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
  30 /*
  31  * We use this list to defer free fdtables that have vmalloced
  32  * sets/arrays. By keeping a per-cpu list, we avoid having to embed
  33  * the work_struct in fdtable itself which avoids a 64 byte (i386) increase in
  34  * this per-task structure.
  35  */
  36 static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
  37
  38 static inline void * alloc_fdmem(unsigned int size)
  39 {
  40         if (size <= PAGE_SIZE)
  41                 return kmalloc(size, GFP_KERNEL);
  42         else
  43                 return vmalloc(size);
  44 }
  45
  46 static inline void free_fdarr(struct fdtable *fdt)
  47 {
  48         if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *)))
  49                 kfree(fdt->fd);
  50         else
  51                 vfree(fdt->fd);
  52 }
  53
  54 static inline void free_fdset(struct fdtable *fdt)
  55 {
  56         if (fdt->max_fds <= (PAGE_SIZE * BITS_PER_BYTE / 2))
  57                 kfree(fdt->open_fds);
  58         else
  59                 vfree(fdt->open_fds);
  60 }
  61
  62 static void free_fdtable_work(struct work_struct *work)
  63 {
  64         struct fdtable_defer *f =
  65                 container_of(work, struct fdtable_defer, wq);
  66         struct fdtable *fdt;
  67
  68         spin_lock_bh(&f->lock);
  69         fdt = f->next;
  70         f->next = NULL;
  71         spin_unlock_bh(&f->lock);
  72         while(fdt) {
  73                 struct fdtable *next = fdt->next;
  74                 vfree(fdt->fd);
  75                 free_fdset(fdt);
  76                 kfree(fdt);
  77                 fdt = next;
  78         }
  79 }
  80
  81 void free_fdtable_rcu(struct rcu_head *rcu)
  82 {
  83         struct fdtable *fdt = container_of(rcu, struct fdtable, rcu);
  84         struct fdtable_defer *fddef;
  85
  86         BUG_ON(!fdt);
  87
  88         if (fdt->max_fds <= NR_OPEN_DEFAULT) {
  89                 /*
  90                  * This fdtable is embedded in the files structure and that
  91                  * structure itself is getting destroyed.
  92                  */
  93                 kmem_cache_free(files_cachep,
  94                                 container_of(fdt, struct files_struct, fdtab));
  95                 return;
  96         }
  97         if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *))) {
  98                 kfree(fdt->fd);
  99                 kfree(fdt->open_fds);
 100                 kfree(fdt);
 101         } else {
 102                 fddef = &get_cpu_var(fdtable_defer_list);
 103                 spin_lock(&fddef->lock);
 104                 fdt->next = fddef->next;
 105                 fddef->next = fdt;
 106                 /* vmallocs are handled from the workqueue context */
 107                 schedule_work(&fddef->wq);
 108                 spin_unlock(&fddef->lock);
 109                 put_cpu_var(fdtable_defer_list);
 110         }
 111 }
 112
 113 /*
 114  * Expand the fdset in the files_struct.  Called with the files spinlock
 115  * held for write.
 116  */
 117 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
 118 {
 119         unsigned int cpy, set;
 120
 121         BUG_ON(nfdt->max_fds < ofdt->max_fds);
 122
 123         cpy = ofdt->max_fds * sizeof(struct file *);
 124         set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
 125         memcpy(nfdt->fd, ofdt->fd, cpy);
 126         memset((char *)(nfdt->fd) + cpy, 0, set);
 127
 128         cpy = ofdt->max_fds / BITS_PER_BYTE;
 129         set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
 130         memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
 131         memset((char *)(nfdt->open_fds) + cpy, 0, set);
 132         memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
 133         memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
 134 }
 135
 136 static struct fdtable * alloc_fdtable(unsigned int nr)
 137 {
 138         struct fdtable *fdt;
 139         char *data;
 140
 141         /*
 142          * Figure out how many fds we actually want to support in this fdtable.
 143          * Allocation steps are keyed to the size of the fdarray, since it
 144          * grows far faster than any of the other dynamic data. We try to fit
 145          * the fdarray into comfortable page-tuned chunks: starting at 1024B
 146          * and growing in powers of two from there on.
 147          */
 148         nr /= (1024 / sizeof(struct file *));
 149         nr = roundup_pow_of_two(nr + 1);
 150         nr *= (1024 / sizeof(struct file *));
 151         /*
 152          * Note that this can drive nr *below* what we had passed if sysctl_nr_open
 153          * had been set lower between the check in expand_files() and here.  Deal
 154          * with that in caller, it's cheaper that way.
 155          *
 156          * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
 157          * bitmaps handling below becomes unpleasant, to put it mildly...
