lib/librte_eal/linuxapp/eal/eal_memory.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   Copyright(c) 2013 6WIND.
   6  *   All rights reserved.
   7  *
   8  *   Redistribution and use in source and binary forms, with or without
   9  *   modification, are permitted provided that the following conditions
  10  *   are met:
  11  *
  12  *     * Redistributions of source code must retain the above copyright
  13  *       notice, this list of conditions and the following disclaimer.
  14  *     * Redistributions in binary form must reproduce the above copyright
  15  *       notice, this list of conditions and the following disclaimer in
  16  *       the documentation and/or other materials provided with the
  17  *       distribution.
  18  *     * Neither the name of Intel Corporation nor the names of its
  19  *       contributors may be used to endorse or promote products derived
  20  *       from this software without specific prior written permission.
  21  *
  22  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  25  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  26  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  27  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  28  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  29  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  30  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  31  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  32  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  33  */
  34
  35 #define _FILE_OFFSET_BITS 64
  36 #include <errno.h>
  37 #include <stdarg.h>
  38 #include <stdbool.h>
  39 #include <stdlib.h>
  40 #include <stdio.h>
  41 #include <stdint.h>
  42 #include <inttypes.h>
  43 #include <string.h>
  44 #include <sys/mman.h>
  45 #include <sys/types.h>
  46 #include <sys/stat.h>
  47 #include <sys/queue.h>
  48 #include <sys/file.h>
  49 #include <unistd.h>
  50 #include <limits.h>
  51 #include <sys/ioctl.h>
  52 #include <sys/time.h>
  53 #include <signal.h>
  54 #include <setjmp.h>
  55 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
  56 #include <numa.h>
  57 #include <numaif.h>
  58 #endif
  59
  60 #include <rte_log.h>
  61 #include <rte_memory.h>
  62 #include <rte_memzone.h>
  63 #include <rte_launch.h>
  64 #include <rte_eal.h>
  65 #include <rte_eal_memconfig.h>
  66 #include <rte_per_lcore.h>
  67 #include <rte_lcore.h>
  68 #include <rte_common.h>
  69 #include <rte_string_fns.h>
  70
  71 #include "eal_private.h"
  72 #include "eal_internal_cfg.h"
  73 #include "eal_filesystem.h"
  74 #include "eal_hugepages.h"
  75
  76 #define PFN_MASK_SIZE   8
  77
  78 /**
  79  * @file
  80  * Huge page mapping under linux
  81  *
  82  * To reserve a big contiguous amount of memory, we use the hugepage
  83  * feature of linux. For that, we need to have hugetlbfs mounted. This
  84  * code will create many files in this directory (one per page) and
  85  * map them in virtual memory. For each page, we will retrieve its
  86  * physical address and remap it in order to have a virtual contiguous
  87  * zone as well as a physical contiguous zone.
  88  */
  89
  90 static uint64_t baseaddr_offset;
  91
  92 static bool phys_addrs_available = true;
  93
  94 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
  95
  96 static void
  97 test_phys_addrs_available(void)
  98 {
  99         uint64_t tmp;
 100         phys_addr_t physaddr;
 101
 102         if (!rte_eal_has_hugepages()) {
 103                 RTE_LOG(ERR, EAL,
 104                         "Started without hugepages support, physical addresses not available\n");
 105                 phys_addrs_available = false;
 106                 return;
 107         }
 108
 109         physaddr = rte_mem_virt2phy(&tmp);
 110         if (physaddr == RTE_BAD_PHYS_ADDR) {
 111                 RTE_LOG(ERR, EAL,
 112                         "Cannot obtain physical addresses: %s. "
 113                         "Only vfio will function.\n",
 114                         strerror(errno));
 115                 phys_addrs_available = false;
 116         }
 117 }
 118
 119 /*
 120  * Get physical address of any mapped virtual address in the current process.
 121  */
 122 phys_addr_t
 123 rte_mem_virt2phy(const void *virtaddr)
 124 {
 125         int fd, retval;
 126         uint64_t page, physaddr;
 127         unsigned long virt_pfn;
 128         int page_size;
 129         off_t offset;
 130
 131         if (rte_eal_iova_mode() == RTE_IOVA_VA)
 132                 return (uintptr_t)virtaddr;
 133
 134         /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
 135         if (!phys_addrs_available)
 136                 return RTE_BAD_PHYS_ADDR;
 137
 138         /* standard page size */
 139         page_size = getpagesize();
 140
 141         fd = open("/proc/self/pagemap", O_RDONLY);
 142         if (fd < 0) {
 143                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 144                         __func__, strerror(errno));
 145                 return RTE_BAD_PHYS_ADDR;
 146         }
 147
 148         virt_pfn = (unsigned long)virtaddr / page_size;
 149         offset = sizeof(uint64_t) * virt_pfn;
 150         if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 151                 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 152                                 __func__, strerror(errno));
 153                 close(fd);
 154                 return RTE_BAD_PHYS_ADDR;
 155         }
 156
 157         retval = read(fd, &page, PFN_MASK_SIZE);
 158         close(fd);
 159         if (retval < 0) {
 160                 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 161                                 __func__, strerror(errno));
 162                 return RTE_BAD_PHYS_ADDR;
 163         } else if (retval != PFN_MASK_SIZE) {
 164                 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
 165                                 "but expected %d:\n",
 166                                 __func__, retval, PFN_MASK_SIZE);
 167                 return RTE_BAD_PHYS_ADDR;
 168         }
 169
 170         /*
 171          * the pfn (page frame number) are bits 0-54 (see
 172          * pagemap.txt in linux Documentation)
 173          */
 174         if ((page & 0x7fffffffffffffULL) == 0)
 175                 return RTE_BAD_PHYS_ADDR;
 176
 177         physaddr = ((page & 0x7fffffffffffffULL) * page_size)
 178                 + ((unsigned long)virtaddr % page_size);
 179
 180         return physaddr;
 181 }
 182
 183 /*
 184  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 185  * it by browsing the /proc/self/pagemap special file.
