Loading...
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/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
25struct fdtable_defer {
26 spinlock_t lock;
27 struct work_struct wq;
28 struct fdtable *next;
29};
30
31int sysctl_nr_open __read_mostly = 1024*1024;
32int sysctl_nr_open_min = BITS_PER_LONG;
33int 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 */
41static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
42
43static void *alloc_fdmem(unsigned int 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
57static void free_fdmem(void *ptr)
58{
59 is_vmalloc_addr(ptr) ? vfree(ptr) : kfree(ptr);
60}
61
62static void __free_fdtable(struct fdtable *fdt)
63{
64 free_fdmem(fdt->fd);
65 free_fdmem(fdt->open_fds);
66 kfree(fdt);
67}
68
69static 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
87void 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 */
123static 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
142static struct fdtable * alloc_fdtable(unsigned int nr)
143{
144 struct fdtable *fdt;
145 char *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 = (struct file **)data;
176 data = alloc_fdmem(max_t(unsigned int,
177 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
178 if (!data)
179 goto out_arr;
180 fdt->open_fds = (fd_set *)data;
181 data += nr / BITS_PER_BYTE;
182 fdt->close_on_exec = (fd_set *)data;
183 fdt->next = NULL;
184
185 return fdt;
186
187out_arr:
188 free_fdmem(fdt->fd);
189out_fdt:
190 kfree(fdt);
191out:
192 return NULL;
193}
194
195/*
196 * Expand the file descriptor table.
197 * This function will allocate a new fdtable and both fd array and fdset, of
198 * the given size.
199 * Return <0 error code on error; 1 on successful completion.
200 * The files->file_lock should be held on entry, and will be held on exit.
201 */
202static int expand_fdtable(struct files_struct *files, int nr)
203 __releases(files->file_lock)
204 __acquires(files->file_lock)
205{
206 struct fdtable *new_fdt, *cur_fdt;
207
208 spin_unlock(&files->file_lock);
209 new_fdt = alloc_fdtable(nr);
210 spin_lock(&files->file_lock);
211 if (!new_fdt)
212 return -ENOMEM;
213 /*
214 * extremely unlikely race - sysctl_nr_open decreased between the check in
215 * caller and alloc_fdtable(). Cheaper to catch it here...
216 */
217 if (unlikely(new_fdt->max_fds <= nr)) {
218 __free_fdtable(new_fdt);
219 return -EMFILE;
220 }
221 /*
222 * Check again since another task may have expanded the fd table while
223 * we dropped the lock
224 */
225 cur_fdt = files_fdtable(files);
226 if (nr >= cur_fdt->max_fds) {
227 /* Continue as planned */
228 copy_fdtable(new_fdt, cur_fdt);
229 rcu_assign_pointer(files->fdt, new_fdt);
230 if (cur_fdt->max_fds > NR_OPEN_DEFAULT)
231 free_fdtable(cur_fdt);
232 } else {
233 /* Somebody else expanded, so undo our attempt */
234 __free_fdtable(new_fdt);
235 }
236 return 1;
237}
238
239/*
240 * Expand files.
241 * This function will expand the file structures, if the requested size exceeds
242 * the current capacity and there is room for expansion.
243 * Return <0 error code on error; 0 when nothing done; 1 when files were
244 * expanded and execution may have blocked.
245 * The files->file_lock should be held on entry, and will be held on exit.
246 */
247int expand_files(struct files_struct *files, int nr)
248{
249 struct fdtable *fdt;
250
251 fdt = files_fdtable(files);
252
253 /*
254 * N.B. For clone tasks sharing a files structure, this test
255 * will limit the total number of files that can be opened.
256 */
257 if (nr >= rlimit(RLIMIT_NOFILE))
258 return -EMFILE;
259
260 /* Do we need to expand? */
261 if (nr < fdt->max_fds)
262 return 0;
263
264 /* Can we expand? */
265 if (nr >= sysctl_nr_open)
266 return -EMFILE;
267
268 /* All good, so we try */
269 return expand_fdtable(files, nr);
270}
271
272static int count_open_files(struct fdtable *fdt)
273{
274 int size = fdt->max_fds;
275 int i;
276
277 /* Find the last open fd */
278 for (i = size/(8*sizeof(long)); i > 0; ) {
279 if (fdt->open_fds->fds_bits[--i])
280 break;
281 }
282 i = (i+1) * 8 * sizeof(long);
283 return i;
284}
285
286/*
287 * Allocate a new files structure and copy contents from the
288 * passed in files structure.
289 * errorp will be valid only when the returned files_struct is NULL.
