4 * Copyright (C) 1998-1999, Stephen Tweedie and Bill Hawes
6 * Manage the dynamic fd arrays in the process files_struct.
9 #include <linux/module.h>
12 #include <linux/time.h>
13 #include <linux/sched.h>
14 #include <linux/slab.h>
15 #include <linux/vmalloc.h>
16 #include <linux/file.h>
17 #include <linux/fdtable.h>
18 #include <linux/bitops.h>
19 #include <linux/interrupt.h>
20 #include <linux/spinlock.h>
21 #include <linux/rcupdate.h>
22 #include <linux/workqueue.h>
24 struct fdtable_defer {
26 struct work_struct wq;
30 int sysctl_nr_open __read_mostly = 1024*1024;
31 int sysctl_nr_open_min = BITS_PER_LONG;
32 int sysctl_nr_open_max = 1024 * 1024; /* raised later */
35 * We use this list to defer free fdtables that have vmalloced
36 * sets/arrays. By keeping a per-cpu list, we avoid having to embed
37 * the work_struct in fdtable itself which avoids a 64 byte (i386) increase in
38 * this per-task structure.
40 static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
42 static inline void * alloc_fdmem(unsigned int size)
44 if (size <= PAGE_SIZE)
45 return kmalloc(size, GFP_KERNEL);
50 static inline void free_fdarr(struct fdtable *fdt)
52 if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *)))
58 static inline void free_fdset(struct fdtable *fdt)
60 if (fdt->max_fds <= (PAGE_SIZE * BITS_PER_BYTE / 2))
66 static void free_fdtable_work(struct work_struct *work)
68 struct fdtable_defer *f =
69 container_of(work, struct fdtable_defer, wq);
72 spin_lock_bh(&f->lock);
75 spin_unlock_bh(&f->lock);
77 struct fdtable *next = fdt->next;
85 void free_fdtable_rcu(struct rcu_head *rcu)
87 struct fdtable *fdt = container_of(rcu, struct fdtable, rcu);
88 struct fdtable_defer *fddef;
92 if (fdt->max_fds <= NR_OPEN_DEFAULT) {
94 * This fdtable is embedded in the files structure and that
95 * structure itself is getting destroyed.
97 kmem_cache_free(files_cachep,
98 container_of(fdt, struct files_struct, fdtab));
101 if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *))) {
103 kfree(fdt->open_fds);
106 fddef = &get_cpu_var(fdtable_defer_list);
107 spin_lock(&fddef->lock);
108 fdt->next = fddef->next;
110 /* vmallocs are handled from the workqueue context */
111 schedule_work(&fddef->wq);
112 spin_unlock(&fddef->lock);
113 put_cpu_var(fdtable_defer_list);
118 * Expand the fdset in the files_struct. Called with the files spinlock
121 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
123 unsigned int cpy, set;
125 BUG_ON(nfdt->max_fds < ofdt->max_fds);
127 cpy = ofdt->max_fds * sizeof(struct file *);
128 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
129 memcpy(nfdt->fd, ofdt->fd, cpy);
130 memset((char *)(nfdt->fd) + cpy, 0, set);
132 cpy = ofdt->max_fds / BITS_PER_BYTE;
133 set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
134 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
135 memset((char *)(nfdt->open_fds) + cpy, 0, set);
136 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
137 memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
140 static struct fdtable * alloc_fdtable(unsigned int nr)
146 * Figure out how many fds we actually want to support in this fdtable.
147 * Allocation steps are keyed to the size of the fdarray, since it
148 * grows far faster than any of the other dynamic data. We try to fit
149 * the fdarray into comfortable page-tuned chunks: starting at 1024B
150 * and growing in powers of two from there on.
152 nr /= (1024 / sizeof(struct file *));
153 nr = roundup_pow_of_two(nr + 1);
154 nr *= (1024 / sizeof(struct file *));
156 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
157 * had been set lower between the check in expand_files() and here. Deal
158 * with that in caller, it's cheaper that way.
160 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
161 * bitmaps handling below becomes unpleasant, to put it mildly...
163 if (unlikely(nr > sysctl_nr_open))
164 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
166 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
170 data = alloc_fdmem(nr * sizeof(struct file *));
173 fdt->fd = (struct file **)data;
174 data = alloc_fdmem(max_t(unsigned int,
175 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
178 fdt->open_fds = (fd_set *)data;
179 data += nr / BITS_PER_BYTE;
180 fdt->close_on_exec = (fd_set *)data;
194 * Expand the file descriptor table.
195 * This function will allocate a new fdtable and both fd array and fdset, of
197 * Return <0 error code on error; 1 on successful completion.
198 * The files->file_lock should be held on entry, and will be held on exit.
200 static int expand_fdtable(struct files_struct *files, int nr)
201 __releases(files->file_lock)
202 __acquires(files->file_lock)
204 struct fdtable *new_fdt, *cur_fdt;
206 spin_unlock(&files->file_lock);
207 new_fdt = alloc_fdtable(nr);
208 spin_lock(&files->file_lock);
212 * extremely unlikely race - sysctl_nr_open decreased between the check in
213 * caller and alloc_fdtable(). Cheaper to catch it here...
215 if (unlikely(new_fdt->max_fds <= nr)) {
222 * Check again since another task may have expanded the fd table while
223 * we dropped the lock
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);
233 /* Somebody else expanded, so undo our attempt */
243 * This function will expand the file structures, if the requested size exceeds
244 * the current capacity and there is room for expansion.
245 * Return <0 error code on error; 0 when nothing done; 1 when files were
246 * expanded and execution may have blocked.
247 * The files->file_lock should be held on entry, and will be held on exit.
