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_flush(struct nfsd_file *nf)
 241 {
 242         if (nf->nf_file && vfs_fsync(nf->nf_file, 1) != 0)
 243                 nfsd_reset_write_verifier(net_generic(nf->nf_net, nfsd_net_id));
 244 }
 245
 246 static void
 247 nfsd_file_do_unhash(struct nfsd_file *nf)
 248 {
 249         lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 250
 251         trace_nfsd_file_unhash(nf);
 252
 253         if (nfsd_file_check_write_error(nf))
 254                 nfsd_reset_write_verifier(net_generic(nf->nf_net, nfsd_net_id));
 255         --nfsd_file_hashtbl[nf->nf_hashval].nfb_count;
 256         hlist_del_rcu(&nf->nf_node);
 257         atomic_long_dec(&nfsd_filecache_count);
 258 }
 259
 260 static bool
 261 nfsd_file_unhash(struct nfsd_file *nf)
 262 {
 263         if (test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
 264                 nfsd_file_do_unhash(nf);
 265                 if (!list_empty(&nf->nf_lru))
 266                         list_lru_del(&nfsd_file_lru, &nf->nf_lru);
 267                 return true;
 268         }
 269         return false;
 270 }
 271
 272 /*
 273  * Return true if the file was unhashed.
 274  */
 275 static bool
 276 nfsd_file_unhash_and_release_locked(struct nfsd_file *nf, struct list_head *dispose)
 277 {
 278         lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 279
 280         trace_nfsd_file_unhash_and_release_locked(nf);
 281         if (!nfsd_file_unhash(nf))
 282                 return false;
 283         /* keep final reference for nfsd_file_lru_dispose */
 284         if (refcount_dec_not_one(&nf->nf_ref))
 285                 return true;
 286
 287         list_add(&nf->nf_lru, dispose);
 288         return true;
 289 }
 290
 291 static void
 292 nfsd_file_put_noref(struct nfsd_file *nf)
 293 {
 294         trace_nfsd_file_put(nf);
 295
 296         if (refcount_dec_and_test(&nf->nf_ref)) {
 297                 WARN_ON(test_bit(NFSD_FILE_HASHED, &nf->nf_flags));
 298                 nfsd_file_free(nf);
 299         }
 300 }
 301
 302 void
 303 nfsd_file_put(struct nfsd_file *nf)
 304 {
 305         set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
 306         if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags) == 0) {
 307                 nfsd_file_flush(nf);
 308                 nfsd_file_put_noref(nf);
 309         } else {
 310                 nfsd_file_put_noref(nf);
 311                 if (nf->nf_file)
 312                         nfsd_file_schedule_laundrette();
 313         }
 314         if (atomic_long_read(&nfsd_filecache_count) >= NFSD_FILE_LRU_LIMIT)
 315                 nfsd_file_gc();
 316 }
 317
 318 struct nfsd_file *
 319 nfsd_file_get(struct nfsd_file *nf)
 320 {
 321         if (likely(refcount_inc_not_zero(&nf->nf_ref)))
 322                 return nf;
 323         return NULL;
 324 }
 325
 326 static void
 327 nfsd_file_dispose_list(struct list_head *dispose)
 328 {
 329         struct nfsd_file *nf;
 330
 331         while(!list_empty(dispose)) {
 332                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 333                 list_del(&nf->nf_lru);
 334                 nfsd_file_flush(nf);
 335                 nfsd_file_put_noref(nf);
 336         }
 337 }
 338
 339 static void
 340 nfsd_file_dispose_list_sync(struct list_head *dispose)
