fs/ceph/inode.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/module.h>
   5 #include <linux/fs.h>
   6 #include <linux/slab.h>
   7 #include <linux/string.h>
   8 #include <linux/uaccess.h>
   9 #include <linux/kernel.h>
  10 #include <linux/writeback.h>
  11 #include <linux/vmalloc.h>
  12 #include <linux/xattr.h>
  13 #include <linux/posix_acl.h>
  14 #include <linux/random.h>
  15 #include <linux/sort.h>
  16
  17 #include "super.h"
  18 #include "mds_client.h"
  19 #include "cache.h"
  20 #include <linux/ceph/decode.h>
  21
  22 /*
  23  * Ceph inode operations
  24  *
  25  * Implement basic inode helpers (get, alloc) and inode ops (getattr,
  26  * setattr, etc.), xattr helpers, and helpers for assimilating
  27  * metadata returned by the MDS into our cache.
  28  *
  29  * Also define helpers for doing asynchronous writeback, invalidation,
  30  * and truncation for the benefit of those who can't afford to block
  31  * (typically because they are in the message handler path).
  32  */
  33
  34 static const struct inode_operations ceph_symlink_iops;
  35
  36 static void ceph_invalidate_work(struct work_struct *work);
  37 static void ceph_writeback_work(struct work_struct *work);
  38 static void ceph_vmtruncate_work(struct work_struct *work);
  39
  40 /*
  41  * find or create an inode, given the ceph ino number
  42  */
  43 static int ceph_set_ino_cb(struct inode *inode, void *data)
  44 {
  45         ceph_inode(inode)->i_vino = *(struct ceph_vino *)data;
  46         inode->i_ino = ceph_vino_to_ino(*(struct ceph_vino *)data);
  47         return 0;
  48 }
  49
  50 struct inode *ceph_get_inode(struct super_block *sb, struct ceph_vino vino)
  51 {
  52         struct inode *inode;
  53         ino_t t = ceph_vino_to_ino(vino);
  54
  55         inode = iget5_locked(sb, t, ceph_ino_compare, ceph_set_ino_cb, &vino);
  56         if (!inode)
  57                 return ERR_PTR(-ENOMEM);
  58         if (inode->i_state & I_NEW) {
  59                 dout("get_inode created new inode %p %llx.%llx ino %llx\n",
  60                      inode, ceph_vinop(inode), (u64)inode->i_ino);
  61                 unlock_new_inode(inode);
  62         }
  63
  64         dout("get_inode on %lu=%llx.%llx got %p\n", inode->i_ino, vino.ino,
  65              vino.snap, inode);
  66         return inode;
  67 }
  68
  69 /*
  70  * get/constuct snapdir inode for a given directory
  71  */
  72 struct inode *ceph_get_snapdir(struct inode *parent)
  73 {
  74         struct ceph_vino vino = {
  75                 .ino = ceph_ino(parent),
  76                 .snap = CEPH_SNAPDIR,
  77         };
  78         struct inode *inode = ceph_get_inode(parent->i_sb, vino);
  79         struct ceph_inode_info *ci = ceph_inode(inode);
  80
  81         BUG_ON(!S_ISDIR(parent->i_mode));
  82         if (IS_ERR(inode))
  83                 return inode;
  84         inode->i_mode = parent->i_mode;
  85         inode->i_uid = parent->i_uid;
  86         inode->i_gid = parent->i_gid;
  87         inode->i_op = &ceph_snapdir_iops;
  88         inode->i_fop = &ceph_snapdir_fops;
  89         ci->i_snap_caps = CEPH_CAP_PIN; /* so we can open */
  90         ci->i_rbytes = 0;
  91         return inode;
  92 }
  93
  94 const struct inode_operations ceph_file_iops = {
  95         .permission = ceph_permission,
  96         .setattr = ceph_setattr,
  97         .getattr = ceph_getattr,
  98         .listxattr = ceph_listxattr,
  99         .get_acl = ceph_get_acl,
 100         .set_acl = ceph_set_acl,
 101 };
 102
 103
 104 /*
 105  * We use a 'frag tree' to keep track of the MDS's directory fragments
 106  * for a given inode (usually there is just a single fragment).  We
 107  * need to know when a child frag is delegated to a new MDS, or when
 108  * it is flagged as replicated, so we can direct our requests
 109  * accordingly.
 110  */
 111
 112 /*
 113  * find/create a frag in the tree
 114  */
 115 static struct ceph_inode_frag *__get_or_create_frag(struct ceph_inode_info *ci,
 116                                                     u32 f)
 117 {
 118         struct rb_node **p;
 119         struct rb_node *parent = NULL;
 120         struct ceph_inode_frag *frag;
 121         int c;
 122
 123         p = &ci->i_fragtree.rb_node;
 124         while (*p) {
 125                 parent = *p;
 126                 frag = rb_entry(parent, struct ceph_inode_frag, node);
 127                 c = ceph_frag_compare(f, frag->frag);
 128                 if (c < 0)
 129                         p = &(*p)->rb_left;
 130                 else if (c > 0)
 131                         p = &(*p)->rb_right;
 132                 else
 133                         return frag;
 134         }
 135
 136         frag = kmalloc(sizeof(*frag), GFP_NOFS);
 137         if (!frag)
 138                 return ERR_PTR(-ENOMEM);
 139
 140         frag->frag = f;
 141         frag->split_by = 0;
 142         frag->mds = -1;
 143         frag->ndist = 0;
 144
 145         rb_link_node(&frag->node, parent, p);
 146         rb_insert_color(&frag->node, &ci->i_fragtree);
 147
 148         dout("get_or_create_frag added %llx.%llx frag %x\n",
 149              ceph_vinop(&ci->vfs_inode), f);
 150         return frag;
 151 }
 152
 153 /*
 154  * find a specific frag @f
 155  */
 156 struct ceph_inode_frag *__ceph_find_frag(struct ceph_inode_info *ci, u32 f)
 157 {
 158         struct rb_node *n = ci->i_fragtree.rb_node;
 159
 160         while (n) {
 161                 struct ceph_inode_frag *frag =
 162                         rb_entry(n, struct ceph_inode_frag, node);
 163                 int c = ceph_frag_compare(f, frag->frag);
 164                 if (c < 0)
 165                         n = n->rb_left;
 166                 else if (c > 0)
 167                         n = n->rb_right;
 168                 else
 169                         return frag;
 170         }
 171         return NULL;
 172 }
 173
 174 /*
 175  * Choose frag containing the given value @v.  If @pfrag is
 176  * specified, copy the frag delegation info to the caller if
 177  * it is present.
 178  */
 179 static u32 __ceph_choose_frag(struct ceph_inode_info *ci, u32 v,
 180                               struct ceph_inode_frag *pfrag, int *found)
 181 {
 182         u32 t = ceph_frag_make(0, 0);
 183         struct ceph_inode_frag *frag;
 184         unsigned nway, i;
 185         u32 n;
 186
 187         if (found)
 188                 *found = 0;
 189
 190         while (1) {
 191                 WARN_ON(!ceph_frag_contains_value(t, v));
 192                 frag = __ceph_find_frag(ci, t);
 193                 if (!frag)
 194                         break; /* t is a leaf */
 195                 if (frag->split_by == 0) {
 196                         if (pfrag)
 197                                 memcpy(pfrag, frag, sizeof(*pfrag));
 198                         if (found)
 199                                 *found = 1;
 200                         break;
 201                 }
 202
 203                 /* choose child */
 204                 nway = 1 << frag->split_by;
 205                 dout("choose_frag(%x) %x splits by %d (%d ways)\n", v, t,
 206                      frag->split_by, nway);
 207                 for (i = 0; i < nway; i++) {
 208                         n = ceph_frag_make_child(t, frag->split_by, i);
 209                         if (ceph_frag_contains_value(n, v)) {
 210                                 t = n;
 211                                 break;
 212                         }
 213                 }
 214                 BUG_ON(i == nway);
 215         }
 216         dout("choose_frag(%x) = %x\n", v, t);
 217
 218         return t;
 219 }
 220
 221 u32 ceph_choose_frag(struct ceph_inode_info *ci, u32 v,
 222                      struct ceph_inode_frag *pfrag, int *found)
 223 {
 224         u32 ret;
 225         mutex_lock(&ci->i_fragtree_mutex);
 226         ret = __ceph_choose_frag(ci, v, pfrag, found);
 227         mutex_unlock(&ci->i_fragtree_mutex);
 228         return ret;
 229 }
 230
 231 /*
 232  * Process dirfrag (delegation) info from the mds.  Include leaf
 233  * fragment in tree ONLY if ndist > 0.  Otherwise, only
 234  * branches/splits are included in i_fragtree)
 235  */
 236 static int ceph_fill_dirfrag(struct inode *inode,
 237                              struct ceph_mds_reply_dirfrag *dirinfo)
 238 {
 239         struct ceph_inode_info *ci = ceph_inode(inode);
 240         struct ceph_inode_frag *frag;
 241         u32 id = le32_to_cpu(dirinfo->frag);
 242         int mds = le32_to_cpu(dirinfo->auth);
 243         int ndist = le32_to_cpu(dirinfo->ndist);
 244         int diri_auth = -1;
 245         int i;
 246         int err = 0;
 247
 248         spin_lock(&ci->i_ceph_lock);
 249         if (ci->i_auth_cap)
 250                 diri_auth = ci->i_auth_cap->mds;
 251         spin_unlock(&ci->i_ceph_lock);
 252
 253         if (mds == -1) /* CDIR_AUTH_PARENT */
 254                 mds = diri_auth;
 255
 256         mutex_lock(&ci->i_fragtree_mutex);
 257         if (ndist == 0 && mds == diri_auth) {
 258                 /* no delegation info needed. */
 259                 frag = __ceph_find_frag(ci, id);
 260                 if (!frag)
 261                         goto out;
 262                 if (frag->split_by == 0) {
 263                         /* tree leaf, remove */
 264                         dout("fill_dirfrag removed %llx.%llx frag %x"
 265                              " (no ref)\n", ceph_vinop(inode), id);
 266                         rb_erase(&frag->node, &ci->i_fragtree);
 267                         kfree(frag);
 268                 } else {
 269                         /* tree branch, keep and clear */
 270                         dout("fill_dirfrag cleared %llx.%llx frag %x"
 271                              " referral\n", ceph_vinop(inode), id);
 272                         frag->mds = -1;
 273                         frag->ndist = 0;
 274                 }
 275                 goto out;
 276         }
 277
 278
 279         /* find/add this frag to store mds delegation info */
 280         frag = __get_or_create_frag(ci, id);
 281         if (IS_ERR(frag)) {
 282                 /* this is not the end of the world; we can continue
 283                    with bad/inaccurate delegation info */
 284                 pr_err("fill_dirfrag ENOMEM on mds ref %llx.%llx fg %x\n",
 285                        ceph_vinop(inode), le32_to_cpu(dirinfo->frag));
 286                 err = -ENOMEM;
 287                 goto out;
 288         }
 289
 290         frag->mds = mds;
 291         frag->ndist = min_t(u32, ndist, CEPH_MAX_DIRFRAG_REP);
 292         for (i = 0; i < frag->ndist; i++)
 293                 frag->dist[i] = le32_to_cpu(dirinfo->dist[i]);
 294         dout("fill_dirfrag %llx.%llx frag %x ndist=%d\n",
 295              ceph_vinop(inode), frag->frag, frag->ndist);
 296
 297 out:
 298         mutex_unlock(&ci->i_fragtree_mutex);
 299         return err;
 300 }
 301
 302 static int frag_tree_split_cmp(const void *l, const void *r)
 303 {
 304         struct ceph_frag_tree_split *ls = (struct ceph_frag_tree_split*)l;
 305         struct ceph_frag_tree_split *rs = (struct ceph_frag_tree_split*)r;
 306         return ceph_frag_compare(le32_to_cpu(ls->frag),
 307                                  le32_to_cpu(rs->frag));
 308 }
 309
