[apple/xnu.git] / bsd / kern / kern_symfile.c

/*
 * Copyright (c) 2000-2006 Apple Computer, 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) 1998 Apple Computer, Inc.  All rights reserved.
 *
 *	File:	bsd/kern/kern_symfile.c
 *
 * HISTORY
 */

#include <mach/vm_param.h>

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/signalvar.h>
#include <sys/resourcevar.h>
#include <sys/namei.h>
#include <sys/vnode_internal.h>
#include <sys/proc_internal.h>
#include <sys/kauth.h>
#include <sys/timeb.h>
#include <sys/times.h>
#include <sys/acct.h>
#include <sys/file_internal.h>
#include <sys/uio.h>
#include <sys/kernel.h>
#include <sys/stat.h>
#include <sys/disk.h>
#include <sys/conf.h>

#include <mach-o/loader.h>
#include <mach-o/nlist.h>

#include <kern/kalloc.h>
#include <vm/vm_kern.h>
#include <pexpert/pexpert.h>
#include <IOKit/IOHibernatePrivate.h>

/* This function is called from kern_sysctl in the current process context;
 * it is exported with the System6.0.exports, but this appears to be a legacy
 * export, as there are no internal consumers.
 */
int
get_kernel_symfile(__unused proc_t p, __unused char const **symfile);
int
get_kernel_symfile(__unused proc_t p, __unused char const **symfile)
{
    return KERN_FAILURE;
}

struct kern_direct_file_io_ref_t
{
    vfs_context_t  ctx;
    struct vnode * vp;
    dev_t          device;
    uint32_t	   blksize;
    off_t          filelength;
    char           pinned;
};


static int file_ioctl(void * p1, void * p2, u_long theIoctl, caddr_t result)
{
    dev_t device = *(dev_t*) p1;

    return ((*bdevsw[major(device)].d_ioctl)
		    (device, theIoctl, result, S_IFBLK, p2));
}

static int device_ioctl(void * p1, __unused void * p2, u_long theIoctl, caddr_t result)
{
    return (VNOP_IOCTL(p1, theIoctl, result, 0, p2));
}

static int
kern_ioctl_file_extents(struct kern_direct_file_io_ref_t * ref, u_long theIoctl, off_t offset, off_t end)
{
    int error;
    int (*do_ioctl)(void * p1, void * p2, u_long theIoctl, caddr_t result);
    void * p1;
    void * p2;
    uint64_t    fileblk;
    size_t      filechunk;
    dk_extent_t  extent;
    dk_unmap_t   unmap;
    _dk_cs_pin_t pin;

    bzero(&extent, sizeof(dk_extent_t));
    bzero(&unmap, sizeof(dk_unmap_t));
    bzero(&pin, sizeof(pin));
    if (ref->vp->v_type == VREG)
    {
	 p1 = &ref->device;
	 p2 = kernproc;
	 do_ioctl = &file_ioctl;
    }
    else
    {
	/* Partition. */
	p1 = ref->vp;
	p2 = ref->ctx;
	do_ioctl = &device_ioctl;
    }
    while (offset < end) 
    {
        if (ref->vp->v_type == VREG)
        {
            daddr64_t blkno;
	    filechunk = 1*1024*1024*1024;
	    if (filechunk > (size_t)(end - offset))
	    filechunk = (size_t)(end - offset);
            error = VNOP_BLOCKMAP(ref->vp, offset, filechunk, &blkno,
								  &filechunk, NULL, VNODE_WRITE, NULL);
			if (error) break;
            fileblk = blkno * ref->blksize;
        }
        else if ((ref->vp->v_type == VBLK) || (ref->vp->v_type == VCHR))
        {
            fileblk = offset;
            filechunk = ref->filelength;
        }

	if (DKIOCUNMAP == theIoctl)
	{
	    extent.offset = fileblk;
	    extent.length = filechunk;
	    unmap.extents = &extent;
	    unmap.extentsCount = 1;
	    error = do_ioctl(p1, p2, theIoctl, (caddr_t)&unmap);
// 	    printf("DKIOCUNMAP(%d) 0x%qx, 0x%qx\n", error, extent.offset, extent.length);
	}
	else if (_DKIOCCSPINEXTENT == theIoctl)
	{
	    pin.cp_extent.offset = fileblk;
	    pin.cp_extent.length = filechunk;
	    pin.cp_flags = _DKIOCCSPINFORHIBERNATION;
	    error = do_ioctl(p1, p2, theIoctl, (caddr_t)&pin);
	    if (error && (ENOTTY != error))
	    {
		printf("_DKIOCCSPINEXTENT(%d) 0x%qx, 0x%qx\n", 
			error, pin.cp_extent.offset, pin.cp_extent.length);
	    }
	}
	else error = EINVAL;

	if (error) break;
        offset += filechunk;
    }
    return (error);
}


struct kern_direct_file_io_ref_t *
kern_open_file_for_direct_io(const char * name, 
                 boolean_t create_file,
			     kern_get_file_extents_callback_t callback, 
			     void * callback_ref,
                             off_t set_file_size,
                             off_t write_file_offset,
                             caddr_t write_file_addr,
                             vm_size_t write_file_len,
			     dev_t * partition_device_result,
			     dev_t * image_device_result,
                             uint64_t * partitionbase_result,
                             uint64_t * maxiocount_result,
                             uint32_t * oflags)
{
    struct kern_direct_file_io_ref_t * ref;

