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_LRU_RESCAN                 (0)
  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 list_head list;
  49         struct work_struct work;
  50         struct net *net;
  51         spinlock_t lock;
  52         struct list_head freeme;
  53         struct rcu_head rcu;
  54 };
  55
  56 struct workqueue_struct *nfsd_filecache_wq __read_mostly;
  57
  58 static struct kmem_cache                *nfsd_file_slab;
  59 static struct kmem_cache                *nfsd_file_mark_slab;
  60 static struct nfsd_fcache_bucket        *nfsd_file_hashtbl;
  61 static struct list_lru                  nfsd_file_lru;
  62 static long                             nfsd_file_lru_flags;
  63 static struct fsnotify_group            *nfsd_file_fsnotify_group;
  64 static atomic_long_t                    nfsd_filecache_count;
  65 static struct delayed_work              nfsd_filecache_laundrette;
  66 static DEFINE_SPINLOCK(laundrette_lock);
  67 static LIST_HEAD(laundrettes);
  68
  69 static void nfsd_file_gc(void);
  70
  71 static void
  72 nfsd_file_schedule_laundrette(void)
  73 {
  74         long count = atomic_long_read(&nfsd_filecache_count);
  75
  76         if (count == 0 || test_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags))
  77                 return;
  78
  79         queue_delayed_work(system_wq, &nfsd_filecache_laundrette,
  80                         NFSD_LAUNDRETTE_DELAY);
  81 }
  82
  83 static void
  84 nfsd_file_slab_free(struct rcu_head *rcu)
  85 {
  86         struct nfsd_file *nf = container_of(rcu, struct nfsd_file, nf_rcu);
  87
  88         put_cred(nf->nf_cred);
  89         kmem_cache_free(nfsd_file_slab, nf);
  90 }
  91
  92 static void
  93 nfsd_file_mark_free(struct fsnotify_mark *mark)
  94 {
  95         struct nfsd_file_mark *nfm = container_of(mark, struct nfsd_file_mark,
  96                                                   nfm_mark);
  97
  98         kmem_cache_free(nfsd_file_mark_slab, nfm);
  99 }
 100
 101 static struct nfsd_file_mark *
 102 nfsd_file_mark_get(struct nfsd_file_mark *nfm)
 103 {
 104         if (!atomic_inc_not_zero(&nfm->nfm_ref))
 105                 return NULL;
 106         return nfm;
 107 }
 108
 109 static void
 110 nfsd_file_mark_put(struct nfsd_file_mark *nfm)
 111 {
 112         if (atomic_dec_and_test(&nfm->nfm_ref)) {
 113
 114                 fsnotify_destroy_mark(&nfm->nfm_mark, nfsd_file_fsnotify_group);
 115                 fsnotify_put_mark(&nfm->nfm_mark);
 116         }
 117 }
 118
 119 static struct nfsd_file_mark *
 120 nfsd_file_mark_find_or_create(struct nfsd_file *nf)
 121 {
 122         int                     err;
 123         struct fsnotify_mark    *mark;
 124         struct nfsd_file_mark   *nfm = NULL, *new;
 125         struct inode *inode = nf->nf_inode;
 126
 127         do {
 128                 mutex_lock(&nfsd_file_fsnotify_group->mark_mutex);
 129                 mark = fsnotify_find_mark(&inode->i_fsnotify_marks,
 130                                 nfsd_file_fsnotify_group);
 131                 if (mark) {
 132                         nfm = nfsd_file_mark_get(container_of(mark,
 133                                                  struct nfsd_file_mark,
 134                                                  nfm_mark));
 135                         mutex_unlock(&nfsd_file_fsnotify_group->mark_mutex);
 136                         if (nfm) {
 137                                 fsnotify_put_mark(mark);
 138                                 break;
 139                         }
 140                         /* Avoid soft lockup race with nfsd_file_mark_put() */
 141                         fsnotify_destroy_mark(mark, nfsd_file_fsnotify_group);
 142                         fsnotify_put_mark(mark);
 143                 } else
 144                         mutex_unlock(&nfsd_file_fsnotify_group->mark_mutex);
 145
 146                 /* allocate a new nfm */
 147                 new = kmem_cache_alloc(nfsd_file_mark_slab, GFP_KERNEL);
 148                 if (!new)
 149                         return NULL;
 150                 fsnotify_init_mark(&new->nfm_mark, nfsd_file_fsnotify_group);
 151                 new->nfm_mark.mask = FS_ATTRIB|FS_DELETE_SELF;
 152                 atomic_set(&new->nfm_ref, 1);
 153
 154                 err = fsnotify_add_inode_mark(&new->nfm_mark, inode, 0);
 155
 156                 /*
 157                  * If the add was successful, then return the object.
 158                  * Otherwise, we need to put the reference we hold on the
 159                  * nfm_mark. The fsnotify code will take a reference and put
 160                  * it on failure, so we can't just free it directly. It's also
 161                  * not safe to call fsnotify_destroy_mark on it as the
 162                  * mark->group will be NULL. Thus, we can't let the nfm_ref
 163                  * counter drive the destruction at this point.
