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 #ifdef RTE_LIBRTE_XEN_DOM0
 101 int is_xen_dom0_supported(void)
 102 {
 103         return internal_config.xen_dom0_support;
 104 }
 105 #endif
 106
 107 /**
 108  * @file
 109  * Huge page mapping under linux
 110  *
 111  * To reserve a big contiguous amount of memory, we use the hugepage
 112  * feature of linux. For that, we need to have hugetlbfs mounted. This
 113  * code will create many files in this directory (one per page) and
 114  * map them in virtual memory. For each page, we will retrieve its
 115  * physical address and remap it in order to have a virtual contiguous
 116  * zone as well as a physical contiguous zone.
 117  */
 118
 119 static uint64_t baseaddr_offset;
 120
 121 static unsigned proc_pagemap_readable;
 122
 123 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
 124
 125 static void
 126 test_proc_pagemap_readable(void)
 127 {
 128         int fd = open("/proc/self/pagemap", O_RDONLY);
 129
 130         if (fd < 0) {
 131                 RTE_LOG(ERR, EAL,
 132                         "Cannot open /proc/self/pagemap: %s. "
 133                         "virt2phys address translation will not work\n",
 134                         strerror(errno));
 135                 return;
 136         }
 137
 138         /* Is readable */
 139         close(fd);
 140         proc_pagemap_readable = 1;
 141 }
 142
 143 /* Lock page in physical memory and prevent from swapping. */
 144 int
 145 rte_mem_lock_page(const void *virt)
 146 {
 147         unsigned long virtual = (unsigned long)virt;
 148         int page_size = getpagesize();
 149         unsigned long aligned = (virtual & ~ (page_size - 1));
 150         return mlock((void*)aligned, page_size);
 151 }
 152
 153 /*
 154  * Get physical address of any mapped virtual address in the current process.
 155  */
 156 phys_addr_t
 157 rte_mem_virt2phy(const void *virtaddr)
 158 {
 159         int fd;
 160         uint64_t page, physaddr;
 161         unsigned long virt_pfn;
 162         int page_size;
 163         off_t offset;
 164
 165         /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
 166         if (!proc_pagemap_readable)
 167                 return RTE_BAD_PHYS_ADDR;
 168
 169         /* standard page size */
 170         page_size = getpagesize();
 171
 172         fd = open("/proc/self/pagemap", O_RDONLY);
 173         if (fd < 0) {
 174                 RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 175                         __func__, strerror(errno));
 176                 return RTE_BAD_PHYS_ADDR;
 177         }
 178
 179         virt_pfn = (unsigned long)virtaddr / page_size;
 180         offset = sizeof(uint64_t) * virt_pfn;
 181         if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 182                 RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 183                                 __func__, strerror(errno));
 184                 close(fd);
 185                 return RTE_BAD_PHYS_ADDR;
 186         }
 187         if (read(fd, &page, sizeof(uint64_t)) < 0) {
 188                 RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 189                                 __func__, strerror(errno));
 190                 close(fd);
 191                 return RTE_BAD_PHYS_ADDR;
 192         }
 193
 194         /*
 195          * the pfn (page frame number) are bits 0-54 (see
 196          * pagemap.txt in linux Documentation)
 197          */
 198         physaddr = ((page & 0x7fffffffffffffULL) * page_size)
 199                 + ((unsigned long)virtaddr % page_size);
 200         close(fd);
 201         return physaddr;
 202 }
 203
 204 /*
 205  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 206  * it by browsing the /proc/self/pagemap special file.
 207  */
 208 static int
 209 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 210 {
 211         unsigned i;
 212         phys_addr_t addr;
 213
 214         for (i = 0; i < hpi->num_pages[0]; i++) {
 215                 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
 216                 if (addr == RTE_BAD_PHYS_ADDR)
 217                         return -1;
 218                 hugepg_tbl[i].physaddr = addr;
 219         }
 220         return 0;
 221 }
 222
 223 /*
 224  * Check whether address-space layout randomization is enabled in
 225  * the kernel. This is important for multi-process as it can prevent
 226  * two processes mapping data to the same virtual address
 227  * Returns:
 228  *    0 - address space randomization disabled
 229  *    1/2 - address space randomization enabled
 230  *    negative error code on error
 231  */
 232 static int
 233 aslr_enabled(void)
 234 {
 235         char c;
 236         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 237         if (fd < 0)
 238                 return -errno;
 239         retval = read(fd, &c, 1);
 240         close(fd);
 241         if (retval < 0)
 242                 return -errno;
 243         if (retval == 0)
 244                 return -EIO;
 245         switch (c) {
 246                 case '0' : return 0;
 247                 case '1' : return 1;
 248                 case '2' : return 2;
 249                 default: return -EINVAL;
 250         }
 251 }
 252
 253 /*
 254  * Try to mmap *size bytes in /dev/zero. If it is successful, return the
 255  * pointer to the mmap'd area and keep *size unmodified. Else, retry
 256  * with a smaller zone: decrease *size by hugepage_sz until it reaches
 257  * 0. In this case, return NULL. Note: this function returns an address
 258  * which is a multiple of hugepage size.
 259  */
 260 static void *
 261 get_virtual_area(size_t *size, size_t hugepage_sz)
 262 {
 263         void *addr;
 264         int fd;
 265         long aligned_addr;
 266
 267         if (internal_config.base_virtaddr != 0) {
 268                 addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
 269                                 baseaddr_offset);
 270         }
 271         else addr = NULL;
 272
 273         RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
 274
 275         fd = open("/dev/zero", O_RDONLY);
 276         if (fd < 0){
 277                 RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
 278                 return NULL;
 279         }
 280         do {
 281                 addr = mmap(addr,
 282                                 (*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
 283                 if (addr == MAP_FAILED)
 284                         *size -= hugepage_sz;
 285         } while (addr == MAP_FAILED && *size > 0);
 286
 287         if (addr == MAP_FAILED) {
 288                 close(fd);
 289                 RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
 290                         strerror(errno));
 291                 return NULL;
 292         }
 293
 294         munmap(addr, (*size) + hugepage_sz);
 295         close(fd);
 296
 297         /* align addr to a huge page size boundary */
 298         aligned_addr = (long)addr;
 299         aligned_addr += (hugepage_sz - 1);
 300         aligned_addr &= (~(hugepage_sz - 1));
 301         addr = (void *)(aligned_addr);
 302
 303         RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
 304                 addr, *size);
 305
 306         /* increment offset */
 307         baseaddr_offset += *size;
 308
 309         return addr;
 310 }
 311
 312 /*
 313  * Mmap all hugepages of hugepage table: it first open a file in
 314  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 315  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 316  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 317  * map continguous physical blocks in contiguous virtual blocks.
