fs/ceph/snap.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/sort.h>
   4 #include <linux/slab.h>
   5
   6 #include "super.h"
   7 #include "mds_client.h"
   8
   9 #include <linux/ceph/decode.h>
  10
  11 /*
  12  * Snapshots in ceph are driven in large part by cooperation from the
  13  * client.  In contrast to local file systems or file servers that
  14  * implement snapshots at a single point in the system, ceph's
  15  * distributed access to storage requires clients to help decide
  16  * whether a write logically occurs before or after a recently created
  17  * snapshot.
  18  *
  19  * This provides a perfect instantanous client-wide snapshot.  Between
  20  * clients, however, snapshots may appear to be applied at slightly
  21  * different points in time, depending on delays in delivering the
  22  * snapshot notification.
  23  *
  24  * Snapshots are _not_ file system-wide.  Instead, each snapshot
  25  * applies to the subdirectory nested beneath some directory.  This
  26  * effectively divides the hierarchy into multiple "realms," where all
  27  * of the files contained by each realm share the same set of
  28  * snapshots.  An individual realm's snap set contains snapshots
  29  * explicitly created on that realm, as well as any snaps in its
  30  * parent's snap set _after_ the point at which the parent became it's
  31  * parent (due to, say, a rename).  Similarly, snaps from prior parents
  32  * during the time intervals during which they were the parent are included.
  33  *
  34  * The client is spared most of this detail, fortunately... it must only
  35  * maintains a hierarchy of realms reflecting the current parent/child
  36  * realm relationship, and for each realm has an explicit list of snaps
  37  * inherited from prior parents.
  38  *
  39  * A snap_realm struct is maintained for realms containing every inode
  40  * with an open cap in the system.  (The needed snap realm information is
  41  * provided by the MDS whenever a cap is issued, i.e., on open.)  A 'seq'
  42  * version number is used to ensure that as realm parameters change (new
  43  * snapshot, new parent, etc.) the client's realm hierarchy is updated.
  44  *
  45  * The realm hierarchy drives the generation of a 'snap context' for each
  46  * realm, which simply lists the resulting set of snaps for the realm.  This
  47  * is attached to any writes sent to OSDs.
  48  */
  49 /*
  50  * Unfortunately error handling is a bit mixed here.  If we get a snap
  51  * update, but don't have enough memory to update our realm hierarchy,
  52  * it's not clear what we can do about it (besides complaining to the
  53  * console).
  54  */
  55
  56
  57 /*
  58  * increase ref count for the realm
  59  *
  60  * caller must hold snap_rwsem for write.
  61  */
  62 void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
  63                          struct ceph_snap_realm *realm)
  64 {
  65         dout("get_realm %p %d -> %d\n", realm,
  66              atomic_read(&realm->nref), atomic_read(&realm->nref)+1);
  67         /*
  68          * since we _only_ increment realm refs or empty the empty
  69          * list with snap_rwsem held, adjusting the empty list here is
  70          * safe.  we do need to protect against concurrent empty list
  71          * additions, however.
  72          */
  73         if (atomic_read(&realm->nref) == 0) {
  74                 spin_lock(&mdsc->snap_empty_lock);
  75                 list_del_init(&realm->empty_item);
  76                 spin_unlock(&mdsc->snap_empty_lock);
  77         }
  78
  79         atomic_inc(&realm->nref);
  80 }
  81
  82 static void __insert_snap_realm(struct rb_root *root,
  83                                 struct ceph_snap_realm *new)
  84 {
  85         struct rb_node **p = &root->rb_node;
  86         struct rb_node *parent = NULL;
  87         struct ceph_snap_realm *r = NULL;
  88
  89         while (*p) {
  90                 parent = *p;
  91                 r = rb_entry(parent, struct ceph_snap_realm, node);
  92                 if (new->ino < r->ino)
  93                         p = &(*p)->rb_left;
  94                 else if (new->ino > r->ino)
  95                         p = &(*p)->rb_right;
  96                 else
  97                         BUG();
  98         }
  99
 100         rb_link_node(&new->node, parent, p);
 101         rb_insert_color(&new->node, root);
 102 }
 103
 104 /*
 105  * create and get the realm rooted at @ino and bump its ref count.
 106  *
 107  * caller must hold snap_rwsem for write.
