xref: /xnu-12377.81.4/bsd/vfs/vfs_cluster.c (revision 043036a2b3718f7f0be807e2870f8f47d3fa0796)

/*
 * Copyright (c) 2000-2020 Apple Inc. All rights reserved.
 *
 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
 *
 * This file contains Original Code and/or Modifications of Original Code
 * as defined in and that are subject to the Apple Public Source License
 * Version 2.0 (the 'License'). You may not use this file except in
 * compliance with the License. The rights granted to you under the License
 * may not be used to create, or enable the creation or redistribution of,
 * unlawful or unlicensed copies of an Apple operating system, or to
 * circumvent, violate, or enable the circumvention or violation of, any
 * terms of an Apple operating system software license agreement.
 *
 * Please obtain a copy of the License at
 * http://www.opensource.apple.com/apsl/ and read it before using this file.
 *
 * The Original Code and all software distributed under the License are
 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
 * Please see the License for the specific language governing rights and
 * limitations under the License.
 *
 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
 */
/* Copyright (c) 1995 NeXT Computer, Inc. All Rights Reserved */
/*
 * Copyright (c) 1993
 *	The Regents of the University of California.  All rights reserved.
 *
 * 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. All advertising materials mentioning features or use of this software
 *    must display the following acknowledgement:
 *	This product includes software developed by the University of
 *	California, Berkeley and its contributors.
 * 4. Neither the name of the University 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 REGENTS 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 REGENTS 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.
 *
 *	@(#)vfs_cluster.c	8.10 (Berkeley) 3/28/95
 */

#include <sys/param.h>
#include <sys/proc_internal.h>
#include <sys/buf_internal.h>
#include <sys/mount_internal.h>
#include <sys/vnode_internal.h>
#include <sys/trace.h>
#include <kern/kalloc.h>
#include <sys/time.h>
#include <sys/kernel.h>
#include <sys/resourcevar.h>
#include <miscfs/specfs/specdev.h>
#include <sys/uio_internal.h>
#include <libkern/libkern.h>
#include <machine/machine_routines.h>
#include <machine/smp.h>

#include <sys/ubc_internal.h>
#include <vm/vnode_pager.h>
#include <vm/vm_upl.h>

#include <mach/mach_types.h>
#include <mach/memory_object_types.h>
#include <mach/vm_map.h>
#include <mach/upl.h>
#include <mach/thread_info.h>
#include <kern/task.h>
#include <kern/policy_internal.h>
#include <kern/thread.h>

#include <vm/vm_kern_xnu.h>
#include <vm/vm_map_xnu.h>
#include <vm/vm_pageout_xnu.h>
#include <vm/vm_fault.h>
#include <vm/vm_ubc.h>

#include <sys/kdebug.h>
#include <sys/kdebug_triage.h>
#include <libkern/OSAtomic.h>

#include <sys/sdt.h>

#include <stdbool.h>

#include <vfs/vfs_disk_conditioner.h>

#if 0
#undef KERNEL_DEBUG
#define KERNEL_DEBUG KERNEL_DEBUG_CONSTANT
#endif


#define CL_READ         0x01
#define CL_WRITE        0x02
#define CL_ASYNC        0x04
#define CL_COMMIT       0x08
#define CL_PAGEOUT      0x10
#define CL_AGE          0x20
#define CL_NOZERO       0x40
#define CL_PAGEIN       0x80
#define CL_DEV_MEMORY   0x100
#define CL_PRESERVE     0x200
#define CL_THROTTLE     0x400
#define CL_KEEPCACHED   0x800
#define CL_DIRECT_IO    0x1000
#define CL_PASSIVE      0x2000
#define CL_IOSTREAMING  0x4000
#define CL_CLOSE        0x8000
#define CL_ENCRYPTED    0x10000
#define CL_RAW_ENCRYPTED        0x20000
#define CL_NOCACHE      0x40000
#define CL_DIRECT_IO_FSBLKSZ    0x80000

#define MAX_VECTOR_UPL_SIZE     (2 * MAX_UPL_SIZE_BYTES)

#define CLUSTER_IO_WAITING              ((buf_t)1)

extern void vector_upl_set_iostate(upl_t, upl_t, vm_offset_t, upl_size_t);

struct clios {
	lck_mtx_t io_mtxp;
	u_int  io_completed;       /* amount of io that has currently completed */
	u_int  io_issued;          /* amount of io that was successfully issued */
	int    io_error;           /* error code of first error encountered */
	int    io_wanted;          /* someone is sleeping waiting for a change in state */
};

struct cl_direct_read_lock {
	LIST_ENTRY(cl_direct_read_lock)         chain;
	int32_t                                                         ref_count;
	vnode_t                                                         vp;
	lck_rw_t                                                        rw_lock;
};

#define CL_DIRECT_READ_LOCK_BUCKETS 61

static LIST_HEAD(cl_direct_read_locks, cl_direct_read_lock)
cl_direct_read_locks[CL_DIRECT_READ_LOCK_BUCKETS];

static LCK_GRP_DECLARE(cl_mtx_grp, "cluster I/O");
static LCK_MTX_DECLARE(cl_transaction_mtxp, &cl_mtx_grp);
static LCK_SPIN_DECLARE(cl_direct_read_spin_lock, &cl_mtx_grp);

static ZONE_DEFINE(cl_rd_zone, "cluster_read",
    sizeof(struct cl_readahead), ZC_ZFREE_CLEARMEM);

static ZONE_DEFINE(cl_wr_zone, "cluster_write",
    sizeof(struct cl_writebehind), ZC_ZFREE_CLEARMEM);

#define IO_UNKNOWN      0
#define IO_DIRECT       1
#define IO_CONTIG       2
#define IO_COPY         3

#define PUSH_DELAY      0x01
#define PUSH_ALL        0x02
#define PUSH_SYNC       0x04


static void cluster_EOT(buf_t cbp_head, buf_t cbp_tail, int zero_offset, size_t verify_block_size);
static void cluster_wait_IO(buf_t cbp_head, int async);
static void cluster_complete_transaction(buf_t *cbp_head, void *callback_arg, int *retval, int flags, int needwait);

static int cluster_io_type(struct uio *uio, int *io_type, u_int32_t *io_length, u_int32_t min_length);

static int cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int non_rounded_size,
    int flags, buf_t real_bp, struct clios *iostate, int (*)(buf_t, void *), void *callback_arg);
static void cluster_iodone_verify_continue(void);
static int cluster_iodone(buf_t bp, void *callback_arg);
static int cluster_iodone_finish(buf_t cbp_head, void *callback_arg);
static int cluster_ioerror(upl_t upl, int upl_offset, int abort_size, int error, int io_flags, vnode_t vp);
static int cluster_is_throttled(vnode_t vp);

static void cluster_iostate_wait(struct clios *iostate, u_int target, const char *wait_name);

static void cluster_syncup(vnode_t vp, off_t newEOF, int (*)(buf_t, void *), void *callback_arg, int flags);

static int cluster_handle_split_pagein(vnode_t vp, upl_t upl, upl_offset_t upl_offset, off_t f_offset,
    u_int io_size, int rounded_size, int local_flags, int (*callback)(buf_t, void *), void *callback_arg);

static void cluster_read_upl_release(upl_t upl, int start_pg, int last_pg, int take_reference);
static int cluster_copy_ubc_data_internal(vnode_t vp, struct uio *uio, int *io_resid, int mark_dirty, int take_reference);

static int cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t filesize, int flags,
    int (*)(buf_t, void *), void *callback_arg) __attribute__((noinline));
static int cluster_read_direct(vnode_t vp, struct uio *uio, off_t filesize, int *read_type, u_int32_t *read_length,
    int flags, int (*)(buf_t, void *), void *callback_arg) __attribute__((noinline));
static int cluster_read_contig(vnode_t vp, struct uio *uio, off_t filesize, int *read_type, u_int32_t *read_length,
    int (*)(buf_t, void *), void *callback_arg, int flags) __attribute__((noinline));

static int cluster_write_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t oldEOF, off_t newEOF,
    off_t headOff, off_t tailOff, int flags, int (*)(buf_t, void *), void *callback_arg) __attribute__((noinline));
static int cluster_write_direct(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, int *write_type, u_int32_t *write_length,
    int flags, int (*callback)(buf_t, void *), void *callback_arg, uint32_t min_io_size) __attribute__((noinline));
static int cluster_write_contig(vnode_t vp, struct uio *uio, off_t newEOF,
    int *write_type, u_int32_t *write_length, int (*)(buf_t, void *), void *callback_arg, int bflag) __attribute__((noinline));

static void cluster_update_state_internal(vnode_t vp, struct cl_extent *cl, int flags, boolean_t defer_writes, boolean_t *first_pass,
    off_t write_off, int write_cnt, off_t newEOF, int (*callback)(buf_t, void *), void *callback_arg, boolean_t vm_initiated);

static int cluster_align_phys_io(vnode_t vp, struct uio *uio, addr64_t usr_paddr, u_int32_t xsize, int flags, int (*)(buf_t, void *), void *callback_arg);

static int      cluster_read_prefetch(vnode_t vp, off_t f_offset, u_int size, off_t filesize, int (*callback)(buf_t, void *), void *callback_arg, int bflag);
static void     cluster_read_ahead(vnode_t vp, struct cl_extent *extent, off_t filesize, struct cl_readahead *ra,
    int (*callback)(buf_t, void *), void *callback_arg, int bflag);

static int      cluster_push_now(vnode_t vp, struct cl_extent *, off_t EOF, int flags, int (*)(buf_t, void *), void *callback_arg, boolean_t vm_ioitiated);

static int      cluster_try_push(struct cl_writebehind *, vnode_t vp, off_t EOF, int push_flag, int flags, int (*)(buf_t, void *),
    void *callback_arg, int *err, boolean_t vm_initiated);

static int      sparse_cluster_switch(struct cl_writebehind *, vnode_t vp, off_t EOF, int (*)(buf_t, void *), void *callback_arg, boolean_t vm_initiated);
static int      sparse_cluster_push(struct cl_writebehind *, void **cmapp, vnode_t vp, off_t EOF, int push_flag,
    int io_flags, int (*)(buf_t, void *), void *callback_arg, boolean_t vm_initiated);
static int      sparse_cluster_add(struct cl_writebehind *, void **cmapp, vnode_t vp, struct cl_extent *, off_t EOF,
    int (*)(buf_t, void *), void *callback_arg, boolean_t vm_initiated);

static kern_return_t vfs_drt_mark_pages(void **cmapp, off_t offset, u_int length, u_int *setcountp);
static kern_return_t vfs_drt_get_cluster(void **cmapp, off_t *offsetp, u_int *lengthp);
static kern_return_t vfs_drt_control(void **cmapp, int op_type);
static kern_return_t vfs_get_scmap_push_behavior_internal(void **cmapp, int *push_flag);


/*
 * For throttled IO to check whether
 * a block is cached by the boot cache
 * and thus it can avoid delaying the IO.
 *
 * bootcache_contains_block is initially
 * NULL. The BootCache will set it while
 * the cache is active and clear it when
 * the cache is jettisoned.
 *
 * Returns 0 if the block is not
 * contained in the cache, 1 if it is
 * contained.
 *
 * The function pointer remains valid
 * after the cache has been evicted even
 * if bootcache_contains_block has been
 * cleared.
 *
 * See rdar://9974130 The new throttling mechanism breaks the boot cache for throttled IOs
 */
int (*bootcache_contains_block)(dev_t device, u_int64_t blkno) = NULL;


/*
 * limit the internal I/O size so that we
 * can represent it in a 32 bit int
 */
#define MAX_IO_REQUEST_SIZE     (1024 * 1024 * 512)
#define MAX_IO_CONTIG_SIZE      MAX_UPL_SIZE_BYTES
#define MAX_VECTS               16
/*
 * The MIN_DIRECT_WRITE_SIZE governs how much I/O should be issued before we consider
 * allowing the caller to bypass the buffer cache.  For small I/Os (less than 16k),
 * we have not historically allowed the write to bypass the UBC.
 */
#define MIN_DIRECT_WRITE_SIZE   (16384)

#define WRITE_THROTTLE          6
#define WRITE_THROTTLE_SSD      2
#define WRITE_BEHIND            1
#define WRITE_BEHIND_SSD        1

#if !defined(XNU_TARGET_OS_OSX)
#define PREFETCH                1
#define PREFETCH_SSD            1
uint32_t speculative_prefetch_max = (2048 * 1024);              /* maximum bytes in a specluative read-ahead */
uint32_t speculative_prefetch_max_iosize = (512 * 1024);        /* maximum I/O size to use in a specluative read-ahead */
#else /* XNU_TARGET_OS_OSX */
#define PREFETCH                3
#define PREFETCH_SSD            2
uint32_t speculative_prefetch_max = (MAX_UPL_SIZE_BYTES * 3);   /* maximum bytes in a specluative read-ahead */
uint32_t speculative_prefetch_max_iosize = (512 * 1024);        /* maximum I/O size to use in a specluative read-ahead on SSDs*/
#endif /* ! XNU_TARGET_OS_OSX */

/* maximum bytes for read-ahead */
uint32_t prefetch_max = (1024 * 1024 * 1024);
/* maximum bytes for outstanding reads */
uint32_t overlapping_read_max = (1024 * 1024 * 1024);
/* maximum bytes for outstanding writes */
uint32_t overlapping_write_max = (1024 * 1024 * 1024);

#define IO_SCALE(vp, base)              (vp->v_mount->mnt_ioscale * (base))
#define MAX_CLUSTER_SIZE(vp)            (cluster_max_io_size(vp->v_mount, CL_WRITE))

int     speculative_reads_disabled = 0;

/*
 * throttle the number of async writes that
 * can be outstanding on a single vnode
 * before we issue a synchronous write
 */
#define THROTTLE_MAXCNT 0

uint32_t throttle_max_iosize = (128 * 1024);

#define THROTTLE_MAX_IOSIZE (throttle_max_iosize)

SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_max_iosize, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_max_iosize, 0, "");

uint32_t split_pgin = 1;
uint32_t split_all_pgin = 1;
uint32_t split_all_pgin_equal = 0;
uint32_t split_pgin_headio = 0;

SYSCTL_INT(_kern, OID_AUTO, split_pagein_io, CTLFLAG_RW | CTLFLAG_LOCKED, &split_pgin, 0, "");
#if DEVELOPMENT || DEBUG
SYSCTL_INT(_kern, OID_AUTO, split_pagein_io_all, CTLFLAG_RW | CTLFLAG_LOCKED, &split_all_pgin, 0, "");
SYSCTL_INT(_kern, OID_AUTO, split_pagein_io_equal, CTLFLAG_RW | CTLFLAG_LOCKED, &split_all_pgin_equal, 0, "");
SYSCTL_INT(_kern, OID_AUTO, split_pagein_do_headio, CTLFLAG_RW | CTLFLAG_LOCKED, &split_pgin_headio, 0, "");
#endif

struct verify_buf {
	TAILQ_ENTRY(verify_buf) vb_entry;
	buf_t vb_cbp;
	void* vb_callback_arg;
	int32_t vb_whichq;
};

TAILQ_HEAD(, verify_buf) verify_free_head;
TAILQ_HEAD(, verify_buf) verify_work_head;

#define MAX_VERIFY_THREADS 4
#define MAX_REQUESTS_PER_THREAD  2

static struct verify_buf verify_bufs[MAX_VERIFY_THREADS * MAX_REQUESTS_PER_THREAD];
/*
 * Each thread needs to check if the item at the head of the queue has a UPL
 * pointer that is any of the threads are currently operating on.
 * slot 0 is for the io completion thread to do the request inline if there are no free
 * queue slots.
 */
static int verify_in_flight = 0;

#if defined(XNU_TARGET_OS_IOS) || defined(XNU_TARGET_OS_XR)
#define NUM_DEFAULT_THREADS 2
#elif defined(XNU_TARGET_OS_OSX)
#define NUM_DEFAULT_THREADS 4
#else
#define NUM_DEFAULT_THREADS 0
#endif

static TUNABLE(uint32_t, num_verify_threads, "num_verify_threads", NUM_DEFAULT_THREADS);
static uint32_t cluster_verify_threads = 0; /* will be launched as needed upto num_verify_threads */

#if __AMP__
static TUNABLE(uint32_t, ecore_verify_threads, "ecore_verify_threads", false);
#endif /* __AMP__ */

static void
cluster_verify_init(void)
{
	TAILQ_INIT(&verify_free_head);
	TAILQ_INIT(&verify_work_head);

	if (num_verify_threads > MAX_VERIFY_THREADS) {
		num_verify_threads = MAX_VERIFY_THREADS;
	}

	for (int i = 0; i < num_verify_threads * MAX_REQUESTS_PER_THREAD; i++) {
		TAILQ_INSERT_TAIL(&verify_free_head, &verify_bufs[i], vb_entry);
	}
}

void
cluster_init(void)
{
	for (int i = 0; i < CL_DIRECT_READ_LOCK_BUCKETS; ++i) {
		LIST_INIT(&cl_direct_read_locks[i]);
	}

	cluster_verify_init();
}

uint32_t
cluster_max_io_size(mount_t mp, int type)
{
	uint32_t        max_io_size;
	uint32_t        segcnt;
	uint32_t        maxcnt;

	switch (type) {
	case CL_READ:
		segcnt = mp->mnt_segreadcnt;
		maxcnt = mp->mnt_maxreadcnt;
		break;
	case CL_WRITE:
		segcnt = mp->mnt_segwritecnt;
		maxcnt = mp->mnt_maxwritecnt;
		break;
	default:
		segcnt = min(mp->mnt_segreadcnt, mp->mnt_segwritecnt);
		maxcnt = min(mp->mnt_maxreadcnt, mp->mnt_maxwritecnt);
		break;
	}
	if (segcnt > (MAX_UPL_SIZE_BYTES >> PAGE_SHIFT)) {
		/*
		 * don't allow a size beyond the max UPL size we can create
		 */
		segcnt = MAX_UPL_SIZE_BYTES >> PAGE_SHIFT;
	}
	max_io_size = min((segcnt * PAGE_SIZE), maxcnt);

	if (max_io_size < MAX_UPL_TRANSFER_BYTES) {
		/*
		 * don't allow a size smaller than the old fixed limit
		 */
		max_io_size = MAX_UPL_TRANSFER_BYTES;
	} else {
		/*
		 * make sure the size specified is a multiple of PAGE_SIZE
		 */
		max_io_size &= ~PAGE_MASK;
	}
	return max_io_size;
}

/*
 * Returns max prefetch value. If the value overflows or exceeds the specified
 * 'prefetch_limit', it will be capped at 'prefetch_limit' value.
 */
static inline uint32_t
cluster_max_prefetch(vnode_t vp, uint32_t max_io_size, uint32_t prefetch_limit)
{
	bool is_ssd = disk_conditioner_mount_is_ssd(vp->v_mount);
	uint32_t io_scale = IO_SCALE(vp, is_ssd ? PREFETCH_SSD : PREFETCH);
	uint32_t prefetch = 0;

	if (__improbable(os_mul_overflow(max_io_size, io_scale, &prefetch) ||
	    (prefetch > prefetch_limit))) {
		prefetch = prefetch_limit;
	}

	return prefetch;
}

static inline uint32_t
calculate_max_throttle_size(vnode_t vp)
{
	bool is_ssd = disk_conditioner_mount_is_ssd(vp->v_mount);
	uint32_t io_scale = IO_SCALE(vp, is_ssd ? 2 : 1);

	return MIN(io_scale * THROTTLE_MAX_IOSIZE, MAX_UPL_TRANSFER_BYTES);
}

static inline uint32_t
calculate_max_throttle_cnt(vnode_t vp)
{
	bool is_ssd = disk_conditioner_mount_is_ssd(vp->v_mount);
	uint32_t io_scale = IO_SCALE(vp, 1);

	return is_ssd ? MIN(io_scale, 4) : THROTTLE_MAXCNT;
}

#define CLW_ALLOCATE            0x01
#define CLW_RETURNLOCKED        0x02
#define CLW_IONOCACHE           0x04
#define CLW_IOPASSIVE   0x08

/*
 * if the read ahead context doesn't yet exist,
 * allocate and initialize it...
 * the vnode lock serializes multiple callers
 * during the actual assignment... first one
 * to grab the lock wins... the other callers
 * will release the now unnecessary storage
 *
 * once the context is present, try to grab (but don't block on)
 * the lock associated with it... if someone
 * else currently owns it, than the read
 * will run without read-ahead.  this allows
 * multiple readers to run in parallel and
 * since there's only 1 read ahead context,
 * there's no real loss in only allowing 1
 * reader to have read-ahead enabled.
 */
static struct cl_readahead *
cluster_get_rap(vnode_t vp)
{
	struct ubc_info         *ubc;
	struct cl_readahead     *rap;

	ubc = vp->v_ubcinfo;

	if ((rap = ubc->cl_rahead) == NULL) {
		rap = zalloc_flags(cl_rd_zone, Z_WAITOK | Z_ZERO);
		rap->cl_lastr = -1;
		lck_mtx_init(&rap->cl_lockr, &cl_mtx_grp, LCK_ATTR_NULL);

		vnode_lock(vp);

		if (ubc->cl_rahead == NULL) {
			ubc->cl_rahead = rap;
		} else {
			lck_mtx_destroy(&rap->cl_lockr, &cl_mtx_grp);
			zfree(cl_rd_zone, rap);
			rap = ubc->cl_rahead;
		}
		vnode_unlock(vp);
	}
	if (lck_mtx_try_lock(&rap->cl_lockr) == TRUE) {
		return rap;
	}

	return (struct cl_readahead *)NULL;
}


/*
 * if the write behind context doesn't yet exist,
 * and CLW_ALLOCATE is specified, allocate and initialize it...
 * the vnode lock serializes multiple callers
 * during the actual assignment... first one
 * to grab the lock wins... the other callers
 * will release the now unnecessary storage
 *
 * if CLW_RETURNLOCKED is set, grab (blocking if necessary)
 * the lock associated with the write behind context before
 * returning
 */

static struct cl_writebehind *
cluster_get_wbp(vnode_t vp, int flags)
{
	struct ubc_info *ubc;
	struct cl_writebehind *wbp;

	ubc = vp->v_ubcinfo;

	if ((wbp = ubc->cl_wbehind) == NULL) {
		if (!(flags & CLW_ALLOCATE)) {
			return (struct cl_writebehind *)NULL;
		}

		wbp = zalloc_flags(cl_wr_zone, Z_WAITOK | Z_ZERO);

		lck_mtx_init(&wbp->cl_lockw, &cl_mtx_grp, LCK_ATTR_NULL);

		vnode_lock(vp);

		if (ubc->cl_wbehind == NULL) {
			ubc->cl_wbehind = wbp;
		} else {
			lck_mtx_destroy(&wbp->cl_lockw, &cl_mtx_grp);
			zfree(cl_wr_zone, wbp);
			wbp = ubc->cl_wbehind;
		}
		vnode_unlock(vp);
	}
	if (flags & CLW_RETURNLOCKED) {
		lck_mtx_lock(&wbp->cl_lockw);
	}

	return wbp;
}


static void
cluster_syncup(vnode_t vp, off_t newEOF, int (*callback)(buf_t, void *), void *callback_arg, int flags)
{
	struct cl_writebehind *wbp;

	if ((wbp = cluster_get_wbp(vp, 0)) != NULL) {
		if (wbp->cl_number) {
			lck_mtx_lock(&wbp->cl_lockw);

			cluster_try_push(wbp, vp, newEOF, PUSH_ALL | flags, 0, callback, callback_arg, NULL, FALSE);

			lck_mtx_unlock(&wbp->cl_lockw);
		}
	}
}


static int
cluster_io_present_in_BC(vnode_t vp, off_t f_offset)
{
	daddr64_t blkno;
	size_t    io_size;
	int (*bootcache_check_fn)(dev_t device, u_int64_t blkno) = bootcache_contains_block;

	if (bootcache_check_fn && vp->v_mount && vp->v_mount->mnt_devvp) {
		if (VNOP_BLOCKMAP(vp, f_offset, PAGE_SIZE, &blkno, &io_size, NULL, VNODE_READ | VNODE_BLOCKMAP_NO_TRACK, NULL)) {
			return 0;
		}

		if (io_size == 0) {
			return 0;
		}

		if (bootcache_check_fn(vp->v_mount->mnt_devvp->v_rdev, blkno)) {
			return 1;
		}
	}
	return 0;
}


static int
cluster_is_throttled(vnode_t vp)
{
	return throttle_io_will_be_throttled(-1, vp->v_mount);
}


static void
cluster_iostate_wait(struct clios *iostate, u_int target, const char *wait_name)
{
	lck_mtx_lock(&iostate->io_mtxp);

	while ((iostate->io_issued - iostate->io_completed) > target) {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_START,
		    iostate->io_issued, iostate->io_completed, target, 0, 0);

		iostate->io_wanted = 1;
		msleep((caddr_t)&iostate->io_wanted, &iostate->io_mtxp, PRIBIO + 1, wait_name, NULL);

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_END,
		    iostate->io_issued, iostate->io_completed, target, 0, 0);
	}
	lck_mtx_unlock(&iostate->io_mtxp);
}


static void
cluster_handle_associated_upl(struct clios *iostate, upl_t upl,
    upl_offset_t upl_offset, upl_size_t size, off_t f_offset)
{
	if (!size) {
		return;
	}

	upl_t associated_upl = upl_associated_upl(upl);

	if (!associated_upl) {
		return;
	}

	/*
	 * The associated upl functions as a "range lock" for the file.
	 *
	 * The associated upl is created and is attached to to the upl in
	 * cluster_io when the direct io write is being started. Since the
	 * upl may be released in parts so the corresponding associated upl
	 * has to be released in parts as well.
	 *
	 * We have the f_offset, upl_offset and size and from that we have figure
	 * out the associated upl offset and length, we are interested in.
	 */
	upl_offset_t assoc_upl_offset, assoc_upl_end;

	/*                        ALIGNED UPL's                            */
	if ((upl_offset & PAGE_MASK) == (f_offset & PAGE_MASK)) {
		assoc_upl_offset = trunc_page_32(upl_offset);
		assoc_upl_end = round_page_32(upl_offset + size);
		goto do_commit;
	}

	/*
	 *                    HANDLE UNALIGNED UPLS
	 *
	 *  ( See also cluster_io where the associated upl is created )
	 *  While we create the upl in one go, we will be dumping the pages in
	 *  the upl in "transaction sized chunks" relative to the upl. Except
	 *  for the first transction, the upl_offset will always be page aligned.
	 *  and when the upl's are not aligned the associated upl offset will not
	 *  be page aligned and so we have to truncate and round up the starting
	 *  and the end of the pages in question and see if they are shared with
	 *  other transctions or not. If two transctions "share" a page in the
	 *  associated upl, the first one to complete "marks" it and skips that
	 *  page and the second  one will include it in the "commit range"
	 *
	 *  As an example, consider the case where 4 transctions are needed (this
	 *  is the worst case).
	 *
	 *  Transaction for 0-1 (size -> PAGE_SIZE - upl_offset)
	 *
	 *  This covers the associated upl from a -> c. a->b is not shared but
	 *  b-c is shared with the next transction so the first one to complete
	 *  will only "mark" it.
	 *
	 *  Transaction for 1-2 (size -> PAGE_SIZE)
	 *
	 *  For transaction 1, assoc_upl_offset would be 0 (corresponding to the
	 *  file offset a or b depending on what file offset the upl_offset
	 *  corrssponds to ) and assoc_upl_end would correspond to the file
	 *  offset c.
	 *
	 *                 (associated_upl - based on f_offset alignment)
	 *       0         a    b    c    d    e     f
	 *       <----|----|----|----|----|----|-----|---->
	 *
	 *
	 *                  (upl - based on user buffer address alignment)
	 *                   <__--|----|----|--__>
	 *
	 *                   0    1    2    3
	 *
	 */
	upl_size_t assoc_upl_size = upl_get_size(associated_upl);
#if 0
	/* knock off the simple case first -> this transaction covers the entire UPL */
	upl_offset_t upl_end = round_page_32(upl_offset + size);
	upl_size_t upl_size = vector_upl_get_size(upl);

	if ((trunc_page_32(upl_offset) == 0) && (upl_end == upl_size)) {
		assoc_upl_offset = 0;
		assoc_upl_end = assoc_upl_size;
		goto do_commit;
	}
#endif
	off_t assoc_upl_start_f_offset = upl_adjusted_offset(associated_upl, PAGE_MASK);

	assoc_upl_offset = (upl_offset_t)trunc_page_64(f_offset - assoc_upl_start_f_offset);
	assoc_upl_end = round_page_64(f_offset + size) - assoc_upl_start_f_offset;

	/*
	 * We can only sanity check the offset returned by upl_adjusted_offset
	 * for the first transaction for this UPL i.e. when (upl_offset < PAGE_SIZE)
	 */
	assertf((upl_offset >= PAGE_SIZE) || ((assoc_upl_start_f_offset == trunc_page_64(f_offset)) && (assoc_upl_offset == 0)),
	    "upl_offset = %d, f_offset = %lld, size = %d, start_f_offset = %lld,  assoc_upl_offset = %d",
	    upl_offset, f_offset, size, assoc_upl_start_f_offset, assoc_upl_offset);

	assertf((upl_offset == assoc_upl_offset) || (upl_offset > assoc_upl_offset && ((upl_offset - assoc_upl_offset) <= PAGE_SIZE)) ||
	    (assoc_upl_offset > upl_offset && ((assoc_upl_offset - upl_offset) <= PAGE_SIZE)),
	    "abs(upl_offset - assoc_upl_offset) >  PAGE_SIZE : "
	    "upl_offset = %d, f_offset = %lld, size = %d, start_f_offset = %lld, assoc_upl_offset = %d",
	    upl_offset, f_offset, size, assoc_upl_start_f_offset, assoc_upl_offset);

	assertf(assoc_upl_end <= assoc_upl_size,
	    "upl_offset = %d, f_offset = %lld, size = %d, start_f_offset = %lld, assoc_upl_size = %d, assoc_upl_offset = %d, assoc_upl_end = %d",
	    upl_offset, f_offset, size, assoc_upl_start_f_offset, assoc_upl_size, assoc_upl_offset, assoc_upl_end);

	assertf((assoc_upl_size > PAGE_SIZE) || (assoc_upl_offset == 0 && assoc_upl_end == PAGE_SIZE),
	    "upl_offset = %d, f_offset = %lld, size = %d, start_f_offset = %lld, assoc_upl_size = %d, assoc_upl_offset = %d, assoc_upl_end = %d",
	    upl_offset, f_offset, size, assoc_upl_start_f_offset, assoc_upl_size, assoc_upl_offset, assoc_upl_end);

	if (assoc_upl_size == PAGE_SIZE) {
		assoc_upl_offset = 0;
		assoc_upl_end = PAGE_SIZE;
		goto do_commit;
	}

	/*
	 * We have to check if the first and last pages of the associated UPL
	 * range could potentially be shared with other transactions and if the
	 * "sharing transactions" are both done. The first one sets the mark bit
	 * and the second one checks it and if set it includes that page in the
	 * pages to be "freed".
	 */
	bool check_first_pg = (assoc_upl_offset != 0) || ((f_offset + size) < (assoc_upl_start_f_offset + PAGE_SIZE));
	bool check_last_pg = (assoc_upl_end != assoc_upl_size) || (f_offset > ((assoc_upl_start_f_offset + assoc_upl_size) - PAGE_SIZE));

	if (check_first_pg || check_last_pg) {
		int first_pg = assoc_upl_offset >> PAGE_SHIFT;
		int last_pg = trunc_page_32(assoc_upl_end - 1) >> PAGE_SHIFT;
		upl_page_info_t *assoc_pl = UPL_GET_INTERNAL_PAGE_LIST(associated_upl);

		lck_mtx_lock_spin(&iostate->io_mtxp);
		if (check_first_pg && !upl_page_get_mark(assoc_pl, first_pg)) {
			/*
			 * The first page isn't marked so let another transaction
			 * completion handle it.
			 */
			upl_page_set_mark(assoc_pl, first_pg, true);
			assoc_upl_offset += PAGE_SIZE;
		}
		if (check_last_pg && !upl_page_get_mark(assoc_pl, last_pg)) {
			/*
			 * The last page isn't marked so mark the page and let another
			 * transaction completion handle it.
			 */
			upl_page_set_mark(assoc_pl, last_pg, true);
			assoc_upl_end -= PAGE_SIZE;
		}
		lck_mtx_unlock(&iostate->io_mtxp);
	}

	if (assoc_upl_end <= assoc_upl_offset) {
		return;
	}

do_commit:
	size = assoc_upl_end - assoc_upl_offset;

	boolean_t empty;

	/*
	 * We can unlock these pages now and as this is for a
	 * direct/uncached write, we want to dump the pages too.
	 */
	kern_return_t kr = upl_abort_range(associated_upl, assoc_upl_offset, size,
	    UPL_ABORT_DUMP_PAGES, &empty);

	assert(!kr);

	if (!kr && empty) {
		upl_set_associated_upl(upl, NULL);
		upl_deallocate(associated_upl);
	}
}

static void
cluster_iodone_verify_continue(void)
{
	lck_mtx_lock_spin(&cl_transaction_mtxp);
	for (;;) {
		struct verify_buf *vb = TAILQ_FIRST(&verify_work_head);

		if (!vb) {
			assert_wait(&verify_work_head, (THREAD_UNINT));
			break;
		}
		buf_t cbp = vb->vb_cbp;
		void* callback_arg = vb->vb_callback_arg;

		TAILQ_REMOVE(&verify_work_head, vb, vb_entry);
		vb->vb_cbp = NULL;
		vb->vb_callback_arg = NULL;
		vb->vb_whichq = 0;
		TAILQ_INSERT_TAIL(&verify_free_head, vb, vb_entry);
		lck_mtx_unlock(&cl_transaction_mtxp);

		(void)cluster_iodone_finish(cbp, callback_arg);
		cbp = NULL;
		lck_mtx_lock_spin(&cl_transaction_mtxp);
	}
	lck_mtx_unlock(&cl_transaction_mtxp);
	thread_block((thread_continue_t)cluster_iodone_verify_continue);
	/* NOT REACHED */
}

static void
cluster_verify_thread(void)
{
	thread_t self = current_thread();

	thread_set_thread_name(self, "cluster_verify_thread");
#if __AMP__
	if (ecore_verify_threads) {
		kern_return_t kr = thread_soft_bind_cluster_type(self, 'E');
		if (kr != KERN_SUCCESS) {
			printf("%s: WARN: failed to bind thread to cluster type; does the hardware topology match expectations?\n", __FUNCTION__);
		}
	}
#endif /* __AMP__ */
#if !defined(__x86_64__)
	thread_group_join_io_storage();
#endif /* __x86_64__ */
	cluster_iodone_verify_continue();
	/* NOT REACHED */
}

static bool
enqueue_buf_for_verify(buf_t cbp, void *callback_arg)
{
	struct verify_buf *vb;

	vb = TAILQ_FIRST(&verify_free_head);
	if (vb) {
		TAILQ_REMOVE(&verify_free_head, vb, vb_entry);
		vb->vb_cbp = cbp;
		vb->vb_callback_arg = callback_arg;
		vb->vb_whichq = 1;
		TAILQ_INSERT_TAIL(&verify_work_head, vb, vb_entry);
		return true;
	} else {
		return false;
	}
}

static int
cluster_handle_verification(buf_t cbp_head, vnode_t vp, upl_t upl, int upl_offset, int transaction_size, int error)
{
	off_t start_off = cbp_head->b_clfoffset;
	void *verify_ctx = cbp_head->b_attr.ba_un.verify_ctx;
	caddr_t verify_buf = NULL;
	uint32_t verify_length = transaction_size;
	vnode_verify_flags_t verify_flags = VNODE_VERIFY_CONTEXT_FREE;
	int verify_error = EAGAIN;

	assert(cbp_head->b_attr.ba_flags & BA_WILL_VERIFY);

	cbp_head->b_attr.ba_un.verify_ctx = NULL;
	if (error) {
		goto free_context;
	}

	/*
	 * If we don't have a precomputed hash, we make a single call to both
	 * verify and free the context. If we have a precomputed hash, then we
	 * make two separate calls - one to verify the hash and the second one to
	 * free. If the filesystem returns EAGAIN we fall back to the non
	 * precomputed hash case.
	 */
	if (cbp_head->b_attr.ba_verify_type && cbp_head->b_attr.ba_flags & BA_VERIFY_VALID) {
		verify_buf = (caddr_t)buf_verifyptr_with_size(cbp_head, transaction_size, &verify_length);
		verify_flags = VNODE_VERIFY_WITH_CONTEXT | VNODE_VERIFY_PRECOMPUTED;

		if (verify_buf && verify_length) {
			verify_error = VNOP_VERIFY(vp, start_off, (uint8_t *)verify_buf, verify_length,
			    NULL, &verify_ctx, verify_flags, NULL, NULL);
		} else {
			verify_error = EAGAIN;
		}

		verify_buf = NULL;
		verify_length = transaction_size;
		verify_flags = VNODE_VERIFY_CONTEXT_FREE;
	}

	if (verify_error != EAGAIN) {
		error = verify_error;
	} else {
		vm_offset_t vaddr;

