1// SPDX-License-Identifier: GPL-2.0 2#include <linux/ceph/ceph_debug.h> 3 4#include <linux/sort.h> 5#include <linux/slab.h> 6#include <linux/iversion.h> 7#include "super.h" 8#include "mds_client.h" 9#include <linux/ceph/decode.h> 10 11/* unused map expires after 5 minutes */ 12#define CEPH_SNAPID_MAP_TIMEOUT (5 * 60 * HZ) 13 14/* 15 * Snapshots in ceph are driven in large part by cooperation from the 16 * client. In contrast to local file systems or file servers that 17 * implement snapshots at a single point in the system, ceph's 18 * distributed access to storage requires clients to help decide 19 * whether a write logically occurs before or after a recently created 20 * snapshot. 21 * 22 * This provides a perfect instantanous client-wide snapshot. Between 23 * clients, however, snapshots may appear to be applied at slightly 24 * different points in time, depending on delays in delivering the 25 * snapshot notification. 26 * 27 * Snapshots are _not_ file system-wide. Instead, each snapshot 28 * applies to the subdirectory nested beneath some directory. This 29 * effectively divides the hierarchy into multiple "realms," where all 30 * of the files contained by each realm share the same set of 31 * snapshots. An individual realm's snap set contains snapshots 32 * explicitly created on that realm, as well as any snaps in its 33 * parent's snap set _after_ the point at which the parent became it's 34 * parent (due to, say, a rename). Similarly, snaps from prior parents 35 * during the time intervals during which they were the parent are included. 36 * 37 * The client is spared most of this detail, fortunately... it must only 38 * maintains a hierarchy of realms reflecting the current parent/child 39 * realm relationship, and for each realm has an explicit list of snaps 40 * inherited from prior parents. 41 * 42 * A snap_realm struct is maintained for realms containing every inode 43 * with an open cap in the system. (The needed snap realm information is 44 * provided by the MDS whenever a cap is issued, i.e., on open.) A 'seq' 45 * version number is used to ensure that as realm parameters change (new 46 * snapshot, new parent, etc.) the client's realm hierarchy is updated. 47 * 48 * The realm hierarchy drives the generation of a 'snap context' for each 49 * realm, which simply lists the resulting set of snaps for the realm. This 50 * is attached to any writes sent to OSDs. 51 */ 52/* 53 * Unfortunately error handling is a bit mixed here. If we get a snap 54 * update, but don't have enough memory to update our realm hierarchy, 55 * it's not clear what we can do about it (besides complaining to the 56 * console). 57 */ 58 59 60/* 61 * increase ref count for the realm 62 * 63 * caller must hold snap_rwsem. 64 */ 65void ceph_get_snap_realm(struct ceph_mds_client *mdsc, 66 struct ceph_snap_realm *realm) 67{ 68 lockdep_assert_held(&mdsc->snap_rwsem); 69 70 /* 71 * The 0->1 and 1->0 transitions must take the snap_empty_lock 72 * atomically with the refcount change. Go ahead and bump the 73 * nref here, unless it's 0, in which case we take the spinlock 74 * and then do the increment and remove it from the list. 75 */ 76 if (atomic_inc_not_zero(&realm->nref)) 77 return; 78 79 spin_lock(&mdsc->snap_empty_lock); 80 if (atomic_inc_return(&realm->nref) == 1) 81 list_del_init(&realm->empty_item); 82 spin_unlock(&mdsc->snap_empty_lock); 83} 84 85static void __insert_snap_realm(struct rb_root *root, 86 struct ceph_snap_realm *new) 87{ 88 struct rb_node **p = &root->rb_node; 89 struct rb_node *parent = NULL; 90 struct ceph_snap_realm *r = NULL; 91 92 while (*p) { 93 parent = *p; 94 r = rb_entry(parent, struct ceph_snap_realm, node); 95 if (new->ino < r->ino) 96 p = &(*p)->rb_left; 97 else if (new->ino > r->ino) 98 p = &(*p)->rb_right; 99 else 100 BUG(); 101 } 102 103 rb_link_node(&new->node, parent, p); 104 rb_insert_color(&new->node, root); 105} 106 107/* 108 * create and get the realm rooted at @ino and bump its ref count. 109 * 110 * caller must hold snap_rwsem for write. 111 */ 112static struct ceph_snap_realm *ceph_create_snap_realm( 113 struct ceph_mds_client *mdsc, 114 u64 ino) 115{ 116 struct ceph_snap_realm *realm; 117 118 lockdep_assert_held_write(&mdsc->snap_rwsem); 119 120 realm = kzalloc(sizeof(*realm), GFP_NOFS); 121 if (!realm) 122 return ERR_PTR(-ENOMEM); 123 124 atomic_set(&realm->nref, 1); /* for caller */ 125 realm->ino = ino; 126 INIT_LIST_HEAD(&realm->children); 127 INIT_LIST_HEAD(&realm->child_item); 128 INIT_LIST_HEAD(&realm->empty_item); 129 INIT_LIST_HEAD(&realm->dirty_item); 130 INIT_LIST_HEAD(&realm->inodes_with_caps); 131 spin_lock_init(&realm->inodes_with_caps_lock); 132 __insert_snap_realm(&mdsc->snap_realms, realm); 133 mdsc->num_snap_realms++; 134 135 dout("create_snap_realm %llx %p\n", realm->ino, realm); 136 return realm; 137} 138 139/* 140 * lookup the realm rooted at @ino. 141 * 142 * caller must hold snap_rwsem. 