    proc_t			p;
    struct vnode_attr		va;
    int				error;
    off_t			f_offset;
    uint64_t                    fileblk;
    size_t                      filechunk;
    uint64_t                    physoffset;
    dev_t			device;
    dev_t			target = 0;
    int			        isssd = 0;
    uint32_t                    flags = 0;
    uint32_t			blksize;
    off_t 			maxiocount, count, segcount;
    boolean_t                   locked = FALSE;

    int (*do_ioctl)(void * p1, void * p2, u_long theIoctl, caddr_t result);
    void * p1 = NULL;
    void * p2 = NULL;

    error = EFAULT;

    ref = (struct kern_direct_file_io_ref_t *) kalloc(sizeof(struct kern_direct_file_io_ref_t));
    if (!ref)
    {
	error = EFAULT;
    	goto out;
    }

    bzero(ref, sizeof(*ref));
    p = kernproc;
    ref->ctx = vfs_context_create(vfs_context_current());

    if ((error = vnode_open(name, (create_file) ? (O_CREAT | FWRITE) : FWRITE, 
                            (0), 0, &ref->vp, ref->ctx)))
        goto out;

    if (ref->vp->v_type == VREG)
    {
        vnode_lock_spin(ref->vp);
        SET(ref->vp->v_flag, VSWAP);
        vnode_unlock(ref->vp);
    }

    if (write_file_addr && write_file_len)
    {
	if ((error = kern_write_file(ref, write_file_offset, write_file_addr, write_file_len, 0)))
	    goto out;
    }

    VATTR_INIT(&va);
    VATTR_WANTED(&va, va_rdev);
    VATTR_WANTED(&va, va_fsid);
    VATTR_WANTED(&va, va_data_size);
    VATTR_WANTED(&va, va_data_alloc);
    VATTR_WANTED(&va, va_nlink);
    error = EFAULT;
    if (vnode_getattr(ref->vp, &va, ref->ctx))
    	goto out;

    kprintf("vp va_rdev major %d minor %d\n", major(va.va_rdev), minor(va.va_rdev));
    kprintf("vp va_fsid major %d minor %d\n", major(va.va_fsid), minor(va.va_fsid));
    kprintf("vp size %qd alloc %qd\n", va.va_data_size, va.va_data_alloc);

    if (ref->vp->v_type == VREG)
    {
		/* Don't dump files with links. */
		if (va.va_nlink != 1)
			goto out;

        device = va.va_fsid;
        ref->filelength = va.va_data_size;

        p1 = &device;
        p2 = p;
        do_ioctl = &file_ioctl;

		if (set_file_size)
	    {
			error = vnode_setsize(ref->vp, set_file_size, 
								  IO_NOZEROFILL | IO_NOAUTH, ref->ctx);
			if (error)
				goto out;
			ref->filelength = set_file_size;
		}
    }
    else if ((ref->vp->v_type == VBLK) || (ref->vp->v_type == VCHR))
    {
	/* Partition. */
        device = va.va_rdev;

        p1 = ref->vp;
        p2 = ref->ctx;
        do_ioctl = &device_ioctl;
    }
    else
    {
	/* Don't dump to non-regular files. */
	error = EFAULT;
        goto out;
    }
    ref->device = device;

    // get block size

    error = do_ioctl(p1, p2, DKIOCGETBLOCKSIZE, (caddr_t) &ref->blksize);
    if (error)
        goto out;

    if (ref->vp->v_type != VREG)
    {
        error = do_ioctl(p1, p2, DKIOCGETBLOCKCOUNT, (caddr_t) &fileblk);
        if (error)
            goto out;
	ref->filelength = fileblk * ref->blksize;    
    }

    // pin logical extents

    error = kern_ioctl_file_extents(ref, _DKIOCCSPINEXTENT, 0, ref->filelength);
    if (error && (ENOTTY != error)) goto out;
    ref->pinned = (error == 0);

    // generate the block list

    error = do_ioctl(p1, p2, DKIOCLOCKPHYSICALEXTENTS, NULL);
    if (error)
        goto out;
    locked = TRUE;

    f_offset = 0;
    while (f_offset < ref->filelength) 
    {
        if (ref->vp->v_type == VREG)
        {
            filechunk = 1*1024*1024*1024;
            daddr64_t blkno;

            error = VNOP_BLOCKMAP(ref->vp, f_offset, filechunk, &blkno,
								  &filechunk, NULL, VNODE_WRITE, NULL);
            if (error)
                goto out;

            fileblk = blkno * ref->blksize;
        }
        else if ((ref->vp->v_type == VBLK) || (ref->vp->v_type == VCHR))
        {
            fileblk = f_offset;
            filechunk = f_offset ? 0 : ref->filelength;
        }

        physoffset = 0;
        while (physoffset < filechunk)
        {
            dk_physical_extent_t getphysreq;
            bzero(&getphysreq, sizeof(getphysreq));

            getphysreq.offset = fileblk + physoffset;
            getphysreq.length = (filechunk - physoffset);
            error = do_ioctl(p1, p2, DKIOCGETPHYSICALEXTENT, (caddr_t) &getphysreq);
            if (error)
                goto out;
            if (!target)
            {
                target = getphysreq.dev;
            }
            else if (target != getphysreq.dev)
            {
                error = ENOTSUP;
                goto out;
            }
#if HIBFRAGMENT
	    uint64_t rev;
	    for (rev = 4096; rev <= getphysreq.length; rev += 4096)
	    {
		callback(callback_ref, getphysreq.offset + getphysreq.length - rev, 4096);
	    }
#else
            callback(callback_ref, getphysreq.offset, getphysreq.length);
#endif
            physoffset += getphysreq.length;
        }
        f_offset += filechunk;
    }
    callback(callback_ref, 0ULL, 0ULL);

    if (ref->vp->v_type == VREG)
        p1 = &target;

