lib/librte_eal/linuxapp/eal/eal_memory.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 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 /*   BSD LICENSE
  34  *
  35  *   Copyright(c) 2013 6WIND.
  36  *
  37  *   Redistribution and use in source and binary forms, with or without
  38  *   modification, are permitted provided that the following conditions
  39  *   are met:
  40  *
  41  *     * Redistributions of source code must retain the above copyright
  42  *       notice, this list of conditions and the following disclaimer.
  43  *     * Redistributions in binary form must reproduce the above copyright
  44  *       notice, this list of conditions and the following disclaimer in
  45  *       the documentation and/or other materials provided with the
  46  *       distribution.
  47  *     * Neither the name of 6WIND S.A. nor the names of its
  48  *       contributors may be used to endorse or promote products derived
  49  *       from this software without specific prior written permission.
  50  *
  51  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  52  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  53  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  54  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  55  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  56  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  57  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  58  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  59  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  60  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  61  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  62  */
  63
  64 #include <errno.h>
  65 #include <stdarg.h>
  66 #include <stdlib.h>
  67 #include <stdio.h>
  68 #include <stdint.h>
  69 #include <inttypes.h>
  70 #include <string.h>
  71 #include <stdarg.h>
  72 #include <sys/mman.h>
  73 #include <sys/types.h>
  74 #include <sys/stat.h>
  75 #include <sys/queue.h>
  76 #include <sys/file.h>
  77 #include <unistd.h>
  78 #include <limits.h>
  79 #include <errno.h>
  80 #include <sys/ioctl.h>
  81 #include <sys/time.h>
  82
  83 #include <rte_log.h>
  84 #include <rte_memory.h>
  85 #include <rte_memzone.h>
  86 #include <rte_launch.h>
  87 #include <rte_tailq.h>
  88 #include <rte_eal.h>
  89 #include <rte_eal_memconfig.h>
  90 #include <rte_per_lcore.h>
  91 #include <rte_lcore.h>
  92 #include <rte_common.h>
  93 #include <rte_string_fns.h>
  94
  95 #include "eal_private.h"
  96 #include "eal_internal_cfg.h"
  97 #include "eal_filesystem.h"
  98 #include "eal_hugepages.h"
  99
 100 /**
 101  * @file
 102  * Huge page mapping under linux
 103  *
 104  * To reserve a big contiguous amount of memory, we use the hugepage
 105  * feature of linux. For that, we need to have hugetlbfs mounted. This
 106  * code will create many files in this directory (one per page) and
 107  * map them in virtual memory. For each page, we will retrieve its
 108  * physical address and remap it in order to have a virtual contiguous
 109  * zone as well as a physical contiguous zone.
 110  */
 111
 112
 113 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
 114
 115 /*
 116  * Check whether address-space layout randomization is enabled in
 117  * the kernel. This is important for multi-process as it can prevent
 118  * two processes mapping data to the same virtual address
 119  * Returns:
 120  *    0 - address space randomization disabled
 121  *    1/2 - address space randomization enabled
 122  *    negative error code on error
 123  */
 124 static int
 125 aslr_enabled(void)
 126 {
 127         char c;
 128         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 129         if (fd < 0)
 130                 return -errno;
 131         retval = read(fd, &c, 1);
 132         close(fd);
 133         if (retval < 0)
 134                 return -errno;
 135         if (retval == 0)
 136                 return -EIO;
 137         switch (c) {
 138                 case '0' : return 0;
 139                 case '1' : return 1;
 140                 case '2' : return 2;
 141                 default: return -EINVAL;
 142         }
 143 }
 144
 145 /*
 146  * Try to mmap *size bytes in /dev/zero. If it is succesful, return the
 147  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 148  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 149  * 0. In this case, return NULL. Note: this function returns an address
 150  * which is a multiple of hugepage size.