 310 static bool is_frag_child(u32 f, struct ceph_inode_frag *frag)
 311 {
 312         if (!frag)
 313                 return f == ceph_frag_make(0, 0);
 314         if (ceph_frag_bits(f) != ceph_frag_bits(frag->frag) + frag->split_by)
 315                 return false;
 316         return ceph_frag_contains_value(frag->frag, ceph_frag_value(f));
 317 }
 318
 319 static int ceph_fill_fragtree(struct inode *inode,
 320                               struct ceph_frag_tree_head *fragtree,
 321                               struct ceph_mds_reply_dirfrag *dirinfo)
 322 {
 323         struct ceph_inode_info *ci = ceph_inode(inode);
 324         struct ceph_inode_frag *frag, *prev_frag = NULL;
 325         struct rb_node *rb_node;
 326         unsigned i, split_by, nsplits;
 327         u32 id;
 328         bool update = false;
 329
 330         mutex_lock(&ci->i_fragtree_mutex);
 331         nsplits = le32_to_cpu(fragtree->nsplits);
 332         if (nsplits != ci->i_fragtree_nsplits) {
 333                 update = true;
 334         } else if (nsplits) {
 335                 i = prandom_u32() % nsplits;
 336                 id = le32_to_cpu(fragtree->splits[i].frag);
 337                 if (!__ceph_find_frag(ci, id))
 338                         update = true;
 339         } else if (!RB_EMPTY_ROOT(&ci->i_fragtree)) {
 340                 rb_node = rb_first(&ci->i_fragtree);
 341                 frag = rb_entry(rb_node, struct ceph_inode_frag, node);
 342                 if (frag->frag != ceph_frag_make(0, 0) || rb_next(rb_node))
 343                         update = true;
 344         }
 345         if (!update && dirinfo) {
 346                 id = le32_to_cpu(dirinfo->frag);
 347                 if (id != __ceph_choose_frag(ci, id, NULL, NULL))
 348                         update = true;
 349         }
 350         if (!update)
 351                 goto out_unlock;
 352
 353         if (nsplits > 1) {
 354                 sort(fragtree->splits, nsplits, sizeof(fragtree->splits[0]),
 355                      frag_tree_split_cmp, NULL);
 356         }
 357
 358         dout("fill_fragtree %llx.%llx\n", ceph_vinop(inode));
 359         rb_node = rb_first(&ci->i_fragtree);
 360         for (i = 0; i < nsplits; i++) {
 361                 id = le32_to_cpu(fragtree->splits[i].frag);
 362                 split_by = le32_to_cpu(fragtree->splits[i].by);
 363                 if (split_by == 0 || ceph_frag_bits(id) + split_by > 24) {
 364                         pr_err("fill_fragtree %llx.%llx invalid split %d/%u, "
 365                                "frag %x split by %d\n", ceph_vinop(inode),
 366                                i, nsplits, id, split_by);
 367                         continue;
 368                 }
 369                 frag = NULL;
 370                 while (rb_node) {
 371                         frag = rb_entry(rb_node, struct ceph_inode_frag, node);
 372                         if (ceph_frag_compare(frag->frag, id) >= 0) {
 373                                 if (frag->frag != id)
 374                                         frag = NULL;
 375                                 else
 376                                         rb_node = rb_next(rb_node);
 377                                 break;
 378                         }
 379                         rb_node = rb_next(rb_node);
 380                         /* delete stale split/leaf node */
 381                         if (frag->split_by > 0 ||
 382                             !is_frag_child(frag->frag, prev_frag)) {
 383                                 rb_erase(&frag->node, &ci->i_fragtree);
 384                                 if (frag->split_by > 0)
 385                                         ci->i_fragtree_nsplits--;
 386                                 kfree(frag);
 387                         }
 388                         frag = NULL;
 389                 }
 390                 if (!frag) {
 391                         frag = __get_or_create_frag(ci, id);
 392                         if (IS_ERR(frag))
 393                                 continue;
 394                 }
 395                 if (frag->split_by == 0)
 396                         ci->i_fragtree_nsplits++;
 397                 frag->split_by = split_by;
 398                 dout(" frag %x split by %d\n", frag->frag, frag->split_by);
 399                 prev_frag = frag;
 400         }
 401         while (rb_node) {
 402                 frag = rb_entry(rb_node, struct ceph_inode_frag, node);
 403                 rb_node = rb_next(rb_node);
 404                 /* delete stale split/leaf node */
 405                 if (frag->split_by > 0 ||
 406                     !is_frag_child(frag->frag, prev_frag)) {
 407                         rb_erase(&frag->node, &ci->i_fragtree);
 408                         if (frag->split_by > 0)
 409                                 ci->i_fragtree_nsplits--;
 410                         kfree(frag);
 411                 }
 412         }
 413 out_unlock:
 414         mutex_unlock(&ci->i_fragtree_mutex);
 415         return 0;
 416 }
 417
 418 /*
 419  * initialize a newly allocated inode.
 420  */
 421 struct inode *ceph_alloc_inode(struct super_block *sb)
 422 {
 423         struct ceph_inode_info *ci;
 424         int i;
 425
 426         ci = kmem_cache_alloc(ceph_inode_cachep, GFP_NOFS);
 427         if (!ci)
 428                 return NULL;
 429
 430         dout("alloc_inode %p\n", &ci->vfs_inode);
 431
 432         spin_lock_init(&ci->i_ceph_lock);
 433
 434         ci->i_version = 0;
 435         ci->i_inline_version = 0;
 436         ci->i_time_warp_seq = 0;
 437         ci->i_ceph_flags = 0;
 438         atomic64_set(&ci->i_ordered_count, 1);
 439         atomic64_set(&ci->i_release_count, 1);
 440         atomic64_set(&ci->i_complete_seq[0], 0);
 441         atomic64_set(&ci->i_complete_seq[1], 0);
 442         ci->i_symlink = NULL;
 443
 444         ci->i_max_bytes = 0;
 445         ci->i_max_files = 0;
 446
 447         memset(&ci->i_dir_layout, 0, sizeof(ci->i_dir_layout));
 448         RCU_INIT_POINTER(ci->i_layout.pool_ns, NULL);
 449
 450         ci->i_fragtree = RB_ROOT;
 451         mutex_init(&ci->i_fragtree_mutex);
 452
 453         ci->i_xattrs.blob = NULL;
 454         ci->i_xattrs.prealloc_blob = NULL;
 455         ci->i_xattrs.dirty = false;
 456         ci->i_xattrs.index = RB_ROOT;
 457         ci->i_xattrs.count = 0;
 458         ci->i_xattrs.names_size = 0;
 459         ci->i_xattrs.vals_size = 0;
 460         ci->i_xattrs.version = 0;
 461         ci->i_xattrs.index_version = 0;
 462
 463         ci->i_caps = RB_ROOT;
 464         ci->i_auth_cap = NULL;
 465         ci->i_dirty_caps = 0;
 466         ci->i_flushing_caps = 0;
 467         INIT_LIST_HEAD(&ci->i_dirty_item);
 468         INIT_LIST_HEAD(&ci->i_flushing_item);
 469         ci->i_prealloc_cap_flush = NULL;
 470         INIT_LIST_HEAD(&ci->i_cap_flush_list);
 471         init_waitqueue_head(&ci->i_cap_wq);
 472         ci->i_hold_caps_min = 0;
 473         ci->i_hold_caps_max = 0;
 474         INIT_LIST_HEAD(&ci->i_cap_delay_list);
 475         INIT_LIST_HEAD(&ci->i_cap_snaps);
 476         ci->i_head_snapc = NULL;
 477         ci->i_snap_caps = 0;
 478
 479         for (i = 0; i < CEPH_FILE_MODE_BITS; i++)
 480                 ci->i_nr_by_mode[i] = 0;
 481
 482         mutex_init(&ci->i_truncate_mutex);
 483         ci->i_truncate_seq = 0;
 484         ci->i_truncate_size = 0;
 485         ci->i_truncate_pending = 0;
 486
 487         ci->i_max_size = 0;
 488         ci->i_reported_size = 0;
 489         ci->i_wanted_max_size = 0;
 490         ci->i_requested_max_size = 0;
 491
 492         ci->i_pin_ref = 0;
 493         ci->i_rd_ref = 0;
 494         ci->i_rdcache_ref = 0;
 495         ci->i_wr_ref = 0;
 496         ci->i_wb_ref = 0;
 497         ci->i_wrbuffer_ref = 0;
 498         ci->i_wrbuffer_ref_head = 0;
 499         atomic_set(&ci->i_filelock_ref, 0);
 500         atomic_set(&ci->i_shared_gen, 0);
 501         ci->i_rdcache_gen = 0;
 502         ci->i_rdcache_revoking = 0;
 503
 504         INIT_LIST_HEAD(&ci->i_unsafe_dirops);
 505         INIT_LIST_HEAD(&ci->i_unsafe_iops);
 506         spin_lock_init(&ci->i_unsafe_lock);
 507
 508         ci->i_snap_realm = NULL;
 509         INIT_LIST_HEAD(&ci->i_snap_realm_item);
 510         INIT_LIST_HEAD(&ci->i_snap_flush_item);
 511
 512         INIT_WORK(&ci->i_wb_work, ceph_writeback_work);
 513         INIT_WORK(&ci->i_pg_inv_work, ceph_invalidate_work);
 514
 515         INIT_WORK(&ci->i_vmtruncate_work, ceph_vmtruncate_work);
 516
 517         ceph_fscache_inode_init(ci);
 518
 519         return &ci->vfs_inode;
 520 }
 521
 522 static void ceph_i_callback(struct rcu_head *head)
 523 {
 524         struct inode *inode = container_of(head, struct inode, i_rcu);
 525         struct ceph_inode_info *ci = ceph_inode(inode);
 526
 527         kfree(ci->i_symlink);
 528         kmem_cache_free(ceph_inode_cachep, ci);
 529 }
 530
 531 void ceph_evict_inode(struct inode *inode)
 532 {
 533         struct ceph_inode_info *ci = ceph_inode(inode);
 534         struct ceph_inode_frag *frag;
 535         struct rb_node *n;
 536
 537         dout("evict_inode %p ino %llx.%llx\n", inode, ceph_vinop(inode));
 538
 539         truncate_inode_pages_final(&inode->i_data);
 540         clear_inode(inode);
 541
 542         ceph_fscache_unregister_inode_cookie(ci);
 543
 544         ceph_queue_caps_release(inode);
 545
 546         if (__ceph_has_any_quota(ci))
 547                 ceph_adjust_quota_realms_count(inode, false);
 548
 549         /*
 550          * we may still have a snap_realm reference if there are stray
 551          * caps in i_snap_caps.
 552          */
 553         if (ci->i_snap_realm) {
 554                 struct ceph_mds_client *mdsc =
 555                         ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
 556                 struct ceph_snap_realm *realm = ci->i_snap_realm;
 557
 558                 dout(" dropping residual ref to snap realm %p\n", realm);
 559                 spin_lock(&realm->inodes_with_caps_lock);
 560                 list_del_init(&ci->i_snap_realm_item);
 561                 ci->i_snap_realm = NULL;
 562                 if (realm->ino == ci->i_vino.ino)
 563                         realm->inode = NULL;
 564                 spin_unlock(&realm->inodes_with_caps_lock);
 565                 ceph_put_snap_realm(mdsc, realm);
 566         }
 567
 568         while ((n = rb_first(&ci->i_fragtree)) != NULL) {
 569                 frag = rb_entry(n, struct ceph_inode_frag, node);
 570                 rb_erase(n, &ci->i_fragtree);
 571                 kfree(frag);
 572         }
 573         ci->i_fragtree_nsplits = 0;
 574
 575         __ceph_destroy_xattrs(ci);
 576         if (ci->i_xattrs.blob)
 577                 ceph_buffer_put(ci->i_xattrs.blob);
 578         if (ci->i_xattrs.prealloc_blob)
 579                 ceph_buffer_put(ci->i_xattrs.prealloc_blob);
 580
 581         ceph_put_string(rcu_dereference_raw(ci->i_layout.pool_ns));
 582 }
 583
 584 void ceph_destroy_inode(struct inode *inode)
 585 {
 586         call_rcu(&inode->i_rcu, ceph_i_callback);
 587 }
 588
 589 int ceph_drop_inode(struct inode *inode)
 590 {
 591         /*
 592          * Positve dentry and corresponding inode are always accompanied
 593          * in MDS reply. So no need to keep inode in the cache after
 594          * dropping all its aliases.