143 */ 144static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc, 145 u64 ino) 146{ 147 struct rb_node *n = mdsc->snap_realms.rb_node; 148 struct ceph_snap_realm *r; 149 150 lockdep_assert_held(&mdsc->snap_rwsem); 151 152 while (n) { 153 r = rb_entry(n, struct ceph_snap_realm, node); 154 if (ino < r->ino) 155 n = n->rb_left; 156 else if (ino > r->ino) 157 n = n->rb_right; 158 else { 159 dout("lookup_snap_realm %llx %p\n", r->ino, r); 160 return r; 161 } 162 } 163 return NULL; 164} 165 166struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc, 167 u64 ino) 168{ 169 struct ceph_snap_realm *r; 170 r = __lookup_snap_realm(mdsc, ino); 171 if (r) 172 ceph_get_snap_realm(mdsc, r); 173 return r; 174} 175 176static void __put_snap_realm(struct ceph_mds_client *mdsc, 177 struct ceph_snap_realm *realm); 178 179/* 180 * called with snap_rwsem (write) 181 */ 182static void __destroy_snap_realm(struct ceph_mds_client *mdsc, 183 struct ceph_snap_realm *realm) 184{ 185 lockdep_assert_held_write(&mdsc->snap_rwsem); 186 187 dout("__destroy_snap_realm %p %llx\n", realm, realm->ino); 188 189 rb_erase(&realm->node, &mdsc->snap_realms); 190 mdsc->num_snap_realms--; 191 192 if (realm->parent) { 193 list_del_init(&realm->child_item); 194 __put_snap_realm(mdsc, realm->parent); 195 } 196 197 kfree(realm->prior_parent_snaps); 198 kfree(realm->snaps); 199 ceph_put_snap_context(realm->cached_context); 200 kfree(realm); 201} 202 203/* 204 * caller holds snap_rwsem (write) 205 */ 206static void __put_snap_realm(struct ceph_mds_client *mdsc, 207 struct ceph_snap_realm *realm) 208{ 209 lockdep_assert_held_write(&mdsc->snap_rwsem); 210 211 /* 212 * We do not require the snap_empty_lock here, as any caller that 213 * increments the value must hold the snap_rwsem. 214 */ 215 if (atomic_dec_and_test(&realm->nref)) 216 __destroy_snap_realm(mdsc, realm); 217} 218 219/* 220 * See comments in ceph_get_snap_realm. Caller needn't hold any locks. 221 */ 222void ceph_put_snap_realm(struct ceph_mds_client *mdsc, 223 struct ceph_snap_realm *realm) 224{ 225 if (!atomic_dec_and_lock(&realm->nref, &mdsc->snap_empty_lock)) 226 return; 227 228 if (down_write_trylock(&mdsc->snap_rwsem)) { 229 spin_unlock(&mdsc->snap_empty_lock); 230 __destroy_snap_realm(mdsc, realm); 231 up_write(&mdsc->snap_rwsem); 232 } else { 233 list_add(&realm->empty_item, &mdsc->snap_empty); 234 spin_unlock(&mdsc->snap_empty_lock); 235 } 236} 237 238/* 239 * Clean up any realms whose ref counts have dropped to zero. Note 240 * that this does not include realms who were created but not yet 241 * used. 242 * 243 * Called under snap_rwsem (write) 244 */ 245static void __cleanup_empty_realms(struct ceph_mds_client *mdsc) 246{ 247 struct ceph_snap_realm *realm; 248 249 lockdep_assert_held_write(&mdsc->snap_rwsem); 250 251 spin_lock(&mdsc->snap_empty_lock); 252 while (!list_empty(&mdsc->snap_empty)) { 253 realm = list_first_entry(&mdsc->snap_empty, 254 struct ceph_snap_realm, empty_item); 255 list_del(&realm->empty_item); 256 spin_unlock(&mdsc->snap_empty_lock); 257 __destroy_snap_realm(mdsc, realm); 258 spin_lock(&mdsc->snap_empty_lock); 259 } 260 spin_unlock(&mdsc->snap_empty_lock); 261} 262 263void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc) 264{ 265 down_write(&mdsc->snap_rwsem); 266 __cleanup_empty_realms(mdsc); 267 up_write(&mdsc->snap_rwsem); 268} 269 270/* 271 * adjust the parent realm of a given @realm. adjust child list, and parent 272 * pointers, and ref counts appropriately. 273 * 274 * return true if parent was changed, 0 if unchanged, <0 on error. 275 * 276 * caller must hold snap_rwsem for write. 