 341 {
 342         bool flush = false;
 343         struct nfsd_file *nf;
 344
 345         while(!list_empty(dispose)) {
 346                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 347                 list_del(&nf->nf_lru);
 348                 nfsd_file_flush(nf);
 349                 if (!refcount_dec_and_test(&nf->nf_ref))
 350                         continue;
 351                 if (nfsd_file_free(nf))
 352                         flush = true;
 353         }
 354         if (flush)
 355                 flush_delayed_fput();
 356 }
 357
 358 static void
 359 nfsd_file_list_remove_disposal(struct list_head *dst,
 360                 struct nfsd_fcache_disposal *l)
 361 {
 362         spin_lock(&l->lock);
 363         list_splice_init(&l->freeme, dst);
 364         spin_unlock(&l->lock);
 365 }
 366
 367 static void
 368 nfsd_file_list_add_disposal(struct list_head *files, struct net *net)
 369 {
 370         struct nfsd_net *nn = net_generic(net, nfsd_net_id);
 371         struct nfsd_fcache_disposal *l = nn->fcache_disposal;
 372
 373         spin_lock(&l->lock);
 374         list_splice_tail_init(files, &l->freeme);
 375         spin_unlock(&l->lock);
 376         queue_work(nfsd_filecache_wq, &l->work);
 377 }
 378
 379 static void
 380 nfsd_file_list_add_pernet(struct list_head *dst, struct list_head *src,
 381                 struct net *net)
 382 {
 383         struct nfsd_file *nf, *tmp;
 384
 385         list_for_each_entry_safe(nf, tmp, src, nf_lru) {
 386                 if (nf->nf_net == net)
 387                         list_move_tail(&nf->nf_lru, dst);
 388         }
 389 }
 390
 391 static void
 392 nfsd_file_dispose_list_delayed(struct list_head *dispose)
 393 {
 394         LIST_HEAD(list);
 395         struct nfsd_file *nf;
 396
 397         while(!list_empty(dispose)) {
 398                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 399                 nfsd_file_list_add_pernet(&list, dispose, nf->nf_net);
 400                 nfsd_file_list_add_disposal(&list, nf->nf_net);
 401         }
 402 }
 403
 404 /*
 405  * Note this can deadlock with nfsd_file_cache_purge.
 406  */
 407 static enum lru_status
 408 nfsd_file_lru_cb(struct list_head *item, struct list_lru_one *lru,
 409                  spinlock_t *lock, void *arg)
 410         __releases(lock)
 411         __acquires(lock)
 412 {
 413         struct list_head *head = arg;
 414         struct nfsd_file *nf = list_entry(item, struct nfsd_file, nf_lru);
 415
 416         /*
 417          * Do a lockless refcount check. The hashtable holds one reference, so
 418          * we look to see if anything else has a reference, or if any have
 419          * been put since the shrinker last ran. Those don't get unhashed and
 420          * released.
 421          *
 422          * Note that in the put path, we set the flag and then decrement the
 423          * counter. Here we check the counter and then test and clear the flag.
 424          * That order is deliberate to ensure that we can do this locklessly.
 425          */
 426         if (refcount_read(&nf->nf_ref) > 1)
 427                 goto out_skip;
 428
 429         /*
 430          * Don't throw out files that are still undergoing I/O or
 431          * that have uncleared errors pending.
