Merge ath-next from git://git.kernel.org/pub/scm/linux/kernel/git/kvalo/ath.git
[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         struct work_struct work;
65         spinlock_t lock;
66         struct list_head freeme;
67 };
68
69 static struct workqueue_struct *nfsd_filecache_wq __read_mostly;
70
71 static struct kmem_cache                *nfsd_file_slab;
72 static struct kmem_cache                *nfsd_file_mark_slab;
73 static struct list_lru                  nfsd_file_lru;
74 static unsigned long                    nfsd_file_flags;
75 static struct fsnotify_group            *nfsd_file_fsnotify_group;
76 static struct delayed_work              nfsd_filecache_laundrette;
77 static struct rhashtable                nfsd_file_rhash_tbl
78                                                 ____cacheline_aligned_in_smp;
79
80 enum nfsd_file_lookup_type {
81         NFSD_FILE_KEY_INODE,
82         NFSD_FILE_KEY_FULL,
83 };
84
85 struct nfsd_file_lookup_key {
86         struct inode                    *inode;
87         struct net                      *net;
88         const struct cred               *cred;
89         unsigned char                   need;
90         bool                            gc;
91         enum nfsd_file_lookup_type      type;
92 };
93
94 /*
95  * The returned hash value is based solely on the address of an in-code
96  * inode, a pointer to a slab-allocated object. The entropy in such a
97  * pointer is concentrated in its middle bits.
98  */
99 static u32 nfsd_file_inode_hash(const struct inode *inode, u32 seed)
100 {
101         unsigned long ptr = (unsigned long)inode;
102         u32 k;
103
104         k = ptr >> L1_CACHE_SHIFT;
105         k &= 0x00ffffff;
106         return jhash2(&k, 1, seed);
107 }
108
109 /**
110  * nfsd_file_key_hashfn - Compute the hash value of a lookup key
111  * @data: key on which to compute the hash value
112  * @len: rhash table's key_len parameter (unused)
113  * @seed: rhash table's random seed of the day
114  *
115  * Return value:
116  *   Computed 32-bit hash value
117  */
118 static u32 nfsd_file_key_hashfn(const void *data, u32 len, u32 seed)
119 {
120         const struct nfsd_file_lookup_key *key = data;
121
122         return nfsd_file_inode_hash(key->inode, seed);
123 }
124
125 /**
126  * nfsd_file_obj_hashfn - Compute the hash value of an nfsd_file
127  * @data: object on which to compute the hash value
128  * @len: rhash table's key_len parameter (unused)
129  * @seed: rhash table's random seed of the day
130  *
131  * Return value:
132  *   Computed 32-bit hash value
133  */
134 static u32 nfsd_file_obj_hashfn(const void *data, u32 len, u32 seed)
135 {
136         const struct nfsd_file *nf = data;
137
138         return nfsd_file_inode_hash(nf->nf_inode, seed);
139 }
140
141 static bool
142 nfsd_match_cred(const struct cred *c1, const struct cred *c2)
143 {
144         int i;
145
146         if (!uid_eq(c1->fsuid, c2->fsuid))
147                 return false;
148         if (!gid_eq(c1->fsgid, c2->fsgid))
149                 return false;
150         if (c1->group_info == NULL || c2->group_info == NULL)
151                 return c1->group_info == c2->group_info;
152         if (c1->group_info->ngroups != c2->group_info->ngroups)
153                 return false;
154         for (i = 0; i < c1->group_info->ngroups; i++) {
155                 if (!gid_eq(c1->group_info->gid[i], c2->group_info->gid[i]))
156                         return false;
157         }
158         return true;
159 }
160
161 /**
162  * nfsd_file_obj_cmpfn - Match a cache item against search criteria
163  * @arg: search criteria
164  * @ptr: cache item to check
165  *
166  * Return values:
167  *   %0 - Item matches search criteria
168  *   %1 - Item does not match search criteria
169  */
170 static int nfsd_file_obj_cmpfn(struct rhashtable_compare_arg *arg,
171                                const void *ptr)
172 {
173         const struct nfsd_file_lookup_key *key = arg->key;
174         const struct nfsd_file *nf = ptr;
175
176         switch (key->type) {
177         case NFSD_FILE_KEY_INODE:
178                 if (nf->nf_inode != key->inode)
179                         return 1;
180                 break;
181         case NFSD_FILE_KEY_FULL:
182                 if (nf->nf_inode != key->inode)
183                         return 1;
184                 if (nf->nf_may != key->need)
185                         return 1;
186                 if (nf->nf_net != key->net)
187                         return 1;
188                 if (!nfsd_match_cred(nf->nf_cred, key->cred))
189                         return 1;
190                 if (!!test_bit(NFSD_FILE_GC, &nf->nf_flags) != key->gc)
191                         return 1;
192                 if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags) == 0)
193                         return 1;
194                 break;
195         }
196         return 0;
197 }
198
199 static const struct rhashtable_params nfsd_file_rhash_params = {
200         .key_len                = sizeof_field(struct nfsd_file, nf_inode),
201         .key_offset             = offsetof(struct nfsd_file, nf_inode),
202         .head_offset            = offsetof(struct nfsd_file, nf_rhash),
203         .hashfn                 = nfsd_file_key_hashfn,
204         .obj_hashfn             = nfsd_file_obj_hashfn,
205         .obj_cmpfn              = nfsd_file_obj_cmpfn,
206         /* Reduce resizing churn on light workloads */
207         .min_size               = 512,          /* buckets */
208         .automatic_shrinking    = true,
209 };
210
211 static void
212 nfsd_file_schedule_laundrette(void)
213 {
214         if (test_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags))
215                 queue_delayed_work(system_wq, &nfsd_filecache_laundrette,
216                                    NFSD_LAUNDRETTE_DELAY);
217 }
218
219 static void
220 nfsd_file_slab_free(struct rcu_head *rcu)
221 {
222         struct nfsd_file *nf = container_of(rcu, struct nfsd_file, nf_rcu);
223
224         put_cred(nf->nf_cred);
225         kmem_cache_free(nfsd_file_slab, nf);
226 }
227
228 static void
229 nfsd_file_mark_free(struct fsnotify_mark *mark)
