blob: 7cb71b99260340fe7302d2a8c5dc4f84f8029aff [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
Al Virofe17f222012-08-21 11:48:11 -04009#include <linux/syscalls.h>
Paul Gortmaker630d9c42011-11-16 23:57:37 -050010#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070011#include <linux/fs.h>
12#include <linux/mm.h>
Andrew Morton6d4831c2011-04-27 15:26:41 -070013#include <linux/mmzone.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/time.h>
Alexey Dobriyand43c36d2009-10-07 17:09:06 +040015#include <linux/sched.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/slab.h>
17#include <linux/vmalloc.h>
18#include <linux/file.h>
Al Viro9f3acc32008-04-24 07:44:08 -040019#include <linux/fdtable.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070020#include <linux/bitops.h>
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070021#include <linux/interrupt.h>
22#include <linux/spinlock.h>
23#include <linux/rcupdate.h>
24#include <linux/workqueue.h>
25
26struct fdtable_defer {
27 spinlock_t lock;
28 struct work_struct wq;
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070029 struct fdtable *next;
30};
31
Eric Dumazet9cfe0152008-02-06 01:37:16 -080032int sysctl_nr_open __read_mostly = 1024*1024;
Al Viroeceea0b2008-05-10 10:08:32 -040033int sysctl_nr_open_min = BITS_PER_LONG;
34int sysctl_nr_open_max = 1024 * 1024; /* raised later */
Eric Dumazet9cfe0152008-02-06 01:37:16 -080035
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070036/*
37 * We use this list to defer free fdtables that have vmalloced
38 * sets/arrays. By keeping a per-cpu list, we avoid having to embed
39 * the work_struct in fdtable itself which avoids a 64 byte (i386) increase in
40 * this per-task structure.
41 */
42static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
David Howells1fd36ad2012-02-16 17:49:54 +000044static void *alloc_fdmem(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -070045{
Andrew Morton6d4831c2011-04-27 15:26:41 -070046 /*
47 * Very large allocations can stress page reclaim, so fall back to
48 * vmalloc() if the allocation size will be considered "large" by the VM.
49 */
50 if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) {
51 void *data = kmalloc(size, GFP_KERNEL|__GFP_NOWARN);
52 if (data != NULL)
53 return data;
54 }
Changli Gaoa892e2d2010-08-10 18:01:35 -070055 return vmalloc(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -070056}
57
Changli Gaoa892e2d2010-08-10 18:01:35 -070058static void free_fdmem(void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -070059{
Changli Gaoa892e2d2010-08-10 18:01:35 -070060 is_vmalloc_addr(ptr) ? vfree(ptr) : kfree(ptr);
Vadim Lobanov5466b452006-12-10 02:21:22 -080061}
62
Changli Gaoa892e2d2010-08-10 18:01:35 -070063static void __free_fdtable(struct fdtable *fdt)
Vadim Lobanov5466b452006-12-10 02:21:22 -080064{
Changli Gaoa892e2d2010-08-10 18:01:35 -070065 free_fdmem(fdt->fd);
66 free_fdmem(fdt->open_fds);
67 kfree(fdt);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070068}
69
David Howells65f27f32006-11-22 14:55:48 +000070static void free_fdtable_work(struct work_struct *work)
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070071{
David Howells65f27f32006-11-22 14:55:48 +000072 struct fdtable_defer *f =
73 container_of(work, struct fdtable_defer, wq);
Dipankar Sarmabadf1662005-09-09 13:04:10 -070074 struct fdtable *fdt;
Linus Torvalds1da177e2005-04-16 15:20:36 -070075
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070076 spin_lock_bh(&f->lock);
77 fdt = f->next;
78 f->next = NULL;
79 spin_unlock_bh(&f->lock);
80 while(fdt) {
81 struct fdtable *next = fdt->next;
Changli Gaoa892e2d2010-08-10 18:01:35 -070082
83 __free_fdtable(fdt);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070084 fdt = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -070085 }
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070086}
87
Al Viro7cf4dc32012-08-15 19:56:12 -040088static void free_fdtable_rcu(struct rcu_head *rcu)
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070089{
90 struct fdtable *fdt = container_of(rcu, struct fdtable, rcu);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070091 struct fdtable_defer *fddef;
92
93 BUG_ON(!fdt);
Al Viro1983e782012-08-15 20:06:36 -040094 BUG_ON(fdt->max_fds <= NR_OPEN_DEFAULT);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070095
Changli Gaoa892e2d2010-08-10 18:01:35 -070096 if (!is_vmalloc_addr(fdt->fd) && !is_vmalloc_addr(fdt->open_fds)) {
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070097 kfree(fdt->fd);
Vadim Lobanov5466b452006-12-10 02:21:22 -080098 kfree(fdt->open_fds);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -070099 kfree(fdt);
100 } else {
101 fddef = &get_cpu_var(fdtable_defer_list);
102 spin_lock(&fddef->lock);
103 fdt->next = fddef->next;
104 fddef->next = fdt;
Tejun Heo593be072006-12-06 20:36:01 -0800105 /* vmallocs are handled from the workqueue context */
106 schedule_work(&fddef->wq);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700107 spin_unlock(&fddef->lock);
108 put_cpu_var(fdtable_defer_list);
109 }
110}
111
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700112/*
113 * Expand the fdset in the files_struct. Called with the files spinlock
114 * held for write.
