1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/ceph/ceph_debug.h> 3 4 #include <linux/fs.h> 5 #include <linux/kernel.h> 6 #include <linux/sched/signal.h> 7 #include <linux/slab.h> 8 #include <linux/vmalloc.h> 9 #include <linux/wait.h> 10 #include <linux/writeback.h> 11 #include <linux/iversion.h> 12 13 #include "super.h" 14 #include "mds_client.h" 15 #include "cache.h" 16 #include <linux/ceph/decode.h> 17 #include <linux/ceph/messenger.h> 18 19 /* 20 * Capability management 21 * 22 * The Ceph metadata servers control client access to inode metadata 23 * and file data by issuing capabilities, granting clients permission 24 * to read and/or write both inode field and file data to OSDs 25 * (storage nodes). Each capability consists of a set of bits 26 * indicating which operations are allowed. 27 * 28 * If the client holds a *_SHARED cap, the client has a coherent value 29 * that can be safely read from the cached inode. 30 * 31 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the 32 * client is allowed to change inode attributes (e.g., file size, 33 * mtime), note its dirty state in the ceph_cap, and asynchronously 34 * flush that metadata change to the MDS. 35 * 36 * In the event of a conflicting operation (perhaps by another 37 * client), the MDS will revoke the conflicting client capabilities. 38 * 39 * In order for a client to cache an inode, it must hold a capability 40 * with at least one MDS server. When inodes are released, release 41 * notifications are batched and periodically sent en masse to the MDS 42 * cluster to release server state. 43 */ 44 45 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc); 46 static void __kick_flushing_caps(struct ceph_mds_client *mdsc, 47 struct ceph_mds_session *session, 48 struct ceph_inode_info *ci, 49 u64 oldest_flush_tid); 50 51 /* 52 * Generate readable cap strings for debugging output. 53 */ 54 #define MAX_CAP_STR 20 55 static char cap_str[MAX_CAP_STR][40]; 56 static DEFINE_SPINLOCK(cap_str_lock); 57 static int last_cap_str; 58 59 static char *gcap_string(char *s, int c) 60 { 61 if (c & CEPH_CAP_GSHARED) 62 *s++ = 's'; 63 if (c & CEPH_CAP_GEXCL) 64 *s++ = 'x'; 65 if (c & CEPH_CAP_GCACHE) 66 *s++ = 'c'; 67 if (c & CEPH_CAP_GRD) 68 *s++ = 'r'; 69 if (c & CEPH_CAP_GWR) 70 *s++ = 'w'; 71 if (c & CEPH_CAP_GBUFFER) 72 *s++ = 'b'; 73 if (c & CEPH_CAP_GWREXTEND) 74 *s++ = 'a'; 75 if (c & CEPH_CAP_GLAZYIO) 76 *s++ = 'l'; 77 return s; 78 } 79 80 const char *ceph_cap_string(int caps) 81 { 82 int i; 83 char *s; 84 int c; 85 86 spin_lock(&cap_str_lock); 87 i = last_cap_str++; 88 if (last_cap_str == MAX_CAP_STR) 89 last_cap_str = 0; 90 spin_unlock(&cap_str_lock); 91 92 s = cap_str[i]; 93 94 if (caps & CEPH_CAP_PIN) 95 *s++ = 'p'; 96 97 c = (caps >> CEPH_CAP_SAUTH) & 3; 98 if (c) { 99 *s++ = 'A'; 100 s = gcap_string(s, c); 101 } 102 103 c = (caps >> CEPH_CAP_SLINK) & 3; 104 if (c) { 105 *s++ = 'L'; 106 s = gcap_string(s, c); 107 } 108 109 c = (caps >> CEPH_CAP_SXATTR) & 3; 110 if (c) { 111 *s++ = 'X'; 112 s = gcap_string(s, c); 113 } 114 115 c = caps >> CEPH_CAP_SFILE; 116 if (c) { 117 *s++ = 'F'; 118 s = gcap_string(s, c); 119 } 120 121 if (s == cap_str[i]) 122 *s++ = '-'; 123 *s = 0; 124 return cap_str[i]; 125 } 126 127 void ceph_caps_init(struct ceph_mds_client *mdsc) 128 { 129 INIT_LIST_HEAD(&mdsc->caps_list); 130 spin_lock_init(&mdsc->caps_list_lock); 131 } 132 133 void ceph_caps_finalize(struct ceph_mds_client *mdsc) 134 { 135 struct ceph_cap *cap; 136 137 spin_lock(&mdsc->caps_list_lock); 138 while (!list_empty(&mdsc->caps_list)) { 139 cap = list_first_entry(&mdsc->caps_list, 140 struct ceph_cap, caps_item); 141 list_del(&cap->caps_item); 142 kmem_cache_free(ceph_cap_cachep, cap); 143 } 144 mdsc->caps_total_count = 0; 145 mdsc->caps_avail_count = 0; 146 mdsc->caps_use_count = 0; 147 mdsc->caps_reserve_count = 0; 148 mdsc->caps_min_count = 0; 149 spin_unlock(&mdsc->caps_list_lock); 150 } 151 152 void ceph_adjust_caps_max_min(struct ceph_mds_client *mdsc, 153 struct ceph_mount_options *fsopt) 154 { 155 spin_lock(&mdsc->caps_list_lock); 156 mdsc->caps_min_count = fsopt->max_readdir; 157 if (mdsc->caps_min_count < 1024) 158 mdsc->caps_min_count = 1024; 159 mdsc->caps_use_max = fsopt->caps_max; 160 if (mdsc->caps_use_max > 0 && 161 mdsc->caps_use_max < mdsc->caps_min_count) 162 mdsc->caps_use_max = mdsc->caps_min_count; 163 spin_unlock(&mdsc->caps_list_lock); 164 } 165 166 static void __ceph_unreserve_caps(struct ceph_mds_client *mdsc, int nr_caps) 167 { 168 struct ceph_cap *cap; 169 int i; 170 171 if (nr_caps) { 172 BUG_ON(mdsc->caps_reserve_count < nr_caps); 173 mdsc->caps_reserve_count -= nr_caps; 174 if (mdsc->caps_avail_count >= 175 mdsc->caps_reserve_count + mdsc->caps_min_count) { 176 mdsc->caps_total_count -= nr_caps; 177 for (i = 0; i < nr_caps; i++) { 178 cap = list_first_entry(&mdsc->caps_list, 179 struct ceph_cap, caps_item); 180 list_del(&cap->caps_item); 181 kmem_cache_free(ceph_cap_cachep, cap); 182 } 183 } else { 184 mdsc->caps_avail_count += nr_caps; 185 } 186 187 dout("%s: caps %d = %d used + %d resv + %d avail\n", 188 __func__, 189 mdsc->caps_total_count, mdsc->caps_use_count, 190 mdsc->caps_reserve_count, mdsc->caps_avail_count); 191 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + 192 mdsc->caps_reserve_count + 193 mdsc->caps_avail_count); 194 } 195 } 196 197 /* 198 * Called under mdsc->mutex. 199 */ 200 int ceph_reserve_caps(struct ceph_mds_client *mdsc, 201 struct ceph_cap_reservation *ctx, int need) 202 { 203 int i, j; 204 struct ceph_cap *cap; 205 int have; 206 int alloc = 0; 207 int max_caps; 208 int err = 0; 209 bool trimmed = false; 210 struct ceph_mds_session *s; 211 LIST_HEAD(newcaps); 212 213 dout("reserve caps ctx=%p need=%d\n", ctx, need); 214 215 /* first reserve any caps that are already allocated */ 216 spin_lock(&mdsc->caps_list_lock); 217 if (mdsc->caps_avail_count >= need) 218 have = need; 219 else 220 have = mdsc->caps_avail_count; 221 mdsc->caps_avail_count -= have; 222 mdsc->caps_reserve_count += have; 223 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + 224 mdsc->caps_reserve_count + 225 mdsc->caps_avail_count); 226 spin_unlock(&mdsc->caps_list_lock); 227 228 for (i = have; i < need; ) { 229 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS); 230 if (cap) { 231 list_add(&cap->caps_item, &newcaps); 232 alloc++; 233 i++; 234 continue; 235 } 236 237 if (!trimmed) { 238 for (j = 0; j < mdsc->max_sessions; j++) { 239 s = __ceph_lookup_mds_session(mdsc, j); 240 if (!s) 241 continue; 242 mutex_unlock(&mdsc->mutex); 243 244 mutex_lock(&s->s_mutex); 245 max_caps = s->s_nr_caps - (need - i); 246 ceph_trim_caps(mdsc, s, max_caps); 247 mutex_unlock(&s->s_mutex); 248 249 ceph_put_mds_session(s); 250 mutex_lock(&mdsc->mutex); 251 } 252 trimmed = true; 253 254 spin_lock(&mdsc->caps_list_lock); 255 if (mdsc->caps_avail_count) { 256 int more_have; 257 if (mdsc->caps_avail_count >= need - i) 258 more_have = need - i; 259 else 260 more_have = mdsc->caps_avail_count; 261 262 i += more_have; 263 have += more_have; 264 mdsc->caps_avail_count -= more_have; 265 mdsc->caps_reserve_count += more_have; 266 267 } 268 spin_unlock(&mdsc->caps_list_lock); 269 270 continue; 271 } 272 273 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n", 274 ctx, need, have + alloc); 275 err = -ENOMEM; 276 break; 277 } 278 279 if (!err) { 280 BUG_ON(have + alloc != need); 281 ctx->count = need; 282 ctx->used = 0; 283 } 284 285 spin_lock(&mdsc->caps_list_lock); 286 mdsc->caps_total_count += alloc; 287 mdsc->caps_reserve_count += alloc; 288 list_splice(&newcaps, &mdsc->caps_list); 289 290 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + 291 mdsc->caps_reserve_count + 292 mdsc->caps_avail_count); 293 294 if (err) 295 __ceph_unreserve_caps(mdsc, have + alloc); 296 297 spin_unlock(&mdsc->caps_list_lock); 298 299 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n", 300 ctx, mdsc->caps_total_count, mdsc->caps_use_count, 301 mdsc->caps_reserve_count, mdsc->caps_avail_count); 302 return err; 303 } 304 305 void ceph_unreserve_caps(struct ceph_mds_client *mdsc, 306 struct ceph_cap_reservation *ctx) 307 { 308 bool reclaim = false; 309 if (!ctx->count) 310 return; 311 312 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count); 313 spin_lock(&mdsc->caps_list_lock); 314 __ceph_unreserve_caps(mdsc, ctx->count); 315 ctx->count = 0; 316 317 if (mdsc->caps_use_max > 0 && 318 mdsc->caps_use_count > mdsc->caps_use_max) 319 reclaim = true; 320 spin_unlock(&mdsc->caps_list_lock); 321 322 if (reclaim) 323 ceph_reclaim_caps_nr(mdsc, ctx->used); 324 } 325 326 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc, 327 struct ceph_cap_reservation *ctx) 328 { 329 struct ceph_cap *cap = NULL; 330 331 /* temporary, until we do something about cap import/export */ 332 if (!ctx) { 333 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS); 334 if (cap) { 335 spin_lock(&mdsc->caps_list_lock); 336 mdsc->caps_use_count++; 337 mdsc->caps_total_count++; 338 spin_unlock(&mdsc->caps_list_lock); 339 } else { 340 spin_lock(&mdsc->caps_list_lock); 341 if (mdsc->caps_avail_count) { 342 BUG_ON(list_empty(&mdsc->caps_list)); 343 344 mdsc->caps_avail_count--; 345 mdsc->caps_use_count++; 346 cap = list_first_entry(&mdsc->caps_list, 347 struct ceph_cap, caps_item); 348 list_del(&cap->caps_item); 349 350 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + 351 mdsc->caps_reserve_count + mdsc->caps_avail_count); 352 } 353 spin_unlock(&mdsc->caps_list_lock); 354 } 355 356 return cap; 357 } 358 359 spin_lock(&mdsc->caps_list_lock); 360 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n", 361 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count, 362 mdsc->caps_reserve_count, mdsc->caps_avail_count); 363 BUG_ON(!ctx->count); 364 BUG_ON(ctx->count > mdsc->caps_reserve_count); 365 BUG_ON(list_empty(&mdsc->caps_list)); 366 367 ctx->count--; 368 ctx->used++; 369 mdsc->caps_reserve_count--; 370 mdsc->caps_use_count++; 371 372 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item); 373 list_del(&cap->caps_item); 374 375 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + 376 mdsc->caps_reserve_count + mdsc->caps_avail_count); 377 spin_unlock(&mdsc->caps_list_lock); 378 return cap; 379 } 380 381 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap) 382 { 383 spin_lock(&mdsc->caps_list_lock); 384 dout("put_cap %p %d = %d used + %d resv + %d avail\n", 385 cap, mdsc->caps_total_count, mdsc->caps_use_count, 386 mdsc->caps_reserve_count, mdsc->caps_avail_count); 387 mdsc->caps_use_count--; 388 /* 389 * Keep some preallocated caps around (ceph_min_count), to 390 * avoid lots of free/alloc churn. 391 */ 392 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count + 393 mdsc->caps_min_count) { 394 mdsc->caps_total_count--; 395 kmem_cache_free(ceph_cap_cachep, cap); 396 } else { 397 mdsc->caps_avail_count++; 398 list_add(&cap->caps_item, &mdsc->caps_list); 399 } 400 401 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + 402 mdsc->caps_reserve_count + mdsc->caps_avail_count); 403 spin_unlock(&mdsc->caps_list_lock); 404 } 405 406 void ceph_reservation_status(struct ceph_fs_client *fsc, 407 int *total, int *avail, int *used, int *reserved, 408 int *min) 409 { 410 struct ceph_mds_client *mdsc = fsc->mdsc; 411 412 spin_lock(&mdsc->caps_list_lock); 413 414 if (total) 415 *total = mdsc->caps_total_count; 416 if (avail) 417 *avail = mdsc->caps_avail_count; 418 if (used) 419 *used = mdsc->caps_use_count; 420 if (reserved) 421 *reserved = mdsc->caps_reserve_count; 422 if (min) 423 *min = mdsc->caps_min_count; 424 425 spin_unlock(&mdsc->caps_list_lock); 426 } 427 428 /* 429 * Find ceph_cap for given mds, if any. 430 * 431 * Called with i_ceph_lock held. 432 */ 433 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds) 434 { 435 struct ceph_cap *cap; 436 struct rb_node *n = ci->i_caps.rb_node; 437 438 while (n) { 439 cap = rb_entry(n, struct ceph_cap, ci_node); 440 if (mds < cap->mds) 441 n = n->rb_left; 442 else if (mds > cap->mds) 443 n = n->rb_right; 444 else 445 return cap; 446 } 447 return NULL; 448 } 449 450 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds) 451 { 452 struct ceph_cap *cap; 453 454 spin_lock(&ci->i_ceph_lock); 455 cap = __get_cap_for_mds(ci, mds); 456 spin_unlock(&ci->i_ceph_lock); 457 return cap; 458 } 459 460 /* 461 * Called under i_ceph_lock. 462 */ 463 static void __insert_cap_node(struct ceph_inode_info *ci, 464 struct ceph_cap *new) 465 { 466 struct rb_node **p = &ci->i_caps.rb_node; 467 struct rb_node *parent = NULL; 468 struct ceph_cap *cap = NULL; 469 470 while (*p) { 471 parent = *p; 472 cap = rb_entry(parent, struct ceph_cap, ci_node); 473 if (new->mds < cap->mds) 474 p = &(*p)->rb_left; 475 else if (new->mds > cap->mds) 476 p = &(*p)->rb_right; 477 else 478 BUG(); 479 } 480 481 rb_link_node(&new->ci_node, parent, p); 482 rb_insert_color(&new->ci_node, &ci->i_caps); 483 } 484 485 /* 486 * (re)set cap hold timeouts, which control the delayed release 487 * of unused caps back to the MDS. Should be called on cap use. 488 */ 489 static void __cap_set_timeouts(struct ceph_mds_client *mdsc, 490 struct ceph_inode_info *ci) 491 { 492 struct ceph_mount_options *opt = mdsc->fsc->mount_options; 493 ci->i_hold_caps_max = round_jiffies(jiffies + 494 opt->caps_wanted_delay_max * HZ); 495 dout("__cap_set_timeouts %p %lu\n", &ci->vfs_inode, 496 ci->i_hold_caps_max - jiffies); 497 } 498 499 /* 500 * (Re)queue cap at the end of the delayed cap release list. 501 * 502 * If I_FLUSH is set, leave the inode at the front of the list. 503 * 504 * Caller holds i_ceph_lock 505 * -> we take mdsc->cap_delay_lock 506 */ 507 static void __cap_delay_requeue(struct ceph_mds_client *mdsc, 508 struct ceph_inode_info *ci) 509 { 510 dout("__cap_delay_requeue %p flags 0x%lx at %lu\n", &ci->vfs_inode, 511 ci->i_ceph_flags, ci->i_hold_caps_max); 512 if (!mdsc->stopping) { 513 spin_lock(&mdsc->cap_delay_lock); 514 if (!list_empty(&ci->i_cap_delay_list)) { 515 if (ci->i_ceph_flags & CEPH_I_FLUSH) 516 goto no_change; 517 list_del_init(&ci->i_cap_delay_list); 518 } 519 __cap_set_timeouts(mdsc, ci); 520 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list); 521 no_change: 522 spin_unlock(&mdsc->cap_delay_lock); 523 } 524 } 525 526 /* 527 * Queue an inode for immediate writeback. Mark inode with I_FLUSH, 528 * indicating we should send a cap message to flush dirty metadata 529 * asap, and move to the front of the delayed cap list. 530 */ 531 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc, 532 struct ceph_inode_info *ci) 533 { 534 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode); 535 spin_lock(&mdsc->cap_delay_lock); 536 ci->i_ceph_flags |= CEPH_I_FLUSH; 537 if (!list_empty(&ci->i_cap_delay_list)) 538 list_del_init(&ci->i_cap_delay_list); 539 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list); 540 spin_unlock(&mdsc->cap_delay_lock); 541 } 542 543 /* 544 * Cancel delayed work on cap. 545 * 546 * Caller must hold i_ceph_lock. 547 */ 548 static void __cap_delay_cancel(struct ceph_mds_client *mdsc, 549 struct ceph_inode_info *ci) 550 { 551 dout("__cap_delay_cancel %p\n", &ci->vfs_inode); 552 if (list_empty(&ci->i_cap_delay_list)) 553 return; 554 spin_lock(&mdsc->cap_delay_lock); 555 list_del_init(&ci->i_cap_delay_list); 556 spin_unlock(&mdsc->cap_delay_lock); 557 } 558 559 /* Common issue checks for add_cap, handle_cap_grant. */ 560 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap, 561 unsigned issued) 562 { 563 unsigned had = __ceph_caps_issued(ci, NULL); 564 565 lockdep_assert_held(&ci->i_ceph_lock); 566 567 /* 568 * Each time we receive FILE_CACHE anew, we increment 569 * i_rdcache_gen. 570 */ 571 if (S_ISREG(ci->vfs_inode.i_mode) && 572 (issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) && 573 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) { 574 ci->i_rdcache_gen++; 575 } 576 577 /* 578 * If FILE_SHARED is newly issued, mark dir not complete. We don't 579 * know what happened to this directory while we didn't have the cap. 580 * If FILE_SHARED is being revoked, also mark dir not complete. It 581 * stops on-going cached readdir. 582 */ 583 if ((issued & CEPH_CAP_FILE_SHARED) != (had & CEPH_CAP_FILE_SHARED)) { 584 if (issued & CEPH_CAP_FILE_SHARED) 585 atomic_inc(&ci->i_shared_gen); 586 if (S_ISDIR(ci->vfs_inode.i_mode)) { 587 dout(" marking %p NOT complete\n", &ci->vfs_inode); 588 __ceph_dir_clear_complete(ci); 589 } 590 } 591 592 /* Wipe saved layout if we're losing DIR_CREATE caps */ 593 if (S_ISDIR(ci->vfs_inode.i_mode) && (had & CEPH_CAP_DIR_CREATE) && 594 !