hammer2 - Refactor flush mechanics

[dragonfly.git] / sys / vfs / hammer2 / hammer2_chain.c

/*
 * Copyright (c) 2011-2014 The DragonFly Project.  All rights reserved.
 *
 * This code is derived from software contributed to The DragonFly Project
 * by Matthew Dillon <dillon@dragonflybsd.org>
 * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 *
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in
 *    the documentation and/or other materials provided with the
 *    distribution.
 * 3. Neither the name of The DragonFly Project nor the names of its
 *    contributors may be used to endorse or promote products derived
 *    from this software without specific, prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 */
/*
 * This subsystem implements most of the core support functions for
 * the hammer2_chain structure.
 *
 * Chains are the in-memory version on media objects (volume header, inodes,
 * indirect blocks, data blocks, etc).  Chains represent a portion of the
 * HAMMER2 topology.
 *
 * A chain is topologically stable once it has been inserted into the
 * in-memory topology.  Modifications which copy, move, or resize the chain
 * are handled via the DELETE-DUPLICATE mechanic where the original chain
 * stays intact but is marked deleted and a new chain is allocated which
 * shares the old chain's children.
 *
 * This sharing is handled via the hammer2_chain_core structure.
 *
 * The DELETE-DUPLICATE mechanism allows the same topological level to contain
 * many overloadings.  However, our RBTREE mechanics require that there be
 * no overlaps so we accomplish the overloading by moving conflicting chains
 * with smaller or equal radii into a sub-RBTREE under the chain being
 * overloaded.
 *
 * DELETE-DUPLICATE is also used when a modification to a chain crosses a
 * flush synchronization boundary, allowing the flush code to continue flushing
 * the older version of the topology and not be disrupted by new frontend
 * operations.
 *
 *                              LIVE VS FLUSH VIEW
 *
 * All lookup and iterate operations and most modifications are done on the
 * live view.  During flushes lookups are not normally done and modifications
 * may be run on the flush view.  However, flushes often needs to allocate
 * blocks and the freemap_alloc/free code issues lookups.  This code is
 * special cased to use the live view when called from a flush.
 *
 * General chain lookup/iteration functions are NOT aware of the flush view,
 * they only know about live views.
 */
#include <sys/cdefs.h>
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/types.h>
#include <sys/lock.h>
#include <sys/kern_syscall.h>
#include <sys/uuid.h>

#include "hammer2.h"

static int hammer2_indirect_optimize;   /* XXX SYSCTL */

static hammer2_chain_t *hammer2_chain_create_indirect(
                hammer2_trans_t *trans, hammer2_chain_t *parent,
                hammer2_key_t key, int keybits, int for_type, int *errorp);
static void hammer2_chain_drop_data(hammer2_chain_t *chain, int lastdrop);
static void adjreadcounter(hammer2_blockref_t *bref, size_t bytes);
static hammer2_chain_t *hammer2_combined_find(
                hammer2_chain_t *parent,
                hammer2_blockref_t *base, int count,
                int *cache_indexp, hammer2_key_t *key_nextp,
                hammer2_key_t key_beg, hammer2_key_t key_end,
                hammer2_blockref_t **bresp);

/*
 * Basic RBTree for chains (core->rbtree and core->dbtree).  Chains cannot
 * overlap in the RB trees.  Deleted chains are moved from rbtree to either
 * dbtree or to dbq.
 *
 * Chains in delete-duplicate sequences can always iterate through core_entry
 * to locate the live version of the chain.
 */
RB_GENERATE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);

int
hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2)
{
        hammer2_key_t c1_beg;
        hammer2_key_t c1_end;
        hammer2_key_t c2_beg;
        hammer2_key_t c2_end;

        /*
         * Compare chains.  Overlaps are not supposed to happen and catch
         * any software issues early we count overlaps as a match.
         */
        c1_beg = chain1->bref.key;
        c1_end = c1_beg + ((hammer2_key_t)1 << chain1->bref.keybits) - 1;
        c2_beg = chain2->bref.key;
        c2_end = c2_beg + ((hammer2_key_t)1 << chain2->bref.keybits) - 1;

        if (c1_end < c2_beg)    /* fully to the left */
                return(-1);
        if (c1_beg > c2_end)    /* fully to the right */
                return(1);
        return(0);              /* overlap (must not cross edge boundary) */
}

static __inline
int
hammer2_isclusterable(hammer2_chain_t *chain)
{
        if (hammer2_cluster_enable) {
                if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
                    chain->bref.type == HAMMER2_BREF_TYPE_INODE ||
                    chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
                        return(1);
                }
        }
        return(0);
}

/*
 * Recursively set update_hi starting at chain up through to the root.
 *
 * This controls top-down visibility for flushes.  The child has just one
 * 'above' core, but the core itself can be multi-homed with parents iterated
 * via core->ownerq.  The last parent is the 'live' parent (all others had to
 * have been delete-duplicated).  We always propagate upward through the live
 * parent.
 *
 * This function is not used during a flush (except when the flush is
 * allocating which requires the live tree).  The flush keeps track of its
 * recursion itself.
 *
 * XXX SMP races
 */
void
hammer2_chain_setsubmod(hammer2_trans_t *trans, hammer2_chain_t *chain)
{
        hammer2_chain_core_t *above;

        if (chain->update_hi < trans->sync_tid)
                chain->update_hi = trans->sync_tid;

        while ((above = chain->above) != NULL) {
                spin_lock(&above->cst.spin);
                chain = TAILQ_LAST(&above->ownerq, h2_core_list);
                if (chain->update_hi < trans->sync_tid)
                        chain->update_hi = trans->sync_tid;
                spin_unlock(&above->cst.spin);
        }
}

/*
 * Allocate a new disconnected chain element representing the specified
 * bref.  chain->refs is set to 1 and the passed bref is copied to
 * chain->bref.  chain->bytes is derived from the bref.
 *
 * chain->core is NOT allocated and the media data and bp pointers are left
 * NULL.  The caller must call chain_core_alloc() to allocate or associate
 * a core with the chain.
 *
 * NOTE: Returns a referenced but unlocked (because there is no core) chain.
 */
hammer2_chain_t *
hammer2_chain_alloc(hammer2_mount_t *hmp, hammer2_pfsmount_t *pmp,
                    hammer2_trans_t *trans, hammer2_blockref_t *bref)
{
        hammer2_chain_t *chain;
        u_int bytes = 1U << (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX);

        /*
         * Construct the appropriate system structure.
         */
        switch(bref->type) {
        case HAMMER2_BREF_TYPE_INODE:
        case HAMMER2_BREF_TYPE_INDIRECT:
        case HAMMER2_BREF_TYPE_FREEMAP_NODE:
        case HAMMER2_BREF_TYPE_DATA:
        case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
                /*
                 * Chain's are really only associated with the hmp but we
                 * maintain a pmp association for per-mount memory tracking
                 * purposes.  The pmp can be NULL.
                 */
                chain = kmalloc(sizeof(*chain), hmp->mchain, M_WAITOK | M_ZERO);
                if (pmp)
                        chain->pmp = pmp;
                break;
        case HAMMER2_BREF_TYPE_VOLUME:
        case HAMMER2_BREF_TYPE_FREEMAP:
                chain = NULL;
                panic("hammer2_chain_alloc volume type illegal for op");
        default:
                chain = NULL;
                panic("hammer2_chain_alloc: unrecognized blockref type: %d",
                      bref->type);
        }

        chain->hmp = hmp;
        chain->bref = *bref;
        chain->bytes = bytes;
        chain->refs = 1;
        chain->flags = HAMMER2_CHAIN_ALLOCATED;
        chain->delete_tid = HAMMER2_MAX_TID;

        /*
         * Set modify_tid if a transaction is creating the inode.
         * Enforce update_lo = 0 so nearby transactions do not think
         * it has been flushed when it hasn't.
         *
         * NOTE: When loading a chain from backing store or creating a
         *       snapshot, trans will be NULL and the caller is responsible
         *       for setting these fields.
         */
        if (trans) {
                chain->modify_tid = trans->sync_tid;
                chain->update_lo = 0;
        }

        return (chain);
}

/*
 * Associate an existing core with the chain or allocate a new core.
 *
 * The core is not locked.  No additional refs on the chain are made.
 * (trans) must not be NULL if (core) is not NULL.
 *
 * When chains are delete-duplicated during flushes we insert nchain on
 * the ownerq after ochain instead of at the end in order to give the
 * drop code visibility in the correct order, otherwise drops can be missed.
 */
void
hammer2_chain_core_alloc(hammer2_trans_t *trans,
                         hammer2_chain_t *nchain, hammer2_chain_t *ochain)
{
        hammer2_chain_core_t *core;

        KKASSERT(nchain->core == NULL);

        if (ochain == NULL) {
                /*
                 * Fresh core under nchain (no multi-homing of ochain's
                 * sub-tree).
                 */
                core = kmalloc(sizeof(*core), nchain->hmp->mchain,
                               M_WAITOK | M_ZERO);
                TAILQ_INIT(&core->ownerq);
                TAILQ_INIT(&core->dbq);
                RB_INIT(&core->rbtree); /* live chains */
                RB_INIT(&core->dbtree); /* deleted original (bmapped) chains */
                core->sharecnt = 1;
                core->good = 0x1234;
                nchain->core = core;
                ccms_cst_init(&core->cst, nchain);
                TAILQ_INSERT_TAIL(&core->ownerq, nchain, core_entry);
        } else {
                /*
                 * Propagate the PFSROOT flag which we set on all subdirs
                 * under the super-root.
                 */
                atomic_set_int(&nchain->flags,
                               ochain->flags & HAMMER2_CHAIN_PFSROOT);

                /*
                 * Duplicating ochain -> nchain.  Set the DUPLICATED flag on
                 * ochain if nchain is not a snapshot.
                 *
                 * It is possible for the DUPLICATED flag to already be
                 * set when called via a flush operation because flush
                 * operations may have to work on elements with delete_tid's
                 * beyond the flush sync_tid.  In this situation we must
                 * ensure that nchain is placed just after ochain in the
                 * ownerq and that the DUPLICATED flag is set on nchain so
                 * 'live' operations skip past it to the correct chain.
                 *
                 * The flusher understands the blockref synchronization state
                 * for any stale chains by observing bref.mirror_tid, which
                 * delete-duplicate replicates.
                 *
                 * WARNING! However, the case is disallowed when the flusher
                 *          is allocating freemap space because this entails
                 *          more than just adjusting a block table.
                 */
                if (ochain->flags & HAMMER2_CHAIN_DUPLICATED) {
                        KKASSERT((trans->flags &
                                  (HAMMER2_TRANS_ISFLUSH |
                                   HAMMER2_TRANS_ISALLOCATING)) ==
                                 HAMMER2_TRANS_ISFLUSH);
                        atomic_set_int(&nchain->flags,
                                       HAMMER2_CHAIN_DUPLICATED);
                }
                if ((nchain->flags & HAMMER2_CHAIN_SNAPSHOT) == 0) {
                        atomic_set_int(&ochain->flags,
                                       HAMMER2_CHAIN_DUPLICATED);
                }
                core = ochain->core;
                atomic_add_int(&core->sharecnt, 1);

                spin_lock(&core->cst.spin);
                nchain->core = core;

                /*
                 * Maintain ordering for refactor test so we don't skip over
                 * a snapshot.  Also, during flushes, delete-duplications
                 * for block-table updates can occur on ochains already
                 * deleted (delete-duplicated by a later transaction), or
                 * on forward-indexed ochains.  We must properly insert
                 * nchain relative to ochain.
                 */
                if (trans && trans->sync_tid < ochain->modify_tid) {
                        TAILQ_INSERT_BEFORE(ochain, nchain, core_entry);
                } else {
                        TAILQ_INSERT_AFTER(&core->ownerq, ochain,
                                           nchain, core_entry);
                }
                spin_unlock(&core->cst.spin);
        }
}

/*
 * Add a reference to a chain element, preventing its destruction.
 */
void
hammer2_chain_ref(hammer2_chain_t *chain)
{
        atomic_add_int(&chain->refs, 1);
}

/*
 * Insert the chain in the core rbtree.
 *
 * Normal insertions are placed in the live rbtree.  Insertion of a deleted
 * chain is a special case used by the flush code that is placed on the
 * unstaged deleted list to avoid confusing the live view.
 */
#define HAMMER2_CHAIN_INSERT_SPIN       0x0001
#define HAMMER2_CHAIN_INSERT_LIVE       0x0002
#define HAMMER2_CHAIN_INSERT_RACE       0x0004

static
int
hammer2_chain_insert(hammer2_chain_core_t *above,
                     hammer2_chain_t *ochain, hammer2_chain_t *nchain,
                     int flags, int generation)
{
        hammer2_chain_t *xchain;
        int error = 0;

        if (flags & HAMMER2_CHAIN_INSERT_SPIN)
                spin_lock(&above->cst.spin);

        /*
         * Interlocked by spinlock, check for race
         */
        if ((flags & HAMMER2_CHAIN_INSERT_RACE) &&
            above->generation != generation) {
                error = EAGAIN;
                goto failed;
        }

        /*
         * Insert nchain
         */
        if (nchain->flags & HAMMER2_CHAIN_DELETED) {
                if (ochain && (ochain->flags & HAMMER2_CHAIN_BMAPPED)) {
                        if (ochain->flags & HAMMER2_CHAIN_ONDBTREE) {
                                RB_REMOVE(hammer2_chain_tree,
                                          &above->dbtree, ochain);
                                atomic_clear_int(&ochain->flags,
                                                 HAMMER2_CHAIN_ONDBTREE);
                                TAILQ_INSERT_TAIL(&above->dbq,
                                                  ochain, db_entry);
                                atomic_set_int(&ochain->flags,
                                                HAMMER2_CHAIN_ONDBQ);
                        }
                        /* clear BMAPPED (DBTREE, sometimes RBTREE) */
                        atomic_clear_int(&ochain->flags, HAMMER2_CHAIN_BMAPPED);

                        xchain = RB_INSERT(hammer2_chain_tree,
                                           &above->dbtree, nchain);
                        KKASSERT(xchain == NULL);
                        atomic_set_int(&nchain->flags,
                                       HAMMER2_CHAIN_ONDBTREE |
                                       HAMMER2_CHAIN_BMAPPED);
                } else {
                        TAILQ_INSERT_TAIL(&above->dbq, nchain, db_entry);
                        atomic_set_int(&nchain->flags, HAMMER2_CHAIN_ONDBQ);
                }
        } else {
                xchain = RB_INSERT(hammer2_chain_tree, &above->rbtree, nchain);
                KASSERT(xchain == NULL,
                        ("hammer2_chain_insert: collision %p", nchain));
                atomic_set_int(&nchain->flags, HAMMER2_CHAIN_ONRBTREE);
        }

        nchain->above = above;
        ++above->chain_count;
        ++above->generation;

        /*
         * We have to keep track of the effective live-view blockref count
         * so the create code knows when to push an indirect block.
         */
        if (flags & HAMMER2_CHAIN_INSERT_LIVE)
                atomic_add_int(&above->live_count, 1);
failed:
        if (flags & HAMMER2_CHAIN_INSERT_SPIN)
                spin_unlock(&above->cst.spin);
        return error;
}

/*
 * Drop the caller's reference to the chain.  When the ref count drops to
 * zero this function will try to disassociate the chain from its parent and
 * deallocate it, then recursely drop the parent using the implied ref
 * from the chain's chain->parent.
 */
static hammer2_chain_t *hammer2_chain_lastdrop(hammer2_chain_t *chain,
                                               struct h2_core_list *delayq);

void
hammer2_chain_drop(hammer2_chain_t *chain)
{
        struct h2_core_list delayq;
        hammer2_chain_t *scan;
        u_int refs;
        u_int need = 0;

        if (hammer2_debug & 0x200000)
                Debugger("drop");

        if (chain->flags & HAMMER2_CHAIN_FLUSH_CREATE)
                ++need;
        if (chain->flags & HAMMER2_CHAIN_FLUSH_DELETE)
                ++need;
        if (chain->flags & HAMMER2_CHAIN_MODIFIED)
                ++need;
        KKASSERT(chain->refs > need);

        TAILQ_INIT(&delayq);

        while (chain) {
                refs = chain->refs;
                cpu_ccfence();
                KKASSERT(refs > 0);

                if (refs == 1) {
                        chain = hammer2_chain_lastdrop(chain, &delayq);
                } else {
                        if (atomic_cmpset_int(&chain->refs, refs, refs - 1))
                                break;
                        /* retry the same chain */
                }

                /*
                 * When we've exhausted lastdrop chaining pull off of delayq.
                 * chains on delayq are dead but are used to placehold other
                 * chains which we added a ref to for the purpose of dropping.
                 */
                if (chain == NULL) {
                        hammer2_mount_t *hmp;

                        if ((scan = TAILQ_FIRST(&delayq)) != NULL) {
                                chain = (void *)scan->data;
                                TAILQ_REMOVE(&delayq, scan, core_entry);
                                scan->flags &= ~HAMMER2_CHAIN_ALLOCATED;
                                hmp = scan->hmp;
                                scan->hmp = NULL;
                                kfree(scan, hmp->mchain);
                        }
                }
        }
}

/*
 * Safe handling of the 1->0 transition on chain.  Returns a chain for
 * recursive drop or NULL, possibly returning the same chain if the atomic
 * op fails.
 *
 * Whem two chains need to be recursively dropped we use the chain
 * we would otherwise free to placehold the additional chain.  It's a bit
 * convoluted but we can't just recurse without potentially blowing out
 * the kernel stack.
 *
 * The chain cannot be freed if it has a non-empty core (children) or
 * it is not at the head of ownerq.
 *
 * The cst spinlock is allowed nest child-to-parent (not parent-to-child).
 */
static
hammer2_chain_t *
hammer2_chain_lastdrop(hammer2_chain_t *chain, struct h2_core_list *delayq)
{
        hammer2_pfsmount_t *pmp;
        hammer2_mount_t *hmp;
        hammer2_chain_core_t *above;
        hammer2_chain_core_t *core;
        hammer2_chain_t *rdrop1;
        hammer2_chain_t *rdrop2;

        /*
         * Spinlock the core and check to see if it is empty.  If it is
         * not empty we leave chain intact with refs == 0.  The elements
         * in core->rbtree are associated with other chains contemporary
         * with ours but not with our chain directly.
         */
        if ((core = chain->core) != NULL) {
                spin_lock(&core->cst.spin);

                /*
                 * We can't free non-stale chains with children until we are
                 * able to free the children because there might be a flush
                 * dependency.  Flushes of stale children (which should also
                 * have their deleted flag set) short-cut recursive flush
                 * dependencies and can be freed here.  Any flushes which run
                 * through stale children due to the flush synchronization
                 * point should have a FLUSH_* bit set in the chain and not
                 * reach lastdrop at this time.
                 *
                 * NOTE: We return (chain) on failure to retry.
                 */
                if (core->chain_count &&
                    (chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0) {
                        if (atomic_cmpset_int(&chain->refs, 1, 0))
                                chain = NULL;   /* success */
                        spin_unlock(&core->cst.spin);
                        return(chain);
                }
                /* no chains left under us */