 186  */
 187 static int
 188 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 189 {
 190         unsigned int i;
 191         phys_addr_t addr;
 192
 193         for (i = 0; i < hpi->num_pages[0]; i++) {
 194                 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
 195                 if (addr == RTE_BAD_PHYS_ADDR)
 196                         return -1;
 197                 hugepg_tbl[i].physaddr = addr;
 198         }
 199         return 0;
 200 }
 201
 202 /*
 203  * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
 204  */
 205 static int
 206 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 207 {
 208         unsigned int i;
 209         static phys_addr_t addr;
 210
 211         for (i = 0; i < hpi->num_pages[0]; i++) {
 212                 hugepg_tbl[i].physaddr = addr;
 213                 addr += hugepg_tbl[i].size;
 214         }
 215         return 0;
 216 }
 217
 218 /*
 219  * Check whether address-space layout randomization is enabled in
 220  * the kernel. This is important for multi-process as it can prevent
 221  * two processes mapping data to the same virtual address
 222  * Returns:
 223  *    0 - address space randomization disabled
 224  *    1/2 - address space randomization enabled
 225  *    negative error code on error
 226  */
 227 static int
 228 aslr_enabled(void)
 229 {
 230         char c;
 231         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 232         if (fd < 0)
 233                 return -errno;
 234         retval = read(fd, &c, 1);
 235         close(fd);
 236         if (retval < 0)
 237                 return -errno;
 238         if (retval == 0)
 239                 return -EIO;
 240         switch (c) {
 241                 case '0' : return 0;
 242                 case '1' : return 1;
 243                 case '2' : return 2;
 244                 default: return -EINVAL;
 245         }
 246 }
 247
 248 /*
 249  * Try to mmap *size bytes in /dev/zero. If it is successful, return the
 250  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 251  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 252  * 0. In this case, return NULL. Note: this function returns an address
 253  * which is a multiple of hugepage size.
 254  */
 255 static void *
 256 get_virtual_area(size_t *size, size_t hugepage_sz)
 257 {
 258         void *addr;
 259         int fd;
 260         long aligned_addr;
 261
 262         if (internal_config.base_virtaddr != 0) {
 263                 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
 264                                 baseaddr_offset);
 265         }
 266         else addr = NULL;
 267
 268         RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
 269
 270         fd = open("/dev/zero", O_RDONLY);
 271         if (fd < 0){
 272                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 273                 return NULL;
 274         }
 275         do {
 276                 addr = mmap(addr,
 277                                 (*size) + hugepage_sz, PROT_READ,
 278 #ifdef RTE_ARCH_PPC_64
 279                                 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
 280 #else
 281                                 MAP_PRIVATE,
 282 #endif
 283                                 fd, 0);
 284                 if (addr == MAP_FAILED)
 285                         *size -= hugepage_sz;
 286         } while (addr == MAP_FAILED && *size > 0);
 287
 288         if (addr == MAP_FAILED) {
 289                 close(fd);
 290                 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
 291                         strerror(errno));
 292                 return NULL;
 293         }
 294
 295         munmap(addr, (*size) + hugepage_sz);
 296         close(fd);
 297
 298         /* align addr to a huge page size boundary */
 299         aligned_addr = (long)addr;
 300         aligned_addr += (hugepage_sz - 1);
 301         aligned_addr &= (~(hugepage_sz - 1));
 302         addr = (void *)(aligned_addr);
 303
 304         RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
 305                 addr, *size);
 306
 307         /* increment offset */
 308         baseaddr_offset += *size;
 309
 310         return addr;
 311 }
 312
 313 static sigjmp_buf huge_jmpenv;
 314
 315 static void huge_sigbus_handler(int signo __rte_unused)
 316 {
 317         siglongjmp(huge_jmpenv, 1);
 318 }
 319
 320 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
 321  * non-static local variable in the stack frame calling sigsetjmp might be
 322  * clobbered by a call to longjmp.
 323  */
 324 static int huge_wrap_sigsetjmp(void)
 325 {
 326         return sigsetjmp(huge_jmpenv, 1);
 327 }
 328
 329 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 330 /* Callback for numa library. */
 331 void numa_error(char *where)
 332 {
 333         RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
 334 }
 335 #endif
 336
 337 /*
 338  * Mmap all hugepages of hugepage table: it first open a file in
 339  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 340  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 341  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 342  * map continguous physical blocks in contiguous virtual blocks.
