lib/librte_eal/linuxapp/eal/eal_hugepage_info.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2013 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <string.h>
  35 #include <sys/types.h>
  36 #include <sys/file.h>
  37 #include <dirent.h>
  38 #include <stdint.h>
  39 #include <stdlib.h>
  40 #include <stdio.h>
  41 #include <fnmatch.h>
  42 #include <inttypes.h>
  43 #include <stdarg.h>
  44 #include <unistd.h>
  45 #include <errno.h>
  46 #include <sys/queue.h>
  47
  48 #include <rte_memory.h>
  49 #include <rte_memzone.h>
  50 #include <rte_tailq.h>
  51 #include <rte_eal.h>
  52 #include <rte_launch.h>
  53 #include <rte_per_lcore.h>
  54 #include <rte_lcore.h>
  55 #include <rte_debug.h>
  56 #include <rte_log.h>
  57 #include <rte_common.h>
  58 #include "rte_string_fns.h"
  59 #include "eal_internal_cfg.h"
  60 #include "eal_hugepages.h"
  61 #include "eal_filesystem.h"
  62
  63 static const char sys_dir_path[] = "/sys/kernel/mm/hugepages";
  64
  65 static int32_t
  66 get_num_hugepages(const char *subdir)
  67 {
  68         char path[PATH_MAX];
  69         long unsigned num_pages = 0;
  70         const char *nr_hp_file;
  71
  72         /* if secondary process, just look at the number of hugepages,
  73          * otherwise look at number of free hugepages */
  74         if (internal_config.process_type == RTE_PROC_SECONDARY)
  75                 nr_hp_file = "nr_hugepages";
  76         else
  77                 nr_hp_file = "free_hugepages";
  78
  79         rte_snprintf(path, sizeof(path), "%s/%s/%s",
  80                         sys_dir_path, subdir, nr_hp_file);
  81
  82         if (eal_parse_sysfs_value(path, &num_pages) < 0)
  83                 return 0;
  84
  85         if (num_pages == 0)
  86                 RTE_LOG(WARNING, EAL, "No free hugepages reported in %s\n",
  87                                 subdir);
  88
  89         return (int32_t)num_pages;
  90 }
  91
  92 static uint64_t
  93 get_default_hp_size(void)
  94 {
  95         const char proc_meminfo[] = "/proc/meminfo";
  96         const char str_hugepagesz[] = "Hugepagesize:";
  97         unsigned hugepagesz_len = sizeof(str_hugepagesz) - 1;
  98         char buffer[256];
  99         unsigned long long size = 0;
 100
 101         FILE *fd = fopen(proc_meminfo, "r");
 102         if (fd == NULL)
 103                 rte_panic("Cannot open %s\n", proc_meminfo);
 104         while(fgets(buffer, sizeof(buffer), fd)){
 105                 if (strncmp(buffer, str_hugepagesz, hugepagesz_len) == 0){
 106                         size = rte_str_to_size(&buffer[hugepagesz_len]);
 107                         break;
 108                 }
 109         }
 110         fclose(fd);
 111         if (size == 0)
 112                 rte_panic("Cannot get default hugepage size from %s\n", proc_meminfo);
 113         return size;
 114 }
 115
 116 static const char *
 117 get_hugepage_dir(uint64_t hugepage_sz)
 118 {
 119         enum proc_mount_fieldnames {
 120                 DEVICE = 0,
 121                 MOUNTPT,
 122                 FSTYPE,
 123                 OPTIONS,
 124                 _FIELDNAME_MAX
 125         };
 126         static uint64_t default_size = 0;
 127         const char proc_mounts[] = "/proc/mounts";
 128         const char hugetlbfs_str[] = "hugetlbfs";
 129         const size_t htlbfs_str_len = sizeof(hugetlbfs_str) - 1;
 130         const char pagesize_opt[] = "pagesize=";
 131         const size_t pagesize_opt_len = sizeof(pagesize_opt) - 1;
 132         const char split_tok = ' ';
 133         char *splitstr[_FIELDNAME_MAX];
 134         char buf[BUFSIZ];
 135         char *retval = NULL;
 136
 137         FILE *fd = fopen(proc_mounts, "r");
 138         if (fd == NULL)
 139                 rte_panic("Cannot open %s\n", proc_mounts);
 140
 141         if (default_size == 0)
 142                 default_size = get_default_hp_size();
 143
 144         while (fgets(buf, sizeof(buf), fd)){
 145                 if (rte_strsplit(buf, sizeof(buf), splitstr, _FIELDNAME_MAX,
 146                                 split_tok) != _FIELDNAME_MAX) {
 147                         RTE_LOG(ERR, EAL, "Error parsing %s\n", proc_mounts);
 148                         break; /* return NULL */
 149                 }
 150
