lib/librte_eal/linuxapp/eal/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 #define _FILE_OFFSET_BITS 64
   7 #include <errno.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 RTE_EAL_NUMA_AWARE_HUGEPAGES
  28 #include <numa.h>
  29 #include <numaif.h>
  30 #endif
  31
  32 #include <rte_errno.h>
  33 #include <rte_log.h>
  34 #include <rte_memory.h>
  35 #include <rte_launch.h>
  36 #include <rte_eal.h>
  37 #include <rte_eal_memconfig.h>
  38 #include <rte_per_lcore.h>
  39 #include <rte_lcore.h>
  40 #include <rte_common.h>
  41 #include <rte_string_fns.h>
  42
  43 #include "eal_private.h"
  44 #include "eal_memalloc.h"
  45 #include "eal_internal_cfg.h"
  46 #include "eal_filesystem.h"
  47 #include "eal_hugepages.h"
  48
  49 #define PFN_MASK_SIZE   8
  50
  51 /**
  52  * @file
  53  * Huge page mapping under linux
  54  *
  55  * To reserve a big contiguous amount of memory, we use the hugepage
  56  * feature of linux. For that, we need to have hugetlbfs mounted. This
  57  * code will create many files in this directory (one per page) and
  58  * map them in virtual memory. For each page, we will retrieve its
  59  * physical address and remap it in order to have a virtual contiguous
  60  * zone as well as a physical contiguous zone.
  61  */
  62
  63 static bool phys_addrs_available = true;
  64
  65 #define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
  66
  67 static void
  68 test_phys_addrs_available(void)
  69 {
  70         uint64_t tmp = 0;
  71         phys_addr_t physaddr;
  72
  73         if (!rte_eal_has_hugepages()) {
  74                 RTE_LOG(ERR, EAL,
  75                         "Started without hugepages support, physical addresses not available\n");
  76                 phys_addrs_available = false;
  77                 return;
  78         }
  79
  80         physaddr = rte_mem_virt2phy(&tmp);
  81         if (physaddr == RTE_BAD_PHYS_ADDR) {
  82                 if (rte_eal_iova_mode() == RTE_IOVA_PA)
  83                         RTE_LOG(ERR, EAL,
  84                                 "Cannot obtain physical addresses: %s. "
  85                                 "Only vfio will function.\n",
  86                                 strerror(errno));
  87                 phys_addrs_available = false;
  88         }
  89 }
  90
  91 /*
  92  * Get physical address of any mapped virtual address in the current process.
  93  */
  94 phys_addr_t
  95 rte_mem_virt2phy(const void *virtaddr)
  96 {
  97         int fd, retval;
  98         uint64_t page, physaddr;
  99         unsigned long virt_pfn;
 100         int page_size;
 101         off_t offset;
 102
 103         /* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
 104         if (!phys_addrs_available)
 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(ERR, 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(ERR, 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(ERR, 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(ERR, 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 = numa_allocate_nodemask();
 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                 if (get_mempolicy(&oldpolicy, oldmask->maskp,
 280                                   oldmask->size + 1, 0, 0) < 0) {
 281                         RTE_LOG(ERR, EAL,
 282                                 "Failed to get current mempolicy: %s. "
 283                                 "Assuming MPOL_DEFAULT.\n", strerror(errno));
 284                         oldpolicy = MPOL_DEFAULT;
 285                 }
 286                 for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 287                         if (internal_config.socket_mem[i])
 288                                 maxnode = i + 1;
 289         }
 290 #endif
 291
 292         for (i = 0; i < hpi->num_pages[0]; i++) {
 293                 struct hugepage_file *hf = &hugepg_tbl[i];
 294                 uint64_t hugepage_sz = hpi->hugepage_sz;
 295
 296 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 297                 if (maxnode) {
 298                         unsigned int j;
 299
 300                         for (j = 0; j < maxnode; j++)
 301                                 if (essential_memory[j])
 302                                         break;
 303
 304                         if (j == maxnode) {
 305                                 node_id = (node_id + 1) % maxnode;
 306                                 while (!internal_config.socket_mem[node_id]) {
 307                                         node_id++;
 308                                         node_id %= maxnode;
 309                                 }
 310                                 essential_prev = 0;
 311                         } else {
 312                                 node_id = j;
 313                                 essential_prev = essential_memory[j];
 314
 315                                 if (essential_memory[j] < hugepage_sz)
 316                                         essential_memory[j] = 0;
 317                                 else
 318                                         essential_memory[j] -= hugepage_sz;
 319                         }
 320
 321                         RTE_LOG(DEBUG, EAL,
 322                                 "Setting policy MPOL_PREFERRED for socket %d\n",
 323                                 node_id);
 324                         numa_set_preferred(node_id);
 325                 }
 326 #endif
 327
 328                 hf->file_id = i;
 329                 hf->size = hugepage_sz;
 330                 eal_get_hugefile_path(hf->filepath, sizeof(hf->filepath),
 331                                 hpi->hugedir, hf->file_id);
 332                 hf->filepath[sizeof(hf->filepath) - 1] = '\0';
 333
 334                 /* try to create hugepage file */
 335                 fd = open(hf->filepath, O_CREAT | O_RDWR, 0600);
 336                 if (fd < 0) {
 337                         RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
 338                                         strerror(errno));
 339                         goto out;
 340                 }
 341
 342                 /* map the segment, and populate page tables,
 343                  * the kernel fills this segment with zeros. we don't care where
 344                  * this gets mapped - we already have contiguous memory areas
 345                  * ready for us to map into.
 346                  */
 347                 virtaddr = mmap(NULL, hugepage_sz, PROT_READ | PROT_WRITE,
 348                                 MAP_SHARED | MAP_POPULATE, fd, 0);
 349                 if (virtaddr == MAP_FAILED) {
 350                         RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
 351                                         strerror(errno));
 352                         close(fd);
 353                         goto out;
 354                 }
 355
 356                 hf->orig_va = virtaddr;
 357
 358                 /* In linux, hugetlb limitations, like cgroup, are
 359                  * enforced at fault time instead of mmap(), even
 360                  * with the option of MAP_POPULATE. Kernel will send
 361                  * a SIGBUS signal. To avoid to be killed, save stack
 362                  * environment here, if SIGBUS happens, we can jump
 363                  * back here.
 364                  */
 365                 if (huge_wrap_sigsetjmp()) {
 366                         RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
 367                                 "hugepages of size %u MB\n",
 368                                 (unsigned int)(hugepage_sz / 0x100000));
 369                         munmap(virtaddr, hugepage_sz);
 370                         close(fd);
 371                         unlink(hugepg_tbl[i].filepath);
 372 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 373                         if (maxnode)
 374                                 essential_memory[node_id] =
 375                                         essential_prev;
 376 #endif
 377                         goto out;
 378                 }
 379                 *(int *)virtaddr = 0;
 380
 381                 /* set shared lock on the file. */
 382                 if (flock(fd, LOCK_SH) < 0) {
 383                         RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
 384                                 __func__, strerror(errno));
 385                         close(fd);
 386                         goto out;
 387                 }
 388
 389                 close(fd);
 390         }
 391
 392 out:
 393 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 394         if (maxnode) {
 395                 RTE_LOG(DEBUG, EAL,
 396                         "Restoring previous memory policy: %d\n", oldpolicy);
 397                 if (oldpolicy == MPOL_DEFAULT) {
 398                         numa_set_localalloc();
 399                 } else if (set_mempolicy(oldpolicy, oldmask->maskp,
 400                                          oldmask->size + 1) < 0) {
 401                         RTE_LOG(ERR, EAL, "Failed to restore mempolicy: %s\n",
 402                                 strerror(errno));
 403                         numa_set_localalloc();
 404                 }
 405         }
 406         numa_free_cpumask(oldmask);
 407 #endif
 408         return i;
 409 }
 410
 411 /*
 412  * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
 413  * page.
