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 static void
 119 test_proc_pagemap_readable(void)
 120 {
 121         int fd = open("/proc/self/pagemap", O_RDONLY);
 122
 123         if (fd < 0) {
 124                 RTE_LOG(ERR, EAL,
 125                         "Cannot open /proc/self/pagemap: %s. "
 126                         "virt2phys address translation will not work\n",
 127                         strerror(errno));
 128                 return;
 129         }
 130
 131         /* Is readable */
 132         close(fd);
 133         proc_pagemap_readable = 1;
 134 }
 135
 136 /* Lock page in physical memory and prevent from swapping. */
 137 int
 138 rte_mem_lock_page(const void *virt)
 139 {
 140         unsigned long virtual = (unsigned long)virt;
 141         int page_size = getpagesize();
 142         unsigned long aligned = (virtual & ~ (page_size - 1));
 143         return mlock((void*)aligned, page_size);
 144 }
 145
 146 /*
 147  * Get physical address of any mapped virtual address in the current process.
 148  */
 149 phys_addr_t
 150 rte_mem_virt2phy(const void *virtaddr)
 151 {
 152         int fd;
 153         uint64_t page, physaddr;
 154         unsigned long virt_pfn;
 155         int page_size;
 156         off_t offset;
 157
 158         /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
 159         if (!proc_pagemap_readable)
 160                 return RTE_BAD_PHYS_ADDR;
 161
 162         /* standard page size */
 163         page_size = getpagesize();
 164
 165         fd = open("/proc/self/pagemap", O_RDONLY);
 166         if (fd < 0) {
 167                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 168                         __func__, strerror(errno));
 169                 return RTE_BAD_PHYS_ADDR;
 170         }
 171
 172         virt_pfn = (unsigned long)virtaddr / page_size;
 173         offset = sizeof(uint64_t) * virt_pfn;
 174         if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 175                 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 176                                 __func__, strerror(errno));
 177                 close(fd);
 178                 return RTE_BAD_PHYS_ADDR;
 179         }
 180         if (read(fd, &page, sizeof(uint64_t)) < 0) {
 181                 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 182                                 __func__, strerror(errno));
 183                 close(fd);
 184                 return RTE_BAD_PHYS_ADDR;
 185         }
 186
 187         /*
 188          * the pfn (page frame number) are bits 0-54 (see
 189          * pagemap.txt in linux Documentation)
 190          */
 191         physaddr = ((page & 0x7fffffffffffffULL) * page_size)
 192                 + ((unsigned long)virtaddr % page_size);
 193         close(fd);
 194         return physaddr;
 195 }
 196
 197 /*
 198  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 199  * it by browsing the /proc/self/pagemap special file.
 200  */
 201 static int
 202 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 203 {
 204         unsigned i;
 205         phys_addr_t addr;
 206
 207         for (i = 0; i < hpi->num_pages[0]; i++) {
 208                 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
 209                 if (addr == RTE_BAD_PHYS_ADDR)
 210                         return -1;
 211                 hugepg_tbl[i].physaddr = addr;
 212         }
 213         return 0;
 214 }
 215
 216 /*
 217  * Check whether address-space layout randomization is enabled in
 218  * the kernel. This is important for multi-process as it can prevent
 219  * two processes mapping data to the same virtual address
 220  * Returns:
 221  *    0 - address space randomization disabled
 222  *    1/2 - address space randomization enabled
 223  *    negative error code on error
 224  */
 225 static int
 226 aslr_enabled(void)
 227 {
 228         char c;
 229         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 230         if (fd < 0)
 231                 return -errno;
 232         retval = read(fd, &c, 1);
 233         close(fd);
 234         if (retval < 0)
 235                 return -errno;
 236         if (retval == 0)
 237                 return -EIO;
 238         switch (c) {
 239                 case '0' : return 0;
 240                 case '1' : return 1;
 241                 case '2' : return 2;
 242                 default: return -EINVAL;
 243         }
 244 }
 245
 246 /*
 247  * Try to mmap *size bytes in /dev/zero. If it is successful, return the
 248  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 249  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 250  * 0. In this case, return NULL. Note: this function returns an address
 251  * which is a multiple of hugepage size.
 252  */
 253 static void *
 254 get_virtual_area(size_t *size, size_t hugepage_sz)
 255 {
 256         void *addr;
 257         int fd;
 258         long aligned_addr;
 259
 260         if (internal_config.base_virtaddr != 0) {
 261                 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
 262                                 baseaddr_offset);
 263         }
 264         else addr = NULL;
 265
 266         RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
 267
 268         fd = open("/dev/zero", O_RDONLY);
 269         if (fd < 0){
 270                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 271                 return NULL;
 272         }
 273         do {
 274                 addr = mmap(addr,
 275                                 (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
 276                 if (addr == MAP_FAILED)
 277                         *size -= hugepage_sz;
 278         } while (addr == MAP_FAILED && *size > 0);
 279
 280         if (addr == MAP_FAILED) {
 281                 close(fd);
 282                 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
 283                         strerror(errno));
 284                 return NULL;
 285         }
 286
 287         munmap(addr, (*size) + hugepage_sz);
 288         close(fd);
 289
 290         /* align addr to a huge page size boundary */
 291         aligned_addr = (long)addr;
 292         aligned_addr += (hugepage_sz - 1);
 293         aligned_addr &= (~(hugepage_sz - 1));
 294         addr = (void *)(aligned_addr);
 295
 296         RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
 297                 addr, *size);
 298
 299         /* increment offset */
 300         baseaddr_offset += *size;
 301
 302         return addr;
 303 }
 304
 305 /*
 306  * Mmap all hugepages of hugepage table: it first open a file in
 307  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 308  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 309  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 310  * map continguous physical blocks in contiguous virtual blocks.
