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/slab.h>
14 #include <linux/vmalloc.h>
15 #include <linux/file.h>
16 #include <linux/fdtable.h>
17 #include <linux/bitops.h>
18 #include <linux/interrupt.h>
19 #include <linux/spinlock.h>
20 #include <linux/rcupdate.h>
21 #include <linux/workqueue.h>
23 struct fdtable_defer {
25 struct work_struct wq;
29 int sysctl_nr_open __read_mostly = 1024*1024;
30 int sysctl_nr_open_min = BITS_PER_LONG;
31 int sysctl_nr_open_max = 1024 * 1024; /* raised later */
34 * We use this list to defer free fdtables that have vmalloced
35 * sets/arrays. By keeping a per-cpu list, we avoid having to embed
36 * the work_struct in fdtable itself which avoids a 64 byte (i386) increase in
37 * this per-task structure.
39 static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
41 static inline void * alloc_fdmem(unsigned int size)
43 if (size <= PAGE_SIZE)
44 return kmalloc(size, GFP_KERNEL);
49 static inline void free_fdarr(struct fdtable *fdt)
51 if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *)))
57 static inline void free_fdset(struct fdtable *fdt)
59 if (fdt->max_fds <= (PAGE_SIZE * BITS_PER_BYTE / 2))
65 static void free_fdtable_work(struct work_struct *work)
67 struct fdtable_defer *f =
68 container_of(work, struct fdtable_defer, wq);
71 spin_lock_bh(&f->lock);
74 spin_unlock_bh(&f->lock);
76 struct fdtable *next = fdt->next;
84 void free_fdtable_rcu(struct rcu_head *rcu)
86 struct fdtable *fdt = container_of(rcu, struct fdtable, rcu);
87 struct fdtable_defer *fddef;
91 if (fdt->max_fds <= NR_OPEN_DEFAULT) {
93 * This fdtable is embedded in the files structure and that
94 * structure itself is getting destroyed.
96 kmem_cache_free(files_cachep,
97 container_of(fdt, struct files_struct, fdtab));
100 if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *))) {
102 kfree(fdt->open_fds);
105 fddef = &get_cpu_var(fdtable_defer_list);
106 spin_lock(&fddef->lock);
107 fdt->next = fddef->next;
109 /* vmallocs are handled from the workqueue context */
110 schedule_work(&fddef->wq);
111 spin_unlock(&fddef->lock);
112 put_cpu_var(fdtable_defer_list);
117 * Expand the fdset in the files_struct. Called with the files spinlock
120 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
122 unsigned int cpy, set;
124 BUG_ON(nfdt->max_fds < ofdt->max_fds);
126 cpy = ofdt->max_fds * sizeof(struct file *);
127 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
128 memcpy(nfdt->fd, ofdt->fd, cpy);
129 memset((char *)(nfdt->fd) + cpy, 0, set);
131 cpy = ofdt->max_fds / BITS_PER_BYTE;
132 set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
133 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
134 memset((char *)(nfdt->open_fds) + cpy, 0, set);
135 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
136 memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
139 static struct fdtable * alloc_fdtable(unsigned int nr)
145 * Figure out how many fds we actually want to support in this fdtable.
146 * Allocation steps are keyed to the size of the fdarray, since it
147 * grows far faster than any of the other dynamic data. We try to fit
148 * the fdarray into comfortable page-tuned chunks: starting at 1024B
149 * and growing in powers of two from there on.
151 nr /= (1024 / sizeof(struct file *));
152 nr = roundup_pow_of_two(nr + 1);
153 nr *= (1024 / sizeof(struct file *));
155 * Note that this can drive nr *below* what we had passed if sysctl_nr_open
156 * had been set lower between the check in expand_files() and here. Deal
157 * with that in caller, it's cheaper that way.
159 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
160 * bitmaps handling below becomes unpleasant, to put it mildly...
162 if (unlikely(nr > sysctl_nr_open))
163 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
165 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
169 data = alloc_fdmem(nr * sizeof(struct file *));
172 fdt->fd = (struct file **)data;
173 data = alloc_fdmem(max_t(unsigned int,
174 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
177 fdt->open_fds = (fd_set *)data;
178 data += nr / BITS_PER_BYTE;
179 fdt->close_on_exec = (fd_set *)data;
180 INIT_RCU_HEAD(&fdt->rcu);
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 >= current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
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 INIT_RCU_HEAD(&new_fdt->rcu);
315 new_fdt->next = NULL;
317 spin_lock(&oldf->file_lock);
318 old_fdt = files_fdtable(oldf);
319 open_files = count_open_files(old_fdt);
322 * Check whether we need to allocate a larger fd array and fd set.
