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_launch.h>
  63 #include <rte_eal.h>
  64 #include <rte_eal_memconfig.h>
  65 #include <rte_per_lcore.h>
  66 #include <rte_lcore.h>
  67 #include <rte_common.h>
  68 #include <rte_string_fns.h>
  69
  70 #include "eal_private.h"
  71 #include "eal_internal_cfg.h"
  72 #include "eal_filesystem.h"
  73 #include "eal_hugepages.h"
  74
  75 #define PFN_MASK_SIZE   8
  76
  77 /**
  78  * @file
  79  * Huge page mapping under linux
  80  *
  81  * To reserve a big contiguous amount of memory, we use the hugepage
  82  * feature of linux. For that, we need to have hugetlbfs mounted. This
  83  * code will create many files in this directory (one per page) and
  84  * map them in virtual memory. For each page, we will retrieve its
  85  * physical address and remap it in order to have a virtual contiguous
  86  * zone as well as a physical contiguous zone.
  87  */
  88
  89 static uint64_t baseaddr_offset;
  90
  91 static bool phys_addrs_available = true;
  92
  93 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
  94
  95 static void
  96 test_phys_addrs_available(void)
  97 {
  98         uint64_t tmp;
  99         phys_addr_t physaddr;
 100
 101         if (!rte_eal_has_hugepages()) {
 102                 RTE_LOG(ERR, EAL,
 103                         "Started without hugepages support, physical addresses not available\n");
 104                 phys_addrs_available = false;
 105                 return;
 106         }
 107
 108         physaddr = rte_mem_virt2phy(&tmp);
 109         if (physaddr == RTE_BAD_PHYS_ADDR) {
 110                 RTE_LOG(ERR, EAL,
 111                         "Cannot obtain physical addresses: %s. "
 112                         "Only vfio will function.\n",
 113                         strerror(errno));
 114                 phys_addrs_available = false;
 115         }
 116 }
 117
 118 /*
 119  * Get physical address of any mapped virtual address in the current process.
 120  */
 121 phys_addr_t
 122 rte_mem_virt2phy(const void *virtaddr)
 123 {
 124         int fd, retval;
 125         uint64_t page, physaddr;
 126         unsigned long virt_pfn;
 127         int page_size;
 128         off_t offset;
 129
 130         if (rte_eal_iova_mode() == RTE_IOVA_VA)
 131                 return (uintptr_t)virtaddr;
 132
 133         /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
 134         if (!phys_addrs_available)
 135                 return RTE_BAD_PHYS_ADDR;
 136
 137         /* standard page size */
 138         page_size = getpagesize();
 139
 140         fd = open("/proc/self/pagemap", O_RDONLY);
 141         if (fd < 0) {
 142                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 143                         __func__, strerror(errno));
 144                 return RTE_BAD_PHYS_ADDR;
 145         }
 146
 147         virt_pfn = (unsigned long)virtaddr / page_size;
 148         offset = sizeof(uint64_t) * virt_pfn;
 149         if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 150                 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 151                                 __func__, strerror(errno));
 152                 close(fd);
 153                 return RTE_BAD_PHYS_ADDR;
 154         }
 155
 156         retval = read(fd, &page, PFN_MASK_SIZE);
 157         close(fd);
 158         if (retval < 0) {
 159                 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 160                                 __func__, strerror(errno));
 161                 return RTE_BAD_PHYS_ADDR;
 162         } else if (retval != PFN_MASK_SIZE) {
 163                 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
 164                                 "but expected %d:\n",
 165                                 __func__, retval, PFN_MASK_SIZE);
 166                 return RTE_BAD_PHYS_ADDR;
 167         }
 168
 169         /*
 170          * the pfn (page frame number) are bits 0-54 (see
 171          * pagemap.txt in linux Documentation)
 172          */
 173         if ((page & 0x7fffffffffffffULL) == 0)
 174                 return RTE_BAD_PHYS_ADDR;
 175
 176         physaddr = ((page & 0x7fffffffffffffULL) * page_size)
 177                 + ((unsigned long)virtaddr % page_size);
 178
 179         return physaddr;
 180 }
 181
 182 /*
 183  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 184  * it by browsing the /proc/self/pagemap special file.
 185  */
 186 static int
 187 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 188 {
 189         unsigned int i;
 190         phys_addr_t addr;
 191
 192         for (i = 0; i < hpi->num_pages[0]; i++) {
 193                 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
 194                 if (addr == RTE_BAD_PHYS_ADDR)
 195                         return -1;
 196                 hugepg_tbl[i].physaddr = addr;
 197         }
 198         return 0;
 199 }
 200
 201 /*
 202  * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
 203  */
 204 static int
 205 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 206 {
 207         unsigned int i;
 208         static phys_addr_t addr;
 209
 210         for (i = 0; i < hpi->num_pages[0]; i++) {
 211                 hugepg_tbl[i].physaddr = addr;
 212                 addr += hugepg_tbl[i].size;
 213         }
 214         return 0;
 215 }
 216
 217 /*
 218  * Check whether address-space layout randomization is enabled in
 219  * the kernel. This is important for multi-process as it can prevent
 220  * two processes mapping data to the same virtual address
 221  * Returns:
 222  *    0 - address space randomization disabled
 223  *    1/2 - address space randomization enabled
 224  *    negative error code on error
 225  */
 226 static int
 227 aslr_enabled(void)
 228 {
 229         char c;
 230         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 231         if (fd < 0)
 232                 return -errno;
 233         retval = read(fd, &c, 1);
 234         close(fd);
 235         if (retval < 0)
 236                 return -errno;
 237         if (retval == 0)
 238                 return -EIO;
 239         switch (c) {
 240                 case '0' : return 0;
 241                 case '1' : return 1;
 242                 case '2' : return 2;
 243                 default: return -EINVAL;
 244         }
 245 }
 246
 247 /*
 248  * Try to mmap *size bytes in /dev/zero. If it is successful, return the
 249  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 250  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 251  * 0. In this case, return NULL. Note: this function returns an address
 252  * which is a multiple of hugepage size.