                /*
                 * Various parts of the code might be holding a ref on a
                 * stale chain as a placemarker which must be iterated to
                 * locate a later non-stale (live) chain.  We must be sure
                 * NOT to free the later non-stale chain (which might have
                 * no refs).  Otherwise mass confusion may result.
                 *
                 * The DUPLICATED flag tells us whether the chain is stale
                 * or not, so the rule is that any chain whos DUPLICATED flag
                 * is NOT set must also be at the head of the ownerq.
                 *
                 * Note that the DELETED flag is not involved.  That is, a
                 * live chain can represent a deletion that has not yet been
                 * flushed (or still has refs).
                 */
#if 0
                if (TAILQ_NEXT(chain, core_entry) == NULL &&
                    TAILQ_FIRST(&core->ownerq) != chain) {
#endif
                if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0 &&
                    TAILQ_FIRST(&core->ownerq) != chain) {
                        if (atomic_cmpset_int(&chain->refs, 1, 0))
                                chain = NULL;   /* success */
                        spin_unlock(&core->cst.spin);
                        return(chain);
                }
        }

        /*
         * chain->core has no children left so no accessors can get to our
         * chain from there.  Now we have to lock the above core to interlock
         * remaining possible accessors that might bump chain's refs before
         * we can safely drop chain's refs with intent to free the chain.
         */
        hmp = chain->hmp;
        pmp = chain->pmp;       /* can be NULL */
        rdrop1 = NULL;
        rdrop2 = NULL;

        /*
         * Spinlock the parent and try to drop the last ref on chain.
         * On success remove chain from its parent, otherwise return NULL.
         *
         * (normal core locks are top-down recursive but we define core
         *  spinlocks as bottom-up recursive, so this is safe).
         */
        if ((above = chain->above) != NULL) {
                spin_lock(&above->cst.spin);
                if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) {
                        /* 1->0 transition failed */
                        spin_unlock(&above->cst.spin);
                        if (core)
                                spin_unlock(&core->cst.spin);
                        return(chain);  /* retry */
                }

                /*
                 * 1->0 transition successful, remove chain from its
                 * above core.
                 */
                switch (chain->flags & (HAMMER2_CHAIN_ONRBTREE |
                                        HAMMER2_CHAIN_ONDBTREE |
                                        HAMMER2_CHAIN_ONDBQ)) {
                case HAMMER2_CHAIN_ONRBTREE:
                        RB_REMOVE(hammer2_chain_tree, &above->rbtree, chain);
                        atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);
                        break;
                case HAMMER2_CHAIN_ONDBTREE:
                        RB_REMOVE(hammer2_chain_tree, &above->dbtree, chain);
                        atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONDBTREE);
                        break;
                case HAMMER2_CHAIN_ONDBQ:
                        TAILQ_REMOVE(&above->dbq, chain, db_entry);
                        atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONDBQ);
                        break;
                default:
                        panic("hammer2_chain_lastdrop: chain %p badflags %08x",
                              chain, chain->flags);
                        break;
                }

                --above->chain_count;
                chain->above = NULL;

                /*
                 * If our chain was the last chain in the parent's core the
                 * core is now empty and its parents might now be droppable.
                 * Try to drop the first multi-homed parent by gaining a
                 * ref on it here and then dropping it below.
                 */
                if (above->chain_count == 0) {
                        rdrop1 = TAILQ_FIRST(&above->ownerq);
                        if (rdrop1 &&
                            atomic_cmpset_int(&rdrop1->refs, 0, 1) == 0) {
                                rdrop1 = NULL;
                        }
                }
                spin_unlock(&above->cst.spin);
                above = NULL;   /* safety */
        }

        /*
         * Successful 1->0 transition and the chain can be destroyed now.
         *
         * We still have the core spinlock (if core is non-NULL), and core's
         * chain_count is 0.  The above spinlock is gone.
         *
         * Remove chain from ownerq.  Once core has no more owners (and no
         * children which is already the case) we can destroy core.
         *
         * If core has more owners we may be able to continue a bottom-up
         * drop with our next sibling.
         */
        if (core) {
                chain->core = NULL;

                TAILQ_REMOVE(&core->ownerq, chain, core_entry);
                rdrop2 = TAILQ_FIRST(&core->ownerq);
                if (rdrop2 && atomic_cmpset_int(&rdrop2->refs, 0, 1) == 0)
                        rdrop2 = NULL;
                spin_unlock(&core->cst.spin);

                /*
                 * We can do the final 1->0 transition with an atomic op
                 * after releasing core's spinlock.
                 */
                if (atomic_fetchadd_int(&core->sharecnt, -1) == 1) {
                        /*
                         * On the 1->0 transition of core we can destroy
                         * it.
                         */
                        KKASSERT(TAILQ_EMPTY(&core->ownerq));
                        KKASSERT(RB_EMPTY(&core->rbtree) &&
                                 RB_EMPTY(&core->dbtree) &&
                                 TAILQ_EMPTY(&core->dbq) &&
                                 core->chain_count == 0);
                        KKASSERT(core->cst.count == 0);
                        KKASSERT(core->cst.upgrade == 0);
                        core->good = 0x5678;
                        kfree(core, hmp->mchain);
                }
                core = NULL;    /* safety */
        }

        /*
         * All spin locks are gone, finish freeing stuff.
         */
        KKASSERT((chain->flags & (HAMMER2_CHAIN_FLUSH_CREATE |
                                  HAMMER2_CHAIN_FLUSH_DELETE |
                                  HAMMER2_CHAIN_MODIFIED)) == 0);
        hammer2_chain_drop_data(chain, 1);

        KKASSERT(chain->dio == NULL);

        /*
         * Once chain resources are gone we can use the now dead chain
         * structure to placehold what might otherwise require a recursive
         * drop, because we have potentially two things to drop and can only
         * return one directly.
         */
        if (rdrop1 && rdrop2) {
                KKASSERT(chain->flags & HAMMER2_CHAIN_ALLOCATED);
                chain->data = (void *)rdrop1;
                TAILQ_INSERT_TAIL(delayq, chain, core_entry);
                rdrop1 = NULL;
        } else if (chain->flags & HAMMER2_CHAIN_ALLOCATED) {
                chain->flags &= ~HAMMER2_CHAIN_ALLOCATED;
                chain->hmp = NULL;
                kfree(chain, hmp->mchain);
        }

        /*
         * Either or both can be NULL.  We already handled the case where
         * both might not have been NULL.
         */
        if (rdrop1)
                return(rdrop1);
        else
                return(rdrop2);
}

/*
 * On either last lock release or last drop
 */
static void
hammer2_chain_drop_data(hammer2_chain_t *chain, int lastdrop)
{
        /*hammer2_mount_t *hmp = chain->hmp;*/

        switch(chain->bref.type) {
        case HAMMER2_BREF_TYPE_VOLUME:
        case HAMMER2_BREF_TYPE_FREEMAP:
                if (lastdrop)
                        chain->data = NULL;
                break;
        default:
                KKASSERT(chain->data == NULL);
                break;
        }
}

/*
 * Ref and lock a chain element, acquiring its data with I/O if necessary,
 * and specify how you would like the data to be resolved.
 *
 * Returns 0 on success or an error code if the data could not be acquired.
 * The chain element is locked on return regardless of whether an error
 * occurred or not.
 *
 * The lock is allowed to recurse, multiple locking ops will aggregate
 * the requested resolve types.  Once data is assigned it will not be
 * removed until the last unlock.
 *
 * HAMMER2_RESOLVE_NEVER - Do not resolve the data element.
 *                         (typically used to avoid device/logical buffer
 *                          aliasing for data)
 *
 * HAMMER2_RESOLVE_MAYBE - Do not resolve data elements for chains in
 *                         the INITIAL-create state (indirect blocks only).
 *
 *                         Do not resolve data elements for DATA chains.
 *                         (typically used to avoid device/logical buffer
 *                          aliasing for data)
 *
 * HAMMER2_RESOLVE_ALWAYS- Always resolve the data element.
 *
 * HAMMER2_RESOLVE_SHARED- (flag) The chain is locked shared, otherwise
 *                         it will be locked exclusive.
 *
 * NOTE: Embedded elements (volume header, inodes) are always resolved
 *       regardless.
 *
 * NOTE: Specifying HAMMER2_RESOLVE_ALWAYS on a newly-created non-embedded
 *       element will instantiate and zero its buffer, and flush it on
 *       release.
 *
 * NOTE: (data) elements are normally locked RESOLVE_NEVER or RESOLVE_MAYBE
 *       so as not to instantiate a device buffer, which could alias against
 *       a logical file buffer.  However, if ALWAYS is specified the
 *       device buffer will be instantiated anyway.
 *
 * WARNING! If data must be fetched a shared lock will temporarily be
 *          upgraded to exclusive.  However, a deadlock can occur if
 *          the caller owns more than one shared lock.
 */
int
hammer2_chain_lock(hammer2_chain_t *chain, int how)
{
        hammer2_mount_t *hmp;
        hammer2_chain_core_t *core;
        hammer2_blockref_t *bref;
        ccms_state_t ostate;
        char *bdata;
        int error;

        /*
         * Ref and lock the element.  Recursive locks are allowed.
         */
        if ((how & HAMMER2_RESOLVE_NOREF) == 0)
                hammer2_chain_ref(chain);
        atomic_add_int(&chain->lockcnt, 1);

        hmp = chain->hmp;
        KKASSERT(hmp != NULL);

        /*
         * Get the appropriate lock.
         */
        core = chain->core;
        if (how & HAMMER2_RESOLVE_SHARED)
                ccms_thread_lock(&core->cst, CCMS_STATE_SHARED);
        else
                ccms_thread_lock(&core->cst, CCMS_STATE_EXCLUSIVE);

        /*
         * If we already have a valid data pointer no further action is
         * necessary.
         */
        if (chain->data)
                return (0);

        /*
         * Do we have to resolve the data?
         */
        switch(how & HAMMER2_RESOLVE_MASK) {
        case HAMMER2_RESOLVE_NEVER:
                return(0);
        case HAMMER2_RESOLVE_MAYBE:
                if (chain->flags & HAMMER2_CHAIN_INITIAL)
                        return(0);
                if (chain->bref.type == HAMMER2_BREF_TYPE_DATA)
                        return(0);
#if 0
                if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE)
                        return(0);
#endif
                if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF)
                        return(0);
                /* fall through */
        case HAMMER2_RESOLVE_ALWAYS:
                break;
        }

        /*
         * Upgrade to an exclusive lock so we can safely manipulate the
         * buffer cache.  If another thread got to it before us we
         * can just return.
         */
        ostate = ccms_thread_lock_upgrade(&core->cst);
        if (chain->data) {
                ccms_thread_lock_downgrade(&core->cst, ostate);
                return (0);
        }

        /*
         * We must resolve to a device buffer, either by issuing I/O or
         * by creating a zero-fill element.  We do not mark the buffer
         * dirty when creating a zero-fill element (the hammer2_chain_modify()
         * API must still be used to do that).
         *
         * The device buffer is variable-sized in powers of 2 down
         * to HAMMER2_MIN_ALLOC (typically 1K).  A 64K physical storage
         * chunk always contains buffers of the same size. (XXX)
         *
         * The minimum physical IO size may be larger than the variable
         * block size.
         */
        bref = &chain->bref;

        /*
         * The getblk() optimization can only be used on newly created
         * elements if the physical block size matches the request.
         */
        if (chain->flags & HAMMER2_CHAIN_INITIAL) {
                error = hammer2_io_new(hmp, bref->data_off, chain->bytes,
                                        &chain->dio);
        } else {
                error = hammer2_io_bread(hmp, bref->data_off, chain->bytes,
                                         &chain->dio);
                adjreadcounter(&chain->bref, chain->bytes);
        }

        if (error) {
                kprintf("hammer2_chain_lock: I/O error %016jx: %d\n",
                        (intmax_t)bref->data_off, error);
                hammer2_io_bqrelse(&chain->dio);
                ccms_thread_lock_downgrade(&core->cst, ostate);
                return (error);
        }

#if 0
        /*
         * No need for this, always require that hammer2_chain_modify()
         * be called before any modifying operations.
         */
        if ((chain->flags & HAMMER2_CHAIN_MODIFIED) &&
            !hammer2_io_isdirty(chain->dio)) {
                hammer2_io_setdirty(chain->dio);
        }
#endif

        /*
         * We can clear the INITIAL state now, we've resolved the buffer
         * to zeros and marked it dirty with hammer2_io_new().
         */
        bdata = hammer2_io_data(chain->dio, chain->bref.data_off);
        if (chain->flags & HAMMER2_CHAIN_INITIAL) {
                atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
        }

        /*
         * Setup the data pointer, either pointing it to an embedded data
         * structure and copying the data from the buffer, or pointing it
         * into the buffer.
         *
         * The buffer is not retained when copying to an embedded data
         * structure in order to avoid potential deadlocks or recursions
         * on the same physical buffer.
         */
        switch (bref->type) {
        case HAMMER2_BREF_TYPE_VOLUME:
        case HAMMER2_BREF_TYPE_FREEMAP:
                /*
                 * Copy data from bp to embedded buffer
                 */
                panic("hammer2_chain_lock: called on unresolved volume header");
                break;
        case HAMMER2_BREF_TYPE_INODE:
        case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
        case HAMMER2_BREF_TYPE_INDIRECT:
        case HAMMER2_BREF_TYPE_DATA:
        case HAMMER2_BREF_TYPE_FREEMAP_NODE:
        default:
                /*
                 * Point data at the device buffer and leave dio intact.
                 */
                chain->data = (void *)bdata;
                break;
        }
        ccms_thread_lock_downgrade(&core->cst, ostate);
        return (0);
}

/*
 * This basically calls hammer2_io_breadcb() but does some pre-processing
 * of the chain first to handle certain cases.
 */
void
hammer2_chain_load_async(hammer2_chain_t *chain,
                         void (*callback)(hammer2_io_t *dio,
                                          hammer2_chain_t *chain,
                                          void *arg_p, off_t arg_o),
                         void *arg_p, off_t arg_o)
{
        hammer2_mount_t *hmp;
        struct hammer2_io *dio;
        hammer2_blockref_t *bref;
        int error;

        if (chain->data) {
                callback(NULL, chain, arg_p, arg_o);
                return;
        }

        /*
         * We must resolve to a device buffer, either by issuing I/O or
         * by creating a zero-fill element.  We do not mark the buffer
         * dirty when creating a zero-fill element (the hammer2_chain_modify()
         * API must still be used to do that).
         *
         * The device buffer is variable-sized in powers of 2 down
         * to HAMMER2_MIN_ALLOC (typically 1K).  A 64K physical storage
         * chunk always contains buffers of the same size. (XXX)
         *
         * The minimum physical IO size may be larger than the variable
         * block size.
         */
        bref = &chain->bref;
        hmp = chain->hmp;

        /*
         * The getblk() optimization can only be used on newly created
         * elements if the physical block size matches the request.
         */
        if ((chain->flags & HAMMER2_CHAIN_INITIAL) &&
            chain->bytes == hammer2_devblksize(chain->bytes)) {
                error = hammer2_io_new(hmp, bref->data_off, chain->bytes, &dio);
                KKASSERT(error == 0);
                callback(dio, chain, arg_p, arg_o);
                return;
        }

        /*
         * Otherwise issue a read
         */
        adjreadcounter(&chain->bref, chain->bytes);
        hammer2_io_breadcb(hmp, bref->data_off, chain->bytes,
                           callback, chain, arg_p, arg_o);
}

/*
 * Unlock and deref a chain element.
 *
 * On the last lock release any non-embedded data (chain->dio) will be
 * retired.
 */
void
hammer2_chain_unlock(hammer2_chain_t *chain)
{
        hammer2_chain_core_t *core = chain->core;
        ccms_state_t ostate;
        long *counterp;
        u_int lockcnt;

        /*
         * The core->cst lock can be shared across several chains so we
         * need to track the per-chain lockcnt separately.
         *
         * If multiple locks are present (or being attempted) on this
         * particular chain we can just unlock, drop refs, and return.
         *
         * Otherwise fall-through on the 1->0 transition.
         */
        for (;;) {
                lockcnt = chain->lockcnt;
                KKASSERT(lockcnt > 0);
                cpu_ccfence();
                if (lockcnt > 1) {
                        if (atomic_cmpset_int(&chain->lockcnt,
                                              lockcnt, lockcnt - 1)) {
                                ccms_thread_unlock(&core->cst);
                                hammer2_chain_drop(chain);
                                return;
                        }
                } else {
                        if (atomic_cmpset_int(&chain->lockcnt, 1, 0))
                                break;
                }
                /* retry */
        }

        /*
         * On the 1->0 transition we upgrade the core lock (if necessary)
         * to exclusive for terminal processing.  If after upgrading we find
         * that lockcnt is non-zero, another thread is racing us and will
         * handle the unload for us later on, so just cleanup and return
         * leaving the data/io intact
         *
         * Otherwise if lockcnt is still 0 it is possible for it to become
         * non-zero and race, but since we hold the core->cst lock
         * exclusively all that will happen is that the chain will be
         * reloaded after we unload it.
         */
        ostate = ccms_thread_lock_upgrade(&core->cst);
        if (chain->lockcnt) {
                ccms_thread_unlock_upgraded(&core->cst, ostate);
                hammer2_chain_drop(chain);
                return;
        }

        /*
         * Shortcut the case if the data is embedded or not resolved.
         *
         * Do NOT NULL out chain->data (e.g. inode data), it might be
         * dirty.
         */
        if (chain->dio == NULL) {
                if ((chain->flags & HAMMER2_CHAIN_MODIFIED) == 0)
                        hammer2_chain_drop_data(chain, 0);
                ccms_thread_unlock_upgraded(&core->cst, ostate);
                hammer2_chain_drop(chain);
                return;
        }

        /*
         * Statistics
         */
        if (hammer2_io_isdirty(chain->dio) == 0) {
                ;
        } else if (chain->flags & HAMMER2_CHAIN_IOFLUSH) {
                switch(chain->bref.type) {
                case HAMMER2_BREF_TYPE_DATA:
                        counterp = &hammer2_ioa_file_write;
                        break;
                case HAMMER2_BREF_TYPE_INODE:
                        counterp = &hammer2_ioa_meta_write;
                        break;
                case HAMMER2_BREF_TYPE_INDIRECT:
                        counterp = &hammer2_ioa_indr_write;
                        break;
                case HAMMER2_BREF_TYPE_FREEMAP_NODE:
                case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
                        counterp = &hammer2_ioa_fmap_write;
                        break;
                default:
                        counterp = &hammer2_ioa_volu_write;
                        break;
                }
                *counterp += chain->bytes;
        } else {
                switch(chain->bref.type) {
                case HAMMER2_BREF_TYPE_DATA:
                        counterp = &hammer2_iod_file_write;
                        break;
                case HAMMER2_BREF_TYPE_INODE:
                        counterp = &hammer2_iod_meta_write;
                        break;
                case HAMMER2_BREF_TYPE_INDIRECT:
                        counterp = &hammer2_iod_indr_write;
                        break;
                case HAMMER2_BREF_TYPE_FREEMAP_NODE:
                case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
                        counterp = &hammer2_iod_fmap_write;
                        break;
                default:
                        counterp = &hammer2_iod_volu_write;
                        break;
                }
                *counterp += chain->bytes;
        }

