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 static uint64_t baseaddr_offset;
 113
 114 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
 115
 116 /*
 117  * Check whether address-space layout randomization is enabled in
 118  * the kernel. This is important for multi-process as it can prevent
 119  * two processes mapping data to the same virtual address
 120  * Returns:
 121  *    0 - address space randomization disabled
 122  *    1/2 - address space randomization enabled
 123  *    negative error code on error
 124  */
 125 static int
 126 aslr_enabled(void)
 127 {
 128         char c;
 129         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 130         if (fd < 0)
 131                 return -errno;
 132         retval = read(fd, &c, 1);
 133         close(fd);
 134         if (retval < 0)
 135                 return -errno;
 136         if (retval == 0)
 137                 return -EIO;
 138         switch (c) {
 139                 case '0' : return 0;
 140                 case '1' : return 1;
 141                 case '2' : return 2;
 142                 default: return -EINVAL;
 143         }
 144 }
 145
 146 /*
 147  * Try to mmap *size bytes in /dev/zero. If it is succesful, return the
 148  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 149  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 150  * 0. In this case, return NULL. Note: this function returns an address
 151  * which is a multiple of hugepage size.
 152  */
 153 static void *
 154 get_virtual_area(size_t *size, size_t hugepage_sz)
 155 {
 156         void *addr;
 157         int fd;
 158         long aligned_addr;
 159
 160         if (internal_config.base_virtaddr != 0) {
 161                 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
 162                                 baseaddr_offset);
 163         }
 164         else addr = NULL;
 165
 166         RTE_LOG(INFO, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
 167
 168         fd = open("/dev/zero", O_RDONLY);
 169         if (fd < 0){
 170                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 171                 return NULL;
 172         }
 173         do {
 174                 addr = mmap(addr,
 175                                 (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
 176                 if (addr == MAP_FAILED)
 177                         *size -= hugepage_sz;
 178         } while (addr == MAP_FAILED && *size > 0);
 179
 180         if (addr == MAP_FAILED) {
 181                 close(fd);
 182                 RTE_LOG(INFO, EAL, "Cannot get a virtual area\n");
 183                 return NULL;
 184         }
 185
 186         munmap(addr, (*size) + hugepage_sz);
 187         close(fd);
 188
 189         /* align addr to a huge page size boundary */
 190         aligned_addr = (long)addr;
 191         aligned_addr += (hugepage_sz - 1);
 192         aligned_addr &= (~(hugepage_sz - 1));
 193         addr = (void *)(aligned_addr);
 194
 195         RTE_LOG(INFO, EAL, "Virtual area found at %p (size = 0x%zx)\n",
 196                 addr, *size);
 197
 198         /* increment offset */
 199         baseaddr_offset += *size;
 200
 201         return addr;
 202 }
 203
 204 /*
 205  * Mmap all hugepages of hugepage table: it first open a file in
 206  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 207  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 208  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 209  * map continguous physical blocks in contiguous virtual blocks.
 210  */
 211 static int
 212 map_all_hugepages(struct hugepage *hugepg_tbl,
 213                 struct hugepage_info *hpi, int orig)
 214 {
 215         int fd;
 216         unsigned i;
 217         void *virtaddr;
 218         void *vma_addr = NULL;
 219         size_t vma_len = 0;
 220
 221         for (i = 0; i < hpi->num_pages[0]; i++) {
 222                 size_t hugepage_sz = hpi->hugepage_sz;
 223
 224                 if (orig) {
 225                         hugepg_tbl[i].file_id = i;
 226                         hugepg_tbl[i].size = hugepage_sz;
 227                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 228                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 229                                         hugepg_tbl[i].file_id);
 230                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 231                 }
 232 #ifndef RTE_ARCH_X86_64
 233                 /* for 32-bit systems, don't remap 1G pages, just reuse original
 234                  * map address as final map address.
