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 static int32_t
  65 get_num_hugepages(const char *subdir)
  66 {
  67         char path[PATH_MAX];
  68         long unsigned resv_pages, num_pages = 0;
  69         const char *nr_hp_file;
  70         const char *nr_rsvd_file = "resv_hugepages";
  71
  72         /* first, check how many reserved pages kernel reports */
  73         snprintf(path, sizeof(path), "%s/%s/%s",
  74                         sys_dir_path, subdir, nr_rsvd_file);
  75
  76         if (eal_parse_sysfs_value(path, &resv_pages) < 0)
  77                 return 0;
  78
  79         /* if secondary process, just look at the number of hugepages,
  80          * otherwise look at number of free hugepages */
  81         if (internal_config.process_type == RTE_PROC_SECONDARY)
  82                 nr_hp_file = "nr_hugepages";
  83         else
  84                 nr_hp_file = "free_hugepages";
  85
  86         snprintf(path, sizeof(path), "%s/%s/%s",
  87                         sys_dir_path, subdir, nr_hp_file);
  88
  89         if (eal_parse_sysfs_value(path, &num_pages) < 0)
  90                 return 0;
  91
  92         if (num_pages == 0)
  93                 RTE_LOG(WARNING, EAL, "No free hugepages reported in %s\n",
  94                                 subdir);
  95
  96         /* adjust num_pages in case of primary process */
  97         if (num_pages > 0 && internal_config.process_type == RTE_PROC_PRIMARY)
  98                 num_pages -= resv_pages;
  99
 100         return (int32_t)num_pages;
 101 }
 102
 103 static uint64_t
 104 get_default_hp_size(void)
 105 {
 106         const char proc_meminfo[] = "/proc/meminfo";
 107         const char str_hugepagesz[] = "Hugepagesize:";
 108         unsigned hugepagesz_len = sizeof(str_hugepagesz) - 1;
 109         char buffer[256];
 110         unsigned long long size = 0;
 111
 112         FILE *fd = fopen(proc_meminfo, "r");
 113         if (fd == NULL)
 114                 rte_panic("Cannot open %s\n", proc_meminfo);
 115         while(fgets(buffer, sizeof(buffer), fd)){
 116                 if (strncmp(buffer, str_hugepagesz, hugepagesz_len) == 0){
 117                         size = rte_str_to_size(&buffer[hugepagesz_len]);
 118                         break;
 119                 }
 120         }
 121         fclose(fd);
 122         if (size == 0)
 123                 rte_panic("Cannot get default hugepage size from %s\n", proc_meminfo);
 124         return size;
 125 }
 126
 127 static const char *
 128 get_hugepage_dir(uint64_t hugepage_sz)
 129 {
 130         enum proc_mount_fieldnames {
 131                 DEVICE = 0,
 132                 MOUNTPT,
 133                 FSTYPE,
 134                 OPTIONS,
 135                 _FIELDNAME_MAX
 136         };
 137         static uint64_t default_size = 0;
 138         const char proc_mounts[] = "/proc/mounts";
 139         const char hugetlbfs_str[] = "hugetlbfs";
 140         const size_t htlbfs_str_len = sizeof(hugetlbfs_str) - 1;
 141         const char pagesize_opt[] = "pagesize=";
 142         const size_t pagesize_opt_len = sizeof(pagesize_opt) - 1;
 143         const char split_tok = ' ';
 144         char *splitstr[_FIELDNAME_MAX];
 145         char buf[BUFSIZ];
 146         char *retval = NULL;
 147
 148         FILE *fd = fopen(proc_mounts, "r");
 149         if (fd == NULL)
 150                 rte_panic("Cannot open %s\n", proc_mounts);
 151
 152         if (default_size == 0)
 153                 default_size = get_default_hp_size();
 154
 155         while (fgets(buf, sizeof(buf), fd)){
 156                 if (rte_strsplit(buf, sizeof(buf), splitstr, _FIELDNAME_MAX,
 157                                 split_tok) != _FIELDNAME_MAX) {
 158                         RTE_LOG(ERR, EAL, "Error parsing %s\n", proc_mounts);
 159                         break; /* return NULL */
 160                 }
 161
 162                 /* we have a specified --huge-dir option, only examine that dir */
 163                 if (internal_config.hugepage_dir != NULL &&
 164                                 strcmp(splitstr[MOUNTPT], internal_config.hugepage_dir) != 0)
 165                         continue;
 166
 167                 if (strncmp(splitstr[FSTYPE], hugetlbfs_str, htlbfs_str_len) == 0){
 168                         const char *pagesz_str = strstr(splitstr[OPTIONS], pagesize_opt);
 169
 170                         /* if no explicit page size, the default page size is compared */
 171                         if (pagesz_str == NULL){
 172                                 if (hugepage_sz == default_size){
 173                                         retval = strdup(splitstr[MOUNTPT]);
 174                                         break;
 175                                 }
 176                         }
 177                         /* there is an explicit page size, so check it */
 178                         else {
 179                                 uint64_t pagesz = rte_str_to_size(&pagesz_str[pagesize_opt_len]);
 180                                 if (pagesz == hugepage_sz) {
 181                                         retval = strdup(splitstr[MOUNTPT]);
 182                                         break;
 183                                 }
 184                         }
 185                 } /* end if strncmp hugetlbfs */
 186         } /* end while fgets */
 187
 188         fclose(fd);
 189         return retval;
 190 }
 191
 192 static inline void
 193 swap_hpi(struct hugepage_info *a, struct hugepage_info *b)
 194 {
 195         char buf[sizeof(*a)];
 196         memcpy(buf, a, sizeof(buf));
 197         memcpy(a, b, sizeof(buf));
 198         memcpy(b, buf, sizeof(buf));
 199 }
 200
 201 /*
 202  * Clear the hugepage directory of whatever hugepage files
 203  * there are. Checks if the file is locked (i.e.
