lib/librte_eal/linuxapp/eal/eal_memory.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33 /*   BSD LICENSE
  34  *
  35  *   Copyright(c) 2013 6WIND.
  36  *
  37  *   Redistribution and use in source and binary forms, with or without
  38  *   modification, are permitted provided that the following conditions
  39  *   are met:
  40  *
  41  *     * Redistributions of source code must retain the above copyright
  42  *       notice, this list of conditions and the following disclaimer.
  43  *     * Redistributions in binary form must reproduce the above copyright
  44  *       notice, this list of conditions and the following disclaimer in
  45  *       the documentation and/or other materials provided with the
  46  *       distribution.
  47  *     * Neither the name of 6WIND S.A. nor the names of its
  48  *       contributors may be used to endorse or promote products derived
  49  *       from this software without specific prior written permission.
  50  *
  51  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  52  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  53  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  54  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  55  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  56  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  57  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  58  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  59  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  60  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  61  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  62  */
  63
  64 #define _FILE_OFFSET_BITS 64
  65 #include <errno.h>
  66 #include <stdarg.h>
  67 #include <stdlib.h>
  68 #include <stdio.h>
  69 #include <stdint.h>
  70 #include <inttypes.h>
  71 #include <string.h>
  72 #include <stdarg.h>
  73 #include <sys/mman.h>
  74 #include <sys/types.h>
  75 #include <sys/stat.h>
  76 #include <sys/queue.h>
  77 #include <sys/file.h>
  78 #include <unistd.h>
  79 #include <limits.h>
  80 #include <errno.h>
  81 #include <sys/ioctl.h>
  82 #include <sys/time.h>
  83
  84 #include <rte_log.h>
  85 #include <rte_memory.h>
  86 #include <rte_memzone.h>
  87 #include <rte_launch.h>
  88 #include <rte_eal.h>
  89 #include <rte_eal_memconfig.h>
  90 #include <rte_per_lcore.h>
  91 #include <rte_lcore.h>
  92 #include <rte_common.h>
  93 #include <rte_string_fns.h>
  94
  95 #include "eal_private.h"
  96 #include "eal_internal_cfg.h"
  97 #include "eal_filesystem.h"
  98 #include "eal_hugepages.h"
  99
 100 /**
 101  * @file
 102  * Huge page mapping under linux
 103  *
 104  * To reserve a big contiguous amount of memory, we use the hugepage
 105  * feature of linux. For that, we need to have hugetlbfs mounted. This
 106  * code will create many files in this directory (one per page) and
 107  * map them in virtual memory. For each page, we will retrieve its
 108  * physical address and remap it in order to have a virtual contiguous
 109  * zone as well as a physical contiguous zone.
 110  */
 111
 112 static uint64_t baseaddr_offset;
 113
 114 static unsigned proc_pagemap_readable;
 115
 116 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
 117
 118 /* Lock page in physical memory and prevent from swapping. */
 119 int
 120 rte_mem_lock_page(const void *virt)
 121 {
 122         unsigned long virtual = (unsigned long)virt;
 123         int page_size = getpagesize();
 124         unsigned long aligned = (virtual & ~ (page_size - 1));
 125         return mlock((void*)aligned, page_size);
 126 }
 127
 128 /*
 129  * Get physical address of any mapped virtual address in the current process.
 130  */
 131 phys_addr_t
 132 rte_mem_virt2phy(const void *virtaddr)
 133 {
 134         int fd;
 135         uint64_t page, physaddr;
 136         unsigned long virt_pfn;
 137         int page_size;
 138         off_t offset;
 139
 140         /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
 141         if (!proc_pagemap_readable)
 142                 return RTE_BAD_PHYS_ADDR;
 143
 144         /* standard page size */
 145         page_size = getpagesize();
 146
 147         fd = open("/proc/self/pagemap", O_RDONLY);
 148         if (fd < 0) {
 149                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 150                         __func__, strerror(errno));
 151                 return RTE_BAD_PHYS_ADDR;
 152         }
 153
 154         virt_pfn = (unsigned long)virtaddr / page_size;
 155         offset = sizeof(uint64_t) * virt_pfn;
 156         if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 157                 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 158                                 __func__, strerror(errno));
 159                 close(fd);
 160                 return RTE_BAD_PHYS_ADDR;
 161         }
 162         if (read(fd, &page, sizeof(uint64_t)) < 0) {
 163                 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 164                                 __func__, strerror(errno));
 165                 close(fd);
 166                 return RTE_BAD_PHYS_ADDR;
 167         }
 168
 169         /*
 170          * the pfn (page frame number) are bits 0-54 (see
 171          * pagemap.txt in linux Documentation)
 172          */
 173         physaddr = ((page & 0x7fffffffffffffULL) * page_size)
 174                 + ((unsigned long)virtaddr % page_size);
 175         close(fd);
 176         return physaddr;
 177 }
 178
 179 /*
 180  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 181  * it by browsing the /proc/self/pagemap special file.
 182  */
 183 static int
 184 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 185 {
 186         unsigned i;
 187         phys_addr_t addr;
 188
 189         for (i = 0; i < hpi->num_pages[0]; i++) {
 190                 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
 191                 if (addr == RTE_BAD_PHYS_ADDR)
 192                         return -1;
 193                 hugepg_tbl[i].physaddr = addr;
 194         }
 195         return 0;
 196 }
 197
 198 /*
 199  * Check whether address-space layout randomization is enabled in
 200  * the kernel. This is important for multi-process as it can prevent
 201  * two processes mapping data to the same virtual address
 202  * Returns:
 203  *    0 - address space randomization disabled
 204  *    1/2 - address space randomization enabled
 205  *    negative error code on error
 206  */
 207 static int
 208 aslr_enabled(void)
 209 {
 210         char c;
 211         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 212         if (fd < 0)
 213                 return -errno;
 214         retval = read(fd, &c, 1);
 215         close(fd);
 216         if (retval < 0)
 217                 return -errno;
 218         if (retval == 0)
 219                 return -EIO;
 220         switch (c) {
 221                 case '0' : return 0;
 222                 case '1' : return 1;
 223                 case '2' : return 2;
 224                 default: return -EINVAL;
 225         }
 226 }
 227
 228 /*
 229  * Try to mmap *size bytes in /dev/zero. If it is successful, return the
 230  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 231  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 232  * 0. In this case, return NULL. Note: this function returns an address
 233  * which is a multiple of hugepage size.