277 */ 278static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc, 279 struct ceph_snap_realm *realm, 280 u64 parentino) 281{ 282 struct ceph_snap_realm *parent; 283 284 lockdep_assert_held_write(&mdsc->snap_rwsem); 285 286 if (realm->parent_ino == parentino) 287 return 0; 288 289 parent = ceph_lookup_snap_realm(mdsc, parentino); 290 if (!parent) { 291 parent = ceph_create_snap_realm(mdsc, parentino); 292 if (IS_ERR(parent)) 293 return PTR_ERR(parent); 294 } 295 dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n", 296 realm->ino, realm, realm->parent_ino, realm->parent, 297 parentino, parent); 298 if (realm->parent) { 299 list_del_init(&realm->child_item); 300 ceph_put_snap_realm(mdsc, realm->parent); 301 } 302 realm->parent_ino = parentino; 303 realm->parent = parent; 304 list_add(&realm->child_item, &parent->children); 305 return 1; 306} 307 308 309static int cmpu64_rev(const void *a, const void *b) 310{ 311 if (*(u64 *)a < *(u64 *)b) 312 return 1; 313 if (*(u64 *)a > *(u64 *)b) 314 return -1; 315 return 0; 316} 317 318 319/* 320 * build the snap context for a given realm. 321 */ 322static int build_snap_context(struct ceph_snap_realm *realm, 323 struct list_head* dirty_realms) 324{ 325 struct ceph_snap_realm *parent = realm->parent; 326 struct ceph_snap_context *snapc; 327 int err = 0; 328 u32 num = realm->num_prior_parent_snaps + realm->num_snaps; 329 330 /* 331 * build parent context, if it hasn't been built. 332 * conservatively estimate that all parent snaps might be 333 * included by us. 334 */ 335 if (parent) { 336 if (!parent->cached_context) { 337 err = build_snap_context(parent, dirty_realms); 338 if (err) 339 goto fail; 340 } 341 num += parent->cached_context->num_snaps; 342 } 343 344 /* do i actually need to update? not if my context seq 345 matches realm seq, and my parents' does to. (this works 346 because we rebuild_snap_realms() works _downward_ in 347 hierarchy after each update.) */ 348 if (realm->cached_context && 349 realm->cached_context->seq == realm->seq && 350 (!parent || 351 realm->cached_context->seq >= parent->cached_context->seq)) { 352 dout("build_snap_context %llx %p: %p seq %lld (%u snaps)" 353 " (unchanged)\n", 354 realm->ino, realm, realm->cached_context, 355 realm->cached_context->seq, 356 (unsigned int)realm->cached_context->num_snaps); 357 return 0; 358 } 359 360 /* alloc new snap context */ 361 err = -ENOMEM; 362 if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64)) 363 goto fail; 364 snapc = ceph_create_snap_context(num, GFP_NOFS); 365 if (!snapc) 366 goto fail; 367 368 /* build (reverse sorted) snap vector */ 369 num = 0; 370 snapc->seq = realm->seq; 371 if (parent) { 372 u32 i; 373 374 /* include any of parent's snaps occurring _after_ my 375 parent became my parent */ 376 for (i = 0; i < parent->cached_context->num_snaps; i++) 377 if (parent->cached_context->snaps[i] >= 378 realm->parent_since) 379 snapc->snaps[num++] = 380 parent->cached_context->snaps[i]; 381 if (parent->cached_context->seq > snapc->seq) 382 snapc->seq = parent->cached_context->seq; 383 } 384 memcpy(snapc->snaps + num, realm->snaps, 385 sizeof(u64)*realm->num_snaps); 386 num += realm->num_snaps; 387 memcpy(snapc->snaps + num, realm->prior_parent_snaps, 388 sizeof(u64)*realm->num_prior_parent_snaps); 389 num += realm->num_prior_parent_snaps; 390 391 sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL); 392 snapc->num_snaps = num; 393 dout("build_snap_context %llx %p: %p seq %lld (%u snaps)\n", 394 realm->ino, realm, snapc, snapc->seq, 395 (unsigned int) snapc->num_snaps); 396 397 ceph_put_snap_context(realm->cached_context); 398 realm->cached_context = snapc; 399 /* queue realm for cap_snap creation */ 400 list_add_tail(&realm->dirty_item, dirty_realms); 401 return 0; 402 403fail: 404 /* 405 * if we fail, clear old (incorrect) cached_context... hopefully 406 * we'll have better luck building it later 407 */ 408 if (realm->cached_context) { 409 ceph_put_snap_context(realm->cached_context); 410 realm->cached_context = NULL; 411 } 412 pr_err("build_snap_context %llx %p fail %d\n", realm->ino, 413 realm, err); 414 return err; 415} 416 417/* 418 * rebuild snap context for the given realm and all of its children. 419 */ 420static void rebuild_snap_realms(struct ceph_snap_realm *realm, 421 struct list_head *dirty_realms) 422{ 423 struct ceph_snap_realm *child; 424 425 dout("rebuild_snap_realms %llx %p\n", realm->ino, realm); 426 build_snap_context(realm, dirty_realms); 427 428 list_for_each_entry(child, &realm->children, child_item) 429 rebuild_snap_realms(child, dirty_realms); 430} 431 432 433/* 434 * helper to allocate and decode an array of snapids. free prior 435 * instance, if any. 436 */ 437static int dup_array(u64 **dst, __le64 *src, u32 num) 438{ 439 u32 i; 440 441 kfree(*dst); 442 if (num) { 443 *dst = kcalloc(num, sizeof(u64), GFP_NOFS); 444 if (!