lib/librte_eal/linuxapp/eal/eal_memory.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2012 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  *
  33  */
  34
  35 #include <errno.h>
  36 #include <stdarg.h>
  37 #include <stdlib.h>
  38 #include <stdio.h>
  39 #include <stdint.h>
  40 #include <inttypes.h>
  41 #include <string.h>
  42 #include <stdarg.h>
  43 #include <sys/mman.h>
  44 #include <sys/types.h>
  45 #include <sys/stat.h>
  46 #include <sys/queue.h>
  47 #include <fcntl.h>
  48 #include <unistd.h>
  49 #include <limits.h>
  50 #include <errno.h>
  51 #include <sys/ioctl.h>
  52
  53 #include <rte_log.h>
  54 #include <rte_memory.h>
  55 #include <rte_memzone.h>
  56 #include <rte_launch.h>
  57 #include <rte_tailq.h>
  58 #include <rte_eal.h>
  59 #include <rte_per_lcore.h>
  60 #include <rte_lcore.h>
  61 #include <rte_common.h>
  62 #include <rte_string_fns.h>
  63
  64 #include "eal_private.h"
  65 #include "eal_internal_cfg.h"
  66 #include "eal_filesystem.h"
  67 #include "eal_hugepages.h"
  68
  69 /**
  70  * @file
  71  * Huge page mapping under linux
  72  *
  73  * To reserve a big contiguous amount of memory, we use the hugepage
  74  * feature of linux. For that, we need to have hugetlbfs mounted. This
  75  * code will create many files in this directory (one per page) and
  76  * map them in virtual memory. For each page, we will retrieve its
  77  * physical address and remap it in order to have a virtual contiguous
  78  * zone as well as a physical contiguous zone.
  79  */
  80
  81
  82 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
  83
  84 /*
  85  * Check whether address-space layout randomization is enabled in
  86  * the kernel. This is important for multi-process as it can prevent
  87  * two processes mapping data to the same virtual address
  88  * Returns:
  89  *    0 - address space randomization disabled
  90  *    1/2 - address space randomization enabled
  91  *    negative error code on error
  92  */
  93 static int
  94 aslr_enabled(void)
  95 {
  96         char c;
  97         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
  98         if (fd < 0)
  99                 return -errno;
 100         retval = read(fd, &c, 1);
 101         close(fd);
 102         if (retval < 0)
 103                 return -errno;
 104         if (retval == 0)
 105                 return -EIO;
 106         switch (c) {
 107                 case '0' : return 0;
 108                 case '1' : return 1;
 109                 case '2' : return 2;
 110                 default: return -EINVAL;
 111         }
 112 }
 113
 114 /*
 115  * Try to mmap *size bytes in /dev/zero. If it is succesful, return the
 116  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 117  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 118  * 0. In this case, return NULL. Note: this function returns an address
 119  * which is a multiple of hugepage size.
 120  */
 121 static void *
 122 get_virtual_area(uint64_t *size, uint64_t hugepage_sz)
 123 {
 124         void *addr;
 125         int fd;
 126         long aligned_addr;
 127
 128         RTE_LOG(INFO, EAL, "Ask a virtual area of 0x%"PRIx64" bytes\n", *size);
 129
 130         fd = open("/dev/zero", O_RDONLY);
 131         if (fd < 0){
 132                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 133                 return NULL;
 134         }
 135         do {
 136                 addr = mmap(NULL, (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
 137                 if (addr == MAP_FAILED)
 138                         *size -= hugepage_sz;
 139         } while (addr == MAP_FAILED && *size > 0);
 140
 141         if (addr == MAP_FAILED) {
 142                 close(fd);
 143                 RTE_LOG(INFO, EAL, "Cannot get a virtual area\n");
 144                 return NULL;
 145         }
 146
 147         munmap(addr, (*size) + hugepage_sz);
 148         close(fd);
 149
 150         /* align addr to a huge page size boundary */
 151         aligned_addr = (long)addr;
 152         aligned_addr += (hugepage_sz - 1);
 153         aligned_addr &= (~(hugepage_sz - 1));
 154         addr = (void *)(aligned_addr);
 155
 156         RTE_LOG(INFO, EAL, "Virtual area found at %p (size = 0x%"PRIx64")\n",
 157                 addr, *size);
 158
 159         return addr;
 160 }
 161
 162 /*
 163  * Mmap all hugepages of hugepage table: it first open a file in
 164  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 165  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 166  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 167  * map continguous physical blocks in contiguous virtual blocks.
