lib/librte_eal/linuxapp/eal/eal_memory.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33 /*   BSD LICENSE
  34  *
  35  *   Copyright(c) 2013 6WIND.
  36  *
  37  *   Redistribution and use in source and binary forms, with or without
  38  *   modification, are permitted provided that the following conditions
  39  *   are met:
  40  *
  41  *     * Redistributions of source code must retain the above copyright
  42  *       notice, this list of conditions and the following disclaimer.
  43  *     * Redistributions in binary form must reproduce the above copyright
  44  *       notice, this list of conditions and the following disclaimer in
  45  *       the documentation and/or other materials provided with the
  46  *       distribution.
  47  *     * Neither the name of 6WIND S.A. nor the names of its
  48  *       contributors may be used to endorse or promote products derived
  49  *       from this software without specific prior written permission.
  50  *
  51  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  52  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  53  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  54  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  55  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  56  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  57  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  58  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  59  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  60  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  61  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  62  */
  63
  64 #define _FILE_OFFSET_BITS 64
  65 #include <errno.h>
  66 #include <stdarg.h>
  67 #include <stdbool.h>
  68 #include <stdlib.h>
  69 #include <stdio.h>
  70 #include <stdint.h>
  71 #include <inttypes.h>
  72 #include <string.h>
  73 #include <stdarg.h>
  74 #include <sys/mman.h>
  75 #include <sys/types.h>
  76 #include <sys/stat.h>
  77 #include <sys/queue.h>
  78 #include <sys/file.h>
  79 #include <unistd.h>
  80 #include <limits.h>
  81 #include <errno.h>
  82 #include <sys/ioctl.h>
  83 #include <sys/time.h>
  84 #include <signal.h>
  85 #include <setjmp.h>
  86
  87 #include <rte_log.h>
  88 #include <rte_memory.h>
  89 #include <rte_memzone.h>
  90 #include <rte_launch.h>
  91 #include <rte_eal.h>
  92 #include <rte_eal_memconfig.h>
  93 #include <rte_per_lcore.h>
  94 #include <rte_lcore.h>
  95 #include <rte_common.h>
  96 #include <rte_string_fns.h>
  97
  98 #include "eal_private.h"
  99 #include "eal_internal_cfg.h"
 100 #include "eal_filesystem.h"
 101 #include "eal_hugepages.h"
 102
 103 #define PFN_MASK_SIZE   8
 104
 105 #ifdef RTE_LIBRTE_XEN_DOM0
 106 int rte_xen_dom0_supported(void)
 107 {
 108         return internal_config.xen_dom0_support;
 109 }
 110 #endif
 111
 112 /**
 113  * @file
 114  * Huge page mapping under linux
 115  *
 116  * To reserve a big contiguous amount of memory, we use the hugepage
 117  * feature of linux. For that, we need to have hugetlbfs mounted. This
 118  * code will create many files in this directory (one per page) and
 119  * map them in virtual memory. For each page, we will retrieve its
 120  * physical address and remap it in order to have a virtual contiguous
 121  * zone as well as a physical contiguous zone.
 122  */
 123
 124 static uint64_t baseaddr_offset;
 125
 126 static bool phys_addrs_available = true;
 127
 128 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
 129
 130 static void
 131 test_phys_addrs_available(void)
 132 {
 133         uint64_t tmp;
 134         phys_addr_t physaddr;
 135
 136         physaddr = rte_mem_virt2phy(&tmp);
 137         if (physaddr == RTE_BAD_PHYS_ADDR) {
 138                 RTE_LOG(ERR, EAL,
 139                         "Cannot obtain physical addresses: %s. "
 140                         "Only vfio will function.\n",
 141                         strerror(errno));
 142                 phys_addrs_available = false;
 143         }
 144 }
 145
 146 /* Lock page in physical memory and prevent from swapping. */
 147 int
 148 rte_mem_lock_page(const void *virt)
 149 {
 150         unsigned long virtual = (unsigned long)virt;
 151         int page_size = getpagesize();
 152         unsigned long aligned = (virtual & ~ (page_size - 1));
 153         return mlock((void*)aligned, page_size);
 154 }
 155
 156 /*
 157  * Get physical address of any mapped virtual address in the current process.