 151  */
 152 static void *
 153 get_virtual_area(size_t *size, size_t hugepage_sz)
 154 {
 155         void *addr;
 156         int fd;
 157         long aligned_addr;
 158
 159         RTE_LOG(INFO, EAL, "Ask a virtual area of 0x%zu bytes\n", *size);
 160
 161         fd = open("/dev/zero", O_RDONLY);
 162         if (fd < 0){
 163                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 164                 return NULL;
 165         }
 166         do {
 167                 addr = mmap(NULL, (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
 168                 if (addr == MAP_FAILED)
 169                         *size -= hugepage_sz;
 170         } while (addr == MAP_FAILED && *size > 0);
 171
 172         if (addr == MAP_FAILED) {
 173                 close(fd);
 174                 RTE_LOG(INFO, EAL, "Cannot get a virtual area\n");
 175                 return NULL;
 176         }
 177
 178         munmap(addr, (*size) + hugepage_sz);
 179         close(fd);
 180
 181         /* align addr to a huge page size boundary */
 182         aligned_addr = (long)addr;
 183         aligned_addr += (hugepage_sz - 1);
 184         aligned_addr &= (~(hugepage_sz - 1));
 185         addr = (void *)(aligned_addr);
 186
 187         RTE_LOG(INFO, EAL, "Virtual area found at %p (size = 0x%zx)\n",
 188                 addr, *size);
 189
 190         return addr;
 191 }
 192
 193 /*
 194  * Mmap all hugepages of hugepage table: it first open a file in
 195  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 196  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 197  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 198  * map continguous physical blocks in contiguous virtual blocks.
 199  */
 200 static int
 201 map_all_hugepages(struct hugepage *hugepg_tbl,
 202                 struct hugepage_info *hpi, int orig)
 203 {
 204         int fd;
 205         unsigned i;
 206         void *virtaddr;
 207         void *vma_addr = NULL;
 208         size_t vma_len = 0;
 209
 210         for (i = 0; i < hpi->num_pages[0]; i++) {
 211                 size_t hugepage_sz = hpi->hugepage_sz;
 212
 213                 if (orig) {
 214                         hugepg_tbl[i].file_id = i;
 215                         hugepg_tbl[i].size = hugepage_sz;
 216                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 217                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 218                                         hugepg_tbl[i].file_id);
 219                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 220                 }
 221 #ifndef RTE_ARCH_X86_64
 222                 /* for 32-bit systems, don't remap 1G pages, just reuse original
 223                  * map address as final map address.
 224                  */
 225                 else if (hugepage_sz == RTE_PGSIZE_1G){
 226                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 227                         hugepg_tbl[i].orig_va = NULL;
 228                         continue;
 229                 }
 230 #endif
 231                 else if (vma_len == 0) {
 232                         unsigned j, num_pages;
 233
 234                         /* reserve a virtual area for next contiguous
 235                          * physical block: count the number of
 236                          * contiguous physical pages. */
 237                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 238                                 if (hugepg_tbl[j].physaddr !=
 239                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 240                                         break;
 241                         }
 242                         num_pages = j - i;
 243                         vma_len = num_pages * hugepage_sz;
 244
 245                         /* get the biggest virtual memory area up to
 246                          * vma_len. If it fails, vma_addr is NULL, so
 247                          * let the kernel provide the address. */
 248                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 249                         if (vma_addr == NULL)
 250                                 vma_len = hugepage_sz;
 251                 }
 252
 253                 /* try to create hugepage file */
 254                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
 255                 if (fd < 0) {
 256                         RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
 257                                         strerror(errno));
 258                         return -1;
 259                 }
 260
 261                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 262                                 MAP_SHARED, fd, 0);
 263                 if (virtaddr == MAP_FAILED) {
 264                         RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__,
 265                                         strerror(errno));
 266                         close(fd);
 267                         return -1;
 268                 }
 269
 270                 if (orig) {
 271                         hugepg_tbl[i].orig_va = virtaddr;
 272                         memset(virtaddr, 0, hugepage_sz);
 273                 }
 274                 else {
 275                         hugepg_tbl[i].final_va = virtaddr;
 276                 }
 277
 278                 /* set shared flock on the file. */
 279                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 280                         RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
 281                                 __func__, strerror(errno));
 282                         close(fd);
 283                         return -1;
 284                 }
 285
 286                 close(fd);
 287
 288                 vma_addr = (char *)vma_addr + hugepage_sz;
 289                 vma_len -= hugepage_sz;
 290         }
 291         return 0;
 292 }
 293
 294 /* Unmap all hugepages from original mapping. */
 295 static int
 296 unmap_all_hugepages_orig(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 297 {
 298         unsigned i;
 299         for (i = 0; i < hpi->num_pages[0]; i++) {
 300                 if (hugepg_tbl[i].orig_va) {
 301                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 302                         hugepg_tbl[i].orig_va = NULL;
 303                 }
 304         }
 305         return 0;
 306 }
 307
 308 /*
 309  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 310  * it by browsing the /proc/self/pagemap special file.
