xref: /linux/fs/nfsd/filecache.c (revision 6c363eafc4d637ac4bd83d4a7dd06dd3cfbe7c5f)
1 /*
2  * Open file cache.
3  *
4  * (c) 2015 - Jeff Layton <jeff.layton@primarydata.com>
5  */
6 
7 #include <linux/hash.h>
8 #include <linux/slab.h>
9 #include <linux/file.h>
10 #include <linux/sched.h>
11 #include <linux/list_lru.h>
12 #include <linux/fsnotify_backend.h>
13 #include <linux/fsnotify.h>
14 #include <linux/seq_file.h>
15 
16 #include "vfs.h"
17 #include "nfsd.h"
18 #include "nfsfh.h"
19 #include "netns.h"
20 #include "filecache.h"
21 #include "trace.h"
22 
23 #define NFSDDBG_FACILITY	NFSDDBG_FH
24 
25 /* FIXME: dynamically size this for the machine somehow? */
26 #define NFSD_FILE_HASH_BITS                   12
27 #define NFSD_FILE_HASH_SIZE                  (1 << NFSD_FILE_HASH_BITS)
28 #define NFSD_LAUNDRETTE_DELAY		     (2 * HZ)
29 
30 #define NFSD_FILE_SHUTDOWN		     (1)
31 #define NFSD_FILE_LRU_THRESHOLD		     (4096UL)
32 #define NFSD_FILE_LRU_LIMIT		     (NFSD_FILE_LRU_THRESHOLD << 2)
33 
34 /* We only care about NFSD_MAY_READ/WRITE for this cache */
35 #define NFSD_FILE_MAY_MASK	(NFSD_MAY_READ|NFSD_MAY_WRITE)
36 
37 struct nfsd_fcache_bucket {
38 	struct hlist_head	nfb_head;
39 	spinlock_t		nfb_lock;
40 	unsigned int		nfb_count;
41 	unsigned int		nfb_maxcount;
42 };
43 
44 static DEFINE_PER_CPU(unsigned long, nfsd_file_cache_hits);
45 
46 struct nfsd_fcache_disposal {
47 	struct list_head list;
48 	struct work_struct work;
49 	struct net *net;
50 	spinlock_t lock;
51 	struct list_head freeme;
52 	struct rcu_head rcu;
53 };
54 
55 static struct workqueue_struct *nfsd_filecache_wq __read_mostly;
56 
57 static struct kmem_cache		*nfsd_file_slab;
58 static struct kmem_cache		*nfsd_file_mark_slab;
59 static struct nfsd_fcache_bucket	*nfsd_file_hashtbl;
60 static struct list_lru			nfsd_file_lru;
61 static long				nfsd_file_lru_flags;
62 static struct fsnotify_group		*nfsd_file_fsnotify_group;
63 static atomic_long_t			nfsd_filecache_count;
64 static struct delayed_work		nfsd_filecache_laundrette;
65 static DEFINE_SPINLOCK(laundrette_lock);
66 static LIST_HEAD(laundrettes);
67 
68 static void nfsd_file_gc(void);
69 
70 static void
71 nfsd_file_schedule_laundrette(void)
72 {
73 	long count = atomic_long_read(&nfsd_filecache_count);
74 
75 	if (count == 0 || test_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags))
76 		return;
77 
78 	queue_delayed_work(system_wq, &nfsd_filecache_laundrette,
79 			NFSD_LAUNDRETTE_DELAY);
80 }
81 
82 static void
83 nfsd_file_slab_free(struct rcu_head *rcu)
84 {
85 	struct nfsd_file *nf = container_of(rcu, struct nfsd_file, nf_rcu);
86 
87 	put_cred(nf->nf_cred);
88 	kmem_cache_free(nfsd_file_slab, nf);
89 }
90 
91 static void
92 nfsd_file_mark_free(struct fsnotify_mark *mark)
93 {
94 	struct nfsd_file_mark *nfm = container_of(mark, struct nfsd_file_mark,
95 						  nfm_mark);
96 
97 	kmem_cache_free(nfsd_file_mark_slab, nfm);
98 }
99 
100 static struct nfsd_file_mark *
101 nfsd_file_mark_get(struct nfsd_file_mark *nfm)
102 {
103 	if (!refcount_inc_not_zero(&nfm->nfm_ref))
104 		return NULL;
105 	return nfm;
106 }
107 
108 static void
109 nfsd_file_mark_put(struct nfsd_file_mark *nfm)
110 {
111 	if (refcount_dec_and_test(&nfm->nfm_ref)) {
112 		fsnotify_destroy_mark(&nfm->nfm_mark, nfsd_file_fsnotify_group);
113 		fsnotify_put_mark(&nfm->nfm_mark);
114 	}
115 }
116 
117 static struct nfsd_file_mark *
118 nfsd_file_mark_find_or_create(struct nfsd_file *nf)
119 {
120 	int			err;
121 	struct fsnotify_mark	*mark;
122 	struct nfsd_file_mark	*nfm = NULL, *new;
123 	struct inode *inode = nf->nf_inode;
124 
125 	do {
126 		mutex_lock(&nfsd_file_fsnotify_group->mark_mutex);
127 		mark = fsnotify_find_mark(&inode->i_fsnotify_marks,
128 				nfsd_file_fsnotify_group);
129 		if (mark) {
130 			nfm = nfsd_file_mark_get(container_of(mark,
131 						 struct nfsd_file_mark,
132 						 nfm_mark));
133 			mutex_unlock(&nfsd_file_fsnotify_group->mark_mutex);
134 			if (nfm) {
135 				fsnotify_put_mark(mark);
136 				break;
137 			}
138 			/* Avoid soft lockup race with nfsd_file_mark_put() */
139 			fsnotify_destroy_mark(mark, nfsd_file_fsnotify_group);
140 			fsnotify_put_mark(mark);
141 		} else
142 			mutex_unlock(&nfsd_file_fsnotify_group->mark_mutex);
143 
144 		/* allocate a new nfm */
145 		new = kmem_cache_alloc(nfsd_file_mark_slab, GFP_KERNEL);
146 		if (!new)
147 			return NULL;
148 		fsnotify_init_mark(&new->nfm_mark, nfsd_file_fsnotify_group);
149 		new->nfm_mark.mask = FS_ATTRIB|FS_DELETE_SELF;
150 		refcount_set(&new->nfm_ref, 1);
151 
152 		err = fsnotify_add_inode_mark(&new->nfm_mark, inode, 0);
153 
154 		/*
155 		 * If the add was successful, then return the object.
