lib/librte_eal/linuxapp/eal/eal_hugepage_info.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  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_eal.h>
  51 #include <rte_launch.h>
  52 #include <rte_per_lcore.h>
  53 #include <rte_lcore.h>
  54 #include <rte_debug.h>
  55 #include <rte_log.h>
  56 #include <rte_common.h>
  57 #include "rte_string_fns.h"
  58 #include "eal_internal_cfg.h"
  59 #include "eal_hugepages.h"
  60 #include "eal_filesystem.h"
  61
  62 static const char sys_dir_path[] = "/sys/kernel/mm/hugepages";
  63
  64 /* this function is only called from eal_hugepage_info_init which itself
  65  * is only called from a primary process */
  66 static int32_t
  67 get_num_hugepages(const char *subdir)
  68 {
  69         char path[PATH_MAX];
  70         long unsigned resv_pages, num_pages = 0;
  71         const char *nr_hp_file = "free_hugepages";
  72         const char *nr_rsvd_file = "resv_hugepages";
  73
  74         /* first, check how many reserved pages kernel reports */
  75         snprintf(path, sizeof(path), "%s/%s/%s",
  76                         sys_dir_path, subdir, nr_rsvd_file);
  77         if (eal_parse_sysfs_value(path, &resv_pages) < 0)
  78                 return 0;
  79
  80         snprintf(path, sizeof(path), "%s/%s/%s",
  81                         sys_dir_path, subdir, nr_hp_file);
  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         /* adjust num_pages */
  90         if (num_pages > 0)
  91                 num_pages -= resv_pages;
  92
  93         return (int32_t)num_pages;
  94 }
  95
  96 static uint64_t
  97 get_default_hp_size(void)
  98 {
  99         const char proc_meminfo[] = "/proc/meminfo";
 100         const char str_hugepagesz[] = "Hugepagesize:";
 101         unsigned hugepagesz_len = sizeof(str_hugepagesz) - 1;
 102         char buffer[256];
 103         unsigned long long size = 0;
 104
 105         FILE *fd = fopen(proc_meminfo, "r");
 106         if (fd == NULL)
 107                 rte_panic("Cannot open %s\n", proc_meminfo);
 108         while(fgets(buffer, sizeof(buffer), fd)){
 109                 if (strncmp(buffer, str_hugepagesz, hugepagesz_len) == 0){
 110                         size = rte_str_to_size(&buffer[hugepagesz_len]);
 111                         break;
 112                 }
 113         }
 114         fclose(fd);
 115         if (size == 0)
 116                 rte_panic("Cannot get default hugepage size from %s\n", proc_meminfo);
 117         return size;
 118 }
 119
 120 static const char *
 121 get_hugepage_dir(uint64_t hugepage_sz)
 122 {
 123         enum proc_mount_fieldnames {
 124                 DEVICE = 0,
 125                 MOUNTPT,
 126                 FSTYPE,
 127                 OPTIONS,
 128                 _FIELDNAME_MAX
 129         };
 130         static uint64_t default_size = 0;
 131         const char proc_mounts[] = "/proc/mounts";
 132         const char hugetlbfs_str[] = "hugetlbfs";
 133         const size_t htlbfs_str_len = sizeof(hugetlbfs_str) - 1;
 134         const char pagesize_opt[] = "pagesize=";
 135         const size_t pagesize_opt_len = sizeof(pagesize_opt) - 1;
 136         const char split_tok = ' ';
 137         char *splitstr[_FIELDNAME_MAX];
 138         char buf[BUFSIZ];
 139         char *retval = NULL;
 140
 141         FILE *fd = fopen(proc_mounts, "r");
 142         if (fd == NULL)
 143                 rte_panic("Cannot open %s\n", proc_mounts);
 144
 145         if (default_size == 0)
 146                 default_size = get_default_hp_size();
 147
 148         while (fgets(buf, sizeof(buf), fd)){
 149                 if (rte_strsplit(buf, sizeof(buf), splitstr, _FIELDNAME_MAX,
 150                                 split_tok) != _FIELDNAME_MAX) {
 151                         RTE_LOG(ERR, EAL, "Error parsing %s\n", proc_mounts);
 152                         break; /* return NULL */
 153                 }
 154
 155                 /* we have a specified --huge-dir option, only examine that dir */
 156                 if (internal_config.hugepage_dir != NULL &&
 157                                 strcmp(splitstr[MOUNTPT], internal_config.hugepage_dir) != 0)
