xref: /src/sys/miscfs/genfs/genfs_io.c

/*	$NetBSD: genfs_io.c,v 1.53.2.19 2014/05/22 11:41:05 yamt Exp $	*/

/*
 * Copyright (c) 1982, 1986, 1989, 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. 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.
 *
 */

#include <sys/cdefs.h>
__KERNEL_RCSID(0, "$NetBSD: genfs_io.c,v 1.53.2.19 2014/05/22 11:41:05 yamt Exp $");

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/proc.h>
#include <sys/kernel.h>
#include <sys/mount.h>
#include <sys/vnode.h>
#include <sys/kmem.h>
#include <sys/kauth.h>
#include <sys/fstrans.h>
#include <sys/buf.h>
#include <sys/radixtree.h>

#include <miscfs/genfs/genfs.h>
#include <miscfs/genfs/genfs_node.h>
#include <miscfs/specfs/specdev.h>
#include <miscfs/syncfs/syncfs.h>

#include <uvm/uvm.h>
#include <uvm/uvm_pager.h>
#include <uvm/uvm_page_array.h>

static int genfs_do_directio(struct vmspace *, vaddr_t, size_t, struct vnode *,
    off_t, enum uio_rw);
static void genfs_dio_iodone(struct buf *);

static int genfs_do_io(struct vnode *, off_t, vaddr_t, size_t, int, enum uio_rw,
    void (*)(struct buf *));
static void genfs_rel_pages(struct vm_page **, unsigned int);
static void genfs_markdirty(struct vnode *);

int genfs_maxdio = MAXPHYS;

static void
genfs_rel_pages(struct vm_page **pgs, unsigned int npages)
{
	unsigned int i;

	for (i = 0; i < npages; i++) {
		struct vm_page *pg = pgs[i];

		if (pg == NULL || pg == PGO_DONTCARE)
			continue;
		KASSERT(uvm_page_locked_p(pg));
		if (pg->flags & PG_FAKE) {
			pg->flags |= PG_RELEASED;
		}
	}
	mutex_enter(&uvm_pageqlock);
	uvm_page_unbusy(pgs, npages);
	mutex_exit(&uvm_pageqlock);
}

static void
genfs_markdirty(struct vnode *vp)
{

	KASSERT(mutex_owned(vp->v_interlock));
	if ((vp->v_iflag & VI_ONWORKLST) == 0) {
		vn_syncer_add_to_worklist(vp, filedelay);
	}
	if ((vp->v_iflag & (VI_WRMAP|VI_WRMAPDIRTY)) == VI_WRMAP) {
		vp->v_iflag |= VI_WRMAPDIRTY;
	}
}

/*
 * generic VM getpages routine.
 * Return PG_BUSY pages for the given range,
 * reading from backing store if necessary.
 */

int
genfs_getpages(void *v)
{
	struct vop_getpages_args /* {
		struct vnode *a_vp;
		voff_t a_offset;
		struct vm_page **a_m;
		int *a_count;
		int a_centeridx;
		vm_prot_t a_access_type;
		int a_advice;
		int a_flags;
	} */ * const ap = v;

	off_t diskeof, memeof;
	int i, error, npages;
	const int flags = ap->a_flags;
	struct vnode * const vp = ap->a_vp;
	struct uvm_object * const uobj = &vp->v_uobj;
	kauth_cred_t const cred = curlwp->l_cred;		/* XXXUBC curlwp */
	const bool async = (flags & PGO_SYNCIO) == 0;
	const bool memwrite = (ap->a_access_type & VM_PROT_WRITE) != 0;
	const bool overwrite = (flags & PGO_OVERWRITE) != 0;
	const bool blockalloc = memwrite && (flags & PGO_NOBLOCKALLOC) == 0;
	const bool glocked = (flags & PGO_GLOCKHELD) != 0;
	const bool need_wapbl = blockalloc && vp->v_mount->mnt_wapbl;
	bool has_trans_wapbl = false;
	UVMHIST_FUNC("genfs_getpages"); UVMHIST_CALLED(ubchist);

	UVMHIST_LOG(ubchist, "vp %p off 0x%x/%x count %d",
	    vp, ap->a_offset >> 32, ap->a_offset, *ap->a_count);

	KASSERT(vp->v_type == VREG || vp->v_type == VDIR ||
	    vp->v_type == VLNK || vp->v_type == VBLK);

startover:
	error = 0;
	const voff_t origvsize = vp->v_size;
	const off_t origoffset = ap->a_offset;
	const int orignpages = *ap->a_count;

	GOP_SIZE(vp, origvsize, &diskeof, 0);
	if (flags & PGO_PASTEOF) {
		off_t newsize;
#if defined(DIAGNOSTIC)
		off_t writeeof;
#endif /* defined(DIAGNOSTIC) */

		newsize = MAX(origvsize,
		    origoffset + (orignpages << PAGE_SHIFT));
		GOP_SIZE(vp, newsize, &memeof, GOP_SIZE_MEM);
#if defined(DIAGNOSTIC)
		GOP_SIZE(vp, vp->v_writesize, &writeeof, GOP_SIZE_MEM);
		if (newsize > round_page(writeeof)) {
			panic("%s: past eof: %" PRId64 " vs. %" PRId64,
			    __func__, newsize, round_page(writeeof));
		}
#endif /* defined(DIAGNOSTIC) */
	} else {
		GOP_SIZE(vp, origvsize, &memeof, GOP_SIZE_MEM);
	}
	KASSERT(ap->a_centeridx >= 0 || ap->a_centeridx <= orignpages);
	KASSERT((origoffset & (PAGE_SIZE - 1)) == 0 && origoffset >= 0);
	KASSERT(orignpages > 0);

	/*
	 * Bounds-check the request.
	 */

	if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= memeof) {
		if ((flags & PGO_LOCKED) == 0) {
			mutex_exit(uobj->vmobjlock);
		}
		UVMHIST_LOG(ubchist, "off 0x%x count %d goes past EOF 0x%x",
		    origoffset, *ap->a_count, memeof,0);
		error = EINVAL;
		goto out_err;
	}

	/* uobj is locked */

	if ((flags & PGO_NOTIMESTAMP) == 0 &&
	    (vp->v_type != VBLK ||
	    (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) {
		int updflags = 0;

		if ((vp->v_mount->mnt_flag & MNT_NOATIME) == 0) {
			updflags = GOP_UPDATE_ACCESSED;
		}
		if (memwrite) {
			updflags |= GOP_UPDATE_MODIFIED;
		}
		if (updflags != 0) {
			GOP_MARKUPDATE(vp, updflags);
		}
	}

	/*
	 * For PGO_LOCKED requests, just return whatever's in memory.
	 */

	if (flags & PGO_LOCKED) {
		int nfound;
		struct vm_page *pg;

		KASSERT(!glocked);
		npages = *ap->a_count;
#if defined(DEBUG)
		for (i = 0; i < npages; i++) {
			pg = ap->a_m[i];
			KASSERT(pg == NULL || pg == PGO_DONTCARE);
		}
#endif /* defined(DEBUG) */
		nfound = uvn_findpages(uobj, origoffset, &npages,
		    ap->a_m, NULL,
		    UFP_NOWAIT|UFP_NOALLOC|(memwrite ? UFP_NORDONLY : 0));
		KASSERT(npages == *ap->a_count);
		if (nfound == 0) {
			error = EBUSY;
			goto out_err;
		}
		/*
		 * lock and unlock g_glock to ensure that no one is truncating
		 * the file behind us.
		 */
		if (!genfs_node_rdtrylock(vp)) {
			genfs_rel_pages(ap->a_m, npages);