*dst) 445 return -ENOMEM; 446 for (i = 0; i < num; i++) 447 (*dst)[i] = get_unaligned_le64(src + i); 448 } else { 449 *dst = NULL; 450 } 451 return 0; 452} 453 454static bool has_new_snaps(struct ceph_snap_context *o, 455 struct ceph_snap_context *n) 456{ 457 if (n->num_snaps == 0) 458 return false; 459 /* snaps are in descending order */ 460 return n->snaps[0] > o->seq; 461} 462 463/* 464 * When a snapshot is applied, the size/mtime inode metadata is queued 465 * in a ceph_cap_snap (one for each snapshot) until writeback 466 * completes and the metadata can be flushed back to the MDS. 467 * 468 * However, if a (sync) write is currently in-progress when we apply 469 * the snapshot, we have to wait until the write succeeds or fails 470 * (and a final size/mtime is known). In this case the 471 * cap_snap->writing = 1, and is said to be "pending." When the write 472 * finishes, we __ceph_finish_cap_snap(). 473 * 474 * Caller must hold snap_rwsem for read (i.e., the realm topology won't 475 * change). 476 */ 477void ceph_queue_cap_snap(struct ceph_inode_info *ci) 478{ 479 struct inode *inode = &ci->vfs_inode; 480 struct ceph_cap_snap *capsnap; 481 struct ceph_snap_context *old_snapc, *new_snapc; 482 struct ceph_buffer *old_blob = NULL; 483 int used, dirty; 484 485 capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS); 486 if (!capsnap) { 487 pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode); 488 return; 489 } 490 capsnap->cap_flush.is_capsnap = true; 491 INIT_LIST_HEAD(&capsnap->cap_flush.i_list); 492 INIT_LIST_HEAD(&capsnap->cap_flush.g_list); 493 494 spin_lock(&ci->i_ceph_lock); 495 used = __ceph_caps_used(ci); 496 dirty = __ceph_caps_dirty(ci); 497 498 old_snapc = ci->i_head_snapc; 499 new_snapc = ci->i_snap_realm->cached_context; 500 501 /* 502 * If there is a write in progress, treat that as a dirty Fw, 503 * even though it hasn't completed yet; by the time we finish 504 * up this capsnap it will be. 505 */ 506 if (used & CEPH_CAP_FILE_WR) 507 dirty |= CEPH_CAP_FILE_WR; 508 509 if (__ceph_have_pending_cap_snap(ci)) { 510 /* there is no point in queuing multiple "pending" cap_snaps, 511 as no new writes are allowed to start when pending, so any 512 writes in progress now were started before the previous 513 cap_snap. lucky us. */ 514 dout("queue_cap_snap %p already pending\n", inode); 515 goto update_snapc; 516 } 517 if (ci->i_wrbuffer_ref_head == 0 && 518 !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) { 519 dout("queue_cap_snap %p nothing dirty|writing\n", inode); 520 goto update_snapc; 521 } 522 523 BUG_ON(!old_snapc); 524 525 /* 526 * There is no need to send FLUSHSNAP message to MDS if there is 527 * no new snapshot. But when there is dirty pages or on-going 528 * writes, we still need to create cap_snap. cap_snap is needed 529 * by the write path and page writeback path. 530 * 531 * also see ceph_try_drop_cap_snap() 532 */ 533 if (has_new_snaps(old_snapc, new_snapc)) { 534 if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR)) 535 capsnap->need_flush = true; 536 } else { 537 if (!(used & CEPH_CAP_FILE_WR) && 538 ci->i_wrbuffer_ref_head == 0) { 539 dout("queue_cap_snap %p " 540 "no new_snap|dirty_page|writing\n", inode); 541 goto update_snapc; 542 } 543 } 544 545 dout("queue_cap_snap %p cap_snap %p queuing under %p %s %s\n", 546 inode, capsnap, old_snapc, ceph_cap_string(dirty), 547 capsnap->need_flush ? "" : "no_flush"); 548 ihold(inode); 549 550 refcount_set(&capsnap->nref, 1); 551 INIT_LIST_HEAD(&capsnap->ci_item); 552 553 capsnap->follows = old_snapc->seq; 554 capsnap->issued = __ceph_caps_issued(ci, NULL); 555 capsnap->dirty = dirty; 556 557 capsnap->mode = inode->i_mode; 558 capsnap->uid = inode->i_uid; 559 capsnap->gid = inode->i_gid; 560 561 if (dirty & CEPH_CAP_XATTR_EXCL) { 562 old_blob = __ceph_build_xattrs_blob(ci); 563 capsnap->xattr_blob = 564 ceph_buffer_get(ci->i_xattrs.blob); 565 capsnap->xattr_version = ci->i_xattrs.version; 566 } else { 567 capsnap->xattr_blob = NULL; 568 capsnap->xattr_version = 0; 569 } 570 571 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE; 572 573 /* dirty page count moved from _head to this cap_snap; 574 all subsequent writes page dirties occur _after_ this 575 snapshot. */ 576 capsnap->dirty_pages = ci->i_wrbuffer_ref_head; 577 ci->i_wrbuffer_ref_head = 0; 578 capsnap->context = old_snapc; 579 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps); 580 581 if (used & CEPH_CAP_FILE_WR) { 582 dout("queue_cap_snap %p cap_snap %p snapc %p" 583 " seq %llu used WR, now pending\n", inode, 584 capsnap, old_snapc, old_snapc->seq); 585 capsnap->writing = 1; 586 } else { 587 /* note mtime, size NOW. */ 588 __ceph_finish_cap_snap(ci, capsnap); 589 } 590 