lib/librte_eal/linuxapp/eal/eal_memory.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2014 Intel Corporation.
   3  * Copyright(c) 2013 6WIND S.A.
   4  */
   5
   6 #define _FILE_OFFSET_BITS 64
   7 #include <errno.h>
   8 #include <stdarg.h>
   9 #include <stdbool.h>
  10 #include <stdlib.h>
  11 #include <stdio.h>
  12 #include <stdint.h>
  13 #include <inttypes.h>
  14 #include <string.h>
  15 #include <sys/mman.h>
  16 #include <sys/types.h>
  17 #include <sys/stat.h>
  18 #include <sys/queue.h>
  19 #include <sys/file.h>
  20 #include <unistd.h>
  21 #include <limits.h>
  22 #include <sys/ioctl.h>
  23 #include <sys/time.h>
  24 #include <signal.h>
  25 #include <setjmp.h>
  26 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
  27 #include <numa.h>
  28 #include <numaif.h>
  29 #endif
  30
  31 #include <rte_log.h>
  32 #include <rte_memory.h>
  33 #include <rte_launch.h>
  34 #include <rte_eal.h>
  35 #include <rte_eal_memconfig.h>
  36 #include <rte_per_lcore.h>
  37 #include <rte_lcore.h>
  38 #include <rte_common.h>
  39 #include <rte_string_fns.h>
  40
  41 #include "eal_private.h"
  42 #include "eal_internal_cfg.h"
  43 #include "eal_filesystem.h"
  44 #include "eal_hugepages.h"
  45
  46 #define PFN_MASK_SIZE   8
  47
  48 /**
  49  * @file
  50  * Huge page mapping under linux
  51  *
  52  * To reserve a big contiguous amount of memory, we use the hugepage
  53  * feature of linux. For that, we need to have hugetlbfs mounted. This
  54  * code will create many files in this directory (one per page) and
  55  * map them in virtual memory. For each page, we will retrieve its
  56  * physical address and remap it in order to have a virtual contiguous
  57  * zone as well as a physical contiguous zone.
  58  */
  59
  60 static uint64_t baseaddr_offset;
  61
  62 static bool phys_addrs_available = true;
  63
  64 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
  65
  66 static void
  67 test_phys_addrs_available(void)
  68 {
  69         uint64_t tmp;
  70         phys_addr_t physaddr;
  71
  72         if (!rte_eal_has_hugepages()) {
  73                 RTE_LOG(ERR, EAL,
  74                         "Started without hugepages support, physical addresses not available\n");
  75                 phys_addrs_available = false;
  76                 return;
  77         }
  78
  79         physaddr = rte_mem_virt2phy(&tmp);
  80         if (physaddr == RTE_BAD_PHYS_ADDR) {
  81                 if (rte_eal_iova_mode() == RTE_IOVA_PA)
  82                         RTE_LOG(ERR, EAL,
  83                                 "Cannot obtain physical addresses: %s. "
  84                                 "Only vfio will function.\n",
  85                                 strerror(errno));
  86                 phys_addrs_available = false;
  87         }
  88 }
  89
  90 /*
  91  * Get physical address of any mapped virtual address in the current process.
  92  */
  93 phys_addr_t
  94 rte_mem_virt2phy(const void *virtaddr)
  95 {
  96         int fd, retval;
  97         uint64_t page, physaddr;
  98         unsigned long virt_pfn;
  99         int page_size;
 100         off_t offset;
 101
 102         /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
 103         if (!phys_addrs_available)
 104                 return RTE_BAD_IOVA;
 105
 106         /* standard page size */
 107         page_size = getpagesize();
 108
 109         fd = open("/proc/self/pagemap", O_RDONLY);
 110         if (fd < 0) {
 111                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 112                         __func__, strerror(errno));
 113                 return RTE_BAD_IOVA;
 114         }
 115
 116         virt_pfn = (unsigned long)virtaddr / page_size;
 117         offset = sizeof(uint64_t) * virt_pfn;
 118         if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 119                 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 120                                 __func__, strerror(errno));
 121                 close(fd);
 122                 return RTE_BAD_IOVA;
 123         }
 124
 125         retval = read(fd, &page, PFN_MASK_SIZE);
 126         close(fd);
 127         if (retval < 0) {
 128                 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 129                                 __func__, strerror(errno));
 130                 return RTE_BAD_IOVA;
 131         } else if (retval != PFN_MASK_SIZE) {
 132                 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
 133                                 "but expected %d:\n",
 134                                 __func__, retval, PFN_MASK_SIZE);
 135                 return RTE_BAD_IOVA;
 136         }
 137
 138         /*
 139          * the pfn (page frame number) are bits 0-54 (see
 140          * pagemap.txt in linux Documentation)
 141          */
 142         if ((page & 0x7fffffffffffffULL) == 0)
 143                 return RTE_BAD_IOVA;
 144
 145         physaddr = ((page & 0x7fffffffffffffULL) * page_size)
 146                 + ((unsigned long)virtaddr % page_size);
 147
 148         return physaddr;
 149 }
 150
 151 rte_iova_t
 152 rte_mem_virt2iova(const void *virtaddr)
 153 {
 154         if (rte_eal_iova_mode() == RTE_IOVA_VA)
 155                 return (uintptr_t)virtaddr;
 156         return rte_mem_virt2phy(virtaddr);
 157 }
 158
 159 /*
 160  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 161  * it by browsing the /proc/self/pagemap special file.
