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/module.h>
  10 #include <linux/fs.h>
  11 #include <linux/mm.h>
  12 #include <linux/time.h>
  13 #include <linux/slab.h>
  14 #include <linux/vmalloc.h>
  15 #include <linux/file.h>
  16 #include <linux/fdtable.h>
  17 #include <linux/bitops.h>
  18 #include <linux/interrupt.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/rcupdate.h>
  21 #include <linux/workqueue.h>
  22
  23 struct fdtable_defer {
  24         spinlock_t lock;
  25         struct work_struct wq;
  26         struct fdtable *next;
  27 };
  28
  29 int sysctl_nr_open __read_mostly = 1024*1024;
  30 int sysctl_nr_open_min = BITS_PER_LONG;
  31 int sysctl_nr_open_max = 1024 * 1024; /* raised later */
  32
  33 /*
  34  * We use this list to defer free fdtables that have vmalloced
  35  * sets/arrays. By keeping a per-cpu list, we avoid having to embed
  36  * the work_struct in fdtable itself which avoids a 64 byte (i386) increase in
  37  * this per-task structure.
  38  */
  39 static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
  40
  41 static inline void * alloc_fdmem(unsigned int size)
  42 {
  43         if (size <= PAGE_SIZE)
  44                 return kmalloc(size, GFP_KERNEL);
  45         else
  46                 return vmalloc(size);
  47 }
  48
  49 static inline void free_fdarr(struct fdtable *fdt)
  50 {
  51         if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *)))
  52                 kfree(fdt->fd);
  53         else
  54                 vfree(fdt->fd);
  55 }
  56
  57 static inline void free_fdset(struct fdtable *fdt)
  58 {
  59         if (fdt->max_fds <= (PAGE_SIZE * BITS_PER_BYTE / 2))
  60                 kfree(fdt->open_fds);
  61         else
  62                 vfree(fdt->open_fds);
  63 }
  64
  65 static void free_fdtable_work(struct work_struct *work)
  66 {
  67         struct fdtable_defer *f =
  68                 container_of(work, struct fdtable_defer, wq);
  69         struct fdtable *fdt;
  70
  71         spin_lock_bh(&f->lock);
  72         fdt = f->next;
  73         f->next = NULL;
  74         spin_unlock_bh(&f->lock);
  75         while(fdt) {
  76                 struct fdtable *next = fdt->next;
  77                 vfree(fdt->fd);
  78                 free_fdset(fdt);
  79                 kfree(fdt);
  80                 fdt = next;
  81         }
  82 }
  83
  84 void free_fdtable_rcu(struct rcu_head *rcu)
  85 {
  86         struct fdtable *fdt = container_of(rcu, struct fdtable, rcu);
  87         struct fdtable_defer *fddef;
  88
  89         BUG_ON(!fdt);
  90
  91         if (fdt->max_fds <= NR_OPEN_DEFAULT) {
  92                 /*
  93                  * This fdtable is embedded in the files structure and that
  94                  * structure itself is getting destroyed.
  95                  */
  96                 kmem_cache_free(files_cachep,
  97                                 container_of(fdt, struct files_struct, fdtab));
  98                 return;
  99         }
 100         if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *))) {
 101                 kfree(fdt->fd);
 102                 kfree(fdt->open_fds);
 103                 kfree(fdt);
 104         } else {
 105                 fddef = &get_cpu_var(fdtable_defer_list);
 106                 spin_lock(&fddef->lock);
 107                 fdt->next = fddef->next;
 108                 fddef->next = fdt;
 109                 /* vmallocs are handled from the workqueue context */
 110                 schedule_work(&fddef->wq);
 111                 spin_unlock(&fddef->lock);
 112                 put_cpu_var(fdtable_defer_list);
 113         }
 114 }
 115
 116 /*
 117  * Expand the fdset in the files_struct.  Called with the files spinlock
 118  * held for write.
 119  */
 120 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
 121 {
 122         unsigned int cpy, set;
 123
 124         BUG_ON(nfdt->max_fds < ofdt->max_fds);
 125
 126         cpy = ofdt->max_fds * sizeof(struct file *);
 127         set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
 128         memcpy(nfdt->fd, ofdt->fd, cpy);
 129         memset((char *)(nfdt->fd) + cpy, 0, set);
 130
 131         cpy = ofdt->max_fds / BITS_PER_BYTE;
 132         set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
 133         memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
 134         memset((char *)(nfdt->open_fds) + cpy, 0, set);
 135         memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
 136         memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
 137 }
 138
 139 static struct fdtable * alloc_fdtable(unsigned int nr)
 140 {
 141         struct fdtable *fdt;
 142         char *data;
 143
 144         /*
 145          * Figure out how many fds we actually want to support in this fdtable.