(issued & CEPH_CAP_DIR_CREATE)) { 595 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns)); 596 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout)); 597 } 598 } 599 600 /** 601 * change_auth_cap_ses - move inode to appropriate lists when auth caps change 602 * @ci: inode to be moved 603 * @session: new auth caps session 604 */ 605 static void change_auth_cap_ses(struct ceph_inode_info *ci, 606 struct ceph_mds_session *session) 607 { 608 lockdep_assert_held(&ci->i_ceph_lock); 609 610 if (list_empty(&ci->i_dirty_item) && list_empty(&ci->i_flushing_item)) 611 return; 612 613 spin_lock(&session->s_mdsc->cap_dirty_lock); 614 if (!list_empty(&ci->i_dirty_item)) 615 list_move(&ci->i_dirty_item, &session->s_cap_dirty); 616 if (!list_empty(&ci->i_flushing_item)) 617 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing); 618 spin_unlock(&session->s_mdsc->cap_dirty_lock); 619 } 620 621 /* 622 * Add a capability under the given MDS session. 623 * 624 * Caller should hold session snap_rwsem (read) and ci->i_ceph_lock 625 * 626 * @fmode is the open file mode, if we are opening a file, otherwise 627 * it is < 0. (This is so we can atomically add the cap and add an 628 * open file reference to it.) 629 */ 630 void ceph_add_cap(struct inode *inode, 631 struct ceph_mds_session *session, u64 cap_id, 632 unsigned issued, unsigned wanted, 633 unsigned seq, unsigned mseq, u64 realmino, int flags, 634 struct ceph_cap **new_cap) 635 { 636 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 637 struct ceph_inode_info *ci = ceph_inode(inode); 638 struct ceph_cap *cap; 639 int mds = session->s_mds; 640 int actual_wanted; 641 u32 gen; 642 643 lockdep_assert_held(&ci->i_ceph_lock); 644 645 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode, 646 session->s_mds, cap_id, ceph_cap_string(issued), seq); 647 648 spin_lock(&session->s_gen_ttl_lock); 649 gen = session->s_cap_gen; 650 spin_unlock(&session->s_gen_ttl_lock); 651 652 cap = __get_cap_for_mds(ci, mds); 653 if (!cap) { 654 cap = *new_cap; 655 *new_cap = NULL; 656 657 cap->issued = 0; 658 cap->implemented = 0; 659 cap->mds = mds; 660 cap->mds_wanted = 0; 661 cap->mseq = 0; 662 663 cap->ci = ci; 664 __insert_cap_node(ci, cap); 665 666 /* add to session cap list */ 667 cap->session = session; 668 spin_lock(&session->s_cap_lock); 669 list_add_tail(&cap->session_caps, &session->s_caps); 670 session->s_nr_caps++; 671 atomic64_inc(&mdsc->metric.total_caps); 672 spin_unlock(&session->s_cap_lock); 673 } else { 674 spin_lock(&session->s_cap_lock); 675 list_move_tail(&cap->session_caps, &session->s_caps); 676 spin_unlock(&session->s_cap_lock); 677 678 if (cap->cap_gen < gen) 679 cap->issued = cap->implemented = CEPH_CAP_PIN; 680 681 /* 682 * auth mds of the inode changed. we received the cap export 683 * message, but still haven't received the cap import message. 684 * handle_cap_export() updated the new auth MDS' cap. 685 * 686 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing 687 * a message that was send before the cap import message. So 688 * don't remove caps. 689 */ 690 if (ceph_seq_cmp(seq, cap->seq) <= 0) { 691 WARN_ON(cap != ci->i_auth_cap); 692 WARN_ON(cap->cap_id != cap_id); 693 seq = cap->seq; 694 mseq = cap->mseq; 695 issued |= cap->issued; 696 flags |= CEPH_CAP_FLAG_AUTH; 697 } 698 } 699 700 if (!ci->i_snap_realm || 701 ((flags & CEPH_CAP_FLAG_AUTH) && 702 realmino != (u64)-1 && ci->i_snap_realm->ino != realmino)) { 703 /* 704 * add this inode to the appropriate snap realm 705 */ 706 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc, 707 realmino); 708 if (realm) { 709 struct ceph_snap_realm *oldrealm = ci->i_snap_realm; 710 if (oldrealm) { 711 spin_lock(&oldrealm->inodes_with_caps_lock); 712 list_del_init(&ci->i_snap_realm_item); 713 spin_unlock(&oldrealm->inodes_with_caps_lock); 714 } 715 716 spin_lock(&realm->inodes_with_caps_lock); 717 list_add(&ci->i_snap_realm_item, 718 &realm->inodes_with_caps); 719 ci->i_snap_realm = realm; 720 if (realm->ino == ci->i_vino.ino) 721 realm->inode = inode; 722 spin_unlock(&realm->inodes_with_caps_lock); 723 724 if (oldrealm) 725 ceph_put_snap_realm(mdsc, oldrealm); 726 } else { 727 pr_err("ceph_add_cap: couldn't find snap realm %llx\n", 728 realmino); 729 WARN_ON(!realm); 730 } 731 } 732 733 __check_cap_issue(ci, cap, issued); 734 735 /* 736 * If we are issued caps we don't want, or the mds' wanted 737 * value appears to be off, queue a check so we'll release 738 * later and/or update the mds wanted value. 739 */ 740 actual_wanted = __ceph_caps_wanted(ci); 741 if ((wanted & ~actual_wanted) || 742 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) { 743 dout(" issued %s, mds wanted %s, actual %s, queueing\n", 744 ceph_cap_string(issued), ceph_cap_string(wanted), 745 ceph_cap_string(actual_wanted)); 746 __cap_delay_requeue(mdsc, ci); 747 } 748 749 if (flags & CEPH_CAP_FLAG_AUTH) { 750 if (!ci->i_auth_cap || 751 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) { 752 if (ci->i_auth_cap && 753 ci->i_auth_cap->session != cap->session) 754 change_auth_cap_ses(ci, cap->session); 755 ci->i_auth_cap = cap; 756 cap->mds_wanted = wanted; 757 } 758 } else { 759 WARN_ON(ci->i_auth_cap == cap); 760 } 761 762 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n", 763 inode, ceph_vinop(inode), cap, ceph_cap_string(issued), 764 ceph_cap_string(issued|cap->issued), seq, mds); 765 cap->cap_id = cap_id; 766 cap->issued = issued; 767 cap->implemented |= issued; 768 if (ceph_seq_cmp(mseq, cap->mseq) > 0) 769 cap->mds_wanted = wanted; 770 else 771 cap->mds_wanted |= wanted; 772 cap->seq = seq; 773 cap->issue_seq = seq; 774 cap->mseq = mseq; 775 cap->cap_gen = gen; 776 } 777 778 /* 779 * Return true if cap has not timed out and belongs to the current 780 * generation of the MDS session (i.e. has not gone 'stale' due to 781 * us losing touch with the mds). 782 */ 783 static int __cap_is_valid(struct ceph_cap *cap) 784 { 785 unsigned long ttl; 786 u32 gen; 787 788 spin_lock(&cap->session->s_gen_ttl_lock); 789 gen = cap->session->s_cap_gen; 790 ttl = cap->session->s_cap_ttl; 791 spin_unlock(&cap->session->s_gen_ttl_lock); 792 793 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) { 794 dout("__cap_is_valid %p cap %p issued %s " 795 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode, 796 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen); 797 return 0; 798 } 799 800 return 1; 801 } 802 803 /* 804 * Return set of valid cap bits issued to us. Note that caps time 805 * out, and may be invalidated in bulk if the client session times out 806 * and session->s_cap_gen is bumped. 807 */ 808 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented) 809 { 810 int have = ci->i_snap_caps; 811 struct ceph_cap *cap; 812 struct rb_node *p; 813 814 if (implemented) 815 *implemented = 0; 816 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 817 cap = rb_entry(p, struct ceph_cap, ci_node); 818 if (!__cap_is_valid(cap)) 819 continue; 820 dout("__ceph_caps_issued %p cap %p issued %s\n", 821 &ci->vfs_inode, cap, ceph_cap_string(cap->issued)); 822 have |= cap->issued; 823 if (implemented) 824 *implemented |= cap->implemented; 825 } 826 /* 827 * exclude caps issued by non-auth MDS, but are been revoking 828 * by the auth MDS. The non-auth MDS should be revoking/exporting 829 * these caps, but the message is delayed. 830 */ 831 if (ci->i_auth_cap) { 832 cap = ci->i_auth_cap; 833 have &= ~cap->implemented | cap->issued; 834 } 835 return have; 836 } 837 838 /* 839 * Get cap bits issued by caps other than @ocap 840 */ 841 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap) 842 { 843 int have = ci->i_snap_caps; 844 struct ceph_cap *cap; 845 struct rb_node *p; 846 847 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 848 cap = rb_entry(p, struct ceph_cap, ci_node); 849 if (cap == ocap) 850 continue; 851 if (!__cap_is_valid(cap)) 852 continue; 853 have |= cap->issued; 854 } 855 return have; 856 } 857 858 /* 859 * Move a cap to the end of the LRU (oldest caps at list head, newest 860 * at list tail). 861 */ 862 static void __touch_cap(struct ceph_cap *cap) 863 { 864 struct ceph_mds_session *s = cap->session; 865 866 spin_lock(&s->s_cap_lock); 867 if (!s->s_cap_iterator) { 868 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap, 869 s->s_mds); 870 list_move_tail(&cap->session_caps, &s->s_caps); 871 } else { 872 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n", 873 &cap->ci->vfs_inode, cap, s->s_mds); 874 } 875 spin_unlock(&s->s_cap_lock); 876 } 877 878 /* 879 * Check if we hold the given mask. If so, move the cap(s) to the 880 * front of their respective LRUs. (This is the preferred way for 881 * callers to check for caps they want.) 882 */ 883 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch) 884 { 885 struct ceph_cap *cap; 886 struct rb_node *p; 887 int have = ci->i_snap_caps; 888 889 if ((have & mask) == mask) { 890 dout("__ceph_caps_issued_mask ino 0x%llx snap issued %s" 891 " (mask %s)\n", ceph_ino(&ci->vfs_inode), 892 ceph_cap_string(have), 893 ceph_cap_string(mask)); 894 return 1; 895 } 896 897 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 898 cap = rb_entry(p, struct ceph_cap, ci_node); 899 if (!__cap_is_valid(cap)) 900 continue; 901 if ((cap->issued & mask) == mask) { 902 dout("__ceph_caps_issued_mask ino 0x%llx cap %p issued %s" 903 " (mask %s)\n", ceph_ino(&ci->vfs_inode), cap, 904 ceph_cap_string(cap->issued), 905 ceph_cap_string(mask)); 906 if (touch) 907 __touch_cap(cap); 908 return 1; 909 } 910 911 /* does a combination of caps satisfy mask? */ 912 have |= cap->issued; 913 if ((have & mask) == mask) { 914 dout("__ceph_caps_issued_mask ino 0x%llx combo issued %s" 915 " (mask %s)\n", ceph_ino(&ci->vfs_inode), 916 ceph_cap_string(cap->issued), 917 ceph_cap_string(mask)); 918 if (touch) { 919 struct rb_node *q; 920 921 /* touch this + preceding caps */ 922 __touch_cap(cap); 923 for (q = rb_first(&ci->i_caps); q != p; 924 q = rb_next(q)) { 925 cap = rb_entry(q, struct ceph_cap, 926 ci_node); 927 if (!__cap_is_valid(cap)) 928 continue; 929 if (cap->issued & mask) 930 __touch_cap(cap); 931 } 932 } 933 return 1; 934 } 935 } 936 937 return 0; 938 } 939 940 int __ceph_caps_issued_mask_metric(struct ceph_inode_info *ci, int mask, 941 int touch) 942 { 943 struct ceph_fs_client *fsc = ceph_sb_to_client(ci->vfs_inode.i_sb); 944 int r; 945 946 r = __ceph_caps_issued_mask(ci, mask, touch); 947 if (r) 948 ceph_update_cap_hit(&fsc->mdsc->metric); 949 else 950 ceph_update_cap_mis(&fsc->mdsc->metric); 951 return r; 952 } 953 954 /* 955 * Return true if mask caps are currently being revoked by an MDS. 956 */ 957 int __ceph_caps_revoking_other(struct ceph_inode_info *ci, 958 struct ceph_cap *ocap, int mask) 959 { 960 struct ceph_cap *cap; 961 struct rb_node *p; 962 963 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 964 cap = rb_entry(p, struct ceph_cap, ci_node); 965 if (cap != ocap && 966 (cap->implemented & ~cap->issued & mask)) 967 return 1; 968 } 969 return 0; 970 } 971 972 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask) 973 { 974 struct inode *inode = &ci->vfs_inode; 975 int ret; 976 977 spin_lock(&ci->i_ceph_lock); 978 ret = __ceph_caps_revoking_other(ci, NULL, mask); 979 spin_unlock(&ci->i_ceph_lock); 980 dout("ceph_caps_revoking %p %s = %d\n", inode, 981 ceph_cap_string(mask), ret); 982 return ret; 983 } 984 985 int __ceph_caps_used(struct ceph_inode_info *ci) 986 { 987 int used = 0; 988 if (ci->i_pin_ref) 989 used |= CEPH_CAP_PIN; 990 if (ci->i_rd_ref) 991 used |= CEPH_CAP_FILE_RD; 992 if (ci->i_rdcache_ref || 993 (S_ISREG(ci->vfs_inode.i_mode) && 994 ci->vfs_inode.i_data.nrpages)) 995 used |= CEPH_CAP_FILE_CACHE; 996 if (ci->i_wr_ref) 997 used |= CEPH_CAP_FILE_WR; 998 if (ci->i_wb_ref || ci->i_wrbuffer_ref) 999 used |= CEPH_CAP_FILE_BUFFER; 1000 if (ci->i_fx_ref) 1001 used |= CEPH_CAP_FILE_EXCL; 1002 return used; 1003 } 1004 1005 #define FMODE_WAIT_BIAS 1000 1006 1007 /* 1008 * wanted, by virtue of open file modes 1009 */ 1010 int __ceph_caps_file_wanted(struct ceph_inode_info *ci) 1011 { 1012 const int PIN_SHIFT = ffs(CEPH_FILE_MODE_PIN); 1013 const int RD_SHIFT = ffs(CEPH_FILE_MODE_RD); 1014 const int WR_SHIFT = ffs(CEPH_FILE_MODE_WR); 1015 const int LAZY_SHIFT = ffs(CEPH_FILE_MODE_LAZY); 1016 struct ceph_mount_options *opt = 1017 ceph_inode_to_client(&ci->vfs_inode)->mount_options; 1018 unsigned long used_cutoff = jiffies - opt->caps_wanted_delay_max * HZ; 1019 unsigned long idle_cutoff = jiffies - opt->caps_wanted_delay_min * HZ; 1020 1021 if (S_ISDIR(ci->vfs_inode.i_mode)) { 1022 int want = 0; 1023 1024 /* use used_cutoff here, to keep dir's wanted caps longer */ 1025 if (ci->i_nr_by_mode[RD_SHIFT] > 0 || 1026 time_after(ci->i_last_rd, used_cutoff)) 1027 want |= CEPH_CAP_ANY_SHARED; 1028 1029 if (ci->i_nr_by_mode[WR_SHIFT] > 0 || 1030 time_after(ci->i_last_wr, used_cutoff)) { 1031 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL; 1032 if (opt->flags & CEPH_MOUNT_OPT_ASYNC_DIROPS) 1033 want |= CEPH_CAP_ANY_DIR_OPS; 1034 } 1035 1036 if (want || ci->i_nr_by_mode[PIN_SHIFT] > 0) 1037 want |= CEPH_CAP_PIN; 1038 1039 return want; 1040 } else { 1041 int bits = 0; 1042 1043 if (ci->i_nr_by_mode[RD_SHIFT] > 0) { 1044 if (ci->i_nr_by_mode[RD_SHIFT] >= FMODE_WAIT_BIAS || 1045 time_after(ci->i_last_rd, used_cutoff)) 1046 bits |= 1 << RD_SHIFT; 1047 } else if (time_after(ci->i_last_rd, idle_cutoff)) { 1048 bits |= 1 << RD_SHIFT; 1049 } 1050 1051 if (ci->i_nr_by_mode[WR_SHIFT] > 0) { 1052 if (ci->i_nr_by_mode[WR_SHIFT] >= FMODE_WAIT_BIAS || 1053 time_after(ci->i_last_wr, used_cutoff)) 1054 bits |= 1 << WR_SHIFT; 1055 } else if (time_after(ci->i_last_wr, idle_cutoff)) { 1056 bits |= 1 << WR_SHIFT; 1057 } 1058 1059 /* check lazyio only when read/write is wanted */ 1060 if ((bits & (CEPH_FILE_MODE_RDWR << 1)) && 1061 ci->i_nr_by_mode[LAZY_SHIFT] > 0) 1062 bits |= 1 << LAZY_SHIFT; 1063 1064 return bits ? ceph_caps_for_mode(bits >> 1) : 0; 1065 } 1066 } 1067 1068 /* 1069 * wanted, by virtue of open file modes AND cap refs (buffered/cached data) 1070 */ 1071 int __ceph_caps_wanted(struct ceph_inode_info *ci) 1072 { 1073 int w = __ceph_caps_file_wanted(ci) | __ceph_caps_used(ci); 1074 if (S_ISDIR(ci->vfs_inode.i_mode)) { 1075 /* we want EXCL if holding caps of dir ops */ 1076 if (w & CEPH_CAP_ANY_DIR_OPS) 1077 w |= CEPH_CAP_FILE_EXCL; 1078 } else { 1079 /* we want EXCL if dirty data */ 1080 if (w & CEPH_CAP_FILE_BUFFER) 1081 w |= CEPH_CAP_FILE_EXCL; 1082 } 1083 return w; 1084 } 1085 1086 /* 1087 * Return caps we have registered with the MDS(s) as 'wanted'. 1088 */ 1089 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check) 1090 { 1091 struct ceph_cap *cap; 1092 struct rb_node *p; 1093 int mds_wanted = 0; 1094 1095 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 1096 cap = rb_entry(p, struct ceph_cap, ci_node); 1097 if (check && !__cap_is_valid(cap)) 1098 continue; 1099 if (cap == ci->i_auth_cap) 1100 mds_wanted |= cap->mds_wanted; 1101 else 1102 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR); 1103 } 1104 return mds_wanted; 1105 } 1106 1107 int ceph_is_any_caps(struct inode *inode) 1108 { 1109 struct ceph_inode_info *ci = ceph_inode(inode); 1110 int ret; 1111 1112 spin_lock(&ci->i_ceph_lock); 1113 ret = __ceph_is_any_real_caps(ci); 1114 spin_unlock(&ci->i_ceph_lock); 1115 1116 return ret; 1117 } 1118 1119 static void drop_inode_snap_realm(struct ceph_inode_info *ci) 1120 { 1121 struct ceph_snap_realm *realm = ci->i_snap_realm; 1122 spin_lock(&realm->inodes_with_caps_lock); 1123 list_del_init(&ci->i_snap_realm_item); 1124 ci->i_snap_realm_counter++; 1125 ci->i_snap_realm = NULL; 1126 if (realm->ino == ci->i_vino.ino) 1127 realm->inode = NULL; 1128 spin_unlock(&realm->inodes_with_caps_lock); 1129 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc, 1130 realm); 1131 } 1132 1133 /* 1134 * Remove a cap. Take steps to deal with a racing iterate_session_caps. 1135 * 1136 * caller should hold i_ceph_lock. 1137 * caller will not hold session s_mutex if called from destroy_inode. 1138 */ 1139 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release) 1140 { 1141 struct ceph_mds_session *session = cap->session; 1142 struct ceph_inode_info *ci = cap->ci; 1143 struct ceph_mds_client *mdsc = 1144 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc; 1145 int removed = 0; 1146 1147 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode); 1148 1149 /* remove from inode's cap rbtree, and clear auth cap */ 1150 rb_erase(&cap->ci_node, &ci->i_caps); 1151 if (ci->i_auth_cap == cap) { 1152 WARN_ON_ONCE(!list_empty(&ci->i_dirty_item)); 1153 ci->i_auth_cap = NULL; 1154 } 1155 1156 /* remove from session list */ 1157 spin_lock(&session->s_cap_lock); 1158 if (session->s_cap_iterator == cap) { 1159 /* not yet, we are iterating over this very cap */ 1160 dout("__ceph_remove_cap delaying %p removal from session %p\n", 1161 cap, cap->session); 1162 } else { 1163 list_del_init(&cap->session_caps); 1164 session->s_nr_caps--; 1165 atomic64_dec(&mdsc->metric.total_caps); 1166 cap->session = NULL; 1167 removed = 1; 1168 } 1169 /* protect backpointer with s_cap_lock: see iterate_session_caps */ 1170 cap->ci = NULL; 1171 1172 /* 1173 * s_cap_reconnect is protected by s_cap_lock. no one changes 1174 * s_cap_gen while session is in the reconnect state. 