 311  */
 312 static int
 313 map_all_hugepages(struct hugepage_file *hugepg_tbl,
 314                 struct hugepage_info *hpi, int orig)
 315 {
 316         int fd;
 317         unsigned i;
 318         void *virtaddr;
 319         void *vma_addr = NULL;
 320         size_t vma_len = 0;
 321
 322 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 323         RTE_SET_USED(vma_len);
 324 #endif
 325
 326         for (i = 0; i < hpi->num_pages[0]; i++) {
 327                 uint64_t hugepage_sz = hpi->hugepage_sz;
 328
 329                 if (orig) {
 330                         hugepg_tbl[i].file_id = i;
 331                         hugepg_tbl[i].size = hugepage_sz;
 332 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 333                         eal_get_hugefile_temp_path(hugepg_tbl[i].filepath,
 334                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 335                                         hugepg_tbl[i].file_id);
 336 #else
 337                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 338                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 339                                         hugepg_tbl[i].file_id);
 340 #endif
 341                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 342                 }
 343 #ifndef RTE_ARCH_64
 344                 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
 345                  * original map address as final map address.
 346                  */
 347                 else if ((hugepage_sz == RTE_PGSIZE_1G)
 348                         || (hugepage_sz == RTE_PGSIZE_16G)) {
 349                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 350                         hugepg_tbl[i].orig_va = NULL;
 351                         continue;
 352                 }
 353 #endif
 354
 355 #ifndef RTE_EAL_SINGLE_FILE_SEGMENTS
 356                 else if (vma_len == 0) {
 357                         unsigned j, num_pages;
 358
 359                         /* reserve a virtual area for next contiguous
 360                          * physical block: count the number of
 361                          * contiguous physical pages. */
 362                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 363 #ifdef RTE_ARCH_PPC_64
 364                                 /* The physical addresses are sorted in
 365                                  * descending order on PPC64 */
 366                                 if (hugepg_tbl[j].physaddr !=
 367                                     hugepg_tbl[j-1].physaddr - hugepage_sz)
 368                                         break;
 369 #else
 370                                 if (hugepg_tbl[j].physaddr !=
 371                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 372                                         break;
 373 #endif
 374                         }
 375                         num_pages = j - i;
 376                         vma_len = num_pages * hugepage_sz;
 377
 378                         /* get the biggest virtual memory area up to
 379                          * vma_len. If it fails, vma_addr is NULL, so
 380                          * let the kernel provide the address. */
 381                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 382                         if (vma_addr == NULL)
 383                                 vma_len = hugepage_sz;
 384                 }
 385 #endif
 386
 387                 /* try to create hugepage file */
 388                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
 389                 if (fd < 0) {
 390                         RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
 391                                         strerror(errno));
 392                         return -1;
 393                 }
 394
 395                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 396                                 MAP_SHARED, fd, 0);
 397                 if (virtaddr == MAP_FAILED) {
 398                         RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__,
 399                                         strerror(errno));
 400                         close(fd);
 401                         return -1;
 402                 }
 403
 404                 if (orig) {
 405                         hugepg_tbl[i].orig_va = virtaddr;
 406                         memset(virtaddr, 0, hugepage_sz);
 407                 }
 408                 else {
 409                         hugepg_tbl[i].final_va = virtaddr;
 410                 }
 411
 412                 /* set shared flock on the file. */
 413                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 414                         RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
 415                                 __func__, strerror(errno));
 416                         close(fd);
 417                         return -1;
 418                 }
 419
 420                 close(fd);
 421
 422                 vma_addr = (char *)vma_addr + hugepage_sz;
 423                 vma_len -= hugepage_sz;
 424         }
 425         return 0;
 426 }
 427
 428 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 429
 430 /*
 431  * Remaps all hugepages into single file segments
 432  */
 433 static int
 434 remap_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 435 {
 436         int fd;
 437         unsigned i = 0, j, num_pages, page_idx = 0;
 438         void *vma_addr = NULL, *old_addr = NULL, *page_addr = NULL;
 439         size_t vma_len = 0;
 440         size_t hugepage_sz = hpi->hugepage_sz;
 441         size_t total_size, offset;
 442         char filepath[MAX_HUGEPAGE_PATH];
 443         phys_addr_t physaddr;
 444         int socket;
 445
 446         while (i < hpi->num_pages[0]) {
 447
 448 #ifndef RTE_ARCH_64
 449                 /* for 32-bit systems, don't remap 1G pages and 16G pages,
 450                  * just reuse original map address as final map address.