 164                  */
 165                 if (likely(!err))
 166                         nfm = new;
 167                 else
 168                         fsnotify_put_mark(&new->nfm_mark);
 169         } while (unlikely(err == -EEXIST));
 170
 171         return nfm;
 172 }
 173
 174 static struct nfsd_file *
 175 nfsd_file_alloc(struct inode *inode, unsigned int may, unsigned int hashval,
 176                 struct net *net)
 177 {
 178         struct nfsd_file *nf;
 179
 180         nf = kmem_cache_alloc(nfsd_file_slab, GFP_KERNEL);
 181         if (nf) {
 182                 INIT_HLIST_NODE(&nf->nf_node);
 183                 INIT_LIST_HEAD(&nf->nf_lru);
 184                 nf->nf_file = NULL;
 185                 nf->nf_cred = get_current_cred();
 186                 nf->nf_net = net;
 187                 nf->nf_flags = 0;
 188                 nf->nf_inode = inode;
 189                 nf->nf_hashval = hashval;
 190                 atomic_set(&nf->nf_ref, 1);
 191                 nf->nf_may = may & NFSD_FILE_MAY_MASK;
 192                 if (may & NFSD_MAY_NOT_BREAK_LEASE) {
 193                         if (may & NFSD_MAY_WRITE)
 194                                 __set_bit(NFSD_FILE_BREAK_WRITE, &nf->nf_flags);
 195                         if (may & NFSD_MAY_READ)
 196                                 __set_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags);
 197                 }
 198                 nf->nf_mark = NULL;
 199                 trace_nfsd_file_alloc(nf);
 200         }
 201         return nf;
 202 }
 203
 204 static bool
 205 nfsd_file_free(struct nfsd_file *nf)
 206 {
 207         bool flush = false;
 208
 209         trace_nfsd_file_put_final(nf);
 210         if (nf->nf_mark)
 211                 nfsd_file_mark_put(nf->nf_mark);
 212         if (nf->nf_file) {
 213                 get_file(nf->nf_file);
 214                 filp_close(nf->nf_file, NULL);
 215                 fput(nf->nf_file);
 216                 flush = true;
 217         }
 218         call_rcu(&nf->nf_rcu, nfsd_file_slab_free);
 219         return flush;
 220 }
 221
 222 static bool
 223 nfsd_file_check_writeback(struct nfsd_file *nf)
 224 {
 225         struct file *file = nf->nf_file;
 226         struct address_space *mapping;
 227
 228         if (!file || !(file->f_mode & FMODE_WRITE))
 229                 return false;
 230         mapping = file->f_mapping;
 231         return mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) ||
 232                 mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK);
 233 }
 234
 235 static int
 236 nfsd_file_check_write_error(struct nfsd_file *nf)
 237 {
 238         struct file *file = nf->nf_file;
 239
 240         if (!file || !(file->f_mode & FMODE_WRITE))
 241                 return 0;
 242         return filemap_check_wb_err(file->f_mapping, READ_ONCE(file->f_wb_err));
 243 }
 244
 245 static bool
 246 nfsd_file_in_use(struct nfsd_file *nf)
 247 {
 248         return nfsd_file_check_writeback(nf) ||
 249                         nfsd_file_check_write_error(nf);
 250 }
 251
 252 static void
 253 nfsd_file_do_unhash(struct nfsd_file *nf)
 254 {
 255         lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 256
 257         trace_nfsd_file_unhash(nf);
 258
 259         if (nfsd_file_check_write_error(nf))
 260                 nfsd_reset_boot_verifier(net_generic(nf->nf_net, nfsd_net_id));
 261         --nfsd_file_hashtbl[nf->nf_hashval].nfb_count;
 262         hlist_del_rcu(&nf->nf_node);
 263         atomic_long_dec(&nfsd_filecache_count);
 264 }
 265
 266 static bool
 267 nfsd_file_unhash(struct nfsd_file *nf)
 268 {
 269         if (test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
 270                 nfsd_file_do_unhash(nf);
 271                 if (!list_empty(&nf->nf_lru))
 272                         list_lru_del(&nfsd_file_lru, &nf->nf_lru);
 273                 return true;
 274         }
 275         return false;
 276 }
 277
 278 /*
 279  * Return true if the file was unhashed.
