fs/file.c

   1 /*
   2  *  linux/fs/file.c
   3  *
   4  *  Copyright (C) 1998-1999, Stephen Tweedie and Bill Hawes
   5  *
   6  *  Manage the dynamic fd arrays in the process files_struct.
   7  */
   8
   9 #include <linux/fs.h>
  10 #include <linux/mm.h>
  11 #include <linux/time.h>
  12 #include <linux/slab.h>
  13 #include <linux/vmalloc.h>
  14 #include <linux/file.h>
  15 #include <linux/fdtable.h>
  16 #include <linux/bitops.h>
  17 #include <linux/interrupt.h>
  18 #include <linux/spinlock.h>
  19 #include <linux/rcupdate.h>
  20 #include <linux/workqueue.h>
  21
  22 struct fdtable_defer {
  23         spinlock_t lock;
  24         struct work_struct wq;
  25         struct fdtable *next;
  26 };
  27
  28 int sysctl_nr_open __read_mostly = 1024*1024;
  29 int sysctl_nr_open_min = BITS_PER_LONG;
  30 int sysctl_nr_open_max = 1024 * 1024; /* raised later */
  31
  32 /*
  33  * We use this list to defer free fdtables that have vmalloced
  34  * sets/arrays. By keeping a per-cpu list, we avoid having to embed
  35  * the work_struct in fdtable itself which avoids a 64 byte (i386) increase in
  36  * this per-task structure.
  37  */
  38 static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
  39
  40 static inline void * alloc_fdmem(unsigned int size)
  41 {
  42         if (size <= PAGE_SIZE)
  43                 return kmalloc(size, GFP_KERNEL);
  44         else
  45                 return vmalloc(size);
  46 }
  47
  48 static inline void free_fdarr(struct fdtable *fdt)
  49 {
  50         if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *)))
  51                 kfree(fdt->fd);
  52         else
  53                 vfree(fdt->fd);
  54 }
  55
  56 static inline void free_fdset(struct fdtable *fdt)
  57 {
  58         if (fdt->max_fds <= (PAGE_SIZE * BITS_PER_BYTE / 2))
  59                 kfree(fdt->open_fds);
  60         else
  61                 vfree(fdt->open_fds);
  62 }
  63
  64 static void free_fdtable_work(struct work_struct *work)
  65 {
  66         struct fdtable_defer *f =
  67                 container_of(work, struct fdtable_defer, wq);
  68         struct fdtable *fdt;
  69
  70         spin_lock_bh(&f->lock);
  71         fdt = f->next;
  72         f->next = NULL;
  73         spin_unlock_bh(&f->lock);
  74         while(fdt) {
  75                 struct fdtable *next = fdt->next;
  76                 vfree(fdt->fd);
  77                 free_fdset(fdt);
  78                 kfree(fdt);
  79                 fdt = next;
  80         }
  81 }
  82
  83 void free_fdtable_rcu(struct rcu_head *rcu)
  84 {
  85         struct fdtable *fdt = container_of(rcu, struct fdtable, rcu);
  86         struct fdtable_defer *fddef;
  87
  88         BUG_ON(!fdt);
  89
  90         if (fdt->max_fds <= NR_OPEN_DEFAULT) {
  91                 /*
  92                  * This fdtable is embedded in the files structure and that
  93                  * structure itself is getting destroyed.
  94                  */
  95                 kmem_cache_free(files_cachep,
  96                                 container_of(fdt, struct files_struct, fdtab));
  97                 return;
  98         }
  99         if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *))) {
 100                 kfree(fdt->fd);
 101                 kfree(fdt->open_fds);
 102                 kfree(fdt);
 103         } else {
 104                 fddef = &get_cpu_var(fdtable_defer_list);
 105                 spin_lock(&fddef->lock);
 106                 fdt->next = fddef->next;
 107                 fddef->next = fdt;
 108                 /* vmallocs are handled from the workqueue context */
 109                 schedule_work(&fddef->wq);
 110                 spin_unlock(&fddef->lock);
 111                 put_cpu_var(fdtable_defer_list);
 112         }
 113 }
 114
 115 /*
 116  * Expand the fdset in the files_struct.  Called with the files spinlock
 117  * held for write.
