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/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_SHUTDOWN                   (1)
  31 #define NFSD_FILE_LRU_THRESHOLD              (4096UL)
  32 #define NFSD_FILE_LRU_LIMIT                  (NFSD_FILE_LRU_THRESHOLD << 2)
  33
  34 /* We only care about NFSD_MAY_READ/WRITE for this cache */
  35 #define NFSD_FILE_MAY_MASK      (NFSD_MAY_READ|NFSD_MAY_WRITE)
  36
  37 struct nfsd_fcache_bucket {
  38         struct hlist_head       nfb_head;
  39         spinlock_t              nfb_lock;
  40         unsigned int            nfb_count;
  41         unsigned int            nfb_maxcount;
  42 };
  43
  44 static DEFINE_PER_CPU(unsigned long, nfsd_file_cache_hits);
  45
  46 struct nfsd_fcache_disposal {
  47         struct list_head list;
  48         struct work_struct work;
  49         struct net *net;
  50         spinlock_t lock;
  51         struct list_head freeme;
  52         struct rcu_head rcu;
  53 };
  54
  55 struct workqueue_struct *nfsd_filecache_wq __read_mostly;
  56
  57 static struct kmem_cache                *nfsd_file_slab;
  58 static struct kmem_cache                *nfsd_file_mark_slab;
  59 static struct nfsd_fcache_bucket        *nfsd_file_hashtbl;
  60 static struct list_lru                  nfsd_file_lru;
  61 static long                             nfsd_file_lru_flags;
  62 static struct fsnotify_group            *nfsd_file_fsnotify_group;
  63 static atomic_long_t                    nfsd_filecache_count;
  64 static struct delayed_work              nfsd_filecache_laundrette;
  65 static DEFINE_SPINLOCK(laundrette_lock);
  66 static LIST_HEAD(laundrettes);
  67
  68 static void nfsd_file_gc(void);
  69
  70 static void
  71 nfsd_file_schedule_laundrette(void)
  72 {
  73         long count = atomic_long_read(&nfsd_filecache_count);
  74
  75         if (count == 0 || test_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags))
  76                 return;
  77
  78         queue_delayed_work(system_wq, &nfsd_filecache_laundrette,
  79                         NFSD_LAUNDRETTE_DELAY);
  80 }
  81
  82 static void
  83 nfsd_file_slab_free(struct rcu_head *rcu)
  84 {
  85         struct nfsd_file *nf = container_of(rcu, struct nfsd_file, nf_rcu);
  86
  87         put_cred(nf->nf_cred);
  88         kmem_cache_free(nfsd_file_slab, nf);
  89 }
  90
  91 static void
  92 nfsd_file_mark_free(struct fsnotify_mark *mark)
  93 {
  94         struct nfsd_file_mark *nfm = container_of(mark, struct nfsd_file_mark,
  95                                                   nfm_mark);
  96
  97         kmem_cache_free(nfsd_file_mark_slab, nfm);
  98 }
  99
 100 static struct nfsd_file_mark *
 101 nfsd_file_mark_get(struct nfsd_file_mark *nfm)
 102 {
 103         if (!atomic_inc_not_zero(&nfm->nfm_ref))
 104                 return NULL;
 105         return nfm;
 106 }
 107
 108 static void
 109 nfsd_file_mark_put(struct nfsd_file_mark *nfm)
 110 {
 111         if (atomic_dec_and_test(&nfm->nfm_ref)) {
 112
 113                 fsnotify_destroy_mark(&nfm->nfm_mark, nfsd_file_fsnotify_group);
 114                 fsnotify_put_mark(&nfm->nfm_mark);
 115         }
 116 }
 117
 118 static struct nfsd_file_mark *
 119 nfsd_file_mark_find_or_create(struct nfsd_file *nf)
 120 {
 121         int                     err;
 122         struct fsnotify_mark    *mark;
 123         struct nfsd_file_mark   *nfm = NULL, *new;
 124         struct inode *inode = nf->nf_inode;
 125
 126         do {
 127                 mutex_lock(&nfsd_file_fsnotify_group->mark_mutex);
 128                 mark = fsnotify_find_mark(&inode->i_fsnotify_marks,
 129                                 nfsd_file_fsnotify_group);
 130                 if (mark) {
 131                         nfm = nfsd_file_mark_get(container_of(mark,
 132                                                  struct nfsd_file_mark,
 133                                                  nfm_mark));
 134                         mutex_unlock(&nfsd_file_fsnotify_group->mark_mutex);
 135                         fsnotify_put_mark(mark);
 136                         if (likely(nfm))
 137                                 break;
 138                 } else
 139                         mutex_unlock(&nfsd_file_fsnotify_group->mark_mutex);
 140
 141                 /* allocate a new nfm */
 142                 new = kmem_cache_alloc(nfsd_file_mark_slab, GFP_KERNEL);
 143                 if (!new)
 144                         return NULL;
 145                 fsnotify_init_mark(&new->nfm_mark, nfsd_file_fsnotify_group);
 146                 new->nfm_mark.mask = FS_ATTRIB|FS_DELETE_SELF;
 147                 atomic_set(&new->nfm_ref, 1);
 148
 149                 err = fsnotify_add_inode_mark(&new->nfm_mark, inode, 0);
 150
 151                 /*
 152                  * If the add was successful, then return the object.