 151                 /* we have a specified --huge-dir option, only examine that dir */
 152                 if (internal_config.hugepage_dir != NULL &&
 153                                 strcmp(splitstr[MOUNTPT], internal_config.hugepage_dir) != 0)
 154                         continue;
 155
 156                 if (strncmp(splitstr[FSTYPE], hugetlbfs_str, htlbfs_str_len) == 0){
 157                         const char *pagesz_str = strstr(splitstr[OPTIONS], pagesize_opt);
 158
 159                         /* if no explicit page size, the default page size is compared */
 160                         if (pagesz_str == NULL){
 161                                 if (hugepage_sz == default_size){
 162                                         retval = strdup(splitstr[MOUNTPT]);
 163                                         break;
 164                                 }
 165                         }
 166                         /* there is an explicit page size, so check it */
 167                         else {
 168                                 uint64_t pagesz = rte_str_to_size(&pagesz_str[pagesize_opt_len]);
 169                                 if (pagesz == hugepage_sz) {
 170                                         retval = strdup(splitstr[MOUNTPT]);
 171                                         break;
 172                                 }
 173                         }
 174                 } /* end if strncmp hugetlbfs */
 175         } /* end while fgets */
 176
 177         fclose(fd);
 178         return retval;
 179 }
 180
 181 static inline void
 182 swap_hpi(struct hugepage_info *a, struct hugepage_info *b)
 183 {
 184         char buf[sizeof(*a)];
 185         memcpy(buf, a, sizeof(buf));
 186         memcpy(a, b, sizeof(buf));
 187         memcpy(b, buf, sizeof(buf));
 188 }
 189
 190 /*
 191  * Clear the hugepage directory of whatever hugepage files
 192  * there are. Checks if the file is locked (i.e.
 193  * if it's in use by another DPDK process).
 194  */
 195 static int
 196 clear_hugedir(const char * hugedir)
 197 {
 198         DIR *dir;
 199         struct dirent *dirent;
 200         int dir_fd, fd, lck_result;
 201         const char filter[] = "*map_*"; /* matches hugepage files */
 202
 203         /* open directory */
 204         dir = opendir(hugedir);
 205         if (!dir) {
 206                 RTE_LOG(INFO, EAL, "Unable to open hugepage directory %s\n",
 207                                 hugedir);
 208                 goto error;
 209         }
 210         dir_fd = dirfd(dir);
 211
 212         dirent = readdir(dir);
 213         if (!dirent) {
 214                 RTE_LOG(INFO, EAL, "Unable to read hugepage directory %s\n",
 215                                 hugedir);
 216                 goto error;
 217         }
 218
 219         while(dirent != NULL){
 220                 /* skip files that don't match the hugepage pattern */
 221                 if (fnmatch(filter, dirent->d_name, 0) > 0) {
 222                         dirent = readdir(dir);
 223                         continue;
 224                 }
 225
 226                 /* try and lock the file */
 227                 fd = openat(dir_fd, dirent->d_name, O_RDONLY);
 228
 229                 /* skip to next file */
 230                 if (fd == -1) {
 231                         dirent = readdir(dir);
 232                         continue;
 233                 }
 234
 235                 /* non-blocking lock */
 236                 lck_result = flock(fd, LOCK_EX | LOCK_NB);
 237
 238                 /* if lock succeeds, unlock and remove the file */
 239                 if (lck_result != -1) {
 240                         flock(fd, LOCK_UN);
 241                         unlinkat(dir_fd, dirent->d_name, 0);
 242                 }
 243                 close (fd);
 244                 dirent = readdir(dir);
 245         }
 246
 247         closedir(dir);
 248         return 0;
 249
 250 error:
 251         if (dir)
 252                 closedir(dir);
 253
 254         RTE_LOG(INFO, EAL, "Error while clearing hugepage dir: %s\n",
 255                 strerror(errno));
 256
 257         return -1;
 258 }
 259
 260 /*
 261  * when we initialize the hugepage info, everything goes
 262  * to socket 0 by default. it will later get sorted by memory
 263  * initialization procedure.