 253  */
 254 static void *
 255 get_virtual_area(size_t *size, size_t hugepage_sz)
 256 {
 257         void *addr;
 258         int fd;
 259         long aligned_addr;
 260
 261         if (internal_config.base_virtaddr != 0) {
 262                 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
 263                                 baseaddr_offset);
 264         }
 265         else addr = NULL;
 266
 267         RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
 268
 269         fd = open("/dev/zero", O_RDONLY);
 270         if (fd < 0){
 271                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 272                 return NULL;
 273         }
 274         do {
 275                 addr = mmap(addr,
 276                                 (*size) + hugepage_sz, PROT_READ,
 277 #ifdef RTE_ARCH_PPC_64
 278                                 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
 279 #else
 280                                 MAP_PRIVATE,
 281 #endif
 282                                 fd, 0);
 283                 if (addr == MAP_FAILED)
 284                         *size -= hugepage_sz;
 285         } while (addr == MAP_FAILED && *size > 0);
 286
 287         if (addr == MAP_FAILED) {
 288                 close(fd);
 289                 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
 290                         strerror(errno));
 291                 return NULL;
 292         }
 293
 294         munmap(addr, (*size) + hugepage_sz);
 295         close(fd);
 296
 297         /* align addr to a huge page size boundary */
 298         aligned_addr = (long)addr;
 299         aligned_addr += (hugepage_sz - 1);
 300         aligned_addr &= (~(hugepage_sz - 1));
 301         addr = (void *)(aligned_addr);
 302
 303         RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
 304                 addr, *size);
 305
 306         /* increment offset */
 307         baseaddr_offset += *size;
 308
 309         return addr;
 310 }
 311
 312 static sigjmp_buf huge_jmpenv;
 313
 314 static void huge_sigbus_handler(int signo __rte_unused)
 315 {
 316         siglongjmp(huge_jmpenv, 1);
 317 }
 318
 319 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
 320  * non-static local variable in the stack frame calling sigsetjmp might be
 321  * clobbered by a call to longjmp.
 322  */
 323 static int huge_wrap_sigsetjmp(void)
 324 {
 325         return sigsetjmp(huge_jmpenv, 1);
 326 }
 327
 328 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 329 /* Callback for numa library. */
 330 void numa_error(char *where)
 331 {
 332         RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
 333 }
 334 #endif
 335
 336 /*
 337  * Mmap all hugepages of hugepage table: it first open a file in
 338  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 339  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 340  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 341  * map continguous physical blocks in contiguous virtual blocks.
 342  */
 343 static unsigned
 344 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
 345                   uint64_t *essential_memory __rte_unused, int orig)
 346 {
 347         int fd;
 348         unsigned i;
 349         void *virtaddr;
 350         void *vma_addr = NULL;
 351         size_t vma_len = 0;
 352 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 353         int node_id = -1;
 354         int essential_prev = 0;
 355         int oldpolicy;
 356         struct bitmask *oldmask = numa_allocate_nodemask();
 357         bool have_numa = true;
 358         unsigned long maxnode = 0;
 359
 360         /* Check if kernel supports NUMA. */
 361         if (numa_available() != 0) {
 362                 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
 363                 have_numa = false;
 364         }
 365
 366         if (orig && have_numa) {
 367                 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
 368                 if (get_mempolicy(&oldpolicy, oldmask->maskp,
 369                                   oldmask->size + 1, 0, 0) < 0) {
 370                         RTE_LOG(ERR, EAL,
 371                                 "Failed to get current mempolicy: %s. "
 372                                 "Assuming MPOL_DEFAULT.\n", strerror(errno));
 373                         oldpolicy = MPOL_DEFAULT;
 374                 }
 375                 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 376                         if (internal_config.socket_mem[i])
 377                                 maxnode = i + 1;
 378         }
 379 #endif
 380
 381         for (i = 0; i < hpi->num_pages[0]; i++) {
 382                 uint64_t hugepage_sz = hpi->hugepage_sz;
 383
 384 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 385                 if (maxnode) {
 386                         unsigned int j;
 387
 388                         for (j = 0; j < maxnode; j++)
 389                                 if (essential_memory[j])
 390                                         break;
 391
 392                         if (j == maxnode) {
 393                                 node_id = (node_id + 1) % maxnode;
 394                                 while (!internal_config.socket_mem[node_id]) {
 395                                         node_id++;
 396                                         node_id %= maxnode;
 397                                 }
 398                                 essential_prev = 0;
 399                         } else {
 400                                 node_id = j;
 401                                 essential_prev = essential_memory[j];
 402
 403                                 if (essential_memory[j] < hugepage_sz)
 404                                         essential_memory[j] = 0;
 405                                 else
 406                                         essential_memory[j] -= hugepage_sz;
 407                         }
 408
 409                         RTE_LOG(DEBUG, EAL,
 410                                 "Setting policy MPOL_PREFERRED for socket %d\n",
 411                                 node_id);
 412                         numa_set_preferred(node_id);
 413                 }
 414 #endif
 415
 416                 if (orig) {
 417                         hugepg_tbl[i].file_id = i;
 418                         hugepg_tbl[i].size = hugepage_sz;
 419                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 420                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 421                                         hugepg_tbl[i].file_id);
 422                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 423                 }
 424 #ifndef RTE_ARCH_64
 425                 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
 426                  * original map address as final map address.