 432          */
 433         if (nfsd_file_check_writeback(nf))
 434                 goto out_skip;
 435
 436         if (test_and_clear_bit(NFSD_FILE_REFERENCED, &nf->nf_flags))
 437                 goto out_skip;
 438
 439         if (!test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags))
 440                 goto out_skip;
 441
 442         list_lru_isolate_move(lru, &nf->nf_lru, head);
 443         return LRU_REMOVED;
 444 out_skip:
 445         return LRU_SKIP;
 446 }
 447
 448 static unsigned long
 449 nfsd_file_lru_walk_list(struct shrink_control *sc)
 450 {
 451         LIST_HEAD(head);
 452         struct nfsd_file *nf;
 453         unsigned long ret;
 454
 455         if (sc)
 456                 ret = list_lru_shrink_walk(&nfsd_file_lru, sc,
 457                                 nfsd_file_lru_cb, &head);
 458         else
 459                 ret = list_lru_walk(&nfsd_file_lru,
 460                                 nfsd_file_lru_cb,
 461                                 &head, LONG_MAX);
 462         list_for_each_entry(nf, &head, nf_lru) {
 463                 spin_lock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 464                 nfsd_file_do_unhash(nf);
 465                 spin_unlock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 466         }
 467         nfsd_file_dispose_list_delayed(&head);
 468         return ret;
 469 }
 470
 471 static void
 472 nfsd_file_gc(void)
 473 {
 474         nfsd_file_lru_walk_list(NULL);
 475 }
 476
 477 static void
 478 nfsd_file_gc_worker(struct work_struct *work)
 479 {
 480         nfsd_file_gc();
 481         nfsd_file_schedule_laundrette();
 482 }
 483
 484 static unsigned long
 485 nfsd_file_lru_count(struct shrinker *s, struct shrink_control *sc)
 486 {
 487         return list_lru_count(&nfsd_file_lru);
 488 }
 489
 490 static unsigned long
 491 nfsd_file_lru_scan(struct shrinker *s, struct shrink_control *sc)
 492 {
 493         return nfsd_file_lru_walk_list(sc);
 494 }
 495
 496 static struct shrinker  nfsd_file_shrinker = {
 497         .scan_objects = nfsd_file_lru_scan,
 498         .count_objects = nfsd_file_lru_count,
 499         .seeks = 1,
 500 };
 501
 502 static void
 503 __nfsd_file_close_inode(struct inode *inode, unsigned int hashval,
 504                         struct list_head *dispose)
 505 {
 506         struct nfsd_file        *nf;
 507         struct hlist_node       *tmp;
 508
 509         spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
 510         hlist_for_each_entry_safe(nf, tmp, &nfsd_file_hashtbl[hashval].nfb_head, nf_node) {
 511                 if (inode == nf->nf_inode)
 512                         nfsd_file_unhash_and_release_locked(nf, dispose);
 513         }
 514         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
 515 }
 516
 517 /**
 518  * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
 519  * @inode: inode of the file to attempt to remove
 520  *
 521  * Walk the whole hash bucket, looking for any files that correspond to "inode".
 522  * If any do, then unhash them and put the hashtable reference to them and
 523  * destroy any that had their last reference put. Also ensure that any of the
 524  * fputs also have their final __fput done as well.
 525  */
 526 void
 527 nfsd_file_close_inode_sync(struct inode *inode)
 528 {
 529         unsigned int            hashval = (unsigned int)hash_long(inode->i_ino,
 530                                                 NFSD_FILE_HASH_BITS);
 531         LIST_HEAD(dispose);
 532
 533         __nfsd_file_close_inode(inode, hashval, &dispose);
 534         trace_nfsd_file_close_inode_sync(inode, hashval, !list_empty(&dispose));
 535         nfsd_file_dispose_list_sync(&dispose);
 536 }
 537
 538 /**
 539  * nfsd_file_close_inode - attempt a delayed close of a nfsd_file
 540  * @inode: inode of the file to attempt to remove
 541  *
 542  * Walk the whole hash bucket, looking for any files that correspond to "inode".
 543  * If any do, then unhash them and put the hashtable reference to them and
 544  * destroy any that had their last reference put.
 545  */
 546 static void
 547 nfsd_file_close_inode(struct inode *inode)
 548 {
 549         unsigned int            hashval = (unsigned int)hash_long(inode->i_ino,
 550                                                 NFSD_FILE_HASH_BITS);
 551         LIST_HEAD(dispose);
 552
 553         __nfsd_file_close_inode(inode, hashval, &dispose);
 554         trace_nfsd_file_close_inode(inode, hashval, !list_empty(&dispose));
 555         nfsd_file_dispose_list_delayed(&dispose);
 556 }
 557
 558 /**
 559  * nfsd_file_delayed_close - close unused nfsd_files
 560  * @work: dummy
 561  *
 562  * Walk the LRU list and close any entries that have not been used since
 563  * the last scan.
