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