lib/librte_eal/linuxapp/eal/eal_memory.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   Copyright(c) 2013 6WIND.
   6  *   All rights reserved.
   7  *
   8  *   Redistribution and use in source and binary forms, with or without
   9  *   modification, are permitted provided that the following conditions
  10  *   are met:
  11  *
  12  *     * Redistributions of source code must retain the above copyright
  13  *       notice, this list of conditions and the following disclaimer.
  14  *     * Redistributions in binary form must reproduce the above copyright
  15  *       notice, this list of conditions and the following disclaimer in
  16  *       the documentation and/or other materials provided with the
  17  *       distribution.
  18  *     * Neither the name of Intel Corporation nor the names of its
  19  *       contributors may be used to endorse or promote products derived
  20  *       from this software without specific prior written permission.
  21  *
  22  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  25  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  26  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  27  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  28  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  29  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  30  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  31  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  32  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  33  */
  34
  35 #define _FILE_OFFSET_BITS 64
  36 #include <errno.h>
  37 #include <stdarg.h>
  38 #include <stdbool.h>
  39 #include <stdlib.h>
  40 #include <stdio.h>
  41 #include <stdint.h>
  42 #include <inttypes.h>
  43 #include <string.h>
  44 #include <sys/mman.h>
  45 #include <sys/types.h>
  46 #include <sys/stat.h>
  47 #include <sys/queue.h>
  48 #include <sys/file.h>
  49 #include <unistd.h>
  50 #include <limits.h>
  51 #include <sys/ioctl.h>
  52 #include <sys/time.h>
  53 #include <signal.h>
  54 #include <setjmp.h>
  55 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
  56 #include <numa.h>
  57 #include <numaif.h>
  58 #endif
  59
  60 #include <rte_log.h>
  61 #include <rte_memory.h>
  62 #include <rte_launch.h>
  63 #include <rte_eal.h>
  64 #include <rte_eal_memconfig.h>
  65 #include <rte_per_lcore.h>
  66 #include <rte_lcore.h>
  67 #include <rte_common.h>
  68 #include <rte_string_fns.h>
  69
  70 #include "eal_private.h"
  71 #include "eal_internal_cfg.h"
  72 #include "eal_filesystem.h"
  73 #include "eal_hugepages.h"
  74
  75 #define PFN_MASK_SIZE   8
  76
  77 /**
  78  * @file
  79  * Huge page mapping under linux
  80  *
  81  * To reserve a big contiguous amount of memory, we use the hugepage
  82  * feature of linux. For that, we need to have hugetlbfs mounted. This
  83  * code will create many files in this directory (one per page) and
  84  * map them in virtual memory. For each page, we will retrieve its
  85  * physical address and remap it in order to have a virtual contiguous
  86  * zone as well as a physical contiguous zone.
  87  */
  88
  89 static uint64_t baseaddr_offset;
  90
  91 static bool phys_addrs_available = true;
  92
  93 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
  94
  95 static void
  96 test_phys_addrs_available(void)
  97 {
  98         uint64_t tmp;
  99         phys_addr_t physaddr;
 100
 101         if (!rte_eal_has_hugepages()) {
 102                 RTE_LOG(ERR, EAL,
 103                         "Started without hugepages support, physical addresses not available\n");
 104                 phys_addrs_available = false;
 105                 return;
 106         }
 107
 108         physaddr = rte_mem_virt2phy(&tmp);
 109         if (physaddr == RTE_BAD_PHYS_ADDR) {
 110                 if (rte_eal_iova_mode() == RTE_IOVA_PA)
 111                         RTE_LOG(ERR, EAL,
 112                                 "Cannot obtain physical addresses: %s. "
 113                                 "Only vfio will function.\n",
 114                                 strerror(errno));
 115                 phys_addrs_available = false;
 116         }
 117 }
 118
 119 /*
 120  * Get physical address of any mapped virtual address in the current process.
 121  */
 122 phys_addr_t
 123 rte_mem_virt2phy(const void *virtaddr)
 124 {
 125         int fd, retval;
 126         uint64_t page, physaddr;
 127         unsigned long virt_pfn;
 128         int page_size;
 129         off_t offset;
 130
 131         /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
 132         if (!phys_addrs_available)
 133                 return RTE_BAD_IOVA;
 134
 135         /* standard page size */
 136         page_size = getpagesize();
 137
 138         fd = open("/proc/self/pagemap", O_RDONLY);
 139         if (fd < 0) {
 140                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 141                         __func__, strerror(errno));
 142                 return RTE_BAD_IOVA;
 143         }
 144
 145         virt_pfn = (unsigned long)virtaddr / page_size;
 146         offset = sizeof(uint64_t) * virt_pfn;
 147         if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 148                 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 149                                 __func__, strerror(errno));
 150                 close(fd);
 151                 return RTE_BAD_IOVA;
 152         }
 153
 154         retval = read(fd, &page, PFN_MASK_SIZE);
 155         close(fd);
 156         if (retval < 0) {
 157                 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 158                                 __func__, strerror(errno));
 159                 return RTE_BAD_IOVA;
 160         } else if (retval != PFN_MASK_SIZE) {
 161                 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
 162                                 "but expected %d:\n",
 163                                 __func__, retval, PFN_MASK_SIZE);
 164                 return RTE_BAD_IOVA;
 165         }
 166
 167         /*
 168          * the pfn (page frame number) are bits 0-54 (see
 169          * pagemap.txt in linux Documentation)
 170          */
 171         if ((page & 0x7fffffffffffffULL) == 0)
 172                 return RTE_BAD_IOVA;
 173
 174         physaddr = ((page & 0x7fffffffffffffULL) * page_size)
 175                 + ((unsigned long)virtaddr % page_size);
 176
 177         return physaddr;
 178 }
 179
 180 rte_iova_t
 181 rte_mem_virt2iova(const void *virtaddr)
 182 {
 183         if (rte_eal_iova_mode() == RTE_IOVA_VA)
 184                 return (uintptr_t)virtaddr;
 185         return rte_mem_virt2phy(virtaddr);
 186 }
 187
 188 /*
 189  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 190  * it by browsing the /proc/self/pagemap special file.
