
 419                 }
 420
 421                 /* if we find this page in our mappings, set socket_id */
 422                 for (i = 0; i < hpi->num_pages[0]; i++) {
 423                         void *va = (void *)(unsigned long)virt_addr;
 424                         if (hugepg_tbl[i].orig_va == va) {
 425                                 hugepg_tbl[i].socket_id = socket_id;
 426                                 hp_count++;
 427                         }
 428                 }
 429         }
 430
 431         if (hp_count < hpi->num_pages[0])
 432                 goto error;
 433
 434         fclose(f);
 435         return 0;
 436
 437 error:
 438         fclose(f);
 439         return -1;
 440 }
 441
 442 /*
 443  * Sort the hugepg_tbl by physical address (lower addresses first). We
 444  * use a slow algorithm, but we won't have millions of pages, and this
 445  * is only done at init time.
 446  */
 447 static int
 448 sort_by_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 449 {
 450         unsigned i, j;
 451         int smallest_idx;
 452         uint64_t smallest_addr;
 453         struct hugepage tmp;
 454
 455         for (i = 0; i < hpi->num_pages[0]; i++) {
 456                 smallest_addr = 0;
 457                 smallest_idx = -1;
 458
 459                 /*
 460                  * browse all entries starting at 'i', and find the
 461                  * entry with the smallest addr
 462                  */
 463                 for (j=i; j< hpi->num_pages[0]; j++) {
 464
 465                         if (smallest_addr == 0 ||
 466                             hugepg_tbl[j].physaddr < smallest_addr) {
 467                                 smallest_addr = hugepg_tbl[j].physaddr;
 468                                 smallest_idx = j;
 469                         }
 470                 }
 471
 472                 /* should not happen */
 473                 if (smallest_idx == -1) {
 474                         RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
 475                         return -1;
 476                 }
 477
 478                 /* swap the 2 entries in the table */
 479                 memcpy(&tmp, &hugepg_tbl[smallest_idx], sizeof(struct hugepage));
 480                 memcpy(&hugepg_tbl[smallest_idx], &hugepg_tbl[i],
 481                                 sizeof(struct hugepage));
 482                 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage));
 483         }
 484         return 0;
 485 }
 486
 487 /*
 488  * Uses mmap to create a shared memory area for storage of data
 489  * Used in this file to store the hugepage file map on disk
 490  */
 491 static void *
 492 create_shared_memory(const char *filename, const size_t mem_size)
 493 {
 494         void *retval;
 495         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 496         if (fd < 0)
 497                 return NULL;
 498         if (ftruncate(fd, mem_size) < 0) {
 499                 close(fd);
 500                 return NULL;
 501         }
 502         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 503         close(fd);
 504         return retval;
 505 }
 506
 507 /*
 508  * this copies *active* hugepages from one hugepage table to another.
 509  * destination is typically the shared memory.
 510  */
 511 static int
 512 copy_hugepages_to_shared_mem(struct hugepage * dst, int dest_size,
 513                 const struct hugepage * src, int src_size)
 514 {
 515         int src_pos, dst_pos = 0;
 516
 517         for (src_pos = 0; src_pos < src_size; src_pos++) {
 518                 if (src[src_pos].final_va != NULL) {
 519                         /* error on overflow attempt */
 520                         if (dst_pos == dest_size)
 521                                 return -1;
 522                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage));
 523                         dst_pos++;
 524                 }
 525         }
 526         return 0;
 527 }
 528
 529 /*
 530  * unmaps hugepages that are not going to be used. since we originally allocate
 531  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 532  */
 533 static int
 534 unmap_unneeded_hugepages(struct hugepage *hugepg_tbl,
 535                 struct hugepage_info *hpi,
 536                 unsigned num_hp_info)
 537 {
 538         unsigned socket, size;
 539         int page, nrpages = 0;
 540
 541         /* get total number of hugepages */
 542         for (size = 0; size < num_hp_info; size++)
 543                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 544                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 545
 546         for (size = 0; size < num_hp_info; size++) {
 547                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 548                         unsigned pages_found = 0;
 549                         /* traverse until we have unmapped all the unused pages */
 550                         for (page = 0; page < nrpages; page++) {
 551                                 struct hugepage *hp = &hugepg_tbl[page];
 552
 553                                 /* find a page that matches the criteria */
 554                                 if ((hp->size == hpi[size].hugepage_sz) &&
 555                                                 (hp->socket_id == (int) socket)) {
 556
 557                                         /* if we skipped enough pages, unmap the rest */
 558                                         if (pages_found == hpi[size].num_pages[socket]) {
 559                                                 munmap(hp->final_va, hp->size);
 560                                                 hp->final_va = NULL;
 561                                         }
 562                                         /* lock the page and skip */
 563                                         else
 564                                                 pages_found++;
 565
 566                                 } /* match page */
 567                         } /* foreach page */
 568                 } /* foreach socket */
 569         } /* foreach pagesize */
 570
 571         return 0;
 572 }
 573
 574 static inline uint64_t
 575 get_socket_mem_size(int socket)
 576 {
 577         uint64_t size = 0;
 578         unsigned i;
 579
 580         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 581                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 582                 if (hpi->hugedir != NULL)
 583                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 584         }
 585
 586         return (size);
 587 }
 588
 589 /*
 590  * This function is a NUMA-aware equivalent of calc_num_pages.