 414  */
 415 static int
 416 find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 417 {
 418         int socket_id;
 419         char *end, *nodestr;
 420         unsigned i, hp_count = 0;
 421         uint64_t virt_addr;
 422         char buf[BUFSIZ];
 423         char hugedir_str[PATH_MAX];
 424         FILE *f;
 425
 426         f = fopen("/proc/self/numa_maps", "r");
 427         if (f == NULL) {
 428                 RTE_LOG(NOTICE, EAL, "NUMA support not available"
 429                         " consider that all memory is in socket_id 0\n");
 430                 return 0;
 431         }
 432
 433         snprintf(hugedir_str, sizeof(hugedir_str),
 434                         "%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
 435
 436         /* parse numa map */
 437         while (fgets(buf, sizeof(buf), f) != NULL) {
 438
 439                 /* ignore non huge page */
 440                 if (strstr(buf, " huge ") == NULL &&
 441                                 strstr(buf, hugedir_str) == NULL)
 442                         continue;
 443
 444                 /* get zone addr */
 445                 virt_addr = strtoull(buf, &end, 16);
 446                 if (virt_addr == 0 || end == buf) {
 447                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 448                         goto error;
 449                 }
 450
 451                 /* get node id (socket id) */
 452                 nodestr = strstr(buf, " N");
 453                 if (nodestr == NULL) {
 454                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 455                         goto error;
 456                 }
 457                 nodestr += 2;
 458                 end = strstr(nodestr, "=");
 459                 if (end == NULL) {
 460                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 461                         goto error;
 462                 }
 463                 end[0] = '\0';
 464                 end = NULL;
 465
 466                 socket_id = strtoul(nodestr, &end, 0);
 467                 if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
 468                         RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
 469                         goto error;
 470                 }
 471
 472                 /* if we find this page in our mappings, set socket_id */
 473                 for (i = 0; i < hpi->num_pages[0]; i++) {
 474                         void *va = (void *)(unsigned long)virt_addr;
 475                         if (hugepg_tbl[i].orig_va == va) {
 476                                 hugepg_tbl[i].socket_id = socket_id;
 477                                 hp_count++;
 478 #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
 479                                 RTE_LOG(DEBUG, EAL,
 480                                         "Hugepage %s is on socket %d\n",
 481                                         hugepg_tbl[i].filepath, socket_id);
 482 #endif
 483                         }
 484                 }
 485         }
 486
 487         if (hp_count < hpi->num_pages[0])
 488                 goto error;
 489
 490         fclose(f);
 491         return 0;
 492
 493 error:
 494         fclose(f);
 495         return -1;
 496 }
 497
 498 static int
 499 cmp_physaddr(const void *a, const void *b)
 500 {
 501 #ifndef RTE_ARCH_PPC_64
 502         const struct hugepage_file *p1 = a;
 503         const struct hugepage_file *p2 = b;
 504 #else
 505         /* PowerPC needs memory sorted in reverse order from x86 */
 506         const struct hugepage_file *p1 = b;
 507         const struct hugepage_file *p2 = a;
 508 #endif
 509         if (p1->physaddr < p2->physaddr)
 510                 return -1;
 511         else if (p1->physaddr > p2->physaddr)
 512                 return 1;
 513         else
 514                 return 0;
 515 }
 516
 517 /*
 518  * Uses mmap to create a shared memory area for storage of data
 519  * Used in this file to store the hugepage file map on disk
 520  */
 521 static void *
 522 create_shared_memory(const char *filename, const size_t mem_size)
 523 {
 524         void *retval;
 525         int fd;
 526
 527         /* if no shared files mode is used, create anonymous memory instead */
 528         if (internal_config.no_shconf) {
 529                 retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE,
 530                                 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 531                 if (retval == MAP_FAILED)
 532                         return NULL;
 533                 return retval;
 534         }
 535
 536         fd = open(filename, O_CREAT | O_RDWR, 0666);
 537         if (fd < 0)
 538                 return NULL;
 539         if (ftruncate(fd, mem_size) < 0) {
 540                 close(fd);
 541                 return NULL;
 542         }
 543         retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 544         close(fd);
 545         if (retval == MAP_FAILED)
 546                 return NULL;
 547         return retval;
 548 }
 549
 550 /*
 551  * this copies *active* hugepages from one hugepage table to another.
 552  * destination is typically the shared memory.
 553  */
 554 static int
 555 copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
 556                 const struct hugepage_file * src, int src_size)
 557 {
 558         int src_pos, dst_pos = 0;
 559
 560         for (src_pos = 0; src_pos < src_size; src_pos++) {
 561                 if (src[src_pos].orig_va != NULL) {
 562                         /* error on overflow attempt */
 563                         if (dst_pos == dest_size)
 564                                 return -1;
 565                         memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
 566                         dst_pos++;
 567                 }
 568         }
 569         return 0;
 570 }
 571
 572 static int
 573 unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
 574                 unsigned num_hp_info)
 575 {
 576         unsigned socket, size;
 577         int page, nrpages = 0;
 578
 579         /* get total number of hugepages */
 580         for (size = 0; size < num_hp_info; size++)
 581                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 582                         nrpages +=
 583                         internal_config.hugepage_info[size].num_pages[socket];
 584
 585         for (page = 0; page < nrpages; page++) {
 586                 struct hugepage_file *hp = &hugepg_tbl[page];
 587
 588                 if (hp->final_va != NULL && unlink(hp->filepath)) {
 589                         RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
 590                                 __func__, hp->filepath, strerror(errno));
 591                 }
 592         }
 593         return 0;
 594 }
 595
 596 /*
 597  * unmaps hugepages that are not going to be used. since we originally allocate
 598  * ALL hugepages (not just those we need), additional unmapping needs to be done.
 599  */
 600 static int
 601 unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
 602                 struct hugepage_info *hpi,
 603                 unsigned num_hp_info)
 604 {
 605         unsigned socket, size;
 606         int page, nrpages = 0;
 607
 608         /* get total number of hugepages */
 609         for (size = 0; size < num_hp_info; size++)
 610                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
 611                         nrpages += internal_config.hugepage_info[size].num_pages[socket];
 612
 613         for (size = 0; size < num_hp_info; size++) {
 614                 for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
 615                         unsigned pages_found = 0;
 616
 617                         /* traverse until we have unmapped all the unused pages */
 618                         for (page = 0; page < nrpages; page++) {
 619                                 struct hugepage_file *hp = &hugepg_tbl[page];
 620
 621                                 /* find a page that matches the criteria */
 622                                 if ((hp->size == hpi[size].hugepage_sz) &&
 623                                                 (hp->socket_id == (int) socket)) {
 624
 625                                         /* if we skipped enough pages, unmap the rest */
 626                                         if (pages_found == hpi[size].num_pages[socket]) {
 627                                                 uint64_t unmap_len;
 628
 629                                                 unmap_len = hp->size;
 630
 631                                                 /* get start addr and len of the remaining segment */
 632                                                 munmap(hp->orig_va,
 633                                                         (size_t)unmap_len);
 634
 635                                                 hp->orig_va = NULL;
 636                                                 if (unlink(hp->filepath) == -1) {
 637                                                         RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
 638                                                                         __func__, hp->filepath, strerror(errno));
 639                                                         return -1;
 640                                                 }
 641                                         } else {
 642                                                 /* lock the page and skip */
 643                                                 pages_found++;
 644                                         }
 645
 646                                 } /* match page */
 647                         } /* foreach page */
 648                 } /* foreach socket */
 649         } /* foreach pagesize */
 650
 651         return 0;
 652 }
 653
 654 static int
 655 remap_segment(struct hugepage_file *hugepages, int seg_start, int seg_end)
 656 {
 657         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 658         struct rte_memseg_list *msl;
 659         struct rte_fbarray *arr;
 660         int cur_page, seg_len;
 661         unsigned int msl_idx;
 662         int ms_idx;
 663         uint64_t page_sz;
 664         size_t memseg_len;
 665         int socket_id;
 666
 667         page_sz = hugepages[seg_start].size;
 668         socket_id = hugepages[seg_start].socket_id;
 669         seg_len = seg_end - seg_start;
 670
 671         RTE_LOG(DEBUG, EAL, "Attempting to map %" PRIu64 "M on socket %i\n",
 672                         (seg_len * page_sz) >> 20ULL, socket_id);
 673
 674         /* find free space in memseg lists */
 675         for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
 676                 bool empty;
 677                 msl = &mcfg->memsegs[msl_idx];
 678                 arr = &msl->memseg_arr;
 679
 680                 if (msl->page_sz != page_sz)
 681                         continue;
 682                 if (msl->socket_id != socket_id)
 683                         continue;
 684
 685                 /* leave space for a hole if array is not empty */
 686                 empty = arr->count == 0;
 687                 ms_idx = rte_fbarray_find_next_n_free(arr, 0,
 688                                 seg_len + (empty ? 0 : 1));
 689
 690                 /* memseg list is full? */
 691                 if (ms_idx < 0)
 692                         continue;
 693
 694                 /* leave some space between memsegs, they are not IOVA
 695                  * contiguous, so they shouldn't be VA contiguous either.
