lib/librte_eal/linux/eal_memory.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2014 Intel Corporation.
   3  * Copyright(c) 2013 6WIND S.A.
   4  */
   5
   6 #include <errno.h>
   7 #include <fcntl.h>
   8 #include <stdarg.h>
   9 #include <stdbool.h>
  10 #include <stdlib.h>
  11 #include <stdio.h>
  12 #include <stdint.h>
  13 #include <inttypes.h>
  14 #include <string.h>
  15 #include <sys/mman.h>
  16 #include <sys/types.h>
  17 #include <sys/stat.h>
  18 #include <sys/queue.h>
  19 #include <sys/file.h>
  20 #include <sys/resource.h>
  21 #include <unistd.h>
  22 #include <limits.h>
  23 #include <sys/ioctl.h>
  24 #include <sys/time.h>
  25 #include <signal.h>
  26 #include <setjmp.h>
  27 #ifdef F_ADD_SEALS /* if file sealing is supported, so is memfd */
  28 #include <linux/memfd.h>
  29 #define MEMFD_SUPPORTED
  30 #endif
  31 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
  32 #include <numa.h>
  33 #include <numaif.h>
  34 #endif
  35
  36 #include <rte_errno.h>
  37 #include <rte_log.h>
  38 #include <rte_memory.h>
  39 #include <rte_launch.h>
  40 #include <rte_eal.h>
  41 #include <rte_per_lcore.h>
  42 #include <rte_lcore.h>
  43 #include <rte_common.h>
  44 #include <rte_string_fns.h>
  45
  46 #include "eal_private.h"
  47 #include "eal_memalloc.h"
  48 #include "eal_memcfg.h"
  49 #include "eal_internal_cfg.h"
  50 #include "eal_filesystem.h"
  51 #include "eal_hugepages.h"
  52 #include "eal_options.h"
  53
  54 #define PFN_MASK_SIZE   8
  55
  56 /**
  57  * @file
  58  * Huge page mapping under linux
  59  *
  60  * To reserve a big contiguous amount of memory, we use the hugepage
  61  * feature of linux. For that, we need to have hugetlbfs mounted. This
  62  * code will create many files in this directory (one per page) and
  63  * map them in virtual memory. For each page, we will retrieve its
  64  * physical address and remap it in order to have a virtual contiguous
  65  * zone as well as a physical contiguous zone.
  66  */
  67
  68 static int phys_addrs_available = -1;
  69
  70 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
  71
  72 uint64_t eal_get_baseaddr(void)
  73 {
  74         /*
  75          * Linux kernel uses a really high address as starting address for
  76          * serving mmaps calls. If there exists addressing limitations and IOVA
  77          * mode is VA, this starting address is likely too high for those
  78          * devices. However, it is possible to use a lower address in the
  79          * process virtual address space as with 64 bits there is a lot of
  80          * available space.
  81          *
  82          * Current known limitations are 39 or 40 bits. Setting the starting
  83          * address at 4GB implies there are 508GB or 1020GB for mapping the
  84          * available hugepages. This is likely enough for most systems, although
  85          * a device with addressing limitations should call
  86          * rte_mem_check_dma_mask for ensuring all memory is within supported
  87          * range.
  88          */
  89         return 0x100000000ULL;
  90 }
  91
  92 /*
  93  * Get physical address of any mapped virtual address in the current process.
  94  */
  95 phys_addr_t
  96 rte_mem_virt2phy(const void *virtaddr)
  97 {
  98         int fd, retval;
  99         uint64_t page, physaddr;
 100         unsigned long virt_pfn;
 101         int page_size;
 102         off_t offset;
 103
 104         if (phys_addrs_available == 0)
 105                 return RTE_BAD_IOVA;
 106
 107         /* standard page size */
 108         page_size = getpagesize();
 109
 110         fd = open("/proc/self/pagemap", O_RDONLY);
 111         if (fd < 0) {
 112                 RTE_LOG(INFO, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
 113                         __func__, strerror(errno));
 114                 return RTE_BAD_IOVA;
 115         }
 116
 117         virt_pfn = (unsigned long)virtaddr / page_size;
 118         offset = sizeof(uint64_t) * virt_pfn;
 119         if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
 120                 RTE_LOG(INFO, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
 121                                 __func__, strerror(errno));
 122                 close(fd);
 123                 return RTE_BAD_IOVA;
 124         }
 125
 126         retval = read(fd, &page, PFN_MASK_SIZE);
 127         close(fd);
 128         if (retval < 0) {
 129                 RTE_LOG(INFO, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
 130                                 __func__, strerror(errno));
 131                 return RTE_BAD_IOVA;
 132         } else if (retval != PFN_MASK_SIZE) {
 133                 RTE_LOG(INFO, EAL, "%s(): read %d bytes from /proc/self/pagemap "
 134                                 "but expected %d:\n",
 135                                 __func__, retval, PFN_MASK_SIZE);
 136                 return RTE_BAD_IOVA;
 137         }
 138
 139         /*
 140          * the pfn (page frame number) are bits 0-54 (see
 141          * pagemap.txt in linux Documentation)
 142          */
 143         if ((page & 0x7fffffffffffffULL) == 0)
 144                 return RTE_BAD_IOVA;
 145
 146         physaddr = ((page & 0x7fffffffffffffULL) * page_size)
 147                 + ((unsigned long)virtaddr % page_size);
 148
 149         return physaddr;
 150 }
 151
 152 rte_iova_t
 153 rte_mem_virt2iova(const void *virtaddr)
 154 {
 155         if (rte_eal_iova_mode() == RTE_IOVA_VA)
 156                 return (uintptr_t)virtaddr;
 157         return rte_mem_virt2phy(virtaddr);
 158 }
 159
 160 /*
 161  * For each hugepage in hugepg_tbl, fill the physaddr value. We find
 162  * it by browsing the /proc/self/pagemap special file.
 163  */
 164 static int
 165 find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 166 {
 167         unsigned int i;
 168         phys_addr_t addr;
 169
 170         for (i = 0; i < hpi->num_pages[0]; i++) {
 171                 addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
 172                 if (addr == RTE_BAD_PHYS_ADDR)
 173                         return -1;
 174                 hugepg_tbl[i].physaddr = addr;
 175         }
 176         return 0;
 177 }
 178
 179 /*
 180  * For each hugepage in hugepg_tbl, fill the physaddr value sequentially.
 181  */
 182 static int
 183 set_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 184 {
 185         unsigned int i;
 186         static phys_addr_t addr;
 187
 188         for (i = 0; i < hpi->num_pages[0]; i++) {
 189                 hugepg_tbl[i].physaddr = addr;
 190                 addr += hugepg_tbl[i].size;
 191         }
 192         return 0;
 193 }
 194
 195 /*
 196  * Check whether address-space layout randomization is enabled in
 197  * the kernel. This is important for multi-process as it can prevent
 198  * two processes mapping data to the same virtual address
 199  * Returns:
 200  *    0 - address space randomization disabled
 201  *    1/2 - address space randomization enabled
 202  *    negative error code on error
 203  */
 204 static int
 205 aslr_enabled(void)
 206 {
 207         char c;
 208         int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
 209         if (fd < 0)
 210                 return -errno;
 211         retval = read(fd, &c, 1);
 212         close(fd);
 213         if (retval < 0)
 214                 return -errno;
 215         if (retval == 0)
 216                 return -EIO;
 217         switch (c) {
 218                 case '0' : return 0;
 219                 case '1' : return 1;
 220                 case '2' : return 2;
 221                 default: return -EINVAL;
 222         }
 223 }
 224
 225 static sigjmp_buf huge_jmpenv;
 226
 227 static void huge_sigbus_handler(int signo __rte_unused)
 228 {
 229         siglongjmp(huge_jmpenv, 1);
 230 }
 231
 232 /* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
 233  * non-static local variable in the stack frame calling sigsetjmp might be
 234  * clobbered by a call to longjmp.
 235  */
 236 static int huge_wrap_sigsetjmp(void)
 237 {
 238         return sigsetjmp(huge_jmpenv, 1);
 239 }
 240
 241 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 242 /* Callback for numa library. */
 243 void numa_error(char *where)
 244 {
 245         RTE_LOG(ERR, EAL, "%s failed: %s\n", where, strerror(errno));
 246 }
 247 #endif
 248
 249 /*
 250  * Mmap all hugepages of hugepage table: it first open a file in
 251  * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
 252  * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
 253  * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
 254  * map contiguous physical blocks in contiguous virtual blocks.
 255  */
 256 static unsigned
 257 map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi,
 258                   uint64_t *essential_memory __rte_unused)
 259 {
 260         int fd;
 261         unsigned i;
 262         void *virtaddr;
 263 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 264         int node_id = -1;
 265         int essential_prev = 0;
 266         int oldpolicy;
 267         struct bitmask *oldmask = NULL;
 268         bool have_numa = true;
 269         unsigned long maxnode = 0;
 270
 271         /* Check if kernel supports NUMA. */
 272         if (numa_available() != 0) {
 273                 RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n");
 274                 have_numa = false;
 275         }
 276
 277         if (have_numa) {
 278                 RTE_LOG(DEBUG, EAL, "Trying to obtain current memory policy.\n");
 279                 oldmask = numa_allocate_nodemask();
 280                 if (get_mempolicy(&oldpolicy, oldmask->maskp,
 281                                   oldmask->size + 1, 0, 0) < 0) {
 282                         RTE_LOG(ERR, EAL,
 283                                 "Failed to get current mempolicy: %s. "
 284                                 "Assuming MPOL_DEFAULT.\n", strerror(errno));
 285                         oldpolicy = MPOL_DEFAULT;
 286                 }
 287                 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 288                         if (internal_config.socket_mem[i])
 289                                 maxnode = i + 1;
 290         }
 291 #endif
 292
 293         for (i = 0; i < hpi->num_pages[0]; i++) {
 294                 struct hugepage_file *hf = &hugepg_tbl[i];
 295                 uint64_t hugepage_sz = hpi->hugepage_sz;
 296
 297 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 298                 if (maxnode) {
 299                         unsigned int j;
 300
 301                         for (j = 0; j < maxnode; j++)
 302                                 if (essential_memory[j])
 303                                         break;
 304
 305                         if (j == maxnode) {
 306                                 node_id = (node_id + 1) % maxnode;
 307                                 while (!internal_config.socket_mem[node_id]) {
 308                                         node_id++;
 309                                         node_id %= maxnode;
 310                                 }
 311                                 essential_prev = 0;
 312                         } else {
 313                                 node_id = j;
 314                                 essential_prev = essential_memory[j];
 315
 316                                 if (essential_memory[j] < hugepage_sz)
 317                                         essential_memory[j] = 0;
 318                                 else
 319                                         essential_memory[j] -= hugepage_sz;
 320                         }
 321
 322                         RTE_LOG(DEBUG, EAL,
 323                                 "Setting policy MPOL_PREFERRED for socket %d\n",
 324                                 node_id);
 325                         numa_set_preferred(node_id);
 326                 }
 327 #endif
 328
 329                 hf->file_id = i;
 330                 hf->size = hugepage_sz;
 331                 eal_get_hugefile_path(hf->filepath, sizeof(hf->filepath),
 332                                 hpi->hugedir, hf->file_id);
 333                 hf->filepath[sizeof(hf->filepath) - 1] = '\0';
 334
 335                 /* try to create hugepage file */
 336                 fd = open(hf->filepath, O_CREAT | O_RDWR, 0600);
 337                 if (fd < 0) {
 338                         RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
 339                                         strerror(errno));
 340                         goto out;
 341                 }
 342
 343                 /* map the segment, and populate page tables,
 344                  * the kernel fills this segment with zeros. we don't care where
 345                  * this gets mapped - we already have contiguous memory areas
 346                  * ready for us to map into.
