lib/librte_eal/linuxapp/eal/eal_memory.c

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