 234  */
 235 static void *
 236 get_virtual_area(size_t *size, size_t hugepage_sz)
 237 {
 238         void *addr;
 239         int fd;
 240         long aligned_addr;
 241
 242         if (internal_config.base_virtaddr != 0) {
 243                 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
 244                                 baseaddr_offset);
 245         }
 246         else addr = NULL;
 247
 248         RTE_LOG(INFO, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
 249
 250         fd = open("/dev/zero", O_RDONLY);
 251         if (fd < 0){
 252                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 253                 return NULL;
 254         }
 255         do {
 256                 addr = mmap(addr,
 257                                 (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
 258                 if (addr == MAP_FAILED)
 259                         *size -= hugepage_sz;
 260         } while (addr == MAP_FAILED && *size > 0);
 261
 262         if (addr == MAP_FAILED) {
 263                 close(fd);
 264                 RTE_LOG(INFO, EAL, "Cannot get a virtual area\n");
 265                 return NULL;
 266         }
 267
 268         munmap(addr, (*size) + hugepage_sz);
 269         close(fd);
 270
 271         /* align addr to a huge page size boundary */
 272         aligned_addr = (long)addr;
 273         aligned_addr += (hugepage_sz - 1);
 274         aligned_addr &= (~(hugepage_sz - 1));
 275         addr = (void *)(aligned_addr);
 276
 277         RTE_LOG(INFO, EAL, "Virtual area found at %p (size = 0x%zx)\n",
 278                 addr, *size);
 279
 280         /* increment offset */
 281         baseaddr_offset += *size;
 282
 283         return addr;
 284 }
 285
 286 /*
 287  * Mmap all hugepages of hugepage table: it first open a file in
 288  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 289  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 290  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 291  * map continguous physical blocks in contiguous virtual blocks.
 292  */
 293 static int
 294 map_all_hugepages(struct hugepage_file *hugepg_tbl,
 295                 struct hugepage_info *hpi, int orig)
 296 {
 297         int fd;
 298         unsigned i;
 299         void *virtaddr;
 300         void *vma_addr = NULL;
 301         size_t vma_len = 0;
 302
 303 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 304         RTE_SET_USED(vma_len);
 305 #endif
 306
 307         for (i = 0; i < hpi->num_pages[0]; i++) {
 308                 uint64_t hugepage_sz = hpi->hugepage_sz;
 309
 310                 if (orig) {
 311                         hugepg_tbl[i].file_id = i;
 312                         hugepg_tbl[i].size = hugepage_sz;
 313 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 314                         eal_get_hugefile_temp_path(hugepg_tbl[i].filepath,
 315                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 316                                         hugepg_tbl[i].file_id);
 317 #else
 318                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 319                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 320                                         hugepg_tbl[i].file_id);
 321 #endif
 322                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 323                 }
 324 #ifndef RTE_ARCH_64
 325                 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
 326                  * original map address as final map address.
 327                  */
 328                 else if ((hugepage_sz == RTE_PGSIZE_1G)
 329                         || (hugepage_sz == RTE_PGSIZE_16G)) {
 330                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 331                         hugepg_tbl[i].orig_va = NULL;
 332                         continue;
 333                 }
 334 #endif
 335
 336 #ifndef RTE_EAL_SINGLE_FILE_SEGMENTS
 337                 else if (vma_len == 0) {
 338                         unsigned j, num_pages;
 339
 340                         /* reserve a virtual area for next contiguous
 341                          * physical block: count the number of
 342                          * contiguous physical pages. */
 343                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 344 #ifdef RTE_ARCH_PPC_64
 345                                 /* The physical addresses are sorted in
 346                                  * descending order on PPC64 */
 347                                 if (hugepg_tbl[j].physaddr !=
 348                                     hugepg_tbl[j-1].physaddr - hugepage_sz)
 349                                         break;
 350 #else
 351                                 if (hugepg_tbl[j].physaddr !=
 352                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 353                                         break;
 354 #endif
 355                         }
 356                         num_pages = j - i;
 357                         vma_len = num_pages * hugepage_sz;
 358
 359                         /* get the biggest virtual memory area up to
 360                          * vma_len. If it fails, vma_addr is NULL, so
 361                          * let the kernel provide the address. */
 362                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 363                         if (vma_addr == NULL)
 364                                 vma_len = hugepage_sz;
 365                 }
 366 #endif
 367
 368                 /* try to create hugepage file */
 369                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
 370                 if (fd < 0) {
 371                         RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
 372                                         strerror(errno));
 373                         return -1;
 374                 }
 375
 376                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 377                                 MAP_SHARED, fd, 0);
 378                 if (virtaddr == MAP_FAILED) {
 379                         RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__,
 380                                         strerror(errno));
 381                         close(fd);
 382                         return -1;
 383                 }
 384
 385                 if (orig) {
 386                         hugepg_tbl[i].orig_va = virtaddr;
 387                         memset(virtaddr, 0, hugepage_sz);
 388                 }
 389                 else {
 390                         hugepg_tbl[i].final_va = virtaddr;
 391                 }
 392
 393                 /* set shared flock on the file. */
 394                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 395                         RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
 396                                 __func__, strerror(errno));
 397                         close(fd);
 398                         return -1;
 399                 }
 400
 401                 close(fd);
 402
 403                 vma_addr = (char *)vma_addr + hugepage_sz;
 404                 vma_len -= hugepage_sz;
 405         }
 406         return 0;
 407 }
 408
 409 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 410
 411 /*
 412  * Remaps all hugepages into single file segments
 413  */
 414 static int
 415 remap_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 416 {
 417         int fd;
 418         unsigned i = 0, j, num_pages, page_idx = 0;
 419         void *vma_addr = NULL, *old_addr = NULL, *page_addr = NULL;
 420         size_t vma_len = 0;
 421         size_t hugepage_sz = hpi->hugepage_sz;
 422         size_t total_size, offset;
 423         char filepath[MAX_HUGEPAGE_PATH];
 424         phys_addr_t physaddr;
 425         int socket;
 426
 427         while (i < hpi->num_pages[0]) {
 428
 429 #ifndef RTE_ARCH_64
 430                 /* for 32-bit systems, don't remap 1G pages and 16G pages,
 431                  * just reuse original map address as final map address.