 153                  * Otherwise, we need to put the reference we hold on the
 154                  * nfm_mark. The fsnotify code will take a reference and put
 155                  * it on failure, so we can't just free it directly. It's also
 156                  * not safe to call fsnotify_destroy_mark on it as the
 157                  * mark->group will be NULL. Thus, we can't let the nfm_ref
 158                  * counter drive the destruction at this point.
 159                  */
 160                 if (likely(!err))
 161                         nfm = new;
 162                 else
 163                         fsnotify_put_mark(&new->nfm_mark);
 164         } while (unlikely(err == -EEXIST));
 165
 166         return nfm;
 167 }
 168
 169 static struct nfsd_file *
 170 nfsd_file_alloc(struct inode *inode, unsigned int may, unsigned int hashval,
 171                 struct net *net)
 172 {
 173         struct nfsd_file *nf;
 174
 175         nf = kmem_cache_alloc(nfsd_file_slab, GFP_KERNEL);
 176         if (nf) {
 177                 INIT_HLIST_NODE(&nf->nf_node);
 178                 INIT_LIST_HEAD(&nf->nf_lru);
 179                 nf->nf_file = NULL;
 180                 nf->nf_cred = get_current_cred();
 181                 nf->nf_net = net;
 182                 nf->nf_flags = 0;
 183                 nf->nf_inode = inode;
 184                 nf->nf_hashval = hashval;
 185                 atomic_set(&nf->nf_ref, 1);
 186                 nf->nf_may = may & NFSD_FILE_MAY_MASK;
 187                 if (may & NFSD_MAY_NOT_BREAK_LEASE) {
 188                         if (may & NFSD_MAY_WRITE)
 189                                 __set_bit(NFSD_FILE_BREAK_WRITE, &nf->nf_flags);
 190                         if (may & NFSD_MAY_READ)
 191                                 __set_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags);
 192                 }
 193                 nf->nf_mark = NULL;
 194                 trace_nfsd_file_alloc(nf);
 195         }
 196         return nf;
 197 }
 198
 199 static bool
 200 nfsd_file_free(struct nfsd_file *nf)
 201 {
 202         bool flush = false;
 203
 204         trace_nfsd_file_put_final(nf);
 205         if (nf->nf_mark)
 206                 nfsd_file_mark_put(nf->nf_mark);
 207         if (nf->nf_file) {
 208                 get_file(nf->nf_file);
 209                 filp_close(nf->nf_file, NULL);
 210                 fput(nf->nf_file);
 211                 flush = true;
 212         }
 213         call_rcu(&nf->nf_rcu, nfsd_file_slab_free);
 214         return flush;
 215 }
 216
 217 static bool
 218 nfsd_file_check_writeback(struct nfsd_file *nf)
 219 {
 220         struct file *file = nf->nf_file;
 221         struct address_space *mapping;
 222
 223         if (!file || !(file->f_mode & FMODE_WRITE))
 224                 return false;
 225         mapping = file->f_mapping;
 226         return mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) ||
 227                 mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK);
 228 }
 229
 230 static int
 231 nfsd_file_check_write_error(struct nfsd_file *nf)
 232 {
 233         struct file *file = nf->nf_file;
 234
 235         if (!file || !(file->f_mode & FMODE_WRITE))
 236                 return 0;
 237         return filemap_check_wb_err(file->f_mapping, READ_ONCE(file->f_wb_err));
 238 }
 239
 240 static void
 241 nfsd_file_do_unhash(struct nfsd_file *nf)
 242 {
 243         lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 244
 245         trace_nfsd_file_unhash(nf);
 246
 247         if (nfsd_file_check_write_error(nf))
 248                 nfsd_reset_boot_verifier(net_generic(nf->nf_net, nfsd_net_id));
 249         --nfsd_file_hashtbl[nf->nf_hashval].nfb_count;
 250         hlist_del_rcu(&nf->nf_node);
 251         atomic_long_dec(&nfsd_filecache_count);
 252 }
 253
 254 static bool
 255 nfsd_file_unhash(struct nfsd_file *nf)
 256 {
 257         if (test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
 258                 nfsd_file_do_unhash(nf);
 259                 if (!list_empty(&nf->nf_lru))
 260                         list_lru_del(&nfsd_file_lru, &nf->nf_lru);
 261                 return true;
 262         }
 263         return false;
 264 }
 265
 266 /*
 267  * Return true if the file was unhashed.