 118  */
 119 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
 120 {
 121         unsigned int cpy, set;
 122
 123         BUG_ON(nfdt->max_fds < ofdt->max_fds);
 124
 125         cpy = ofdt->max_fds * sizeof(struct file *);
 126         set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
 127         memcpy(nfdt->fd, ofdt->fd, cpy);
 128         memset((char *)(nfdt->fd) + cpy, 0, set);
 129
 130         cpy = ofdt->max_fds / BITS_PER_BYTE;
 131         set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
 132         memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
 133         memset((char *)(nfdt->open_fds) + cpy, 0, set);
 134         memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
 135         memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
 136 }
 137
 138 static struct fdtable * alloc_fdtable(unsigned int nr)
 139 {
 140         struct fdtable *fdt;
 141         char *data;
 142
 143         /*
 144          * Figure out how many fds we actually want to support in this fdtable.
 145          * Allocation steps are keyed to the size of the fdarray, since it
 146          * grows far faster than any of the other dynamic data. We try to fit
 147          * the fdarray into comfortable page-tuned chunks: starting at 1024B
 148          * and growing in powers of two from there on.
 149          */
 150         nr /= (1024 / sizeof(struct file *));
 151         nr = roundup_pow_of_two(nr + 1);
 152         nr *= (1024 / sizeof(struct file *));
 153         /*
 154          * Note that this can drive nr *below* what we had passed if sysctl_nr_open
 155          * had been set lower between the check in expand_files() and here.  Deal
 156          * with that in caller, it's cheaper that way.
 157          *
 158          * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
 159          * bitmaps handling below becomes unpleasant, to put it mildly...
 160          */
 161         if (unlikely(nr > sysctl_nr_open))
 162                 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
 163
 164         fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
 165         if (!fdt)
 166                 goto out;
 167         fdt->max_fds = nr;
 168         data = alloc_fdmem(nr * sizeof(struct file *));
 169         if (!data)
 170                 goto out_fdt;
 171         fdt->fd = (struct file **)data;
 172         data = alloc_fdmem(max_t(unsigned int,
 173                                  2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
 174         if (!data)
 175                 goto out_arr;
 176         fdt->open_fds = (fd_set *)data;
 177         data += nr / BITS_PER_BYTE;
 178         fdt->close_on_exec = (fd_set *)data;
 179         INIT_RCU_HEAD(&fdt->rcu);
 180         fdt->next = NULL;
 181
 182         return fdt;
 183
 184 out_arr:
 185         free_fdarr(fdt);
 186 out_fdt:
 187         kfree(fdt);
 188 out:
 189         return NULL;
 190 }
 191
 192 /*
 193  * Expand the file descriptor table.
 194  * This function will allocate a new fdtable and both fd array and fdset, of
 195  * the given size.
 196  * Return <0 error code on error; 1 on successful completion.
 197  * The files->file_lock should be held on entry, and will be held on exit.
 198  */
 199 static int expand_fdtable(struct files_struct *files, int nr)
 200         __releases(files->file_lock)
 201         __acquires(files->file_lock)
 202 {
 203         struct fdtable *new_fdt, *cur_fdt;
 204
 205         spin_unlock(&files->file_lock);
 206         new_fdt = alloc_fdtable(nr);
 207         spin_lock(&files->file_lock);
 208         if (!new_fdt)
 209                 return -ENOMEM;
 210         /*
 211          * extremely unlikely race - sysctl_nr_open decreased between the check in
 212          * caller and alloc_fdtable().  Cheaper to catch it here...