1175 */ 1176 if (queue_release && 1177 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) { 1178 cap->queue_release = 1; 1179 if (removed) { 1180 __ceph_queue_cap_release(session, cap); 1181 removed = 0; 1182 } 1183 } else { 1184 cap->queue_release = 0; 1185 } 1186 cap->cap_ino = ci->i_vino.ino; 1187 1188 spin_unlock(&session->s_cap_lock); 1189 1190 if (removed) 1191 ceph_put_cap(mdsc, cap); 1192 1193 if (!__ceph_is_any_real_caps(ci)) { 1194 /* when reconnect denied, we remove session caps forcibly, 1195 * i_wr_ref can be non-zero. If there are ongoing write, 1196 * keep i_snap_realm. 1197 */ 1198 if (ci->i_wr_ref == 0 && ci->i_snap_realm) 1199 drop_inode_snap_realm(ci); 1200 1201 __cap_delay_cancel(mdsc, ci); 1202 } 1203 } 1204 1205 struct cap_msg_args { 1206 struct ceph_mds_session *session; 1207 u64 ino, cid, follows; 1208 u64 flush_tid, oldest_flush_tid, size, max_size; 1209 u64 xattr_version; 1210 u64 change_attr; 1211 struct ceph_buffer *xattr_buf; 1212 struct ceph_buffer *old_xattr_buf; 1213 struct timespec64 atime, mtime, ctime, btime; 1214 int op, caps, wanted, dirty; 1215 u32 seq, issue_seq, mseq, time_warp_seq; 1216 u32 flags; 1217 kuid_t uid; 1218 kgid_t gid; 1219 umode_t mode; 1220 bool inline_data; 1221 bool wake; 1222 }; 1223 1224 /* 1225 * cap struct size + flock buffer size + inline version + inline data size + 1226 * osd_epoch_barrier + oldest_flush_tid 1227 */ 1228 #define CAP_MSG_SIZE (sizeof(struct ceph_mds_caps) + \ 1229 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4) 1230 1231 /* Marshal up the cap msg to the MDS */ 1232 static void encode_cap_msg(struct ceph_msg *msg, struct cap_msg_args *arg) 1233 { 1234 struct ceph_mds_caps *fc; 1235 void *p; 1236 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc; 1237 1238 dout("%s %s %llx %llx caps %s wanted %s dirty %s seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu xattr_ver %llu xattr_len %d\n", 1239 __func__, ceph_cap_op_name(arg->op), arg->cid, arg->ino, 1240 ceph_cap_string(arg->caps), ceph_cap_string(arg->wanted), 1241 ceph_cap_string(arg->dirty), arg->seq, arg->issue_seq, 1242 arg->flush_tid, arg->oldest_flush_tid, arg->mseq, arg->follows, 1243 arg->size, arg->max_size, arg->xattr_version, 1244 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0); 1245 1246 msg->hdr.version = cpu_to_le16(10); 1247 msg->hdr.tid = cpu_to_le64(arg->flush_tid); 1248 1249 fc = msg->front.iov_base; 1250 memset(fc, 0, sizeof(*fc)); 1251 1252 fc->cap_id = cpu_to_le64(arg->cid); 1253 fc->op = cpu_to_le32(arg->op); 1254 fc->seq = cpu_to_le32(arg->seq); 1255 fc->issue_seq = cpu_to_le32(arg->issue_seq); 1256 fc->migrate_seq = cpu_to_le32(arg->mseq); 1257 fc->caps = cpu_to_le32(arg->caps); 1258 fc->wanted = cpu_to_le32(arg->wanted); 1259 fc->dirty = cpu_to_le32(arg->dirty); 1260 fc->ino = cpu_to_le64(arg->ino); 1261 fc->snap_follows = cpu_to_le64(arg->follows); 1262 1263 fc->size = cpu_to_le64(arg->size); 1264 fc->max_size = cpu_to_le64(arg->max_size); 1265 ceph_encode_timespec64(&fc->mtime, &arg->mtime); 1266 ceph_encode_timespec64(&fc->atime, &arg->atime); 1267 ceph_encode_timespec64(&fc->ctime, &arg->ctime); 1268 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq); 1269 1270 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid)); 1271 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid)); 1272 fc->mode = cpu_to_le32(arg->mode); 1273 1274 fc->xattr_version = cpu_to_le64(arg->xattr_version); 1275 if (arg->xattr_buf) { 1276 msg->middle = ceph_buffer_get(arg->xattr_buf); 1277 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len); 1278 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len); 1279 } 1280 1281 p = fc + 1; 1282 /* flock buffer size (version 2) */ 1283 ceph_encode_32(&p, 0); 1284 /* inline version (version 4) */ 1285 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE); 1286 /* inline data size */ 1287 ceph_encode_32(&p, 0); 1288 /* 1289 * osd_epoch_barrier (version 5) 1290 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in 1291 * case it was recently changed 1292 */ 1293 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier)); 1294 /* oldest_flush_tid (version 6) */ 1295 ceph_encode_64(&p, arg->oldest_flush_tid); 1296 1297 /* 1298 * caller_uid/caller_gid (version 7) 1299 * 1300 * Currently, we don't properly track which caller dirtied the caps 1301 * last, and force a flush of them when there is a conflict. For now, 1302 * just set this to 0:0, to emulate how the MDS has worked up to now. 1303 */ 1304 ceph_encode_32(&p, 0); 1305 ceph_encode_32(&p, 0); 1306 1307 /* pool namespace (version 8) (mds always ignores this) */ 1308 ceph_encode_32(&p, 0); 1309 1310 /* btime and change_attr (version 9) */ 1311 ceph_encode_timespec64(p, &arg->btime); 1312 p += sizeof(struct ceph_timespec); 1313 ceph_encode_64(&p, arg->change_attr); 1314 1315 /* Advisory flags (version 10) */ 1316 ceph_encode_32(&p, arg->flags); 1317 } 1318 1319 /* 1320 * Queue cap releases when an inode is dropped from our cache. 1321 */ 1322 void __ceph_remove_caps(struct ceph_inode_info *ci) 1323 { 1324 struct rb_node *p; 1325 1326 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU) 1327 * may call __ceph_caps_issued_mask() on a freeing inode. */ 1328 spin_lock(&ci->i_ceph_lock); 1329 p = rb_first(&ci->i_caps); 1330 while (p) { 1331 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node); 1332 p = rb_next(p); 1333 __ceph_remove_cap(cap, true); 1334 } 1335 spin_unlock(&ci->i_ceph_lock); 1336 } 1337 1338 /* 1339 * Prepare to send a cap message to an MDS. Update the cap state, and populate 1340 * the arg struct with the parameters that will need to be sent. This should 1341 * be done under the i_ceph_lock to guard against changes to cap state. 1342 * 1343 * Make note of max_size reported/requested from mds, revoked caps 1344 * that have now been implemented. 1345 */ 1346 static void __prep_cap(struct cap_msg_args *arg, struct ceph_cap *cap, 1347 int op, int flags, int used, int want, int retain, 1348 int flushing, u64 flush_tid, u64 oldest_flush_tid) 1349 { 1350 struct ceph_inode_info *ci = cap->ci; 1351 struct inode *inode = &ci->vfs_inode; 1352 int held, revoking; 1353 1354 lockdep_assert_held(&ci->i_ceph_lock); 1355 1356 held = cap->issued | cap->implemented; 1357 revoking = cap->implemented & ~cap->issued; 1358 retain &= ~revoking; 1359 1360 dout("%s %p cap %p session %p %s -> %s (revoking %s)\n", 1361 __func__, inode, cap, cap->session, 1362 ceph_cap_string(held), ceph_cap_string(held & retain), 1363 ceph_cap_string(revoking)); 1364 BUG_ON((retain & CEPH_CAP_PIN) == 0); 1365 1366 ci->i_ceph_flags &= ~CEPH_I_FLUSH; 1367 1368 cap->issued &= retain; /* drop bits we don't want */ 1369 /* 1370 * Wake up any waiters on wanted -> needed transition. This is due to 1371 * the weird transition from buffered to sync IO... we need to flush 1372 * dirty pages _before_ allowing sync writes to avoid reordering. 1373 */ 1374 arg->wake = cap->implemented & ~cap->issued; 1375 cap->implemented &= cap->issued | used; 1376 cap->mds_wanted = want; 1377 1378 arg->session = cap->session; 1379 arg->ino = ceph_vino(inode).ino; 1380 arg->cid = cap->cap_id; 1381 arg->follows = flushing ? ci->i_head_snapc->seq : 0; 1382 arg->flush_tid = flush_tid; 1383 arg->oldest_flush_tid = oldest_flush_tid; 1384 1385 arg->size = inode->i_size; 1386 ci->i_reported_size = arg->size; 1387 arg->max_size = ci->i_wanted_max_size; 1388 if (cap == ci->i_auth_cap) { 1389 if (want & CEPH_CAP_ANY_FILE_WR) 1390 ci->i_requested_max_size = arg->max_size; 1391 else 1392 ci->i_requested_max_size = 0; 1393 } 1394 1395 if (flushing & CEPH_CAP_XATTR_EXCL) { 1396 arg->old_xattr_buf = __ceph_build_xattrs_blob(ci); 1397 arg->xattr_version = ci->i_xattrs.version; 1398 arg->xattr_buf = ci->i_xattrs.blob; 1399 } else { 1400 arg->xattr_buf = NULL; 1401 arg->old_xattr_buf = NULL; 1402 } 1403 1404 arg->mtime = inode->i_mtime; 1405 arg->atime = inode->i_atime; 1406 arg->ctime = inode->i_ctime; 1407 arg->btime = ci->i_btime; 1408 arg->change_attr = inode_peek_iversion_raw(inode); 1409 1410 arg->op = op; 1411 arg->caps = cap->implemented; 1412 arg->wanted = want; 1413 arg->dirty = flushing; 1414 1415 arg->seq = cap->seq; 1416 arg->issue_seq = cap->issue_seq; 1417 arg->mseq = cap->mseq; 1418 arg->time_warp_seq = ci->i_time_warp_seq; 1419 1420 arg->uid = inode->i_uid; 1421 arg->gid = inode->i_gid; 1422 arg->mode = inode->i_mode; 1423 1424 arg->inline_data = ci->i_inline_version != CEPH_INLINE_NONE; 1425 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) && 1426 !list_empty(&ci->i_cap_snaps)) { 1427 struct ceph_cap_snap *capsnap; 1428 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) { 1429 if (capsnap->cap_flush.tid) 1430 break; 1431 if (capsnap->need_flush) { 1432 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP; 1433 break; 1434 } 1435 } 1436 } 1437 arg->flags = flags; 1438 } 1439 1440 /* 1441 * Send a cap msg on the given inode. 1442 * 1443 * Caller should hold snap_rwsem (read), s_mutex. 1444 */ 1445 static void __send_cap(struct cap_msg_args *arg, struct ceph_inode_info *ci) 1446 { 1447 struct ceph_msg *msg; 1448 struct inode *inode = &ci->vfs_inode; 1449 1450 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false); 1451 if (!msg) { 1452 pr_err("error allocating cap msg: ino (%llx.%llx) flushing %s tid %llu, requeuing cap.\n", 1453 ceph_vinop(inode), ceph_cap_string(arg->dirty), 1454 arg->flush_tid); 1455 spin_lock(&ci->i_ceph_lock); 1456 __cap_delay_requeue(arg->session->s_mdsc, ci); 1457 spin_unlock(&ci->i_ceph_lock); 1458 return; 1459 } 1460 1461 encode_cap_msg(msg, arg); 1462 ceph_con_send(&arg->session->s_con, msg); 1463 ceph_buffer_put(arg->old_xattr_buf); 1464 if (arg->wake) 1465 wake_up_all(&ci->i_cap_wq); 1466 } 1467 1468 static inline int __send_flush_snap(struct inode *inode, 1469 struct ceph_mds_session *session, 1470 struct ceph_cap_snap *capsnap, 1471 u32 mseq, u64 oldest_flush_tid) 1472 { 1473 struct cap_msg_args arg; 1474 struct ceph_msg *msg; 1475 1476 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false); 1477 if (!msg) 1478 return -ENOMEM; 1479 1480 arg.session = session; 1481 arg.ino = ceph_vino(inode).ino; 1482 arg.cid = 0; 1483 arg.follows = capsnap->follows; 1484 arg.flush_tid = capsnap->cap_flush.tid; 1485 arg.oldest_flush_tid = oldest_flush_tid; 1486 1487 arg.size = capsnap->size; 1488 arg.max_size = 0; 1489 arg.xattr_version = capsnap->xattr_version; 1490 arg.xattr_buf = capsnap->xattr_blob; 1491 arg.old_xattr_buf = NULL; 1492 1493 arg.atime = capsnap->atime; 1494 arg.mtime = capsnap->mtime; 1495 arg.ctime = capsnap->ctime; 1496 arg.btime = capsnap->btime; 1497 arg.change_attr = capsnap->change_attr; 1498 1499 arg.op = CEPH_CAP_OP_FLUSHSNAP; 1500 arg.caps = capsnap->issued; 1501 arg.wanted = 0; 1502 arg.dirty = capsnap->dirty; 1503 1504 arg.seq = 0; 1505 arg.issue_seq = 0; 1506 arg.mseq = mseq; 1507 arg.time_warp_seq = capsnap->time_warp_seq; 1508 1509 arg.uid = capsnap->uid; 1510 arg.gid = capsnap->gid; 1511 arg.mode = capsnap->mode; 1512 1513 arg.inline_data = capsnap->inline_data; 1514 arg.flags = 0; 1515 arg.wake = false; 1516 1517 encode_cap_msg(msg, &arg); 1518 ceph_con_send(&arg.session->s_con, msg); 1519 return 0; 1520 } 1521 1522 /* 1523 * When a snapshot is taken, clients accumulate dirty metadata on 1524 * inodes with capabilities in ceph_cap_snaps to describe the file 1525 * state at the time the snapshot was taken. This must be flushed 1526 * asynchronously back to the MDS once sync writes complete and dirty 1527 * data is written out. 1528 * 1529 * Called under i_ceph_lock. Takes s_mutex as needed. 1530 */ 1531 static void __ceph_flush_snaps(struct ceph_inode_info *ci, 1532 struct ceph_mds_session *session) 1533 __releases(ci->i_ceph_lock) 1534 __acquires(ci->i_ceph_lock) 1535 { 1536 struct inode *inode = &ci->vfs_inode; 1537 struct ceph_mds_client *mdsc = session->s_mdsc; 1538 struct ceph_cap_snap *capsnap; 1539 u64 oldest_flush_tid = 0; 1540 u64 first_tid = 1, last_tid = 0; 1541 1542 dout("__flush_snaps %p session %p\n", inode, session); 1543 1544 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 1545 /* 1546 * we need to wait for sync writes to complete and for dirty 1547 * pages to be written out. 1548 */ 1549 if (capsnap->dirty_pages || capsnap->writing) 1550 break; 1551 1552 /* should be removed by ceph_try_drop_cap_snap() */ 1553 BUG_ON(!capsnap->need_flush); 1554 1555 /* only flush each capsnap once */ 1556 if (capsnap->cap_flush.tid > 0) { 1557 dout(" already flushed %p, skipping\n", capsnap); 1558 continue; 1559 } 1560 1561 spin_lock(&mdsc->cap_dirty_lock); 1562 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid; 1563 list_add_tail(&capsnap->cap_flush.g_list, 1564 &mdsc->cap_flush_list); 1565 if (oldest_flush_tid == 0) 1566 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 1567 if (list_empty(&ci->i_flushing_item)) { 1568 list_add_tail(&ci->i_flushing_item, 1569 &session->s_cap_flushing); 1570 } 1571 spin_unlock(&mdsc->cap_dirty_lock); 1572 1573 list_add_tail(&capsnap->cap_flush.i_list, 1574 &ci->i_cap_flush_list); 1575 1576 if (first_tid == 1) 1577 first_tid = capsnap->cap_flush.tid; 1578 last_tid = capsnap->cap_flush.tid; 1579 } 1580 1581 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS; 1582 1583 while (first_tid <= last_tid) { 1584 struct ceph_cap *cap = ci->i_auth_cap; 1585 struct ceph_cap_flush *cf; 1586 int ret; 1587 1588 if (!(cap && cap->session == session)) { 1589 dout("__flush_snaps %p auth cap %p not mds%d, " 1590 "stop\n", inode, cap, session->s_mds); 1591 break; 1592 } 1593 1594 ret = -ENOENT; 1595 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) { 1596 if (cf->tid >= first_tid) { 1597 ret = 0; 1598 break; 1599 } 1600 } 1601 if (ret < 0) 1602 break; 1603 1604 first_tid = cf->tid + 1; 1605 1606 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush); 1607 refcount_inc(&capsnap->nref); 1608 spin_unlock(&ci->i_ceph_lock); 1609 1610 dout("__flush_snaps %p capsnap %p tid %llu %s\n", 1611 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty)); 1612 1613 ret = __send_flush_snap(inode, session, capsnap, cap->mseq, 1614 oldest_flush_tid); 1615 if (ret < 0) { 1616 pr_err("__flush_snaps: error sending cap flushsnap, " 1617 "ino (%llx.%llx) tid %llu follows %llu\n", 1618 ceph_vinop(inode), cf->tid, capsnap->follows); 1619 } 1620 1621 ceph_put_cap_snap(capsnap); 1622 spin_lock(&ci->i_ceph_lock); 1623 } 1624 } 1625 1626 void ceph_flush_snaps(struct ceph_inode_info *ci, 1627 struct ceph_mds_session **psession) 1628 { 1629 struct inode *inode = &ci->vfs_inode; 1630 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 1631 struct ceph_mds_session *session = NULL; 1632 int mds; 1633 1634 dout("ceph_flush_snaps %p\n", inode); 1635 if (psession) 1636 session = *psession; 1637 retry: 1638 spin_lock(&ci->i_ceph_lock); 1639 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) { 1640 dout(" no capsnap needs flush, doing nothing\n"); 1641 goto out; 1642 } 1643 if (!ci->i_auth_cap) { 1644 dout(" no auth cap (migrating?), doing nothing\n"); 1645 goto out; 1646 } 1647 1648 mds = ci->i_auth_cap->session->s_mds; 1649 if (session && session->s_mds != mds) { 1650 dout(" oops, wrong session %p mutex\n", session); 1651 mutex_unlock(&session->s_mutex); 1652 ceph_put_mds_session(session); 1653 session = NULL; 1654 } 1655 if (!session) { 1656 spin_unlock(&ci->i_ceph_lock); 1657 mutex_lock(&mdsc->mutex); 1658 session = __ceph_lookup_mds_session(mdsc, mds); 1659 mutex_unlock(&mdsc->mutex); 1660 if (session) { 1661 dout(" inverting session/ino locks on %p\n", session); 1662 mutex_lock(&session->s_mutex); 1663 } 1664 goto retry; 1665 } 1666 1667 // make sure flushsnap messages are sent in proper order. 1668 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) 1669 __kick_flushing_caps(mdsc, session, ci, 0); 1670 1671 __ceph_flush_snaps(ci, session); 1672 out: 1673 spin_unlock(&ci->i_ceph_lock); 1674 1675 if (psession) { 1676 *psession = session; 1677 } else if (session) { 1678 mutex_unlock(&session->s_mutex); 1679 ceph_put_mds_session(session); 1680 } 1681 /* we flushed them all; remove this inode from the queue */ 1682 spin_lock(&mdsc->snap_flush_lock); 1683 list_del_init(&ci->i_snap_flush_item); 1684 spin_unlock(&mdsc->snap_flush_lock); 1685 } 1686 1687 /* 1688 * Mark caps dirty. If inode is newly dirty, return the dirty flags. 1689 * Caller is then responsible for calling __mark_inode_dirty with the 1690 * returned flags value. 1691 */ 1692 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask, 1693 struct ceph_cap_flush **pcf) 1694 { 1695 struct ceph_mds_client *mdsc = 1696 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc; 1697 struct inode *inode = &ci->vfs_inode; 1698 int was = ci->i_dirty_caps; 1699 int dirty = 0; 1700 1701 lockdep_assert_held(&ci->i_ceph_lock); 1702 1703 if (!ci->i_auth_cap) { 1704 pr_warn("__mark_dirty_caps %p %llx mask %s, " 1705 "but no auth cap (session was closed?)