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;
27 int sysctl_nr_open __read_mostly = 1024*1024;
30 * We use this list to defer free fdtables that have vmalloced
31 * sets/arrays. By keeping a per-cpu list, we avoid having to embed
32 * the work_struct in fdtable itself which avoids a 64 byte (i386) increase in
33 * this per-task structure.
35 static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
37 static inline void * alloc_fdmem(unsigned int size)
39 if (size <= PAGE_SIZE)
40 return kmalloc(size, GFP_KERNEL);
45 static inline void free_fdarr(struct fdtable *fdt)
47 if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *)))
53 static inline void free_fdset(struct fdtable *fdt)
55 if (fdt->max_fds <= (PAGE_SIZE * BITS_PER_BYTE / 2))
61 static void free_fdtable_work(struct work_struct *work)
63 struct fdtable_defer *f =
64 container_of(work, struct fdtable_defer, wq);
67 spin_lock_bh(&f->lock);
70 spin_unlock_bh(&f->lock);
72 struct fdtable *next = fdt->next;
80 void free_fdtable_rcu(struct rcu_head *rcu)
82 struct fdtable *fdt = container_of(rcu, struct fdtable, rcu);
83 struct fdtable_defer *fddef;
87 if (fdt->max_fds <= NR_OPEN_DEFAULT) {
89 * This fdtable is embedded in the files structure and that
90 * structure itself is getting destroyed.
92 kmem_cache_free(files_cachep,
93 container_of(fdt, struct files_struct, fdtab));
96 if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *))) {
101 fddef = &get_cpu_var(fdtable_defer_list);
102 spin_lock(&fddef->lock);
103 fdt->next = fddef->next;
105 /* vmallocs are handled from the workqueue context */
106 schedule_work(&fddef->wq);
107 spin_unlock(&fddef->lock);
108 put_cpu_var(fdtable_defer_list);
113 * Expand the fdset in the files_struct. Called with the files spinlock
116 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
118 unsigned int cpy, set;
120 BUG_ON(nfdt->max_fds < ofdt->max_fds);
121 if (ofdt->max_fds == 0)
124 cpy = ofdt->max_fds * sizeof(struct file *);
125 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
126 memcpy(nfdt->fd, ofdt->fd, cpy);
127 memset((char *)(nfdt->fd) + cpy, 0, set);
129 cpy = ofdt->max_fds / BITS_PER_BYTE;
130 set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
131 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
132 memset((char *)(nfdt->open_fds) + cpy, 0, set);
133 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
134 memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
137 static struct fdtable * alloc_fdtable(unsigned int nr)
143 * Figure out how many fds we actually want to support in this fdtable.
144 * Allocation steps are keyed to the size of the fdarray, since it
145 * grows far faster than any of the other dynamic data. We try to fit
146 * the fdarray into comfortable page-tuned chunks: starting at 1024B
147 * and growing in powers of two from there on.
149 nr /= (1024 / sizeof(struct file *));
150 nr = roundup_pow_of_two(nr + 1);
151 nr *= (1024 / sizeof(struct file *));
152 if (nr > sysctl_nr_open)
155 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
159 data = alloc_fdmem(nr * sizeof(struct file *));
162 fdt->fd = (struct file **)data;
163 data = alloc_fdmem(max_t(unsigned int,
164 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
167 fdt->open_fds = (fd_set *)data;
168 data += nr / BITS_PER_BYTE;
169 fdt->close_on_exec = (fd_set *)data;
170 INIT_RCU_HEAD(&fdt->rcu);
184 * Expand the file descriptor table.
185 * This function will allocate a new fdtable and both fd array and fdset, of
187 * Return <0 error code on error; 1 on successful completion.
188 * The files->file_lock should be held on entry, and will be held on exit.
190 static int expand_fdtable(struct files_struct *files, int nr)
191 __releases(files->file_lock)
192 __acquires(files->file_lock)
194 struct fdtable *new_fdt, *cur_fdt;
196 spin_unlock(&files->file_lock);
197 new_fdt = alloc_fdtable(nr);
198 spin_lock(&files->file_lock);
202 * Check again since another task may have expanded the fd table while
203 * we dropped the lock
205 cur_fdt = files_fdtable(files);
206 if (nr >= cur_fdt->max_fds) {
207 /* Continue as planned */
208 copy_fdtable(new_fdt, cur_fdt);
209 rcu_assign_pointer(files->fdt, new_fdt);
210 if (cur_fdt->max_fds > NR_OPEN_DEFAULT)
211 free_fdtable(cur_fdt);
213 /* Somebody else expanded, so undo our attempt */
223 * This function will expand the file structures, if the requested size exceeds
224 * the current capacity and there is room for expansion.
225 * Return <0 error code on error; 0 when nothing done; 1 when files were
226 * expanded and execution may have blocked.
227 * The files->file_lock should be held on entry, and will be held on exit.
229 int expand_files(struct files_struct *files, int nr)
233 fdt = files_fdtable(files);
234 /* Do we need to expand? */
235 if (nr < fdt->max_fds)
238 if (nr >= sysctl_nr_open)
241 /* All good, so we try */
242 return expand_fdtable(files, nr);
245 static void __devinit fdtable_defer_list_init(int cpu)
247 struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
248 spin_lock_init(&fddef->lock);
249 INIT_WORK(&fddef->wq, free_fdtable_work);
253 void __init files_defer_init(void)
256 for_each_possible_cpu(i)
257 fdtable_defer_list_init(i);