 168  */
 169 static int
 170 map_all_hugepages(struct hugepage *hugepg_tbl,
 171                 struct hugepage_info *hpi, int orig)
 172 {
 173         int fd;
 174         unsigned i;
 175         void *virtaddr;
 176         void *vma_addr = NULL;
 177         uint64_t vma_len = 0;
 178
 179         for (i = 0; i < hpi->num_pages[0]; i++) {
 180                 uint64_t hugepage_sz = hpi->hugepage_sz;
 181
 182                 if (orig) {
 183                         hugepg_tbl[i].file_id = i;
 184                         hugepg_tbl[i].size = hugepage_sz;
 185                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 186                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 187                                         hugepg_tbl[i].file_id);
 188                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 189                 }
 190 #ifndef RTE_ARCH_X86_64
 191                 /* for 32-bit systems, don't remap 1G pages, just reuse original
 192                  * map address as final map address.
 193                  */
 194                 else if (hugepage_sz == RTE_PGSIZE_1G){
 195                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 196                         hugepg_tbl[i].orig_va = NULL;
 197                         continue;
 198                 }
 199 #endif
 200                 else if (vma_len == 0) {
 201                         unsigned j, num_pages;
 202
 203                         /* reserve a virtual area for next contiguous
 204                          * physical block: count the number of
 205                          * contiguous physical pages. */
 206                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 207                                 if (hugepg_tbl[j].physaddr !=
 208                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 209                                         break;
 210                         }
 211                         num_pages = j - i;
 212                         vma_len = num_pages * hugepage_sz;
 213
 214                         /* get the biggest virtual memory area up to
 215                          * vma_len. If it fails, vma_addr is NULL, so
 216                          * let the kernel provide the address. */
 217                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 218                         if (vma_addr == NULL)
 219                                 vma_len = hugepage_sz;
 220                 }
 221
 222                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
 223                 if (fd < 0) {
 224                         RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
 225                                         strerror(errno));
 226                         return -1;
 227                 }
 228
 229                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 230                                 MAP_SHARED, fd, 0);
 231                 if (virtaddr == MAP_FAILED) {
 232                         RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__,
 233                                         strerror(errno));
 234                         close(fd);
 235                         return -1;
 236                 }
 237
 238                 if (orig) {
 239                         hugepg_tbl[i].orig_va = virtaddr;
 240                         memset(virtaddr, 0, hugepage_sz);
 241                 }
 242                 else {
 243                         hugepg_tbl[i].final_va = virtaddr;
 244                 }
 245
 246                 vma_addr = (char *)vma_addr + hugepage_sz;
 247                 vma_len -= hugepage_sz;
 248                 close(fd);
 249         }
 250         return 0;
 251 }
 252
 253 /* Unmap all hugepages from original mapping. */
 254 static int
 255 unmap_all_hugepages_orig(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 256 {
 257         unsigned i;
 258         for (i = 0; i < hpi->num_pages[0]; i++) {
 259                 if (hugepg_tbl[i].orig_va) {
 260                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 261                         hugepg_tbl[i].orig_va = NULL;
 262                 }
 263         }
 264         return 0;
 265 }
 266
 267 /*
 268  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 269  * it by browsing the /proc/self/pagemap special file.