\n", 1706 inode, ceph_ino(inode), ceph_cap_string(mask)); 1707 return 0; 1708 } 1709 1710 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode, 1711 ceph_cap_string(mask), ceph_cap_string(was), 1712 ceph_cap_string(was | mask)); 1713 ci->i_dirty_caps |= mask; 1714 if (was == 0) { 1715 struct ceph_mds_session *session = ci->i_auth_cap->session; 1716 1717 WARN_ON_ONCE(ci->i_prealloc_cap_flush); 1718 swap(ci->i_prealloc_cap_flush, *pcf); 1719 1720 if (!ci->i_head_snapc) { 1721 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem)); 1722 ci->i_head_snapc = ceph_get_snap_context( 1723 ci->i_snap_realm->cached_context); 1724 } 1725 dout(" inode %p now dirty snapc %p auth cap %p\n", 1726 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap); 1727 BUG_ON(!list_empty(&ci->i_dirty_item)); 1728 spin_lock(&mdsc->cap_dirty_lock); 1729 list_add(&ci->i_dirty_item, &session->s_cap_dirty); 1730 spin_unlock(&mdsc->cap_dirty_lock); 1731 if (ci->i_flushing_caps == 0) { 1732 ihold(inode); 1733 dirty |= I_DIRTY_SYNC; 1734 } 1735 } else { 1736 WARN_ON_ONCE(!ci->i_prealloc_cap_flush); 1737 } 1738 BUG_ON(list_empty(&ci->i_dirty_item)); 1739 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) && 1740 (mask & CEPH_CAP_FILE_BUFFER)) 1741 dirty |= I_DIRTY_DATASYNC; 1742 __cap_delay_requeue(mdsc, ci); 1743 return dirty; 1744 } 1745 1746 struct ceph_cap_flush *ceph_alloc_cap_flush(void) 1747 { 1748 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL); 1749 } 1750 1751 void ceph_free_cap_flush(struct ceph_cap_flush *cf) 1752 { 1753 if (cf) 1754 kmem_cache_free(ceph_cap_flush_cachep, cf); 1755 } 1756 1757 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc) 1758 { 1759 if (!list_empty(&mdsc->cap_flush_list)) { 1760 struct ceph_cap_flush *cf = 1761 list_first_entry(&mdsc->cap_flush_list, 1762 struct ceph_cap_flush, g_list); 1763 return cf->tid; 1764 } 1765 return 0; 1766 } 1767 1768 /* 1769 * Remove cap_flush from the mdsc's or inode's flushing cap list. 1770 * Return true if caller needs to wake up flush waiters. 1771 */ 1772 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc, 1773 struct ceph_cap_flush *cf) 1774 { 1775 struct ceph_cap_flush *prev; 1776 bool wake = cf->wake; 1777 1778 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) { 1779 prev = list_prev_entry(cf, g_list); 1780 prev->wake = true; 1781 wake = false; 1782 } 1783 list_del(&cf->g_list); 1784 return wake; 1785 } 1786 1787 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci, 1788 struct ceph_cap_flush *cf) 1789 { 1790 struct ceph_cap_flush *prev; 1791 bool wake = cf->wake; 1792 1793 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) { 1794 prev = list_prev_entry(cf, i_list); 1795 prev->wake = true; 1796 wake = false; 1797 } 1798 list_del(&cf->i_list); 1799 return wake; 1800 } 1801 1802 /* 1803 * Add dirty inode to the flushing list. Assigned a seq number so we 1804 * can wait for caps to flush without starving. 1805 * 1806 * Called under i_ceph_lock. Returns the flush tid. 1807 */ 1808 static u64 __mark_caps_flushing(struct inode *inode, 1809 struct ceph_mds_session *session, bool wake, 1810 u64 *oldest_flush_tid) 1811 { 1812 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 1813 struct ceph_inode_info *ci = ceph_inode(inode); 1814 struct ceph_cap_flush *cf = NULL; 1815 int flushing; 1816 1817 lockdep_assert_held(&ci->i_ceph_lock); 1818 BUG_ON(ci->i_dirty_caps == 0); 1819 BUG_ON(list_empty(&ci->i_dirty_item)); 1820 BUG_ON(!ci->i_prealloc_cap_flush); 1821 1822 flushing = ci->i_dirty_caps; 1823 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n", 1824 ceph_cap_string(flushing), 1825 ceph_cap_string(ci->i_flushing_caps), 1826 ceph_cap_string(ci->i_flushing_caps | flushing)); 1827 ci->i_flushing_caps |= flushing; 1828 ci->i_dirty_caps = 0; 1829 dout(" inode %p now !dirty\n", inode); 1830 1831 swap(cf, ci->i_prealloc_cap_flush); 1832 cf->caps = flushing; 1833 cf->wake = wake; 1834 1835 spin_lock(&mdsc->cap_dirty_lock); 1836 list_del_init(&ci->i_dirty_item); 1837 1838 cf->tid = ++mdsc->last_cap_flush_tid; 1839 list_add_tail(&cf->g_list, &mdsc->cap_flush_list); 1840 *oldest_flush_tid = __get_oldest_flush_tid(mdsc); 1841 1842 if (list_empty(&ci->i_flushing_item)) { 1843 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing); 1844 mdsc->num_cap_flushing++; 1845 } 1846 spin_unlock(&mdsc->cap_dirty_lock); 1847 1848 list_add_tail(&cf->i_list, &ci->i_cap_flush_list); 1849 1850 return cf->tid; 1851 } 1852 1853 /* 1854 * try to invalidate mapping pages without blocking. 1855 */ 1856 static int try_nonblocking_invalidate(struct inode *inode) 1857 { 1858 struct ceph_inode_info *ci = ceph_inode(inode); 1859 u32 invalidating_gen = ci->i_rdcache_gen; 1860 1861 spin_unlock(&ci->i_ceph_lock); 1862 invalidate_mapping_pages(&inode->i_data, 0, -1); 1863 spin_lock(&ci->i_ceph_lock); 1864 1865 if (inode->i_data.nrpages == 0 && 1866 invalidating_gen == ci->i_rdcache_gen) { 1867 /* success. */ 1868 dout("try_nonblocking_invalidate %p success\n", inode); 1869 /* save any racing async invalidate some trouble */ 1870 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1; 1871 return 0; 1872 } 1873 dout("try_nonblocking_invalidate %p failed\n", inode); 1874 return -1; 1875 } 1876 1877 bool __ceph_should_report_size(struct ceph_inode_info *ci) 1878 { 1879 loff_t size = ci->vfs_inode.i_size; 1880 /* mds will adjust max size according to the reported size */ 1881 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR) 1882 return false; 1883 if (size >= ci->i_max_size) 1884 return true; 1885 /* half of previous max_size increment has been used */ 1886 if (ci->i_max_size > ci->i_reported_size && 1887 (size << 1) >= ci->i_max_size + ci->i_reported_size) 1888 return true; 1889 return false; 1890 } 1891 1892 /* 1893 * Swiss army knife function to examine currently used and wanted 1894 * versus held caps. Release, flush, ack revoked caps to mds as 1895 * appropriate. 1896 * 1897 * CHECK_CAPS_AUTHONLY - we should only check the auth cap 1898 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without 1899 * further delay. 1900 */ 1901 void ceph_check_caps(struct ceph_inode_info *ci, int flags, 1902 struct ceph_mds_session *session) 1903 { 1904 struct inode *inode = &ci->vfs_inode; 1905 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb); 1906 struct ceph_cap *cap; 1907 u64 flush_tid, oldest_flush_tid; 1908 int file_wanted, used, cap_used; 1909 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */ 1910 int issued, implemented, want, retain, revoking, flushing = 0; 1911 int mds = -1; /* keep track of how far we've gone through i_caps list 1912 to avoid an infinite loop on retry */ 1913 struct rb_node *p; 1914 bool queue_invalidate = false; 1915 bool tried_invalidate = false; 1916 1917 spin_lock(&ci->i_ceph_lock); 1918 if (ci->i_ceph_flags & CEPH_I_FLUSH) 1919 flags |= CHECK_CAPS_FLUSH; 1920 1921 goto retry_locked; 1922 retry: 1923 spin_lock(&ci->i_ceph_lock); 1924 retry_locked: 1925 /* Caps wanted by virtue of active open files. */ 1926 file_wanted = __ceph_caps_file_wanted(ci); 1927 1928 /* Caps which have active references against them */ 1929 used = __ceph_caps_used(ci); 1930 1931 /* 1932 * "issued" represents the current caps that the MDS wants us to have. 1933 * "implemented" is the set that we have been granted, and includes the 1934 * ones that have not yet been returned to the MDS (the "revoking" set, 1935 * usually because they have outstanding references). 1936 */ 1937 issued = __ceph_caps_issued(ci, &implemented); 1938 revoking = implemented & ~issued; 1939 1940 want = file_wanted; 1941 1942 /* The ones we currently want to retain (may be adjusted below) */ 1943 retain = file_wanted | used | CEPH_CAP_PIN; 1944 if (!mdsc->stopping && inode->i_nlink > 0) { 1945 if (file_wanted) { 1946 retain |= CEPH_CAP_ANY; /* be greedy */ 1947 } else if (S_ISDIR(inode->i_mode) && 1948 (issued & CEPH_CAP_FILE_SHARED) && 1949 __ceph_dir_is_complete(ci)) { 1950 /* 1951 * If a directory is complete, we want to keep 1952 * the exclusive cap. So that MDS does not end up 1953 * revoking the shared cap on every create/unlink 1954 * operation. 1955 */ 1956 if (IS_RDONLY(inode)) { 1957 want = CEPH_CAP_ANY_SHARED; 1958 } else { 1959 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL; 1960 } 1961 retain |= want; 1962 } else { 1963 1964 retain |= CEPH_CAP_ANY_SHARED; 1965 /* 1966 * keep RD only if we didn't have the file open RW, 1967 * because then the mds would revoke it anyway to 1968 * journal max_size=0. 1969 */ 1970 if (ci->i_max_size == 0) 1971 retain |= CEPH_CAP_ANY_RD; 1972 } 1973 } 1974 1975 dout("check_caps %p file_want %s used %s dirty %s flushing %s" 1976 " issued %s revoking %s retain %s %s%s\n", inode, 1977 ceph_cap_string(file_wanted), 1978 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps), 1979 ceph_cap_string(ci->i_flushing_caps), 1980 ceph_cap_string(issued), ceph_cap_string(revoking), 1981 ceph_cap_string(retain), 1982 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "", 1983 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : ""); 1984 1985 /* 1986 * If we no longer need to hold onto old our caps, and we may 1987 * have cached pages, but don't want them, then try to invalidate. 1988 * If we fail, it's because pages are locked.... try again later. 1989 */ 1990 if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) && 1991 S_ISREG(inode->i_mode) && 1992 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */ 1993 inode->i_data.nrpages && /* have cached pages */ 1994 (revoking & (CEPH_CAP_FILE_CACHE| 1995 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */ 1996 !tried_invalidate) { 1997 dout("check_caps trying to invalidate on %p\n", inode); 1998 if (try_nonblocking_invalidate(inode) < 0) { 1999 dout("check_caps queuing invalidate\n"); 2000 queue_invalidate = true; 2001 ci->i_rdcache_revoking = ci->i_rdcache_gen; 2002 } 2003 tried_invalidate = true; 2004 goto retry_locked; 2005 } 2006 2007 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 2008 int mflags = 0; 2009 struct cap_msg_args arg; 2010 2011 cap = rb_entry(p, struct ceph_cap, ci_node); 2012 2013 /* avoid looping forever */ 2014 if (mds >= cap->mds || 2015 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap)) 2016 continue; 2017 2018 /* NOTE: no side-effects allowed, until we take s_mutex */ 2019 2020 /* 2021 * If we have an auth cap, we don't need to consider any 2022 * overlapping caps as used. 2023 */ 2024 cap_used = used; 2025 if (ci->i_auth_cap && cap != ci->i_auth_cap) 2026 cap_used &= ~ci->i_auth_cap->issued; 2027 2028 revoking = cap->implemented & ~cap->issued; 2029 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n", 2030 cap->mds, cap, ceph_cap_string(cap_used), 2031 ceph_cap_string(cap->issued), 2032 ceph_cap_string(cap->implemented), 2033 ceph_cap_string(revoking)); 2034 2035 if (cap == ci->i_auth_cap && 2036 (cap->issued & CEPH_CAP_FILE_WR)) { 2037 /* request larger max_size from MDS? */ 2038 if (ci->i_wanted_max_size > ci->i_max_size && 2039 ci->i_wanted_max_size > ci->i_requested_max_size) { 2040 dout("requesting new max_size\n"); 2041 goto ack; 2042 } 2043 2044 /* approaching file_max? */ 2045 if (__ceph_should_report_size(ci)) { 2046 dout("i_size approaching max_size\n"); 2047 goto ack; 2048 } 2049 } 2050 /* flush anything dirty? */ 2051 if (cap == ci->i_auth_cap) { 2052 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) { 2053 dout("flushing dirty caps\n"); 2054 goto ack; 2055 } 2056 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) { 2057 dout("flushing snap caps\n"); 2058 goto ack; 2059 } 2060 } 2061 2062 /* completed revocation? going down and there are no caps? */ 2063 if (revoking && (revoking & cap_used) == 0) { 2064 dout("completed revocation of %s\n", 2065 ceph_cap_string(cap->implemented & ~cap->issued)); 2066 goto ack; 2067 } 2068 2069 /* want more caps from mds? */ 2070 if (want & ~cap->mds_wanted) { 2071 if (want & ~(cap->mds_wanted | cap->issued)) 2072 goto ack; 2073 if (!__cap_is_valid(cap)) 2074 goto ack; 2075 } 2076 2077 /* things we might delay */ 2078 if ((cap->issued & ~retain) == 0) 2079 continue; /* nope, all good */ 2080 2081 ack: 2082 if (session && session != cap->session) { 2083 dout("oops, wrong session %p mutex\n", session); 2084 mutex_unlock(&session->s_mutex); 2085 session = NULL; 2086 } 2087 if (!session) { 2088 session = cap->session; 2089 if (mutex_trylock(&session->s_mutex) == 0) { 2090 dout("inverting session/ino locks on %p\n", 2091 session); 2092 session = ceph_get_mds_session(session); 2093 spin_unlock(&ci->i_ceph_lock); 2094 if (took_snap_rwsem) { 2095 up_read(&mdsc->snap_rwsem); 2096 took_snap_rwsem = 0; 2097 } 2098 if (session) { 2099 mutex_lock(&session->s_mutex); 2100 ceph_put_mds_session(session); 2101 } else { 2102 /* 2103 * Because we take the reference while 2104 * holding the i_ceph_lock, it should 2105 * never be NULL. Throw a warning if it 2106 * ever is. 2107 */ 2108 WARN_ON_ONCE(true); 2109 } 2110 goto retry; 2111 } 2112 } 2113 2114 /* kick flushing and flush snaps before sending normal 2115 * cap message */ 2116 if (cap == ci->i_auth_cap && 2117 (ci->i_ceph_flags & 2118 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) { 2119 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) 2120 __kick_flushing_caps(mdsc, session, ci, 0); 2121 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) 2122 __ceph_flush_snaps(ci, session); 2123 2124 goto retry_locked; 2125 } 2126 2127 /* take snap_rwsem after session mutex */ 2128 if (!took_snap_rwsem) { 2129 if (down_read_trylock(&mdsc->snap_rwsem) == 0) { 2130 dout("inverting snap/in locks on %p\n", 2131 inode); 2132 spin_unlock(&ci->i_ceph_lock); 2133 down_read(&mdsc->snap_rwsem); 2134 took_snap_rwsem = 1; 2135 goto retry; 2136 } 2137 took_snap_rwsem = 1; 2138 } 2139 2140 if (cap == ci->i_auth_cap && ci->i_dirty_caps) { 2141 flushing = ci->i_dirty_caps; 2142 flush_tid = __mark_caps_flushing(inode, session, false, 2143 &oldest_flush_tid); 2144 if (flags & CHECK_CAPS_FLUSH && 2145 list_empty(&session->s_cap_dirty)) 2146 mflags |= CEPH_CLIENT_CAPS_SYNC; 2147 } else { 2148 flushing = 0; 2149 flush_tid = 0; 2150 spin_lock(&mdsc->cap_dirty_lock); 2151 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 2152 spin_unlock(&mdsc->cap_dirty_lock); 2153 } 2154 2155 mds = cap->mds; /* remember mds, so we don't repeat */ 2156 2157 __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used, 2158 want, retain, flushing, flush_tid, oldest_flush_tid); 2159 spin_unlock(&ci->i_ceph_lock); 2160 2161 __send_cap(&arg, ci); 2162 2163 goto retry; /* retake i_ceph_lock and restart our cap scan. */ 2164 } 2165 2166 /* periodically re-calculate caps wanted by open files */ 2167 if (__ceph_is_any_real_caps(ci) && 2168 list_empty(&ci->i_cap_delay_list) && 2169 (file_wanted & ~CEPH_CAP_PIN) && 2170 !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) { 2171 __cap_delay_requeue(mdsc, ci); 2172 } 2173 2174 spin_unlock(&ci->i_ceph_lock); 2175 2176 if (queue_invalidate) 2177 ceph_queue_invalidate(inode); 2178 2179 if (session) 2180 mutex_unlock(&session->s_mutex); 2181 if (took_snap_rwsem) 2182 up_read(&mdsc->snap_rwsem); 2183 } 2184 2185 /* 2186 * Try to flush dirty caps back to the auth mds. 2187 */ 2188 static int try_flush_caps(struct inode *inode, u64 *ptid) 2189 { 2190 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 2191 struct ceph_inode_info *ci = ceph_inode(inode); 2192 struct ceph_mds_session *session = NULL; 2193 int flushing = 0; 2194 u64 flush_tid = 0, oldest_flush_tid = 0; 2195 2196 retry: 2197 spin_lock(&ci->i_ceph_lock); 2198 retry_locked: 2199 if (ci->i_dirty_caps && ci->i_auth_cap) { 2200 struct ceph_cap *cap = ci->i_auth_cap; 2201 struct cap_msg_args arg; 2202 2203 if (session != cap->session) { 2204 spin_unlock(&ci->i_ceph_lock); 2205 if (session) 2206 mutex_unlock(&session->s_mutex); 2207 session = cap->session; 2208 mutex_lock(&session->s_mutex); 2209 goto retry; 2210 } 2211 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) { 2212 spin_unlock(&ci->i_ceph_lock); 2213 goto out; 2214 } 2215 2216 if (ci->i_ceph_flags & 2217 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) { 2218 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) 2219 __kick_flushing_caps(mdsc, session, ci, 0); 2220 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) 2221 __ceph_flush_snaps(ci, session); 2222 goto retry_locked; 2223 } 2224 2225 flushing = ci->i_dirty_caps; 2226 flush_tid = __mark_caps_flushing(inode, session, true, 2227 &oldest_flush_tid); 2228 2229 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC, 2230 __ceph_caps_used(ci), __ceph_caps_wanted(ci), 2231 (cap->issued | cap->implemented), 2232 flushing, flush_tid, oldest_flush_tid); 2233 spin_unlock(&ci->i_ceph_lock); 2234 2235 __send_cap(&arg, ci); 2236 } else { 2237 if (!list_empty(&ci->i_cap_flush_list)) { 2238 struct ceph_cap_flush *cf = 2239 list_last_entry(&ci->i_cap_flush_list, 2240 struct ceph_cap_flush, i_list); 2241 cf->wake = true; 2242 flush_tid = cf->tid; 2243 } 2244 flushing = ci->i_flushing_caps; 2245 spin_unlock(&ci->i_ceph_lock); 2246 } 2247 out: 2248 if (session) 2249 mutex_unlock(&session->s_mutex); 2250 2251 *ptid = flush_tid; 2252 return flushing; 2253 } 2254 2255 /* 2256 * Return true if we've flushed caps through the given flush_tid. 2257 */ 2258 static int caps_are_flushed(struct inode *inode, u64 flush_tid) 2259 { 2260 struct ceph_inode_info *ci = ceph_inode(inode); 2261 int ret = 1; 2262 2263 spin_lock(&ci->i_ceph_lock); 2264 if (!list_empty(&ci->i_cap_flush_list)) { 2265 struct ceph_cap_flush * cf = 2266 list_first_entry(&ci->i_cap_flush_list, 2267 struct ceph_cap_flush, i_list); 2268 if (cf->tid <= flush_tid) 2269 ret = 0; 2270 } 2271 spin_unlock(&ci->i_ceph_lock); 2272 return ret; 2273 } 2274 2275 /* 2276 * wait for any unsafe requests to complete. 