 280  */
 281 static bool
 282 nfsd_file_unhash_and_release_locked(struct nfsd_file *nf, struct list_head *dispose)
 283 {
 284         lockdep_assert_held(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 285
 286         trace_nfsd_file_unhash_and_release_locked(nf);
 287         if (!nfsd_file_unhash(nf))
 288                 return false;
 289         /* keep final reference for nfsd_file_lru_dispose */
 290         if (atomic_add_unless(&nf->nf_ref, -1, 1))
 291                 return true;
 292
 293         list_add(&nf->nf_lru, dispose);
 294         return true;
 295 }
 296
 297 static int
 298 nfsd_file_put_noref(struct nfsd_file *nf)
 299 {
 300         int count;
 301         trace_nfsd_file_put(nf);
 302
 303         count = atomic_dec_return(&nf->nf_ref);
 304         if (!count) {
 305                 WARN_ON(test_bit(NFSD_FILE_HASHED, &nf->nf_flags));
 306                 nfsd_file_free(nf);
 307         }
 308         return count;
 309 }
 310
 311 void
 312 nfsd_file_put(struct nfsd_file *nf)
 313 {
 314         bool is_hashed = test_bit(NFSD_FILE_HASHED, &nf->nf_flags) != 0;
 315         bool unused = !nfsd_file_in_use(nf);
 316
 317         set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
 318         if (nfsd_file_put_noref(nf) == 1 && is_hashed && unused)
 319                 nfsd_file_schedule_laundrette();
 320         if (atomic_long_read(&nfsd_filecache_count) >= NFSD_FILE_LRU_LIMIT)
 321                 nfsd_file_gc();
 322 }
 323
 324 struct nfsd_file *
 325 nfsd_file_get(struct nfsd_file *nf)
 326 {
 327         if (likely(atomic_inc_not_zero(&nf->nf_ref)))
 328                 return nf;
 329         return NULL;
 330 }
 331
 332 static void
 333 nfsd_file_dispose_list(struct list_head *dispose)
 334 {
 335         struct nfsd_file *nf;
 336
 337         while(!list_empty(dispose)) {
 338                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 339                 list_del(&nf->nf_lru);
 340                 nfsd_file_put_noref(nf);
 341         }
 342 }
 343
 344 static void
 345 nfsd_file_dispose_list_sync(struct list_head *dispose)
 346 {
 347         bool flush = false;
 348         struct nfsd_file *nf;
 349
 350         while(!list_empty(dispose)) {
 351                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 352                 list_del(&nf->nf_lru);
 353                 if (!atomic_dec_and_test(&nf->nf_ref))
 354                         continue;
 355                 if (nfsd_file_free(nf))
 356                         flush = true;
 357         }
 358         if (flush)
 359                 flush_delayed_fput();
 360 }
 361
 362 static void
 363 nfsd_file_list_remove_disposal(struct list_head *dst,
 364                 struct nfsd_fcache_disposal *l)
 365 {
 366         spin_lock(&l->lock);
 367         list_splice_init(&l->freeme, dst);
 368         spin_unlock(&l->lock);
 369 }
 370
 371 static void
 372 nfsd_file_list_add_disposal(struct list_head *files, struct net *net)
 373 {
 374         struct nfsd_fcache_disposal *l;
 375
 376         rcu_read_lock();
 377         list_for_each_entry_rcu(l, &laundrettes, list) {
 378                 if (l->net == net) {
 379                         spin_lock(&l->lock);
 380                         list_splice_tail_init(files, &l->freeme);
 381                         spin_unlock(&l->lock);
 382                         queue_work(nfsd_filecache_wq, &l->work);
 383                         break;
 384                 }
 385         }
 386         rcu_read_unlock();
 387 }
 388
 389 static void
 390 nfsd_file_list_add_pernet(struct list_head *dst, struct list_head *src,
 391                 struct net *net)
 392 {
 393         struct nfsd_file *nf, *tmp;
 394
 395         list_for_each_entry_safe(nf, tmp, src, nf_lru) {
 396                 if (nf->nf_net == net)
 397                         list_move_tail(&nf->nf_lru, dst);
 398         }
 399 }
 400
 401 static void
 402 nfsd_file_dispose_list_delayed(struct list_head *dispose)
 403 {
 404         LIST_HEAD(list);
 405         struct nfsd_file *nf;
 406
 407         while(!list_empty(dispose)) {
 408                 nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
 409                 nfsd_file_list_add_pernet(&list, dispose, nf->nf_net);
 410                 nfsd_file_list_add_disposal(&list, nf->nf_net);
 411         }
 412 }
 413
 414 /*
 415  * Note this can deadlock with nfsd_file_cache_purge.
 416  */
 417 static enum lru_status
 418 nfsd_file_lru_cb(struct list_head *item, struct list_lru_one *lru,
 419                  spinlock_t *lock, void *arg)
 420         __releases(lock)
 421         __acquires(lock)
 422 {
 423         struct list_head *head = arg;
 424         struct nfsd_file *nf = list_entry(item, struct nfsd_file, nf_lru);
 425
 426         /*
 427          * Do a lockless refcount check. The hashtable holds one reference, so
 428          * we look to see if anything else has a reference, or if any have
 429          * been put since the shrinker last ran. Those don't get unhashed and
 430          * released.
 431          *
 432          * Note that in the put path, we set the flag and then decrement the
 433          * counter. Here we check the counter and then test and clear the flag.
 434          * That order is deliberate to ensure that we can do this locklessly.
 435          */
 436         if (atomic_read(&nf->nf_ref) > 1)
 437                 goto out_skip;
 438
 439         /*
 440          * Don't throw out files that are still undergoing I/O or
 441          * that have uncleared errors pending.