 347                  */
 348                 virtaddr = mmap(NULL, hugepage_sz, PROT_READ | PROT_WRITE,
 349                                 MAP_SHARED | MAP_POPULATE, fd, 0);
 350                 if (virtaddr == MAP_FAILED) {
 351                         RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
 352                                         strerror(errno));
 353                         close(fd);
 354                         goto out;
 355                 }
 356
 357                 hf->orig_va = virtaddr;
 358
 359                 /* In linux, hugetlb limitations, like cgroup, are
 360                  * enforced at fault time instead of mmap(), even
 361                  * with the option of MAP_POPULATE. Kernel will send
 362                  * a SIGBUS signal. To avoid to be killed, save stack
 363                  * environment here, if SIGBUS happens, we can jump
 364                  * back here.
 365                  */
 366                 if (huge_wrap_sigsetjmp()) {
 367                         RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
 368                                 "hugepages of size %u MB\n",
 369                                 (unsigned int)(hugepage_sz / 0x100000));
 370                         munmap(virtaddr, hugepage_sz);
 371                         close(fd);
 372                         unlink(hugepg_tbl[i].filepath);
 373 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 374                         if (maxnode)
 375                                 essential_memory[node_id] =
 376                                         essential_prev;
 377 #endif
 378                         goto out;
 379                 }
 380                 *(int *)virtaddr = 0;
 381
 382                 /* set shared lock on the file. */
 383                 if (flock(fd, LOCK_SH) < 0) {
 384                         RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
 385                                 __func__, strerror(errno));
 386                         close(fd);
 387                         goto out;
 388                 }
 389
 390                 close(fd);
 391         }
 392
 393 out:
 394 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 395         if (maxnode) {
 396                 RTE_LOG(DEBUG, EAL,
 397                         "Restoring previous memory policy: %d\n", oldpolicy);
 398                 if (oldpolicy == MPOL_DEFAULT) {
 399                         numa_set_localalloc();
 400                 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
 401                                          oldmask->size + 1) < 0) {
 402                         RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
 403                                 strerror(errno));
 404                         numa_set_localalloc();
 405                 }
 406         }
 407         if (oldmask != NULL)
 408                 numa_free_cpumask(oldmask);
 409 #endif
 410         return i;
 411 }
 412
 413 /*
 414  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 415  * page.
 416  */
 417 static int
 418 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 419 {
 420         int socket_id;
 421         char *end, *nodestr;
 422         unsigned i, hp_count = 0;
 423         uint64_t virt_addr;
 424         char buf[BUFSIZ];
 425         char hugedir_str[PATH_MAX];
 426         FILE *f;
 427
 428         f = fopen("/proc/self/numa_maps", "r");
 429         if (f == NULL) {
 430                 RTE_LOG(NOTICE, EAL, "NUMA support not available"
 431                         " consider that all memory is in socket_id 0\n");
 432                 return 0;
 433         }
 434
 435         snprintf(hugedir_str, sizeof(hugedir_str),
 436                         "%s/%s", hpi->hugedir, eal_get_hugefile_prefix());
 437
 438         /* parse numa map */
 439         while (fgets(buf, sizeof(buf), f) != NULL) {
 440
 441                 /* ignore non huge page */
 442                 if (strstr(buf, " huge ") == NULL &&
 443                                 strstr(buf, hugedir_str) == NULL)
 444                         continue;
 445
 446                 /* get zone addr */
 447                 virt_addr = strtoull(buf, &end, 16);
 448                 if (virt_addr == 0 || end == buf) {
 449                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 450                         goto error;
 451                 }
 452
 453                 /* get node id (socket id) */
 454                 nodestr = strstr(buf, " N");
 455                 if (nodestr == NULL) {
 456                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 457                         goto error;
 458                 }
 459                 nodestr += 2;
 460                 end = strstr(nodestr, "=");
 461                 if (end == NULL) {
 462                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 463                         goto error;
 464                 }
 465                 end[0] = '\0';
 466                 end = NULL;
 467
 468                 socket_id = strtoul(nodestr, &end, 0);
 469                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 470                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 471                         goto error;
 472                 }
 473
 474                 /* if we find this page in our mappings, set socket_id */
 475                 for (i = 0; i < hpi->num_pages[0]; i++) {
 476                         void *va = (void *)(unsigned long)virt_addr;
 477                         if (hugepg_tbl[i].orig_va == va) {
 478                                 hugepg_tbl[i].socket_id = socket_id;
 479                                 hp_count++;
 480 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 481                                 RTE_LOG(DEBUG, EAL,
 482                                         "Hugepage %s is on socket %d\n",
 483                                         hugepg_tbl[i].filepath, socket_id);
 484 #endif
 485                         }
 486                 }
 487         }
 488
 489         if (hp_count < hpi->num_pages[0])
 490                 goto error;
 491
 492         fclose(f);
 493         return 0;
 494
 495 error:
 496         fclose(f);
 497         return -1;
 498 }
 499
 500 static int
 501 cmp_physaddr(const void *a, const void *b)
 502 {
 503 #ifndef RTE_ARCH_PPC_64
 504         const struct hugepage_file *p1 = a;
 505         const struct hugepage_file *p2 = b;
 506 #else
 507         /* PowerPC needs memory sorted in reverse order from x86 */
 508         const struct hugepage_file *p1 = b;
 509         const struct hugepage_file *p2 = a;
 510 #endif
 511         if (p1->physaddr < p2->physaddr)
 512                 return -1;
 513         else if (p1->physaddr > p2->physaddr)
 514                 return 1;
 515         else
 516                 return 0;
 517 }
 518
 519 /*
 520  * Uses mmap to create a shared memory area for storage of data
 521  * Used in this file to store the hugepage file map on disk
 522  */
 523 static void *
 524 create_shared_memory(const char *filename, const size_t mem_size)
 525 {
 526         void *retval;
 527         int fd;
 528
 529         /* if no shared files mode is used, create anonymous memory instead */
 530         if (internal_config.no_shconf) {
 531                 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE,
 532                                 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 533                 if (retval == MAP_FAILED)
 534                         return NULL;
 535                 return retval;
 536         }
 537
 538         fd = open(filename, O_CREAT | O_RDWR, 0600);
 539         if (fd < 0)
 540                 return NULL;
 541         if (ftruncate(fd, mem_size) < 0) {
 542                 close(fd);
 543                 return NULL;
 544         }
 545         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 546         close(fd);
 547         if (retval == MAP_FAILED)
 548                 return NULL;
 549         return retval;
 550 }
 551
 552 /*
 553  * this copies *active* hugepages from one hugepage table to another.
 554  * destination is typically the shared memory.
 555  */
 556 static int
 557 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
 558                 const struct hugepage_file * src, int src_size)
 559 {
 560         int src_pos, dst_pos = 0;
 561
 562         for (src_pos = 0; src_pos < src_size; src_pos++) {
 563                 if (src[src_pos].orig_va != NULL) {
 564                         /* error on overflow attempt */
 565                         if (dst_pos == dest_size)
 566                                 return -1;
 567                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
 568                         dst_pos++;
 569                 }
 570         }
 571         return 0;
 572 }
 573
 574 static int
 575 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
 576                 unsigned num_hp_info)
 577 {
 578         unsigned socket, size;
 579         int page, nrpages = 0;
 580
 581         /* get total number of hugepages */
 582         for (size = 0; size < num_hp_info; size++)
 583                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 584                         nrpages +=
 585                         internal_config.hugepage_info[size].num_pages[socket];
 586
 587         for (page = 0; page < nrpages; page++) {
 588                 struct hugepage_file *hp = &hugepg_tbl[page];
 589
 590                 if (hp->orig_va != NULL && unlink(hp->filepath)) {
 591                         RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
 592                                 __func__, hp->filepath, strerror(errno));
 593                 }
 594         }
 595         return 0;
 596 }
 597
 598 /*
 599  * unmaps hugepages that are not going to be used. since we originally allocate
 600  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 601  */
 602 static int
 603 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
 604                 struct hugepage_info *hpi,
 605                 unsigned num_hp_info)
 606 {
 607         unsigned socket, size;
 608         int page, nrpages = 0;
 609
 610         /* get total number of hugepages */
 611         for (size = 0; size < num_hp_info; size++)
 612                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 613                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 614
 615         for (size = 0; size < num_hp_info; size++) {
 616                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 617                         unsigned pages_found = 0;
 618
 619                         /* traverse until we have unmapped all the unused pages */
 620                         for (page = 0; page < nrpages; page++) {
 621                                 struct hugepage_file *hp = &hugepg_tbl[page];
 622
 623                                 /* find a page that matches the criteria */
 624                                 if ((hp->size == hpi[size].hugepage_sz) &&
 625                                                 (hp->socket_id == (int) socket)) {
 626
 627                                         /* if we skipped enough pages, unmap the rest */
 628                                         if (pages_found == hpi[size].num_pages[socket]) {
 629                                                 uint64_t unmap_len;
 630
 631                                                 unmap_len = hp->size;
 632
 633                                                 /* get start addr and len of the remaining segment */
 634                                                 munmap(hp->orig_va,
 635                                                         (size_t)unmap_len);
 636
 637                                                 hp->orig_va = NULL;
 638                                                 if (unlink(hp->filepath) == -1) {
 639                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 640                                                                         __func__, hp->filepath, strerror(errno));
 641                                                         return -1;
 642                                                 }
 643                                         } else {
 644                                                 /* lock the page and skip */
 645                                                 pages_found++;
 646                                         }
 647
 648                                 } /* match page */
 649                         } /* foreach page */
 650                 } /* foreach socket */
 651         } /* foreach pagesize */
 652
 653         return 0;
 654 }
 655
 656 static int
 657 remap_segment(struct hugepage_file *hugepages, int seg_start, int seg_end)
 658 {
 659         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 660         struct rte_memseg_list *msl;
 661         struct rte_fbarray *arr;
 662         int cur_page, seg_len;
 663         unsigned int msl_idx;
 664         int ms_idx;
 665         uint64_t page_sz;
 666         size_t memseg_len;
 667         int socket_id;
 668
 669         page_sz = hugepages[seg_start].size;
 670         socket_id = hugepages[seg_start].socket_id;
 671         seg_len = seg_end - seg_start;
 672
 673         RTE_LOG(DEBUG, EAL, "Attempting to map %" PRIu64 "M on socket %i\n",
 674                         (seg_len * page_sz) >> 20ULL, socket_id);
 675
 676         /* find free space in memseg lists */
 677         for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
 678                 bool empty;
 679                 msl = &mcfg->memsegs[msl_idx];
 680                 arr = &msl->memseg_arr;
 681
 682                 if (msl->page_sz != page_sz)
 683                         continue;
 684                 if (msl->socket_id != socket_id)
 685                         continue;
 686
 687                 /* leave space for a hole if array is not empty */
 688                 empty = arr->count == 0;
 689                 ms_idx = rte_fbarray_find_next_n_free(arr, 0,
 690                                 seg_len + (empty ? 0 : 1));
 691
 692                 /* memseg list is full? */
 693                 if (ms_idx < 0)
 694                         continue;
 695
 696                 /* leave some space between memsegs, they are not IOVA
 697                  * contiguous, so they shouldn't be VA contiguous either.
