sys/vfs/hammer2/hammer2_flush.c

   1 /*
   2  * Copyright (c) 2011-2015 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@dragonflybsd.org>
   6  * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 /*
  36  *                      TRANSACTION AND FLUSH HANDLING
  37  *
  38  * Deceptively simple but actually fairly difficult to implement properly is
  39  * how I would describe it.
  40  *
  41  * Flushing generally occurs bottom-up but requires a top-down scan to
  42  * locate chains with MODIFIED and/or UPDATE bits set.  The ONFLUSH flag
  43  * tells how to recurse downward to find these chains.
  44  */
  45
  46 #include <sys/cdefs.h>
  47 #include <sys/param.h>
  48 #include <sys/systm.h>
  49 #include <sys/types.h>
  50 #include <sys/lock.h>
  51 #include <sys/uuid.h>
  52
  53 #include "hammer2.h"
  54
  55 #define FLUSH_DEBUG 0
  56
  57 #define HAMMER2_FLUSH_DEPTH_LIMIT       60      /* stack recursion limit */
  58
  59
  60 /*
  61  * Recursively flush the specified chain.  The chain is locked and
  62  * referenced by the caller and will remain so on return.  The chain
  63  * will remain referenced throughout but can temporarily lose its
  64  * lock during the recursion to avoid unnecessarily stalling user
  65  * processes.
  66  */
  67 struct hammer2_flush_info {
  68         hammer2_chain_t *parent;
  69         int             depth;
  70         long            diddeferral;
  71         int             error;                  /* cumulative error */
  72         int             flags;
  73 #ifdef HAMMER2_SCAN_DEBUG
  74         long            scan_count;
  75         long            scan_mod_count;
  76         long            scan_upd_count;
  77         long            scan_onf_count;
  78         long            scan_del_count;
  79         long            scan_btype[7];
  80         long            flushq_count;
  81 #endif
  82         struct h2_flush_list flushq;
  83         hammer2_chain_t *debug;
  84 };
  85
  86 typedef struct hammer2_flush_info hammer2_flush_info_t;
  87
  88 static void hammer2_flush_core(hammer2_flush_info_t *info,
  89                                 hammer2_chain_t *chain, int flags);
  90 static int hammer2_flush_recurse(hammer2_chain_t *child, void *data);
  91
  92 /*
  93  * Any per-pfs transaction initialization goes here.
  94  */
  95 void
  96 hammer2_trans_manage_init(hammer2_pfs_t *pmp)
  97 {
  98 }
  99
 100 /*
 101  * Transaction support for any modifying operation.  Transactions are used
 102  * in the pmp layer by the frontend and in the spmp layer by the backend.
 103  *
 104  * 0                    - Normal transaction, interlocked against flush
 105  *                        transaction.
 106  *
 107  * TRANS_ISFLUSH        - Flush transaction, interlocked against normal
 108  *                        transaction.
 109  *
 110  * TRANS_BUFCACHE       - Buffer cache transaction, no interlock.
 111  *
 112  * Initializing a new transaction allocates a transaction ID.  Typically
 113  * passed a pmp (hmp passed as NULL), indicating a cluster transaction.  Can
 114  * be passed a NULL pmp and non-NULL hmp to indicate a transaction on a single
 115  * media target.  The latter mode is used by the recovery code.
 116  *
 117  * TWO TRANSACTION IDs can run concurrently, where one is a flush and the
 118  * other is a set of any number of concurrent filesystem operations.  We
 119  * can either have <running_fs_ops> + <waiting_flush> + <blocked_fs_ops>
 120  * or we can have <running_flush> + <concurrent_fs_ops>.
 121  *
 122  * During a flush, new fs_ops are only blocked until the fs_ops prior to
 123  * the flush complete.  The new fs_ops can then run concurrent with the flush.
 124  *
 125  * Buffer-cache transactions operate as fs_ops but never block.  A
 126  * buffer-cache flush will run either before or after the current pending
 127  * flush depending on its state.
 128  */
 129 void
 130 hammer2_trans_init(hammer2_pfs_t *pmp, uint32_t flags)
 131 {
 132         uint32_t oflags;
 133         uint32_t nflags;
 134         int dowait;
 135
 136         for (;;) {
 137                 oflags = pmp->trans.flags;
 138                 cpu_ccfence();
 139                 dowait = 0;
 140
 141                 if (flags & HAMMER2_TRANS_ISFLUSH) {
 142                         /*
 143                          * Requesting flush transaction.  Wait for all
 144                          * currently running transactions to finish.
 145                          * Afterwords, normal transactions will be
 146                          * interlocked.
 147                          */
 148                         if (oflags & HAMMER2_TRANS_MASK) {
 149                                 nflags = oflags | HAMMER2_TRANS_FPENDING |
 150                                                   HAMMER2_TRANS_WAITING;
 151                                 dowait = 1;
 152                         } else {
 153                                 nflags = (oflags | flags) + 1;
 154                         }
 155                 } else if (flags & HAMMER2_TRANS_BUFCACHE) {
 156                         /*
 157                          * Requesting strategy transaction from buffer-cache,
 158                          * or a VM getpages/putpages through the buffer cache.
 159                          * We must allow such transactions in all situations
 160                          * to avoid deadlocks.
 161                          */
 162                         nflags = (oflags | flags) + 1;
 163 #if 0
 164                         /*
 165                          * (old) previous code interlocked against the main
 166                          *       flush pass.
 167                          */
 168                         if ((oflags & (HAMMER2_TRANS_ISFLUSH |
 169                                        HAMMER2_TRANS_PREFLUSH)) ==
 170                             HAMMER2_TRANS_ISFLUSH) {
 171                                 nflags = oflags | HAMMER2_TRANS_WAITING;
 172                                 dowait = 1;
 173                         } else {
 174                                 nflags = (oflags | flags) + 1;
 175                         }
 176 #endif
 177                 } else {
 178                         /*
 179                          * Requesting normal modifying transaction (read-only
 180                          * operations do not use transactions).  Waits for
 181                          * any flush to finish before allowing.  Multiple
 182                          * modifying transactions can run concurrently.
 183                          */
 184                         if (oflags & HAMMER2_TRANS_ISFLUSH) {
 185                                 nflags = oflags | HAMMER2_TRANS_WAITING;
 186                                 dowait = 1;
 187                         } else {
 188                                 nflags = (oflags | flags) + 1;
 189                         }
 190                 }
 191                 if (dowait)
 192                         tsleep_interlock(&pmp->trans.sync_wait, 0);
 193                 if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
 194                         if (dowait == 0)
 195                                 break;
 196                         tsleep(&pmp->trans.sync_wait, PINTERLOCKED,
 197                                "h2trans", hz);
 198                 } else {
 199                         cpu_pause();
 200                 }
 201                 /* retry */
 202         }
 203 }
 204
 205 /*
 206  * Start a sub-transaction, there is no 'subdone' function.  This will
 207  * issue a new modify_tid (mtid) for the current transaction, which is a
 208  * CLC (cluster level change) id and not a per-node id.
 209  *
 210  * This function must be called for each XOP when multiple XOPs are run in
 211  * sequence within a transaction.
 212  *
 213  * Callers typically update the inode with the transaction mtid manually
 214  * to enforce sequencing.