 270  */
 271 static int
 272 find_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 273 {
 274         int fd;
 275         unsigned i;
 276         uint64_t page;
 277         unsigned long virt_pfn;
 278         int page_size;
 279
 280         /* standard page size */
 281         page_size = getpagesize();
 282
 283         fd = open("/proc/self/pagemap", O_RDONLY);
 284         if (fd < 0) {
 285                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 286                         __func__, strerror(errno));
 287                 return -1;
 288         }
 289
 290         for (i = 0; i < hpi->num_pages[0]; i++) {
 291                 off_t offset;
 292                 virt_pfn = (unsigned long)hugepg_tbl[i].orig_va /
 293                         page_size;
 294                 offset = sizeof(uint64_t) * virt_pfn;
 295                 if (lseek(fd, offset, SEEK_SET) != offset){
 296                         RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 297                                         __func__, strerror(errno));
 298                         close(fd);
 299                         return -1;
 300                 }
 301                 if (read(fd, &page, sizeof(uint64_t)) < 0) {
 302                         RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 303                                         __func__, strerror(errno));
 304                         close(fd);
 305                         return -1;
 306                 }
 307
 308                 /*
 309                  * the pfn (page frame number) are bits 0-54 (see
 310                  * pagemap.txt in linux Documentation)
 311                  */
 312                 hugepg_tbl[i].physaddr = ((page & 0x7fffffffffffffULL) * page_size);
 313         }
 314         close(fd);
 315         return 0;
 316 }
 317
 318 /*
 319  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 320  * page.
 321  */
 322 static int
 323 find_numasocket(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 324 {
 325         int socket_id;
 326         char *end, *nodestr;
 327         unsigned i, hp_count = 0;
 328         uint64_t virt_addr;
 329         char buf[BUFSIZ];
 330         char hugedir_str[PATH_MAX];
 331         FILE *f;
 332
 333         f = fopen("/proc/self/numa_maps", "r");
 334         if (f == NULL) {
 335                 RTE_LOG(INFO, EAL, "cannot open /proc/self/numa_maps,"
 336                                 " consider that all memory is in socket_id 0\n");
 337                 return 0;
 338         }
 339
 340         rte_snprintf(hugedir_str, sizeof(hugedir_str),
 341                         "%s/", hpi->hugedir);
 342
 343         /* parse numa map */
 344         while (fgets(buf, sizeof(buf), f) != NULL) {
 345
 346                 /* ignore non huge page */
 347                 if (strstr(buf, " huge ") == NULL &&
 348                                 strstr(buf, hugedir_str) == NULL)
 349                         continue;
 350
 351                 /* get zone addr */
 352                 virt_addr = strtoull(buf, &end, 16);
 353                 if (virt_addr == 0 || end == buf) {
 354                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 355                         goto error;
 356                 }
 357
 358                 /* get node id (socket id) */
 359                 nodestr = strstr(buf, " N");
 360                 if (nodestr == NULL) {
 361                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 362                         goto error;
 363                 }
 364                 nodestr += 2;
 365                 end = strstr(nodestr, "=");
 366                 if (end == NULL) {
 367                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 368                         goto error;
 369                 }
 370                 end[0] = '\0';
 371                 end = NULL;
 372
 373                 socket_id = strtoul(nodestr, &end, 0);
 374                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 375                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 376                         goto error;
 377                 }
 378
 379                 /* if we find this page in our mappings, set socket_id */
 380                 for (i = 0; i < hpi->num_pages[0]; i++) {
 381                         void *va = (void *)(unsigned long)virt_addr;
 382                         if (hugepg_tbl[i].orig_va == va) {
 383                                 hugepg_tbl[i].socket_id = socket_id;
 384                                 hp_count++;
 385                         }
 386                 }
 387         }
 388
 389         if (hp_count < hpi->num_pages[0])
 390                 goto error;
 391
 392         fclose(f);
 393         return 0;
 394
 395 error:
 396         fclose(f);
 397         return -1;
 398 }
 399
 400 /*
 401  * Sort the hugepg_tbl by physical address (lower addresses first). We
 402  * use a slow algorithm, but we won't have millions of pages, and this
 403  * is only done at init time.
