sys/vfs/hammer2/hammer2_flush.c

   1 /*
   2  * Copyright (c) 2011-2018 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/vnode.h>
  52 #include <sys/buf.h>
  53
  54 #include "hammer2.h"
  55
  56 #define FLUSH_DEBUG 0
  57
  58 #define HAMMER2_FLUSH_DEPTH_LIMIT       60      /* stack recursion limit */
  59
  60
  61 /*
  62  * Recursively flush the specified chain.  The chain is locked and
  63  * referenced by the caller and will remain so on return.  The chain
  64  * will remain referenced throughout but can temporarily lose its
  65  * lock during the recursion to avoid unnecessarily stalling user
  66  * processes.
  67  */
  68 struct hammer2_flush_info {
  69         hammer2_chain_t *parent;
  70         int             depth;
  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 #endif
  81         hammer2_chain_t *debug;
  82 };
  83
  84 typedef struct hammer2_flush_info hammer2_flush_info_t;
  85
  86 static int hammer2_flush_core(hammer2_flush_info_t *info,
  87                                 hammer2_chain_t *chain, int flags);
  88 static int hammer2_flush_recurse(hammer2_chain_t *child, void *data);
  89
  90 /*
  91  * Any per-pfs transaction initialization goes here.
  92  */
  93 void
  94 hammer2_trans_manage_init(hammer2_pfs_t *pmp)
  95 {
  96 }
  97
  98 /*
  99  * Transaction support for any modifying operation.  Transactions are used
 100  * in the pmp layer by the frontend and in the spmp layer by the backend.
 101  *
 102  * 0                    - Normal transaction.  Interlocks against just the
 103  *                        COPYQ portion of an ISFLUSH transaction.
 104  *
 105  * TRANS_ISFLUSH        - Flush transaction.  Interlocks against other flush
 106  *                        transactions.
 107  *
 108  *                        When COPYQ is also specified, waits for the count
 109  *                        to drop to 1.
 110  *
 111  * TRANS_BUFCACHE       - Buffer cache transaction.  No interlock.
 112  *
 113  * TRANS_SIDEQ          - Run the sideq (only tested in trans_done())
 114  *
 115  * Initializing a new transaction allocates a transaction ID.  Typically
 116  * passed a pmp (hmp passed as NULL), indicating a cluster transaction.  Can
 117  * be passed a NULL pmp and non-NULL hmp to indicate a transaction on a single
 118  * media target.  The latter mode is used by the recovery code.
 119  */
 120 void
 121 hammer2_trans_init(hammer2_pfs_t *pmp, uint32_t flags)
 122 {
 123         uint32_t oflags;
 124         uint32_t nflags;
 125         int dowait;
 126
 127         for (;;) {
 128                 oflags = pmp->trans.flags;
 129                 cpu_ccfence();
 130                 dowait = 0;
 131
 132                 if (flags & HAMMER2_TRANS_ISFLUSH) {
 133                         /*
 134                          * Interlock against other flush transactions.
 135                          */
 136                         if (oflags & HAMMER2_TRANS_ISFLUSH) {
 137                                 nflags = oflags | HAMMER2_TRANS_WAITING;
 138                                 dowait = 1;
 139                         } else {
 140                                 nflags = (oflags | flags) + 1;
 141                         }
 142                 } else if (flags & HAMMER2_TRANS_BUFCACHE) {
 143                         /*
 144                          * Requesting strategy transaction from buffer-cache,
 145                          * or a VM getpages/putpages through the buffer cache.
 146                          * We must allow such transactions in all situations
 147                          * to avoid deadlocks.
 148                          */
 149                         nflags = (oflags | flags) + 1;
 150                 } else {
 151                         /*
 152                          * Normal transaction.  We do not interlock against
 153                          * BUFCACHE or ISFLUSH.
 154                          *
 155                          * Note that vnode locks may be held going into
 156                          * this call.
 157                          *
 158                          * NOTE: Remember that non-modifying operations
 159                          *       such as read, stat, readdir, etc, do
 160                          *       not use transactions.
 161                          */
 162                         nflags = (oflags | flags) + 1;
 163                 }
 164                 if (dowait)
 165                         tsleep_interlock(&pmp->trans.sync_wait, 0);
 166                 if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
 167                         if (dowait == 0)
 168                                 break;
 169                         tsleep(&pmp->trans.sync_wait, PINTERLOCKED,
 170                                "h2trans", hz);
 171                         /* retry */
 172                 } else {
 173                         cpu_pause();
 174                         /* retry */
 175                 }
 176                 /* retry */
 177         }
 178
 179 #if 0
 180         /*
 181          * When entering a FLUSH transaction with COPYQ set, wait for the
 182          * transaction count to drop to 1 (our flush transaction only)
 183          * before proceeding.
 184          *
 185          * This waits for all non-flush transactions to complete and blocks
 186          * new non-flush transactions from starting until COPYQ is cleared.
 187          * (the flush will then proceed after clearing COPYQ).  This should
 188          * be a very short stall on modifying operations.
 189          */
 190         while ((flags & HAMMER2_TRANS_ISFLUSH) &&
 191                (flags & HAMMER2_TRANS_COPYQ)) {
 192                 oflags = pmp->trans.flags;
 193                 cpu_ccfence();
 194                 if ((oflags & HAMMER2_TRANS_MASK) == 1)
 195                         break;
 196                 nflags = oflags | HAMMER2_TRANS_WAITING;
 197                 tsleep_interlock(&pmp->trans.sync_wait, 0);
 198                 if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
 199                         tsleep(&pmp->trans.sync_wait, PINTERLOCKED,
 200                                "h2trans2", hz);
 201                 }
 202         }
 203 #endif
 204 }
 205
 206 /*
 207  * Start a sub-transaction, there is no 'subdone' function.  This will
 208  * issue a new modify_tid (mtid) for the current transaction, which is a
 209  * CLC (cluster level change) id and not a per-node id.
 210  *
 211  * This function must be called for each XOP when multiple XOPs are run in
 212  * sequence within a transaction.
 213  *
 214  * Callers typically update the inode with the transaction mtid manually
 215  * to enforce sequencing.
 216  */
 217 hammer2_tid_t
 218 hammer2_trans_sub(hammer2_pfs_t *pmp)
 219 {
 220         hammer2_tid_t mtid;
 221
 222         mtid = atomic_fetchadd_64(&pmp->modify_tid, 1);
 223
 224         return (mtid);
 225 }
 226
 227 void
 228 hammer2_trans_setflags(hammer2_pfs_t *pmp, uint32_t flags)
 229 {
 230         atomic_set_int(&pmp->trans.flags, flags);
 231 }
 232
 233 /*
 234  * Typically used to clear trans flags asynchronously.  If TRANS_WAITING
 235  * is in the mask, and was previously set, this function will wake up
 236  * any waiters.
 237  */
 238 void
 239 hammer2_trans_clearflags(hammer2_pfs_t *pmp, uint32_t flags)
 240 {
 241         uint32_t oflags;
 242         uint32_t nflags;
 243
 244         for (;;) {
 245                 oflags = pmp->trans.flags;
 246                 cpu_ccfence();
 247                 nflags = oflags & ~flags;
 248                 if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
 249                         if ((oflags ^ nflags) & HAMMER2_TRANS_WAITING)
 250                                 wakeup(&pmp->trans.sync_wait);
 251                         break;
 252                 }
 253                 cpu_pause();
 254                 /* retry */
 255         }
 256 }
 257
 258 void
 259 hammer2_trans_done(hammer2_pfs_t *pmp, uint32_t flags)
 260 {
 261         uint32_t oflags;
 262         uint32_t nflags;
 263
 264 #if 0
 265         /*
 266          * Modifying ops on the front-end can cause dirty inodes to
 267          * build up in the sideq.  We don't flush these on inactive/reclaim
 268          * due to potential deadlocks, so we have to deal with them from
 269          * inside other nominal modifying front-end transactions.
