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 static 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 (!refcount_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 (refcount_dec_and_test(&nfm->nfm_ref)) {
 112                 fsnotify_destroy_mark(&nfm->nfm_mark, nfsd_file_fsnotify_group);
 113                 fsnotify_put_mark(&nfm->nfm_mark);
 114         }
 115 }
 116
 117 static struct nfsd_file_mark *
 118 nfsd_file_mark_find_or_create(struct nfsd_file *nf)
 119 {
 120         int                     err;
 121         struct fsnotify_mark    *mark;
 122         struct nfsd_file_mark   *nfm = NULL, *new;
 123         struct inode *inode = nf->nf_inode;
 124
 125         do {
 126                 mutex_lock(&nfsd_file_fsnotify_group->mark_mutex);
 127                 mark = fsnotify_find_mark(&inode->i_fsnotify_marks,
 128                                 nfsd_file_fsnotify_group);
 129                 if (mark) {
 130                         nfm = nfsd_file_mark_get(container_of(mark,
 131                                                  struct nfsd_file_mark,
 132                                                  nfm_mark));
 133                         mutex_unlock(&nfsd_file_fsnotify_group->mark_mutex);
 134                         if (nfm) {
 135                                 fsnotify_put_mark(mark);
 136                                 break;
 137                         }
 138                         /* Avoid soft lockup race with nfsd_file_mark_put() */
 139                         fsnotify_destroy_mark(mark, nfsd_file_fsnotify_group);
 140                         fsnotify_put_mark(mark);
 141                 } else
 142                         mutex_unlock(&nfsd_file_fsnotify_group->mark_mutex);
 143
 144                 /* allocate a new nfm */
 145                 new = kmem_cache_alloc(nfsd_file_mark_slab, GFP_KERNEL);
 146                 if (!new)
 147                         return NULL;
 148                 fsnotify_init_mark(&new->nfm_mark, nfsd_file_fsnotify_group);
 149                 new->nfm_mark.mask = FS_ATTRIB|FS_DELETE_SELF;
 150                 refcount_set(&new->nfm_ref, 1);
 151
 152                 err = fsnotify_add_inode_mark(&new->nfm_mark, inode, 0);
 153
 154                 /*
 155                  * If the add was successful, then return the object.
 156                  * Otherwise, we need to put the reference we hold on the
 157                  * nfm_mark. The fsnotify code will take a reference and put
 158                  * it on failure, so we can't just free it directly. It's also
 159                  * not safe to call fsnotify_destroy_mark on it as the
 160                  * mark->group will be NULL. Thus, we can't let the nfm_ref
 161                  * counter drive the destruction at this point.
 162                  */
 163                 if (likely(!err))
 164                         nfm = new;
 165                 else
 166                         fsnotify_put_mark(&new->nfm_mark);
 167         } while (unlikely(err == -EEXIST));
 168
 169         return nfm;
 170 }
 171
 172 static struct nfsd_file *
 173 nfsd_file_alloc(struct inode *inode, unsigned int may, unsigned int hashval,
 174                 struct net *net)
 175 {
 176         struct nfsd_file *nf;
 177
 178         nf = kmem_cache_alloc(nfsd_file_slab, GFP_KERNEL);
 179         if (nf) {
 180                 INIT_HLIST_NODE(&nf->nf_node);
 181                 INIT_LIST_HEAD(&nf->nf_lru);
 182                 nf->nf_file = NULL;
 183                 nf->nf_cred = get_current_cred();
 184                 nf->nf_net = net;
 185                 nf->nf_flags = 0;
 186                 nf->nf_inode = inode;
 187                 nf->nf_hashval = hashval;
 188                 refcount_set(&nf->nf_ref, 1);
 189                 nf->nf_may = may & NFSD_FILE_MAY_MASK;
 190                 nf->nf_mark = NULL;