 591  * It takes in the list of hugepage sizes and the
 592  * number of pages thereof, and calculates the best number of
 593  * pages of each size to fulfill the request for <memory> ram
 594  */
 595 static int
 596 calc_num_pages_per_socket(uint64_t * memory,
 597                 struct hugepage_info *hp_info,
 598                 struct hugepage_info *hp_used,
 599                 unsigned num_hp_info)
 600 {
 601         unsigned socket, j, i = 0;
 602         unsigned requested, available;
 603         int total_num_pages = 0;
 604         uint64_t remaining_mem, cur_mem;
 605         uint64_t total_mem = internal_config.memory;
 606
 607         if (num_hp_info == 0)
 608                 return -1;
 609
 610         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 611                 /* if specific memory amounts per socket weren't requested */
 612                 if (internal_config.force_sockets == 0) {
 613                         /* take whatever is available */
 614                         memory[socket] = RTE_MIN(get_socket_mem_size(socket),
 615                                         total_mem);
 616                 }
 617                 /* skips if the memory on specific socket wasn't requested */
 618                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 619                         hp_used[i].hugedir = hp_info[i].hugedir;
 620                         hp_used[i].num_pages[socket] = RTE_MIN(
 621                                         memory[socket] / hp_info[i].hugepage_sz,
 622                                         hp_info[i].num_pages[socket]);
 623
 624                         cur_mem = hp_used[i].num_pages[socket] *
 625                                         hp_used[i].hugepage_sz;
 626
 627                         memory[socket] -= cur_mem;
 628                         total_mem -= cur_mem;
 629
 630                         total_num_pages += hp_used[i].num_pages[socket];
 631
 632                         /* check if we have met all memory requests */
 633                         if (memory[socket] == 0)
 634                                 break;
 635
 636                         /* check if we have any more pages left at this size, if so
 637                          * move on to next size */
 638                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
 639                                 continue;
 640                         /* At this point we know that there are more pages available that are
 641                          * bigger than the memory we want, so lets see if we can get enough
 642                          * from other page sizes.
 643                          */
 644                         remaining_mem = 0;
 645                         for (j = i+1; j < num_hp_info; j++)
 646                                 remaining_mem += hp_info[j].hugepage_sz *
 647                                 hp_info[j].num_pages[socket];
 648
 649                         /* is there enough other memory, if not allocate another page and quit */
 650                         if (remaining_mem < memory[socket]){
 651                                 cur_mem = RTE_MIN(memory[socket],
 652                                                 hp_info[i].hugepage_sz);
 653                                 memory[socket] -= cur_mem;
 654                                 total_mem -= cur_mem;
 655                                 hp_used[i].num_pages[socket]++;
 656                                 total_num_pages++;
 657                                 break; /* we are done with this socket*/
 658                         }
 659                 }
 660                 /* if we didn't satisfy all memory requirements per socket */
 661                 if (memory[socket] > 0) {
 662                         /* to prevent icc errors */
 663                         requested = (unsigned) (internal_config.socket_mem[socket] /
 664                                         0x100000);
 665                         available = requested -
 666                                         ((unsigned) (memory[socket] / 0x100000));
 667                         RTE_LOG(INFO, EAL, "Not enough memory available on socket %u! "
 668                                         "Requested: %uMB, available: %uMB\n", socket,
 669                                         requested, available);
 670                         return -1;
 671                 }
 672         }
 673
 674         /* if we didn't satisfy total memory requirements */
 675         if (total_mem > 0) {
 676                 requested = (unsigned) (internal_config.memory / 0x100000);
 677                 available = requested - (unsigned) (total_mem / 0x100000);
 678                 RTE_LOG(INFO, EAL, "Not enough memory available! Requested: %uMB,"
 679                                 " available: %uMB\n", requested, available);
 680                 return -1;
 681         }
 682         return total_num_pages;
 683 }
 684
 685 /*
 686  * Prepare physical memory mapping: fill configuration structure with
 687  * these infos, return 0 on success.