230 {
231         struct nfsd_file_mark *nfm = container_of(mark, struct nfsd_file_mark,
232                                                   nfm_mark);
233
234         kmem_cache_free(nfsd_file_mark_slab, nfm);
235 }
236
237 static struct nfsd_file_mark *
238 nfsd_file_mark_get(struct nfsd_file_mark *nfm)
239 {
240         if (!refcount_inc_not_zero(&nfm->nfm_ref))
241                 return NULL;
242         return nfm;
243 }
244
245 static void
246 nfsd_file_mark_put(struct nfsd_file_mark *nfm)
247 {
248         if (refcount_dec_and_test(&nfm->nfm_ref)) {
249                 fsnotify_destroy_mark(&nfm->nfm_mark, nfsd_file_fsnotify_group);
250                 fsnotify_put_mark(&nfm->nfm_mark);
251         }
252 }
253
254 static struct nfsd_file_mark *
255 nfsd_file_mark_find_or_create(struct nfsd_file *nf, struct inode *inode)
256 {
257         int                     err;
258         struct fsnotify_mark    *mark;
259         struct nfsd_file_mark   *nfm = NULL, *new;
260
261         do {
262                 fsnotify_group_lock(nfsd_file_fsnotify_group);
263                 mark = fsnotify_find_mark(&inode->i_fsnotify_marks,
264                                           nfsd_file_fsnotify_group);
265                 if (mark) {
266                         nfm = nfsd_file_mark_get(container_of(mark,
267                                                  struct nfsd_file_mark,
268                                                  nfm_mark));
269                         fsnotify_group_unlock(nfsd_file_fsnotify_group);
270                         if (nfm) {
271                                 fsnotify_put_mark(mark);
272                                 break;
273                         }
274                         /* Avoid soft lockup race with nfsd_file_mark_put() */
275                         fsnotify_destroy_mark(mark, nfsd_file_fsnotify_group);
276                         fsnotify_put_mark(mark);
277                 } else {
278                         fsnotify_group_unlock(nfsd_file_fsnotify_group);
279                 }
280
281                 /* allocate a new nfm */
282                 new = kmem_cache_alloc(nfsd_file_mark_slab, GFP_KERNEL);
283                 if (!new)
284                         return NULL;
285                 fsnotify_init_mark(&new->nfm_mark, nfsd_file_fsnotify_group);
286                 new->nfm_mark.mask = FS_ATTRIB|FS_DELETE_SELF;
287                 refcount_set(&new->nfm_ref, 1);
288
289                 err = fsnotify_add_inode_mark(&new->nfm_mark, inode, 0);
290
291                 /*
292                  * If the add was successful, then return the object.
293                  * Otherwise, we need to put the reference we hold on the
294                  * nfm_mark. The fsnotify code will take a reference and put
295                  * it on failure, so we can't just free it directly. It's also
296                  * not safe to call fsnotify_destroy_mark on it as the
297                  * mark->group will be NULL. Thus, we can't let the nfm_ref
298                  * counter drive the destruction at this point.
299                  */
300                 if (likely(!err))
301                         nfm = new;
302                 else
303                         fsnotify_put_mark(&new->nfm_mark);
304         } while (unlikely(err == -EEXIST));
305
306         return nfm;
307 }
308
309 static struct nfsd_file *
310 nfsd_file_alloc(struct nfsd_file_lookup_key *key, unsigned int may)
311 {
312         struct nfsd_file *nf;
313
314         nf = kmem_cache_alloc(nfsd_file_slab, GFP_KERNEL);
315         if (nf) {
316                 INIT_LIST_HEAD(&nf->nf_lru);
317                 nf->nf_birthtime = ktime_get();
318                 nf->nf_file = NULL;
319                 nf->nf_cred = get_current_cred();
320                 nf->nf_net = key->net;
321                 nf->nf_flags = 0;
322                 __set_bit(NFSD_FILE_HASHED, &nf->nf_flags);
323                 __set_bit(NFSD_FILE_PENDING, &nf->nf_flags);
324                 if (key->gc)
325                         __set_bit(NFSD_FILE_GC, &nf->nf_flags);
326                 nf->nf_inode = key->inode;
327                 refcount_set(&nf->nf_ref, 1);
328                 nf->nf_may = key->need;
329                 nf->nf_mark = NULL;
330         }
331         return nf;
332 }
333
334 /**
335  * nfsd_file_check_write_error - check for writeback errors on a file
336  * @nf: nfsd_file to check for writeback errors
337  *
338  * Check whether a nfsd_file has an unseen error. Reset the write
339  * verifier if so.
340  */
341 static void
342 nfsd_file_check_write_error(struct nfsd_file *nf)
343 {
344         struct file *file = nf->nf_file;
345
346         if ((file->f_mode & FMODE_WRITE) &&
347             filemap_check_wb_err(file->f_mapping, READ_ONCE(file->f_wb_err)))
348                 nfsd_reset_write_verifier(net_generic(nf->nf_net, nfsd_net_id));
349 }
350
351 static void
352 nfsd_file_hash_remove(struct nfsd_file *nf)
353 {
354         trace_nfsd_file_unhash(nf);
355         rhashtable_remove_fast(&nfsd_file_rhash_tbl, &nf->nf_rhash,
356                                nfsd_file_rhash_params);
357 }
358
359 static bool
360 nfsd_file_unhash(struct nfsd_file *nf)
361 {
362         if (test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
363                 nfsd_file_hash_remove(nf);
364                 return true;
365         }
366         return false;
367 }
368
369 static void
370 nfsd_file_free(struct nfsd_file *nf)
371 {
372         s64 age = ktime_to_ms(ktime_sub(ktime_get(), nf->nf_birthtime));
373
374         trace_nfsd_file_free(nf);
375
376         this_cpu_inc(nfsd_file_releases);
377         this_cpu_add(nfsd_file_total_age, age);
378
379         nfsd_file_unhash(nf);
380         if (nf->nf_mark)
381                 nfsd_file_mark_put(nf->nf_mark);
382         if (nf->nf_file) {
383                 get_file(nf->nf_file);
384                 filp_close(nf->nf_file, NULL);
385                 nfsd_file_check_write_error(nf);
386                 fput(nf->nf_file);
387         }
388
389         /*
390          * If this item is still linked via nf_lru, that's a bug.
