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 <stdarg.h>
  45 #include <sys/mman.h>
  46 #include <sys/types.h>
  47 #include <sys/stat.h>
  48 #include <sys/queue.h>
  49 #include <sys/file.h>
  50 #include <unistd.h>
  51 #include <limits.h>
  52 #include <errno.h>
  53 #include <sys/ioctl.h>
  54 #include <sys/time.h>
  55 #include <signal.h>
  56 #include <setjmp.h>
  57
  58 #include <rte_log.h>
  59 #include <rte_memory.h>
  60 #include <rte_memzone.h>
  61 #include <rte_launch.h>
  62 #include <rte_eal.h>
  63 #include <rte_eal_memconfig.h>
  64 #include <rte_per_lcore.h>
  65 #include <rte_lcore.h>
  66 #include <rte_common.h>
  67 #include <rte_string_fns.h>
  68
  69 #include "eal_private.h"
  70 #include "eal_internal_cfg.h"
  71 #include "eal_filesystem.h"
  72 #include "eal_hugepages.h"
  73
  74 #define PFN_MASK_SIZE   8
  75
  76 #ifdef RTE_LIBRTE_XEN_DOM0
  77 int rte_xen_dom0_supported(void)
  78 {
  79         return internal_config.xen_dom0_support;
  80 }
  81 #endif
  82
  83 /**
  84  * @file
  85  * Huge page mapping under linux
  86  *
  87  * To reserve a big contiguous amount of memory, we use the hugepage
  88  * feature of linux. For that, we need to have hugetlbfs mounted. This
  89  * code will create many files in this directory (one per page) and
  90  * map them in virtual memory. For each page, we will retrieve its
  91  * physical address and remap it in order to have a virtual contiguous
  92  * zone as well as a physical contiguous zone.
  93  */
  94
  95 static uint64_t baseaddr_offset;
  96
  97 static bool phys_addrs_available = true;
  98
  99 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
 100
 101 static void
 102 test_phys_addrs_available(void)
 103 {
 104         uint64_t tmp;
 105         phys_addr_t physaddr;
 106
 107         /* For dom0, phys addresses can always be available */
 108         if (rte_xen_dom0_supported())
 109                 return;
 110
 111         physaddr = rte_mem_virt2phy(&tmp);
 112         if (physaddr == RTE_BAD_PHYS_ADDR) {
 113                 RTE_LOG(ERR, EAL,
 114                         "Cannot obtain physical addresses: %s. "
 115                         "Only vfio will function.\n",
 116                         strerror(errno));
 117                 phys_addrs_available = false;
 118         }
 119 }
 120
 121 /* Lock page in physical memory and prevent from swapping. */
 122 int
 123 rte_mem_lock_page(const void *virt)
 124 {
 125         unsigned long virtual = (unsigned long)virt;
 126         int page_size = getpagesize();
 127         unsigned long aligned = (virtual & ~ (page_size - 1));
 128         return mlock((void*)aligned, page_size);
 129 }
 130
 131 /*
 132  * Get physical address of any mapped virtual address in the current process.
 133  */
 134 phys_addr_t
 135 rte_mem_virt2phy(const void *virtaddr)
 136 {
 137         int fd, retval;
 138         uint64_t page, physaddr;
 139         unsigned long virt_pfn;
 140         int page_size;
 141         off_t offset;
 142
 143         /* when using dom0, /proc/self/pagemap always returns 0, check in
 144          * dpdk memory by browsing the memsegs */
 145         if (rte_xen_dom0_supported()) {
 146                 struct rte_mem_config *mcfg;
 147                 struct rte_memseg *memseg;
 148                 unsigned i;
 149
 150                 mcfg = rte_eal_get_configuration()->mem_config;
 151                 for (i = 0; i < RTE_MAX_MEMSEG; i++) {
 152                         memseg = &mcfg->memseg[i];
 153                         if (memseg->addr == NULL)
 154                                 break;
 155                         if (virtaddr > memseg->addr &&
 156                                         virtaddr < RTE_PTR_ADD(memseg->addr,
 157                                                 memseg->len)) {
 158                                 return memseg->phys_addr +
 159                                         RTE_PTR_DIFF(virtaddr, memseg->addr);
 160                         }
 161                 }
 162
 163                 return RTE_BAD_PHYS_ADDR;
 164         }
 165
 166         /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
 167         if (!phys_addrs_available)
 168                 return RTE_BAD_PHYS_ADDR;
 169
 170         /* standard page size */
 171         page_size = getpagesize();
 172
 173         fd = open("/proc/self/pagemap", O_RDONLY);
 174         if (fd < 0) {
 175                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 176                         __func__, strerror(errno));
 177                 return RTE_BAD_PHYS_ADDR;
 178         }
 179
 180         virt_pfn = (unsigned long)virtaddr / page_size;
 181         offset = sizeof(uint64_t) * virt_pfn;
 182         if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 183                 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 184                                 __func__, strerror(errno));
 185                 close(fd);
 186                 return RTE_BAD_PHYS_ADDR;
 187         }
 188
 189         retval = read(fd, &page, PFN_MASK_SIZE);
 190         close(fd);
 191         if (retval < 0) {
 192                 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 193                                 __func__, strerror(errno));
 194                 return RTE_BAD_PHYS_ADDR;
 195         } else if (retval != PFN_MASK_SIZE) {
 196                 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
 197                                 "but expected %d:\n",
 198                                 __func__, retval, PFN_MASK_SIZE);
 199                 return RTE_BAD_PHYS_ADDR;
 200         }
 201
 202         /*
 203          * the pfn (page frame number) are bits 0-54 (see
 204          * pagemap.txt in linux Documentation)
 205          */
 206         if ((page & 0x7fffffffffffffULL) == 0)
 207                 return RTE_BAD_PHYS_ADDR;
 208
 209         physaddr = ((page & 0x7fffffffffffffULL) * page_size)
 210                 + ((unsigned long)virtaddr % page_size);
 211
 212         return physaddr;
 213 }
 214
 215 /*
 216  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 217  * it by browsing the /proc/self/pagemap special file.