 146          * Allocation steps are keyed to the size of the fdarray, since it
 147          * grows far faster than any of the other dynamic data. We try to fit
 148          * the fdarray into comfortable page-tuned chunks: starting at 1024B
 149          * and growing in powers of two from there on.
 150          */
 151         nr /= (1024 / sizeof(struct file *));
 152         nr = roundup_pow_of_two(nr + 1);
 153         nr *= (1024 / sizeof(struct file *));
 154         /*
 155          * Note that this can drive nr *below* what we had passed if sysctl_nr_open
 156          * had been set lower between the check in expand_files() and here.  Deal
 157          * with that in caller, it's cheaper that way.
 158          *
 159          * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
 160          * bitmaps handling below becomes unpleasant, to put it mildly...
 161          */
 162         if (unlikely(nr > sysctl_nr_open))
 163                 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
 164
 165         fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
 166         if (!fdt)
 167                 goto out;
 168         fdt->max_fds = nr;
 169         data = alloc_fdmem(nr * sizeof(struct file *));
 170         if (!data)
 171                 goto out_fdt;
 172         fdt->fd = (struct file **)data;
 173         data = alloc_fdmem(max_t(unsigned int,
 174                                  2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
 175         if (!data)
 176                 goto out_arr;
 177         fdt->open_fds = (fd_set *)data;
 178         data += nr / BITS_PER_BYTE;
 179         fdt->close_on_exec = (fd_set *)data;
 180         INIT_RCU_HEAD(&fdt->rcu);
 181         fdt->next = NULL;
 182
 183         return fdt;
 184
 185 out_arr:
 186         free_fdarr(fdt);
 187 out_fdt:
 188         kfree(fdt);
 189 out:
 190         return NULL;
 191 }
 192
 193 /*
 194  * Expand the file descriptor table.
 195  * This function will allocate a new fdtable and both fd array and fdset, of
 196  * the given size.
 197  * Return <0 error code on error; 1 on successful completion.
 198  * The files->file_lock should be held on entry, and will be held on exit.
 199  */
 200 static int expand_fdtable(struct files_struct *files, int nr)
 201         __releases(files->file_lock)
 202         __acquires(files->file_lock)
 203 {
 204         struct fdtable *new_fdt, *cur_fdt;
 205
 206         spin_unlock(&files->file_lock);
 207         new_fdt = alloc_fdtable(nr);
 208         spin_lock(&files->file_lock);
 209         if (!new_fdt)
 210                 return -ENOMEM;
 211         /*
 212          * extremely unlikely race - sysctl_nr_open decreased between the check in
 213          * caller and alloc_fdtable().  Cheaper to catch it here...
 214          */
 215         if (unlikely(new_fdt->max_fds <= nr)) {
 216                 free_fdarr(new_fdt);
 217                 free_fdset(new_fdt);
 218                 kfree(new_fdt);
 219                 return -EMFILE;
 220         }
 221         /*
 222          * Check again since another task may have expanded the fd table while
 223          * we dropped the lock
 224          */
 225         cur_fdt = files_fdtable(files);
 226         if (nr >= cur_fdt->max_fds) {
 227                 /* Continue as planned */
 228                 copy_fdtable(new_fdt, cur_fdt);
 229                 rcu_assign_pointer(files->fdt, new_fdt);
 230                 if (cur_fdt->max_fds > NR_OPEN_DEFAULT)
 231                         free_fdtable(cur_fdt);
 232         } else {
 233                 /* Somebody else expanded, so undo our attempt */
 234                 free_fdarr(new_fdt);
 235                 free_fdset(new_fdt);
 236                 kfree(new_fdt);
 237         }
 238         return 1;
 239 }
 240
 241 /*
 242  * Expand files.
 243  * This function will expand the file structures, if the requested size exceeds
 244  * the current capacity and there is room for expansion.
 245  * Return <0 error code on error; 0 when nothing done; 1 when files were
 246  * expanded and execution may have blocked.
 247  * The files->file_lock should be held on entry, and will be held on exit.