 235                  */
 236                 else if (hugepage_sz == RTE_PGSIZE_1G){
 237                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 238                         hugepg_tbl[i].orig_va = NULL;
 239                         continue;
 240                 }
 241 #endif
 242                 else if (vma_len == 0) {
 243                         unsigned j, num_pages;
 244
 245                         /* reserve a virtual area for next contiguous
 246                          * physical block: count the number of
 247                          * contiguous physical pages. */
 248                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 249                                 if (hugepg_tbl[j].physaddr !=
 250                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 251                                         break;
 252                         }
 253                         num_pages = j - i;
 254                         vma_len = num_pages * hugepage_sz;
 255
 256                         /* get the biggest virtual memory area up to
 257                          * vma_len. If it fails, vma_addr is NULL, so
 258                          * let the kernel provide the address. */
 259                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 260                         if (vma_addr == NULL)
 261                                 vma_len = hugepage_sz;
 262                 }
 263
 264                 /* try to create hugepage file */
 265                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
 266                 if (fd < 0) {
 267                         RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
 268                                         strerror(errno));
 269                         return -1;
 270                 }
 271
 272                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 273                                 MAP_SHARED, fd, 0);
 274                 if (virtaddr == MAP_FAILED) {
 275                         RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__,
 276                                         strerror(errno));
 277                         close(fd);
 278                         return -1;
 279                 }
 280
 281                 if (orig) {
 282                         hugepg_tbl[i].orig_va = virtaddr;
 283                         memset(virtaddr, 0, hugepage_sz);
 284                 }
 285                 else {
 286                         hugepg_tbl[i].final_va = virtaddr;
 287                 }
 288
 289                 /* set shared flock on the file. */
 290                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 291                         RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
 292                                 __func__, strerror(errno));
 293                         close(fd);
 294                         return -1;
 295                 }
 296
 297                 close(fd);
 298
 299                 vma_addr = (char *)vma_addr + hugepage_sz;
 300                 vma_len -= hugepage_sz;
 301         }
 302         return 0;
 303 }
 304
 305 /* Unmap all hugepages from original mapping. */
 306 static int
 307 unmap_all_hugepages_orig(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 308 {
 309         unsigned i;
 310         for (i = 0; i < hpi->num_pages[0]; i++) {
 311                 if (hugepg_tbl[i].orig_va) {
 312                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 313                         hugepg_tbl[i].orig_va = NULL;
 314                 }
 315         }
 316         return 0;
 317 }
 318
 319 /*
 320  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 321  * it by browsing the /proc/self/pagemap special file.
 322  */
 323 static int
 324 find_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 325 {
 326         int fd;
 327         unsigned i;
 328         uint64_t page;
 329         unsigned long virt_pfn;
 330         int page_size;
 331
 332         /* standard page size */
 333         page_size = getpagesize();
 334
 335         fd = open("/proc/self/pagemap", O_RDONLY);
 336         if (fd < 0) {
 337                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 338                         __func__, strerror(errno));
 339                 return -1;
 340         }
 341
 342         for (i = 0; i < hpi->num_pages[0]; i++) {
 343                 off_t offset;
 344                 virt_pfn = (unsigned long)hugepg_tbl[i].orig_va /
 345                         page_size;
 346                 offset = sizeof(uint64_t) * virt_pfn;
 347                 if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 348                         RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 349                                         __func__, strerror(errno));
 350                         close(fd);
 351                         return -1;
 352                 }
 353                 if (read(fd, &page, sizeof(uint64_t)) < 0) {
 354                         RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 355                                         __func__, strerror(errno));
 356                         close(fd);
 357                         return -1;
 358                 }
 359
 360                 /*
 361                  * the pfn (page frame number) are bits 0-54 (see
 362                  * pagemap.txt in linux Documentation)
 363                  */
 364                 hugepg_tbl[i].physaddr = ((page & 0x7fffffffffffffULL) * page_size);
 365         }
 366         close(fd);
 367         return 0;
 368 }
 369
 370 /*
 371  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 372  * page.