 698                  */
 699                 if (!empty)
 700                         ms_idx++;
 701                 break;
 702         }
 703         if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
 704                 RTE_LOG(ERR, EAL, "Could not find space for memseg. Please increase %s and/or %s in configuration.\n",
 705                                 RTE_STR(CONFIG_RTE_MAX_MEMSEG_PER_TYPE),
 706                                 RTE_STR(CONFIG_RTE_MAX_MEM_PER_TYPE));
 707                 return -1;
 708         }
 709
 710 #ifdef RTE_ARCH_PPC_64
 711         /* for PPC64 we go through the list backwards */
 712         for (cur_page = seg_end - 1; cur_page >= seg_start;
 713                         cur_page--, ms_idx++) {
 714 #else
 715         for (cur_page = seg_start; cur_page < seg_end; cur_page++, ms_idx++) {
 716 #endif
 717                 struct hugepage_file *hfile = &hugepages[cur_page];
 718                 struct rte_memseg *ms = rte_fbarray_get(arr, ms_idx);
 719                 void *addr;
 720                 int fd;
 721
 722                 fd = open(hfile->filepath, O_RDWR);
 723                 if (fd < 0) {
 724                         RTE_LOG(ERR, EAL, "Could not open '%s': %s\n",
 725                                         hfile->filepath, strerror(errno));
 726                         return -1;
 727                 }
 728                 /* set shared lock on the file. */
 729                 if (flock(fd, LOCK_SH) < 0) {
 730                         RTE_LOG(DEBUG, EAL, "Could not lock '%s': %s\n",
 731                                         hfile->filepath, strerror(errno));
 732                         close(fd);
 733                         return -1;
 734                 }
 735                 memseg_len = (size_t)page_sz;
 736                 addr = RTE_PTR_ADD(msl->base_va, ms_idx * memseg_len);
 737
 738                 /* we know this address is already mmapped by memseg list, so
 739                  * using MAP_FIXED here is safe
 740                  */
 741                 addr = mmap(addr, page_sz, PROT_READ | PROT_WRITE,
 742                                 MAP_SHARED | MAP_POPULATE | MAP_FIXED, fd, 0);
 743                 if (addr == MAP_FAILED) {
 744                         RTE_LOG(ERR, EAL, "Couldn't remap '%s': %s\n",
 745                                         hfile->filepath, strerror(errno));
 746                         close(fd);
 747                         return -1;
 748                 }
 749
 750                 /* we have a new address, so unmap previous one */
 751 #ifndef RTE_ARCH_64
 752                 /* in 32-bit legacy mode, we have already unmapped the page */
 753                 if (!internal_config.legacy_mem)
 754                         munmap(hfile->orig_va, page_sz);
 755 #else
 756                 munmap(hfile->orig_va, page_sz);
 757 #endif
 758
 759                 hfile->orig_va = NULL;
 760                 hfile->final_va = addr;
 761
 762                 /* rewrite physical addresses in IOVA as VA mode */
 763                 if (rte_eal_iova_mode() == RTE_IOVA_VA)
 764                         hfile->physaddr = (uintptr_t)addr;
 765
 766                 /* set up memseg data */
 767                 ms->addr = addr;
 768                 ms->hugepage_sz = page_sz;
 769                 ms->len = memseg_len;
 770                 ms->iova = hfile->physaddr;
 771                 ms->socket_id = hfile->socket_id;
 772                 ms->nchannel = rte_memory_get_nchannel();
 773                 ms->nrank = rte_memory_get_nrank();
 774
 775                 rte_fbarray_set_used(arr, ms_idx);
 776
 777                 /* store segment fd internally */
 778                 if (eal_memalloc_set_seg_fd(msl_idx, ms_idx, fd) < 0)
 779                         RTE_LOG(ERR, EAL, "Could not store segment fd: %s\n",
 780                                 rte_strerror(rte_errno));
 781         }
 782         RTE_LOG(DEBUG, EAL, "Allocated %" PRIu64 "M on socket %i\n",
 783                         (seg_len * page_sz) >> 20, socket_id);
 784         return 0;
 785 }
 786
 787 static uint64_t
 788 get_mem_amount(uint64_t page_sz, uint64_t max_mem)
 789 {
 790         uint64_t area_sz, max_pages;
 791
 792         /* limit to RTE_MAX_MEMSEG_PER_LIST pages or RTE_MAX_MEM_MB_PER_LIST */
 793         max_pages = RTE_MAX_MEMSEG_PER_LIST;
 794         max_mem = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20, max_mem);
 795
 796         area_sz = RTE_MIN(page_sz * max_pages, max_mem);
 797
 798         /* make sure the list isn't smaller than the page size */
 799         area_sz = RTE_MAX(area_sz, page_sz);
 800
 801         return RTE_ALIGN(area_sz, page_sz);
 802 }
 803
 804 static int
 805 memseg_list_free(struct rte_memseg_list *msl)
 806 {
 807         if (rte_fbarray_destroy(&msl->memseg_arr)) {
 808                 RTE_LOG(ERR, EAL, "Cannot destroy memseg list\n");
 809                 return -1;
 810         }
 811         memset(msl, 0, sizeof(*msl));
 812         return 0;
 813 }
 814
 815 static int
 816 memseg_list_init(struct rte_memseg_list *msl, uint64_t page_sz,
 817                 int n_segs, int socket_id, int type_msl_idx)
 818 {
 819         return eal_memseg_list_init(
 820                 msl, page_sz, n_segs, socket_id, type_msl_idx, true);
 821 }
 822
 823 static int
 824 memseg_list_alloc(struct rte_memseg_list *msl)
 825 {
 826         return eal_memseg_list_alloc(msl, 0);
 827 }
 828
 829 /*
 830  * Our VA space is not preallocated yet, so preallocate it here. We need to know
 831  * how many segments there are in order to map all pages into one address space,
 832  * and leave appropriate holes between segments so that rte_malloc does not
 833  * concatenate them into one big segment.
 834  *
 835  * we also need to unmap original pages to free up address space.
 836  */
 837 static int __rte_unused
 838 prealloc_segments(struct hugepage_file *hugepages, int n_pages)
 839 {
 840         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 841         int cur_page, seg_start_page, end_seg, new_memseg;
 842         unsigned int hpi_idx, socket, i;
 843         int n_contig_segs, n_segs;
 844         int msl_idx;
 845
 846         /* before we preallocate segments, we need to free up our VA space.
 847          * we're not removing files, and we already have information about
 848          * PA-contiguousness, so it is safe to unmap everything.
 849          */
 850         for (cur_page = 0; cur_page < n_pages; cur_page++) {
 851                 struct hugepage_file *hpi = &hugepages[cur_page];
 852                 munmap(hpi->orig_va, hpi->size);
 853                 hpi->orig_va = NULL;
 854         }
 855
 856         /* we cannot know how many page sizes and sockets we have discovered, so
 857          * loop over all of them
 858          */
 859         for (hpi_idx = 0; hpi_idx < internal_config.num_hugepage_sizes;
 860                         hpi_idx++) {
 861                 uint64_t page_sz =
 862                         internal_config.hugepage_info[hpi_idx].hugepage_sz;
 863
 864                 for (i = 0; i < rte_socket_count(); i++) {
 865                         struct rte_memseg_list *msl;
 866
 867                         socket = rte_socket_id_by_idx(i);
 868                         n_contig_segs = 0;
 869                         n_segs = 0;
 870                         seg_start_page = -1;
 871
 872                         for (cur_page = 0; cur_page < n_pages; cur_page++) {
 873                                 struct hugepage_file *prev, *cur;
 874                                 int prev_seg_start_page = -1;
 875
 876                                 cur = &hugepages[cur_page];
 877                                 prev = cur_page == 0 ? NULL :
 878                                                 &hugepages[cur_page - 1];
 879
 880                                 new_memseg = 0;
 881                                 end_seg = 0;
 882
 883                                 if (cur->size == 0)
 884                                         end_seg = 1;
 885                                 else if (cur->socket_id != (int) socket)
 886                                         end_seg = 1;
 887                                 else if (cur->size != page_sz)
 888                                         end_seg = 1;
 889                                 else if (cur_page == 0)
 890                                         new_memseg = 1;
 891 #ifdef RTE_ARCH_PPC_64
 892                                 /* On PPC64 architecture, the mmap always start
 893                                  * from higher address to lower address. Here,
 894                                  * physical addresses are in descending order.
 895                                  */
 896                                 else if ((prev->physaddr - cur->physaddr) !=
 897                                                 cur->size)
 898                                         new_memseg = 1;
 899 #else
 900                                 else if ((cur->physaddr - prev->physaddr) !=
 901                                                 cur->size)
 902                                         new_memseg = 1;
 903 #endif
 904                                 if (new_memseg) {
 905                                         /* if we're already inside a segment,
 906                                          * new segment means end of current one
 907                                          */
 908                                         if (seg_start_page != -1) {
 909                                                 end_seg = 1;
 910                                                 prev_seg_start_page =
 911                                                                 seg_start_page;
 912                                         }
 913                                         seg_start_page = cur_page;
 914                                 }
 915
 916                                 if (end_seg) {
 917                                         if (prev_seg_start_page != -1) {
 918                                                 /* we've found a new segment */
 919                                                 n_contig_segs++;
 920                                                 n_segs += cur_page -
 921                                                         prev_seg_start_page;
 922                                         } else if (seg_start_page != -1) {
 923                                                 /* we didn't find new segment,
 924                                                  * but did end current one
 925                                                  */
 926                                                 n_contig_segs++;
 927                                                 n_segs += cur_page -
 928                                                                 seg_start_page;
 929                                                 seg_start_page = -1;
 930                                                 continue;
 931                                         } else {
 932                                                 /* we're skipping this page */
 933                                                 continue;
 934                                         }
 935                                 }
 936                                 /* segment continues */
 937                         }
 938                         /* check if we missed last segment */
 939                         if (seg_start_page != -1) {
 940                                 n_contig_segs++;
 941                                 n_segs += cur_page - seg_start_page;
 942                         }
 943
 944                         /* if no segments were found, do not preallocate */
 945                         if (n_segs == 0)
 946                                 continue;
 947
 948                         /* we now have total number of pages that we will
 949                          * allocate for this segment list. add separator pages
 950                          * to the total count, and preallocate VA space.
