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// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/file.c
4 *
5 * Copyright (C) 1998-1999, Stephen Tweedie and Bill Hawes
6 *
7 * Manage the dynamic fd arrays in the process files_struct.
8 */
9
10#include <linux/syscalls.h>
11#include <linux/export.h>
12#include <linux/fs.h>
13#include <linux/kernel.h>
14#include <linux/mm.h>
15#include <linux/sched/signal.h>
16#include <linux/slab.h>
17#include <linux/file.h>
18#include <linux/fdtable.h>
19#include <linux/bitops.h>
20#include <linux/spinlock.h>
21#include <linux/rcupdate.h>
22#include <linux/close_range.h>
23#include <net/sock.h>
24
25#include "internal.h"
26
27unsigned int sysctl_nr_open __read_mostly = 1024*1024;
28unsigned int sysctl_nr_open_min = BITS_PER_LONG;
29/* our min() is unusable in constant expressions ;-/ */
30#define __const_min(x, y) ((x) < (y) ? (x) : (y))
31unsigned int sysctl_nr_open_max =
32 __const_min(INT_MAX, ~(size_t)0/sizeof(void *)) & -BITS_PER_LONG;
33
34static void __free_fdtable(struct fdtable *fdt)
35{
36 kvfree(fdt->fd);
37 kvfree(fdt->open_fds);
38 kfree(fdt);
39}
40
41static void free_fdtable_rcu(struct rcu_head *rcu)
42{
43 __free_fdtable(container_of(rcu, struct fdtable, rcu));
44}
45
46#define BITBIT_NR(nr) BITS_TO_LONGS(BITS_TO_LONGS(nr))
47#define BITBIT_SIZE(nr) (BITBIT_NR(nr) * sizeof(long))
48
49/*
50 * Copy 'count' fd bits from the old table to the new table and clear the extra
51 * space if any. This does not copy the file pointers. Called with the files
52 * spinlock held for write.
53 */
54static void copy_fd_bitmaps(struct fdtable *nfdt, struct fdtable *ofdt,
55 unsigned int count)
56{
57 unsigned int cpy, set;
58
59 cpy = count / BITS_PER_BYTE;
60 set = (nfdt->max_fds - count) / BITS_PER_BYTE;
61 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
62 memset((char *)nfdt->open_fds + cpy, 0, set);
63 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
64 memset((char *)nfdt->close_on_exec + cpy, 0, set);
65
66 cpy = BITBIT_SIZE(count);
67 set = BITBIT_SIZE(nfdt->max_fds) - cpy;
68 memcpy(nfdt->full_fds_bits, ofdt->full_fds_bits, cpy);
69 memset((char *)nfdt->full_fds_bits + cpy, 0, set);
70}
71
72/*
73 * Copy all file descriptors from the old table to the new, expanded table and
74 * clear the extra space. Called with the files spinlock held for write.
75 */
76static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
77{
78 size_t cpy, set;
79
80 BUG_ON(nfdt->max_fds < ofdt->max_fds);
81
82 cpy = ofdt->max_fds * sizeof(struct file *);
83 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
84 memcpy(nfdt->fd, ofdt->fd, cpy);
85 memset((char *)nfdt->fd + cpy, 0, set);
86
87 copy_fd_bitmaps(nfdt, ofdt, ofdt->max_fds);
88}
89
90/*
91 * Note how the fdtable bitmap allocations very much have to be a multiple of
92 * BITS_PER_LONG. This is not only because we walk those things in chunks of
93 * 'unsigned long' in some places, but simply because that is how the Linux
94 * kernel bitmaps are defined to work: they are not "bits in an array of bytes",
95 * they are very much "bits in an array of unsigned long".
96 *
97 * The ALIGN(nr, BITS_PER_LONG) here is for clarity: since we just multiplied
98 * by that "1024/sizeof(ptr)" before, we already know there are sufficient
99 * clear low bits. Clang seems to realize that, gcc ends up being confused.
100 *
101 * On a 128-bit machine, the ALIGN() would actually matter. In the meantime,
102 * let's consider it documentation (and maybe a test-case for gcc to improve
103 * its code generation ;)
104 */
105static struct fdtable * alloc_fdtable(unsigned int nr)
106{
107 struct fdtable *fdt;
108 void *data;
109
110 /*
111 * Figure out how many fds we actually want to support in this fdtable.
112 * Allocation steps are keyed to the size of the fdarray, since it
113 * grows far faster than any of the other dynamic data. We try to fit
114 * the fdarray into comfortable page-tuned chunks: starting at 1024B
115 * and growing in powers of two from there on.
116 */
117 nr /= (1024 / sizeof(struct file *));
118 nr = roundup_pow_of_two(nr + 1);
119 nr *= (1024 / sizeof(struct file *));
120 nr = ALIGN(nr, BITS_PER_LONG);
121 /*
122 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
123 * had been set lower between the check in expand_files() and here. Deal
124 * with that in caller, it's cheaper that way.
125 *
126 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
127 * bitmaps handling below becomes unpleasant, to put it mildly...