391          * WARN and leak it to preserve system stability.
392          */
393         if (WARN_ON_ONCE(!list_empty(&nf->nf_lru)))
394                 return;
395
396         call_rcu(&nf->nf_rcu, nfsd_file_slab_free);
397 }
398
399 static bool
400 nfsd_file_check_writeback(struct nfsd_file *nf)
401 {
402         struct file *file = nf->nf_file;
403         struct address_space *mapping;
404
405         if (!file || !(file->f_mode & FMODE_WRITE))
406                 return false;
407         mapping = file->f_mapping;
408         return mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) ||
409                 mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK);
410 }
411
412 static bool nfsd_file_lru_add(struct nfsd_file *nf)
413 {
414         set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
415         if (list_lru_add(&nfsd_file_lru, &nf->nf_lru)) {
416                 trace_nfsd_file_lru_add(nf);
417                 return true;
418         }
419         return false;
420 }
421
422 static bool nfsd_file_lru_remove(struct nfsd_file *nf)
423 {
424         if (list_lru_del(&nfsd_file_lru, &nf->nf_lru)) {
425                 trace_nfsd_file_lru_del(nf);
426                 return true;
427         }
428         return false;
429 }
430
431 struct nfsd_file *
432 nfsd_file_get(struct nfsd_file *nf)
433 {
434         if (nf && refcount_inc_not_zero(&nf->nf_ref))
435                 return nf;
436         return NULL;
437 }
438
439 /**
440  * nfsd_file_put - put the reference to a nfsd_file
441  * @nf: nfsd_file of which to put the reference
442  *
443  * Put a reference to a nfsd_file. In the non-GC case, we just put the
444  * reference immediately. In the GC case, if the reference would be
445  * the last one, the put it on the LRU instead to be cleaned up later.
446  */
447 void
448 nfsd_file_put(struct nfsd_file *nf)
449 {
450         might_sleep();
451         trace_nfsd_file_put(nf);
452
453         if (test_bit(NFSD_FILE_GC, &nf->nf_flags) &&
454             test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
455                 /*
456                  * If this is the last reference (nf_ref == 1), then try to
457                  * transfer it to the LRU.
458                  */
459                 if (refcount_dec_not_one(&nf->nf_ref))
460                         return;
461
462                 /* Try to add it to the LRU.  If that fails, decrement. */
463                 if (nfsd_file_lru_add(nf)) {
464                         /* If it's still hashed, we're done */
465                         if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
466                                 nfsd_file_schedule_laundrette();
467                                 return;
468                         }
469
470                         /*
471                          * We're racing with unhashing, so try to remove it from
472                          * the LRU. If removal fails, then someone else already
473                          * has our reference.
474                          */
475                         if (!nfsd_file_lru_remove(nf))
476                                 return;
477                 }
478         }
479         if (refcount_dec_and_test(&nf->nf_ref))
480                 nfsd_file_free(nf);
481 }
482
483 static void
484 nfsd_file_dispose_list(struct list_head *dispose)
485 {
486         struct nfsd_file *nf;
487
488         while (!list_empty(dispose)) {
489                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
490                 list_del_init(&nf->nf_lru);
491                 nfsd_file_free(nf);
492         }
493 }
494
495 static void
496 nfsd_file_list_remove_disposal(struct list_head *dst,
497                 struct nfsd_fcache_disposal *l)
498 {
499         spin_lock(&l->lock);
500         list_splice_init(&l->freeme, dst);
501         spin_unlock(&l->lock);
502 }
503
504 static void
505 nfsd_file_list_add_disposal(struct list_head *files, struct net *net)
506 {
507         struct nfsd_net *nn = net_generic(net, nfsd_net_id);
508         struct nfsd_fcache_disposal *l = nn->fcache_disposal;
509
510         spin_lock(&l->lock);
511         list_splice_tail_init(files, &l->freeme);
512         spin_unlock(&l->lock);
513         queue_work(nfsd_filecache_wq, &l->work);
514 }
515
516 static void
517 nfsd_file_list_add_pernet(struct list_head *dst, struct list_head *src,
518                 struct net *net)
519 {
520         struct nfsd_file *nf, *tmp;
521
522         list_for_each_entry_safe(nf, tmp, src, nf_lru) {
523                 if (nf->nf_net == net)
524                         list_move_tail(&nf->nf_lru, dst);
525         }
526 }
527
528 static void
529 nfsd_file_dispose_list_delayed(struct list_head *dispose)
530 {
531         LIST_HEAD(list);
532         struct nfsd_file *nf;
533
534         while(!list_empty(dispose)) {
535                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
536                 nfsd_file_list_add_pernet(&list, dispose, nf->nf_net);
537                 nfsd_file_list_add_disposal(&list, nf->nf_net);
538         }
539 }
540
541 /**
542  * nfsd_file_lru_cb - Examine an entry on the LRU list
543  * @item: LRU entry to examine
544  * @lru: controlling LRU
545  * @lock: LRU list lock (unused)
546  * @arg: dispose list
547  *
548  * Return values:
549  *   %LRU_REMOVED: @item was removed from the LRU
550  *   %LRU_ROTATE: @item is to be moved to the LRU tail
551  *   %LRU_SKIP: @item cannot be evicted
552  */
553 static enum lru_status
554 nfsd_file_lru_cb(struct list_head *item, struct list_lru_one *lru,
555                  spinlock_t *lock, void *arg)
556         __releases(lock)
557         __acquires(lock)
558 {
559         struct list_head *head = arg;
560         struct nfsd_file *nf = list_entry(item, struct nfsd_file, nf_lru);
561
562         /* We should only be dealing with GC entries here */
563         WARN_ON_ONCE(!test_bit(NFSD_FILE_GC, &nf->nf_flags));
564
565         /*
566          * Don't throw out files that are still undergoing I/O or
567          * that have uncleared errors pending.