 158                         continue;
 159
 160                 if (strncmp(splitstr[FSTYPE], hugetlbfs_str, htlbfs_str_len) == 0){
 161                         const char *pagesz_str = strstr(splitstr[OPTIONS], pagesize_opt);
 162
 163                         /* if no explicit page size, the default page size is compared */
 164                         if (pagesz_str == NULL){
 165                                 if (hugepage_sz == default_size){
 166                                         retval = strdup(splitstr[MOUNTPT]);
 167                                         break;
 168                                 }
 169                         }
 170                         /* there is an explicit page size, so check it */
 171                         else {
 172                                 uint64_t pagesz = rte_str_to_size(&pagesz_str[pagesize_opt_len]);
 173                                 if (pagesz == hugepage_sz) {
 174                                         retval = strdup(splitstr[MOUNTPT]);
 175                                         break;
 176                                 }
 177                         }
 178                 } /* end if strncmp hugetlbfs */
 179         } /* end while fgets */
 180
 181         fclose(fd);
 182         return retval;
 183 }
 184
 185 static inline void
 186 swap_hpi(struct hugepage_info *a, struct hugepage_info *b)
 187 {
 188         char buf[sizeof(*a)];
 189         memcpy(buf, a, sizeof(buf));
 190         memcpy(a, b, sizeof(buf));
 191         memcpy(b, buf, sizeof(buf));
 192 }
 193
 194 /*
 195  * Clear the hugepage directory of whatever hugepage files
 196  * there are. Checks if the file is locked (i.e.
 197  * if it's in use by another DPDK process).
 198  */
 199 static int
 200 clear_hugedir(const char * hugedir)
 201 {
 202         DIR *dir;
 203         struct dirent *dirent;
 204         int dir_fd, fd, lck_result;
 205         const char filter[] = "*map_*"; /* matches hugepage files */
 206
 207         /* open directory */
 208         dir = opendir(hugedir);
 209         if (!dir) {
 210                 RTE_LOG(INFO, EAL, "Unable to open hugepage directory %s\n",
 211                                 hugedir);
 212                 goto error;
 213         }
 214         dir_fd = dirfd(dir);
 215
 216         dirent = readdir(dir);
 217         if (!dirent) {
 218                 RTE_LOG(INFO, EAL, "Unable to read hugepage directory %s\n",
 219                                 hugedir);
 220                 goto error;
 221         }
 222
 223         while(dirent != NULL){
 224                 /* skip files that don't match the hugepage pattern */
 225                 if (fnmatch(filter, dirent->d_name, 0) > 0) {
 226                         dirent = readdir(dir);
 227                         continue;
 228                 }
 229
 230                 /* try and lock the file */
 231                 fd = openat(dir_fd, dirent->d_name, O_RDONLY);
 232
 233                 /* skip to next file */
 234                 if (fd == -1) {
 235                         dirent = readdir(dir);
 236                         continue;
 237                 }
 238
 239                 /* non-blocking lock */
 240                 lck_result = flock(fd, LOCK_EX | LOCK_NB);
 241
 242                 /* if lock succeeds, unlock and remove the file */
 243                 if (lck_result != -1) {
 244                         flock(fd, LOCK_UN);
 245                         unlinkat(dir_fd, dirent->d_name, 0);
 246                 }
 247                 close (fd);
 248                 dirent = readdir(dir);
 249         }
 250
 251         closedir(dir);
 252         return 0;
 253
 254 error:
 255         if (dir)
 256                 closedir(dir);
 257
 258         RTE_LOG(INFO, EAL, "Error while clearing hugepage dir: %s\n",
 259                 strerror(errno));
 260
 261         return -1;
 262 }
 263
 264 /*
 265  * when we initialize the hugepage info, everything goes
 266  * to socket 0 by default. it will later get sorted by memory
 267  * initialization procedure.