128 */
129 if (unlikely(nr > sysctl_nr_open))
130 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
131
132 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL_ACCOUNT);
133 if (!fdt)
134 goto out;
135 fdt->max_fds = nr;
136 data = kvmalloc_array(nr, sizeof(struct file *), GFP_KERNEL_ACCOUNT);
137 if (!data)
138 goto out_fdt;
139 fdt->fd = data;
140
141 data = kvmalloc(max_t(size_t,
142 2 * nr / BITS_PER_BYTE + BITBIT_SIZE(nr), L1_CACHE_BYTES),
143 GFP_KERNEL_ACCOUNT);
144 if (!data)
145 goto out_arr;
146 fdt->open_fds = data;
147 data += nr / BITS_PER_BYTE;
148 fdt->close_on_exec = data;
149 data += nr / BITS_PER_BYTE;
150 fdt->full_fds_bits = data;
151
152 return fdt;
153
154out_arr:
155 kvfree(fdt->fd);
156out_fdt:
157 kfree(fdt);
158out:
159 return NULL;
160}
161
162/*
163 * Expand the file descriptor table.
164 * This function will allocate a new fdtable and both fd array and fdset, of
165 * the given size.
166 * Return <0 error code on error; 1 on successful completion.
167 * The files->file_lock should be held on entry, and will be held on exit.
168 */
169static int expand_fdtable(struct files_struct *files, unsigned int nr)
170 __releases(files->file_lock)
171 __acquires(files->file_lock)
172{
173 struct fdtable *new_fdt, *cur_fdt;
174
175 spin_unlock(&files->file_lock);
176 new_fdt = alloc_fdtable(nr);
177
178 /* make sure all fd_install() have seen resize_in_progress
179 * or have finished their rcu_read_lock_sched() section.
180 */
181 if (atomic_read(&files->count) > 1)
182 synchronize_rcu();
183
184 spin_lock(&files->file_lock);
185 if (!new_fdt)
186 return -ENOMEM;
187 /*
188 * extremely unlikely race - sysctl_nr_open decreased between the check in
189 * caller and alloc_fdtable(). Cheaper to catch it here...
190 */
191 if (unlikely(new_fdt->max_fds <= nr)) {
192 __free_fdtable(new_fdt);
193 return -EMFILE;
194 }
195 cur_fdt = files_fdtable(files);
196 BUG_ON(nr < cur_fdt->max_fds);
197 copy_fdtable(new_fdt, cur_fdt);
198 rcu_assign_pointer(files->fdt, new_fdt);
199 if (cur_fdt != &files->fdtab)
200 call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
201 /* coupled with smp_rmb() in fd_install() */
202 smp_wmb();
203 return 1;
204}
205
206/*
207 * Expand files.
208 * This function will expand the file structures, if the requested size exceeds
209 * the current capacity and there is room for expansion.
210 * Return <0 error code on error; 0 when nothing done; 1 when files were
211 * expanded and execution may have blocked.
212 * The files->file_lock should be held on entry, and will be held on exit.
213 */
214static int expand_files(struct files_struct *files, unsigned int nr)
215 __releases(files->file_lock)
216 __acquires(files->file_lock)
217{
218 struct fdtable *fdt;
219 int expanded = 0;
220
221repeat:
222 fdt = files_fdtable(files);
223
224 /* Do we need to expand? */
225 if (nr < fdt->max_fds)
226 return expanded;
227
228 /* Can we expand? */
229 if (nr >= sysctl_nr_open)
230 return -EMFILE;
231
232 if (unlikely(files->resize_in_progress)) {
233 spin_unlock(&files->file_lock);
234 expanded = 1;
235 wait_event(files->resize_wait, !files->resize_in_progress);
236 spin_lock(&files->file_lock);
237 goto repeat;
238 }
239
240 /* All good, so we try */
241 files->resize_in_progress = true;
242 expanded = expand_fdtable(files, nr);
243 files->resize_in_progress = false;
244
245 wake_up_all(&files->resize_wait);
246 return expanded;
247}
248
249static inline void __set_close_on_exec(unsigned int fd, struct fdtable *fdt)
250{
251 __set_bit(fd, fdt->close_on_exec);
252}
253
254static inline void __clear_close_on_exec(unsigned int fd, struct fdtable *fdt)
255{
256 if (test_bit(fd, fdt->close_on_exec))
257 __clear_bit(fd, fdt->close_on_exec);
258}
259
260static inline void __set_open_fd(unsigned int fd, struct fdtable *fdt)
261{
262 __set_bit(fd, fdt->open_fds);
263 fd /= BITS_PER_LONG;
264 if (!~fdt->open_fds[fd])
265 __set_bit(fd, fdt->full_fds_bits);
266}
267
268static inline void __clear_open_fd(unsigned int fd, struct fdtable *fdt)
269{
270 __clear_bit(fd, fdt->open_fds);
271 __clear_bit(fd / BITS_PER_LONG, fdt->full_fds_bits);
272}
273
274static unsigned int count_open_files(struct fdtable *fdt)
275{
276 unsigned int size = fdt->max_fds;
277 unsigned int i;
278
279 /* Find the last open fd */
280 for (i = size / BITS_PER_LONG; i > 0; ) {
281 if (fdt->open_fds[--i])
282 break;
283 }
284 i = (i + 1) * BITS_PER_LONG;
285 return i;
286}
287
288/*
289 * Note that a sane fdtable size always has to be a multiple of
290 * BITS_PER_LONG, since we have bitmaps that are sized by this.
291 *
292 * 'max_fds' will normally already be properly aligned, but it
293 * turns out that in the close_range() -> __close_range() ->
294 * unshare_fd() -> dup_fd() -> sane_fdtable_size() we can end
295 * up having a 'max_fds' value that isn't already aligned.