 108  */
 109 static struct ceph_snap_realm *ceph_create_snap_realm(
 110         struct ceph_mds_client *mdsc,
 111         u64 ino)
 112 {
 113         struct ceph_snap_realm *realm;
 114
 115         realm = kzalloc(sizeof(*realm), GFP_NOFS);
 116         if (!realm)
 117                 return ERR_PTR(-ENOMEM);
 118
 119         atomic_set(&realm->nref, 0);    /* tree does not take a ref */
 120         realm->ino = ino;
 121         INIT_LIST_HEAD(&realm->children);
 122         INIT_LIST_HEAD(&realm->child_item);
 123         INIT_LIST_HEAD(&realm->empty_item);
 124         INIT_LIST_HEAD(&realm->dirty_item);
 125         INIT_LIST_HEAD(&realm->inodes_with_caps);
 126         spin_lock_init(&realm->inodes_with_caps_lock);
 127         __insert_snap_realm(&mdsc->snap_realms, realm);
 128         dout("create_snap_realm %llx %p\n", realm->ino, realm);
 129         return realm;
 130 }
 131
 132 /*
 133  * lookup the realm rooted at @ino.
 134  *
 135  * caller must hold snap_rwsem for write.
 136  */
 137 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
 138                                                u64 ino)
 139 {
 140         struct rb_node *n = mdsc->snap_realms.rb_node;
 141         struct ceph_snap_realm *r;
 142
 143         while (n) {
 144                 r = rb_entry(n, struct ceph_snap_realm, node);
 145                 if (ino < r->ino)
 146                         n = n->rb_left;
 147                 else if (ino > r->ino)
 148                         n = n->rb_right;
 149                 else {
 150                         dout("lookup_snap_realm %llx %p\n", r->ino, r);
 151                         return r;
 152                 }
 153         }
 154         return NULL;
 155 }
 156
 157 static void __put_snap_realm(struct ceph_mds_client *mdsc,
 158                              struct ceph_snap_realm *realm);
 159
 160 /*
 161  * called with snap_rwsem (write)
 162  */
 163 static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
 164                                  struct ceph_snap_realm *realm)
 165 {
 166         dout("__destroy_snap_realm %p %llx\n", realm, realm->ino);
 167
 168         rb_erase(&realm->node, &mdsc->snap_realms);
 169
 170         if (realm->parent) {
 171                 list_del_init(&realm->child_item);
 172                 __put_snap_realm(mdsc, realm->parent);
 173         }
 174
 175         kfree(realm->prior_parent_snaps);
 176         kfree(realm->snaps);
 177         ceph_put_snap_context(realm->cached_context);
 178         kfree(realm);
 179 }
 180
 181 /*
 182  * caller holds snap_rwsem (write)
 183  */
 184 static void __put_snap_realm(struct ceph_mds_client *mdsc,
 185                              struct ceph_snap_realm *realm)
 186 {
 187         dout("__put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
 188              atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
 189         if (atomic_dec_and_test(&realm->nref))
 190                 __destroy_snap_realm(mdsc, realm);
 191 }
 192
 193 /*
 194  * caller needn't hold any locks
 195  */
 196 void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
 197                          struct ceph_snap_realm *realm)
 198 {
 199         dout("put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
 200              atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
 201         if (!atomic_dec_and_test(&realm->nref))
 202                 return;
 203
 204         if (down_write_trylock(&mdsc->snap_rwsem)) {
 205                 __destroy_snap_realm(mdsc, realm);
 206                 up_write(&mdsc->snap_rwsem);
 207         } else {
 208                 spin_lock(&mdsc->snap_empty_lock);
 209                 list_add(&realm->empty_item, &mdsc->snap_empty);
 210                 spin_unlock(&mdsc->snap_empty_lock);
 211         }
 212 }
 213
 214 /*
 215  * Clean up any realms whose ref counts have dropped to zero.  Note
 216  * that this does not include realms who were created but not yet
 217  * used.
 218  *
 219  * Called under snap_rwsem (write)
 220  */
 221 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
 222 {
 223         struct ceph_snap_realm *realm;
 224
 225         spin_lock(&mdsc->snap_empty_lock);
 226         while (!list_empty(&mdsc->snap_empty)) {
 227                 realm = list_first_entry(&mdsc->snap_empty,
 228                                    struct ceph_snap_realm, empty_item);
 229                 list_del(&realm->empty_item);
 230                 spin_unlock(&mdsc->snap_empty_lock);
 231                 __destroy_snap_realm(mdsc, realm);
 232                 spin_lock(&mdsc->snap_empty_lock);
 233         }
 234         spin_unlock(&mdsc->snap_empty_lock);
 235 }
 236
 237 void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc)
 238 {
 239         down_write(&mdsc->snap_rwsem);
 240         __cleanup_empty_realms(mdsc);
 241         up_write(&mdsc->snap_rwsem);
 242 }
 243
 244 /*
 245  * adjust the parent realm of a given @realm.  adjust child list, and parent
 246  * pointers, and ref counts appropriately.
