]>
Commit | Line | Data |
---|---|---|
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/syscalls.h> | |
10 | #include <linux/export.h> | |
11 | #include <linux/fs.h> | |
12 | #include <linux/mm.h> | |
13 | #include <linux/mmzone.h> | |
14 | #include <linux/time.h> | |
15 | #include <linux/sched.h> | |
16 | #include <linux/slab.h> | |
17 | #include <linux/vmalloc.h> | |
18 | #include <linux/file.h> | |
19 | #include <linux/fdtable.h> | |
20 | #include <linux/bitops.h> | |
21 | #include <linux/interrupt.h> | |
22 | #include <linux/spinlock.h> | |
23 | #include <linux/rcupdate.h> | |
24 | #include <linux/workqueue.h> | |
25 | ||
26 | struct fdtable_defer { | |
27 | spinlock_t lock; | |
28 | struct work_struct wq; | |
29 | struct fdtable *next; | |
30 | }; | |
31 | ||
32 | int sysctl_nr_open __read_mostly = 1024*1024; | |
33 | int sysctl_nr_open_min = BITS_PER_LONG; | |
34 | int sysctl_nr_open_max = 1024 * 1024; /* raised later */ | |
35 | ||
36 | /* | |
37 | * We use this list to defer free fdtables that have vmalloced | |
38 | * sets/arrays. By keeping a per-cpu list, we avoid having to embed | |
39 | * the work_struct in fdtable itself which avoids a 64 byte (i386) increase in | |
40 | * this per-task structure. | |
41 | */ | |
42 | static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list); | |
43 | ||
44 | static void *alloc_fdmem(size_t size) | |
45 | { | |
46 | /* | |
47 | * Very large allocations can stress page reclaim, so fall back to | |
48 | * vmalloc() if the allocation size will be considered "large" by the VM. | |
49 | */ | |
50 | if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) { | |
51 | void *data = kmalloc(size, GFP_KERNEL|__GFP_NOWARN); | |
52 | if (data != NULL) | |
53 | return data; | |
54 | } | |
55 | return vmalloc(size); | |
56 | } | |
57 | ||
58 | static void free_fdmem(void *ptr) | |
59 | { | |
60 | is_vmalloc_addr(ptr) ? vfree(ptr) : kfree(ptr); | |
61 | } | |
62 | ||
63 | static void __free_fdtable(struct fdtable *fdt) | |
64 | { | |
65 | free_fdmem(fdt->fd); | |
66 | free_fdmem(fdt->open_fds); | |
67 | kfree(fdt); | |
68 | } | |
69 | ||
70 | static void free_fdtable_work(struct work_struct *work) | |
71 | { | |
72 | struct fdtable_defer *f = | |
73 | container_of(work, struct fdtable_defer, wq); | |
74 | struct fdtable *fdt; | |
75 | ||
76 | spin_lock_bh(&f->lock); | |
77 | fdt = f->next; | |
78 | f->next = NULL; | |
79 | spin_unlock_bh(&f->lock); | |
80 | while(fdt) { | |
81 | struct fdtable *next = fdt->next; | |
82 | ||
83 | __free_fdtable(fdt); | |
84 | fdt = next; | |
85 | } | |
86 | } | |
87 | ||
88 | static void free_fdtable_rcu(struct rcu_head *rcu) | |
89 | { | |
90 | struct fdtable *fdt = container_of(rcu, struct fdtable, rcu); | |
91 | struct fdtable_defer *fddef; | |
92 | ||
93 | BUG_ON(!fdt); | |
94 | BUG_ON(fdt->max_fds <= NR_OPEN_DEFAULT); | |
95 | ||
96 | if (!is_vmalloc_addr(fdt->fd) && !is_vmalloc_addr(fdt->open_fds)) { | |
97 | kfree(fdt->fd); | |
98 | kfree(fdt->open_fds); | |
99 | kfree(fdt); | |
100 | } else { | |
101 | fddef = &get_cpu_var(fdtable_defer_list); | |
102 | spin_lock(&fddef->lock); | |
103 | fdt->next = fddef->next; | |
104 | fddef->next = fdt; | |
105 | /* vmallocs are handled from the workqueue context */ | |
106 | schedule_work(&fddef->wq); | |
107 | spin_unlock(&fddef->lock); | |
108 | put_cpu_var(fdtable_defer_list); | |
109 | } | |
110 | } | |
111 | ||
112 | /* | |
113 | * Expand the fdset in the files_struct. Called with the files spinlock | |
114 | * held for write. | |
115 | */ | |
116 | static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt) | |
117 | { | |
118 | unsigned int cpy, set; | |
119 | ||
120 | BUG_ON(nfdt->max_fds < ofdt->max_fds); | |
121 | ||
122 | cpy = ofdt->max_fds * sizeof(struct file *); | |
123 | set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *); | |
124 | memcpy(nfdt->fd, ofdt->fd, cpy); | |
125 | memset((char *)(nfdt->fd) + cpy, 0, set); | |
