Merge tag 'nfsd-6.9' of git://git.kernel.org/pub/scm/linux/kernel/git/cel/linux
[linux-block.git] / fs / nfsd / filecache.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * The NFSD open file cache.
4  *
5  * (c) 2015 - Jeff Layton <jeff.layton@primarydata.com>
6  *
7  * An nfsd_file object is a per-file collection of open state that binds
8  * together:
9  *   - a struct file *
10  *   - a user credential
11  *   - a network namespace
12  *   - a read-ahead context
13  *   - monitoring for writeback errors
14  *
15  * nfsd_file objects are reference-counted. Consumers acquire a new
16  * object via the nfsd_file_acquire API. They manage their interest in
17  * the acquired object, and hence the object's reference count, via
18  * nfsd_file_get and nfsd_file_put. There are two varieties of nfsd_file
19  * object:
20  *
21  *  * non-garbage-collected: When a consumer wants to precisely control
22  *    the lifetime of a file's open state, it acquires a non-garbage-
23  *    collected nfsd_file. The final nfsd_file_put releases the open
24  *    state immediately.
25  *
26  *  * garbage-collected: When a consumer does not control the lifetime
27  *    of open state, it acquires a garbage-collected nfsd_file. The
28  *    final nfsd_file_put allows the open state to linger for a period
29  *    during which it may be re-used.
30  */
31
32 #include <linux/hash.h>
33 #include <linux/slab.h>
34 #include <linux/file.h>
35 #include <linux/pagemap.h>
36 #include <linux/sched.h>
37 #include <linux/list_lru.h>
38 #include <linux/fsnotify_backend.h>
39 #include <linux/fsnotify.h>
40 #include <linux/seq_file.h>
41 #include <linux/rhashtable.h>
42
43 #include "vfs.h"
44 #include "nfsd.h"
45 #include "nfsfh.h"
46 #include "netns.h"
47 #include "filecache.h"
48 #include "trace.h"
49
50 #define NFSD_LAUNDRETTE_DELAY                (2 * HZ)
51
52 #define NFSD_FILE_CACHE_UP                   (0)
53
54 /* We only care about NFSD_MAY_READ/WRITE for this cache */
55 #define NFSD_FILE_MAY_MASK      (NFSD_MAY_READ|NFSD_MAY_WRITE)
56
57 static DEFINE_PER_CPU(unsigned long, nfsd_file_cache_hits);
58 static DEFINE_PER_CPU(unsigned long, nfsd_file_acquisitions);
59 static DEFINE_PER_CPU(unsigned long, nfsd_file_releases);
60 static DEFINE_PER_CPU(unsigned long, nfsd_file_total_age);
61 static DEFINE_PER_CPU(unsigned long, nfsd_file_evictions);
62
63 struct nfsd_fcache_disposal {
64         spinlock_t lock;
65         struct list_head freeme;
66 };
67
68 static struct kmem_cache                *nfsd_file_slab;
69 static struct kmem_cache                *nfsd_file_mark_slab;
70 static struct list_lru                  nfsd_file_lru;
71 static unsigned long                    nfsd_file_flags;
72 static struct fsnotify_group            *nfsd_file_fsnotify_group;
73 static struct delayed_work              nfsd_filecache_laundrette;
74 static struct rhltable                  nfsd_file_rhltable
75                                                 ____cacheline_aligned_in_smp;
76
77 static bool
78 nfsd_match_cred(const struct cred *c1, const struct cred *c2)
79 {
80         int i;
81
82         if (!uid_eq(c1->fsuid, c2->fsuid))
83                 return false;
84         if (!gid_eq(c1->fsgid, c2->fsgid))
85                 return false;
86         if (c1->group_info == NULL || c2->group_info == NULL)
87                 return c1->group_info == c2->group_info;
88         if (c1->group_info->ngroups != c2->group_info->ngroups)
89                 return false;
90         for (i = 0; i < c1->group_info->ngroups; i++) {
91                 if (!gid_eq(c1->group_info->gid[i], c2->group_info->gid[i]))
92                         return false;
93         }
94         return true;
95 }
96
97 static const struct rhashtable_params nfsd_file_rhash_params = {
98         .key_len                = sizeof_field(struct nfsd_file, nf_inode),
99         .key_offset             = offsetof(struct nfsd_file, nf_inode),
100         .head_offset            = offsetof(struct nfsd_file, nf_rlist),
101
102         /*
103          * Start with a single page hash table to reduce resizing churn
104          * on light workloads.
