1 /* 2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 #include <vm/vm_page2.h> 36 37 #include "hammer.h" 38 39 static int hammer_unload_inode(struct hammer_inode *ip); 40 static void hammer_free_inode(hammer_inode_t ip); 41 static void hammer_flush_inode_core(hammer_inode_t ip, 42 hammer_flush_group_t flg, int flags); 43 static int hammer_setup_child_callback(hammer_record_t rec, void *data); 44 #if 0 45 static int hammer_syncgrp_child_callback(hammer_record_t rec, void *data); 46 #endif 47 static int hammer_setup_parent_inodes(hammer_inode_t ip, int depth, 48 hammer_flush_group_t flg); 49 static int hammer_setup_parent_inodes_helper(hammer_record_t record, 50 int depth, hammer_flush_group_t flg); 51 static void hammer_inode_wakereclaims(hammer_inode_t ip); 52 static struct hammer_inostats *hammer_inode_inostats(hammer_mount_t hmp, 53 pid_t pid); 54 55 #ifdef DEBUG_TRUNCATE 56 extern struct hammer_inode *HammerTruncIp; 57 #endif 58 59 struct krate hammer_gen_krate = { 1 }; 60 61 /* 62 * RB-Tree support for inode structures 63 */ 64 int 65 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2) 66 { 67 if (ip1->obj_localization < ip2->obj_localization) 68 return(-1); 69 if (ip1->obj_localization > ip2->obj_localization) 70 return(1); 71 if (ip1->obj_id < ip2->obj_id) 72 return(-1); 73 if (ip1->obj_id > ip2->obj_id) 74 return(1); 75 if (ip1->obj_asof < ip2->obj_asof) 76 return(-1); 77 if (ip1->obj_asof > ip2->obj_asof) 78 return(1); 79 return(0); 80 } 81 82 int 83 hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2) 84 { 85 if (ip1->redo_fifo_start < ip2->redo_fifo_start) 86 return(-1); 87 if (ip1->redo_fifo_start > ip2->redo_fifo_start) 88 return(1); 89 return(0); 90 } 91 92 /* 93 * RB-Tree support for inode structures / special LOOKUP_INFO 94 */ 95 static int 96 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip) 97 { 98 if (info->obj_localization < ip->obj_localization) 99 return(-1); 100 if (info->obj_localization > ip->obj_localization) 101 return(1); 102 if (info->obj_id < ip->obj_id) 103 return(-1); 104 if (info->obj_id > ip->obj_id) 105 return(1); 106 if (info->obj_asof < ip->obj_asof) 107 return(-1); 108 if (info->obj_asof > ip->obj_asof) 109 return(1); 110 return(0); 111 } 112 113 /* 114 * Used by hammer_scan_inode_snapshots() to locate all of an object's 115 * snapshots. Note that the asof field is not tested, which we can get 116 * away with because it is the lowest-priority field. 117 */ 118 static int 119 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data) 120 { 121 hammer_inode_info_t info = data; 122 123 if (ip->obj_localization > info->obj_localization) 124 return(1); 125 if (ip->obj_localization < info->obj_localization) 126 return(-1); 127 if (ip->obj_id > info->obj_id) 128 return(1); 129 if (ip->obj_id < info->obj_id) 130 return(-1); 131 return(0); 132 } 133 134 /* 135 * Used by hammer_unload_pseudofs() to locate all inodes associated with 136 * a particular PFS. 137 */ 138 static int 139 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data) 140 { 141 u_int32_t localization = *(u_int32_t *)data; 142 if (ip->obj_localization > localization) 143 return(1); 144 if (ip->obj_localization < localization) 145 return(-1); 146 return(0); 147 } 148 149 /* 150 * RB-Tree support for pseudofs structures 151 */ 152 static int 153 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2) 154 { 155 if (p1->localization < p2->localization) 156 return(-1); 157 if (p1->localization > p2->localization) 158 return(1); 159 return(0); 160 } 161 162 163 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare); 164 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node, 165 hammer_inode_info_cmp, hammer_inode_info_t); 166 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node, 167 hammer_pfs_rb_compare, u_int32_t, localization); 168 169 /* 170 * The kernel is not actively referencing this vnode but is still holding 171 * it cached. 172 * 173 * This is called from the frontend. 174 * 175 * MPALMOSTSAFE 176 */ 177 int 178 hammer_vop_inactive(struct vop_inactive_args *ap) 179 { 180 struct hammer_inode *ip = VTOI(ap->a_vp); 181 hammer_mount_t hmp; 182 183 /* 184 * Degenerate case 185 */ 186 if (ip == NULL) { 187 vrecycle(ap->a_vp); 188 return(0); 189 } 190 191 /* 192 * If the inode no longer has visibility in the filesystem try to 193 * recycle it immediately, even if the inode is dirty. Recycling 194 * it quickly allows the system to reclaim buffer cache and VM 195 * resources which can matter a lot in a heavily loaded system. 196 * 197 * This can deadlock in vfsync() if we aren't careful. 198 * 199 * Do not queue the inode to the flusher if we still have visibility, 200 * otherwise namespace calls such as chmod will unnecessarily generate 201 * multiple inode updates. 202 */ 203 if (ip->ino_data.nlinks == 0) { 204 hmp = ip->hmp; 205 lwkt_gettoken(&hmp->fs_token); 206 hammer_inode_unloadable_check(ip, 0); 207 if (ip->flags & HAMMER_INODE_MODMASK) 208 hammer_flush_inode(ip, 0); 209 lwkt_reltoken(&hmp->fs_token); 210 vrecycle(ap->a_vp); 211 } 212 return(0); 213 } 214 215 /* 216 * Release the vnode association. This is typically (but not always) 217 * the last reference on the inode. 218 * 219 * Once the association is lost we are on our own with regards to 220 * flushing the inode. 221 * 222 * We must interlock ip->vp so hammer_get_vnode() can avoid races. 223 */ 224 int 225 hammer_vop_reclaim(struct vop_reclaim_args *ap) 226 { 227 struct hammer_inode *ip; 228 hammer_mount_t hmp; 229 struct vnode *vp; 230 231 vp = ap->a_vp; 232 233 if ((ip = vp->v_data) != NULL) { 234 hmp = ip->hmp; 235 lwkt_gettoken(&hmp->fs_token); 236 hammer_lock_ex(&ip->lock); 237 vp->v_data = NULL; 238 ip->vp = NULL; 239 240 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) { 241 ++hammer_count_reclaims; 242 ++hmp->count_reclaims; 243 ip->flags |= HAMMER_INODE_RECLAIM; 244 } 245 hammer_unlock(&ip->lock); 246 vclrisdirty(vp); 247 hammer_rel_inode(ip, 1); 248 lwkt_reltoken(&hmp->fs_token); 249 } 250 return(0); 251 } 252 253 /* 254 * Inform the kernel that the inode is dirty. This will be checked 255 * by vn_unlock(). 256 * 257 * Theoretically in order to reclaim a vnode the hammer_vop_reclaim() 258 * must be called which will interlock against our inode lock, so 259 * if VRECLAIMED is not set vp->v_mount (as used by vsetisdirty()) 260 * should be stable without having to acquire any new locks. 261 */ 262 void 263 hammer_inode_dirty(struct hammer_inode *ip) 264 { 265 struct vnode *vp; 266 267 if ((ip->flags & HAMMER_INODE_MODMASK) && 268 (vp = ip->vp) != NULL && 269 (vp->v_flag & (VRECLAIMED | VISDIRTY)) == 0) { 270 vsetisdirty(vp); 271 } 272 } 273 274 /* 275 * Return a locked vnode for the specified inode. The inode must be 276 * referenced but NOT LOCKED on entry and will remain referenced on 277 * return. 278 * 279 * Called from the frontend. 280 */ 281 int 282 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp) 283 { 284 hammer_mount_t hmp; 285 struct vnode *vp; 286 int error = 0; 287 u_int8_t obj_type; 288 289 hmp = ip->hmp; 290 291 for (;;) { 292 if ((vp = ip->vp) == NULL) { 293 error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0); 294 if (error) 295 break; 296 hammer_lock_ex(&ip->lock); 297 if (ip->vp != NULL) { 298 hammer_unlock(&ip->lock); 299 vp = *vpp; 300 vp->v_type = VBAD; 301 vx_put(vp); 302 continue; 303 } 304 hammer_ref(&ip->lock); 305 vp = *vpp; 306 ip->vp = vp; 307 308 obj_type = ip->ino_data.obj_type; 309 vp->v_type = hammer_get_vnode_type(obj_type); 310 311 hammer_inode_wakereclaims(ip); 312 313 switch(ip->ino_data.obj_type) { 314 case HAMMER_OBJTYPE_CDEV: 315 case HAMMER_OBJTYPE_BDEV: 316 vp->v_ops = &hmp->mp->mnt_vn_spec_ops; 317 addaliasu(vp, ip->ino_data.rmajor, 318 ip->ino_data.rminor); 319 break; 320 case HAMMER_OBJTYPE_FIFO: 321 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops; 322 break; 323 case HAMMER_OBJTYPE_REGFILE: 324 break; 325 default: 326 break; 327 } 328 329 /* 330 * Only mark as the root vnode if the ip is not 331 * historical, otherwise the VFS cache will get 332 * confused. The other half of the special handling 333 * is in hammer_vop_nlookupdotdot(). 334 * 335 * Pseudo-filesystem roots can be accessed via 336 * non-root filesystem paths and setting VROOT may 337 * confuse the namecache. Set VPFSROOT instead. 338 */ 339 if (ip->obj_id == HAMMER_OBJID_ROOT) { 340 if (ip->obj_asof == hmp->asof) { 341 if (ip->obj_localization == 0) 342 vsetflags(vp, VROOT); 343 else 344 vsetflags(vp, VPFSROOT); 345 } else { 346 vsetflags(vp, VPFSROOT); 347 } 348 } 349 350 vp->v_data = (void *)ip; 351 /* vnode locked by getnewvnode() */ 352 /* make related vnode dirty if inode dirty? */ 353 hammer_unlock(&ip->lock); 354 if (vp->v_type == VREG) { 355 vinitvmio(vp, ip->ino_data.size, 356 hammer_blocksize(ip->ino_data.size), 357 hammer_blockoff(ip->ino_data.size)); 358 } 359 break; 360 } 361 362 /* 363 * Interlock vnode clearing. This does not prevent the 364 * vnode from going into a reclaimed state but it does 365 * prevent it from being destroyed or reused so the vget() 366 * will properly fail. 367 */ 368 hammer_lock_ex(&ip->lock); 369 if ((vp = ip->vp) == NULL) { 370 hammer_unlock(&ip->lock); 371 continue; 372 } 373 vhold(vp); 374 hammer_unlock(&ip->lock); 375 376 /* 377 * loop if the vget fails (aka races), or if the vp 378 * no longer matches ip->vp. 379 */ 380 if (vget(vp, LK_EXCLUSIVE) == 0) { 381 if (vp == ip->vp) { 382 vdrop(vp); 383 break; 384 } 385 vput(vp); 386 } 387 vdrop(vp); 388 } 389 *vpp = vp; 390 return(error); 391 } 392 393 /* 394 * Locate all copies of the inode for obj_id compatible with the specified 395 * asof, reference, and issue the related call-back. This routine is used 396 * for direct-io invalidation and does not create any new inodes. 397 */ 398 void 399 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo, 400 int (*callback)(hammer_inode_t ip, void *data), 401 void *data) 402 { 403 hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root, 404 hammer_inode_info_cmp_all_history, 405 callback, iinfo); 406 } 407 408 /* 409 * Acquire a HAMMER inode. The returned inode is not locked. These functions 410 * do not attach or detach the related vnode (use hammer_get_vnode() for 411 * that). 412 * 413 * The flags argument is only applied for newly created inodes, and only 414 * certain flags are inherited. 415 * 416 * Called from the frontend. 417 */ 418 struct hammer_inode * 419 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip, 420 int64_t obj_id, hammer_tid_t asof, u_int32_t localization, 421 int flags, int *errorp) 422 { 423 hammer_mount_t hmp = trans->hmp; 424 struct hammer_node_cache *cachep; 425 struct hammer_inode_info iinfo; 426 struct hammer_cursor cursor; 427 struct hammer_inode *ip; 428 429 430 /* 431 * Determine if we already have an inode cached. If we do then 432 * we are golden. 433 * 434 * If we find an inode with no vnode we have to mark the 435 * transaction such that hammer_inode_waitreclaims() is 436 * called later on to avoid building up an infinite number 437 * of inodes. Otherwise we can continue to * add new inodes 438 * faster then they can be disposed of, even with the tsleep 439 * delay. 440 * 441 * If we find a dummy inode we return a failure so dounlink 442 * (which does another lookup) doesn't try to mess with the 443 * link count. hammer_vop_nresolve() uses hammer_get_dummy_inode() 444 * to ref dummy inodes. 445 */ 446 iinfo.obj_id = obj_id; 447 iinfo.obj_asof = asof; 448 iinfo.obj_localization = localization; 449 loop: 450 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo); 451 if (ip) { 452 if (ip->flags & HAMMER_INODE_DUMMY) { 453 *errorp = ENOENT; 454 return(NULL); 455 } 456 hammer_ref(&ip->lock); 457 *errorp = 0; 458 return(ip); 459 } 460 461 /* 462 * Allocate a new inode structure and deal with races later. 463 */ 464 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO); 465 ++hammer_count_inodes; 466 ++hmp->count_inodes; 467 ip->obj_id = obj_id; 468 ip->obj_asof = iinfo.obj_asof; 469 ip->obj_localization = localization; 470 ip->hmp = hmp; 471 ip->flags = flags & HAMMER_INODE_RO; 472 ip->cache[0].ip = ip; 473 ip->cache[1].ip = ip; 474 ip->cache[2].ip = ip; 475 ip->cache[3].ip = ip; 476 if (hmp->ronly) 477 ip->flags |= HAMMER_INODE_RO; 478 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off = 479 0x7FFFFFFFFFFFFFFFLL; 480 RB_INIT(&ip->rec_tree); 481 TAILQ_INIT(&ip->target_list); 482 hammer_ref(&ip->lock); 483 484 /* 485 * Locate the on-disk inode. If this is a PFS root we always 486 * access the current version of the root inode and (if it is not 487 * a master) always access information under it with a snapshot 488 * TID. 489 * 490 * We cache recent inode lookups in this directory in dip->cache[2]. 491 * If we can't find it we assume the inode we are looking for is 492 * close to the directory inode. 