lib/librte_eal/linuxapp/eal/eal_memory.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   Copyright(c) 2013 6WIND.
   6  *   All rights reserved.
   7  *
   8  *   Redistribution and use in source and binary forms, with or without
   9  *   modification, are permitted provided that the following conditions
  10  *   are met:
  11  *
  12  *     * Redistributions of source code must retain the above copyright
  13  *       notice, this list of conditions and the following disclaimer.
  14  *     * Redistributions in binary form must reproduce the above copyright
  15  *       notice, this list of conditions and the following disclaimer in
  16  *       the documentation and/or other materials provided with the
  17  *       distribution.
  18  *     * Neither the name of Intel Corporation nor the names of its
  19  *       contributors may be used to endorse or promote products derived
  20  *       from this software without specific prior written permission.
  21  *
  22  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  25  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  26  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  27  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  28  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  29  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  30  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  31  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  32  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  33  */
  34
  35 #define _FILE_OFFSET_BITS 64
  36 #include <errno.h>
  37 #include <stdarg.h>
  38 #include <stdbool.h>
  39 #include <stdlib.h>
  40 #include <stdio.h>
  41 #include <stdint.h>
  42 #include <inttypes.h>
  43 #include <string.h>
  44 #include <stdarg.h>
  45 #include <sys/mman.h>
  46 #include <sys/types.h>
  47 #include <sys/stat.h>
  48 #include <sys/queue.h>
  49 #include <sys/file.h>
  50 #include <unistd.h>
  51 #include <limits.h>
  52 #include <errno.h>
  53 #include <sys/ioctl.h>
  54 #include <sys/time.h>
  55 #include <signal.h>
  56 #include <setjmp.h>
  57 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
  58 #include <numa.h>
  59 #include <numaif.h>
  60 #endif
  61
  62 #include <rte_log.h>
  63 #include <rte_memory.h>
  64 #include <rte_memzone.h>
  65 #include <rte_launch.h>
  66 #include <rte_eal.h>
  67 #include <rte_eal_memconfig.h>
  68 #include <rte_per_lcore.h>
  69 #include <rte_lcore.h>
  70 #include <rte_common.h>
  71 #include <rte_string_fns.h>
  72
  73 #include "eal_private.h"
  74 #include "eal_internal_cfg.h"
  75 #include "eal_filesystem.h"
  76 #include "eal_hugepages.h"
  77
  78 #define PFN_MASK_SIZE   8
  79
  80 #ifdef RTE_LIBRTE_XEN_DOM0
  81 int rte_xen_dom0_supported(void)
  82 {
  83         return internal_config.xen_dom0_support;
  84 }
  85 #endif
  86
  87 /**
  88  * @file
  89  * Huge page mapping under linux
  90  *
  91  * To reserve a big contiguous amount of memory, we use the hugepage
  92  * feature of linux. For that, we need to have hugetlbfs mounted. This
  93  * code will create many files in this directory (one per page) and
  94  * map them in virtual memory. For each page, we will retrieve its
  95  * physical address and remap it in order to have a virtual contiguous
  96  * zone as well as a physical contiguous zone.
  97  */
  98
  99 static uint64_t baseaddr_offset;
 100
 101 static bool phys_addrs_available = true;
 102
 103 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
 104
 105 static void
 106 test_phys_addrs_available(void)
 107 {
 108         uint64_t tmp;
 109         phys_addr_t physaddr;
 110
 111         /* For dom0, phys addresses can always be available */
 112         if (rte_xen_dom0_supported())
 113                 return;
 114
 115         physaddr = rte_mem_virt2phy(&tmp);
 116         if (physaddr == RTE_BAD_PHYS_ADDR) {
 117                 RTE_LOG(ERR, EAL,
 118                         "Cannot obtain physical addresses: %s. "
 119                         "Only vfio will function.\n",
 120                         strerror(errno));
 121                 phys_addrs_available = false;
 122         }
 123 }
 124
 125 /*
 126  * Get physical address of any mapped virtual address in the current process.
 127  */
 128 phys_addr_t
 129 rte_mem_virt2phy(const void *virtaddr)
 130 {
 131         int fd, retval;
 132         uint64_t page, physaddr;
 133         unsigned long virt_pfn;
 134         int page_size;
 135         off_t offset;
 136
 137         /* when using dom0, /proc/self/pagemap always returns 0, check in
 138          * dpdk memory by browsing the memsegs */
 139         if (rte_xen_dom0_supported()) {
 140                 struct rte_mem_config *mcfg;
 141                 struct rte_memseg *memseg;
 142                 unsigned i;
 143
 144                 mcfg = rte_eal_get_configuration()->mem_config;
 145                 for (i = 0; i < RTE_MAX_MEMSEG; i++) {
 146                         memseg = &mcfg->memseg[i];
 147                         if (memseg->addr == NULL)
 148                                 break;
 149                         if (virtaddr > memseg->addr &&
 150                                         virtaddr < RTE_PTR_ADD(memseg->addr,
 151                                                 memseg->len)) {
 152                                 return memseg->phys_addr +
 153                                         RTE_PTR_DIFF(virtaddr, memseg->addr);
 154                         }
 155                 }
 156
 157                 return RTE_BAD_PHYS_ADDR;
 158         }
 159
 160         /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
 161         if (!phys_addrs_available)
 162                 return RTE_BAD_PHYS_ADDR;
 163
 164         /* standard page size */
 165         page_size = getpagesize();
 166
 167         fd = open("/proc/self/pagemap", O_RDONLY);
 168         if (fd < 0) {
 169                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 170                         __func__, strerror(errno));
 171                 return RTE_BAD_PHYS_ADDR;
 172         }
 173
 174         virt_pfn = (unsigned long)virtaddr / page_size;
 175         offset = sizeof(uint64_t) * virt_pfn;
 176         if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 177                 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 178                                 __func__, strerror(errno));
 179                 close(fd);
 180                 return RTE_BAD_PHYS_ADDR;
 181         }
 182
 183         retval = read(fd, &page, PFN_MASK_SIZE);
 184         close(fd);
 185         if (retval < 0) {
 186                 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 187                                 __func__, strerror(errno));
 188                 return RTE_BAD_PHYS_ADDR;
 189         } else if (retval != PFN_MASK_SIZE) {
 190                 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
 191                                 "but expected %d:\n",
 192                                 __func__, retval, PFN_MASK_SIZE);
 193                 return RTE_BAD_PHYS_ADDR;
 194         }
 195
 196         /*
 197          * the pfn (page frame number) are bits 0-54 (see
 198          * pagemap.txt in linux Documentation)
 199          */
 200         if ((page & 0x7fffffffffffffULL) == 0)
 201                 return RTE_BAD_PHYS_ADDR;
 202
 203         physaddr = ((page & 0x7fffffffffffffULL) * page_size)
 204                 + ((unsigned long)virtaddr % page_size);
 205
 206         return physaddr;
 207 }
 208
 209 /*
 210  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 211  * it by browsing the /proc/self/pagemap special file.