 951                          */
 952                         n_segs += n_contig_segs - 1;
 953
 954                         /* now, preallocate VA space for these segments */
 955
 956                         /* first, find suitable memseg list for this */
 957                         for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS;
 958                                         msl_idx++) {
 959                                 msl = &mcfg->memsegs[msl_idx];
 960
 961                                 if (msl->base_va != NULL)
 962                                         continue;
 963                                 break;
 964                         }
 965                         if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
 966                                 RTE_LOG(ERR, EAL, "Not enough space in memseg lists, please increase %s\n",
 967                                         RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
 968                                 return -1;
 969                         }
 970
 971                         /* now, allocate fbarray itself */
 972                         if (memseg_list_init(msl, page_sz, n_segs, socket,
 973                                                 msl_idx) < 0)
 974                                 return -1;
 975
 976                         /* finally, allocate VA space */
 977                         if (memseg_list_alloc(msl) < 0) {
 978                                 RTE_LOG(ERR, EAL, "Cannot preallocate 0x%"PRIx64"kB hugepages\n",
 979                                         page_sz >> 10);
 980                                 return -1;
 981                         }
 982                 }
 983         }
 984         return 0;
 985 }
 986
 987 /*
 988  * We cannot reallocate memseg lists on the fly because PPC64 stores pages
 989  * backwards, therefore we have to process the entire memseg first before
 990  * remapping it into memseg list VA space.
 991  */
 992 static int
 993 remap_needed_hugepages(struct hugepage_file *hugepages, int n_pages)
 994 {
 995         int cur_page, seg_start_page, new_memseg, ret;
 996
 997         seg_start_page = 0;
 998         for (cur_page = 0; cur_page < n_pages; cur_page++) {
 999                 struct hugepage_file *prev, *cur;
1000
1001                 new_memseg = 0;
1002
1003                 cur = &hugepages[cur_page];
1004                 prev = cur_page == 0 ? NULL : &hugepages[cur_page - 1];
1005
1006                 /* if size is zero, no more pages left */
1007                 if (cur->size == 0)
1008                         break;
1009
1010                 if (cur_page == 0)
1011                         new_memseg = 1;
1012                 else if (cur->socket_id != prev->socket_id)
1013                         new_memseg = 1;
1014                 else if (cur->size != prev->size)
1015                         new_memseg = 1;
1016 #ifdef RTE_ARCH_PPC_64
1017                 /* On PPC64 architecture, the mmap always start from higher
1018                  * address to lower address. Here, physical addresses are in
1019                  * descending order.
1020                  */
1021                 else if ((prev->physaddr - cur->physaddr) != cur->size)
1022                         new_memseg = 1;
1023 #else
1024                 else if ((cur->physaddr - prev->physaddr) != cur->size)
1025                         new_memseg = 1;
1026 #endif
1027
1028                 if (new_memseg) {
1029                         /* if this isn't the first time, remap segment */
1030                         if (cur_page != 0) {
1031                                 ret = remap_segment(hugepages, seg_start_page,
1032                                                 cur_page);
1033                                 if (ret != 0)
1034                                         return -1;
1035                         }
1036                         /* remember where we started */
1037                         seg_start_page = cur_page;
1038                 }
1039                 /* continuation of previous memseg */
1040         }
1041         /* we were stopped, but we didn't remap the last segment, do it now */
1042         if (cur_page != 0) {
1043                 ret = remap_segment(hugepages, seg_start_page,
1044                                 cur_page);
1045                 if (ret != 0)
1046                         return -1;
1047         }
1048         return 0;
1049 }
1050
1051 __rte_unused /* function is unused on 32-bit builds */
1052 static inline uint64_t
1053 get_socket_mem_size(int socket)
1054 {
1055         uint64_t size = 0;
1056         unsigned i;
1057
1058         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
1059                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
1060                 size += hpi->hugepage_sz * hpi->num_pages[socket];
1061         }
1062
1063         return size;
1064 }
1065
1066 /*
1067  * This function is a NUMA-aware equivalent of calc_num_pages.
1068  * It takes in the list of hugepage sizes and the
1069  * number of pages thereof, and calculates the best number of
1070  * pages of each size to fulfill the request for <memory> ram
1071  */
1072 static int
1073 calc_num_pages_per_socket(uint64_t * memory,
1074                 struct hugepage_info *hp_info,
1075                 struct hugepage_info *hp_used,
1076                 unsigned num_hp_info)
1077 {
1078         unsigned socket, j, i = 0;
1079         unsigned requested, available;
1080         int total_num_pages = 0;
1081         uint64_t remaining_mem, cur_mem;
1082         uint64_t total_mem = internal_config.memory;
1083
1084         if (num_hp_info == 0)
1085                 return -1;
1086
1087         /* if specific memory amounts per socket weren't requested */
1088         if (internal_config.force_sockets == 0) {
1089                 size_t total_size;
1090 #ifdef RTE_ARCH_64
1091                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
1092                 size_t default_size;
1093                 unsigned lcore_id;
1094
1095                 /* Compute number of cores per socket */
1096                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
1097                 RTE_LCORE_FOREACH(lcore_id) {
1098                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
1099                 }
1100
1101                 /*
1102                  * Automatically spread requested memory amongst detected sockets according
1103                  * to number of cores from cpu mask present on each socket
1104                  */
1105                 total_size = internal_config.memory;
1106                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
1107
1108                         /* Set memory amount per socket */
1109                         default_size = (internal_config.memory * cpu_per_socket[socket])
1110                                         / rte_lcore_count();
1111
1112                         /* Limit to maximum available memory on socket */
1113                         default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
1114
1115                         /* Update sizes */
1116                         memory[socket] = default_size;
1117                         total_size -= default_size;
1118                 }
1119
1120                 /*
1121                  * If some memory is remaining, try to allocate it by getting all
1122                  * available memory from sockets, one after the other
1123                  */
1124                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
1125                         /* take whatever is available */
1126                         default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
1127                                                total_size);
1128
1129                         /* Update sizes */
1130                         memory[socket] += default_size;
1131                         total_size -= default_size;
1132                 }
1133 #else
1134                 /* in 32-bit mode, allocate all of the memory only on master
1135                  * lcore socket
1136                  */
1137                 total_size = internal_config.memory;
1138                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0;
1139                                 socket++) {
1140                         struct rte_config *cfg = rte_eal_get_configuration();
1141                         unsigned int master_lcore_socket;
1142
1143                         master_lcore_socket =
1144                                 rte_lcore_to_socket_id(cfg->master_lcore);
1145
1146                         if (master_lcore_socket != socket)
1147                                 continue;
1148
1149                         /* Update sizes */
1150                         memory[socket] = total_size;
1151                         break;
1152                 }
1153 #endif
1154         }
1155
1156         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
1157                 /* skips if the memory on specific socket wasn't requested */
1158                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
1159                         strlcpy(hp_used[i].hugedir, hp_info[i].hugedir,
1160                                 sizeof(hp_used[i].hugedir));
1161                         hp_used[i].num_pages[socket] = RTE_MIN(
1162                                         memory[socket] / hp_info[i].hugepage_sz,
1163                                         hp_info[i].num_pages[socket]);
1164
1165                         cur_mem = hp_used[i].num_pages[socket] *
1166                                         hp_used[i].hugepage_sz;
1167
1168                         memory[socket] -= cur_mem;
1169                         total_mem -= cur_mem;
1170
1171                         total_num_pages += hp_used[i].num_pages[socket];
1172
1173                         /* check if we have met all memory requests */
1174                         if (memory[socket] == 0)
1175                                 break;
1176
1177                         /* check if we have any more pages left at this size, if so
1178                          * move on to next size */
1179                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
1180                                 continue;
1181                         /* At this point we know that there are more pages available that are
1182                          * bigger than the memory we want, so lets see if we can get enough
1183                          * from other page sizes.
1184                          */
1185                         remaining_mem = 0;
1186                         for (j = i+1; j < num_hp_info; j++)
1187                                 remaining_mem += hp_info[j].hugepage_sz *
1188                                 hp_info[j].num_pages[socket];
1189
1190                         /* is there enough other memory, if not allocate another page and quit */
1191                         if (remaining_mem < memory[socket]){
1192                                 cur_mem = RTE_MIN(memory[socket],
1193                                                 hp_info[i].hugepage_sz);
1194                                 memory[socket] -= cur_mem;
1195                                 total_mem -= cur_mem;
1196                                 hp_used[i].num_pages[socket]++;
1197                                 total_num_pages++;
1198                                 break; /* we are done with this socket*/
1199                         }
1200                 }
1201                 /* if we didn't satisfy all memory requirements per socket */
1202                 if (memory[socket] > 0 &&
1203                                 internal_config.socket_mem[socket] != 0) {
1204                         /* to prevent icc errors */
1205                         requested = (unsigned) (internal_config.socket_mem[socket] /
1206                                         0x100000);
1207                         available = requested -
1208                                         ((unsigned) (memory[socket] / 0x100000));
1209                         RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
1210                                         "Requested: %uMB, available: %uMB\n", socket,
1211                                         requested, available);
1212                         return -1;
1213                 }
1214         }
1215
1216         /* if we didn't satisfy total memory requirements */
1217         if (total_mem > 0) {
1218                 requested = (unsigned) (internal_config.memory / 0x100000);
1219                 available = requested - (unsigned) (total_mem / 0x100000);
1220                 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
1221                                 " available: %uMB\n", requested, available);
1222                 return -1;
1223         }
1224         return total_num_pages;
1225 }
1226
1227 static inline size_t
1228 eal_get_hugepage_mem_size(void)
1229 {
1230         uint64_t size = 0;
1231         unsigned i, j;
1232
1233         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
1234                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
1235                 if (strnlen(hpi->hugedir, sizeof(hpi->hugedir)) != 0) {
1236                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1237                                 size += hpi->hugepage_sz * hpi->num_pages[j];
1238                         }
1239                 }
1240         }
1241
1242         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
1243 }
1244
1245 static struct sigaction huge_action_old;
1246 static int huge_need_recover;
1247
1248 static void
1249 huge_register_sigbus(void)
1250 {
1251         sigset_t mask;
1252         struct sigaction action;
1253
1254         sigemptyset(&mask);
1255         sigaddset(&mask, SIGBUS);
1256         action.sa_flags = 0;
1257         action.sa_mask = mask;
1258         action.sa_handler = huge_sigbus_handler;
1259
1260         huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
1261 }
1262
1263 static void
1264 huge_recover_sigbus(void)
1265 {
1266         if (huge_need_recover) {
1267                 sigaction(SIGBUS, &huge_action_old, NULL);
1268                 huge_need_recover = 0;
1269         }
1270 }
1271
1272 /*
1273  * Prepare physical memory mapping: fill configuration structure with
1274  * these infos, return 0 on success.
