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