 373  */
 374 static int
 375 find_numasocket(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 376 {
 377         int socket_id;
 378         char *end, *nodestr;
 379         unsigned i, hp_count = 0;
 380         uint64_t virt_addr;
 381         char buf[BUFSIZ];
 382         char hugedir_str[PATH_MAX];
 383         FILE *f;
 384
 385         f = fopen("/proc/self/numa_maps", "r");
 386         if (f == NULL) {
 387                 RTE_LOG(INFO, EAL, "cannot open /proc/self/numa_maps,"
 388                                 " consider that all memory is in socket_id 0\n");
 389                 return 0;
 390         }
 391
 392         rte_snprintf(hugedir_str, sizeof(hugedir_str),
 393                         "%s/", hpi->hugedir);
 394
 395         /* parse numa map */
 396         while (fgets(buf, sizeof(buf), f) != NULL) {
 397
 398                 /* ignore non huge page */
 399                 if (strstr(buf, " huge ") == NULL &&
 400                                 strstr(buf, hugedir_str) == NULL)
 401                         continue;
 402
 403                 /* get zone addr */
 404                 virt_addr = strtoull(buf, &end, 16);
 405                 if (virt_addr == 0 || end == buf) {
 406                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 407                         goto error;
 408                 }
 409
 410                 /* get node id (socket id) */
 411                 nodestr = strstr(buf, " N");
 412                 if (nodestr == NULL) {
 413                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 414                         goto error;
 415                 }
 416                 nodestr += 2;
 417                 end = strstr(nodestr, "=");
 418                 if (end == NULL) {
 419                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 420                         goto error;
 421                 }
 422                 end[0] = '\0';
 423                 end = NULL;
 424
 425                 socket_id = strtoul(nodestr, &end, 0);
 426                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 427                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 428                         goto error;
 429                 }
 430
 431                 /* if we find this page in our mappings, set socket_id */
 432                 for (i = 0; i < hpi->num_pages[0]; i++) {
 433                         void *va = (void *)(unsigned long)virt_addr;
 434                         if (hugepg_tbl[i].orig_va == va) {
 435                                 hugepg_tbl[i].socket_id = socket_id;
 436                                 hp_count++;
 437                         }
 438                 }
 439         }
 440
 441         if (hp_count < hpi->num_pages[0])
 442                 goto error;
 443
 444         fclose(f);
 445         return 0;
 446
 447 error:
 448         fclose(f);
 449         return -1;
 450 }
 451
 452 /*
 453  * Sort the hugepg_tbl by physical address (lower addresses first). We
 454  * use a slow algorithm, but we won't have millions of pages, and this
 455  * is only done at init time.
 456  */
 457 static int
 458 sort_by_physaddr(struct hugepage *hugepg_tbl, struct hugepage_info *hpi)
 459 {
 460         unsigned i, j;
 461         int smallest_idx;
 462         uint64_t smallest_addr;
 463         struct hugepage tmp;
 464
 465         for (i = 0; i < hpi->num_pages[0]; i++) {
 466                 smallest_addr = 0;
 467                 smallest_idx = -1;
 468
 469                 /*
 470                  * browse all entries starting at 'i', and find the
 471                  * entry with the smallest addr
 472                  */
 473                 for (j=i; j< hpi->num_pages[0]; j++) {
 474
 475                         if (smallest_addr == 0 ||
 476                             hugepg_tbl[j].physaddr < smallest_addr) {
 477                                 smallest_addr = hugepg_tbl[j].physaddr;
 478                                 smallest_idx = j;
 479                         }
 480                 }
 481
 482                 /* should not happen */
 483                 if (smallest_idx == -1) {
 484                         RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
 485                         return -1;
 486                 }
 487
 488                 /* swap the 2 entries in the table */
 489                 memcpy(&tmp, &hugepg_tbl[smallest_idx], sizeof(struct hugepage));
 490                 memcpy(&hugepg_tbl[smallest_idx], &hugepg_tbl[i],
 491                                 sizeof(struct hugepage));
 492                 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage));
 493         }
 494         return 0;
 495 }
 496
 497 /*
 498  * Uses mmap to create a shared memory area for storage of data
 499  * Used in this file to store the hugepage file map on disk
 500  */
 501 static void *
 502 create_shared_memory(const char *filename, const size_t mem_size)
 503 {
 504         void *retval;
 505         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 506         if (fd < 0)
 507                 return NULL;
 508         if (ftruncate(fd, mem_size) < 0) {
 509                 close(fd);
 510                 return NULL;
 511         }
 512         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 513         close(fd);
 514         return retval;
 515 }
 516
 517 /*
 518  * this copies *active* hugepages from one hugepage table to another.