493 */ 494 retry: 495 cachep = NULL; 496 if (dip) { 497 if (dip->cache[2].node) 498 cachep = &dip->cache[2]; 499 else 500 cachep = &dip->cache[0]; 501 } 502 hammer_init_cursor(trans, &cursor, cachep, NULL); 503 cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE; 504 cursor.key_beg.obj_id = ip->obj_id; 505 cursor.key_beg.key = 0; 506 cursor.key_beg.create_tid = 0; 507 cursor.key_beg.delete_tid = 0; 508 cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE; 509 cursor.key_beg.obj_type = 0; 510 511 cursor.asof = iinfo.obj_asof; 512 cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA | 513 HAMMER_CURSOR_ASOF; 514 515 *errorp = hammer_btree_lookup(&cursor); 516 if (*errorp == EDEADLK) { 517 hammer_done_cursor(&cursor); 518 goto retry; 519 } 520 521 /* 522 * On success the B-Tree lookup will hold the appropriate 523 * buffer cache buffers and provide a pointer to the requested 524 * information. Copy the information to the in-memory inode 525 * and cache the B-Tree node to improve future operations. 526 */ 527 if (*errorp == 0) { 528 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf; 529 ip->ino_data = cursor.data->inode; 530 531 /* 532 * cache[0] tries to cache the location of the object inode. 533 * The assumption is that it is near the directory inode. 534 * 535 * cache[1] tries to cache the location of the object data. 536 * We might have something in the governing directory from 537 * scan optimizations (see the strategy code in 538 * hammer_vnops.c). 539 * 540 * We update dip->cache[2], if possible, with the location 541 * of the object inode for future directory shortcuts. 542 */ 543 hammer_cache_node(&ip->cache[0], cursor.node); 544 if (dip) { 545 if (dip->cache[3].node) { 546 hammer_cache_node(&ip->cache[1], 547 dip->cache[3].node); 548 } 549 hammer_cache_node(&dip->cache[2], cursor.node); 550 } 551 552 /* 553 * The file should not contain any data past the file size 554 * stored in the inode. Setting save_trunc_off to the 555 * file size instead of max reduces B-Tree lookup overheads 556 * on append by allowing the flusher to avoid checking for 557 * record overwrites. 558 */ 559 ip->save_trunc_off = ip->ino_data.size; 560 561 /* 562 * Locate and assign the pseudofs management structure to 563 * the inode. 564 */ 565 if (dip && dip->obj_localization == ip->obj_localization) { 566 ip->pfsm = dip->pfsm; 567 hammer_ref(&ip->pfsm->lock); 568 } else { 569 ip->pfsm = hammer_load_pseudofs(trans, 570 ip->obj_localization, 571 errorp); 572 *errorp = 0; /* ignore ENOENT */ 573 } 574 } 575 576 /* 577 * The inode is placed on the red-black tree and will be synced to 578 * the media when flushed or by the filesystem sync. If this races 579 * another instantiation/lookup the insertion will fail. 580 */ 581 if (*errorp == 0) { 582 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) { 583 hammer_free_inode(ip); 584 hammer_done_cursor(&cursor); 585 goto loop; 586 } 587 ip->flags |= HAMMER_INODE_ONDISK; 588 } else { 589 if (ip->flags & HAMMER_INODE_RSV_INODES) { 590 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */ 591 --hmp->rsv_inodes; 592 } 593 594 hammer_free_inode(ip); 595 ip = NULL; 596 } 597 hammer_done_cursor(&cursor); 598 599 /* 600 * NEWINODE is only set if the inode becomes dirty later, 601 * setting it here just leads to unnecessary stalls. 602 * 603 * trans->flags |= HAMMER_TRANSF_NEWINODE; 604 */ 605 return (ip); 606 } 607 608 /* 609 * Get a dummy inode to placemark a broken directory entry. 610 */ 611 struct hammer_inode * 612 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip, 613 int64_t obj_id, hammer_tid_t asof, u_int32_t localization, 614 int flags, int *errorp) 615 { 616 hammer_mount_t hmp = trans->hmp; 617 struct hammer_inode_info iinfo; 618 struct hammer_inode *ip; 619 620 /* 621 * Determine if we already have an inode cached. If we do then 622 * we are golden. 623 * 624 * If we find an inode with no vnode we have to mark the 625 * transaction such that hammer_inode_waitreclaims() is 626 * called later on to avoid building up an infinite number 627 * of inodes. Otherwise we can continue to * add new inodes 628 * faster then they can be disposed of, even with the tsleep 629 * delay. 630 * 631 * If we find a non-fake inode we return an error. Only fake 632 * inodes can be returned by this routine. 633 */ 634 iinfo.obj_id = obj_id; 635 iinfo.obj_asof = asof; 636 iinfo.obj_localization = localization; 637 loop: 638 *errorp = 0; 639 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo); 640 if (ip) { 641 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) { 642 *errorp = ENOENT; 643 return(NULL); 644 } 645 hammer_ref(&ip->lock); 646 return(ip); 647 } 648 649 /* 650 * Allocate a new inode structure and deal with races later. 651 */ 652 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO); 653 ++hammer_count_inodes; 654 ++hmp->count_inodes; 655 ip->obj_id = obj_id; 656 ip->obj_asof = iinfo.obj_asof; 657 ip->obj_localization = localization; 658 ip->hmp = hmp; 659 ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY; 660 ip->cache[0].ip = ip; 661 ip->cache[1].ip = ip; 662 ip->cache[2].ip = ip; 663 ip->cache[3].ip = ip; 664 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off = 665 0x7FFFFFFFFFFFFFFFLL; 666 RB_INIT(&ip->rec_tree); 667 TAILQ_INIT(&ip->target_list); 668 hammer_ref(&ip->lock); 669 670 /* 671 * Populate the dummy inode. Leave everything zero'd out. 672 * 673 * (ip->ino_leaf and ip->ino_data) 674 * 675 * Make the dummy inode a FIFO object which most copy programs 676 * will properly ignore. 677 */ 678 ip->save_trunc_off = ip->ino_data.size; 679 ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO; 680 681 /* 682 * Locate and assign the pseudofs management structure to 683 * the inode. 684 */ 685 if (dip && dip->obj_localization == ip->obj_localization) { 686 ip->pfsm = dip->pfsm; 687 hammer_ref(&ip->pfsm->lock); 688 } else { 689 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization, 690 errorp); 691 *errorp = 0; /* ignore ENOENT */ 692 } 693 694 /* 695 * The inode is placed on the red-black tree and will be synced to 696 * the media when flushed or by the filesystem sync. If this races 697 * another instantiation/lookup the insertion will fail. 698 * 699 * NOTE: Do not set HAMMER_INODE_ONDISK. The inode is a fake. 700 */ 701 if (*errorp == 0) { 702 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) { 703 hammer_free_inode(ip); 704 goto loop; 705 } 706 } else { 707 if (ip->flags & HAMMER_INODE_RSV_INODES) { 708 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */ 709 --hmp->rsv_inodes; 710 } 711 hammer_free_inode(ip); 712 ip = NULL; 713 } 714 trans->flags |= HAMMER_TRANSF_NEWINODE; 715 return (ip); 716 } 717 718 /* 719 * Return a referenced inode only if it is in our inode cache. 720 * 721 * Dummy inodes do not count. 722 */ 723 struct hammer_inode * 724 hammer_find_inode(hammer_transaction_t trans, int64_t obj_id, 725 hammer_tid_t asof, u_int32_t localization) 726 { 727 hammer_mount_t hmp = trans->hmp; 728 struct hammer_inode_info iinfo; 729 struct hammer_inode *ip; 730 731 iinfo.obj_id = obj_id; 732 iinfo.obj_asof = asof; 733 iinfo.obj_localization = localization; 734 735 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo); 736 if (ip) { 737 if (ip->flags & HAMMER_INODE_DUMMY) 738 ip = NULL; 739 else 740 hammer_ref(&ip->lock); 741 } 742 return(ip); 743 } 744 745 /* 746 * Create a new filesystem object, returning the inode in *ipp. The 747 * returned inode will be referenced. The inode is created in-memory. 748 * 749 * If pfsm is non-NULL the caller wishes to create the root inode for 750 * a master PFS. 751 */ 752 int 753 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap, 754 struct ucred *cred, 755 hammer_inode_t dip, const char *name, int namelen, 756 hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp) 757 { 758 hammer_mount_t hmp; 759 hammer_inode_t ip; 760 uid_t xuid; 761 int error; 762 int64_t namekey; 763 u_int32_t dummy; 764 765 hmp = trans->hmp; 766 767 /* 768 * Disallow the creation of new inodes in directories which 769 * have been deleted. In HAMMER, this will cause a record 770 * syncing assertion later on in the flush code. 771 */ 772 if (dip && dip->ino_data.nlinks == 0) { 773 *ipp = NULL; 774 return (EINVAL); 775 } 776 777 /* 778 * Allocate inode 779 */ 780 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO); 781 ++hammer_count_inodes; 782 ++hmp->count_inodes; 783 trans->flags |= HAMMER_TRANSF_NEWINODE; 784 785 if (pfsm) { 786 KKASSERT(pfsm->localization != 0); 787 ip->obj_id = HAMMER_OBJID_ROOT; 788 ip->obj_localization = pfsm->localization; 789 } else { 790 KKASSERT(dip != NULL); 791 namekey = hammer_directory_namekey(dip, name, namelen, &dummy); 792 ip->obj_id = hammer_alloc_objid(hmp, dip, namekey); 793 ip->obj_localization = dip->obj_localization; 794 } 795 796 KKASSERT(ip->obj_id != 0); 797 ip->obj_asof = hmp->asof; 798 ip->hmp = hmp; 799 ip->flush_state = HAMMER_FST_IDLE; 800 ip->flags = HAMMER_INODE_DDIRTY | 801 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME; 802 ip->cache[0].ip = ip; 803 ip->cache[1].ip = ip; 804 ip->cache[2].ip = ip; 805 ip->cache[3].ip = ip; 806 807 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL; 808 /* ip->save_trunc_off = 0; (already zero) */ 809 RB_INIT(&ip->rec_tree); 810 TAILQ_INIT(&ip->target_list); 811 812 ip->ino_data.atime = trans->time; 813 ip->ino_data.mtime = trans->time; 814 ip->ino_data.size = 0; 815 ip->ino_data.nlinks = 0; 816 817 /* 818 * A nohistory designator on the parent directory is inherited by 819 * the child. We will do this even for pseudo-fs creation... the 820 * sysad can turn it off. 821 */ 822 if (dip) { 823 ip->ino_data.uflags = dip->ino_data.uflags & 824 (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP); 825 } 826 827 ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD; 828 ip->ino_leaf.base.localization = ip->obj_localization + 829 HAMMER_LOCALIZE_INODE; 830 ip->ino_leaf.base.obj_id = ip->obj_id; 831 ip->ino_leaf.base.key = 0; 832 ip->ino_leaf.base.create_tid = 0; 833 ip->ino_leaf.base.delete_tid = 0; 834 ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE; 835 ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type); 836 837 ip->ino_data.obj_type = ip->ino_leaf.base.obj_type; 838 ip->ino_data.version = HAMMER_INODE_DATA_VERSION; 839 ip->ino_data.mode = vap->va_mode; 840 ip->ino_data.ctime = trans->time; 841 842 /* 843 * If we are running version 2 or greater directory entries are 844 * inode-localized instead of data-localized. 845 */ 846 if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) { 847 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) { 848 ip->ino_data.cap_flags |= 849 HAMMER_INODE_CAP_DIR_LOCAL_INO; 850 } 851 } 852 if (trans->hmp->version >= HAMMER_VOL_VERSION_SIX) { 853 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) { 854 ip->ino_data.cap_flags |= 855 HAMMER_INODE_CAP_DIRHASH_ALG1; 856 } 857 } 858 859 /* 860 * Setup the ".." pointer. This only needs to be done for directories 861 * but we do it for all objects as a recovery aid if dip exists. 862 * The inode is probably a PFS root if dip is NULL. 863 */ 864 if (dip) 865 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id; 866 #if 0 867 /* 868 * The parent_obj_localization field only applies to pseudo-fs roots. 869 * XXX this is no longer applicable, PFSs are no longer directly 870 * tied into the parent's directory structure. 871 */ 872 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY && 873 ip->obj_id == HAMMER_OBJID_ROOT) { 874 ip->ino_data.ext.obj.parent_obj_localization = 875 dip->obj_localization; 876 } 877 #endif 878 879 switch(ip->ino_leaf.base.obj_type) { 880 case HAMMER_OBJTYPE_CDEV: 881 case HAMMER_OBJTYPE_BDEV: 882 ip->ino_data.rmajor = vap->va_rmajor; 883 ip->ino_data.rminor = vap->va_rminor; 884 break; 885 default: 886 break; 887 } 888 889 /* 890 * Calculate default uid/gid and overwrite with information from 891 * the vap. 892 */ 893 if (dip) { 894 xuid = hammer_to_unix_xid(&dip->ino_data.uid); 895 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode, 896 xuid, cred, &vap->va_mode); 897 } else { 898 xuid = 0; 899 } 900 ip->ino_data.mode = vap->va_mode; 901 902 if (vap->va_vaflags & VA_UID_UUID_VALID) 903 ip->ino_data.uid = vap->va_uid_uuid; 904 else if (vap->va_uid != (uid_t)VNOVAL) 905 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid); 906 else 907 hammer_guid_to_uuid(&ip->ino_data.uid, xuid); 908 909 if (vap->va_vaflags & VA_GID_UUID_VALID) 910 ip->ino_data.gid = vap->va_gid_uuid; 911 else if (vap->va_gid != (gid_t)VNOVAL) 912 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid); 913 else if (dip) 914 ip->ino_data.gid = dip->ino_data.gid; 915 916 hammer_ref(&ip->lock); 917 918 if (pfsm) { 919 ip->pfsm = pfsm; 920 hammer_ref(&pfsm->lock); 921 error = 0; 922 } else if (dip->obj_localization == ip->obj_localization) { 923 ip->pfsm = dip->pfsm; 924 hammer_ref(&ip->pfsm->lock); 925 error = 0; 926 } else { 927 ip->pfsm = hammer_load_pseudofs(trans, 928 ip->obj_localization, 929 &error); 930 error = 0; /* ignore ENOENT */ 931 } 932 933 if (error) { 934 hammer_free_inode(ip); 935 ip = NULL; 936 } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) { 937 hpanic("duplicate obj_id %llx", (long long)ip->obj_id); 938 /* not reached */ 939 hammer_free_inode(ip); 940 } 941 *ipp = ip; 942 return(error); 943 } 944 945 /* 946 * Final cleanup / freeing of an inode structure 947 */ 948 static void 949 hammer_free_inode(hammer_inode_t ip) 950 { 951 struct hammer_mount *hmp; 952 953 hmp = ip->hmp; 954 KKASSERT(hammer_oneref(&ip->lock)); 955 hammer_uncache_node(&ip->cache[0]); 956 hammer_uncache_node(&ip->cache[1]); 957 hammer_uncache_node(&ip->cache[2]); 958 hammer_uncache_node(&ip->cache[3]); 959 hammer_inode_wakereclaims(ip); 960 if (ip->objid_cache) 961 hammer_clear_objid(ip); 962 --hammer_count_inodes; 963 --hmp->count_inodes; 964 if (ip->pfsm) { 965 hammer_rel_pseudofs(hmp, ip->pfsm); 966 ip->pfsm = NULL; 967 } 968 kfree(ip, hmp->m_inodes); 969 } 970 971 /* 972 * Retrieve pseudo-fs data. NULL will never be returned. 