296 *
297 * Rather than make close_range() have to worry about this,
298 * just make that BITS_PER_LONG alignment be part of a sane
299 * fdtable size. Becuase that's really what it is.
300 */
301static unsigned int sane_fdtable_size(struct fdtable *fdt, unsigned int max_fds)
302{
303 unsigned int count;
304
305 count = count_open_files(fdt);
306 if (max_fds < NR_OPEN_DEFAULT)
307 max_fds = NR_OPEN_DEFAULT;
308 return ALIGN(min(count, max_fds), BITS_PER_LONG);
309}
310
311/*
312 * Allocate a new files structure and copy contents from the
313 * passed in files structure.
314 * errorp will be valid only when the returned files_struct is NULL.
315 */
316struct files_struct *dup_fd(struct files_struct *oldf, unsigned int max_fds, int *errorp)
317{
318 struct files_struct *newf;
319 struct file **old_fds, **new_fds;
320 unsigned int open_files, i;
321 struct fdtable *old_fdt, *new_fdt;
322
323 *errorp = -ENOMEM;
324 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
325 if (!newf)
326 goto out;
327
328 atomic_set(&newf->count, 1);
329
330 spin_lock_init(&newf->file_lock);
331 newf->resize_in_progress = false;
332 init_waitqueue_head(&newf->resize_wait);
333 newf->next_fd = 0;
334 new_fdt = &newf->fdtab;
335 new_fdt->max_fds = NR_OPEN_DEFAULT;
336 new_fdt->close_on_exec = newf->close_on_exec_init;
337 new_fdt->open_fds = newf->open_fds_init;
338 new_fdt->full_fds_bits = newf->full_fds_bits_init;
339 new_fdt->fd = &newf->fd_array[0];
340
341 spin_lock(&oldf->file_lock);
342 old_fdt = files_fdtable(oldf);
343 open_files = sane_fdtable_size(old_fdt, max_fds);
344
345 /*
346 * Check whether we need to allocate a larger fd array and fd set.
347 */
348 while (unlikely(open_files > new_fdt->max_fds)) {
349 spin_unlock(&oldf->file_lock);
350
351 if (new_fdt != &newf->fdtab)
352 __free_fdtable(new_fdt);
353
354 new_fdt = alloc_fdtable(open_files - 1);
355 if (!new_fdt) {
356 *errorp = -ENOMEM;
357 goto out_release;
358 }
359
360 /* beyond sysctl_nr_open; nothing to do */
361 if (unlikely(new_fdt->max_fds < open_files)) {
362 __free_fdtable(new_fdt);
363 *errorp = -EMFILE;
364 goto out_release;
365 }
366
367 /*
368 * Reacquire the oldf lock and a pointer to its fd table
369 * who knows it may have a new bigger fd table. We need
370 * the latest pointer.
371 */
372 spin_lock(&oldf->file_lock);
373 old_fdt = files_fdtable(oldf);
374 open_files = sane_fdtable_size(old_fdt, max_fds);
375 }
376
377 copy_fd_bitmaps(new_fdt, old_fdt, open_files);
378
379 old_fds = old_fdt->fd;
380 new_fds = new_fdt->fd;
381
382 for (i = open_files; i != 0; i--) {
383 struct file *f = *old_fds++;
384 if (f) {
385 get_file(f);
386 } else {
387 /*
388 * The fd may be claimed in the fd bitmap but not yet
389 * instantiated in the files array if a sibling thread
390 * is partway through open(). So make sure that this
391 * fd is available to the new process.
392 */
393 __clear_open_fd(open_files - i, new_fdt);
394 }
395 rcu_assign_pointer(*new_fds++, f);
396 }
397 spin_unlock(&oldf->file_lock);
398
399 /* clear the remainder */
400 memset(new_fds, 0, (new_fdt->max_fds - open_files) * sizeof(struct file *));
401
402 rcu_assign_pointer(newf->fdt, new_fdt);
403
404 return newf;
405
406out_release:
407 kmem_cache_free(files_cachep, newf);
408out:
409 return NULL;
410}
411
412static struct fdtable *close_files(struct files_struct * files)
413{
414 /*
415 * It is safe to dereference the fd table without RCU or
416 * ->file_lock because this is the last reference to the
417 * files structure.