 564  *
 565  * Note this can deadlock with nfsd_file_cache_purge.
 566  */
 567 static void
 568 nfsd_file_delayed_close(struct work_struct *work)
 569 {
 570         LIST_HEAD(head);
 571         struct nfsd_fcache_disposal *l = container_of(work,
 572                         struct nfsd_fcache_disposal, work);
 573
 574         nfsd_file_list_remove_disposal(&head, l);
 575         nfsd_file_dispose_list(&head);
 576 }
 577
 578 static int
 579 nfsd_file_lease_notifier_call(struct notifier_block *nb, unsigned long arg,
 580                             void *data)
 581 {
 582         struct file_lock *fl = data;
 583
 584         /* Only close files for F_SETLEASE leases */
 585         if (fl->fl_flags & FL_LEASE)
 586                 nfsd_file_close_inode_sync(file_inode(fl->fl_file));
 587         return 0;
 588 }
 589
 590 static struct notifier_block nfsd_file_lease_notifier = {
 591         .notifier_call = nfsd_file_lease_notifier_call,
 592 };
 593
 594 static int
 595 nfsd_file_fsnotify_handle_event(struct fsnotify_mark *mark, u32 mask,
 596                                 struct inode *inode, struct inode *dir,
 597                                 const struct qstr *name, u32 cookie)
 598 {
 599         if (WARN_ON_ONCE(!inode))
 600                 return 0;
 601
 602         trace_nfsd_file_fsnotify_handle_event(inode, mask);
 603
 604         /* Should be no marks on non-regular files */
 605         if (!S_ISREG(inode->i_mode)) {
 606                 WARN_ON_ONCE(1);
 607                 return 0;
 608         }
 609
 610         /* don't close files if this was not the last link */
 611         if (mask & FS_ATTRIB) {
 612                 if (inode->i_nlink)
 613                         return 0;
 614         }
 615
 616         nfsd_file_close_inode(inode);
 617         return 0;
 618 }
 619
 620
 621 static const struct fsnotify_ops nfsd_file_fsnotify_ops = {
 622         .handle_inode_event = nfsd_file_fsnotify_handle_event,
 623         .free_mark = nfsd_file_mark_free,
 624 };
 625
 626 int
 627 nfsd_file_cache_init(void)
 628 {
 629         int             ret = -ENOMEM;
 630         unsigned int    i;
 631
 632         clear_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
 633
 634         if (nfsd_file_hashtbl)
 635                 return 0;
 636
 637         nfsd_filecache_wq = alloc_workqueue("nfsd_filecache", 0, 0);
 638         if (!nfsd_filecache_wq)
 639                 goto out;
 640
 641         nfsd_file_hashtbl = kvcalloc(NFSD_FILE_HASH_SIZE,
 642                                 sizeof(*nfsd_file_hashtbl), GFP_KERNEL);
 643         if (!nfsd_file_hashtbl) {
 644                 pr_err("nfsd: unable to allocate nfsd_file_hashtbl\n");
 645                 goto out_err;
 646         }
 647
 648         nfsd_file_slab = kmem_cache_create("nfsd_file",
 649                                 sizeof(struct nfsd_file), 0, 0, NULL);
 650         if (!nfsd_file_slab) {
 651                 pr_err("nfsd: unable to create nfsd_file_slab\n");
 652                 goto out_err;
 653         }
 654
 655         nfsd_file_mark_slab = kmem_cache_create("nfsd_file_mark",
 656                                         sizeof(struct nfsd_file_mark), 0, 0, NULL);
 657         if (!nfsd_file_mark_slab) {
