sys/vfs/hammer/hammer_inode.c

   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 static struct hammer_inode *__hammer_find_inode(hammer_transaction_t trans,
  55                                         int64_t obj_id, hammer_tid_t asof,
  56                                         uint32_t localization);
  57
  58 struct krate hammer_gen_krate = { 1 };
  59
  60 /*
  61  * RB-Tree support for inode structures
  62  */
  63 int
  64 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
  65 {
  66         if (ip1->obj_localization < ip2->obj_localization)
  67                 return(-1);
  68         if (ip1->obj_localization > ip2->obj_localization)
  69                 return(1);
  70         if (ip1->obj_id < ip2->obj_id)
  71                 return(-1);
  72         if (ip1->obj_id > ip2->obj_id)
  73                 return(1);
  74         if (ip1->obj_asof < ip2->obj_asof)
  75                 return(-1);
  76         if (ip1->obj_asof > ip2->obj_asof)
  77                 return(1);
  78         return(0);
  79 }
  80
  81 int
  82 hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
  83 {
  84         if (ip1->redo_fifo_start < ip2->redo_fifo_start)
  85                 return(-1);
  86         if (ip1->redo_fifo_start > ip2->redo_fifo_start)
  87                 return(1);
  88         return(0);
  89 }
  90
  91 /*
  92  * RB-Tree support for inode structures / special LOOKUP_INFO
  93  */
  94 static int
  95 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
  96 {
  97         if (info->obj_localization < ip->obj_localization)
  98                 return(-1);
  99         if (info->obj_localization > ip->obj_localization)
 100                 return(1);
 101         if (info->obj_id < ip->obj_id)
 102                 return(-1);
 103         if (info->obj_id > ip->obj_id)
 104                 return(1);
 105         if (info->obj_asof < ip->obj_asof)
 106                 return(-1);
 107         if (info->obj_asof > ip->obj_asof)
 108                 return(1);
 109         return(0);
 110 }
 111
 112 /*
 113  * Used by hammer_scan_inode_snapshots() to locate all of an object's
 114  * snapshots.  Note that the asof field is not tested, which we can get
 115  * away with because it is the lowest-priority field.
 116  */
 117 static int
 118 hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
 119 {
 120         hammer_inode_info_t info = data;
 121
 122         if (ip->obj_localization > info->obj_localization)
 123                 return(1);
 124         if (ip->obj_localization < info->obj_localization)
 125                 return(-1);
 126         if (ip->obj_id > info->obj_id)
 127                 return(1);
 128         if (ip->obj_id < info->obj_id)
 129                 return(-1);
 130         return(0);
 131 }
 132
 133 /*
 134  * Used by hammer_unload_pseudofs() to locate all inodes associated with
 135  * a particular PFS.
 136  */
 137 static int
 138 hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
 139 {
 140         uint32_t localization = *(uint32_t *)data;
 141         if (ip->obj_localization > localization)
 142                 return(1);
 143         if (ip->obj_localization < localization)
 144                 return(-1);
 145         return(0);
 146 }
 147
 148 /*
 149  * RB-Tree support for pseudofs structures
 150  */
 151 static int
 152 hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
 153 {
 154         if (p1->localization < p2->localization)
 155                 return(-1);
 156         if (p1->localization > p2->localization)
 157                 return(1);
 158         return(0);
 159 }
 160
 161
 162 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
 163 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 164                 hammer_inode_info_cmp, hammer_inode_info_t);
 165 RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 166              hammer_pfs_rb_compare, uint32_t, localization);
 167
 168 /*
 169  * The kernel is not actively referencing this vnode but is still holding
 170  * it cached.
 171  *
 172  * This is called from the frontend.
 173  *
 174  * MPALMOSTSAFE
 175  */
 176 int
 177 hammer_vop_inactive(struct vop_inactive_args *ap)
 178 {
 179         struct hammer_inode *ip = VTOI(ap->a_vp);
 180         hammer_mount_t hmp;
 181
 182         /*
 183          * Degenerate case
 184          */
 185         if (ip == NULL) {
 186                 vrecycle(ap->a_vp);
 187                 return(0);
 188         }
 189
 190         /*
 191          * If the inode no longer has visibility in the filesystem try to
 192          * recycle it immediately, even if the inode is dirty.  Recycling
 193          * it quickly allows the system to reclaim buffer cache and VM
 194          * resources which can matter a lot in a heavily loaded system.
 195          *
 196          * This can deadlock in vfsync() if we aren't careful.
 197          *
 198          * Do not queue the inode to the flusher if we still have visibility,
 199          * otherwise namespace calls such as chmod will unnecessarily generate
 200          * multiple inode updates.
 201          */
 202         if (ip->ino_data.nlinks == 0) {
 203                 hmp = ip->hmp;
 204                 lwkt_gettoken(&hmp->fs_token);
 205                 hammer_inode_unloadable_check(ip, 0);
 206                 if (ip->flags & HAMMER_INODE_MODMASK)
 207                         hammer_flush_inode(ip, 0);
 208                 lwkt_reltoken(&hmp->fs_token);
 209                 vrecycle(ap->a_vp);
 210         }
 211         return(0);
 212 }
 213
 214 /*
 215  * Release the vnode association.  This is typically (but not always)
 216  * the last reference on the inode.
 217  *
 218  * Once the association is lost we are on our own with regards to
 219  * flushing the inode.
 220  *
 221  * We must interlock ip->vp so hammer_get_vnode() can avoid races.
 222  */
 223 int
 224 hammer_vop_reclaim(struct vop_reclaim_args *ap)
 225 {
 226         struct hammer_inode *ip;
 227         hammer_mount_t hmp;
 228         struct vnode *vp;
 229
 230         vp = ap->a_vp;
 231
 232         if ((ip = vp->v_data) != NULL) {
 233                 hmp = ip->hmp;
 234                 lwkt_gettoken(&hmp->fs_token);
 235                 hammer_lock_ex(&ip->lock);
 236                 vp->v_data = NULL;
 237                 ip->vp = NULL;
 238
 239                 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
 240                         ++hammer_count_reclaims;
 241                         ++hmp->count_reclaims;
 242                         ip->flags |= HAMMER_INODE_RECLAIM;
 243                 }
 244                 hammer_unlock(&ip->lock);
 245                 vclrisdirty(vp);
 246                 hammer_rel_inode(ip, 1);
 247                 lwkt_reltoken(&hmp->fs_token);
 248         }
 249         return(0);
 250 }
 251
 252 /*
 253  * Inform the kernel that the inode is dirty.  This will be checked
 254  * by vn_unlock().
 255  *
 256  * Theoretically in order to reclaim a vnode the hammer_vop_reclaim()
 257  * must be called which will interlock against our inode lock, so
 258  * if VRECLAIMED is not set vp->v_mount (as used by vsetisdirty())
 259  * should be stable without having to acquire any new locks.
 260  */
 261 void
 262 hammer_inode_dirty(struct hammer_inode *ip)
 263 {
 264         struct vnode *vp;
 265
 266         if ((ip->flags & HAMMER_INODE_MODMASK) &&
 267             (vp = ip->vp) != NULL &&
 268             (vp->v_flag & (VRECLAIMED | VISDIRTY)) == 0) {
 269                 vsetisdirty(vp);
 270         }
 271 }
 272
 273 /*
 274  * Return a locked vnode for the specified inode.  The inode must be
 275  * referenced but NOT LOCKED on entry and will remain referenced on
 276  * return.
 277  *
 278  * Called from the frontend.
 279  */
 280 int
 281 hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
 282 {
 283         hammer_mount_t hmp;
 284         struct vnode *vp;
 285         int error = 0;
 286         uint8_t obj_type;
 287
 288         hmp = ip->hmp;
 289
 290         for (;;) {
 291                 if ((vp = ip->vp) == NULL) {
 292                         error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
 293                         if (error)
 294                                 break;
 295                         hammer_lock_ex(&ip->lock);
 296                         if (ip->vp != NULL) {
 297                                 hammer_unlock(&ip->lock);
 298                                 vp = *vpp;
 299                                 vp->v_type = VBAD;
 300                                 vx_put(vp);
 301                                 continue;
 302                         }
 303                         hammer_ref(&ip->lock);
 304                         vp = *vpp;
 305                         ip->vp = vp;
 306
 307                         obj_type = ip->ino_data.obj_type;
 308                         vp->v_type = hammer_get_vnode_type(obj_type);
 309
 310                         hammer_inode_wakereclaims(ip);
 311
 312                         switch(ip->ino_data.obj_type) {
 313                         case HAMMER_OBJTYPE_CDEV:
 314                         case HAMMER_OBJTYPE_BDEV:
 315                                 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
 316                                 addaliasu(vp, ip->ino_data.rmajor,
 317                                           ip->ino_data.rminor);
 318                                 break;
 319                         case HAMMER_OBJTYPE_FIFO:
 320                                 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
 321                                 break;
 322                         case HAMMER_OBJTYPE_REGFILE:
 323                                 break;
 324                         default:
 325                                 break;
 326                         }
 327
 328                         /*
 329                          * Only mark as the root vnode if the ip is not
 330                          * historical, otherwise the VFS cache will get
 331                          * confused.  The other half of the special handling
 332                          * is in hammer_vop_nlookupdotdot().
 333                          *
 334                          * Pseudo-filesystem roots can be accessed via
 335                          * non-root filesystem paths and setting VROOT may
 336                          * confuse the namecache.  Set VPFSROOT instead.
 337                          */
 338                         if (ip->obj_id == HAMMER_OBJID_ROOT) {
 339                                 if (ip->obj_asof == hmp->asof) {
 340                                         if (ip->obj_localization == 0)
 341                                                 vsetflags(vp, VROOT);
 342                                         else
 343                                                 vsetflags(vp, VPFSROOT);
 344                                 } else {
 345                                         vsetflags(vp, VPFSROOT);
 346                                 }
 347                         }
 348
 349                         vp->v_data = (void *)ip;
 350                         /* vnode locked by getnewvnode() */
 351                         /* make related vnode dirty if inode dirty? */
 352                         hammer_unlock(&ip->lock);
 353                         if (vp->v_type == VREG) {
 354                                 vinitvmio(vp, ip->ino_data.size,
 355                                           hammer_blocksize(ip->ino_data.size),
 356                                           hammer_blockoff(ip->ino_data.size));
 357                         }
 358                         break;
 359                 }
 360
 361                 /*
 362                  * Interlock vnode clearing.  This does not prevent the
 363                  * vnode from going into a reclaimed state but it does
 364                  * prevent it from being destroyed or reused so the vget()
 365                  * will properly fail.
 366                  */
 367                 hammer_lock_ex(&ip->lock);
 368                 if ((vp = ip->vp) == NULL) {
 369                         hammer_unlock(&ip->lock);
 370                         continue;
 371                 }
 372                 vhold(vp);
 373                 hammer_unlock(&ip->lock);
 374
 375                 /*
 376                  * loop if the vget fails (aka races), or if the vp
 377                  * no longer matches ip->vp.
 378                  */
 379                 if (vget(vp, LK_EXCLUSIVE) == 0) {
 380                         if (vp == ip->vp) {
 381                                 vdrop(vp);
 382                                 break;
 383                         }
 384                         vput(vp);
 385                 }
 386                 vdrop(vp);
 387         }
 388         *vpp = vp;
 389         return(error);
 390 }
 391
 392 /*
 393  * Locate all copies of the inode for obj_id compatible with the specified
 394  * asof, reference, and issue the related call-back.  This routine is used
 395  * for direct-io invalidation and does not create any new inodes.
 396  */
 397 void
 398 hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
 399                             int (*callback)(hammer_inode_t ip, void *data),
 400                             void *data)
 401 {
 402         hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
 403                                    hammer_inode_info_cmp_all_history,
 404                                    callback, iinfo);
 405 }
 406
 407 /*
 408  * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
 409  * do not attach or detach the related vnode (use hammer_get_vnode() for
 410  * that).
 411  *
 412  * The flags argument is only applied for newly created inodes, and only
 413  * certain flags are inherited.
 414  *
 415  * Called from the frontend.
 416  */
 417 struct hammer_inode *
 418 hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
 419                  int64_t obj_id, hammer_tid_t asof, uint32_t localization,
 420                  int flags, int *errorp)
 421 {
 422         hammer_mount_t hmp = trans->hmp;
 423         struct hammer_node_cache *cachep;
 424         struct hammer_cursor cursor;
 425         struct hammer_inode *ip;
 426
 427
 428         /*
 429          * Determine if we already have an inode cached.  If we do then
 430          * we are golden.
 431          *
 432          * If we find an inode with no vnode we have to mark the
 433          * transaction such that hammer_inode_waitreclaims() is
 434          * called later on to avoid building up an infinite number
 435          * of inodes.  Otherwise we can continue to * add new inodes
 436          * faster then they can be disposed of, even with the tsleep
 437          * delay.
 438          *
 439          * If we find a dummy inode we return a failure so dounlink
 440          * (which does another lookup) doesn't try to mess with the
 441          * link count.  hammer_vop_nresolve() uses hammer_get_dummy_inode()
 442          * to ref dummy inodes.
 443          */
 444 loop:
 445         *errorp = 0;
 446         ip = __hammer_find_inode(trans, obj_id, asof, localization);
 447         if (ip) {
 448                 if (ip->flags & HAMMER_INODE_DUMMY) {
 449                         *errorp = ENOENT;
 450                         return(NULL);
 451                 }
 452                 hammer_ref(&ip->lock);
 453                 return(ip);
 454         }
 455
 456         /*
 457          * Allocate a new inode structure and deal with races later.
 458          */
 459         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
 460         ++hammer_count_inodes;
 461         ++hmp->count_inodes;
 462         ip->obj_id = obj_id;
 463         ip->obj_asof = asof;
 464         ip->obj_localization = localization;
 465         ip->hmp = hmp;
 466         ip->flags = flags & HAMMER_INODE_RO;
 467         ip->cache[0].ip = ip;
 468         ip->cache[1].ip = ip;
 469         ip->cache[2].ip = ip;
 470         ip->cache[3].ip = ip;
 471         if (hmp->ronly)
 472                 ip->flags |= HAMMER_INODE_RO;
 473         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
 474                 0x7FFFFFFFFFFFFFFFLL;
 475         RB_INIT(&ip->rec_tree);
 476         TAILQ_INIT(&ip->target_list);
 477         hammer_ref(&ip->lock);
 478
 479         /*
 480          * Locate the on-disk inode.  If this is a PFS root we always
 481          * access the current version of the root inode and (if it is not
 482          * a master) always access information under it with a snapshot
 483          * TID.
 484          *
 485          * We cache recent inode lookups in this directory in dip->cache[2].
 486          * If we can't find it we assume the inode we are looking for is
 487          * close to the directory inode.