 442          */
 443         if (nfsd_file_check_writeback(nf))
 444                 goto out_skip;
 445
 446         if (test_and_clear_bit(NFSD_FILE_REFERENCED, &nf->nf_flags))
 447                 goto out_rescan;
 448
 449         if (!test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags))
 450                 goto out_skip;
 451
 452         list_lru_isolate_move(lru, &nf->nf_lru, head);
 453         return LRU_REMOVED;
 454 out_rescan:
 455         set_bit(NFSD_FILE_LRU_RESCAN, &nfsd_file_lru_flags);
 456 out_skip:
 457         return LRU_SKIP;
 458 }
 459
 460 static unsigned long
 461 nfsd_file_lru_walk_list(struct shrink_control *sc)
 462 {
 463         LIST_HEAD(head);
 464         struct nfsd_file *nf;
 465         unsigned long ret;
 466
 467         if (sc)
 468                 ret = list_lru_shrink_walk(&nfsd_file_lru, sc,
 469                                 nfsd_file_lru_cb, &head);
 470         else
 471                 ret = list_lru_walk(&nfsd_file_lru,
 472                                 nfsd_file_lru_cb,
 473                                 &head, LONG_MAX);
 474         list_for_each_entry(nf, &head, nf_lru) {
 475                 spin_lock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 476                 nfsd_file_do_unhash(nf);
 477                 spin_unlock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
 478         }
 479         nfsd_file_dispose_list_delayed(&head);
 480         return ret;
 481 }
 482
 483 static void
 484 nfsd_file_gc(void)
 485 {
 486         nfsd_file_lru_walk_list(NULL);
 487 }
 488
 489 static void
 490 nfsd_file_gc_worker(struct work_struct *work)
 491 {
 492         nfsd_file_gc();
 493         nfsd_file_schedule_laundrette();
 494 }
 495
 496 static unsigned long
 497 nfsd_file_lru_count(struct shrinker *s, struct shrink_control *sc)
 498 {
 499         return list_lru_count(&nfsd_file_lru);
 500 }
 501
 502 static unsigned long
 503 nfsd_file_lru_scan(struct shrinker *s, struct shrink_control *sc)
 504 {
 505         return nfsd_file_lru_walk_list(sc);
 506 }
 507
 508 static struct shrinker  nfsd_file_shrinker = {
 509         .scan_objects = nfsd_file_lru_scan,
 510         .count_objects = nfsd_file_lru_count,
 511         .seeks = 1,
 512 };
 513
 514 static void
 515 __nfsd_file_close_inode(struct inode *inode, unsigned int hashval,
 516                         struct list_head *dispose)
 517 {
 518         struct nfsd_file        *nf;
 519         struct hlist_node       *tmp;
 520
 521         spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
 522         hlist_for_each_entry_safe(nf, tmp, &nfsd_file_hashtbl[hashval].nfb_head, nf_node) {
 523                 if (inode == nf->nf_inode)
 524                         nfsd_file_unhash_and_release_locked(nf, dispose);
 525         }
 526         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
 527 }
 528
 529 /**
 530  * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
 531  * @inode: inode of the file to attempt to remove
 532  *
 533  * Walk the whole hash bucket, looking for any files that correspond to "inode".
 534  * If any do, then unhash them and put the hashtable reference to them and
 535  * destroy any that had their last reference put. Also ensure that any of the
 536  * fputs also have their final __fput done as well.
 537  */
 538 void
 539 nfsd_file_close_inode_sync(struct inode *inode)
 540 {
 541         unsigned int            hashval = (unsigned int)hash_long(inode->i_ino,
 542                                                 NFSD_FILE_HASH_BITS);
 543         LIST_HEAD(dispose);
 544
 545         __nfsd_file_close_inode(inode, hashval, &dispose);
 546         trace_nfsd_file_close_inode_sync(inode, hashval, !list_empty(&dispose));
 547         nfsd_file_dispose_list_sync(&dispose);
 548 }
 549
 550 /**
 551  * nfsd_file_close_inode_sync - attempt to forcibly close a nfsd_file
 552  * @inode: inode of the file to attempt to remove
 553  *
 554  * Walk the whole hash bucket, looking for any files that correspond to "inode".
 555  * If any do, then unhash them and put the hashtable reference to them and
 556  * destroy any that had their last reference put.
 557  */
 558 static void
 559 nfsd_file_close_inode(struct inode *inode)
 560 {
 561         unsigned int            hashval = (unsigned int)hash_long(inode->i_ino,
 562                                                 NFSD_FILE_HASH_BITS);
 563         LIST_HEAD(dispose);
 564
 565         __nfsd_file_close_inode(inode, hashval, &dispose);
 566         trace_nfsd_file_close_inode(inode, hashval, !list_empty(&dispose));
 567         nfsd_file_dispose_list_delayed(&dispose);
 568 }
 569
 570 /**
 571  * nfsd_file_delayed_close - close unused nfsd_files
 572  * @work: dummy
 573  *
 574  * Walk the LRU list and close any entries that have not been used since
 575  * the last scan.
 576  *
 577  * Note this can deadlock with nfsd_file_cache_purge.