 404  */
 405 static int
 406 sort_by_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 407 {
 408         unsigned i, j;
 409         int smallest_idx;
 410         uint64_t smallest_addr;
 411         struct hugepage tmp;
 412
 413         for (i = 0; i < hpi->num_pages[0]; i++) {
 414                 smallest_addr = 0;
 415                 smallest_idx = -1;
 416
 417                 /*
 418                  * browse all entries starting at 'i', and find the
 419                  * entry with the smallest addr
 420                  */
 421                 for (j=i; j< hpi->num_pages[0]; j++) {
 422
 423                         if (smallest_addr == 0 ||
 424                             hugepg_tbl[j].physaddr < smallest_addr) {
 425                                 smallest_addr = hugepg_tbl[j].physaddr;
 426                                 smallest_idx = j;
 427                         }
 428                 }
 429
 430                 /* should not happen */
 431                 if (smallest_idx == -1) {
 432                         RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
 433                         return -1;
 434                 }
 435
 436                 /* swap the 2 entries in the table */
 437                 memcpy(&tmp, &hugepg_tbl[smallest_idx], sizeof(struct hugepage));
 438                 memcpy(&hugepg_tbl[smallest_idx], &hugepg_tbl[i],
 439                                 sizeof(struct hugepage));
 440                 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage));
 441         }
 442         return 0;
 443 }
 444
 445 /*
 446  * Uses mmap to create a shared memory area for storage of data
 447  * Used in this file to store the hugepage file map on disk
 448  */
 449 static void *
 450 create_shared_memory(const char *filename, const size_t mem_size)
 451 {
 452         void *retval;
 453         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 454         if (fd < 0)
 455                 return NULL;
 456         if (ftruncate(fd, mem_size) < 0) {
 457                 close(fd);
 458                 return NULL;
 459         }
 460         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 461         close(fd);
 462         return retval;
 463 }
 464
 465 /*
 466  * this copies *active* hugepages from one hugepage table to another.
 467  * destination is typically the shared memory.
 468  */
 469 static int
 470 copy_hugepages_to_shared_mem(struct hugepage * dst, int dest_size,
 471                 const struct hugepage * src, int src_size)
 472 {
 473         int src_pos, dst_pos = 0;
 474
 475         for (src_pos = 0; src_pos < src_size; src_pos++) {
 476                 if (src[src_pos].final_va != NULL) {
 477                         /* error on overflow attempt */
 478                         if (dst_pos == dest_size)
 479                                 return -1;
 480                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage));
 481                         dst_pos++;
 482                 }
 483         }
 484         return 0;
 485 }
 486
 487 /*
 488  * unmaps hugepages that are not going to be used. since we originally allocate
 489  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 490  */
 491 static int
 492 unmap_unneeded_hugepages(struct hugepage *hugepg_tbl,
 493                 struct hugepage_info *hpi,
 494                 unsigned num_hp_info)
 495 {
 496         unsigned socket, size;
 497         int page, nrpages = 0;
 498         int fd;
 499
 500         /* get total number of hugepages */
 501         for (size = 0; size < num_hp_info; size++)
 502                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 503                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 504
 505         for (size = 0; size < num_hp_info; size++) {
 506                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 507                         unsigned pages_found = 0;
 508                         /* traverse until we have unmapped all the unused pages */
 509                         for (page = 0; page < nrpages; page++) {
 510                                 struct hugepage *hp = &hugepg_tbl[page];
 511
 512                                 /* find a page that matches the criteria */
 513                                 if ((hp->size == hpi[size].hugepage_sz) &&
 514                                                 (hp->socket_id == (int) socket)) {
 515
 516                                         /* if we skipped enough pages, unmap the rest */
 517                                         if (pages_found == hpi[size].num_pages[socket]) {
 518                                                 munmap(hp->final_va, hp->size);
 519                                                 hp->final_va = NULL;
 520                                         }
 521                                         else {
 522                                                 pages_found++;
 523                                         }
 524                                 } /* match page */
 525                         } /* foreach page */
 526                 } /* foreach socket */
 527         } /* foreach pagesize */
 528
 529         return 0;
 530 }
 531
 532 static inline uint64_t
 533 get_socket_mem_size(int socket)
 534 {
 535         uint64_t size = 0;
 536         unsigned i;
 537
 538         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 539                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 540                 if (hpi->hugedir != NULL)
 541                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 542         }
 543
 544         return (size);
 545 }
 546
 547 /*
 548  * This function is a NUMA-aware equivalent of calc_num_pages.