 191                 trace_nfsd_file_alloc(nf);
 192         }
 193         return nf;
 194 }
 195
 196 static bool
 197 nfsd_file_free(struct nfsd_file *nf)
 198 {
 199         bool flush = false;
 200
 201         trace_nfsd_file_put_final(nf);
 202         if (nf->nf_mark)
 203                 nfsd_file_mark_put(nf->nf_mark);
 204         if (nf->nf_file) {
 205                 get_file(nf->nf_file);
 206                 filp_close(nf->nf_file, NULL);
 207                 fput(nf->nf_file);
 208                 flush = true;
 209         }
 210         call_rcu(&nf->nf_rcu, nfsd_file_slab_free);
 211         return flush;
 212 }
 213
 214 static bool
 215 nfsd_file_check_writeback(struct nfsd_file *nf)
 216 {
 217         struct file *file = nf->nf_file;
 218         struct address_space *mapping;
 219
 220         if (!file || !(file->f_mode & FMODE_WRITE))
 221                 return false;
 222         mapping = file->f_mapping;
 223         return mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) ||
 224                 mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK);
 225 }
 226
 227 static int
 228 nfsd_file_check_write_error(struct nfsd_file *nf)
 229 {
 230         struct file *file = nf->nf_file;
 231
 232         if (!file || !(file->f_mode & FMODE_WRITE))
 233                 return 0;
 234         return filemap_check_wb_err(file->f_mapping, READ_ONCE(file->f_wb_err));
 235 }
 236
 237 static void
 238 nfsd_file_do_unhash(struct nfsd_file *nf)
 239 {
 240         lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 241
 242         trace_nfsd_file_unhash(nf);
 243
 244         if (nfsd_file_check_write_error(nf))
 245                 nfsd_reset_boot_verifier(net_generic(nf->nf_net, nfsd_net_id));
 246         --nfsd_file_hashtbl[nf->nf_hashval].nfb_count;
 247         hlist_del_rcu(&nf->nf_node);
 248         atomic_long_dec(&nfsd_filecache_count);
 249 }
 250
 251 static bool
 252 nfsd_file_unhash(struct nfsd_file *nf)
 253 {
 254         if (test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
 255                 nfsd_file_do_unhash(nf);
 256                 if (!list_empty(&nf->nf_lru))
 257                         list_lru_del(&nfsd_file_lru, &nf->nf_lru);
 258                 return true;
 259         }
 260         return false;
 261 }
 262
 263 /*
 264  * Return true if the file was unhashed.
 265  */
 266 static bool
 267 nfsd_file_unhash_and_release_locked(struct nfsd_file *nf, struct list_head *dispose)
 268 {
 269         lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 270
 271         trace_nfsd_file_unhash_and_release_locked(nf);
 272         if (!nfsd_file_unhash(nf))
 273                 return false;
 274         /* keep final reference for nfsd_file_lru_dispose */
 275         if (refcount_dec_not_one(&nf->nf_ref))
 276                 return true;
 277
 278         list_add(&nf->nf_lru, dispose);
 279         return true;
 280 }
 281
 282 static void
 283 nfsd_file_put_noref(struct nfsd_file *nf)
 284 {
 285         trace_nfsd_file_put(nf);
 286
 287         if (refcount_dec_and_test(&nf->nf_ref)) {
 288                 WARN_ON(test_bit(NFSD_FILE_HASHED, &nf->nf_flags));
 289                 nfsd_file_free(nf);
 290         }
 291 }
 292
 293 void
 294 nfsd_file_put(struct nfsd_file *nf)
 295 {
 296         bool is_hashed;
 297
 298         set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
 299         if (refcount_read(&nf->nf_ref) > 2 || !nf->nf_file) {
 300                 nfsd_file_put_noref(nf);
 301                 return;
 302         }
 303
 304         filemap_flush(nf->nf_file->f_mapping);
 305         is_hashed = test_bit(NFSD_FILE_HASHED, &nf->nf_flags) != 0;
 306         nfsd_file_put_noref(nf);
 307         if (is_hashed)
 308                 nfsd_file_schedule_laundrette();
 309         if (atomic_long_read(&nfsd_filecache_count) >= NFSD_FILE_LRU_LIMIT)
 310                 nfsd_file_gc();
 311 }