 268  */
 269 static bool
 270 nfsd_file_unhash_and_release_locked(struct nfsd_file *nf, struct list_head *dispose)
 271 {
 272         lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 273
 274         trace_nfsd_file_unhash_and_release_locked(nf);
 275         if (!nfsd_file_unhash(nf))
 276                 return false;
 277         /* keep final reference for nfsd_file_lru_dispose */
 278         if (atomic_add_unless(&nf->nf_ref, -1, 1))
 279                 return true;
 280
 281         list_add(&nf->nf_lru, dispose);
 282         return true;
 283 }
 284
 285 static int
 286 nfsd_file_put_noref(struct nfsd_file *nf)
 287 {
 288         int count;
 289         trace_nfsd_file_put(nf);
 290
 291         count = atomic_dec_return(&nf->nf_ref);
 292         if (!count) {
 293                 WARN_ON(test_bit(NFSD_FILE_HASHED, &nf->nf_flags));
 294                 nfsd_file_free(nf);
 295         }
 296         return count;
 297 }
 298
 299 void
 300 nfsd_file_put(struct nfsd_file *nf)
 301 {
 302         bool is_hashed = test_bit(NFSD_FILE_HASHED, &nf->nf_flags) != 0;
 303
 304         set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
 305         if (nfsd_file_put_noref(nf) == 1 && is_hashed)
 306                 nfsd_file_schedule_laundrette();
 307         if (atomic_long_read(&nfsd_filecache_count) >= NFSD_FILE_LRU_LIMIT)
 308                 nfsd_file_gc();
 309 }
 310
 311 struct nfsd_file *
 312 nfsd_file_get(struct nfsd_file *nf)
 313 {
 314         if (likely(atomic_inc_not_zero(&nf->nf_ref)))
 315                 return nf;
 316         return NULL;
 317 }
 318
 319 static void
 320 nfsd_file_dispose_list(struct list_head *dispose)
 321 {
 322         struct nfsd_file *nf;
 323
 324         while(!list_empty(dispose)) {
 325                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 326                 list_del(&nf->nf_lru);
 327                 nfsd_file_put_noref(nf);
 328         }
 329 }
 330
 331 static void
 332 nfsd_file_dispose_list_sync(struct list_head *dispose)
 333 {
 334         bool flush = false;
 335         struct nfsd_file *nf;
 336
 337         while(!list_empty(dispose)) {
 338                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 339                 list_del(&nf->nf_lru);
 340                 if (!atomic_dec_and_test(&nf->nf_ref))
 341                         continue;
 342                 if (nfsd_file_free(nf))
 343                         flush = true;
 344         }
 345         if (flush)
 346                 flush_delayed_fput();
 347 }
 348
 349 static void
 350 nfsd_file_list_remove_disposal(struct list_head *dst,
 351                 struct nfsd_fcache_disposal *l)
 352 {
 353         spin_lock(&l->lock);
 354         list_splice_init(&l->freeme, dst);
 355         spin_unlock(&l->lock);
 356 }
 357
 358 static void
 359 nfsd_file_list_add_disposal(struct list_head *files, struct net *net)
 360 {
 361         struct nfsd_fcache_disposal *l;
 362
 363         rcu_read_lock();
 364         list_for_each_entry_rcu(l, &laundrettes, list) {
 365                 if (l->net == net) {
 366                         spin_lock(&l->lock);
 367                         list_splice_tail_init(files, &l->freeme);
 368                         spin_unlock(&l->lock);
 369                         queue_work(nfsd_filecache_wq, &l->work);
 370                         break;
 371                 }
 372         }
 373         rcu_read_unlock();
 374 }
 375
 376 static void
 377 nfsd_file_list_add_pernet(struct list_head *dst, struct list_head *src,
 378                 struct net *net)
 379 {
 380         struct nfsd_file *nf, *tmp;
 381
 382         list_for_each_entry_safe(nf, tmp, src, nf_lru) {
 383                 if (nf->nf_net == net)
 384                         list_move_tail(&nf->nf_lru, dst);
 385         }
 386 }
 387
 388 static void
 389 nfsd_file_dispose_list_delayed(struct list_head *dispose)
 390 {
 391         LIST_HEAD(list);
 392         struct nfsd_file *nf;
 393
 394         while(!list_empty(dispose)) {
 395                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 396                 nfsd_file_list_add_pernet(&list, dispose, nf->nf_net);
 397                 nfsd_file_list_add_disposal(&list, nf->nf_net);
 398         }
 399 }
 400
 401 /*
 402  * Note this can deadlock with nfsd_file_cache_purge.