126 | ||
127 | cpy = ofdt->max_fds / BITS_PER_BYTE; | |
128 | set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE; | |
129 | memcpy(nfdt->open_fds, ofdt->open_fds, cpy); | |
130 | memset((char *)(nfdt->open_fds) + cpy, 0, set); | |
131 | memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy); | |
132 | memset((char *)(nfdt->close_on_exec) + cpy, 0, set); | |
133 | } | |
134 | ||
135 | static struct fdtable * alloc_fdtable(unsigned int nr) | |
136 | { | |
137 | struct fdtable *fdt; | |
138 | void *data; | |
139 | ||
140 | /* | |
141 | * Figure out how many fds we actually want to support in this fdtable. | |
142 | * Allocation steps are keyed to the size of the fdarray, since it | |
143 | * grows far faster than any of the other dynamic data. We try to fit | |
144 | * the fdarray into comfortable page-tuned chunks: starting at 1024B | |
145 | * and growing in powers of two from there on. | |
146 | */ | |
147 | nr /= (1024 / sizeof(struct file *)); | |
148 | nr = roundup_pow_of_two(nr + 1); | |
149 | nr *= (1024 / sizeof(struct file *)); | |
150 | /* | |
151 | * Note that this can drive nr *below* what we had passed if sysctl_nr_open | |
152 | * had been set lower between the check in expand_files() and here. Deal | |
153 | * with that in caller, it's cheaper that way. | |
154 | * | |
155 | * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise | |
156 | * bitmaps handling below becomes unpleasant, to put it mildly... | |
157 | */ | |
158 | if (unlikely(nr > sysctl_nr_open)) | |
159 | nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1; | |
160 | ||
161 | fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL); | |
162 | if (!fdt) | |
163 | goto out; | |
164 | fdt->max_fds = nr; | |
165 | data = alloc_fdmem(nr * sizeof(struct file *)); | |
166 | if (!data) | |
167 | goto out_fdt; | |
168 | fdt->fd = data; | |
169 | ||
170 | data = alloc_fdmem(max_t(size_t, | |
171 | 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES)); | |
172 | if (!data) | |
173 | goto out_arr; | |
174 | fdt->open_fds = data; | |
175 | data += nr / BITS_PER_BYTE; | |
176 | fdt->close_on_exec = data; | |
177 | fdt->next = NULL; | |
178 | ||
179 | return fdt; | |
180 | ||
181 | out_arr: | |
182 | free_fdmem(fdt->fd); | |
183 | out_fdt: | |
184 | kfree(fdt); | |
185 | out: | |
186 | return NULL; | |
187 | } | |
188 | ||
189 | /* | |
190 | * Expand the file descriptor table. | |
191 | * This function will allocate a new fdtable and both fd array and fdset, of | |
192 | * the given size. | |
193 | * Return <0 error code on error; 1 on successful completion. | |
194 | * The files->file_lock should be held on entry, and will be held on exit. | |
195 | */ | |
196 | static int expand_fdtable(struct files_struct *files, int nr) | |
197 | __releases(files->file_lock) | |
198 | __acquires(files->file_lock) | |
199 | { | |
200 | struct fdtable *new_fdt, *cur_fdt; | |
201 | ||
202 | spin_unlock(&files->file_lock); | |
203 | new_fdt = alloc_fdtable(nr); | |
204 | spin_lock(&files->file_lock); | |
205 | if (!new_fdt) | |
206 | return -ENOMEM; | |
207 | /* | |
208 | * extremely unlikely race - sysctl_nr_open decreased between the check in | |
209 | * caller and alloc_fdtable(). Cheaper to catch it here... | |
210 | */ | |
211 | if (unlikely(new_fdt->max_fds <= nr)) { | |
212 | __free_fdtable(new_fdt); | |
213 | return -EMFILE; | |
214 | } | |
215 | /* | |
216 | * Check again since another task may have expanded the fd table while | |
217 | * we dropped the lock | |
218 | */ | |
219 | cur_fdt = files_fdtable(files); | |
220 | if (nr >= cur_fdt->max_fds) { | |
221 | /* Continue as planned */ | |
222 | copy_fdtable(new_fdt, cur_fdt); | |
223 | rcu_assign_pointer(files->fdt, new_fdt); | |
224 | if (cur_fdt->max_fds > NR_OPEN_DEFAULT) | |
225 | call_rcu(&cur_fdt->rcu, free_fdtable_rcu); | |
226 | } else { | |
227 | /* Somebody else expanded, so undo our attempt */ | |
