blob: d8773b19fe47fd23d6bbec2149679b123a79c139 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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/fs.h>
10#include <linux/mm.h>
11#include <linux/time.h>
12#include <linux/slab.h>
13#include <linux/vmalloc.h>
14#include <linux/file.h>
Al Viro9f3acc32008-04-24 07:44:08 -040015#include <linux/fdtable.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/bitops.h>
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070017#include <linux/interrupt.h>
18#include <linux/spinlock.h>
19#include <linux/rcupdate.h>
20#include <linux/workqueue.h>
21
22struct fdtable_defer {
23 spinlock_t lock;
24 struct work_struct wq;
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070025 struct fdtable *next;
26};
27
Eric Dumazet9cfe0152008-02-06 01:37:16 -080028int sysctl_nr_open __read_mostly = 1024*1024;
Al Viroeceea0b2008-05-10 10:08:32 -040029int sysctl_nr_open_min = BITS_PER_LONG;
30int sysctl_nr_open_max = 1024 * 1024; /* raised later */
Eric Dumazet9cfe0152008-02-06 01:37:16 -080031
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070032/*
33 * We use this list to defer free fdtables that have vmalloced
34 * sets/arrays. By keeping a per-cpu list, we avoid having to embed
35 * the work_struct in fdtable itself which avoids a 64 byte (i386) increase in
36 * this per-task structure.
37 */
38static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -070039
Vadim Lobanov5466b452006-12-10 02:21:22 -080040static inline void * alloc_fdmem(unsigned int size)
Linus Torvalds1da177e2005-04-16 15:20:36 -070041{
Linus Torvalds1da177e2005-04-16 15:20:36 -070042 if (size <= PAGE_SIZE)
Vadim Lobanov5466b452006-12-10 02:21:22 -080043 return kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -070044 else
Vadim Lobanov5466b452006-12-10 02:21:22 -080045 return vmalloc(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -070046}
47
Vadim Lobanov5466b452006-12-10 02:21:22 -080048static inline void free_fdarr(struct fdtable *fdt)
Linus Torvalds1da177e2005-04-16 15:20:36 -070049{
Vadim Lobanov5466b452006-12-10 02:21:22 -080050 if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *)))
51 kfree(fdt->fd);
52 else
53 vfree(fdt->fd);
54}
55
56static inline void free_fdset(struct fdtable *fdt)
57{
58 if (fdt->max_fds <= (PAGE_SIZE * BITS_PER_BYTE / 2))
59 kfree(fdt->open_fds);
60 else
61 vfree(fdt->open_fds);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070062}
63
David Howells65f27f32006-11-22 14:55:48 +000064static void free_fdtable_work(struct work_struct *work)
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070065{
David Howells65f27f32006-11-22 14:55:48 +000066 struct fdtable_defer *f =
67 container_of(work, struct fdtable_defer, wq);
Dipankar Sarmabadf1662005-09-09 13:04:10 -070068 struct fdtable *fdt;
Linus Torvalds1da177e2005-04-16 15:20:36 -070069
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070070 spin_lock_bh(&f->lock);
71 fdt = f->next;
72 f->next = NULL;
73 spin_unlock_bh(&f->lock);
74 while(fdt) {
75 struct fdtable *next = fdt->next;
Vadim Lobanov5466b452006-12-10 02:21:22 -080076 vfree(fdt->fd);
77 free_fdset(fdt);
78 kfree(fdt);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070079 fdt = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -070080 }
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070081}
82
Vadim Lobanov4fd45812006-12-10 02:21:17 -080083void free_fdtable_rcu(struct rcu_head *rcu)
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070084{
85 struct fdtable *fdt = container_of(rcu, struct fdtable, rcu);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070086 struct fdtable_defer *fddef;
87
88 BUG_ON(!fdt);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070089
Vadim Lobanov4fd45812006-12-10 02:21:17 -080090 if (fdt->max_fds <= NR_OPEN_DEFAULT) {
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070091 /*
Vadim Lobanov4fd45812006-12-10 02:21:17 -080092 * This fdtable is embedded in the files structure and that
93 * structure itself is getting destroyed.