    // get partition base

    if (partitionbase_result) 
    {
        error = do_ioctl(p1, p2, DKIOCGETBASE, (caddr_t) partitionbase_result);
        if (error)
            goto out;
    }

    // get block size & constraints

    error = do_ioctl(p1, p2, DKIOCGETBLOCKSIZE, (caddr_t) &blksize);
    if (error)
        goto out;

    maxiocount = 1*1024*1024*1024;

    error = do_ioctl(p1, p2, DKIOCGETMAXBLOCKCOUNTREAD, (caddr_t) &count);
    if (error)
        count = 0;
    count *= blksize;
    if (count && (count < maxiocount))
        maxiocount = count;

    error = do_ioctl(p1, p2, DKIOCGETMAXBLOCKCOUNTWRITE, (caddr_t) &count);
    if (error)
        count = 0;
    count *= blksize;
    if (count && (count < maxiocount))
        maxiocount = count;

    error = do_ioctl(p1, p2, DKIOCGETMAXBYTECOUNTREAD, (caddr_t) &count);
    if (error)
        count = 0;
    if (count && (count < maxiocount))
        maxiocount = count;

    error = do_ioctl(p1, p2, DKIOCGETMAXBYTECOUNTWRITE, (caddr_t) &count);
    if (error)
        count = 0;
    if (count && (count < maxiocount))
        maxiocount = count;

    error = do_ioctl(p1, p2, DKIOCGETMAXSEGMENTBYTECOUNTREAD, (caddr_t) &count);
    if (!error)
	error = do_ioctl(p1, p2, DKIOCGETMAXSEGMENTCOUNTREAD, (caddr_t) &segcount);
    if (error)
        count = segcount = 0;
    count *= segcount;
    if (count && (count < maxiocount))
        maxiocount = count;

    error = do_ioctl(p1, p2, DKIOCGETMAXSEGMENTBYTECOUNTWRITE, (caddr_t) &count);
    if (!error)
	error = do_ioctl(p1, p2, DKIOCGETMAXSEGMENTCOUNTWRITE, (caddr_t) &segcount);
    if (error)
        count = segcount = 0;
    count *= segcount;
    if (count && (count < maxiocount))
        maxiocount = count;

    kprintf("max io 0x%qx bytes\n", maxiocount);
    if (maxiocount_result)
        *maxiocount_result = maxiocount;

    error = do_ioctl(p1, p2, DKIOCISSOLIDSTATE, (caddr_t)&isssd);
    if (!error && isssd)
        flags |= kIOHibernateOptionSSD;

    if (partition_device_result)
        *partition_device_result = device;
    if (image_device_result)
        *image_device_result = target;
    if (oflags)
        *oflags = flags;

out:
    kprintf("kern_open_file_for_direct_io(%d)\n", error);

    if (error && locked)
    {
        p1 = &device;
        (void) do_ioctl(p1, p2, DKIOCUNLOCKPHYSICALEXTENTS, NULL);
    }

    if (error && ref)
    {
    if (ref->pinned)
    {
        _dk_cs_pin_t pin;
        bzero(&pin, sizeof(pin));

	    pin.cp_flags = _DKIOCCSPINDISCARDBLACKLIST;
        p1 = &device;
        (void) do_ioctl(p1, p2, _DKIOCCSUNPINEXTENT, (caddr_t)&pin);
    }
	if (ref->vp)
	{
	    vnode_close(ref->vp, FWRITE, ref->ctx);
	    ref->vp = NULLVP;
	}
	vfs_context_rele(ref->ctx);
	kfree(ref, sizeof(struct kern_direct_file_io_ref_t));
	ref = NULL;
    }

    return(ref);
}

int
kern_write_file(struct kern_direct_file_io_ref_t * ref, off_t offset, caddr_t addr, vm_size_t len, int ioflag)
{
    return (vn_rdwr(UIO_WRITE, ref->vp,
			addr, len, offset,
			UIO_SYSSPACE, ioflag|IO_SYNC|IO_NODELOCKED|IO_UNIT, 
                        vfs_context_ucred(ref->ctx), (int *) 0,
			vfs_context_proc(ref->ctx)));
}


void
kern_close_file_for_direct_io(struct kern_direct_file_io_ref_t * ref,
			      off_t write_offset, caddr_t addr, vm_size_t write_length,
			      off_t discard_offset, off_t discard_end)
{
    int error;
    _dk_cs_pin_t pin;
    kprintf("kern_close_file_for_direct_io\n");