973 * 974 * If an error occurs *errorp will be set and a default template is returned, 975 * otherwise *errorp is set to 0. Typically when an error occurs it will 976 * be ENOENT. 977 */ 978 hammer_pseudofs_inmem_t 979 hammer_load_pseudofs(hammer_transaction_t trans, 980 u_int32_t localization, int *errorp) 981 { 982 hammer_mount_t hmp = trans->hmp; 983 hammer_inode_t ip; 984 hammer_pseudofs_inmem_t pfsm; 985 struct hammer_cursor cursor; 986 int bytes; 987 988 retry: 989 pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization); 990 if (pfsm) { 991 hammer_ref(&pfsm->lock); 992 *errorp = 0; 993 return(pfsm); 994 } 995 996 /* 997 * PFS records are associated with the root inode (not the PFS root 998 * inode, but the real root). Avoid an infinite recursion if loading 999 * the PFS for the real root. 1000 */ 1001 if (localization) { 1002 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, 1003 HAMMER_MAX_TID, 1004 HAMMER_DEF_LOCALIZATION, 0, errorp); 1005 } else { 1006 ip = NULL; 1007 } 1008 1009 pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO); 1010 pfsm->localization = localization; 1011 pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid; 1012 pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid; 1013 1014 hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip); 1015 cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION + 1016 HAMMER_LOCALIZE_MISC; 1017 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT; 1018 cursor.key_beg.create_tid = 0; 1019 cursor.key_beg.delete_tid = 0; 1020 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS; 1021 cursor.key_beg.obj_type = 0; 1022 cursor.key_beg.key = localization; 1023 cursor.asof = HAMMER_MAX_TID; 1024 cursor.flags |= HAMMER_CURSOR_ASOF; 1025 1026 if (ip) 1027 *errorp = hammer_ip_lookup(&cursor); 1028 else 1029 *errorp = hammer_btree_lookup(&cursor); 1030 if (*errorp == 0) { 1031 *errorp = hammer_ip_resolve_data(&cursor); 1032 if (*errorp == 0) { 1033 if (cursor.data->pfsd.mirror_flags & 1034 HAMMER_PFSD_DELETED) { 1035 *errorp = ENOENT; 1036 } else { 1037 bytes = cursor.leaf->data_len; 1038 if (bytes > sizeof(pfsm->pfsd)) 1039 bytes = sizeof(pfsm->pfsd); 1040 bcopy(cursor.data, &pfsm->pfsd, bytes); 1041 } 1042 } 1043 } 1044 hammer_done_cursor(&cursor); 1045 1046 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid); 1047 hammer_ref(&pfsm->lock); 1048 if (ip) 1049 hammer_rel_inode(ip, 0); 1050 if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) { 1051 kfree(pfsm, hmp->m_misc); 1052 goto retry; 1053 } 1054 return(pfsm); 1055 } 1056 1057 /* 1058 * Store pseudo-fs data. The backend will automatically delete any prior 1059 * on-disk pseudo-fs data but we have to delete in-memory versions. 1060 */ 1061 int 1062 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm) 1063 { 1064 struct hammer_cursor cursor; 1065 hammer_record_t record; 1066 hammer_inode_t ip; 1067 int error; 1068 1069 /* 1070 * PFS records are associated with the root inode (not the PFS root 1071 * inode, but the real root). 1072 */ 1073 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID, 1074 HAMMER_DEF_LOCALIZATION, 0, &error); 1075 retry: 1076 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid); 1077 hammer_init_cursor(trans, &cursor, &ip->cache[1], ip); 1078 cursor.key_beg.localization = ip->obj_localization + 1079 HAMMER_LOCALIZE_MISC; 1080 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT; 1081 cursor.key_beg.create_tid = 0; 1082 cursor.key_beg.delete_tid = 0; 1083 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS; 1084 cursor.key_beg.obj_type = 0; 1085 cursor.key_beg.key = pfsm->localization; 1086 cursor.asof = HAMMER_MAX_TID; 1087 cursor.flags |= HAMMER_CURSOR_ASOF; 1088 1089 /* 1090 * Replace any in-memory version of the record. 1091 */ 1092 error = hammer_ip_lookup(&cursor); 1093 if (error == 0 && hammer_cursor_inmem(&cursor)) { 1094 record = cursor.iprec; 1095 if (record->flags & HAMMER_RECF_INTERLOCK_BE) { 1096 KKASSERT(cursor.deadlk_rec == NULL); 1097 hammer_ref(&record->lock); 1098 cursor.deadlk_rec = record; 1099 error = EDEADLK; 1100 } else { 1101 record->flags |= HAMMER_RECF_DELETED_FE; 1102 error = 0; 1103 } 1104 } 1105 1106 /* 1107 * Allocate replacement general record. The backend flush will 1108 * delete any on-disk version of the record. 1109 */ 1110 if (error == 0 || error == ENOENT) { 1111 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd)); 1112 record->type = HAMMER_MEM_RECORD_GENERAL; 1113 1114 record->leaf.base.localization = ip->obj_localization + 1115 HAMMER_LOCALIZE_MISC; 1116 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS; 1117 record->leaf.base.key = pfsm->localization; 1118 record->leaf.data_len = sizeof(pfsm->pfsd); 1119 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd)); 1120 error = hammer_ip_add_record(trans, record); 1121 } 1122 hammer_done_cursor(&cursor); 1123 if (error == EDEADLK) 1124 goto retry; 1125 hammer_rel_inode(ip, 0); 1126 return(error); 1127 } 1128 1129 /* 1130 * Create a root directory for a PFS if one does not alredy exist. 1131 * 1132 * The PFS root stands alone so we must also bump the nlinks count 1133 * to prevent it from being destroyed on release. 1134 */ 1135 int 1136 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred, 1137 hammer_pseudofs_inmem_t pfsm) 1138 { 1139 hammer_inode_t ip; 1140 struct vattr vap; 1141 int error; 1142 1143 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID, 1144 pfsm->localization, 0, &error); 1145 if (ip == NULL) { 1146 vattr_null(&vap); 1147 vap.va_mode = 0755; 1148 vap.va_type = VDIR; 1149 error = hammer_create_inode(trans, &vap, cred, 1150 NULL, NULL, 0, 1151 pfsm, &ip); 1152 if (error == 0) { 1153 ++ip->ino_data.nlinks; 1154 hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY); 1155 } 1156 } 1157 if (ip) 1158 hammer_rel_inode(ip, 0); 1159 return(error); 1160 } 1161 1162 /* 1163 * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY 1164 * if we are unable to disassociate all the inodes. 1165 */ 1166 static 1167 int 1168 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data) 1169 { 1170 int res; 1171 1172 hammer_ref(&ip->lock); 1173 if (ip->vp && (ip->vp->v_flag & VPFSROOT)) { 1174 /* 1175 * The hammer pfs-upgrade directive itself might have the 1176 * root of the pfs open. Just allow it. 1177 */ 1178 res = 0; 1179 } else { 1180 /* 1181 * Don't allow any subdirectories or files to be open. 1182 */ 1183 if (hammer_isactive(&ip->lock) == 2 && ip->vp) 1184 vclean_unlocked(ip->vp); 1185 if (hammer_isactive(&ip->lock) == 1 && ip->vp == NULL) 1186 res = 0; 1187 else 1188 res = -1; /* stop, someone is using the inode */ 1189 } 1190 hammer_rel_inode(ip, 0); 1191 return(res); 1192 } 1193 1194 int 1195 hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization) 1196 { 1197 int res; 1198 int try; 1199 1200 for (try = res = 0; try < 4; ++try) { 1201 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root, 1202 hammer_inode_pfs_cmp, 1203 hammer_unload_pseudofs_callback, 1204 &localization); 1205 if (res == 0 && try > 1) 1206 break; 1207 hammer_flusher_sync(trans->hmp); 1208 } 1209 if (res != 0) 1210 res = ENOTEMPTY; 1211 return(res); 1212 } 1213 1214 1215 /* 1216 * Release a reference on a PFS 1217 */ 1218 void 1219 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm) 1220 { 1221 hammer_rel(&pfsm->lock); 1222 if (hammer_norefs(&pfsm->lock)) { 1223 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm); 1224 kfree(pfsm, hmp->m_misc); 1225 } 1226 } 1227 1228 /* 1229 * Called by hammer_sync_inode(). 1230 */ 1231 static int 1232 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip) 1233 { 1234 hammer_transaction_t trans = cursor->trans; 1235 hammer_record_t record; 1236 int error; 1237 int redirty; 1238 1239 retry: 1240 error = 0; 1241 1242 /* 1243 * If the inode has a presence on-disk then locate it and mark 1244 * it deleted, setting DELONDISK. 1245 * 1246 * The record may or may not be physically deleted, depending on 1247 * the retention policy. 1248 */ 1249 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) == 1250 HAMMER_INODE_ONDISK) { 1251 hammer_normalize_cursor(cursor); 1252 cursor->key_beg.localization = ip->obj_localization + 1253 HAMMER_LOCALIZE_INODE; 1254 cursor->key_beg.obj_id = ip->obj_id; 1255 cursor->key_beg.key = 0; 1256 cursor->key_beg.create_tid = 0; 1257 cursor->key_beg.delete_tid = 0; 1258 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE; 1259 cursor->key_beg.obj_type = 0; 1260 cursor->asof = ip->obj_asof; 1261 cursor->flags &= ~HAMMER_CURSOR_INITMASK; 1262 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF; 1263 cursor->flags |= HAMMER_CURSOR_BACKEND; 1264 1265 error = hammer_btree_lookup(cursor); 1266 if (hammer_debug_inode) 1267 hdkprintf("IPDEL %p %08x %d", ip, ip->flags, error); 1268 1269 if (error == 0) { 1270 error = hammer_ip_delete_record(cursor, ip, trans->tid); 1271 if (hammer_debug_inode) 1272 hdkprintf("error %d\n", error); 1273 if (error == 0) { 1274 ip->flags |= HAMMER_INODE_DELONDISK; 1275 } 1276 if (cursor->node) 1277 hammer_cache_node(&ip->cache[0], cursor->node); 1278 } 1279 if (error == EDEADLK) { 1280 hammer_done_cursor(cursor); 1281 error = hammer_init_cursor(trans, cursor, 1282 &ip->cache[0], ip); 1283 if (hammer_debug_inode) 1284 hdkprintf("IPDED %p %d\n", ip, error); 1285 if (error == 0) 1286 goto retry; 1287 } 1288 } 1289 1290 /* 1291 * Ok, write out the initial record or a new record (after deleting 1292 * the old one), unless the DELETED flag is set. This routine will 1293 * clear DELONDISK if it writes out a record. 1294 * 1295 * Update our inode statistics if this is the first application of 1296 * the inode on-disk. 1297 */ 1298 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) { 1299 /* 1300 * Generate a record and write it to the media. We clean-up 1301 * the state before releasing so we do not have to set-up 1302 * a flush_group. 1303 */ 1304 record = hammer_alloc_mem_record(ip, 0); 1305 record->type = HAMMER_MEM_RECORD_INODE; 1306 record->flush_state = HAMMER_FST_FLUSH; 1307 record->leaf = ip->sync_ino_leaf; 1308 record->leaf.base.create_tid = trans->tid; 1309 record->leaf.data_len = sizeof(ip->sync_ino_data); 1310 record->leaf.create_ts = trans->time32; 1311 record->data = (void *)&ip->sync_ino_data; 1312 record->flags |= HAMMER_RECF_INTERLOCK_BE; 1313 1314 /* 1315 * If this flag is set we cannot sync the new file size 1316 * because we haven't finished related truncations. The 1317 * inode will be flushed in another flush group to finish 1318 * the job. 1319 */ 1320 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) && 1321 ip->sync_ino_data.size != ip->ino_data.size) { 1322 redirty = 1; 1323 ip->sync_ino_data.size = ip->ino_data.size; 1324 } else { 1325 redirty = 0; 1326 } 1327 1328 for (;;) { 1329 error = hammer_ip_sync_record_cursor(cursor, record); 1330 if (hammer_debug_inode) 1331 hdkprintf("GENREC %p rec %08x %d\n", 1332 ip, record->flags, error); 1333 if (error != EDEADLK) 1334 break; 1335 hammer_done_cursor(cursor); 1336 error = hammer_init_cursor(trans, cursor, 1337 &ip->cache[0], ip); 1338 if (hammer_debug_inode) 1339 hdkprintf("GENREC reinit %d\n", error); 1340 if (error) 1341 break; 1342 } 1343 1344 /* 1345 * Note: The record was never on the inode's record tree 1346 * so just wave our hands importantly and destroy it. 1347 */ 1348 record->flags |= HAMMER_RECF_COMMITTED; 1349 record->flags &= ~HAMMER_RECF_INTERLOCK_BE; 1350 record->flush_state = HAMMER_FST_IDLE; 1351 ++ip->rec_generation; 1352 hammer_rel_mem_record(record); 1353 1354 /* 1355 * Finish up. 1356 */ 1357 if (error == 0) { 1358 if (hammer_debug_inode) 1359 hdkprintf("CLEANDELOND %p %08x\n", ip, ip->flags); 1360 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | 1361 HAMMER_INODE_SDIRTY | 1362 HAMMER_INODE_ATIME | 1363 HAMMER_INODE_MTIME); 1364 ip->flags &= ~HAMMER_INODE_DELONDISK; 1365 if (redirty) 1366 ip->sync_flags |= HAMMER_INODE_DDIRTY; 1367 1368 /* 1369 * Root volume count of inodes 1370 */ 1371 hammer_sync_lock_sh(trans); 1372 if ((ip->flags & HAMMER_INODE_ONDISK) == 0) { 1373 hammer_modify_volume_field(trans, 1374 trans->rootvol, 1375 vol0_stat_inodes); 1376 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes; 1377 hammer_modify_volume_done(trans->rootvol); 1378 ip->flags |= HAMMER_INODE_ONDISK; 1379 if (hammer_debug_inode) 1380 hdkprintf("NOWONDISK %p\n", ip); 1381 } 1382 hammer_sync_unlock(trans); 1383 } 1384 } 1385 1386 /* 1387 * If the inode has been destroyed, clean out any left-over flags 1388 * that may have been set by the frontend. 1389 */ 1390 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) { 1391 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | 1392 HAMMER_INODE_SDIRTY | 1393 HAMMER_INODE_ATIME | 1394 HAMMER_INODE_MTIME); 1395 } 1396 return(error); 1397 } 1398 1399 /* 1400 * Update only the itimes fields. 1401 * 1402 * ATIME can be updated without generating any UNDO. MTIME is updated 1403 * with UNDO so it is guaranteed to be synchronized properly in case of 1404 * a crash. 