 158  */
 159 phys_addr_t
 160 rte_mem_virt2phy(const void *virtaddr)
 161 {
 162         int fd, retval;
 163         uint64_t page, physaddr;
 164         unsigned long virt_pfn;
 165         int page_size;
 166         off_t offset;
 167
 168         /* when using dom0, /proc/self/pagemap always returns 0, check in
 169          * dpdk memory by browsing the memsegs */
 170         if (rte_xen_dom0_supported()) {
 171                 struct rte_mem_config *mcfg;
 172                 struct rte_memseg *memseg;
 173                 unsigned i;
 174
 175                 mcfg = rte_eal_get_configuration()->mem_config;
 176                 for (i = 0; i < RTE_MAX_MEMSEG; i++) {
 177                         memseg = &mcfg->memseg[i];
 178                         if (memseg->addr == NULL)
 179                                 break;
 180                         if (virtaddr > memseg->addr &&
 181                                         virtaddr < RTE_PTR_ADD(memseg->addr,
 182                                                 memseg->len)) {
 183                                 return memseg->phys_addr +
 184                                         RTE_PTR_DIFF(virtaddr, memseg->addr);
 185                         }
 186                 }
 187
 188                 return RTE_BAD_PHYS_ADDR;
 189         }
 190
 191         /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
 192         if (!phys_addrs_available)
 193                 return RTE_BAD_PHYS_ADDR;
 194
 195         /* standard page size */
 196         page_size = getpagesize();
 197
 198         fd = open("/proc/self/pagemap", O_RDONLY);
 199         if (fd < 0) {
 200                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 201                         __func__, strerror(errno));
 202                 return RTE_BAD_PHYS_ADDR;
 203         }
 204
 205         virt_pfn = (unsigned long)virtaddr / page_size;
 206         offset = sizeof(uint64_t) * virt_pfn;
 207         if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 208                 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 209                                 __func__, strerror(errno));
 210                 close(fd);
 211                 return RTE_BAD_PHYS_ADDR;
 212         }
 213
 214         retval = read(fd, &page, PFN_MASK_SIZE);
 215         close(fd);
 216         if (retval < 0) {
 217                 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 218                                 __func__, strerror(errno));
 219                 return RTE_BAD_PHYS_ADDR;
 220         } else if (retval != PFN_MASK_SIZE) {
 221                 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
 222                                 "but expected %d:\n",
 223                                 __func__, retval, PFN_MASK_SIZE);
 224                 return RTE_BAD_PHYS_ADDR;
 225         }
 226
 227         /*
 228          * the pfn (page frame number) are bits 0-54 (see
 229          * pagemap.txt in linux Documentation)
 230          */
 231         if ((page & 0x7fffffffffffffULL) == 0)
 232                 return RTE_BAD_PHYS_ADDR;
 233
 234         physaddr = ((page & 0x7fffffffffffffULL) * page_size)
 235                 + ((unsigned long)virtaddr % page_size);
 236
 237         return physaddr;
 238 }
 239
 240 /*
 241  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 242  * it by browsing the /proc/self/pagemap special file.
 243  */
 244 static int
 245 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 246 {
 247         unsigned int i;
 248         phys_addr_t addr;
 249
 250         for (i = 0; i < hpi->num_pages[0]; i++) {
 251                 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
 252                 if (addr == RTE_BAD_PHYS_ADDR)
 253                         return -1;
 254                 hugepg_tbl[i].physaddr = addr;
 255         }
 256         return 0;
 257 }
 258
 259 /*
 260  * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
 261  */
 262 static int
 263 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 264 {
 265         unsigned int i;
 266         static phys_addr_t addr;
 267
 268         for (i = 0; i < hpi->num_pages[0]; i++) {
 269                 hugepg_tbl[i].physaddr = addr;
 270                 addr += hugepg_tbl[i].size;
 271         }
 272         return 0;
 273 }
 274
 275 /*
 276  * Check whether address-space layout randomization is enabled in
 277  * the kernel. This is important for multi-process as it can prevent
 278  * two processes mapping data to the same virtual address
 279  * Returns:
 280  *    0 - address space randomization disabled
 281  *    1/2 - address space randomization enabled
 282  *    negative error code on error
 283  */
 284 static int
 285 aslr_enabled(void)
 286 {
 287         char c;
 288         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 289         if (fd < 0)
 290                 return -errno;
 291         retval = read(fd, &c, 1);
 292         close(fd);
 293         if (retval < 0)
 294                 return -errno;
 295         if (retval == 0)
 296                 return -EIO;
 297         switch (c) {
 298                 case '0' : return 0;
 299                 case '1' : return 1;
 300                 case '2' : return 2;
 301                 default: return -EINVAL;
 302         }
 303 }
 304
 305 /*
 306  * Try to mmap *size bytes in /dev/zero. If it is successful, return the
 307  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 308  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 309  * 0. In this case, return NULL. Note: this function returns an address
 310  * which is a multiple of hugepage size.
 311  */
 312 static void *
 313 get_virtual_area(size_t *size, size_t hugepage_sz)
 314 {
 315         void *addr;
 316         int fd;
 317         long aligned_addr;
 318
 319         if (internal_config.base_virtaddr != 0) {
 320                 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
 321                                 baseaddr_offset);
 322         }
 323         else addr = NULL;
 324
 325         RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
 326
 327         fd = open("/dev/zero", O_RDONLY);
 328         if (fd < 0){
 329                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 330                 return NULL;
 331         }
 332         do {
 333                 addr = mmap(addr,
 334                                 (*size) + hugepage_sz, PROT_READ,
 335 #ifdef RTE_ARCH_PPC_64
 336                                 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
 337 #else
 338                                 MAP_PRIVATE,
 339 #endif
 340                                 fd, 0);
 341                 if (addr == MAP_FAILED)
 342                         *size -= hugepage_sz;
 343         } while (addr == MAP_FAILED && *size > 0);
 344
 345         if (addr == MAP_FAILED) {
 346                 close(fd);
 347                 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
 348                         strerror(errno));
 349                 return NULL;
 350         }
 351
 352         munmap(addr, (*size) + hugepage_sz);
 353         close(fd);
 354
 355         /* align addr to a huge page size boundary */
 356         aligned_addr = (long)addr;
 357         aligned_addr += (hugepage_sz - 1);
 358         aligned_addr &= (~(hugepage_sz - 1));
 359         addr = (void *)(aligned_addr);
 360
 361         RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
 362                 addr, *size);
 363
 364         /* increment offset */
 365         baseaddr_offset += *size;
 366
 367         return addr;
 368 }
 369
 370 static sigjmp_buf huge_jmpenv;
 371
 372 static void huge_sigbus_handler(int signo __rte_unused)
 373 {
 374         siglongjmp(huge_jmpenv, 1);
 375 }
 376
 377 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
 378  * non-static local variable in the stack frame calling sigsetjmp might be
 379  * clobbered by a call to longjmp.
 380  */
 381 static int huge_wrap_sigsetjmp(void)
 382 {
 383         return sigsetjmp(huge_jmpenv, 1);
 384 }
 385
 386 /*
 387  * Mmap all hugepages of hugepage table: it first open a file in
 388  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 389  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 390  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 391  * map continguous physical blocks in contiguous virtual blocks.