115 */
Vadim Lobanov5466b452006-12-10 02:21:22 -0800116static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700117{
Vadim Lobanov5466b452006-12-10 02:21:22 -0800118 unsigned int cpy, set;
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700119
Vadim Lobanov5466b452006-12-10 02:21:22 -0800120 BUG_ON(nfdt->max_fds < ofdt->max_fds);
Vadim Lobanov5466b452006-12-10 02:21:22 -0800121
122 cpy = ofdt->max_fds * sizeof(struct file *);
123 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
124 memcpy(nfdt->fd, ofdt->fd, cpy);
125 memset((char *)(nfdt->fd) + cpy, 0, set);
126
127 cpy = ofdt->max_fds / BITS_PER_BYTE;
128 set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
129 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
130 memset((char *)(nfdt->open_fds) + cpy, 0, set);
131 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
132 memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133}
134
Vadim Lobanov5466b452006-12-10 02:21:22 -0800135static struct fdtable * alloc_fdtable(unsigned int nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136{
Vadim Lobanov5466b452006-12-10 02:21:22 -0800137 struct fdtable *fdt;
David Howells1fd36ad2012-02-16 17:49:54 +0000138 void *data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700140 /*
Vadim Lobanov5466b452006-12-10 02:21:22 -0800141 * Figure out how many fds we actually want to support in this fdtable.
142 * Allocation steps are keyed to the size of the fdarray, since it
143 * grows far faster than any of the other dynamic data. We try to fit
144 * the fdarray into comfortable page-tuned chunks: starting at 1024B
145 * and growing in powers of two from there on.
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700146 */
Vadim Lobanov5466b452006-12-10 02:21:22 -0800147 nr /= (1024 / sizeof(struct file *));
148 nr = roundup_pow_of_two(nr + 1);
149 nr *= (1024 / sizeof(struct file *));
Al Viro5c598b32008-04-27 20:04:15 -0400150 /*
151 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
152 * had been set lower between the check in expand_files() and here. Deal
153 * with that in caller, it's cheaper that way.
154 *
155 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
156 * bitmaps handling below becomes unpleasant, to put it mildly...
157 */
158 if (unlikely(nr > sysctl_nr_open))
159 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
Vadim Lobanovbbea9f62006-12-10 02:21:12 -0800160
Vadim Lobanov5466b452006-12-10 02:21:22 -0800161 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
162 if (!fdt)
Vadim Lobanovbbea9f62006-12-10 02:21:12 -0800163 goto out;
Vadim Lobanov5466b452006-12-10 02:21:22 -0800164 fdt->max_fds = nr;
165 data = alloc_fdmem(nr * sizeof(struct file *));
166 if (!data)
167 goto out_fdt;
David Howells1fd36ad2012-02-16 17:49:54 +0000168 fdt->fd = data;
169
170 data = alloc_fdmem(max_t(size_t,
Vadim Lobanov5466b452006-12-10 02:21:22 -0800171 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
172 if (!data)
173 goto out_arr;
David Howells1fd36ad2012-02-16 17:49:54 +0000174 fdt->open_fds = data;
Vadim Lobanov5466b452006-12-10 02:21:22 -0800175 data += nr / BITS_PER_BYTE;
David Howells1fd36ad2012-02-16 17:49:54 +0000176 fdt->close_on_exec = data;
Vadim Lobanov5466b452006-12-10 02:21:22 -0800177 fdt->next = NULL;
178
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700179 return fdt;
Vadim Lobanov5466b452006-12-10 02:21:22 -0800180
181out_arr:
Changli Gaoa892e2d2010-08-10 18:01:35 -0700182 free_fdmem(fdt->fd);
Vadim Lobanov5466b452006-12-10 02:21:22 -0800183out_fdt:
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700184 kfree(fdt);
Vadim Lobanov5466b452006-12-10 02:21:22 -0800185out:
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700186 return NULL;
187}
188
189/*
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700190 * Expand the file descriptor table.
191 * This function will allocate a new fdtable and both fd array and fdset, of
192 * the given size.
193 * Return <0 error code on error; 1 on successful completion.
194 * The files->file_lock should be held on entry, and will be held on exit.
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700195 */
196static int expand_fdtable(struct files_struct *files, int nr)
197 __releases(files->file_lock)
198 __acquires(files->file_lock)
199{
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700200 struct fdtable *new_fdt, *cur_fdt;
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700201
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202 spin_unlock(&files->file_lock);
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700203 new_fdt = alloc_fdtable(nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204 spin_lock(&files->file_lock);
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700205 if (!new_fdt)
206 return -ENOMEM;
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700207 /*
Al Viro5c598b32008-04-27 20:04:15 -0400208 * extremely unlikely race - sysctl_nr_open decreased between the check in
209 * caller and alloc_fdtable(). Cheaper to catch it here...