 218  */
 219 static int
 220 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 221 {
 222         unsigned int i;
 223         phys_addr_t addr;
 224
 225         for (i = 0; i < hpi->num_pages[0]; i++) {
 226                 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
 227                 if (addr == RTE_BAD_PHYS_ADDR)
 228                         return -1;
 229                 hugepg_tbl[i].physaddr = addr;
 230         }
 231         return 0;
 232 }
 233
 234 /*
 235  * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
 236  */
 237 static int
 238 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 239 {
 240         unsigned int i;
 241         static phys_addr_t addr;
 242
 243         for (i = 0; i < hpi->num_pages[0]; i++) {
 244                 hugepg_tbl[i].physaddr = addr;
 245                 addr += hugepg_tbl[i].size;
 246         }
 247         return 0;
 248 }
 249
 250 /*
 251  * Check whether address-space layout randomization is enabled in
 252  * the kernel. This is important for multi-process as it can prevent
 253  * two processes mapping data to the same virtual address
 254  * Returns:
 255  *    0 - address space randomization disabled
 256  *    1/2 - address space randomization enabled
 257  *    negative error code on error
 258  */
 259 static int
 260 aslr_enabled(void)
 261 {
 262         char c;
 263         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 264         if (fd < 0)
 265                 return -errno;
 266         retval = read(fd, &c, 1);
 267         close(fd);
 268         if (retval < 0)
 269                 return -errno;
 270         if (retval == 0)
 271                 return -EIO;
 272         switch (c) {
 273                 case '0' : return 0;
 274                 case '1' : return 1;
 275                 case '2' : return 2;
 276                 default: return -EINVAL;
 277         }
 278 }
 279
 280 /*
 281  * Try to mmap *size bytes in /dev/zero. If it is successful, return the
 282  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 283  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 284  * 0. In this case, return NULL. Note: this function returns an address
 285  * which is a multiple of hugepage size.
 286  */
 287 static void *
 288 get_virtual_area(size_t *size, size_t hugepage_sz)
 289 {
 290         void *addr;
 291         int fd;
 292         long aligned_addr;
 293
 294         if (internal_config.base_virtaddr != 0) {
 295                 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
 296                                 baseaddr_offset);
 297         }
 298         else addr = NULL;
 299
 300         RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
 301
 302         fd = open("/dev/zero", O_RDONLY);
 303         if (fd < 0){
 304                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 305                 return NULL;
 306         }
 307         do {
 308                 addr = mmap(addr,
 309                                 (*size) + hugepage_sz, PROT_READ,
 310 #ifdef RTE_ARCH_PPC_64
 311                                 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
 312 #else
 313                                 MAP_PRIVATE,
 314 #endif
 315                                 fd, 0);
 316                 if (addr == MAP_FAILED)
 317                         *size -= hugepage_sz;
 318         } while (addr == MAP_FAILED && *size > 0);
 319
 320         if (addr == MAP_FAILED) {
 321                 close(fd);
 322                 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
 323                         strerror(errno));
 324                 return NULL;
 325         }
 326
 327         munmap(addr, (*size) + hugepage_sz);
 328         close(fd);
 329
 330         /* align addr to a huge page size boundary */
 331         aligned_addr = (long)addr;
 332         aligned_addr += (hugepage_sz - 1);
 333         aligned_addr &= (~(hugepage_sz - 1));
 334         addr = (void *)(aligned_addr);
 335
 336         RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
 337                 addr, *size);
 338
 339         /* increment offset */
 340         baseaddr_offset += *size;
 341
 342         return addr;
 343 }
 344
 345 static sigjmp_buf huge_jmpenv;
 346
 347 static void huge_sigbus_handler(int signo __rte_unused)
 348 {
 349         siglongjmp(huge_jmpenv, 1);
 350 }
 351
 352 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
 353  * non-static local variable in the stack frame calling sigsetjmp might be
 354  * clobbered by a call to longjmp.
 355  */
 356 static int huge_wrap_sigsetjmp(void)
 357 {
 358         return sigsetjmp(huge_jmpenv, 1);
 359 }
 360
 361 /*
 362  * Mmap all hugepages of hugepage table: it first open a file in
 363  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 364  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 365  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 366  * map continguous physical blocks in contiguous virtual blocks.
 367  */
 368 static unsigned
 369 map_all_hugepages(struct hugepage_file *hugepg_tbl,
 370                 struct hugepage_info *hpi, int orig)
 371 {
 372         int fd;
 373         unsigned i;
 374         void *virtaddr;
 375         void *vma_addr = NULL;
 376         size_t vma_len = 0;
 377
 378         for (i = 0; i < hpi->num_pages[0]; i++) {
 379                 uint64_t hugepage_sz = hpi->hugepage_sz;
 380
 381                 if (orig) {
 382                         hugepg_tbl[i].file_id = i;
 383                         hugepg_tbl[i].size = hugepage_sz;
 384                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 385                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 386                                         hugepg_tbl[i].file_id);
 387                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 388                 }
 389 #ifndef RTE_ARCH_64
 390                 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
 391                  * original map address as final map address.