290 */
291struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
292{
293 struct files_struct *newf;
294 struct file **old_fds, **new_fds;
295 int open_files, size, i;
296 struct fdtable *old_fdt, *new_fdt;
297
298 *errorp = -ENOMEM;
299 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
300 if (!newf)
301 goto out;
302
303 atomic_set(&newf->count, 1);
304
305 spin_lock_init(&newf->file_lock);
306 newf->next_fd = 0;
307 new_fdt = &newf->fdtab;
308 new_fdt->max_fds = NR_OPEN_DEFAULT;
309 new_fdt->close_on_exec = (fd_set *)&newf->close_on_exec_init;
310 new_fdt->open_fds = (fd_set *)&newf->open_fds_init;
311 new_fdt->fd = &newf->fd_array[0];
312 new_fdt->next = NULL;
313
314 spin_lock(&oldf->file_lock);
315 old_fdt = files_fdtable(oldf);
316 open_files = count_open_files(old_fdt);
317
318 /*
319 * Check whether we need to allocate a larger fd array and fd set.
320 */
321 while (unlikely(open_files > new_fdt->max_fds)) {
322 spin_unlock(&oldf->file_lock);
323
324 if (new_fdt != &newf->fdtab)
325 __free_fdtable(new_fdt);
326
327 new_fdt = alloc_fdtable(open_files - 1);
328 if (!new_fdt) {
329 *errorp = -ENOMEM;
330 goto out_release;
331 }
332
333 /* beyond sysctl_nr_open; nothing to do */
334 if (unlikely(new_fdt->max_fds < open_files)) {
335 __free_fdtable(new_fdt);
336 *errorp = -EMFILE;
337 goto out_release;
338 }
339
340 /*
341 * Reacquire the oldf lock and a pointer to its fd table
342 * who knows it may have a new bigger fd table. We need
343 * the latest pointer.
344 */
345 spin_lock(&oldf->file_lock);
346 old_fdt = files_fdtable(oldf);
347 open_files = count_open_files(old_fdt);
348 }
349
350 old_fds = old_fdt->fd;
351 new_fds = new_fdt->fd;
352
353 memcpy(new_fdt->open_fds->fds_bits,
354 old_fdt->open_fds->fds_bits, open_files/8);
355 memcpy(new_fdt->close_on_exec->fds_bits,
356 old_fdt->close_on_exec->fds_bits, open_files/8);
357
358 for (i = open_files; i != 0; i--) {
359 struct file *f = *old_fds++;
360 if (f) {
361 get_file(f);
362 } else {
363 /*
364 * The fd may be claimed in the fd bitmap but not yet
365 * instantiated in the files array if a sibling thread
366 * is partway through open(). So make sure that this
367 * fd is available to the new process.
368 */
369 FD_CLR(open_files - i, new_fdt->open_fds);
370 }
371 rcu_assign_pointer(*new_fds++, f);
372 }
373 spin_unlock(&oldf->file_lock);
374
375 /* compute the remainder to be cleared */
376 size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
377
378 /* This is long word aligned thus could use a optimized version */
379 memset(new_fds, 0, size);
380
381 if (new_fdt->max_fds > open_files) {
382 int left = (new_fdt->max_fds-open_files)/8;
383 int start = open_files / (8 * sizeof(unsigned long));
384
385 memset(&new_fdt->open_fds->fds_bits[start], 0, left);
386 memset(&new_fdt->close_on_exec->fds_bits[start], 0, left);
387 }
388
389 rcu_assign_pointer(newf->fdt, new_fdt);
390
391 return newf;
392
393out_release:
394 kmem_cache_free(files_cachep, newf);
395out:
396 return NULL;
397}
398
399static void __devinit fdtable_defer_list_init(int cpu)
400{
401 struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
402 spin_lock_init(&fddef->lock);
403 INIT_WORK(&fddef->wq, free_fdtable_work);
404 fddef->next = NULL;
405}
406
407void __init files_defer_init(void)
408{
409 int i;
410 for_each_possible_cpu(i)
411 fdtable_defer_list_init(i);
412 sysctl_nr_open_max = min((size_t)INT_MAX, ~(size_t)0/sizeof(void *)) &
413 -BITS_PER_LONG;
414}
415
416struct files_struct init_files = {
417 .count = ATOMIC_INIT(1),
418 .fdt = &init_files.fdtab,
419 .fdtab = {
420 .max_fds = NR_OPEN_DEFAULT,
421 .fd = &init_files.fd_array[0],
422 .close_on_exec = (fd_set *)&init_files.close_on_exec_init,
423 .open_fds = (fd_set *)&init_files.open_fds_init,
424 },
425 .file_lock = __SPIN_LOCK_UNLOCKED(init_task.file_lock),
426};
427
428/*
429 * allocate a file descriptor, mark it busy.