 451                  */
 452                 if ((hugepage_sz == RTE_PGSIZE_1G)
 453                         || (hugepage_sz == RTE_PGSIZE_16G)) {
 454                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 455                         hugepg_tbl[i].orig_va = NULL;
 456                         i++;
 457                         continue;
 458                 }
 459 #endif
 460
 461                 /* reserve a virtual area for next contiguous
 462                  * physical block: count the number of
 463                  * contiguous physical pages. */
 464                 for (j = i+1; j < hpi->num_pages[0] ; j++) {
 465 #ifdef RTE_ARCH_PPC_64
 466                         /* The physical addresses are sorted in descending
 467                          * order on PPC64 */
 468                         if (hugepg_tbl[j].physaddr !=
 469                                 hugepg_tbl[j-1].physaddr - hugepage_sz)
 470                                 break;
 471 #else
 472                         if (hugepg_tbl[j].physaddr !=
 473                                 hugepg_tbl[j-1].physaddr + hugepage_sz)
 474                                 break;
 475 #endif
 476                 }
 477                 num_pages = j - i;
 478                 vma_len = num_pages * hugepage_sz;
 479
 480                 socket = hugepg_tbl[i].socket_id;
 481
 482                 /* get the biggest virtual memory area up to
 483                  * vma_len. If it fails, vma_addr is NULL, so
 484                  * let the kernel provide the address. */
 485                 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 486
 487                 /* If we can't find a big enough virtual area, work out how many pages
 488                  * we are going to get */
 489                 if (vma_addr == NULL)
 490                         j = i + 1;
 491                 else if (vma_len != num_pages * hugepage_sz) {
 492                         num_pages = vma_len / hugepage_sz;
 493                         j = i + num_pages;
 494
 495                 }
 496
 497                 hugepg_tbl[page_idx].file_id = page_idx;
 498                 eal_get_hugefile_path(filepath,
 499                                 sizeof(filepath),
 500                                 hpi->hugedir,
 501                                 hugepg_tbl[page_idx].file_id);
 502
 503                 /* try to create hugepage file */
 504                 fd = open(filepath, O_CREAT | O_RDWR, 0755);
 505                 if (fd < 0) {
 506                         RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__, strerror(errno));
 507                         return -1;
 508                 }
 509
 510                 total_size = 0;
 511                 for (;i < j; i++) {
 512
 513                         /* unmap current segment */
 514                         if (total_size > 0)
 515                                 munmap(vma_addr, total_size);
 516
 517                         /* unmap original page */
 518                         munmap(hugepg_tbl[i].orig_va, hugepage_sz);
 519                         unlink(hugepg_tbl[i].filepath);
 520
 521                         total_size += hugepage_sz;
 522
 523                         old_addr = vma_addr;
 524
 525                         /* map new, bigger segment */
 526                         vma_addr = mmap(vma_addr, total_size,
 527                                         PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 528
 529                         if (vma_addr == MAP_FAILED || vma_addr != old_addr) {
 530                                 RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__, strerror(errno));
 531                                 close(fd);
 532                                 return -1;
 533                         }
 534
 535                         /* touch the page. this is needed because kernel postpones mapping
 536                          * creation until the first page fault. with this, we pin down
 537                          * the page and it is marked as used and gets into process' pagemap.
 538                          */
 539                         for (offset = 0; offset < total_size; offset += hugepage_sz)
 540                                 *((volatile uint8_t*) RTE_PTR_ADD(vma_addr, offset));
 541                 }
 542
 543                 /* set shared flock on the file. */
 544                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 545                         RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
 546                                 __func__, strerror(errno));
 547                         close(fd);
 548                         return -1;
 549                 }
 550
 551                 snprintf(hugepg_tbl[page_idx].filepath, MAX_HUGEPAGE_PATH, "%s",
 552                                 filepath);
 553
 554                 physaddr = rte_mem_virt2phy(vma_addr);
 555
 556                 if (physaddr == RTE_BAD_PHYS_ADDR)
 557                         return -1;
 558
 559                 hugepg_tbl[page_idx].final_va = vma_addr;
 560
 561                 hugepg_tbl[page_idx].physaddr = physaddr;
 562
 563                 hugepg_tbl[page_idx].repeated = num_pages;
 564
 565                 hugepg_tbl[page_idx].socket_id = socket;
 566
 567                 close(fd);
 568
 569                 /* verify the memory segment - that is, check that every VA corresponds
 570                  * to the physical address we expect to see
 571                  */
 572                 for (offset = 0; offset < vma_len; offset += hugepage_sz) {
 573                         uint64_t expected_physaddr;
 574
 575                         expected_physaddr = hugepg_tbl[page_idx].physaddr + offset;
 576                         page_addr = RTE_PTR_ADD(vma_addr, offset);
 577                         physaddr = rte_mem_virt2phy(page_addr);
 578
 579                         if (physaddr != expected_physaddr) {
 580                                 RTE_LOG(ERR, EAL, "Segment sanity check failed: wrong physaddr "
 581                                                 "at %p (offset 0x%" PRIx64 ": 0x%" PRIx64
 582                                                 " (expected 0x%" PRIx64 ")\n",
 583                                                 page_addr, offset, physaddr, expected_physaddr);
 584                                 return -1;
 585                         }
 586                 }
 587
 588                 /* zero out the whole segment */
 589                 memset(hugepg_tbl[page_idx].final_va, 0, total_size);
 590
 591                 page_idx++;
 592         }
 593
 594         /* zero out the rest */
 595         memset(&hugepg_tbl[page_idx], 0, (hpi->num_pages[0] - page_idx) * sizeof(struct hugepage_file));
 596         return page_idx;
 597 }
 598 #else/* RTE_EAL_SINGLE_FILE_SEGMENTS=n */
 599
 600 /* Unmap all hugepages from original mapping */
 601 static int
 602 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 603 {
 604         unsigned i;
 605         for (i = 0; i < hpi->num_pages[0]; i++) {
 606                 if (hugepg_tbl[i].orig_va) {
 607                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 608                         hugepg_tbl[i].orig_va = NULL;
 609                 }
 610         }
 611         return 0;
 612 }
 613 #endif /* RTE_EAL_SINGLE_FILE_SEGMENTS */
 614
 615 /*
 616  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 617  * page.