    if (!ref) return;

    if (ref->vp)
    {
        int (*do_ioctl)(void * p1, void * p2, u_long theIoctl, caddr_t result);
        void * p1;
        void * p2;

        if (ref->vp->v_type == VREG)
        {
            p1 = &ref->device;
            p2 = kernproc;
            do_ioctl = &file_ioctl;
        }
        else
        {
            /* Partition. */
            p1 = ref->vp;
            p2 = ref->ctx;
            do_ioctl = &device_ioctl;
        }
        (void) do_ioctl(p1, p2, DKIOCUNLOCKPHYSICALEXTENTS, NULL);

        if (ref->pinned)
        {
            bzero(&pin, sizeof(pin));
            pin.cp_flags = _DKIOCCSPINDISCARDBLACKLIST;
            (void) do_ioctl(p1, p2, _DKIOCCSUNPINEXTENT, (caddr_t)&pin);
        }

        
        if (discard_offset && discard_end && !ref->pinned)
        {
            (void) kern_ioctl_file_extents(ref, DKIOCUNMAP, discard_offset, discard_end);
        }
        if (addr && write_length)
        {
            (void) kern_write_file(ref, write_offset, addr, write_length, 0);
        }

        error = vnode_close(ref->vp, FWRITE, ref->ctx);