		/*
		 * Map it in.
		 *
		 * ubc_upl_map_range unfortunately cannot handle concurrent map
		 * requests for the same UPL and returns failures when it can't
		 * map. The map exclusive mechanism enforces mutual exclusion
		 * for concurrent requests.
		 */
		verify_error = 0;
		os_atomic_inc(&verify_in_flight, relaxed);
		upl_set_map_exclusive(upl);
		error = ubc_upl_map_range(upl, upl_offset, round_page(transaction_size), VM_PROT_DEFAULT, &vaddr);
		if (error) {
			upl_clear_map_exclusive(upl);
			printf("ubc_upl_map_range returned error %d upl = %p, upl_offset = %d, size = %d",
			    error, upl, (int)upl_offset, (int)round_page(transaction_size));
			error  = EIO;
			if (os_atomic_dec_orig(&verify_in_flight, relaxed) == 0) {
				panic("verify_in_flight underflow");
			}
		} else {
			verify_buf = (caddr_t)vaddr;
			verify_flags |= VNODE_VERIFY_WITH_CONTEXT;
		}
	}

free_context:
	verify_error = VNOP_VERIFY(vp, start_off, (uint8_t *)verify_buf, verify_length,
	    NULL, &verify_ctx, verify_flags, NULL, NULL);
	if (!error) {
		error = verify_error;
	}

	if (verify_buf) {
		(void)ubc_upl_unmap_range(upl, upl_offset, round_page(transaction_size));
		upl_clear_map_exclusive(upl);
		verify_buf = NULL;
		if (os_atomic_dec_orig(&verify_in_flight, relaxed) == 0) {
			panic("verify_in_flight underflow");
		}
	}

	return error;
}

static int
cluster_ioerror(upl_t upl, int upl_offset, int abort_size, int error, int io_flags, vnode_t vp)
{
	int upl_abort_code = 0;
	int page_in  = 0;
	int page_out = 0;

	if ((io_flags & (B_PHYS | B_CACHE)) == (B_PHYS | B_CACHE)) {
		/*
		 * direct write of any flavor, or a direct read that wasn't aligned
		 */
		ubc_upl_commit_range(upl, upl_offset, abort_size, UPL_COMMIT_FREE_ON_EMPTY);
	} else {
		if (io_flags & B_PAGEIO) {
			if (io_flags & B_READ) {
				page_in  = 1;
			} else {
				page_out = 1;
			}
		}
		if (io_flags & B_CACHE) {
			/*
			 * leave pages in the cache unchanged on error
			 */
			upl_abort_code = UPL_ABORT_FREE_ON_EMPTY;
		} else if (((io_flags & B_READ) == 0) && ((error != ENXIO) || vnode_isswap(vp))) {
			/*
			 * transient error on pageout/write path... leave pages unchanged
			 */
			upl_abort_code = UPL_ABORT_FREE_ON_EMPTY;
		} else if (page_in) {
			upl_abort_code = UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_ERROR;
		} else {
			upl_abort_code = UPL_ABORT_FREE_ON_EMPTY /* | UPL_ABORT_DUMP_PAGES */;
		}

		ubc_upl_abort_range(upl, upl_offset, abort_size, upl_abort_code);
	}
	return upl_abort_code;
}


static int
cluster_iodone(buf_t bp, void *callback_arg)
{
	buf_t   cbp;
	buf_t   cbp_head;
	int     error = 0;
	boolean_t       transaction_complete = FALSE;
	bool async;

	__IGNORE_WCASTALIGN(cbp_head = (buf_t)(bp->b_trans_head));

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_START,
	    cbp_head, bp->b_lblkno, bp->b_bcount, bp->b_flags, 0);

	async = cluster_verify_threads &&
	    (os_atomic_load(&cbp_head->b_attr.ba_flags, acquire) & BA_ASYNC_VERIFY);

	assert(!async || cbp_head->b_attr.ba_un.verify_ctx);

	if (cbp_head->b_trans_next || !(cbp_head->b_flags & B_EOT)) {
		lck_mtx_lock_spin(&cl_transaction_mtxp);

		bp->b_flags |= B_TDONE;

		for (cbp = cbp_head; cbp; cbp = cbp->b_trans_next) {
			/*
			 * all I/O requests that are part of this transaction
			 * have to complete before we can process it
			 */
			if (!(cbp->b_flags & B_TDONE)) {
				KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END,
				    cbp_head, cbp, cbp->b_bcount, cbp->b_flags, 0);

				lck_mtx_unlock(&cl_transaction_mtxp);

				return 0;
			}

			if (cbp->b_trans_next == CLUSTER_IO_WAITING) {
				KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END,
				    cbp_head, cbp, cbp->b_bcount, cbp->b_flags, 0);

				lck_mtx_unlock(&cl_transaction_mtxp);
				wakeup(cbp);

				return 0;
			}

			if (cbp->b_flags & B_EOT) {
				transaction_complete = TRUE;

				if (async) {
					async = enqueue_buf_for_verify(cbp_head, callback_arg);
				}
			}
		}
		lck_mtx_unlock(&cl_transaction_mtxp);

		if (transaction_complete == FALSE) {
			KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END,
			    cbp_head, 0, 0, 0, 0);
			return 0;
		}
	} else if (async) {
		lck_mtx_lock_spin(&cl_transaction_mtxp);
		async = enqueue_buf_for_verify(cbp_head, callback_arg);
		lck_mtx_unlock(&cl_transaction_mtxp);
	}

	if (async) {
		wakeup(&verify_work_head);
	} else {
		error = cluster_iodone_finish(cbp_head, callback_arg);
	}

	return error;
}

static int
cluster_iodone_finish(buf_t cbp_head, void *callback_arg)
{
	int     b_flags;
	int     error;
	int     total_size;
	int     total_resid;
	int     upl_offset;
	int     zero_offset;
	int     pg_offset = 0;
	int     commit_size = 0;
	int     upl_flags = 0;
	int     transaction_size = 0;
	upl_t   upl;
	buf_t   cbp;
	buf_t   cbp_next;
	buf_t   real_bp;
	vnode_t vp;
	struct  clios *iostate;

	error       = 0;
	total_size  = 0;
	total_resid = 0;

	cbp        = cbp_head;
	vp         = cbp->b_vp;
	upl_offset = cbp->b_uploffset;
	upl        = cbp->b_upl;
	b_flags    = cbp->b_flags;
	real_bp    = cbp->b_real_bp;
	zero_offset = cbp->b_validend;
	iostate    = (struct clios *)cbp->b_iostate;

	if (real_bp) {
		real_bp->b_dev = cbp->b_dev;
	}

	while (cbp) {
		if ((cbp->b_flags & B_ERROR) && error == 0) {
			error = cbp->b_error;
		}

		total_resid += cbp->b_resid;
		total_size  += cbp->b_bcount;

		cbp_next = cbp->b_trans_next;

		if (cbp_next == NULL) {
			/*
			 * compute the overall size of the transaction
			 * in case we created one that has 'holes' in it
			 * 'total_size' represents the amount of I/O we
			 * did, not the span of the transaction w/r to the UPL
			 */
			transaction_size = cbp->b_uploffset + cbp->b_bcount - upl_offset;
		}

		cbp = cbp_next;
	}

	if (ISSET(b_flags, B_COMMIT_UPL)) {
		cluster_handle_associated_upl(iostate,
		    cbp_head->b_upl,
		    upl_offset,
		    transaction_size,
		    cbp_head->b_clfoffset);
	}

	if (error == 0 && total_resid) {
		error = EIO;
	}

	if (error == 0) {
		int     (*cliodone_func)(buf_t, void *) = (int (*)(buf_t, void *))(cbp_head->b_cliodone);

		if (cliodone_func != NULL) {
			cbp_head->b_bcount = transaction_size;

			error = (*cliodone_func)(cbp_head, callback_arg);
		}
	}
	if (zero_offset) {
		cluster_zero(upl, zero_offset, PAGE_SIZE - (zero_offset & PAGE_MASK), real_bp);
	}

	if (cbp_head->b_attr.ba_un.verify_ctx) {
		error = cluster_handle_verification(cbp_head, vp, upl, upl_offset, transaction_size, error);
	} else if (cbp_head->b_attr.ba_flags & BA_WILL_VERIFY) {
		error = EBADMSG;
	}

	if (iostate) {
		int need_wakeup = 0;

		/*
		 * someone has issued multiple I/Os asynchrounsly
		 * and is waiting for them to complete (streaming)
		 */
		lck_mtx_lock_spin(&iostate->io_mtxp);

		if (error && iostate->io_error == 0) {
			iostate->io_error = error;
		}

		iostate->io_completed += total_size;

		if (iostate->io_wanted) {
			/*
			 * someone is waiting for the state of
			 * this io stream to change
			 */
			iostate->io_wanted = 0;
			need_wakeup = 1;
		}
		lck_mtx_unlock(&iostate->io_mtxp);

		if (need_wakeup) {
			wakeup((caddr_t)&iostate->io_wanted);
		}
	}

	if (b_flags & B_COMMIT_UPL) {
		pg_offset   = upl_offset & PAGE_MASK;
		commit_size = (pg_offset + transaction_size + (PAGE_SIZE - 1)) & ~PAGE_MASK;

		if (error) {
			upl_set_iodone_error(upl, error);

			upl_flags = cluster_ioerror(upl, upl_offset - pg_offset, commit_size, error, b_flags, vp);
		} else {
			upl_flags = UPL_COMMIT_FREE_ON_EMPTY;

			if ((b_flags & B_PHYS) && (b_flags & B_READ)) {
				upl_flags |= UPL_COMMIT_SET_DIRTY;
			}

			if (b_flags & B_AGE) {
				upl_flags |= UPL_COMMIT_INACTIVATE;
			}

			ubc_upl_commit_range(upl, upl_offset - pg_offset, commit_size, upl_flags);
		}
	}

	cbp = cbp_head->b_trans_next;
	while (cbp) {
		cbp_next = cbp->b_trans_next;

		if (cbp != cbp_head) {
			free_io_buf(cbp);
		}

		cbp = cbp_next;
	}
	free_io_buf(cbp_head);

	if (real_bp) {
		if (error) {
			real_bp->b_flags |= B_ERROR;
			real_bp->b_error = error;
		}
		real_bp->b_resid = total_resid;

		buf_biodone(real_bp);
	}
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END,
	    upl, upl_offset - pg_offset, commit_size, (error << 24) | upl_flags, 0);

	return error;
}


uint32_t
cluster_throttle_io_limit(vnode_t vp, uint32_t *limit)
{
	if (cluster_is_throttled(vp)) {
		*limit = calculate_max_throttle_size(vp);
		return 1;
	}
	return 0;
}


void
cluster_zero(upl_t upl, upl_offset_t upl_offset, int size, buf_t bp)
{
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 23)) | DBG_FUNC_START,
	    upl_offset, size, bp, 0, 0);

	if (bp == NULL || bp->b_datap == 0) {
		upl_page_info_t *pl;
		addr64_t        zero_addr;

		pl = ubc_upl_pageinfo(upl);

		if (upl_device_page(pl) == TRUE) {
			zero_addr = ((addr64_t)upl_phys_page(pl, 0) << PAGE_SHIFT) + upl_offset;

			bzero_phys_nc(zero_addr, size);
		} else {
			while (size) {
				int     page_offset;
				int     page_index;
				int     zero_cnt;

				page_index  = upl_offset / PAGE_SIZE;
				page_offset = upl_offset & PAGE_MASK;

				zero_addr = ((addr64_t)upl_phys_page(pl, page_index) << PAGE_SHIFT) + page_offset;
				zero_cnt  = min(PAGE_SIZE - page_offset, size);

				bzero_phys(zero_addr, zero_cnt);

				size       -= zero_cnt;
				upl_offset += zero_cnt;
			}
		}
	} else {
		bzero((caddr_t)((vm_offset_t)bp->b_datap + upl_offset), size);
	}

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 23)) | DBG_FUNC_END,
	    upl_offset, size, 0, 0, 0);
}


static void
cluster_EOT(buf_t cbp_head, buf_t cbp_tail, int zero_offset, size_t verify_block_size)
{
	/*
	 * We will assign a verification context to cbp_head.
	 * This will be passed back to the filesystem  when
	 * verifying (in cluster_iodone).
	 */
	if (verify_block_size) {
		off_t start_off = cbp_head->b_clfoffset;
		size_t length;
		void *verify_ctx = NULL;
		int error = 0;
		vnode_t vp = buf_vnode(cbp_head);

		if (cbp_head == cbp_tail) {
			length = cbp_head->b_bcount;
		} else {
			length = (cbp_tail->b_clfoffset + cbp_tail->b_bcount) - start_off;
		}

		/*
		 * zero_offset is non zero for the transaction containing the EOF
		 * (if the filesize is not page aligned). In that case we might
		 * have the transaction size not be page/verify block size aligned
		 */
		if ((zero_offset == 0) &&
		    ((length < verify_block_size) || (length % verify_block_size)) != 0) {
			panic("%s length = %zu, verify_block_size = %zu",
			    __FUNCTION__, length, verify_block_size);
		}

		error = VNOP_VERIFY(vp, start_off, NULL, length,
		    &verify_block_size, &verify_ctx, VNODE_VERIFY_CONTEXT_ALLOC, NULL, NULL);

		assert(!(error && verify_ctx));

		if (verify_ctx) {
			if (num_verify_threads && (os_atomic_load(&cluster_verify_threads, relaxed) == 0)) {
				if (os_atomic_inc_orig(&cluster_verify_threads, relaxed) == 0) {
					thread_t thread;
					int i;

					for (i = 0; i < num_verify_threads && i < MAX_VERIFY_THREADS; i++) {
						kernel_thread_start((thread_continue_t)cluster_verify_thread, NULL, &thread);
						thread_deallocate(thread);
					}
					os_atomic_store(&cluster_verify_threads, i, relaxed);
				} else {
					os_atomic_dec(&cluster_verify_threads, relaxed);
				}
			}
			cbp_head->b_attr.ba_un.verify_ctx = verify_ctx;
			/*
			 * At least one thread is busy (at the time we
			 * checked), so we can let it get queued for
			 * async processing. It's fine if we occasionally get
			 * this wrong.
			 */
			if (os_atomic_load(&verify_in_flight, relaxed)) {
				/* This flag and the setting of ba_un.verify_ctx needs to be ordered */
				os_atomic_or(&cbp_head->b_attr.ba_flags, BA_ASYNC_VERIFY, release);
			}
		}
	} else {
		cbp_head->b_attr.ba_un.verify_ctx = NULL;
	}

	cbp_head->b_validend = zero_offset;
	cbp_tail->b_flags |= B_EOT;
}

static void
cluster_wait_IO(buf_t cbp_head, int async)
{
	buf_t   cbp;

	if (async) {
		/*
		 * Async callback completion will not normally generate a
		 * wakeup upon I/O completion.  To get woken up, we set
		 * b_trans_next (which is safe for us to modify) on the last
		 * buffer to CLUSTER_IO_WAITING so that cluster_iodone knows
		 * to wake us up when all buffers as part of this transaction
		 * are completed.  This is done under the umbrella of
		 * cl_transaction_mtxp which is also taken in cluster_iodone.
		 */
		bool done = true;
		buf_t last = NULL;

		lck_mtx_lock_spin(&cl_transaction_mtxp);

		for (cbp = cbp_head; cbp; last = cbp, cbp = cbp->b_trans_next) {
			if (!ISSET(cbp->b_flags, B_TDONE)) {
				done = false;
			}
		}

		if (!done) {
			last->b_trans_next = CLUSTER_IO_WAITING;

			DTRACE_IO1(wait__start, buf_t, last);
			do {
				msleep(last, &cl_transaction_mtxp, PSPIN | (PRIBIO + 1), "cluster_wait_IO", NULL);

				/*
				 * We should only have been woken up if all the
				 * buffers are completed, but just in case...
				 */
				done = true;
				for (cbp = cbp_head; cbp != CLUSTER_IO_WAITING; cbp = cbp->b_trans_next) {
					if (!ISSET(cbp->b_flags, B_TDONE)) {
						done = false;
						break;
					}
				}
			} while (!done);
			DTRACE_IO1(wait__done, buf_t, last);

			last->b_trans_next = NULL;
		}

		lck_mtx_unlock(&cl_transaction_mtxp);
	} else { // !async
		for (cbp = cbp_head; cbp; cbp = cbp->b_trans_next) {
			buf_biowait(cbp);
		}
	}
}

static void
cluster_complete_transaction(buf_t *cbp_head, void *callback_arg, int *retval, int flags, int needwait)
{
	buf_t   cbp;
	int     error;
	boolean_t isswapout = FALSE;

	/*
	 * cluster_complete_transaction will
	 * only be called if we've issued a complete chain in synchronous mode
	 * or, we've already done a cluster_wait_IO on an incomplete chain
	 */
	if (needwait) {
		for (cbp = *cbp_head; cbp; cbp = cbp->b_trans_next) {
			buf_biowait(cbp);
		}
	}
	/*
	 * we've already waited on all of the I/Os in this transaction,
	 * so mark all of the buf_t's in this transaction as B_TDONE
	 * so that cluster_iodone sees the transaction as completed
	 */
	for (cbp = *cbp_head; cbp; cbp = cbp->b_trans_next) {
		cbp->b_flags |= B_TDONE;
		cbp->b_attr.ba_flags &= ~BA_ASYNC_VERIFY;
	}
	cbp = *cbp_head;

	if ((flags & (CL_ASYNC | CL_PAGEOUT)) == CL_PAGEOUT && vnode_isswap(cbp->b_vp)) {
		isswapout = TRUE;
	}

	error = cluster_iodone(cbp, callback_arg);

	if (!(flags & CL_ASYNC) && error && *retval == 0) {
		if (((flags & (CL_PAGEOUT | CL_KEEPCACHED)) != CL_PAGEOUT) || (error != ENXIO)) {
			*retval = error;
		} else if (isswapout == TRUE) {
			*retval = error;
		}
	}
	*cbp_head = (buf_t)NULL;
}

uint64_t cluster_direct_write_wired = 0;

static int
cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int non_rounded_size,
    int flags, buf_t real_bp, struct clios *iostate, int (*callback)(buf_t, void *), void *callback_arg)
{
	buf_t   cbp;
	u_int   size;
	u_int   io_size;
	int     io_flags;
	int     bmap_flags;
	int     error = 0;
	int     retval = 0;
	buf_t   cbp_head = NULL;
	buf_t   cbp_tail = NULL;
	int     trans_count = 0;
	int     max_trans_count;
	u_int   pg_count;
	int     pg_offset;
	u_int   max_iosize;
	u_int   max_vectors;
	int     priv;
	int     zero_offset = 0;
	int     async_throttle = 0;
	mount_t mp;
	size_t verify_block_size = 0;
	vm_offset_t upl_end_offset;
	vnode_verify_kind_t verify_kind = VK_HASH_NONE;
	boolean_t   need_EOT = FALSE;

	/*
	 * we currently don't support buffers larger than a page
	 */
	if (real_bp && non_rounded_size > PAGE_SIZE) {
		panic("%s(): Called with real buffer of size %d bytes which "
		    "is greater than the maximum allowed size of "
		    "%d bytes (the system PAGE_SIZE).\n",
		    __FUNCTION__, non_rounded_size, PAGE_SIZE);
	}

	mp = vp->v_mount;

	/*
	 * we don't want to do any funny rounding of the size for IO requests
	 * coming through the DIRECT or CONTIGUOUS paths...  those pages don't
	 * belong to us... we can't extend (nor do we need to) the I/O to fill
	 * out a page
	 */
	if (mp->mnt_devblocksize > 1 && !(flags & (CL_DEV_MEMORY | CL_DIRECT_IO))) {
		/*
		 * round the requested size up so that this I/O ends on a
		 * page boundary in case this is a 'write'... if the filesystem
		 * has blocks allocated to back the page beyond the EOF, we want to
		 * make sure to write out the zero's that are sitting beyond the EOF
		 * so that in case the filesystem doesn't explicitly zero this area
		 * if a hole is created via a lseek/write beyond the current EOF,
		 * it will return zeros when it's read back from the disk.  If the
		 * physical allocation doesn't extend for the whole page, we'll
		 * only write/read from the disk up to the end of this allocation
		 * via the extent info returned from the VNOP_BLOCKMAP call.
		 */
		pg_offset = upl_offset & PAGE_MASK;

		size = (((non_rounded_size + pg_offset) + (PAGE_SIZE - 1)) & ~PAGE_MASK) - pg_offset;
	} else {
		/*
		 * anyone advertising a blocksize of 1 byte probably
		 * can't deal with us rounding up the request size
		 * AFP is one such filesystem/device
		 */
		size = non_rounded_size;
	}
	upl_end_offset = upl_offset + size;

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 22)) | DBG_FUNC_START, (int)f_offset, size, upl_offset, flags, 0);

	/*
	 * Set the maximum transaction size to the maximum desired number of
	 * buffers.
	 */
	max_trans_count = 8;
	if (flags & CL_DEV_MEMORY) {
		max_trans_count = 16;
	}

	if (flags & CL_READ) {
		io_flags = B_READ;
		bmap_flags = VNODE_READ;

		max_iosize  = mp->mnt_maxreadcnt;
		max_vectors = mp->mnt_segreadcnt;

		/* See if we can do cluster verification (pageins and aligned reads) */
		if ((flags & CL_PAGEIN || cluster_verify_threads) &&
		    !(mp->mnt_kern_flag & MNTK_VIRTUALDEV) &&
		    (VNOP_VERIFY(vp, f_offset, NULL, 0, &verify_block_size, NULL, VNODE_VERIFY_DEFAULT, NULL, &verify_kind) == 0) &&
		    verify_block_size) {
			if (verify_block_size != PAGE_SIZE) {
				verify_block_size = 0;
			}
			if (real_bp && verify_block_size) {
				panic("%s(): Called with real buffer and needs verification ",
				    __FUNCTION__);
			}
			/*
			 * For reads, only allow cluster verification if f_offset
			 * and upl_offset are both page aligned. Additionally, for direct reads,
			 * require that the length of the write also be page aligned.
			 * If they are not  page aligned, leave it to the filesystem to do verification.
			 * Strictly speaking, the alignments need to be for verify_block_size
			 * but since the only verify_block_size that is currently supported
			 * is page size, we check against page alignment.
			 */
			if (verify_block_size && !(flags & CL_PAGEIN) &&
			    ((f_offset & PAGE_MASK) || (upl_offset & PAGE_MASK) ||
			    ((flags & CL_DIRECT_IO) && (non_rounded_size & PAGE_MASK)))) {
				verify_block_size = 0;
				verify_kind = VK_HASH_NONE;
			}
			if (verify_block_size && verify_kind && !upl_has_fs_verify_info(upl)) {
				upl_set_fs_verify_info(upl,
				    (upl_adjusted_size(upl, PAGE_MASK) / mp->mnt_devblocksize) * get_num_bytes_for_verify_kind(verify_kind));
			}
		}
	} else {
		io_flags = B_WRITE;
		bmap_flags = VNODE_WRITE;

		max_iosize  = mp->mnt_maxwritecnt;
		max_vectors = mp->mnt_segwritecnt;
	}
	if (verify_block_size) {
		bmap_flags |= VNODE_CLUSTER_VERIFY;
	}
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 22)) | DBG_FUNC_NONE, max_iosize, max_vectors, mp->mnt_devblocksize, 0, 0);

	/*
	 * make sure the maximum iosize is a
	 * multiple of the page size
	 */
	max_iosize  &= ~PAGE_MASK;

	/*
	 * Ensure the maximum iosize is sensible.
	 */
	if (!max_iosize) {
		max_iosize = PAGE_SIZE;
	}

	if (flags & CL_THROTTLE) {
		if (!(flags & CL_PAGEOUT) && cluster_is_throttled(vp)) {
			uint32_t max_throttle_size = calculate_max_throttle_size(vp);

			if (max_iosize > max_throttle_size) {
				max_iosize = max_throttle_size;
			}
			async_throttle = calculate_max_throttle_cnt(vp);
		} else {
			if ((flags & CL_DEV_MEMORY)) {
				async_throttle = IO_SCALE(vp, VNODE_ASYNC_THROTTLE);
			} else {
				u_int max_cluster;
				u_int max_cluster_size;
				u_int scale;

				if (vp->v_mount->mnt_minsaturationbytecount) {
					max_cluster_size = vp->v_mount->mnt_minsaturationbytecount;

					scale = 1;
				} else {
					max_cluster_size = MAX_CLUSTER_SIZE(vp);

					if (disk_conditioner_mount_is_ssd(vp->v_mount)) {
						scale = WRITE_THROTTLE_SSD;
					} else {
						scale = WRITE_THROTTLE;
					}
				}
				if (max_iosize > max_cluster_size) {
					max_cluster = max_cluster_size;
				} else {
					max_cluster = max_iosize;
				}

				if (size < max_cluster) {
					max_cluster = size;
				}

				if (flags & CL_CLOSE) {
					scale += MAX_CLUSTERS;
				}

				async_throttle = min(IO_SCALE(vp, VNODE_ASYNC_THROTTLE), ((scale * max_cluster_size) / max_cluster) - 1);
			}
		}
	}
	if (flags & CL_AGE) {
		io_flags |= B_AGE;
	}
	if (flags & (CL_PAGEIN | CL_PAGEOUT)) {
		io_flags |= B_PAGEIO;
	}
	if (flags & (CL_IOSTREAMING)) {
		io_flags |= B_IOSTREAMING;
	}
	if (flags & CL_COMMIT) {
		io_flags |= B_COMMIT_UPL;
	}
	if (flags & CL_DIRECT_IO) {
		io_flags |= B_PHYS;
	}
	if (flags & (CL_PRESERVE | CL_KEEPCACHED)) {
		io_flags |= B_CACHE;
	}
	if (flags & CL_PASSIVE) {
		io_flags |= B_PASSIVE;
	}
	if (flags & CL_ENCRYPTED) {
		io_flags |= B_ENCRYPTED_IO;
	}

	if (vp->v_flag & VSYSTEM) {
		io_flags |= B_META;
	}

	if ((flags & CL_READ) && ((upl_offset + non_rounded_size) & PAGE_MASK) && (!(flags & CL_NOZERO))) {
		/*
		 * then we are going to end up
		 * with a page that we can't complete (the file size wasn't a multiple
		 * of PAGE_SIZE and we're trying to read to the end of the file
		 * so we'll go ahead and zero out the portion of the page we can't
		 * read in from the file
		 */
		zero_offset = (int)(upl_offset + non_rounded_size);
	} else if (!ISSET(flags, CL_READ) && ISSET(flags, CL_DIRECT_IO)) {
		assert(ISSET(flags, CL_COMMIT));

		// For a direct/uncached write, we need to lock pages...
		upl_t cached_upl = NULL;
		upl_page_info_t *cached_pl;

		assert(upl_offset < PAGE_SIZE);

		/*
		 *
		 *                       f_offset = b
		 *                      upl_offset = 8K
		 *
		 *                       (cached_upl - based on f_offset alignment)
		 *       0         a    b              c
		 *       <----|----|----|----|----|----|-----|---->
		 *
		 *
		 *                          (upl - based on user buffer address alignment)
		 *                   <__--|----|----|--__>
		 *
		 *                   0    1x   2x  3x
		 *
		 */
		const off_t cached_upl_f_offset = trunc_page_64(f_offset);
		const int cached_upl_size = round_page_32((f_offset - cached_upl_f_offset) + non_rounded_size);
		int num_retries = 0;

		/*
		 * Create a UPL to lock the pages in the cache whilst the
		 * write is in progress.
		 */
create_cached_upl:
		ubc_create_upl_kernel(vp, cached_upl_f_offset, cached_upl_size, &cached_upl,
		    &cached_pl, UPL_SET_LITE | UPL_WILL_MODIFY, VM_KERN_MEMORY_FILE);
		if (cached_upl && upl_has_wired_pages(cached_upl)) {
			/*
			 * Pages in this UPL would contain stale data after our direct write
			 * (which is intended to overwrite these pages on disk).  The UPL is
			 * just holding these pages "busy" to synchronize with any other I/O
			 * or mmap() access and we have to dump these pages when the direct
			 * write is done.
			 * But we can't do that for wired pages, so let's release this UPL
			 * and fall back to the "cached" path.
			 */
//			printf("*******  FBDP %s:%d vp %p offset 0x%llx size 0x%llx - switching from direct to cached write\n", __FUNCTION__, __LINE__, vp, cached_upl_f_offset, (uint64_t)cached_upl_size);
			ubc_upl_abort_range(cached_upl, 0, cached_upl_size, UPL_ABORT_FREE_ON_EMPTY);
			cached_upl = NULL;
			cached_pl = NULL;
			cluster_direct_write_wired++;
			return ENOTSUP;
		}

		/*
		 * If we are not overwriting the first and last pages completely
		 * we need to write them out first if they are dirty. These pages
		 * will be discarded after the write completes so we might lose
		 * the writes for the parts that are not overwrrtten.
		 */
		bool first_page_needs_sync = false;
		bool last_page_needs_sync = false;

		if (cached_upl && (cached_upl_f_offset < f_offset) && upl_dirty_page(cached_pl, 0)) {
			first_page_needs_sync = true;
		}

		if (cached_upl && (cached_upl_f_offset + cached_upl_size) > (f_offset + non_rounded_size)) {
			int last_page = (cached_upl_size / PAGE_SIZE) - 1;

			if ((last_page != 0 || !first_page_needs_sync) && upl_dirty_page(cached_pl, last_page)) {
				last_page_needs_sync = true;
			}
		}

		if (first_page_needs_sync || last_page_needs_sync) {
			ubc_upl_abort_range(cached_upl, 0, cached_upl_size, UPL_ABORT_FREE_ON_EMPTY);
			cached_upl = NULL;
			cached_pl = NULL;
			if (first_page_needs_sync) {
				ubc_msync(vp, cached_upl_f_offset, cached_upl_f_offset + PAGE_SIZE, NULL, UBC_PUSHALL | UBC_INVALIDATE | UBC_SYNC);
			}
			if (last_page_needs_sync) {
				off_t cached_upl_end_offset = cached_upl_f_offset + cached_upl_size;

				ubc_msync(vp, cached_upl_end_offset - PAGE_SIZE, cached_upl_end_offset, NULL, UBC_PUSHALL | UBC_INVALIDATE | UBC_SYNC);
			}
			if (++num_retries < 16) {
				goto create_cached_upl;
			}
			printf("%s : Number of retries for syncing first or last page reached %d\n", __FUNCTION__, num_retries);
			assertf(num_retries < 16, "%s : Number of retries for syncing first or last page reached %d\n", __FUNCTION__, num_retries);
		}

		/*
		 * Attach this UPL to the other UPL so that we can find it
		 * later.
		 */
		upl_set_associated_upl(upl, cached_upl);
		assertf(!cached_upl ||
		    (upl_adjusted_offset(cached_upl, PAGE_MASK) == cached_upl_f_offset),
		    "upl_adjusted_offset(cached_upl, PAGE_MASK) = %lld, cached_upl_f_offset = %lld",
		    upl_adjusted_offset(cached_upl, PAGE_MASK), cached_upl_f_offset);
	}

	while (size) {
		daddr64_t blkno;
		daddr64_t lblkno;
		size_t  io_size_tmp;
		u_int   io_size_wanted;

		if (size > max_iosize) {
			io_size = max_iosize;
		} else {
			io_size = size;
		}

		io_size_wanted = io_size;
		io_size_tmp = (size_t)io_size;

		if ((error = VNOP_BLOCKMAP(vp, f_offset, io_size, &blkno, &io_size_tmp, NULL, bmap_flags, NULL))) {
			break;
		}

		if (io_size_tmp > io_size_wanted) {
			io_size = io_size_wanted;
		} else {
			io_size = (u_int)io_size_tmp;
		}

		if (real_bp && (real_bp->b_blkno == real_bp->b_lblkno)) {
			real_bp->b_blkno = blkno;
		}

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 24)) | DBG_FUNC_NONE,
		    (int)f_offset, (int)(blkno >> 32), (int)blkno, io_size, 0);

		if (io_size == 0) {
			/*
			 * vnop_blockmap didn't return an error... however, it did
			 * return an extent size of 0 which means we can't
			 * make forward progress on this I/O... a hole in the
			 * file would be returned as a blkno of -1 with a non-zero io_size
			 * a real extent is returned with a blkno != -1 and a non-zero io_size
			 */
			error = EINVAL;
			break;
		}
		if (!(flags & CL_READ) && blkno == -1) {
			off_t   e_offset;
			int     pageout_flags;

			if (upl_get_internal_vectorupl(upl)) {
				panic("Vector UPLs should not take this code-path");
			}
			/*
			 * we're writing into a 'hole'
			 */
			if (flags & CL_PAGEOUT) {
				/*
				 * if we got here via cluster_pageout
				 * then just error the request and return
				 * the 'hole' should already have been covered
				 */
				error = EINVAL;
				break;
			}
			/*
			 * we can get here if the cluster code happens to
			 * pick up a page that was dirtied via mmap vs
			 * a 'write' and the page targets a 'hole'...
			 * i.e. the writes to the cluster were sparse
			 * and the file was being written for the first time
			 *
			 * we can also get here if the filesystem supports
			 * 'holes' that are less than PAGE_SIZE.... because
			 * we can't know if the range in the page that covers
			 * the 'hole' has been dirtied via an mmap or not,
			 * we have to assume the worst and try to push the
			 * entire page to storage.
			 *
			 * Try paging out the page individually before
			 * giving up entirely and dumping it (the pageout
			 * path will insure that the zero extent accounting
			 * has been taken care of before we get back into cluster_io)
			 *
			 * go direct to vnode_pageout so that we don't have to
			 * unbusy the page from the UPL... we used to do this
			 * so that we could call ubc_msync, but that results
			 * in a potential deadlock if someone else races us to acquire
			 * that page and wins and in addition needs one of the pages
			 * we're continuing to hold in the UPL
			 */
			pageout_flags = UPL_MSYNC | UPL_VNODE_PAGER | UPL_NESTED_PAGEOUT;

			if (!(flags & CL_ASYNC)) {
				pageout_flags |= UPL_IOSYNC;
			}
			if (!(flags & CL_COMMIT)) {
				pageout_flags |= UPL_NOCOMMIT;
			}

			if (cbp_head) {
				buf_t prev_cbp;
				uint32_t   bytes_in_last_page;

				/*
				 * first we have to wait for the the current outstanding I/Os
				 * to complete... EOT hasn't been set yet on this transaction
				 * so the pages won't be released
				 */
				cluster_wait_IO(cbp_head, (flags & CL_ASYNC));

				bytes_in_last_page = cbp_head->b_uploffset & PAGE_MASK;
				for (cbp = cbp_head; cbp; cbp = cbp->b_trans_next) {
					bytes_in_last_page += cbp->b_bcount;
				}
				bytes_in_last_page &= PAGE_MASK;

				while (bytes_in_last_page) {
					/*
					 * we've got a transcation that
					 * includes the page we're about to push out through vnode_pageout...
					 * find the bp's in the list which intersect this page and either
					 * remove them entirely from the transaction (there could be multiple bp's), or
					 * round it's iosize down to the page boundary (there can only be one)...
					 *
					 * find the last bp in the list and act on it
					 */
					for (prev_cbp = cbp = cbp_head; cbp->b_trans_next; cbp = cbp->b_trans_next) {
						prev_cbp = cbp;
					}

					if (bytes_in_last_page >= cbp->b_bcount) {
						/*
						 * this buf no longer has any I/O associated with it
						 */
						bytes_in_last_page -= cbp->b_bcount;
						cbp->b_bcount = 0;

						free_io_buf(cbp);

						if (cbp == cbp_head) {
							assert(bytes_in_last_page == 0);
							/*
							 * the buf we just freed was the only buf in
							 * this transaction... so there's no I/O to do
							 */
							cbp_head = NULL;
							cbp_tail = NULL;
						} else {
							/*
							 * remove the buf we just freed from
							 * the transaction list
							 */
							prev_cbp->b_trans_next = NULL;
							cbp_tail = prev_cbp;
						}
					} else {
						/*
						 * this is the last bp that has I/O
						 * intersecting the page of interest
						 * only some of the I/O is in the intersection
						 * so clip the size but keep it in the transaction list
						 */
						cbp->b_bcount -= bytes_in_last_page;
						cbp_tail = cbp;
						bytes_in_last_page = 0;
					}
				}
				if (cbp_head) {
					/*
					 * there was more to the current transaction
					 * than just the page we are pushing out via vnode_pageout...
					 * mark it as finished and complete it... we've already
					 * waited for the I/Os to complete above in the call to cluster_wait_IO
					 */
					cluster_EOT(cbp_head, cbp_tail, 0, 0);

					cluster_complete_transaction(&cbp_head, callback_arg, &retval, flags, 0);

					trans_count = 0;
				}
			}
			if (vnode_pageout(vp, upl, (upl_offset_t)trunc_page(upl_offset), trunc_page_64(f_offset), PAGE_SIZE, pageout_flags, NULL) != PAGER_SUCCESS) {
				error = EINVAL;
			}
			e_offset = round_page_64(f_offset + 1);
			io_size = (u_int)(e_offset - f_offset);

			f_offset   += io_size;
			upl_offset += io_size;

			if (size >= io_size) {
				size -= io_size;
			} else {
				size = 0;
			}
			/*
			 * keep track of how much of the original request
			 * that we've actually completed... non_rounded_size
			 * may go negative due to us rounding the request
			 * to a page size multiple (i.e.  size > non_rounded_size)
			 */
			non_rounded_size -= io_size;

			if (non_rounded_size <= 0) {
				/*
				 * we've transferred all of the data in the original
				 * request, but we were unable to complete the tail
				 * of the last page because the file didn't have
				 * an allocation to back that portion... this is ok.
				 */
				size = 0;
			}
			if (error) {
				if (size == 0) {
					flags &= ~CL_COMMIT;
				}
				break;
			}
			continue;
		}

		lblkno = (daddr64_t)(f_offset / CLUSTER_IO_BLOCK_SIZE);