 658                 pr_err("nfsd: unable to create nfsd_file_mark_slab\n");
 659                 goto out_err;
 660         }
 661
 662
 663         ret = list_lru_init(&nfsd_file_lru);
 664         if (ret) {
 665                 pr_err("nfsd: failed to init nfsd_file_lru: %d\n", ret);
 666                 goto out_err;
 667         }
 668
 669         ret = register_shrinker(&nfsd_file_shrinker);
 670         if (ret) {
 671                 pr_err("nfsd: failed to register nfsd_file_shrinker: %d\n", ret);
 672                 goto out_lru;
 673         }
 674
 675         ret = lease_register_notifier(&nfsd_file_lease_notifier);
 676         if (ret) {
 677                 pr_err("nfsd: unable to register lease notifier: %d\n", ret);
 678                 goto out_shrinker;
 679         }
 680
 681         nfsd_file_fsnotify_group = fsnotify_alloc_group(&nfsd_file_fsnotify_ops);
 682         if (IS_ERR(nfsd_file_fsnotify_group)) {
 683                 pr_err("nfsd: unable to create fsnotify group: %ld\n",
 684                         PTR_ERR(nfsd_file_fsnotify_group));
 685                 ret = PTR_ERR(nfsd_file_fsnotify_group);
 686                 nfsd_file_fsnotify_group = NULL;
 687                 goto out_notifier;
 688         }
 689
 690         for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
 691                 INIT_HLIST_HEAD(&nfsd_file_hashtbl[i].nfb_head);
 692                 spin_lock_init(&nfsd_file_hashtbl[i].nfb_lock);
 693         }
 694
 695         INIT_DELAYED_WORK(&nfsd_filecache_laundrette, nfsd_file_gc_worker);
 696 out:
 697         return ret;
 698 out_notifier:
 699         lease_unregister_notifier(&nfsd_file_lease_notifier);
 700 out_shrinker:
 701         unregister_shrinker(&nfsd_file_shrinker);
 702 out_lru:
 703         list_lru_destroy(&nfsd_file_lru);
 704 out_err:
 705         kmem_cache_destroy(nfsd_file_slab);
 706         nfsd_file_slab = NULL;
 707         kmem_cache_destroy(nfsd_file_mark_slab);
 708         nfsd_file_mark_slab = NULL;
 709         kvfree(nfsd_file_hashtbl);
 710         nfsd_file_hashtbl = NULL;
 711         destroy_workqueue(nfsd_filecache_wq);
 712         nfsd_filecache_wq = NULL;
 713         goto out;
 714 }
 715
 716 /*
 717  * Note this can deadlock with nfsd_file_lru_cb.
 718  */
 719 void
 720 nfsd_file_cache_purge(struct net *net)
 721 {
 722         unsigned int            i;
 723         struct nfsd_file        *nf;
 724         struct hlist_node       *next;
 725         LIST_HEAD(dispose);
 726         bool del;
 727
 728         if (!nfsd_file_hashtbl)
 729                 return;
 730
 731         for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
 732                 struct nfsd_fcache_bucket *nfb = &nfsd_file_hashtbl[i];
 733
 734                 spin_lock(&nfb->nfb_lock);
 735                 hlist_for_each_entry_safe(nf, next, &nfb->nfb_head, nf_node) {
 736                         if (net && nf->nf_net != net)
 737                                 continue;
 738                         del = nfsd_file_unhash_and_release_locked(nf, &dispose);
 739
 740                         /*
 741                          * Deadlock detected! Something marked this entry as
 742                          * unhased, but hasn't removed it from the hash list.