 215  */
 216 hammer2_tid_t
 217 hammer2_trans_sub(hammer2_pfs_t *pmp)
 218 {
 219         hammer2_tid_t mtid;
 220
 221         mtid = atomic_fetchadd_64(&pmp->modify_tid, 1);
 222
 223         return (mtid);
 224 }
 225
 226 void
 227 hammer2_trans_done(hammer2_pfs_t *pmp)
 228 {
 229         uint32_t oflags;
 230         uint32_t nflags;
 231
 232         for (;;) {
 233                 oflags = pmp->trans.flags;
 234                 cpu_ccfence();
 235                 KKASSERT(oflags & HAMMER2_TRANS_MASK);
 236                 if ((oflags & HAMMER2_TRANS_MASK) == 1) {
 237                         /*
 238                          * This was the last transaction
 239                          */
 240                         nflags = (oflags - 1) & ~(HAMMER2_TRANS_ISFLUSH |
 241                                                   HAMMER2_TRANS_BUFCACHE |
 242                                                   HAMMER2_TRANS_FPENDING |
 243                                                   HAMMER2_TRANS_WAITING);
 244                 } else {
 245                         /*
 246                          * Still transactions pending
 247                          */
 248                         nflags = oflags - 1;
 249                 }
 250                 if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
 251                         if ((nflags & HAMMER2_TRANS_MASK) == 0 &&
 252                             (oflags & HAMMER2_TRANS_WAITING)) {
 253                                 wakeup(&pmp->trans.sync_wait);
 254                         }
 255                         break;
 256                 } else {
 257                         cpu_pause();
 258                 }
 259                 /* retry */
 260         }
 261 }
 262
 263 /*
 264  * Obtain new, unique inode number (not serialized by caller).
 265  */
 266 hammer2_tid_t
 267 hammer2_trans_newinum(hammer2_pfs_t *pmp)
 268 {
 269         hammer2_tid_t tid;
 270
 271         tid = atomic_fetchadd_64(&pmp->inode_tid, 1);
 272
 273         return tid;
 274 }
 275
 276 /*
 277  * Assert that a strategy call is ok here.  Currently we allow strategy
 278  * calls in all situations, including during flushes.  Previously:
 279  *      (old) (1) In a normal transaction.
 280  *      (old) (2) In a flush transaction only if PREFLUSH is also set.
 281  */
 282 void
 283 hammer2_trans_assert_strategy(hammer2_pfs_t *pmp)
 284 {
 285 #if 0
 286         KKASSERT((pmp->trans.flags & HAMMER2_TRANS_ISFLUSH) == 0 ||
 287                  (pmp->trans.flags & HAMMER2_TRANS_PREFLUSH));
 288 #endif
 289 }
 290
 291
 292 /*
 293  * Chains undergoing destruction are removed from the in-memory topology.
 294  * To avoid getting lost these chains are placed on the delayed flush
 295  * queue which will properly dispose of them.
 296  *
 297  * We do this instead of issuing an immediate flush in order to give
 298  * recursive deletions (rm -rf, etc) a chance to remove more of the
 299  * hierarchy, potentially allowing an enormous amount of write I/O to
 300  * be avoided.
 301  *
 302  * NOTE: The flush code tests HAMMER2_CHAIN_DESTROY to differentiate
 303  *       between these chains and the deep-recursion requeue.
 304  */
 305 void
 306 hammer2_delayed_flush(hammer2_chain_t *chain)
 307 {
 308         if ((chain->flags & HAMMER2_CHAIN_DELAYED) == 0) {
 309                 hammer2_spin_ex(&chain->hmp->list_spin);
 310                 if ((chain->flags & (HAMMER2_CHAIN_DELAYED |
 311                                      HAMMER2_CHAIN_DEFERRED)) == 0) {
 312                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELAYED |
 313                                                       HAMMER2_CHAIN_DEFERRED);
 314                         TAILQ_INSERT_TAIL(&chain->hmp->flushq,
 315                                           chain, flush_node);
 316                         hammer2_chain_ref(chain);
 317                 }
 318                 hammer2_spin_unex(&chain->hmp->list_spin);
 319                 hammer2_voldata_modify(chain->hmp);
 320         }
 321 }
 322
 323 /*
 324  * Flush the chain and all modified sub-chains through the specified
 325  * synchronization point, propagating blockref updates back up.  As
 326  * part of this propagation, mirror_tid and inode/data usage statistics
 327  * propagates back upward.
 328  *
 329  * Returns a HAMMER2 error code, 0 if no error.  Note that I/O errors from
 330  * buffers dirtied during the flush operation can occur later.
 331  *
 332  * modify_tid (clc - cluster level change) is not propagated.
 333  *
 334  * update_tid (clc) is used for validation and is not propagated by this
 335  * function.
 336  *
 337  * This routine can be called from several places but the most important
 338  * is from VFS_SYNC (frontend) via hammer2_inode_xop_flush (backend).
 339  *
 340  * chain is locked on call and will remain locked on return.  The chain's
 341  * UPDATE flag indicates that its parent's block table (which is not yet
 342  * part of the flush) should be updated.
 343  *
 344  * flags:
 345  *      HAMMER2_FLUSH_TOP       Indicates that this is the top of the flush.
 346  *                              Is cleared for the recursion.
 347  *
 348  *      HAMMER2_FLUSH_ALL       Recurse everything
 349  *
 350  *      HAMMER2_FLUSH_INODE_RECURSE
 351  *                              Recurse one inode level, flush includes
 352  *                              sub-inodes but do not go deeper (thus UPDATE
 353  *                              can wind up remaining set).
 354  */
 355 int
 356 hammer2_flush(hammer2_chain_t *chain, int flags)
 357 {
 358         hammer2_chain_t *scan;
 359         hammer2_flush_info_t info;
 360         hammer2_dev_t *hmp;
 361         int loops;
 362
 363         /*
 364          * Execute the recursive flush and handle deferrals.
 365          *
 366          * Chains can be ridiculously long (thousands deep), so to
 367          * avoid blowing out the kernel stack the recursive flush has a
 368          * depth limit.  Elements at the limit are placed on a list
 369          * for re-execution after the stack has been popped.
 370          */
 371         bzero(&info, sizeof(info));
 372         TAILQ_INIT(&info.flushq);
 373         info.flags = flags & ~HAMMER2_FLUSH_TOP;
 374
 375         /*
 376          * Calculate parent (can be NULL), if not NULL the flush core
 377          * expects the parent to be referenced so it can easily lock/unlock
 378          * it without it getting ripped up.
 379          */
 380         if ((info.parent = chain->parent) != NULL)
 381                 hammer2_chain_ref(info.parent);
 382
 383         /*
 384          * Extra ref needed because flush_core expects it when replacing
 385          * chain.
 386          */
 387         hammer2_chain_ref(chain);
 388         hmp = chain->hmp;
 389         loops = 0;
 390
 391         for (;;) {
 392                 /*
 393                  * Move hmp->flushq to info.flushq if non-empty so it can
 394                  * be processed.
 395                  */
 396                 if (TAILQ_FIRST(&hmp->flushq) != NULL) {
 397                         hammer2_spin_ex(&chain->hmp->list_spin);
 398                         TAILQ_CONCAT(&info.flushq, &hmp->flushq, flush_node);
 399                         hammer2_spin_unex(&chain->hmp->list_spin);
 400                 }
 401
 402                 /*
 403                  * Unwind deep recursions which had been deferred.  This
 404                  * can leave the FLUSH_* bits set for these chains, which
 405                  * will be handled when we [re]flush chain after the unwind.
 406                  */
 407                 while ((scan = TAILQ_FIRST(&info.flushq)) != NULL) {
 408                         KKASSERT(scan->flags & HAMMER2_CHAIN_DEFERRED);
 409                         TAILQ_REMOVE(&info.flushq, scan, flush_node);
 410 #ifdef HAMMER2_SCAN_DEBUG
 411                         ++info.flushq_count;
 412 #endif
 413                         atomic_clear_int(&scan->flags, HAMMER2_CHAIN_DEFERRED |
 414                                                        HAMMER2_CHAIN_DELAYED);
 415
 416                         /*
 417                          * Now that we've popped back up we can do a secondary
 418                          * recursion on the deferred elements.