228 | __free_fdtable(new_fdt); | |
229 | } | |
230 | return 1; | |
231 | } | |
232 | ||
233 | /* | |
234 | * Expand files. | |
235 | * This function will expand the file structures, if the requested size exceeds | |
236 | * the current capacity and there is room for expansion. | |
237 | * Return <0 error code on error; 0 when nothing done; 1 when files were | |
238 | * expanded and execution may have blocked. | |
239 | * The files->file_lock should be held on entry, and will be held on exit. | |
240 | */ | |
241 | static int expand_files(struct files_struct *files, int nr) | |
242 | { | |
243 | struct fdtable *fdt; | |
244 | ||
245 | fdt = files_fdtable(files); | |
246 | ||
247 | /* Do we need to expand? */ | |
248 | if (nr < fdt->max_fds) | |
249 | return 0; | |
250 | ||
251 | /* Can we expand? */ | |
252 | if (nr >= sysctl_nr_open) | |
253 | return -EMFILE; | |
254 | ||
255 | /* All good, so we try */ | |
256 | return expand_fdtable(files, nr); | |
257 | } | |
258 | ||
259 | static inline void __set_close_on_exec(int fd, struct fdtable *fdt) | |
260 | { | |
261 | __set_bit(fd, fdt->close_on_exec); | |
262 | } | |
263 | ||
264 | static inline void __clear_close_on_exec(int fd, struct fdtable *fdt) | |
265 | { | |
266 | __clear_bit(fd, fdt->close_on_exec); | |
267 | } | |
268 | ||
269 | static inline void __set_open_fd(int fd, struct fdtable *fdt) | |
270 | { | |
271 | __set_bit(fd, fdt->open_fds); | |
272 | } | |
273 | ||
274 | static inline void __clear_open_fd(int fd, struct fdtable *fdt) | |
275 | { | |
276 | __clear_bit(fd, fdt->open_fds); | |
277 | } | |
278 | ||
279 | static int count_open_files(struct fdtable *fdt) | |
280 | { | |
281 | int size = fdt->max_fds; | |
282 | int i; | |
283 | ||
284 | /* Find the last open fd */ | |
285 | for (i = size / BITS_PER_LONG; i > 0; ) { | |
286 | if (fdt->open_fds[--i]) | |
287 | break; | |
288 | } | |
289 | i = (i + 1) * BITS_PER_LONG; | |
290 | return i; | |
291 | } | |
292 | ||
293 | /* | |
294 | * Allocate a new files structure and copy contents from the | |
295 | * passed in files structure. | |
296 | * errorp will be valid only when the returned files_struct is NULL. | |
297 | */ | |
298 | struct files_struct *dup_fd(struct files_struct *oldf, int *errorp) | |
299 | { | |
300 | struct files_struct *newf; | |
301 | struct file **old_fds, **new_fds; | |
302 | int open_files, size, i; | |
303 | struct fdtable *old_fdt, *new_fdt; | |
304 | ||
305 | *errorp = -ENOMEM; | |
306 | newf = kmem_cache_alloc(files_cachep, GFP_KERNEL); | |
307 | if (!newf) | |
308 | goto out; | |
309 | ||
310 | atomic_set(&newf->count, 1); | |
311 | ||
312 | spin_lock_init(&newf->file_lock); | |
313 | newf->next_fd = 0; | |
314 | new_fdt = &newf->fdtab; | |
315 | new_fdt->max_fds = NR_OPEN_DEFAULT; | |
316 | new_fdt->close_on_exec = newf->close_on_exec_init; | |
317 | new_fdt->open_fds = newf->open_fds_init; | |
318 | new_fdt->fd = &newf->fd_array[0]; | |
319 | new_fdt->next = NULL; | |
320 | ||
321 | spin_lock(&oldf->file_lock); | |
322 | old_fdt = files_fdtable(oldf); | |
323 | open_files = count_open_files(old_fdt); | |
324 | ||
325 | /* | |
326 | * Check whether we need to allocate a larger fd array and fd set. | |
327 | */ | |
328 | while (unlikely(open_files > new_fdt->max_fds)) { | |
329 | spin_unlock(&oldf->file_lock); | |
330 | ||
331 | if (new_fdt != &newf->fdtab) | |
332 | __free_fdtable(new_fdt); | |
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_fdtable(new_fdt); | |
343 | *errorp = -EMFILE; | |
344 | goto out_release; | |
345 | } | |
346 | ||
347 | /* | |
348 | * Reacquire the oldf lock and a pointer to its fd table | |
349 | * who knows it may have a new bigger fd table. We need | |
350 | * the latest pointer. | |
351 | */ | |
352 | spin_lock(&oldf->file_lock); | |
353 | old_fdt = files_fdtable(oldf); | |
354 | open_files = count_open_files(old_fdt); | |
355 | } | |
356 | ||
357 | old_fds = old_fdt->fd; | |