 392                  */
 393                 else if ((hugepage_sz == RTE_PGSIZE_1G)
 394                         || (hugepage_sz == RTE_PGSIZE_16G)) {
 395                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 396                         hugepg_tbl[i].orig_va = NULL;
 397                         continue;
 398                 }
 399 #endif
 400                 else if (vma_len == 0) {
 401                         unsigned j, num_pages;
 402
 403                         /* reserve a virtual area for next contiguous
 404                          * physical block: count the number of
 405                          * contiguous physical pages. */
 406                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 407 #ifdef RTE_ARCH_PPC_64
 408                                 /* The physical addresses are sorted in
 409                                  * descending order on PPC64 */
 410                                 if (hugepg_tbl[j].physaddr !=
 411                                     hugepg_tbl[j-1].physaddr - hugepage_sz)
 412                                         break;
 413 #else
 414                                 if (hugepg_tbl[j].physaddr !=
 415                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 416                                         break;
 417 #endif
 418                         }
 419                         num_pages = j - i;
 420                         vma_len = num_pages * hugepage_sz;
 421
 422                         /* get the biggest virtual memory area up to
 423                          * vma_len. If it fails, vma_addr is NULL, so
 424                          * let the kernel provide the address. */
 425                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 426                         if (vma_addr == NULL)
 427                                 vma_len = hugepage_sz;
 428                 }
 429
 430                 /* try to create hugepage file */
 431                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
 432                 if (fd < 0) {
 433                         RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
 434                                         strerror(errno));
 435                         return i;
 436                 }
 437
 438                 /* map the segment, and populate page tables,
 439                  * the kernel fills this segment with zeros */
 440                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 441                                 MAP_SHARED | MAP_POPULATE, fd, 0);
 442                 if (virtaddr == MAP_FAILED) {
 443                         RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
 444                                         strerror(errno));
 445                         close(fd);
 446                         return i;
 447                 }
 448
 449                 if (orig) {
 450                         hugepg_tbl[i].orig_va = virtaddr;
 451                 }
 452                 else {
 453                         hugepg_tbl[i].final_va = virtaddr;
 454                 }
 455
 456                 if (orig) {
 457                         /* In linux, hugetlb limitations, like cgroup, are
 458                          * enforced at fault time instead of mmap(), even
 459                          * with the option of MAP_POPULATE. Kernel will send
 460                          * a SIGBUS signal. To avoid to be killed, save stack
 461                          * environment here, if SIGBUS happens, we can jump
 462                          * back here.
 463                          */
 464                         if (huge_wrap_sigsetjmp()) {
 465                                 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
 466                                         "hugepages of size %u MB\n",
 467                                         (unsigned)(hugepage_sz / 0x100000));
 468                                 munmap(virtaddr, hugepage_sz);
 469                                 close(fd);
 470                                 unlink(hugepg_tbl[i].filepath);
 471                                 return i;
 472                         }
 473                         *(int *)virtaddr = 0;
 474                 }
 475
 476
 477                 /* set shared flock on the file. */
 478                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 479                         RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
 480                                 __func__, strerror(errno));
 481                         close(fd);
 482                         return i;
 483                 }
 484
 485                 close(fd);
 486
 487                 vma_addr = (char *)vma_addr + hugepage_sz;
 488                 vma_len -= hugepage_sz;
 489         }
 490
 491         return i;
 492 }
 493
 494 /* Unmap all hugepages from original mapping */
 495 static int
 496 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 497 {
 498         unsigned i;
 499         for (i = 0; i < hpi->num_pages[0]; i++) {
 500                 if (hugepg_tbl[i].orig_va) {
 501                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 502                         hugepg_tbl[i].orig_va = NULL;
 503                 }
 504         }
 505         return 0;
 506 }
 507
 508 /*
 509  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 510  * page.
 511  */
 512 static int
 513 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 514 {
 515         int socket_id;
 516         char *end, *nodestr;
 517         unsigned i, hp_count = 0;
 518         uint64_t virt_addr;
 519         char buf[BUFSIZ];
 520         char hugedir_str[PATH_MAX];
 521         FILE *f;
 522
 523         f = fopen("/proc/self/numa_maps", "r");
 524         if (f == NULL) {
 525                 RTE_LOG(NOTICE, EAL, "cannot open /proc/self/numa_maps,"
 526                                 " consider that all memory is in socket_id 0\n");
 527                 return 0;
 528         }
 529
 530         snprintf(hugedir_str, sizeof(hugedir_str),
 531                         "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
 532
 533         /* parse numa map */
 534         while (fgets(buf, sizeof(buf), f) != NULL) {
 535
 536                 /* ignore non huge page */
 537                 if (strstr(buf, " huge ") == NULL &&
 538                                 strstr(buf, hugedir_str) == NULL)
 539                         continue;
 540
 541                 /* get zone addr */
 542                 virt_addr = strtoull(buf, &end, 16);
 543                 if (virt_addr == 0 || end == buf) {
 544                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 545                         goto error;
 546                 }
 547
 548                 /* get node id (socket id) */
 549                 nodestr = strstr(buf, " N");
 550                 if (nodestr == NULL) {
 551                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 552                         goto error;
 553                 }
 554                 nodestr += 2;
 555                 end = strstr(nodestr, "=");
 556                 if (end == NULL) {
 557                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 558                         goto error;
 559                 }
 560                 end[0] = '\0';
 561                 end = NULL;