        ref->vp = NULLVP;
        kprintf("vnode_close(%d)\n", error);
    }
    vfs_context_rele(ref->ctx);
    ref->ctx = NULL;
    kfree(ref, sizeof(struct kern_direct_file_io_ref_t));
}
Commit	Line	Data
1c79356b	1	/*
2d21ac55	2	* Copyright (c) 2000-2006 Apple Computer, Inc. All rights reserved.
5d5c5d0d	3	*
2d21ac55	4	* @APPLE_OSREFERENCE_LICENSE_HEADER_START@
1c79356b	5	*
2d21ac55 A	6	* This file contains Original Code and/or Modifications of Original Code
	7	* as defined in and that are subject to the Apple Public Source License
	8	* Version 2.0 (the 'License'). You may not use this file except in
	9	* compliance with the License. The rights granted to you under the License
	10	* may not be used to create, or enable the creation or redistribution of,
	11	* unlawful or unlicensed copies of an Apple operating system, or to
	12	* circumvent, violate, or enable the circumvention or violation of, any
	13	* terms of an Apple operating system software license agreement.
8f6c56a5	14	*
2d21ac55 A	15	* Please obtain a copy of the License at
	16	* http://www.opensource.apple.com/apsl/ and read it before using this file.
	17	*
	18	* The Original Code and all software distributed under the License are
	19	* distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
8f6c56a5 A	20	* EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
8f6c56a5 A	21	* INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
2d21ac55 A	22	* FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
	23	* Please see the License for the specific language governing rights and
	24	* limitations under the License.
8f6c56a5	25	*
2d21ac55	26	* @APPLE_OSREFERENCE_LICENSE_HEADER_END@
1c79356b A	27	*/
	28	/* Copyright (c) 1998 Apple Computer, Inc. All rights reserved.
	29	*
	30	* File: bsd/kern/kern_symfile.c
	31	*
1c79356b	32	* HISTORY
1c79356b A	33	*/
	34
	35	#include <mach/vm_param.h>
	36
	37	#include <sys/param.h>
	38	#include <sys/systm.h>
	39	#include <sys/signalvar.h>
	40	#include <sys/resourcevar.h>
	41	#include <sys/namei.h>
91447636 A	42	#include <sys/vnode_internal.h>
	43	#include <sys/proc_internal.h>
	44	#include <sys/kauth.h>
1c79356b A	45	#include <sys/timeb.h>
1c79356b A	46	#include <sys/times.h>
1c79356b	47	#include <sys/acct.h>
91447636	48	#include <sys/file_internal.h>
1c79356b A	49	#include <sys/uio.h>
	50	#include <sys/kernel.h>
	51	#include <sys/stat.h>
91447636 A	52	#include <sys/disk.h>
91447636 A	53	#include <sys/conf.h>
1c79356b A	54
	55	#include <mach-o/loader.h>
	56	#include <mach-o/nlist.h>
	57
91447636	58	#include <kern/kalloc.h>
1c79356b	59	#include <vm/vm_kern.h>
91447636	60	#include <pexpert/pexpert.h>
3a60a9f5	61	#include <IOKit/IOHibernatePrivate.h>
1c79356b	62
2d21ac55 A	63	/* This function is called from kern_sysctl in the current process context;
	64	* it is exported with the System6.0.exports, but this appears to be a legacy
	65	* export, as there are no internal consumers.
1c79356b	66	*/
2d21ac55	67	int
7ddcb079 A	68	get_kernel_symfile(__unused proc_t p, __unused char const **symfile);
7ddcb079 A	69	int
2d21ac55	70	get_kernel_symfile(__unused proc_t p, __unused char const **symfile)
1c79356b	71	{
2d21ac55	72	return KERN_FAILURE;
0b4e3aa0	73	}
91447636	74
3a60a9f5 A	75	struct kern_direct_file_io_ref_t
3a60a9f5 A	76	{
6d2010ae A	77	vfs_context_t ctx;
	78	struct vnode * vp;
	79	dev_t device;
7ddcb079	80	uint32_t blksize;
316670eb A	81	off_t filelength;
316670eb A	82	char pinned;
3a60a9f5 A	83	};
	84
	85
6d2010ae	86	static int file_ioctl(void * p1, void * p2, u_long theIoctl, caddr_t result)
3a60a9f5	87	{
b0d623f7	88	dev_t device = (dev_t) p1;
3a60a9f5 A	89
	90	return ((*bdevsw[major(device)].d_ioctl)
	91	(device, theIoctl, result, S_IFBLK, p2));
	92	}
	93
6d2010ae	94	static int device_ioctl(void * p1, __unused void * p2, u_long theIoctl, caddr_t result)
3a60a9f5 A	95	{
	96	return (VNOP_IOCTL(p1, theIoctl, result, 0, p2));
	97	}
	98
316670eb A	99	static int
	100	kern_ioctl_file_extents(struct kern_direct_file_io_ref_t * ref, u_long theIoctl, off_t offset, off_t end)
	101	{
	102	int error;
	103	int (do_ioctl)(void p1, void * p2, u_long theIoctl, caddr_t result);
	104	void * p1;
	105	void * p2;
	106	uint64_t fileblk;
	107	size_t filechunk;
	108	dk_extent_t extent;
	109	dk_unmap_t unmap;
	110	_dk_cs_pin_t pin;
	111
	112	bzero(&extent, sizeof(dk_extent_t));
	113	bzero(&unmap, sizeof(dk_unmap_t));