1275  *  1. map N huge pages in separate files in hugetlbfs
1276  *  2. find associated physical addr
1277  *  3. find associated NUMA socket ID
1278  *  4. sort all huge pages by physical address
1279  *  5. remap these N huge pages in the correct order
1280  *  6. unmap the first mapping
1281  *  7. fill memsegs in configuration with contiguous zones
1282  */
1283 static int
1284 eal_legacy_hugepage_init(void)
1285 {
1286         struct rte_mem_config *mcfg;
1287         struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1288         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1289
1290         uint64_t memory[RTE_MAX_NUMA_NODES];
1291
1292         unsigned hp_offset;
1293         int i, j;
1294         int nr_hugefiles, nr_hugepages = 0;
1295         void *addr;
1296
1297         memset(used_hp, 0, sizeof(used_hp));
1298
1299         /* get pointer to global configuration */
1300         mcfg = rte_eal_get_configuration()->mem_config;
1301
1302         /* hugetlbfs can be disabled */
1303         if (internal_config.no_hugetlbfs) {
1304                 void *prealloc_addr;
1305                 size_t mem_sz;
1306                 struct rte_memseg_list *msl;
1307                 int n_segs, fd, flags;
1308 #ifdef MEMFD_SUPPORTED
1309                 int memfd;
1310 #endif
1311                 uint64_t page_sz;
1312
1313                 /* nohuge mode is legacy mode */
1314                 internal_config.legacy_mem = 1;
1315
1316                 /* nohuge mode is single-file segments mode */
1317                 internal_config.single_file_segments = 1;
1318
1319                 /* create a memseg list */
1320                 msl = &mcfg->memsegs[0];
1321
1322                 mem_sz = internal_config.memory;
1323                 page_sz = RTE_PGSIZE_4K;
1324                 n_segs = mem_sz / page_sz;
1325
1326                 if (eal_memseg_list_init_named(
1327                                 msl, "nohugemem", page_sz, n_segs, 0, true)) {
1328                         return -1;
1329                 }
1330
1331                 /* set up parameters for anonymous mmap */
1332                 fd = -1;
1333                 flags = MAP_PRIVATE | MAP_ANONYMOUS;
1334
1335 #ifdef MEMFD_SUPPORTED
1336                 /* create a memfd and store it in the segment fd table */
1337                 memfd = memfd_create("nohuge", 0);
1338                 if (memfd < 0) {
1339                         RTE_LOG(DEBUG, EAL, "Cannot create memfd: %s\n",
1340                                         strerror(errno));
1341                         RTE_LOG(DEBUG, EAL, "Falling back to anonymous map\n");
1342                 } else {
1343                         /* we got an fd - now resize it */
1344                         if (ftruncate(memfd, internal_config.memory) < 0) {
1345                                 RTE_LOG(ERR, EAL, "Cannot resize memfd: %s\n",
1346                                                 strerror(errno));
1347                                 RTE_LOG(ERR, EAL, "Falling back to anonymous map\n");
1348                                 close(memfd);
1349                         } else {
1350                                 /* creating memfd-backed file was successful.
1351                                  * we want changes to memfd to be visible to
1352                                  * other processes (such as vhost backend), so
1353                                  * map it as shared memory.
1354                                  */
1355                                 RTE_LOG(DEBUG, EAL, "Using memfd for anonymous memory\n");
1356                                 fd = memfd;
1357                                 flags = MAP_SHARED;
1358                         }
1359                 }
1360 #endif
1361                 /* preallocate address space for the memory, so that it can be
1362                  * fit into the DMA mask.
1363                  */
1364                 if (eal_memseg_list_alloc(msl, 0)) {
1365                         RTE_LOG(ERR, EAL, "Cannot preallocate VA space for hugepage memory\n");
1366                         return -1;
1367                 }
1368
1369                 prealloc_addr = msl->base_va;
1370                 addr = mmap(prealloc_addr, mem_sz, PROT_READ | PROT_WRITE,
1371                                 flags | MAP_FIXED, fd, 0);
1372                 if (addr == MAP_FAILED || addr != prealloc_addr) {
1373                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1374                                         strerror(errno));
1375                         munmap(prealloc_addr, mem_sz);
1376                         return -1;
1377                 }
1378
1379                 /* we're in single-file segments mode, so only the segment list
1380                  * fd needs to be set up.
1381                  */
1382                 if (fd != -1) {
1383                         if (eal_memalloc_set_seg_list_fd(0, fd) < 0) {
1384                                 RTE_LOG(ERR, EAL, "Cannot set up segment list fd\n");
1385                                 /* not a serious error, proceed */
1386                         }
1387                 }
1388
1389                 eal_memseg_list_populate(msl, addr, n_segs);
1390
1391                 if (mcfg->dma_maskbits &&
1392                     rte_mem_check_dma_mask_thread_unsafe(mcfg->dma_maskbits)) {
1393                         RTE_LOG(ERR, EAL,
1394                                 "%s(): couldn't allocate memory due to IOVA exceeding limits of current DMA mask.\n",
1395                                 __func__);
1396                         if (rte_eal_iova_mode() == RTE_IOVA_VA &&
1397                             rte_eal_using_phys_addrs())
1398                                 RTE_LOG(ERR, EAL,
1399                                         "%s(): Please try initializing EAL with --iova-mode=pa parameter.\n",
1400                                         __func__);
1401                         goto fail;
1402                 }
1403                 return 0;
1404         }
1405
1406         /* calculate total number of hugepages available. at this point we haven't
1407          * yet started sorting them so they all are on socket 0 */
1408         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1409                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1410                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1411
1412                 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1413         }
1414
1415         /*
1416          * allocate a memory area for hugepage table.
1417          * this isn't shared memory yet. due to the fact that we need some
1418          * processing done on these pages, shared memory will be created
1419          * at a later stage.
1420          */
1421         tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1422         if (tmp_hp == NULL)
1423                 goto fail;
1424
1425         memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1426
1427         hp_offset = 0; /* where we start the current page size entries */
1428
1429         huge_register_sigbus();
1430
1431         /* make a copy of socket_mem, needed for balanced allocation. */
1432         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1433                 memory[i] = internal_config.socket_mem[i];
1434
1435         /* map all hugepages and sort them */
1436         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1437                 unsigned pages_old, pages_new;
1438                 struct hugepage_info *hpi;
1439
1440                 /*
1441                  * we don't yet mark hugepages as used at this stage, so
1442                  * we just map all hugepages available to the system
1443                  * all hugepages are still located on socket 0
1444                  */
1445                 hpi = &internal_config.hugepage_info[i];
1446
1447                 if (hpi->num_pages[0] == 0)
1448                         continue;
1449
1450                 /* map all hugepages available */
1451                 pages_old = hpi->num_pages[0];
1452                 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, memory);
1453                 if (pages_new < pages_old) {
1454                         RTE_LOG(DEBUG, EAL,
1455                                 "%d not %d hugepages of size %u MB allocated\n",
1456                                 pages_new, pages_old,
1457                                 (unsigned)(hpi->hugepage_sz / 0x100000));
1458
1459                         int pages = pages_old - pages_new;
1460
1461                         nr_hugepages -= pages;
1462                         hpi->num_pages[0] = pages_new;
1463                         if (pages_new == 0)
1464                                 continue;
1465                 }
1466
1467                 if (rte_eal_using_phys_addrs() &&
1468                                 rte_eal_iova_mode() != RTE_IOVA_VA) {
1469                         /* find physical addresses for each hugepage */
1470                         if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1471                                 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1472                                         "for %u MB pages\n",
1473                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1474                                 goto fail;
1475                         }
1476                 } else {
1477                         /* set physical addresses for each hugepage */
1478                         if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1479                                 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1480                                         "for %u MB pages\n",
1481                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1482                                 goto fail;
1483                         }
1484                 }
1485
1486                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1487                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1488                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1489                         goto fail;
1490                 }
1491
1492                 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1493                       sizeof(struct hugepage_file), cmp_physaddr);
1494
1495                 /* we have processed a num of hugepages of this size, so inc offset */
1496                 hp_offset += hpi->num_pages[0];
1497         }
1498
1499         huge_recover_sigbus();
1500
1501         if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1502                 internal_config.memory = eal_get_hugepage_mem_size();
1503
1504         nr_hugefiles = nr_hugepages;
1505
1506
1507         /* clean out the numbers of pages */
1508         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1509                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1510                         internal_config.hugepage_info[i].num_pages[j] = 0;
1511
1512         /* get hugepages for each socket */
1513         for (i = 0; i < nr_hugefiles; i++) {
1514                 int socket = tmp_hp[i].socket_id;
1515
1516                 /* find a hugepage info with right size and increment num_pages */
1517                 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1518                                 (int)internal_config.num_hugepage_sizes);
1519                 for (j = 0; j < nb_hpsizes; j++) {
1520                         if (tmp_hp[i].size ==
1521                                         internal_config.hugepage_info[j].hugepage_sz) {
1522                                 internal_config.hugepage_info[j].num_pages[socket]++;
1523                         }
1524                 }
1525         }
1526
1527         /* make a copy of socket_mem, needed for number of pages calculation */
1528         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1529                 memory[i] = internal_config.socket_mem[i];
1530
1531         /* calculate final number of pages */
1532         nr_hugepages = calc_num_pages_per_socket(memory,
1533                         internal_config.hugepage_info, used_hp,
1534                         internal_config.num_hugepage_sizes);
1535
1536         /* error if not enough memory available */
1537         if (nr_hugepages < 0)
1538                 goto fail;
1539
1540         /* reporting in! */
1541         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1542                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1543                         if (used_hp[i].num_pages[j] > 0) {
1544                                 RTE_LOG(DEBUG, EAL,
1545                                         "Requesting %u pages of size %uMB"
1546                                         " from socket %i\n",
1547                                         used_hp[i].num_pages[j],
1548                                         (unsigned)
1549                                         (used_hp[i].hugepage_sz / 0x100000),
1550                                         j);
1551                         }
1552                 }
1553         }
1554
1555         /* create shared memory */
1556         hugepage = create_shared_memory(eal_hugepage_data_path(),
1557                         nr_hugefiles * sizeof(struct hugepage_file));
1558
1559         if (hugepage == NULL) {
1560                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1561                 goto fail;
1562         }
1563         memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1564
1565         /*
1566          * unmap pages that we won't need (looks at used_hp).
1567          * also, sets final_va to NULL on pages that were unmapped.
1568          */
1569         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1570                         internal_config.num_hugepage_sizes) < 0) {
1571                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1572                 goto fail;
1573         }
1574
1575         /*
1576          * copy stuff from malloc'd hugepage* to the actual shared memory.
1577          * this procedure only copies those hugepages that have orig_va
1578          * not NULL. has overflow protection.