568          */
569         if (nfsd_file_check_writeback(nf)) {
570                 trace_nfsd_file_gc_writeback(nf);
571                 return LRU_SKIP;
572         }
573
574         /* If it was recently added to the list, skip it */
575         if (test_and_clear_bit(NFSD_FILE_REFERENCED, &nf->nf_flags)) {
576                 trace_nfsd_file_gc_referenced(nf);
577                 return LRU_ROTATE;
578         }
579
580         /*
581          * Put the reference held on behalf of the LRU. If it wasn't the last
582          * one, then just remove it from the LRU and ignore it.
583          */
584         if (!refcount_dec_and_test(&nf->nf_ref)) {
585                 trace_nfsd_file_gc_in_use(nf);
586                 list_lru_isolate(lru, &nf->nf_lru);
587                 return LRU_REMOVED;
588         }
589
590         /* Refcount went to zero. Unhash it and queue it to the dispose list */
591         nfsd_file_unhash(nf);
592         list_lru_isolate_move(lru, &nf->nf_lru, head);
593         this_cpu_inc(nfsd_file_evictions);
594         trace_nfsd_file_gc_disposed(nf);
595         return LRU_REMOVED;
596 }
597
598 static void
599 nfsd_file_gc(void)
600 {
601         LIST_HEAD(dispose);
602         unsigned long ret;
603
604         ret = list_lru_walk(&nfsd_file_lru, nfsd_file_lru_cb,
605                             &dispose, list_lru_count(&nfsd_file_lru));
606         trace_nfsd_file_gc_removed(ret, list_lru_count(&nfsd_file_lru));
607         nfsd_file_dispose_list_delayed(&dispose);
608 }
609
610 static void
611 nfsd_file_gc_worker(struct work_struct *work)
612 {
613         nfsd_file_gc();
614         if (list_lru_count(&nfsd_file_lru))
615                 nfsd_file_schedule_laundrette();
616 }
617
618 static unsigned long
619 nfsd_file_lru_count(struct shrinker *s, struct shrink_control *sc)
620 {
621         return list_lru_count(&nfsd_file_lru);
622 }
623
624 static unsigned long
625 nfsd_file_lru_scan(struct shrinker *s, struct shrink_control *sc)
626 {
627         LIST_HEAD(dispose);
628         unsigned long ret;
629
630         ret = list_lru_shrink_walk(&nfsd_file_lru, sc,
631                                    nfsd_file_lru_cb, &dispose);
632         trace_nfsd_file_shrinker_removed(ret, list_lru_count(&nfsd_file_lru));
633         nfsd_file_dispose_list_delayed(&dispose);
634         return ret;
635 }
636
637 static struct shrinker  nfsd_file_shrinker = {
638         .scan_objects = nfsd_file_lru_scan,
639         .count_objects = nfsd_file_lru_count,
640         .seeks = 1,
641 };
642
643 /**
644  * nfsd_file_cond_queue - conditionally unhash and queue a nfsd_file
645  * @nf: nfsd_file to attempt to queue
646  * @dispose: private list to queue successfully-put objects
647  *
648  * Unhash an nfsd_file, try to get a reference to it, and then put that
649  * reference. If it's the last reference, queue it to the dispose list.
650  */
651 static void
652 nfsd_file_cond_queue(struct nfsd_file *nf, struct list_head *dispose)
653         __must_hold(RCU)
654 {
655         int decrement = 1;
656
657         /* If we raced with someone else unhashing, ignore it */
658         if (!nfsd_file_unhash(nf))
659                 return;
660
661         /* If we can't get a reference, ignore it */
662         if (!nfsd_file_get(nf))
663                 return;
664
665         /* Extra decrement if we remove from the LRU */
666         if (nfsd_file_lru_remove(nf))
667                 ++decrement;
668
669         /* If refcount goes to 0, then put on the dispose list */
670         if (refcount_sub_and_test(decrement, &nf->nf_ref)) {
671                 list_add(&nf->nf_lru, dispose);
672                 trace_nfsd_file_closing(nf);
673         }
674 }
675
676 /**
677  * nfsd_file_queue_for_close: try to close out any open nfsd_files for an inode
678  * @inode:   inode on which to close out nfsd_files
679  * @dispose: list on which to gather nfsd_files to close out
680  *
681  * An nfsd_file represents a struct file being held open on behalf of nfsd. An
682  * open file however can block other activity (such as leases), or cause
683  * undesirable behavior (e.g. spurious silly-renames when reexporting NFS).
684  *
685  * This function is intended to find open nfsd_files when this sort of
686  * conflicting access occurs and then attempt to close those files out.
687  *
688  * Populates the dispose list with entries that have already had their
689  * refcounts go to zero. The actual free of an nfsd_file can be expensive,
690  * so we leave it up to the caller whether it wants to wait or not.
