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