 318  */
 319 static int
 320 map_all_hugepages(struct hugepage_file *hugepg_tbl,
 321                 struct hugepage_info *hpi, int orig)
 322 {
 323         int fd;
 324         unsigned i;
 325         void *virtaddr;
 326         void *vma_addr = NULL;
 327         size_t vma_len = 0;
 328
 329 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 330         RTE_SET_USED(vma_len);
 331 #endif
 332
 333         for (i = 0; i < hpi->num_pages[0]; i++) {
 334                 uint64_t hugepage_sz = hpi->hugepage_sz;
 335
 336                 if (orig) {
 337                         hugepg_tbl[i].file_id = i;
 338                         hugepg_tbl[i].size = hugepage_sz;
 339 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 340                         eal_get_hugefile_temp_path(hugepg_tbl[i].filepath,
 341                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 342                                         hugepg_tbl[i].file_id);
 343 #else
 344                         eal_get_hugefile_path(hugepg_tbl[i].filepath,
 345                                         sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
 346                                         hugepg_tbl[i].file_id);
 347 #endif
 348                         hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
 349                 }
 350 #ifndef RTE_ARCH_64
 351                 /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
 352                  * original map address as final map address.
 353                  */
 354                 else if ((hugepage_sz == RTE_PGSIZE_1G)
 355                         || (hugepage_sz == RTE_PGSIZE_16G)) {
 356                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 357                         hugepg_tbl[i].orig_va = NULL;
 358                         continue;
 359                 }
 360 #endif
 361
 362 #ifndef RTE_EAL_SINGLE_FILE_SEGMENTS
 363                 else if (vma_len == 0) {
 364                         unsigned j, num_pages;
 365
 366                         /* reserve a virtual area for next contiguous
 367                          * physical block: count the number of
 368                          * contiguous physical pages. */
 369                         for (j = i+1; j < hpi->num_pages[0] ; j++) {
 370 #ifdef RTE_ARCH_PPC_64
 371                                 /* The physical addresses are sorted in
 372                                  * descending order on PPC64 */
 373                                 if (hugepg_tbl[j].physaddr !=
 374                                     hugepg_tbl[j-1].physaddr - hugepage_sz)
 375                                         break;
 376 #else
 377                                 if (hugepg_tbl[j].physaddr !=
 378                                     hugepg_tbl[j-1].physaddr + hugepage_sz)
 379                                         break;
 380 #endif
 381                         }
 382                         num_pages = j - i;
 383                         vma_len = num_pages * hugepage_sz;
 384
 385                         /* get the biggest virtual memory area up to
 386                          * vma_len. If it fails, vma_addr is NULL, so
 387                          * let the kernel provide the address. */
 388                         vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 389                         if (vma_addr == NULL)
 390                                 vma_len = hugepage_sz;
 391                 }
 392 #endif
 393
 394                 /* try to create hugepage file */
 395                 fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
 396                 if (fd < 0) {
 397                         RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__,
 398                                         strerror(errno));
 399                         return -1;
 400                 }
 401
 402                 virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
 403                                 MAP_SHARED, fd, 0);
 404                 if (virtaddr == MAP_FAILED) {
 405                         RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__,
 406                                         strerror(errno));
 407                         close(fd);
 408                         return -1;
 409                 }
 410
 411                 if (orig) {
 412                         hugepg_tbl[i].orig_va = virtaddr;
 413                         memset(virtaddr, 0, hugepage_sz);
 414                 }
 415                 else {
 416                         hugepg_tbl[i].final_va = virtaddr;
 417                 }
 418
 419                 /* set shared flock on the file. */
 420                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 421                         RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
 422                                 __func__, strerror(errno));
 423                         close(fd);
 424                         return -1;
 425                 }
 426
 427                 close(fd);
 428
 429                 vma_addr = (char *)vma_addr + hugepage_sz;
 430                 vma_len -= hugepage_sz;
 431         }
 432         return 0;
 433 }
 434
 435 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 436
 437 /*
 438  * Remaps all hugepages into single file segments
 439  */
 440 static int
 441 remap_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 442 {
 443         int fd;
 444         unsigned i = 0, j, num_pages, page_idx = 0;
 445         void *vma_addr = NULL, *old_addr = NULL, *page_addr = NULL;
 446         size_t vma_len = 0;
 447         size_t hugepage_sz = hpi->hugepage_sz;
 448         size_t total_size, offset;
 449         char filepath[MAX_HUGEPAGE_PATH];
 450         phys_addr_t physaddr;
 451         int socket;
 452
 453         while (i < hpi->num_pages[0]) {
 454
 455 #ifndef RTE_ARCH_64
 456                 /* for 32-bit systems, don't remap 1G pages and 16G pages,
 457                  * just reuse original map address as final map address.
