lib/librte_eal/linuxapp/eal/eal_memory.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   Copyright(c) 2013 6WIND.
   6  *   All rights reserved.
   7  *
   8  *   Redistribution and use in source and binary forms, with or without
   9  *   modification, are permitted provided that the following conditions
  10  *   are met:
  11  *
  12  *     * Redistributions of source code must retain the above copyright
  13  *       notice, this list of conditions and the following disclaimer.
  14  *     * Redistributions in binary form must reproduce the above copyright
  15  *       notice, this list of conditions and the following disclaimer in
  16  *       the documentation and/or other materials provided with the
  17  *       distribution.
  18  *     * Neither the name of Intel Corporation nor the names of its
  19  *       contributors may be used to endorse or promote products derived
  20  *       from this software without specific prior written permission.
  21  *
  22  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  25  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  26  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  27  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  28  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  29  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  30  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  31  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  32  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  33  */
  34
  35 #define _FILE_OFFSET_BITS 64
  36 #include <errno.h>
  37 #include <stdarg.h>
  38 #include <stdbool.h>
  39 #include <stdlib.h>
  40 #include <stdio.h>
  41 #include <stdint.h>
  42 #include <inttypes.h>
  43 #include <string.h>
  44 #include <sys/mman.h>
  45 #include <sys/types.h>
  46 #include <sys/stat.h>
  47 #include <sys/queue.h>
  48 #include <sys/file.h>
  49 #include <unistd.h>
  50 #include <limits.h>
  51 #include <sys/ioctl.h>
  52 #include <sys/time.h>
  53 #include <signal.h>
  54 #include <setjmp.h>
  55 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
  56 #include <numa.h>
  57 #include <numaif.h>
  58 #endif
  59
  60 #include <rte_log.h>
  61 #include <rte_memory.h>
  62 #include <rte_memzone.h>
  63 #include <rte_launch.h>
  64 #include <rte_eal.h>
  65 #include <rte_eal_memconfig.h>
  66 #include <rte_per_lcore.h>
  67 #include <rte_lcore.h>
  68 #include <rte_common.h>
  69 #include <rte_string_fns.h>
  70
  71 #include "eal_private.h"
  72 #include "eal_internal_cfg.h"
  73 #include "eal_filesystem.h"
  74 #include "eal_hugepages.h"
  75
  76 #define PFN_MASK_SIZE   8
  77
  78 #ifdef RTE_LIBRTE_XEN_DOM0
  79 int rte_xen_dom0_supported(void)
  80 {
  81         return internal_config.xen_dom0_support;
  82 }
  83 #endif
  84
  85 /**
  86  * @file
  87  * Huge page mapping under linux
  88  *
  89  * To reserve a big contiguous amount of memory, we use the hugepage
  90  * feature of linux. For that, we need to have hugetlbfs mounted. This
  91  * code will create many files in this directory (one per page) and
  92  * map them in virtual memory. For each page, we will retrieve its
  93  * physical address and remap it in order to have a virtual contiguous
  94  * zone as well as a physical contiguous zone.
  95  */
  96
  97 static uint64_t baseaddr_offset;
  98
  99 static bool phys_addrs_available = true;
 100
 101 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
 102
 103 static void
 104 test_phys_addrs_available(void)
 105 {
 106         uint64_t tmp;
 107         phys_addr_t physaddr;
 108
 109         /* For dom0, phys addresses can always be available */
 110         if (rte_xen_dom0_supported())
 111                 return;
 112
 113         if (!rte_eal_has_hugepages()) {
 114                 RTE_LOG(ERR, EAL,
 115                         "Started without hugepages support, physical addresses not available\n");
 116                 phys_addrs_available = false;
 117                 return;
 118         }
 119
 120         physaddr = rte_mem_virt2phy(&tmp);
 121         if (physaddr == RTE_BAD_PHYS_ADDR) {
 122                 RTE_LOG(ERR, EAL,
 123                         "Cannot obtain physical addresses: %s. "
 124                         "Only vfio will function.\n",
 125                         strerror(errno));
 126                 phys_addrs_available = false;
 127         }
 128 }
 129
 130 /*
 131  * Get physical address of any mapped virtual address in the current process.
 132  */
 133 phys_addr_t
 134 rte_mem_virt2phy(const void *virtaddr)
 135 {
 136         int fd, retval;
 137         uint64_t page, physaddr;
 138         unsigned long virt_pfn;
 139         int page_size;
 140         off_t offset;
 141
 142         if (rte_eal_iova_mode() == RTE_IOVA_VA)
 143                 return (uintptr_t)virtaddr;
 144
 145         /* when using dom0, /proc/self/pagemap always returns 0, check in
 146          * dpdk memory by browsing the memsegs */
 147         if (rte_xen_dom0_supported()) {
 148                 struct rte_mem_config *mcfg;
 149                 struct rte_memseg *memseg;
 150                 unsigned i;
 151
 152                 mcfg = rte_eal_get_configuration()->mem_config;
 153                 for (i = 0; i < RTE_MAX_MEMSEG; i++) {
 154                         memseg = &mcfg->memseg[i];
 155                         if (memseg->addr == NULL)
 156                                 break;
 157                         if (virtaddr > memseg->addr &&
 158                                         virtaddr < RTE_PTR_ADD(memseg->addr,
 159                                                 memseg->len)) {
 160                                 return memseg->phys_addr +
 161                                         RTE_PTR_DIFF(virtaddr, memseg->addr);
 162                         }
 163                 }
 164
 165                 return RTE_BAD_PHYS_ADDR;
 166         }
 167
 168         /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
 169         if (!phys_addrs_available)
 170                 return RTE_BAD_PHYS_ADDR;
 171
 172         /* standard page size */
 173         page_size = getpagesize();
 174
 175         fd = open("/proc/self/pagemap", O_RDONLY);
 176         if (fd < 0) {
 177                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 178                         __func__, strerror(errno));
 179                 return RTE_BAD_PHYS_ADDR;
 180         }
 181
 182         virt_pfn = (unsigned long)virtaddr / page_size;
 183         offset = sizeof(uint64_t) * virt_pfn;
 184         if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 185                 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 186                                 __func__, strerror(errno));
 187                 close(fd);
 188                 return RTE_BAD_PHYS_ADDR;
 189         }
 190
 191         retval = read(fd, &page, PFN_MASK_SIZE);
 192         close(fd);
 193         if (retval < 0) {
 194                 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 195                                 __func__, strerror(errno));
 196                 return RTE_BAD_PHYS_ADDR;
 197         } else if (retval != PFN_MASK_SIZE) {
 198                 RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
 199                                 "but expected %d:\n",
 200                                 __func__, retval, PFN_MASK_SIZE);
 201                 return RTE_BAD_PHYS_ADDR;
 202         }
 203
 204         /*
 205          * the pfn (page frame number) are bits 0-54 (see
 206          * pagemap.txt in linux Documentation)
 207          */
 208         if ((page & 0x7fffffffffffffULL) == 0)
 209                 return RTE_BAD_PHYS_ADDR;
 210
 211         physaddr = ((page & 0x7fffffffffffffULL) * page_size)
 212                 + ((unsigned long)virtaddr % page_size);
 213
 214         return physaddr;
 215 }
 216
 217 /*
 218  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 219  * it by browsing the /proc/self/pagemap special file.
