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