			/*
			 * restore the array.
			 */

			for (i = 0; i < npages; i++) {
				pg = ap->a_m[i];

				if (pg != NULL && pg != PGO_DONTCARE) {
					ap->a_m[i] = NULL;
				}
				KASSERT(ap->a_m[i] == NULL ||
				    ap->a_m[i] == PGO_DONTCARE);
			}
		} else {
			genfs_node_unlock(vp);
		}
		error = (ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0);
		if (error == 0 && memwrite) {
			for (i = 0; i < npages; i++) {
				pg = ap->a_m[i];
				if (pg == NULL || pg == PGO_DONTCARE) {
					continue;
				}
				if (uvm_pagegetdirty(pg) ==
				    UVM_PAGE_STATUS_CLEAN) {
					uvm_pagemarkdirty(pg,
					    UVM_PAGE_STATUS_UNKNOWN);
				}
			}
			genfs_markdirty(vp);
		}
		goto out_err;
	}
	mutex_exit(uobj->vmobjlock);

	/*
	 * find the requested pages and make some simple checks.
	 * leave space in the page array for a whole block.
	 */

	const int fs_bshift = (vp->v_type != VBLK) ?
	    vp->v_mount->mnt_fs_bshift : DEV_BSHIFT;
	const int dev_bshift = (vp->v_type != VBLK) ?
	    vp->v_mount->mnt_dev_bshift : DEV_BSHIFT;
	const int fs_bsize = 1 << fs_bshift;
#define	blk_mask	(fs_bsize - 1)
#define	trunc_blk(x)	((x) & ~blk_mask)
#define	round_blk(x)	(((x) + blk_mask) & ~blk_mask)

	const int orignmempages = MIN(orignpages,
	    round_page(memeof - origoffset) >> PAGE_SHIFT);
	npages = orignmempages;
	const off_t startoffset = trunc_blk(origoffset);
	const off_t endoffset = MIN(
	    round_page(round_blk(origoffset + (npages << PAGE_SHIFT))),
	    round_page(memeof));
	const int ridx = (origoffset - startoffset) >> PAGE_SHIFT;

	const int pgs_size = sizeof(struct vm_page *) *
	    ((endoffset - startoffset) >> PAGE_SHIFT);
	struct vm_page **pgs, *pgs_onstack[UBC_MAX_PAGES];

	if (pgs_size > sizeof(pgs_onstack)) {
		pgs = kmem_zalloc(pgs_size, async ? KM_NOSLEEP : KM_SLEEP);
		if (pgs == NULL) {
			pgs = pgs_onstack;
			error = ENOMEM;
			goto out_err;
		}
	} else {
		pgs = pgs_onstack;
		(void)memset(pgs, 0, pgs_size);
	}

	UVMHIST_LOG(ubchist, "ridx %d npages %d startoff %ld endoff %ld",
	    ridx, npages, startoffset, endoffset);

	if (!has_trans_wapbl) {
		fstrans_start(vp->v_mount, FSTRANS_SHARED);
		/*
		 * XXX: This assumes that we come here only via
		 * the mmio path
		 */
		if (need_wapbl) {
			error = WAPBL_BEGIN(vp->v_mount);
			if (error) {
				fstrans_done(vp->v_mount);
				goto out_err_free;
			}
		}
		has_trans_wapbl = true;
	}

	/*
	 * hold g_glock to prevent a race with truncate.
	 *
	 * check if our idea of v_size is still valid.
	 */

	KASSERT(!glocked || genfs_node_wrlocked(vp));
	if (!glocked) {
		if (blockalloc) {
			genfs_node_wrlock(vp);
		} else {
			genfs_node_rdlock(vp);
		}
	}
	mutex_enter(uobj->vmobjlock);
	if (vp->v_size < origvsize) {
		if (!glocked) {
			genfs_node_unlock(vp);
		}
		if (pgs != pgs_onstack)
			kmem_free(pgs, pgs_size);
		goto startover;
	}

	if (uvn_findpages(uobj, origoffset, &npages, &pgs[ridx], NULL,
	    async ? UFP_NOWAIT : UFP_ALL) != orignmempages) {
		if (!glocked) {
			genfs_node_unlock(vp);
		}
		KASSERT(async != 0);
		genfs_rel_pages(&pgs[ridx], orignmempages);
		mutex_exit(uobj->vmobjlock);
		error = EBUSY;
		goto out_err_free;
	}

	/*
	 * if PGO_OVERWRITE is set, don't bother reading the pages.
	 */

	if (overwrite) {
		if (!glocked) {
			genfs_node_unlock(vp);
		}
		UVMHIST_LOG(ubchist, "PGO_OVERWRITE",0,0,0,0);

		for (i = 0; i < npages; i++) {
			struct vm_page *pg = pgs[ridx + i];

			/*
			 * it's caller's responsibility to allocate blocks
			 * beforehand for the overwrite case.
			 */
			pg->flags &= ~PG_RDONLY;
			/*
			 * mark the page DIRTY.
			 * otherwise another thread can do putpages and pull
			 * our vnode from syncer's queue before our caller does
			 * ubc_release.  note that putpages won't see CLEAN
			 * pages even if they are BUSY.
			 */
			uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
		}
		npages += ridx;
		goto out;
	}

	/*
	 * if the pages are already resident, just return them.
	 */

	for (i = 0; i < npages; i++) {
		struct vm_page *pg = pgs[ridx + i];

		if ((pg->flags & PG_FAKE) ||
		    (memwrite && (pg->flags & PG_RDONLY) != 0)) {
			break;
		}
	}
	if (i == npages) {
		if (!glocked) {
			genfs_node_unlock(vp);
		}
		UVMHIST_LOG(ubchist, "returning cached pages", 0,0,0,0);
		npages += ridx;
		goto out;
	}

	/*
	 * the page wasn't resident and we're not overwriting,
	 * so we're going to have to do some i/o.
	 * find any additional pages needed to cover the expanded range.
	 */

	npages = (endoffset - startoffset) >> PAGE_SHIFT;
	if (startoffset != origoffset || npages != orignmempages) {
		int npgs;

		/*
		 * we need to avoid deadlocks caused by locking
		 * additional pages at lower offsets than pages we
		 * already have locked.  unlock them all and start over.
		 */

		genfs_rel_pages(&pgs[ridx], orignmempages);
		memset(pgs, 0, pgs_size);

		UVMHIST_LOG(ubchist, "reset npages start 0x%x end 0x%x",
		    startoffset, endoffset, 0,0);
		npgs = npages;
		if (uvn_findpages(uobj, startoffset, &npgs, pgs, NULL,
		    async ? UFP_NOWAIT : UFP_ALL) != npages) {
			if (!glocked) {
				genfs_node_unlock(vp);
			}
			KASSERT(async != 0);
			genfs_rel_pages(pgs, npages);
			mutex_exit(uobj->vmobjlock);
			error = EBUSY;
			goto out_err_free;
		}
	}

	mutex_exit(uobj->vmobjlock);

    {
	size_t bytes, iobytes, tailstart, tailbytes, totalbytes, skipbytes;
	vaddr_t kva;
	struct buf *bp, *mbp;
	bool sawhole = false;