		/*
		 * we have now figured out how much I/O we can do - this is in 'io_size'
		 * pg_offset is the starting point in the first page for the I/O
		 * pg_count is the number of full and partial pages that 'io_size' encompasses
		 */
		pg_offset = upl_offset & PAGE_MASK;

		if (flags & CL_DEV_MEMORY) {
			/*
			 * treat physical requests as one 'giant' page
			 */
			pg_count = 1;
		} else {
			pg_count  = (io_size + pg_offset + (PAGE_SIZE - 1)) / PAGE_SIZE;
		}

		if ((flags & CL_READ) && blkno == -1) {
			vm_offset_t  commit_offset;
			int bytes_to_zero;
			int complete_transaction_now = 0;

			/*
			 * if we're reading and blkno == -1, then we've got a
			 * 'hole' in the file that we need to deal with by zeroing
			 * out the affected area in the upl
			 */
			if (io_size >= (u_int)non_rounded_size) {
				/*
				 * if this upl contains the EOF and it is not a multiple of PAGE_SIZE
				 * than 'zero_offset' will be non-zero
				 * if the 'hole' returned by vnop_blockmap extends all the way to the eof
				 * (indicated by the io_size finishing off the I/O request for this UPL)
				 * than we're not going to issue an I/O for the
				 * last page in this upl... we need to zero both the hole and the tail
				 * of the page beyond the EOF, since the delayed zero-fill won't kick in
				 */
				bytes_to_zero = non_rounded_size;
				if (!(flags & CL_NOZERO)) {
					bytes_to_zero = (int)((((upl_offset + io_size) + (PAGE_SIZE - 1)) & ~PAGE_MASK) - upl_offset);
				}

				zero_offset = 0;
			} else {
				bytes_to_zero = io_size;
			}

			pg_count = 0;

			cluster_zero(upl, (upl_offset_t)upl_offset, bytes_to_zero, real_bp);

			if (cbp_head) {
				int     pg_resid;

				/*
				 * if there is a current I/O chain pending
				 * then the first page of the group we just zero'd
				 * will be handled by the I/O completion if the zero
				 * fill started in the middle of the page
				 */
				commit_offset = (upl_offset + (PAGE_SIZE - 1)) & ~PAGE_MASK;

				pg_resid = (int)(commit_offset - upl_offset);

				if (bytes_to_zero >= pg_resid) {
					/*
					 * the last page of the current I/O
					 * has been completed...
					 * compute the number of fully zero'd
					 * pages that are beyond it
					 * plus the last page if its partial
					 * and we have no more I/O to issue...
					 * otherwise a partial page is left
					 * to begin the next I/O
					 */
					if ((int)io_size >= non_rounded_size) {
						pg_count = (bytes_to_zero - pg_resid + (PAGE_SIZE - 1)) / PAGE_SIZE;
					} else {
						pg_count = (bytes_to_zero - pg_resid) / PAGE_SIZE;
					}

					complete_transaction_now = 1;
				}
			} else {
				/*
				 * no pending I/O to deal with
				 * so, commit all of the fully zero'd pages
				 * plus the last page if its partial
				 * and we have no more I/O to issue...
				 * otherwise a partial page is left
				 * to begin the next I/O
				 */
				if ((int)io_size >= non_rounded_size) {
					pg_count = (pg_offset + bytes_to_zero + (PAGE_SIZE - 1)) / PAGE_SIZE;
				} else {
					pg_count = (pg_offset + bytes_to_zero) / PAGE_SIZE;
				}

				commit_offset = upl_offset & ~PAGE_MASK;
			}

			// Associated UPL is currently only used in the direct write path
			assert(!upl_associated_upl(upl));

			if ((flags & CL_COMMIT) && pg_count) {
				ubc_upl_commit_range(upl, (upl_offset_t)commit_offset,
				    pg_count * PAGE_SIZE,
				    UPL_COMMIT_CLEAR_DIRTY | UPL_COMMIT_FREE_ON_EMPTY);
			}
			upl_offset += io_size;
			f_offset   += io_size;
			size       -= io_size;

			/*
			 * keep track of how much of the original request
			 * that we've actually completed... non_rounded_size
			 * may go negative due to us rounding the request
			 * to a page size multiple (i.e.  size > non_rounded_size)
			 */
			non_rounded_size -= io_size;

			if (non_rounded_size <= 0) {
				/*
				 * we've transferred all of the data in the original
				 * request, but we were unable to complete the tail
				 * of the last page because the file didn't have
				 * an allocation to back that portion... this is ok.
				 */
				size = 0;
			}
			if (cbp_head && (complete_transaction_now || size == 0)) {
				cluster_wait_IO(cbp_head, (flags & CL_ASYNC));

				cluster_EOT(cbp_head, cbp_tail, size == 0 ? zero_offset : 0, verify_block_size);

				cluster_complete_transaction(&cbp_head, callback_arg, &retval, flags, 0);

				trans_count = 0;
			}
			continue;
		}
		if (pg_count > max_vectors) {
			if (((pg_count - max_vectors) * PAGE_SIZE) > io_size) {
				io_size = PAGE_SIZE - pg_offset;
				pg_count = 1;
			} else {
				io_size -= (pg_count - max_vectors) * PAGE_SIZE;
				pg_count = max_vectors;
			}
		}
		/*
		 * If the transaction is going to reach the maximum number of
		 * desired elements, truncate the i/o to the nearest page so
		 * that the actual i/o is initiated after this buffer is
		 * created and added to the i/o chain.
		 *
		 * I/O directed to physically contiguous memory
		 * doesn't have a requirement to make sure we 'fill' a page
		 */
		if (!(flags & CL_DEV_MEMORY) && trans_count >= max_trans_count &&
		    ((upl_offset + io_size) & PAGE_MASK)) {
			vm_offset_t aligned_ofs;

			aligned_ofs = (upl_offset + io_size) & ~PAGE_MASK;
			/*
			 * If the io_size does not actually finish off even a
			 * single page we have to keep adding buffers to the
			 * transaction despite having reached the desired limit.
			 *
			 * Eventually we get here with the page being finished
			 * off (and exceeded) and then we truncate the size of
			 * this i/o request so that it is page aligned so that
			 * we can finally issue the i/o on the transaction.
			 */
			if (aligned_ofs > upl_offset) {
				io_size = (u_int)(aligned_ofs - upl_offset);
				pg_count--;
			}
		}

		if (!(mp->mnt_kern_flag & MNTK_VIRTUALDEV)) {
			/*
			 * if we're not targeting a virtual device i.e. a disk image
			 * it's safe to dip into the reserve pool since real devices
			 * can complete this I/O request without requiring additional
			 * bufs from the alloc_io_buf pool
			 */
			priv = 1;
		} else if ((flags & CL_ASYNC) && !(flags & CL_PAGEOUT) && !cbp_head) {
			/*
			 * Throttle the speculative IO
			 *
			 * We can only throttle this if it is the first iobuf
			 * for the transaction. alloc_io_buf implements
			 * additional restrictions for diskimages anyway.
			 */
			priv = 0;
		} else {
			priv = 1;
		}

		cbp = alloc_io_buf(vp, priv);

		if (flags & CL_PAGEOUT) {
			u_int i;

			/*
			 * since blocks are in offsets of CLUSTER_IO_BLOCK_SIZE, scale
			 * iteration to (PAGE_SIZE * pg_count) of blks.
			 */
			for (i = 0; i < (PAGE_SIZE * pg_count) / CLUSTER_IO_BLOCK_SIZE; i++) {
				if (buf_invalblkno(vp, lblkno + i, 0) == EBUSY) {
					panic("BUSY bp found in cluster_io");
				}
			}
		}
		if (flags & CL_ASYNC) {
			if (buf_setcallback(cbp, (void *)cluster_iodone, callback_arg)) {
				panic("buf_setcallback failed");
			}
		}
		cbp->b_cliodone = (void *)callback;
		cbp->b_flags |= io_flags;
		if (flags & CL_NOCACHE) {
			cbp->b_attr.ba_flags |= BA_NOCACHE;
		}
		if (verify_block_size) {
			cbp->b_attr.ba_flags |= BA_WILL_VERIFY;
			if (verify_kind) {
				cbp->b_attr.ba_verify_type = verify_kind;
			}
		}

		cbp->b_lblkno = lblkno;
		cbp->b_clfoffset = f_offset;
		cbp->b_blkno  = blkno;
		cbp->b_bcount = io_size;

		if (buf_setupl(cbp, upl, (uint32_t)upl_offset)) {
			panic("buf_setupl failed");
		}
#if CONFIG_IOSCHED
		upl_set_blkno(upl, upl_offset, io_size, blkno);
#endif
		cbp->b_trans_next = (buf_t)NULL;

		if ((cbp->b_iostate = (void *)iostate)) {
			/*
			 * caller wants to track the state of this
			 * io... bump the amount issued against this stream
			 */
			iostate->io_issued += io_size;
		}

		if (flags & CL_READ) {
			KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 26)) | DBG_FUNC_NONE,
			    (int)cbp->b_lblkno, (int)cbp->b_blkno, upl_offset, io_size, 0);
		} else {
			KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 27)) | DBG_FUNC_NONE,
			    (int)cbp->b_lblkno, (int)cbp->b_blkno, upl_offset, io_size, 0);
		}

		if (cbp_head) {
			cbp_tail->b_trans_next = cbp;
			cbp_tail = cbp;
		} else {
			cbp_head = cbp;
			cbp_tail = cbp;

			if ((cbp_head->b_real_bp = real_bp)) {
				real_bp = (buf_t)NULL;
			}
		}
		*(buf_t *)(&cbp->b_trans_head) = cbp_head;

		trans_count++;

		upl_offset += io_size;
		f_offset   += io_size;
		size       -= io_size;
		/*
		 * keep track of how much of the original request
		 * that we've actually completed... non_rounded_size
		 * may go negative due to us rounding the request
		 * to a page size multiple (i.e.  size > non_rounded_size)
		 */
		non_rounded_size -= io_size;

		if (non_rounded_size <= 0) {
			/*
			 * we've transferred all of the data in the original
			 * request, but we were unable to complete the tail
			 * of the last page because the file didn't have
			 * an allocation to back that portion... this is ok.
			 */
			size = 0;
		}
		if (size == 0) {
			/*
			 * we have no more I/O to issue, so go
			 * finish the final transaction
			 */
			need_EOT = TRUE;
		} else if (((flags & CL_DEV_MEMORY) || (upl_offset & PAGE_MASK) == 0) &&
		    ((flags & CL_ASYNC) || trans_count > max_trans_count)) {
			/*
			 * I/O directed to physically contiguous memory...
			 * which doesn't have a requirement to make sure we 'fill' a page
			 * or...
			 * the current I/O we've prepared fully
			 * completes the last page in this request
			 * and ...
			 * it's either an ASYNC request or
			 * we've already accumulated more than 8 I/O's into
			 * this transaction so mark it as complete so that
			 * it can finish asynchronously or via the cluster_complete_transaction
			 * below if the request is synchronous
			 */
			need_EOT = TRUE;
		}
		if (need_EOT == TRUE) {
			cluster_EOT(cbp_head, cbp_tail, size == 0 ? zero_offset : 0, verify_block_size);
		}

		if (flags & CL_THROTTLE) {
			(void)vnode_waitforwrites(vp, async_throttle, 0, 0, "cluster_io");
		}

		if (!(io_flags & B_READ)) {
			vnode_startwrite(vp);
		}

		if (flags & CL_RAW_ENCRYPTED) {
			/*
			 * User requested raw encrypted bytes.
			 * Twiddle the bit in the ba_flags for the buffer
			 */
			cbp->b_attr.ba_flags |= BA_RAW_ENCRYPTED_IO;
		}

		(void) VNOP_STRATEGY(cbp);

		if (need_EOT == TRUE) {
			if (!(flags & CL_ASYNC)) {
				cluster_complete_transaction(&cbp_head, callback_arg, &retval, flags, 1);
			}

			need_EOT = FALSE;
			trans_count = 0;
			cbp_head = NULL;
		}
	}
	if (error) {
		int abort_size;

		io_size = 0;

		if (cbp_head) {
			/*
			 * Wait until all of the outstanding I/O
			 * for this partial transaction has completed
			 */
			cluster_wait_IO(cbp_head, (flags & CL_ASYNC));

			/*
			 * Rewind the upl offset to the beginning of the
			 * transaction.
			 */
			upl_offset = cbp_head->b_uploffset;
		}

		if (ISSET(flags, CL_COMMIT)) {
			cluster_handle_associated_upl(iostate, upl,
			    (upl_offset_t)upl_offset,
			    (upl_size_t)(upl_end_offset - upl_offset),
			    cbp_head ? cbp_head->b_clfoffset : f_offset);
		}

		// Free all the IO buffers in this transaction
		for (cbp = cbp_head; cbp;) {
			buf_t   cbp_next;

			size       += cbp->b_bcount;
			io_size    += cbp->b_bcount;

			cbp_next = cbp->b_trans_next;
			free_io_buf(cbp);
			cbp = cbp_next;
		}

		if (iostate) {
			int need_wakeup = 0;

			/*
			 * update the error condition for this stream
			 * since we never really issued the io
			 * just go ahead and adjust it back
			 */
			lck_mtx_lock_spin(&iostate->io_mtxp);

			if (iostate->io_error == 0) {
				iostate->io_error = error;
			}
			iostate->io_issued -= io_size;

			if (iostate->io_wanted) {
				/*
				 * someone is waiting for the state of
				 * this io stream to change
				 */
				iostate->io_wanted = 0;
				need_wakeup = 1;
			}
			lck_mtx_unlock(&iostate->io_mtxp);

			if (need_wakeup) {
				wakeup((caddr_t)&iostate->io_wanted);
			}
		}

		if (flags & CL_COMMIT) {
			int     upl_flags;

			pg_offset  = upl_offset & PAGE_MASK;
			abort_size = (int)((upl_end_offset - upl_offset + PAGE_MASK) & ~PAGE_MASK);

			upl_flags = cluster_ioerror(upl, (int)(upl_offset - pg_offset),
			    abort_size, error, io_flags, vp);

			KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 28)) | DBG_FUNC_NONE,
			    upl, upl_offset - pg_offset, abort_size, (error << 24) | upl_flags, 0);
		}
		if (retval == 0) {
			retval = error;
		}
	} else if (cbp_head) {
		panic("%s(): cbp_head is not NULL.", __FUNCTION__);
	}

	if (real_bp) {
		/*
		 * can get here if we either encountered an error
		 * or we completely zero-filled the request and
		 * no I/O was issued
		 */
		if (error) {
			real_bp->b_flags |= B_ERROR;
			real_bp->b_error = error;
		}
		buf_biodone(real_bp);
	}
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 22)) | DBG_FUNC_END, (int)f_offset, size, upl_offset, retval, 0);

	return retval;
}

#define reset_vector_run_state()                                                                                \
	issueVectorUPL = vector_upl_offset = vector_upl_index = vector_upl_iosize = vector_upl_size = 0;

static int
vector_cluster_io(vnode_t vp, upl_t vector_upl, vm_offset_t vector_upl_offset, off_t v_upl_uio_offset, int vector_upl_iosize,
    int io_flag, buf_t real_bp, struct clios *iostate, int (*callback)(buf_t, void *), void *callback_arg)
{
	vector_upl_set_pagelist(vector_upl);

	if (io_flag & CL_READ) {
		if (vector_upl_offset == 0 && ((vector_upl_iosize & PAGE_MASK) == 0)) {
			io_flag &= ~CL_PRESERVE; /*don't zero fill*/
		} else {
			io_flag |= CL_PRESERVE; /*zero fill*/
		}
	}
	return cluster_io(vp, vector_upl, vector_upl_offset, v_upl_uio_offset, vector_upl_iosize, io_flag, real_bp, iostate, callback, callback_arg);
}

static int
cluster_read_prefetch(vnode_t vp, off_t f_offset, u_int size, off_t filesize, int (*callback)(buf_t, void *), void *callback_arg, int bflag)
{
	int           pages_in_prefetch;

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 49)) | DBG_FUNC_START,
	    (int)f_offset, size, (int)filesize, 0, 0);

	if (f_offset >= filesize) {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 49)) | DBG_FUNC_END,
		    (int)f_offset, 0, 0, 0, 0);
		return 0;
	}
	if ((off_t)size > (filesize - f_offset)) {
		size = (u_int)(filesize - f_offset);
	}
	pages_in_prefetch = (size + (PAGE_SIZE - 1)) / PAGE_SIZE;

	advisory_read_ext(vp, filesize, f_offset, size, callback, callback_arg, bflag);

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 49)) | DBG_FUNC_END,
	    (int)f_offset + size, pages_in_prefetch, 0, 1, 0);

	return pages_in_prefetch;
}



static void
cluster_read_ahead(vnode_t vp, struct cl_extent *extent, off_t filesize, struct cl_readahead *rap, int (*callback)(buf_t, void *), void *callback_arg,
    int bflag)
{
	daddr64_t       r_addr;
	off_t           f_offset;
	int             size_of_prefetch;
	u_int           max_prefetch;


	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 48)) | DBG_FUNC_START,
	    (int)extent->b_addr, (int)extent->e_addr, (int)rap->cl_lastr, 0, 0);

	if (extent->b_addr == rap->cl_lastr && extent->b_addr == extent->e_addr) {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 48)) | DBG_FUNC_END,
		    rap->cl_ralen, (int)rap->cl_maxra, (int)rap->cl_lastr, 0, 0);
		return;
	}
	if (rap->cl_lastr == -1 || (extent->b_addr != rap->cl_lastr && extent->b_addr != (rap->cl_lastr + 1))) {
		rap->cl_ralen = 0;
		rap->cl_maxra = 0;

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 48)) | DBG_FUNC_END,
		    rap->cl_ralen, (int)rap->cl_maxra, (int)rap->cl_lastr, 1, 0);

		return;
	}

	max_prefetch = cluster_max_prefetch(vp,
	    cluster_max_io_size(vp->v_mount, CL_READ), speculative_prefetch_max);

	if (max_prefetch <= PAGE_SIZE) {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 48)) | DBG_FUNC_END,
		    rap->cl_ralen, (int)rap->cl_maxra, (int)rap->cl_lastr, 6, 0);
		return;
	}
	if (extent->e_addr < rap->cl_maxra && rap->cl_ralen >= 4) {
		if ((rap->cl_maxra - extent->e_addr) > (rap->cl_ralen / 4)) {
			KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 48)) | DBG_FUNC_END,
			    rap->cl_ralen, (int)rap->cl_maxra, (int)rap->cl_lastr, 2, 0);
			return;
		}
	}
	r_addr = MAX(extent->e_addr, rap->cl_maxra) + 1;
	f_offset = (off_t)(r_addr * PAGE_SIZE_64);

	size_of_prefetch = 0;

	ubc_range_op(vp, f_offset, f_offset + PAGE_SIZE_64, UPL_ROP_PRESENT, &size_of_prefetch);

	if (size_of_prefetch) {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 48)) | DBG_FUNC_END,
		    rap->cl_ralen, (int)rap->cl_maxra, (int)rap->cl_lastr, 3, 0);
		return;
	}
	if (f_offset < filesize) {
		daddr64_t read_size;

		rap->cl_ralen = rap->cl_ralen ? min(max_prefetch / PAGE_SIZE, rap->cl_ralen << 1) : 1;

		read_size = (extent->e_addr + 1) - extent->b_addr;

		if (read_size > rap->cl_ralen) {
			if (read_size > max_prefetch / PAGE_SIZE) {
				rap->cl_ralen = max_prefetch / PAGE_SIZE;
			} else {
				rap->cl_ralen = (int)read_size;
			}
		}
		size_of_prefetch = cluster_read_prefetch(vp, f_offset, rap->cl_ralen * PAGE_SIZE, filesize, callback, callback_arg, bflag);

		if (size_of_prefetch) {
			rap->cl_maxra = (r_addr + size_of_prefetch) - 1;
		}
	}
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 48)) | DBG_FUNC_END,
	    rap->cl_ralen, (int)rap->cl_maxra, (int)rap->cl_lastr, 4, 0);
}


int
cluster_pageout(vnode_t vp, upl_t upl, upl_offset_t upl_offset, off_t f_offset,
    int size, off_t filesize, int flags)
{
	return cluster_pageout_ext(vp, upl, upl_offset, f_offset, size, filesize, flags, NULL, NULL);
}


int
cluster_pageout_ext(vnode_t vp, upl_t upl, upl_offset_t upl_offset, off_t f_offset,
    int size, off_t filesize, int flags, int (*callback)(buf_t, void *), void *callback_arg)
{
	int           io_size;
	int           rounded_size;
	off_t         max_size;
	int           local_flags;

	local_flags = CL_PAGEOUT | CL_THROTTLE;

	if ((flags & UPL_IOSYNC) == 0) {
		local_flags |= CL_ASYNC;
	}
	if ((flags & UPL_NOCOMMIT) == 0) {
		local_flags |= CL_COMMIT;
	}
	if ((flags & UPL_KEEPCACHED)) {
		local_flags |= CL_KEEPCACHED;
	}
	if (flags & UPL_PAGING_ENCRYPTED) {
		local_flags |= CL_ENCRYPTED;
	}


	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 52)) | DBG_FUNC_NONE,
	    (int)f_offset, size, (int)filesize, local_flags, 0);

	/*
	 * If they didn't specify any I/O, then we are done...
	 * we can't issue an abort because we don't know how
	 * big the upl really is
	 */
	if (size <= 0) {
		return EINVAL;
	}

	if (vp->v_mount->mnt_flag & MNT_RDONLY) {
		if (local_flags & CL_COMMIT) {
			ubc_upl_abort_range(upl, upl_offset, size, UPL_ABORT_FREE_ON_EMPTY);
		}
		return EROFS;
	}
	/*
	 * can't page-in from a negative offset
	 * or if we're starting beyond the EOF
	 * or if the file offset isn't page aligned
	 * or the size requested isn't a multiple of PAGE_SIZE
	 */
	if (f_offset < 0 || f_offset >= filesize ||
	    (f_offset & PAGE_MASK_64) || (size & PAGE_MASK)) {
		if (local_flags & CL_COMMIT) {
			ubc_upl_abort_range(upl, upl_offset, size, UPL_ABORT_FREE_ON_EMPTY);
		}
		return EINVAL;
	}
	max_size = filesize - f_offset;

	if (size < max_size) {
		io_size = size;
	} else {
		io_size = (int)max_size;
	}

	rounded_size = (io_size + (PAGE_SIZE - 1)) & ~PAGE_MASK;

	if (size > rounded_size) {
		if (local_flags & CL_COMMIT) {
			ubc_upl_abort_range(upl, upl_offset + rounded_size, size - rounded_size,
			    UPL_ABORT_FREE_ON_EMPTY);
		}
	}
	return cluster_io(vp, upl, upl_offset, f_offset, io_size,
	           local_flags, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg);
}


int
cluster_pagein(vnode_t vp, upl_t upl, upl_offset_t upl_offset, off_t f_offset,
    int size, off_t filesize, int flags)
{
	return cluster_pagein_ext(vp, upl, upl_offset, f_offset, size, filesize, flags, NULL, NULL);
}

#define SPLIT_PAGEIN_MAX_IOSIZE 32768

/*
 * Do a big pagein request as multiple I/Os - the first I/O will be for
 * SPLIT_PAGEIN_MAX_IOSIZE (32K)sized which includes the page that the caused
 * the fault and then i/o will be initiated for the remaining.
 */
static int
cluster_handle_split_pagein(vnode_t vp, upl_t upl, upl_offset_t upl_offset, off_t f_offset,
    u_int io_size, int rounded_size, int local_flags, int (*callback)(buf_t, void *), void *callback_arg)
{
	upl_page_info_t *pl = ubc_upl_pageinfo(upl);
	const off_t start_f_offset = f_offset;
	const upl_offset_t start_upl_offset = upl_offset;
	const int start_pg = upl_offset >> PAGE_SHIFT;
	const int last_pg = ((upl_offset + rounded_size) >> PAGE_SHIFT) - 1;
	u_int split_io_size = SPLIT_PAGEIN_MAX_IOSIZE;
	u_int head_io_size = 0;
	int retval = 0;
	int error = 0;
	int pg;

	assert(SPLIT_PAGEIN_MAX_IOSIZE >= (2 * PAGE_SIZE));

	for (pg = start_pg; (pg <= last_pg) && !(upl_page_is_needed(pl, pg)); pg++) {
		;
	}

	/*
	 * The global variables affecting behaviour
	 * split_all_pgin -> Split pageins even if we don't find the needed page.
	 * split_pgin_headio -> for a pagein in which there is a head calculated,
	 *                   do the head i/o or not.
	 *
	 * split_all_pgin_equal -> split the entire bug request into equal sized small i/os of 32K.
	 *
	 * Whichever way the i/o is split, the i/o for the needed page always happens first and then we decide
	 * whether we have to do i/o for the head and then if we need to issue equal sized i/o.
	 *
	 * By default we are set up to do only the i/o for the needed page, followed by a "unsplit" tail.
	 */
	if ((pg > start_pg) && (pg <= last_pg)) {
		head_io_size = ((pg - start_pg) * PAGE_SIZE);

		if (head_io_size < SPLIT_PAGEIN_MAX_IOSIZE) {
			head_io_size = 0;
		} else if (!split_all_pgin) {
			goto out;
		} else if ((rounded_size - head_io_size) <= SPLIT_PAGEIN_MAX_IOSIZE) {
			head_io_size = (rounded_size - SPLIT_PAGEIN_MAX_IOSIZE);
		} else {
			head_io_size &= ~(SPLIT_PAGEIN_MAX_IOSIZE - 1);
		}

		assertf(io_size > head_io_size, "io_size is %d, head_io_size = %d", io_size, head_io_size);

		if (head_io_size) {
			upl_offset += head_io_size;
			f_offset += head_io_size;
			io_size -= head_io_size;

			if (!split_pgin_headio) {
				if (local_flags & CL_COMMIT) {
					ubc_upl_abort_range(upl, start_upl_offset, head_io_size,
					    UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_ERROR);
				}
				head_io_size = 0;
			}

			split_io_size = MIN(SPLIT_PAGEIN_MAX_IOSIZE, io_size);
		}

		assertf(io_size >= split_io_size, "io_size is %d, split_io_size = %d", io_size, split_io_size);
	} else if ((pg > last_pg) && !split_all_pgin) {
		goto out;
	}

	/* This is the 32K i/o for the "needed" page */
	retval = cluster_io(vp, upl, upl_offset, f_offset, split_io_size,
	    local_flags | CL_READ | CL_PAGEIN, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg);

	io_size -= split_io_size;

	if (io_size) {
		upl_offset += split_io_size;
		f_offset += split_io_size;
	} else if (head_io_size) {
		io_size = head_io_size;
		head_io_size = 0;
		upl_offset = start_upl_offset;
		f_offset = start_f_offset;
	}

	while (io_size) {
		if (split_all_pgin_equal && (io_size > SPLIT_PAGEIN_MAX_IOSIZE)) {
			split_io_size = SPLIT_PAGEIN_MAX_IOSIZE;
		} else {
			split_io_size = io_size;
		}

		assertf(io_size >= split_io_size, "io_size is %d, split_io_size = %d", io_size, split_io_size);

		/* We have to issue this i/o anyway even if we get an error from any of the previous ones */
		error = cluster_io(vp, upl, upl_offset, f_offset, split_io_size,
		    local_flags | CL_READ | CL_PAGEIN, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg);
		if (!retval) {
			retval = error;
		}

		io_size -= split_io_size;

		if ((io_size == 0) && head_io_size) {
			io_size = head_io_size;
			head_io_size = 0;
			upl_offset = start_upl_offset;
			f_offset = start_f_offset;
		} else if (io_size) {
			upl_offset += split_io_size;
			f_offset += split_io_size;
		}
	}

	return retval;
out:
	return cluster_io(vp, upl, upl_offset, f_offset, io_size,
	           local_flags | CL_READ | CL_PAGEIN, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg);
}

int
cluster_pagein_ext(vnode_t vp, upl_t upl, upl_offset_t upl_offset, off_t f_offset,
    int size, off_t filesize, int flags, int (*callback)(buf_t, void *), void *callback_arg)
{
	u_int         io_size;
	int           rounded_size;
	off_t         max_size;
	int           retval;
	int           local_flags = 0;

	if (upl == NULL || size < 0) {
		panic("cluster_pagein: NULL upl passed in");
	}

	if ((flags & UPL_IOSYNC) == 0) {
		local_flags |= CL_ASYNC;
	}
	if ((flags & UPL_NOCOMMIT) == 0) {
		local_flags |= CL_COMMIT;
	}
	if (flags & UPL_IOSTREAMING) {
		local_flags |= CL_IOSTREAMING;
	}
	if (flags & UPL_PAGING_ENCRYPTED) {
		local_flags |= CL_ENCRYPTED;
	}


	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 56)) | DBG_FUNC_NONE,
	    (int)f_offset, size, (int)filesize, local_flags, 0);

	/*
	 * can't page-in from a negative offset
	 * or if we're starting beyond the EOF
	 * or if the file offset isn't page aligned
	 * or the size requested isn't a multiple of PAGE_SIZE
	 */
	if (f_offset < 0 || f_offset >= filesize ||
	    (f_offset & PAGE_MASK_64) || (size & PAGE_MASK) || (upl_offset & PAGE_MASK)) {
		if (local_flags & CL_COMMIT) {
			ubc_upl_abort_range(upl, upl_offset, size, UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_ERROR);
		}

		if (f_offset >= filesize) {
			ktriage_record(thread_tid(current_thread()), KDBG_TRIAGE_EVENTID(KDBG_TRIAGE_SUBSYS_CLUSTER, KDBG_TRIAGE_RESERVED, KDBG_TRIAGE_CL_PGIN_PAST_EOF), 0 /* arg */);
		}

		return EINVAL;
	}
	max_size = filesize - f_offset;

	if (size < max_size) {
		io_size = size;
	} else {
		io_size = (int)max_size;
	}

	rounded_size = (io_size + (PAGE_SIZE - 1)) & ~PAGE_MASK;

	if (size > rounded_size && (local_flags & CL_COMMIT)) {
		ubc_upl_abort_range(upl, upl_offset + rounded_size,
		    size - rounded_size, UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_ERROR);
	}

	if ((io_size > SPLIT_PAGEIN_MAX_IOSIZE) && vnode_isonssd(vp) && split_pgin) {
		return cluster_handle_split_pagein(vp, upl, upl_offset, f_offset, io_size,
		           rounded_size, local_flags, callback, callback_arg);
	}

	retval = cluster_io(vp, upl, upl_offset, f_offset, io_size,
	    local_flags | CL_READ | CL_PAGEIN, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg);

	return retval;
}


int
cluster_bp(buf_t bp)
{
	return cluster_bp_ext(bp, NULL, NULL);
}


int
cluster_bp_ext(buf_t bp, int (*callback)(buf_t, void *), void *callback_arg)
{
	off_t  f_offset;
	int    flags;

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 19)) | DBG_FUNC_START,
	    bp, (int)bp->b_lblkno, bp->b_bcount, bp->b_flags, 0);

	if (bp->b_flags & B_READ) {
		flags = CL_ASYNC | CL_READ;
	} else {
		flags = CL_ASYNC;
	}
	if (bp->b_flags & B_PASSIVE) {
		flags |= CL_PASSIVE;
	}

	f_offset = ubc_blktooff(bp->b_vp, bp->b_lblkno);

	return cluster_io(bp->b_vp, bp->b_upl, 0, f_offset, bp->b_bcount, flags, bp, (struct clios *)NULL, callback, callback_arg);
}



int
cluster_write(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t headOff, off_t tailOff, int xflags)
{
	return cluster_write_ext(vp, uio, oldEOF, newEOF, headOff, tailOff, xflags, NULL, NULL);
}


int
cluster_write_ext(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t headOff, off_t tailOff,
    int xflags, int (*callback)(buf_t, void *), void *callback_arg)
{
	user_ssize_t    cur_resid;
	int             retval = 0;
	int             flags;
	int             zflags;
	int             bflag;
	int             write_type = IO_COPY;
	u_int32_t       write_length = 0, saved_write_length;
	uint32_t        min_direct_size = MIN_DIRECT_WRITE_SIZE;

	flags = xflags;

	if (flags & IO_PASSIVE) {
		bflag = CL_PASSIVE;
	} else {
		bflag = 0;
	}

	if (vp->v_flag & VNOCACHE_DATA) {
		flags |= IO_NOCACHE;
		bflag |= CL_NOCACHE;
	}
	if (uio == NULL) {
		/*
		 * no user data...
		 * this call is being made to zero-fill some range in the file
		 */
		retval = cluster_write_copy(vp, NULL, (u_int32_t)0, oldEOF, newEOF, headOff, tailOff, flags, callback, callback_arg);

		return retval;
	}
	/*
	 * do a write through the cache if one of the following is true....
	 *   NOCACHE is not true or NODIRECT is true
	 *   the uio request doesn't target USERSPACE
	 * otherwise, find out if we want the direct or contig variant for
	 * the first vector in the uio request
	 */
	if (((flags & (IO_NOCACHE | IO_NODIRECT)) == IO_NOCACHE) && UIO_SEG_IS_USER_SPACE(uio->uio_segflg)) {
		if (flags & IO_NOCACHE_SWRITE) {
			uint32_t fs_bsize = vp->v_mount->mnt_vfsstat.f_bsize;

			if (fs_bsize && (fs_bsize < MIN_DIRECT_WRITE_SIZE) &&
			    ((fs_bsize & (fs_bsize - 1)) == 0)) {
				min_direct_size = fs_bsize;
			}
		}
		retval = cluster_io_type(uio, &write_type, &write_length, min_direct_size);
	}

	if ((flags & (IO_TAILZEROFILL | IO_HEADZEROFILL)) && write_type == IO_DIRECT) {
		/*
		 * must go through the cached variant in this case
		 */
		write_type = IO_COPY;
	}

	while ((cur_resid = uio_resid(uio)) && uio->uio_offset < newEOF && retval == 0) {
		switch (write_type) {
		case IO_COPY:
			/*
			 * make sure the uio_resid isn't too big...
			 * internally, we want to handle all of the I/O in
			 * chunk sizes that fit in a 32 bit int
			 */
			if (cur_resid > (user_ssize_t)(MAX_IO_REQUEST_SIZE)) {
				/*
				 * we're going to have to call cluster_write_copy
				 * more than once...
				 *
				 * only want the last call to cluster_write_copy to
				 * have the IO_TAILZEROFILL flag set and only the
				 * first call should have IO_HEADZEROFILL
				 */
				zflags = flags & ~IO_TAILZEROFILL;
				flags &= ~IO_HEADZEROFILL;

				write_length = MAX_IO_REQUEST_SIZE;
			} else {
				/*
				 * last call to cluster_write_copy
				 */
				zflags = flags;

				write_length = (u_int32_t)cur_resid;
			}
			retval = cluster_write_copy(vp, uio, write_length, oldEOF, newEOF, headOff, tailOff, zflags, callback, callback_arg);
			break;

		case IO_CONTIG:
			zflags = flags & ~(IO_TAILZEROFILL | IO_HEADZEROFILL);

			if (flags & IO_HEADZEROFILL) {
				/*
				 * only do this once per request
				 */
				flags &= ~IO_HEADZEROFILL;

				retval = cluster_write_copy(vp, (struct uio *)0, (u_int32_t)0, (off_t)0, uio->uio_offset,
				    headOff, (off_t)0, zflags | IO_HEADZEROFILL | IO_SYNC, callback, callback_arg);
				if (retval) {
					break;
				}
			}
			retval = cluster_write_contig(vp, uio, newEOF, &write_type, &write_length, callback, callback_arg, bflag);

			if (retval == 0 && (flags & IO_TAILZEROFILL) && uio_resid(uio) == 0) {
				/*
				 * we're done with the data from the user specified buffer(s)
				 * and we've been requested to zero fill at the tail
				 * treat this as an IO_HEADZEROFILL which doesn't require a uio
				 * by rearranging the args and passing in IO_HEADZEROFILL
				 */