 696                  */
 697                 if (!empty)
 698                         ms_idx++;
 699                 break;
 700         }
 701         if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
 702                 RTE_LOG(ERR, EAL, "Could not find space for memseg. Please increase %s and/or %s in configuration.\n",
 703                                 RTE_STR(CONFIG_RTE_MAX_MEMSEG_PER_TYPE),
 704                                 RTE_STR(CONFIG_RTE_MAX_MEM_PER_TYPE));
 705                 return -1;
 706         }
 707
 708 #ifdef RTE_ARCH_PPC64
 709         /* for PPC64 we go through the list backwards */
 710         for (cur_page = seg_end - 1; cur_page >= seg_start;
 711                         cur_page--, ms_idx++) {
 712 #else
 713         for (cur_page = seg_start; cur_page < seg_end; cur_page++, ms_idx++) {
 714 #endif
 715                 struct hugepage_file *hfile = &hugepages[cur_page];
 716                 struct rte_memseg *ms = rte_fbarray_get(arr, ms_idx);
 717                 void *addr;
 718                 int fd;
 719
 720                 fd = open(hfile->filepath, O_RDWR);
 721                 if (fd < 0) {
 722                         RTE_LOG(ERR, EAL, "Could not open '%s': %s\n",
 723                                         hfile->filepath, strerror(errno));
 724                         return -1;
 725                 }
 726                 /* set shared lock on the file. */
 727                 if (flock(fd, LOCK_SH) < 0) {
 728                         RTE_LOG(DEBUG, EAL, "Could not lock '%s': %s\n",
 729                                         hfile->filepath, strerror(errno));
 730                         close(fd);
 731                         return -1;
 732                 }
 733                 memseg_len = (size_t)page_sz;
 734                 addr = RTE_PTR_ADD(msl->base_va, ms_idx * memseg_len);
 735
 736                 /* we know this address is already mmapped by memseg list, so
 737                  * using MAP_FIXED here is safe
 738                  */
 739                 addr = mmap(addr, page_sz, PROT_READ | PROT_WRITE,
 740                                 MAP_SHARED | MAP_POPULATE | MAP_FIXED, fd, 0);
 741                 if (addr == MAP_FAILED) {
 742                         RTE_LOG(ERR, EAL, "Couldn't remap '%s': %s\n",
 743                                         hfile->filepath, strerror(errno));
 744                         close(fd);
 745                         return -1;
 746                 }
 747
 748                 /* we have a new address, so unmap previous one */
 749 #ifndef RTE_ARCH_64
 750                 /* in 32-bit legacy mode, we have already unmapped the page */
 751                 if (!internal_config.legacy_mem)
 752                         munmap(hfile->orig_va, page_sz);
 753 #else
 754                 munmap(hfile->orig_va, page_sz);
 755 #endif
 756
 757                 hfile->orig_va = NULL;
 758                 hfile->final_va = addr;
 759
 760                 /* rewrite physical addresses in IOVA as VA mode */
 761                 if (rte_eal_iova_mode() == RTE_IOVA_VA)
 762                         hfile->physaddr = (uintptr_t)addr;
 763
 764                 /* set up memseg data */
 765                 ms->addr = addr;
 766                 ms->hugepage_sz = page_sz;
 767                 ms->len = memseg_len;
 768                 ms->iova = hfile->physaddr;
 769                 ms->socket_id = hfile->socket_id;
 770                 ms->nchannel = rte_memory_get_nchannel();
 771                 ms->nrank = rte_memory_get_nrank();
 772
 773                 rte_fbarray_set_used(arr, ms_idx);
 774
 775                 close(fd);
 776         }
 777         RTE_LOG(DEBUG, EAL, "Allocated %" PRIu64 "M on socket %i\n",
 778                         (seg_len * page_sz) >> 20, socket_id);
 779         return 0;
 780 }
 781
 782 static uint64_t
 783 get_mem_amount(uint64_t page_sz, uint64_t max_mem)
 784 {
 785         uint64_t area_sz, max_pages;
 786
 787         /* limit to RTE_MAX_MEMSEG_PER_LIST pages or RTE_MAX_MEM_MB_PER_LIST */
 788         max_pages = RTE_MAX_MEMSEG_PER_LIST;
 789         max_mem = RTE_MIN((uint64_t)RTE_MAX_MEM_MB_PER_LIST << 20, max_mem);
 790
 791         area_sz = RTE_MIN(page_sz * max_pages, max_mem);
 792
 793         /* make sure the list isn't smaller than the page size */
 794         area_sz = RTE_MAX(area_sz, page_sz);
 795
 796         return RTE_ALIGN(area_sz, page_sz);
 797 }
 798
 799 static int
 800 free_memseg_list(struct rte_memseg_list *msl)
 801 {
 802         if (rte_fbarray_destroy(&msl->memseg_arr)) {
 803                 RTE_LOG(ERR, EAL, "Cannot destroy memseg list\n");
 804                 return -1;
 805         }
 806         memset(msl, 0, sizeof(*msl));
 807         return 0;
 808 }
 809
 810 #define MEMSEG_LIST_FMT "memseg-%" PRIu64 "k-%i-%i"
 811 static int
 812 alloc_memseg_list(struct rte_memseg_list *msl, uint64_t page_sz,
 813                 int n_segs, int socket_id, int type_msl_idx)
 814 {
 815         char name[RTE_FBARRAY_NAME_LEN];
 816
 817         snprintf(name, sizeof(name), MEMSEG_LIST_FMT, page_sz >> 10, socket_id,
 818                  type_msl_idx);
 819         if (rte_fbarray_init(&msl->memseg_arr, name, n_segs,
 820                         sizeof(struct rte_memseg))) {
 821                 RTE_LOG(ERR, EAL, "Cannot allocate memseg list: %s\n",
 822                         rte_strerror(rte_errno));
 823                 return -1;
 824         }
 825
 826         msl->page_sz = page_sz;
 827         msl->socket_id = socket_id;
 828         msl->base_va = NULL;
 829
 830         RTE_LOG(DEBUG, EAL, "Memseg list allocated: 0x%zxkB at socket %i\n",
 831                         (size_t)page_sz >> 10, socket_id);
 832
 833         return 0;
 834 }
 835
 836 static int
 837 alloc_va_space(struct rte_memseg_list *msl)
 838 {
 839         uint64_t page_sz;
 840         size_t mem_sz;
 841         void *addr;
 842         int flags = 0;
 843
 844 #ifdef RTE_ARCH_PPC_64
 845         flags |= MAP_HUGETLB;
 846 #endif
 847
 848         page_sz = msl->page_sz;
 849         mem_sz = page_sz * msl->memseg_arr.len;
 850
 851         addr = eal_get_virtual_area(msl->base_va, &mem_sz, page_sz, 0, flags);
 852         if (addr == NULL) {
 853                 if (rte_errno == EADDRNOTAVAIL)
 854                         RTE_LOG(ERR, EAL, "Could not mmap %llu bytes at [%p] - please use '--base-virtaddr' option\n",
 855                                 (unsigned long long)mem_sz, msl->base_va);
 856                 else
 857                         RTE_LOG(ERR, EAL, "Cannot reserve memory\n");
 858                 return -1;
 859         }
 860         msl->base_va = addr;
 861
 862         return 0;
 863 }
 864
 865 /*
 866  * Our VA space is not preallocated yet, so preallocate it here. We need to know
 867  * how many segments there are in order to map all pages into one address space,
 868  * and leave appropriate holes between segments so that rte_malloc does not
 869  * concatenate them into one big segment.
 870  *
 871  * we also need to unmap original pages to free up address space.
 872  */
 873 static int __rte_unused
 874 prealloc_segments(struct hugepage_file *hugepages, int n_pages)
 875 {
 876         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 877         int cur_page, seg_start_page, end_seg, new_memseg;
 878         unsigned int hpi_idx, socket, i;
 879         int n_contig_segs, n_segs;
 880         int msl_idx;
 881
 882         /* before we preallocate segments, we need to free up our VA space.
 883          * we're not removing files, and we already have information about
 884          * PA-contiguousness, so it is safe to unmap everything.
 885          */
 886         for (cur_page = 0; cur_page < n_pages; cur_page++) {
 887                 struct hugepage_file *hpi = &hugepages[cur_page];
 888                 munmap(hpi->orig_va, hpi->size);
 889                 hpi->orig_va = NULL;
 890         }
 891
 892         /* we cannot know how many page sizes and sockets we have discovered, so
 893          * loop over all of them
 894          */
 895         for (hpi_idx = 0; hpi_idx < internal_config.num_hugepage_sizes;
 896                         hpi_idx++) {
 897                 uint64_t page_sz =
 898                         internal_config.hugepage_info[hpi_idx].hugepage_sz;
 899
 900                 for (i = 0; i < rte_socket_count(); i++) {
 901                         struct rte_memseg_list *msl;
 902
 903                         socket = rte_socket_id_by_idx(i);
 904                         n_contig_segs = 0;
 905                         n_segs = 0;
 906                         seg_start_page = -1;
 907
 908                         for (cur_page = 0; cur_page < n_pages; cur_page++) {
 909                                 struct hugepage_file *prev, *cur;
 910                                 int prev_seg_start_page = -1;
 911
 912                                 cur = &hugepages[cur_page];
 913                                 prev = cur_page == 0 ? NULL :
 914                                                 &hugepages[cur_page - 1];
 915
 916                                 new_memseg = 0;
 917                                 end_seg = 0;
 918
 919                                 if (cur->size == 0)
 920                                         end_seg = 1;
 921                                 else if (cur->socket_id != (int) socket)
 922                                         end_seg = 1;
 923                                 else if (cur->size != page_sz)
 924                                         end_seg = 1;
 925                                 else if (cur_page == 0)
 926                                         new_memseg = 1;
 927 #ifdef RTE_ARCH_PPC_64
 928                                 /* On PPC64 architecture, the mmap always start
 929                                  * from higher address to lower address. Here,
 930                                  * physical addresses are in descending order.