 458                  */
 459                 if ((hugepage_sz == RTE_PGSIZE_1G)
 460                         || (hugepage_sz == RTE_PGSIZE_16G)) {
 461                         hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
 462                         hugepg_tbl[i].orig_va = NULL;
 463                         i++;
 464                         continue;
 465                 }
 466 #endif
 467
 468                 /* reserve a virtual area for next contiguous
 469                  * physical block: count the number of
 470                  * contiguous physical pages. */
 471                 for (j = i+1; j < hpi->num_pages[0] ; j++) {
 472 #ifdef RTE_ARCH_PPC_64
 473                         /* The physical addresses are sorted in descending
 474                          * order on PPC64 */
 475                         if (hugepg_tbl[j].physaddr !=
 476                                 hugepg_tbl[j-1].physaddr - hugepage_sz)
 477                                 break;
 478 #else
 479                         if (hugepg_tbl[j].physaddr !=
 480                                 hugepg_tbl[j-1].physaddr + hugepage_sz)
 481                                 break;
 482 #endif
 483                 }
 484                 num_pages = j - i;
 485                 vma_len = num_pages * hugepage_sz;
 486
 487                 socket = hugepg_tbl[i].socket_id;
 488
 489                 /* get the biggest virtual memory area up to
 490                  * vma_len. If it fails, vma_addr is NULL, so
 491                  * let the kernel provide the address. */
 492                 vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
 493
 494                 /* If we can't find a big enough virtual area, work out how many pages
 495                  * we are going to get */
 496                 if (vma_addr == NULL)
 497                         j = i + 1;
 498                 else if (vma_len != num_pages * hugepage_sz) {
 499                         num_pages = vma_len / hugepage_sz;
 500                         j = i + num_pages;
 501
 502                 }
 503
 504                 hugepg_tbl[page_idx].file_id = page_idx;
 505                 eal_get_hugefile_path(filepath,
 506                                 sizeof(filepath),
 507                                 hpi->hugedir,
 508                                 hugepg_tbl[page_idx].file_id);
 509
 510                 /* try to create hugepage file */
 511                 fd = open(filepath, O_CREAT | O_RDWR, 0755);
 512                 if (fd < 0) {
 513                         RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__, strerror(errno));
 514                         return -1;
 515                 }
 516
 517                 total_size = 0;
 518                 for (;i < j; i++) {
 519
 520                         /* unmap current segment */
 521                         if (total_size > 0)
 522                                 munmap(vma_addr, total_size);
 523
 524                         /* unmap original page */
 525                         munmap(hugepg_tbl[i].orig_va, hugepage_sz);
 526                         unlink(hugepg_tbl[i].filepath);
 527
 528                         total_size += hugepage_sz;
 529
 530                         old_addr = vma_addr;
 531
 532                         /* map new, bigger segment */
 533                         vma_addr = mmap(vma_addr, total_size,
 534                                         PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 535
 536                         if (vma_addr == MAP_FAILED || vma_addr != old_addr) {
 537                                 RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__, strerror(errno));
 538                                 close(fd);
 539                                 return -1;
 540                         }
 541
 542                         /* touch the page. this is needed because kernel postpones mapping
 543                          * creation until the first page fault. with this, we pin down
 544                          * the page and it is marked as used and gets into process' pagemap.
 545                          */
 546                         for (offset = 0; offset < total_size; offset += hugepage_sz)
 547                                 *((volatile uint8_t*) RTE_PTR_ADD(vma_addr, offset));
 548                 }
 549
 550                 /* set shared flock on the file. */
 551                 if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
 552                         RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
 553                                 __func__, strerror(errno));
 554                         close(fd);
 555                         return -1;
 556                 }
 557
 558                 snprintf(hugepg_tbl[page_idx].filepath, MAX_HUGEPAGE_PATH, "%s",
 559                                 filepath);
 560
 561                 physaddr = rte_mem_virt2phy(vma_addr);
 562
 563                 if (physaddr == RTE_BAD_PHYS_ADDR)
 564                         return -1;
 565
 566                 hugepg_tbl[page_idx].final_va = vma_addr;
 567
 568                 hugepg_tbl[page_idx].physaddr = physaddr;
 569
 570                 hugepg_tbl[page_idx].repeated = num_pages;
 571
 572                 hugepg_tbl[page_idx].socket_id = socket;
 573
 574                 close(fd);
 575
 576                 /* verify the memory segment - that is, check that every VA corresponds
 577                  * to the physical address we expect to see
 578                  */
 579                 for (offset = 0; offset < vma_len; offset += hugepage_sz) {
 580                         uint64_t expected_physaddr;
 581
 582                         expected_physaddr = hugepg_tbl[page_idx].physaddr + offset;
 583                         page_addr = RTE_PTR_ADD(vma_addr, offset);
 584                         physaddr = rte_mem_virt2phy(page_addr);
 585
 586                         if (physaddr != expected_physaddr) {
 587                                 RTE_LOG(ERR, EAL, "Segment sanity check failed: wrong physaddr "
 588                                                 "at %p (offset 0x%" PRIx64 ": 0x%" PRIx64
 589                                                 " (expected 0x%" PRIx64 ")\n",
 590                                                 page_addr, offset, physaddr, expected_physaddr);
 591                                 return -1;
 592                         }
 593                 }
 594
 595                 /* zero out the whole segment */
 596                 memset(hugepg_tbl[page_idx].final_va, 0, total_size);
 597
 598                 page_idx++;
 599         }
 600
 601         /* zero out the rest */
 602         memset(&hugepg_tbl[page_idx], 0, (hpi->num_pages[0] - page_idx) * sizeof(struct hugepage_file));
 603         return page_idx;
 604 }
 605 #else/* RTE_EAL_SINGLE_FILE_SEGMENTS=n */
 606
 607 /* Unmap all hugepages from original mapping */
 608 static int
 609 unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 610 {
 611         unsigned i;
 612         for (i = 0; i < hpi->num_pages[0]; i++) {
 613                 if (hugepg_tbl[i].orig_va) {
 614                         munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
 615                         hugepg_tbl[i].orig_va = NULL;
 616                 }
 617         }
 618         return 0;
 619 }
 620 #endif /* RTE_EAL_SINGLE_FILE_SEGMENTS */
 621
 622 /*
 623  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 624  * page.