430 */
431int alloc_fd(unsigned start, unsigned flags)
432{
433 struct files_struct *files = current->files;
434 unsigned int fd;
435 int error;
436 struct fdtable *fdt;
437
438 spin_lock(&files->file_lock);
439repeat:
440 fdt = files_fdtable(files);
441 fd = start;
442 if (fd < files->next_fd)
443 fd = files->next_fd;
444
445 if (fd < fdt->max_fds)
446 fd = find_next_zero_bit(fdt->open_fds->fds_bits,
447 fdt->max_fds, fd);
448
449 error = expand_files(files, fd);
450 if (error < 0)
451 goto out;
452
453 /*
454 * If we needed to expand the fs array we
455 * might have blocked - try again.
456 */
457 if (error)
458 goto repeat;
459
460 if (start <= files->next_fd)
461 files->next_fd = fd + 1;
462
463 FD_SET(fd, fdt->open_fds);
464 if (flags & O_CLOEXEC)
465 FD_SET(fd, fdt->close_on_exec);
466 else
467 FD_CLR(fd, fdt->close_on_exec);
468 error = fd;
469#if 1
470 /* Sanity check */
471 if (rcu_dereference_raw(fdt->fd[fd]) != NULL) {
472 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
473 rcu_assign_pointer(fdt->fd[fd], NULL);
474 }
475#endif
476
477out:
478 spin_unlock(&files->file_lock);
479 return error;
480}
481
482int get_unused_fd(void)
483{
484 return alloc_fd(0, 0);
485}
486EXPORT_SYMBOL(get_unused_fd);
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
26int sysctl_nr_open __read_mostly = 1024*1024;
27int sysctl_nr_open_min = BITS_PER_LONG;
28/* our max() is unusable in constant expressions ;-/ */
29#define __const_max(x, y) ((x) < (y) ? (x) : (y))
30int sysctl_nr_open_max = __const_max(INT_MAX, ~(size_t)0/sizeof(void *)) &
31 -BITS_PER_LONG;
32
33static void *alloc_fdmem(size_t size)
34{
35 /*
36 * Very large allocations can stress page reclaim, so fall back to
37 * vmalloc() if the allocation size will be considered "large" by the VM.
38 */
39 if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) {
40 void *data = kmalloc(size, GFP_KERNEL|__GFP_NOWARN|__GFP_NORETRY);
41 if (data != NULL)
42 return data;
43 }
44 return vmalloc(size);
45}
46
47static void free_fdmem(void *ptr)
48{
49 is_vmalloc_addr(ptr) ? vfree(ptr) : kfree(ptr);
50}
51
52static void __free_fdtable(struct fdtable *fdt)
53{
54 free_fdmem(fdt->fd);
55 free_fdmem(fdt->open_fds);
56 kfree(fdt);
57}
58
59static void free_fdtable_rcu(struct rcu_head *rcu)
60{
61 __free_fdtable(container_of(rcu, struct fdtable, rcu));
62}
63
64/*
65 * Expand the fdset in the files_struct. Called with the files spinlock
66 * held for write.
67 */
68static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
69{
70 unsigned int cpy, set;
71
72 BUG_ON(nfdt->max_fds < ofdt->max_fds);
73
74 cpy = ofdt->max_fds * sizeof(struct file *);
75 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
76 memcpy(nfdt->fd, ofdt->fd, cpy);
77 memset((char *)(nfdt->fd) + cpy, 0, set);
78
79 cpy = ofdt->max_fds / BITS_PER_BYTE;
80 set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
81 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
82 memset((char *)(nfdt->open_fds) + cpy, 0, set);
83 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
84 memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
85}
86
87static struct fdtable * alloc_fdtable(unsigned int nr)
88{
89 struct fdtable *fdt;
90 void *data;
91
92 /*
93 * Figure out how many fds we actually want to support in this fdtable.
94 * Allocation steps are keyed to the size of the fdarray, since it
95 * grows far faster than any of the other dynamic data. We try to fit
96 * the fdarray into comfortable page-tuned chunks: starting at 1024B
97 * and growing in powers of two from there on.
98 */
99 nr /= (1024 / sizeof(struct file *));
100 nr = roundup_pow_of_two(nr + 1);
101 nr *= (1024 / sizeof(struct file *));
102 /*
103 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
104 * had been set lower between the check in expand_files() and here. Deal
105 * with that in caller, it's cheaper that way.
106 *
107 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
108 * bitmaps handling below becomes unpleasant, to put it mildly...
109 */
110 if (unlikely(nr > sysctl_nr_open))
111 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
112
113 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
114 if (!fdt)
115 goto out;
116 fdt->max_fds = nr;
117 data = alloc_fdmem(nr * sizeof(struct file *));
118 if (!data)
119 goto out_fdt;
120 fdt->fd = data;
121
122 data = alloc_fdmem(max_t(size_t,
123 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
124 if (!data)
125 goto out_arr;
126 fdt->open_fds = data;
127 data += nr / BITS_PER_BYTE;
128 fdt->close_on_exec = data;
129
130 return fdt;
131
132out_arr:
133 free_fdmem(fdt->fd);
134out_fdt:
135 kfree(fdt);
136out:
137 return NULL;
138}
139
140/*
141 * Expand the file descriptor table.