1405 * 1406 * Neither field is included in the B-Tree leaf element's CRC, which is how 1407 * we can get away with updating ATIME the way we do. 1408 */ 1409 static int 1410 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip) 1411 { 1412 hammer_transaction_t trans = cursor->trans; 1413 int error; 1414 1415 retry: 1416 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) != 1417 HAMMER_INODE_ONDISK) { 1418 return(0); 1419 } 1420 1421 hammer_normalize_cursor(cursor); 1422 cursor->key_beg.localization = ip->obj_localization + 1423 HAMMER_LOCALIZE_INODE; 1424 cursor->key_beg.obj_id = ip->obj_id; 1425 cursor->key_beg.key = 0; 1426 cursor->key_beg.create_tid = 0; 1427 cursor->key_beg.delete_tid = 0; 1428 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE; 1429 cursor->key_beg.obj_type = 0; 1430 cursor->asof = ip->obj_asof; 1431 cursor->flags &= ~HAMMER_CURSOR_INITMASK; 1432 cursor->flags |= HAMMER_CURSOR_ASOF; 1433 cursor->flags |= HAMMER_CURSOR_GET_LEAF; 1434 cursor->flags |= HAMMER_CURSOR_GET_DATA; 1435 cursor->flags |= HAMMER_CURSOR_BACKEND; 1436 1437 error = hammer_btree_lookup(cursor); 1438 if (error == 0) { 1439 hammer_cache_node(&ip->cache[0], cursor->node); 1440 if (ip->sync_flags & HAMMER_INODE_MTIME) { 1441 /* 1442 * Updating MTIME requires an UNDO. Just cover 1443 * both atime and mtime. 1444 */ 1445 hammer_sync_lock_sh(trans); 1446 hammer_modify_buffer(trans, cursor->data_buffer, 1447 HAMMER_ITIMES_BASE(&cursor->data->inode), 1448 HAMMER_ITIMES_BYTES); 1449 cursor->data->inode.atime = ip->sync_ino_data.atime; 1450 cursor->data->inode.mtime = ip->sync_ino_data.mtime; 1451 hammer_modify_buffer_done(cursor->data_buffer); 1452 hammer_sync_unlock(trans); 1453 } else if (ip->sync_flags & HAMMER_INODE_ATIME) { 1454 /* 1455 * Updating atime only can be done in-place with 1456 * no UNDO. 1457 */ 1458 hammer_sync_lock_sh(trans); 1459 hammer_modify_buffer(trans, cursor->data_buffer, 1460 NULL, 0); 1461 cursor->data->inode.atime = ip->sync_ino_data.atime; 1462 hammer_modify_buffer_done(cursor->data_buffer); 1463 hammer_sync_unlock(trans); 1464 } 1465 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME); 1466 } 1467 if (error == EDEADLK) { 1468 hammer_done_cursor(cursor); 1469 error = hammer_init_cursor(trans, cursor, &ip->cache[0], ip); 1470 if (error == 0) 1471 goto retry; 1472 } 1473 return(error); 1474 } 1475 1476 /* 1477 * Release a reference on an inode, flush as requested. 1478 * 1479 * On the last reference we queue the inode to the flusher for its final 1480 * disposition. 1481 */ 1482 void 1483 hammer_rel_inode(struct hammer_inode *ip, int flush) 1484 { 1485 /* 1486 * Handle disposition when dropping the last ref. 1487 */ 1488 for (;;) { 1489 if (hammer_oneref(&ip->lock)) { 1490 /* 1491 * Determine whether on-disk action is needed for 1492 * the inode's final disposition. 1493 */ 1494 KKASSERT(ip->vp == NULL); 1495 hammer_inode_unloadable_check(ip, 0); 1496 if (ip->flags & HAMMER_INODE_MODMASK) { 1497 hammer_flush_inode(ip, 0); 1498 } else if (hammer_oneref(&ip->lock)) { 1499 hammer_unload_inode(ip); 1500 break; 1501 } 1502 } else { 1503 if (flush) 1504 hammer_flush_inode(ip, 0); 1505 1506 /* 1507 * The inode still has multiple refs, try to drop 1508 * one ref. 1509 */ 1510 KKASSERT(hammer_isactive(&ip->lock) >= 1); 1511 if (hammer_isactive(&ip->lock) > 1) { 1512 hammer_rel(&ip->lock); 1513 break; 1514 } 1515 } 1516 } 1517 } 1518 1519 /* 1520 * Unload and destroy the specified inode. Must be called with one remaining 1521 * reference. The reference is disposed of. 1522 * 1523 * The inode must be completely clean. 1524 */ 1525 static int 1526 hammer_unload_inode(struct hammer_inode *ip) 1527 { 1528 hammer_mount_t hmp = ip->hmp; 1529 1530 KASSERT(hammer_oneref(&ip->lock), 1531 ("hammer_unload_inode: %d refs", hammer_isactive(&ip->lock))); 1532 KKASSERT(ip->vp == NULL); 1533 KKASSERT(ip->flush_state == HAMMER_FST_IDLE); 1534 KKASSERT(ip->cursor_ip_refs == 0); 1535 KKASSERT(hammer_notlocked(&ip->lock)); 1536 KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0); 1537 1538 KKASSERT(RB_EMPTY(&ip->rec_tree)); 1539 KKASSERT(TAILQ_EMPTY(&ip->target_list)); 1540 1541 if (ip->flags & HAMMER_INODE_RDIRTY) { 1542 RB_REMOVE(hammer_redo_rb_tree, &hmp->rb_redo_root, ip); 1543 ip->flags &= ~HAMMER_INODE_RDIRTY; 1544 } 1545 RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip); 1546 1547 hammer_free_inode(ip); 1548 return(0); 1549 } 1550 1551 /* 1552 * Called during unmounting if a critical error occured. The in-memory 1553 * inode and all related structures are destroyed. 1554 * 1555 * If a critical error did not occur the unmount code calls the standard 1556 * release and asserts that the inode is gone. 1557 */ 1558 int 1559 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused) 1560 { 1561 hammer_record_t rec; 1562 1563 /* 1564 * Get rid of the inodes in-memory records, regardless of their 1565 * state, and clear the mod-mask. 1566 */ 1567 while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) { 1568 TAILQ_REMOVE(&ip->target_list, rec, target_entry); 1569 rec->target_ip = NULL; 1570 if (rec->flush_state == HAMMER_FST_SETUP) 1571 rec->flush_state = HAMMER_FST_IDLE; 1572 } 1573 while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) { 1574 if (rec->flush_state == HAMMER_FST_FLUSH) 1575 --rec->flush_group->refs; 1576 else 1577 hammer_ref(&rec->lock); 1578 KKASSERT(hammer_oneref(&rec->lock)); 1579 rec->flush_state = HAMMER_FST_IDLE; 1580 rec->flush_group = NULL; 1581 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */ 1582 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */ 1583 ++ip->rec_generation; 1584 hammer_rel_mem_record(rec); 1585 } 1586 ip->flags &= ~HAMMER_INODE_MODMASK; 1587 ip->sync_flags &= ~HAMMER_INODE_MODMASK; 1588 KKASSERT(ip->vp == NULL); 1589 1590 /* 1591 * Remove the inode from any flush group, force it idle. FLUSH 1592 * and SETUP states have an inode ref. 1593 */ 1594 switch(ip->flush_state) { 1595 case HAMMER_FST_FLUSH: 1596 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip); 1597 --ip->flush_group->refs; 1598 ip->flush_group = NULL; 1599 /* fall through */ 1600 case HAMMER_FST_SETUP: 1601 hammer_rel(&ip->lock); 1602 ip->flush_state = HAMMER_FST_IDLE; 1603 /* fall through */ 1604 case HAMMER_FST_IDLE: 1605 break; 1606 } 1607 1608 /* 1609 * There shouldn't be any associated vnode. The unload needs at 1610 * least one ref, if we do have a vp steal its ip ref. 1611 */ 1612 if (ip->vp) { 1613 hdkprintf("Unexpected vnode association ip %p vp %p\n", 1614 ip, ip->vp); 1615 ip->vp->v_data = NULL; 1616 ip->vp = NULL; 1617 } else { 1618 hammer_ref(&ip->lock); 1619 } 1620 hammer_unload_inode(ip); 1621 return(0); 1622 } 1623 1624 /* 1625 * Called on mount -u when switching from RW to RO or vise-versa. Adjust 1626 * the read-only flag for cached inodes. 1627 * 1628 * This routine is called from a RB_SCAN(). 1629 */ 1630 int 1631 hammer_reload_inode(hammer_inode_t ip, void *arg __unused) 1632 { 1633 hammer_mount_t hmp = ip->hmp; 1634 1635 if (hmp->ronly || hmp->asof != HAMMER_MAX_TID) 1636 ip->flags |= HAMMER_INODE_RO; 1637 else 1638 ip->flags &= ~HAMMER_INODE_RO; 1639 return(0); 1640 } 1641 1642 /* 1643 * A transaction has modified an inode, requiring updates as specified by 1644 * the passed flags. 1645 * 1646 * HAMMER_INODE_DDIRTY: Inode data has been updated, not incl mtime/atime, 1647 * and not including size changes due to write-append 1648 * (but other size changes are included). 1649 * HAMMER_INODE_SDIRTY: Inode data has been updated, size changes due to 1650 * write-append. 1651 * HAMMER_INODE_XDIRTY: Dirty in-memory records 1652 * HAMMER_INODE_BUFS: Dirty buffer cache buffers 1653 * HAMMER_INODE_DELETED: Inode record/data must be deleted 1654 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated 1655 */ 1656 void 1657 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags) 1658 { 1659 /* 1660 * ronly of 0 or 2 does not trigger assertion. 1661 * 2 is a special error state 1662 */ 1663 KKASSERT(ip->hmp->ronly != 1 || 1664 (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY | 1665 HAMMER_INODE_SDIRTY | 1666 HAMMER_INODE_BUFS | HAMMER_INODE_DELETED | 1667 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0); 1668 if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) { 1669 ip->flags |= HAMMER_INODE_RSV_INODES; 1670 ++ip->hmp->rsv_inodes; 1671 } 1672 1673 /* 1674 * Set the NEWINODE flag in the transaction if the inode 1675 * transitions to a dirty state. This is used to track 1676 * the load on the inode cache. 1677 */ 1678 if (trans && 1679 (ip->flags & HAMMER_INODE_MODMASK) == 0 && 1680 (flags & HAMMER_INODE_MODMASK)) { 1681 trans->flags |= HAMMER_TRANSF_NEWINODE; 1682 } 1683 if (flags & HAMMER_INODE_MODMASK) 1684 hammer_inode_dirty(ip); 1685 ip->flags |= flags; 1686 } 1687 1688 /* 1689 * Attempt to quickly update the atime for a hammer inode. Return 0 on 1690 * success, -1 on failure. 1691 * 1692 * We attempt to update the atime with only the ip lock and not the 1693 * whole filesystem lock in order to improve concurrency. We can only 1694 * do this safely if the ATIME flag is already pending on the inode. 1695 * 1696 * This function is called via a vnops path (ip pointer is stable) without 1697 * fs_token held. 1698 */ 1699 int 1700 hammer_update_atime_quick(hammer_inode_t ip) 1701 { 1702 struct timeval tv; 1703 int res = -1; 1704 1705 if ((ip->flags & HAMMER_INODE_RO) || 1706 (ip->hmp->mp->mnt_flag & MNT_NOATIME)) { 1707 /* 1708 * Silently indicate success on read-only mount/snap 1709 */ 1710 res = 0; 1711 } else if (ip->flags & HAMMER_INODE_ATIME) { 1712 /* 1713 * Double check with inode lock held against backend. This 1714 * is only safe if all we need to do is update 1715 * ino_data.atime. 1716 */ 1717 getmicrotime(&tv); 1718 hammer_lock_ex(&ip->lock); 1719 if (ip->flags & HAMMER_INODE_ATIME) { 1720 ip->ino_data.atime = 1721 (unsigned long)tv.tv_sec * 1000000ULL + tv.tv_usec; 1722 res = 0; 1723 } 1724 hammer_unlock(&ip->lock); 1725 } 1726 return res; 1727 } 1728 1729 /* 1730 * Request that an inode be flushed. This whole mess cannot block and may 1731 * recurse (if not synchronous). Once requested HAMMER will attempt to 1732 * actively flush the inode until the flush can be done. 1733 * 1734 * The inode may already be flushing, or may be in a setup state. We can 1735 * place the inode in a flushing state if it is currently idle and flag it 1736 * to reflush if it is currently flushing. 1737 * 1738 * Upon return if the inode could not be flushed due to a setup 1739 * dependancy, then it will be automatically flushed when the dependancy 1740 * is satisfied. 1741 */ 1742 void 1743 hammer_flush_inode(hammer_inode_t ip, int flags) 1744 { 1745 hammer_mount_t hmp; 1746 hammer_flush_group_t flg; 1747 int good; 1748 1749 /* 1750 * fill_flush_group is the first flush group we may be able to 1751 * continue filling, it may be open or closed but it will always 1752 * be past the currently flushing (running) flg. 1753 * 1754 * next_flush_group is the next open flush group. 1755 */ 1756 hmp = ip->hmp; 1757 while ((flg = hmp->fill_flush_group) != NULL) { 1758 KKASSERT(flg->running == 0); 1759 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit && 1760 flg->total_count <= hammer_autoflush) { 1761 break; 1762 } 1763 hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry); 1764 hammer_flusher_async(ip->hmp, flg); 1765 } 1766 if (flg == NULL) { 1767 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO); 1768 flg->seq = hmp->flusher.next++; 1769 if (hmp->next_flush_group == NULL) 1770 hmp->next_flush_group = flg; 1771 if (hmp->fill_flush_group == NULL) 1772 hmp->fill_flush_group = flg; 1773 RB_INIT(&flg->flush_tree); 1774 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry); 1775 } 1776 1777 /* 1778 * Trivial 'nothing to flush' case. If the inode is in a SETUP 1779 * state we have to put it back into an IDLE state so we can 1780 * drop the extra ref. 1781 * 1782 * If we have a parent dependancy we must still fall through 1783 * so we can run it. 1784 */ 1785 if ((ip->flags & HAMMER_INODE_MODMASK) == 0) { 1786 if (ip->flush_state == HAMMER_FST_SETUP && 1787 TAILQ_EMPTY(&ip->target_list)) { 1788 ip->flush_state = HAMMER_FST_IDLE; 1789 hammer_rel_inode(ip, 0); 1790 } 1791 if (ip->flush_state == HAMMER_FST_IDLE) 1792 return; 1793 } 1794 1795 /* 1796 * Our flush action will depend on the current state. 1797 */ 1798 switch(ip->flush_state) { 1799 case HAMMER_FST_IDLE: 1800 /* 1801 * We have no dependancies and can flush immediately. Some 1802 * our children may not be flushable so we have to re-test 1803 * with that additional knowledge. 1804 */ 1805 hammer_flush_inode_core(ip, flg, flags); 1806 break; 1807 case HAMMER_FST_SETUP: 1808 /* 1809 * Recurse upwards through dependancies via target_list 1810 * and start their flusher actions going if possible. 1811 * 1812 * 'good' is our connectivity. -1 means we have none and 1813 * can't flush, 0 means there weren't any dependancies, and 1814 * 1 means we have good connectivity. 1815 */ 1816 good = hammer_setup_parent_inodes(ip, 0, flg); 1817 1818 if (good >= 0) { 1819 /* 1820 * We can continue if good >= 0. Determine how 1821 * many records under our inode can be flushed (and 1822 * mark them). 