 213          */
 214         if (unlikely(new_fdt->max_fds <= nr)) {
 215                 free_fdarr(new_fdt);
 216                 free_fdset(new_fdt);
 217                 kfree(new_fdt);
 218                 return -EMFILE;
 219         }
 220         /*
 221          * Check again since another task may have expanded the fd table while
 222          * we dropped the lock
 223          */
 224         cur_fdt = files_fdtable(files);
 225         if (nr >= cur_fdt->max_fds) {
 226                 /* Continue as planned */
 227                 copy_fdtable(new_fdt, cur_fdt);
 228                 rcu_assign_pointer(files->fdt, new_fdt);
 229                 if (cur_fdt->max_fds > NR_OPEN_DEFAULT)
 230                         free_fdtable(cur_fdt);
 231         } else {
 232                 /* Somebody else expanded, so undo our attempt */
 233                 free_fdarr(new_fdt);
 234                 free_fdset(new_fdt);
 235                 kfree(new_fdt);
 236         }
 237         return 1;
 238 }
 239
 240 /*
 241  * Expand files.
 242  * This function will expand the file structures, if the requested size exceeds
 243  * the current capacity and there is room for expansion.
 244  * Return <0 error code on error; 0 when nothing done; 1 when files were
 245  * expanded and execution may have blocked.
 246  * The files->file_lock should be held on entry, and will be held on exit.
 247  */
 248 int expand_files(struct files_struct *files, int nr)
 249 {
 250         struct fdtable *fdt;
 251
 252         fdt = files_fdtable(files);
 253         /* Do we need to expand? */
 254         if (nr < fdt->max_fds)
 255                 return 0;
 256         /* Can we expand? */
 257         if (nr >= sysctl_nr_open)
 258                 return -EMFILE;
 259
 260         /* All good, so we try */
 261         return expand_fdtable(files, nr);
 262 }
 263
 264 static int count_open_files(struct fdtable *fdt)
 265 {
 266         int size = fdt->max_fds;
 267         int i;
 268
 269         /* Find the last open fd */
 270         for (i = size/(8*sizeof(long)); i > 0; ) {
 271                 if (fdt->open_fds->fds_bits[--i])
 272                         break;
 273         }
 274         i = (i+1) * 8 * sizeof(long);
 275         return i;
 276 }
 277
 278 /*
 279  * Allocate a new files structure and copy contents from the
 280  * passed in files structure.
 281  * errorp will be valid only when the returned files_struct is NULL.
 282  */
 283 struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
 284 {
 285         struct files_struct *newf;
 286         struct file **old_fds, **new_fds;
 287         int open_files, size, i;
 288         struct fdtable *old_fdt, *new_fdt;
 289
 290         *errorp = -ENOMEM;
 291         newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
 292         if (!newf)
 293                 goto out;
 294
 295         atomic_set(&newf->count, 1);
 296
 297         spin_lock_init(&newf->file_lock);
 298         newf->next_fd = 0;
 299         new_fdt = &newf->fdtab;
 300         new_fdt->max_fds = NR_OPEN_DEFAULT;
 301         new_fdt->close_on_exec = (fd_set *)&newf->close_on_exec_init;
 302         new_fdt->open_fds = (fd_set *)&newf->open_fds_init;
 303         new_fdt->fd = &newf->fd_array[0];
 304         INIT_RCU_HEAD(&new_fdt->rcu);
 305         new_fdt->next = NULL;
 306
 307         spin_lock(&oldf->file_lock);
 308         old_fdt = files_fdtable(oldf);
 309         open_files = count_open_files(old_fdt);
 310
 311         /*
 312          * Check whether we need to allocate a larger fd array and fd set.