210 */
211 if (unlikely(new_fdt->max_fds <= nr)) {
Changli Gaoa892e2d2010-08-10 18:01:35 -0700212 __free_fdtable(new_fdt);
Al Viro5c598b32008-04-27 20:04:15 -0400213 return -EMFILE;
214 }
215 /*
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700216 * Check again since another task may have expanded the fd table while
217 * we dropped the lock
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700218 */
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700219 cur_fdt = files_fdtable(files);
Vadim Lobanovbbea9f62006-12-10 02:21:12 -0800220 if (nr >= cur_fdt->max_fds) {
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700221 /* Continue as planned */
222 copy_fdtable(new_fdt, cur_fdt);
223 rcu_assign_pointer(files->fdt, new_fdt);
Vadim Lobanov4fd45812006-12-10 02:21:17 -0800224 if (cur_fdt->max_fds > NR_OPEN_DEFAULT)
Al Viro1983e782012-08-15 20:06:36 -0400225 call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700226 } else {
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700227 /* Somebody else expanded, so undo our attempt */
Changli Gaoa892e2d2010-08-10 18:01:35 -0700228 __free_fdtable(new_fdt);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700229 }
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700230 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231}
232
233/*
234 * Expand files.
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700235 * This function will expand the file structures, if the requested size exceeds
236 * the current capacity and there is room for expansion.
237 * Return <0 error code on error; 0 when nothing done; 1 when files were
238 * expanded and execution may have blocked.
239 * The files->file_lock should be held on entry, and will be held on exit.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240 */
Al Viroad47bd72012-08-21 20:11:34 -0400241static int expand_files(struct files_struct *files, int nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242{
Dipankar Sarmabadf1662005-09-09 13:04:10 -0700243 struct fdtable *fdt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244
Dipankar Sarmabadf1662005-09-09 13:04:10 -0700245 fdt = files_fdtable(files);
Al Viro4e1e0182008-07-26 16:01:20 -0400246
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700247 /* Do we need to expand? */
Vadim Lobanovbbea9f62006-12-10 02:21:12 -0800248 if (nr < fdt->max_fds)
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700249 return 0;
Al Viro4e1e0182008-07-26 16:01:20 -0400250
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700251 /* Can we expand? */
Eric Dumazet9cfe0152008-02-06 01:37:16 -0800252 if (nr >= sysctl_nr_open)
Vadim Lobanov74d392a2006-09-29 02:01:43 -0700253 return -EMFILE;
254
255 /* All good, so we try */
256 return expand_fdtable(files, nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257}
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700258
Al Virob8318b02012-08-21 20:09:42 -0400259static inline void __set_close_on_exec(int fd, struct fdtable *fdt)
260{
261 __set_bit(fd, fdt->close_on_exec);
262}
263
264static inline void __clear_close_on_exec(int fd, struct fdtable *fdt)
265{
266 __clear_bit(fd, fdt->close_on_exec);
267}
268
269static inline void __set_open_fd(int fd, struct fdtable *fdt)
270{
271 __set_bit(fd, fdt->open_fds);
272}
273
274static inline void __clear_open_fd(int fd, struct fdtable *fdt)
275{
276 __clear_bit(fd, fdt->open_fds);
277}
278
Al Viro02afc6262008-05-08 19:42:56 -0400279static int count_open_files(struct fdtable *fdt)
280{
281 int size = fdt->max_fds;
282 int i;
283
284 /* Find the last open fd */
David Howells1fd36ad2012-02-16 17:49:54 +0000285 for (i = size / BITS_PER_LONG; i > 0; ) {
286 if (fdt->open_fds[--i])
Al Viro02afc6262008-05-08 19:42:56 -0400287 break;
288 }
David Howells1fd36ad2012-02-16 17:49:54 +0000289 i = (i + 1) * BITS_PER_LONG;
Al Viro02afc6262008-05-08 19:42:56 -0400290 return i;
291}
292
Al Viro02afc6262008-05-08 19:42:56 -0400293/*
294 * Allocate a new files structure and copy contents from the
295 * passed in files structure.
296 * errorp will be valid only when the returned files_struct is NULL.
297 */
298struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
299{
300 struct files_struct *newf;
301 struct file **old_fds, **new_fds;
302 int open_files, size, i;
303 struct fdtable *old_fdt, *new_fdt;
304
305 *errorp = -ENOMEM;
Al Viroafbec7f2008-05-08 21:11:17 -0400306 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
Al Viro02afc6262008-05-08 19:42:56 -0400307 if (!newf)
308 goto out;
309
Al Viroafbec7f2008-05-08 21:11:17 -0400310 atomic_set(&newf->count, 1);
311
312 spin_lock_init(&newf->file_lock);
313 newf->next_fd = 0;
314 new_fdt = &newf->fdtab;
315 new_fdt->max_fds = NR_OPEN_DEFAULT;
David Howells1fd36ad2012-02-16 17:49:54 +0000316 new_fdt->close_on_exec = newf->close_on_exec_init;
317 new_fdt->open_fds = newf->open_fds_init;
Al Viroafbec7f2008-05-08 21:11:17 -0400318 new_fdt->fd = &newf->fd_array[0];
Al Viroafbec7f2008-05-08 21:11:17 -0400319 new_fdt->next = NULL;
320
Al Viro02afc6262008-05-08 19:42:56 -0400321 spin_lock(&oldf->file_lock);
322 old_fdt = files_fdtable(oldf);
Al Viro02afc6262008-05-08 19:42:56 -0400323 open_files = count_open_files(old_fdt);
324
325 /*
326 * Check whether we need to allocate a larger fd array and fd set.
