Loading...
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|NFSD_MAY_LOCALIO)
56
57static DEFINE_PER_CPU(unsigned long, nfsd_file_cache_hits);
58static DEFINE_PER_CPU(unsigned long, nfsd_file_acquisitions);
59static DEFINE_PER_CPU(unsigned long, nfsd_file_allocations);
60static DEFINE_PER_CPU(unsigned long, nfsd_file_releases);
61static DEFINE_PER_CPU(unsigned long, nfsd_file_total_age);
62static DEFINE_PER_CPU(unsigned long, nfsd_file_evictions);
63
64struct nfsd_fcache_disposal {
65 spinlock_t lock;
66 struct list_head freeme;
67};
68
69static struct kmem_cache *nfsd_file_slab;
70static struct kmem_cache *nfsd_file_mark_slab;
71static struct list_lru nfsd_file_lru;
72static unsigned long nfsd_file_flags;
73static struct fsnotify_group *nfsd_file_fsnotify_group;
74static struct delayed_work nfsd_filecache_laundrette;
75static struct rhltable nfsd_file_rhltable
76 ____cacheline_aligned_in_smp;
77
78static bool
79nfsd_match_cred(const struct cred *c1, const struct cred *c2)
80{
81 int i;
82
83 if (!uid_eq(c1->fsuid, c2->fsuid))
84 return false;
85 if (!gid_eq(c1->fsgid, c2->fsgid))
86 return false;
87 if (c1->group_info == NULL || c2->group_info == NULL)
88 return c1->group_info == c2->group_info;
89 if (c1->group_info->ngroups != c2->group_info->ngroups)
90 return false;
91 for (i = 0; i < c1->group_info->ngroups; i++) {
92 if (!gid_eq(c1->group_info->gid[i], c2->group_info->gid[i]))
93 return false;
94 }
95 return true;
96}
97
98static const struct rhashtable_params nfsd_file_rhash_params = {
99 .key_len = sizeof_field(struct nfsd_file, nf_inode),
100 .key_offset = offsetof(struct nfsd_file, nf_inode),
101 .head_offset = offsetof(struct nfsd_file, nf_rlist),
102
103 /*
104 * Start with a single page hash table to reduce resizing churn
105 * on light workloads.
106 */
107 .min_size = 256,
108 .automatic_shrinking = true,
109};
110
111static void
112nfsd_file_schedule_laundrette(void)
113{
114 if (test_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags))
115 queue_delayed_work(system_unbound_wq, &nfsd_filecache_laundrette,
116 NFSD_LAUNDRETTE_DELAY);
117}
118
119static void
120nfsd_file_slab_free(struct rcu_head *rcu)
121{
122 struct nfsd_file *nf = container_of(rcu, struct nfsd_file, nf_rcu);
123
124 put_cred(nf->nf_cred);
125 kmem_cache_free(nfsd_file_slab, nf);
126}
127
128static void
129nfsd_file_mark_free(struct fsnotify_mark *mark)
130{
131 struct nfsd_file_mark *nfm = container_of(mark, struct nfsd_file_mark,
132 nfm_mark);
133
134 kmem_cache_free(nfsd_file_mark_slab, nfm);
135}
136
137static struct nfsd_file_mark *
138nfsd_file_mark_get(struct nfsd_file_mark *nfm)
139{
140 if (!refcount_inc_not_zero(&nfm->nfm_ref))
141 return NULL;
142 return nfm;
143}
144
145static void
146nfsd_file_mark_put(struct nfsd_file_mark *nfm)
147{
148 if (refcount_dec_and_test(&nfm->nfm_ref)) {
149 fsnotify_destroy_mark(&nfm->nfm_mark, nfsd_file_fsnotify_group);
150 fsnotify_put_mark(&nfm->nfm_mark);
151 }
152}
153
154static struct nfsd_file_mark *
155nfsd_file_mark_find_or_create(struct inode *inode)
156{
157 int err;
158 struct fsnotify_mark *mark;
159 struct nfsd_file_mark *nfm = NULL, *new;
160
161 do {
162 fsnotify_group_lock(nfsd_file_fsnotify_group);
163 mark = fsnotify_find_inode_mark(inode,
164 nfsd_file_fsnotify_group);
165 if (mark) {
166 nfm = nfsd_file_mark_get(container_of(mark,
167 struct nfsd_file_mark,
168 nfm_mark));
169 fsnotify_group_unlock(nfsd_file_fsnotify_group);
170 if (nfm) {
171 fsnotify_put_mark(mark);
172 break;
173 }
174 /* Avoid soft lockup race with nfsd_file_mark_put() */
175 fsnotify_destroy_mark(mark, nfsd_file_fsnotify_group);
176 fsnotify_put_mark(mark);
177 } else {
178 fsnotify_group_unlock(nfsd_file_fsnotify_group);
179 }
180
181 /* allocate a new nfm */
182 new = kmem_cache_alloc(nfsd_file_mark_slab, GFP_KERNEL);
183 if (!new)
184 return NULL;
185 fsnotify_init_mark(&new->nfm_mark, nfsd_file_fsnotify_group);
186 new->nfm_mark.mask = FS_ATTRIB|FS_DELETE_SELF;
187 refcount_set(&new->nfm_ref, 1);
188
189 err = fsnotify_add_inode_mark(&new->nfm_mark, inode, 0);
190
191 /*
192 * If the add was successful, then return the object.
193 * Otherwise, we need to put the reference we hold on the
194 * nfm_mark. The fsnotify code will take a reference and put
195 * it on failure, so we can't just free it directly. It's also
196 * not safe to call fsnotify_destroy_mark on it as the
197 * mark->group will be NULL. Thus, we can't let the nfm_ref
198 * counter drive the destruction at this point.
199 */
200 if (likely(!err))
201 nfm = new;
202 else
203 fsnotify_put_mark(&new->nfm_mark);
204 } while (unlikely(err == -EEXIST));
205
206 return nfm;
207}
208
209static struct nfsd_file *
210nfsd_file_alloc(struct net *net, struct inode *inode, unsigned char need,
211 bool want_gc)
212{
213 struct nfsd_file *nf;
214
215 nf = kmem_cache_alloc(nfsd_file_slab, GFP_KERNEL);
216 if (unlikely(!nf))
217 return NULL;
218
219 this_cpu_inc(nfsd_file_allocations);
220 INIT_LIST_HEAD(&nf->nf_lru);
221 INIT_LIST_HEAD(&nf->nf_gc);
222 nf->nf_birthtime = ktime_get();
223 nf->nf_file = NULL;
224 nf->nf_cred = get_current_cred();
225 nf->nf_net = net;
226 nf->nf_flags = want_gc ?
227 BIT(NFSD_FILE_HASHED) | BIT(NFSD_FILE_PENDING) | BIT(NFSD_FILE_GC) :
228 BIT(NFSD_FILE_HASHED) | BIT(NFSD_FILE_PENDING);
229 nf->nf_inode = inode;
230 refcount_set(&nf->nf_ref, 1);
231 nf->nf_may = need;
232 nf->nf_mark = NULL;
233 return nf;
234}
235
236/**
237 * nfsd_file_check_write_error - check for writeback errors on a file
238 * @nf: nfsd_file to check for writeback errors
239 *
240 * Check whether a nfsd_file has an unseen error. Reset the write
241 * verifier if so.
242 */
243static void
244nfsd_file_check_write_error(struct nfsd_file *nf)
245{
246 struct file *file = nf->nf_file;
247
248 if ((file->f_mode & FMODE_WRITE) &&
249 filemap_check_wb_err(file->f_mapping, READ_ONCE(file->f_wb_err)))
250 nfsd_reset_write_verifier(net_generic(nf->nf_net, nfsd_net_id));
251}
252
253static void
254nfsd_file_hash_remove(struct nfsd_file *nf)
255{
256 trace_nfsd_file_unhash(nf);
257 rhltable_remove(&nfsd_file_rhltable, &nf->nf_rlist,
258 nfsd_file_rhash_params);
259}
260
261static bool
262nfsd_file_unhash(struct nfsd_file *nf)
263{
264 if (test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
265 nfsd_file_hash_remove(nf);
266 return true;
267 }
268 return false;
269}
270
271static void
272nfsd_file_free(struct nfsd_file *nf)
273{
274 s64 age = ktime_to_ms(ktime_sub(ktime_get(), nf->nf_birthtime));
275
276 trace_nfsd_file_free(nf);
277
278 this_cpu_inc(nfsd_file_releases);
279 this_cpu_add(nfsd_file_total_age, age);
280
281 nfsd_file_unhash(nf);
282 if (nf->nf_mark)
283 nfsd_file_mark_put(nf->nf_mark);
284 if (nf->nf_file) {
285 nfsd_file_check_write_error(nf);
286 nfsd_filp_close(nf->nf_file);
287 }
288
289 /*
290 * If this item is still linked via nf_lru, that's a bug.
291 * WARN and leak it to preserve system stability.
292 */
293 if (WARN_ON_ONCE(!list_empty(&nf->nf_lru)))
294 return;
295
296 call_rcu(&nf->nf_rcu, nfsd_file_slab_free);
297}
298
299static bool
300nfsd_file_check_writeback(struct nfsd_file *nf)
301{
302 struct file *file = nf->nf_file;
303 struct address_space *mapping;
304
305 /* File not open for write? */
306 if (!(file->f_mode & FMODE_WRITE))
307 return false;
308
309 /*
310 * Some filesystems (e.g. NFS) flush all dirty data on close.
311 * On others, there is no need to wait for writeback.
