fs/nfsd/filecache.c

   1 /*
   2  * Open file cache.
   3  *
   4  * (c) 2015 - Jeff Layton <[email protected]>
   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                 fsnotify_group_lock(nfsd_file_fsnotify_group);
 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                         fsnotify_group_unlock(nfsd_file_fsnotify_group);
 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                         fsnotify_group_unlock(nfsd_file_fsnotify_group);
 139                 }
 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                 refcount_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                 refcount_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_flush(struct nfsd_file *nf)
 242 {
 243         if (nf->nf_file && vfs_fsync(nf->nf_file, 1) != 0)
 244                 nfsd_reset_write_verifier(net_generic(nf->nf_net, nfsd_net_id));
 245 }
 246
 247 static void
 248 nfsd_file_do_unhash(struct nfsd_file *nf)
 249 {
 250         lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 251
 252         trace_nfsd_file_unhash(nf);
 253
 254         if (nfsd_file_check_write_error(nf))
 255                 nfsd_reset_write_verifier(net_generic(nf->nf_net, nfsd_net_id));
 256         --nfsd_file_hashtbl[nf->nf_hashval].nfb_count;
 257         hlist_del_rcu(&nf->nf_node);
 258         atomic_long_dec(&nfsd_filecache_count);
 259 }
 260
 261 static bool
 262 nfsd_file_unhash(struct nfsd_file *nf)
 263 {
 264         if (test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
 265                 nfsd_file_do_unhash(nf);
 266                 if (!list_empty(&nf->nf_lru))
 267                         list_lru_del(&nfsd_file_lru, &nf->nf_lru);
 268                 return true;
 269         }
 270         return false;
 271 }
 272
 273 /*
 274  * Return true if the file was unhashed.
 275  */
 276 static bool
 277 nfsd_file_unhash_and_release_locked(struct nfsd_file *nf, struct list_head *dispose)
 278 {
 279         lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 280
 281         trace_nfsd_file_unhash_and_release_locked(nf);
 282         if (!nfsd_file_unhash(nf))
 283                 return false;
 284         /* keep final reference for nfsd_file_lru_dispose */
 285         if (refcount_dec_not_one(&nf->nf_ref))
 286                 return true;
 287
 288         list_add(&nf->nf_lru, dispose);
 289         return true;
 290 }
 291
 292 static void
 293 nfsd_file_put_noref(struct nfsd_file *nf)
 294 {
 295         trace_nfsd_file_put(nf);
 296
 297         if (refcount_dec_and_test(&nf->nf_ref)) {
 298                 WARN_ON(test_bit(NFSD_FILE_HASHED, &nf->nf_flags));
 299                 nfsd_file_free(nf);
 300         }
 301 }
 302
 303 void
 304 nfsd_file_put(struct nfsd_file *nf)
 305 {
 306         might_sleep();
 307
 308         set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
 309         if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags) == 0) {
 310                 nfsd_file_flush(nf);
 311                 nfsd_file_put_noref(nf);
 312         } else {
 313                 nfsd_file_put_noref(nf);
 314                 if (nf->nf_file)
 315                         nfsd_file_schedule_laundrette();
 316         }
 317         if (atomic_long_read(&nfsd_filecache_count) >= NFSD_FILE_LRU_LIMIT)
 318                 nfsd_file_gc();
 319 }
 320
 321 struct nfsd_file *
 322 nfsd_file_get(struct nfsd_file *nf)
 323 {
 324         if (likely(refcount_inc_not_zero(&nf->nf_ref)))
 325                 return nf;
 326         return NULL;
 327 }
 328
 329 static void
 330 nfsd_file_dispose_list(struct list_head *dispose)
 331 {
