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_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 resv_pages, num_pages = 0;
  70         const char *nr_hp_file;
  71         const char *nr_rsvd_file = "resv_hugepages";
  72
  73         /* first, check how many reserved pages kernel reports */
  74         rte_snprintf(path, sizeof(path), "%s/%s/%s",
  75                         sys_dir_path, subdir, nr_rsvd_file);
  76
  77         if (eal_parse_sysfs_value(path, &resv_pages) < 0)
  78                 return 0;
  79
  80         /* if secondary process, just look at the number of hugepages,
  81          * otherwise look at number of free hugepages */
  82         if (internal_config.process_type == RTE_PROC_SECONDARY)
  83                 nr_hp_file = "nr_hugepages";
  84         else
  85                 nr_hp_file = "free_hugepages";
  86
  87         memset(path, 0, sizeof(path));
  88
  89         rte_snprintf(path, sizeof(path), "%s/%s/%s",
  90                         sys_dir_path, subdir, nr_hp_file);
  91
  92         if (eal_parse_sysfs_value(path, &num_pages) < 0)
  93                 return 0;
  94
  95         if (num_pages == 0)
  96                 RTE_LOG(WARNING, EAL, "No free hugepages reported in %s\n",
  97                                 subdir);
  98
  99         /* adjust num_pages in case of primary process */
 100         if (num_pages > 0 && internal_config.process_type == RTE_PROC_PRIMARY)
 101                 num_pages -= resv_pages;
 102
 103         return (int32_t)num_pages;
 104 }
 105
 106 static uint64_t
 107 get_default_hp_size(void)
 108 {
 109         const char proc_meminfo[] = "/proc/meminfo";
 110         const char str_hugepagesz[] = "Hugepagesize:";
 111         unsigned hugepagesz_len = sizeof(str_hugepagesz) - 1;
 112         char buffer[256];
 113         unsigned long long size = 0;
 114
 115         FILE *fd = fopen(proc_meminfo, "r");
 116         if (fd == NULL)
 117                 rte_panic("Cannot open %s\n", proc_meminfo);
 118         while(fgets(buffer, sizeof(buffer), fd)){
 119                 if (strncmp(buffer, str_hugepagesz, hugepagesz_len) == 0){
 120                         size = rte_str_to_size(&buffer[hugepagesz_len]);
 121                         break;
 122                 }
 123         }
 124         fclose(fd);
 125         if (size == 0)
 126                 rte_panic("Cannot get default hugepage size from %s\n", proc_meminfo);
 127         return size;
 128 }
 129
 130 static const char *
 131 get_hugepage_dir(uint64_t hugepage_sz)
 132 {
 133         enum proc_mount_fieldnames {
 134                 DEVICE = 0,
 135                 MOUNTPT,
 136                 FSTYPE,
 137                 OPTIONS,
 138                 _FIELDNAME_MAX
 139         };
 140         static uint64_t default_size = 0;
 141         const char proc_mounts[] = "/proc/mounts";
 142         const char hugetlbfs_str[] = "hugetlbfs";
 143         const size_t htlbfs_str_len = sizeof(hugetlbfs_str) - 1;
 144         const char pagesize_opt[] = "pagesize=";
 145         const size_t pagesize_opt_len = sizeof(pagesize_opt) - 1;
 146         const char split_tok = ' ';
 147         char *splitstr[_FIELDNAME_MAX];
 148         char buf[BUFSIZ];
 149         char *retval = NULL;
 150
 151         FILE *fd = fopen(proc_mounts, "r");
 152         if (fd == NULL)
 153                 rte_panic("Cannot open %s\n", proc_mounts);
 154
 155         if (default_size == 0)
 156                 default_size = get_default_hp_size();
 157
 158         while (fgets(buf, sizeof(buf), fd)){
 159                 if (rte_strsplit(buf, sizeof(buf), splitstr, _FIELDNAME_MAX,
 160                                 split_tok) != _FIELDNAME_MAX) {
 161                         RTE_LOG(ERR, EAL, "Error parsing %s\n", proc_mounts);
 162                         break; /* return NULL */
 163                 }
 164
 165                 /* we have a specified --huge-dir option, only examine that dir */
 166                 if (internal_config.hugepage_dir != NULL &&
 167                                 strcmp(splitstr[MOUNTPT], internal_config.hugepage_dir) != 0)