 392  */
 393 static unsigned
 394 map_all_hugepages(struct hugepage_file *hugepg_tbl,
 395                 struct hugepage_info *hpi, int orig)
 396 {
 397         int fd;
 398         unsigned i;
 399         void *virtaddr;
 400         void *vma_addr = NULL;
 401         size_t vma_len = 0;
 402
 403         for (i = 0; i < hpi->num_pages[0]; i++) {
 404                 uint64_t hugepage_sz = hpi->hugepage_sz;
 405
 406                 if (orig) {
 407                         hugepg_tbl[i].file_id = i;
 408                         hugepg_tbl[i].size = hugepage_sz;
 409                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 410                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 411                                         hugepg_tbl[i].file_id);
 412                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 413                 }
 414 #ifndef RTE_ARCH_64
 415                 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
 416                  * original map address as final map address.
 417                  */
 418                 else if ((hugepage_sz == RTE_PGSIZE_1G)
 419                         || (hugepage_sz == RTE_PGSIZE_16G)) {
 420                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 421                         hugepg_tbl[i].orig_va = NULL;
 422                         continue;
 423                 }
 424 #endif
 425                 else if (vma_len == 0) {
 426                         unsigned j, num_pages;
 427
 428                         /* reserve a virtual area for next contiguous
 429                          * physical block: count the number of
 430                          * contiguous physical pages. */
 431                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 432 #ifdef RTE_ARCH_PPC_64
 433                                 /* The physical addresses are sorted in
 434                                  * descending order on PPC64 */
 435                                 if (hugepg_tbl[j].physaddr !=
 436                                     hugepg_tbl[j-1].physaddr - hugepage_sz)
 437                                         break;
 438 #else
 439                                 if (hugepg_tbl[j].physaddr !=
 440                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 441                                         break;
 442 #endif
 443                         }
 444                         num_pages = j - i;
 445                         vma_len = num_pages * hugepage_sz;
 446
 447                         /* get the biggest virtual memory area up to
 448                          * vma_len. If it fails, vma_addr is NULL, so
 449                          * let the kernel provide the address. */
 450                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 451                         if (vma_addr == NULL)
 452                                 vma_len = hugepage_sz;
 453                 }
 454
 455                 /* try to create hugepage file */
 456                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
 457                 if (fd < 0) {
 458                         RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
 459                                         strerror(errno));
 460                         return i;
 461                 }
 462
 463                 /* map the segment, and populate page tables,
 464                  * the kernel fills this segment with zeros */
 465                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 466                                 MAP_SHARED | MAP_POPULATE, fd, 0);
 467                 if (virtaddr == MAP_FAILED) {
 468                         RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
 469                                         strerror(errno));
 470                         close(fd);
 471                         return i;
 472                 }
 473
 474                 if (orig) {
 475                         hugepg_tbl[i].orig_va = virtaddr;
 476                 }
 477                 else {
 478                         hugepg_tbl[i].final_va = virtaddr;
 479                 }
 480
 481                 if (orig) {
 482                         /* In linux, hugetlb limitations, like cgroup, are
 483                          * enforced at fault time instead of mmap(), even
 484                          * with the option of MAP_POPULATE. Kernel will send
 485                          * a SIGBUS signal. To avoid to be killed, save stack
 486                          * environment here, if SIGBUS happens, we can jump
 487                          * back here.
 488                          */
 489                         if (huge_wrap_sigsetjmp()) {
 490                                 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
 491                                         "hugepages of size %u MB\n",
 492                                         (unsigned)(hugepage_sz / 0x100000));
 493                                 munmap(virtaddr, hugepage_sz);
 494                                 close(fd);
 495                                 unlink(hugepg_tbl[i].filepath);
 496                                 return i;
 497                         }
 498                         *(int *)virtaddr = 0;
 499                 }
 500
 501
 502                 /* set shared flock on the file. */
 503                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 504                         RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
 505                                 __func__, strerror(errno));
 506                         close(fd);
 507                         return i;
 508                 }
 509
 510                 close(fd);
 511
 512                 vma_addr = (char *)vma_addr + hugepage_sz;
 513                 vma_len -= hugepage_sz;
 514         }
 515
 516         return i;
 517 }
 518
 519 /* Unmap all hugepages from original mapping */
 520 static int
 521 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 522 {
 523         unsigned i;
 524         for (i = 0; i < hpi->num_pages[0]; i++) {
 525                 if (hugepg_tbl[i].orig_va) {
 526                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 527                         hugepg_tbl[i].orig_va = NULL;
 528                 }
 529         }
 530         return 0;
 531 }
 532
 533 /*
 534  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 535  * page.