        /*
         * Clean out the dio.
         *
         * If a device buffer was used for data be sure to destroy the
         * buffer when we are done to avoid aliases (XXX what about the
         * underlying VM pages?).
         *
         * NOTE: Freemap leaf's use reserved blocks and thus no aliasing
         *       is possible.
         *
         * NOTE: The isdirty check tracks whether we have to bdwrite() the
         *       buffer or not.  The buffer might already be dirty.  The
         *       flag is re-set when chain_modify() is called, even if
         *       MODIFIED is already set, allowing the OS to retire the
         *       buffer independent of a hammer2 flush.
         */
        chain->data = NULL;
        if ((chain->flags & HAMMER2_CHAIN_IOFLUSH) &&
            hammer2_io_isdirty(chain->dio)) {
                hammer2_io_bawrite(&chain->dio);
        } else {
                hammer2_io_bqrelse(&chain->dio);
        }
        ccms_thread_unlock_upgraded(&core->cst, ostate);
        hammer2_chain_drop(chain);
}

/*
 * This counts the number of live blockrefs in a block array and
 * also calculates the point at which all remaining blockrefs are empty.
 * This routine can only be called on a live chain (DUPLICATED flag not set).
 *
 * NOTE: Flag is not set until after the count is complete, allowing
 *       callers to test the flag without holding the spinlock.
 *
 * NOTE: If base is NULL the related chain is still in the INITIAL
 *       state and there are no blockrefs to count.
 *
 * NOTE: live_count may already have some counts accumulated due to
 *       creation and deletion and could even be initially negative.
 */
void
hammer2_chain_countbrefs(hammer2_chain_t *chain,
                         hammer2_blockref_t *base, int count)
{
        hammer2_chain_core_t *core = chain->core;

        KKASSERT((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0);

        spin_lock(&core->cst.spin);
        if ((core->flags & HAMMER2_CORE_COUNTEDBREFS) == 0) {
                if (base) {
                        while (--count >= 0) {
                                if (base[count].type)
                                        break;
                        }
                        core->live_zero = count + 1;
                        while (count >= 0) {
                                if (base[count].type)
                                        atomic_add_int(&core->live_count, 1);
                                --count;
                        }
                } else {
                        core->live_zero = 0;
                }
                /* else do not modify live_count */
                atomic_set_int(&core->flags, HAMMER2_CORE_COUNTEDBREFS);
        }
        spin_unlock(&core->cst.spin);
}

/*
 * Resize the chain's physical storage allocation in-place.  This may
 * replace the passed-in chain with a new chain.
 *
 * Chains can be resized smaller without reallocating the storage.
 * Resizing larger will reallocate the storage.
 *
 * Must be passed an exclusively locked parent and chain, returns a new
 * exclusively locked chain at the same index and unlocks the old chain.
 * Flushes the buffer if necessary.
 *
 * This function is mostly used with DATA blocks locked RESOLVE_NEVER in order
 * to avoid instantiating a device buffer that conflicts with the vnode
 * data buffer.  That is, the passed-in bp is a logical buffer, whereas
 * any chain-oriented bp would be a device buffer.
 *
 * XXX return error if cannot resize.
 */
void
hammer2_chain_resize(hammer2_trans_t *trans, hammer2_inode_t *ip,
                     hammer2_chain_t *parent, hammer2_chain_t **chainp,
                     int nradix, int flags)
{
        hammer2_mount_t *hmp;
        hammer2_chain_t *chain;
        size_t obytes;
        size_t nbytes;

        chain = *chainp;
        hmp = chain->hmp;

        /*
         * Only data and indirect blocks can be resized for now.
         * (The volu root, inodes, and freemap elements use a fixed size).
         */
        KKASSERT(chain != &hmp->vchain);
        KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA ||
                 chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT);

        /*
         * Nothing to do if the element is already the proper size
         */
        obytes = chain->bytes;
        nbytes = 1U << nradix;
        if (obytes == nbytes)
                return;

        /*
         * Delete the old chain and duplicate it at the same (parent, index),
         * returning a new chain.  This allows the old chain to still be
         * used by the flush code.  The new chain will be returned in a
         * modified state.
         *
         * The parent does not have to be locked for the delete/duplicate call,
         * but is in this particular code path.
         *
         * NOTE: If we are not crossing a synchronization point the
         *       duplication code will simply reuse the existing chain
         *       structure.
         */
        hammer2_chain_delete_duplicate(trans, &chain, 0);

        /*
         * Relocate the block, even if making it smaller (because different
         * block sizes may be in different regions).
         *
         * (data blocks only, we aren't copying the storage here).
         */
        hammer2_freemap_alloc(trans, chain, nbytes);
        chain->bytes = nbytes;
        atomic_clear_int(&chain->flags, HAMMER2_CHAIN_FORCECOW);
        /*ip->delta_dcount += (ssize_t)(nbytes - obytes);*/ /* XXX atomic */

        /*
         * For now just support it on DATA chains (and not on indirect
         * blocks).
         */
        KKASSERT(chain->dio == NULL);

        *chainp = chain;
}

/*
 * Set a chain modified, making it read-write and duplicating it if necessary.
 * This function will assign a new physical block to the chain if necessary
 *
 * Duplication of already-modified chains is possible when the modification
 * crosses a flush synchronization boundary.
 *
 * Non-data blocks - The chain should be locked to at least the RESOLVE_MAYBE
 *                   level or the COW operation will not work.
 *
 * Data blocks     - The chain is usually locked RESOLVE_NEVER so as not to
 *                   run the data through the device buffers.
 *
 * This function may return a different chain than was passed, in which case
 * the old chain will be unlocked and the new chain will be locked.
 *
 * ip->chain may be adjusted by hammer2_chain_modify_ip().
 */
hammer2_inode_data_t *
hammer2_chain_modify_ip(hammer2_trans_t *trans, hammer2_inode_t *ip,
                        hammer2_chain_t **chainp, int flags)
{
        atomic_set_int(&ip->flags, HAMMER2_INODE_MODIFIED);
        hammer2_chain_modify(trans, chainp, flags);
        if (ip->chain != *chainp)
                hammer2_inode_repoint(ip, NULL, *chainp);
        if (ip->vp)
                vsetisdirty(ip->vp);
        return(&ip->chain->data->ipdata);
}

void
hammer2_chain_modify(hammer2_trans_t *trans, hammer2_chain_t **chainp,
                     int flags)
{
        hammer2_mount_t *hmp;
        hammer2_chain_t *chain;
        hammer2_io_t *dio;
        int error;
        int wasinitial;
        char *bdata;

        chain = *chainp;
        hmp = chain->hmp;

        KKASSERT(chain->bref.mirror_tid != trans->sync_tid ||
                 (chain->flags & HAMMER2_CHAIN_MODIFIED));

        /*
         * data is not optional for freemap chains (we must always be sure
         * to copy the data on COW storage allocations).
         */
        if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
            chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
                KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) ||
                         (flags & HAMMER2_MODIFY_OPTDATA) == 0);
        }

        /*
         * Determine if a delete-duplicate is needed.
         *
         * (a) Modify_tid is part of a prior flush
         * (b) Transaction is concurrent with a flush (has higher tid)
         * (c) and chain is not in the initial state (freshly created)
         * (d) and caller didn't request an in-place modification.
         *
         * The freemap and volume header special chains are never D-Dd.
         */
        if (chain->modify_tid != trans->sync_tid &&        /* cross boundary */
            (flags & HAMMER2_MODIFY_INPLACE) == 0) {       /* from d-d */
                if (chain != &hmp->fchain && chain != &hmp->vchain) {
                        KKASSERT((flags & HAMMER2_MODIFY_ASSERTNOCOPY) == 0);
                        hammer2_chain_delete_duplicate(trans, chainp, 0);
                        chain = *chainp;
                }
        }

        /*
         * Data must be resolved if already assigned unless explicitly
         * flagged otherwise.
         */
        if (chain->data == NULL && (flags & HAMMER2_MODIFY_OPTDATA) == 0 &&
            (chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX)) {
                hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
                hammer2_chain_unlock(chain);
        }

        /*
         * Otherwise do initial-chain handling.  Set MODIFIED to indicate
         * that the chain has been modified.  Set FLUSH_CREATE to flush
         * the new blockref (the D-D set FLUSH_DELETE on the old chain to
         * delete the old blockref).
         */
        if ((chain->flags & HAMMER2_CHAIN_MODIFIED) == 0) {
                atomic_set_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
                hammer2_chain_ref(chain);
                hammer2_chain_memory_inc(chain->pmp);
        }
        if ((chain->flags & HAMMER2_CHAIN_FLUSH_CREATE) == 0) {
                atomic_set_int(&chain->flags, HAMMER2_CHAIN_FLUSH_CREATE);
                hammer2_chain_ref(chain);
        }

        /*
         * The modification or re-modification requires an allocation and
         * possible COW.
         *
         * We normally always allocate new storage here.  If storage exists
         * and MODIFY_NOREALLOC is passed in, we do not allocate new storage.
         */
        if (chain != &hmp->vchain && chain != &hmp->fchain) {
                if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0 ||
                     ((flags & HAMMER2_MODIFY_NOREALLOC) == 0 &&
                      chain->modify_tid != trans->sync_tid)
                ) {
                        hammer2_freemap_alloc(trans, chain, chain->bytes);
                        /* XXX failed allocation */
                } else if (chain->flags & HAMMER2_CHAIN_FORCECOW) {
                        hammer2_freemap_alloc(trans, chain, chain->bytes);
                        /* XXX failed allocation */
                }
                atomic_clear_int(&chain->flags, HAMMER2_CHAIN_FORCECOW);
        }

        /*
         * Update modify_tid.  XXX special-case vchain/fchain because they
         * are always modified in-place.  Otherwise the chain being modified
         * must not be part of a future transaction.
         */
        if (chain == &hmp->vchain || chain == &hmp->fchain) {
                if (chain->modify_tid <= trans->sync_tid)
                        chain->modify_tid = trans->sync_tid;
        } else {
                KKASSERT(chain->modify_tid <= trans->sync_tid);
                chain->modify_tid = trans->sync_tid;
        }

        if ((flags & HAMMER2_MODIFY_NO_MODIFY_TID) == 0)
                chain->bref.modify_tid = trans->sync_tid;

        /*
         * Do not COW BREF_TYPE_DATA when OPTDATA is set.  This is because
         * data modifications are done via the logical buffer cache so COWing
         * it here would result in unnecessary extra copies (and possibly extra
         * block reallocations).  The INITIAL flag remains unchanged in this
         * situation.
         *
         * (This is a bit of a hack).
         */
        if (chain->bref.type == HAMMER2_BREF_TYPE_DATA &&
            (flags & HAMMER2_MODIFY_OPTDATA)) {
                goto skip2;
        }

        /*
         * Clearing the INITIAL flag (for indirect blocks) indicates that
         * we've processed the uninitialized storage allocation.
         *
         * If this flag is already clear we are likely in a copy-on-write
         * situation but we have to be sure NOT to bzero the storage if
         * no data is present.
         */
        if (chain->flags & HAMMER2_CHAIN_INITIAL) {
                atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
                wasinitial = 1;
        } else {
                wasinitial = 0;
        }

        /*
         * Instantiate data buffer and possibly execute COW operation
         */
        switch(chain->bref.type) {
        case HAMMER2_BREF_TYPE_VOLUME:
        case HAMMER2_BREF_TYPE_FREEMAP:
                /*
                 * The data is embedded, no copy-on-write operation is
                 * needed.
                 */
                KKASSERT(chain->dio == NULL);
                break;
        case HAMMER2_BREF_TYPE_INODE:
        case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
        case HAMMER2_BREF_TYPE_DATA:
        case HAMMER2_BREF_TYPE_INDIRECT:
        case HAMMER2_BREF_TYPE_FREEMAP_NODE:
                /*
                 * Perform the copy-on-write operation
                 *
                 * zero-fill or copy-on-write depending on whether
                 * chain->data exists or not and set the dirty state for
                 * the new buffer.  hammer2_io_new() will handle the
                 * zero-fill.
                 */
                KKASSERT(chain != &hmp->vchain && chain != &hmp->fchain);

                if (wasinitial) {
                        error = hammer2_io_new(hmp, chain->bref.data_off,
                                               chain->bytes, &dio);
                } else {
                        error = hammer2_io_bread(hmp, chain->bref.data_off,
                                                 chain->bytes, &dio);
                }
                adjreadcounter(&chain->bref, chain->bytes);
                KKASSERT(error == 0);

                bdata = hammer2_io_data(dio, chain->bref.data_off);

                if (chain->data) {
                        KKASSERT(chain->dio != NULL);
                        if (chain->data != (void *)bdata) {
                                bcopy(chain->data, bdata, chain->bytes);
                        }
                } else if (wasinitial == 0) {
                        /*
                         * We have a problem.  We were asked to COW but
                         * we don't have any data to COW with!
                         */
                        panic("hammer2_chain_modify: having a COW %p\n",
                              chain);
                }

                /*
                 * Retire the old buffer, replace with the new
                 */
                if (chain->dio)
                        hammer2_io_brelse(&chain->dio);
                chain->data = (void *)bdata;
                chain->dio = dio;
                hammer2_io_setdirty(dio);       /* modified by bcopy above */
                break;
        default:
                panic("hammer2_chain_modify: illegal non-embedded type %d",
                      chain->bref.type);
                break;

        }
skip2:
        hammer2_chain_setsubmod(trans, chain);
}

/*
 * Mark the volume as having been modified.  This short-cut version
 * does not have to lock the volume's chain, which allows the ioctl
 * code to make adjustments to connections without deadlocking.  XXX
 *
 * No ref is made on vchain when flagging it MODIFIED.
 */
void
hammer2_modify_volume(hammer2_mount_t *hmp)
{
        hammer2_voldata_lock(hmp);
        hammer2_voldata_unlock(hmp, 1);
}

/*
 * This function returns the chain at the nearest key within the specified
 * range with the highest delete_tid.  The core spinlock must be held on
 * call and the returned chain will be referenced but not locked.
 *
 * The returned chain may or may not be in a deleted state.  Note that
 * live chains have a delete_tid = MAX_TID.
 *
 * This function will recurse through chain->rbtree as necessary and will
 * return a *key_nextp suitable for iteration.  *key_nextp is only set if
 * the iteration value is less than the current value of *key_nextp.
 *
 * The caller should use (*key_nextp) to calculate the actual range of
 * the returned element, which will be (key_beg to *key_nextp - 1), because
 * there might be another element which is superior to the returned element
 * and overlaps it.
 *
 * (*key_nextp) can be passed as key_beg in an iteration only while non-NULL
 * chains continue to be returned.  On EOF (*key_nextp) may overflow since
 * it will wind up being (key_end + 1).
 */
struct hammer2_chain_find_info {
        hammer2_chain_t         *best;
        hammer2_key_t           key_beg;
        hammer2_key_t           key_end;
        hammer2_key_t           key_next;
};

static int hammer2_chain_find_cmp(hammer2_chain_t *child, void *data);
static int hammer2_chain_find_callback(hammer2_chain_t *child, void *data);

static
hammer2_chain_t *
hammer2_chain_find(hammer2_chain_t *parent, hammer2_key_t *key_nextp,
                          hammer2_key_t key_beg, hammer2_key_t key_end)
{
        struct hammer2_chain_find_info info;

        info.best = NULL;
        info.key_beg = key_beg;
        info.key_end = key_end;
        info.key_next = *key_nextp;

        KKASSERT(parent->core->good == 0x1234);
        RB_SCAN(hammer2_chain_tree, &parent->core->rbtree,
                hammer2_chain_find_cmp, hammer2_chain_find_callback,
                &info);
        *key_nextp = info.key_next;
#if 0
        kprintf("chain_find %p %016jx:%016jx next=%016jx\n",
                parent, key_beg, key_end, *key_nextp);
#endif

        return (info.best);
}

/*
 * Find a deleted chain covering a block table entry.  Be careful to deal
 * with the race condition where the block table has been updated but the
 * chain has not yet been removed from dbtree (due to multiple parents having
 * to be updated).
 */
static
hammer2_chain_t *
hammer2_chain_find_deleted(hammer2_chain_t *parent,
                          hammer2_key_t key_beg, hammer2_key_t key_end)
{
        struct hammer2_chain_find_info info;
        hammer2_chain_t *child;

        info.best = NULL;
        info.key_beg = key_beg;
        info.key_end = key_end;
        info.key_next = 0;

        KKASSERT(parent->core->good == 0x1234);
        RB_SCAN(hammer2_chain_tree, &parent->core->dbtree,
                hammer2_chain_find_cmp, hammer2_chain_find_callback,
                &info);
        if ((child = info.best) != NULL) {
                if (child->delete_tid <= parent->update_lo)
                        child = NULL;
        }
        return child;
}

static
int
hammer2_chain_find_cmp(hammer2_chain_t *child, void *data)
{
        struct hammer2_chain_find_info *info = data;
        hammer2_key_t child_beg;
        hammer2_key_t child_end;

        child_beg = child->bref.key;
        child_end = child_beg + ((hammer2_key_t)1 << child->bref.keybits) - 1;

        if (child_end < info->key_beg)
                return(-1);
        if (child_beg > info->key_end)
                return(1);
        return(0);
}

static
int
hammer2_chain_find_callback(hammer2_chain_t *child, void *data)
{
        struct hammer2_chain_find_info *info = data;
        hammer2_chain_t *best;
        hammer2_key_t child_end;

        /*
         * WARNING! Do not discard DUPLICATED chains, it is possible that
         *          we are catching an insertion half-way done.  If a
         *          duplicated chain turns out to be the best choice the
         *          caller will re-check its flags after locking it.
         *
         * WARNING! Layerq is scanned forwards, exact matches should keep
         *          the existing info->best.
         */
        if ((best = info->best) == NULL) {
                /*
                 * No previous best.  Assign best
                 */
                info->best = child;
        } else if (best->bref.key <= info->key_beg &&
                   child->bref.key <= info->key_beg) {
                /*
                 * If our current best is flush with key_beg and child is
                 * also flush with key_beg choose based on delete_tid.
                 *
                 * key_next will automatically be limited to the smaller of
                 * the two end-points.
                 */
                if (child->delete_tid > best->delete_tid)
                        info->best = child;
        } else if (child->bref.key < best->bref.key) {
                /*
                 * Child has a nearer key and best is not flush with key_beg.
                 * Truncate key_next to the old best key iff it had a better
                 * delete_tid.
                 */
                info->best = child;
                if (best->delete_tid >= child->delete_tid &&
                    (info->key_next > best->bref.key || info->key_next == 0))
                        info->key_next = best->bref.key;
        } else if (child->bref.key == best->bref.key) {
                /*
                 * If our current best is flush with the child then choose
                 * based on delete_tid.
                 *
                 * key_next will automatically be limited to the smaller of
                 * the two end-points.
                 */
                if (child->delete_tid > best->delete_tid)
                        info->best = child;
        } else {
                /*
                 * Keep the current best but truncate key_next to the child's
                 * base iff the child has a higher delete_tid.
                 *
                 * key_next will also automatically be limited to the smaller
                 * of the two end-points (probably not necessary for this case
                 * but we do it anyway).
                 */
                if (child->delete_tid >= best->delete_tid &&
                    (info->key_next > child->bref.key || info->key_next == 0))
                        info->key_next = child->bref.key;
        }