 625  */
 626 static int
 627 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 628 {
 629         int socket_id;
 630         char *end, *nodestr;
 631         unsigned i, hp_count = 0;
 632         uint64_t virt_addr;
 633         char buf[BUFSIZ];
 634         char hugedir_str[PATH_MAX];
 635         FILE *f;
 636
 637         f = fopen("/proc/self/numa_maps", "r");
 638         if (f == NULL) {
 639                 RTE_LOG(NOTICE, EAL, "cannot open /proc/self/numa_maps,"
 640                                 " consider that all memory is in socket_id 0\n");
 641                 return 0;
 642         }
 643
 644         snprintf(hugedir_str, sizeof(hugedir_str),
 645                         "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
 646
 647         /* parse numa map */
 648         while (fgets(buf, sizeof(buf), f) != NULL) {
 649
 650                 /* ignore non huge page */
 651                 if (strstr(buf, " huge ") == NULL &&
 652                                 strstr(buf, hugedir_str) == NULL)
 653                         continue;
 654
 655                 /* get zone addr */
 656                 virt_addr = strtoull(buf, &end, 16);
 657                 if (virt_addr == 0 || end == buf) {
 658                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 659                         goto error;
 660                 }
 661
 662                 /* get node id (socket id) */
 663                 nodestr = strstr(buf, " N");
 664                 if (nodestr == NULL) {
 665                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 666                         goto error;
 667                 }
 668                 nodestr += 2;
 669                 end = strstr(nodestr, "=");
 670                 if (end == NULL) {
 671                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 672                         goto error;
 673                 }
 674                 end[0] = '\0';
 675                 end = NULL;
 676
 677                 socket_id = strtoul(nodestr, &end, 0);
 678                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 679                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 680                         goto error;
 681                 }
 682
 683                 /* if we find this page in our mappings, set socket_id */
 684                 for (i = 0; i < hpi->num_pages[0]; i++) {
 685                         void *va = (void *)(unsigned long)virt_addr;
 686                         if (hugepg_tbl[i].orig_va == va) {
 687                                 hugepg_tbl[i].socket_id = socket_id;
 688                                 hp_count++;
 689                         }
 690                 }
 691         }
 692
 693         if (hp_count < hpi->num_pages[0])
 694                 goto error;
 695
 696         fclose(f);
 697         return 0;
 698
 699 error:
 700         fclose(f);
 701         return -1;
 702 }
 703
 704 /*
 705  * Sort the hugepg_tbl by physical address (lower addresses first on x86,
 706  * higher address first on powerpc). We use a slow algorithm, but we won't
 707  * have millions of pages, and this is only done at init time.
 708  */
 709 static int
 710 sort_by_physaddr(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 711 {
 712         unsigned i, j;
 713         int compare_idx;
 714         uint64_t compare_addr;
 715         struct hugepage_file tmp;
 716
 717         for (i = 0; i < hpi->num_pages[0]; i++) {
 718                 compare_addr = 0;
 719                 compare_idx = -1;
 720
 721                 /*
 722                  * browse all entries starting at 'i', and find the
 723                  * entry with the smallest addr
 724                  */
 725                 for (j=i; j< hpi->num_pages[0]; j++) {
 726
 727                         if (compare_addr == 0 ||
 728 #ifdef RTE_ARCH_PPC_64
 729                                 hugepg_tbl[j].physaddr > compare_addr) {
 730 #else
 731                                 hugepg_tbl[j].physaddr < compare_addr) {
 732 #endif
 733                                 compare_addr = hugepg_tbl[j].physaddr;
 734                                 compare_idx = j;
 735                         }
 736                 }
 737
 738                 /* should not happen */
 739                 if (compare_idx == -1) {
 740                         RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
 741                         return -1;
 742                 }
 743
 744                 /* swap the 2 entries in the table */
 745                 memcpy(&tmp, &hugepg_tbl[compare_idx],
 746                         sizeof(struct hugepage_file));
 747                 memcpy(&hugepg_tbl[compare_idx], &hugepg_tbl[i],
 748                         sizeof(struct hugepage_file));
 749                 memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage_file));
 750         }
 751         return 0;
 752 }
 753
 754 /*
 755  * Uses mmap to create a shared memory area for storage of data
 756  * Used in this file to store the hugepage file map on disk
 757  */
 758 static void *
 759 create_shared_memory(const char *filename, const size_t mem_size)
 760 {
 761         void *retval;
 762         int fd = open(filename, O_CREAT | O_RDWR, 0666);
 763         if (fd < 0)
 764                 return NULL;
 765         if (ftruncate(fd, mem_size) < 0) {
 766                 close(fd);
 767                 return NULL;
 768         }
 769         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 770         close(fd);
 771         return retval;
 772 }
 773
 774 /*
 775  * this copies *active* hugepages from one hugepage table to another.
 776  * destination is typically the shared memory.
 777  */
 778 static int
 779 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
 780                 const struct hugepage_file * src, int src_size)
 781 {
 782         int src_pos, dst_pos = 0;
 783
 784         for (src_pos = 0; src_pos < src_size; src_pos++) {
 785                 if (src[src_pos].final_va != NULL) {
 786                         /* error on overflow attempt */
 787                         if (dst_pos == dest_size)
 788                                 return -1;
 789                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
 790                         dst_pos++;
 791                 }
 792         }
 793         return 0;
 794 }
 795
 796 static int
 797 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
 798                 unsigned num_hp_info)
 799 {
 800         unsigned socket, size;
 801         int page, nrpages = 0;
 802
 803         /* get total number of hugepages */
 804         for (size = 0; size < num_hp_info; size++)
 805                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 806                         nrpages +=
 807                         internal_config.hugepage_info[size].num_pages[socket];
 808
 809         for (page = 0; page < nrpages; page++) {
 810                 struct hugepage_file *hp = &hugepg_tbl[page];
 811
 812                 if (hp->final_va != NULL && unlink(hp->filepath)) {
 813                         RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
 814                                 __func__, hp->filepath, strerror(errno));
 815                 }
 816         }
 817         return 0;
 818 }
 819
 820 /*
 821  * unmaps hugepages that are not going to be used. since we originally allocate
 822  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 823  */
 824 static int
 825 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
 826                 struct hugepage_info *hpi,
 827                 unsigned num_hp_info)
 828 {
 829         unsigned socket, size;
 830         int page, nrpages = 0;
 831
 832         /* get total number of hugepages */
 833         for (size = 0; size < num_hp_info; size++)
 834                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 835                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 836
 837         for (size = 0; size < num_hp_info; size++) {
 838                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 839                         unsigned pages_found = 0;
 840
 841                         /* traverse until we have unmapped all the unused pages */
 842                         for (page = 0; page < nrpages; page++) {
 843                                 struct hugepage_file *hp = &hugepg_tbl[page];
 844
 845 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 846                                 /* if this page was already cleared */
 847                                 if (hp->final_va == NULL)
 848                                         continue;
 849 #endif
 850
 851                                 /* find a page that matches the criteria */
 852                                 if ((hp->size == hpi[size].hugepage_sz) &&
 853                                                 (hp->socket_id == (int) socket)) {
 854
 855                                         /* if we skipped enough pages, unmap the rest */
 856                                         if (pages_found == hpi[size].num_pages[socket]) {
 857                                                 uint64_t unmap_len;
 858