 432                  */
 433                 if ((hugepage_sz == RTE_PGSIZE_1G)
 434                         || (hugepage_sz == RTE_PGSIZE_16G)) {
 435                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 436                         hugepg_tbl[i].orig_va = NULL;
 437                         i++;
 438                         continue;
 439                 }
 440 #endif
 441
 442                 /* reserve a virtual area for next contiguous
 443                  * physical block: count the number of
 444                  * contiguous physical pages. */
 445                 for (j = i+1; j < hpi->num_pages[0] ; j++) {
 446 #ifdef RTE_ARCH_PPC_64
 447                         /* The physical addresses are sorted in descending
 448                          * order on PPC64 */
 449                         if (hugepg_tbl[j].physaddr !=
 450                                 hugepg_tbl[j-1].physaddr - hugepage_sz)
 451                                 break;
 452 #else
 453                         if (hugepg_tbl[j].physaddr !=
 454                                 hugepg_tbl[j-1].physaddr + hugepage_sz)
 455                                 break;
 456 #endif
 457                 }
 458                 num_pages = j - i;
 459                 vma_len = num_pages * hugepage_sz;
 460
 461                 socket = hugepg_tbl[i].socket_id;
 462
 463                 /* get the biggest virtual memory area up to
 464                  * vma_len. If it fails, vma_addr is NULL, so
 465                  * let the kernel provide the address. */
 466                 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 467
 468                 /* If we can't find a big enough virtual area, work out how many pages
 469                  * we are going to get */
 470                 if (vma_addr == NULL)
 471                         j = i + 1;
 472                 else if (vma_len != num_pages * hugepage_sz) {
 473                         num_pages = vma_len / hugepage_sz;
 474                         j = i + num_pages;
 475
 476                 }
 477
 478                 hugepg_tbl[page_idx].file_id = page_idx;
 479                 eal_get_hugefile_path(filepath,
 480                                 sizeof(filepath),
 481                                 hpi->hugedir,
 482                                 hugepg_tbl[page_idx].file_id);
 483
 484                 /* try to create hugepage file */
 485                 fd = open(filepath, O_CREAT | O_RDWR, 0755);
 486                 if (fd < 0) {
 487                         RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__, strerror(errno));
 488                         return -1;
 489                 }
 490
 491                 total_size = 0;
 492                 for (;i < j; i++) {
 493
 494                         /* unmap current segment */
 495                         if (total_size > 0)
 496                                 munmap(vma_addr, total_size);
 497
 498                         /* unmap original page */
 499                         munmap(hugepg_tbl[i].orig_va, hugepage_sz);
 500                         unlink(hugepg_tbl[i].filepath);
 501
 502                         total_size += hugepage_sz;
 503
 504                         old_addr = vma_addr;
 505
 506                         /* map new, bigger segment */
 507                         vma_addr = mmap(vma_addr, total_size,
 508                                         PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 509
 510                         if (vma_addr == MAP_FAILED || vma_addr != old_addr) {
 511                                 RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__, strerror(errno));
 512                                 close(fd);
 513                                 return -1;
 514                         }
 515
 516                         /* touch the page. this is needed because kernel postpones mapping
 517                          * creation until the first page fault. with this, we pin down
 518                          * the page and it is marked as used and gets into process' pagemap.
 519                          */
 520                         for (offset = 0; offset < total_size; offset += hugepage_sz)
 521                                 *((volatile uint8_t*) RTE_PTR_ADD(vma_addr, offset));
 522                 }
 523
 524                 /* set shared flock on the file. */
 525                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 526                         RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
 527                                 __func__, strerror(errno));
 528                         close(fd);
 529                         return -1;
 530                 }
 531
 532                 snprintf(hugepg_tbl[page_idx].filepath, MAX_HUGEPAGE_PATH, "%s",
 533                                 filepath);
 534
 535                 physaddr = rte_mem_virt2phy(vma_addr);
 536
 537                 if (physaddr == RTE_BAD_PHYS_ADDR)
 538                         return -1;
 539
 540                 hugepg_tbl[page_idx].final_va = vma_addr;
 541
 542                 hugepg_tbl[page_idx].physaddr = physaddr;
 543
 544                 hugepg_tbl[page_idx].repeated = num_pages;
 545
 546                 hugepg_tbl[page_idx].socket_id = socket;
 547
 548                 close(fd);
 549
 550                 /* verify the memory segment - that is, check that every VA corresponds
 551                  * to the physical address we expect to see
 552                  */
 553                 for (offset = 0; offset < vma_len; offset += hugepage_sz) {
 554                         uint64_t expected_physaddr;
 555
 556                         expected_physaddr = hugepg_tbl[page_idx].physaddr + offset;
 557                         page_addr = RTE_PTR_ADD(vma_addr, offset);
 558                         physaddr = rte_mem_virt2phy(page_addr);
 559
 560                         if (physaddr != expected_physaddr) {
 561                                 RTE_LOG(ERR, EAL, "Segment sanity check failed: wrong physaddr "
 562                                                 "at %p (offset 0x%" PRIx64 ": 0x%" PRIx64
 563                                                 " (expected 0x%" PRIx64 ")\n",
 564                                                 page_addr, offset, physaddr, expected_physaddr);
 565                                 return -1;
 566                         }
 567                 }
 568
 569                 /* zero out the whole segment */
 570                 memset(hugepg_tbl[page_idx].final_va, 0, total_size);
 571
 572                 page_idx++;
 573         }
 574
 575         /* zero out the rest */
 576         memset(&hugepg_tbl[page_idx], 0, (hpi->num_pages[0] - page_idx) * sizeof(struct hugepage_file));
 577         return page_idx;
 578 }
 579 #else/* RTE_EAL_SINGLE_FILE_SEGMENTS=n */
 580
 581 /* Unmap all hugepages from original mapping */
 582 static int
 583 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 584 {
 585         unsigned i;
 586         for (i = 0; i < hpi->num_pages[0]; i++) {
 587                 if (hugepg_tbl[i].orig_va) {
 588                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 589                         hugepg_tbl[i].orig_va = NULL;
 590                 }
 591         }
 592         return 0;
 593 }
 594 #endif /* RTE_EAL_SINGLE_FILE_SEGMENTS */
 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(INFO, EAL, "cannot open /proc/self/numa_maps,"
 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                         }
 664                 }
 665         }
 666
 667         if (hp_count < hpi->num_pages[0])
 668                 goto error;
 669
 670         fclose(f);
 671         return 0;
 672
 673 error:
 674         fclose(f);
 675         return -1;
 676 }
 677
 678 /*
 679  * Sort the hugepg_tbl by physical address (lower addresses first on x86,
 680  * higher address first on powerpc). We use a slow algorithm, but we won't
 681  * have millions of pages, and this is only done at init time.