 743                          */
 744                         WARN_ON_ONCE(!del);
 745                 }
 746                 spin_unlock(&nfb->nfb_lock);
 747                 nfsd_file_dispose_list(&dispose);
 748         }
 749 }
 750
 751 static struct nfsd_fcache_disposal *
 752 nfsd_alloc_fcache_disposal(void)
 753 {
 754         struct nfsd_fcache_disposal *l;
 755
 756         l = kmalloc(sizeof(*l), GFP_KERNEL);
 757         if (!l)
 758                 return NULL;
 759         INIT_WORK(&l->work, nfsd_file_delayed_close);
 760         spin_lock_init(&l->lock);
 761         INIT_LIST_HEAD(&l->freeme);
 762         return l;
 763 }
 764
 765 static void
 766 nfsd_free_fcache_disposal(struct nfsd_fcache_disposal *l)
 767 {
 768         cancel_work_sync(&l->work);
 769         nfsd_file_dispose_list(&l->freeme);
 770         kfree(l);
 771 }
 772
 773 static void
 774 nfsd_free_fcache_disposal_net(struct net *net)
 775 {
 776         struct nfsd_net *nn = net_generic(net, nfsd_net_id);
 777         struct nfsd_fcache_disposal *l = nn->fcache_disposal;
 778
 779         nfsd_free_fcache_disposal(l);
 780 }
 781
 782 int
 783 nfsd_file_cache_start_net(struct net *net)
 784 {
 785         struct nfsd_net *nn = net_generic(net, nfsd_net_id);
 786
 787         nn->fcache_disposal = nfsd_alloc_fcache_disposal();
 788         return nn->fcache_disposal ? 0 : -ENOMEM;
 789 }
 790
 791 void
 792 nfsd_file_cache_shutdown_net(struct net *net)
 793 {
 794         nfsd_file_cache_purge(net);
 795         nfsd_free_fcache_disposal_net(net);
 796 }
 797
 798 void
 799 nfsd_file_cache_shutdown(void)
 800 {
 801         set_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
 802
 803         lease_unregister_notifier(&nfsd_file_lease_notifier);
 804         unregister_shrinker(&nfsd_file_shrinker);
 805         /*
 806          * make sure all callers of nfsd_file_lru_cb are done before
 807          * calling nfsd_file_cache_purge
 808          */
 809         cancel_delayed_work_sync(&nfsd_filecache_laundrette);
 810         nfsd_file_cache_purge(NULL);
 811         list_lru_destroy(&nfsd_file_lru);
 812         rcu_barrier();
 813         fsnotify_put_group(nfsd_file_fsnotify_group);
 814         nfsd_file_fsnotify_group = NULL;
 815         kmem_cache_destroy(nfsd_file_slab);
 816         nfsd_file_slab = NULL;
 817         fsnotify_wait_marks_destroyed();
 818         kmem_cache_destroy(nfsd_file_mark_slab);
 819         nfsd_file_mark_slab = NULL;
 820         kvfree(nfsd_file_hashtbl);
 821         nfsd_file_hashtbl = NULL;
 822         destroy_workqueue(nfsd_filecache_wq);
 823         nfsd_filecache_wq = NULL;
 824 }
 825
 826 static bool
 827 nfsd_match_cred(const struct cred *c1, const struct cred *c2)
 828 {
 829         int i;
 830
 831         if (!uid_eq(c1->fsuid, c2->fsuid))
 832                 return false;
 833         if (!gid_eq(c1->fsgid, c2->fsgid))
 834                 return false;
 835         if (c1->group_info == NULL || c2->group_info == NULL)
 836                 return c1->group_info == c2->group_info;
 837         if (c1->group_info->ngroups != c2->group_info->ngroups)
 838                 return false;
 839         for (i = 0; i < c1->group_info->ngroups; i++) {
 840                 if (!gid_eq(c1->group_info->gid[i], c2->group_info->gid[i]))
 841                         return false;
 842         }
 843         return true;
 844 }
 845
 846 static struct nfsd_file *
 847 nfsd_file_find_locked(struct inode *inode, unsigned int may_flags,
 848                         unsigned int hashval, struct net *net)
 849 {
 850         struct nfsd_file *nf;
 851         unsigned char need = may_flags & NFSD_FILE_MAY_MASK;
 852
 853         hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
 854                                  nf_node, lockdep_is_held(&nfsd_file_hashtbl[hashval].nfb_lock)) {
 855                 if (nf->nf_may != need)
 856                         continue;
 857                 if (nf->nf_inode != inode)
 858                         continue;
 859                 if (nf->nf_net != net)
 860                         continue;
 861                 if (!nfsd_match_cred(nf->nf_cred, current_cred()))
 862                         continue;
 863                 if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags))
 864                         continue;
 865                 if (nfsd_file_get(nf) != NULL)
 866                         return nf;
 867         }
 868         return NULL;
 869 }
 870
 871 /**
 872  * nfsd_file_is_cached - are there any cached open files for this fh?