 519  * destination is typically the shared memory.
 520  */
 521 static int
 522 copy_hugepages_to_shared_mem(struct hugepage * dst, int dest_size,
 523                 const struct hugepage * src, int src_size)
 524 {
 525         int src_pos, dst_pos = 0;
 526
 527         for (src_pos = 0; src_pos < src_size; src_pos++) {
 528                 if (src[src_pos].final_va != NULL) {
 529                         /* error on overflow attempt */
 530                         if (dst_pos == dest_size)
 531                                 return -1;
 532                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage));
 533                         dst_pos++;
 534                 }
 535         }
 536         return 0;
 537 }
 538
 539 /*
 540  * unmaps hugepages that are not going to be used. since we originally allocate
 541  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 542  */
 543 static int
 544 unmap_unneeded_hugepages(struct hugepage *hugepg_tbl,
 545                 struct hugepage_info *hpi,
 546                 unsigned num_hp_info)
 547 {
 548         unsigned socket, size;
 549         int page, nrpages = 0;
 550
 551         /* get total number of hugepages */
 552         for (size = 0; size < num_hp_info; size++)
 553                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 554                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 555
 556         for (size = 0; size < num_hp_info; size++) {
 557                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 558                         unsigned pages_found = 0;
 559                         /* traverse until we have unmapped all the unused pages */
 560                         for (page = 0; page < nrpages; page++) {
 561                                 struct hugepage *hp = &hugepg_tbl[page];
 562
 563                                 /* find a page that matches the criteria */
 564                                 if ((hp->size == hpi[size].hugepage_sz) &&
 565                                                 (hp->socket_id == (int) socket)) {
 566
 567                                         /* if we skipped enough pages, unmap the rest */
 568                                         if (pages_found == hpi[size].num_pages[socket]) {
 569                                                 munmap(hp->final_va, hp->size);
 570                                                 hp->final_va = NULL;
 571                                                 if (remove(hp->filepath) == -1) {
 572                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 573                                                                         __func__, hp->filepath, strerror(errno));
 574                                                         return -1;
 575                                                 }
 576                                         }
 577                                         /* lock the page and skip */
 578                                         else
 579                                                 pages_found++;
 580
 581                                 } /* match page */
 582                         } /* foreach page */
 583                 } /* foreach socket */
 584         } /* foreach pagesize */
 585
 586         return 0;
 587 }
 588
 589 static inline uint64_t
 590 get_socket_mem_size(int socket)
 591 {
 592         uint64_t size = 0;
 593         unsigned i;
 594
 595         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 596                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 597                 if (hpi->hugedir != NULL)
 598                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 599         }
 600
 601         return (size);
 602 }
 603
 604 /*
 605  * This function is a NUMA-aware equivalent of calc_num_pages.
 606  * It takes in the list of hugepage sizes and the
 607  * number of pages thereof, and calculates the best number of
 608  * pages of each size to fulfill the request for <memory> ram
 609  */
 610 static int
 611 calc_num_pages_per_socket(uint64_t * memory,
 612                 struct hugepage_info *hp_info,
 613                 struct hugepage_info *hp_used,
 614                 unsigned num_hp_info)
 615 {
 616         unsigned socket, j, i = 0;
 617         unsigned requested, available;
 618         int total_num_pages = 0;
 619         uint64_t remaining_mem, cur_mem;
 620         uint64_t total_mem = internal_config.memory;
 621
 622         if (num_hp_info == 0)
 623                 return -1;
 624
 625         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 626                 /* if specific memory amounts per socket weren't requested */
 627                 if (internal_config.force_sockets == 0) {
 628                         /* take whatever is available */
 629                         memory[socket] = RTE_MIN(get_socket_mem_size(socket),
 630                                         total_mem);
 631                 }
 632                 /* skips if the memory on specific socket wasn't requested */
 633                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 634                         hp_used[i].hugedir = hp_info[i].hugedir;
 635                         hp_used[i].num_pages[socket] = RTE_MIN(
 636                                         memory[socket] / hp_info[i].hugepage_sz,
 637                                         hp_info[i].num_pages[socket]);
 638
 639                         cur_mem = hp_used[i].num_pages[socket] *
 640                                         hp_used[i].hugepage_sz;
 641
 642                         memory[socket] -= cur_mem;
 643                         total_mem -= cur_mem;
 644
 645                         total_num_pages += hp_used[i].num_pages[socket];
 646
 647                         /* check if we have met all memory requests */
 648                         if (memory[socket] == 0)
 649                                 break;
 650
 651                         /* check if we have any more pages left at this size, if so
 652                          * move on to next size */
 653                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
 654                                 continue;
 655                         /* At this point we know that there are more pages available that are
 656                          * bigger than the memory we want, so lets see if we can get enough
 657                          * from other page sizes.