1579          */
1580         if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1581                         tmp_hp, nr_hugefiles) < 0) {
1582                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1583                 goto fail;
1584         }
1585
1586 #ifndef RTE_ARCH_64
1587         /* for legacy 32-bit mode, we did not preallocate VA space, so do it */
1588         if (internal_config.legacy_mem &&
1589                         prealloc_segments(hugepage, nr_hugefiles)) {
1590                 RTE_LOG(ERR, EAL, "Could not preallocate VA space for hugepages\n");
1591                 goto fail;
1592         }
1593 #endif
1594
1595         /* remap all pages we do need into memseg list VA space, so that those
1596          * pages become first-class citizens in DPDK memory subsystem
1597          */
1598         if (remap_needed_hugepages(hugepage, nr_hugefiles)) {
1599                 RTE_LOG(ERR, EAL, "Couldn't remap hugepage files into memseg lists\n");
1600                 goto fail;
1601         }
1602
1603         /* free the hugepage backing files */
1604         if (internal_config.hugepage_unlink &&
1605                 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1606                 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1607                 goto fail;
1608         }
1609
1610         /* free the temporary hugepage table */
1611         free(tmp_hp);
1612         tmp_hp = NULL;
1613
1614         munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1615         hugepage = NULL;
1616
1617         /* we're not going to allocate more pages, so release VA space for
1618          * unused memseg lists
1619          */
1620         for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
1621                 struct rte_memseg_list *msl = &mcfg->memsegs[i];
1622                 size_t mem_sz;
1623
1624                 /* skip inactive lists */
1625                 if (msl->base_va == NULL)
1626                         continue;
1627                 /* skip lists where there is at least one page allocated */
1628                 if (msl->memseg_arr.count > 0)
1629                         continue;
1630                 /* this is an unused list, deallocate it */
1631                 mem_sz = msl->len;
1632                 munmap(msl->base_va, mem_sz);
1633                 msl->base_va = NULL;
1634                 msl->heap = 0;
1635
1636                 /* destroy backing fbarray */
1637                 rte_fbarray_destroy(&msl->memseg_arr);
1638         }
1639
1640         if (mcfg->dma_maskbits &&
1641             rte_mem_check_dma_mask_thread_unsafe(mcfg->dma_maskbits)) {
1642                 RTE_LOG(ERR, EAL,
1643                         "%s(): couldn't allocate memory due to IOVA exceeding limits of current DMA mask.\n",
1644                         __func__);
1645                 goto fail;
1646         }
1647
1648         return 0;
1649
1650 fail:
1651         huge_recover_sigbus();
1652         free(tmp_hp);
1653         if (hugepage != NULL)
1654                 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1655
1656         return -1;
1657 }
1658
1659 static int __rte_unused
1660 hugepage_count_walk(const struct rte_memseg_list *msl, void *arg)
1661 {
1662         struct hugepage_info *hpi = arg;
1663
1664         if (msl->page_sz != hpi->hugepage_sz)
1665                 return 0;
1666
1667         hpi->num_pages[msl->socket_id] += msl->memseg_arr.len;
1668         return 0;
1669 }
1670
1671 static int
1672 limits_callback(int socket_id, size_t cur_limit, size_t new_len)
1673 {
1674         RTE_SET_USED(socket_id);
1675         RTE_SET_USED(cur_limit);
1676         RTE_SET_USED(new_len);
1677         return -1;
1678 }
1679
1680 static int
1681 eal_hugepage_init(void)
1682 {
1683         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1684         uint64_t memory[RTE_MAX_NUMA_NODES];
1685         int hp_sz_idx, socket_id;
1686
1687         memset(used_hp, 0, sizeof(used_hp));
1688
1689         for (hp_sz_idx = 0;
1690                         hp_sz_idx < (int) internal_config.num_hugepage_sizes;
1691                         hp_sz_idx++) {
1692 #ifndef RTE_ARCH_64
1693                 struct hugepage_info dummy;
1694                 unsigned int i;
1695 #endif
1696                 /* also initialize used_hp hugepage sizes in used_hp */
1697                 struct hugepage_info *hpi;
1698                 hpi = &internal_config.hugepage_info[hp_sz_idx];
1699                 used_hp[hp_sz_idx].hugepage_sz = hpi->hugepage_sz;
1700
1701 #ifndef RTE_ARCH_64
1702                 /* for 32-bit, limit number of pages on socket to whatever we've
1703                  * preallocated, as we cannot allocate more.
1704                  */
1705                 memset(&dummy, 0, sizeof(dummy));
1706                 dummy.hugepage_sz = hpi->hugepage_sz;
1707                 if (rte_memseg_list_walk(hugepage_count_walk, &dummy) < 0)
1708                         return -1;
1709
1710                 for (i = 0; i < RTE_DIM(dummy.num_pages); i++) {
1711                         hpi->num_pages[i] = RTE_MIN(hpi->num_pages[i],
1712                                         dummy.num_pages[i]);
1713                 }
1714 #endif
1715         }
1716
1717         /* make a copy of socket_mem, needed for balanced allocation. */
1718         for (hp_sz_idx = 0; hp_sz_idx < RTE_MAX_NUMA_NODES; hp_sz_idx++)
1719                 memory[hp_sz_idx] = internal_config.socket_mem[hp_sz_idx];
1720
1721         /* calculate final number of pages */
1722         if (calc_num_pages_per_socket(memory,
1723                         internal_config.hugepage_info, used_hp,
1724                         internal_config.num_hugepage_sizes) < 0)
1725                 return -1;
1726
1727         for (hp_sz_idx = 0;
1728                         hp_sz_idx < (int)internal_config.num_hugepage_sizes;
1729                         hp_sz_idx++) {
1730                 for (socket_id = 0; socket_id < RTE_MAX_NUMA_NODES;
1731                                 socket_id++) {
1732                         struct rte_memseg **pages;
1733                         struct hugepage_info *hpi = &used_hp[hp_sz_idx];
1734                         unsigned int num_pages = hpi->num_pages[socket_id];
1735                         unsigned int num_pages_alloc;
1736
1737                         if (num_pages == 0)
1738                                 continue;
1739
1740                         RTE_LOG(DEBUG, EAL, "Allocating %u pages of size %" PRIu64 "M on socket %i\n",
1741                                 num_pages, hpi->hugepage_sz >> 20, socket_id);
1742
1743                         /* we may not be able to allocate all pages in one go,
1744                          * because we break up our memory map into multiple
1745                          * memseg lists. therefore, try allocating multiple
1746                          * times and see if we can get the desired number of
1747                          * pages from multiple allocations.
1748                          */
1749
1750                         num_pages_alloc = 0;
1751                         do {
1752                                 int i, cur_pages, needed;
1753
1754                                 needed = num_pages - num_pages_alloc;
1755
1756                                 pages = malloc(sizeof(*pages) * needed);
1757
1758                                 /* do not request exact number of pages */
1759                                 cur_pages = eal_memalloc_alloc_seg_bulk(pages,
1760                                                 needed, hpi->hugepage_sz,
1761                                                 socket_id, false);
1762                                 if (cur_pages <= 0) {
1763                                         free(pages);
1764                                         return -1;
1765                                 }
1766
1767                                 /* mark preallocated pages as unfreeable */
1768                                 for (i = 0; i < cur_pages; i++) {
1769                                         struct rte_memseg *ms = pages[i];
1770                                         ms->flags |= RTE_MEMSEG_FLAG_DO_NOT_FREE;
1771                                 }
1772                                 free(pages);
1773
1774                                 num_pages_alloc += cur_pages;
1775                         } while (num_pages_alloc != num_pages);
1776                 }
1777         }
1778         /* if socket limits were specified, set them */
1779         if (internal_config.force_socket_limits) {
1780                 unsigned int i;
1781                 for (i = 0; i < RTE_MAX_NUMA_NODES; i++) {
1782                         uint64_t limit = internal_config.socket_limit[i];
1783                         if (limit == 0)
1784                                 continue;
1785                         if (rte_mem_alloc_validator_register("socket-limit",
1786                                         limits_callback, i, limit))
1787                                 RTE_LOG(ERR, EAL, "Failed to register socket limits validator callback\n");
1788                 }
1789         }
1790         return 0;
1791 }
1792
1793 /*
1794  * uses fstat to report the size of a file on disk
1795  */
1796 static off_t
1797 getFileSize(int fd)
1798 {
1799         struct stat st;
1800         if (fstat(fd, &st) < 0)
1801                 return 0;
1802         return st.st_size;
1803 }
1804
1805 /*
1806  * This creates the memory mappings in the secondary process to match that of
1807  * the server process. It goes through each memory segment in the DPDK runtime
1808  * configuration and finds the hugepages which form that segment, mapping them
1809  * in order to form a contiguous block in the virtual memory space
1810  */
1811 static int
1812 eal_legacy_hugepage_attach(void)
1813 {
1814         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1815         struct hugepage_file *hp = NULL;
1816         unsigned int num_hp = 0;
1817         unsigned int i = 0;
1818         unsigned int cur_seg;
1819         off_t size = 0;
1820         int fd, fd_hugepage = -1;
1821
1822         if (aslr_enabled() > 0) {
1823                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1824                                 "(ASLR) is enabled in the kernel.\n");
1825                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1826                                 "into secondary processes\n");
1827         }
1828
1829         fd_hugepage = open(eal_hugepage_data_path(), O_RDONLY);
1830         if (fd_hugepage < 0) {
1831                 RTE_LOG(ERR, EAL, "Could not open %s\n",
1832                                 eal_hugepage_data_path());
1833                 goto error;
1834         }
1835
1836         size = getFileSize(fd_hugepage);
1837         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1838         if (hp == MAP_FAILED) {
1839                 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1840                                 eal_hugepage_data_path());
1841                 goto error;
1842         }
1843
1844         num_hp = size / sizeof(struct hugepage_file);
1845         RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1846
1847         /* map all segments into memory to make sure we get the addrs. the
1848          * segments themselves are already in memseg list (which is shared and
1849          * has its VA space already preallocated), so we just need to map
1850          * everything into correct addresses.