 220  */
 221 static int
 222 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 223 {
 224         unsigned int i;
 225         phys_addr_t addr;
 226
 227         for (i = 0; i < hpi->num_pages[0]; i++) {
 228                 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
 229                 if (addr == RTE_BAD_PHYS_ADDR)
 230                         return -1;
 231                 hugepg_tbl[i].physaddr = addr;
 232         }
 233         return 0;
 234 }
 235
 236 /*
 237  * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
 238  */
 239 static int
 240 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 241 {
 242         unsigned int i;
 243         static phys_addr_t addr;
 244
 245         for (i = 0; i < hpi->num_pages[0]; i++) {
 246                 hugepg_tbl[i].physaddr = addr;
 247                 addr += hugepg_tbl[i].size;
 248         }
 249         return 0;
 250 }
 251
 252 /*
 253  * Check whether address-space layout randomization is enabled in
 254  * the kernel. This is important for multi-process as it can prevent
 255  * two processes mapping data to the same virtual address
 256  * Returns:
 257  *    0 - address space randomization disabled
 258  *    1/2 - address space randomization enabled
 259  *    negative error code on error
 260  */
 261 static int
 262 aslr_enabled(void)
 263 {
 264         char c;
 265         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 266         if (fd < 0)
 267                 return -errno;
 268         retval = read(fd, &c, 1);
 269         close(fd);
 270         if (retval < 0)
 271                 return -errno;
 272         if (retval == 0)
 273                 return -EIO;
 274         switch (c) {
 275                 case '0' : return 0;
 276                 case '1' : return 1;
 277                 case '2' : return 2;
 278                 default: return -EINVAL;
 279         }
 280 }
 281
 282 /*
 283  * Try to mmap *size bytes in /dev/zero. If it is successful, return the
 284  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 285  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 286  * 0. In this case, return NULL. Note: this function returns an address
 287  * which is a multiple of hugepage size.
 288  */
 289 static void *
 290 get_virtual_area(size_t *size, size_t hugepage_sz)
 291 {
 292         void *addr;
 293         int fd;
 294         long aligned_addr;
 295
 296         if (internal_config.base_virtaddr != 0) {
 297                 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
 298                                 baseaddr_offset);
 299         }
 300         else addr = NULL;
 301
 302         RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
 303
 304         fd = open("/dev/zero", O_RDONLY);
 305         if (fd < 0){
 306                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 307                 return NULL;
 308         }
 309         do {
 310                 addr = mmap(addr,
 311                                 (*size) + hugepage_sz, PROT_READ,
 312 #ifdef RTE_ARCH_PPC_64
 313                                 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
 314 #else
 315                                 MAP_PRIVATE,
 316 #endif
 317                                 fd, 0);
 318                 if (addr == MAP_FAILED)
 319                         *size -= hugepage_sz;
 320         } while (addr == MAP_FAILED && *size > 0);
 321
 322         if (addr == MAP_FAILED) {
 323                 close(fd);
 324                 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
 325                         strerror(errno));
 326                 return NULL;
 327         }
 328
 329         munmap(addr, (*size) + hugepage_sz);
 330         close(fd);
 331
 332         /* align addr to a huge page size boundary */
 333         aligned_addr = (long)addr;
 334         aligned_addr += (hugepage_sz - 1);
 335         aligned_addr &= (~(hugepage_sz - 1));
 336         addr = (void *)(aligned_addr);
 337
 338         RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
 339                 addr, *size);
 340
 341         /* increment offset */
 342         baseaddr_offset += *size;
 343
 344         return addr;
 345 }
 346
 347 static sigjmp_buf huge_jmpenv;
 348
 349 static void huge_sigbus_handler(int signo __rte_unused)
 350 {
 351         siglongjmp(huge_jmpenv, 1);
 352 }
 353
 354 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
 355  * non-static local variable in the stack frame calling sigsetjmp might be
 356  * clobbered by a call to longjmp.
 357  */
 358 static int huge_wrap_sigsetjmp(void)
 359 {
 360         return sigsetjmp(huge_jmpenv, 1);
 361 }
 362
 363 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 364 /* Callback for numa library. */
 365 void numa_error(char *where)
 366 {
 367         RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
 368 }
 369 #endif
 370
 371 /*
 372  * Mmap all hugepages of hugepage table: it first open a file in
 373  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 374  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 375  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 376  * map continguous physical blocks in contiguous virtual blocks.