        /*
         * Always truncate key_next based on child's end-of-range.
         */
        child_end = child->bref.key + ((hammer2_key_t)1 << child->bref.keybits);
        if (child_end && (info->key_next > child_end || info->key_next == 0))
                info->key_next = child_end;

        return(0);
}

/*
 * Retrieve the specified chain from a media blockref, creating the
 * in-memory chain structure which reflects it.  modify_tid will be
 * left 0 which forces any modifications to issue a delete-duplicate.
 *
 * To handle insertion races pass the INSERT_RACE flag along with the
 * generation number of the core.  NULL will be returned if the generation
 * number changes before we have a chance to insert the chain.  Insert
 * races can occur because the parent might be held shared.
 *
 * Caller must hold the parent locked shared or exclusive since we may
 * need the parent's bref array to find our block.
 */
hammer2_chain_t *
hammer2_chain_get(hammer2_chain_t *parent, int generation,
                  hammer2_blockref_t *bref)
{
        hammer2_mount_t *hmp = parent->hmp;
        hammer2_chain_core_t *above = parent->core;
        hammer2_chain_t *chain;
        int error;

        /*
         * Allocate a chain structure representing the existing media
         * entry.  Resulting chain has one ref and is not locked.
         */
        chain = hammer2_chain_alloc(hmp, parent->pmp, NULL, bref);
        hammer2_chain_core_alloc(NULL, chain, NULL);
        /* ref'd chain returned */

        /*
         * Set modify_tid and update_lo to the chain's synchronization
         * point from the media.
         */
        chain->modify_tid = chain->bref.mirror_tid;
        chain->update_lo = chain->bref.mirror_tid;
        atomic_set_int(&chain->flags, HAMMER2_CHAIN_BMAPPED);

        /*
         * Link the chain into its parent.  A spinlock is required to safely
         * access the RBTREE, and it is possible to collide with another
         * hammer2_chain_get() operation because the caller might only hold
         * a shared lock on the parent.
         */
        KKASSERT(parent->refs > 0);
        error = hammer2_chain_insert(above, NULL, chain,
                                     HAMMER2_CHAIN_INSERT_SPIN |
                                     HAMMER2_CHAIN_INSERT_RACE,
                                     generation);
        if (error) {
                KKASSERT((chain->flags & (HAMMER2_CHAIN_ONRBTREE |
                                          HAMMER2_CHAIN_ONDBTREE |
                                          HAMMER2_CHAIN_ONDBQ)) == 0);
                kprintf("chain %p get race\n", chain);
                hammer2_chain_drop(chain);
                chain = NULL;
        } else {
                KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
        }

        /*
         * Return our new chain referenced but not locked, or NULL if
         * a race occurred.
         */
        return (chain);
}

/*
 * Lookup initialization/completion API
 */
hammer2_chain_t *
hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags)
{
        if (flags & HAMMER2_LOOKUP_SHARED) {
                hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
                                           HAMMER2_RESOLVE_SHARED);
        } else {
                hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
        }
        return (parent);
}

void
hammer2_chain_lookup_done(hammer2_chain_t *parent)
{
        if (parent)
                hammer2_chain_unlock(parent);
}

static
hammer2_chain_t *
hammer2_chain_getparent(hammer2_chain_t **parentp, int how)
{
        hammer2_chain_t *oparent;
        hammer2_chain_t *bparent;
        hammer2_chain_t *nparent;
        hammer2_chain_core_t *above;

        oparent = *parentp;
        above = oparent->above;

        spin_lock(&above->cst.spin);
        bparent = TAILQ_FIRST(&above->ownerq);
        hammer2_chain_ref(bparent);

        /*
         * Be careful of order, oparent must be unlocked before nparent
         * is locked below to avoid a deadlock.  We might as well delay its
         * unlocking until we conveniently no longer have the spinlock (instead
         * of cycling the spinlock).
         *
         * Theoretically our ref on bparent should prevent elements of the
         * following chain from going away and prevent above from going away,
         * but we still need the spinlock to safely scan the list.
         */
        for (;;) {
                nparent = bparent;
                while (nparent->flags & HAMMER2_CHAIN_DUPLICATED)
                        nparent = TAILQ_NEXT(nparent, core_entry);
                hammer2_chain_ref(nparent);
                spin_unlock(&above->cst.spin);

                if (oparent) {
                        hammer2_chain_unlock(oparent);
                        oparent = NULL;
                }
                hammer2_chain_lock(nparent, how | HAMMER2_RESOLVE_NOREF);
                hammer2_chain_drop(bparent);

                /*
                 * We might have raced a delete-duplicate.
                 */
                if ((nparent->flags & HAMMER2_CHAIN_DUPLICATED) == 0)
                        break;
                bparent = nparent;
                hammer2_chain_ref(bparent);
                hammer2_chain_unlock(nparent);
                spin_lock(&above->cst.spin);
                /* retry */
        }
        *parentp = nparent;

        return (nparent);
}

/*
 * Locate the first chain whos key range overlaps (key_beg, key_end) inclusive.
 * (*parentp) typically points to an inode but can also point to a related
 * indirect block and this function will recurse upwards and find the inode
 * again.
 *
 * (*parentp) must be exclusively locked and referenced and can be an inode
 * or an existing indirect block within the inode.
 *
 * On return (*parentp) will be modified to point at the deepest parent chain
 * element encountered during the search, as a helper for an insertion or
 * deletion.   The new (*parentp) will be locked and referenced and the old
 * will be unlocked and dereferenced (no change if they are both the same).
 *
 * The matching chain will be returned exclusively locked.  If NOLOCK is
 * requested the chain will be returned only referenced.
 *
 * NULL is returned if no match was found, but (*parentp) will still
 * potentially be adjusted.
 *
 * On return (*key_nextp) will point to an iterative value for key_beg.
 * (If NULL is returned (*key_nextp) is set to key_end).
 *
 * This function will also recurse up the chain if the key is not within the
 * current parent's range.  (*parentp) can never be set to NULL.  An iteration
 * can simply allow (*parentp) to float inside the loop.
 *
 * NOTE!  chain->data is not always resolved.  By default it will not be
 *        resolved for BREF_TYPE_DATA, FREEMAP_NODE, or FREEMAP_LEAF.  Use
 *        HAMMER2_LOOKUP_ALWAYS to force resolution (but be careful w/
 *        BREF_TYPE_DATA as the device buffer can alias the logical file
 *        buffer).
 */
hammer2_chain_t *
hammer2_chain_lookup(hammer2_chain_t **parentp, hammer2_key_t *key_nextp,
                     hammer2_key_t key_beg, hammer2_key_t key_end,
                     int *cache_indexp, int flags)
{
        hammer2_mount_t *hmp;
        hammer2_chain_t *parent;
        hammer2_chain_t *chain;
        hammer2_blockref_t *base;
        hammer2_blockref_t *bref;
        hammer2_blockref_t bcopy;
        hammer2_key_t scan_beg;
        hammer2_key_t scan_end;
        hammer2_chain_core_t *above;
        int count = 0;
        int how_always = HAMMER2_RESOLVE_ALWAYS;
        int how_maybe = HAMMER2_RESOLVE_MAYBE;
        int how;
        int generation;
        int maxloops = 300000;
        int wasdup;

        if (flags & HAMMER2_LOOKUP_ALWAYS) {
                how_maybe = how_always;
                how = HAMMER2_RESOLVE_ALWAYS;
        } else if (flags & (HAMMER2_LOOKUP_NODATA | HAMMER2_LOOKUP_NOLOCK)) {
                how = HAMMER2_RESOLVE_NEVER;
        } else {
                how = HAMMER2_RESOLVE_MAYBE;
        }
        if (flags & (HAMMER2_LOOKUP_SHARED | HAMMER2_LOOKUP_NOLOCK)) {
                how_maybe |= HAMMER2_RESOLVE_SHARED;
                how_always |= HAMMER2_RESOLVE_SHARED;
                how |= HAMMER2_RESOLVE_SHARED;
        }

        /*
         * Recurse (*parentp) upward if necessary until the parent completely
         * encloses the key range or we hit the inode.
         *
         * This function handles races against the flusher doing a delete-
         * duplicate above us and re-homes the parent to the duplicate in
         * that case, otherwise we'd wind up recursing down a stale chain.
         */
        parent = *parentp;
        hmp = parent->hmp;

        while (parent->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
               parent->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
                scan_beg = parent->bref.key;
                scan_end = scan_beg +
                           ((hammer2_key_t)1 << parent->bref.keybits) - 1;
                if (key_beg >= scan_beg && key_end <= scan_end)
                        break;
                parent = hammer2_chain_getparent(parentp, how_maybe);
        }

again:
        if (--maxloops == 0)
                panic("hammer2_chain_lookup: maxloops");
        /*
         * Locate the blockref array.  Currently we do a fully associative
         * search through the array.
         */
        switch(parent->bref.type) {
        case HAMMER2_BREF_TYPE_INODE:
                /*
                 * Special shortcut for embedded data returns the inode
                 * itself.  Callers must detect this condition and access
                 * the embedded data (the strategy code does this for us).
                 *
                 * This is only applicable to regular files and softlinks.
                 */
                if (parent->data->ipdata.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
                        if (flags & HAMMER2_LOOKUP_NOLOCK)
                                hammer2_chain_ref(parent);
                        else
                                hammer2_chain_lock(parent, how_always);
                        *key_nextp = key_end + 1;
                        return (parent);
                }
                base = &parent->data->ipdata.u.blockset.blockref[0];
                count = HAMMER2_SET_COUNT;
                break;
        case HAMMER2_BREF_TYPE_FREEMAP_NODE:
        case HAMMER2_BREF_TYPE_INDIRECT:
                /*
                 * Handle MATCHIND on the parent
                 */
                if (flags & HAMMER2_LOOKUP_MATCHIND) {
                        scan_beg = parent->bref.key;
                        scan_end = scan_beg +
                               ((hammer2_key_t)1 << parent->bref.keybits) - 1;
                        if (key_beg == scan_beg && key_end == scan_end) {
                                chain = parent;
                                hammer2_chain_lock(chain, how_maybe);
                                *key_nextp = scan_end + 1;
                                goto done;
                        }
                }
                /*
                 * Optimize indirect blocks in the INITIAL state to avoid
                 * I/O.
                 */
                if (parent->flags & HAMMER2_CHAIN_INITIAL) {
                        base = NULL;
                } else {
                        if (parent->data == NULL)
                                panic("parent->data is NULL");
                        base = &parent->data->npdata[0];
                }
                count = parent->bytes / sizeof(hammer2_blockref_t);
                break;
        case HAMMER2_BREF_TYPE_VOLUME:
                base = &hmp->voldata.sroot_blockset.blockref[0];
                count = HAMMER2_SET_COUNT;
                break;
        case HAMMER2_BREF_TYPE_FREEMAP:
                base = &hmp->voldata.freemap_blockset.blockref[0];
                count = HAMMER2_SET_COUNT;
                break;
        default:
                panic("hammer2_chain_lookup: unrecognized blockref type: %d",
                      parent->bref.type);
                base = NULL;    /* safety */
                count = 0;      /* safety */
        }

        /*
         * Merged scan to find next candidate.
         *
         * hammer2_base_*() functions require the above->live_* fields
         * to be synchronized.
         *
         * We need to hold the spinlock to access the block array and RB tree
         * and to interlock chain creation.
         */
        above = parent->core;
        if ((parent->core->flags & HAMMER2_CORE_COUNTEDBREFS) == 0)
                hammer2_chain_countbrefs(parent, base, count);

        /*
         * Combined search
         */
        spin_lock(&above->cst.spin);
        chain = hammer2_combined_find(parent, base, count,
                                      cache_indexp, key_nextp,
                                      key_beg, key_end,
                                      &bref);
        generation = above->generation;

        /*
         * Exhausted parent chain, iterate.
         */
        if (bref == NULL) {
                spin_unlock(&above->cst.spin);
                if (key_beg == key_end) /* short cut single-key case */
                        return (NULL);
                return (hammer2_chain_next(parentp, NULL, key_nextp,
                                           key_beg, key_end,
                                           cache_indexp, flags));
        }

        /*
         * Selected from blockref or in-memory chain.
         */
        if (chain == NULL) {
                bcopy = *bref;
                spin_unlock(&above->cst.spin);
                chain = hammer2_chain_get(parent, generation,
                                          &bcopy);
                if (chain == NULL) {
                        kprintf("retry lookup parent %p keys %016jx:%016jx\n",
                                parent, key_beg, key_end);
                        goto again;
                }
                if (bcmp(&bcopy, bref, sizeof(bcopy))) {
                        hammer2_chain_drop(chain);
                        goto again;
                }
                wasdup = 0;
        } else {
                hammer2_chain_ref(chain);
                wasdup = ((chain->flags & HAMMER2_CHAIN_DUPLICATED) != 0);
                spin_unlock(&above->cst.spin);
        }

        /*
         * chain is referenced but not locked.  We must lock the chain
         * to obtain definitive DUPLICATED/DELETED state
         */
        if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
            chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
                hammer2_chain_lock(chain, how_maybe | HAMMER2_RESOLVE_NOREF);
        } else {
                hammer2_chain_lock(chain, how | HAMMER2_RESOLVE_NOREF);
        }

        /*
         * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
         *
         * NOTE: Chain's key range is not relevant as there might be
         *       one-offs within the range that are not deleted.
         *
         * NOTE: Lookups can race delete-duplicate because
         *       delete-duplicate does not lock the parent's core
         *       (they just use the spinlock on the core).  We must
         *       check for races by comparing the DUPLICATED flag before
         *       releasing the spinlock with the flag after locking the
         *       chain.
         */
        if (chain->flags & HAMMER2_CHAIN_DELETED) {
                hammer2_chain_unlock(chain);
                if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0 || wasdup) {
                        key_beg = *key_nextp;
                        if (key_beg == 0 || key_beg > key_end)
                                return(NULL);
                }
                goto again;
        }

        /*
         * If the chain element is an indirect block it becomes the new
         * parent and we loop on it.  We must maintain our top-down locks
         * to prevent the flusher from interfering (i.e. doing a
         * delete-duplicate and leaving us recursing down a deleted chain).
         *
         * The parent always has to be locked with at least RESOLVE_MAYBE
         * so we can access its data.  It might need a fixup if the caller
         * passed incompatible flags.  Be careful not to cause a deadlock
         * as a data-load requires an exclusive lock.
         *
         * If HAMMER2_LOOKUP_MATCHIND is set and the indirect block's key
         * range is within the requested key range we return the indirect
         * block and do NOT loop.  This is usually only used to acquire
         * freemap nodes.
         */
        if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT ||
            chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) {
                hammer2_chain_unlock(parent);
                *parentp = parent = chain;
                goto again;
        }
done:
        /*
         * All done, return the chain
         */
        return (chain);
}

/*
 * After having issued a lookup we can iterate all matching keys.
 *
 * If chain is non-NULL we continue the iteration from just after it's index.
 *
 * If chain is NULL we assume the parent was exhausted and continue the
 * iteration at the next parent.
 *
 * parent must be locked on entry and remains locked throughout.  chain's
 * lock status must match flags.  Chain is always at least referenced.
 *
 * WARNING!  The MATCHIND flag does not apply to this function.
 */
hammer2_chain_t *
hammer2_chain_next(hammer2_chain_t **parentp, hammer2_chain_t *chain,
                   hammer2_key_t *key_nextp,
                   hammer2_key_t key_beg, hammer2_key_t key_end,
                   int *cache_indexp, int flags)
{
        hammer2_chain_t *parent;
        int how_maybe;

        /*
         * Calculate locking flags for upward recursion.
         */
        how_maybe = HAMMER2_RESOLVE_MAYBE;
        if (flags & (HAMMER2_LOOKUP_SHARED | HAMMER2_LOOKUP_NOLOCK))
                how_maybe |= HAMMER2_RESOLVE_SHARED;

        parent = *parentp;

        /*
         * Calculate the next index and recalculate the parent if necessary.
         */
        if (chain) {
                key_beg = chain->bref.key +
                          ((hammer2_key_t)1 << chain->bref.keybits);
                if (flags & HAMMER2_LOOKUP_NOLOCK)
                        hammer2_chain_drop(chain);
                else
                        hammer2_chain_unlock(chain);

                /*
                 * Any scan where the lookup returned degenerate data embedded
                 * in the inode has an invalid index and must terminate.
                 */
                if (chain == parent)
                        return(NULL);
                if (key_beg == 0 || key_beg > key_end)
                        return(NULL);
                chain = NULL;
        } else if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT &&
                   parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) {
                /*
                 * We reached the end of the iteration.
                 */
                return (NULL);
        } else {
                /*
                 * Continue iteration with next parent unless the current
                 * parent covers the range.
                 */
                key_beg = parent->bref.key +
                          ((hammer2_key_t)1 << parent->bref.keybits);
                if (key_beg == 0 || key_beg > key_end)
                        return (NULL);
                parent = hammer2_chain_getparent(parentp, how_maybe);
        }

        /*
         * And execute
         */
        return (hammer2_chain_lookup(parentp, key_nextp,
                                     key_beg, key_end,
                                     cache_indexp, flags));
}

/*
 * The raw scan function is similar to lookup/next but does not seek to a key.
 * Blockrefs are iterated via first_chain = (parent, NULL) and
 * next_chain = (parent, chain).
 *
 * The passed-in parent must be locked and its data resolved.  The returned
 * chain will be locked.  Pass chain == NULL to acquire the first sub-chain
 * under parent and then iterate with the passed-in chain (which this
 * function will unlock).
 */
hammer2_chain_t *
hammer2_chain_scan(hammer2_chain_t *parent, hammer2_chain_t *chain,
                   int *cache_indexp, int flags)
{
        hammer2_mount_t *hmp;
        hammer2_blockref_t *base;
        hammer2_blockref_t *bref;
        hammer2_blockref_t bcopy;
        hammer2_chain_core_t *above;
        hammer2_key_t key;
        hammer2_key_t next_key;
        int count = 0;
        int how_always = HAMMER2_RESOLVE_ALWAYS;
        int how_maybe = HAMMER2_RESOLVE_MAYBE;
        int how;
        int generation;
        int maxloops = 300000;
        int wasdup;

        hmp = parent->hmp;

        /*
         * Scan flags borrowed from lookup
         */
        if (flags & HAMMER2_LOOKUP_ALWAYS) {
                how_maybe = how_always;
                how = HAMMER2_RESOLVE_ALWAYS;
        } else if (flags & (HAMMER2_LOOKUP_NODATA | HAMMER2_LOOKUP_NOLOCK)) {
                how = HAMMER2_RESOLVE_NEVER;
        } else {
                how = HAMMER2_RESOLVE_MAYBE;
        }
        if (flags & (HAMMER2_LOOKUP_SHARED | HAMMER2_LOOKUP_NOLOCK)) {
                how_maybe |= HAMMER2_RESOLVE_SHARED;
                how_always |= HAMMER2_RESOLVE_SHARED;
                how |= HAMMER2_RESOLVE_SHARED;
        }