 658                          */
 659                         remaining_mem = 0;
 660                         for (j = i+1; j < num_hp_info; j++)
 661                                 remaining_mem += hp_info[j].hugepage_sz *
 662                                 hp_info[j].num_pages[socket];
 663
 664                         /* is there enough other memory, if not allocate another page and quit */
 665                         if (remaining_mem < memory[socket]){
 666                                 cur_mem = RTE_MIN(memory[socket],
 667                                                 hp_info[i].hugepage_sz);
 668                                 memory[socket] -= cur_mem;
 669                                 total_mem -= cur_mem;
 670                                 hp_used[i].num_pages[socket]++;
 671                                 total_num_pages++;
 672                                 break; /* we are done with this socket*/
 673                         }
 674                 }
 675                 /* if we didn't satisfy all memory requirements per socket */
 676                 if (memory[socket] > 0) {
 677                         /* to prevent icc errors */
 678                         requested = (unsigned) (internal_config.socket_mem[socket] /
 679                                         0x100000);
 680                         available = requested -
 681                                         ((unsigned) (memory[socket] / 0x100000));
 682                         RTE_LOG(INFO, EAL, "Not enough memory available on socket %u! "
 683                                         "Requested: %uMB, available: %uMB\n", socket,
 684                                         requested, available);
 685                         return -1;
 686                 }
 687         }
 688
 689         /* if we didn't satisfy total memory requirements */
 690         if (total_mem > 0) {
 691                 requested = (unsigned) (internal_config.memory / 0x100000);
 692                 available = requested - (unsigned) (total_mem / 0x100000);
 693                 RTE_LOG(INFO, EAL, "Not enough memory available! Requested: %uMB,"
 694                                 " available: %uMB\n", requested, available);
 695                 return -1;
 696         }
 697         return total_num_pages;
 698 }
 699
 700 /*
 701  * Prepare physical memory mapping: fill configuration structure with
 702  * these infos, return 0 on success.
 703  *  1. map N huge pages in separate files in hugetlbfs
 704  *  2. find associated physical addr
 705  *  3. find associated NUMA socket ID
 706  *  4. sort all huge pages by physical address
 707  *  5. remap these N huge pages in the correct order
 708  *  6. unmap the first mapping
 709  *  7. fill memsegs in configuration with contiguous zones
 710  */
 711 static int
 712 rte_eal_hugepage_init(void)
 713 {
 714         struct rte_mem_config *mcfg;
 715         struct hugepage *hugepage, *tmp_hp = NULL;
 716         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
 717
 718         uint64_t memory[RTE_MAX_NUMA_NODES];
 719
 720         unsigned hp_offset;
 721         int i, j, new_memseg;
 722         int nrpages, total_pages = 0;
 723         void *addr;
 724
 725         memset(used_hp, 0, sizeof(used_hp));
 726
 727         /* get pointer to global configuration */
 728         mcfg = rte_eal_get_configuration()->mem_config;
 729
 730         /* for debug purposes, hugetlbfs can be disabled */
 731         if (internal_config.no_hugetlbfs) {
 732                 addr = malloc(internal_config.memory);
 733                 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
 734                 mcfg->memseg[0].addr = addr;
 735                 mcfg->memseg[0].len = internal_config.memory;
 736                 mcfg->memseg[0].socket_id = 0;
 737                 return 0;
 738         }
 739
 740
 741         /* calculate total number of hugepages available. at this point we haven't
 742          * yet started sorting them so they all are on socket 0 */
 743         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
 744                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
 745                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
 746
 747                 total_pages += internal_config.hugepage_info[i].num_pages[0];
 748         }
 749
 750         /*
 751          * allocate a memory area for hugepage table.