 377  */
 378 static unsigned
 379 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
 380                   uint64_t *essential_memory __rte_unused, int orig)
 381 {
 382         int fd;
 383         unsigned i;
 384         void *virtaddr;
 385         void *vma_addr = NULL;
 386         size_t vma_len = 0;
 387 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 388         int node_id = -1;
 389         int essential_prev = 0;
 390         int oldpolicy;
 391         struct bitmask *oldmask = numa_allocate_nodemask();
 392         bool have_numa = true;
 393         unsigned long maxnode = 0;
 394
 395         /* Check if kernel supports NUMA. */
 396         if (numa_available() != 0) {
 397                 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
 398                 have_numa = false;
 399         }
 400
 401         if (orig && have_numa) {
 402                 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
 403                 if (get_mempolicy(&oldpolicy, oldmask->maskp,
 404                                   oldmask->size + 1, 0, 0) < 0) {
 405                         RTE_LOG(ERR, EAL,
 406                                 "Failed to get current mempolicy: %s. "
 407                                 "Assuming MPOL_DEFAULT.\n", strerror(errno));
 408                         oldpolicy = MPOL_DEFAULT;
 409                 }
 410                 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 411                         if (internal_config.socket_mem[i])
 412                                 maxnode = i + 1;
 413         }
 414 #endif
 415
 416         for (i = 0; i < hpi->num_pages[0]; i++) {
 417                 uint64_t hugepage_sz = hpi->hugepage_sz;
 418
 419 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 420                 if (maxnode) {
 421                         unsigned int j;
 422
 423                         for (j = 0; j < maxnode; j++)
 424                                 if (essential_memory[j])
 425                                         break;
 426
 427                         if (j == maxnode) {
 428                                 node_id = (node_id + 1) % maxnode;
 429                                 while (!internal_config.socket_mem[node_id]) {
 430                                         node_id++;
 431                                         node_id %= maxnode;
 432                                 }
 433                                 essential_prev = 0;
 434                         } else {
 435                                 node_id = j;
 436                                 essential_prev = essential_memory[j];
 437
 438                                 if (essential_memory[j] < hugepage_sz)
 439                                         essential_memory[j] = 0;
 440                                 else
 441                                         essential_memory[j] -= hugepage_sz;
 442                         }
 443
 444                         RTE_LOG(DEBUG, EAL,
 445                                 "Setting policy MPOL_PREFERRED for socket %d\n",
 446                                 node_id);
 447                         numa_set_preferred(node_id);
 448                 }
 449 #endif
 450
 451                 if (orig) {
 452                         hugepg_tbl[i].file_id = i;
 453                         hugepg_tbl[i].size = hugepage_sz;
 454                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 455                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 456                                         hugepg_tbl[i].file_id);
 457                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 458                 }
 459 #ifndef RTE_ARCH_64
 460                 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
 461                  * original map address as final map address.
 462                  */
 463                 else if ((hugepage_sz == RTE_PGSIZE_1G)
 464                         || (hugepage_sz == RTE_PGSIZE_16G)) {
 465                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 466                         hugepg_tbl[i].orig_va = NULL;
 467                         continue;
 468                 }
 469 #endif
 470                 else if (vma_len == 0) {
 471                         unsigned j, num_pages;
 472
 473                         /* reserve a virtual area for next contiguous
 474                          * physical block: count the number of
 475                          * contiguous physical pages. */
 476                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 477 #ifdef RTE_ARCH_PPC_64
 478                                 /* The physical addresses are sorted in
 479                                  * descending order on PPC64 */
 480                                 if (hugepg_tbl[j].physaddr !=
 481                                     hugepg_tbl[j-1].physaddr - hugepage_sz)
 482                                         break;
 483 #else
 484                                 if (hugepg_tbl[j].physaddr !=
 485                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 486                                         break;
 487 #endif
 488                         }
 489                         num_pages = j - i;
 490                         vma_len = num_pages * hugepage_sz;
 491
 492                         /* get the biggest virtual memory area up to
 493                          * vma_len. If it fails, vma_addr is NULL, so
 494                          * let the kernel provide the address. */
 495                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 496                         if (vma_addr == NULL)
 497                                 vma_len = hugepage_sz;
 498                 }
 499
 500                 /* try to create hugepage file */
 501                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
 502                 if (fd < 0) {
 503                         RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
 504                                         strerror(errno));
 505                         goto out;
 506                 }
 507
 508                 /* map the segment, and populate page tables,
 509                  * the kernel fills this segment with zeros */
 510                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 511                                 MAP_SHARED | MAP_POPULATE, fd, 0);
 512                 if (virtaddr == MAP_FAILED) {
 513                         RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
 514                                         strerror(errno));
 515                         close(fd);
 516                         goto out;
 517                 }
 518
 519                 if (orig) {
 520                         hugepg_tbl[i].orig_va = virtaddr;
 521                 }
 522                 else {
 523                         hugepg_tbl[i].final_va = virtaddr;
 524                 }
 525
 526                 if (orig) {
 527                         /* In linux, hugetlb limitations, like cgroup, are
 528                          * enforced at fault time instead of mmap(), even
 529                          * with the option of MAP_POPULATE. Kernel will send
 530                          * a SIGBUS signal. To avoid to be killed, save stack
 531                          * environment here, if SIGBUS happens, we can jump
 532                          * back here.
 533                          */
 534                         if (huge_wrap_sigsetjmp()) {
 535                                 RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
 536                                         "hugepages of size %u MB\n",
 537                                         (unsigned)(hugepage_sz / 0x100000));
 538                                 munmap(virtaddr, hugepage_sz);
 539                                 close(fd);
 540                                 unlink(hugepg_tbl[i].filepath);
 541 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 542                                 if (maxnode)
 543                                         essential_memory[node_id] =
 544                                                 essential_prev;
 545 #endif
 546                                 goto out;
 547                         }
 548                         *(int *)virtaddr = 0;
 549                 }
 550
 551
 552                 /* set shared flock on the file. */
 553                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 554                         RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
 555                                 __func__, strerror(errno));
 556                         close(fd);
 557                         goto out;
 558                 }
 559
 560                 close(fd);
 561
 562                 vma_addr = (char *)vma_addr + hugepage_sz;
 563                 vma_len -= hugepage_sz;
 564         }
 565
 566 out:
 567 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 568         if (maxnode) {
 569                 RTE_LOG(DEBUG, EAL,
 570                         "Restoring previous memory policy: %d\n", oldpolicy);
 571                 if (oldpolicy == MPOL_DEFAULT) {
 572                         numa_set_localalloc();
 573                 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
 574                                          oldmask->size + 1) < 0) {
 575                         RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
 576                                 strerror(errno));
 577                         numa_set_localalloc();
 578                 }
 579         }
 580         numa_free_cpumask(oldmask);
 581 #endif
 582         return i;
 583 }
 584
 585 /* Unmap all hugepages from original mapping */
 586 static int
 587 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 588 {
 589         unsigned i;
 590         for (i = 0; i < hpi->num_pages[0]; i++) {
 591                 if (hugepg_tbl[i].orig_va) {
 592                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 593                         hugepg_tbl[i].orig_va = NULL;
 594                 }
 595         }
 596         return 0;
 597 }
 598
 599 /*
 600  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 601  * page.