358 | new_fds = new_fdt->fd; | |
359 | ||
360 | memcpy(new_fdt->open_fds, old_fdt->open_fds, open_files / 8); | |
361 | memcpy(new_fdt->close_on_exec, old_fdt->close_on_exec, 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 | __clear_open_fd(open_files - i, new_fdt); | |
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 / BITS_PER_LONG; | |
389 | ||
390 | memset(&new_fdt->open_fds[start], 0, left); | |
391 | memset(&new_fdt->close_on_exec[start], 0, left); | |
392 | } | |
393 | ||
394 | rcu_assign_pointer(newf->fdt, new_fdt); | |
395 | ||
396 | return newf; | |
397 | ||
398 | out_release: | |
399 | kmem_cache_free(files_cachep, newf); | |
400 | out: | |
401 | return NULL; | |
402 | } | |
403 | ||
404 | static void close_files(struct files_struct * files) | |
405 | { | |
406 | int i, j; | |
407 | struct fdtable *fdt; | |
408 | ||
409 | j = 0; | |
410 | ||
411 | /* | |
412 | * It is safe to dereference the fd table without RCU or | |
413 | * ->file_lock because this is the last reference to the | |
414 | * files structure. But use RCU to shut RCU-lockdep up. | |
415 | */ | |
416 | rcu_read_lock(); | |
417 | fdt = files_fdtable(files); | |
418 | rcu_read_unlock(); | |
419 | for (;;) { | |
420 | unsigned long set; | |
421 | i = j * BITS_PER_LONG; | |
422 | if (i >= fdt->max_fds) | |
423 | break; | |
424 | set = fdt->open_fds[j++]; | |
425 | while (set) { | |
426 | if (set & 1) { | |
427 | struct file * file = xchg(&fdt->fd[i], NULL); | |
428 | if (file) { | |
429 | filp_close(file, files); | |
430 | cond_resched(); | |
431 | } | |
432 | } | |
433 | i++; | |
434 | set >>= 1; | |
435 | } | |
436 | } | |
437 | } | |
438 | ||
439 | struct files_struct *get_files_struct(struct task_struct *task) | |
440 | { | |
441 | struct files_struct *files; | |
442 | ||
443 | task_lock(task); | |
444 | files = task->files; | |
445 | if (files) | |
446 | atomic_inc(&files->count); | |
447 | task_unlock(task); | |
448 | ||
449 | return files; | |
450 | } | |
451 | ||
452 | void put_files_struct(struct files_struct *files) | |
453 | { | |
454 | struct fdtable *fdt; | |
455 | ||
456 | if (atomic_dec_and_test(&files->count)) { | |
457 | close_files(files); | |
458 | /* not really needed, since nobody can see us */ | |
459 | rcu_read_lock(); | |
460 | fdt = files_fdtable(files); | |
461 | rcu_read_unlock(); | |
462 | /* free the arrays if they are not embedded */ | |
463 | if (fdt != &files->fdtab) | |
464 | __free_fdtable(fdt); | |
465 | kmem_cache_free(files_cachep, files); | |
466 | } | |
467 | } | |
468 | ||
469 | void reset_files_struct(struct files_struct *files) | |
470 | { | |
471 | struct task_struct *tsk = current; | |
472 | struct files_struct *old; | |
473 | ||
474 | old = tsk->files; | |
475 | task_lock(tsk); | |
476 | tsk->files = files; | |
477 | task_unlock(tsk); | |
478 | put_files_struct(old); | |
479 | } | |
480 | ||
481 | void exit_files(struct task_struct *tsk) | |
482 | { | |
483 | struct files_struct * files = tsk->files; | |
484 | ||
485 | if (files) { | |
486 | task_lock(tsk); | |
487 | tsk->files = NULL; | |
488 | task_unlock(tsk); | |
489 | put_files_struct(files); | |
490 | } | |
491 | } | |
492 | ||
493 | static void fdtable_defer_list_init(int cpu) | |
494 | { | |
495 | struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu); | |
496 | spin_lock_init(&fddef->lock); | |
497 | INIT_WORK(&fddef->wq, free_fdtable_work); | |
498 | fddef->next = NULL; | |
499 | } | |
500 | ||
501 | void __init files_defer_init(void) | |
502 | { | |
503 | int i; | |
504 | for_each_possible_cpu(i) | |
505 | fdtable_defer_list_init(i); | |
506 | sysctl_nr_open_max = min((size_t)INT_MAX, ~(size_t)0/sizeof(void *)) & | |
507 | -BITS_PER_LONG; | |
508 | } | |
509 | ||
510 | struct files_struct init_files = { | |
511 | .count = ATOMIC_INIT(1), | |
512 | .fdt = &init_files.fdtab, | |
513 | .fdtab = { | |
514 | .max_fds = NR_OPEN_DEFAULT, | |
515 | .fd = &init_files.fd_array[0], | |