 419                          *
 420                          * NOTE: hmp->flushq chains (marked DESTROY) must be
 421                          *       handled unconditionally so they can be cleaned
 422                          *       out.
 423                          *
 424                          * NOTE: hammer2_flush() may replace scan.
 425                          */
 426                         if (hammer2_debug & 0x0040)
 427                                 kprintf("deferred flush %p\n", scan);
 428                         hammer2_chain_lock(scan, HAMMER2_RESOLVE_MAYBE);
 429                         if (scan->error == 0) {
 430                                 if (scan->flags & HAMMER2_CHAIN_DESTROY) {
 431                                         hammer2_flush(scan,
 432                                                     flags |
 433                                                     HAMMER2_FLUSH_TOP |
 434                                                     HAMMER2_FLUSH_ALL);
 435                                 } else {
 436                                         hammer2_flush(scan,
 437                                                     flags & ~HAMMER2_FLUSH_TOP);
 438                                 }
 439                         } else {
 440                                 info.error |= scan->error;
 441                         }
 442                         hammer2_chain_unlock(scan);
 443                         hammer2_chain_drop(scan);/* ref from defer */
 444                 }
 445
 446                 /*
 447                  * [re]flush chain as the deep recursion may have generated
 448                  * additional modifications.
 449                  */
 450                 info.diddeferral = 0;
 451                 if (info.parent != chain->parent) {
 452                         kprintf("LOST CHILD4 %p->%p (actual parent %p)\n",
 453                                 info.parent, chain, chain->parent);
 454                         hammer2_chain_drop(info.parent);
 455                         info.parent = chain->parent;
 456                         hammer2_chain_ref(info.parent);
 457                 }
 458                 hammer2_flush_core(&info, chain, flags);
 459
 460                 /*
 461                  * Only loop if deep recursions have been deferred.
 462                  */
 463                 if (TAILQ_EMPTY(&info.flushq))
 464                         break;
 465
 466                 if (++loops % 1000 == 0) {
 467                         kprintf("hammer2_flush: excessive loops on %p\n",
 468                                 chain);
 469                         if (hammer2_debug & 0x100000)
 470                                 Debugger("hell4");
 471                 }
 472         }
 473 #ifdef HAMMER2_SCAN_DEBUG
 474         if (info.scan_count >= 10)
 475         kprintf("hammer2_flush: scan_count %ld (%ld,%ld,%ld,%ld) "
 476                 "bt(%ld,%ld,%ld,%ld,%ld,%ld) flushq %ld\n",
 477                 info.scan_count,
 478                 info.scan_mod_count,
 479                 info.scan_upd_count,
 480                 info.scan_onf_count,
 481                 info.scan_del_count,
 482                 info.scan_btype[1],
 483                 info.scan_btype[2],
 484                 info.scan_btype[3],
 485                 info.scan_btype[4],
 486                 info.scan_btype[5],
 487                 info.scan_btype[6],
 488                 info.flushq_count);
 489 #endif
 490         hammer2_chain_drop(chain);
 491         if (info.parent)
 492                 hammer2_chain_drop(info.parent);
 493         return (info.error);
 494 }
 495
 496 /*
 497  * This is the core of the chain flushing code.  The chain is locked by the
 498  * caller and must also have an extra ref on it by the caller, and remains
 499  * locked and will have an extra ref on return.  info.parent is referenced
 500  * but not locked.
 501  *
 502  * Upon return, the caller can test the UPDATE bit on the chain to determine
 503  * if the parent needs updating.
 504  *
 505  * (1) Determine if this node is a candidate for the flush, return if it is
 506  *     not.  fchain and vchain are always candidates for the flush.
 507  *
 508  * (2) If we recurse too deep the chain is entered onto the deferral list and
 509  *     the current flush stack is aborted until after the deferral list is
 510  *     run.
 511  *
 512  * (3) Recursively flush live children (rbtree).  This can create deferrals.
 513  *     A successful flush clears the MODIFIED and UPDATE bits on the children
 514  *     and typically causes the parent to be marked MODIFIED as the children
 515  *     update the parent's block table.  A parent might already be marked
 516  *     MODIFIED due to a deletion (whos blocktable update in the parent is
 517  *     handled by the frontend), or if the parent itself is modified by the
 518  *     frontend for other reasons.
 519  *
 520  * (4) Permanently disconnected sub-trees are cleaned up by the front-end.
 521  *     Deleted-but-open inodes can still be individually flushed via the
 522  *     filesystem syncer.
 523  *
 524  * (5) Delete parents on the way back up if they are normal indirect blocks
 525  *     and have no children.
 526  *
 527  * (6) Note that an unmodified child may still need the block table in its
 528  *     parent updated (e.g. rename/move).  The child will have UPDATE set
 529  *     in this case.
 530  *
 531  *                      WARNING ON BREF MODIFY_TID/MIRROR_TID
 532  *
 533  * blockref.modify_tid is consistent only within a PFS, and will not be
 534  * consistent during synchronization.  mirror_tid is consistent across the
 535  * block device regardless of the PFS.
 536  */
 537 static void
 538 hammer2_flush_core(hammer2_flush_info_t *info, hammer2_chain_t *chain,
 539                    int flags)
 540 {
 541         hammer2_chain_t *parent;
 542         hammer2_dev_t *hmp;
 543         int save_error;
 544
 545         /*
 546          * (1) Optimize downward recursion to locate nodes needing action.
 547          *     Nothing to do if none of these flags are set.
 548          */
 549         if ((chain->flags & HAMMER2_CHAIN_FLUSH_MASK) == 0) {
 550                 if (hammer2_debug & 0x200) {
 551                         if (info->debug == NULL)
 552                                 info->debug = chain;
 553                 } else {
 554                         return;
 555                 }
 556         }
 557
 558         hmp = chain->hmp;
 559
 560         /*
 561          * NOTE: parent can be NULL, usually due to destroy races.
 562          */
 563         parent = info->parent;
 564         KKASSERT(chain->parent == parent);
 565
 566         /*
 567          * Downward search recursion
 568          *
 569          * We must be careful on cold stops.  If CHAIN_UPDATE is set and
 570          * we stop cold (verses a deferral which will re-run the chain later),
 571          * the update can wind up never being applied.  This situation most
 572          * typically occurs on inode boundaries due to the way
 573          * hammer2_vfs_sync() breaks-up the flush.  As a safety, we
 574          * flush-through such situations.
 575          */
 576         if (chain->flags & (HAMMER2_CHAIN_DEFERRED | HAMMER2_CHAIN_DELAYED)) {
 577                 /*
 578                  * Already deferred.
 579                  */
 580                 ++info->diddeferral;
 581         } else if ((chain->flags & HAMMER2_CHAIN_PFSBOUNDARY) &&
 582                    (chain->flags & HAMMER2_CHAIN_UPDATE) == 0 &&
 583                    (flags & HAMMER2_FLUSH_ALL) == 0 &&
 584                    (flags & HAMMER2_FLUSH_TOP) == 0 &&
 585                    chain->pmp && chain->pmp->mp) {
 586                 /*
 587                  * If FLUSH_ALL is not specified the caller does not want
 588                  * to recurse through PFS roots that have been mounted.
 589                  *
 590                  * (If the PFS has not been mounted there may not be
 591                  *  anything monitoring its chains and its up to us
 592                  *  to flush it).
 593                  *
 594                  * The typical sequence is to flush dirty PFS's starting at
 595                  * their root downward, then flush the device root (vchain).