 162  */
 163 static int
 164 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 165 {
 166         unsigned int i;
 167         phys_addr_t addr;
 168
 169         for (i = 0; i < hpi->num_pages[0]; i++) {
 170                 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
 171                 if (addr == RTE_BAD_PHYS_ADDR)
 172                         return -1;
 173                 hugepg_tbl[i].physaddr = addr;
 174         }
 175         return 0;
 176 }
 177
 178 /*
 179  * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
 180  */
 181 static int
 182 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 183 {
 184         unsigned int i;
 185         static phys_addr_t addr;
 186
 187         for (i = 0; i < hpi->num_pages[0]; i++) {
 188                 hugepg_tbl[i].physaddr = addr;
 189                 addr += hugepg_tbl[i].size;
 190         }
 191         return 0;
 192 }
 193
 194 /*
 195  * Check whether address-space layout randomization is enabled in
 196  * the kernel. This is important for multi-process as it can prevent
 197  * two processes mapping data to the same virtual address
 198  * Returns:
 199  *    0 - address space randomization disabled
 200  *    1/2 - address space randomization enabled
 201  *    negative error code on error
 202  */
 203 static int
 204 aslr_enabled(void)
 205 {
 206         char c;
 207         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 208         if (fd < 0)
 209                 return -errno;
 210         retval = read(fd, &c, 1);
 211         close(fd);
 212         if (retval < 0)
 213                 return -errno;
 214         if (retval == 0)
 215                 return -EIO;
 216         switch (c) {
 217                 case '0' : return 0;
 218                 case '1' : return 1;
 219                 case '2' : return 2;
 220                 default: return -EINVAL;
 221         }
 222 }
 223
 224 /*
 225  * Try to mmap *size bytes in /dev/zero. If it is successful, return the
 226  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 227  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 228  * 0. In this case, return NULL. Note: this function returns an address
 229  * which is a multiple of hugepage size.
 230  */
 231 static void *
 232 get_virtual_area(size_t *size, size_t hugepage_sz)
 233 {
 234         void *addr;
 235         void *addr_hint;
 236         int fd;
 237         long aligned_addr;
 238
 239         if (internal_config.base_virtaddr != 0) {
 240                 int page_size = sysconf(_SC_PAGE_SIZE);
 241                 addr_hint = (void *) (uintptr_t)
 242                         (internal_config.base_virtaddr + baseaddr_offset);
 243                 addr_hint = RTE_PTR_ALIGN_FLOOR(addr_hint, page_size);
 244         } else {
 245                 addr_hint = NULL;
 246         }
 247
 248         RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
 249
 250
 251         fd = open("/dev/zero", O_RDONLY);
 252         if (fd < 0){
 253                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 254                 return NULL;
 255         }
 256         do {
 257                 addr = mmap(addr_hint, (*size) + hugepage_sz, PROT_READ,
 258 #ifdef RTE_ARCH_PPC_64
 259                                 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
 260 #else
 261                                 MAP_PRIVATE,
 262 #endif
 263                                 fd, 0);
 264                 if (addr == MAP_FAILED) {
 265                         *size -= hugepage_sz;
 266                 } else if (addr_hint != NULL && addr != addr_hint) {
 267                         RTE_LOG(WARNING, EAL, "WARNING! Base virtual address "
 268                                 "hint (%p != %p) not respected!\n",
 269                                 addr_hint, addr);
 270                         RTE_LOG(WARNING, EAL, "   This may cause issues with "
 271                                 "mapping memory into secondary processes\n");
 272                 }
 273         } while (addr == MAP_FAILED && *size > 0);
 274
 275         if (addr == MAP_FAILED) {
 276                 close(fd);
 277                 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
 278                         strerror(errno));
 279                 return NULL;
 280         }
 281
 282         munmap(addr, (*size) + hugepage_sz);
 283         close(fd);
 284
 285         /* align addr to a huge page size boundary */
 286         aligned_addr = (long)addr;
 287         aligned_addr += (hugepage_sz - 1);
 288         aligned_addr &= (~(hugepage_sz - 1));
 289         addr = (void *)(aligned_addr);
 290
 291         RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
 292                 addr, *size);
 293
 294         /* increment offset */
 295         baseaddr_offset += *size;
 296
 297         return addr;
 298 }
 299
 300 static sigjmp_buf huge_jmpenv;
 301
 302 static void huge_sigbus_handler(int signo __rte_unused)
 303 {
 304         siglongjmp(huge_jmpenv, 1);
 305 }
 306
 307 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
 308  * non-static local variable in the stack frame calling sigsetjmp might be
 309  * clobbered by a call to longjmp.
 310  */
 311 static int huge_wrap_sigsetjmp(void)
 312 {
 313         return sigsetjmp(huge_jmpenv, 1);
 314 }
 315
 316 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 317 /* Callback for numa library. */
 318 void numa_error(char *where)
 319 {
 320         RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
 321 }
 322 #endif
 323
 324 /*
 325  * Mmap all hugepages of hugepage table: it first open a file in
 326  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 327  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 328  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 329  * map contiguous physical blocks in contiguous virtual blocks.