 536  */
 537 static int
 538 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 539 {
 540         int socket_id;
 541         char *end, *nodestr;
 542         unsigned i, hp_count = 0;
 543         uint64_t virt_addr;
 544         char buf[BUFSIZ];
 545         char hugedir_str[PATH_MAX];
 546         FILE *f;
 547
 548         f = fopen("/proc/self/numa_maps", "r");
 549         if (f == NULL) {
 550                 RTE_LOG(NOTICE, EAL, "cannot open /proc/self/numa_maps,"
 551                                 " consider that all memory is in socket_id 0\n");
 552                 return 0;
 553         }
 554
 555         snprintf(hugedir_str, sizeof(hugedir_str),
 556                         "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
 557
 558         /* parse numa map */
 559         while (fgets(buf, sizeof(buf), f) != NULL) {
 560
 561                 /* ignore non huge page */
 562                 if (strstr(buf, " huge ") == NULL &&
 563                                 strstr(buf, hugedir_str) == NULL)
 564                         continue;
 565
 566                 /* get zone addr */
 567                 virt_addr = strtoull(buf, &end, 16);
 568                 if (virt_addr == 0 || end == buf) {
 569                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 570                         goto error;
 571                 }
 572
 573                 /* get node id (socket id) */
 574                 nodestr = strstr(buf, " N");
 575                 if (nodestr == NULL) {
 576                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 577                         goto error;
 578                 }
 579                 nodestr += 2;
 580                 end = strstr(nodestr, "=");
 581                 if (end == NULL) {
 582                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 583                         goto error;
 584                 }
 585                 end[0] = '\0';
 586                 end = NULL;
 587
 588                 socket_id = strtoul(nodestr, &end, 0);
 589                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 590                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 591                         goto error;
 592                 }
 593
 594                 /* if we find this page in our mappings, set socket_id */
 595                 for (i = 0; i < hpi->num_pages[0]; i++) {
 596                         void *va = (void *)(unsigned long)virt_addr;
 597                         if (hugepg_tbl[i].orig_va == va) {
 598                                 hugepg_tbl[i].socket_id = socket_id;
 599                                 hp_count++;
 600                         }
 601                 }
 602         }
 603
 604         if (hp_count < hpi->num_pages[0])
 605                 goto error;
 606
 607         fclose(f);
 608         return 0;
 609
 610 error:
 611         fclose(f);
 612         return -1;
 613 }
 614
 615 static int
 616 cmp_physaddr(const void *a, const void *b)
 617 {
 618 #ifndef RTE_ARCH_PPC_64
 619         const struct hugepage_file *p1 = a;
 620         const struct hugepage_file *p2 = b;
 621 #else
 622         /* PowerPC needs memory sorted in reverse order from x86 */
 623         const struct hugepage_file *p1 = b;
 624         const struct hugepage_file *p2 = a;
 625 #endif
 626         if (p1->physaddr < p2->physaddr)
 627                 return -1;
 628         else if (p1->physaddr > p2->physaddr)
 629                 return 1;
 630         else
 631                 return 0;
 632 }
 633
 634 /*
 635  * Uses mmap to create a shared memory area for storage of data
 636  * Used in this file to store the hugepage file map on disk
 637  */
 638 static void *
 639 create_shared_memory(const char *filename, const size_t mem_size)
 640 {
 641         void *retval;
 642         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 643         if (fd < 0)
 644                 return NULL;
 645         if (ftruncate(fd, mem_size) < 0) {
 646                 close(fd);
 647                 return NULL;
 648         }
 649         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 650         close(fd);
 651         return retval;
 652 }
 653
 654 /*
 655  * this copies *active* hugepages from one hugepage table to another.
 656  * destination is typically the shared memory.
 657  */
 658 static int
 659 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
 660                 const struct hugepage_file * src, int src_size)
 661 {
 662         int src_pos, dst_pos = 0;
 663
 664         for (src_pos = 0; src_pos < src_size; src_pos++) {
 665                 if (src[src_pos].final_va != NULL) {
 666                         /* error on overflow attempt */
 667                         if (dst_pos == dest_size)
 668                                 return -1;
 669                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
 670                         dst_pos++;
 671                 }
 672         }
 673         return 0;
 674 }
 675
 676 static int
 677 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
 678                 unsigned num_hp_info)
 679 {
 680         unsigned socket, size;
 681         int page, nrpages = 0;
 682
 683         /* get total number of hugepages */
 684         for (size = 0; size < num_hp_info; size++)
 685                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 686                         nrpages +=
 687                         internal_config.hugepage_info[size].num_pages[socket];
 688
 689         for (page = 0; page < nrpages; page++) {
 690                 struct hugepage_file *hp = &hugepg_tbl[page];
 691
 692                 if (hp->final_va != NULL && unlink(hp->filepath)) {
 693                         RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
 694                                 __func__, hp->filepath, strerror(errno));
 695                 }
 696         }
 697         return 0;
 698 }
 699
 700 /*
 701  * unmaps hugepages that are not going to be used. since we originally allocate
 702  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 703  */
 704 static int
 705 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
 706                 struct hugepage_info *hpi,
 707                 unsigned num_hp_info)
 708 {
 709         unsigned socket, size;
 710         int page, nrpages = 0;
 711
 712         /* get total number of hugepages */
 713         for (size = 0; size < num_hp_info; size++)
 714                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 715                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 716
 717         for (size = 0; size < num_hp_info; size++) {
 718                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 719                         unsigned pages_found = 0;
 720
 721                         /* traverse until we have unmapped all the unused pages */
 722                         for (page = 0; page < nrpages; page++) {
 723                                 struct hugepage_file *hp = &hugepg_tbl[page];
 724
 725                                 /* find a page that matches the criteria */
 726                                 if ((hp->size == hpi[size].hugepage_sz) &&
 727                                                 (hp->socket_id == (int) socket)) {
 728
 729                                         /* if we skipped enough pages, unmap the rest */
 730                                         if (pages_found == hpi[size].num_pages[socket]) {
 731                                                 uint64_t unmap_len;
 732
 733                                                 unmap_len = hp->size;
 734
 735                                                 /* get start addr and len of the remaining segment */
 736                                                 munmap(hp->final_va, (size_t) unmap_len);
 737
 738                                                 hp->final_va = NULL;
 739                                                 if (unlink(hp->filepath) == -1) {
 740                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 741                                                                         __func__, hp->filepath, strerror(errno));
 742                                                         return -1;
 743                                                 }
 744                                         } else {
 745                                                 /* lock the page and skip */
 746                                                 pages_found++;
 747                                         }
 748
 749                                 } /* match page */
 750                         } /* foreach page */
 751                 } /* foreach socket */
 752         } /* foreach pagesize */
 753
 754         return 0;
 755 }
 756
 757 static inline uint64_t
 758 get_socket_mem_size(int socket)
 759 {
 760         uint64_t size = 0;
 761         unsigned i;
 762
 763         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 764                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 765                 if (hpi->hugedir != NULL)
 766                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 767         }
 768
 769         return size;
 770 }
 771
 772 /*
 773  * This function is a NUMA-aware equivalent of calc_num_pages.