 191  */
 192 static int
 193 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 194 {
 195         unsigned int i;
 196         phys_addr_t addr;
 197
 198         for (i = 0; i < hpi->num_pages[0]; i++) {
 199                 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
 200                 if (addr == RTE_BAD_PHYS_ADDR)
 201                         return -1;
 202                 hugepg_tbl[i].physaddr = addr;
 203         }
 204         return 0;
 205 }
 206
 207 /*
 208  * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
 209  */
 210 static int
 211 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 212 {
 213         unsigned int i;
 214         static phys_addr_t addr;
 215
 216         for (i = 0; i < hpi->num_pages[0]; i++) {
 217                 hugepg_tbl[i].physaddr = addr;
 218                 addr += hugepg_tbl[i].size;
 219         }
 220         return 0;
 221 }
 222
 223 /*
 224  * Check whether address-space layout randomization is enabled in
 225  * the kernel. This is important for multi-process as it can prevent
 226  * two processes mapping data to the same virtual address
 227  * Returns:
 228  *    0 - address space randomization disabled
 229  *    1/2 - address space randomization enabled
 230  *    negative error code on error
 231  */
 232 static int
 233 aslr_enabled(void)
 234 {
 235         char c;
 236         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 237         if (fd < 0)
 238                 return -errno;
 239         retval = read(fd, &c, 1);
 240         close(fd);
 241         if (retval < 0)
 242                 return -errno;
 243         if (retval == 0)
 244                 return -EIO;
 245         switch (c) {
 246                 case '0' : return 0;
 247                 case '1' : return 1;
 248                 case '2' : return 2;
 249                 default: return -EINVAL;
 250         }
 251 }
 252
 253 /*
 254  * Try to mmap *size bytes in /dev/zero. If it is successful, return the
 255  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 256  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 257  * 0. In this case, return NULL. Note: this function returns an address
 258  * which is a multiple of hugepage size.
 259  */
 260 static void *
 261 get_virtual_area(size_t *size, size_t hugepage_sz)
 262 {
 263         void *addr;
 264         int fd;
 265         long aligned_addr;
 266
 267         if (internal_config.base_virtaddr != 0) {
 268                 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
 269                                 baseaddr_offset);
 270         }
 271         else addr = NULL;
 272
 273         RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
 274
 275         fd = open("/dev/zero", O_RDONLY);
 276         if (fd < 0){
 277                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 278                 return NULL;
 279         }
 280         do {
 281                 addr = mmap(addr,
 282                                 (*size) + hugepage_sz, PROT_READ,
 283 #ifdef RTE_ARCH_PPC_64
 284                                 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
 285 #else
 286                                 MAP_PRIVATE,
 287 #endif
 288                                 fd, 0);
 289                 if (addr == MAP_FAILED)
 290                         *size -= hugepage_sz;
 291         } while (addr == MAP_FAILED && *size > 0);
 292
 293         if (addr == MAP_FAILED) {
 294                 close(fd);
 295                 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
 296                         strerror(errno));
 297                 return NULL;
 298         }
 299
 300         munmap(addr, (*size) + hugepage_sz);
 301         close(fd);
 302
 303         /* align addr to a huge page size boundary */
 304         aligned_addr = (long)addr;
 305         aligned_addr += (hugepage_sz - 1);
 306         aligned_addr &= (~(hugepage_sz - 1));
 307         addr = (void *)(aligned_addr);
 308
 309         RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
 310                 addr, *size);
 311
 312         /* increment offset */
 313         baseaddr_offset += *size;
 314
 315         return addr;
 316 }
 317
 318 static sigjmp_buf huge_jmpenv;
 319
 320 static void huge_sigbus_handler(int signo __rte_unused)
 321 {
 322         siglongjmp(huge_jmpenv, 1);
 323 }
 324
 325 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
 326  * non-static local variable in the stack frame calling sigsetjmp might be
 327  * clobbered by a call to longjmp.
 328  */
 329 static int huge_wrap_sigsetjmp(void)
 330 {
 331         return sigsetjmp(huge_jmpenv, 1);
 332 }
 333
 334 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 335 /* Callback for numa library. */
 336 void numa_error(char *where)
 337 {
 338         RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
 339 }
 340 #endif
 341
 342 /*
 343  * Mmap all hugepages of hugepage table: it first open a file in
 344  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 345  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 346  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 347  * map contiguous physical blocks in contiguous virtual blocks.