 562
 563                 socket_id = strtoul(nodestr, &end, 0);
 564                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 565                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 566                         goto error;
 567                 }
 568
 569                 /* if we find this page in our mappings, set socket_id */
 570                 for (i = 0; i < hpi->num_pages[0]; i++) {
 571                         void *va = (void *)(unsigned long)virt_addr;
 572                         if (hugepg_tbl[i].orig_va == va) {
 573                                 hugepg_tbl[i].socket_id = socket_id;
 574                                 hp_count++;
 575                         }
 576                 }
 577         }
 578
 579         if (hp_count < hpi->num_pages[0])
 580                 goto error;
 581
 582         fclose(f);
 583         return 0;
 584
 585 error:
 586         fclose(f);
 587         return -1;
 588 }
 589
 590 static int
 591 cmp_physaddr(const void *a, const void *b)
 592 {
 593 #ifndef RTE_ARCH_PPC_64
 594         const struct hugepage_file *p1 = a;
 595         const struct hugepage_file *p2 = b;
 596 #else
 597         /* PowerPC needs memory sorted in reverse order from x86 */
 598         const struct hugepage_file *p1 = b;
 599         const struct hugepage_file *p2 = a;
 600 #endif
 601         if (p1->physaddr < p2->physaddr)
 602                 return -1;
 603         else if (p1->physaddr > p2->physaddr)
 604                 return 1;
 605         else
 606                 return 0;
 607 }
 608
 609 /*
 610  * Uses mmap to create a shared memory area for storage of data
 611  * Used in this file to store the hugepage file map on disk
 612  */
 613 static void *
 614 create_shared_memory(const char *filename, const size_t mem_size)
 615 {
 616         void *retval;
 617         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 618         if (fd < 0)
 619                 return NULL;
 620         if (ftruncate(fd, mem_size) < 0) {
 621                 close(fd);
 622                 return NULL;
 623         }
 624         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 625         close(fd);
 626         return retval;
 627 }
 628
 629 /*
 630  * this copies *active* hugepages from one hugepage table to another.
 631  * destination is typically the shared memory.
 632  */
 633 static int
 634 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
 635                 const struct hugepage_file * src, int src_size)
 636 {
 637         int src_pos, dst_pos = 0;
 638
 639         for (src_pos = 0; src_pos < src_size; src_pos++) {
 640                 if (src[src_pos].final_va != NULL) {
 641                         /* error on overflow attempt */
 642                         if (dst_pos == dest_size)
 643                                 return -1;
 644                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
 645                         dst_pos++;
 646                 }
 647         }
 648         return 0;
 649 }
 650
 651 static int
 652 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
 653                 unsigned num_hp_info)
 654 {
 655         unsigned socket, size;
 656         int page, nrpages = 0;
 657
 658         /* get total number of hugepages */
 659         for (size = 0; size < num_hp_info; size++)
 660                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 661                         nrpages +=
 662                         internal_config.hugepage_info[size].num_pages[socket];
 663
 664         for (page = 0; page < nrpages; page++) {
 665                 struct hugepage_file *hp = &hugepg_tbl[page];
 666
 667                 if (hp->final_va != NULL && unlink(hp->filepath)) {
 668                         RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
 669                                 __func__, hp->filepath, strerror(errno));
 670                 }
 671         }
 672         return 0;
 673 }
 674
 675 /*
 676  * unmaps hugepages that are not going to be used. since we originally allocate
 677  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 678  */
 679 static int
 680 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
 681                 struct hugepage_info *hpi,
 682                 unsigned num_hp_info)
 683 {
 684         unsigned socket, size;
 685         int page, nrpages = 0;
 686
 687         /* get total number of hugepages */
 688         for (size = 0; size < num_hp_info; size++)
 689                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 690                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 691
 692         for (size = 0; size < num_hp_info; size++) {
 693                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 694                         unsigned pages_found = 0;
 695
 696                         /* traverse until we have unmapped all the unused pages */
 697                         for (page = 0; page < nrpages; page++) {
 698                                 struct hugepage_file *hp = &hugepg_tbl[page];
 699
 700                                 /* find a page that matches the criteria */
 701                                 if ((hp->size == hpi[size].hugepage_sz) &&
 702                                                 (hp->socket_id == (int) socket)) {
 703
 704                                         /* if we skipped enough pages, unmap the rest */
 705                                         if (pages_found == hpi[size].num_pages[socket]) {
 706                                                 uint64_t unmap_len;
 707
 708                                                 unmap_len = hp->size;
 709
 710                                                 /* get start addr and len of the remaining segment */
 711                                                 munmap(hp->final_va, (size_t) unmap_len);
 712
 713                                                 hp->final_va = NULL;
 714                                                 if (unlink(hp->filepath) == -1) {
 715                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 716                                                                         __func__, hp->filepath, strerror(errno));
 717                                                         return -1;
 718                                                 }
 719                                         } else {
 720                                                 /* lock the page and skip */
 721                                                 pages_found++;
 722                                         }
 723
 724                                 } /* match page */
 725                         } /* foreach page */
 726                 } /* foreach socket */
 727         } /* foreach pagesize */
 728
 729         return 0;
 730 }
 731
 732 static inline uint64_t
 733 get_socket_mem_size(int socket)
 734 {
 735         uint64_t size = 0;
 736         unsigned i;
 737
 738         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 739                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 740                 if (hpi->hugedir != NULL)
 741                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 742         }
 743
 744         return size;
 745 }
 746
 747 /*
 748  * This function is a NUMA-aware equivalent of calc_num_pages.