 343  */
 344 static unsigned
 345 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
 346                   uint64_t *essential_memory __rte_unused, int orig)
 347 {
 348         int fd;
 349         unsigned i;
 350         void *virtaddr;
 351         void *vma_addr = NULL;
 352         size_t vma_len = 0;
 353 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 354         int node_id = -1;
 355         int essential_prev = 0;
 356         int oldpolicy;
 357         struct bitmask *oldmask = numa_allocate_nodemask();
 358         bool have_numa = true;
 359         unsigned long maxnode = 0;
 360
 361         /* Check if kernel supports NUMA. */
 362         if (numa_available() != 0) {
 363                 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
 364                 have_numa = false;
 365         }
 366
 367         if (orig && have_numa) {
 368                 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
 369                 if (get_mempolicy(&oldpolicy, oldmask->maskp,
 370                                   oldmask->size + 1, 0, 0) < 0) {
 371                         RTE_LOG(ERR, EAL,
 372                                 "Failed to get current mempolicy: %s. "
 373                                 "Assuming MPOL_DEFAULT.\n", strerror(errno));
 374                         oldpolicy = MPOL_DEFAULT;
 375                 }
 376                 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 377                         if (internal_config.socket_mem[i])
 378                                 maxnode = i + 1;
 379         }
 380 #endif
 381
 382         for (i = 0; i < hpi->num_pages[0]; i++) {
 383                 uint64_t hugepage_sz = hpi->hugepage_sz;
 384
 385 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 386                 if (maxnode) {
 387                         unsigned int j;
 388
 389                         for (j = 0; j < maxnode; j++)
 390                                 if (essential_memory[j])
 391                                         break;
 392
 393                         if (j == maxnode) {
 394                                 node_id = (node_id + 1) % maxnode;
 395                                 while (!internal_config.socket_mem[node_id]) {
 396                                         node_id++;
 397                                         node_id %= maxnode;
 398                                 }
 399                                 essential_prev = 0;
 400                         } else {
 401                                 node_id = j;
 402                                 essential_prev = essential_memory[j];
 403
 404                                 if (essential_memory[j] < hugepage_sz)
 405                                         essential_memory[j] = 0;
 406                                 else
 407                                         essential_memory[j] -= hugepage_sz;
 408                         }
 409
 410                         RTE_LOG(DEBUG, EAL,
 411                                 "Setting policy MPOL_PREFERRED for socket %d\n",
 412                                 node_id);
 413                         numa_set_preferred(node_id);
 414                 }
 415 #endif
 416
 417                 if (orig) {
 418                         hugepg_tbl[i].file_id = i;
 419                         hugepg_tbl[i].size = hugepage_sz;
 420                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 421                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 422                                         hugepg_tbl[i].file_id);
 423                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 424                 }
 425 #ifndef RTE_ARCH_64
 426                 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
 427                  * original map address as final map address.
 428                  */
 429                 else if ((hugepage_sz == RTE_PGSIZE_1G)
 430                         || (hugepage_sz == RTE_PGSIZE_16G)) {
 431                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 432                         hugepg_tbl[i].orig_va = NULL;
 433                         continue;
 434                 }
 435 #endif
 436                 else if (vma_len == 0) {
 437                         unsigned j, num_pages;
 438
 439                         /* reserve a virtual area for next contiguous
 440                          * physical block: count the number of
 441                          * contiguous physical pages. */
 442                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 443 #ifdef RTE_ARCH_PPC_64
 444                                 /* The physical addresses are sorted in
 445                                  * descending order on PPC64 */
 446                                 if (hugepg_tbl[j].physaddr !=
 447                                     hugepg_tbl[j-1].physaddr - hugepage_sz)
 448                                         break;
 449 #else
 450                                 if (hugepg_tbl[j].physaddr !=
 451                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 452                                         break;
 453 #endif
 454                         }
 455                         num_pages = j - i;
 456                         vma_len = num_pages * hugepage_sz;
 457
 458                         /* get the biggest virtual memory area up to
 459                          * vma_len. If it fails, vma_addr is NULL, so
 460                          * let the kernel provide the address. */
 461                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 462                         if (vma_addr == NULL)
 463                                 vma_len = hugepage_sz;
 464                 }
 465
 466                 /* try to create hugepage file */
 467                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
 468                 if (fd < 0) {
 469                         RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
 470                                         strerror(errno));
 471                         goto out;
 472                 }
 473
 474                 /* map the segment, and populate page tables,
 475                  * the kernel fills this segment with zeros */
 476                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 477                                 MAP_SHARED | MAP_POPULATE, fd, 0);
 478                 if (virtaddr == MAP_FAILED) {
 479                         RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
 480                                         strerror(errno));
 481                         close(fd);
 482                         goto out;
 483                 }
 484
 485                 if (orig) {
 486                         hugepg_tbl[i].orig_va = virtaddr;
 487                 }
 488                 else {
 489                         hugepg_tbl[i].final_va = virtaddr;
 490                 }
 491
 492                 if (orig) {
 493                         /* In linux, hugetlb limitations, like cgroup, are
 494                          * enforced at fault time instead of mmap(), even
 495                          * with the option of MAP_POPULATE. Kernel will send
 496                          * a SIGBUS signal. To avoid to be killed, save stack
 497                          * environment here, if SIGBUS happens, we can jump
 498                          * back here.
 499                          */
 500                         if (huge_wrap_sigsetjmp()) {
 501                                 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
 502                                         "hugepages of size %u MB\n",
 503                                         (unsigned)(hugepage_sz / 0x100000));
 504                                 munmap(virtaddr, hugepage_sz);
 505                                 close(fd);
 506                                 unlink(hugepg_tbl[i].filepath);
 507 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 508                                 if (maxnode)
 509                                         essential_memory[node_id] =
 510                                                 essential_prev;
 511 #endif
 512                                 goto out;
 513                         }
 514                         *(int *)virtaddr = 0;
 515                 }
 516
 517
 518                 /* set shared flock on the file. */
 519                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 520                         RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
 521                                 __func__, strerror(errno));
 522                         close(fd);
 523                         goto out;
 524                 }
 525
 526                 close(fd);
 527
 528                 vma_addr = (char *)vma_addr + hugepage_sz;
 529                 vma_len -= hugepage_sz;
 530         }
 531
 532 out:
 533 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 534         if (maxnode) {
 535                 RTE_LOG(DEBUG, EAL,
 536                         "Restoring previous memory policy: %d\n", oldpolicy);
 537                 if (oldpolicy == MPOL_DEFAULT) {
 538                         numa_set_localalloc();
 539                 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
 540                                          oldmask->size + 1) < 0) {
 541                         RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
 542                                 strerror(errno));
 543                         numa_set_localalloc();
 544                 }
 545         }
 546         numa_free_cpumask(oldmask);
 547 #endif
 548         return i;
 549 }
 550
 551 /* Unmap all hugepages from original mapping */
 552 static int
 553 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 554 {
 555         unsigned i;
 556         for (i = 0; i < hpi->num_pages[0]; i++) {
 557                 if (hugepg_tbl[i].orig_va) {
 558                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 559                         hugepg_tbl[i].orig_va = NULL;
 560                 }
 561         }
 562         return 0;
 563 }
 564
 565 /*
 566  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 567  * page.