 348  */
 349 static unsigned
 350 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
 351                   uint64_t *essential_memory __rte_unused, int orig)
 352 {
 353         int fd;
 354         unsigned i;
 355         void *virtaddr;
 356         void *vma_addr = NULL;
 357         size_t vma_len = 0;
 358 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 359         int node_id = -1;
 360         int essential_prev = 0;
 361         int oldpolicy;
 362         struct bitmask *oldmask = numa_allocate_nodemask();
 363         bool have_numa = true;
 364         unsigned long maxnode = 0;
 365
 366         /* Check if kernel supports NUMA. */
 367         if (numa_available() != 0) {
 368                 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
 369                 have_numa = false;
 370         }
 371
 372         if (orig && have_numa) {
 373                 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
 374                 if (get_mempolicy(&oldpolicy, oldmask->maskp,
 375                                   oldmask->size + 1, 0, 0) < 0) {
 376                         RTE_LOG(ERR, EAL,
 377                                 "Failed to get current mempolicy: %s. "
 378                                 "Assuming MPOL_DEFAULT.\n", strerror(errno));
 379                         oldpolicy = MPOL_DEFAULT;
 380                 }
 381                 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 382                         if (internal_config.socket_mem[i])
 383                                 maxnode = i + 1;
 384         }
 385 #endif
 386
 387         for (i = 0; i < hpi->num_pages[0]; i++) {
 388                 uint64_t hugepage_sz = hpi->hugepage_sz;
 389
 390 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 391                 if (maxnode) {
 392                         unsigned int j;
 393
 394                         for (j = 0; j < maxnode; j++)
 395                                 if (essential_memory[j])
 396                                         break;
 397
 398                         if (j == maxnode) {
 399                                 node_id = (node_id + 1) % maxnode;
 400                                 while (!internal_config.socket_mem[node_id]) {
 401                                         node_id++;
 402                                         node_id %= maxnode;
 403                                 }
 404                                 essential_prev = 0;
 405                         } else {
 406                                 node_id = j;
 407                                 essential_prev = essential_memory[j];
 408
 409                                 if (essential_memory[j] < hugepage_sz)
 410                                         essential_memory[j] = 0;
 411                                 else
 412                                         essential_memory[j] -= hugepage_sz;
 413                         }
 414
 415                         RTE_LOG(DEBUG, EAL,
 416                                 "Setting policy MPOL_PREFERRED for socket %d\n",
 417                                 node_id);
 418                         numa_set_preferred(node_id);
 419                 }
 420 #endif
 421
 422                 if (orig) {
 423                         hugepg_tbl[i].file_id = i;
 424                         hugepg_tbl[i].size = hugepage_sz;
 425                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 426                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 427                                         hugepg_tbl[i].file_id);
 428                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 429                 }
 430 #ifndef RTE_ARCH_64
 431                 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
 432                  * original map address as final map address.
 433                  */
 434                 else if ((hugepage_sz == RTE_PGSIZE_1G)
 435                         || (hugepage_sz == RTE_PGSIZE_16G)) {
 436                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 437                         hugepg_tbl[i].orig_va = NULL;
 438                         continue;
 439                 }
 440 #endif
 441                 else if (vma_len == 0) {
 442                         unsigned j, num_pages;
 443
 444                         /* reserve a virtual area for next contiguous
 445                          * physical block: count the number of
 446                          * contiguous physical pages. */
 447                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 448 #ifdef RTE_ARCH_PPC_64
 449                                 /* The physical addresses are sorted in
 450                                  * descending order on PPC64 */
 451                                 if (hugepg_tbl[j].physaddr !=
 452                                     hugepg_tbl[j-1].physaddr - hugepage_sz)
 453                                         break;
 454 #else
 455                                 if (hugepg_tbl[j].physaddr !=
 456                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 457                                         break;
 458 #endif
 459                         }
 460                         num_pages = j - i;
 461                         vma_len = num_pages * hugepage_sz;
 462
 463                         /* get the biggest virtual memory area up to
 464                          * vma_len. If it fails, vma_addr is NULL, so
 465                          * let the kernel provide the address. */
 466                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 467                         if (vma_addr == NULL)
 468                                 vma_len = hugepage_sz;
 469                 }
 470
 471                 /* try to create hugepage file */
 472                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
 473                 if (fd < 0) {
 474                         RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
 475                                         strerror(errno));
 476                         goto out;
 477                 }
 478
 479                 /* map the segment, and populate page tables,
 480                  * the kernel fills this segment with zeros */
 481                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 482                                 MAP_SHARED | MAP_POPULATE, fd, 0);
 483                 if (virtaddr == MAP_FAILED) {
 484                         RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
 485                                         strerror(errno));
 486                         close(fd);
 487                         goto out;
 488                 }
 489
 490                 if (orig) {
 491                         hugepg_tbl[i].orig_va = virtaddr;
 492                 }
 493                 else {
 494                         hugepg_tbl[i].final_va = virtaddr;
 495                 }
 496
 497                 if (orig) {
 498                         /* In linux, hugetlb limitations, like cgroup, are
 499                          * enforced at fault time instead of mmap(), even
 500                          * with the option of MAP_POPULATE. Kernel will send
 501                          * a SIGBUS signal. To avoid to be killed, save stack
 502                          * environment here, if SIGBUS happens, we can jump
 503                          * back here.
 504                          */
 505                         if (huge_wrap_sigsetjmp()) {
 506                                 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
 507                                         "hugepages of size %u MB\n",
 508                                         (unsigned)(hugepage_sz / 0x100000));
 509                                 munmap(virtaddr, hugepage_sz);
 510                                 close(fd);
 511                                 unlink(hugepg_tbl[i].filepath);
 512 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 513                                 if (maxnode)
 514                                         essential_memory[node_id] =
 515                                                 essential_prev;
 516 #endif
 517                                 goto out;
 518                         }
 519                         *(int *)virtaddr = 0;
 520                 }
 521
 522
 523                 /* set shared flock on the file. */
 524                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 525                         RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
 526                                 __func__, strerror(errno));
 527                         close(fd);
 528                         goto out;
 529                 }
 530
 531                 close(fd);
 532
 533                 vma_addr = (char *)vma_addr + hugepage_sz;
 534                 vma_len -= hugepage_sz;
 535         }
 536
 537 out:
 538 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 539         if (maxnode) {
 540                 RTE_LOG(DEBUG, EAL,
 541                         "Restoring previous memory policy: %d\n", oldpolicy);
 542                 if (oldpolicy == MPOL_DEFAULT) {
 543                         numa_set_localalloc();
 544                 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
 545                                          oldmask->size + 1) < 0) {
 546                         RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
 547                                 strerror(errno));
 548                         numa_set_localalloc();
 549                 }
 550         }
 551         numa_free_cpumask(oldmask);
 552 #endif
 553         return i;
 554 }
 555
 556 /* Unmap all hugepages from original mapping */
 557 static int
 558 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 559 {
 560         unsigned i;
 561         for (i = 0; i < hpi->num_pages[0]; i++) {
 562                 if (hugepg_tbl[i].orig_va) {
 563                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 564                         hugepg_tbl[i].orig_va = NULL;
 565                 }
 566         }
 567         return 0;
 568 }
 569
 570 /*
 571  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 572  * page.