 596                  * It is this second flush that typically leaves out the
 597                  * ALL flag.
 598                  *
 599                  * However we must still process the PFSROOT chains for block
 600                  * table updates in their parent (which IS part of our flush).
 601                  *
 602                  * NOTE: The volume root, vchain, does not set PFSBOUNDARY.
 603                  *
 604                  * NOTE: This test must be done before the depth-limit test,
 605                  *       else it might become the top on a flushq iteration.
 606                  *
 607                  * NOTE: We must re-set ONFLUSH in the parent to retain if
 608                  *       this chain (that we are skipping) requires work.
 609                  */
 610                 if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
 611                                     HAMMER2_CHAIN_DESTROY |
 612                                     HAMMER2_CHAIN_MODIFIED)) {
 613                         hammer2_chain_setflush(parent);
 614                 }
 615         } else if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
 616                    (chain->flags & HAMMER2_CHAIN_UPDATE) == 0 &&
 617                    (flags & HAMMER2_FLUSH_INODE_STOP) &&
 618                    (flags & HAMMER2_FLUSH_ALL) == 0 &&
 619                    (flags & HAMMER2_FLUSH_TOP) == 0 &&
 620                    chain->pmp && chain->pmp->mp) {
 621                 /*
 622                  * If FLUSH_INODE_STOP is specified and both ALL and TOP
 623                  * are clear, we must not flush the chain.  The chain should
 624                  * have already been flushed and any further ONFLUSH/UPDATE
 625                  * setting will be related to the next flush.
 626                  *
 627                  * This features allows us to flush inodes independently of
 628                  * each other and meta-data above the inodes separately.
 629                  */
 630                 if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
 631                                     HAMMER2_CHAIN_DESTROY |
 632                                     HAMMER2_CHAIN_MODIFIED)) {
 633                         if (parent)
 634                                 hammer2_chain_setflush(parent);
 635                 }
 636         } else if (info->depth == HAMMER2_FLUSH_DEPTH_LIMIT) {
 637                 /*
 638                  * Recursion depth reached.
 639                  */
 640                 KKASSERT((chain->flags & HAMMER2_CHAIN_DELAYED) == 0);
 641                 hammer2_chain_ref(chain);
 642                 TAILQ_INSERT_TAIL(&info->flushq, chain, flush_node);
 643                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEFERRED);
 644                 ++info->diddeferral;
 645         } else if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
 646                                    HAMMER2_CHAIN_DESTROY)) {
 647                 /*
 648                  * Downward recursion search (actual flush occurs bottom-up).
 649                  * pre-clear ONFLUSH.  It can get set again due to races or
 650                  * flush errors, which we want so the scan finds us again in
 651                  * the next flush.
 652                  *
 653                  * We must also recurse if DESTROY is set so we can finally
 654                  * get rid of the related children, otherwise the node will
 655                  * just get re-flushed on lastdrop.
 656                  *
 657                  * WARNING!  The recursion will unlock/relock info->parent
 658                  *           (which is 'chain'), potentially allowing it
 659                  *           to be ripped up.
 660                  */
 661                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
 662                 save_error = info->error;
 663                 info->error = 0;
 664                 info->parent = chain;
 665
 666                 /*
 667                  * We may have to do this twice to catch any indirect
 668                  * block maintenance that occurs.  Other conditions which
 669                  * can keep setting ONFLUSH (such as deferrals) ought to
 670                  * be handled by the flushq code.  XXX needs more help
 671                  */
 672                 hammer2_spin_ex(&chain->core.spin);
 673                 RB_SCAN(hammer2_chain_tree, &chain->core.rbtree,
 674                         NULL, hammer2_flush_recurse, info);
 675                 if (chain->flags & HAMMER2_CHAIN_ONFLUSH) {
 676                         atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
 677                         RB_SCAN(hammer2_chain_tree, &chain->core.rbtree,
 678                                 NULL, hammer2_flush_recurse, info);
 679                 }
 680                 hammer2_spin_unex(&chain->core.spin);
 681                 info->parent = parent;
 682
 683                 /*
 684                  * Re-set the flush bits if the flush was incomplete or
 685                  * an error occurred.  If an error occurs it is typically
 686                  * an allocation error.  Errors do not cause deferrals.
 687                  */
 688                 if (info->error)
 689                         hammer2_chain_setflush(chain);
 690                 info->error |= save_error;
 691                 if (info->diddeferral)
 692                         hammer2_chain_setflush(chain);
 693
 694                 /*
 695                  * If we lost the parent->chain association we have to
 696                  * stop processing this chain because it is no longer
 697                  * in this recursion.  If it moved, it will be handled
 698                  * by the ONFLUSH flag elsewhere.
 699                  */
 700                 if (chain->parent != parent) {
 701                         kprintf("LOST CHILD2 %p->%p (actual parent %p)\n",
 702                                 parent, chain, chain->parent);
 703                         goto done;
 704                 }
 705         }
 706
 707         /*
 708          * Now we are in the bottom-up part of the recursion.
 709          *
 710          * Do not update chain if lower layers were deferred.  We continue
 711          * to try to update the chain on lower-level errors, but the flush
 712          * code may decide not to flush the volume root.
 713          *
 714          * XXX should we continue to try to update the chain if an error
 715          *     occurred?
 716          */
 717         if (info->diddeferral)
 718                 goto done;
 719
 720         /*
 721          * Both parent and chain must be locked in order to flush chain,
 722          * in order to properly update the parent under certain conditions.
 723          *
 724          * In addition, we can't safely unlock/relock the chain once we
 725          * start flushing the chain itself, which we would have to do later
 726          * on in order to lock the parent if we didn't do that now.
 727          */
 728         hammer2_chain_ref_hold(chain);
 729         hammer2_chain_unlock(chain);
 730         if (parent)
 731                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
 732         hammer2_chain_lock(chain, HAMMER2_RESOLVE_MAYBE);
 733         hammer2_chain_drop_unhold(chain);
 734
 735         /*
 736          * Can't process if we can't access their content.
 737          */
 738         if ((parent && parent->error) || chain->error) {
 739                 kprintf("hammer2: chain error during flush\n");
 740                 info->error |= chain->error;
 741                 if (parent) {
 742                         info->error |= parent->error;
 743                         hammer2_chain_unlock(parent);
 744                 }
 745                 goto done;
 746         }
 747
 748         if (chain->parent != parent) {
 749                 kprintf("LOST CHILD3 %p->%p (actual parent %p)\n",
 750                         parent, chain, chain->parent);
 751                 KKASSERT(parent != NULL);
 752                 hammer2_chain_unlock(parent);
 753                 if ((chain->flags & HAMMER2_CHAIN_DELAYED) == 0) {
 754                         hammer2_chain_ref(chain);
 755                         TAILQ_INSERT_TAIL(&info->flushq, chain, flush_node);
 756                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEFERRED);
 757                         ++info->diddeferral;
 758                 }
 759                 goto done;
 760         }
 761
 762         /*
 763          * Propagate the DESTROY flag downwards.  This dummies up the flush
 764          * code and tries to invalidate related buffer cache buffers to
 765          * avoid the disk write.
 766          */
 767         if (parent && (parent->flags & HAMMER2_CHAIN_DESTROY))
 768                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
 769
 770         /*
 771          * Dispose of the modified bit.
 772          *
 773          * If parent is present, the UPDATE bit should already be set.
 774          * UPDATE should already be set.
 775          * bref.mirror_tid should already be set.