	/*
	 * read the desired page(s).
	 */

	totalbytes = npages << PAGE_SHIFT;
	bytes = MIN(totalbytes, MAX(diskeof - startoffset, 0));
	tailbytes = totalbytes - bytes;
	skipbytes = 0;

	kva = uvm_pagermapin(pgs, npages,
	    UVMPAGER_MAPIN_READ | (async ? 0 : UVMPAGER_MAPIN_WAITOK));
	if (kva == 0) {
		error = EBUSY;
		goto mapin_fail;
	}

	mbp = getiobuf(vp, true);
	mbp->b_bufsize = totalbytes;
	mbp->b_data = (void *)kva;
	mbp->b_resid = mbp->b_bcount = bytes;
	mbp->b_cflags = BC_BUSY;
	if (async) {
		mbp->b_flags = B_READ | B_ASYNC;
		mbp->b_iodone = uvm_aio_biodone;
	} else {
		mbp->b_flags = B_READ;
		mbp->b_iodone = NULL;
	}
	if (async)
		BIO_SETPRIO(mbp, BPRIO_TIMELIMITED);
	else
		BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL);

	/*
	 * if EOF is in the middle of the range, zero the part past EOF.
	 * skip over pages which are not PG_FAKE since in that case they have
	 * valid data that we need to preserve.
	 */

	tailstart = bytes;
	while (tailbytes > 0) {
		const int len = PAGE_SIZE - (tailstart & PAGE_MASK);

		KASSERT(len <= tailbytes);
		if ((pgs[tailstart >> PAGE_SHIFT]->flags & PG_FAKE) != 0) {
			memset((void *)(kva + tailstart), 0, len);
			UVMHIST_LOG(ubchist, "tailbytes %p 0x%x 0x%x",
			    kva, tailstart, len, 0);
		}
		tailstart += len;
		tailbytes -= len;
	}

	/*
	 * now loop over the pages, reading as needed.
	 */

	bp = NULL;
	off_t offset;
	for (offset = startoffset;
	    bytes > 0;
	    offset += iobytes, bytes -= iobytes) {
		int run;
		daddr_t lbn, blkno;
		int pidx;
		struct vnode *devvp;

		/*
		 * skip pages which don't need to be read.
		 */

		pidx = (offset - startoffset) >> PAGE_SHIFT;
		while ((pgs[pidx]->flags & PG_FAKE) == 0) {
			size_t b;

			KASSERT((offset & (PAGE_SIZE - 1)) == 0);
			if ((pgs[pidx]->flags & PG_RDONLY)) {
				sawhole = true;
			}
			b = MIN(PAGE_SIZE, bytes);
			offset += b;
			bytes -= b;
			skipbytes += b;
			pidx++;
			UVMHIST_LOG(ubchist, "skipping, new offset 0x%x",
			    offset, 0,0,0);
			if (bytes == 0) {
				goto loopdone;
			}
		}

		/*
		 * bmap the file to find out the blkno to read from and
		 * how much we can read in one i/o.  if bmap returns an error,
		 * skip the rest of the top-level i/o.
		 */

		lbn = offset >> fs_bshift;
		error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run);
		if (error) {
			UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%x -> %d\n",
			    lbn,error,0,0);
			skipbytes += bytes;
			bytes = 0;
			goto loopdone;
		}

		/*
		 * see how many pages can be read with this i/o.
		 * reduce the i/o size if necessary to avoid
		 * overwriting pages with valid data.
		 */

		iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset,
		    bytes);
		if (offset + iobytes > round_page(offset)) {
			int pcount;

			pcount = 1;
			while (pidx + pcount < npages &&
			    pgs[pidx + pcount]->flags & PG_FAKE) {
				pcount++;
			}
			iobytes = MIN(iobytes, (pcount << PAGE_SHIFT) -
			    (offset - trunc_page(offset)));
		}

		/*
		 * if this block isn't allocated, zero it instead of
		 * reading it.  unless we are going to allocate blocks,
		 * mark the pages we zeroed PG_RDONLY.
		 */

		if (blkno == (daddr_t)-1) {
			int holepages = (round_page(offset + iobytes) -
			    trunc_page(offset)) >> PAGE_SHIFT;
			UVMHIST_LOG(ubchist, "lbn 0x%x -> HOLE", lbn,0,0,0);

			sawhole = true;
			memset((char *)kva + (offset - startoffset), 0,
			    iobytes);
			skipbytes += iobytes;

			if (!blockalloc) {
				mutex_enter(uobj->vmobjlock);
				for (i = 0; i < holepages; i++) {
					pgs[pidx + i]->flags |= PG_RDONLY;
				}
				mutex_exit(uobj->vmobjlock);
			}
			continue;
		}

		/*
		 * allocate a sub-buf for this piece of the i/o
		 * (or just use mbp if there's only 1 piece),
		 * and start it going.
		 */

		if (offset == startoffset && iobytes == bytes) {
			bp = mbp;
		} else {
			UVMHIST_LOG(ubchist, "vp %p bp %p num now %d",
			    vp, bp, vp->v_numoutput, 0);
			bp = getiobuf(vp, true);
			nestiobuf_setup(mbp, bp, offset - startoffset, iobytes);
		}
		bp->b_lblkno = 0;

		/* adjust physical blkno for partial blocks */
		bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >>
		    dev_bshift);

		UVMHIST_LOG(ubchist,
		    "bp %p offset 0x%x bcount 0x%x blkno 0x%x",
		    bp, offset, bp->b_bcount, bp->b_blkno);

		VOP_STRATEGY(devvp, bp);
	}

loopdone:
	nestiobuf_done(mbp, skipbytes, error);
	if (async) {
		UVMHIST_LOG(ubchist, "returning 0 (async)",0,0,0,0);
		if (!glocked) {
			genfs_node_unlock(vp);
		}
		error = 0;
		goto out_err_free;
	}
	if (bp != NULL) {
		error = biowait(mbp);
	}

	/* Remove the mapping (make KVA available as soon as possible) */
	uvm_pagermapout(kva, npages);

	/*
	 * if this we encountered a hole then we have to do a little more work.
	 * if blockalloc is false, we marked the page PG_RDONLY so that future
	 * write accesses to the page will fault again.
	 * if blockalloc is true, we must make sure that the backing store for
	 * the page is completely allocated while the pages are locked.
	 */

	if (!error && sawhole && blockalloc) {
		error = GOP_ALLOC(vp, startoffset,
		    npages << PAGE_SHIFT, 0, cred);
		UVMHIST_LOG(ubchist, "gop_alloc off 0x%x/0x%x -> %d",
		    startoffset, npages << PAGE_SHIFT, error,0);
		if (!error) {
			mutex_enter(uobj->vmobjlock);
			for (i = 0; i < npages; i++) {
				struct vm_page *pg = pgs[i];

				if (pg == NULL) {
					continue;
				}
				pg->flags &= ~PG_RDONLY;
				uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
				UVMHIST_LOG(ubchist, "mark dirty pg %p",
				    pg,0,0,0);
			}
			mutex_exit(uobj->vmobjlock);
		}
	}

	putiobuf(mbp);
    }

mapin_fail:
	if (!glocked) {
		genfs_node_unlock(vp);
	}
	mutex_enter(uobj->vmobjlock);