 618  */
 619 static int
 620 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 621 {
 622         int socket_id;
 623         char *end, *nodestr;
 624         unsigned i, hp_count = 0;
 625         uint64_t virt_addr;
 626         char buf[BUFSIZ];
 627         char hugedir_str[PATH_MAX];
 628         FILE *f;
 629
 630         f = fopen("/proc/self/numa_maps", "r");
 631         if (f == NULL) {
 632                 RTE_LOG(NOTICE, EAL, "cannot open /proc/self/numa_maps,"
 633                                 " consider that all memory is in socket_id 0\n");
 634                 return 0;
 635         }
 636
 637         snprintf(hugedir_str, sizeof(hugedir_str),
 638                         "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
 639
 640         /* parse numa map */
 641         while (fgets(buf, sizeof(buf), f) != NULL) {
 642
 643                 /* ignore non huge page */
 644                 if (strstr(buf, " huge ") == NULL &&
 645                                 strstr(buf, hugedir_str) == NULL)
 646                         continue;
 647
 648                 /* get zone addr */
 649                 virt_addr = strtoull(buf, &end, 16);
 650                 if (virt_addr == 0 || end == buf) {
 651                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 652                         goto error;
 653                 }
 654
 655                 /* get node id (socket id) */
 656                 nodestr = strstr(buf, " N");
 657                 if (nodestr == NULL) {
 658                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 659                         goto error;
 660                 }
 661                 nodestr += 2;
 662                 end = strstr(nodestr, "=");
 663                 if (end == NULL) {
 664                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 665                         goto error;
 666                 }
 667                 end[0] = '\0';
 668                 end = NULL;
 669
 670                 socket_id = strtoul(nodestr, &end, 0);
 671                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 672                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 673                         goto error;
 674                 }
 675
 676                 /* if we find this page in our mappings, set socket_id */
 677                 for (i = 0; i < hpi->num_pages[0]; i++) {
 678                         void *va = (void *)(unsigned long)virt_addr;
 679                         if (hugepg_tbl[i].orig_va == va) {
 680                                 hugepg_tbl[i].socket_id = socket_id;
 681                                 hp_count++;
 682                         }
 683                 }
 684         }
 685
 686         if (hp_count < hpi->num_pages[0])
 687                 goto error;
 688
 689         fclose(f);
 690         return 0;
 691
 692 error:
 693         fclose(f);
 694         return -1;
 695 }
 696
 697 /*
 698  * Sort the hugepg_tbl by physical address (lower addresses first on x86,
 699  * higher address first on powerpc). We use a slow algorithm, but we won't
 700  * have millions of pages, and this is only done at init time.
 701  */
 702 static int
 703 sort_by_physaddr(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 704 {
 705         unsigned i, j;
 706         int compare_idx;
 707         uint64_t compare_addr;
 708         struct hugepage_file tmp;
 709
 710         for (i = 0; i < hpi->num_pages[0]; i++) {
 711                 compare_addr = 0;
 712                 compare_idx = -1;
 713
 714                 /*
 715                  * browse all entries starting at 'i', and find the
 716                  * entry with the smallest addr
 717                  */
 718                 for (j=i; j< hpi->num_pages[0]; j++) {
 719
 720                         if (compare_addr == 0 ||
 721 #ifdef RTE_ARCH_PPC_64
 722                                 hugepg_tbl[j].physaddr > compare_addr) {
 723 #else
 724                                 hugepg_tbl[j].physaddr < compare_addr) {
 725 #endif
 726                                 compare_addr = hugepg_tbl[j].physaddr;
 727                                 compare_idx = j;
 728                         }
 729                 }
 730
 731                 /* should not happen */
 732                 if (compare_idx == -1) {
 733                         RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
 734                         return -1;
 735                 }
 736
 737                 /* swap the 2 entries in the table */
 738                 memcpy(&tmp, &hugepg_tbl[compare_idx],
 739                         sizeof(struct hugepage_file));
 740                 memcpy(&hugepg_tbl[compare_idx], &hugepg_tbl[i],
 741                         sizeof(struct hugepage_file));
 742                 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage_file));
 743         }
 744         return 0;
 745 }
 746
 747 /*
 748  * Uses mmap to create a shared memory area for storage of data
 749  * Used in this file to store the hugepage file map on disk
 750  */
 751 static void *
 752 create_shared_memory(const char *filename, const size_t mem_size)
 753 {
 754         void *retval;
 755         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 756         if (fd < 0)
 757                 return NULL;
 758         if (ftruncate(fd, mem_size) < 0) {
 759                 close(fd);
 760                 return NULL;
 761         }
 762         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 763         close(fd);
 764         return retval;
 765 }
 766
 767 /*
 768  * this copies *active* hugepages from one hugepage table to another.
 769  * destination is typically the shared memory.
 770  */
 771 static int
 772 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
 773                 const struct hugepage_file * src, int src_size)
 774 {
 775         int src_pos, dst_pos = 0;
 776
 777         for (src_pos = 0; src_pos < src_size; src_pos++) {
 778                 if (src[src_pos].final_va != NULL) {
 779                         /* error on overflow attempt */
 780                         if (dst_pos == dest_size)
 781                                 return -1;
 782                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
 783                         dst_pos++;
 784                 }
 785         }
 786         return 0;
 787 }
 788
 789 /*
 790  * unmaps hugepages that are not going to be used. since we originally allocate
 791  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 792  */
 793 static int
 794 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
 795                 struct hugepage_info *hpi,
 796                 unsigned num_hp_info)
 797 {
 798         unsigned socket, size;
 799         int page, nrpages = 0;
 800
 801         /* get total number of hugepages */
 802         for (size = 0; size < num_hp_info; size++)
 803                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 804                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 805
 806         for (size = 0; size < num_hp_info; size++) {
 807                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 808                         unsigned pages_found = 0;
 809
 810                         /* traverse until we have unmapped all the unused pages */
 811                         for (page = 0; page < nrpages; page++) {
 812                                 struct hugepage_file *hp = &hugepg_tbl[page];
 813
 814 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 815                                 /* if this page was already cleared */
 816                                 if (hp->final_va == NULL)
 817                                         continue;
 818 #endif
 819
 820                                 /* find a page that matches the criteria */
 821                                 if ((hp->size == hpi[size].hugepage_sz) &&
 822                                                 (hp->socket_id == (int) socket)) {
 823
 824                                         /* if we skipped enough pages, unmap the rest */
 825                                         if (pages_found == hpi[size].num_pages[socket]) {
 826                                                 uint64_t unmap_len;
 827
 828 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 829                                                 unmap_len = hp->size * hp->repeated;
 830 #else
 831                                                 unmap_len = hp->size;
 832 #endif
 833
 834                                                 /* get start addr and len of the remaining segment */
 835                                                 munmap(hp->final_va, (size_t) unmap_len);
 836
 837                                                 hp->final_va = NULL;