 776          */
 777         if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
 778                 KKASSERT((chain->flags & HAMMER2_CHAIN_UPDATE) ||
 779                          chain->parent == NULL);
 780                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
 781                 atomic_add_long(&hammer2_count_modified_chains, -1);
 782
 783                 /*
 784                  * Manage threads waiting for excessive dirty memory to
 785                  * be retired.
 786                  */
 787                 if (chain->pmp)
 788                         hammer2_pfs_memory_wakeup(chain->pmp);
 789
 790 #if 0
 791                 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0 &&
 792                     chain != &hmp->vchain &&
 793                     chain != &hmp->fchain) {
 794                         /*
 795                          * Set UPDATE bit indicating that the parent block
 796                          * table requires updating.
 797                          */
 798                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
 799                 }
 800 #endif
 801
 802                 /*
 803                  * Issue the flush.  This is indirect via the DIO.
 804                  *
 805                  * NOTE: A DELETED node that reaches this point must be
 806                  *       flushed for synchronization point consistency.
 807                  *
 808                  * NOTE: Even though MODIFIED was already set, the related DIO
 809                  *       might not be dirty due to a system buffer cache
 810                  *       flush and must be set dirty if we are going to make
 811                  *       further modifications to the buffer.  Chains with
 812                  *       embedded data don't need this.
 813                  */
 814                 if (hammer2_debug & 0x1000) {
 815                         kprintf("Flush %p.%d %016jx/%d data=%016jx\n",
 816                                 chain, chain->bref.type,
 817                                 (uintmax_t)chain->bref.key,
 818                                 chain->bref.keybits,
 819                                 (uintmax_t)chain->bref.data_off);
 820                 }
 821                 if (hammer2_debug & 0x2000) {
 822                         Debugger("Flush hell");
 823                 }
 824
 825                 /*
 826                  * Update chain CRCs for flush.
 827                  *
 828                  * NOTE: Volume headers are NOT flushed here as they require
 829                  *       special processing.
 830                  */
 831                 switch(chain->bref.type) {
 832                 case HAMMER2_BREF_TYPE_FREEMAP:
 833                         /*
 834                          * Update the volume header's freemap_tid to the
 835                          * freemap's flushing mirror_tid.
 836                          *
 837                          * (note: embedded data, do not call setdirty)
 838                          */
 839                         KKASSERT(hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED);
 840                         KKASSERT(chain == &hmp->fchain);
 841                         hmp->voldata.freemap_tid = chain->bref.mirror_tid;
 842                         if (hammer2_debug & 0x8000) {
 843                                 /* debug only, avoid syslogd loop */
 844                                 kprintf("sync freemap mirror_tid %08jx\n",
 845                                         (intmax_t)chain->bref.mirror_tid);
 846                         }
 847
 848                         /*
 849                          * The freemap can be flushed independently of the
 850                          * main topology, but for the case where it is
 851                          * flushed in the same transaction, and flushed
 852                          * before vchain (a case we want to allow for
 853                          * performance reasons), make sure modifications
 854                          * made during the flush under vchain use a new
 855                          * transaction id.
 856                          *
 857                          * Otherwise the mount recovery code will get confused.
 858                          */
 859                         ++hmp->voldata.mirror_tid;
 860                         break;
 861                 case HAMMER2_BREF_TYPE_VOLUME:
 862                         /*
 863                          * The free block table is flushed by
 864                          * hammer2_vfs_sync() before it flushes vchain.
 865                          * We must still hold fchain locked while copying
 866                          * voldata to volsync, however.
 867                          *
 868                          * These do not error per-say since their data does
 869                          * not need to be re-read from media on lock.
 870                          *
 871                          * (note: embedded data, do not call setdirty)
 872                          */
 873                         hammer2_chain_lock(&hmp->fchain,
 874                                            HAMMER2_RESOLVE_ALWAYS);
 875                         hammer2_voldata_lock(hmp);
 876                         if (hammer2_debug & 0x8000) {
 877                                 /* debug only, avoid syslogd loop */
 878                                 kprintf("sync volume  mirror_tid %08jx\n",
 879                                         (intmax_t)chain->bref.mirror_tid);
 880                         }
 881
 882                         /*
 883                          * Update the volume header's mirror_tid to the
 884                          * main topology's flushing mirror_tid.  It is
 885                          * possible that voldata.mirror_tid is already
 886                          * beyond bref.mirror_tid due to the bump we made
 887                          * above in BREF_TYPE_FREEMAP.
 888                          */
 889                         if (hmp->voldata.mirror_tid < chain->bref.mirror_tid) {
 890                                 hmp->voldata.mirror_tid =
 891                                         chain->bref.mirror_tid;
 892                         }
 893
 894                         /*
 895                          * The volume header is flushed manually by the
 896                          * syncer, not here.  All we do here is adjust the
 897                          * crc's.
 898                          */
 899                         KKASSERT(chain->data != NULL);
 900                         KKASSERT(chain->dio == NULL);
 901
 902                         hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT1]=
 903                                 hammer2_icrc32(
 904                                         (char *)&hmp->voldata +
 905                                          HAMMER2_VOLUME_ICRC1_OFF,
 906                                         HAMMER2_VOLUME_ICRC1_SIZE);
 907                         hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT0]=
 908                                 hammer2_icrc32(
 909                                         (char *)&hmp->voldata +
 910                                          HAMMER2_VOLUME_ICRC0_OFF,
 911                                         HAMMER2_VOLUME_ICRC0_SIZE);
 912                         hmp->voldata.icrc_volheader =
 913                                 hammer2_icrc32(
 914                                         (char *)&hmp->voldata +
 915                                          HAMMER2_VOLUME_ICRCVH_OFF,
 916                                         HAMMER2_VOLUME_ICRCVH_SIZE);
 917
 918                         if (hammer2_debug & 0x8000) {
 919                                 /* debug only, avoid syslogd loop */
 920                                 kprintf("syncvolhdr %016jx %016jx\n",
 921                                         hmp->voldata.mirror_tid,
 922                                         hmp->vchain.bref.mirror_tid);
 923                         }
 924                         hmp->volsync = hmp->voldata;
 925                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_VOLUMESYNC);
 926                         hammer2_voldata_unlock(hmp);
 927                         hammer2_chain_unlock(&hmp->fchain);
 928                         break;
 929                 case HAMMER2_BREF_TYPE_DATA:
 930                         /*
 931                          * Data elements have already been flushed via the
 932                          * logical file buffer cache.  Their hash was set in
 933                          * the bref by the vop_write code.  Do not re-dirty.
 934                          *
 935                          * Make sure any device buffer(s) have been flushed
 936                          * out here (there aren't usually any to flush) XXX.
 937                          */
 938                         break;
 939                 case HAMMER2_BREF_TYPE_INDIRECT:
 940                 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
 941                 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
 942                         /*
 943                          * Buffer I/O will be cleaned up when the volume is
 944                          * flushed (but the kernel is free to flush it before
 945                          * then, as well).
 946                          */
 947                         KKASSERT((chain->flags & HAMMER2_CHAIN_EMBEDDED) == 0);
 948                         hammer2_chain_setcheck(chain, chain->data);
 949                         break;
 950                 case HAMMER2_BREF_TYPE_DIRENT:
 951                         /*
 952                          * A directory entry can use the check area to store
 953                          * the filename for filenames <= 64 bytes, don't blow
 954                          * it up!
 955                          */
 956                         KKASSERT((chain->flags & HAMMER2_CHAIN_EMBEDDED) == 0);
 957                         if (chain->bytes)
 958                                 hammer2_chain_setcheck(chain, chain->data);
 959                         break;
 960                 case HAMMER2_BREF_TYPE_INODE:
 961                         /*
 962                          * NOTE: We must call io_setdirty() to make any late
 963                          *       changes to the inode data, the system might
 964                          *       have already flushed the buffer.