	/*
	 * we're almost done!  release the pages...
	 * for errors, we free the pages.
	 * otherwise we activate them and mark them as valid and clean.
	 * also, unbusy pages that were not actually requested.
	 */

	if (error) {
		genfs_rel_pages(pgs, npages);
		mutex_exit(uobj->vmobjlock);
		UVMHIST_LOG(ubchist, "returning error %d", error,0,0,0);
		goto out_err_free;
	}

out:
	UVMHIST_LOG(ubchist, "succeeding, npages %d", npages,0,0,0);
	error = 0;
	mutex_enter(&uvm_pageqlock);
	for (i = 0; i < npages; i++) {
		struct vm_page *pg = pgs[i];
		if (pg == NULL) {
			continue;
		}
		UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x",
		    pg, pg->flags, 0,0);
		if (pg->flags & PG_FAKE && !overwrite) {
			/*
			 * we've read page's contents from the backing storage.
			 *
			 * for a read fault, we keep them CLEAN.
			 */
			KASSERT(uvm_pagegetdirty(pg) == UVM_PAGE_STATUS_CLEAN);
			pg->flags &= ~PG_FAKE;
		}
		KASSERT(!blockalloc || (pg->flags & PG_RDONLY) == 0);
		if (i < ridx || i >= ridx + orignmempages || async) {
			UVMHIST_LOG(ubchist, "unbusy pg %p offset 0x%x",
			    pg, pg->offset,0,0);
			KASSERT(!overwrite);
			if (pg->flags & PG_WANTED) {
				wakeup(pg);
			}
			if (pg->flags & PG_FAKE && overwrite) {
				uvm_pagezero(pg);
			}
			if (pg->flags & PG_RELEASED) {
				uvm_pagefree(pg);
				continue;
			}
			uvm_pageenqueue(pg);
			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
			UVM_PAGE_OWN(pg, NULL);
		} else if (memwrite && !overwrite &&
		    uvm_pagegetdirty(pg) == UVM_PAGE_STATUS_CLEAN) {
			/*
			 * for a write fault, start dirtiness tracking of
			 * requested pages.
			 */
			uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_UNKNOWN);
		}
	}
	mutex_exit(&uvm_pageqlock);
	if (memwrite) {
		genfs_markdirty(vp);
	}
	mutex_exit(uobj->vmobjlock);
	if (ap->a_m != NULL) {
		memcpy(ap->a_m, &pgs[ridx],
		    orignmempages * sizeof(struct vm_page *));
	}

out_err_free:
	if (pgs != NULL && pgs != pgs_onstack)
		kmem_free(pgs, pgs_size);
out_err:
	if (has_trans_wapbl) {
		if (need_wapbl)
			WAPBL_END(vp->v_mount);
		fstrans_done(vp->v_mount);
	}
	return error;
}

/*
 * generic VM putpages routine.
 * Write the given range of pages to backing store.
 *
 * => "offhi == 0" means flush all pages at or after "offlo".
 * => object should be locked by caller.  we return with the
 *      object unlocked.
 * => if PGO_CLEANIT or PGO_SYNCIO is set, we may block (due to I/O).
 *	thus, a caller might want to unlock higher level resources
 *	(e.g. vm_map) before calling flush.
 * => if neither PGO_CLEANIT nor PGO_SYNCIO is set, we will not block
 * => if PGO_ALLPAGES is set, then all pages in the object will be processed.
 * => NOTE: we are allowed to lock the page queues, so the caller
 *	must not be holding the page queue lock.
 *
 * note on "cleaning" object and PG_BUSY pages:
 *	this routine is holding the lock on the object.   the only time
 *	that it can run into a PG_BUSY page that it does not own is if
 *	some other process has started I/O on the page (e.g. either
 *	a pagein, or a pageout).  if the PG_BUSY page is being paged
 *	in, then it can not be dirty (!UVM_PAGE_STATUS_CLEAN) because no
 *	one has	had a chance to modify it yet.  if the PG_BUSY page is
 *	being paged out then it means that someone else has already started
 *	cleaning the page for us (how nice!).  in this case, if we
 *	have syncio specified, then after we make our pass through the
 *	object we need to wait for the other PG_BUSY pages to clear
 *	off (i.e. we need to do an iosync).   also note that once a
 *	page is PG_BUSY it must stay in its object until it is un-busyed.
 */

int
genfs_putpages(void *v)
{
	struct vop_putpages_args /* {
		struct vnode *a_vp;
		voff_t a_offlo;
		voff_t a_offhi;
		int a_flags;
	} */ * const ap = v;

	return genfs_do_putpages(ap->a_vp, ap->a_offlo, ap->a_offhi,
	    ap->a_flags, NULL);
}

int
genfs_do_putpages(struct vnode *vp, off_t startoff, off_t endoff,
    int origflags, struct vm_page **busypg)
{
	struct uvm_object * const uobj = &vp->v_uobj;
	kmutex_t * const slock = uobj->vmobjlock;
	off_t nextoff;
	/* Even for strange MAXPHYS, the shift rounds down to a page */
#define maxpages (MAXPHYS >> PAGE_SHIFT)
	unsigned int i;
	unsigned int npages, nback;
	unsigned int freeflag;
	int error;
	struct vm_page *pgs[maxpages], *pg;
	struct uvm_page_array a;
	bool wasclean, needs_clean;
	bool pagedaemon = curlwp == uvm.pagedaemon_lwp;
	struct lwp * const l = curlwp ? curlwp : &lwp0;
	int flags;
	bool written;		/* if we write out any pages */
	bool need_wapbl;
	bool has_trans;
	bool tryclean;		/* try to pull off from the syncer's list */
	bool onworklst;
	const bool integrity_sync =
	    (origflags & (PGO_LAZY|PGO_SYNCIO|PGO_CLEANIT)) ==
	    (PGO_SYNCIO|PGO_CLEANIT);
	const bool dirtyonly = (origflags & (PGO_DEACTIVATE|PGO_FREE)) == 0;

	UVMHIST_FUNC("genfs_putpages"); UVMHIST_CALLED(ubchist);

	KASSERT(origflags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE));
	KASSERT((startoff & PAGE_MASK) == 0 && (endoff & PAGE_MASK) == 0);
	KASSERT(startoff < endoff || endoff == 0);

	UVMHIST_LOG(ubchist, "vp %p pages %d off 0x%x len 0x%x",
	    vp, uobj->uo_npages, startoff, endoff - startoff);

	has_trans = false;
	need_wapbl = (!pagedaemon && vp->v_mount && vp->v_mount->mnt_wapbl &&
	    (origflags & PGO_JOURNALLOCKED) == 0);

retry:
	flags = origflags;
	KASSERT((vp->v_iflag & VI_ONWORKLST) != 0 ||
	    (vp->v_iflag & VI_WRMAPDIRTY) == 0);

	/*
	 * shortcut if we have no pages to process.
	 */

	if (uobj->uo_npages == 0 || (dirtyonly &&
	    radix_tree_empty_tagged_tree_p(&uobj->uo_pages,
	    UVM_PAGE_DIRTY_TAG))) {
		if (vp->v_iflag & VI_ONWORKLST) {
			vp->v_iflag &= ~VI_WRMAPDIRTY;
			if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL)
				vn_syncer_remove_from_worklist(vp);
		}
		if (has_trans) {
			if (need_wapbl)
				WAPBL_END(vp->v_mount);
			fstrans_done(vp->v_mount);
		}
		mutex_exit(slock);
		return (0);
	}

	/*
	 * the vnode has pages, set up to process the request.
	 */

	if (!has_trans && (flags & PGO_CLEANIT) != 0) {
		mutex_exit(slock);
		if (pagedaemon) {
			error = fstrans_start_nowait(vp->v_mount, FSTRANS_LAZY);
			if (error)
				return error;
		} else
			fstrans_start(vp->v_mount, FSTRANS_LAZY);
		if (need_wapbl) {
			error = WAPBL_BEGIN(vp->v_mount);
			if (error) {
				fstrans_done(vp->v_mount);
				return error;
			}
		}
		has_trans = true;
		mutex_enter(slock);
		goto retry;
	}

	error = 0;
	wasclean = (vp->v_numoutput == 0);

