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