 931                                  */
 932                                 else if ((prev->physaddr - cur->physaddr) !=
 933                                                 cur->size)
 934                                         new_memseg = 1;
 935 #else
 936                                 else if ((cur->physaddr - prev->physaddr) !=
 937                                                 cur->size)
 938                                         new_memseg = 1;
 939 #endif
 940                                 if (new_memseg) {
 941                                         /* if we're already inside a segment,
 942                                          * new segment means end of current one
 943                                          */
 944                                         if (seg_start_page != -1) {
 945                                                 end_seg = 1;
 946                                                 prev_seg_start_page =
 947                                                                 seg_start_page;
 948                                         }
 949                                         seg_start_page = cur_page;
 950                                 }
 951
 952                                 if (end_seg) {
 953                                         if (prev_seg_start_page != -1) {
 954                                                 /* we've found a new segment */
 955                                                 n_contig_segs++;
 956                                                 n_segs += cur_page -
 957                                                         prev_seg_start_page;
 958                                         } else if (seg_start_page != -1) {
 959                                                 /* we didn't find new segment,
 960                                                  * but did end current one
 961                                                  */
 962                                                 n_contig_segs++;
 963                                                 n_segs += cur_page -
 964                                                                 seg_start_page;
 965                                                 seg_start_page = -1;
 966                                                 continue;
 967                                         } else {
 968                                                 /* we're skipping this page */
 969                                                 continue;
 970                                         }
 971                                 }
 972                                 /* segment continues */
 973                         }
 974                         /* check if we missed last segment */
 975                         if (seg_start_page != -1) {
 976                                 n_contig_segs++;
 977                                 n_segs += cur_page - seg_start_page;
 978                         }
 979
 980                         /* if no segments were found, do not preallocate */
 981                         if (n_segs == 0)
 982                                 continue;
 983
 984                         /* we now have total number of pages that we will
 985                          * allocate for this segment list. add separator pages
 986                          * to the total count, and preallocate VA space.
 987                          */
 988                         n_segs += n_contig_segs - 1;
 989
 990                         /* now, preallocate VA space for these segments */
 991
 992                         /* first, find suitable memseg list for this */
 993                         for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS;
 994                                         msl_idx++) {
 995                                 msl = &mcfg->memsegs[msl_idx];
 996
 997                                 if (msl->base_va != NULL)
 998                                         continue;
 999                                 break;
1000                         }
1001                         if (msl_idx == RTE_MAX_MEMSEG_LISTS) {
1002                                 RTE_LOG(ERR, EAL, "Not enough space in memseg lists, please increase %s\n",
1003                                         RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
1004                                 return -1;
1005                         }
1006
1007                         /* now, allocate fbarray itself */
1008                         if (alloc_memseg_list(msl, page_sz, n_segs, socket,
1009                                                 msl_idx) < 0)
1010                                 return -1;
1011
1012                         /* finally, allocate VA space */
1013                         if (alloc_va_space(msl) < 0)
1014                                 return -1;
1015                 }
1016         }
1017         return 0;
1018 }
1019
1020 /*
1021  * We cannot reallocate memseg lists on the fly because PPC64 stores pages
1022  * backwards, therefore we have to process the entire memseg first before
1023  * remapping it into memseg list VA space.
1024  */
1025 static int
1026 remap_needed_hugepages(struct hugepage_file *hugepages, int n_pages)
1027 {
1028         int cur_page, seg_start_page, new_memseg, ret;
1029
1030         seg_start_page = 0;
1031         for (cur_page = 0; cur_page < n_pages; cur_page++) {
1032                 struct hugepage_file *prev, *cur;
1033
1034                 new_memseg = 0;
1035
1036                 cur = &hugepages[cur_page];
1037                 prev = cur_page == 0 ? NULL : &hugepages[cur_page - 1];
1038
1039                 /* if size is zero, no more pages left */
1040                 if (cur->size == 0)
1041                         break;
1042
1043                 if (cur_page == 0)
1044                         new_memseg = 1;
1045                 else if (cur->socket_id != prev->socket_id)
1046                         new_memseg = 1;
1047                 else if (cur->size != prev->size)
1048                         new_memseg = 1;
1049 #ifdef RTE_ARCH_PPC_64
1050                 /* On PPC64 architecture, the mmap always start from higher
1051                  * address to lower address. Here, physical addresses are in
1052                  * descending order.
1053                  */
1054                 else if ((prev->physaddr - cur->physaddr) != cur->size)
1055                         new_memseg = 1;
1056 #else
1057                 else if ((cur->physaddr - prev->physaddr) != cur->size)
1058                         new_memseg = 1;
1059 #endif
1060
1061                 if (new_memseg) {
1062                         /* if this isn't the first time, remap segment */
1063                         if (cur_page != 0) {
1064                                 ret = remap_segment(hugepages, seg_start_page,
1065                                                 cur_page);
1066                                 if (ret != 0)
1067                                         return -1;
1068                         }
1069                         /* remember where we started */
1070                         seg_start_page = cur_page;
1071                 }
1072                 /* continuation of previous memseg */
1073         }
1074         /* we were stopped, but we didn't remap the last segment, do it now */
1075         if (cur_page != 0) {
1076                 ret = remap_segment(hugepages, seg_start_page,
1077                                 cur_page);
1078                 if (ret != 0)
1079                         return -1;
1080         }
1081         return 0;
1082 }
1083
1084 static inline uint64_t
1085 get_socket_mem_size(int socket)
1086 {
1087         uint64_t size = 0;
1088         unsigned i;
1089
1090         for (i = 0; i < internal_config.num_hugepage_sizes; i++){
1091                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
1092                 size += hpi->hugepage_sz * hpi->num_pages[socket];
1093         }
1094
1095         return size;
1096 }
1097
1098 /*
1099  * This function is a NUMA-aware equivalent of calc_num_pages.
1100  * It takes in the list of hugepage sizes and the
1101  * number of pages thereof, and calculates the best number of
1102  * pages of each size to fulfill the request for <memory> ram
1103  */
1104 static int
1105 calc_num_pages_per_socket(uint64_t * memory,
1106                 struct hugepage_info *hp_info,
1107                 struct hugepage_info *hp_used,
1108                 unsigned num_hp_info)
1109 {
1110         unsigned socket, j, i = 0;
1111         unsigned requested, available;
1112         int total_num_pages = 0;
1113         uint64_t remaining_mem, cur_mem;
1114         uint64_t total_mem = internal_config.memory;
1115
1116         if (num_hp_info == 0)
1117                 return -1;
1118
1119         /* if specific memory amounts per socket weren't requested */
1120         if (internal_config.force_sockets == 0) {
1121                 size_t total_size;
1122 #ifdef RTE_ARCH_64
1123                 int cpu_per_socket[RTE_MAX_NUMA_NODES];
1124                 size_t default_size;
1125                 unsigned lcore_id;
1126
1127                 /* Compute number of cores per socket */
1128                 memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
1129                 RTE_LCORE_FOREACH(lcore_id) {
1130                         cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
1131                 }
1132
1133                 /*
1134                  * Automatically spread requested memory amongst detected sockets according
1135                  * to number of cores from cpu mask present on each socket
1136                  */
1137                 total_size = internal_config.memory;
1138                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
1139
1140                         /* Set memory amount per socket */
1141                         default_size = (internal_config.memory * cpu_per_socket[socket])
1142                                         / rte_lcore_count();
1143
1144                         /* Limit to maximum available memory on socket */
1145                         default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
1146
1147                         /* Update sizes */
1148                         memory[socket] = default_size;
1149                         total_size -= default_size;
1150                 }
1151
1152                 /*
1153                  * If some memory is remaining, try to allocate it by getting all
1154                  * available memory from sockets, one after the other
1155                  */
1156                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
1157                         /* take whatever is available */
1158                         default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
1159                                                total_size);
1160
1161                         /* Update sizes */
1162                         memory[socket] += default_size;
1163                         total_size -= default_size;
1164                 }
1165 #else
1166                 /* in 32-bit mode, allocate all of the memory only on master
1167                  * lcore socket
1168                  */
1169                 total_size = internal_config.memory;
1170                 for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0;
1171                                 socket++) {
1172                         struct rte_config *cfg = rte_eal_get_configuration();
1173                         unsigned int master_lcore_socket;
1174
1175                         master_lcore_socket =
1176                                 rte_lcore_to_socket_id(cfg->master_lcore);
1177
1178                         if (master_lcore_socket != socket)
1179                                 continue;
1180
1181                         /* Update sizes */
1182                         memory[socket] = total_size;
1183                         break;
1184                 }
1185 #endif
1186         }
1187
1188         for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
1189                 /* skips if the memory on specific socket wasn't requested */
1190                 for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
1191                         strlcpy(hp_used[i].hugedir, hp_info[i].hugedir,
1192                                 sizeof(hp_used[i].hugedir));
1193                         hp_used[i].num_pages[socket] = RTE_MIN(
1194                                         memory[socket] / hp_info[i].hugepage_sz,
1195                                         hp_info[i].num_pages[socket]);
1196
1197                         cur_mem = hp_used[i].num_pages[socket] *
1198                                         hp_used[i].hugepage_sz;
1199
1200                         memory[socket] -= cur_mem;
1201                         total_mem -= cur_mem;
1202
1203                         total_num_pages += hp_used[i].num_pages[socket];
1204
1205                         /* check if we have met all memory requests */
1206                         if (memory[socket] == 0)
1207                                 break;
1208
1209                         /* check if we have any more pages left at this size, if so
1210                          * move on to next size */
1211                         if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
1212                                 continue;
1213                         /* At this point we know that there are more pages available that are
1214                          * bigger than the memory we want, so lets see if we can get enough
1215                          * from other page sizes.