 752          * this isn't shared memory yet. due to the fact that we need some
 753          * processing done on these pages, shared memory will be created
 754          * at a later stage.
 755          */
 756         tmp_hp = malloc(total_pages * sizeof(struct hugepage));
 757         if (tmp_hp == NULL)
 758                 goto fail;
 759
 760         memset(tmp_hp, 0, total_pages * sizeof(struct hugepage));
 761
 762         hp_offset = 0; /* where we start the current page size entries */
 763
 764         /* map all hugepages and sort them */
 765         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
 766                 struct hugepage_info *hpi;
 767
 768                 /*
 769                  * we don't yet mark hugepages as used at this stage, so
 770                  * we just map all hugepages available to the system
 771                  * all hugepages are still located on socket 0
 772                  */
 773                 hpi = &internal_config.hugepage_info[i];
 774
 775                 if (hpi->num_pages == 0)
 776                         continue;
 777
 778                 /* map all hugepages available */
 779                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 1) < 0){
 780                         RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
 781                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 782                         goto fail;
 783                 }
 784
 785                 /* find physical addresses and sockets for each hugepage */
 786                 if (find_physaddr(&tmp_hp[hp_offset], hpi) < 0){
 787                         RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
 788                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 789                         goto fail;
 790                 }
 791
 792                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
 793                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
 794                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 795                         goto fail;
 796                 }
 797
 798                 if (sort_by_physaddr(&tmp_hp[hp_offset], hpi) < 0)
 799                         goto fail;
 800
 801                 /* remap all hugepages */
 802                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) < 0){
 803                         RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
 804                                         (unsigned)(hpi->hugepage_sz / 0x100000));
 805                         goto fail;
 806                 }
 807
 808                 /* unmap original mappings */
 809                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
 810                         goto fail;
 811
 812                 /* we have processed a num of hugepages of this size, so inc offset */
 813                 hp_offset += hpi->num_pages[0];
 814         }
 815
 816         /* clean out the numbers of pages */
 817         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
 818                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
 819                         internal_config.hugepage_info[i].num_pages[j] = 0;
 820
 821         /* get hugepages for each socket */
 822         for (i = 0; i < total_pages; i++) {
 823                 int socket = tmp_hp[i].socket_id;
 824
 825                 /* find a hugepage info with right size and increment num_pages */
 826                 for (j = 0; j < (int) internal_config.num_hugepage_sizes; j++) {
 827                         if (tmp_hp[i].size ==
 828                                         internal_config.hugepage_info[j].hugepage_sz) {
 829                                 internal_config.hugepage_info[j].num_pages[socket]++;
 830                         }
 831                 }
 832         }
 833
 834         /* make a copy of socket_mem, needed for number of pages calculation */
 835         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 836                 memory[i] = internal_config.socket_mem[i];
 837
 838         /* calculate final number of pages */
 839         nrpages = calc_num_pages_per_socket(memory,
 840                         internal_config.hugepage_info, used_hp,
 841                         internal_config.num_hugepage_sizes);
 842
 843         /* error if not enough memory available */
 844         if (nrpages < 0)
 845                 goto fail;
 846
 847         /* reporting in! */
 848         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
 849                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 850                         if (used_hp[i].num_pages[j] > 0) {
 851                                 RTE_LOG(INFO, EAL,
 852                                                 "Requesting %u pages of size %uMB"
 853                                                 " from socket %i\n",
 854                                                 used_hp[i].num_pages[j],
 855                                                 (unsigned)
 856                                                         (used_hp[i].hugepage_sz / 0x100000),
 857                                                 j);
 858                         }
 859                 }
 860         }
 861
 862         /* create shared memory */
 863         hugepage = create_shared_memory(eal_hugepage_info_path(),
 864                                         nrpages * sizeof(struct hugepage));
 865
 866         if (hugepage == NULL) {
 867                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
 868                 goto fail;
 869         }
 870
 871         /*
 872          * unmap pages that we won't need (looks at used_hp).