 488          */
 489 retry:
 490         cachep = NULL;
 491         if (dip) {
 492                 if (dip->cache[2].node)
 493                         cachep = &dip->cache[2];
 494                 else
 495                         cachep = &dip->cache[0];
 496         }
 497         hammer_init_cursor(trans, &cursor, cachep, NULL);
 498         cursor.key_beg.localization = localization | HAMMER_LOCALIZE_INODE;
 499         cursor.key_beg.obj_id = ip->obj_id;
 500         cursor.key_beg.key = 0;
 501         cursor.key_beg.create_tid = 0;
 502         cursor.key_beg.delete_tid = 0;
 503         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
 504         cursor.key_beg.obj_type = 0;
 505
 506         cursor.asof = asof;
 507         cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
 508                        HAMMER_CURSOR_ASOF;
 509
 510         *errorp = hammer_btree_lookup(&cursor);
 511         if (*errorp == EDEADLK) {
 512                 hammer_done_cursor(&cursor);
 513                 goto retry;
 514         }
 515
 516         /*
 517          * On success the B-Tree lookup will hold the appropriate
 518          * buffer cache buffers and provide a pointer to the requested
 519          * information.  Copy the information to the in-memory inode
 520          * and cache the B-Tree node to improve future operations.
 521          */
 522         if (*errorp == 0) {
 523                 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
 524                 ip->ino_data = cursor.data->inode;
 525
 526                 /*
 527                  * cache[0] tries to cache the location of the object inode.
 528                  * The assumption is that it is near the directory inode.
 529                  *
 530                  * cache[1] tries to cache the location of the object data.
 531                  * We might have something in the governing directory from
 532                  * scan optimizations (see the strategy code in
 533                  * hammer_vnops.c).
 534                  *
 535                  * We update dip->cache[2], if possible, with the location
 536                  * of the object inode for future directory shortcuts.
 537                  */
 538                 hammer_cache_node(&ip->cache[0], cursor.node);
 539                 if (dip) {
 540                         if (dip->cache[3].node) {
 541                                 hammer_cache_node(&ip->cache[1],
 542                                                   dip->cache[3].node);
 543                         }
 544                         hammer_cache_node(&dip->cache[2], cursor.node);
 545                 }
 546
 547                 /*
 548                  * The file should not contain any data past the file size
 549                  * stored in the inode.  Setting save_trunc_off to the
 550                  * file size instead of max reduces B-Tree lookup overheads
 551                  * on append by allowing the flusher to avoid checking for
 552                  * record overwrites.
 553                  */
 554                 ip->save_trunc_off = ip->ino_data.size;
 555
 556                 /*
 557                  * Locate and assign the pseudofs management structure to
 558                  * the inode.
 559                  */
 560                 if (dip && dip->obj_localization == ip->obj_localization) {
 561                         ip->pfsm = dip->pfsm;
 562                         hammer_ref(&ip->pfsm->lock);
 563                 } else {
 564                         ip->pfsm = hammer_load_pseudofs(trans,
 565                                                         ip->obj_localization,
 566                                                         errorp);
 567                         *errorp = 0;    /* ignore ENOENT */
 568                 }
 569         }
 570
 571         /*
 572          * The inode is placed on the red-black tree and will be synced to
 573          * the media when flushed or by the filesystem sync.  If this races
 574          * another instantiation/lookup the insertion will fail.
 575          */
 576         if (*errorp == 0) {
 577                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 578                         hammer_free_inode(ip);
 579                         hammer_done_cursor(&cursor);
 580                         goto loop;
 581                 }
 582                 ip->flags |= HAMMER_INODE_ONDISK;
 583         } else {
 584                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
 585                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
 586                         --hmp->rsv_inodes;
 587                 }
 588
 589                 hammer_free_inode(ip);
 590                 ip = NULL;
 591         }
 592         hammer_done_cursor(&cursor);
 593
 594         /*
 595          * NEWINODE is only set if the inode becomes dirty later,
 596          * setting it here just leads to unnecessary stalls.
 597          *
 598          * trans->flags |= HAMMER_TRANSF_NEWINODE;
 599          */
 600         return (ip);
 601 }
 602
 603 /*
 604  * Get a dummy inode to placemark a broken directory entry.
 605  */
 606 struct hammer_inode *
 607 hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
 608                  int64_t obj_id, hammer_tid_t asof, uint32_t localization,
 609                  int flags, int *errorp)
 610 {
 611         hammer_mount_t hmp = trans->hmp;
 612         struct hammer_inode *ip;
 613
 614         /*
 615          * Determine if we already have an inode cached.  If we do then
 616          * we are golden.
 617          *
 618          * If we find an inode with no vnode we have to mark the
 619          * transaction such that hammer_inode_waitreclaims() is
 620          * called later on to avoid building up an infinite number
 621          * of inodes.  Otherwise we can continue to * add new inodes
 622          * faster then they can be disposed of, even with the tsleep
 623          * delay.
 624          *
 625          * If we find a non-fake inode we return an error.  Only fake
 626          * inodes can be returned by this routine.
 627          */
 628 loop:
 629         *errorp = 0;
 630         ip = __hammer_find_inode(trans, obj_id, asof, localization);
 631         if (ip) {
 632                 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
 633                         *errorp = ENOENT;
 634                         return(NULL);
 635                 }
 636                 hammer_ref(&ip->lock);
 637                 return(ip);
 638         }
 639
 640         /*
 641          * Allocate a new inode structure and deal with races later.
 642          */
 643         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
 644         ++hammer_count_inodes;
 645         ++hmp->count_inodes;
 646         ip->obj_id = obj_id;
 647         ip->obj_asof = asof;
 648         ip->obj_localization = localization;
 649         ip->hmp = hmp;
 650         ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
 651         ip->cache[0].ip = ip;
 652         ip->cache[1].ip = ip;
 653         ip->cache[2].ip = ip;
 654         ip->cache[3].ip = ip;
 655         ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
 656                 0x7FFFFFFFFFFFFFFFLL;
 657         RB_INIT(&ip->rec_tree);
 658         TAILQ_INIT(&ip->target_list);
 659         hammer_ref(&ip->lock);
 660
 661         /*
 662          * Populate the dummy inode.  Leave everything zero'd out.
 663          *
 664          * (ip->ino_leaf and ip->ino_data)
 665          *
 666          * Make the dummy inode a FIFO object which most copy programs
 667          * will properly ignore.
 668          */
 669         ip->save_trunc_off = ip->ino_data.size;
 670         ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
 671
 672         /*
 673          * Locate and assign the pseudofs management structure to
 674          * the inode.
 675          */
 676         if (dip && dip->obj_localization == ip->obj_localization) {
 677                 ip->pfsm = dip->pfsm;
 678                 hammer_ref(&ip->pfsm->lock);
 679         } else {
 680                 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
 681                                                 errorp);
 682                 *errorp = 0;    /* ignore ENOENT */
 683         }
 684
 685         /*
 686          * The inode is placed on the red-black tree and will be synced to
 687          * the media when flushed or by the filesystem sync.  If this races
 688          * another instantiation/lookup the insertion will fail.
 689          *
 690          * NOTE: Do not set HAMMER_INODE_ONDISK.  The inode is a fake.
 691          */
 692         if (*errorp == 0) {
 693                 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 694                         hammer_free_inode(ip);
 695                         goto loop;
 696                 }
 697         } else {
 698                 if (ip->flags & HAMMER_INODE_RSV_INODES) {
 699                         ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
 700                         --hmp->rsv_inodes;
 701                 }
 702                 hammer_free_inode(ip);
 703                 ip = NULL;
 704         }
 705         trans->flags |= HAMMER_TRANSF_NEWINODE;
 706         return (ip);
 707 }
 708
 709 /*
 710  * Return a referenced inode only if it is in our inode cache.
 711  * Dummy inodes do not count.
 712  */
 713 struct hammer_inode *
 714 hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
 715                   hammer_tid_t asof, uint32_t localization)
 716 {
 717         struct hammer_inode *ip;
 718
 719         ip = __hammer_find_inode(trans, obj_id, asof, localization);
 720         if (ip) {
 721                 if (ip->flags & HAMMER_INODE_DUMMY)
 722                         ip = NULL;
 723                 else
 724                         hammer_ref(&ip->lock);
 725         }
 726         return(ip);
 727 }
 728
 729 /*
 730  * Return a referenced inode only if it is in our inode cache.
 731  * This function does not reference inode.
 732  */
 733 static struct hammer_inode *
 734 __hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
 735                   hammer_tid_t asof, uint32_t localization)
 736 {
 737         hammer_mount_t hmp = trans->hmp;
 738         struct hammer_inode_info iinfo;
 739         struct hammer_inode *ip;
 740
 741         iinfo.obj_id = obj_id;
 742         iinfo.obj_asof = asof;
 743         iinfo.obj_localization = localization;
 744
 745         ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
 746
 747         return(ip);
 748 }
 749
 750 /*
 751  * Create a new filesystem object, returning the inode in *ipp.  The
 752  * returned inode will be referenced.  The inode is created in-memory.
 753  *
 754  * If pfsm is non-NULL the caller wishes to create the root inode for
 755  * a master PFS.
 756  */
 757 int
 758 hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
 759                     struct ucred *cred,
 760                     hammer_inode_t dip, const char *name, int namelen,
 761                     hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
 762 {
 763         hammer_mount_t hmp;
 764         hammer_inode_t ip;
 765         uid_t xuid;
 766         int error;
 767         int64_t namekey;
 768         uint32_t dummy;
 769
 770         hmp = trans->hmp;
 771
 772         /*
 773          * Disallow the creation of new inodes in directories which
 774          * have been deleted.  In HAMMER, this will cause a record
 775          * syncing assertion later on in the flush code.
 776          */
 777         if (dip && dip->ino_data.nlinks == 0) {
 778                 *ipp = NULL;
 779                 return (EINVAL);
 780         }
 781
 782         /*
 783          * Allocate inode
 784          */
 785         ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
 786         ++hammer_count_inodes;
 787         ++hmp->count_inodes;
 788         trans->flags |= HAMMER_TRANSF_NEWINODE;
 789
 790         if (pfsm) {
 791                 KKASSERT(pfsm->localization != 0);
 792                 ip->obj_id = HAMMER_OBJID_ROOT;
 793                 ip->obj_localization = pfsm->localization;
 794         } else {
 795                 KKASSERT(dip != NULL);
 796                 namekey = hammer_directory_namekey(dip, name, namelen, &dummy);
 797                 ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
 798                 ip->obj_localization = dip->obj_localization;
 799         }
 800
 801         KKASSERT(ip->obj_id != 0);
 802         ip->obj_asof = hmp->asof;
 803         ip->hmp = hmp;
 804         ip->flush_state = HAMMER_FST_IDLE;
 805         ip->flags = HAMMER_INODE_DDIRTY |
 806                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
 807         ip->cache[0].ip = ip;
 808         ip->cache[1].ip = ip;
 809         ip->cache[2].ip = ip;
 810         ip->cache[3].ip = ip;
 811
 812         ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
 813         /* ip->save_trunc_off = 0; (already zero) */
 814         RB_INIT(&ip->rec_tree);
 815         TAILQ_INIT(&ip->target_list);
 816
 817         ip->ino_data.atime = trans->time;
 818         ip->ino_data.mtime = trans->time;
 819         ip->ino_data.size = 0;
 820         ip->ino_data.nlinks = 0;
 821
 822         /*
 823          * A nohistory designator on the parent directory is inherited by
 824          * the child.  We will do this even for pseudo-fs creation... the
 825          * sysad can turn it off.
 826          */
 827         if (dip) {
 828                 ip->ino_data.uflags = dip->ino_data.uflags &
 829                                       (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
 830         }
 831
 832         ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 833         ip->ino_leaf.base.localization = ip->obj_localization |
 834                                          HAMMER_LOCALIZE_INODE;
 835         ip->ino_leaf.base.obj_id = ip->obj_id;
 836         ip->ino_leaf.base.key = 0;
 837         ip->ino_leaf.base.create_tid = 0;
 838         ip->ino_leaf.base.delete_tid = 0;
 839         ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
 840         ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
 841
 842         ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
 843         ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
 844         ip->ino_data.mode = vap->va_mode;
 845         ip->ino_data.ctime = trans->time;
 846
 847         /*
 848          * If we are running version 2 or greater directory entries are
 849          * inode-localized instead of data-localized.
 850          */
 851         if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
 852                 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
 853                         ip->ino_data.cap_flags |=
 854                                 HAMMER_INODE_CAP_DIR_LOCAL_INO;
 855                 }
 856         }
 857         if (trans->hmp->version >= HAMMER_VOL_VERSION_SIX) {
 858                 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
 859                         ip->ino_data.cap_flags |=
 860                                 HAMMER_INODE_CAP_DIRHASH_ALG1;
 861                 }
 862         }
 863
 864         /*
 865          * Setup the ".." pointer.  This only needs to be done for directories
 866          * but we do it for all objects as a recovery aid if dip exists.
 867          * The inode is probably a PFS root if dip is NULL.
 868          */
 869         if (dip)
 870                 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
 871 #if 0
 872         /*
 873          * The parent_obj_localization field only applies to pseudo-fs roots.
 874          * XXX this is no longer applicable, PFSs are no longer directly
 875          * tied into the parent's directory structure.
 876          */
 877         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
 878             ip->obj_id == HAMMER_OBJID_ROOT) {
 879                 ip->ino_data.ext.obj.parent_obj_localization =
 880                                                 dip->obj_localization;
 881         }
 882 #endif
 883
 884         switch(ip->ino_leaf.base.obj_type) {
 885         case HAMMER_OBJTYPE_CDEV:
 886         case HAMMER_OBJTYPE_BDEV:
 887                 ip->ino_data.rmajor = vap->va_rmajor;
 888                 ip->ino_data.rminor = vap->va_rminor;
 889                 break;
 890         default:
 891                 break;
 892         }
 893
 894         /*
 895          * Calculate default uid/gid and overwrite with information from
 896          * the vap.