156 		 * Otherwise, we need to put the reference we hold on the
157 		 * nfm_mark. The fsnotify code will take a reference and put
158 		 * it on failure, so we can't just free it directly. It's also
159 		 * not safe to call fsnotify_destroy_mark on it as the
160 		 * mark->group will be NULL. Thus, we can't let the nfm_ref
161 		 * counter drive the destruction at this point.
162 		 */
163 		if (likely(!err))
164 			nfm = new;
165 		else
166 			fsnotify_put_mark(&new->nfm_mark);
167 	} while (unlikely(err == -EEXIST));
168 
169 	return nfm;
170 }
171 
172 static struct nfsd_file *
173 nfsd_file_alloc(struct inode *inode, unsigned int may, unsigned int hashval,
174 		struct net *net)
175 {
176 	struct nfsd_file *nf;
177 
178 	nf = kmem_cache_alloc(nfsd_file_slab, GFP_KERNEL);
179 	if (nf) {
180 		INIT_HLIST_NODE(&nf->nf_node);
181 		INIT_LIST_HEAD(&nf->nf_lru);
182 		nf->nf_file = NULL;
183 		nf->nf_cred = get_current_cred();
184 		nf->nf_net = net;
185 		nf->nf_flags = 0;
186 		nf->nf_inode = inode;
187 		nf->nf_hashval = hashval;
188 		refcount_set(&nf->nf_ref, 1);
189 		nf->nf_may = may & NFSD_FILE_MAY_MASK;
190 		if (may & NFSD_MAY_NOT_BREAK_LEASE) {
191 			if (may & NFSD_MAY_WRITE)
192 				__set_bit(NFSD_FILE_BREAK_WRITE, &nf->nf_flags);
193 			if (may & NFSD_MAY_READ)
194 				__set_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags);
195 		}
196 		nf->nf_mark = NULL;
197 		init_rwsem(&nf->nf_rwsem);
198 		trace_nfsd_file_alloc(nf);
199 	}
200 	return nf;
201 }
202 
203 static bool
204 nfsd_file_free(struct nfsd_file *nf)
205 {
206 	bool flush = false;
207 
208 	trace_nfsd_file_put_final(nf);
209 	if (nf->nf_mark)
210 		nfsd_file_mark_put(nf->nf_mark);
211 	if (nf->nf_file) {
212 		get_file(nf->nf_file);
213 		filp_close(nf->nf_file, NULL);
214 		fput(nf->nf_file);
215 		flush = true;
216 	}
217 	call_rcu(&nf->nf_rcu, nfsd_file_slab_free);
218 	return flush;
219 }
220 
221 static bool
222 nfsd_file_check_writeback(struct nfsd_file *nf)
223 {
224 	struct file *file = nf->nf_file;
225 	struct address_space *mapping;
226 
227 	if (!file || !(file->f_mode & FMODE_WRITE))
228 		return false;
229 	mapping = file->f_mapping;
230 	return mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) ||
231 		mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK);
232 }
233 
234 static int
235 nfsd_file_check_write_error(struct nfsd_file *nf)
236 {
237 	struct file *file = nf->nf_file;
238 
239 	if (!file || !(file->f_mode & FMODE_WRITE))
240 		return 0;
241 	return filemap_check_wb_err(file->f_mapping, READ_ONCE(file->f_wb_err));
242 }
243 
244 static void
245 nfsd_file_do_unhash(struct nfsd_file *nf)
246 {
247 	lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
248 
249 	trace_nfsd_file_unhash(nf);
250 
251 	if (nfsd_file_check_write_error(nf))
252 		nfsd_reset_boot_verifier(net_generic(nf->nf_net, nfsd_net_id));
253 	--nfsd_file_hashtbl[nf->nf_hashval].nfb_count;
254 	hlist_del_rcu(&nf->nf_node);
255 	atomic_long_dec(&nfsd_filecache_count);
256 }
257 
258 static bool
259 nfsd_file_unhash(struct nfsd_file *nf)
260 {
261 	if (test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
262 		nfsd_file_do_unhash(nf);
263 		if (!list_empty(&nf->nf_lru))
264 			list_lru_del(&nfsd_file_lru, &nf->nf_lru);
265 		return true;
266 	}
267 	return false;
268 }
269 
270 /*
271  * Return true if the file was unhashed.