 212  */
 213 static int
 214 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 215 {
 216         unsigned int i;
 217         phys_addr_t addr;
 218
 219         for (i = 0; i < hpi->num_pages[0]; i++) {
 220                 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
 221                 if (addr == RTE_BAD_PHYS_ADDR)
 222                         return -1;
 223                 hugepg_tbl[i].physaddr = addr;
 224         }
 225         return 0;
 226 }
 227
 228 /*
 229  * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
 230  */
 231 static int
 232 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 233 {
 234         unsigned int i;
 235         static phys_addr_t addr;
 236
 237         for (i = 0; i < hpi->num_pages[0]; i++) {
 238                 hugepg_tbl[i].physaddr = addr;
 239                 addr += hugepg_tbl[i].size;
 240         }
 241         return 0;
 242 }
 243
 244 /*
 245  * Check whether address-space layout randomization is enabled in
 246  * the kernel. This is important for multi-process as it can prevent
 247  * two processes mapping data to the same virtual address
 248  * Returns:
 249  *    0 - address space randomization disabled
 250  *    1/2 - address space randomization enabled
 251  *    negative error code on error
 252  */
 253 static int
 254 aslr_enabled(void)
 255 {
 256         char c;
 257         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 258         if (fd < 0)
 259                 return -errno;
 260         retval = read(fd, &c, 1);
 261         close(fd);
 262         if (retval < 0)
 263                 return -errno;
 264         if (retval == 0)
 265                 return -EIO;
 266         switch (c) {
 267                 case '0' : return 0;
 268                 case '1' : return 1;
 269                 case '2' : return 2;
 270                 default: return -EINVAL;
 271         }
 272 }
 273
 274 /*
 275  * Try to mmap *size bytes in /dev/zero. If it is successful, return the
 276  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 277  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 278  * 0. In this case, return NULL. Note: this function returns an address
 279  * which is a multiple of hugepage size.
 280  */
 281 static void *
 282 get_virtual_area(size_t *size, size_t hugepage_sz)
 283 {
 284         void *addr;
 285         int fd;
 286         long aligned_addr;
 287
 288         if (internal_config.base_virtaddr != 0) {
 289                 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
 290                                 baseaddr_offset);
 291         }
 292         else addr = NULL;
 293
 294         RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
 295
 296         fd = open("/dev/zero", O_RDONLY);
 297         if (fd < 0){
 298                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 299                 return NULL;
 300         }
 301         do {
 302                 addr = mmap(addr,
 303                                 (*size) + hugepage_sz, PROT_READ,
 304 #ifdef RTE_ARCH_PPC_64
 305                                 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
 306 #else
 307                                 MAP_PRIVATE,
 308 #endif
 309                                 fd, 0);
 310                 if (addr == MAP_FAILED)
 311                         *size -= hugepage_sz;
 312         } while (addr == MAP_FAILED && *size > 0);
 313
 314         if (addr == MAP_FAILED) {
 315                 close(fd);
 316                 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
 317                         strerror(errno));
 318                 return NULL;
 319         }
 320
 321         munmap(addr, (*size) + hugepage_sz);
 322         close(fd);
 323
 324         /* align addr to a huge page size boundary */
 325         aligned_addr = (long)addr;
 326         aligned_addr += (hugepage_sz - 1);
 327         aligned_addr &= (~(hugepage_sz - 1));
 328         addr = (void *)(aligned_addr);
 329
 330         RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
 331                 addr, *size);
 332
 333         /* increment offset */
 334         baseaddr_offset += *size;
 335
 336         return addr;
 337 }
 338
 339 static sigjmp_buf huge_jmpenv;
 340
 341 static void huge_sigbus_handler(int signo __rte_unused)
 342 {
 343         siglongjmp(huge_jmpenv, 1);
 344 }
 345
 346 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
 347  * non-static local variable in the stack frame calling sigsetjmp might be
 348  * clobbered by a call to longjmp.
 349  */
 350 static int huge_wrap_sigsetjmp(void)
 351 {
 352         return sigsetjmp(huge_jmpenv, 1);
 353 }
 354
 355 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 356 /* Callback for numa library. */
 357 void numa_error(char *where)
 358 {
 359         RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
 360 }
 361 #endif
 362
 363 /*
 364  * Mmap all hugepages of hugepage table: it first open a file in
 365  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 366  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 367  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 368  * map continguous physical blocks in contiguous virtual blocks.