        /*
         * Calculate key to locate first/next element, unlocking the previous
         * element as we go.  Be careful, the key calculation can overflow.
         */
        if (chain) {
                key = chain->bref.key +
                      ((hammer2_key_t)1 << chain->bref.keybits);
                hammer2_chain_unlock(chain);
                chain = NULL;
                if (key == 0)
                        goto done;
        } else {
                key = 0;
        }

again:
        if (--maxloops == 0)
                panic("hammer2_chain_scan: maxloops");
        /*
         * Locate the blockref array.  Currently we do a fully associative
         * search through the array.
         */
        switch(parent->bref.type) {
        case HAMMER2_BREF_TYPE_INODE:
                /*
                 * An inode with embedded data has no sub-chains.
                 */
                if (parent->data->ipdata.op_flags & HAMMER2_OPFLAG_DIRECTDATA)
                        goto done;
                base = &parent->data->ipdata.u.blockset.blockref[0];
                count = HAMMER2_SET_COUNT;
                break;
        case HAMMER2_BREF_TYPE_FREEMAP_NODE:
        case HAMMER2_BREF_TYPE_INDIRECT:
                /*
                 * Optimize indirect blocks in the INITIAL state to avoid
                 * I/O.
                 */
                if (parent->flags & HAMMER2_CHAIN_INITIAL) {
                        base = NULL;
                } else {
                        if (parent->data == NULL)
                                panic("parent->data is NULL");
                        base = &parent->data->npdata[0];
                }
                count = parent->bytes / sizeof(hammer2_blockref_t);
                break;
        case HAMMER2_BREF_TYPE_VOLUME:
                base = &hmp->voldata.sroot_blockset.blockref[0];
                count = HAMMER2_SET_COUNT;
                break;
        case HAMMER2_BREF_TYPE_FREEMAP:
                base = &hmp->voldata.freemap_blockset.blockref[0];
                count = HAMMER2_SET_COUNT;
                break;
        default:
                panic("hammer2_chain_lookup: unrecognized blockref type: %d",
                      parent->bref.type);
                base = NULL;    /* safety */
                count = 0;      /* safety */
        }

        /*
         * Merged scan to find next candidate.
         *
         * hammer2_base_*() functions require the above->live_* fields
         * to be synchronized.
         *
         * We need to hold the spinlock to access the block array and RB tree
         * and to interlock chain creation.
         */
        if ((parent->core->flags & HAMMER2_CORE_COUNTEDBREFS) == 0)
                hammer2_chain_countbrefs(parent, base, count);

        above = parent->core;
        next_key = 0;
        spin_lock(&above->cst.spin);
        chain = hammer2_combined_find(parent, base, count,
                                      cache_indexp, &next_key,
                                      key, HAMMER2_MAX_KEY,
                                      &bref);
        generation = above->generation;

        /*
         * Exhausted parent chain, we're done.
         */
        if (bref == NULL) {
                spin_unlock(&above->cst.spin);
                KKASSERT(chain == NULL);
                goto done;
        }

        /*
         * Selected from blockref or in-memory chain.
         */
        if (chain == NULL) {
                bcopy = *bref;
                spin_unlock(&above->cst.spin);
                chain = hammer2_chain_get(parent, generation, &bcopy);
                if (chain == NULL) {
                        kprintf("retry scan parent %p keys %016jx\n",
                                parent, key);
                        goto again;
                }
                if (bcmp(&bcopy, bref, sizeof(bcopy))) {
                        hammer2_chain_drop(chain);
                        chain = NULL;
                        goto again;
                }
                wasdup = 0;
        } else {
                hammer2_chain_ref(chain);
                wasdup = ((chain->flags & HAMMER2_CHAIN_DUPLICATED) != 0);
                spin_unlock(&above->cst.spin);
        }

        /*
         * chain is referenced but not locked.  We must lock the chain
         * to obtain definitive DUPLICATED/DELETED state
         */
        hammer2_chain_lock(chain, how | HAMMER2_RESOLVE_NOREF);

        /*
         * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX)
         *
         * NOTE: chain's key range is not relevant as there might be
         *       one-offs within the range that are not deleted.
         *
         * NOTE: XXX this could create problems with scans used in
         *       situations other than mount-time recovery.
         *
         * NOTE: Lookups can race delete-duplicate because
         *       delete-duplicate does not lock the parent's core
         *       (they just use the spinlock on the core).  We must
         *       check for races by comparing the DUPLICATED flag before
         *       releasing the spinlock with the flag after locking the
         *       chain.
         */
        if (chain->flags & HAMMER2_CHAIN_DELETED) {
                hammer2_chain_unlock(chain);
                chain = NULL;

                if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0 || wasdup) {
                        key = next_key;
                        if (key == 0)
                                goto done;
                }
                goto again;
        }

done:
        /*
         * All done, return the chain or NULL
         */
        return (chain);
}

/*
 * Create and return a new hammer2 system memory structure of the specified
 * key, type and size and insert it under (*parentp).  This is a full
 * insertion, based on the supplied key/keybits, and may involve creating
 * indirect blocks and moving other chains around via delete/duplicate.
 *
 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (*parentp) TO THE INSERTION
 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
 * FULL.  This typically means that the caller is creating the chain after
 * doing a hammer2_chain_lookup().
 *
 * (*parentp) must be exclusive locked and may be replaced on return
 * depending on how much work the function had to do.
 *
 * (*chainp) usually starts out NULL and returns the newly created chain,
 * but if the caller desires the caller may allocate a disconnected chain
 * and pass it in instead.  (It is also possible for the caller to use
 * chain_duplicate() to create a disconnected chain, manipulate it, then
 * pass it into this function to insert it).
 *
 * This function should NOT be used to insert INDIRECT blocks.  It is
 * typically used to create/insert inodes and data blocks.
 *
 * Caller must pass-in an exclusively locked parent the new chain is to
 * be inserted under, and optionally pass-in a disconnected, exclusively
 * locked chain to insert (else we create a new chain).  The function will
 * adjust (*parentp) as necessary, create or connect the chain, and
 * return an exclusively locked chain in *chainp.
 */
int
hammer2_chain_create(hammer2_trans_t *trans, hammer2_chain_t **parentp,
                     hammer2_chain_t **chainp,
                     hammer2_key_t key, int keybits, int type, size_t bytes)
{
        hammer2_mount_t *hmp;
        hammer2_chain_t *chain;
        hammer2_chain_t *parent = *parentp;
        hammer2_chain_core_t *above;
        hammer2_blockref_t *base;
        hammer2_blockref_t dummy;
        int allocated = 0;
        int error = 0;
        int count;
        int maxloops = 300000;

        above = parent->core;
        KKASSERT(ccms_thread_lock_owned(&above->cst));
        hmp = parent->hmp;
        chain = *chainp;

        if (chain == NULL) {
                /*
                 * First allocate media space and construct the dummy bref,
                 * then allocate the in-memory chain structure.  Set the
                 * INITIAL flag for fresh chains which do not have embedded
                 * data.
                 */
                bzero(&dummy, sizeof(dummy));
                dummy.type = type;
                dummy.key = key;
                dummy.keybits = keybits;
                dummy.data_off = hammer2_getradix(bytes);
                dummy.methods = parent->bref.methods;
                chain = hammer2_chain_alloc(hmp, parent->pmp, trans, &dummy);
                hammer2_chain_core_alloc(trans, chain, NULL);

                /*
                 * Lock the chain manually, chain_lock will load the chain
                 * which we do NOT want to do.  (note: chain->refs is set
                 * to 1 by chain_alloc() for us, but lockcnt is not).
                 */
                chain->lockcnt = 1;
                ccms_thread_lock(&chain->core->cst, CCMS_STATE_EXCLUSIVE);
                allocated = 1;

                /*
                 * We do NOT set INITIAL here (yet).  INITIAL is only
                 * used for indirect blocks.
                 *
                 * Recalculate bytes to reflect the actual media block
                 * allocation.
                 */
                bytes = (hammer2_off_t)1 <<
                        (int)(chain->bref.data_off & HAMMER2_OFF_MASK_RADIX);
                chain->bytes = bytes;

                switch(type) {
                case HAMMER2_BREF_TYPE_VOLUME:
                case HAMMER2_BREF_TYPE_FREEMAP:
                        panic("hammer2_chain_create: called with volume type");
                        break;
                case HAMMER2_BREF_TYPE_INDIRECT:
                        panic("hammer2_chain_create: cannot be used to"
                              "create indirect block");
                        break;
                case HAMMER2_BREF_TYPE_FREEMAP_NODE:
                        panic("hammer2_chain_create: cannot be used to"
                              "create freemap root or node");
                        break;
                case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
                        KKASSERT(bytes == sizeof(chain->data->bmdata));
                        /* fall through */
                case HAMMER2_BREF_TYPE_INODE:
                case HAMMER2_BREF_TYPE_DATA:
                default:
                        /*
                         * leave chain->data NULL, set INITIAL
                         */
                        KKASSERT(chain->data == NULL);
                        atomic_set_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
                        break;
                }
        } else {
                /*
                 * We are reattaching a chain that has been duplicated and
                 * left disconnected under a DIFFERENT parent with potentially
                 * different key/keybits.
                 *
                 * The chain must be modified in the current transaction
                 * (the duplication code should have done that for us),
                 * and it's modify_tid should be greater than the parent's
                 * bref.mirror_tid.  This should cause it to be created under
                 * the new parent.
                 *
                 * If deleted in the same transaction, the create/delete TIDs
                 * will be the same and effective the chain will not have
                 * existed at all from the point of view of the parent.
                 *
                 * Do NOT mess with the current state of the INITIAL flag.
                 */
                KKASSERT(chain->modify_tid == trans->sync_tid);
                chain->bref.key = key;
                chain->bref.keybits = keybits;
                KKASSERT(chain->above == NULL);
        }

        /*
         * Calculate how many entries we have in the blockref array and
         * determine if an indirect block is required.
         */
again:
        if (--maxloops == 0)
                panic("hammer2_chain_create: maxloops");
        above = parent->core;

        switch(parent->bref.type) {
        case HAMMER2_BREF_TYPE_INODE:
                KKASSERT((parent->data->ipdata.op_flags &
                          HAMMER2_OPFLAG_DIRECTDATA) == 0);
                KKASSERT(parent->data != NULL);
                base = &parent->data->ipdata.u.blockset.blockref[0];
                count = HAMMER2_SET_COUNT;
                break;
        case HAMMER2_BREF_TYPE_INDIRECT:
        case HAMMER2_BREF_TYPE_FREEMAP_NODE:
                if (parent->flags & HAMMER2_CHAIN_INITIAL)
                        base = NULL;
                else
                        base = &parent->data->npdata[0];
                count = parent->bytes / sizeof(hammer2_blockref_t);
                break;
        case HAMMER2_BREF_TYPE_VOLUME:
                KKASSERT(parent->data != NULL);
                base = &hmp->voldata.sroot_blockset.blockref[0];
                count = HAMMER2_SET_COUNT;
                break;
        case HAMMER2_BREF_TYPE_FREEMAP:
                KKASSERT(parent->data != NULL);
                base = &hmp->voldata.freemap_blockset.blockref[0];
                count = HAMMER2_SET_COUNT;
                break;
        default:
                panic("hammer2_chain_create: unrecognized blockref type: %d",
                      parent->bref.type);
                base = NULL;
                count = 0;
                break;
        }

        /*
         * Make sure we've counted the brefs
         */
        if ((parent->core->flags & HAMMER2_CORE_COUNTEDBREFS) == 0)
                hammer2_chain_countbrefs(parent, base, count);

        KKASSERT(above->live_count >= 0 && above->live_count <= count);

        /*
         * If no free blockref could be found we must create an indirect
         * block and move a number of blockrefs into it.  With the parent
         * locked we can safely lock each child in order to delete+duplicate
         * it without causing a deadlock.
         *
         * This may return the new indirect block or the old parent depending
         * on where the key falls.  NULL is returned on error.
         */
        if (above->live_count == count) {
                hammer2_chain_t *nparent;

                nparent = hammer2_chain_create_indirect(trans, parent,
                                                        key, keybits,
                                                        type, &error);
                if (nparent == NULL) {
                        if (allocated)
                                hammer2_chain_drop(chain);
                        chain = NULL;
                        goto done;
                }
                if (parent != nparent) {
                        hammer2_chain_unlock(parent);
                        parent = *parentp = nparent;
                }
                goto again;
        }

        /*
         * Link the chain into its parent.
         */
        if (chain->above != NULL)
                panic("hammer2: hammer2_chain_create: chain already connected");
        KKASSERT(chain->above == NULL);
        hammer2_chain_insert(above, NULL, chain,
                             HAMMER2_CHAIN_INSERT_SPIN |
                             HAMMER2_CHAIN_INSERT_LIVE,
                             0);

        if (allocated) {
                /*
                 * Mark the newly created chain modified.  This will cause
                 * FLUSH_CREATE to be set.
                 *
                 * Device buffers are not instantiated for DATA elements
                 * as these are handled by logical buffers.
                 *
                 * Indirect and freemap node indirect blocks are handled
                 * by hammer2_chain_create_indirect() and not by this
                 * function.
                 *
                 * Data for all other bref types is expected to be
                 * instantiated (INODE, LEAF).
                 */
                switch(chain->bref.type) {
                case HAMMER2_BREF_TYPE_DATA:
                case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
                case HAMMER2_BREF_TYPE_INODE:
                        hammer2_chain_modify(trans, &chain,
                                             HAMMER2_MODIFY_OPTDATA |
                                             HAMMER2_MODIFY_ASSERTNOCOPY);
                        break;
                default:
                        /*
                         * Remaining types are not supported by this function.
                         * In particular, INDIRECT and LEAF_NODE types are
                         * handled by create_indirect().
                         */
                        panic("hammer2_chain_create: bad type: %d",
                              chain->bref.type);
                        /* NOT REACHED */
                        break;
                }
        } else {
                /*
                 * When reconnecting a chain we must set FLUSH_CREATE and
                 * setsubmod so the flush recognizes that it must update
                 * the bref in the parent.
                 */
                if ((chain->flags & HAMMER2_CHAIN_FLUSH_CREATE) == 0) {
                        hammer2_chain_ref(chain);
                        atomic_set_int(&chain->flags,
                                       HAMMER2_CHAIN_FLUSH_CREATE);
                }
        }
        hammer2_chain_setsubmod(trans, chain);

done:
        *chainp = chain;

        return (error);
}

/*
 * Replace (*chainp) with a duplicate in-memory chain structure which shares
 * the same core and media state as the orignal.  The original *chainp is
 * unlocked and the replacement will be returned locked.  The duplicated
 * chain is inserted under (*parentp).
 *
 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (*parentp) TO THE INSERTION
 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING
 * FULL.  This typically means that the caller is creating the chain after
 * doing a hammer2_chain_lookup().
 *
 * The old chain must be in a DELETED state unless snapshot is non-zero.
 *
 * The new chain will be live (i.e. not deleted), and modified.
 *
 * If (parent) is non-NULL then the new duplicated chain is inserted under
 * the parent.
 *
 * If (parent) is NULL then the newly duplicated chain is not inserted
 * anywhere, similar to if it had just been chain_alloc()'d (suitable for
 * passing into hammer2_chain_create() after this function returns).
 *
 * WARNING! This function cannot take snapshots all by itself.  The caller
 *          needs to do other massaging for snapshots.
 *
 * WARNING! This function calls create which means it can insert indirect
 *          blocks.  Callers may have to refactor locked chains held across
 *          the call (other than the ones passed into the call).
 */
void
hammer2_chain_duplicate(hammer2_trans_t *trans, hammer2_chain_t **parentp,
                        hammer2_chain_t **chainp, hammer2_blockref_t *bref,
                        int snapshot, int duplicate_reason)
{
        hammer2_mount_t *hmp;
        hammer2_chain_t *parent;
        hammer2_chain_t *ochain;
        hammer2_chain_t *nchain;
        hammer2_chain_core_t *above;
        size_t bytes;

        /*
         * We want nchain to be our go-to live chain, but ochain may be in
         * a MODIFIED state within the current flush synchronization segment.
         * Force any further modifications of ochain to do another COW
         * operation even if modify_tid indicates that one is not needed.
         *
         * We don't want to set FORCECOW on nchain simply as an optimization,
         * as many duplication calls simply move chains into ichains and
         * then delete the original.
         *
         * WARNING!  We should never resolve DATA to device buffers
         *           (XXX allow it if the caller did?), and since
         *           we currently do not have the logical buffer cache
         *           buffer in-hand to fix its cached physical offset
         *           we also force the modify code to not COW it. XXX
         */
        ochain = *chainp;
        hmp = ochain->hmp;
        KKASSERT(snapshot == 1 || (ochain->flags & HAMMER2_CHAIN_DELETED));

        /*
         * Now create a duplicate of the chain structure, associating
         * it with the same core, making it the same size, pointing it
         * to the same bref (the same media block).
         *
         * Give nchain the same modify_tid that we previously ensured was
         * sufficiently advanced to trigger a block table insertion on flush.
         *
         * nchain copies ochain's data and must inherit ochain->update_lo.
         *
         * NOTE: bref.mirror_tid duplicated by virtue of bref copy in
         *       hammer2_chain_alloc()
         */
        if (bref == NULL)
                bref = &ochain->bref;
        if (snapshot) {
                nchain = hammer2_chain_alloc(hmp, NULL, trans, bref);
                atomic_set_int(&nchain->flags, HAMMER2_CHAIN_SNAPSHOT);
        } else {
                nchain = hammer2_chain_alloc(hmp, ochain->pmp, trans, bref);
        }
        hammer2_chain_core_alloc(trans, nchain, ochain);
        bytes = (hammer2_off_t)1 <<
                (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
        nchain->bytes = bytes;
        nchain->modify_tid = ochain->modify_tid;
        nchain->update_lo = ochain->update_lo;
        nchain->inode_reason = ochain->inode_reason + 0x100000;
        atomic_set_int(&nchain->flags,
                       ochain->flags & (HAMMER2_CHAIN_INITIAL |
                                        HAMMER2_CHAIN_FORCECOW |
                                        HAMMER2_CHAIN_UNLINKED));
        if (ochain->modify_tid == trans->sync_tid)
                atomic_set_int(&ochain->flags, HAMMER2_CHAIN_FORCECOW);

        /*
         * Switch from ochain to nchain
         */
        hammer2_chain_lock(nchain, HAMMER2_RESOLVE_NEVER |
                                   HAMMER2_RESOLVE_NOREF);
        /* nchain has 1 ref */
        hammer2_chain_unlock(ochain);