 312
 313 struct nfsd_file *
 314 nfsd_file_get(struct nfsd_file *nf)
 315 {
 316         if (likely(refcount_inc_not_zero(&nf->nf_ref)))
 317                 return nf;
 318         return NULL;
 319 }
 320
 321 static void
 322 nfsd_file_dispose_list(struct list_head *dispose)
 323 {
 324         struct nfsd_file *nf;
 325
 326         while(!list_empty(dispose)) {
 327                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 328                 list_del(&nf->nf_lru);
 329                 nfsd_file_put_noref(nf);
 330         }
 331 }
 332
 333 static void
 334 nfsd_file_dispose_list_sync(struct list_head *dispose)
 335 {
 336         bool flush = false;
 337         struct nfsd_file *nf;
 338
 339         while(!list_empty(dispose)) {
 340                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 341                 list_del(&nf->nf_lru);
 342                 if (!refcount_dec_and_test(&nf->nf_ref))
 343                         continue;
 344                 if (nfsd_file_free(nf))
 345                         flush = true;
 346         }
 347         if (flush)
 348                 flush_delayed_fput();
 349 }
 350
 351 static void
 352 nfsd_file_list_remove_disposal(struct list_head *dst,
 353                 struct nfsd_fcache_disposal *l)
 354 {
 355         spin_lock(&l->lock);
 356         list_splice_init(&l->freeme, dst);
 357         spin_unlock(&l->lock);
 358 }
 359
 360 static void
 361 nfsd_file_list_add_disposal(struct list_head *files, struct net *net)
 362 {
 363         struct nfsd_fcache_disposal *l;
 364
 365         rcu_read_lock();
 366         list_for_each_entry_rcu(l, &laundrettes, list) {
 367                 if (l->net == net) {
 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                         break;
 373                 }
 374         }
 375         rcu_read_unlock();
 376 }
 377
 378 static void
 379 nfsd_file_list_add_pernet(struct list_head *dst, struct list_head *src,
 380                 struct net *net)
 381 {
 382         struct nfsd_file *nf, *tmp;
 383
 384         list_for_each_entry_safe(nf, tmp, src, nf_lru) {
 385                 if (nf->nf_net == net)
 386                         list_move_tail(&nf->nf_lru, dst);
 387         }
 388 }
 389
 390 static void
 391 nfsd_file_dispose_list_delayed(struct list_head *dispose)
 392 {
 393         LIST_HEAD(list);
 394         struct nfsd_file *nf;
 395
 396         while(!list_empty(dispose)) {
 397                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 398                 nfsd_file_list_add_pernet(&list, dispose, nf->nf_net);
 399                 nfsd_file_list_add_disposal(&list, nf->nf_net);
 400         }
 401 }
 402
 403 /*
 404  * Note this can deadlock with nfsd_file_cache_purge.
 405  */
 406 static enum lru_status
 407 nfsd_file_lru_cb(struct list_head *item, struct list_lru_one *lru,
 408                  spinlock_t *lock, void *arg)
 409         __releases(lock)
 410         __acquires(lock)
 411 {
 412         struct list_head *head = arg;
 413         struct nfsd_file *nf = list_entry(item, struct nfsd_file, nf_lru);
 414
 415         /*
 416          * Do a lockless refcount check. The hashtable holds one reference, so
 417          * we look to see if anything else has a reference, or if any have
 418          * been put since the shrinker last ran. Those don't get unhashed and
 419          * released.
 420          *
 421          * Note that in the put path, we set the flag and then decrement the
 422          * counter. Here we check the counter and then test and clear the flag.
 423          * That order is deliberate to ensure that we can do this locklessly.
 424          */
 425         if (refcount_read(&nf->nf_ref) > 1)
 426                 goto out_skip;
 427
 428         /*
 429          * Don't throw out files that are still undergoing I/O or
 430          * that have uncleared errors pending.