142 * This function will allocate a new fdtable and both fd array and fdset, of
143 * the given size.
144 * Return <0 error code on error; 1 on successful completion.
145 * The files->file_lock should be held on entry, and will be held on exit.
146 */
147static int expand_fdtable(struct files_struct *files, int nr)
148 __releases(files->file_lock)
149 __acquires(files->file_lock)
150{
151 struct fdtable *new_fdt, *cur_fdt;
152
153 spin_unlock(&files->file_lock);
154 new_fdt = alloc_fdtable(nr);
155 spin_lock(&files->file_lock);
156 if (!new_fdt)
157 return -ENOMEM;
158 /*
159 * extremely unlikely race - sysctl_nr_open decreased between the check in
160 * caller and alloc_fdtable(). Cheaper to catch it here...
161 */
162 if (unlikely(new_fdt->max_fds <= nr)) {
163 __free_fdtable(new_fdt);
164 return -EMFILE;
165 }
166 /*
167 * Check again since another task may have expanded the fd table while
168 * we dropped the lock
169 */
170 cur_fdt = files_fdtable(files);
171 if (nr >= cur_fdt->max_fds) {
172 /* Continue as planned */
173 copy_fdtable(new_fdt, cur_fdt);
174 rcu_assign_pointer(files->fdt, new_fdt);
175 if (cur_fdt != &files->fdtab)
176 call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
177 } else {
178 /* Somebody else expanded, so undo our attempt */
179 __free_fdtable(new_fdt);
180 }
181 return 1;
182}
183
184/*
185 * Expand files.
186 * This function will expand the file structures, if the requested size exceeds
187 * the current capacity and there is room for expansion.
188 * Return <0 error code on error; 0 when nothing done; 1 when files were
189 * expanded and execution may have blocked.
190 * The files->file_lock should be held on entry, and will be held on exit.
191 */
192static int expand_files(struct files_struct *files, int nr)
193{
194 struct fdtable *fdt;
195
196 fdt = files_fdtable(files);
197
198 /* Do we need to expand? */
199 if (nr < fdt->max_fds)
200 return 0;
201
202 /* Can we expand? */
203 if (nr >= sysctl_nr_open)
204 return -EMFILE;
205
206 /* All good, so we try */
207 return expand_fdtable(files, nr);
208}
209
210static inline void __set_close_on_exec(int fd, struct fdtable *fdt)
211{
212 __set_bit(fd, fdt->close_on_exec);
213}
214
215static inline void __clear_close_on_exec(int fd, struct fdtable *fdt)
216{
217 __clear_bit(fd, fdt->close_on_exec);
218}
219
220static inline void __set_open_fd(int fd, struct fdtable *fdt)
221{
222 __set_bit(fd, fdt->open_fds);
223}
224
225static inline void __clear_open_fd(int fd, struct fdtable *fdt)
226{
227 __clear_bit(fd, fdt->open_fds);
228}
229
230static int count_open_files(struct fdtable *fdt)
231{
232 int size = fdt->max_fds;
233 int i;
234
235 /* Find the last open fd */
236 for (i = size / BITS_PER_LONG; i > 0; ) {
237 if (fdt->open_fds[--i])
238 break;
239 }
240 i = (i + 1) * BITS_PER_LONG;
241 return i;
242}
243
244/*
245 * Allocate a new files structure and copy contents from the
246 * passed in files structure.
247 * errorp will be valid only when the returned files_struct is NULL.
248 */
249struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
250{
251 struct files_struct *newf;
252 struct file **old_fds, **new_fds;
253 int open_files, size, i;
254 struct fdtable *old_fdt, *new_fdt;
255
256 *errorp = -ENOMEM;
257 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
258 if (!newf)
259 goto out;
260
261 atomic_set(&newf->count, 1);
262
263 spin_lock_init(&newf->file_lock);
264 newf->next_fd = 0;
265 new_fdt = &newf->fdtab;
266 new_fdt->max_fds = NR_OPEN_DEFAULT;
267 new_fdt->close_on_exec = newf->close_on_exec_init;
268 new_fdt->open_fds = newf->open_fds_init;
269 new_fdt->fd = &newf->fd_array[0];
270
271 spin_lock(&oldf->file_lock);
272 old_fdt = files_fdtable(oldf);
273 open_files = count_open_files(old_fdt);
274
275 /*
276 * Check whether we need to allocate a larger fd array and fd set.