 595          */
 596         return 1;
 597 }
 598
 599 static inline blkcnt_t calc_inode_blocks(u64 size)
 600 {
 601         return (size + (1<<9) - 1) >> 9;
 602 }
 603
 604 /*
 605  * Helpers to fill in size, ctime, mtime, and atime.  We have to be
 606  * careful because either the client or MDS may have more up to date
 607  * info, depending on which capabilities are held, and whether
 608  * time_warp_seq or truncate_seq have increased.  (Ordinarily, mtime
 609  * and size are monotonically increasing, except when utimes() or
 610  * truncate() increments the corresponding _seq values.)
 611  */
 612 int ceph_fill_file_size(struct inode *inode, int issued,
 613                         u32 truncate_seq, u64 truncate_size, u64 size)
 614 {
 615         struct ceph_inode_info *ci = ceph_inode(inode);
 616         int queue_trunc = 0;
 617
 618         if (ceph_seq_cmp(truncate_seq, ci->i_truncate_seq) > 0 ||
 619             (truncate_seq == ci->i_truncate_seq && size > inode->i_size)) {
 620                 dout("size %lld -> %llu\n", inode->i_size, size);
 621                 if (size > 0 && S_ISDIR(inode->i_mode)) {
 622                         pr_err("fill_file_size non-zero size for directory\n");
 623                         size = 0;
 624                 }
 625                 i_size_write(inode, size);
 626                 inode->i_blocks = calc_inode_blocks(size);
 627                 ci->i_reported_size = size;
 628                 if (truncate_seq != ci->i_truncate_seq) {
 629                         dout("truncate_seq %u -> %u\n",
 630                              ci->i_truncate_seq, truncate_seq);
 631                         ci->i_truncate_seq = truncate_seq;
 632
 633                         /* the MDS should have revoked these caps */
 634                         WARN_ON_ONCE(issued & (CEPH_CAP_FILE_EXCL |
 635                                                CEPH_CAP_FILE_RD |
 636                                                CEPH_CAP_FILE_WR |
 637                                                CEPH_CAP_FILE_LAZYIO));
 638                         /*
 639                          * If we hold relevant caps, or in the case where we're
 640                          * not the only client referencing this file and we
 641                          * don't hold those caps, then we need to check whether
 642                          * the file is either opened or mmaped
 643                          */
 644                         if ((issued & (CEPH_CAP_FILE_CACHE|
 645                                        CEPH_CAP_FILE_BUFFER)) ||
 646                             mapping_mapped(inode->i_mapping) ||
 647                             __ceph_caps_file_wanted(ci)) {
 648                                 ci->i_truncate_pending++;
 649                                 queue_trunc = 1;
 650                         }
 651                 }
 652         }
 653         if (ceph_seq_cmp(truncate_seq, ci->i_truncate_seq) >= 0 &&
 654             ci->i_truncate_size != truncate_size) {
 655                 dout("truncate_size %lld -> %llu\n", ci->i_truncate_size,
 656                      truncate_size);
 657                 ci->i_truncate_size = truncate_size;
 658         }
 659
 660         if (queue_trunc)
 661                 ceph_fscache_invalidate(inode);
 662
 663         return queue_trunc;
 664 }
 665
 666 void ceph_fill_file_time(struct inode *inode, int issued,
 667                          u64 time_warp_seq, struct timespec64 *ctime,
 668                          struct timespec64 *mtime, struct timespec64 *atime)
 669 {
 670         struct ceph_inode_info *ci = ceph_inode(inode);
 671         int warn = 0;
 672
 673         if (issued & (CEPH_CAP_FILE_EXCL|
 674                       CEPH_CAP_FILE_WR|
 675                       CEPH_CAP_FILE_BUFFER|
 676                       CEPH_CAP_AUTH_EXCL|
 677                       CEPH_CAP_XATTR_EXCL)) {
 678                 if (ci->i_version == 0 ||
 679                     timespec64_compare(ctime, &inode->i_ctime) > 0) {
 680                         dout("ctime %lld.%09ld -> %lld.%09ld inc w/ cap\n",
 681                              inode->i_ctime.tv_sec, inode->i_ctime.tv_nsec,
 682                              ctime->tv_sec, ctime->tv_nsec);
 683                         inode->i_ctime = *ctime;
 684                 }
 685                 if (ci->i_version == 0 ||
 686                     ceph_seq_cmp(time_warp_seq, ci->i_time_warp_seq) > 0) {
 687                         /* the MDS did a utimes() */
 688                         dout("mtime %lld.%09ld -> %lld.%09ld "
 689                              "tw %d -> %d\n",
 690                              inode->i_mtime.tv_sec, inode->i_mtime.tv_nsec,
 691                              mtime->tv_sec, mtime->tv_nsec,
 692                              ci->i_time_warp_seq, (int)time_warp_seq);
 693
 694                         inode->i_mtime = *mtime;
 695                         inode->i_atime = *atime;
 696                         ci->i_time_warp_seq = time_warp_seq;
 697                 } else if (time_warp_seq == ci->i_time_warp_seq) {
 698                         /* nobody did utimes(); take the max */
 699                         if (timespec64_compare(mtime, &inode->i_mtime) > 0) {
 700                                 dout("mtime %lld.%09ld -> %lld.%09ld inc\n",
 701                                      inode->i_mtime.tv_sec,
 702                                      inode->i_mtime.tv_nsec,
 703                                      mtime->tv_sec, mtime->tv_nsec);
 704                                 inode->i_mtime = *mtime;
 705                         }
 706                         if (timespec64_compare(atime, &inode->i_atime) > 0) {
 707                                 dout("atime %lld.%09ld -> %lld.%09ld inc\n",
 708                                      inode->i_atime.tv_sec,
 709                                      inode->i_atime.tv_nsec,
 710                                      atime->tv_sec, atime->tv_nsec);
 711                                 inode->i_atime = *atime;
 712                         }
 713                 } else if (issued & CEPH_CAP_FILE_EXCL) {
 714                         /* we did a utimes(); ignore mds values */
 715                 } else {
 716                         warn = 1;
 717                 }
 718         } else {
 719                 /* we have no write|excl caps; whatever the MDS says is true */
 720                 if (ceph_seq_cmp(time_warp_seq, ci->i_time_warp_seq) >= 0) {
 721                         inode->i_ctime = *ctime;
 722                         inode->i_mtime = *mtime;
 723                         inode->i_atime = *atime;
 724                         ci->i_time_warp_seq = time_warp_seq;
 725                 } else {
 726                         warn = 1;
 727                 }
 728         }
 729         if (warn) /* time_warp_seq shouldn't go backwards */
 730                 dout("%p mds time_warp_seq %llu < %u\n",
 731                      inode, time_warp_seq, ci->i_time_warp_seq);
 732 }
 733
 734 /*
 735  * Populate an inode based on info from mds.  May be called on new or
 736  * existing inodes.
 737  */
 738 static int fill_inode(struct inode *inode, struct page *locked_page,
 739                       struct ceph_mds_reply_info_in *iinfo,
 740                       struct ceph_mds_reply_dirfrag *dirinfo,
 741                       struct ceph_mds_session *session,
 742                       unsigned long ttl_from, int cap_fmode,
 743                       struct ceph_cap_reservation *caps_reservation)
 744 {
 745         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
 746         struct ceph_mds_reply_inode *info = iinfo->in;
 747         struct ceph_inode_info *ci = ceph_inode(inode);
 748         int issued, new_issued, info_caps;
 749         struct timespec64 mtime, atime, ctime;
 750         struct ceph_buffer *xattr_blob = NULL;
 751         struct ceph_buffer *old_blob = NULL;
 752         struct ceph_string *pool_ns = NULL;
 753         struct ceph_cap *new_cap = NULL;
 754         int err = 0;
 755         bool wake = false;
 756         bool queue_trunc = false;
 757         bool new_version = false;
 758         bool fill_inline = false;
 759
 760         dout("fill_inode %p ino %llx.%llx v %llu had %llu\n",
 761              inode, ceph_vinop(inode), le64_to_cpu(info->version),
 762              ci->i_version);
 763
 764         info_caps = le32_to_cpu(info->cap.caps);
 765
 766         /* prealloc new cap struct */
 767         if (info_caps && ceph_snap(inode) == CEPH_NOSNAP) {
 768                 new_cap = ceph_get_cap(mdsc, caps_reservation);
 769                 if (!new_cap)
 770                         return -ENOMEM;
 771         }
 772
 773         /*
 774          * prealloc xattr data, if it looks like we'll need it.  only
 775          * if len > 4 (meaning there are actually xattrs; the first 4
 776          * bytes are the xattr count).
 777          */
 778         if (iinfo->xattr_len > 4) {
 779                 xattr_blob = ceph_buffer_new(iinfo->xattr_len, GFP_NOFS);
 780                 if (!xattr_blob)
 781                         pr_err("fill_inode ENOMEM xattr blob %d bytes\n",
 782                                iinfo->xattr_len);
 783         }
 784
 785         if (iinfo->pool_ns_len > 0)
 786                 pool_ns = ceph_find_or_create_string(iinfo->pool_ns_data,
 787                                                      iinfo->pool_ns_len);
 788
 789         spin_lock(&ci->i_ceph_lock);
 790
 791         /*
 792          * provided version will be odd if inode value is projected,
 793          * even if stable.  skip the update if we have newer stable
 794          * info (ours>=theirs, e.g. due to racing mds replies), unless
 795          * we are getting projected (unstable) info (in which case the
 796          * version is odd, and we want ours>theirs).