				/*
				 * Update the oldEOF to reflect the current EOF. If the UPL page
				 * to zero-fill is not valid (when F_NOCACHE is set), the
				 * cluster_write_copy() will perform RMW on the UPL page when
				 * the oldEOF is not aligned on page boundary due to unaligned
				 * write.
				 */
				if (uio->uio_offset > oldEOF) {
					oldEOF = uio->uio_offset;
				}
				retval = cluster_write_copy(vp, (struct uio *)0, (u_int32_t)0, (off_t)oldEOF, tailOff, uio->uio_offset,
				    (off_t)0, zflags | IO_HEADZEROFILL | IO_SYNC, callback, callback_arg);
			}
			break;

		case IO_DIRECT:
			/*
			 * cluster_write_direct is never called with IO_TAILZEROFILL || IO_HEADZEROFILL
			 */
			saved_write_length = write_length;
			retval = cluster_write_direct(vp, uio, oldEOF, newEOF, &write_type, &write_length, flags, callback, callback_arg, min_direct_size);
			if (retval == ENOTSUP) {
				/* direct I/O didn't work; retry with cached I/O */
//				printf("*******  FBDP %s:%d ENOTSUP cnt %d resid 0x%llx offset 0x%llx write_length 0x%x -> 0x%x\n", __FUNCTION__, __LINE__, uio_iovcnt(uio), (uint64_t) uio_resid(uio), uio_offset(uio), write_length, saved_write_length);
				write_length = saved_write_length;
				write_type = IO_COPY;
				retval = 0;
			}
			break;

		case IO_UNKNOWN:
			retval = cluster_io_type(uio, &write_type, &write_length, min_direct_size);
			break;
		}
		/*
		 * in case we end up calling cluster_write_copy (from cluster_write_direct)
		 * multiple times to service a multi-vector request that is not aligned properly
		 * we need to update the oldEOF so that we
		 * don't zero-fill the head of a page if we've successfully written
		 * data to that area... 'cluster_write_copy' will zero-fill the head of a
		 * page that is beyond the oldEOF if the write is unaligned... we only
		 * want that to happen for the very first page of the cluster_write,
		 * NOT the first page of each vector making up a multi-vector write.
		 */
		if (uio->uio_offset > oldEOF) {
			oldEOF = uio->uio_offset;
		}
	}
	return retval;
}


static int
cluster_write_direct(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, int *write_type, u_int32_t *write_length,
    int flags, int (*callback)(buf_t, void *), void *callback_arg, uint32_t min_io_size)
{
	upl_t            upl = NULL;
	upl_page_info_t  *pl;
	vm_offset_t      upl_offset;
	vm_offset_t      vector_upl_offset = 0;
	u_int32_t        io_req_size;
	u_int32_t        offset_in_file;
	u_int32_t        offset_in_iovbase;
	u_int32_t        io_size;
	int              io_flag = 0;
	upl_size_t       upl_size = 0, vector_upl_size = 0;
	vm_size_t        upl_needed_size;
	mach_msg_type_number_t  pages_in_pl = 0;
	upl_control_flags_t upl_flags;
	kern_return_t    kret = KERN_SUCCESS;
	mach_msg_type_number_t  i = 0;
	int              force_data_sync;
	int              retval = 0;
	int              first_IO = 1;
	struct clios     iostate;
	user_addr_t      iov_base;
	u_int32_t        mem_alignment_mask;
	u_int32_t        devblocksize;
	u_int32_t        max_io_size;
	u_int32_t        max_upl_size;
	u_int32_t        max_vector_size;
	u_int32_t        bytes_outstanding_limit;
	boolean_t        io_throttled = FALSE;

	u_int32_t        vector_upl_iosize = 0;
	int              issueVectorUPL = 0, useVectorUPL = (uio->uio_iovcnt > 1);
	off_t            v_upl_uio_offset = 0;
	int              vector_upl_index = 0;
	upl_t            vector_upl = NULL;
	uio_t            snapshot_uio = NULL;

	uint32_t         io_align_mask;

	/*
	 * When we enter this routine, we know
	 *  -- the resid will not exceed iov_len
	 */
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 75)) | DBG_FUNC_START,
	    (int)uio->uio_offset, *write_length, (int)newEOF, 0, 0);

	assert(vm_map_page_shift(current_map()) >= PAGE_SHIFT);

	max_upl_size = cluster_max_io_size(vp->v_mount, CL_WRITE);

	io_flag = CL_ASYNC | CL_PRESERVE | CL_COMMIT | CL_THROTTLE | CL_DIRECT_IO;

	if (flags & IO_PASSIVE) {
		io_flag |= CL_PASSIVE;
	}

	if (flags & IO_NOCACHE) {
		io_flag |= CL_NOCACHE;
	}

	if (flags & IO_SKIP_ENCRYPTION) {
		io_flag |= CL_ENCRYPTED;
	}

	iostate.io_completed = 0;
	iostate.io_issued = 0;
	iostate.io_error = 0;
	iostate.io_wanted = 0;

	lck_mtx_init(&iostate.io_mtxp, &cl_mtx_grp, LCK_ATTR_NULL);

	mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask;
	devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize;

	if (devblocksize == 1) {
		/*
		 * the AFP client advertises a devblocksize of 1
		 * however, its BLOCKMAP routine maps to physical
		 * blocks that are PAGE_SIZE in size...
		 * therefore we can't ask for I/Os that aren't page aligned
		 * or aren't multiples of PAGE_SIZE in size
		 * by setting devblocksize to PAGE_SIZE, we re-instate
		 * the old behavior we had before the mem_alignment_mask
		 * changes went in...
		 */
		devblocksize = PAGE_SIZE;
	}

	io_align_mask = PAGE_MASK;
	if (min_io_size < MIN_DIRECT_WRITE_SIZE) {
		/* The process has opted into fs blocksize direct io writes */
		assert((min_io_size & (min_io_size - 1)) == 0);
		io_align_mask = min_io_size - 1;
		io_flag |= CL_DIRECT_IO_FSBLKSZ;
	}

	if (uio_iovcnt(uio) > 1) {
		/* vector uio -> take a snapshot so we can rollback if needed */
		if (snapshot_uio) {
			uio_free(snapshot_uio);
			snapshot_uio = NULL;
		}
		snapshot_uio = uio_duplicate(uio);
	}

next_dwrite:
	io_req_size = *write_length;
	iov_base = uio_curriovbase(uio);

	offset_in_file = (u_int32_t)(uio->uio_offset & io_align_mask);
	offset_in_iovbase = (u_int32_t)(iov_base & mem_alignment_mask);

	if (offset_in_file || offset_in_iovbase) {
		/*
		 * one of the 2 important offsets is misaligned
		 * so fire an I/O through the cache for this entire vector
		 */
		goto wait_for_dwrites;
	}
	if (iov_base & (devblocksize - 1)) {
		/*
		 * the offset in memory must be on a device block boundary
		 * so that we can guarantee that we can generate an
		 * I/O that ends on a page boundary in cluster_io
		 */
		goto wait_for_dwrites;
	}

	task_update_logical_writes(current_task(), (io_req_size & ~PAGE_MASK), TASK_WRITE_IMMEDIATE, vp);
	while ((io_req_size >= PAGE_SIZE || io_req_size >= min_io_size) && uio->uio_offset < newEOF && retval == 0) {
		int     throttle_type;

		if ((throttle_type = cluster_is_throttled(vp))) {
			uint32_t max_throttle_size = calculate_max_throttle_size(vp);

			/*
			 * we're in the throttle window, at the very least
			 * we want to limit the size of the I/O we're about
			 * to issue
			 */
			if ((flags & IO_RETURN_ON_THROTTLE) && throttle_type == THROTTLE_NOW) {
				/*
				 * we're in the throttle window and at least 1 I/O
				 * has already been issued by a throttleable thread
				 * in this window, so return with EAGAIN to indicate
				 * to the FS issuing the cluster_write call that it
				 * should now throttle after dropping any locks
				 */
				throttle_info_update_by_mount(vp->v_mount);

				io_throttled = TRUE;
				goto wait_for_dwrites;
			}
			max_vector_size = max_throttle_size;
			max_io_size = max_throttle_size;
		} else {
			max_vector_size = MAX_VECTOR_UPL_SIZE;
			max_io_size = max_upl_size;
		}

		if (first_IO) {
			cluster_syncup(vp, newEOF, callback, callback_arg, callback ? PUSH_SYNC : 0);
			first_IO = 0;
		}
		io_size  = io_req_size & ~io_align_mask;
		iov_base = uio_curriovbase(uio);

		if (io_size > max_io_size) {
			io_size = max_io_size;
		}

		if (useVectorUPL && (iov_base & PAGE_MASK)) {
			/*
			 * We have an iov_base that's not page-aligned.
			 * Issue all I/O's that have been collected within
			 * this Vectored UPL.
			 */
			if (vector_upl_index) {
				retval = vector_cluster_io(vp, vector_upl, vector_upl_offset, v_upl_uio_offset, vector_upl_iosize, io_flag, (buf_t)NULL, &iostate, callback, callback_arg);
				if (retval == ENOTSUP) {
					goto enotsup;
				}
				reset_vector_run_state();
			}

			/*
			 * After this point, if we are using the Vector UPL path and the base is
			 * not page-aligned then the UPL with that base will be the first in the vector UPL.
			 */
		}

		upl_offset = (vm_offset_t)((u_int32_t)iov_base & PAGE_MASK);
		upl_needed_size = (upl_offset + io_size + (PAGE_SIZE - 1)) & ~PAGE_MASK;

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 76)) | DBG_FUNC_START,
		    (int)upl_offset, upl_needed_size, (int)iov_base, io_size, 0);

		vm_map_t map = UIO_SEG_IS_USER_SPACE(uio->uio_segflg) ? current_map() : kernel_map;
		for (force_data_sync = 0; force_data_sync < 3; force_data_sync++) {
			pages_in_pl = 0;
			upl_size = (upl_size_t)upl_needed_size;
			upl_flags = UPL_FILE_IO | UPL_COPYOUT_FROM | UPL_NO_SYNC |
			    UPL_CLEAN_IN_PLACE | UPL_SET_INTERNAL | UPL_SET_LITE | UPL_SET_IO_WIRE;

			kret = vm_map_get_upl(map,
#if HAS_MTE || HAS_MTE_EMULATION_SHIMS
			    vm_memtag_canonicalize(map, (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK))),
#else /* HAS_MTE || HAS_MTE_EMULATION_SHIMS */
			    (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK)),
#endif /* HAS_MTE || HAS_MTE_EMULATION_SHIMS */
			    &upl_size,
			    &upl,
			    NULL,
			    &pages_in_pl,
			    &upl_flags,
			    VM_KERN_MEMORY_FILE,
			    force_data_sync);

			if (kret != KERN_SUCCESS) {
				KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 76)) | DBG_FUNC_END,
				    0, 0, 0, kret, 0);
				/*
				 * failed to get pagelist
				 *
				 * we may have already spun some portion of this request
				 * off as async requests... we need to wait for the I/O
				 * to complete before returning
				 */
				goto wait_for_dwrites;
			}
			pl = UPL_GET_INTERNAL_PAGE_LIST(upl);
			pages_in_pl = upl_size / PAGE_SIZE;

			for (i = 0; i < pages_in_pl; i++) {
				if (!upl_valid_page(pl, i)) {
					break;
				}
			}
			if (i == pages_in_pl) {
				break;
			}

			/*
			 * didn't get all the pages back that we
			 * needed... release this upl and try again
			 */
			ubc_upl_abort(upl, 0);
		}
		if (force_data_sync >= 3) {
			KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 76)) | DBG_FUNC_END,
			    i, pages_in_pl, upl_size, kret, 0);
			/*
			 * for some reason, we couldn't acquire a hold on all
			 * the pages needed in the user's address space
			 *
			 * we may have already spun some portion of this request
			 * off as async requests... we need to wait for the I/O
			 * to complete before returning
			 */
			goto wait_for_dwrites;
		}

		/*
		 * Consider the possibility that upl_size wasn't satisfied.
		 */
		if (upl_size < upl_needed_size) {
			if (upl_size && upl_offset == 0) {
				io_size = upl_size;
			} else {
				io_size = 0;
			}
		}
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 76)) | DBG_FUNC_END,
		    (int)upl_offset, upl_size, (int)iov_base, io_size, 0);

		if (io_size == 0) {
			ubc_upl_abort(upl, 0);
			upl = NULL;
			/*
			 * we may have already spun some portion of this request
			 * off as async requests... we need to wait for the I/O
			 * to complete before returning
			 */
			goto wait_for_dwrites;
		}

		if (useVectorUPL) {
			vm_offset_t end_off = ((iov_base + io_size) & PAGE_MASK);
			if (end_off) {
				issueVectorUPL = 1;
			}
			/*
			 * After this point, if we are using a vector UPL, then
			 * either all the UPL elements end on a page boundary OR
			 * this UPL is the last element because it does not end
			 * on a page boundary.
			 */
		}

		/*
		 * we want push out these writes asynchronously so that we can overlap
		 * the preparation of the next I/O
		 * if there are already too many outstanding writes
		 * wait until some complete before issuing the next
		 */
		if (vp->v_mount->mnt_minsaturationbytecount) {
			bytes_outstanding_limit = vp->v_mount->mnt_minsaturationbytecount;
		} else {
			if (__improbable(os_mul_overflow(max_upl_size, IO_SCALE(vp, 2),
			    &bytes_outstanding_limit) ||
			    (bytes_outstanding_limit > overlapping_write_max))) {
				bytes_outstanding_limit = overlapping_write_max;
			}
		}

		cluster_iostate_wait(&iostate, bytes_outstanding_limit, "cluster_write_direct");

		if (iostate.io_error) {
			/*
			 * one of the earlier writes we issued ran into a hard error
			 * don't issue any more writes, cleanup the UPL
			 * that was just created but not used, then
			 * go wait for all writes that are part of this stream
			 * to complete before returning the error to the caller
			 */
			ubc_upl_abort(upl, 0);
			upl = NULL;

			goto wait_for_dwrites;
		}

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 77)) | DBG_FUNC_START,
		    (int)upl_offset, (int)uio->uio_offset, io_size, io_flag, 0);

		if (!useVectorUPL) {
			retval = cluster_io(vp, upl, upl_offset, uio->uio_offset,
			    io_size, io_flag, (buf_t)NULL, &iostate, callback, callback_arg);
		} else {
			if (!vector_upl_index) {
				vector_upl = vector_upl_create(upl_offset, uio->uio_iovcnt);
				v_upl_uio_offset = uio->uio_offset;
				vector_upl_offset = upl_offset;
			}

			vector_upl_set_subupl(vector_upl, upl, upl_size);
			vector_upl_set_iostate(vector_upl, upl, vector_upl_size, upl_size);
			vector_upl_index++;
			vector_upl_iosize += io_size;
			vector_upl_size += upl_size;

			if (issueVectorUPL || vector_upl_index == vector_upl_max_upls(vector_upl) || vector_upl_size >= max_vector_size) {
				retval = vector_cluster_io(vp, vector_upl, vector_upl_offset, v_upl_uio_offset, vector_upl_iosize, io_flag, (buf_t)NULL, &iostate, callback, callback_arg);
				if (retval != ENOTSUP) {
					reset_vector_run_state();
				}
			}
		}
		if (retval == ENOTSUP) {
enotsup:
			/*
			 * Can't do direct I/O.  Try again with cached I/O.
			 */
//			printf("*******  FBDP %s:%d ENOTSUP io_size 0%x resid 0x%llx\n", __FUNCTION__, __LINE__, io_size, uio_resid(uio));
			io_size = 0;
			if (snapshot_uio) {
				int restore_error;

				/*
				 * We've been collecting UPLs for this vector UPL and
				 * moving the uio along.  We need to undo that so that
				 * the I/O can continue where it actually stopped...
				 */
				restore_error = uio_restore(uio, snapshot_uio);
				assert(!restore_error);
				uio_free(snapshot_uio);
				snapshot_uio = NULL;
			}
			if (vector_upl_index) {
				ubc_upl_abort(vector_upl, 0);
				vector_upl = NULL;
			} else {
				ubc_upl_abort(upl, 0);
				upl = NULL;
			}
			goto wait_for_dwrites;
		}

		/*
		 * update the uio structure to
		 * reflect the I/O that we just issued
		 */
		uio_update(uio, (user_size_t)io_size);

		/*
		 * in case we end up calling through to cluster_write_copy to finish
		 * the tail of this request, we need to update the oldEOF so that we
		 * don't zero-fill the head of a page if we've successfully written
		 * data to that area... 'cluster_write_copy' will zero-fill the head of a
		 * page that is beyond the oldEOF if the write is unaligned... we only
		 * want that to happen for the very first page of the cluster_write,
		 * NOT the first page of each vector making up a multi-vector write.
		 */
		if (uio->uio_offset > oldEOF) {
			oldEOF = uio->uio_offset;
		}

		io_req_size -= io_size;

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 77)) | DBG_FUNC_END,
		    (int)upl_offset, (int)uio->uio_offset, io_req_size, retval, 0);
	} /* end while */

	if (retval == 0 && iostate.io_error == 0 && io_req_size == 0) {
		retval = cluster_io_type(uio, write_type, write_length, min_io_size);

		if (retval == 0 && *write_type == IO_DIRECT) {
			KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 75)) | DBG_FUNC_NONE,
			    (int)uio->uio_offset, *write_length, (int)newEOF, 0, 0);

			goto next_dwrite;
		}
	}

wait_for_dwrites:

	if (retval == 0 && iostate.io_error == 0 && useVectorUPL && vector_upl_index) {
		retval = vector_cluster_io(vp, vector_upl, vector_upl_offset, v_upl_uio_offset, vector_upl_iosize, io_flag, (buf_t)NULL, &iostate, callback, callback_arg);
		reset_vector_run_state();
	}
	/*
	 * make sure all async writes issued as part of this stream
	 * have completed before we return
	 */
	cluster_iostate_wait(&iostate, 0, "cluster_write_direct");

	if (iostate.io_error) {
		retval = iostate.io_error;
	}

	lck_mtx_destroy(&iostate.io_mtxp, &cl_mtx_grp);

	if (io_throttled == TRUE && retval == 0) {
		retval = EAGAIN;
	}

	if (io_req_size && retval == 0) {
		/*
		 * we couldn't handle the tail of this request in DIRECT mode
		 * so fire it through the copy path
		 *
		 * note that flags will never have IO_HEADZEROFILL or IO_TAILZEROFILL set
		 * so we can just pass 0 in for the headOff and tailOff
		 */
		if (uio->uio_offset > oldEOF) {
			oldEOF = uio->uio_offset;
		}

		retval = cluster_write_copy(vp, uio, io_req_size, oldEOF, newEOF, (off_t)0, (off_t)0, flags, callback, callback_arg);

		*write_type = IO_UNKNOWN;
	}

	if (snapshot_uio) {
		uio_free(snapshot_uio);
		snapshot_uio = NULL;
	}

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 75)) | DBG_FUNC_END,
	    (int)uio->uio_offset, io_req_size, retval, 4, 0);

	return retval;
}


static int
cluster_write_contig(vnode_t vp, struct uio *uio, off_t newEOF, int *write_type, u_int32_t *write_length,
    int (*callback)(buf_t, void *), void *callback_arg, int bflag)
{
	upl_page_info_t *pl;
	addr64_t         src_paddr = 0;
	upl_t            upl[MAX_VECTS];
	vm_offset_t      upl_offset;
	u_int32_t        tail_size = 0;
	u_int32_t        io_size;
	u_int32_t        xsize;
	upl_size_t       upl_size;
	vm_size_t        upl_needed_size;
	mach_msg_type_number_t  pages_in_pl;
	upl_control_flags_t upl_flags;
	kern_return_t    kret;
	struct clios     iostate;
	int              error  = 0;
	int              cur_upl = 0;
	int              num_upl = 0;
	int              n;
	user_addr_t      iov_base;
	u_int32_t        devblocksize;
	u_int32_t        mem_alignment_mask;

	/*
	 * When we enter this routine, we know
	 *  -- the io_req_size will not exceed iov_len
	 *  -- the target address is physically contiguous
	 */
	cluster_syncup(vp, newEOF, callback, callback_arg, callback ? PUSH_SYNC : 0);

	devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize;
	mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask;

	iostate.io_completed = 0;
	iostate.io_issued = 0;
	iostate.io_error = 0;
	iostate.io_wanted = 0;

	lck_mtx_init(&iostate.io_mtxp, &cl_mtx_grp, LCK_ATTR_NULL);

next_cwrite:
	io_size = *write_length;

	iov_base = uio_curriovbase(uio);

	upl_offset = (vm_offset_t)((u_int32_t)iov_base & PAGE_MASK);
	upl_needed_size = upl_offset + io_size;

	pages_in_pl = 0;
	upl_size = (upl_size_t)upl_needed_size;
	upl_flags = UPL_FILE_IO | UPL_COPYOUT_FROM | UPL_NO_SYNC |
	    UPL_CLEAN_IN_PLACE | UPL_SET_INTERNAL | UPL_SET_LITE | UPL_SET_IO_WIRE;

	vm_map_t map = UIO_SEG_IS_USER_SPACE(uio->uio_segflg) ? current_map() : kernel_map;
	kret = vm_map_get_upl(map,
#if HAS_MTE || HAS_MTE_EMULATION_SHIMS
	    vm_memtag_canonicalize(map, vm_map_trunc_page(iov_base, vm_map_page_mask(map))),
#else /* HAS_MTE || HAS_MTE_EMULATION_SHIMS */
	    vm_map_trunc_page(iov_base, vm_map_page_mask(map)),
#endif /* HAS_MTE || HAS_MTE_EMULATION_SHIMS */
	    &upl_size, &upl[cur_upl], NULL, &pages_in_pl, &upl_flags, VM_KERN_MEMORY_FILE, 0);

	if (kret != KERN_SUCCESS) {
		/*
		 * failed to get pagelist
		 */
		error = EINVAL;
		goto wait_for_cwrites;
	}
	num_upl++;

	if (!(upl_flags & UPL_PHYS_CONTIG)) {
		/*
		 * The created UPL needs to have the UPL_PHYS_CONTIG flag.
		 */
		error = EINVAL;
		goto wait_for_cwrites;
	}

	/*
	 * Consider the possibility that upl_size wasn't satisfied.
	 */
	if (upl_size < upl_needed_size) {
		/*
		 * This is a failure in the physical memory case.
		 */
		error = EINVAL;
		goto wait_for_cwrites;
	}
	pl = ubc_upl_pageinfo(upl[cur_upl]);

	src_paddr = ((addr64_t)upl_phys_page(pl, 0) << PAGE_SHIFT) + (addr64_t)upl_offset;

	while (((uio->uio_offset & (devblocksize - 1)) || io_size < devblocksize) && io_size) {
		u_int32_t   head_size;

		head_size = devblocksize - (u_int32_t)(uio->uio_offset & (devblocksize - 1));

		if (head_size > io_size) {
			head_size = io_size;
		}

		error = cluster_align_phys_io(vp, uio, src_paddr, head_size, 0, callback, callback_arg);

		if (error) {
			goto wait_for_cwrites;
		}

		upl_offset += head_size;
		src_paddr  += head_size;
		io_size    -= head_size;

		iov_base   += head_size;
	}
	if ((u_int32_t)iov_base & mem_alignment_mask) {
		/*
		 * request doesn't set up on a memory boundary
		 * the underlying DMA engine can handle...
		 * return an error instead of going through
		 * the slow copy path since the intent of this
		 * path is direct I/O from device memory
		 */
		error = EINVAL;
		goto wait_for_cwrites;
	}

	tail_size = io_size & (devblocksize - 1);
	io_size  -= tail_size;

	while (io_size && error == 0) {
		if (io_size > MAX_IO_CONTIG_SIZE) {
			xsize = MAX_IO_CONTIG_SIZE;
		} else {
			xsize = io_size;
		}
		/*
		 * request asynchronously so that we can overlap
		 * the preparation of the next I/O... we'll do
		 * the commit after all the I/O has completed
		 * since its all issued against the same UPL
		 * if there are already too many outstanding writes
		 * wait until some have completed before issuing the next
		 */
		cluster_iostate_wait(&iostate, MAX_IO_CONTIG_SIZE * IO_SCALE(vp, 2), "cluster_write_contig");

		if (iostate.io_error) {
			/*
			 * one of the earlier writes we issued ran into a hard error
			 * don't issue any more writes...
			 * go wait for all writes that are part of this stream
			 * to complete before returning the error to the caller
			 */
			goto wait_for_cwrites;
		}
		/*
		 * issue an asynchronous write to cluster_io
		 */
		error = cluster_io(vp, upl[cur_upl], upl_offset, uio->uio_offset,
		    xsize, CL_DEV_MEMORY | CL_ASYNC | bflag, (buf_t)NULL, (struct clios *)&iostate, callback, callback_arg);

		if (error == 0) {
			/*
			 * The cluster_io write completed successfully,
			 * update the uio structure
			 */
			uio_update(uio, (user_size_t)xsize);

			upl_offset += xsize;
			src_paddr  += xsize;
			io_size    -= xsize;
		}
	}
	if (error == 0 && iostate.io_error == 0 && tail_size == 0 && num_upl < MAX_VECTS) {
		error = cluster_io_type(uio, write_type, write_length, 0);

		if (error == 0 && *write_type == IO_CONTIG) {
			cur_upl++;
			goto next_cwrite;
		}
	} else {
		*write_type = IO_UNKNOWN;
	}

wait_for_cwrites:
	/*
	 * make sure all async writes that are part of this stream
	 * have completed before we proceed
	 */
	cluster_iostate_wait(&iostate, 0, "cluster_write_contig");

	if (iostate.io_error) {
		error = iostate.io_error;
	}

	lck_mtx_destroy(&iostate.io_mtxp, &cl_mtx_grp);

	if (error == 0 && tail_size) {
		error = cluster_align_phys_io(vp, uio, src_paddr, tail_size, 0, callback, callback_arg);
	}

	for (n = 0; n < num_upl; n++) {
		/*
		 * just release our hold on each physically contiguous
		 * region without changing any state
		 */
		ubc_upl_abort(upl[n], 0);
	}

	return error;
}


/*
 * need to avoid a race between an msync of a range of pages dirtied via mmap
 * vs a filesystem such as HFS deciding to write a 'hole' to disk via cluster_write's
 * zerofill mechanism before it has seen the VNOP_PAGEOUTs for the pages being msync'd
 *
 * we should never force-zero-fill pages that are already valid in the cache...
 * the entire page contains valid data (either from disk, zero-filled or dirtied
 * via an mmap) so we can only do damage by trying to zero-fill
 *
 */
static int
cluster_zero_range(upl_t upl, upl_page_info_t *pl, int flags, int io_offset, off_t zero_off, off_t upl_f_offset, int bytes_to_zero)
{
	int zero_pg_index;
	boolean_t need_cluster_zero = TRUE;

	if ((flags & (IO_NOZEROVALID | IO_NOZERODIRTY))) {
		bytes_to_zero = min(bytes_to_zero, PAGE_SIZE - (int)(zero_off & PAGE_MASK_64));
		zero_pg_index = (int)((zero_off - upl_f_offset) / PAGE_SIZE_64);

		if (upl_valid_page(pl, zero_pg_index)) {
			/*
			 * never force zero valid pages - dirty or clean
			 * we'll leave these in the UPL for cluster_write_copy to deal with
			 */
			need_cluster_zero = FALSE;
		}
	}
	if (need_cluster_zero == TRUE) {
		cluster_zero(upl, io_offset, bytes_to_zero, NULL);
	}

	return bytes_to_zero;
}


void
cluster_update_state(vnode_t vp, vm_object_offset_t s_offset, vm_object_offset_t e_offset, boolean_t vm_initiated)
{
	struct cl_extent cl;
	boolean_t first_pass = TRUE;

	assert(s_offset < e_offset);
	assert((s_offset & PAGE_MASK_64) == 0);
	assert((e_offset & PAGE_MASK_64) == 0);

	cl.b_addr = (daddr64_t)(s_offset / PAGE_SIZE_64);
	cl.e_addr = (daddr64_t)(e_offset / PAGE_SIZE_64);

	cluster_update_state_internal(vp, &cl, 0, TRUE, &first_pass, s_offset, (int)(e_offset - s_offset),
	    vp->v_un.vu_ubcinfo->ui_size, NULL, NULL, vm_initiated);
}


static void
cluster_update_state_internal(vnode_t vp, struct cl_extent *cl, int flags, boolean_t defer_writes,
    boolean_t *first_pass, off_t write_off, int write_cnt, off_t newEOF,
    int (*callback)(buf_t, void *), void *callback_arg, boolean_t vm_initiated)
{
	struct cl_writebehind *wbp;
	int     cl_index;
	int     ret_cluster_try_push;
	u_int   max_cluster_pgcount;


	max_cluster_pgcount = MAX_CLUSTER_SIZE(vp) / PAGE_SIZE;

	/*
	 * take the lock to protect our accesses
	 * of the writebehind and sparse cluster state
	 */
	wbp = cluster_get_wbp(vp, CLW_ALLOCATE | CLW_RETURNLOCKED);

	if (wbp->cl_scmap) {
		if (!(flags & IO_NOCACHE)) {
			/*
			 * we've fallen into the sparse
			 * cluster method of delaying dirty pages
			 */
			sparse_cluster_add(wbp, &(wbp->cl_scmap), vp, cl, newEOF, callback, callback_arg, vm_initiated);

			lck_mtx_unlock(&wbp->cl_lockw);
			return;
		}
		/*
		 * must have done cached writes that fell into
		 * the sparse cluster mechanism... we've switched
		 * to uncached writes on the file, so go ahead
		 * and push whatever's in the sparse map
		 * and switch back to normal clustering
		 */
		wbp->cl_number = 0;

		sparse_cluster_push(wbp, &(wbp->cl_scmap), vp, newEOF, PUSH_ALL, 0, callback, callback_arg, vm_initiated);
		/*
		 * no clusters of either type present at this point
		 * so just go directly to start_new_cluster since
		 * we know we need to delay this I/O since we've
		 * already released the pages back into the cache
		 * to avoid the deadlock with sparse_cluster_push
		 */
		goto start_new_cluster;
	}
	if (*first_pass == TRUE) {
		if (write_off == wbp->cl_last_write) {
			wbp->cl_seq_written += write_cnt;
		} else {
			wbp->cl_seq_written = write_cnt;
		}

		wbp->cl_last_write = write_off + write_cnt;

		*first_pass = FALSE;
	}
	if (wbp->cl_number == 0) {
		/*
		 * no clusters currently present
		 */
		goto start_new_cluster;
	}

	for (cl_index = 0; cl_index < wbp->cl_number; cl_index++) {
		/*
		 * check each cluster that we currently hold
		 * try to merge some or all of this write into
		 * one or more of the existing clusters... if
		 * any portion of the write remains, start a
		 * new cluster
		 */
		if (cl->b_addr >= wbp->cl_clusters[cl_index].b_addr) {
			/*
			 * the current write starts at or after the current cluster
			 */
			if (cl->e_addr <= (wbp->cl_clusters[cl_index].b_addr + max_cluster_pgcount)) {
				/*
				 * we have a write that fits entirely
				 * within the existing cluster limits
				 */
				if (cl->e_addr > wbp->cl_clusters[cl_index].e_addr) {
					/*
					 * update our idea of where the cluster ends
					 */
					wbp->cl_clusters[cl_index].e_addr = cl->e_addr;
				}
				break;
			}
			if (cl->b_addr < (wbp->cl_clusters[cl_index].b_addr + max_cluster_pgcount)) {
				/*
				 * we have a write that starts in the middle of the current cluster
				 * but extends beyond the cluster's limit... we know this because
				 * of the previous checks
				 * we'll extend the current cluster to the max
				 * and update the b_addr for the current write to reflect that
				 * the head of it was absorbed into this cluster...
				 * note that we'll always have a leftover tail in this case since
				 * full absorbtion would have occurred in the clause above
				 */
				wbp->cl_clusters[cl_index].e_addr = wbp->cl_clusters[cl_index].b_addr + max_cluster_pgcount;

				cl->b_addr = wbp->cl_clusters[cl_index].e_addr;
			}
			/*
			 * we come here for the case where the current write starts
			 * beyond the limit of the existing cluster or we have a leftover
			 * tail after a partial absorbtion
			 *
			 * in either case, we'll check the remaining clusters before
			 * starting a new one
			 */
		} else {
			/*
			 * the current write starts in front of the cluster we're currently considering
			 */
			if ((wbp->cl_clusters[cl_index].e_addr - cl->b_addr) <= max_cluster_pgcount) {
				/*
				 * we can just merge the new request into
				 * this cluster and leave it in the cache
				 * since the resulting cluster is still
				 * less than the maximum allowable size
				 */
				wbp->cl_clusters[cl_index].b_addr = cl->b_addr;

				if (cl->e_addr > wbp->cl_clusters[cl_index].e_addr) {
					/*
					 * the current write completely
					 * envelops the existing cluster and since
					 * each write is limited to at most max_cluster_pgcount pages
					 * we can just use the start and last blocknos of the write
					 * to generate the cluster limits
					 */
					wbp->cl_clusters[cl_index].e_addr = cl->e_addr;
				}
				break;
			}
			/*
			 * if we were to combine this write with the current cluster
			 * we would exceed the cluster size limit.... so,
			 * let's see if there's any overlap of the new I/O with
			 * the cluster we're currently considering... in fact, we'll
			 * stretch the cluster out to it's full limit and see if we
			 * get an intersection with the current write
			 *
			 */
			if (cl->e_addr > wbp->cl_clusters[cl_index].e_addr - max_cluster_pgcount) {
				/*
				 * the current write extends into the proposed cluster
				 * clip the length of the current write after first combining it's
				 * tail with the newly shaped cluster
				 */
				wbp->cl_clusters[cl_index].b_addr = wbp->cl_clusters[cl_index].e_addr - max_cluster_pgcount;

				cl->e_addr = wbp->cl_clusters[cl_index].b_addr;
			}
			/*
			 * if we get here, there was no way to merge
			 * any portion of this write with this cluster
			 * or we could only merge part of it which
			 * will leave a tail...
			 * we'll check the remaining clusters before starting a new one
			 */
		}
	}
	if (cl_index < wbp->cl_number) {
		/*
		 * we found an existing cluster(s) that we
		 * could entirely merge this I/O into
		 */
		goto delay_io;
	}

	if (defer_writes == FALSE &&
	    wbp->cl_number == MAX_CLUSTERS &&
	    wbp->cl_seq_written >= (MAX_CLUSTERS * (max_cluster_pgcount * PAGE_SIZE))) {
		uint32_t        n;

		if (vp->v_mount->mnt_minsaturationbytecount) {
			n = vp->v_mount->mnt_minsaturationbytecount / MAX_CLUSTER_SIZE(vp);

			if (n > MAX_CLUSTERS) {
				n = MAX_CLUSTERS;
			}
		} else {
			n = 0;
		}

		if (n == 0) {
			if (disk_conditioner_mount_is_ssd(vp->v_mount)) {
				n = WRITE_BEHIND_SSD;
			} else {
				n = WRITE_BEHIND;
			}
		}
		while (n--) {
			cluster_try_push(wbp, vp, newEOF, 0, 0, callback, callback_arg, NULL, vm_initiated);
		}
	}
	if (wbp->cl_number < MAX_CLUSTERS) {
		/*
		 * we didn't find an existing cluster to
		 * merge into, but there's room to start
		 * a new one
		 */
		goto start_new_cluster;
	}
	/*
	 * no exisitng cluster to merge with and no
	 * room to start a new one... we'll try
	 * pushing one of the existing ones... if none of
	 * them are able to be pushed, we'll switch
	 * to the sparse cluster mechanism
	 * cluster_try_push updates cl_number to the
	 * number of remaining clusters... and
	 * returns the number of currently unused clusters
	 */
	ret_cluster_try_push = 0;