249 int expand_files(struct files_struct *files, int nr)
253 fdt = files_fdtable(files);
256 * N.B. For clone tasks sharing a files structure, this test
257 * will limit the total number of files that can be opened.
259 if (nr >= rlimit(RLIMIT_NOFILE))
262 /* Do we need to expand? */
263 if (nr < fdt->max_fds)
267 if (nr >= sysctl_nr_open)
270 /* All good, so we try */
271 return expand_fdtable(files, nr);
274 static int count_open_files(struct fdtable *fdt)
276 int size = fdt->max_fds;
279 /* Find the last open fd */
280 for (i = size/(8*sizeof(long)); i > 0; ) {
281 if (fdt->open_fds->fds_bits[--i])
284 i = (i+1) * 8 * sizeof(long);
289 * Allocate a new files structure and copy contents from the
290 * passed in files structure.
291 * errorp will be valid only when the returned files_struct is NULL.
293 struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
295 struct files_struct *newf;
296 struct file **old_fds, **new_fds;
297 int open_files, size, i;
298 struct fdtable *old_fdt, *new_fdt;
301 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
305 atomic_set(&newf->count, 1);
307 spin_lock_init(&newf->file_lock);
309 new_fdt = &newf->fdtab;
310 new_fdt->max_fds = NR_OPEN_DEFAULT;
311 new_fdt->close_on_exec = (fd_set *)&newf->close_on_exec_init;
312 new_fdt->open_fds = (fd_set *)&newf->open_fds_init;
313 new_fdt->fd = &newf->fd_array[0];
314 new_fdt->next = NULL;
316 spin_lock(&oldf->file_lock);
317 old_fdt = files_fdtable(oldf);
318 open_files = count_open_files(old_fdt);
321 * Check whether we need to allocate a larger fd array and fd set.
323 while (unlikely(open_files > new_fdt->max_fds)) {
324 spin_unlock(&oldf->file_lock);
326 if (new_fdt != &newf->fdtab) {
332 new_fdt = alloc_fdtable(open_files - 1);
338 /* beyond sysctl_nr_open; nothing to do */
339 if (unlikely(new_fdt->max_fds < open_files)) {
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.
352 spin_lock(&oldf->file_lock);
353 old_fdt = files_fdtable(oldf);
354 open_files = count_open_files(old_fdt);
357 old_fds = old_fdt->fd;
358 new_fds = new_fdt->fd;
360 memcpy(new_fdt->open_fds->fds_bits,
361 old_fdt->open_fds->fds_bits, open_files/8);
362 memcpy(new_fdt->close_on_exec->fds_bits,
363 old_fdt->close_on_exec->fds_bits, open_files/8);
365 for (i = open_files; i != 0; i--) {
366 struct file *f = *old_fds++;
371 * The fd may be claimed in the fd bitmap but not yet
372 * instantiated in the files array if a sibling thread
373 * is partway through open(). So make sure that this
374 * fd is available to the new process.
376 FD_CLR(open_files - i, new_fdt->open_fds);
378 rcu_assign_pointer(*new_fds++, f);
380 spin_unlock(&oldf->file_lock);
382 /* compute the remainder to be cleared */
383 size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
385 /* This is long word aligned thus could use a optimized version */
386 memset(new_fds, 0, size);
388 if (new_fdt->max_fds > open_files) {
389 int left = (new_fdt->max_fds-open_files)/8;
390 int start = open_files / (8 * sizeof(unsigned long));
392 memset(&new_fdt->open_fds->fds_bits[start], 0, left);
393 memset(&new_fdt->close_on_exec->fds_bits[start], 0, left);
396 rcu_assign_pointer(newf->fdt, new_fdt);
401 kmem_cache_free(files_cachep, newf);
406 static void __devinit fdtable_defer_list_init(int cpu)
408 struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
409 spin_lock_init(&fddef->lock);
410 INIT_WORK(&fddef->wq, free_fdtable_work);
414 void __init files_defer_init(void)
417 for_each_possible_cpu(i)
418 fdtable_defer_list_init(i);
419 sysctl_nr_open_max = min((size_t)INT_MAX, ~(size_t)0/sizeof(void *)) &
423 struct files_struct init_files = {
424 .count = ATOMIC_INIT(1),
425 .fdt = &init_files.fdtab,
427 .max_fds = NR_OPEN_DEFAULT,
428 .fd = &init_files.fd_array[0],
429 .close_on_exec = (fd_set *)&init_files.close_on_exec_init,
430 .open_fds = (fd_set *)&init_files.open_fds_init,
432 .file_lock = __SPIN_LOCK_UNLOCKED(init_task.file_lock),
436 * allocate a file descriptor, mark it busy.
438 int alloc_fd(unsigned start, unsigned flags)
440 struct files_struct *files = current->files;
445 spin_lock(&files->file_lock);
447 fdt = files_fdtable(files);
449 if (fd < files->next_fd)
452 if (fd < fdt->max_fds)
453 fd = find_next_zero_bit(fdt->open_fds->fds_bits,
456 error = expand_files(files, fd);
461 * If we needed to expand the fs array we
462 * might have blocked - try again.
467 if (start <= files->next_fd)
468 files->next_fd = fd + 1;
470 FD_SET(fd, fdt->open_fds);
471 if (flags & O_CLOEXEC)
472 FD_SET(fd, fdt->close_on_exec);
474 FD_CLR(fd, fdt->close_on_exec);
478 if (rcu_dereference_raw(fdt->fd[fd]) != NULL) {
479 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
480 rcu_assign_pointer(fdt->fd[fd], NULL);
485 spin_unlock(&files->file_lock);
489 int get_unused_fd(void)
491 return alloc_fd(0, 0);
493 EXPORT_SYMBOL(get_unused_fd);