 774  * It takes in the list of hugepage sizes and the
 775  * number of pages thereof, and calculates the best number of
 776  * pages of each size to fulfill the request for <memory> ram
 777  */
 778 static int
 779 calc_num_pages_per_socket(uint64_t * memory,
 780                 struct hugepage_info *hp_info,
 781                 struct hugepage_info *hp_used,
 782                 unsigned num_hp_info)
 783 {
 784         unsigned socket, j, i = 0;
 785         unsigned requested, available;
 786         int total_num_pages = 0;
 787         uint64_t remaining_mem, cur_mem;
 788         uint64_t total_mem = internal_config.memory;
 789
 790         if (num_hp_info == 0)
 791                 return -1;
 792
 793         /* if specific memory amounts per socket weren't requested */
 794         if (internal_config.force_sockets == 0) {
 795                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
 796                 size_t default_size, total_size;
 797                 unsigned lcore_id;
 798
 799                 /* Compute number of cores per socket */
 800                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
 801                 RTE_LCORE_FOREACH(lcore_id) {
 802                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
 803                 }
 804
 805                 /*
 806                  * Automatically spread requested memory amongst detected sockets according
 807                  * to number of cores from cpu mask present on each socket
 808                  */
 809                 total_size = internal_config.memory;
 810                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 811
 812                         /* Set memory amount per socket */
 813                         default_size = (internal_config.memory * cpu_per_socket[socket])
 814                                         / rte_lcore_count();
 815
 816                         /* Limit to maximum available memory on socket */
 817                         default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
 818
 819                         /* Update sizes */
 820                         memory[socket] = default_size;
 821                         total_size -= default_size;
 822                 }
 823
 824                 /*
 825                  * If some memory is remaining, try to allocate it by getting all
 826                  * available memory from sockets, one after the other
 827                  */
 828                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 829                         /* take whatever is available */
 830                         default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
 831                                                total_size);
 832
 833                         /* Update sizes */
 834                         memory[socket] += default_size;
 835                         total_size -= default_size;
 836                 }
 837         }
 838
 839         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 840                 /* skips if the memory on specific socket wasn't requested */
 841                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 842                         hp_used[i].hugedir = hp_info[i].hugedir;
 843                         hp_used[i].num_pages[socket] = RTE_MIN(
 844                                         memory[socket] / hp_info[i].hugepage_sz,
 845                                         hp_info[i].num_pages[socket]);
 846
 847                         cur_mem = hp_used[i].num_pages[socket] *
 848                                         hp_used[i].hugepage_sz;
 849
 850                         memory[socket] -= cur_mem;
 851                         total_mem -= cur_mem;
 852
 853                         total_num_pages += hp_used[i].num_pages[socket];
 854
 855                         /* check if we have met all memory requests */
 856                         if (memory[socket] == 0)
 857                                 break;
 858
 859                         /* check if we have any more pages left at this size, if so
 860                          * move on to next size */
 861                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
 862                                 continue;
 863                         /* At this point we know that there are more pages available that are
 864                          * bigger than the memory we want, so lets see if we can get enough
 865                          * from other page sizes.
 866                          */
 867                         remaining_mem = 0;
 868                         for (j = i+1; j < num_hp_info; j++)
 869                                 remaining_mem += hp_info[j].hugepage_sz *
 870                                 hp_info[j].num_pages[socket];
 871
 872                         /* is there enough other memory, if not allocate another page and quit */
 873                         if (remaining_mem < memory[socket]){
 874                                 cur_mem = RTE_MIN(memory[socket],
 875                                                 hp_info[i].hugepage_sz);
 876                                 memory[socket] -= cur_mem;
 877                                 total_mem -= cur_mem;
 878                                 hp_used[i].num_pages[socket]++;
 879                                 total_num_pages++;
 880                                 break; /* we are done with this socket*/
 881                         }
 882                 }
 883                 /* if we didn't satisfy all memory requirements per socket */
 884                 if (memory[socket] > 0) {
 885                         /* to prevent icc errors */
 886                         requested = (unsigned) (internal_config.socket_mem[socket] /
 887                                         0x100000);
 888                         available = requested -
 889                                         ((unsigned) (memory[socket] / 0x100000));
 890                         RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
 891                                         "Requested: %uMB, available: %uMB\n", socket,
 892                                         requested, available);
 893                         return -1;
 894                 }
 895         }
 896
 897         /* if we didn't satisfy total memory requirements */
 898         if (total_mem > 0) {
 899                 requested = (unsigned) (internal_config.memory / 0x100000);
 900                 available = requested - (unsigned) (total_mem / 0x100000);
 901                 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
 902                                 " available: %uMB\n", requested, available);
 903                 return -1;
 904         }
 905         return total_num_pages;
 906 }
 907
 908 static inline size_t
 909 eal_get_hugepage_mem_size(void)
 910 {
 911         uint64_t size = 0;
 912         unsigned i, j;
 913
 914         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 915                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 916                 if (hpi->hugedir != NULL) {
 917                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 918                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 919                         }
 920                 }
 921         }
 922
 923         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 924 }
 925
 926 static struct sigaction huge_action_old;
 927 static int huge_need_recover;
 928
 929 static void
 930 huge_register_sigbus(void)
 931 {
 932         sigset_t mask;
 933         struct sigaction action;
 934
 935         sigemptyset(&mask);
 936         sigaddset(&mask, SIGBUS);
 937         action.sa_flags = 0;
 938         action.sa_mask = mask;
 939         action.sa_handler = huge_sigbus_handler;
 940
 941         huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
 942 }
 943
 944 static void
 945 huge_recover_sigbus(void)
 946 {
 947         if (huge_need_recover) {
 948                 sigaction(SIGBUS, &huge_action_old, NULL);
 949                 huge_need_recover = 0;
 950         }
 951 }
 952
 953 /*
 954  * Prepare physical memory mapping: fill configuration structure with
 955  * these infos, return 0 on success.