 797          *   us   them
 798          *   2    2     skip
 799          *   3    2     skip
 800          *   3    3     update
 801          */
 802         if (ci->i_version == 0 ||
 803             ((info->cap.flags & CEPH_CAP_FLAG_AUTH) &&
 804              le64_to_cpu(info->version) > (ci->i_version & ~1)))
 805                 new_version = true;
 806
 807         __ceph_caps_issued(ci, &issued);
 808         issued |= __ceph_caps_dirty(ci);
 809         new_issued = ~issued & info_caps;
 810
 811         /* update inode */
 812         inode->i_rdev = le32_to_cpu(info->rdev);
 813         /* directories have fl_stripe_unit set to zero */
 814         if (le32_to_cpu(info->layout.fl_stripe_unit))
 815                 inode->i_blkbits =
 816                         fls(le32_to_cpu(info->layout.fl_stripe_unit)) - 1;
 817         else
 818                 inode->i_blkbits = CEPH_BLOCK_SHIFT;
 819
 820         __ceph_update_quota(ci, iinfo->max_bytes, iinfo->max_files);
 821
 822         if ((new_version || (new_issued & CEPH_CAP_AUTH_SHARED)) &&
 823             (issued & CEPH_CAP_AUTH_EXCL) == 0) {
 824                 inode->i_mode = le32_to_cpu(info->mode);
 825                 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(info->uid));
 826                 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(info->gid));
 827                 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
 828                      from_kuid(&init_user_ns, inode->i_uid),
 829                      from_kgid(&init_user_ns, inode->i_gid));
 830         }
 831
 832         if ((new_version || (new_issued & CEPH_CAP_LINK_SHARED)) &&
 833             (issued & CEPH_CAP_LINK_EXCL) == 0)
 834                 set_nlink(inode, le32_to_cpu(info->nlink));
 835
 836         if (new_version || (new_issued & CEPH_CAP_ANY_RD)) {
 837                 /* be careful with mtime, atime, size */
 838                 ceph_decode_timespec64(&atime, &info->atime);
 839                 ceph_decode_timespec64(&mtime, &info->mtime);
 840                 ceph_decode_timespec64(&ctime, &info->ctime);
 841                 ceph_fill_file_time(inode, issued,
 842                                 le32_to_cpu(info->time_warp_seq),
 843                                 &ctime, &mtime, &atime);
 844         }
 845
 846         if (new_version || (info_caps & CEPH_CAP_FILE_SHARED)) {
 847                 ci->i_files = le64_to_cpu(info->files);
 848                 ci->i_subdirs = le64_to_cpu(info->subdirs);
 849         }
 850
 851         if (new_version ||
 852             (new_issued & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
 853                 s64 old_pool = ci->i_layout.pool_id;
 854                 struct ceph_string *old_ns;
 855
 856                 ceph_file_layout_from_legacy(&ci->i_layout, &info->layout);
 857                 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
 858                                         lockdep_is_held(&ci->i_ceph_lock));
 859                 rcu_assign_pointer(ci->i_layout.pool_ns, pool_ns);
 860
 861                 if (ci->i_layout.pool_id != old_pool || pool_ns != old_ns)
 862                         ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
 863
 864                 pool_ns = old_ns;
 865
 866                 queue_trunc = ceph_fill_file_size(inode, issued,
 867                                         le32_to_cpu(info->truncate_seq),
 868                                         le64_to_cpu(info->truncate_size),
 869                                         le64_to_cpu(info->size));
 870                 /* only update max_size on auth cap */
 871                 if ((info->cap.flags & CEPH_CAP_FLAG_AUTH) &&
 872                     ci->i_max_size != le64_to_cpu(info->max_size)) {
 873                         dout("max_size %lld -> %llu\n", ci->i_max_size,
 874                                         le64_to_cpu(info->max_size));
 875                         ci->i_max_size = le64_to_cpu(info->max_size);
 876                 }
 877         }
 878
 879         /* layout and rstat are not tracked by capability, update them if
 880          * the inode info is from auth mds */
 881         if (new_version || (info->cap.flags & CEPH_CAP_FLAG_AUTH)) {
 882                 if (S_ISDIR(inode->i_mode)) {
 883                         ci->i_dir_layout = iinfo->dir_layout;
 884                         ci->i_rbytes = le64_to_cpu(info->rbytes);
 885                         ci->i_rfiles = le64_to_cpu(info->rfiles);
 886                         ci->i_rsubdirs = le64_to_cpu(info->rsubdirs);
 887                         ceph_decode_timespec64(&ci->i_rctime, &info->rctime);
 888                 }
 889         }
 890
 891         /* xattrs */
 892         /* note that if i_xattrs.len <= 4, i_xattrs.data will still be NULL. */
 893         if ((ci->i_xattrs.version == 0 || !(issued & CEPH_CAP_XATTR_EXCL))  &&
 894             le64_to_cpu(info->xattr_version) > ci->i_xattrs.version) {
 895                 if (ci->i_xattrs.blob)
 896                         old_blob = ci->i_xattrs.blob;
 897                 ci->i_xattrs.blob = xattr_blob;
 898                 if (xattr_blob)
 899                         memcpy(ci->i_xattrs.blob->vec.iov_base,
 900                                iinfo->xattr_data, iinfo->xattr_len);
 901                 ci->i_xattrs.version = le64_to_cpu(info->xattr_version);
 902                 ceph_forget_all_cached_acls(inode);
 903                 xattr_blob = NULL;
 904         }
 905
 906         /* finally update i_version */
 907         if (le64_to_cpu(info->version) > ci->i_version)
 908                 ci->i_version = le64_to_cpu(info->version);
 909
 910         inode->i_mapping->a_ops = &ceph_aops;
 911
 912         switch (inode->i_mode & S_IFMT) {
 913         case S_IFIFO:
 914         case S_IFBLK:
 915         case S_IFCHR:
 916         case S_IFSOCK:
 917                 init_special_inode(inode, inode->i_mode, inode->i_rdev);
 918                 inode->i_op = &ceph_file_iops;
 919                 break;
 920         case S_IFREG:
 921                 inode->i_op = &ceph_file_iops;
 922                 inode->i_fop = &ceph_file_fops;
 923                 break;
 924         case S_IFLNK:
 925                 inode->i_op = &ceph_symlink_iops;
 926                 if (!ci->i_symlink) {
 927                         u32 symlen = iinfo->symlink_len;
 928                         char *sym;
 929
 930                         spin_unlock(&ci->i_ceph_lock);
 931
 932                         if (symlen != i_size_read(inode)) {
 933                                 pr_err("fill_inode %llx.%llx BAD symlink "
 934                                         "size %lld\n", ceph_vinop(inode),
 935                                         i_size_read(inode));
 936                                 i_size_write(inode, symlen);
 937                                 inode->i_blocks = calc_inode_blocks(symlen);
 938                         }
 939
 940                         err = -ENOMEM;
 941                         sym = kstrndup(iinfo->symlink, symlen, GFP_NOFS);
 942                         if (!sym)
 943                                 goto out;
 944
 945                         spin_lock(&ci->i_ceph_lock);
 946                         if (!ci->i_symlink)
 947                                 ci->i_symlink = sym;
 948                         else
 949                                 kfree(sym); /* lost a race */
 950                 }
 951                 inode->i_link = ci->i_symlink;
 952                 break;
 953         case S_IFDIR:
 954                 inode->i_op = &ceph_dir_iops;
 955                 inode->i_fop = &ceph_dir_fops;
 956                 break;
 957         default:
 958                 pr_err("fill_inode %llx.%llx BAD mode 0%o\n",
 959                        ceph_vinop(inode), inode->i_mode);
 960         }
 961
 962         /* were we issued a capability? */
 963         if (info_caps) {
 964                 if (ceph_snap(inode) == CEPH_NOSNAP) {
 965                         ceph_add_cap(inode, session,
 966                                      le64_to_cpu(info->cap.cap_id),
 967                                      cap_fmode, info_caps,
 968                                      le32_to_cpu(info->cap.wanted),
 969                                      le32_to_cpu(info->cap.seq),
 970                                      le32_to_cpu(info->cap.mseq),
 971                                      le64_to_cpu(info->cap.realm),
 972                                      info->cap.flags, &new_cap);
 973
 974                         /* set dir completion flag? */
 975                         if (S_ISDIR(inode->i_mode) &&
 976                             ci->i_files == 0 && ci->i_subdirs == 0 &&
 977                             (info_caps & CEPH_CAP_FILE_SHARED) &&
 978                             (issued & CEPH_CAP_FILE_EXCL) == 0 &&
 979                             !__ceph_dir_is_complete(ci)) {
 980                                 dout(" marking %p complete (empty)\n", inode);
 981                                 i_size_write(inode, 0);
 982                                 __ceph_dir_set_complete(ci,
 983                                         atomic64_read(&ci->i_release_count),
 984                                         atomic64_read(&ci->i_ordered_count));
 985                         }
 986
 987                         wake = true;
 988                 } else {
 989                         dout(" %p got snap_caps %s\n", inode,
 990                              ceph_cap_string(info_caps));
 991                         ci->i_snap_caps |= info_caps;
 992                         if (cap_fmode >= 0)
 993                                 __ceph_get_fmode(ci, cap_fmode);
 994                 }
 995         } else if (cap_fmode >= 0) {
 996                 pr_warn("mds issued no caps on %llx.%llx\n",
 997                            ceph_vinop(inode));
 998                 __ceph_get_fmode(ci, cap_fmode);
 999         }
1000
1001         if (iinfo->inline_version > 0 &&
1002             iinfo->inline_version >= ci->i_inline_version) {
1003                 int cache_caps = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1004                 ci->i_inline_version = iinfo->inline_version;
1005                 if (ci->i_inline_version != CEPH_INLINE_NONE &&
1006                     (locked_page || (info_caps & cache_caps)))
1007                         fill_inline = true;
1008         }
1009
1010         spin_unlock(&ci->i_ceph_lock);
1011
1012         if (fill_inline)
1013                 ceph_fill_inline_data(inode, locked_page,
1014                                       iinfo->inline_data, iinfo->inline_len);
1015
1016         if (wake)
1017                 wake_up_all(&ci->i_cap_wq);
1018
1019         /* queue truncate if we saw i_size decrease */
1020         if (queue_trunc)
1021                 ceph_queue_vmtruncate(inode);
1022
1023         /* populate frag tree */
1024         if (S_ISDIR(inode->i_mode))
1025                 ceph_fill_fragtree(inode, &info->fragtree, dirinfo);
1026
1027         /* update delegation info? */
1028         if (dirinfo)
1029                 ceph_fill_dirfrag(inode, dirinfo);
1030
1031         err = 0;
1032 out:
1033         if (new_cap)
1034                 ceph_put_cap(mdsc, new_cap);
1035         ceph_buffer_put(old_blob);
1036         ceph_buffer_put(xattr_blob);
1037         ceph_put_string(pool_ns);
1038         return err;
1039 }
1040
1041 /*
1042  * caller should hold session s_mutex.
1043  */
1044 static void update_dentry_lease(struct dentry *dentry,
1045                                 struct ceph_mds_reply_lease *lease,
1046                                 struct ceph_mds_session *session,
1047                                 unsigned long from_time,
1048                                 struct ceph_vino *tgt_vino,
1049                                 struct ceph_vino *dir_vino)
1050 {
1051         struct ceph_dentry_info *di = ceph_dentry(dentry);
1052         long unsigned duration = le32_to_cpu(lease->duration_ms);
1053         long unsigned ttl = from_time + (duration * HZ) / 1000;
1054         long unsigned half_ttl = from_time + (duration * HZ / 2) / 1000;
1055         struct inode *dir;
1056         struct ceph_mds_session *old_lease_session = NULL;
1057
1058         /*
1059          * Make sure dentry's inode matches tgt_vino. NULL tgt_vino means that
1060          * we expect a negative dentry.
1061          */
1062         if (!tgt_vino && d_really_is_positive(dentry))
1063                 return;
1064
1065         if (tgt_vino && (d_really_is_negative(dentry) ||
1066                         !ceph_ino_compare(d_inode(dentry), tgt_vino)))
1067                 return;
1068
1069         spin_lock(&dentry->d_lock);
1070         dout("update_dentry_lease %p duration %lu ms ttl %lu\n",
1071              dentry, duration, ttl);
1072
1073         dir = d_inode(dentry->d_parent);
1074
1075         /* make sure parent matches dir_vino */
1076         if (!ceph_ino_compare(dir, dir_vino))
1077                 goto out_unlock;
1078
1079         /* only track leases on regular dentries */
1080         if (ceph_snap(dir) != CEPH_NOSNAP)
1081                 goto out_unlock;
1082
1083         di->lease_shared_gen = atomic_read(&ceph_inode(dir)->i_shared_gen);
1084
1085         if (duration == 0)
1086                 goto out_unlock;
1087
1088         if (di->lease_gen == session->s_cap_gen &&
1089             time_before(ttl, di->time))
1090                 goto out_unlock;  /* we already have a newer lease. */
1091
1092         if (di->lease_session && di->lease_session != session) {
1093                 old_lease_session = di->lease_session;
1094                 di->lease_session = NULL;
1095         }
1096
1097         ceph_dentry_lru_touch(dentry);
1098
1099         if (!di->lease_session)
1100                 di->lease_session = ceph_get_mds_session(session);
1101         di->lease_gen = session->s_cap_gen;
1102         di->lease_seq = le32_to_cpu(lease->seq);
1103         di->lease_renew_after = half_ttl;
1104         di->lease_renew_from = 0;
1105         di->time = ttl;
1106 out_unlock:
1107         spin_unlock(&dentry->d_lock);
1108         if (old_lease_session)
1109                 ceph_put_mds_session(old_lease_session);
1110 }
1111
1112 /*
1113  * splice a dentry to an inode.
1114  * caller must hold directory i_mutex for this to be safe.
1115  */
1116 static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)
1117 {
1118         struct dentry *realdn;
1119
1120         BUG_ON(d_inode(dn));
1121
1122         if (S_ISDIR(in->i_mode)) {
1123                 /* If inode is directory, d_splice_alias() below will remove
1124                  * 'realdn' from its origin parent. We need to ensure that
1125                  * origin parent's readdir cache will not reference 'realdn'
1126                  */
1127                 realdn = d_find_any_alias(in);
1128                 if (realdn) {
1129                         struct ceph_dentry_info *di = ceph_dentry(realdn);
1130                         spin_lock(&realdn->d_lock);
1131
1132                         realdn->d_op->d_prune(realdn);
1133
1134                         di->time = jiffies;
1135                         di->lease_shared_gen = 0;
1136                         di->offset = 0;
1137
1138                         spin_unlock(&realdn->d_lock);
1139                         dput(realdn);
1140                 }
1141         }
1142
1143         /* dn must be unhashed */
1144         if (!d_unhashed(dn))
1145                 d_drop(dn);
1146         realdn = d_splice_alias(in, dn);
1147         if (IS_ERR(realdn)) {
1148                 pr_err("splice_dentry error %ld %p inode %p ino %llx.%llx\n",
1149                        PTR_ERR(realdn), dn, in, ceph_vinop(in));
1150                 dn = realdn;
1151                 /*
1152                  * Caller should release 'dn' in the case of error.