 573  */
 574 static int
 575 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 576 {
 577         int socket_id;
 578         char *end, *nodestr;
 579         unsigned i, hp_count = 0;
 580         uint64_t virt_addr;
 581         char buf[BUFSIZ];
 582         char hugedir_str[PATH_MAX];
 583         FILE *f;
 584
 585         f = fopen("/proc/self/numa_maps", "r");
 586         if (f == NULL) {
 587                 RTE_LOG(NOTICE, EAL, "NUMA support not available"
 588                         " consider that all memory is in socket_id 0\n");
 589                 return 0;
 590         }
 591
 592         snprintf(hugedir_str, sizeof(hugedir_str),
 593                         "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
 594
 595         /* parse numa map */
 596         while (fgets(buf, sizeof(buf), f) != NULL) {
 597
 598                 /* ignore non huge page */
 599                 if (strstr(buf, " huge ") == NULL &&
 600                                 strstr(buf, hugedir_str) == NULL)
 601                         continue;
 602
 603                 /* get zone addr */
 604                 virt_addr = strtoull(buf, &end, 16);
 605                 if (virt_addr == 0 || end == buf) {
 606                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 607                         goto error;
 608                 }
 609
 610                 /* get node id (socket id) */
 611                 nodestr = strstr(buf, " N");
 612                 if (nodestr == NULL) {
 613                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 614                         goto error;
 615                 }
 616                 nodestr += 2;
 617                 end = strstr(nodestr, "=");
 618                 if (end == NULL) {
 619                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 620                         goto error;
 621                 }
 622                 end[0] = '\0';
 623                 end = NULL;
 624
 625                 socket_id = strtoul(nodestr, &end, 0);
 626                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 627                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 628                         goto error;
 629                 }
 630
 631                 /* if we find this page in our mappings, set socket_id */
 632                 for (i = 0; i < hpi->num_pages[0]; i++) {
 633                         void *va = (void *)(unsigned long)virt_addr;
 634                         if (hugepg_tbl[i].orig_va == va) {
 635                                 hugepg_tbl[i].socket_id = socket_id;
 636                                 hp_count++;
 637 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 638                                 RTE_LOG(DEBUG, EAL,
 639                                         "Hugepage %s is on socket %d\n",
 640                                         hugepg_tbl[i].filepath, socket_id);
 641 #endif
 642                         }
 643                 }
 644         }
 645
 646         if (hp_count < hpi->num_pages[0])
 647                 goto error;
 648
 649         fclose(f);
 650         return 0;
 651
 652 error:
 653         fclose(f);
 654         return -1;
 655 }
 656
 657 static int
 658 cmp_physaddr(const void *a, const void *b)
 659 {
 660 #ifndef RTE_ARCH_PPC_64
 661         const struct hugepage_file *p1 = a;
 662         const struct hugepage_file *p2 = b;
 663 #else
 664         /* PowerPC needs memory sorted in reverse order from x86 */
 665         const struct hugepage_file *p1 = b;
 666         const struct hugepage_file *p2 = a;
 667 #endif
 668         if (p1->physaddr < p2->physaddr)
 669                 return -1;
 670         else if (p1->physaddr > p2->physaddr)
 671                 return 1;
 672         else
 673                 return 0;
 674 }
 675
 676 /*
 677  * Uses mmap to create a shared memory area for storage of data
 678  * Used in this file to store the hugepage file map on disk
 679  */
 680 static void *
 681 create_shared_memory(const char *filename, const size_t mem_size)
 682 {
 683         void *retval;
 684         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 685         if (fd < 0)
 686                 return NULL;
 687         if (ftruncate(fd, mem_size) < 0) {
 688                 close(fd);
 689                 return NULL;
 690         }
 691         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 692         close(fd);
 693         if (retval == MAP_FAILED)
 694                 return NULL;
 695         return retval;
 696 }
 697
 698 /*
 699  * this copies *active* hugepages from one hugepage table to another.
 700  * destination is typically the shared memory.
 701  */
 702 static int
 703 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
 704                 const struct hugepage_file * src, int src_size)
 705 {
 706         int src_pos, dst_pos = 0;
 707
 708         for (src_pos = 0; src_pos < src_size; src_pos++) {
 709                 if (src[src_pos].final_va != NULL) {
 710                         /* error on overflow attempt */
 711                         if (dst_pos == dest_size)
 712                                 return -1;
 713                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
 714                         dst_pos++;
 715                 }
 716         }
 717         return 0;
 718 }
 719
 720 static int
 721 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
 722                 unsigned num_hp_info)
 723 {
 724         unsigned socket, size;
 725         int page, nrpages = 0;
 726
 727         /* get total number of hugepages */
 728         for (size = 0; size < num_hp_info; size++)
 729                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 730                         nrpages +=
 731                         internal_config.hugepage_info[size].num_pages[socket];
 732
 733         for (page = 0; page < nrpages; page++) {
 734                 struct hugepage_file *hp = &hugepg_tbl[page];
 735
 736                 if (hp->final_va != NULL && unlink(hp->filepath)) {
 737                         RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
 738                                 __func__, hp->filepath, strerror(errno));
 739                 }
 740         }
 741         return 0;
 742 }
 743
 744 /*
 745  * unmaps hugepages that are not going to be used. since we originally allocate
 746  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 747  */
 748 static int
 749 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
 750                 struct hugepage_info *hpi,
 751                 unsigned num_hp_info)
 752 {
 753         unsigned socket, size;
 754         int page, nrpages = 0;
 755
 756         /* get total number of hugepages */
 757         for (size = 0; size < num_hp_info; size++)
 758                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 759                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 760
 761         for (size = 0; size < num_hp_info; size++) {
 762                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 763                         unsigned pages_found = 0;
 764
 765                         /* traverse until we have unmapped all the unused pages */
 766                         for (page = 0; page < nrpages; page++) {
 767                                 struct hugepage_file *hp = &hugepg_tbl[page];
 768
 769                                 /* find a page that matches the criteria */
 770                                 if ((hp->size == hpi[size].hugepage_sz) &&
 771                                                 (hp->socket_id == (int) socket)) {
 772
 773                                         /* if we skipped enough pages, unmap the rest */
 774                                         if (pages_found == hpi[size].num_pages[socket]) {
 775                                                 uint64_t unmap_len;
 776
 777                                                 unmap_len = hp->size;
 778
 779                                                 /* get start addr and len of the remaining segment */
 780                                                 munmap(hp->final_va, (size_t) unmap_len);
 781
 782                                                 hp->final_va = NULL;
 783                                                 if (unlink(hp->filepath) == -1) {
 784                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 785                                                                         __func__, hp->filepath, strerror(errno));
 786                                                         return -1;
 787                                                 }
 788                                         } else {
 789                                                 /* lock the page and skip */
 790                                                 pages_found++;
 791                                         }
 792
 793                                 } /* match page */
 794                         } /* foreach page */
 795                 } /* foreach socket */
 796         } /* foreach pagesize */
 797
 798         return 0;
 799 }
 800
 801 static inline uint64_t
 802 get_socket_mem_size(int socket)
 803 {
 804         uint64_t size = 0;
 805         unsigned i;
 806
 807         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 808                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 809                 if (hpi->hugedir != NULL)
 810                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 811         }
 812
 813         return size;
 814 }
 815
 816 /*
 817  * This function is a NUMA-aware equivalent of calc_num_pages.