 568  */
 569 static int
 570 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 571 {
 572         int socket_id;
 573         char *end, *nodestr;
 574         unsigned i, hp_count = 0;
 575         uint64_t virt_addr;
 576         char buf[BUFSIZ];
 577         char hugedir_str[PATH_MAX];
 578         FILE *f;
 579
 580         f = fopen("/proc/self/numa_maps", "r");
 581         if (f == NULL) {
 582                 RTE_LOG(NOTICE, EAL, "NUMA support not available"
 583                         " consider that all memory is in socket_id 0\n");
 584                 return 0;
 585         }
 586
 587         snprintf(hugedir_str, sizeof(hugedir_str),
 588                         "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
 589
 590         /* parse numa map */
 591         while (fgets(buf, sizeof(buf), f) != NULL) {
 592
 593                 /* ignore non huge page */
 594                 if (strstr(buf, " huge ") == NULL &&
 595                                 strstr(buf, hugedir_str) == NULL)
 596                         continue;
 597
 598                 /* get zone addr */
 599                 virt_addr = strtoull(buf, &end, 16);
 600                 if (virt_addr == 0 || end == buf) {
 601                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 602                         goto error;
 603                 }
 604
 605                 /* get node id (socket id) */
 606                 nodestr = strstr(buf, " N");
 607                 if (nodestr == NULL) {
 608                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 609                         goto error;
 610                 }
 611                 nodestr += 2;
 612                 end = strstr(nodestr, "=");
 613                 if (end == NULL) {
 614                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 615                         goto error;
 616                 }
 617                 end[0] = '\0';
 618                 end = NULL;
 619
 620                 socket_id = strtoul(nodestr, &end, 0);
 621                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 622                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 623                         goto error;
 624                 }
 625
 626                 /* if we find this page in our mappings, set socket_id */
 627                 for (i = 0; i < hpi->num_pages[0]; i++) {
 628                         void *va = (void *)(unsigned long)virt_addr;
 629                         if (hugepg_tbl[i].orig_va == va) {
 630                                 hugepg_tbl[i].socket_id = socket_id;
 631                                 hp_count++;
 632 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 633                                 RTE_LOG(DEBUG, EAL,
 634                                         "Hugepage %s is on socket %d\n",
 635                                         hugepg_tbl[i].filepath, socket_id);
 636 #endif
 637                         }
 638                 }
 639         }
 640
 641         if (hp_count < hpi->num_pages[0])
 642                 goto error;
 643
 644         fclose(f);
 645         return 0;
 646
 647 error:
 648         fclose(f);
 649         return -1;
 650 }
 651
 652 static int
 653 cmp_physaddr(const void *a, const void *b)
 654 {
 655 #ifndef RTE_ARCH_PPC_64
 656         const struct hugepage_file *p1 = a;
 657         const struct hugepage_file *p2 = b;
 658 #else
 659         /* PowerPC needs memory sorted in reverse order from x86 */
 660         const struct hugepage_file *p1 = b;
 661         const struct hugepage_file *p2 = a;
 662 #endif
 663         if (p1->physaddr < p2->physaddr)
 664                 return -1;
 665         else if (p1->physaddr > p2->physaddr)
 666                 return 1;
 667         else
 668                 return 0;
 669 }
 670
 671 /*
 672  * Uses mmap to create a shared memory area for storage of data
 673  * Used in this file to store the hugepage file map on disk
 674  */
 675 static void *
 676 create_shared_memory(const char *filename, const size_t mem_size)
 677 {
 678         void *retval;
 679         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 680         if (fd < 0)
 681                 return NULL;
 682         if (ftruncate(fd, mem_size) < 0) {
 683                 close(fd);
 684                 return NULL;
 685         }
 686         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 687         close(fd);
 688         return retval;
 689 }
 690
 691 /*
 692  * this copies *active* hugepages from one hugepage table to another.
 693  * destination is typically the shared memory.
 694  */
 695 static int
 696 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
 697                 const struct hugepage_file * src, int src_size)
 698 {
 699         int src_pos, dst_pos = 0;
 700
 701         for (src_pos = 0; src_pos < src_size; src_pos++) {
 702                 if (src[src_pos].final_va != NULL) {
 703                         /* error on overflow attempt */
 704                         if (dst_pos == dest_size)
 705                                 return -1;
 706                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
 707                         dst_pos++;
 708                 }
 709         }
 710         return 0;
 711 }
 712
 713 static int
 714 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
 715                 unsigned num_hp_info)
 716 {
 717         unsigned socket, size;
 718         int page, nrpages = 0;
 719
 720         /* get total number of hugepages */
 721         for (size = 0; size < num_hp_info; size++)
 722                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 723                         nrpages +=
 724                         internal_config.hugepage_info[size].num_pages[socket];
 725
 726         for (page = 0; page < nrpages; page++) {
 727                 struct hugepage_file *hp = &hugepg_tbl[page];
 728
 729                 if (hp->final_va != NULL && unlink(hp->filepath)) {
 730                         RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
 731                                 __func__, hp->filepath, strerror(errno));
 732                 }
 733         }
 734         return 0;
 735 }
 736
 737 /*
 738  * unmaps hugepages that are not going to be used. since we originally allocate
 739  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 740  */
 741 static int
 742 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
 743                 struct hugepage_info *hpi,
 744                 unsigned num_hp_info)
 745 {
 746         unsigned socket, size;
 747         int page, nrpages = 0;
 748
 749         /* get total number of hugepages */
 750         for (size = 0; size < num_hp_info; size++)
 751                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 752                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 753
 754         for (size = 0; size < num_hp_info; size++) {
 755                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 756                         unsigned pages_found = 0;
 757
 758                         /* traverse until we have unmapped all the unused pages */
 759                         for (page = 0; page < nrpages; page++) {
 760                                 struct hugepage_file *hp = &hugepg_tbl[page];
 761
 762                                 /* find a page that matches the criteria */
 763                                 if ((hp->size == hpi[size].hugepage_sz) &&
 764                                                 (hp->socket_id == (int) socket)) {
 765
 766                                         /* if we skipped enough pages, unmap the rest */
 767                                         if (pages_found == hpi[size].num_pages[socket]) {
 768                                                 uint64_t unmap_len;
 769
 770                                                 unmap_len = hp->size;
 771
 772                                                 /* get start addr and len of the remaining segment */
 773                                                 munmap(hp->final_va, (size_t) unmap_len);
 774
 775                                                 hp->final_va = NULL;
 776                                                 if (unlink(hp->filepath) == -1) {
 777                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 778                                                                         __func__, hp->filepath, strerror(errno));
 779                                                         return -1;
 780                                                 }
 781                                         } else {
 782                                                 /* lock the page and skip */
 783                                                 pages_found++;
 784                                         }
 785
 786                                 } /* match page */
 787                         } /* foreach page */
 788                 } /* foreach socket */
 789         } /* foreach pagesize */
 790
 791         return 0;
 792 }
 793
 794 static inline uint64_t
 795 get_socket_mem_size(int socket)
 796 {
 797         uint64_t size = 0;
 798         unsigned i;
 799
 800         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 801                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 802                 if (hpi->hugedir != NULL)
 803                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 804         }
 805
 806         return size;
 807 }
 808
 809 /*
 810  * This function is a NUMA-aware equivalent of calc_num_pages.