 682  */
 683 static int
 684 sort_by_physaddr(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 685 {
 686         unsigned i, j;
 687         int compare_idx;
 688         uint64_t compare_addr;
 689         struct hugepage_file tmp;
 690
 691         for (i = 0; i < hpi->num_pages[0]; i++) {
 692                 compare_addr = 0;
 693                 compare_idx = -1;
 694
 695                 /*
 696                  * browse all entries starting at 'i', and find the
 697                  * entry with the smallest addr
 698                  */
 699                 for (j=i; j< hpi->num_pages[0]; j++) {
 700
 701                         if (compare_addr == 0 ||
 702 #ifdef RTE_ARCH_PPC_64
 703                                 hugepg_tbl[j].physaddr > compare_addr) {
 704 #else
 705                                 hugepg_tbl[j].physaddr < compare_addr) {
 706 #endif
 707                                 compare_addr = hugepg_tbl[j].physaddr;
 708                                 compare_idx = j;
 709                         }
 710                 }
 711
 712                 /* should not happen */
 713                 if (compare_idx == -1) {
 714                         RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
 715                         return -1;
 716                 }
 717
 718                 /* swap the 2 entries in the table */
 719                 memcpy(&tmp, &hugepg_tbl[compare_idx],
 720                         sizeof(struct hugepage_file));
 721                 memcpy(&hugepg_tbl[compare_idx], &hugepg_tbl[i],
 722                         sizeof(struct hugepage_file));
 723                 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage_file));
 724         }
 725         return 0;
 726 }
 727
 728 /*
 729  * Uses mmap to create a shared memory area for storage of data
 730  * Used in this file to store the hugepage file map on disk
 731  */
 732 static void *
 733 create_shared_memory(const char *filename, const size_t mem_size)
 734 {
 735         void *retval;
 736         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 737         if (fd < 0)
 738                 return NULL;
 739         if (ftruncate(fd, mem_size) < 0) {
 740                 close(fd);
 741                 return NULL;
 742         }
 743         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 744         close(fd);
 745         return retval;
 746 }
 747
 748 /*
 749  * this copies *active* hugepages from one hugepage table to another.
 750  * destination is typically the shared memory.
 751  */
 752 static int
 753 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
 754                 const struct hugepage_file * src, int src_size)
 755 {
 756         int src_pos, dst_pos = 0;
 757
 758         for (src_pos = 0; src_pos < src_size; src_pos++) {
 759                 if (src[src_pos].final_va != NULL) {
 760                         /* error on overflow attempt */
 761                         if (dst_pos == dest_size)
 762                                 return -1;
 763                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
 764                         dst_pos++;
 765                 }
 766         }
 767         return 0;
 768 }
 769
 770 /*
 771  * unmaps hugepages that are not going to be used. since we originally allocate
 772  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 773  */
 774 static int
 775 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
 776                 struct hugepage_info *hpi,
 777                 unsigned num_hp_info)
 778 {
 779         unsigned socket, size;
 780         int page, nrpages = 0;
 781
 782         /* get total number of hugepages */
 783         for (size = 0; size < num_hp_info; size++)
 784                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 785                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 786
 787         for (size = 0; size < num_hp_info; size++) {
 788                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 789                         unsigned pages_found = 0;
 790
 791                         /* traverse until we have unmapped all the unused pages */
 792                         for (page = 0; page < nrpages; page++) {
 793                                 struct hugepage_file *hp = &hugepg_tbl[page];
 794
 795 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 796                                 /* if this page was already cleared */
 797                                 if (hp->final_va == NULL)
 798                                         continue;
 799 #endif
 800
 801                                 /* find a page that matches the criteria */
 802                                 if ((hp->size == hpi[size].hugepage_sz) &&
 803                                                 (hp->socket_id == (int) socket)) {
 804
 805                                         /* if we skipped enough pages, unmap the rest */
 806                                         if (pages_found == hpi[size].num_pages[socket]) {
 807                                                 uint64_t unmap_len;
 808
 809 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 810                                                 unmap_len = hp->size * hp->repeated;
 811 #else
 812                                                 unmap_len = hp->size;
 813 #endif
 814
 815                                                 /* get start addr and len of the remaining segment */
 816                                                 munmap(hp->final_va, (size_t) unmap_len);
 817
 818                                                 hp->final_va = NULL;
 819                                                 if (unlink(hp->filepath) == -1) {
 820                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 821                                                                         __func__, hp->filepath, strerror(errno));
 822                                                         return -1;
 823                                                 }
 824                                         }
 825 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 826                                         /* else, check how much do we need to map */
 827                                         else {
 828                                                 int nr_pg_left =
 829                                                                 hpi[size].num_pages[socket] - pages_found;
 830
 831                                                 /* if we need enough memory to fit into the segment */
 832                                                 if (hp->repeated <= nr_pg_left) {
 833                                                         pages_found += hp->repeated;
 834                                                 }
 835                                                 /* truncate the segment */
 836                                                 else {
 837                                                         uint64_t final_size = nr_pg_left * hp->size;
 838                                                         uint64_t seg_size = hp->repeated * hp->size;
 839
 840                                                         void * unmap_va = RTE_PTR_ADD(hp->final_va,
 841                                                                         final_size);
 842                                                         int fd;
 843
 844                                                         munmap(unmap_va, seg_size - final_size);
 845
 846                                                         fd = open(hp->filepath, O_RDWR);
 847                                                         if (fd < 0) {
 848                                                                 RTE_LOG(ERR, EAL, "Cannot open %s: %s\n",
 849                                                                                 hp->filepath, strerror(errno));
 850                                                                 return -1;
 851                                                         }
 852                                                         if (ftruncate(fd, final_size) < 0) {
 853                                                                 RTE_LOG(ERR, EAL, "Cannot truncate %s: %s\n",
 854                                                                                 hp->filepath, strerror(errno));