 859 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 860                                                 unmap_len = hp->size * hp->repeated;
 861 #else
 862                                                 unmap_len = hp->size;
 863 #endif
 864
 865                                                 /* get start addr and len of the remaining segment */
 866                                                 munmap(hp->final_va, (size_t) unmap_len);
 867
 868                                                 hp->final_va = NULL;
 869                                                 if (unlink(hp->filepath) == -1) {
 870                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 871                                                                         __func__, hp->filepath, strerror(errno));
 872                                                         return -1;
 873                                                 }
 874                                         }
 875 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
 876                                         /* else, check how much do we need to map */
 877                                         else {
 878                                                 int nr_pg_left =
 879                                                                 hpi[size].num_pages[socket] - pages_found;
 880
 881                                                 /* if we need enough memory to fit into the segment */
 882                                                 if (hp->repeated <= nr_pg_left) {
 883                                                         pages_found += hp->repeated;
 884                                                 }
 885                                                 /* truncate the segment */
 886                                                 else {
 887                                                         uint64_t final_size = nr_pg_left * hp->size;
 888                                                         uint64_t seg_size = hp->repeated * hp->size;
 889
 890                                                         void * unmap_va = RTE_PTR_ADD(hp->final_va,
 891                                                                         final_size);
 892                                                         int fd;
 893
 894                                                         munmap(unmap_va, seg_size - final_size);
 895
 896                                                         fd = open(hp->filepath, O_RDWR);
 897                                                         if (fd < 0) {
 898                                                                 RTE_LOG(ERR, EAL, "Cannot open %s: %s\n",
 899                                                                                 hp->filepath, strerror(errno));
 900                                                                 return -1;
 901                                                         }
 902                                                         if (ftruncate(fd, final_size) < 0) {
 903                                                                 RTE_LOG(ERR, EAL, "Cannot truncate %s: %s\n",
 904                                                                                 hp->filepath, strerror(errno));
 905                                                                 return -1;
 906                                                         }
 907                                                         close(fd);
 908
 909                                                         pages_found += nr_pg_left;
 910                                                         hp->repeated = nr_pg_left;
 911                                                 }
 912                                         }
 913 #else
 914                                         /* else, lock the page and skip */
 915                                         else
 916                                                 pages_found++;
 917 #endif
 918
 919                                 } /* match page */
 920                         } /* foreach page */
 921                 } /* foreach socket */
 922         } /* foreach pagesize */
 923
 924         return 0;
 925 }
 926
 927 static inline uint64_t
 928 get_socket_mem_size(int socket)
 929 {
 930         uint64_t size = 0;
 931         unsigned i;
 932
 933         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
 934                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 935                 if (hpi->hugedir != NULL)
 936                         size += hpi->hugepage_sz * hpi->num_pages[socket];
 937         }
 938
 939         return size;
 940 }
 941
 942 /*
 943  * This function is a NUMA-aware equivalent of calc_num_pages.
 944  * It takes in the list of hugepage sizes and the
 945  * number of pages thereof, and calculates the best number of
 946  * pages of each size to fulfill the request for <memory> ram
 947  */
 948 static int
 949 calc_num_pages_per_socket(uint64_t * memory,
 950                 struct hugepage_info *hp_info,
 951                 struct hugepage_info *hp_used,
 952                 unsigned num_hp_info)
 953 {
 954         unsigned socket, j, i = 0;
 955         unsigned requested, available;
 956         int total_num_pages = 0;
 957         uint64_t remaining_mem, cur_mem;
 958         uint64_t total_mem = internal_config.memory;
 959
 960         if (num_hp_info == 0)
 961                 return -1;
 962
 963         /* if specific memory amounts per socket weren't requested */
 964         if (internal_config.force_sockets == 0) {
 965                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
 966                 size_t default_size, total_size;
 967                 unsigned lcore_id;
 968
 969                 /* Compute number of cores per socket */
 970                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
 971                 RTE_LCORE_FOREACH(lcore_id) {
 972                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
 973                 }
 974
 975                 /*
 976                  * Automatically spread requested memory amongst detected sockets according
 977                  * to number of cores from cpu mask present on each socket
 978                  */
 979                 total_size = internal_config.memory;
 980                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 981
 982                         /* Set memory amount per socket */
 983                         default_size = (internal_config.memory * cpu_per_socket[socket])
 984                                         / rte_lcore_count();
 985
 986                         /* Limit to maximum available memory on socket */
 987                         default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
 988
 989                         /* Update sizes */
 990                         memory[socket] = default_size;
 991                         total_size -= default_size;
 992                 }
 993
 994                 /*
 995                  * If some memory is remaining, try to allocate it by getting all
 996                  * available memory from sockets, one after the other
 997                  */
 998                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
 999                         /* take whatever is available */
1000                         default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
1001                                                total_size);
1002
1003                         /* Update sizes */
1004                         memory[socket] += default_size;
1005                         total_size -= default_size;
1006                 }
1007         }
1008
1009         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
1010                 /* skips if the memory on specific socket wasn't requested */
1011                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
1012                         hp_used[i].hugedir = hp_info[i].hugedir;
1013                         hp_used[i].num_pages[socket] = RTE_MIN(
1014                                         memory[socket] / hp_info[i].hugepage_sz,
1015                                         hp_info[i].num_pages[socket]);
1016
1017                         cur_mem = hp_used[i].num_pages[socket] *
1018                                         hp_used[i].hugepage_sz;
1019
1020                         memory[socket] -= cur_mem;
1021                         total_mem -= cur_mem;
1022
1023                         total_num_pages += hp_used[i].num_pages[socket];
1024
1025                         /* check if we have met all memory requests */
1026                         if (memory[socket] == 0)
1027                                 break;
1028
1029                         /* check if we have any more pages left at this size, if so
1030                          * move on to next size */
1031                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
1032                                 continue;
1033                         /* At this point we know that there are more pages available that are
1034                          * bigger than the memory we want, so lets see if we can get enough
1035                          * from other page sizes.
