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/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>
25 struct fdtable_defer {
27 struct work_struct wq;
31 int sysctl_nr_open __read_mostly = 1024*1024;
32 int sysctl_nr_open_min = BITS_PER_LONG;
33 int sysctl_nr_open_max = 1024 * 1024; /* raised later */
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.
41 static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
43 static void *alloc_fdmem(unsigned int size)
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.
49 if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) {
50 void *data = kmalloc(size, GFP_KERNEL|__GFP_NOWARN);
57 static void free_fdmem(void *ptr)
59 is_vmalloc_addr(ptr) ? vfree(ptr) : kfree(ptr);
62 static void __free_fdtable(struct fdtable *fdt)
65 free_fdmem(fdt->open_fds);
69 static void free_fdtable_work(struct work_struct *work)
71 struct fdtable_defer *f =
72 container_of(work, struct fdtable_defer, wq);
75 spin_lock_bh(&f->lock);
78 spin_unlock_bh(&f->lock);
80 struct fdtable *next = fdt->next;
87 void free_fdtable_rcu(struct rcu_head *rcu)
89 struct fdtable *fdt = container_of(rcu, struct fdtable, rcu);
90 struct fdtable_defer *fddef;
94 if (fdt->max_fds <= NR_OPEN_DEFAULT) {
96 * This fdtable is embedded in the files structure and that
97 * structure itself is getting destroyed.
99 kmem_cache_free(files_cachep,
100 container_of(fdt, struct files_struct, fdtab));
103 if (!is_vmalloc_addr(fdt->fd) && !is_vmalloc_addr(fdt->open_fds)) {
105 kfree(fdt->open_fds);
108 fddef = &get_cpu_var(fdtable_defer_list);
109 spin_lock(&fddef->lock);
110 fdt->next = fddef->next;
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);
120 * Expand the fdset in the files_struct. Called with the files spinlock
123 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
125 unsigned int cpy, set;
127 BUG_ON(nfdt->max_fds < ofdt->max_fds);
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);
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);
142 static struct fdtable * alloc_fdtable(unsigned int nr)
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.
154 nr /= (1024 / sizeof(struct file *));
155 nr = roundup_pow_of_two(nr + 1);
156 nr *= (1024 / sizeof(struct file *));
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.
162 * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise
163 * bitmaps handling below becomes unpleasant, to put it mildly...
165 if (unlikely(nr > sysctl_nr_open))
166 nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1;
168 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
172 data = alloc_fdmem(nr * sizeof(struct file *));
175 fdt->fd = (struct file **)data;
176 data = alloc_fdmem(max_t(unsigned int,
177 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
180 fdt->open_fds = (fd_set *)data;
181 data += nr / BITS_PER_BYTE;
182 fdt->close_on_exec = (fd_set *)data;
196 * Expand the file descriptor table.
197 * This function will allocate a new fdtable and both fd array and fdset, of
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.
202 static int expand_fdtable(struct files_struct *files, int nr)
203 __releases(files->file_lock)
204 __acquires(files->file_lock)
206 struct fdtable *new_fdt, *cur_fdt;
208 spin_unlock(&files->file_lock);
209 new_fdt = alloc_fdtable(nr);
210 spin_lock(&files->file_lock);
214 * extremely unlikely race - sysctl_nr_open decreased between the check in
215 * caller and alloc_fdtable(). Cheaper to catch it here...
217 if (unlikely(new_fdt->max_fds <= nr)) {
218 __free_fdtable(new_fdt);
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 */
234 __free_fdtable(new_fdt);
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.
247 int expand_files(struct files_struct *files, int nr)
251 fdt = files_fdtable(files);
254 * N.B. For clone tasks sharing a files structure, this test
255 * will limit the total number of files that can be opened.
257 if (nr >= rlimit(RLIMIT_NOFILE))
260 /* Do we need to expand? */
261 if (nr < fdt->max_fds)
265 if (nr >= sysctl_nr_open)
268 /* All good, so we try */
269 return expand_fdtable(files, nr);
272 static int count_open_files(struct fdtable *fdt)
274 int size = fdt->max_fds;
277 /* Find the last open fd */
278 for (i = size/(8*sizeof(long)); i > 0; ) {
279 if (fdt->open_fds->fds_bits[--i])
282 i = (i+1) * 8 * sizeof(long);
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.
291 struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
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;
299 newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
303 atomic_set(&newf->count, 1);
305 spin_lock_init(&newf->file_lock);
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;
314 spin_lock(&oldf->file_lock);
315 old_fdt = files_fdtable(oldf);
316 open_files = count_open_files(old_fdt);
319 * Check whether we need to allocate a larger fd array and fd set.
321 while (unlikely(open_files > new_fdt->max_fds)) {
322 spin_unlock(&oldf->file_lock);
324 if (new_fdt != &newf->fdtab)
325 __free_fdtable(new_fdt);
327 new_fdt = alloc_fdtable(open_files - 1);
333 /* beyond sysctl_nr_open; nothing to do */
334 if (unlikely(new_fdt->max_fds < open_files)) {
335 __free_fdtable(new_fdt);
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.
345 spin_lock(&oldf->file_lock);
346 old_fdt = files_fdtable(oldf);
347 open_files = count_open_files(old_fdt);
350 old_fds = old_fdt->fd;
351 new_fds = new_fdt->fd;
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);
358 for (i = open_files; i != 0; i--) {
359 struct file *f = *old_fds++;
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.
369 FD_CLR(open_files - i, new_fdt->open_fds);
371 rcu_assign_pointer(*new_fds++, f);
373 spin_unlock(&oldf->file_lock);
375 /* compute the remainder to be cleared */
376 size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
378 /* This is long word aligned thus could use a optimized version */
379 memset(new_fds, 0, size);
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));
385 memset(&new_fdt->open_fds->fds_bits[start], 0, left);
386 memset(&new_fdt->close_on_exec->fds_bits[start], 0, left);
389 rcu_assign_pointer(newf->fdt, new_fdt);
394 kmem_cache_free(files_cachep, newf);
399 static void __devinit fdtable_defer_list_init(int cpu)
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);
407 void __init files_defer_init(void)
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 *)) &
416 struct files_struct init_files = {
417 .count = ATOMIC_INIT(1),
418 .fdt = &init_files.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,
425 .file_lock = __SPIN_LOCK_UNLOCKED(init_task.file_lock),
429 * allocate a file descriptor, mark it busy.
431 int alloc_fd(unsigned start, unsigned flags)
433 struct files_struct *files = current->files;
438 spin_lock(&files->file_lock);
440 fdt = files_fdtable(files);
442 if (fd < files->next_fd)
445 if (fd < fdt->max_fds)
446 fd = find_next_zero_bit(fdt->open_fds->fds_bits,
449 error = expand_files(files, fd);
454 * If we needed to expand the fs array we
455 * might have blocked - try again.
460 if (start <= files->next_fd)
461 files->next_fd = fd + 1;
463 FD_SET(fd, fdt->open_fds);
464 if (flags & O_CLOEXEC)
465 FD_SET(fd, fdt->close_on_exec);
467 FD_CLR(fd, fdt->close_on_exec);
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);
478 spin_unlock(&files->file_lock);
482 int get_unused_fd(void)
484 return alloc_fd(0, 0);
486 EXPORT_SYMBOL(get_unused_fd);