1216                          */
1217                         remaining_mem = 0;
1218                         for (j = i+1; j < num_hp_info; j++)
1219                                 remaining_mem += hp_info[j].hugepage_sz *
1220                                 hp_info[j].num_pages[socket];
1221
1222                         /* is there enough other memory, if not allocate another page and quit */
1223                         if (remaining_mem < memory[socket]){
1224                                 cur_mem = RTE_MIN(memory[socket],
1225                                                 hp_info[i].hugepage_sz);
1226                                 memory[socket] -= cur_mem;
1227                                 total_mem -= cur_mem;
1228                                 hp_used[i].num_pages[socket]++;
1229                                 total_num_pages++;
1230                                 break; /* we are done with this socket*/
1231                         }
1232                 }
1233                 /* if we didn't satisfy all memory requirements per socket */
1234                 if (memory[socket] > 0 &&
1235                                 internal_config.socket_mem[socket] != 0) {
1236                         /* to prevent icc errors */
1237                         requested = (unsigned) (internal_config.socket_mem[socket] /
1238                                         0x100000);
1239                         available = requested -
1240                                         ((unsigned) (memory[socket] / 0x100000));
1241                         RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
1242                                         "Requested: %uMB, available: %uMB\n", socket,
1243                                         requested, available);
1244                         return -1;
1245                 }
1246         }
1247
1248         /* if we didn't satisfy total memory requirements */
1249         if (total_mem > 0) {
1250                 requested = (unsigned) (internal_config.memory / 0x100000);
1251                 available = requested - (unsigned) (total_mem / 0x100000);
1252                 RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
1253                                 " available: %uMB\n", requested, available);
1254                 return -1;
1255         }
1256         return total_num_pages;
1257 }
1258
1259 static inline size_t
1260 eal_get_hugepage_mem_size(void)
1261 {
1262         uint64_t size = 0;
1263         unsigned i, j;
1264
1265         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
1266                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
1267                 if (strnlen(hpi->hugedir, sizeof(hpi->hugedir)) != 0) {
1268                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1269                                 size += hpi->hugepage_sz * hpi->num_pages[j];
1270                         }
1271                 }
1272         }
1273
1274         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
1275 }
1276
1277 static struct sigaction huge_action_old;
1278 static int huge_need_recover;
1279
1280 static void
1281 huge_register_sigbus(void)
1282 {
1283         sigset_t mask;
1284         struct sigaction action;
1285
1286         sigemptyset(&mask);
1287         sigaddset(&mask, SIGBUS);
1288         action.sa_flags = 0;
1289         action.sa_mask = mask;
1290         action.sa_handler = huge_sigbus_handler;
1291
1292         huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
1293 }
1294
1295 static void
1296 huge_recover_sigbus(void)
1297 {
1298         if (huge_need_recover) {
1299                 sigaction(SIGBUS, &huge_action_old, NULL);
1300                 huge_need_recover = 0;
1301         }
1302 }
1303
1304 /*
1305  * Prepare physical memory mapping: fill configuration structure with
1306  * these infos, return 0 on success.
1307  *  1. map N huge pages in separate files in hugetlbfs
1308  *  2. find associated physical addr
1309  *  3. find associated NUMA socket ID
1310  *  4. sort all huge pages by physical address
1311  *  5. remap these N huge pages in the correct order
1312  *  6. unmap the first mapping
1313  *  7. fill memsegs in configuration with contiguous zones
1314  */
1315 static int
1316 eal_legacy_hugepage_init(void)
1317 {
1318         struct rte_mem_config *mcfg;
1319         struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1320         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1321         struct rte_fbarray *arr;
1322         struct rte_memseg *ms;
1323
1324         uint64_t memory[RTE_MAX_NUMA_NODES];
1325
1326         unsigned hp_offset;
1327         int i, j;
1328         int nr_hugefiles, nr_hugepages = 0;
1329         void *addr;
1330
1331         test_phys_addrs_available();
1332
1333         memset(used_hp, 0, sizeof(used_hp));
1334
1335         /* get pointer to global configuration */
1336         mcfg = rte_eal_get_configuration()->mem_config;
1337
1338         /* hugetlbfs can be disabled */
1339         if (internal_config.no_hugetlbfs) {
1340                 struct rte_memseg_list *msl;
1341                 uint64_t page_sz;
1342                 int n_segs, cur_seg;
1343
1344                 /* nohuge mode is legacy mode */
1345                 internal_config.legacy_mem = 1;
1346
1347                 /* create a memseg list */
1348                 msl = &mcfg->memsegs[0];
1349
1350                 page_sz = RTE_PGSIZE_4K;
1351                 n_segs = internal_config.memory / page_sz;
1352
1353                 if (rte_fbarray_init(&msl->memseg_arr, "nohugemem", n_segs,
1354                                         sizeof(struct rte_memseg))) {
1355                         RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
1356                         return -1;
1357                 }
1358
1359                 addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1360                                 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
1361                 if (addr == MAP_FAILED) {
1362                         RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1363                                         strerror(errno));
1364                         return -1;
1365                 }
1366                 msl->base_va = addr;
1367                 msl->page_sz = page_sz;
1368                 msl->socket_id = 0;
1369
1370                 /* populate memsegs. each memseg is one page long */
1371                 for (cur_seg = 0; cur_seg < n_segs; cur_seg++) {
1372                         arr = &msl->memseg_arr;
1373
1374                         ms = rte_fbarray_get(arr, cur_seg);
1375                         if (rte_eal_iova_mode() == RTE_IOVA_VA)
1376                                 ms->iova = (uintptr_t)addr;
1377                         else
1378                                 ms->iova = RTE_BAD_IOVA;
1379                         ms->addr = addr;
1380                         ms->hugepage_sz = page_sz;
1381                         ms->socket_id = 0;
1382                         ms->len = page_sz;
1383
1384                         rte_fbarray_set_used(arr, cur_seg);
1385
1386                         addr = RTE_PTR_ADD(addr, (size_t)page_sz);
1387                 }
1388                 return 0;
1389         }
1390
1391         /* calculate total number of hugepages available. at this point we haven't
1392          * yet started sorting them so they all are on socket 0 */
1393         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1394                 /* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1395                 used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1396
1397                 nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1398         }
1399
1400         /*
1401          * allocate a memory area for hugepage table.
1402          * this isn't shared memory yet. due to the fact that we need some
1403          * processing done on these pages, shared memory will be created
1404          * at a later stage.
1405          */
1406         tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1407         if (tmp_hp == NULL)
1408                 goto fail;
1409
1410         memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1411
1412         hp_offset = 0; /* where we start the current page size entries */
1413
1414         huge_register_sigbus();
1415
1416         /* make a copy of socket_mem, needed for balanced allocation. */
1417         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1418                 memory[i] = internal_config.socket_mem[i];
1419
1420         /* map all hugepages and sort them */
1421         for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1422                 unsigned pages_old, pages_new;
1423                 struct hugepage_info *hpi;
1424
1425                 /*
1426                  * we don't yet mark hugepages as used at this stage, so
1427                  * we just map all hugepages available to the system
1428                  * all hugepages are still located on socket 0
1429                  */
1430                 hpi = &internal_config.hugepage_info[i];
1431
1432                 if (hpi->num_pages[0] == 0)
1433                         continue;
1434
1435                 /* map all hugepages available */
1436                 pages_old = hpi->num_pages[0];
1437                 pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, memory);
1438                 if (pages_new < pages_old) {
1439                         RTE_LOG(DEBUG, EAL,
1440                                 "%d not %d hugepages of size %u MB allocated\n",
1441                                 pages_new, pages_old,
1442                                 (unsigned)(hpi->hugepage_sz / 0x100000));
1443
1444                         int pages = pages_old - pages_new;
1445
1446                         nr_hugepages -= pages;
1447                         hpi->num_pages[0] = pages_new;
1448                         if (pages_new == 0)
1449                                 continue;
1450                 }
1451
1452                 if (phys_addrs_available &&
1453                                 rte_eal_iova_mode() != RTE_IOVA_VA) {
1454                         /* find physical addresses for each hugepage */
1455                         if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1456                                 RTE_LOG(DEBUG, EAL, "Failed to find phys addr "
1457                                         "for %u MB pages\n",
1458                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1459                                 goto fail;
1460                         }
1461                 } else {
1462                         /* set physical addresses for each hugepage */
1463                         if (set_physaddrs(&tmp_hp[hp_offset], hpi) < 0) {
1464                                 RTE_LOG(DEBUG, EAL, "Failed to set phys addr "
1465                                         "for %u MB pages\n",
1466                                         (unsigned int)(hpi->hugepage_sz / 0x100000));
1467                                 goto fail;
1468                         }
1469                 }
1470
1471                 if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1472                         RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1473                                         (unsigned)(hpi->hugepage_sz / 0x100000));
1474                         goto fail;
1475                 }
1476
1477                 qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1478                       sizeof(struct hugepage_file), cmp_physaddr);
1479
1480                 /* we have processed a num of hugepages of this size, so inc offset */
1481                 hp_offset += hpi->num_pages[0];
1482         }
1483
1484         huge_recover_sigbus();
1485
1486         if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1487                 internal_config.memory = eal_get_hugepage_mem_size();
1488
1489         nr_hugefiles = nr_hugepages;
1490
1491
1492         /* clean out the numbers of pages */
1493         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1494                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1495                         internal_config.hugepage_info[i].num_pages[j] = 0;
1496
1497         /* get hugepages for each socket */
1498         for (i = 0; i < nr_hugefiles; i++) {
1499                 int socket = tmp_hp[i].socket_id;
1500
1501                 /* find a hugepage info with right size and increment num_pages */
1502                 const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1503                                 (int)internal_config.num_hugepage_sizes);
1504                 for (j = 0; j < nb_hpsizes; j++) {
1505                         if (tmp_hp[i].size ==
1506                                         internal_config.hugepage_info[j].hugepage_sz) {
1507                                 internal_config.hugepage_info[j].num_pages[socket]++;
1508                         }
1509                 }
1510         }
1511
1512         /* make a copy of socket_mem, needed for number of pages calculation */
1513         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1514                 memory[i] = internal_config.socket_mem[i];
1515
1516         /* calculate final number of pages */
1517         nr_hugepages = calc_num_pages_per_socket(memory,
1518                         internal_config.hugepage_info, used_hp,
1519                         internal_config.num_hugepage_sizes);
1520
1521         /* error if not enough memory available */
1522         if (nr_hugepages < 0)
1523                 goto fail;
1524
1525         /* reporting in! */
1526         for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1527                 for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1528                         if (used_hp[i].num_pages[j] > 0) {
1529                                 RTE_LOG(DEBUG, EAL,
1530                                         "Requesting %u pages of size %uMB"
1531                                         " from socket %i\n",
1532                                         used_hp[i].num_pages[j],
1533                                         (unsigned)
1534                                         (used_hp[i].hugepage_sz / 0x100000),
1535                                         j);
1536                         }
1537                 }
1538         }
1539
1540         /* create shared memory */
1541         hugepage = create_shared_memory(eal_hugepage_data_path(),
1542                         nr_hugefiles * sizeof(struct hugepage_file));
1543
1544         if (hugepage == NULL) {
1545                 RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1546                 goto fail;
1547         }
1548         memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1549
1550         /*
1551          * unmap pages that we won't need (looks at used_hp).
