]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
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 | ||
1027abe8 | 9 | #include <linux/module.h> |
1da177e4 LT |
10 | #include <linux/fs.h> |
11 | #include <linux/mm.h> | |
12 | #include <linux/time.h> | |
d43c36dc | 13 | #include <linux/sched.h> |
1da177e4 LT |
14 | #include <linux/slab.h> |
15 | #include <linux/vmalloc.h> | |
16 | #include <linux/file.h> | |
9f3acc31 | 17 | #include <linux/fdtable.h> |
1da177e4 | 18 | #include <linux/bitops.h> |
ab2af1f5 DS |
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; | |
ab2af1f5 DS |
27 | struct fdtable *next; |
28 | }; | |
29 | ||
9cfe015a | 30 | int sysctl_nr_open __read_mostly = 1024*1024; |
eceea0b3 AV |
31 | int sysctl_nr_open_min = BITS_PER_LONG; |
32 | int sysctl_nr_open_max = 1024 * 1024; /* raised later */ | |
9cfe015a | 33 | |
ab2af1f5 DS |
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); | |
1da177e4 | 41 | |
5466b456 | 42 | static inline void * alloc_fdmem(unsigned int size) |
1da177e4 | 43 | { |
1da177e4 | 44 | if (size <= PAGE_SIZE) |
5466b456 VL |
45 | return kmalloc(size, GFP_KERNEL); |
46 | else | |
47 | return vmalloc(size); | |
1da177e4 LT |
48 | } |
49 | ||
5466b456 | 50 | static inline void free_fdarr(struct fdtable *fdt) |
1da177e4 | 51 | { |
5466b456 VL |
52 | if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *))) |
53 | kfree(fdt->fd); | |
1da177e4 | 54 | else |
5466b456 | 55 | vfree(fdt->fd); |
1da177e4 LT |
56 | } |
57 | ||
5466b456 | 58 | static inline void free_fdset(struct fdtable *fdt) |
1da177e4 | 59 | { |
5466b456 VL |
60 | if (fdt->max_fds <= (PAGE_SIZE * BITS_PER_BYTE / 2)) |
61 | kfree(fdt->open_fds); | |
62 | else | |
63 | vfree(fdt->open_fds); | |
ab2af1f5 | 64 | } |
1da177e4 | 65 | |
65f27f38 | 66 | static void free_fdtable_work(struct work_struct *work) |
ab2af1f5 | 67 | { |
65f27f38 DH |
68 | struct fdtable_defer *f = |
69 | container_of(work, struct fdtable_defer, wq); | |
ab2af1f5 | 70 | struct fdtable *fdt; |
1da177e4 | 71 | |
ab2af1f5 DS |
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; | |
5466b456 VL |
78 | vfree(fdt->fd); |
79 | free_fdset(fdt); | |
80 | kfree(fdt); | |
ab2af1f5 DS |
81 | fdt = next; |
82 | } | |
83 | } | |
1da177e4 | 84 | |
4fd45812 | 85 | void free_fdtable_rcu(struct rcu_head *rcu) |
ab2af1f5 DS |
86 | { |
87 | struct fdtable *fdt = container_of(rcu, struct fdtable, rcu); | |
ab2af1f5 | 88 | struct fdtable_defer *fddef; |
1da177e4 | 89 | |
ab2af1f5 | 90 | BUG_ON(!fdt); |
ab2af1f5 | 91 | |
4fd45812 | 92 | if (fdt->max_fds <= NR_OPEN_DEFAULT) { |
ab2af1f5 | 93 | /* |
4fd45812 VL |
94 | * This fdtable is embedded in the files structure and that |
95 | * structure itself is getting destroyed. | |
ab2af1f5 | 96 | */ |
4fd45812 VL |
97 | kmem_cache_free(files_cachep, |
98 | container_of(fdt, struct files_struct, fdtab)); | |
ab2af1f5 DS |
99 | return; |
100 | } | |
5466b456 | 101 | if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *))) { |
ab2af1f5 | 102 | kfree(fdt->fd); |
5466b456 | 103 | kfree(fdt->open_fds); |
ab2af1f5 | 104 | kfree(fdt); |
1da177e4 | 105 | } else { |
ab2af1f5 DS |
106 | fddef = &get_cpu_var(fdtable_defer_list); |
107 | spin_lock(&fddef->lock); | |
108 | fdt->next = fddef->next; | |
109 | fddef->next = fdt; | |
593be07a TH |
110 | /* vmallocs are handled from the workqueue context */ |