1036                          */
1037                         remaining_mem = 0;
1038                         for (j = i+1; j < num_hp_info; j++)
1039                                 remaining_mem += hp_info[j].hugepage_sz *
1040                                 hp_info[j].num_pages[socket];
1041
1042                         /* is there enough other memory, if not allocate another page and quit */
1043                         if (remaining_mem < memory[socket]){
1044                                 cur_mem = RTE_MIN(memory[socket],
1045                                                 hp_info[i].hugepage_sz);
1046                                 memory[socket] -= cur_mem;
1047                                 total_mem -= cur_mem;
1048                                 hp_used[i].num_pages[socket]++;
1049                                 total_num_pages++;
1050                                 break; /* we are done with this socket*/
1051                         }
1052                 }
1053                 /* if we didn't satisfy all memory requirements per socket */
1054                 if (memory[socket] > 0) {
1055                         /* to prevent icc errors */
1056                         requested = (unsigned) (internal_config.socket_mem[socket] /
1057                                         0x100000);
1058                         available = requested -
1059                                         ((unsigned) (memory[socket] / 0x100000));
1060                         RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
1061                                         "Requested: %uMB, available: %uMB\n", socket,
1062                                         requested, available);
1063                         return -1;
1064                 }
1065         }
1066
1067         /* if we didn't satisfy total memory requirements */
1068         if (total_mem > 0) {
1069                 requested = (unsigned) (internal_config.memory / 0x100000);
1070                 available = requested - (unsigned) (total_mem / 0x100000);
1071                 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
1072                                 " available: %uMB\n", requested, available);
1073                 return -1;
1074         }
1075         return total_num_pages;
1076 }
1077
1078 /*
1079  * Prepare physical memory mapping: fill configuration structure with
1080  * these infos, return 0 on success.
1081  *  1. map N huge pages in separate files in hugetlbfs
1082  *  2. find associated physical addr
1083  *  3. find associated NUMA socket ID
1084  *  4. sort all huge pages by physical address
1085  *  5. remap these N huge pages in the correct order
1086  *  6. unmap the first mapping
1087  *  7. fill memsegs in configuration with contiguous zones
1088  */
1089 int
1090 rte_eal_hugepage_init(void)
1091 {
1092         struct rte_mem_config *mcfg;
1093         struct hugepage_file *hugepage, *tmp_hp = NULL;
1094         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1095
1096         uint64_t memory[RTE_MAX_NUMA_NODES];
1097
1098         unsigned hp_offset;
1099         int i, j, new_memseg;
1100         int nr_hugefiles, nr_hugepages = 0;
1101         void *addr;
1102 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1103         int new_pages_count[MAX_HUGEPAGE_SIZES];
1104 #endif
1105
1106         test_proc_pagemap_readable();
1107
1108         memset(used_hp, 0, sizeof(used_hp));
1109
1110         /* get pointer to global configuration */
1111         mcfg = rte_eal_get_configuration()->mem_config;
1112
1113         /* hugetlbfs can be disabled */
1114         if (internal_config.no_hugetlbfs) {
1115                 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1116                                 MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1117                 if (addr == MAP_FAILED) {
1118                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1119                                         strerror(errno));
1120                         return -1;
1121                 }
1122                 mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
1123                 mcfg->memseg[0].addr = addr;
1124                 mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1125                 mcfg->memseg[0].len = internal_config.memory;
1126                 mcfg->memseg[0].socket_id = 0;
1127                 return 0;
1128         }
1129
1130 /* check if app runs on Xen Dom0 */
1131         if (internal_config.xen_dom0_support) {
1132 #ifdef RTE_LIBRTE_XEN_DOM0
1133                 /* use dom0_mm kernel driver to init memory */
1134                 if (rte_xen_dom0_memory_init() < 0)
1135                         return -1;
1136                 else
1137                         return 0;
1138 #endif
1139         }
1140
1141         /* calculate total number of hugepages available. at this point we haven't
1142          * yet started sorting them so they all are on socket 0 */
1143         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1144                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1145                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1146
1147                 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1148         }
1149
1150         /*
1151          * allocate a memory area for hugepage table.
1152          * this isn't shared memory yet. due to the fact that we need some
1153          * processing done on these pages, shared memory will be created
1154          * at a later stage.
1155          */
1156         tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1157         if (tmp_hp == NULL)
1158                 goto fail;
1159
1160         memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1161
1162         hp_offset = 0; /* where we start the current page size entries */
1163
1164         /* map all hugepages and sort them */
1165         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1166                 struct hugepage_info *hpi;
1167
1168                 /*
1169                  * we don't yet mark hugepages as used at this stage, so
1170                  * we just map all hugepages available to the system
1171                  * all hugepages are still located on socket 0
1172                  */
1173                 hpi = &internal_config.hugepage_info[i];
1174
1175                 if (hpi->num_pages[0] == 0)
1176                         continue;
1177
1178                 /* map all hugepages available */
1179                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 1) < 0){
1180                         RTE_LOG(DEBUG, EAL, "Failed to mmap %u MB hugepages\n",
1181                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1182                         goto fail;
1183                 }
1184
1185                 /* find physical addresses and sockets for each hugepage */
1186                 if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0){
1187                         RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
1188                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1189                         goto fail;
1190                 }
1191
1192                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1193                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1194                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1195                         goto fail;
1196                 }
1197
1198                 if (sort_by_physaddr(&tmp_hp[hp_offset], hpi) < 0)
1199                         goto fail;
1200
1201 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1202                 /* remap all hugepages into single file segments */
1203                 new_pages_count[i] = remap_all_hugepages(&tmp_hp[hp_offset], hpi);
1204                 if (new_pages_count[i] < 0){
1205                         RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1206                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1207                         goto fail;
1208                 }
1209
1210                 /* we have processed a num of hugepages of this size, so inc offset */
1211                 hp_offset += new_pages_count[i];
1212 #else
1213                 /* remap all hugepages */
1214                 if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) < 0){
1215                         RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1216                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1217                         goto fail;