105          */
106         .min_size               = 256,
107         .automatic_shrinking    = true,
108 };
109
110 static void
111 nfsd_file_schedule_laundrette(void)
112 {
113         if (test_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags))
114                 queue_delayed_work(system_wq, &nfsd_filecache_laundrette,
115                                    NFSD_LAUNDRETTE_DELAY);
116 }
117
118 static void
119 nfsd_file_slab_free(struct rcu_head *rcu)
120 {
121         struct nfsd_file *nf = container_of(rcu, struct nfsd_file, nf_rcu);
122
123         put_cred(nf->nf_cred);
124         kmem_cache_free(nfsd_file_slab, nf);
125 }
126
127 static void
128 nfsd_file_mark_free(struct fsnotify_mark *mark)
129 {
130         struct nfsd_file_mark *nfm = container_of(mark, struct nfsd_file_mark,
131                                                   nfm_mark);
132
133         kmem_cache_free(nfsd_file_mark_slab, nfm);
134 }
135
136 static struct nfsd_file_mark *
137 nfsd_file_mark_get(struct nfsd_file_mark *nfm)
138 {
139         if (!refcount_inc_not_zero(&nfm->nfm_ref))
140                 return NULL;
141         return nfm;
142 }
143
144 static void
145 nfsd_file_mark_put(struct nfsd_file_mark *nfm)
146 {
147         if (refcount_dec_and_test(&nfm->nfm_ref)) {
148                 fsnotify_destroy_mark(&nfm->nfm_mark, nfsd_file_fsnotify_group);
149                 fsnotify_put_mark(&nfm->nfm_mark);
150         }
151 }
152
153 static struct nfsd_file_mark *
154 nfsd_file_mark_find_or_create(struct nfsd_file *nf, struct inode *inode)
155 {
156         int                     err;
157         struct fsnotify_mark    *mark;
158         struct nfsd_file_mark   *nfm = NULL, *new;
159
160         do {
161                 fsnotify_group_lock(nfsd_file_fsnotify_group);
162                 mark = fsnotify_find_mark(&inode->i_fsnotify_marks,
163                                           nfsd_file_fsnotify_group);
164                 if (mark) {
165                         nfm = nfsd_file_mark_get(container_of(mark,
166                                                  struct nfsd_file_mark,
167                                                  nfm_mark));
168                         fsnotify_group_unlock(nfsd_file_fsnotify_group);
169                         if (nfm) {
170                                 fsnotify_put_mark(mark);
171                                 break;
172                         }
173                         /* Avoid soft lockup race with nfsd_file_mark_put() */
174                         fsnotify_destroy_mark(mark, nfsd_file_fsnotify_group);
175                         fsnotify_put_mark(mark);
176                 } else {
177                         fsnotify_group_unlock(nfsd_file_fsnotify_group);
178                 }
179
180                 /* allocate a new nfm */
181                 new = kmem_cache_alloc(nfsd_file_mark_slab, GFP_KERNEL);
182                 if (!new)
183                         return NULL;
184                 fsnotify_init_mark(&new->nfm_mark, nfsd_file_fsnotify_group);
185                 new->nfm_mark.mask = FS_ATTRIB|FS_DELETE_SELF;
186                 refcount_set(&new->nfm_ref, 1);
187
188                 err = fsnotify_add_inode_mark(&new->nfm_mark, inode, 0);
189
190                 /*
191                  * If the add was successful, then return the object.
192                  * Otherwise, we need to put the reference we hold on the
193                  * nfm_mark. The fsnotify code will take a reference and put
194                  * it on failure, so we can't just free it directly. It's also
195                  * not safe to call fsnotify_destroy_mark on it as the
196                  * mark->group will be NULL. Thus, we can't let the nfm_ref
197                  * counter drive the destruction at this point.
198                  */
199                 if (likely(!err))
200                         nfm = new;
201                 else
202                         fsnotify_put_mark(&new->nfm_mark);
203         } while (unlikely(err == -EEXIST));
204
205         return nfm;
206 }
207
208 static struct nfsd_file *
209 nfsd_file_alloc(struct net *net, struct inode *inode, unsigned char need,
210                 bool want_gc)
211 {
212         struct nfsd_file *nf;
213
214         nf = kmem_cache_alloc(nfsd_file_slab, GFP_KERNEL);
215         if (unlikely(!nf))
216                 return NULL;
217
218         INIT_LIST_HEAD(&nf->nf_lru);
219         nf->nf_birthtime = ktime_get();
220         nf->nf_file = NULL;
221         nf->nf_cred = get_current_cred();
222         nf->nf_net = net;
223         nf->nf_flags = want_gc ?
224                 BIT(NFSD_FILE_HASHED) | BIT(NFSD_FILE_PENDING) | BIT(NFSD_FILE_GC) :
225                 BIT(NFSD_FILE_HASHED) | BIT(NFSD_FILE_PENDING);
226         nf->nf_inode = inode;
227         refcount_set(&nf->nf_ref, 1);
228         nf->nf_may = need;
229         nf->nf_mark = NULL;
230         return nf;
231 }
232
233 /**
234  * nfsd_file_check_write_error - check for writeback errors on a file
235  * @nf: nfsd_file to check for writeback errors
236  *
237  * Check whether a nfsd_file has an unseen error. Reset the write
238  * verifier if so.
239  */
240 static void
241 nfsd_file_check_write_error(struct nfsd_file *nf)
242 {
243         struct file *file = nf->nf_file;
244
245         if ((file->f_mode & FMODE_WRITE) &&
246             filemap_check_wb_err(file->f_mapping, READ_ONCE(file->f_wb_err)))
247                 nfsd_reset_write_verifier(net_generic(nf->nf_net, nfsd_net_id));
248 }
249
250 static void
251 nfsd_file_hash_remove(struct nfsd_file *nf)
252 {
253         trace_nfsd_file_unhash(nf);
254         rhltable_remove(&nfsd_file_rhltable, &nf->nf_rlist,
255                         nfsd_file_rhash_params);
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_hash_remove(nf);
263                 return true;
264         }
265         return false;
266 }
267
268 static void
269 nfsd_file_free(struct nfsd_file *nf)
270 {
271         s64 age = ktime_to_ms(ktime_sub(ktime_get(), nf->nf_birthtime));
272
273         trace_nfsd_file_free(nf);
274
275         this_cpu_inc(nfsd_file_releases);
276         this_cpu_add(nfsd_file_total_age, age);
277
278         nfsd_file_unhash(nf);
279         if (nf->nf_mark)
280                 nfsd_file_mark_put(nf->nf_mark);
281         if (nf->nf_file) {
282                 nfsd_file_check_write_error(nf);
283                 nfsd_filp_close(nf->nf_file);
284         }
285
286         /*
287          * If this item is still linked via nf_lru, that's a bug.
288          * WARN and leak it to preserve system stability.
289          */
290         if (WARN_ON_ONCE(!list_empty(&nf->nf_lru)))
291                 return;
292
293         call_rcu(&nf->nf_rcu, nfsd_file_slab_free);
294 }
295
296 static bool
297 nfsd_file_check_writeback(struct nfsd_file *nf)
298 {
299         struct file *file = nf->nf_file;
300         struct address_space *mapping;
301
302         /* File not open for write? */
303         if (!(file->f_mode & FMODE_WRITE))
304                 return false;
305
306         /*
307          * Some filesystems (e.g. NFS) flush all dirty data on close.