	114	bzero(&pin, sizeof(pin));
	115	if (ref->vp->v_type == VREG)
	116	{
	117	p1 = &ref->device;
	118	p2 = kernproc;
	119	do_ioctl = &file_ioctl;
	120	}
	121	else
	122	{
	123	/* Partition. */
	124	p1 = ref->vp;
	125	p2 = ref->ctx;
	126	do_ioctl = &device_ioctl;
	127	}
	128	while (offset < end)
	129	{
	130	if (ref->vp->v_type == VREG)
	131	{
	132	daddr64_t blkno;
	133	filechunk = 110241024*1024;
	134	if (filechunk > (size_t)(end - offset))
	135	filechunk = (size_t)(end - offset);
fe8ab488 A	136	error = VNOP_BLOCKMAP(ref->vp, offset, filechunk, &blkno,
fe8ab488 A	137	&filechunk, NULL, VNODE_WRITE, NULL);
316670eb A	138	if (error) break;
	139	fileblk = blkno * ref->blksize;
	140	}
	141	else if ((ref->vp->v_type == VBLK) \|\| (ref->vp->v_type == VCHR))
	142	{
	143	fileblk = offset;
	144	filechunk = ref->filelength;
	145	}
	146
	147	if (DKIOCUNMAP == theIoctl)
	148	{
	149	extent.offset = fileblk;
	150	extent.length = filechunk;
	151	unmap.extents = &extent;
	152	unmap.extentsCount = 1;
	153	error = do_ioctl(p1, p2, theIoctl, (caddr_t)&unmap);
	154	// printf("DKIOCUNMAP(%d) 0x%qx, 0x%qx\n", error, extent.offset, extent.length);
	155	}
	156	else if (_DKIOCCSPINEXTENT == theIoctl)
	157	{
	158	pin.cp_extent.offset = fileblk;
	159	pin.cp_extent.length = filechunk;
fe8ab488	160	pin.cp_flags = _DKIOCCSPINFORHIBERNATION;
316670eb A	161	error = do_ioctl(p1, p2, theIoctl, (caddr_t)&pin);
	162	if (error && (ENOTTY != error))
	163	{
	164	printf("_DKIOCCSPINEXTENT(%d) 0x%qx, 0x%qx\n",
	165	error, pin.cp_extent.offset, pin.cp_extent.length);
	166	}
	167	}
	168	else error = EINVAL;
	169
	170	if (error) break;
	171	offset += filechunk;
	172	}
	173	return (error);
	174	}
	175
7ddcb079	176
3a60a9f5 A	177	struct kern_direct_file_io_ref_t *
3a60a9f5 A	178	kern_open_file_for_direct_io(const char * name,
fe8ab488	179	boolean_t create_file,
3a60a9f5 A	180	kern_get_file_extents_callback_t callback,
3a60a9f5 A	181	void * callback_ref,
db609669 A	182	off_t set_file_size,
	183	off_t write_file_offset,
	184	caddr_t write_file_addr,
	185	vm_size_t write_file_len,
6d2010ae A	186	dev_t * partition_device_result,
6d2010ae A	187	dev_t * image_device_result,
3a60a9f5	188	uint64_t * partitionbase_result,
0b4c1975	189	uint64_t * maxiocount_result,
db609669	190	uint32_t * oflags)
3a60a9f5 A	191	{
	192	struct kern_direct_file_io_ref_t * ref;
	193
2d21ac55	194	proc_t p;
3a60a9f5 A	195	struct vnode_attr va;
	196	int error;
	197	off_t f_offset;
6d2010ae A	198	uint64_t fileblk;
	199	size_t filechunk;
	200	uint64_t physoffset;
3a60a9f5	201	dev_t device;
6d2010ae A	202	dev_t target = 0;
	203	int isssd = 0;
	204	uint32_t flags = 0;
	205	uint32_t blksize;
bd504ef0	206	off_t maxiocount, count, segcount;
6d2010ae	207	boolean_t locked = FALSE;
3a60a9f5	208
6d2010ae A	209	int (do_ioctl)(void p1, void * p2, u_long theIoctl, caddr_t result);
	210	void * p1 = NULL;
	211	void * p2 = NULL;
3a60a9f5 A	212
	213	error = EFAULT;
	214
	215	ref = (struct kern_direct_file_io_ref_t *) kalloc(sizeof(struct kern_direct_file_io_ref_t));
	216	if (!ref)
	217	{
	218	error = EFAULT;
	219	goto out;
	220	}
	221
316670eb	222	bzero(ref, sizeof(*ref));
6d2010ae	223	p = kernproc;
2d21ac55	224	ref->ctx = vfs_context_create(vfs_context_current());
3a60a9f5	225
fe8ab488 A	226	if ((error = vnode_open(name, (create_file) ? (O_CREAT \| FWRITE) : FWRITE,
fe8ab488 A	227	(0), 0, &ref->vp, ref->ctx)))
3a60a9f5 A	228	goto out;
3a60a9f5 A	229
39236c6e A	230	if (ref->vp->v_type == VREG)
	231	{
	232	vnode_lock_spin(ref->vp);
	233	SET(ref->vp->v_flag, VSWAP);
	234	vnode_unlock(ref->vp);
	235	}
	236
db609669	237	if (write_file_addr && write_file_len)
6d2010ae	238	{
fe8ab488	239	if ((error = kern_write_file(ref, write_file_offset, write_file_addr, write_file_len, 0)))
6d2010ae A	240	goto out;
	241	}
	242
3a60a9f5 A	243	VATTR_INIT(&va);
	244	VATTR_WANTED(&va, va_rdev);
	245	VATTR_WANTED(&va, va_fsid);
	246	VATTR_WANTED(&va, va_data_size);
39236c6e	247	VATTR_WANTED(&va, va_data_alloc);
3a60a9f5 A	248	VATTR_WANTED(&va, va_nlink);
3a60a9f5 A	249	error = EFAULT;
2d21ac55	250	if (vnode_getattr(ref->vp, &va, ref->ctx))
3a60a9f5 A	251	goto out;
	252
	253	kprintf("vp va_rdev major %d minor %d\n", major(va.va_rdev), minor(va.va_rdev));
	254	kprintf("vp va_fsid major %d minor %d\n", major(va.va_fsid), minor(va.va_fsid));
39236c6e	255	kprintf("vp size %qd alloc %qd\n", va.va_data_size, va.va_data_alloc);