1552          * also, sets final_va to NULL on pages that were unmapped.
1553          */
1554         if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1555                         internal_config.num_hugepage_sizes) < 0) {
1556                 RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1557                 goto fail;
1558         }
1559
1560         /*
1561          * copy stuff from malloc'd hugepage* to the actual shared memory.
1562          * this procedure only copies those hugepages that have orig_va
1563          * not NULL. has overflow protection.
1564          */
1565         if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1566                         tmp_hp, nr_hugefiles) < 0) {
1567                 RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1568                 goto fail;
1569         }
1570
1571 #ifndef RTE_ARCH_64
1572         /* for legacy 32-bit mode, we did not preallocate VA space, so do it */
1573         if (internal_config.legacy_mem &&
1574                         prealloc_segments(hugepage, nr_hugefiles)) {
1575                 RTE_LOG(ERR, EAL, "Could not preallocate VA space for hugepages\n");
1576                 goto fail;
1577         }
1578 #endif
1579
1580         /* remap all pages we do need into memseg list VA space, so that those
1581          * pages become first-class citizens in DPDK memory subsystem
1582          */
1583         if (remap_needed_hugepages(hugepage, nr_hugefiles)) {
1584                 RTE_LOG(ERR, EAL, "Couldn't remap hugepage files into memseg lists\n");
1585                 goto fail;
1586         }
1587
1588         /* free the hugepage backing files */
1589         if (internal_config.hugepage_unlink &&
1590                 unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1591                 RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1592                 goto fail;
1593         }
1594
1595         /* free the temporary hugepage table */
1596         free(tmp_hp);
1597         tmp_hp = NULL;
1598
1599         munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1600
1601         /* we're not going to allocate more pages, so release VA space for
1602          * unused memseg lists
1603          */
1604         for (i = 0; i < RTE_MAX_MEMSEG_LISTS; i++) {
1605                 struct rte_memseg_list *msl = &mcfg->memsegs[i];
1606                 size_t mem_sz;
1607
1608                 /* skip inactive lists */
1609                 if (msl->base_va == NULL)
1610                         continue;
1611                 /* skip lists where there is at least one page allocated */
1612                 if (msl->memseg_arr.count > 0)
1613                         continue;
1614                 /* this is an unused list, deallocate it */
1615                 mem_sz = (size_t)msl->page_sz * msl->memseg_arr.len;
1616                 munmap(msl->base_va, mem_sz);
1617                 msl->base_va = NULL;
1618
1619                 /* destroy backing fbarray */
1620                 rte_fbarray_destroy(&msl->memseg_arr);
1621         }
1622
1623         return 0;
1624
1625 fail:
1626         huge_recover_sigbus();
1627         free(tmp_hp);
1628         if (hugepage != NULL)
1629                 munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1630
1631         return -1;
1632 }
1633
1634 static int __rte_unused
1635 hugepage_count_walk(const struct rte_memseg_list *msl, void *arg)
1636 {
1637         struct hugepage_info *hpi = arg;
1638
1639         if (msl->page_sz != hpi->hugepage_sz)
1640                 return 0;
1641
1642         hpi->num_pages[msl->socket_id] += msl->memseg_arr.len;
1643         return 0;
1644 }
1645
1646 static int
1647 limits_callback(int socket_id, size_t cur_limit, size_t new_len)
1648 {
1649         RTE_SET_USED(socket_id);
1650         RTE_SET_USED(cur_limit);
1651         RTE_SET_USED(new_len);
1652         return -1;
1653 }
1654
1655 static int
1656 eal_hugepage_init(void)
1657 {
1658         struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1659         uint64_t memory[RTE_MAX_NUMA_NODES];
1660         int hp_sz_idx, socket_id;
1661
1662         test_phys_addrs_available();
1663
1664         memset(used_hp, 0, sizeof(used_hp));
1665
1666         for (hp_sz_idx = 0;
1667                         hp_sz_idx < (int) internal_config.num_hugepage_sizes;
1668                         hp_sz_idx++) {
1669 #ifndef RTE_ARCH_64
1670                 struct hugepage_info dummy;
1671                 unsigned int i;
1672 #endif
1673                 /* also initialize used_hp hugepage sizes in used_hp */
1674                 struct hugepage_info *hpi;
1675                 hpi = &internal_config.hugepage_info[hp_sz_idx];
1676                 used_hp[hp_sz_idx].hugepage_sz = hpi->hugepage_sz;
1677
1678 #ifndef RTE_ARCH_64
1679                 /* for 32-bit, limit number of pages on socket to whatever we've
1680                  * preallocated, as we cannot allocate more.
1681                  */
1682                 memset(&dummy, 0, sizeof(dummy));
1683                 dummy.hugepage_sz = hpi->hugepage_sz;
1684                 if (rte_memseg_list_walk(hugepage_count_walk, &dummy) < 0)
1685                         return -1;
1686
1687                 for (i = 0; i < RTE_DIM(dummy.num_pages); i++) {
1688                         hpi->num_pages[i] = RTE_MIN(hpi->num_pages[i],
1689                                         dummy.num_pages[i]);
1690                 }
1691 #endif
1692         }
1693
1694         /* make a copy of socket_mem, needed for balanced allocation. */
1695         for (hp_sz_idx = 0; hp_sz_idx < RTE_MAX_NUMA_NODES; hp_sz_idx++)
1696                 memory[hp_sz_idx] = internal_config.socket_mem[hp_sz_idx];
1697
1698         /* calculate final number of pages */
1699         if (calc_num_pages_per_socket(memory,
1700                         internal_config.hugepage_info, used_hp,
1701                         internal_config.num_hugepage_sizes) < 0)
1702                 return -1;
1703
1704         for (hp_sz_idx = 0;
1705                         hp_sz_idx < (int)internal_config.num_hugepage_sizes;
1706                         hp_sz_idx++) {
1707                 for (socket_id = 0; socket_id < RTE_MAX_NUMA_NODES;
1708                                 socket_id++) {
1709                         struct rte_memseg **pages;
1710                         struct hugepage_info *hpi = &used_hp[hp_sz_idx];
1711                         unsigned int num_pages = hpi->num_pages[socket_id];
1712                         int num_pages_alloc, i;
1713
1714                         if (num_pages == 0)
1715                                 continue;
1716
1717                         pages = malloc(sizeof(*pages) * num_pages);
1718
1719                         RTE_LOG(DEBUG, EAL, "Allocating %u pages of size %" PRIu64 "M on socket %i\n",
1720                                 num_pages, hpi->hugepage_sz >> 20, socket_id);
1721
1722                         num_pages_alloc = eal_memalloc_alloc_seg_bulk(pages,
1723                                         num_pages, hpi->hugepage_sz,
1724                                         socket_id, true);
1725                         if (num_pages_alloc < 0) {
1726                                 free(pages);
1727                                 return -1;
1728                         }
1729
1730                         /* mark preallocated pages as unfreeable */
1731                         for (i = 0; i < num_pages_alloc; i++) {
1732                                 struct rte_memseg *ms = pages[i];
1733                                 ms->flags |= RTE_MEMSEG_FLAG_DO_NOT_FREE;
1734                         }
1735                         free(pages);
1736                 }
1737         }
1738         /* if socket limits were specified, set them */
1739         if (internal_config.force_socket_limits) {
1740                 unsigned int i;
1741                 for (i = 0; i < RTE_MAX_NUMA_NODES; i++) {
1742                         uint64_t limit = internal_config.socket_limit[i];
1743                         if (limit == 0)
1744                                 continue;
1745                         if (rte_mem_alloc_validator_register("socket-limit",
1746                                         limits_callback, i, limit))
1747                                 RTE_LOG(ERR, EAL, "Failed to register socket limits validator callback\n");
1748                 }
1749         }
1750         return 0;
1751 }
1752
1753 /*
1754  * uses fstat to report the size of a file on disk
1755  */
1756 static off_t
1757 getFileSize(int fd)
1758 {
1759         struct stat st;
1760         if (fstat(fd, &st) < 0)
1761                 return 0;
1762         return st.st_size;
1763 }
1764
1765 /*
1766  * This creates the memory mappings in the secondary process to match that of
1767  * the server process. It goes through each memory segment in the DPDK runtime
1768  * configuration and finds the hugepages which form that segment, mapping them
1769  * in order to form a contiguous block in the virtual memory space
1770  */
1771 static int
1772 eal_legacy_hugepage_attach(void)
1773 {
1774         struct hugepage_file *hp = NULL;
1775         unsigned int num_hp = 0;
1776         unsigned int i = 0;
1777         unsigned int cur_seg;
1778         off_t size = 0;
1779         int fd, fd_hugepage = -1;
1780
1781         if (aslr_enabled() > 0) {
1782                 RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1783                                 "(ASLR) is enabled in the kernel.\n");
1784                 RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1785                                 "into secondary processes\n");
1786         }
1787
1788         test_phys_addrs_available();
1789
1790         fd_hugepage = open(eal_hugepage_data_path(), O_RDONLY);
1791         if (fd_hugepage < 0) {
1792                 RTE_LOG(ERR, EAL, "Could not open %s\n",
1793                                 eal_hugepage_data_path());
1794                 goto error;
1795         }
1796
1797         size = getFileSize(fd_hugepage);
1798         hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1799         if (hp == MAP_FAILED) {
1800                 RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1801                                 eal_hugepage_data_path());
1802                 goto error;
1803         }
1804
1805         num_hp = size / sizeof(struct hugepage_file);
1806         RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1807
1808         /* map all segments into memory to make sure we get the addrs. the
1809          * segments themselves are already in memseg list (which is shared and
1810          * has its VA space already preallocated), so we just need to map
1811          * everything into correct addresses.