 811  * It takes in the list of hugepage sizes and the
 812  * number of pages thereof, and calculates the best number of
 813  * pages of each size to fulfill the request for <memory> ram
 814  */
 815 static int
 816 calc_num_pages_per_socket(uint64_t * memory,
 817                 struct hugepage_info *hp_info,
 818                 struct hugepage_info *hp_used,
 819                 unsigned num_hp_info)
 820 {
 821         unsigned socket, j, i = 0;
 822         unsigned requested, available;
 823         int total_num_pages = 0;
 824         uint64_t remaining_mem, cur_mem;
 825         uint64_t total_mem = internal_config.memory;
 826
 827         if (num_hp_info == 0)
 828                 return -1;
 829
 830         /* if specific memory amounts per socket weren't requested */
 831         if (internal_config.force_sockets == 0) {
 832                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
 833                 size_t default_size, total_size;
 834                 unsigned lcore_id;
 835
 836                 /* Compute number of cores per socket */
 837                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
 838                 RTE_LCORE_FOREACH(lcore_id) {
 839                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
 840                 }
 841
 842                 /*
 843                  * Automatically spread requested memory amongst detected sockets according
 844                  * to number of cores from cpu mask present on each socket
 845                  */
 846                 total_size = internal_config.memory;
 847                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 848
 849                         /* Set memory amount per socket */
 850                         default_size = (internal_config.memory * cpu_per_socket[socket])
 851                                         / rte_lcore_count();
 852
 853                         /* Limit to maximum available memory on socket */
 854                         default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
 855
 856                         /* Update sizes */
 857                         memory[socket] = default_size;
 858                         total_size -= default_size;
 859                 }
 860
 861                 /*
 862                  * If some memory is remaining, try to allocate it by getting all
 863                  * available memory from sockets, one after the other
 864                  */
 865                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 866                         /* take whatever is available */
 867                         default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
 868                                                total_size);
 869
 870                         /* Update sizes */
 871                         memory[socket] += default_size;
 872                         total_size -= default_size;
 873                 }
 874         }
 875
 876         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 877                 /* skips if the memory on specific socket wasn't requested */
 878                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 879                         hp_used[i].hugedir = hp_info[i].hugedir;
 880                         hp_used[i].num_pages[socket] = RTE_MIN(
 881                                         memory[socket] / hp_info[i].hugepage_sz,
 882                                         hp_info[i].num_pages[socket]);
 883
 884                         cur_mem = hp_used[i].num_pages[socket] *
 885                                         hp_used[i].hugepage_sz;
 886
 887                         memory[socket] -= cur_mem;
 888                         total_mem -= cur_mem;
 889
 890                         total_num_pages += hp_used[i].num_pages[socket];
 891
 892                         /* check if we have met all memory requests */
 893                         if (memory[socket] == 0)
 894                                 break;
 895
 896                         /* check if we have any more pages left at this size, if so
 897                          * move on to next size */
 898                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
 899                                 continue;
 900                         /* At this point we know that there are more pages available that are
 901                          * bigger than the memory we want, so lets see if we can get enough
 902                          * from other page sizes.