 332         struct nfsd_file *nf;
 333
 334         while(!list_empty(dispose)) {
 335                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 336                 list_del(&nf->nf_lru);
 337                 nfsd_file_flush(nf);
 338                 nfsd_file_put_noref(nf);
 339         }
 340 }
 341
 342 static void
 343 nfsd_file_dispose_list_sync(struct list_head *dispose)
 344 {
 345         bool flush = false;
 346         struct nfsd_file *nf;
 347
 348         while(!list_empty(dispose)) {
 349                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 350                 list_del(&nf->nf_lru);
 351                 nfsd_file_flush(nf);
 352                 if (!refcount_dec_and_test(&nf->nf_ref))
 353                         continue;
 354                 if (nfsd_file_free(nf))
 355                         flush = true;
 356         }
 357         if (flush)
 358                 flush_delayed_fput();
 359 }
 360
 361 static void
 362 nfsd_file_list_remove_disposal(struct list_head *dst,
 363                 struct nfsd_fcache_disposal *l)
 364 {
 365         spin_lock(&l->lock);
 366         list_splice_init(&l->freeme, dst);
 367         spin_unlock(&l->lock);
 368 }
 369
 370 static void
 371 nfsd_file_list_add_disposal(struct list_head *files, struct net *net)
 372 {
 373         struct nfsd_net *nn = net_generic(net, nfsd_net_id);
 374         struct nfsd_fcache_disposal *l = nn->fcache_disposal;
 375
 376         spin_lock(&l->lock);
 377         list_splice_tail_init(files, &l->freeme);
 378         spin_unlock(&l->lock);
 379         queue_work(nfsd_filecache_wq, &l->work);
 380 }
 381
 382 static void
 383 nfsd_file_list_add_pernet(struct list_head *dst, struct list_head *src,
 384                 struct net *net)
 385 {
 386         struct nfsd_file *nf, *tmp;
 387
 388         list_for_each_entry_safe(nf, tmp, src, nf_lru) {
 389                 if (nf->nf_net == net)
 390                         list_move_tail(&nf->nf_lru, dst);
 391         }
 392 }
 393
 394 static void
 395 nfsd_file_dispose_list_delayed(struct list_head *dispose)
 396 {
 397         LIST_HEAD(list);
 398         struct nfsd_file *nf;
 399
 400         while(!list_empty(dispose)) {
 401                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 402                 nfsd_file_list_add_pernet(&list, dispose, nf->nf_net);
 403                 nfsd_file_list_add_disposal(&list, nf->nf_net);
 404         }
 405 }
 406
 407 /*
 408  * Note this can deadlock with nfsd_file_cache_purge.
 409  */
 410 static enum lru_status
 411 nfsd_file_lru_cb(struct list_head *item, struct list_lru_one *lru,
 412                  spinlock_t *lock, void *arg)
 413         __releases(lock)
 414         __acquires(lock)
 415 {
 416         struct list_head *head = arg;
 417         struct nfsd_file *nf = list_entry(item, struct nfsd_file, nf_lru);
 418
 419         /*
 420          * Do a lockless refcount check. The hashtable holds one reference, so
 421          * we look to see if anything else has a reference, or if any have
 422          * been put since the shrinker last ran. Those don't get unhashed and
 423          * released.
 424          *
 425          * Note that in the put path, we set the flag and then decrement the
 426          * counter. Here we check the counter and then test and clear the flag.
 427          * That order is deliberate to ensure that we can do this locklessly.
 428          */
 429         if (refcount_read(&nf->nf_ref) > 1)
 430                 goto out_skip;
 431
 432         /*
 433          * Don't throw out files that are still undergoing I/O or
 434          * that have uncleared errors pending.