 168                         continue;
 169
 170                 if (strncmp(splitstr[FSTYPE], hugetlbfs_str, htlbfs_str_len) == 0){
 171                         const char *pagesz_str = strstr(splitstr[OPTIONS], pagesize_opt);
 172
 173                         /* if no explicit page size, the default page size is compared */
 174                         if (pagesz_str == NULL){
 175                                 if (hugepage_sz == default_size){
 176                                         retval = strdup(splitstr[MOUNTPT]);
 177                                         break;
 178                                 }
 179                         }
 180                         /* there is an explicit page size, so check it */
 181                         else {
 182                                 uint64_t pagesz = rte_str_to_size(&pagesz_str[pagesize_opt_len]);
 183                                 if (pagesz == hugepage_sz) {
 184                                         retval = strdup(splitstr[MOUNTPT]);
 185                                         break;
 186                                 }
 187                         }
 188                 } /* end if strncmp hugetlbfs */
 189         } /* end while fgets */
 190
 191         fclose(fd);
 192         return retval;
 193 }
 194
 195 static inline void
 196 swap_hpi(struct hugepage_info *a, struct hugepage_info *b)
 197 {
 198         char buf[sizeof(*a)];
 199         memcpy(buf, a, sizeof(buf));
 200         memcpy(a, b, sizeof(buf));
 201         memcpy(b, buf, sizeof(buf));
 202 }
 203
 204 /*
 205  * Clear the hugepage directory of whatever hugepage files
 206  * there are. Checks if the file is locked (i.e.
 207  * if it's in use by another DPDK process).
 208  */
 209 static int
 210 clear_hugedir(const char * hugedir)
 211 {
 212         DIR *dir;
 213         struct dirent *dirent;
 214         int dir_fd, fd, lck_result;
 215         const char filter[] = "*map_*"; /* matches hugepage files */
 216
 217         /* open directory */
 218         dir = opendir(hugedir);
 219         if (!dir) {
 220                 RTE_LOG(INFO, EAL, "Unable to open hugepage directory %s\n",
 221                                 hugedir);
 222                 goto error;
 223         }
 224         dir_fd = dirfd(dir);
 225
 226         dirent = readdir(dir);
 227         if (!dirent) {
 228                 RTE_LOG(INFO, EAL, "Unable to read hugepage directory %s\n",
 229                                 hugedir);
 230                 goto error;
 231         }
 232
 233         while(dirent != NULL){
 234                 /* skip files that don't match the hugepage pattern */
 235                 if (fnmatch(filter, dirent->d_name, 0) > 0) {
 236                         dirent = readdir(dir);
 237                         continue;
 238                 }
 239
 240                 /* try and lock the file */
 241                 fd = openat(dir_fd, dirent->d_name, O_RDONLY);
 242
 243                 /* skip to next file */
 244                 if (fd == -1) {
 245                         dirent = readdir(dir);
 246                         continue;
 247                 }
 248
 249                 /* non-blocking lock */
 250                 lck_result = flock(fd, LOCK_EX | LOCK_NB);
 251
 252                 /* if lock succeeds, unlock and remove the file */
 253                 if (lck_result != -1) {
 254                         flock(fd, LOCK_UN);
 255                         unlinkat(dir_fd, dirent->d_name, 0);
 256                 }
 257                 close (fd);
 258                 dirent = readdir(dir);
 259         }
 260
 261         closedir(dir);
 262         return 0;
 263
 264 error:
 265         if (dir)
 266                 closedir(dir);
 267
 268         RTE_LOG(INFO, EAL, "Error while clearing hugepage dir: %s\n",
 269                 strerror(errno));
 270
 271         return -1;
 272 }
 273
 274 /*
 275  * when we initialize the hugepage info, everything goes
 276  * to socket 0 by default. it will later get sorted by memory
 277  * initialization procedure.