312 */
313 if (!(file_inode(file)->i_sb->s_export_op->flags & EXPORT_OP_FLUSH_ON_CLOSE))
314 return false;
315
316 mapping = file->f_mapping;
317 return mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) ||
318 mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK);
319}
320
321
322static bool nfsd_file_lru_add(struct nfsd_file *nf)
323{
324 set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
325 if (list_lru_add_obj(&nfsd_file_lru, &nf->nf_lru)) {
326 trace_nfsd_file_lru_add(nf);
327 return true;
328 }
329 return false;
330}
331
332static bool nfsd_file_lru_remove(struct nfsd_file *nf)
333{
334 if (list_lru_del_obj(&nfsd_file_lru, &nf->nf_lru)) {
335 trace_nfsd_file_lru_del(nf);
336 return true;
337 }
338 return false;
339}
340
341struct nfsd_file *
342nfsd_file_get(struct nfsd_file *nf)
343{
344 if (nf && refcount_inc_not_zero(&nf->nf_ref))
345 return nf;
346 return NULL;
347}
348
349/**
350 * nfsd_file_put - put the reference to a nfsd_file
351 * @nf: nfsd_file of which to put the reference
352 *
353 * Put a reference to a nfsd_file. In the non-GC case, we just put the
354 * reference immediately. In the GC case, if the reference would be
355 * the last one, the put it on the LRU instead to be cleaned up later.
356 */
357void
358nfsd_file_put(struct nfsd_file *nf)
359{
360 might_sleep();
361 trace_nfsd_file_put(nf);
362
363 if (test_bit(NFSD_FILE_GC, &nf->nf_flags) &&
364 test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
365 /*
366 * If this is the last reference (nf_ref == 1), then try to
367 * transfer it to the LRU.
368 */
369 if (refcount_dec_not_one(&nf->nf_ref))
370 return;
371
372 /* Try to add it to the LRU. If that fails, decrement. */
373 if (nfsd_file_lru_add(nf)) {
374 /* If it's still hashed, we're done */
375 if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
376 nfsd_file_schedule_laundrette();
377 return;
378 }
379
380 /*
381 * We're racing with unhashing, so try to remove it from
382 * the LRU. If removal fails, then someone else already
383 * has our reference.
384 */
385 if (!nfsd_file_lru_remove(nf))
386 return;
387 }
388 }
389 if (refcount_dec_and_test(&nf->nf_ref))
390 nfsd_file_free(nf);
391}
392
393/**
394 * nfsd_file_put_local - put nfsd_file reference and arm nfsd_serv_put in caller
395 * @nf: nfsd_file of which to put the reference
396 *
397 * First save the associated net to return to caller, then put
398 * the reference of the nfsd_file.
399 */
400struct net *
401nfsd_file_put_local(struct nfsd_file *nf)
402{
403 struct net *net = nf->nf_net;
404
405 nfsd_file_put(nf);
406 return net;
407}
408
409/**
410 * nfsd_file_file - get the backing file of an nfsd_file
411 * @nf: nfsd_file of which to access the backing file.
412 *
413 * Return backing file for @nf.
414 */
415struct file *
416nfsd_file_file(struct nfsd_file *nf)
417{
418 return nf->nf_file;
419}
420
421static void
422nfsd_file_dispose_list(struct list_head *dispose)
423{
424 struct nfsd_file *nf;
425
426 while (!list_empty(dispose)) {
427 nf = list_first_entry(dispose, struct nfsd_file, nf_gc);
428 list_del_init(&nf->nf_gc);
429 nfsd_file_free(nf);
430 }
431}
432
433/**
434 * nfsd_file_dispose_list_delayed - move list of dead files to net's freeme list
435 * @dispose: list of nfsd_files to be disposed
436 *
437 * Transfers each file to the "freeme" list for its nfsd_net, to eventually
438 * be disposed of by the per-net garbage collector.
439 */
440static void
441nfsd_file_dispose_list_delayed(struct list_head *dispose)
442{
443 while(!list_empty(dispose)) {
444 struct nfsd_file *nf = list_first_entry(dispose,
445 struct nfsd_file, nf_gc);
446 struct nfsd_net *nn = net_generic(nf->nf_net, nfsd_net_id);
447 struct nfsd_fcache_disposal *l = nn->fcache_disposal;
448 struct svc_serv *serv;
449
450 spin_lock(&l->lock);
451 list_move_tail(&nf->nf_gc, &l->freeme);
452 spin_unlock(&l->lock);
453
454 /*
455 * The filecache laundrette is shut down after the
456 * nn->nfsd_serv pointer is cleared, but before the
457 * svc_serv is freed.
458 */
459 serv = nn->nfsd_serv;
460 if (serv)
461 svc_wake_up(serv);
462 }
463}
464
465/**
466 * nfsd_file_net_dispose - deal with nfsd_files waiting to be disposed.
467 * @nn: nfsd_net in which to find files to be disposed.
468 *
469 * When files held open for nfsv3 are removed from the filecache, whether
470 * due to memory pressure or garbage collection, they are queued to
471 * a per-net-ns queue. This function completes the disposal, either
472 * directly or by waking another nfsd thread to help with the work.
473 */
474void nfsd_file_net_dispose(struct nfsd_net *nn)
475{
476 struct nfsd_fcache_disposal *l = nn->fcache_disposal;
477
478 if (!list_empty(&l->freeme)) {
479 LIST_HEAD(dispose);
480 int i;
481
482 spin_lock(&l->lock);
483 for (i = 0; i < 8 && !list_empty(&l->freeme); i++)
484 list_move(l->freeme.next, &dispose);
485 spin_unlock(&l->lock);
486 if (!list_empty(&l->freeme))
487 /* Wake up another thread to share the work
488 * *before* doing any actual disposing.
489 */
490 svc_wake_up(nn->nfsd_serv);
491 nfsd_file_dispose_list(&dispose);
492 }
493}
494
495/**
496 * nfsd_file_lru_cb - Examine an entry on the LRU list
497 * @item: LRU entry to examine
498 * @lru: controlling LRU
499 * @arg: dispose list
500 *
501 * Return values:
502 * %LRU_REMOVED: @item was removed from the LRU
503 * %LRU_ROTATE: @item is to be moved to the LRU tail
504 * %LRU_SKIP: @item cannot be evicted
505 */
506static enum lru_status
507nfsd_file_lru_cb(struct list_head *item, struct list_lru_one *lru,
508 void *arg)
509{
510 struct list_head *head = arg;
511 struct nfsd_file *nf = list_entry(item, struct nfsd_file, nf_lru);
512
513 /* We should only be dealing with GC entries here */
514 WARN_ON_ONCE(!test_bit(NFSD_FILE_GC, &nf->nf_flags));
515
516 /*
517 * Don't throw out files that are still undergoing I/O or
518 * that have uncleared errors pending.
519 */
520 if (nfsd_file_check_writeback(nf)) {
521 trace_nfsd_file_gc_writeback(nf);
522 return LRU_SKIP;
523 }
524
525 /* If it was recently added to the list, skip it */
526 if (test_and_clear_bit(NFSD_FILE_REFERENCED, &nf->nf_flags)) {
527 trace_nfsd_file_gc_referenced(nf);
528 return LRU_ROTATE;
529 }
530
531 /*
532 * Put the reference held on behalf of the LRU. If it wasn't the last
533 * one, then just remove it from the LRU and ignore it.
534 */
535 if (!refcount_dec_and_test(&nf->nf_ref)) {
536 trace_nfsd_file_gc_in_use(nf);
537 list_lru_isolate(lru, &nf->nf_lru);
538 return LRU_REMOVED;
539 }
540
541 /* Refcount went to zero. Unhash it and queue it to the dispose list */
542 nfsd_file_unhash(nf);
543 list_lru_isolate(lru, &nf->nf_lru);
544 list_add(&nf->nf_gc, head);
545 this_cpu_inc(nfsd_file_evictions);
546 trace_nfsd_file_gc_disposed(nf);
547 return LRU_REMOVED;
548}
549
550static void
551nfsd_file_gc(void)
552{
553 LIST_HEAD(dispose);
554 unsigned long ret;
555
556 ret = list_lru_walk(&nfsd_file_lru, nfsd_file_lru_cb,
557 &dispose, list_lru_count(&nfsd_file_lru));
558 trace_nfsd_file_gc_removed(ret, list_lru_count(&nfsd_file_lru));
559 nfsd_file_dispose_list_delayed(&dispose);
560}
561
562static void
563nfsd_file_gc_worker(struct work_struct *work)
564{
565 nfsd_file_gc();
566 if (list_lru_count(&nfsd_file_lru))
567 nfsd_file_schedule_laundrette();
568}
569
570static unsigned long
571nfsd_file_lru_count(struct shrinker *s, struct shrink_control *sc)
572{
573 return list_lru_count(&nfsd_file_lru);
574}
575
576static unsigned long
577nfsd_file_lru_scan(struct shrinker *s, struct shrink_control *sc)
578{
579 LIST_HEAD(dispose);
580 unsigned long ret;
581
582 ret = list_lru_shrink_walk(&nfsd_file_lru, sc,
583 nfsd_file_lru_cb, &dispose);
584 trace_nfsd_file_shrinker_removed(ret, list_lru_count(&nfsd_file_lru));
585 nfsd_file_dispose_list_delayed(&dispose);
586 return ret;
587}
588
589static struct shrinker *nfsd_file_shrinker;
590
591/**
592 * nfsd_file_cond_queue - conditionally unhash and queue a nfsd_file
593 * @nf: nfsd_file to attempt to queue
594 * @dispose: private list to queue successfully-put objects
595 *
596 * Unhash an nfsd_file, try to get a reference to it, and then put that
597 * reference. If it's the last reference, queue it to the dispose list.
598 */
599static void
600nfsd_file_cond_queue(struct nfsd_file *nf, struct list_head *dispose)
601 __must_hold(RCU)
602{
603 int decrement = 1;
604
605 /* If we raced with someone else unhashing, ignore it */
606 if (!nfsd_file_unhash(nf))
607 return;
608
609 /* If we can't get a reference, ignore it */
610 if (!nfsd_file_get(nf))
611 return;
612
613 /* Extra decrement if we remove from the LRU */
614 if (nfsd_file_lru_remove(nf))
615 ++decrement;
616
617 /* If refcount goes to 0, then put on the dispose list */
618 if (refcount_sub_and_test(decrement, &nf->nf_ref)) {
619 list_add(&nf->nf_gc, dispose);
620 trace_nfsd_file_closing(nf);
621 }
622}
623
624/**
625 * nfsd_file_queue_for_close: try to close out any open nfsd_files for an inode
626 * @inode: inode on which to close out nfsd_files
627 * @dispose: list on which to gather nfsd_files to close out
628 *
629 * An nfsd_file represents a struct file being held open on behalf of nfsd.