 330  */
 331 static unsigned
 332 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
 333                   uint64_t *essential_memory __rte_unused, int orig)
 334 {
 335         int fd;
 336         unsigned i;
 337         void *virtaddr;
 338         void *vma_addr = NULL;
 339         size_t vma_len = 0;
 340 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 341         int node_id = -1;
 342         int essential_prev = 0;
 343         int oldpolicy;
 344         struct bitmask *oldmask = numa_allocate_nodemask();
 345         bool have_numa = true;
 346         unsigned long maxnode = 0;
 347
 348         /* Check if kernel supports NUMA. */
 349         if (numa_available() != 0) {
 350                 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
 351                 have_numa = false;
 352         }
 353
 354         if (orig && have_numa) {
 355                 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
 356                 if (get_mempolicy(&oldpolicy, oldmask->maskp,
 357                                   oldmask->size + 1, 0, 0) < 0) {
 358                         RTE_LOG(ERR, EAL,
 359                                 "Failed to get current mempolicy: %s. "
 360                                 "Assuming MPOL_DEFAULT.\n", strerror(errno));
 361                         oldpolicy = MPOL_DEFAULT;
 362                 }
 363                 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 364                         if (internal_config.socket_mem[i])
 365                                 maxnode = i + 1;
 366         }
 367 #endif
 368
 369         for (i = 0; i < hpi->num_pages[0]; i++) {
 370                 uint64_t hugepage_sz = hpi->hugepage_sz;
 371
 372 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 373                 if (maxnode) {
 374                         unsigned int j;
 375
 376                         for (j = 0; j < maxnode; j++)
 377                                 if (essential_memory[j])
 378                                         break;
 379
 380                         if (j == maxnode) {
 381                                 node_id = (node_id + 1) % maxnode;
 382                                 while (!internal_config.socket_mem[node_id]) {
 383                                         node_id++;
 384                                         node_id %= maxnode;
 385                                 }
 386                                 essential_prev = 0;
 387                         } else {
 388                                 node_id = j;
 389                                 essential_prev = essential_memory[j];
 390
 391                                 if (essential_memory[j] < hugepage_sz)
 392                                         essential_memory[j] = 0;
 393                                 else
 394                                         essential_memory[j] -= hugepage_sz;
 395                         }
 396
 397                         RTE_LOG(DEBUG, EAL,
 398                                 "Setting policy MPOL_PREFERRED for socket %d\n",
 399                                 node_id);
 400                         numa_set_preferred(node_id);
 401                 }
 402 #endif
 403
 404                 if (orig) {
 405                         hugepg_tbl[i].file_id = i;
 406                         hugepg_tbl[i].size = hugepage_sz;
 407                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 408                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 409                                         hugepg_tbl[i].file_id);
 410                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 411                 }
 412 #ifndef RTE_ARCH_64
 413                 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
 414                  * original map address as final map address.
 415                  */
 416                 else if ((hugepage_sz == RTE_PGSIZE_1G)
 417                         || (hugepage_sz == RTE_PGSIZE_16G)) {
 418                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 419                         hugepg_tbl[i].orig_va = NULL;
 420                         continue;
 421                 }
 422 #endif
 423                 else if (vma_len == 0) {
 424                         unsigned j, num_pages;
 425
 426                         /* reserve a virtual area for next contiguous
 427                          * physical block: count the number of
 428                          * contiguous physical pages. */
 429                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 430 #ifdef RTE_ARCH_PPC_64
 431                                 /* The physical addresses are sorted in
 432                                  * descending order on PPC64 */
 433                                 if (hugepg_tbl[j].physaddr !=
 434                                     hugepg_tbl[j-1].physaddr - hugepage_sz)
 435                                         break;
 436 #else
 437                                 if (hugepg_tbl[j].physaddr !=
 438                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 439                                         break;
 440 #endif
 441                         }
 442                         num_pages = j - i;
 443                         vma_len = num_pages * hugepage_sz;
 444
 445                         /* get the biggest virtual memory area up to
 446                          * vma_len. If it fails, vma_addr is NULL, so
 447                          * let the kernel provide the address. */
 448                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 449                         if (vma_addr == NULL)
 450                                 vma_len = hugepage_sz;
 451                 }
 452
 453                 /* try to create hugepage file */
 454                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
 455                 if (fd < 0) {
 456                         RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
 457                                         strerror(errno));
 458                         goto out;
 459                 }
 460
 461                 /* map the segment, and populate page tables,
 462                  * the kernel fills this segment with zeros */
 463                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 464                                 MAP_SHARED | MAP_POPULATE, fd, 0);
 465                 if (virtaddr == MAP_FAILED) {
 466                         RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
 467                                         strerror(errno));
 468                         close(fd);
 469                         goto out;
 470                 }
 471
 472                 if (orig) {
 473                         hugepg_tbl[i].orig_va = virtaddr;
 474                 }
 475                 else {
 476                         /* rewrite physical addresses in IOVA as VA mode */
 477                         if (rte_eal_iova_mode() == RTE_IOVA_VA)
 478                                 hugepg_tbl[i].physaddr = (uintptr_t)virtaddr;
 479                         hugepg_tbl[i].final_va = virtaddr;
 480                 }
 481
 482                 if (orig) {
 483                         /* In linux, hugetlb limitations, like cgroup, are
 484                          * enforced at fault time instead of mmap(), even
 485                          * with the option of MAP_POPULATE. Kernel will send
 486                          * a SIGBUS signal. To avoid to be killed, save stack
 487                          * environment here, if SIGBUS happens, we can jump
 488                          * back here.
 489                          */
 490                         if (huge_wrap_sigsetjmp()) {
 491                                 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
 492                                         "hugepages of size %u MB\n",
 493                                         (unsigned)(hugepage_sz / 0x100000));
 494                                 munmap(virtaddr, hugepage_sz);
 495                                 close(fd);
 496                                 unlink(hugepg_tbl[i].filepath);
 497 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 498                                 if (maxnode)
 499                                         essential_memory[node_id] =
 500                                                 essential_prev;
 501 #endif
 502                                 goto out;
 503                         }
 504                         *(int *)virtaddr = 0;
 505                 }
 506
 507
 508                 /* set shared flock on the file. */
 509                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 510                         RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
 511                                 __func__, strerror(errno));
 512                         close(fd);
 513                         goto out;
 514                 }
 515
 516                 close(fd);
 517
 518                 vma_addr = (char *)vma_addr + hugepage_sz;
 519                 vma_len -= hugepage_sz;
 520         }
 521
 522 out:
 523 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 524         if (maxnode) {
 525                 RTE_LOG(DEBUG, EAL,
 526                         "Restoring previous memory policy: %d\n", oldpolicy);
 527                 if (oldpolicy == MPOL_DEFAULT) {
 528                         numa_set_localalloc();
 529                 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
 530                                          oldmask->size + 1) < 0) {
 531                         RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
 532                                 strerror(errno));
 533                         numa_set_localalloc();
 534                 }
 535         }
 536         numa_free_cpumask(oldmask);
 537 #endif
 538         return i;
 539 }
 540
 541 /* Unmap all hugepages from original mapping */
 542 static int
 543 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 544 {
 545         unsigned i;
 546         for (i = 0; i < hpi->num_pages[0]; i++) {
 547                 if (hugepg_tbl[i].orig_va) {
 548                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 549                         hugepg_tbl[i].orig_va = NULL;
 550                 }
 551         }
 552         return 0;
 553 }
 554
 555 /*
 556  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 557  * page.