 270          */
 271         if ((flags & HAMMER2_TRANS_SIDEQ) &&
 272             pmp->sideq_count > hammer2_limit_dirty_inodes / 2 &&
 273             pmp->sideq_count > (pmp->inum_count >> 3) &&
 274             pmp->mp) {
 275                 speedup_syncer(pmp->mp);
 276         }
 277 #endif
 278
 279         /*
 280          * Clean-up the transaction.  Wakeup any waiters when finishing
 281          * a flush transaction or transitioning the non-flush transaction
 282          * count from 2->1 while a flush transaction is pending.
 283          */
 284         for (;;) {
 285                 oflags = pmp->trans.flags;
 286                 cpu_ccfence();
 287                 KKASSERT(oflags & HAMMER2_TRANS_MASK);
 288
 289                 nflags = (oflags - 1) & ~flags;
 290                 if (flags & HAMMER2_TRANS_ISFLUSH) {
 291                         nflags &= ~HAMMER2_TRANS_WAITING;
 292                 }
 293                 if ((oflags & (HAMMER2_TRANS_ISFLUSH|HAMMER2_TRANS_MASK)) ==
 294                     (HAMMER2_TRANS_ISFLUSH|2)) {
 295                         nflags &= ~HAMMER2_TRANS_WAITING;
 296                 }
 297                 if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
 298                         if ((oflags ^ nflags) & HAMMER2_TRANS_WAITING)
 299                                 wakeup(&pmp->trans.sync_wait);
 300                         break;
 301                 }
 302                 cpu_pause();
 303                 /* retry */
 304         }
 305 }
 306
 307 /*
 308  * Obtain new, unique inode number (not serialized by caller).
 309  */
 310 hammer2_tid_t
 311 hammer2_trans_newinum(hammer2_pfs_t *pmp)
 312 {
 313         hammer2_tid_t tid;
 314
 315         tid = atomic_fetchadd_64(&pmp->inode_tid, 1);
 316
 317         return tid;
 318 }
 319
 320 /*
 321  * Assert that a strategy call is ok here.  Currently we allow strategy
 322  * calls in all situations, including during flushes.  Previously:
 323  *      (old) (1) In a normal transaction.
 324  *      (old) (2) In a flush transaction only if PREFLUSH is also set.
 325  */
 326 void
 327 hammer2_trans_assert_strategy(hammer2_pfs_t *pmp)
 328 {
 329 #if 0
 330         KKASSERT((pmp->trans.flags & HAMMER2_TRANS_ISFLUSH) == 0 ||
 331                  (pmp->trans.flags & HAMMER2_TRANS_PREFLUSH));
 332 #endif
 333 }
 334
 335 /*
 336  * Flush the chain and all modified sub-chains through the specified
 337  * synchronization point, propagating blockref updates back up.  As
 338  * part of this propagation, mirror_tid and inode/data usage statistics
 339  * propagates back upward.
 340  *
 341  * Returns a HAMMER2 error code, 0 if no error.  Note that I/O errors from
 342  * buffers dirtied during the flush operation can occur later.
 343  *
 344  * modify_tid (clc - cluster level change) is not propagated.
 345  *
 346  * update_tid (clc) is used for validation and is not propagated by this
 347  * function.
 348  *
 349  * This routine can be called from several places but the most important
 350  * is from VFS_SYNC (frontend) via hammer2_xop_inode_flush (backend).
 351  *
 352  * chain is locked on call and will remain locked on return.  The chain's
 353  * UPDATE flag indicates that its parent's block table (which is not yet
 354  * part of the flush) should be updated.
 355  *
 356  * flags:
 357  *      HAMMER2_FLUSH_TOP       Indicates that this is the top of the flush.
 358  *                              Is cleared for the recursion.
 359  *
 360  *      HAMMER2_FLUSH_ALL       Recurse everything
 361  *
 362  *      HAMMER2_FLUSH_INODE_STOP
 363  *                              Stop at PFS inode or normal inode boundary
 364  */
 365 int
 366 hammer2_flush(hammer2_chain_t *chain, int flags)
 367 {
 368         hammer2_flush_info_t info;
 369         hammer2_dev_t *hmp;
 370         int loops;
 371
 372         /*
 373          * Execute the recursive flush and handle deferrals.
 374          *
 375          * Chains can be ridiculously long (thousands deep), so to
 376          * avoid blowing out the kernel stack the recursive flush has a
 377          * depth limit.  Elements at the limit are placed on a list
 378          * for re-execution after the stack has been popped.
 379          */
 380         bzero(&info, sizeof(info));
 381         info.flags = flags & ~HAMMER2_FLUSH_TOP;
 382
 383         /*
 384          * Calculate parent (can be NULL), if not NULL the flush core
 385          * expects the parent to be referenced so it can easily lock/unlock
 386          * it without it getting ripped up.
 387          */
 388         if ((info.parent = chain->parent) != NULL)
 389                 hammer2_chain_ref(info.parent);
 390
 391         /*
 392          * Extra ref needed because flush_core expects it when replacing
 393          * chain.
 394          */
 395         hammer2_chain_ref(chain);
 396         hmp = chain->hmp;
 397         loops = 0;
 398
 399         for (;;) {
 400                 /*
 401                  * [re]flush chain as the deep recursion may have generated
 402                  * additional modifications.
 403                  */
 404                 if (info.parent != chain->parent) {
 405                         if (hammer2_debug & 0x0040) {
 406                                 kprintf("LOST CHILD4 %p->%p "
 407                                         "(actual parent %p)\n",
 408                                         info.parent, chain, chain->parent);
 409                         }
 410                         hammer2_chain_drop(info.parent);
 411                         info.parent = chain->parent;
 412                         hammer2_chain_ref(info.parent);
 413                 }
 414                 if (hammer2_flush_core(&info, chain, flags) == 0)
 415                         break;
 416
 417                 if (++loops % 1000 == 0) {
 418                         kprintf("hammer2_flush: excessive loops on %p\n",
 419                                 chain);
 420                         if (hammer2_debug & 0x100000)
 421                                 Debugger("hell4");
 422                 }
 423         }
 424 #ifdef HAMMER2_SCAN_DEBUG
 425         if (info.scan_count >= 10)
 426         kprintf("hammer2_flush: scan_count %ld (%ld,%ld,%ld,%ld) "
 427                 "bt(%ld,%ld,%ld,%ld,%ld,%ld)\n",
 428                 info.scan_count,
 429                 info.scan_mod_count,
 430                 info.scan_upd_count,
 431                 info.scan_onf_count,
 432                 info.scan_del_count,
 433                 info.scan_btype[1],
 434                 info.scan_btype[2],
 435                 info.scan_btype[3],
 436                 info.scan_btype[4],
 437                 info.scan_btype[5],
 438                 info.scan_btype[6]);
 439 #endif
 440         hammer2_chain_drop(chain);
 441         if (info.parent)
 442                 hammer2_chain_drop(info.parent);
 443         return (info.error);
 444 }
 445
 446 /*
 447  * This is the core of the chain flushing code.  The chain is locked by the
 448  * caller and must also have an extra ref on it by the caller, and remains
 449  * locked and will have an extra ref on return.  info.parent is referenced
 450  * but not locked.