1851          */
1852         for (i = 0; i < num_hp; i++) {
1853                 struct hugepage_file *hf = &hp[i];
1854                 size_t map_sz = hf->size;
1855                 void *map_addr = hf->final_va;
1856                 int msl_idx, ms_idx;
1857                 struct rte_memseg_list *msl;
1858                 struct rte_memseg *ms;
1859
1860                 /* if size is zero, no more pages left */
1861                 if (map_sz == 0)
1862                         break;
1863
1864                 fd = open(hf->filepath, O_RDWR);
1865                 if (fd < 0) {
1866                         RTE_LOG(ERR, EAL, "Could not open %s: %s\n",
1867                                 hf->filepath, strerror(errno));
1868                         goto error;
1869                 }
1870
1871                 map_addr = mmap(map_addr, map_sz, PROT_READ | PROT_WRITE,
1872                                 MAP_SHARED | MAP_FIXED, fd, 0);
1873                 if (map_addr == MAP_FAILED) {
1874                         RTE_LOG(ERR, EAL, "Could not map %s: %s\n",
1875                                 hf->filepath, strerror(errno));
1876                         goto fd_error;
1877                 }
1878
1879                 /* set shared lock on the file. */
1880                 if (flock(fd, LOCK_SH) < 0) {
1881                         RTE_LOG(DEBUG, EAL, "%s(): Locking file failed: %s\n",
1882                                 __func__, strerror(errno));
1883                         goto mmap_error;
1884                 }
1885
1886                 /* find segment data */
1887                 msl = rte_mem_virt2memseg_list(map_addr);
1888                 if (msl == NULL) {
1889                         RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg list\n",
1890                                 __func__);
1891                         goto mmap_error;
1892                 }
1893                 ms = rte_mem_virt2memseg(map_addr, msl);
1894                 if (ms == NULL) {
1895                         RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg\n",
1896                                 __func__);
1897                         goto mmap_error;
1898                 }
1899
1900                 msl_idx = msl - mcfg->memsegs;
1901                 ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
1902                 if (ms_idx < 0) {
1903                         RTE_LOG(DEBUG, EAL, "%s(): Cannot find memseg idx\n",
1904                                 __func__);
1905                         goto mmap_error;
1906                 }
1907
1908                 /* store segment fd internally */
1909                 if (eal_memalloc_set_seg_fd(msl_idx, ms_idx, fd) < 0)
1910                         RTE_LOG(ERR, EAL, "Could not store segment fd: %s\n",
1911                                 rte_strerror(rte_errno));
1912         }
1913         /* unmap the hugepage config file, since we are done using it */
1914         munmap(hp, size);
1915         close(fd_hugepage);
1916         return 0;
1917
1918 mmap_error:
1919         munmap(hp[i].final_va, hp[i].size);
1920 fd_error:
1921         close(fd);
1922 error:
1923         /* unwind mmap's done so far */
1924         for (cur_seg = 0; cur_seg < i; cur_seg++)
1925                 munmap(hp[cur_seg].final_va, hp[cur_seg].size);
1926
1927         if (hp != NULL && hp != MAP_FAILED)
1928                 munmap(hp, size);
1929         if (fd_hugepage >= 0)
1930                 close(fd_hugepage);
1931         return -1;
1932 }
1933
1934 static int
1935 eal_hugepage_attach(void)
1936 {
1937         if (eal_memalloc_sync_with_primary()) {
1938                 RTE_LOG(ERR, EAL, "Could not map memory from primary process\n");
1939                 if (aslr_enabled() > 0)
1940                         RTE_LOG(ERR, EAL, "It is recommended to disable ASLR in the kernel and retry running both primary and secondary processes\n");
1941                 return -1;
1942         }
1943         return 0;
1944 }
1945
1946 int
1947 rte_eal_hugepage_init(void)
1948 {
1949         return internal_config.legacy_mem ?
1950                         eal_legacy_hugepage_init() :
1951                         eal_hugepage_init();
1952 }
1953
1954 int
1955 rte_eal_hugepage_attach(void)
1956 {
1957         return internal_config.legacy_mem ?
1958                         eal_legacy_hugepage_attach() :
1959                         eal_hugepage_attach();
1960 }
1961
1962 int
1963 rte_eal_using_phys_addrs(void)
1964 {
1965         if (phys_addrs_available == -1) {
1966                 uint64_t tmp = 0;
1967
1968                 if (rte_eal_has_hugepages() != 0 &&
1969                     rte_mem_virt2phy(&tmp) != RTE_BAD_PHYS_ADDR)
1970                         phys_addrs_available = 1;
1971                 else
1972                         phys_addrs_available = 0;
1973         }
1974         return phys_addrs_available;
1975 }
1976
1977 static int __rte_unused
1978 memseg_primary_init_32(void)
1979 {
1980         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1981         int active_sockets, hpi_idx, msl_idx = 0;
1982         unsigned int socket_id, i;
1983         struct rte_memseg_list *msl;
1984         uint64_t extra_mem_per_socket, total_extra_mem, total_requested_mem;
1985         uint64_t max_mem;
1986
1987         /* no-huge does not need this at all */
1988         if (internal_config.no_hugetlbfs)
1989                 return 0;
1990
1991         /* this is a giant hack, but desperate times call for desperate
1992          * measures. in legacy 32-bit mode, we cannot preallocate VA space,
1993          * because having upwards of 2 gigabytes of VA space already mapped will
1994          * interfere with our ability to map and sort hugepages.
1995          *
1996          * therefore, in legacy 32-bit mode, we will be initializing memseg
1997          * lists much later - in eal_memory.c, right after we unmap all the
1998          * unneeded pages. this will not affect secondary processes, as those
1999          * should be able to mmap the space without (too many) problems.
2000          */
2001         if (internal_config.legacy_mem)
2002                 return 0;
2003
2004         /* 32-bit mode is a very special case. we cannot know in advance where
2005          * the user will want to allocate their memory, so we have to do some
2006          * heuristics.
2007          */
2008         active_sockets = 0;
2009         total_requested_mem = 0;
2010         if (internal_config.force_sockets)
2011                 for (i = 0; i < rte_socket_count(); i++) {
2012                         uint64_t mem;
2013
2014                         socket_id = rte_socket_id_by_idx(i);
2015                         mem = internal_config.socket_mem[socket_id];
2016
2017                         if (mem == 0)
2018                                 continue;
2019
2020                         active_sockets++;
2021                         total_requested_mem += mem;
2022                 }
2023         else
2024                 total_requested_mem = internal_config.memory;
2025
2026         max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
2027         if (total_requested_mem > max_mem) {
2028                 RTE_LOG(ERR, EAL, "Invalid parameters: 32-bit process can at most use %uM of memory\n",
2029                                 (unsigned int)(max_mem >> 20));
2030                 return -1;
2031         }
2032         total_extra_mem = max_mem - total_requested_mem;
2033         extra_mem_per_socket = active_sockets == 0 ? total_extra_mem :
2034                         total_extra_mem / active_sockets;
2035
2036         /* the allocation logic is a little bit convoluted, but here's how it
2037          * works, in a nutshell:
2038          *  - if user hasn't specified on which sockets to allocate memory via
2039          *    --socket-mem, we allocate all of our memory on master core socket.
2040          *  - if user has specified sockets to allocate memory on, there may be
2041          *    some "unused" memory left (e.g. if user has specified --socket-mem
2042          *    such that not all memory adds up to 2 gigabytes), so add it to all
2043          *    sockets that are in use equally.
2044          *
2045          * page sizes are sorted by size in descending order, so we can safely
2046          * assume that we dispense with bigger page sizes first.
2047          */
2048
2049         /* create memseg lists */
2050         for (i = 0; i < rte_socket_count(); i++) {
2051                 int hp_sizes = (int) internal_config.num_hugepage_sizes;
2052                 uint64_t max_socket_mem, cur_socket_mem;
2053                 unsigned int master_lcore_socket;
2054                 struct rte_config *cfg = rte_eal_get_configuration();
2055                 bool skip;
2056
2057                 socket_id = rte_socket_id_by_idx(i);
2058
2059 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
2060                 /* we can still sort pages by socket in legacy mode */
2061                 if (!internal_config.legacy_mem && socket_id > 0)
2062                         break;
2063 #endif
2064
2065                 /* if we didn't specifically request memory on this socket */
2066                 skip = active_sockets != 0 &&
2067                                 internal_config.socket_mem[socket_id] == 0;
2068                 /* ...or if we didn't specifically request memory on *any*
2069                  * socket, and this is not master lcore
2070                  */
2071                 master_lcore_socket = rte_lcore_to_socket_id(cfg->master_lcore);
2072                 skip |= active_sockets == 0 && socket_id != master_lcore_socket;
2073
2074                 if (skip) {
2075                         RTE_LOG(DEBUG, EAL, "Will not preallocate memory on socket %u\n",
2076                                         socket_id);
2077                         continue;
2078                 }
2079
2080                 /* max amount of memory on this socket */
2081                 max_socket_mem = (active_sockets != 0 ?
2082                                         internal_config.socket_mem[socket_id] :
2083                                         internal_config.memory) +
2084                                         extra_mem_per_socket;
2085                 cur_socket_mem = 0;
2086
2087                 for (hpi_idx = 0; hpi_idx < hp_sizes; hpi_idx++) {
2088                         uint64_t max_pagesz_mem, cur_pagesz_mem = 0;
2089                         uint64_t hugepage_sz;
2090                         struct hugepage_info *hpi;
2091                         int type_msl_idx, max_segs, total_segs = 0;
2092
2093                         hpi = &internal_config.hugepage_info[hpi_idx];
2094                         hugepage_sz = hpi->hugepage_sz;
2095
2096                         /* check if pages are actually available */
2097                         if (hpi->num_pages[socket_id] == 0)
2098                                 continue;
2099
2100                         max_segs = RTE_MAX_MEMSEG_PER_TYPE;
2101                         max_pagesz_mem = max_socket_mem - cur_socket_mem;
2102
2103                         /* make it multiple of page size */
2104                         max_pagesz_mem = RTE_ALIGN_FLOOR(max_pagesz_mem,
2105                                         hugepage_sz);
2106
2107                         RTE_LOG(DEBUG, EAL, "Attempting to preallocate "
2108                                         "%" PRIu64 "M on socket %i\n",
2109                                         max_pagesz_mem >> 20, socket_id);
2110
2111                         type_msl_idx = 0;
2112                         while (cur_pagesz_mem < max_pagesz_mem &&
2113                                         total_segs < max_segs) {
2114                                 uint64_t cur_mem;
2115                                 unsigned int n_segs;
2116
2117                                 if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
2118                                         RTE_LOG(ERR, EAL,
2119                                                 "No more space in memseg lists, please increase %s\n",
2120                                                 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2121                                         return -1;
2122                                 }
2123
2124                                 msl = &mcfg->memsegs[msl_idx];
2125
2126                                 cur_mem = get_mem_amount(hugepage_sz,
2127                                                 max_pagesz_mem);
2128                                 n_segs = cur_mem / hugepage_sz;
2129
2130                                 if (memseg_list_init(msl, hugepage_sz, n_segs,
2131                                                 socket_id, type_msl_idx)) {
2132                                         /* failing to allocate a memseg list is
2133                                          * a serious error.
2134                                          */
2135                                         RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
2136                                         return -1;
2137                                 }
2138
2139                                 if (memseg_list_alloc(msl)) {
2140                                         /* if we couldn't allocate VA space, we
2141                                          * can try with smaller page sizes.