 427                  */
 428                 else if ((hugepage_sz == RTE_PGSIZE_1G)
 429                         || (hugepage_sz == RTE_PGSIZE_16G)) {
 430                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 431                         hugepg_tbl[i].orig_va = NULL;
 432                         continue;
 433                 }
 434 #endif
 435                 else if (vma_len == 0) {
 436                         unsigned j, num_pages;
 437
 438                         /* reserve a virtual area for next contiguous
 439                          * physical block: count the number of
 440                          * contiguous physical pages. */
 441                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 442 #ifdef RTE_ARCH_PPC_64
 443                                 /* The physical addresses are sorted in
 444                                  * descending order on PPC64 */
 445                                 if (hugepg_tbl[j].physaddr !=
 446                                     hugepg_tbl[j-1].physaddr - hugepage_sz)
 447                                         break;
 448 #else
 449                                 if (hugepg_tbl[j].physaddr !=
 450                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 451                                         break;
 452 #endif
 453                         }
 454                         num_pages = j - i;
 455                         vma_len = num_pages * hugepage_sz;
 456
 457                         /* get the biggest virtual memory area up to
 458                          * vma_len. If it fails, vma_addr is NULL, so
 459                          * let the kernel provide the address. */
 460                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 461                         if (vma_addr == NULL)
 462                                 vma_len = hugepage_sz;
 463                 }
 464
 465                 /* try to create hugepage file */
 466                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
 467                 if (fd < 0) {
 468                         RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
 469                                         strerror(errno));
 470                         goto out;
 471                 }
 472
 473                 /* map the segment, and populate page tables,
 474                  * the kernel fills this segment with zeros */
 475                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 476                                 MAP_SHARED | MAP_POPULATE, fd, 0);
 477                 if (virtaddr == MAP_FAILED) {
 478                         RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
 479                                         strerror(errno));
 480                         close(fd);
 481                         goto out;
 482                 }
 483
 484                 if (orig) {
 485                         hugepg_tbl[i].orig_va = virtaddr;
 486                 }
 487                 else {
 488                         hugepg_tbl[i].final_va = virtaddr;
 489                 }
 490
 491                 if (orig) {
 492                         /* In linux, hugetlb limitations, like cgroup, are
 493                          * enforced at fault time instead of mmap(), even
 494                          * with the option of MAP_POPULATE. Kernel will send
 495                          * a SIGBUS signal. To avoid to be killed, save stack
 496                          * environment here, if SIGBUS happens, we can jump
 497                          * back here.
 498                          */
 499                         if (huge_wrap_sigsetjmp()) {
 500                                 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
 501                                         "hugepages of size %u MB\n",
 502                                         (unsigned)(hugepage_sz / 0x100000));
 503                                 munmap(virtaddr, hugepage_sz);
 504                                 close(fd);
 505                                 unlink(hugepg_tbl[i].filepath);
 506 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 507                                 if (maxnode)
 508                                         essential_memory[node_id] =
 509                                                 essential_prev;
 510 #endif
 511                                 goto out;
 512                         }
 513                         *(int *)virtaddr = 0;
 514                 }
 515
 516
 517                 /* set shared flock on the file. */
 518                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 519                         RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
 520                                 __func__, strerror(errno));
 521                         close(fd);
 522                         goto out;
 523                 }
 524
 525                 close(fd);
 526
 527                 vma_addr = (char *)vma_addr + hugepage_sz;
 528                 vma_len -= hugepage_sz;
 529         }
 530
 531 out:
 532 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 533         if (maxnode) {
 534                 RTE_LOG(DEBUG, EAL,
 535                         "Restoring previous memory policy: %d\n", oldpolicy);
 536                 if (oldpolicy == MPOL_DEFAULT) {
 537                         numa_set_localalloc();
 538                 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
 539                                          oldmask->size + 1) < 0) {
 540                         RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
 541                                 strerror(errno));
 542                         numa_set_localalloc();
 543                 }
 544         }
 545         numa_free_cpumask(oldmask);
 546 #endif
 547         return i;
 548 }
 549
 550 /* Unmap all hugepages from original mapping */
 551 static int
 552 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 553 {
 554         unsigned i;
 555         for (i = 0; i < hpi->num_pages[0]; i++) {
 556                 if (hugepg_tbl[i].orig_va) {
 557                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 558                         hugepg_tbl[i].orig_va = NULL;
 559                 }
 560         }
 561         return 0;
 562 }
 563
 564 /*
 565  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 566  * page.