 956  *  1. map N huge pages in separate files in hugetlbfs
 957  *  2. find associated physical addr
 958  *  3. find associated NUMA socket ID
 959  *  4. sort all huge pages by physical address
 960  *  5. remap these N huge pages in the correct order
 961  *  6. unmap the first mapping
 962  *  7. fill memsegs in configuration with contiguous zones
 963  */
 964 int
 965 rte_eal_hugepage_init(void)
 966 {
 967         struct rte_mem_config *mcfg;
 968         struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
 969         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
 970
 971         uint64_t memory[RTE_MAX_NUMA_NODES];
 972
 973         unsigned hp_offset;
 974         int i, j, new_memseg;
 975         int nr_hugefiles, nr_hugepages = 0;
 976         void *addr;
 977
 978         test_phys_addrs_available();
 979
 980         memset(used_hp, 0, sizeof(used_hp));
 981
 982         /* get pointer to global configuration */
 983         mcfg = rte_eal_get_configuration()->mem_config;
 984
 985         /* hugetlbfs can be disabled */
 986         if (internal_config.no_hugetlbfs) {
 987                 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
 988                                 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
 989                 if (addr == MAP_FAILED) {
 990                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
 991                                         strerror(errno));
 992                         return -1;
 993                 }
 994                 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
 995                 mcfg->memseg[0].addr = addr;
 996                 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
 997                 mcfg->memseg[0].len = internal_config.memory;
 998                 mcfg->memseg[0].socket_id = 0;
 999                 return 0;
1000         }
1001
1002 /* check if app runs on Xen Dom0 */
1003         if (internal_config.xen_dom0_support) {
1004 #ifdef RTE_LIBRTE_XEN_DOM0
1005                 /* use dom0_mm kernel driver to init memory */
1006                 if (rte_xen_dom0_memory_init() < 0)
1007                         return -1;
1008                 else
1009                         return 0;
1010 #endif
1011         }
1012
1013         /* calculate total number of hugepages available. at this point we haven't
1014          * yet started sorting them so they all are on socket 0 */
1015         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1016                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1017                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1018
1019                 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1020         }
1021
1022         /*
1023          * allocate a memory area for hugepage table.
1024          * this isn't shared memory yet. due to the fact that we need some
1025          * processing done on these pages, shared memory will be created
1026          * at a later stage.
1027          */
1028         tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1029         if (tmp_hp == NULL)
1030                 goto fail;
1031
1032         memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1033
1034         hp_offset = 0; /* where we start the current page size entries */
1035
1036         huge_register_sigbus();
1037
1038         /* map all hugepages and sort them */
1039         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1040                 unsigned pages_old, pages_new;
1041                 struct hugepage_info *hpi;
1042
1043                 /*
1044                  * we don't yet mark hugepages as used at this stage, so
1045                  * we just map all hugepages available to the system
1046                  * all hugepages are still located on socket 0
1047                  */
1048                 hpi = &internal_config.hugepage_info[i];
1049
1050                 if (hpi->num_pages[0] == 0)
1051                         continue;
1052
1053                 /* map all hugepages available */
1054                 pages_old = hpi->num_pages[0];
1055                 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, 1);
1056                 if (pages_new < pages_old) {
1057                         RTE_LOG(DEBUG, EAL,
1058                                 "%d not %d hugepages of size %u MB allocated\n",
1059                                 pages_new, pages_old,
1060                                 (unsigned)(hpi->hugepage_sz / 0x100000));
1061
1062                         int pages = pages_old - pages_new;
1063
1064                         nr_hugepages -= pages;
1065                         hpi->num_pages[0] = pages_new;
1066                         if (pages_new == 0)
1067                                 continue;
1068                 }
1069
1070                 if (phys_addrs_available) {
1071                         /* find physical addresses for each hugepage */
1072                         if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1073                                 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1074                                         "for %u MB pages\n",
1075                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1076                                 goto fail;
1077                         }
1078                 } else {
1079                         /* set physical addresses for each hugepage */
1080                         if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1081                                 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1082                                         "for %u MB pages\n",
1083                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1084                                 goto fail;
1085                         }
1086                 }
1087
1088                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1089                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1090                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1091                         goto fail;
1092                 }
1093
1094                 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1095                       sizeof(struct hugepage_file), cmp_physaddr);
1096
1097                 /* remap all hugepages */
1098                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) !=
1099                     hpi->num_pages[0]) {
1100                         RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1101                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1102                         goto fail;
1103                 }
1104
1105                 /* unmap original mappings */
1106                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1107                         goto fail;
1108
1109                 /* we have processed a num of hugepages of this size, so inc offset */
1110                 hp_offset += hpi->num_pages[0];
1111         }
1112
1113         huge_recover_sigbus();
1114
1115         if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1116                 internal_config.memory = eal_get_hugepage_mem_size();
1117
1118         nr_hugefiles = nr_hugepages;
1119
1120
1121         /* clean out the numbers of pages */
1122         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1123                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1124                         internal_config.hugepage_info[i].num_pages[j] = 0;
1125
1126         /* get hugepages for each socket */
1127         for (i = 0; i < nr_hugefiles; i++) {
1128                 int socket = tmp_hp[i].socket_id;
1129
1130                 /* find a hugepage info with right size and increment num_pages */
1131                 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1132                                 (int)internal_config.num_hugepage_sizes);
1133                 for (j = 0; j < nb_hpsizes; j++) {
1134                         if (tmp_hp[i].size ==
1135                                         internal_config.hugepage_info[j].hugepage_sz) {
1136                                 internal_config.hugepage_info[j].num_pages[socket]++;
1137                         }
1138                 }
1139         }
1140
1141         /* make a copy of socket_mem, needed for number of pages calculation */
1142         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1143                 memory[i] = internal_config.socket_mem[i];
1144
1145         /* calculate final number of pages */
1146         nr_hugepages = calc_num_pages_per_socket(memory,
1147                         internal_config.hugepage_info, used_hp,
1148                         internal_config.num_hugepage_sizes);
1149