308          * On others, there is no need to wait for writeback.
309          */
310         if (!(file_inode(file)->i_sb->s_export_op->flags & EXPORT_OP_FLUSH_ON_CLOSE))
311                 return false;
312
313         mapping = file->f_mapping;
314         return mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) ||
315                 mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK);
316 }
317
318
319 static bool nfsd_file_lru_add(struct nfsd_file *nf)
320 {
321         set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
322         if (list_lru_add_obj(&nfsd_file_lru, &nf->nf_lru)) {
323                 trace_nfsd_file_lru_add(nf);
324                 return true;
325         }
326         return false;
327 }
328
329 static bool nfsd_file_lru_remove(struct nfsd_file *nf)
330 {
331         if (list_lru_del_obj(&nfsd_file_lru, &nf->nf_lru)) {
332                 trace_nfsd_file_lru_del(nf);
333                 return true;
334         }
335         return false;
336 }
337
338 struct nfsd_file *
339 nfsd_file_get(struct nfsd_file *nf)
340 {
341         if (nf && refcount_inc_not_zero(&nf->nf_ref))
342                 return nf;
343         return NULL;
344 }
345
346 /**
347  * nfsd_file_put - put the reference to a nfsd_file
348  * @nf: nfsd_file of which to put the reference
349  *
350  * Put a reference to a nfsd_file. In the non-GC case, we just put the
351  * reference immediately. In the GC case, if the reference would be
352  * the last one, the put it on the LRU instead to be cleaned up later.
353  */
354 void
355 nfsd_file_put(struct nfsd_file *nf)
356 {
357         might_sleep();
358         trace_nfsd_file_put(nf);
359
360         if (test_bit(NFSD_FILE_GC, &nf->nf_flags) &&
361             test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
362                 /*
363                  * If this is the last reference (nf_ref == 1), then try to
364                  * transfer it to the LRU.
365                  */
366                 if (refcount_dec_not_one(&nf->nf_ref))
367                         return;
368
369                 /* Try to add it to the LRU.  If that fails, decrement. */
370                 if (nfsd_file_lru_add(nf)) {
371                         /* If it's still hashed, we're done */
372                         if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
373                                 nfsd_file_schedule_laundrette();
374                                 return;
375                         }
376
377                         /*
378                          * We're racing with unhashing, so try to remove it from
379                          * the LRU. If removal fails, then someone else already
380                          * has our reference.
381                          */
382                         if (!nfsd_file_lru_remove(nf))
383                                 return;
384                 }
385         }
386         if (refcount_dec_and_test(&nf->nf_ref))
387                 nfsd_file_free(nf);
388 }
389
390 static void
391 nfsd_file_dispose_list(struct list_head *dispose)
392 {
393         struct nfsd_file *nf;
394
395         while (!list_empty(dispose)) {
396                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
397                 list_del_init(&nf->nf_lru);
398                 nfsd_file_free(nf);
399         }
400 }
401
402 /**
403  * nfsd_file_dispose_list_delayed - move list of dead files to net's freeme list
404  * @dispose: list of nfsd_files to be disposed
405  *
406  * Transfers each file to the "freeme" list for its nfsd_net, to eventually
407  * be disposed of by the per-net garbage collector.
408  */
409 static void
410 nfsd_file_dispose_list_delayed(struct list_head *dispose)
411 {
412         while(!list_empty(dispose)) {
413                 struct nfsd_file *nf = list_first_entry(dispose,
414                                                 struct nfsd_file, nf_lru);
415                 struct nfsd_net *nn = net_generic(nf->nf_net, nfsd_net_id);
416                 struct nfsd_fcache_disposal *l = nn->fcache_disposal;
417
418                 spin_lock(&l->lock);
419                 list_move_tail(&nf->nf_lru, &l->freeme);
420                 spin_unlock(&l->lock);
421                 svc_wake_up(nn->nfsd_serv);
422         }
423 }
424
425 /**
426  * nfsd_file_net_dispose - deal with nfsd_files waiting to be disposed.
427  * @nn: nfsd_net in which to find files to be disposed.
428  *
429  * When files held open for nfsv3 are removed from the filecache, whether
430  * due to memory pressure or garbage collection, they are queued to
431  * a per-net-ns queue.  This function completes the disposal, either
432  * directly or by waking another nfsd thread to help with the work.
433  */
434 void nfsd_file_net_dispose(struct nfsd_net *nn)
435 {
436         struct nfsd_fcache_disposal *l = nn->fcache_disposal;
437
438         if (!list_empty(&l->freeme)) {
439                 LIST_HEAD(dispose);
440                 int i;
441
442                 spin_lock(&l->lock);
443                 for (i = 0; i < 8 && !list_empty(&l->freeme); i++)
444                         list_move(l->freeme.next, &dispose);
445                 spin_unlock(&l->lock);
446                 if (!list_empty(&l->freeme))
447                         /* Wake up another thread to share the work
448                          * *before* doing any actual disposing.
449                          */
450                         svc_wake_up(nn->nfsd_serv);
451                 nfsd_file_dispose_list(&dispose);
452         }
453 }
454
455 /**
456  * nfsd_file_lru_cb - Examine an entry on the LRU list
457  * @item: LRU entry to examine
458  * @lru: controlling LRU
459  * @lock: LRU list lock (unused)
460  * @arg: dispose list
461  *
462  * Return values:
463  *   %LRU_REMOVED: @item was removed from the LRU
464  *   %LRU_ROTATE: @item is to be moved to the LRU tail
465  *   %LRU_SKIP: @item cannot be evicted
466  */
467 static enum lru_status
468 nfsd_file_lru_cb(struct list_head *item, struct list_lru_one *lru,
469                  spinlock_t *lock, void *arg)
470         __releases(lock)
471         __acquires(lock)
472 {
473         struct list_head *head = arg;
474         struct nfsd_file *nf = list_entry(item, struct nfsd_file, nf_lru);
475
476         /* We should only be dealing with GC entries here */
477         WARN_ON_ONCE(!test_bit(NFSD_FILE_GC, &nf->nf_flags));
478
479         /*
480          * Don't throw out files that are still undergoing I/O or
481          * that have uncleared errors pending.
