xref: /dragonfly/sys/vfs/hammer2/hammer2_flush.c (revision 3ea159d2)

/*
 * Copyright (c) 2011-2018 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.
 */
/*
 *			TRANSACTION AND FLUSH HANDLING
 *
 * Deceptively simple but actually fairly difficult to implement properly is
 * how I would describe it.
 *
 * Flushing generally occurs bottom-up but requires a top-down scan to
 * locate chains with MODIFIED and/or UPDATE bits set.  The ONFLUSH flag
 * tells how to recurse downward to find these chains.
 */

#include <sys/cdefs.h>
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/types.h>
#include <sys/lock.h>
#include <sys/uuid.h>

#include "hammer2.h"

#define FLUSH_DEBUG 0

#define HAMMER2_FLUSH_DEPTH_LIMIT	60      /* stack recursion limit */


/*
 * Recursively flush the specified chain.  The chain is locked and
 * referenced by the caller and will remain so on return.  The chain
 * will remain referenced throughout but can temporarily lose its
 * lock during the recursion to avoid unnecessarily stalling user
 * processes.
 */
struct hammer2_flush_info {
	hammer2_chain_t *parent;
	int		depth;
	int		error;			/* cumulative error */
	int		flags;
#ifdef HAMMER2_SCAN_DEBUG
	long		scan_count;
	long		scan_mod_count;
	long		scan_upd_count;
	long		scan_onf_count;
	long		scan_del_count;
	long		scan_btype[7];
#endif
	hammer2_chain_t	*debug;
};

typedef struct hammer2_flush_info hammer2_flush_info_t;

static int hammer2_flush_core(hammer2_flush_info_t *info,
				hammer2_chain_t *chain, int flags);
static int hammer2_flush_recurse(hammer2_chain_t *child, void *data);

/*
 * Any per-pfs transaction initialization goes here.
 */
void
hammer2_trans_manage_init(hammer2_pfs_t *pmp)
{
}

/*
 * Transaction support for any modifying operation.  Transactions are used
 * in the pmp layer by the frontend and in the spmp layer by the backend.
 *
 * 0			- Normal transaction.  Interlocks against just the
 *			  COPYQ portion of an ISFLUSH transaction.
 *
 * TRANS_ISFLUSH	- Flush transaction.  Interlocks against other flush
 *			  transactions.
 *
 *			  When COPYQ is also specified, waits for the count
 *			  to drop to 1.
 *
 * TRANS_BUFCACHE	- Buffer cache transaction.  No interlock.
 *
 * TRANS_SIDEQ		- Run the sideq (only tested in trans_done())
 *
 * Initializing a new transaction allocates a transaction ID.  Typically
 * passed a pmp (hmp passed as NULL), indicating a cluster transaction.  Can
 * be passed a NULL pmp and non-NULL hmp to indicate a transaction on a single
 * media target.  The latter mode is used by the recovery code.
 */
void
hammer2_trans_init(hammer2_pfs_t *pmp, uint32_t flags)
{
	uint32_t oflags;
	uint32_t nflags;
	int dowait;

	for (;;) {
		oflags = pmp->trans.flags;
		cpu_ccfence();
		dowait = 0;

		if (flags & HAMMER2_TRANS_ISFLUSH) {
			/*
			 * Interlock against other flush transactions.
			 */
			if (oflags & HAMMER2_TRANS_ISFLUSH) {
				nflags = oflags | HAMMER2_TRANS_WAITING;
				dowait = 1;
			} else {
				nflags = (oflags | flags) + 1;
			}
		} else if (flags & HAMMER2_TRANS_BUFCACHE) {
			/*
			 * Requesting strategy transaction from buffer-cache,
			 * or a VM getpages/putpages through the buffer cache.
			 * We must allow such transactions in all situations
			 * to avoid deadlocks.
			 */
			nflags = (oflags | flags) + 1;
		} else {
			/*
			 * Normal transaction.  We do not interlock against
			 * BUFCACHE or ISFLUSH.
			 *
			 * Note that vnode locks may be held going into
			 * this call.
			 *
			 * NOTE: Remember that non-modifying operations
			 *	 such as read, stat, readdir, etc, do
			 *	 not use transactions.
			 */
			nflags = (oflags | flags) + 1;
		}
		if (dowait)
			tsleep_interlock(&pmp->trans.sync_wait, 0);
		if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
			if (dowait == 0)
				break;
			tsleep(&pmp->trans.sync_wait, PINTERLOCKED,
			       "h2trans", hz);
			/* retry */
		} else {
			cpu_pause();
			/* retry */
		}
		/* retry */
	}

#if 0
	/*
	 * When entering a FLUSH transaction with COPYQ set, wait for the
	 * transaction count to drop to 1 (our flush transaction only)
	 * before proceeding.
	 *
	 * This waits for all non-flush transactions to complete and blocks
	 * new non-flush transactions from starting until COPYQ is cleared.
	 * (the flush will then proceed after clearing COPYQ).  This should
	 * be a very short stall on modifying operations.
	 */
	while ((flags & HAMMER2_TRANS_ISFLUSH) &&
	       (flags & HAMMER2_TRANS_COPYQ)) {
		oflags = pmp->trans.flags;
		cpu_ccfence();
		if ((oflags & HAMMER2_TRANS_MASK) == 1)
			break;
		nflags = oflags | HAMMER2_TRANS_WAITING;
		tsleep_interlock(&pmp->trans.sync_wait, 0);
		if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
			tsleep(&pmp->trans.sync_wait, PINTERLOCKED,
			       "h2trans2", hz);
		}
	}
#endif
}

/*
 * Start a sub-transaction, there is no 'subdone' function.  This will
 * issue a new modify_tid (mtid) for the current transaction, which is a
 * CLC (cluster level change) id and not a per-node id.
 *
 * This function must be called for each XOP when multiple XOPs are run in
 * sequence within a transaction.
 *
 * Callers typically update the inode with the transaction mtid manually
 * to enforce sequencing.
 */
hammer2_tid_t
hammer2_trans_sub(hammer2_pfs_t *pmp)
{
	hammer2_tid_t mtid;

	mtid = atomic_fetchadd_64(&pmp->modify_tid, 1);

	return (mtid);
}

void
hammer2_trans_setflags(hammer2_pfs_t *pmp, uint32_t flags)
{
	atomic_set_int(&pmp->trans.flags, flags);
}