 602  */
 603 static int
 604 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 605 {
 606         int socket_id;
 607         char *end, *nodestr;
 608         unsigned i, hp_count = 0;
 609         uint64_t virt_addr;
 610         char buf[BUFSIZ];
 611         char hugedir_str[PATH_MAX];
 612         FILE *f;
 613
 614         f = fopen("/proc/self/numa_maps", "r");
 615         if (f == NULL) {
 616                 RTE_LOG(NOTICE, EAL, "NUMA support not available"
 617                         " consider that all memory is in socket_id 0\n");
 618                 return 0;
 619         }
 620
 621         snprintf(hugedir_str, sizeof(hugedir_str),
 622                         "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
 623
 624         /* parse numa map */
 625         while (fgets(buf, sizeof(buf), f) != NULL) {
 626
 627                 /* ignore non huge page */
 628                 if (strstr(buf, " huge ") == NULL &&
 629                                 strstr(buf, hugedir_str) == NULL)
 630                         continue;
 631
 632                 /* get zone addr */
 633                 virt_addr = strtoull(buf, &end, 16);
 634                 if (virt_addr == 0 || end == buf) {
 635                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 636                         goto error;
 637                 }
 638
 639                 /* get node id (socket id) */
 640                 nodestr = strstr(buf, " N");
 641                 if (nodestr == NULL) {
 642                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 643                         goto error;
 644                 }
 645                 nodestr += 2;
 646                 end = strstr(nodestr, "=");
 647                 if (end == NULL) {
 648                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 649                         goto error;
 650                 }
 651                 end[0] = '\0';
 652                 end = NULL;
 653
 654                 socket_id = strtoul(nodestr, &end, 0);
 655                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 656                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 657                         goto error;
 658                 }
 659
 660                 /* if we find this page in our mappings, set socket_id */
 661                 for (i = 0; i < hpi->num_pages[0]; i++) {
 662                         void *va = (void *)(unsigned long)virt_addr;
 663                         if (hugepg_tbl[i].orig_va == va) {
 664                                 hugepg_tbl[i].socket_id = socket_id;
 665                                 hp_count++;
 666 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 667                                 RTE_LOG(DEBUG, EAL,
 668                                         "Hugepage %s is on socket %d\n",
 669                                         hugepg_tbl[i].filepath, socket_id);
 670 #endif
 671                         }
 672                 }
 673         }
 674
 675         if (hp_count < hpi->num_pages[0])
 676                 goto error;
 677
 678         fclose(f);
 679         return 0;
 680
 681 error:
 682         fclose(f);
 683         return -1;
 684 }
 685
 686 static int
 687 cmp_physaddr(const void *a, const void *b)
 688 {
 689 #ifndef RTE_ARCH_PPC_64
 690         const struct hugepage_file *p1 = a;
 691         const struct hugepage_file *p2 = b;
 692 #else
 693         /* PowerPC needs memory sorted in reverse order from x86 */
 694         const struct hugepage_file *p1 = b;
 695         const struct hugepage_file *p2 = a;
 696 #endif
 697         if (p1->physaddr < p2->physaddr)
 698                 return -1;
 699         else if (p1->physaddr > p2->physaddr)
 700                 return 1;
 701         else
 702                 return 0;
 703 }
 704
 705 /*
 706  * Uses mmap to create a shared memory area for storage of data
 707  * Used in this file to store the hugepage file map on disk
 708  */
 709 static void *
 710 create_shared_memory(const char *filename, const size_t mem_size)
 711 {
 712         void *retval;
 713         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 714         if (fd < 0)
 715                 return NULL;
 716         if (ftruncate(fd, mem_size) < 0) {
 717                 close(fd);
 718                 return NULL;
 719         }
 720         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 721         close(fd);
 722         return retval;
 723 }
 724
 725 /*
 726  * this copies *active* hugepages from one hugepage table to another.
 727  * destination is typically the shared memory.