 578  */
 579 static void
 580 nfsd_file_delayed_close(struct work_struct *work)
 581 {
 582         LIST_HEAD(head);
 583         struct nfsd_fcache_disposal *l = container_of(work,
 584                         struct nfsd_fcache_disposal, work);
 585
 586         nfsd_file_list_remove_disposal(&head, l);
 587         nfsd_file_dispose_list(&head);
 588 }
 589
 590 static int
 591 nfsd_file_lease_notifier_call(struct notifier_block *nb, unsigned long arg,
 592                             void *data)
 593 {
 594         struct file_lock *fl = data;
 595
 596         /* Only close files for F_SETLEASE leases */
 597         if (fl->fl_flags & FL_LEASE)
 598                 nfsd_file_close_inode_sync(file_inode(fl->fl_file));
 599         return 0;
 600 }
 601
 602 static struct notifier_block nfsd_file_lease_notifier = {
 603         .notifier_call = nfsd_file_lease_notifier_call,
 604 };
 605
 606 static int
 607 nfsd_file_fsnotify_handle_event(struct fsnotify_group *group,
 608                                 struct inode *inode,
 609                                 u32 mask, const void *data, int data_type,
 610                                 const struct qstr *file_name, u32 cookie,
 611                                 struct fsnotify_iter_info *iter_info)
 612 {
 613         trace_nfsd_file_fsnotify_handle_event(inode, mask);
 614
 615         /* Should be no marks on non-regular files */
 616         if (!S_ISREG(inode->i_mode)) {
 617                 WARN_ON_ONCE(1);
 618                 return 0;
 619         }
 620
 621         /* don't close files if this was not the last link */
 622         if (mask & FS_ATTRIB) {
 623                 if (inode->i_nlink)
 624                         return 0;
 625         }
 626
 627         nfsd_file_close_inode(inode);
 628         return 0;
 629 }
 630
 631
 632 static const struct fsnotify_ops nfsd_file_fsnotify_ops = {
 633         .handle_event = nfsd_file_fsnotify_handle_event,
 634         .free_mark = nfsd_file_mark_free,
 635 };
 636
 637 int
 638 nfsd_file_cache_init(void)
 639 {
 640         int             ret = -ENOMEM;
 641         unsigned int    i;
 642
 643         clear_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
 644
 645         if (nfsd_file_hashtbl)
 646                 return 0;
 647
 648         nfsd_filecache_wq = alloc_workqueue("nfsd_filecache", 0, 0);
 649         if (!nfsd_filecache_wq)
 650                 goto out;
 651
 652         nfsd_file_hashtbl = kcalloc(NFSD_FILE_HASH_SIZE,
 653                                 sizeof(*nfsd_file_hashtbl), GFP_KERNEL);
 654         if (!nfsd_file_hashtbl) {
 655                 pr_err("nfsd: unable to allocate nfsd_file_hashtbl\n");
 656                 goto out_err;
 657         }
 658
 659         nfsd_file_slab = kmem_cache_create("nfsd_file",
 660                                 sizeof(struct nfsd_file), 0, 0, NULL);
 661         if (!nfsd_file_slab) {
 662                 pr_err("nfsd: unable to create nfsd_file_slab\n");
 663                 goto out_err;
 664         }
 665
 666         nfsd_file_mark_slab = kmem_cache_create("nfsd_file_mark",
 667                                         sizeof(struct nfsd_file_mark), 0, 0, NULL);
 668         if (!nfsd_file_mark_slab) {
 669                 pr_err("nfsd: unable to create nfsd_file_mark_slab\n");
 670                 goto out_err;
 671         }
 672
 673
 674         ret = list_lru_init(&nfsd_file_lru);
 675         if (ret) {
 676                 pr_err("nfsd: failed to init nfsd_file_lru: %d\n", ret);
 677                 goto out_err;
 678         }
 679
 680         ret = register_shrinker(&nfsd_file_shrinker);
 681         if (ret) {
 682                 pr_err("nfsd: failed to register nfsd_file_shrinker: %d\n", ret);
 683                 goto out_lru;
 684         }
 685
 686         ret = lease_register_notifier(&nfsd_file_lease_notifier);
 687         if (ret) {
 688                 pr_err("nfsd: unable to register lease notifier: %d\n", ret);
 689                 goto out_shrinker;
 690         }
 691
 692         nfsd_file_fsnotify_group = fsnotify_alloc_group(&nfsd_file_fsnotify_ops);
 693         if (IS_ERR(nfsd_file_fsnotify_group)) {
 694                 pr_err("nfsd: unable to create fsnotify group: %ld\n",
 695                         PTR_ERR(nfsd_file_fsnotify_group));
 696                 nfsd_file_fsnotify_group = NULL;
 697                 goto out_notifier;
 698         }
 699
 700         for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
 701                 INIT_HLIST_HEAD(&nfsd_file_hashtbl[i].nfb_head);
 702                 spin_lock_init(&nfsd_file_hashtbl[i].nfb_lock);
 703         }
 704
 705         INIT_DELAYED_WORK(&nfsd_filecache_laundrette, nfsd_file_gc_worker);
 706 out:
 707         return ret;
 708 out_notifier:
 709         lease_unregister_notifier(&nfsd_file_lease_notifier);
 710 out_shrinker:
 711         unregister_shrinker(&nfsd_file_shrinker);
 712 out_lru:
 713         list_lru_destroy(&nfsd_file_lru);
 714 out_err:
 715         kmem_cache_destroy(nfsd_file_slab);
 716         nfsd_file_slab = NULL;
 717         kmem_cache_destroy(nfsd_file_mark_slab);
 718         nfsd_file_mark_slab = NULL;
 719         kfree(nfsd_file_hashtbl);
 720         nfsd_file_hashtbl = NULL;
 721         destroy_workqueue(nfsd_filecache_wq);
 722         nfsd_filecache_wq = NULL;
 723         goto out;
 724 }
 725
 726 /*
 727  * Note this can deadlock with nfsd_file_lru_cb.