	/*
	 * if this vnode is known not to have dirty pages,
	 * don't bother to clean it out.
	 */

	if ((vp->v_iflag & VI_ONWORKLST) == 0) {
#if !defined(DEBUG)
		if (dirtyonly) {
			goto skip_scan;
		}
#endif /* !defined(DEBUG) */
		flags &= ~PGO_CLEANIT;
	}

	/*
	 * start the loop to scan pages.
	 */

	written = false;
	nextoff = startoff;
	if (endoff == 0 || flags & PGO_ALLPAGES) {
		endoff = trunc_page(LLONG_MAX);
	}
	freeflag = pagedaemon ? PG_PAGEOUT : PG_RELEASED;
	tryclean = true;
	uvm_page_array_init(&a);
	for (;;) {
		bool protected;

		/*
		 * if !dirtyonly, iterate over all resident pages in the range.
		 *
		 * if dirtyonly, only possibly dirty pages are interested.
		 * however, if we are asked to sync for integrity, we should
		 * wait on pages being written back by another threads as well.
		 */

		pg = uvm_page_array_fill_and_peek(&a, uobj, nextoff, 0,
		    dirtyonly ? (UVM_PAGE_ARRAY_FILL_DIRTY |
		    (integrity_sync ? UVM_PAGE_ARRAY_FILL_WRITEBACK : 0)) : 0);
		if (pg == NULL) {
			break;
		}

		KASSERT(pg->uobject == uobj);
		KASSERT((pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 ||
		    (pg->flags & (PG_BUSY)) != 0);
		KASSERT(pg->offset >= startoff);
		KASSERT(pg->offset >= nextoff);
		KASSERT(!dirtyonly ||
		    uvm_pagegetdirty(pg) != UVM_PAGE_STATUS_CLEAN ||
		    radix_tree_get_tag(&uobj->uo_pages,
			pg->offset >> PAGE_SHIFT, UVM_PAGE_WRITEBACK_TAG));
		if (pg->offset >= endoff) {
			break;
		}

		/*
		 * a preempt point.
		 */

		if ((l->l_cpu->ci_schedstate.spc_flags & SPCF_SHOULDYIELD)
		    != 0) {
			nextoff = pg->offset; /* visit this page again */
			mutex_exit(slock);
			preempt();
			/*
			 * as we dropped the object lock, our cached pages can
			 * be stale.
			 */
			uvm_page_array_clear(&a);
			mutex_enter(slock);
			continue;
		}

		/*
		 * if the current page is busy, wait for it to become unbusy.
		 */

		if ((pg->flags & PG_BUSY) != 0) {
			UVMHIST_LOG(ubchist, "busy %p", pg,0,0,0);
			if ((pg->flags & (PG_RELEASED|PG_PAGEOUT)) != 0
			    && (flags & PGO_BUSYFAIL) != 0) {
				UVMHIST_LOG(ubchist, "busyfail %p", pg,
				    0,0,0);
				error = EDEADLK;
				if (busypg != NULL)
					*busypg = pg;
				break;
			}
			if (pagedaemon) {
				/*
				 * someone has taken the page while we
				 * dropped the lock for fstrans_start.
				 */
				break;
			}
			/*
			 * don't bother to wait on other's activities
			 * unless we are asked to sync for integrity.
			 */
			if (!integrity_sync && (flags & PGO_RECLAIM) == 0) {
				wasclean = false;
				nextoff = pg->offset + PAGE_SIZE;
				uvm_page_array_advance(&a);
				continue;
			}
			nextoff = pg->offset; /* visit this page again */
			pg->flags |= PG_WANTED;
			UVM_UNLOCK_AND_WAIT(pg, slock, 0, "genput", 0);
			/*
			 * as we dropped the object lock, our cached pages can
			 * be stale.
			 */
			uvm_page_array_clear(&a);
			mutex_enter(slock);
			continue;
		}

		nextoff = pg->offset + PAGE_SIZE;
		uvm_page_array_advance(&a);

		/*
		 * if we're freeing, remove all mappings of the page now.
		 * if we're cleaning, check if the page needs to be cleaned.
		 */

		protected = false;
		if (flags & PGO_FREE) {
			pmap_page_protect(pg, VM_PROT_NONE);
			protected = true;
		} else if (flags & PGO_CLEANIT) {

			/*
			 * if we still have some hope to pull this vnode off
			 * from the syncer queue, write-protect the page.
			 */

			if (tryclean && wasclean) {

				/*
				 * uobj pages get wired only by uvm_fault
				 * where uobj is locked.
				 */

				if (pg->wire_count == 0) {
					pmap_page_protect(pg,
					    VM_PROT_READ|VM_PROT_EXECUTE);
					protected = true;
				} else {
					/*
					 * give up.
					 */
					tryclean = false;
				}
			}
		}

		if (flags & PGO_CLEANIT) {
			needs_clean = uvm_pagecheckdirty(pg, protected);
		} else {
			needs_clean = false;
		}

		/*
		 * if we're cleaning, build a cluster.
		 * the cluster will consist of pages which are currently dirty.
		 * if not cleaning, just operate on the one page.
		 */

		if (needs_clean) {
			unsigned int nforw;
			unsigned int fpflags;

			KDASSERT((vp->v_iflag & VI_ONWORKLST));
			wasclean = false;
			memset(pgs, 0, sizeof(pgs));
			pg->flags |= PG_BUSY;
			UVM_PAGE_OWN(pg, "genfs_putpages");

			fpflags = UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY;

			/*
			 * XXX PG_PAGER1 incompatibility check.
			 *
			 * this is a kludge for nfs.  nfs has two kind of dirty
			 * pages:
			 *	- not written to the server yet
			 *	- written to the server but not committed yet
			 * the latter is marked as PG_NEEDCOMMIT. (== PG_PAGER1)
			 * nfs doesn't want them being clustered together.
			 *
			 * probably it's better to make PG_NEEDCOMMIT a first
			 * level citizen for uvm/genfs.
			 */
			if ((pg->flags & PG_PAGER1) != 0) {
				fpflags |= UFP_ONLYPAGER1;
			} else {
				fpflags |= UFP_NOPAGER1;
			}

			/*
			 * first look backward.
			 *
			 * because we always scan pages in the ascending order,
			 * backward scan can be useful only for the first page
			 * in the range.
			 */
			if (startoff == pg->offset) {
				npages = MIN(maxpages >> 1,
				    pg->offset >> PAGE_SHIFT);
				nback = npages;
				uvn_findpages(uobj, pg->offset - PAGE_SIZE,
				    &nback, &pgs[0], NULL,
				    fpflags | UFP_BACKWARD);
				if (nback) {
					memmove(&pgs[0], &pgs[npages - nback],
					    nback * sizeof(pgs[0]));
					if (npages - nback < nback)
						memset(&pgs[nback], 0,
						    (npages - nback) *
						    sizeof(pgs[0]));
					else
						memset(&pgs[npages - nback], 0,
						    nback * sizeof(pgs[0]));
				}
			} else {
				nback = 0;
			}

			/*
			 * then plug in our page of interest.
			 */

			pgs[nback] = pg;

			/*
			 * then look forward to fill in the remaining space in
			 * the array of pages.
			 *
			 * pass our cached array of pages so that hopefully
			 * uvn_findpages can find some good pages in it.
			 * the array a was filled above with the one of
			 * following sets of flags:
			 *	0
			 *	UVM_PAGE_ARRAY_FILL_DIRTY
			 *	UVM_PAGE_ARRAY_FILL_DIRTY|WRITEBACK
			 */

			nforw = maxpages - nback - 1;
			uvn_findpages(uobj, pg->offset + PAGE_SIZE,
			    &nforw, &pgs[nback + 1], &a, fpflags);
			npages = nback + 1 + nforw;
		} else {
			pgs[0] = pg;
			npages = 1;
			nback = 0;
		}