630 * An open file however can block other activity (such as leases), or cause
631 * undesirable behavior (e.g. spurious silly-renames when reexporting NFS).
632 *
633 * This function is intended to find open nfsd_files when this sort of
634 * conflicting access occurs and then attempt to close those files out.
635 *
636 * Populates the dispose list with entries that have already had their
637 * refcounts go to zero. The actual free of an nfsd_file can be expensive,
638 * so we leave it up to the caller whether it wants to wait or not.
639 */
640static void
641nfsd_file_queue_for_close(struct inode *inode, struct list_head *dispose)
642{
643 struct rhlist_head *tmp, *list;
644 struct nfsd_file *nf;
645
646 rcu_read_lock();
647 list = rhltable_lookup(&nfsd_file_rhltable, &inode,
648 nfsd_file_rhash_params);
649 rhl_for_each_entry_rcu(nf, tmp, list, nf_rlist) {
650 if (!test_bit(NFSD_FILE_GC, &nf->nf_flags))
651 continue;
652 nfsd_file_cond_queue(nf, dispose);
653 }
654 rcu_read_unlock();
655}
656
657/**
658 * nfsd_file_close_inode - attempt a delayed close of a nfsd_file
659 * @inode: inode of the file to attempt to remove
660 *
661 * Close out any open nfsd_files that can be reaped for @inode. The
662 * actual freeing is deferred to the dispose_list_delayed infrastructure.
663 *
664 * This is used by the fsnotify callbacks and setlease notifier.
665 */
666static void
667nfsd_file_close_inode(struct inode *inode)
668{
669 LIST_HEAD(dispose);
670
671 nfsd_file_queue_for_close(inode, &dispose);
672 nfsd_file_dispose_list_delayed(&dispose);
673}
674
675/**
676 * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
677 * @inode: inode of the file to attempt to remove
678 *
679 * Close out any open nfsd_files that can be reaped for @inode. The
680 * nfsd_files are closed out synchronously.
681 *
682 * This is called from nfsd_rename and nfsd_unlink to avoid silly-renames
683 * when reexporting NFS.
684 */
685void
686nfsd_file_close_inode_sync(struct inode *inode)
687{
688 struct nfsd_file *nf;
689 LIST_HEAD(dispose);
690
691 trace_nfsd_file_close(inode);
692
693 nfsd_file_queue_for_close(inode, &dispose);
694 while (!list_empty(&dispose)) {
695 nf = list_first_entry(&dispose, struct nfsd_file, nf_gc);
696 list_del_init(&nf->nf_gc);
697 nfsd_file_free(nf);
698 }
699}
700
701static int
702nfsd_file_lease_notifier_call(struct notifier_block *nb, unsigned long arg,
703 void *data)
704{
705 struct file_lease *fl = data;
706
707 /* Only close files for F_SETLEASE leases */
708 if (fl->c.flc_flags & FL_LEASE)
709 nfsd_file_close_inode(file_inode(fl->c.flc_file));
710 return 0;
711}
712
713static struct notifier_block nfsd_file_lease_notifier = {
714 .notifier_call = nfsd_file_lease_notifier_call,
715};
716
717static int
718nfsd_file_fsnotify_handle_event(struct fsnotify_mark *mark, u32 mask,
719 struct inode *inode, struct inode *dir,
720 const struct qstr *name, u32 cookie)
721{
722 if (WARN_ON_ONCE(!inode))
723 return 0;
724
725 trace_nfsd_file_fsnotify_handle_event(inode, mask);
726
727 /* Should be no marks on non-regular files */
728 if (!S_ISREG(inode->i_mode)) {
729 WARN_ON_ONCE(1);
730 return 0;
731 }
732
733 /* don't close files if this was not the last link */
734 if (mask & FS_ATTRIB) {
735 if (inode->i_nlink)
736 return 0;
737 }
738
739 nfsd_file_close_inode(inode);
740 return 0;
741}
742
743
744static const struct fsnotify_ops nfsd_file_fsnotify_ops = {
745 .handle_inode_event = nfsd_file_fsnotify_handle_event,
746 .free_mark = nfsd_file_mark_free,
747};
748
749int
750nfsd_file_cache_init(void)
751{
752 int ret;
753
754 lockdep_assert_held(&nfsd_mutex);
755 if (test_and_set_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags) == 1)
756 return 0;
757
758 ret = rhltable_init(&nfsd_file_rhltable, &nfsd_file_rhash_params);
759 if (ret)
760 goto out;
761
762 ret = -ENOMEM;
763 nfsd_file_slab = KMEM_CACHE(nfsd_file, 0);
764 if (!nfsd_file_slab) {
765 pr_err("nfsd: unable to create nfsd_file_slab\n");
766 goto out_err;
767 }
768
769 nfsd_file_mark_slab = KMEM_CACHE(nfsd_file_mark, 0);
770 if (!nfsd_file_mark_slab) {
771 pr_err("nfsd: unable to create nfsd_file_mark_slab\n");
772 goto out_err;
773 }
774
775 ret = list_lru_init(&nfsd_file_lru);
776 if (ret) {
777 pr_err("nfsd: failed to init nfsd_file_lru: %d\n", ret);
778 goto out_err;
779 }
780
781 nfsd_file_shrinker = shrinker_alloc(0, "nfsd-filecache");
782 if (!nfsd_file_shrinker) {
783 ret = -ENOMEM;
784 pr_err("nfsd: failed to allocate nfsd_file_shrinker\n");
785 goto out_lru;
786 }
787
788 nfsd_file_shrinker->count_objects = nfsd_file_lru_count;
789 nfsd_file_shrinker->scan_objects = nfsd_file_lru_scan;
790 nfsd_file_shrinker->seeks = 1;
791
792 shrinker_register(nfsd_file_shrinker);
793
794 ret = lease_register_notifier(&nfsd_file_lease_notifier);
795 if (ret) {
796 pr_err("nfsd: unable to register lease notifier: %d\n", ret);
797 goto out_shrinker;
798 }
799
800 nfsd_file_fsnotify_group = fsnotify_alloc_group(&nfsd_file_fsnotify_ops,
801 0);
802 if (IS_ERR(nfsd_file_fsnotify_group)) {
803 pr_err("nfsd: unable to create fsnotify group: %ld\n",
804 PTR_ERR(nfsd_file_fsnotify_group));
805 ret = PTR_ERR(nfsd_file_fsnotify_group);
806 nfsd_file_fsnotify_group = NULL;
807 goto out_notifier;
808 }
809
810 INIT_DELAYED_WORK(&nfsd_filecache_laundrette, nfsd_file_gc_worker);
811out:
812 if (ret)
813 clear_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags);
814 return ret;
815out_notifier:
816 lease_unregister_notifier(&nfsd_file_lease_notifier);
817out_shrinker:
818 shrinker_free(nfsd_file_shrinker);
819out_lru:
820 list_lru_destroy(&nfsd_file_lru);
821out_err:
822 kmem_cache_destroy(nfsd_file_slab);
823 nfsd_file_slab = NULL;
824 kmem_cache_destroy(nfsd_file_mark_slab);
825 nfsd_file_mark_slab = NULL;
826 rhltable_destroy(&nfsd_file_rhltable);
827 goto out;
828}
829
830/**
831 * __nfsd_file_cache_purge: clean out the cache for shutdown
832 * @net: net-namespace to shut down the cache (may be NULL)
833 *
834 * Walk the nfsd_file cache and close out any that match @net. If @net is NULL,
835 * then close out everything. Called when an nfsd instance is being shut down,
836 * and when the exports table is flushed.