 855                                                                 return -1;
 856                                                         }
 857                                                         close(fd);
 858
 859                                                         pages_found += nr_pg_left;
 860                                                         hp->repeated = nr_pg_left;
 861                                                 }
 862                                         }
 863 #else
 864                                         /* else, lock the page and skip */
 865                                         else
 866                                                 pages_found++;
 867 #endif
 868
 869                                 } /* match page */
 870                         } /* foreach page */
 871                 } /* foreach socket */
 872         } /* foreach pagesize */
 873
 874         return 0;
 875 }
 876
 877 static inline uint64_t
 878 get_socket_mem_size(int socket)
 879 {
 880         uint64_t size = 0;
 881         unsigned i;
 882
 883         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 884                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 885                 if (hpi->hugedir != NULL)
 886                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 887         }
 888
 889         return size;
 890 }
 891
 892 /*
 893  * This function is a NUMA-aware equivalent of calc_num_pages.
 894  * It takes in the list of hugepage sizes and the
 895  * number of pages thereof, and calculates the best number of
 896  * pages of each size to fulfill the request for <memory> ram
 897  */
 898 static int
 899 calc_num_pages_per_socket(uint64_t * memory,
 900                 struct hugepage_info *hp_info,
 901                 struct hugepage_info *hp_used,
 902                 unsigned num_hp_info)
 903 {
 904         unsigned socket, j, i = 0;
 905         unsigned requested, available;
 906         int total_num_pages = 0;
 907         uint64_t remaining_mem, cur_mem;
 908         uint64_t total_mem = internal_config.memory;
 909
 910         if (num_hp_info == 0)
 911                 return -1;
 912
 913         /* if specific memory amounts per socket weren't requested */
 914         if (internal_config.force_sockets == 0) {
 915                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
 916                 size_t default_size, total_size;
 917                 unsigned lcore_id;
 918
 919                 /* Compute number of cores per socket */
 920                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
 921                 RTE_LCORE_FOREACH(lcore_id) {
 922                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
 923                 }
 924
 925                 /*
 926                  * Automatically spread requested memory amongst detected sockets according
 927                  * to number of cores from cpu mask present on each socket
 928                  */
 929                 total_size = internal_config.memory;
 930                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 931
 932                         /* Set memory amount per socket */
 933                         default_size = (internal_config.memory * cpu_per_socket[socket])
 934                                         / rte_lcore_count();
 935
 936                         /* Limit to maximum available memory on socket */
 937                         default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
 938
 939                         /* Update sizes */
 940                         memory[socket] = default_size;
 941                         total_size -= default_size;
 942                 }
 943
 944                 /*
 945                  * If some memory is remaining, try to allocate it by getting all
 946                  * available memory from sockets, one after the other
 947                  */
 948                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 949                         /* take whatever is available */
 950                         default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
 951                                                total_size);
 952
 953                         /* Update sizes */
 954                         memory[socket] += default_size;
 955                         total_size -= default_size;
 956                 }
 957         }
 958
 959         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 960                 /* skips if the memory on specific socket wasn't requested */
 961                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 962                         hp_used[i].hugedir = hp_info[i].hugedir;
 963                         hp_used[i].num_pages[socket] = RTE_MIN(
 964                                         memory[socket] / hp_info[i].hugepage_sz,
 965                                         hp_info[i].num_pages[socket]);
 966
 967                         cur_mem = hp_used[i].num_pages[socket] *
 968                                         hp_used[i].hugepage_sz;
 969
 970                         memory[socket] -= cur_mem;
 971                         total_mem -= cur_mem;
 972
 973                         total_num_pages += hp_used[i].num_pages[socket];
 974
 975                         /* check if we have met all memory requests */
 976                         if (memory[socket] == 0)
 977                                 break;
 978
 979                         /* check if we have any more pages left at this size, if so
 980                          * move on to next size */
 981                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
 982                                 continue;
 983                         /* At this point we know that there are more pages available that are
 984                          * bigger than the memory we want, so lets see if we can get enough
 985                          * from other page sizes.
 986                          */
 987                         remaining_mem = 0;
 988                         for (j = i+1; j < num_hp_info; j++)
 989                                 remaining_mem += hp_info[j].hugepage_sz *
 990                                 hp_info[j].num_pages[socket];
 991
 992                         /* is there enough other memory, if not allocate another page and quit */
 993                         if (remaining_mem < memory[socket]){
 994                                 cur_mem = RTE_MIN(memory[socket],
 995                                                 hp_info[i].hugepage_sz);
 996                                 memory[socket] -= cur_mem;
 997                                 total_mem -= cur_mem;
 998                                 hp_used[i].num_pages[socket]++;
 999                                 total_num_pages++;
1000                                 break; /* we are done with this socket*/
1001                         }
1002                 }
1003                 /* if we didn't satisfy all memory requirements per socket */
1004                 if (memory[socket] > 0) {
1005                         /* to prevent icc errors */
1006                         requested = (unsigned) (internal_config.socket_mem[socket] /
1007                                         0x100000);
1008                         available = requested -
1009                                         ((unsigned) (memory[socket] / 0x100000));
1010                         RTE_LOG(INFO, EAL, "Not enough memory available on socket %u! "
1011                                         "Requested: %uMB, available: %uMB\n", socket,
1012                                         requested, available);
1013                         return -1;
1014                 }
1015         }
1016
1017         /* if we didn't satisfy total memory requirements */
1018         if (total_mem > 0) {
1019                 requested = (unsigned) (internal_config.memory / 0x100000);
1020                 available = requested - (unsigned) (total_mem / 0x100000);
1021                 RTE_LOG(INFO, EAL, "Not enough memory available! Requested: %uMB,"
1022                                 " available: %uMB\n", requested, available);
1023                 return -1;
1024         }
1025         return total_num_pages;
1026 }
1027
1028 /*
1029  * Prepare physical memory mapping: fill configuration structure with
1030  * these infos, return 0 on success.