 728  */
 729 static int
 730 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
 731                 const struct hugepage_file * src, int src_size)
 732 {
 733         int src_pos, dst_pos = 0;
 734
 735         for (src_pos = 0; src_pos < src_size; src_pos++) {
 736                 if (src[src_pos].final_va != NULL) {
 737                         /* error on overflow attempt */
 738                         if (dst_pos == dest_size)
 739                                 return -1;
 740                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
 741                         dst_pos++;
 742                 }
 743         }
 744         return 0;
 745 }
 746
 747 static int
 748 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
 749                 unsigned num_hp_info)
 750 {
 751         unsigned socket, size;
 752         int page, nrpages = 0;
 753
 754         /* get total number of hugepages */
 755         for (size = 0; size < num_hp_info; size++)
 756                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 757                         nrpages +=
 758                         internal_config.hugepage_info[size].num_pages[socket];
 759
 760         for (page = 0; page < nrpages; page++) {
 761                 struct hugepage_file *hp = &hugepg_tbl[page];
 762
 763                 if (hp->final_va != NULL && unlink(hp->filepath)) {
 764                         RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
 765                                 __func__, hp->filepath, strerror(errno));
 766                 }
 767         }
 768         return 0;
 769 }
 770
 771 /*
 772  * unmaps hugepages that are not going to be used. since we originally allocate
 773  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 774  */
 775 static int
 776 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
 777                 struct hugepage_info *hpi,
 778                 unsigned num_hp_info)
 779 {
 780         unsigned socket, size;
 781         int page, nrpages = 0;
 782
 783         /* get total number of hugepages */
 784         for (size = 0; size < num_hp_info; size++)
 785                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 786                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 787
 788         for (size = 0; size < num_hp_info; size++) {
 789                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 790                         unsigned pages_found = 0;
 791
 792                         /* traverse until we have unmapped all the unused pages */
 793                         for (page = 0; page < nrpages; page++) {
 794                                 struct hugepage_file *hp = &hugepg_tbl[page];
 795
 796                                 /* find a page that matches the criteria */
 797                                 if ((hp->size == hpi[size].hugepage_sz) &&
 798                                                 (hp->socket_id == (int) socket)) {
 799
 800                                         /* if we skipped enough pages, unmap the rest */
 801                                         if (pages_found == hpi[size].num_pages[socket]) {
 802                                                 uint64_t unmap_len;
 803
 804                                                 unmap_len = hp->size;
 805
 806                                                 /* get start addr and len of the remaining segment */
 807                                                 munmap(hp->final_va, (size_t) unmap_len);
 808
 809                                                 hp->final_va = NULL;
 810                                                 if (unlink(hp->filepath) == -1) {
 811                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 812                                                                         __func__, hp->filepath, strerror(errno));
 813                                                         return -1;
 814                                                 }
 815                                         } else {
 816                                                 /* lock the page and skip */
 817                                                 pages_found++;
 818                                         }
 819
 820                                 } /* match page */
 821                         } /* foreach page */
 822                 } /* foreach socket */
 823         } /* foreach pagesize */
 824
 825         return 0;
 826 }
 827
 828 static inline uint64_t
 829 get_socket_mem_size(int socket)
 830 {
 831         uint64_t size = 0;
 832         unsigned i;
 833
 834         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 835                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 836                 if (hpi->hugedir != NULL)
 837                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 838         }
 839
 840         return size;
 841 }
 842
 843 /*
 844  * This function is a NUMA-aware equivalent of calc_num_pages.
 845  * It takes in the list of hugepage sizes and the
 846  * number of pages thereof, and calculates the best number of
 847  * pages of each size to fulfill the request for <memory> ram
 848  */
 849 static int
 850 calc_num_pages_per_socket(uint64_t * memory,
 851                 struct hugepage_info *hp_info,
 852                 struct hugepage_info *hp_used,
 853                 unsigned num_hp_info)
 854 {
 855         unsigned socket, j, i = 0;
 856         unsigned requested, available;
 857         int total_num_pages = 0;
 858         uint64_t remaining_mem, cur_mem;
 859         uint64_t total_mem = internal_config.memory;
 860
 861         if (num_hp_info == 0)
 862                 return -1;
 863
 864         /* if specific memory amounts per socket weren't requested */
 865         if (internal_config.force_sockets == 0) {
 866                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
 867                 size_t default_size, total_size;
 868                 unsigned lcore_id;
 869
 870                 /* Compute number of cores per socket */
 871                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
 872                 RTE_LCORE_FOREACH(lcore_id) {
 873                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
 874                 }
 875
 876                 /*
 877                  * Automatically spread requested memory amongst detected sockets according
 878                  * to number of cores from cpu mask present on each socket
 879                  */
 880                 total_size = internal_config.memory;
 881                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 882
 883                         /* Set memory amount per socket */
 884                         default_size = (internal_config.memory * cpu_per_socket[socket])
 885                                         / rte_lcore_count();
 886
 887                         /* Limit to maximum available memory on socket */
 888                         default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
 889
 890                         /* Update sizes */
 891                         memory[socket] = default_size;
 892                         total_size -= default_size;
 893                 }
 894
 895                 /*
 896                  * If some memory is remaining, try to allocate it by getting all
 897                  * available memory from sockets, one after the other
 898                  */
 899                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 900                         /* take whatever is available */
 901                         default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
 902                                                total_size);
 903
 904                         /* Update sizes */
 905                         memory[socket] += default_size;
 906                         total_size -= default_size;
 907                 }
 908         }
 909
 910         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 911                 /* skips if the memory on specific socket wasn't requested */
 912                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 913                         hp_used[i].hugedir = hp_info[i].hugedir;
 914                         hp_used[i].num_pages[socket] = RTE_MIN(
 915                                         memory[socket] / hp_info[i].hugepage_sz,
 916                                         hp_info[i].num_pages[socket]);
 917
 918                         cur_mem = hp_used[i].num_pages[socket] *
 919                                         hp_used[i].hugepage_sz;
 920
 921                         memory[socket] -= cur_mem;
 922                         total_mem -= cur_mem;
 923
 924                         total_num_pages += hp_used[i].num_pages[socket];
 925
 926                         /* check if we have met all memory requests */
 927                         if (memory[socket] == 0)
 928                                 break;
 929
 930                         /* check if we have any more pages left at this size, if so
 931                          * move on to next size */
 932                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
 933                                 continue;
 934                         /* At this point we know that there are more pages available that are
 935                          * bigger than the memory we want, so lets see if we can get enough
 936                          * from other page sizes.