 431          */
 432         if (nfsd_file_check_writeback(nf))
 433                 goto out_skip;
 434
 435         if (test_and_clear_bit(NFSD_FILE_REFERENCED, &nf->nf_flags))
 436                 goto out_skip;
 437
 438         if (!test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags))
 439                 goto out_skip;
 440
 441         list_lru_isolate_move(lru, &nf->nf_lru, head);
 442         return LRU_REMOVED;
 443 out_skip:
 444         return LRU_SKIP;
 445 }
 446
 447 static unsigned long
 448 nfsd_file_lru_walk_list(struct shrink_control *sc)
 449 {
 450         LIST_HEAD(head);
 451         struct nfsd_file *nf;
 452         unsigned long ret;
 453
 454         if (sc)
 455                 ret = list_lru_shrink_walk(&nfsd_file_lru, sc,
 456                                 nfsd_file_lru_cb, &head);
 457         else
 458                 ret = list_lru_walk(&nfsd_file_lru,
 459                                 nfsd_file_lru_cb,
 460                                 &head, LONG_MAX);
 461         list_for_each_entry(nf, &head, nf_lru) {
 462                 spin_lock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 463                 nfsd_file_do_unhash(nf);
 464                 spin_unlock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 465         }
 466         nfsd_file_dispose_list_delayed(&head);
 467         return ret;
 468 }
 469
 470 static void
 471 nfsd_file_gc(void)
 472 {
 473         nfsd_file_lru_walk_list(NULL);
 474 }
 475
 476 static void
 477 nfsd_file_gc_worker(struct work_struct *work)
 478 {
 479         nfsd_file_gc();
 480         nfsd_file_schedule_laundrette();
 481 }
 482
 483 static unsigned long
 484 nfsd_file_lru_count(struct shrinker *s, struct shrink_control *sc)
 485 {
 486         return list_lru_count(&nfsd_file_lru);
 487 }
 488
 489 static unsigned long
 490 nfsd_file_lru_scan(struct shrinker *s, struct shrink_control *sc)
 491 {
 492         return nfsd_file_lru_walk_list(sc);
 493 }
 494
 495 static struct shrinker  nfsd_file_shrinker = {
 496         .scan_objects = nfsd_file_lru_scan,
 497         .count_objects = nfsd_file_lru_count,
 498         .seeks = 1,
 499 };
 500
 501 static void
 502 __nfsd_file_close_inode(struct inode *inode, unsigned int hashval,
 503                         struct list_head *dispose)
 504 {
 505         struct nfsd_file        *nf;
 506         struct hlist_node       *tmp;
 507
 508         spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
 509         hlist_for_each_entry_safe(nf, tmp, &nfsd_file_hashtbl[hashval].nfb_head, nf_node) {
 510                 if (inode == nf->nf_inode)
 511                         nfsd_file_unhash_and_release_locked(nf, dispose);
 512         }
 513         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
 514 }
 515
 516 /**
 517  * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
 518  * @inode: inode of the file to attempt to remove
 519  *
 520  * Walk the whole hash bucket, looking for any files that correspond to "inode".
 521  * If any do, then unhash them and put the hashtable reference to them and
 522  * destroy any that had their last reference put. Also ensure that any of the
 523  * fputs also have their final __fput done as well.
 524  */
 525 void
 526 nfsd_file_close_inode_sync(struct inode *inode)
 527 {
 528         unsigned int            hashval = (unsigned int)hash_long(inode->i_ino,
 529                                                 NFSD_FILE_HASH_BITS);
 530         LIST_HEAD(dispose);
 531
 532         __nfsd_file_close_inode(inode, hashval, &dispose);
 533         trace_nfsd_file_close_inode_sync(inode, hashval, !list_empty(&dispose));
 534         nfsd_file_dispose_list_sync(&dispose);
 535 }
 536
 537 /**
 538  * nfsd_file_close_inode - attempt a delayed close of a nfsd_file
 539  * @inode: inode of the file to attempt to remove
 540  *
 541  * Walk the whole hash bucket, looking for any files that correspond to "inode".
 542  * If any do, then unhash them and put the hashtable reference to them and
 543  * destroy any that had their last reference put.
 544  */
 545 static void
 546 nfsd_file_close_inode(struct inode *inode)
 547 {
 548         unsigned int            hashval = (unsigned int)hash_long(inode->i_ino,
 549                                                 NFSD_FILE_HASH_BITS);
 550         LIST_HEAD(dispose);
 551
 552         __nfsd_file_close_inode(inode, hashval, &dispose);
 553         trace_nfsd_file_close_inode(inode, hashval, !list_empty(&dispose));
 554         nfsd_file_dispose_list_delayed(&dispose);
 555 }
 556
 557 /**
 558  * nfsd_file_delayed_close - close unused nfsd_files
 559  * @work: dummy
 560  *
 561  * Walk the LRU list and close any entries that have not been used since
 562  * the last scan.