1153                  * If 'req->r_dentry' is passed to this function,
1154                  * caller should leave 'req->r_dentry' untouched.
1155                  */
1156                 goto out;
1157         } else if (realdn) {
1158                 dout("dn %p (%d) spliced with %p (%d) "
1159                      "inode %p ino %llx.%llx\n",
1160                      dn, d_count(dn),
1161                      realdn, d_count(realdn),
1162                      d_inode(realdn), ceph_vinop(d_inode(realdn)));
1163                 dput(dn);
1164                 dn = realdn;
1165         } else {
1166                 BUG_ON(!ceph_dentry(dn));
1167                 dout("dn %p attached to %p ino %llx.%llx\n",
1168                      dn, d_inode(dn), ceph_vinop(d_inode(dn)));
1169         }
1170 out:
1171         return dn;
1172 }
1173
1174 /*
1175  * Incorporate results into the local cache.  This is either just
1176  * one inode, or a directory, dentry, and possibly linked-to inode (e.g.,
1177  * after a lookup).
1178  *
1179  * A reply may contain
1180  *         a directory inode along with a dentry.
1181  *  and/or a target inode
1182  *
1183  * Called with snap_rwsem (read).
1184  */
1185 int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req)
1186 {
1187         struct ceph_mds_session *session = req->r_session;
1188         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1189         struct inode *in = NULL;
1190         struct ceph_vino tvino, dvino;
1191         struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
1192         int err = 0;
1193
1194         dout("fill_trace %p is_dentry %d is_target %d\n", req,
1195              rinfo->head->is_dentry, rinfo->head->is_target);
1196
1197         if (!rinfo->head->is_target && !rinfo->head->is_dentry) {
1198                 dout("fill_trace reply is empty!\n");
1199                 if (rinfo->head->result == 0 && req->r_parent)
1200                         ceph_invalidate_dir_request(req);
1201                 return 0;
1202         }
1203
1204         if (rinfo->head->is_dentry) {
1205                 struct inode *dir = req->r_parent;
1206
1207                 if (dir) {
1208                         err = fill_inode(dir, NULL,
1209                                          &rinfo->diri, rinfo->dirfrag,
1210                                          session, req->r_request_started, -1,
1211                                          &req->r_caps_reservation);
1212                         if (err < 0)
1213                                 goto done;
1214                 } else {
1215                         WARN_ON_ONCE(1);
1216                 }
1217
1218                 if (dir && req->r_op == CEPH_MDS_OP_LOOKUPNAME) {
1219                         struct qstr dname;
1220                         struct dentry *dn, *parent;
1221
1222                         BUG_ON(!rinfo->head->is_target);
1223                         BUG_ON(req->r_dentry);
1224
1225                         parent = d_find_any_alias(dir);
1226                         BUG_ON(!parent);
1227
1228                         dname.name = rinfo->dname;
1229                         dname.len = rinfo->dname_len;
1230                         dname.hash = full_name_hash(parent, dname.name, dname.len);
1231                         tvino.ino = le64_to_cpu(rinfo->targeti.in->ino);
1232                         tvino.snap = le64_to_cpu(rinfo->targeti.in->snapid);
1233 retry_lookup:
1234                         dn = d_lookup(parent, &dname);
1235                         dout("d_lookup on parent=%p name=%.*s got %p\n",
1236                              parent, dname.len, dname.name, dn);
1237
1238                         if (!dn) {
1239                                 dn = d_alloc(parent, &dname);
1240                                 dout("d_alloc %p '%.*s' = %p\n", parent,
1241                                      dname.len, dname.name, dn);
1242                                 if (!dn) {
1243                                         dput(parent);
1244                                         err = -ENOMEM;
1245                                         goto done;
1246                                 }
1247                                 err = 0;
1248                         } else if (d_really_is_positive(dn) &&
1249                                    (ceph_ino(d_inode(dn)) != tvino.ino ||
1250                                     ceph_snap(d_inode(dn)) != tvino.snap)) {
1251                                 dout(" dn %p points to wrong inode %p\n",
1252                                      dn, d_inode(dn));
1253                                 ceph_dir_clear_ordered(dir);
1254                                 d_delete(dn);
1255                                 dput(dn);
1256                                 goto retry_lookup;
1257                         }
1258
1259                         req->r_dentry = dn;
1260                         dput(parent);
1261                 }
1262         }
1263
1264         if (rinfo->head->is_target) {
1265                 tvino.ino = le64_to_cpu(rinfo->targeti.in->ino);
1266                 tvino.snap = le64_to_cpu(rinfo->targeti.in->snapid);
1267
1268                 in = ceph_get_inode(sb, tvino);
1269                 if (IS_ERR(in)) {
1270                         err = PTR_ERR(in);
1271                         goto done;
1272                 }
1273                 req->r_target_inode = in;
1274
1275                 err = fill_inode(in, req->r_locked_page, &rinfo->targeti, NULL,
1276                                 session, req->r_request_started,
1277                                 (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags) &&
1278                                 rinfo->head->result == 0) ?  req->r_fmode : -1,
1279                                 &req->r_caps_reservation);
1280                 if (err < 0) {
1281                         pr_err("fill_inode badness %p %llx.%llx\n",
1282                                 in, ceph_vinop(in));
1283                         goto done;
1284                 }
1285         }
1286
1287         /*
1288          * ignore null lease/binding on snapdir ENOENT, or else we
1289          * will have trouble splicing in the virtual snapdir later
1290          */
1291         if (rinfo->head->is_dentry &&
1292             !test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags) &&
1293             test_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags) &&
1294             (rinfo->head->is_target || strncmp(req->r_dentry->d_name.name,
1295                                                fsc->mount_options->snapdir_name,
1296                                                req->r_dentry->d_name.len))) {
1297                 /*
1298                  * lookup link rename   : null -> possibly existing inode
1299                  * mknod symlink mkdir  : null -> new inode
1300                  * unlink               : linked -> null
1301                  */
1302                 struct inode *dir = req->r_parent;
1303                 struct dentry *dn = req->r_dentry;
1304                 bool have_dir_cap, have_lease;
1305
1306                 BUG_ON(!dn);
1307                 BUG_ON(!dir);
1308                 BUG_ON(d_inode(dn->d_parent) != dir);
1309
1310                 dvino.ino = le64_to_cpu(rinfo->diri.in->ino);
1311                 dvino.snap = le64_to_cpu(rinfo->diri.in->snapid);
1312
1313                 BUG_ON(ceph_ino(dir) != dvino.ino);
1314                 BUG_ON(ceph_snap(dir) != dvino.snap);
1315
1316                 /* do we have a lease on the whole dir? */
1317                 have_dir_cap =
1318                         (le32_to_cpu(rinfo->diri.in->cap.caps) &
1319                          CEPH_CAP_FILE_SHARED);
1320
1321                 /* do we have a dn lease? */
1322                 have_lease = have_dir_cap ||
1323                         le32_to_cpu(rinfo->dlease->duration_ms);
1324                 if (!have_lease)
1325                         dout("fill_trace  no dentry lease or dir cap\n");
1326
1327                 /* rename? */
1328                 if (req->r_old_dentry && req->r_op == CEPH_MDS_OP_RENAME) {
1329                         struct inode *olddir = req->r_old_dentry_dir;
1330                         BUG_ON(!olddir);
1331
1332                         dout(" src %p '%pd' dst %p '%pd'\n",
1333                              req->r_old_dentry,
1334                              req->r_old_dentry,
1335                              dn, dn);
1336                         dout("fill_trace doing d_move %p -> %p\n",
1337                              req->r_old_dentry, dn);
1338
1339                         /* d_move screws up sibling dentries' offsets */
1340                         ceph_dir_clear_ordered(dir);
1341                         ceph_dir_clear_ordered(olddir);
1342
1343                         d_move(req->r_old_dentry, dn);
1344                         dout(" src %p '%pd' dst %p '%pd'\n",
1345                              req->r_old_dentry,
1346                              req->r_old_dentry,
1347                              dn, dn);
1348
1349                         /* ensure target dentry is invalidated, despite
1350                            rehashing bug in vfs_rename_dir */
1351                         ceph_invalidate_dentry_lease(dn);
1352
1353                         dout("dn %p gets new offset %lld\n", req->r_old_dentry,
1354                              ceph_dentry(req->r_old_dentry)->offset);
1355
1356                         dn = req->r_old_dentry;  /* use old_dentry */
1357                 }
1358
1359                 /* null dentry? */
1360                 if (!rinfo->head->is_target) {
1361                         dout("fill_trace null dentry\n");
1362                         if (d_really_is_positive(dn)) {
1363                                 dout("d_delete %p\n", dn);
1364                                 ceph_dir_clear_ordered(dir);
1365                                 d_delete(dn);
1366                         } else if (have_lease) {
1367                                 if (d_unhashed(dn))
1368                                         d_add(dn, NULL);
1369                                 update_dentry_lease(dn, rinfo->dlease,
1370                                                     session,
1371                                                     req->r_request_started,
1372                                                     NULL, &dvino);
1373                         }
1374                         goto done;
1375                 }
1376
1377                 /* attach proper inode */
1378                 if (d_really_is_negative(dn)) {
1379                         ceph_dir_clear_ordered(dir);
1380                         ihold(in);
1381                         dn = splice_dentry(dn, in);
1382                         if (IS_ERR(dn)) {
1383                                 err = PTR_ERR(dn);
1384                                 goto done;
1385                         }
1386                         req->r_dentry = dn;  /* may have spliced */
1387                 } else if (d_really_is_positive(dn) && d_inode(dn) != in) {
1388                         dout(" %p links to %p %llx.%llx, not %llx.%llx\n",
1389                              dn, d_inode(dn), ceph_vinop(d_inode(dn)),
1390                              ceph_vinop(in));
1391                         d_invalidate(dn);
1392                         have_lease = false;
1393                 }
1394
1395                 if (have_lease) {
1396                         tvino.ino = le64_to_cpu(rinfo->targeti.in->ino);
1397                         tvino.snap = le64_to_cpu(rinfo->targeti.in->snapid);
1398                         update_dentry_lease(dn, rinfo->dlease, session,
1399                                             req->r_request_started,
1400                                             &tvino, &dvino);
1401                 }
1402                 dout(" final dn %p\n", dn);
1403         } else if ((req->r_op == CEPH_MDS_OP_LOOKUPSNAP ||
1404                     req->r_op == CEPH_MDS_OP_MKSNAP) &&
1405                    test_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags) &&
1406                    !test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
1407                 struct dentry *dn = req->r_dentry;
1408                 struct inode *dir = req->r_parent;
1409
1410                 /* fill out a snapdir LOOKUPSNAP dentry */
1411                 BUG_ON(!dn);
1412                 BUG_ON(!dir);
1413                 BUG_ON(ceph_snap(dir) != CEPH_SNAPDIR);
1414                 dout(" linking snapped dir %p to dn %p\n", in, dn);
1415                 ceph_dir_clear_ordered(dir);
1416                 ihold(in);
1417                 dn = splice_dentry(dn, in);
1418                 if (IS_ERR(dn)) {
1419                         err = PTR_ERR(dn);
1420                         goto done;
1421                 }
1422                 req->r_dentry = dn;  /* may have spliced */
1423         } else if (rinfo->head->is_dentry) {
1424                 struct ceph_vino *ptvino = NULL;
1425
1426                 if ((le32_to_cpu(rinfo->diri.in->cap.caps) & CEPH_CAP_FILE_SHARED) ||
1427                     le32_to_cpu(rinfo->dlease->duration_ms)) {
1428                         dvino.ino = le64_to_cpu(rinfo->diri.in->ino);
1429                         dvino.snap = le64_to_cpu(rinfo->diri.in->snapid);
1430
1431                         if (rinfo->head->is_target) {
1432                                 tvino.ino = le64_to_cpu(rinfo->targeti.in->ino);
1433                                 tvino.snap = le64_to_cpu(rinfo->targeti.in->snapid);
1434                                 ptvino = &tvino;
1435                         }
1436
1437                         update_dentry_lease(req->r_dentry, rinfo->dlease,
1438                                 session, req->r_request_started, ptvino,
1439                                 &dvino);
1440                 } else {
1441                         dout("%s: no dentry lease or dir cap\n", __func__);
1442                 }
1443         }
1444 done:
1445         dout("fill_trace done err=%d\n", err);
1446         return err;
1447 }
1448
1449 /*
1450  * Prepopulate our cache with readdir results, leases, etc.