Al Viro02afc6262008-05-08 19:42:56 -0400327 */
Al Viroadbecb12008-05-08 21:19:42 -0400328 while (unlikely(open_files > new_fdt->max_fds)) {
Al Viro02afc6262008-05-08 19:42:56 -0400329 spin_unlock(&oldf->file_lock);
Al Viro9dec3c42008-05-08 21:02:45 -0400330
Changli Gaoa892e2d2010-08-10 18:01:35 -0700331 if (new_fdt != &newf->fdtab)
332 __free_fdtable(new_fdt);
Al Viroadbecb12008-05-08 21:19:42 -0400333
Al Viro9dec3c42008-05-08 21:02:45 -0400334 new_fdt = alloc_fdtable(open_files - 1);
335 if (!new_fdt) {
336 *errorp = -ENOMEM;
Al Viro02afc6262008-05-08 19:42:56 -0400337 goto out_release;
Al Viro9dec3c42008-05-08 21:02:45 -0400338 }
339
340 /* beyond sysctl_nr_open; nothing to do */
341 if (unlikely(new_fdt->max_fds < open_files)) {
Changli Gaoa892e2d2010-08-10 18:01:35 -0700342 __free_fdtable(new_fdt);
Al Viro9dec3c42008-05-08 21:02:45 -0400343 *errorp = -EMFILE;
344 goto out_release;
345 }
Al Viro9dec3c42008-05-08 21:02:45 -0400346
Al Viro02afc6262008-05-08 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 Viro02afc6262008-05-08 19:42:56 -0400355 }
356
357 old_fds = old_fdt->fd;
358 new_fds = new_fdt->fd;
359
David Howells1fd36ad2012-02-16 17:49:54 +0000360 memcpy(new_fdt->open_fds, old_fdt->open_fds, open_files / 8);
361 memcpy(new_fdt->close_on_exec, old_fdt->close_on_exec, open_files / 8);
Al Viro02afc6262008-05-08 19:42:56 -0400362
363 for (i = open_files; i != 0; i--) {
364 struct file *f = *old_fds++;
365 if (f) {
366 get_file(f);
367 } else {
368 /*
369 * The fd may be claimed in the fd bitmap but not yet
370 * instantiated in the files array if a sibling thread
371 * is partway through open(). So make sure that this
372 * fd is available to the new process.
373 */
David Howells1dce27c2012-02-16 17:49:42 +0000374 __clear_open_fd(open_files - i, new_fdt);
Al Viro02afc6262008-05-08 19:42:56 -0400375 }
376 rcu_assign_pointer(*new_fds++, f);
377 }
378 spin_unlock(&oldf->file_lock);
379
380 /* compute the remainder to be cleared */
381 size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
382
383 /* This is long word aligned thus could use a optimized version */
384 memset(new_fds, 0, size);
385
386 if (new_fdt->max_fds > open_files) {
David Howells1fd36ad2012-02-16 17:49:54 +0000387 int left = (new_fdt->max_fds - open_files) / 8;
388 int start = open_files / BITS_PER_LONG;
Al Viro02afc6262008-05-08 19:42:56 -0400389
David Howells1fd36ad2012-02-16 17:49:54 +0000390 memset(&new_fdt->open_fds[start], 0, left);
391 memset(&new_fdt->close_on_exec[start], 0, left);
Al Viro02afc6262008-05-08 19:42:56 -0400392 }
393
Al Viroafbec7f2008-05-08 21:11:17 -0400394 rcu_assign_pointer(newf->fdt, new_fdt);
395
Al Viro02afc6262008-05-08 19:42:56 -0400396 return newf;
397
398out_release:
399 kmem_cache_free(files_cachep, newf);
400out:
401 return NULL;
402}
403
Al Viro7cf4dc32012-08-15 19:56:12 -0400404static void close_files(struct files_struct * files)
405{
406 int i, j;
407 struct fdtable *fdt;
408
409 j = 0;
410
411 /*
412 * It is safe to dereference the fd table without RCU or
413 * ->file_lock because this is the last reference to the
414 * files structure. But use RCU to shut RCU-lockdep up.