691  */
692 static void
693 nfsd_file_queue_for_close(struct inode *inode, struct list_head *dispose)
694 {
695         struct nfsd_file_lookup_key key = {
696                 .type   = NFSD_FILE_KEY_INODE,
697                 .inode  = inode,
698         };
699         struct nfsd_file *nf;
700
701         rcu_read_lock();
702         do {
703                 nf = rhashtable_lookup(&nfsd_file_rhash_tbl, &key,
704                                        nfsd_file_rhash_params);
705                 if (!nf)
706                         break;
707                 nfsd_file_cond_queue(nf, dispose);
708         } while (1);
709         rcu_read_unlock();
710 }
711
712 /**
713  * nfsd_file_close_inode - attempt a delayed close of a nfsd_file
714  * @inode: inode of the file to attempt to remove
715  *
716  * Close out any open nfsd_files that can be reaped for @inode. The
717  * actual freeing is deferred to the dispose_list_delayed infrastructure.
718  *
719  * This is used by the fsnotify callbacks and setlease notifier.
720  */
721 static void
722 nfsd_file_close_inode(struct inode *inode)
723 {
724         LIST_HEAD(dispose);
725
726         nfsd_file_queue_for_close(inode, &dispose);
727         nfsd_file_dispose_list_delayed(&dispose);
728 }
729
730 /**
731  * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
732  * @inode: inode of the file to attempt to remove
733  *
734  * Close out any open nfsd_files that can be reaped for @inode. The
735  * nfsd_files are closed out synchronously.
736  *
737  * This is called from nfsd_rename and nfsd_unlink to avoid silly-renames
738  * when reexporting NFS.
739  */
740 void
741 nfsd_file_close_inode_sync(struct inode *inode)
742 {
743         struct nfsd_file *nf;
744         LIST_HEAD(dispose);
745
746         trace_nfsd_file_close(inode);
747
748         nfsd_file_queue_for_close(inode, &dispose);
749         while (!list_empty(&dispose)) {
750                 nf = list_first_entry(&dispose, struct nfsd_file, nf_lru);
751                 list_del_init(&nf->nf_lru);
752                 nfsd_file_free(nf);
753         }
754         flush_delayed_fput();
755 }
756
757 /**
758  * nfsd_file_delayed_close - close unused nfsd_files
759  * @work: dummy
760  *
761  * Walk the LRU list and destroy any entries that have not been used since
762  * the last scan.
763  */
764 static void
765 nfsd_file_delayed_close(struct work_struct *work)
766 {
767         LIST_HEAD(head);
768         struct nfsd_fcache_disposal *l = container_of(work,
769                         struct nfsd_fcache_disposal, work);
770
771         nfsd_file_list_remove_disposal(&head, l);
772         nfsd_file_dispose_list(&head);
773 }
774
775 static int
776 nfsd_file_lease_notifier_call(struct notifier_block *nb, unsigned long arg,
777                             void *data)
778 {
779         struct file_lock *fl = data;
780
781         /* Only close files for F_SETLEASE leases */
782         if (fl->fl_flags & FL_LEASE)
783                 nfsd_file_close_inode(file_inode(fl->fl_file));
784         return 0;
785 }
786
787 static struct notifier_block nfsd_file_lease_notifier = {
788         .notifier_call = nfsd_file_lease_notifier_call,
789 };
790
791 static int
792 nfsd_file_fsnotify_handle_event(struct fsnotify_mark *mark, u32 mask,
793                                 struct inode *inode, struct inode *dir,
794                                 const struct qstr *name, u32 cookie)
795 {
796         if (WARN_ON_ONCE(!inode))
797                 return 0;
798
799         trace_nfsd_file_fsnotify_handle_event(inode, mask);
800
801         /* Should be no marks on non-regular files */
802         if (!S_ISREG(inode->i_mode)) {
803                 WARN_ON_ONCE(1);
804                 return 0;
805         }
806
807         /* don't close files if this was not the last link */
808         if (mask & FS_ATTRIB) {
809                 if (inode->i_nlink)
810                         return 0;
811         }
812
813         nfsd_file_close_inode(inode);
814         return 0;
815 }
816
817
818 static const struct fsnotify_ops nfsd_file_fsnotify_ops = {
819         .handle_inode_event = nfsd_file_fsnotify_handle_event,
820         .free_mark = nfsd_file_mark_free,
821 };
822
823 int
824 nfsd_file_cache_init(void)
825 {
826         int ret;
827
828         lockdep_assert_held(&nfsd_mutex);
829         if (test_and_set_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags) == 1)
830                 return 0;
831
832         ret = rhashtable_init(&nfsd_file_rhash_tbl, &nfsd_file_rhash_params);
833         if (ret)
834                 return ret;
835
836         ret = -ENOMEM;
837         nfsd_filecache_wq = alloc_workqueue("nfsd_filecache", 0, 0);
838         if (!nfsd_filecache_wq)
839                 goto out;
840
841         nfsd_file_slab = kmem_cache_create("nfsd_file",
842                                 sizeof(struct nfsd_file), 0, 0, NULL);
843         if (!nfsd_file_slab) {
844                 pr_err("nfsd: unable to create nfsd_file_slab\n");
845                 goto out_err;
846         }
847
848         nfsd_file_mark_slab = kmem_cache_create("nfsd_file_mark",
849                                         sizeof(struct nfsd_file_mark), 0, 0, NULL);
850         if (!nfsd_file_mark_slab) {
851                 pr_err("nfsd: unable to create nfsd_file_mark_slab\n");
852                 goto out_err;
853         }
854
855
856         ret = list_lru_init(&nfsd_file_lru);
857         if (ret) {
858                 pr_err("nfsd: failed to init nfsd_file_lru: %d\n", ret);
859                 goto out_err;
860         }
861
862         ret = register_shrinker(&nfsd_file_shrinker, "nfsd-filecache");
863         if (ret) {
864                 pr_err("nfsd: failed to register nfsd_file_shrinker: %d\n", ret);
865                 goto out_lru;
866         }
867