1031  *  1. map N huge pages in separate files in hugetlbfs
1032  *  2. find associated physical addr
1033  *  3. find associated NUMA socket ID
1034  *  4. sort all huge pages by physical address
1035  *  5. remap these N huge pages in the correct order
1036  *  6. unmap the first mapping
1037  *  7. fill memsegs in configuration with contiguous zones
1038  */
1039 static int
1040 rte_eal_hugepage_init(void)
1041 {
1042         struct rte_mem_config *mcfg;
1043         struct hugepage_file *hugepage, *tmp_hp = NULL;
1044         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1045
1046         uint64_t memory[RTE_MAX_NUMA_NODES];
1047
1048         unsigned hp_offset;
1049         int i, j, new_memseg;
1050         int nr_hugefiles, nr_hugepages = 0;
1051         void *addr;
1052 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1053         int new_pages_count[MAX_HUGEPAGE_SIZES];
1054 #endif
1055
1056         memset(used_hp, 0, sizeof(used_hp));
1057
1058         /* get pointer to global configuration */
1059         mcfg = rte_eal_get_configuration()->mem_config;
1060
1061         /* hugetlbfs can be disabled */
1062         if (internal_config.no_hugetlbfs) {
1063                 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1064                                 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1065                 if (addr == MAP_FAILED) {
1066                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1067                                         strerror(errno));
1068                         return -1;
1069                 }
1070                 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
1071                 mcfg->memseg[0].addr = addr;
1072                 mcfg->memseg[0].len = internal_config.memory;
1073                 mcfg->memseg[0].socket_id = SOCKET_ID_ANY;
1074                 return 0;
1075         }
1076
1077 /* check if app runs on Xen Dom0 */
1078         if (internal_config.xen_dom0_support) {
1079 #ifdef RTE_LIBRTE_XEN_DOM0
1080                 /* use dom0_mm kernel driver to init memory */
1081                 if (rte_xen_dom0_memory_init() < 0)
1082                         return -1;
1083                 else
1084                         return 0;
1085 #endif
1086         }
1087
1088
1089         /* calculate total number of hugepages available. at this point we haven't
1090          * yet started sorting them so they all are on socket 0 */
1091         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1092                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1093                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1094
1095                 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1096         }
1097
1098         /*
1099          * allocate a memory area for hugepage table.
1100          * this isn't shared memory yet. due to the fact that we need some
1101          * processing done on these pages, shared memory will be created
1102          * at a later stage.
1103          */
1104         tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1105         if (tmp_hp == NULL)
1106                 goto fail;
1107
1108         memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1109
1110         hp_offset = 0; /* where we start the current page size entries */
1111
1112         /* map all hugepages and sort them */
1113         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
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                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 1) < 0){
1128                         RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
1129                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1130                         goto fail;
1131                 }
1132
1133                 /* find physical addresses and sockets for each hugepage */
1134                 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0){
1135                         RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
1136                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1137                         goto fail;
1138                 }
1139
1140                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1141                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1142                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1143                         goto fail;
1144                 }
1145
1146                 if (sort_by_physaddr(&tmp_hp[hp_offset], hpi) < 0)
1147                         goto fail;
1148
1149 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1150                 /* remap all hugepages into single file segments */
1151                 new_pages_count[i] = remap_all_hugepages(&tmp_hp[hp_offset], hpi);
1152                 if (new_pages_count[i] < 0){
1153                         RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1154                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1155                         goto fail;
1156                 }
1157
1158                 /* we have processed a num of hugepages of this size, so inc offset */
1159                 hp_offset += new_pages_count[i];
1160 #else
1161                 /* remap all hugepages */
1162                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) < 0){
1163                         RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1164                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1165                         goto fail;
1166                 }
1167
1168                 /* unmap original mappings */
1169                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1170                         goto fail;
1171
1172                 /* we have processed a num of hugepages of this size, so inc offset */
1173                 hp_offset += hpi->num_pages[0];
1174 #endif
1175         }
1176
1177 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1178         nr_hugefiles = 0;
1179         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1180                 nr_hugefiles += new_pages_count[i];
1181         }
1182 #else
1183         nr_hugefiles = nr_hugepages;
1184 #endif
1185
1186
1187         /* clean out the numbers of pages */
1188         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1189                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1190                         internal_config.hugepage_info[i].num_pages[j] = 0;
1191
1192         /* get hugepages for each socket */
1193         for (i = 0; i < nr_hugefiles; i++) {
1194                 int socket = tmp_hp[i].socket_id;
1195
1196                 /* find a hugepage info with right size and increment num_pages */
1197                 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1198                                 (int)internal_config.num_hugepage_sizes);
1199                 for (j = 0; j < nb_hpsizes; j++) {
1200                         if (tmp_hp[i].size ==
1201                                         internal_config.hugepage_info[j].hugepage_sz) {
1202 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1203                                         internal_config.hugepage_info[j].num_pages[socket] +=
1204                                                 tmp_hp[i].repeated;
1205 #else
1206                                 internal_config.hugepage_info[j].num_pages[socket]++;
1207 #endif
1208                         }
1209                 }
1210         }
1211
1212         /* make a copy of socket_mem, needed for number of pages calculation */
1213         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1214                 memory[i] = internal_config.socket_mem[i];
1215
1216         /* calculate final number of pages */
1217         nr_hugepages = calc_num_pages_per_socket(memory,
1218                         internal_config.hugepage_info, used_hp,
1219                         internal_config.num_hugepage_sizes);
1220
1221         /* error if not enough memory available */
1222         if (nr_hugepages < 0)
1223                 goto fail;
1224
1225         /* reporting in! */
1226         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1227                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1228                         if (used_hp[i].num_pages[j] > 0) {
1229                                 RTE_LOG(INFO, EAL,
1230                                                 "Requesting %u pages of size %uMB"
1231                                                 " from socket %i\n",
1232                                                 used_hp[i].num_pages[j],
1233                                                 (unsigned)
1234                                                         (used_hp[i].hugepage_sz / 0x100000),
1235                                                 j);
1236                         }
1237                 }
1238         }
1239
1240         /* create shared memory */
1241         hugepage = create_shared_memory(eal_hugepage_info_path(),
1242                         nr_hugefiles * sizeof(struct hugepage_file));
1243
1244         if (hugepage == NULL) {
1245                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1246                 goto fail;
1247         }
1248         memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1249
1250         /*
1251          * unmap pages that we won't need (looks at used_hp).
