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