 451  *
 452  * Upon return, the caller can test the UPDATE bit on the chain to determine
 453  * if the parent needs updating.
 454  *
 455  * If non-zero is returned, the chain's parent changed during the flush and
 456  * the caller must retry the operation.
 457  *
 458  * (1) Determine if this node is a candidate for the flush, return if it is
 459  *     not.  fchain and vchain are always candidates for the flush.
 460  *
 461  * (2) If we recurse too deep the chain is entered onto the deferral list and
 462  *     the current flush stack is aborted until after the deferral list is
 463  *     run.
 464  *
 465  * (3) Recursively flush live children (rbtree).  This can create deferrals.
 466  *     A successful flush clears the MODIFIED and UPDATE bits on the children
 467  *     and typically causes the parent to be marked MODIFIED as the children
 468  *     update the parent's block table.  A parent might already be marked
 469  *     MODIFIED due to a deletion (whos blocktable update in the parent is
 470  *     handled by the frontend), or if the parent itself is modified by the
 471  *     frontend for other reasons.
 472  *
 473  * (4) Permanently disconnected sub-trees are cleaned up by the front-end.
 474  *     Deleted-but-open inodes can still be individually flushed via the
 475  *     filesystem syncer.
 476  *
 477  * (5) Delete parents on the way back up if they are normal indirect blocks
 478  *     and have no children.
 479  *
 480  * (6) Note that an unmodified child may still need the block table in its
 481  *     parent updated (e.g. rename/move).  The child will have UPDATE set
 482  *     in this case.
 483  *
 484  *                      WARNING ON BREF MODIFY_TID/MIRROR_TID
 485  *
 486  * blockref.modify_tid is consistent only within a PFS, and will not be
 487  * consistent during synchronization.  mirror_tid is consistent across the
 488  * block device regardless of the PFS.
 489  */
 490 static int
 491 hammer2_flush_core(hammer2_flush_info_t *info, hammer2_chain_t *chain,
 492                    int flags)
 493 {
 494         hammer2_chain_t *parent;
 495         hammer2_dev_t *hmp;
 496         int save_error;
 497         int retry;
 498
 499         retry = 0;
 500
 501         /*
 502          * (1) Optimize downward recursion to locate nodes needing action.
 503          *     Nothing to do if none of these flags are set.
 504          */
 505         if ((chain->flags & HAMMER2_CHAIN_FLUSH_MASK) == 0) {
 506                 if (hammer2_debug & 0x200) {
 507                         if (info->debug == NULL)
 508                                 info->debug = chain;
 509                 } else {
 510                         return 0;
 511                 }
 512         }
 513
 514         hmp = chain->hmp;
 515
 516         /*
 517          * NOTE: parent can be NULL, usually due to destroy races.
 518          */
 519         parent = info->parent;
 520         KKASSERT(chain->parent == parent);
 521
 522         /*
 523          * Downward search recursion
 524          *
 525          * We must be careful on cold stops, which often occur on inode
 526          * boundaries due to the way hammer2_vfs_sync() sequences the flush.
 527          * Be sure to issue an appropriate chain_setflush()
 528          */
 529         if ((chain->flags & HAMMER2_CHAIN_PFSBOUNDARY) &&
 530             (flags & HAMMER2_FLUSH_ALL) == 0 &&
 531             (flags & HAMMER2_FLUSH_TOP) == 0 &&
 532             chain->pmp && chain->pmp->mp) {
 533                 /*
 534                  * If FLUSH_ALL is not specified the caller does not want
 535                  * to recurse through PFS roots that have been mounted.
 536                  *
 537                  * (If the PFS has not been mounted there may not be
 538                  *  anything monitoring its chains and its up to us
 539                  *  to flush it).
 540                  *
 541                  * The typical sequence is to flush dirty PFS's starting at
 542                  * their root downward, then flush the device root (vchain).
 543                  * It is this second flush that typically leaves out the
 544                  * ALL flag.
 545                  *
 546                  * However we must still process the PFSROOT chains for block
 547                  * table updates in their parent (which IS part of our flush).
 548                  *
 549                  * NOTE: The volume root, vchain, does not set PFSBOUNDARY.
 550                  *
 551                  * NOTE: We must re-set ONFLUSH in the parent to retain if
 552                  *       this chain (that we are skipping) requires work.
 553                  */
 554                 if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
 555                                     HAMMER2_CHAIN_DESTROY |
 556                                     HAMMER2_CHAIN_MODIFIED)) {
 557                         hammer2_chain_setflush(parent);
 558                 }
 559                 goto done;
 560         } else if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
 561                    (flags & HAMMER2_FLUSH_INODE_STOP) &&
 562                    (flags & HAMMER2_FLUSH_ALL) == 0 &&
 563                    (flags & HAMMER2_FLUSH_TOP) == 0 &&
 564                    chain->pmp && chain->pmp->mp) {
 565                 /*
 566                  * When FLUSH_INODE_STOP is specified we are being asked not
 567                  * to include any inode changes for inodes we encounter,
 568                  * with the exception of the inode that the flush began with.
 569                  * So: INODE, INODE_STOP, and TOP==0 basically.
 570                  *
 571                  * Dirty inodes are flushed based on the hammer2_inode
 572                  * in-memory structure, issuing a chain_setflush() here
 573                  * will only cause unnecessary traversals of the topology.
 574                  */
 575                 goto done;
 576 #if 0
 577                 /*
 578                  * If FLUSH_INODE_STOP is specified and both ALL and TOP
 579                  * are clear, we must not flush the chain.  The chain should
 580                  * have already been flushed and any further ONFLUSH/UPDATE
 581                  * setting will be related to the next flush.
 582                  *
 583                  * This features allows us to flush inodes independently of
 584                  * each other and meta-data above the inodes separately.
 585                  */
 586                 if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
 587                                     HAMMER2_CHAIN_DESTROY |
 588                                     HAMMER2_CHAIN_MODIFIED)) {
 589                         if (parent)
 590                                 hammer2_chain_setflush(parent);
 591                 }
 592 #endif
 593         } else if (info->depth == HAMMER2_FLUSH_DEPTH_LIMIT) {
 594                 /*
 595                  * Recursion depth reached.
 596                  */
 597                 panic("hammer2: flush depth limit");
 598         } else if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
 599                                    HAMMER2_CHAIN_DESTROY)) {
 600                 /*
 601                  * Downward recursion search (actual flush occurs bottom-up).
 602                  * pre-clear ONFLUSH.  It can get set again due to races or
 603                  * flush errors, which we want so the scan finds us again in
 604                  * the next flush.
 605                  *
 606                  * We must also recurse if DESTROY is set so we can finally
 607                  * get rid of the related children, otherwise the node will
 608                  * just get re-flushed on lastdrop.
 609                  *
 610                  * WARNING!  The recursion will unlock/relock info->parent
 611                  *           (which is 'chain'), potentially allowing it
 612                  *           to be ripped up.
 613                  */
 614                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
 615                 save_error = info->error;
 616                 info->error = 0;
 617                 info->parent = chain;
 618
 619                 /*
 620                  * We may have to do this twice to catch any indirect
 621                  * block maintenance that occurs.
