4 * Copyright (C) 1998-1999, Stephen Tweedie and Bill Hawes
6 * Manage the dynamic fd arrays in the process files_struct.
11 #include <linux/time.h>
12 #include <linux/slab.h>
13 #include <linux/vmalloc.h>
14 #include <linux/file.h>
15 #include <linux/bitops.h>
16 #include <linux/interrupt.h>
17 #include <linux/spinlock.h>
18 #include <linux/rcupdate.h>
19 #include <linux/workqueue.h>
21 struct fdtable_defer {
23 struct work_struct wq;
24 struct timer_list timer;
29 * We use this list to defer free fdtables that have vmalloced
30 * sets/arrays. By keeping a per-cpu list, we avoid having to embed
31 * the work_struct in fdtable itself which avoids a 64 byte (i386) increase in
32 * this per-task structure.
34 static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
38 * Allocate an fd array, using kmalloc or vmalloc.
39 * Note: the array isn't cleared at allocation time.
41 struct file ** alloc_fd_array(int num)
43 struct file **new_fds;
44 int size = num * sizeof(struct file *);
46 if (size <= PAGE_SIZE)
47 new_fds = (struct file **) kmalloc(size, GFP_KERNEL);
49 new_fds = (struct file **) vmalloc(size);
53 void free_fd_array(struct file **array, int num)
55 int size = num * sizeof(struct file *);
58 printk (KERN_ERR "free_fd_array: array = 0 (num = %d)\n", num);
62 if (num <= NR_OPEN_DEFAULT) /* Don't free the embedded fd array! */
64 else if (size <= PAGE_SIZE)
70 static void __free_fdtable(struct fdtable *fdt)
72 int fdset_size, fdarray_size;
74 fdset_size = fdt->max_fdset / 8;
75 fdarray_size = fdt->max_fds * sizeof(struct file *);
76 free_fdset(fdt->open_fds, fdset_size);
77 free_fdset(fdt->close_on_exec, fdset_size);
78 free_fd_array(fdt->fd, fdarray_size);
82 static void fdtable_timer(unsigned long data)
84 struct fdtable_defer *fddef = (struct fdtable_defer *)data;
86 spin_lock(&fddef->lock);
88 * If someone already emptied the queue return.
92 if (!schedule_work(&fddef->wq))
93 mod_timer(&fddef->timer, 5);
95 spin_unlock(&fddef->lock);
98 static void free_fdtable_work(struct fdtable_defer *f)
102 spin_lock_bh(&f->lock);
105 spin_unlock_bh(&f->lock);
107 struct fdtable *next = fdt->next;
113 static void free_fdtable_rcu(struct rcu_head *rcu)
115 struct fdtable *fdt = container_of(rcu, struct fdtable, rcu);
116 int fdset_size, fdarray_size;
117 struct fdtable_defer *fddef;
120 fdset_size = fdt->max_fdset / 8;
121 fdarray_size = fdt->max_fds * sizeof(struct file *);
123 if (fdt->free_files) {
125 * The this fdtable was embedded in the files structure
126 * and the files structure itself was getting destroyed.
127 * It is now safe to free the files structure.
129 kmem_cache_free(files_cachep, fdt->free_files);
132 if (fdt->max_fdset <= __FD_SETSIZE && fdt->max_fds <= NR_OPEN_DEFAULT) {
134 * The fdtable was embedded
138 if (fdset_size <= PAGE_SIZE && fdarray_size <= PAGE_SIZE) {
139 kfree(fdt->open_fds);
140 kfree(fdt->close_on_exec);
144 fddef = &get_cpu_var(fdtable_defer_list);
145 spin_lock(&fddef->lock);
146 fdt->next = fddef->next;
149 * vmallocs are handled from the workqueue context.
150 * If the per-cpu workqueue is running, then we
151 * defer work scheduling through a timer.
153 if (!schedule_work(&fddef->wq))
154 mod_timer(&fddef->timer, 5);
155 spin_unlock(&fddef->lock);
156 put_cpu_var(fdtable_defer_list);
160 void free_fdtable(struct fdtable *fdt)
162 if (fdt->free_files || fdt->max_fdset > __FD_SETSIZE ||
163 fdt->max_fds > NR_OPEN_DEFAULT)
164 call_rcu(&fdt->rcu, free_fdtable_rcu);
168 * Expand the fdset in the files_struct. Called with the files spinlock
171 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *fdt)
176 BUG_ON(nfdt->max_fdset < fdt->max_fdset);
177 BUG_ON(nfdt->max_fds < fdt->max_fds);
178 /* Copy the existing tables and install the new pointers */
180 i = fdt->max_fdset / (sizeof(unsigned long) * 8);
181 count = (nfdt->max_fdset - fdt->max_fdset) / 8;
184 * Don't copy the entire array if the current fdset is
185 * not yet initialised.