 268  */
 269 int
 270 eal_hugepage_info_init(void)
 271 {
 272         const char dirent_start_text[] = "hugepages-";
 273         const size_t dirent_start_len = sizeof(dirent_start_text) - 1;
 274         unsigned i, num_sizes = 0;
 275
 276         DIR *dir = opendir(sys_dir_path);
 277         if (dir == NULL)
 278                 rte_panic("Cannot open directory %s to read system hugepage info\n",
 279                                 sys_dir_path);
 280
 281         struct dirent *dirent = readdir(dir);
 282         while(dirent != NULL){
 283                 if (strncmp(dirent->d_name, dirent_start_text, dirent_start_len) == 0){
 284                         struct hugepage_info *hpi = \
 285                                         &internal_config.hugepage_info[num_sizes];
 286                         hpi->hugepage_sz = rte_str_to_size(&dirent->d_name[dirent_start_len]);
 287                         hpi->hugedir = get_hugepage_dir(hpi->hugepage_sz);
 288
 289                         /* first, check if we have a mountpoint */
 290                         if (hpi->hugedir == NULL){
 291                                 int32_t num_pages;
 292                                 if ((num_pages = get_num_hugepages(dirent->d_name)) > 0)
 293                                         RTE_LOG(INFO, EAL, "%u hugepages of size %llu reserved, "\
 294                                                         "but no mounted hugetlbfs found for that size\n",
 295                                                         (unsigned)num_pages,
 296                                                         (unsigned long long)hpi->hugepage_sz);
 297                         } else {
 298                                 /* try to obtain a writelock */
 299                                 hpi->lock_descriptor = open(hpi->hugedir, O_RDONLY);
 300
 301                                 /* if blocking lock failed */
 302                                 if (flock(hpi->lock_descriptor, LOCK_EX) == -1) {
 303                                         RTE_LOG(CRIT, EAL, "Failed to lock hugepage directory!\n");
 304                                         closedir(dir);
 305                                         return -1;
 306                                 }
 307                                 /* clear out the hugepages dir from unused pages */
 308                                 if (clear_hugedir(hpi->hugedir) == -1) {
 309                                         closedir(dir);
 310                                         return -1;
 311                                 }
 312
 313                                 /* for now, put all pages into socket 0,
 314                                  * later they will be sorted */
 315                                 hpi->num_pages[0] = get_num_hugepages(dirent->d_name);
 316
 317 #ifndef RTE_ARCH_64
 318                                 /* for 32-bit systems, limit number of hugepages to 1GB per page size */
 319                                 hpi->num_pages[0] = RTE_MIN(hpi->num_pages[0],
 320                                                 RTE_PGSIZE_1G / hpi->hugepage_sz);
 321 #endif
 322
 323                                 num_sizes++;
 324                         }
 325                 }
 326                 dirent = readdir(dir);
 327         }
 328         closedir(dir);
 329         internal_config.num_hugepage_sizes = num_sizes;
 330
 331         /* sort the page directory entries by size, largest to smallest */
 332         for (i = 0; i < num_sizes; i++){
 333                 unsigned j;
 334                 for (j = i+1; j < num_sizes; j++)
 335                         if (internal_config.hugepage_info[j-1].hugepage_sz < \
 336                                         internal_config.hugepage_info[j].hugepage_sz)
 337                                 swap_hpi(&internal_config.hugepage_info[j-1],
 338                                                 &internal_config.hugepage_info[j]);
 339         }
 340
 341         /* now we have all info, check we have at least one valid size */
 342         for (i = 0; i < num_sizes; i++)
 343                 if (internal_config.hugepage_info[i].hugedir != NULL &&
 344                                 internal_config.hugepage_info[i].num_pages[0] > 0)
 345                         return 0;
 346
 347         /* no valid hugepage mounts available, return error */
 348         return -1;
 349 }