 688  *  1. map N huge pages in separate files in hugetlbfs
 689  *  2. find associated physical addr
 690  *  3. find associated NUMA socket ID
 691  *  4. sort all huge pages by physical address
 692  *  5. remap these N huge pages in the correct order
 693  *  6. unmap the first mapping
 694  *  7. fill memsegs in configuration with contiguous zones
 695  */
 696 static int
 697 rte_eal_hugepage_init(void)
 698 {
 699         struct rte_mem_config *mcfg;
 700         struct hugepage *hugepage, *tmp_hp = NULL;
 701         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
 702
 703         uint64_t memory[RTE_MAX_NUMA_NODES];
 704
 705         unsigned hp_offset;
 706         int i, j, new_memseg;
 707         int nrpages, total_pages = 0;
 708         void *addr;
 709
 710         memset(used_hp, 0, sizeof(used_hp));
 711
 712         /* get pointer to global configuration */
 713         mcfg = rte_eal_get_configuration()->mem_config;
 714
 715         /* for debug purposes, hugetlbfs can be disabled */
 716         if (internal_config.no_hugetlbfs) {
 717                 addr = malloc(internal_config.memory);
 718                 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
 719                 mcfg->memseg[0].addr = addr;
 720                 mcfg->memseg[0].len = internal_config.memory;
 721                 mcfg->memseg[0].socket_id = 0;
 722                 return 0;
 723         }
 724
 725
 726         /* calculate total number of hugepages available. at this point we haven't
 727          * yet started sorting them so they all are on socket 0 */
 728         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
 729                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
 730                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
 731
 732                 total_pages += internal_config.hugepage_info[i].num_pages[0];
 733         }
 734
 735         /*
 736          * allocate a memory area for hugepage table.
 737          * this isn't shared memory yet. due to the fact that we need some
 738          * processing done on these pages, shared memory will be created
 739          * at a later stage.
 740          */
 741         tmp_hp = malloc(total_pages * sizeof(struct hugepage));
 742         if (tmp_hp == NULL)
 743                 goto fail;
 744
 745         memset(tmp_hp, 0, total_pages * sizeof(struct hugepage));
 746
 747         hp_offset = 0; /* where we start the current page size entries */
 748
 749         /* map all hugepages and sort them */
 750         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
 751                 struct hugepage_info *hpi;
 752
 753                 /*
 754                  * we don't yet mark hugepages as used at this stage, so
 755                  * we just map all hugepages available to the system
 756                  * all hugepages are still located on socket 0
 757                  */
 758                 hpi = &internal_config.hugepage_info[i];
 759
 760                 if (hpi->num_pages == 0)
 761                         continue;
 762
 763                 /* map all hugepages available */
 764                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 1) < 0){
 765                         RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
 766                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 767                         goto fail;
 768                 }
 769
 770                 /* find physical addresses and sockets for each hugepage */
 771                 if (find_physaddr(&tmp_hp[hp_offset], hpi) < 0){
 772                         RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
 773                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 774                         goto fail;
 775                 }
 776
 777                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
 778                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
 779                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 780                         goto fail;
 781                 }
 782
 783                 if (sort_by_physaddr(&tmp_hp[hp_offset], hpi) < 0)
 784                         goto fail;
 785
 786                 /* remap all hugepages */
 787                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) < 0){
 788                         RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
 789                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 790                         goto fail;
 791                 }
 792
 793                 /* unmap original mappings */
 794                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
 795                         goto fail;
 796
 797                 /* we have processed a num of hugepages of this size, so inc offset */
 798                 hp_offset += hpi->num_pages[0];
 799         }
 800
 801         /* clean out the numbers of pages */
 802         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
 803                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
 804                         internal_config.hugepage_info[i].num_pages[j] = 0;
 805
 806         /* get hugepages for each socket */
 807         for (i = 0; i < total_pages; i++) {
 808                 int socket = tmp_hp[i].socket_id;
 809
 810                 /* find a hugepage info with right size and increment num_pages */
 811                 for (j = 0; j < (int) internal_config.num_hugepage_sizes; j++) {
 812                         if (tmp_hp[i].size ==
 813                                         internal_config.hugepage_info[j].hugepage_sz) {
 814                                 internal_config.hugepage_info[j].num_pages[socket]++;
 815                         }
 816                 }
 817         }
 818
 819         /* make a copy of socket_mem, needed for number of pages calculation */
 820         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 821                 memory[i] = internal_config.socket_mem[i];
 822
 823         /* calculate final number of pages */
 824         nrpages = calc_num_pages_per_socket(memory,
 825                         internal_config.hugepage_info, used_hp,
 826                         internal_config.num_hugepage_sizes);
 827
 828         /* error if not enough memory available */
 829         if (nrpages < 0)
 830                 goto fail;
 831
 832         /* reporting in! */
 833         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
 834                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 835                         if (used_hp[i].num_pages[j] > 0) {
 836                                 RTE_LOG(INFO, EAL,
 837                                                 "Requesting %u pages of size %uMB"
 838                                                 " from socket %i\n",
 839                                                 used_hp[i].num_pages[j],
 840                                                 (unsigned)
 841                                                         (used_hp[i].hugepage_sz / 0x100000),
 842                                                 j);
 843                         }
 844                 }
 845         }
 846
 847         /* create shared memory */
 848         hugepage = create_shared_memory(eal_hugepage_info_path(),
 849                                         nrpages * sizeof(struct hugepage));
 850
 851         if (hugepage == NULL) {
 852                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
 853                 goto fail;
 854         }
 855
 856         /*
 857          * unmap pages that we won't need (looks at used_hp).