 435          */
 436         if (nfsd_file_check_writeback(nf))
 437                 goto out_skip;
 438
 439         if (test_and_clear_bit(NFSD_FILE_REFERENCED, &nf->nf_flags))
 440                 goto out_skip;
 441
 442         if (!test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags))
 443                 goto out_skip;
 444
 445         list_lru_isolate_move(lru, &nf->nf_lru, head);
 446         return LRU_REMOVED;
 447 out_skip:
 448         return LRU_SKIP;
 449 }
 450
 451 static unsigned long
 452 nfsd_file_lru_walk_list(struct shrink_control *sc)
 453 {
 454         LIST_HEAD(head);
 455         struct nfsd_file *nf;
 456         unsigned long ret;
 457
 458         if (sc)
 459                 ret = list_lru_shrink_walk(&nfsd_file_lru, sc,
 460                                 nfsd_file_lru_cb, &head);
 461         else
 462                 ret = list_lru_walk(&nfsd_file_lru,
 463                                 nfsd_file_lru_cb,
 464                                 &head, LONG_MAX);
 465         list_for_each_entry(nf, &head, nf_lru) {
 466                 spin_lock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 467                 nfsd_file_do_unhash(nf);
 468                 spin_unlock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 469         }
 470         nfsd_file_dispose_list_delayed(&head);
 471         return ret;
 472 }
 473
 474 static void
 475 nfsd_file_gc(void)
 476 {
 477         nfsd_file_lru_walk_list(NULL);
 478 }
 479
 480 static void
 481 nfsd_file_gc_worker(struct work_struct *work)
 482 {
 483         nfsd_file_gc();
 484         nfsd_file_schedule_laundrette();
 485 }
 486
 487 static unsigned long
 488 nfsd_file_lru_count(struct shrinker *s, struct shrink_control *sc)
 489 {
 490         return list_lru_count(&nfsd_file_lru);
 491 }
 492
 493 static unsigned long
 494 nfsd_file_lru_scan(struct shrinker *s, struct shrink_control *sc)
 495 {
 496         return nfsd_file_lru_walk_list(sc);
 497 }
 498
 499 static struct shrinker  nfsd_file_shrinker = {
 500         .scan_objects = nfsd_file_lru_scan,
 501         .count_objects = nfsd_file_lru_count,
 502         .seeks = 1,
 503 };
 504
 505 static void
 506 __nfsd_file_close_inode(struct inode *inode, unsigned int hashval,
 507                         struct list_head *dispose)
 508 {
 509         struct nfsd_file        *nf;
 510         struct hlist_node       *tmp;
 511
 512         spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
 513         hlist_for_each_entry_safe(nf, tmp, &nfsd_file_hashtbl[hashval].nfb_head, nf_node) {
 514                 if (inode == nf->nf_inode)
 515                         nfsd_file_unhash_and_release_locked(nf, dispose);
 516         }
 517         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
 518 }
 519
 520 /**
 521  * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
 522  * @inode: inode of the file to attempt to remove
 523  *
 524  * Walk the whole hash bucket, looking for any files that correspond to "inode".
 525  * If any do, then unhash them and put the hashtable reference to them and
 526  * destroy any that had their last reference put. Also ensure that any of the
 527  * fputs also have their final __fput done as well.
 528  */
 529 void
 530 nfsd_file_close_inode_sync(struct inode *inode)
 531 {
 532         unsigned int            hashval = (unsigned int)hash_long(inode->i_ino,
 533                                                 NFSD_FILE_HASH_BITS);
 534         LIST_HEAD(dispose);
 535
 536         __nfsd_file_close_inode(inode, hashval, &dispose);
 537         trace_nfsd_file_close_inode_sync(inode, hashval, !list_empty(&dispose));
 538         nfsd_file_dispose_list_sync(&dispose);
 539 }
 540
 541 /**
 542  * nfsd_file_close_inode - attempt a delayed close of a nfsd_file
 543  * @inode: inode of the file to attempt to remove
 544  *
 545  * Walk the whole hash bucket, looking for any files that correspond to "inode".
 546  * If any do, then unhash them and put the hashtable reference to them and
 547  * destroy any that had their last reference put.
 548  */
 549 static void
 550 nfsd_file_close_inode(struct inode *inode)
 551 {
 552         unsigned int            hashval = (unsigned int)hash_long(inode->i_ino,
 553                                                 NFSD_FILE_HASH_BITS);
 554         LIST_HEAD(dispose);
 555
 556         __nfsd_file_close_inode(inode, hashval, &dispose);
 557         trace_nfsd_file_close_inode(inode, hashval, !list_empty(&dispose));
 558         nfsd_file_dispose_list_delayed(&dispose);
 559 }
 560
 561 /**
 562  * nfsd_file_delayed_close - close unused nfsd_files
 563  * @work: dummy
 564  *
 565  * Walk the LRU list and close any entries that have not been used since
 566  * the last scan.
 567  *
 568  * Note this can deadlock with nfsd_file_cache_purge.