 818  * It takes in the list of hugepage sizes and the
 819  * number of pages thereof, and calculates the best number of
 820  * pages of each size to fulfill the request for <memory> ram
 821  */
 822 static int
 823 calc_num_pages_per_socket(uint64_t * memory,
 824                 struct hugepage_info *hp_info,
 825                 struct hugepage_info *hp_used,
 826                 unsigned num_hp_info)
 827 {
 828         unsigned socket, j, i = 0;
 829         unsigned requested, available;
 830         int total_num_pages = 0;
 831         uint64_t remaining_mem, cur_mem;
 832         uint64_t total_mem = internal_config.memory;
 833
 834         if (num_hp_info == 0)
 835                 return -1;
 836
 837         /* if specific memory amounts per socket weren't requested */
 838         if (internal_config.force_sockets == 0) {
 839                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
 840                 size_t default_size, total_size;
 841                 unsigned lcore_id;
 842
 843                 /* Compute number of cores per socket */
 844                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
 845                 RTE_LCORE_FOREACH(lcore_id) {
 846                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
 847                 }
 848
 849                 /*
 850                  * Automatically spread requested memory amongst detected sockets according
 851                  * to number of cores from cpu mask present on each socket
 852                  */
 853                 total_size = internal_config.memory;
 854                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 855
 856                         /* Set memory amount per socket */
 857                         default_size = (internal_config.memory * cpu_per_socket[socket])
 858                                         / rte_lcore_count();
 859
 860                         /* Limit to maximum available memory on socket */
 861                         default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
 862
 863                         /* Update sizes */
 864                         memory[socket] = default_size;
 865                         total_size -= default_size;
 866                 }
 867
 868                 /*
 869                  * If some memory is remaining, try to allocate it by getting all
 870                  * available memory from sockets, one after the other
 871                  */
 872                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 873                         /* take whatever is available */
 874                         default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
 875                                                total_size);
 876
 877                         /* Update sizes */
 878                         memory[socket] += default_size;
 879                         total_size -= default_size;
 880                 }
 881         }
 882
 883         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 884                 /* skips if the memory on specific socket wasn't requested */
 885                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 886                         hp_used[i].hugedir = hp_info[i].hugedir;
 887                         hp_used[i].num_pages[socket] = RTE_MIN(
 888                                         memory[socket] / hp_info[i].hugepage_sz,
 889                                         hp_info[i].num_pages[socket]);
 890
 891                         cur_mem = hp_used[i].num_pages[socket] *
 892                                         hp_used[i].hugepage_sz;
 893
 894                         memory[socket] -= cur_mem;
 895                         total_mem -= cur_mem;
 896
 897                         total_num_pages += hp_used[i].num_pages[socket];
 898
 899                         /* check if we have met all memory requests */
 900                         if (memory[socket] == 0)
 901                                 break;
 902
 903                         /* check if we have any more pages left at this size, if so
 904                          * move on to next size */
 905                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
 906                                 continue;
 907                         /* At this point we know that there are more pages available that are
 908                          * bigger than the memory we want, so lets see if we can get enough
 909                          * from other page sizes.