415 */
416 rcu_read_lock();
417 fdt = files_fdtable(files);
418 rcu_read_unlock();
419 for (;;) {
420 unsigned long set;
421 i = j * BITS_PER_LONG;
422 if (i >= fdt->max_fds)
423 break;
424 set = fdt->open_fds[j++];
425 while (set) {
426 if (set & 1) {
427 struct file * file = xchg(&fdt->fd[i], NULL);
428 if (file) {
429 filp_close(file, files);
430 cond_resched();
431 }
432 }
433 i++;
434 set >>= 1;
435 }
436 }
437}
438
439struct files_struct *get_files_struct(struct task_struct *task)
440{
441 struct files_struct *files;
442
443 task_lock(task);
444 files = task->files;
445 if (files)
446 atomic_inc(&files->count);
447 task_unlock(task);
448
449 return files;
450}
451
452void put_files_struct(struct files_struct *files)
453{
454 struct fdtable *fdt;
455
456 if (atomic_dec_and_test(&files->count)) {
457 close_files(files);
Al Virob9e02af2012-08-15 20:00:58 -0400458 /* not really needed, since nobody can see us */
Al Viro7cf4dc32012-08-15 19:56:12 -0400459 rcu_read_lock();
460 fdt = files_fdtable(files);
Al Viro7cf4dc32012-08-15 19:56:12 -0400461 rcu_read_unlock();
Al Virob9e02af2012-08-15 20:00:58 -0400462 /* free the arrays if they are not embedded */
463 if (fdt != &files->fdtab)
464 __free_fdtable(fdt);
465 kmem_cache_free(files_cachep, files);
Al Viro7cf4dc32012-08-15 19:56:12 -0400466 }
467}
468
469void reset_files_struct(struct files_struct *files)
470{
471 struct task_struct *tsk = current;
472 struct files_struct *old;
473
474 old = tsk->files;
475 task_lock(tsk);
476 tsk->files = files;
477 task_unlock(tsk);
478 put_files_struct(old);
479}
480
481void exit_files(struct task_struct *tsk)
482{
483 struct files_struct * files = tsk->files;
484
485 if (files) {
486 task_lock(tsk);
487 tsk->files = NULL;
488 task_unlock(tsk);
489 put_files_struct(files);
490 }
491}
492
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700493static void __devinit fdtable_defer_list_init(int cpu)
494{
495 struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
496 spin_lock_init(&fddef->lock);
David Howells65f27f32006-11-22 14:55:48 +0000497 INIT_WORK(&fddef->wq, free_fdtable_work);
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700498 fddef->next = NULL;
499}
500
501void __init files_defer_init(void)
502{
503 int i;
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -0800504 for_each_possible_cpu(i)
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700505 fdtable_defer_list_init(i);
Al Viroeceea0b2008-05-10 10:08:32 -0400506 sysctl_nr_open_max = min((size_t)INT_MAX, ~(size_t)0/sizeof(void *)) &
507 -BITS_PER_LONG;
Dipankar Sarmaab2af1f2005-09-09 13:04:13 -0700508}
Al Virof52111b2008-05-08 18:19:16 -0400509
510struct files_struct init_files = {
511 .count = ATOMIC_INIT(1),
512 .fdt = &init_files.fdtab,
513 .fdtab = {
514 .max_fds = NR_OPEN_DEFAULT,
515 .fd = &init_files.fd_array[0],
David Howells1fd36ad2012-02-16 17:49:54 +0000516 .close_on_exec = init_files.close_on_exec_init,
517 .open_fds = init_files.open_fds_init,
Al Virof52111b2008-05-08 18:19:16 -0400518 },
519 .file_lock = __SPIN_LOCK_UNLOCKED(init_task.file_lock),
520};
Al Viro1027abe2008-07-30 04:13:04 -0400521
Al Viro864bdb32012-08-22 18:42:10 -0400522void daemonize_descriptors(void)
523{
524 atomic_inc(&init_files.count);
525 reset_files_struct(&init_files);
526}
527
Al Viro1027abe2008-07-30 04:13:04 -0400528/*
529 * allocate a file descriptor, mark it busy.
530 */
Al Virodcfadfa2012-08-12 17:27:30 -0400531int __alloc_fd(struct files_struct *files,
532 unsigned start, unsigned end, unsigned flags)
Al Viro1027abe2008-07-30 04:13:04 -0400533{
Al Viro1027abe2008-07-30 04:13:04 -0400534 unsigned int fd;
535 int error;
536 struct fdtable *fdt;
537
538 spin_lock(&files->file_lock);
539repeat:
540 fdt = files_fdtable(files);
541 fd = start;
542 if (fd < files->next_fd)
543 fd = files->next_fd;
544
545 if (fd < fdt->max_fds)
David Howells1fd36ad2012-02-16 17:49:54 +0000546 fd = find_next_zero_bit(fdt->open_fds, fdt->max_fds, fd);
Al Viro1027abe2008-07-30 04:13:04 -0400547
Al Virof33ff992012-08-12 16:17:59 -0400548 /*
549 * N.B. For clone tasks sharing a files structure, this test
550 * will limit the total number of files that can be opened.
551 */
552 error = -EMFILE;
553 if (fd >= end)
554 goto out;
555
Al Viro1027abe2008-07-30 04:13:04 -0400556 error = expand_files(files, fd);
557 if (error < 0)
558 goto out;
559
560 /*
561 * If we needed to expand the fs array we
562 * might have blocked - try again.