 567  */
 568 static int
 569 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 570 {
 571         int socket_id;
 572         char *end, *nodestr;
 573         unsigned i, hp_count = 0;
 574         uint64_t virt_addr;
 575         char buf[BUFSIZ];
 576         char hugedir_str[PATH_MAX];
 577         FILE *f;
 578
 579         f = fopen("/proc/self/numa_maps", "r");
 580         if (f == NULL) {
 581                 RTE_LOG(NOTICE, EAL, "NUMA support not available"
 582                         " consider that all memory is in socket_id 0\n");
 583                 return 0;
 584         }
 585
 586         snprintf(hugedir_str, sizeof(hugedir_str),
 587                         "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
 588
 589         /* parse numa map */
 590         while (fgets(buf, sizeof(buf), f) != NULL) {
 591
 592                 /* ignore non huge page */
 593                 if (strstr(buf, " huge ") == NULL &&
 594                                 strstr(buf, hugedir_str) == NULL)
 595                         continue;
 596
 597                 /* get zone addr */
 598                 virt_addr = strtoull(buf, &end, 16);
 599                 if (virt_addr == 0 || end == buf) {
 600                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 601                         goto error;
 602                 }
 603
 604                 /* get node id (socket id) */
 605                 nodestr = strstr(buf, " N");
 606                 if (nodestr == NULL) {
 607                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 608                         goto error;
 609                 }
 610                 nodestr += 2;
 611                 end = strstr(nodestr, "=");
 612                 if (end == NULL) {
 613                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 614                         goto error;
 615                 }
 616                 end[0] = '\0';
 617                 end = NULL;
 618
 619                 socket_id = strtoul(nodestr, &end, 0);
 620                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 621                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 622                         goto error;
 623                 }
 624
 625                 /* if we find this page in our mappings, set socket_id */
 626                 for (i = 0; i < hpi->num_pages[0]; i++) {
 627                         void *va = (void *)(unsigned long)virt_addr;
 628                         if (hugepg_tbl[i].orig_va == va) {
 629                                 hugepg_tbl[i].socket_id = socket_id;
 630                                 hp_count++;
 631 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 632                                 RTE_LOG(DEBUG, EAL,
 633                                         "Hugepage %s is on socket %d\n",
 634                                         hugepg_tbl[i].filepath, socket_id);
 635 #endif
 636                         }
 637                 }
 638         }
 639
 640         if (hp_count < hpi->num_pages[0])
 641                 goto error;
 642
 643         fclose(f);
 644         return 0;
 645
 646 error:
 647         fclose(f);
 648         return -1;
 649 }
 650
 651 static int
 652 cmp_physaddr(const void *a, const void *b)
 653 {
 654 #ifndef RTE_ARCH_PPC_64
 655         const struct hugepage_file *p1 = a;
 656         const struct hugepage_file *p2 = b;
 657 #else
 658         /* PowerPC needs memory sorted in reverse order from x86 */
 659         const struct hugepage_file *p1 = b;
 660         const struct hugepage_file *p2 = a;
 661 #endif
 662         if (p1->physaddr < p2->physaddr)
 663                 return -1;
 664         else if (p1->physaddr > p2->physaddr)
 665                 return 1;
 666         else
 667                 return 0;
 668 }
 669
 670 /*
 671  * Uses mmap to create a shared memory area for storage of data
 672  * Used in this file to store the hugepage file map on disk
 673  */
 674 static void *
 675 create_shared_memory(const char *filename, const size_t mem_size)
 676 {
 677         void *retval;
 678         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 679         if (fd < 0)
 680                 return NULL;
 681         if (ftruncate(fd, mem_size) < 0) {
 682                 close(fd);
 683                 return NULL;
 684         }
 685         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 686         close(fd);
 687         if (retval == MAP_FAILED)
 688                 return NULL;
 689         return retval;
 690 }
 691
 692 /*
 693  * this copies *active* hugepages from one hugepage table to another.
 694  * destination is typically the shared memory.
 695  */
 696 static int
 697 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
 698                 const struct hugepage_file * src, int src_size)
 699 {
 700         int src_pos, dst_pos = 0;
 701
 702         for (src_pos = 0; src_pos < src_size; src_pos++) {
 703                 if (src[src_pos].final_va != NULL) {
 704                         /* error on overflow attempt */
 705                         if (dst_pos == dest_size)
 706                                 return -1;
 707                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
 708                         dst_pos++;
 709                 }
 710         }
 711         return 0;
 712 }
 713
 714 static int
 715 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
 716                 unsigned num_hp_info)
 717 {
 718         unsigned socket, size;
 719         int page, nrpages = 0;
 720
 721         /* get total number of hugepages */
 722         for (size = 0; size < num_hp_info; size++)
 723                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 724                         nrpages +=
 725                         internal_config.hugepage_info[size].num_pages[socket];
 726
 727         for (page = 0; page < nrpages; page++) {
 728                 struct hugepage_file *hp = &hugepg_tbl[page];
 729
 730                 if (hp->final_va != NULL && unlink(hp->filepath)) {
 731                         RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
 732                                 __func__, hp->filepath, strerror(errno));
 733                 }
 734         }
 735         return 0;
 736 }
 737
 738 /*
 739  * unmaps hugepages that are not going to be used. since we originally allocate
 740  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 741  */
 742 static int
 743 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
 744                 struct hugepage_info *hpi,
 745                 unsigned num_hp_info)
 746 {
 747         unsigned socket, size;
 748         int page, nrpages = 0;
 749
 750         /* get total number of hugepages */
 751         for (size = 0; size < num_hp_info; size++)
 752                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 753                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 754
 755         for (size = 0; size < num_hp_info; size++) {
 756                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 757                         unsigned pages_found = 0;
 758
 759                         /* traverse until we have unmapped all the unused pages */
 760                         for (page = 0; page < nrpages; page++) {
 761                                 struct hugepage_file *hp = &hugepg_tbl[page];
 762
 763                                 /* find a page that matches the criteria */
 764                                 if ((hp->size == hpi[size].hugepage_sz) &&
 765                                                 (hp->socket_id == (int) socket)) {
 766
 767                                         /* if we skipped enough pages, unmap the rest */
 768                                         if (pages_found == hpi[size].num_pages[socket]) {
 769                                                 uint64_t unmap_len;
 770
 771                                                 unmap_len = hp->size;
 772
 773                                                 /* get start addr and len of the remaining segment */
 774                                                 munmap(hp->final_va, (size_t) unmap_len);
 775
 776                                                 hp->final_va = NULL;
 777                                                 if (unlink(hp->filepath) == -1) {
 778                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 779                                                                         __func__, hp->filepath, strerror(errno));
 780                                                         return -1;
 781                                                 }
 782                                         } else {
 783                                                 /* lock the page and skip */
 784                                                 pages_found++;
 785                                         }
 786
 787                                 } /* match page */
 788                         } /* foreach page */
 789                 } /* foreach socket */
 790         } /* foreach pagesize */
 791
 792         return 0;
 793 }
 794
 795 static inline uint64_t
 796 get_socket_mem_size(int socket)
 797 {
 798         uint64_t size = 0;
 799         unsigned i;
 800
 801         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 802                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 803                 if (hpi->hugedir != NULL)
 804                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 805         }
 806
 807         return size;
 808 }
 809
 810 /*
 811  * This function is a NUMA-aware equivalent of calc_num_pages.