 858          * also, sets final_va to NULL on pages that were unmapped.
 859          */
 860         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
 861                         internal_config.num_hugepage_sizes) < 0) {
 862                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
 863                 goto fail;
 864         }
 865
 866         /*
 867          * copy stuff from malloc'd hugepage* to the actual shared memory.
 868          * this procedure only copies those hugepages that have final_va
 869          * not NULL. has overflow protection.
 870          */
 871         if (copy_hugepages_to_shared_mem(hugepage, nrpages,
 872                         tmp_hp, total_pages) < 0) {
 873                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
 874                 goto fail;
 875         }
 876
 877         /* free the temporary hugepage table */
 878         free(tmp_hp);
 879         tmp_hp = NULL;
 880
 881         memset(mcfg->memseg, 0, sizeof(mcfg->memseg));
 882         j = -1;
 883         for (i = 0; i < nrpages; i++) {
 884                 new_memseg = 0;
 885
 886                 /* if this is a new section, create a new memseg */
 887                 if (i == 0)
 888                         new_memseg = 1;
 889                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
 890                         new_memseg = 1;
 891                 else if (hugepage[i].size != hugepage[i-1].size)
 892                         new_memseg = 1;
 893                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
 894                     hugepage[i].size)
 895                         new_memseg = 1;
 896                 else if (((unsigned long)hugepage[i].final_va -
 897                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
 898                         new_memseg = 1;
 899
 900                 if (new_memseg) {
 901                         j += 1;
 902                         if (j == RTE_MAX_MEMSEG)
 903                                 break;
 904
 905                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
 906                         mcfg->memseg[j].addr = hugepage[i].final_va;
 907                         mcfg->memseg[j].len = hugepage[i].size;
 908                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
 909                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
 910                 }
 911                 /* continuation of previous memseg */
 912                 else {
 913                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
 914                 }
 915                 hugepage[i].memseg_id = j;
 916         }
 917
 918         if (i < nrpages) {
 919                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
 920                         "from %d requested\n"
 921                         "Current %s=%d is not enough\n"
 922                         "Please either increase it or request less amount "
 923                         "of memory.\n",
 924                         i, nrpages, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
 925                         RTE_MAX_MEMSEG);
 926                 return (-ENOMEM);
 927         }
 928
 929
 930         return 0;
 931
 932
 933 fail:
 934         if (tmp_hp)
 935                 free(tmp_hp);
 936         return -1;
 937 }
 938
 939 /*
 940  * uses fstat to report the size of a file on disk
 941  */
 942 static off_t
 943 getFileSize(int fd)
 944 {
 945         struct stat st;
 946         if (fstat(fd, &st) < 0)
 947                 return 0;
 948         return st.st_size;
 949 }
 950
 951 /*
 952  * This creates the memory mappings in the secondary process to match that of
 953  * the server process. It goes through each memory segment in the DPDK runtime
 954  * configuration and finds the hugepages which form that segment, mapping them
 955  * in order to form a contiguous block in the virtual memory space
 956  */
 957 static int
 958 rte_eal_hugepage_attach(void)
 959 {
 960         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 961         const struct hugepage *hp = NULL;
 962         unsigned num_hp = 0;
 963         unsigned i, s = 0; /* s used to track the segment number */
 964         off_t size;
 965         int fd, fd_zero = -1, fd_hugepage = -1;
 966
 967         if (aslr_enabled() > 0) {
 968                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
 969                                 "(ASLR) is enabled in the kernel.\n");
 970                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
 971                                 "into secondary processes\n");
 972         }
 973
 974         fd_zero = open("/dev/zero", O_RDONLY);
 975         if (fd_zero < 0) {
 976                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
 977                 goto error;
 978         }
 979         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
 980         if (fd_hugepage < 0) {
 981                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
 982                 goto error;
 983         }
 984
 985         /* map all segments into memory to make sure we get the addrs */
 986         for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
 987                 void *base_addr;
 988
 989                 /*
 990                  * the first memory segment with len==0 is the one that
 991                  * follows the last valid segment.