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070094 */
Vadim Lobanov4fd45812006-12-10 02:21:17 -080095 kmem_cache_free(files_cachep,
96 container_of(fdt, struct files_struct, fdtab));
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070097 return;
98 }
Vadim Lobanov5466b452006-12-10 02:21:22 -080099 if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *))) {
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700100 kfree(fdt->fd);
Vadim Lobanov5466b452006-12-10 02:21:22 -0800101 kfree(fdt->open_fds);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700102 kfree(fdt);
103 } else {
104 fddef = &get_cpu_var(fdtable_defer_list);
105 spin_lock(&fddef->lock);
106 fdt->next = fddef->next;
107 fddef->next = fdt;
Tejun Heo593be072006-12-06 20:36:01 -0800108 /* vmallocs are handled from the workqueue context */
109 schedule_work(&fddef->wq);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700110 spin_unlock(&fddef->lock);
111 put_cpu_var(fdtable_defer_list);
112 }
113}
114
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700115/*
116 * Expand the fdset in the files_struct. Called with the files spinlock
117 * held for write.
118 */
Vadim Lobanov5466b452006-12-10 02:21:22 -0800119static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700120{
Vadim Lobanov5466b452006-12-10 02:21:22 -0800121 unsigned int cpy, set;
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700122
Vadim Lobanov5466b452006-12-10 02:21:22 -0800123 BUG_ON(nfdt->max_fds < ofdt->max_fds);
Vadim Lobanov5466b452006-12-10 02:21:22 -0800124
125 cpy = ofdt->max_fds * sizeof(struct file *);
126 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
127 memcpy(nfdt->fd, ofdt->fd, cpy);
128 memset((char *)(nfdt->fd) + cpy, 0, set);
129
130 cpy = ofdt->max_fds / BITS_PER_BYTE;
131 set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
132 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
133 memset((char *)(nfdt->open_fds) + cpy, 0, set);
134 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
135 memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136}
137
Vadim Lobanov5466b452006-12-10 02:21:22 -0800138static struct fdtable * alloc_fdtable(unsigned int nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139{
Vadim Lobanov5466b452006-12-10 02:21:22 -0800140 struct fdtable *fdt;
141 char *data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700143 /*
Vadim Lobanov5466b452006-12-10 02:21:22 -0800144 * Figure out how many fds we actually want to support in this fdtable.
145 * Allocation steps are keyed to the size of the fdarray, since it
146 * grows far faster than any of the other dynamic data. We try to fit
147 * the fdarray into comfortable page-tuned chunks: starting at 1024B
148 * and growing in powers of two from there on.
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700149 */
Vadim Lobanov5466b452006-12-10 02:21:22 -0800150 nr /= (1024 / sizeof(struct file *));
151 nr = roundup_pow_of_two(nr + 1);
152 nr *= (1024 / sizeof(struct file *));
Al Viro5c598b32008-04-27 20:04:15 -0400153 /*
154 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
155 * had been set lower between the check in expand_files() and here. Deal
156 * with that in caller, it's cheaper that way.
157 *
158 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
159 * bitmaps handling below becomes unpleasant, to put it mildly...