1218                 }
1219
1220                 /* unmap original mappings */
1221                 if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1222                         goto fail;
1223
1224                 /* we have processed a num of hugepages of this size, so inc offset */
1225                 hp_offset += hpi->num_pages[0];
1226 #endif
1227         }
1228
1229 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1230         nr_hugefiles = 0;
1231         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1232                 nr_hugefiles += new_pages_count[i];
1233         }
1234 #else
1235         nr_hugefiles = nr_hugepages;
1236 #endif
1237
1238
1239         /* clean out the numbers of pages */
1240         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1241                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1242                         internal_config.hugepage_info[i].num_pages[j] = 0;
1243
1244         /* get hugepages for each socket */
1245         for (i = 0; i < nr_hugefiles; i++) {
1246                 int socket = tmp_hp[i].socket_id;
1247
1248                 /* find a hugepage info with right size and increment num_pages */
1249                 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1250                                 (int)internal_config.num_hugepage_sizes);
1251                 for (j = 0; j < nb_hpsizes; j++) {
1252                         if (tmp_hp[i].size ==
1253                                         internal_config.hugepage_info[j].hugepage_sz) {
1254 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1255                                         internal_config.hugepage_info[j].num_pages[socket] +=
1256                                                 tmp_hp[i].repeated;
1257 #else
1258                                 internal_config.hugepage_info[j].num_pages[socket]++;
1259 #endif
1260                         }
1261                 }
1262         }
1263
1264         /* make a copy of socket_mem, needed for number of pages calculation */
1265         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1266                 memory[i] = internal_config.socket_mem[i];
1267
1268         /* calculate final number of pages */
1269         nr_hugepages = calc_num_pages_per_socket(memory,
1270                         internal_config.hugepage_info, used_hp,
1271                         internal_config.num_hugepage_sizes);
1272
1273         /* error if not enough memory available */
1274         if (nr_hugepages < 0)
1275                 goto fail;
1276
1277         /* reporting in! */
1278         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1279                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1280                         if (used_hp[i].num_pages[j] > 0) {
1281                                 RTE_LOG(DEBUG, EAL,
1282                                         "Requesting %u pages of size %uMB"
1283                                         " from socket %i\n",
1284                                         used_hp[i].num_pages[j],
1285                                         (unsigned)
1286                                         (used_hp[i].hugepage_sz / 0x100000),
1287                                         j);
1288                         }
1289                 }
1290         }
1291
1292         /* create shared memory */
1293         hugepage = create_shared_memory(eal_hugepage_info_path(),
1294                         nr_hugefiles * sizeof(struct hugepage_file));
1295
1296         if (hugepage == NULL) {
1297                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1298                 goto fail;
1299         }
1300         memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1301
1302         /*
1303          * unmap pages that we won't need (looks at used_hp).
1304          * also, sets final_va to NULL on pages that were unmapped.
1305          */
1306         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1307                         internal_config.num_hugepage_sizes) < 0) {
1308                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1309                 goto fail;
1310         }
1311
1312         /*
1313          * copy stuff from malloc'd hugepage* to the actual shared memory.
1314          * this procedure only copies those hugepages that have final_va
1315          * not NULL. has overflow protection.
1316          */
1317         if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1318                         tmp_hp, nr_hugefiles) < 0) {
1319                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1320                 goto fail;
1321         }
1322
1323         /* free the hugepage backing files */
1324         if (internal_config.hugepage_unlink &&
1325                 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1326                 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1327                 goto fail;
1328         }
1329
1330         /* free the temporary hugepage table */
1331         free(tmp_hp);
1332         tmp_hp = NULL;
1333
1334         /* find earliest free memseg - this is needed because in case of IVSHMEM,
1335          * segments might have already been initialized */
1336         for (j = 0; j < RTE_MAX_MEMSEG; j++)
1337                 if (mcfg->memseg[j].addr == NULL) {
1338                         /* move to previous segment and exit loop */
1339                         j--;
1340                         break;
1341                 }
1342
1343         for (i = 0; i < nr_hugefiles; i++) {
1344                 new_memseg = 0;
1345
1346                 /* if this is a new section, create a new memseg */
1347                 if (i == 0)
1348                         new_memseg = 1;
1349                 else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1350                         new_memseg = 1;
1351                 else if (hugepage[i].size != hugepage[i-1].size)
1352                         new_memseg = 1;
1353
1354 #ifdef RTE_ARCH_PPC_64
1355                 /* On PPC64 architecture, the mmap always start from higher
1356                  * virtual address to lower address. Here, both the physical
1357                  * address and virtual address are in descending order */
1358                 else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1359                     hugepage[i].size)
1360                         new_memseg = 1;
1361                 else if (((unsigned long)hugepage[i-1].final_va -
1362                     (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1363                         new_memseg = 1;
1364 #else
1365                 else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1366                     hugepage[i].size)
1367                         new_memseg = 1;
1368                 else if (((unsigned long)hugepage[i].final_va -
1369                     (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1370                         new_memseg = 1;
1371 #endif
1372
1373                 if (new_memseg) {
1374                         j += 1;
1375                         if (j == RTE_MAX_MEMSEG)
1376                                 break;
1377
1378                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1379                         mcfg->memseg[j].addr = hugepage[i].final_va;
1380 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1381                         mcfg->memseg[j].len = hugepage[i].size * hugepage[i].repeated;
1382 #else
1383                         mcfg->memseg[j].len = hugepage[i].size;
1384 #endif
1385                         mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1386                         mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1387                 }
1388                 /* continuation of previous memseg */
1389                 else {
1390 #ifdef RTE_ARCH_PPC_64
1391                 /* Use the phy and virt address of the last page as segment
1392                  * address for IBM Power architecture */
1393                         mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1394                         mcfg->memseg[j].addr = hugepage[i].final_va;
1395 #endif
1396                         mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1397                 }
1398                 hugepage[i].memseg_id = j;
1399         }
1400
1401         if (i < nr_hugefiles) {
1402                 RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1403                         "from %d requested\n"
1404                         "Current %s=%d is not enough\n"