 873          * also, sets final_va to NULL on pages that were unmapped.
 874          */
 875         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
 876                         internal_config.num_hugepage_sizes) < 0) {
 877                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
 878                 goto fail;
 879         }
 880
 881         /*
 882          * copy stuff from malloc'd hugepage* to the actual shared memory.
 883          * this procedure only copies those hugepages that have final_va
 884          * not NULL. has overflow protection.
 885          */
 886         if (copy_hugepages_to_shared_mem(hugepage, nrpages,
 887                         tmp_hp, total_pages) < 0) {
 888                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
 889                 goto fail;
 890         }
 891
 892         /* free the temporary hugepage table */
 893         free(tmp_hp);
 894         tmp_hp = NULL;
 895
 896         memset(mcfg->memseg, 0, sizeof(mcfg->memseg));
 897         j = -1;
 898         for (i = 0; i < nrpages; i++) {
 899                 new_memseg = 0;
 900
 901                 /* if this is a new section, create a new memseg */
 902                 if (i == 0)
 903                         new_memseg = 1;
 904                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
 905                         new_memseg = 1;
 906                 else if (hugepage[i].size != hugepage[i-1].size)
 907                         new_memseg = 1;
 908                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
 909                     hugepage[i].size)
 910                         new_memseg = 1;
 911                 else if (((unsigned long)hugepage[i].final_va -
 912                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
 913                         new_memseg = 1;
 914
 915                 if (new_memseg) {
 916                         j += 1;
 917                         if (j == RTE_MAX_MEMSEG)
 918                                 break;
 919
 920                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
 921                         mcfg->memseg[j].addr = hugepage[i].final_va;
 922                         mcfg->memseg[j].len = hugepage[i].size;
 923                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
 924                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
 925                 }
 926                 /* continuation of previous memseg */
 927                 else {
 928                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
 929                 }
 930                 hugepage[i].memseg_id = j;
 931         }
 932
 933         if (i < nrpages) {
 934                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
 935                         "from %d requested\n"
 936                         "Current %s=%d is not enough\n"
 937                         "Please either increase it or request less amount "
 938                         "of memory.\n",
 939                         i, nrpages, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
 940                         RTE_MAX_MEMSEG);
 941                 return (-ENOMEM);
 942         }
 943
 944
 945         return 0;
 946
 947
 948 fail:
 949         if (tmp_hp)
 950                 free(tmp_hp);
 951         return -1;
 952 }
 953
 954 /*
 955  * uses fstat to report the size of a file on disk
 956  */
 957 static off_t
 958 getFileSize(int fd)
 959 {
 960         struct stat st;
 961         if (fstat(fd, &st) < 0)
 962                 return 0;
 963         return st.st_size;
 964 }
 965
 966 /*
 967  * This creates the memory mappings in the secondary process to match that of
 968  * the server process. It goes through each memory segment in the DPDK runtime
 969  * configuration and finds the hugepages which form that segment, mapping them
 970  * in order to form a contiguous block in the virtual memory space
 971  */
 972 static int
 973 rte_eal_hugepage_attach(void)
 974 {
 975         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 976         const struct hugepage *hp = NULL;
 977         unsigned num_hp = 0;
 978         unsigned i, s = 0; /* s used to track the segment number */
 979         off_t size;
 980         int fd, fd_zero = -1, fd_hugepage = -1;
 981
 982         if (aslr_enabled() > 0) {
 983                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
 984                                 "(ASLR) is enabled in the kernel.\n");
 985                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
 986                                 "into secondary processes\n");
 987         }
 988
 989         fd_zero = open("/dev/zero", O_RDONLY);
 990         if (fd_zero < 0) {
 991                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
 992                 goto error;
 993         }
 994         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
 995         if (fd_hugepage < 0) {
 996                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
 997                 goto error;
 998         }
 999
1000         /* map all segments into memory to make sure we get the addrs */
1001         for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1002                 void *base_addr;
1003
1004                 /*
1005                  * the first memory segment with len==0 is the one that
1006                  * follows the last valid segment.