2277 */ 2278 static int unsafe_request_wait(struct inode *inode) 2279 { 2280 struct ceph_inode_info *ci = ceph_inode(inode); 2281 struct ceph_mds_request *req1 = NULL, *req2 = NULL; 2282 int ret, err = 0; 2283 2284 spin_lock(&ci->i_unsafe_lock); 2285 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) { 2286 req1 = list_last_entry(&ci->i_unsafe_dirops, 2287 struct ceph_mds_request, 2288 r_unsafe_dir_item); 2289 ceph_mdsc_get_request(req1); 2290 } 2291 if (!list_empty(&ci->i_unsafe_iops)) { 2292 req2 = list_last_entry(&ci->i_unsafe_iops, 2293 struct ceph_mds_request, 2294 r_unsafe_target_item); 2295 ceph_mdsc_get_request(req2); 2296 } 2297 spin_unlock(&ci->i_unsafe_lock); 2298 2299 dout("unsafe_request_wait %p wait on tid %llu %llu\n", 2300 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL); 2301 if (req1) { 2302 ret = !wait_for_completion_timeout(&req1->r_safe_completion, 2303 ceph_timeout_jiffies(req1->r_timeout)); 2304 if (ret) 2305 err = -EIO; 2306 ceph_mdsc_put_request(req1); 2307 } 2308 if (req2) { 2309 ret = !wait_for_completion_timeout(&req2->r_safe_completion, 2310 ceph_timeout_jiffies(req2->r_timeout)); 2311 if (ret) 2312 err = -EIO; 2313 ceph_mdsc_put_request(req2); 2314 } 2315 return err; 2316 } 2317 2318 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync) 2319 { 2320 struct ceph_file_info *fi = file->private_data; 2321 struct inode *inode = file->f_mapping->host; 2322 struct ceph_inode_info *ci = ceph_inode(inode); 2323 u64 flush_tid; 2324 int ret, err; 2325 int dirty; 2326 2327 dout("fsync %p%s\n", inode, datasync ? " datasync" : ""); 2328 2329 ret = file_write_and_wait_range(file, start, end); 2330 if (datasync) 2331 goto out; 2332 2333 ret = ceph_wait_on_async_create(inode); 2334 if (ret) 2335 goto out; 2336 2337 dirty = try_flush_caps(inode, &flush_tid); 2338 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty)); 2339 2340 err = unsafe_request_wait(inode); 2341 2342 /* 2343 * only wait on non-file metadata writeback (the mds 2344 * can recover size and mtime, so we don't need to 2345 * wait for that) 2346 */ 2347 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) { 2348 err = wait_event_interruptible(ci->i_cap_wq, 2349 caps_are_flushed(inode, flush_tid)); 2350 } 2351 2352 if (err < 0) 2353 ret = err; 2354 2355 if (errseq_check(&ci->i_meta_err, READ_ONCE(fi->meta_err))) { 2356 spin_lock(&file->f_lock); 2357 err = errseq_check_and_advance(&ci->i_meta_err, 2358 &fi->meta_err); 2359 spin_unlock(&file->f_lock); 2360 if (err < 0) 2361 ret = err; 2362 } 2363 out: 2364 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret); 2365 return ret; 2366 } 2367 2368 /* 2369 * Flush any dirty caps back to the mds. If we aren't asked to wait, 2370 * queue inode for flush but don't do so immediately, because we can 2371 * get by with fewer MDS messages if we wait for data writeback to 2372 * complete first. 2373 */ 2374 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc) 2375 { 2376 struct ceph_inode_info *ci = ceph_inode(inode); 2377 u64 flush_tid; 2378 int err = 0; 2379 int dirty; 2380 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync); 2381 2382 dout("write_inode %p wait=%d\n", inode, wait); 2383 if (wait) { 2384 dirty = try_flush_caps(inode, &flush_tid); 2385 if (dirty) 2386 err = wait_event_interruptible(ci->i_cap_wq, 2387 caps_are_flushed(inode, flush_tid)); 2388 } else { 2389 struct ceph_mds_client *mdsc = 2390 ceph_sb_to_client(inode->i_sb)->mdsc; 2391 2392 spin_lock(&ci->i_ceph_lock); 2393 if (__ceph_caps_dirty(ci)) 2394 __cap_delay_requeue_front(mdsc, ci); 2395 spin_unlock(&ci->i_ceph_lock); 2396 } 2397 return err; 2398 } 2399 2400 static void __kick_flushing_caps(struct ceph_mds_client *mdsc, 2401 struct ceph_mds_session *session, 2402 struct ceph_inode_info *ci, 2403 u64 oldest_flush_tid) 2404 __releases(ci->i_ceph_lock) 2405 __acquires(ci->i_ceph_lock) 2406 { 2407 struct inode *inode = &ci->vfs_inode; 2408 struct ceph_cap *cap; 2409 struct ceph_cap_flush *cf; 2410 int ret; 2411 u64 first_tid = 0; 2412 u64 last_snap_flush = 0; 2413 2414 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH; 2415 2416 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) { 2417 if (!cf->caps) { 2418 last_snap_flush = cf->tid; 2419 break; 2420 } 2421 } 2422 2423 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) { 2424 if (cf->tid < first_tid) 2425 continue; 2426 2427 cap = ci->i_auth_cap; 2428 if (!(cap && cap->session == session)) { 2429 pr_err("%p auth cap %p not mds%d ???\n", 2430 inode, cap, session->s_mds); 2431 break; 2432 } 2433 2434 first_tid = cf->tid + 1; 2435 2436 if (cf->caps) { 2437 struct cap_msg_args arg; 2438 2439 dout("kick_flushing_caps %p cap %p tid %llu %s\n", 2440 inode, cap, cf->tid, ceph_cap_string(cf->caps)); 2441 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, 2442 (cf->tid < last_snap_flush ? 2443 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0), 2444 __ceph_caps_used(ci), 2445 __ceph_caps_wanted(ci), 2446 (cap->issued | cap->implemented), 2447 cf->caps, cf->tid, oldest_flush_tid); 2448 spin_unlock(&ci->i_ceph_lock); 2449 __send_cap(&arg, ci); 2450 } else { 2451 struct ceph_cap_snap *capsnap = 2452 container_of(cf, struct ceph_cap_snap, 2453 cap_flush); 2454 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n", 2455 inode, capsnap, cf->tid, 2456 ceph_cap_string(capsnap->dirty)); 2457 2458 refcount_inc(&capsnap->nref); 2459 spin_unlock(&ci->i_ceph_lock); 2460 2461 ret = __send_flush_snap(inode, session, capsnap, cap->mseq, 2462 oldest_flush_tid); 2463 if (ret < 0) { 2464 pr_err("kick_flushing_caps: error sending " 2465 "cap flushsnap, ino (%llx.%llx) " 2466 "tid %llu follows %llu\n", 2467 ceph_vinop(inode), cf->tid, 2468 capsnap->follows); 2469 } 2470 2471 ceph_put_cap_snap(capsnap); 2472 } 2473 2474 spin_lock(&ci->i_ceph_lock); 2475 } 2476 } 2477 2478 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc, 2479 struct ceph_mds_session *session) 2480 { 2481 struct ceph_inode_info *ci; 2482 struct ceph_cap *cap; 2483 u64 oldest_flush_tid; 2484 2485 dout("early_kick_flushing_caps mds%d\n", session->s_mds); 2486 2487 spin_lock(&mdsc->cap_dirty_lock); 2488 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 2489 spin_unlock(&mdsc->cap_dirty_lock); 2490 2491 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) { 2492 spin_lock(&ci->i_ceph_lock); 2493 cap = ci->i_auth_cap; 2494 if (!(cap && cap->session == session)) { 2495 pr_err("%p auth cap %p not mds%d ???\n", 2496 &ci->vfs_inode, cap, session->s_mds); 2497 spin_unlock(&ci->i_ceph_lock); 2498 continue; 2499 } 2500 2501 2502 /* 2503 * if flushing caps were revoked, we re-send the cap flush 2504 * in client reconnect stage. This guarantees MDS * processes 2505 * the cap flush message before issuing the flushing caps to 2506 * other client. 2507 */ 2508 if ((cap->issued & ci->i_flushing_caps) != 2509 ci->i_flushing_caps) { 2510 /* encode_caps_cb() also will reset these sequence 2511 * numbers. make sure sequence numbers in cap flush 2512 * message match later reconnect message */ 2513 cap->seq = 0; 2514 cap->issue_seq = 0; 2515 cap->mseq = 0; 2516 __kick_flushing_caps(mdsc, session, ci, 2517 oldest_flush_tid); 2518 } else { 2519 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH; 2520 } 2521 2522 spin_unlock(&ci->i_ceph_lock); 2523 } 2524 } 2525 2526 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc, 2527 struct ceph_mds_session *session) 2528 { 2529 struct ceph_inode_info *ci; 2530 struct ceph_cap *cap; 2531 u64 oldest_flush_tid; 2532 2533 lockdep_assert_held(&session->s_mutex); 2534 2535 dout("kick_flushing_caps mds%d\n", session->s_mds); 2536 2537 spin_lock(&mdsc->cap_dirty_lock); 2538 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 2539 spin_unlock(&mdsc->cap_dirty_lock); 2540 2541 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) { 2542 spin_lock(&ci->i_ceph_lock); 2543 cap = ci->i_auth_cap; 2544 if (!(cap && cap->session == session)) { 2545 pr_err("%p auth cap %p not mds%d ???\n", 2546 &ci->vfs_inode, cap, session->s_mds); 2547 spin_unlock(&ci->i_ceph_lock); 2548 continue; 2549 } 2550 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) { 2551 __kick_flushing_caps(mdsc, session, ci, 2552 oldest_flush_tid); 2553 } 2554 spin_unlock(&ci->i_ceph_lock); 2555 } 2556 } 2557 2558 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session, 2559 struct ceph_inode_info *ci) 2560 { 2561 struct ceph_mds_client *mdsc = session->s_mdsc; 2562 struct ceph_cap *cap = ci->i_auth_cap; 2563 2564 lockdep_assert_held(&ci->i_ceph_lock); 2565 2566 dout("%s %p flushing %s\n", __func__, &ci->vfs_inode, 2567 ceph_cap_string(ci->i_flushing_caps)); 2568 2569 if (!list_empty(&ci->i_cap_flush_list)) { 2570 u64 oldest_flush_tid; 2571 spin_lock(&mdsc->cap_dirty_lock); 2572 list_move_tail(&ci->i_flushing_item, 2573 &cap->session->s_cap_flushing); 2574 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 2575 spin_unlock(&mdsc->cap_dirty_lock); 2576 2577 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid); 2578 } 2579 } 2580 2581 2582 /* 2583 * Take references to capabilities we hold, so that we don't release 2584 * them to the MDS prematurely. 2585 */ 2586 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got, 2587 bool snap_rwsem_locked) 2588 { 2589 lockdep_assert_held(&ci->i_ceph_lock); 2590 2591 if (got & CEPH_CAP_PIN) 2592 ci->i_pin_ref++; 2593 if (got & CEPH_CAP_FILE_RD) 2594 ci->i_rd_ref++; 2595 if (got & CEPH_CAP_FILE_CACHE) 2596 ci->i_rdcache_ref++; 2597 if (got & CEPH_CAP_FILE_EXCL) 2598 ci->i_fx_ref++; 2599 if (got & CEPH_CAP_FILE_WR) { 2600 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) { 2601 BUG_ON(!snap_rwsem_locked); 2602 ci->i_head_snapc = ceph_get_snap_context( 2603 ci->i_snap_realm->cached_context); 2604 } 2605 ci->i_wr_ref++; 2606 } 2607 if (got & CEPH_CAP_FILE_BUFFER) { 2608 if (ci->i_wb_ref == 0) 2609 ihold(&ci->vfs_inode); 2610 ci->i_wb_ref++; 2611 dout("%s %p wb %d -> %d (?)\n", __func__, 2612 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref); 2613 } 2614 } 2615 2616 /* 2617 * Try to grab cap references. Specify those refs we @want, and the 2618 * minimal set we @need. Also include the larger offset we are writing 2619 * to (when applicable), and check against max_size here as well. 2620 * Note that caller is responsible for ensuring max_size increases are 2621 * requested from the MDS. 2622 * 2623 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed, 2624 * or a negative error code. There are 3 speical error codes: 2625 * -EAGAIN: need to sleep but non-blocking is specified 2626 * -EFBIG: ask caller to call check_max_size() and try again. 2627 * -ESTALE: ask caller to call ceph_renew_caps() and try again. 2628 */ 2629 enum { 2630 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */ 2631 NON_BLOCKING = (1 << 8), 2632 CHECK_FILELOCK = (1 << 9), 2633 }; 2634 2635 static int try_get_cap_refs(struct inode *inode, int need, int want, 2636 loff_t endoff, int flags, int *got) 2637 { 2638 struct ceph_inode_info *ci = ceph_inode(inode); 2639 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 2640 int ret = 0; 2641 int have, implemented; 2642 bool snap_rwsem_locked = false; 2643 2644 dout("get_cap_refs %p need %s want %s\n", inode, 2645 ceph_cap_string(need), ceph_cap_string(want)); 2646 2647 again: 2648 spin_lock(&ci->i_ceph_lock); 2649 2650 if ((flags & CHECK_FILELOCK) && 2651 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) { 2652 dout("try_get_cap_refs %p error filelock\n", inode); 2653 ret = -EIO; 2654 goto out_unlock; 2655 } 2656 2657 /* finish pending truncate */ 2658 while (ci->i_truncate_pending) { 2659 spin_unlock(&ci->i_ceph_lock); 2660 if (snap_rwsem_locked) { 2661 up_read(&mdsc->snap_rwsem); 2662 snap_rwsem_locked = false; 2663 } 2664 __ceph_do_pending_vmtruncate(inode); 2665 spin_lock(&ci->i_ceph_lock); 2666 } 2667 2668 have = __ceph_caps_issued(ci, &implemented); 2669 2670 if (have & need & CEPH_CAP_FILE_WR) { 2671 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) { 2672 dout("get_cap_refs %p endoff %llu > maxsize %llu\n", 2673 inode, endoff, ci->i_max_size); 2674 if (endoff > ci->i_requested_max_size) 2675 ret = ci->i_auth_cap ? -EFBIG : -ESTALE; 2676 goto out_unlock; 2677 } 2678 /* 2679 * If a sync write is in progress, we must wait, so that we 2680 * can get a final snapshot value for size+mtime. 2681 */ 2682 if (__ceph_have_pending_cap_snap(ci)) { 2683 dout("get_cap_refs %p cap_snap_pending\n", inode); 2684 goto out_unlock; 2685 } 2686 } 2687 2688 if ((have & need) == need) { 2689 /* 2690 * Look at (implemented & ~have & not) so that we keep waiting 2691 * on transition from wanted -> needed caps. This is needed 2692 * for WRBUFFER|WR -> WR to avoid a new WR sync write from 2693 * going before a prior buffered writeback happens. 2694 */ 2695 int not = want & ~(have & need); 2696 int revoking = implemented & ~have; 2697 dout("get_cap_refs %p have %s but not %s (revoking %s)\n", 2698 inode, ceph_cap_string(have), ceph_cap_string(not), 2699 ceph_cap_string(revoking)); 2700 if ((revoking & not) == 0) { 2701 if (!snap_rwsem_locked && 2702 !ci->i_head_snapc && 2703 (need & CEPH_CAP_FILE_WR)) { 2704 if (!down_read_trylock(&mdsc->snap_rwsem)) { 2705 /* 2706 * we can not call down_read() when 2707 * task isn't in TASK_RUNNING state 2708 */ 2709 if (flags & NON_BLOCKING) { 2710 ret = -EAGAIN; 2711 goto out_unlock; 2712 } 2713 2714 spin_unlock(&ci->i_ceph_lock); 2715 down_read(&mdsc->snap_rwsem); 2716 snap_rwsem_locked = true; 2717 goto again; 2718 } 2719 snap_rwsem_locked = true; 2720 } 2721 if ((have & want) == want) 2722 *got = need | want; 2723 else 2724 *got = need; 2725 if (S_ISREG(inode->i_mode) && 2726 (need & CEPH_CAP_FILE_RD) && 2727 !(*got & CEPH_CAP_FILE_CACHE)) 2728 ceph_disable_fscache_readpage(ci); 2729 ceph_take_cap_refs(ci, *got, true); 2730 ret = 1; 2731 } 2732 } else { 2733 int session_readonly = false; 2734 int mds_wanted; 2735 if (ci->i_auth_cap && 2736 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) { 2737 struct ceph_mds_session *s = ci->i_auth_cap->session; 2738 spin_lock(&s->s_cap_lock); 2739 session_readonly = s->s_readonly; 2740 spin_unlock(&s->s_cap_lock); 2741 } 2742 if (session_readonly) { 2743 dout("get_cap_refs %p need %s but mds%d readonly\n", 2744 inode, ceph_cap_string(need), ci->i_auth_cap->mds); 2745 ret = -EROFS; 2746 goto out_unlock; 2747 } 2748 2749 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) { 2750 dout("get_cap_refs %p forced umount\n", inode); 2751 ret = -EIO; 2752 goto out_unlock; 2753 } 2754 mds_wanted = __ceph_caps_mds_wanted(ci, false); 2755 if (need & ~mds_wanted) { 2756 dout("get_cap_refs %p need %s > mds_wanted %s\n", 2757 inode, ceph_cap_string(need), 2758 ceph_cap_string(mds_wanted)); 2759 ret = -ESTALE; 2760 goto out_unlock; 2761 } 2762 2763 dout("get_cap_refs %p have %s need %s\n", inode, 2764 ceph_cap_string(have), ceph_cap_string(need)); 2765 } 2766 out_unlock: 2767 2768 __ceph_touch_fmode(ci, mdsc, flags); 2769 2770 spin_unlock(&ci->i_ceph_lock); 2771 if (snap_rwsem_locked) 2772 up_read(&mdsc->snap_rwsem); 2773 2774 if (!ret) 2775 ceph_update_cap_mis(&mdsc->metric); 2776 else if (ret == 1) 2777 ceph_update_cap_hit(&mdsc->metric); 2778 2779 dout("get_cap_refs %p ret %d got %s\n", inode, 2780 ret, ceph_cap_string(*got)); 2781 return ret; 2782 } 2783 2784 /* 2785 * Check the offset we are writing up to against our current 2786 * max_size. If necessary, tell the MDS we want to write to 2787 * a larger offset. 