 812  * It takes in the list of hugepage sizes and the
 813  * number of pages thereof, and calculates the best number of
 814  * pages of each size to fulfill the request for <memory> ram
 815  */
 816 static int
 817 calc_num_pages_per_socket(uint64_t * memory,
 818                 struct hugepage_info *hp_info,
 819                 struct hugepage_info *hp_used,
 820                 unsigned num_hp_info)
 821 {
 822         unsigned socket, j, i = 0;
 823         unsigned requested, available;
 824         int total_num_pages = 0;
 825         uint64_t remaining_mem, cur_mem;
 826         uint64_t total_mem = internal_config.memory;
 827
 828         if (num_hp_info == 0)
 829                 return -1;
 830
 831         /* if specific memory amounts per socket weren't requested */
 832         if (internal_config.force_sockets == 0) {
 833                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
 834                 size_t default_size, total_size;
 835                 unsigned lcore_id;
 836
 837                 /* Compute number of cores per socket */
 838                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
 839                 RTE_LCORE_FOREACH(lcore_id) {
 840                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
 841                 }
 842
 843                 /*
 844                  * Automatically spread requested memory amongst detected sockets according
 845                  * to number of cores from cpu mask present on each socket
 846                  */
 847                 total_size = internal_config.memory;
 848                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 849
 850                         /* Set memory amount per socket */
 851                         default_size = (internal_config.memory * cpu_per_socket[socket])
 852                                         / rte_lcore_count();
 853
 854                         /* Limit to maximum available memory on socket */
 855                         default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
 856
 857                         /* Update sizes */
 858                         memory[socket] = default_size;
 859                         total_size -= default_size;
 860                 }
 861
 862                 /*
 863                  * If some memory is remaining, try to allocate it by getting all
 864                  * available memory from sockets, one after the other
 865                  */
 866                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 867                         /* take whatever is available */
 868                         default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
 869                                                total_size);
 870
 871                         /* Update sizes */
 872                         memory[socket] += default_size;
 873                         total_size -= default_size;
 874                 }
 875         }
 876
 877         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 878                 /* skips if the memory on specific socket wasn't requested */
 879                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 880                         hp_used[i].hugedir = hp_info[i].hugedir;
 881                         hp_used[i].num_pages[socket] = RTE_MIN(
 882                                         memory[socket] / hp_info[i].hugepage_sz,
 883                                         hp_info[i].num_pages[socket]);
 884
 885                         cur_mem = hp_used[i].num_pages[socket] *
 886                                         hp_used[i].hugepage_sz;
 887
 888                         memory[socket] -= cur_mem;
 889                         total_mem -= cur_mem;
 890
 891                         total_num_pages += hp_used[i].num_pages[socket];
 892
 893                         /* check if we have met all memory requests */
 894                         if (memory[socket] == 0)
 895                                 break;
 896
 897                         /* check if we have any more pages left at this size, if so
 898                          * move on to next size */
 899                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
 900                                 continue;
 901                         /* At this point we know that there are more pages available that are
 902                          * bigger than the memory we want, so lets see if we can get enough
 903                          * from other page sizes.
 904                          */
 905                         remaining_mem = 0;
 906                         for (j = i+1; j < num_hp_info; j++)
 907                                 remaining_mem += hp_info[j].hugepage_sz *
 908                                 hp_info[j].num_pages[socket];
 909
 910                         /* is there enough other memory, if not allocate another page and quit */
 911                         if (remaining_mem < memory[socket]){
 912                                 cur_mem = RTE_MIN(memory[socket],
 913                                                 hp_info[i].hugepage_sz);
 914                                 memory[socket] -= cur_mem;
 915                                 total_mem -= cur_mem;
 916                                 hp_used[i].num_pages[socket]++;
 917                                 total_num_pages++;
 918                                 break; /* we are done with this socket*/
 919                         }
 920                 }
 921                 /* if we didn't satisfy all memory requirements per socket */
 922                 if (memory[socket] > 0) {
 923                         /* to prevent icc errors */
 924                         requested = (unsigned) (internal_config.socket_mem[socket] /
 925                                         0x100000);
 926                         available = requested -
 927                                         ((unsigned) (memory[socket] / 0x100000));
 928                         RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
 929                                         "Requested: %uMB, available: %uMB\n", socket,
 930                                         requested, available);
 931                         return -1;
 932                 }
 933         }
 934
 935         /* if we didn't satisfy total memory requirements */
 936         if (total_mem > 0) {
 937                 requested = (unsigned) (internal_config.memory / 0x100000);
 938                 available = requested - (unsigned) (total_mem / 0x100000);
 939                 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
 940                                 " available: %uMB\n", requested, available);
 941                 return -1;
 942         }
 943         return total_num_pages;
 944 }
 945
 946 static inline size_t
 947 eal_get_hugepage_mem_size(void)
 948 {
 949         uint64_t size = 0;
 950         unsigned i, j;
 951
 952         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 953                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 954                 if (hpi->hugedir != NULL) {
 955                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 956                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 957                         }
 958                 }
 959         }
 960
 961         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 962 }
 963
 964 static struct sigaction huge_action_old;
 965 static int huge_need_recover;
 966
 967 static void
 968 huge_register_sigbus(void)
 969 {
 970         sigset_t mask;
 971         struct sigaction action;
 972
 973         sigemptyset(&mask);
 974         sigaddset(&mask, SIGBUS);
 975         action.sa_flags = 0;
 976         action.sa_mask = mask;
 977         action.sa_handler = huge_sigbus_handler;
 978
 979         huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
 980 }
 981
 982 static void
 983 huge_recover_sigbus(void)
 984 {
 985         if (huge_need_recover) {
 986                 sigaction(SIGBUS, &huge_action_old, NULL);
 987                 huge_need_recover = 0;
 988         }
 989 }
 990
 991 /*
 992  * Prepare physical memory mapping: fill configuration structure with
 993  * these infos, return 0 on success.