418 */
419 struct fdtable *fdt = rcu_dereference_raw(files->fdt);
420 unsigned int i, j = 0;
421
422 for (;;) {
423 unsigned long set;
424 i = j * BITS_PER_LONG;
425 if (i >= fdt->max_fds)
426 break;
427 set = fdt->open_fds[j++];
428 while (set) {
429 if (set & 1) {
430 struct file * file = xchg(&fdt->fd[i], NULL);
431 if (file) {
432 filp_close(file, files);
433 cond_resched();
434 }
435 }
436 i++;
437 set >>= 1;
438 }
439 }
440
441 return fdt;
442}
443
444void put_files_struct(struct files_struct *files)
445{
446 if (atomic_dec_and_test(&files->count)) {
447 struct fdtable *fdt = close_files(files);
448
449 /* free the arrays if they are not embedded */
450 if (fdt != &files->fdtab)
451 __free_fdtable(fdt);
452 kmem_cache_free(files_cachep, files);
453 }
454}
455
456void exit_files(struct task_struct *tsk)
457{
458 struct files_struct * files = tsk->files;
459
460 if (files) {
461 task_lock(tsk);
462 tsk->files = NULL;
463 task_unlock(tsk);
464 put_files_struct(files);
465 }
466}
467
468struct files_struct init_files = {
469 .count = ATOMIC_INIT(1),
470 .fdt = &init_files.fdtab,
471 .fdtab = {
472 .max_fds = NR_OPEN_DEFAULT,
473 .fd = &init_files.fd_array[0],
474 .close_on_exec = init_files.close_on_exec_init,
475 .open_fds = init_files.open_fds_init,
476 .full_fds_bits = init_files.full_fds_bits_init,
477 },
478 .file_lock = __SPIN_LOCK_UNLOCKED(init_files.file_lock),
479 .resize_wait = __WAIT_QUEUE_HEAD_INITIALIZER(init_files.resize_wait),
480};
481
482static unsigned int find_next_fd(struct fdtable *fdt, unsigned int start)
483{
484 unsigned int maxfd = fdt->max_fds;
485 unsigned int maxbit = maxfd / BITS_PER_LONG;
486 unsigned int bitbit = start / BITS_PER_LONG;
487
488 bitbit = find_next_zero_bit(fdt->full_fds_bits, maxbit, bitbit) * BITS_PER_LONG;
489 if (bitbit > maxfd)
490 return maxfd;
491 if (bitbit > start)
492 start = bitbit;
493 return find_next_zero_bit(fdt->open_fds, maxfd, start);
494}
495
496/*
497 * allocate a file descriptor, mark it busy.
498 */
499static int alloc_fd(unsigned start, unsigned end, unsigned flags)
500{
501 struct files_struct *files = current->files;
502 unsigned int fd;
503 int error;
504 struct fdtable *fdt;
505
506 spin_lock(&files->file_lock);
507repeat:
508 fdt = files_fdtable(files);
509 fd = start;
510 if (fd < files->next_fd)
511 fd = files->next_fd;
512
513 if (fd < fdt->max_fds)
514 fd = find_next_fd(fdt, fd);
515
516 /*
517 * N.B. For clone tasks sharing a files structure, this test
518 * will limit the total number of files that can be opened.
519 */
520 error = -EMFILE;
521 if (fd >= end)
522 goto out;
523
524 error = expand_files(files, fd);
525 if (error < 0)
526 goto out;
527
528 /*
529 * If we needed to expand the fs array we
530 * might have blocked - try again.
531 */
532 if (error)
533 goto repeat;
534
535 if (start <= files->next_fd)
536 files->next_fd = fd + 1;
537
538 __set_open_fd(fd, fdt);
539 if (flags & O_CLOEXEC)
540 __set_close_on_exec(fd, fdt);
541 else
542 __clear_close_on_exec(fd, fdt);
543 error = fd;
544#if 1
545 /* Sanity check */
546 if (rcu_access_pointer(fdt->fd[fd]) != NULL) {
547 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
548 rcu_assign_pointer(fdt->fd[fd], NULL);
549 }
550#endif
551
552out:
553 spin_unlock(&files->file_lock);
554 return error;
555}
556
557int __get_unused_fd_flags(unsigned flags, unsigned long nofile)
558{
559 return alloc_fd(0, nofile, flags);
560}
561
562int get_unused_fd_flags(unsigned flags)
563{
564 return __get_unused_fd_flags(flags, rlimit(RLIMIT_NOFILE));
565}
566EXPORT_SYMBOL(get_unused_fd_flags);
567
568static void __put_unused_fd(struct files_struct *files, unsigned int fd)
569{
570 struct fdtable *fdt = files_fdtable(files);
571 __clear_open_fd(fd, fdt);
572 if (fd < files->next_fd)
573 files->next_fd = fd;
574}
575
576void put_unused_fd(unsigned int fd)
577{
578 struct files_struct *files = current->files;
579 spin_lock(&files->file_lock);
580 __put_unused_fd(files, fd);
581 spin_unlock(&files->file_lock);
582}
583
584EXPORT_SYMBOL(put_unused_fd);
585
586/*
587 * Install a file pointer in the fd array.
588 *
589 * The VFS is full of places where we drop the files lock between
590 * setting the open_fds bitmap and installing the file in the file
591 * array. At any such point, we are vulnerable to a dup2() race
592 * installing a file in the array before us. We need to detect this and
593 * fput() the struct file we are about to overwrite in this case.
594 *
595 * It should never happen - if we allow dup2() do it, _really_ bad things
596 * will follow.
597 *
598 * This consumes the "file" refcount, so callers should treat it
599 * as if they had called fput(file).
600 */
601
602void fd_install(unsigned int fd, struct file *file)
603{
604 struct files_struct *files = current->files;
605 struct fdtable *fdt;
606
607 rcu_read_lock_sched();
608
609 if (unlikely(files->resize_in_progress)) {
610 rcu_read_unlock_sched();
611 spin_lock(&files->file_lock);
612 fdt = files_fdtable(files);
613 BUG_ON(fdt->fd[fd] != NULL);
614 rcu_assign_pointer(fdt->fd[fd], file);
615 spin_unlock(&files->file_lock);
616 return;
617 }
618 /* coupled with smp_wmb() in expand_fdtable() */
619 smp_rmb();
620 fdt = rcu_dereference_sched(files->fdt);
621 BUG_ON(fdt->fd[fd] != NULL);
622 rcu_assign_pointer(fdt->fd[fd], file);
623 rcu_read_unlock_sched();
624}
625
626EXPORT_SYMBOL(fd_install);
627
628/**
629 * pick_file - return file associatd with fd
630 * @files: file struct to retrieve file from
631 * @fd: file descriptor to retrieve file for
632 *
633 * Context: files_lock must be held.