/*
 * Typically used to clear trans flags asynchronously.  If TRANS_WAITING
 * is in the mask, and was previously set, this function will wake up
 * any waiters.
 */
void
hammer2_trans_clearflags(hammer2_pfs_t *pmp, uint32_t flags)
{
	uint32_t oflags;
	uint32_t nflags;

	for (;;) {
		oflags = pmp->trans.flags;
		cpu_ccfence();
		nflags = oflags & ~flags;
		if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
			if ((oflags ^ nflags) & HAMMER2_TRANS_WAITING)
				wakeup(&pmp->trans.sync_wait);
			break;
		}
		cpu_pause();
		/* retry */
	}
}

void
hammer2_trans_done(hammer2_pfs_t *pmp, uint32_t flags)
{
	uint32_t oflags;
	uint32_t nflags;

#if 0
	/*
	 * Modifying ops on the front-end can cause dirty inodes to
	 * build up in the sideq.  We don't flush these on inactive/reclaim
	 * due to potential deadlocks, so we have to deal with them from
	 * inside other nominal modifying front-end transactions.
	 */
	if ((flags & HAMMER2_TRANS_SIDEQ) &&
	    pmp->sideq_count > hammer2_limit_dirty_inodes / 2 &&
	    pmp->sideq_count > (pmp->inum_count >> 3) &&
	    pmp->mp) {
		speedup_syncer(pmp->mp);
	}
#endif

	/*
	 * Clean-up the transaction.  Wakeup any waiters when finishing
	 * a flush transaction or transitioning the non-flush transaction
	 * count from 2->1 while a flush transaction is pending.
	 */
	for (;;) {
		oflags = pmp->trans.flags;
		cpu_ccfence();
		KKASSERT(oflags & HAMMER2_TRANS_MASK);

		nflags = (oflags - 1) & ~flags;
		if (flags & HAMMER2_TRANS_ISFLUSH) {
			nflags &= ~HAMMER2_TRANS_WAITING;
		}
		if ((oflags & (HAMMER2_TRANS_ISFLUSH|HAMMER2_TRANS_MASK)) ==
		    (HAMMER2_TRANS_ISFLUSH|2)) {
			nflags &= ~HAMMER2_TRANS_WAITING;
		}
		if (atomic_cmpset_int(&pmp->trans.flags, oflags, nflags)) {
			if ((oflags ^ nflags) & HAMMER2_TRANS_WAITING)
				wakeup(&pmp->trans.sync_wait);
			break;
		}
		cpu_pause();
		/* retry */
	}
}

/*
 * Obtain new, unique inode number (not serialized by caller).
 */
hammer2_tid_t
hammer2_trans_newinum(hammer2_pfs_t *pmp)
{
	hammer2_tid_t tid;

	tid = atomic_fetchadd_64(&pmp->inode_tid, 1);

	return tid;
}

/*
 * Assert that a strategy call is ok here.  Currently we allow strategy
 * calls in all situations, including during flushes.  Previously:
 *	(old) (1) In a normal transaction.
 *	(old) (2) In a flush transaction only if PREFLUSH is also set.
 */
void
hammer2_trans_assert_strategy(hammer2_pfs_t *pmp)
{
#if 0
	KKASSERT((pmp->trans.flags & HAMMER2_TRANS_ISFLUSH) == 0 ||
		 (pmp->trans.flags & HAMMER2_TRANS_PREFLUSH));
#endif
}

/*
 * Flush the chain and all modified sub-chains through the specified
 * synchronization point, propagating blockref updates back up.  As
 * part of this propagation, mirror_tid and inode/data usage statistics
 * propagates back upward.
 *
 * Returns a HAMMER2 error code, 0 if no error.  Note that I/O errors from
 * buffers dirtied during the flush operation can occur later.
 *
 * modify_tid (clc - cluster level change) is not propagated.
 *
 * update_tid (clc) is used for validation and is not propagated by this
 * function.
 *
 * This routine can be called from several places but the most important
 * is from VFS_SYNC (frontend) via hammer2_xop_inode_flush (backend).
 *
 * chain is locked on call and will remain locked on return.  The chain's
 * UPDATE flag indicates that its parent's block table (which is not yet
 * part of the flush) should be updated.
 *
 * flags:
 *	HAMMER2_FLUSH_TOP	Indicates that this is the top of the flush.
 *				Is cleared for the recursion.
 *
 *	HAMMER2_FLUSH_ALL	Recurse everything
 *
 *	HAMMER2_FLUSH_INODE_STOP
 *				Stop at PFS inode or normal inode boundary
 */
int
hammer2_flush(hammer2_chain_t *chain, int flags)
{
	hammer2_flush_info_t info;
	hammer2_dev_t *hmp;
	int loops;

	/*
	 * Execute the recursive flush and handle deferrals.
	 *
	 * Chains can be ridiculously long (thousands deep), so to
	 * avoid blowing out the kernel stack the recursive flush has a
	 * depth limit.  Elements at the limit are placed on a list
	 * for re-execution after the stack has been popped.
	 */
	bzero(&info, sizeof(info));
	info.flags = flags & ~HAMMER2_FLUSH_TOP;

	/*
	 * Calculate parent (can be NULL), if not NULL the flush core
	 * expects the parent to be referenced so it can easily lock/unlock
	 * it without it getting ripped up.
	 */
	if ((info.parent = chain->parent) != NULL)
		hammer2_chain_ref(info.parent);

	/*
	 * Extra ref needed because flush_core expects it when replacing
	 * chain.
	 */
	hammer2_chain_ref(chain);
	hmp = chain->hmp;
	loops = 0;

	for (;;) {
		/*
		 * [re]flush chain as the deep recursion may have generated
		 * additional modifications.
		 */
		if (info.parent != chain->parent) {
			if (hammer2_debug & 0x0040) {
				kprintf("LOST CHILD4 %p->%p "
					"(actual parent %p)\n",
					info.parent, chain, chain->parent);
			}
			hammer2_chain_drop(info.parent);
			info.parent = chain->parent;
			hammer2_chain_ref(info.parent);
		}
		if (hammer2_flush_core(&info, chain, flags) == 0)
			break;

		if (++loops % 1000 == 0) {
			kprintf("hammer2_flush: excessive loops on %p\n",
				chain);
			if (hammer2_debug & 0x100000)
				Debugger("hell4");
		}
	}
#ifdef HAMMER2_SCAN_DEBUG
	if (info.scan_count >= 10)
	kprintf("hammer2_flush: scan_count %ld (%ld,%ld,%ld,%ld) "
		"bt(%ld,%ld,%ld,%ld,%ld,%ld)\n",
		info.scan_count,
		info.scan_mod_count,
		info.scan_upd_count,
		info.scan_onf_count,
		info.scan_del_count,
		info.scan_btype[1],
		info.scan_btype[2],
		info.scan_btype[3],
		info.scan_btype[4],
		info.scan_btype[5],
		info.scan_btype[6]);
#endif
	hammer2_chain_drop(chain);
	if (info.parent)
		hammer2_chain_drop(info.parent);
	return (info.error);
}