277 */
278 while (unlikely(open_files > new_fdt->max_fds)) {
279 spin_unlock(&oldf->file_lock);
280
281 if (new_fdt != &newf->fdtab)
282 __free_fdtable(new_fdt);
283
284 new_fdt = alloc_fdtable(open_files - 1);
285 if (!new_fdt) {
286 *errorp = -ENOMEM;
287 goto out_release;
288 }
289
290 /* beyond sysctl_nr_open; nothing to do */
291 if (unlikely(new_fdt->max_fds < open_files)) {
292 __free_fdtable(new_fdt);
293 *errorp = -EMFILE;
294 goto out_release;
295 }
296
297 /*
298 * Reacquire the oldf lock and a pointer to its fd table
299 * who knows it may have a new bigger fd table. We need
300 * the latest pointer.
301 */
302 spin_lock(&oldf->file_lock);
303 old_fdt = files_fdtable(oldf);
304 open_files = count_open_files(old_fdt);
305 }
306
307 old_fds = old_fdt->fd;
308 new_fds = new_fdt->fd;
309
310 memcpy(new_fdt->open_fds, old_fdt->open_fds, open_files / 8);
311 memcpy(new_fdt->close_on_exec, old_fdt->close_on_exec, open_files / 8);
312
313 for (i = open_files; i != 0; i--) {
314 struct file *f = *old_fds++;
315 if (f) {
316 get_file(f);
317 } else {
318 /*
319 * The fd may be claimed in the fd bitmap but not yet
320 * instantiated in the files array if a sibling thread
321 * is partway through open(). So make sure that this
322 * fd is available to the new process.
323 */
324 __clear_open_fd(open_files - i, new_fdt);
325 }
326 rcu_assign_pointer(*new_fds++, f);
327 }
328 spin_unlock(&oldf->file_lock);
329
330 /* compute the remainder to be cleared */
331 size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
332
333 /* This is long word aligned thus could use a optimized version */
334 memset(new_fds, 0, size);
335
336 if (new_fdt->max_fds > open_files) {
337 int left = (new_fdt->max_fds - open_files) / 8;
338 int start = open_files / BITS_PER_LONG;
339
340 memset(&new_fdt->open_fds[start], 0, left);
341 memset(&new_fdt->close_on_exec[start], 0, left);
342 }
343
344 rcu_assign_pointer(newf->fdt, new_fdt);
345
346 return newf;
347
348out_release:
349 kmem_cache_free(files_cachep, newf);
350out:
351 return NULL;
352}
353
354static struct fdtable *close_files(struct files_struct * files)
355{
356 /*
357 * It is safe to dereference the fd table without RCU or
358 * ->file_lock because this is the last reference to the
359 * files structure.
360 */
361 struct fdtable *fdt = rcu_dereference_raw(files->fdt);
362 int i, j = 0;
363
364 for (;;) {
365 unsigned long set;
366 i = j * BITS_PER_LONG;
367 if (i >= fdt->max_fds)
368 break;
369 set = fdt->open_fds[j++];
370 while (set) {
371 if (set & 1) {
372 struct file * file = xchg(&fdt->fd[i], NULL);
373 if (file) {
374 filp_close(file, files);
375 cond_resched();
376 }
377 }
378 i++;
379 set >>= 1;
380 }
381 }
382
383 return fdt;
384}
385
386struct files_struct *get_files_struct(struct task_struct *task)
387{
388 struct files_struct *files;
389
390 task_lock(task);
391 files = task->files;
392 if (files)
393 atomic_inc(&files->count);
394 task_unlock(task);
395
396 return files;
397}
398
399void put_files_struct(struct files_struct *files)
400{
401 if (atomic_dec_and_test(&files->count)) {
402 struct fdtable *fdt = close_files(files);
403
404 /* free the arrays if they are not embedded */
405 if (fdt != &files->fdtab)
406 __free_fdtable(fdt);
407 kmem_cache_free(files_cachep, files);
408 }
409}
410
411void reset_files_struct(struct files_struct *files)
412{
413 struct task_struct *tsk = current;
414 struct files_struct *old;
415
416 old = tsk->files;
417 task_lock(tsk);
418 tsk->files = files;
419 task_unlock(tsk);
420 put_files_struct(old);
421}
422
423void exit_files(struct task_struct *tsk)
424{
425 struct files_struct * files = tsk->files;
426
427 if (files) {
428 task_lock(tsk);
429 tsk->files = NULL;
430 task_unlock(tsk);
431 put_files_struct(files);
432 }
433}
434
435struct files_struct init_files = {
436 .count = ATOMIC_INIT(1),
437 .fdt = &init_files.fdtab,
438 .fdtab = {
439 .max_fds = NR_OPEN_DEFAULT,
440 .fd = &init_files.fd_array[0],
441 .close_on_exec = init_files.close_on_exec_init,
442 .open_fds = init_files.open_fds_init,
443 },
444 .file_lock = __SPIN_LOCK_UNLOCKED(init_files.file_lock),
445};
446
447/*
448 * allocate a file descriptor, mark it busy.
