Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/fs/file.c | |
3 | * | |
4 | * Copyright (C) 1998-1999, Stephen Tweedie and Bill Hawes | |
5 | * | |
6 | * Manage the dynamic fd arrays in the process files_struct. | |
7 | */ | |
8 | ||
9 | #include <linux/fs.h> | |
10 | #include <linux/mm.h> | |
11 | #include <linux/time.h> | |
12 | #include <linux/slab.h> | |
13 | #include <linux/vmalloc.h> | |
14 | #include <linux/file.h> | |
9f3acc31 | 15 | #include <linux/fdtable.h> |
1da177e4 | 16 | #include <linux/bitops.h> |
ab2af1f5 DS |
17 | #include <linux/interrupt.h> |
18 | #include <linux/spinlock.h> | |
19 | #include <linux/rcupdate.h> | |
20 | #include <linux/workqueue.h> | |
21 | ||
22 | struct fdtable_defer { | |
23 | spinlock_t lock; | |
24 | struct work_struct wq; | |
ab2af1f5 DS |
25 | struct fdtable *next; |
26 | }; | |
27 | ||
9cfe015a ED |
28 | int sysctl_nr_open __read_mostly = 1024*1024; |
29 | ||
ab2af1f5 DS |
30 | /* |
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. | |
35 | */ | |
36 | static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list); | |
1da177e4 | 37 | |
5466b456 | 38 | static inline void * alloc_fdmem(unsigned int size) |
1da177e4 | 39 | { |
1da177e4 | 40 | if (size <= PAGE_SIZE) |
5466b456 VL |
41 | return kmalloc(size, GFP_KERNEL); |
42 | else | |
43 | return vmalloc(size); | |
1da177e4 LT |
44 | } |
45 | ||
5466b456 | 46 | static inline void free_fdarr(struct fdtable *fdt) |
1da177e4 | 47 | { |
5466b456 VL |
48 | if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *))) |
49 | kfree(fdt->fd); | |
1da177e4 | 50 | else |
5466b456 | 51 | vfree(fdt->fd); |
1da177e4 LT |
52 | } |
53 | ||
5466b456 | 54 | static inline void free_fdset(struct fdtable *fdt) |
1da177e4 | 55 | { |
5466b456 VL |
56 | if (fdt->max_fds <= (PAGE_SIZE * BITS_PER_BYTE / 2)) |
57 | kfree(fdt->open_fds); | |
58 | else | |
59 | vfree(fdt->open_fds); | |
ab2af1f5 | 60 | } |
1da177e4 | 61 | |
65f27f38 | 62 | static void free_fdtable_work(struct work_struct *work) |
ab2af1f5 | 63 | { |
65f27f38 DH |
64 | struct fdtable_defer *f = |
65 | container_of(work, struct fdtable_defer, wq); | |
ab2af1f5 | 66 | struct fdtable *fdt; |
1da177e4 | 67 | |
ab2af1f5 DS |
68 | spin_lock_bh(&f->lock); |
69 | fdt = f->next; | |
70 | f->next = NULL; | |
71 | spin_unlock_bh(&f->lock); | |
72 | while(fdt) { | |
73 | struct fdtable *next = fdt->next; | |
5466b456 VL |
74 | vfree(fdt->fd); |
75 | free_fdset(fdt); | |
76 | kfree(fdt); | |
ab2af1f5 DS |
77 | fdt = next; |
78 | } | |
79 | } | |
1da177e4 | 80 | |
4fd45812 | 81 | void free_fdtable_rcu(struct rcu_head *rcu) |
ab2af1f5 DS |
82 | { |
83 | struct fdtable *fdt = container_of(rcu, struct fdtable, rcu); | |
ab2af1f5 | 84 | struct fdtable_defer *fddef; |
1da177e4 | 85 | |
ab2af1f5 | 86 | BUG_ON(!fdt); |
ab2af1f5 | 87 | |
4fd45812 | 88 | if (fdt->max_fds <= NR_OPEN_DEFAULT) { |
ab2af1f5 | 89 | /* |
4fd45812 VL |
90 | * This fdtable is embedded in the files structure and that |
91 | * structure itself is getting destroyed. | |
ab2af1f5 | 92 | */ |
4fd45812 VL |
93 | kmem_cache_free(files_cachep, |