1150         /* error if not enough memory available */
1151         if (nr_hugepages < 0)
1152                 goto fail;
1153
1154         /* reporting in! */
1155         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1156                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1157                         if (used_hp[i].num_pages[j] > 0) {
1158                                 RTE_LOG(DEBUG, EAL,
1159                                         "Requesting %u pages of size %uMB"
1160                                         " from socket %i\n",
1161                                         used_hp[i].num_pages[j],
1162                                         (unsigned)
1163                                         (used_hp[i].hugepage_sz / 0x100000),
1164                                         j);
1165                         }
1166                 }
1167         }
1168
1169         /* create shared memory */
1170         hugepage = create_shared_memory(eal_hugepage_info_path(),
1171                         nr_hugefiles * sizeof(struct hugepage_file));
1172
1173         if (hugepage == NULL) {
1174                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1175                 goto fail;
1176         }
1177         memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1178
1179         /*
1180          * unmap pages that we won't need (looks at used_hp).
1181          * also, sets final_va to NULL on pages that were unmapped.
1182          */
1183         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1184                         internal_config.num_hugepage_sizes) < 0) {
1185                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1186                 goto fail;
1187         }
1188
1189         /*
1190          * copy stuff from malloc'd hugepage* to the actual shared memory.
1191          * this procedure only copies those hugepages that have final_va
1192          * not NULL. has overflow protection.
1193          */
1194         if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1195                         tmp_hp, nr_hugefiles) < 0) {
1196                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1197                 goto fail;
1198         }
1199
1200         /* free the hugepage backing files */
1201         if (internal_config.hugepage_unlink &&
1202                 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1203                 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1204                 goto fail;
1205         }
1206
1207         /* free the temporary hugepage table */
1208         free(tmp_hp);
1209         tmp_hp = NULL;
1210
1211         /* first memseg index shall be 0 after incrementing it below */
1212         j = -1;
1213         for (i = 0; i < nr_hugefiles; i++) {
1214                 new_memseg = 0;
1215
1216                 /* if this is a new section, create a new memseg */
1217                 if (i == 0)
1218                         new_memseg = 1;
1219                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1220                         new_memseg = 1;
1221                 else if (hugepage[i].size != hugepage[i-1].size)
1222                         new_memseg = 1;
1223
1224 #ifdef RTE_ARCH_PPC_64
1225                 /* On PPC64 architecture, the mmap always start from higher
1226                  * virtual address to lower address. Here, both the physical
1227                  * address and virtual address are in descending order */
1228                 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1229                     hugepage[i].size)
1230                         new_memseg = 1;
1231                 else if (((unsigned long)hugepage[i-1].final_va -
1232                     (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1233                         new_memseg = 1;
1234 #else
1235                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1236                     hugepage[i].size)
1237                         new_memseg = 1;
1238                 else if (((unsigned long)hugepage[i].final_va -
1239                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1240                         new_memseg = 1;
1241 #endif
1242
1243                 if (new_memseg) {
1244                         j += 1;
1245                         if (j == RTE_MAX_MEMSEG)
1246                                 break;
1247
1248                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1249                         mcfg->memseg[j].addr = hugepage[i].final_va;
1250                         mcfg->memseg[j].len = hugepage[i].size;
1251                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1252                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1253                 }
1254                 /* continuation of previous memseg */
1255                 else {
1256 #ifdef RTE_ARCH_PPC_64
1257                 /* Use the phy and virt address of the last page as segment
1258                  * address for IBM Power architecture */
1259                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1260                         mcfg->memseg[j].addr = hugepage[i].final_va;
1261 #endif
1262                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1263                 }
1264                 hugepage[i].memseg_id = j;
1265         }
1266
1267         if (i < nr_hugefiles) {
1268                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1269                         "from %d requested\n"
1270                         "Current %s=%d is not enough\n"
1271                         "Please either increase it or request less amount "
1272                         "of memory.\n",
1273                         i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1274                         RTE_MAX_MEMSEG);
1275                 goto fail;
1276         }
1277
1278         munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1279
1280         return 0;
1281
1282 fail:
1283         huge_recover_sigbus();
1284         free(tmp_hp);
1285         if (hugepage != NULL)
1286                 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1287
1288         return -1;
1289 }
1290
1291 /*
1292  * uses fstat to report the size of a file on disk
1293  */
1294 static off_t
1295 getFileSize(int fd)
1296 {
1297         struct stat st;
1298         if (fstat(fd, &st) < 0)
1299                 return 0;
1300         return st.st_size;
1301 }
1302
1303 /*
1304  * This creates the memory mappings in the secondary process to match that of
1305  * the server process. It goes through each memory segment in the DPDK runtime
1306  * configuration and finds the hugepages which form that segment, mapping them
1307  * in order to form a contiguous block in the virtual memory space
1308  */
1309 int
1310 rte_eal_hugepage_attach(void)
1311 {
1312         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1313         struct hugepage_file *hp = NULL;
1314         unsigned num_hp = 0;
1315         unsigned i, s = 0; /* s used to track the segment number */
1316         unsigned max_seg = RTE_MAX_MEMSEG;
1317         off_t size = 0;
1318         int fd, fd_zero = -1, fd_hugepage = -1;
1319
1320         if (aslr_enabled() > 0) {
1321                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1322                                 "(ASLR) is enabled in the kernel.\n");
1323                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1324                                 "into secondary processes\n");
1325         }
1326
1327         test_phys_addrs_available();
1328
1329         if (internal_config.xen_dom0_support) {
1330 #ifdef RTE_LIBRTE_XEN_DOM0
1331                 if (rte_xen_dom0_memory_attach() < 0) {
1332                         RTE_LOG(ERR, EAL, "Failed to attach memory segments of primary "
1333                                         "process\n");
1334                         return -1;
1335                 }
1336                 return 0;
1337 #endif
1338         }
1339
1340         fd_zero = open("/dev/zero", O_RDONLY);
1341         if (fd_zero < 0) {
1342                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1343                 goto error;
1344         }
1345         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1346         if (fd_hugepage < 0) {
1347                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1348                 goto error;
1349         }
1350
1351         /* map all segments into memory to make sure we get the addrs */
1352         for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1353                 void *base_addr;
1354
1355                 /*
1356                  * the first memory segment with len==0 is the one that
1357                  * follows the last valid segment.