449 */
450int __alloc_fd(struct files_struct *files,
451 unsigned start, unsigned end, unsigned flags)
452{
453 unsigned int fd;
454 int error;
455 struct fdtable *fdt;
456
457 spin_lock(&files->file_lock);
458repeat:
459 fdt = files_fdtable(files);
460 fd = start;
461 if (fd < files->next_fd)
462 fd = files->next_fd;
463
464 if (fd < fdt->max_fds)
465 fd = find_next_zero_bit(fdt->open_fds, fdt->max_fds, fd);
466
467 /*
468 * N.B. For clone tasks sharing a files structure, this test
469 * will limit the total number of files that can be opened.
470 */
471 error = -EMFILE;
472 if (fd >= end)
473 goto out;
474
475 error = expand_files(files, fd);
476 if (error < 0)
477 goto out;
478
479 /*
480 * If we needed to expand the fs array we
481 * might have blocked - try again.
482 */
483 if (error)
484 goto repeat;
485
486 if (start <= files->next_fd)
487 files->next_fd = fd + 1;
488
489 __set_open_fd(fd, fdt);
490 if (flags & O_CLOEXEC)
491 __set_close_on_exec(fd, fdt);
492 else
493 __clear_close_on_exec(fd, fdt);
494 error = fd;
495#if 1
496 /* Sanity check */
497 if (rcu_access_pointer(fdt->fd[fd]) != NULL) {
498 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
499 rcu_assign_pointer(fdt->fd[fd], NULL);
500 }
501#endif
502
503out:
504 spin_unlock(&files->file_lock);
505 return error;
506}
507
508static int alloc_fd(unsigned start, unsigned flags)
509{
510 return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
511}
512
513int get_unused_fd_flags(unsigned flags)
514{
515 return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags);
516}
517EXPORT_SYMBOL(get_unused_fd_flags);
518
519static void __put_unused_fd(struct files_struct *files, unsigned int fd)
520{
521 struct fdtable *fdt = files_fdtable(files);
522 __clear_open_fd(fd, fdt);
523 if (fd < files->next_fd)
524 files->next_fd = fd;
525}
526
527void put_unused_fd(unsigned int fd)
528{
529 struct files_struct *files = current->files;
530 spin_lock(&files->file_lock);
531 __put_unused_fd(files, fd);
532 spin_unlock(&files->file_lock);
533}
534
535EXPORT_SYMBOL(put_unused_fd);
536
537/*
538 * Install a file pointer in the fd array.
539 *
540 * The VFS is full of places where we drop the files lock between
541 * setting the open_fds bitmap and installing the file in the file
542 * array. At any such point, we are vulnerable to a dup2() race
543 * installing a file in the array before us. We need to detect this and
544 * fput() the struct file we are about to overwrite in this case.
545 *
546 * It should never happen - if we allow dup2() do it, _really_ bad things
547 * will follow.
548 *
549 * NOTE: __fd_install() variant is really, really low-level; don't
550 * use it unless you are forced to by truly lousy API shoved down
551 * your throat. 'files' *MUST* be either current->files or obtained
552 * by get_files_struct(current) done by whoever had given it to you,
553 * or really bad things will happen. Normally you want to use
554 * fd_install() instead.
555 */
556
557void __fd_install(struct files_struct *files, unsigned int fd,
558 struct file *file)
559{
560 struct fdtable *fdt;
561 spin_lock(&files->file_lock);
562 fdt = files_fdtable(files);
563 BUG_ON(fdt->fd[fd] != NULL);
564 rcu_assign_pointer(fdt->fd[fd], file);
565 spin_unlock(&files->file_lock);
566}
567
568void fd_install(unsigned int fd, struct file *file)
569{
570 __fd_install(current->files, fd, file);
571}
572
573EXPORT_SYMBOL(fd_install);
574
575/*
576 * The same warnings as for __alloc_fd()/__fd_install() apply here...