482          */
483         if (nfsd_file_check_writeback(nf)) {
484                 trace_nfsd_file_gc_writeback(nf);
485                 return LRU_SKIP;
486         }
487
488         /* If it was recently added to the list, skip it */
489         if (test_and_clear_bit(NFSD_FILE_REFERENCED, &nf->nf_flags)) {
490                 trace_nfsd_file_gc_referenced(nf);
491                 return LRU_ROTATE;
492         }
493
494         /*
495          * Put the reference held on behalf of the LRU. If it wasn't the last
496          * one, then just remove it from the LRU and ignore it.
497          */
498         if (!refcount_dec_and_test(&nf->nf_ref)) {
499                 trace_nfsd_file_gc_in_use(nf);
500                 list_lru_isolate(lru, &nf->nf_lru);
501                 return LRU_REMOVED;
502         }
503
504         /* Refcount went to zero. Unhash it and queue it to the dispose list */
505         nfsd_file_unhash(nf);
506         list_lru_isolate_move(lru, &nf->nf_lru, head);
507         this_cpu_inc(nfsd_file_evictions);
508         trace_nfsd_file_gc_disposed(nf);
509         return LRU_REMOVED;
510 }
511
512 static void
513 nfsd_file_gc(void)
514 {
515         LIST_HEAD(dispose);
516         unsigned long ret;
517
518         ret = list_lru_walk(&nfsd_file_lru, nfsd_file_lru_cb,
519                             &dispose, list_lru_count(&nfsd_file_lru));
520         trace_nfsd_file_gc_removed(ret, list_lru_count(&nfsd_file_lru));
521         nfsd_file_dispose_list_delayed(&dispose);
522 }
523
524 static void
525 nfsd_file_gc_worker(struct work_struct *work)
526 {
527         nfsd_file_gc();
528         if (list_lru_count(&nfsd_file_lru))
529                 nfsd_file_schedule_laundrette();
530 }
531
532 static unsigned long
533 nfsd_file_lru_count(struct shrinker *s, struct shrink_control *sc)
534 {
535         return list_lru_count(&nfsd_file_lru);
536 }
537
538 static unsigned long
539 nfsd_file_lru_scan(struct shrinker *s, struct shrink_control *sc)
540 {
541         LIST_HEAD(dispose);
542         unsigned long ret;
543
544         ret = list_lru_shrink_walk(&nfsd_file_lru, sc,
545                                    nfsd_file_lru_cb, &dispose);
546         trace_nfsd_file_shrinker_removed(ret, list_lru_count(&nfsd_file_lru));
547         nfsd_file_dispose_list_delayed(&dispose);
548         return ret;
549 }
550
551 static struct shrinker *nfsd_file_shrinker;
552
553 /**
554  * nfsd_file_cond_queue - conditionally unhash and queue a nfsd_file
555  * @nf: nfsd_file to attempt to queue
556  * @dispose: private list to queue successfully-put objects
557  *
558  * Unhash an nfsd_file, try to get a reference to it, and then put that
559  * reference. If it's the last reference, queue it to the dispose list.
560  */
561 static void
562 nfsd_file_cond_queue(struct nfsd_file *nf, struct list_head *dispose)
563         __must_hold(RCU)
564 {
565         int decrement = 1;
566
567         /* If we raced with someone else unhashing, ignore it */
568         if (!nfsd_file_unhash(nf))
569                 return;
570
571         /* If we can't get a reference, ignore it */
572         if (!nfsd_file_get(nf))
573                 return;
574
575         /* Extra decrement if we remove from the LRU */
576         if (nfsd_file_lru_remove(nf))
577                 ++decrement;
578
579         /* If refcount goes to 0, then put on the dispose list */
580         if (refcount_sub_and_test(decrement, &nf->nf_ref)) {
581                 list_add(&nf->nf_lru, dispose);
582                 trace_nfsd_file_closing(nf);
583         }
584 }
585
586 /**
587  * nfsd_file_queue_for_close: try to close out any open nfsd_files for an inode
588  * @inode:   inode on which to close out nfsd_files
589  * @dispose: list on which to gather nfsd_files to close out
590  *
591  * An nfsd_file represents a struct file being held open on behalf of nfsd.
592  * An open file however can block other activity (such as leases), or cause
593  * undesirable behavior (e.g. spurious silly-renames when reexporting NFS).
594  *
595  * This function is intended to find open nfsd_files when this sort of
596  * conflicting access occurs and then attempt to close those files out.
597  *
598  * Populates the dispose list with entries that have already had their
599  * refcounts go to zero. The actual free of an nfsd_file can be expensive,
600  * so we leave it up to the caller whether it wants to wait or not.
601  */
602 static void
603 nfsd_file_queue_for_close(struct inode *inode, struct list_head *dispose)
604 {
605         struct rhlist_head *tmp, *list;
606         struct nfsd_file *nf;
607
608         rcu_read_lock();
609         list = rhltable_lookup(&nfsd_file_rhltable, &inode,
610                                nfsd_file_rhash_params);
611         rhl_for_each_entry_rcu(nf, tmp, list, nf_rlist) {
612                 if (!test_bit(NFSD_FILE_GC, &nf->nf_flags))
613                         continue;
614                 nfsd_file_cond_queue(nf, dispose);
615         }
616         rcu_read_unlock();
617 }
618
619 /**
620  * nfsd_file_close_inode - attempt a delayed close of a nfsd_file
621  * @inode: inode of the file to attempt to remove
622  *
623  * Close out any open nfsd_files that can be reaped for @inode. The
624  * actual freeing is deferred to the dispose_list_delayed infrastructure.