 897          */
 898         if (dip) {
 899                 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
 900                 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
 901                                              xuid, cred, &vap->va_mode);
 902         } else {
 903                 xuid = 0;
 904         }
 905         ip->ino_data.mode = vap->va_mode;
 906
 907         if (vap->va_vaflags & VA_UID_UUID_VALID)
 908                 ip->ino_data.uid = vap->va_uid_uuid;
 909         else if (vap->va_uid != (uid_t)VNOVAL)
 910                 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
 911         else
 912                 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
 913
 914         if (vap->va_vaflags & VA_GID_UUID_VALID)
 915                 ip->ino_data.gid = vap->va_gid_uuid;
 916         else if (vap->va_gid != (gid_t)VNOVAL)
 917                 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
 918         else if (dip)
 919                 ip->ino_data.gid = dip->ino_data.gid;
 920
 921         hammer_ref(&ip->lock);
 922
 923         if (pfsm) {
 924                 ip->pfsm = pfsm;
 925                 hammer_ref(&pfsm->lock);
 926                 error = 0;
 927         } else if (dip->obj_localization == ip->obj_localization) {
 928                 ip->pfsm = dip->pfsm;
 929                 hammer_ref(&ip->pfsm->lock);
 930                 error = 0;
 931         } else {
 932                 ip->pfsm = hammer_load_pseudofs(trans,
 933                                                 ip->obj_localization,
 934                                                 &error);
 935                 error = 0;      /* ignore ENOENT */
 936         }
 937
 938         if (error) {
 939                 hammer_free_inode(ip);
 940                 ip = NULL;
 941         } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
 942                 hpanic("duplicate obj_id %llx", (long long)ip->obj_id);
 943                 /* not reached */
 944                 hammer_free_inode(ip);
 945         }
 946         *ipp = ip;
 947         return(error);
 948 }
 949
 950 /*
 951  * Final cleanup / freeing of an inode structure
 952  */
 953 static void
 954 hammer_free_inode(hammer_inode_t ip)
 955 {
 956         struct hammer_mount *hmp;
 957
 958         hmp = ip->hmp;
 959         KKASSERT(hammer_oneref(&ip->lock));
 960         hammer_uncache_node(&ip->cache[0]);
 961         hammer_uncache_node(&ip->cache[1]);
 962         hammer_uncache_node(&ip->cache[2]);
 963         hammer_uncache_node(&ip->cache[3]);
 964         hammer_inode_wakereclaims(ip);
 965         if (ip->objid_cache)
 966                 hammer_clear_objid(ip);
 967         --hammer_count_inodes;
 968         --hmp->count_inodes;
 969         if (ip->pfsm) {
 970                 hammer_rel_pseudofs(hmp, ip->pfsm);
 971                 ip->pfsm = NULL;
 972         }
 973         kfree(ip, hmp->m_inodes);
 974 }
 975
 976 /*
 977  * Retrieve pseudo-fs data.  NULL will never be returned.
 978  *
 979  * If an error occurs *errorp will be set and a default template is returned,
 980  * otherwise *errorp is set to 0.  Typically when an error occurs it will
 981  * be ENOENT.
 982  */
 983 hammer_pseudofs_inmem_t
 984 hammer_load_pseudofs(hammer_transaction_t trans,
 985                      uint32_t localization, int *errorp)
 986 {
 987         hammer_mount_t hmp = trans->hmp;
 988         hammer_inode_t ip;
 989         hammer_pseudofs_inmem_t pfsm;
 990         struct hammer_cursor cursor;
 991         int bytes;
 992
 993 retry:
 994         pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
 995         if (pfsm) {
 996                 hammer_ref(&pfsm->lock);
 997                 *errorp = 0;
 998                 return(pfsm);
 999         }
1000
1001         /*
1002          * PFS records are associated with the root inode (not the PFS root
1003          * inode, but the real root).  Avoid an infinite recursion if loading
1004          * the PFS for the real root.
1005          */
1006         if (localization) {
1007                 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
1008                                       HAMMER_MAX_TID,
1009                                       HAMMER_DEF_LOCALIZATION, 0, errorp);
1010         } else {
1011                 ip = NULL;
1012         }
1013
1014         pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
1015         pfsm->localization = localization;
1016         pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
1017         pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
1018
1019         hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
1020         cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION |
1021                                       HAMMER_LOCALIZE_MISC;
1022         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
1023         cursor.key_beg.create_tid = 0;
1024         cursor.key_beg.delete_tid = 0;
1025         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
1026         cursor.key_beg.obj_type = 0;
1027         cursor.key_beg.key = localization;
1028         cursor.asof = HAMMER_MAX_TID;
1029         cursor.flags |= HAMMER_CURSOR_ASOF;
1030
1031         if (ip)
1032                 *errorp = hammer_ip_lookup(&cursor);
1033         else
1034                 *errorp = hammer_btree_lookup(&cursor);
1035         if (*errorp == 0) {
1036                 *errorp = hammer_ip_resolve_data(&cursor);
1037                 if (*errorp == 0) {
1038                         if (cursor.data->pfsd.mirror_flags &
1039                             HAMMER_PFSD_DELETED) {
1040                                 *errorp = ENOENT;
1041                         } else {
1042                                 bytes = cursor.leaf->data_len;
1043                                 if (bytes > sizeof(pfsm->pfsd))
1044                                         bytes = sizeof(pfsm->pfsd);
1045                                 bcopy(cursor.data, &pfsm->pfsd, bytes);
1046                         }
1047                 }
1048         }
1049         hammer_done_cursor(&cursor);
1050
1051         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1052         hammer_ref(&pfsm->lock);
1053         if (ip)
1054                 hammer_rel_inode(ip, 0);
1055         if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
1056                 kfree(pfsm, hmp->m_misc);
1057                 goto retry;
1058         }
1059         return(pfsm);
1060 }
1061
1062 /*
1063  * Store pseudo-fs data.  The backend will automatically delete any prior
1064  * on-disk pseudo-fs data but we have to delete in-memory versions.
1065  */
1066 int
1067 hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
1068 {
1069         struct hammer_cursor cursor;
1070         hammer_record_t record;
1071         hammer_inode_t ip;
1072         int error;
1073
1074         /*
1075          * PFS records are associated with the root inode (not the PFS root
1076          * inode, but the real root).
1077          */
1078         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1079                               HAMMER_DEF_LOCALIZATION, 0, &error);
1080 retry:
1081         pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1082         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
1083         cursor.key_beg.localization = ip->obj_localization |
1084                                       HAMMER_LOCALIZE_MISC;
1085         cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
1086         cursor.key_beg.create_tid = 0;
1087         cursor.key_beg.delete_tid = 0;
1088         cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
1089         cursor.key_beg.obj_type = 0;
1090         cursor.key_beg.key = pfsm->localization;
1091         cursor.asof = HAMMER_MAX_TID;
1092         cursor.flags |= HAMMER_CURSOR_ASOF;
1093
1094         /*
1095          * Replace any in-memory version of the record.
1096          */
1097         error = hammer_ip_lookup(&cursor);
1098         if (error == 0 && hammer_cursor_inmem(&cursor)) {
1099                 record = cursor.iprec;
1100                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
1101                         KKASSERT(cursor.deadlk_rec == NULL);
1102                         hammer_ref(&record->lock);
1103                         cursor.deadlk_rec = record;
1104                         error = EDEADLK;
1105                 } else {
1106                         record->flags |= HAMMER_RECF_DELETED_FE;
1107                         error = 0;
1108                 }
1109         }
1110
1111         /*
1112          * Allocate replacement general record.  The backend flush will
1113          * delete any on-disk version of the record.
1114          */
1115         if (error == 0 || error == ENOENT) {
1116                 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1117                 record->type = HAMMER_MEM_RECORD_GENERAL;
1118
1119                 record->leaf.base.localization = ip->obj_localization |
1120                                                  HAMMER_LOCALIZE_MISC;
1121                 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1122                 record->leaf.base.key = pfsm->localization;
1123                 record->leaf.data_len = sizeof(pfsm->pfsd);
1124                 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1125                 error = hammer_ip_add_record(trans, record);
1126         }
1127         hammer_done_cursor(&cursor);
1128         if (error == EDEADLK)
1129                 goto retry;
1130         hammer_rel_inode(ip, 0);
1131         return(error);
1132 }
1133
1134 /*
1135  * Create a root directory for a PFS if one does not alredy exist.
1136  *
1137  * The PFS root stands alone so we must also bump the nlinks count
1138  * to prevent it from being destroyed on release.
1139  */
1140 int
1141 hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1142                        hammer_pseudofs_inmem_t pfsm)
1143 {
1144         hammer_inode_t ip;
1145         struct vattr vap;
1146         int error;
1147
1148         ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1149                               pfsm->localization, 0, &error);
1150         if (ip == NULL) {
1151                 vattr_null(&vap);
1152                 vap.va_mode = 0755;
1153                 vap.va_type = VDIR;
1154                 error = hammer_create_inode(trans, &vap, cred,
1155                                             NULL, NULL, 0,
1156                                             pfsm, &ip);
1157                 if (error == 0) {
1158                         ++ip->ino_data.nlinks;
1159                         hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
1160                 }
1161         }
1162         if (ip)
1163                 hammer_rel_inode(ip, 0);
1164         return(error);
1165 }
1166
1167 /*
1168  * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1169  * if we are unable to disassociate all the inodes.
1170  */
1171 static
1172 int
1173 hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1174 {
1175         int res;
1176
1177         hammer_ref(&ip->lock);
1178         if (ip->vp && (ip->vp->v_flag & VPFSROOT)) {
1179                 /*
1180                  * The hammer pfs-upgrade directive itself might have the
1181                  * root of the pfs open.  Just allow it.
1182                  */
1183                 res = 0;
1184         } else {
1185                 /*
1186                  * Don't allow any subdirectories or files to be open.
1187                  */
1188                 if (hammer_isactive(&ip->lock) == 2 && ip->vp)
1189                         vclean_unlocked(ip->vp);
1190                 if (hammer_isactive(&ip->lock) == 1 && ip->vp == NULL)
1191                         res = 0;
1192                 else
1193                         res = -1;       /* stop, someone is using the inode */
1194         }
1195         hammer_rel_inode(ip, 0);
1196         return(res);
1197 }
1198
1199 int
1200 hammer_unload_pseudofs(hammer_transaction_t trans, uint32_t localization)
1201 {
1202         int res;
1203         int try;
1204
1205         for (try = res = 0; try < 4; ++try) {
1206                 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1207                                            hammer_inode_pfs_cmp,
1208                                            hammer_unload_pseudofs_callback,
1209                                            &localization);
1210                 if (res == 0 && try > 1)
1211                         break;
1212                 hammer_flusher_sync(trans->hmp);
1213         }
1214         if (res != 0)
1215                 res = ENOTEMPTY;
1216         return(res);
1217 }
1218
1219
1220 /*
1221  * Release a reference on a PFS
1222  */
1223 void
1224 hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1225 {
1226         hammer_rel(&pfsm->lock);
1227         if (hammer_norefs(&pfsm->lock)) {
1228                 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
1229                 kfree(pfsm, hmp->m_misc);
1230         }
1231 }
1232
1233 /*
1234  * Called by hammer_sync_inode().
1235  */
1236 static int
1237 hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
1238 {
1239         hammer_transaction_t trans = cursor->trans;
1240         hammer_record_t record;
1241         int error;
1242         int redirty;
1243
1244 retry:
1245         error = 0;
1246
1247         /*
1248          * If the inode has a presence on-disk then locate it and mark
1249          * it deleted, setting DELONDISK.
1250          *
1251          * The record may or may not be physically deleted, depending on
1252          * the retention policy.
1253          */
1254         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1255             HAMMER_INODE_ONDISK) {
1256                 hammer_normalize_cursor(cursor);
1257                 cursor->key_beg.localization = ip->obj_localization |
1258                                                HAMMER_LOCALIZE_INODE;
1259                 cursor->key_beg.obj_id = ip->obj_id;
1260                 cursor->key_beg.key = 0;
1261                 cursor->key_beg.create_tid = 0;
1262                 cursor->key_beg.delete_tid = 0;
1263                 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1264                 cursor->key_beg.obj_type = 0;
1265                 cursor->asof = ip->obj_asof;
1266                 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1267                 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
1268                 cursor->flags |= HAMMER_CURSOR_BACKEND;
1269
1270                 error = hammer_btree_lookup(cursor);
1271                 if (hammer_debug_inode)
1272                         hdkprintf("IPDEL %p %08x %d", ip, ip->flags, error);
1273
1274                 if (error == 0) {
1275                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1276                         if (hammer_debug_inode)
1277                                 hdkprintf("error %d\n", error);
1278                         if (error == 0) {
1279                                 ip->flags |= HAMMER_INODE_DELONDISK;
1280                         }
1281                         if (cursor->node)
1282                                 hammer_cache_node(&ip->cache[0], cursor->node);
1283                 }
1284                 if (error == EDEADLK) {
1285                         hammer_done_cursor(cursor);
1286                         error = hammer_init_cursor(trans, cursor,
1287                                                    &ip->cache[0], ip);
1288                         if (hammer_debug_inode)
1289                                 hdkprintf("IPDED %p %d\n", ip, error);
1290                         if (error == 0)
1291                                 goto retry;
1292                 }
1293         }
1294
1295         /*
1296          * Ok, write out the initial record or a new record (after deleting
1297          * the old one), unless the DELETED flag is set.  This routine will
1298          * clear DELONDISK if it writes out a record.
1299          *
1300          * Update our inode statistics if this is the first application of
1301          * the inode on-disk.
1302          */
1303         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1304                 /*
1305                  * Generate a record and write it to the media.  We clean-up
1306                  * the state before releasing so we do not have to set-up
1307                  * a flush_group.
1308                  */
1309                 record = hammer_alloc_mem_record(ip, 0);
1310                 record->type = HAMMER_MEM_RECORD_INODE;
1311                 record->flush_state = HAMMER_FST_FLUSH;
1312                 record->leaf = ip->sync_ino_leaf;
1313                 record->leaf.base.create_tid = trans->tid;
1314                 record->leaf.data_len = sizeof(ip->sync_ino_data);
1315                 record->leaf.create_ts = trans->time32;
1316                 record->data = (void *)&ip->sync_ino_data;
1317                 record->flags |= HAMMER_RECF_INTERLOCK_BE;
1318
1319                 /*
1320                  * If this flag is set we cannot sync the new file size
1321                  * because we haven't finished related truncations.  The
1322                  * inode will be flushed in another flush group to finish
1323                  * the job.
1324                  */
1325                 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1326                     ip->sync_ino_data.size != ip->ino_data.size) {
1327                         redirty = 1;
1328                         ip->sync_ino_data.size = ip->ino_data.size;
1329                 } else {
1330                         redirty = 0;
1331                 }
1332
1333                 for (;;) {
1334                         error = hammer_ip_sync_record_cursor(cursor, record);
1335                         if (hammer_debug_inode)
1336                                 hdkprintf("GENREC %p rec %08x %d\n",
1337                                         ip, record->flags, error);
1338                         if (error != EDEADLK)
1339                                 break;
1340                         hammer_done_cursor(cursor);
1341                         error = hammer_init_cursor(trans, cursor,
1342                                                    &ip->cache[0], ip);
1343                         if (hammer_debug_inode)
1344                                 hdkprintf("GENREC reinit %d\n", error);
1345                         if (error)
1346                                 break;
1347                 }
1348
1349                 /*
1350                  * Note:  The record was never on the inode's record tree
1351                  * so just wave our hands importantly and destroy it.
1352                  */
1353                 record->flags |= HAMMER_RECF_COMMITTED;
1354                 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1355                 record->flush_state = HAMMER_FST_IDLE;
1356                 ++ip->rec_generation;
1357                 hammer_rel_mem_record(record);
1358
1359                 /*
1360                  * Finish up.