 247  *
 248  * return true if parent was changed, 0 if unchanged, <0 on error.
 249  *
 250  * caller must hold snap_rwsem for write.
 251  */
 252 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
 253                                     struct ceph_snap_realm *realm,
 254                                     u64 parentino)
 255 {
 256         struct ceph_snap_realm *parent;
 257
 258         if (realm->parent_ino == parentino)
 259                 return 0;
 260
 261         parent = ceph_lookup_snap_realm(mdsc, parentino);
 262         if (!parent) {
 263                 parent = ceph_create_snap_realm(mdsc, parentino);
 264                 if (IS_ERR(parent))
 265                         return PTR_ERR(parent);
 266         }
 267         dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n",
 268              realm->ino, realm, realm->parent_ino, realm->parent,
 269              parentino, parent);
 270         if (realm->parent) {
 271                 list_del_init(&realm->child_item);
 272                 ceph_put_snap_realm(mdsc, realm->parent);
 273         }
 274         realm->parent_ino = parentino;
 275         realm->parent = parent;
 276         ceph_get_snap_realm(mdsc, parent);
 277         list_add(&realm->child_item, &parent->children);
 278         return 1;
 279 }
 280
 281
 282 static int cmpu64_rev(const void *a, const void *b)
 283 {
 284         if (*(u64 *)a < *(u64 *)b)
 285                 return 1;
 286         if (*(u64 *)a > *(u64 *)b)
 287                 return -1;
 288         return 0;
 289 }
 290
 291
 292 static struct ceph_snap_context *empty_snapc;
 293
 294 /*
 295  * build the snap context for a given realm.
 296  */
 297 static int build_snap_context(struct ceph_snap_realm *realm)
 298 {
 299         struct ceph_snap_realm *parent = realm->parent;
 300         struct ceph_snap_context *snapc;
 301         int err = 0;
 302         u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
 303
 304         /*
 305          * build parent context, if it hasn't been built.
 306          * conservatively estimate that all parent snaps might be
 307          * included by us.
 308          */
 309         if (parent) {
 310                 if (!parent->cached_context) {
 311                         err = build_snap_context(parent);
 312                         if (err)
 313                                 goto fail;
 314                 }
 315                 num += parent->cached_context->num_snaps;
 316         }
 317
 318         /* do i actually need to update?  not if my context seq
 319            matches realm seq, and my parents' does to.  (this works
 320            because we rebuild_snap_realms() works _downward_ in
 321            hierarchy after each update.) */
 322         if (realm->cached_context &&
 323             realm->cached_context->seq == realm->seq &&
 324             (!parent ||
 325              realm->cached_context->seq >= parent->cached_context->seq)) {
 326                 dout("build_snap_context %llx %p: %p seq %lld (%u snaps)"
 327                      " (unchanged)\n",
 328                      realm->ino, realm, realm->cached_context,
 329                      realm->cached_context->seq,
 330                      (unsigned int) realm->cached_context->num_snaps);
 331                 return 0;
 332         }
 333
 334         if (num == 0 && realm->seq == empty_snapc->seq) {
 335                 ceph_get_snap_context(empty_snapc);
 336                 snapc = empty_snapc;
 337                 goto done;
 338         }
 339
 340         /* alloc new snap context */
 341         err = -ENOMEM;
 342         if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
 343                 goto fail;
 344         snapc = ceph_create_snap_context(num, GFP_NOFS);
 345         if (!snapc)
 346                 goto fail;
 347
 348         /* build (reverse sorted) snap vector */
 349         num = 0;
 350         snapc->seq = realm->seq;
 351         if (parent) {
 352                 u32 i;
 353
 354                 /* include any of parent's snaps occurring _after_ my
 355                    parent became my parent */
 356                 for (i = 0; i < parent->cached_context->num_snaps; i++)
 357                         if (parent->cached_context->snaps[i] >=
 358                             realm->parent_since)
 359                                 snapc->snaps[num++] =
 360                                         parent->cached_context->snaps[i];
 361                 if (parent->cached_context->seq > snapc->seq)
 362                         snapc->seq = parent->cached_context->seq;
 363         }
 364         memcpy(snapc->snaps + num, realm->snaps,
 365                sizeof(u64)*realm->num_snaps);
 366         num += realm->num_snaps;
 367         memcpy(snapc->snaps + num, realm->prior_parent_snaps,
 368                sizeof(u64)*realm->num_prior_parent_snaps);
 369         num += realm->num_prior_parent_snaps;
 370
 371         sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
 372         snapc->num_snaps = num;
 373         dout("build_snap_context %llx %p: %p seq %lld (%u snaps)\n",
 374              realm->ino, realm, snapc, snapc->seq,
 375              (unsigned int) snapc->num_snaps);
 376
 377 done:
 378         ceph_put_snap_context(realm->cached_context);
 379         realm->cached_context = snapc;
 380         return 0;
 381
 382 fail:
 383         /*
 384          * if we fail, clear old (incorrect) cached_context... hopefully
 385          * we'll have better luck building it later
 386          */
 387         if (realm->cached_context) {
 388                 ceph_put_snap_context(realm->cached_context);
 389                 realm->cached_context = NULL;
 390         }
 391         pr_err("build_snap_context %llx %p fail %d\n", realm->ino,
 392                realm, err);
 393         return err;
 394 }
 395
 396 /*
 397  * rebuild snap context for the given realm and all of its children.