 369  */
 370 static unsigned
 371 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
 372                   uint64_t *essential_memory __rte_unused, int orig)
 373 {
 374         int fd;
 375         unsigned i;
 376         void *virtaddr;
 377         void *vma_addr = NULL;
 378         size_t vma_len = 0;
 379 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 380         int node_id = -1;
 381         int essential_prev = 0;
 382         int oldpolicy;
 383         struct bitmask *oldmask = numa_allocate_nodemask();
 384         bool have_numa = true;
 385         unsigned long maxnode = 0;
 386
 387         /* Check if kernel supports NUMA. */
 388         if (numa_available() != 0) {
 389                 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
 390                 have_numa = false;
 391         }
 392
 393         if (orig && have_numa) {
 394                 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
 395                 if (get_mempolicy(&oldpolicy, oldmask->maskp,
 396                                   oldmask->size + 1, 0, 0) < 0) {
 397                         RTE_LOG(ERR, EAL,
 398                                 "Failed to get current mempolicy: %s. "
 399                                 "Assuming MPOL_DEFAULT.\n", strerror(errno));
 400                         oldpolicy = MPOL_DEFAULT;
 401                 }
 402                 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 403                         if (internal_config.socket_mem[i])
 404                                 maxnode = i + 1;
 405         }
 406 #endif
 407
 408         for (i = 0; i < hpi->num_pages[0]; i++) {
 409                 uint64_t hugepage_sz = hpi->hugepage_sz;
 410
 411 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 412                 if (maxnode) {
 413                         unsigned int j;
 414
 415                         for (j = 0; j < maxnode; j++)
 416                                 if (essential_memory[j])
 417                                         break;
 418
 419                         if (j == maxnode) {
 420                                 node_id = (node_id + 1) % maxnode;
 421                                 while (!internal_config.socket_mem[node_id]) {
 422                                         node_id++;
 423                                         node_id %= maxnode;
 424                                 }
 425                                 essential_prev = 0;
 426                         } else {
 427                                 node_id = j;
 428                                 essential_prev = essential_memory[j];
 429
 430                                 if (essential_memory[j] < hugepage_sz)
 431                                         essential_memory[j] = 0;
 432                                 else
 433                                         essential_memory[j] -= hugepage_sz;
 434                         }
 435
 436                         RTE_LOG(DEBUG, EAL,
 437                                 "Setting policy MPOL_PREFERRED for socket %d\n",
 438                                 node_id);
 439                         numa_set_preferred(node_id);
 440                 }
 441 #endif
 442
 443                 if (orig) {
 444                         hugepg_tbl[i].file_id = i;
 445                         hugepg_tbl[i].size = hugepage_sz;
 446                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 447                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 448                                         hugepg_tbl[i].file_id);
 449                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 450                 }
 451 #ifndef RTE_ARCH_64
 452                 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
 453                  * original map address as final map address.
 454                  */
 455                 else if ((hugepage_sz == RTE_PGSIZE_1G)
 456                         || (hugepage_sz == RTE_PGSIZE_16G)) {
 457                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 458                         hugepg_tbl[i].orig_va = NULL;
 459                         continue;
 460                 }
 461 #endif
 462                 else if (vma_len == 0) {
 463                         unsigned j, num_pages;
 464
 465                         /* reserve a virtual area for next contiguous
 466                          * physical block: count the number of
 467                          * contiguous physical pages. */
 468                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 469 #ifdef RTE_ARCH_PPC_64
 470                                 /* The physical addresses are sorted in
 471                                  * descending order on PPC64 */
 472                                 if (hugepg_tbl[j].physaddr !=
 473                                     hugepg_tbl[j-1].physaddr - hugepage_sz)
 474                                         break;
 475 #else
 476                                 if (hugepg_tbl[j].physaddr !=
 477                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 478                                         break;
 479 #endif
 480                         }
 481                         num_pages = j - i;
 482                         vma_len = num_pages * hugepage_sz;
 483
 484                         /* get the biggest virtual memory area up to
 485                          * vma_len. If it fails, vma_addr is NULL, so
 486                          * let the kernel provide the address. */
 487                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 488                         if (vma_addr == NULL)
 489                                 vma_len = hugepage_sz;
 490                 }
 491
 492                 /* try to create hugepage file */
 493                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
 494                 if (fd < 0) {
 495                         RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
 496                                         strerror(errno));
 497                         goto out;
 498                 }
 499
 500                 /* map the segment, and populate page tables,
 501                  * the kernel fills this segment with zeros */
 502                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 503                                 MAP_SHARED | MAP_POPULATE, fd, 0);
 504                 if (virtaddr == MAP_FAILED) {
 505                         RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
 506                                         strerror(errno));
 507                         close(fd);
 508                         goto out;
 509                 }
 510
 511                 if (orig) {
 512                         hugepg_tbl[i].orig_va = virtaddr;
 513                 }
 514                 else {
 515                         hugepg_tbl[i].final_va = virtaddr;
 516                 }
 517
 518                 if (orig) {
 519                         /* In linux, hugetlb limitations, like cgroup, are
 520                          * enforced at fault time instead of mmap(), even
 521                          * with the option of MAP_POPULATE. Kernel will send
 522                          * a SIGBUS signal. To avoid to be killed, save stack
 523                          * environment here, if SIGBUS happens, we can jump
 524                          * back here.
 525                          */
 526                         if (huge_wrap_sigsetjmp()) {
 527                                 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
 528                                         "hugepages of size %u MB\n",
 529                                         (unsigned)(hugepage_sz / 0x100000));
 530                                 munmap(virtaddr, hugepage_sz);
 531                                 close(fd);
 532                                 unlink(hugepg_tbl[i].filepath);
 533 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 534                                 if (maxnode)
 535                                         essential_memory[node_id] =
 536                                                 essential_prev;
 537 #endif
 538                                 goto out;
 539                         }
 540                         *(int *)virtaddr = 0;
 541                 }
 542
 543
 544                 /* set shared flock on the file. */
 545                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 546                         RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
 547                                 __func__, strerror(errno));
 548                         close(fd);
 549                         goto out;
 550                 }
 551
 552                 close(fd);
 553
 554                 vma_addr = (char *)vma_addr + hugepage_sz;
 555                 vma_len -= hugepage_sz;
 556         }
 557
 558 out:
 559 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 560         if (maxnode) {
 561                 RTE_LOG(DEBUG, EAL,
 562                         "Restoring previous memory policy: %d\n", oldpolicy);
 563                 if (oldpolicy == MPOL_DEFAULT) {
 564                         numa_set_localalloc();
 565                 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
 566                                          oldmask->size + 1) < 0) {
 567                         RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
 568                                 strerror(errno));
 569                         numa_set_localalloc();
 570                 }
 571         }
 572         numa_free_cpumask(oldmask);
 573 #endif
 574         return i;
 575 }
 576
 577 /* Unmap all hugepages from original mapping */
 578 static int
 579 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 580 {
 581         unsigned i;
 582         for (i = 0; i < hpi->num_pages[0]; i++) {
 583                 if (hugepg_tbl[i].orig_va) {
 584                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 585                         hugepg_tbl[i].orig_va = NULL;
 586                 }
 587         }
 588         return 0;
 589 }
 590
 591 /*
 592  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 593  * page.