94 | container_of(fdt, struct files_struct, fdtab)); | |
ab2af1f5 DS |
95 | return; |
96 | } | |
5466b456 | 97 | if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *))) { |
ab2af1f5 | 98 | kfree(fdt->fd); |
5466b456 | 99 | kfree(fdt->open_fds); |
ab2af1f5 | 100 | kfree(fdt); |
1da177e4 | 101 | } else { |
ab2af1f5 DS |
102 | fddef = &get_cpu_var(fdtable_defer_list); |
103 | spin_lock(&fddef->lock); | |
104 | fdt->next = fddef->next; | |
105 | fddef->next = fdt; | |
593be07a TH |
106 | /* vmallocs are handled from the workqueue context */ |
107 | schedule_work(&fddef->wq); | |
ab2af1f5 DS |
108 | spin_unlock(&fddef->lock); |
109 | put_cpu_var(fdtable_defer_list); | |
1da177e4 | 110 | } |
ab2af1f5 DS |
111 | } |
112 | ||
ab2af1f5 DS |
113 | /* |
114 | * Expand the fdset in the files_struct. Called with the files spinlock | |
115 | * held for write. | |
116 | */ | |
5466b456 | 117 | static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt) |
ab2af1f5 | 118 | { |
5466b456 | 119 | unsigned int cpy, set; |
ab2af1f5 | 120 | |
5466b456 | 121 | BUG_ON(nfdt->max_fds < ofdt->max_fds); |
5466b456 VL |
122 | |
123 | cpy = ofdt->max_fds * sizeof(struct file *); | |
124 | set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *); | |
125 | memcpy(nfdt->fd, ofdt->fd, cpy); | |
126 | memset((char *)(nfdt->fd) + cpy, 0, set); | |
127 | ||
128 | cpy = ofdt->max_fds / BITS_PER_BYTE; | |
129 | set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE; | |
130 | memcpy(nfdt->open_fds, ofdt->open_fds, cpy); | |
131 | memset((char *)(nfdt->open_fds) + cpy, 0, set); | |
132 | memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy); | |
133 | memset((char *)(nfdt->close_on_exec) + cpy, 0, set); | |
1da177e4 LT |
134 | } |
135 | ||
5466b456 | 136 | static struct fdtable * alloc_fdtable(unsigned int nr) |
1da177e4 | 137 | { |
5466b456 VL |
138 | struct fdtable *fdt; |
139 | char *data; | |
1da177e4 | 140 | |
ab2af1f5 | 141 | /* |
5466b456 VL |
142 | * Figure out how many fds we actually want to support in this fdtable. |
143 | * Allocation steps are keyed to the size of the fdarray, since it | |
144 | * grows far faster than any of the other dynamic data. We try to fit | |
145 | * the fdarray into comfortable page-tuned chunks: starting at 1024B | |
146 | * and growing in powers of two from there on. | |
ab2af1f5 | 147 | */ |
5466b456 VL |
148 | nr /= (1024 / sizeof(struct file *)); |
149 | nr = roundup_pow_of_two(nr + 1); | |
150 | nr *= (1024 / sizeof(struct file *)); | |
5c598b34 AV |
151 | /* |
152 | * Note that this can drive nr *below* what we had passed if sysctl_nr_open | |
153 | * had been set lower between the check in expand_files() and here. Deal | |
154 | * with that in caller, it's cheaper that way. | |
155 | * | |
156 | * We make sure that nr remains a multiple of BITS_PER_LONG - otherwise | |
157 | * bitmaps handling below becomes unpleasant, to put it mildly... | |
158 | */ | |
159 | if (unlikely(nr > sysctl_nr_open)) | |
160 | nr = ((sysctl_nr_open - 1) | (BITS_PER_LONG - 1)) + 1; | |
bbea9f69 | 161 | |
5466b456 VL |