 558  */
 559 static int
 560 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 561 {
 562         int socket_id;
 563         char *end, *nodestr;
 564         unsigned i, hp_count = 0;
 565         uint64_t virt_addr;
 566         char buf[BUFSIZ];
 567         char hugedir_str[PATH_MAX];
 568         FILE *f;
 569
 570         f = fopen("/proc/self/numa_maps", "r");
 571         if (f == NULL) {
 572                 RTE_LOG(NOTICE, EAL, "NUMA support not available"
 573                         " consider that all memory is in socket_id 0\n");
 574                 return 0;
 575         }
 576
 577         snprintf(hugedir_str, sizeof(hugedir_str),
 578                         "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
 579
 580         /* parse numa map */
 581         while (fgets(buf, sizeof(buf), f) != NULL) {
 582
 583                 /* ignore non huge page */
 584                 if (strstr(buf, " huge ") == NULL &&
 585                                 strstr(buf, hugedir_str) == NULL)
 586                         continue;
 587
 588                 /* get zone addr */
 589                 virt_addr = strtoull(buf, &end, 16);
 590                 if (virt_addr == 0 || end == buf) {
 591                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 592                         goto error;
 593                 }
 594
 595                 /* get node id (socket id) */
 596                 nodestr = strstr(buf, " N");
 597                 if (nodestr == NULL) {
 598                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 599                         goto error;
 600                 }
 601                 nodestr += 2;
 602                 end = strstr(nodestr, "=");
 603                 if (end == NULL) {
 604                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 605                         goto error;
 606                 }
 607                 end[0] = '\0';
 608                 end = NULL;
 609
 610                 socket_id = strtoul(nodestr, &end, 0);
 611                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 612                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 613                         goto error;
 614                 }
 615
 616                 /* if we find this page in our mappings, set socket_id */
 617                 for (i = 0; i < hpi->num_pages[0]; i++) {
 618                         void *va = (void *)(unsigned long)virt_addr;
 619                         if (hugepg_tbl[i].orig_va == va) {
 620                                 hugepg_tbl[i].socket_id = socket_id;
 621                                 hp_count++;
 622 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 623                                 RTE_LOG(DEBUG, EAL,
 624                                         "Hugepage %s is on socket %d\n",
 625                                         hugepg_tbl[i].filepath, socket_id);
 626 #endif
 627                         }
 628                 }
 629         }
 630
 631         if (hp_count < hpi->num_pages[0])
 632                 goto error;
 633
 634         fclose(f);
 635         return 0;
 636
 637 error:
 638         fclose(f);
 639         return -1;
 640 }
 641
 642 static int
 643 cmp_physaddr(const void *a, const void *b)
 644 {
 645 #ifndef RTE_ARCH_PPC_64
 646         const struct hugepage_file *p1 = a;
 647         const struct hugepage_file *p2 = b;
 648 #else
 649         /* PowerPC needs memory sorted in reverse order from x86 */
 650         const struct hugepage_file *p1 = b;
 651         const struct hugepage_file *p2 = a;
 652 #endif
 653         if (p1->physaddr < p2->physaddr)
 654                 return -1;
 655         else if (p1->physaddr > p2->physaddr)
 656                 return 1;
 657         else
 658                 return 0;
 659 }
 660
 661 /*
 662  * Uses mmap to create a shared memory area for storage of data
 663  * Used in this file to store the hugepage file map on disk
 664  */
 665 static void *
 666 create_shared_memory(const char *filename, const size_t mem_size)
 667 {
 668         void *retval;
 669         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 670         if (fd < 0)
 671                 return NULL;
 672         if (ftruncate(fd, mem_size) < 0) {
 673                 close(fd);
 674                 return NULL;
 675         }
 676         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 677         close(fd);
 678         if (retval == MAP_FAILED)
 679                 return NULL;
 680         return retval;
 681 }
 682
 683 /*
 684  * this copies *active* hugepages from one hugepage table to another.
 685  * destination is typically the shared memory.
 686  */
 687 static int
 688 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
 689                 const struct hugepage_file * src, int src_size)
 690 {
 691         int src_pos, dst_pos = 0;
 692
 693         for (src_pos = 0; src_pos < src_size; src_pos++) {
 694                 if (src[src_pos].final_va != NULL) {
 695                         /* error on overflow attempt */
 696                         if (dst_pos == dest_size)
 697                                 return -1;
 698                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
 699                         dst_pos++;
 700                 }
 701         }
 702         return 0;
 703 }
 704
 705 static int
 706 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
 707                 unsigned num_hp_info)
 708 {
 709         unsigned socket, size;
 710         int page, nrpages = 0;
 711
 712         /* get total number of hugepages */
 713         for (size = 0; size < num_hp_info; size++)
 714                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 715                         nrpages +=
 716                         internal_config.hugepage_info[size].num_pages[socket];
 717
 718         for (page = 0; page < nrpages; page++) {
 719                 struct hugepage_file *hp = &hugepg_tbl[page];
 720
 721                 if (hp->final_va != NULL && unlink(hp->filepath)) {
 722                         RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
 723                                 __func__, hp->filepath, strerror(errno));
 724                 }
 725         }
 726         return 0;
 727 }
 728
 729 /*
 730  * unmaps hugepages that are not going to be used. since we originally allocate
 731  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 732  */
 733 static int
 734 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
 735                 struct hugepage_info *hpi,
 736                 unsigned num_hp_info)
 737 {
 738         unsigned socket, size;
 739         int page, nrpages = 0;
 740
 741         /* get total number of hugepages */
 742         for (size = 0; size < num_hp_info; size++)
 743                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 744                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 745
 746         for (size = 0; size < num_hp_info; size++) {
 747                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 748                         unsigned pages_found = 0;
 749
 750                         /* traverse until we have unmapped all the unused pages */
 751                         for (page = 0; page < nrpages; page++) {
 752                                 struct hugepage_file *hp = &hugepg_tbl[page];
 753
 754                                 /* find a page that matches the criteria */
 755                                 if ((hp->size == hpi[size].hugepage_sz) &&
 756                                                 (hp->socket_id == (int) socket)) {
 757
 758                                         /* if we skipped enough pages, unmap the rest */
 759                                         if (pages_found == hpi[size].num_pages[socket]) {
 760                                                 uint64_t unmap_len;
 761
 762                                                 unmap_len = hp->size;
 763
 764                                                 /* get start addr and len of the remaining segment */
 765                                                 munmap(hp->final_va, (size_t) unmap_len);
 766
 767                                                 hp->final_va = NULL;
 768                                                 if (unlink(hp->filepath) == -1) {
 769                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 770                                                                         __func__, hp->filepath, strerror(errno));
 771                                                         return -1;
 772                                                 }
 773                                         } else {
 774                                                 /* lock the page and skip */
 775                                                 pages_found++;
 776                                         }
 777
 778                                 } /* match page */
 779                         } /* foreach page */
 780                 } /* foreach socket */
 781         } /* foreach pagesize */
 782
 783         return 0;
 784 }
 785
 786 static inline uint64_t
 787 get_socket_mem_size(int socket)
 788 {
 789         uint64_t size = 0;
 790         unsigned i;
 791
 792         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 793                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 794                 if (hpi->hugedir != NULL)
 795                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 796         }
 797
 798         return size;
 799 }
 800
 801 /*
 802  * This function is a NUMA-aware equivalent of calc_num_pages.