2788 */ 2789 static void check_max_size(struct inode *inode, loff_t endoff) 2790 { 2791 struct ceph_inode_info *ci = ceph_inode(inode); 2792 int check = 0; 2793 2794 /* do we need to explicitly request a larger max_size? */ 2795 spin_lock(&ci->i_ceph_lock); 2796 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) { 2797 dout("write %p at large endoff %llu, req max_size\n", 2798 inode, endoff); 2799 ci->i_wanted_max_size = endoff; 2800 } 2801 /* duplicate ceph_check_caps()'s logic */ 2802 if (ci->i_auth_cap && 2803 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) && 2804 ci->i_wanted_max_size > ci->i_max_size && 2805 ci->i_wanted_max_size > ci->i_requested_max_size) 2806 check = 1; 2807 spin_unlock(&ci->i_ceph_lock); 2808 if (check) 2809 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL); 2810 } 2811 2812 static inline int get_used_fmode(int caps) 2813 { 2814 int fmode = 0; 2815 if (caps & CEPH_CAP_FILE_RD) 2816 fmode |= CEPH_FILE_MODE_RD; 2817 if (caps & CEPH_CAP_FILE_WR) 2818 fmode |= CEPH_FILE_MODE_WR; 2819 return fmode; 2820 } 2821 2822 int ceph_try_get_caps(struct inode *inode, int need, int want, 2823 bool nonblock, int *got) 2824 { 2825 int ret, flags; 2826 2827 BUG_ON(need & ~CEPH_CAP_FILE_RD); 2828 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO | 2829 CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL | 2830 CEPH_CAP_ANY_DIR_OPS)); 2831 if (need) { 2832 ret = ceph_pool_perm_check(inode, need); 2833 if (ret < 0) 2834 return ret; 2835 } 2836 2837 flags = get_used_fmode(need | want); 2838 if (nonblock) 2839 flags |= NON_BLOCKING; 2840 2841 ret = try_get_cap_refs(inode, need, want, 0, flags, got); 2842 /* three special error codes */ 2843 if (ret == -EAGAIN || ret == -EFBIG || ret == -ESTALE) 2844 ret = 0; 2845 return ret; 2846 } 2847 2848 /* 2849 * Wait for caps, and take cap references. If we can't get a WR cap 2850 * due to a small max_size, make sure we check_max_size (and possibly 2851 * ask the mds) so we don't get hung up indefinitely. 2852 */ 2853 int ceph_get_caps(struct file *filp, int need, int want, 2854 loff_t endoff, int *got, struct page **pinned_page) 2855 { 2856 struct ceph_file_info *fi = filp->private_data; 2857 struct inode *inode = file_inode(filp); 2858 struct ceph_inode_info *ci = ceph_inode(inode); 2859 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 2860 int ret, _got, flags; 2861 2862 ret = ceph_pool_perm_check(inode, need); 2863 if (ret < 0) 2864 return ret; 2865 2866 if ((fi->fmode & CEPH_FILE_MODE_WR) && 2867 fi->filp_gen != READ_ONCE(fsc->filp_gen)) 2868 return -EBADF; 2869 2870 flags = get_used_fmode(need | want); 2871 2872 while (true) { 2873 flags &= CEPH_FILE_MODE_MASK; 2874 if (atomic_read(&fi->num_locks)) 2875 flags |= CHECK_FILELOCK; 2876 _got = 0; 2877 ret = try_get_cap_refs(inode, need, want, endoff, 2878 flags, &_got); 2879 WARN_ON_ONCE(ret == -EAGAIN); 2880 if (!ret) { 2881 struct ceph_mds_client *mdsc = fsc->mdsc; 2882 struct cap_wait cw; 2883 DEFINE_WAIT_FUNC(wait, woken_wake_function); 2884 2885 cw.ino = ceph_ino(inode); 2886 cw.tgid = current->tgid; 2887 cw.need = need; 2888 cw.want = want; 2889 2890 spin_lock(&mdsc->caps_list_lock); 2891 list_add(&cw.list, &mdsc->cap_wait_list); 2892 spin_unlock(&mdsc->caps_list_lock); 2893 2894 /* make sure used fmode not timeout */ 2895 ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS); 2896 add_wait_queue(&ci->i_cap_wq, &wait); 2897 2898 flags |= NON_BLOCKING; 2899 while (!(ret = try_get_cap_refs(inode, need, want, 2900 endoff, flags, &_got))) { 2901 if (signal_pending(current)) { 2902 ret = -ERESTARTSYS; 2903 break; 2904 } 2905 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2906 } 2907 2908 remove_wait_queue(&ci->i_cap_wq, &wait); 2909 ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS); 2910 2911 spin_lock(&mdsc->caps_list_lock); 2912 list_del(&cw.list); 2913 spin_unlock(&mdsc->caps_list_lock); 2914 2915 if (ret == -EAGAIN) 2916 continue; 2917 } 2918 2919 if ((fi->fmode & CEPH_FILE_MODE_WR) && 2920 fi->filp_gen != READ_ONCE(fsc->filp_gen)) { 2921 if (ret >= 0 && _got) 2922 ceph_put_cap_refs(ci, _got); 2923 return -EBADF; 2924 } 2925 2926 if (ret < 0) { 2927 if (ret == -EFBIG || ret == -ESTALE) { 2928 int ret2 = ceph_wait_on_async_create(inode); 2929 if (ret2 < 0) 2930 return ret2; 2931 } 2932 if (ret == -EFBIG) { 2933 check_max_size(inode, endoff); 2934 continue; 2935 } 2936 if (ret == -ESTALE) { 2937 /* session was killed, try renew caps */ 2938 ret = ceph_renew_caps(inode, flags); 2939 if (ret == 0) 2940 continue; 2941 } 2942 return ret; 2943 } 2944 2945 if (S_ISREG(ci->vfs_inode.i_mode) && 2946 ci->i_inline_version != CEPH_INLINE_NONE && 2947 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) && 2948 i_size_read(inode) > 0) { 2949 struct page *page = 2950 find_get_page(inode->i_mapping, 0); 2951 if (page) { 2952 if (PageUptodate(page)) { 2953 *pinned_page = page; 2954 break; 2955 } 2956 put_page(page); 2957 } 2958 /* 2959 * drop cap refs first because getattr while 2960 * holding * caps refs can cause deadlock. 2961 */ 2962 ceph_put_cap_refs(ci, _got); 2963 _got = 0; 2964 2965 /* 2966 * getattr request will bring inline data into 2967 * page cache 2968 */ 2969 ret = __ceph_do_getattr(inode, NULL, 2970 CEPH_STAT_CAP_INLINE_DATA, 2971 true); 2972 if (ret < 0) 2973 return ret; 2974 continue; 2975 } 2976 break; 2977 } 2978 2979 if (S_ISREG(ci->vfs_inode.i_mode) && 2980 (_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE)) 2981 ceph_fscache_revalidate_cookie(ci); 2982 2983 *got = _got; 2984 return 0; 2985 } 2986 2987 /* 2988 * Take cap refs. Caller must already know we hold at least one ref 2989 * on the caps in question or we don't know this is safe. 2990 */ 2991 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps) 2992 { 2993 spin_lock(&ci->i_ceph_lock); 2994 ceph_take_cap_refs(ci, caps, false); 2995 spin_unlock(&ci->i_ceph_lock); 2996 } 2997 2998 2999 /* 3000 * drop cap_snap that is not associated with any snapshot. 3001 * we don't need to send FLUSHSNAP message for it. 3002 */ 3003 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci, 3004 struct ceph_cap_snap *capsnap) 3005 { 3006 if (!capsnap->need_flush && 3007 !capsnap->writing && !capsnap->dirty_pages) { 3008 dout("dropping cap_snap %p follows %llu\n", 3009 capsnap, capsnap->follows); 3010 BUG_ON(capsnap->cap_flush.tid > 0); 3011 ceph_put_snap_context(capsnap->context); 3012 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps)) 3013 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS; 3014 3015 list_del(&capsnap->ci_item); 3016 ceph_put_cap_snap(capsnap); 3017 return 1; 3018 } 3019 return 0; 3020 } 3021 3022 /* 3023 * Release cap refs. 3024 * 3025 * If we released the last ref on any given cap, call ceph_check_caps 3026 * to release (or schedule a release). 3027 * 3028 * If we are releasing a WR cap (from a sync write), finalize any affected 3029 * cap_snap, and wake up any waiters. 3030 */ 3031 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had, 3032 bool skip_checking_caps) 3033 { 3034 struct inode *inode = &ci->vfs_inode; 3035 int last = 0, put = 0, flushsnaps = 0, wake = 0; 3036 3037 spin_lock(&ci->i_ceph_lock); 3038 if (had & CEPH_CAP_PIN) 3039 --ci->i_pin_ref; 3040 if (had & CEPH_CAP_FILE_RD) 3041 if (--ci->i_rd_ref == 0) 3042 last++; 3043 if (had & CEPH_CAP_FILE_CACHE) 3044 if (--ci->i_rdcache_ref == 0) 3045 last++; 3046 if (had & CEPH_CAP_FILE_EXCL) 3047 if (--ci->i_fx_ref == 0) 3048 last++; 3049 if (had & CEPH_CAP_FILE_BUFFER) { 3050 if (--ci->i_wb_ref == 0) { 3051 last++; 3052 put++; 3053 } 3054 dout("put_cap_refs %p wb %d -> %d (?)\n", 3055 inode, ci->i_wb_ref+1, ci->i_wb_ref); 3056 } 3057 if (had & CEPH_CAP_FILE_WR) 3058 if (--ci->i_wr_ref == 0) { 3059 last++; 3060 if (__ceph_have_pending_cap_snap(ci)) { 3061 struct ceph_cap_snap *capsnap = 3062 list_last_entry(&ci->i_cap_snaps, 3063 struct ceph_cap_snap, 3064 ci_item); 3065 capsnap->writing = 0; 3066 if (ceph_try_drop_cap_snap(ci, capsnap)) 3067 put++; 3068 else if (__ceph_finish_cap_snap(ci, capsnap)) 3069 flushsnaps = 1; 3070 wake = 1; 3071 } 3072 if (ci->i_wrbuffer_ref_head == 0 && 3073 ci->i_dirty_caps == 0 && 3074 ci->i_flushing_caps == 0) { 3075 BUG_ON(!ci->i_head_snapc); 3076 ceph_put_snap_context(ci->i_head_snapc); 3077 ci->i_head_snapc = NULL; 3078 } 3079 /* see comment in __ceph_remove_cap() */ 3080 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm) 3081 drop_inode_snap_realm(ci); 3082 } 3083 spin_unlock(&ci->i_ceph_lock); 3084 3085 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had), 3086 last ? " last" : "", put ? " put" : ""); 3087 3088 if (last && !skip_checking_caps) 3089 ceph_check_caps(ci, 0, NULL); 3090 else if (flushsnaps) 3091 ceph_flush_snaps(ci, NULL); 3092 if (wake) 3093 wake_up_all(&ci->i_cap_wq); 3094 while (put-- > 0) 3095 iput(inode); 3096 } 3097 3098 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had) 3099 { 3100 __ceph_put_cap_refs(ci, had, false); 3101 } 3102 3103 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had) 3104 { 3105 __ceph_put_cap_refs(ci, had, true); 3106 } 3107 3108 /* 3109 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap 3110 * context. Adjust per-snap dirty page accounting as appropriate. 3111 * Once all dirty data for a cap_snap is flushed, flush snapped file 3112 * metadata back to the MDS. If we dropped the last ref, call 3113 * ceph_check_caps. 3114 */ 3115 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr, 3116 struct ceph_snap_context *snapc) 3117 { 3118 struct inode *inode = &ci->vfs_inode; 3119 struct ceph_cap_snap *capsnap = NULL; 3120 int put = 0; 3121 bool last = false; 3122 bool found = false; 3123 bool flush_snaps = false; 3124 bool complete_capsnap = false; 3125 3126 spin_lock(&ci->i_ceph_lock); 3127 ci->i_wrbuffer_ref -= nr; 3128 if (ci->i_wrbuffer_ref == 0) { 3129 last = true; 3130 put++; 3131 } 3132 3133 if (ci->i_head_snapc == snapc) { 3134 ci->i_wrbuffer_ref_head -= nr; 3135 if (ci->i_wrbuffer_ref_head == 0 && 3136 ci->i_wr_ref == 0 && 3137 ci->i_dirty_caps == 0 && 3138 ci->i_flushing_caps == 0) { 3139 BUG_ON(!ci->i_head_snapc); 3140 ceph_put_snap_context(ci->i_head_snapc); 3141 ci->i_head_snapc = NULL; 3142 } 3143 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n", 3144 inode, 3145 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr, 3146 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head, 3147 last ? " LAST" : ""); 3148 } else { 3149 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 3150 if (capsnap->context == snapc) { 3151 found = true; 3152 break; 3153 } 3154 } 3155 BUG_ON(!found); 3156 capsnap->dirty_pages -= nr; 3157 if (capsnap->dirty_pages == 0) { 3158 complete_capsnap = true; 3159 if (!capsnap->writing) { 3160 if (ceph_try_drop_cap_snap(ci, capsnap)) { 3161 put++; 3162 } else { 3163 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS; 3164 flush_snaps = true; 3165 } 3166 } 3167 } 3168 dout("put_wrbuffer_cap_refs on %p cap_snap %p " 3169 " snap %lld %d/%d -> %d/%d %s%s\n", 3170 inode, capsnap, capsnap->context->seq, 3171 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr, 3172 ci->i_wrbuffer_ref, capsnap->dirty_pages, 3173 last ? " (wrbuffer last)" : "", 3174 complete_capsnap ? " (complete capsnap)" : ""); 3175 } 3176 3177 spin_unlock(&ci->i_ceph_lock); 3178 3179 if (last) { 3180 ceph_check_caps(ci, 0, NULL); 3181 } else if (flush_snaps) { 3182 ceph_flush_snaps(ci, NULL); 3183 } 3184 if (complete_capsnap) 3185 wake_up_all(&ci->i_cap_wq); 3186 while (put-- > 0) { 3187 /* avoid calling iput_final() in osd dispatch threads */ 3188 ceph_async_iput(inode); 3189 } 3190 } 3191 3192 /* 3193 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP. 3194 */ 3195 static void invalidate_aliases(struct inode *inode) 3196 { 3197 struct dentry *dn, *prev = NULL; 3198 3199 dout("invalidate_aliases inode %p\n", inode); 3200 d_prune_aliases(inode); 3201 /* 3202 * For non-directory inode, d_find_alias() only returns 3203 * hashed dentry. After calling d_invalidate(), the 3204 * dentry becomes unhashed. 3205 * 3206 * For directory inode, d_find_alias() can return 3207 * unhashed dentry. But directory inode should have 3208 * one alias at most. 3209 */ 3210 while ((dn = d_find_alias(inode))) { 3211 if (dn == prev) { 3212 dput(dn); 3213 break; 3214 } 3215 d_invalidate(dn); 3216 if (prev) 3217 dput(prev); 3218 prev = dn; 3219 } 3220 if (prev) 3221 dput(prev); 3222 } 3223 3224 struct cap_extra_info { 3225 struct ceph_string *pool_ns; 3226 /* inline data */ 3227 u64 inline_version; 3228 void *inline_data; 3229 u32 inline_len; 3230 /* dirstat */ 3231 bool dirstat_valid; 3232 u64 nfiles; 3233 u64 nsubdirs; 3234 u64 change_attr; 3235 /* currently issued */ 3236 int issued; 3237 struct timespec64 btime; 3238 }; 3239 3240 /* 3241 * Handle a cap GRANT message from the MDS. (Note that a GRANT may 3242 * actually be a revocation if it specifies a smaller cap set.) 3243 * 3244 * caller holds s_mutex and i_ceph_lock, we drop both. 3245 */ 3246 static void handle_cap_grant(struct inode *inode, 3247 struct ceph_mds_session *session, 3248 struct ceph_cap *cap, 3249 struct ceph_mds_caps *grant, 3250 struct ceph_buffer *xattr_buf, 3251 struct cap_extra_info *extra_info) 3252 __releases(ci->i_ceph_lock) 3253 __releases(session->s_mdsc->snap_rwsem) 3254 { 3255 struct ceph_inode_info *ci = ceph_inode(inode); 3256 int seq = le32_to_cpu(grant->seq); 3257 int newcaps = le32_to_cpu(grant->caps); 3258 int used, wanted, dirty; 3259 u64 size = le64_to_cpu(grant->size); 3260 u64 max_size = le64_to_cpu(grant->max_size); 3261 unsigned char check_caps = 0; 3262 bool was_stale = cap->cap_gen < session->s_cap_gen; 3263 bool wake = false; 3264 bool writeback = false; 3265 bool queue_trunc = false; 3266 bool queue_invalidate = false; 3267 bool deleted_inode = false; 3268 bool fill_inline = false; 3269 3270 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n", 3271 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps)); 3272 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size, 3273 inode->i_size); 3274 3275 3276 /* 3277 * If CACHE is being revoked, and we have no dirty buffers, 3278 * try to invalidate (once). (If there are dirty buffers, we 3279 * will invalidate _after_ writeback.) 3280 */ 3281 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */ 3282 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) && 3283 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 && 3284 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) { 3285 if (try_nonblocking_invalidate(inode)) { 3286 /* there were locked pages.. invalidate later 3287 in a separate thread. */ 3288 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) { 3289 queue_invalidate = true; 3290 ci->i_rdcache_revoking = ci->i_rdcache_gen; 3291 } 3292 } 3293 } 3294 3295 if (was_stale) 3296 cap->issued = cap->implemented = CEPH_CAP_PIN; 3297 3298 /* 3299 * auth mds of the inode changed. we received the cap export message, 3300 * but still haven't received the cap import message. handle_cap_export 3301 * updated the new auth MDS' cap. 3302 * 3303 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message 3304 * that was sent before the cap import message. So don't remove caps. 3305 */ 3306 if (ceph_seq_cmp(seq, cap->seq) <= 0) { 3307 WARN_ON(cap != ci->i_auth_cap); 3308 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id)); 3309 seq = cap->seq; 3310 newcaps |= cap->issued; 3311 } 3312 3313 /* side effects now are allowed */ 3314 cap->cap_gen = session->s_cap_gen; 3315 cap->seq = seq; 3316 3317 __check_cap_issue(ci, cap, newcaps); 3318 3319 inode_set_max_iversion_raw(inode, extra_info->change_attr); 3320 3321 if ((newcaps & CEPH_CAP_AUTH_SHARED) && 3322 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) { 3323 inode->i_mode = le32_to_cpu(grant->mode); 3324 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid)); 3325 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid)); 3326 ci->i_btime = extra_info->btime; 3327 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode, 3328 from_kuid(&init_user_ns, inode->i_uid), 3329 from_kgid(&init_user_ns, inode->i_gid)); 3330 } 3331 3332 if ((newcaps & CEPH_CAP_LINK_SHARED) && 3333 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) { 3334 set_nlink(inode, le32_to_cpu(grant->nlink)); 3335 if (inode->i_nlink == 0 && 3336 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL))) 3337 deleted_inode = true; 3338 } 3339 3340 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 && 3341 grant->xattr_len) { 3342 int len = le32_to_cpu(grant->xattr_len); 3343 u64 version = le64_to_cpu(grant->xattr_version); 3344 3345 if (version > ci->i_xattrs.version) { 3346 dout(" got new xattrs v%llu on %p len %d\n", 3347 version, inode, len); 3348 if (ci->i_xattrs.blob) 3349 ceph_buffer_put(ci->i_xattrs.blob); 3350 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf); 3351 ci->i_xattrs.version = version; 3352 ceph_forget_all_cached_acls(inode); 3353 ceph_security_invalidate_secctx(inode); 3354 } 3355 } 3356 3357 if (newcaps & CEPH_CAP_ANY_RD) { 3358 struct timespec64 mtime, atime, ctime; 3359 /* ctime/mtime/atime? */ 3360 ceph_decode_timespec64(&mtime, &grant->mtime); 3361 ceph_decode_timespec64(&atime, &grant->atime); 3362 ceph_decode_timespec64(&ctime, &grant->ctime); 3363 ceph_fill_file_time(inode, extra_info->issued, 3364 le32_to_cpu(grant->time_warp_seq), 3365 &ctime, &mtime, &atime); 3366 } 3367 3368 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) { 3369 ci->i_files = extra_info->nfiles; 3370 ci->i_subdirs = extra_info->nsubdirs; 3371 } 3372 3373 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) { 3374 /* file layout may have changed */ 3375 s64 old_pool = ci->i_layout.pool_id; 3376 struct ceph_string *old_ns; 3377 3378 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout); 3379 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns, 3380 lockdep_is_held(&ci->i_ceph_lock)); 3381 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns); 3382 3383 if (ci->i_layout.pool_id != old_pool || 3384 extra_info->pool_ns != old_ns) 3385 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM; 3386 3387 extra_info->pool_ns = old_ns; 3388 3389 /* size/truncate_seq? */ 3390 queue_trunc = ceph_fill_file_size(inode, extra_info->issued, 3391 le32_to_cpu(grant->truncate_seq), 3392 le64_to_cpu(grant->truncate_size), 3393 size); 3394 } 3395 3396 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) { 3397 if (max_size != ci->i_max_size) { 3398 dout("max_size %lld -> %llu\n", 3399 ci->i_max_size, max_size); 3400 ci->i_max_size = max_size; 3401 if (max_size >= ci->i_wanted_max_size) { 3402 ci->i_wanted_max_size = 0; /* reset */ 3403 ci->i_requested_max_size = 0; 3404 } 3405 wake = true; 3406 } 3407 } 3408 3409 /* check cap bits */ 3410 wanted = __ceph_caps_wanted(ci); 3411 used = __ceph_caps_used(ci); 3412 dirty = __ceph_caps_dirty(ci); 3413 dout(" my wanted = %s, used = %s, dirty %s\n", 3414 ceph_cap_string(wanted), 3415 ceph_cap_string(used), 3416 ceph_cap_string(dirty)); 3417 3418 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) && 3419 (wanted & ~(cap->mds_wanted | newcaps))) { 3420 /* 3421 * If mds is importing cap, prior cap messages that update 3422 * 'wanted' may get dropped by mds (migrate seq mismatch). 