160 */
161 if (unlikely(nr > sysctl_nr_open))
162 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
Vadim Lobanovbbea9f62006-12-10 02:21:12 -0800163
Vadim Lobanov5466b452006-12-10 02:21:22 -0800164 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
165 if (!fdt)
Vadim Lobanovbbea9f62006-12-10 02:21:12 -0800166 goto out;
Vadim Lobanov5466b452006-12-10 02:21:22 -0800167 fdt->max_fds = nr;
168 data = alloc_fdmem(nr * sizeof(struct file *));
169 if (!data)
170 goto out_fdt;
171 fdt->fd = (struct file **)data;
172 data = alloc_fdmem(max_t(unsigned int,
173 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
174 if (!data)
175 goto out_arr;
176 fdt->open_fds = (fd_set *)data;
177 data += nr / BITS_PER_BYTE;
178 fdt->close_on_exec = (fd_set *)data;
179 INIT_RCU_HEAD(&fdt->rcu);
180 fdt->next = NULL;
181
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700182 return fdt;
Vadim Lobanov5466b452006-12-10 02:21:22 -0800183
184out_arr:
185 free_fdarr(fdt);
186out_fdt:
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700187 kfree(fdt);
Vadim Lobanov5466b452006-12-10 02:21:22 -0800188out:
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700189 return NULL;
190}
191
192/*
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700193 * Expand the file descriptor table.
194 * This function will allocate a new fdtable and both fd array and fdset, of
195 * the given size.
196 * Return <0 error code on error; 1 on successful completion.
197 * The files->file_lock should be held on entry, and will be held on exit.
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700198 */
199static int expand_fdtable(struct files_struct *files, int nr)
200 __releases(files->file_lock)
201 __acquires(files->file_lock)
202{
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700203 struct fdtable *new_fdt, *cur_fdt;
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700204
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205 spin_unlock(&files->file_lock);
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700206 new_fdt = alloc_fdtable(nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207 spin_lock(&files->file_lock);
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700208 if (!new_fdt)
209 return -ENOMEM;
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700210 /*
Al Viro5c598b32008-04-27 20:04:15 -0400211 * extremely unlikely race - sysctl_nr_open decreased between the check in
212 * caller and alloc_fdtable(). Cheaper to catch it here...
213 */
214 if (unlikely(new_fdt->max_fds <= nr)) {
215 free_fdarr(new_fdt);
216 free_fdset(new_fdt);
217 kfree(new_fdt);
218 return -EMFILE;
219 }
220 /*
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700221 * Check again since another task may have expanded the fd table while
222 * we dropped the lock
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700223 */
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700224 cur_fdt = files_fdtable(files);
Vadim Lobanovbbea9f62006-12-10 02:21:12 -0800225 if (nr >= cur_fdt->max_fds) {
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700226 /* Continue as planned */
227 copy_fdtable(new_fdt, cur_fdt);
228 rcu_assign_pointer(files->fdt, new_fdt);
Vadim Lobanov4fd45812006-12-10 02:21:17 -0800229 if (cur_fdt->max_fds > NR_OPEN_DEFAULT)
Vadim Lobanov01b2d932006-12-22 01:10:43 -0800230 free_fdtable(cur_fdt);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700231 } else {
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700232 /* Somebody else expanded, so undo our attempt */
Vadim Lobanov5466b452006-12-10 02:21:22 -0800233 free_fdarr(new_fdt);
234 free_fdset(new_fdt);
235 kfree(new_fdt);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700236 }
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700237 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238}
239
240/*
241 * Expand files.