 311  */
 312 static int
 313 find_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 314 {
 315         int fd;
 316         unsigned i;
 317         uint64_t page;
 318         unsigned long virt_pfn;
 319         int page_size;
 320
 321         /* standard page size */
 322         page_size = getpagesize();
 323
 324         fd = open("/proc/self/pagemap", O_RDONLY);
 325         if (fd < 0) {
 326                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 327                         __func__, strerror(errno));
 328                 return -1;
 329         }
 330
 331         for (i = 0; i < hpi->num_pages[0]; i++) {
 332                 off_t offset;
 333                 virt_pfn = (unsigned long)hugepg_tbl[i].orig_va /
 334                         page_size;
 335                 offset = sizeof(uint64_t) * virt_pfn;
 336                 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 337                         RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 338                                         __func__, strerror(errno));
 339                         close(fd);
 340                         return -1;
 341                 }
 342                 if (read(fd, &page, sizeof(uint64_t)) < 0) {
 343                         RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 344                                         __func__, strerror(errno));
 345                         close(fd);
 346                         return -1;
 347                 }
 348
 349                 /*
 350                  * the pfn (page frame number) are bits 0-54 (see
 351                  * pagemap.txt in linux Documentation)
 352                  */
 353                 hugepg_tbl[i].physaddr = ((page & 0x7fffffffffffffULL) * page_size);
 354         }
 355         close(fd);
 356         return 0;
 357 }
 358
 359 /*
 360  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 361  * page.
 362  */
 363 static int
 364 find_numasocket(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 365 {
 366         int socket_id;
 367         char *end, *nodestr;
 368         unsigned i, hp_count = 0;
 369         uint64_t virt_addr;
 370         char buf[BUFSIZ];
 371         char hugedir_str[PATH_MAX];
 372         FILE *f;
 373
 374         f = fopen("/proc/self/numa_maps", "r");
 375         if (f == NULL) {
 376                 RTE_LOG(INFO, EAL, "cannot open /proc/self/numa_maps,"
 377                                 " consider that all memory is in socket_id 0\n");
 378                 return 0;
 379         }
 380
 381         rte_snprintf(hugedir_str, sizeof(hugedir_str),
 382                         "%s/", hpi->hugedir);
 383
 384         /* parse numa map */
 385         while (fgets(buf, sizeof(buf), f) != NULL) {
 386
 387                 /* ignore non huge page */
 388                 if (strstr(buf, " huge ") == NULL &&
 389                                 strstr(buf, hugedir_str) == NULL)
 390                         continue;
 391
 392                 /* get zone addr */
 393                 virt_addr = strtoull(buf, &end, 16);
 394                 if (virt_addr == 0 || end == buf) {
 395                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 396                         goto error;
 397                 }
 398
 399                 /* get node id (socket id) */
 400                 nodestr = strstr(buf, " N");
 401                 if (nodestr == NULL) {
 402                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 403                         goto error;
 404                 }
 405                 nodestr += 2;
 406                 end = strstr(nodestr, "=");
 407                 if (end == NULL) {
 408                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 409                         goto error;
 410                 }
 411                 end[0] = '\0';
 412                 end = NULL;
 413
 414                 socket_id = strtoul(nodestr, &end, 0);
 415                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 416                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 417                         goto error;
 418                 }
 419
 420                 /* if we find this page in our mappings, set socket_id */
 421                 for (i = 0; i < hpi->num_pages[0]; i++) {
 422                         void *va = (void *)(unsigned long)virt_addr;
 423                         if (hugepg_tbl[i].orig_va == va) {
 424                                 hugepg_tbl[i].socket_id = socket_id;
 425                                 hp_count++;
 426                         }
 427                 }
 428         }
 429
 430         if (hp_count < hpi->num_pages[0])
 431                 goto error;
 432
 433         fclose(f);
 434         return 0;
 435
 436 error:
 437         fclose(f);
 438         return -1;
 439 }
 440
 441 /*
 442  * Sort the hugepg_tbl by physical address (lower addresses first). We
 443  * use a slow algorithm, but we won't have millions of pages, and this
 444  * is only done at init time.