111 | schedule_work(&fddef->wq); | |
ab2af1f5 DS |
112 | spin_unlock(&fddef->lock); |
113 | put_cpu_var(fdtable_defer_list); | |
1da177e4 | 114 | } |
ab2af1f5 DS |
115 | } |
116 | ||
ab2af1f5 DS |
117 | /* |
118 | * Expand the fdset in the files_struct. Called with the files spinlock | |
119 | * held for write. | |
120 | */ | |
5466b456 | 121 | static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt) |
ab2af1f5 | 122 | { |
5466b456 | 123 | unsigned int cpy, set; |
ab2af1f5 | 124 | |
5466b456 | 125 | BUG_ON(nfdt->max_fds < ofdt->max_fds); |
5466b456 VL |
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); | |
1da177e4 LT |
138 | } |
139 | ||
5466b456 | 140 | static struct fdtable * alloc_fdtable(unsigned int nr) |
1da177e4 | 141 | { |
5466b456 VL |
142 | struct fdtable *fdt; |
143 | char *data; | |
1da177e4 | 144 | |
ab2af1f5 | 145 | /* |
5466b456 VL |
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. | |
ab2af1f5 | 151 | */ |
5466b456 VL |
152 | nr /= (1024 / sizeof(struct file *)); |
153 | nr = roundup_pow_of_two(nr + 1); | |
154 | nr *= (1024 / sizeof(struct file *)); | |
5c598b34 AV |
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; | |
bbea9f69 | 165 | |
5466b456 VL |
166 | fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL); |
167 | if (!fdt) | |
bbea9f69 | 168 | goto out; |
5466b456 VL |
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; | |
5466b456 VL |
181 | fdt->next = NULL; |
182 | ||
ab2af1f5 | 183 | return fdt; |
5466b456 VL |
184 | |
185 | out_arr: | |
186 | free_fdarr(fdt); | |
187 | out_fdt: | |
ab2af1f5 | 188 | kfree(fdt); |
5466b456 | 189 | out: |
ab2af1f5 DS |
190 | return NULL; |
191 | } | |
1da177e4 | 192 | |
ab2af1f5 | 193 | /* |
74d392aa VL |
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. | |
ab2af1f5 DS |
199 | */ |
200 | static int expand_fdtable(struct files_struct *files, int nr) | |
201 | __releases(files->file_lock) | |
202 | __acquires(files->file_lock) | |
203 | { | |
74d392aa | 204 | struct fdtable *new_fdt, *cur_fdt; |
ab2af1f5 DS |
205 | |
206 | spin_unlock(&files->file_lock); | |
74d392aa | 207 | new_fdt = alloc_fdtable(nr); |
ab2af1f5 | 208 | spin_lock(&files->file_lock); |
74d392aa VL |
209 | if (!new_fdt) |
210 | return -ENOMEM; | |
5c598b34 AV |
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 | } | |
ab2af1f5 | 221 | /* |
74d392aa VL |
222 | * Check again since another task may have expanded the fd table while |
223 | * we dropped the lock | |
ab2af1f5 | 224 | */ |
74d392aa | 225 | cur_fdt = files_fdtable(files); |
bbea9f69 | 226 | if (nr >= cur_fdt->max_fds) { |
74d392aa VL |
227 | /* Continue as planned */ |
228 | copy_fdtable(new_fdt, cur_fdt); | |
229 | rcu_assign_pointer(files->fdt, new_fdt); | |
4fd45812 | 230 | if (cur_fdt->max_fds > NR_OPEN_DEFAULT) |
01b2d93c | 231 | free_fdtable(cur_fdt); |
ab2af1f5 | 232 | } else { |
74d392aa | 233 | /* Somebody else expanded, so undo our attempt */ |
5466b456 VL |
234 | free_fdarr(new_fdt); |
235 | free_fdset(new_fdt); | |
236 | kfree(new_fdt); | |
ab2af1f5 | 237 | } |
74d392aa | 238 | return 1; |
1da177e4 LT |
239 | } |
240 | ||
241 | /* | |
242 | * Expand files. | |
74d392aa VL |
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. | |
1da177e4 LT |
248 | */ |
249 | int expand_files(struct files_struct *files, int nr) | |
250 | { | |
badf1662 | 251 | struct fdtable *fdt; |
1da177e4 | 252 | |