272  */
273 static bool
274 nfsd_file_unhash_and_release_locked(struct nfsd_file *nf, struct list_head *dispose)
275 {
276 	lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
277 
278 	trace_nfsd_file_unhash_and_release_locked(nf);
279 	if (!nfsd_file_unhash(nf))
280 		return false;
281 	/* keep final reference for nfsd_file_lru_dispose */
282 	if (refcount_dec_not_one(&nf->nf_ref))
283 		return true;
284 
285 	list_add(&nf->nf_lru, dispose);
286 	return true;
287 }
288 
289 static void
290 nfsd_file_put_noref(struct nfsd_file *nf)
291 {
292 	trace_nfsd_file_put(nf);
293 
294 	if (refcount_dec_and_test(&nf->nf_ref)) {
295 		WARN_ON(test_bit(NFSD_FILE_HASHED, &nf->nf_flags));
296 		nfsd_file_free(nf);
297 	}
298 }
299 
300 void
301 nfsd_file_put(struct nfsd_file *nf)
302 {
303 	bool is_hashed;
304 
305 	set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
306 	if (refcount_read(&nf->nf_ref) > 2 || !nf->nf_file) {
307 		nfsd_file_put_noref(nf);
308 		return;
309 	}
310 
311 	filemap_flush(nf->nf_file->f_mapping);
312 	is_hashed = test_bit(NFSD_FILE_HASHED, &nf->nf_flags) != 0;
313 	nfsd_file_put_noref(nf);
314 	if (is_hashed)
315 		nfsd_file_schedule_laundrette();
316 	if (atomic_long_read(&nfsd_filecache_count) >= NFSD_FILE_LRU_LIMIT)
317 		nfsd_file_gc();
318 }
319 
320 struct nfsd_file *
321 nfsd_file_get(struct nfsd_file *nf)
322 {
323 	if (likely(refcount_inc_not_zero(&nf->nf_ref)))
324 		return nf;
325 	return NULL;
326 }
327 
328 static void
329 nfsd_file_dispose_list(struct list_head *dispose)
330 {
331 	struct nfsd_file *nf;
332 
333 	while(!list_empty(dispose)) {
334 		nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
335 		list_del(&nf->nf_lru);
336 		nfsd_file_put_noref(nf);
337 	}
338 }
339 
340 static void
341 nfsd_file_dispose_list_sync(struct list_head *dispose)
342 {
343 	bool flush = false;
344 	struct nfsd_file *nf;
345 
346 	while(!list_empty(dispose)) {
347 		nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
348 		list_del(&nf->nf_lru);
349 		if (!refcount_dec_and_test(&nf->nf_ref))
350 			continue;
351 		if (nfsd_file_free(nf))
352 			flush = true;
353 	}
354 	if (flush)
355 		flush_delayed_fput();
356 }
357 
358 static void
359 nfsd_file_list_remove_disposal(struct list_head *dst,
360 		struct nfsd_fcache_disposal *l)
361 {
362 	spin_lock(&l->lock);
363 	list_splice_init(&l->freeme, dst);
364 	spin_unlock(&l->lock);
365 }
366 
367 static void
368 nfsd_file_list_add_disposal(struct list_head *files, struct net *net)
369 {
370 	struct nfsd_fcache_disposal *l;
371 
372 	rcu_read_lock();
373 	list_for_each_entry_rcu(l, &laundrettes, list) {
374 		if (l->net == net) {
375 			spin_lock(&l->lock);
376 			list_splice_tail_init(files, &l->freeme);
377 			spin_unlock(&l->lock);
378 			queue_work(nfsd_filecache_wq, &l->work);
379 			break;
380 		}
381 	}
382 	rcu_read_unlock();
383 }
384 
385 static void
386 nfsd_file_list_add_pernet(struct list_head *dst, struct list_head *src,
387 		struct net *net)
388 {
389 	struct nfsd_file *nf, *tmp;
390 
391 	list_for_each_entry_safe(nf, tmp, src, nf_lru) {
392 		if (nf->nf_net == net)
393 			list_move_tail(&nf->nf_lru, dst);
394 	}
395 }
396 
397 static void
398 nfsd_file_dispose_list_delayed(struct list_head *dispose)
399 {
400 	LIST_HEAD(list);
401 	struct nfsd_file *nf;
402 
403 	while(!list_empty(dispose)) {
404 		nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
405 		nfsd_file_list_add_pernet(&list, dispose, nf->nf_net);
406 		nfsd_file_list_add_disposal(&list, nf->nf_net);
407 	}
408 }
409 
410 /*
411  * Note this can deadlock with nfsd_file_cache_purge.