 903                          */
 904                         remaining_mem = 0;
 905                         for (j = i+1; j < num_hp_info; j++)
 906                                 remaining_mem += hp_info[j].hugepage_sz *
 907                                 hp_info[j].num_pages[socket];
 908
 909                         /* is there enough other memory, if not allocate another page and quit */
 910                         if (remaining_mem < memory[socket]){
 911                                 cur_mem = RTE_MIN(memory[socket],
 912                                                 hp_info[i].hugepage_sz);
 913                                 memory[socket] -= cur_mem;
 914                                 total_mem -= cur_mem;
 915                                 hp_used[i].num_pages[socket]++;
 916                                 total_num_pages++;
 917                                 break; /* we are done with this socket*/
 918                         }
 919                 }
 920                 /* if we didn't satisfy all memory requirements per socket */
 921                 if (memory[socket] > 0) {
 922                         /* to prevent icc errors */
 923                         requested = (unsigned) (internal_config.socket_mem[socket] /
 924                                         0x100000);
 925                         available = requested -
 926                                         ((unsigned) (memory[socket] / 0x100000));
 927                         RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
 928                                         "Requested: %uMB, available: %uMB\n", socket,
 929                                         requested, available);
 930                         return -1;
 931                 }
 932         }
 933
 934         /* if we didn't satisfy total memory requirements */
 935         if (total_mem > 0) {
 936                 requested = (unsigned) (internal_config.memory / 0x100000);
 937                 available = requested - (unsigned) (total_mem / 0x100000);
 938                 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
 939                                 " available: %uMB\n", requested, available);
 940                 return -1;
 941         }
 942         return total_num_pages;
 943 }
 944
 945 static inline size_t
 946 eal_get_hugepage_mem_size(void)
 947 {
 948         uint64_t size = 0;
 949         unsigned i, j;
 950
 951         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 952                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 953                 if (hpi->hugedir != NULL) {
 954                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 955                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 956                         }
 957                 }
 958         }
 959
 960         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 961 }
 962
 963 static struct sigaction huge_action_old;
 964 static int huge_need_recover;
 965
 966 static void
 967 huge_register_sigbus(void)
 968 {
 969         sigset_t mask;
 970         struct sigaction action;
 971
 972         sigemptyset(&mask);
 973         sigaddset(&mask, SIGBUS);
 974         action.sa_flags = 0;
 975         action.sa_mask = mask;
 976         action.sa_handler = huge_sigbus_handler;
 977
 978         huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
 979 }
 980
 981 static void
 982 huge_recover_sigbus(void)
 983 {
 984         if (huge_need_recover) {
 985                 sigaction(SIGBUS, &huge_action_old, NULL);
 986                 huge_need_recover = 0;
 987         }
 988 }
 989
 990 /*
 991  * Prepare physical memory mapping: fill configuration structure with
 992  * these infos, return 0 on success.
 993  *  1. map N huge pages in separate files in hugetlbfs
 994  *  2. find associated physical addr
 995  *  3. find associated NUMA socket ID
 996  *  4. sort all huge pages by physical address
 997  *  5. remap these N huge pages in the correct order
 998  *  6. unmap the first mapping
 999  *  7. fill memsegs in configuration with contiguous zones
1000  */
1001 int
1002 rte_eal_hugepage_init(void)
1003 {
1004         struct rte_mem_config *mcfg;
1005         struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1006         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1007
1008         uint64_t memory[RTE_MAX_NUMA_NODES];
1009
1010         unsigned hp_offset;
1011         int i, j, new_memseg;
1012         int nr_hugefiles, nr_hugepages = 0;
1013         void *addr;
1014
1015         test_phys_addrs_available();
1016
1017         memset(used_hp, 0, sizeof(used_hp));
1018
1019         /* get pointer to global configuration */
1020         mcfg = rte_eal_get_configuration()->mem_config;
1021
1022         /* hugetlbfs can be disabled */
1023         if (internal_config.no_hugetlbfs) {
1024                 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1025                                 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1026                 if (addr == MAP_FAILED) {
1027                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1028                                         strerror(errno));
1029                         return -1;
1030                 }
1031                 mcfg->memseg[0].phys_addr = RTE_BAD_PHYS_ADDR;
1032                 mcfg->memseg[0].addr = addr;
1033                 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1034                 mcfg->memseg[0].len = internal_config.memory;
1035                 mcfg->memseg[0].socket_id = 0;
1036                 return 0;
1037         }
1038
1039         /* calculate total number of hugepages available. at this point we haven't
1040          * yet started sorting them so they all are on socket 0 */
1041         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1042                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1043                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1044
1045                 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1046         }
1047
1048         /*
1049          * allocate a memory area for hugepage table.
1050          * this isn't shared memory yet. due to the fact that we need some
1051          * processing done on these pages, shared memory will be created
1052          * at a later stage.
1053          */
1054         tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1055         if (tmp_hp == NULL)
1056                 goto fail;
1057
1058         memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1059
1060         hp_offset = 0; /* where we start the current page size entries */
1061
1062         huge_register_sigbus();
1063
1064         /* make a copy of socket_mem, needed for balanced allocation. */
1065         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1066                 memory[i] = internal_config.socket_mem[i];
1067
1068
1069         /* map all hugepages and sort them */
1070         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1071                 unsigned pages_old, pages_new;
1072                 struct hugepage_info *hpi;
1073
1074                 /*
1075                  * we don't yet mark hugepages as used at this stage, so
1076                  * we just map all hugepages available to the system
1077                  * all hugepages are still located on socket 0
1078                  */
1079                 hpi = &internal_config.hugepage_info[i];
1080
1081                 if (hpi->num_pages[0] == 0)
1082                         continue;
1083
1084                 /* map all hugepages available */
1085                 pages_old = hpi->num_pages[0];
1086                 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi,
1087                                               memory, 1);
1088                 if (pages_new < pages_old) {
1089                         RTE_LOG(DEBUG, EAL,
1090                                 "%d not %d hugepages of size %u MB allocated\n",
1091                                 pages_new, pages_old,
1092                                 (unsigned)(hpi->hugepage_sz / 0x100000));
1093
1094                         int pages = pages_old - pages_new;
1095
1096                         nr_hugepages -= pages;
1097                         hpi->num_pages[0] = pages_new;
1098                         if (pages_new == 0)
1099                                 continue;
1100                 }
1101
1102                 if (phys_addrs_available) {
1103                         /* find physical addresses for each hugepage */
1104                         if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1105                                 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1106                                         "for %u MB pages\n",
1107                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1108                                 goto fail;