 749  * It takes in the list of hugepage sizes and the
 750  * number of pages thereof, and calculates the best number of
 751  * pages of each size to fulfill the request for <memory> ram
 752  */
 753 static int
 754 calc_num_pages_per_socket(uint64_t * memory,
 755                 struct hugepage_info *hp_info,
 756                 struct hugepage_info *hp_used,
 757                 unsigned num_hp_info)
 758 {
 759         unsigned socket, j, i = 0;
 760         unsigned requested, available;
 761         int total_num_pages = 0;
 762         uint64_t remaining_mem, cur_mem;
 763         uint64_t total_mem = internal_config.memory;
 764
 765         if (num_hp_info == 0)
 766                 return -1;
 767
 768         /* if specific memory amounts per socket weren't requested */
 769         if (internal_config.force_sockets == 0) {
 770                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
 771                 size_t default_size, total_size;
 772                 unsigned lcore_id;
 773
 774                 /* Compute number of cores per socket */
 775                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
 776                 RTE_LCORE_FOREACH(lcore_id) {
 777                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
 778                 }
 779
 780                 /*
 781                  * Automatically spread requested memory amongst detected sockets according
 782                  * to number of cores from cpu mask present on each socket
 783                  */
 784                 total_size = internal_config.memory;
 785                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 786
 787                         /* Set memory amount per socket */
 788                         default_size = (internal_config.memory * cpu_per_socket[socket])
 789                                         / rte_lcore_count();
 790
 791                         /* Limit to maximum available memory on socket */
 792                         default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
 793
 794                         /* Update sizes */
 795                         memory[socket] = default_size;
 796                         total_size -= default_size;
 797                 }
 798
 799                 /*
 800                  * If some memory is remaining, try to allocate it by getting all
 801                  * available memory from sockets, one after the other
 802                  */
 803                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 804                         /* take whatever is available */
 805                         default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
 806                                                total_size);
 807
 808                         /* Update sizes */
 809                         memory[socket] += default_size;
 810                         total_size -= default_size;
 811                 }
 812         }
 813
 814         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 815                 /* skips if the memory on specific socket wasn't requested */
 816                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 817                         hp_used[i].hugedir = hp_info[i].hugedir;
 818                         hp_used[i].num_pages[socket] = RTE_MIN(
 819                                         memory[socket] / hp_info[i].hugepage_sz,
 820                                         hp_info[i].num_pages[socket]);
 821
 822                         cur_mem = hp_used[i].num_pages[socket] *
 823                                         hp_used[i].hugepage_sz;
 824
 825                         memory[socket] -= cur_mem;
 826                         total_mem -= cur_mem;
 827
 828                         total_num_pages += hp_used[i].num_pages[socket];
 829
 830                         /* check if we have met all memory requests */
 831                         if (memory[socket] == 0)
 832                                 break;
 833
 834                         /* check if we have any more pages left at this size, if so
 835                          * move on to next size */
 836                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
 837                                 continue;
 838                         /* At this point we know that there are more pages available that are
 839                          * bigger than the memory we want, so lets see if we can get enough
 840                          * from other page sizes.
 841                          */
 842                         remaining_mem = 0;
 843                         for (j = i+1; j < num_hp_info; j++)
 844                                 remaining_mem += hp_info[j].hugepage_sz *
 845                                 hp_info[j].num_pages[socket];
 846
 847                         /* is there enough other memory, if not allocate another page and quit */
 848                         if (remaining_mem < memory[socket]){
 849                                 cur_mem = RTE_MIN(memory[socket],
 850                                                 hp_info[i].hugepage_sz);
 851                                 memory[socket] -= cur_mem;
 852                                 total_mem -= cur_mem;
 853                                 hp_used[i].num_pages[socket]++;
 854                                 total_num_pages++;
 855                                 break; /* we are done with this socket*/
 856                         }
 857                 }
 858                 /* if we didn't satisfy all memory requirements per socket */
 859                 if (memory[socket] > 0) {
 860                         /* to prevent icc errors */
 861                         requested = (unsigned) (internal_config.socket_mem[socket] /
 862                                         0x100000);
 863                         available = requested -
 864                                         ((unsigned) (memory[socket] / 0x100000));
 865                         RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
 866                                         "Requested: %uMB, available: %uMB\n", socket,
 867                                         requested, available);
 868                         return -1;
 869                 }
 870         }
 871
 872         /* if we didn't satisfy total memory requirements */
 873         if (total_mem > 0) {
 874                 requested = (unsigned) (internal_config.memory / 0x100000);
 875                 available = requested - (unsigned) (total_mem / 0x100000);
 876                 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
 877                                 " available: %uMB\n", requested, available);
 878                 return -1;
 879         }
 880         return total_num_pages;
 881 }
 882
 883 static inline size_t
 884 eal_get_hugepage_mem_size(void)
 885 {
 886         uint64_t size = 0;
 887         unsigned i, j;
 888
 889         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 890                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 891                 if (hpi->hugedir != NULL) {
 892                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 893                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 894                         }
 895                 }
 896         }
 897
 898         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 899 }
 900
 901 static struct sigaction huge_action_old;
 902 static int huge_need_recover;
 903
 904 static void
 905 huge_register_sigbus(void)
 906 {
 907         sigset_t mask;
 908         struct sigaction action;
 909
 910         sigemptyset(&mask);
 911         sigaddset(&mask, SIGBUS);
 912         action.sa_flags = 0;
 913         action.sa_mask = mask;
 914         action.sa_handler = huge_sigbus_handler;
 915
 916         huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
 917 }
 918
 919 static void
 920 huge_recover_sigbus(void)
 921 {
 922         if (huge_need_recover) {
 923                 sigaction(SIGBUS, &huge_action_old, NULL);
 924                 huge_need_recover = 0;
 925         }
 926 }
 927
 928 /*
 929  * Prepare physical memory mapping: fill configuration structure with
 930  * these infos, return 0 on success.