 248  */
 249 int expand_files(struct files_struct *files, int nr)
 250 {
 251         struct fdtable *fdt;
 252
 253         fdt = files_fdtable(files);
 254
 255         /*
 256          * N.B. For clone tasks sharing a files structure, this test
 257          * will limit the total number of files that can be opened.
 258          */
 259         if (nr >= current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
 260                 return -EMFILE;
 261
 262         /* Do we need to expand? */
 263         if (nr < fdt->max_fds)
 264                 return 0;
 265
 266         /* Can we expand? */
 267         if (nr >= sysctl_nr_open)
 268                 return -EMFILE;
 269
 270         /* All good, so we try */
 271         return expand_fdtable(files, nr);
 272 }
 273
 274 static int count_open_files(struct fdtable *fdt)
 275 {
 276         int size = fdt->max_fds;
 277         int i;
 278
 279         /* Find the last open fd */
 280         for (i = size/(8*sizeof(long)); i > 0; ) {
 281                 if (fdt->open_fds->fds_bits[--i])
 282                         break;
 283         }
 284         i = (i+1) * 8 * sizeof(long);
 285         return i;
 286 }
 287
 288 /*
 289  * Allocate a new files structure and copy contents from the
 290  * passed in files structure.
 291  * errorp will be valid only when the returned files_struct is NULL.
 292  */
 293 struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
 294 {
 295         struct files_struct *newf;
 296         struct file **old_fds, **new_fds;
 297         int open_files, size, i;
 298         struct fdtable *old_fdt, *new_fdt;
 299
 300         *errorp = -ENOMEM;
 301         newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
 302         if (!newf)
 303                 goto out;
 304
 305         atomic_set(&newf->count, 1);
 306
 307         spin_lock_init(&newf->file_lock);
 308         newf->next_fd = 0;
 309         new_fdt = &newf->fdtab;
 310         new_fdt->max_fds = NR_OPEN_DEFAULT;
 311         new_fdt->close_on_exec = (fd_set *)&newf->close_on_exec_init;
 312         new_fdt->open_fds = (fd_set *)&newf->open_fds_init;
 313         new_fdt->fd = &newf->fd_array[0];
 314         INIT_RCU_HEAD(&new_fdt->rcu);
 315         new_fdt->next = NULL;
 316
 317         spin_lock(&oldf->file_lock);
 318         old_fdt = files_fdtable(oldf);
 319         open_files = count_open_files(old_fdt);
 320
 321         /*
 322          * Check whether we need to allocate a larger fd array and fd set.
 323          */
 324         while (unlikely(open_files > new_fdt->max_fds)) {
 325                 spin_unlock(&oldf->file_lock);
 326
 327                 if (new_fdt != &newf->fdtab) {
 328                         free_fdarr(new_fdt);
 329                         free_fdset(new_fdt);
 330                         kfree(new_fdt);
 331                 }
 332
 333                 new_fdt = alloc_fdtable(open_files - 1);
 334                 if (!new_fdt) {
 335                         *errorp = -ENOMEM;
 336                         goto out_release;
 337                 }
 338
 339                 /* beyond sysctl_nr_open; nothing to do */
 340                 if (unlikely(new_fdt->max_fds < open_files)) {
 341                         free_fdarr(new_fdt);
 342                         free_fdset(new_fdt);
 343                         kfree(new_fdt);
 344                         *errorp = -EMFILE;
 345                         goto out_release;
 346                 }
 347
 348                 /*
 349                  * Reacquire the oldf lock and a pointer to its fd table
 350                  * who knows it may have a new bigger fd table. We need
 351                  * the latest pointer.
 352                  */
 353                 spin_lock(&oldf->file_lock);
 354                 old_fdt = files_fdtable(oldf);
 355                 open_files = count_open_files(old_fdt);
 356         }
 357
 358         old_fds = old_fdt->fd;
 359         new_fds = new_fdt->fd;
 360
 361         memcpy(new_fdt->open_fds->fds_bits,
 362                 old_fdt->open_fds->fds_bits, open_files/8);
 363         memcpy(new_fdt->close_on_exec->fds_bits,
 364                 old_fdt->close_on_exec->fds_bits, open_files/8);
 365
 366         for (i = open_files; i != 0; i--) {
 367                 struct file *f = *old_fds++;
 368                 if (f) {
 369                         get_file(f);
 370                 } else {
 371                         /*
 372                          * The fd may be claimed in the fd bitmap but not yet
 373                          * instantiated in the files array if a sibling thread
 374                          * is partway through open().  So make sure that this
 375                          * fd is available to the new process.