 728  */
 729 void
 730 nfsd_file_cache_purge(struct net *net)
 731 {
 732         unsigned int            i;
 733         struct nfsd_file        *nf;
 734         struct hlist_node       *next;
 735         LIST_HEAD(dispose);
 736         bool del;
 737
 738         if (!nfsd_file_hashtbl)
 739                 return;
 740
 741         for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
 742                 struct nfsd_fcache_bucket *nfb = &nfsd_file_hashtbl[i];
 743
 744                 spin_lock(&nfb->nfb_lock);
 745                 hlist_for_each_entry_safe(nf, next, &nfb->nfb_head, nf_node) {
 746                         if (net && nf->nf_net != net)
 747                                 continue;
 748                         del = nfsd_file_unhash_and_release_locked(nf, &dispose);
 749
 750                         /*
 751                          * Deadlock detected! Something marked this entry as
 752                          * unhased, but hasn't removed it from the hash list.
 753                          */
 754                         WARN_ON_ONCE(!del);
 755                 }
 756                 spin_unlock(&nfb->nfb_lock);
 757                 nfsd_file_dispose_list(&dispose);
 758         }
 759 }
 760
 761 static struct nfsd_fcache_disposal *
 762 nfsd_alloc_fcache_disposal(struct net *net)
 763 {
 764         struct nfsd_fcache_disposal *l;
 765
 766         l = kmalloc(sizeof(*l), GFP_KERNEL);
 767         if (!l)
 768                 return NULL;
 769         INIT_WORK(&l->work, nfsd_file_delayed_close);
 770         l->net = net;
 771         spin_lock_init(&l->lock);
 772         INIT_LIST_HEAD(&l->freeme);
 773         return l;
 774 }
 775
 776 static void
 777 nfsd_free_fcache_disposal(struct nfsd_fcache_disposal *l)
 778 {
 779         rcu_assign_pointer(l->net, NULL);
 780         cancel_work_sync(&l->work);
 781         nfsd_file_dispose_list(&l->freeme);
 782         kfree_rcu(l, rcu);
 783 }
 784
 785 static void
 786 nfsd_add_fcache_disposal(struct nfsd_fcache_disposal *l)
 787 {
 788         spin_lock(&laundrette_lock);
 789         list_add_tail_rcu(&l->list, &laundrettes);
 790         spin_unlock(&laundrette_lock);
 791 }
 792
 793 static void
 794 nfsd_del_fcache_disposal(struct nfsd_fcache_disposal *l)
 795 {
 796         spin_lock(&laundrette_lock);
 797         list_del_rcu(&l->list);
 798         spin_unlock(&laundrette_lock);
 799 }
 800
 801 static int
 802 nfsd_alloc_fcache_disposal_net(struct net *net)
 803 {
 804         struct nfsd_fcache_disposal *l;
 805
 806         l = nfsd_alloc_fcache_disposal(net);
 807         if (!l)
 808                 return -ENOMEM;
 809         nfsd_add_fcache_disposal(l);
 810         return 0;
 811 }
 812
 813 static void
 814 nfsd_free_fcache_disposal_net(struct net *net)
 815 {
 816         struct nfsd_fcache_disposal *l;
 817
 818         rcu_read_lock();
 819         list_for_each_entry_rcu(l, &laundrettes, list) {
 820                 if (l->net != net)
 821                         continue;
 822                 nfsd_del_fcache_disposal(l);
 823                 rcu_read_unlock();
 824                 nfsd_free_fcache_disposal(l);
 825                 return;
 826         }
 827         rcu_read_unlock();
 828 }
 829
 830 int
 831 nfsd_file_cache_start_net(struct net *net)
 832 {
 833         return nfsd_alloc_fcache_disposal_net(net);
 834 }
 835
 836 void
 837 nfsd_file_cache_shutdown_net(struct net *net)
 838 {
 839         nfsd_file_cache_purge(net);
 840         nfsd_free_fcache_disposal_net(net);
 841 }
 842
 843 void
 844 nfsd_file_cache_shutdown(void)
 845 {
 846         LIST_HEAD(dispose);
 847
 848         set_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags);
 849
 850         lease_unregister_notifier(&nfsd_file_lease_notifier);
 851         unregister_shrinker(&nfsd_file_shrinker);
 852         /*
 853          * make sure all callers of nfsd_file_lru_cb are done before
 854          * calling nfsd_file_cache_purge
 855          */
 856         cancel_delayed_work_sync(&nfsd_filecache_laundrette);
 857         nfsd_file_cache_purge(NULL);
 858         list_lru_destroy(&nfsd_file_lru);
 859         rcu_barrier();
 860         fsnotify_put_group(nfsd_file_fsnotify_group);
 861         nfsd_file_fsnotify_group = NULL;
 862         kmem_cache_destroy(nfsd_file_slab);
 863         nfsd_file_slab = NULL;
 864         fsnotify_wait_marks_destroyed();
 865         kmem_cache_destroy(nfsd_file_mark_slab);
 866         nfsd_file_mark_slab = NULL;
 867         kfree(nfsd_file_hashtbl);
 868         nfsd_file_hashtbl = NULL;
 869         destroy_workqueue(nfsd_filecache_wq);
 870         nfsd_filecache_wq = NULL;
 871 }
 872
 873 static bool