1823 */ 1824 hammer_flush_inode_core(ip, flg, flags); 1825 } else { 1826 /* 1827 * Parent has no connectivity, tell it to flush 1828 * us as soon as it does. 1829 * 1830 * The REFLUSH flag is also needed to trigger 1831 * dependancy wakeups. 1832 */ 1833 ip->flags |= HAMMER_INODE_CONN_DOWN | 1834 HAMMER_INODE_REFLUSH; 1835 if (flags & HAMMER_FLUSH_SIGNAL) { 1836 ip->flags |= HAMMER_INODE_RESIGNAL; 1837 hammer_flusher_async(ip->hmp, flg); 1838 } 1839 } 1840 break; 1841 case HAMMER_FST_FLUSH: 1842 /* 1843 * We are already flushing, flag the inode to reflush 1844 * if needed after it completes its current flush. 1845 * 1846 * The REFLUSH flag is also needed to trigger 1847 * dependancy wakeups. 1848 */ 1849 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0) 1850 ip->flags |= HAMMER_INODE_REFLUSH; 1851 if (flags & HAMMER_FLUSH_SIGNAL) { 1852 ip->flags |= HAMMER_INODE_RESIGNAL; 1853 hammer_flusher_async(ip->hmp, flg); 1854 } 1855 break; 1856 } 1857 } 1858 1859 /* 1860 * Scan ip->target_list, which is a list of records owned by PARENTS to our 1861 * ip which reference our ip. 1862 * 1863 * XXX This is a huge mess of recursive code, but not one bit of it blocks 1864 * so for now do not ref/deref the structures. Note that if we use the 1865 * ref/rel code later, the rel CAN block. 1866 */ 1867 static int 1868 hammer_setup_parent_inodes(hammer_inode_t ip, int depth, 1869 hammer_flush_group_t flg) 1870 { 1871 hammer_record_t depend; 1872 int good; 1873 int r; 1874 1875 /* 1876 * If we hit our recursion limit and we have parent dependencies 1877 * We cannot continue. Returning < 0 will cause us to be flagged 1878 * for reflush. Returning -2 cuts off additional dependency checks 1879 * because they are likely to also hit the depth limit. 1880 * 1881 * We cannot return < 0 if there are no dependencies or there might 1882 * not be anything to wakeup (ip). 1883 */ 1884 if (depth == 20 && TAILQ_FIRST(&ip->target_list)) { 1885 if (hammer_debug_general & 0x10000) 1886 hkrateprintf(&hammer_gen_krate, 1887 "Warning: depth limit reached on " 1888 "setup recursion, inode %p %016llx\n", 1889 ip, (long long)ip->obj_id); 1890 return(-2); 1891 } 1892 1893 /* 1894 * Scan dependencies 1895 */ 1896 good = 0; 1897 TAILQ_FOREACH(depend, &ip->target_list, target_entry) { 1898 r = hammer_setup_parent_inodes_helper(depend, depth, flg); 1899 KKASSERT(depend->target_ip == ip); 1900 if (r < 0 && good == 0) 1901 good = -1; 1902 if (r > 0) 1903 good = 1; 1904 1905 /* 1906 * If we failed due to the recursion depth limit then stop 1907 * now. 1908 */ 1909 if (r == -2) 1910 break; 1911 } 1912 return(good); 1913 } 1914 1915 /* 1916 * This helper function takes a record representing the dependancy between 1917 * the parent inode and child inode. 1918 * 1919 * record = record in question (*rec in below) 1920 * record->ip = parent inode (*pip in below) 1921 * record->target_ip = child inode (*ip in below) 1922 * 1923 * *pip--------------\ 1924 * ^ \rec_tree 1925 * \ \ 1926 * \ip /\\\\\ rbtree of recs from parent inode's view 1927 * \ //\\\\\\ 1928 * \ / ........ 1929 * \ / 1930 * \------*rec------target_ip------>*ip 1931 * ...target_entry<----...----->target_list<---... 1932 * list of recs from inode's view 1933 * 1934 * We are asked to recurse upwards and convert the record from SETUP 1935 * to FLUSH if possible. 1936 * 1937 * Return 1 if the record gives us connectivity 1938 * 1939 * Return 0 if the record is not relevant 1940 * 1941 * Return -1 if we can't resolve the dependancy and there is no connectivity. 1942 */ 1943 static int 1944 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth, 1945 hammer_flush_group_t flg) 1946 { 1947 hammer_inode_t pip; 1948 int good; 1949 1950 KKASSERT(record->flush_state != HAMMER_FST_IDLE); 1951 pip = record->ip; 1952 1953 /* 1954 * If the record is already flushing, is it in our flush group? 1955 * 1956 * If it is in our flush group but it is a general record or a 1957 * delete-on-disk, it does not improve our connectivity (return 0), 1958 * and if the target inode is not trying to destroy itself we can't 1959 * allow the operation yet anyway (the second return -1). 1960 */ 1961 if (record->flush_state == HAMMER_FST_FLUSH) { 1962 /* 1963 * If not in our flush group ask the parent to reflush 1964 * us as soon as possible. 1965 */ 1966 if (record->flush_group != flg) { 1967 pip->flags |= HAMMER_INODE_REFLUSH; 1968 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN; 1969 return(-1); 1970 } 1971 1972 /* 1973 * If in our flush group everything is already set up, 1974 * just return whether the record will improve our 1975 * visibility or not. 1976 */ 1977 if (record->type == HAMMER_MEM_RECORD_ADD) 1978 return(1); 1979 return(0); 1980 } 1981 1982 /* 1983 * It must be a setup record. Try to resolve the setup dependancies 1984 * by recursing upwards so we can place ip on the flush list. 1985 * 1986 * Limit ourselves to 20 levels of recursion to avoid blowing out 1987 * the kernel stack. If we hit the recursion limit we can't flush 1988 * until the parent flushes. The parent will flush independantly 1989 * on its own and ultimately a deep recursion will be resolved. 1990 */ 1991 KKASSERT(record->flush_state == HAMMER_FST_SETUP); 1992 1993 good = hammer_setup_parent_inodes(pip, depth + 1, flg); 1994 1995 /* 1996 * If good < 0 the parent has no connectivity and we cannot safely 1997 * flush the directory entry, which also means we can't flush our 1998 * ip. Flag us for downward recursion once the parent's 1999 * connectivity is resolved. Flag the parent for [re]flush or it 2000 * may not check for downward recursions. 2001 */ 2002 if (good < 0) { 2003 pip->flags |= HAMMER_INODE_REFLUSH; 2004 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN; 2005 return(good); 2006 } 2007 2008 /* 2009 * We are go, place the parent inode in a flushing state so we can 2010 * place its record in a flushing state. Note that the parent 2011 * may already be flushing. The record must be in the same flush 2012 * group as the parent. 2013 */ 2014 if (pip->flush_state != HAMMER_FST_FLUSH) 2015 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION); 2016 KKASSERT(pip->flush_state == HAMMER_FST_FLUSH); 2017 2018 /* 2019 * It is possible for a rename to create a loop in the recursion 2020 * and revisit a record. This will result in the record being 2021 * placed in a flush state unexpectedly. This check deals with 2022 * the case. 2023 */ 2024 if (record->flush_state == HAMMER_FST_FLUSH) { 2025 if (record->type == HAMMER_MEM_RECORD_ADD) 2026 return(1); 2027 return(0); 2028 } 2029 2030 KKASSERT(record->flush_state == HAMMER_FST_SETUP); 2031 2032 #if 0 2033 if (record->type == HAMMER_MEM_RECORD_DEL && 2034 (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) { 2035 /* 2036 * Regardless of flushing state we cannot sync this path if the 2037 * record represents a delete-on-disk but the target inode 2038 * is not ready to sync its own deletion. 2039 * 2040 * XXX need to count effective nlinks to determine whether 2041 * the flush is ok, otherwise removing a hardlink will 2042 * just leave the DEL record to rot. 2043 */ 2044 record->target_ip->flags |= HAMMER_INODE_REFLUSH; 2045 return(-1); 2046 } else 2047 #endif 2048 if (pip->flush_group == flg) { 2049 /* 2050 * Because we have not calculated nlinks yet we can just 2051 * set records to the flush state if the parent is in 2052 * the same flush group as we are. 2053 */ 2054 record->flush_state = HAMMER_FST_FLUSH; 2055 record->flush_group = flg; 2056 ++record->flush_group->refs; 2057 hammer_ref(&record->lock); 2058 2059 /* 2060 * A general directory-add contributes to our visibility. 2061 * 2062 * Otherwise it is probably a directory-delete or 2063 * delete-on-disk record and does not contribute to our 2064 * visbility (but we can still flush it). 2065 */ 2066 if (record->type == HAMMER_MEM_RECORD_ADD) 2067 return(1); 2068 return(0); 2069 } else { 2070 /* 2071 * If the parent is not in our flush group we cannot 2072 * flush this record yet, there is no visibility. 2073 * We tell the parent to reflush and mark ourselves 2074 * so the parent knows it should flush us too. 2075 */ 2076 pip->flags |= HAMMER_INODE_REFLUSH; 2077 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN; 2078 return(-1); 2079 } 2080 } 2081 2082 /* 2083 * This is the core routine placing an inode into the FST_FLUSH state. 2084 */ 2085 static void 2086 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags) 2087 { 2088 hammer_mount_t hmp = ip->hmp; 2089 int go_count; 2090 2091 /* 2092 * Set flush state and prevent the flusher from cycling into 2093 * the next flush group. Do not place the ip on the list yet. 2094 * Inodes not in the idle state get an extra reference. 2095 */ 2096 KKASSERT(ip->flush_state != HAMMER_FST_FLUSH); 2097 if (ip->flush_state == HAMMER_FST_IDLE) 2098 hammer_ref(&ip->lock); 2099 ip->flush_state = HAMMER_FST_FLUSH; 2100 ip->flush_group = flg; 2101 ++hmp->flusher.group_lock; 2102 ++hmp->count_iqueued; 2103 ++hammer_count_iqueued; 2104 ++flg->total_count; 2105 hammer_redo_fifo_start_flush(ip); 2106 2107 #if 0 2108 /* 2109 * We need to be able to vfsync/truncate from the backend. 2110 * 2111 * XXX Any truncation from the backend will acquire the vnode 2112 * independently. 2113 */ 2114 KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0); 2115 if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) { 2116 ip->flags |= HAMMER_INODE_VHELD; 2117 vref(ip->vp); 2118 } 2119 #endif 2120 2121 /* 2122 * Figure out how many in-memory records we can actually flush 2123 * (not including inode meta-data, buffers, etc). 2124 */ 2125 KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0); 2126 if (flags & HAMMER_FLUSH_RECURSION) { 2127 /* 2128 * If this is a upwards recursion we do not want to 2129 * recurse down again! 2130 */ 2131 go_count = 1; 2132 #if 0 2133 } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) { 2134 /* 2135 * No new records are added if we must complete a flush 2136 * from a previous cycle, but we do have to move the records 2137 * from the previous cycle to the current one. 2138 */ 2139 #if 0 2140 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL, 2141 hammer_syncgrp_child_callback, NULL); 2142 #endif 2143 go_count = 1; 2144 #endif 2145 } else { 2146 /* 2147 * Normal flush, scan records and bring them into the flush. 2148 * Directory adds and deletes are usually skipped (they are 2149 * grouped with the related inode rather then with the 2150 * directory). 2151 * 2152 * go_count can be negative, which means the scan aborted 2153 * due to the flush group being over-full and we should 2154 * flush what we have. 2155 */ 2156 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL, 2157 hammer_setup_child_callback, NULL); 2158 } 2159 2160 /* 2161 * This is a more involved test that includes go_count. If we 2162 * can't flush, flag the inode and return. If go_count is 0 we 2163 * were are unable to flush any records in our rec_tree and 2164 * must ignore the XDIRTY flag. 2165 */ 2166 if (go_count == 0) { 2167 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) { 2168 --hmp->count_iqueued; 2169 --hammer_count_iqueued; 2170 2171 --flg->total_count; 2172 ip->flush_state = HAMMER_FST_SETUP; 2173 ip->flush_group = NULL; 2174 if (flags & HAMMER_FLUSH_SIGNAL) { 2175 ip->flags |= HAMMER_INODE_REFLUSH | 2176 HAMMER_INODE_RESIGNAL; 2177 } else { 2178 ip->flags |= HAMMER_INODE_REFLUSH; 2179 } 2180 #if 0 2181 if (ip->flags & HAMMER_INODE_VHELD) { 2182 ip->flags &= ~HAMMER_INODE_VHELD; 2183 vrele(ip->vp); 2184 } 2185 #endif 2186 2187 /* 2188 * REFLUSH is needed to trigger dependancy wakeups 2189 * when an inode is in SETUP. 2190 */ 2191 ip->flags |= HAMMER_INODE_REFLUSH; 2192 if (--hmp->flusher.group_lock == 0) 2193 wakeup(&hmp->flusher.group_lock); 2194 return; 2195 } 2196 } 2197 2198 /* 2199 * Snapshot the state of the inode for the backend flusher. 2200 * 2201 * We continue to retain save_trunc_off even when all truncations 2202 * have been resolved as an optimization to determine if we can 2203 * skip the B-Tree lookup for overwrite deletions. 2204 * 2205 * NOTE: The DELETING flag is a mod flag, but it is also sticky, 2206 * and stays in ip->flags. Once set, it stays set until the 2207 * inode is destroyed. 2208 */ 2209 if (ip->flags & HAMMER_INODE_TRUNCATED) { 2210 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0); 2211 ip->sync_trunc_off = ip->trunc_off; 2212 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL; 2213 ip->flags &= ~HAMMER_INODE_TRUNCATED; 2214 ip->sync_flags |= HAMMER_INODE_TRUNCATED; 2215 2216 /* 2217 * The save_trunc_off used to cache whether the B-Tree 2218 * holds any records past that point is not used until 2219 * after the truncation has succeeded, so we can safely 2220 * set it now. 2221 */ 2222 if (ip->save_trunc_off > ip->sync_trunc_off) 2223 ip->save_trunc_off = ip->sync_trunc_off; 2224 } 2225 ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK & 2226 ~HAMMER_INODE_TRUNCATED); 2227 ip->sync_ino_leaf = ip->ino_leaf; 2228 ip->sync_ino_data = ip->ino_data; 2229 ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED; 2230 #ifdef DEBUG_TRUNCATE 2231 if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp) 2232 hdkprintf("truncate %016llx\n", ip->sync_trunc_off); 2233 #endif 2234 2235 /* 2236 * The flusher list inherits our inode and reference. 