/*
 * This is the core of the chain flushing code.  The chain is locked by the
 * caller and must also have an extra ref on it by the caller, and remains
 * locked and will have an extra ref on return.  info.parent is referenced
 * but not locked.
 *
 * Upon return, the caller can test the UPDATE bit on the chain to determine
 * if the parent needs updating.
 *
 * If non-zero is returned, the chain's parent changed during the flush and
 * the caller must retry the operation.
 *
 * (1) Determine if this node is a candidate for the flush, return if it is
 *     not.  fchain and vchain are always candidates for the flush.
 *
 * (2) If we recurse too deep the chain is entered onto the deferral list and
 *     the current flush stack is aborted until after the deferral list is
 *     run.
 *
 * (3) Recursively flush live children (rbtree).  This can create deferrals.
 *     A successful flush clears the MODIFIED and UPDATE bits on the children
 *     and typically causes the parent to be marked MODIFIED as the children
 *     update the parent's block table.  A parent might already be marked
 *     MODIFIED due to a deletion (whos blocktable update in the parent is
 *     handled by the frontend), or if the parent itself is modified by the
 *     frontend for other reasons.
 *
 * (4) Permanently disconnected sub-trees are cleaned up by the front-end.
 *     Deleted-but-open inodes can still be individually flushed via the
 *     filesystem syncer.
 *
 * (5) Delete parents on the way back up if they are normal indirect blocks
 *     and have no children.
 *
 * (6) Note that an unmodified child may still need the block table in its
 *     parent updated (e.g. rename/move).  The child will have UPDATE set
 *     in this case.
 *
 *			WARNING ON BREF MODIFY_TID/MIRROR_TID
 *
 * blockref.modify_tid is consistent only within a PFS, and will not be
 * consistent during synchronization.  mirror_tid is consistent across the
 * block device regardless of the PFS.
 */
static int
hammer2_flush_core(hammer2_flush_info_t *info, hammer2_chain_t *chain,
		   int flags)
{
	hammer2_chain_t *parent;
	hammer2_dev_t *hmp;
	int save_error;
	int retry;

	retry = 0;

	/*
	 * (1) Optimize downward recursion to locate nodes needing action.
	 *     Nothing to do if none of these flags are set.
	 */
	if ((chain->flags & HAMMER2_CHAIN_FLUSH_MASK) == 0) {
		if (hammer2_debug & 0x200) {
			if (info->debug == NULL)
				info->debug = chain;
		} else {
			return 0;
		}
	}

	hmp = chain->hmp;

	/*
	 * NOTE: parent can be NULL, usually due to destroy races.
	 */
	parent = info->parent;
	KKASSERT(chain->parent == parent);

	/*
	 * Downward search recursion
	 *
	 * We must be careful on cold stops, which often occur on inode
	 * boundaries due to the way hammer2_vfs_sync() sequences the flush.
	 * Be sure to issue an appropriate chain_setflush()
	 */
	if ((chain->flags & HAMMER2_CHAIN_PFSBOUNDARY) &&
	    (flags & HAMMER2_FLUSH_ALL) == 0 &&
	    (flags & HAMMER2_FLUSH_TOP) == 0 &&
	    chain->pmp && chain->pmp->mp) {
		/*
		 * If FLUSH_ALL is not specified the caller does not want
		 * to recurse through PFS roots that have been mounted.
		 *
		 * (If the PFS has not been mounted there may not be
		 *  anything monitoring its chains and its up to us
		 *  to flush it).
		 *
		 * The typical sequence is to flush dirty PFS's starting at
		 * their root downward, then flush the device root (vchain).
		 * It is this second flush that typically leaves out the
		 * ALL flag.
		 *
		 * However we must still process the PFSROOT chains for block
		 * table updates in their parent (which IS part of our flush).
		 *
		 * NOTE: The volume root, vchain, does not set PFSBOUNDARY.
		 *
		 * NOTE: We must re-set ONFLUSH in the parent to retain if
		 *	 this chain (that we are skipping) requires work.
		 */
		if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
				    HAMMER2_CHAIN_DESTROY |
				    HAMMER2_CHAIN_MODIFIED)) {
			hammer2_chain_setflush(parent);
		}
		goto done;
	} else if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
		   (flags & HAMMER2_FLUSH_INODE_STOP) &&
		   (flags & HAMMER2_FLUSH_ALL) == 0 &&
		   (flags & HAMMER2_FLUSH_TOP) == 0 &&
		   chain->pmp && chain->pmp->mp) {
		/*
		 * When FLUSH_INODE_STOP is specified we are being asked not
		 * to include any inode changes for inodes we encounter,
		 * with the exception of the inode that the flush began with.
		 * So: INODE, INODE_STOP, and TOP==0 basically.
		 *
		 * Dirty inodes are flushed based on the hammer2_inode
		 * in-memory structure, issuing a chain_setflush() here
		 * will only cause unnecessary traversals of the topology.
		 */
		goto done;
#if 0
		/*
		 * If FLUSH_INODE_STOP is specified and both ALL and TOP
		 * are clear, we must not flush the chain.  The chain should
		 * have already been flushed and any further ONFLUSH/UPDATE
		 * setting will be related to the next flush.
		 *
		 * This features allows us to flush inodes independently of
		 * each other and meta-data above the inodes separately.
		 */
		if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
				    HAMMER2_CHAIN_DESTROY |
				    HAMMER2_CHAIN_MODIFIED)) {
			if (parent)
				hammer2_chain_setflush(parent);
		}
#endif
	} else if (info->depth == HAMMER2_FLUSH_DEPTH_LIMIT) {
		/*
		 * Recursion depth reached.
		 */
		panic("hammer2: flush depth limit");
	} else if (chain->flags & (HAMMER2_CHAIN_ONFLUSH |
				   HAMMER2_CHAIN_DESTROY)) {
		/*
		 * Downward recursion search (actual flush occurs bottom-up).
		 * pre-clear ONFLUSH.  It can get set again due to races or
		 * flush errors, which we want so the scan finds us again in
		 * the next flush.
		 *
		 * We must also recurse if DESTROY is set so we can finally
		 * get rid of the related children, otherwise the node will
		 * just get re-flushed on lastdrop.
		 *
		 * WARNING!  The recursion will unlock/relock info->parent
		 *	     (which is 'chain'), potentially allowing it
		 *	     to be ripped up.
		 */
		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
		save_error = info->error;
		info->error = 0;
		info->parent = chain;