 874 nfsd_match_cred(const struct cred *c1, const struct cred *c2)
 875 {
 876         int i;
 877
 878         if (!uid_eq(c1->fsuid, c2->fsuid))
 879                 return false;
 880         if (!gid_eq(c1->fsgid, c2->fsgid))
 881                 return false;
 882         if (c1->group_info == NULL || c2->group_info == NULL)
 883                 return c1->group_info == c2->group_info;
 884         if (c1->group_info->ngroups != c2->group_info->ngroups)
 885                 return false;
 886         for (i = 0; i < c1->group_info->ngroups; i++) {
 887                 if (!gid_eq(c1->group_info->gid[i], c2->group_info->gid[i]))
 888                         return false;
 889         }
 890         return true;
 891 }
 892
 893 static struct nfsd_file *
 894 nfsd_file_find_locked(struct inode *inode, unsigned int may_flags,
 895                         unsigned int hashval, struct net *net)
 896 {
 897         struct nfsd_file *nf;
 898         unsigned char need = may_flags & NFSD_FILE_MAY_MASK;
 899
 900         hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
 901                                  nf_node) {
 902                 if ((need & nf->nf_may) != need)
 903                         continue;
 904                 if (nf->nf_inode != inode)
 905                         continue;
 906                 if (nf->nf_net != net)
 907                         continue;
 908                 if (!nfsd_match_cred(nf->nf_cred, current_cred()))
 909                         continue;
 910                 if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags))
 911                         continue;
 912                 if (nfsd_file_get(nf) != NULL)
 913                         return nf;
 914         }
 915         return NULL;
 916 }
 917
 918 /**
 919  * nfsd_file_is_cached - are there any cached open files for this fh?
 920  * @inode: inode of the file to check
 921  *
 922  * Scan the hashtable for open files that match this fh. Returns true if there
 923  * are any, and false if not.
 924  */
 925 bool
 926 nfsd_file_is_cached(struct inode *inode)
 927 {
 928         bool                    ret = false;
 929         struct nfsd_file        *nf;
 930         unsigned int            hashval;
 931
 932         hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
 933
 934         rcu_read_lock();
 935         hlist_for_each_entry_rcu(nf, &nfsd_file_hashtbl[hashval].nfb_head,
 936                                  nf_node) {
 937                 if (inode == nf->nf_inode) {
 938                         ret = true;
 939                         break;
 940                 }
 941         }
 942         rcu_read_unlock();
 943         trace_nfsd_file_is_cached(inode, hashval, (int)ret);
 944         return ret;
 945 }
 946
 947 __be32
 948 nfsd_file_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 949                   unsigned int may_flags, struct nfsd_file **pnf)
 950 {
 951         __be32  status;
 952         struct net *net = SVC_NET(rqstp);
 953         struct nfsd_file *nf, *new;
 954         struct inode *inode;
 955         unsigned int hashval;
 956         bool retry = true;
 957
 958         /* FIXME: skip this if fh_dentry is already set? */
 959         status = fh_verify(rqstp, fhp, S_IFREG,
 960                                 may_flags|NFSD_MAY_OWNER_OVERRIDE);
 961         if (status != nfs_ok)
 962                 return status;
 963
 964         inode = d_inode(fhp->fh_dentry);
 965         hashval = (unsigned int)hash_long(inode->i_ino, NFSD_FILE_HASH_BITS);
 966 retry:
 967         rcu_read_lock();
 968         nf = nfsd_file_find_locked(inode, may_flags, hashval, net);
 969         rcu_read_unlock();
 970         if (nf)
 971                 goto wait_for_construction;
 972
 973         new = nfsd_file_alloc(inode, may_flags, hashval, net);
 974         if (!new) {
 975                 trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags,
 976                                         NULL, nfserr_jukebox);
 977                 return nfserr_jukebox;
 978         }
 979
 980         spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
 981         nf = nfsd_file_find_locked(inode, may_flags, hashval, net);
 982         if (nf == NULL)
 983                 goto open_file;
 984         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
 985         nfsd_file_slab_free(&new->nf_rcu);
 986
 987 wait_for_construction:
 988         wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
 989
 990         /* Did construction of this file fail? */
 991         if (!test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
 992                 if (!retry) {
 993                         status = nfserr_jukebox;