563 */
564 if (error)
565 goto repeat;
566
567 if (start <= files->next_fd)
568 files->next_fd = fd + 1;
569
David Howells1dce27c2012-02-16 17:49:42 +0000570 __set_open_fd(fd, fdt);
Al Viro1027abe2008-07-30 04:13:04 -0400571 if (flags & O_CLOEXEC)
David Howells1dce27c2012-02-16 17:49:42 +0000572 __set_close_on_exec(fd, fdt);
Al Viro1027abe2008-07-30 04:13:04 -0400573 else
David Howells1dce27c2012-02-16 17:49:42 +0000574 __clear_close_on_exec(fd, fdt);
Al Viro1027abe2008-07-30 04:13:04 -0400575 error = fd;
576#if 1
577 /* Sanity check */
Paul E. McKenney7dc52152010-02-22 17:04:52 -0800578 if (rcu_dereference_raw(fdt->fd[fd]) != NULL) {
Al Viro1027abe2008-07-30 04:13:04 -0400579 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
580 rcu_assign_pointer(fdt->fd[fd], NULL);
581 }
582#endif
583
584out:
585 spin_unlock(&files->file_lock);
586 return error;
587}
588
Al Viroad47bd72012-08-21 20:11:34 -0400589static int alloc_fd(unsigned start, unsigned flags)
Al Virodcfadfa2012-08-12 17:27:30 -0400590{
591 return __alloc_fd(current->files, start, rlimit(RLIMIT_NOFILE), flags);
592}
593
Al Viro1a7bd222012-08-12 17:18:05 -0400594int get_unused_fd_flags(unsigned flags)
Al Viro1027abe2008-07-30 04:13:04 -0400595{
Al Virodcfadfa2012-08-12 17:27:30 -0400596 return __alloc_fd(current->files, 0, rlimit(RLIMIT_NOFILE), flags);
Al Viro1027abe2008-07-30 04:13:04 -0400597}
Al Viro1a7bd222012-08-12 17:18:05 -0400598EXPORT_SYMBOL(get_unused_fd_flags);
Al Viro56007ca2012-08-15 21:03:26 -0400599
600static void __put_unused_fd(struct files_struct *files, unsigned int fd)
601{
602 struct fdtable *fdt = files_fdtable(files);
603 __clear_open_fd(fd, fdt);
604 if (fd < files->next_fd)
605 files->next_fd = fd;
606}
607
608void put_unused_fd(unsigned int fd)
609{
610 struct files_struct *files = current->files;
611 spin_lock(&files->file_lock);
612 __put_unused_fd(files, fd);
613 spin_unlock(&files->file_lock);
614}
615
616EXPORT_SYMBOL(put_unused_fd);
617
618/*
619 * Install a file pointer in the fd array.
620 *
621 * The VFS is full of places where we drop the files lock between
622 * setting the open_fds bitmap and installing the file in the file
623 * array. At any such point, we are vulnerable to a dup2() race
624 * installing a file in the array before us. We need to detect this and
625 * fput() the struct file we are about to overwrite in this case.
626 *
627 * It should never happen - if we allow dup2() do it, _really_ bad things
628 * will follow.
Al Virof869e8a2012-08-15 21:06:33 -0400629 *
630 * NOTE: __fd_install() variant is really, really low-level; don't
631 * use it unless you are forced to by truly lousy API shoved down
632 * your throat. 'files' *MUST* be either current->files or obtained
633 * by get_files_struct(current) done by whoever had given it to you,
634 * or really bad things will happen. Normally you want to use
635 * fd_install() instead.
Al Viro56007ca2012-08-15 21:03:26 -0400636 */
637
Al Virof869e8a2012-08-15 21:06:33 -0400638void __fd_install(struct files_struct *files, unsigned int fd,
639 struct file *file)
Al Viro56007ca2012-08-15 21:03:26 -0400640{
Al Viro56007ca2012-08-15 21:03:26 -0400641 struct fdtable *fdt;
642 spin_lock(&files->file_lock);
643 fdt = files_fdtable(files);
644 BUG_ON(fdt->fd[fd] != NULL);
645 rcu_assign_pointer(fdt->fd[fd], file);
646 spin_unlock(&files->file_lock);
647}
648
Al Virof869e8a2012-08-15 21:06:33 -0400649void fd_install(unsigned int fd, struct file *file)
650{
651 __fd_install(current->files, fd, file);
652}
653
Al Viro56007ca2012-08-15 21:03:26 -0400654EXPORT_SYMBOL(fd_install);
Al Viro0ee8cdf2012-08-15 21:12:10 -0400655
Al Viro483ce1d2012-08-19 12:04:24 -0400656/*
657 * The same warnings as for __alloc_fd()/__fd_install() apply here...