        /*
         * Place nchain in the modified state, instantiate media data
         * if necessary.  Because modify_tid is already completely
         * synchronized this should not result in a delete-duplicate.
         *
         * We want nchain at the target to look like a new insertion.
         * Forcing the modification to be INPLACE accomplishes this
         * because we get the same nchain with an updated modify_tid.
         */
        if (nchain->bref.type == HAMMER2_BREF_TYPE_DATA) {
                hammer2_chain_modify(trans, &nchain,
                                     HAMMER2_MODIFY_OPTDATA |
                                     HAMMER2_MODIFY_NOREALLOC |
                                     HAMMER2_MODIFY_INPLACE);
        } else if (nchain->flags & HAMMER2_CHAIN_INITIAL) {
                hammer2_chain_modify(trans, &nchain,
                                     HAMMER2_MODIFY_OPTDATA |
                                     HAMMER2_MODIFY_INPLACE);
        } else {
                hammer2_chain_modify(trans, &nchain,
                                     HAMMER2_MODIFY_INPLACE);
        }

        /*
         * If parent is not NULL the duplicated chain will be entered under
         * the parent and the FLUSH_CREATE bit set to tell flush to update
         * the blockref.
         *
         * Having both chains locked is extremely important for atomicy.
         */
        if (parentp && (parent = *parentp) != NULL) {
                above = parent->core;
                KKASSERT(ccms_thread_lock_owned(&above->cst));
                KKASSERT((nchain->flags & HAMMER2_CHAIN_DELETED) == 0);
                KKASSERT(parent->refs > 0);

                hammer2_chain_create(trans, parentp, &nchain,
                                     nchain->bref.key, nchain->bref.keybits,
                                     nchain->bref.type, nchain->bytes);
                parent = NULL;

                KKASSERT(nchain->flags & HAMMER2_CHAIN_FLUSH_CREATE);
                hammer2_chain_setsubmod(trans, nchain);
        }

        *chainp = nchain;
}

/*
 * Helper function for deleting chains.
 *
 * The chain is removed from the live view (the RBTREE).
 *
 * If appropriate, the chain is added to the shadow topology and FLUSH_DELETE
 * is set for flusher visbility.  The caller is responsible for calling
 * setsubmod on chain, so we do not adjust update_hi here.
 */
static void
_hammer2_chain_delete_helper(hammer2_trans_t *trans,
                             hammer2_chain_core_t *above,
                             hammer2_chain_t *chain)
{
        hammer2_mount_t *hmp;
        hammer2_chain_t *xchain;

        KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE);
        KKASSERT(trans->sync_tid >= chain->modify_tid);
        KKASSERT((chain->flags & (HAMMER2_CHAIN_DELETED |
                                  HAMMER2_CHAIN_ONDBQ |
                                  HAMMER2_CHAIN_ONDBTREE |
                                  HAMMER2_CHAIN_FLUSH_DELETE)) == 0);

        /*
         * Flag as deleted, reduce live_count and bump the above core's
         * generation.
         */
        chain->delete_tid = trans->sync_tid;
        atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED);
        atomic_add_int(&above->live_count, -1);
        ++above->generation;
        hmp = chain->hmp;

        /*
         * Remove from live tree
         */
        RB_REMOVE(hammer2_chain_tree, &above->rbtree, chain);
        atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE);

        if (chain->flags & HAMMER2_CHAIN_BMAPPED) {
                /*
                 * If the chain was originally bmapped we must place on the
                 * deleted tree and set FLUSH_DELETE (+ref) to prevent
                 * destruction of the chain until the flush can reconcile
                 * the parent's block table.
                 *
                 * NOTE! DBTREE is only representitive of the live view,
                 *       the flush must check both DBTREE and DBQ.
                 */
                xchain = RB_INSERT(hammer2_chain_tree, &above->dbtree, chain);
                KKASSERT(xchain == NULL);
                atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONDBTREE);

                atomic_set_int(&chain->flags, HAMMER2_CHAIN_FLUSH_DELETE);
                hammer2_chain_ref(chain);
        } else {
                /*
                 * If the chain no longer (and never had) an actual blockmap
                 * entry we must place it on the dbq list and set FLUSH_DELETE
                 * (+ref) to prevent destruction of the chain until the flush
                 * can reconcile the parent's block table.
                 *
                 * NOTE! DBTREE is only representitive of the live view,
                 *       the flush must check both DBTREE and DBQ.
                 */
                TAILQ_INSERT_TAIL(&above->dbq, chain, db_entry);
                atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONDBQ);

                atomic_set_int(&chain->flags, HAMMER2_CHAIN_FLUSH_DELETE);
                hammer2_chain_ref(chain);
        }
}

/*
 * Special in-place delete-duplicate sequence which does not require a
 * locked parent.  (*chainp) is marked DELETED and atomically replaced
 * with a duplicate.  Atomicy is at the very-fine spin-lock level in
 * order to ensure that lookups do not race us.
 *
 * The flush code will sometimes call this function with a deleted chain.
 * In this situation the old chain's memory is reallocated without
 * duplicating it.
 *
 * The new chain will be marked modified for the current transaction.
 */
void
hammer2_chain_delete_duplicate(hammer2_trans_t *trans, hammer2_chain_t **chainp,
                               int flags)
{
        hammer2_mount_t *hmp;
        hammer2_chain_t *ochain;
        hammer2_chain_t *nchain;
        hammer2_chain_core_t *above;
        size_t bytes;

        if (hammer2_debug & 0x20000)
                Debugger("dd");

        /*
         * Note that we do not have to call setsubmod on ochain, calling it
         * on nchain is sufficient.
         */
        ochain = *chainp;
        hmp = ochain->hmp;

        if (ochain->bref.type == HAMMER2_BREF_TYPE_INODE) {
                KKASSERT(ochain->data);
        }

        /*
         * First create a duplicate of the chain structure.
         * (nchain is allocated with one ref).
         *
         * In the case where nchain inherits ochains core, nchain is
         * effectively locked due to ochain being locked (and sharing the
         * core), until we can give nchain its own official ock.
         *
         * WARNING! Flusher concurrency can create two cases.  The first is
         *          that the flusher might be working on a chain that has
         *          been deleted in the live view but is live in the flusher's
         *          view.  In the second case the flusher may be duplicating
         *          a forward-transacted chain.  In both situations nchain
         *          must be marked deleted.
         *
         * WARNING! hammer2_chain_core_alloc() also acts on these issues.
         */
        nchain = hammer2_chain_alloc(hmp, ochain->pmp, trans, &ochain->bref);
        if ((ochain->flags & HAMMER2_CHAIN_DELETED) ||
            (ochain->modify_tid > trans->sync_tid)) {
                atomic_set_int(&nchain->flags, HAMMER2_CHAIN_DELETED);
        }
        if (flags & HAMMER2_DELDUP_RECORE)
                hammer2_chain_core_alloc(trans, nchain, NULL);
        else
                hammer2_chain_core_alloc(trans, nchain, ochain);
        above = ochain->above;

        bytes = (hammer2_off_t)1 <<
                (int)(ochain->bref.data_off & HAMMER2_OFF_MASK_RADIX);
        nchain->bytes = bytes;

        /*
         * nchain inherits ochain's live state including its modification
         * state.  This function disposes of the original.  Because we are
         * doing this in-place under the same parent the block array
         * inserted/deleted state does not change.
         *
         * nchain copies ochain's data and must inherit ochain->update_lo.
         *
         * If ochain was previously marked FORCECOW we also flag nchain
         * FORCECOW (used during hardlink splits).  FORCECOW forces a
         * reallocation of the block when we modify the chain a little later,
         * it does not force another delete-duplicate.
         *
         * NOTE: bref.mirror_tid duplicated by virtue of bref copy in
         *       hammer2_chain_alloc()
         */
        nchain->data_count += ochain->data_count;
        nchain->inode_count += ochain->inode_count;
        atomic_set_int(&nchain->flags,
                       ochain->flags & (HAMMER2_CHAIN_INITIAL |
                                        HAMMER2_CHAIN_FORCECOW |
                                        HAMMER2_CHAIN_UNLINKED));
        if (ochain->modify_tid == trans->sync_tid)
                atomic_set_int(&ochain->flags, HAMMER2_CHAIN_FORCECOW);
        nchain->inode_reason = ochain->inode_reason + 0x1000;
        nchain->update_lo = ochain->update_lo;
        nchain->dsrc = ochain->bref;                    /* DEBUG */
        nchain->dsrc_dupfromat = trans->sync_tid;       /* DEBUG */
        nchain->dsrc_dupfromflags = trans->flags;       /* DEBUG */
        nchain->dsrc_reason = ochain->inode_reason;     /* DEBUG */
        nchain->dsrc_ninserts = ochain->ninserts;       /* DEBUG */
        nchain->dsrc_flags = ochain->flags;             /* DEBUG */
        nchain->dsrc_modify = ochain->modify_tid;       /* DEBUG */
        nchain->dsrc_delete = ochain->delete_tid;       /* DEBUG */
        nchain->dsrc_update_lo = ochain->update_lo;     /* DEBUG */
        nchain->dsrc_original = ochain;                 /* DEBUG */

        /*
         * Lock nchain so both chains are now locked (extremely important
         * for atomicy).  The shared core allows us to unlock ochain without
         * actually unlocking ochain.
         */
        hammer2_chain_lock(nchain, HAMMER2_RESOLVE_NEVER);
        /* extra ref still present from original allocation */

        KKASSERT(ochain->flags & (HAMMER2_CHAIN_ONRBTREE |
                                  HAMMER2_CHAIN_ONDBTREE |
                                  HAMMER2_CHAIN_ONDBQ));
        spin_lock(&above->cst.spin);

        nchain->modify_tid = ochain->modify_tid;
        nchain->delete_tid = HAMMER2_MAX_TID;

        if (nchain->flags & HAMMER2_CHAIN_DELETED) {
                /*
                 * Special case, used by the flush code only in two cases:
                 *
                 * (1) The flush must operate on a chain that is visible to
                 *     the flush but deleted in the live view.
                 *
                 * (2) The flush must operate on a forward-indexed chain in
                 *     the live view, which typically
                 *
                 * In these situations nchain will be marked deleted and
                 * insert before ochain.  nchain must inherit certain features
                 * of ochain such as the BMAPPED state.
                 */
                KKASSERT(trans->flags & HAMMER2_TRANS_ISFLUSH);
                KKASSERT(ochain->modify_tid < trans->sync_tid);
                KKASSERT(ochain->delete_tid > trans->sync_tid);
                atomic_set_int(&nchain->flags, HAMMER2_CHAIN_FLUSH_TEMPORARY);
                hammer2_chain_insert(above, ochain, nchain, 0, 0);

                if ((ochain->flags & HAMMER2_CHAIN_DELETED) &&
                    ochain->modify_tid < trans->sync_tid) {
                        nchain->delete_tid = ochain->delete_tid;
                        ochain->delete_tid = trans->sync_tid;
                } else if (ochain->modify_tid > trans->sync_tid) {
                        nchain->delete_tid = ochain->modify_tid;
                }
        } else {
                /*
                 * Normal case, delete-duplicate deletes ochain and nchain
                 * is the new live chain.
                 */
                _hammer2_chain_delete_helper(trans, above, ochain);
                hammer2_chain_insert(above, ochain, nchain,
                                     HAMMER2_CHAIN_INSERT_LIVE, 0);
        }
        spin_unlock(&above->cst.spin);

        /*
         * ochain must be unlocked because ochain and nchain might share
         * a buffer cache buffer, so we need to release it so nchain can
         * potentially obtain it.
         */
        hammer2_chain_setsubmod(trans, ochain);
        hammer2_chain_unlock(ochain);

        /*
         * Finishing fixing up nchain.  A new block will be allocated if
         * crossing a synchronization point (meta-data only).
         *
         * Calling hammer2_chain_modify() will update modify_tid to
         * (typically) trans->sync_tid.
         */
        if (nchain->bref.type == HAMMER2_BREF_TYPE_DATA) {
                hammer2_chain_modify(trans, &nchain,
                                     HAMMER2_MODIFY_OPTDATA |
                                     HAMMER2_MODIFY_NOREALLOC |
                                     HAMMER2_MODIFY_INPLACE);
        } else if (nchain->flags & HAMMER2_CHAIN_INITIAL) {
                hammer2_chain_modify(trans, &nchain,
                                     HAMMER2_MODIFY_OPTDATA |
                                     HAMMER2_MODIFY_INPLACE);
        } else {
                hammer2_chain_modify(trans, &nchain,
                                     HAMMER2_MODIFY_INPLACE);
        }
        hammer2_chain_drop(nchain);

        /*
         * Unconditionally set FLUSH_CREATE to force the parent blockrefs to
         * update as the chain_modify() above won't necessarily do it.
         */
        if ((nchain->flags & HAMMER2_CHAIN_FLUSH_CREATE) == 0) {
                atomic_set_int(&nchain->flags, HAMMER2_CHAIN_FLUSH_CREATE);
                hammer2_chain_ref(nchain);
        }

        /*
         * If nchain is in a DELETED state we must set FLUSH_DELETE
         */
        if (nchain->flags & HAMMER2_CHAIN_DELETED)
                KKASSERT((nchain->flags & HAMMER2_CHAIN_FLUSH_DELETE) == 0);
#if 1
        if ((nchain->flags & HAMMER2_CHAIN_FLUSH_DELETE) == 0 &&
            (nchain->flags & HAMMER2_CHAIN_DELETED)) {
                atomic_set_int(&nchain->flags, HAMMER2_CHAIN_FLUSH_DELETE);
                hammer2_chain_ref(nchain);
        }
#endif
        hammer2_chain_setsubmod(trans, nchain);
        *chainp = nchain;
}

/*
 * Create a snapshot of the specified {parent, ochain} with the specified
 * label.  The originating hammer2_inode must be exclusively locked for
 * safety.
 *
 * The ioctl code has already synced the filesystem.
 */
int
hammer2_chain_snapshot(hammer2_trans_t *trans, hammer2_chain_t **ochainp,
                       hammer2_ioc_pfs_t *pfs)
{
        hammer2_mount_t *hmp;
        hammer2_chain_t *ochain = *ochainp;
        hammer2_chain_t *nchain;
        hammer2_inode_data_t *ipdata;
        hammer2_inode_t *nip;
        size_t name_len;
        hammer2_key_t lhc;
        struct vattr vat;
        uuid_t opfs_clid;
        int error;

        kprintf("snapshot %s ochain->refs %d ochain->flags %08x\n",
                pfs->name, ochain->refs, ochain->flags);

        name_len = strlen(pfs->name);
        lhc = hammer2_dirhash(pfs->name, name_len);

        hmp = ochain->hmp;
        opfs_clid = ochain->data->ipdata.pfs_clid;

        *ochainp = ochain;

        /*
         * Create the snapshot directory under the super-root
         *
         * Set PFS type, generate a unique filesystem id, and generate
         * a cluster id.  Use the same clid when snapshotting a PFS root,
         * which theoretically allows the snapshot to be used as part of
         * the same cluster (perhaps as a cache).
         *
         * Copy the (flushed) ochain's blockref array.  Theoretically we
         * could use chain_duplicate() but it becomes difficult to disentangle
         * the shared core so for now just brute-force it.
         */
        VATTR_NULL(&vat);
        vat.va_type = VDIR;
        vat.va_mode = 0755;
        nchain = NULL;
        nip = hammer2_inode_create(trans, hmp->sroot, &vat, proc0.p_ucred,
                                   pfs->name, name_len, &nchain, &error);

        if (nip) {
                ipdata = hammer2_chain_modify_ip(trans, nip, &nchain, 0);
                ipdata->pfs_type = HAMMER2_PFSTYPE_SNAPSHOT;
                kern_uuidgen(&ipdata->pfs_fsid, 1);
                if (ochain->flags & HAMMER2_CHAIN_PFSROOT)
                        ipdata->pfs_clid = opfs_clid;
                else
                        kern_uuidgen(&ipdata->pfs_clid, 1);
                atomic_set_int(&nchain->flags, HAMMER2_CHAIN_PFSROOT);
                ipdata->u.blockset = ochain->data->ipdata.u.blockset;

                hammer2_inode_unlock_ex(nip, nchain);
        }
        return (error);
}

/*
 * Create an indirect block that covers one or more of the elements in the
 * current parent.  Either returns the existing parent with no locking or
 * ref changes or returns the new indirect block locked and referenced
 * and leaving the original parent lock/ref intact as well.
 *
 * If an error occurs, NULL is returned and *errorp is set to the error.
 *
 * The returned chain depends on where the specified key falls.
 *
 * The key/keybits for the indirect mode only needs to follow three rules:
 *
 * (1) That all elements underneath it fit within its key space and
 *
 * (2) That all elements outside it are outside its key space.
 *
 * (3) When creating the new indirect block any elements in the current
 *     parent that fit within the new indirect block's keyspace must be
 *     moved into the new indirect block.
 *
 * (4) The keyspace chosen for the inserted indirect block CAN cover a wider
 *     keyspace the the current parent, but lookup/iteration rules will
 *     ensure (and must ensure) that rule (2) for all parents leading up
 *     to the nearest inode or the root volume header is adhered to.  This
 *     is accomplished by always recursing through matching keyspaces in
 *     the hammer2_chain_lookup() and hammer2_chain_next() API.
 *
 * The current implementation calculates the current worst-case keyspace by
 * iterating the current parent and then divides it into two halves, choosing
 * whichever half has the most elements (not necessarily the half containing
 * the requested key).
 *
 * We can also opt to use the half with the least number of elements.  This
 * causes lower-numbered keys (aka logical file offsets) to recurse through
 * fewer indirect blocks and higher-numbered keys to recurse through more.
 * This also has the risk of not moving enough elements to the new indirect
 * block and being forced to create several indirect blocks before the element
 * can be inserted.
 *
 * Must be called with an exclusively locked parent.
 */
static int hammer2_chain_indkey_freemap(hammer2_chain_t *parent,
                                hammer2_key_t *keyp, int keybits,
                                hammer2_blockref_t *base, int count);
static int hammer2_chain_indkey_normal(hammer2_chain_t *parent,
                                hammer2_key_t *keyp, int keybits,
                                hammer2_blockref_t *base, int count);
static
hammer2_chain_t *
hammer2_chain_create_indirect(hammer2_trans_t *trans, hammer2_chain_t *parent,
                              hammer2_key_t create_key, int create_bits,
                              int for_type, int *errorp)
{
        hammer2_mount_t *hmp;
        hammer2_chain_core_t *above;
        hammer2_chain_core_t *icore;
        hammer2_blockref_t *base;
        hammer2_blockref_t *bref;
        hammer2_blockref_t bcopy;
        hammer2_chain_t *chain;
        hammer2_chain_t *ichain;
        hammer2_chain_t dummy;
        hammer2_key_t key = create_key;
        hammer2_key_t key_beg;
        hammer2_key_t key_end;
        hammer2_key_t key_next;
        int keybits = create_bits;
        int count;
        int nbytes;
        int cache_index;
        int loops;
        int reason;
        int generation;
        int maxloops = 300000;
        int retry_same;
        int wasdup;

        /*
         * Calculate the base blockref pointer or NULL if the chain
         * is known to be empty.  We need to calculate the array count
         * for RB lookups either way.
         */
        hmp = parent->hmp;
        *errorp = 0;
        KKASSERT(ccms_thread_lock_owned(&parent->core->cst));
        above = parent->core;