capsnap = NULL; 591 old_snapc = NULL; 592 593update_snapc: 594 if (ci->i_wrbuffer_ref_head == 0 && 595 ci->i_wr_ref == 0 && 596 ci->i_dirty_caps == 0 && 597 ci->i_flushing_caps == 0) { 598 ci->i_head_snapc = NULL; 599 } else { 600 ci->i_head_snapc = ceph_get_snap_context(new_snapc); 601 dout(" new snapc is %p\n", new_snapc); 602 } 603 spin_unlock(&ci->i_ceph_lock); 604 605 ceph_buffer_put(old_blob); 606 kfree(capsnap); 607 ceph_put_snap_context(old_snapc); 608} 609 610/* 611 * Finalize the size, mtime for a cap_snap.. that is, settle on final values 612 * to be used for the snapshot, to be flushed back to the mds. 613 * 614 * If capsnap can now be flushed, add to snap_flush list, and return 1. 615 * 616 * Caller must hold i_ceph_lock. 617 */ 618int __ceph_finish_cap_snap(struct ceph_inode_info *ci, 619 struct ceph_cap_snap *capsnap) 620{ 621 struct inode *inode = &ci->vfs_inode; 622 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb); 623 624 BUG_ON(capsnap->writing); 625 capsnap->size = inode->i_size; 626 capsnap->mtime = inode->i_mtime; 627 capsnap->atime = inode->i_atime; 628 capsnap->ctime = inode->i_ctime; 629 capsnap->btime = ci->i_btime; 630 capsnap->change_attr = inode_peek_iversion_raw(inode); 631 capsnap->time_warp_seq = ci->i_time_warp_seq; 632 capsnap->truncate_size = ci->i_truncate_size; 633 capsnap->truncate_seq = ci->i_truncate_seq; 634 if (capsnap->dirty_pages) { 635 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu " 636 "still has %d dirty pages\n", inode, capsnap, 637 capsnap->context, capsnap->context->seq, 638 ceph_cap_string(capsnap->dirty), capsnap->size, 639 capsnap->dirty_pages); 640 return 0; 641 } 642 643 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS; 644 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n", 645 inode, capsnap, capsnap->context, 646 capsnap->context->seq, ceph_cap_string(capsnap->dirty), 647 capsnap->size); 648 649 spin_lock(&mdsc->snap_flush_lock); 650 if (list_empty(&ci->i_snap_flush_item)) { 651 ihold(inode); 652 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list); 653 } 654 spin_unlock(&mdsc->snap_flush_lock); 655 return 1; /* caller may want to ceph_flush_snaps */ 656} 657 658/* 659 * Queue cap_snaps for snap writeback for this realm and its children. 660 * Called under snap_rwsem, so realm topology won't change. 661 */ 662static void queue_realm_cap_snaps(struct ceph_snap_realm *realm) 663{ 664 struct ceph_inode_info *ci; 665 struct inode *lastinode = NULL; 666 667 dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino); 668 669 spin_lock(&realm->inodes_with_caps_lock); 670 list_for_each_entry(ci, &realm->inodes_with_caps, i_snap_realm_item) { 671 struct inode *inode = igrab(&ci->vfs_inode); 672 if (!inode) 673 continue; 674 spin_unlock(&realm->inodes_with_caps_lock); 675 /* avoid calling iput_final() while holding 676 * mdsc->snap_rwsem or in mds dispatch threads */ 677 ceph_async_iput(lastinode); 678 lastinode = inode; 679 ceph_queue_cap_snap(ci); 680 spin_lock(&realm->inodes_with_caps_lock); 681 } 682 spin_unlock(&realm->inodes_with_caps_lock); 683 ceph_async_iput(lastinode); 684 685 dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino); 686} 687 688/* 689 * Parse and apply a snapblob "snap trace" from the MDS. This specifies 690 * the snap realm parameters from a given realm and all of its ancestors, 691 * up to the root. 692 * 693 * Caller must hold snap_rwsem for write. 694 */ 695int ceph_update_snap_trace(struct ceph_mds_client *mdsc, 696 void *p, void *e, bool deletion, 697 struct ceph_snap_realm **realm_ret) 698{ 699 struct ceph_mds_snap_realm *ri; /* encoded */ 700 __le64 *snaps; /* encoded */ 701 __le64 *prior_parent_snaps; /* encoded */ 702 struct ceph_snap_realm *realm; 703 struct ceph_snap_realm *first_realm = NULL; 704 struct ceph_snap_realm *realm_to_rebuild = NULL; 705 int rebuild_snapcs; 706 int err = -ENOMEM; 707 LIST_HEAD(dirty_realms); 708 709 lockdep_assert_held_write(&mdsc->snap_rwsem); 710 711 dout("update_snap_trace deletion=%d\n", deletion); 712more: 713 realm = NULL; 714 rebuild_snapcs = 0; 715 ceph_decode_need(&p, e, sizeof(*ri), bad); 716 ri = p; 717 p += sizeof(*ri); 718 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) + 719 le32_to_cpu(ri->num_prior_parent_snaps)), bad); 720 snaps = p; 721 p += sizeof(u64) * le32_to_cpu(ri->num_snaps); 722 prior_parent_snaps = p; 723 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps); 724 725 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino)); 726 if (!realm) { 727 