 931  *  1. map N huge pages in separate files in hugetlbfs
 932  *  2. find associated physical addr
 933  *  3. find associated NUMA socket ID
 934  *  4. sort all huge pages by physical address
 935  *  5. remap these N huge pages in the correct order
 936  *  6. unmap the first mapping
 937  *  7. fill memsegs in configuration with contiguous zones
 938  */
 939 int
 940 rte_eal_hugepage_init(void)
 941 {
 942         struct rte_mem_config *mcfg;
 943         struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
 944         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
 945
 946         uint64_t memory[RTE_MAX_NUMA_NODES];
 947
 948         unsigned hp_offset;
 949         int i, j, new_memseg;
 950         int nr_hugefiles, nr_hugepages = 0;
 951         void *addr;
 952
 953         test_phys_addrs_available();
 954
 955         memset(used_hp, 0, sizeof(used_hp));
 956
 957         /* get pointer to global configuration */
 958         mcfg = rte_eal_get_configuration()->mem_config;
 959
 960         /* hugetlbfs can be disabled */
 961         if (internal_config.no_hugetlbfs) {
 962                 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
 963                                 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
 964                 if (addr == MAP_FAILED) {
 965                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
 966                                         strerror(errno));
 967                         return -1;
 968                 }
 969                 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
 970                 mcfg->memseg[0].addr = addr;
 971                 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
 972                 mcfg->memseg[0].len = internal_config.memory;
 973                 mcfg->memseg[0].socket_id = 0;
 974                 return 0;
 975         }
 976
 977 /* check if app runs on Xen Dom0 */
 978         if (internal_config.xen_dom0_support) {
 979 #ifdef RTE_LIBRTE_XEN_DOM0
 980                 /* use dom0_mm kernel driver to init memory */
 981                 if (rte_xen_dom0_memory_init() < 0)
 982                         return -1;
 983                 else
 984                         return 0;
 985 #endif
 986         }
 987
 988         /* calculate total number of hugepages available. at this point we haven't
 989          * yet started sorting them so they all are on socket 0 */
 990         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
 991                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
 992                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
 993
 994                 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
 995         }
 996
 997         /*
 998          * allocate a memory area for hugepage table.
 999          * this isn't shared memory yet. due to the fact that we need some
1000          * processing done on these pages, shared memory will be created
1001          * at a later stage.
1002          */
1003         tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1004         if (tmp_hp == NULL)
1005                 goto fail;
1006
1007         memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1008
1009         hp_offset = 0; /* where we start the current page size entries */
1010
1011         huge_register_sigbus();
1012
1013         /* map all hugepages and sort them */
1014         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1015                 unsigned pages_old, pages_new;
1016                 struct hugepage_info *hpi;
1017
1018                 /*
1019                  * we don't yet mark hugepages as used at this stage, so
1020                  * we just map all hugepages available to the system
1021                  * all hugepages are still located on socket 0
1022                  */
1023                 hpi = &internal_config.hugepage_info[i];
1024
1025                 if (hpi->num_pages[0] == 0)
1026                         continue;
1027
1028                 /* map all hugepages available */
1029                 pages_old = hpi->num_pages[0];
1030                 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, 1);
1031                 if (pages_new < pages_old) {
1032                         RTE_LOG(DEBUG, EAL,
1033                                 "%d not %d hugepages of size %u MB allocated\n",
1034                                 pages_new, pages_old,
1035                                 (unsigned)(hpi->hugepage_sz / 0x100000));
1036
1037                         int pages = pages_old - pages_new;
1038
1039                         nr_hugepages -= pages;
1040                         hpi->num_pages[0] = pages_new;
1041                         if (pages_new == 0)
1042                                 continue;
1043                 }
1044
1045                 if (phys_addrs_available) {
1046                         /* find physical addresses for each hugepage */
1047                         if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1048                                 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1049                                         "for %u MB pages\n",
1050                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1051                                 goto fail;
1052                         }
1053                 } else {
1054                         /* set physical addresses for each hugepage */
1055                         if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1056                                 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1057                                         "for %u MB pages\n",
1058                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1059                                 goto fail;
1060                         }
1061                 }
1062
1063                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1064                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1065                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1066                         goto fail;
1067                 }
1068
1069                 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1070                       sizeof(struct hugepage_file), cmp_physaddr);
1071
1072                 /* remap all hugepages */
1073                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) !=
1074                     hpi->num_pages[0]) {
1075                         RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1076                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1077                         goto fail;
1078                 }
1079
1080                 /* unmap original mappings */
1081                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1082                         goto fail;
1083
1084                 /* we have processed a num of hugepages of this size, so inc offset */
1085                 hp_offset += hpi->num_pages[0];
1086         }
1087
1088         huge_recover_sigbus();
1089
1090         if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1091                 internal_config.memory = eal_get_hugepage_mem_size();
1092
1093         nr_hugefiles = nr_hugepages;
1094
1095
1096         /* clean out the numbers of pages */
1097         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1098                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1099                         internal_config.hugepage_info[i].num_pages[j] = 0;
1100
1101         /* get hugepages for each socket */
1102         for (i = 0; i < nr_hugefiles; i++) {
1103                 int socket = tmp_hp[i].socket_id;
1104
1105                 /* find a hugepage info with right size and increment num_pages */
1106                 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1107                                 (int)internal_config.num_hugepage_sizes);
1108                 for (j = 0; j < nb_hpsizes; j++) {
1109                         if (tmp_hp[i].size ==
1110                                         internal_config.hugepage_info[j].hugepage_sz) {
1111                                 internal_config.hugepage_info[j].num_pages[socket]++;
1112                         }
1113                 }
1114         }
1115
1116         /* make a copy of socket_mem, needed for number of pages calculation */
1117         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1118                 memory[i] = internal_config.socket_mem[i];
1119
1120         /* calculate final number of pages */
1121         nr_hugepages = calc_num_pages_per_socket(memory,
1122                         internal_config.hugepage_info, used_hp,
1123                         internal_config.num_hugepage_sizes);
1124