412  */
413 static enum lru_status
414 nfsd_file_lru_cb(struct list_head *item, struct list_lru_one *lru,
415 		 spinlock_t *lock, void *arg)
416 	__releases(lock)
417 	__acquires(lock)
418 {
419 	struct list_head *head = arg;
420 	struct nfsd_file *nf = list_entry(item, struct nfsd_file, nf_lru);
421 
422 	/*
423 	 * Do a lockless refcount check. The hashtable holds one reference, so
424 	 * we look to see if anything else has a reference, or if any have
425 	 * been put since the shrinker last ran. Those don't get unhashed and
426 	 * released.
427 	 *
428 	 * Note that in the put path, we set the flag and then decrement the
429 	 * counter. Here we check the counter and then test and clear the flag.
430 	 * That order is deliberate to ensure that we can do this locklessly.
431 	 */
432 	if (refcount_read(&nf->nf_ref) > 1)
433 		goto out_skip;
434 
435 	/*
436 	 * Don't throw out files that are still undergoing I/O or
437 	 * that have uncleared errors pending.
438 	 */
439 	if (nfsd_file_check_writeback(nf))
440 		goto out_skip;
441 
442 	if (test_and_clear_bit(NFSD_FILE_REFERENCED, &nf->nf_flags))
443 		goto out_skip;
444 
445 	if (!test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags))
446 		goto out_skip;
447 
448 	list_lru_isolate_move(lru, &nf->nf_lru, head);
449 	return LRU_REMOVED;
450 out_skip:
451 	return LRU_SKIP;
452 }
453 
454 static unsigned long
455 nfsd_file_lru_walk_list(struct shrink_control *sc)
456 {
457 	LIST_HEAD(head);
458 	struct nfsd_file *nf;
459 	unsigned long ret;
460 
461 	if (sc)
462 		ret = list_lru_shrink_walk(&nfsd_file_lru, sc,
463 				nfsd_file_lru_cb, &head);
464 	else
465 		ret = list_lru_walk(&nfsd_file_lru,
466 				nfsd_file_lru_cb,
467 				&head, LONG_MAX);
468 	list_for_each_entry(nf, &head, nf_lru) {
469 		spin_lock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
470 		nfsd_file_do_unhash(nf);
471 		spin_unlock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
472 	}
473 	nfsd_file_dispose_list_delayed(&head);
474 	return ret;
475 }
476 
477 static void
478 nfsd_file_gc(void)
479 {
480 	nfsd_file_lru_walk_list(NULL);
481 }
482 
483 static void
484 nfsd_file_gc_worker(struct work_struct *work)
485 {
486 	nfsd_file_gc();
487 	nfsd_file_schedule_laundrette();
488 }
489 
490 static unsigned long
491 nfsd_file_lru_count(struct shrinker *s, struct shrink_control *sc)
492 {
493 	return list_lru_count(&nfsd_file_lru);
494 }
495 
496 static unsigned long
497 nfsd_file_lru_scan(struct shrinker *s, struct shrink_control *sc)
498 {
499 	return nfsd_file_lru_walk_list(sc);
500 }
501 
502 static struct shrinker	nfsd_file_shrinker = {
503 	.scan_objects = nfsd_file_lru_scan,
504 	.count_objects = nfsd_file_lru_count,
505 	.seeks = 1,
506 };
507 
508 static void
509 __nfsd_file_close_inode(struct inode *inode, unsigned int hashval,
510 			struct list_head *dispose)
511 {
512 	struct nfsd_file	*nf;
513 	struct hlist_node	*tmp;
514 
515 	spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
516 	hlist_for_each_entry_safe(nf, tmp, &nfsd_file_hashtbl[hashval].nfb_head, nf_node) {
517 		if (inode == nf->nf_inode)
518 			nfsd_file_unhash_and_release_locked(nf, dispose);
519 	}
520 	spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
521 }
522 
523 /**
524  * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
525  * @inode: inode of the file to attempt to remove
526  *
527  * Walk the whole hash bucket, looking for any files that correspond to "inode".
528  * If any do, then unhash them and put the hashtable reference to them and
529  * destroy any that had their last reference put. Also ensure that any of the
530  * fputs also have their final __fput done as well.
531  */
532 void
533 nfsd_file_close_inode_sync(struct inode *inode)
534 {
535 	unsigned int		hashval = (unsigned int)hash_long(inode->i_ino,
536 						NFSD_FILE_HASH_BITS);
537 	LIST_HEAD(dispose);
538 
539 	__nfsd_file_close_inode(inode, hashval, &dispose);
540 	trace_nfsd_file_close_inode_sync(inode, hashval, !list_empty(&dispose));
541 	nfsd_file_dispose_list_sync(&dispose);
542 }
543 
544 /**
545  * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
546  * @inode: inode of the file to attempt to remove
547  *
548  * Walk the whole hash bucket, looking for any files that correspond to "inode".
549  * If any do, then unhash them and put the hashtable reference to them and
550  * destroy any that had their last reference put.