1361                  */
1362                 if (error == 0) {
1363                         if (hammer_debug_inode)
1364                                 hdkprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
1365                         ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1366                                             HAMMER_INODE_SDIRTY |
1367                                             HAMMER_INODE_ATIME |
1368                                             HAMMER_INODE_MTIME);
1369                         ip->flags &= ~HAMMER_INODE_DELONDISK;
1370                         if (redirty)
1371                                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1372
1373                         /*
1374                          * Root volume count of inodes
1375                          */
1376                         hammer_sync_lock_sh(trans);
1377                         if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
1378                                 hammer_modify_volume_field(trans,
1379                                                            trans->rootvol,
1380                                                            vol0_stat_inodes);
1381                                 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
1382                                 hammer_modify_volume_done(trans->rootvol);
1383                                 ip->flags |= HAMMER_INODE_ONDISK;
1384                                 if (hammer_debug_inode)
1385                                         hdkprintf("NOWONDISK %p\n", ip);
1386                         }
1387                         hammer_sync_unlock(trans);
1388                 }
1389         }
1390
1391         /*
1392          * If the inode has been destroyed, clean out any left-over flags
1393          * that may have been set by the frontend.
1394          */
1395         if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
1396                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
1397                                     HAMMER_INODE_SDIRTY |
1398                                     HAMMER_INODE_ATIME |
1399                                     HAMMER_INODE_MTIME);
1400         }
1401         return(error);
1402 }
1403
1404 /*
1405  * Update only the itimes fields.
1406  *
1407  * ATIME can be updated without generating any UNDO.  MTIME is updated
1408  * with UNDO so it is guaranteed to be synchronized properly in case of
1409  * a crash.
1410  *
1411  * Neither field is included in the B-Tree leaf element's CRC, which is how
1412  * we can get away with updating ATIME the way we do.
1413  */
1414 static int
1415 hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
1416 {
1417         hammer_transaction_t trans = cursor->trans;
1418         int error;
1419
1420 retry:
1421         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
1422             HAMMER_INODE_ONDISK) {
1423                 return(0);
1424         }
1425
1426         hammer_normalize_cursor(cursor);
1427         cursor->key_beg.localization = ip->obj_localization |
1428                                        HAMMER_LOCALIZE_INODE;
1429         cursor->key_beg.obj_id = ip->obj_id;
1430         cursor->key_beg.key = 0;
1431         cursor->key_beg.create_tid = 0;
1432         cursor->key_beg.delete_tid = 0;
1433         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1434         cursor->key_beg.obj_type = 0;
1435         cursor->asof = ip->obj_asof;
1436         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1437         cursor->flags |= HAMMER_CURSOR_ASOF;
1438         cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1439         cursor->flags |= HAMMER_CURSOR_GET_DATA;
1440         cursor->flags |= HAMMER_CURSOR_BACKEND;
1441
1442         error = hammer_btree_lookup(cursor);
1443         if (error == 0) {
1444                 hammer_cache_node(&ip->cache[0], cursor->node);
1445                 if (ip->sync_flags & HAMMER_INODE_MTIME) {
1446                         /*
1447                          * Updating MTIME requires an UNDO.  Just cover
1448                          * both atime and mtime.
1449                          */
1450                         hammer_sync_lock_sh(trans);
1451                         hammer_modify_buffer(trans, cursor->data_buffer,
1452                                 &cursor->data->inode.mtime,
1453                                 sizeof(cursor->data->inode.atime) +
1454                                 sizeof(cursor->data->inode.mtime));
1455                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1456                         cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1457                         hammer_modify_buffer_done(cursor->data_buffer);
1458                         hammer_sync_unlock(trans);
1459                 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1460                         /*
1461                          * Updating atime only can be done in-place with
1462                          * no UNDO.
1463                          */
1464                         hammer_sync_lock_sh(trans);
1465                         hammer_modify_buffer_noundo(trans, cursor->data_buffer);
1466                         cursor->data->inode.atime = ip->sync_ino_data.atime;
1467                         hammer_modify_buffer_done(cursor->data_buffer);
1468                         hammer_sync_unlock(trans);
1469                 }
1470                 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1471         }
1472         if (error == EDEADLK) {
1473                 hammer_done_cursor(cursor);
1474                 error = hammer_init_cursor(trans, cursor, &ip->cache[0], ip);
1475                 if (error == 0)
1476                         goto retry;
1477         }
1478         return(error);
1479 }
1480
1481 /*
1482  * Release a reference on an inode, flush as requested.
1483  *
1484  * On the last reference we queue the inode to the flusher for its final
1485  * disposition.
1486  */
1487 void
1488 hammer_rel_inode(struct hammer_inode *ip, int flush)
1489 {
1490         /*
1491          * Handle disposition when dropping the last ref.
1492          */
1493         for (;;) {
1494                 if (hammer_oneref(&ip->lock)) {
1495                         /*
1496                          * Determine whether on-disk action is needed for
1497                          * the inode's final disposition.
1498                          */
1499                         KKASSERT(ip->vp == NULL);
1500                         hammer_inode_unloadable_check(ip, 0);
1501                         if (ip->flags & HAMMER_INODE_MODMASK) {
1502                                 hammer_flush_inode(ip, 0);
1503                         } else if (hammer_oneref(&ip->lock)) {
1504                                 hammer_unload_inode(ip);
1505                                 break;
1506                         }
1507                 } else {
1508                         if (flush)
1509                                 hammer_flush_inode(ip, 0);
1510
1511                         /*
1512                          * The inode still has multiple refs, try to drop
1513                          * one ref.
1514                          */
1515                         KKASSERT(hammer_isactive(&ip->lock) >= 1);
1516                         if (hammer_isactive(&ip->lock) > 1) {
1517                                 hammer_rel(&ip->lock);
1518                                 break;
1519                         }
1520                 }
1521         }
1522 }
1523
1524 /*
1525  * Unload and destroy the specified inode.  Must be called with one remaining
1526  * reference.  The reference is disposed of.
1527  *
1528  * The inode must be completely clean.
1529  */
1530 static int
1531 hammer_unload_inode(struct hammer_inode *ip)
1532 {
1533         hammer_mount_t hmp = ip->hmp;
1534
1535         KASSERT(hammer_oneref(&ip->lock),
1536                 ("hammer_unload_inode: %d refs", hammer_isactive(&ip->lock)));
1537         KKASSERT(ip->vp == NULL);
1538         KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1539         KKASSERT(ip->cursor_ip_refs == 0);
1540         KKASSERT(hammer_notlocked(&ip->lock));
1541         KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1542
1543         KKASSERT(RB_EMPTY(&ip->rec_tree));
1544         KKASSERT(TAILQ_EMPTY(&ip->target_list));
1545
1546         if (ip->flags & HAMMER_INODE_RDIRTY) {
1547                 RB_REMOVE(hammer_redo_rb_tree, &hmp->rb_redo_root, ip);
1548                 ip->flags &= ~HAMMER_INODE_RDIRTY;
1549         }
1550         RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
1551
1552         hammer_free_inode(ip);
1553         return(0);
1554 }
1555
1556 /*
1557  * Called during unmounting if a critical error occured.  The in-memory
1558  * inode and all related structures are destroyed.
1559  *
1560  * If a critical error did not occur the unmount code calls the standard
1561  * release and asserts that the inode is gone.
1562  */
1563 int
1564 hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1565 {
1566         hammer_record_t rec;
1567
1568         /*
1569          * Get rid of the inodes in-memory records, regardless of their
1570          * state, and clear the mod-mask.
1571          */
1572         while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1573                 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1574                 rec->target_ip = NULL;
1575                 if (rec->flush_state == HAMMER_FST_SETUP)
1576                         rec->flush_state = HAMMER_FST_IDLE;
1577         }
1578         while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1579                 if (rec->flush_state == HAMMER_FST_FLUSH)
1580                         --rec->flush_group->refs;
1581                 else
1582                         hammer_ref(&rec->lock);
1583                 KKASSERT(hammer_oneref(&rec->lock));
1584                 rec->flush_state = HAMMER_FST_IDLE;
1585                 rec->flush_group = NULL;
1586                 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1587                 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1588                 ++ip->rec_generation;
1589                 hammer_rel_mem_record(rec);
1590         }
1591         ip->flags &= ~HAMMER_INODE_MODMASK;
1592         ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1593         KKASSERT(ip->vp == NULL);
1594
1595         /*
1596          * Remove the inode from any flush group, force it idle.  FLUSH
1597          * and SETUP states have an inode ref.
1598          */
1599         switch(ip->flush_state) {
1600         case HAMMER_FST_FLUSH:
1601                 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
1602                 --ip->flush_group->refs;
1603                 ip->flush_group = NULL;
1604                 /* fall through */
1605         case HAMMER_FST_SETUP:
1606                 hammer_rel(&ip->lock);
1607                 ip->flush_state = HAMMER_FST_IDLE;
1608                 /* fall through */
1609         case HAMMER_FST_IDLE:
1610                 break;
1611         }
1612
1613         /*
1614          * There shouldn't be any associated vnode.  The unload needs at
1615          * least one ref, if we do have a vp steal its ip ref.
1616          */
1617         if (ip->vp) {
1618                 hdkprintf("Unexpected vnode association ip %p vp %p\n",
1619                         ip, ip->vp);
1620                 ip->vp->v_data = NULL;
1621                 ip->vp = NULL;
1622         } else {
1623                 hammer_ref(&ip->lock);
1624         }
1625         hammer_unload_inode(ip);
1626         return(0);
1627 }
1628
1629 /*
1630  * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
1631  * the read-only flag for cached inodes.
1632  *
1633  * This routine is called from a RB_SCAN().
1634  */
1635 int
1636 hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1637 {
1638         hammer_mount_t hmp = ip->hmp;
1639
1640         if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1641                 ip->flags |= HAMMER_INODE_RO;
1642         else
1643                 ip->flags &= ~HAMMER_INODE_RO;
1644         return(0);
1645 }
1646
1647 /*
1648  * A transaction has modified an inode, requiring updates as specified by
1649  * the passed flags.
1650  *
1651  * HAMMER_INODE_DDIRTY: Inode data has been updated, not incl mtime/atime,
1652  *                      and not including size changes due to write-append
1653  *                      (but other size changes are included).
1654  * HAMMER_INODE_SDIRTY: Inode data has been updated, size changes due to
1655  *                      write-append.
1656  * HAMMER_INODE_XDIRTY: Dirty in-memory records
1657  * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
1658  * HAMMER_INODE_DELETED: Inode record/data must be deleted
1659  * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
1660  */
1661 void
1662 hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
1663 {
1664         /*
1665          * ronly of 0 or 2 does not trigger assertion.
1666          * 2 is a special error state
1667          */
1668         KKASSERT(ip->hmp->ronly != 1 ||
1669                   (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
1670                             HAMMER_INODE_SDIRTY |
1671                             HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1672                             HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
1673         if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1674                 ip->flags |= HAMMER_INODE_RSV_INODES;
1675                 ++ip->hmp->rsv_inodes;
1676         }
1677
1678         /*
1679          * Set the NEWINODE flag in the transaction if the inode
1680          * transitions to a dirty state.  This is used to track
1681          * the load on the inode cache.
1682          */
1683         if (trans &&
1684             (ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1685             (flags & HAMMER_INODE_MODMASK)) {
1686                 trans->flags |= HAMMER_TRANSF_NEWINODE;
1687         }
1688         if (flags & HAMMER_INODE_MODMASK)
1689                 hammer_inode_dirty(ip);
1690         ip->flags |= flags;
1691 }
1692
1693 /*
1694  * Attempt to quickly update the atime for a hammer inode.  Return 0 on
1695  * success, -1 on failure.
1696  *
1697  * We attempt to update the atime with only the ip lock and not the
1698  * whole filesystem lock in order to improve concurrency.  We can only
1699  * do this safely if the ATIME flag is already pending on the inode.
1700  *
1701  * This function is called via a vnops path (ip pointer is stable) without
1702  * fs_token held.
1703  */
1704 int
1705 hammer_update_atime_quick(hammer_inode_t ip)
1706 {
1707         struct timeval tv;
1708         int res = -1;
1709
1710         if ((ip->flags & HAMMER_INODE_RO) ||
1711             (ip->hmp->mp->mnt_flag & MNT_NOATIME)) {
1712                 /*
1713                  * Silently indicate success on read-only mount/snap
1714                  */
1715                 res = 0;
1716         } else if (ip->flags & HAMMER_INODE_ATIME) {
1717                 /*
1718                  * Double check with inode lock held against backend.  This
1719                  * is only safe if all we need to do is update
1720                  * ino_data.atime.
1721                  */
1722                 getmicrotime(&tv);
1723                 hammer_lock_ex(&ip->lock);
1724                 if (ip->flags & HAMMER_INODE_ATIME) {
1725                         ip->ino_data.atime =
1726                             (unsigned long)tv.tv_sec * 1000000ULL + tv.tv_usec;
1727                         res = 0;
1728                 }
1729                 hammer_unlock(&ip->lock);
1730         }
1731         return res;
1732 }
1733
1734 /*
1735  * Request that an inode be flushed.  This whole mess cannot block and may
1736  * recurse (if not synchronous).  Once requested HAMMER will attempt to
1737  * actively flush the inode until the flush can be done.
1738  *
1739  * The inode may already be flushing, or may be in a setup state.  We can
1740  * place the inode in a flushing state if it is currently idle and flag it
1741  * to reflush if it is currently flushing.
1742  *
1743  * Upon return if the inode could not be flushed due to a setup
1744  * dependancy, then it will be automatically flushed when the dependancy
1745  * is satisfied.
1746  */
1747 void
1748 hammer_flush_inode(hammer_inode_t ip, int flags)
1749 {
1750         hammer_mount_t hmp;
1751         hammer_flush_group_t flg;
1752         int good;
1753
1754         /*
1755          * fill_flush_group is the first flush group we may be able to
1756          * continue filling, it may be open or closed but it will always
1757          * be past the currently flushing (running) flg.
1758          *
1759          * next_flush_group is the next open flush group.
1760          */
1761         hmp = ip->hmp;
1762         while ((flg = hmp->fill_flush_group) != NULL) {
1763                 KKASSERT(flg->running == 0);
1764                 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit &&
1765                     flg->total_count <= hammer_autoflush) {
1766                         break;
1767                 }
1768                 hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
1769                 hammer_flusher_async(ip->hmp, flg);
1770         }
1771         if (flg == NULL) {
1772                 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
1773                 flg->seq = hmp->flusher.next++;
1774                 if (hmp->next_flush_group == NULL)
1775                         hmp->next_flush_group = flg;
1776                 if (hmp->fill_flush_group == NULL)
1777                         hmp->fill_flush_group = flg;
1778                 RB_INIT(&flg->flush_tree);
1779                 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1780         }
1781
1782         /*
1783          * Trivial 'nothing to flush' case.  If the inode is in a SETUP
1784          * state we have to put it back into an IDLE state so we can
1785          * drop the extra ref.
1786          *
1787          * If we have a parent dependancy we must still fall through
1788          * so we can run it.