 994  *  1. map N huge pages in separate files in hugetlbfs
 995  *  2. find associated physical addr
 996  *  3. find associated NUMA socket ID
 997  *  4. sort all huge pages by physical address
 998  *  5. remap these N huge pages in the correct order
 999  *  6. unmap the first mapping
1000  *  7. fill memsegs in configuration with contiguous zones
1001  */
1002 int
1003 rte_eal_hugepage_init(void)
1004 {
1005         struct rte_mem_config *mcfg;
1006         struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1007         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1008
1009         uint64_t memory[RTE_MAX_NUMA_NODES];
1010
1011         unsigned hp_offset;
1012         int i, j, new_memseg;
1013         int nr_hugefiles, nr_hugepages = 0;
1014         void *addr;
1015
1016         test_phys_addrs_available();
1017
1018         memset(used_hp, 0, sizeof(used_hp));
1019
1020         /* get pointer to global configuration */
1021         mcfg = rte_eal_get_configuration()->mem_config;
1022
1023         /* hugetlbfs can be disabled */
1024         if (internal_config.no_hugetlbfs) {
1025                 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1026                                 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1027                 if (addr == MAP_FAILED) {
1028                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1029                                         strerror(errno));
1030                         return -1;
1031                 }
1032                 if (rte_eal_iova_mode() == RTE_IOVA_VA)
1033                         mcfg->memseg[0].phys_addr = (uintptr_t)addr;
1034                 else
1035                         mcfg->memseg[0].phys_addr = RTE_BAD_PHYS_ADDR;
1036                 mcfg->memseg[0].addr = addr;
1037                 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1038                 mcfg->memseg[0].len = internal_config.memory;
1039                 mcfg->memseg[0].socket_id = 0;
1040                 return 0;
1041         }
1042
1043         /* calculate total number of hugepages available. at this point we haven't
1044          * yet started sorting them so they all are on socket 0 */
1045         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1046                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1047                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1048
1049                 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1050         }
1051
1052         /*
1053          * allocate a memory area for hugepage table.
1054          * this isn't shared memory yet. due to the fact that we need some
1055          * processing done on these pages, shared memory will be created
1056          * at a later stage.
1057          */
1058         tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1059         if (tmp_hp == NULL)
1060                 goto fail;
1061
1062         memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1063
1064         hp_offset = 0; /* where we start the current page size entries */
1065
1066         huge_register_sigbus();
1067
1068         /* make a copy of socket_mem, needed for balanced allocation. */
1069         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1070                 memory[i] = internal_config.socket_mem[i];
1071
1072
1073         /* map all hugepages and sort them */
1074         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1075                 unsigned pages_old, pages_new;
1076                 struct hugepage_info *hpi;
1077
1078                 /*
1079                  * we don't yet mark hugepages as used at this stage, so
1080                  * we just map all hugepages available to the system
1081                  * all hugepages are still located on socket 0
1082                  */
1083                 hpi = &internal_config.hugepage_info[i];
1084
1085                 if (hpi->num_pages[0] == 0)
1086                         continue;
1087
1088                 /* map all hugepages available */
1089                 pages_old = hpi->num_pages[0];
1090                 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi,
1091                                               memory, 1);
1092                 if (pages_new < pages_old) {
1093                         RTE_LOG(DEBUG, EAL,
1094                                 "%d not %d hugepages of size %u MB allocated\n",
1095                                 pages_new, pages_old,
1096                                 (unsigned)(hpi->hugepage_sz / 0x100000));
1097
1098                         int pages = pages_old - pages_new;
1099
1100                         nr_hugepages -= pages;
1101                         hpi->num_pages[0] = pages_new;
1102                         if (pages_new == 0)
1103                                 continue;
1104                 }
1105
1106                 if (phys_addrs_available) {
1107                         /* find physical addresses for each hugepage */
1108                         if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1109                                 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1110                                         "for %u MB pages\n",
1111                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1112                                 goto fail;
1113                         }
1114                 } else {
1115                         /* set physical addresses for each hugepage */
1116                         if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1117                                 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1118                                         "for %u MB pages\n",
1119                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1120                                 goto fail;
1121                         }
1122                 }
1123
1124                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1125                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1126                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1127                         goto fail;
1128                 }
1129
1130                 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1131                       sizeof(struct hugepage_file), cmp_physaddr);
1132
1133                 /* remap all hugepages */
1134                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, NULL, 0) !=
1135                     hpi->num_pages[0]) {
1136                         RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1137                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1138                         goto fail;
1139                 }
1140
1141                 /* unmap original mappings */
1142                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1143                         goto fail;
1144
1145                 /* we have processed a num of hugepages of this size, so inc offset */
1146                 hp_offset += hpi->num_pages[0];
1147         }
1148
1149         huge_recover_sigbus();
1150
1151         if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1152                 internal_config.memory = eal_get_hugepage_mem_size();
1153
1154         nr_hugefiles = nr_hugepages;
1155
1156
1157         /* clean out the numbers of pages */
1158         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1159                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1160                         internal_config.hugepage_info[i].num_pages[j] = 0;
1161
1162         /* get hugepages for each socket */
1163         for (i = 0; i < nr_hugefiles; i++) {
1164                 int socket = tmp_hp[i].socket_id;
1165
1166                 /* find a hugepage info with right size and increment num_pages */
1167                 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1168                                 (int)internal_config.num_hugepage_sizes);
1169                 for (j = 0; j < nb_hpsizes; j++) {
1170                         if (tmp_hp[i].size ==
1171                                         internal_config.hugepage_info[j].hugepage_sz) {
1172                                 internal_config.hugepage_info[j].num_pages[socket]++;
1173                         }
1174                 }
1175         }
1176
1177         /* make a copy of socket_mem, needed for number of pages calculation */
1178         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1179                 memory[i] = internal_config.socket_mem[i];
1180
1181         /* calculate final number of pages */