1451  */
1452 static int readdir_prepopulate_inodes_only(struct ceph_mds_request *req,
1453                                            struct ceph_mds_session *session)
1454 {
1455         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1456         int i, err = 0;
1457
1458         for (i = 0; i < rinfo->dir_nr; i++) {
1459                 struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
1460                 struct ceph_vino vino;
1461                 struct inode *in;
1462                 int rc;
1463
1464                 vino.ino = le64_to_cpu(rde->inode.in->ino);
1465                 vino.snap = le64_to_cpu(rde->inode.in->snapid);
1466
1467                 in = ceph_get_inode(req->r_dentry->d_sb, vino);
1468                 if (IS_ERR(in)) {
1469                         err = PTR_ERR(in);
1470                         dout("new_inode badness got %d\n", err);
1471                         continue;
1472                 }
1473                 rc = fill_inode(in, NULL, &rde->inode, NULL, session,
1474                                 req->r_request_started, -1,
1475                                 &req->r_caps_reservation);
1476                 if (rc < 0) {
1477                         pr_err("fill_inode badness on %p got %d\n", in, rc);
1478                         err = rc;
1479                 }
1480                 iput(in);
1481         }
1482
1483         return err;
1484 }
1485
1486 void ceph_readdir_cache_release(struct ceph_readdir_cache_control *ctl)
1487 {
1488         if (ctl->page) {
1489                 kunmap(ctl->page);
1490                 put_page(ctl->page);
1491                 ctl->page = NULL;
1492         }
1493 }
1494
1495 static int fill_readdir_cache(struct inode *dir, struct dentry *dn,
1496                               struct ceph_readdir_cache_control *ctl,
1497                               struct ceph_mds_request *req)
1498 {
1499         struct ceph_inode_info *ci = ceph_inode(dir);
1500         unsigned nsize = PAGE_SIZE / sizeof(struct dentry*);
1501         unsigned idx = ctl->index % nsize;
1502         pgoff_t pgoff = ctl->index / nsize;
1503
1504         if (!ctl->page || pgoff != page_index(ctl->page)) {
1505                 ceph_readdir_cache_release(ctl);
1506                 if (idx == 0)
1507                         ctl->page = grab_cache_page(&dir->i_data, pgoff);
1508                 else
1509                         ctl->page = find_lock_page(&dir->i_data, pgoff);
1510                 if (!ctl->page) {
1511                         ctl->index = -1;
1512                         return idx == 0 ? -ENOMEM : 0;
1513                 }
1514                 /* reading/filling the cache are serialized by
1515                  * i_mutex, no need to use page lock */
1516                 unlock_page(ctl->page);
1517                 ctl->dentries = kmap(ctl->page);
1518                 if (idx == 0)
1519                         memset(ctl->dentries, 0, PAGE_SIZE);
1520         }
1521
1522         if (req->r_dir_release_cnt == atomic64_read(&ci->i_release_count) &&
1523             req->r_dir_ordered_cnt == atomic64_read(&ci->i_ordered_count)) {
1524                 dout("readdir cache dn %p idx %d\n", dn, ctl->index);
1525                 ctl->dentries[idx] = dn;
1526                 ctl->index++;
1527         } else {
1528                 dout("disable readdir cache\n");
1529                 ctl->index = -1;
1530         }
1531         return 0;
1532 }
1533
1534 int ceph_readdir_prepopulate(struct ceph_mds_request *req,
1535                              struct ceph_mds_session *session)
1536 {
1537         struct dentry *parent = req->r_dentry;
1538         struct ceph_inode_info *ci = ceph_inode(d_inode(parent));
1539         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1540         struct qstr dname;
1541         struct dentry *dn;
1542         struct inode *in;
1543         int err = 0, skipped = 0, ret, i;
1544         struct ceph_mds_request_head *rhead = req->r_request->front.iov_base;
1545         u32 frag = le32_to_cpu(rhead->args.readdir.frag);
1546         u32 last_hash = 0;
1547         u32 fpos_offset;
1548         struct ceph_readdir_cache_control cache_ctl = {};
1549
1550         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
1551                 return readdir_prepopulate_inodes_only(req, session);
1552
1553         if (rinfo->hash_order) {
1554                 if (req->r_path2) {
1555                         last_hash = ceph_str_hash(ci->i_dir_layout.dl_dir_hash,
1556                                                   req->r_path2,
1557                                                   strlen(req->r_path2));
1558                         last_hash = ceph_frag_value(last_hash);
1559                 } else if (rinfo->offset_hash) {
1560                         /* mds understands offset_hash */
1561                         WARN_ON_ONCE(req->r_readdir_offset != 2);
1562                         last_hash = le32_to_cpu(rhead->args.readdir.offset_hash);
1563                 }
1564         }
1565
1566         if (rinfo->dir_dir &&
1567             le32_to_cpu(rinfo->dir_dir->frag) != frag) {
1568                 dout("readdir_prepopulate got new frag %x -> %x\n",
1569                      frag, le32_to_cpu(rinfo->dir_dir->frag));
1570                 frag = le32_to_cpu(rinfo->dir_dir->frag);
1571                 if (!rinfo->hash_order)
1572                         req->r_readdir_offset = 2;
1573         }
1574
1575         if (le32_to_cpu(rinfo->head->op) == CEPH_MDS_OP_LSSNAP) {
1576                 dout("readdir_prepopulate %d items under SNAPDIR dn %p\n",
1577                      rinfo->dir_nr, parent);
1578         } else {
1579                 dout("readdir_prepopulate %d items under dn %p\n",
1580                      rinfo->dir_nr, parent);
1581                 if (rinfo->dir_dir)
1582                         ceph_fill_dirfrag(d_inode(parent), rinfo->dir_dir);
1583
1584                 if (ceph_frag_is_leftmost(frag) &&
1585                     req->r_readdir_offset == 2 &&
1586                     !(rinfo->hash_order && last_hash)) {
1587                         /* note dir version at start of readdir so we can
1588                          * tell if any dentries get dropped */
1589                         req->r_dir_release_cnt =
1590                                 atomic64_read(&ci->i_release_count);
1591                         req->r_dir_ordered_cnt =
1592                                 atomic64_read(&ci->i_ordered_count);
1593                         req->r_readdir_cache_idx = 0;
1594                 }
1595         }
1596
1597         cache_ctl.index = req->r_readdir_cache_idx;
1598         fpos_offset = req->r_readdir_offset;
1599
1600         /* FIXME: release caps/leases if error occurs */
1601         for (i = 0; i < rinfo->dir_nr; i++) {
1602                 struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
1603                 struct ceph_vino tvino, dvino;
1604
1605                 dname.name = rde->name;
1606                 dname.len = rde->name_len;
1607                 dname.hash = full_name_hash(parent, dname.name, dname.len);
1608
1609                 tvino.ino = le64_to_cpu(rde->inode.in->ino);
1610                 tvino.snap = le64_to_cpu(rde->inode.in->snapid);
1611
1612                 if (rinfo->hash_order) {
1613                         u32 hash = ceph_str_hash(ci->i_dir_layout.dl_dir_hash,
1614                                                  rde->name, rde->name_len);
1615                         hash = ceph_frag_value(hash);
1616                         if (hash != last_hash)
1617                                 fpos_offset = 2;
1618                         last_hash = hash;
1619                         rde->offset = ceph_make_fpos(hash, fpos_offset++, true);
1620                 } else {
1621                         rde->offset = ceph_make_fpos(frag, fpos_offset++, false);
1622                 }
1623
1624 retry_lookup:
1625                 dn = d_lookup(parent, &dname);
1626                 dout("d_lookup on parent=%p name=%.*s got %p\n",
1627                      parent, dname.len, dname.name, dn);
1628
1629                 if (!dn) {
1630                         dn = d_alloc(parent, &dname);
1631                         dout("d_alloc %p '%.*s' = %p\n", parent,
1632                              dname.len, dname.name, dn);
1633                         if (!dn) {
1634                                 dout("d_alloc badness\n");
1635                                 err = -ENOMEM;
1636                                 goto out;
1637                         }
1638                 } else if (d_really_is_positive(dn) &&
1639                            (ceph_ino(d_inode(dn)) != tvino.ino ||
1640                             ceph_snap(d_inode(dn)) != tvino.snap)) {
1641                         struct ceph_dentry_info *di = ceph_dentry(dn);
1642                         dout(" dn %p points to wrong inode %p\n",
1643                              dn, d_inode(dn));
1644
1645                         spin_lock(&dn->d_lock);
1646                         if (di->offset > 0 &&
1647                             di->lease_shared_gen ==
1648                             atomic_read(&ci->i_shared_gen)) {
1649                                 __ceph_dir_clear_ordered(ci);
1650                                 di->offset = 0;
1651                         }
1652                         spin_unlock(&dn->d_lock);
1653
1654                         d_delete(dn);
1655                         dput(dn);
1656                         goto retry_lookup;
1657                 }
1658
1659                 /* inode */
1660                 if (d_really_is_positive(dn)) {
1661                         in = d_inode(dn);
1662                 } else {
1663                         in = ceph_get_inode(parent->d_sb, tvino);
1664                         if (IS_ERR(in)) {
1665                                 dout("new_inode badness\n");
1666                                 d_drop(dn);
1667                                 dput(dn);
1668                                 err = PTR_ERR(in);
1669                                 goto out;
1670                         }
1671                 }
1672
1673                 ret = fill_inode(in, NULL, &rde->inode, NULL, session,
1674                                  req->r_request_started, -1,
1675                                  &req->r_caps_reservation);
1676                 if (ret < 0) {
1677                         pr_err("fill_inode badness on %p\n", in);
1678                         if (d_really_is_negative(dn))
1679                                 iput(in);
1680                         d_drop(dn);
1681                         err = ret;
1682                         goto next_item;
1683                 }
1684
1685                 if (d_really_is_negative(dn)) {
1686                         struct dentry *realdn;
1687
1688                         if (ceph_security_xattr_deadlock(in)) {
1689                                 dout(" skip splicing dn %p to inode %p"
1690                                      " (security xattr deadlock)\n", dn, in);
1691                                 iput(in);
1692                                 skipped++;
1693                                 goto next_item;
1694                         }
1695
1696                         realdn = splice_dentry(dn, in);
1697                         if (IS_ERR(realdn)) {
1698                                 err = PTR_ERR(realdn);
1699                                 d_drop(dn);
1700                                 goto next_item;
1701                         }
1702                         dn = realdn;
1703                 }
1704
1705                 ceph_dentry(dn)->offset = rde->offset;
1706
1707                 dvino = ceph_vino(d_inode(parent));
1708                 update_dentry_lease(dn, rde->lease, req->r_session,
1709                                     req->r_request_started, &tvino, &dvino);
1710
1711                 if (err == 0 && skipped == 0 && cache_ctl.index >= 0) {
1712                         ret = fill_readdir_cache(d_inode(parent), dn,
1713                                                  &cache_ctl, req);
1714                         if (ret < 0)
1715                                 err = ret;
1716                 }
1717 next_item:
1718                 if (dn)
1719                         dput(dn);
1720         }
1721 out:
1722         if (err == 0 && skipped == 0) {
1723                 set_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags);
1724                 req->r_readdir_cache_idx = cache_ctl.index;
1725         }
1726         ceph_readdir_cache_release(&cache_ctl);
1727         dout("readdir_prepopulate done\n");
1728         return err;
1729 }
1730
1731 bool ceph_inode_set_size(struct inode *inode, loff_t size)
1732 {
1733         struct ceph_inode_info *ci = ceph_inode(inode);
1734         bool ret;
1735
1736         spin_lock(&ci->i_ceph_lock);
1737         dout("set_size %p %llu -> %llu\n", inode, inode->i_size, size);
1738         i_size_write(inode, size);
1739         inode->i_blocks = calc_inode_blocks(size);
1740
1741         ret = __ceph_should_report_size(ci);
1742
1743         spin_unlock(&ci->i_ceph_lock);
1744         return ret;
1745 }
1746
1747 /*
1748  * Write back inode data in a worker thread.  (This can't be done
1749  * in the message handler context.)