1789          */
1790         if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
1791                 if (ip->flush_state == HAMMER_FST_SETUP &&
1792                     TAILQ_EMPTY(&ip->target_list)) {
1793                         ip->flush_state = HAMMER_FST_IDLE;
1794                         hammer_rel_inode(ip, 0);
1795                 }
1796                 if (ip->flush_state == HAMMER_FST_IDLE)
1797                         return;
1798         }
1799
1800         /*
1801          * Our flush action will depend on the current state.
1802          */
1803         switch(ip->flush_state) {
1804         case HAMMER_FST_IDLE:
1805                 /*
1806                  * We have no dependancies and can flush immediately.  Some
1807                  * our children may not be flushable so we have to re-test
1808                  * with that additional knowledge.
1809                  */
1810                 hammer_flush_inode_core(ip, flg, flags);
1811                 break;
1812         case HAMMER_FST_SETUP:
1813                 /*
1814                  * Recurse upwards through dependancies via target_list
1815                  * and start their flusher actions going if possible.
1816                  *
1817                  * 'good' is our connectivity.  -1 means we have none and
1818                  * can't flush, 0 means there weren't any dependancies, and
1819                  * 1 means we have good connectivity.
1820                  */
1821                 good = hammer_setup_parent_inodes(ip, 0, flg);
1822
1823                 if (good >= 0) {
1824                         /*
1825                          * We can continue if good >= 0.  Determine how
1826                          * many records under our inode can be flushed (and
1827                          * mark them).
1828                          */
1829                         hammer_flush_inode_core(ip, flg, flags);
1830                 } else {
1831                         /*
1832                          * Parent has no connectivity, tell it to flush
1833                          * us as soon as it does.
1834                          *
1835                          * The REFLUSH flag is also needed to trigger
1836                          * dependancy wakeups.
1837                          */
1838                         ip->flags |= HAMMER_INODE_CONN_DOWN |
1839                                      HAMMER_INODE_REFLUSH;
1840                         if (flags & HAMMER_FLUSH_SIGNAL) {
1841                                 ip->flags |= HAMMER_INODE_RESIGNAL;
1842                                 hammer_flusher_async(ip->hmp, flg);
1843                         }
1844                 }
1845                 break;
1846         case HAMMER_FST_FLUSH:
1847                 /*
1848                  * We are already flushing, flag the inode to reflush
1849                  * if needed after it completes its current flush.
1850                  *
1851                  * The REFLUSH flag is also needed to trigger
1852                  * dependancy wakeups.
1853                  */
1854                 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1855                         ip->flags |= HAMMER_INODE_REFLUSH;
1856                 if (flags & HAMMER_FLUSH_SIGNAL) {
1857                         ip->flags |= HAMMER_INODE_RESIGNAL;
1858                         hammer_flusher_async(ip->hmp, flg);
1859                 }
1860                 break;
1861         }
1862 }
1863
1864 /*
1865  * Scan ip->target_list, which is a list of records owned by PARENTS to our
1866  * ip which reference our ip.
1867  *
1868  * XXX This is a huge mess of recursive code, but not one bit of it blocks
1869  *     so for now do not ref/deref the structures.  Note that if we use the
1870  *     ref/rel code later, the rel CAN block.
1871  */
1872 static int
1873 hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1874                            hammer_flush_group_t flg)
1875 {
1876         hammer_record_t depend;
1877         int good;
1878         int r;
1879
1880         /*
1881          * If we hit our recursion limit and we have parent dependencies
1882          * We cannot continue.  Returning < 0 will cause us to be flagged
1883          * for reflush.  Returning -2 cuts off additional dependency checks
1884          * because they are likely to also hit the depth limit.
1885          *
1886          * We cannot return < 0 if there are no dependencies or there might
1887          * not be anything to wakeup (ip).
1888          */
1889         if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1890                 if (hammer_debug_general & 0x10000)
1891                         hkrateprintf(&hammer_gen_krate,
1892                             "Warning: depth limit reached on "
1893                             "setup recursion, inode %p %016llx\n",
1894                             ip, (long long)ip->obj_id);
1895                 return(-2);
1896         }
1897
1898         /*
1899          * Scan dependencies
1900          */
1901         good = 0;
1902         TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
1903                 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
1904                 KKASSERT(depend->target_ip == ip);
1905                 if (r < 0 && good == 0)
1906                         good = -1;
1907                 if (r > 0)
1908                         good = 1;
1909
1910                 /*
1911                  * If we failed due to the recursion depth limit then stop
1912                  * now.
1913                  */
1914                 if (r == -2)
1915                         break;
1916         }
1917         return(good);
1918 }
1919
1920 /*
1921  * This helper function takes a record representing the dependancy between
1922  * the parent inode and child inode.
1923  *
1924  * record               = record in question (*rec in below)
1925  * record->ip           = parent inode (*pip in below)
1926  * record->target_ip    = child inode (*ip in below)
1927  *
1928  * *pip--------------\
1929  *    ^               \rec_tree
1930  *     \               \
1931  *      \ip            /\\\\\ rbtree of recs from parent inode's view
1932  *       \            //\\\\\\
1933  *        \          / ........
1934  *         \        /
1935  *          \------*rec------target_ip------>*ip
1936  *               ...target_entry<----...----->target_list<---...
1937  *                                            list of recs from inode's view
1938  *
1939  * We are asked to recurse upwards and convert the record from SETUP
1940  * to FLUSH if possible.
1941  *
1942  * Return 1 if the record gives us connectivity
1943  *
1944  * Return 0 if the record is not relevant
1945  *
1946  * Return -1 if we can't resolve the dependancy and there is no connectivity.
1947  */
1948 static int
1949 hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
1950                                   hammer_flush_group_t flg)
1951 {
1952         hammer_inode_t pip;
1953         int good;
1954
1955         KKASSERT(record->flush_state != HAMMER_FST_IDLE);
1956         pip = record->ip;
1957
1958         /*
1959          * If the record is already flushing, is it in our flush group?
1960          *
1961          * If it is in our flush group but it is a general record or a
1962          * delete-on-disk, it does not improve our connectivity (return 0),
1963          * and if the target inode is not trying to destroy itself we can't
1964          * allow the operation yet anyway (the second return -1).
1965          */
1966         if (record->flush_state == HAMMER_FST_FLUSH) {
1967                 /*
1968                  * If not in our flush group ask the parent to reflush
1969                  * us as soon as possible.
1970                  */
1971                 if (record->flush_group != flg) {
1972                         pip->flags |= HAMMER_INODE_REFLUSH;
1973                         record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1974                         return(-1);
1975                 }
1976
1977                 /*
1978                  * If in our flush group everything is already set up,
1979                  * just return whether the record will improve our
1980                  * visibility or not.
1981                  */
1982                 if (record->type == HAMMER_MEM_RECORD_ADD)
1983                         return(1);
1984                 return(0);
1985         }
1986
1987         /*
1988          * It must be a setup record.  Try to resolve the setup dependancies
1989          * by recursing upwards so we can place ip on the flush list.
1990          *
1991          * Limit ourselves to 20 levels of recursion to avoid blowing out
1992          * the kernel stack.  If we hit the recursion limit we can't flush
1993          * until the parent flushes.  The parent will flush independantly
1994          * on its own and ultimately a deep recursion will be resolved.
1995          */
1996         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1997
1998         good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1999
2000         /*
2001          * If good < 0 the parent has no connectivity and we cannot safely
2002          * flush the directory entry, which also means we can't flush our
2003          * ip.  Flag us for downward recursion once the parent's
2004          * connectivity is resolved.  Flag the parent for [re]flush or it
2005          * may not check for downward recursions.
2006          */
2007         if (good < 0) {
2008                 pip->flags |= HAMMER_INODE_REFLUSH;
2009                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
2010                 return(good);
2011         }
2012
2013         /*
2014          * We are go, place the parent inode in a flushing state so we can
2015          * place its record in a flushing state.  Note that the parent
2016          * may already be flushing.  The record must be in the same flush
2017          * group as the parent.
2018          */
2019         if (pip->flush_state != HAMMER_FST_FLUSH)
2020                 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
2021         KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
2022
2023         /*
2024          * It is possible for a rename to create a loop in the recursion
2025          * and revisit a record.  This will result in the record being
2026          * placed in a flush state unexpectedly.  This check deals with
2027          * the case.
2028          */
2029         if (record->flush_state == HAMMER_FST_FLUSH) {
2030                 if (record->type == HAMMER_MEM_RECORD_ADD)
2031                         return(1);
2032                 return(0);
2033         }
2034
2035         KKASSERT(record->flush_state == HAMMER_FST_SETUP);
2036
2037 #if 0
2038         if (record->type == HAMMER_MEM_RECORD_DEL &&
2039             (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
2040                 /*
2041                  * Regardless of flushing state we cannot sync this path if the
2042                  * record represents a delete-on-disk but the target inode
2043                  * is not ready to sync its own deletion.
2044                  *
2045                  * XXX need to count effective nlinks to determine whether
2046                  * the flush is ok, otherwise removing a hardlink will
2047                  * just leave the DEL record to rot.
2048                  */
2049                 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
2050                 return(-1);
2051         } else
2052 #endif
2053         if (pip->flush_group == flg) {
2054                 /*
2055                  * Because we have not calculated nlinks yet we can just
2056                  * set records to the flush state if the parent is in
2057                  * the same flush group as we are.
2058                  */
2059                 record->flush_state = HAMMER_FST_FLUSH;
2060                 record->flush_group = flg;
2061                 ++record->flush_group->refs;
2062                 hammer_ref(&record->lock);
2063
2064                 /*
2065                  * A general directory-add contributes to our visibility.
2066                  *
2067                  * Otherwise it is probably a directory-delete or
2068                  * delete-on-disk record and does not contribute to our
2069                  * visbility (but we can still flush it).
2070                  */
2071                 if (record->type == HAMMER_MEM_RECORD_ADD)
2072                         return(1);
2073                 return(0);
2074         } else {
2075                 /*
2076                  * If the parent is not in our flush group we cannot
2077                  * flush this record yet, there is no visibility.
2078                  * We tell the parent to reflush and mark ourselves
2079                  * so the parent knows it should flush us too.
2080                  */
2081                 pip->flags |= HAMMER_INODE_REFLUSH;
2082                 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
2083                 return(-1);
2084         }
2085 }
2086
2087 /*
2088  * This is the core routine placing an inode into the FST_FLUSH state.
2089  */
2090 static void
2091 hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
2092 {
2093         hammer_mount_t hmp = ip->hmp;
2094         int go_count;
2095
2096         /*
2097          * Set flush state and prevent the flusher from cycling into
2098          * the next flush group.  Do not place the ip on the list yet.
2099          * Inodes not in the idle state get an extra reference.
2100          */
2101         KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
2102         if (ip->flush_state == HAMMER_FST_IDLE)
2103                 hammer_ref(&ip->lock);
2104         ip->flush_state = HAMMER_FST_FLUSH;
2105         ip->flush_group = flg;
2106         ++hmp->flusher.group_lock;
2107         ++hmp->count_iqueued;
2108         ++hammer_count_iqueued;
2109         ++flg->total_count;
2110         hammer_redo_fifo_start_flush(ip);
2111
2112 #if 0
2113         /*
2114          * We need to be able to vfsync/truncate from the backend.
2115          *
2116          * XXX Any truncation from the backend will acquire the vnode
2117          *     independently.
2118          */
2119         KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
2120         if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
2121                 ip->flags |= HAMMER_INODE_VHELD;
2122                 vref(ip->vp);
2123         }
2124 #endif
2125
2126         /*
2127          * Figure out how many in-memory records we can actually flush
2128          * (not including inode meta-data, buffers, etc).
2129          */
2130         KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
2131         if (flags & HAMMER_FLUSH_RECURSION) {
2132                 /*
2133                  * If this is a upwards recursion we do not want to
2134                  * recurse down again!
2135                  */
2136                 go_count = 1;
2137 #if 0
2138         } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2139                 /*
2140                  * No new records are added if we must complete a flush
2141                  * from a previous cycle, but we do have to move the records
2142                  * from the previous cycle to the current one.
2143                  */
2144 #if 0
2145                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2146                                    hammer_syncgrp_child_callback, NULL);
2147 #endif
2148                 go_count = 1;
2149 #endif
2150         } else {
2151                 /*
2152                  * Normal flush, scan records and bring them into the flush.
2153                  * Directory adds and deletes are usually skipped (they are
2154                  * grouped with the related inode rather then with the
2155                  * directory).
2156                  *
2157                  * go_count can be negative, which means the scan aborted
2158                  * due to the flush group being over-full and we should
2159                  * flush what we have.
2160                  */
2161                 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2162                                    hammer_setup_child_callback, NULL);
2163         }
2164
2165         /*
2166          * This is a more involved test that includes go_count.  If we
2167          * can't flush, flag the inode and return.  If go_count is 0 we
2168          * were are unable to flush any records in our rec_tree and
2169          * must ignore the XDIRTY flag.
2170          */
2171         if (go_count == 0) {
2172                 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
2173                         --hmp->count_iqueued;
2174                         --hammer_count_iqueued;
2175
2176                         --flg->total_count;
2177                         ip->flush_state = HAMMER_FST_SETUP;
2178                         ip->flush_group = NULL;
2179                         if (flags & HAMMER_FLUSH_SIGNAL) {
2180                                 ip->flags |= HAMMER_INODE_REFLUSH |
2181                                              HAMMER_INODE_RESIGNAL;
2182                         } else {
2183                                 ip->flags |= HAMMER_INODE_REFLUSH;
2184                         }
2185 #if 0
2186                         if (ip->flags & HAMMER_INODE_VHELD) {
2187                                 ip->flags &= ~HAMMER_INODE_VHELD;
2188                                 vrele(ip->vp);
2189                         }
2190 #endif
2191
2192                         /*
2193                          * REFLUSH is needed to trigger dependancy wakeups
2194                          * when an inode is in SETUP.
2195                          */
2196                         ip->flags |= HAMMER_INODE_REFLUSH;
2197                         if (--hmp->flusher.group_lock == 0)
2198                                 wakeup(&hmp->flusher.group_lock);
2199                         return;
2200                 }
2201         }
2202
2203         /*
2204          * Snapshot the state of the inode for the backend flusher.
2205          *
2206          * We continue to retain save_trunc_off even when all truncations
2207          * have been resolved as an optimization to determine if we can
2208          * skip the B-Tree lookup for overwrite deletions.
2209          *
2210          * NOTE: The DELETING flag is a mod flag, but it is also sticky,
2211          * and stays in ip->flags.  Once set, it stays set until the
2212          * inode is destroyed.
2213          */
2214         if (ip->flags & HAMMER_INODE_TRUNCATED) {
2215                 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2216                 ip->sync_trunc_off = ip->trunc_off;
2217                 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2218                 ip->flags &= ~HAMMER_INODE_TRUNCATED;
2219                 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
2220
2221                 /*
2222                  * The save_trunc_off used to cache whether the B-Tree
2223                  * holds any records past that point is not used until
2224                  * after the truncation has succeeded, so we can safely
2225                  * set it now.