 873  * @inode: inode of the file to check
 874  *
 875  * Scan the hashtable for open files that match this fh. Returns true if there
 876  * are any, and false if not.
 877  */
 878 bool
 879 nfsd_file_is_cached(struct inode *inode)
 880 {
 881         bool                    ret = false;
 882         struct nfsd_file        *nf;
 883         unsigned int            hashval;
 884
 885         hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
 886
 887         rcu_read_lock();
 888         hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
 889                                  nf_node) {
 890                 if (inode == nf->nf_inode) {
 891                         ret = true;
 892                         break;
 893                 }
 894         }
 895         rcu_read_unlock();
 896         trace_nfsd_file_is_cached(inode, hashval, (int)ret);
 897         return ret;
 898 }
 899
 900 __be32
 901 nfsd_file_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 902                   unsigned int may_flags, struct nfsd_file **pnf)
 903 {
 904         __be32  status;
 905         struct net *net = SVC_NET(rqstp);
 906         struct nfsd_file *nf, *new;
 907         struct inode *inode;
 908         unsigned int hashval;
 909         bool retry = true;
 910
 911         /* FIXME: skip this if fh_dentry is already set? */
 912         status = fh_verify(rqstp, fhp, S_IFREG,
 913                                 may_flags|NFSD_MAY_OWNER_OVERRIDE);
 914         if (status != nfs_ok)
 915                 return status;
 916
 917         inode = d_inode(fhp->fh_dentry);
 918         hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
 919 retry:
 920         rcu_read_lock();
 921         nf = nfsd_file_find_locked(inode, may_flags, hashval, net);
 922         rcu_read_unlock();
 923         if (nf)
 924                 goto wait_for_construction;
 925
 926         new = nfsd_file_alloc(inode, may_flags, hashval, net);
 927         if (!new) {
 928                 trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags,
 929                                         NULL, nfserr_jukebox);
 930                 return nfserr_jukebox;
 931         }
 932
 933         spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
 934         nf = nfsd_file_find_locked(inode, may_flags, hashval, net);
 935         if (nf == NULL)
 936                 goto open_file;
 937         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
 938         nfsd_file_slab_free(&new->nf_rcu);
 939
 940 wait_for_construction:
 941         wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
 942
 943         /* Did construction of this file fail? */
 944         if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
 945                 if (!retry) {
 946                         status = nfserr_jukebox;
 947                         goto out;
 948                 }
 949                 retry = false;
 950                 nfsd_file_put_noref(nf);
 951                 goto retry;
 952         }
 953
 954         this_cpu_inc(nfsd_file_cache_hits);
 955
 956         if (!(may_flags & NFSD_MAY_NOT_BREAK_LEASE)) {
 957                 bool write = (may_flags & NFSD_MAY_WRITE);
 958
 959                 if (test_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags) ||
 960                     (test_bit(NFSD_FILE_BREAK_WRITE, &nf->nf_flags) && write)) {
 961                         status = nfserrno(nfsd_open_break_lease(
 962                                         file_inode(nf->nf_file), may_flags));
 963                         if (status == nfs_ok) {
 964                                 clear_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags);
 965                                 if (write)
 966                                         clear_bit(NFSD_FILE_BREAK_WRITE,