837 */
838static void
839__nfsd_file_cache_purge(struct net *net)
840{
841 struct rhashtable_iter iter;
842 struct nfsd_file *nf;
843 LIST_HEAD(dispose);
844
845 rhltable_walk_enter(&nfsd_file_rhltable, &iter);
846 do {
847 rhashtable_walk_start(&iter);
848
849 nf = rhashtable_walk_next(&iter);
850 while (!IS_ERR_OR_NULL(nf)) {
851 if (!net || nf->nf_net == net)
852 nfsd_file_cond_queue(nf, &dispose);
853 nf = rhashtable_walk_next(&iter);
854 }
855
856 rhashtable_walk_stop(&iter);
857 } while (nf == ERR_PTR(-EAGAIN));
858 rhashtable_walk_exit(&iter);
859
860 nfsd_file_dispose_list(&dispose);
861}
862
863static struct nfsd_fcache_disposal *
864nfsd_alloc_fcache_disposal(void)
865{
866 struct nfsd_fcache_disposal *l;
867
868 l = kmalloc(sizeof(*l), GFP_KERNEL);
869 if (!l)
870 return NULL;
871 spin_lock_init(&l->lock);
872 INIT_LIST_HEAD(&l->freeme);
873 return l;
874}
875
876static void
877nfsd_free_fcache_disposal(struct nfsd_fcache_disposal *l)
878{
879 nfsd_file_dispose_list(&l->freeme);
880 kfree(l);
881}
882
883static void
884nfsd_free_fcache_disposal_net(struct net *net)
885{
886 struct nfsd_net *nn = net_generic(net, nfsd_net_id);
887 struct nfsd_fcache_disposal *l = nn->fcache_disposal;
888
889 nfsd_free_fcache_disposal(l);
890}
891
892int
893nfsd_file_cache_start_net(struct net *net)
894{
895 struct nfsd_net *nn = net_generic(net, nfsd_net_id);
896
897 nn->fcache_disposal = nfsd_alloc_fcache_disposal();
898 return nn->fcache_disposal ? 0 : -ENOMEM;
899}
900
901/**
902 * nfsd_file_cache_purge - Remove all cache items associated with @net
903 * @net: target net namespace
904 *
905 */
906void
907nfsd_file_cache_purge(struct net *net)
908{
909 lockdep_assert_held(&nfsd_mutex);
910 if (test_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags) == 1)
911 __nfsd_file_cache_purge(net);
912}
913
914void
915nfsd_file_cache_shutdown_net(struct net *net)
916{
917 nfsd_file_cache_purge(net);
918 nfsd_free_fcache_disposal_net(net);
919}
920
921void
922nfsd_file_cache_shutdown(void)
923{
924 int i;
925
926 lockdep_assert_held(&nfsd_mutex);
927 if (test_and_clear_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags) == 0)
928 return;
929
930 lease_unregister_notifier(&nfsd_file_lease_notifier);
931 shrinker_free(nfsd_file_shrinker);
932 /*
933 * make sure all callers of nfsd_file_lru_cb are done before
934 * calling nfsd_file_cache_purge
935 */
936 cancel_delayed_work_sync(&nfsd_filecache_laundrette);
937 __nfsd_file_cache_purge(NULL);
938 list_lru_destroy(&nfsd_file_lru);
939 rcu_barrier();
940 fsnotify_put_group(nfsd_file_fsnotify_group);
941 nfsd_file_fsnotify_group = NULL;
942 kmem_cache_destroy(nfsd_file_slab);
943 nfsd_file_slab = NULL;
944 fsnotify_wait_marks_destroyed();
945 kmem_cache_destroy(nfsd_file_mark_slab);
946 nfsd_file_mark_slab = NULL;
947 rhltable_destroy(&nfsd_file_rhltable);
948
949 for_each_possible_cpu(i) {
950 per_cpu(nfsd_file_cache_hits, i) = 0;
951 per_cpu(nfsd_file_acquisitions, i) = 0;
952 per_cpu(nfsd_file_allocations, i) = 0;
953 per_cpu(nfsd_file_releases, i) = 0;
954 per_cpu(nfsd_file_total_age, i) = 0;
955 per_cpu(nfsd_file_evictions, i) = 0;
956 }
957}
958
959static struct nfsd_file *
960nfsd_file_lookup_locked(const struct net *net, const struct cred *cred,
961 struct inode *inode, unsigned char need,
962 bool want_gc)
963{
964 struct rhlist_head *tmp, *list;
965 struct nfsd_file *nf;
966
967 list = rhltable_lookup(&nfsd_file_rhltable, &inode,
968 nfsd_file_rhash_params);
969 rhl_for_each_entry_rcu(nf, tmp, list, nf_rlist) {
970 if (nf->nf_may != need)
971 continue;
972 if (nf->nf_net != net)
973 continue;
974 if (!nfsd_match_cred(nf->nf_cred, cred))
975 continue;
976 if (test_bit(NFSD_FILE_GC, &nf->nf_flags) != want_gc)
977 continue;
978 if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags) == 0)
979 continue;
980
981 if (!nfsd_file_get(nf))
982 continue;
983 return nf;
984 }
985 return NULL;
986}
987
988/**
989 * nfsd_file_is_cached - are there any cached open files for this inode?
990 * @inode: inode to check
991 *
992 * The lookup matches inodes in all net namespaces and is atomic wrt
993 * nfsd_file_acquire().
994 *
995 * Return values:
996 * %true: filecache contains at least one file matching this inode
997 * %false: filecache contains no files matching this inode
998 */
999bool
1000nfsd_file_is_cached(struct inode *inode)
1001{
1002 struct rhlist_head *tmp, *list;
1003 struct nfsd_file *nf;
1004 bool ret = false;
1005
1006 rcu_read_lock();
1007 list = rhltable_lookup(&nfsd_file_rhltable, &inode,
1008 nfsd_file_rhash_params);
1009 rhl_for_each_entry_rcu(nf, tmp, list, nf_rlist)
1010 if (test_bit(NFSD_FILE_GC, &nf->nf_flags)) {
1011 ret = true;
1012 break;
1013 }
1014 rcu_read_unlock();
1015
1016 trace_nfsd_file_is_cached(inode, (int)ret);
1017 return ret;
1018}
1019
1020static __be32
1021nfsd_file_do_acquire(struct svc_rqst *rqstp, struct net *net,
1022 struct svc_cred *cred,
1023 struct auth_domain *client,
1024 struct svc_fh *fhp,
1025 unsigned int may_flags, struct file *file,
1026 struct nfsd_file **pnf, bool want_gc)
1027{
1028 unsigned char need = may_flags & NFSD_FILE_MAY_MASK;
1029 struct nfsd_file *new, *nf;
1030 bool stale_retry = true;
1031 bool open_retry = true;
1032 struct inode *inode;
1033 __be32 status;
1034 int ret;
1035
1036retry:
1037 if (rqstp) {
1038 status = fh_verify(rqstp, fhp, S_IFREG,
1039 may_flags|NFSD_MAY_OWNER_OVERRIDE);
1040 } else {
1041 status = fh_verify_local(net, cred, client, fhp, S_IFREG,
1042 may_flags|NFSD_MAY_OWNER_OVERRIDE);
1043 }
1044 if (status != nfs_ok)
1045 return status;
1046 inode = d_inode(fhp->fh_dentry);
1047
1048 rcu_read_lock();
1049 nf = nfsd_file_lookup_locked(net, current_cred(), inode, need, want_gc);
1050 rcu_read_unlock();
1051
1052 if (nf) {
1053 /*
1054 * If the nf is on the LRU then it holds an extra reference
1055 * that must be put if it's removed. It had better not be
1056 * the last one however, since we should hold another.
1057 */
1058 if (nfsd_file_lru_remove(nf))
1059 refcount_dec(&nf->nf_ref);
1060 goto wait_for_construction;
1061 }
1062
1063 new = nfsd_file_alloc(net, inode, need, want_gc);
1064 if (!new) {
1065 status = nfserr_jukebox;
1066 goto out;
1067 }
1068
1069 rcu_read_lock();
1070 spin_lock(&inode->i_lock);
1071 nf = nfsd_file_lookup_locked(net, current_cred(), inode, need, want_gc);
1072 if (unlikely(nf)) {
1073 spin_unlock(&inode->i_lock);
1074 rcu_read_unlock();
1075 nfsd_file_free(new);
1076 goto wait_for_construction;
1077 }
1078 nf = new;
1079 ret = rhltable_insert(&nfsd_file_rhltable, &nf->nf_rlist,
1080 nfsd_file_rhash_params);
1081 spin_unlock(&inode->i_lock);
1082 rcu_read_unlock();
1083 if (likely(ret == 0))
1084 goto open_file;
1085
1086 trace_nfsd_file_insert_err(rqstp, inode, may_flags, ret);
1087 status = nfserr_jukebox;
1088 goto construction_err;
1089
1090wait_for_construction:
1091 wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
1092
1093 /* Did construction of this file fail? */
1094 if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
1095 trace_nfsd_file_cons_err(rqstp, inode, may_flags, nf);
1096 if (!open_retry) {
1097 status = nfserr_jukebox;
1098 goto construction_err;
1099 }
1100 nfsd_file_put(nf);
1101 open_retry = false;
1102 fh_put(fhp);
1103 goto retry;
1104 }
1105 this_cpu_inc(nfsd_file_cache_hits);
1106
1107 status = nfserrno(nfsd_open_break_lease(file_inode(nf->nf_file), may_flags));
1108 if (status != nfs_ok) {
1109 nfsd_file_put(nf);
1110 nf = NULL;
1111 }
1112
1113out:
1114 if (status == nfs_ok) {
1115 this_cpu_inc(nfsd_file_acquisitions);
1116 nfsd_file_check_write_error(nf);
1117 *pnf = nf;
1118 }
1119 trace_nfsd_file_acquire(rqstp, inode, may_flags, nf, status);
1120 return status;
1121
1122open_file:
1123 trace_nfsd_file_alloc(nf);
1124 nf->nf_mark = nfsd_file_mark_find_or_create(inode);
1125 if (nf->nf_mark) {
1126 if (file) {
1127 get_file(file);
1128 nf->nf_file = file;
1129 status = nfs_ok;
1130 trace_nfsd_file_opened(nf, status);
1131 } else {
1132 ret = nfsd_open_verified(fhp, may_flags, &nf->nf_file);
1133 if (ret == -EOPENSTALE && stale_retry) {
1134 stale_retry = false;
1135 nfsd_file_unhash(nf);
1136 clear_and_wake_up_bit(NFSD_FILE_PENDING,
1137 &nf->nf_flags);
1138 if (refcount_dec_and_test(&nf->nf_ref))
1139 nfsd_file_free(nf);
1140 nf = NULL;
1141 fh_put(fhp);
1142 goto retry;
1143 }
1144 status = nfserrno(ret);
1145 trace_nfsd_file_open(nf, status);
1146 }
1147 } else
1148 status = nfserr_jukebox;
1149 /*
1150 * If construction failed, or we raced with a call to unlink()
1151 * then unhash.
1152 */
1153 if (status != nfs_ok || inode->i_nlink == 0)
1154 nfsd_file_unhash(nf);
1155 clear_and_wake_up_bit(NFSD_FILE_PENDING, &nf->nf_flags);
1156 if (status == nfs_ok)
1157 goto out;
1158
1159construction_err:
1160 if (refcount_dec_and_test(&nf->nf_ref))
1161 nfsd_file_free(nf);
1162 nf = NULL;
1163 goto out;
1164}
1165
1166/**
1167 * nfsd_file_acquire_gc - Get a struct nfsd_file with an open file
1168 * @rqstp: the RPC transaction being executed
1169 * @fhp: the NFS filehandle of the file to be opened
1170 * @may_flags: NFSD_MAY_ settings for the file
1171 * @pnf: OUT: new or found "struct nfsd_file" object
1172 *
1173 * The nfsd_file object returned by this API is reference-counted
1174 * and garbage-collected. The object is retained for a few
1175 * seconds after the final nfsd_file_put() in case the caller
1176 * wants to re-use it.