1125         /* error if not enough memory available */
1126         if (nr_hugepages < 0)
1127                 goto fail;
1128
1129         /* reporting in! */
1130         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1131                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1132                         if (used_hp[i].num_pages[j] > 0) {
1133                                 RTE_LOG(DEBUG, EAL,
1134                                         "Requesting %u pages of size %uMB"
1135                                         " from socket %i\n",
1136                                         used_hp[i].num_pages[j],
1137                                         (unsigned)
1138                                         (used_hp[i].hugepage_sz / 0x100000),
1139                                         j);
1140                         }
1141                 }
1142         }
1143
1144         /* create shared memory */
1145         hugepage = create_shared_memory(eal_hugepage_info_path(),
1146                         nr_hugefiles * sizeof(struct hugepage_file));
1147
1148         if (hugepage == NULL) {
1149                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1150                 goto fail;
1151         }
1152         memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1153
1154         /*
1155          * unmap pages that we won't need (looks at used_hp).
1156          * also, sets final_va to NULL on pages that were unmapped.
1157          */
1158         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1159                         internal_config.num_hugepage_sizes) < 0) {
1160                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1161                 goto fail;
1162         }
1163
1164         /*
1165          * copy stuff from malloc'd hugepage* to the actual shared memory.
1166          * this procedure only copies those hugepages that have final_va
1167          * not NULL. has overflow protection.
1168          */
1169         if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1170                         tmp_hp, nr_hugefiles) < 0) {
1171                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1172                 goto fail;
1173         }
1174
1175         /* free the hugepage backing files */
1176         if (internal_config.hugepage_unlink &&
1177                 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1178                 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1179                 goto fail;
1180         }
1181
1182         /* free the temporary hugepage table */
1183         free(tmp_hp);
1184         tmp_hp = NULL;
1185
1186         /* first memseg index shall be 0 after incrementing it below */
1187         j = -1;
1188         for (i = 0; i < nr_hugefiles; i++) {
1189                 new_memseg = 0;
1190
1191                 /* if this is a new section, create a new memseg */
1192                 if (i == 0)
1193                         new_memseg = 1;
1194                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1195                         new_memseg = 1;
1196                 else if (hugepage[i].size != hugepage[i-1].size)
1197                         new_memseg = 1;
1198
1199 #ifdef RTE_ARCH_PPC_64
1200                 /* On PPC64 architecture, the mmap always start from higher
1201                  * virtual address to lower address. Here, both the physical
1202                  * address and virtual address are in descending order */
1203                 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1204                     hugepage[i].size)
1205                         new_memseg = 1;
1206                 else if (((unsigned long)hugepage[i-1].final_va -
1207                     (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1208                         new_memseg = 1;
1209 #else
1210                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1211                     hugepage[i].size)
1212                         new_memseg = 1;
1213                 else if (((unsigned long)hugepage[i].final_va -
1214                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1215                         new_memseg = 1;
1216 #endif
1217
1218                 if (new_memseg) {
1219                         j += 1;
1220                         if (j == RTE_MAX_MEMSEG)
1221                                 break;
1222
1223                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1224                         mcfg->memseg[j].addr = hugepage[i].final_va;
1225                         mcfg->memseg[j].len = hugepage[i].size;
1226                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1227                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1228                 }
1229                 /* continuation of previous memseg */
1230                 else {
1231 #ifdef RTE_ARCH_PPC_64
1232                 /* Use the phy and virt address of the last page as segment
1233                  * address for IBM Power architecture */
1234                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1235                         mcfg->memseg[j].addr = hugepage[i].final_va;
1236 #endif
1237                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1238                 }
1239                 hugepage[i].memseg_id = j;
1240         }
1241
1242         if (i < nr_hugefiles) {
1243                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1244                         "from %d requested\n"
1245                         "Current %s=%d is not enough\n"
1246                         "Please either increase it or request less amount "
1247                         "of memory.\n",
1248                         i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1249                         RTE_MAX_MEMSEG);
1250                 goto fail;
1251         }
1252
1253         munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1254
1255         return 0;
1256
1257 fail:
1258         huge_recover_sigbus();
1259         free(tmp_hp);
1260         if (hugepage != NULL)
1261                 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1262
1263         return -1;
1264 }
1265
1266 /*
1267  * uses fstat to report the size of a file on disk
1268  */
1269 static off_t
1270 getFileSize(int fd)
1271 {
1272         struct stat st;
1273         if (fstat(fd, &st) < 0)
1274                 return 0;
1275         return st.st_size;
1276 }
1277
1278 /*
1279  * This creates the memory mappings in the secondary process to match that of
1280  * the server process. It goes through each memory segment in the DPDK runtime
1281  * configuration and finds the hugepages which form that segment, mapping them
1282  * in order to form a contiguous block in the virtual memory space
1283  */
1284 int
1285 rte_eal_hugepage_attach(void)
1286 {
1287         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1288         struct hugepage_file *hp = NULL;
1289         unsigned num_hp = 0;
1290         unsigned i, s = 0; /* s used to track the segment number */
1291         unsigned max_seg = RTE_MAX_MEMSEG;
1292         off_t size = 0;
1293         int fd, fd_zero = -1, fd_hugepage = -1;
1294
1295         if (aslr_enabled() > 0) {
1296                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1297                                 "(ASLR) is enabled in the kernel.\n");
1298                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1299                                 "into secondary processes\n");
1300         }
1301
1302         test_phys_addrs_available();
1303
1304         if (internal_config.xen_dom0_support) {
1305 #ifdef RTE_LIBRTE_XEN_DOM0
1306                 if (rte_xen_dom0_memory_attach() < 0) {
1307                         RTE_LOG(ERR, EAL, "Failed to attach memory segments of primary "
1308                                         "process\n");
1309                         return -1;
1310                 }
1311                 return 0;
1312 #endif
1313         }
1314
1315         fd_zero = open("/dev/zero", O_RDONLY);
1316         if (fd_zero < 0) {
1317                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1318                 goto error;
1319         }
1320         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1321         if (fd_hugepage < 0) {
1322                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1323                 goto error;
1324         }
1325
1326         /* map all segments into memory to make sure we get the addrs */
1327         for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1328                 void *base_addr;
1329
1330                 /*
1331                  * the first memory segment with len==0 is the one that
1332                  * follows the last valid segment.