 403  */
 404 static enum lru_status
 405 nfsd_file_lru_cb(struct list_head *item, struct list_lru_one *lru,
 406                  spinlock_t *lock, void *arg)
 407         __releases(lock)
 408         __acquires(lock)
 409 {
 410         struct list_head *head = arg;
 411         struct nfsd_file *nf = list_entry(item, struct nfsd_file, nf_lru);
 412
 413         /*
 414          * Do a lockless refcount check. The hashtable holds one reference, so
 415          * we look to see if anything else has a reference, or if any have
 416          * been put since the shrinker last ran. Those don't get unhashed and
 417          * released.
 418          *
 419          * Note that in the put path, we set the flag and then decrement the
 420          * counter. Here we check the counter and then test and clear the flag.
 421          * That order is deliberate to ensure that we can do this locklessly.
 422          */
 423         if (atomic_read(&nf->nf_ref) > 1)
 424                 goto out_skip;
 425
 426         /*
 427          * Don't throw out files that are still undergoing I/O or
 428          * that have uncleared errors pending.
 429          */
 430         if (nfsd_file_check_writeback(nf))
 431                 goto out_skip;
 432
 433         if (test_and_clear_bit(NFSD_FILE_REFERENCED, &nf->nf_flags))
 434                 goto out_skip;
 435
 436         if (!test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags))
 437                 goto out_skip;
 438
 439         list_lru_isolate_move(lru, &nf->nf_lru, head);
 440         return LRU_REMOVED;
 441 out_skip:
 442         return LRU_SKIP;
 443 }
 444
 445 static unsigned long
 446 nfsd_file_lru_walk_list(struct shrink_control *sc)
 447 {
 448         LIST_HEAD(head);
 449         struct nfsd_file *nf;
 450         unsigned long ret;
 451
 452         if (sc)
 453                 ret = list_lru_shrink_walk(&nfsd_file_lru, sc,
 454                                 nfsd_file_lru_cb, &head);
 455         else
 456                 ret = list_lru_walk(&nfsd_file_lru,
 457                                 nfsd_file_lru_cb,
 458                                 &head, LONG_MAX);
 459         list_for_each_entry(nf, &head, nf_lru) {
 460                 spin_lock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 461                 nfsd_file_do_unhash(nf);
 462                 spin_unlock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 463         }
 464         nfsd_file_dispose_list_delayed(&head);
 465         return ret;
 466 }
 467
 468 static void
 469 nfsd_file_gc(void)
 470 {
 471         nfsd_file_lru_walk_list(NULL);
 472 }
 473
 474 static void
 475 nfsd_file_gc_worker(struct work_struct *work)
 476 {
 477         nfsd_file_gc();
 478         nfsd_file_schedule_laundrette();
 479 }
 480
 481 static unsigned long
 482 nfsd_file_lru_count(struct shrinker *s, struct shrink_control *sc)
 483 {
 484         return list_lru_count(&nfsd_file_lru);
 485 }
 486
 487 static unsigned long
 488 nfsd_file_lru_scan(struct shrinker *s, struct shrink_control *sc)
 489 {
 490         return nfsd_file_lru_walk_list(sc);
 491 }
 492
 493 static struct shrinker  nfsd_file_shrinker = {
 494         .scan_objects = nfsd_file_lru_scan,
 495         .count_objects = nfsd_file_lru_count,
 496         .seeks = 1,
 497 };
 498
 499 static void
 500 __nfsd_file_close_inode(struct inode *inode, unsigned int hashval,
 501                         struct list_head *dispose)
 502 {
 503         struct nfsd_file        *nf;
 504         struct hlist_node       *tmp;
 505
 506         spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
 507         hlist_for_each_entry_safe(nf, tmp, &nfsd_file_hashtbl[hashval].nfb_head, nf_node) {
 508                 if (inode == nf->nf_inode)
 509                         nfsd_file_unhash_and_release_locked(nf, dispose);
 510         }
 511         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
 512 }
 513
 514 /**
 515  * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
 516  * @inode: inode of the file to attempt to remove
 517  *
 518  * Walk the whole hash bucket, looking for any files that correspond to "inode".
 519  * If any do, then unhash them and put the hashtable reference to them and
 520  * destroy any that had their last reference put. Also ensure that any of the
 521  * fputs also have their final __fput done as well.