2226                  */
2227                 if (ip->save_trunc_off > ip->sync_trunc_off)
2228                         ip->save_trunc_off = ip->sync_trunc_off;
2229         }
2230         ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2231                            ~HAMMER_INODE_TRUNCATED);
2232         ip->sync_ino_leaf = ip->ino_leaf;
2233         ip->sync_ino_data = ip->ino_data;
2234         ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
2235
2236         /*
2237          * The flusher list inherits our inode and reference.
2238          */
2239         KKASSERT(flg->running == 0);
2240         RB_INSERT(hammer_fls_rb_tree, &flg->flush_tree, ip);
2241         if (--hmp->flusher.group_lock == 0)
2242                 wakeup(&hmp->flusher.group_lock);
2243
2244         /*
2245          * Auto-flush the group if it grows too large.  Make sure the
2246          * inode reclaim wait pipeline continues to work.
2247          */
2248         if (flg->total_count >= hammer_autoflush ||
2249             flg->total_count >= hammer_limit_reclaims / 4) {
2250                 if (hmp->fill_flush_group == flg)
2251                         hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
2252                 hammer_flusher_async(hmp, flg);
2253         }
2254 }
2255
2256 /*
2257  * Callback for scan of ip->rec_tree.  Try to include each record in our
2258  * flush.  ip->flush_group has been set but the inode has not yet been
2259  * moved into a flushing state.
2260  *
2261  * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2262  * both inodes.
2263  *
2264  * We return 1 for any record placed or found in FST_FLUSH, which prevents
2265  * the caller from shortcutting the flush.
2266  */
2267 static int
2268 hammer_setup_child_callback(hammer_record_t rec, void *data)
2269 {
2270         hammer_flush_group_t flg;
2271         hammer_inode_t target_ip;
2272         hammer_inode_t ip;
2273         int r;
2274
2275         /*
2276          * Records deleted or committed by the backend are ignored.
2277          * Note that the flush detects deleted frontend records at
2278          * multiple points to deal with races.  This is just the first
2279          * line of defense.  The only time HAMMER_RECF_DELETED_FE cannot
2280          * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2281          * messes up link-count calculations.
2282          *
2283          * NOTE: Don't get confused between record deletion and, say,
2284          * directory entry deletion.  The deletion of a directory entry
2285          * which is on-media has nothing to do with the record deletion
2286          * flags.
2287          */
2288         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2289                           HAMMER_RECF_COMMITTED)) {
2290                 if (rec->flush_state == HAMMER_FST_FLUSH) {
2291                         KKASSERT(rec->flush_group == rec->ip->flush_group);
2292                         r = 1;
2293                 } else {
2294                         r = 0;
2295                 }
2296                 return(r);
2297         }
2298
2299         /*
2300          * If the record is in an idle state it has no dependancies and
2301          * can be flushed.
2302          */
2303         ip = rec->ip;
2304         flg = ip->flush_group;
2305         r = 0;
2306
2307         switch(rec->flush_state) {
2308         case HAMMER_FST_IDLE:
2309                 /*
2310                  * The record has no setup dependancy, we can flush it.
2311                  */
2312                 KKASSERT(rec->target_ip == NULL);
2313                 rec->flush_state = HAMMER_FST_FLUSH;
2314                 rec->flush_group = flg;
2315                 ++flg->refs;
2316                 hammer_ref(&rec->lock);
2317                 r = 1;
2318                 break;
2319         case HAMMER_FST_SETUP:
2320                 /*
2321                  * The record has a setup dependancy.  These are typically
2322                  * directory entry adds and deletes.  Such entries will be
2323                  * flushed when their inodes are flushed so we do not
2324                  * usually have to add them to the flush here.  However,
2325                  * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2326                  * it is asking us to flush this record (and it).
2327                  */
2328                 target_ip = rec->target_ip;
2329                 KKASSERT(target_ip != NULL);
2330                 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
2331
2332                 /*
2333                  * If the target IP is already flushing in our group
2334                  * we could associate the record, but target_ip has
2335                  * already synced ino_data to sync_ino_data and we
2336                  * would also have to adjust nlinks.   Plus there are
2337                  * ordering issues for adds and deletes.
2338                  *
2339                  * Reflush downward if this is an ADD, and upward if
2340                  * this is a DEL.
2341                  */
2342                 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
2343                         if (rec->type == HAMMER_MEM_RECORD_ADD)
2344                                 ip->flags |= HAMMER_INODE_REFLUSH;
2345                         else
2346                                 target_ip->flags |= HAMMER_INODE_REFLUSH;
2347                         break;
2348                 }
2349
2350                 /*
2351                  * Target IP is not yet flushing.  This can get complex
2352                  * because we have to be careful about the recursion.
2353                  *
2354                  * Directories create an issue for us in that if a flush
2355                  * of a directory is requested the expectation is to flush
2356                  * any pending directory entries, but this will cause the
2357                  * related inodes to recursively flush as well.  We can't
2358                  * really defer the operation so just get as many as we
2359                  * can and
2360                  */
2361 #if 0
2362                 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
2363                     (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
2364                         /*
2365                          * We aren't reclaiming and the target ip was not
2366                          * previously prevented from flushing due to this
2367                          * record dependancy.  Do not flush this record.
2368                          */
2369                         /*r = 0;*/
2370                 } else
2371 #endif
2372                 if (flg->total_count + flg->refs >
2373                            ip->hmp->undo_rec_limit) {
2374                         /*
2375                          * Our flush group is over-full and we risk blowing
2376                          * out the UNDO FIFO.  Stop the scan, flush what we
2377                          * have, then reflush the directory.
2378                          *
2379                          * The directory may be forced through multiple
2380                          * flush groups before it can be completely
2381                          * flushed.
2382                          */
2383                         ip->flags |= HAMMER_INODE_RESIGNAL |
2384                                      HAMMER_INODE_REFLUSH;
2385                         r = -1;
2386                 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2387                         /*
2388                          * If the target IP is not flushing we can force
2389                          * it to flush, even if it is unable to write out
2390                          * any of its own records we have at least one in
2391                          * hand that we CAN deal with.
2392                          */
2393                         rec->flush_state = HAMMER_FST_FLUSH;
2394                         rec->flush_group = flg;
2395                         ++flg->refs;
2396                         hammer_ref(&rec->lock);
2397                         hammer_flush_inode_core(target_ip, flg,
2398                                                 HAMMER_FLUSH_RECURSION);
2399                         r = 1;
2400                 } else {
2401                         /*
2402                          * General or delete-on-disk record.
2403                          *
2404                          * XXX this needs help.  If a delete-on-disk we could
2405                          * disconnect the target.  If the target has its own
2406                          * dependancies they really need to be flushed.
2407                          *
2408                          * XXX
2409                          */
2410                         rec->flush_state = HAMMER_FST_FLUSH;
2411                         rec->flush_group = flg;
2412                         ++flg->refs;
2413                         hammer_ref(&rec->lock);
2414                         hammer_flush_inode_core(target_ip, flg,
2415                                                 HAMMER_FLUSH_RECURSION);
2416                         r = 1;
2417                 }
2418                 break;
2419         case HAMMER_FST_FLUSH:
2420                 /*
2421                  * The record could be part of a previous flush group if the
2422                  * inode is a directory (the record being a directory entry).
2423                  * Once the flush group was closed a hammer_test_inode()
2424                  * function can cause a new flush group to be setup, placing
2425                  * the directory inode itself in a new flush group.
2426                  *
2427                  * When associated with a previous flush group we count it
2428                  * as if it were in our current flush group, since it will
2429                  * effectively be flushed by the time we flush our current
2430                  * flush group.
2431                  */
2432                 KKASSERT(
2433                     rec->ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY ||
2434                     rec->flush_group == flg);
2435                 r = 1;
2436                 break;
2437         }
2438         return(r);
2439 }
2440
2441 #if 0
2442 /*
2443  * This version just moves records already in a flush state to the new
2444  * flush group and that is it.
2445  */
2446 static int
2447 hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2448 {
2449         hammer_inode_t ip = rec->ip;
2450
2451         switch(rec->flush_state) {
2452         case HAMMER_FST_FLUSH:
2453                 KKASSERT(rec->flush_group == ip->flush_group);
2454                 break;
2455         default:
2456                 break;
2457         }
2458         return(0);
2459 }
2460 #endif
2461
2462 /*
2463  * Wait for a previously queued flush to complete.
2464  *
2465  * If a critical error occured we don't try to wait.
2466  */
2467 void
2468 hammer_wait_inode(hammer_inode_t ip)
2469 {
2470         /*
2471          * The inode can be in a SETUP state in which case RESIGNAL
2472          * should be set.  If RESIGNAL is not set then the previous
2473          * flush completed and a later operation placed the inode
2474          * in a passive setup state again, so we're done.
2475          *
2476          * The inode can be in a FLUSH state in which case we
2477          * can just wait for completion.
2478          */
2479         while (ip->flush_state == HAMMER_FST_FLUSH ||
2480             (ip->flush_state == HAMMER_FST_SETUP &&
2481              (ip->flags & HAMMER_INODE_RESIGNAL))) {
2482                 /*
2483                  * Don't try to flush on a critical error
2484                  */
2485                 if (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
2486                         break;
2487
2488                 /*
2489                  * If the inode was already being flushed its flg
2490                  * may not have been queued to the backend.  We
2491                  * have to make sure it gets queued or we can wind
2492                  * up blocked or deadlocked (particularly if we are
2493                  * the vnlru thread).
2494                  */
2495                 if (ip->flush_state == HAMMER_FST_FLUSH) {
2496                         KKASSERT(ip->flush_group);
2497                         if (ip->flush_group->closed == 0) {
2498                                 if (hammer_debug_inode) {
2499                                         hkprintf("debug: forcing "
2500                                                 "async flush ip %016jx\n",
2501                                                 (intmax_t)ip->obj_id);
2502                                 }
2503                                 hammer_flusher_async(ip->hmp, ip->flush_group);
2504                                 continue; /* retest */
2505                         }
2506                 }
2507
2508                 /*
2509                  * In a flush state with the flg queued to the backend
2510                  * or in a setup state with RESIGNAL set, we can safely
2511                  * wait.
2512                  */
2513                 ip->flags |= HAMMER_INODE_FLUSHW;
2514                 tsleep(&ip->flags, 0, "hmrwin", 0);
2515         }
2516
2517 #if 0
2518         /*
2519          * The inode may have been in a passive setup state,
2520          * call flush to make sure we get signaled.
2521          */
2522         if (ip->flush_state == HAMMER_FST_SETUP)
2523                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2524 #endif
2525
2526 }
2527
2528 /*
2529  * Called by the backend code when a flush has been completed.
2530  * The inode has already been removed from the flush list.
2531  *
2532  * A pipelined flush can occur, in which case we must re-enter the
2533  * inode on the list and re-copy its fields.
2534  */
2535 void
2536 hammer_flush_inode_done(hammer_inode_t ip, int error)
2537 {
2538         hammer_mount_t hmp;
2539         int dorel;
2540
2541         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2542
2543         hmp = ip->hmp;
2544
2545         /*
2546          * Auto-reflush if the backend could not completely flush
2547          * the inode.  This fixes a case where a deferred buffer flush
2548          * could cause fsync to return early.
2549          */
2550         if (ip->sync_flags & HAMMER_INODE_MODMASK)
2551                 ip->flags |= HAMMER_INODE_REFLUSH;
2552
2553         /*
2554          * Merge left-over flags back into the frontend and fix the state.
2555          * Incomplete truncations are retained by the backend.
2556          */
2557         ip->error = error;
2558         ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2559         ip->sync_flags &= HAMMER_INODE_TRUNCATED;
2560
2561         /*
2562          * The backend may have adjusted nlinks, so if the adjusted nlinks
2563          * does not match the fronttend set the frontend's DDIRTY flag again.
2564          */
2565         if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2566                 ip->flags |= HAMMER_INODE_DDIRTY;
2567
2568         /*
2569          * Fix up the dirty buffer status.
2570          */
2571         if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
2572                 ip->flags |= HAMMER_INODE_BUFS;
2573         }
2574         hammer_redo_fifo_end_flush(ip);
2575
2576         /*
2577          * Re-set the XDIRTY flag if some of the inode's in-memory records
2578          * could not be flushed.
2579          */
2580         KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2581                   (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2582                  (!RB_EMPTY(&ip->rec_tree) &&
2583                   (ip->flags & HAMMER_INODE_XDIRTY) != 0));
2584
2585         /*
2586          * Do not lose track of inodes which no longer have vnode
2587          * assocations, otherwise they may never get flushed again.
2588          *
2589          * The reflush flag can be set superfluously, causing extra pain
2590          * for no reason.  If the inode is no longer modified it no longer
2591          * needs to be flushed.
2592          */
2593         if (ip->flags & HAMMER_INODE_MODMASK) {
2594                 if (ip->vp == NULL)
2595                         ip->flags |= HAMMER_INODE_REFLUSH;
2596         } else {
2597                 ip->flags &= ~HAMMER_INODE_REFLUSH;
2598         }
2599
2600         /*
2601          * The fs token is held but the inode lock is not held.  Because this
2602          * is a backend flush it is possible that the vnode has no references
2603          * and cause a reclaim race inside vsetisdirty() if/when it blocks.
2604          *
2605          * Therefore, we must lock the inode around this particular dirtying
2606          * operation.  We don't have to around other dirtying operations
2607          * where the vnode is implicitly or explicitly held.
2608          */
2609         if (ip->flags & HAMMER_INODE_MODMASK) {
2610                 hammer_lock_ex(&ip->lock);
2611                 hammer_inode_dirty(ip);
2612                 hammer_unlock(&ip->lock);
2613         }
2614
2615         /*
2616          * Adjust the flush state.
2617          */
2618         if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
2619                 /*
2620                  * We were unable to flush out all our records, leave the
2621                  * inode in a flush state and in the current flush group.
2622                  * The flush group will be re-run.
2623                  *
2624                  * This occurs if the UNDO block gets too full or there is
2625                  * too much dirty meta-data and allows the flusher to
2626                  * finalize the UNDO block and then re-flush.
2627                  */
2628                 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
2629                 dorel = 0;
2630         } else {
2631                 /*
2632                  * Remove from the flush_group
2633                  */
2634                 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
2635                 ip->flush_group = NULL;
2636
2637 #if 0
2638                 /*
2639                  * Clean up the vnode ref and tracking counts.
2640                  */
2641                 if (ip->flags & HAMMER_INODE_VHELD) {
2642                         ip->flags &= ~HAMMER_INODE_VHELD;
2643                         vrele(ip->vp);
2644                 }
2645 #endif
2646                 --hmp->count_iqueued;
2647                 --hammer_count_iqueued;
2648
2649                 /*
2650                  * And adjust the state.
2651                  */
2652                 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2653                         ip->flush_state = HAMMER_FST_IDLE;
2654                         dorel = 1;
2655                 } else {
2656                         ip->flush_state = HAMMER_FST_SETUP;
2657                         dorel = 0;
2658                 }
2659
2660                 /*
2661                  * If the frontend is waiting for a flush to complete,
2662                  * wake it up.