634 *
635 * Returns: The file associated with @fd (NULL if @fd is not open)
636 */
637static struct file *pick_file(struct files_struct *files, unsigned fd)
638{
639 struct fdtable *fdt = files_fdtable(files);
640 struct file *file;
641
642 if (fd >= fdt->max_fds)
643 return NULL;
644
645 file = fdt->fd[fd];
646 if (file) {
647 rcu_assign_pointer(fdt->fd[fd], NULL);
648 __put_unused_fd(files, fd);
649 }
650 return file;
651}
652
653int close_fd(unsigned fd)
654{
655 struct files_struct *files = current->files;
656 struct file *file;
657
658 spin_lock(&files->file_lock);
659 file = pick_file(files, fd);
660 spin_unlock(&files->file_lock);
661 if (!file)
662 return -EBADF;
663
664 return filp_close(file, files);
665}
666EXPORT_SYMBOL(close_fd); /* for ksys_close() */
667
668/**
669 * last_fd - return last valid index into fd table
670 * @cur_fds: files struct
671 *
672 * Context: Either rcu read lock or files_lock must be held.
673 *
674 * Returns: Last valid index into fdtable.
675 */
676static inline unsigned last_fd(struct fdtable *fdt)
677{
678 return fdt->max_fds - 1;
679}
680
681static inline void __range_cloexec(struct files_struct *cur_fds,
682 unsigned int fd, unsigned int max_fd)
683{
684 struct fdtable *fdt;
685
686 /* make sure we're using the correct maximum value */
687 spin_lock(&cur_fds->file_lock);
688 fdt = files_fdtable(cur_fds);
689 max_fd = min(last_fd(fdt), max_fd);
690 if (fd <= max_fd)
691 bitmap_set(fdt->close_on_exec, fd, max_fd - fd + 1);
692 spin_unlock(&cur_fds->file_lock);
693}
694
695static inline void __range_close(struct files_struct *cur_fds, unsigned int fd,
696 unsigned int max_fd)
697{
698 unsigned n;
699
700 rcu_read_lock();
701 n = last_fd(files_fdtable(cur_fds));
702 rcu_read_unlock();
703 max_fd = min(max_fd, n);
704
705 while (fd <= max_fd) {
706 struct file *file;
707
708 spin_lock(&cur_fds->file_lock);
709 file = pick_file(cur_fds, fd++);
710 spin_unlock(&cur_fds->file_lock);
711
712 if (file) {
713 /* found a valid file to close */
714 filp_close(file, cur_fds);
715 cond_resched();
716 }
717 }
718}
719
720/**
721 * __close_range() - Close all file descriptors in a given range.
722 *
723 * @fd: starting file descriptor to close
724 * @max_fd: last file descriptor to close
725 *
726 * This closes a range of file descriptors. All file descriptors
727 * from @fd up to and including @max_fd are closed.
728 */
729int __close_range(unsigned fd, unsigned max_fd, unsigned int flags)
730{
731 struct task_struct *me = current;
732 struct files_struct *cur_fds = me->files, *fds = NULL;
733
734 if (flags & ~(CLOSE_RANGE_UNSHARE | CLOSE_RANGE_CLOEXEC))
735 return -EINVAL;
736
737 if (fd > max_fd)
738 return -EINVAL;
739
740 if (flags & CLOSE_RANGE_UNSHARE) {
741 int ret;
742 unsigned int max_unshare_fds = NR_OPEN_MAX;
743
744 /*
745 * If the caller requested all fds to be made cloexec we always
746 * copy all of the file descriptors since they still want to
747 * use them.
748 */
749 if (!(flags & CLOSE_RANGE_CLOEXEC)) {
750 /*
751 * If the requested range is greater than the current
752 * maximum, we're closing everything so only copy all
753 * file descriptors beneath the lowest file descriptor.
754 */
755 rcu_read_lock();
756 if (max_fd >= last_fd(files_fdtable(cur_fds)))
757 max_unshare_fds = fd;
758 rcu_read_unlock();
759 }
760
761 ret = unshare_fd(CLONE_FILES, max_unshare_fds, &fds);
762 if (ret)
763 return ret;
764
765 /*
766 * We used to share our file descriptor table, and have now
767 * created a private one, make sure we're using it below.
768 */
769 if (fds)
770 swap(cur_fds, fds);
771 }
772
773 if (flags & CLOSE_RANGE_CLOEXEC)
774 __range_cloexec(cur_fds, fd, max_fd);
775 else
776 __range_close(cur_fds, fd, max_fd);
777
778 if (fds) {
779 /*
780 * We're done closing the files we were supposed to. Time to install
781 * the new file descriptor table and drop the old one.
782 */
783 task_lock(me);
784 me->files = cur_fds;
785 task_unlock(me);
786 put_files_struct(fds);
787 }
788
789 return 0;
790}
791
792/*
793 * See close_fd_get_file() below, this variant assumes current->files->file_lock
794 * is held.