1812          */
1813         for (i = 0; i < num_hp; i++) {
1814                 struct hugepage_file *hf = &hp[i];
1815                 size_t map_sz = hf->size;
1816                 void *map_addr = hf->final_va;
1817
1818                 /* if size is zero, no more pages left */
1819                 if (map_sz == 0)
1820                         break;
1821
1822                 fd = open(hf->filepath, O_RDWR);
1823                 if (fd < 0) {
1824                         RTE_LOG(ERR, EAL, "Could not open %s: %s\n",
1825                                 hf->filepath, strerror(errno));
1826                         goto error;
1827                 }
1828
1829                 map_addr = mmap(map_addr, map_sz, PROT_READ | PROT_WRITE,
1830                                 MAP_SHARED | MAP_FIXED, fd, 0);
1831                 if (map_addr == MAP_FAILED) {
1832                         RTE_LOG(ERR, EAL, "Could not map %s: %s\n",
1833                                 hf->filepath, strerror(errno));
1834                         close(fd);
1835                         goto error;
1836                 }
1837
1838                 /* set shared lock on the file. */
1839                 if (flock(fd, LOCK_SH) < 0) {
1840                         RTE_LOG(DEBUG, EAL, "%s(): Locking file failed: %s\n",
1841                                 __func__, strerror(errno));
1842                         close(fd);
1843                         goto error;
1844                 }
1845
1846                 close(fd);
1847         }
1848         /* unmap the hugepage config file, since we are done using it */
1849         munmap(hp, size);
1850         close(fd_hugepage);
1851         return 0;
1852
1853 error:
1854         /* map all segments into memory to make sure we get the addrs */
1855         cur_seg = 0;
1856         for (cur_seg = 0; cur_seg < i; cur_seg++) {
1857                 struct hugepage_file *hf = &hp[i];
1858                 size_t map_sz = hf->size;
1859                 void *map_addr = hf->final_va;
1860
1861                 munmap(map_addr, map_sz);
1862         }
1863         if (hp != NULL && hp != MAP_FAILED)
1864                 munmap(hp, size);
1865         if (fd_hugepage >= 0)
1866                 close(fd_hugepage);
1867         return -1;
1868 }
1869
1870 static int
1871 eal_hugepage_attach(void)
1872 {
1873         if (eal_memalloc_sync_with_primary()) {
1874                 RTE_LOG(ERR, EAL, "Could not map memory from primary process\n");
1875                 if (aslr_enabled() > 0)
1876                         RTE_LOG(ERR, EAL, "It is recommended to disable ASLR in the kernel and retry running both primary and secondary processes\n");
1877                 return -1;
1878         }
1879         return 0;
1880 }
1881
1882 int
1883 rte_eal_hugepage_init(void)
1884 {
1885         return internal_config.legacy_mem ?
1886                         eal_legacy_hugepage_init() :
1887                         eal_hugepage_init();
1888 }
1889
1890 int
1891 rte_eal_hugepage_attach(void)
1892 {
1893         return internal_config.legacy_mem ?
1894                         eal_legacy_hugepage_attach() :
1895                         eal_hugepage_attach();
1896 }
1897
1898 int
1899 rte_eal_using_phys_addrs(void)
1900 {
1901         return phys_addrs_available;
1902 }
1903
1904 static int __rte_unused
1905 memseg_primary_init_32(void)
1906 {
1907         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1908         int active_sockets, hpi_idx, msl_idx = 0;
1909         unsigned int socket_id, i;
1910         struct rte_memseg_list *msl;
1911         uint64_t extra_mem_per_socket, total_extra_mem, total_requested_mem;
1912         uint64_t max_mem;
1913
1914         /* no-huge does not need this at all */
1915         if (internal_config.no_hugetlbfs)
1916                 return 0;
1917
1918         /* this is a giant hack, but desperate times call for desperate
1919          * measures. in legacy 32-bit mode, we cannot preallocate VA space,
1920          * because having upwards of 2 gigabytes of VA space already mapped will
1921          * interfere with our ability to map and sort hugepages.
1922          *
1923          * therefore, in legacy 32-bit mode, we will be initializing memseg
1924          * lists much later - in eal_memory.c, right after we unmap all the
1925          * unneeded pages. this will not affect secondary processes, as those
1926          * should be able to mmap the space without (too many) problems.
1927          */
1928         if (internal_config.legacy_mem)
1929                 return 0;
1930
1931         /* 32-bit mode is a very special case. we cannot know in advance where
1932          * the user will want to allocate their memory, so we have to do some
1933          * heuristics.
1934          */
1935         active_sockets = 0;
1936         total_requested_mem = 0;
1937         if (internal_config.force_sockets)
1938                 for (i = 0; i < rte_socket_count(); i++) {
1939                         uint64_t mem;
1940
1941                         socket_id = rte_socket_id_by_idx(i);
1942                         mem = internal_config.socket_mem[socket_id];
1943
1944                         if (mem == 0)
1945                                 continue;
1946
1947                         active_sockets++;
1948                         total_requested_mem += mem;
1949                 }
1950         else
1951                 total_requested_mem = internal_config.memory;
1952
1953         max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
1954         if (total_requested_mem > max_mem) {
1955                 RTE_LOG(ERR, EAL, "Invalid parameters: 32-bit process can at most use %uM of memory\n",
1956                                 (unsigned int)(max_mem >> 20));
1957                 return -1;
1958         }
1959         total_extra_mem = max_mem - total_requested_mem;
1960         extra_mem_per_socket = active_sockets == 0 ? total_extra_mem :
1961                         total_extra_mem / active_sockets;
1962
1963         /* the allocation logic is a little bit convoluted, but here's how it
1964          * works, in a nutshell:
1965          *  - if user hasn't specified on which sockets to allocate memory via
1966          *    --socket-mem, we allocate all of our memory on master core socket.
1967          *  - if user has specified sockets to allocate memory on, there may be
1968          *    some "unused" memory left (e.g. if user has specified --socket-mem
1969          *    such that not all memory adds up to 2 gigabytes), so add it to all
1970          *    sockets that are in use equally.
1971          *
1972          * page sizes are sorted by size in descending order, so we can safely
1973          * assume that we dispense with bigger page sizes first.
1974          */
1975
1976         /* create memseg lists */
1977         for (i = 0; i < rte_socket_count(); i++) {
1978                 int hp_sizes = (int) internal_config.num_hugepage_sizes;
1979                 uint64_t max_socket_mem, cur_socket_mem;
1980                 unsigned int master_lcore_socket;
1981                 struct rte_config *cfg = rte_eal_get_configuration();
1982                 bool skip;
1983
1984                 socket_id = rte_socket_id_by_idx(i);
1985
1986 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
1987                 if (socket_id > 0)
1988                         break;
1989 #endif
1990
1991                 /* if we didn't specifically request memory on this socket */
1992                 skip = active_sockets != 0 &&
1993                                 internal_config.socket_mem[socket_id] == 0;
1994                 /* ...or if we didn't specifically request memory on *any*
1995                  * socket, and this is not master lcore
1996                  */
1997                 master_lcore_socket = rte_lcore_to_socket_id(cfg->master_lcore);
1998                 skip |= active_sockets == 0 && socket_id != master_lcore_socket;
1999
2000                 if (skip) {
2001                         RTE_LOG(DEBUG, EAL, "Will not preallocate memory on socket %u\n",
2002                                         socket_id);
2003                         continue;
2004                 }
2005
2006                 /* max amount of memory on this socket */
2007                 max_socket_mem = (active_sockets != 0 ?