3a60a9f5 A	256
	257	if (ref->vp->v_type == VREG)
	258	{
fe8ab488 A	259	/* Don't dump files with links. */
	260	if (va.va_nlink != 1)
	261	goto out;
3a60a9f5 A	262
3a60a9f5 A	263	device = va.va_fsid;
db609669 A	264	ref->filelength = va.va_data_size;
db609669 A	265
b0d623f7	266	p1 = &device;
3a60a9f5 A	267	p2 = p;
3a60a9f5 A	268	do_ioctl = &file_ioctl;
db609669	269
fe8ab488 A	270	if (set_file_size)
	271	{
	272	error = vnode_setsize(ref->vp, set_file_size,
	273	IO_NOZEROFILL \| IO_NOAUTH, ref->ctx);
	274	if (error)
	275	goto out;
	276	ref->filelength = set_file_size;
	277	}
3a60a9f5 A	278	}
	279	else if ((ref->vp->v_type == VBLK) \|\| (ref->vp->v_type == VCHR))
	280	{
	281	/* Partition. */
	282	device = va.va_rdev;
	283
	284	p1 = ref->vp;
2d21ac55	285	p2 = ref->ctx;
3a60a9f5 A	286	do_ioctl = &device_ioctl;
	287	}
	288	else
	289	{
	290	/* Don't dump to non-regular files. */
	291	error = EFAULT;
	292	goto out;
	293	}
6d2010ae A	294	ref->device = device;
6d2010ae A	295
6d2010ae A	296	// get block size
6d2010ae A	297
7ddcb079	298	error = do_ioctl(p1, p2, DKIOCGETBLOCKSIZE, (caddr_t) &ref->blksize);
6d2010ae A	299	if (error)
	300	goto out;
	301
db609669	302	if (ref->vp->v_type != VREG)
6d2010ae A	303	{
	304	error = do_ioctl(p1, p2, DKIOCGETBLOCKCOUNT, (caddr_t) &fileblk);
	305	if (error)
	306	goto out;
7ddcb079	307	ref->filelength = fileblk * ref->blksize;
6d2010ae A	308	}
6d2010ae A	309
316670eb A	310	// pin logical extents
	311
	312	error = kern_ioctl_file_extents(ref, _DKIOCCSPINEXTENT, 0, ref->filelength);
	313	if (error && (ENOTTY != error)) goto out;
	314	ref->pinned = (error == 0);
	315
	316	// generate the block list
	317
	318	error = do_ioctl(p1, p2, DKIOCLOCKPHYSICALEXTENTS, NULL);
	319	if (error)
	320	goto out;
	321	locked = TRUE;
	322
6d2010ae	323	f_offset = 0;
7ddcb079	324	while (f_offset < ref->filelength)
6d2010ae A	325	{
	326	if (ref->vp->v_type == VREG)
	327	{
	328	filechunk = 110241024*1024;
	329	daddr64_t blkno;
	330
fe8ab488 A	331	error = VNOP_BLOCKMAP(ref->vp, f_offset, filechunk, &blkno,
fe8ab488 A	332	&filechunk, NULL, VNODE_WRITE, NULL);
6d2010ae A	333	if (error)
	334	goto out;
	335
7ddcb079	336	fileblk = blkno * ref->blksize;
6d2010ae A	337	}
	338	else if ((ref->vp->v_type == VBLK) \|\| (ref->vp->v_type == VCHR))
	339	{
	340	fileblk = f_offset;
7ddcb079	341	filechunk = f_offset ? 0 : ref->filelength;
6d2010ae A	342	}
	343
	344	physoffset = 0;
	345	while (physoffset < filechunk)
	346	{
	347	dk_physical_extent_t getphysreq;
	348	bzero(&getphysreq, sizeof(getphysreq));
	349
	350	getphysreq.offset = fileblk + physoffset;
	351	getphysreq.length = (filechunk - physoffset);
	352	error = do_ioctl(p1, p2, DKIOCGETPHYSICALEXTENT, (caddr_t) &getphysreq);
	353	if (error)
	354	goto out;
	355	if (!target)
	356	{
	357	target = getphysreq.dev;
	358	}
	359	else if (target != getphysreq.dev)
	360	{
	361	error = ENOTSUP;
	362	goto out;
	363	}
39236c6e A	364	#if HIBFRAGMENT
	365	uint64_t rev;
	366	for (rev = 4096; rev <= getphysreq.length; rev += 4096)
	367	{
	368	callback(callback_ref, getphysreq.offset + getphysreq.length - rev, 4096);
	369	}
	370	#else
6d2010ae	371	callback(callback_ref, getphysreq.offset, getphysreq.length);
39236c6e	372	#endif
6d2010ae A	373	physoffset += getphysreq.length;
	374	}
	375	f_offset += filechunk;
	376	}
	377	callback(callback_ref, 0ULL, 0ULL);
	378
	379	if (ref->vp->v_type == VREG)
	380	p1 = &target;
3a60a9f5 A	381
	382	// get partition base
	383
fe8ab488 A	384	if (partitionbase_result)
	385	{
	386	error = do_ioctl(p1, p2, DKIOCGETBASE, (caddr_t) partitionbase_result);
	387	if (error)
	388	goto out;
	389	}
3a60a9f5 A	390
	391	// get block size & constraints
	392
	393	error = do_ioctl(p1, p2, DKIOCGETBLOCKSIZE, (caddr_t) &blksize);
	394	if (error)
	395	goto out;
	396
	397	maxiocount = 110241024*1024;
	398
	399	error = do_ioctl(p1, p2, DKIOCGETMAXBLOCKCOUNTREAD, (caddr_t) &count);
	400	if (error)
	401	count = 0;
	402	count *= blksize;
	403	if (count && (count < maxiocount))
	404	maxiocount = count;
	405
	406	error = do_ioctl(p1, p2, DKIOCGETMAXBLOCKCOUNTWRITE, (caddr_t) &count);
	407	if (error)
	408	count = 0;
	409	count *= blksize;
	410	if (count && (count < maxiocount))
	411	maxiocount = count;
	412
	413	error = do_ioctl(p1, p2, DKIOCGETMAXBYTECOUNTREAD, (caddr_t) &count);
	414	if (error)
	415	count = 0;
	416	if (count && (count < maxiocount))
	417	maxiocount = count;