868         ret = lease_register_notifier(&nfsd_file_lease_notifier);
869         if (ret) {
870                 pr_err("nfsd: unable to register lease notifier: %d\n", ret);
871                 goto out_shrinker;
872         }
873
874         nfsd_file_fsnotify_group = fsnotify_alloc_group(&nfsd_file_fsnotify_ops,
875                                                         FSNOTIFY_GROUP_NOFS);
876         if (IS_ERR(nfsd_file_fsnotify_group)) {
877                 pr_err("nfsd: unable to create fsnotify group: %ld\n",
878                         PTR_ERR(nfsd_file_fsnotify_group));
879                 ret = PTR_ERR(nfsd_file_fsnotify_group);
880                 nfsd_file_fsnotify_group = NULL;
881                 goto out_notifier;
882         }
883
884         INIT_DELAYED_WORK(&nfsd_filecache_laundrette, nfsd_file_gc_worker);
885 out:
886         return ret;
887 out_notifier:
888         lease_unregister_notifier(&nfsd_file_lease_notifier);
889 out_shrinker:
890         unregister_shrinker(&nfsd_file_shrinker);
891 out_lru:
892         list_lru_destroy(&nfsd_file_lru);
893 out_err:
894         kmem_cache_destroy(nfsd_file_slab);
895         nfsd_file_slab = NULL;
896         kmem_cache_destroy(nfsd_file_mark_slab);
897         nfsd_file_mark_slab = NULL;
898         destroy_workqueue(nfsd_filecache_wq);
899         nfsd_filecache_wq = NULL;
900         rhashtable_destroy(&nfsd_file_rhash_tbl);
901         goto out;
902 }
903
904 /**
905  * __nfsd_file_cache_purge: clean out the cache for shutdown
906  * @net: net-namespace to shut down the cache (may be NULL)
907  *
908  * Walk the nfsd_file cache and close out any that match @net. If @net is NULL,
909  * then close out everything. Called when an nfsd instance is being shut down.
910  */
911 static void
912 __nfsd_file_cache_purge(struct net *net)
913 {
914         struct rhashtable_iter iter;
915         struct nfsd_file *nf;
916         LIST_HEAD(dispose);
917
918         rhashtable_walk_enter(&nfsd_file_rhash_tbl, &iter);
919         do {
920                 rhashtable_walk_start(&iter);
921
922                 nf = rhashtable_walk_next(&iter);
923                 while (!IS_ERR_OR_NULL(nf)) {
924                         if (!net || nf->nf_net == net)
925                                 nfsd_file_cond_queue(nf, &dispose);
926                         nf = rhashtable_walk_next(&iter);
927                 }
928
929                 rhashtable_walk_stop(&iter);
930         } while (nf == ERR_PTR(-EAGAIN));
931         rhashtable_walk_exit(&iter);
932
933         nfsd_file_dispose_list(&dispose);
934 }
935
936 static struct nfsd_fcache_disposal *
937 nfsd_alloc_fcache_disposal(void)
938 {
939         struct nfsd_fcache_disposal *l;
940
941         l = kmalloc(sizeof(*l), GFP_KERNEL);
942         if (!l)
943                 return NULL;
944         INIT_WORK(&l->work, nfsd_file_delayed_close);
945         spin_lock_init(&l->lock);
946         INIT_LIST_HEAD(&l->freeme);
947         return l;
948 }
949
950 static void
951 nfsd_free_fcache_disposal(struct nfsd_fcache_disposal *l)
952 {
953         cancel_work_sync(&l->work);
954         nfsd_file_dispose_list(&l->freeme);
955         kfree(l);
956 }
957
958 static void
959 nfsd_free_fcache_disposal_net(struct net *net)
960 {
961         struct nfsd_net *nn = net_generic(net, nfsd_net_id);
962         struct nfsd_fcache_disposal *l = nn->fcache_disposal;
963
964         nfsd_free_fcache_disposal(l);
965 }
966
967 int
968 nfsd_file_cache_start_net(struct net *net)
969 {
970         struct nfsd_net *nn = net_generic(net, nfsd_net_id);
971
972         nn->fcache_disposal = nfsd_alloc_fcache_disposal();
973         return nn->fcache_disposal ? 0 : -ENOMEM;
974 }
975
976 /**
977  * nfsd_file_cache_purge - Remove all cache items associated with @net
978  * @net: target net namespace
979  *
980  */
981 void
982 nfsd_file_cache_purge(struct net *net)
983 {
984         lockdep_assert_held(&nfsd_mutex);
985         if (test_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags) == 1)
986                 __nfsd_file_cache_purge(net);
987 }
988
989 void
990 nfsd_file_cache_shutdown_net(struct net *net)
991 {
992         nfsd_file_cache_purge(net);
993         nfsd_free_fcache_disposal_net(net);
994 }
995
996 void
997 nfsd_file_cache_shutdown(void)
998 {
999         int i;
1000
1001         lockdep_assert_held(&nfsd_mutex);
1002         if (test_and_clear_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags) == 0)
1003                 return;
1004
1005         lease_unregister_notifier(&nfsd_file_lease_notifier);
1006         unregister_shrinker(&nfsd_file_shrinker);
1007         /*
1008          * make sure all callers of nfsd_file_lru_cb are done before
1009          * calling nfsd_file_cache_purge
1010          */
1011         cancel_delayed_work_sync(&nfsd_filecache_laundrette);
1012         __nfsd_file_cache_purge(NULL);
1013         list_lru_destroy(&nfsd_file_lru);
1014         rcu_barrier();
1015         fsnotify_put_group(nfsd_file_fsnotify_group);
1016         nfsd_file_fsnotify_group = NULL;
1017         kmem_cache_destroy(nfsd_file_slab);
1018         nfsd_file_slab = NULL;
1019         fsnotify_wait_marks_destroyed();
1020         kmem_cache_destroy(nfsd_file_mark_slab);
1021         nfsd_file_mark_slab = NULL;
1022         destroy_workqueue(nfsd_filecache_wq);
1023         nfsd_filecache_wq = NULL;
1024         rhashtable_destroy(&nfsd_file_rhash_tbl);
1025
1026         for_each_possible_cpu(i) {
1027                 per_cpu(nfsd_file_cache_hits, i) = 0;
1028                 per_cpu(nfsd_file_acquisitions, i) = 0;
1029                 per_cpu(nfsd_file_releases, i) = 0;
1030                 per_cpu(nfsd_file_total_age, i) = 0;
1031                 per_cpu(nfsd_file_evictions, i) = 0;
1032         }
1033 }
1034
1035 /**
1036  * nfsd_file_is_cached - are there any cached open files for this inode?