 158          */
 159         if (unlikely(nr > sysctl_nr_open))
 160                 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
 161
 162         fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
 163         if (!fdt)
 164                 goto out;
 165         fdt->max_fds = nr;
 166         data = alloc_fdmem(nr * sizeof(struct file *));
 167         if (!data)
 168                 goto out_fdt;
 169         fdt->fd = (struct file **)data;
 170         data = alloc_fdmem(max_t(unsigned int,
 171                                  2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
 172         if (!data)
 173                 goto out_arr;
 174         fdt->open_fds = (fd_set *)data;
 175         data += nr / BITS_PER_BYTE;
 176         fdt->close_on_exec = (fd_set *)data;
 177         INIT_RCU_HEAD(&fdt->rcu);
 178         fdt->next = NULL;
 179
 180         return fdt;
 181
 182 out_arr:
 183         free_fdarr(fdt);
 184 out_fdt:
 185         kfree(fdt);
 186 out:
 187         return NULL;
 188 }
 189
 190 /*
 191  * Expand the file descriptor table.
 192  * This function will allocate a new fdtable and both fd array and fdset, of
 193  * the given size.
 194  * Return <0 error code on error; 1 on successful completion.
 195  * The files->file_lock should be held on entry, and will be held on exit.
 196  */
 197 static int expand_fdtable(struct files_struct *files, int nr)
 198         __releases(files->file_lock)
 199         __acquires(files->file_lock)
 200 {
 201         struct fdtable *new_fdt, *cur_fdt;
 202
 203         spin_unlock(&files->file_lock);
 204         new_fdt = alloc_fdtable(nr);
 205         spin_lock(&files->file_lock);
 206         if (!new_fdt)
 207                 return -ENOMEM;
 208         /*
 209          * extremely unlikely race - sysctl_nr_open decreased between the check in
 210          * caller and alloc_fdtable().  Cheaper to catch it here...
 211          */
 212         if (unlikely(new_fdt->max_fds <= nr)) {
 213                 free_fdarr(new_fdt);
 214                 free_fdset(new_fdt);
 215                 kfree(new_fdt);
 216                 return -EMFILE;
 217         }
 218         /*
 219          * Check again since another task may have expanded the fd table while
 220          * we dropped the lock
 221          */
 222         cur_fdt = files_fdtable(files);
 223         if (nr >= cur_fdt->max_fds) {
 224                 /* Continue as planned */
 225                 copy_fdtable(new_fdt, cur_fdt);
 226                 rcu_assign_pointer(files->fdt, new_fdt);
 227                 if (cur_fdt->max_fds > NR_OPEN_DEFAULT)
 228                         free_fdtable(cur_fdt);
 229         } else {
 230                 /* Somebody else expanded, so undo our attempt */
 231                 free_fdarr(new_fdt);
 232                 free_fdset(new_fdt);
 233                 kfree(new_fdt);
 234         }
 235         return 1;
 236 }
 237
 238 /*
 239  * Expand files.
 240  * This function will expand the file structures, if the requested size exceeds
 241  * the current capacity and there is room for expansion.
 242  * Return <0 error code on error; 0 when nothing done; 1 when files were
 243  * expanded and execution may have blocked.
 244  * The files->file_lock should be held on entry, and will be held on exit.
 245  */
 246 int expand_files(struct files_struct *files, int nr)
 247 {
 248         struct fdtable *fdt;
 249
 250         fdt = files_fdtable(files);
 251         /* Do we need to expand? */
 252         if (nr < fdt->max_fds)
 253                 return 0;
 254         /* Can we expand? */
 255         if (nr >= sysctl_nr_open)
 256                 return -EMFILE;
 257
 258         /* All good, so we try */
 259         return expand_fdtable(files, nr);
 260 }
 261
 262 static int count_open_files(struct fdtable *fdt)
 263 {
 264         int size = fdt->max_fds;
 265         int i;
 266
 267         /* Find the last open fd */
 268         for (i = size/(8*sizeof(long)); i > 0; ) {
 269                 if (fdt->open_fds->fds_bits[--i])
 270                         break;
 271         }
 272         i = (i+1) * 8 * sizeof(long);
 273         return i;
 274 }
 275
 276 static struct files_struct *alloc_files(void)
 277 {
 278         struct files_struct *newf;
 279         struct fdtable *fdt;
 280
 281         newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
 282         if (!newf)
 283                 goto out;
 284
 285         atomic_set(&newf->count, 1);
 286
 287         spin_lock_init(&newf->file_lock);
 288         newf->next_fd = 0;
 289         fdt = &newf->fdtab;
 290         fdt->max_fds = NR_OPEN_DEFAULT;
 291         fdt->close_on_exec = (fd_set *)&newf->close_on_exec_init;
 292         fdt->open_fds = (fd_set *)&newf->open_fds_init;
 293         fdt->fd = &newf->fd_array[0];
 294         INIT_RCU_HEAD(&fdt->rcu);
 295         fdt->next = NULL;
 296         rcu_assign_pointer(newf->fdt, fdt);
 297 out:
 298         return newf;
 299 }
 300
 301 /*
 302  * Allocate a new files structure and copy contents from the
 303  * passed in files structure.