795 */
796struct file *__close_fd_get_file(unsigned int fd)
797{
798 return pick_file(current->files, fd);
799}
800
801/*
802 * variant of close_fd that gets a ref on the file for later fput.
803 * The caller must ensure that filp_close() called on the file.
804 */
805struct file *close_fd_get_file(unsigned int fd)
806{
807 struct files_struct *files = current->files;
808 struct file *file;
809
810 spin_lock(&files->file_lock);
811 file = pick_file(files, fd);
812 spin_unlock(&files->file_lock);
813
814 return file;
815}
816
817void do_close_on_exec(struct files_struct *files)
818{
819 unsigned i;
820 struct fdtable *fdt;
821
822 /* exec unshares first */
823 spin_lock(&files->file_lock);
824 for (i = 0; ; i++) {
825 unsigned long set;
826 unsigned fd = i * BITS_PER_LONG;
827 fdt = files_fdtable(files);
828 if (fd >= fdt->max_fds)
829 break;
830 set = fdt->close_on_exec[i];
831 if (!set)
832 continue;
833 fdt->close_on_exec[i] = 0;
834 for ( ; set ; fd++, set >>= 1) {
835 struct file *file;
836 if (!(set & 1))
837 continue;
838 file = fdt->fd[fd];
839 if (!file)
840 continue;
841 rcu_assign_pointer(fdt->fd[fd], NULL);
842 __put_unused_fd(files, fd);
843 spin_unlock(&files->file_lock);
844 filp_close(file, files);
845 cond_resched();
846 spin_lock(&files->file_lock);
847 }
848
849 }
850 spin_unlock(&files->file_lock);
851}
852
853static inline struct file *__fget_files_rcu(struct files_struct *files,
854 unsigned int fd, fmode_t mask)
855{
856 for (;;) {
857 struct file *file;
858 struct fdtable *fdt = rcu_dereference_raw(files->fdt);
859 struct file __rcu **fdentry;
860
861 if (unlikely(fd >= fdt->max_fds))
862 return NULL;
863
864 fdentry = fdt->fd + array_index_nospec(fd, fdt->max_fds);
865 file = rcu_dereference_raw(*fdentry);
866 if (unlikely(!file))
867 return NULL;
868
869 if (unlikely(file->f_mode & mask))
870 return NULL;
871
872 /*
873 * Ok, we have a file pointer. However, because we do
874 * this all locklessly under RCU, we may be racing with
875 * that file being closed.
876 *
877 * Such a race can take two forms:
878 *
879 * (a) the file ref already went down to zero,
880 * and get_file_rcu() fails. Just try again:
881 */
882 if (unlikely(!get_file_rcu(file)))
883 continue;
884
885 /*
886 * (b) the file table entry has changed under us.
887 * Note that we don't need to re-check the 'fdt->fd'
888 * pointer having changed, because it always goes
889 * hand-in-hand with 'fdt'.
890 *
891 * If so, we need to put our ref and try again.
892 */
893 if (unlikely(rcu_dereference_raw(files->fdt) != fdt) ||
894 unlikely(rcu_dereference_raw(*fdentry) != file)) {
895 fput(file);
896 continue;
897 }
898
899 /*
900 * Ok, we have a ref to the file, and checked that it
901 * still exists.
902 */
903 return file;
904 }
905}
906
907static struct file *__fget_files(struct files_struct *files, unsigned int fd,
908 fmode_t mask)
909{
910 struct file *file;
911
912 rcu_read_lock();
913 file = __fget_files_rcu(files, fd, mask);
914 rcu_read_unlock();
915
916 return file;
917}
918
919static inline struct file *__fget(unsigned int fd, fmode_t mask)
920{
921 return __fget_files(current->files, fd, mask);
922}
923
924struct file *fget(unsigned int fd)
925{
926 return __fget(fd, FMODE_PATH);
927}
928EXPORT_SYMBOL(fget);
929
930struct file *fget_raw(unsigned int fd)
931{
932 return __fget(fd, 0);
933}
934EXPORT_SYMBOL(fget_raw);
935
936struct file *fget_task(struct task_struct *task, unsigned int fd)
937{
938 struct file *file = NULL;
939
940 task_lock(task);
941 if (task->files)
942 file = __fget_files(task->files, fd, 0);
943 task_unlock(task);
944
945 return file;
946}
947
948struct file *task_lookup_fd_rcu(struct task_struct *task, unsigned int fd)
949{
950 /* Must be called with rcu_read_lock held */
951 struct files_struct *files;
952 struct file *file = NULL;
953
954 task_lock(task);
955 files = task->files;
956 if (files)
957 file = files_lookup_fd_rcu(files, fd);
958 task_unlock(task);
959
960 return file;
961}
962
963struct file *task_lookup_next_fd_rcu(struct task_struct *task, unsigned int *ret_fd)
964{
965 /* Must be called with rcu_read_lock held */
966 struct files_struct *files;
967 unsigned int fd = *ret_fd;
968 struct file *file = NULL;
969
970 task_lock(task);
971 files = task->files;
972 if (files) {
973 for (; fd < files_fdtable(files)->max_fds; fd++) {
974 file = files_lookup_fd_rcu(files, fd);
975 if (file)
976 break;
977 }
978 }
979 task_unlock(task);
980 *ret_fd = fd;
981 return file;
982}
983EXPORT_SYMBOL(task_lookup_next_fd_rcu);
984
985/*
986 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
987 *
988 * You can use this instead of fget if you satisfy all of the following
989 * conditions:
990 * 1) You must call fput_light before exiting the syscall and returning control
991 * to userspace (i.e. you cannot remember the returned struct file * after
992 * returning to userspace).