 569  */
 570 static void
 571 nfsd_file_delayed_close(struct work_struct *work)
 572 {
 573         LIST_HEAD(head);
 574         struct nfsd_fcache_disposal *l = container_of(work,
 575                         struct nfsd_fcache_disposal, work);
 576
 577         nfsd_file_list_remove_disposal(&head, l);
 578         nfsd_file_dispose_list(&head);
 579 }
 580
 581 static int
 582 nfsd_file_lease_notifier_call(struct notifier_block *nb, unsigned long arg,
 583                             void *data)
 584 {
 585         struct file_lock *fl = data;
 586
 587         /* Only close files for F_SETLEASE leases */
 588         if (fl->fl_flags & FL_LEASE)
 589                 nfsd_file_close_inode_sync(file_inode(fl->fl_file));
 590         return 0;
 591 }
 592
 593 static struct notifier_block nfsd_file_lease_notifier = {
 594         .notifier_call = nfsd_file_lease_notifier_call,
 595 };
 596
 597 static int
 598 nfsd_file_fsnotify_handle_event(struct fsnotify_mark *mark, u32 mask,
 599                                 struct inode *inode, struct inode *dir,
 600                                 const struct qstr *name, u32 cookie)
 601 {
 602         if (WARN_ON_ONCE(!inode))
 603                 return 0;
 604
 605         trace_nfsd_file_fsnotify_handle_event(inode, mask);
 606
 607         /* Should be no marks on non-regular files */
 608         if (!S_ISREG(inode->i_mode)) {
 609                 WARN_ON_ONCE(1);
 610                 return 0;
 611         }
 612
 613         /* don't close files if this was not the last link */
 614         if (mask & FS_ATTRIB) {
 615                 if (inode->i_nlink)
 616                         return 0;
 617         }
 618
 619         nfsd_file_close_inode(inode);
 620         return 0;
 621 }
 622
 623
 624 static const struct fsnotify_ops nfsd_file_fsnotify_ops = {
 625         .handle_inode_event = nfsd_file_fsnotify_handle_event,
 626         .free_mark = nfsd_file_mark_free,
 627 };
 628
 629 int
 630 nfsd_file_cache_init(void)
 631 {
 632         int             ret = -ENOMEM;
 633         unsigned int    i;
 634
 635         clear_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
 636
 637         if (nfsd_file_hashtbl)
 638                 return 0;
 639
 640         nfsd_filecache_wq = alloc_workqueue("nfsd_filecache", 0, 0);
 641         if (!nfsd_filecache_wq)
 642                 goto out;
 643
 644         nfsd_file_hashtbl = kvcalloc(NFSD_FILE_HASH_SIZE,
 645                                 sizeof(*nfsd_file_hashtbl), GFP_KERNEL);
 646         if (!nfsd_file_hashtbl) {
 647                 pr_err("nfsd: unable to allocate nfsd_file_hashtbl\n");
 648                 goto out_err;
 649         }
 650
 651         nfsd_file_slab = kmem_cache_create("nfsd_file",
 652                                 sizeof(struct nfsd_file), 0, 0, NULL);
 653         if (!nfsd_file_slab) {
 654                 pr_err("nfsd: unable to create nfsd_file_slab\n");
 655                 goto out_err;
 656         }
 657
 658         nfsd_file_mark_slab = kmem_cache_create("nfsd_file_mark",
 659                                         sizeof(struct nfsd_file_mark), 0, 0, NULL);
 660         if (!nfsd_file_mark_slab) {
 661                 pr_err("nfsd: unable to create nfsd_file_mark_slab\n");
 662                 goto out_err;
 663         }
 664
 665
 666         ret = list_lru_init(&nfsd_file_lru);
 667         if (ret) {
 668                 pr_err("nfsd: failed to init nfsd_file_lru: %d\n", ret);
 669                 goto out_err;
 670         }
 671
 672         ret = register_shrinker(&nfsd_file_shrinker);
 673         if (ret) {
 674                 pr_err("nfsd: failed to register nfsd_file_shrinker: %d\n", ret);
 675                 goto out_lru;
 676         }
 677