 549  * It takes in the list of hugepage sizes and the
 550  * number of pages thereof, and calculates the best number of
 551  * pages of each size to fulfill the request for <memory> ram
 552  */
 553 static int
 554 calc_num_pages_per_socket(uint64_t * memory,
 555                 struct hugepage_info *hp_info,
 556                 struct hugepage_info *hp_used,
 557                 unsigned num_hp_info)
 558 {
 559         unsigned socket, j, i = 0;
 560         unsigned requested, available;
 561         int total_num_pages = 0;
 562         uint64_t remaining_mem, cur_mem;
 563         uint64_t total_mem = internal_config.memory;
 564
 565         if (num_hp_info == 0)
 566                 return -1;
 567
 568         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 569                 /* if specific memory amounts per socket weren't requested */
 570                 if (internal_config.force_sockets == 0) {
 571                         /* take whatever is available */
 572                         memory[socket] = RTE_MIN(get_socket_mem_size(socket),
 573                                         total_mem);
 574                 }
 575                 /* skips if the memory on specific socket wasn't requested */
 576                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 577                         hp_used[i].hugedir = hp_info[i].hugedir;
 578                         hp_used[i].num_pages[socket] = RTE_MIN(
 579                                         memory[socket] / hp_info[i].hugepage_sz,
 580                                         hp_info[i].num_pages[socket]);
 581
 582                         cur_mem = hp_used[i].num_pages[socket] *
 583                                         hp_used[i].hugepage_sz;
 584
 585                         memory[socket] -= cur_mem;
 586                         total_mem -= cur_mem;
 587
 588                         total_num_pages += hp_used[i].num_pages[socket];
 589
 590                         /* check if we have met all memory requests */
 591                         if (memory[socket] == 0)
 592                                 break;
 593
 594                         /* check if we have any more pages left at this size, if so
 595                          * move on to next size */
 596                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
 597                                 continue;
 598                         /* At this point we know that there are more pages available that are
 599                          * bigger than the memory we want, so lets see if we can get enough
 600                          * from other page sizes.
 601                          */
 602                         remaining_mem = 0;
 603                         for (j = i+1; j < num_hp_info; j++)
 604                                 remaining_mem += hp_info[j].hugepage_sz *
 605                                 hp_info[j].num_pages[socket];
 606
 607                         /* is there enough other memory, if not allocate another page and quit */
 608                         if (remaining_mem < memory[socket]){
 609                                 cur_mem = RTE_MIN(memory[socket],
 610                                                 hp_info[i].hugepage_sz);
 611                                 memory[socket] -= cur_mem;
 612                                 total_mem -= cur_mem;
 613                                 hp_used[i].num_pages[socket]++;
 614                                 total_num_pages++;
 615                                 break; /* we are done with this socket*/
 616                         }
 617                 }
 618                 /* if we didn't satisfy all memory requirements per socket */
 619                 if (memory[socket] > 0) {
 620                         /* to prevent icc errors */
 621                         requested = (unsigned) (internal_config.socket_mem[socket] /
 622                                         0x100000);
 623                         available = requested -
 624                                         ((unsigned) (memory[socket] / 0x100000));
 625                         RTE_LOG(INFO, EAL, "Not enough memory available on socket %u! "
 626                                         "Requested: %uMB, available: %uMB\n", socket,
 627                                         requested, available);
 628                         return -1;
 629                 }
 630         }
 631
 632         /* if we didn't satisfy total memory requirements */
 633         if (total_mem > 0) {
 634                 requested = (unsigned) (internal_config.memory / 0x100000);
 635                 available = requested - (unsigned) (total_mem / 0x100000);
 636                 RTE_LOG(INFO, EAL, "Not enough memory available! Requested: %uMB,"
 637                                 " available: %uMB\n", requested, available);
 638                 return -1;
 639         }
 640         return total_num_pages;
 641 }
 642
 643 /*
 644  * Prepare physical memory mapping: fill configuration structure with
 645  * these infos, return 0 on success.