 937                          */
 938                         remaining_mem = 0;
 939                         for (j = i+1; j < num_hp_info; j++)
 940                                 remaining_mem += hp_info[j].hugepage_sz *
 941                                 hp_info[j].num_pages[socket];
 942
 943                         /* is there enough other memory, if not allocate another page and quit */
 944                         if (remaining_mem < memory[socket]){
 945                                 cur_mem = RTE_MIN(memory[socket],
 946                                                 hp_info[i].hugepage_sz);
 947                                 memory[socket] -= cur_mem;
 948                                 total_mem -= cur_mem;
 949                                 hp_used[i].num_pages[socket]++;
 950                                 total_num_pages++;
 951                                 break; /* we are done with this socket*/
 952                         }
 953                 }
 954                 /* if we didn't satisfy all memory requirements per socket */
 955                 if (memory[socket] > 0) {
 956                         /* to prevent icc errors */
 957                         requested = (unsigned) (internal_config.socket_mem[socket] /
 958                                         0x100000);
 959                         available = requested -
 960                                         ((unsigned) (memory[socket] / 0x100000));
 961                         RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
 962                                         "Requested: %uMB, available: %uMB\n", socket,
 963                                         requested, available);
 964                         return -1;
 965                 }
 966         }
 967
 968         /* if we didn't satisfy total memory requirements */
 969         if (total_mem > 0) {
 970                 requested = (unsigned) (internal_config.memory / 0x100000);
 971                 available = requested - (unsigned) (total_mem / 0x100000);
 972                 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
 973                                 " available: %uMB\n", requested, available);
 974                 return -1;
 975         }
 976         return total_num_pages;
 977 }
 978
 979 static inline size_t
 980 eal_get_hugepage_mem_size(void)
 981 {
 982         uint64_t size = 0;
 983         unsigned i, j;
 984
 985         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 986                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 987                 if (hpi->hugedir != NULL) {
 988                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 989                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 990                         }
 991                 }
 992         }
 993
 994         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 995 }
 996
 997 static struct sigaction huge_action_old;
 998 static int huge_need_recover;
 999
1000 static void
1001 huge_register_sigbus(void)
1002 {
1003         sigset_t mask;
1004         struct sigaction action;
1005
1006         sigemptyset(&mask);
1007         sigaddset(&mask, SIGBUS);
1008         action.sa_flags = 0;
1009         action.sa_mask = mask;
1010         action.sa_handler = huge_sigbus_handler;
1011
1012         huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
1013 }
1014
1015 static void
1016 huge_recover_sigbus(void)
1017 {
1018         if (huge_need_recover) {
1019                 sigaction(SIGBUS, &huge_action_old, NULL);
1020                 huge_need_recover = 0;
1021         }
1022 }
1023
1024 /*
1025  * Prepare physical memory mapping: fill configuration structure with
1026  * these infos, return 0 on success.
1027  *  1. map N huge pages in separate files in hugetlbfs
1028  *  2. find associated physical addr
1029  *  3. find associated NUMA socket ID
1030  *  4. sort all huge pages by physical address
1031  *  5. remap these N huge pages in the correct order
1032  *  6. unmap the first mapping
1033  *  7. fill memsegs in configuration with contiguous zones
1034  */
1035 int
1036 rte_eal_hugepage_init(void)
1037 {
1038         struct rte_mem_config *mcfg;
1039         struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1040         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1041
1042         uint64_t memory[RTE_MAX_NUMA_NODES];
1043
1044         unsigned hp_offset;
1045         int i, j, new_memseg;
1046         int nr_hugefiles, nr_hugepages = 0;
1047         void *addr;
1048
1049         test_phys_addrs_available();
1050
1051         memset(used_hp, 0, sizeof(used_hp));
1052
1053         /* get pointer to global configuration */
1054         mcfg = rte_eal_get_configuration()->mem_config;
1055
1056         /* hugetlbfs can be disabled */
1057         if (internal_config.no_hugetlbfs) {
1058                 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1059                                 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1060                 if (addr == MAP_FAILED) {
1061                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1062                                         strerror(errno));
1063                         return -1;
1064                 }
1065                 mcfg->memseg[0].phys_addr = RTE_BAD_PHYS_ADDR;
1066                 mcfg->memseg[0].addr = addr;
1067                 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1068                 mcfg->memseg[0].len = internal_config.memory;
1069                 mcfg->memseg[0].socket_id = 0;
1070                 return 0;
1071         }
1072
1073 /* check if app runs on Xen Dom0 */
1074         if (internal_config.xen_dom0_support) {
1075 #ifdef RTE_LIBRTE_XEN_DOM0
1076                 /* use dom0_mm kernel driver to init memory */
1077                 if (rte_xen_dom0_memory_init() < 0)
1078                         return -1;
1079                 else
1080                         return 0;
1081 #endif
1082         }
1083
1084         /* calculate total number of hugepages available. at this point we haven't
1085          * yet started sorting them so they all are on socket 0 */
1086         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1087                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1088                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1089
1090                 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1091         }
1092
1093         /*
1094          * allocate a memory area for hugepage table.
1095          * this isn't shared memory yet. due to the fact that we need some
1096          * processing done on these pages, shared memory will be created
1097          * at a later stage.
1098          */
1099         tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1100         if (tmp_hp == NULL)
1101                 goto fail;
1102
1103         memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1104
1105         hp_offset = 0; /* where we start the current page size entries */
1106
1107         huge_register_sigbus();
1108
1109         /* make a copy of socket_mem, needed for balanced allocation. */
1110         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1111                 memory[i] = internal_config.socket_mem[i];
1112
1113
1114         /* map all hugepages and sort them */
1115         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1116                 unsigned pages_old, pages_new;
1117                 struct hugepage_info *hpi;
1118
1119                 /*
1120                  * we don't yet mark hugepages as used at this stage, so
1121                  * we just map all hugepages available to the system
1122                  * all hugepages are still located on socket 0
1123                  */
1124                 hpi = &internal_config.hugepage_info[i];
1125
1126                 if (hpi->num_pages[0] == 0)
1127                         continue;
1128
1129                 /* map all hugepages available */
1130                 pages_old = hpi->num_pages[0];
1131                 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi,
1132                                               memory, 1);
1133                 if (pages_new < pages_old) {
1134                         RTE_LOG(DEBUG, EAL,
1135                                 "%d not %d hugepages of size %u MB allocated\n",
1136                                 pages_new, pages_old,
1137                                 (unsigned)(hpi->hugepage_sz / 0x100000));
1138
1139                         int pages = pages_old - pages_new;
1140
1141                         nr_hugepages -= pages;
1142                         hpi->num_pages[0] = pages_new;
1143                         if (pages_new == 0)
1144                                 continue;
1145                 }
1146
1147                 if (phys_addrs_available) {
1148                         /* find physical addresses for each hugepage */
1149                         if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1150                                 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1151                                         "for %u MB pages\n",