 445  */
 446 static int
 447 sort_by_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 448 {
 449         unsigned i, j;
 450         int smallest_idx;
 451         uint64_t smallest_addr;
 452         struct hugepage tmp;
 453
 454         for (i = 0; i < hpi->num_pages[0]; i++) {
 455                 smallest_addr = 0;
 456                 smallest_idx = -1;
 457
 458                 /*
 459                  * browse all entries starting at 'i', and find the
 460                  * entry with the smallest addr
 461                  */
 462                 for (j=i; j< hpi->num_pages[0]; j++) {
 463
 464                         if (smallest_addr == 0 ||
 465                             hugepg_tbl[j].physaddr < smallest_addr) {
 466                                 smallest_addr = hugepg_tbl[j].physaddr;
 467                                 smallest_idx = j;
 468                         }
 469                 }
 470
 471                 /* should not happen */
 472                 if (smallest_idx == -1) {
 473                         RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
 474                         return -1;
 475                 }
 476
 477                 /* swap the 2 entries in the table */
 478                 memcpy(&tmp, &hugepg_tbl[smallest_idx], sizeof(struct hugepage));
 479                 memcpy(&hugepg_tbl[smallest_idx], &hugepg_tbl[i],
 480                                 sizeof(struct hugepage));
 481                 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage));
 482         }
 483         return 0;
 484 }
 485
 486 /*
 487  * Uses mmap to create a shared memory area for storage of data
 488  * Used in this file to store the hugepage file map on disk
 489  */
 490 static void *
 491 create_shared_memory(const char *filename, const size_t mem_size)
 492 {
 493         void *retval;
 494         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 495         if (fd < 0)
 496                 return NULL;
 497         if (ftruncate(fd, mem_size) < 0) {
 498                 close(fd);
 499                 return NULL;
 500         }
 501         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 502         close(fd);
 503         return retval;
 504 }
 505
 506 /*
 507  * this copies *active* hugepages from one hugepage table to another.
 508  * destination is typically the shared memory.
 509  */
 510 static int
 511 copy_hugepages_to_shared_mem(struct hugepage * dst, int dest_size,
 512                 const struct hugepage * src, int src_size)
 513 {
 514         int src_pos, dst_pos = 0;
 515
 516         for (src_pos = 0; src_pos < src_size; src_pos++) {
 517                 if (src[src_pos].final_va != NULL) {
 518                         /* error on overflow attempt */
 519                         if (dst_pos == dest_size)
 520                                 return -1;
 521                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage));
 522                         dst_pos++;
 523                 }
 524         }
 525         return 0;
 526 }
 527
 528 /*
 529  * unmaps hugepages that are not going to be used. since we originally allocate
 530  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 531  */
 532 static int
 533 unmap_unneeded_hugepages(struct hugepage *hugepg_tbl,
 534                 struct hugepage_info *hpi,
 535                 unsigned num_hp_info)
 536 {
 537         unsigned socket, size;
 538         int page, nrpages = 0;
 539
 540         /* get total number of hugepages */
 541         for (size = 0; size < num_hp_info; size++)
 542                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 543                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 544
 545         for (size = 0; size < num_hp_info; size++) {
 546                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 547                         unsigned pages_found = 0;
 548                         /* traverse until we have unmapped all the unused pages */
 549                         for (page = 0; page < nrpages; page++) {
 550                                 struct hugepage *hp = &hugepg_tbl[page];
 551
 552                                 /* find a page that matches the criteria */
 553                                 if ((hp->size == hpi[size].hugepage_sz) &&
 554                                                 (hp->socket_id == (int) socket)) {
 555
 556                                         /* if we skipped enough pages, unmap the rest */
 557                                         if (pages_found == hpi[size].num_pages[socket]) {
 558                                                 munmap(hp->final_va, hp->size);
 559                                                 hp->final_va = NULL;
 560                                                 if (remove(hp->filepath) == -1) {
 561                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 562                                                                         __func__, hp->filepath, strerror(errno));
 563                                                         return -1;
 564                                                 }
 565                                         }
 566                                         /* lock the page and skip */
 567                                         else
 568                                                 pages_found++;
 569
 570                                 } /* match page */
 571                         } /* foreach page */
 572                 } /* foreach socket */
 573         } /* foreach pagesize */
 574
 575         return 0;
 576 }
 577
 578 static inline uint64_t
 579 get_socket_mem_size(int socket)
 580 {
 581         uint64_t size = 0;
 582         unsigned i;
 583
 584         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 585                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 586                 if (hpi->hugedir != NULL)
 587                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 588         }
 589
 590         return (size);
 591 }
 592
 593 /*
 594  * This function is a NUMA-aware equivalent of calc_num_pages.