 838                                                 if (unlink(hp->filepath) == -1) {
 839                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 840                                                                         __func__, hp->filepath, strerror(errno));
 841                                                         return -1;
 842                                                 }
 843                                         }
 844 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 845                                         /* else, check how much do we need to map */
 846                                         else {
 847                                                 int nr_pg_left =
 848                                                                 hpi[size].num_pages[socket] - pages_found;
 849
 850                                                 /* if we need enough memory to fit into the segment */
 851                                                 if (hp->repeated <= nr_pg_left) {
 852                                                         pages_found += hp->repeated;
 853                                                 }
 854                                                 /* truncate the segment */
 855                                                 else {
 856                                                         uint64_t final_size = nr_pg_left * hp->size;
 857                                                         uint64_t seg_size = hp->repeated * hp->size;
 858
 859                                                         void * unmap_va = RTE_PTR_ADD(hp->final_va,
 860                                                                         final_size);
 861                                                         int fd;
 862
 863                                                         munmap(unmap_va, seg_size - final_size);
 864
 865                                                         fd = open(hp->filepath, O_RDWR);
 866                                                         if (fd < 0) {
 867                                                                 RTE_LOG(ERR, EAL, "Cannot open %s: %s\n",
 868                                                                                 hp->filepath, strerror(errno));
 869                                                                 return -1;
 870                                                         }
 871                                                         if (ftruncate(fd, final_size) < 0) {
 872                                                                 RTE_LOG(ERR, EAL, "Cannot truncate %s: %s\n",
 873                                                                                 hp->filepath, strerror(errno));
 874                                                                 return -1;
 875                                                         }
 876                                                         close(fd);
 877
 878                                                         pages_found += nr_pg_left;
 879                                                         hp->repeated = nr_pg_left;
 880                                                 }
 881                                         }
 882 #else
 883                                         /* else, lock the page and skip */
 884                                         else
 885                                                 pages_found++;
 886 #endif
 887
 888                                 } /* match page */
 889                         } /* foreach page */
 890                 } /* foreach socket */
 891         } /* foreach pagesize */
 892
 893         return 0;
 894 }
 895
 896 static inline uint64_t
 897 get_socket_mem_size(int socket)
 898 {
 899         uint64_t size = 0;
 900         unsigned i;
 901
 902         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 903                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 904                 if (hpi->hugedir != NULL)
 905                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 906         }
 907
 908         return size;
 909 }
 910
 911 /*
 912  * This function is a NUMA-aware equivalent of calc_num_pages.
 913  * It takes in the list of hugepage sizes and the
 914  * number of pages thereof, and calculates the best number of
 915  * pages of each size to fulfill the request for <memory> ram
 916  */
 917 static int
 918 calc_num_pages_per_socket(uint64_t * memory,
 919                 struct hugepage_info *hp_info,
 920                 struct hugepage_info *hp_used,
 921                 unsigned num_hp_info)
 922 {
 923         unsigned socket, j, i = 0;
 924         unsigned requested, available;
 925         int total_num_pages = 0;
 926         uint64_t remaining_mem, cur_mem;
 927         uint64_t total_mem = internal_config.memory;
 928
 929         if (num_hp_info == 0)
 930                 return -1;
 931
 932         /* if specific memory amounts per socket weren't requested */
 933         if (internal_config.force_sockets == 0) {
 934                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
 935                 size_t default_size, total_size;
 936                 unsigned lcore_id;
 937
 938                 /* Compute number of cores per socket */
 939                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
 940                 RTE_LCORE_FOREACH(lcore_id) {
 941                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
 942                 }
 943
 944                 /*
 945                  * Automatically spread requested memory amongst detected sockets according
 946                  * to number of cores from cpu mask present on each socket
 947                  */
 948                 total_size = internal_config.memory;
 949                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 950
 951                         /* Set memory amount per socket */
 952                         default_size = (internal_config.memory * cpu_per_socket[socket])
 953                                         / rte_lcore_count();
 954
 955                         /* Limit to maximum available memory on socket */
 956                         default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
 957
 958                         /* Update sizes */
 959                         memory[socket] = default_size;
 960                         total_size -= default_size;
 961                 }
 962
 963                 /*
 964                  * If some memory is remaining, try to allocate it by getting all
 965                  * available memory from sockets, one after the other
 966                  */
 967                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 968                         /* take whatever is available */
 969                         default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
 970                                                total_size);
 971
 972                         /* Update sizes */
 973                         memory[socket] += default_size;
 974                         total_size -= default_size;
 975                 }
 976         }
 977
 978         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
 979                 /* skips if the memory on specific socket wasn't requested */
 980                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
 981                         hp_used[i].hugedir = hp_info[i].hugedir;
 982                         hp_used[i].num_pages[socket] = RTE_MIN(
 983                                         memory[socket] / hp_info[i].hugepage_sz,
 984                                         hp_info[i].num_pages[socket]);
 985
 986                         cur_mem = hp_used[i].num_pages[socket] *
 987                                         hp_used[i].hugepage_sz;
 988
 989                         memory[socket] -= cur_mem;
 990                         total_mem -= cur_mem;
 991
 992                         total_num_pages += hp_used[i].num_pages[socket];
 993
 994                         /* check if we have met all memory requests */
 995                         if (memory[socket] == 0)
 996                                 break;
 997
 998                         /* check if we have any more pages left at this size, if so
 999                          * move on to next size */
1000                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
1001                                 continue;
1002                         /* At this point we know that there are more pages available that are
1003                          * bigger than the memory we want, so lets see if we can get enough
1004                          * from other page sizes.