1152                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1153                                 goto fail;
1154                         }
1155                 } else {
1156                         /* set physical addresses for each hugepage */
1157                         if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1158                                 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1159                                         "for %u MB pages\n",
1160                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1161                                 goto fail;
1162                         }
1163                 }
1164
1165                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1166                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1167                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1168                         goto fail;
1169                 }
1170
1171                 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1172                       sizeof(struct hugepage_file), cmp_physaddr);
1173
1174                 /* remap all hugepages */
1175                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, NULL, 0) !=
1176                     hpi->num_pages[0]) {
1177                         RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1178                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1179                         goto fail;
1180                 }
1181
1182                 /* unmap original mappings */
1183                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1184                         goto fail;
1185
1186                 /* we have processed a num of hugepages of this size, so inc offset */
1187                 hp_offset += hpi->num_pages[0];
1188         }
1189
1190         huge_recover_sigbus();
1191
1192         if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1193                 internal_config.memory = eal_get_hugepage_mem_size();
1194
1195         nr_hugefiles = nr_hugepages;
1196
1197
1198         /* clean out the numbers of pages */
1199         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1200                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1201                         internal_config.hugepage_info[i].num_pages[j] = 0;
1202
1203         /* get hugepages for each socket */
1204         for (i = 0; i < nr_hugefiles; i++) {
1205                 int socket = tmp_hp[i].socket_id;
1206
1207                 /* find a hugepage info with right size and increment num_pages */
1208                 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1209                                 (int)internal_config.num_hugepage_sizes);
1210                 for (j = 0; j < nb_hpsizes; j++) {
1211                         if (tmp_hp[i].size ==
1212                                         internal_config.hugepage_info[j].hugepage_sz) {
1213                                 internal_config.hugepage_info[j].num_pages[socket]++;
1214                         }
1215                 }
1216         }
1217
1218         /* make a copy of socket_mem, needed for number of pages calculation */
1219         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1220                 memory[i] = internal_config.socket_mem[i];
1221
1222         /* calculate final number of pages */
1223         nr_hugepages = calc_num_pages_per_socket(memory,
1224                         internal_config.hugepage_info, used_hp,
1225                         internal_config.num_hugepage_sizes);
1226
1227         /* error if not enough memory available */
1228         if (nr_hugepages < 0)
1229                 goto fail;
1230
1231         /* reporting in! */
1232         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1233                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1234                         if (used_hp[i].num_pages[j] > 0) {
1235                                 RTE_LOG(DEBUG, EAL,
1236                                         "Requesting %u pages of size %uMB"
1237                                         " from socket %i\n",
1238                                         used_hp[i].num_pages[j],
1239                                         (unsigned)
1240                                         (used_hp[i].hugepage_sz / 0x100000),
1241                                         j);
1242                         }
1243                 }
1244         }
1245
1246         /* create shared memory */
1247         hugepage = create_shared_memory(eal_hugepage_info_path(),
1248                         nr_hugefiles * sizeof(struct hugepage_file));
1249
1250         if (hugepage == NULL) {
1251                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1252                 goto fail;
1253         }
1254         memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1255
1256         /*
1257          * unmap pages that we won't need (looks at used_hp).
1258          * also, sets final_va to NULL on pages that were unmapped.
1259          */
1260         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1261                         internal_config.num_hugepage_sizes) < 0) {
1262                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1263                 goto fail;
1264         }
1265
1266         /*
1267          * copy stuff from malloc'd hugepage* to the actual shared memory.
1268          * this procedure only copies those hugepages that have final_va
1269          * not NULL. has overflow protection.
1270          */
1271         if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1272                         tmp_hp, nr_hugefiles) < 0) {
1273                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1274                 goto fail;
1275         }
1276
1277         /* free the hugepage backing files */
1278         if (internal_config.hugepage_unlink &&
1279                 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1280                 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1281                 goto fail;
1282         }
1283
1284         /* free the temporary hugepage table */
1285         free(tmp_hp);
1286         tmp_hp = NULL;
1287
1288         /* first memseg index shall be 0 after incrementing it below */
1289         j = -1;
1290         for (i = 0; i < nr_hugefiles; i++) {
1291                 new_memseg = 0;
1292
1293                 /* if this is a new section, create a new memseg */
1294                 if (i == 0)
1295                         new_memseg = 1;
1296                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1297                         new_memseg = 1;
1298                 else if (hugepage[i].size != hugepage[i-1].size)
1299                         new_memseg = 1;
1300
1301 #ifdef RTE_ARCH_PPC_64
1302                 /* On PPC64 architecture, the mmap always start from higher
1303                  * virtual address to lower address. Here, both the physical
1304                  * address and virtual address are in descending order */
1305                 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1306                     hugepage[i].size)
1307                         new_memseg = 1;
1308                 else if (((unsigned long)hugepage[i-1].final_va -
1309                     (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1310                         new_memseg = 1;
1311 #else
1312                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1313                     hugepage[i].size)
1314                         new_memseg = 1;
1315                 else if (((unsigned long)hugepage[i].final_va -
1316                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1317                         new_memseg = 1;
1318 #endif
1319
1320                 if (new_memseg) {
1321                         j += 1;
1322                         if (j == RTE_MAX_MEMSEG)
1323                                 break;
1324
1325                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1326                         mcfg->memseg[j].addr = hugepage[i].final_va;
1327                         mcfg->memseg[j].len = hugepage[i].size;
1328                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1329                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1330                 }
1331                 /* continuation of previous memseg */
1332                 else {
1333 #ifdef RTE_ARCH_PPC_64
1334                 /* Use the phy and virt address of the last page as segment
1335                  * address for IBM Power architecture */
1336                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1337                         mcfg->memseg[j].addr = hugepage[i].final_va;
1338 #endif
1339                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1340                 }
1341                 hugepage[i].memseg_id = j;
1342         }
1343
1344         if (i < nr_hugefiles) {
1345                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1346                         "from %d requested\n"