        /*hammer2_chain_modify(trans, &parent, HAMMER2_MODIFY_OPTDATA);*/
        if (parent->flags & HAMMER2_CHAIN_INITIAL) {
                base = NULL;

                switch(parent->bref.type) {
                case HAMMER2_BREF_TYPE_INODE:
                        count = HAMMER2_SET_COUNT;
                        break;
                case HAMMER2_BREF_TYPE_INDIRECT:
                case HAMMER2_BREF_TYPE_FREEMAP_NODE:
                        count = parent->bytes / sizeof(hammer2_blockref_t);
                        break;
                case HAMMER2_BREF_TYPE_VOLUME:
                        count = HAMMER2_SET_COUNT;
                        break;
                case HAMMER2_BREF_TYPE_FREEMAP:
                        count = HAMMER2_SET_COUNT;
                        break;
                default:
                        panic("hammer2_chain_create_indirect: "
                              "unrecognized blockref type: %d",
                              parent->bref.type);
                        count = 0;
                        break;
                }
        } else {
                switch(parent->bref.type) {
                case HAMMER2_BREF_TYPE_INODE:
                        base = &parent->data->ipdata.u.blockset.blockref[0];
                        count = HAMMER2_SET_COUNT;
                        break;
                case HAMMER2_BREF_TYPE_INDIRECT:
                case HAMMER2_BREF_TYPE_FREEMAP_NODE:
                        base = &parent->data->npdata[0];
                        count = parent->bytes / sizeof(hammer2_blockref_t);
                        break;
                case HAMMER2_BREF_TYPE_VOLUME:
                        base = &hmp->voldata.sroot_blockset.blockref[0];
                        count = HAMMER2_SET_COUNT;
                        break;
                case HAMMER2_BREF_TYPE_FREEMAP:
                        base = &hmp->voldata.freemap_blockset.blockref[0];
                        count = HAMMER2_SET_COUNT;
                        break;
                default:
                        panic("hammer2_chain_create_indirect: "
                              "unrecognized blockref type: %d",
                              parent->bref.type);
                        count = 0;
                        break;
                }
        }

        /*
         * dummy used in later chain allocation (no longer used for lookups).
         */
        bzero(&dummy, sizeof(dummy));
        dummy.delete_tid = HAMMER2_MAX_TID;

        /*
         * When creating an indirect block for a freemap node or leaf
         * the key/keybits must be fitted to static radix levels because
         * particular radix levels use particular reserved blocks in the
         * related zone.
         *
         * This routine calculates the key/radix of the indirect block
         * we need to create, and whether it is on the high-side or the
         * low-side.
         */
        if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
            for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
                keybits = hammer2_chain_indkey_freemap(parent, &key, keybits,
                                                       base, count);
        } else {
                keybits = hammer2_chain_indkey_normal(parent, &key, keybits,
                                                      base, count);
        }

        /*
         * Normalize the key for the radix being represented, keeping the
         * high bits and throwing away the low bits.
         */
        key &= ~(((hammer2_key_t)1 << keybits) - 1);

        /*
         * How big should our new indirect block be?  It has to be at least
         * as large as its parent.
         */
        if (parent->bref.type == HAMMER2_BREF_TYPE_INODE)
                nbytes = HAMMER2_IND_BYTES_MIN;
        else
                nbytes = HAMMER2_IND_BYTES_MAX;
        if (nbytes < count * sizeof(hammer2_blockref_t))
                nbytes = count * sizeof(hammer2_blockref_t);

        /*
         * Ok, create our new indirect block
         */
        if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
            for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
                dummy.bref.type = HAMMER2_BREF_TYPE_FREEMAP_NODE;
        } else {
                dummy.bref.type = HAMMER2_BREF_TYPE_INDIRECT;
        }
        dummy.bref.key = key;
        dummy.bref.keybits = keybits;
        dummy.bref.data_off = hammer2_getradix(nbytes);
        dummy.bref.methods = parent->bref.methods;

        ichain = hammer2_chain_alloc(hmp, parent->pmp, trans, &dummy.bref);
        atomic_set_int(&ichain->flags, HAMMER2_CHAIN_INITIAL);
        hammer2_chain_core_alloc(trans, ichain, NULL);
        icore = ichain->core;
        hammer2_chain_lock(ichain, HAMMER2_RESOLVE_MAYBE);
        hammer2_chain_drop(ichain);     /* excess ref from alloc */

        /*
         * We have to mark it modified to allocate its block, but use
         * OPTDATA to allow it to remain in the INITIAL state.  Otherwise
         * it won't be acted upon by the flush code.
         */
        hammer2_chain_modify(trans, &ichain, HAMMER2_MODIFY_OPTDATA);

        /*
         * Iterate the original parent and move the matching brefs into
         * the new indirect block.
         *
         * XXX handle flushes.
         */
        key_beg = 0;
        key_end = HAMMER2_MAX_KEY;
        cache_index = 0;
        spin_lock(&above->cst.spin);
        loops = 0;
        reason = 0;
        retry_same = 0;

        for (;;) {
                if (++loops > 100000) {
                    spin_unlock(&above->cst.spin);
                    panic("excessive loops r=%d p=%p base/count %p:%d %016jx\n",
                          reason, parent, base, count, key_next);
                }

                /*
                 * NOTE: spinlock stays intact, returned chain (if not NULL)
                 *       is not referenced or locked which means that we
                 *       cannot safely check its flagged / deletion status
                 *       until we lock it.
                 */
                chain = hammer2_combined_find(parent, base, count,
                                              &cache_index, &key_next,
                                              key_beg, key_end,
                                              &bref);
                generation = above->generation;
                if (bref == NULL)
                        break;
                key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);

                /*
                 * Skip keys that are not within the key/radix of the new
                 * indirect block.  They stay in the parent.
                 */
                if ((~(((hammer2_key_t)1 << keybits) - 1) &
                    (key ^ bref->key)) != 0) {
                        goto next_key_spinlocked;
                }

                /*
                 * Load the new indirect block by acquiring the related
                 * chains (potentially from media as it might not be
                 * in-memory).  Then move it to the new parent (ichain)
                 * via DELETE-DUPLICATE.
                 *
                 * chain is referenced but not locked.  We must lock the
                 * chain to obtain definitive DUPLICATED/DELETED state
                 */
                if (chain) {
                        /*
                         * Use chain already present in the RBTREE
                         */
                        hammer2_chain_ref(chain);
                        wasdup = ((chain->flags &
                                   HAMMER2_CHAIN_DUPLICATED) != 0);
                        spin_unlock(&above->cst.spin);
                        hammer2_chain_lock(chain, HAMMER2_RESOLVE_NEVER |
                                                  HAMMER2_RESOLVE_NOREF);
                } else {
                        /*
                         * Get chain for blockref element.  _get returns NULL
                         * on insertion race.
                         */
                        bcopy = *bref;
                        spin_unlock(&above->cst.spin);
                        chain = hammer2_chain_get(parent, generation, &bcopy);
                        if (chain == NULL) {
                                reason = 1;
                                spin_lock(&above->cst.spin);
                                continue;
                        }
                        if (bcmp(&bcopy, bref, sizeof(bcopy))) {
                                reason = 2;
                                hammer2_chain_drop(chain);
                                spin_lock(&above->cst.spin);
                                continue;
                        }
                        hammer2_chain_lock(chain, HAMMER2_RESOLVE_NEVER |
                                                  HAMMER2_RESOLVE_NOREF);
                        wasdup = 0;
                }

                /*
                 * This is always live so if the chain has been delete-
                 * duplicated we raced someone and we have to retry.
                 *
                 * NOTE: Lookups can race delete-duplicate because
                 *       delete-duplicate does not lock the parent's core
                 *       (they just use the spinlock on the core).  We must
                 *       check for races by comparing the DUPLICATED flag before
                 *       releasing the spinlock with the flag after locking the
                 *       chain.
                 *
                 *       (note reversed logic for this one)
                 */
                if (chain->flags & HAMMER2_CHAIN_DELETED) {
                        hammer2_chain_unlock(chain);
                        if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) &&
                            wasdup == 0) {
                                retry_same = 1;
                        }
                        goto next_key;
                }

                /*
                 * Shift the chain to the indirect block.
                 *
                 * WARNING! Can cause held-over chains to require a refactor.
                 *          Fortunately we have none (our locked chains are
                 *          passed into and modified by the call).
                 */
                hammer2_chain_delete(trans, chain, 0);
                hammer2_chain_duplicate(trans, &ichain, &chain, NULL, 0, 1);
                hammer2_chain_unlock(chain);
                KKASSERT(parent->refs > 0);
                chain = NULL;
next_key:
                spin_lock(&above->cst.spin);
next_key_spinlocked:
                if (--maxloops == 0)
                        panic("hammer2_chain_create_indirect: maxloops");
                reason = 4;
                if (retry_same == 0) {
                        if (key_next == 0 || key_next > key_end)
                                break;
                        key_beg = key_next;
                }
                /* loop */
        }
        spin_unlock(&above->cst.spin);

        /*
         * Insert the new indirect block into the parent now that we've
         * cleared out some entries in the parent.  We calculated a good
         * insertion index in the loop above (ichain->index).
         *
         * We don't have to set FLUSH_CREATE here because we mark ichain
         * modified down below (so the normal modified -> flush -> set-moved
         * sequence applies).
         *
         * The insertion shouldn't race as this is a completely new block
         * and the parent is locked.
         */
        KKASSERT((ichain->flags & HAMMER2_CHAIN_ONRBTREE) == 0);
        hammer2_chain_insert(above, NULL, ichain,
                             HAMMER2_CHAIN_INSERT_SPIN |
                             HAMMER2_CHAIN_INSERT_LIVE,
                             0);

        /*
         * Mark the new indirect block modified after insertion, which
         * will propagate up through parent all the way to the root and
         * also allocate the physical block in ichain for our caller,
         * and assign ichain->data to a pre-zero'd space (because there
         * is not prior data to copy into it).
         */
        /*hammer2_chain_modify(trans, &ichain, HAMMER2_MODIFY_OPTDATA);*/
        hammer2_chain_setsubmod(trans, ichain);

        /*
         * Figure out what to return.
         */
        if (~(((hammer2_key_t)1 << keybits) - 1) &
                   (create_key ^ key)) {
                /*
                 * Key being created is outside the key range,
                 * return the original parent.
                 */
                hammer2_chain_unlock(ichain);
        } else {
                /*
                 * Otherwise its in the range, return the new parent.
                 * (leave both the new and old parent locked).
                 */
                parent = ichain;
        }

        return(parent);
}

/*
 * Calculate the keybits and highside/lowside of the freemap node the
 * caller is creating.
 *
 * This routine will specify the next higher-level freemap key/radix
 * representing the lowest-ordered set.  By doing so, eventually all
 * low-ordered sets will be moved one level down.
 *
 * We have to be careful here because the freemap reserves a limited
 * number of blocks for a limited number of levels.  So we can't just
 * push indiscriminately.
 */
int
hammer2_chain_indkey_freemap(hammer2_chain_t *parent, hammer2_key_t *keyp,
                             int keybits, hammer2_blockref_t *base, int count)
{
        hammer2_chain_core_t *above;
        hammer2_chain_t *chain;
        hammer2_blockref_t *bref;
        hammer2_key_t key;
        hammer2_key_t key_beg;
        hammer2_key_t key_end;
        hammer2_key_t key_next;
        int cache_index;
        int locount;
        int hicount;
        int maxloops = 300000;

        key = *keyp;
        above = parent->core;
        locount = 0;
        hicount = 0;
        keybits = 64;

        /*
         * Calculate the range of keys in the array being careful to skip
         * slots which are overridden with a deletion.
         */
        key_beg = 0;
        key_end = HAMMER2_MAX_KEY;
        cache_index = 0;
        spin_lock(&above->cst.spin);

        for (;;) {
                if (--maxloops == 0) {
                        panic("indkey_freemap shit %p %p:%d\n",
                              parent, base, count);
                }
                chain = hammer2_combined_find(parent, base, count,
                                              &cache_index, &key_next,
                                              key_beg, key_end,
                                              &bref);

                /*
                 * Exhausted search
                 */
                if (bref == NULL)
                        break;

                /*
                 * NOTE: No need to check DUPLICATED here because we do
                 *       not release the spinlock.
                 */
                if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
                        if (key_next == 0 || key_next > key_end)
                                break;
                        key_beg = key_next;
                        continue;
                }

                /*
                 * Use the full live (not deleted) element for the scan
                 * iteration.  HAMMER2 does not allow partial replacements.
                 *
                 * XXX should be built into hammer2_combined_find().
                 */
                key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);

                if (keybits > bref->keybits) {
                        key = bref->key;
                        keybits = bref->keybits;
                } else if (keybits == bref->keybits && bref->key < key) {
                        key = bref->key;
                }
                if (key_next == 0)
                        break;
                key_beg = key_next;
        }
        spin_unlock(&above->cst.spin);

        /*
         * Return the keybits for a higher-level FREEMAP_NODE covering
         * this node.
         */
        switch(keybits) {
        case HAMMER2_FREEMAP_LEVEL0_RADIX:
                keybits = HAMMER2_FREEMAP_LEVEL1_RADIX;
                break;
        case HAMMER2_FREEMAP_LEVEL1_RADIX:
                keybits = HAMMER2_FREEMAP_LEVEL2_RADIX;
                break;
        case HAMMER2_FREEMAP_LEVEL2_RADIX:
                keybits = HAMMER2_FREEMAP_LEVEL3_RADIX;
                break;
        case HAMMER2_FREEMAP_LEVEL3_RADIX:
                keybits = HAMMER2_FREEMAP_LEVEL4_RADIX;
                break;
        case HAMMER2_FREEMAP_LEVEL4_RADIX:
                panic("hammer2_chain_indkey_freemap: level too high");
                break;
        default:
                panic("hammer2_chain_indkey_freemap: bad radix");
                break;
        }
        *keyp = key;

        return (keybits);
}

/*
 * Calculate the keybits and highside/lowside of the indirect block the
 * caller is creating.
 */
static int
hammer2_chain_indkey_normal(hammer2_chain_t *parent, hammer2_key_t *keyp,
                            int keybits, hammer2_blockref_t *base, int count)
{
        hammer2_chain_core_t *above;
        hammer2_blockref_t *bref;
        hammer2_chain_t *chain;
        hammer2_key_t key_beg;
        hammer2_key_t key_end;
        hammer2_key_t key_next;
        hammer2_key_t key;
        int nkeybits;
        int locount;
        int hicount;
        int cache_index;
        int maxloops = 300000;

        key = *keyp;
        above = parent->core;
        locount = 0;
        hicount = 0;

        /*
         * Calculate the range of keys in the array being careful to skip
         * slots which are overridden with a deletion.  Once the scan
         * completes we will cut the key range in half and shift half the
         * range into the new indirect block.
         */
        key_beg = 0;
        key_end = HAMMER2_MAX_KEY;
        cache_index = 0;
        spin_lock(&above->cst.spin);

        for (;;) {
                if (--maxloops == 0) {
                        panic("indkey_freemap shit %p %p:%d\n",
                              parent, base, count);
                }
                chain = hammer2_combined_find(parent, base, count,
                                              &cache_index, &key_next,
                                              key_beg, key_end,
                                              &bref);

                /*
                 * Exhausted search
                 */
                if (bref == NULL)
                        break;

                /*
                 * NOTE: No need to check DUPLICATED here because we do
                 *       not release the spinlock.
                 */
                if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) {
                        if (key_next == 0 || key_next > key_end)
                                break;
                        key_beg = key_next;
                        continue;
                }

                /*
                 * Use the full live (not deleted) element for the scan
                 * iteration.  HAMMER2 does not allow partial replacements.
                 *
                 * XXX should be built into hammer2_combined_find().
                 */
                key_next = bref->key + ((hammer2_key_t)1 << bref->keybits);

                /*
                 * Expand our calculated key range (key, keybits) to fit
                 * the scanned key.  nkeybits represents the full range
                 * that we will later cut in half (two halves @ nkeybits - 1).
                 */
                nkeybits = keybits;
                if (nkeybits < bref->keybits) {
                        if (bref->keybits > 64) {
                                kprintf("bad bref chain %p bref %p\n",
                                        chain, bref);
                                Debugger("fubar");
                        }
                        nkeybits = bref->keybits;
                }
                while (nkeybits < 64 &&
                       (~(((hammer2_key_t)1 << nkeybits) - 1) &
                        (key ^ bref->key)) != 0) {
                        ++nkeybits;
                }

                /*
                 * If the new key range is larger we have to determine
                 * which side of the new key range the existing keys fall
                 * under by checking the high bit, then collapsing the
                 * locount into the hicount or vise-versa.
                 */
                if (keybits != nkeybits) {
                        if (((hammer2_key_t)1 << (nkeybits - 1)) & key) {
                                hicount += locount;
                                locount = 0;
                        } else {
                                locount += hicount;
                                hicount = 0;
                        }
                        keybits = nkeybits;
                }

                /*
                 * The newly scanned key will be in the lower half or the
                 * upper half of the (new) key range.
                 */
                if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key)
                        ++hicount;
                else
                        ++locount;

                if (key_next == 0)
                        break;
                key_beg = key_next;
        }
        spin_unlock(&above->cst.spin);
        bref = NULL;    /* now invalid (safety) */

        /*
         * Adjust keybits to represent half of the full range calculated
         * above (radix 63 max)
         */
        --keybits;

        /*
         * Select whichever half contains the most elements.  Theoretically
         * we can select either side as long as it contains at least one
         * element (in order to ensure that a free slot is present to hold
         * the indirect block).
         */
        if (hammer2_indirect_optimize) {
                /*
                 * Insert node for least number of keys, this will arrange
                 * the first few blocks of a large file or the first few
                 * inodes in a directory with fewer indirect blocks when
                 * created linearly.
                 */
                if (hicount < locount && hicount != 0)
                        key |= (hammer2_key_t)1 << keybits;
                else
                        key &= ~(hammer2_key_t)1 << keybits;
        } else {
                /*
                 * Insert node for most number of keys, best for heavily
                 * fragmented files.
                 */
                if (hicount > locount)
                        key |= (hammer2_key_t)1 << keybits;
                else
                        key &= ~(hammer2_key_t)1 << keybits;
        }
        *keyp = key;

        return (keybits);
}

/*
 * Sets CHAIN_DELETED and CHAIN_FLUSH_DELETE in the chain being deleted and
 * set chain->delete_tid.  The chain is not actually marked possibly-free
 * in the freemap until the deletion is completely flushed out (because
 * a flush which doesn't cover the entire deletion is flushing the deleted
 * chain as if it were live).
 *
 * This function does NOT generate a modification to the parent.  It
 * would be nearly impossible to figure out which parent to modify anyway.
 * Such modifications are handled top-down by the flush code and are
 * properly merged using the flush synchronization point.
 *
 * The find/get code will properly overload the RBTREE check on top of
 * the bref check to detect deleted entries.
 *
 * This function is NOT recursive.  Any entity already pushed into the
 * chain (such as an inode) may still need visibility into its contents,
 * as well as the ability to read and modify the contents.  For example,
 * for an unlinked file which is still open.
 *
 * NOTE: This function does NOT set chain->modify_tid, allowing future
 *       code to distinguish between live and deleted chains by testing
 *       trans->sync_tid vs chain->modify_tid and chain->delete_tid.
 *
 * NOTE: Deletions normally do not occur in the middle of a duplication
 *       chain but we use a trick for hardlink migration that refactors
 *       the originating inode without deleting it, so we make no assumptions
 *       here.
 */
void
hammer2_chain_delete(hammer2_trans_t *trans, hammer2_chain_t *chain, int flags)
{
        KKASSERT(ccms_thread_lock_owned(&chain->core->cst));