 646  *  1. map N huge pages in separate files in hugetlbfs
 647  *  2. find associated physical addr
 648  *  3. find associated NUMA socket ID
 649  *  4. sort all huge pages by physical address
 650  *  5. remap these N huge pages in the correct order
 651  *  6. unmap the first mapping
 652  *  7. fill memsegs in configuration with contiguous zones
 653  */
 654 static int
 655 rte_eal_hugepage_init(void)
 656 {
 657         struct rte_mem_config *mcfg;
 658         struct hugepage *hugepage, *tmp_hp = NULL;
 659         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
 660
 661         uint64_t memory[RTE_MAX_NUMA_NODES];
 662
 663         unsigned hp_offset;
 664         int i, j, new_memseg;
 665         int nrpages, total_pages = 0;
 666         void *addr;
 667
 668         memset(used_hp, 0, sizeof(used_hp));
 669
 670         /* get pointer to global configuration */
 671         mcfg = rte_eal_get_configuration()->mem_config;
 672
 673         /* for debug purposes, hugetlbfs can be disabled */
 674         if (internal_config.no_hugetlbfs) {
 675                 addr = malloc(internal_config.memory);
 676                 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
 677                 mcfg->memseg[0].addr = addr;
 678                 mcfg->memseg[0].len = internal_config.memory;
 679                 mcfg->memseg[0].socket_id = 0;
 680                 return 0;
 681         }
 682
 683
 684         /* calculate total number of hugepages available. at this point we haven't
 685          * yet started sorting them so they all are on socket 0 */
 686         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
 687                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
 688                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
 689
 690                 total_pages += internal_config.hugepage_info[i].num_pages[0];
 691         }
 692
 693         /*
 694          * allocate a memory area for hugepage table.
 695          * this isn't shared memory yet. due to the fact that we need some
 696          * processing done on these pages, shared memory will be created
 697          * at a later stage.
 698          */
 699         tmp_hp = malloc(total_pages * sizeof(struct hugepage));
 700         if (tmp_hp == NULL)
 701                 goto fail;
 702
 703         memset(tmp_hp, 0, total_pages * sizeof(struct hugepage));
 704
 705         hp_offset = 0; /* where we start the current page size entries */
 706
 707         /* map all hugepages and sort them */
 708         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
 709                 struct hugepage_info *hpi;
 710
 711                 /*
 712                  * we don't yet mark hugepages as used at this stage, so
 713                  * we just map all hugepages available to the system
 714                  * all hugepages are still located on socket 0
 715                  */
 716                 hpi = &internal_config.hugepage_info[i];
 717
 718                 if (hpi->num_pages == 0)
 719                         continue;
 720
 721                 /* map all hugepages available */
 722                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 1) < 0){
 723                         RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
 724                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 725                         goto fail;
 726                 }
 727
 728                 /* find physical addresses and sockets for each hugepage */
 729                 if (find_physaddr(&tmp_hp[hp_offset], hpi) < 0){
 730                         RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
 731                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 732                         goto fail;
 733                 }
 734
 735                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
 736                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
 737                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 738                         goto fail;
 739                 }
 740
 741                 if (sort_by_physaddr(&tmp_hp[hp_offset], hpi) < 0)
 742                         goto fail;
 743
 744                 /* remap all hugepages */
 745                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) < 0){
 746                         RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
 747                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 748                         goto fail;
 749                 }
 750
 751                 /* unmap original mappings */
 752                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
 753                         goto fail;
 754
 755                 /* we have processed a num of hugepages of this size, so inc offset */
 756                 hp_offset += hpi->num_pages[0];
 757         }
 758
 759         /* clean out the numbers of pages */
 760         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
 761                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
 762                         internal_config.hugepage_info[i].num_pages[j] = 0;
 763
 764         /* get hugepages for each socket */
 765         for (i = 0; i < total_pages; i++) {
 766                 int socket = tmp_hp[i].socket_id;
 767
 768                 /* find a hugepage info with right size and increment num_pages */
 769                 for (j = 0; j < (int) internal_config.num_hugepage_sizes; j++) {
 770                         if (tmp_hp[i].size ==
 771                                         internal_config.hugepage_info[j].hugepage_sz) {
 772                                 internal_config.hugepage_info[j].num_pages[socket]++;
 773                         }
 774                 }
 775         }
 776
 777         /* make a copy of socket_mem, needed for number of pages calculation */
 778         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 779                 memory[i] = internal_config.socket_mem[i];
 780
 781         /* calculate final number of pages */
 782         nrpages = calc_num_pages_per_socket(memory,
 783                         internal_config.hugepage_info, used_hp,
 784                         internal_config.num_hugepage_sizes);
 785
 786         /* error if not enough memory available */
 787         if (nrpages < 0)
 788                 goto fail;
 789
 790         /* reporting in! */
 791         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