 563  *
 564  * Note this can deadlock with nfsd_file_cache_purge.
 565  */
 566 static void
 567 nfsd_file_delayed_close(struct work_struct *work)
 568 {
 569         LIST_HEAD(head);
 570         struct nfsd_fcache_disposal *l = container_of(work,
 571                         struct nfsd_fcache_disposal, work);
 572
 573         nfsd_file_list_remove_disposal(&head, l);
 574         nfsd_file_dispose_list(&head);
 575 }
 576
 577 static int
 578 nfsd_file_lease_notifier_call(struct notifier_block *nb, unsigned long arg,
 579                             void *data)
 580 {
 581         struct file_lock *fl = data;
 582
 583         /* Only close files for F_SETLEASE leases */
 584         if (fl->fl_flags & FL_LEASE)
 585                 nfsd_file_close_inode_sync(file_inode(fl->fl_file));
 586         return 0;
 587 }
 588
 589 static struct notifier_block nfsd_file_lease_notifier = {
 590         .notifier_call = nfsd_file_lease_notifier_call,
 591 };
 592
 593 static int
 594 nfsd_file_fsnotify_handle_event(struct fsnotify_mark *mark, u32 mask,
 595                                 struct inode *inode, struct inode *dir,
 596                                 const struct qstr *name, u32 cookie)
 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_inode_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 = kvcalloc(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                 ret = PTR_ERR(nfsd_file_fsnotify_group);
 682                 nfsd_file_fsnotify_group = NULL;
 683                 goto out_notifier;
 684         }
 685
 686         for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
 687                 INIT_HLIST_HEAD(&nfsd_file_hashtbl[i].nfb_head);
 688                 spin_lock_init(&nfsd_file_hashtbl[i].nfb_lock);
 689         }
 690
 691         INIT_DELAYED_WORK(&nfsd_filecache_laundrette, nfsd_file_gc_worker);
 692 out:
 693         return ret;
 694 out_notifier:
 695         lease_unregister_notifier(&nfsd_file_lease_notifier);
 696 out_shrinker:
 697         unregister_shrinker(&nfsd_file_shrinker);
 698 out_lru:
 699         list_lru_destroy(&nfsd_file_lru);
 700 out_err:
 701         kmem_cache_destroy(nfsd_file_slab);
 702         nfsd_file_slab = NULL;
 703         kmem_cache_destroy(nfsd_file_mark_slab);
 704         nfsd_file_mark_slab = NULL;
 705         kvfree(nfsd_file_hashtbl);
 706         nfsd_file_hashtbl = NULL;
 707         destroy_workqueue(nfsd_filecache_wq);
 708         nfsd_filecache_wq = NULL;
 709         goto out;
 710 }
 711
 712 /*
 713  * Note this can deadlock with nfsd_file_lru_cb.
 714  */
 715 void
 716 nfsd_file_cache_purge(struct net *net)
 717 {
 718         unsigned int            i;
 719         struct nfsd_file        *nf;
 720         struct hlist_node       *next;
 721         LIST_HEAD(dispose);
 722         bool del;
 723
 724         if (!nfsd_file_hashtbl)
 725                 return;
 726
 727         for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
 728                 struct nfsd_fcache_bucket *nfb = &nfsd_file_hashtbl[i];
 729
 730                 spin_lock(&nfb->nfb_lock);
 731                 hlist_for_each_entry_safe(nf, next, &nfb->nfb_head, nf_node) {
 732                         if (net && nf->nf_net != net)
 733                                 continue;
 734                         del = nfsd_file_unhash_and_release_locked(nf, &dispose);
 735
 736                         /*
 737                          * Deadlock detected! Something marked this entry as
 738                          * unhased, but hasn't removed it from the hash list.