1037  * @inode: inode to check
1038  *
1039  * The lookup matches inodes in all net namespaces and is atomic wrt
1040  * nfsd_file_acquire().
1041  *
1042  * Return values:
1043  *   %true: filecache contains at least one file matching this inode
1044  *   %false: filecache contains no files matching this inode
1045  */
1046 bool
1047 nfsd_file_is_cached(struct inode *inode)
1048 {
1049         struct nfsd_file_lookup_key key = {
1050                 .type   = NFSD_FILE_KEY_INODE,
1051                 .inode  = inode,
1052         };
1053         bool ret = false;
1054
1055         if (rhashtable_lookup_fast(&nfsd_file_rhash_tbl, &key,
1056                                    nfsd_file_rhash_params) != NULL)
1057                 ret = true;
1058         trace_nfsd_file_is_cached(inode, (int)ret);
1059         return ret;
1060 }
1061
1062 static __be32
1063 nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
1064                      unsigned int may_flags, struct file *file,
1065                      struct nfsd_file **pnf, bool want_gc)
1066 {
1067         struct nfsd_file_lookup_key key = {
1068                 .type   = NFSD_FILE_KEY_FULL,
1069                 .need   = may_flags & NFSD_FILE_MAY_MASK,
1070                 .net    = SVC_NET(rqstp),
1071                 .gc     = want_gc,
1072         };
1073         bool open_retry = true;
1074         struct nfsd_file *nf;
1075         __be32 status;
1076         int ret;
1077
1078         status = fh_verify(rqstp, fhp, S_IFREG,
1079                                 may_flags|NFSD_MAY_OWNER_OVERRIDE);
1080         if (status != nfs_ok)
1081                 return status;
1082         key.inode = d_inode(fhp->fh_dentry);
1083         key.cred = get_current_cred();
1084
1085 retry:
1086         rcu_read_lock();
1087         nf = rhashtable_lookup(&nfsd_file_rhash_tbl, &key,
1088                                nfsd_file_rhash_params);
1089         nf = nfsd_file_get(nf);
1090         rcu_read_unlock();
1091
1092         if (nf) {
1093                 if (nfsd_file_lru_remove(nf))
1094                         WARN_ON_ONCE(refcount_dec_and_test(&nf->nf_ref));
1095                 goto wait_for_construction;
1096         }
1097
1098         nf = nfsd_file_alloc(&key, may_flags);
1099         if (!nf) {
1100                 status = nfserr_jukebox;
1101                 goto out_status;
1102         }
1103
1104         ret = rhashtable_lookup_insert_key(&nfsd_file_rhash_tbl,
1105                                            &key, &nf->nf_rhash,
1106                                            nfsd_file_rhash_params);
1107         if (likely(ret == 0))
1108                 goto open_file;
1109
1110         nfsd_file_slab_free(&nf->nf_rcu);
1111         nf = NULL;
1112         if (ret == -EEXIST)
1113                 goto retry;
1114         trace_nfsd_file_insert_err(rqstp, key.inode, may_flags, ret);
1115         status = nfserr_jukebox;
1116         goto out_status;
1117
1118 wait_for_construction:
1119         wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
1120
1121         /* Did construction of this file fail? */
1122         if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
1123                 trace_nfsd_file_cons_err(rqstp, key.inode, may_flags, nf);
1124                 if (!open_retry) {
1125                         status = nfserr_jukebox;
1126                         goto out;
1127                 }
1128                 open_retry = false;
1129                 if (refcount_dec_and_test(&nf->nf_ref))
1130                         nfsd_file_free(nf);
1131                 goto retry;
1132         }
1133
1134         this_cpu_inc(nfsd_file_cache_hits);
1135
1136         status = nfserrno(nfsd_open_break_lease(file_inode(nf->nf_file), may_flags));
1137 out:
1138         if (status == nfs_ok) {
1139                 this_cpu_inc(nfsd_file_acquisitions);
1140                 nfsd_file_check_write_error(nf);
1141                 *pnf = nf;
1142         } else {
1143                 if (refcount_dec_and_test(&nf->nf_ref))
1144                         nfsd_file_free(nf);
1145                 nf = NULL;
1146         }
1147
1148 out_status:
1149         put_cred(key.cred);
1150         trace_nfsd_file_acquire(rqstp, key.inode, may_flags, nf, status);
1151         return status;
1152
1153 open_file:
1154         trace_nfsd_file_alloc(nf);
1155         nf->nf_mark = nfsd_file_mark_find_or_create(nf, key.inode);
1156         if (nf->nf_mark) {
1157                 if (file) {
1158                         get_file(file);
1159                         nf->nf_file = file;
1160                         status = nfs_ok;
1161                         trace_nfsd_file_opened(nf, status);
1162                 } else {
1163                         status = nfsd_open_verified(rqstp, fhp, may_flags,
1164                                                     &nf->nf_file);
1165                         trace_nfsd_file_open(nf, status);
1166                 }
1167         } else
1168                 status = nfserr_jukebox;
1169         /*
1170          * If construction failed, or we raced with a call to unlink()
1171          * then unhash.