188 memcpy (nfdt->open_fds, fdt->open_fds,
190 memcpy (nfdt->close_on_exec, fdt->close_on_exec,
192 memset (&nfdt->open_fds->fds_bits[i], 0, count);
193 memset (&nfdt->close_on_exec->fds_bits[i], 0, count);
196 /* Don't copy/clear the array if we are creating a new
197 fd array for fork() */
199 memcpy(nfdt->fd, fdt->fd,
200 fdt->max_fds * sizeof(struct file *));
201 /* clear the remainder of the array */
202 memset(&nfdt->fd[fdt->max_fds], 0,
203 (nfdt->max_fds - fdt->max_fds) *
204 sizeof(struct file *));
206 nfdt->next_fd = fdt->next_fd;
210 * Allocate an fdset array, using kmalloc or vmalloc.
211 * Note: the array isn't cleared at allocation time.
213 fd_set * alloc_fdset(int num)
218 if (size <= PAGE_SIZE)
219 new_fdset = (fd_set *) kmalloc(size, GFP_KERNEL);
221 new_fdset = (fd_set *) vmalloc(size);
225 void free_fdset(fd_set *array, int num)
229 if (num <= __FD_SETSIZE) /* Don't free an embedded fdset */
231 else if (size <= PAGE_SIZE)
237 static struct fdtable *alloc_fdtable(int nr)
239 struct fdtable *fdt = NULL;
241 fd_set *new_openset = NULL, *new_execset = NULL;
242 struct file **new_fds;
244 fdt = kmalloc(sizeof(*fdt), GFP_KERNEL);
247 memset(fdt, 0, sizeof(*fdt));
250 /* Expand to the max in easy steps */
252 if (nfds < (PAGE_SIZE * 8))
253 nfds = PAGE_SIZE * 8;
259 } while (nfds <= nr);
261 new_openset = alloc_fdset(nfds);
262 new_execset = alloc_fdset(nfds);
263 if (!new_openset || !new_execset)
265 fdt->open_fds = new_openset;
266 fdt->close_on_exec = new_execset;
267 fdt->max_fdset = nfds;
269 nfds = NR_OPEN_DEFAULT;
271 * Expand to the max in easy steps, and keep expanding it until
272 * we have enough for the requested fd array size.
275 #if NR_OPEN_DEFAULT < 256
280 if (nfds < (PAGE_SIZE / sizeof(struct file *)))
281 nfds = PAGE_SIZE / sizeof(struct file *);
287 } while (nfds <= nr);
288 new_fds = alloc_fd_array(nfds);
293 fdt->free_files = NULL;
297 free_fdset(new_openset, nfds);
299 free_fdset(new_execset, nfds);
305 * Expands the file descriptor table - it will allocate a new fdtable and
306 * both fd array and fdset. It is expected to be called with the
309 static int expand_fdtable(struct files_struct *files, int nr)
310 __releases(files->file_lock)
311 __acquires(files->file_lock)
315 struct fdtable *nfdt = NULL;
317 spin_unlock(&files->file_lock);
318 nfdt = alloc_fdtable(nr);
321 spin_lock(&files->file_lock);
325 spin_lock(&files->file_lock);
326 fdt = files_fdtable(files);
328 * Check again since another task may have expanded the
329 * fd table while we dropped the lock
331 if (nr >= fdt->max_fds || nr >= fdt->max_fdset) {
332 copy_fdtable(nfdt, fdt);
334 /* Somebody expanded while we dropped file_lock */
335 spin_unlock(&files->file_lock);
336 __free_fdtable(nfdt);
337 spin_lock(&files->file_lock);
340 rcu_assign_pointer(files->fdt, nfdt);
348 * Return <0 on error; 0 nothing done; 1 files expanded, we may have blocked.
349 * Should be called with the files->file_lock spinlock held for write.
351 int expand_files(struct files_struct *files, int nr)
356 fdt = files_fdtable(files);
357 if (nr >= fdt->max_fdset || nr >= fdt->max_fds) {
358 if (fdt->max_fdset >= NR_OPEN ||
359 fdt->max_fds >= NR_OPEN || nr >= NR_OPEN) {
364 if ((err = expand_fdtable(files, nr)))
372 static void __devinit fdtable_defer_list_init(int cpu)
374 struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
375 spin_lock_init(&fddef->lock);
376 INIT_WORK(&fddef->wq, (void (*)(void *))free_fdtable_work, fddef);
377 init_timer(&fddef->timer);
378 fddef->timer.data = (unsigned long)fddef;
379 fddef->timer.function = fdtable_timer;
383 void __init files_defer_init(void)
386 /* Really early - can't use for_each_cpu */
387 for (i = 0; i < NR_CPUS; i++)
388 fdtable_defer_list_init(i);