577 */
578int __close_fd(struct files_struct *files, unsigned fd)
579{
580 struct file *file;
581 struct fdtable *fdt;
582
583 spin_lock(&files->file_lock);
584 fdt = files_fdtable(files);
585 if (fd >= fdt->max_fds)
586 goto out_unlock;
587 file = fdt->fd[fd];
588 if (!file)
589 goto out_unlock;
590 rcu_assign_pointer(fdt->fd[fd], NULL);
591 __clear_close_on_exec(fd, fdt);
592 __put_unused_fd(files, fd);
593 spin_unlock(&files->file_lock);
594 return filp_close(file, files);
595
596out_unlock:
597 spin_unlock(&files->file_lock);
598 return -EBADF;
599}
600
601void do_close_on_exec(struct files_struct *files)
602{
603 unsigned i;
604 struct fdtable *fdt;
605
606 /* exec unshares first */
607 spin_lock(&files->file_lock);
608 for (i = 0; ; i++) {
609 unsigned long set;
610 unsigned fd = i * BITS_PER_LONG;
611 fdt = files_fdtable(files);
612 if (fd >= fdt->max_fds)
613 break;
614 set = fdt->close_on_exec[i];
615 if (!set)
616 continue;
617 fdt->close_on_exec[i] = 0;
618 for ( ; set ; fd++, set >>= 1) {
619 struct file *file;
620 if (!(set & 1))
621 continue;
622 file = fdt->fd[fd];
623 if (!file)
624 continue;
625 rcu_assign_pointer(fdt->fd[fd], NULL);
626 __put_unused_fd(files, fd);
627 spin_unlock(&files->file_lock);
628 filp_close(file, files);
629 cond_resched();
630 spin_lock(&files->file_lock);
631 }
632
633 }
634 spin_unlock(&files->file_lock);
635}
636
637static struct file *__fget(unsigned int fd, fmode_t mask)
638{
639 struct files_struct *files = current->files;
640 struct file *file;
641
642 rcu_read_lock();
643 file = fcheck_files(files, fd);
644 if (file) {
645 /* File object ref couldn't be taken */
646 if ((file->f_mode & mask) ||
647 !atomic_long_inc_not_zero(&file->f_count))
648 file = NULL;
649 }
650 rcu_read_unlock();
651
652 return file;
653}
654
655struct file *fget(unsigned int fd)
656{
657 return __fget(fd, FMODE_PATH);
658}
659EXPORT_SYMBOL(fget);
660
661struct file *fget_raw(unsigned int fd)
662{
663 return __fget(fd, 0);
664}
665EXPORT_SYMBOL(fget_raw);
666
667/*
668 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
669 *
670 * You can use this instead of fget if you satisfy all of the following
671 * conditions:
672 * 1) You must call fput_light before exiting the syscall and returning control
673 * to userspace (i.e. you cannot remember the returned struct file * after
674 * returning to userspace).
675 * 2) You must not call filp_close on the returned struct file * in between
676 * calls to fget_light and fput_light.
677 * 3) You must not clone the current task in between the calls to fget_light
678 * and fput_light.
679 *
680 * The fput_needed flag returned by fget_light should be passed to the
681 * corresponding fput_light.
682 */
683static unsigned long __fget_light(unsigned int fd, fmode_t mask)
684{
685 struct files_struct *files = current->files;
686 struct file *file;
687
688 if (atomic_read(&files->count) == 1) {
689 file = __fcheck_files(files, fd);
690 if (!file || unlikely(file->f_mode & mask))
691 return 0;
692 return (unsigned long)file;
693 } else {
694 file = __fget(fd, mask);
695 if (!file)
696 return 0;
697 return FDPUT_FPUT | (unsigned long)file;
698 }
699}
700unsigned long __fdget(unsigned int fd)
701{
702 return __fget_light(fd, FMODE_PATH);
703}
704EXPORT_SYMBOL(__fdget);
705
706unsigned long __fdget_raw(unsigned int fd)
707{
708 return __fget_light(fd, 0);
709}
710
711unsigned long __fdget_pos(unsigned int fd)
712{
713 unsigned long v = __fdget(fd);
714 struct file *file = (struct file *)(v & ~3);
715
716 if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
717 if (file_count(file) > 1) {
718 v |= FDPUT_POS_UNLOCK;
719 mutex_lock(&file->f_pos_lock);
720 }
721 }
722 return v;
723}
724
725/*
726 * We only lock f_pos if we have threads or if the file might be
727 * shared with another process. In both cases we'll have an elevated
728 * file count (done either by fdget() or by fork()).
729 */
730
731void set_close_on_exec(unsigned int fd, int flag)
732{
733 struct files_struct *files = current->files;
734 struct fdtable *fdt;
735 spin_lock(&files->file_lock);
736 fdt = files_fdtable(files);
737 if (flag)
738 __set_close_on_exec(fd, fdt);
739 else
740 __clear_close_on_exec(fd, fdt);
741 spin_unlock(&files->file_lock);
742}
743
744bool get_close_on_exec(unsigned int fd)
745{
746 struct files_struct *files = current->files;
747 struct fdtable *fdt;
748 bool res;
749 rcu_read_lock();
750 fdt = files_fdtable(files);
751 res = close_on_exec(fd, fdt);
752 rcu_read_unlock();
753 return res;
754}
755
756static int do_dup2(struct files_struct *files,
757 struct file *file, unsigned fd, unsigned flags)
758{
759 struct file *tofree;
760 struct fdtable *fdt;
761
762 /*
763 * We need to detect attempts to do dup2() over allocated but still
764 * not finished descriptor. NB: OpenBSD avoids that at the price of
765 * extra work in their equivalent of fget() - they insert struct
766 * file immediately after grabbing descriptor, mark it larval if
767 * more work (e.g. actual opening) is needed and make sure that
768 * fget() treats larval files as absent. Potentially interesting,
769 * but while extra work in fget() is trivial, locking implications
770 * and amount of surgery on open()-related paths in VFS are not.