 622                  */
 623                 hammer2_spin_ex(&chain->core.spin);
 624                 RB_SCAN(hammer2_chain_tree, &chain->core.rbtree,
 625                         NULL, hammer2_flush_recurse, info);
 626                 if (chain->flags & HAMMER2_CHAIN_ONFLUSH) {
 627                         atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
 628                         RB_SCAN(hammer2_chain_tree, &chain->core.rbtree,
 629                                 NULL, hammer2_flush_recurse, info);
 630                 }
 631                 hammer2_spin_unex(&chain->core.spin);
 632                 info->parent = parent;
 633
 634                 /*
 635                  * Re-set the flush bits if the flush was incomplete or
 636                  * an error occurred.  If an error occurs it is typically
 637                  * an allocation error.  Errors do not cause deferrals.
 638                  */
 639                 if (info->error)
 640                         hammer2_chain_setflush(chain);
 641                 info->error |= save_error;
 642
 643                 /*
 644                  * If we lost the parent->chain association we have to
 645                  * stop processing this chain because it is no longer
 646                  * in this recursion.  If it moved, it will be handled
 647                  * by the ONFLUSH flag elsewhere.
 648                  */
 649                 if (chain->parent != parent) {
 650                         kprintf("LOST CHILD2 %p->%p (actual parent %p)\n",
 651                                 parent, chain, chain->parent);
 652                         goto done;
 653                 }
 654         }
 655
 656         /*
 657          * Now we are in the bottom-up part of the recursion.
 658          *
 659          * We continue to try to update the chain on lower-level errors, but
 660          * the flush code may decide not to flush the volume root.
 661          *
 662          * XXX should we continue to try to update the chain if an error
 663          *     occurred?
 664          */
 665
 666         /*
 667          * Both parent and chain must be locked in order to flush chain,
 668          * in order to properly update the parent under certain conditions.
 669          *
 670          * In addition, we can't safely unlock/relock the chain once we
 671          * start flushing the chain itself, which we would have to do later
 672          * on in order to lock the parent if we didn't do that now.
 673          */
 674         hammer2_chain_ref_hold(chain);
 675         hammer2_chain_unlock(chain);
 676         if (parent)
 677                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
 678         hammer2_chain_lock(chain, HAMMER2_RESOLVE_MAYBE);
 679         hammer2_chain_drop_unhold(chain);
 680
 681         /*
 682          * Can't process if we can't access their content.
 683          */
 684         if ((parent && parent->error) || chain->error) {
 685                 kprintf("hammer2: chain error during flush\n");
 686                 info->error |= chain->error;
 687                 if (parent) {
 688                         info->error |= parent->error;
 689                         hammer2_chain_unlock(parent);
 690                 }
 691                 goto done;
 692         }
 693
 694         if (chain->parent != parent) {
 695                 if (hammer2_debug & 0x0040) {
 696                         kprintf("LOST CHILD3 %p->%p (actual parent %p)\n",
 697                                 parent, chain, chain->parent);
 698                 }
 699                 KKASSERT(parent != NULL);
 700                 hammer2_chain_unlock(parent);
 701                 retry = 1;
 702                 goto done;
 703         }
 704
 705         /*
 706          * Propagate the DESTROY flag downwards.  This dummies up the flush
 707          * code and tries to invalidate related buffer cache buffers to
 708          * avoid the disk write.
 709          */
 710         if (parent && (parent->flags & HAMMER2_CHAIN_DESTROY))
 711                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);
 712
 713         /*
 714          * Dispose of the modified bit.
 715          *
 716          * If parent is present, the UPDATE bit should already be set.
 717          * UPDATE should already be set.
 718          * bref.mirror_tid should already be set.
 719          */
 720         if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
 721                 KKASSERT((chain->flags & HAMMER2_CHAIN_UPDATE) ||
 722                          chain->parent == NULL);
 723                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
 724                 atomic_add_long(&hammer2_count_modified_chains, -1);
 725
 726                 /*
 727                  * Manage threads waiting for excessive dirty memory to
 728                  * be retired.
 729                  */
 730                 if (chain->pmp)
 731                         hammer2_pfs_memory_wakeup(chain->pmp, -1);
 732
 733 #if 0
 734                 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0 &&
 735                     chain != &hmp->vchain &&
 736                     chain != &hmp->fchain) {
 737                         /*
 738                          * Set UPDATE bit indicating that the parent block
 739                          * table requires updating.
 740                          */
 741                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
 742                 }
 743 #endif
 744
 745                 /*
 746                  * Issue the flush.  This is indirect via the DIO.
 747                  *
 748                  * NOTE: A DELETED node that reaches this point must be
 749                  *       flushed for synchronization point consistency.
 750                  *
 751                  * NOTE: Even though MODIFIED was already set, the related DIO
 752                  *       might not be dirty due to a system buffer cache
 753                  *       flush and must be set dirty if we are going to make
 754                  *       further modifications to the buffer.  Chains with
 755                  *       embedded data don't need this.
 756                  */
 757                 if (hammer2_debug & 0x1000) {
 758                         kprintf("Flush %p.%d %016jx/%d data=%016jx\n",
 759                                 chain, chain->bref.type,
 760                                 (uintmax_t)chain->bref.key,
 761                                 chain->bref.keybits,
 762                                 (uintmax_t)chain->bref.data_off);
 763                 }
 764
 765                 /*
 766                  * Update chain CRCs for flush.
 767                  *
 768                  * NOTE: Volume headers are NOT flushed here as they require
 769                  *       special processing.
 770                  */
 771                 switch(chain->bref.type) {
 772                 case HAMMER2_BREF_TYPE_FREEMAP:
 773                         /*
 774                          * Update the volume header's freemap_tid to the
 775                          * freemap's flushing mirror_tid.
 776                          *
 777                          * (note: embedded data, do not call setdirty)
 778                          */
 779                         KKASSERT(hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED);
 780                         KKASSERT(chain == &hmp->fchain);
 781                         hmp->voldata.freemap_tid = chain->bref.mirror_tid;
 782                         if (hammer2_debug & 0x8000) {
 783                                 /* debug only, avoid syslogd loop */
 784                                 kprintf("sync freemap mirror_tid %08jx\n",
 785                                         (intmax_t)chain->bref.mirror_tid);
 786                         }
 787
 788                         /*
 789                          * The freemap can be flushed independently of the
 790                          * main topology, but for the case where it is
 791                          * flushed in the same transaction, and flushed
 792                          * before vchain (a case we want to allow for
 793                          * performance reasons), make sure modifications
 794                          * made during the flush under vchain use a new
 795                          * transaction id.
 796                          *
 797                          * Otherwise the mount recovery code will get confused.
 798                          */
 799                         ++hmp->voldata.mirror_tid;
 800                         break;
 801                 case HAMMER2_BREF_TYPE_VOLUME:
 802                         /*
 803                          * The free block table is flushed by
 804                          * hammer2_vfs_sync() before it flushes vchain.
 805                          * We must still hold fchain locked while copying
 806                          * voldata to volsync, however.
 807                          *
 808                          * These do not error per-say since their data does
 809                          * not need to be re-read from media on lock.
 810                          *
 811                          * (note: embedded data, do not call setdirty)
 812                          */
 813                         hammer2_chain_lock(&hmp->fchain,
 814                                            HAMMER2_RESOLVE_ALWAYS);
 815                         hammer2_voldata_lock(hmp);
 816                         if (hammer2_debug & 0x8000) {
 817                                 /* debug only, avoid syslogd loop */
 818                                 kprintf("sync volume  mirror_tid %08jx\n",
 819                                         (intmax_t)chain->bref.mirror_tid);
 820                         }
 821
 822                         /*
 823                          * Update the volume header's mirror_tid to the
 824                          * main topology's flushing mirror_tid.  It is
 825                          * possible that voldata.mirror_tid is already
 826                          * beyond bref.mirror_tid due to the bump we made
 827                          * above in BREF_TYPE_FREEMAP.