badf1662 | 253 | fdt = files_fdtable(files); |
4e1e018e AV |
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 | */ | |
d554ed89 | 259 | if (nr >= rlimit(RLIMIT_NOFILE)) |
4e1e018e AV |
260 | return -EMFILE; |
261 | ||
74d392aa | 262 | /* Do we need to expand? */ |
bbea9f69 | 263 | if (nr < fdt->max_fds) |
74d392aa | 264 | return 0; |
4e1e018e | 265 | |
74d392aa | 266 | /* Can we expand? */ |
9cfe015a | 267 | if (nr >= sysctl_nr_open) |
74d392aa VL |
268 | return -EMFILE; |
269 | ||
270 | /* All good, so we try */ | |
271 | return expand_fdtable(files, nr); | |
1da177e4 | 272 | } |
ab2af1f5 | 273 | |
02afc626 AV |
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 | ||
02afc626 AV |
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; | |
afbec7ff | 301 | newf = kmem_cache_alloc(files_cachep, GFP_KERNEL); |
02afc626 AV |
302 | if (!newf) |
303 | goto out; | |
304 | ||
afbec7ff AV |
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]; | |
afbec7ff AV |
314 | new_fdt->next = NULL; |
315 | ||
02afc626 AV |
316 | spin_lock(&oldf->file_lock); |
317 | old_fdt = files_fdtable(oldf); | |
02afc626 AV |
318 | open_files = count_open_files(old_fdt); |
319 | ||
320 | /* | |
321 | * Check whether we need to allocate a larger fd array and fd set. | |
02afc626 | 322 | */ |
adbecb12 | 323 | while (unlikely(open_files > new_fdt->max_fds)) { |
02afc626 | 324 | spin_unlock(&oldf->file_lock); |
9dec3c4d | 325 | |
adbecb12 AV |
326 | if (new_fdt != &newf->fdtab) { |
327 | free_fdarr(new_fdt); | |
328 | free_fdset(new_fdt); | |
329 | kfree(new_fdt); | |
330 | } | |
331 | ||
9dec3c4d AV |
332 | new_fdt = alloc_fdtable(open_files - 1); |
333 | if (!new_fdt) { | |
334 | *errorp = -ENOMEM; | |
335 | goto out_release; | |
336 | } | |
337 | ||
338 | /* beyond sysctl_nr_open; nothing to do */ | |
339 | if (unlikely(new_fdt->max_fds < open_files)) { | |
340 | free_fdarr(new_fdt); | |
341 | free_fdset(new_fdt); | |
342 | kfree(new_fdt); | |
343 | *errorp = -EMFILE; | |
02afc626 | 344 | goto out_release; |
9dec3c4d | 345 | } |
9dec3c4d | 346 | |
02afc626 AV |
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); | |
adbecb12 | 354 | open_files = count_open_files(old_fdt); |
02afc626 AV |
355 | } |
356 | ||
357 | old_fds = old_fdt->fd; | |
358 | new_fds = new_fdt->fd; | |
359 | ||
360 | memcpy(new_fdt->open_fds->fds_bits, | |
361 | old_fdt->open_fds->fds_bits, open_files/8); | |
362 | memcpy(new_fdt->close_on_exec->fds_bits, | |
363 | old_fdt->close_on_exec->fds_bits, open_files/8); | |
364 | ||
365 | for (i = open_files; i != 0; i--) { | |
366 | struct file *f = *old_fds++; | |
367 | if (f) { | |
368 | get_file(f); | |
369 | } else { | |
370 | /* | |
371 | * The fd may be claimed in the fd bitmap but not yet | |
372 | * instantiated in the files array if a sibling thread | |
373 | * is partway through open(). So make sure that this | |
374 | * fd is available to the new process. | |
375 | */ | |
376 | FD_CLR(open_files - i, new_fdt->open_fds); | |
377 | } | |
378 | rcu_assign_pointer(*new_fds++, f); | |
379 | } | |
380 | spin_unlock(&oldf->file_lock); | |
381 | ||
382 | /* compute the remainder to be cleared */ | |
383 | size = (new_fdt->max_fds - open_files) * sizeof(struct file *); | |
384 | ||
385 | /* This is long word aligned thus could use a optimized version */ | |
386 | memset(new_fds, 0, size); | |
387 | ||