 398  */
 399 static void rebuild_snap_realms(struct ceph_snap_realm *realm)
 400 {
 401         struct ceph_snap_realm *child;
 402
 403         dout("rebuild_snap_realms %llx %p\n", realm->ino, realm);
 404         build_snap_context(realm);
 405
 406         list_for_each_entry(child, &realm->children, child_item)
 407                 rebuild_snap_realms(child);
 408 }
 409
 410
 411 /*
 412  * helper to allocate and decode an array of snapids.  free prior
 413  * instance, if any.
 414  */
 415 static int dup_array(u64 **dst, __le64 *src, u32 num)
 416 {
 417         u32 i;
 418
 419         kfree(*dst);
 420         if (num) {
 421                 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
 422                 if (!*dst)
 423                         return -ENOMEM;
 424                 for (i = 0; i < num; i++)
 425                         (*dst)[i] = get_unaligned_le64(src + i);
 426         } else {
 427                 *dst = NULL;
 428         }
 429         return 0;
 430 }
 431
 432
 433 /*
 434  * When a snapshot is applied, the size/mtime inode metadata is queued
 435  * in a ceph_cap_snap (one for each snapshot) until writeback
 436  * completes and the metadata can be flushed back to the MDS.
 437  *
 438  * However, if a (sync) write is currently in-progress when we apply
 439  * the snapshot, we have to wait until the write succeeds or fails
 440  * (and a final size/mtime is known).  In this case the
 441  * cap_snap->writing = 1, and is said to be "pending."  When the write
 442  * finishes, we __ceph_finish_cap_snap().
 443  *
 444  * Caller must hold snap_rwsem for read (i.e., the realm topology won't
 445  * change).
 446  */
 447 void ceph_queue_cap_snap(struct ceph_inode_info *ci)
 448 {
 449         struct inode *inode = &ci->vfs_inode;
 450         struct ceph_cap_snap *capsnap;
 451         int used, dirty;
 452
 453         capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
 454         if (!capsnap) {
 455                 pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode);
 456                 return;
 457         }
 458
 459         spin_lock(&ci->i_ceph_lock);
 460         used = __ceph_caps_used(ci);
 461         dirty = __ceph_caps_dirty(ci);
 462
 463         /*
 464          * If there is a write in progress, treat that as a dirty Fw,
 465          * even though it hasn't completed yet; by the time we finish
 466          * up this capsnap it will be.
 467          */
 468         if (used & CEPH_CAP_FILE_WR)
 469                 dirty |= CEPH_CAP_FILE_WR;
 470
 471         if (__ceph_have_pending_cap_snap(ci)) {
 472                 /* there is no point in queuing multiple "pending" cap_snaps,
 473                    as no new writes are allowed to start when pending, so any
 474                    writes in progress now were started before the previous
 475                    cap_snap.  lucky us. */
 476                 dout("queue_cap_snap %p already pending\n", inode);
 477                 kfree(capsnap);
 478         } else if (ci->i_snap_realm->cached_context == empty_snapc) {
 479                 dout("queue_cap_snap %p empty snapc\n", inode);
 480                 kfree(capsnap);
 481         } else if (dirty & (CEPH_CAP_AUTH_EXCL|CEPH_CAP_XATTR_EXCL|
 482                             CEPH_CAP_FILE_EXCL|CEPH_CAP_FILE_WR)) {
 483                 struct ceph_snap_context *snapc = ci->i_head_snapc;
 484
 485                 /*
 486                  * if we are a sync write, we may need to go to the snaprealm
 487                  * to get the current snapc.