993 * 2) You must not call filp_close on the returned struct file * in between
994 * calls to fget_light and fput_light.
995 * 3) You must not clone the current task in between the calls to fget_light
996 * and fput_light.
997 *
998 * The fput_needed flag returned by fget_light should be passed to the
999 * corresponding fput_light.
1000 */
1001static unsigned long __fget_light(unsigned int fd, fmode_t mask)
1002{
1003 struct files_struct *files = current->files;
1004 struct file *file;
1005
1006 /*
1007 * If another thread is concurrently calling close_fd() followed
1008 * by put_files_struct(), we must not observe the old table
1009 * entry combined with the new refcount - otherwise we could
1010 * return a file that is concurrently being freed.
1011 *
1012 * atomic_read_acquire() pairs with atomic_dec_and_test() in
1013 * put_files_struct().
1014 */
1015 if (atomic_read_acquire(&files->count) == 1) {
1016 file = files_lookup_fd_raw(files, fd);
1017 if (!file || unlikely(file->f_mode & mask))
1018 return 0;
1019 return (unsigned long)file;
1020 } else {
1021 file = __fget(fd, mask);
1022 if (!file)
1023 return 0;
1024 return FDPUT_FPUT | (unsigned long)file;
1025 }
1026}
1027unsigned long __fdget(unsigned int fd)
1028{
1029 return __fget_light(fd, FMODE_PATH);
1030}
1031EXPORT_SYMBOL(__fdget);
1032
1033unsigned long __fdget_raw(unsigned int fd)
1034{
1035 return __fget_light(fd, 0);
1036}
1037
1038unsigned long __fdget_pos(unsigned int fd)
1039{
1040 unsigned long v = __fdget(fd);
1041 struct file *file = (struct file *)(v & ~3);
1042
1043 if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
1044 if (file_count(file) > 1) {
1045 v |= FDPUT_POS_UNLOCK;
1046 mutex_lock(&file->f_pos_lock);
1047 }
1048 }
1049 return v;
1050}
1051
1052void __f_unlock_pos(struct file *f)
1053{
1054 mutex_unlock(&f->f_pos_lock);
1055}
1056
1057/*
1058 * We only lock f_pos if we have threads or if the file might be
1059 * shared with another process. In both cases we'll have an elevated
1060 * file count (done either by fdget() or by fork()).
1061 */
1062
1063void set_close_on_exec(unsigned int fd, int flag)
1064{
1065 struct files_struct *files = current->files;
1066 struct fdtable *fdt;
1067 spin_lock(&files->file_lock);
1068 fdt = files_fdtable(files);
1069 if (flag)
1070 __set_close_on_exec(fd, fdt);
1071 else
1072 __clear_close_on_exec(fd, fdt);
1073 spin_unlock(&files->file_lock);
1074}
1075
1076bool get_close_on_exec(unsigned int fd)
1077{
1078 struct files_struct *files = current->files;
1079 struct fdtable *fdt;
1080 bool res;
1081 rcu_read_lock();
1082 fdt = files_fdtable(files);
1083 res = close_on_exec(fd, fdt);
1084 rcu_read_unlock();
1085 return res;
1086}
1087
1088static int do_dup2(struct files_struct *files,
1089 struct file *file, unsigned fd, unsigned flags)
1090__releases(&files->file_lock)
1091{
1092 struct file *tofree;
1093 struct fdtable *fdt;
1094
1095 /*
1096 * We need to detect attempts to do dup2() over allocated but still
1097 * not finished descriptor. NB: OpenBSD avoids that at the price of
1098 * extra work in their equivalent of fget() - they insert struct
1099 * file immediately after grabbing descriptor, mark it larval if
1100 * more work (e.g. actual opening) is needed and make sure that
1101 * fget() treats larval files as absent. Potentially interesting,
1102 * but while extra work in fget() is trivial, locking implications
1103 * and amount of surgery on open()-related paths in VFS are not.
1104 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
1105 * deadlocks in rather amusing ways, AFAICS. All of that is out of
1106 * scope of POSIX or SUS, since neither considers shared descriptor
1107 * tables and this condition does not arise without those.
1108 */
1109 fdt = files_fdtable(files);
1110 tofree = fdt->fd[fd];
1111 if (!tofree && fd_is_open(fd, fdt))
1112 goto Ebusy;
1113 get_file(file);
1114 rcu_assign_pointer(fdt->fd[fd], file);
1115 __set_open_fd(fd, fdt);
1116 if (flags & O_CLOEXEC)
1117 __set_close_on_exec(fd, fdt);
1118 else
1119 __clear_close_on_exec(fd, fdt);
1120 spin_unlock(&files->file_lock);
1121
1122 if (tofree)
1123 filp_close(tofree, files);
1124
1125 return fd;
1126
1127Ebusy:
1128 spin_unlock(&files->file_lock);
1129 return -EBUSY;
1130}
1131
1132int replace_fd(unsigned fd, struct file *file, unsigned flags)
1133{
1134 int err;
1135 struct files_struct *files = current->files;
1136
1137 if (!file)
1138 return close_fd(fd);
1139
1140 if (fd >= rlimit(RLIMIT_NOFILE))
1141 return -EBADF;
1142
1143 spin_lock(&files->file_lock);
1144 err = expand_files(files, fd);
1145 if (unlikely(err < 0))
1146 goto out_unlock;
1147 return do_dup2(files, file, fd, flags);
1148
1149out_unlock:
1150 spin_unlock(&files->file_lock);
1151 return err;
1152}
1153
1154/**
1155 * __receive_fd() - Install received file into file descriptor table
1156 * @file: struct file that was received from another process
1157 * @ufd: __user pointer to write new fd number to
1158 * @o_flags: the O_* flags to apply to the new fd entry
1159 *
1160 * Installs a received file into the file descriptor table, with appropriate
1161 * checks and count updates. Optionally writes the fd number to userspace, if
1162 * @ufd is non-NULL.