		/*
		 * We may have to do this twice to catch any indirect
		 * block maintenance that occurs.
		 */
		hammer2_spin_ex(&chain->core.spin);
		RB_SCAN(hammer2_chain_tree, &chain->core.rbtree,
			NULL, hammer2_flush_recurse, info);
		if (chain->flags & HAMMER2_CHAIN_ONFLUSH) {
			atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH);
			RB_SCAN(hammer2_chain_tree, &chain->core.rbtree,
				NULL, hammer2_flush_recurse, info);
		}
		hammer2_spin_unex(&chain->core.spin);
		info->parent = parent;

		/*
		 * Re-set the flush bits if the flush was incomplete or
		 * an error occurred.  If an error occurs it is typically
		 * an allocation error.  Errors do not cause deferrals.
		 */
		if (info->error)
			hammer2_chain_setflush(chain);
		info->error |= save_error;

		/*
		 * If we lost the parent->chain association we have to
		 * stop processing this chain because it is no longer
		 * in this recursion.  If it moved, it will be handled
		 * by the ONFLUSH flag elsewhere.
		 */
		if (chain->parent != parent) {
			kprintf("LOST CHILD2 %p->%p (actual parent %p)\n",
				parent, chain, chain->parent);
			goto done;
		}
	}

	/*
	 * Now we are in the bottom-up part of the recursion.
	 *
	 * We continue to try to update the chain on lower-level errors, but
	 * the flush code may decide not to flush the volume root.
	 *
	 * XXX should we continue to try to update the chain if an error
	 *     occurred?
	 */

	/*
	 * Both parent and chain must be locked in order to flush chain,
	 * in order to properly update the parent under certain conditions.
	 *
	 * In addition, we can't safely unlock/relock the chain once we
	 * start flushing the chain itself, which we would have to do later
	 * on in order to lock the parent if we didn't do that now.
	 */
	hammer2_chain_ref_hold(chain);
	hammer2_chain_unlock(chain);
	if (parent)
		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
	hammer2_chain_lock(chain, HAMMER2_RESOLVE_MAYBE);
	hammer2_chain_drop_unhold(chain);

	/*
	 * Can't process if we can't access their content.
	 */
	if ((parent && parent->error) || chain->error) {
		kprintf("hammer2: chain error during flush\n");
		info->error |= chain->error;
		if (parent) {
			info->error |= parent->error;
			hammer2_chain_unlock(parent);
		}
		goto done;
	}

	if (chain->parent != parent) {
		if (hammer2_debug & 0x0040) {
			kprintf("LOST CHILD3 %p->%p (actual parent %p)\n",
				parent, chain, chain->parent);
		}
		KKASSERT(parent != NULL);
		hammer2_chain_unlock(parent);
		retry = 1;
		goto done;
	}

	/*
	 * Propagate the DESTROY flag downwards.  This dummies up the flush
	 * code and tries to invalidate related buffer cache buffers to
	 * avoid the disk write.
	 */
	if (parent && (parent->flags & HAMMER2_CHAIN_DESTROY))
		atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY);

	/*
	 * Dispose of the modified bit.
	 *
	 * If parent is present, the UPDATE bit should already be set.
	 * UPDATE should already be set.
	 * bref.mirror_tid should already be set.
	 */
	if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
		KKASSERT((chain->flags & HAMMER2_CHAIN_UPDATE) ||
			 chain->parent == NULL);
		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
		atomic_add_long(&hammer2_count_modified_chains, -1);

		/*
		 * Manage threads waiting for excessive dirty memory to
		 * be retired.
		 */
		if (chain->pmp)
			hammer2_pfs_memory_wakeup(chain->pmp, -1);

#if 0
		if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0 &&
		    chain != &hmp->vchain &&
		    chain != &hmp->fchain) {
			/*
			 * Set UPDATE bit indicating that the parent block
			 * table requires updating.
			 */
			atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
		}
#endif

		/*
		 * Issue the flush.  This is indirect via the DIO.
		 *
		 * NOTE: A DELETED node that reaches this point must be
		 *	 flushed for synchronization point consistency.
		 *
		 * NOTE: Even though MODIFIED was already set, the related DIO
		 *	 might not be dirty due to a system buffer cache
		 *	 flush and must be set dirty if we are going to make
		 *	 further modifications to the buffer.  Chains with
		 *	 embedded data don't need this.
		 */
		if (hammer2_debug & 0x1000) {
			kprintf("Flush %p.%d %016jx/%d data=%016jx\n",
				chain, chain->bref.type,
				(uintmax_t)chain->bref.key,
				chain->bref.keybits,
				(uintmax_t)chain->bref.data_off);
		}

		/*
		 * Update chain CRCs for flush.
		 *
		 * NOTE: Volume headers are NOT flushed here as they require
		 *	 special processing.
		 */
		switch(chain->bref.type) {
		case HAMMER2_BREF_TYPE_FREEMAP:
			/*
			 * Update the volume header's freemap_tid to the
			 * freemap's flushing mirror_tid.
			 *
			 * (note: embedded data, do not call setdirty)
			 */
			KKASSERT(hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED);
			KKASSERT(chain == &hmp->fchain);
			hmp->voldata.freemap_tid = chain->bref.mirror_tid;
			if (hammer2_debug & 0x8000) {
				/* debug only, avoid syslogd loop */
				kprintf("sync freemap mirror_tid %08jx\n",
					(intmax_t)chain->bref.mirror_tid);
			}

			/*
			 * The freemap can be flushed independently of the
			 * main topology, but for the case where it is
			 * flushed in the same transaction, and flushed
			 * before vchain (a case we want to allow for
			 * performance reasons), make sure modifications
			 * made during the flush under vchain use a new
			 * transaction id.
			 *
			 * Otherwise the mount recovery code will get confused.
			 */
			++hmp->voldata.mirror_tid;
			break;
		case HAMMER2_BREF_TYPE_VOLUME:
			/*
			 * The free block table is flushed by
			 * hammer2_vfs_sync() before it flushes vchain.
			 * We must still hold fchain locked while copying
			 * voldata to volsync, however.
			 *
			 * These do not error per-say since their data does
			 * not need to be re-read from media on lock.
			 *
			 * (note: embedded data, do not call setdirty)
			 */
			hammer2_chain_lock(&hmp->fchain,
					   HAMMER2_RESOLVE_ALWAYS);
			hammer2_voldata_lock(hmp);
			if (hammer2_debug & 0x8000) {
				/* debug only, avoid syslogd loop */
				kprintf("sync volume  mirror_tid %08jx\n",
					(intmax_t)chain->bref.mirror_tid);
			}