2663                  */
2664                 if (ip->flags & HAMMER_INODE_FLUSHW) {
2665                         ip->flags &= ~HAMMER_INODE_FLUSHW;
2666                         wakeup(&ip->flags);
2667                 }
2668
2669                 /*
2670                  * If the frontend made more changes and requested another
2671                  * flush, then try to get it running.
2672                  *
2673                  * Reflushes are aborted when the inode is errored out.
2674                  */
2675                 if (ip->flags & HAMMER_INODE_REFLUSH) {
2676                         ip->flags &= ~HAMMER_INODE_REFLUSH;
2677                         if (ip->flags & HAMMER_INODE_RESIGNAL) {
2678                                 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2679                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2680                         } else {
2681                                 hammer_flush_inode(ip, 0);
2682                         }
2683                 }
2684         }
2685
2686         /*
2687          * If we have no parent dependancies we can clear CONN_DOWN
2688          */
2689         if (TAILQ_EMPTY(&ip->target_list))
2690                 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2691
2692         /*
2693          * If the inode is now clean drop the space reservation.
2694          */
2695         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2696             (ip->flags & HAMMER_INODE_RSV_INODES)) {
2697                 ip->flags &= ~HAMMER_INODE_RSV_INODES;
2698                 --hmp->rsv_inodes;
2699         }
2700
2701         ip->flags &= ~HAMMER_INODE_SLAVEFLUSH;
2702
2703         if (dorel)
2704                 hammer_rel_inode(ip, 0);
2705 }
2706
2707 /*
2708  * Called from hammer_sync_inode() to synchronize in-memory records
2709  * to the media.
2710  */
2711 static int
2712 hammer_sync_record_callback(hammer_record_t record, void *data)
2713 {
2714         hammer_cursor_t cursor = data;
2715         hammer_transaction_t trans = cursor->trans;
2716         hammer_mount_t hmp = trans->hmp;
2717         int error;
2718
2719         /*
2720          * Skip records that do not belong to the current flush.
2721          */
2722         ++hammer_stats_record_iterations;
2723         if (record->flush_state != HAMMER_FST_FLUSH)
2724                 return(0);
2725
2726         if (record->flush_group != record->ip->flush_group) {
2727                 hdkprintf("rec %p ip %p bad flush group %p %p\n",
2728                         record,
2729                         record->ip,
2730                         record->flush_group,
2731                         record->ip->flush_group);
2732                 if (hammer_debug_critical)
2733                         Debugger("blah2");
2734                 return(0);
2735         }
2736         KKASSERT(record->flush_group == record->ip->flush_group);
2737
2738         /*
2739          * Interlock the record using the BE flag.  Once BE is set the
2740          * frontend cannot change the state of FE.
2741          *
2742          * NOTE: If FE is set prior to us setting BE we still sync the
2743          * record out, but the flush completion code converts it to
2744          * a delete-on-disk record instead of destroying it.
2745          */
2746         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
2747         record->flags |= HAMMER_RECF_INTERLOCK_BE;
2748
2749         /*
2750          * The backend has already disposed of the record.
2751          */
2752         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
2753                 error = 0;
2754                 goto done;
2755         }
2756
2757         /*
2758          * If the whole inode is being deleted and all on-disk records will
2759          * be deleted very soon, we can't sync any new records to disk
2760          * because they will be deleted in the same transaction they were
2761          * created in (delete_tid == create_tid), which will assert.
2762          *
2763          * XXX There may be a case with RECORD_ADD with DELETED_FE set
2764          * that we currently panic on.
2765          */
2766         if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
2767                 switch(record->type) {
2768                 case HAMMER_MEM_RECORD_DATA:
2769                         /*
2770                          * We don't have to do anything, if the record was
2771                          * committed the space will have been accounted for
2772                          * in the blockmap.
2773                          */
2774                         /* fall through */
2775                 case HAMMER_MEM_RECORD_GENERAL:
2776                         /*
2777                          * Set deleted-by-backend flag.  Do not set the
2778                          * backend committed flag, because we are throwing
2779                          * the record away.
2780                          */
2781                         record->flags |= HAMMER_RECF_DELETED_BE;
2782                         ++record->ip->rec_generation;
2783                         error = 0;
2784                         goto done;
2785                 case HAMMER_MEM_RECORD_ADD:
2786                         hpanic("illegal add during inode deletion record %p",
2787                                 record);
2788                         break; /* NOT REACHED */
2789                 case HAMMER_MEM_RECORD_INODE:
2790                         hpanic("attempt to sync inode record %p?", record);
2791                         break; /* NOT REACHED */
2792                 case HAMMER_MEM_RECORD_DEL:
2793                         /*
2794                          * Follow through and issue the on-disk deletion
2795                          */
2796                         break;
2797                 }
2798         }
2799
2800         /*
2801          * If DELETED_FE is set special handling is needed for directory
2802          * entries.  Dependant pieces related to the directory entry may
2803          * have already been synced to disk.  If this occurs we have to
2804          * sync the directory entry and then change the in-memory record
2805          * from an ADD to a DELETE to cover the fact that it's been
2806          * deleted by the frontend.
2807          *
2808          * A directory delete covering record (MEM_RECORD_DEL) can never
2809          * be deleted by the frontend.
2810          *
2811          * Any other record type (aka DATA) can be deleted by the frontend.
2812          * XXX At the moment the flusher must skip it because there may
2813          * be another data record in the flush group for the same block,
2814          * meaning that some frontend data changes can leak into the backend's
2815          * synchronization point.
2816          */
2817         if (record->flags & HAMMER_RECF_DELETED_FE) {
2818                 if (record->type == HAMMER_MEM_RECORD_ADD) {
2819                         /*
2820                          * Convert a front-end deleted directory-add to
2821                          * a directory-delete entry later.
2822                          */
2823                         record->flags |= HAMMER_RECF_CONVERT_DELETE;
2824                 } else {
2825                         /*
2826                          * Dispose of the record (race case).  Mark as
2827                          * deleted by backend (and not committed).
2828                          */
2829                         KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
2830                         record->flags |= HAMMER_RECF_DELETED_BE;
2831                         ++record->ip->rec_generation;
2832                         error = 0;
2833                         goto done;
2834                 }
2835         }
2836
2837         /*
2838          * Assign the create_tid for new records.  Deletions already
2839          * have the record's entire key properly set up.
2840          */
2841         if (record->type != HAMMER_MEM_RECORD_DEL) {
2842                 record->leaf.base.create_tid = trans->tid;
2843                 record->leaf.create_ts = trans->time32;
2844         }
2845
2846         /*
2847          * This actually moves the record to the on-media B-Tree.  We
2848          * must also generate REDO_TERM entries in the UNDO/REDO FIFO
2849          * indicating that the related REDO_WRITE(s) have been committed.
2850          *
2851          * During recovery any REDO_TERM's within the nominal recovery span
2852          * are ignored since the related meta-data is being undone, causing
2853          * any matching REDO_WRITEs to execute.  The REDO_TERMs outside
2854          * the nominal recovery span will match against REDO_WRITEs and
2855          * prevent them from being executed (because the meta-data has
2856          * already been synchronized).
2857          */
2858         if (record->flags & HAMMER_RECF_REDO) {
2859                 KKASSERT(record->type == HAMMER_MEM_RECORD_DATA);
2860                 hammer_generate_redo(trans, record->ip,
2861                                      record->leaf.base.key -
2862                                          record->leaf.data_len,
2863                                      HAMMER_REDO_TERM_WRITE,
2864                                      NULL,
2865                                      record->leaf.data_len);
2866         }
2867
2868         for (;;) {
2869                 error = hammer_ip_sync_record_cursor(cursor, record);
2870                 if (error != EDEADLK)
2871                         break;
2872                 hammer_done_cursor(cursor);
2873                 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2874                                            record->ip);
2875                 if (error)
2876                         break;
2877         }
2878         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
2879
2880         if (error)
2881                 error = -error;
2882 done:
2883         hammer_flush_record_done(record, error);
2884
2885         /*
2886          * Do partial finalization if we have built up too many dirty
2887          * buffers.  Otherwise a buffer cache deadlock can occur when
2888          * doing things like creating tens of thousands of tiny files.
2889          *
2890          * We must release our cursor lock to avoid a 3-way deadlock
2891          * due to the exclusive sync lock the finalizer must get.
2892          *
2893          * WARNING: See warnings in hammer_unlock_cursor() function.
2894          */
2895         if (hammer_flusher_meta_limit(hmp) ||
2896             vm_page_count_severe()) {
2897                 hammer_unlock_cursor(cursor);
2898                 hammer_flusher_finalize(trans, 0);
2899                 hammer_lock_cursor(cursor);
2900         }
2901         return(error);
2902 }
2903
2904 /*
2905  * Backend function called by the flusher to sync an inode to media.
2906  */
2907 int
2908 hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
2909 {
2910         struct hammer_cursor cursor;
2911         hammer_node_t tmp_node;
2912         hammer_record_t depend;
2913         hammer_record_t next;
2914         int error, tmp_error;
2915         uint64_t nlinks;
2916
2917         if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
2918                 return(0);
2919
2920         error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2921         if (error)
2922                 goto done;
2923
2924         /*
2925          * Any directory records referencing this inode which are not in
2926          * our current flush group must adjust our nlink count for the
2927          * purposes of synchronizating to disk.
2928          *
2929          * Records which are in our flush group can be unlinked from our
2930          * inode now, potentially allowing the inode to be physically
2931          * deleted.
2932          *
2933          * This cannot block.
2934          */
2935         nlinks = ip->ino_data.nlinks;
2936         next = TAILQ_FIRST(&ip->target_list);
2937         while ((depend = next) != NULL) {
2938                 next = TAILQ_NEXT(depend, target_entry);
2939                 if (depend->flush_state == HAMMER_FST_FLUSH &&
2940                     depend->flush_group == ip->flush_group) {
2941                         /*
2942                          * If this is an ADD that was deleted by the frontend
2943                          * the frontend nlinks count will have already been
2944                          * decremented, but the backend is going to sync its
2945                          * directory entry and must account for it.  The
2946                          * record will be converted to a delete-on-disk when
2947                          * it gets synced.
2948                          *
2949                          * If the ADD was not deleted by the frontend we
2950                          * can remove the dependancy from our target_list.
2951                          */
2952                         if (depend->flags & HAMMER_RECF_DELETED_FE) {
2953                                 ++nlinks;
2954                         } else {
2955                                 TAILQ_REMOVE(&ip->target_list, depend,
2956                                              target_entry);
2957                                 depend->target_ip = NULL;
2958                         }
2959                 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
2960                         /*
2961                          * Not part of our flush group and not deleted by
2962                          * the front-end, adjust the link count synced to
2963                          * the media (undo what the frontend did when it
2964                          * queued the record).
2965                          */
2966                         KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
2967                         switch(depend->type) {
2968                         case HAMMER_MEM_RECORD_ADD:
2969                                 --nlinks;
2970                                 break;
2971                         case HAMMER_MEM_RECORD_DEL:
2972                                 ++nlinks;
2973                                 break;
2974                         default:
2975                                 break;
2976                         }
2977                 }
2978         }
2979
2980         /*
2981          * Set dirty if we had to modify the link count.
2982          */
2983         if (ip->sync_ino_data.nlinks != nlinks) {
2984                 KKASSERT((int64_t)nlinks >= 0);
2985                 ip->sync_ino_data.nlinks = nlinks;
2986                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
2987         }
2988
2989         /*
2990          * If there is a trunction queued destroy any data past the (aligned)
2991          * truncation point.  Userland will have dealt with the buffer
2992          * containing the truncation point for us.
2993          *
2994          * We don't flush pending frontend data buffers until after we've
2995          * dealt with the truncation.
2996          */
2997         if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2998                 /*
2999                  * Interlock trunc_off.  The VOP front-end may continue to
3000                  * make adjustments to it while we are blocked.
3001                  */
3002                 off_t trunc_off;
3003                 off_t aligned_trunc_off;
3004                 int blkmask;
3005
3006                 trunc_off = ip->sync_trunc_off;
3007                 blkmask = hammer_blocksize(trunc_off) - 1;
3008                 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
3009
3010                 /*
3011                  * Delete any whole blocks on-media.  The front-end has
3012                  * already cleaned out any partial block and made it
3013                  * pending.  The front-end may have updated trunc_off
3014                  * while we were blocked so we only use sync_trunc_off.
3015                  *
3016                  * This operation can blow out the buffer cache, EWOULDBLOCK
3017                  * means we were unable to complete the deletion.  The
3018                  * deletion will update sync_trunc_off in that case.
3019                  */
3020                 error = hammer_ip_delete_range(&cursor, ip,
3021                                                 aligned_trunc_off,
3022                                                 0x7FFFFFFFFFFFFFFFLL, 2);
3023                 if (error == EWOULDBLOCK) {
3024                         ip->flags |= HAMMER_INODE_WOULDBLOCK;
3025                         error = 0;
3026                         goto defer_buffer_flush;
3027                 }
3028
3029                 if (error)
3030                         goto done;
3031
3032                 /*
3033                  * Generate a REDO_TERM_TRUNC entry in the UNDO/REDO FIFO.
3034                  *
3035                  * XXX we do this even if we did not previously generate
3036                  * a REDO_TRUNC record.  This operation may enclosed the
3037                  * range for multiple prior truncation entries in the REDO
3038                  * log.
3039                  */
3040                 if (trans->hmp->version >= HAMMER_VOL_VERSION_FOUR &&
3041                     (ip->flags & HAMMER_INODE_RDIRTY)) {
3042                         hammer_generate_redo(trans, ip, aligned_trunc_off,
3043                                              HAMMER_REDO_TERM_TRUNC,
3044                                              NULL, 0);
3045                 }
3046
3047                 /*
3048                  * Clear the truncation flag on the backend after we have
3049                  * completed the deletions.  Backend data is now good again
3050                  * (including new records we are about to sync, below).
3051                  *
3052                  * Leave sync_trunc_off intact.  As we write additional
3053                  * records the backend will update sync_trunc_off.  This
3054                  * tells the backend whether it can skip the overwrite
3055                  * test.  This should work properly even when the backend
3056                  * writes full blocks where the truncation point straddles
3057                  * the block because the comparison is against the base
3058                  * offset of the record.
3059                  */
3060                 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
3061                 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
3062         } else {
3063                 error = 0;
3064         }
3065
3066         /*
3067          * Now sync related records.  These will typically be directory
3068          * entries, records tracking direct-writes, or delete-on-disk records.
3069          */
3070         if (error == 0) {
3071                 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
3072                                     hammer_sync_record_callback, &cursor);
3073                 if (tmp_error < 0)
3074                         tmp_error = -error;
3075                 if (tmp_error)
3076                         error = tmp_error;
3077         }
3078         hammer_cache_node(&ip->cache[1], cursor.node);
3079
3080         /*
3081          * Re-seek for inode update, assuming our cache hasn't been ripped
3082          * out from under us.