 910                          */
 911                         remaining_mem = 0;
 912                         for (j = i+1; j < num_hp_info; j++)
 913                                 remaining_mem += hp_info[j].hugepage_sz *
 914                                 hp_info[j].num_pages[socket];
 915
 916                         /* is there enough other memory, if not allocate another page and quit */
 917                         if (remaining_mem < memory[socket]){
 918                                 cur_mem = RTE_MIN(memory[socket],
 919                                                 hp_info[i].hugepage_sz);
 920                                 memory[socket] -= cur_mem;
 921                                 total_mem -= cur_mem;
 922                                 hp_used[i].num_pages[socket]++;
 923                                 total_num_pages++;
 924                                 break; /* we are done with this socket*/
 925                         }
 926                 }
 927                 /* if we didn't satisfy all memory requirements per socket */
 928                 if (memory[socket] > 0) {
 929                         /* to prevent icc errors */
 930                         requested = (unsigned) (internal_config.socket_mem[socket] /
 931                                         0x100000);
 932                         available = requested -
 933                                         ((unsigned) (memory[socket] / 0x100000));
 934                         RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
 935                                         "Requested: %uMB, available: %uMB\n", socket,
 936                                         requested, available);
 937                         return -1;
 938                 }
 939         }
 940
 941         /* if we didn't satisfy total memory requirements */
 942         if (total_mem > 0) {
 943                 requested = (unsigned) (internal_config.memory / 0x100000);
 944                 available = requested - (unsigned) (total_mem / 0x100000);
 945                 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
 946                                 " available: %uMB\n", requested, available);
 947                 return -1;
 948         }
 949         return total_num_pages;
 950 }
 951
 952 static inline size_t
 953 eal_get_hugepage_mem_size(void)
 954 {
 955         uint64_t size = 0;
 956         unsigned i, j;
 957
 958         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 959                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 960                 if (hpi->hugedir != NULL) {
 961                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 962                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 963                         }
 964                 }
 965         }
 966
 967         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 968 }
 969
 970 static struct sigaction huge_action_old;
 971 static int huge_need_recover;
 972
 973 static void
 974 huge_register_sigbus(void)
 975 {
 976         sigset_t mask;
 977         struct sigaction action;
 978
 979         sigemptyset(&mask);
 980         sigaddset(&mask, SIGBUS);
 981         action.sa_flags = 0;
 982         action.sa_mask = mask;
 983         action.sa_handler = huge_sigbus_handler;
 984
 985         huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
 986 }
 987
 988 static void
 989 huge_recover_sigbus(void)
 990 {
 991         if (huge_need_recover) {
 992                 sigaction(SIGBUS, &huge_action_old, NULL);
 993                 huge_need_recover = 0;
 994         }
 995 }
 996
 997 /*
 998  * Prepare physical memory mapping: fill configuration structure with
 999  * these infos, return 0 on success.
1000  *  1. map N huge pages in separate files in hugetlbfs
1001  *  2. find associated physical addr
1002  *  3. find associated NUMA socket ID
1003  *  4. sort all huge pages by physical address
1004  *  5. remap these N huge pages in the correct order
1005  *  6. unmap the first mapping
1006  *  7. fill memsegs in configuration with contiguous zones
1007  */
1008 int
1009 rte_eal_hugepage_init(void)
1010 {
1011         struct rte_mem_config *mcfg;
1012         struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1013         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1014
1015         uint64_t memory[RTE_MAX_NUMA_NODES];
1016
1017         unsigned hp_offset;
1018         int i, j, new_memseg;
1019         int nr_hugefiles, nr_hugepages = 0;
1020         void *addr;
1021
1022         test_phys_addrs_available();
1023
1024         memset(used_hp, 0, sizeof(used_hp));
1025
1026         /* get pointer to global configuration */
1027         mcfg = rte_eal_get_configuration()->mem_config;
1028
1029         /* hugetlbfs can be disabled */
1030         if (internal_config.no_hugetlbfs) {
1031                 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1032                                 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1033                 if (addr == MAP_FAILED) {
1034                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1035                                         strerror(errno));
1036                         return -1;
1037                 }
1038                 if (rte_eal_iova_mode() == RTE_IOVA_VA)
1039                         mcfg->memseg[0].iova = (uintptr_t)addr;
1040                 else
1041                         mcfg->memseg[0].iova = RTE_BAD_IOVA;
1042                 mcfg->memseg[0].addr = addr;
1043                 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1044                 mcfg->memseg[0].len = internal_config.memory;
1045                 mcfg->memseg[0].socket_id = 0;
1046                 return 0;
1047         }
1048
1049         /* calculate total number of hugepages available. at this point we haven't
1050          * yet started sorting them so they all are on socket 0 */
1051         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1052                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1053                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1054
1055                 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1056         }
1057
1058         /*
1059          * allocate a memory area for hugepage table.
1060          * this isn't shared memory yet. due to the fact that we need some
1061          * processing done on these pages, shared memory will be created
1062          * at a later stage.
1063          */
1064         tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1065         if (tmp_hp == NULL)
1066                 goto fail;
1067
1068         memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1069
1070         hp_offset = 0; /* where we start the current page size entries */
1071
1072         huge_register_sigbus();
1073
1074         /* make a copy of socket_mem, needed for balanced allocation. */
1075         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1076                 memory[i] = internal_config.socket_mem[i];
1077
1078
1079         /* map all hugepages and sort them */
1080         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1081                 unsigned pages_old, pages_new;
1082                 struct hugepage_info *hpi;
1083
1084                 /*
1085                  * we don't yet mark hugepages as used at this stage, so
1086                  * we just map all hugepages available to the system
1087                  * all hugepages are still located on socket 0
1088                  */
1089                 hpi = &internal_config.hugepage_info[i];
1090
1091                 if (hpi->num_pages[0] == 0)
1092                         continue;
1093
1094                 /* map all hugepages available */
1095                 pages_old = hpi->num_pages[0];
1096                 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi,
1097                                               memory, 1);
1098                 if (pages_new < pages_old) {
1099                         RTE_LOG(DEBUG, EAL,
1100                                 "%d not %d hugepages of size %u MB allocated\n",
1101                                 pages_new, pages_old,
1102                                 (unsigned)(hpi->hugepage_sz / 0x100000));
1103
1104                         int pages = pages_old - pages_new;
1105
1106                         nr_hugepages -= pages;
1107                         hpi->num_pages[0] = pages_new;
1108                         if (pages_new == 0)
1109                                 continue;
1110                 }
1111
1112                 if (phys_addrs_available) {
1113                         /* find physical addresses for each hugepage */
1114                         if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1115                                 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1116                                         "for %u MB pages\n",