 595  * It takes in the list of hugepage sizes and the
 596  * number of pages thereof, and calculates the best number of
 597  * pages of each size to fulfill the request for <memory> ram
 598  */
 599 static int
 600 calc_num_pages_per_socket(uint64_t * memory,
 601                 struct hugepage_info *hp_info,
 602                 struct hugepage_info *hp_used,
 603                 unsigned num_hp_info)
 604 {
 605         unsigned socket, j, i = 0;
 606         unsigned requested, available;
 607         int total_num_pages = 0;
 608         uint64_t remaining_mem, cur_mem;
 609         uint64_t total_mem = internal_config.memory;
 610
 611         if (num_hp_info == 0)
 612                 return -1;
 613
 614         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 615                 /* if specific memory amounts per socket weren't requested */
 616                 if (internal_config.force_sockets == 0) {
 617                         /* take whatever is available */
 618                         memory[socket] = RTE_MIN(get_socket_mem_size(socket),
 619                                         total_mem);
 620                 }
 621                 /* skips if the memory on specific socket wasn't requested */
 622                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 623                         hp_used[i].hugedir = hp_info[i].hugedir;
 624                         hp_used[i].num_pages[socket] = RTE_MIN(
 625                                         memory[socket] / hp_info[i].hugepage_sz,
 626                                         hp_info[i].num_pages[socket]);
 627
 628                         cur_mem = hp_used[i].num_pages[socket] *
 629                                         hp_used[i].hugepage_sz;
 630
 631                         memory[socket] -= cur_mem;
 632                         total_mem -= cur_mem;
 633
 634                         total_num_pages += hp_used[i].num_pages[socket];
 635
 636                         /* check if we have met all memory requests */
 637                         if (memory[socket] == 0)
 638                                 break;
 639
 640                         /* check if we have any more pages left at this size, if so
 641                          * move on to next size */
 642                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
 643                                 continue;
 644                         /* At this point we know that there are more pages available that are
 645                          * bigger than the memory we want, so lets see if we can get enough
 646                          * from other page sizes.
 647                          */
 648                         remaining_mem = 0;
 649                         for (j = i+1; j < num_hp_info; j++)
 650                                 remaining_mem += hp_info[j].hugepage_sz *
 651                                 hp_info[j].num_pages[socket];
 652
 653                         /* is there enough other memory, if not allocate another page and quit */
 654                         if (remaining_mem < memory[socket]){
 655                                 cur_mem = RTE_MIN(memory[socket],
 656                                                 hp_info[i].hugepage_sz);
 657                                 memory[socket] -= cur_mem;
 658                                 total_mem -= cur_mem;
 659                                 hp_used[i].num_pages[socket]++;
 660                                 total_num_pages++;
 661                                 break; /* we are done with this socket*/
 662                         }
 663                 }
 664                 /* if we didn't satisfy all memory requirements per socket */
 665                 if (memory[socket] > 0) {
 666                         /* to prevent icc errors */
 667                         requested = (unsigned) (internal_config.socket_mem[socket] /
 668                                         0x100000);
 669                         available = requested -
 670                                         ((unsigned) (memory[socket] / 0x100000));
 671                         RTE_LOG(INFO, EAL, "Not enough memory available on socket %u! "
 672                                         "Requested: %uMB, available: %uMB\n", socket,
 673                                         requested, available);
 674                         return -1;
 675                 }
 676         }
 677
 678         /* if we didn't satisfy total memory requirements */
 679         if (total_mem > 0) {
 680                 requested = (unsigned) (internal_config.memory / 0x100000);
 681                 available = requested - (unsigned) (total_mem / 0x100000);
 682                 RTE_LOG(INFO, EAL, "Not enough memory available! Requested: %uMB,"
 683                                 " available: %uMB\n", requested, available);
 684                 return -1;
 685         }
 686         return total_num_pages;
 687 }
 688
 689 /*
 690  * Prepare physical memory mapping: fill configuration structure with
 691  * these infos, return 0 on success.