1347                         "Current %s=%d is not enough\n"
1348                         "Please either increase it or request less amount "
1349                         "of memory.\n",
1350                         i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1351                         RTE_MAX_MEMSEG);
1352                 goto fail;
1353         }
1354
1355         munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1356
1357         return 0;
1358
1359 fail:
1360         huge_recover_sigbus();
1361         free(tmp_hp);
1362         if (hugepage != NULL)
1363                 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1364
1365         return -1;
1366 }
1367
1368 /*
1369  * uses fstat to report the size of a file on disk
1370  */
1371 static off_t
1372 getFileSize(int fd)
1373 {
1374         struct stat st;
1375         if (fstat(fd, &st) < 0)
1376                 return 0;
1377         return st.st_size;
1378 }
1379
1380 /*
1381  * This creates the memory mappings in the secondary process to match that of
1382  * the server process. It goes through each memory segment in the DPDK runtime
1383  * configuration and finds the hugepages which form that segment, mapping them
1384  * in order to form a contiguous block in the virtual memory space
1385  */
1386 int
1387 rte_eal_hugepage_attach(void)
1388 {
1389         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1390         struct hugepage_file *hp = NULL;
1391         unsigned num_hp = 0;
1392         unsigned i, s = 0; /* s used to track the segment number */
1393         unsigned max_seg = RTE_MAX_MEMSEG;
1394         off_t size = 0;
1395         int fd, fd_zero = -1, fd_hugepage = -1;
1396
1397         if (aslr_enabled() > 0) {
1398                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1399                                 "(ASLR) is enabled in the kernel.\n");
1400                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1401                                 "into secondary processes\n");
1402         }
1403
1404         test_phys_addrs_available();
1405
1406         if (internal_config.xen_dom0_support) {
1407 #ifdef RTE_LIBRTE_XEN_DOM0
1408                 if (rte_xen_dom0_memory_attach() < 0) {
1409                         RTE_LOG(ERR, EAL, "Failed to attach memory segments of primary "
1410                                         "process\n");
1411                         return -1;
1412                 }
1413                 return 0;
1414 #endif
1415         }
1416
1417         fd_zero = open("/dev/zero", O_RDONLY);
1418         if (fd_zero < 0) {
1419                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1420                 goto error;
1421         }
1422         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1423         if (fd_hugepage < 0) {
1424                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1425                 goto error;
1426         }
1427
1428         /* map all segments into memory to make sure we get the addrs */
1429         for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1430                 void *base_addr;
1431
1432                 /*
1433                  * the first memory segment with len==0 is the one that
1434                  * follows the last valid segment.
1435                  */
1436                 if (mcfg->memseg[s].len == 0)
1437                         break;
1438
1439                 /*
1440                  * fdzero is mmapped to get a contiguous block of virtual
1441                  * addresses of the appropriate memseg size.
1442                  * use mmap to get identical addresses as the primary process.
1443                  */
1444                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1445                                  PROT_READ,
1446 #ifdef RTE_ARCH_PPC_64
1447                                  MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB,
1448 #else
1449                                  MAP_PRIVATE,
1450 #endif
1451                                  fd_zero, 0);
1452                 if (base_addr == MAP_FAILED ||
1453                     base_addr != mcfg->memseg[s].addr) {
1454                         max_seg = s;
1455                         if (base_addr != MAP_FAILED) {
1456                                 /* errno is stale, don't use */
1457                                 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1458                                         "in /dev/zero at [%p], got [%p] - "
1459                                         "please use '--base-virtaddr' option\n",
1460                                         (unsigned long long)mcfg->memseg[s].len,
1461                                         mcfg->memseg[s].addr, base_addr);
1462                                 munmap(base_addr, mcfg->memseg[s].len);
1463                         } else {
1464                                 RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1465                                         "in /dev/zero at [%p]: '%s'\n",
1466                                         (unsigned long long)mcfg->memseg[s].len,
1467                                         mcfg->memseg[s].addr, strerror(errno));
1468                         }
1469                         if (aslr_enabled() > 0) {
1470                                 RTE_LOG(ERR, EAL, "It is recommended to "
1471                                         "disable ASLR in the kernel "
1472                                         "and retry running both primary "
1473                                         "and secondary processes\n");
1474                         }
1475                         goto error;
1476                 }
1477         }
1478
1479         size = getFileSize(fd_hugepage);
1480         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1481         if (hp == MAP_FAILED) {
1482                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1483                 goto error;
1484         }
1485
1486         num_hp = size / sizeof(struct hugepage_file);
1487         RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1488
1489         s = 0;
1490         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1491                 void *addr, *base_addr;
1492                 uintptr_t offset = 0;
1493                 size_t mapping_size;
1494                 /*
1495                  * free previously mapped memory so we can map the
1496                  * hugepages into the space
1497                  */
1498                 base_addr = mcfg->memseg[s].addr;
1499                 munmap(base_addr, mcfg->memseg[s].len);
1500
1501                 /* find the hugepages for this segment and map them
1502                  * we don't need to worry about order, as the server sorted the
1503                  * entries before it did the second mmap of them */
1504                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1505                         if (hp[i].memseg_id == (int)s){
1506                                 fd = open(hp[i].filepath, O_RDWR);
1507                                 if (fd < 0) {
1508                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1509                                                 hp[i].filepath);
1510                                         goto error;
1511                                 }
1512                                 mapping_size = hp[i].size;
1513                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1514                                                 mapping_size, PROT_READ | PROT_WRITE,
1515                                                 MAP_SHARED, fd, 0);
1516                                 close(fd); /* close file both on success and on failure */
1517                                 if (addr == MAP_FAILED ||
1518                                                 addr != RTE_PTR_ADD(base_addr, offset)) {
1519                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1520                                                 hp[i].filepath);
1521                                         goto error;
1522                                 }
1523                                 offset+=mapping_size;
1524                         }
1525                 }
1526                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1527                                 (unsigned long long)mcfg->memseg[s].len);
1528                 s++;
1529         }
1530         /* unmap the hugepage config file, since we are done using it */
1531         munmap(hp, size);
1532         close(fd_zero);
1533         close(fd_hugepage);
1534         return 0;
1535
1536 error:
1537         for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1538                 munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1539         if (hp != NULL && hp != MAP_FAILED)
1540                 munmap(hp, size);
1541         if (fd_zero >= 0)
1542                 close(fd_zero);
1543         if (fd_hugepage >= 0)
1544                 close(fd_hugepage);
1545         return -1;
1546 }
1547
1548 bool
1549 rte_eal_using_phys_addrs(void)
1550 {
1551         return phys_addrs_available;
1552 }