625  *
626  * This is used by the fsnotify callbacks and setlease notifier.
627  */
628 static void
629 nfsd_file_close_inode(struct inode *inode)
630 {
631         LIST_HEAD(dispose);
632
633         nfsd_file_queue_for_close(inode, &dispose);
634         nfsd_file_dispose_list_delayed(&dispose);
635 }
636
637 /**
638  * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
639  * @inode: inode of the file to attempt to remove
640  *
641  * Close out any open nfsd_files that can be reaped for @inode. The
642  * nfsd_files are closed out synchronously.
643  *
644  * This is called from nfsd_rename and nfsd_unlink to avoid silly-renames
645  * when reexporting NFS.
646  */
647 void
648 nfsd_file_close_inode_sync(struct inode *inode)
649 {
650         struct nfsd_file *nf;
651         LIST_HEAD(dispose);
652
653         trace_nfsd_file_close(inode);
654
655         nfsd_file_queue_for_close(inode, &dispose);
656         while (!list_empty(&dispose)) {
657                 nf = list_first_entry(&dispose, struct nfsd_file, nf_lru);
658                 list_del_init(&nf->nf_lru);
659                 nfsd_file_free(nf);
660         }
661 }
662
663 static int
664 nfsd_file_lease_notifier_call(struct notifier_block *nb, unsigned long arg,
665                             void *data)
666 {
667         struct file_lock *fl = data;
668
669         /* Only close files for F_SETLEASE leases */
670         if (fl->c.flc_flags & FL_LEASE)
671                 nfsd_file_close_inode(file_inode(fl->c.flc_file));
672         return 0;
673 }
674
675 static struct notifier_block nfsd_file_lease_notifier = {
676         .notifier_call = nfsd_file_lease_notifier_call,
677 };
678
679 static int
680 nfsd_file_fsnotify_handle_event(struct fsnotify_mark *mark, u32 mask,
681                                 struct inode *inode, struct inode *dir,
682                                 const struct qstr *name, u32 cookie)
683 {
684         if (WARN_ON_ONCE(!inode))
685                 return 0;
686
687         trace_nfsd_file_fsnotify_handle_event(inode, mask);
688
689         /* Should be no marks on non-regular files */
690         if (!S_ISREG(inode->i_mode)) {
691                 WARN_ON_ONCE(1);
692                 return 0;
693         }
694
695         /* don't close files if this was not the last link */
696         if (mask & FS_ATTRIB) {
697                 if (inode->i_nlink)
698                         return 0;
699         }
700
701         nfsd_file_close_inode(inode);
702         return 0;
703 }
704
705
706 static const struct fsnotify_ops nfsd_file_fsnotify_ops = {
707         .handle_inode_event = nfsd_file_fsnotify_handle_event,
708         .free_mark = nfsd_file_mark_free,
709 };
710
711 int
712 nfsd_file_cache_init(void)
713 {
714         int ret;
715
716         lockdep_assert_held(&nfsd_mutex);
717         if (test_and_set_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags) == 1)
718                 return 0;
719
720         ret = rhltable_init(&nfsd_file_rhltable, &nfsd_file_rhash_params);
721         if (ret)
722                 return ret;
723
724         ret = -ENOMEM;
725         nfsd_file_slab = KMEM_CACHE(nfsd_file, 0);
726         if (!nfsd_file_slab) {
727                 pr_err("nfsd: unable to create nfsd_file_slab\n");
728                 goto out_err;
729         }
730
731         nfsd_file_mark_slab = KMEM_CACHE(nfsd_file_mark, 0);
732         if (!nfsd_file_mark_slab) {
733                 pr_err("nfsd: unable to create nfsd_file_mark_slab\n");
734                 goto out_err;
735         }
736
737         ret = list_lru_init(&nfsd_file_lru);
738         if (ret) {
739                 pr_err("nfsd: failed to init nfsd_file_lru: %d\n", ret);
740                 goto out_err;
741         }
742
743         nfsd_file_shrinker = shrinker_alloc(0, "nfsd-filecache");
744         if (!nfsd_file_shrinker) {
745                 ret = -ENOMEM;
746                 pr_err("nfsd: failed to allocate nfsd_file_shrinker\n");
747                 goto out_lru;
748         }
749
750         nfsd_file_shrinker->count_objects = nfsd_file_lru_count;
751         nfsd_file_shrinker->scan_objects = nfsd_file_lru_scan;
752         nfsd_file_shrinker->seeks = 1;
753
754         shrinker_register(nfsd_file_shrinker);
755
756         ret = lease_register_notifier(&nfsd_file_lease_notifier);
757         if (ret) {
758                 pr_err("nfsd: unable to register lease notifier: %d\n", ret);
759                 goto out_shrinker;
760         }
761
762         nfsd_file_fsnotify_group = fsnotify_alloc_group(&nfsd_file_fsnotify_ops,
763                                                         FSNOTIFY_GROUP_NOFS);
764         if (IS_ERR(nfsd_file_fsnotify_group)) {
765                 pr_err("nfsd: unable to create fsnotify group: %ld\n",
766                         PTR_ERR(nfsd_file_fsnotify_group));
767                 ret = PTR_ERR(nfsd_file_fsnotify_group);
768                 nfsd_file_fsnotify_group = NULL;
769                 goto out_notifier;
770         }
771
772         INIT_DELAYED_WORK(&nfsd_filecache_laundrette, nfsd_file_gc_worker);
773 out:
774         return ret;
775 out_notifier:
776         lease_unregister_notifier(&nfsd_file_lease_notifier);
777 out_shrinker:
778         shrinker_free(nfsd_file_shrinker);
779 out_lru:
780         list_lru_destroy(&nfsd_file_lru);
781 out_err:
782         kmem_cache_destroy(nfsd_file_slab);
783         nfsd_file_slab = NULL;
784         kmem_cache_destroy(nfsd_file_mark_slab);
785         nfsd_file_mark_slab = NULL;
786         rhltable_destroy(&nfsd_file_rhltable);
787         goto out;
788 }
789
790 /**
791  * __nfsd_file_cache_purge: clean out the cache for shutdown
792  * @net: net-namespace to shut down the cache (may be NULL)
793  *
794  * Walk the nfsd_file cache and close out any that match @net. If @net is NULL,
795  * then close out everything. Called when an nfsd instance is being shut down,
796  * and when the exports table is flushed.