388 | if (new_fdt->max_fds > open_files) { | |
389 | int left = (new_fdt->max_fds-open_files)/8; | |
390 | int start = open_files / (8 * sizeof(unsigned long)); | |
391 | ||
392 | memset(&new_fdt->open_fds->fds_bits[start], 0, left); | |
393 | memset(&new_fdt->close_on_exec->fds_bits[start], 0, left); | |
394 | } | |
395 | ||
afbec7ff AV |
396 | rcu_assign_pointer(newf->fdt, new_fdt); |
397 | ||
02afc626 AV |
398 | return newf; |
399 | ||
400 | out_release: | |
401 | kmem_cache_free(files_cachep, newf); | |
402 | out: | |
403 | return NULL; | |
404 | } | |
405 | ||
ab2af1f5 DS |
406 | static void __devinit fdtable_defer_list_init(int cpu) |
407 | { | |
408 | struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu); | |
409 | spin_lock_init(&fddef->lock); | |
65f27f38 | 410 | INIT_WORK(&fddef->wq, free_fdtable_work); |
ab2af1f5 DS |
411 | fddef->next = NULL; |
412 | } | |
413 | ||
414 | void __init files_defer_init(void) | |
415 | { | |
416 | int i; | |
0a945022 | 417 | for_each_possible_cpu(i) |
ab2af1f5 | 418 | fdtable_defer_list_init(i); |
eceea0b3 AV |
419 | sysctl_nr_open_max = min((size_t)INT_MAX, ~(size_t)0/sizeof(void *)) & |
420 | -BITS_PER_LONG; | |
ab2af1f5 | 421 | } |
f52111b1 AV |
422 | |
423 | struct files_struct init_files = { | |
424 | .count = ATOMIC_INIT(1), | |
425 | .fdt = &init_files.fdtab, | |
426 | .fdtab = { | |
427 | .max_fds = NR_OPEN_DEFAULT, | |
428 | .fd = &init_files.fd_array[0], | |
429 | .close_on_exec = (fd_set *)&init_files.close_on_exec_init, | |
430 | .open_fds = (fd_set *)&init_files.open_fds_init, | |
f52111b1 AV |
431 | }, |
432 | .file_lock = __SPIN_LOCK_UNLOCKED(init_task.file_lock), | |
433 | }; | |
1027abe8 AV |
434 | |
435 | /* | |
436 | * allocate a file descriptor, mark it busy. | |
437 | */ | |
438 | int alloc_fd(unsigned start, unsigned flags) | |
439 | { | |
440 | struct files_struct *files = current->files; | |
441 | unsigned int fd; | |
442 | int error; | |
443 | struct fdtable *fdt; | |
444 | ||
445 | spin_lock(&files->file_lock); | |
446 | repeat: | |
447 | fdt = files_fdtable(files); | |
448 | fd = start; | |
449 | if (fd < files->next_fd) | |
450 | fd = files->next_fd; | |
451 | ||
452 | if (fd < fdt->max_fds) | |
453 | fd = find_next_zero_bit(fdt->open_fds->fds_bits, | |
454 | fdt->max_fds, fd); | |
455 | ||
456 | error = expand_files(files, fd); | |
457 | if (error < 0) | |
458 | goto out; | |
459 | ||
460 | /* | |
461 | * If we needed to expand the fs array we | |
462 | * might have blocked - try again. | |
463 | */ | |
464 | if (error) | |
465 | goto repeat; | |
466 | ||
467 | if (start <= files->next_fd) | |
468 | files->next_fd = fd + 1; | |
469 | ||
470 | FD_SET(fd, fdt->open_fds); | |
471 | if (flags & O_CLOEXEC) | |
472 | FD_SET(fd, fdt->close_on_exec); | |
473 | else | |
474 | FD_CLR(fd, fdt->close_on_exec); | |
475 | error = fd; | |
476 | #if 1 | |
477 | /* Sanity check */ | |
7dc52157 | 478 | if (rcu_dereference_raw(fdt->fd[fd]) != NULL) { |
1027abe8 AV |
479 | printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd); |
480 | rcu_assign_pointer(fdt->fd[fd], NULL); | |
481 | } | |
482 | #endif | |
483 | ||
484 | out: | |
485 | spin_unlock(&files->file_lock); | |
486 | return error; | |
487 | } | |
488 | ||
489 | int get_unused_fd(void) | |
490 | { | |
491 | return alloc_fd(0, 0); | |
492 | } | |
493 | EXPORT_SYMBOL(get_unused_fd); |