 488                  */
 489                 if (!snapc)
 490                         snapc = ci->i_snap_realm->cached_context;
 491
 492                 dout("queue_cap_snap %p cap_snap %p queuing under %p %s\n",
 493                      inode, capsnap, snapc, ceph_cap_string(dirty));
 494                 ihold(inode);
 495
 496                 atomic_set(&capsnap->nref, 1);
 497                 capsnap->ci = ci;
 498                 INIT_LIST_HEAD(&capsnap->ci_item);
 499                 INIT_LIST_HEAD(&capsnap->flushing_item);
 500
 501                 capsnap->follows = snapc->seq;
 502                 capsnap->issued = __ceph_caps_issued(ci, NULL);
 503                 capsnap->dirty = dirty;
 504
 505                 capsnap->mode = inode->i_mode;
 506                 capsnap->uid = inode->i_uid;
 507                 capsnap->gid = inode->i_gid;
 508
 509                 if (dirty & CEPH_CAP_XATTR_EXCL) {
 510                         __ceph_build_xattrs_blob(ci);
 511                         capsnap->xattr_blob =
 512                                 ceph_buffer_get(ci->i_xattrs.blob);
 513                         capsnap->xattr_version = ci->i_xattrs.version;
 514                 } else {
 515                         capsnap->xattr_blob = NULL;
 516                         capsnap->xattr_version = 0;
 517                 }
 518
 519                 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
 520
 521                 /* dirty page count moved from _head to this cap_snap;
 522                    all subsequent writes page dirties occur _after_ this
 523                    snapshot. */
 524                 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
 525                 ci->i_wrbuffer_ref_head = 0;
 526                 capsnap->context = snapc;
 527                 ci->i_head_snapc =
 528                         ceph_get_snap_context(ci->i_snap_realm->cached_context);
 529                 dout(" new snapc is %p\n", ci->i_head_snapc);
 530                 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
 531
 532                 if (used & CEPH_CAP_FILE_WR) {
 533                         dout("queue_cap_snap %p cap_snap %p snapc %p"
 534                              " seq %llu used WR, now pending\n", inode,
 535                              capsnap, snapc, snapc->seq);
 536                         capsnap->writing = 1;
 537                 } else {
 538                         /* note mtime, size NOW. */
 539                         __ceph_finish_cap_snap(ci, capsnap);
 540                 }
 541         } else {
 542                 dout("queue_cap_snap %p nothing dirty|writing\n", inode);
 543                 kfree(capsnap);
 544         }
 545
 546         spin_unlock(&ci->i_ceph_lock);
 547 }
 548
 549 /*
 550  * Finalize the size, mtime for a cap_snap.. that is, settle on final values
 551  * to be used for the snapshot, to be flushed back to the mds.
 552  *
 553  * If capsnap can now be flushed, add to snap_flush list, and return 1.
 554  *
 555  * Caller must hold i_ceph_lock.
 556  */
 557 int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
 558                             struct ceph_cap_snap *capsnap)
 559 {
 560         struct inode *inode = &ci->vfs_inode;
 561         struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
 562
 563         BUG_ON(capsnap->writing);
 564         capsnap->size = inode->i_size;
 565         capsnap->mtime = inode->i_mtime;
 566         capsnap->atime = inode->i_atime;
 567         capsnap->ctime = inode->i_ctime;
 568         capsnap->time_warp_seq = ci->i_time_warp_seq;
 569         if (capsnap->dirty_pages) {
 570                 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu "
 571                      "still has %d dirty pages\n", inode, capsnap,
 572                      capsnap->context, capsnap->context->seq,
 573                      ceph_cap_string(capsnap->dirty), capsnap->size,
 574                      capsnap->dirty_pages);
 575                 return 0;
 576         }
 577         dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n",
 578              inode, capsnap, capsnap->context,
 579              capsnap->context->seq, ceph_cap_string(capsnap->dirty),
 580              capsnap->size);
 581
 582         spin_lock(&mdsc->snap_flush_lock);
 583         list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
 584         spin_unlock(&mdsc->snap_flush_lock);
 585         return 1;  /* caller may want to ceph_flush_snaps */
 586 }
 587
 588 /*
 589  * Queue cap_snaps for snap writeback for this realm and its children.
 590  * Called under snap_rwsem, so realm topology won't change.