 739                          */
 740                         WARN_ON_ONCE(!del);
 741                 }
 742                 spin_unlock(&nfb->nfb_lock);
 743                 nfsd_file_dispose_list(&dispose);
 744         }
 745 }
 746
 747 static struct nfsd_fcache_disposal *
 748 nfsd_alloc_fcache_disposal(struct net *net)
 749 {
 750         struct nfsd_fcache_disposal *l;
 751
 752         l = kmalloc(sizeof(*l), GFP_KERNEL);
 753         if (!l)
 754                 return NULL;
 755         INIT_WORK(&l->work, nfsd_file_delayed_close);
 756         l->net = net;
 757         spin_lock_init(&l->lock);
 758         INIT_LIST_HEAD(&l->freeme);
 759         return l;
 760 }
 761
 762 static void
 763 nfsd_free_fcache_disposal(struct nfsd_fcache_disposal *l)
 764 {
 765         rcu_assign_pointer(l->net, NULL);
 766         cancel_work_sync(&l->work);
 767         nfsd_file_dispose_list(&l->freeme);
 768         kfree_rcu(l, rcu);
 769 }
 770
 771 static void
 772 nfsd_add_fcache_disposal(struct nfsd_fcache_disposal *l)
 773 {
 774         spin_lock(&laundrette_lock);
 775         list_add_tail_rcu(&l->list, &laundrettes);
 776         spin_unlock(&laundrette_lock);
 777 }
 778
 779 static void
 780 nfsd_del_fcache_disposal(struct nfsd_fcache_disposal *l)
 781 {
 782         spin_lock(&laundrette_lock);
 783         list_del_rcu(&l->list);
 784         spin_unlock(&laundrette_lock);
 785 }
 786
 787 static int
 788 nfsd_alloc_fcache_disposal_net(struct net *net)
 789 {
 790         struct nfsd_fcache_disposal *l;
 791
 792         l = nfsd_alloc_fcache_disposal(net);
 793         if (!l)
 794                 return -ENOMEM;
 795         nfsd_add_fcache_disposal(l);
 796         return 0;
 797 }
 798
 799 static void
 800 nfsd_free_fcache_disposal_net(struct net *net)
 801 {
 802         struct nfsd_fcache_disposal *l;
 803
 804         rcu_read_lock();
 805         list_for_each_entry_rcu(l, &laundrettes, list) {
 806                 if (l->net != net)
 807                         continue;
 808                 nfsd_del_fcache_disposal(l);
 809                 rcu_read_unlock();
 810                 nfsd_free_fcache_disposal(l);
 811                 return;
 812         }
 813         rcu_read_unlock();
 814 }
 815
 816 int
 817 nfsd_file_cache_start_net(struct net *net)
 818 {
 819         return nfsd_alloc_fcache_disposal_net(net);
 820 }
 821
 822 void
 823 nfsd_file_cache_shutdown_net(struct net *net)
 824 {
 825         nfsd_file_cache_purge(net);
 826         nfsd_free_fcache_disposal_net(net);
 827 }
 828
 829 void
 830 nfsd_file_cache_shutdown(void)
 831 {
 832         set_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
 833
 834         lease_unregister_notifier(&nfsd_file_lease_notifier);
 835         unregister_shrinker(&nfsd_file_shrinker);
 836         /*
 837          * make sure all callers of nfsd_file_lru_cb are done before
 838          * calling nfsd_file_cache_purge
 839          */
 840         cancel_delayed_work_sync(&nfsd_filecache_laundrette);
 841         nfsd_file_cache_purge(NULL);
 842         list_lru_destroy(&nfsd_file_lru);
 843         rcu_barrier();
 844         fsnotify_put_group(nfsd_file_fsnotify_group);
 845         nfsd_file_fsnotify_group = NULL;
 846         kmem_cache_destroy(nfsd_file_slab);
 847         nfsd_file_slab = NULL;
 848         fsnotify_wait_marks_destroyed();
 849         kmem_cache_destroy(nfsd_file_mark_slab);
 850         nfsd_file_mark_slab = NULL;
 851         kvfree(nfsd_file_hashtbl);
 852         nfsd_file_hashtbl = NULL;
 853         destroy_workqueue(nfsd_filecache_wq);
 854         nfsd_filecache_wq = NULL;
 855 }
 856
 857 static bool
 858 nfsd_match_cred(const struct cred *c1, const struct cred *c2)
 859 {
 860         int i;
 861
 862         if (!uid_eq(c1->fsuid, c2->fsuid))
 863                 return false;
 864         if (!gid_eq(c1->fsgid, c2->fsgid))
 865                 return false;
 866         if (c1->group_info == NULL || c2->group_info == NULL)
 867                 return c1->group_info == c2->group_info;
 868         if (c1->group_info->ngroups != c2->group_info->ngroups)