1177 *
1178 * Return values:
1179 * %nfs_ok - @pnf points to an nfsd_file with its reference
1180 * count boosted.
1181 *
1182 * On error, an nfsstat value in network byte order is returned.
1183 */
1184__be32
1185nfsd_file_acquire_gc(struct svc_rqst *rqstp, struct svc_fh *fhp,
1186 unsigned int may_flags, struct nfsd_file **pnf)
1187{
1188 return nfsd_file_do_acquire(rqstp, SVC_NET(rqstp), NULL, NULL,
1189 fhp, may_flags, NULL, pnf, true);
1190}
1191
1192/**
1193 * nfsd_file_acquire - Get a struct nfsd_file with an open file
1194 * @rqstp: the RPC transaction being executed
1195 * @fhp: the NFS filehandle of the file to be opened
1196 * @may_flags: NFSD_MAY_ settings for the file
1197 * @pnf: OUT: new or found "struct nfsd_file" object
1198 *
1199 * The nfsd_file_object returned by this API is reference-counted
1200 * but not garbage-collected. The object is unhashed after the
1201 * final nfsd_file_put().
1202 *
1203 * Return values:
1204 * %nfs_ok - @pnf points to an nfsd_file with its reference
1205 * count boosted.
1206 *
1207 * On error, an nfsstat value in network byte order is returned.
1208 */
1209__be32
1210nfsd_file_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
1211 unsigned int may_flags, struct nfsd_file **pnf)
1212{
1213 return nfsd_file_do_acquire(rqstp, SVC_NET(rqstp), NULL, NULL,
1214 fhp, may_flags, NULL, pnf, false);
1215}
1216
1217/**
1218 * nfsd_file_acquire_local - Get a struct nfsd_file with an open file for localio
1219 * @net: The network namespace in which to perform a lookup
1220 * @cred: the user credential with which to validate access
1221 * @client: the auth_domain for LOCALIO lookup
1222 * @fhp: the NFS filehandle of the file to be opened
1223 * @may_flags: NFSD_MAY_ settings for the file
1224 * @pnf: OUT: new or found "struct nfsd_file" object
1225 *
1226 * This file lookup interface provide access to a file given the
1227 * filehandle and credential. No connection-based authorisation
1228 * is performed and in that way it is quite different to other
1229 * file access mediated by nfsd. It allows a kernel module such as the NFS
1230 * client to reach across network and filesystem namespaces to access
1231 * a file. The security implications of this should be carefully
1232 * considered before use.
1233 *
1234 * The nfsd_file object returned by this API is reference-counted
1235 * and garbage-collected. The object is retained for a few
1236 * seconds after the final nfsd_file_put() in case the caller
1237 * wants to re-use it.
1238 *
1239 * Return values:
1240 * %nfs_ok - @pnf points to an nfsd_file with its reference
1241 * count boosted.
1242 *
1243 * On error, an nfsstat value in network byte order is returned.
1244 */
1245__be32
1246nfsd_file_acquire_local(struct net *net, struct svc_cred *cred,
1247 struct auth_domain *client, struct svc_fh *fhp,
1248 unsigned int may_flags, struct nfsd_file **pnf)
1249{
1250 /*
1251 * Save creds before calling nfsd_file_do_acquire() (which calls
1252 * nfsd_setuser). Important because caller (LOCALIO) is from
1253 * client context.
1254 */
1255 const struct cred *save_cred = get_current_cred();
1256 __be32 beres;
1257
1258 beres = nfsd_file_do_acquire(NULL, net, cred, client,
1259 fhp, may_flags, NULL, pnf, true);
1260 revert_creds(save_cred);
1261 return beres;
1262}
1263
1264/**
1265 * nfsd_file_acquire_opened - Get a struct nfsd_file using existing open file
1266 * @rqstp: the RPC transaction being executed
1267 * @fhp: the NFS filehandle of the file just created
1268 * @may_flags: NFSD_MAY_ settings for the file
1269 * @file: cached, already-open file (may be NULL)
1270 * @pnf: OUT: new or found "struct nfsd_file" object
1271 *
1272 * Acquire a nfsd_file object that is not GC'ed. If one doesn't already exist,
1273 * and @file is non-NULL, use it to instantiate a new nfsd_file instead of
1274 * opening a new one.
1275 *
1276 * Return values:
1277 * %nfs_ok - @pnf points to an nfsd_file with its reference
1278 * count boosted.
1279 *
1280 * On error, an nfsstat value in network byte order is returned.
1281 */
1282__be32
1283nfsd_file_acquire_opened(struct svc_rqst *rqstp, struct svc_fh *fhp,
1284 unsigned int may_flags, struct file *file,
1285 struct nfsd_file **pnf)
1286{
1287 return nfsd_file_do_acquire(rqstp, SVC_NET(rqstp), NULL, NULL,
1288 fhp, may_flags, file, pnf, false);
1289}
1290
1291/*
1292 * Note that fields may be added, removed or reordered in the future. Programs
1293 * scraping this file for info should test the labels to ensure they're
1294 * getting the correct field.
1295 */
1296int nfsd_file_cache_stats_show(struct seq_file *m, void *v)
1297{
1298 unsigned long allocations = 0, releases = 0, evictions = 0;
1299 unsigned long hits = 0, acquisitions = 0;
1300 unsigned int i, count = 0, buckets = 0;
1301 unsigned long lru = 0, total_age = 0;
1302
1303 /* Serialize with server shutdown */
1304 mutex_lock(&nfsd_mutex);
1305 if (test_bit(NFSD_FILE_CACHE_UP, &nfsd_file_flags) == 1) {
1306 struct bucket_table *tbl;
1307 struct rhashtable *ht;
1308
1309 lru = list_lru_count(&nfsd_file_lru);
1310
1311 rcu_read_lock();
1312 ht = &nfsd_file_rhltable.ht;
1313 count = atomic_read(&ht->nelems);
1314 tbl = rht_dereference_rcu(ht->tbl, ht);
1315 buckets = tbl->size;
1316 rcu_read_unlock();
1317 }
1318 mutex_unlock(&nfsd_mutex);
1319
1320 for_each_possible_cpu(i) {
1321 hits += per_cpu(nfsd_file_cache_hits, i);
1322 acquisitions += per_cpu(nfsd_file_acquisitions, i);
1323 allocations += per_cpu(nfsd_file_allocations, i);
1324 releases += per_cpu(nfsd_file_releases, i);
1325 total_age += per_cpu(nfsd_file_total_age, i);
1326 evictions += per_cpu(nfsd_file_evictions, i);
1327 }
1328
1329 seq_printf(m, "total inodes: %u\n", count);
1330 seq_printf(m, "hash buckets: %u\n", buckets);
1331 seq_printf(m, "lru entries: %lu\n", lru);
1332 seq_printf(m, "cache hits: %lu\n", hits);
1333 seq_printf(m, "acquisitions: %lu\n", acquisitions);
1334 seq_printf(m, "allocations: %lu\n", allocations);
1335 seq_printf(m, "releases: %lu\n", releases);
1336 seq_printf(m, "evictions: %lu\n", evictions);
1337 if (releases)
1338 seq_printf(m, "mean age (ms): %ld\n", total_age / releases);
1339 else
1340 seq_printf(m, "mean age (ms): -\n");
1341 return 0;
1342}
1/*
2 * Open file cache.