1252          * also, sets final_va to NULL on pages that were unmapped.
1253          */
1254         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1255                         internal_config.num_hugepage_sizes) < 0) {
1256                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1257                 goto fail;
1258         }
1259
1260         /*
1261          * copy stuff from malloc'd hugepage* to the actual shared memory.
1262          * this procedure only copies those hugepages that have final_va
1263          * not NULL. has overflow protection.
1264          */
1265         if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1266                         tmp_hp, nr_hugefiles) < 0) {
1267                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1268                 goto fail;
1269         }
1270
1271         /* free the temporary hugepage table */
1272         free(tmp_hp);
1273         tmp_hp = NULL;
1274
1275         /* find earliest free memseg - this is needed because in case of IVSHMEM,
1276          * segments might have already been initialized */
1277         for (j = 0; j < RTE_MAX_MEMSEG; j++)
1278                 if (mcfg->memseg[j].addr == NULL) {
1279                         /* move to previous segment and exit loop */
1280                         j--;
1281                         break;
1282                 }
1283
1284         for (i = 0; i < nr_hugefiles; i++) {
1285                 new_memseg = 0;
1286
1287                 /* if this is a new section, create a new memseg */
1288                 if (i == 0)
1289                         new_memseg = 1;
1290                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1291                         new_memseg = 1;
1292                 else if (hugepage[i].size != hugepage[i-1].size)
1293                         new_memseg = 1;
1294
1295 #ifdef RTE_ARCH_PPC_64
1296                 /* On PPC64 architecture, the mmap always start from higher
1297                  * virtual address to lower address. Here, both the physical
1298                  * address and virtual address are in descending order */
1299                 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1300                     hugepage[i].size)
1301                         new_memseg = 1;
1302                 else if (((unsigned long)hugepage[i-1].final_va -
1303                     (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1304                         new_memseg = 1;
1305 #else
1306                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1307                     hugepage[i].size)
1308                         new_memseg = 1;
1309                 else if (((unsigned long)hugepage[i].final_va -
1310                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1311                         new_memseg = 1;
1312 #endif
1313
1314                 if (new_memseg) {
1315                         j += 1;
1316                         if (j == RTE_MAX_MEMSEG)
1317                                 break;
1318
1319                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1320                         mcfg->memseg[j].addr = hugepage[i].final_va;
1321 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1322                         mcfg->memseg[j].len = hugepage[i].size * hugepage[i].repeated;
1323 #else
1324                         mcfg->memseg[j].len = hugepage[i].size;
1325 #endif
1326                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1327                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1328                 }
1329                 /* continuation of previous memseg */
1330                 else {
1331 #ifdef RTE_ARCH_PPC_64
1332                 /* Use the phy and virt address of the last page as segment
1333                  * address for IBM Power architecture */
1334                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1335                         mcfg->memseg[j].addr = hugepage[i].final_va;
1336 #endif
1337                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1338                 }
1339                 hugepage[i].memseg_id = j;
1340         }
1341
1342         if (i < nr_hugefiles) {
1343                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1344                         "from %d requested\n"
1345                         "Current %s=%d is not enough\n"
1346                         "Please either increase it or request less amount "
1347                         "of memory.\n",
1348                         i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1349                         RTE_MAX_MEMSEG);
1350                 return -ENOMEM;
1351         }
1352
1353         return 0;
1354
1355 fail:
1356         if (tmp_hp)
1357                 free(tmp_hp);
1358         return -1;
1359 }
1360
1361 /*
1362  * uses fstat to report the size of a file on disk
1363  */
1364 static off_t
1365 getFileSize(int fd)
1366 {
1367         struct stat st;
1368         if (fstat(fd, &st) < 0)
1369                 return 0;
1370         return st.st_size;
1371 }
1372
1373 /*
1374  * This creates the memory mappings in the secondary process to match that of
1375  * the server process. It goes through each memory segment in the DPDK runtime
1376  * configuration and finds the hugepages which form that segment, mapping them
1377  * in order to form a contiguous block in the virtual memory space
1378  */
1379 static int
1380 rte_eal_hugepage_attach(void)
1381 {
1382         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1383         const struct hugepage_file *hp = NULL;
1384         unsigned num_hp = 0;
1385         unsigned i, s = 0; /* s used to track the segment number */
1386         off_t size;
1387         int fd, fd_zero = -1, fd_hugepage = -1;
1388
1389         if (aslr_enabled() > 0) {
1390                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1391                                 "(ASLR) is enabled in the kernel.\n");
1392                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1393                                 "into secondary processes\n");
1394         }
1395
1396         if (internal_config.xen_dom0_support) {
1397 #ifdef RTE_LIBRTE_XEN_DOM0
1398                 if (rte_xen_dom0_memory_attach() < 0) {
1399                         RTE_LOG(ERR, EAL,"Failed to attach memory setments of primay "
1400                                         "process\n");
1401                         return -1;
1402                 }
1403                 return 0;
1404 #endif
1405         }
1406
1407         fd_zero = open("/dev/zero", O_RDONLY);
1408         if (fd_zero < 0) {
1409                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1410                 goto error;
1411         }
1412         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1413         if (fd_hugepage < 0) {
1414                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1415                 goto error;
1416         }
1417
1418         /* map all segments into memory to make sure we get the addrs */
1419         for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1420                 void *base_addr;
1421
1422                 /*
1423                  * the first memory segment with len==0 is the one that
1424                  * follows the last valid segment.