3083          */
3084         if (error == 0) {
3085                 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
3086                 if (tmp_node) {
3087                         hammer_cursor_downgrade(&cursor);
3088                         hammer_lock_sh(&tmp_node->lock);
3089                         if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
3090                                 hammer_cursor_seek(&cursor, tmp_node, 0);
3091                         hammer_unlock(&tmp_node->lock);
3092                         hammer_rel_node(tmp_node);
3093                 }
3094                 error = 0;
3095         }
3096
3097         /*
3098          * If we are deleting the inode the frontend had better not have
3099          * any active references on elements making up the inode.
3100          *
3101          * The call to hammer_ip_delete_clean() cleans up auxillary records
3102          * but not DB or DATA records.  Those must have already been deleted
3103          * by the normal truncation mechanic.
3104          */
3105         if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
3106                 RB_EMPTY(&ip->rec_tree)  &&
3107             (ip->sync_flags & HAMMER_INODE_DELETING) &&
3108             (ip->flags & HAMMER_INODE_DELETED) == 0) {
3109                 int count1 = 0;
3110
3111                 error = hammer_ip_delete_clean(&cursor, ip, &count1);
3112                 if (error == 0) {
3113                         ip->flags |= HAMMER_INODE_DELETED;
3114                         ip->sync_flags &= ~HAMMER_INODE_DELETING;
3115                         ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
3116                         KKASSERT(RB_EMPTY(&ip->rec_tree));
3117
3118                         /*
3119                          * Set delete_tid in both the frontend and backend
3120                          * copy of the inode record.  The DELETED flag handles
3121                          * this, do not set DDIRTY.
3122                          */
3123                         ip->ino_leaf.base.delete_tid = trans->tid;
3124                         ip->sync_ino_leaf.base.delete_tid = trans->tid;
3125                         ip->ino_leaf.delete_ts = trans->time32;
3126                         ip->sync_ino_leaf.delete_ts = trans->time32;
3127
3128
3129                         /*
3130                          * Adjust the inode count in the volume header
3131                          */
3132                         hammer_sync_lock_sh(trans);
3133                         if (ip->flags & HAMMER_INODE_ONDISK) {
3134                                 hammer_modify_volume_field(trans,
3135                                                            trans->rootvol,
3136                                                            vol0_stat_inodes);
3137                                 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
3138                                 hammer_modify_volume_done(trans->rootvol);
3139                         }
3140                         hammer_sync_unlock(trans);
3141                 }
3142         }
3143
3144         if (error)
3145                 goto done;
3146         ip->sync_flags &= ~HAMMER_INODE_BUFS;
3147
3148 defer_buffer_flush:
3149         /*
3150          * Now update the inode's on-disk inode-data and/or on-disk record.
3151          * DELETED and ONDISK are managed only in ip->flags.
3152          *
3153          * In the case of a defered buffer flush we still update the on-disk
3154          * inode to satisfy visibility requirements if there happen to be
3155          * directory dependancies.
3156          */
3157         switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
3158         case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
3159                 /*
3160                  * If deleted and on-disk, don't set any additional flags.
3161                  * the delete flag takes care of things.
3162                  *
3163                  * Clear flags which may have been set by the frontend.
3164                  */
3165                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3166                                     HAMMER_INODE_SDIRTY |
3167                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3168                                     HAMMER_INODE_DELETING);
3169                 break;
3170         case HAMMER_INODE_DELETED:
3171                 /*
3172                  * Take care of the case where a deleted inode was never
3173                  * flushed to the disk in the first place.
3174                  *
3175                  * Clear flags which may have been set by the frontend.
3176                  */
3177                 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
3178                                     HAMMER_INODE_SDIRTY |
3179                                     HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
3180                                     HAMMER_INODE_DELETING);
3181                 while (RB_ROOT(&ip->rec_tree)) {
3182                         hammer_record_t record = RB_ROOT(&ip->rec_tree);
3183                         hammer_ref(&record->lock);
3184                         KKASSERT(hammer_oneref(&record->lock));
3185                         record->flags |= HAMMER_RECF_DELETED_BE;
3186                         ++record->ip->rec_generation;
3187                         hammer_rel_mem_record(record);
3188                 }
3189                 break;
3190         case HAMMER_INODE_ONDISK:
3191                 /*
3192                  * If already on-disk, do not set any additional flags.
3193                  */
3194                 break;
3195         default:
3196                 /*
3197                  * If not on-disk and not deleted, set DDIRTY to force
3198                  * an initial record to be written.
3199                  *
3200                  * Also set the create_tid in both the frontend and backend
3201                  * copy of the inode record.
3202                  */
3203                 ip->ino_leaf.base.create_tid = trans->tid;
3204                 ip->ino_leaf.create_ts = trans->time32;
3205                 ip->sync_ino_leaf.base.create_tid = trans->tid;
3206                 ip->sync_ino_leaf.create_ts = trans->time32;
3207                 ip->sync_flags |= HAMMER_INODE_DDIRTY;
3208                 break;
3209         }
3210
3211         /*
3212          * If DDIRTY or SDIRTY is set, write out a new record.
3213          * If the inode is already on-disk the old record is marked as
3214          * deleted.
3215          *
3216          * If DELETED is set hammer_update_inode() will delete the existing
3217          * record without writing out a new one.
3218          */
3219         if (ip->flags & HAMMER_INODE_DELETED) {
3220                 error = hammer_update_inode(&cursor, ip);
3221         } else
3222         if (!(ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY)) &&
3223             (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
3224                 error = hammer_update_itimes(&cursor, ip);
3225         } else
3226         if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY |
3227                               HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
3228                 error = hammer_update_inode(&cursor, ip);
3229         }
3230 done:
3231         if (ip->flags & HAMMER_INODE_MODMASK)
3232                 hammer_inode_dirty(ip);
3233         if (error) {
3234                 hammer_critical_error(ip->hmp, ip, error,
3235                                       "while syncing inode");
3236         }
3237         hammer_done_cursor(&cursor);
3238         return(error);
3239 }
3240
3241 /*
3242  * This routine is called when the OS is no longer actively referencing
3243  * the inode (but might still be keeping it cached), or when releasing
3244  * the last reference to an inode.
3245  *
3246  * At this point if the inode's nlinks count is zero we want to destroy
3247  * it, which may mean destroying it on-media too.
3248  */
3249 void
3250 hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
3251 {
3252         struct vnode *vp;
3253
3254         /*
3255          * Set the DELETING flag when the link count drops to 0 and the
3256          * OS no longer has any opens on the inode.
3257          *
3258          * The backend will clear DELETING (a mod flag) and set DELETED
3259          * (a state flag) when it is actually able to perform the
3260          * operation.
3261          *
3262          * Don't reflag the deletion if the flusher is currently syncing
3263          * one that was already flagged.  A previously set DELETING flag
3264          * may bounce around flags and sync_flags until the operation is
3265          * completely done.
3266          *
3267          * Do not attempt to modify a snapshot inode (one set to read-only).
3268          */
3269         if (ip->ino_data.nlinks == 0 &&
3270             ((ip->flags | ip->sync_flags) & (HAMMER_INODE_RO|HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
3271                 ip->flags |= HAMMER_INODE_DELETING;
3272                 ip->flags |= HAMMER_INODE_TRUNCATED;
3273                 ip->trunc_off = 0;
3274                 vp = NULL;
3275                 if (getvp) {
3276                         if (hammer_get_vnode(ip, &vp) != 0)
3277                                 return;
3278                 }
3279
3280                 /*
3281                  * Final cleanup
3282                  */
3283                 if (ip->vp)
3284                         nvtruncbuf(ip->vp, 0, HAMMER_BUFSIZE, 0, 0);
3285                 if (ip->flags & HAMMER_INODE_MODMASK)
3286                         hammer_inode_dirty(ip);
3287                 if (getvp)
3288                         vput(vp);
3289         }
3290 }
3291
3292 /*
3293  * After potentially resolving a dependancy the inode is tested
3294  * to determine whether it needs to be reflushed.
3295  */
3296 void
3297 hammer_test_inode(hammer_inode_t ip)
3298 {
3299         if (ip->flags & HAMMER_INODE_REFLUSH) {
3300                 ip->flags &= ~HAMMER_INODE_REFLUSH;
3301                 hammer_ref(&ip->lock);
3302                 if (ip->flags & HAMMER_INODE_RESIGNAL) {
3303                         ip->flags &= ~HAMMER_INODE_RESIGNAL;
3304                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
3305                 } else {
3306                         hammer_flush_inode(ip, 0);
3307                 }
3308                 hammer_rel_inode(ip, 0);
3309         }
3310 }
3311
3312 /*
3313  * Clear the RECLAIM flag on an inode.  This occurs when the inode is
3314  * reassociated with a vp or just before it gets freed.
3315  *
3316  * Pipeline wakeups to threads blocked due to an excessive number of
3317  * detached inodes.  This typically occurs when atime updates accumulate
3318  * while scanning a directory tree.
3319  */
3320 static void
3321 hammer_inode_wakereclaims(hammer_inode_t ip)
3322 {
3323         struct hammer_reclaim *reclaim;
3324         hammer_mount_t hmp = ip->hmp;
3325
3326         if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
3327                 return;
3328
3329         --hammer_count_reclaims;
3330         --hmp->count_reclaims;
3331         ip->flags &= ~HAMMER_INODE_RECLAIM;
3332
3333         if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
3334                 KKASSERT(reclaim->count > 0);
3335                 if (--reclaim->count == 0) {
3336                         TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3337                         wakeup(reclaim);
3338                 }
3339         }
3340 }
3341
3342 /*
3343  * Setup our reclaim pipeline.  We only let so many detached (and dirty)
3344  * inodes build up before we start blocking.  This routine is called
3345  * if a new inode is created or an inode is loaded from media.
3346  *
3347  * When we block we don't care *which* inode has finished reclaiming,
3348  * as long as one does.
3349  *
3350  * The reclaim pipeline is primarily governed by the auto-flush which is
3351  * 1/4 hammer_limit_reclaims.  We don't want to block if the count is
3352  * less than 1/2 hammer_limit_reclaims.  From 1/2 to full count is
3353  * dynamically governed.
3354  */
3355 void
3356 hammer_inode_waitreclaims(hammer_transaction_t trans)
3357 {
3358         hammer_mount_t hmp = trans->hmp;
3359         struct hammer_reclaim reclaim;
3360         int lower_limit;
3361
3362         /*
3363          * Track inode load, delay if the number of reclaiming inodes is
3364          * between 2/4 and 4/4 hammer_limit_reclaims, depending.
3365          */
3366         if (curthread->td_proc) {
3367                 struct hammer_inostats *stats;
3368
3369                 stats = hammer_inode_inostats(hmp, curthread->td_proc->p_pid);
3370                 ++stats->count;
3371
3372                 if (stats->count > hammer_limit_reclaims / 2)
3373                         stats->count = hammer_limit_reclaims / 2;
3374                 lower_limit = hammer_limit_reclaims - stats->count;
3375                 if (hammer_debug_general & 0x10000) {
3376                         hdkprintf("pid %5d limit %d\n",
3377                                 (int)curthread->td_proc->p_pid, lower_limit);
3378                 }
3379         } else {
3380                 lower_limit = hammer_limit_reclaims * 3 / 4;
3381         }
3382         if (hmp->count_reclaims >= lower_limit) {
3383                 reclaim.count = 1;
3384                 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3385                 tsleep(&reclaim, 0, "hmrrcm", hz);
3386                 if (reclaim.count > 0)
3387                         TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
3388         }
3389 }
3390
3391 /*
3392  * Keep track of reclaim statistics on a per-pid basis using a loose
3393  * 4-way set associative hash table.  Collisions inherit the count of
3394  * the previous entry.
3395  *
3396  * NOTE: We want to be careful here to limit the chain size.  If the chain
3397  *       size is too large a pid will spread its stats out over too many
3398  *       entries under certain types of heavy filesystem activity and
3399  *       wind up not delaying long enough.
3400  */
3401 static
3402 struct hammer_inostats *
3403 hammer_inode_inostats(hammer_mount_t hmp, pid_t pid)
3404 {
3405         struct hammer_inostats *stats;
3406         int delta;
3407         int chain;
3408         static volatile int iterator;   /* we don't care about MP races */
3409
3410         /*
3411          * Chain up to 4 times to find our entry.
3412          */
3413         for (chain = 0; chain < 4; ++chain) {
3414                 stats = &hmp->inostats[(pid + chain) & HAMMER_INOSTATS_HMASK];
3415                 if (stats->pid == pid)
3416                         break;
3417         }
3418
3419         /*
3420          * Replace one of the four chaining entries with our new entry.
3421          */
3422         if (chain == 4) {
3423                 stats = &hmp->inostats[(pid + (iterator++ & 3)) &
3424                                        HAMMER_INOSTATS_HMASK];
3425                 stats->pid = pid;
3426         }
3427
3428         /*
3429          * Decay the entry
3430          */
3431         if (stats->count && stats->ltick != ticks) {
3432                 delta = ticks - stats->ltick;
3433                 stats->ltick = ticks;
3434                 if (delta <= 0 || delta > hz * 60)
3435                         stats->count = 0;
3436                 else
3437                         stats->count = stats->count * hz / (hz + delta);
3438         }
3439         if (hammer_debug_general & 0x10000)
3440                 hdkprintf("pid %5d stats %d\n", (int)pid, stats->count);
3441         return (stats);
3442 }
3443
3444 #if 0
3445
3446 /*
3447  * XXX not used, doesn't work very well due to the large batching nature
3448  * of flushes.
3449  *
3450  * A larger then normal backlog of inodes is sitting in the flusher,
3451  * enforce a general slowdown to let it catch up.  This routine is only
3452  * called on completion of a non-flusher-related transaction which
3453  * performed B-Tree node I/O.
3454  *
3455  * It is possible for the flusher to stall in a continuous load.
3456  * blogbench -i1000 -o seems to do a good job generating this sort of load.
3457  * If the flusher is unable to catch up the inode count can bloat until
3458  * we run out of kvm.
3459  *
3460  * This is a bit of a hack.
3461  */
3462 void
3463 hammer_inode_waithard(hammer_mount_t hmp)
3464 {
3465         /*
3466          * Hysteresis.
3467          */
3468         if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
3469                 if (hmp->count_reclaims < hammer_limit_reclaims / 2 &&
3470                     hmp->count_iqueued < hmp->count_inodes / 20) {
3471                         hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
3472                         return;
3473                 }
3474         } else {
3475                 if (hmp->count_reclaims < hammer_limit_reclaims ||
3476                     hmp->count_iqueued < hmp->count_inodes / 10) {
3477                         return;
3478                 }
3479                 hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
3480         }
3481
3482         /*
3483          * Block for one flush cycle.
3484          */
3485         hammer_flusher_wait_next(hmp);
3486 }
3487
3488 #endif