658 */
659int __close_fd(struct files_struct *files, unsigned fd)
660{
661 struct file *file;
662 struct fdtable *fdt;
663
664 spin_lock(&files->file_lock);
665 fdt = files_fdtable(files);
666 if (fd >= fdt->max_fds)
667 goto out_unlock;
668 file = fdt->fd[fd];
669 if (!file)
670 goto out_unlock;
671 rcu_assign_pointer(fdt->fd[fd], NULL);
672 __clear_close_on_exec(fd, fdt);
673 __put_unused_fd(files, fd);
674 spin_unlock(&files->file_lock);
675 return filp_close(file, files);
676
677out_unlock:
678 spin_unlock(&files->file_lock);
679 return -EBADF;
680}
681
Al Viro6a6d27d2012-08-21 09:56:33 -0400682void do_close_on_exec(struct files_struct *files)
683{
684 unsigned i;
685 struct fdtable *fdt;
686
687 /* exec unshares first */
Al Viro6a6d27d2012-08-21 09:56:33 -0400688 spin_lock(&files->file_lock);
689 for (i = 0; ; i++) {
690 unsigned long set;
691 unsigned fd = i * BITS_PER_LONG;
692 fdt = files_fdtable(files);
693 if (fd >= fdt->max_fds)
694 break;
695 set = fdt->close_on_exec[i];
696 if (!set)
697 continue;
698 fdt->close_on_exec[i] = 0;
699 for ( ; set ; fd++, set >>= 1) {
700 struct file *file;
701 if (!(set & 1))
702 continue;
703 file = fdt->fd[fd];
704 if (!file)
705 continue;
706 rcu_assign_pointer(fdt->fd[fd], NULL);
707 __put_unused_fd(files, fd);
708 spin_unlock(&files->file_lock);
709 filp_close(file, files);
710 cond_resched();
711 spin_lock(&files->file_lock);
712 }
713
714 }
715 spin_unlock(&files->file_lock);
716}
717
Al Viro0ee8cdf2012-08-15 21:12:10 -0400718struct file *fget(unsigned int fd)
719{
720 struct file *file;
721 struct files_struct *files = current->files;
722
723 rcu_read_lock();
724 file = fcheck_files(files, fd);
725 if (file) {
726 /* File object ref couldn't be taken */
727 if (file->f_mode & FMODE_PATH ||
728 !atomic_long_inc_not_zero(&file->f_count))
729 file = NULL;
730 }
731 rcu_read_unlock();
732
733 return file;
734}
735
736EXPORT_SYMBOL(fget);
737
738struct file *fget_raw(unsigned int fd)
739{
740 struct file *file;
741 struct files_struct *files = current->files;
742
743 rcu_read_lock();
744 file = fcheck_files(files, fd);
745 if (file) {
746 /* File object ref couldn't be taken */
747 if (!atomic_long_inc_not_zero(&file->f_count))
748 file = NULL;
749 }
750 rcu_read_unlock();
751
752 return file;
753}
754
755EXPORT_SYMBOL(fget_raw);
756
757/*
758 * Lightweight file lookup - no refcnt increment if fd table isn't shared.
759 *
760 * You can use this instead of fget if you satisfy all of the following
761 * conditions:
762 * 1) You must call fput_light before exiting the syscall and returning control
763 * to userspace (i.e. you cannot remember the returned struct file * after
764 * returning to userspace).
765 * 2) You must not call filp_close on the returned struct file * in between
766 * calls to fget_light and fput_light.
767 * 3) You must not clone the current task in between the calls to fget_light
768 * and fput_light.
769 *
770 * The fput_needed flag returned by fget_light should be passed to the
771 * corresponding fput_light.
772 */
773struct file *fget_light(unsigned int fd, int *fput_needed)
774{
775 struct file *file;
776 struct files_struct *files = current->files;
777
778 *fput_needed = 0;
779 if (atomic_read(&files->count) == 1) {
780 file = fcheck_files(files, fd);
781 if (file && (file->f_mode & FMODE_PATH))
782 file = NULL;
783 } else {
784 rcu_read_lock();
785 file = fcheck_files(files, fd);
786 if (file) {
787 if (!(file->f_mode & FMODE_PATH) &&
788 atomic_long_inc_not_zero(&file->f_count))
789 *fput_needed = 1;
790 else
791 /* Didn't get the reference, someone's freed */
792 file = NULL;
793 }
794 rcu_read_unlock();
795 }
796
797 return file;
798}
Al Viro4557c662012-08-28 10:19:41 -0400799EXPORT_SYMBOL(fget_light);
Al Viro0ee8cdf2012-08-15 21:12:10 -0400800
801struct file *fget_raw_light(unsigned int fd, int *fput_needed)
802{
803 struct file *file;
804 struct files_struct *files = current->files;
805
806 *fput_needed = 0;
807 if (atomic_read(&files->count) == 1) {
808 file = fcheck_files(files, fd);
809 } else {
810 rcu_read_lock();
811 file = fcheck_files(files, fd);
812 if (file) {
813 if (atomic_long_inc_not_zero(&file->f_count))
814 *fput_needed = 1;
815 else
816 /* Didn't get the reference, someone's freed */
817 file = NULL;
818 }
819 rcu_read_unlock();
820 }
821
822 return file;
823}
Al Virofe17f222012-08-21 11:48:11 -0400824
825void set_close_on_exec(unsigned int fd, int flag)
826{
827 struct files_struct *files = current->files;
828 struct fdtable *fdt;
829 spin_lock(&files->file_lock);
830 fdt = files_fdtable(files);
831 if (flag)
832 __set_close_on_exec(fd, fdt);
833 else
834 __clear_close_on_exec(fd, fdt);
835 spin_unlock(&files->file_lock);
836}
837
838bool get_close_on_exec(unsigned int fd)
839{
840 struct files_struct *files = current->files;
841 struct fdtable *fdt;
842 bool res;
843 rcu_read_lock();
844 fdt = files_fdtable(files);
845 res = close_on_exec(fd, fdt);
846 rcu_read_unlock();
847 return res;
848}
849
Al Viro8280d162012-08-21 12:11:46 -0400850static int do_dup2(struct files_struct *files,
851 struct file *file, unsigned fd, unsigned flags)
852{
853 struct file *tofree;
854 struct fdtable *fdt;
855
856 /*
857 * We need to detect attempts to do dup2() over allocated but still
858 * not finished descriptor. NB: OpenBSD avoids that at the price of
859 * extra work in their equivalent of fget() - they insert struct
860 * file immediately after grabbing descriptor, mark it larval if
861 * more work (e.g. actual opening) is needed and make sure that
862 * fget() treats larval files as absent. Potentially interesting,
863 * but while extra work in fget() is trivial, locking implications
864 * and amount of surgery on open()-related paths in VFS are not.