 692  *  1. map N huge pages in separate files in hugetlbfs
 693  *  2. find associated physical addr
 694  *  3. find associated NUMA socket ID
 695  *  4. sort all huge pages by physical address
 696  *  5. remap these N huge pages in the correct order
 697  *  6. unmap the first mapping
 698  *  7. fill memsegs in configuration with contiguous zones
 699  */
 700 static int
 701 rte_eal_hugepage_init(void)
 702 {
 703         struct rte_mem_config *mcfg;
 704         struct hugepage *hugepage, *tmp_hp = NULL;
 705         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
 706
 707         uint64_t memory[RTE_MAX_NUMA_NODES];
 708
 709         unsigned hp_offset;
 710         int i, j, new_memseg;
 711         int nrpages, total_pages = 0;
 712         void *addr;
 713
 714         memset(used_hp, 0, sizeof(used_hp));
 715
 716         /* get pointer to global configuration */
 717         mcfg = rte_eal_get_configuration()->mem_config;
 718
 719         /* for debug purposes, hugetlbfs can be disabled */
 720         if (internal_config.no_hugetlbfs) {
 721                 addr = malloc(internal_config.memory);
 722                 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
 723                 mcfg->memseg[0].addr = addr;
 724                 mcfg->memseg[0].len = internal_config.memory;
 725                 mcfg->memseg[0].socket_id = 0;
 726                 return 0;
 727         }
 728
 729
 730         /* calculate total number of hugepages available. at this point we haven't
 731          * yet started sorting them so they all are on socket 0 */
 732         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
 733                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
 734                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
 735
 736                 total_pages += internal_config.hugepage_info[i].num_pages[0];
 737         }
 738
 739         /*
 740          * allocate a memory area for hugepage table.
 741          * this isn't shared memory yet. due to the fact that we need some
 742          * processing done on these pages, shared memory will be created
 743          * at a later stage.
 744          */
 745         tmp_hp = malloc(total_pages * sizeof(struct hugepage));
 746         if (tmp_hp == NULL)
 747                 goto fail;
 748
 749         memset(tmp_hp, 0, total_pages * sizeof(struct hugepage));
 750
 751         hp_offset = 0; /* where we start the current page size entries */
 752
 753         /* map all hugepages and sort them */
 754         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
 755                 struct hugepage_info *hpi;
 756
 757                 /*
 758                  * we don't yet mark hugepages as used at this stage, so
 759                  * we just map all hugepages available to the system
 760                  * all hugepages are still located on socket 0
 761                  */
 762                 hpi = &internal_config.hugepage_info[i];
 763
 764                 if (hpi->num_pages == 0)
 765                         continue;
 766
 767                 /* map all hugepages available */
 768                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 1) < 0){
 769                         RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
 770                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 771                         goto fail;
 772                 }
 773
 774                 /* find physical addresses and sockets for each hugepage */
 775                 if (find_physaddr(&tmp_hp[hp_offset], hpi) < 0){
 776                         RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
 777                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 778                         goto fail;
 779                 }
 780
 781                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
 782                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
 783                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 784                         goto fail;
 785                 }
 786
 787                 if (sort_by_physaddr(&tmp_hp[hp_offset], hpi) < 0)
 788                         goto fail;
 789
 790                 /* remap all hugepages */
 791                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) < 0){
 792                         RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
 793                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 794                         goto fail;
 795                 }
 796
 797                 /* unmap original mappings */
 798                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
 799                         goto fail;
 800
 801                 /* we have processed a num of hugepages of this size, so inc offset */
 802                 hp_offset += hpi->num_pages[0];
 803         }
 804
 805         /* clean out the numbers of pages */
 806         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
 807                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
 808                         internal_config.hugepage_info[i].num_pages[j] = 0;
 809
 810         /* get hugepages for each socket */
 811         for (i = 0; i < total_pages; i++) {
 812                 int socket = tmp_hp[i].socket_id;
 813
 814                 /* find a hugepage info with right size and increment num_pages */
 815                 for (j = 0; j < (int) internal_config.num_hugepage_sizes; j++) {
 816                         if (tmp_hp[i].size ==
 817                                         internal_config.hugepage_info[j].hugepage_sz) {
 818                                 internal_config.hugepage_info[j].num_pages[socket]++;
 819                         }
 820                 }
 821         }
 822
 823         /* make a copy of socket_mem, needed for number of pages calculation */
 824         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 825                 memory[i] = internal_config.socket_mem[i];
 826
 827         /* calculate final number of pages */
 828         nrpages = calc_num_pages_per_socket(memory,
 829                         internal_config.hugepage_info, used_hp,
 830                         internal_config.num_hugepage_sizes);
 831
 832         /* error if not enough memory available */
 833         if (nrpages < 0)
 834                 goto fail;
 835
 836         /* reporting in! */
 837         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
 838                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 839                         if (used_hp[i].num_pages[j] > 0) {
 840                                 RTE_LOG(INFO, EAL,
 841                                                 "Requesting %u pages of size %uMB"
 842                                                 " from socket %i\n",
 843                                                 used_hp[i].num_pages[j],
 844                                                 (unsigned)
 845                                                         (used_hp[i].hugepage_sz / 0x100000),
 846                                                 j);
 847                         }
 848                 }
 849         }
 850
 851         /* create shared memory */
 852         hugepage = create_shared_memory(eal_hugepage_info_path(),
 853                                         nrpages * sizeof(struct hugepage));
 854
 855         if (hugepage == NULL) {
 856                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
 857                 goto fail;
 858         }
 859
 860         /*
 861          * unmap pages that we won't need (looks at used_hp).