1117                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1118                                 goto fail;
1119                         }
1120                 } else {
1121                         /* set physical addresses for each hugepage */
1122                         if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1123                                 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1124                                         "for %u MB pages\n",
1125                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1126                                 goto fail;
1127                         }
1128                 }
1129
1130                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1131                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1132                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1133                         goto fail;
1134                 }
1135
1136                 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1137                       sizeof(struct hugepage_file), cmp_physaddr);
1138
1139                 /* remap all hugepages */
1140                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, NULL, 0) !=
1141                     hpi->num_pages[0]) {
1142                         RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1143                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1144                         goto fail;
1145                 }
1146
1147                 /* unmap original mappings */
1148                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1149                         goto fail;
1150
1151                 /* we have processed a num of hugepages of this size, so inc offset */
1152                 hp_offset += hpi->num_pages[0];
1153         }
1154
1155         huge_recover_sigbus();
1156
1157         if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1158                 internal_config.memory = eal_get_hugepage_mem_size();
1159
1160         nr_hugefiles = nr_hugepages;
1161
1162
1163         /* clean out the numbers of pages */
1164         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1165                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1166                         internal_config.hugepage_info[i].num_pages[j] = 0;
1167
1168         /* get hugepages for each socket */
1169         for (i = 0; i < nr_hugefiles; i++) {
1170                 int socket = tmp_hp[i].socket_id;
1171
1172                 /* find a hugepage info with right size and increment num_pages */
1173                 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1174                                 (int)internal_config.num_hugepage_sizes);
1175                 for (j = 0; j < nb_hpsizes; j++) {
1176                         if (tmp_hp[i].size ==
1177                                         internal_config.hugepage_info[j].hugepage_sz) {
1178                                 internal_config.hugepage_info[j].num_pages[socket]++;
1179                         }
1180                 }
1181         }
1182
1183         /* make a copy of socket_mem, needed for number of pages calculation */
1184         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1185                 memory[i] = internal_config.socket_mem[i];
1186
1187         /* calculate final number of pages */
1188         nr_hugepages = calc_num_pages_per_socket(memory,
1189                         internal_config.hugepage_info, used_hp,
1190                         internal_config.num_hugepage_sizes);
1191
1192         /* error if not enough memory available */
1193         if (nr_hugepages < 0)
1194                 goto fail;
1195
1196         /* reporting in! */
1197         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1198                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1199                         if (used_hp[i].num_pages[j] > 0) {
1200                                 RTE_LOG(DEBUG, EAL,
1201                                         "Requesting %u pages of size %uMB"
1202                                         " from socket %i\n",
1203                                         used_hp[i].num_pages[j],
1204                                         (unsigned)
1205                                         (used_hp[i].hugepage_sz / 0x100000),
1206                                         j);
1207                         }
1208                 }
1209         }
1210
1211         /* create shared memory */
1212         hugepage = create_shared_memory(eal_hugepage_info_path(),
1213                         nr_hugefiles * sizeof(struct hugepage_file));
1214
1215         if (hugepage == NULL) {
1216                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1217                 goto fail;
1218         }
1219         memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1220
1221         /*
1222          * unmap pages that we won't need (looks at used_hp).
1223          * also, sets final_va to NULL on pages that were unmapped.
1224          */
1225         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1226                         internal_config.num_hugepage_sizes) < 0) {
1227                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1228                 goto fail;
1229         }
1230
1231         /*
1232          * copy stuff from malloc'd hugepage* to the actual shared memory.
1233          * this procedure only copies those hugepages that have final_va
1234          * not NULL. has overflow protection.
1235          */
1236         if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1237                         tmp_hp, nr_hugefiles) < 0) {
1238                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1239                 goto fail;
1240         }
1241
1242         /* free the hugepage backing files */
1243         if (internal_config.hugepage_unlink &&
1244                 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1245                 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1246                 goto fail;
1247         }
1248
1249         /* free the temporary hugepage table */
1250         free(tmp_hp);
1251         tmp_hp = NULL;
1252
1253         /* first memseg index shall be 0 after incrementing it below */
1254         j = -1;
1255         for (i = 0; i < nr_hugefiles; i++) {
1256                 new_memseg = 0;
1257
1258                 /* if this is a new section, create a new memseg */
1259                 if (i == 0)
1260                         new_memseg = 1;
1261                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1262                         new_memseg = 1;
1263                 else if (hugepage[i].size != hugepage[i-1].size)
1264                         new_memseg = 1;
1265
1266 #ifdef RTE_ARCH_PPC_64
1267                 /* On PPC64 architecture, the mmap always start from higher
1268                  * virtual address to lower address. Here, both the physical
1269                  * address and virtual address are in descending order */
1270                 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1271                     hugepage[i].size)
1272                         new_memseg = 1;
1273                 else if (((unsigned long)hugepage[i-1].final_va -
1274                     (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1275                         new_memseg = 1;
1276 #else
1277                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1278                     hugepage[i].size)
1279                         new_memseg = 1;
1280                 else if (((unsigned long)hugepage[i].final_va -
1281                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1282                         new_memseg = 1;
1283 #endif
1284
1285                 if (new_memseg) {
1286                         j += 1;
1287                         if (j == RTE_MAX_MEMSEG)
1288                                 break;
1289
1290                         mcfg->memseg[j].iova = hugepage[i].physaddr;
1291                         mcfg->memseg[j].addr = hugepage[i].final_va;
1292                         mcfg->memseg[j].len = hugepage[i].size;
1293                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1294                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1295                 }
1296                 /* continuation of previous memseg */
1297                 else {
1298 #ifdef RTE_ARCH_PPC_64
1299                 /* Use the phy and virt address of the last page as segment
1300                  * address for IBM Power architecture */
1301                         mcfg->memseg[j].iova = hugepage[i].physaddr;
1302                         mcfg->memseg[j].addr = hugepage[i].final_va;
1303 #endif
1304                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1305                 }