1005                          */
1006                         remaining_mem = 0;
1007                         for (j = i+1; j < num_hp_info; j++)
1008                                 remaining_mem += hp_info[j].hugepage_sz *
1009                                 hp_info[j].num_pages[socket];
1010
1011                         /* is there enough other memory, if not allocate another page and quit */
1012                         if (remaining_mem < memory[socket]){
1013                                 cur_mem = RTE_MIN(memory[socket],
1014                                                 hp_info[i].hugepage_sz);
1015                                 memory[socket] -= cur_mem;
1016                                 total_mem -= cur_mem;
1017                                 hp_used[i].num_pages[socket]++;
1018                                 total_num_pages++;
1019                                 break; /* we are done with this socket*/
1020                         }
1021                 }
1022                 /* if we didn't satisfy all memory requirements per socket */
1023                 if (memory[socket] > 0) {
1024                         /* to prevent icc errors */
1025                         requested = (unsigned) (internal_config.socket_mem[socket] /
1026                                         0x100000);
1027                         available = requested -
1028                                         ((unsigned) (memory[socket] / 0x100000));
1029                         RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
1030                                         "Requested: %uMB, available: %uMB\n", socket,
1031                                         requested, available);
1032                         return -1;
1033                 }
1034         }
1035
1036         /* if we didn't satisfy total memory requirements */
1037         if (total_mem > 0) {
1038                 requested = (unsigned) (internal_config.memory / 0x100000);
1039                 available = requested - (unsigned) (total_mem / 0x100000);
1040                 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
1041                                 " available: %uMB\n", requested, available);
1042                 return -1;
1043         }
1044         return total_num_pages;
1045 }
1046
1047 /*
1048  * Prepare physical memory mapping: fill configuration structure with
1049  * these infos, return 0 on success.
1050  *  1. map N huge pages in separate files in hugetlbfs
1051  *  2. find associated physical addr
1052  *  3. find associated NUMA socket ID
1053  *  4. sort all huge pages by physical address
1054  *  5. remap these N huge pages in the correct order
1055  *  6. unmap the first mapping
1056  *  7. fill memsegs in configuration with contiguous zones
1057  */
1058 int
1059 rte_eal_hugepage_init(void)
1060 {
1061         struct rte_mem_config *mcfg;
1062         struct hugepage_file *hugepage, *tmp_hp = NULL;
1063         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1064
1065         uint64_t memory[RTE_MAX_NUMA_NODES];
1066
1067         unsigned hp_offset;
1068         int i, j, new_memseg;
1069         int nr_hugefiles, nr_hugepages = 0;
1070         void *addr;
1071 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1072         int new_pages_count[MAX_HUGEPAGE_SIZES];
1073 #endif
1074
1075         test_proc_pagemap_readable();
1076
1077         memset(used_hp, 0, sizeof(used_hp));
1078
1079         /* get pointer to global configuration */
1080         mcfg = rte_eal_get_configuration()->mem_config;
1081
1082         /* hugetlbfs can be disabled */
1083         if (internal_config.no_hugetlbfs) {
1084                 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1085                                 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1086                 if (addr == MAP_FAILED) {
1087                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1088                                         strerror(errno));
1089                         return -1;
1090                 }
1091                 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
1092                 mcfg->memseg[0].addr = addr;
1093                 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1094                 mcfg->memseg[0].len = internal_config.memory;
1095                 mcfg->memseg[0].socket_id = 0;
1096                 return 0;
1097         }
1098
1099 /* check if app runs on Xen Dom0 */
1100         if (internal_config.xen_dom0_support) {
1101 #ifdef RTE_LIBRTE_XEN_DOM0
1102                 /* use dom0_mm kernel driver to init memory */
1103                 if (rte_xen_dom0_memory_init() < 0)
1104                         return -1;
1105                 else
1106                         return 0;
1107 #endif
1108         }
1109
1110         /* calculate total number of hugepages available. at this point we haven't
1111          * yet started sorting them so they all are on socket 0 */
1112         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1113                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1114                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1115
1116                 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1117         }
1118
1119         /*
1120          * allocate a memory area for hugepage table.
1121          * this isn't shared memory yet. due to the fact that we need some
1122          * processing done on these pages, shared memory will be created
1123          * at a later stage.
1124          */
1125         tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1126         if (tmp_hp == NULL)
1127                 goto fail;
1128
1129         memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1130
1131         hp_offset = 0; /* where we start the current page size entries */
1132
1133         /* map all hugepages and sort them */
1134         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1135                 struct hugepage_info *hpi;
1136
1137                 /*
1138                  * we don't yet mark hugepages as used at this stage, so
1139                  * we just map all hugepages available to the system
1140                  * all hugepages are still located on socket 0
1141                  */
1142                 hpi = &internal_config.hugepage_info[i];
1143
1144                 if (hpi->num_pages[0] == 0)
1145                         continue;
1146
1147                 /* map all hugepages available */
1148                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 1) < 0){
1149                         RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
1150                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1151                         goto fail;
1152                 }
1153
1154                 /* find physical addresses and sockets for each hugepage */
1155                 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0){
1156                         RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
1157                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1158                         goto fail;
1159                 }
1160
1161                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1162                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1163                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1164                         goto fail;
1165                 }
1166
1167                 if (sort_by_physaddr(&tmp_hp[hp_offset], hpi) < 0)
1168                         goto fail;
1169
1170 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1171                 /* remap all hugepages into single file segments */
1172                 new_pages_count[i] = remap_all_hugepages(&tmp_hp[hp_offset], hpi);
1173                 if (new_pages_count[i] < 0){
1174                         RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1175                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1176                         goto fail;
1177                 }
1178
1179                 /* we have processed a num of hugepages of this size, so inc offset */
1180                 hp_offset += new_pages_count[i];
1181 #else
1182                 /* remap all hugepages */
1183                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) < 0){