1163 *
1164 * This helper handles its own reference counting of the incoming
1165 * struct file.
1166 *
1167 * Returns newly install fd or -ve on error.
1168 */
1169int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags)
1170{
1171 int new_fd;
1172 int error;
1173
1174 error = security_file_receive(file);
1175 if (error)
1176 return error;
1177
1178 new_fd = get_unused_fd_flags(o_flags);
1179 if (new_fd < 0)
1180 return new_fd;
1181
1182 if (ufd) {
1183 error = put_user(new_fd, ufd);
1184 if (error) {
1185 put_unused_fd(new_fd);
1186 return error;
1187 }
1188 }
1189
1190 fd_install(new_fd, get_file(file));
1191 __receive_sock(file);
1192 return new_fd;
1193}
1194
1195int receive_fd_replace(int new_fd, struct file *file, unsigned int o_flags)
1196{
1197 int error;
1198
1199 error = security_file_receive(file);
1200 if (error)
1201 return error;
1202 error = replace_fd(new_fd, file, o_flags);
1203 if (error)
1204 return error;
1205 __receive_sock(file);
1206 return new_fd;
1207}
1208
1209int receive_fd(struct file *file, unsigned int o_flags)
1210{
1211 return __receive_fd(file, NULL, o_flags);
1212}
1213EXPORT_SYMBOL_GPL(receive_fd);
1214
1215static int ksys_dup3(unsigned int oldfd, unsigned int newfd, int flags)
1216{
1217 int err = -EBADF;
1218 struct file *file;
1219 struct files_struct *files = current->files;
1220
1221 if ((flags & ~O_CLOEXEC) != 0)
1222 return -EINVAL;
1223
1224 if (unlikely(oldfd == newfd))
1225 return -EINVAL;
1226
1227 if (newfd >= rlimit(RLIMIT_NOFILE))
1228 return -EBADF;
1229
1230 spin_lock(&files->file_lock);
1231 err = expand_files(files, newfd);
1232 file = files_lookup_fd_locked(files, oldfd);
1233 if (unlikely(!file))
1234 goto Ebadf;
1235 if (unlikely(err < 0)) {
1236 if (err == -EMFILE)
1237 goto Ebadf;
1238 goto out_unlock;
1239 }
1240 return do_dup2(files, file, newfd, flags);
1241
1242Ebadf:
1243 err = -EBADF;
1244out_unlock:
1245 spin_unlock(&files->file_lock);
1246 return err;
1247}
1248
1249SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
1250{
1251 return ksys_dup3(oldfd, newfd, flags);
1252}
1253
1254SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
1255{
1256 if (unlikely(newfd == oldfd)) { /* corner case */
1257 struct files_struct *files = current->files;
1258 int retval = oldfd;
1259
1260 rcu_read_lock();
1261 if (!files_lookup_fd_rcu(files, oldfd))
1262 retval = -EBADF;
1263 rcu_read_unlock();
1264 return retval;
1265 }
1266 return ksys_dup3(oldfd, newfd, 0);
1267}
1268
1269SYSCALL_DEFINE1(dup, unsigned int, fildes)
1270{
1271 int ret = -EBADF;
1272 struct file *file = fget_raw(fildes);
1273
1274 if (file) {
1275 ret = get_unused_fd_flags(0);
1276 if (ret >= 0)
1277 fd_install(ret, file);
1278 else
1279 fput(file);
1280 }
1281 return ret;
1282}
1283
1284int f_dupfd(unsigned int from, struct file *file, unsigned flags)
1285{
1286 unsigned long nofile = rlimit(RLIMIT_NOFILE);
1287 int err;
1288 if (from >= nofile)
1289 return -EINVAL;
1290 err = alloc_fd(from, nofile, flags);
1291 if (err >= 0) {
1292 get_file(file);
1293 fd_install(err, file);
1294 }
1295 return err;
1296}
1297
1298int iterate_fd(struct files_struct *files, unsigned n,
1299 int (*f)(const void *, struct file *, unsigned),
1300 const void *p)
1301{
1302 struct fdtable *fdt;
1303 int res = 0;
1304 if (!files)
1305 return 0;
1306 spin_lock(&files->file_lock);
1307 for (fdt = files_fdtable(files); n < fdt->max_fds; n++) {
1308 struct file *file;
1309 file = rcu_dereference_check_fdtable(files, fdt->fd[n]);
1310 if (!file)
1311 continue;
1312 res = f(p, file, n);
1313 if (res)
1314 break;
1315 }
1316 spin_unlock(&files->file_lock);
1317 return res;
1318}
1319EXPORT_SYMBOL(iterate_fd);