 591  */
 592 static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
 593 {
 594         struct ceph_inode_info *ci;
 595         struct inode *lastinode = NULL;
 596         struct ceph_snap_realm *child;
 597
 598         dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino);
 599
 600         spin_lock(&realm->inodes_with_caps_lock);
 601         list_for_each_entry(ci, &realm->inodes_with_caps,
 602                             i_snap_realm_item) {
 603                 struct inode *inode = igrab(&ci->vfs_inode);
 604                 if (!inode)
 605                         continue;
 606                 spin_unlock(&realm->inodes_with_caps_lock);
 607                 iput(lastinode);
 608                 lastinode = inode;
 609                 ceph_queue_cap_snap(ci);
 610                 spin_lock(&realm->inodes_with_caps_lock);
 611         }
 612         spin_unlock(&realm->inodes_with_caps_lock);
 613         iput(lastinode);
 614
 615         list_for_each_entry(child, &realm->children, child_item) {
 616                 dout("queue_realm_cap_snaps %p %llx queue child %p %llx\n",
 617                      realm, realm->ino, child, child->ino);
 618                 list_del_init(&child->dirty_item);
 619                 list_add(&child->dirty_item, &realm->dirty_item);
 620         }
 621
 622         list_del_init(&realm->dirty_item);
 623         dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino);
 624 }
 625
 626 /*
 627  * Parse and apply a snapblob "snap trace" from the MDS.  This specifies
 628  * the snap realm parameters from a given realm and all of its ancestors,
 629  * up to the root.
 630  *
 631  * Caller must hold snap_rwsem for write.
 632  */
 633 int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
 634                            void *p, void *e, bool deletion)
 635 {
 636         struct ceph_mds_snap_realm *ri;    /* encoded */
 637         __le64 *snaps;                     /* encoded */
 638         __le64 *prior_parent_snaps;        /* encoded */
 639         struct ceph_snap_realm *realm;
 640         int invalidate = 0;
 641         int err = -ENOMEM;
 642         LIST_HEAD(dirty_realms);
 643
 644         dout("update_snap_trace deletion=%d\n", deletion);
 645 more:
 646         ceph_decode_need(&p, e, sizeof(*ri), bad);
 647         ri = p;
 648         p += sizeof(*ri);
 649         ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
 650                             le32_to_cpu(ri->num_prior_parent_snaps)), bad);
 651         snaps = p;
 652         p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
 653         prior_parent_snaps = p;
 654         p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
 655
 656         realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
 657         if (!realm) {
 658                 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
 659                 if (IS_ERR(realm)) {
 660                         err = PTR_ERR(realm);
 661                         goto fail;
 662                 }
 663         }
 664
 665         /* ensure the parent is correct */
 666         err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
 667         if (err < 0)
 668                 goto fail;
 669         invalidate += err;
 670
 671         if (le64_to_cpu(ri->seq) > realm->seq) {
 672                 dout("update_snap_trace updating %llx %p %lld -> %lld\n",
 673                      realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
 674                 /* update realm parameters, snap lists */
 675                 realm->seq = le64_to_cpu(ri->seq);
 676                 realm->created = le64_to_cpu(ri->created);
 677                 realm->parent_since = le64_to_cpu(ri->parent_since);
 678
 679                 realm->num_snaps = le32_to_cpu(ri->num_snaps);
 680                 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
 681                 if (err < 0)
 682                         goto fail;
 683
 684                 realm->num_prior_parent_snaps =
 685                         le32_to_cpu(ri->num_prior_parent_snaps);
 686                 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
 687                                 realm->num_prior_parent_snaps);
 688                 if (err < 0)
 689                         goto fail;
 690
 691                 /* queue realm for cap_snap creation */
 692                 list_add(&realm->dirty_item, &dirty_realms);
 693
 694                 invalidate = 1;
 695         } else if (!realm->cached_context) {
 696                 dout("update_snap_trace %llx %p seq %lld new\n",
 697                      realm->ino, realm, realm->seq);
 698                 invalidate = 1;
 699         } else {
 700                 dout("update_snap_trace %llx %p seq %lld unchanged\n",
 701                      realm->ino, realm, realm->seq);
 702         }
 703
 704         dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino,
 705              realm, invalidate, p, e);
 706
 707         if (p < e)
 708                 goto more;
 709
 710         /* invalidate when we reach the _end_ (root) of the trace */
 711         if (invalidate)
 712                 rebuild_snap_realms(realm);
 713
 714         /*
 715          * queue cap snaps _after_ we've built the new snap contexts,
 716          * so that i_head_snapc can be set appropriately.