			/*
			 * Update the volume header's mirror_tid to the
			 * main topology's flushing mirror_tid.  It is
			 * possible that voldata.mirror_tid is already
			 * beyond bref.mirror_tid due to the bump we made
			 * above in BREF_TYPE_FREEMAP.
			 */
			if (hmp->voldata.mirror_tid < chain->bref.mirror_tid) {
				hmp->voldata.mirror_tid =
					chain->bref.mirror_tid;
			}

			/*
			 * The volume header is flushed manually by the
			 * syncer, not here.  All we do here is adjust the
			 * crc's.
			 */
			KKASSERT(chain->data != NULL);
			KKASSERT(chain->dio == NULL);

			hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT1]=
				hammer2_icrc32(
					(char *)&hmp->voldata +
					 HAMMER2_VOLUME_ICRC1_OFF,
					HAMMER2_VOLUME_ICRC1_SIZE);
			hmp->voldata.icrc_sects[HAMMER2_VOL_ICRC_SECT0]=
				hammer2_icrc32(
					(char *)&hmp->voldata +
					 HAMMER2_VOLUME_ICRC0_OFF,
					HAMMER2_VOLUME_ICRC0_SIZE);
			hmp->voldata.icrc_volheader =
				hammer2_icrc32(
					(char *)&hmp->voldata +
					 HAMMER2_VOLUME_ICRCVH_OFF,
					HAMMER2_VOLUME_ICRCVH_SIZE);

			if (hammer2_debug & 0x8000) {
				/* debug only, avoid syslogd loop */
				kprintf("syncvolhdr %016jx %016jx\n",
					hmp->voldata.mirror_tid,
					hmp->vchain.bref.mirror_tid);
			}
			hmp->volsync = hmp->voldata;
			atomic_set_int(&chain->flags, HAMMER2_CHAIN_VOLUMESYNC);
			hammer2_voldata_unlock(hmp);
			hammer2_chain_unlock(&hmp->fchain);
			break;
		case HAMMER2_BREF_TYPE_DATA:
			/*
			 * Data elements have already been flushed via the
			 * logical file buffer cache.  Their hash was set in
			 * the bref by the vop_write code.  Do not re-dirty.
			 *
			 * Make sure any device buffer(s) have been flushed
			 * out here (there aren't usually any to flush) XXX.
			 */
			break;
		case HAMMER2_BREF_TYPE_INDIRECT:
		case HAMMER2_BREF_TYPE_FREEMAP_NODE:
		case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
			/*
			 * Buffer I/O will be cleaned up when the volume is
			 * flushed (but the kernel is free to flush it before
			 * then, as well).
			 */
			hammer2_chain_setcheck(chain, chain->data);
			break;
		case HAMMER2_BREF_TYPE_DIRENT:
			/*
			 * A directory entry can use the check area to store
			 * the filename for filenames <= 64 bytes, don't blow
			 * it up!
			 */
			if (chain->bytes)
				hammer2_chain_setcheck(chain, chain->data);
			break;
		case HAMMER2_BREF_TYPE_INODE:
			/*
			 * NOTE: We must call io_setdirty() to make any late
			 *	 changes to the inode data, the system might
			 *	 have already flushed the buffer.
			 */
			if (chain->data->ipdata.meta.op_flags &
			    HAMMER2_OPFLAG_PFSROOT) {
				/*
				 * non-NULL pmp if mounted as a PFS.  We must
				 * sync fields cached in the pmp? XXX
				 */
				hammer2_inode_data_t *ipdata;

				hammer2_io_setdirty(chain->dio);
				ipdata = &chain->data->ipdata;
				if (chain->pmp) {
					ipdata->meta.pfs_inum =
						chain->pmp->inode_tid;
				}
			} else {
				/* can't be mounted as a PFS */
			}

			hammer2_chain_setcheck(chain, chain->data);
			break;
		default:
			panic("hammer2_flush_core: unsupported "
			      "embedded bref %d",
			      chain->bref.type);
			/* NOT REACHED */
		}

		/*
		 * If the chain was destroyed try to avoid unnecessary I/O
		 * that might not have yet occurred.  Remove the data range
		 * from dedup candidacy and attempt to invalidation that
		 * potentially dirty portion of the I/O buffer.
		 */
		if (chain->flags & HAMMER2_CHAIN_DESTROY) {
			hammer2_io_dedup_delete(hmp,
						chain->bref.type,
						chain->bref.data_off,
						chain->bytes);
#if 0
			hammer2_io_t *dio;
			if (chain->dio) {
				hammer2_io_inval(chain->dio,
						 chain->bref.data_off,
						 chain->bytes);
			} else if ((dio = hammer2_io_getquick(hmp,
						  chain->bref.data_off,
						  chain->bytes,
						  1)) != NULL) {
				hammer2_io_inval(dio,
						 chain->bref.data_off,
						 chain->bytes);
				hammer2_io_putblk(&dio);
			}
#endif
		}
	}

	/*
	 * If UPDATE is set the parent block table may need to be updated.
	 * This can fail if the hammer2_chain_modify() fails.
	 *
	 * NOTE: UPDATE may be set on vchain or fchain in which case
	 *	 parent could be NULL, or on an inode that has not yet
	 *	 been inserted into the radix tree.  It's easiest to allow
	 *	 the case and test for NULL.  parent can also wind up being
	 *	 NULL due to a deletion so we need to handle the case anyway.
	 *
	 * NOTE: UPDATE can be set when chains are renamed into or out of
	 *	 an indirect block, without the chain itself being flagged
	 *	 MODIFIED.
	 *
	 * If no parent exists we can just clear the UPDATE bit.  If the
	 * chain gets reattached later on the bit will simply get set
	 * again.
	 */
	if ((chain->flags & HAMMER2_CHAIN_UPDATE) && parent == NULL)
		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);