 313          */
 314         while (unlikely(open_files > new_fdt->max_fds)) {
 315                 spin_unlock(&oldf->file_lock);
 316
 317                 if (new_fdt != &newf->fdtab) {
 318                         free_fdarr(new_fdt);
 319                         free_fdset(new_fdt);
 320                         kfree(new_fdt);
 321                 }
 322
 323                 new_fdt = alloc_fdtable(open_files - 1);
 324                 if (!new_fdt) {
 325                         *errorp = -ENOMEM;
 326                         goto out_release;
 327                 }
 328
 329                 /* beyond sysctl_nr_open; nothing to do */
 330                 if (unlikely(new_fdt->max_fds < open_files)) {
 331                         free_fdarr(new_fdt);
 332                         free_fdset(new_fdt);
 333                         kfree(new_fdt);
 334                         *errorp = -EMFILE;
 335                         goto out_release;
 336                 }
 337
 338                 /*
 339                  * Reacquire the oldf lock and a pointer to its fd table
 340                  * who knows it may have a new bigger fd table. We need
 341                  * the latest pointer.
 342                  */
 343                 spin_lock(&oldf->file_lock);
 344                 old_fdt = files_fdtable(oldf);
 345                 open_files = count_open_files(old_fdt);
 346         }
 347
 348         old_fds = old_fdt->fd;
 349         new_fds = new_fdt->fd;
 350
 351         memcpy(new_fdt->open_fds->fds_bits,
 352                 old_fdt->open_fds->fds_bits, open_files/8);
 353         memcpy(new_fdt->close_on_exec->fds_bits,
 354                 old_fdt->close_on_exec->fds_bits, open_files/8);
 355
 356         for (i = open_files; i != 0; i--) {
 357                 struct file *f = *old_fds++;
 358                 if (f) {
 359                         get_file(f);
 360                 } else {
 361                         /*
 362                          * The fd may be claimed in the fd bitmap but not yet
 363                          * instantiated in the files array if a sibling thread
 364                          * is partway through open().  So make sure that this
 365                          * fd is available to the new process.
 366                          */
 367                         FD_CLR(open_files - i, new_fdt->open_fds);
 368                 }
 369                 rcu_assign_pointer(*new_fds++, f);
 370         }
 371         spin_unlock(&oldf->file_lock);
 372
 373         /* compute the remainder to be cleared */
 374         size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
 375
 376         /* This is long word aligned thus could use a optimized version */
 377         memset(new_fds, 0, size);
 378
 379         if (new_fdt->max_fds > open_files) {
 380                 int left = (new_fdt->max_fds-open_files)/8;
 381                 int start = open_files / (8 * sizeof(unsigned long));
 382
 383                 memset(&new_fdt->open_fds->fds_bits[start], 0, left);
 384                 memset(&new_fdt->close_on_exec->fds_bits[start], 0, left);
 385         }
 386
 387         rcu_assign_pointer(newf->fdt, new_fdt);
 388
 389         return newf;
 390
 391 out_release:
 392         kmem_cache_free(files_cachep, newf);
 393 out:
 394         return NULL;
 395 }
 396
 397 static void __devinit fdtable_defer_list_init(int cpu)
 398 {
 399         struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
 400         spin_lock_init(&fddef->lock);
 401         INIT_WORK(&fddef->wq, free_fdtable_work);
 402         fddef->next = NULL;
 403 }
 404
 405 void __init files_defer_init(void)
 406 {
 407         int i;
 408         for_each_possible_cpu(i)
 409                 fdtable_defer_list_init(i);
 410         sysctl_nr_open_max = min((size_t)INT_MAX, ~(size_t)0/sizeof(void *)) &
 411                              -BITS_PER_LONG;
 412 }
 413
 414 struct files_struct init_files = {
 415         .count          = ATOMIC_INIT(1),
 416         .fdt            = &init_files.fdtab,
 417         .fdtab          = {
 418                 .max_fds        = NR_OPEN_DEFAULT,
 419                 .fd             = &init_files.fd_array[0],
 420                 .close_on_exec  = (fd_set *)&init_files.close_on_exec_init,
 421                 .open_fds       = (fd_set *)&init_files.open_fds_init,
 422                 .rcu            = RCU_HEAD_INIT,
 423         },
 424         .file_lock      = __SPIN_LOCK_UNLOCKED(init_task.file_lock),
 425 };