1109                         }
1110                 } else {
1111                         /* set physical addresses for each hugepage */
1112                         if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1113                                 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1114                                         "for %u MB pages\n",
1115                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1116                                 goto fail;
1117                         }
1118                 }
1119
1120                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1121                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1122                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1123                         goto fail;
1124                 }
1125
1126                 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1127                       sizeof(struct hugepage_file), cmp_physaddr);
1128
1129                 /* remap all hugepages */
1130                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, NULL, 0) !=
1131                     hpi->num_pages[0]) {
1132                         RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1133                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1134                         goto fail;
1135                 }
1136
1137                 /* unmap original mappings */
1138                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1139                         goto fail;
1140
1141                 /* we have processed a num of hugepages of this size, so inc offset */
1142                 hp_offset += hpi->num_pages[0];
1143         }
1144
1145         huge_recover_sigbus();
1146
1147         if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1148                 internal_config.memory = eal_get_hugepage_mem_size();
1149
1150         nr_hugefiles = nr_hugepages;
1151
1152
1153         /* clean out the numbers of pages */
1154         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1155                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1156                         internal_config.hugepage_info[i].num_pages[j] = 0;
1157
1158         /* get hugepages for each socket */
1159         for (i = 0; i < nr_hugefiles; i++) {
1160                 int socket = tmp_hp[i].socket_id;
1161
1162                 /* find a hugepage info with right size and increment num_pages */
1163                 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1164                                 (int)internal_config.num_hugepage_sizes);
1165                 for (j = 0; j < nb_hpsizes; j++) {
1166                         if (tmp_hp[i].size ==
1167                                         internal_config.hugepage_info[j].hugepage_sz) {
1168                                 internal_config.hugepage_info[j].num_pages[socket]++;
1169                         }
1170                 }
1171         }
1172
1173         /* make a copy of socket_mem, needed for number of pages calculation */
1174         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1175                 memory[i] = internal_config.socket_mem[i];
1176
1177         /* calculate final number of pages */
1178         nr_hugepages = calc_num_pages_per_socket(memory,
1179                         internal_config.hugepage_info, used_hp,
1180                         internal_config.num_hugepage_sizes);
1181
1182         /* error if not enough memory available */
1183         if (nr_hugepages < 0)
1184                 goto fail;
1185
1186         /* reporting in! */
1187         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1188                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1189                         if (used_hp[i].num_pages[j] > 0) {
1190                                 RTE_LOG(DEBUG, EAL,
1191                                         "Requesting %u pages of size %uMB"
1192                                         " from socket %i\n",
1193                                         used_hp[i].num_pages[j],
1194                                         (unsigned)
1195                                         (used_hp[i].hugepage_sz / 0x100000),
1196                                         j);
1197                         }
1198                 }
1199         }
1200
1201         /* create shared memory */
1202         hugepage = create_shared_memory(eal_hugepage_info_path(),
1203                         nr_hugefiles * sizeof(struct hugepage_file));
1204
1205         if (hugepage == NULL) {
1206                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1207                 goto fail;
1208         }
1209         memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1210
1211         /*
1212          * unmap pages that we won't need (looks at used_hp).
1213          * also, sets final_va to NULL on pages that were unmapped.
1214          */
1215         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1216                         internal_config.num_hugepage_sizes) < 0) {
1217                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1218                 goto fail;
1219         }
1220
1221         /*
1222          * copy stuff from malloc'd hugepage* to the actual shared memory.
1223          * this procedure only copies those hugepages that have final_va
1224          * not NULL. has overflow protection.
1225          */
1226         if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1227                         tmp_hp, nr_hugefiles) < 0) {
1228                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1229                 goto fail;
1230         }
1231
1232         /* free the hugepage backing files */
1233         if (internal_config.hugepage_unlink &&
1234                 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1235                 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1236                 goto fail;
1237         }
1238
1239         /* free the temporary hugepage table */
1240         free(tmp_hp);
1241         tmp_hp = NULL;
1242
1243         /* first memseg index shall be 0 after incrementing it below */
1244         j = -1;
1245         for (i = 0; i < nr_hugefiles; i++) {
1246                 new_memseg = 0;
1247
1248                 /* if this is a new section, create a new memseg */
1249                 if (i == 0)
1250                         new_memseg = 1;
1251                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1252                         new_memseg = 1;
1253                 else if (hugepage[i].size != hugepage[i-1].size)
1254                         new_memseg = 1;
1255
1256 #ifdef RTE_ARCH_PPC_64
1257                 /* On PPC64 architecture, the mmap always start from higher
1258                  * virtual address to lower address. Here, both the physical
1259                  * address and virtual address are in descending order */
1260                 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1261                     hugepage[i].size)
1262                         new_memseg = 1;
1263                 else if (((unsigned long)hugepage[i-1].final_va -
1264                     (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1265                         new_memseg = 1;
1266 #else
1267                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1268                     hugepage[i].size)
1269                         new_memseg = 1;
1270                 else if (((unsigned long)hugepage[i].final_va -
1271                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1272                         new_memseg = 1;
1273 #endif
1274
1275                 if (new_memseg) {
1276                         j += 1;
1277                         if (j == RTE_MAX_MEMSEG)
1278                                 break;
1279
1280                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1281                         mcfg->memseg[j].addr = hugepage[i].final_va;
1282                         mcfg->memseg[j].len = hugepage[i].size;
1283                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1284                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1285                 }
1286                 /* continuation of previous memseg */
1287                 else {
1288 #ifdef RTE_ARCH_PPC_64
1289                 /* Use the phy and virt address of the last page as segment
1290                  * address for IBM Power architecture */
1291                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1292                         mcfg->memseg[j].addr = hugepage[i].final_va;
1293 #endif
1294                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1295                 }
1296                 hugepage[i].memseg_id = j;
1297         }
1298