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino)); 728 if (IS_ERR(realm)) { 729 err = PTR_ERR(realm); 730 goto fail; 731 } 732 } 733 734 /* ensure the parent is correct */ 735 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent)); 736 if (err < 0) 737 goto fail; 738 rebuild_snapcs += err; 739 740 if (le64_to_cpu(ri->seq) > realm->seq) { 741 dout("update_snap_trace updating %llx %p %lld -> %lld\n", 742 realm->ino, realm, realm->seq, le64_to_cpu(ri->seq)); 743 /* update realm parameters, snap lists */ 744 realm->seq = le64_to_cpu(ri->seq); 745 realm->created = le64_to_cpu(ri->created); 746 realm->parent_since = le64_to_cpu(ri->parent_since); 747 748 realm->num_snaps = le32_to_cpu(ri->num_snaps); 749 err = dup_array(&realm->snaps, snaps, realm->num_snaps); 750 if (err < 0) 751 goto fail; 752 753 realm->num_prior_parent_snaps = 754 le32_to_cpu(ri->num_prior_parent_snaps); 755 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps, 756 realm->num_prior_parent_snaps); 757 if (err < 0) 758 goto fail; 759 760 if (realm->seq > mdsc->last_snap_seq) 761 mdsc->last_snap_seq = realm->seq; 762 763 rebuild_snapcs = 1; 764 } else if (!realm->cached_context) { 765 dout("update_snap_trace %llx %p seq %lld new\n", 766 realm->ino, realm, realm->seq); 767 rebuild_snapcs = 1; 768 } else { 769 dout("update_snap_trace %llx %p seq %lld unchanged\n", 770 realm->ino, realm, realm->seq); 771 } 772 773 dout("done with %llx %p, rebuild_snapcs=%d, %p %p\n", realm->ino, 774 realm, rebuild_snapcs, p, e); 775 776 /* 777 * this will always track the uppest parent realm from which 778 * we need to rebuild the snapshot contexts _downward_ in 779 * hierarchy. 780 */ 781 if (rebuild_snapcs) 782 realm_to_rebuild = realm; 783 784 /* rebuild_snapcs when we reach the _end_ (root) of the trace */ 785 if (realm_to_rebuild && p >= e) 786 rebuild_snap_realms(realm_to_rebuild, &dirty_realms); 787 788 if (!first_realm) 789 first_realm = realm; 790 else 791 ceph_put_snap_realm(mdsc, realm); 792 793 if (p < e) 794 goto more; 795 796 /* 797 * queue cap snaps _after_ we've built the new snap contexts, 798 * so that i_head_snapc can be set appropriately. 799 */ 800 while (!list_empty(&dirty_realms)) { 801 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm, 802 dirty_item); 803 list_del_init(&realm->dirty_item); 804 queue_realm_cap_snaps(realm); 805 } 806 807 if (realm_ret) 808 *realm_ret = first_realm; 809 else 810 ceph_put_snap_realm(mdsc, first_realm); 811 812 __cleanup_empty_realms(mdsc); 813 return 0; 814 815bad: 816 err = -EINVAL; 817fail: 818 if (realm && !IS_ERR(realm)) 819 ceph_put_snap_realm(mdsc, realm); 820 if (first_realm) 821 ceph_put_snap_realm(mdsc, first_realm); 822 pr_err("update_snap_trace error %d\n", err); 823 return err; 824} 825 826 827/* 828 * Send any cap_snaps that are queued for flush. Try to carry 829 * s_mutex across multiple snap flushes to avoid locking overhead. 830 * 831 * Caller holds no locks. 832 */ 833static void flush_snaps(struct ceph_mds_client *mdsc) 834{ 835 struct ceph_inode_info *ci; 836 struct inode *inode; 837 struct ceph_mds_session *session = NULL; 838 839 dout("flush_snaps\n"); 840 spin_lock(&mdsc->snap_flush_lock); 841 while (!list_empty(&mdsc->snap_flush_list)) { 842 ci = list_first_entry(&mdsc->snap_flush_list, 843 struct ceph_inode_info, i_snap_flush_item); 844 inode = &ci->vfs_inode; 845 ihold(inode); 846 spin_unlock(&mdsc->snap_flush_lock); 847 ceph_flush_snaps(ci, &session); 848 /* avoid calling iput_final() while holding 849 * session->s_mutex or in mds dispatch threads */ 850 ceph_async_iput(inode); 851 spin_lock(&mdsc->snap_flush_lock); 852 } 853 spin_unlock(&mdsc->snap_flush_lock); 854 855 if (session) { 856 mutex_unlock(&session->s_mutex); 857 ceph_put_mds_session(session); 858 } 859 dout("flush_snaps done\n"); 860} 861 862 863/* 864 * Handle a snap notification from the MDS. 865 * 866 * This can take two basic forms: the simplest is just a snap creation 867 * or deletion notification on an existing realm. This should update the 868 * realm and its children. 869 * 870 * The more difficult case is realm creation, due to snap creation at a 871 * new point in the file hierarchy, or due to a rename that moves a file or 872 * directory into another realm. 