	/*
	 * if writes are not deferred, call cluster push immediately
	 */
	if (defer_writes == FALSE) {
		ret_cluster_try_push = cluster_try_push(wbp, vp, newEOF, (flags & IO_NOCACHE) ? 0 : PUSH_DELAY, 0, callback, callback_arg, NULL, vm_initiated);
	}
	/*
	 * execute following regardless of writes being deferred or not
	 */
	if (ret_cluster_try_push == 0) {
		/*
		 * no more room in the normal cluster mechanism
		 * so let's switch to the more expansive but expensive
		 * sparse mechanism....
		 */
		sparse_cluster_switch(wbp, vp, newEOF, callback, callback_arg, vm_initiated);
		sparse_cluster_add(wbp, &(wbp->cl_scmap), vp, cl, newEOF, callback, callback_arg, vm_initiated);

		lck_mtx_unlock(&wbp->cl_lockw);
		return;
	}
start_new_cluster:
	wbp->cl_clusters[wbp->cl_number].b_addr = cl->b_addr;
	wbp->cl_clusters[wbp->cl_number].e_addr = cl->e_addr;

	wbp->cl_clusters[wbp->cl_number].io_flags = 0;

	if (flags & IO_NOCACHE) {
		wbp->cl_clusters[wbp->cl_number].io_flags |= CLW_IONOCACHE;
	}

	if (flags & IO_PASSIVE) {
		wbp->cl_clusters[wbp->cl_number].io_flags |= CLW_IOPASSIVE;
	}

	wbp->cl_number++;
delay_io:
	lck_mtx_unlock(&wbp->cl_lockw);
	return;
}


static int
cluster_write_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t oldEOF, off_t newEOF, off_t headOff,
    off_t tailOff, int flags, int (*callback)(buf_t, void *), void *callback_arg)
{
	upl_page_info_t *pl;
	upl_t            upl;
	vm_offset_t      upl_offset = 0;
	vm_size_t        upl_size;
	off_t            upl_f_offset;
	int              pages_in_upl;
	int              start_offset;
	int              xfer_resid;
	int              io_size;
	int              io_offset;
	int              bytes_to_zero;
	int              bytes_to_move;
	kern_return_t    kret;
	int              retval = 0;
	int              io_resid;
	long long        total_size;
	long long        zero_cnt;
	off_t            zero_off;
	long long        zero_cnt1;
	off_t            zero_off1;
	off_t            write_off = 0;
	int              write_cnt = 0;
	boolean_t        first_pass = FALSE;
	struct cl_extent cl;
	int              bflag;
	u_int            max_io_size;

	if (uio) {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 40)) | DBG_FUNC_START,
		    (int)uio->uio_offset, io_req_size, (int)oldEOF, (int)newEOF, 0);

		io_resid = io_req_size;
	} else {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 40)) | DBG_FUNC_START,
		    0, 0, (int)oldEOF, (int)newEOF, 0);

		io_resid = 0;
	}
	if (flags & IO_PASSIVE) {
		bflag = CL_PASSIVE;
	} else {
		bflag = 0;
	}
	if (flags & IO_NOCACHE) {
		bflag |= CL_NOCACHE;
	}

	if (flags & IO_SKIP_ENCRYPTION) {
		bflag |= CL_ENCRYPTED;
	}

	zero_cnt  = 0;
	zero_cnt1 = 0;
	zero_off  = 0;
	zero_off1 = 0;

	max_io_size = cluster_max_io_size(vp->v_mount, CL_WRITE);

	if (flags & IO_HEADZEROFILL) {
		/*
		 * some filesystems (HFS is one) don't support unallocated holes within a file...
		 * so we zero fill the intervening space between the old EOF and the offset
		 * where the next chunk of real data begins.... ftruncate will also use this
		 * routine to zero fill to the new EOF when growing a file... in this case, the
		 * uio structure will not be provided
		 */
		if (uio) {
			if (headOff < uio->uio_offset) {
				zero_cnt = uio->uio_offset - headOff;
				zero_off = headOff;
			}
		} else if (headOff < newEOF) {
			zero_cnt = newEOF - headOff;
			zero_off = headOff;
		}
	} else {
		if (uio && uio->uio_offset > oldEOF) {
			zero_off = uio->uio_offset & ~PAGE_MASK_64;

			if (zero_off >= oldEOF) {
				zero_cnt = uio->uio_offset - zero_off;

				flags |= IO_HEADZEROFILL;
			}
		}
	}
	if (flags & IO_TAILZEROFILL) {
		if (uio) {
			zero_off1 = uio->uio_offset + io_req_size;

			if (zero_off1 < tailOff) {
				zero_cnt1 = tailOff - zero_off1;
			}
		}
	} else {
		if (uio && newEOF > oldEOF) {
			zero_off1 = uio->uio_offset + io_req_size;

			if (zero_off1 == newEOF && (zero_off1 & PAGE_MASK_64)) {
				zero_cnt1 = PAGE_SIZE_64 - (zero_off1 & PAGE_MASK_64);

				flags |= IO_TAILZEROFILL;
			}
		}
	}
	if (zero_cnt == 0 && uio == (struct uio *) 0) {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 40)) | DBG_FUNC_END,
		    retval, 0, 0, 0, 0);
		return 0;
	}
	if (uio) {
		write_off = uio->uio_offset;
		write_cnt = (int)uio_resid(uio);
		/*
		 * delay updating the sequential write info
		 * in the control block until we've obtained
		 * the lock for it
		 */
		first_pass = TRUE;
	}
	while ((total_size = (io_resid + zero_cnt + zero_cnt1)) && retval == 0) {
		/*
		 * for this iteration of the loop, figure out where our starting point is
		 */
		if (zero_cnt) {
			start_offset = (int)(zero_off & PAGE_MASK_64);
			upl_f_offset = zero_off - start_offset;
		} else if (io_resid) {
			start_offset = (int)(uio->uio_offset & PAGE_MASK_64);
			upl_f_offset = uio->uio_offset - start_offset;
		} else {
			start_offset = (int)(zero_off1 & PAGE_MASK_64);
			upl_f_offset = zero_off1 - start_offset;
		}
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 46)) | DBG_FUNC_NONE,
		    (int)zero_off, (int)zero_cnt, (int)zero_off1, (int)zero_cnt1, 0);

		if (total_size > max_io_size) {
			total_size = max_io_size;
		}

		cl.b_addr = (daddr64_t)(upl_f_offset / PAGE_SIZE_64);

		if (uio && ((flags & (IO_SYNC | IO_HEADZEROFILL | IO_TAILZEROFILL)) == 0)) {
			/*
			 * assumption... total_size <= io_resid
			 * because IO_HEADZEROFILL and IO_TAILZEROFILL not set
			 */
			if ((start_offset + total_size) > max_io_size) {
				total_size = max_io_size - start_offset;
			}
			xfer_resid = (int)total_size;

			retval = cluster_copy_ubc_data_internal(vp, uio, &xfer_resid, 1, 1);

			if (retval) {
				break;
			}

			io_resid    -= (total_size - xfer_resid);
			total_size   = xfer_resid;
			start_offset = (int)(uio->uio_offset & PAGE_MASK_64);
			upl_f_offset = uio->uio_offset - start_offset;

			if (total_size == 0) {
				if (start_offset) {
					/*
					 * the write did not finish on a page boundary
					 * which will leave upl_f_offset pointing to the
					 * beginning of the last page written instead of
					 * the page beyond it... bump it in this case
					 * so that the cluster code records the last page
					 * written as dirty
					 */
					upl_f_offset += PAGE_SIZE_64;
				}
				upl_size = 0;

				goto check_cluster;
			}
		}
		/*
		 * compute the size of the upl needed to encompass
		 * the requested write... limit each call to cluster_io
		 * to the maximum UPL size... cluster_io will clip if
		 * this exceeds the maximum io_size for the device,
		 * make sure to account for
		 * a starting offset that's not page aligned
		 */
		upl_size = (start_offset + total_size + (PAGE_SIZE - 1)) & ~PAGE_MASK;

		if (upl_size > max_io_size) {
			upl_size = max_io_size;
		}

		pages_in_upl = (int)(upl_size / PAGE_SIZE);
		io_size      = (int)(upl_size - start_offset);

		if ((long long)io_size > total_size) {
			io_size = (int)total_size;
		}

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 41)) | DBG_FUNC_START, upl_size, io_size, total_size, 0, 0);


		/*
		 * Gather the pages from the buffer cache.
		 * The UPL_WILL_MODIFY flag lets the UPL subsystem know
		 * that we intend to modify these pages.
		 */
		kret = ubc_create_upl_kernel(vp,
		    upl_f_offset,
		    (int)upl_size,
		    &upl,
		    &pl,
		    UPL_SET_LITE | ((uio != NULL && (uio->uio_flags & UIO_FLAGS_IS_COMPRESSED_FILE)) ? 0 : UPL_WILL_MODIFY),
		    VM_KERN_MEMORY_FILE);
		if (kret != KERN_SUCCESS) {
			panic("cluster_write_copy: failed to get pagelist");
		}

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 41)) | DBG_FUNC_END,
		    upl, (int)upl_f_offset, start_offset, 0, 0);

		if (start_offset && upl_f_offset < oldEOF && !upl_valid_page(pl, 0)) {
			int   read_size;

			/*
			 * we're starting in the middle of the first page of the upl
			 * and the page isn't currently valid, so we're going to have
			 * to read it in first... this is a synchronous operation
			 */
			read_size = PAGE_SIZE;

			if ((upl_f_offset + read_size) > oldEOF) {
				read_size = (int)(oldEOF - upl_f_offset);
			}

			retval = cluster_io(vp, upl, 0, upl_f_offset, read_size,
			    CL_READ | bflag, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg);
			if (retval) {
				/*
				 * we had an error during the read which causes us to abort
				 * the current cluster_write request... before we do, we need
				 * to release the rest of the pages in the upl without modifying
				 * there state and mark the failed page in error
				 */
				ubc_upl_abort_range(upl, 0, PAGE_SIZE, UPL_ABORT_DUMP_PAGES | UPL_ABORT_FREE_ON_EMPTY);

				if (upl_size > PAGE_SIZE) {
					ubc_upl_abort_range(upl, 0, (upl_size_t)upl_size,
					    UPL_ABORT_FREE_ON_EMPTY);
				}

				KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 45)) | DBG_FUNC_NONE,
				    upl, 0, 0, retval, 0);
				break;
			}
		}
		if ((start_offset == 0 || upl_size > PAGE_SIZE) && ((start_offset + io_size) & PAGE_MASK)) {
			/*
			 * the last offset we're writing to in this upl does not end on a page
			 * boundary... if it's not beyond the old EOF, then we'll also need to
			 * pre-read this page in if it isn't already valid
			 */
			upl_offset = upl_size - PAGE_SIZE;

			if ((upl_f_offset + start_offset + io_size) < oldEOF &&
			    !upl_valid_page(pl, (int)(upl_offset / PAGE_SIZE))) {
				int   read_size;

				read_size = PAGE_SIZE;

				if ((off_t)(upl_f_offset + upl_offset + read_size) > oldEOF) {
					read_size = (int)(oldEOF - (upl_f_offset + upl_offset));
				}

				retval = cluster_io(vp, upl, upl_offset, upl_f_offset + upl_offset, read_size,
				    CL_READ | bflag, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg);
				if (retval) {
					/*
					 * we had an error during the read which causes us to abort
					 * the current cluster_write request... before we do, we
					 * need to release the rest of the pages in the upl without
					 * modifying there state and mark the failed page in error
					 */
					ubc_upl_abort_range(upl, (upl_offset_t)upl_offset, PAGE_SIZE, UPL_ABORT_DUMP_PAGES | UPL_ABORT_FREE_ON_EMPTY);

					if (upl_size > PAGE_SIZE) {
						ubc_upl_abort_range(upl, 0, (upl_size_t)upl_size, UPL_ABORT_FREE_ON_EMPTY);
					}

					KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 45)) | DBG_FUNC_NONE,
					    upl, 0, 0, retval, 0);
					break;
				}
			}
		}
		xfer_resid = io_size;
		io_offset = start_offset;

		while (zero_cnt && xfer_resid) {
			if (zero_cnt < (long long)xfer_resid) {
				bytes_to_zero = (int)zero_cnt;
			} else {
				bytes_to_zero = xfer_resid;
			}

			bytes_to_zero = cluster_zero_range(upl, pl, flags, io_offset, zero_off, upl_f_offset, bytes_to_zero);

			xfer_resid -= bytes_to_zero;
			zero_cnt   -= bytes_to_zero;
			zero_off   += bytes_to_zero;
			io_offset  += bytes_to_zero;
		}
		if (xfer_resid && io_resid) {
			u_int32_t  io_requested;

			bytes_to_move = min(io_resid, xfer_resid);
			io_requested = bytes_to_move;

			retval = cluster_copy_upl_data(uio, upl, io_offset, (int *)&io_requested);

			if (retval) {
				ubc_upl_abort_range(upl, 0, (upl_size_t)upl_size, UPL_ABORT_FREE_ON_EMPTY);

				KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 45)) | DBG_FUNC_NONE,
				    upl, 0, 0, retval, 0);
			} else {
				io_resid   -= bytes_to_move;
				xfer_resid -= bytes_to_move;
				io_offset  += bytes_to_move;
			}
		}
		while (xfer_resid && zero_cnt1 && retval == 0) {
			if (zero_cnt1 < (long long)xfer_resid) {
				bytes_to_zero = (int)zero_cnt1;
			} else {
				bytes_to_zero = xfer_resid;
			}

			bytes_to_zero = cluster_zero_range(upl, pl, flags, io_offset, zero_off1, upl_f_offset, bytes_to_zero);

			xfer_resid -= bytes_to_zero;
			zero_cnt1  -= bytes_to_zero;
			zero_off1  += bytes_to_zero;
			io_offset  += bytes_to_zero;
		}
		if (retval == 0) {
			int do_zeroing = 1;

			io_size += start_offset;

			/* Force more restrictive zeroing behavior only on APFS */
			if ((vnode_tag(vp) == VT_APFS) && (newEOF < oldEOF)) {
				do_zeroing = 0;
			}

			if (do_zeroing && (upl_f_offset + io_size) >= newEOF && (u_int)io_size < upl_size) {
				/*
				 * if we're extending the file with this write
				 * we'll zero fill the rest of the page so that
				 * if the file gets extended again in such a way as to leave a
				 * hole starting at this EOF, we'll have zero's in the correct spot
				 */
				cluster_zero(upl, io_size, (int)(upl_size - io_size), NULL);
			}
			/*
			 * release the upl now if we hold one since...
			 * 1) pages in it may be present in the sparse cluster map
			 *    and may span 2 separate buckets there... if they do and
			 *    we happen to have to flush a bucket to make room and it intersects
			 *    this upl, a deadlock may result on page BUSY
			 * 2) we're delaying the I/O... from this point forward we're just updating
			 *    the cluster state... no need to hold the pages, so commit them
			 * 3) IO_SYNC is set...
			 *    because we had to ask for a UPL that provides currenty non-present pages, the
			 *    UPL has been automatically set to clear the dirty flags (both software and hardware)
			 *    upon committing it... this is not the behavior we want since it's possible for
			 *    pages currently present as part of a mapped file to be dirtied while the I/O is in flight.
			 *    we'll pick these pages back up later with the correct behavior specified.
			 * 4) we don't want to hold pages busy in a UPL and then block on the cluster lock... if a flush
			 *    of this vnode is in progress, we will deadlock if the pages being flushed intersect the pages
			 *    we hold since the flushing context is holding the cluster lock.
			 */
			ubc_upl_commit_range(upl, 0, (upl_size_t)upl_size,
			    UPL_COMMIT_SET_DIRTY | UPL_COMMIT_INACTIVATE | UPL_COMMIT_FREE_ON_EMPTY);
check_cluster:
			/*
			 * calculate the last logical block number
			 * that this delayed I/O encompassed
			 */
			cl.e_addr = (daddr64_t)((upl_f_offset + (off_t)upl_size) / PAGE_SIZE_64);

			if (flags & IO_SYNC) {
				/*
				 * if the IO_SYNC flag is set than we need to bypass
				 * any clustering and immediately issue the I/O
				 *
				 * we don't hold the lock at this point
				 *
				 * we've already dropped the current upl, so pick it back up with COPYOUT_FROM set
				 * so that we correctly deal with a change in state of the hardware modify bit...
				 * we do this via cluster_push_now... by passing along the IO_SYNC flag, we force
				 * cluster_push_now to wait until all the I/Os have completed... cluster_push_now is also
				 * responsible for generating the correct sized I/O(s)
				 */
				retval = cluster_push_now(vp, &cl, newEOF, flags, callback, callback_arg, FALSE);
			} else {
				boolean_t defer_writes = FALSE;

				if (vfs_flags(vp->v_mount) & MNT_DEFWRITE) {
					defer_writes = TRUE;
				}

				cluster_update_state_internal(vp, &cl, flags, defer_writes, &first_pass,
				    write_off, write_cnt, newEOF, callback, callback_arg, FALSE);
			}
		}
	}
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 40)) | DBG_FUNC_END, retval, 0, io_resid, 0, 0);

	return retval;
}



int
cluster_read(vnode_t vp, struct uio *uio, off_t filesize, int xflags)
{
	return cluster_read_ext(vp, uio, filesize, xflags, NULL, NULL);
}


int
cluster_read_ext(vnode_t vp, struct uio *uio, off_t filesize, int xflags, int (*callback)(buf_t, void *), void *callback_arg)
{
	int             retval = 0;
	int             flags;
	user_ssize_t    cur_resid;
	u_int32_t       io_size;
	u_int32_t       read_length = 0;
	int             read_type = IO_COPY;
	bool            check_io_type;

	flags = xflags;

	if (vp->v_flag & VNOCACHE_DATA) {
		flags |= IO_NOCACHE;
	}
	if ((vp->v_flag & VRAOFF) || speculative_reads_disabled) {
		flags |= IO_RAOFF;
	}

	if (flags & IO_SKIP_ENCRYPTION) {
		flags |= IO_ENCRYPTED;
	}

	/*
	 * do a read through the cache if one of the following is true....
	 *   NOCACHE is not true
	 *   the uio request doesn't target USERSPACE (unless IO_NOCACHE_SYSSPACE is also set)
	 * Alternatively, if IO_ENCRYPTED is set, then we want to bypass the cache as well.
	 * Reading encrypted data from a CP filesystem should never result in the data touching
	 * the UBC.
	 *
	 * otherwise, find out if we want the direct or contig variant for
	 * the first vector in the uio request
	 */
	check_io_type = false;
	if (flags & IO_NOCACHE) {
		if (UIO_SEG_IS_USER_SPACE(uio->uio_segflg)) {
			/*
			 * no-cache to user-space: ok to consider IO_DIRECT.
			 */
			check_io_type = true;
		} else if (uio->uio_segflg == UIO_SYSSPACE &&
		    (flags & IO_NOCACHE_SYSSPACE)) {
			/*
			 * no-cache to kernel-space but w/ IO_NOCACHE_SYSSPACE:
			 * ok to consider IO_DIRECT.
			 * The caller should make sure to target kernel buffer
			 * that is backed by regular anonymous memory (i.e.
			 * not backed by the kernel object or an external
			 * memory manager like device memory or a file).
			 */
			check_io_type = true;
		}
	} else if (flags & IO_ENCRYPTED) {
		check_io_type = true;
	}
	if (check_io_type) {
		retval = cluster_io_type(uio, &read_type, &read_length, 0);
	}

	while ((cur_resid = uio_resid(uio)) && uio->uio_offset < filesize && retval == 0) {
		switch (read_type) {
		case IO_COPY:
			/*
			 * make sure the uio_resid isn't too big...
			 * internally, we want to handle all of the I/O in
			 * chunk sizes that fit in a 32 bit int
			 */
			if (cur_resid > (user_ssize_t)(MAX_IO_REQUEST_SIZE)) {
				io_size = MAX_IO_REQUEST_SIZE;
			} else {
				io_size = (u_int32_t)cur_resid;
			}

			retval = cluster_read_copy(vp, uio, io_size, filesize, flags, callback, callback_arg);
			break;

		case IO_DIRECT:
			retval = cluster_read_direct(vp, uio, filesize, &read_type, &read_length, flags, callback, callback_arg);
			break;

		case IO_CONTIG:
			retval = cluster_read_contig(vp, uio, filesize, &read_type, &read_length, callback, callback_arg, flags);
			break;

		case IO_UNKNOWN:
			retval = cluster_io_type(uio, &read_type, &read_length, 0);
			break;
		}
	}
	return retval;
}



static void
cluster_read_upl_release(upl_t upl, int start_pg, int last_pg, int take_reference)
{
	int range;
	int abort_flags = UPL_ABORT_FREE_ON_EMPTY;

	if ((range = last_pg - start_pg)) {
		if (take_reference) {
			abort_flags |= UPL_ABORT_REFERENCE;
		}

		ubc_upl_abort_range(upl, start_pg * PAGE_SIZE, range * PAGE_SIZE, abort_flags);
	}
}


static int
cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t filesize, int flags, int (*callback)(buf_t, void *), void *callback_arg)
{
	upl_page_info_t *pl;
	upl_t            upl = NULL;
	vm_offset_t      upl_offset;
	u_int32_t        upl_size;
	off_t            upl_f_offset;
	int              start_offset;
	int              start_pg;
	int              last_pg;
	int              uio_last = 0;
	int              pages_in_upl;
	off_t            max_size;
	off_t            last_ioread_offset;
	off_t            last_request_offset;
	kern_return_t    kret;
	int              error  = 0;
	int              retval = 0;
	u_int32_t        size_of_prefetch;
	u_int32_t        xsize;
	u_int32_t        io_size;
	u_int32_t        max_rd_size;
	u_int32_t        max_io_size;
	u_int32_t        max_prefetch;
	u_int            rd_ahead_enabled = 1;
	u_int            prefetch_enabled = 1;
	struct cl_readahead *   rap;
	struct clios            iostate;
	struct cl_extent        extent;
	int              bflag;
	int              take_reference = 1;
	int              policy = IOPOL_DEFAULT;
	boolean_t        iolock_inited = FALSE;

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 32)) | DBG_FUNC_START,
	    (int)uio->uio_offset, io_req_size, (int)filesize, flags, 0);

	if (flags & IO_ENCRYPTED) {
		panic("encrypted blocks will hit UBC!");
	}

	policy = throttle_get_io_policy(NULL);

	if (policy == THROTTLE_LEVEL_TIER3 || policy == THROTTLE_LEVEL_TIER2 || (flags & IO_NOCACHE)) {
		take_reference = 0;
	}

	if (flags & IO_PASSIVE) {
		bflag = CL_PASSIVE;
	} else {
		bflag = 0;
	}

	if (flags & IO_NOCACHE) {
		bflag |= CL_NOCACHE;
	}

	if (flags & IO_SKIP_ENCRYPTION) {
		bflag |= CL_ENCRYPTED;
	}

	max_io_size = cluster_max_io_size(vp->v_mount, CL_READ);
	max_prefetch = cluster_max_prefetch(vp, max_io_size, prefetch_max);
	max_rd_size = max_prefetch;

	last_request_offset = uio->uio_offset + io_req_size;

	if (last_request_offset > filesize) {
		last_request_offset = filesize;
	}

	if ((flags & (IO_RAOFF | IO_NOCACHE)) || ((last_request_offset & ~PAGE_MASK_64) == (uio->uio_offset & ~PAGE_MASK_64))) {
		rd_ahead_enabled = 0;
		rap = NULL;
	} else {
		if (cluster_is_throttled(vp)) {
			/*
			 * we're in the throttle window, at the very least
			 * we want to limit the size of the I/O we're about
			 * to issue
			 */
			rd_ahead_enabled = 0;
			prefetch_enabled = 0;

			max_rd_size = calculate_max_throttle_size(vp);
		}
		if ((rap = cluster_get_rap(vp)) == NULL) {
			rd_ahead_enabled = 0;
		} else {
			extent.b_addr = uio->uio_offset / PAGE_SIZE_64;
			extent.e_addr = (last_request_offset - 1) / PAGE_SIZE_64;
		}
	}
	if (rap != NULL && rap->cl_ralen && (rap->cl_lastr == extent.b_addr || (rap->cl_lastr + 1) == extent.b_addr)) {
		/*
		 * determine if we already have a read-ahead in the pipe courtesy of the
		 * last read systemcall that was issued...
		 * if so, pick up it's extent to determine where we should start
		 * with respect to any read-ahead that might be necessary to
		 * garner all the data needed to complete this read systemcall
		 */
		last_ioread_offset = (rap->cl_maxra * PAGE_SIZE_64) + PAGE_SIZE_64;

		if (last_ioread_offset < uio->uio_offset) {
			last_ioread_offset = (off_t)0;
		} else if (last_ioread_offset > last_request_offset) {
			last_ioread_offset = last_request_offset;
		}
	} else {
		last_ioread_offset = (off_t)0;
	}

	while (io_req_size && uio->uio_offset < filesize && retval == 0) {
		max_size = filesize - uio->uio_offset;
		bool leftover_upl_aborted = false;

		if ((off_t)(io_req_size) < max_size) {
			io_size = io_req_size;
		} else {
			io_size = (u_int32_t)max_size;
		}

		if (!(flags & IO_NOCACHE)) {
			while (io_size) {
				u_int32_t io_resid;
				u_int32_t io_requested;

				/*
				 * if we keep finding the pages we need already in the cache, then
				 * don't bother to call cluster_read_prefetch since it costs CPU cycles
				 * to determine that we have all the pages we need... once we miss in
				 * the cache and have issued an I/O, than we'll assume that we're likely
				 * to continue to miss in the cache and it's to our advantage to try and prefetch
				 */
				if (last_request_offset && last_ioread_offset && (size_of_prefetch = (u_int32_t)(last_request_offset - last_ioread_offset))) {
					if ((last_ioread_offset - uio->uio_offset) <= max_rd_size && prefetch_enabled) {
						/*
						 * we've already issued I/O for this request and
						 * there's still work to do and
						 * our prefetch stream is running dry, so issue a
						 * pre-fetch I/O... the I/O latency will overlap
						 * with the copying of the data
						 */
						if (size_of_prefetch > max_rd_size) {
							size_of_prefetch = max_rd_size;
						}

						size_of_prefetch = cluster_read_prefetch(vp, last_ioread_offset, size_of_prefetch, filesize, callback, callback_arg, bflag);

						last_ioread_offset += (off_t)(size_of_prefetch * PAGE_SIZE);

						if (last_ioread_offset > last_request_offset) {
							last_ioread_offset = last_request_offset;
						}
					}
				}
				/*
				 * limit the size of the copy we're about to do so that
				 * we can notice that our I/O pipe is running dry and
				 * get the next I/O issued before it does go dry
				 */
				if (last_ioread_offset && io_size > (max_io_size / 4)) {
					io_resid = (max_io_size / 4);
				} else {
					io_resid = io_size;
				}

				io_requested = io_resid;

				retval = cluster_copy_ubc_data_internal(vp, uio, (int *)&io_resid, 0, take_reference);

				xsize = io_requested - io_resid;

				io_size -= xsize;
				io_req_size -= xsize;

				if (retval || io_resid) {
					/*
					 * if we run into a real error or
					 * a page that is not in the cache
					 * we need to leave streaming mode
					 */
					break;
				}

				if (rd_ahead_enabled && (io_size == 0 || last_ioread_offset == last_request_offset)) {
					/*
					 * we're already finished the I/O for this read request
					 * let's see if we should do a read-ahead
					 */
					cluster_read_ahead(vp, &extent, filesize, rap, callback, callback_arg, bflag);
				}
			}
			if (retval) {
				break;
			}
			if (io_size == 0) {
				if (rap != NULL) {
					if (extent.e_addr < rap->cl_lastr) {
						rap->cl_maxra = 0;
					}
					rap->cl_lastr = extent.e_addr;
				}
				break;
			}
			/*
			 * recompute max_size since cluster_copy_ubc_data_internal
			 * may have advanced uio->uio_offset
			 */
			max_size = filesize - uio->uio_offset;
		}

		iostate.io_completed = 0;
		iostate.io_issued = 0;
		iostate.io_error = 0;
		iostate.io_wanted = 0;

		if ((flags & IO_RETURN_ON_THROTTLE)) {
			if (cluster_is_throttled(vp) == THROTTLE_NOW) {
				if (!cluster_io_present_in_BC(vp, uio->uio_offset)) {
					/*
					 * we're in the throttle window and at least 1 I/O
					 * has already been issued by a throttleable thread
					 * in this window, so return with EAGAIN to indicate
					 * to the FS issuing the cluster_read call that it
					 * should now throttle after dropping any locks
					 */
					throttle_info_update_by_mount(vp->v_mount);

					retval = EAGAIN;
					break;
				}
			}
		}

		/*
		 * compute the size of the upl needed to encompass
		 * the requested read... limit each call to cluster_io
		 * to the maximum UPL size... cluster_io will clip if
		 * this exceeds the maximum io_size for the device,
		 * make sure to account for
		 * a starting offset that's not page aligned
		 */
		start_offset = (int)(uio->uio_offset & PAGE_MASK_64);
		upl_f_offset = uio->uio_offset - (off_t)start_offset;

		if (io_size > max_rd_size) {
			io_size = max_rd_size;
		}

		upl_size = (start_offset + io_size + (PAGE_SIZE - 1)) & ~PAGE_MASK;

		if (flags & IO_NOCACHE) {
			if (upl_size > max_io_size) {
				upl_size = max_io_size;
			}
		} else {
			if (upl_size > max_io_size / 4) {
				upl_size = max_io_size / 4;
				upl_size &= ~PAGE_MASK;

				if (upl_size == 0) {
					upl_size = PAGE_SIZE;
				}
			}
		}
		pages_in_upl = upl_size / PAGE_SIZE;

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 33)) | DBG_FUNC_START,
		    upl, (int)upl_f_offset, upl_size, start_offset, 0);

		kret = ubc_create_upl_kernel(vp,
		    upl_f_offset,
		    upl_size,
		    &upl,
		    &pl,
		    UPL_FILE_IO | UPL_SET_LITE,
		    VM_KERN_MEMORY_FILE);
		if (kret != KERN_SUCCESS) {
			panic("cluster_read_copy: failed to get pagelist");
		}

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 33)) | DBG_FUNC_END,
		    upl, (int)upl_f_offset, upl_size, start_offset, 0);

		/*
		 * scan from the beginning of the upl looking for the first
		 * non-valid page.... this will become the first page in
		 * the request we're going to make to 'cluster_io'... if all
		 * of the pages are valid, we won't call through to 'cluster_io'
		 */
		for (start_pg = 0; start_pg < pages_in_upl; start_pg++) {
			if (!upl_valid_page(pl, start_pg)) {
				break;
			}
		}

		/*
		 * scan from the starting invalid page looking for a valid
		 * page before the end of the upl is reached, if we
		 * find one, then it will be the last page of the request to
		 * 'cluster_io'
		 */
		for (last_pg = start_pg; last_pg < pages_in_upl; last_pg++) {
			if (upl_valid_page(pl, last_pg)) {
				break;
			}
		}

		if (start_pg < last_pg) {
			/*
			 * we found a range of 'invalid' pages that must be filled
			 * if the last page in this range is the last page of the file
			 * we may have to clip the size of it to keep from reading past
			 * the end of the last physical block associated with the file
			 */
			if (iolock_inited == FALSE) {
				lck_mtx_init(&iostate.io_mtxp, &cl_mtx_grp, LCK_ATTR_NULL);

				iolock_inited = TRUE;
			}
			upl_offset = start_pg * PAGE_SIZE;
			io_size    = (last_pg - start_pg) * PAGE_SIZE;

			if ((off_t)(upl_f_offset + upl_offset + io_size) > filesize) {
				io_size = (u_int32_t)(filesize - (upl_f_offset + upl_offset));
			}

			/*
			 * Find out if this needs verification, we'll have to manage the UPL
			 * diffrently if so. Note that this call only lets us know if
			 * verification is enabled on this mount point, the actual verification
			 * is performed in the File system.
			 */
			size_t verify_block_size = 0;
			if ((VNOP_VERIFY(vp, start_offset, NULL, 0, &verify_block_size, NULL, VNODE_VERIFY_DEFAULT, NULL, NULL) == 0) /* && verify_block_size */) {
				for (uio_last = last_pg; uio_last < pages_in_upl; uio_last++) {
					if (!upl_valid_page(pl, uio_last)) {
						break;
					}
				}
				if (uio_last < pages_in_upl) {
					/*
					 * there were some invalid pages beyond the valid pages
					 * that we didn't issue an I/O for, just release them
					 * unchanged now, so that any prefetch/readahed can
					 * include them
					 */
					ubc_upl_abort_range(upl, uio_last * PAGE_SIZE,
					    (pages_in_upl - uio_last) * PAGE_SIZE, UPL_ABORT_FREE_ON_EMPTY);
					leftover_upl_aborted = true;
				}
			}

			/*
			 * issue an asynchronous read to cluster_io
			 */

			error = cluster_io(vp, upl, upl_offset, upl_f_offset + upl_offset,
			    io_size, CL_READ | CL_ASYNC | bflag, (buf_t)NULL, &iostate, callback, callback_arg);

			if (rap) {
				if (extent.e_addr < rap->cl_maxra) {
					/*
					 * we've just issued a read for a block that should have been
					 * in the cache courtesy of the read-ahead engine... something
					 * has gone wrong with the pipeline, so reset the read-ahead
					 * logic which will cause us to restart from scratch
					 */
					rap->cl_maxra = 0;
				}
			}
		}
		if (error == 0) {
			/*
			 * if the read completed successfully, or there was no I/O request
			 * issued, than copy the data into user land via 'cluster_upl_copy_data'
			 * we'll first add on any 'valid'
			 * pages that were present in the upl when we acquired it.
			 */
			u_int  val_size;

			if (!leftover_upl_aborted) {
				for (uio_last = last_pg; uio_last < pages_in_upl; uio_last++) {
					if (!upl_valid_page(pl, uio_last)) {
						break;
					}
				}
				if (uio_last < pages_in_upl) {
					/*
					 * there were some invalid pages beyond the valid pages
					 * that we didn't issue an I/O for, just release them
					 * unchanged now, so that any prefetch/readahed can
					 * include them
					 */
					ubc_upl_abort_range(upl, uio_last * PAGE_SIZE,
					    (pages_in_upl - uio_last) * PAGE_SIZE, UPL_ABORT_FREE_ON_EMPTY);
				}
			}

			/*
			 * compute size to transfer this round,  if io_req_size is
			 * still non-zero after this attempt, we'll loop around and
			 * set up for another I/O.
			 */
			val_size = (uio_last * PAGE_SIZE) - start_offset;

			if (val_size > max_size) {
				val_size = (u_int)max_size;
			}

			if (val_size > io_req_size) {
				val_size = io_req_size;
			}

			if ((uio->uio_offset + val_size) > last_ioread_offset) {
				last_ioread_offset = uio->uio_offset + val_size;
			}

			if ((size_of_prefetch = (u_int32_t)(last_request_offset - last_ioread_offset)) && prefetch_enabled) {
				if ((last_ioread_offset - (uio->uio_offset + val_size)) <= upl_size) {
					/*
					 * if there's still I/O left to do for this request, and...
					 * we're not in hard throttle mode, and...
					 * we're close to using up the previous prefetch, then issue a
					 * new pre-fetch I/O... the I/O latency will overlap
					 * with the copying of the data
					 */
					if (size_of_prefetch > max_rd_size) {
						size_of_prefetch = max_rd_size;
					}

					size_of_prefetch = cluster_read_prefetch(vp, last_ioread_offset, size_of_prefetch, filesize, callback, callback_arg, bflag);

					last_ioread_offset += (off_t)(size_of_prefetch * PAGE_SIZE);

					if (last_ioread_offset > last_request_offset) {
						last_ioread_offset = last_request_offset;
					}
				}
			} else if ((uio->uio_offset + val_size) == last_request_offset) {
				/*
				 * this transfer will finish this request, so...
				 * let's try to read ahead if we're in
				 * a sequential access pattern and we haven't
				 * explicitly disabled it
				 */
				if (rd_ahead_enabled) {
					cluster_read_ahead(vp, &extent, filesize, rap, callback, callback_arg, bflag);
				}

				if (rap != NULL) {
					if (extent.e_addr < rap->cl_lastr) {
						rap->cl_maxra = 0;
					}
					rap->cl_lastr = extent.e_addr;
				}
			}
			if (iolock_inited == TRUE) {
				cluster_iostate_wait(&iostate, 0, "cluster_read_copy");
			}

			if (iostate.io_error) {
				error = iostate.io_error;
			} else {
				u_int32_t io_requested;

				io_requested = val_size;

				retval = cluster_copy_upl_data(uio, upl, start_offset, (int *)&io_requested);

				io_req_size -= (val_size - io_requested);
			}
		} else {
			if (iolock_inited == TRUE) {
				cluster_iostate_wait(&iostate, 0, "cluster_read_copy");
			}
		}
		if (start_pg < last_pg) {
			/*
			 * compute the range of pages that we actually issued an I/O for
			 * and either commit them as valid if the I/O succeeded
			 * or abort them if the I/O failed or we're not supposed to
			 * keep them in the cache
			 */
			io_size = (last_pg - start_pg) * PAGE_SIZE;

			KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 35)) | DBG_FUNC_START, upl, start_pg * PAGE_SIZE, io_size, error, 0);

			if (error || (flags & IO_NOCACHE)) {
				ubc_upl_abort_range(upl, start_pg * PAGE_SIZE, io_size,
				    UPL_ABORT_DUMP_PAGES | UPL_ABORT_FREE_ON_EMPTY);
			} else {
				int     commit_flags = UPL_COMMIT_CLEAR_DIRTY | UPL_COMMIT_FREE_ON_EMPTY;

				if (take_reference) {
					commit_flags |= UPL_COMMIT_INACTIVATE;
				} else {
					commit_flags |= UPL_COMMIT_SPECULATE;
				}

				ubc_upl_commit_range(upl, start_pg * PAGE_SIZE, io_size, commit_flags);
			}
			KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 35)) | DBG_FUNC_END, upl, start_pg * PAGE_SIZE, io_size, error, 0);
		}
		if ((last_pg - start_pg) < pages_in_upl) {
			/*
			 * the set of pages that we issued an I/O for did not encompass
			 * the entire upl... so just release these without modifying
			 * their state
			 */
			if (error) {
				if (leftover_upl_aborted) {
					ubc_upl_abort_range(upl, start_pg * PAGE_SIZE, (uio_last - start_pg) * PAGE_SIZE,
					    UPL_ABORT_FREE_ON_EMPTY);
				} else {
					ubc_upl_abort_range(upl, 0, upl_size, UPL_ABORT_FREE_ON_EMPTY);
				}
			} else {
				KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 35)) | DBG_FUNC_START,
				    upl, -1, pages_in_upl - (last_pg - start_pg), 0, 0);

				/*
				 * handle any valid pages at the beginning of
				 * the upl... release these appropriately
				 */
				cluster_read_upl_release(upl, 0, start_pg, take_reference);

				/*
				 * handle any valid pages immediately after the
				 * pages we issued I/O for... ... release these appropriately
				 */
				cluster_read_upl_release(upl, last_pg, uio_last, take_reference);

				KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 35)) | DBG_FUNC_END, upl, -1, -1, 0, 0);
			}
		}
		if (retval == 0) {
			retval = error;
		}

		if (io_req_size) {
			uint32_t max_throttle_size = calculate_max_throttle_size(vp);

			if (cluster_is_throttled(vp)) {
				/*
				 * we're in the throttle window, at the very least
				 * we want to limit the size of the I/O we're about
				 * to issue
				 */
				rd_ahead_enabled = 0;
				prefetch_enabled = 0;
				max_rd_size = max_throttle_size;
			} else {
				if (max_rd_size == max_throttle_size) {
					/*
					 * coming out of throttled state
					 */
					if (policy != THROTTLE_LEVEL_TIER3 && policy != THROTTLE_LEVEL_TIER2) {
						if (rap != NULL) {
							rd_ahead_enabled = 1;
						}
						prefetch_enabled = 1;
					}
					max_rd_size = max_prefetch;
					last_ioread_offset = 0;
				}
			}
		}
	}
	if (iolock_inited == TRUE) {
		/*
		 * cluster_io returned an error after it
		 * had already issued some I/O.  we need
		 * to wait for that I/O to complete before
		 * we can destroy the iostate mutex...
		 * 'retval' already contains the early error
		 * so no need to pick it up from iostate.io_error
		 */
		cluster_iostate_wait(&iostate, 0, "cluster_read_copy");

		lck_mtx_destroy(&iostate.io_mtxp, &cl_mtx_grp);
	}
	if (rap != NULL) {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 32)) | DBG_FUNC_END,
		    (int)uio->uio_offset, io_req_size, rap->cl_lastr, retval, 0);

		lck_mtx_unlock(&rap->cl_lockr);
	} else {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 32)) | DBG_FUNC_END,
		    (int)uio->uio_offset, io_req_size, 0, retval, 0);
	}

	return retval;
}

/*
 * We don't want another read/write lock for every vnode in the system
 * so we keep a hash of them here.  There should never be very many of
 * these around at any point in time.
 */
cl_direct_read_lock_t *
cluster_lock_direct_read(vnode_t vp, lck_rw_type_t type)
{
	struct cl_direct_read_locks *head
	        = &cl_direct_read_locks[(uintptr_t)vp / sizeof(*vp)
	    % CL_DIRECT_READ_LOCK_BUCKETS];

	struct cl_direct_read_lock *lck, *new_lck = NULL;

	for (;;) {
		lck_spin_lock(&cl_direct_read_spin_lock);

		LIST_FOREACH(lck, head, chain) {
			if (lck->vp == vp) {
				++lck->ref_count;
				lck_spin_unlock(&cl_direct_read_spin_lock);
				if (new_lck) {
					// Someone beat us to it, ditch the allocation
					lck_rw_destroy(&new_lck->rw_lock, &cl_mtx_grp);
					kfree_type(cl_direct_read_lock_t, new_lck);
				}
				lck_rw_lock(&lck->rw_lock, type);
				return lck;
			}
		}

		if (new_lck) {
			// Use the lock we allocated
			LIST_INSERT_HEAD(head, new_lck, chain);
			lck_spin_unlock(&cl_direct_read_spin_lock);
			lck_rw_lock(&new_lck->rw_lock, type);
			return new_lck;
		}

		lck_spin_unlock(&cl_direct_read_spin_lock);

		// Allocate a new lock
		new_lck = kalloc_type(cl_direct_read_lock_t, Z_WAITOK);
		lck_rw_init(&new_lck->rw_lock, &cl_mtx_grp, LCK_ATTR_NULL);
		new_lck->vp = vp;
		new_lck->ref_count = 1;

		// Got to go round again
	}
}

void
cluster_unlock_direct_read(cl_direct_read_lock_t *lck)
{
	lck_rw_done(&lck->rw_lock);

	lck_spin_lock(&cl_direct_read_spin_lock);
	if (lck->ref_count == 1) {
		LIST_REMOVE(lck, chain);
		lck_spin_unlock(&cl_direct_read_spin_lock);
		lck_rw_destroy(&lck->rw_lock, &cl_mtx_grp);
		kfree_type(cl_direct_read_lock_t, lck);
	} else {
		--lck->ref_count;
		lck_spin_unlock(&cl_direct_read_spin_lock);
	}
}

static int
cluster_read_direct(vnode_t vp, struct uio *uio, off_t filesize, int *read_type, u_int32_t *read_length,
    int flags, int (*callback)(buf_t, void *), void *callback_arg)
{
	upl_t            upl = NULL;
	upl_page_info_t  *pl;
	off_t            max_io_size;
	size_t           verify_block_size = 0;
	vm_offset_t      upl_offset, vector_upl_offset = 0;
	upl_size_t       upl_size = 0, vector_upl_size = 0;
	vm_size_t        upl_needed_size;
	unsigned int     pages_in_pl;
	upl_control_flags_t upl_flags;
	kern_return_t    kret = KERN_SUCCESS;
	unsigned int     i;
	int              force_data_sync;
	int              retval = 0;
	int              no_zero_fill = 0;
	int              io_flag = 0;
	int              misaligned = 0;
	struct clios     iostate;
	user_addr_t      iov_base;
	u_int32_t        io_req_size;
	u_int32_t        offset_in_file;
	u_int32_t        offset_in_iovbase;
	u_int32_t        io_size;
	u_int32_t        io_min;
	u_int32_t        xsize;
	u_int32_t        devblocksize;
	u_int32_t        mem_alignment_mask;
	u_int32_t        max_upl_size;
	u_int32_t        max_rd_size;
	u_int32_t        max_rd_ahead;
	u_int32_t        max_vector_size;
	boolean_t        io_throttled = FALSE;

	u_int32_t        vector_upl_iosize = 0;
	int              issueVectorUPL = 0, useVectorUPL = (uio->uio_iovcnt > 1);
	off_t            v_upl_uio_offset = 0;
	int              vector_upl_index = 0;
	upl_t            vector_upl = NULL;
	cl_direct_read_lock_t *lock = NULL;
	uint32_t verify_mask = 0;

	assert(vm_map_page_shift(current_map()) >= PAGE_SHIFT);

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 70)) | DBG_FUNC_START,
	    (int)uio->uio_offset, (int)filesize, *read_type, *read_length, 0);

	max_upl_size = cluster_max_io_size(vp->v_mount, CL_READ);

	max_rd_size = max_upl_size;

	if (__improbable(os_mul_overflow(max_rd_size, IO_SCALE(vp, 2),
	    &max_rd_ahead) || (max_rd_ahead > overlapping_read_max))) {
		max_rd_ahead = overlapping_read_max;
	}

	io_flag = CL_COMMIT | CL_READ | CL_ASYNC | CL_NOZERO | CL_DIRECT_IO;

	if (flags & IO_PASSIVE) {
		io_flag |= CL_PASSIVE;
	}

	if (flags & IO_ENCRYPTED) {
		io_flag |= CL_RAW_ENCRYPTED;
	}

	if (flags & IO_NOCACHE) {
		io_flag |= CL_NOCACHE;
	}

	if (flags & IO_SKIP_ENCRYPTION) {
		io_flag |= CL_ENCRYPTED;
	}

	iostate.io_completed = 0;
	iostate.io_issued = 0;
	iostate.io_error = 0;
	iostate.io_wanted = 0;

	lck_mtx_init(&iostate.io_mtxp, &cl_mtx_grp, LCK_ATTR_NULL);

	devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize;
	mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask;

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 70)) | DBG_FUNC_NONE,
	    (int)devblocksize, (int)mem_alignment_mask, 0, 0, 0);

	if (devblocksize == 1) {
		/*
		 * the AFP client advertises a devblocksize of 1
		 * however, its BLOCKMAP routine maps to physical
		 * blocks that are PAGE_SIZE in size...
		 * therefore we can't ask for I/Os that aren't page aligned
		 * or aren't multiples of PAGE_SIZE in size
		 * by setting devblocksize to PAGE_SIZE, we re-instate
		 * the old behavior we had before the mem_alignment_mask
		 * changes went in...
		 */
		devblocksize = PAGE_SIZE;
	}

	/*
	 * We are going to need this uio for the prefaulting later
	 * especially for the cases where multiple non-contiguous
	 * iovs are passed into this routine.
	 *
	 * Note that we only want to prefault for direct IOs to userspace buffers,
	 * not kernel buffers.
	 */
	uio_t uio_acct = NULL;
	if (uio->uio_segflg != UIO_SYSSPACE) {
		uio_acct = uio_duplicate(uio);
	}

	retval = VNOP_VERIFY(vp, 0, NULL, 0, &verify_block_size, NULL, VNODE_VERIFY_DEFAULT, NULL, NULL);
	if (retval) {
		verify_block_size = 0;
	} else if (verify_block_size) {
		assert((verify_block_size & (verify_block_size - 1)) == 0);
		verify_mask = verify_block_size - 1;
	}

next_dread:
	io_req_size = *read_length;
	iov_base = uio_curriovbase(uio);

	offset_in_file = (u_int32_t)uio->uio_offset & (devblocksize - 1);
	offset_in_iovbase = (u_int32_t)iov_base & mem_alignment_mask;

	if (vm_map_page_mask(current_map()) < PAGE_MASK) {
		/*
		 * XXX TODO4K
		 * Direct I/O might not work as expected from a 16k kernel space
		 * to a 4k user space because each 4k chunk might point to
		 * a different 16k physical page...
		 * Let's go the "misaligned" way.
		 */
		if (!misaligned) {
			DEBUG4K_VFS("forcing misaligned\n");
		}
		misaligned = 1;
	}

	if (offset_in_file || offset_in_iovbase) {
		/*
		 * one of the 2 important offsets is misaligned
		 * so fire an I/O through the cache for this entire vector
		 */
		misaligned = 1;
	}
	if (iov_base & (devblocksize - 1)) {
		/*
		 * the offset in memory must be on a device block boundary
		 * so that we can guarantee that we can generate an
		 * I/O that ends on a page boundary in cluster_io
		 */
		misaligned = 1;
	}

	if (verify_block_size && !misaligned && ((uio_offset(uio) & verify_mask) || (uio_resid(uio) & verify_mask))) {
		/*
		 * If the offset is not aligned to the verification block size
		 * or the size is not aligned to the verification block size,
		 * we simply send this through the cached i/o path as that is
		 * what the Filesystem will end up doing anyway i.e. it will
		 * read all the remaining data in order to verify it and then
		 * discard the data it has read.
		 */
		misaligned = 1;
	}

	max_io_size = filesize - uio->uio_offset;

	/*
	 * The user must request IO in aligned chunks.  If the
	 * offset into the file is bad, or the userland pointer
	 * is non-aligned, then we cannot service the encrypted IO request.
	 */
	if (flags & IO_ENCRYPTED) {
		if (misaligned || (io_req_size & (devblocksize - 1))) {
			retval = EINVAL;
		}

		max_io_size = roundup(max_io_size, devblocksize);
	}

	if ((off_t)io_req_size > max_io_size) {
		io_req_size = (u_int32_t)max_io_size;
	}

	/*
	 * When we get to this point, we know...
	 *  -- the offset into the file is on a devblocksize boundary
	 */

	while (io_req_size && retval == 0) {
		u_int32_t io_start;

		if (cluster_is_throttled(vp)) {
			uint32_t max_throttle_size = calculate_max_throttle_size(vp);

			/*
			 * we're in the throttle window, at the very least
			 * we want to limit the size of the I/O we're about
			 * to issue
			 */
			max_rd_size  = max_throttle_size;
			max_rd_ahead = max_throttle_size - 1;
			max_vector_size = max_throttle_size;
		} else {
			max_rd_size  = max_upl_size;
			max_rd_ahead = max_rd_size * IO_SCALE(vp, 2);
			max_vector_size = MAX_VECTOR_UPL_SIZE;
		}
		io_start = io_size = io_req_size;

		/*
		 * First look for pages already in the cache
		 * and move them to user space.  But only do this
		 * check if we are not retrieving encrypted data directly
		 * from the filesystem;  those blocks should never
		 * be in the UBC.
		 *
		 * cluster_copy_ubc_data returns the resid
		 * in io_size
		 */
		if ((flags & IO_ENCRYPTED) == 0) {
			retval = cluster_copy_ubc_data_internal(vp, uio, (int *)&io_size, 0, 0);
		}
		/*
		 * calculate the number of bytes actually copied
		 * starting size - residual
		 */
		xsize = io_start - io_size;

		io_req_size -= xsize;

		if (useVectorUPL && (xsize || (iov_base & PAGE_MASK))) {
			/*
			 * We found something in the cache or we have an iov_base that's not
			 * page-aligned.
			 *
			 * Issue all I/O's that have been collected within this Vectored UPL.
			 */
			if (vector_upl_index) {
				retval = vector_cluster_io(vp, vector_upl, vector_upl_offset, v_upl_uio_offset, vector_upl_iosize, io_flag, (buf_t)NULL, &iostate, callback, callback_arg);
				reset_vector_run_state();
			}

			if (xsize) {
				useVectorUPL = 0;
			}

			/*
			 * After this point, if we are using the Vector UPL path and the base is
			 * not page-aligned then the UPL with that base will be the first in the vector UPL.
			 */
		}

		/*
		 * check to see if we are finished with this request.
		 *
		 * If we satisfied this IO already, then io_req_size will be 0.
		 * Otherwise, see if the IO was mis-aligned and needs to go through
		 * the UBC to deal with the 'tail'.
		 *
		 */
		if (io_req_size == 0 || (misaligned)) {
			/*
			 * see if there's another uio vector to
			 * process that's of type IO_DIRECT
			 *
			 * break out of while loop to get there
			 */
			break;
		}
		/*
		 * assume the request ends on a device block boundary
		 */
		io_min = devblocksize;

		/*
		 * we can handle I/O's in multiples of the device block size
		 * however, if io_size isn't a multiple of devblocksize we
		 * want to clip it back to the nearest page boundary since
		 * we are going to have to go through cluster_read_copy to
		 * deal with the 'overhang'... by clipping it to a PAGE_SIZE
		 * multiple, we avoid asking the drive for the same physical
		 * blocks twice.. once for the partial page at the end of the
		 * request and a 2nd time for the page we read into the cache
		 * (which overlaps the end of the direct read) in order to
		 * get at the overhang bytes
		 */
		if (io_size & (devblocksize - 1)) {
			assert(!(flags & IO_ENCRYPTED));
			/*
			 * Clip the request to the previous page size boundary
			 * since request does NOT end on a device block boundary
			 */
			io_size &= ~PAGE_MASK;
			io_min = PAGE_SIZE;
		}
		if (retval || io_size < io_min) {
			/*
			 * either an error or we only have the tail left to
			 * complete via the copy path...
			 * we may have already spun some portion of this request
			 * off as async requests... we need to wait for the I/O
			 * to complete before returning
			 */
			goto wait_for_dreads;
		}

		/*
		 * Don't re-check the UBC data if we are looking for uncached IO
		 * or asking for encrypted blocks.
		 */
		if ((flags & IO_ENCRYPTED) == 0) {
			if ((xsize = io_size) > max_rd_size) {
				xsize = max_rd_size;
			}

			io_size = 0;

			if (!lock) {
				/*
				 * We hold a lock here between the time we check the
				 * cache and the time we issue I/O.  This saves us
				 * from having to lock the pages in the cache.  Not
				 * all clients will care about this lock but some
				 * clients may want to guarantee stability between
				 * here and when the I/O is issued in which case they
				 * will take the lock exclusively.
				 */
				lock = cluster_lock_direct_read(vp, LCK_RW_TYPE_SHARED);
			}

			ubc_range_op(vp, uio->uio_offset, uio->uio_offset + xsize, UPL_ROP_ABSENT, (int *)&io_size);

			if (io_size == 0) {
				/*
				 * a page must have just come into the cache
				 * since the first page in this range is no
				 * longer absent, go back and re-evaluate
				 */
				continue;
			}
		}
		if ((flags & IO_RETURN_ON_THROTTLE)) {
			if (cluster_is_throttled(vp) == THROTTLE_NOW) {
				if (!cluster_io_present_in_BC(vp, uio->uio_offset)) {
					/*
					 * we're in the throttle window and at least 1 I/O
					 * has already been issued by a throttleable thread
					 * in this window, so return with EAGAIN to indicate
					 * to the FS issuing the cluster_read call that it
					 * should now throttle after dropping any locks
					 */
					throttle_info_update_by_mount(vp->v_mount);

					io_throttled = TRUE;
					goto wait_for_dreads;
				}
			}
		}
		if (io_size > max_rd_size) {
			io_size = max_rd_size;
		}

		iov_base = uio_curriovbase(uio);

		upl_offset = (vm_offset_t)((u_int32_t)iov_base & PAGE_MASK);
		upl_needed_size = (upl_offset + io_size + (PAGE_SIZE - 1)) & ~PAGE_MASK;

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 72)) | DBG_FUNC_START,
		    (int)upl_offset, upl_needed_size, (int)iov_base, io_size, 0);

		if (upl_offset == 0 && ((io_size & PAGE_MASK) == 0)) {
			no_zero_fill = 1;
		} else {
			no_zero_fill = 0;
		}

		vm_map_t map = UIO_SEG_IS_USER_SPACE(uio->uio_segflg) ? current_map() : kernel_map;
		for (force_data_sync = 0; force_data_sync < 3; force_data_sync++) {
			pages_in_pl = 0;
			upl_size = (upl_size_t)upl_needed_size;
			upl_flags = UPL_FILE_IO | UPL_NO_SYNC | UPL_SET_INTERNAL | UPL_SET_LITE | UPL_SET_IO_WIRE;
			if (no_zero_fill) {
				upl_flags |= UPL_NOZEROFILL;
			}
			if (force_data_sync) {
				upl_flags |= UPL_FORCE_DATA_SYNC;
			}

			kret = vm_map_create_upl(map,
			    (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK)),
			    &upl_size, &upl, NULL, &pages_in_pl, &upl_flags, VM_KERN_MEMORY_FILE);

			if (kret != KERN_SUCCESS) {
				KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 72)) | DBG_FUNC_END,
				    (int)upl_offset, upl_size, io_size, kret, 0);
				/*
				 * failed to get pagelist
				 *
				 * we may have already spun some portion of this request
				 * off as async requests... we need to wait for the I/O
				 * to complete before returning
				 */
				goto wait_for_dreads;
			}
			pages_in_pl = upl_size / PAGE_SIZE;
			pl = UPL_GET_INTERNAL_PAGE_LIST(upl);

			for (i = 0; i < pages_in_pl; i++) {
				if (!upl_page_present(pl, i)) {
					break;
				}
			}
			if (i == pages_in_pl) {
				break;
			}

			ubc_upl_abort(upl, 0);
		}
		if (force_data_sync >= 3) {
			KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 72)) | DBG_FUNC_END,
			    (int)upl_offset, upl_size, io_size, kret, 0);

			goto wait_for_dreads;
		}
		/*
		 * Consider the possibility that upl_size wasn't satisfied.
		 */
		if (upl_size < upl_needed_size) {
			if (upl_size && upl_offset == 0) {
				io_size = upl_size;
			} else {
				io_size = 0;
			}
		}
		if (io_size == 0) {
			ubc_upl_abort(upl, 0);
			goto wait_for_dreads;
		}
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 72)) | DBG_FUNC_END,
		    (int)upl_offset, upl_size, io_size, kret, 0);

		if (useVectorUPL) {
			vm_offset_t end_off = ((iov_base + io_size) & PAGE_MASK);
			if (end_off) {
				issueVectorUPL = 1;
			}
			/*
			 * After this point, if we are using a vector UPL, then
			 * either all the UPL elements end on a page boundary OR
			 * this UPL is the last element because it does not end
			 * on a page boundary.
			 */
		}

		/*
		 * request asynchronously so that we can overlap
		 * the preparation of the next I/O
		 * if there are already too many outstanding reads
		 * wait until some have completed before issuing the next read
		 */
		cluster_iostate_wait(&iostate, max_rd_ahead, "cluster_read_direct");

		if (iostate.io_error) {
			/*
			 * one of the earlier reads we issued ran into a hard error
			 * don't issue any more reads, cleanup the UPL
			 * that was just created but not used, then
			 * go wait for any other reads to complete before
			 * returning the error to the caller
			 */
			ubc_upl_abort(upl, 0);

			goto wait_for_dreads;
		}
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 73)) | DBG_FUNC_START,
		    upl, (int)upl_offset, (int)uio->uio_offset, io_size, 0);

		if (!useVectorUPL) {
			if (no_zero_fill) {
				io_flag &= ~CL_PRESERVE;
			} else {
				io_flag |= CL_PRESERVE;
			}

			retval = cluster_io(vp, upl, upl_offset, uio->uio_offset, io_size, io_flag, (buf_t)NULL, &iostate, callback, callback_arg);
		} else {
			if (!vector_upl_index) {
				vector_upl = vector_upl_create(upl_offset, uio->uio_iovcnt);
				v_upl_uio_offset = uio->uio_offset;
				vector_upl_offset = upl_offset;
			}

			vector_upl_set_subupl(vector_upl, upl, upl_size);
			vector_upl_set_iostate(vector_upl, upl, vector_upl_size, upl_size);
			vector_upl_index++;
			vector_upl_size += upl_size;
			vector_upl_iosize += io_size;

			if (issueVectorUPL || vector_upl_index == vector_upl_max_upls(vector_upl) || vector_upl_size >= max_vector_size) {
				retval = vector_cluster_io(vp, vector_upl, vector_upl_offset, v_upl_uio_offset, vector_upl_iosize, io_flag, (buf_t)NULL, &iostate, callback, callback_arg);
				reset_vector_run_state();
			}
		}

		if (lock) {
			// We don't need to wait for the I/O to complete
			cluster_unlock_direct_read(lock);
			lock = NULL;
		}

		/*
		 * update the uio structure
		 */
		if ((flags & IO_ENCRYPTED) && (max_io_size < io_size)) {
			uio_update(uio, (user_size_t)max_io_size);
		} else {
			uio_update(uio, (user_size_t)io_size);
		}

		io_req_size -= io_size;

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 73)) | DBG_FUNC_END,
		    upl, (int)uio->uio_offset, io_req_size, retval, 0);
	} /* end while */

	if (retval == 0 && iostate.io_error == 0 && io_req_size == 0 && uio->uio_offset < filesize) {
		retval = cluster_io_type(uio, read_type, read_length, 0);

		if (retval == 0 && *read_type == IO_DIRECT) {
			KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 70)) | DBG_FUNC_NONE,
			    (int)uio->uio_offset, (int)filesize, *read_type, *read_length, 0);

			goto next_dread;
		}
	}

wait_for_dreads:

	if (retval == 0 && iostate.io_error == 0 && useVectorUPL && vector_upl_index) {
		retval = vector_cluster_io(vp, vector_upl, vector_upl_offset, v_upl_uio_offset, vector_upl_iosize, io_flag, (buf_t)NULL, &iostate, callback, callback_arg);
		reset_vector_run_state();
	}

	// We don't need to wait for the I/O to complete
	if (lock) {
		cluster_unlock_direct_read(lock);
	}

	/*
	 * make sure all async reads that are part of this stream
	 * have completed before we return
	 */
	cluster_iostate_wait(&iostate, 0, "cluster_read_direct");

	if (iostate.io_error) {
		retval = iostate.io_error;
	}

	lck_mtx_destroy(&iostate.io_mtxp, &cl_mtx_grp);

	if (io_throttled == TRUE && retval == 0) {
		retval = EAGAIN;
	}

	vm_map_offset_t current_page_size, current_page_mask;
	current_page_size = vm_map_page_size(current_map());
	current_page_mask = vm_map_page_mask(current_map());
	if (uio_acct) {
		assert(uio_acct->uio_segflg != UIO_SYSSPACE);
		off_t bytes_to_prefault = 0, bytes_prefaulted = 0;
		user_addr_t curr_iov_base = 0;
		user_addr_t curr_iov_end = 0;
		user_size_t curr_iov_len = 0;

		bytes_to_prefault = uio_offset(uio) - uio_offset(uio_acct);

		for (; bytes_prefaulted < bytes_to_prefault;) {
			curr_iov_base = uio_curriovbase(uio_acct);
			curr_iov_len = MIN(uio_curriovlen(uio_acct), bytes_to_prefault - bytes_prefaulted);
			curr_iov_end = curr_iov_base + curr_iov_len;

			for (; curr_iov_base < curr_iov_end;) {
				/*
				 * This is specifically done for pmap accounting purposes.
				 * vm_pre_fault() will call vm_fault() to enter the page into
				 * the pmap if there isn't _a_ physical page for that VA already.
				 */
				vm_pre_fault(vm_map_trunc_page(curr_iov_base, current_page_mask), VM_PROT_READ);
				curr_iov_base += current_page_size;
				bytes_prefaulted += current_page_size;
			}
			/*
			 * Use update instead of advance so we can see how many iovs we processed.
			 */
			uio_update(uio_acct, curr_iov_len);
		}
		uio_free(uio_acct);
		uio_acct = NULL;
	}

	if (io_req_size && retval == 0) {
		/*
		 * we couldn't handle the tail of this request in DIRECT mode
		 * so fire it through the copy path
		 */
		if (flags & IO_ENCRYPTED) {
			/*
			 * We cannot fall back to the copy path for encrypted I/O. If this
			 * happens, there is something wrong with the user buffer passed
			 * down.
			 */
			retval = EFAULT;
		} else {
			retval = cluster_read_copy(vp, uio, io_req_size, filesize, flags, callback, callback_arg);
		}

		*read_type = IO_UNKNOWN;
	}
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 70)) | DBG_FUNC_END,
	    (int)uio->uio_offset, (int)uio_resid(uio), io_req_size, retval, 0);

	return retval;
}


static int
cluster_read_contig(vnode_t vp, struct uio *uio, off_t filesize, int *read_type, u_int32_t *read_length,
    int (*callback)(buf_t, void *), void *callback_arg, int flags)
{
	upl_page_info_t *pl;
	upl_t            upl[MAX_VECTS];
	vm_offset_t      upl_offset;
	addr64_t         dst_paddr = 0;
	user_addr_t      iov_base;
	off_t            max_size;
	upl_size_t       upl_size;
	vm_size_t        upl_needed_size;
	mach_msg_type_number_t  pages_in_pl;
	upl_control_flags_t upl_flags;
	kern_return_t    kret;
	struct clios     iostate;
	int              error = 0;
	int              cur_upl = 0;
	int              num_upl = 0;
	int              n;
	u_int32_t        xsize;
	u_int32_t        io_size;
	u_int32_t        devblocksize;
	u_int32_t        mem_alignment_mask;
	u_int32_t        tail_size = 0;
	int              bflag;

	if (flags & IO_PASSIVE) {
		bflag = CL_PASSIVE;
	} else {
		bflag = 0;
	}

	if (flags & IO_NOCACHE) {
		bflag |= CL_NOCACHE;
	}

	/*
	 * When we enter this routine, we know
	 *  -- the read_length will not exceed the current iov_len
	 *  -- the target address is physically contiguous for read_length
	 */
	cluster_syncup(vp, filesize, callback, callback_arg, PUSH_SYNC);

	devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize;
	mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask;

	iostate.io_completed = 0;
	iostate.io_issued = 0;
	iostate.io_error = 0;
	iostate.io_wanted = 0;

	lck_mtx_init(&iostate.io_mtxp, &cl_mtx_grp, LCK_ATTR_NULL);

next_cread:
	io_size = *read_length;

	max_size = filesize - uio->uio_offset;

	if (io_size > max_size) {
		io_size = (u_int32_t)max_size;
	}

	iov_base = uio_curriovbase(uio);

	upl_offset = (vm_offset_t)((u_int32_t)iov_base & PAGE_MASK);
	upl_needed_size = upl_offset + io_size;

	pages_in_pl = 0;
	upl_size = (upl_size_t)upl_needed_size;
	upl_flags = UPL_FILE_IO | UPL_NO_SYNC | UPL_CLEAN_IN_PLACE | UPL_SET_INTERNAL | UPL_SET_LITE | UPL_SET_IO_WIRE;


	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 92)) | DBG_FUNC_START,
	    (int)upl_offset, (int)upl_size, (int)iov_base, io_size, 0);

	vm_map_t map = UIO_SEG_IS_USER_SPACE(uio->uio_segflg) ? current_map() : kernel_map;
	kret = vm_map_get_upl(map,
#if HAS_MTE || HAS_MTE_EMULATION_SHIMS
	    vm_memtag_canonicalize(map, vm_map_trunc_page(iov_base, vm_map_page_mask(map))),
#else /* HAS_MTE || HAS_MTE_EMULATION_SHIMS */
	    vm_map_trunc_page(iov_base, vm_map_page_mask(map)),
#endif /* HAS_MTE || HAS_MTE_EMULATION_SHIMS */
	    &upl_size, &upl[cur_upl], NULL, &pages_in_pl, &upl_flags, VM_KERN_MEMORY_FILE, 0);

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 92)) | DBG_FUNC_END,
	    (int)upl_offset, upl_size, io_size, kret, 0);

	if (kret != KERN_SUCCESS) {
		/*
		 * failed to get pagelist
		 */
		error = EINVAL;
		goto wait_for_creads;
	}
	num_upl++;

	if (!(upl_flags & UPL_PHYS_CONTIG)) {
		/*
		 * The created UPL needs to have the UPL_PHYS_CONTIG flag.
		 */
		error = EINVAL;
		goto wait_for_creads;
	}

	if (upl_size < upl_needed_size) {
		/*
		 * The upl_size wasn't satisfied.
		 */
		error = EINVAL;
		goto wait_for_creads;
	}
	pl = ubc_upl_pageinfo(upl[cur_upl]);

	dst_paddr = ((addr64_t)upl_phys_page(pl, 0) << PAGE_SHIFT) + (addr64_t)upl_offset;

	while (((uio->uio_offset & (devblocksize - 1)) || io_size < devblocksize) && io_size) {
		u_int32_t   head_size;

		head_size = devblocksize - (u_int32_t)(uio->uio_offset & (devblocksize - 1));

		if (head_size > io_size) {
			head_size = io_size;
		}

		error = cluster_align_phys_io(vp, uio, dst_paddr, head_size, CL_READ, callback, callback_arg);

		if (error) {
			goto wait_for_creads;
		}

		upl_offset += head_size;
		dst_paddr  += head_size;
		io_size    -= head_size;

		iov_base   += head_size;
	}
	if ((u_int32_t)iov_base & mem_alignment_mask) {
		/*
		 * request doesn't set up on a memory boundary
		 * the underlying DMA engine can handle...
		 * return an error instead of going through
		 * the slow copy path since the intent of this
		 * path is direct I/O to device memory
		 */
		error = EINVAL;
		goto wait_for_creads;
	}

	tail_size = io_size & (devblocksize - 1);

	io_size  -= tail_size;

	while (io_size && error == 0) {
		if (io_size > MAX_IO_CONTIG_SIZE) {
			xsize = MAX_IO_CONTIG_SIZE;
		} else {
			xsize = io_size;
		}
		/*
		 * request asynchronously so that we can overlap
		 * the preparation of the next I/O... we'll do
		 * the commit after all the I/O has completed
		 * since its all issued against the same UPL
		 * if there are already too many outstanding reads
		 * wait until some have completed before issuing the next
		 */
		cluster_iostate_wait(&iostate, MAX_IO_CONTIG_SIZE * IO_SCALE(vp, 2), "cluster_read_contig");

		if (iostate.io_error) {
			/*
			 * one of the earlier reads we issued ran into a hard error
			 * don't issue any more reads...
			 * go wait for any other reads to complete before
			 * returning the error to the caller
			 */
			goto wait_for_creads;
		}
		error = cluster_io(vp, upl[cur_upl], upl_offset, uio->uio_offset, xsize,
		    CL_READ | CL_NOZERO | CL_DEV_MEMORY | CL_ASYNC | bflag,
		    (buf_t)NULL, &iostate, callback, callback_arg);
		/*
		 * The cluster_io read was issued successfully,
		 * update the uio structure
		 */
		if (error == 0) {
			uio_update(uio, (user_size_t)xsize);

			dst_paddr  += xsize;
			upl_offset += xsize;
			io_size    -= xsize;
		}
	}
	if (error == 0 && iostate.io_error == 0 && tail_size == 0 && num_upl < MAX_VECTS && uio->uio_offset < filesize) {
		error = cluster_io_type(uio, read_type, read_length, 0);

		if (error == 0 && *read_type == IO_CONTIG) {
			cur_upl++;
			goto next_cread;
		}
	} else {
		*read_type = IO_UNKNOWN;
	}

wait_for_creads:
	/*
	 * make sure all async reads that are part of this stream
	 * have completed before we proceed
	 */
	cluster_iostate_wait(&iostate, 0, "cluster_read_contig");

	if (iostate.io_error) {
		error = iostate.io_error;
	}

	lck_mtx_destroy(&iostate.io_mtxp, &cl_mtx_grp);

	if (error == 0 && tail_size) {
		error = cluster_align_phys_io(vp, uio, dst_paddr, tail_size, CL_READ, callback, callback_arg);
	}

	for (n = 0; n < num_upl; n++) {
		/*
		 * just release our hold on each physically contiguous
		 * region without changing any state
		 */
		ubc_upl_abort(upl[n], 0);
	}

	return error;
}


static int
cluster_io_type(struct uio *uio, int *io_type, u_int32_t *io_length, u_int32_t min_length)
{
	user_size_t      iov_len;
	user_addr_t      iov_base = 0;
	upl_t            upl;
	upl_size_t       upl_size;
	upl_control_flags_t upl_flags;
	int              retval = 0;

	/*
	 * skip over any emtpy vectors
	 */
	uio_update(uio, (user_size_t)0);

	iov_len = MIN(uio_curriovlen(uio), uio_resid(uio));

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 94)) | DBG_FUNC_START, uio, (int)iov_len, 0, 0, 0);

	if (iov_len) {
		iov_base = uio_curriovbase(uio);
		/*
		 * make sure the size of the vector isn't too big...
		 * internally, we want to handle all of the I/O in
		 * chunk sizes that fit in a 32 bit int
		 */
		if (iov_len > (user_size_t)MAX_IO_REQUEST_SIZE) {
			upl_size = MAX_IO_REQUEST_SIZE;
		} else {
			upl_size = (u_int32_t)iov_len;
		}

		upl_flags = UPL_QUERY_OBJECT_TYPE;

		vm_map_t map = UIO_SEG_IS_USER_SPACE(uio->uio_segflg) ? current_map() : kernel_map;
		if ((vm_map_get_upl(map,
#if HAS_MTE || HAS_MTE_EMULATION_SHIMS
		    vm_memtag_canonicalize(map, vm_map_trunc_page(iov_base, vm_map_page_mask(map))),
#else /* HAS_MTE || HAS_MTE_EMULATION_SHIMS */
		    vm_map_trunc_page(iov_base, vm_map_page_mask(map)),
#endif /* HAS_MTE || HAS_MTE_EMULATION_SHIMS */
		    &upl_size, &upl, NULL, NULL, &upl_flags, VM_KERN_MEMORY_FILE, 0)) != KERN_SUCCESS) {
			/*
			 * the user app must have passed in an invalid address
			 */
			retval = EFAULT;
		}
		if (upl_size == 0) {
			retval = EFAULT;
		}