        /*
         * Nothing to do if already marked.
         */
        if (chain->flags & HAMMER2_CHAIN_DELETED)
                return;

        /*
         * The setting of DELETED causes finds, lookups, and _next iterations
         * to no longer recognize the chain.  RB_SCAN()s will still have
         * visibility (needed for flush serialization points).
         *
         * We need the spinlock on the core whos RBTREE contains chain
         * to protect against races.
         */
        spin_lock(&chain->above->cst.spin);
        _hammer2_chain_delete_helper(trans, chain->above, chain);
        spin_unlock(&chain->above->cst.spin);

        hammer2_chain_setsubmod(trans, chain);
}

/*
 * Returns the index of the nearest element in the blockref array >= elm.
 * Returns (count) if no element could be found.  If delete_filter is non-zero
 * the scan filters out any blockrefs which match deleted chains on dbtree.
 *
 * Sets *key_nextp to the next key for loop purposes but does not modify
 * it if the next key would be higher than the current value of *key_nextp.
 * Note that *key_nexp can overflow to 0, which should be tested by the
 * caller.
 *
 * (*cache_indexp) is a heuristic and can be any value without effecting
 * the result.
 *
 * The spin lock on the related chain must be held.
 */
int
hammer2_base_find(hammer2_chain_t *parent,
                  hammer2_blockref_t *base, int count,
                  int *cache_indexp, hammer2_key_t *key_nextp,
                  hammer2_key_t key_beg, hammer2_key_t key_end,
                  int delete_filter)
{
        hammer2_chain_core_t *core = parent->core;
        hammer2_blockref_t *scan;
        hammer2_key_t scan_end;
        int i;
        int limit;

        /*
         * Require the live chain's already have their core's counted
         * so we can optimize operations.
         */
        KKASSERT((parent->flags & HAMMER2_CHAIN_DUPLICATED) ||
                 core->flags & HAMMER2_CORE_COUNTEDBREFS);

        /*
         * Degenerate case
         */
        if (count == 0 || base == NULL)
                return(count);

        /*
         * Sequential optimization using *cache_indexp.  This is the most
         * likely scenario.
         *
         * We can avoid trailing empty entries on live chains, otherwise
         * we might have to check the whole block array.
         */
        i = *cache_indexp;
        cpu_ccfence();
        if (parent->flags & HAMMER2_CHAIN_DUPLICATED)
                limit = count;
        else
                limit = core->live_zero;
        if (i >= limit)
                i = limit - 1;
        if (i < 0)
                i = 0;
        KKASSERT(i < count);

        /*
         * Search backwards
         */
        scan = &base[i];
        while (i > 0 && (scan->type == 0 || scan->key > key_beg)) {
                --scan;
                --i;
        }
        *cache_indexp = i;

        /*
         * Search forwards, stop when we find a scan element which
         * encloses the key or until we know that there are no further
         * elements.
         */
        while (i < count) {
                if (scan->type != 0) {
                        scan_end = scan->key +
                                   ((hammer2_key_t)1 << scan->keybits) - 1;
                        if (scan->key > key_beg || scan_end >= key_beg) {
                                /*
                                 * Check to see if the entry is covered by
                                 * a deleted chain and ignore the entry if
                                 * it is and delete_filter != 0.
                                 */
                                if (delete_filter == 0)
                                        break;
                                if (hammer2_chain_find_deleted(
                                        parent, scan->key, scan_end) == NULL) {
                                        break;
                                }
                        }
                }
                if (i >= limit)
                        return (count);
                ++scan;
                ++i;
        }
        if (i != count) {
                *cache_indexp = i;
                if (i >= limit) {
                        i = count;
                } else {
                        scan_end = scan->key +
                                   ((hammer2_key_t)1 << scan->keybits);
                        if (scan_end && (*key_nextp > scan_end ||
                                         *key_nextp == 0)) {
                                *key_nextp = scan_end;
                        }
                }
        }
        return (i);
}

/*
 * Do a combined search and return the next match either from the blockref
 * array or from the in-memory chain.  Sets *bresp to the returned bref in
 * both cases, or sets it to NULL if the search exhausted.  Only returns
 * a non-NULL chain if the search matched from the in-memory chain.
 *
 * When no in-memory chain has been found and a non-NULL bref is returned
 * in *bresp.
 *
 * Must be called with above's spinlock held.  Spinlock remains held
 * through the operation.
 *
 * The returned chain is not locked or referenced.  Use the returned bref
 * to determine if the search exhausted or not.  Iterate if the base find
 * is chosen but matches a deleted chain.
 */
static hammer2_chain_t *
hammer2_combined_find(hammer2_chain_t *parent,
                      hammer2_blockref_t *base, int count,
                      int *cache_indexp, hammer2_key_t *key_nextp,
                      hammer2_key_t key_beg, hammer2_key_t key_end,
                      hammer2_blockref_t **bresp)
{
        hammer2_blockref_t *bref;
        hammer2_chain_t *chain;
        int i;

        /*
         * Lookup in block array and in rbtree.
         */
        *key_nextp = key_end + 1;
        i = hammer2_base_find(parent, base, count, cache_indexp,
                              key_nextp, key_beg, key_end, 1);
        chain = hammer2_chain_find(parent, key_nextp, key_beg, key_end);

        /*
         * Neither matched
         */
        if (i == count && chain == NULL) {
                *bresp = NULL;
                return(NULL);
        }

        /*
         * Only chain matched.
         */
        if (i == count) {
                bref = &chain->bref;
                goto found;
        }

        /*
         * Only blockref matched.
         */
        if (chain == NULL) {
                bref = &base[i];
                goto found;
        }

        /*
         * Both in-memory and blockref matched, select the nearer element.
         *
         * If both are flush with the left-hand side or both are the
         * same distance away, select the chain.  In this situation the
         * chain must have been loaded from the matching blockmap.
         */
        if ((chain->bref.key <= key_beg && base[i].key <= key_beg) ||
            chain->bref.key == base[i].key) {
                KKASSERT(chain->bref.key == base[i].key);
                if ((chain->flags & HAMMER2_CHAIN_BMAPPED) == 0) {
                        kprintf("chain not bmapped %p.%d %08x\n", chain, chain->bref.type, chain->flags);
                        kprintf("in chain mod/del %016jx %016jx\n", chain->modify_tid, chain->delete_tid);
                        kprintf("and updlo/hi %016jx %016jx\n", chain->update_lo, chain->update_hi);
                }
                KKASSERT(chain->flags & HAMMER2_CHAIN_BMAPPED);
                bref = &chain->bref;
                goto found;
        }

        /*
         * Select the nearer key
         */
        if (chain->bref.key < base[i].key) {
                bref = &chain->bref;
        } else {
                bref = &base[i];
                chain = NULL;
        }

        /*
         * If the bref is out of bounds we've exhausted our search.
         */
found:
        if (bref->key > key_end) {
                *bresp = NULL;
                chain = NULL;
        } else {
                *bresp = bref;
        }
        return(chain);
}

/*
 * Locate the specified block array element and delete it.  The element
 * must exist.
 *
 * The spin lock on the related chain must be held.
 *
 * NOTE: live_count was adjusted when the chain was deleted, so it does not
 *       need to be adjusted when we commit the media change.
 */
void
hammer2_base_delete(hammer2_trans_t *trans, hammer2_chain_t *parent,
                    hammer2_blockref_t *base, int count,
                    int *cache_indexp, hammer2_chain_t *child)
{
        hammer2_blockref_t *elm = &child->bref;
        hammer2_chain_core_t *core = parent->core;
        hammer2_key_t key_next;
        int i;

        /*
         * Delete element.  Expect the element to exist.
         *
         * XXX see caller, flush code not yet sophisticated enough to prevent
         *     re-flushed in some cases.
         */
        key_next = 0; /* max range */
        i = hammer2_base_find(parent, base, count, cache_indexp,
                              &key_next, elm->key, elm->key, 0);
        if (i == count || base[i].type == 0 ||
            base[i].key != elm->key || base[i].keybits != elm->keybits) {
                spin_unlock(&core->cst.spin);
                panic("delete base %p element not found at %d/%d elm %p\n"
                      "child ino_reason=%08x\n",
                      base, i, count, elm,
                      child->inode_reason);
                return;
        }
        bzero(&base[i], sizeof(*base));
        base[i].mirror_tid = (intptr_t)parent;          /* MEDIA DEBUG */
        base[i].modify_tid = (intptr_t)child;           /* MEDIA DEBUG */
        base[i].check.debug.sync_tid = trans->sync_tid; /* MEDIA DEBUG */
        ++parent->nremoves;                             /* DEBUG */

        /*
         * We can only optimize core->live_zero for live chains.
         */
        if ((parent->flags & HAMMER2_CHAIN_DUPLICATED) == 0) {
                if (core->live_zero == i + 1) {
                        while (--i >= 0 && base[i].type == 0)
                                ;
                        core->live_zero = i + 1;
                }
        }
}

/*
 * Insert the specified element.  The block array must not already have the
 * element and must have space available for the insertion.
 *
 * The spin lock on the related chain must be held.
 *
 * NOTE: live_count was adjusted when the chain was deleted, so it does not
 *       need to be adjusted when we commit the media change.
 */
void
hammer2_base_insert(hammer2_trans_t *trans __unused, hammer2_chain_t *parent,
                    hammer2_blockref_t *base, int count,
                    int *cache_indexp, hammer2_chain_t *child)
{
        hammer2_blockref_t *elm = &child->bref;
        hammer2_chain_core_t *core = parent->core;
        hammer2_key_t key_next;
        hammer2_key_t xkey;
        int i;
        int j;
        int k;
        int l;
        int u = 1;

        /*
         * Insert new element.  Expect the element to not already exist
         * unless we are replacing it.
         *
         * XXX see caller, flush code not yet sophisticated enough to prevent
         *     re-flushed in some cases.
         */
        key_next = 0; /* max range */
        i = hammer2_base_find(parent, base, count, cache_indexp,
                              &key_next, elm->key, elm->key, 0);

        /*
         * Shortcut fill optimization, typical ordered insertion(s) may not
         * require a search.
         */
        KKASSERT(i >= 0 && i <= count);

        /*
         * We can only optimize core->live_zero for live chains.
         */
        if (i == count && core->live_zero < count) {
                if ((parent->flags & HAMMER2_CHAIN_DUPLICATED) == 0) {
                        i = core->live_zero++;
                        base[i] = *elm;
        ++parent->ninserts;     /* DEBUG */
                        return;
                }
        }

        xkey = elm->key + ((hammer2_key_t)1 << elm->keybits) - 1;
        if (i != count && (base[i].key < elm->key || xkey >= base[i].key)) {
                if (child->flags & HAMMER2_CHAIN_FLUSH_TEMPORARY) {
                        kprintf("child %p special replace\n", child);
                        base[i] = *elm;
                        return;
                } else {
                        spin_unlock(&core->cst.spin);
                        panic("insert base %p overlapping "
                              "elements at %d elm %p\n",
                              base, i, elm);
                }
        }

        /*
         * Try to find an empty slot before or after.
         */
        j = i;
        k = i;
        while (j > 0 || k < count) {
                --j;
                if (j >= 0 && base[j].type == 0) {
                        if (j == i - 1) {
                                base[j] = *elm;
                        } else {
                                bcopy(&base[j+1], &base[j],
                                      (i - j - 1) * sizeof(*base));
                                base[i - 1] = *elm;
                        }
        ++parent->ninserts;     /* DEBUG */
                        goto validate;
                }
                ++k;
                if (k < count && base[k].type == 0) {
                        bcopy(&base[i], &base[i+1],
                              (k - i) * sizeof(hammer2_blockref_t));
                        base[i] = *elm;

                        /*
                         * We can only update core->live_zero for live
                         * chains.
                         */
                        if ((parent->flags & HAMMER2_CHAIN_DUPLICATED) == 0) {
                                if (core->live_zero <= k)
                                        core->live_zero = k + 1;
                        }
                        u = 2;
        ++parent->ninserts;     /* DEBUG */
                        goto validate;
                }
        }
        panic("hammer2_base_insert: no room!");

        /*
         * Debugging
         */
validate:
        key_next = 0;
        for (l = 0; l < count; ++l) {
                if (base[l].type) {
                        key_next = base[l].key +
                                   ((hammer2_key_t)1 << base[l].keybits) - 1;
                        break;
                }
        }
        while (++l < count) {
                if (base[l].type) {
                        if (base[l].key <= key_next)
                                panic("base_insert %d %d,%d,%d fail %p:%d", u, i, j, k, base, l);
                        key_next = base[l].key +
                                   ((hammer2_key_t)1 << base[l].keybits) - 1;

                }
        }

}

#if 0

/*
 * Sort the blockref array for the chain.  Used by the flush code to
 * sort the blockref[] array.
 *
 * The chain must be exclusively locked AND spin-locked.
 */
typedef hammer2_blockref_t *hammer2_blockref_p;

static
int
hammer2_base_sort_callback(const void *v1, const void *v2)
{
        hammer2_blockref_p bref1 = *(const hammer2_blockref_p *)v1;
        hammer2_blockref_p bref2 = *(const hammer2_blockref_p *)v2;

        /*
         * Make sure empty elements are placed at the end of the array
         */
        if (bref1->type == 0) {
                if (bref2->type == 0)
                        return(0);
                return(1);
        } else if (bref2->type == 0) {
                return(-1);
        }

        /*
         * Sort by key
         */
        if (bref1->key < bref2->key)
                return(-1);
        if (bref1->key > bref2->key)
                return(1);
        return(0);
}

void
hammer2_base_sort(hammer2_chain_t *chain)
{
        hammer2_blockref_t *base;
        int count;

        switch(chain->bref.type) {
        case HAMMER2_BREF_TYPE_INODE:
                /*
                 * Special shortcut for embedded data returns the inode
                 * itself.  Callers must detect this condition and access
                 * the embedded data (the strategy code does this for us).
                 *
                 * This is only applicable to regular files and softlinks.
                 */
                if (chain->data->ipdata.op_flags & HAMMER2_OPFLAG_DIRECTDATA)
                        return;
                base = &chain->data->ipdata.u.blockset.blockref[0];
                count = HAMMER2_SET_COUNT;
                break;
        case HAMMER2_BREF_TYPE_FREEMAP_NODE:
        case HAMMER2_BREF_TYPE_INDIRECT:
                /*
                 * Optimize indirect blocks in the INITIAL state to avoid
                 * I/O.
                 */
                KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) == 0);
                base = &chain->data->npdata[0];
                count = chain->bytes / sizeof(hammer2_blockref_t);
                break;
        case HAMMER2_BREF_TYPE_VOLUME:
                base = &chain->hmp->voldata.sroot_blockset.blockref[0];
                count = HAMMER2_SET_COUNT;
                break;
        case HAMMER2_BREF_TYPE_FREEMAP:
                base = &chain->hmp->voldata.freemap_blockset.blockref[0];
                count = HAMMER2_SET_COUNT;
                break;
        default:
                panic("hammer2_chain_lookup: unrecognized blockref type: %d",
                      chain->bref.type);
                base = NULL;    /* safety */
                count = 0;      /* safety */
        }
        kqsort(base, count, sizeof(*base), hammer2_base_sort_callback);
}

#endif

/*
 * Chain memory management
 */
void
hammer2_chain_wait(hammer2_chain_t *chain)
{
        tsleep(chain, 0, "chnflw", 1);
}

/*
 * Manage excessive memory resource use for chain and related
 * structures.
 */
void
hammer2_chain_memory_wait(hammer2_pfsmount_t *pmp)
{
        long waiting;
        long count;
        long limit;
#if 0
        static int zzticks;
#endif

        /*
         * Atomic check condition and wait.  Also do an early speedup of
         * the syncer to try to avoid hitting the wait.
         */
        for (;;) {
                waiting = pmp->inmem_dirty_chains;
                cpu_ccfence();
                count = waiting & HAMMER2_DIRTYCHAIN_MASK;

                limit = pmp->mp->mnt_nvnodelistsize / 10;
                if (limit < hammer2_limit_dirty_chains)
                        limit = hammer2_limit_dirty_chains;
                if (limit < 1000)
                        limit = 1000;

#if 0
                if ((int)(ticks - zzticks) > hz) {
                        zzticks = ticks;
                        kprintf("count %ld %ld\n", count, limit);
                }
#endif

                /*
                 * Block if there are too many dirty chains present, wait
                 * for the flush to clean some out.
                 */
                if (count > limit) {
                        tsleep_interlock(&pmp->inmem_dirty_chains, 0);
                        if (atomic_cmpset_long(&pmp->inmem_dirty_chains,
                                               waiting,
                                       waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
                                speedup_syncer(pmp->mp);
                                tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
                                       "chnmem", hz);
                        }
                        continue;       /* loop on success or fail */
                }

                /*
                 * Try to start an early flush before we are forced to block.
                 */
                if (count > limit * 7 / 10)
                        speedup_syncer(pmp->mp);
                break;
        }
}

void
hammer2_chain_memory_inc(hammer2_pfsmount_t *pmp)
{
        if (pmp)
                atomic_add_long(&pmp->inmem_dirty_chains, 1);
}

void
hammer2_chain_memory_wakeup(hammer2_pfsmount_t *pmp)
{
        long waiting;

        if (pmp == NULL)
                return;

        for (;;) {
                waiting = pmp->inmem_dirty_chains;
                cpu_ccfence();
                if (atomic_cmpset_long(&pmp->inmem_dirty_chains,
                                       waiting,
                                       (waiting - 1) &
                                        ~HAMMER2_DIRTYCHAIN_WAITING)) {
                        break;
                }
        }

        if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
                wakeup(&pmp->inmem_dirty_chains);
}

static
void
adjreadcounter(hammer2_blockref_t *bref, size_t bytes)
{
        long *counterp;

        switch(bref->type) {
        case HAMMER2_BREF_TYPE_DATA:
                counterp = &hammer2_iod_file_read;
                break;
        case HAMMER2_BREF_TYPE_INODE:
                counterp = &hammer2_iod_meta_read;
                break;
        case HAMMER2_BREF_TYPE_INDIRECT:
                counterp = &hammer2_iod_indr_read;
                break;
        case HAMMER2_BREF_TYPE_FREEMAP_NODE:
        case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
                counterp = &hammer2_iod_fmap_read;
                break;
        default:
                counterp = &hammer2_iod_volu_read;
                break;
        }
        *counterp += bytes;
}