3423 * 3424 * We don't send cap message to update 'wanted' if what we 3425 * want are already issued. If mds revokes caps, cap message 3426 * that releases caps also tells mds what we want. But if 3427 * caps got revoked by mds forcedly (session stale). We may 3428 * haven't told mds what we want. 3429 */ 3430 check_caps = 1; 3431 } 3432 3433 /* revocation, grant, or no-op? */ 3434 if (cap->issued & ~newcaps) { 3435 int revoking = cap->issued & ~newcaps; 3436 3437 dout("revocation: %s -> %s (revoking %s)\n", 3438 ceph_cap_string(cap->issued), 3439 ceph_cap_string(newcaps), 3440 ceph_cap_string(revoking)); 3441 if (S_ISREG(inode->i_mode) && 3442 (revoking & used & CEPH_CAP_FILE_BUFFER)) 3443 writeback = true; /* initiate writeback; will delay ack */ 3444 else if (queue_invalidate && 3445 revoking == CEPH_CAP_FILE_CACHE && 3446 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0) 3447 ; /* do nothing yet, invalidation will be queued */ 3448 else if (cap == ci->i_auth_cap) 3449 check_caps = 1; /* check auth cap only */ 3450 else 3451 check_caps = 2; /* check all caps */ 3452 cap->issued = newcaps; 3453 cap->implemented |= newcaps; 3454 } else if (cap->issued == newcaps) { 3455 dout("caps unchanged: %s -> %s\n", 3456 ceph_cap_string(cap->issued), ceph_cap_string(newcaps)); 3457 } else { 3458 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued), 3459 ceph_cap_string(newcaps)); 3460 /* non-auth MDS is revoking the newly grant caps ? */ 3461 if (cap == ci->i_auth_cap && 3462 __ceph_caps_revoking_other(ci, cap, newcaps)) 3463 check_caps = 2; 3464 3465 cap->issued = newcaps; 3466 cap->implemented |= newcaps; /* add bits only, to 3467 * avoid stepping on a 3468 * pending revocation */ 3469 wake = true; 3470 } 3471 BUG_ON(cap->issued & ~cap->implemented); 3472 3473 if (extra_info->inline_version > 0 && 3474 extra_info->inline_version >= ci->i_inline_version) { 3475 ci->i_inline_version = extra_info->inline_version; 3476 if (ci->i_inline_version != CEPH_INLINE_NONE && 3477 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO))) 3478 fill_inline = true; 3479 } 3480 3481 if (ci->i_auth_cap == cap && 3482 le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) { 3483 if (newcaps & ~extra_info->issued) 3484 wake = true; 3485 3486 if (ci->i_requested_max_size > max_size || 3487 !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) { 3488 /* re-request max_size if necessary */ 3489 ci->i_requested_max_size = 0; 3490 wake = true; 3491 } 3492 3493 ceph_kick_flushing_inode_caps(session, ci); 3494 spin_unlock(&ci->i_ceph_lock); 3495 up_read(&session->s_mdsc->snap_rwsem); 3496 } else { 3497 spin_unlock(&ci->i_ceph_lock); 3498 } 3499 3500 if (fill_inline) 3501 ceph_fill_inline_data(inode, NULL, extra_info->inline_data, 3502 extra_info->inline_len); 3503 3504 if (queue_trunc) 3505 ceph_queue_vmtruncate(inode); 3506 3507 if (writeback) 3508 /* 3509 * queue inode for writeback: we can't actually call 3510 * filemap_write_and_wait, etc. from message handler 3511 * context. 3512 */ 3513 ceph_queue_writeback(inode); 3514 if (queue_invalidate) 3515 ceph_queue_invalidate(inode); 3516 if (deleted_inode) 3517 invalidate_aliases(inode); 3518 if (wake) 3519 wake_up_all(&ci->i_cap_wq); 3520 3521 if (check_caps == 1) 3522 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL, 3523 session); 3524 else if (check_caps == 2) 3525 ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session); 3526 else 3527 mutex_unlock(&session->s_mutex); 3528 } 3529 3530 /* 3531 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the 3532 * MDS has been safely committed. 3533 */ 3534 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid, 3535 struct ceph_mds_caps *m, 3536 struct ceph_mds_session *session, 3537 struct ceph_cap *cap) 3538 __releases(ci->i_ceph_lock) 3539 { 3540 struct ceph_inode_info *ci = ceph_inode(inode); 3541 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 3542 struct ceph_cap_flush *cf, *tmp_cf; 3543 LIST_HEAD(to_remove); 3544 unsigned seq = le32_to_cpu(m->seq); 3545 int dirty = le32_to_cpu(m->dirty); 3546 int cleaned = 0; 3547 bool drop = false; 3548 bool wake_ci = false; 3549 bool wake_mdsc = false; 3550 3551 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) { 3552 /* Is this the one that was flushed? */ 3553 if (cf->tid == flush_tid) 3554 cleaned = cf->caps; 3555 3556 /* Is this a capsnap? */ 3557 if (cf->caps == 0) 3558 continue; 3559 3560 if (cf->tid <= flush_tid) { 3561 /* 3562 * An earlier or current tid. The FLUSH_ACK should 3563 * represent a superset of this flush's caps. 3564 */ 3565 wake_ci |= __detach_cap_flush_from_ci(ci, cf); 3566 list_add_tail(&cf->i_list, &to_remove); 3567 } else { 3568 /* 3569 * This is a later one. Any caps in it are still dirty 3570 * so don't count them as cleaned. 3571 */ 3572 cleaned &= ~cf->caps; 3573 if (!cleaned) 3574 break; 3575 } 3576 } 3577 3578 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s," 3579 " flushing %s -> %s\n", 3580 inode, session->s_mds, seq, ceph_cap_string(dirty), 3581 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps), 3582 ceph_cap_string(ci->i_flushing_caps & ~cleaned)); 3583 3584 if (list_empty(&to_remove) && !cleaned) 3585 goto out; 3586 3587 ci->i_flushing_caps &= ~cleaned; 3588 3589 spin_lock(&mdsc->cap_dirty_lock); 3590 3591 list_for_each_entry(cf, &to_remove, i_list) 3592 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf); 3593 3594 if (ci->i_flushing_caps == 0) { 3595 if (list_empty(&ci->i_cap_flush_list)) { 3596 list_del_init(&ci->i_flushing_item); 3597 if (!list_empty(&session->s_cap_flushing)) { 3598 dout(" mds%d still flushing cap on %p\n", 3599 session->s_mds, 3600 &list_first_entry(&session->s_cap_flushing, 3601 struct ceph_inode_info, 3602 i_flushing_item)->vfs_inode); 3603 } 3604 } 3605 mdsc->num_cap_flushing--; 3606 dout(" inode %p now !flushing\n", inode); 3607 3608 if (ci->i_dirty_caps == 0) { 3609 dout(" inode %p now clean\n", inode); 3610 BUG_ON(!list_empty(&ci->i_dirty_item)); 3611 drop = true; 3612 if (ci->i_wr_ref == 0 && 3613 ci->i_wrbuffer_ref_head == 0) { 3614 BUG_ON(!ci->i_head_snapc); 3615 ceph_put_snap_context(ci->i_head_snapc); 3616 ci->i_head_snapc = NULL; 3617 } 3618 } else { 3619 BUG_ON(list_empty(&ci->i_dirty_item)); 3620 } 3621 } 3622 spin_unlock(&mdsc->cap_dirty_lock); 3623 3624 out: 3625 spin_unlock(&ci->i_ceph_lock); 3626 3627 while (!list_empty(&to_remove)) { 3628 cf = list_first_entry(&to_remove, 3629 struct ceph_cap_flush, i_list); 3630 list_del(&cf->i_list); 3631 ceph_free_cap_flush(cf); 3632 } 3633 3634 if (wake_ci) 3635 wake_up_all(&ci->i_cap_wq); 3636 if (wake_mdsc) 3637 wake_up_all(&mdsc->cap_flushing_wq); 3638 if (drop) 3639 iput(inode); 3640 } 3641 3642 /* 3643 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can 3644 * throw away our cap_snap. 3645 * 3646 * Caller hold s_mutex. 3647 */ 3648 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid, 3649 struct ceph_mds_caps *m, 3650 struct ceph_mds_session *session) 3651 { 3652 struct ceph_inode_info *ci = ceph_inode(inode); 3653 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 3654 u64 follows = le64_to_cpu(m->snap_follows); 3655 struct ceph_cap_snap *capsnap; 3656 bool flushed = false; 3657 bool wake_ci = false; 3658 bool wake_mdsc = false; 3659 3660 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n", 3661 inode, ci, session->s_mds, follows); 3662 3663 spin_lock(&ci->i_ceph_lock); 3664 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 3665 if (capsnap->follows == follows) { 3666 if (capsnap->cap_flush.tid != flush_tid) { 3667 dout(" cap_snap %p follows %lld tid %lld !=" 3668 " %lld\n", capsnap, follows, 3669 flush_tid, capsnap->cap_flush.tid); 3670 break; 3671 } 3672 flushed = true; 3673 break; 3674 } else { 3675 dout(" skipping cap_snap %p follows %lld\n", 3676 capsnap, capsnap->follows); 3677 } 3678 } 3679 if (flushed) { 3680 WARN_ON(capsnap->dirty_pages || capsnap->writing); 3681 dout(" removing %p cap_snap %p follows %lld\n", 3682 inode, capsnap, follows); 3683 list_del(&capsnap->ci_item); 3684 wake_ci |= __detach_cap_flush_from_ci(ci, &capsnap->cap_flush); 3685 3686 spin_lock(&mdsc->cap_dirty_lock); 3687 3688 if (list_empty(&ci->i_cap_flush_list)) 3689 list_del_init(&ci->i_flushing_item); 3690 3691 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, 3692 &capsnap->cap_flush); 3693 spin_unlock(&mdsc->cap_dirty_lock); 3694 } 3695 spin_unlock(&ci->i_ceph_lock); 3696 if (flushed) { 3697 ceph_put_snap_context(capsnap->context); 3698 ceph_put_cap_snap(capsnap); 3699 if (wake_ci) 3700 wake_up_all(&ci->i_cap_wq); 3701 if (wake_mdsc) 3702 wake_up_all(&mdsc->cap_flushing_wq); 3703 iput(inode); 3704 } 3705 } 3706 3707 /* 3708 * Handle TRUNC from MDS, indicating file truncation. 3709 * 3710 * caller hold s_mutex. 3711 */ 3712 static bool handle_cap_trunc(struct inode *inode, 3713 struct ceph_mds_caps *trunc, 3714 struct ceph_mds_session *session) 3715 { 3716 struct ceph_inode_info *ci = ceph_inode(inode); 3717 int mds = session->s_mds; 3718 int seq = le32_to_cpu(trunc->seq); 3719 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq); 3720 u64 truncate_size = le64_to_cpu(trunc->truncate_size); 3721 u64 size = le64_to_cpu(trunc->size); 3722 int implemented = 0; 3723 int dirty = __ceph_caps_dirty(ci); 3724 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented); 3725 bool queue_trunc = false; 3726 3727 lockdep_assert_held(&ci->i_ceph_lock); 3728 3729 issued |= implemented | dirty; 3730 3731 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n", 3732 inode, mds, seq, truncate_size, truncate_seq); 3733 queue_trunc = ceph_fill_file_size(inode, issued, 3734 truncate_seq, truncate_size, size); 3735 return queue_trunc; 3736 } 3737 3738 /* 3739 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a 3740 * different one. If we are the most recent migration we've seen (as 3741 * indicated by mseq), make note of the migrating cap bits for the 3742 * duration (until we see the corresponding IMPORT). 3743 * 3744 * caller holds s_mutex 3745 */ 3746 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex, 3747 struct ceph_mds_cap_peer *ph, 3748 struct ceph_mds_session *session) 3749 { 3750 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 3751 struct ceph_mds_session *tsession = NULL; 3752 struct ceph_cap *cap, *tcap, *new_cap = NULL; 3753 struct ceph_inode_info *ci = ceph_inode(inode); 3754 u64 t_cap_id; 3755 unsigned mseq = le32_to_cpu(ex->migrate_seq); 3756 unsigned t_seq, t_mseq; 3757 int target, issued; 3758 int mds = session->s_mds; 3759 3760 if (ph) { 3761 t_cap_id = le64_to_cpu(ph->cap_id); 3762 t_seq = le32_to_cpu(ph->seq); 3763 t_mseq = le32_to_cpu(ph->mseq); 3764 target = le32_to_cpu(ph->mds); 3765 } else { 3766 t_cap_id = t_seq = t_mseq = 0; 3767 target = -1; 3768 } 3769 3770 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n", 3771 inode, ci, mds, mseq, target); 3772 retry: 3773 spin_lock(&ci->i_ceph_lock); 3774 cap = __get_cap_for_mds(ci, mds); 3775 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id)) 3776 goto out_unlock; 3777 3778 if (target < 0) { 3779 __ceph_remove_cap(cap, false); 3780 goto out_unlock; 3781 } 3782 3783 /* 3784 * now we know we haven't received the cap import message yet 3785 * because the exported cap still exist. 3786 */ 3787 3788 issued = cap->issued; 3789 if (issued != cap->implemented) 3790 pr_err_ratelimited("handle_cap_export: issued != implemented: " 3791 "ino (%llx.%llx) mds%d seq %d mseq %d " 3792 "issued %s implemented %s\n", 3793 ceph_vinop(inode), mds, cap->seq, cap->mseq, 3794 ceph_cap_string(issued), 3795 ceph_cap_string(cap->implemented)); 3796 3797 3798 tcap = __get_cap_for_mds(ci, target); 3799 if (tcap) { 3800 /* already have caps from the target */ 3801 if (tcap->cap_id == t_cap_id && 3802 ceph_seq_cmp(tcap->seq, t_seq) < 0) { 3803 dout(" updating import cap %p mds%d\n", tcap, target); 3804 tcap->cap_id = t_cap_id; 3805 tcap->seq = t_seq - 1; 3806 tcap->issue_seq = t_seq - 1; 3807 tcap->issued |= issued; 3808 tcap->implemented |= issued; 3809 if (cap == ci->i_auth_cap) { 3810 ci->i_auth_cap = tcap; 3811 change_auth_cap_ses(ci, tcap->session); 3812 } 3813 } 3814 __ceph_remove_cap(cap, false); 3815 goto out_unlock; 3816 } else if (tsession) { 3817 /* add placeholder for the export tagert */ 3818 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0; 3819 tcap = new_cap; 3820 ceph_add_cap(inode, tsession, t_cap_id, issued, 0, 3821 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap); 3822 3823 if (!list_empty(&ci->i_cap_flush_list) && 3824 ci->i_auth_cap == tcap) { 3825 spin_lock(&mdsc->cap_dirty_lock); 3826 list_move_tail(&ci->i_flushing_item, 3827 &tcap->session->s_cap_flushing); 3828 spin_unlock(&mdsc->cap_dirty_lock); 3829 } 3830 3831 __ceph_remove_cap(cap, false); 3832 goto out_unlock; 3833 } 3834 3835 spin_unlock(&ci->i_ceph_lock); 3836 mutex_unlock(&session->s_mutex); 3837 3838 /* open target session */ 3839 tsession = ceph_mdsc_open_export_target_session(mdsc, target); 3840 if (!IS_ERR(tsession)) { 3841 if (mds > target) { 3842 mutex_lock(&session->s_mutex); 3843 mutex_lock_nested(&tsession->s_mutex, 3844 SINGLE_DEPTH_NESTING); 3845 } else { 3846 mutex_lock(&tsession->s_mutex); 3847 mutex_lock_nested(&session->s_mutex, 3848 SINGLE_DEPTH_NESTING); 3849 } 3850 new_cap = ceph_get_cap(mdsc, NULL); 3851 } else { 3852 WARN_ON(1); 3853 tsession = NULL; 3854 target = -1; 3855 mutex_lock(&session->s_mutex); 3856 } 3857 goto retry; 3858 3859 out_unlock: 3860 spin_unlock(&ci->i_ceph_lock); 3861 mutex_unlock(&session->s_mutex); 3862 if (tsession) { 3863 mutex_unlock(&tsession->s_mutex); 3864 ceph_put_mds_session(tsession); 3865 } 3866 if (new_cap) 3867 ceph_put_cap(mdsc, new_cap); 3868 } 3869 3870 /* 3871 * Handle cap IMPORT. 3872 * 3873 * caller holds s_mutex. acquires i_ceph_lock 3874 */ 3875 static void handle_cap_import(struct ceph_mds_client *mdsc, 3876 struct inode *inode, struct ceph_mds_caps *im, 3877 struct ceph_mds_cap_peer *ph, 3878 struct ceph_mds_session *session, 3879 struct ceph_cap **target_cap, int *old_issued) 3880 { 3881 struct ceph_inode_info *ci = ceph_inode(inode); 3882 struct ceph_cap *cap, *ocap, *new_cap = NULL; 3883 int mds = session->s_mds; 3884 int issued; 3885 unsigned caps = le32_to_cpu(im->caps); 3886 unsigned wanted = le32_to_cpu(im->wanted); 3887 unsigned seq = le32_to_cpu(im->seq); 3888 unsigned mseq = le32_to_cpu(im->migrate_seq); 3889 u64 realmino = le64_to_cpu(im->realm); 3890 u64 cap_id = le64_to_cpu(im->cap_id); 3891 u64 p_cap_id; 3892 int peer; 3893 3894 if (ph) { 3895 p_cap_id = le64_to_cpu(ph->cap_id); 3896 peer = le32_to_cpu(ph->mds); 3897 } else { 3898 p_cap_id = 0; 3899 peer = -1; 3900 } 3901 3902 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n", 3903 inode, ci, mds, mseq, peer); 3904 retry: 3905 cap = __get_cap_for_mds(ci, mds); 3906 if (!cap) { 3907 if (!new_cap) { 3908 spin_unlock(&ci->i_ceph_lock); 3909 new_cap = ceph_get_cap(mdsc, NULL); 3910 spin_lock(&ci->i_ceph_lock); 3911 goto retry; 3912 } 3913 cap = new_cap; 3914 } else { 3915 if (new_cap) { 3916 ceph_put_cap(mdsc, new_cap); 3917 new_cap = NULL; 3918 } 3919 } 3920 3921 __ceph_caps_issued(ci, &issued); 3922 issued |= __ceph_caps_dirty(ci); 3923 3924 ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq, 3925 realmino, CEPH_CAP_FLAG_AUTH, &new_cap); 3926 3927 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL; 3928 if (ocap && ocap->cap_id == p_cap_id) { 3929 dout(" remove export cap %p mds%d flags %d\n", 3930 ocap, peer, ph->flags); 3931 if ((ph->flags & CEPH_CAP_FLAG_AUTH) && 3932 (ocap->seq != le32_to_cpu(ph->seq) || 3933 ocap->mseq != le32_to_cpu(ph->mseq))) { 3934 pr_err_ratelimited("handle_cap_import: " 3935 "mismatched seq/mseq: ino (%llx.