 522  */
 523 void
 524 nfsd_file_close_inode_sync(struct inode *inode)
 525 {
 526         unsigned int            hashval = (unsigned int)hash_long(inode->i_ino,
 527                                                 NFSD_FILE_HASH_BITS);
 528         LIST_HEAD(dispose);
 529
 530         __nfsd_file_close_inode(inode, hashval, &dispose);
 531         trace_nfsd_file_close_inode_sync(inode, hashval, !list_empty(&dispose));
 532         nfsd_file_dispose_list_sync(&dispose);
 533 }
 534
 535 /**
 536  * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
 537  * @inode: inode of the file to attempt to remove
 538  *
 539  * Walk the whole hash bucket, looking for any files that correspond to "inode".
 540  * If any do, then unhash them and put the hashtable reference to them and
 541  * destroy any that had their last reference put.
 542  */
 543 static void
 544 nfsd_file_close_inode(struct inode *inode)
 545 {
 546         unsigned int            hashval = (unsigned int)hash_long(inode->i_ino,
 547                                                 NFSD_FILE_HASH_BITS);
 548         LIST_HEAD(dispose);
 549
 550         __nfsd_file_close_inode(inode, hashval, &dispose);
 551         trace_nfsd_file_close_inode(inode, hashval, !list_empty(&dispose));
 552         nfsd_file_dispose_list_delayed(&dispose);
 553 }
 554
 555 /**
 556  * nfsd_file_delayed_close - close unused nfsd_files
 557  * @work: dummy
 558  *
 559  * Walk the LRU list and close any entries that have not been used since
 560  * the last scan.
 561  *
 562  * Note this can deadlock with nfsd_file_cache_purge.
 563  */
 564 static void
 565 nfsd_file_delayed_close(struct work_struct *work)
 566 {
 567         LIST_HEAD(head);
 568         struct nfsd_fcache_disposal *l = container_of(work,
 569                         struct nfsd_fcache_disposal, work);
 570
 571         nfsd_file_list_remove_disposal(&head, l);
 572         nfsd_file_dispose_list(&head);
 573 }
 574
 575 static int
 576 nfsd_file_lease_notifier_call(struct notifier_block *nb, unsigned long arg,
 577                             void *data)
 578 {
 579         struct file_lock *fl = data;
 580
 581         /* Only close files for F_SETLEASE leases */
 582         if (fl->fl_flags & FL_LEASE)
 583                 nfsd_file_close_inode_sync(file_inode(fl->fl_file));
 584         return 0;
 585 }
 586
 587 static struct notifier_block nfsd_file_lease_notifier = {
 588         .notifier_call = nfsd_file_lease_notifier_call,
 589 };
 590
 591 static int
 592 nfsd_file_fsnotify_handle_event(struct fsnotify_group *group,
 593                                 struct inode *inode,
 594                                 u32 mask, const void *data, int data_type,
 595                                 const struct qstr *file_name, u32 cookie,
 596                                 struct fsnotify_iter_info *iter_info)
 597 {
 598         trace_nfsd_file_fsnotify_handle_event(inode, mask);
 599
 600         /* Should be no marks on non-regular files */
 601         if (!S_ISREG(inode->i_mode)) {
 602                 WARN_ON_ONCE(1);
 603                 return 0;
 604         }
 605
 606         /* don't close files if this was not the last link */
 607         if (mask & FS_ATTRIB) {
 608                 if (inode->i_nlink)
 609                         return 0;
 610         }
 611
 612         nfsd_file_close_inode(inode);
 613         return 0;
 614 }
 615
 616
 617 static const struct fsnotify_ops nfsd_file_fsnotify_ops = {
 618         .handle_event = nfsd_file_fsnotify_handle_event,
 619         .free_mark = nfsd_file_mark_free,
 620 };
 621
 622 int
 623 nfsd_file_cache_init(void)
 624 {
 625         int             ret = -ENOMEM;
 626         unsigned int    i;
 627
 628         clear_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
 629
 630         if (nfsd_file_hashtbl)
 631                 return 0;
 632
 633         nfsd_filecache_wq = alloc_workqueue("nfsd_filecache", 0, 0);
 634         if (!nfsd_filecache_wq)
 635                 goto out;
 636
 637         nfsd_file_hashtbl = kcalloc(NFSD_FILE_HASH_SIZE,
 638                                 sizeof(*nfsd_file_hashtbl), GFP_KERNEL);
 639         if (!nfsd_file_hashtbl) {
 640                 pr_err("nfsd: unable to allocate nfsd_file_hashtbl\n");
 641                 goto out_err;
 642         }
 643
 644         nfsd_file_slab = kmem_cache_create("nfsd_file",
 645                                 sizeof(struct nfsd_file), 0, 0, NULL);
 646         if (!nfsd_file_slab) {
 647                 pr_err("nfsd: unable to create nfsd_file_slab\n");
 648                 goto out_err;