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700242 * 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.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 */
248int expand_files(struct files_struct *files, int nr)
249{
Dipankar Sarmabadf1662005-09-09 13:04:10 -0700250 struct fdtable *fdt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251
Dipankar Sarmabadf1662005-09-09 13:04:10 -0700252 fdt = files_fdtable(files);
Al Viro4e1e0182008-07-26 16:01:20 -0400253
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 >= current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
259 return -EMFILE;
260
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700261 /* Do we need to expand? */
Vadim Lobanovbbea9f62006-12-10 02:21:12 -0800262 if (nr < fdt->max_fds)
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700263 return 0;
Al Viro4e1e0182008-07-26 16:01:20 -0400264
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700265 /* Can we expand? */
Eric Dumazet9cfe0152008-02-06 01:37:16 -0800266 if (nr >= sysctl_nr_open)
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700267 return -EMFILE;
268
269 /* All good, so we try */
270 return expand_fdtable(files, nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271}
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700272
Al Viro02afc622008-05-08 19:42:56 -0400273static 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/(8*sizeof(long)); i > 0; ) {
280 if (fdt->open_fds->fds_bits[--i])
281 break;
282 }
283 i = (i+1) * 8 * sizeof(long);
284 return i;
285}
286
Al Viro02afc622008-05-08 19:42:56 -0400287/*
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 */
292struct 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;
Al Viroafbec7f2008-05-08 21:11:17 -0400300 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
Al Viro02afc622008-05-08 19:42:56 -0400301 if (!newf)
302 goto out;
303
Al Viroafbec7f2008-05-08 21:11:17 -0400304 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 = (fd_set *)&newf->close_on_exec_init;
311 new_fdt->open_fds = (fd_set *)&newf->open_fds_init;
312 new_fdt->fd = &newf->fd_array[0];
313 INIT_RCU_HEAD(&new_fdt->rcu);
314 new_fdt->next = NULL;
315
Al Viro02afc622008-05-08 19:42:56 -0400316 spin_lock(&oldf->file_lock);
317 old_fdt = files_fdtable(oldf);
Al Viro02afc622008-05-08 19:42:56 -0400318 open_files = count_open_files(old_fdt);
319
320 /*
321 * Check whether we need to allocate a larger fd array and fd set.
Al Viro02afc622008-05-08 19:42:56 -0400322 */
Al Viroadbecb12008-05-08 21:19:42 -0400323 while (unlikely(open_files > new_fdt->max_fds)) {
Al Viro02afc622008-05-08 19:42:56 -0400324 spin_unlock(&oldf->file_lock);
Al Viro9dec3c42008-05-08 21:02:45 -0400325
Al Viroadbecb12008-05-08 21:19:42 -0400326 if (new_fdt != &newf->fdtab) {
327 free_fdarr(new_fdt);
328 free_fdset(new_fdt);
329 kfree(new_fdt);
330 }
331
Al Viro9dec3c42008-05-08 21:02:45 -0400332 new_fdt = alloc_fdtable(open_files - 1);
333 if (!new_fdt) {
334 *errorp = -ENOMEM;
Al Viro02afc622008-05-08 19:42:56 -0400335 goto out_release;
Al Viro9dec3c42008-05-08 21:02:45 -0400336 }
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;
344 goto out_release;
345 }
Al Viro9dec3c42008-05-08 21:02:45 -0400346
Al Viro02afc622008-05-08 19:42:56 -0400347 /*
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);
Al Viroadbecb12008-05-08 21:19:42 -0400354 open_files = count_open_files(old_fdt);
Al Viro02afc622008-05-08 19:42:56 -0400355 }
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
Al Viroafbec7f2008-05-08 21:11:17 -0400396 rcu_assign_pointer(newf->fdt, new_fdt);
397
Al Viro02afc622008-05-08 19:42:56 -0400398 return newf;
399
400out_release:
401 kmem_cache_free(files_cachep, newf);
402out:
403 return NULL;
404}
405
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700406static 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);
David Howells65f27f32006-11-22 14:55:48 +0000410 INIT_WORK(&fddef->wq, free_fdtable_work);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700411 fddef->next = NULL;
412}
413
414void __init files_defer_init(void)
415{
416 int i;
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -0800417 for_each_possible_cpu(i)
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700418 fdtable_defer_list_init(i);
Al Viroeceea0b2008-05-10 10:08:32 -0400419 sysctl_nr_open_max = min((size_t)INT_MAX, ~(size_t)0/sizeof(void *)) &
420 -BITS_PER_LONG;
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700421}
Al Virof52111b2008-05-08 18:19:16 -0400422
423struct 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,
431 .rcu = RCU_HEAD_INIT,
432 },
433 .file_lock = __SPIN_LOCK_UNLOCKED(init_task.file_lock),
434};