516 | .close_on_exec = init_files.close_on_exec_init, | |
517 | .open_fds = init_files.open_fds_init, | |
518 | }, | |
519 | .file_lock = __SPIN_LOCK_UNLOCKED(init_files.file_lock), | |
520 | }; | |
521 | ||
522 | /* | |
523 | * allocate a file descriptor, mark it busy. | |
524 | */ | |
525 | int __alloc_fd(struct files_struct *files, | |
526 | unsigned start, unsigned end, unsigned flags) | |
527 | { | |
528 | unsigned int fd; | |
529 | int error; | |
530 | struct fdtable *fdt; | |
531 | ||
532 | spin_lock(&files->file_lock); | |
533 | repeat: | |
534 | fdt = files_fdtable(files); | |
535 | fd = start; | |
536 | if (fd < files->next_fd) | |
537 | fd = files->next_fd; | |
538 | ||
539 | if (fd < fdt->max_fds) | |
540 | fd = find_next_zero_bit(fdt->open_fds, fdt->max_fds, fd); | |
541 | ||
542 | /* | |
543 | * N.B. For clone tasks sharing a files structure, this test | |
544 | * will limit the total number of files that can be opened. | |
545 | */ | |
546 | error = -EMFILE; | |
547 | if (fd >= end) | |
548 | goto out; | |
549 | ||
550 | error = expand_files(files, fd); | |
551 | if (error < 0) | |
552 | goto out; | |
553 | ||
554 | /* | |
555 | * If we needed to expand the fs array we | |
556 | * might have blocked - try again. | |
557 | */ | |
558 | if (error) | |
559 | goto repeat; | |
560 | ||
561 | if (start <= files->next_fd) | |
562 | files->next_fd = fd + 1; | |
563 | ||
564 | __set_open_fd(fd, fdt); | |
565 | if (flags & O_CLOEXEC) | |
566 | __set_close_on_exec(fd, fdt); | |
567 | else | |
568 | __clear_close_on_exec(fd, fdt); | |
569 | error = fd; | |
570 | #if 1 | |
571 | /* Sanity check */ | |
572 | if (rcu_dereference_raw(fdt->fd[fd]) != NULL) { | |
573 | printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd); | |
574 | rcu_assign_pointer(fdt->fd[fd], NULL); | |
575 | } | |
576 | #endif | |
577 | ||
578 | out: | |
579 | spin_unlock(&files->file_lock); | |
580 | return error; | |
581 | } | |
582 | ||
583 | static int alloc_fd(unsigned start, unsigned flags) | |
584 | { | |
585 | return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags); | |
586 | } | |
587 | ||
588 | int get_unused_fd_flags(unsigned flags) | |
589 | { | |
590 | return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags); | |
591 | } | |
592 | EXPORT_SYMBOL(get_unused_fd_flags); | |
593 | ||
594 | static void __put_unused_fd(struct files_struct *files, unsigned int fd) | |
595 | { | |
596 | struct fdtable *fdt = files_fdtable(files); | |
597 | __clear_open_fd(fd, fdt); | |
598 | if (fd < files->next_fd) | |
599 | files->next_fd = fd; | |
600 | } | |
601 | ||
602 | void put_unused_fd(unsigned int fd) | |
603 | { | |
604 | struct files_struct *files = current->files; | |
605 | spin_lock(&files->file_lock); | |
606 | __put_unused_fd(files, fd); | |
607 | spin_unlock(&files->file_lock); | |
608 | } | |
609 | ||
610 | EXPORT_SYMBOL(put_unused_fd); | |
611 | ||
612 | /* | |
613 | * Install a file pointer in the fd array. | |
614 | * | |
615 | * The VFS is full of places where we drop the files lock between | |
616 | * setting the open_fds bitmap and installing the file in the file | |
617 | * array. At any such point, we are vulnerable to a dup2() race | |
618 | * installing a file in the array before us. We need to detect this and | |
619 | * fput() the struct file we are about to overwrite in this case. | |
620 | * | |
621 | * It should never happen - if we allow dup2() do it, _really_ bad things | |
622 | * will follow. | |
623 | * | |
624 | * NOTE: __fd_install() variant is really, really low-level; don't | |
625 | * use it unless you are forced to by truly lousy API shoved down | |
626 | * your throat. 'files' *MUST* be either current->files or obtained | |
627 | * by get_files_struct(current) done by whoever had given it to you, | |
628 | * or really bad things will happen. Normally you want to use | |
629 | * fd_install() instead. | |