162 | fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL); |
163 | if (!fdt) | |
bbea9f69 | 164 | goto out; |
5466b456 VL |
165 | fdt->max_fds = nr; |
166 | data = alloc_fdmem(nr * sizeof(struct file *)); | |
167 | if (!data) | |
168 | goto out_fdt; | |
169 | fdt->fd = (struct file **)data; | |
170 | data = alloc_fdmem(max_t(unsigned int, | |
171 | 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES)); | |
172 | if (!data) | |
173 | goto out_arr; | |
174 | fdt->open_fds = (fd_set *)data; | |
175 | data += nr / BITS_PER_BYTE; | |
176 | fdt->close_on_exec = (fd_set *)data; | |
177 | INIT_RCU_HEAD(&fdt->rcu); | |
178 | fdt->next = NULL; | |
179 | ||
ab2af1f5 | 180 | return fdt; |
5466b456 VL |
181 | |
182 | out_arr: | |
183 | free_fdarr(fdt); | |
184 | out_fdt: | |
ab2af1f5 | 185 | kfree(fdt); |
5466b456 | 186 | out: |
ab2af1f5 DS |
187 | return NULL; |
188 | } | |
1da177e4 | 189 | |
ab2af1f5 | 190 | /* |
74d392aa VL |
191 | * Expand the file descriptor table. |
192 | * This function will allocate a new fdtable and both fd array and fdset, of | |
193 | * the given size. | |
194 | * Return <0 error code on error; 1 on successful completion. | |
195 | * The files->file_lock should be held on entry, and will be held on exit. | |
ab2af1f5 DS |
196 | */ |
197 | static int expand_fdtable(struct files_struct *files, int nr) | |
198 | __releases(files->file_lock) | |
199 | __acquires(files->file_lock) | |
200 | { | |
74d392aa | 201 | struct fdtable *new_fdt, *cur_fdt; |
ab2af1f5 DS |
202 | |
203 | spin_unlock(&files->file_lock); | |
74d392aa | 204 | new_fdt = alloc_fdtable(nr); |
ab2af1f5 | 205 | spin_lock(&files->file_lock); |
74d392aa VL |
206 | if (!new_fdt) |
207 | return -ENOMEM; | |
5c598b34 AV |
208 | /* |
209 | * extremely unlikely race - sysctl_nr_open decreased between the check in | |
210 | * caller and alloc_fdtable(). Cheaper to catch it here... | |
211 | */ | |
212 | if (unlikely(new_fdt->max_fds <= nr)) { | |
213 | free_fdarr(new_fdt); | |
214 | free_fdset(new_fdt); | |
215 | kfree(new_fdt); | |
216 | return -EMFILE; | |
217 | } | |
ab2af1f5 | 218 | /* |
74d392aa VL |
219 | * Check again since another task may have expanded the fd table while |
220 | * we dropped the lock | |
ab2af1f5 | 221 | */ |
74d392aa | 222 | cur_fdt = files_fdtable(files); |
bbea9f69 | 223 | if (nr >= cur_fdt->max_fds) { |
74d392aa VL |
224 | /* Continue as planned */ |
225 | copy_fdtable(new_fdt, cur_fdt); | |
226 | rcu_assign_pointer(files->fdt, new_fdt); | |
4fd45812 | 227 | if (cur_fdt->max_fds > NR_OPEN_DEFAULT) |
01b2d93c | 228 | free_fdtable(cur_fdt); |
ab2af1f5 | 229 | } else { |
74d392aa | 230 | /* Somebody else expanded, so undo our attempt */ |
5466b456 VL |
231 | free_fdarr(new_fdt); |
232 | free_fdset(new_fdt); | |
233 | kfree(new_fdt); | |
ab2af1f5 | 234 | } |
74d392aa | 235 | return 1; |
1da177e4 LT |
236 | } |
237 | ||
238 | /* | |
239 | * Expand files. | |
74d392aa VL |
240 | * This function will expand the file structures, if the requested size exceeds |
241 | * the current capacity and there is room for expansion. | |