771 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
772 * deadlocks in rather amusing ways, AFAICS. All of that is out of
773 * scope of POSIX or SUS, since neither considers shared descriptor
774 * tables and this condition does not arise without those.
775 */
776 fdt = files_fdtable(files);
777 tofree = fdt->fd[fd];
778 if (!tofree && fd_is_open(fd, fdt))
779 goto Ebusy;
780 get_file(file);
781 rcu_assign_pointer(fdt->fd[fd], file);
782 __set_open_fd(fd, fdt);
783 if (flags & O_CLOEXEC)
784 __set_close_on_exec(fd, fdt);
785 else
786 __clear_close_on_exec(fd, fdt);
787 spin_unlock(&files->file_lock);
788
789 if (tofree)
790 filp_close(tofree, files);
791
792 return fd;
793
794Ebusy:
795 spin_unlock(&files->file_lock);
796 return -EBUSY;
797}
798
799int replace_fd(unsigned fd, struct file *file, unsigned flags)
800{
801 int err;
802 struct files_struct *files = current->files;
803
804 if (!file)
805 return __close_fd(files, fd);
806
807 if (fd >= rlimit(RLIMIT_NOFILE))
808 return -EBADF;
809
810 spin_lock(&files->file_lock);
811 err = expand_files(files, fd);
812 if (unlikely(err < 0))
813 goto out_unlock;
814 return do_dup2(files, file, fd, flags);
815
816out_unlock:
817 spin_unlock(&files->file_lock);
818 return err;
819}
820
821SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
822{
823 int err = -EBADF;
824 struct file *file;
825 struct files_struct *files = current->files;
826
827 if ((flags & ~O_CLOEXEC) != 0)
828 return -EINVAL;
829
830 if (unlikely(oldfd == newfd))
831 return -EINVAL;
832
833 if (newfd >= rlimit(RLIMIT_NOFILE))
834 return -EBADF;
835
836 spin_lock(&files->file_lock);
837 err = expand_files(files, newfd);
838 file = fcheck(oldfd);
839 if (unlikely(!file))
840 goto Ebadf;
841 if (unlikely(err < 0)) {
842 if (err == -EMFILE)
843 goto Ebadf;
844 goto out_unlock;
845 }
846 return do_dup2(files, file, newfd, flags);
847
848Ebadf:
849 err = -EBADF;
850out_unlock:
851 spin_unlock(&files->file_lock);
852 return err;
853}
854
855SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
856{
857 if (unlikely(newfd == oldfd)) { /* corner case */
858 struct files_struct *files = current->files;
859 int retval = oldfd;
860
861 rcu_read_lock();
862 if (!fcheck_files(files, oldfd))
863 retval = -EBADF;
864 rcu_read_unlock();
865 return retval;
866 }
867 return sys_dup3(oldfd, newfd, 0);
868}
869
870SYSCALL_DEFINE1(dup, unsigned int, fildes)
871{
872 int ret = -EBADF;
873 struct file *file = fget_raw(fildes);
874
875 if (file) {
876 ret = get_unused_fd();
877 if (ret >= 0)
878 fd_install(ret, file);
879 else
880 fput(file);
881 }
882 return ret;
883}
884
885int f_dupfd(unsigned int from, struct file *file, unsigned flags)
886{
887 int err;
888 if (from >= rlimit(RLIMIT_NOFILE))
889 return -EINVAL;
890 err = alloc_fd(from, flags);
891 if (err >= 0) {
892 get_file(file);
893 fd_install(err, file);
894 }
895 return err;
896}
897
898int iterate_fd(struct files_struct *files, unsigned n,
899 int (*f)(const void *, struct file *, unsigned),
900 const void *p)
901{
902 struct fdtable *fdt;
903 int res = 0;
904 if (!files)
905 return 0;
906 spin_lock(&files->file_lock);
907 for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
908 struct file *file;
909 file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
910 if (!file)
911 continue;
912 res = f(p, file, n);
913 if (res)
914 break;
915 }
916 spin_unlock(&files->file_lock);
917 return res;
918}
919EXPORT_SYMBOL(iterate_fd);