 278  */
 279 int
 280 eal_hugepage_info_init(void)
 281 {
 282         const char dirent_start_text[] = "hugepages-";
 283         const size_t dirent_start_len = sizeof(dirent_start_text) - 1;
 284         unsigned i, num_sizes = 0;
 285
 286         DIR *dir = opendir(sys_dir_path);
 287         if (dir == NULL)
 288                 rte_panic("Cannot open directory %s to read system hugepage info\n",
 289                                 sys_dir_path);
 290
 291         struct dirent *dirent = readdir(dir);
 292         while(dirent != NULL){
 293                 if (strncmp(dirent->d_name, dirent_start_text, dirent_start_len) == 0){
 294                         struct hugepage_info *hpi = \
 295                                         &internal_config.hugepage_info[num_sizes];
 296                         hpi->hugepage_sz = rte_str_to_size(&dirent->d_name[dirent_start_len]);
 297                         hpi->hugedir = get_hugepage_dir(hpi->hugepage_sz);
 298
 299                         /* first, check if we have a mountpoint */
 300                         if (hpi->hugedir == NULL){
 301                                 int32_t num_pages;
 302                                 if ((num_pages = get_num_hugepages(dirent->d_name)) > 0)
 303                                         RTE_LOG(INFO, EAL, "%u hugepages of size %llu reserved, "\
 304                                                         "but no mounted hugetlbfs found for that size\n",
 305                                                         (unsigned)num_pages,
 306                                                         (unsigned long long)hpi->hugepage_sz);
 307                         } else {
 308                                 /* try to obtain a writelock */
 309                                 hpi->lock_descriptor = open(hpi->hugedir, O_RDONLY);
 310
 311                                 /* if blocking lock failed */
 312                                 if (flock(hpi->lock_descriptor, LOCK_EX) == -1) {
 313                                         RTE_LOG(CRIT, EAL, "Failed to lock hugepage directory!\n");
 314                                         return -1;
 315                                 }
 316                                 /* clear out the hugepages dir from unused pages */
 317                                 if (clear_hugedir(hpi->hugedir) == -1)
 318                                         return -1;
 319
 320                                 /* for now, put all pages into socket 0,
 321                                  * later they will be sorted */
 322                                 hpi->num_pages[0] = get_num_hugepages(dirent->d_name);
 323
 324 #ifndef RTE_ARCH_X86_64
 325                                 /* for 32-bit systems, limit number of hugepages to 1GB per page size */
 326                                 hpi->num_pages[0] = RTE_MIN(hpi->num_pages[0],
 327                                                 RTE_PGSIZE_1G / hpi->hugepage_sz);
 328 #endif
 329
 330                                 num_sizes++;
 331                         }
 332                 }
 333                 dirent = readdir(dir);
 334         }
 335         closedir(dir);
 336         internal_config.num_hugepage_sizes = num_sizes;
 337
 338         /* sort the page directory entries by size, largest to smallest */
 339         for (i = 0; i < num_sizes; i++){
 340                 unsigned j;
 341                 for (j = i+1; j < num_sizes; j++)
 342                         if (internal_config.hugepage_info[j-1].hugepage_sz < \
 343                                         internal_config.hugepage_info[j].hugepage_sz)
 344                                 swap_hpi(&internal_config.hugepage_info[j-1],
 345                                                 &internal_config.hugepage_info[j]);
 346         }
 347
 348         /* now we have all info, check we have at least one valid size */
 349         for (i = 0; i < num_sizes; i++)
 350                 if (internal_config.hugepage_info[i].hugedir != NULL &&
 351                                 internal_config.hugepage_info[i].num_pages[0] > 0)
 352                         return 0;
 353
 354         /* no valid hugepage mounts available, return error */
 355         return -1;
 356 }