 965                          */
 966                         if (chain->data->ipdata.meta.op_flags &
 967                             HAMMER2_OPFLAG_PFSROOT) {
 968                                 /*
 969                                  * non-NULL pmp if mounted as a PFS.  We must
 970                                  * sync fields cached in the pmp? XXX
 971                                  */
 972                                 hammer2_inode_data_t *ipdata;
 973
 974                                 hammer2_io_setdirty(chain->dio);
 975                                 ipdata = &chain->data->ipdata;
 976                                 if (chain->pmp) {
 977                                         ipdata->meta.pfs_inum =
 978                                                 chain->pmp->inode_tid;
 979                                 }
 980                         } else {
 981                                 /* can't be mounted as a PFS */
 982                         }
 983
 984                         KKASSERT((chain->flags & HAMMER2_CHAIN_EMBEDDED) == 0);
 985                         hammer2_chain_setcheck(chain, chain->data);
 986
 987                                 hammer2_inode_data_t *ipdata;
 988                         ipdata = &chain->data->ipdata;
 989                         break;
 990                 default:
 991                         KKASSERT(chain->flags & HAMMER2_CHAIN_EMBEDDED);
 992                         panic("hammer2_flush_core: unsupported "
 993                               "embedded bref %d",
 994                               chain->bref.type);
 995                         /* NOT REACHED */
 996                 }
 997
 998                 /*
 999                  * If the chain was destroyed try to avoid unnecessary I/O
1000                  * that might not have yet occurred.  Remove the data range
1001                  * from dedup candidacy and attempt to invalidation that
1002                  * potentially dirty portion of the I/O buffer.
1003                  */
1004                 if (chain->flags & HAMMER2_CHAIN_DESTROY) {
1005                         hammer2_io_dedup_delete(hmp,
1006                                                 chain->bref.type,
1007                                                 chain->bref.data_off,
1008                                                 chain->bytes);
1009 #if 0
1010                         hammer2_io_t *dio;
1011                         if (chain->dio) {
1012                                 hammer2_io_inval(chain->dio,
1013                                                  chain->bref.data_off,
1014                                                  chain->bytes);
1015                         } else if ((dio = hammer2_io_getquick(hmp,
1016                                                   chain->bref.data_off,
1017                                                   chain->bytes,
1018                                                   1)) != NULL) {
1019                                 hammer2_io_inval(dio,
1020                                                  chain->bref.data_off,
1021                                                  chain->bytes);
1022                                 hammer2_io_putblk(&dio);
1023                         }
1024 #endif
1025                 }
1026         }
1027
1028         /*
1029          * If UPDATE is set the parent block table may need to be updated.
1030          * This can fail if the hammer2_chain_modify() fails.
1031          *
1032          * NOTE: UPDATE may be set on vchain or fchain in which case
1033          *       parent could be NULL.  It's easiest to allow the case
1034          *       and test for NULL.  parent can also wind up being NULL
1035          *       due to a deletion so we need to handle the case anyway.
1036          *
1037          * If no parent exists we can just clear the UPDATE bit.  If the
1038          * chain gets reattached later on the bit will simply get set
1039          * again.
1040          */
1041         if ((chain->flags & HAMMER2_CHAIN_UPDATE) && parent == NULL)
1042                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1043
1044         /*
1045          * The chain may need its blockrefs updated in the parent.
1046          */
1047         if (chain->flags & HAMMER2_CHAIN_UPDATE) {
1048                 hammer2_blockref_t *base;
1049                 int count;
1050
1051                 /*
1052                  * Clear UPDATE flag, mark parent modified, update its
1053                  * modify_tid if necessary, and adjust the parent blockmap.
1054                  */
1055                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1056
1057                 /*
1058                  * (optional code)
1059                  *
1060                  * Avoid actually modifying and updating the parent if it
1061                  * was flagged for destruction.  This can greatly reduce
1062                  * disk I/O in large tree removals because the
1063                  * hammer2_io_setinval() call in the upward recursion
1064                  * (see MODIFIED code above) can only handle a few cases.
1065                  */
1066                 if (parent->flags & HAMMER2_CHAIN_DESTROY) {
1067                         if (parent->bref.modify_tid < chain->bref.modify_tid) {
1068                                 parent->bref.modify_tid =
1069                                         chain->bref.modify_tid;
1070                         }
1071                         atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BMAPPED |
1072                                                         HAMMER2_CHAIN_BMAPUPD);
1073                         goto skipupdate;
1074                 }
1075
1076                 /*
1077                  * The flusher is responsible for deleting empty indirect
1078                  * blocks at this point.  If we don't do this, no major harm
1079                  * will be done but the empty indirect blocks will stay in
1080                  * the topology and make it a messy and inefficient.
1081                  *
1082                  * The flusher is also responsible for collapsing the
1083                  * content of an indirect block into its parent whenever
1084                  * possible (with some hysteresis).  Not doing this will also
1085                  * not harm the topology, but would make it messy and
1086                  * inefficient.
1087                  */
1088                 if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT) {
1089                         if (hammer2_chain_indirect_maintenance(parent, chain))
1090                                 goto skipupdate;
1091                 }
1092
1093                 /*
1094                  * We are updating the parent's blockmap, the parent must
1095                  * be set modified.  If this fails we re-set the UPDATE flag
1096                  * in the child.
1097                  *
1098                  * NOTE! A modification error can be ENOSPC.  We still want
1099                  *       to flush modified chains recursively, not break out,
1100                  *       so we just skip the update in this situation and
1101                  *       continue.  That is, we still need to try to clean
1102                  *       out dirty chains and buffers.
1103                  *
1104                  *       This may not help bulkfree though. XXX
1105                  */
1106                 save_error = hammer2_chain_modify(parent, 0, 0, 0);
1107                 if (save_error) {
1108                         info->error |= save_error;
1109                         kprintf("hammer2_flush: %016jx.%02x error=%08x\n",
1110                                 parent->bref.data_off, parent->bref.type,
1111                                 save_error);
1112                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1113                         goto skipupdate;
1114                 }
1115                 if (parent->bref.modify_tid < chain->bref.modify_tid)
1116                         parent->bref.modify_tid = chain->bref.modify_tid;
1117
1118                 /*
1119                  * Calculate blockmap pointer
1120                  */
1121                 switch(parent->bref.type) {
1122                 case HAMMER2_BREF_TYPE_INODE:
1123                         /*
1124                          * Access the inode's block array.  However, there is
1125                          * no block array if the inode is flagged DIRECTDATA.
1126                          */
1127                         if (parent->data &&
1128                             (parent->data->ipdata.meta.op_flags &
1129                              HAMMER2_OPFLAG_DIRECTDATA) == 0) {
1130                                 base = &parent->data->
1131                                         ipdata.u.blockset.blockref[0];
1132                         } else {
1133                                 base = NULL;
1134                         }
1135                         count = HAMMER2_SET_COUNT;
1136                         break;
1137                 case HAMMER2_BREF_TYPE_INDIRECT:
1138                 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1139                         if (parent->data)
1140                                 base = &parent->data->npdata[0];
1141                         else
1142                                 base = NULL;
1143                         count = parent->bytes / sizeof(hammer2_blockref_t);
1144                         break;
1145                 case HAMMER2_BREF_TYPE_VOLUME:
1146                         base = &chain->hmp->voldata.sroot_blockset.blockref[0];
1147                         count = HAMMER2_SET_COUNT;
1148                         break;
1149                 case HAMMER2_BREF_TYPE_FREEMAP:
1150                         base = &parent->data->npdata[0];
1151                         count = HAMMER2_SET_COUNT;
1152                         break;
1153                 default:
1154                         base = NULL;
1155                         count = 0;
1156                         panic("hammer2_flush_core: "
1157                               "unrecognized blockref type: %d",
1158                               parent->bref.type);
1159                 }
1160
1161                 /*
1162                  * Blocktable updates
1163                  *
1164                  * We synchronize pending statistics at this time.  Delta
1165                  * adjustments designated for the current and upper level
1166                  * are synchronized.