 828                          */
 829                         if (hmp->voldata.mirror_tid < chain->bref.mirror_tid) {
 830                                 hmp->voldata.mirror_tid =
 831                                         chain->bref.mirror_tid;
 832                         }
 833
 834                         /*
 835                          * The volume header is flushed manually by the
 836                          * syncer, not here.  All we do here is adjust the
 837                          * crc's.
 838                          */
 839                         KKASSERT(chain->data != NULL);
 840                         KKASSERT(chain->dio == NULL);
 841
 842                         hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT1]=
 843                                 hammer2_icrc32(
 844                                         (char *)&hmp->voldata +
 845                                          HAMMER2_VOLUME_ICRC1_OFF,
 846                                         HAMMER2_VOLUME_ICRC1_SIZE);
 847                         hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT0]=
 848                                 hammer2_icrc32(
 849                                         (char *)&hmp->voldata +
 850                                          HAMMER2_VOLUME_ICRC0_OFF,
 851                                         HAMMER2_VOLUME_ICRC0_SIZE);
 852                         hmp->voldata.icrc_volheader =
 853                                 hammer2_icrc32(
 854                                         (char *)&hmp->voldata +
 855                                          HAMMER2_VOLUME_ICRCVH_OFF,
 856                                         HAMMER2_VOLUME_ICRCVH_SIZE);
 857
 858                         if (hammer2_debug & 0x8000) {
 859                                 /* debug only, avoid syslogd loop */
 860                                 kprintf("syncvolhdr %016jx %016jx\n",
 861                                         hmp->voldata.mirror_tid,
 862                                         hmp->vchain.bref.mirror_tid);
 863                         }
 864                         hmp->volsync = hmp->voldata;
 865                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_VOLUMESYNC);
 866                         hammer2_voldata_unlock(hmp);
 867                         hammer2_chain_unlock(&hmp->fchain);
 868                         break;
 869                 case HAMMER2_BREF_TYPE_DATA:
 870                         /*
 871                          * Data elements have already been flushed via the
 872                          * logical file buffer cache.  Their hash was set in
 873                          * the bref by the vop_write code.  Do not re-dirty.
 874                          *
 875                          * Make sure any device buffer(s) have been flushed
 876                          * out here (there aren't usually any to flush) XXX.
 877                          */
 878                         break;
 879                 case HAMMER2_BREF_TYPE_INDIRECT:
 880                 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
 881                 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
 882                         /*
 883                          * Buffer I/O will be cleaned up when the volume is
 884                          * flushed (but the kernel is free to flush it before
 885                          * then, as well).
 886                          */
 887                         hammer2_chain_setcheck(chain, chain->data);
 888                         break;
 889                 case HAMMER2_BREF_TYPE_DIRENT:
 890                         /*
 891                          * A directory entry can use the check area to store
 892                          * the filename for filenames <= 64 bytes, don't blow
 893                          * it up!
 894                          */
 895                         if (chain->bytes)
 896                                 hammer2_chain_setcheck(chain, chain->data);
 897                         break;
 898                 case HAMMER2_BREF_TYPE_INODE:
 899                         /*
 900                          * NOTE: We must call io_setdirty() to make any late
 901                          *       changes to the inode data, the system might
 902                          *       have already flushed the buffer.
 903                          */
 904                         if (chain->data->ipdata.meta.op_flags &
 905                             HAMMER2_OPFLAG_PFSROOT) {
 906                                 /*
 907                                  * non-NULL pmp if mounted as a PFS.  We must
 908                                  * sync fields cached in the pmp? XXX
 909                                  */
 910                                 hammer2_inode_data_t *ipdata;
 911
 912                                 hammer2_io_setdirty(chain->dio);
 913                                 ipdata = &chain->data->ipdata;
 914                                 if (chain->pmp) {
 915                                         ipdata->meta.pfs_inum =
 916                                                 chain->pmp->inode_tid;
 917                                 }
 918                         } else {
 919                                 /* can't be mounted as a PFS */
 920                         }
 921
 922                         hammer2_chain_setcheck(chain, chain->data);
 923                         break;
 924                 default:
 925                         panic("hammer2_flush_core: unsupported "
 926                               "embedded bref %d",
 927                               chain->bref.type);
 928                         /* NOT REACHED */
 929                 }
 930
 931                 /*
 932                  * If the chain was destroyed try to avoid unnecessary I/O
 933                  * that might not have yet occurred.  Remove the data range
 934                  * from dedup candidacy and attempt to invalidation that
 935                  * potentially dirty portion of the I/O buffer.
 936                  */
 937                 if (chain->flags & HAMMER2_CHAIN_DESTROY) {
 938                         hammer2_io_dedup_delete(hmp,
 939                                                 chain->bref.type,
 940                                                 chain->bref.data_off,
 941                                                 chain->bytes);
 942 #if 0
 943                         hammer2_io_t *dio;
 944                         if (chain->dio) {
 945                                 hammer2_io_inval(chain->dio,
 946                                                  chain->bref.data_off,
 947                                                  chain->bytes);
 948                         } else if ((dio = hammer2_io_getquick(hmp,
 949                                                   chain->bref.data_off,
 950                                                   chain->bytes,
 951                                                   1)) != NULL) {
 952                                 hammer2_io_inval(dio,
 953                                                  chain->bref.data_off,
 954                                                  chain->bytes);
 955                                 hammer2_io_putblk(&dio);
 956                         }
 957 #endif
 958                 }
 959         }
 960
 961         /*
 962          * If UPDATE is set the parent block table may need to be updated.
 963          * This can fail if the hammer2_chain_modify() fails.
 964          *
 965          * NOTE: UPDATE may be set on vchain or fchain in which case
 966          *       parent could be NULL, or on an inode that has not yet
 967          *       been inserted into the radix tree.  It's easiest to allow
 968          *       the case and test for NULL.  parent can also wind up being
 969          *       NULL due to a deletion so we need to handle the case anyway.
 970          *
 971          * NOTE: UPDATE can be set when chains are renamed into or out of
 972          *       an indirect block, without the chain itself being flagged
 973          *       MODIFIED.
 974          *
 975          * If no parent exists we can just clear the UPDATE bit.  If the
 976          * chain gets reattached later on the bit will simply get set
 977          * again.
 978          */
 979         if ((chain->flags & HAMMER2_CHAIN_UPDATE) && parent == NULL)
 980                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
 981
 982         /*
 983          * When flushing an inode outside of a FLUSH_FSSYNC we must NOT
 984          * update the parent block table to point at the flushed inode.
 985          * The block table should only ever be updated by the filesystem
 986          * sync code.  If we do, inode<->inode dependencies (such as
 987          * directory entries vs inode nlink count) can wind up not being
 988          * flushed together and result in a broken topology if a crash/reboot
 989          * occurs at the wrong time.