		*io_length = upl_size;

		if (upl_flags & UPL_PHYS_CONTIG) {
			*io_type = IO_CONTIG;
		} else if (iov_len >= min_length) {
			*io_type = IO_DIRECT;
		} else {
			*io_type = IO_COPY;
		}
	} else {
		/*
		 * nothing left to do for this uio
		 */
		*io_length = 0;
		*io_type   = IO_UNKNOWN;
	}
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 94)) | DBG_FUNC_END, iov_base, *io_type, *io_length, retval, 0);

	if (*io_type == IO_DIRECT &&
	    vm_map_page_shift(current_map()) < PAGE_SHIFT) {
		/* no direct I/O for sub-page-size address spaces */
		DEBUG4K_VFS("io_type IO_DIRECT -> IO_COPY\n");
		*io_type = IO_COPY;
	}

	return retval;
}


/*
 * generate advisory I/O's in the largest chunks possible
 * the completed pages will be released into the VM cache
 */
int
advisory_read(vnode_t vp, off_t filesize, off_t f_offset, int resid)
{
	return advisory_read_ext(vp, filesize, f_offset, resid, NULL, NULL, CL_PASSIVE);
}

int
advisory_read_ext(vnode_t vp, off_t filesize, off_t f_offset, int resid, int (*callback)(buf_t, void *), void *callback_arg, int bflag)
{
	upl_page_info_t *pl;
	upl_t            upl = NULL;
	vm_offset_t      upl_offset;
	int              upl_size;
	off_t            upl_f_offset;
	int              start_offset;
	int              start_pg;
	int              last_pg;
	int              pages_in_upl;
	off_t            max_size;
	int              io_size;
	kern_return_t    kret;
	int              retval = 0;
	int              issued_io;
	int              skip_range;
	uint32_t         max_io_size;


	if (!UBCINFOEXISTS(vp)) {
		return EINVAL;
	}

	if (f_offset < 0 || resid < 0) {
		return EINVAL;
	}

	max_io_size = cluster_max_io_size(vp->v_mount, CL_READ);

	if (disk_conditioner_mount_is_ssd(vp->v_mount)) {
		if (max_io_size > speculative_prefetch_max_iosize) {
			max_io_size = speculative_prefetch_max_iosize;
		}
	}

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 60)) | DBG_FUNC_START,
	    (int)f_offset, resid, (int)filesize, 0, 0);

	while (resid && f_offset < filesize && retval == 0) {
		/*
		 * compute the size of the upl needed to encompass
		 * the requested read... limit each call to cluster_io
		 * to the maximum UPL size... cluster_io will clip if
		 * this exceeds the maximum io_size for the device,
		 * make sure to account for
		 * a starting offset that's not page aligned
		 */
		start_offset = (int)(f_offset & PAGE_MASK_64);
		upl_f_offset = f_offset - (off_t)start_offset;
		max_size     = filesize - f_offset;

		if (resid < max_size) {
			io_size = resid;
		} else {
			io_size = (int)max_size;
		}

		upl_size = (start_offset + io_size + (PAGE_SIZE - 1)) & ~PAGE_MASK;
		if ((uint32_t)upl_size > max_io_size) {
			upl_size = max_io_size;
		}

		skip_range = 0;
		/*
		 * return the number of contiguously present pages in the cache
		 * starting at upl_f_offset within the file
		 */
		ubc_range_op(vp, upl_f_offset, upl_f_offset + upl_size, UPL_ROP_PRESENT, &skip_range);

		if (skip_range) {
			/*
			 * skip over pages already present in the cache
			 */
			io_size = skip_range - start_offset;

			f_offset += io_size;
			resid    -= io_size;

			if (skip_range == upl_size) {
				continue;
			}
			/*
			 * have to issue some real I/O
			 * at this point, we know it's starting on a page boundary
			 * because we've skipped over at least the first page in the request
			 */
			start_offset = 0;
			upl_f_offset += skip_range;
			upl_size     -= skip_range;
		}
		pages_in_upl = upl_size / PAGE_SIZE;

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 61)) | DBG_FUNC_START,
		    upl, (int)upl_f_offset, upl_size, start_offset, 0);

		kret = ubc_create_upl_kernel(vp,
		    upl_f_offset,
		    upl_size,
		    &upl,
		    &pl,
		    UPL_RET_ONLY_ABSENT | UPL_SET_LITE,
		    VM_KERN_MEMORY_FILE);
		if (kret != KERN_SUCCESS) {
			return retval;
		}
		issued_io = 0;

		/*
		 * before we start marching forward, we must make sure we end on
		 * a present page, otherwise we will be working with a freed
		 * upl
		 */
		for (last_pg = pages_in_upl - 1; last_pg >= 0; last_pg--) {
			if (upl_page_present(pl, last_pg)) {
				break;
			}
		}
		pages_in_upl = last_pg + 1;


		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 61)) | DBG_FUNC_END,
		    upl, (int)upl_f_offset, upl_size, start_offset, 0);


		for (last_pg = 0; last_pg < pages_in_upl;) {
			/*
			 * scan from the beginning of the upl looking for the first
			 * page that is present.... this will become the first page in
			 * the request we're going to make to 'cluster_io'... if all
			 * of the pages are absent, we won't call through to 'cluster_io'
			 */
			for (start_pg = last_pg; start_pg < pages_in_upl; start_pg++) {
				if (upl_page_present(pl, start_pg)) {
					break;
				}
			}

			/*
			 * scan from the starting present page looking for an absent
			 * page before the end of the upl is reached, if we
			 * find one, then it will terminate the range of pages being
			 * presented to 'cluster_io'
			 */
			for (last_pg = start_pg; last_pg < pages_in_upl; last_pg++) {
				if (!upl_page_present(pl, last_pg)) {
					break;
				}
			}

			if (last_pg > start_pg) {
				/*
				 * we found a range of pages that must be filled
				 * if the last page in this range is the last page of the file
				 * we may have to clip the size of it to keep from reading past
				 * the end of the last physical block associated with the file
				 */
				upl_offset = start_pg * PAGE_SIZE;
				io_size    = (last_pg - start_pg) * PAGE_SIZE;

				if ((off_t)(upl_f_offset + upl_offset + io_size) > filesize) {
					io_size = (int)(filesize - (upl_f_offset + upl_offset));
				}

				/*
				 * issue an asynchronous read to cluster_io
				 */
				retval = cluster_io(vp, upl, upl_offset, upl_f_offset + upl_offset, io_size,
				    CL_ASYNC | CL_READ | CL_COMMIT | CL_AGE | bflag, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg);

				issued_io = 1;
			}
		}
		if (issued_io == 0) {
			ubc_upl_abort(upl, 0);
		}

		io_size = upl_size - start_offset;

		if (io_size > resid) {
			io_size = resid;
		}
		f_offset += io_size;
		resid    -= io_size;
	}

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 60)) | DBG_FUNC_END,
	    (int)f_offset, resid, retval, 0, 0);

	return retval;
}


int
cluster_push(vnode_t vp, int flags)
{
	return cluster_push_ext(vp, flags, NULL, NULL);
}


int
cluster_push_ext(vnode_t vp, int flags, int (*callback)(buf_t, void *), void *callback_arg)
{
	return cluster_push_err(vp, flags, callback, callback_arg, NULL);
}

/* write errors via err, but return the number of clusters written */
extern uint32_t system_inshutdown;
uint32_t cl_sparse_push_error = 0;
int
cluster_push_err(vnode_t vp, int flags, int (*callback)(buf_t, void *), void *callback_arg, int *err)
{
	int     retval;
	int     my_sparse_wait = 0;
	struct  cl_writebehind *wbp;
	int     local_err = 0;

	if (err) {
		*err = 0;
	}

	if (!UBCINFOEXISTS(vp)) {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_NONE, kdebug_vnode(vp), flags, 0, -1, 0);
		return 0;
	}
	/* return if deferred write is set */
	if (((unsigned int)vfs_flags(vp->v_mount) & MNT_DEFWRITE) && (flags & IO_DEFWRITE)) {
		return 0;
	}
	if ((wbp = cluster_get_wbp(vp, CLW_RETURNLOCKED)) == NULL) {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_NONE, kdebug_vnode(vp), flags, 0, -2, 0);
		return 0;
	}
	if (!ISSET(flags, IO_SYNC) && wbp->cl_number == 0 && wbp->cl_scmap == NULL) {
		lck_mtx_unlock(&wbp->cl_lockw);

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_NONE, kdebug_vnode(vp), flags, 0, -3, 0);
		return 0;
	}
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_START,
	    wbp->cl_scmap, wbp->cl_number, flags, 0, 0);

	/*
	 * if we have an fsync in progress, we don't want to allow any additional
	 * sync/fsync/close(s) to occur until it finishes.
	 * note that its possible for writes to continue to occur to this file
	 * while we're waiting and also once the fsync starts to clean if we're
	 * in the sparse map case
	 */
	while (wbp->cl_sparse_wait) {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 97)) | DBG_FUNC_START, kdebug_vnode(vp), 0, 0, 0, 0);

		msleep((caddr_t)&wbp->cl_sparse_wait, &wbp->cl_lockw, PRIBIO + 1, "cluster_push_ext", NULL);

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 97)) | DBG_FUNC_END, kdebug_vnode(vp), 0, 0, 0, 0);
	}
	if (flags & IO_SYNC) {
		my_sparse_wait = 1;
		wbp->cl_sparse_wait = 1;

		/*
		 * this is an fsync (or equivalent)... we must wait for any existing async
		 * cleaning operations to complete before we evaulate the current state
		 * and finish cleaning... this insures that all writes issued before this
		 * fsync actually get cleaned to the disk before this fsync returns
		 */
		while (wbp->cl_sparse_pushes) {
			KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 98)) | DBG_FUNC_START, kdebug_vnode(vp), 0, 0, 0, 0);

			msleep((caddr_t)&wbp->cl_sparse_pushes, &wbp->cl_lockw, PRIBIO + 1, "cluster_push_ext", NULL);

			KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 98)) | DBG_FUNC_END, kdebug_vnode(vp), 0, 0, 0, 0);
		}
	}
	if (wbp->cl_scmap) {
		void    *scmap;

		if (wbp->cl_sparse_pushes < SPARSE_PUSH_LIMIT) {
			scmap = wbp->cl_scmap;
			wbp->cl_scmap = NULL;

			wbp->cl_sparse_pushes++;

			lck_mtx_unlock(&wbp->cl_lockw);

			retval = sparse_cluster_push(wbp, &scmap, vp, ubc_getsize(vp), PUSH_ALL, flags, callback, callback_arg, FALSE);

			lck_mtx_lock(&wbp->cl_lockw);

			wbp->cl_sparse_pushes--;

			if (retval) {
				if (wbp->cl_scmap != NULL) {
					/*
					 * panic("cluster_push_err: Expected NULL cl_scmap\n");
					 *
					 * This can happen if we get an error from the underlying FS
					 * e.g. ENOSPC, EPERM or EIO etc. We hope that these errors
					 * are transient and the I/Os will succeed at a later point.
					 *
					 * The tricky part here is that a new sparse cluster has been
					 * allocated and tracking a different set of dirty pages. So these
					 * pages are not going to be pushed out with the next sparse_cluster_push.
					 * An explicit msync or file close will, however, push the pages out.
					 *
					 * What if those calls still don't work? And so, during shutdown we keep
					 * trying till we succeed...
					 */

					if (system_inshutdown) {
						if ((retval == ENOSPC) && (vp->v_mount->mnt_flag & (MNT_LOCAL | MNT_REMOVABLE)) == MNT_LOCAL) {
							os_atomic_inc(&cl_sparse_push_error, relaxed);
						}
					} else {
						vfs_drt_control(&scmap, 0); /* emit stats and free this memory. Dirty pages stay intact. */
						scmap = NULL;
					}
				} else {
					wbp->cl_scmap = scmap;
				}
			}

			if (wbp->cl_sparse_wait && wbp->cl_sparse_pushes == 0) {
				wakeup((caddr_t)&wbp->cl_sparse_pushes);
			}
		} else {
			retval = sparse_cluster_push(wbp, &(wbp->cl_scmap), vp, ubc_getsize(vp), PUSH_ALL, flags, callback, callback_arg, FALSE);
		}

		local_err = retval;

		if (err) {
			*err = retval;
		}
		retval = 1;
	} else {
		retval = cluster_try_push(wbp, vp, ubc_getsize(vp), PUSH_ALL, flags, callback, callback_arg, &local_err, FALSE);
		if (err) {
			*err = local_err;
		}
	}
	lck_mtx_unlock(&wbp->cl_lockw);

	if (flags & IO_SYNC) {
		(void)vnode_waitforwrites(vp, 0, 0, 0, "cluster_push");
	}

	if (my_sparse_wait) {
		/*
		 * I'm the owner of the serialization token
		 * clear it and wakeup anyone that is waiting
		 * for me to finish
		 */
		lck_mtx_lock(&wbp->cl_lockw);

		wbp->cl_sparse_wait = 0;
		wakeup((caddr_t)&wbp->cl_sparse_wait);

		lck_mtx_unlock(&wbp->cl_lockw);
	}
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_END,
	    wbp->cl_scmap, wbp->cl_number, retval, local_err, 0);

	return retval;
}


__private_extern__ void
cluster_release(struct ubc_info *ubc)
{
	struct cl_writebehind *wbp;
	struct cl_readahead   *rap;

	if ((wbp = ubc->cl_wbehind)) {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 81)) | DBG_FUNC_START, ubc, wbp->cl_scmap, 0, 0, 0);

		if (wbp->cl_scmap) {
			vfs_drt_control(&(wbp->cl_scmap), 0);
		}
		lck_mtx_destroy(&wbp->cl_lockw, &cl_mtx_grp);
		zfree(cl_wr_zone, wbp);
		ubc->cl_wbehind = NULL;
	} else {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 81)) | DBG_FUNC_START, ubc, 0, 0, 0, 0);
	}

	if ((rap = ubc->cl_rahead)) {
		lck_mtx_destroy(&rap->cl_lockr, &cl_mtx_grp);
		zfree(cl_rd_zone, rap);
		ubc->cl_rahead  = NULL;
	}

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 81)) | DBG_FUNC_END, ubc, rap, wbp, 0, 0);
}


static int
cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int push_flag, int io_flags, int (*callback)(buf_t, void *), void *callback_arg, int *err, boolean_t vm_initiated)
{
	int cl_index;
	int cl_index1;
	int min_index;
	int cl_len;
	int cl_pushed = 0;
	struct cl_wextent l_clusters[MAX_CLUSTERS];
	u_int  max_cluster_pgcount;
	int error = 0;

	max_cluster_pgcount = MAX_CLUSTER_SIZE(vp) / PAGE_SIZE;
	/*
	 * the write behind context exists and has
	 * already been locked...
	 */
	if (wbp->cl_number == 0) {
		/*
		 * no clusters to push
		 * return number of empty slots
		 */
		return MAX_CLUSTERS;
	}

	/*
	 * make a local 'sorted' copy of the clusters
	 * and clear wbp->cl_number so that new clusters can
	 * be developed
	 */
	for (cl_index = 0; cl_index < wbp->cl_number; cl_index++) {
		for (min_index = -1, cl_index1 = 0; cl_index1 < wbp->cl_number; cl_index1++) {
			if (wbp->cl_clusters[cl_index1].b_addr == wbp->cl_clusters[cl_index1].e_addr) {
				continue;
			}
			if (min_index == -1) {
				min_index = cl_index1;
			} else if (wbp->cl_clusters[cl_index1].b_addr < wbp->cl_clusters[min_index].b_addr) {
				min_index = cl_index1;
			}
		}
		if (min_index == -1) {
			break;
		}

		l_clusters[cl_index].b_addr = wbp->cl_clusters[min_index].b_addr;
		l_clusters[cl_index].e_addr = wbp->cl_clusters[min_index].e_addr;
		l_clusters[cl_index].io_flags = wbp->cl_clusters[min_index].io_flags;

		wbp->cl_clusters[min_index].b_addr = wbp->cl_clusters[min_index].e_addr;
	}
	wbp->cl_number = 0;

	cl_len = cl_index;

	/* skip switching to the sparse cluster mechanism if on diskimage */
	if (((push_flag & PUSH_DELAY) && cl_len == MAX_CLUSTERS) &&
	    !(vp->v_mount->mnt_kern_flag & MNTK_VIRTUALDEV)) {
		int   i;

		/*
		 * determine if we appear to be writing the file sequentially
		 * if not, by returning without having pushed any clusters
		 * we will cause this vnode to be pushed into the sparse cluster mechanism
		 * used for managing more random I/O patterns
		 *
		 * we know that we've got all clusters currently in use and the next write doesn't fit into one of them...
		 * that's why we're in try_push with PUSH_DELAY...
		 *
		 * check to make sure that all the clusters except the last one are 'full'... and that each cluster
		 * is adjacent to the next (i.e. we're looking for sequential writes) they were sorted above
		 * so we can just make a simple pass through, up to, but not including the last one...
		 * note that e_addr is not inclusive, so it will be equal to the b_addr of the next cluster if they
		 * are sequential
		 *
		 * we let the last one be partial as long as it was adjacent to the previous one...
		 * we need to do this to deal with multi-threaded servers that might write an I/O or 2 out
		 * of order... if this occurs at the tail of the last cluster, we don't want to fall into the sparse cluster world...
		 */
		for (i = 0; i < MAX_CLUSTERS - 1; i++) {
			if ((l_clusters[i].e_addr - l_clusters[i].b_addr) != max_cluster_pgcount) {
				goto dont_try;
			}
			if (l_clusters[i].e_addr != l_clusters[i + 1].b_addr) {
				goto dont_try;
			}
		}
	}
	if (vm_initiated == TRUE) {
		lck_mtx_unlock(&wbp->cl_lockw);
	}

	for (cl_index = 0; cl_index < cl_len; cl_index++) {
		int     flags;
		struct  cl_extent cl;
		int retval;

		flags = io_flags & (IO_PASSIVE | IO_CLOSE);

		/*
		 * try to push each cluster in turn...
		 */
		if (l_clusters[cl_index].io_flags & CLW_IONOCACHE) {
			flags |= IO_NOCACHE;
		}

		if (l_clusters[cl_index].io_flags & CLW_IOPASSIVE) {
			flags |= IO_PASSIVE;
		}

		if (push_flag & PUSH_SYNC) {
			flags |= IO_SYNC;
		}

		cl.b_addr = l_clusters[cl_index].b_addr;
		cl.e_addr = l_clusters[cl_index].e_addr;

		retval = cluster_push_now(vp, &cl, EOF, flags, callback, callback_arg, vm_initiated);

		if (retval == 0) {
			cl_pushed++;

			l_clusters[cl_index].b_addr = 0;
			l_clusters[cl_index].e_addr = 0;
		} else if (error == 0) {
			error = retval;
		}

		if (!(push_flag & PUSH_ALL)) {
			break;
		}
	}
	if (vm_initiated == TRUE) {
		lck_mtx_lock(&wbp->cl_lockw);
	}

	if (err) {
		*err = error;
	}

dont_try:
	if (cl_len > cl_pushed) {
		/*
		 * we didn't push all of the clusters, so
		 * lets try to merge them back in to the vnode
		 */
		if ((MAX_CLUSTERS - wbp->cl_number) < (cl_len - cl_pushed)) {
			/*
			 * we picked up some new clusters while we were trying to
			 * push the old ones... this can happen because I've dropped
			 * the vnode lock... the sum of the
			 * leftovers plus the new cluster count exceeds our ability
			 * to represent them, so switch to the sparse cluster mechanism
			 *
			 * collect the active public clusters...
			 */
			sparse_cluster_switch(wbp, vp, EOF, callback, callback_arg, vm_initiated);

			for (cl_index = 0, cl_index1 = 0; cl_index < cl_len; cl_index++) {
				if (l_clusters[cl_index].b_addr == l_clusters[cl_index].e_addr) {
					continue;
				}
				wbp->cl_clusters[cl_index1].b_addr = l_clusters[cl_index].b_addr;
				wbp->cl_clusters[cl_index1].e_addr = l_clusters[cl_index].e_addr;
				wbp->cl_clusters[cl_index1].io_flags = l_clusters[cl_index].io_flags;

				cl_index1++;
			}
			/*
			 * update the cluster count
			 */
			wbp->cl_number = cl_index1;

			/*
			 * and collect the original clusters that were moved into the
			 * local storage for sorting purposes
			 */
			sparse_cluster_switch(wbp, vp, EOF, callback, callback_arg, vm_initiated);
		} else {
			/*
			 * we've got room to merge the leftovers back in
			 * just append them starting at the next 'hole'
			 * represented by wbp->cl_number
			 */
			for (cl_index = 0, cl_index1 = wbp->cl_number; cl_index < cl_len; cl_index++) {
				if (l_clusters[cl_index].b_addr == l_clusters[cl_index].e_addr) {
					continue;
				}

				wbp->cl_clusters[cl_index1].b_addr = l_clusters[cl_index].b_addr;
				wbp->cl_clusters[cl_index1].e_addr = l_clusters[cl_index].e_addr;
				wbp->cl_clusters[cl_index1].io_flags = l_clusters[cl_index].io_flags;

				cl_index1++;
			}
			/*
			 * update the cluster count
			 */
			wbp->cl_number = cl_index1;
		}
	}
	return MAX_CLUSTERS - wbp->cl_number;
}



static int
cluster_push_now(vnode_t vp, struct cl_extent *cl, off_t EOF, int flags,
    int (*callback)(buf_t, void *), void *callback_arg, boolean_t vm_initiated)
{
	upl_page_info_t *pl;
	upl_t            upl;
	vm_offset_t      upl_offset;
	int              upl_size;
	off_t            upl_f_offset;
	int              pages_in_upl;
	int              start_pg;
	int              last_pg;
	int              io_size;
	int              io_flags;
	int              upl_flags;
	int              bflag;
	int              size;
	int              error = 0;
	int              retval;
	kern_return_t    kret;

	if (flags & IO_PASSIVE) {
		bflag = CL_PASSIVE;
	} else {
		bflag = 0;
	}

	if (flags & IO_SKIP_ENCRYPTION) {
		bflag |= CL_ENCRYPTED;
	}

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 51)) | DBG_FUNC_START,
	    (int)cl->b_addr, (int)cl->e_addr, (int)EOF, flags, 0);

	if ((pages_in_upl = (int)(cl->e_addr - cl->b_addr)) == 0) {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 51)) | DBG_FUNC_END, 1, 0, 0, 0, 0);

		return 0;
	}
	upl_size = pages_in_upl * PAGE_SIZE;
	upl_f_offset = (off_t)(cl->b_addr * PAGE_SIZE_64);

	if (upl_f_offset + upl_size >= EOF) {
		if (upl_f_offset >= EOF) {
			/*
			 * must have truncated the file and missed
			 * clearing a dangling cluster (i.e. it's completely
			 * beyond the new EOF
			 */
			KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 51)) | DBG_FUNC_END, 1, 1, 0, 0, 0);

			return 0;
		}
		size = (int)(EOF - upl_f_offset);

		upl_size = (size + (PAGE_SIZE - 1)) & ~PAGE_MASK;
		pages_in_upl = upl_size / PAGE_SIZE;
	} else {
		size = upl_size;
	}


	if (vm_initiated) {
		vnode_pageout(vp, NULL, (upl_offset_t)0, upl_f_offset, (upl_size_t)upl_size,
		    UPL_MSYNC | UPL_VNODE_PAGER | UPL_KEEPCACHED, &error);

		return error;
	}
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 41)) | DBG_FUNC_START, upl_size, size, 0, 0, 0);

	/*
	 * by asking for UPL_COPYOUT_FROM and UPL_RET_ONLY_DIRTY, we get the following desirable behavior
	 *
	 * - only pages that are currently dirty are returned... these are the ones we need to clean
	 * - the hardware dirty bit is cleared when the page is gathered into the UPL... the software dirty bit is set
	 * - if we have to abort the I/O for some reason, the software dirty bit is left set since we didn't clean the page
	 * - when we commit the page, the software dirty bit is cleared... the hardware dirty bit is untouched so that if
	 *   someone dirties this page while the I/O is in progress, we don't lose track of the new state
	 *
	 * when the I/O completes, we no longer ask for an explicit clear of the DIRTY state (either soft or hard)
	 */

	if ((vp->v_flag & VNOCACHE_DATA) || (flags & IO_NOCACHE)) {
		upl_flags = UPL_COPYOUT_FROM | UPL_RET_ONLY_DIRTY | UPL_SET_LITE | UPL_WILL_BE_DUMPED;
	} else {
		upl_flags = UPL_COPYOUT_FROM | UPL_RET_ONLY_DIRTY | UPL_SET_LITE;
	}

	kret = ubc_create_upl_kernel(vp,
	    upl_f_offset,
	    upl_size,
	    &upl,
	    &pl,
	    upl_flags,
	    VM_KERN_MEMORY_FILE);
	if (kret != KERN_SUCCESS) {
		panic("cluster_push: failed to get pagelist");
	}

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 41)) | DBG_FUNC_END, upl, upl_f_offset, 0, 0, 0);

	/*
	 * since we only asked for the dirty pages back
	 * it's possible that we may only get a few or even none, so...
	 * before we start marching forward, we must make sure we know
	 * where the last present page is in the UPL, otherwise we could
	 * end up working with a freed upl due to the FREE_ON_EMPTY semantics
	 * employed by commit_range and abort_range.
	 */
	for (last_pg = pages_in_upl - 1; last_pg >= 0; last_pg--) {
		if (upl_page_present(pl, last_pg)) {
			break;
		}
	}
	pages_in_upl = last_pg + 1;

	if (pages_in_upl == 0) {
		ubc_upl_abort(upl, 0);

		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 51)) | DBG_FUNC_END, 1, 2, 0, 0, 0);
		return 0;
	}

	for (last_pg = 0; last_pg < pages_in_upl;) {
		/*
		 * find the next dirty page in the UPL
		 * this will become the first page in the
		 * next I/O to generate
		 */
		for (start_pg = last_pg; start_pg < pages_in_upl; start_pg++) {
			if (upl_dirty_page(pl, start_pg)) {
				break;
			}
			if (upl_page_present(pl, start_pg)) {
				/*
				 * RET_ONLY_DIRTY will return non-dirty 'precious' pages
				 * just release these unchanged since we're not going
				 * to steal them or change their state
				 */
				ubc_upl_abort_range(upl, start_pg * PAGE_SIZE, PAGE_SIZE, UPL_ABORT_FREE_ON_EMPTY);
			}
		}
		if (start_pg >= pages_in_upl) {
			/*
			 * done... no more dirty pages to push
			 */
			break;
		}
		if (start_pg > last_pg) {
			/*
			 * skipped over some non-dirty pages
			 */
			size -= ((start_pg - last_pg) * PAGE_SIZE);
		}

		/*
		 * find a range of dirty pages to write
		 */
		for (last_pg = start_pg; last_pg < pages_in_upl; last_pg++) {
			if (!upl_dirty_page(pl, last_pg)) {
				break;
			}
		}
		upl_offset = start_pg * PAGE_SIZE;

		io_size = min(size, (last_pg - start_pg) * PAGE_SIZE);

		io_flags = CL_THROTTLE | CL_COMMIT | CL_AGE | bflag;

		if (!(flags & IO_SYNC)) {
			io_flags |= CL_ASYNC;
		}

		if (flags & IO_CLOSE) {
			io_flags |= CL_CLOSE;
		}

		if (flags & IO_NOCACHE) {
			io_flags |= CL_NOCACHE;
		}

		retval = cluster_io(vp, upl, upl_offset, upl_f_offset + upl_offset, io_size,
		    io_flags, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg);

		if (error == 0 && retval) {
			error = retval;
		}

		size -= io_size;
	}
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 51)) | DBG_FUNC_END, 1, 3, error, 0, 0);

	return error;
}


/*
 * sparse_cluster_switch is called with the write behind lock held
 */
static int
sparse_cluster_switch(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int (*callback)(buf_t, void *), void *callback_arg, boolean_t vm_initiated)
{
	int     cl_index;
	int     error = 0;

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 78)) | DBG_FUNC_START, kdebug_vnode(vp), wbp->cl_scmap, wbp->cl_number, 0, 0);

	for (cl_index = 0; cl_index < wbp->cl_number; cl_index++) {
		int       flags;
		struct cl_extent cl;

		for (cl.b_addr = wbp->cl_clusters[cl_index].b_addr; cl.b_addr < wbp->cl_clusters[cl_index].e_addr; cl.b_addr++) {
			if (ubc_page_op(vp, (off_t)(cl.b_addr * PAGE_SIZE_64), 0, NULL, &flags) == KERN_SUCCESS) {
				if (flags & UPL_POP_DIRTY) {
					cl.e_addr = cl.b_addr + 1;

					error = sparse_cluster_add(wbp, &(wbp->cl_scmap), vp, &cl, EOF, callback, callback_arg, vm_initiated);

					if (error) {
						break;
					}
				}
			}
		}
	}
	wbp->cl_number -= cl_index;

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 78)) | DBG_FUNC_END, kdebug_vnode(vp), wbp->cl_scmap, wbp->cl_number, error, 0);

	return error;
}


/*
 * sparse_cluster_push must be called with the write-behind lock held if the scmap is
 * still associated with the write-behind context... however, if the scmap has been disassociated
 * from the write-behind context (the cluster_push case), the wb lock is not held
 */
static int
sparse_cluster_push(struct cl_writebehind *wbp, void **scmap, vnode_t vp, off_t EOF, int push_flag,
    int io_flags, int (*callback)(buf_t, void *), void *callback_arg, boolean_t vm_initiated)
{
	struct cl_extent cl;
	off_t           offset;
	u_int           length;
	void            *l_scmap;
	int error = 0;

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 79)) | DBG_FUNC_START, kdebug_vnode(vp), (*scmap), 0, push_flag, 0);

	if (push_flag & PUSH_ALL) {
		vfs_drt_control(scmap, 1);
	}

	l_scmap = *scmap;

	for (;;) {
		int retval;

		if (vfs_drt_get_cluster(scmap, &offset, &length) != KERN_SUCCESS) {
			/*
			 * Not finding anything to push will return KERN_FAILURE.
			 * Confusing since it isn't really a failure. But that's the
			 * reason we don't set 'error' here like we do below.
			 */
			break;
		}

		if (vm_initiated == TRUE) {
			lck_mtx_unlock(&wbp->cl_lockw);
		}

		cl.b_addr = (daddr64_t)(offset / PAGE_SIZE_64);
		cl.e_addr = (daddr64_t)((offset + length) / PAGE_SIZE_64);

		retval = cluster_push_now(vp, &cl, EOF, io_flags, callback, callback_arg, vm_initiated);
		if (error == 0 && retval) {
			error = retval;
		}

		if (vm_initiated == TRUE) {
			lck_mtx_lock(&wbp->cl_lockw);

			if (*scmap != l_scmap) {
				break;
			}
		}

		if (error) {
			if (vfs_drt_mark_pages(scmap, offset, length, NULL) != KERN_SUCCESS) {
				panic("Failed to restore dirty state on failure");
			}

			break;
		}

		if (!(push_flag & PUSH_ALL)) {
			break;
		}
	}
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 79)) | DBG_FUNC_END, kdebug_vnode(vp), (*scmap), error, 0, 0);

	return error;
}


/*
 * sparse_cluster_add is called with the write behind lock held
 */
static int
sparse_cluster_add(struct cl_writebehind *wbp, void **scmap, vnode_t vp, struct cl_extent *cl, off_t EOF,
    int (*callback)(buf_t, void *), void *callback_arg, boolean_t vm_initiated)
{
	u_int   new_dirty;
	u_int   length;
	off_t   offset;
	int     error = 0;
	int     push_flag = 0; /* Is this a valid value? */

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 80)) | DBG_FUNC_START, (*scmap), 0, cl->b_addr, (int)cl->e_addr, 0);

	offset = (off_t)(cl->b_addr * PAGE_SIZE_64);
	length = ((u_int)(cl->e_addr - cl->b_addr)) * PAGE_SIZE;

	while (vfs_drt_mark_pages(scmap, offset, length, &new_dirty) != KERN_SUCCESS) {
		/*
		 * no room left in the map
		 * only a partial update was done
		 * push out some pages and try again
		 */

		if (vfs_get_scmap_push_behavior_internal(scmap, &push_flag)) {
			push_flag = 0;
		}

		error = sparse_cluster_push(wbp, scmap, vp, EOF, push_flag, 0, callback, callback_arg, vm_initiated);

		if (error) {
			break;
		}

		offset += (new_dirty * PAGE_SIZE_64);
		length -= (new_dirty * PAGE_SIZE);
	}
	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 80)) | DBG_FUNC_END, kdebug_vnode(vp), (*scmap), error, 0, 0);

	return error;
}


static int
cluster_align_phys_io(vnode_t vp, struct uio *uio, addr64_t usr_paddr, u_int32_t xsize, int flags, int (*callback)(buf_t, void *), void *callback_arg)
{
	upl_page_info_t  *pl;
	upl_t            upl;
	addr64_t         ubc_paddr;
	kern_return_t    kret;
	int              error = 0;
	int              did_read = 0;
	int              abort_flags;
	int              upl_flags;
	int              bflag;

	if (flags & IO_PASSIVE) {
		bflag = CL_PASSIVE;
	} else {
		bflag = 0;
	}

	if (flags & IO_NOCACHE) {
		bflag |= CL_NOCACHE;
	}

	upl_flags = UPL_SET_LITE;

	if (!(flags & CL_READ)) {
		/*
		 * "write" operation:  let the UPL subsystem know
		 * that we intend to modify the buffer cache pages
		 * we're gathering.
		 */
		upl_flags |= UPL_WILL_MODIFY;
	} else {
		/*
		 * indicate that there is no need to pull the
		 * mapping for this page... we're only going
		 * to read from it, not modify it.
		 */
		upl_flags |= UPL_FILE_IO;
	}
	kret = ubc_create_upl_kernel(vp,
	    uio->uio_offset & ~PAGE_MASK_64,
	    PAGE_SIZE,
	    &upl,
	    &pl,
	    upl_flags,
	    VM_KERN_MEMORY_FILE);

	if (kret != KERN_SUCCESS) {
		return EINVAL;
	}

	if (!upl_valid_page(pl, 0)) {
		/*
		 * issue a synchronous read to cluster_io
		 */
		error = cluster_io(vp, upl, 0, uio->uio_offset & ~PAGE_MASK_64, PAGE_SIZE,
		    CL_READ | bflag, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg);
		if (error) {
			ubc_upl_abort_range(upl, 0, PAGE_SIZE, UPL_ABORT_DUMP_PAGES | UPL_ABORT_FREE_ON_EMPTY);

			return error;
		}
		did_read = 1;
	}
	ubc_paddr = ((addr64_t)upl_phys_page(pl, 0) << PAGE_SHIFT) + (addr64_t)(uio->uio_offset & PAGE_MASK_64);

/*
 *	NOTE:  There is no prototype for the following in BSD. It, and the definitions
 *	of the defines for cppvPsrc, cppvPsnk, cppvFsnk, and cppvFsrc will be found in
 *	osfmk/ppc/mappings.h.  They are not included here because there appears to be no
 *	way to do so without exporting them to kexts as well.
 */
	if (flags & CL_READ) {
//		copypv(ubc_paddr, usr_paddr, xsize, cppvPsrc | cppvPsnk | cppvFsnk);	/* Copy physical to physical and flush the destination */
		copypv(ubc_paddr, usr_paddr, xsize, 2 |        1 |        4);           /* Copy physical to physical and flush the destination */
	} else {
//		copypv(usr_paddr, ubc_paddr, xsize, cppvPsrc | cppvPsnk | cppvFsrc);	/* Copy physical to physical and flush the source */
		copypv(usr_paddr, ubc_paddr, xsize, 2 |        1 |        8);           /* Copy physical to physical and flush the source */
	}
	if (!(flags & CL_READ) || (upl_valid_page(pl, 0) && upl_dirty_page(pl, 0))) {
		/*
		 * issue a synchronous write to cluster_io
		 */
		error = cluster_io(vp, upl, 0, uio->uio_offset & ~PAGE_MASK_64, PAGE_SIZE,
		    bflag, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg);
	}
	if (error == 0) {
		uio_update(uio, (user_size_t)xsize);
	}

	if (did_read) {
		abort_flags = UPL_ABORT_FREE_ON_EMPTY;
	} else {
		abort_flags = UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_DUMP_PAGES;
	}

	ubc_upl_abort_range(upl, 0, PAGE_SIZE, abort_flags);

	return error;
}

int
cluster_copy_upl_data(struct uio *uio, upl_t upl, int upl_offset, int *io_resid)
{
	int       pg_offset;
	int       pg_index;
	int       csize;
	int       segflg;
	int       retval = 0;
	int       xsize;
	upl_page_info_t *pl;
	int       dirty_count;

	xsize = *io_resid;

