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