 990          */
 991         if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
 992             (flags & HAMMER2_FLUSH_INODE_STOP) &&
 993             (flags & HAMMER2_FLUSH_FSSYNC) == 0 &&
 994             (flags & HAMMER2_FLUSH_ALL) == 0 &&
 995             chain->pmp && chain->pmp->mp) {
 996 #ifdef HAMMER2_DEBUG_SYNC
 997                 kprintf("inum %ld do not update parent, non-fssync\n",
 998                         (long)chain->bref.key);
 999 #endif
1000                 goto skipupdate;
1001         }
1002 #ifdef HAMMER2_DEBUG_SYNC
1003         if (chain->bref.type == HAMMER2_BREF_TYPE_INODE)
1004                 kprintf("inum %ld update parent\n", (long)chain->bref.key);
1005 #endif
1006
1007         /*
1008          * The chain may need its blockrefs updated in the parent, normal
1009          * path.
1010          */
1011         if (chain->flags & HAMMER2_CHAIN_UPDATE) {
1012                 hammer2_blockref_t *base;
1013                 int count;
1014
1015                 /*
1016                  * Clear UPDATE flag, mark parent modified, update its
1017                  * modify_tid if necessary, and adjust the parent blockmap.
1018                  */
1019                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1020
1021                 /*
1022                  * (optional code)
1023                  *
1024                  * Avoid actually modifying and updating the parent if it
1025                  * was flagged for destruction.  This can greatly reduce
1026                  * disk I/O in large tree removals because the
1027                  * hammer2_io_setinval() call in the upward recursion
1028                  * (see MODIFIED code above) can only handle a few cases.
1029                  */
1030                 if (parent->flags & HAMMER2_CHAIN_DESTROY) {
1031                         if (parent->bref.modify_tid < chain->bref.modify_tid) {
1032                                 parent->bref.modify_tid =
1033                                         chain->bref.modify_tid;
1034                         }
1035                         atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BMAPPED |
1036                                                         HAMMER2_CHAIN_BMAPUPD);
1037                         goto skipupdate;
1038                 }
1039
1040                 /*
1041                  * The flusher is responsible for deleting empty indirect
1042                  * blocks at this point.  If we don't do this, no major harm
1043                  * will be done but the empty indirect blocks will stay in
1044                  * the topology and make it a messy and inefficient.
1045                  *
1046                  * The flusher is also responsible for collapsing the
1047                  * content of an indirect block into its parent whenever
1048                  * possible (with some hysteresis).  Not doing this will also
1049                  * not harm the topology, but would make it messy and
1050                  * inefficient.
1051                  */
1052                 if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT) {
1053                         if (hammer2_chain_indirect_maintenance(parent, chain))
1054                                 goto skipupdate;
1055                 }
1056
1057                 /*
1058                  * We are updating the parent's blockmap, the parent must
1059                  * be set modified.  If this fails we re-set the UPDATE flag
1060                  * in the child.
1061                  *
1062                  * NOTE! A modification error can be ENOSPC.  We still want
1063                  *       to flush modified chains recursively, not break out,
1064                  *       so we just skip the update in this situation and
1065                  *       continue.  That is, we still need to try to clean
1066                  *       out dirty chains and buffers.
1067                  *
1068                  *       This may not help bulkfree though. XXX
1069                  */
1070                 save_error = hammer2_chain_modify(parent, 0, 0, 0);
1071                 if (save_error) {
1072                         info->error |= save_error;
1073                         kprintf("hammer2_flush: %016jx.%02x error=%08x\n",
1074                                 parent->bref.data_off, parent->bref.type,
1075                                 save_error);
1076                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
1077                         goto skipupdate;
1078                 }
1079                 if (parent->bref.modify_tid < chain->bref.modify_tid)
1080                         parent->bref.modify_tid = chain->bref.modify_tid;
1081
1082                 /*
1083                  * Calculate blockmap pointer
1084                  */
1085                 switch(parent->bref.type) {
1086                 case HAMMER2_BREF_TYPE_INODE:
1087                         /*
1088                          * Access the inode's block array.  However, there is
1089                          * no block array if the inode is flagged DIRECTDATA.
1090                          */
1091                         if (parent->data &&
1092                             (parent->data->ipdata.meta.op_flags &
1093                              HAMMER2_OPFLAG_DIRECTDATA) == 0) {
1094                                 base = &parent->data->
1095                                         ipdata.u.blockset.blockref[0];
1096                         } else {
1097                                 base = NULL;
1098                         }
1099                         count = HAMMER2_SET_COUNT;
1100                         break;
1101                 case HAMMER2_BREF_TYPE_INDIRECT:
1102                 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1103                         if (parent->data)
1104                                 base = &parent->data->npdata[0];
1105                         else
1106                                 base = NULL;
1107                         count = parent->bytes / sizeof(hammer2_blockref_t);
1108                         break;
1109                 case HAMMER2_BREF_TYPE_VOLUME:
1110                         base = &chain->hmp->voldata.sroot_blockset.blockref[0];
1111                         count = HAMMER2_SET_COUNT;
1112                         break;
1113                 case HAMMER2_BREF_TYPE_FREEMAP:
1114                         base = &parent->data->npdata[0];
1115                         count = HAMMER2_SET_COUNT;
1116                         break;
1117                 default:
1118                         base = NULL;
1119                         count = 0;
1120                         panic("hammer2_flush_core: "
1121                               "unrecognized blockref type: %d",
1122                               parent->bref.type);
1123                         break;
1124                 }
1125
1126                 /*
1127                  * Blocktable updates
1128                  *
1129                  * We synchronize pending statistics at this time.  Delta
1130                  * adjustments designated for the current and upper level
1131                  * are synchronized.
1132                  */
1133                 if (base && (chain->flags & HAMMER2_CHAIN_BMAPUPD)) {
1134                         if (chain->flags & HAMMER2_CHAIN_BMAPPED) {
1135                                 hammer2_spin_ex(&parent->core.spin);
1136                                 hammer2_base_delete(parent, base, count, chain,
1137                                                     NULL);
1138                                 hammer2_spin_unex(&parent->core.spin);
1139                                 /* base_delete clears both bits */
1140                         } else {
1141                                 atomic_clear_int(&chain->flags,
1142                                                  HAMMER2_CHAIN_BMAPUPD);
1143                         }
1144                 }
1145                 if (base && (chain->flags & HAMMER2_CHAIN_BMAPPED) == 0) {
1146                         hammer2_spin_ex(&parent->core.spin);
1147                         hammer2_base_insert(parent, base, count,
1148                                             chain, &chain->bref);
1149                         hammer2_spin_unex(&parent->core.spin);
1150                         /* base_insert sets BMAPPED */
1151                 }
1152         }
1153 skipupdate:
1154         if (parent)
1155                 hammer2_chain_unlock(parent);
1156
1157         /*
1158          * Final cleanup after flush
1159          */
1160 done:
1161         KKASSERT(chain->refs > 0);
1162         if (hammer2_debug & 0x200) {
1163                 if (info->debug == chain)
1164                         info->debug = NULL;
1165         }
1166         return retry;
1167 }
1168
1169 /*
1170  * Flush recursion helper, called from flush_core, calls flush_core.
1171  *
1172  * Flushes the children of the caller's chain (info->parent), restricted
1173  * by sync_tid.
1174  *
1175  * This function may set info->error as a side effect.
1176  *
1177  * WARNING! If we do not call hammer2_flush_core() we must update
1178  *          bref.mirror_tid ourselves to indicate that the flush has
1179  *          processed the child.
1180  *
1181  * WARNING! parent->core spinlock is held on entry and return.