242 | * Return <0 error code on error; 0 when nothing done; 1 when files were | |
243 | * expanded and execution may have blocked. | |
244 | * The files->file_lock should be held on entry, and will be held on exit. | |
1da177e4 LT |
245 | */ |
246 | int expand_files(struct files_struct *files, int nr) | |
247 | { | |
badf1662 | 248 | struct fdtable *fdt; |
1da177e4 | 249 | |
badf1662 | 250 | fdt = files_fdtable(files); |
74d392aa | 251 | /* Do we need to expand? */ |
bbea9f69 | 252 | if (nr < fdt->max_fds) |
74d392aa VL |
253 | return 0; |
254 | /* Can we expand? */ | |
9cfe015a | 255 | if (nr >= sysctl_nr_open) |
74d392aa VL |
256 | return -EMFILE; |
257 | ||
258 | /* All good, so we try */ | |
259 | return expand_fdtable(files, nr); | |
1da177e4 | 260 | } |
ab2af1f5 | 261 | |
02afc626 AV |
262 | static int count_open_files(struct fdtable *fdt) |
263 | { | |
264 | int size = fdt->max_fds; | |
265 | int i; | |
266 | ||
267 | /* Find the last open fd */ | |
268 | for (i = size/(8*sizeof(long)); i > 0; ) { | |
269 | if (fdt->open_fds->fds_bits[--i]) | |
270 | break; | |
271 | } | |
272 | i = (i+1) * 8 * sizeof(long); | |
273 | return i; | |
274 | } | |
275 | ||
02afc626 AV |
276 | /* |
277 | * Allocate a new files structure and copy contents from the | |
278 | * passed in files structure. | |
279 | * errorp will be valid only when the returned files_struct is NULL. | |
280 | */ | |
281 | struct files_struct *dup_fd(struct files_struct *oldf, int *errorp) | |
282 | { | |
283 | struct files_struct *newf; | |
284 | struct file **old_fds, **new_fds; | |
285 | int open_files, size, i; | |
286 | struct fdtable *old_fdt, *new_fdt; | |
287 | ||
288 | *errorp = -ENOMEM; | |
afbec7ff | 289 | newf = kmem_cache_alloc(files_cachep, GFP_KERNEL); |
02afc626 AV |
290 | if (!newf) |
291 | goto out; | |
292 | ||
afbec7ff AV |
293 | atomic_set(&newf->count, 1); |
294 | ||
295 | spin_lock_init(&newf->file_lock); | |
296 | newf->next_fd = 0; | |
297 | new_fdt = &newf->fdtab; | |
298 | new_fdt->max_fds = NR_OPEN_DEFAULT; | |
299 | new_fdt->close_on_exec = (fd_set *)&newf->close_on_exec_init; | |
300 | new_fdt->open_fds = (fd_set *)&newf->open_fds_init; | |
301 | new_fdt->fd = &newf->fd_array[0]; | |
302 | INIT_RCU_HEAD(&new_fdt->rcu); | |
303 | new_fdt->next = NULL; | |
304 | ||
02afc626 AV |
305 | spin_lock(&oldf->file_lock); |
306 | old_fdt = files_fdtable(oldf); | |
02afc626 AV |
307 | open_files = count_open_files(old_fdt); |
308 | ||
309 | /* | |
310 | * Check whether we need to allocate a larger fd array and fd set. | |
311 | * Note: we're not a clone task, so the open count won't change. | |
312 | */ | |
313 | if (open_files > new_fdt->max_fds) { | |
02afc626 | 314 | spin_unlock(&oldf->file_lock); |
9dec3c4d AV |
315 | |
316 | new_fdt = alloc_fdtable(open_files - 1); | |
317 | if (!new_fdt) { | |
318 | *errorp = -ENOMEM; | |
319 | goto out_release; | |
320 | } | |
321 | ||
322 | /* beyond sysctl_nr_open; nothing to do */ | |
323 | if (unlikely(new_fdt->max_fds < open_files)) { | |
324 | free_fdarr(new_fdt); | |
325 | free_fdset(new_fdt); | |
326 | kfree(new_fdt); | |