1358                  */
1359                 if (mcfg->memseg[s].len == 0)
1360                         break;
1361
1362                 /*
1363                  * fdzero is mmapped to get a contiguous block of virtual
1364                  * addresses of the appropriate memseg size.
1365                  * use mmap to get identical addresses as the primary process.
1366                  */
1367                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1368                                  PROT_READ,
1369 #ifdef RTE_ARCH_PPC_64
1370                                  MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
1371 #else
1372                                  MAP_PRIVATE,
1373 #endif
1374                                  fd_zero, 0);
1375                 if (base_addr == MAP_FAILED ||
1376                     base_addr != mcfg->memseg[s].addr) {
1377                         max_seg = s;
1378                         if (base_addr != MAP_FAILED) {
1379                                 /* errno is stale, don't use */
1380                                 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1381                                         "in /dev/zero at [%p], got [%p] - "
1382                                         "please use '--base-virtaddr' option\n",
1383                                         (unsigned long long)mcfg->memseg[s].len,
1384                                         mcfg->memseg[s].addr, base_addr);
1385                                 munmap(base_addr, mcfg->memseg[s].len);
1386                         } else {
1387                                 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1388                                         "in /dev/zero at [%p]: '%s'\n",
1389                                         (unsigned long long)mcfg->memseg[s].len,
1390                                         mcfg->memseg[s].addr, strerror(errno));
1391                         }
1392                         if (aslr_enabled() > 0) {
1393                                 RTE_LOG(ERR, EAL, "It is recommended to "
1394                                         "disable ASLR in the kernel "
1395                                         "and retry running both primary "
1396                                         "and secondary processes\n");
1397                         }
1398                         goto error;
1399                 }
1400         }
1401
1402         size = getFileSize(fd_hugepage);
1403         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1404         if (hp == MAP_FAILED) {
1405                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1406                 goto error;
1407         }
1408
1409         num_hp = size / sizeof(struct hugepage_file);
1410         RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1411
1412         s = 0;
1413         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1414                 void *addr, *base_addr;
1415                 uintptr_t offset = 0;
1416                 size_t mapping_size;
1417                 /*
1418                  * free previously mapped memory so we can map the
1419                  * hugepages into the space
1420                  */
1421                 base_addr = mcfg->memseg[s].addr;
1422                 munmap(base_addr, mcfg->memseg[s].len);
1423
1424                 /* find the hugepages for this segment and map them
1425                  * we don't need to worry about order, as the server sorted the
1426                  * entries before it did the second mmap of them */
1427                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1428                         if (hp[i].memseg_id == (int)s){
1429                                 fd = open(hp[i].filepath, O_RDWR);
1430                                 if (fd < 0) {
1431                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1432                                                 hp[i].filepath);
1433                                         goto error;
1434                                 }
1435                                 mapping_size = hp[i].size;
1436                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1437                                                 mapping_size, PROT_READ | PROT_WRITE,
1438                                                 MAP_SHARED, fd, 0);
1439                                 close(fd); /* close file both on success and on failure */
1440                                 if (addr == MAP_FAILED ||
1441                                                 addr != RTE_PTR_ADD(base_addr, offset)) {
1442                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1443                                                 hp[i].filepath);
1444                                         goto error;
1445                                 }
1446                                 offset+=mapping_size;
1447                         }
1448                 }
1449                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1450                                 (unsigned long long)mcfg->memseg[s].len);
1451                 s++;
1452         }
1453         /* unmap the hugepage config file, since we are done using it */
1454         munmap(hp, size);
1455         close(fd_zero);
1456         close(fd_hugepage);
1457         return 0;
1458
1459 error:
1460         for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1461                 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1462         if (hp != NULL && hp != MAP_FAILED)
1463                 munmap(hp, size);
1464         if (fd_zero >= 0)
1465                 close(fd_zero);
1466         if (fd_hugepage >= 0)
1467                 close(fd_hugepage);
1468         return -1;
1469 }
1470
1471 bool
1472 rte_eal_using_phys_addrs(void)
1473 {
1474         return phys_addrs_available;
1475 }