1750  */
1751 void ceph_queue_writeback(struct inode *inode)
1752 {
1753         ihold(inode);
1754         if (queue_work(ceph_inode_to_client(inode)->wb_wq,
1755                        &ceph_inode(inode)->i_wb_work)) {
1756                 dout("ceph_queue_writeback %p\n", inode);
1757         } else {
1758                 dout("ceph_queue_writeback %p failed\n", inode);
1759                 iput(inode);
1760         }
1761 }
1762
1763 static void ceph_writeback_work(struct work_struct *work)
1764 {
1765         struct ceph_inode_info *ci = container_of(work, struct ceph_inode_info,
1766                                                   i_wb_work);
1767         struct inode *inode = &ci->vfs_inode;
1768
1769         dout("writeback %p\n", inode);
1770         filemap_fdatawrite(&inode->i_data);
1771         iput(inode);
1772 }
1773
1774 /*
1775  * queue an async invalidation
1776  */
1777 void ceph_queue_invalidate(struct inode *inode)
1778 {
1779         ihold(inode);
1780         if (queue_work(ceph_inode_to_client(inode)->pg_inv_wq,
1781                        &ceph_inode(inode)->i_pg_inv_work)) {
1782                 dout("ceph_queue_invalidate %p\n", inode);
1783         } else {
1784                 dout("ceph_queue_invalidate %p failed\n", inode);
1785                 iput(inode);
1786         }
1787 }
1788
1789 /*
1790  * Invalidate inode pages in a worker thread.  (This can't be done
1791  * in the message handler context.)
1792  */
1793 static void ceph_invalidate_work(struct work_struct *work)
1794 {
1795         struct ceph_inode_info *ci = container_of(work, struct ceph_inode_info,
1796                                                   i_pg_inv_work);
1797         struct inode *inode = &ci->vfs_inode;
1798         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1799         u32 orig_gen;
1800         int check = 0;
1801
1802         mutex_lock(&ci->i_truncate_mutex);
1803
1804         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1805                 pr_warn_ratelimited("invalidate_pages %p %lld forced umount\n",
1806                                     inode, ceph_ino(inode));
1807                 mapping_set_error(inode->i_mapping, -EIO);
1808                 truncate_pagecache(inode, 0);
1809                 mutex_unlock(&ci->i_truncate_mutex);
1810                 goto out;
1811         }
1812
1813         spin_lock(&ci->i_ceph_lock);
1814         dout("invalidate_pages %p gen %d revoking %d\n", inode,
1815              ci->i_rdcache_gen, ci->i_rdcache_revoking);
1816         if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
1817                 if (__ceph_caps_revoking_other(ci, NULL, CEPH_CAP_FILE_CACHE))
1818                         check = 1;
1819                 spin_unlock(&ci->i_ceph_lock);
1820                 mutex_unlock(&ci->i_truncate_mutex);
1821                 goto out;
1822         }
1823         orig_gen = ci->i_rdcache_gen;
1824         spin_unlock(&ci->i_ceph_lock);
1825
1826         ceph_fscache_invalidate(inode);
1827         if (invalidate_inode_pages2(inode->i_mapping) < 0) {
1828                 pr_err("invalidate_pages %p fails\n", inode);
1829         }
1830
1831         spin_lock(&ci->i_ceph_lock);
1832         if (orig_gen == ci->i_rdcache_gen &&
1833             orig_gen == ci->i_rdcache_revoking) {
1834                 dout("invalidate_pages %p gen %d successful\n", inode,
1835                      ci->i_rdcache_gen);
1836                 ci->i_rdcache_revoking--;
1837                 check = 1;
1838         } else {
1839                 dout("invalidate_pages %p gen %d raced, now %d revoking %d\n",
1840                      inode, orig_gen, ci->i_rdcache_gen,
1841                      ci->i_rdcache_revoking);
1842                 if (__ceph_caps_revoking_other(ci, NULL, CEPH_CAP_FILE_CACHE))
1843                         check = 1;
1844         }
1845         spin_unlock(&ci->i_ceph_lock);
1846         mutex_unlock(&ci->i_truncate_mutex);
1847 out:
1848         if (check)
1849                 ceph_check_caps(ci, 0, NULL);
1850         iput(inode);
1851 }
1852
1853
1854 /*
1855  * called by trunc_wq;
1856  *
1857  * We also truncate in a separate thread as well.
1858  */
1859 static void ceph_vmtruncate_work(struct work_struct *work)
1860 {
1861         struct ceph_inode_info *ci = container_of(work, struct ceph_inode_info,
1862                                                   i_vmtruncate_work);
1863         struct inode *inode = &ci->vfs_inode;
1864
1865         dout("vmtruncate_work %p\n", inode);
1866         __ceph_do_pending_vmtruncate(inode);
1867         iput(inode);
1868 }
1869
1870 /*
1871  * Queue an async vmtruncate.  If we fail to queue work, we will handle
1872  * the truncation the next time we call __ceph_do_pending_vmtruncate.
1873  */
1874 void ceph_queue_vmtruncate(struct inode *inode)
1875 {
1876         struct ceph_inode_info *ci = ceph_inode(inode);
1877
1878         ihold(inode);
1879
1880         if (queue_work(ceph_sb_to_client(inode->i_sb)->trunc_wq,
1881                        &ci->i_vmtruncate_work)) {
1882                 dout("ceph_queue_vmtruncate %p\n", inode);
1883         } else {
1884                 dout("ceph_queue_vmtruncate %p failed, pending=%d\n",
1885                      inode, ci->i_truncate_pending);
1886                 iput(inode);
1887         }
1888 }
1889
1890 /*
1891  * Make sure any pending truncation is applied before doing anything
1892  * that may depend on it.
1893  */
1894 void __ceph_do_pending_vmtruncate(struct inode *inode)
1895 {
1896         struct ceph_inode_info *ci = ceph_inode(inode);
1897         u64 to;
1898         int wrbuffer_refs, finish = 0;
1899
1900         mutex_lock(&ci->i_truncate_mutex);
1901 retry:
1902         spin_lock(&ci->i_ceph_lock);
1903         if (ci->i_truncate_pending == 0) {
1904                 dout("__do_pending_vmtruncate %p none pending\n", inode);
1905                 spin_unlock(&ci->i_ceph_lock);
1906                 mutex_unlock(&ci->i_truncate_mutex);
1907                 return;
1908         }
1909
1910         /*
1911          * make sure any dirty snapped pages are flushed before we
1912          * possibly truncate them.. so write AND block!
1913          */
1914         if (ci->i_wrbuffer_ref_head < ci->i_wrbuffer_ref) {
1915                 spin_unlock(&ci->i_ceph_lock);
1916                 dout("__do_pending_vmtruncate %p flushing snaps first\n",
1917                      inode);
1918                 filemap_write_and_wait_range(&inode->i_data, 0,
1919                                              inode->i_sb->s_maxbytes);
1920                 goto retry;
1921         }
1922
1923         /* there should be no reader or writer */
1924         WARN_ON_ONCE(ci->i_rd_ref || ci->i_wr_ref);
1925
1926         to = ci->i_truncate_size;
1927         wrbuffer_refs = ci->i_wrbuffer_ref;
1928         dout("__do_pending_vmtruncate %p (%d) to %lld\n", inode,
1929              ci->i_truncate_pending, to);
1930         spin_unlock(&ci->i_ceph_lock);
1931
1932         truncate_pagecache(inode, to);
1933
1934         spin_lock(&ci->i_ceph_lock);
1935         if (to == ci->i_truncate_size) {
1936                 ci->i_truncate_pending = 0;
1937                 finish = 1;
1938         }
1939         spin_unlock(&ci->i_ceph_lock);
1940         if (!finish)
1941                 goto retry;
1942
1943         mutex_unlock(&ci->i_truncate_mutex);
1944
1945         if (wrbuffer_refs == 0)
1946                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
1947
1948         wake_up_all(&ci->i_cap_wq);
1949 }
1950
1951 /*
1952  * symlinks
1953  */
1954 static const struct inode_operations ceph_symlink_iops = {
1955         .get_link = simple_get_link,
1956         .setattr = ceph_setattr,
1957         .getattr = ceph_getattr,
1958         .listxattr = ceph_listxattr,
1959 };
1960
1961 int __ceph_setattr(struct inode *inode, struct iattr *attr)
1962 {
1963         struct ceph_inode_info *ci = ceph_inode(inode);
1964         const unsigned int ia_valid = attr->ia_valid;
1965         struct ceph_mds_request *req;
1966         struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1967         struct ceph_cap_flush *prealloc_cf;
1968         int issued;
1969         int release = 0, dirtied = 0;
1970         int mask = 0;
1971         int err = 0;
1972         int inode_dirty_flags = 0;
1973         bool lock_snap_rwsem = false;
1974
1975         prealloc_cf = ceph_alloc_cap_flush();
1976         if (!prealloc_cf)
1977                 return -ENOMEM;
1978
1979         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SETATTR,
1980                                        USE_AUTH_MDS);
1981         if (IS_ERR(req)) {
1982                 ceph_free_cap_flush(prealloc_cf);
1983                 return PTR_ERR(req);
1984         }
1985
1986         spin_lock(&ci->i_ceph_lock);
1987         issued = __ceph_caps_issued(ci, NULL);
1988
1989         if (!ci->i_head_snapc &&
1990             (issued & (CEPH_CAP_ANY_EXCL | CEPH_CAP_FILE_WR))) {
1991                 lock_snap_rwsem = true;
1992                 if (!down_read_trylock(&mdsc->snap_rwsem)) {
1993                         spin_unlock(&ci->i_ceph_lock);
1994                         down_read(&mdsc->snap_rwsem);
1995                         spin_lock(&ci->i_ceph_lock);
1996                         issued = __ceph_caps_issued(ci, NULL);
1997                 }
1998         }
1999
2000         dout("setattr %p issued %s\n", inode, ceph_cap_string(issued));
2001
2002         if (ia_valid & ATTR_UID) {
2003                 dout("setattr %p uid %d -> %d\n", inode,
2004                      from_kuid(&init_user_ns, inode->i_uid),
2005                      from_kuid(&init_user_ns, attr->ia_uid));
2006                 if (issued & CEPH_CAP_AUTH_EXCL) {
2007                         inode->i_uid = attr->ia_uid;
2008                         dirtied |= CEPH_CAP_AUTH_EXCL;
2009                 } else if ((issued & CEPH_CAP_AUTH_SHARED) == 0 ||
2010                            !uid_eq(attr->ia_uid, inode->i_uid)) {
2011                         req->r_args.setattr.uid = cpu_to_le32(
2012                                 from_kuid(&init_user_ns, attr->ia_uid));
2013                         mask |= CEPH_SETATTR_UID;
2014                         release |= CEPH_CAP_AUTH_SHARED;
2015                 }
2016         }
2017         if (ia_valid & ATTR_GID) {
2018                 dout("setattr %p gid %d -> %d\n", inode,
2019                      from_kgid(&init_user_ns, inode->i_gid),
2020                      from_kgid(&init_user_ns, attr->ia_gid));
2021                 if (issued & CEPH_CAP_AUTH_EXCL) {
2022                         inode->i_gid = attr->ia_gid;
2023                         dirtied |= CEPH_CAP_AUTH_EXCL;
2024                 } else if ((issued & CEPH_CAP_AUTH_SHARED) == 0 ||
2025                            !gid_eq(attr->ia_gid, inode->i_gid)) {
2026                         req->r_args.setattr.gid = cpu_to_le32(
2027                                 from_kgid(&init_user_ns, attr->ia_gid));
2028                         mask |= CEPH_SETATTR_GID;
2029                         release |= CEPH_CAP_AUTH_SHARED;
2030                 }
2031         }
2032         if (ia_valid & ATTR_MODE) {
2033                 dout("setattr %p mode 0%o -> 0%o\n", inode, inode->i_mode,
2034                      attr->ia_mode);
2035                 if (issued & CEPH_CAP_AUTH_EXCL) {
2036                         inode->i_mode = attr->ia_mode;
2037                         dirtied |= CEPH_CAP_AUTH_EXCL;
2038                 } else if ((issued & CEPH_CAP_AUTH_SHARED) == 0 ||
2039                            attr->ia_mode != inode->i_mode) {
2040                         inode->i_mode = attr->ia_mode;
2041                         req->r_args.setattr.mode = cpu_to_le32(attr->ia_mode);
2042                         mask |= CEPH_SETATTR_MODE;
2043                         release |= CEPH_CAP_AUTH_SHARED;
2044                 }
2045         }
2046
2047         if (ia_valid & ATTR_ATIME) {