 869                 return false;
 870         for (i = 0; i < c1->group_info->ngroups; i++) {
 871                 if (!gid_eq(c1->group_info->gid[i], c2->group_info->gid[i]))
 872                         return false;
 873         }
 874         return true;
 875 }
 876
 877 static struct nfsd_file *
 878 nfsd_file_find_locked(struct inode *inode, unsigned int may_flags,
 879                         unsigned int hashval, struct net *net)
 880 {
 881         struct nfsd_file *nf;
 882         unsigned char need = may_flags & NFSD_FILE_MAY_MASK;
 883
 884         hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
 885                                  nf_node, lockdep_is_held(&nfsd_file_hashtbl[hashval].nfb_lock)) {
 886                 if (nf->nf_may != need)
 887                         continue;
 888                 if (nf->nf_inode != inode)
 889                         continue;
 890                 if (nf->nf_net != net)
 891                         continue;
 892                 if (!nfsd_match_cred(nf->nf_cred, current_cred()))
 893                         continue;
 894                 if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags))
 895                         continue;
 896                 if (nfsd_file_get(nf) != NULL)
 897                         return nf;
 898         }
 899         return NULL;
 900 }
 901
 902 /**
 903  * nfsd_file_is_cached - are there any cached open files for this fh?
 904  * @inode: inode of the file to check
 905  *
 906  * Scan the hashtable for open files that match this fh. Returns true if there
 907  * are any, and false if not.
 908  */
 909 bool
 910 nfsd_file_is_cached(struct inode *inode)
 911 {
 912         bool                    ret = false;
 913         struct nfsd_file        *nf;
 914         unsigned int            hashval;
 915
 916         hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
 917
 918         rcu_read_lock();
 919         hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
 920                                  nf_node) {
 921                 if (inode == nf->nf_inode) {
 922                         ret = true;
 923                         break;
 924                 }
 925         }
 926         rcu_read_unlock();
 927         trace_nfsd_file_is_cached(inode, hashval, (int)ret);
 928         return ret;
 929 }
 930
 931 __be32
 932 nfsd_file_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 933                   unsigned int may_flags, struct nfsd_file **pnf)
 934 {
 935         __be32  status;
 936         struct net *net = SVC_NET(rqstp);
 937         struct nfsd_file *nf, *new;
 938         struct inode *inode;
 939         unsigned int hashval;
 940         bool retry = true;
 941
 942         /* FIXME: skip this if fh_dentry is already set? */
 943         status = fh_verify(rqstp, fhp, S_IFREG,
 944                                 may_flags|NFSD_MAY_OWNER_OVERRIDE);
 945         if (status != nfs_ok)
 946                 return status;
 947
 948         inode = d_inode(fhp->fh_dentry);
 949         hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
 950 retry:
 951         rcu_read_lock();
 952         nf = nfsd_file_find_locked(inode, may_flags, hashval, net);
 953         rcu_read_unlock();
 954         if (nf)
 955                 goto wait_for_construction;
 956
 957         new = nfsd_file_alloc(inode, may_flags, hashval, net);
 958         if (!new) {
 959                 trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags,
 960                                         NULL, nfserr_jukebox);
 961                 return nfserr_jukebox;
 962         }
 963
 964         spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
 965         nf = nfsd_file_find_locked(inode, may_flags, hashval, net);
 966         if (nf == NULL)
 967                 goto open_file;
 968         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
 969         nfsd_file_slab_free(&new->nf_rcu);
 970
 971 wait_for_construction:
 972         wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
 973
 974         /* Did construction of this file fail? */