1299         if (i < nr_hugefiles) {
1300                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1301                         "from %d requested\n"
1302                         "Current %s=%d is not enough\n"
1303                         "Please either increase it or request less amount "
1304                         "of memory.\n",
1305                         i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1306                         RTE_MAX_MEMSEG);
1307                 goto fail;
1308         }
1309
1310         munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1311
1312         return 0;
1313
1314 fail:
1315         huge_recover_sigbus();
1316         free(tmp_hp);
1317         if (hugepage != NULL)
1318                 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1319
1320         return -1;
1321 }
1322
1323 /*
1324  * uses fstat to report the size of a file on disk
1325  */
1326 static off_t
1327 getFileSize(int fd)
1328 {
1329         struct stat st;
1330         if (fstat(fd, &st) < 0)
1331                 return 0;
1332         return st.st_size;
1333 }
1334
1335 /*
1336  * This creates the memory mappings in the secondary process to match that of
1337  * the server process. It goes through each memory segment in the DPDK runtime
1338  * configuration and finds the hugepages which form that segment, mapping them
1339  * in order to form a contiguous block in the virtual memory space
1340  */
1341 int
1342 rte_eal_hugepage_attach(void)
1343 {
1344         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1345         struct hugepage_file *hp = NULL;
1346         unsigned num_hp = 0;
1347         unsigned i, s = 0; /* s used to track the segment number */
1348         unsigned max_seg = RTE_MAX_MEMSEG;
1349         off_t size = 0;
1350         int fd, fd_zero = -1, fd_hugepage = -1;
1351
1352         if (aslr_enabled() > 0) {
1353                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1354                                 "(ASLR) is enabled in the kernel.\n");
1355                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1356                                 "into secondary processes\n");
1357         }
1358
1359         test_phys_addrs_available();
1360
1361         fd_zero = open("/dev/zero", O_RDONLY);
1362         if (fd_zero < 0) {
1363                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1364                 goto error;
1365         }
1366         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1367         if (fd_hugepage < 0) {
1368                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1369                 goto error;
1370         }
1371
1372         /* map all segments into memory to make sure we get the addrs */
1373         for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1374                 void *base_addr;
1375
1376                 /*
1377                  * the first memory segment with len==0 is the one that
1378                  * follows the last valid segment.
1379                  */
1380                 if (mcfg->memseg[s].len == 0)
1381                         break;
1382
1383                 /*
1384                  * fdzero is mmapped to get a contiguous block of virtual
1385                  * addresses of the appropriate memseg size.
1386                  * use mmap to get identical addresses as the primary process.
1387                  */
1388                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1389                                  PROT_READ,
1390 #ifdef RTE_ARCH_PPC_64
1391                                  MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
1392 #else
1393                                  MAP_PRIVATE,
1394 #endif
1395                                  fd_zero, 0);
1396                 if (base_addr == MAP_FAILED ||
1397                     base_addr != mcfg->memseg[s].addr) {
1398                         max_seg = s;
1399                         if (base_addr != MAP_FAILED) {
1400                                 /* errno is stale, don't use */
1401                                 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1402                                         "in /dev/zero at [%p], got [%p] - "
1403                                         "please use '--base-virtaddr' option\n",
1404                                         (unsigned long long)mcfg->memseg[s].len,
1405                                         mcfg->memseg[s].addr, base_addr);
1406                                 munmap(base_addr, mcfg->memseg[s].len);
1407                         } else {
1408                                 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1409                                         "in /dev/zero at [%p]: '%s'\n",
1410                                         (unsigned long long)mcfg->memseg[s].len,
1411                                         mcfg->memseg[s].addr, strerror(errno));
1412                         }
1413                         if (aslr_enabled() > 0) {
1414                                 RTE_LOG(ERR, EAL, "It is recommended to "
1415                                         "disable ASLR in the kernel "
1416                                         "and retry running both primary "
1417                                         "and secondary processes\n");
1418                         }
1419                         goto error;
1420                 }
1421         }
1422
1423         size = getFileSize(fd_hugepage);
1424         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1425         if (hp == MAP_FAILED) {
1426                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1427                 goto error;
1428         }
1429
1430         num_hp = size / sizeof(struct hugepage_file);
1431         RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1432
1433         s = 0;
1434         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1435                 void *addr, *base_addr;
1436                 uintptr_t offset = 0;
1437                 size_t mapping_size;
1438                 /*
1439                  * free previously mapped memory so we can map the
1440                  * hugepages into the space
1441                  */
1442                 base_addr = mcfg->memseg[s].addr;
1443                 munmap(base_addr, mcfg->memseg[s].len);
1444
1445                 /* find the hugepages for this segment and map them
1446                  * we don't need to worry about order, as the server sorted the
1447                  * entries before it did the second mmap of them */
1448                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1449                         if (hp[i].memseg_id == (int)s){
1450                                 fd = open(hp[i].filepath, O_RDWR);
1451                                 if (fd < 0) {
1452                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1453                                                 hp[i].filepath);
1454                                         goto error;
1455                                 }
1456                                 mapping_size = hp[i].size;
1457                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1458                                                 mapping_size, PROT_READ | PROT_WRITE,
1459                                                 MAP_SHARED, fd, 0);
1460                                 close(fd); /* close file both on success and on failure */
1461                                 if (addr == MAP_FAILED ||
1462                                                 addr != RTE_PTR_ADD(base_addr, offset)) {
1463                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1464                                                 hp[i].filepath);
1465                                         goto error;
1466                                 }
1467                                 offset+=mapping_size;
1468                         }
1469                 }
1470                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1471                                 (unsigned long long)mcfg->memseg[s].len);
1472                 s++;
1473         }
1474         /* unmap the hugepage config file, since we are done using it */
1475         munmap(hp, size);
1476         close(fd_zero);
1477         close(fd_hugepage);
1478         return 0;
1479
1480 error:
1481         for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1482                 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1483         if (hp != NULL && hp != MAP_FAILED)
1484                 munmap(hp, size);
1485         if (fd_zero >= 0)
1486                 close(fd_zero);
1487         if (fd_hugepage >= 0)
1488                 close(fd_hugepage);
1489         return -1;
1490 }
1491
1492 bool
1493 rte_eal_using_phys_addrs(void)
1494 {
1495         return phys_addrs_available;
1496 }