873 */ 874void ceph_handle_snap(struct ceph_mds_client *mdsc, 875 struct ceph_mds_session *session, 876 struct ceph_msg *msg) 877{ 878 struct super_block *sb = mdsc->fsc->sb; 879 int mds = session->s_mds; 880 u64 split; 881 int op; 882 int trace_len; 883 struct ceph_snap_realm *realm = NULL; 884 void *p = msg->front.iov_base; 885 void *e = p + msg->front.iov_len; 886 struct ceph_mds_snap_head *h; 887 int num_split_inos, num_split_realms; 888 __le64 *split_inos = NULL, *split_realms = NULL; 889 int i; 890 int locked_rwsem = 0; 891 892 /* decode */ 893 if (msg->front.iov_len < sizeof(*h)) 894 goto bad; 895 h = p; 896 op = le32_to_cpu(h->op); 897 split = le64_to_cpu(h->split); /* non-zero if we are splitting an 898 * existing realm */ 899 num_split_inos = le32_to_cpu(h->num_split_inos); 900 num_split_realms = le32_to_cpu(h->num_split_realms); 901 trace_len = le32_to_cpu(h->trace_len); 902 p += sizeof(*h); 903 904 dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds, 905 ceph_snap_op_name(op), split, trace_len); 906 907 mutex_lock(&session->s_mutex); 908 inc_session_sequence(session); 909 mutex_unlock(&session->s_mutex); 910 911 down_write(&mdsc->snap_rwsem); 912 locked_rwsem = 1; 913 914 if (op == CEPH_SNAP_OP_SPLIT) { 915 struct ceph_mds_snap_realm *ri; 916 917 /* 918 * A "split" breaks part of an existing realm off into 919 * a new realm. The MDS provides a list of inodes 920 * (with caps) and child realms that belong to the new 921 * child. 922 */ 923 split_inos = p; 924 p += sizeof(u64) * num_split_inos; 925 split_realms = p; 926 p += sizeof(u64) * num_split_realms; 927 ceph_decode_need(&p, e, sizeof(*ri), bad); 928 /* we will peek at realm info here, but will _not_ 929 * advance p, as the realm update will occur below in 930 * ceph_update_snap_trace. */ 931 ri = p; 932 933 realm = ceph_lookup_snap_realm(mdsc, split); 934 if (!realm) { 935 realm = ceph_create_snap_realm(mdsc, split); 936 if (IS_ERR(realm)) 937 goto out; 938 } 939 940 dout("splitting snap_realm %llx %p\n", realm->ino, realm); 941 for (i = 0; i < num_split_inos; i++) { 942 struct ceph_vino vino = { 943 .ino = le64_to_cpu(split_inos[i]), 944 .snap = CEPH_NOSNAP, 945 }; 946 struct inode *inode = ceph_find_inode(sb, vino); 947 struct ceph_inode_info *ci; 948 struct ceph_snap_realm *oldrealm; 949 950 if (!inode) 951 continue; 952 ci = ceph_inode(inode); 953 954 spin_lock(&ci->i_ceph_lock); 955 if (!ci->i_snap_realm) 956 goto skip_inode; 957 /* 958 * If this inode belongs to a realm that was 959 * created after our new realm, we experienced 960 * a race (due to another split notifications 961 * arriving from a different MDS). So skip 962 * this inode. 963 */ 964 if (ci->i_snap_realm->created > 965 le64_to_cpu(ri->created)) { 966 dout(" leaving %p in newer realm %llx %p\n", 967 inode, ci->i_snap_realm->ino, 968 ci->i_snap_realm); 969 goto skip_inode; 970 } 971 dout(" will move %p to split realm %llx %p\n", 972 inode, realm->ino, realm); 973 /* 974 * Move the inode to the new realm 975 */ 976 oldrealm = ci->i_snap_realm; 977 spin_lock(&oldrealm->inodes_with_caps_lock); 978 list_del_init(&ci->i_snap_realm_item); 979 spin_unlock(&oldrealm->inodes_with_caps_lock); 980 981 spin_lock(&realm->inodes_with_caps_lock); 982 list_add(&ci->i_snap_realm_item, 983 &realm->inodes_with_caps); 984 ci->i_snap_realm = realm; 985 if (realm->ino == ci->i_vino.ino) 986 realm->inode = inode; 987 spin_unlock(&realm->inodes_with_caps_lock); 988 989 spin_unlock(&ci->i_ceph_lock); 990 991 ceph_get_snap_realm(mdsc, realm); 992 ceph_put_snap_realm(mdsc, oldrealm); 993 994 /* avoid calling iput_final() while holding 995 * mdsc->snap_rwsem or mds in dispatch threads */ 996 ceph_async_iput(inode); 997 continue; 998 999skip_inode: 1000 spin_unlock(&ci->i_ceph_lock); 1001 ceph_async_iput(inode); 1002 } 1003 1004 /* we may have taken some of the old realm's children. */ 1005 for (i = 0; i < num_split_realms; i++) { 1006 struct ceph_snap_realm *child = 1007 __lookup_snap_realm(mdsc, 1008 le64_to_cpu(split_realms[i])); 1009 if (!child) 1010 continue; 1011 adjust_snap_realm_parent(mdsc, child, realm->ino); 1012 } 1013 } else { 1014 /* 1015 * In the non-split case both 'num_split_inos' and 1016 * 'num_split_realms' should be 0, making this a no-op. 1017 * However the MDS happens to populate 'split_realms' list 1018 * in one of the UPDATE op cases by mistake. 1019 * 1020 * Skip both lists just in case to ensure that 'p' is 1021 * positioned at the start of realm info, as expected by 1022 * ceph_update_snap_trace(). 1023 */ 1024 p += sizeof(u64) * num_split_inos; 1025 p += sizeof(u64) * num_split_realms; 1026 } 1027 1028 /* 1029 * update using the provided snap trace. if we are deleting a 1030 * snap, we can avoid queueing cap_snaps. 