 649         }
 650
 651         nfsd_file_mark_slab = kmem_cache_create("nfsd_file_mark",
 652                                         sizeof(struct nfsd_file_mark), 0, 0, NULL);
 653         if (!nfsd_file_mark_slab) {
 654                 pr_err("nfsd: unable to create nfsd_file_mark_slab\n");
 655                 goto out_err;
 656         }
 657
 658
 659         ret = list_lru_init(&nfsd_file_lru);
 660         if (ret) {
 661                 pr_err("nfsd: failed to init nfsd_file_lru: %d\n", ret);
 662                 goto out_err;
 663         }
 664
 665         ret = register_shrinker(&nfsd_file_shrinker);
 666         if (ret) {
 667                 pr_err("nfsd: failed to register nfsd_file_shrinker: %d\n", ret);
 668                 goto out_lru;
 669         }
 670
 671         ret = lease_register_notifier(&nfsd_file_lease_notifier);
 672         if (ret) {
 673                 pr_err("nfsd: unable to register lease notifier: %d\n", ret);
 674                 goto out_shrinker;
 675         }
 676
 677         nfsd_file_fsnotify_group = fsnotify_alloc_group(&nfsd_file_fsnotify_ops);
 678         if (IS_ERR(nfsd_file_fsnotify_group)) {
 679                 pr_err("nfsd: unable to create fsnotify group: %ld\n",
 680                         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         kfree(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(struct net *net)
 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         l->net = net;
 756         spin_lock_init(&l->lock);
 757         INIT_LIST_HEAD(&l->freeme);
 758         return l;
 759 }
 760
 761 static void
 762 nfsd_free_fcache_disposal(struct nfsd_fcache_disposal *l)
 763 {
 764         rcu_assign_pointer(l->net, NULL);
 765         cancel_work_sync(&l->work);
 766         nfsd_file_dispose_list(&l->freeme);
 767         kfree_rcu(l, rcu);
 768 }
 769
 770 static void
 771 nfsd_add_fcache_disposal(struct nfsd_fcache_disposal *l)
 772 {
 773         spin_lock(&laundrette_lock);
 774         list_add_tail_rcu(&l->list, &laundrettes);
 775         spin_unlock(&laundrette_lock);
 776 }
 777
 778 static void
 779 nfsd_del_fcache_disposal(struct nfsd_fcache_disposal *l)
 780 {
 781         spin_lock(&laundrette_lock);
 782         list_del_rcu(&l->list);
 783         spin_unlock(&laundrette_lock);
 784 }
 785
 786 static int
 787 nfsd_alloc_fcache_disposal_net(struct net *net)
 788 {
 789         struct nfsd_fcache_disposal *l;
 790
 791         l = nfsd_alloc_fcache_disposal(net);
 792         if (!l)
 793                 return -ENOMEM;
 794         nfsd_add_fcache_disposal(l);
 795         return 0;
 796 }
 797
 798 static void
 799 nfsd_free_fcache_disposal_net(struct net *net)
 800 {
 801         struct nfsd_fcache_disposal *l;
 802
 803         rcu_read_lock();
 804         list_for_each_entry_rcu(l, &laundrettes, list) {
 805                 if (l->net != net)
 806                         continue;
 807                 nfsd_del_fcache_disposal(l);
 808                 rcu_read_unlock();
 809                 nfsd_free_fcache_disposal(l);
 810                 return;
 811         }
 812         rcu_read_unlock();
 813 }
 814
 815 int
 816 nfsd_file_cache_start_net(struct net *net)
 817 {
 818         return nfsd_alloc_fcache_disposal_net(net);
 819 }
 820
 821 void
 822 nfsd_file_cache_shutdown_net(struct net *net)
 823 {
 824         nfsd_file_cache_purge(net);
 825         nfsd_free_fcache_disposal_net(net);
 826 }
 827
 828 void
 829 nfsd_file_cache_shutdown(void)
 830 {
 831         set_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
 832
 833         lease_unregister_notifier(&nfsd_file_lease_notifier);
 834         unregister_shrinker(&nfsd_file_shrinker);
 835         /*
 836          * make sure all callers of nfsd_file_lru_cb are done before
 837          * calling nfsd_file_cache_purge
 838          */
 839         cancel_delayed_work_sync(&nfsd_filecache_laundrette);
 840         nfsd_file_cache_purge(NULL);
 841         list_lru_destroy(&nfsd_file_lru);
 842         rcu_barrier();
 843         fsnotify_put_group(nfsd_file_fsnotify_group);
 844         nfsd_file_fsnotify_group = NULL;
 845         kmem_cache_destroy(nfsd_file_slab);
 846         nfsd_file_slab = NULL;
 847         fsnotify_wait_marks_destroyed();
 848         kmem_cache_destroy(nfsd_file_mark_slab);
 849         nfsd_file_mark_slab = NULL;
 850         kfree(nfsd_file_hashtbl);
 851         nfsd_file_hashtbl = NULL;
 852         destroy_workqueue(nfsd_filecache_wq);
 853         nfsd_filecache_wq = NULL;
 854 }
 855
 856 static bool
 857 nfsd_match_cred(const struct cred *c1, const struct cred *c2)