 862          * also, sets final_va to NULL on pages that were unmapped.
 863          */
 864         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
 865                         internal_config.num_hugepage_sizes) < 0) {
 866                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
 867                 goto fail;
 868         }
 869
 870         /*
 871          * copy stuff from malloc'd hugepage* to the actual shared memory.
 872          * this procedure only copies those hugepages that have final_va
 873          * not NULL. has overflow protection.
 874          */
 875         if (copy_hugepages_to_shared_mem(hugepage, nrpages,
 876                         tmp_hp, total_pages) < 0) {
 877                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
 878                 goto fail;
 879         }
 880
 881         /* free the temporary hugepage table */
 882         free(tmp_hp);
 883         tmp_hp = NULL;
 884
 885         memset(mcfg->memseg, 0, sizeof(mcfg->memseg));
 886         j = -1;
 887         for (i = 0; i < nrpages; i++) {
 888                 new_memseg = 0;
 889
 890                 /* if this is a new section, create a new memseg */
 891                 if (i == 0)
 892                         new_memseg = 1;
 893                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
 894                         new_memseg = 1;
 895                 else if (hugepage[i].size != hugepage[i-1].size)
 896                         new_memseg = 1;
 897                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
 898                     hugepage[i].size)
 899                         new_memseg = 1;
 900                 else if (((unsigned long)hugepage[i].final_va -
 901                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
 902                         new_memseg = 1;
 903
 904                 if (new_memseg) {
 905                         j += 1;
 906                         if (j == RTE_MAX_MEMSEG)
 907                                 break;
 908
 909                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
 910                         mcfg->memseg[j].addr = hugepage[i].final_va;
 911                         mcfg->memseg[j].len = hugepage[i].size;
 912                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
 913                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
 914                 }
 915                 /* continuation of previous memseg */
 916                 else {
 917                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
 918                 }
 919                 hugepage[i].memseg_id = j;
 920         }
 921
 922         if (i < nrpages) {
 923                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
 924                         "from %d requested\n"
 925                         "Current %s=%d is not enough\n"
 926                         "Please either increase it or request less amount "
 927                         "of memory.\n",
 928                         i, nrpages, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
 929                         RTE_MAX_MEMSEG);
 930                 return (-ENOMEM);
 931         }
 932
 933
 934         return 0;
 935
 936
 937 fail:
 938         if (tmp_hp)
 939                 free(tmp_hp);
 940         return -1;
 941 }
 942
 943 /*
 944  * uses fstat to report the size of a file on disk
 945  */
 946 static off_t
 947 getFileSize(int fd)
 948 {
 949         struct stat st;
 950         if (fstat(fd, &st) < 0)
 951                 return 0;
 952         return st.st_size;
 953 }
 954
 955 /*
 956  * This creates the memory mappings in the secondary process to match that of
 957  * the server process. It goes through each memory segment in the DPDK runtime
 958  * configuration and finds the hugepages which form that segment, mapping them
 959  * in order to form a contiguous block in the virtual memory space
 960  */
 961 static int
 962 rte_eal_hugepage_attach(void)
 963 {
 964         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 965         const struct hugepage *hp = NULL;
 966         unsigned num_hp = 0;
 967         unsigned i, s = 0; /* s used to track the segment number */
 968         off_t size;
 969         int fd, fd_zero = -1, fd_hugepage = -1;
 970
 971         if (aslr_enabled() > 0) {
 972                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
 973                                 "(ASLR) is enabled in the kernel.\n");
 974                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
 975                                 "into secondary processes\n");
 976         }
 977
 978         fd_zero = open("/dev/zero", O_RDONLY);
 979         if (fd_zero < 0) {
 980                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
 981                 goto error;
 982         }
 983         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
 984         if (fd_hugepage < 0) {
 985                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
 986                 goto error;
 987         }
 988
 989         /* map all segments into memory to make sure we get the addrs */
 990         for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
 991                 void *base_addr;
 992
 993                 /*
 994                  * the first memory segment with len==0 is the one that
 995                  * follows the last valid segment.