 803  * It takes in the list of hugepage sizes and the
 804  * number of pages thereof, and calculates the best number of
 805  * pages of each size to fulfill the request for <memory> ram
 806  */
 807 static int
 808 calc_num_pages_per_socket(uint64_t * memory,
 809                 struct hugepage_info *hp_info,
 810                 struct hugepage_info *hp_used,
 811                 unsigned num_hp_info)
 812 {
 813         unsigned socket, j, i = 0;
 814         unsigned requested, available;
 815         int total_num_pages = 0;
 816         uint64_t remaining_mem, cur_mem;
 817         uint64_t total_mem = internal_config.memory;
 818
 819         if (num_hp_info == 0)
 820                 return -1;
 821
 822         /* if specific memory amounts per socket weren't requested */
 823         if (internal_config.force_sockets == 0) {
 824                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
 825                 size_t default_size, total_size;
 826                 unsigned lcore_id;
 827
 828                 /* Compute number of cores per socket */
 829                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
 830                 RTE_LCORE_FOREACH(lcore_id) {
 831                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
 832                 }
 833
 834                 /*
 835                  * Automatically spread requested memory amongst detected sockets according
 836                  * to number of cores from cpu mask present on each socket
 837                  */
 838                 total_size = internal_config.memory;
 839                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 840
 841                         /* Set memory amount per socket */
 842                         default_size = (internal_config.memory * cpu_per_socket[socket])
 843                                         / rte_lcore_count();
 844
 845                         /* Limit to maximum available memory on socket */
 846                         default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
 847
 848                         /* Update sizes */
 849                         memory[socket] = default_size;
 850                         total_size -= default_size;
 851                 }
 852
 853                 /*
 854                  * If some memory is remaining, try to allocate it by getting all
 855                  * available memory from sockets, one after the other
 856                  */
 857                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 858                         /* take whatever is available */
 859                         default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
 860                                                total_size);
 861
 862                         /* Update sizes */
 863                         memory[socket] += default_size;
 864                         total_size -= default_size;
 865                 }
 866         }
 867
 868         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 869                 /* skips if the memory on specific socket wasn't requested */
 870                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 871                         hp_used[i].hugedir = hp_info[i].hugedir;
 872                         hp_used[i].num_pages[socket] = RTE_MIN(
 873                                         memory[socket] / hp_info[i].hugepage_sz,
 874                                         hp_info[i].num_pages[socket]);
 875
 876                         cur_mem = hp_used[i].num_pages[socket] *
 877                                         hp_used[i].hugepage_sz;
 878
 879                         memory[socket] -= cur_mem;
 880                         total_mem -= cur_mem;
 881
 882                         total_num_pages += hp_used[i].num_pages[socket];
 883
 884                         /* check if we have met all memory requests */
 885                         if (memory[socket] == 0)
 886                                 break;
 887
 888                         /* check if we have any more pages left at this size, if so
 889                          * move on to next size */
 890                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
 891                                 continue;
 892                         /* At this point we know that there are more pages available that are
 893                          * bigger than the memory we want, so lets see if we can get enough
 894                          * from other page sizes.