 792                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 793                         if (used_hp[i].num_pages[j] > 0) {
 794                                 RTE_LOG(INFO, EAL,
 795                                                 "Requesting %u pages of size %uMB"
 796                                                 " from socket %i\n",
 797                                                 used_hp[i].num_pages[j],
 798                                                 (unsigned)
 799                                                         (used_hp[i].hugepage_sz / 0x100000),
 800                                                 j);
 801                         }
 802                 }
 803         }
 804
 805         /* create shared memory */
 806         hugepage = create_shared_memory(eal_hugepage_info_path(),
 807                                         nrpages * sizeof(struct hugepage));
 808
 809         if (hugepage == NULL) {
 810                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
 811                 goto fail;
 812         }
 813
 814         /*
 815          * unmap pages that we won't need (looks at used_hp).
 816          * also, sets final_va to NULL on pages that were unmapped.
 817          */
 818         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
 819                         internal_config.num_hugepage_sizes) < 0) {
 820                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
 821                 goto fail;
 822         }
 823
 824         /*
 825          * copy stuff from malloc'd hugepage* to the actual shared memory.
 826          * this procedure only copies those hugepages that have final_va
 827          * not NULL. has overflow protection.
 828          */
 829         if (copy_hugepages_to_shared_mem(hugepage, nrpages,
 830                         tmp_hp, total_pages) < 0) {
 831                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
 832                 goto fail;
 833         }
 834
 835         /* free the temporary hugepage table */
 836         free(tmp_hp);
 837         tmp_hp = NULL;
 838
 839         memset(mcfg->memseg, 0, sizeof(mcfg->memseg));
 840         j = -1;
 841         for (i = 0; i < nrpages; i++) {
 842                 new_memseg = 0;
 843
 844                 /* if this is a new section, create a new memseg */
 845                 if (i == 0)
 846                         new_memseg = 1;
 847                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
 848                         new_memseg = 1;
 849                 else if (hugepage[i].size != hugepage[i-1].size)
 850                         new_memseg = 1;
 851                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
 852                     hugepage[i].size)
 853                         new_memseg = 1;
 854                 else if (((unsigned long)hugepage[i].final_va -
 855                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
 856                         new_memseg = 1;
 857
 858                 if (new_memseg) {
 859                         j += 1;
 860                         if (j == RTE_MAX_MEMSEG)
 861                                 break;
 862
 863                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
 864                         mcfg->memseg[j].addr = hugepage[i].final_va;
 865                         mcfg->memseg[j].len = hugepage[i].size;
 866                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
 867                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
 868                 }
 869                 /* continuation of previous memseg */
 870                 else {
 871                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
 872                 }
 873                 hugepage[i].memseg_id = j;
 874         }
 875
 876         return 0;
 877
 878
 879 fail:
 880         if (tmp_hp)
 881                 free(tmp_hp);
 882         return -1;
 883 }
 884
 885 /*
 886  * uses fstat to report the size of a file on disk
 887  */
 888 static off_t
 889 getFileSize(int fd)
 890 {
 891         struct stat st;
 892         if (fstat(fd, &st) < 0)
 893                 return 0;
 894         return st.st_size;
 895 }
 896
 897 /*
 898  * This creates the memory mappings in the secondary process to match that of
 899  * the server process. It goes through each memory segment in the DPDK runtime
 900  * configuration and finds the hugepages which form that segment, mapping them
 901  * in order to form a contiguous block in the virtual memory space
 902  */
 903 static int
 904 rte_eal_hugepage_attach(void)
 905 {
 906         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 907         const struct hugepage *hp = NULL;
 908         unsigned num_hp = 0;
 909         unsigned i, s = 0; /* s used to track the segment number */
 910         off_t size;
 911         int fd, fd_zero = -1, fd_hugepage = -1;
 912
 913         if (aslr_enabled() > 0) {
 914                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
 915                                 "(ASLR) is enabled in the kernel.\n");
 916                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
 917                                 "into secondary processes\n");
 918         }
 919
 920         fd_zero = open("/dev/zero", O_RDONLY);
 921         if (fd_zero < 0) {
 922                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
 923                 goto error;
 924         }
 925         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
 926         if (fd_hugepage < 0) {
 927                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
 928                 goto error;
 929         }
 930
 931         size = getFileSize(fd_hugepage);
 932         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
 933         if (hp == NULL) {
 934                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
 935                 goto error;
 936         }
 937
 938         num_hp = size / sizeof(struct hugepage);
 939         RTE_LOG(DEBUG, EAL, "Analysing %u hugepages\n", num_hp);
 940
 941         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
 942                 void *addr, *base_addr;
 943                 uintptr_t offset = 0;
 944
 945                 /* fdzero is mmapped to get a contiguous block of virtual addresses
 946                  * get a block of free memory of the appropriate size -
 947                  * use mmap to attempt to get an identical address as server.