324 while (unlikely(open_files > new_fdt->max_fds)) {
325 spin_unlock(&oldf->file_lock);
327 if (new_fdt != &newf->fdtab) {
333 new_fdt = alloc_fdtable(open_files - 1);
339 /* beyond sysctl_nr_open; nothing to do */
340 if (unlikely(new_fdt->max_fds < open_files)) {
349 * Reacquire the oldf lock and a pointer to its fd table
350 * who knows it may have a new bigger fd table. We need
351 * the latest pointer.
353 spin_lock(&oldf->file_lock);
354 old_fdt = files_fdtable(oldf);
355 open_files = count_open_files(old_fdt);
358 old_fds = old_fdt->fd;
359 new_fds = new_fdt->fd;
361 memcpy(new_fdt->open_fds->fds_bits,
362 old_fdt->open_fds->fds_bits, open_files/8);
363 memcpy(new_fdt->close_on_exec->fds_bits,
364 old_fdt->close_on_exec->fds_bits, open_files/8);
366 for (i = open_files; i != 0; i--) {
367 struct file *f = *old_fds++;
372 * The fd may be claimed in the fd bitmap but not yet
373 * instantiated in the files array if a sibling thread
374 * is partway through open(). So make sure that this
375 * fd is available to the new process.
377 FD_CLR(open_files - i, new_fdt->open_fds);
379 rcu_assign_pointer(*new_fds++, f);
381 spin_unlock(&oldf->file_lock);
383 /* compute the remainder to be cleared */
384 size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
386 /* This is long word aligned thus could use a optimized version */
387 memset(new_fds, 0, size);
389 if (new_fdt->max_fds > open_files) {
390 int left = (new_fdt->max_fds-open_files)/8;
391 int start = open_files / (8 * sizeof(unsigned long));
393 memset(&new_fdt->open_fds->fds_bits[start], 0, left);
394 memset(&new_fdt->close_on_exec->fds_bits[start], 0, left);
397 rcu_assign_pointer(newf->fdt, new_fdt);
402 kmem_cache_free(files_cachep, newf);
407 static void __devinit fdtable_defer_list_init(int cpu)
409 struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
410 spin_lock_init(&fddef->lock);
411 INIT_WORK(&fddef->wq, free_fdtable_work);
415 void __init files_defer_init(void)
418 for_each_possible_cpu(i)
419 fdtable_defer_list_init(i);
420 sysctl_nr_open_max = min((size_t)INT_MAX, ~(size_t)0/sizeof(void *)) &
424 struct files_struct init_files = {
425 .count = ATOMIC_INIT(1),
426 .fdt = &init_files.fdtab,
428 .max_fds = NR_OPEN_DEFAULT,
429 .fd = &init_files.fd_array[0],
430 .close_on_exec = (fd_set *)&init_files.close_on_exec_init,
431 .open_fds = (fd_set *)&init_files.open_fds_init,
432 .rcu = RCU_HEAD_INIT,
434 .file_lock = __SPIN_LOCK_UNLOCKED(init_task.file_lock),
438 * allocate a file descriptor, mark it busy.
440 int alloc_fd(unsigned start, unsigned flags)
442 struct files_struct *files = current->files;
447 spin_lock(&files->file_lock);
449 fdt = files_fdtable(files);
451 if (fd < files->next_fd)
454 if (fd < fdt->max_fds)
455 fd = find_next_zero_bit(fdt->open_fds->fds_bits,
458 error = expand_files(files, fd);
463 * If we needed to expand the fs array we
464 * might have blocked - try again.
469 if (start <= files->next_fd)
470 files->next_fd = fd + 1;
472 FD_SET(fd, fdt->open_fds);
473 if (flags & O_CLOEXEC)
474 FD_SET(fd, fdt->close_on_exec);
476 FD_CLR(fd, fdt->close_on_exec);
480 if (rcu_dereference(fdt->fd[fd]) != NULL) {
481 printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
482 rcu_assign_pointer(fdt->fd[fd], NULL);
487 spin_unlock(&files->file_lock);
491 int get_unused_fd(void)
493 return alloc_fd(0, 0);
495 EXPORT_SYMBOL(get_unused_fd);