 895                          */
 896                         remaining_mem = 0;
 897                         for (j = i+1; j < num_hp_info; j++)
 898                                 remaining_mem += hp_info[j].hugepage_sz *
 899                                 hp_info[j].num_pages[socket];
 900
 901                         /* is there enough other memory, if not allocate another page and quit */
 902                         if (remaining_mem < memory[socket]){
 903                                 cur_mem = RTE_MIN(memory[socket],
 904                                                 hp_info[i].hugepage_sz);
 905                                 memory[socket] -= cur_mem;
 906                                 total_mem -= cur_mem;
 907                                 hp_used[i].num_pages[socket]++;
 908                                 total_num_pages++;
 909                                 break; /* we are done with this socket*/
 910                         }
 911                 }
 912                 /* if we didn't satisfy all memory requirements per socket */
 913                 if (memory[socket] > 0) {
 914                         /* to prevent icc errors */
 915                         requested = (unsigned) (internal_config.socket_mem[socket] /
 916                                         0x100000);
 917                         available = requested -
 918                                         ((unsigned) (memory[socket] / 0x100000));
 919                         RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
 920                                         "Requested: %uMB, available: %uMB\n", socket,
 921                                         requested, available);
 922                         return -1;
 923                 }
 924         }
 925
 926         /* if we didn't satisfy total memory requirements */
 927         if (total_mem > 0) {
 928                 requested = (unsigned) (internal_config.memory / 0x100000);
 929                 available = requested - (unsigned) (total_mem / 0x100000);
 930                 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
 931                                 " available: %uMB\n", requested, available);
 932                 return -1;
 933         }
 934         return total_num_pages;
 935 }
 936
 937 static inline size_t
 938 eal_get_hugepage_mem_size(void)
 939 {
 940         uint64_t size = 0;
 941         unsigned i, j;
 942
 943         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 944                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 945                 if (hpi->hugedir != NULL) {
 946                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 947                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 948                         }
 949                 }
 950         }
 951
 952         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 953 }
 954
 955 static struct sigaction huge_action_old;
 956 static int huge_need_recover;
 957
 958 static void
 959 huge_register_sigbus(void)
 960 {
 961         sigset_t mask;
 962         struct sigaction action;
 963
 964         sigemptyset(&mask);
 965         sigaddset(&mask, SIGBUS);
 966         action.sa_flags = 0;
 967         action.sa_mask = mask;
 968         action.sa_handler = huge_sigbus_handler;
 969
 970         huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
 971 }
 972
 973 static void
 974 huge_recover_sigbus(void)
 975 {
 976         if (huge_need_recover) {
 977                 sigaction(SIGBUS, &huge_action_old, NULL);
 978                 huge_need_recover = 0;
 979         }
 980 }
 981
 982 /*
 983  * Prepare physical memory mapping: fill configuration structure with
 984  * these infos, return 0 on success.
 985  *  1. map N huge pages in separate files in hugetlbfs
 986  *  2. find associated physical addr
 987  *  3. find associated NUMA socket ID
 988  *  4. sort all huge pages by physical address
 989  *  5. remap these N huge pages in the correct order
 990  *  6. unmap the first mapping
 991  *  7. fill memsegs in configuration with contiguous zones
 992  */
 993 int
 994 rte_eal_hugepage_init(void)
 995 {
 996         struct rte_mem_config *mcfg;
 997         struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
 998         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
 999
1000         uint64_t memory[RTE_MAX_NUMA_NODES];
1001
1002         unsigned hp_offset;
1003         int i, j, new_memseg;
1004         int nr_hugefiles, nr_hugepages = 0;
1005         void *addr;
1006
1007         test_phys_addrs_available();
1008
1009         memset(used_hp, 0, sizeof(used_hp));
1010
1011         /* get pointer to global configuration */
1012         mcfg = rte_eal_get_configuration()->mem_config;
1013
1014         /* hugetlbfs can be disabled */
1015         if (internal_config.no_hugetlbfs) {
1016                 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1017                                 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1018                 if (addr == MAP_FAILED) {
1019                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1020                                         strerror(errno));
1021                         return -1;
1022                 }
1023                 if (rte_eal_iova_mode() == RTE_IOVA_VA)
1024                         mcfg->memseg[0].iova = (uintptr_t)addr;
1025                 else
1026                         mcfg->memseg[0].iova = RTE_BAD_IOVA;
1027                 mcfg->memseg[0].addr = addr;
1028                 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1029                 mcfg->memseg[0].len = internal_config.memory;
1030                 mcfg->memseg[0].socket_id = 0;
1031                 return 0;
1032         }
1033
1034         /* calculate total number of hugepages available. at this point we haven't
1035          * yet started sorting them so they all are on socket 0 */
1036         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1037                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1038                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1039
1040                 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1041         }
1042
1043         /*
1044          * allocate a memory area for hugepage table.
1045          * this isn't shared memory yet. due to the fact that we need some
1046          * processing done on these pages, shared memory will be created
1047          * at a later stage.
1048          */
1049         tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1050         if (tmp_hp == NULL)
1051                 goto fail;
1052
1053         memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1054
1055         hp_offset = 0; /* where we start the current page size entries */
1056
1057         huge_register_sigbus();
1058
1059         /* make a copy of socket_mem, needed for balanced allocation. */
1060         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1061                 memory[i] = internal_config.socket_mem[i];
1062
1063
1064         /* map all hugepages and sort them */
1065         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1066                 unsigned pages_old, pages_new;
1067                 struct hugepage_info *hpi;
1068
1069                 /*
1070                  * we don't yet mark hugepages as used at this stage, so
1071                  * we just map all hugepages available to the system
1072                  * all hugepages are still located on socket 0
1073                  */
1074                 hpi = &internal_config.hugepage_info[i];
1075
1076                 if (hpi->num_pages[0] == 0)
1077                         continue;
1078
1079                 /* map all hugepages available */
1080                 pages_old = hpi->num_pages[0];
1081                 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi,
1082                                               memory, 1);
1083                 if (pages_new < pages_old) {
1084                         RTE_LOG(DEBUG, EAL,
1085                                 "%d not %d hugepages of size %u MB allocated\n",
1086                                 pages_new, pages_old,
1087                                 (unsigned)(hpi->hugepage_sz / 0x100000));
1088
1089                         int pages = pages_old - pages_new;
1090
1091                         nr_hugepages -= pages;
1092                         hpi->num_pages[0] = pages_new;
1093                         if (pages_new == 0)
1094                                 continue;
1095                 }
1096
1097                 if (phys_addrs_available &&
1098                                 rte_eal_iova_mode() != RTE_IOVA_VA) {
1099                         /* find physical addresses for each hugepage */
1100                         if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1101                                 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1102                                         "for %u MB pages\n",