 948                  */
 949                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
 950                                 PROT_READ, MAP_PRIVATE, fd_zero, 0);
 951                 if (base_addr == MAP_FAILED || base_addr != mcfg->memseg[s].addr) {
 952                         RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
 953                                 "in /dev/zero to requested address [%p]\n",
 954                                 (unsigned long long)mcfg->memseg[s].len,
 955                                 mcfg->memseg[s].addr);
 956                         if (aslr_enabled() > 0)
 957                                 RTE_LOG(ERR, EAL, "It is recommended to disable ASLR in the kernel "
 958                                                 "and retry running both primary and secondary processes\n");
 959                         goto error;
 960                 }
 961                 /* free memory so we can map the hugepages into the space */
 962                 munmap(base_addr, mcfg->memseg[s].len);
 963
 964                 /* find the hugepages for this segment and map them
 965                  * we don't need to worry about order, as the server sorted the
 966                  * entries before it did the second mmap of them */
 967                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
 968                         if (hp[i].memseg_id == (int)s){
 969                                 fd = open(hp[i].filepath, O_RDWR);
 970                                 if (fd < 0) {
 971                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
 972                                                 hp[i].filepath);
 973                                         goto error;
 974                                 }
 975                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
 976                                                 hp[i].size, PROT_READ | PROT_WRITE,
 977                                                 MAP_SHARED | MAP_FIXED, fd, 0);
 978                                 close(fd); /* close file both on success and on failure */
 979                                 if (addr == MAP_FAILED) {
 980                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
 981                                                 hp[i].filepath);
 982                                         goto error;
 983                                 }
 984                                 offset+=hp[i].size;
 985                         }
 986                 }
 987                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
 988                                 (unsigned long long)mcfg->memseg[s].len);
 989                 s++;
 990         }
 991         close(fd_zero);
 992         close(fd_hugepage);
 993         return 0;
 994
 995 error:
 996         if (fd_zero >= 0)
 997                 close(fd_zero);
 998         if (fd_hugepage >= 0)
 999                 close(fd_hugepage);
1000         return -1;
1001 }
1002
1003 static int
1004 rte_eal_memdevice_init(void)
1005 {
1006         struct rte_config *config;
1007
1008         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1009                 return 0;
1010
1011         config = rte_eal_get_configuration();
1012         config->mem_config->nchannel = internal_config.force_nchannel;
1013         config->mem_config->nrank = internal_config.force_nrank;
1014
1015         return 0;
1016 }
1017
1018
1019 /* init memory subsystem */
1020 int
1021 rte_eal_memory_init(void)
1022 {
1023         RTE_LOG(INFO, EAL, "Setting up hugepage memory...\n");
1024         const int retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
1025                         rte_eal_hugepage_init() :
1026                         rte_eal_hugepage_attach();
1027         if (retval < 0)
1028                 return -1;
1029
1030         if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)
1031                 return -1;
1032
1033         return 0;
1034 }