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