2008                                         internal_config.socket_mem[socket_id] :
2009                                         internal_config.memory) +
2010                                         extra_mem_per_socket;
2011                 cur_socket_mem = 0;
2012
2013                 for (hpi_idx = 0; hpi_idx < hp_sizes; hpi_idx++) {
2014                         uint64_t max_pagesz_mem, cur_pagesz_mem = 0;
2015                         uint64_t hugepage_sz;
2016                         struct hugepage_info *hpi;
2017                         int type_msl_idx, max_segs, total_segs = 0;
2018
2019                         hpi = &internal_config.hugepage_info[hpi_idx];
2020                         hugepage_sz = hpi->hugepage_sz;
2021
2022                         /* check if pages are actually available */
2023                         if (hpi->num_pages[socket_id] == 0)
2024                                 continue;
2025
2026                         max_segs = RTE_MAX_MEMSEG_PER_TYPE;
2027                         max_pagesz_mem = max_socket_mem - cur_socket_mem;
2028
2029                         /* make it multiple of page size */
2030                         max_pagesz_mem = RTE_ALIGN_FLOOR(max_pagesz_mem,
2031                                         hugepage_sz);
2032
2033                         RTE_LOG(DEBUG, EAL, "Attempting to preallocate "
2034                                         "%" PRIu64 "M on socket %i\n",
2035                                         max_pagesz_mem >> 20, socket_id);
2036
2037                         type_msl_idx = 0;
2038                         while (cur_pagesz_mem < max_pagesz_mem &&
2039                                         total_segs < max_segs) {
2040                                 uint64_t cur_mem;
2041                                 unsigned int n_segs;
2042
2043                                 if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
2044                                         RTE_LOG(ERR, EAL,
2045                                                 "No more space in memseg lists, please increase %s\n",
2046                                                 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2047                                         return -1;
2048                                 }
2049
2050                                 msl = &mcfg->memsegs[msl_idx];
2051
2052                                 cur_mem = get_mem_amount(hugepage_sz,
2053                                                 max_pagesz_mem);
2054                                 n_segs = cur_mem / hugepage_sz;
2055
2056                                 if (alloc_memseg_list(msl, hugepage_sz, n_segs,
2057                                                 socket_id, type_msl_idx)) {
2058                                         /* failing to allocate a memseg list is
2059                                          * a serious error.
2060                                          */
2061                                         RTE_LOG(ERR, EAL, "Cannot allocate memseg list\n");
2062                                         return -1;
2063                                 }
2064
2065                                 if (alloc_va_space(msl)) {
2066                                         /* if we couldn't allocate VA space, we
2067                                          * can try with smaller page sizes.
2068                                          */
2069                                         RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list, retrying with different page size\n");
2070                                         /* deallocate memseg list */
2071                                         if (free_memseg_list(msl))
2072                                                 return -1;
2073                                         break;
2074                                 }
2075
2076                                 total_segs += msl->memseg_arr.len;
2077                                 cur_pagesz_mem = total_segs * hugepage_sz;
2078                                 type_msl_idx++;
2079                                 msl_idx++;
2080                         }
2081                         cur_socket_mem += cur_pagesz_mem;
2082                 }
2083                 if (cur_socket_mem == 0) {
2084                         RTE_LOG(ERR, EAL, "Cannot allocate VA space on socket %u\n",
2085                                 socket_id);
2086                         return -1;
2087                 }
2088         }
2089
2090         return 0;
2091 }
2092
2093 static int __rte_unused
2094 memseg_primary_init(void)
2095 {
2096         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
2097         int i, socket_id, hpi_idx, msl_idx = 0;
2098         struct rte_memseg_list *msl;
2099         uint64_t max_mem, total_mem;
2100
2101         /* no-huge does not need this at all */
2102         if (internal_config.no_hugetlbfs)
2103                 return 0;
2104
2105         max_mem = (uint64_t)RTE_MAX_MEM_MB << 20;
2106         total_mem = 0;
2107
2108         /* create memseg lists */
2109         for (hpi_idx = 0; hpi_idx < (int) internal_config.num_hugepage_sizes;
2110                         hpi_idx++) {
2111                 struct hugepage_info *hpi;
2112                 uint64_t hugepage_sz;
2113
2114                 hpi = &internal_config.hugepage_info[hpi_idx];
2115                 hugepage_sz = hpi->hugepage_sz;
2116
2117                 for (i = 0; i < (int) rte_socket_count(); i++) {
2118                         uint64_t max_type_mem, total_type_mem = 0;
2119                         int type_msl_idx, max_segs, total_segs = 0;
2120
2121                         socket_id = rte_socket_id_by_idx(i);
2122
2123 #ifndef RTE_EAL_NUMA_AWARE_HUGEPAGES
2124                         if (socket_id > 0)
2125                                 break;
2126 #endif
2127
2128                         if (total_mem >= max_mem)
2129                                 break;
2130
2131                         max_type_mem = RTE_MIN(max_mem - total_mem,
2132                                 (uint64_t)RTE_MAX_MEM_MB_PER_TYPE << 20);
2133                         max_segs = RTE_MAX_MEMSEG_PER_TYPE;
2134
2135                         type_msl_idx = 0;
2136                         while (total_type_mem < max_type_mem &&
2137                                         total_segs < max_segs) {
2138                                 uint64_t cur_max_mem, cur_mem;
2139                                 unsigned int n_segs;
2140
2141                                 if (msl_idx >= RTE_MAX_MEMSEG_LISTS) {
2142                                         RTE_LOG(ERR, EAL,
2143                                                 "No more space in memseg lists, please increase %s\n",
2144                                                 RTE_STR(CONFIG_RTE_MAX_MEMSEG_LISTS));
2145                                         return -1;
2146                                 }
2147
2148                                 msl = &mcfg->memsegs[msl_idx++];
2149
2150                                 cur_max_mem = max_type_mem - total_type_mem;
2151
2152                                 cur_mem = get_mem_amount(hugepage_sz,
2153                                                 cur_max_mem);
2154                                 n_segs = cur_mem / hugepage_sz;
2155
2156                                 if (alloc_memseg_list(msl, hugepage_sz, n_segs,
2157                                                 socket_id, type_msl_idx))
2158                                         return -1;
2159
2160                                 total_segs += msl->memseg_arr.len;
2161                                 total_type_mem = total_segs * hugepage_sz;
2162                                 type_msl_idx++;
2163
2164                                 if (alloc_va_space(msl)) {
2165                                         RTE_LOG(ERR, EAL, "Cannot allocate VA space for memseg list\n");
2166                                         return -1;
2167                                 }
2168                         }
2169                         total_mem += total_type_mem;
2170                 }
2171         }
2172         return 0;
2173 }
2174
2175 static int
2176 memseg_secondary_init(void)
2177 {
2178         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
2179         int msl_idx = 0;
2180         struct rte_memseg_list *msl;
2181
2182         for (msl_idx = 0; msl_idx < RTE_MAX_MEMSEG_LISTS; msl_idx++) {
2183
2184                 msl = &mcfg->memsegs[msl_idx];
2185
2186                 /* skip empty memseg lists */
2187                 if (msl->memseg_arr.len == 0)
2188                         continue;
2189
2190                 if (rte_fbarray_attach(&msl->memseg_arr)) {
2191                         RTE_LOG(ERR, EAL, "Cannot attach to primary process memseg lists\n");
2192                         return -1;
2193                 }
2194
2195                 /* preallocate VA space */
2196                 if (alloc_va_space(msl)) {
2197                         RTE_LOG(ERR, EAL, "Cannot preallocate VA space for hugepage memory\n");
2198                         return -1;
2199                 }
2200         }
2201
2202         return 0;
2203 }
2204
2205 int
2206 rte_eal_memseg_init(void)
2207 {
2208         /* increase rlimit to maximum */
2209         struct rlimit lim;
2210
2211         if (getrlimit(RLIMIT_NOFILE, &lim) == 0) {
2212                 /* set limit to maximum */
2213                 lim.rlim_cur = lim.rlim_max;
2214
2215                 if (setrlimit(RLIMIT_NOFILE, &lim) < 0) {
2216                         RTE_LOG(DEBUG, EAL, "Setting maximum number of open files failed: %s\n",
2217                                         strerror(errno));
2218                 } else {
2219                         RTE_LOG(DEBUG, EAL, "Setting maximum number of open files to %"
2220                                         PRIu64 "\n",
2221                                         (uint64_t)lim.rlim_cur);
2222                 }
2223         } else {
2224                 RTE_LOG(ERR, EAL, "Cannot get current resource limits\n");
2225         }
2226
2227         return rte_eal_process_type() == RTE_PROC_PRIMARY ?
2228 #ifndef RTE_ARCH_64
2229                         memseg_primary_init_32() :
2230 #else
2231                         memseg_primary_init() :
2232 #endif
2233                         memseg_secondary_init();
2234 }