1103                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1104                                 goto fail;
1105                         }
1106                 } else {
1107                         /* set physical addresses for each hugepage */
1108                         if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1109                                 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1110                                         "for %u MB pages\n",
1111                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1112                                 goto fail;
1113                         }
1114                 }
1115
1116                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1117                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1118                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1119                         goto fail;
1120                 }
1121
1122                 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1123                       sizeof(struct hugepage_file), cmp_physaddr);
1124
1125                 /* remap all hugepages */
1126                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, NULL, 0) !=
1127                     hpi->num_pages[0]) {
1128                         RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1129                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1130                         goto fail;
1131                 }
1132
1133                 /* unmap original mappings */
1134                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1135                         goto fail;
1136
1137                 /* we have processed a num of hugepages of this size, so inc offset */
1138                 hp_offset += hpi->num_pages[0];
1139         }
1140
1141         huge_recover_sigbus();
1142
1143         if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1144                 internal_config.memory = eal_get_hugepage_mem_size();
1145
1146         nr_hugefiles = nr_hugepages;
1147
1148
1149         /* clean out the numbers of pages */
1150         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1151                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1152                         internal_config.hugepage_info[i].num_pages[j] = 0;
1153
1154         /* get hugepages for each socket */
1155         for (i = 0; i < nr_hugefiles; i++) {
1156                 int socket = tmp_hp[i].socket_id;
1157
1158                 /* find a hugepage info with right size and increment num_pages */
1159                 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1160                                 (int)internal_config.num_hugepage_sizes);
1161                 for (j = 0; j < nb_hpsizes; j++) {
1162                         if (tmp_hp[i].size ==
1163                                         internal_config.hugepage_info[j].hugepage_sz) {
1164                                 internal_config.hugepage_info[j].num_pages[socket]++;
1165                         }
1166                 }
1167         }
1168
1169         /* make a copy of socket_mem, needed for number of pages calculation */
1170         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1171                 memory[i] = internal_config.socket_mem[i];
1172
1173         /* calculate final number of pages */
1174         nr_hugepages = calc_num_pages_per_socket(memory,
1175                         internal_config.hugepage_info, used_hp,
1176                         internal_config.num_hugepage_sizes);
1177
1178         /* error if not enough memory available */
1179         if (nr_hugepages < 0)
1180                 goto fail;
1181
1182         /* reporting in! */
1183         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1184                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1185                         if (used_hp[i].num_pages[j] > 0) {
1186                                 RTE_LOG(DEBUG, EAL,
1187                                         "Requesting %u pages of size %uMB"
1188                                         " from socket %i\n",
1189                                         used_hp[i].num_pages[j],
1190                                         (unsigned)
1191                                         (used_hp[i].hugepage_sz / 0x100000),
1192                                         j);
1193                         }
1194                 }
1195         }
1196
1197         /* create shared memory */
1198         hugepage = create_shared_memory(eal_hugepage_info_path(),
1199                         nr_hugefiles * sizeof(struct hugepage_file));
1200
1201         if (hugepage == NULL) {
1202                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1203                 goto fail;
1204         }
1205         memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1206
1207         /*
1208          * unmap pages that we won't need (looks at used_hp).
1209          * also, sets final_va to NULL on pages that were unmapped.
1210          */
1211         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1212                         internal_config.num_hugepage_sizes) < 0) {
1213                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1214                 goto fail;
1215         }
1216
1217         /*
1218          * copy stuff from malloc'd hugepage* to the actual shared memory.
1219          * this procedure only copies those hugepages that have final_va
1220          * not NULL. has overflow protection.
1221          */
1222         if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1223                         tmp_hp, nr_hugefiles) < 0) {
1224                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1225                 goto fail;
1226         }
1227
1228         /* free the hugepage backing files */
1229         if (internal_config.hugepage_unlink &&
1230                 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1231                 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1232                 goto fail;
1233         }
1234
1235         /* free the temporary hugepage table */
1236         free(tmp_hp);
1237         tmp_hp = NULL;
1238
1239         /* first memseg index shall be 0 after incrementing it below */
1240         j = -1;
1241         for (i = 0; i < nr_hugefiles; i++) {
1242                 new_memseg = 0;
1243
1244                 /* if this is a new section, create a new memseg */
1245                 if (i == 0)
1246                         new_memseg = 1;
1247                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1248                         new_memseg = 1;
1249                 else if (hugepage[i].size != hugepage[i-1].size)
1250                         new_memseg = 1;
1251
1252 #ifdef RTE_ARCH_PPC_64
1253                 /* On PPC64 architecture, the mmap always start from higher
1254                  * virtual address to lower address. Here, both the physical
1255                  * address and virtual address are in descending order */
1256                 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1257                     hugepage[i].size)
1258                         new_memseg = 1;
1259                 else if (((unsigned long)hugepage[i-1].final_va -
1260                     (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1261                         new_memseg = 1;
1262 #else
1263                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1264                     hugepage[i].size)
1265                         new_memseg = 1;
1266                 else if (((unsigned long)hugepage[i].final_va -
1267                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1268                         new_memseg = 1;
1269 #endif
1270
1271                 if (new_memseg) {
1272                         j += 1;
1273                         if (j == RTE_MAX_MEMSEG)
1274                                 break;
1275
1276                         mcfg->memseg[j].iova = hugepage[i].physaddr;
1277                         mcfg->memseg[j].addr = hugepage[i].final_va;
1278                         mcfg->memseg[j].len = hugepage[i].size;
1279                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1280                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1281                 }
1282                 /* continuation of previous memseg */
1283                 else {
1284 #ifdef RTE_ARCH_PPC_64
1285                 /* Use the phy and virt address of the last page as segment
1286                  * address for IBM Power architecture */
1287                         mcfg->memseg[j].iova = hugepage[i].physaddr;
1288                         mcfg->memseg[j].addr = hugepage[i].final_va;
1289 #endif
1290                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1291                 }