797  */
798 static void
799 __nfsd_file_cache_purge(struct net *net)
800 {
801         struct rhashtable_iter iter;
802         struct nfsd_file *nf;
803         LIST_HEAD(dispose);
804
805         rhltable_walk_enter(&nfsd_file_rhltable, &iter);
806         do {
807                 rhashtable_walk_start(&iter);
808
809                 nf = rhashtable_walk_next(&iter);
810                 while (!IS_ERR_OR_NULL(nf)) {
811                         if (!net || nf->nf_net == net)
812                                 nfsd_file_cond_queue(nf, &dispose);
813                         nf = rhashtable_walk_next(&iter);
814                 }
815
816                 rhashtable_walk_stop(&iter);
817         } while (nf == ERR_PTR(-EAGAIN));
818         rhashtable_walk_exit(&iter);
819
820         nfsd_file_dispose_list(&dispose);
821 }
822
823 static struct nfsd_fcache_disposal *
824 nfsd_alloc_fcache_disposal(void)
825 {
826         struct nfsd_fcache_disposal *l;
827
828         l = kmalloc(sizeof(*l), GFP_KERNEL);
829         if (!l)
830                 return NULL;
831         spin_lock_init(&l->lock);
832         INIT_LIST_HEAD(&l->freeme);
833         return l;
834 }
835
836 static void
837 nfsd_free_fcache_disposal(struct nfsd_fcache_disposal *l)
838 {
839         nfsd_file_dispose_list(&l->freeme);
840         kfree(l);
841 }
842
843 static void
844 nfsd_free_fcache_disposal_net(struct net *net)
845 {
846         struct nfsd_net *nn = net_generic(net, nfsd_net_id);
847         struct nfsd_fcache_disposal *l = nn->fcache_disposal;
848
849         nfsd_free_fcache_disposal(l);
850 }
851
852 int
853 nfsd_file_cache_start_net(struct net *net)
854 {
855         struct nfsd_net *nn = net_generic(net, nfsd_net_id);
856
857         nn->fcache_disposal = nfsd_alloc_fcache_disposal();
858         return nn->fcache_disposal ? 0 : -ENOMEM;
859 }
860
861 /**
862  * nfsd_file_cache_purge - Remove all cache items associated with @net
863  * @net: target net namespace
864  *
865  */
866 void
867 nfsd_file_cache_purge(struct net *net)
868 {
869         lockdep_assert_held(&nfsd_mutex);
870         if (test_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags) == 1)
871                 __nfsd_file_cache_purge(net);
872 }
873
874 void
875 nfsd_file_cache_shutdown_net(struct net *net)
876 {
877         nfsd_file_cache_purge(net);
878         nfsd_free_fcache_disposal_net(net);
879 }
880
881 void
882 nfsd_file_cache_shutdown(void)
883 {
884         int i;
885
886         lockdep_assert_held(&nfsd_mutex);
887         if (test_and_clear_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags) == 0)
888                 return;
889
890         lease_unregister_notifier(&nfsd_file_lease_notifier);
891         shrinker_free(nfsd_file_shrinker);
892         /*
893          * make sure all callers of nfsd_file_lru_cb are done before
894          * calling nfsd_file_cache_purge
895          */
896         cancel_delayed_work_sync(&nfsd_filecache_laundrette);
897         __nfsd_file_cache_purge(NULL);
898         list_lru_destroy(&nfsd_file_lru);
899         rcu_barrier();
900         fsnotify_put_group(nfsd_file_fsnotify_group);
901         nfsd_file_fsnotify_group = NULL;
902         kmem_cache_destroy(nfsd_file_slab);
903         nfsd_file_slab = NULL;
904         fsnotify_wait_marks_destroyed();
905         kmem_cache_destroy(nfsd_file_mark_slab);
906         nfsd_file_mark_slab = NULL;
907         rhltable_destroy(&nfsd_file_rhltable);
908
909         for_each_possible_cpu(i) {
910                 per_cpu(nfsd_file_cache_hits, i) = 0;
911                 per_cpu(nfsd_file_acquisitions, i) = 0;
912                 per_cpu(nfsd_file_releases, i) = 0;
913                 per_cpu(nfsd_file_total_age, i) = 0;
914                 per_cpu(nfsd_file_evictions, i) = 0;
915         }
916 }
917
918 static struct nfsd_file *
919 nfsd_file_lookup_locked(const struct net *net, const struct cred *cred,
920                         struct inode *inode, unsigned char need,
921                         bool want_gc)
922 {
923         struct rhlist_head *tmp, *list;
924         struct nfsd_file *nf;
925
926         list = rhltable_lookup(&nfsd_file_rhltable, &inode,
927                                nfsd_file_rhash_params);
928         rhl_for_each_entry_rcu(nf, tmp, list, nf_rlist) {
929                 if (nf->nf_may != need)
930                         continue;
931                 if (nf->nf_net != net)
932                         continue;
933                 if (!nfsd_match_cred(nf->nf_cred, cred))
934                         continue;
935                 if (test_bit(NFSD_FILE_GC, &nf->nf_flags) != want_gc)
936                         continue;
937                 if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags) == 0)
938                         continue;
939
940                 if (!nfsd_file_get(nf))
941                         continue;
942                 return nf;
943         }
944         return NULL;
945 }
946
947 /**
948  * nfsd_file_is_cached - are there any cached open files for this inode?
949  * @inode: inode to check
950  *
951  * The lookup matches inodes in all net namespaces and is atomic wrt
952  * nfsd_file_acquire().