2237 */ 2238 KKASSERT(flg->running == 0); 2239 RB_INSERT(hammer_fls_rb_tree, &flg->flush_tree, ip); 2240 if (--hmp->flusher.group_lock == 0) 2241 wakeup(&hmp->flusher.group_lock); 2242 2243 /* 2244 * Auto-flush the group if it grows too large. Make sure the 2245 * inode reclaim wait pipeline continues to work. 2246 */ 2247 if (flg->total_count >= hammer_autoflush || 2248 flg->total_count >= hammer_limit_reclaims / 4) { 2249 if (hmp->fill_flush_group == flg) 2250 hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry); 2251 hammer_flusher_async(hmp, flg); 2252 } 2253 } 2254 2255 /* 2256 * Callback for scan of ip->rec_tree. Try to include each record in our 2257 * flush. ip->flush_group has been set but the inode has not yet been 2258 * moved into a flushing state. 2259 * 2260 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on 2261 * both inodes. 2262 * 2263 * We return 1 for any record placed or found in FST_FLUSH, which prevents 2264 * the caller from shortcutting the flush. 2265 */ 2266 static int 2267 hammer_setup_child_callback(hammer_record_t rec, void *data) 2268 { 2269 hammer_flush_group_t flg; 2270 hammer_inode_t target_ip; 2271 hammer_inode_t ip; 2272 int r; 2273 2274 /* 2275 * Records deleted or committed by the backend are ignored. 2276 * Note that the flush detects deleted frontend records at 2277 * multiple points to deal with races. This is just the first 2278 * line of defense. The only time HAMMER_RECF_DELETED_FE cannot 2279 * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it 2280 * messes up link-count calculations. 2281 * 2282 * NOTE: Don't get confused between record deletion and, say, 2283 * directory entry deletion. The deletion of a directory entry 2284 * which is on-media has nothing to do with the record deletion 2285 * flags. 2286 */ 2287 if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE | 2288 HAMMER_RECF_COMMITTED)) { 2289 if (rec->flush_state == HAMMER_FST_FLUSH) { 2290 KKASSERT(rec->flush_group == rec->ip->flush_group); 2291 r = 1; 2292 } else { 2293 r = 0; 2294 } 2295 return(r); 2296 } 2297 2298 /* 2299 * If the record is in an idle state it has no dependancies and 2300 * can be flushed. 2301 */ 2302 ip = rec->ip; 2303 flg = ip->flush_group; 2304 r = 0; 2305 2306 switch(rec->flush_state) { 2307 case HAMMER_FST_IDLE: 2308 /* 2309 * The record has no setup dependancy, we can flush it. 2310 */ 2311 KKASSERT(rec->target_ip == NULL); 2312 rec->flush_state = HAMMER_FST_FLUSH; 2313 rec->flush_group = flg; 2314 ++flg->refs; 2315 hammer_ref(&rec->lock); 2316 r = 1; 2317 break; 2318 case HAMMER_FST_SETUP: 2319 /* 2320 * The record has a setup dependancy. These are typically 2321 * directory entry adds and deletes. Such entries will be 2322 * flushed when their inodes are flushed so we do not 2323 * usually have to add them to the flush here. However, 2324 * if the target_ip has set HAMMER_INODE_CONN_DOWN then 2325 * it is asking us to flush this record (and it). 2326 */ 2327 target_ip = rec->target_ip; 2328 KKASSERT(target_ip != NULL); 2329 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE); 2330 2331 /* 2332 * If the target IP is already flushing in our group 2333 * we could associate the record, but target_ip has 2334 * already synced ino_data to sync_ino_data and we 2335 * would also have to adjust nlinks. Plus there are 2336 * ordering issues for adds and deletes. 2337 * 2338 * Reflush downward if this is an ADD, and upward if 2339 * this is a DEL. 2340 */ 2341 if (target_ip->flush_state == HAMMER_FST_FLUSH) { 2342 if (rec->type == HAMMER_MEM_RECORD_ADD) 2343 ip->flags |= HAMMER_INODE_REFLUSH; 2344 else 2345 target_ip->flags |= HAMMER_INODE_REFLUSH; 2346 break; 2347 } 2348 2349 /* 2350 * Target IP is not yet flushing. This can get complex 2351 * because we have to be careful about the recursion. 2352 * 2353 * Directories create an issue for us in that if a flush 2354 * of a directory is requested the expectation is to flush 2355 * any pending directory entries, but this will cause the 2356 * related inodes to recursively flush as well. We can't 2357 * really defer the operation so just get as many as we 2358 * can and 2359 */ 2360 #if 0 2361 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 && 2362 (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) { 2363 /* 2364 * We aren't reclaiming and the target ip was not 2365 * previously prevented from flushing due to this 2366 * record dependancy. Do not flush this record. 2367 */ 2368 /*r = 0;*/ 2369 } else 2370 #endif 2371 if (flg->total_count + flg->refs > 2372 ip->hmp->undo_rec_limit) { 2373 /* 2374 * Our flush group is over-full and we risk blowing 2375 * out the UNDO FIFO. Stop the scan, flush what we 2376 * have, then reflush the directory. 2377 * 2378 * The directory may be forced through multiple 2379 * flush groups before it can be completely 2380 * flushed. 2381 */ 2382 ip->flags |= HAMMER_INODE_RESIGNAL | 2383 HAMMER_INODE_REFLUSH; 2384 r = -1; 2385 } else if (rec->type == HAMMER_MEM_RECORD_ADD) { 2386 /* 2387 * If the target IP is not flushing we can force 2388 * it to flush, even if it is unable to write out 2389 * any of its own records we have at least one in 2390 * hand that we CAN deal with. 2391 */ 2392 rec->flush_state = HAMMER_FST_FLUSH; 2393 rec->flush_group = flg; 2394 ++flg->refs; 2395 hammer_ref(&rec->lock); 2396 hammer_flush_inode_core(target_ip, flg, 2397 HAMMER_FLUSH_RECURSION); 2398 r = 1; 2399 } else { 2400 /* 2401 * General or delete-on-disk record. 2402 * 2403 * XXX this needs help. If a delete-on-disk we could 2404 * disconnect the target. If the target has its own 2405 * dependancies they really need to be flushed. 2406 * 2407 * XXX 2408 */ 2409 rec->flush_state = HAMMER_FST_FLUSH; 2410 rec->flush_group = flg; 2411 ++flg->refs; 2412 hammer_ref(&rec->lock); 2413 hammer_flush_inode_core(target_ip, flg, 2414 HAMMER_FLUSH_RECURSION); 2415 r = 1; 2416 } 2417 break; 2418 case HAMMER_FST_FLUSH: 2419 /* 2420 * The record could be part of a previous flush group if the 2421 * inode is a directory (the record being a directory entry). 2422 * Once the flush group was closed a hammer_test_inode() 2423 * function can cause a new flush group to be setup, placing 2424 * the directory inode itself in a new flush group. 2425 * 2426 * When associated with a previous flush group we count it 2427 * as if it were in our current flush group, since it will 2428 * effectively be flushed by the time we flush our current 2429 * flush group. 2430 */ 2431 KKASSERT( 2432 rec->ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY || 2433 rec->flush_group == flg); 2434 r = 1; 2435 break; 2436 } 2437 return(r); 2438 } 2439 2440 #if 0 2441 /* 2442 * This version just moves records already in a flush state to the new 2443 * flush group and that is it. 2444 */ 2445 static int 2446 hammer_syncgrp_child_callback(hammer_record_t rec, void *data) 2447 { 2448 hammer_inode_t ip = rec->ip; 2449 2450 switch(rec->flush_state) { 2451 case HAMMER_FST_FLUSH: 2452 KKASSERT(rec->flush_group == ip->flush_group); 2453 break; 2454 default: 2455 break; 2456 } 2457 return(0); 2458 } 2459 #endif 2460 2461 /* 2462 * Wait for a previously queued flush to complete. 2463 * 2464 * If a critical error occured we don't try to wait. 2465 */ 2466 void 2467 hammer_wait_inode(hammer_inode_t ip) 2468 { 2469 /* 2470 * The inode can be in a SETUP state in which case RESIGNAL 2471 * should be set. If RESIGNAL is not set then the previous 2472 * flush completed and a later operation placed the inode 2473 * in a passive setup state again, so we're done. 2474 * 2475 * The inode can be in a FLUSH state in which case we 2476 * can just wait for completion. 2477 */ 2478 while (ip->flush_state == HAMMER_FST_FLUSH || 2479 (ip->flush_state == HAMMER_FST_SETUP && 2480 (ip->flags & HAMMER_INODE_RESIGNAL))) { 2481 /* 2482 * Don't try to flush on a critical error 2483 */ 2484 if (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) 2485 break; 2486 2487 /* 2488 * If the inode was already being flushed its flg 2489 * may not have been queued to the backend. We 2490 * have to make sure it gets queued or we can wind 2491 * up blocked or deadlocked (particularly if we are 2492 * the vnlru thread). 2493 */ 2494 if (ip->flush_state == HAMMER_FST_FLUSH) { 2495 KKASSERT(ip->flush_group); 2496 if (ip->flush_group->closed == 0) { 2497 if (hammer_debug_inode) { 2498 hkprintf("debug: forcing " 2499 "async flush ip %016jx\n", 2500 (intmax_t)ip->obj_id); 2501 } 2502 hammer_flusher_async(ip->hmp, ip->flush_group); 2503 continue; /* retest */ 2504 } 2505 } 2506 2507 /* 2508 * In a flush state with the flg queued to the backend 2509 * or in a setup state with RESIGNAL set, we can safely 2510 * wait. 2511 */ 2512 ip->flags |= HAMMER_INODE_FLUSHW; 2513 tsleep(&ip->flags, 0, "hmrwin", 0); 2514 } 2515 2516 #if 0 2517 /* 2518 * The inode may have been in a passive setup state, 2519 * call flush to make sure we get signaled. 2520 */ 2521 if (ip->flush_state == HAMMER_FST_SETUP) 2522 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL); 2523 #endif 2524 2525 } 2526 2527 /* 2528 * Called by the backend code when a flush has been completed. 2529 * The inode has already been removed from the flush list. 2530 * 2531 * A pipelined flush can occur, in which case we must re-enter the 2532 * inode on the list and re-copy its fields. 2533 */ 2534 void 2535 hammer_flush_inode_done(hammer_inode_t ip, int error) 2536 { 2537 hammer_mount_t hmp; 2538 int dorel; 2539 2540 KKASSERT(ip->flush_state == HAMMER_FST_FLUSH); 2541 2542 hmp = ip->hmp; 2543 2544 /* 2545 * Auto-reflush if the backend could not completely flush 2546 * the inode. This fixes a case where a deferred buffer flush 2547 * could cause fsync to return early. 2548 */ 2549 if (ip->sync_flags & HAMMER_INODE_MODMASK) 2550 ip->flags |= HAMMER_INODE_REFLUSH; 2551 2552 /* 2553 * Merge left-over flags back into the frontend and fix the state. 2554 * Incomplete truncations are retained by the backend. 2555 */ 2556 ip->error = error; 2557 ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED; 2558 ip->sync_flags &= HAMMER_INODE_TRUNCATED; 2559 2560 /* 2561 * The backend may have adjusted nlinks, so if the adjusted nlinks 2562 * does not match the fronttend set the frontend's DDIRTY flag again. 2563 */ 2564 if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks) 2565 ip->flags |= HAMMER_INODE_DDIRTY; 2566 2567 /* 2568 * Fix up the dirty buffer status. 2569 */ 2570 if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) { 2571 ip->flags |= HAMMER_INODE_BUFS; 2572 } 2573 hammer_redo_fifo_end_flush(ip); 2574 2575 /* 2576 * Re-set the XDIRTY flag if some of the inode's in-memory records 2577 * could not be flushed. 2578 */ 2579 KKASSERT((RB_EMPTY(&ip->rec_tree) && 2580 (ip->flags & HAMMER_INODE_XDIRTY) == 0) || 2581 (!RB_EMPTY(&ip->rec_tree) && 2582 (ip->flags & HAMMER_INODE_XDIRTY) != 0)); 2583 2584 /* 2585 * Do not lose track of inodes which no longer have vnode 2586 * assocations, otherwise they may never get flushed again. 2587 * 2588 * The reflush flag can be set superfluously, causing extra pain 2589 * for no reason. If the inode is no longer modified it no longer 2590 * needs to be flushed. 2591 */ 2592 if (ip->flags & HAMMER_INODE_MODMASK) { 2593 if (ip->vp == NULL) 2594 ip->flags |= HAMMER_INODE_REFLUSH; 2595 } else { 2596 ip->flags &= ~HAMMER_INODE_REFLUSH; 2597 } 2598 2599 /* 2600 * The fs token is held but the inode lock is not held. Because this 2601 * is a backend flush it is possible that the vnode has no references 2602 * and cause a reclaim race inside vsetisdirty() if/when it blocks. 2603 * 2604 * Therefore, we must lock the inode around this particular dirtying 2605 * operation. We don't have to around other dirtying operations 2606 * where the vnode is implicitly or explicitly held. 2607 */ 2608 if (ip->flags & HAMMER_INODE_MODMASK) { 2609 hammer_lock_ex(&ip->lock); 2610 hammer_inode_dirty(ip); 2611 hammer_unlock(&ip->lock); 2612 } 2613 2614 /* 2615 * Adjust the flush state. 2616 */ 2617 if (ip->flags & HAMMER_INODE_WOULDBLOCK) { 2618 /* 2619 * We were unable to flush out all our records, leave the 2620 * inode in a flush state and in the current flush group. 2621 * The flush group will be re-run. 2622 * 2623 * This occurs if the UNDO block gets too full or there is 2624 * too much dirty meta-data and allows the flusher to 2625 * finalize the UNDO block and then re-flush. 2626 */ 2627 ip->flags &= ~HAMMER_INODE_WOULDBLOCK; 2628 dorel = 0; 2629 } else { 2630 /* 2631 * Remove from the flush_group 2632 */ 2633 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip); 2634 ip->flush_group = NULL; 2635 2636 #if 0 2637 /* 2638 * Clean up the vnode ref and tracking counts. 2639 */ 2640 if (ip->flags & HAMMER_INODE_VHELD) { 2641 ip->flags &= ~HAMMER_INODE_VHELD; 2642 vrele(ip->vp); 2643 } 2644 #endif 2645 --hmp->count_iqueued; 2646 --hammer_count_iqueued; 2647 2648 /* 2649 * And adjust the state. 2650 */ 2651 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) { 2652 ip->flush_state = HAMMER_FST_IDLE; 2653 dorel = 1; 2654 } else { 2655 ip->flush_state = HAMMER_FST_SETUP; 2656 dorel = 0; 2657 } 2658 2659 /* 2660 * If the frontend is waiting for a flush to complete, 2661 * wake it up. 2662 */ 2663 if (ip->flags & HAMMER_INODE_FLUSHW) { 2664 ip->flags &= ~HAMMER_INODE_FLUSHW; 2665 wakeup(&ip->flags); 2666 } 2667 2668 /* 2669 * If the frontend made more changes and requested another 2670 * flush, then try to get it running. 