2142                                          */
2143                                         RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list, retrying with different page size\n");
2144                                         /* deallocate memseg list */
2145                                         if (memseg_list_free(msl))
2146                                                 return -1;
2147                                         break;
2148                                 }
2149
2150                                 total_segs += msl->memseg_arr.len;
2151                                 cur_pagesz_mem = total_segs * hugepage_sz;
2152                                 type_msl_idx++;
2153                                 msl_idx++;
2154                         }
2155                         cur_socket_mem += cur_pagesz_mem;
2156                 }
2157                 if (cur_socket_mem == 0) {
2158                         RTE_LOG(ERR, EAL, "Cannot allocate VA space on socket %u\n",
2159                                 socket_id);
2160                         return -1;
2161                 }
2162         }
2163
2164         return 0;
2165 }
2166
2167 static int __rte_unused
2168 memseg_primary_init(void)
2169 {
2170         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
2171         struct memtype {
2172                 uint64_t page_sz;
2173                 int socket_id;
2174         } *memtypes = NULL;
2175         int i, hpi_idx, msl_idx, ret = -1; /* fail unless told to succeed */
2176         struct rte_memseg_list *msl;
2177         uint64_t max_mem, max_mem_per_type;
2178         unsigned int max_seglists_per_type;
2179         unsigned int n_memtypes, cur_type;
2180
2181         /* no-huge does not need this at all */
2182         if (internal_config.no_hugetlbfs)
2183                 return 0;
2184
2185         /*
2186          * figuring out amount of memory we're going to have is a long and very
2187          * involved process. the basic element we're operating with is a memory
2188          * type, defined as a combination of NUMA node ID and page size (so that
2189          * e.g. 2 sockets with 2 page sizes yield 4 memory types in total).
2190          *
2191          * deciding amount of memory going towards each memory type is a
2192          * balancing act between maximum segments per type, maximum memory per
2193          * type, and number of detected NUMA nodes. the goal is to make sure
2194          * each memory type gets at least one memseg list.
2195          *
2196          * the total amount of memory is limited by RTE_MAX_MEM_MB value.
2197          *
2198          * the total amount of memory per type is limited by either
2199          * RTE_MAX_MEM_MB_PER_TYPE, or by RTE_MAX_MEM_MB divided by the number
2200          * of detected NUMA nodes. additionally, maximum number of segments per
2201          * type is also limited by RTE_MAX_MEMSEG_PER_TYPE. this is because for
2202          * smaller page sizes, it can take hundreds of thousands of segments to
2203          * reach the above specified per-type memory limits.
2204          *
2205          * additionally, each type may have multiple memseg lists associated
2206          * with it, each limited by either RTE_MAX_MEM_MB_PER_LIST for bigger
2207          * page sizes, or RTE_MAX_MEMSEG_PER_LIST segments for smaller ones.
2208          *
2209          * the number of memseg lists per type is decided based on the above
2210          * limits, and also taking number of detected NUMA nodes, to make sure
2211          * that we don't run out of memseg lists before we populate all NUMA
2212          * nodes with memory.
2213          *
2214          * we do this in three stages. first, we collect the number of types.
2215          * then, we figure out memory constraints and populate the list of
2216          * would-be memseg lists. then, we go ahead and allocate the memseg
2217          * lists.
2218          */
2219
2220         /* create space for mem types */
2221         n_memtypes = internal_config.num_hugepage_sizes * rte_socket_count();
2222         memtypes = calloc(n_memtypes, sizeof(*memtypes));
2223         if (memtypes == NULL) {
2224                 RTE_LOG(ERR, EAL, "Cannot allocate space for memory types\n");
2225                 return -1;
2226         }
2227
2228         /* populate mem types */
2229         cur_type = 0;
2230         for (hpi_idx = 0; hpi_idx < (int) internal_config.num_hugepage_sizes;
2231                         hpi_idx++) {
2232                 struct hugepage_info *hpi;
2233                 uint64_t hugepage_sz;
2234
2235                 hpi = &internal_config.hugepage_info[hpi_idx];
2236                 hugepage_sz = hpi->hugepage_sz;
2237
2238                 for (i = 0; i < (int) rte_socket_count(); i++, cur_type++) {
2239                         int socket_id = rte_socket_id_by_idx(i);
2240
2241 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
2242                         /* we can still sort pages by socket in legacy mode */
2243                         if (!internal_config.legacy_mem && socket_id > 0)
2244                                 break;
2245 #endif
2246                         memtypes[cur_type].page_sz = hugepage_sz;
2247                         memtypes[cur_type].socket_id = socket_id;
2248
2249                         RTE_LOG(DEBUG, EAL, "Detected memory type: "
2250                                 "socket_id:%u hugepage_sz:%" PRIu64 "\n",
2251                                 socket_id, hugepage_sz);
2252                 }
2253         }
2254         /* number of memtypes could have been lower due to no NUMA support */
2255         n_memtypes = cur_type;
2256
2257         /* set up limits for types */
2258         max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
2259         max_mem_per_type = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_TYPE << 20,
2260                         max_mem / n_memtypes);
2261         /*
2262          * limit maximum number of segment lists per type to ensure there's
2263          * space for memseg lists for all NUMA nodes with all page sizes
2264          */
2265         max_seglists_per_type = RTE_MAX_MEMSEG_LISTS / n_memtypes;
2266
2267         if (max_seglists_per_type == 0) {
2268                 RTE_LOG(ERR, EAL, "Cannot accommodate all memory types, please increase %s\n",
2269                         RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2270                 goto out;
2271         }
2272
2273         /* go through all mem types and create segment lists */
2274         msl_idx = 0;
2275         for (cur_type = 0; cur_type < n_memtypes; cur_type++) {
2276                 unsigned int cur_seglist, n_seglists, n_segs;
2277                 unsigned int max_segs_per_type, max_segs_per_list;
2278                 struct memtype *type = &memtypes[cur_type];
2279                 uint64_t max_mem_per_list, pagesz;
2280                 int socket_id;
2281
2282                 pagesz = type->page_sz;
2283                 socket_id = type->socket_id;
2284
2285                 /*
2286                  * we need to create segment lists for this type. we must take
2287                  * into account the following things:
2288                  *
2289                  * 1. total amount of memory we can use for this memory type
2290                  * 2. total amount of memory per memseg list allowed
2291                  * 3. number of segments needed to fit the amount of memory
2292                  * 4. number of segments allowed per type
2293                  * 5. number of segments allowed per memseg list
2294                  * 6. number of memseg lists we are allowed to take up
2295                  */
2296
2297                 /* calculate how much segments we will need in total */
2298                 max_segs_per_type = max_mem_per_type / pagesz;
2299                 /* limit number of segments to maximum allowed per type */
2300                 max_segs_per_type = RTE_MIN(max_segs_per_type,
2301                                 (unsigned int)RTE_MAX_MEMSEG_PER_TYPE);
2302                 /* limit number of segments to maximum allowed per list */
2303                 max_segs_per_list = RTE_MIN(max_segs_per_type,
2304                                 (unsigned int)RTE_MAX_MEMSEG_PER_LIST);
2305
2306                 /* calculate how much memory we can have per segment list */
2307                 max_mem_per_list = RTE_MIN(max_segs_per_list * pagesz,
2308                                 (uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20);
2309
2310                 /* calculate how many segments each segment list will have */
2311                 n_segs = RTE_MIN(max_segs_per_list, max_mem_per_list / pagesz);
2312
2313                 /* calculate how many segment lists we can have */
2314                 n_seglists = RTE_MIN(max_segs_per_type / n_segs,
2315                                 max_mem_per_type / max_mem_per_list);
2316
2317                 /* limit number of segment lists according to our maximum */
2318                 n_seglists = RTE_MIN(n_seglists, max_seglists_per_type);
2319
2320                 RTE_LOG(DEBUG, EAL, "Creating %i segment lists: "
2321                                 "n_segs:%i socket_id:%i hugepage_sz:%" PRIu64 "\n",
2322                         n_seglists, n_segs, socket_id, pagesz);
2323
2324                 /* create all segment lists */
2325                 for (cur_seglist = 0; cur_seglist < n_seglists; cur_seglist++) {
2326                         if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
2327                                 RTE_LOG(ERR, EAL,
2328                                         "No more space in memseg lists, please increase %s\n",
2329                                         RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2330                                 goto out;
2331                         }
2332                         msl = &mcfg->memsegs[msl_idx++];
2333
2334                         if (memseg_list_init(msl, pagesz, n_segs,
2335                                         socket_id, cur_seglist))
2336                                 goto out;
2337
2338                         if (memseg_list_alloc(msl)) {
2339                                 RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list\n");
2340                                 goto out;
2341                         }
2342                 }
2343         }
2344         /* we're successful */
2345         ret = 0;
2346 out:
2347         free(memtypes);
2348         return ret;
2349 }
2350
2351 static int
2352 memseg_secondary_init(void)
2353 {
2354         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
2355         int msl_idx = 0;
2356         struct rte_memseg_list *msl;
2357
2358         for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
2359
2360                 msl = &mcfg->memsegs[msl_idx];
2361
2362                 /* skip empty memseg lists */
2363                 if (msl->memseg_arr.len == 0)
2364                         continue;
2365
2366                 if (rte_fbarray_attach(&msl->memseg_arr)) {
2367                         RTE_LOG(ERR, EAL, "Cannot attach to primary process memseg lists\n");
2368                         return -1;
2369                 }
2370
2371                 /* preallocate VA space */
2372                 if (memseg_list_alloc(msl)) {
2373                         RTE_LOG(ERR, EAL, "Cannot preallocate VA space for hugepage memory\n");
2374                         return -1;
2375                 }
2376         }
2377
2378         return 0;
2379 }
2380
2381 int
2382 rte_eal_memseg_init(void)
2383 {
2384         /* increase rlimit to maximum */
2385         struct rlimit lim;
2386
2387         if (getrlimit(RLIMIT_NOFILE, &lim) == 0) {
2388                 /* set limit to maximum */
2389                 lim.rlim_cur = lim.rlim_max;
2390
2391                 if (setrlimit(RLIMIT_NOFILE, &lim) < 0) {
2392                         RTE_LOG(DEBUG, EAL, "Setting maximum number of open files failed: %s\n",
2393                                         strerror(errno));
2394                 } else {
2395                         RTE_LOG(DEBUG, EAL, "Setting maximum number of open files to %"
2396                                         PRIu64 "\n",
2397                                         (uint64_t)lim.rlim_cur);
2398                 }
2399         } else {
2400                 RTE_LOG(ERR, EAL, "Cannot get current resource limits\n");
2401         }
2402 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
2403         if (!internal_config.legacy_mem && rte_socket_count() > 1) {
2404                 RTE_LOG(WARNING, EAL, "DPDK is running on a NUMA system, but is compiled without NUMA support.\n");
2405                 RTE_LOG(WARNING, EAL, "This will have adverse consequences for performance and usability.\n");
2406                 RTE_LOG(WARNING, EAL, "Please use --"OPT_LEGACY_MEM" option, or recompile with NUMA support.\n");
2407         }
2408 #endif
2409
2410         return rte_eal_process_type() == RTE_PROC_PRIMARY ?
2411 #ifndef RTE_ARCH_64
2412                         memseg_primary_init_32() :
2413 #else
2414                         memseg_primary_init() :
2415 #endif
2416                         memseg_secondary_init();
2417 }