	/*
	 * When flushing an inode outside of a FLUSH_FSSYNC we must NOT
	 * update the parent block table to point at the flushed inode.
	 * The block table should only ever be updated by the filesystem
	 * sync code.  If we do, inode<->inode dependencies (such as
	 * directory entries vs inode nlink count) can wind up not being
	 * flushed together and result in a broken topology if a crash/reboot
	 * occurs at the wrong time.
	 */
	if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
	    (flags & HAMMER2_FLUSH_INODE_STOP) &&
	    (flags & HAMMER2_FLUSH_FSSYNC) == 0 &&
	    (flags & HAMMER2_FLUSH_ALL) == 0 &&
	    chain->pmp && chain->pmp->mp) {
#ifdef HAMMER2_DEBUG_SYNC
		kprintf("inum %ld do not update parent, non-fssync\n",
			(long)chain->bref.key);
#endif
		goto skipupdate;
	}
#ifdef HAMMER2_DEBUG_SYNC
	if (chain->bref.type == HAMMER2_BREF_TYPE_INODE)
		kprintf("inum %ld update parent\n", (long)chain->bref.key);
#endif

	/*
	 * The chain may need its blockrefs updated in the parent, normal
	 * path.
	 */
	if (chain->flags & HAMMER2_CHAIN_UPDATE) {
		hammer2_blockref_t *base;
		int count;

		/*
		 * Clear UPDATE flag, mark parent modified, update its
		 * modify_tid if necessary, and adjust the parent blockmap.
		 */
		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE);

		/*
		 * (optional code)
		 *
		 * Avoid actually modifying and updating the parent if it
		 * was flagged for destruction.  This can greatly reduce
		 * disk I/O in large tree removals because the
		 * hammer2_io_setinval() call in the upward recursion
		 * (see MODIFIED code above) can only handle a few cases.
		 */
		if (parent->flags & HAMMER2_CHAIN_DESTROY) {
			if (parent->bref.modify_tid < chain->bref.modify_tid) {
				parent->bref.modify_tid =
					chain->bref.modify_tid;
			}
			atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BMAPPED |
							HAMMER2_CHAIN_BMAPUPD);
			goto skipupdate;
		}

		/*
		 * The flusher is responsible for deleting empty indirect
		 * blocks at this point.  If we don't do this, no major harm
		 * will be done but the empty indirect blocks will stay in
		 * the topology and make it a messy and inefficient.
		 *
		 * The flusher is also responsible for collapsing the
		 * content of an indirect block into its parent whenever
		 * possible (with some hysteresis).  Not doing this will also
		 * not harm the topology, but would make it messy and
		 * inefficient.
		 */
		if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT) {
			if (hammer2_chain_indirect_maintenance(parent, chain))
				goto skipupdate;
		}

		/*
		 * We are updating the parent's blockmap, the parent must
		 * be set modified.  If this fails we re-set the UPDATE flag
		 * in the child.
		 *
		 * NOTE! A modification error can be ENOSPC.  We still want
		 *	 to flush modified chains recursively, not break out,
		 *	 so we just skip the update in this situation and
		 *	 continue.  That is, we still need to try to clean
		 *	 out dirty chains and buffers.
		 *
		 *	 This may not help bulkfree though. XXX
		 */
		save_error = hammer2_chain_modify(parent, 0, 0, 0);
		if (save_error) {
			info->error |= save_error;
			kprintf("hammer2_flush: %016jx.%02x error=%08x\n",
				parent->bref.data_off, parent->bref.type,
				save_error);
			atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE);
			goto skipupdate;
		}
		if (parent->bref.modify_tid < chain->bref.modify_tid)
			parent->bref.modify_tid = chain->bref.modify_tid;

		/*
		 * Calculate blockmap pointer
		 */
		switch(parent->bref.type) {
		case HAMMER2_BREF_TYPE_INODE:
			/*
			 * Access the inode's block array.  However, there is
			 * no block array if the inode is flagged DIRECTDATA.
			 */
			if (parent->data &&
			    (parent->data->ipdata.meta.op_flags &
			     HAMMER2_OPFLAG_DIRECTDATA) == 0) {
				base = &parent->data->
					ipdata.u.blockset.blockref[0];
			} else {
				base = NULL;
			}
			count = HAMMER2_SET_COUNT;
			break;
		case HAMMER2_BREF_TYPE_INDIRECT:
		case HAMMER2_BREF_TYPE_FREEMAP_NODE:
			if (parent->data)
				base = &parent->data->npdata[0];
			else
				base = NULL;
			count = parent->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 = &parent->data->npdata[0];
			count = HAMMER2_SET_COUNT;
			break;
		default:
			base = NULL;
			count = 0;
			panic("hammer2_flush_core: "
			      "unrecognized blockref type: %d",
			      parent->bref.type);
			break;
		}

		/*
		 * Blocktable updates
		 *
		 * We synchronize pending statistics at this time.  Delta
		 * adjustments designated for the current and upper level
		 * are synchronized.
		 */
		if (base && (chain->flags & HAMMER2_CHAIN_BMAPUPD)) {
			if (chain->flags & HAMMER2_CHAIN_BMAPPED) {
				hammer2_spin_ex(&parent->core.spin);
				hammer2_base_delete(parent, base, count, chain,
						    NULL);
				hammer2_spin_unex(&parent->core.spin);
				/* base_delete clears both bits */
			} else {
				atomic_clear_int(&chain->flags,
						 HAMMER2_CHAIN_BMAPUPD);
			}
		}
		if (base && (chain->flags & HAMMER2_CHAIN_BMAPPED) == 0) {
			hammer2_spin_ex(&parent->core.spin);
			hammer2_base_insert(parent, base, count,
					    chain, &chain->bref);
			hammer2_spin_unex(&parent->core.spin);
			/* base_insert sets BMAPPED */
		}
	}
skipupdate:
	if (parent)
		hammer2_chain_unlock(parent);

	/*
	 * Final cleanup after flush
	 */
done:
	KKASSERT(chain->refs > 0);
	if (hammer2_debug & 0x200) {
		if (info->debug == chain)
			info->debug = NULL;
	}
	return retry;
}

/*
 * Flush recursion helper, called from flush_core, calls flush_core.
 *
 * Flushes the children of the caller's chain (info->parent), restricted
 * by sync_tid.
 *
 * This function may set info->error as a side effect.
 *
 * WARNING! If we do not call hammer2_flush_core() we must update
 *	    bref.mirror_tid ourselves to indicate that the flush has
 *	    processed the child.
 *
 * WARNING! parent->core spinlock is held on entry and return.
 */
static int
hammer2_flush_recurse(hammer2_chain_t *child, void *data)
{
	hammer2_flush_info_t *info = data;
	hammer2_chain_t *parent = info->parent;