1031 */ 1032 ceph_update_snap_trace(mdsc, p, e, 1033 op == CEPH_SNAP_OP_DESTROY, NULL); 1034 1035 if (op == CEPH_SNAP_OP_SPLIT) 1036 /* we took a reference when we created the realm, above */ 1037 ceph_put_snap_realm(mdsc, realm); 1038 1039 __cleanup_empty_realms(mdsc); 1040 1041 up_write(&mdsc->snap_rwsem); 1042 1043 flush_snaps(mdsc); 1044 return; 1045 1046bad: 1047 pr_err("corrupt snap message from mds%d\n", mds); 1048 ceph_msg_dump(msg); 1049out: 1050 if (locked_rwsem) 1051 up_write(&mdsc->snap_rwsem); 1052 return; 1053} 1054 1055struct ceph_snapid_map* ceph_get_snapid_map(struct ceph_mds_client *mdsc, 1056 u64 snap) 1057{ 1058 struct ceph_snapid_map *sm, *exist; 1059 struct rb_node **p, *parent; 1060 int ret; 1061 1062 exist = NULL; 1063 spin_lock(&mdsc->snapid_map_lock); 1064 p = &mdsc->snapid_map_tree.rb_node; 1065 while (*p) { 1066 exist = rb_entry(*p, struct ceph_snapid_map, node); 1067 if (snap > exist->snap) { 1068 p = &(*p)->rb_left; 1069 } else if (snap < exist->snap) { 1070 p = &(*p)->rb_right; 1071 } else { 1072 if (atomic_inc_return(&exist->ref) == 1) 1073 list_del_init(&exist->lru); 1074 break; 1075 } 1076 exist = NULL; 1077 } 1078 spin_unlock(&mdsc->snapid_map_lock); 1079 if (exist) { 1080 dout("found snapid map %llx -> %x\n", exist->snap, exist->dev); 1081 return exist; 1082 } 1083 1084 sm = kmalloc(sizeof(*sm), GFP_NOFS); 1085 if (!sm) 1086 return NULL; 1087 1088 ret = get_anon_bdev(&sm->dev); 1089 if (ret < 0) { 1090 kfree(sm); 1091 return NULL; 1092 } 1093 1094 INIT_LIST_HEAD(&sm->lru); 1095 atomic_set(&sm->ref, 1); 1096 sm->snap = snap; 1097 1098 exist = NULL; 1099 parent = NULL; 1100 p = &mdsc->snapid_map_tree.rb_node; 1101 spin_lock(&mdsc->snapid_map_lock); 1102 while (*p) { 1103 parent = *p; 1104 exist = rb_entry(*p, struct ceph_snapid_map, node); 1105 if (snap > exist->snap) 1106 p = &(*p)->rb_left; 1107 else if (snap < exist->snap) 1108 p = &(*p)->rb_right; 1109 else 1110 break; 1111 exist = NULL; 1112 } 1113 if (exist) { 1114 if (atomic_inc_return(&exist->ref) == 1) 1115 list_del_init(&exist->lru); 1116 } else { 1117 rb_link_node(&sm->node, parent, p); 1118 rb_insert_color(&sm->node, &mdsc->snapid_map_tree); 1119 } 1120 spin_unlock(&mdsc->snapid_map_lock); 1121 if (exist) { 1122 free_anon_bdev(sm->dev); 1123 kfree(sm); 1124 dout("found snapid map %llx -> %x\n", exist->snap, exist->dev); 1125 return exist; 1126 } 1127 1128 dout("create snapid map %llx -> %x\n", sm->snap, sm->dev); 1129 return sm; 1130} 1131 1132void ceph_put_snapid_map(struct ceph_mds_client* mdsc, 1133 struct ceph_snapid_map *sm) 1134{ 1135 if (!sm) 1136 return; 1137 if (atomic_dec_and_lock(&sm->ref, &mdsc->snapid_map_lock)) { 1138 if (!RB_EMPTY_NODE(&sm->node)) { 1139 sm->last_used = jiffies; 1140 list_add_tail(&sm->lru, &mdsc->snapid_map_lru); 1141 spin_unlock(&mdsc->snapid_map_lock); 1142 } else { 1143 /* already cleaned up by 1144 * ceph_cleanup_snapid_map() */ 1145 spin_unlock(&mdsc->snapid_map_lock); 1146 kfree(sm); 1147 } 1148 } 1149} 1150 1151void ceph_trim_snapid_map(struct ceph_mds_client *mdsc) 1152{ 1153 struct ceph_snapid_map *sm; 1154 unsigned long now; 1155 LIST_HEAD(to_free); 1156 1157 spin_lock(&mdsc->snapid_map_lock); 1158 now = jiffies; 1159 1160 while (!list_empty(&mdsc->snapid_map_lru)) { 1161 sm = list_first_entry(&mdsc->snapid_map_lru, 1162 struct ceph_snapid_map, lru); 1163 if (time_after(sm->last_used + CEPH_SNAPID_MAP_TIMEOUT, now)) 1164 break; 1165 1166 rb_erase(&sm->node, &mdsc->snapid_map_tree); 1167 list_move(&sm->lru, &to_free); 1168 } 1169 spin_unlock(&mdsc->snapid_map_lock); 1170 1171 while (!list_empty(&to_free)) { 1172 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru); 1173 list_del(&sm->lru); 1174 dout("trim snapid map %llx -> %x\n", sm->snap, sm->dev); 1175 free_anon_bdev(sm->dev); 1176 kfree(sm); 1177 } 1178} 1179 1180void ceph_cleanup_snapid_map(struct ceph_mds_client *mdsc) 1181{ 1182 struct ceph_snapid_map *sm; 1183 struct rb_node *p; 1184 LIST_HEAD(to_free); 1185 1186 spin_lock(&mdsc->snapid_map_lock); 1187 while ((p = rb_first(&mdsc->snapid_map_tree))) { 1188 sm = rb_entry(p, struct ceph_snapid_map, node); 1189 rb_erase(p, &mdsc->snapid_map_tree); 1190 RB_CLEAR_NODE(p); 1191 list_move(&sm->lru, &to_free); 1192 } 1193 spin_unlock(&mdsc->snapid_map_lock); 1194 1195 while (!list_empty(&to_free)) { 1196 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru); 1197 list_del(&sm->lru); 1198 free_anon_bdev(sm->dev); 1199 if (WARN_ON_ONCE(atomic_read(&sm->ref))) { 1200 pr_err("snapid map %llx -> %x still in use\n", 1201 sm->snap, sm->dev); 1202 } 1203 kfree(sm); 1204 } 1205} 1206