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 34)) | DBG_FUNC_START,
	    (int)uio->uio_offset, upl_offset, xsize, 0, 0);

	segflg = uio->uio_segflg;

	switch (segflg) {
	case UIO_USERSPACE32:
	case UIO_USERISPACE32:
		uio->uio_segflg = UIO_PHYS_USERSPACE32;
		break;

	case UIO_USERSPACE:
	case UIO_USERISPACE:
		uio->uio_segflg = UIO_PHYS_USERSPACE;
		break;

	case UIO_USERSPACE64:
	case UIO_USERISPACE64:
		uio->uio_segflg = UIO_PHYS_USERSPACE64;
		break;

	case UIO_SYSSPACE:
		uio->uio_segflg = UIO_PHYS_SYSSPACE;
		break;
	}
	pl = ubc_upl_pageinfo(upl);

	pg_index  = upl_offset / PAGE_SIZE;
	pg_offset = upl_offset & PAGE_MASK;
	csize     = min(PAGE_SIZE - pg_offset, xsize);

	dirty_count = 0;
	while (xsize && retval == 0) {
		addr64_t  paddr;
		ppnum_t pn = upl_phys_page(pl, pg_index);

		paddr = ((addr64_t)pn << PAGE_SHIFT) + pg_offset;
		if ((uio->uio_rw == UIO_WRITE) && (upl_dirty_page(pl, pg_index) == FALSE)) {
			dirty_count++;
		}

		/* such phyiscal pages should never be restricted pages */
		if (pmap_is_page_restricted(pn)) {
			panic("%s: cannot uiomove64 into a restricted page", __func__);
		}

		retval = uiomove64(paddr, csize, uio);

		pg_index += 1;
		pg_offset = 0;
		xsize    -= csize;
		csize     = min(PAGE_SIZE, xsize);
	}
	*io_resid = xsize;

	uio->uio_segflg = segflg;

	if (dirty_count) {
		task_update_logical_writes(current_task(), (dirty_count * PAGE_SIZE), TASK_WRITE_DEFERRED, upl_lookup_vnode(upl));
	}

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 34)) | DBG_FUNC_END,
	    (int)uio->uio_offset, xsize, retval, segflg, 0);

	return retval;
}


int
cluster_copy_ubc_data(vnode_t vp, struct uio *uio, int *io_resid, int mark_dirty)
{
	return cluster_copy_ubc_data_internal(vp, uio, io_resid, mark_dirty, 1);
}


static int
cluster_copy_ubc_data_internal(vnode_t vp, struct uio *uio, int *io_resid, int mark_dirty, int take_reference)
{
	int       segflg;
	int       io_size;
	int       xsize;
	int       start_offset;
	int       retval = 0;
	memory_object_control_t  control;

	io_size = *io_resid;

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 34)) | DBG_FUNC_START,
	    (int)uio->uio_offset, io_size, mark_dirty, take_reference, 0);

	control = ubc_getobject(vp, UBC_FLAGS_NONE);

	if (control == MEMORY_OBJECT_CONTROL_NULL) {
		KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 34)) | DBG_FUNC_END,
		    (int)uio->uio_offset, io_size, retval, 3, 0);

		return 0;
	}
	segflg = uio->uio_segflg;

	switch (segflg) {
	case UIO_USERSPACE32:
	case UIO_USERISPACE32:
		uio->uio_segflg = UIO_PHYS_USERSPACE32;
		break;

	case UIO_USERSPACE64:
	case UIO_USERISPACE64:
		uio->uio_segflg = UIO_PHYS_USERSPACE64;
		break;

	case UIO_USERSPACE:
	case UIO_USERISPACE:
		uio->uio_segflg = UIO_PHYS_USERSPACE;
		break;

	case UIO_SYSSPACE:
		uio->uio_segflg = UIO_PHYS_SYSSPACE;
		break;
	}

	if ((io_size = *io_resid)) {
		start_offset = (int)(uio->uio_offset & PAGE_MASK_64);
		xsize = (int)uio_resid(uio);

		retval = memory_object_control_uiomove(control, uio->uio_offset - start_offset, uio,
		    start_offset, io_size, mark_dirty, take_reference);
		xsize -= uio_resid(uio);

		int num_bytes_copied = xsize;
		if (num_bytes_copied && uio_rw(uio)) {
			task_update_logical_writes(current_task(), num_bytes_copied, TASK_WRITE_DEFERRED, vp);
		}
		io_size -= xsize;
	}
	uio->uio_segflg = segflg;
	*io_resid       = io_size;

	KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 34)) | DBG_FUNC_END,
	    (int)uio->uio_offset, io_size, retval, 0x80000000 | segflg, 0);

	return retval;
}


int
is_file_clean(vnode_t vp, off_t filesize)
{
	off_t f_offset;
	int   flags;
	int   total_dirty = 0;

	for (f_offset = 0; f_offset < filesize; f_offset += PAGE_SIZE_64) {
		if (ubc_page_op(vp, f_offset, 0, NULL, &flags) == KERN_SUCCESS) {
			if (flags & UPL_POP_DIRTY) {
				total_dirty++;
			}
		}
	}
	if (total_dirty) {
		return EINVAL;
	}

	return 0;
}



/*
 * Dirty region tracking/clustering mechanism.
 *
 * This code (vfs_drt_*) provides a mechanism for tracking and clustering
 * dirty regions within a larger space (file).  It is primarily intended to
 * support clustering in large files with many dirty areas.
 *
 * The implementation assumes that the dirty regions are pages.
 *
 * To represent dirty pages within the file, we store bit vectors in a
 * variable-size circular hash.
 */

/*
 * Bitvector size.  This determines the number of pages we group in a
 * single hashtable entry.  Each hashtable entry is aligned to this
 * size within the file.
 */
#define DRT_BITVECTOR_PAGES             ((1024 * 256) / PAGE_SIZE)

/*
 * File offset handling.
 *
 * DRT_ADDRESS_MASK is dependent on DRT_BITVECTOR_PAGES;
 * the correct formula is  (~((DRT_BITVECTOR_PAGES * PAGE_SIZE) - 1))
 */
#define DRT_ADDRESS_MASK                (~((DRT_BITVECTOR_PAGES * PAGE_SIZE) - 1))
#define DRT_ALIGN_ADDRESS(addr)         ((addr) & DRT_ADDRESS_MASK)

/*
 * Hashtable address field handling.
 *
 * The low-order bits of the hashtable address are used to conserve
 * space.
 *
 * DRT_HASH_COUNT_MASK must be large enough to store the range
 * 0-DRT_BITVECTOR_PAGES inclusive, as well as have one value
 * to indicate that the bucket is actually unoccupied.
 */
#define DRT_HASH_GET_ADDRESS(scm, i)    ((scm)->scm_hashtable[(i)].dhe_control & DRT_ADDRESS_MASK)
#define DRT_HASH_SET_ADDRESS(scm, i, a)                                                                 \
	do {                                                                                            \
	        (scm)->scm_hashtable[(i)].dhe_control =                                                 \
	            ((scm)->scm_hashtable[(i)].dhe_control & ~DRT_ADDRESS_MASK) | DRT_ALIGN_ADDRESS(a); \
	} while (0)
#define DRT_HASH_COUNT_MASK             0x1ff
#define DRT_HASH_GET_COUNT(scm, i)      ((scm)->scm_hashtable[(i)].dhe_control & DRT_HASH_COUNT_MASK)
#define DRT_HASH_SET_COUNT(scm, i, c)                                                                                   \
	do {                                                                                                            \
	        (scm)->scm_hashtable[(i)].dhe_control =                                                                 \
	            ((scm)->scm_hashtable[(i)].dhe_control & ~DRT_HASH_COUNT_MASK) | ((c) & DRT_HASH_COUNT_MASK);       \
	} while (0)
#define DRT_HASH_CLEAR(scm, i)                                                                                          \
	do {                                                                                                            \
	        (scm)->scm_hashtable[(i)].dhe_control =	0;                                                              \
	} while (0)
#define DRT_HASH_VACATE(scm, i)         DRT_HASH_SET_COUNT((scm), (i), DRT_HASH_COUNT_MASK)
#define DRT_HASH_VACANT(scm, i)         (DRT_HASH_GET_COUNT((scm), (i)) == DRT_HASH_COUNT_MASK)
#define DRT_HASH_COPY(oscm, oi, scm, i)                                                                 \
	do {                                                                                            \
	        (scm)->scm_hashtable[(i)].dhe_control = (oscm)->scm_hashtable[(oi)].dhe_control;        \
	        DRT_BITVECTOR_COPY(oscm, oi, scm, i);                                                   \
	} while(0);


#if !defined(XNU_TARGET_OS_OSX)
/*
 * Hash table moduli.
 *
 * Since the hashtable entry's size is dependent on the size of
 * the bitvector, and since the hashtable size is constrained to
 * both being prime and fitting within the desired allocation
 * size, these values need to be manually determined.
 *
 * For DRT_BITVECTOR_SIZE = 64, the entry size is 16 bytes.
 *
 * The small hashtable allocation is 4096 bytes, so the modulus is 251.
 * The large hashtable allocation is 32768 bytes, so the modulus is 2039.
 * The xlarge hashtable allocation is 131072 bytes, so the modulus is 8179.
 */

#define DRT_HASH_SMALL_MODULUS  251
#define DRT_HASH_LARGE_MODULUS  2039
#define DRT_HASH_XLARGE_MODULUS  8179

/*
 * Physical memory required before the large hash modulus is permitted.
 *
 * On small memory systems, the large hash modulus can lead to phsyical
 * memory starvation, so we avoid using it there.
 */
#define DRT_HASH_LARGE_MEMORY_REQUIRED  (1024LL * 1024LL * 1024LL)      /* 1GiB */
#define DRT_HASH_XLARGE_MEMORY_REQUIRED  (8 * 1024LL * 1024LL * 1024LL)  /* 8GiB */

#define DRT_SMALL_ALLOCATION    4096    /* 80 bytes spare */
#define DRT_LARGE_ALLOCATION    32768   /* 144 bytes spare */
#define DRT_XLARGE_ALLOCATION    131072  /* 208 bytes spare */

#else /* XNU_TARGET_OS_OSX */
/*
 * Hash table moduli.
 *
 * Since the hashtable entry's size is dependent on the size of
 * the bitvector, and since the hashtable size is constrained to
 * both being prime and fitting within the desired allocation
 * size, these values need to be manually determined.
 *
 * For DRT_BITVECTOR_SIZE = 64, the entry size is 16 bytes.
 *
 * The small hashtable allocation is 16384 bytes, so the modulus is 1019.
 * The large hashtable allocation is 131072 bytes, so the modulus is 8179.
 * The xlarge hashtable allocation is 524288 bytes, so the modulus is 32749.
 */

#define DRT_HASH_SMALL_MODULUS  1019
#define DRT_HASH_LARGE_MODULUS  8179
#define DRT_HASH_XLARGE_MODULUS  32749

/*
 * Physical memory required before the large hash modulus is permitted.
 *
 * On small memory systems, the large hash modulus can lead to phsyical
 * memory starvation, so we avoid using it there.
 */
#define DRT_HASH_LARGE_MEMORY_REQUIRED  (4 * 1024LL * 1024LL * 1024LL)  /* 4GiB */
#define DRT_HASH_XLARGE_MEMORY_REQUIRED  (32 * 1024LL * 1024LL * 1024LL)  /* 32GiB */

#define DRT_SMALL_ALLOCATION    16384   /* 80 bytes spare */
#define DRT_LARGE_ALLOCATION    131072  /* 208 bytes spare */
#define DRT_XLARGE_ALLOCATION   524288  /* 304 bytes spare */

#endif /* ! XNU_TARGET_OS_OSX */

/* *** nothing below here has secret dependencies on DRT_BITVECTOR_PAGES *** */

/*
 * Hashtable entry.
 */
struct vfs_drt_hashentry {
	u_int64_t       dhe_control;
/*
 * dhe_bitvector was declared as dhe_bitvector[DRT_BITVECTOR_PAGES / 32];
 * DRT_BITVECTOR_PAGES is defined as ((1024 * 256) / PAGE_SIZE)
 * Since PAGE_SIZE is only known at boot time,
 *	-define MAX_DRT_BITVECTOR_PAGES for smallest supported page size (4k)
 *	-declare dhe_bitvector array for largest possible length
 */
#define MAX_DRT_BITVECTOR_PAGES (1024 * 256)/( 4 * 1024)
	u_int32_t       dhe_bitvector[MAX_DRT_BITVECTOR_PAGES / 32];
};

/*
 * Hashtable bitvector handling.
 *
 * Bitvector fields are 32 bits long.
 */

#define DRT_HASH_SET_BIT(scm, i, bit)                           \
	(scm)->scm_hashtable[(i)].dhe_bitvector[(bit) / 32] |= (1 << ((bit) % 32))

#define DRT_HASH_CLEAR_BIT(scm, i, bit)                         \
	(scm)->scm_hashtable[(i)].dhe_bitvector[(bit) / 32] &= ~(1 << ((bit) % 32))

#define DRT_HASH_TEST_BIT(scm, i, bit)                          \
	((scm)->scm_hashtable[(i)].dhe_bitvector[(bit) / 32] & (1 << ((bit) % 32)))

#define DRT_BITVECTOR_CLEAR(scm, i)                             \
	bzero(&(scm)->scm_hashtable[(i)].dhe_bitvector[0], (MAX_DRT_BITVECTOR_PAGES / 32) * sizeof(u_int32_t))

#define DRT_BITVECTOR_COPY(oscm, oi, scm, i)                    \
	bcopy(&(oscm)->scm_hashtable[(oi)].dhe_bitvector[0],    \
	    &(scm)->scm_hashtable[(i)].dhe_bitvector[0],        \
	    (MAX_DRT_BITVECTOR_PAGES / 32) * sizeof(u_int32_t))

/*
 * Dirty Region Tracking structure.
 *
 * The hashtable is allocated entirely inside the DRT structure.
 *
 * The hash is a simple circular prime modulus arrangement, the structure
 * is resized from small to large if it overflows.
 */

struct vfs_drt_clustermap {
	u_int32_t               scm_magic;      /* sanity/detection */
#define DRT_SCM_MAGIC           0x12020003
	u_int32_t               scm_modulus;    /* current ring size */
	u_int32_t               scm_buckets;    /* number of occupied buckets */
	u_int32_t               scm_lastclean;  /* last entry we cleaned */
	u_int32_t               scm_iskips;     /* number of slot skips */

	struct vfs_drt_hashentry scm_hashtable[0];
};


#define DRT_HASH(scm, addr)             ((addr) % (scm)->scm_modulus)
#define DRT_HASH_NEXT(scm, addr)        (((addr) + 1) % (scm)->scm_modulus)

/*
 * Debugging codes and arguments.
 */
#define DRT_DEBUG_EMPTYFREE     (FSDBG_CODE(DBG_FSRW, 82)) /* nil */
#define DRT_DEBUG_RETCLUSTER    (FSDBG_CODE(DBG_FSRW, 83)) /* offset, length */
#define DRT_DEBUG_ALLOC         (FSDBG_CODE(DBG_FSRW, 84)) /* copycount */
#define DRT_DEBUG_INSERT        (FSDBG_CODE(DBG_FSRW, 85)) /* offset, iskip */
#define DRT_DEBUG_MARK          (FSDBG_CODE(DBG_FSRW, 86)) /* offset, length,
	                                                    * dirty */
                                                           /* 0, setcount */
                                                           /* 1 (clean, no map) */
                                                           /* 2 (map alloc fail) */
                                                           /* 3, resid (partial) */
#define DRT_DEBUG_6             (FSDBG_CODE(DBG_FSRW, 87))
#define DRT_DEBUG_SCMDATA       (FSDBG_CODE(DBG_FSRW, 88)) /* modulus, buckets,
	                                                    * lastclean, iskips */


static kern_return_t    vfs_drt_alloc_map(struct vfs_drt_clustermap **cmapp);
static kern_return_t    vfs_drt_free_map(struct vfs_drt_clustermap *cmap);
static kern_return_t    vfs_drt_search_index(struct vfs_drt_clustermap *cmap,
    u_int64_t offset, int *indexp);
static kern_return_t    vfs_drt_get_index(struct vfs_drt_clustermap **cmapp,
    u_int64_t offset,
    int *indexp,
    int recursed);
static kern_return_t    vfs_drt_do_mark_pages(
	void            **cmapp,
	u_int64_t       offset,
	u_int           length,
	u_int           *setcountp,
	int             dirty);
static void             vfs_drt_trace(
	struct vfs_drt_clustermap *cmap,
	int code,
	int arg1,
	int arg2,
	int arg3,
	int arg4);


/*
 * Allocate and initialise a sparse cluster map.
 *
 * Will allocate a new map, resize or compact an existing map.
 *
 * XXX we should probably have at least one intermediate map size,
 * as the 1:16 ratio seems a bit drastic.
 */
static kern_return_t
vfs_drt_alloc_map(struct vfs_drt_clustermap **cmapp)
{
	struct vfs_drt_clustermap *cmap = NULL, *ocmap = NULL;
	kern_return_t   kret = KERN_SUCCESS;
	u_int64_t       offset = 0;
	u_int32_t       i = 0;
	int             modulus_size = 0, map_size = 0, active_buckets = 0, index = 0, copycount = 0;

	ocmap = NULL;
	if (cmapp != NULL) {
		ocmap = *cmapp;
	}

	/*
	 * Decide on the size of the new map.
	 */
	if (ocmap == NULL) {
		modulus_size = DRT_HASH_SMALL_MODULUS;
		map_size = DRT_SMALL_ALLOCATION;
	} else {
		/* count the number of active buckets in the old map */
		active_buckets = 0;
		for (i = 0; i < ocmap->scm_modulus; i++) {
			if (!DRT_HASH_VACANT(ocmap, i) &&
			    (DRT_HASH_GET_COUNT(ocmap, i) != 0)) {
				active_buckets++;
			}
		}
		/*
		 * If we're currently using the small allocation, check to
		 * see whether we should grow to the large one.
		 */
		if (ocmap->scm_modulus == DRT_HASH_SMALL_MODULUS) {
			/*
			 * If the ring is nearly full and we are allowed to
			 * use the large modulus, upgrade.
			 */
			if ((active_buckets > (DRT_HASH_SMALL_MODULUS - 5)) &&
			    (max_mem >= DRT_HASH_LARGE_MEMORY_REQUIRED)) {
				modulus_size = DRT_HASH_LARGE_MODULUS;
				map_size = DRT_LARGE_ALLOCATION;
			} else {
				modulus_size = DRT_HASH_SMALL_MODULUS;
				map_size = DRT_SMALL_ALLOCATION;
			}
		} else if (ocmap->scm_modulus == DRT_HASH_LARGE_MODULUS) {
			if ((active_buckets > (DRT_HASH_LARGE_MODULUS - 5)) &&
			    (max_mem >= DRT_HASH_XLARGE_MEMORY_REQUIRED)) {
				modulus_size = DRT_HASH_XLARGE_MODULUS;
				map_size = DRT_XLARGE_ALLOCATION;
			} else {
				/*
				 * If the ring is completely full and we can't
				 * expand, there's nothing useful for us to do.
				 * Behave as though we had compacted into the new
				 * array and return.
				 */
				return KERN_SUCCESS;
			}
		} else {
			/* already using the xlarge modulus */
			modulus_size = DRT_HASH_XLARGE_MODULUS;
			map_size = DRT_XLARGE_ALLOCATION;

			/*
			 * If the ring is completely full, there's
			 * nothing useful for us to do.  Behave as
			 * though we had compacted into the new
			 * array and return.
			 */
			if (active_buckets >= DRT_HASH_XLARGE_MODULUS) {
				return KERN_SUCCESS;
			}
		}
	}

	/*
	 * Allocate and initialise the new map.
	 */

	kret = kmem_alloc(kernel_map, (vm_offset_t *)&cmap, map_size,
	    KMA_DATA, VM_KERN_MEMORY_FILE);
	if (kret != KERN_SUCCESS) {
		return kret;
	}
	cmap->scm_magic = DRT_SCM_MAGIC;
	cmap->scm_modulus = modulus_size;
	cmap->scm_buckets = 0;
	cmap->scm_lastclean = 0;
	cmap->scm_iskips = 0;
	for (i = 0; i < cmap->scm_modulus; i++) {
		DRT_HASH_CLEAR(cmap, i);
		DRT_HASH_VACATE(cmap, i);
		DRT_BITVECTOR_CLEAR(cmap, i);
	}

	/*
	 * If there's an old map, re-hash entries from it into the new map.
	 */
	copycount = 0;
	if (ocmap != NULL) {
		for (i = 0; i < ocmap->scm_modulus; i++) {
			/* skip empty buckets */
			if (DRT_HASH_VACANT(ocmap, i) ||
			    (DRT_HASH_GET_COUNT(ocmap, i) == 0)) {
				continue;
			}
			/* get new index */
			offset = DRT_HASH_GET_ADDRESS(ocmap, i);
			kret = vfs_drt_get_index(&cmap, offset, &index, 1);
			if (kret != KERN_SUCCESS) {
				/* XXX need to bail out gracefully here */
				panic("vfs_drt: new cluster map mysteriously too small");
				index = 0;
			}
			/* copy */
			DRT_HASH_COPY(ocmap, i, cmap, index);
			copycount++;
		}
	}

	/* log what we've done */
	vfs_drt_trace(cmap, DRT_DEBUG_ALLOC, copycount, 0, 0, 0);

	/*
	 * It's important to ensure that *cmapp always points to
	 * a valid map, so we must overwrite it before freeing
	 * the old map.
	 */
	*cmapp = cmap;
	if (ocmap != NULL) {
		/* emit stats into trace buffer */
		vfs_drt_trace(ocmap, DRT_DEBUG_SCMDATA,
		    ocmap->scm_modulus,
		    ocmap->scm_buckets,
		    ocmap->scm_lastclean,
		    ocmap->scm_iskips);

		vfs_drt_free_map(ocmap);
	}
	return KERN_SUCCESS;
}


/*
 * Free a sparse cluster map.
 */
static kern_return_t
vfs_drt_free_map(struct vfs_drt_clustermap *cmap)
{
	vm_size_t map_size = 0;

	if (cmap->scm_modulus == DRT_HASH_SMALL_MODULUS) {
		map_size = DRT_SMALL_ALLOCATION;
	} else if (cmap->scm_modulus == DRT_HASH_LARGE_MODULUS) {
		map_size = DRT_LARGE_ALLOCATION;
	} else if (cmap->scm_modulus == DRT_HASH_XLARGE_MODULUS) {
		map_size = DRT_XLARGE_ALLOCATION;
	} else {
		panic("vfs_drt_free_map: Invalid modulus %d", cmap->scm_modulus);
	}

	kmem_free(kernel_map, (vm_offset_t)cmap, map_size);
	return KERN_SUCCESS;
}


/*
 * Find the hashtable slot currently occupied by an entry for the supplied offset.
 */
static kern_return_t
vfs_drt_search_index(struct vfs_drt_clustermap *cmap, u_int64_t offset, int *indexp)
{
	int             index;
	u_int32_t       i;

	offset = DRT_ALIGN_ADDRESS(offset);
	index = DRT_HASH(cmap, offset);

	/* traverse the hashtable */
	for (i = 0; i < cmap->scm_modulus; i++) {
		/*
		 * If the slot is vacant, we can stop.
		 */
		if (DRT_HASH_VACANT(cmap, index)) {
			break;
		}

		/*
		 * If the address matches our offset, we have success.
		 */
		if (DRT_HASH_GET_ADDRESS(cmap, index) == offset) {
			*indexp = index;
			return KERN_SUCCESS;
		}

		/*
		 * Move to the next slot, try again.
		 */
		index = DRT_HASH_NEXT(cmap, index);
	}
	/*
	 * It's not there.
	 */
	return KERN_FAILURE;
}

/*
 * Find the hashtable slot for the supplied offset.  If we haven't allocated
 * one yet, allocate one and populate the address field.  Note that it will
 * not have a nonzero page count and thus will still technically be free, so
 * in the case where we are called to clean pages, the slot will remain free.
 */
static kern_return_t
vfs_drt_get_index(struct vfs_drt_clustermap **cmapp, u_int64_t offset, int *indexp, int recursed)
{
	struct vfs_drt_clustermap *cmap;
	kern_return_t   kret;
	u_int32_t       index;
	u_int32_t       i;

	cmap = *cmapp;

	/* look for an existing entry */
	kret = vfs_drt_search_index(cmap, offset, indexp);
	if (kret == KERN_SUCCESS) {
		return kret;
	}

	/* need to allocate an entry */
	offset = DRT_ALIGN_ADDRESS(offset);
	index = DRT_HASH(cmap, offset);

	/* scan from the index forwards looking for a vacant slot */
	for (i = 0; i < cmap->scm_modulus; i++) {
		/* slot vacant? */
		if (DRT_HASH_VACANT(cmap, index) || DRT_HASH_GET_COUNT(cmap, index) == 0) {
			cmap->scm_buckets++;
			if (index < cmap->scm_lastclean) {
				cmap->scm_lastclean = index;
			}
			DRT_HASH_SET_ADDRESS(cmap, index, offset);
			DRT_HASH_SET_COUNT(cmap, index, 0);
			DRT_BITVECTOR_CLEAR(cmap, index);
			*indexp = index;
			vfs_drt_trace(cmap, DRT_DEBUG_INSERT, (int)offset, i, 0, 0);
			return KERN_SUCCESS;
		}
		cmap->scm_iskips += i;
		index = DRT_HASH_NEXT(cmap, index);
	}

	/*
	 * We haven't found a vacant slot, so the map is full.  If we're not
	 * already recursed, try reallocating/compacting it.
	 */
	if (recursed) {
		return KERN_FAILURE;
	}
	kret = vfs_drt_alloc_map(cmapp);
	if (kret == KERN_SUCCESS) {
		/* now try to insert again */
		kret = vfs_drt_get_index(cmapp, offset, indexp, 1);
	}
	return kret;
}

/*
 * Implementation of set dirty/clean.
 *
 * In the 'clean' case, not finding a map is OK.
 */
static kern_return_t
vfs_drt_do_mark_pages(
	void            **private,
	u_int64_t       offset,
	u_int           length,
	u_int           *setcountp,
	int             dirty)
{
	struct vfs_drt_clustermap *cmap, **cmapp;
	kern_return_t   kret;
	int             i, index, pgoff, pgcount, setcount, ecount;

	cmapp = (struct vfs_drt_clustermap **)private;
	cmap = *cmapp;

	vfs_drt_trace(cmap, DRT_DEBUG_MARK | DBG_FUNC_START, (int)offset, (int)length, dirty, 0);

	if (setcountp != NULL) {
		*setcountp = 0;
	}

	/* allocate a cluster map if we don't already have one */
	if (cmap == NULL) {
		/* no cluster map, nothing to clean */
		if (!dirty) {
			vfs_drt_trace(cmap, DRT_DEBUG_MARK | DBG_FUNC_END, 1, 0, 0, 0);
			return KERN_SUCCESS;
		}
		kret = vfs_drt_alloc_map(cmapp);
		if (kret != KERN_SUCCESS) {
			vfs_drt_trace(cmap, DRT_DEBUG_MARK | DBG_FUNC_END, 2, 0, 0, 0);
			return kret;
		}
	}
	setcount = 0;

	/*
	 * Iterate over the length of the region.
	 */
	while (length > 0) {
		/*
		 * Get the hashtable index for this offset.
		 *
		 * XXX this will add blank entries if we are clearing a range
		 * that hasn't been dirtied.
		 */
		kret = vfs_drt_get_index(cmapp, offset, &index, 0);
		cmap = *cmapp;  /* may have changed! */
		/* this may be a partial-success return */
		if (kret != KERN_SUCCESS) {
			if (setcountp != NULL) {
				*setcountp = setcount;
			}
			vfs_drt_trace(cmap, DRT_DEBUG_MARK | DBG_FUNC_END, 3, (int)length, 0, 0);

			return kret;
		}

		/*
		 * Work out how many pages we're modifying in this
		 * hashtable entry.
		 */
		pgoff = (int)((offset - DRT_ALIGN_ADDRESS(offset)) / PAGE_SIZE);
		pgcount = min((length / PAGE_SIZE), (DRT_BITVECTOR_PAGES - pgoff));

		/*
		 * Iterate over pages, dirty/clearing as we go.
		 */
		ecount = DRT_HASH_GET_COUNT(cmap, index);
		for (i = 0; i < pgcount; i++) {
			if (dirty) {
				if (!DRT_HASH_TEST_BIT(cmap, index, pgoff + i)) {
					if (ecount >= DRT_BITVECTOR_PAGES) {
						panic("ecount >= DRT_BITVECTOR_PAGES, cmap = %p, index = %d, bit = %d", cmap, index, pgoff + i);
					}
					DRT_HASH_SET_BIT(cmap, index, pgoff + i);
					ecount++;
					setcount++;
				}
			} else {
				if (DRT_HASH_TEST_BIT(cmap, index, pgoff + i)) {
					if (ecount <= 0) {
						panic("ecount <= 0, cmap = %p, index = %d, bit = %d", cmap, index, pgoff + i);
					}
					assert(ecount > 0);
					DRT_HASH_CLEAR_BIT(cmap, index, pgoff + i);
					ecount--;
					setcount++;
				}
			}
		}
		DRT_HASH_SET_COUNT(cmap, index, ecount);

		offset += pgcount * PAGE_SIZE;
		length -= pgcount * PAGE_SIZE;
	}
	if (setcountp != NULL) {
		*setcountp = setcount;
	}

	vfs_drt_trace(cmap, DRT_DEBUG_MARK | DBG_FUNC_END, 0, setcount, 0, 0);

	return KERN_SUCCESS;
}

/*
 * Mark a set of pages as dirty/clean.
 *
 * This is a public interface.
 *
 * cmapp
 *	Pointer to storage suitable for holding a pointer.  Note that
 *	this must either be NULL or a value set by this function.
 *
 * size
 *	Current file size in bytes.
 *
 * offset
 *	Offset of the first page to be marked as dirty, in bytes.  Must be
 *	page-aligned.
 *
 * length
 *	Length of dirty region, in bytes.  Must be a multiple of PAGE_SIZE.
 *
 * setcountp
 *	Number of pages newly marked dirty by this call (optional).
 *
 * Returns KERN_SUCCESS if all the pages were successfully marked.
 */
static kern_return_t
vfs_drt_mark_pages(void **cmapp, off_t offset, u_int length, u_int *setcountp)
{
	/* XXX size unused, drop from interface */
	return vfs_drt_do_mark_pages(cmapp, offset, length, setcountp, 1);
}

#if 0
static kern_return_t
vfs_drt_unmark_pages(void **cmapp, off_t offset, u_int length)
{
	return vfs_drt_do_mark_pages(cmapp, offset, length, NULL, 0);
}
#endif

/*
 * Get a cluster of dirty pages.
 *
 * This is a public interface.
 *
 * cmapp
 *	Pointer to storage managed by drt_mark_pages.  Note that this must
 *	be NULL or a value set by drt_mark_pages.
 *
 * offsetp
 *	Returns the byte offset into the file of the first page in the cluster.
 *
 * lengthp
 *	Returns the length in bytes of the cluster of dirty pages.
 *
 * Returns success if a cluster was found.  If KERN_FAILURE is returned, there
 * are no dirty pages meeting the minmum size criteria.  Private storage will
 * be released if there are no more dirty pages left in the map
 *
 */
static kern_return_t
vfs_drt_get_cluster(void **cmapp, off_t *offsetp, u_int *lengthp)
{
	struct vfs_drt_clustermap *cmap;
	u_int64_t       offset;
	u_int           length;
	u_int32_t       j;
	int             index, i, fs, ls;

	/* sanity */
	if ((cmapp == NULL) || (*cmapp == NULL)) {
		return KERN_FAILURE;
	}
	cmap = *cmapp;

	/* walk the hashtable */
	for (offset = 0, j = 0; j < cmap->scm_modulus; offset += (DRT_BITVECTOR_PAGES * PAGE_SIZE), j++) {
		index = DRT_HASH(cmap, offset);

		if (DRT_HASH_VACANT(cmap, index) || (DRT_HASH_GET_COUNT(cmap, index) == 0)) {
			continue;
		}

		/* scan the bitfield for a string of bits */
		fs = -1;

		for (i = 0; i < DRT_BITVECTOR_PAGES; i++) {
			if (DRT_HASH_TEST_BIT(cmap, index, i)) {
				fs = i;
				break;
			}
		}
		if (fs == -1) {
			/*  didn't find any bits set */
			panic("vfs_drt: entry summary count > 0 but no bits set in map, cmap = %p, index = %d, count = %lld",
			    cmap, index, DRT_HASH_GET_COUNT(cmap, index));
		}
		for (ls = 0; i < DRT_BITVECTOR_PAGES; i++, ls++) {
			if (!DRT_HASH_TEST_BIT(cmap, index, i)) {
				break;
			}
		}

		/* compute offset and length, mark pages clean */
		offset = DRT_HASH_GET_ADDRESS(cmap, index) + (PAGE_SIZE * fs);
		length = ls * PAGE_SIZE;
		vfs_drt_do_mark_pages(cmapp, offset, length, NULL, 0);
		cmap->scm_lastclean = index;

		/* return successful */
		*offsetp = (off_t)offset;
		*lengthp = length;

		vfs_drt_trace(cmap, DRT_DEBUG_RETCLUSTER, (int)offset, (int)length, 0, 0);
		return KERN_SUCCESS;
	}
	/*
	 * We didn't find anything... hashtable is empty
	 * emit stats into trace buffer and
	 * then free it
	 */
	vfs_drt_trace(cmap, DRT_DEBUG_SCMDATA,
	    cmap->scm_modulus,
	    cmap->scm_buckets,
	    cmap->scm_lastclean,
	    cmap->scm_iskips);

	vfs_drt_free_map(cmap);
	*cmapp = NULL;

	return KERN_FAILURE;
}


static kern_return_t
vfs_drt_control(void **cmapp, int op_type)
{
	struct vfs_drt_clustermap *cmap;

	/* sanity */
	if ((cmapp == NULL) || (*cmapp == NULL)) {
		return KERN_FAILURE;
	}
	cmap = *cmapp;

	switch (op_type) {
	case 0:
		/* emit stats into trace buffer */
		vfs_drt_trace(cmap, DRT_DEBUG_SCMDATA,
		    cmap->scm_modulus,
		    cmap->scm_buckets,
		    cmap->scm_lastclean,
		    cmap->scm_iskips);

		vfs_drt_free_map(cmap);
		*cmapp = NULL;
		break;

	case 1:
		cmap->scm_lastclean = 0;
		break;
	}
	return KERN_SUCCESS;
}



/*
 * Emit a summary of the state of the clustermap into the trace buffer
 * along with some caller-provided data.
 */
#if KDEBUG
static void
vfs_drt_trace(__unused struct vfs_drt_clustermap *cmap, int code, int arg1, int arg2, int arg3, int arg4)
{
	KERNEL_DEBUG(code, arg1, arg2, arg3, arg4, 0);
}
#else
static void
vfs_drt_trace(__unused struct vfs_drt_clustermap *cmap, __unused int code,
    __unused int arg1, __unused int arg2, __unused int arg3,
    __unused int arg4)
{
}
#endif

#if 0
/*
 * Perform basic sanity check on the hash entry summary count
 * vs. the actual bits set in the entry.
 */
static void
vfs_drt_sanity(struct vfs_drt_clustermap *cmap)
{
	int index, i;
	int bits_on;

	for (index = 0; index < cmap->scm_modulus; index++) {
		if (DRT_HASH_VACANT(cmap, index)) {
			continue;
		}

		for (bits_on = 0, i = 0; i < DRT_BITVECTOR_PAGES; i++) {
			if (DRT_HASH_TEST_BIT(cmap, index, i)) {
				bits_on++;
			}
		}
		if (bits_on != DRT_HASH_GET_COUNT(cmap, index)) {
			panic("bits_on = %d,  index = %d", bits_on, index);
		}
	}
}
#endif

/*
 * Internal interface only.
 */
static kern_return_t
vfs_get_scmap_push_behavior_internal(void **cmapp, int *push_flag)
{
	struct vfs_drt_clustermap *cmap;

	/* sanity */
	if ((cmapp == NULL) || (*cmapp == NULL) || (push_flag == NULL)) {
		return KERN_FAILURE;
	}
	cmap = *cmapp;

	if (cmap->scm_modulus == DRT_HASH_XLARGE_MODULUS) {
		/*
		 * If we have a full xlarge sparse cluster,
		 * we push it out all at once so the cluster
		 * map can be available to absorb more I/Os.
		 * This is done on large memory configs so
		 * the small I/Os don't interfere with the
		 * pro workloads.
		 */
		*push_flag = PUSH_ALL;
	}
	return KERN_SUCCESS;
}