 858 {
 859         int i;
 860
 861         if (!uid_eq(c1->fsuid, c2->fsuid))
 862                 return false;
 863         if (!gid_eq(c1->fsgid, c2->fsgid))
 864                 return false;
 865         if (c1->group_info == NULL || c2->group_info == NULL)
 866                 return c1->group_info == c2->group_info;
 867         if (c1->group_info->ngroups != c2->group_info->ngroups)
 868                 return false;
 869         for (i = 0; i < c1->group_info->ngroups; i++) {
 870                 if (!gid_eq(c1->group_info->gid[i], c2->group_info->gid[i]))
 871                         return false;
 872         }
 873         return true;
 874 }
 875
 876 static struct nfsd_file *
 877 nfsd_file_find_locked(struct inode *inode, unsigned int may_flags,
 878                         unsigned int hashval, struct net *net)
 879 {
 880         struct nfsd_file *nf;
 881         unsigned char need = may_flags & NFSD_FILE_MAY_MASK;
 882
 883         hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
 884                                  nf_node) {
 885                 if ((need & nf->nf_may) != need)
 886                         continue;
 887                 if (nf->nf_inode != inode)
 888                         continue;
 889                 if (nf->nf_net != net)
 890                         continue;
 891                 if (!nfsd_match_cred(nf->nf_cred, current_cred()))
 892                         continue;
 893                 if (nfsd_file_get(nf) != NULL)
 894                         return nf;
 895         }
 896         return NULL;
 897 }
 898
 899 /**
 900  * nfsd_file_is_cached - are there any cached open files for this fh?
 901  * @inode: inode of the file to check
 902  *
 903  * Scan the hashtable for open files that match this fh. Returns true if there
 904  * are any, and false if not.
 905  */
 906 bool
 907 nfsd_file_is_cached(struct inode *inode)
 908 {
 909         bool                    ret = false;
 910         struct nfsd_file        *nf;
 911         unsigned int            hashval;
 912
 913         hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
 914
 915         rcu_read_lock();
 916         hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
 917                                  nf_node) {
 918                 if (inode == nf->nf_inode) {
 919                         ret = true;
 920                         break;
 921                 }
 922         }
 923         rcu_read_unlock();
 924         trace_nfsd_file_is_cached(inode, hashval, (int)ret);
 925         return ret;
 926 }
 927
 928 __be32
 929 nfsd_file_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 930                   unsigned int may_flags, struct nfsd_file **pnf)
 931 {
 932         __be32  status;
 933         struct net *net = SVC_NET(rqstp);
 934         struct nfsd_file *nf, *new;
 935         struct inode *inode;
 936         unsigned int hashval;
 937         bool retry = true;
 938
 939         /* FIXME: skip this if fh_dentry is already set? */
 940         status = fh_verify(rqstp, fhp, S_IFREG,
 941                                 may_flags|NFSD_MAY_OWNER_OVERRIDE);
 942         if (status != nfs_ok)
 943                 return status;
 944
 945         inode = d_inode(fhp->fh_dentry);
 946         hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
 947 retry:
 948         rcu_read_lock();
 949         nf = nfsd_file_find_locked(inode, may_flags, hashval, net);
 950         rcu_read_unlock();
 951         if (nf)
 952                 goto wait_for_construction;
 953
 954         new = nfsd_file_alloc(inode, may_flags, hashval, net);
 955         if (!new) {
 956                 trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags,
 957                                         NULL, nfserr_jukebox);
 958                 return nfserr_jukebox;
 959         }
 960
 961         spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
 962         nf = nfsd_file_find_locked(inode, may_flags, hashval, net);
 963         if (nf == NULL)
 964                 goto open_file;
 965         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
 966         nfsd_file_slab_free(&new->nf_rcu);
 967
 968 wait_for_construction:
 969         wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
 970
 971         /* Did construction of this file fail? */
 972         if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
 973                 if (!retry) {
 974                         status = nfserr_jukebox;
 975                         goto out;
 976                 }
 977                 retry = false;
 978                 nfsd_file_put_noref(nf);