1333                  */
1334                 if (mcfg->memseg[s].len == 0)
1335                         break;
1336
1337                 /*
1338                  * fdzero is mmapped to get a contiguous block of virtual
1339                  * addresses of the appropriate memseg size.
1340                  * use mmap to get identical addresses as the primary process.
1341                  */
1342                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1343                                  PROT_READ,
1344 #ifdef RTE_ARCH_PPC_64
1345                                  MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
1346 #else
1347                                  MAP_PRIVATE,
1348 #endif
1349                                  fd_zero, 0);
1350                 if (base_addr == MAP_FAILED ||
1351                     base_addr != mcfg->memseg[s].addr) {
1352                         max_seg = s;
1353                         if (base_addr != MAP_FAILED) {
1354                                 /* errno is stale, don't use */
1355                                 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1356                                         "in /dev/zero at [%p], got [%p] - "
1357                                         "please use '--base-virtaddr' option\n",
1358                                         (unsigned long long)mcfg->memseg[s].len,
1359                                         mcfg->memseg[s].addr, base_addr);
1360                                 munmap(base_addr, mcfg->memseg[s].len);
1361                         } else {
1362                                 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1363                                         "in /dev/zero at [%p]: '%s'\n",
1364                                         (unsigned long long)mcfg->memseg[s].len,
1365                                         mcfg->memseg[s].addr, strerror(errno));
1366                         }
1367                         if (aslr_enabled() > 0) {
1368                                 RTE_LOG(ERR, EAL, "It is recommended to "
1369                                         "disable ASLR in the kernel "
1370                                         "and retry running both primary "
1371                                         "and secondary processes\n");
1372                         }
1373                         goto error;
1374                 }
1375         }
1376
1377         size = getFileSize(fd_hugepage);
1378         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1379         if (hp == MAP_FAILED) {
1380                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1381                 goto error;
1382         }
1383
1384         num_hp = size / sizeof(struct hugepage_file);
1385         RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1386
1387         s = 0;
1388         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1389                 void *addr, *base_addr;
1390                 uintptr_t offset = 0;
1391                 size_t mapping_size;
1392                 /*
1393                  * free previously mapped memory so we can map the
1394                  * hugepages into the space
1395                  */
1396                 base_addr = mcfg->memseg[s].addr;
1397                 munmap(base_addr, mcfg->memseg[s].len);
1398
1399                 /* find the hugepages for this segment and map them
1400                  * we don't need to worry about order, as the server sorted the
1401                  * entries before it did the second mmap of them */
1402                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1403                         if (hp[i].memseg_id == (int)s){
1404                                 fd = open(hp[i].filepath, O_RDWR);
1405                                 if (fd < 0) {
1406                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1407                                                 hp[i].filepath);
1408                                         goto error;
1409                                 }
1410                                 mapping_size = hp[i].size;
1411                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1412                                                 mapping_size, PROT_READ | PROT_WRITE,
1413                                                 MAP_SHARED, fd, 0);
1414                                 close(fd); /* close file both on success and on failure */
1415                                 if (addr == MAP_FAILED ||
1416                                                 addr != RTE_PTR_ADD(base_addr, offset)) {
1417                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1418                                                 hp[i].filepath);
1419                                         goto error;
1420                                 }
1421                                 offset+=mapping_size;
1422                         }
1423                 }
1424                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1425                                 (unsigned long long)mcfg->memseg[s].len);
1426                 s++;
1427         }
1428         /* unmap the hugepage config file, since we are done using it */
1429         munmap(hp, size);
1430         close(fd_zero);
1431         close(fd_hugepage);
1432         return 0;
1433
1434 error:
1435         for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1436                 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1437         if (hp != NULL && hp != MAP_FAILED)
1438                 munmap(hp, size);
1439         if (fd_zero >= 0)
1440                 close(fd_zero);
1441         if (fd_hugepage >= 0)
1442                 close(fd_hugepage);
1443         return -1;
1444 }
1445
1446 bool
1447 rte_eal_using_phys_addrs(void)
1448 {
1449         return phys_addrs_available;
1450 }