551  */
552 static void
553 nfsd_file_close_inode(struct inode *inode)
554 {
555 	unsigned int		hashval = (unsigned int)hash_long(inode->i_ino,
556 						NFSD_FILE_HASH_BITS);
557 	LIST_HEAD(dispose);
558 
559 	__nfsd_file_close_inode(inode, hashval, &dispose);
560 	trace_nfsd_file_close_inode(inode, hashval, !list_empty(&dispose));
561 	nfsd_file_dispose_list_delayed(&dispose);
562 }
563 
564 /**
565  * nfsd_file_delayed_close - close unused nfsd_files
566  * @work: dummy
567  *
568  * Walk the LRU list and close any entries that have not been used since
569  * the last scan.
570  *
571  * Note this can deadlock with nfsd_file_cache_purge.
572  */
573 static void
574 nfsd_file_delayed_close(struct work_struct *work)
575 {
576 	LIST_HEAD(head);
577 	struct nfsd_fcache_disposal *l = container_of(work,
578 			struct nfsd_fcache_disposal, work);
579 
580 	nfsd_file_list_remove_disposal(&head, l);
581 	nfsd_file_dispose_list(&head);
582 }
583 
584 static int
585 nfsd_file_lease_notifier_call(struct notifier_block *nb, unsigned long arg,
586 			    void *data)
587 {
588 	struct file_lock *fl = data;
589 
590 	/* Only close files for F_SETLEASE leases */
591 	if (fl->fl_flags & FL_LEASE)
592 		nfsd_file_close_inode_sync(file_inode(fl->fl_file));
593 	return 0;
594 }
595 
596 static struct notifier_block nfsd_file_lease_notifier = {
597 	.notifier_call = nfsd_file_lease_notifier_call,
598 };
599 
600 static int
601 nfsd_file_fsnotify_handle_event(struct fsnotify_mark *mark, u32 mask,
602 				struct inode *inode, struct inode *dir,
603 				const struct qstr *name, u32 cookie)
604 {
605 	trace_nfsd_file_fsnotify_handle_event(inode, mask);
606 
607 	/* Should be no marks on non-regular files */
608 	if (!S_ISREG(inode->i_mode)) {
609 		WARN_ON_ONCE(1);
610 		return 0;
611 	}
612 
613 	/* don't close files if this was not the last link */
614 	if (mask & FS_ATTRIB) {
615 		if (inode->i_nlink)
616 			return 0;
617 	}
618 
619 	nfsd_file_close_inode(inode);
620 	return 0;
621 }
622 
623 
624 static const struct fsnotify_ops nfsd_file_fsnotify_ops = {
625 	.handle_inode_event = nfsd_file_fsnotify_handle_event,
626 	.free_mark = nfsd_file_mark_free,
627 };
628 
629 int
630 nfsd_file_cache_init(void)
631 {
632 	int		ret = -ENOMEM;
633 	unsigned int	i;
634 
635 	clear_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
636 
637 	if (nfsd_file_hashtbl)
638 		return 0;
639 
640 	nfsd_filecache_wq = alloc_workqueue("nfsd_filecache", 0, 0);
641 	if (!nfsd_filecache_wq)
642 		goto out;
643 
644 	nfsd_file_hashtbl = kcalloc(NFSD_FILE_HASH_SIZE,
645 				sizeof(*nfsd_file_hashtbl), GFP_KERNEL);
646 	if (!nfsd_file_hashtbl) {
647 		pr_err("nfsd: unable to allocate nfsd_file_hashtbl\n");
648 		goto out_err;
649 	}
650 
651 	nfsd_file_slab = kmem_cache_create("nfsd_file",
652 				sizeof(struct nfsd_file), 0, 0, NULL);
653 	if (!nfsd_file_slab) {
654 		pr_err("nfsd: unable to create nfsd_file_slab\n");
655 		goto out_err;
656 	}
657 
658 	nfsd_file_mark_slab = kmem_cache_create("nfsd_file_mark",
659 					sizeof(struct nfsd_file_mark), 0, 0, NULL);
660 	if (!nfsd_file_mark_slab) {
661 		pr_err("nfsd: unable to create nfsd_file_mark_slab\n");
662 		goto out_err;
663 	}
664 
665 
666 	ret = list_lru_init(&nfsd_file_lru);
667 	if (ret) {
668 		pr_err("nfsd: failed to init nfsd_file_lru: %d\n", ret);
669 		goto out_err;
670 	}
671 
672 	ret = register_shrinker(&nfsd_file_shrinker);
673 	if (ret) {
674 		pr_err("nfsd: failed to register nfsd_file_shrinker: %d\n", ret);
675 		goto out_lru;
676 	}
677 
678 	ret = lease_register_notifier(&nfsd_file_lease_notifier);
679 	if (ret) {
680 		pr_err("nfsd: unable to register lease notifier: %d\n", ret);
681 		goto out_shrinker;
682 	}
683 
684 	nfsd_file_fsnotify_group = fsnotify_alloc_group(&nfsd_file_fsnotify_ops);
685 	if (IS_ERR(nfsd_file_fsnotify_group)) {
686 		pr_err("nfsd: unable to create fsnotify group: %ld\n",
687 			PTR_ERR(nfsd_file_fsnotify_group));
688 		ret = PTR_ERR(nfsd_file_fsnotify_group);