630 | */ | |
631 | ||
632 | void __fd_install(struct files_struct *files, unsigned int fd, | |
633 | struct file *file) | |
634 | { | |
635 | struct fdtable *fdt; | |
636 | spin_lock(&files->file_lock); | |
637 | fdt = files_fdtable(files); | |
638 | BUG_ON(fdt->fd[fd] != NULL); | |
639 | rcu_assign_pointer(fdt->fd[fd], file); | |
640 | spin_unlock(&files->file_lock); | |
641 | } | |
642 | ||
643 | void fd_install(unsigned int fd, struct file *file) | |
644 | { | |
645 | __fd_install(current->files, fd, file); | |
646 | } | |
647 | ||
648 | EXPORT_SYMBOL(fd_install); | |
649 | ||
650 | /* | |
651 | * The same warnings as for __alloc_fd()/__fd_install() apply here... | |
652 | */ | |
653 | int __close_fd(struct files_struct *files, unsigned fd) | |
654 | { | |
655 | struct file *file; | |
656 | struct fdtable *fdt; | |
657 | ||
658 | spin_lock(&files->file_lock); | |
659 | fdt = files_fdtable(files); | |
660 | if (fd >= fdt->max_fds) | |
661 | goto out_unlock; | |
662 | file = fdt->fd[fd]; | |
663 | if (!file) | |
664 | goto out_unlock; | |
665 | rcu_assign_pointer(fdt->fd[fd], NULL); | |
666 | __clear_close_on_exec(fd, fdt); | |
667 | __put_unused_fd(files, fd); | |
668 | spin_unlock(&files->file_lock); | |
669 | return filp_close(file, files); | |
670 | ||
671 | out_unlock: | |
672 | spin_unlock(&files->file_lock); | |
673 | return -EBADF; | |
674 | } | |
675 | ||
676 | void do_close_on_exec(struct files_struct *files) | |
677 | { | |
678 | unsigned i; | |
679 | struct fdtable *fdt; | |
680 | ||
681 | /* exec unshares first */ | |
682 | spin_lock(&files->file_lock); | |
683 | for (i = 0; ; i++) { | |
684 | unsigned long set; | |
685 | unsigned fd = i * BITS_PER_LONG; | |
686 | fdt = files_fdtable(files); | |
687 | if (fd >= fdt->max_fds) | |
688 | break; | |
689 | set = fdt->close_on_exec[i]; | |
690 | if (!set) | |
691 | continue; | |
692 | fdt->close_on_exec[i] = 0; | |
693 | for ( ; set ; fd++, set >>= 1) { | |
694 | struct file *file; | |
695 | if (!(set & 1)) | |
696 | continue; | |
697 | file = fdt->fd[fd]; | |
698 | if (!file) | |
699 | continue; | |
700 | rcu_assign_pointer(fdt->fd[fd], NULL); | |
701 | __put_unused_fd(files, fd); | |
702 | spin_unlock(&files->file_lock); | |
703 | filp_close(file, files); | |
704 | cond_resched(); | |
705 | spin_lock(&files->file_lock); | |
706 | } | |
707 | ||
708 | } | |
709 | spin_unlock(&files->file_lock); | |
710 | } | |
711 | ||
712 | struct file *fget(unsigned int fd) | |
713 | { | |
714 | struct file *file; | |
715 | struct files_struct *files = current->files; | |
716 | ||
717 | rcu_read_lock(); | |
718 | file = fcheck_files(files, fd); | |
719 | if (file) { | |
720 | /* File object ref couldn't be taken */ | |
721 | if (file->f_mode & FMODE_PATH || | |
722 | !atomic_long_inc_not_zero(&file->f_count)) | |
723 | file = NULL; | |
724 | } | |
725 | rcu_read_unlock(); | |
726 | ||
727 | return file; | |
728 | } | |
729 | ||
730 | EXPORT_SYMBOL(fget); | |
731 | ||
732 | struct file *fget_raw(unsigned int fd) | |
733 | { | |
734 | struct file *file; | |
735 | struct files_struct *files = current->files; | |
736 | ||
737 | rcu_read_lock(); | |
738 | file = fcheck_files(files, fd); | |
739 | if (file) { | |
740 | /* File object ref couldn't be taken */ | |
741 | if (!atomic_long_inc_not_zero(&file->f_count)) | |
742 | file = NULL; | |
743 | } | |
744 | rcu_read_unlock(); | |
745 | ||
746 | return file; | |
747 | } | |
748 | ||
749 | EXPORT_SYMBOL(fget_raw); | |
750 | ||
751 | /* | |
752 | * Lightweight file lookup - no refcnt increment if fd table isn't shared. | |
753 | * | |
754 | * You can use this instead of fget if you satisfy all of the following | |
755 | * conditions: | |
756 | * 1) You must call fput_light before exiting the syscall and returning control | |
757 | * to userspace (i.e. you cannot remember the returned struct file * after | |
758 | * returning to userspace). | |