 594  */
 595 static int
 596 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 597 {
 598         int socket_id;
 599         char *end, *nodestr;
 600         unsigned i, hp_count = 0;
 601         uint64_t virt_addr;
 602         char buf[BUFSIZ];
 603         char hugedir_str[PATH_MAX];
 604         FILE *f;
 605
 606         f = fopen("/proc/self/numa_maps", "r");
 607         if (f == NULL) {
 608                 RTE_LOG(NOTICE, EAL, "cannot open /proc/self/numa_maps,"
 609                                 " consider that all memory is in socket_id 0\n");
 610                 return 0;
 611         }
 612
 613         snprintf(hugedir_str, sizeof(hugedir_str),
 614                         "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
 615
 616         /* parse numa map */
 617         while (fgets(buf, sizeof(buf), f) != NULL) {
 618
 619                 /* ignore non huge page */
 620                 if (strstr(buf, " huge ") == NULL &&
 621                                 strstr(buf, hugedir_str) == NULL)
 622                         continue;
 623
 624                 /* get zone addr */
 625                 virt_addr = strtoull(buf, &end, 16);
 626                 if (virt_addr == 0 || end == buf) {
 627                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 628                         goto error;
 629                 }
 630
 631                 /* get node id (socket id) */
 632                 nodestr = strstr(buf, " N");
 633                 if (nodestr == NULL) {
 634                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 635                         goto error;
 636                 }
 637                 nodestr += 2;
 638                 end = strstr(nodestr, "=");
 639                 if (end == NULL) {
 640                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 641                         goto error;
 642                 }
 643                 end[0] = '\0';
 644                 end = NULL;
 645
 646                 socket_id = strtoul(nodestr, &end, 0);
 647                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 648                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 649                         goto error;
 650                 }
 651
 652                 /* if we find this page in our mappings, set socket_id */
 653                 for (i = 0; i < hpi->num_pages[0]; i++) {
 654                         void *va = (void *)(unsigned long)virt_addr;
 655                         if (hugepg_tbl[i].orig_va == va) {
 656                                 hugepg_tbl[i].socket_id = socket_id;
 657                                 hp_count++;
 658 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 659                                 RTE_LOG(DEBUG, EAL,
 660                                         "Hugepage %s is on socket %d\n",
 661                                         hugepg_tbl[i].filepath, socket_id);
 662 #endif
 663                         }
 664                 }
 665         }
 666
 667         if (hp_count < hpi->num_pages[0])
 668                 goto error;
 669
 670         fclose(f);
 671         return 0;
 672
 673 error:
 674         fclose(f);
 675         return -1;
 676 }
 677
 678 static int
 679 cmp_physaddr(const void *a, const void *b)
 680 {
 681 #ifndef RTE_ARCH_PPC_64
 682         const struct hugepage_file *p1 = a;
 683         const struct hugepage_file *p2 = b;
 684 #else
 685         /* PowerPC needs memory sorted in reverse order from x86 */
 686         const struct hugepage_file *p1 = b;
 687         const struct hugepage_file *p2 = a;
 688 #endif
 689         if (p1->physaddr < p2->physaddr)
 690                 return -1;
 691         else if (p1->physaddr > p2->physaddr)
 692                 return 1;
 693         else
 694                 return 0;
 695 }
 696
 697 /*
 698  * Uses mmap to create a shared memory area for storage of data
 699  * Used in this file to store the hugepage file map on disk
 700  */
 701 static void *
 702 create_shared_memory(const char *filename, const size_t mem_size)
 703 {
 704         void *retval;
 705         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 706         if (fd < 0)
 707                 return NULL;
 708         if (ftruncate(fd, mem_size) < 0) {
 709                 close(fd);
 710                 return NULL;
 711         }
 712         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 713         close(fd);
 714         return retval;
 715 }
 716
 717 /*
 718  * this copies *active* hugepages from one hugepage table to another.
 719  * destination is typically the shared memory.