		/*
		 * apply FREE or DEACTIVATE options if requested.
		 */

		if (flags & (PGO_DEACTIVATE|PGO_FREE)) {
			mutex_enter(&uvm_pageqlock);
		}
		for (i = 0; i < npages; i++) {
			struct vm_page *tpg = pgs[i];

			KASSERT(tpg->uobject == uobj);
			KASSERT(i == 0 ||
			    pgs[i-1]->offset + PAGE_SIZE == tpg->offset);
			KASSERT(!needs_clean || uvm_pagegetdirty(pgs[i]) !=
			    UVM_PAGE_STATUS_DIRTY);
			if (needs_clean) {
				/*
				 * mark pages as WRITEBACK so that concurrent
				 * fsync can find and wait for our activities.
				 */
				radix_tree_set_tag(&uobj->uo_pages,
				    pgs[i]->offset >> PAGE_SHIFT,
				    UVM_PAGE_WRITEBACK_TAG);
			}
			if (tpg->offset < startoff || tpg->offset >= endoff)
				continue;
			if (flags & PGO_DEACTIVATE && tpg->wire_count == 0) {
				uvm_pagedeactivate(tpg);
			} else if (flags & PGO_FREE) {
				pmap_page_protect(tpg, VM_PROT_NONE);
				if (tpg->flags & PG_BUSY) {
					tpg->flags |= freeflag;
					if (pagedaemon) {
						uvm_pageout_start(1);
						uvm_pagedequeue(tpg);
					}
				} else {

					/*
					 * ``page is not busy''
					 * implies that npages is 1
					 * and needs_clean is false.
					 */

					KASSERT(npages == 1);
					KASSERT(!needs_clean);
					KASSERT(pg == tpg);
					KASSERT(nextoff ==
					    tpg->offset + PAGE_SIZE);
					uvm_pagefree(tpg);
					if (pagedaemon)
						uvmexp.pdfreed++;
				}
			}
		}
		if (flags & (PGO_DEACTIVATE|PGO_FREE)) {
			mutex_exit(&uvm_pageqlock);
		}
		if (needs_clean) {
			mutex_exit(slock);
			KASSERT(nextoff == pg->offset + PAGE_SIZE);
			KASSERT(nback < npages);
			nextoff = pg->offset + ((npages - nback) << PAGE_SHIFT);
			KASSERT(pgs[nback] == pg);
			KASSERT(nextoff == pgs[npages - 1]->offset + PAGE_SIZE);

			/*
			 * start the i/o.
			 */
			error = GOP_WRITE(vp, pgs, npages, flags);
			written = true;
			/*
			 * as we dropped the object lock, our cached pages can
			 * be stale.
			 */
			uvm_page_array_clear(&a);
			mutex_enter(slock);
			if (error) {
				break;
			}
		}
	}
	uvm_page_array_fini(&a);

	/*
	 * update ctime/mtime if the modification we started writing out might
	 * be from mmap'ed write.
	 *
	 * this is necessary when an application keeps a file mmaped and
	 * repeatedly modifies it via the window.  note that, because we
	 * don't always write-protect pages when cleaning, such modifications
	 * might not involve any page faults.
	 */

	if (written && (vp->v_iflag & VI_WRMAPDIRTY) != 0 &&
	    (vp->v_type != VBLK ||
	    (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) {
		GOP_MARKUPDATE(vp, GOP_UPDATE_MODIFIED);
	}

	/*
	 * if we no longer have any possibly dirty pages, take us off the
	 * syncer list.
	 */

	if ((vp->v_iflag & VI_ONWORKLST) != 0 &&
	    radix_tree_empty_tagged_tree_p(&uobj->uo_pages,
	    UVM_PAGE_DIRTY_TAG)) {
		vp->v_iflag &= ~VI_WRMAPDIRTY;
		if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL)
			vn_syncer_remove_from_worklist(vp);
	}

#if !defined(DEBUG)
skip_scan:
#endif /* !defined(DEBUG) */

	/*
	 * if we found or started any i/o and we're asked to sync for integrity,
	 * wait for all writes to finish.
	 */

	if (!wasclean && integrity_sync) {
		while (vp->v_numoutput != 0)
			cv_wait(&vp->v_cv, slock);
	}
	onworklst = (vp->v_iflag & VI_ONWORKLST) != 0;
	mutex_exit(slock);

	if ((flags & PGO_RECLAIM) != 0 && onworklst) {
		/*
		 * in the case of PGO_RECLAIM, ensure to make the vnode clean.
		 * retrying is not a big deal because, in many cases,
		 * uobj->uo_npages is already 0 here.
		 */
		mutex_enter(slock);
		goto retry;
	}

	if (has_trans) {
		if (need_wapbl)
			WAPBL_END(vp->v_mount);
		fstrans_done(vp->v_mount);
	}

	return (error);
}

int
genfs_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, int flags)
{
	off_t off;
	vaddr_t kva;
	size_t len;
	int error;
	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);

	UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x",
	    vp, pgs, npages, flags);

	off = pgs[0]->offset;
	kva = uvm_pagermapin(pgs, npages,
	    UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK);
	len = npages << PAGE_SHIFT;

	error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE,
			    uvm_aio_biodone);

	return error;
}

/*
 * genfs_gop_write_rwmap:
 *
 * a variant of genfs_gop_write.  it's used by UDF for its directory buffers.
 * this maps pages with PROT_WRITE so that VOP_STRATEGY can modifies
 * the contents before writing it out to the underlying storage.
 */

int
genfs_gop_write_rwmap(struct vnode *vp, struct vm_page **pgs, int npages,
    int flags)
{
	off_t off;
	vaddr_t kva;
	size_t len;
	int error;
	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);

	UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x",
	    vp, pgs, npages, flags);

	off = pgs[0]->offset;
	kva = uvm_pagermapin(pgs, npages,
	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
	len = npages << PAGE_SHIFT;

	error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE,
			    uvm_aio_biodone);

	return error;
}

/*
 * Backend routine for doing I/O to vnode pages.  Pages are already locked
 * and mapped into kernel memory.  Here we just look up the underlying
 * device block addresses and call the strategy routine.
 */

static int
genfs_do_io(struct vnode *vp, off_t off, vaddr_t kva, size_t len, int flags,
    enum uio_rw rw, void (*iodone)(struct buf *))
{
	int s, error;
	int fs_bshift, dev_bshift;
	off_t eof, offset, startoffset;
	size_t bytes, iobytes, skipbytes;
	struct buf *mbp, *bp;
	const bool async = (flags & PGO_SYNCIO) == 0;
	const bool lazy = (flags & PGO_LAZY) == 0;
	const bool iowrite = rw == UIO_WRITE;
	const int brw = iowrite ? B_WRITE : B_READ;
	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);

	UVMHIST_LOG(ubchist, "vp %p kva %p len 0x%x flags 0x%x",
	    vp, kva, len, flags);