1425                  */
1426                 if (mcfg->memseg[s].len == 0)
1427                         break;
1428
1429 #ifdef RTE_LIBRTE_IVSHMEM
1430                 /*
1431                  * if segment has ioremap address set, it's an IVSHMEM segment and
1432                  * doesn't need mapping as it was already mapped earlier
1433                  */
1434                 if (mcfg->memseg[s].ioremap_addr != 0)
1435                         continue;
1436 #endif
1437
1438                 /*
1439                  * fdzero is mmapped to get a contiguous block of virtual
1440                  * addresses of the appropriate memseg size.
1441                  * use mmap to get identical addresses as the primary process.
1442                  */
1443                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1444                                  PROT_READ, MAP_PRIVATE, fd_zero, 0);
1445                 if (base_addr == MAP_FAILED ||
1446                     base_addr != mcfg->memseg[s].addr) {
1447                         RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1448                                 "in /dev/zero to requested address [%p]: '%s'\n",
1449                                 (unsigned long long)mcfg->memseg[s].len,
1450                                 mcfg->memseg[s].addr, strerror(errno));
1451                         if (aslr_enabled() > 0) {
1452                                 RTE_LOG(ERR, EAL, "It is recommended to "
1453                                         "disable ASLR in the kernel "
1454                                         "and retry running both primary "
1455                                         "and secondary processes\n");
1456                         }
1457                         goto error;
1458                 }
1459         }
1460
1461         size = getFileSize(fd_hugepage);
1462         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1463         if (hp == NULL) {
1464                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1465                 goto error;
1466         }
1467
1468         num_hp = size / sizeof(struct hugepage_file);
1469         RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1470
1471         s = 0;
1472         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1473                 void *addr, *base_addr;
1474                 uintptr_t offset = 0;
1475                 size_t mapping_size;
1476 #ifdef RTE_LIBRTE_IVSHMEM
1477                 /*
1478                  * if segment has ioremap address set, it's an IVSHMEM segment and
1479                  * doesn't need mapping as it was already mapped earlier
1480                  */
1481                 if (mcfg->memseg[s].ioremap_addr != 0) {
1482                         s++;
1483                         continue;
1484                 }
1485 #endif
1486                 /*
1487                  * free previously mapped memory so we can map the
1488                  * hugepages into the space
1489                  */
1490                 base_addr = mcfg->memseg[s].addr;
1491                 munmap(base_addr, mcfg->memseg[s].len);
1492
1493                 /* find the hugepages for this segment and map them
1494                  * we don't need to worry about order, as the server sorted the
1495                  * entries before it did the second mmap of them */
1496                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1497                         if (hp[i].memseg_id == (int)s){
1498                                 fd = open(hp[i].filepath, O_RDWR);
1499                                 if (fd < 0) {
1500                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1501                                                 hp[i].filepath);
1502                                         goto error;
1503                                 }
1504 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1505                                 mapping_size = hp[i].size * hp[i].repeated;
1506 #else
1507                                 mapping_size = hp[i].size;
1508 #endif
1509                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1510                                                 mapping_size, PROT_READ | PROT_WRITE,
1511                                                 MAP_SHARED, fd, 0);
1512                                 close(fd); /* close file both on success and on failure */
1513                                 if (addr == MAP_FAILED ||
1514                                                 addr != RTE_PTR_ADD(base_addr, offset)) {
1515                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1516                                                 hp[i].filepath);
1517                                         goto error;
1518                                 }
1519                                 offset+=mapping_size;
1520                         }
1521                 }
1522                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1523                                 (unsigned long long)mcfg->memseg[s].len);
1524                 s++;
1525         }
1526         /* unmap the hugepage config file, since we are done using it */
1527         munmap((void *)(uintptr_t)hp, size);
1528         close(fd_zero);
1529         close(fd_hugepage);
1530         return 0;
1531
1532 error:
1533         if (fd_zero >= 0)
1534                 close(fd_zero);
1535         if (fd_hugepage >= 0)
1536                 close(fd_hugepage);
1537         return -1;
1538 }
1539
1540 static int
1541 rte_eal_memdevice_init(void)
1542 {
1543         struct rte_config *config;
1544
1545         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1546                 return 0;
1547
1548         config = rte_eal_get_configuration();
1549         config->mem_config->nchannel = internal_config.force_nchannel;
1550         config->mem_config->nrank = internal_config.force_nrank;
1551
1552         return 0;
1553 }
1554
1555 static int
1556 test_proc_pagemap_readable(void)
1557 {
1558         int fd = open("/proc/self/pagemap", O_RDONLY);
1559
1560         if (fd < 0)
1561                 return 0;
1562         /* Is readable */
1563         close(fd);
1564
1565         return 1;
1566 }
1567
1568 /* init memory subsystem */
1569 int
1570 rte_eal_memory_init(void)
1571 {
1572         RTE_LOG(INFO, EAL, "Setting up memory...\n");
1573
1574         proc_pagemap_readable = test_proc_pagemap_readable();
1575         if (!proc_pagemap_readable)
1576                 RTE_LOG(ERR, EAL,
1577                         "Cannot open /proc/self/pagemap: %s. "
1578                         "virt2phys address translation will not work\n",
1579                         strerror(errno));
1580
1581         const int retval = rte_eal_process_type() == RTE_PROC_PRIMARY ?
1582                         rte_eal_hugepage_init() :
1583                         rte_eal_hugepage_attach();
1584         if (retval < 0)
1585                 return -1;
1586
1587         if (internal_config.no_shconf == 0 && rte_eal_memdevice_init() < 0)
1588                 return -1;
1589
1590         return 0;
1591 }