 979                 goto retry;
 980         }
 981
 982         this_cpu_inc(nfsd_file_cache_hits);
 983
 984         if (!(may_flags & NFSD_MAY_NOT_BREAK_LEASE)) {
 985                 bool write = (may_flags & NFSD_MAY_WRITE);
 986
 987                 if (test_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags) ||
 988                     (test_bit(NFSD_FILE_BREAK_WRITE, &nf->nf_flags) && write)) {
 989                         status = nfserrno(nfsd_open_break_lease(
 990                                         file_inode(nf->nf_file), may_flags));
 991                         if (status == nfs_ok) {
 992                                 clear_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags);
 993                                 if (write)
 994                                         clear_bit(NFSD_FILE_BREAK_WRITE,
 995                                                   &nf->nf_flags);
 996                         }
 997                 }
 998         }
 999 out:
1000         if (status == nfs_ok) {
1001                 *pnf = nf;
1002         } else {
1003                 nfsd_file_put(nf);
1004                 nf = NULL;
1005         }
1006
1007         trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags, nf, status);
1008         return status;
1009 open_file:
1010         nf = new;
1011         /* Take reference for the hashtable */
1012         atomic_inc(&nf->nf_ref);
1013         __set_bit(NFSD_FILE_HASHED, &nf->nf_flags);
1014         __set_bit(NFSD_FILE_PENDING, &nf->nf_flags);
1015         list_lru_add(&nfsd_file_lru, &nf->nf_lru);
1016         hlist_add_head_rcu(&nf->nf_node, &nfsd_file_hashtbl[hashval].nfb_head);
1017         ++nfsd_file_hashtbl[hashval].nfb_count;
1018         nfsd_file_hashtbl[hashval].nfb_maxcount = max(nfsd_file_hashtbl[hashval].nfb_maxcount,
1019                         nfsd_file_hashtbl[hashval].nfb_count);
1020         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
1021         if (atomic_long_inc_return(&nfsd_filecache_count) >= NFSD_FILE_LRU_THRESHOLD)
1022                 nfsd_file_gc();
1023
1024         nf->nf_mark = nfsd_file_mark_find_or_create(nf);
1025         if (nf->nf_mark)
1026                 status = nfsd_open_verified(rqstp, fhp, S_IFREG,
1027                                 may_flags, &nf->nf_file);
1028         else
1029                 status = nfserr_jukebox;
1030         /*
1031          * If construction failed, or we raced with a call to unlink()
1032          * then unhash.
1033          */
1034         if (status != nfs_ok || inode->i_nlink == 0) {
1035                 bool do_free;
1036                 spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
1037                 do_free = nfsd_file_unhash(nf);
1038                 spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
1039                 if (do_free)
1040                         nfsd_file_put_noref(nf);
1041         }
1042         clear_bit_unlock(NFSD_FILE_PENDING, &nf->nf_flags);
1043         smp_mb__after_atomic();
1044         wake_up_bit(&nf->nf_flags, NFSD_FILE_PENDING);
1045         goto out;
1046 }
1047
1048 /*
1049  * Note that fields may be added, removed or reordered in the future. Programs
1050  * scraping this file for info should test the labels to ensure they're
1051  * getting the correct field.
1052  */
1053 static int nfsd_file_cache_stats_show(struct seq_file *m, void *v)
1054 {
1055         unsigned int i, count = 0, longest = 0;
1056         unsigned long hits = 0;
1057
1058         /*
1059          * No need for spinlocks here since we're not terribly interested in
1060          * accuracy. We do take the nfsd_mutex simply to ensure that we
1061          * don't end up racing with server shutdown
1062          */
1063         mutex_lock(&nfsd_mutex);
1064         if (nfsd_file_hashtbl) {
1065                 for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
1066                         count += nfsd_file_hashtbl[i].nfb_count;
1067                         longest = max(longest, nfsd_file_hashtbl[i].nfb_count);
1068                 }
1069         }
1070         mutex_unlock(&nfsd_mutex);
1071
1072         for_each_possible_cpu(i)
1073                 hits += per_cpu(nfsd_file_cache_hits, i);
1074
1075         seq_printf(m, "total entries: %u\n", count);
1076         seq_printf(m, "longest chain: %u\n", longest);
1077         seq_printf(m, "cache hits:    %lu\n", hits);
1078         return 0;
1079 }
1080
1081 int nfsd_file_cache_stats_open(struct inode *inode, struct file *file)
1082 {
1083         return single_open(file, nfsd_file_cache_stats_show, NULL);
1084 }