 204  * if it's in use by another DPDK process).
 205  */
 206 static int
 207 clear_hugedir(const char * hugedir)
 208 {
 209         DIR *dir;
 210         struct dirent *dirent;
 211         int dir_fd, fd, lck_result;
 212         const char filter[] = "*map_*"; /* matches hugepage files */
 213
 214         /* open directory */
 215         dir = opendir(hugedir);
 216         if (!dir) {
 217                 RTE_LOG(INFO, EAL, "Unable to open hugepage directory %s\n",
 218                                 hugedir);
 219                 goto error;
 220         }
 221         dir_fd = dirfd(dir);
 222
 223         dirent = readdir(dir);
 224         if (!dirent) {
 225                 RTE_LOG(INFO, EAL, "Unable to read hugepage directory %s\n",
 226                                 hugedir);
 227                 goto error;
 228         }
 229
 230         while(dirent != NULL){
 231                 /* skip files that don't match the hugepage pattern */
 232                 if (fnmatch(filter, dirent->d_name, 0) > 0) {
 233                         dirent = readdir(dir);
 234                         continue;
 235                 }
 236
 237                 /* try and lock the file */
 238                 fd = openat(dir_fd, dirent->d_name, O_RDONLY);
 239
 240                 /* skip to next file */
 241                 if (fd == -1) {
 242                         dirent = readdir(dir);
 243                         continue;
 244                 }
 245
 246                 /* non-blocking lock */
 247                 lck_result = flock(fd, LOCK_EX | LOCK_NB);
 248
 249                 /* if lock succeeds, unlock and remove the file */
 250                 if (lck_result != -1) {
 251                         flock(fd, LOCK_UN);
 252                         unlinkat(dir_fd, dirent->d_name, 0);
 253                 }
 254                 close (fd);
 255                 dirent = readdir(dir);
 256         }
 257
 258         closedir(dir);
 259         return 0;
 260
 261 error:
 262         if (dir)
 263                 closedir(dir);
 264
 265         RTE_LOG(INFO, EAL, "Error while clearing hugepage dir: %s\n",
 266                 strerror(errno));
 267
 268         return -1;
 269 }
 270
 271 /*
 272  * when we initialize the hugepage info, everything goes
 273  * to socket 0 by default. it will later get sorted by memory
 274  * initialization procedure.
 275  */
 276 int
 277 eal_hugepage_info_init(void)
 278 {
 279         const char dirent_start_text[] = "hugepages-";
 280         const size_t dirent_start_len = sizeof(dirent_start_text) - 1;
 281         unsigned i, num_sizes = 0;
 282
 283         DIR *dir = opendir(sys_dir_path);
 284         if (dir == NULL)
 285                 rte_panic("Cannot open directory %s to read system hugepage info\n",
 286                                 sys_dir_path);
 287
 288         struct dirent *dirent = readdir(dir);
 289         while(dirent != NULL){
 290                 if (strncmp(dirent->d_name, dirent_start_text, dirent_start_len) == 0){
 291                         struct hugepage_info *hpi = \
 292                                         &internal_config.hugepage_info[num_sizes];
 293                         hpi->hugepage_sz = rte_str_to_size(&dirent->d_name[dirent_start_len]);
 294                         hpi->hugedir = get_hugepage_dir(hpi->hugepage_sz);
 295
 296                         /* first, check if we have a mountpoint */
 297                         if (hpi->hugedir == NULL){
 298                                 int32_t num_pages;
 299                                 if ((num_pages = get_num_hugepages(dirent->d_name)) > 0)
 300                                         RTE_LOG(INFO, EAL, "%u hugepages of size %llu reserved, "\
 301                                                         "but no mounted hugetlbfs found for that size\n",
 302                                                         (unsigned)num_pages,
 303                                                         (unsigned long long)hpi->hugepage_sz);
 304                         } else {
 305                                 /* try to obtain a writelock */
 306                                 hpi->lock_descriptor = open(hpi->hugedir, O_RDONLY);
 307
 308                                 /* if blocking lock failed */
 309                                 if (flock(hpi->lock_descriptor, LOCK_EX) == -1) {
 310                                         RTE_LOG(CRIT, EAL, "Failed to lock hugepage directory!\n");
 311                                         closedir(dir);
 312                                         return -1;
 313                                 }
 314                                 /* clear out the hugepages dir from unused pages */
 315                                 if (clear_hugedir(hpi->hugedir) == -1) {
 316                                         closedir(dir);
 317                                         return -1;
 318                                 }
 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_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 }