 996                  */
 997                 if (mcfg->memseg[s].len == 0)
 998                         break;
 999
1000                 /*
1001                  * fdzero is mmapped to get a contiguous block of virtual
1002                  * addresses of the appropriate memseg size.
1003                  * use mmap to get identical addresses as the primary process.
1004                  */
1005                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1006                                  PROT_READ, MAP_PRIVATE, fd_zero, 0);
1007                 if (base_addr == MAP_FAILED ||
1008                     base_addr != mcfg->memseg[s].addr) {
1009                         RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1010                                 "in /dev/zero to requested address [%p]\n",
1011                                 (unsigned long long)mcfg->memseg[s].len,
1012                                 mcfg->memseg[s].addr);
1013                         if (aslr_enabled() > 0) {
1014                                 RTE_LOG(ERR, EAL, "It is recommended to "
1015                                         "disable ASLR in the kernel "
1016                                         "and retry running both primary "
1017                                         "and secondary processes\n");
1018                         }
1019                         goto error;
1020                 }
1021         }
1022
1023         size = getFileSize(fd_hugepage);
1024         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1025         if (hp == NULL) {
1026                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1027                 goto error;
1028         }
1029
1030         num_hp = size / sizeof(struct hugepage);
1031         RTE_LOG(DEBUG, EAL, "Analysing %u hugepages\n", num_hp);
1032
1033         s = 0;
1034         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1035                 void *addr, *base_addr;
1036                 uintptr_t offset = 0;
1037
1038                 /*
1039                  * free previously mapped memory so we can map the
1040                  * hugepages into the space
1041                  */
1042                 base_addr = mcfg->memseg[s].addr;
1043                 munmap(base_addr, mcfg->memseg[s].len);
1044
1045                 /* find the hugepages for this segment and map them
1046                  * we don't need to worry about order, as the server sorted the
1047                  * entries before it did the second mmap of them */
1048                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1049                         if (hp[i].memseg_id == (int)s){
1050                                 fd = open(hp[i].filepath, O_RDWR);
1051                                 if (fd < 0) {
1052                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1053                                                 hp[i].filepath);
1054                                         goto error;
1055                                 }
1056                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1057                                                 hp[i].size, PROT_READ | PROT_WRITE,
1058                                                 MAP_SHARED | MAP_FIXED, fd, 0);
1059                                 close(fd); /* close file both on success and on failure */
1060                                 if (addr == MAP_FAILED) {
1061                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1062                                                 hp[i].filepath);
1063                                         goto error;
1064                                 }
1065                                 offset+=hp[i].size;
1066                         }
1067                 }
1068                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1069                                 (unsigned long long)mcfg->memseg[s].len);
1070                 s++;
1071         }
1072         /* unmap the hugepage config file, since we are done using it */
1073         munmap((void *)(uintptr_t)hp, size);
1074         close(fd_zero);
1075         close(fd_hugepage);
1076         return 0;
1077
1078 error:
1079         if (fd_zero >= 0)
1080                 close(fd_zero);
1081         if (fd_hugepage >= 0)
1082                 close(fd_hugepage);
1083         return -1;
1084 }
1085
1086 static int
1087 rte_eal_memdevice_init(void)
1088 {
1089         struct rte_config *config;
1090
1091         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1092                 return 0;
1093
1094         config = rte_eal_get_configuration();
1095         config->mem_config->nchannel = internal_config.force_nchannel;
1096         config->mem_config->nrank = internal_config.force_nrank;
1097
1098         return 0;
1099 }
1100
1101
1102 /* init memory subsystem */
1103 int
1104 rte_eal_memory_init(void)
1105 {
1106         RTE_LOG(INFO, EAL, "Setting up hugepage memory...\n");
1107         const int retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
1108                         rte_eal_hugepage_init() :
1109                         rte_eal_hugepage_attach();
1110         if (retval < 0)
1111                 return -1;
1112
1113         if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)
1114                 return -1;
1115
1116         return 0;
1117 }