3 *
4 * (c) 2015 - Jeff Layton <jeff.layton@primarydata.com>
5 */
6
7#include <linux/hash.h>
8#include <linux/slab.h>
9#include <linux/file.h>
10#include <linux/sched.h>
11#include <linux/list_lru.h>
12#include <linux/fsnotify_backend.h>
13#include <linux/fsnotify.h>
14#include <linux/seq_file.h>
15
16#include "vfs.h"
17#include "nfsd.h"
18#include "nfsfh.h"
19#include "netns.h"
20#include "filecache.h"
21#include "trace.h"
22
23#define NFSDDBG_FACILITY NFSDDBG_FH
24
25/* FIXME: dynamically size this for the machine somehow? */
26#define NFSD_FILE_HASH_BITS 12
27#define NFSD_FILE_HASH_SIZE (1 << NFSD_FILE_HASH_BITS)
28#define NFSD_LAUNDRETTE_DELAY (2 * HZ)
29
30#define NFSD_FILE_LRU_RESCAN (0)
31#define NFSD_FILE_SHUTDOWN (1)
32#define NFSD_FILE_LRU_THRESHOLD (4096UL)
33#define NFSD_FILE_LRU_LIMIT (NFSD_FILE_LRU_THRESHOLD << 2)
34
35/* We only care about NFSD_MAY_READ/WRITE for this cache */
36#define NFSD_FILE_MAY_MASK (NFSD_MAY_READ|NFSD_MAY_WRITE)
37
38struct nfsd_fcache_bucket {
39 struct hlist_head nfb_head;
40 spinlock_t nfb_lock;
41 unsigned int nfb_count;
42 unsigned int nfb_maxcount;
43};
44
45static DEFINE_PER_CPU(unsigned long, nfsd_file_cache_hits);
46
47static struct kmem_cache *nfsd_file_slab;
48static struct kmem_cache *nfsd_file_mark_slab;
49static struct nfsd_fcache_bucket *nfsd_file_hashtbl;
50static struct list_lru nfsd_file_lru;
51static long nfsd_file_lru_flags;
52static struct fsnotify_group *nfsd_file_fsnotify_group;
53static atomic_long_t nfsd_filecache_count;
54static struct delayed_work nfsd_filecache_laundrette;
55
56enum nfsd_file_laundrette_ctl {
57 NFSD_FILE_LAUNDRETTE_NOFLUSH = 0,
58 NFSD_FILE_LAUNDRETTE_MAY_FLUSH
59};
60
61static void
62nfsd_file_schedule_laundrette(enum nfsd_file_laundrette_ctl ctl)
63{
64 long count = atomic_long_read(&nfsd_filecache_count);
65
66 if (count == 0 || test_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags))
67 return;
68
69 /* Be more aggressive about scanning if over the threshold */
70 if (count > NFSD_FILE_LRU_THRESHOLD)
71 mod_delayed_work(system_wq, &nfsd_filecache_laundrette, 0);
72 else
73 schedule_delayed_work(&nfsd_filecache_laundrette, NFSD_LAUNDRETTE_DELAY);
74
75 if (ctl == NFSD_FILE_LAUNDRETTE_NOFLUSH)
76 return;
77
78 /* ...and don't delay flushing if we're out of control */
79 if (count >= NFSD_FILE_LRU_LIMIT)
80 flush_delayed_work(&nfsd_filecache_laundrette);
81}
82
83static void
84nfsd_file_slab_free(struct rcu_head *rcu)
85{
86 struct nfsd_file *nf = container_of(rcu, struct nfsd_file, nf_rcu);
87
88 put_cred(nf->nf_cred);
89 kmem_cache_free(nfsd_file_slab, nf);
90}
91
92static void
93nfsd_file_mark_free(struct fsnotify_mark *mark)
94{
95 struct nfsd_file_mark *nfm = container_of(mark, struct nfsd_file_mark,
96 nfm_mark);
97
98 kmem_cache_free(nfsd_file_mark_slab, nfm);
99}
100
101static struct nfsd_file_mark *
102nfsd_file_mark_get(struct nfsd_file_mark *nfm)
103{
104 if (!atomic_inc_not_zero(&nfm->nfm_ref))
105 return NULL;
106 return nfm;
107}
108
109static void
110nfsd_file_mark_put(struct nfsd_file_mark *nfm)
111{
112 if (atomic_dec_and_test(&nfm->nfm_ref)) {
113
114 fsnotify_destroy_mark(&nfm->nfm_mark, nfsd_file_fsnotify_group);
115 fsnotify_put_mark(&nfm->nfm_mark);
116 }
117}
118
119static struct nfsd_file_mark *
120nfsd_file_mark_find_or_create(struct nfsd_file *nf)
121{
122 int err;
123 struct fsnotify_mark *mark;
124 struct nfsd_file_mark *nfm = NULL, *new;
125 struct inode *inode = nf->nf_inode;
126
127 do {
128 mutex_lock(&nfsd_file_fsnotify_group->mark_mutex);
129 mark = fsnotify_find_mark(&inode->i_fsnotify_marks,
130 nfsd_file_fsnotify_group);
131 if (mark) {
132 nfm = nfsd_file_mark_get(container_of(mark,
133 struct nfsd_file_mark,
134 nfm_mark));
135 mutex_unlock(&nfsd_file_fsnotify_group->mark_mutex);
136 fsnotify_put_mark(mark);
137 if (likely(nfm))
138 break;
139 } else
140 mutex_unlock(&nfsd_file_fsnotify_group->mark_mutex);
141
142 /* allocate a new nfm */
143 new = kmem_cache_alloc(nfsd_file_mark_slab, GFP_KERNEL);
144 if (!new)
145 return NULL;
146 fsnotify_init_mark(&new->nfm_mark, nfsd_file_fsnotify_group);
147 new->nfm_mark.mask = FS_ATTRIB|FS_DELETE_SELF;
148 atomic_set(&new->nfm_ref, 1);
149
150 err = fsnotify_add_inode_mark(&new->nfm_mark, inode, 0);
151
152 /*
153 * If the add was successful, then return the object.
154 * Otherwise, we need to put the reference we hold on the
155 * nfm_mark. The fsnotify code will take a reference and put
156 * it on failure, so we can't just free it directly. It's also
157 * not safe to call fsnotify_destroy_mark on it as the
158 * mark->group will be NULL. Thus, we can't let the nfm_ref
159 * counter drive the destruction at this point.
160 */
161 if (likely(!err))
162 nfm = new;
163 else
164 fsnotify_put_mark(&new->nfm_mark);
165 } while (unlikely(err == -EEXIST));
166
167 return nfm;
168}
169
170static struct nfsd_file *
171nfsd_file_alloc(struct inode *inode, unsigned int may, unsigned int hashval,
172 struct net *net)
173{
174 struct nfsd_file *nf;
175
176 nf = kmem_cache_alloc(nfsd_file_slab, GFP_KERNEL);
177 if (nf) {
178 INIT_HLIST_NODE(&nf->nf_node);
179 INIT_LIST_HEAD(&nf->nf_lru);
180 nf->nf_file = NULL;
181 nf->nf_cred = get_current_cred();
182 nf->nf_net = net;
183 nf->nf_flags = 0;
184 nf->nf_inode = inode;
185 nf->nf_hashval = hashval;
186 atomic_set(&nf->nf_ref, 1);
187 nf->nf_may = may & NFSD_FILE_MAY_MASK;
188 if (may & NFSD_MAY_NOT_BREAK_LEASE) {
189 if (may & NFSD_MAY_WRITE)
190 __set_bit(NFSD_FILE_BREAK_WRITE, &nf->nf_flags);
191 if (may & NFSD_MAY_READ)
192 __set_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags);
193 }
194 nf->nf_mark = NULL;
195 trace_nfsd_file_alloc(nf);
196 }
197 return nf;
198}
199
200static bool
201nfsd_file_free(struct nfsd_file *nf)
202{
203 bool flush = false;
204
205 trace_nfsd_file_put_final(nf);
206 if (nf->nf_mark)
207 nfsd_file_mark_put(nf->nf_mark);
208 if (nf->nf_file) {
209 get_file(nf->nf_file);
210 filp_close(nf->nf_file, NULL);
211 fput(nf->nf_file);
212 flush = true;
213 }
214 call_rcu(&nf->nf_rcu, nfsd_file_slab_free);
215 return flush;
216}
217
218static bool
219nfsd_file_check_writeback(struct nfsd_file *nf)
220{
221 struct file *file = nf->nf_file;
222 struct address_space *mapping;
223
224 if (!file || !(file->f_mode & FMODE_WRITE))
225 return false;
226 mapping = file->f_mapping;
227 return mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) ||
228 mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK);
229}
230
231static int
232nfsd_file_check_write_error(struct nfsd_file *nf)
233{
234 struct file *file = nf->nf_file;
235
236 if (!file || !(file->f_mode & FMODE_WRITE))
237 return 0;
238 return filemap_check_wb_err(file->f_mapping, READ_ONCE(file->f_wb_err));
239}
240
241static bool
242nfsd_file_in_use(struct nfsd_file *nf)
243{
244 return nfsd_file_check_writeback(nf) ||
245 nfsd_file_check_write_error(nf);
246}
247
248static void
249nfsd_file_do_unhash(struct nfsd_file *nf)
250{
251 lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
252
253 trace_nfsd_file_unhash(nf);
254
255 if (nfsd_file_check_write_error(nf))
256 nfsd_reset_boot_verifier(net_generic(nf->nf_net, nfsd_net_id));
257 --nfsd_file_hashtbl[nf->nf_hashval].nfb_count;
258 hlist_del_rcu(&nf->nf_node);
259 if (!list_empty(&nf->nf_lru))
260 list_lru_del(&nfsd_file_lru, &nf->nf_lru);
261 atomic_long_dec(&nfsd_filecache_count);
262}
263
264static bool
265nfsd_file_unhash(struct nfsd_file *nf)
266{
267 if (test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
268 nfsd_file_do_unhash(nf);
269 return true;
270 }
271 return false;
272}
273
274/*
275 * Return true if the file was unhashed.
276 */
277static bool
278nfsd_file_unhash_and_release_locked(struct nfsd_file *nf, struct list_head *dispose)
279{
280 lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
281
282 trace_nfsd_file_unhash_and_release_locked(nf);
283 if (!nfsd_file_unhash(nf))
284 return false;
285 /* keep final reference for nfsd_file_lru_dispose */
286 if (atomic_add_unless(&nf->nf_ref, -1, 1))
287 return true;
288
289 list_add(&nf->nf_lru, dispose);
290 return true;
291}
292
293static int
294nfsd_file_put_noref(struct nfsd_file *nf)
295{
296 int count;
297 trace_nfsd_file_put(nf);
298
299 count = atomic_dec_return(&nf->nf_ref);
300 if (!count) {
301 WARN_ON(test_bit(NFSD_FILE_HASHED, &nf->nf_flags));
302 nfsd_file_free(nf);
303 }
304 return count;
305}
306
307void
308nfsd_file_put(struct nfsd_file *nf)
309{
310 bool is_hashed = test_bit(NFSD_FILE_HASHED, &nf->nf_flags) != 0;
311 bool unused = !nfsd_file_in_use(nf);
312
313 set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
314 if (nfsd_file_put_noref(nf) == 1 && is_hashed && unused)
315 nfsd_file_schedule_laundrette(NFSD_FILE_LAUNDRETTE_MAY_FLUSH);
316}
317
318struct nfsd_file *
319nfsd_file_get(struct nfsd_file *nf)
320{
321 if (likely(atomic_inc_not_zero(&nf->nf_ref)))
322 return nf;
323 return NULL;
324}
325
326static void
327nfsd_file_dispose_list(struct list_head *dispose)
328{
329 struct nfsd_file *nf;
330
331 while(!list_empty(dispose)) {
332 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
333 list_del(&nf->nf_lru);
334 nfsd_file_put_noref(nf);
335 }
336}
337
338static void
339nfsd_file_dispose_list_sync(struct list_head *dispose)
340{
341 bool flush = false;
342 struct nfsd_file *nf;
343
344 while(!list_empty(dispose)) {
345 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
346 list_del(&nf->nf_lru);
347 if (!atomic_dec_and_test(&nf->nf_ref))
348 continue;
349 if (nfsd_file_free(nf))
350 flush = true;
351 }
352 if (flush)
353 flush_delayed_fput();
354}
355
356/*
357 * Note this can deadlock with nfsd_file_cache_purge.