	KASSERT(vp->v_size <= vp->v_writesize);
	GOP_SIZE(vp, vp->v_writesize, &eof, 0);
	if (vp->v_type != VBLK) {
		fs_bshift = vp->v_mount->mnt_fs_bshift;
		dev_bshift = vp->v_mount->mnt_dev_bshift;
	} else {
		fs_bshift = DEV_BSHIFT;
		dev_bshift = DEV_BSHIFT;
	}
	error = 0;
	startoffset = off;
	bytes = MIN(len, eof - startoffset);
	skipbytes = 0;
	KASSERT(bytes != 0);

	if (iowrite) {
		/*
		 * why += 2?
		 * 1 for biodone, 1 for uvm_aio_aiodone.
		 */
		mutex_enter(vp->v_interlock);
		vp->v_numoutput += 2;
		mutex_exit(vp->v_interlock);
	}
	mbp = getiobuf(vp, true);
	UVMHIST_LOG(ubchist, "vp %p mbp %p num now %d bytes 0x%x",
	    vp, mbp, vp->v_numoutput, bytes);
	mbp->b_bufsize = len;
	mbp->b_data = (void *)kva;
	mbp->b_resid = mbp->b_bcount = bytes;
	mbp->b_cflags = BC_BUSY | BC_AGE;
	if (async) {
		mbp->b_flags = brw | B_ASYNC;
		mbp->b_iodone = iodone;
	} else {
		mbp->b_flags = brw;
		mbp->b_iodone = NULL;
	}
	if (curlwp == uvm.pagedaemon_lwp)
		BIO_SETPRIO(mbp, BPRIO_TIMELIMITED);
	else if (async || lazy)
		BIO_SETPRIO(mbp, BPRIO_TIMENONCRITICAL);
	else
		BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL);

	bp = NULL;
	for (offset = startoffset;
	    bytes > 0;
	    offset += iobytes, bytes -= iobytes) {
		int run;
		daddr_t lbn, blkno;
		struct vnode *devvp;

		/*
		 * bmap the file to find out the blkno to read from and
		 * how much we can read in one i/o.  if bmap returns an error,
		 * skip the rest of the top-level i/o.
		 */

		lbn = offset >> fs_bshift;
		error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run);
		if (error) {
			UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%x -> %d\n",
			    lbn,error,0,0);
			skipbytes += bytes;
			bytes = 0;
			goto loopdone;
		}

		/*
		 * see how many pages can be read with this i/o.
		 * reduce the i/o size if necessary to avoid
		 * overwriting pages with valid data.
		 */

		iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset,
		    bytes);

		/*
		 * if this block isn't allocated, zero it instead of
		 * reading it.  unless we are going to allocate blocks,
		 * mark the pages we zeroed PG_RDONLY.
		 */

		if (blkno == (daddr_t)-1) {
			if (!iowrite) {
				memset((char *)kva + (offset - startoffset), 0,
				    iobytes);
			}
			skipbytes += iobytes;
			continue;
		}

		/*
		 * allocate a sub-buf for this piece of the i/o
		 * (or just use mbp if there's only 1 piece),
		 * and start it going.
		 */

		if (offset == startoffset && iobytes == bytes) {
			bp = mbp;
		} else {
			UVMHIST_LOG(ubchist, "vp %p bp %p num now %d",
			    vp, bp, vp->v_numoutput, 0);
			bp = getiobuf(vp, true);
			nestiobuf_setup(mbp, bp, offset - startoffset, iobytes);
		}
		bp->b_lblkno = 0;

		/* adjust physical blkno for partial blocks */
		bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >>
		    dev_bshift);

		UVMHIST_LOG(ubchist,
		    "bp %p offset 0x%x bcount 0x%x blkno 0x%x",
		    bp, offset, bp->b_bcount, bp->b_blkno);

		VOP_STRATEGY(devvp, bp);
	}

loopdone:
	if (skipbytes) {
		UVMHIST_LOG(ubchist, "skipbytes %d", skipbytes, 0,0,0);
	}
	nestiobuf_done(mbp, skipbytes, error);
	if (async) {
		UVMHIST_LOG(ubchist, "returning 0 (async)", 0,0,0,0);
		return (0);
	}
	UVMHIST_LOG(ubchist, "waiting for mbp %p", mbp,0,0,0);
	error = biowait(mbp);
	s = splbio();
	(*iodone)(mbp);
	splx(s);
	UVMHIST_LOG(ubchist, "returning, error %d", error,0,0,0);
	return (error);
}

int
genfs_compat_getpages(void *v)
{
	struct vop_getpages_args /* {
		struct vnode *a_vp;
		voff_t a_offset;
		struct vm_page **a_m;
		int *a_count;
		int a_centeridx;
		vm_prot_t a_access_type;
		int a_advice;
		int a_flags;
	} */ *ap = v;

	off_t origoffset;
	struct vnode *vp = ap->a_vp;
	struct uvm_object *uobj = &vp->v_uobj;
	struct vm_page *pg, **pgs;
	vaddr_t kva;
	int i, error, orignpages, npages;
	struct iovec iov;
	struct uio uio;
	kauth_cred_t cred = curlwp->l_cred;
	const bool memwrite = (ap->a_access_type & VM_PROT_WRITE) != 0;

	error = 0;
	origoffset = ap->a_offset;
	orignpages = *ap->a_count;
	pgs = ap->a_m;

	if (ap->a_flags & PGO_LOCKED) {
		uvn_findpages(uobj, origoffset, ap->a_count, ap->a_m, NULL,
		    UFP_NOWAIT|UFP_NOALLOC| (memwrite ? UFP_NORDONLY : 0));

		error = ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0;
		if (error == 0 && memwrite) {
			genfs_markdirty(vp);
		}
		return error;
	}
	if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= vp->v_size) {
		mutex_exit(uobj->vmobjlock);
		return EINVAL;
	}
	if ((ap->a_flags & PGO_SYNCIO) == 0) {
		mutex_exit(uobj->vmobjlock);
		return 0;
	}
	npages = orignpages;
	uvn_findpages(uobj, origoffset, &npages, pgs, NULL, UFP_ALL);
	mutex_exit(uobj->vmobjlock);
	kva = uvm_pagermapin(pgs, npages,
	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
	for (i = 0; i < npages; i++) {
		pg = pgs[i];
		if ((pg->flags & PG_FAKE) == 0) {
			continue;
		}
		iov.iov_base = (char *)kva + (i << PAGE_SHIFT);
		iov.iov_len = PAGE_SIZE;
		uio.uio_iov = &iov;
		uio.uio_iovcnt = 1;
		uio.uio_offset = origoffset + (i << PAGE_SHIFT);
		uio.uio_rw = UIO_READ;
		uio.uio_resid = PAGE_SIZE;
		UIO_SETUP_SYSSPACE(&uio);
		/* XXX vn_lock */
		error = VOP_READ(vp, &uio, 0, cred);
		if (error) {
			break;
		}
		if (uio.uio_resid) {
			memset(iov.iov_base, 0, uio.uio_resid);
		}
	}
	uvm_pagermapout(kva, npages);
	mutex_enter(uobj->vmobjlock);
	mutex_enter(&uvm_pageqlock);
	for (i = 0; i < npages; i++) {
		pg = pgs[i];
		if (error && (pg->flags & PG_FAKE) != 0) {
			pg->flags |= PG_RELEASED;
		} else {
			uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_UNKNOWN);
			uvm_pageactivate(pg);
		}
	}
	if (error) {
		uvm_page_unbusy(pgs, npages);
	}
	mutex_exit(&uvm_pageqlock);
	if (error == 0 && memwrite) {
		genfs_markdirty(vp);
	}
	mutex_exit(uobj->vmobjlock);
	return error;
}

int
genfs_compat_gop_write(struct vnode *vp, struct vm_page **pgs, int npages,
    int flags)
{
	off_t offset;
	struct iovec iov;
	struct uio uio;
	kauth_cred_t cred = curlwp->l_cred;
	struct buf *bp;
	vaddr_t kva;
	int error;

	offset = pgs[0]->offset;
	kva = uvm_pagermapin(pgs, npages,
	    UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK);

	iov.iov_base = (void *)kva;
	iov.iov_len = npages << PAGE_SHIFT;
	uio.uio_iov = &iov;
	uio.uio_iovcnt = 1;
	uio.uio_offset = offset;
	uio.uio_rw = UIO_WRITE;
	uio.uio_resid = npages << PAGE_SHIFT;
	UIO_SETUP_SYSSPACE(&uio);
	/* XXX vn_lock */
	error = VOP_WRITE(vp, &uio, 0, cred);

	mutex_enter(vp->v_interlock);
	vp->v_numoutput++;
	mutex_exit(vp->v_interlock);

	bp = getiobuf(vp, true);
	bp->b_cflags = BC_BUSY | BC_AGE;
	bp->b_lblkno = offset >> vp->v_mount->mnt_fs_bshift;
	bp->b_data = (char *)kva;
	bp->b_bcount = npages << PAGE_SHIFT;
	bp->b_bufsize = npages << PAGE_SHIFT;
	bp->b_resid = 0;
	bp->b_error = error;
	uvm_aio_aiodone(bp);
	return (error);
}