1182         nr_hugepages = calc_num_pages_per_socket(memory,
1183                         internal_config.hugepage_info, used_hp,
1184                         internal_config.num_hugepage_sizes);
1185
1186         /* error if not enough memory available */
1187         if (nr_hugepages < 0)
1188                 goto fail;
1189
1190         /* reporting in! */
1191         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1192                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1193                         if (used_hp[i].num_pages[j] > 0) {
1194                                 RTE_LOG(DEBUG, EAL,
1195                                         "Requesting %u pages of size %uMB"
1196                                         " from socket %i\n",
1197                                         used_hp[i].num_pages[j],
1198                                         (unsigned)
1199                                         (used_hp[i].hugepage_sz / 0x100000),
1200                                         j);
1201                         }
1202                 }
1203         }
1204
1205         /* create shared memory */
1206         hugepage = create_shared_memory(eal_hugepage_info_path(),
1207                         nr_hugefiles * sizeof(struct hugepage_file));
1208
1209         if (hugepage == NULL) {
1210                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1211                 goto fail;
1212         }
1213         memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1214
1215         /*
1216          * unmap pages that we won't need (looks at used_hp).
1217          * also, sets final_va to NULL on pages that were unmapped.
1218          */
1219         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1220                         internal_config.num_hugepage_sizes) < 0) {
1221                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1222                 goto fail;
1223         }
1224
1225         /*
1226          * copy stuff from malloc'd hugepage* to the actual shared memory.
1227          * this procedure only copies those hugepages that have final_va
1228          * not NULL. has overflow protection.
1229          */
1230         if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1231                         tmp_hp, nr_hugefiles) < 0) {
1232                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1233                 goto fail;
1234         }
1235
1236         /* free the hugepage backing files */
1237         if (internal_config.hugepage_unlink &&
1238                 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1239                 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1240                 goto fail;
1241         }
1242
1243         /* free the temporary hugepage table */
1244         free(tmp_hp);
1245         tmp_hp = NULL;
1246
1247         /* first memseg index shall be 0 after incrementing it below */
1248         j = -1;
1249         for (i = 0; i < nr_hugefiles; i++) {
1250                 new_memseg = 0;
1251
1252                 /* if this is a new section, create a new memseg */
1253                 if (i == 0)
1254                         new_memseg = 1;
1255                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1256                         new_memseg = 1;
1257                 else if (hugepage[i].size != hugepage[i-1].size)
1258                         new_memseg = 1;
1259
1260 #ifdef RTE_ARCH_PPC_64
1261                 /* On PPC64 architecture, the mmap always start from higher
1262                  * virtual address to lower address. Here, both the physical
1263                  * address and virtual address are in descending order */
1264                 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1265                     hugepage[i].size)
1266                         new_memseg = 1;
1267                 else if (((unsigned long)hugepage[i-1].final_va -
1268                     (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1269                         new_memseg = 1;
1270 #else
1271                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1272                     hugepage[i].size)
1273                         new_memseg = 1;
1274                 else if (((unsigned long)hugepage[i].final_va -
1275                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1276                         new_memseg = 1;
1277 #endif
1278
1279                 if (new_memseg) {
1280                         j += 1;
1281                         if (j == RTE_MAX_MEMSEG)
1282                                 break;
1283
1284                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1285                         mcfg->memseg[j].addr = hugepage[i].final_va;
1286                         mcfg->memseg[j].len = hugepage[i].size;
1287                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1288                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1289                 }
1290                 /* continuation of previous memseg */
1291                 else {
1292 #ifdef RTE_ARCH_PPC_64
1293                 /* Use the phy and virt address of the last page as segment
1294                  * address for IBM Power architecture */
1295                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1296                         mcfg->memseg[j].addr = hugepage[i].final_va;
1297 #endif
1298                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1299                 }
1300                 hugepage[i].memseg_id = j;
1301         }
1302
1303         if (i < nr_hugefiles) {
1304                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1305                         "from %d requested\n"
1306                         "Current %s=%d is not enough\n"
1307                         "Please either increase it or request less amount "
1308                         "of memory.\n",
1309                         i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1310                         RTE_MAX_MEMSEG);
1311                 goto fail;
1312         }
1313
1314         munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1315
1316         return 0;
1317
1318 fail:
1319         huge_recover_sigbus();
1320         free(tmp_hp);
1321         if (hugepage != NULL)
1322                 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1323
1324         return -1;
1325 }
1326
1327 /*
1328  * uses fstat to report the size of a file on disk
1329  */
1330 static off_t
1331 getFileSize(int fd)
1332 {
1333         struct stat st;
1334         if (fstat(fd, &st) < 0)
1335                 return 0;
1336         return st.st_size;
1337 }
1338
1339 /*
1340  * This creates the memory mappings in the secondary process to match that of
1341  * the server process. It goes through each memory segment in the DPDK runtime
1342  * configuration and finds the hugepages which form that segment, mapping them
1343  * in order to form a contiguous block in the virtual memory space
1344  */
1345 int
1346 rte_eal_hugepage_attach(void)
1347 {
1348         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1349         struct hugepage_file *hp = NULL;
1350         unsigned num_hp = 0;
1351         unsigned i, s = 0; /* s used to track the segment number */
1352         unsigned max_seg = RTE_MAX_MEMSEG;
1353         off_t size = 0;
1354         int fd, fd_zero = -1, fd_hugepage = -1;
1355
1356         if (aslr_enabled() > 0) {
1357                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1358                                 "(ASLR) is enabled in the kernel.\n");
1359                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1360                                 "into secondary processes\n");
1361         }
1362
1363         test_phys_addrs_available();
1364
1365         fd_zero = open("/dev/zero", O_RDONLY);
1366         if (fd_zero < 0) {
1367                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1368                 goto error;
1369         }
1370         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1371         if (fd_hugepage < 0) {
1372                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1373                 goto error;
1374         }
1375
1376         /* map all segments into memory to make sure we get the addrs */
1377         for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1378                 void *base_addr;
1379
1380                 /*
1381                  * the first memory segment with len==0 is the one that
1382                  * follows the last valid segment.