	418
	419	error = do_ioctl(p1, p2, DKIOCGETMAXBYTECOUNTWRITE, (caddr_t) &count);
	420	if (error)
	421	count = 0;
	422	if (count && (count < maxiocount))
	423	maxiocount = count;
	424
	425	error = do_ioctl(p1, p2, DKIOCGETMAXSEGMENTBYTECOUNTREAD, (caddr_t) &count);
bd504ef0 A	426	if (!error)
bd504ef0 A	427	error = do_ioctl(p1, p2, DKIOCGETMAXSEGMENTCOUNTREAD, (caddr_t) &segcount);
3a60a9f5	428	if (error)
bd504ef0 A	429	count = segcount = 0;
bd504ef0 A	430	count *= segcount;
3a60a9f5 A	431	if (count && (count < maxiocount))
	432	maxiocount = count;
	433
	434	error = do_ioctl(p1, p2, DKIOCGETMAXSEGMENTBYTECOUNTWRITE, (caddr_t) &count);
bd504ef0 A	435	if (!error)
bd504ef0 A	436	error = do_ioctl(p1, p2, DKIOCGETMAXSEGMENTCOUNTWRITE, (caddr_t) &segcount);
3a60a9f5	437	if (error)
bd504ef0 A	438	count = segcount = 0;
bd504ef0 A	439	count *= segcount;
3a60a9f5 A	440	if (count && (count < maxiocount))
	441	maxiocount = count;
	442
	443	kprintf("max io 0x%qx bytes\n", maxiocount);
	444	if (maxiocount_result)
	445	*maxiocount_result = maxiocount;
	446
6d2010ae A	447	error = do_ioctl(p1, p2, DKIOCISSOLIDSTATE, (caddr_t)&isssd);
	448	if (!error && isssd)
	449	flags \|= kIOHibernateOptionSSD;
3a60a9f5	450
6d2010ae A	451	if (partition_device_result)
	452	*partition_device_result = device;
	453	if (image_device_result)
	454	*image_device_result = target;
fe8ab488	455	if (oflags)
6d2010ae	456	*oflags = flags;
3a60a9f5 A	457
	458	out:
	459	kprintf("kern_open_file_for_direct_io(%d)\n", error);
	460
6d2010ae A	461	if (error && locked)
	462	{
	463	p1 = &device;
	464	(void) do_ioctl(p1, p2, DKIOCUNLOCKPHYSICALEXTENTS, NULL);
	465	}
	466
	467	if (error && ref)
	468	{
fe8ab488 A	469	if (ref->pinned)
	470	{
	471	_dk_cs_pin_t pin;
	472	bzero(&pin, sizeof(pin));
	473
	474	pin.cp_flags = _DKIOCCSPINDISCARDBLACKLIST;
	475	p1 = &device;
	476	(void) do_ioctl(p1, p2, _DKIOCCSUNPINEXTENT, (caddr_t)&pin);
	477	}
6d2010ae A	478	if (ref->vp)
6d2010ae A	479	{
2d21ac55 A	480	vnode_close(ref->vp, FWRITE, ref->ctx);
	481	ref->vp = NULLVP;
	482	}
2d21ac55 A	483	vfs_context_rele(ref->ctx);
	484	kfree(ref, sizeof(struct kern_direct_file_io_ref_t));
	485	ref = NULL;
3a60a9f5	486	}
db609669	487
3a60a9f5 A	488	return(ref);
	489	}
	490
	491	int
fe8ab488	492	kern_write_file(struct kern_direct_file_io_ref_t * ref, off_t offset, caddr_t addr, vm_size_t len, int ioflag)
3a60a9f5 A	493	{
	494	return (vn_rdwr(UIO_WRITE, ref->vp,
	495	addr, len, offset,
fe8ab488	496	UIO_SYSSPACE, ioflag\|IO_SYNC\|IO_NODELOCKED\|IO_UNIT,
2d21ac55 A	497	vfs_context_ucred(ref->ctx), (int *) 0,
2d21ac55 A	498	vfs_context_proc(ref->ctx)));
3a60a9f5 A	499	}
3a60a9f5 A	500
7ddcb079	501
3a60a9f5	502	void
6d2010ae	503	kern_close_file_for_direct_io(struct kern_direct_file_io_ref_t * ref,
7ddcb079 A	504	off_t write_offset, caddr_t addr, vm_size_t write_length,
7ddcb079 A	505	off_t discard_offset, off_t discard_end)
3a60a9f5	506	{
6d2010ae	507	int error;
fe8ab488	508	_dk_cs_pin_t pin;
3a60a9f5 A	509	kprintf("kern_close_file_for_direct_io\n");
3a60a9f5 A	510
6d2010ae	511	if (!ref) return;
3a60a9f5	512
6d2010ae A	513	if (ref->vp)
	514	{
	515	int (do_ioctl)(void p1, void * p2, u_long theIoctl, caddr_t result);
	516	void * p1;
	517	void * p2;
	518
	519	if (ref->vp->v_type == VREG)
	520	{
	521	p1 = &ref->device;
	522	p2 = kernproc;
	523	do_ioctl = &file_ioctl;
	524	}
	525	else
	526	{
	527	/* Partition. */
	528	p1 = ref->vp;
	529	p2 = ref->ctx;
	530	do_ioctl = &device_ioctl;
	531	}
	532	(void) do_ioctl(p1, p2, DKIOCUNLOCKPHYSICALEXTENTS, NULL);
fe8ab488 A	533
	534	if (ref->pinned)
	535	{
	536	bzero(&pin, sizeof(pin));
	537	pin.cp_flags = _DKIOCCSPINDISCARDBLACKLIST;
	538	(void) do_ioctl(p1, p2, _DKIOCCSUNPINEXTENT, (caddr_t)&pin);
	539	}
	540
6d2010ae	541
316670eb	542	if (discard_offset && discard_end && !ref->pinned)
6d2010ae	543	{
316670eb	544	(void) kern_ioctl_file_extents(ref, DKIOCUNMAP, discard_offset, discard_end);
6d2010ae	545	}
39236c6e A	546	if (addr && write_length)
39236c6e A	547	{
fe8ab488	548	(void) kern_write_file(ref, write_offset, addr, write_length, 0);
39236c6e	549	}
6d2010ae A	550
	551	error = vnode_close(ref->vp, FWRITE, ref->ctx);
	552
	553	ref->vp = NULLVP;
	554	kprintf("vnode_close(%d)\n", error);
3a60a9f5	555	}
6d2010ae A	556	vfs_context_rele(ref->ctx);
	557	ref->ctx = NULL;
	558	kfree(ref, sizeof(struct kern_direct_file_io_ref_t));
3a60a9f5	559	}