 264  */
 265 int
 266 eal_hugepage_info_init(void)
 267 {
 268         const char dirent_start_text[] = "hugepages-";
 269         const size_t dirent_start_len = sizeof(dirent_start_text) - 1;
 270         unsigned i, num_sizes = 0;
 271
 272         DIR *dir = opendir(sys_dir_path);
 273         if (dir == NULL)
 274                 rte_panic("Cannot open directory %s to read system hugepage info\n",
 275                                 sys_dir_path);
 276
 277         struct dirent *dirent = readdir(dir);
 278         while(dirent != NULL){
 279                 if (strncmp(dirent->d_name, dirent_start_text, dirent_start_len) == 0){
 280                         struct hugepage_info *hpi = \
 281                                         &internal_config.hugepage_info[num_sizes];
 282                         hpi->hugepage_sz = rte_str_to_size(&dirent->d_name[dirent_start_len]);
 283                         hpi->hugedir = get_hugepage_dir(hpi->hugepage_sz);
 284
 285                         /* first, check if we have a mountpoint */
 286                         if (hpi->hugedir == NULL){
 287                                 int32_t num_pages;
 288                                 if ((num_pages = get_num_hugepages(dirent->d_name)) > 0)
 289                                         RTE_LOG(INFO, EAL, "%u hugepages of size %llu reserved, "\
 290                                                         "but no mounted hugetlbfs found for that size\n",
 291                                                         (unsigned)num_pages,
 292                                                         (unsigned long long)hpi->hugepage_sz);
 293                         } else {
 294                                 /* try to obtain a writelock */
 295                                 hpi->lock_descriptor = open(hpi->hugedir, O_RDONLY);
 296
 297                                 /* if blocking lock failed */
 298                                 if (flock(hpi->lock_descriptor, LOCK_EX) == -1) {
 299                                         RTE_LOG(CRIT, EAL, "Failed to lock hugepage directory!\n");
 300                                         return -1;
 301                                 }
 302                                 /* clear out the hugepages dir from unused pages */
 303                                 if (clear_hugedir(hpi->hugedir) == -1)
 304                                         return -1;
 305
 306                                 /* for now, put all pages into socket 0,
 307                                  * later they will be sorted */
 308                                 hpi->num_pages[0] = get_num_hugepages(dirent->d_name);
 309
 310 #ifndef RTE_ARCH_X86_64
 311                                 /* for 32-bit systems, limit number of hugepages to 1GB per page size */
 312                                 hpi->num_pages[0] = RTE_MIN(hpi->num_pages[0],
 313                                                 RTE_PGSIZE_1G / hpi->hugepage_sz);
 314 #endif
 315
 316                                 num_sizes++;
 317                         }
 318                 }
 319                 dirent = readdir(dir);
 320         }
 321         closedir(dir);
 322         internal_config.num_hugepage_sizes = num_sizes;
 323
 324         /* sort the page directory entries by size, largest to smallest */
 325         for (i = 0; i < num_sizes; i++){
 326                 unsigned j;
 327                 for (j = i+1; j < num_sizes; j++)
 328                         if (internal_config.hugepage_info[j-1].hugepage_sz < \
 329                                         internal_config.hugepage_info[j].hugepage_sz)
 330                                 swap_hpi(&internal_config.hugepage_info[j-1],
 331                                                 &internal_config.hugepage_info[j]);
 332         }
 333
 334         /* now we have all info, check we have at least one valid size */
 335         for (i = 0; i < num_sizes; i++)
 336                 if (internal_config.hugepage_info[i].hugedir != NULL &&
 337                                 internal_config.hugepage_info[i].num_pages[0] > 0)
 338                         return 0;
 339
 340         /* no valid hugepage mounts available, return error */
 341         return -1;
 342 }