1383                  */
1384                 if (mcfg->memseg[s].len == 0)
1385                         break;
1386
1387                 /*
1388                  * fdzero is mmapped to get a contiguous block of virtual
1389                  * addresses of the appropriate memseg size.
1390                  * use mmap to get identical addresses as the primary process.
1391                  */
1392                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1393                                  PROT_READ,
1394 #ifdef RTE_ARCH_PPC_64
1395                                  MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
1396 #else
1397                                  MAP_PRIVATE,
1398 #endif
1399                                  fd_zero, 0);
1400                 if (base_addr == MAP_FAILED ||
1401                     base_addr != mcfg->memseg[s].addr) {
1402                         max_seg = s;
1403                         if (base_addr != MAP_FAILED) {
1404                                 /* errno is stale, don't use */
1405                                 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1406                                         "in /dev/zero at [%p], got [%p] - "
1407                                         "please use '--base-virtaddr' option\n",
1408                                         (unsigned long long)mcfg->memseg[s].len,
1409                                         mcfg->memseg[s].addr, base_addr);
1410                                 munmap(base_addr, mcfg->memseg[s].len);
1411                         } else {
1412                                 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1413                                         "in /dev/zero at [%p]: '%s'\n",
1414                                         (unsigned long long)mcfg->memseg[s].len,
1415                                         mcfg->memseg[s].addr, strerror(errno));
1416                         }
1417                         if (aslr_enabled() > 0) {
1418                                 RTE_LOG(ERR, EAL, "It is recommended to "
1419                                         "disable ASLR in the kernel "
1420                                         "and retry running both primary "
1421                                         "and secondary processes\n");
1422                         }
1423                         goto error;
1424                 }
1425         }
1426
1427         size = getFileSize(fd_hugepage);
1428         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1429         if (hp == MAP_FAILED) {
1430                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1431                 goto error;
1432         }
1433
1434         num_hp = size / sizeof(struct hugepage_file);
1435         RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1436
1437         s = 0;
1438         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1439                 void *addr, *base_addr;
1440                 uintptr_t offset = 0;
1441                 size_t mapping_size;
1442                 /*
1443                  * free previously mapped memory so we can map the
1444                  * hugepages into the space
1445                  */
1446                 base_addr = mcfg->memseg[s].addr;
1447                 munmap(base_addr, mcfg->memseg[s].len);
1448
1449                 /* find the hugepages for this segment and map them
1450                  * we don't need to worry about order, as the server sorted the
1451                  * entries before it did the second mmap of them */
1452                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1453                         if (hp[i].memseg_id == (int)s){
1454                                 fd = open(hp[i].filepath, O_RDWR);
1455                                 if (fd < 0) {
1456                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1457                                                 hp[i].filepath);
1458                                         goto error;
1459                                 }
1460                                 mapping_size = hp[i].size;
1461                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1462                                                 mapping_size, PROT_READ | PROT_WRITE,
1463                                                 MAP_SHARED, fd, 0);
1464                                 close(fd); /* close file both on success and on failure */
1465                                 if (addr == MAP_FAILED ||
1466                                                 addr != RTE_PTR_ADD(base_addr, offset)) {
1467                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1468                                                 hp[i].filepath);
1469                                         goto error;
1470                                 }
1471                                 offset+=mapping_size;
1472                         }
1473                 }
1474                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1475                                 (unsigned long long)mcfg->memseg[s].len);
1476                 s++;
1477         }
1478         /* unmap the hugepage config file, since we are done using it */
1479         munmap(hp, size);
1480         close(fd_zero);
1481         close(fd_hugepage);
1482         return 0;
1483
1484 error:
1485         for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1486                 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1487         if (hp != NULL && hp != MAP_FAILED)
1488                 munmap(hp, size);
1489         if (fd_zero >= 0)
1490                 close(fd_zero);
1491         if (fd_hugepage >= 0)
1492                 close(fd_hugepage);
1493         return -1;
1494 }
1495
1496 int
1497 rte_eal_using_phys_addrs(void)
1498 {
1499         return phys_addrs_available;
1500 }