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