953  *
954  * Return values:
955  *   %true: filecache contains at least one file matching this inode
956  *   %false: filecache contains no files matching this inode
957  */
958 bool
959 nfsd_file_is_cached(struct inode *inode)
960 {
961         struct rhlist_head *tmp, *list;
962         struct nfsd_file *nf;
963         bool ret = false;
964
965         rcu_read_lock();
966         list = rhltable_lookup(&nfsd_file_rhltable, &inode,
967                                nfsd_file_rhash_params);
968         rhl_for_each_entry_rcu(nf, tmp, list, nf_rlist)
969                 if (test_bit(NFSD_FILE_GC, &nf->nf_flags)) {
970                         ret = true;
971                         break;
972                 }
973         rcu_read_unlock();
974
975         trace_nfsd_file_is_cached(inode, (int)ret);
976         return ret;
977 }
978
979 static __be32
980 nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
981                      unsigned int may_flags, struct file *file,
982                      struct nfsd_file **pnf, bool want_gc)
983 {
984         unsigned char need = may_flags & NFSD_FILE_MAY_MASK;
985         struct net *net = SVC_NET(rqstp);
986         struct nfsd_file *new, *nf;
987         bool stale_retry = true;
988         bool open_retry = true;
989         struct inode *inode;
990         __be32 status;
991         int ret;
992
993 retry:
994         status = fh_verify(rqstp, fhp, S_IFREG,
995                                 may_flags|NFSD_MAY_OWNER_OVERRIDE);
996         if (status != nfs_ok)
997                 return status;
998         inode = d_inode(fhp->fh_dentry);
999
1000         rcu_read_lock();
1001         nf = nfsd_file_lookup_locked(net, current_cred(), inode, need, want_gc);
1002         rcu_read_unlock();
1003
1004         if (nf) {
1005                 /*
1006                  * If the nf is on the LRU then it holds an extra reference
1007                  * that must be put if it's removed. It had better not be
1008                  * the last one however, since we should hold another.
1009                  */
1010                 if (nfsd_file_lru_remove(nf))
1011                         WARN_ON_ONCE(refcount_dec_and_test(&nf->nf_ref));
1012                 goto wait_for_construction;
1013         }
1014
1015         new = nfsd_file_alloc(net, inode, need, want_gc);
1016         if (!new) {
1017                 status = nfserr_jukebox;
1018                 goto out;
1019         }
1020
1021         rcu_read_lock();
1022         spin_lock(&inode->i_lock);
1023         nf = nfsd_file_lookup_locked(net, current_cred(), inode, need, want_gc);
1024         if (unlikely(nf)) {
1025                 spin_unlock(&inode->i_lock);
1026                 rcu_read_unlock();
1027                 nfsd_file_slab_free(&new->nf_rcu);
1028                 goto wait_for_construction;
1029         }
1030         nf = new;
1031         ret = rhltable_insert(&nfsd_file_rhltable, &nf->nf_rlist,
1032                               nfsd_file_rhash_params);
1033         spin_unlock(&inode->i_lock);
1034         rcu_read_unlock();
1035         if (likely(ret == 0))
1036                 goto open_file;
1037
1038         if (ret == -EEXIST)
1039                 goto retry;
1040         trace_nfsd_file_insert_err(rqstp, inode, may_flags, ret);
1041         status = nfserr_jukebox;
1042         goto construction_err;
1043
1044 wait_for_construction:
1045         wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
1046
1047         /* Did construction of this file fail? */
1048         if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
1049                 trace_nfsd_file_cons_err(rqstp, inode, may_flags, nf);
1050                 if (!open_retry) {
1051                         status = nfserr_jukebox;
1052                         goto construction_err;
1053                 }
1054                 open_retry = false;
1055                 fh_put(fhp);
1056                 goto retry;
1057         }
1058         this_cpu_inc(nfsd_file_cache_hits);
1059
1060         status = nfserrno(nfsd_open_break_lease(file_inode(nf->nf_file), may_flags));
1061         if (status != nfs_ok) {
1062                 nfsd_file_put(nf);
1063                 nf = NULL;
1064         }
1065
1066 out:
1067         if (status == nfs_ok) {
1068                 this_cpu_inc(nfsd_file_acquisitions);
1069                 nfsd_file_check_write_error(nf);
1070                 *pnf = nf;
1071         }
1072         trace_nfsd_file_acquire(rqstp, inode, may_flags, nf, status);
1073         return status;
1074
1075 open_file:
1076         trace_nfsd_file_alloc(nf);
1077         nf->nf_mark = nfsd_file_mark_find_or_create(nf, inode);
1078         if (nf->nf_mark) {
1079                 if (file) {
1080                         get_file(file);
1081                         nf->nf_file = file;
1082                         status = nfs_ok;
1083                         trace_nfsd_file_opened(nf, status);
1084                 } else {
1085                         ret = nfsd_open_verified(rqstp, fhp, may_flags,
1086                                                  &nf->nf_file);
1087                         if (ret == -EOPENSTALE && stale_retry) {
1088                                 stale_retry = false;
1089                                 nfsd_file_unhash(nf);
1090                                 clear_and_wake_up_bit(NFSD_FILE_PENDING,
1091                                                       &nf->nf_flags);
1092                                 if (refcount_dec_and_test(&nf->nf_ref))
1093                                         nfsd_file_free(nf);
1094                                 nf = NULL;
1095                                 fh_put(fhp);
1096                                 goto retry;
1097                         }
1098                         status = nfserrno(ret);
1099                         trace_nfsd_file_open(nf, status);
1100                 }
1101         } else
1102                 status = nfserr_jukebox;
1103         /*
1104          * If construction failed, or we raced with a call to unlink()
1105          * then unhash.