1292                 hugepage[i].memseg_id = j;
1293         }
1294
1295         if (i < nr_hugefiles) {
1296                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1297                         "from %d requested\n"
1298                         "Current %s=%d is not enough\n"
1299                         "Please either increase it or request less amount "
1300                         "of memory.\n",
1301                         i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1302                         RTE_MAX_MEMSEG);
1303                 goto fail;
1304         }
1305
1306         munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1307
1308         return 0;
1309
1310 fail:
1311         huge_recover_sigbus();
1312         free(tmp_hp);
1313         if (hugepage != NULL)
1314                 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1315
1316         return -1;
1317 }
1318
1319 /*
1320  * uses fstat to report the size of a file on disk
1321  */
1322 static off_t
1323 getFileSize(int fd)
1324 {
1325         struct stat st;
1326         if (fstat(fd, &st) < 0)
1327                 return 0;
1328         return st.st_size;
1329 }
1330
1331 /*
1332  * This creates the memory mappings in the secondary process to match that of
1333  * the server process. It goes through each memory segment in the DPDK runtime
1334  * configuration and finds the hugepages which form that segment, mapping them
1335  * in order to form a contiguous block in the virtual memory space
1336  */
1337 int
1338 rte_eal_hugepage_attach(void)
1339 {
1340         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1341         struct hugepage_file *hp = NULL;
1342         unsigned num_hp = 0;
1343         unsigned i, s = 0; /* s used to track the segment number */
1344         unsigned max_seg = RTE_MAX_MEMSEG;
1345         off_t size = 0;
1346         int fd, fd_zero = -1, fd_hugepage = -1;
1347
1348         if (aslr_enabled() > 0) {
1349                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1350                                 "(ASLR) is enabled in the kernel.\n");
1351                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1352                                 "into secondary processes\n");
1353         }
1354
1355         test_phys_addrs_available();
1356
1357         fd_zero = open("/dev/zero", O_RDONLY);
1358         if (fd_zero < 0) {
1359                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1360                 goto error;
1361         }
1362         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1363         if (fd_hugepage < 0) {
1364                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1365                 goto error;
1366         }
1367
1368         /* map all segments into memory to make sure we get the addrs */
1369         for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1370                 void *base_addr;
1371
1372                 /*
1373                  * the first memory segment with len==0 is the one that
1374                  * follows the last valid segment.
1375                  */
1376                 if (mcfg->memseg[s].len == 0)
1377                         break;
1378
1379                 /*
1380                  * fdzero is mmapped to get a contiguous block of virtual
1381                  * addresses of the appropriate memseg size.
1382                  * use mmap to get identical addresses as the primary process.
1383                  */
1384                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1385                                  PROT_READ,
1386 #ifdef RTE_ARCH_PPC_64
1387                                  MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
1388 #else
1389                                  MAP_PRIVATE,
1390 #endif
1391                                  fd_zero, 0);
1392                 if (base_addr == MAP_FAILED ||
1393                     base_addr != mcfg->memseg[s].addr) {
1394                         max_seg = s;
1395                         if (base_addr != MAP_FAILED) {
1396                                 /* errno is stale, don't use */
1397                                 RTE_LOG(ERR, EAL, "Could not mmap %zu bytes "
1398                                         "in /dev/zero at [%p], got [%p] - "
1399                                         "please use '--base-virtaddr' option\n",
1400                                         mcfg->memseg[s].len,
1401                                         mcfg->memseg[s].addr, base_addr);
1402                                 munmap(base_addr, mcfg->memseg[s].len);
1403                         } else {
1404                                 RTE_LOG(ERR, EAL, "Could not mmap %zu bytes "
1405                                         "in /dev/zero at [%p]: '%s'\n",
1406                                         mcfg->memseg[s].len,
1407                                         mcfg->memseg[s].addr, strerror(errno));
1408                         }
1409                         if (aslr_enabled() > 0) {
1410                                 RTE_LOG(ERR, EAL, "It is recommended to "
1411                                         "disable ASLR in the kernel "
1412                                         "and retry running both primary "
1413                                         "and secondary processes\n");
1414                         }
1415                         goto error;
1416                 }
1417         }
1418
1419         size = getFileSize(fd_hugepage);
1420         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1421         if (hp == MAP_FAILED) {
1422                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1423                 goto error;
1424         }
1425
1426         num_hp = size / sizeof(struct hugepage_file);
1427         RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1428
1429         s = 0;
1430         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1431                 void *addr, *base_addr;
1432                 uintptr_t offset = 0;
1433                 size_t mapping_size;
1434                 /*
1435                  * free previously mapped memory so we can map the
1436                  * hugepages into the space
1437                  */
1438                 base_addr = mcfg->memseg[s].addr;
1439                 munmap(base_addr, mcfg->memseg[s].len);
1440
1441                 /* find the hugepages for this segment and map them
1442                  * we don't need to worry about order, as the server sorted the
1443                  * entries before it did the second mmap of them */
1444                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1445                         if (hp[i].memseg_id == (int)s){
1446                                 fd = open(hp[i].filepath, O_RDWR);
1447                                 if (fd < 0) {
1448                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1449                                                 hp[i].filepath);
1450                                         goto error;
1451                                 }
1452                                 mapping_size = hp[i].size;
1453                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1454                                                 mapping_size, PROT_READ | PROT_WRITE,
1455                                                 MAP_SHARED, fd, 0);
1456                                 close(fd); /* close file both on success and on failure */
1457                                 if (addr == MAP_FAILED ||
1458                                                 addr != RTE_PTR_ADD(base_addr, offset)) {
1459                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1460                                                 hp[i].filepath);
1461                                         goto error;
1462                                 }
1463                                 offset+=mapping_size;
1464                         }
1465                 }
1466                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1467                                 (unsigned long long)mcfg->memseg[s].len);
1468                 s++;
1469         }
1470         /* unmap the hugepage config file, since we are done using it */
1471         munmap(hp, size);
1472         close(fd_zero);
1473         close(fd_hugepage);
1474         return 0;
1475
1476 error:
1477         for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1478                 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1479         if (hp != NULL && hp != MAP_FAILED)
1480                 munmap(hp, size);
1481         if (fd_zero >= 0)
1482                 close(fd_zero);
1483         if (fd_hugepage >= 0)
1484                 close(fd_hugepage);
1485         return -1;
1486 }
1487
1488 int
1489 rte_eal_using_phys_addrs(void)
1490 {
1491         return phys_addrs_available;
1492 }