 376                          */
 377                         FD_CLR(open_files - i, new_fdt->open_fds);
 378                 }
 379                 rcu_assign_pointer(*new_fds++, f);
 380         }
 381         spin_unlock(&oldf->file_lock);
 382
 383         /* compute the remainder to be cleared */
 384         size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
 385
 386         /* This is long word aligned thus could use a optimized version */
 387         memset(new_fds, 0, size);
 388
 389         if (new_fdt->max_fds > open_files) {
 390                 int left = (new_fdt->max_fds-open_files)/8;
 391                 int start = open_files / (8 * sizeof(unsigned long));
 392
 393                 memset(&new_fdt->open_fds->fds_bits[start], 0, left);
 394                 memset(&new_fdt->close_on_exec->fds_bits[start], 0, left);
 395         }
 396
 397         rcu_assign_pointer(newf->fdt, new_fdt);
 398
 399         return newf;
 400
 401 out_release:
 402         kmem_cache_free(files_cachep, newf);
 403 out:
 404         return NULL;
 405 }
 406
 407 static void __devinit fdtable_defer_list_init(int cpu)
 408 {
 409         struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
 410         spin_lock_init(&fddef->lock);
 411         INIT_WORK(&fddef->wq, free_fdtable_work);
 412         fddef->next = NULL;
 413 }
 414
 415 void __init files_defer_init(void)
 416 {
 417         int i;
 418         for_each_possible_cpu(i)
 419                 fdtable_defer_list_init(i);
 420         sysctl_nr_open_max = min((size_t)INT_MAX, ~(size_t)0/sizeof(void *)) &
 421                              -BITS_PER_LONG;
 422 }
 423
 424 struct files_struct init_files = {
 425         .count          = ATOMIC_INIT(1),
 426         .fdt            = &init_files.fdtab,
 427         .fdtab          = {
 428                 .max_fds        = NR_OPEN_DEFAULT,
 429                 .fd             = &init_files.fd_array[0],
 430                 .close_on_exec  = (fd_set *)&init_files.close_on_exec_init,
 431                 .open_fds       = (fd_set *)&init_files.open_fds_init,
 432                 .rcu            = RCU_HEAD_INIT,
 433         },
 434         .file_lock      = __SPIN_LOCK_UNLOCKED(init_task.file_lock),
 435 };
 436
 437 /*
 438  * allocate a file descriptor, mark it busy.
 439  */
 440 int alloc_fd(unsigned start, unsigned flags)
 441 {
 442         struct files_struct *files = current->files;
 443         unsigned int fd;
 444         int error;
 445         struct fdtable *fdt;
 446
 447         spin_lock(&files->file_lock);
 448 repeat:
 449         fdt = files_fdtable(files);
 450         fd = start;
 451         if (fd < files->next_fd)
 452                 fd = files->next_fd;
 453
 454         if (fd < fdt->max_fds)
 455                 fd = find_next_zero_bit(fdt->open_fds->fds_bits,
 456                                            fdt->max_fds, fd);
 457
 458         error = expand_files(files, fd);
 459         if (error < 0)
 460                 goto out;
 461
 462         /*
 463          * If we needed to expand the fs array we
 464          * might have blocked - try again.
 465          */
 466         if (error)
 467                 goto repeat;
 468
 469         if (start <= files->next_fd)
 470                 files->next_fd = fd + 1;
 471
 472         FD_SET(fd, fdt->open_fds);
 473         if (flags & O_CLOEXEC)
 474                 FD_SET(fd, fdt->close_on_exec);
 475         else
 476                 FD_CLR(fd, fdt->close_on_exec);
 477         error = fd;
 478 #if 1
 479         /* Sanity check */
 480         if (rcu_dereference(fdt->fd[fd]) != NULL) {
 481                 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
 482                 rcu_assign_pointer(fdt->fd[fd], NULL);
 483         }
 484 #endif
 485
 486 out:
 487         spin_unlock(&files->file_lock);
 488         return error;
 489 }
 490
 491 int get_unused_fd(void)
 492 {
 493         return alloc_fd(0, 0);
 494 }
 495 EXPORT_SYMBOL(get_unused_fd);