1184                         RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1185                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1186                         goto fail;
1187                 }
1188
1189                 /* unmap original mappings */
1190                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1191                         goto fail;
1192
1193                 /* we have processed a num of hugepages of this size, so inc offset */
1194                 hp_offset += hpi->num_pages[0];
1195 #endif
1196         }
1197
1198 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1199         nr_hugefiles = 0;
1200         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1201                 nr_hugefiles += new_pages_count[i];
1202         }
1203 #else
1204         nr_hugefiles = nr_hugepages;
1205 #endif
1206
1207
1208         /* clean out the numbers of pages */
1209         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1210                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1211                         internal_config.hugepage_info[i].num_pages[j] = 0;
1212
1213         /* get hugepages for each socket */
1214         for (i = 0; i < nr_hugefiles; i++) {
1215                 int socket = tmp_hp[i].socket_id;
1216
1217                 /* find a hugepage info with right size and increment num_pages */
1218                 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1219                                 (int)internal_config.num_hugepage_sizes);
1220                 for (j = 0; j < nb_hpsizes; j++) {
1221                         if (tmp_hp[i].size ==
1222                                         internal_config.hugepage_info[j].hugepage_sz) {
1223 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1224                                         internal_config.hugepage_info[j].num_pages[socket] +=
1225                                                 tmp_hp[i].repeated;
1226 #else
1227                                 internal_config.hugepage_info[j].num_pages[socket]++;
1228 #endif
1229                         }
1230                 }
1231         }
1232
1233         /* make a copy of socket_mem, needed for number of pages calculation */
1234         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1235                 memory[i] = internal_config.socket_mem[i];
1236
1237         /* calculate final number of pages */
1238         nr_hugepages = calc_num_pages_per_socket(memory,
1239                         internal_config.hugepage_info, used_hp,
1240                         internal_config.num_hugepage_sizes);
1241
1242         /* error if not enough memory available */
1243         if (nr_hugepages < 0)
1244                 goto fail;
1245
1246         /* reporting in! */
1247         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1248                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1249                         if (used_hp[i].num_pages[j] > 0) {
1250                                 RTE_LOG(DEBUG, EAL,
1251                                         "Requesting %u pages of size %uMB"
1252                                         " from socket %i\n",
1253                                         used_hp[i].num_pages[j],
1254                                         (unsigned)
1255                                         (used_hp[i].hugepage_sz / 0x100000),
1256                                         j);
1257                         }
1258                 }
1259         }
1260
1261         /* create shared memory */
1262         hugepage = create_shared_memory(eal_hugepage_info_path(),
1263                         nr_hugefiles * sizeof(struct hugepage_file));
1264
1265         if (hugepage == NULL) {
1266                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1267                 goto fail;
1268         }
1269         memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1270
1271         /*
1272          * unmap pages that we won't need (looks at used_hp).
1273          * also, sets final_va to NULL on pages that were unmapped.
1274          */
1275         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1276                         internal_config.num_hugepage_sizes) < 0) {
1277                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1278                 goto fail;
1279         }
1280
1281         /*
1282          * copy stuff from malloc'd hugepage* to the actual shared memory.
1283          * this procedure only copies those hugepages that have final_va
1284          * not NULL. has overflow protection.
1285          */
1286         if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1287                         tmp_hp, nr_hugefiles) < 0) {
1288                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1289                 goto fail;
1290         }
1291
1292         /* free the temporary hugepage table */
1293         free(tmp_hp);
1294         tmp_hp = NULL;
1295
1296         /* find earliest free memseg - this is needed because in case of IVSHMEM,
1297          * segments might have already been initialized */
1298         for (j = 0; j < RTE_MAX_MEMSEG; j++)
1299                 if (mcfg->memseg[j].addr == NULL) {
1300                         /* move to previous segment and exit loop */
1301                         j--;
1302                         break;
1303                 }
1304
1305         for (i = 0; i < nr_hugefiles; i++) {
1306                 new_memseg = 0;
1307
1308                 /* if this is a new section, create a new memseg */
1309                 if (i == 0)
1310                         new_memseg = 1;
1311                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1312                         new_memseg = 1;
1313                 else if (hugepage[i].size != hugepage[i-1].size)
1314                         new_memseg = 1;
1315
1316 #ifdef RTE_ARCH_PPC_64
1317                 /* On PPC64 architecture, the mmap always start from higher
1318                  * virtual address to lower address. Here, both the physical
1319                  * address and virtual address are in descending order */
1320                 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1321                     hugepage[i].size)
1322                         new_memseg = 1;
1323                 else if (((unsigned long)hugepage[i-1].final_va -
1324                     (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1325                         new_memseg = 1;
1326 #else
1327                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1328                     hugepage[i].size)
1329                         new_memseg = 1;
1330                 else if (((unsigned long)hugepage[i].final_va -
1331                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1332                         new_memseg = 1;
1333 #endif
1334
1335                 if (new_memseg) {
1336                         j += 1;
1337                         if (j == RTE_MAX_MEMSEG)
1338                                 break;
1339
1340                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1341                         mcfg->memseg[j].addr = hugepage[i].final_va;
1342 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1343                         mcfg->memseg[j].len = hugepage[i].size * hugepage[i].repeated;
1344 #else
1345                         mcfg->memseg[j].len = hugepage[i].size;
1346 #endif
1347                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1348                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1349                 }
1350                 /* continuation of previous memseg */
1351                 else {
1352 #ifdef RTE_ARCH_PPC_64
1353                 /* Use the phy and virt address of the last page as segment
1354                  * address for IBM Power architecture */
1355                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1356                         mcfg->memseg[j].addr = hugepage[i].final_va;
1357 #endif
1358                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1359                 }
1360                 hugepage[i].memseg_id = j;
1361         }
1362
1363         if (i < nr_hugefiles) {
1364                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1365                         "from %d requested\n"
1366                         "Current %s=%d is not enough\n"
1367                         "Please either increase it or request less amount "