 717          */
 718         while (!list_empty(&dirty_realms)) {
 719                 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
 720                                          dirty_item);
 721                 queue_realm_cap_snaps(realm);
 722         }
 723
 724         __cleanup_empty_realms(mdsc);
 725         return 0;
 726
 727 bad:
 728         err = -EINVAL;
 729 fail:
 730         pr_err("update_snap_trace error %d\n", err);
 731         return err;
 732 }
 733
 734
 735 /*
 736  * Send any cap_snaps that are queued for flush.  Try to carry
 737  * s_mutex across multiple snap flushes to avoid locking overhead.
 738  *
 739  * Caller holds no locks.
 740  */
 741 static void flush_snaps(struct ceph_mds_client *mdsc)
 742 {
 743         struct ceph_inode_info *ci;
 744         struct inode *inode;
 745         struct ceph_mds_session *session = NULL;
 746
 747         dout("flush_snaps\n");
 748         spin_lock(&mdsc->snap_flush_lock);
 749         while (!list_empty(&mdsc->snap_flush_list)) {
 750                 ci = list_first_entry(&mdsc->snap_flush_list,
 751                                 struct ceph_inode_info, i_snap_flush_item);
 752                 inode = &ci->vfs_inode;
 753                 ihold(inode);
 754                 spin_unlock(&mdsc->snap_flush_lock);
 755                 spin_lock(&ci->i_ceph_lock);
 756                 __ceph_flush_snaps(ci, &session, 0);
 757                 spin_unlock(&ci->i_ceph_lock);
 758                 iput(inode);
 759                 spin_lock(&mdsc->snap_flush_lock);
 760         }
 761         spin_unlock(&mdsc->snap_flush_lock);
 762
 763         if (session) {
 764                 mutex_unlock(&session->s_mutex);
 765                 ceph_put_mds_session(session);
 766         }
 767         dout("flush_snaps done\n");
 768 }
 769
 770
 771 /*
 772  * Handle a snap notification from the MDS.
 773  *
 774  * This can take two basic forms: the simplest is just a snap creation
 775  * or deletion notification on an existing realm.  This should update the
 776  * realm and its children.
 777  *
 778  * The more difficult case is realm creation, due to snap creation at a
 779  * new point in the file hierarchy, or due to a rename that moves a file or
 780  * directory into another realm.
 781  */
 782 void ceph_handle_snap(struct ceph_mds_client *mdsc,
 783                       struct ceph_mds_session *session,
 784                       struct ceph_msg *msg)
 785 {
 786         struct super_block *sb = mdsc->fsc->sb;
 787         int mds = session->s_mds;
 788         u64 split;
 789         int op;
 790         int trace_len;
 791         struct ceph_snap_realm *realm = NULL;
 792         void *p = msg->front.iov_base;
 793         void *e = p + msg->front.iov_len;
 794         struct ceph_mds_snap_head *h;
 795         int num_split_inos, num_split_realms;
 796         __le64 *split_inos = NULL, *split_realms = NULL;
 797         int i;
 798         int locked_rwsem = 0;
 799
 800         /* decode */
 801         if (msg->front.iov_len < sizeof(*h))
 802                 goto bad;
 803         h = p;
 804         op = le32_to_cpu(h->op);
 805         split = le64_to_cpu(h->split);   /* non-zero if we are splitting an
 806                                           * existing realm */
 807         num_split_inos = le32_to_cpu(h->num_split_inos);
 808         num_split_realms = le32_to_cpu(h->num_split_realms);
 809         trace_len = le32_to_cpu(h->trace_len);
 810         p += sizeof(*h);
 811
 812         dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds,
 813              ceph_snap_op_name(op), split, trace_len);
 814
 815         mutex_lock(&session->s_mutex);
 816         session->s_seq++;
 817         mutex_unlock(&session->s_mutex);
 818
 819         down_write(&mdsc->snap_rwsem);
 820         locked_rwsem = 1;
 821
 822         if (op == CEPH_SNAP_OP_SPLIT) {
 823                 struct ceph_mds_snap_realm *ri;
 824
 825                 /*
 826                  * A "split" breaks part of an existing realm off into
 827                  * a new realm.  The MDS provides a list of inodes
 828                  * (with caps) and child realms that belong to the new
 829                  * child.