1182  */
1183 static int
1184 hammer2_flush_recurse(hammer2_chain_t *child, void *data)
1185 {
1186         hammer2_flush_info_t *info = data;
1187         hammer2_chain_t *parent = info->parent;
1188
1189 #ifdef HAMMER2_SCAN_DEBUG
1190         ++info->scan_count;
1191         if (child->flags & HAMMER2_CHAIN_MODIFIED)
1192                 ++info->scan_mod_count;
1193         if (child->flags & HAMMER2_CHAIN_UPDATE)
1194                 ++info->scan_upd_count;
1195         if (child->flags & HAMMER2_CHAIN_ONFLUSH)
1196                 ++info->scan_onf_count;
1197 #endif
1198
1199         /*
1200          * (child can never be fchain or vchain so a special check isn't
1201          *  needed).
1202          *
1203          * We must ref the child before unlocking the spinlock.
1204          *
1205          * The caller has added a ref to the parent so we can temporarily
1206          * unlock it in order to lock the child.  However, if it no longer
1207          * winds up being the child of the parent we must skip this child.
1208          *
1209          * NOTE! chain locking errors are fatal.  They are never out-of-space
1210          *       errors.
1211          */
1212         hammer2_chain_ref(child);
1213         hammer2_spin_unex(&parent->core.spin);
1214
1215         hammer2_chain_ref_hold(parent);
1216         hammer2_chain_unlock(parent);
1217         hammer2_chain_lock(child, HAMMER2_RESOLVE_MAYBE);
1218         if (child->parent != parent) {
1219                 kprintf("LOST CHILD1 %p->%p (actual parent %p)\n",
1220                         parent, child, child->parent);
1221                 goto done;
1222         }
1223         if (child->error) {
1224                 kprintf("CHILD ERROR DURING FLUSH LOCK %p->%p\n",
1225                         parent, child);
1226                 info->error |= child->error;
1227                 goto done;
1228         }
1229
1230         /*
1231          * Must propagate the DESTROY flag downwards, otherwise the
1232          * parent could end up never being removed because it will
1233          * be requeued to the flusher if it survives this run due to
1234          * the flag.
1235          */
1236         if (parent && (parent->flags & HAMMER2_CHAIN_DESTROY))
1237                 atomic_set_int(&child->flags, HAMMER2_CHAIN_DESTROY);
1238 #ifdef HAMMER2_SCAN_DEBUG
1239         if (child->flags & HAMMER2_CHAIN_DESTROY)
1240                 ++info->scan_del_count;
1241 #endif
1242         /*
1243          * Special handling of the root inode.  Because the root inode
1244          * contains an index of all the inodes in the PFS in addition to
1245          * its normal directory entries, any flush that is not part of a
1246          * filesystem sync must only flush the directory entries, and not
1247          * anything else.
1248          *
1249          * The child might be an indirect block, but H2 guarantees that
1250          * the key-range will fully partition the inode index from the
1251          * directory entries so the case just works naturally.
1252          */
1253         if ((parent->bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
1254             (child->flags & HAMMER2_CHAIN_DESTROY) == 0 &&
1255             parent->bref.type == HAMMER2_BREF_TYPE_INODE &&
1256             (info->flags & HAMMER2_FLUSH_FSSYNC) == 0) {
1257                 if ((child->bref.key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1258                         if (child->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1259                                 hammer2_chain_setflush(parent);
1260                         }
1261                         goto done;
1262                 }
1263         }
1264
1265         /*
1266          * Recurse and collect deferral data.  We're in the media flush,
1267          * this can cross PFS boundaries.
1268          */
1269         if (child->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1270 #ifdef HAMMER2_SCAN_DEBUG
1271                 if (child->bref.type < 7)
1272                         ++info->scan_btype[child->bref.type];
1273 #endif
1274                 ++info->depth;
1275                 hammer2_flush_core(info, child, info->flags);
1276                 --info->depth;
1277         } else if (hammer2_debug & 0x200) {
1278                 if (info->debug == NULL)
1279                         info->debug = child;
1280                 ++info->depth;
1281                 hammer2_flush_core(info, child, info->flags);
1282                 --info->depth;
1283                 if (info->debug == child)
1284                         info->debug = NULL;
1285         }
1286
1287 done:
1288         /*
1289          * Relock to continue the loop.
1290          */
1291         hammer2_chain_unlock(child);
1292         hammer2_chain_lock(parent, HAMMER2_RESOLVE_MAYBE);
1293         hammer2_chain_drop_unhold(parent);
1294         if (parent->error) {
1295                 kprintf("PARENT ERROR DURING FLUSH LOCK %p->%p\n",
1296                         parent, child);
1297                 info->error |= parent->error;
1298         }
1299         hammer2_chain_drop(child);
1300         KKASSERT(info->parent == parent);
1301         hammer2_spin_ex(&parent->core.spin);
1302
1303         return (0);
1304 }
1305
1306 /*
1307  * flush helper (backend threaded)
1308  *
1309  * Flushes chain topology for the specified inode.
1310  *
1311  * HAMMER2_XOP_INODE_STOP       The flush recursion stops at inode boundaries.
1312  *                              Inodes belonging to the same flush are flushed
1313  *                              separately.
1314  *
1315  * chain->parent can be NULL, usually due to destroy races or detached inodes.
1316  *
1317  * Primarily called from vfs_sync().
1318  */
1319 void
1320 hammer2_xop_inode_flush(hammer2_xop_t *arg, void *scratch __unused, int clindex)
1321 {
1322         hammer2_xop_flush_t *xop = &arg->xop_flush;
1323         hammer2_chain_t *chain;
1324         hammer2_inode_t *ip;
1325         hammer2_dev_t *hmp;
1326         hammer2_pfs_t *pmp;
1327         hammer2_devvp_t *e;
1328         struct vnode *devvp;
1329         int flush_error = 0;
1330         int fsync_error = 0;
1331         int total_error = 0;
1332         int j;
1333         int xflags;
1334         int ispfsroot = 0;
1335
1336         xflags = HAMMER2_FLUSH_TOP;
1337         if (xop->head.flags & HAMMER2_XOP_INODE_STOP)
1338                 xflags |= HAMMER2_FLUSH_INODE_STOP;
1339         if (xop->head.flags & HAMMER2_XOP_FSSYNC)
1340                 xflags |= HAMMER2_FLUSH_FSSYNC;
1341
1342         /*
1343          * Flush core chains
1344          */
1345         ip = xop->head.ip1;
1346         pmp = ip->pmp;
1347         chain = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS);
1348         if (chain) {
1349                 hmp = chain->hmp;
1350                 if (chain->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1351                         /*
1352                          * Due to flush partitioning the chain topology
1353                          * above the inode's chain may no longer be flagged.
1354                          * When asked to flush an inode, remark the topology
1355                          * leading to that inode.
1356                          */
1357                         if (chain->parent)
1358                                 hammer2_chain_setflush(chain->parent);
1359                         hammer2_flush(chain, xflags);
1360
1361                         /* XXX cluster */
1362                         if (ip == pmp->iroot && pmp != hmp->spmp) {
1363                                 hammer2_spin_ex(&pmp->inum_spin);
1364                                 pmp->pfs_iroot_blocksets[clindex] =
1365                                         chain->data->ipdata.u.blockset;
1366                                 hammer2_spin_unex(&pmp->inum_spin);
1367                         }
1368
1369 #if 0
1370                         /*
1371                          * Propogate upwards but only cross an inode boundary
1372                          * for inodes associated with the current filesystem
1373                          * sync.