865 * FreeBSD fails with -EBADF in the same situation, NetBSD "solution"
866 * deadlocks in rather amusing ways, AFAICS. All of that is out of
867 * scope of POSIX or SUS, since neither considers shared descriptor
868 * tables and this condition does not arise without those.
869 */
870 fdt = files_fdtable(files);
871 tofree = fdt->fd[fd];
872 if (!tofree && fd_is_open(fd, fdt))
873 goto Ebusy;
874 get_file(file);
875 rcu_assign_pointer(fdt->fd[fd], file);
876 __set_open_fd(fd, fdt);
877 if (flags & O_CLOEXEC)
878 __set_close_on_exec(fd, fdt);
879 else
880 __clear_close_on_exec(fd, fdt);
881 spin_unlock(&files->file_lock);
882
883 if (tofree)
884 filp_close(tofree, files);
885
886 return fd;
887
888Ebusy:
889 spin_unlock(&files->file_lock);
890 return -EBUSY;
891}
892
893int replace_fd(unsigned fd, struct file *file, unsigned flags)
894{
895 int err;
896 struct files_struct *files = current->files;
897
898 if (!file)
899 return __close_fd(files, fd);
900
901 if (fd >= rlimit(RLIMIT_NOFILE))
Al Viro08f05c42012-10-31 03:37:48 +0000902 return -EBADF;
Al Viro8280d162012-08-21 12:11:46 -0400903
904 spin_lock(&files->file_lock);
905 err = expand_files(files, fd);
906 if (unlikely(err < 0))
907 goto out_unlock;
908 return do_dup2(files, file, fd, flags);
909
910out_unlock:
911 spin_unlock(&files->file_lock);
912 return err;
913}
914
Al Virofe17f222012-08-21 11:48:11 -0400915SYSCALL_DEFINE3(dup3, unsigned int, oldfd, unsigned int, newfd, int, flags)
916{
917 int err = -EBADF;
Al Viro8280d162012-08-21 12:11:46 -0400918 struct file *file;
919 struct files_struct *files = current->files;
Al Virofe17f222012-08-21 11:48:11 -0400920
921 if ((flags & ~O_CLOEXEC) != 0)
922 return -EINVAL;
923
Richard W.M. Jonesaed97642012-10-09 15:27:43 +0100924 if (unlikely(oldfd == newfd))
925 return -EINVAL;
926
Al Virofe17f222012-08-21 11:48:11 -0400927 if (newfd >= rlimit(RLIMIT_NOFILE))
Al Viro08f05c42012-10-31 03:37:48 +0000928 return -EBADF;
Al Virofe17f222012-08-21 11:48:11 -0400929
930 spin_lock(&files->file_lock);
931 err = expand_files(files, newfd);
932 file = fcheck(oldfd);
933 if (unlikely(!file))
934 goto Ebadf;
935 if (unlikely(err < 0)) {
936 if (err == -EMFILE)
937 goto Ebadf;
938 goto out_unlock;
939 }
Al Viro8280d162012-08-21 12:11:46 -0400940 return do_dup2(files, file, newfd, flags);
Al Virofe17f222012-08-21 11:48:11 -0400941
942Ebadf:
943 err = -EBADF;
944out_unlock:
945 spin_unlock(&files->file_lock);
946 return err;
947}
948
949SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
950{
951 if (unlikely(newfd == oldfd)) { /* corner case */
952 struct files_struct *files = current->files;
953 int retval = oldfd;
954
955 rcu_read_lock();
956 if (!fcheck_files(files, oldfd))
957 retval = -EBADF;
958 rcu_read_unlock();
959 return retval;
960 }
961 return sys_dup3(oldfd, newfd, 0);
962}
963
964SYSCALL_DEFINE1(dup, unsigned int, fildes)
965{
966 int ret = -EBADF;
967 struct file *file = fget_raw(fildes);
968
969 if (file) {
970 ret = get_unused_fd();
971 if (ret >= 0)
972 fd_install(ret, file);
973 else
974 fput(file);
975 }
976 return ret;
977}
978
979int f_dupfd(unsigned int from, struct file *file, unsigned flags)
980{
981 int err;
982 if (from >= rlimit(RLIMIT_NOFILE))
983 return -EINVAL;
984 err = alloc_fd(from, flags);
985 if (err >= 0) {
986 get_file(file);
987 fd_install(err, file);
988 }
989 return err;
990}
Al Viroc3c073f2012-08-21 22:32:06 -0400991
992int iterate_fd(struct files_struct *files, unsigned n,
993 int (*f)(const void *, struct file *, unsigned),
994 const void *p)
995{
996 struct fdtable *fdt;
997 struct file *file;
998 int res = 0;
999 if (!files)
1000 return 0;
1001 spin_lock(&files->file_lock);
1002 fdt = files_fdtable(files);
1003 while (!res && n < fdt->max_fds) {
1004 file = rcu_dereference_check_fdtable(files, fdt->fd[n++]);
1005 if (file)
1006 res = f(p, file, n);
1007 }
1008 spin_unlock(&files->file_lock);
1009 return res;
1010}
1011EXPORT_SYMBOL(iterate_fd);