1106          */
1107         if (status != nfs_ok || inode->i_nlink == 0)
1108                 nfsd_file_unhash(nf);
1109         clear_and_wake_up_bit(NFSD_FILE_PENDING, &nf->nf_flags);
1110         if (status == nfs_ok)
1111                 goto out;
1112
1113 construction_err:
1114         if (refcount_dec_and_test(&nf->nf_ref))
1115                 nfsd_file_free(nf);
1116         nf = NULL;
1117         goto out;
1118 }
1119
1120 /**
1121  * nfsd_file_acquire_gc - Get a struct nfsd_file with an open file
1122  * @rqstp: the RPC transaction being executed
1123  * @fhp: the NFS filehandle of the file to be opened
1124  * @may_flags: NFSD_MAY_ settings for the file
1125  * @pnf: OUT: new or found "struct nfsd_file" object
1126  *
1127  * The nfsd_file object returned by this API is reference-counted
1128  * and garbage-collected. The object is retained for a few
1129  * seconds after the final nfsd_file_put() in case the caller
1130  * wants to re-use it.
1131  *
1132  * Return values:
1133  *   %nfs_ok - @pnf points to an nfsd_file with its reference
1134  *   count boosted.
1135  *
1136  * On error, an nfsstat value in network byte order is returned.
1137  */
1138 __be32
1139 nfsd_file_acquire_gc(struct svc_rqst *rqstp, struct svc_fh *fhp,
1140                      unsigned int may_flags, struct nfsd_file **pnf)
1141 {
1142         return nfsd_file_do_acquire(rqstp, fhp, may_flags, NULL, pnf, true);
1143 }
1144
1145 /**
1146  * nfsd_file_acquire - Get a struct nfsd_file with an open file
1147  * @rqstp: the RPC transaction being executed
1148  * @fhp: the NFS filehandle of the file to be opened
1149  * @may_flags: NFSD_MAY_ settings for the file
1150  * @pnf: OUT: new or found "struct nfsd_file" object
1151  *
1152  * The nfsd_file_object returned by this API is reference-counted
1153  * but not garbage-collected. The object is unhashed after the
1154  * final nfsd_file_put().
1155  *
1156  * Return values:
1157  *   %nfs_ok - @pnf points to an nfsd_file with its reference
1158  *   count boosted.
1159  *
1160  * On error, an nfsstat value in network byte order is returned.
1161  */
1162 __be32
1163 nfsd_file_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
1164                   unsigned int may_flags, struct nfsd_file **pnf)
1165 {
1166         return nfsd_file_do_acquire(rqstp, fhp, may_flags, NULL, pnf, false);
1167 }
1168
1169 /**
1170  * nfsd_file_acquire_opened - Get a struct nfsd_file using existing open file
1171  * @rqstp: the RPC transaction being executed
1172  * @fhp: the NFS filehandle of the file just created
1173  * @may_flags: NFSD_MAY_ settings for the file
1174  * @file: cached, already-open file (may be NULL)
1175  * @pnf: OUT: new or found "struct nfsd_file" object
1176  *
1177  * Acquire a nfsd_file object that is not GC'ed. If one doesn't already exist,
1178  * and @file is non-NULL, use it to instantiate a new nfsd_file instead of
1179  * opening a new one.
1180  *
1181  * Return values:
1182  *   %nfs_ok - @pnf points to an nfsd_file with its reference
1183  *   count boosted.
1184  *
1185  * On error, an nfsstat value in network byte order is returned.
1186  */
1187 __be32
1188 nfsd_file_acquire_opened(struct svc_rqst *rqstp, struct svc_fh *fhp,
1189                          unsigned int may_flags, struct file *file,
1190                          struct nfsd_file **pnf)
1191 {
1192         return nfsd_file_do_acquire(rqstp, fhp, may_flags, file, pnf, false);
1193 }
1194
1195 /*
1196  * Note that fields may be added, removed or reordered in the future. Programs
1197  * scraping this file for info should test the labels to ensure they're
1198  * getting the correct field.
1199  */
1200 int nfsd_file_cache_stats_show(struct seq_file *m, void *v)
1201 {
1202         unsigned long releases = 0, evictions = 0;
1203         unsigned long hits = 0, acquisitions = 0;
1204         unsigned int i, count = 0, buckets = 0;
1205         unsigned long lru = 0, total_age = 0;
1206
1207         /* Serialize with server shutdown */
1208         mutex_lock(&nfsd_mutex);
1209         if (test_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags) == 1) {
1210                 struct bucket_table *tbl;
1211                 struct rhashtable *ht;
1212
1213                 lru = list_lru_count(&nfsd_file_lru);
1214
1215                 rcu_read_lock();
1216                 ht = &nfsd_file_rhltable.ht;
1217                 count = atomic_read(&ht->nelems);
1218                 tbl = rht_dereference_rcu(ht->tbl, ht);
1219                 buckets = tbl->size;
1220                 rcu_read_unlock();
1221         }
1222         mutex_unlock(&nfsd_mutex);
1223
1224         for_each_possible_cpu(i) {
1225                 hits += per_cpu(nfsd_file_cache_hits, i);
1226                 acquisitions += per_cpu(nfsd_file_acquisitions, i);
1227                 releases += per_cpu(nfsd_file_releases, i);
1228                 total_age += per_cpu(nfsd_file_total_age, i);
1229                 evictions += per_cpu(nfsd_file_evictions, i);
1230         }
1231
1232         seq_printf(m, "total inodes:  %u\n", count);
1233         seq_printf(m, "hash buckets:  %u\n", buckets);
1234         seq_printf(m, "lru entries:   %lu\n", lru);
1235         seq_printf(m, "cache hits:    %lu\n", hits);
1236         seq_printf(m, "acquisitions:  %lu\n", acquisitions);
1237         seq_printf(m, "releases:      %lu\n", releases);
1238         seq_printf(m, "evictions:     %lu\n", evictions);
1239         if (releases)
1240                 seq_printf(m, "mean age (ms): %ld\n", total_age / releases);
1241         else
1242                 seq_printf(m, "mean age (ms): -\n");
1243         return 0;
1244 }