358 */
359static enum lru_status
360nfsd_file_lru_cb(struct list_head *item, struct list_lru_one *lru,
361 spinlock_t *lock, void *arg)
362 __releases(lock)
363 __acquires(lock)
364{
365 struct list_head *head = arg;
366 struct nfsd_file *nf = list_entry(item, struct nfsd_file, nf_lru);
367
368 /*
369 * Do a lockless refcount check. The hashtable holds one reference, so
370 * we look to see if anything else has a reference, or if any have
371 * been put since the shrinker last ran. Those don't get unhashed and
372 * released.
373 *
374 * Note that in the put path, we set the flag and then decrement the
375 * counter. Here we check the counter and then test and clear the flag.
376 * That order is deliberate to ensure that we can do this locklessly.
377 */
378 if (atomic_read(&nf->nf_ref) > 1)
379 goto out_skip;
380
381 /*
382 * Don't throw out files that are still undergoing I/O or
383 * that have uncleared errors pending.
384 */
385 if (nfsd_file_check_writeback(nf))
386 goto out_skip;
387
388 if (test_and_clear_bit(NFSD_FILE_REFERENCED, &nf->nf_flags))
389 goto out_rescan;
390
391 if (!test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags))
392 goto out_skip;
393
394 list_lru_isolate_move(lru, &nf->nf_lru, head);
395 return LRU_REMOVED;
396out_rescan:
397 set_bit(NFSD_FILE_LRU_RESCAN, &nfsd_file_lru_flags);
398out_skip:
399 return LRU_SKIP;
400}
401
402static void
403nfsd_file_lru_dispose(struct list_head *head)
404{
405 while(!list_empty(head)) {
406 struct nfsd_file *nf = list_first_entry(head,
407 struct nfsd_file, nf_lru);
408 list_del_init(&nf->nf_lru);
409 spin_lock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
410 nfsd_file_do_unhash(nf);
411 spin_unlock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
412 nfsd_file_put_noref(nf);
413 }
414}
415
416static unsigned long
417nfsd_file_lru_count(struct shrinker *s, struct shrink_control *sc)
418{
419 return list_lru_count(&nfsd_file_lru);
420}
421
422static unsigned long
423nfsd_file_lru_scan(struct shrinker *s, struct shrink_control *sc)
424{
425 LIST_HEAD(head);
426 unsigned long ret;
427
428 ret = list_lru_shrink_walk(&nfsd_file_lru, sc, nfsd_file_lru_cb, &head);
429 nfsd_file_lru_dispose(&head);
430 return ret;
431}
432
433static struct shrinker nfsd_file_shrinker = {
434 .scan_objects = nfsd_file_lru_scan,
435 .count_objects = nfsd_file_lru_count,
436 .seeks = 1,
437};
438
439static void
440__nfsd_file_close_inode(struct inode *inode, unsigned int hashval,
441 struct list_head *dispose)
442{
443 struct nfsd_file *nf;
444 struct hlist_node *tmp;
445
446 spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
447 hlist_for_each_entry_safe(nf, tmp, &nfsd_file_hashtbl[hashval].nfb_head, nf_node) {
448 if (inode == nf->nf_inode)
449 nfsd_file_unhash_and_release_locked(nf, dispose);
450 }
451 spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
452}
453
454/**
455 * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
456 * @inode: inode of the file to attempt to remove
457 *
458 * Walk the whole hash bucket, looking for any files that correspond to "inode".
459 * If any do, then unhash them and put the hashtable reference to them and
460 * destroy any that had their last reference put. Also ensure that any of the
461 * fputs also have their final __fput done as well.
462 */
463void
464nfsd_file_close_inode_sync(struct inode *inode)
465{
466 unsigned int hashval = (unsigned int)hash_long(inode->i_ino,
467 NFSD_FILE_HASH_BITS);
468 LIST_HEAD(dispose);
469
470 __nfsd_file_close_inode(inode, hashval, &dispose);
471 trace_nfsd_file_close_inode_sync(inode, hashval, !list_empty(&dispose));
472 nfsd_file_dispose_list_sync(&dispose);
473}
474
475/**
476 * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
477 * @inode: inode of the file to attempt to remove
478 *
479 * Walk the whole hash bucket, looking for any files that correspond to "inode".
480 * If any do, then unhash them and put the hashtable reference to them and
481 * destroy any that had their last reference put.
482 */
483static void
484nfsd_file_close_inode(struct inode *inode)
485{
486 unsigned int hashval = (unsigned int)hash_long(inode->i_ino,
487 NFSD_FILE_HASH_BITS);
488 LIST_HEAD(dispose);
489
490 __nfsd_file_close_inode(inode, hashval, &dispose);
491 trace_nfsd_file_close_inode(inode, hashval, !list_empty(&dispose));
492 nfsd_file_dispose_list(&dispose);
493}
494
495/**
496 * nfsd_file_delayed_close - close unused nfsd_files
497 * @work: dummy
498 *
499 * Walk the LRU list and close any entries that have not been used since
500 * the last scan.
501 *
502 * Note this can deadlock with nfsd_file_cache_purge.
503 */
504static void
505nfsd_file_delayed_close(struct work_struct *work)
506{
507 LIST_HEAD(head);
508
509 list_lru_walk(&nfsd_file_lru, nfsd_file_lru_cb, &head, LONG_MAX);
510
511 if (test_and_clear_bit(NFSD_FILE_LRU_RESCAN, &nfsd_file_lru_flags))
512 nfsd_file_schedule_laundrette(NFSD_FILE_LAUNDRETTE_NOFLUSH);
513
514 if (!list_empty(&head)) {
515 nfsd_file_lru_dispose(&head);
516 flush_delayed_fput();
517 }
518}
519
520static int
521nfsd_file_lease_notifier_call(struct notifier_block *nb, unsigned long arg,
522 void *data)
523{
524 struct file_lock *fl = data;
525
526 /* Only close files for F_SETLEASE leases */
527 if (fl->fl_flags & FL_LEASE)
528 nfsd_file_close_inode_sync(file_inode(fl->fl_file));
529 return 0;
530}
531
532static struct notifier_block nfsd_file_lease_notifier = {
533 .notifier_call = nfsd_file_lease_notifier_call,
534};
535
536static int
537nfsd_file_fsnotify_handle_event(struct fsnotify_group *group,
538 struct inode *inode,
539 u32 mask, const void *data, int data_type,
540 const struct qstr *file_name, u32 cookie,
541 struct fsnotify_iter_info *iter_info)
542{
543 trace_nfsd_file_fsnotify_handle_event(inode, mask);
544
545 /* Should be no marks on non-regular files */
546 if (!S_ISREG(inode->i_mode)) {
547 WARN_ON_ONCE(1);
548 return 0;
549 }
550
551 /* don't close files if this was not the last link */
552 if (mask & FS_ATTRIB) {
553 if (inode->i_nlink)
554 return 0;
555 }
556
557 nfsd_file_close_inode(inode);
558 return 0;
559}
560
561
562static const struct fsnotify_ops nfsd_file_fsnotify_ops = {
563 .handle_event = nfsd_file_fsnotify_handle_event,
564 .free_mark = nfsd_file_mark_free,
565};
566
567int
568nfsd_file_cache_init(void)
569{
570 int ret = -ENOMEM;
571 unsigned int i;
572
573 clear_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
574
575 if (nfsd_file_hashtbl)
576 return 0;
577
578 nfsd_file_hashtbl = kcalloc(NFSD_FILE_HASH_SIZE,
579 sizeof(*nfsd_file_hashtbl), GFP_KERNEL);
580 if (!nfsd_file_hashtbl) {
581 pr_err("nfsd: unable to allocate nfsd_file_hashtbl\n");
582 goto out_err;
583 }
584
585 nfsd_file_slab = kmem_cache_create("nfsd_file",
586 sizeof(struct nfsd_file), 0, 0, NULL);
587 if (!nfsd_file_slab) {
588 pr_err("nfsd: unable to create nfsd_file_slab\n");
589 goto out_err;
590 }
591
592 nfsd_file_mark_slab = kmem_cache_create("nfsd_file_mark",
593 sizeof(struct nfsd_file_mark), 0, 0, NULL);
594 if (!nfsd_file_mark_slab) {
595 pr_err("nfsd: unable to create nfsd_file_mark_slab\n");
596 goto out_err;
597 }
598
599
600 ret = list_lru_init(&nfsd_file_lru);
601 if (ret) {
602 pr_err("nfsd: failed to init nfsd_file_lru: %d\n", ret);
603 goto out_err;
604 }
605
606 ret = register_shrinker(&nfsd_file_shrinker);
607 if (ret) {
608 pr_err("nfsd: failed to register nfsd_file_shrinker: %d\n", ret);
609 goto out_lru;
610 }
611
612 ret = lease_register_notifier(&nfsd_file_lease_notifier);
613 if (ret) {
614 pr_err("nfsd: unable to register lease notifier: %d\n", ret);
615 goto out_shrinker;
616 }
617
618 nfsd_file_fsnotify_group = fsnotify_alloc_group(&nfsd_file_fsnotify_ops);
619 if (IS_ERR(nfsd_file_fsnotify_group)) {
620 pr_err("nfsd: unable to create fsnotify group: %ld\n",
621 PTR_ERR(nfsd_file_fsnotify_group));
622 nfsd_file_fsnotify_group = NULL;
623 goto out_notifier;
624 }
625
626 for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
627 INIT_HLIST_HEAD(&nfsd_file_hashtbl[i].nfb_head);
628 spin_lock_init(&nfsd_file_hashtbl[i].nfb_lock);
629 }
630
631 INIT_DELAYED_WORK(&nfsd_filecache_laundrette, nfsd_file_delayed_close);
632out:
633 return ret;
634out_notifier:
635 lease_unregister_notifier(&nfsd_file_lease_notifier);
636out_shrinker:
637 unregister_shrinker(&nfsd_file_shrinker);
638out_lru:
639 list_lru_destroy(&nfsd_file_lru);
640out_err:
641 kmem_cache_destroy(nfsd_file_slab);
642 nfsd_file_slab = NULL;
643 kmem_cache_destroy(nfsd_file_mark_slab);
644 nfsd_file_mark_slab = NULL;
645 kfree(nfsd_file_hashtbl);
646 nfsd_file_hashtbl = NULL;
647 goto out;
648}
649
650/*
651 * Note this can deadlock with nfsd_file_lru_cb.