 720  */
 721 static int
 722 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
 723                 const struct hugepage_file * src, int src_size)
 724 {
 725         int src_pos, dst_pos = 0;
 726
 727         for (src_pos = 0; src_pos < src_size; src_pos++) {
 728                 if (src[src_pos].final_va != NULL) {
 729                         /* error on overflow attempt */
 730                         if (dst_pos == dest_size)
 731                                 return -1;
 732                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
 733                         dst_pos++;
 734                 }
 735         }
 736         return 0;
 737 }
 738
 739 static int
 740 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
 741                 unsigned num_hp_info)
 742 {
 743         unsigned socket, size;
 744         int page, nrpages = 0;
 745
 746         /* get total number of hugepages */
 747         for (size = 0; size < num_hp_info; size++)
 748                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 749                         nrpages +=
 750                         internal_config.hugepage_info[size].num_pages[socket];
 751
 752         for (page = 0; page < nrpages; page++) {
 753                 struct hugepage_file *hp = &hugepg_tbl[page];
 754
 755                 if (hp->final_va != NULL && unlink(hp->filepath)) {
 756                         RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
 757                                 __func__, hp->filepath, strerror(errno));
 758                 }
 759         }
 760         return 0;
 761 }
 762
 763 /*
 764  * unmaps hugepages that are not going to be used. since we originally allocate
 765  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 766  */
 767 static int
 768 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
 769                 struct hugepage_info *hpi,
 770                 unsigned num_hp_info)
 771 {
 772         unsigned socket, size;
 773         int page, nrpages = 0;
 774
 775         /* get total number of hugepages */
 776         for (size = 0; size < num_hp_info; size++)
 777                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 778                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 779
 780         for (size = 0; size < num_hp_info; size++) {
 781                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 782                         unsigned pages_found = 0;
 783
 784                         /* traverse until we have unmapped all the unused pages */
 785                         for (page = 0; page < nrpages; page++) {
 786                                 struct hugepage_file *hp = &hugepg_tbl[page];
 787
 788                                 /* find a page that matches the criteria */
 789                                 if ((hp->size == hpi[size].hugepage_sz) &&
 790                                                 (hp->socket_id == (int) socket)) {
 791
 792                                         /* if we skipped enough pages, unmap the rest */
 793                                         if (pages_found == hpi[size].num_pages[socket]) {
 794                                                 uint64_t unmap_len;
 795
 796                                                 unmap_len = hp->size;
 797
 798                                                 /* get start addr and len of the remaining segment */
 799                                                 munmap(hp->final_va, (size_t) unmap_len);
 800
 801                                                 hp->final_va = NULL;
 802                                                 if (unlink(hp->filepath) == -1) {
 803                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 804                                                                         __func__, hp->filepath, strerror(errno));
 805                                                         return -1;
 806                                                 }
 807                                         } else {
 808                                                 /* lock the page and skip */
 809                                                 pages_found++;
 810                                         }
 811
 812                                 } /* match page */
 813                         } /* foreach page */
 814                 } /* foreach socket */
 815         } /* foreach pagesize */
 816
 817         return 0;
 818 }
 819
 820 static inline uint64_t
 821 get_socket_mem_size(int socket)
 822 {
 823         uint64_t size = 0;
 824         unsigned i;
 825
 826         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 827                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 828                 if (hpi->hugedir != NULL)
 829                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 830         }
 831
 832         return size;
 833 }
 834
 835 /*
 836  * This function is a NUMA-aware equivalent of calc_num_pages.
 837  * It takes in the list of hugepage sizes and the
 838  * number of pages thereof, and calculates the best number of
 839  * pages of each size to fulfill the request for <memory> ram
 840  */
 841 static int
 842 calc_num_pages_per_socket(uint64_t * memory,
 843                 struct hugepage_info *hp_info,
 844                 struct hugepage_info *hp_used,
 845                 unsigned num_hp_info)
 846 {
 847         unsigned socket, j, i = 0;
 848         unsigned requested, available;
 849         int total_num_pages = 0;
 850         uint64_t remaining_mem, cur_mem;
 851         uint64_t total_mem = internal_config.memory;
 852
 853         if (num_hp_info == 0)
 854                 return -1;
 855
 856         /* if specific memory amounts per socket weren't requested */
 857         if (internal_config.force_sockets == 0) {
 858                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
 859                 size_t default_size, total_size;
 860                 unsigned lcore_id;
 861
 862                 /* Compute number of cores per socket */
 863                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
 864                 RTE_LCORE_FOREACH(lcore_id) {
 865                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
 866                 }
 867
 868                 /*
 869                  * Automatically spread requested memory amongst detected sockets according
 870                  * to number of cores from cpu mask present on each socket
 871                  */
 872                 total_size = internal_config.memory;
 873                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 874
 875                         /* Set memory amount per socket */
 876                         default_size = (internal_config.memory * cpu_per_socket[socket])
 877                                         / rte_lcore_count();
 878
 879                         /* Limit to maximum available memory on socket */
 880                         default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
 881
 882                         /* Update sizes */
 883                         memory[socket] = default_size;
 884                         total_size -= default_size;
 885                 }
 886
 887                 /*
 888                  * If some memory is remaining, try to allocate it by getting all
 889                  * available memory from sockets, one after the other
 890                  */
 891                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 892                         /* take whatever is available */
 893                         default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
 894                                                total_size);
 895
 896                         /* Update sizes */
 897                         memory[socket] += default_size;
 898                         total_size -= default_size;
 899                 }
 900         }
 901
 902         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 903                 /* skips if the memory on specific socket wasn't requested */
 904                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 905                         hp_used[i].hugedir = hp_info[i].hugedir;
 906                         hp_used[i].num_pages[socket] = RTE_MIN(
 907                                         memory[socket] / hp_info[i].hugepage_sz,
 908                                         hp_info[i].num_pages[socket]);
 909
 910                         cur_mem = hp_used[i].num_pages[socket] *
 911                                         hp_used[i].hugepage_sz;
 912
 913                         memory[socket] -= cur_mem;
 914                         total_mem -= cur_mem;
 915
 916                         total_num_pages += hp_used[i].num_pages[socket];
 917
 918                         /* check if we have met all memory requests */
 919                         if (memory[socket] == 0)
 920                                 break;
 921
 922                         /* check if we have any more pages left at this size, if so
 923                          * move on to next size */
 924                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
 925                                 continue;
 926                         /* At this point we know that there are more pages available that are
 927                          * bigger than the memory we want, so lets see if we can get enough
 928                          * from other page sizes.