/*
 * Process a uio using direct I/O.  If we reach a part of the request
 * which cannot be processed in this fashion for some reason, just return.
 * The caller must handle some additional part of the request using
 * buffered I/O before trying direct I/O again.
 */

void
genfs_directio(struct vnode *vp, struct uio *uio, int ioflag)
{
	struct vmspace *vs;
	struct iovec *iov;
	vaddr_t va;
	size_t len;
	const int mask = DEV_BSIZE - 1;
	int error;
	bool need_wapbl = (vp->v_mount && vp->v_mount->mnt_wapbl &&
	    (ioflag & IO_JOURNALLOCKED) == 0);

	/*
	 * We only support direct I/O to user space for now.
	 */

	if (VMSPACE_IS_KERNEL_P(uio->uio_vmspace)) {
		return;
	}

	/*
	 * If the vnode is mapped, we would need to get the getpages lock
	 * to stabilize the bmap, but then we would get into trouble while
	 * locking the pages if the pages belong to this same vnode (or a
	 * multi-vnode cascade to the same effect).  Just fall back to
	 * buffered I/O if the vnode is mapped to avoid this mess.
	 */

	if (vp->v_vflag & VV_MAPPED) {
		return;
	}

	if (need_wapbl) {
		error = WAPBL_BEGIN(vp->v_mount);
		if (error)
			return;
	}

	/*
	 * Do as much of the uio as possible with direct I/O.
	 */

	vs = uio->uio_vmspace;
	while (uio->uio_resid) {
		iov = uio->uio_iov;
		if (iov->iov_len == 0) {
			uio->uio_iov++;
			uio->uio_iovcnt--;
			continue;
		}
		va = (vaddr_t)iov->iov_base;
		len = MIN(iov->iov_len, genfs_maxdio);
		len &= ~mask;

		/*
		 * If the next chunk is smaller than DEV_BSIZE or extends past
		 * the current EOF, then fall back to buffered I/O.
		 */

		if (len == 0 || uio->uio_offset + len > vp->v_size) {
			break;
		}

		/*
		 * Check alignment.  The file offset must be at least
		 * sector-aligned.  The exact constraint on memory alignment
		 * is very hardware-dependent, but requiring sector-aligned
		 * addresses there too is safe.
		 */

		if (uio->uio_offset & mask || va & mask) {
			break;
		}
		error = genfs_do_directio(vs, va, len, vp, uio->uio_offset,
					  uio->uio_rw);
		if (error) {
			break;
		}
		iov->iov_base = (char *)iov->iov_base + len;
		iov->iov_len -= len;
		uio->uio_offset += len;
		uio->uio_resid -= len;
	}

	if (need_wapbl)
		WAPBL_END(vp->v_mount);
}

/*
 * Iodone routine for direct I/O.  We don't do much here since the request is
 * always synchronous, so the caller will do most of the work after biowait().
 */

static void
genfs_dio_iodone(struct buf *bp)
{

	KASSERT((bp->b_flags & B_ASYNC) == 0);
	if ((bp->b_flags & B_READ) == 0 && (bp->b_cflags & BC_AGE) != 0) {
		mutex_enter(bp->b_objlock);
		vwakeup(bp);
		mutex_exit(bp->b_objlock);
	}
	putiobuf(bp);
}

/*
 * Process one chunk of a direct I/O request.
 */

static int
genfs_do_directio(struct vmspace *vs, vaddr_t uva, size_t len, struct vnode *vp,
    off_t off, enum uio_rw rw)
{
	struct vm_map *map;
	struct pmap *upm, *kpm __unused;
	size_t klen = round_page(uva + len) - trunc_page(uva);
	off_t spoff, epoff;
	vaddr_t kva, puva;
	paddr_t pa;
	vm_prot_t prot;
	int error, rv __diagused, poff, koff;
	const int pgoflags = PGO_CLEANIT | PGO_SYNCIO | PGO_JOURNALLOCKED |
		(rw == UIO_WRITE ? PGO_FREE : 0);

	/*
	 * For writes, verify that this range of the file already has fully
	 * allocated backing store.  If there are any holes, just punt and
	 * make the caller take the buffered write path.
	 */

	if (rw == UIO_WRITE) {
		daddr_t lbn, elbn, blkno;
		int bsize, bshift, run;

		bshift = vp->v_mount->mnt_fs_bshift;
		bsize = 1 << bshift;
		lbn = off >> bshift;
		elbn = (off + len + bsize - 1) >> bshift;
		while (lbn < elbn) {
			error = VOP_BMAP(vp, lbn, NULL, &blkno, &run);
			if (error) {
				return error;
			}
			if (blkno == (daddr_t)-1) {
				return ENOSPC;
			}
			lbn += 1 + run;
		}
	}

	/*
	 * Flush any cached pages for parts of the file that we're about to
	 * access.  If we're writing, invalidate pages as well.
	 */

	spoff = trunc_page(off);
	epoff = round_page(off + len);
	mutex_enter(vp->v_interlock);
	error = VOP_PUTPAGES(vp, spoff, epoff, pgoflags);
	if (error) {
		return error;
	}

	/*
	 * Wire the user pages and remap them into kernel memory.
	 */

	prot = rw == UIO_READ ? VM_PROT_READ | VM_PROT_WRITE : VM_PROT_READ;
	error = uvm_vslock(vs, (void *)uva, len, prot);
	if (error) {
		return error;
	}

	map = &vs->vm_map;
	upm = vm_map_pmap(map);
	kpm = vm_map_pmap(kernel_map);
	puva = trunc_page(uva);
	kva = uvm_km_alloc(kernel_map, klen, atop(puva) & uvmexp.colormask,
	    UVM_KMF_VAONLY | UVM_KMF_WAITVA | UVM_KMF_COLORMATCH);
	for (poff = 0; poff < klen; poff += PAGE_SIZE) {
		rv = pmap_extract(upm, puva + poff, &pa);
		KASSERT(rv);
		pmap_kenter_pa(kva + poff, pa, prot, PMAP_WIRED);
	}
	pmap_update(kpm);

	/*
	 * Do the I/O.
	 */

	koff = uva - trunc_page(uva);
	error = genfs_do_io(vp, off, kva + koff, len, PGO_SYNCIO, rw,
			    genfs_dio_iodone);

	/*
	 * Tear down the kernel mapping.
	 */

	pmap_kremove(kva, klen);
	pmap_update(kpm);
	uvm_km_free(kernel_map, kva, klen, UVM_KMF_VAONLY);

	/*
	 * Unwire the user pages.
	 */

	uvm_vsunlock(vs, (void *)uva, len);
	return error;
}