759 | * 2) You must not call filp_close on the returned struct file * in between | |
760 | * calls to fget_light and fput_light. | |
761 | * 3) You must not clone the current task in between the calls to fget_light | |
762 | * and fput_light. | |
763 | * | |
764 | * The fput_needed flag returned by fget_light should be passed to the | |
765 | * corresponding fput_light. | |
766 | */ | |
767 | struct file *fget_light(unsigned int fd, int *fput_needed) | |
768 | { | |
769 | struct file *file; | |
770 | struct files_struct *files = current->files; | |
771 | ||
772 | *fput_needed = 0; | |
773 | if (atomic_read(&files->count) == 1) { | |
774 | file = fcheck_files(files, fd); | |
775 | if (file && (file->f_mode & FMODE_PATH)) | |
776 | file = NULL; | |
777 | } else { | |
778 | rcu_read_lock(); | |
779 | file = fcheck_files(files, fd); | |
780 | if (file) { | |
781 | if (!(file->f_mode & FMODE_PATH) && | |
782 | atomic_long_inc_not_zero(&file->f_count)) | |
783 | *fput_needed = 1; | |
784 | else | |
785 | /* Didn't get the reference, someone's freed */ | |
786 | file = NULL; | |
787 | } | |
788 | rcu_read_unlock(); | |
789 | } | |
790 | ||
791 | return file; | |
792 | } | |
793 | EXPORT_SYMBOL(fget_light); | |
794 | ||
795 | struct file *fget_raw_light(unsigned int fd, int *fput_needed) | |
796 | { | |
797 | struct file *file; | |
798 | struct files_struct *files = current->files; | |
799 | ||
800 | *fput_needed = 0; | |
801 | if (atomic_read(&files->count) == 1) { | |
802 | file = fcheck_files(files, fd); | |
803 | } else { | |
804 | rcu_read_lock(); | |
805 | file = fcheck_files(files, fd); | |
806 | if (file) { | |
807 | if (atomic_long_inc_not_zero(&file->f_count)) | |
808 | *fput_needed = 1; | |
809 | else | |
810 | /* Didn't get the reference, someone's freed */ | |
811 | file = NULL; | |
812 | } | |
813 | rcu_read_unlock(); | |
814 | } | |
815 | ||
816 | return file; | |
817 | } | |
818 | ||
819 | void set_close_on_exec(unsigned int fd, int flag) | |
820 | { | |
821 | struct files_struct *files = current->files; | |
822 | struct fdtable *fdt; | |
823 | spin_lock(&files->file_lock); | |
824 | fdt = files_fdtable(files); | |
825 | if (flag) | |
826 | __set_close_on_exec(fd, fdt); | |
827 | else | |
828 | __clear_close_on_exec(fd, fdt); | |
829 | spin_unlock(&files->file_lock); | |
830 | } | |
831 | ||
832 | bool get_close_on_exec(unsigned int fd) | |
833 | { | |
834 | struct files_struct *files = current->files; | |
835 | struct fdtable *fdt; | |
836 | bool res; | |
837 | rcu_read_lock(); | |
838 | fdt = files_fdtable(files); | |
839 | res = close_on_exec(fd, fdt); | |
840 | rcu_read_unlock(); | |
841 | return res; | |
842 | } | |
843 | ||
844 | static int do_dup2(struct files_struct *files, | |
845 | struct file *file, unsigned fd, unsigned flags) | |
846 | { | |
847 | struct file *tofree; | |
848 | struct fdtable *fdt; | |
849 | ||
850 | /* | |
851 | * We need to detect attempts to do dup2() over allocated but still | |
852 | * not finished descriptor. NB: OpenBSD avoids that at the price of | |
853 | * extra work in their equivalent of fget() - they insert struct | |
854 | * file immediately after grabbing descriptor, mark it larval if | |
855 | * more work (e.g. actual opening) is needed and make sure that | |
856 | * fget() treats larval files as absent. Potentially interesting, | |
857 | * but while extra work in fget() is trivial, locking implications | |
858 | * and amount of surgery on open()-related paths in VFS are not. | |
859 | * FreeBSD fails with -EBADF in the same situation, NetBSD "solution" | |
860 | * deadlocks in rather amusing ways, AFAICS. All of that is out of | |
861 | * scope of POSIX or SUS, since neither considers shared descriptor | |
862 | * tables and this condition does not arise without those. | |
863 | */ | |
864 | fdt = files_fdtable(files); | |
865 | tofree = fdt->fd[fd]; | |
866 | if (!tofree && fd_is_open(fd, fdt)) | |
867 | goto Ebusy; | |
