fs/nfsd/filecache.c

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