1167                  */
1168                 if (base && (chain->flags & HAMMER2_CHAIN_BMAPUPD)) {
1169                         if (chain->flags & HAMMER2_CHAIN_BMAPPED) {
1170                                 hammer2_spin_ex(&parent->core.spin);
1171                                 hammer2_base_delete(parent, base, count, chain);
1172                                 hammer2_spin_unex(&parent->core.spin);
1173                                 /* base_delete clears both bits */
1174                         } else {
1175                                 atomic_clear_int(&chain->flags,
1176                                                  HAMMER2_CHAIN_BMAPUPD);
1177                         }
1178                 }
1179                 if (base && (chain->flags & HAMMER2_CHAIN_BMAPPED) == 0) {
1180                         hammer2_spin_ex(&parent->core.spin);
1181                         hammer2_base_insert(parent, base, count,
1182                                             chain, &chain->bref);
1183                         hammer2_spin_unex(&parent->core.spin);
1184                         /* base_insert sets BMAPPED */
1185                 }
1186         }
1187 skipupdate:
1188         if (parent)
1189                 hammer2_chain_unlock(parent);
1190
1191         /*
1192          * Final cleanup after flush
1193          */
1194 done:
1195         KKASSERT(chain->refs > 0);
1196         if (hammer2_debug & 0x200) {
1197                 if (info->debug == chain)
1198                         info->debug = NULL;
1199         }
1200 }
1201
1202 /*
1203  * Flush recursion helper, called from flush_core, calls flush_core.
1204  *
1205  * Flushes the children of the caller's chain (info->parent), restricted
1206  * by sync_tid.  Set info->domodify if the child's blockref must propagate
1207  * back up to the parent.
1208  *
1209  * This function may set info->error as a side effect.
1210  *
1211  * Ripouts can move child from rbtree to dbtree or dbq but the caller's
1212  * flush scan order prevents any chains from being lost.  A child can be
1213  * executes more than once.
1214  *
1215  * WARNING! If we do not call hammer2_flush_core() we must update
1216  *          bref.mirror_tid ourselves to indicate that the flush has
1217  *          processed the child.
1218  *
1219  * WARNING! parent->core spinlock is held on entry and return.
1220  */
1221 static int
1222 hammer2_flush_recurse(hammer2_chain_t *child, void *data)
1223 {
1224         hammer2_flush_info_t *info = data;
1225         hammer2_chain_t *parent = info->parent;
1226
1227 #ifdef HAMMER2_SCAN_DEBUG
1228         ++info->scan_count;
1229         if (child->flags & HAMMER2_CHAIN_MODIFIED)
1230                 ++info->scan_mod_count;
1231         if (child->flags & HAMMER2_CHAIN_UPDATE)
1232                 ++info->scan_upd_count;
1233         if (child->flags & HAMMER2_CHAIN_ONFLUSH)
1234                 ++info->scan_onf_count;
1235 #endif
1236
1237         /*
1238          * (child can never be fchain or vchain so a special check isn't
1239          *  needed).
1240          *
1241          * We must ref the child before unlocking the spinlock.
1242          *
1243          * The caller has added a ref to the parent so we can temporarily
1244          * unlock it in order to lock the child.  However, if it no longer
1245          * winds up being the child of the parent we must skip this child.
1246          *
1247          * NOTE! chain locking errors are fatal.  They are never out-of-space
1248          *       errors.
1249          */
1250         hammer2_chain_ref(child);
1251         hammer2_spin_unex(&parent->core.spin);
1252
1253         hammer2_chain_ref_hold(parent);
1254         hammer2_chain_unlock(parent);
1255         hammer2_chain_lock(child, HAMMER2_RESOLVE_MAYBE);
1256         if (child->parent != parent) {
1257                 kprintf("LOST CHILD1 %p->%p (actual parent %p)\n",
1258                         parent, child, child->parent);
1259                 goto done;
1260         }
1261         if (child->error) {
1262                 kprintf("CHILD ERROR DURING FLUSH LOCK %p->%p\n",
1263                         parent, child);
1264                 info->error |= child->error;
1265                 goto done;
1266         }
1267
1268         /*
1269          * Must propagate the DESTROY flag downwards, otherwise the
1270          * parent could end up never being removed because it will
1271          * be requeued to the flusher if it survives this run due to
1272          * the flag.
1273          */
1274         if (parent && (parent->flags & HAMMER2_CHAIN_DESTROY))
1275                 atomic_set_int(&child->flags, HAMMER2_CHAIN_DESTROY);
1276 #ifdef HAMMER2_SCAN_DEBUG
1277         if (child->flags & HAMMER2_CHAIN_DESTROY)
1278                 ++info->scan_del_count;
1279 #endif
1280
1281         /*
1282          * Recurse and collect deferral data.  We're in the media flush,
1283          * this can cross PFS boundaries.
1284          */
1285         if (child->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1286 #ifdef HAMMER2_SCAN_DEBUG
1287                 if (child->bref.type < 7)
1288                         ++info->scan_btype[child->bref.type];
1289 #endif
1290                 ++info->depth;
1291                 hammer2_flush_core(info, child, info->flags);
1292                 --info->depth;
1293         } else if (hammer2_debug & 0x200) {
1294                 if (info->debug == NULL)
1295                         info->debug = child;
1296                 ++info->depth;
1297                 hammer2_flush_core(info, child, info->flags);
1298                 --info->depth;
1299                 if (info->debug == child)
1300                         info->debug = NULL;
1301         }
1302
1303 done:
1304         /*
1305          * Relock to continue the loop.
1306          */
1307         hammer2_chain_unlock(child);
1308         hammer2_chain_lock(parent, HAMMER2_RESOLVE_MAYBE);
1309         hammer2_chain_drop_unhold(parent);
1310         if (parent->error) {
1311                 kprintf("PARENT ERROR DURING FLUSH LOCK %p->%p\n",
1312                         parent, child);
1313                 info->error |= parent->error;
1314         }
1315         hammer2_chain_drop(child);
1316         KKASSERT(info->parent == parent);
1317         hammer2_spin_ex(&parent->core.spin);
1318
1319         return (0);
1320 }
1321
1322 /*
1323  * flush helper (backend threaded)
1324  *
1325  * Flushes chain topology for the specified inode.
1326  *
1327  * If HAMMER2_XOP_FLUSH is set we flush all chains from the current inode
1328  * through but stop at sub-inodes (we flush the inode chains for sub-inodes,
1329  * but do not go further as deeper modifications do not belong to the current
1330  * flush cycle).
1331  *
1332  * If HAMMER2_XOP_FLUSH is not set we flush the current inode's chains only
1333  * and do not recurse through sub-inodes, including not including those
1334  * sub-inodes.
1335  *
1336  * Remember that HAMMER2 is currently using a flat inode model, so directory
1337  * hierarchies do not translate to inode hierarchies.  PFS ROOTs, however,
1338  * do.
1339  *
1340  * chain->parent can be NULL, usually due to destroy races.
1341  *
1342  * Primarily called from vfs_sync().