868 | get_file(file); | |
869 | rcu_assign_pointer(fdt->fd[fd], file); | |
870 | __set_open_fd(fd, fdt); | |
871 | if (flags & O_CLOEXEC) | |
872 | __set_close_on_exec(fd, fdt); | |
873 | else | |
874 | __clear_close_on_exec(fd, fdt); | |
875 | spin_unlock(&files->file_lock); | |
876 | ||
877 | if (tofree) | |
878 | filp_close(tofree, files); | |
879 | ||
880 | return fd; | |
881 | ||
882 | Ebusy: | |
883 | spin_unlock(&files->file_lock); | |
884 | return -EBUSY; | |
885 | } | |
886 | ||
887 | int replace_fd(unsigned fd, struct file *file, unsigned flags) | |
888 | { | |
889 | int err; | |
890 | struct files_struct *files = current->files; | |
891 | ||
892 | if (!file) | |
893 | return __close_fd(files, fd); | |
894 | ||
895 | if (fd >= rlimit(RLIMIT_NOFILE)) | |
896 | return -EBADF; | |
897 | ||
898 | spin_lock(&files->file_lock); | |
899 | err = expand_files(files, fd); | |
900 | if (unlikely(err < 0)) | |
901 | goto out_unlock; | |
902 | return do_dup2(files, file, fd, flags); | |
903 | ||
904 | out_unlock: | |
905 | spin_unlock(&files->file_lock); | |
906 | return err; | |
907 | } | |
908 | ||
909 | SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags) | |
910 | { | |
911 | int err = -EBADF; | |
912 | struct file *file; | |
913 | struct files_struct *files = current->files; | |
914 | ||
915 | if ((flags & ~O_CLOEXEC) != 0) | |
916 | return -EINVAL; | |
917 | ||
918 | if (unlikely(oldfd == newfd)) | |
919 | return -EINVAL; | |
920 | ||
921 | if (newfd >= rlimit(RLIMIT_NOFILE)) | |
922 | return -EBADF; | |
923 | ||
924 | spin_lock(&files->file_lock); | |
925 | err = expand_files(files, newfd); | |
926 | file = fcheck(oldfd); | |
927 | if (unlikely(!file)) | |
928 | goto Ebadf; | |
929 | if (unlikely(err < 0)) { | |
930 | if (err == -EMFILE) | |
931 | goto Ebadf; | |
932 | goto out_unlock; | |
933 | } | |
934 | return do_dup2(files, file, newfd, flags); | |
935 | ||
936 | Ebadf: | |
937 | err = -EBADF; | |
938 | out_unlock: | |
939 | spin_unlock(&files->file_lock); | |
940 | return err; | |
941 | } | |
942 | ||
943 | SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd) | |
944 | { | |
945 | if (unlikely(newfd == oldfd)) { /* corner case */ | |
946 | struct files_struct *files = current->files; | |
947 | int retval = oldfd; | |
948 | ||
949 | rcu_read_lock(); | |
950 | if (!fcheck_files(files, oldfd)) | |
951 | retval = -EBADF; | |
952 | rcu_read_unlock(); | |
953 | return retval; | |
954 | } | |
955 | return sys_dup3(oldfd, newfd, 0); | |
956 | } | |
957 | ||
958 | SYSCALL_DEFINE1(dup, unsigned int, fildes) | |
959 | { | |
960 | int ret = -EBADF; | |
961 | struct file *file = fget_raw(fildes); | |
962 | ||
963 | if (file) { | |
964 | ret = get_unused_fd(); | |
965 | if (ret >= 0) | |
966 | fd_install(ret, file); | |
967 | else | |
968 | fput(file); | |
969 | } | |
970 | return ret; | |
971 | } | |
972 | ||
973 | int f_dupfd(unsigned int from, struct file *file, unsigned flags) | |
974 | { | |
975 | int err; | |
976 | if (from >= rlimit(RLIMIT_NOFILE)) | |
977 | return -EINVAL; | |
978 | err = alloc_fd(from, flags); | |
979 | if (err >= 0) { | |
980 | get_file(file); | |
981 | fd_install(err, file); | |
982 | } | |
983 | return err; | |
984 | } | |
985 | ||
986 | int iterate_fd(struct files_struct *files, unsigned n, | |
987 | int (*f)(const void *, struct file *, unsigned), | |
988 | const void *p) | |
989 | { | |
990 | struct fdtable *fdt; | |
991 | int res = 0; | |
992 | if (!files) | |
993 | return 0; | |
994 | spin_lock(&files->file_lock); | |
995 | for (fdt = files_fdtable(files); n < fdt->max_fds; n++) { | |
996 | struct file *file; | |
997 | file = rcu_dereference_check_fdtable(files, fdt->fd[n]); | |
998 | if (!file) | |
999 | continue; | |
1000 | res = f(p, file, n); | |
1001 | if (res) | |
1002 | break; | |
1003 | } | |
1004 | spin_unlock(&files->file_lock); | |
1005 | return res; | |
1006 | } | |
1007 | EXPORT_SYMBOL(iterate_fd); |