 929                          */
 930                         remaining_mem = 0;
 931                         for (j = i+1; j < num_hp_info; j++)
 932                                 remaining_mem += hp_info[j].hugepage_sz *
 933                                 hp_info[j].num_pages[socket];
 934
 935                         /* is there enough other memory, if not allocate another page and quit */
 936                         if (remaining_mem < memory[socket]){
 937                                 cur_mem = RTE_MIN(memory[socket],
 938                                                 hp_info[i].hugepage_sz);
 939                                 memory[socket] -= cur_mem;
 940                                 total_mem -= cur_mem;
 941                                 hp_used[i].num_pages[socket]++;
 942                                 total_num_pages++;
 943                                 break; /* we are done with this socket*/
 944                         }
 945                 }
 946                 /* if we didn't satisfy all memory requirements per socket */
 947                 if (memory[socket] > 0) {
 948                         /* to prevent icc errors */
 949                         requested = (unsigned) (internal_config.socket_mem[socket] /
 950                                         0x100000);
 951                         available = requested -
 952                                         ((unsigned) (memory[socket] / 0x100000));
 953                         RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
 954                                         "Requested: %uMB, available: %uMB\n", socket,
 955                                         requested, available);
 956                         return -1;
 957                 }
 958         }
 959
 960         /* if we didn't satisfy total memory requirements */
 961         if (total_mem > 0) {
 962                 requested = (unsigned) (internal_config.memory / 0x100000);
 963                 available = requested - (unsigned) (total_mem / 0x100000);
 964                 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
 965                                 " available: %uMB\n", requested, available);
 966                 return -1;
 967         }
 968         return total_num_pages;
 969 }
 970
 971 static inline size_t
 972 eal_get_hugepage_mem_size(void)
 973 {
 974         uint64_t size = 0;
 975         unsigned i, j;
 976
 977         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 978                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 979                 if (hpi->hugedir != NULL) {
 980                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 981                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 982                         }
 983                 }
 984         }
 985
 986         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 987 }
 988
 989 static struct sigaction huge_action_old;
 990 static int huge_need_recover;
 991
 992 static void
 993 huge_register_sigbus(void)
 994 {
 995         sigset_t mask;
 996         struct sigaction action;
 997
 998         sigemptyset(&mask);
 999         sigaddset(&mask, SIGBUS);
1000         action.sa_flags = 0;
1001         action.sa_mask = mask;
1002         action.sa_handler = huge_sigbus_handler;
1003
1004         huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
1005 }
1006
1007 static void
1008 huge_recover_sigbus(void)
1009 {
1010         if (huge_need_recover) {
1011                 sigaction(SIGBUS, &huge_action_old, NULL);
1012                 huge_need_recover = 0;
1013         }
1014 }
1015
1016 /*
1017  * Prepare physical memory mapping: fill configuration structure with
1018  * these infos, return 0 on success.
1019  *  1. map N huge pages in separate files in hugetlbfs
1020  *  2. find associated physical addr
1021  *  3. find associated NUMA socket ID
1022  *  4. sort all huge pages by physical address
1023  *  5. remap these N huge pages in the correct order
1024  *  6. unmap the first mapping
1025  *  7. fill memsegs in configuration with contiguous zones
1026  */
1027 int
1028 rte_eal_hugepage_init(void)
1029 {
1030         struct rte_mem_config *mcfg;
1031         struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1032         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1033
1034         uint64_t memory[RTE_MAX_NUMA_NODES];
1035
1036         unsigned hp_offset;
1037         int i, j, new_memseg;
1038         int nr_hugefiles, nr_hugepages = 0;
1039         void *addr;
1040
1041         test_phys_addrs_available();
1042
1043         memset(used_hp, 0, sizeof(used_hp));
1044
1045         /* get pointer to global configuration */
1046         mcfg = rte_eal_get_configuration()->mem_config;
1047
1048         /* hugetlbfs can be disabled */
1049         if (internal_config.no_hugetlbfs) {
1050                 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1051                                 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1052                 if (addr == MAP_FAILED) {
1053                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1054                                         strerror(errno));
1055                         return -1;
1056                 }
1057                 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
1058                 mcfg->memseg[0].addr = addr;
1059                 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1060                 mcfg->memseg[0].len = internal_config.memory;
1061                 mcfg->memseg[0].socket_id = 0;
1062                 return 0;
1063         }
1064
1065 /* check if app runs on Xen Dom0 */
1066         if (internal_config.xen_dom0_support) {
1067 #ifdef RTE_LIBRTE_XEN_DOM0
1068                 /* use dom0_mm kernel driver to init memory */
1069                 if (rte_xen_dom0_memory_init() < 0)
1070                         return -1;
1071                 else
1072                         return 0;
1073 #endif
1074         }
1075
1076         /* calculate total number of hugepages available. at this point we haven't
1077          * yet started sorting them so they all are on socket 0 */
1078         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1079                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1080                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1081
1082                 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1083         }
1084
1085         /*
1086          * allocate a memory area for hugepage table.
1087          * this isn't shared memory yet. due to the fact that we need some
1088          * processing done on these pages, shared memory will be created
1089          * at a later stage.
1090          */
1091         tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1092         if (tmp_hp == NULL)
1093                 goto fail;
1094
1095         memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1096
1097         hp_offset = 0; /* where we start the current page size entries */
1098
1099         huge_register_sigbus();
1100
1101         /* make a copy of socket_mem, needed for balanced allocation. */
1102         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1103                 memory[i] = internal_config.socket_mem[i];
1104
1105
1106         /* map all hugepages and sort them */
1107         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1108                 unsigned pages_old, pages_new;
1109                 struct hugepage_info *hpi;
1110
1111                 /*
1112                  * we don't yet mark hugepages as used at this stage, so
1113                  * we just map all hugepages available to the system
1114                  * all hugepages are still located on socket 0
1115                  */
1116                 hpi = &internal_config.hugepage_info[i];
1117
1118                 if (hpi->num_pages[0] == 0)
1119                         continue;
1120
1121                 /* map all hugepages available */
1122                 pages_old = hpi->num_pages[0];
1123                 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi,
1124                                               memory, 1);
1125                 if (pages_new < pages_old) {
1126                         RTE_LOG(DEBUG, EAL,
1127                                 "%d not %d hugepages of size %u MB allocated\n",
1128                                 pages_new, pages_old,
1129                                 (unsigned)(hpi->hugepage_sz / 0x100000));
1130
1131                         int pages = pages_old - pages_new;
1132
1133                         nr_hugepages -= pages;
1134                         hpi->num_pages[0] = pages_new;
1135                         if (pages_new == 0)
1136                                 continue;
1137                 }
1138
1139                 if (phys_addrs_available) {
1140                         /* find physical addresses for each hugepage */
1141                         if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1142                                 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1143                                         "for %u MB pages\n",