 830                  */
 831                 split_inos = p;
 832                 p += sizeof(u64) * num_split_inos;
 833                 split_realms = p;
 834                 p += sizeof(u64) * num_split_realms;
 835                 ceph_decode_need(&p, e, sizeof(*ri), bad);
 836                 /* we will peek at realm info here, but will _not_
 837                  * advance p, as the realm update will occur below in
 838                  * ceph_update_snap_trace. */
 839                 ri = p;
 840
 841                 realm = ceph_lookup_snap_realm(mdsc, split);
 842                 if (!realm) {
 843                         realm = ceph_create_snap_realm(mdsc, split);
 844                         if (IS_ERR(realm))
 845                                 goto out;
 846                 }
 847                 ceph_get_snap_realm(mdsc, realm);
 848
 849                 dout("splitting snap_realm %llx %p\n", realm->ino, realm);
 850                 for (i = 0; i < num_split_inos; i++) {
 851                         struct ceph_vino vino = {
 852                                 .ino = le64_to_cpu(split_inos[i]),
 853                                 .snap = CEPH_NOSNAP,
 854                         };
 855                         struct inode *inode = ceph_find_inode(sb, vino);
 856                         struct ceph_inode_info *ci;
 857                         struct ceph_snap_realm *oldrealm;
 858
 859                         if (!inode)
 860                                 continue;
 861                         ci = ceph_inode(inode);
 862
 863                         spin_lock(&ci->i_ceph_lock);
 864                         if (!ci->i_snap_realm)
 865                                 goto skip_inode;
 866                         /*
 867                          * If this inode belongs to a realm that was
 868                          * created after our new realm, we experienced
 869                          * a race (due to another split notifications
 870                          * arriving from a different MDS).  So skip
 871                          * this inode.
 872                          */
 873                         if (ci->i_snap_realm->created >
 874                             le64_to_cpu(ri->created)) {
 875                                 dout(" leaving %p in newer realm %llx %p\n",
 876                                      inode, ci->i_snap_realm->ino,
 877                                      ci->i_snap_realm);
 878                                 goto skip_inode;
 879                         }
 880                         dout(" will move %p to split realm %llx %p\n",
 881                              inode, realm->ino, realm);
 882                         /*
 883                          * Move the inode to the new realm
 884                          */
 885                         spin_lock(&realm->inodes_with_caps_lock);
 886                         list_del_init(&ci->i_snap_realm_item);
 887                         list_add(&ci->i_snap_realm_item,
 888                                  &realm->inodes_with_caps);
 889                         oldrealm = ci->i_snap_realm;
 890                         ci->i_snap_realm = realm;
 891                         spin_unlock(&realm->inodes_with_caps_lock);
 892                         spin_unlock(&ci->i_ceph_lock);
 893
 894                         ceph_get_snap_realm(mdsc, realm);
 895                         ceph_put_snap_realm(mdsc, oldrealm);
 896
 897                         iput(inode);
 898                         continue;
 899
 900 skip_inode:
 901                         spin_unlock(&ci->i_ceph_lock);
 902                         iput(inode);
 903                 }
 904
 905                 /* we may have taken some of the old realm's children. */
 906                 for (i = 0; i < num_split_realms; i++) {
 907                         struct ceph_snap_realm *child =
 908                                 ceph_lookup_snap_realm(mdsc,
 909                                            le64_to_cpu(split_realms[i]));
 910                         if (!child)
 911                                 continue;
 912                         adjust_snap_realm_parent(mdsc, child, realm->ino);
 913                 }
 914         }
 915
 916         /*
 917          * update using the provided snap trace. if we are deleting a
 918          * snap, we can avoid queueing cap_snaps.
 919          */
 920         ceph_update_snap_trace(mdsc, p, e,
 921                                op == CEPH_SNAP_OP_DESTROY);
 922
 923         if (op == CEPH_SNAP_OP_SPLIT)
 924                 /* we took a reference when we created the realm, above */
 925                 ceph_put_snap_realm(mdsc, realm);
 926
 927         __cleanup_empty_realms(mdsc);
 928
 929         up_write(&mdsc->snap_rwsem);
 930
 931         flush_snaps(mdsc);
 932         return;
 933
 934 bad:
 935         pr_err("corrupt snap message from mds%d\n", mds);
 936         ceph_msg_dump(msg);
 937 out:
 938         if (locked_rwsem)
 939                 up_write(&mdsc->snap_rwsem);
 940         return;
 941 }
 942
 943 int __init ceph_snap_init(void)
 944 {
 945         empty_snapc = ceph_create_snap_context(0, GFP_NOFS);
 946         if (!empty_snapc)
 947                 return -ENOMEM;
 948         empty_snapc->seq = 1;
 949         return 0;
 950 }
 951
 952 void ceph_snap_exit(void)
 953 {
 954         ceph_put_snap_context(empty_snapc);
 955 }