1405                         "Please either increase it or request less amount "
1406                         "of memory.\n",
1407                         i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1408                         RTE_MAX_MEMSEG);
1409                 return -ENOMEM;
1410         }
1411
1412         return 0;
1413
1414 fail:
1415         if (tmp_hp)
1416                 free(tmp_hp);
1417         return -1;
1418 }
1419
1420 /*
1421  * uses fstat to report the size of a file on disk
1422  */
1423 static off_t
1424 getFileSize(int fd)
1425 {
1426         struct stat st;
1427         if (fstat(fd, &st) < 0)
1428                 return 0;
1429         return st.st_size;
1430 }
1431
1432 /*
1433  * This creates the memory mappings in the secondary process to match that of
1434  * the server process. It goes through each memory segment in the DPDK runtime
1435  * configuration and finds the hugepages which form that segment, mapping them
1436  * in order to form a contiguous block in the virtual memory space
1437  */
1438 int
1439 rte_eal_hugepage_attach(void)
1440 {
1441         const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1442         const struct hugepage_file *hp = NULL;
1443         unsigned num_hp = 0;
1444         unsigned i, s = 0; /* s used to track the segment number */
1445         off_t size;
1446         int fd, fd_zero = -1, fd_hugepage = -1;
1447
1448         if (aslr_enabled() > 0) {
1449                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1450                                 "(ASLR) is enabled in the kernel.\n");
1451                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1452                                 "into secondary processes\n");
1453         }
1454
1455         test_proc_pagemap_readable();
1456
1457         if (internal_config.xen_dom0_support) {
1458 #ifdef RTE_LIBRTE_XEN_DOM0
1459                 if (rte_xen_dom0_memory_attach() < 0) {
1460                         RTE_LOG(ERR, EAL,"Failed to attach memory setments of primay "
1461                                         "process\n");
1462                         return -1;
1463                 }
1464                 return 0;
1465 #endif
1466         }
1467
1468         fd_zero = open("/dev/zero", O_RDONLY);
1469         if (fd_zero < 0) {
1470                 RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1471                 goto error;
1472         }
1473         fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1474         if (fd_hugepage < 0) {
1475                 RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1476                 goto error;
1477         }
1478
1479         /* map all segments into memory to make sure we get the addrs */
1480         for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1481                 void *base_addr;
1482
1483                 /*
1484                  * the first memory segment with len==0 is the one that
1485                  * follows the last valid segment.
1486                  */
1487                 if (mcfg->memseg[s].len == 0)
1488                         break;
1489
1490 #ifdef RTE_LIBRTE_IVSHMEM
1491                 /*
1492                  * if segment has ioremap address set, it's an IVSHMEM segment and
1493                  * doesn't need mapping as it was already mapped earlier
1494                  */
1495                 if (mcfg->memseg[s].ioremap_addr != 0)
1496                         continue;
1497 #endif
1498
1499                 /*
1500                  * fdzero is mmapped to get a contiguous block of virtual
1501                  * addresses of the appropriate memseg size.
1502                  * use mmap to get identical addresses as the primary process.
1503                  */
1504                 base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1505                                  PROT_READ, MAP_PRIVATE, fd_zero, 0);
1506                 if (base_addr == MAP_FAILED ||
1507                     base_addr != mcfg->memseg[s].addr) {
1508                         RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1509                                 "in /dev/zero to requested address [%p]: '%s'\n",
1510                                 (unsigned long long)mcfg->memseg[s].len,
1511                                 mcfg->memseg[s].addr, strerror(errno));
1512                         if (aslr_enabled() > 0) {
1513                                 RTE_LOG(ERR, EAL, "It is recommended to "
1514                                         "disable ASLR in the kernel "
1515                                         "and retry running both primary "
1516                                         "and secondary processes\n");
1517                         }
1518                         goto error;
1519                 }
1520         }
1521
1522         size = getFileSize(fd_hugepage);
1523         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1524         if (hp == NULL) {
1525                 RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1526                 goto error;
1527         }
1528
1529         num_hp = size / sizeof(struct hugepage_file);
1530         RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1531
1532         s = 0;
1533         while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1534                 void *addr, *base_addr;
1535                 uintptr_t offset = 0;
1536                 size_t mapping_size;
1537 #ifdef RTE_LIBRTE_IVSHMEM
1538                 /*
1539                  * if segment has ioremap address set, it's an IVSHMEM segment and
1540                  * doesn't need mapping as it was already mapped earlier
1541                  */
1542                 if (mcfg->memseg[s].ioremap_addr != 0) {
1543                         s++;
1544                         continue;
1545                 }
1546 #endif
1547                 /*
1548                  * free previously mapped memory so we can map the
1549                  * hugepages into the space
1550                  */
1551                 base_addr = mcfg->memseg[s].addr;
1552                 munmap(base_addr, mcfg->memseg[s].len);
1553
1554                 /* find the hugepages for this segment and map them
1555                  * we don't need to worry about order, as the server sorted the
1556                  * entries before it did the second mmap of them */
1557                 for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1558                         if (hp[i].memseg_id == (int)s){
1559                                 fd = open(hp[i].filepath, O_RDWR);
1560                                 if (fd < 0) {
1561                                         RTE_LOG(ERR, EAL, "Could not open %s\n",
1562                                                 hp[i].filepath);
1563                                         goto error;
1564                                 }
1565 #ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1566                                 mapping_size = hp[i].size * hp[i].repeated;
1567 #else
1568                                 mapping_size = hp[i].size;
1569 #endif
1570                                 addr = mmap(RTE_PTR_ADD(base_addr, offset),
1571                                                 mapping_size, PROT_READ | PROT_WRITE,
1572                                                 MAP_SHARED, fd, 0);
1573                                 close(fd); /* close file both on success and on failure */
1574                                 if (addr == MAP_FAILED ||
1575                                                 addr != RTE_PTR_ADD(base_addr, offset)) {
1576                                         RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1577                                                 hp[i].filepath);
1578                                         goto error;
1579                                 }
1580                                 offset+=mapping_size;
1581                         }
1582                 }
1583                 RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1584                                 (unsigned long long)mcfg->memseg[s].len);
1585                 s++;
1586         }
1587         /* unmap the hugepage config file, since we are done using it */
1588         munmap((void *)(uintptr_t)hp, size);
1589         close(fd_zero);
1590         close(fd_hugepage);
1591         return 0;
1592
1593 error:
1594         if (fd_zero >= 0)
1595                 close(fd_zero);
1596         if (fd_hugepage >= 0)
1597                 close(fd_hugepage);
1598         return -1;
1599 }