#ifdef HAMMER2_SCAN_DEBUG
	++info->scan_count;
	if (child->flags & HAMMER2_CHAIN_MODIFIED)
		++info->scan_mod_count;
	if (child->flags & HAMMER2_CHAIN_UPDATE)
		++info->scan_upd_count;
	if (child->flags & HAMMER2_CHAIN_ONFLUSH)
		++info->scan_onf_count;
#endif

	/*
	 * (child can never be fchain or vchain so a special check isn't
	 *  needed).
	 *
	 * We must ref the child before unlocking the spinlock.
	 *
	 * The caller has added a ref to the parent so we can temporarily
	 * unlock it in order to lock the child.  However, if it no longer
	 * winds up being the child of the parent we must skip this child.
	 *
	 * NOTE! chain locking errors are fatal.  They are never out-of-space
	 *	 errors.
	 */
	hammer2_chain_ref(child);
	hammer2_spin_unex(&parent->core.spin);

	hammer2_chain_ref_hold(parent);
	hammer2_chain_unlock(parent);
	hammer2_chain_lock(child, HAMMER2_RESOLVE_MAYBE);
	if (child->parent != parent) {
		kprintf("LOST CHILD1 %p->%p (actual parent %p)\n",
			parent, child, child->parent);
		goto done;
	}
	if (child->error) {
		kprintf("CHILD ERROR DURING FLUSH LOCK %p->%p\n",
			parent, child);
		info->error |= child->error;
		goto done;
	}

	/*
	 * Must propagate the DESTROY flag downwards, otherwise the
	 * parent could end up never being removed because it will
	 * be requeued to the flusher if it survives this run due to
	 * the flag.
	 */
	if (parent && (parent->flags & HAMMER2_CHAIN_DESTROY))
		atomic_set_int(&child->flags, HAMMER2_CHAIN_DESTROY);
#ifdef HAMMER2_SCAN_DEBUG
	if (child->flags & HAMMER2_CHAIN_DESTROY)
		++info->scan_del_count;
#endif
	/*
	 * Special handling of the root inode.  Because the root inode
	 * contains an index of all the inodes in the PFS in addition to
	 * its normal directory entries, any flush that is not part of a
	 * filesystem sync must only flush the directory entries, and not
	 * anything else.
	 *
	 * The child might be an indirect block, but H2 guarantees that
	 * the key-range will fully partition the inode index from the
	 * directory entries so the case just works naturally.
	 */
	if ((parent->bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
	    (child->flags & HAMMER2_CHAIN_DESTROY) == 0 &&
	    parent->bref.type == HAMMER2_BREF_TYPE_INODE &&
	    (info->flags & HAMMER2_FLUSH_FSSYNC) == 0) {
		if ((child->bref.key & HAMMER2_DIRHASH_VISIBLE) == 0) {
			if (child->flags & HAMMER2_CHAIN_FLUSH_MASK) {
				hammer2_chain_setflush(parent);
			}
			goto done;
		}
	}

	/*
	 * Recurse and collect deferral data.  We're in the media flush,
	 * this can cross PFS boundaries.
	 */
	if (child->flags & HAMMER2_CHAIN_FLUSH_MASK) {
#ifdef HAMMER2_SCAN_DEBUG
		if (child->bref.type < 7)
			++info->scan_btype[child->bref.type];
#endif
		++info->depth;
		hammer2_flush_core(info, child, info->flags);
		--info->depth;
	} else if (hammer2_debug & 0x200) {
		if (info->debug == NULL)
			info->debug = child;
		++info->depth;
		hammer2_flush_core(info, child, info->flags);
		--info->depth;
		if (info->debug == child)
			info->debug = NULL;
	}

done:
	/*
	 * Relock to continue the loop.
	 */
	hammer2_chain_unlock(child);
	hammer2_chain_lock(parent, HAMMER2_RESOLVE_MAYBE);
	hammer2_chain_drop_unhold(parent);
	if (parent->error) {
		kprintf("PARENT ERROR DURING FLUSH LOCK %p->%p\n",
			parent, child);
		info->error |= parent->error;
	}
	hammer2_chain_drop(child);
	KKASSERT(info->parent == parent);
	hammer2_spin_ex(&parent->core.spin);

	return (0);
}

/*
 * flush helper (backend threaded)
 *
 * Flushes chain topology for the specified inode.
 *
 * HAMMER2_XOP_INODE_STOP	The flush recursion stops at inode boundaries.
 *				Inodes belonging to the same flush are flushed
 *				separately.
 *
 * chain->parent can be NULL, usually due to destroy races or detached inodes.
 *
 * Primarily called from vfs_sync().
 */
void
hammer2_xop_inode_flush(hammer2_xop_t *arg, void *scratch __unused, int clindex)
{
	hammer2_xop_flush_t *xop = &arg->xop_flush;
	hammer2_chain_t *chain;
	hammer2_inode_t *ip;
	hammer2_dev_t *hmp;
	hammer2_pfs_t *pmp;
	hammer2_devvp_t *e;
	struct vnode *devvp;
	int flush_error = 0;
	int fsync_error = 0;
	int total_error = 0;
	int j;
	int xflags;
	int ispfsroot = 0;

	xflags = HAMMER2_FLUSH_TOP;
	if (xop->head.flags & HAMMER2_XOP_INODE_STOP)
		xflags |= HAMMER2_FLUSH_INODE_STOP;
	if (xop->head.flags & HAMMER2_XOP_FSSYNC)
		xflags |= HAMMER2_FLUSH_FSSYNC;

	/*
	 * Flush core chains
	 */
	ip = xop->head.ip1;
	pmp = ip->pmp;
	chain = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS);
	if (chain) {
		hmp = chain->hmp;
		if (chain->flags & HAMMER2_CHAIN_FLUSH_MASK) {
			/*
			 * Due to flush partitioning the chain topology
			 * above the inode's chain may no longer be flagged.
			 * When asked to flush an inode, remark the topology
			 * leading to that inode.
			 */
			if (chain->parent)
				hammer2_chain_setflush(chain->parent);
			hammer2_flush(chain, xflags);

			/* XXX cluster */
			if (ip == pmp->iroot && pmp != hmp->spmp) {
				hammer2_spin_ex(&pmp->inum_spin);
				pmp->pfs_iroot_blocksets[clindex] =
					chain->data->ipdata.u.blockset;
				hammer2_spin_unex(&pmp->inum_spin);
			}