 304  * errorp will be valid only when the returned files_struct is NULL.
 305  */
 306 struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
 307 {
 308         struct files_struct *newf;
 309         struct file **old_fds, **new_fds;
 310         int open_files, size, i;
 311         struct fdtable *old_fdt, *new_fdt;
 312
 313         *errorp = -ENOMEM;
 314         newf = alloc_files();
 315         if (!newf)
 316                 goto out;
 317
 318         spin_lock(&oldf->file_lock);
 319         old_fdt = files_fdtable(oldf);
 320         new_fdt = files_fdtable(newf);
 321         open_files = count_open_files(old_fdt);
 322
 323         /*
 324          * Check whether we need to allocate a larger fd array and fd set.
 325          * Note: we're not a clone task, so the open count won't change.
 326          */
 327         if (open_files > new_fdt->max_fds) {
 328                 spin_unlock(&oldf->file_lock);
 329
 330                 new_fdt = alloc_fdtable(open_files - 1);
 331                 if (!new_fdt) {
 332                         *errorp = -ENOMEM;
 333                         goto out_release;
 334                 }
 335
 336                 /* beyond sysctl_nr_open; nothing to do */
 337                 if (unlikely(new_fdt->max_fds < open_files)) {
 338                         free_fdarr(new_fdt);
 339                         free_fdset(new_fdt);
 340                         kfree(new_fdt);
 341                         *errorp = -EMFILE;
 342                         goto out_release;
 343                 }
 344                 rcu_assign_pointer(files->fdt, new_fdt);
 345
 346                 /*
 347                  * Reacquire the oldf lock and a pointer to its fd table
 348                  * who knows it may have a new bigger fd table. We need
 349                  * the latest pointer.
 350                  */
 351                 spin_lock(&oldf->file_lock);
 352                 old_fdt = files_fdtable(oldf);
 353         }
 354
 355         old_fds = old_fdt->fd;
 356         new_fds = new_fdt->fd;
 357
 358         memcpy(new_fdt->open_fds->fds_bits,
 359                 old_fdt->open_fds->fds_bits, open_files/8);
 360         memcpy(new_fdt->close_on_exec->fds_bits,
 361                 old_fdt->close_on_exec->fds_bits, open_files/8);
 362
 363         for (i = open_files; i != 0; i--) {
 364                 struct file *f = *old_fds++;
 365                 if (f) {
 366                         get_file(f);
 367                 } else {
 368                         /*
 369                          * The fd may be claimed in the fd bitmap but not yet
 370                          * instantiated in the files array if a sibling thread
 371                          * is partway through open().  So make sure that this
 372                          * fd is available to the new process.
 373                          */
 374                         FD_CLR(open_files - i, new_fdt->open_fds);
 375                 }
 376                 rcu_assign_pointer(*new_fds++, f);
 377         }
 378         spin_unlock(&oldf->file_lock);
 379
 380         /* compute the remainder to be cleared */
 381         size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
 382
 383         /* This is long word aligned thus could use a optimized version */
 384         memset(new_fds, 0, size);
 385
 386         if (new_fdt->max_fds > open_files) {
 387                 int left = (new_fdt->max_fds-open_files)/8;
 388                 int start = open_files / (8 * sizeof(unsigned long));
 389
 390                 memset(&new_fdt->open_fds->fds_bits[start], 0, left);
 391                 memset(&new_fdt->close_on_exec->fds_bits[start], 0, left);
 392         }
 393
 394         return newf;
 395
 396 out_release:
 397         kmem_cache_free(files_cachep, newf);
 398 out:
 399         return NULL;
 400 }
 401
 402 static void __devinit fdtable_defer_list_init(int cpu)
 403 {
 404         struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
 405         spin_lock_init(&fddef->lock);
 406         INIT_WORK(&fddef->wq, free_fdtable_work);
 407         fddef->next = NULL;
 408 }
 409
 410 void __init files_defer_init(void)
 411 {
 412         int i;
 413         for_each_possible_cpu(i)
 414                 fdtable_defer_list_init(i);
 415 }
 416
 417 struct files_struct init_files = {
 418         .count          = ATOMIC_INIT(1),
 419         .fdt            = &init_files.fdtab,
 420         .fdtab          = {
 421                 .max_fds        = NR_OPEN_DEFAULT,
 422                 .fd             = &init_files.fd_array[0],
 423                 .close_on_exec  = (fd_set *)&init_files.close_on_exec_init,
 424                 .open_fds       = (fd_set *)&init_files.open_fds_init,
 425                 .rcu            = RCU_HEAD_INIT,
 426         },
 427         .file_lock      = __SPIN_LOCK_UNLOCKED(init_task.file_lock),
 428 };