 678         ret = lease_register_notifier(&nfsd_file_lease_notifier);
 679         if (ret) {
 680                 pr_err("nfsd: unable to register lease notifier: %d\n", ret);
 681                 goto out_shrinker;
 682         }
 683
 684         nfsd_file_fsnotify_group = fsnotify_alloc_group(&nfsd_file_fsnotify_ops,
 685                                                         FSNOTIFY_GROUP_NOFS);
 686         if (IS_ERR(nfsd_file_fsnotify_group)) {
 687                 pr_err("nfsd: unable to create fsnotify group: %ld\n",
 688                         PTR_ERR(nfsd_file_fsnotify_group));
 689                 ret = PTR_ERR(nfsd_file_fsnotify_group);
 690                 nfsd_file_fsnotify_group = NULL;
 691                 goto out_notifier;
 692         }
 693
 694         for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
 695                 INIT_HLIST_HEAD(&nfsd_file_hashtbl[i].nfb_head);
 696                 spin_lock_init(&nfsd_file_hashtbl[i].nfb_lock);
 697         }
 698
 699         INIT_DELAYED_WORK(&nfsd_filecache_laundrette, nfsd_file_gc_worker);
 700 out:
 701         return ret;
 702 out_notifier:
 703         lease_unregister_notifier(&nfsd_file_lease_notifier);
 704 out_shrinker:
 705         unregister_shrinker(&nfsd_file_shrinker);
 706 out_lru:
 707         list_lru_destroy(&nfsd_file_lru);
 708 out_err:
 709         kmem_cache_destroy(nfsd_file_slab);
 710         nfsd_file_slab = NULL;
 711         kmem_cache_destroy(nfsd_file_mark_slab);
 712         nfsd_file_mark_slab = NULL;
 713         kvfree(nfsd_file_hashtbl);
 714         nfsd_file_hashtbl = NULL;
 715         destroy_workqueue(nfsd_filecache_wq);
 716         nfsd_filecache_wq = NULL;
 717         goto out;
 718 }
 719
 720 /*
 721  * Note this can deadlock with nfsd_file_lru_cb.
 722  */
 723 void
 724 nfsd_file_cache_purge(struct net *net)
 725 {
 726         unsigned int            i;
 727         struct nfsd_file        *nf;
 728         struct hlist_node       *next;
 729         LIST_HEAD(dispose);
 730         bool del;
 731
 732         if (!nfsd_file_hashtbl)
 733                 return;
 734
 735         for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
 736                 struct nfsd_fcache_bucket *nfb = &nfsd_file_hashtbl[i];
 737
 738                 spin_lock(&nfb->nfb_lock);
 739                 hlist_for_each_entry_safe(nf, next, &nfb->nfb_head, nf_node) {
 740                         if (net && nf->nf_net != net)
 741                                 continue;
 742                         del = nfsd_file_unhash_and_release_locked(nf, &dispose);
 743
 744                         /*
 745                          * Deadlock detected! Something marked this entry as
 746                          * unhased, but hasn't removed it from the hash list.
 747                          */
 748                         WARN_ON_ONCE(!del);
 749                 }
 750                 spin_unlock(&nfb->nfb_lock);
 751                 nfsd_file_dispose_list(&dispose);
 752         }
 753 }
 754
 755 static struct nfsd_fcache_disposal *
 756 nfsd_alloc_fcache_disposal(void)
 757 {
 758         struct nfsd_fcache_disposal *l;
 759
 760         l = kmalloc(sizeof(*l), GFP_KERNEL);
 761         if (!l)
 762                 return NULL;
 763         INIT_WORK(&l->work, nfsd_file_delayed_close);
 764         spin_lock_init(&l->lock);
 765         INIT_LIST_HEAD(&l->freeme);
 766         return l;
 767 }
 768
 769 static void
 770 nfsd_free_fcache_disposal(struct nfsd_fcache_disposal *l)
 771 {
 772         cancel_work_sync(&l->work);
 773         nfsd_file_dispose_list(&l->freeme);
 774         kfree(l);
 775 }
 776
 777 static void
 778 nfsd_free_fcache_disposal_net(struct net *net)
 779 {
 780         struct nfsd_net *nn = net_generic(net, nfsd_net_id);