1172          */
1173         if (status == nfs_ok && key.inode->i_nlink == 0)
1174                 status = nfserr_jukebox;
1175         if (status != nfs_ok)
1176                 nfsd_file_unhash(nf);
1177         clear_bit_unlock(NFSD_FILE_PENDING, &nf->nf_flags);
1178         smp_mb__after_atomic();
1179         wake_up_bit(&nf->nf_flags, NFSD_FILE_PENDING);
1180         goto out;
1181 }
1182
1183 /**
1184  * nfsd_file_acquire_gc - Get a struct nfsd_file with an open file
1185  * @rqstp: the RPC transaction being executed
1186  * @fhp: the NFS filehandle of the file to be opened
1187  * @may_flags: NFSD_MAY_ settings for the file
1188  * @pnf: OUT: new or found "struct nfsd_file" object
1189  *
1190  * The nfsd_file object returned by this API is reference-counted
1191  * and garbage-collected. The object is retained for a few
1192  * seconds after the final nfsd_file_put() in case the caller
1193  * wants to re-use it.
1194  *
1195  * Returns nfs_ok and sets @pnf on success; otherwise an nfsstat in
1196  * network byte order is returned.
1197  */
1198 __be32
1199 nfsd_file_acquire_gc(struct svc_rqst *rqstp, struct svc_fh *fhp,
1200                      unsigned int may_flags, struct nfsd_file **pnf)
1201 {
1202         return nfsd_file_do_acquire(rqstp, fhp, may_flags, NULL, pnf, true);
1203 }
1204
1205 /**
1206  * nfsd_file_acquire - Get a struct nfsd_file with an open file
1207  * @rqstp: the RPC transaction being executed
1208  * @fhp: the NFS filehandle of the file to be opened
1209  * @may_flags: NFSD_MAY_ settings for the file
1210  * @pnf: OUT: new or found "struct nfsd_file" object
1211  *
1212  * The nfsd_file_object returned by this API is reference-counted
1213  * but not garbage-collected. The object is unhashed after the
1214  * final nfsd_file_put().
1215  *
1216  * Returns nfs_ok and sets @pnf on success; otherwise an nfsstat in
1217  * network byte order is returned.
1218  */
1219 __be32
1220 nfsd_file_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
1221                   unsigned int may_flags, struct nfsd_file **pnf)
1222 {
1223         return nfsd_file_do_acquire(rqstp, fhp, may_flags, NULL, pnf, false);
1224 }
1225
1226 /**
1227  * nfsd_file_acquire_opened - Get a struct nfsd_file using existing open file
1228  * @rqstp: the RPC transaction being executed
1229  * @fhp: the NFS filehandle of the file just created
1230  * @may_flags: NFSD_MAY_ settings for the file
1231  * @file: cached, already-open file (may be NULL)
1232  * @pnf: OUT: new or found "struct nfsd_file" object
1233  *
1234  * Acquire a nfsd_file object that is not GC'ed. If one doesn't already exist,
1235  * and @file is non-NULL, use it to instantiate a new nfsd_file instead of
1236  * opening a new one.
1237  *
1238  * Returns nfs_ok and sets @pnf on success; otherwise an nfsstat in
1239  * network byte order is returned.
1240  */
1241 __be32
1242 nfsd_file_acquire_opened(struct svc_rqst *rqstp, struct svc_fh *fhp,
1243                          unsigned int may_flags, struct file *file,
1244                          struct nfsd_file **pnf)
1245 {
1246         return nfsd_file_do_acquire(rqstp, fhp, may_flags, file, pnf, false);
1247 }
1248
1249 /*
1250  * Note that fields may be added, removed or reordered in the future. Programs
1251  * scraping this file for info should test the labels to ensure they're
1252  * getting the correct field.
1253  */
1254 int nfsd_file_cache_stats_show(struct seq_file *m, void *v)
1255 {
1256         unsigned long releases = 0, evictions = 0;
1257         unsigned long hits = 0, acquisitions = 0;
1258         unsigned int i, count = 0, buckets = 0;
1259         unsigned long lru = 0, total_age = 0;
1260
1261         /* Serialize with server shutdown */
1262         mutex_lock(&nfsd_mutex);
1263         if (test_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags) == 1) {
1264                 struct bucket_table *tbl;
1265                 struct rhashtable *ht;
1266
1267                 lru = list_lru_count(&nfsd_file_lru);
1268
1269                 rcu_read_lock();
1270                 ht = &nfsd_file_rhash_tbl;
1271                 count = atomic_read(&ht->nelems);
1272                 tbl = rht_dereference_rcu(ht->tbl, ht);
1273                 buckets = tbl->size;
1274                 rcu_read_unlock();
1275         }
1276         mutex_unlock(&nfsd_mutex);
1277
1278         for_each_possible_cpu(i) {
1279                 hits += per_cpu(nfsd_file_cache_hits, i);
1280                 acquisitions += per_cpu(nfsd_file_acquisitions, i);
1281                 releases += per_cpu(nfsd_file_releases, i);
1282                 total_age += per_cpu(nfsd_file_total_age, i);
1283                 evictions += per_cpu(nfsd_file_evictions, i);
1284         }
1285
1286         seq_printf(m, "total entries: %u\n", count);
1287         seq_printf(m, "hash buckets:  %u\n", buckets);
1288         seq_printf(m, "lru entries:   %lu\n", lru);
1289         seq_printf(m, "cache hits:    %lu\n", hits);
1290         seq_printf(m, "acquisitions:  %lu\n", acquisitions);
1291         seq_printf(m, "releases:      %lu\n", releases);
1292         seq_printf(m, "evictions:     %lu\n", evictions);
1293         if (releases)
1294                 seq_printf(m, "mean age (ms): %ld\n", total_age / releases);
1295         else
1296                 seq_printf(m, "mean age (ms): -\n");
1297         return 0;
1298 }