%llx) " 3936 "mds%d seq %d mseq %d importer mds%d " 3937 "has peer seq %d mseq %d\n", 3938 ceph_vinop(inode), peer, ocap->seq, 3939 ocap->mseq, mds, le32_to_cpu(ph->seq), 3940 le32_to_cpu(ph->mseq)); 3941 } 3942 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE)); 3943 } 3944 3945 *old_issued = issued; 3946 *target_cap = cap; 3947 } 3948 3949 /* 3950 * Handle a caps message from the MDS. 3951 * 3952 * Identify the appropriate session, inode, and call the right handler 3953 * based on the cap op. 3954 */ 3955 void ceph_handle_caps(struct ceph_mds_session *session, 3956 struct ceph_msg *msg) 3957 { 3958 struct ceph_mds_client *mdsc = session->s_mdsc; 3959 struct inode *inode; 3960 struct ceph_inode_info *ci; 3961 struct ceph_cap *cap; 3962 struct ceph_mds_caps *h; 3963 struct ceph_mds_cap_peer *peer = NULL; 3964 struct ceph_snap_realm *realm = NULL; 3965 int op; 3966 int msg_version = le16_to_cpu(msg->hdr.version); 3967 u32 seq, mseq; 3968 struct ceph_vino vino; 3969 void *snaptrace; 3970 size_t snaptrace_len; 3971 void *p, *end; 3972 struct cap_extra_info extra_info = {}; 3973 bool queue_trunc; 3974 3975 dout("handle_caps from mds%d\n", session->s_mds); 3976 3977 /* decode */ 3978 end = msg->front.iov_base + msg->front.iov_len; 3979 if (msg->front.iov_len < sizeof(*h)) 3980 goto bad; 3981 h = msg->front.iov_base; 3982 op = le32_to_cpu(h->op); 3983 vino.ino = le64_to_cpu(h->ino); 3984 vino.snap = CEPH_NOSNAP; 3985 seq = le32_to_cpu(h->seq); 3986 mseq = le32_to_cpu(h->migrate_seq); 3987 3988 snaptrace = h + 1; 3989 snaptrace_len = le32_to_cpu(h->snap_trace_len); 3990 p = snaptrace + snaptrace_len; 3991 3992 if (msg_version >= 2) { 3993 u32 flock_len; 3994 ceph_decode_32_safe(&p, end, flock_len, bad); 3995 if (p + flock_len > end) 3996 goto bad; 3997 p += flock_len; 3998 } 3999 4000 if (msg_version >= 3) { 4001 if (op == CEPH_CAP_OP_IMPORT) { 4002 if (p + sizeof(*peer) > end) 4003 goto bad; 4004 peer = p; 4005 p += sizeof(*peer); 4006 } else if (op == CEPH_CAP_OP_EXPORT) { 4007 /* recorded in unused fields */ 4008 peer = (void *)&h->size; 4009 } 4010 } 4011 4012 if (msg_version >= 4) { 4013 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad); 4014 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad); 4015 if (p + extra_info.inline_len > end) 4016 goto bad; 4017 extra_info.inline_data = p; 4018 p += extra_info.inline_len; 4019 } 4020 4021 if (msg_version >= 5) { 4022 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc; 4023 u32 epoch_barrier; 4024 4025 ceph_decode_32_safe(&p, end, epoch_barrier, bad); 4026 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier); 4027 } 4028 4029 if (msg_version >= 8) { 4030 u64 flush_tid; 4031 u32 caller_uid, caller_gid; 4032 u32 pool_ns_len; 4033 4034 /* version >= 6 */ 4035 ceph_decode_64_safe(&p, end, flush_tid, bad); 4036 /* version >= 7 */ 4037 ceph_decode_32_safe(&p, end, caller_uid, bad); 4038 ceph_decode_32_safe(&p, end, caller_gid, bad); 4039 /* version >= 8 */ 4040 ceph_decode_32_safe(&p, end, pool_ns_len, bad); 4041 if (pool_ns_len > 0) { 4042 ceph_decode_need(&p, end, pool_ns_len, bad); 4043 extra_info.pool_ns = 4044 ceph_find_or_create_string(p, pool_ns_len); 4045 p += pool_ns_len; 4046 } 4047 } 4048 4049 if (msg_version >= 9) { 4050 struct ceph_timespec *btime; 4051 4052 if (p + sizeof(*btime) > end) 4053 goto bad; 4054 btime = p; 4055 ceph_decode_timespec64(&extra_info.btime, btime); 4056 p += sizeof(*btime); 4057 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad); 4058 } 4059 4060 if (msg_version >= 11) { 4061 u32 flags; 4062 /* version >= 10 */ 4063 ceph_decode_32_safe(&p, end, flags, bad); 4064 /* version >= 11 */ 4065 extra_info.dirstat_valid = true; 4066 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad); 4067 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad); 4068 } 4069 4070 /* lookup ino */ 4071 inode = ceph_find_inode(mdsc->fsc->sb, vino); 4072 ci = ceph_inode(inode); 4073 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino, 4074 vino.snap, inode); 4075 4076 mutex_lock(&session->s_mutex); 4077 session->s_seq++; 4078 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq, 4079 (unsigned)seq); 4080 4081 if (!inode) { 4082 dout(" i don't have ino %llx\n", vino.ino); 4083 4084 if (op == CEPH_CAP_OP_IMPORT) { 4085 cap = ceph_get_cap(mdsc, NULL); 4086 cap->cap_ino = vino.ino; 4087 cap->queue_release = 1; 4088 cap->cap_id = le64_to_cpu(h->cap_id); 4089 cap->mseq = mseq; 4090 cap->seq = seq; 4091 cap->issue_seq = seq; 4092 spin_lock(&session->s_cap_lock); 4093 __ceph_queue_cap_release(session, cap); 4094 spin_unlock(&session->s_cap_lock); 4095 } 4096 goto flush_cap_releases; 4097 } 4098 4099 /* these will work even if we don't have a cap yet */ 4100 switch (op) { 4101 case CEPH_CAP_OP_FLUSHSNAP_ACK: 4102 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid), 4103 h, session); 4104 goto done; 4105 4106 case CEPH_CAP_OP_EXPORT: 4107 handle_cap_export(inode, h, peer, session); 4108 goto done_unlocked; 4109 4110 case CEPH_CAP_OP_IMPORT: 4111 realm = NULL; 4112 if (snaptrace_len) { 4113 down_write(&mdsc->snap_rwsem); 4114 ceph_update_snap_trace(mdsc, snaptrace, 4115 snaptrace + snaptrace_len, 4116 false, &realm); 4117 downgrade_write(&mdsc->snap_rwsem); 4118 } else { 4119 down_read(&mdsc->snap_rwsem); 4120 } 4121 spin_lock(&ci->i_ceph_lock); 4122 handle_cap_import(mdsc, inode, h, peer, session, 4123 &cap, &extra_info.issued); 4124 handle_cap_grant(inode, session, cap, 4125 h, msg->middle, &extra_info); 4126 if (realm) 4127 ceph_put_snap_realm(mdsc, realm); 4128 goto done_unlocked; 4129 } 4130 4131 /* the rest require a cap */ 4132 spin_lock(&ci->i_ceph_lock); 4133 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds); 4134 if (!cap) { 4135 dout(" no cap on %p ino %llx.%llx from mds%d\n", 4136 inode, ceph_ino(inode), ceph_snap(inode), 4137 session->s_mds); 4138 spin_unlock(&ci->i_ceph_lock); 4139 goto flush_cap_releases; 4140 } 4141 4142 /* note that each of these drops i_ceph_lock for us */ 4143 switch (op) { 4144 case CEPH_CAP_OP_REVOKE: 4145 case CEPH_CAP_OP_GRANT: 4146 __ceph_caps_issued(ci, &extra_info.issued); 4147 extra_info.issued |= __ceph_caps_dirty(ci); 4148 handle_cap_grant(inode, session, cap, 4149 h, msg->middle, &extra_info); 4150 goto done_unlocked; 4151 4152 case CEPH_CAP_OP_FLUSH_ACK: 4153 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid), 4154 h, session, cap); 4155 break; 4156 4157 case CEPH_CAP_OP_TRUNC: 4158 queue_trunc = handle_cap_trunc(inode, h, session); 4159 spin_unlock(&ci->i_ceph_lock); 4160 if (queue_trunc) 4161 ceph_queue_vmtruncate(inode); 4162 break; 4163 4164 default: 4165 spin_unlock(&ci->i_ceph_lock); 4166 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op, 4167 ceph_cap_op_name(op)); 4168 } 4169 4170 done: 4171 mutex_unlock(&session->s_mutex); 4172 done_unlocked: 4173 ceph_put_string(extra_info.pool_ns); 4174 /* avoid calling iput_final() in mds dispatch threads */ 4175 ceph_async_iput(inode); 4176 return; 4177 4178 flush_cap_releases: 4179 /* 4180 * send any cap release message to try to move things 4181 * along for the mds (who clearly thinks we still have this 4182 * cap). 4183 */ 4184 ceph_flush_cap_releases(mdsc, session); 4185 goto done; 4186 4187 bad: 4188 pr_err("ceph_handle_caps: corrupt message\n"); 4189 ceph_msg_dump(msg); 4190 return; 4191 } 4192 4193 /* 4194 * Delayed work handler to process end of delayed cap release LRU list. 4195 */ 4196 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc) 4197 { 4198 struct inode *inode; 4199 struct ceph_inode_info *ci; 4200 4201 dout("check_delayed_caps\n"); 4202 spin_lock(&mdsc->cap_delay_lock); 4203 while (!list_empty(&mdsc->cap_delay_list)) { 4204 ci = list_first_entry(&mdsc->cap_delay_list, 4205 struct ceph_inode_info, 4206 i_cap_delay_list); 4207 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 && 4208 time_before(jiffies, ci->i_hold_caps_max)) 4209 break; 4210 list_del_init(&ci->i_cap_delay_list); 4211 4212 inode = igrab(&ci->vfs_inode); 4213 if (inode) { 4214 spin_unlock(&mdsc->cap_delay_lock); 4215 dout("check_delayed_caps on %p\n", inode); 4216 ceph_check_caps(ci, 0, NULL); 4217 /* avoid calling iput_final() in tick thread */ 4218 ceph_async_iput(inode); 4219 spin_lock(&mdsc->cap_delay_lock); 4220 } 4221 } 4222 spin_unlock(&mdsc->cap_delay_lock); 4223 } 4224 4225 /* 4226 * Flush all dirty caps to the mds 4227 */ 4228 static void flush_dirty_session_caps(struct ceph_mds_session *s) 4229 { 4230 struct ceph_mds_client *mdsc = s->s_mdsc; 4231 struct ceph_inode_info *ci; 4232 struct inode *inode; 4233 4234 dout("flush_dirty_caps\n"); 4235 spin_lock(&mdsc->cap_dirty_lock); 4236 while (!list_empty(&s->s_cap_dirty)) { 4237 ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info, 4238 i_dirty_item); 4239 inode = &ci->vfs_inode; 4240 ihold(inode); 4241 dout("flush_dirty_caps %p\n", inode); 4242 spin_unlock(&mdsc->cap_dirty_lock); 4243 ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL); 4244 iput(inode); 4245 spin_lock(&mdsc->cap_dirty_lock); 4246 } 4247 spin_unlock(&mdsc->cap_dirty_lock); 4248 dout("flush_dirty_caps done\n"); 4249 } 4250 4251 static void iterate_sessions(struct ceph_mds_client *mdsc, 4252 void (*cb)(struct ceph_mds_session *)) 4253 { 4254 int mds; 4255 4256 mutex_lock(&mdsc->mutex); 4257 for (mds = 0; mds < mdsc->max_sessions; ++mds) { 4258 struct ceph_mds_session *s; 4259 4260 if (!mdsc->sessions[mds]) 4261 continue; 4262 4263 s = ceph_get_mds_session(mdsc->sessions[mds]); 4264 if (!s) 4265 continue; 4266 4267 mutex_unlock(&mdsc->mutex); 4268 cb(s); 4269 ceph_put_mds_session(s); 4270 mutex_lock(&mdsc->mutex); 4271 } 4272 mutex_unlock(&mdsc->mutex); 4273 } 4274 4275 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc) 4276 { 4277 iterate_sessions(mdsc, flush_dirty_session_caps); 4278 } 4279 4280 void __ceph_touch_fmode(struct ceph_inode_info *ci, 4281 struct ceph_mds_client *mdsc, int fmode) 4282 { 4283 unsigned long now = jiffies; 4284 if (fmode & CEPH_FILE_MODE_RD) 4285 ci->i_last_rd = now; 4286 if (fmode & CEPH_FILE_MODE_WR) 4287 ci->i_last_wr = now; 4288 /* queue periodic check */ 4289 if (fmode && 4290 __ceph_is_any_real_caps(ci) && 4291 list_empty(&ci->i_cap_delay_list)) 4292 __cap_delay_requeue(mdsc, ci); 4293 } 4294 4295 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count) 4296 { 4297 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb); 4298 int bits = (fmode << 1) | 1; 4299 bool is_opened = false; 4300 int i; 4301 4302 if (count == 1) 4303 atomic64_inc(&mdsc->metric.opened_files); 4304 4305 spin_lock(&ci->i_ceph_lock); 4306 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) { 4307 if (bits & (1 << i)) 4308 ci->i_nr_by_mode[i] += count; 4309 4310 /* 4311 * If any of the mode ref is larger than 1, 4312 * that means it has been already opened by 4313 * others. Just skip checking the PIN ref. 4314 */ 4315 if (i && ci->i_nr_by_mode[i] > 1) 4316 is_opened = true; 4317 } 4318 4319 if (!is_opened) 4320 percpu_counter_inc(&mdsc->metric.opened_inodes); 4321 spin_unlock(&ci->i_ceph_lock); 4322 } 4323 4324 /* 4325 * Drop open file reference. If we were the last open file, 4326 * we may need to release capabilities to the MDS (or schedule 4327 * their delayed release). 4328 */ 4329 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count) 4330 { 4331 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb); 4332 int bits = (fmode << 1) | 1; 4333 bool is_closed = true; 4334 int i; 4335 4336 if (count == 1) 4337 atomic64_dec(&mdsc->metric.opened_files); 4338 4339 spin_lock(&ci->i_ceph_lock); 4340 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) { 4341 if (bits & (1 << i)) { 4342 BUG_ON(ci->i_nr_by_mode[i] < count); 4343 ci->i_nr_by_mode[i] -= count; 4344 } 4345 4346 /* 4347 * If any of the mode ref is not 0 after 4348 * decreased, that means it is still opened 4349 * by others. Just skip checking the PIN ref. 4350 */ 4351 if (i && ci->i_nr_by_mode[i]) 4352 is_closed = false; 4353 } 4354 4355 if (is_closed) 4356 percpu_counter_dec(&mdsc->metric.opened_inodes); 4357 spin_unlock(&ci->i_ceph_lock); 4358 } 4359 4360 /* 4361 * For a soon-to-be unlinked file, drop the LINK caps. If it 4362 * looks like the link count will hit 0, drop any other caps (other 4363 * than PIN) we don't specifically want (due to the file still being 4364 * open). 4365 */ 4366 int ceph_drop_caps_for_unlink(struct inode *inode) 4367 { 4368 struct ceph_inode_info *ci = ceph_inode(inode); 4369 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL; 4370 4371 spin_lock(&ci->i_ceph_lock); 4372 if (inode->i_nlink == 1) { 4373 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN); 4374 4375 if (__ceph_caps_dirty(ci)) { 4376 struct ceph_mds_client *mdsc = 4377 ceph_inode_to_client(inode)->mdsc; 4378 __cap_delay_requeue_front(mdsc, ci); 4379 } 4380 } 4381 spin_unlock(&ci->i_ceph_lock); 4382 return drop; 4383 } 4384 4385 /* 4386 * Helpers for embedding cap and dentry lease releases into mds 4387 * requests. 4388 * 4389 * @force is used by dentry_release (below) to force inclusion of a 4390 * record for the directory inode, even when there aren't any caps to 4391 * drop. 4392 */ 4393 int ceph_encode_inode_release(void **p, struct inode *inode, 4394 int mds, int drop, int unless, int force) 4395 { 4396 struct ceph_inode_info *ci = ceph_inode(inode); 4397 struct ceph_cap *cap; 4398 struct ceph_mds_request_release *rel = *p; 4399 int used, dirty; 4400 int ret = 0; 4401 4402 spin_lock(&ci->i_ceph_lock); 4403 used = __ceph_caps_used(ci); 4404 dirty = __ceph_caps_dirty(ci); 4405 4406 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n", 4407 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop), 4408 ceph_cap_string(unless)); 4409 4410 /* only drop unused, clean caps */ 4411 drop &= ~(used | dirty); 4412 4413 cap = __get_cap_for_mds(ci, mds); 4414 if (cap && __cap_is_valid(cap)) { 4415 unless &= cap->issued; 4416 if (unless) { 4417 if (unless & CEPH_CAP_AUTH_EXCL) 4418 drop &= ~CEPH_CAP_AUTH_SHARED; 4419 if (unless & CEPH_CAP_LINK_EXCL) 4420 drop &= ~CEPH_CAP_LINK_SHARED; 4421 if (unless & CEPH_CAP_XATTR_EXCL) 4422 drop &= ~CEPH_CAP_XATTR_SHARED; 4423 if (unless & CEPH_CAP_FILE_EXCL) 4424 drop &= ~CEPH_CAP_FILE_SHARED; 4425 } 4426 4427 if (force || (cap->issued & drop)) { 4428 if (cap->issued & drop) { 4429 int wanted = __ceph_caps_wanted(ci); 4430 dout("encode_inode_release %p cap %p " 4431 "%s -> %s, wanted %s -> %s\n", inode, cap, 4432 ceph_cap_string(cap->issued), 4433 ceph_cap_string(cap->issued & ~drop), 4434 ceph_cap_string(cap->mds_wanted), 4435 ceph_cap_string(wanted)); 4436 4437 cap->issued &= ~drop; 4438 cap->implemented &= ~drop; 4439 cap->mds_wanted = wanted; 4440 if (cap == ci->i_auth_cap && 4441 !(wanted & CEPH_CAP_ANY_FILE_WR)) 4442 ci->i_requested_max_size = 0; 4443 } else { 4444 dout("encode_inode_release %p cap %p %s" 4445 " (force)\n", inode, cap, 4446 ceph_cap_string(cap->issued)); 4447 } 4448 4449 rel->ino = cpu_to_le64(ceph_ino(inode)); 4450 rel->cap_id = cpu_to_le64(cap->cap_id); 4451 rel->seq = cpu_to_le32(cap->seq); 4452 rel->issue_seq = cpu_to_le32(cap->issue_seq); 4453 rel->mseq = cpu_to_le32(cap->mseq); 4454 rel->caps = cpu_to_le32(cap->implemented); 4455 rel->wanted = cpu_to_le32(cap->mds_wanted); 4456 rel->dname_len = 0; 4457 rel->dname_seq = 0; 4458 *p += sizeof(*rel); 4459 ret = 1; 4460 } else { 4461 dout("encode_inode_release %p cap %p %s (noop)\n", 4462 inode, cap, ceph_cap_string(cap->issued)); 4463 } 4464 } 4465 spin_unlock(&ci->i_ceph_lock); 4466 return ret; 4467 } 4468 4469 int ceph_encode_dentry_release(void **p, struct dentry *dentry, 4470 struct inode *dir, 4471 int mds, int drop, int unless) 4472 { 4473 struct dentry *parent = NULL; 4474 struct ceph_mds_request_release *rel = *p; 4475 struct ceph_dentry_info *di = ceph_dentry(dentry); 4476 int force = 0; 4477 int ret; 4478 4479 /* 4480 * force an record for the directory caps if we have a dentry lease. 4481 * this is racy (can't take i_ceph_lock and d_lock together), but it 4482 * doesn't have to be perfect; the mds will revoke anything we don't 4483 * release. 4484 */ 4485 spin_lock(&dentry->d_lock); 4486 if (di->lease_session && di->lease_session->s_mds == mds) 4487 force = 1; 4488 if (!dir) { 4489 parent = dget(dentry->d_parent); 4490 dir = d_inode(parent); 4491 } 4492 spin_unlock(&dentry->d_lock); 4493 4494 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force); 4495 dput(parent); 4496 4497 spin_lock(&dentry->d_lock); 4498 if (ret && di->lease_session && di->lease_session->s_mds == mds) { 4499 dout("encode_dentry_release %p mds%d seq %d\n", 4500 dentry, mds, (int)di->lease_seq); 4501 rel->dname_len = cpu_to_le32(dentry->d_name.len); 4502 memcpy(*p, dentry->d_name.name, dentry->d_name.len); 4503 *p += dentry->d_name.len; 4504 rel->dname_seq = cpu_to_le32(di->lease_seq); 4505 __ceph_mdsc_drop_dentry_lease(dentry); 4506 } 4507 spin_unlock(&dentry->d_lock); 4508 return ret; 4509 } 4510