 994                         goto out;
 995                 }
 996                 retry = false;
 997                 nfsd_file_put_noref(nf);
 998                 goto retry;
 999         }
1000
1001         this_cpu_inc(nfsd_file_cache_hits);
1002
1003         if (!(may_flags & NFSD_MAY_NOT_BREAK_LEASE)) {
1004                 bool write = (may_flags & NFSD_MAY_WRITE);
1005
1006                 if (test_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags) ||
1007                     (test_bit(NFSD_FILE_BREAK_WRITE, &nf->nf_flags) && write)) {
1008                         status = nfserrno(nfsd_open_break_lease(
1009                                         file_inode(nf->nf_file), may_flags));
1010                         if (status == nfs_ok) {
1011                                 clear_bit(NFSD_FILE_BREAK_READ, &nf->nf_flags);
1012                                 if (write)
1013                                         clear_bit(NFSD_FILE_BREAK_WRITE,
1014                                                   &nf->nf_flags);
1015                         }
1016                 }
1017         }
1018 out:
1019         if (status == nfs_ok) {
1020                 *pnf = nf;
1021         } else {
1022                 nfsd_file_put(nf);
1023                 nf = NULL;
1024         }
1025
1026         trace_nfsd_file_acquire(rqstp, hashval, inode, may_flags, nf, status);
1027         return status;
1028 open_file:
1029         nf = new;
1030         /* Take reference for the hashtable */
1031         atomic_inc(&nf->nf_ref);
1032         __set_bit(NFSD_FILE_HASHED, &nf->nf_flags);
1033         __set_bit(NFSD_FILE_PENDING, &nf->nf_flags);
1034         list_lru_add(&nfsd_file_lru, &nf->nf_lru);
1035         hlist_add_head_rcu(&nf->nf_node, &nfsd_file_hashtbl[hashval].nfb_head);
1036         ++nfsd_file_hashtbl[hashval].nfb_count;
1037         nfsd_file_hashtbl[hashval].nfb_maxcount = max(nfsd_file_hashtbl[hashval].nfb_maxcount,
1038                         nfsd_file_hashtbl[hashval].nfb_count);
1039         spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
1040         if (atomic_long_inc_return(&nfsd_filecache_count) >= NFSD_FILE_LRU_THRESHOLD)
1041                 nfsd_file_gc();
1042
1043         nf->nf_mark = nfsd_file_mark_find_or_create(nf);
1044         if (nf->nf_mark)
1045                 status = nfsd_open_verified(rqstp, fhp, S_IFREG,
1046                                 may_flags, &nf->nf_file);
1047         else
1048                 status = nfserr_jukebox;
1049         /*
1050          * If construction failed, or we raced with a call to unlink()
1051          * then unhash.
1052          */
1053         if (status != nfs_ok || inode->i_nlink == 0) {
1054                 bool do_free;
1055                 spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
1056                 do_free = nfsd_file_unhash(nf);
1057                 spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
1058                 if (do_free)
1059                         nfsd_file_put_noref(nf);
1060         }
1061         clear_bit_unlock(NFSD_FILE_PENDING, &nf->nf_flags);
1062         smp_mb__after_atomic();
1063         wake_up_bit(&nf->nf_flags, NFSD_FILE_PENDING);
1064         goto out;
1065 }
1066
1067 /*
1068  * Note that fields may be added, removed or reordered in the future. Programs
1069  * scraping this file for info should test the labels to ensure they're
1070  * getting the correct field.
1071  */
1072 static int nfsd_file_cache_stats_show(struct seq_file *m, void *v)
1073 {
1074         unsigned int i, count = 0, longest = 0;
1075         unsigned long hits = 0;
1076
1077         /*
1078          * No need for spinlocks here since we're not terribly interested in
1079          * accuracy. We do take the nfsd_mutex simply to ensure that we
1080          * don't end up racing with server shutdown
1081          */
1082         mutex_lock(&nfsd_mutex);
1083         if (nfsd_file_hashtbl) {
1084                 for (i = 0; i < NFSD_FILE_HASH_SIZE; i++) {
1085                         count += nfsd_file_hashtbl[i].nfb_count;
1086                         longest = max(longest, nfsd_file_hashtbl[i].nfb_count);
1087                 }
1088         }
1089         mutex_unlock(&nfsd_mutex);
1090
1091         for_each_possible_cpu(i)
1092                 hits += per_cpu(nfsd_file_cache_hits, i);
1093
1094         seq_printf(m, "total entries: %u\n", count);
1095         seq_printf(m, "longest chain: %u\n", longest);
1096         seq_printf(m, "cache hits:    %lu\n", hits);
1097         return 0;
1098 }
1099
1100 int nfsd_file_cache_stats_open(struct inode *inode, struct file *file)
1101 {
1102         return single_open(file, nfsd_file_cache_stats_show, NULL);
1103 }