2671 * 2672 * Reflushes are aborted when the inode is errored out. 2673 */ 2674 if (ip->flags & HAMMER_INODE_REFLUSH) { 2675 ip->flags &= ~HAMMER_INODE_REFLUSH; 2676 if (ip->flags & HAMMER_INODE_RESIGNAL) { 2677 ip->flags &= ~HAMMER_INODE_RESIGNAL; 2678 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL); 2679 } else { 2680 hammer_flush_inode(ip, 0); 2681 } 2682 } 2683 } 2684 2685 /* 2686 * If we have no parent dependancies we can clear CONN_DOWN 2687 */ 2688 if (TAILQ_EMPTY(&ip->target_list)) 2689 ip->flags &= ~HAMMER_INODE_CONN_DOWN; 2690 2691 /* 2692 * If the inode is now clean drop the space reservation. 2693 */ 2694 if ((ip->flags & HAMMER_INODE_MODMASK) == 0 && 2695 (ip->flags & HAMMER_INODE_RSV_INODES)) { 2696 ip->flags &= ~HAMMER_INODE_RSV_INODES; 2697 --hmp->rsv_inodes; 2698 } 2699 2700 ip->flags &= ~HAMMER_INODE_SLAVEFLUSH; 2701 2702 if (dorel) 2703 hammer_rel_inode(ip, 0); 2704 } 2705 2706 /* 2707 * Called from hammer_sync_inode() to synchronize in-memory records 2708 * to the media. 2709 */ 2710 static int 2711 hammer_sync_record_callback(hammer_record_t record, void *data) 2712 { 2713 hammer_cursor_t cursor = data; 2714 hammer_transaction_t trans = cursor->trans; 2715 hammer_mount_t hmp = trans->hmp; 2716 int error; 2717 2718 /* 2719 * Skip records that do not belong to the current flush. 2720 */ 2721 ++hammer_stats_record_iterations; 2722 if (record->flush_state != HAMMER_FST_FLUSH) 2723 return(0); 2724 2725 if (record->flush_group != record->ip->flush_group) { 2726 hdkprintf("rec %p ip %p bad flush group %p %p\n", 2727 record, 2728 record->ip, 2729 record->flush_group, 2730 record->ip->flush_group); 2731 if (hammer_debug_critical) 2732 Debugger("blah2"); 2733 return(0); 2734 } 2735 KKASSERT(record->flush_group == record->ip->flush_group); 2736 2737 /* 2738 * Interlock the record using the BE flag. Once BE is set the 2739 * frontend cannot change the state of FE. 2740 * 2741 * NOTE: If FE is set prior to us setting BE we still sync the 2742 * record out, but the flush completion code converts it to 2743 * a delete-on-disk record instead of destroying it. 2744 */ 2745 KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0); 2746 record->flags |= HAMMER_RECF_INTERLOCK_BE; 2747 2748 /* 2749 * The backend has already disposed of the record. 2750 */ 2751 if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) { 2752 error = 0; 2753 goto done; 2754 } 2755 2756 /* 2757 * If the whole inode is being deleted and all on-disk records will 2758 * be deleted very soon, we can't sync any new records to disk 2759 * because they will be deleted in the same transaction they were 2760 * created in (delete_tid == create_tid), which will assert. 2761 * 2762 * XXX There may be a case with RECORD_ADD with DELETED_FE set 2763 * that we currently panic on. 2764 */ 2765 if (record->ip->sync_flags & HAMMER_INODE_DELETING) { 2766 switch(record->type) { 2767 case HAMMER_MEM_RECORD_DATA: 2768 /* 2769 * We don't have to do anything, if the record was 2770 * committed the space will have been accounted for 2771 * in the blockmap. 2772 */ 2773 /* fall through */ 2774 case HAMMER_MEM_RECORD_GENERAL: 2775 /* 2776 * Set deleted-by-backend flag. Do not set the 2777 * backend committed flag, because we are throwing 2778 * the record away. 2779 */ 2780 record->flags |= HAMMER_RECF_DELETED_BE; 2781 ++record->ip->rec_generation; 2782 error = 0; 2783 goto done; 2784 case HAMMER_MEM_RECORD_ADD: 2785 hpanic("illegal add during inode deletion record %p", 2786 record); 2787 break; /* NOT REACHED */ 2788 case HAMMER_MEM_RECORD_INODE: 2789 hpanic("attempt to sync inode record %p?", record); 2790 break; /* NOT REACHED */ 2791 case HAMMER_MEM_RECORD_DEL: 2792 /* 2793 * Follow through and issue the on-disk deletion 2794 */ 2795 break; 2796 } 2797 } 2798 2799 /* 2800 * If DELETED_FE is set special handling is needed for directory 2801 * entries. Dependant pieces related to the directory entry may 2802 * have already been synced to disk. If this occurs we have to 2803 * sync the directory entry and then change the in-memory record 2804 * from an ADD to a DELETE to cover the fact that it's been 2805 * deleted by the frontend. 2806 * 2807 * A directory delete covering record (MEM_RECORD_DEL) can never 2808 * be deleted by the frontend. 2809 * 2810 * Any other record type (aka DATA) can be deleted by the frontend. 2811 * XXX At the moment the flusher must skip it because there may 2812 * be another data record in the flush group for the same block, 2813 * meaning that some frontend data changes can leak into the backend's 2814 * synchronization point. 2815 */ 2816 if (record->flags & HAMMER_RECF_DELETED_FE) { 2817 if (record->type == HAMMER_MEM_RECORD_ADD) { 2818 /* 2819 * Convert a front-end deleted directory-add to 2820 * a directory-delete entry later. 2821 */ 2822 record->flags |= HAMMER_RECF_CONVERT_DELETE; 2823 } else { 2824 /* 2825 * Dispose of the record (race case). Mark as 2826 * deleted by backend (and not committed). 2827 */ 2828 KKASSERT(record->type != HAMMER_MEM_RECORD_DEL); 2829 record->flags |= HAMMER_RECF_DELETED_BE; 2830 ++record->ip->rec_generation; 2831 error = 0; 2832 goto done; 2833 } 2834 } 2835 2836 /* 2837 * Assign the create_tid for new records. Deletions already 2838 * have the record's entire key properly set up. 2839 */ 2840 if (record->type != HAMMER_MEM_RECORD_DEL) { 2841 record->leaf.base.create_tid = trans->tid; 2842 record->leaf.create_ts = trans->time32; 2843 } 2844 2845 /* 2846 * This actually moves the record to the on-media B-Tree. We 2847 * must also generate REDO_TERM entries in the UNDO/REDO FIFO 2848 * indicating that the related REDO_WRITE(s) have been committed. 2849 * 2850 * During recovery any REDO_TERM's within the nominal recovery span 2851 * are ignored since the related meta-data is being undone, causing 2852 * any matching REDO_WRITEs to execute. The REDO_TERMs outside 2853 * the nominal recovery span will match against REDO_WRITEs and 2854 * prevent them from being executed (because the meta-data has 2855 * already been synchronized). 2856 */ 2857 if (record->flags & HAMMER_RECF_REDO) { 2858 KKASSERT(record->type == HAMMER_MEM_RECORD_DATA); 2859 hammer_generate_redo(trans, record->ip, 2860 record->leaf.base.key - 2861 record->leaf.data_len, 2862 HAMMER_REDO_TERM_WRITE, 2863 NULL, 2864 record->leaf.data_len); 2865 } 2866 2867 for (;;) { 2868 error = hammer_ip_sync_record_cursor(cursor, record); 2869 if (error != EDEADLK) 2870 break; 2871 hammer_done_cursor(cursor); 2872 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0], 2873 record->ip); 2874 if (error) 2875 break; 2876 } 2877 record->flags &= ~HAMMER_RECF_CONVERT_DELETE; 2878 2879 if (error) 2880 error = -error; 2881 done: 2882 hammer_flush_record_done(record, error); 2883 2884 /* 2885 * Do partial finalization if we have built up too many dirty 2886 * buffers. Otherwise a buffer cache deadlock can occur when 2887 * doing things like creating tens of thousands of tiny files. 2888 * 2889 * We must release our cursor lock to avoid a 3-way deadlock 2890 * due to the exclusive sync lock the finalizer must get. 2891 * 2892 * WARNING: See warnings in hammer_unlock_cursor() function. 2893 */ 2894 if (hammer_flusher_meta_limit(hmp) || 2895 vm_page_count_severe()) { 2896 hammer_unlock_cursor(cursor); 2897 hammer_flusher_finalize(trans, 0); 2898 hammer_lock_cursor(cursor); 2899 } 2900 return(error); 2901 } 2902 2903 /* 2904 * Backend function called by the flusher to sync an inode to media. 2905 */ 2906 int 2907 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip) 2908 { 2909 struct hammer_cursor cursor; 2910 hammer_node_t tmp_node; 2911 hammer_record_t depend; 2912 hammer_record_t next; 2913 int error, tmp_error; 2914 u_int64_t nlinks; 2915 2916 if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0) 2917 return(0); 2918 2919 error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip); 2920 if (error) 2921 goto done; 2922 2923 /* 2924 * Any directory records referencing this inode which are not in 2925 * our current flush group must adjust our nlink count for the 2926 * purposes of synchronizating to disk. 2927 * 2928 * Records which are in our flush group can be unlinked from our 2929 * inode now, potentially allowing the inode to be physically 2930 * deleted. 2931 * 2932 * This cannot block. 2933 */ 2934 nlinks = ip->ino_data.nlinks; 2935 next = TAILQ_FIRST(&ip->target_list); 2936 while ((depend = next) != NULL) { 2937 next = TAILQ_NEXT(depend, target_entry); 2938 if (depend->flush_state == HAMMER_FST_FLUSH && 2939 depend->flush_group == ip->flush_group) { 2940 /* 2941 * If this is an ADD that was deleted by the frontend 2942 * the frontend nlinks count will have already been 2943 * decremented, but the backend is going to sync its 2944 * directory entry and must account for it. The 2945 * record will be converted to a delete-on-disk when 2946 * it gets synced. 2947 * 2948 * If the ADD was not deleted by the frontend we 2949 * can remove the dependancy from our target_list. 2950 */ 2951 if (depend->flags & HAMMER_RECF_DELETED_FE) { 2952 ++nlinks; 2953 } else { 2954 TAILQ_REMOVE(&ip->target_list, depend, 2955 target_entry); 2956 depend->target_ip = NULL; 2957 } 2958 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) { 2959 /* 2960 * Not part of our flush group and not deleted by 2961 * the front-end, adjust the link count synced to 2962 * the media (undo what the frontend did when it 2963 * queued the record). 2964 */ 2965 KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0); 2966 switch(depend->type) { 2967 case HAMMER_MEM_RECORD_ADD: 2968 --nlinks; 2969 break; 2970 case HAMMER_MEM_RECORD_DEL: 2971 ++nlinks; 2972 break; 2973 default: 2974 break; 2975 } 2976 } 2977 } 2978 2979 /* 2980 * Set dirty if we had to modify the link count. 2981 */ 2982 if (ip->sync_ino_data.nlinks != nlinks) { 2983 KKASSERT((int64_t)nlinks >= 0); 2984 ip->sync_ino_data.nlinks = nlinks; 2985 ip->sync_flags |= HAMMER_INODE_DDIRTY; 2986 } 2987 2988 /* 2989 * If there is a trunction queued destroy any data past the (aligned) 2990 * truncation point. Userland will have dealt with the buffer 2991 * containing the truncation point for us. 2992 * 2993 * We don't flush pending frontend data buffers until after we've 2994 * dealt with the truncation. 2995 */ 2996 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) { 2997 /* 2998 * Interlock trunc_off. The VOP front-end may continue to 2999 * make adjustments to it while we are blocked. 3000 */ 3001 off_t trunc_off; 3002 off_t aligned_trunc_off; 3003 int blkmask; 3004 3005 trunc_off = ip->sync_trunc_off; 3006 blkmask = hammer_blocksize(trunc_off) - 1; 3007 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask; 3008 3009 /* 3010 * Delete any whole blocks on-media. The front-end has 3011 * already cleaned out any partial block and made it 3012 * pending. The front-end may have updated trunc_off 3013 * while we were blocked so we only use sync_trunc_off. 3014 * 3015 * This operation can blow out the buffer cache, EWOULDBLOCK 3016 * means we were unable to complete the deletion. The 3017 * deletion will update sync_trunc_off in that case. 3018 */ 3019 error = hammer_ip_delete_range(&cursor, ip, 3020 aligned_trunc_off, 3021 0x7FFFFFFFFFFFFFFFLL, 2); 3022 if (error == EWOULDBLOCK) { 3023 ip->flags |= HAMMER_INODE_WOULDBLOCK; 3024 error = 0; 3025 goto defer_buffer_flush; 3026 } 3027 3028 if (error) 3029 goto done; 3030 3031 /* 3032 * Generate a REDO_TERM_TRUNC entry in the UNDO/REDO FIFO. 3033 * 3034 * XXX we do this even if we did not previously generate 3035 * a REDO_TRUNC record. This operation may enclosed the 3036 * range for multiple prior truncation entries in the REDO 3037 * log. 3038 */ 3039 if (trans->hmp->version >= HAMMER_VOL_VERSION_FOUR && 3040 (ip->flags & HAMMER_INODE_RDIRTY)) { 3041 hammer_generate_redo(trans, ip, aligned_trunc_off, 3042 HAMMER_REDO_TERM_TRUNC, 3043 NULL, 0); 3044 } 3045 3046 /* 3047 * Clear the truncation flag on the backend after we have 3048 * completed the deletions. Backend data is now good again 3049 * (including new records we are about to sync, below). 3050 * 3051 * Leave sync_trunc_off intact. As we write additional 3052 * records the backend will update sync_trunc_off. This 3053 * tells the backend whether it can skip the overwrite 3054 * test. This should work properly even when the backend 3055 * writes full blocks where the truncation point straddles 3056 * the block because the comparison is against the base 3057 * offset of the record. 3058 */ 3059 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED; 3060 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */ 3061 } else { 3062 error = 0; 3063 } 3064 3065 /* 3066 * Now sync related records. These will typically be directory 3067 * entries, records tracking direct-writes, or delete-on-disk records. 