 781         struct nfsd_fcache_disposal *l = nn->fcache_disposal;
 782
 783         nfsd_free_fcache_disposal(l);
 784 }
 785
 786 int
 787 nfsd_file_cache_start_net(struct net *net)
 788 {
 789         struct nfsd_net *nn = net_generic(net, nfsd_net_id);
 790
 791         nn->fcache_disposal = nfsd_alloc_fcache_disposal();
 792         return nn->fcache_disposal ? 0 : -ENOMEM;
 793 }
 794
 795 void
 796 nfsd_file_cache_shutdown_net(struct net *net)
 797 {
 798         nfsd_file_cache_purge(net);
 799         nfsd_free_fcache_disposal_net(net);
 800 }
 801
 802 void
 803 nfsd_file_cache_shutdown(void)
 804 {
 805         set_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
 806
 807         lease_unregister_notifier(&nfsd_file_lease_notifier);
 808         unregister_shrinker(&nfsd_file_shrinker);
 809         /*
 810          * make sure all callers of nfsd_file_lru_cb are done before
 811          * calling nfsd_file_cache_purge
 812          */
 813         cancel_delayed_work_sync(&nfsd_filecache_laundrette);
 814         nfsd_file_cache_purge(NULL);
 815         list_lru_destroy(&nfsd_file_lru);
 816         rcu_barrier();
 817         fsnotify_put_group(nfsd_file_fsnotify_group);
 818         nfsd_file_fsnotify_group = NULL;
 819         kmem_cache_destroy(nfsd_file_slab);
 820         nfsd_file_slab = NULL;
 821         fsnotify_wait_marks_destroyed();
 822         kmem_cache_destroy(nfsd_file_mark_slab);
 823         nfsd_file_mark_slab = NULL;
 824         kvfree(nfsd_file_hashtbl);
 825         nfsd_file_hashtbl = NULL;
 826         destroy_workqueue(nfsd_filecache_wq);
 827         nfsd_filecache_wq = NULL;
 828 }
 829
 830 static bool
 831 nfsd_match_cred(const struct cred *c1, const struct cred *c2)
 832 {
 833         int i;
 834
 835         if (!uid_eq(c1->fsuid, c2->fsuid))
 836                 return false;
 837         if (!gid_eq(c1->fsgid, c2->fsgid))
 838                 return false;
 839         if (c1->group_info == NULL || c2->group_info == NULL)
 840                 return c1->group_info == c2->group_info;
 841         if (c1->group_info->ngroups != c2->group_info->ngroups)
 842                 return false;
 843         for (i = 0; i < c1->group_info->ngroups; i++) {
 844                 if (!gid_eq(c1->group_info->gid[i], c2->group_info->gid[i]))
 845                         return false;
 846         }
 847         return true;
 848 }
 849
 850 static struct nfsd_file *
 851 nfsd_file_find_locked(struct inode *inode, unsigned int may_flags,
 852                         unsigned int hashval, struct net *net)
 853 {
 854         struct nfsd_file *nf;
 855         unsigned char need = may_flags & NFSD_FILE_MAY_MASK;
 856
 857         hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
 858                                  nf_node, lockdep_is_held(&nfsd_file_hashtbl[hashval].nfb_lock)) {
 859                 if (nf->nf_may != need)
 860                         continue;
 861                 if (nf->nf_inode != inode)
 862                         continue;
 863                 if (nf->nf_net != net)
 864                         continue;
 865                 if (!nfsd_match_cred(nf->nf_cred, current_cred()))
 866                         continue;
 867                 if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags))
 868                         continue;
 869                 if (nfsd_file_get(nf) != NULL)
 870                         return nf;
 871         }
 872         return NULL;
 873 }
 874
 875 /**
 876  * nfsd_file_is_cached - are there any cached open files for this fh?
 877  * @inode: inode of the file to check
 878  *
 879  * Scan the hashtable for open files that match this fh. Returns true if there
 880  * are any, and false if not.