1306                 hugepage[i].memseg_id = j;
1307         }
1308
1309         if (i < nr_hugefiles) {
1310                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1311                         "from %d requested\n"
1312                         "Current %s=%d is not enough\n"
1313                         "Please either increase it or request less amount "
1314                         "of memory.\n",
1315                         i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1316                         RTE_MAX_MEMSEG);
1317                 goto fail;
1318         }
1319
1320         munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1321
1322         return 0;
1323
1324 fail:
1325         huge_recover_sigbus();
1326         free(tmp_hp);
1327         if (hugepage != NULL)
1328                 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1329
1330         return -1;
1331 }
1332
1333 /*
1334  * uses fstat to report the size of a file on disk
1335  */
1336 static off_t
1337 getFileSize(int fd)
1338 {
1339         struct stat st;
1340         if (fstat(fd, &st) < 0)
1341                 return 0;
1342         return st.st_size;
1343 }
1344
1345 /*
1346  * This creates the memory mappings in the secondary process to match that of
1347  * the server process. It goes through each memory segment in the DPDK runtime
1348  * configuration and finds the hugepages which form that segment, mapping them
1349  * in order to form a contiguous block in the virtual memory space
1350  */
1351 int
1352 rte_eal_hugepage_attach(void)
1353 {
1354         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1355         struct hugepage_file *hp = NULL;
1356         unsigned num_hp = 0;
1357         unsigned i, s = 0; /* s used to track the segment number */
1358         unsigned max_seg = RTE_MAX_MEMSEG;
1359         off_t size = 0;
1360         int fd, fd_zero = -1, fd_hugepage = -1;
1361
1362         if (aslr_enabled() > 0) {
1363                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1364                                 "(ASLR) is enabled in the kernel.\n");
1365                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1366                                 "into secondary processes\n");
1367         }
1368
1369         test_phys_addrs_available();
1370
1371         fd_zero = open("/dev/zero", O_RDONLY);
1372         if (fd_zero < 0) {
1373                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1374                 goto error;
1375         }
1376         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1377         if (fd_hugepage < 0) {
1378                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1379                 goto error;
1380         }
1381
1382         /* map all segments into memory to make sure we get the addrs */
1383         for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1384                 void *base_addr;
1385
1386                 /*
1387                  * the first memory segment with len==0 is the one that
1388                  * follows the last valid segment.
1389                  */
1390                 if (mcfg->memseg[s].len == 0)
1391                         break;
1392
1393                 /*
1394                  * fdzero is mmapped to get a contiguous block of virtual
1395                  * addresses of the appropriate memseg size.
1396                  * use mmap to get identical addresses as the primary process.
1397                  */
1398                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1399                                  PROT_READ,
1400 #ifdef RTE_ARCH_PPC_64
1401                                  MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
1402 #else
1403                                  MAP_PRIVATE,
1404 #endif
1405                                  fd_zero, 0);
1406                 if (base_addr == MAP_FAILED ||
1407                     base_addr != mcfg->memseg[s].addr) {
1408                         max_seg = s;
1409                         if (base_addr != MAP_FAILED) {
1410                                 /* errno is stale, don't use */
1411                                 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1412                                         "in /dev/zero at [%p], got [%p] - "
1413                                         "please use '--base-virtaddr' option\n",
1414                                         (unsigned long long)mcfg->memseg[s].len,
1415                                         mcfg->memseg[s].addr, base_addr);
1416                                 munmap(base_addr, mcfg->memseg[s].len);
1417                         } else {
1418                                 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1419                                         "in /dev/zero at [%p]: '%s'\n",
1420                                         (unsigned long long)mcfg->memseg[s].len,
1421                                         mcfg->memseg[s].addr, strerror(errno));
1422                         }
1423                         if (aslr_enabled() > 0) {
1424                                 RTE_LOG(ERR, EAL, "It is recommended to "
1425                                         "disable ASLR in the kernel "
1426                                         "and retry running both primary "
1427                                         "and secondary processes\n");
1428                         }
1429                         goto error;
1430                 }
1431         }
1432
1433         size = getFileSize(fd_hugepage);
1434         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1435         if (hp == MAP_FAILED) {
1436                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1437                 goto error;
1438         }
1439
1440         num_hp = size / sizeof(struct hugepage_file);
1441         RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1442
1443         s = 0;
1444         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1445                 void *addr, *base_addr;
1446                 uintptr_t offset = 0;
1447                 size_t mapping_size;
1448                 /*
1449                  * free previously mapped memory so we can map the
1450                  * hugepages into the space
1451                  */
1452                 base_addr = mcfg->memseg[s].addr;
1453                 munmap(base_addr, mcfg->memseg[s].len);
1454
1455                 /* find the hugepages for this segment and map them
1456                  * we don't need to worry about order, as the server sorted the
1457                  * entries before it did the second mmap of them */
1458                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1459                         if (hp[i].memseg_id == (int)s){
1460                                 fd = open(hp[i].filepath, O_RDWR);
1461                                 if (fd < 0) {
1462                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1463                                                 hp[i].filepath);
1464                                         goto error;
1465                                 }
1466                                 mapping_size = hp[i].size;
1467                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1468                                                 mapping_size, PROT_READ | PROT_WRITE,
1469                                                 MAP_SHARED, fd, 0);
1470                                 close(fd); /* close file both on success and on failure */
1471                                 if (addr == MAP_FAILED ||
1472                                                 addr != RTE_PTR_ADD(base_addr, offset)) {
1473                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1474                                                 hp[i].filepath);
1475                                         goto error;
1476                                 }
1477                                 offset+=mapping_size;
1478                         }
1479                 }
1480                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1481                                 (unsigned long long)mcfg->memseg[s].len);
1482                 s++;
1483         }
1484         /* unmap the hugepage config file, since we are done using it */
1485         munmap(hp, size);
1486         close(fd_zero);
1487         close(fd_hugepage);
1488         return 0;
1489
1490 error:
1491         for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1492                 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1493         if (hp != NULL && hp != MAP_FAILED)
1494                 munmap(hp, size);
1495         if (fd_zero >= 0)
1496                 close(fd_zero);
1497         if (fd_hugepage >= 0)
1498                 close(fd_hugepage);
1499         return -1;
1500 }
1501
1502 int
1503 rte_eal_using_phys_addrs(void)
1504 {
1505         return phys_addrs_available;
1506 }