652 */
653void
654nfsd_file_cache_purge(struct net *net)
655{
656 unsigned int i;
657 struct nfsd_file *nf;
658 struct hlist_node *next;
659 LIST_HEAD(dispose);
660 bool del;
661
662 if (!nfsd_file_hashtbl)
663 return;
664
665 for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
666 struct nfsd_fcache_bucket *nfb = &nfsd_file_hashtbl[i];
667
668 spin_lock(&nfb->nfb_lock);
669 hlist_for_each_entry_safe(nf, next, &nfb->nfb_head, nf_node) {
670 if (net && nf->nf_net != net)
671 continue;
672 del = nfsd_file_unhash_and_release_locked(nf, &dispose);
673
674 /*
675 * Deadlock detected! Something marked this entry as
676 * unhased, but hasn't removed it from the hash list.
677 */
678 WARN_ON_ONCE(!del);
679 }
680 spin_unlock(&nfb->nfb_lock);
681 nfsd_file_dispose_list(&dispose);
682 }
683}
684
685void
686nfsd_file_cache_shutdown(void)
687{
688 LIST_HEAD(dispose);
689
690 set_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
691
692 lease_unregister_notifier(&nfsd_file_lease_notifier);
693 unregister_shrinker(&nfsd_file_shrinker);
694 /*
695 * make sure all callers of nfsd_file_lru_cb are done before
696 * calling nfsd_file_cache_purge
697 */
698 cancel_delayed_work_sync(&nfsd_filecache_laundrette);
699 nfsd_file_cache_purge(NULL);
700 list_lru_destroy(&nfsd_file_lru);
701 rcu_barrier();
702 fsnotify_put_group(nfsd_file_fsnotify_group);
703 nfsd_file_fsnotify_group = NULL;
704 kmem_cache_destroy(nfsd_file_slab);
705 nfsd_file_slab = NULL;
706 fsnotify_wait_marks_destroyed();
707 kmem_cache_destroy(nfsd_file_mark_slab);
708 nfsd_file_mark_slab = NULL;
709 kfree(nfsd_file_hashtbl);
710 nfsd_file_hashtbl = NULL;
711}
712
713static bool
714nfsd_match_cred(const struct cred *c1, const struct cred *c2)
715{
716 int i;
717
718 if (!uid_eq(c1->fsuid, c2->fsuid))
719 return false;
720 if (!gid_eq(c1->fsgid, c2->fsgid))
721 return false;
722 if (c1->group_info == NULL || c2->group_info == NULL)
723 return c1->group_info == c2->group_info;
724 if (c1->group_info->ngroups != c2->group_info->ngroups)
725 return false;
726 for (i = 0; i < c1->group_info->ngroups; i++) {
727 if (!gid_eq(c1->group_info->gid[i], c2->group_info->gid[i]))
728 return false;
729 }
730 return true;
731}
732
733static struct nfsd_file *
734nfsd_file_find_locked(struct inode *inode, unsigned int may_flags,
735 unsigned int hashval, struct net *net)
736{
737 struct nfsd_file *nf;
738 unsigned char need = may_flags & NFSD_FILE_MAY_MASK;
739
740 hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
741 nf_node) {
742 if ((need & nf->nf_may) != need)
743 continue;
744 if (nf->nf_inode != inode)
745 continue;
746 if (nf->nf_net != net)
747 continue;
748 if (!nfsd_match_cred(nf->nf_cred, current_cred()))
749 continue;
750 if (nfsd_file_get(nf) != NULL)
751 return nf;
752 }
753 return NULL;
754}
755
756/**
757 * nfsd_file_is_cached - are there any cached open files for this fh?
758 * @inode: inode of the file to check
759 *
760 * Scan the hashtable for open files that match this fh. Returns true if there
761 * are any, and false if not.
762 */
763bool
764nfsd_file_is_cached(struct inode *inode)
765{
766 bool ret = false;
767 struct nfsd_file *nf;
768 unsigned int hashval;
769
770 hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
771
772 rcu_read_lock();
773 hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
774 nf_node) {
775 if (inode == nf->nf_inode) {
776 ret = true;
777 break;
778 }
779 }
780 rcu_read_unlock();
781 trace_nfsd_file_is_cached(inode, hashval, (int)ret);
782 return ret;
783}
784
785__be32
786nfsd_file_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
787 unsigned int may_flags, struct nfsd_file **pnf)
788{
789 __be32 status;
790 struct net *net = SVC_NET(rqstp);
791 struct nfsd_file *nf, *new;
792 struct inode *inode;
793 unsigned int hashval;
794
795 /* FIXME: skip this if fh_dentry is already set? */
796 status = fh_verify(rqstp, fhp, S_IFREG,
797 may_flags|NFSD_MAY_OWNER_OVERRIDE);
798 if (status != nfs_ok)
799 return status;
800
801 inode = d_inode(fhp->fh_dentry);
802 hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
803retry:
804 rcu_read_lock();
805 nf = nfsd_file_find_locked(inode, may_flags, hashval, net);
806 rcu_read_unlock();
807 if (nf)
808 goto wait_for_construction;
809
810 new = nfsd_file_alloc(inode, may_flags, hashval, net);
811 if (!new) {
812 trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags,
813 NULL, nfserr_jukebox);
814 return nfserr_jukebox;
815 }
816
817 spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
818 nf = nfsd_file_find_locked(inode, may_flags, hashval, net);
819 if (nf == NULL)
820 goto open_file;
821 spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
822 nfsd_file_slab_free(&new->nf_rcu);
823
824wait_for_construction:
825 wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
826
827 /* Did construction of this file fail? */
828 if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
829 nfsd_file_put_noref(nf);
830 goto retry;
831 }
832
833 this_cpu_inc(nfsd_file_cache_hits);
834
835 if (!(may_flags & NFSD_MAY_NOT_BREAK_LEASE)) {
836 bool write = (may_flags & NFSD_MAY_WRITE);
837
838 if (test_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags) ||
839 (test_bit(NFSD_FILE_BREAK_WRITE, &nf->nf_flags) && write)) {
840 status = nfserrno(nfsd_open_break_lease(
841 file_inode(nf->nf_file), may_flags));
842 if (status == nfs_ok) {
843 clear_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags);
844 if (write)
845 clear_bit(NFSD_FILE_BREAK_WRITE,
846 &nf->nf_flags);
847 }
848 }
849 }
850out:
851 if (status == nfs_ok) {
852 *pnf = nf;
853 } else {
854 nfsd_file_put(nf);
855 nf = NULL;
856 }
857
858 trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags, nf, status);
859 return status;
860open_file:
861 nf = new;
862 /* Take reference for the hashtable */
863 atomic_inc(&nf->nf_ref);
864 __set_bit(NFSD_FILE_HASHED, &nf->nf_flags);
865 __set_bit(NFSD_FILE_PENDING, &nf->nf_flags);
866 list_lru_add(&nfsd_file_lru, &nf->nf_lru);
867 hlist_add_head_rcu(&nf->nf_node, &nfsd_file_hashtbl[hashval].nfb_head);
868 ++nfsd_file_hashtbl[hashval].nfb_count;
869 nfsd_file_hashtbl[hashval].nfb_maxcount = max(nfsd_file_hashtbl[hashval].nfb_maxcount,
870 nfsd_file_hashtbl[hashval].nfb_count);
871 spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
872 atomic_long_inc(&nfsd_filecache_count);
873
874 nf->nf_mark = nfsd_file_mark_find_or_create(nf);
875 if (nf->nf_mark)
876 status = nfsd_open_verified(rqstp, fhp, S_IFREG,
877 may_flags, &nf->nf_file);
878 else
879 status = nfserr_jukebox;
880 /*
881 * If construction failed, or we raced with a call to unlink()
882 * then unhash.
883 */
884 if (status != nfs_ok || inode->i_nlink == 0) {
885 bool do_free;
886 spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
887 do_free = nfsd_file_unhash(nf);
888 spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
889 if (do_free)
890 nfsd_file_put_noref(nf);
891 }
892 clear_bit_unlock(NFSD_FILE_PENDING, &nf->nf_flags);
893 smp_mb__after_atomic();
894 wake_up_bit(&nf->nf_flags, NFSD_FILE_PENDING);
895 goto out;
896}
897
898/*
899 * Note that fields may be added, removed or reordered in the future. Programs
900 * scraping this file for info should test the labels to ensure they're
901 * getting the correct field.
902 */
903static int nfsd_file_cache_stats_show(struct seq_file *m, void *v)
904{
905 unsigned int i, count = 0, longest = 0;
906 unsigned long hits = 0;
907
908 /*
909 * No need for spinlocks here since we're not terribly interested in
910 * accuracy. We do take the nfsd_mutex simply to ensure that we
911 * don't end up racing with server shutdown
912 */
913 mutex_lock(&nfsd_mutex);
914 if (nfsd_file_hashtbl) {
915 for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
916 count += nfsd_file_hashtbl[i].nfb_count;
917 longest = max(longest, nfsd_file_hashtbl[i].nfb_count);
918 }
919 }
920 mutex_unlock(&nfsd_mutex);
921
922 for_each_possible_cpu(i)
923 hits += per_cpu(nfsd_file_cache_hits, i);
924
925 seq_printf(m, "total entries: %u\n", count);
926 seq_printf(m, "longest chain: %u\n", longest);
927 seq_printf(m, "cache hits: %lu\n", hits);
928 return 0;
929}
930
931int nfsd_file_cache_stats_open(struct inode *inode, struct file *file)
932{
933 return single_open(file, nfsd_file_cache_stats_show, NULL);
934}