1007                  */
1008                 if (mcfg->memseg[s].len == 0)
1009                         break;
1010
1011                 /*
1012                  * fdzero is mmapped to get a contiguous block of virtual
1013                  * addresses of the appropriate memseg size.
1014                  * use mmap to get identical addresses as the primary process.
1015                  */
1016                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1017                                  PROT_READ, MAP_PRIVATE, fd_zero, 0);
1018                 if (base_addr == MAP_FAILED ||
1019                     base_addr != mcfg->memseg[s].addr) {
1020                         RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1021                                 "in /dev/zero to requested address [%p]\n",
1022                                 (unsigned long long)mcfg->memseg[s].len,
1023                                 mcfg->memseg[s].addr);
1024                         if (aslr_enabled() > 0) {
1025                                 RTE_LOG(ERR, EAL, "It is recommended to "
1026                                         "disable ASLR in the kernel "
1027                                         "and retry running both primary "
1028                                         "and secondary processes\n");
1029                         }
1030                         goto error;
1031                 }
1032         }
1033
1034         size = getFileSize(fd_hugepage);
1035         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1036         if (hp == NULL) {
1037                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1038                 goto error;
1039         }
1040
1041         num_hp = size / sizeof(struct hugepage);
1042         RTE_LOG(DEBUG, EAL, "Analysing %u hugepages\n", num_hp);
1043
1044         s = 0;
1045         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1046                 void *addr, *base_addr;
1047                 uintptr_t offset = 0;
1048
1049                 /*
1050                  * free previously mapped memory so we can map the
1051                  * hugepages into the space
1052                  */
1053                 base_addr = mcfg->memseg[s].addr;
1054                 munmap(base_addr, mcfg->memseg[s].len);
1055
1056                 /* find the hugepages for this segment and map them
1057                  * we don't need to worry about order, as the server sorted the
1058                  * entries before it did the second mmap of them */
1059                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1060                         if (hp[i].memseg_id == (int)s){
1061                                 fd = open(hp[i].filepath, O_RDWR);
1062                                 if (fd < 0) {
1063                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1064                                                 hp[i].filepath);
1065                                         goto error;
1066                                 }
1067                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1068                                                 hp[i].size, PROT_READ | PROT_WRITE,
1069                                                 MAP_SHARED | MAP_FIXED, fd, 0);
1070                                 close(fd); /* close file both on success and on failure */
1071                                 if (addr == MAP_FAILED) {
1072                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1073                                                 hp[i].filepath);
1074                                         goto error;
1075                                 }
1076                                 offset+=hp[i].size;
1077                         }
1078                 }
1079                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1080                                 (unsigned long long)mcfg->memseg[s].len);
1081                 s++;
1082         }
1083         /* unmap the hugepage config file, since we are done using it */
1084         munmap((void *)(uintptr_t)hp, size);
1085         close(fd_zero);
1086         close(fd_hugepage);
1087         return 0;
1088
1089 error:
1090         if (fd_zero >= 0)
1091                 close(fd_zero);
1092         if (fd_hugepage >= 0)
1093                 close(fd_hugepage);
1094         return -1;
1095 }
1096
1097 static int
1098 rte_eal_memdevice_init(void)
1099 {
1100         struct rte_config *config;
1101
1102         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1103                 return 0;
1104
1105         config = rte_eal_get_configuration();
1106         config->mem_config->nchannel = internal_config.force_nchannel;
1107         config->mem_config->nrank = internal_config.force_nrank;
1108
1109         return 0;
1110 }
1111
1112
1113 /* init memory subsystem */
1114 int
1115 rte_eal_memory_init(void)
1116 {
1117         RTE_LOG(INFO, EAL, "Setting up hugepage memory...\n");
1118         const int retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
1119                         rte_eal_hugepage_init() :
1120                         rte_eal_hugepage_attach();
1121         if (retval < 0)
1122                 return -1;
1123
1124         if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)
1125                 return -1;
1126
1127         return 0;
1128 }