 975         if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
 976                 if (!retry) {
 977                         status = nfserr_jukebox;
 978                         goto out;
 979                 }
 980                 retry = false;
 981                 nfsd_file_put_noref(nf);
 982                 goto retry;
 983         }
 984
 985         this_cpu_inc(nfsd_file_cache_hits);
 986
 987         status = nfserrno(nfsd_open_break_lease(file_inode(nf->nf_file), may_flags));
 988 out:
 989         if (status == nfs_ok) {
 990                 *pnf = nf;
 991         } else {
 992                 nfsd_file_put(nf);
 993                 nf = NULL;
 994         }
 995
 996         trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags, nf, status);
 997         return status;
 998 open_file:
 999         nf = new;
1000         /* Take reference for the hashtable */
1001         refcount_inc(&nf->nf_ref);
1002         __set_bit(NFSD_FILE_HASHED, &nf->nf_flags);
1003         __set_bit(NFSD_FILE_PENDING, &nf->nf_flags);
1004         list_lru_add(&nfsd_file_lru, &nf->nf_lru);
1005         hlist_add_head_rcu(&nf->nf_node, &nfsd_file_hashtbl[hashval].nfb_head);
1006         ++nfsd_file_hashtbl[hashval].nfb_count;
1007         nfsd_file_hashtbl[hashval].nfb_maxcount = max(nfsd_file_hashtbl[hashval].nfb_maxcount,
1008                         nfsd_file_hashtbl[hashval].nfb_count);
1009         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
1010         if (atomic_long_inc_return(&nfsd_filecache_count) >= NFSD_FILE_LRU_THRESHOLD)
1011                 nfsd_file_gc();
1012
1013         nf->nf_mark = nfsd_file_mark_find_or_create(nf);
1014         if (nf->nf_mark)
1015                 status = nfsd_open_verified(rqstp, fhp, S_IFREG,
1016                                 may_flags, &nf->nf_file);
1017         else
1018                 status = nfserr_jukebox;
1019         /*
1020          * If construction failed, or we raced with a call to unlink()
1021          * then unhash.
1022          */
1023         if (status != nfs_ok || inode->i_nlink == 0) {
1024                 bool do_free;
1025                 spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
1026                 do_free = nfsd_file_unhash(nf);
1027                 spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
1028                 if (do_free)
1029                         nfsd_file_put_noref(nf);
1030         }
1031         clear_bit_unlock(NFSD_FILE_PENDING, &nf->nf_flags);
1032         smp_mb__after_atomic();
1033         wake_up_bit(&nf->nf_flags, NFSD_FILE_PENDING);
1034         goto out;
1035 }
1036
1037 /*
1038  * Note that fields may be added, removed or reordered in the future. Programs
1039  * scraping this file for info should test the labels to ensure they're
1040  * getting the correct field.
1041  */
1042 static int nfsd_file_cache_stats_show(struct seq_file *m, void *v)
1043 {
1044         unsigned int i, count = 0, longest = 0;
1045         unsigned long hits = 0;
1046
1047         /*
1048          * No need for spinlocks here since we're not terribly interested in
1049          * accuracy. We do take the nfsd_mutex simply to ensure that we
1050          * don't end up racing with server shutdown
1051          */
1052         mutex_lock(&nfsd_mutex);
1053         if (nfsd_file_hashtbl) {
1054                 for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
1055                         count += nfsd_file_hashtbl[i].nfb_count;
1056                         longest = max(longest, nfsd_file_hashtbl[i].nfb_count);
1057                 }
1058         }
1059         mutex_unlock(&nfsd_mutex);
1060
1061         for_each_possible_cpu(i)
1062                 hits += per_cpu(nfsd_file_cache_hits, i);
1063
1064         seq_printf(m, "total entries: %u\n", count);
1065         seq_printf(m, "longest chain: %u\n", longest);
1066         seq_printf(m, "cache hits:    %lu\n", hits);
1067         return 0;
1068 }
1069
1070 int nfsd_file_cache_stats_open(struct inode *inode, struct file *file)
1071 {
1072         return single_open(file, nfsd_file_cache_stats_show, NULL);
1073 }