1368                         "of memory.\n",
1369                         i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1370                         RTE_MAX_MEMSEG);
1371                 return -ENOMEM;
1372         }
1373
1374         return 0;
1375
1376 fail:
1377         if (tmp_hp)
1378                 free(tmp_hp);
1379         return -1;
1380 }
1381
1382 /*
1383  * uses fstat to report the size of a file on disk
1384  */
1385 static off_t
1386 getFileSize(int fd)
1387 {
1388         struct stat st;
1389         if (fstat(fd, &st) < 0)
1390                 return 0;
1391         return st.st_size;
1392 }
1393
1394 /*
1395  * This creates the memory mappings in the secondary process to match that of
1396  * the server process. It goes through each memory segment in the DPDK runtime
1397  * configuration and finds the hugepages which form that segment, mapping them
1398  * in order to form a contiguous block in the virtual memory space
1399  */
1400 int
1401 rte_eal_hugepage_attach(void)
1402 {
1403         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1404         const struct hugepage_file *hp = NULL;
1405         unsigned num_hp = 0;
1406         unsigned i, s = 0; /* s used to track the segment number */
1407         off_t size;
1408         int fd, fd_zero = -1, fd_hugepage = -1;
1409
1410         if (aslr_enabled() > 0) {
1411                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1412                                 "(ASLR) is enabled in the kernel.\n");
1413                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1414                                 "into secondary processes\n");
1415         }
1416
1417         test_proc_pagemap_readable();
1418
1419         if (internal_config.xen_dom0_support) {
1420 #ifdef RTE_LIBRTE_XEN_DOM0
1421                 if (rte_xen_dom0_memory_attach() < 0) {
1422                         RTE_LOG(ERR, EAL,"Failed to attach memory setments of primay "
1423                                         "process\n");
1424                         return -1;
1425                 }
1426                 return 0;
1427 #endif
1428         }
1429
1430         fd_zero = open("/dev/zero", O_RDONLY);
1431         if (fd_zero < 0) {
1432                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1433                 goto error;
1434         }
1435         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1436         if (fd_hugepage < 0) {
1437                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1438                 goto error;
1439         }
1440
1441         /* map all segments into memory to make sure we get the addrs */
1442         for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1443                 void *base_addr;
1444
1445                 /*
1446                  * the first memory segment with len==0 is the one that
1447                  * follows the last valid segment.
1448                  */
1449                 if (mcfg->memseg[s].len == 0)
1450                         break;
1451
1452 #ifdef RTE_LIBRTE_IVSHMEM
1453                 /*
1454                  * if segment has ioremap address set, it's an IVSHMEM segment and
1455                  * doesn't need mapping as it was already mapped earlier
1456                  */
1457                 if (mcfg->memseg[s].ioremap_addr != 0)
1458                         continue;
1459 #endif
1460
1461                 /*
1462                  * fdzero is mmapped to get a contiguous block of virtual
1463                  * addresses of the appropriate memseg size.
1464                  * use mmap to get identical addresses as the primary process.
1465                  */
1466                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1467                                  PROT_READ, MAP_PRIVATE, fd_zero, 0);
1468                 if (base_addr == MAP_FAILED ||
1469                     base_addr != mcfg->memseg[s].addr) {
1470                         RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1471                                 "in /dev/zero to requested address [%p]: '%s'\n",
1472                                 (unsigned long long)mcfg->memseg[s].len,
1473                                 mcfg->memseg[s].addr, strerror(errno));
1474                         if (aslr_enabled() > 0) {
1475                                 RTE_LOG(ERR, EAL, "It is recommended to "
1476                                         "disable ASLR in the kernel "
1477                                         "and retry running both primary "
1478                                         "and secondary processes\n");
1479                         }
1480                         goto error;
1481                 }
1482         }
1483
1484         size = getFileSize(fd_hugepage);
1485         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1486         if (hp == NULL) {
1487                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1488                 goto error;
1489         }
1490
1491         num_hp = size / sizeof(struct hugepage_file);
1492         RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1493
1494         s = 0;
1495         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1496                 void *addr, *base_addr;
1497                 uintptr_t offset = 0;
1498                 size_t mapping_size;
1499 #ifdef RTE_LIBRTE_IVSHMEM
1500                 /*
1501                  * if segment has ioremap address set, it's an IVSHMEM segment and
1502                  * doesn't need mapping as it was already mapped earlier
1503                  */
1504                 if (mcfg->memseg[s].ioremap_addr != 0) {
1505                         s++;
1506                         continue;
1507                 }
1508 #endif
1509                 /*
1510                  * free previously mapped memory so we can map the
1511                  * hugepages into the space
1512                  */
1513                 base_addr = mcfg->memseg[s].addr;
1514                 munmap(base_addr, mcfg->memseg[s].len);
1515
1516                 /* find the hugepages for this segment and map them
1517                  * we don't need to worry about order, as the server sorted the
1518                  * entries before it did the second mmap of them */
1519                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1520                         if (hp[i].memseg_id == (int)s){
1521                                 fd = open(hp[i].filepath, O_RDWR);
1522                                 if (fd < 0) {
1523                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1524                                                 hp[i].filepath);
1525                                         goto error;
1526                                 }
1527 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1528                                 mapping_size = hp[i].size * hp[i].repeated;
1529 #else
1530                                 mapping_size = hp[i].size;
1531 #endif
1532                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1533                                                 mapping_size, PROT_READ | PROT_WRITE,
1534                                                 MAP_SHARED, fd, 0);
1535                                 close(fd); /* close file both on success and on failure */
1536                                 if (addr == MAP_FAILED ||
1537                                                 addr != RTE_PTR_ADD(base_addr, offset)) {
1538                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1539                                                 hp[i].filepath);
1540                                         goto error;
1541                                 }
1542                                 offset+=mapping_size;
1543                         }
1544                 }
1545                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1546                                 (unsigned long long)mcfg->memseg[s].len);
1547                 s++;
1548         }
1549         /* unmap the hugepage config file, since we are done using it */
1550         munmap((void *)(uintptr_t)hp, size);
1551         close(fd_zero);
1552         close(fd_hugepage);
1553         return 0;
1554
1555 error:
1556         if (fd_zero >= 0)
1557                 close(fd_zero);
1558         if (fd_hugepage >= 0)
1559                 close(fd_hugepage);
1560         return -1;
1561 }