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/fdtable.h>
16 #include <linux/bitops.h>
17 #include <linux/interrupt.h>
18 #include <linux/spinlock.h>
19 #include <linux/rcupdate.h>
20 #include <linux/workqueue.h>
22 struct fdtable_defer {
24 struct work_struct wq;
28 int sysctl_nr_open __read_mostly = 1024*1024;
31 * We use this list to defer free fdtables that have vmalloced
32 * sets/arrays. By keeping a per-cpu list, we avoid having to embed
33 * the work_struct in fdtable itself which avoids a 64 byte (i386) increase in
34 * this per-task structure.
36 static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
38 static inline void * alloc_fdmem(unsigned int size)
40 if (size <= PAGE_SIZE)
41 return kmalloc(size, GFP_KERNEL);
46 static inline void free_fdarr(struct fdtable *fdt)
48 if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *)))
54 static inline void free_fdset(struct fdtable *fdt)
56 if (fdt->max_fds <= (PAGE_SIZE * BITS_PER_BYTE / 2))
62 static void free_fdtable_work(struct work_struct *work)
64 struct fdtable_defer *f =
65 container_of(work, struct fdtable_defer, wq);
68 spin_lock_bh(&f->lock);
71 spin_unlock_bh(&f->lock);
73 struct fdtable *next = fdt->next;
81 void free_fdtable_rcu(struct rcu_head *rcu)
83 struct fdtable *fdt = container_of(rcu, struct fdtable, rcu);
84 struct fdtable_defer *fddef;
88 if (fdt->max_fds <= NR_OPEN_DEFAULT) {
90 * This fdtable is embedded in the files structure and that
91 * structure itself is getting destroyed.
93 kmem_cache_free(files_cachep,
94 container_of(fdt, struct files_struct, fdtab));
97 if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *))) {
102 fddef = &get_cpu_var(fdtable_defer_list);
103 spin_lock(&fddef->lock);
104 fdt->next = fddef->next;
106 /* vmallocs are handled from the workqueue context */
107 schedule_work(&fddef->wq);
108 spin_unlock(&fddef->lock);
109 put_cpu_var(fdtable_defer_list);
114 * Expand the fdset in the files_struct. Called with the files spinlock
117 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
119 unsigned int cpy, set;
121 BUG_ON(nfdt->max_fds < ofdt->max_fds);
122 if (ofdt->max_fds == 0)
125 cpy = ofdt->max_fds * sizeof(struct file *);
126 set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
127 memcpy(nfdt->fd, ofdt->fd, cpy);
128 memset((char *)(nfdt->fd) + cpy, 0, set);
130 cpy = ofdt->max_fds / BITS_PER_BYTE;
131 set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
132 memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
133 memset((char *)(nfdt->open_fds) + cpy, 0, set);
134 memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
135 memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
138 static struct fdtable * alloc_fdtable(unsigned int nr)
144 * Figure out how many fds we actually want to support in this fdtable.
145 * Allocation steps are keyed to the size of the fdarray, since it
146 * grows far faster than any of the other dynamic data. We try to fit
147 * the fdarray into comfortable page-tuned chunks: starting at 1024B
148 * and growing in powers of two from there on.
150 nr /= (1024 / sizeof(struct file *));
151 nr = roundup_pow_of_two(nr + 1);
152 nr *= (1024 / sizeof(struct file *));
153 if (nr > sysctl_nr_open)
156 fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
160 data = alloc_fdmem(nr * sizeof(struct file *));
163 fdt->fd = (struct file **)data;
164 data = alloc_fdmem(max_t(unsigned int,
165 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
168 fdt->open_fds = (fd_set *)data;
169 data += nr / BITS_PER_BYTE;
170 fdt->close_on_exec = (fd_set *)data;
171 INIT_RCU_HEAD(&fdt->rcu);
185 * Expand the file descriptor table.
186 * This function will allocate a new fdtable and both fd array and fdset, of
188 * Return <0 error code on error; 1 on successful completion.
189 * The files->file_lock should be held on entry, and will be held on exit.
191 static int expand_fdtable(struct files_struct *files, int nr)
192 __releases(files->file_lock)
193 __acquires(files->file_lock)
195 struct fdtable *new_fdt, *cur_fdt;
197 spin_unlock(&files->file_lock);
198 new_fdt = alloc_fdtable(nr);
199 spin_lock(&files->file_lock);
203 * Check again since another task may have expanded the fd table while
204 * we dropped the lock
206 cur_fdt = files_fdtable(files);
207 if (nr >= cur_fdt->max_fds) {
208 /* Continue as planned */
209 copy_fdtable(new_fdt, cur_fdt);
210 rcu_assign_pointer(files->fdt, new_fdt);
211 if (cur_fdt->max_fds > NR_OPEN_DEFAULT)
212 free_fdtable(cur_fdt);
214 /* Somebody else expanded, so undo our attempt */
224 * This function will expand the file structures, if the requested size exceeds
225 * the current capacity and there is room for expansion.
226 * Return <0 error code on error; 0 when nothing done; 1 when files were
227 * expanded and execution may have blocked.
228 * The files->file_lock should be held on entry, and will be held on exit.
230 int expand_files(struct files_struct *files, int nr)
234 fdt = files_fdtable(files);
235 /* Do we need to expand? */
236 if (nr < fdt->max_fds)
239 if (nr >= sysctl_nr_open)
242 /* All good, so we try */
243 return expand_fdtable(files, nr);
246 static void __devinit fdtable_defer_list_init(int cpu)
248 struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
249 spin_lock_init(&fddef->lock);
250 INIT_WORK(&fddef->wq, free_fdtable_work);
254 void __init files_defer_init(void)
257 for_each_possible_cpu(i)
258 fdtable_defer_list_init(i);