#if 0
			/*
			 * Propogate upwards but only cross an inode boundary
			 * for inodes associated with the current filesystem
			 * sync.
			 */
			if ((xop->head.flags & HAMMER2_XOP_PARENTONFLUSH) ||
			    chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
				parent = chain->parent;
				if (parent)
					hammer2_chain_setflush(parent);
			}
#endif
		}
		if (chain->flags & HAMMER2_CHAIN_PFSBOUNDARY)
			ispfsroot = 1;
		hammer2_chain_unlock(chain);
		hammer2_chain_drop(chain);
		chain = NULL;
	} else {
		hmp = NULL;
	}

	/*
	 * Only flush the volume header if asked to, plus the inode must also
	 * be the PFS root.
	 */
	if ((xop->head.flags & HAMMER2_XOP_VOLHDR) == 0)
		goto skip;
	if (ispfsroot == 0)
		goto skip;

	/*
	 * Flush volume roots.  Avoid replication, we only want to
	 * flush each hammer2_dev (hmp) once.
	 */
	for (j = clindex - 1; j >= 0; --j) {
		if ((chain = ip->cluster.array[j].chain) != NULL) {
			if (chain->hmp == hmp) {
				chain = NULL;	/* safety */
				goto skip;
			}
		}
	}
	chain = NULL;	/* safety */

	/*
	 * spmp transaction.  The super-root is never directly mounted so
	 * there shouldn't be any vnodes, let alone any dirty vnodes
	 * associated with it, so we shouldn't have to mess around with any
	 * vnode flushes here.
	 */
	hammer2_trans_init(hmp->spmp, HAMMER2_TRANS_ISFLUSH);

	/*
	 * We must flush the superroot down to the PFS iroot.  Remember
	 * that hammer2_chain_setflush() stops at inode boundaries, so
	 * the pmp->iroot has been flushed and flagged down to the superroot,
	 * but the volume root (vchain) probably has not yet been flagged.
	 */
	if (hmp->spmp->iroot) {
		chain = hmp->spmp->iroot->cluster.array[0].chain;
		if (chain) {
			hammer2_chain_ref(chain);
			hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
			flush_error |=
				hammer2_flush(chain,
					      HAMMER2_FLUSH_TOP |
					      HAMMER2_FLUSH_INODE_STOP |
					      HAMMER2_FLUSH_FSSYNC);
			hammer2_chain_unlock(chain);
			hammer2_chain_drop(chain);
		}
	}

	/*
	 * Media mounts have two 'roots', vchain for the topology
	 * and fchain for the free block table.  Flush both.
	 *
	 * Note that the topology and free block table are handled
	 * independently, so the free block table can wind up being
	 * ahead of the topology.  We depend on the bulk free scan
	 * code to deal with any loose ends.
	 *
	 * vchain and fchain do not error on-lock since their data does
	 * not have to be re-read from media.
	 */
	hammer2_chain_ref(&hmp->vchain);
	hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
	hammer2_chain_ref(&hmp->fchain);
	hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
	if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
		/*
		 * This will also modify vchain as a side effect,
		 * mark vchain as modified now.
		 */
		hammer2_voldata_modify(hmp);
		chain = &hmp->fchain;
		flush_error |= hammer2_flush(chain, HAMMER2_FLUSH_TOP);
		KKASSERT(chain == &hmp->fchain);
	}
	hammer2_chain_unlock(&hmp->fchain);
	hammer2_chain_unlock(&hmp->vchain);
	hammer2_chain_drop(&hmp->fchain);
	/* vchain dropped down below */

	hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
	if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
		chain = &hmp->vchain;
		flush_error |= hammer2_flush(chain, HAMMER2_FLUSH_TOP);
		KKASSERT(chain == &hmp->vchain);
	}
	hammer2_chain_unlock(&hmp->vchain);
	hammer2_chain_drop(&hmp->vchain);

	/*
	 * We can't safely flush the volume header until we have
	 * flushed any device buffers which have built up.
	 *
	 * XXX this isn't being incremental
	 */
	TAILQ_FOREACH(e, &hmp->devvpl, entry) {
		devvp = e->devvp;
		KKASSERT(devvp);
		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
		fsync_error = VOP_FSYNC(devvp, MNT_WAIT, 0);
		vn_unlock(devvp);
		if (fsync_error || flush_error) {
			kprintf("hammer2: sync error fsync=%d h2flush=0x%04x dev=%s\n",
				fsync_error, flush_error, e->path);
		}
	}

	/*
	 * The flush code sets CHAIN_VOLUMESYNC to indicate that the
	 * volume header needs synchronization via hmp->volsync.
	 *
	 * XXX synchronize the flag & data with only this flush XXX
	 */
	if (fsync_error == 0 && flush_error == 0 &&
	    (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
		struct buf *bp;
		int vol_error = 0;

		/*
		 * Synchronize the disk before flushing the volume
		 * header.
		 */
		bp = getpbuf(NULL);
		bp->b_bio1.bio_offset = 0;
		bp->b_bufsize = 0;
		bp->b_bcount = 0;
		bp->b_cmd = BUF_CMD_FLUSH;
		bp->b_bio1.bio_done = biodone_sync;
		bp->b_bio1.bio_flags |= BIO_SYNC;
		vn_strategy(hmp->devvp, &bp->b_bio1);
		fsync_error = biowait(&bp->b_bio1, "h2vol");
		relpbuf(bp, NULL);

		/*
		 * Then we can safely flush the version of the
		 * volume header synchronized by the flush code.
		 */
		j = hmp->volhdrno + 1;
		if (j < 0)
			j = 0;
		if (j >= HAMMER2_NUM_VOLHDRS)
			j = 0;
		if (j * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
		    hmp->volsync.volu_size) {
			j = 0;
		}
		if (hammer2_debug & 0x8000) {
			/* debug only, avoid syslogd loop */
			kprintf("sync volhdr %d %jd\n",
				j, (intmax_t)hmp->volsync.volu_size);
		}
		bp = getblk(hmp->devvp, j * HAMMER2_ZONE_BYTES64,
			    HAMMER2_PBUFSIZE, GETBLK_KVABIO, 0);
		atomic_clear_int(&hmp->vchain.flags,
				 HAMMER2_CHAIN_VOLUMESYNC);
		bkvasync(bp);
		bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
		vol_error = bwrite(bp);
		hmp->volhdrno = j;
		if (vol_error)
			fsync_error = vol_error;
	}
	if (flush_error)
		total_error = flush_error;
	if (fsync_error)
		total_error = hammer2_errno_to_error(fsync_error);

	/* spmp trans */
	hammer2_trans_done(hmp->spmp, HAMMER2_TRANS_ISFLUSH);
skip:
	hammer2_xop_feed(&xop->head, NULL, clindex, total_error);
}