1144                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1145                                 goto fail;
1146                         }
1147                 } else {
1148                         /* set physical addresses for each hugepage */
1149                         if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1150                                 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1151                                         "for %u MB pages\n",
1152                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1153                                 goto fail;
1154                         }
1155                 }
1156
1157                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1158                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1159                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1160                         goto fail;
1161                 }
1162
1163                 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1164                       sizeof(struct hugepage_file), cmp_physaddr);
1165
1166                 /* remap all hugepages */
1167                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, NULL, 0) !=
1168                     hpi->num_pages[0]) {
1169                         RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1170                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1171                         goto fail;
1172                 }
1173
1174                 /* unmap original mappings */
1175                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1176                         goto fail;
1177
1178                 /* we have processed a num of hugepages of this size, so inc offset */
1179                 hp_offset += hpi->num_pages[0];
1180         }
1181
1182         huge_recover_sigbus();
1183
1184         if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1185                 internal_config.memory = eal_get_hugepage_mem_size();
1186
1187         nr_hugefiles = nr_hugepages;
1188
1189
1190         /* clean out the numbers of pages */
1191         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1192                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1193                         internal_config.hugepage_info[i].num_pages[j] = 0;
1194
1195         /* get hugepages for each socket */
1196         for (i = 0; i < nr_hugefiles; i++) {
1197                 int socket = tmp_hp[i].socket_id;
1198
1199                 /* find a hugepage info with right size and increment num_pages */
1200                 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1201                                 (int)internal_config.num_hugepage_sizes);
1202                 for (j = 0; j < nb_hpsizes; j++) {
1203                         if (tmp_hp[i].size ==
1204                                         internal_config.hugepage_info[j].hugepage_sz) {
1205                                 internal_config.hugepage_info[j].num_pages[socket]++;
1206                         }
1207                 }
1208         }
1209
1210         /* make a copy of socket_mem, needed for number of pages calculation */
1211         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1212                 memory[i] = internal_config.socket_mem[i];
1213
1214         /* calculate final number of pages */
1215         nr_hugepages = calc_num_pages_per_socket(memory,
1216                         internal_config.hugepage_info, used_hp,
1217                         internal_config.num_hugepage_sizes);
1218
1219         /* error if not enough memory available */
1220         if (nr_hugepages < 0)
1221                 goto fail;
1222
1223         /* reporting in! */
1224         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1225                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1226                         if (used_hp[i].num_pages[j] > 0) {
1227                                 RTE_LOG(DEBUG, EAL,
1228                                         "Requesting %u pages of size %uMB"
1229                                         " from socket %i\n",
1230                                         used_hp[i].num_pages[j],
1231                                         (unsigned)
1232                                         (used_hp[i].hugepage_sz / 0x100000),
1233                                         j);
1234                         }
1235                 }
1236         }
1237
1238         /* create shared memory */
1239         hugepage = create_shared_memory(eal_hugepage_info_path(),
1240                         nr_hugefiles * sizeof(struct hugepage_file));
1241
1242         if (hugepage == NULL) {
1243                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1244                 goto fail;
1245         }
1246         memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1247
1248         /*
1249          * unmap pages that we won't need (looks at used_hp).
1250          * also, sets final_va to NULL on pages that were unmapped.
1251          */
1252         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1253                         internal_config.num_hugepage_sizes) < 0) {
1254                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1255                 goto fail;
1256         }
1257
1258         /*
1259          * copy stuff from malloc'd hugepage* to the actual shared memory.
1260          * this procedure only copies those hugepages that have final_va
1261          * not NULL. has overflow protection.
1262          */
1263         if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1264                         tmp_hp, nr_hugefiles) < 0) {
1265                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1266                 goto fail;
1267         }
1268
1269         /* free the hugepage backing files */
1270         if (internal_config.hugepage_unlink &&
1271                 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1272                 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1273                 goto fail;
1274         }
1275
1276         /* free the temporary hugepage table */
1277         free(tmp_hp);
1278         tmp_hp = NULL;
1279
1280         /* first memseg index shall be 0 after incrementing it below */
1281         j = -1;
1282         for (i = 0; i < nr_hugefiles; i++) {
1283                 new_memseg = 0;
1284
1285                 /* if this is a new section, create a new memseg */
1286                 if (i == 0)
1287                         new_memseg = 1;
1288                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1289                         new_memseg = 1;
1290                 else if (hugepage[i].size != hugepage[i-1].size)
1291                         new_memseg = 1;
1292
1293 #ifdef RTE_ARCH_PPC_64
1294                 /* On PPC64 architecture, the mmap always start from higher
1295                  * virtual address to lower address. Here, both the physical
1296                  * address and virtual address are in descending order */
1297                 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1298                     hugepage[i].size)
1299                         new_memseg = 1;
1300                 else if (((unsigned long)hugepage[i-1].final_va -
1301                     (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1302                         new_memseg = 1;
1303 #else
1304                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1305                     hugepage[i].size)
1306                         new_memseg = 1;
1307                 else if (((unsigned long)hugepage[i].final_va -
1308                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1309                         new_memseg = 1;
1310 #endif
1311
1312                 if (new_memseg) {
1313                         j += 1;
1314                         if (j == RTE_MAX_MEMSEG)
1315                                 break;
1316
1317                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1318                         mcfg->memseg[j].addr = hugepage[i].final_va;
1319                         mcfg->memseg[j].len = hugepage[i].size;
1320                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1321                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1322                 }
1323                 /* continuation of previous memseg */
1324                 else {
1325 #ifdef RTE_ARCH_PPC_64
1326                 /* Use the phy and virt address of the last page as segment
1327                  * address for IBM Power architecture */
1328                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1329                         mcfg->memseg[j].addr = hugepage[i].final_va;
1330 #endif
1331                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1332                 }
1333                 hugepage[i].memseg_id = j;
1334         }
1335
1336         if (i < nr_hugefiles) {
1337                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1338                         "from %d requested\n"