 881  */
 882 bool
 883 nfsd_file_is_cached(struct inode *inode)
 884 {
 885         bool                    ret = false;
 886         struct nfsd_file        *nf;
 887         unsigned int            hashval;
 888
 889         hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
 890
 891         rcu_read_lock();
 892         hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
 893                                  nf_node) {
 894                 if (inode == nf->nf_inode) {
 895                         ret = true;
 896                         break;
 897                 }
 898         }
 899         rcu_read_unlock();
 900         trace_nfsd_file_is_cached(inode, hashval, (int)ret);
 901         return ret;
 902 }
 903
 904 static __be32
 905 nfsd_do_file_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 906                      unsigned int may_flags, struct nfsd_file **pnf, bool open)
 907 {
 908         __be32  status;
 909         struct net *net = SVC_NET(rqstp);
 910         struct nfsd_file *nf, *new;
 911         struct inode *inode;
 912         unsigned int hashval;
 913         bool retry = true;
 914
 915         /* FIXME: skip this if fh_dentry is already set? */
 916         status = fh_verify(rqstp, fhp, S_IFREG,
 917                                 may_flags|NFSD_MAY_OWNER_OVERRIDE);
 918         if (status != nfs_ok)
 919                 return status;
 920
 921         inode = d_inode(fhp->fh_dentry);
 922         hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
 923 retry:
 924         rcu_read_lock();
 925         nf = nfsd_file_find_locked(inode, may_flags, hashval, net);
 926         rcu_read_unlock();
 927         if (nf)
 928                 goto wait_for_construction;
 929
 930         new = nfsd_file_alloc(inode, may_flags, hashval, net);
 931         if (!new) {
 932                 trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags,
 933                                         NULL, nfserr_jukebox);
 934                 return nfserr_jukebox;
 935         }
 936
 937         spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
 938         nf = nfsd_file_find_locked(inode, may_flags, hashval, net);
 939         if (nf == NULL)
 940                 goto open_file;
 941         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
 942         nfsd_file_slab_free(&new->nf_rcu);
 943
 944 wait_for_construction:
 945         wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
 946
 947         /* Did construction of this file fail? */
 948         if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
 949                 if (!retry) {
 950                         status = nfserr_jukebox;
 951                         goto out;
 952                 }
 953                 retry = false;
 954                 nfsd_file_put_noref(nf);
 955                 goto retry;
 956         }
 957
 958         this_cpu_inc(nfsd_file_cache_hits);
 959
 960         if (!(may_flags & NFSD_MAY_NOT_BREAK_LEASE)) {
 961                 bool write = (may_flags & NFSD_MAY_WRITE);
 962
 963                 if (test_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags) ||
 964                     (test_bit(NFSD_FILE_BREAK_WRITE, &nf->nf_flags) && write)) {
 965                         status = nfserrno(nfsd_open_break_lease(
 966                                         file_inode(nf->nf_file), may_flags));
 967                         if (status == nfs_ok) {
 968                                 clear_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags);
 969                                 if (write)
 970                                         clear_bit(NFSD_FILE_BREAK_WRITE,
 971                                                   &nf->nf_flags);
 972                         }
 973                 }
 974         }
 975 out:
 976         if (status == nfs_ok) {
 977                 *pnf = nf;
 978         } else {
 979                 nfsd_file_put(nf);
 980                 nf = NULL;
 981         }
 982
 983         trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags, nf, status);
 984         return status;
 985 open_file:
 986         nf = new;
 987         /* Take reference for the hashtable */
 988         refcount_inc(&nf->nf_ref);
 989         __set_bit(NFSD_FILE_HASHED, &nf->nf_flags);
 990         __set_bit(NFSD_FILE_PENDING, &nf->nf_flags);
 991         list_lru_add(&nfsd_file_lru, &nf->nf_lru);
 992         hlist_add_head_rcu(&nf->nf_node, &nfsd_file_hashtbl[hashval].nfb_head);
 993         ++nfsd_file_hashtbl[hashval].nfb_count;
 994         nfsd_file_hashtbl[hashval].nfb_maxcount = max(nfsd_file_hashtbl[hashval].nfb_maxcount,
 995                         nfsd_file_hashtbl[hashval].nfb_count);
 996         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
 997         if (atomic_long_inc_return(&nfsd_filecache_count) >= NFSD_FILE_LRU_THRESHOLD)
 998                 nfsd_file_gc();
 999
1000         nf->nf_mark = nfsd_file_mark_find_or_create(nf);
1001         if (nf->nf_mark) {
1002                 if (open) {
1003                         status = nfsd_open_verified(rqstp, fhp, may_flags,
1004                                                     &nf->nf_file);
1005                         trace_nfsd_file_open(nf, status);
1006                 } else
1007                         status = nfs_ok;
1008         } else
1009                 status = nfserr_jukebox;
1010         /*
1011          * If construction failed, or we raced with a call to unlink()
1012          * then unhash.