689 		nfsd_file_fsnotify_group = NULL;
690 		goto out_notifier;
691 	}
692 
693 	for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
694 		INIT_HLIST_HEAD(&nfsd_file_hashtbl[i].nfb_head);
695 		spin_lock_init(&nfsd_file_hashtbl[i].nfb_lock);
696 	}
697 
698 	INIT_DELAYED_WORK(&nfsd_filecache_laundrette, nfsd_file_gc_worker);
699 out:
700 	return ret;
701 out_notifier:
702 	lease_unregister_notifier(&nfsd_file_lease_notifier);
703 out_shrinker:
704 	unregister_shrinker(&nfsd_file_shrinker);
705 out_lru:
706 	list_lru_destroy(&nfsd_file_lru);
707 out_err:
708 	kmem_cache_destroy(nfsd_file_slab);
709 	nfsd_file_slab = NULL;
710 	kmem_cache_destroy(nfsd_file_mark_slab);
711 	nfsd_file_mark_slab = NULL;
712 	kfree(nfsd_file_hashtbl);
713 	nfsd_file_hashtbl = NULL;
714 	destroy_workqueue(nfsd_filecache_wq);
715 	nfsd_filecache_wq = NULL;
716 	goto out;
717 }
718 
719 /*
720  * Note this can deadlock with nfsd_file_lru_cb.
721  */
722 void
723 nfsd_file_cache_purge(struct net *net)
724 {
725 	unsigned int		i;
726 	struct nfsd_file	*nf;
727 	struct hlist_node	*next;
728 	LIST_HEAD(dispose);
729 	bool del;
730 
731 	if (!nfsd_file_hashtbl)
732 		return;
733 
734 	for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
735 		struct nfsd_fcache_bucket *nfb = &nfsd_file_hashtbl[i];
736 
737 		spin_lock(&nfb->nfb_lock);
738 		hlist_for_each_entry_safe(nf, next, &nfb->nfb_head, nf_node) {
739 			if (net && nf->nf_net != net)
740 				continue;
741 			del = nfsd_file_unhash_and_release_locked(nf, &dispose);
742 
743 			/*
744 			 * Deadlock detected! Something marked this entry as
745 			 * unhased, but hasn't removed it from the hash list.
746 			 */
747 			WARN_ON_ONCE(!del);
748 		}
749 		spin_unlock(&nfb->nfb_lock);
750 		nfsd_file_dispose_list(&dispose);
751 	}
752 }
753 
754 static struct nfsd_fcache_disposal *
755 nfsd_alloc_fcache_disposal(struct net *net)
756 {
757 	struct nfsd_fcache_disposal *l;
758 
759 	l = kmalloc(sizeof(*l), GFP_KERNEL);
760 	if (!l)
761 		return NULL;
762 	INIT_WORK(&l->work, nfsd_file_delayed_close);
763 	l->net = net;
764 	spin_lock_init(&l->lock);
765 	INIT_LIST_HEAD(&l->freeme);
766 	return l;
767 }
768 
769 static void
770 nfsd_free_fcache_disposal(struct nfsd_fcache_disposal *l)
771 {
772 	rcu_assign_pointer(l->net, NULL);
773 	cancel_work_sync(&l->work);
774 	nfsd_file_dispose_list(&l->freeme);
775 	kfree_rcu(l, rcu);
776 }
777 
778 static void
779 nfsd_add_fcache_disposal(struct nfsd_fcache_disposal *l)
780 {
781 	spin_lock(&laundrette_lock);
782 	list_add_tail_rcu(&l->list, &laundrettes);
783 	spin_unlock(&laundrette_lock);
784 }
785 
786 static void
787 nfsd_del_fcache_disposal(struct nfsd_fcache_disposal *l)
788 {
789 	spin_lock(&laundrette_lock);
790 	list_del_rcu(&l->list);
791 	spin_unlock(&laundrette_lock);
792 }
793 
794 static int
795 nfsd_alloc_fcache_disposal_net(struct net *net)
796 {
797 	struct nfsd_fcache_disposal *l;
798 
799 	l = nfsd_alloc_fcache_disposal(net);
800 	if (!l)
801 		return -ENOMEM;
802 	nfsd_add_fcache_disposal(l);
803 	return 0;
804 }
805 
806 static void
807 nfsd_free_fcache_disposal_net(struct net *net)
808 {
809 	struct nfsd_fcache_disposal *l;
810 
811 	rcu_read_lock();
812 	list_for_each_entry_rcu(l, &laundrettes, list) {
813 		if (l->net != net)
814 			continue;
815 		nfsd_del_fcache_disposal(l);
816 		rcu_read_unlock();
817 		nfsd_free_fcache_disposal(l);
818 		return;
819 	}
820 	rcu_read_unlock();
821 }
822 
823 int
824 nfsd_file_cache_start_net(struct net *net)
825 {
826 	return nfsd_alloc_fcache_disposal_net(net);
827 }
828 
829 void
830 nfsd_file_cache_shutdown_net(struct net *net)
831 {
832 	nfsd_file_cache_purge(net);
833 	nfsd_free_fcache_disposal_net(net);
834 }
835 
836 void
837 nfsd_file_cache_shutdown(void)
838 {
839 	set_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
840 
841 	lease_unregister_notifier(&nfsd_file_lease_notifier);
842 	unregister_shrinker(&nfsd_file_shrinker);
843 	/*
844 	 * make sure all callers of nfsd_file_lru_cb are done before