2048                 dout("setattr %p atime %lld.%ld -> %lld.%ld\n", inode,
2049                      inode->i_atime.tv_sec, inode->i_atime.tv_nsec,
2050                      attr->ia_atime.tv_sec, attr->ia_atime.tv_nsec);
2051                 if (issued & CEPH_CAP_FILE_EXCL) {
2052                         ci->i_time_warp_seq++;
2053                         inode->i_atime = attr->ia_atime;
2054                         dirtied |= CEPH_CAP_FILE_EXCL;
2055                 } else if ((issued & CEPH_CAP_FILE_WR) &&
2056                            timespec64_compare(&inode->i_atime,
2057                                             &attr->ia_atime) < 0) {
2058                         inode->i_atime = attr->ia_atime;
2059                         dirtied |= CEPH_CAP_FILE_WR;
2060                 } else if ((issued & CEPH_CAP_FILE_SHARED) == 0 ||
2061                            !timespec64_equal(&inode->i_atime, &attr->ia_atime)) {
2062                         ceph_encode_timespec64(&req->r_args.setattr.atime,
2063                                                &attr->ia_atime);
2064                         mask |= CEPH_SETATTR_ATIME;
2065                         release |= CEPH_CAP_FILE_SHARED |
2066                                    CEPH_CAP_FILE_RD | CEPH_CAP_FILE_WR;
2067                 }
2068         }
2069         if (ia_valid & ATTR_MTIME) {
2070                 dout("setattr %p mtime %lld.%ld -> %lld.%ld\n", inode,
2071                      inode->i_mtime.tv_sec, inode->i_mtime.tv_nsec,
2072                      attr->ia_mtime.tv_sec, attr->ia_mtime.tv_nsec);
2073                 if (issued & CEPH_CAP_FILE_EXCL) {
2074                         ci->i_time_warp_seq++;
2075                         inode->i_mtime = attr->ia_mtime;
2076                         dirtied |= CEPH_CAP_FILE_EXCL;
2077                 } else if ((issued & CEPH_CAP_FILE_WR) &&
2078                            timespec64_compare(&inode->i_mtime,
2079                                             &attr->ia_mtime) < 0) {
2080                         inode->i_mtime = attr->ia_mtime;
2081                         dirtied |= CEPH_CAP_FILE_WR;
2082                 } else if ((issued & CEPH_CAP_FILE_SHARED) == 0 ||
2083                            !timespec64_equal(&inode->i_mtime, &attr->ia_mtime)) {
2084                         ceph_encode_timespec64(&req->r_args.setattr.mtime,
2085                                                &attr->ia_mtime);
2086                         mask |= CEPH_SETATTR_MTIME;
2087                         release |= CEPH_CAP_FILE_SHARED |
2088                                    CEPH_CAP_FILE_RD | CEPH_CAP_FILE_WR;
2089                 }
2090         }
2091         if (ia_valid & ATTR_SIZE) {
2092                 dout("setattr %p size %lld -> %lld\n", inode,
2093                      inode->i_size, attr->ia_size);
2094                 if ((issued & CEPH_CAP_FILE_EXCL) &&
2095                     attr->ia_size > inode->i_size) {
2096                         i_size_write(inode, attr->ia_size);
2097                         inode->i_blocks = calc_inode_blocks(attr->ia_size);
2098                         ci->i_reported_size = attr->ia_size;
2099                         dirtied |= CEPH_CAP_FILE_EXCL;
2100                 } else if ((issued & CEPH_CAP_FILE_SHARED) == 0 ||
2101                            attr->ia_size != inode->i_size) {
2102                         req->r_args.setattr.size = cpu_to_le64(attr->ia_size);
2103                         req->r_args.setattr.old_size =
2104                                 cpu_to_le64(inode->i_size);
2105                         mask |= CEPH_SETATTR_SIZE;
2106                         release |= CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2107                                    CEPH_CAP_FILE_RD | CEPH_CAP_FILE_WR;
2108                 }
2109         }
2110
2111         /* these do nothing */
2112         if (ia_valid & ATTR_CTIME) {
2113                 bool only = (ia_valid & (ATTR_SIZE|ATTR_MTIME|ATTR_ATIME|
2114                                          ATTR_MODE|ATTR_UID|ATTR_GID)) == 0;
2115                 dout("setattr %p ctime %lld.%ld -> %lld.%ld (%s)\n", inode,
2116                      inode->i_ctime.tv_sec, inode->i_ctime.tv_nsec,
2117                      attr->ia_ctime.tv_sec, attr->ia_ctime.tv_nsec,
2118                      only ? "ctime only" : "ignored");
2119                 if (only) {
2120                         /*
2121                          * if kernel wants to dirty ctime but nothing else,
2122                          * we need to choose a cap to dirty under, or do
2123                          * a almost-no-op setattr
2124                          */
2125                         if (issued & CEPH_CAP_AUTH_EXCL)
2126                                 dirtied |= CEPH_CAP_AUTH_EXCL;
2127                         else if (issued & CEPH_CAP_FILE_EXCL)
2128                                 dirtied |= CEPH_CAP_FILE_EXCL;
2129                         else if (issued & CEPH_CAP_XATTR_EXCL)
2130                                 dirtied |= CEPH_CAP_XATTR_EXCL;
2131                         else
2132                                 mask |= CEPH_SETATTR_CTIME;
2133                 }
2134         }
2135         if (ia_valid & ATTR_FILE)
2136                 dout("setattr %p ATTR_FILE ... hrm!\n", inode);
2137
2138         if (dirtied) {
2139                 inode_dirty_flags = __ceph_mark_dirty_caps(ci, dirtied,
2140                                                            &prealloc_cf);
2141                 inode->i_ctime = attr->ia_ctime;
2142         }
2143
2144         release &= issued;
2145         spin_unlock(&ci->i_ceph_lock);
2146         if (lock_snap_rwsem)
2147                 up_read(&mdsc->snap_rwsem);
2148
2149         if (inode_dirty_flags)
2150                 __mark_inode_dirty(inode, inode_dirty_flags);
2151
2152
2153         if (mask) {
2154                 req->r_inode = inode;
2155                 ihold(inode);
2156                 req->r_inode_drop = release;
2157                 req->r_args.setattr.mask = cpu_to_le32(mask);
2158                 req->r_num_caps = 1;
2159                 req->r_stamp = attr->ia_ctime;
2160                 err = ceph_mdsc_do_request(mdsc, NULL, req);
2161         }
2162         dout("setattr %p result=%d (%s locally, %d remote)\n", inode, err,
2163              ceph_cap_string(dirtied), mask);
2164
2165         ceph_mdsc_put_request(req);
2166         ceph_free_cap_flush(prealloc_cf);
2167
2168         if (err >= 0 && (mask & CEPH_SETATTR_SIZE))
2169                 __ceph_do_pending_vmtruncate(inode);
2170
2171         return err;
2172 }
2173
2174 /*
2175  * setattr
2176  */
2177 int ceph_setattr(struct dentry *dentry, struct iattr *attr)
2178 {
2179         struct inode *inode = d_inode(dentry);
2180         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2181         int err;
2182
2183         if (ceph_snap(inode) != CEPH_NOSNAP)
2184                 return -EROFS;
2185
2186         err = setattr_prepare(dentry, attr);
2187         if (err != 0)
2188                 return err;
2189
2190         if ((attr->ia_valid & ATTR_SIZE) &&
2191             attr->ia_size > max(inode->i_size, fsc->max_file_size))
2192                 return -EFBIG;
2193
2194         if ((attr->ia_valid & ATTR_SIZE) &&
2195             ceph_quota_is_max_bytes_exceeded(inode, attr->ia_size))
2196                 return -EDQUOT;
2197
2198         err = __ceph_setattr(inode, attr);
2199
2200         if (err >= 0 && (attr->ia_valid & ATTR_MODE))
2201                 err = posix_acl_chmod(inode, attr->ia_mode);
2202
2203         return err;
2204 }
2205
2206 /*
2207  * Verify that we have a lease on the given mask.  If not,
2208  * do a getattr against an mds.
2209  */
2210 int __ceph_do_getattr(struct inode *inode, struct page *locked_page,
2211                       int mask, bool force)
2212 {
2213         struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
2214         struct ceph_mds_client *mdsc = fsc->mdsc;
2215         struct ceph_mds_request *req;
2216         int mode;
2217         int err;
2218
2219         if (ceph_snap(inode) == CEPH_SNAPDIR) {
2220                 dout("do_getattr inode %p SNAPDIR\n", inode);
2221                 return 0;
2222         }
2223
2224         dout("do_getattr inode %p mask %s mode 0%o\n",
2225              inode, ceph_cap_string(mask), inode->i_mode);
2226         if (!force && ceph_caps_issued_mask(ceph_inode(inode), mask, 1))
2227                 return 0;
2228
2229         mode = (mask & CEPH_STAT_RSTAT) ? USE_AUTH_MDS : USE_ANY_MDS;
2230         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, mode);
2231         if (IS_ERR(req))
2232                 return PTR_ERR(req);
2233         req->r_inode = inode;
2234         ihold(inode);
2235         req->r_num_caps = 1;
2236         req->r_args.getattr.mask = cpu_to_le32(mask);
2237         req->r_locked_page = locked_page;
2238         err = ceph_mdsc_do_request(mdsc, NULL, req);
2239         if (locked_page && err == 0) {
2240                 u64 inline_version = req->r_reply_info.targeti.inline_version;
2241                 if (inline_version == 0) {
2242                         /* the reply is supposed to contain inline data */
2243                         err = -EINVAL;
2244                 } else if (inline_version == CEPH_INLINE_NONE) {
2245                         err = -ENODATA;
2246                 } else {
2247                         err = req->r_reply_info.targeti.inline_len;
2248                 }
2249         }
2250         ceph_mdsc_put_request(req);
2251         dout("do_getattr result=%d\n", err);
2252         return err;
2253 }
2254
2255
2256 /*
2257  * Check inode permissions.  We verify we have a valid value for
2258  * the AUTH cap, then call the generic handler.
2259  */
2260 int ceph_permission(struct inode *inode, int mask)
2261 {
2262         int err;
2263
2264         if (mask & MAY_NOT_BLOCK)
2265                 return -ECHILD;
2266
2267         err = ceph_do_getattr(inode, CEPH_CAP_AUTH_SHARED, false);
2268
2269         if (!err)
2270                 err = generic_permission(inode, mask);
2271         return err;
2272 }
2273
2274 /*
2275  * Get all attributes.  Hopefully somedata we'll have a statlite()
2276  * and can limit the fields we require to be accurate.
2277  */
2278 int ceph_getattr(const struct path *path, struct kstat *stat,
2279                  u32 request_mask, unsigned int flags)
2280 {
2281         struct inode *inode = d_inode(path->dentry);
2282         struct ceph_inode_info *ci = ceph_inode(inode);
2283         int err;
2284
2285         err = ceph_do_getattr(inode, CEPH_STAT_CAP_INODE_ALL, false);
2286         if (!err) {
2287                 generic_fillattr(inode, stat);
2288                 stat->ino = ceph_translate_ino(inode->i_sb, inode->i_ino);
2289                 if (ceph_snap(inode) != CEPH_NOSNAP)
2290                         stat->dev = ceph_snap(inode);
2291                 else
2292                         stat->dev = 0;
2293                 if (S_ISDIR(inode->i_mode)) {
2294                         if (ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb),
2295                                                 RBYTES))
2296                                 stat->size = ci->i_rbytes;
2297                         else
2298                                 stat->size = ci->i_files + ci->i_subdirs;
2299                         stat->blocks = 0;
2300                         stat->blksize = 65536;
2301                         /*
2302                          * Some applications rely on the number of st_nlink
2303                          * value on directories to be either 0 (if unlinked)
2304                          * or 2 + number of subdirectories.
2305                          */
2306                         if (stat->nlink == 1)
2307                                 /* '.' + '..' + subdirs */
2308                                 stat->nlink = 1 + 1 + ci->i_subdirs;
2309                 }
2310         }
2311         return err;
2312 }