1013          */
1014         if (status != nfs_ok || inode->i_nlink == 0) {
1015                 bool do_free;
1016                 spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
1017                 do_free = nfsd_file_unhash(nf);
1018                 spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
1019                 if (do_free)
1020                         nfsd_file_put_noref(nf);
1021         }
1022         clear_bit_unlock(NFSD_FILE_PENDING, &nf->nf_flags);
1023         smp_mb__after_atomic();
1024         wake_up_bit(&nf->nf_flags, NFSD_FILE_PENDING);
1025         goto out;
1026 }
1027
1028 /**
1029  * nfsd_file_acquire - Get a struct nfsd_file with an open file
1030  * @rqstp: the RPC transaction being executed
1031  * @fhp: the NFS filehandle of the file to be opened
1032  * @may_flags: NFSD_MAY_ settings for the file
1033  * @pnf: OUT: new or found "struct nfsd_file" object
1034  *
1035  * Returns nfs_ok and sets @pnf on success; otherwise an nfsstat in
1036  * network byte order is returned.
1037  */
1038 __be32
1039 nfsd_file_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
1040                   unsigned int may_flags, struct nfsd_file **pnf)
1041 {
1042         return nfsd_do_file_acquire(rqstp, fhp, may_flags, pnf, true);
1043 }
1044
1045 /**
1046  * nfsd_file_create - Get a struct nfsd_file, do not open
1047  * @rqstp: the RPC transaction being executed
1048  * @fhp: the NFS filehandle of the file just created
1049  * @may_flags: NFSD_MAY_ settings for the file
1050  * @pnf: OUT: new or found "struct nfsd_file" object
1051  *
1052  * Returns nfs_ok and sets @pnf on success; otherwise an nfsstat in
1053  * network byte order is returned.
1054  */
1055 __be32
1056 nfsd_file_create(struct svc_rqst *rqstp, struct svc_fh *fhp,
1057                  unsigned int may_flags, struct nfsd_file **pnf)
1058 {
1059         return nfsd_do_file_acquire(rqstp, fhp, may_flags, pnf, false);
1060 }
1061
1062 /*
1063  * Note that fields may be added, removed or reordered in the future. Programs
1064  * scraping this file for info should test the labels to ensure they're
1065  * getting the correct field.
1066  */
1067 static int nfsd_file_cache_stats_show(struct seq_file *m, void *v)
1068 {
1069         unsigned int i, count = 0, longest = 0;
1070         unsigned long hits = 0;
1071
1072         /*
1073          * No need for spinlocks here since we're not terribly interested in
1074          * accuracy. We do take the nfsd_mutex simply to ensure that we
1075          * don't end up racing with server shutdown
1076          */
1077         mutex_lock(&nfsd_mutex);
1078         if (nfsd_file_hashtbl) {
1079                 for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
1080                         count += nfsd_file_hashtbl[i].nfb_count;
1081                         longest = max(longest, nfsd_file_hashtbl[i].nfb_count);
1082                 }
1083         }
1084         mutex_unlock(&nfsd_mutex);
1085
1086         for_each_possible_cpu(i)
1087                 hits += per_cpu(nfsd_file_cache_hits, i);
1088
1089         seq_printf(m, "total entries: %u\n", count);
1090         seq_printf(m, "longest chain: %u\n", longest);
1091         seq_printf(m, "cache hits:    %lu\n", hits);
1092         return 0;
1093 }
1094
1095 int nfsd_file_cache_stats_open(struct inode *inode, struct file *file)
1096 {
1097         return single_open(file, nfsd_file_cache_stats_show, NULL);
1098 }