845 	 * calling nfsd_file_cache_purge
846 	 */
847 	cancel_delayed_work_sync(&nfsd_filecache_laundrette);
848 	nfsd_file_cache_purge(NULL);
849 	list_lru_destroy(&nfsd_file_lru);
850 	rcu_barrier();
851 	fsnotify_put_group(nfsd_file_fsnotify_group);
852 	nfsd_file_fsnotify_group = NULL;
853 	kmem_cache_destroy(nfsd_file_slab);
854 	nfsd_file_slab = NULL;
855 	fsnotify_wait_marks_destroyed();
856 	kmem_cache_destroy(nfsd_file_mark_slab);
857 	nfsd_file_mark_slab = NULL;
858 	kfree(nfsd_file_hashtbl);
859 	nfsd_file_hashtbl = NULL;
860 	destroy_workqueue(nfsd_filecache_wq);
861 	nfsd_filecache_wq = NULL;
862 }
863 
864 static bool
865 nfsd_match_cred(const struct cred *c1, const struct cred *c2)
866 {
867 	int i;
868 
869 	if (!uid_eq(c1->fsuid, c2->fsuid))
870 		return false;
871 	if (!gid_eq(c1->fsgid, c2->fsgid))
872 		return false;
873 	if (c1->group_info == NULL || c2->group_info == NULL)
874 		return c1->group_info == c2->group_info;
875 	if (c1->group_info->ngroups != c2->group_info->ngroups)
876 		return false;
877 	for (i = 0; i < c1->group_info->ngroups; i++) {
878 		if (!gid_eq(c1->group_info->gid[i], c2->group_info->gid[i]))
879 			return false;
880 	}
881 	return true;
882 }
883 
884 static struct nfsd_file *
885 nfsd_file_find_locked(struct inode *inode, unsigned int may_flags,
886 			unsigned int hashval, struct net *net)
887 {
888 	struct nfsd_file *nf;
889 	unsigned char need = may_flags & NFSD_FILE_MAY_MASK;
890 
891 	hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
892 				 nf_node, lockdep_is_held(&nfsd_file_hashtbl[hashval].nfb_lock)) {
893 		if (nf->nf_may != need)
894 			continue;
895 		if (nf->nf_inode != inode)
896 			continue;
897 		if (nf->nf_net != net)
898 			continue;
899 		if (!nfsd_match_cred(nf->nf_cred, current_cred()))
900 			continue;
901 		if (nfsd_file_get(nf) != NULL)
902 			return nf;
903 	}
904 	return NULL;
905 }
906 
907 /**
908  * nfsd_file_is_cached - are there any cached open files for this fh?
909  * @inode: inode of the file to check
910  *
911  * Scan the hashtable for open files that match this fh. Returns true if there
912  * are any, and false if not.
913  */
914 bool
915 nfsd_file_is_cached(struct inode *inode)
916 {
917 	bool			ret = false;
918 	struct nfsd_file	*nf;
919 	unsigned int		hashval;
920 
921         hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
922 
923 	rcu_read_lock();
924 	hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
925 				 nf_node) {
926 		if (inode == nf->nf_inode) {
927 			ret = true;
928 			break;
929 		}
930 	}
931 	rcu_read_unlock();
932 	trace_nfsd_file_is_cached(inode, hashval, (int)ret);
933 	return ret;
934 }
935 
936 __be32
937 nfsd_file_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
938 		  unsigned int may_flags, struct nfsd_file **pnf)
939 {
940 	__be32	status;
941 	struct net *net = SVC_NET(rqstp);
942 	struct nfsd_file *nf, *new;
943 	struct inode *inode;
944 	unsigned int hashval;
945 	bool retry = true;
946 
947 	/* FIXME: skip this if fh_dentry is already set? */
948 	status = fh_verify(rqstp, fhp, S_IFREG,
949 				may_flags|NFSD_MAY_OWNER_OVERRIDE);
950 	if (status != nfs_ok)
951 		return status;
952 
953 	inode = d_inode(fhp->fh_dentry);
954 	hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
955 retry:
956 	rcu_read_lock();
957 	nf = nfsd_file_find_locked(inode, may_flags, hashval, net);
958 	rcu_read_unlock();
959 	if (nf)
960 		goto wait_for_construction;
961 
962 	new = nfsd_file_alloc(inode, may_flags, hashval, net);
963 	if (!new) {
964 		trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags,
965 					NULL, nfserr_jukebox);
966 		return nfserr_jukebox;
967 	}
968 
969 	spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
970 	nf = nfsd_file_find_locked(inode, may_flags, hashval, net);
971 	if (nf == NULL)
972 		goto open_file;
973 	spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
974 	nfsd_file_slab_free(&new->nf_rcu);
975 
976 wait_for_construction:
977 	wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
978 
979 	/* Did construction of this file fail? */
980 	if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
981 		if (!retry) {
982 			status = nfserr_jukebox;
983 			goto out;
984 		}
985 		retry = false;
986 		nfsd_file_put_noref(nf);
987 		goto retry;
988 	}
989 
990 	this_cpu_inc(nfsd_file_cache_hits);
991 
992 	if (!(may_flags & NFSD_MAY_NOT_BREAK_LEASE)) {
993 		bool write = (may_flags & NFSD_MAY_WRITE);
994 
995 		if (test_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags) ||
996 		    (test_bit(NFSD_FILE_BREAK_WRITE, &nf->nf_flags) && write)) {
997 			status = nfserrno(nfsd_open_break_lease(
998 					file_inode(nf->nf_file), may_flags));
999 			if (status == nfs_ok) {
1000 				clear_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags);
1001 				if (write)
1002 					clear_bit(NFSD_FILE_BREAK_WRITE,
1003 						  &nf->nf_flags);
1004 			}
1005 		}
1006 	}
1007 out:
1008 	if (status == nfs_ok) {
1009 		*pnf = nf;
1010 	} else {
1011 		nfsd_file_put(nf);
1012 		nf = NULL;
1013 	}
1014 
1015 	trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags, nf, status);
1016 	return status;
1017 open_file:
1018 	nf = new;
1019 	/* Take reference for the hashtable */
1020 	refcount_inc(&nf->nf_ref);
1021 	__set_bit(NFSD_FILE_HASHED, &nf->nf_flags);
1022 	__set_bit(NFSD_FILE_PENDING, &nf->nf_flags);
1023 	list_lru_add(&nfsd_file_lru, &nf->nf_lru);
1024 	hlist_add_head_rcu(&nf->nf_node, &nfsd_file_hashtbl[hashval].nfb_head);
1025 	++nfsd_file_hashtbl[hashval].nfb_count;
1026 	nfsd_file_hashtbl[hashval].nfb_maxcount = max(nfsd_file_hashtbl[hashval].nfb_maxcount,
1027 			nfsd_file_hashtbl[hashval].nfb_count);
1028 	spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
1029 	if (atomic_long_inc_return(&nfsd_filecache_count) >= NFSD_FILE_LRU_THRESHOLD)
1030 		nfsd_file_gc();
1031 
1032 	nf->nf_mark = nfsd_file_mark_find_or_create(nf);
1033 	if (nf->nf_mark)
1034 		status = nfsd_open_verified(rqstp, fhp, S_IFREG,
1035 				may_flags, &nf->nf_file);
1036 	else
1037 		status = nfserr_jukebox;
1038 	/*
1039 	 * If construction failed, or we raced with a call to unlink()
1040 	 * then unhash.
1041 	 */
1042 	if (status != nfs_ok || inode->i_nlink == 0) {
1043 		bool do_free;
1044 		spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
1045 		do_free = nfsd_file_unhash(nf);
1046 		spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
1047 		if (do_free)
1048 			nfsd_file_put_noref(nf);
1049 	}
1050 	clear_bit_unlock(NFSD_FILE_PENDING, &nf->nf_flags);
1051 	smp_mb__after_atomic();
1052 	wake_up_bit(&nf->nf_flags, NFSD_FILE_PENDING);
1053 	goto out;
1054 }
1055 
1056 /*
1057  * Note that fields may be added, removed or reordered in the future. Programs
1058  * scraping this file for info should test the labels to ensure they're
1059  * getting the correct field.
1060  */
1061 static int nfsd_file_cache_stats_show(struct seq_file *m, void *v)
1062 {
1063 	unsigned int i, count = 0, longest = 0;
1064 	unsigned long hits = 0;
1065 
1066 	/*
1067 	 * No need for spinlocks here since we're not terribly interested in
1068 	 * accuracy. We do take the nfsd_mutex simply to ensure that we
1069 	 * don't end up racing with server shutdown
1070 	 */
1071 	mutex_lock(&nfsd_mutex);
1072 	if (nfsd_file_hashtbl) {
1073 		for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
1074 			count += nfsd_file_hashtbl[i].nfb_count;
1075 			longest = max(longest, nfsd_file_hashtbl[i].nfb_count);
1076 		}
1077 	}
1078 	mutex_unlock(&nfsd_mutex);
1079 
1080 	for_each_possible_cpu(i)
1081 		hits += per_cpu(nfsd_file_cache_hits, i);
1082 
1083 	seq_printf(m, "total entries: %u\n", count);
1084 	seq_printf(m, "longest chain: %u\n", longest);
1085 	seq_printf(m, "cache hits:    %lu\n", hits);
1086 	return 0;
1087 }
1088 
1089 int nfsd_file_cache_stats_open(struct inode *inode, struct file *file)
1090 {
1091 	return single_open(file, nfsd_file_cache_stats_show, NULL);
1092 }
1093