1343  */
1344 void
1345 hammer2_inode_xop_flush(hammer2_thread_t *thr, hammer2_xop_t *arg)
1346 {
1347         hammer2_xop_flush_t *xop = &arg->xop_flush;
1348         hammer2_chain_t *chain;
1349         hammer2_chain_t *parent;
1350         hammer2_dev_t *hmp;
1351         int flush_error = 0;
1352         int fsync_error = 0;
1353         int total_error = 0;
1354         int j;
1355         int xflags;
1356         int ispfsroot = 0;
1357
1358         xflags = HAMMER2_FLUSH_TOP;
1359         if (xop->head.flags & HAMMER2_XOP_INODE_STOP)
1360                 xflags |= HAMMER2_FLUSH_INODE_STOP;
1361
1362         /*
1363          * Flush core chains
1364          */
1365         chain = hammer2_inode_chain(xop->head.ip1, thr->clindex,
1366                                     HAMMER2_RESOLVE_ALWAYS);
1367         if (chain) {
1368                 hmp = chain->hmp;
1369                 if ((chain->flags & HAMMER2_CHAIN_FLUSH_MASK) ||
1370                     TAILQ_FIRST(&hmp->flushq) != NULL) {
1371                         hammer2_flush(chain, xflags);
1372                         parent = chain->parent;
1373                         if (parent)
1374                                 hammer2_chain_setflush(parent);
1375                 }
1376                 if (chain->flags & HAMMER2_CHAIN_PFSBOUNDARY)
1377                         ispfsroot = 1;
1378                 hammer2_chain_unlock(chain);
1379                 hammer2_chain_drop(chain);
1380                 chain = NULL;
1381         } else {
1382                 hmp = NULL;
1383         }
1384
1385         /*
1386          * Don't flush from the volume root to the PFSROOT unless ip was
1387          * a PFSROOT.  If it isn't then this flush is probably related to
1388          * a VOP_FSYNC.
1389          */
1390         if (ispfsroot == 0)
1391                 goto skip;
1392
1393         /*
1394          * Flush volume roots.  Avoid replication, we only want to
1395          * flush each hammer2_dev (hmp) once.
1396          */
1397         for (j = thr->clindex - 1; j >= 0; --j) {
1398                 if ((chain = xop->head.ip1->cluster.array[j].chain) != NULL) {
1399                         if (chain->hmp == hmp) {
1400                                 chain = NULL;   /* safety */
1401                                 goto skip;
1402                         }
1403                 }
1404         }
1405         chain = NULL;   /* safety */
1406
1407         /*
1408          * spmp transaction.  The super-root is never directly mounted so
1409          * there shouldn't be any vnodes, let alone any dirty vnodes
1410          * associated with it, so we shouldn't have to mess around with any
1411          * vnode flushes here.
1412          */
1413         hammer2_trans_init(hmp->spmp, HAMMER2_TRANS_ISFLUSH);
1414
1415         /*
1416          * Media mounts have two 'roots', vchain for the topology
1417          * and fchain for the free block table.  Flush both.
1418          *
1419          * Note that the topology and free block table are handled
1420          * independently, so the free block table can wind up being
1421          * ahead of the topology.  We depend on the bulk free scan
1422          * code to deal with any loose ends.
1423          *
1424          * vchain and fchain do not error on-lock since their data does
1425          * not have to be re-read from media.
1426          */
1427         hammer2_chain_ref(&hmp->vchain);
1428         hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1429         hammer2_chain_ref(&hmp->fchain);
1430         hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1431         if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1432                 /*
1433                  * This will also modify vchain as a side effect,
1434                  * mark vchain as modified now.
1435                  */
1436                 hammer2_voldata_modify(hmp);
1437                 chain = &hmp->fchain;
1438                 flush_error |= hammer2_flush(chain, HAMMER2_FLUSH_TOP);
1439                 KKASSERT(chain == &hmp->fchain);
1440         }
1441         hammer2_chain_unlock(&hmp->fchain);
1442         hammer2_chain_unlock(&hmp->vchain);
1443         hammer2_chain_drop(&hmp->fchain);
1444         /* vchain dropped down below */
1445
1446         hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1447         if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1448                 chain = &hmp->vchain;
1449                 flush_error |= hammer2_flush(chain, HAMMER2_FLUSH_TOP);
1450                 KKASSERT(chain == &hmp->vchain);
1451         }
1452         hammer2_chain_unlock(&hmp->vchain);
1453         hammer2_chain_drop(&hmp->vchain);
1454
1455         /*
1456          * We can't safely flush the volume header until we have
1457          * flushed any device buffers which have built up.
1458          *
1459          * XXX this isn't being incremental
1460          */
1461         vn_lock(hmp->devvp, LK_EXCLUSIVE | LK_RETRY);
1462         fsync_error = VOP_FSYNC(hmp->devvp, MNT_WAIT, 0);
1463         vn_unlock(hmp->devvp);
1464         if (fsync_error || flush_error) {
1465                 kprintf("hammer2: sync error fsync=%d h2flush=0x%04x dev=%s\n",
1466                         fsync_error, flush_error, hmp->devrepname);
1467         }
1468
1469         /*
1470          * The flush code sets CHAIN_VOLUMESYNC to indicate that the
1471          * volume header needs synchronization via hmp->volsync.
1472          *
1473          * XXX synchronize the flag & data with only this flush XXX
1474          */
1475         if (fsync_error == 0 && flush_error == 0 &&
1476             (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
1477                 struct buf *bp;
1478                 int vol_error = 0;
1479
1480                 /*
1481                  * Synchronize the disk before flushing the volume
1482                  * header.
1483                  */
1484                 bp = getpbuf(NULL);
1485                 bp->b_bio1.bio_offset = 0;
1486                 bp->b_bufsize = 0;
1487                 bp->b_bcount = 0;
1488                 bp->b_cmd = BUF_CMD_FLUSH;
1489                 bp->b_bio1.bio_done = biodone_sync;
1490                 bp->b_bio1.bio_flags |= BIO_SYNC;
1491                 vn_strategy(hmp->devvp, &bp->b_bio1);
1492                 fsync_error = biowait(&bp->b_bio1, "h2vol");
1493                 relpbuf(bp, NULL);
1494
1495                 /*
1496                  * Then we can safely flush the version of the
1497                  * volume header synchronized by the flush code.
1498                  */
1499                 j = hmp->volhdrno + 1;
1500                 if (j < 0)
1501                         j = 0;
1502                 if (j >= HAMMER2_NUM_VOLHDRS)
1503                         j = 0;
1504                 if (j * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
1505                     hmp->volsync.volu_size) {
1506                         j = 0;
1507                 }
1508                 if (hammer2_debug & 0x8000) {
1509                         /* debug only, avoid syslogd loop */
1510                         kprintf("sync volhdr %d %jd\n",
1511                                 j, (intmax_t)hmp->volsync.volu_size);
1512                 }
1513                 bp = getblk(hmp->devvp, j * HAMMER2_ZONE_BYTES64,
1514                             HAMMER2_PBUFSIZE, GETBLK_KVABIO, 0);
1515                 atomic_clear_int(&hmp->vchain.flags,
1516                                  HAMMER2_CHAIN_VOLUMESYNC);
1517                 bkvasync(bp);
1518                 bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
1519                 vol_error = bwrite(bp);
1520                 hmp->volhdrno = j;
1521                 if (vol_error)
1522                         fsync_error = vol_error;
1523         }
1524         if (flush_error)
1525                 total_error = flush_error;
1526         if (fsync_error)
1527                 total_error = hammer2_errno_to_error(fsync_error);
1528
1529         hammer2_trans_done(hmp->spmp);  /* spmp trans */
1530 skip:
1531         hammer2_xop_feed(&xop->head, NULL, thr->clindex, total_error);
1532 }