 992                  */
 993                 if (mcfg->memseg[s].len == 0)
 994                         break;
 995
 996                 /*
 997                  * fdzero is mmapped to get a contiguous block of virtual
 998                  * addresses of the appropriate memseg size.
 999                  * use mmap to get identical addresses as the primary process.
1000                  */
1001                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1002                                  PROT_READ, MAP_PRIVATE, fd_zero, 0);
1003                 if (base_addr == MAP_FAILED ||
1004                     base_addr != mcfg->memseg[s].addr) {
1005                         RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1006                                 "in /dev/zero to requested address [%p]\n",
1007                                 (unsigned long long)mcfg->memseg[s].len,
1008                                 mcfg->memseg[s].addr);
1009                         if (aslr_enabled() > 0) {
1010                                 RTE_LOG(ERR, EAL, "It is recommended to "
1011                                         "disable ASLR in the kernel "
1012                                         "and retry running both primary "
1013                                         "and secondary processes\n");
1014                         }
1015                         goto error;
1016                 }
1017         }
1018
1019         size = getFileSize(fd_hugepage);
1020         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1021         if (hp == NULL) {
1022                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1023                 goto error;
1024         }
1025
1026         num_hp = size / sizeof(struct hugepage);
1027         RTE_LOG(DEBUG, EAL, "Analysing %u hugepages\n", num_hp);
1028
1029         s = 0;
1030         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1031                 void *addr, *base_addr;
1032                 uintptr_t offset = 0;
1033
1034                 /*
1035                  * free previously mapped memory so we can map the
1036                  * hugepages into the space
1037                  */
1038                 base_addr = mcfg->memseg[s].addr;
1039                 munmap(base_addr, mcfg->memseg[s].len);
1040
1041                 /* find the hugepages for this segment and map them
1042                  * we don't need to worry about order, as the server sorted the
1043                  * entries before it did the second mmap of them */
1044                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1045                         if (hp[i].memseg_id == (int)s){
1046                                 fd = open(hp[i].filepath, O_RDWR);
1047                                 if (fd < 0) {
1048                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1049                                                 hp[i].filepath);
1050                                         goto error;
1051                                 }
1052                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1053                                                 hp[i].size, PROT_READ | PROT_WRITE,
1054                                                 MAP_SHARED | MAP_FIXED, fd, 0);
1055                                 close(fd); /* close file both on success and on failure */
1056                                 if (addr == MAP_FAILED) {
1057                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1058                                                 hp[i].filepath);
1059                                         goto error;
1060                                 }
1061                                 offset+=hp[i].size;
1062                         }
1063                 }
1064                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1065                                 (unsigned long long)mcfg->memseg[s].len);
1066                 s++;
1067         }
1068         /* unmap the hugepage config file, since we are done using it */
1069         munmap((void *)(uintptr_t)hp, size);
1070         close(fd_zero);
1071         close(fd_hugepage);
1072         return 0;
1073
1074 error:
1075         if (fd_zero >= 0)
1076                 close(fd_zero);
1077         if (fd_hugepage >= 0)
1078                 close(fd_hugepage);
1079         return -1;
1080 }
1081
1082 static int
1083 rte_eal_memdevice_init(void)
1084 {
1085         struct rte_config *config;
1086
1087         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1088                 return 0;
1089
1090         config = rte_eal_get_configuration();
1091         config->mem_config->nchannel = internal_config.force_nchannel;
1092         config->mem_config->nrank = internal_config.force_nrank;
1093
1094         return 0;
1095 }
1096
1097
1098 /* init memory subsystem */
1099 int
1100 rte_eal_memory_init(void)
1101 {
1102         RTE_LOG(INFO, EAL, "Setting up hugepage memory...\n");
1103         const int retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
1104                         rte_eal_hugepage_init() :
1105                         rte_eal_hugepage_attach();
1106         if (retval < 0)
1107                 return -1;
1108
1109         if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)
1110                 return -1;
1111
1112         return 0;
1113 }