1374                          */
1375                         if ((xop->head.flags & HAMMER2_XOP_PARENTONFLUSH) ||
1376                             chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
1377                                 parent = chain->parent;
1378                                 if (parent)
1379                                         hammer2_chain_setflush(parent);
1380                         }
1381 #endif
1382                 }
1383                 if (chain->flags & HAMMER2_CHAIN_PFSBOUNDARY)
1384                         ispfsroot = 1;
1385                 hammer2_chain_unlock(chain);
1386                 hammer2_chain_drop(chain);
1387                 chain = NULL;
1388         } else {
1389                 hmp = NULL;
1390         }
1391
1392         /*
1393          * Only flush the volume header if asked to, plus the inode must also
1394          * be the PFS root.
1395          */
1396         if ((xop->head.flags & HAMMER2_XOP_VOLHDR) == 0)
1397                 goto skip;
1398         if (ispfsroot == 0)
1399                 goto skip;
1400
1401         /*
1402          * Flush volume roots.  Avoid replication, we only want to
1403          * flush each hammer2_dev (hmp) once.
1404          */
1405         for (j = clindex - 1; j >= 0; --j) {
1406                 if ((chain = ip->cluster.array[j].chain) != NULL) {
1407                         if (chain->hmp == hmp) {
1408                                 chain = NULL;   /* safety */
1409                                 goto skip;
1410                         }
1411                 }
1412         }
1413         chain = NULL;   /* safety */
1414
1415         /*
1416          * spmp transaction.  The super-root is never directly mounted so
1417          * there shouldn't be any vnodes, let alone any dirty vnodes
1418          * associated with it, so we shouldn't have to mess around with any
1419          * vnode flushes here.
1420          */
1421         hammer2_trans_init(hmp->spmp, HAMMER2_TRANS_ISFLUSH);
1422
1423         /*
1424          * We must flush the superroot down to the PFS iroot.  Remember
1425          * that hammer2_chain_setflush() stops at inode boundaries, so
1426          * the pmp->iroot has been flushed and flagged down to the superroot,
1427          * but the volume root (vchain) probably has not yet been flagged.
1428          */
1429         if (hmp->spmp->iroot) {
1430                 chain = hmp->spmp->iroot->cluster.array[0].chain;
1431                 if (chain) {
1432                         hammer2_chain_ref(chain);
1433                         hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
1434                         flush_error |=
1435                                 hammer2_flush(chain,
1436                                               HAMMER2_FLUSH_TOP |
1437                                               HAMMER2_FLUSH_INODE_STOP |
1438                                               HAMMER2_FLUSH_FSSYNC);
1439                         hammer2_chain_unlock(chain);
1440                         hammer2_chain_drop(chain);
1441                 }
1442         }
1443
1444         /*
1445          * Media mounts have two 'roots', vchain for the topology
1446          * and fchain for the free block table.  Flush both.
1447          *
1448          * Note that the topology and free block table are handled
1449          * independently, so the free block table can wind up being
1450          * ahead of the topology.  We depend on the bulk free scan
1451          * code to deal with any loose ends.
1452          *
1453          * vchain and fchain do not error on-lock since their data does
1454          * not have to be re-read from media.
1455          */
1456         hammer2_chain_ref(&hmp->vchain);
1457         hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1458         hammer2_chain_ref(&hmp->fchain);
1459         hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1460         if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1461                 /*
1462                  * This will also modify vchain as a side effect,
1463                  * mark vchain as modified now.
1464                  */
1465                 hammer2_voldata_modify(hmp);
1466                 chain = &hmp->fchain;
1467                 flush_error |= hammer2_flush(chain, HAMMER2_FLUSH_TOP);
1468                 KKASSERT(chain == &hmp->fchain);
1469         }
1470         hammer2_chain_unlock(&hmp->fchain);
1471         hammer2_chain_unlock(&hmp->vchain);
1472         hammer2_chain_drop(&hmp->fchain);
1473         /* vchain dropped down below */
1474
1475         hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1476         if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1477                 chain = &hmp->vchain;
1478                 flush_error |= hammer2_flush(chain, HAMMER2_FLUSH_TOP);
1479                 KKASSERT(chain == &hmp->vchain);
1480         }
1481         hammer2_chain_unlock(&hmp->vchain);
1482         hammer2_chain_drop(&hmp->vchain);
1483
1484         /*
1485          * We can't safely flush the volume header until we have
1486          * flushed any device buffers which have built up.
1487          *
1488          * XXX this isn't being incremental
1489          */
1490         TAILQ_FOREACH(e, &hmp->devvpl, entry) {
1491                 devvp = e->devvp;
1492                 KKASSERT(devvp);
1493                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1494                 fsync_error = VOP_FSYNC(devvp, MNT_WAIT, 0);
1495                 vn_unlock(devvp);
1496                 if (fsync_error || flush_error) {
1497                         kprintf("hammer2: sync error fsync=%d h2flush=0x%04x dev=%s\n",
1498                                 fsync_error, flush_error, e->path);
1499                 }
1500         }
1501
1502         /*
1503          * The flush code sets CHAIN_VOLUMESYNC to indicate that the
1504          * volume header needs synchronization via hmp->volsync.
1505          *
1506          * XXX synchronize the flag & data with only this flush XXX
1507          */
1508         if (fsync_error == 0 && flush_error == 0 &&
1509             (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
1510                 struct buf *bp;
1511                 int vol_error = 0;
1512
1513                 /*
1514                  * Synchronize the disk before flushing the volume
1515                  * header.
1516                  */
1517                 bp = getpbuf(NULL);
1518                 bp->b_bio1.bio_offset = 0;
1519                 bp->b_bufsize = 0;
1520                 bp->b_bcount = 0;
1521                 bp->b_cmd = BUF_CMD_FLUSH;
1522                 bp->b_bio1.bio_done = biodone_sync;
1523                 bp->b_bio1.bio_flags |= BIO_SYNC;
1524                 vn_strategy(hmp->devvp, &bp->b_bio1);
1525                 fsync_error = biowait(&bp->b_bio1, "h2vol");
1526                 relpbuf(bp, NULL);
1527
1528                 /*
1529                  * Then we can safely flush the version of the
1530                  * volume header synchronized by the flush code.
1531                  */
1532                 j = hmp->volhdrno + 1;
1533                 if (j < 0)
1534                         j = 0;
1535                 if (j >= HAMMER2_NUM_VOLHDRS)
1536                         j = 0;
1537                 if (j * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
1538                     hmp->volsync.volu_size) {
1539                         j = 0;
1540                 }
1541                 if (hammer2_debug & 0x8000) {
1542                         /* debug only, avoid syslogd loop */
1543                         kprintf("sync volhdr %d %jd\n",
1544                                 j, (intmax_t)hmp->volsync.volu_size);
1545                 }
1546                 bp = getblk(hmp->devvp, j * HAMMER2_ZONE_BYTES64,
1547                             HAMMER2_PBUFSIZE, GETBLK_KVABIO, 0);
1548                 atomic_clear_int(&hmp->vchain.flags,
1549                                  HAMMER2_CHAIN_VOLUMESYNC);
1550                 bkvasync(bp);
1551                 bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
1552                 vol_error = bwrite(bp);
1553                 hmp->volhdrno = j;
1554                 if (vol_error)
1555                         fsync_error = vol_error;
1556         }
1557         if (flush_error)
1558                 total_error = flush_error;
1559         if (fsync_error)
1560                 total_error = hammer2_errno_to_error(fsync_error);
1561
1562         /* spmp trans */
1563         hammer2_trans_done(hmp->spmp, HAMMER2_TRANS_ISFLUSH);
1564 skip:
1565         hammer2_xop_feed(&xop->head, NULL, clindex, total_error);
1566 }