327 | *errorp = -EMFILE; | |
02afc626 | 328 | goto out_release; |
9dec3c4d | 329 | } |
9dec3c4d | 330 | |
02afc626 AV |
331 | /* |
332 | * Reacquire the oldf lock and a pointer to its fd table | |
333 | * who knows it may have a new bigger fd table. We need | |
334 | * the latest pointer. | |
335 | */ | |
336 | spin_lock(&oldf->file_lock); | |
337 | old_fdt = files_fdtable(oldf); | |
338 | } | |
339 | ||
340 | old_fds = old_fdt->fd; | |
341 | new_fds = new_fdt->fd; | |
342 | ||
343 | memcpy(new_fdt->open_fds->fds_bits, | |
344 | old_fdt->open_fds->fds_bits, open_files/8); | |
345 | memcpy(new_fdt->close_on_exec->fds_bits, | |
346 | old_fdt->close_on_exec->fds_bits, open_files/8); | |
347 | ||
348 | for (i = open_files; i != 0; i--) { | |
349 | struct file *f = *old_fds++; | |
350 | if (f) { | |
351 | get_file(f); | |
352 | } else { | |
353 | /* | |
354 | * The fd may be claimed in the fd bitmap but not yet | |
355 | * instantiated in the files array if a sibling thread | |
356 | * is partway through open(). So make sure that this | |
357 | * fd is available to the new process. | |
358 | */ | |
359 | FD_CLR(open_files - i, new_fdt->open_fds); | |
360 | } | |
361 | rcu_assign_pointer(*new_fds++, f); | |
362 | } | |
363 | spin_unlock(&oldf->file_lock); | |
364 | ||
365 | /* compute the remainder to be cleared */ | |
366 | size = (new_fdt->max_fds - open_files) * sizeof(struct file *); | |
367 | ||
368 | /* This is long word aligned thus could use a optimized version */ | |
369 | memset(new_fds, 0, size); | |
370 | ||
371 | if (new_fdt->max_fds > open_files) { | |
372 | int left = (new_fdt->max_fds-open_files)/8; | |
373 | int start = open_files / (8 * sizeof(unsigned long)); | |
374 | ||
375 | memset(&new_fdt->open_fds->fds_bits[start], 0, left); | |
376 | memset(&new_fdt->close_on_exec->fds_bits[start], 0, left); | |
377 | } | |
378 | ||
afbec7ff AV |
379 | rcu_assign_pointer(newf->fdt, new_fdt); |
380 | ||
02afc626 AV |
381 | return newf; |
382 | ||
383 | out_release: | |
384 | kmem_cache_free(files_cachep, newf); | |
385 | out: | |
386 | return NULL; | |
387 | } | |
388 | ||
ab2af1f5 DS |
389 | static void __devinit fdtable_defer_list_init(int cpu) |
390 | { | |
391 | struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu); | |
392 | spin_lock_init(&fddef->lock); | |
65f27f38 | 393 | INIT_WORK(&fddef->wq, free_fdtable_work); |
ab2af1f5 DS |
394 | fddef->next = NULL; |
395 | } | |
396 | ||
397 | void __init files_defer_init(void) | |
398 | { | |
399 | int i; | |
0a945022 | 400 | for_each_possible_cpu(i) |
ab2af1f5 DS |
401 | fdtable_defer_list_init(i); |
402 | } | |
f52111b1 AV |
403 | |
404 | struct files_struct init_files = { | |
405 | .count = ATOMIC_INIT(1), | |
406 | .fdt = &init_files.fdtab, | |
407 | .fdtab = { | |
408 | .max_fds = NR_OPEN_DEFAULT, | |
409 | .fd = &init_files.fd_array[0], | |
410 | .close_on_exec = (fd_set *)&init_files.close_on_exec_init, | |
411 | .open_fds = (fd_set *)&init_files.open_fds_init, | |
412 | .rcu = RCU_HEAD_INIT, | |
413 | }, | |
414 | .file_lock = __SPIN_LOCK_UNLOCKED(init_task.file_lock), | |
415 | }; |