1339                         "Current %s=%d is not enough\n"
1340                         "Please either increase it or request less amount "
1341                         "of memory.\n",
1342                         i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1343                         RTE_MAX_MEMSEG);
1344                 goto fail;
1345         }
1346
1347         munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1348
1349         return 0;
1350
1351 fail:
1352         huge_recover_sigbus();
1353         free(tmp_hp);
1354         if (hugepage != NULL)
1355                 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1356
1357         return -1;
1358 }
1359
1360 /*
1361  * uses fstat to report the size of a file on disk
1362  */
1363 static off_t
1364 getFileSize(int fd)
1365 {
1366         struct stat st;
1367         if (fstat(fd, &st) < 0)
1368                 return 0;
1369         return st.st_size;
1370 }
1371
1372 /*
1373  * This creates the memory mappings in the secondary process to match that of
1374  * the server process. It goes through each memory segment in the DPDK runtime
1375  * configuration and finds the hugepages which form that segment, mapping them
1376  * in order to form a contiguous block in the virtual memory space
1377  */
1378 int
1379 rte_eal_hugepage_attach(void)
1380 {
1381         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1382         struct hugepage_file *hp = NULL;
1383         unsigned num_hp = 0;
1384         unsigned i, s = 0; /* s used to track the segment number */
1385         unsigned max_seg = RTE_MAX_MEMSEG;
1386         off_t size = 0;
1387         int fd, fd_zero = -1, fd_hugepage = -1;
1388
1389         if (aslr_enabled() > 0) {
1390                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1391                                 "(ASLR) is enabled in the kernel.\n");
1392                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1393                                 "into secondary processes\n");
1394         }
1395
1396         test_phys_addrs_available();
1397
1398         if (internal_config.xen_dom0_support) {
1399 #ifdef RTE_LIBRTE_XEN_DOM0
1400                 if (rte_xen_dom0_memory_attach() < 0) {
1401                         RTE_LOG(ERR, EAL, "Failed to attach memory segments of primary "
1402                                         "process\n");
1403                         return -1;
1404                 }
1405                 return 0;
1406 #endif
1407         }
1408
1409         fd_zero = open("/dev/zero", O_RDONLY);
1410         if (fd_zero < 0) {
1411                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1412                 goto error;
1413         }
1414         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1415         if (fd_hugepage < 0) {
1416                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1417                 goto error;
1418         }
1419
1420         /* map all segments into memory to make sure we get the addrs */
1421         for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1422                 void *base_addr;
1423
1424                 /*
1425                  * the first memory segment with len==0 is the one that
1426                  * follows the last valid segment.
1427                  */
1428                 if (mcfg->memseg[s].len == 0)
1429                         break;
1430
1431                 /*
1432                  * fdzero is mmapped to get a contiguous block of virtual
1433                  * addresses of the appropriate memseg size.
1434                  * use mmap to get identical addresses as the primary process.
1435                  */
1436                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1437                                  PROT_READ,
1438 #ifdef RTE_ARCH_PPC_64
1439                                  MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
1440 #else
1441                                  MAP_PRIVATE,
1442 #endif
1443                                  fd_zero, 0);
1444                 if (base_addr == MAP_FAILED ||
1445                     base_addr != mcfg->memseg[s].addr) {
1446                         max_seg = s;
1447                         if (base_addr != MAP_FAILED) {
1448                                 /* errno is stale, don't use */
1449                                 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1450                                         "in /dev/zero at [%p], got [%p] - "
1451                                         "please use '--base-virtaddr' option\n",
1452                                         (unsigned long long)mcfg->memseg[s].len,
1453                                         mcfg->memseg[s].addr, base_addr);
1454                                 munmap(base_addr, mcfg->memseg[s].len);
1455                         } else {
1456                                 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1457                                         "in /dev/zero at [%p]: '%s'\n",
1458                                         (unsigned long long)mcfg->memseg[s].len,
1459                                         mcfg->memseg[s].addr, strerror(errno));
1460                         }
1461                         if (aslr_enabled() > 0) {
1462                                 RTE_LOG(ERR, EAL, "It is recommended to "
1463                                         "disable ASLR in the kernel "
1464                                         "and retry running both primary "
1465                                         "and secondary processes\n");
1466                         }
1467                         goto error;
1468                 }
1469         }
1470
1471         size = getFileSize(fd_hugepage);
1472         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1473         if (hp == MAP_FAILED) {
1474                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1475                 goto error;
1476         }
1477
1478         num_hp = size / sizeof(struct hugepage_file);
1479         RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1480
1481         s = 0;
1482         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1483                 void *addr, *base_addr;
1484                 uintptr_t offset = 0;
1485                 size_t mapping_size;
1486                 /*
1487                  * free previously mapped memory so we can map the
1488                  * hugepages into the space
1489                  */
1490                 base_addr = mcfg->memseg[s].addr;
1491                 munmap(base_addr, mcfg->memseg[s].len);
1492
1493                 /* find the hugepages for this segment and map them
1494                  * we don't need to worry about order, as the server sorted the
1495                  * entries before it did the second mmap of them */
1496                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1497                         if (hp[i].memseg_id == (int)s){
1498                                 fd = open(hp[i].filepath, O_RDWR);
1499                                 if (fd < 0) {
1500                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1501                                                 hp[i].filepath);
1502                                         goto error;
1503                                 }
1504                                 mapping_size = hp[i].size;
1505                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1506                                                 mapping_size, PROT_READ | PROT_WRITE,
1507                                                 MAP_SHARED, fd, 0);
1508                                 close(fd); /* close file both on success and on failure */
1509                                 if (addr == MAP_FAILED ||
1510                                                 addr != RTE_PTR_ADD(base_addr, offset)) {
1511                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1512                                                 hp[i].filepath);
1513                                         goto error;
1514                                 }
1515                                 offset+=mapping_size;
1516                         }
1517                 }
1518                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1519                                 (unsigned long long)mcfg->memseg[s].len);
1520                 s++;
1521         }
1522         /* unmap the hugepage config file, since we are done using it */
1523         munmap(hp, size);
1524         close(fd_zero);
1525         close(fd_hugepage);
1526         return 0;
1527
1528 error:
1529         for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1530                 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1531         if (hp != NULL && hp != MAP_FAILED)
1532                 munmap(hp, size);
1533         if (fd_zero >= 0)
1534                 close(fd_zero);
1535         if (fd_hugepage >= 0)
1536                 close(fd_hugepage);
1537         return -1;
1538 }
1539
1540 bool
1541 rte_eal_using_phys_addrs(void)
1542 {
1543         return phys_addrs_available;
1544 }