3068 */ 3069 if (error == 0) { 3070 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL, 3071 hammer_sync_record_callback, &cursor); 3072 if (tmp_error < 0) 3073 tmp_error = -error; 3074 if (tmp_error) 3075 error = tmp_error; 3076 } 3077 hammer_cache_node(&ip->cache[1], cursor.node); 3078 3079 /* 3080 * Re-seek for inode update, assuming our cache hasn't been ripped 3081 * out from under us. 3082 */ 3083 if (error == 0) { 3084 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error); 3085 if (tmp_node) { 3086 hammer_cursor_downgrade(&cursor); 3087 hammer_lock_sh(&tmp_node->lock); 3088 if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0) 3089 hammer_cursor_seek(&cursor, tmp_node, 0); 3090 hammer_unlock(&tmp_node->lock); 3091 hammer_rel_node(tmp_node); 3092 } 3093 error = 0; 3094 } 3095 3096 /* 3097 * If we are deleting the inode the frontend had better not have 3098 * any active references on elements making up the inode. 3099 * 3100 * The call to hammer_ip_delete_clean() cleans up auxillary records 3101 * but not DB or DATA records. Those must have already been deleted 3102 * by the normal truncation mechanic. 3103 */ 3104 if (error == 0 && ip->sync_ino_data.nlinks == 0 && 3105 RB_EMPTY(&ip->rec_tree) && 3106 (ip->sync_flags & HAMMER_INODE_DELETING) && 3107 (ip->flags & HAMMER_INODE_DELETED) == 0) { 3108 int count1 = 0; 3109 3110 error = hammer_ip_delete_clean(&cursor, ip, &count1); 3111 if (error == 0) { 3112 ip->flags |= HAMMER_INODE_DELETED; 3113 ip->sync_flags &= ~HAMMER_INODE_DELETING; 3114 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED; 3115 KKASSERT(RB_EMPTY(&ip->rec_tree)); 3116 3117 /* 3118 * Set delete_tid in both the frontend and backend 3119 * copy of the inode record. The DELETED flag handles 3120 * this, do not set DDIRTY. 3121 */ 3122 ip->ino_leaf.base.delete_tid = trans->tid; 3123 ip->sync_ino_leaf.base.delete_tid = trans->tid; 3124 ip->ino_leaf.delete_ts = trans->time32; 3125 ip->sync_ino_leaf.delete_ts = trans->time32; 3126 3127 3128 /* 3129 * Adjust the inode count in the volume header 3130 */ 3131 hammer_sync_lock_sh(trans); 3132 if (ip->flags & HAMMER_INODE_ONDISK) { 3133 hammer_modify_volume_field(trans, 3134 trans->rootvol, 3135 vol0_stat_inodes); 3136 --ip->hmp->rootvol->ondisk->vol0_stat_inodes; 3137 hammer_modify_volume_done(trans->rootvol); 3138 } 3139 hammer_sync_unlock(trans); 3140 } 3141 } 3142 3143 if (error) 3144 goto done; 3145 ip->sync_flags &= ~HAMMER_INODE_BUFS; 3146 3147 defer_buffer_flush: 3148 /* 3149 * Now update the inode's on-disk inode-data and/or on-disk record. 3150 * DELETED and ONDISK are managed only in ip->flags. 3151 * 3152 * In the case of a defered buffer flush we still update the on-disk 3153 * inode to satisfy visibility requirements if there happen to be 3154 * directory dependancies. 3155 */ 3156 switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) { 3157 case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK: 3158 /* 3159 * If deleted and on-disk, don't set any additional flags. 3160 * the delete flag takes care of things. 3161 * 3162 * Clear flags which may have been set by the frontend. 3163 */ 3164 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY | 3165 HAMMER_INODE_SDIRTY | 3166 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME | 3167 HAMMER_INODE_DELETING); 3168 break; 3169 case HAMMER_INODE_DELETED: 3170 /* 3171 * Take care of the case where a deleted inode was never 3172 * flushed to the disk in the first place. 3173 * 3174 * Clear flags which may have been set by the frontend. 3175 */ 3176 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY | 3177 HAMMER_INODE_SDIRTY | 3178 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME | 3179 HAMMER_INODE_DELETING); 3180 while (RB_ROOT(&ip->rec_tree)) { 3181 hammer_record_t record = RB_ROOT(&ip->rec_tree); 3182 hammer_ref(&record->lock); 3183 KKASSERT(hammer_oneref(&record->lock)); 3184 record->flags |= HAMMER_RECF_DELETED_BE; 3185 ++record->ip->rec_generation; 3186 hammer_rel_mem_record(record); 3187 } 3188 break; 3189 case HAMMER_INODE_ONDISK: 3190 /* 3191 * If already on-disk, do not set any additional flags. 3192 */ 3193 break; 3194 default: 3195 /* 3196 * If not on-disk and not deleted, set DDIRTY to force 3197 * an initial record to be written. 3198 * 3199 * Also set the create_tid in both the frontend and backend 3200 * copy of the inode record. 3201 */ 3202 ip->ino_leaf.base.create_tid = trans->tid; 3203 ip->ino_leaf.create_ts = trans->time32; 3204 ip->sync_ino_leaf.base.create_tid = trans->tid; 3205 ip->sync_ino_leaf.create_ts = trans->time32; 3206 ip->sync_flags |= HAMMER_INODE_DDIRTY; 3207 break; 3208 } 3209 3210 /* 3211 * If DDIRTY or SDIRTY is set, write out a new record. 3212 * If the inode is already on-disk the old record is marked as 3213 * deleted. 3214 * 3215 * If DELETED is set hammer_update_inode() will delete the existing 3216 * record without writing out a new one. 3217 * 3218 * If *ONLY* the ITIMES flag is set we can update the record in-place. 3219 */ 3220 if (ip->flags & HAMMER_INODE_DELETED) { 3221 error = hammer_update_inode(&cursor, ip); 3222 } else 3223 if (!(ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY)) && 3224 (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) { 3225 error = hammer_update_itimes(&cursor, ip); 3226 } else 3227 if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY | 3228 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) { 3229 error = hammer_update_inode(&cursor, ip); 3230 } 3231 done: 3232 if (ip->flags & HAMMER_INODE_MODMASK) 3233 hammer_inode_dirty(ip); 3234 if (error) { 3235 hammer_critical_error(ip->hmp, ip, error, 3236 "while syncing inode"); 3237 } 3238 hammer_done_cursor(&cursor); 3239 return(error); 3240 } 3241 3242 /* 3243 * This routine is called when the OS is no longer actively referencing 3244 * the inode (but might still be keeping it cached), or when releasing 3245 * the last reference to an inode. 3246 * 3247 * At this point if the inode's nlinks count is zero we want to destroy 3248 * it, which may mean destroying it on-media too. 3249 */ 3250 void 3251 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp) 3252 { 3253 struct vnode *vp; 3254 3255 /* 3256 * Set the DELETING flag when the link count drops to 0 and the 3257 * OS no longer has any opens on the inode. 3258 * 3259 * The backend will clear DELETING (a mod flag) and set DELETED 3260 * (a state flag) when it is actually able to perform the 3261 * operation. 3262 * 3263 * Don't reflag the deletion if the flusher is currently syncing 3264 * one that was already flagged. A previously set DELETING flag 3265 * may bounce around flags and sync_flags until the operation is 3266 * completely done. 3267 * 3268 * Do not attempt to modify a snapshot inode (one set to read-only). 3269 */ 3270 if (ip->ino_data.nlinks == 0 && 3271 ((ip->flags | ip->sync_flags) & (HAMMER_INODE_RO|HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) { 3272 ip->flags |= HAMMER_INODE_DELETING; 3273 ip->flags |= HAMMER_INODE_TRUNCATED; 3274 ip->trunc_off = 0; 3275 vp = NULL; 3276 if (getvp) { 3277 if (hammer_get_vnode(ip, &vp) != 0) 3278 return; 3279 } 3280 3281 /* 3282 * Final cleanup 3283 */ 3284 if (ip->vp) 3285 nvtruncbuf(ip->vp, 0, HAMMER_BUFSIZE, 0, 0); 3286 if (ip->flags & HAMMER_INODE_MODMASK) 3287 hammer_inode_dirty(ip); 3288 if (getvp) 3289 vput(vp); 3290 } 3291 } 3292 3293 /* 3294 * After potentially resolving a dependancy the inode is tested 3295 * to determine whether it needs to be reflushed. 3296 */ 3297 void 3298 hammer_test_inode(hammer_inode_t ip) 3299 { 3300 if (ip->flags & HAMMER_INODE_REFLUSH) { 3301 ip->flags &= ~HAMMER_INODE_REFLUSH; 3302 hammer_ref(&ip->lock); 3303 if (ip->flags & HAMMER_INODE_RESIGNAL) { 3304 ip->flags &= ~HAMMER_INODE_RESIGNAL; 3305 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL); 3306 } else { 3307 hammer_flush_inode(ip, 0); 3308 } 3309 hammer_rel_inode(ip, 0); 3310 } 3311 } 3312 3313 /* 3314 * Clear the RECLAIM flag on an inode. This occurs when the inode is 3315 * reassociated with a vp or just before it gets freed. 3316 * 3317 * Pipeline wakeups to threads blocked due to an excessive number of 3318 * detached inodes. This typically occurs when atime updates accumulate 3319 * while scanning a directory tree. 3320 */ 3321 static void 3322 hammer_inode_wakereclaims(hammer_inode_t ip) 3323 { 3324 struct hammer_reclaim *reclaim; 3325 hammer_mount_t hmp = ip->hmp; 3326 3327 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) 3328 return; 3329 3330 --hammer_count_reclaims; 3331 --hmp->count_reclaims; 3332 ip->flags &= ~HAMMER_INODE_RECLAIM; 3333 3334 if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) { 3335 KKASSERT(reclaim->count > 0); 3336 if (--reclaim->count == 0) { 3337 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry); 3338 wakeup(reclaim); 3339 } 3340 } 3341 } 3342 3343 /* 3344 * Setup our reclaim pipeline. We only let so many detached (and dirty) 3345 * inodes build up before we start blocking. This routine is called 3346 * if a new inode is created or an inode is loaded from media. 3347 * 3348 * When we block we don't care *which* inode has finished reclaiming, 3349 * as long as one does. 3350 * 3351 * The reclaim pipeline is primarily governed by the auto-flush which is 3352 * 1/4 hammer_limit_reclaims. We don't want to block if the count is 3353 * less than 1/2 hammer_limit_reclaims. From 1/2 to full count is 3354 * dynamically governed. 3355 */ 3356 void 3357 hammer_inode_waitreclaims(hammer_transaction_t trans) 3358 { 3359 hammer_mount_t hmp = trans->hmp; 3360 struct hammer_reclaim reclaim; 3361 int lower_limit; 3362 3363 /* 3364 * Track inode load, delay if the number of reclaiming inodes is 3365 * between 2/4 and 4/4 hammer_limit_reclaims, depending. 3366 */ 3367 if (curthread->td_proc) { 3368 struct hammer_inostats *stats; 3369 3370 stats = hammer_inode_inostats(hmp, curthread->td_proc->p_pid); 3371 ++stats->count; 3372 3373 if (stats->count > hammer_limit_reclaims / 2) 3374 stats->count = hammer_limit_reclaims / 2; 3375 lower_limit = hammer_limit_reclaims - stats->count; 3376 if (hammer_debug_general & 0x10000) { 3377 hdkprintf("pid %5d limit %d\n", 3378 (int)curthread->td_proc->p_pid, lower_limit); 3379 } 3380 } else { 3381 lower_limit = hammer_limit_reclaims * 3 / 4; 3382 } 3383 if (hmp->count_reclaims >= lower_limit) { 3384 reclaim.count = 1; 3385 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry); 3386 tsleep(&reclaim, 0, "hmrrcm", hz); 3387 if (reclaim.count > 0) 3388 TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry); 3389 } 3390 } 3391 3392 /* 3393 * Keep track of reclaim statistics on a per-pid basis using a loose 3394 * 4-way set associative hash table. Collisions inherit the count of 3395 * the previous entry. 3396 * 3397 * NOTE: We want to be careful here to limit the chain size. If the chain 3398 * size is too large a pid will spread its stats out over too many 3399 * entries under certain types of heavy filesystem activity and 3400 * wind up not delaying long enough. 3401 */ 3402 static 3403 struct hammer_inostats * 3404 hammer_inode_inostats(hammer_mount_t hmp, pid_t pid) 3405 { 3406 struct hammer_inostats *stats; 3407 int delta; 3408 int chain; 3409 static volatile int iterator; /* we don't care about MP races */ 3410 3411 /* 3412 * Chain up to 4 times to find our entry. 3413 */ 3414 for (chain = 0; chain < 4; ++chain) { 3415 stats = &hmp->inostats[(pid + chain) & HAMMER_INOSTATS_HMASK]; 3416 if (stats->pid == pid) 3417 break; 3418 } 3419 3420 /* 3421 * Replace one of the four chaining entries with our new entry. 3422 */ 3423 if (chain == 4) { 3424 stats = &hmp->inostats[(pid + (iterator++ & 3)) & 3425 HAMMER_INOSTATS_HMASK]; 3426 stats->pid = pid; 3427 } 3428 3429 /* 3430 * Decay the entry 3431 */ 3432 if (stats->count && stats->ltick != ticks) { 3433 delta = ticks - stats->ltick; 3434 stats->ltick = ticks; 3435 if (delta <= 0 || delta > hz * 60) 3436 stats->count = 0; 3437 else 3438 stats->count = stats->count * hz / (hz + delta); 3439 } 3440 if (hammer_debug_general & 0x10000) 3441 hdkprintf("pid %5d stats %d\n", (int)pid, stats->count); 3442 return (stats); 3443 } 3444 3445 #if 0 3446 3447 /* 3448 * XXX not used, doesn't work very well due to the large batching nature 3449 * of flushes. 3450 * 3451 * A larger then normal backlog of inodes is sitting in the flusher, 3452 * enforce a general slowdown to let it catch up. This routine is only 3453 * called on completion of a non-flusher-related transaction which 3454 * performed B-Tree node I/O. 3455 * 3456 * It is possible for the flusher to stall in a continuous load. 3457 * blogbench -i1000 -o seems to do a good job generating this sort of load. 3458 * If the flusher is unable to catch up the inode count can bloat until 3459 * we run out of kvm. 3460 * 3461 * This is a bit of a hack. 3462 */ 3463 void 3464 hammer_inode_waithard(hammer_mount_t hmp) 3465 { 3466 /* 3467 * Hysteresis. 3468 */ 3469 if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) { 3470 if (hmp->count_reclaims < hammer_limit_reclaims / 2 && 3471 hmp->count_iqueued < hmp->count_inodes / 20) { 3472 hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY; 3473 return; 3474 } 3475 } else { 3476 if (hmp->count_reclaims < hammer_limit_reclaims || 3477 hmp->count_iqueued < hmp->count_inodes / 10) { 3478 return; 3479 } 3480 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY; 3481 } 3482 3483 /* 3484 * Block for one flush cycle. 3485 */ 3486 hammer_flusher_wait_next(hmp); 3487 } 3488 3489 #endif 3490