 967                                                   &nf->nf_flags);
 968                         }
 969                 }
 970         }
 971 out:
 972         if (status == nfs_ok) {
 973                 *pnf = nf;
 974         } else {
 975                 nfsd_file_put(nf);
 976                 nf = NULL;
 977         }
 978
 979         trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags, nf, status);
 980         return status;
 981 open_file:
 982         nf = new;
 983         /* Take reference for the hashtable */
 984         refcount_inc(&nf->nf_ref);
 985         __set_bit(NFSD_FILE_HASHED, &nf->nf_flags);
 986         __set_bit(NFSD_FILE_PENDING, &nf->nf_flags);
 987         list_lru_add(&nfsd_file_lru, &nf->nf_lru);
 988         hlist_add_head_rcu(&nf->nf_node, &nfsd_file_hashtbl[hashval].nfb_head);
 989         ++nfsd_file_hashtbl[hashval].nfb_count;
 990         nfsd_file_hashtbl[hashval].nfb_maxcount = max(nfsd_file_hashtbl[hashval].nfb_maxcount,
 991                         nfsd_file_hashtbl[hashval].nfb_count);
 992         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
 993         if (atomic_long_inc_return(&nfsd_filecache_count) >= NFSD_FILE_LRU_THRESHOLD)
 994                 nfsd_file_gc();
 995
 996         nf->nf_mark = nfsd_file_mark_find_or_create(nf);
 997         if (nf->nf_mark)
 998                 status = nfsd_open_verified(rqstp, fhp, S_IFREG,
 999                                 may_flags, &nf->nf_file);
1000         else
1001                 status = nfserr_jukebox;
1002         /*
1003          * If construction failed, or we raced with a call to unlink()
1004          * then unhash.
1005          */
1006         if (status != nfs_ok || inode->i_nlink == 0) {
1007                 bool do_free;
1008                 spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
1009                 do_free = nfsd_file_unhash(nf);
1010                 spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
1011                 if (do_free)
1012                         nfsd_file_put_noref(nf);
1013         }
1014         clear_bit_unlock(NFSD_FILE_PENDING, &nf->nf_flags);
1015         smp_mb__after_atomic();
1016         wake_up_bit(&nf->nf_flags, NFSD_FILE_PENDING);
1017         goto out;
1018 }
1019
1020 /*
1021  * Note that fields may be added, removed or reordered in the future. Programs
1022  * scraping this file for info should test the labels to ensure they're
1023  * getting the correct field.
1024  */
1025 static int nfsd_file_cache_stats_show(struct seq_file *m, void *v)
1026 {
1027         unsigned int i, count = 0, longest = 0;
1028         unsigned long hits = 0;
1029
1030         /*
1031          * No need for spinlocks here since we're not terribly interested in
1032          * accuracy. We do take the nfsd_mutex simply to ensure that we
1033          * don't end up racing with server shutdown
1034          */
1035         mutex_lock(&nfsd_mutex);
1036         if (nfsd_file_hashtbl) {
1037                 for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
1038                         count += nfsd_file_hashtbl[i].nfb_count;
1039                         longest = max(longest, nfsd_file_hashtbl[i].nfb_count);
1040                 }
1041         }
1042         mutex_unlock(&nfsd_mutex);
1043
1044         for_each_possible_cpu(i)
1045                 hits += per_cpu(nfsd_file_cache_hits, i);
1046
1047         seq_printf(m, "total entries: %u\n", count);
1048         seq_printf(m, "longest chain: %u\n", longest);
1049         seq_printf(m, "cache hits:    %lu\n", hits);
1050         return 0;
1051 }
1052
1053 int nfsd_file_cache_stats_open(struct inode *inode, struct file *file)
1054 {
1055         return single_open(file, nfsd_file_cache_stats_show, NULL);
1056 }