xref: /netbsd-src/sys/miscfs/genfs/genfs_io.c (revision b1c86f5f087524e68db12794ee9c3e3da1ab17a0)
1 /*	$NetBSD: genfs_io.c,v 1.40 2010/09/01 16:56:19 chs Exp $	*/
2 
3 /*
4  * Copyright (c) 1982, 1986, 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  */
32 
33 #include <sys/cdefs.h>
34 __KERNEL_RCSID(0, "$NetBSD: genfs_io.c,v 1.40 2010/09/01 16:56:19 chs Exp $");
35 
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/proc.h>
39 #include <sys/kernel.h>
40 #include <sys/mount.h>
41 #include <sys/namei.h>
42 #include <sys/vnode.h>
43 #include <sys/fcntl.h>
44 #include <sys/kmem.h>
45 #include <sys/poll.h>
46 #include <sys/mman.h>
47 #include <sys/file.h>
48 #include <sys/kauth.h>
49 #include <sys/fstrans.h>
50 #include <sys/buf.h>
51 
52 #include <miscfs/genfs/genfs.h>
53 #include <miscfs/genfs/genfs_node.h>
54 #include <miscfs/specfs/specdev.h>
55 
56 #include <uvm/uvm.h>
57 #include <uvm/uvm_pager.h>
58 
59 static int genfs_do_directio(struct vmspace *, vaddr_t, size_t, struct vnode *,
60     off_t, enum uio_rw);
61 static void genfs_dio_iodone(struct buf *);
62 
63 static int genfs_do_io(struct vnode *, off_t, vaddr_t, size_t, int, enum uio_rw,
64     void (*)(struct buf *));
65 static void genfs_rel_pages(struct vm_page **, int);
66 static void genfs_markdirty(struct vnode *);
67 
68 int genfs_maxdio = MAXPHYS;
69 
70 static void
71 genfs_rel_pages(struct vm_page **pgs, int npages)
72 {
73 	int i;
74 
75 	for (i = 0; i < npages; i++) {
76 		struct vm_page *pg = pgs[i];
77 
78 		if (pg == NULL || pg == PGO_DONTCARE)
79 			continue;
80 		if (pg->flags & PG_FAKE) {
81 			pg->flags |= PG_RELEASED;
82 		}
83 	}
84 	mutex_enter(&uvm_pageqlock);
85 	uvm_page_unbusy(pgs, npages);
86 	mutex_exit(&uvm_pageqlock);
87 }
88 
89 static void
90 genfs_markdirty(struct vnode *vp)
91 {
92 	struct genfs_node * const gp = VTOG(vp);
93 
94 	KASSERT(mutex_owned(&vp->v_interlock));
95 	gp->g_dirtygen++;
96 	if ((vp->v_iflag & VI_ONWORKLST) == 0) {
97 		vn_syncer_add_to_worklist(vp, filedelay);
98 	}
99 	if ((vp->v_iflag & (VI_WRMAP|VI_WRMAPDIRTY)) == VI_WRMAP) {
100 		vp->v_iflag |= VI_WRMAPDIRTY;
101 	}
102 }
103 
104 /*
105  * generic VM getpages routine.
106  * Return PG_BUSY pages for the given range,
107  * reading from backing store if necessary.
108  */
109 
110 int
111 genfs_getpages(void *v)
112 {
113 	struct vop_getpages_args /* {
114 		struct vnode *a_vp;
115 		voff_t a_offset;
116 		struct vm_page **a_m;
117 		int *a_count;
118 		int a_centeridx;
119 		vm_prot_t a_access_type;
120 		int a_advice;
121 		int a_flags;
122 	} */ * const ap = v;
123 
124 	off_t diskeof, memeof;
125 	int i, error, npages;
126 	const int flags = ap->a_flags;
127 	struct vnode * const vp = ap->a_vp;
128 	struct uvm_object * const uobj = &vp->v_uobj;
129 	kauth_cred_t const cred = curlwp->l_cred;		/* XXXUBC curlwp */
130 	const bool async = (flags & PGO_SYNCIO) == 0;
131 	const bool memwrite = (ap->a_access_type & VM_PROT_WRITE) != 0;
132 	bool has_trans = false;
133 	const bool overwrite = (flags & PGO_OVERWRITE) != 0;
134 	const bool blockalloc = memwrite && (flags & PGO_NOBLOCKALLOC) == 0;
135 	const bool glocked = (flags & PGO_GLOCKHELD) != 0;
136 	UVMHIST_FUNC("genfs_getpages"); UVMHIST_CALLED(ubchist);
137 
138 	UVMHIST_LOG(ubchist, "vp %p off 0x%x/%x count %d",
139 	    vp, ap->a_offset >> 32, ap->a_offset, *ap->a_count);
140 
141 	KASSERT(vp->v_type == VREG || vp->v_type == VDIR ||
142 	    vp->v_type == VLNK || vp->v_type == VBLK);
143 
144 startover:
145 	error = 0;
146 	const voff_t origvsize = vp->v_size;
147 	const off_t origoffset = ap->a_offset;
148 	const int orignpages = *ap->a_count;
149 
150 	GOP_SIZE(vp, origvsize, &diskeof, 0);
151 	if (flags & PGO_PASTEOF) {
152 		off_t newsize;
153 #if defined(DIAGNOSTIC)
154 		off_t writeeof;
155 #endif /* defined(DIAGNOSTIC) */
156 
157 		newsize = MAX(origvsize,
158 		    origoffset + (orignpages << PAGE_SHIFT));
159 		GOP_SIZE(vp, newsize, &memeof, GOP_SIZE_MEM);
160 #if defined(DIAGNOSTIC)
161 		GOP_SIZE(vp, vp->v_writesize, &writeeof, GOP_SIZE_MEM);
162 		if (newsize > round_page(writeeof)) {
163 			panic("%s: past eof: %" PRId64 " vs. %" PRId64,
164 			    __func__, newsize, round_page(writeeof));
165 		}
166 #endif /* defined(DIAGNOSTIC) */
167 	} else {
168 		GOP_SIZE(vp, origvsize, &memeof, GOP_SIZE_MEM);
169 	}
170 	KASSERT(ap->a_centeridx >= 0 || ap->a_centeridx <= orignpages);
171 	KASSERT((origoffset & (PAGE_SIZE - 1)) == 0 && origoffset >= 0);
172 	KASSERT(orignpages > 0);
173 
174 	/*
175 	 * Bounds-check the request.
176 	 */
177 
178 	if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= memeof) {
179 		if ((flags & PGO_LOCKED) == 0) {
180 			mutex_exit(&uobj->vmobjlock);
181 		}
182 		UVMHIST_LOG(ubchist, "off 0x%x count %d goes past EOF 0x%x",
183 		    origoffset, *ap->a_count, memeof,0);
184 		error = EINVAL;
185 		goto out_err;
186 	}
187 
188 	/* uobj is locked */
189 
190 	if ((flags & PGO_NOTIMESTAMP) == 0 &&
191 	    (vp->v_type != VBLK ||
192 	    (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) {
193 		int updflags = 0;
194 
195 		if ((vp->v_mount->mnt_flag & MNT_NOATIME) == 0) {
196 			updflags = GOP_UPDATE_ACCESSED;
197 		}
198 		if (memwrite) {
199 			updflags |= GOP_UPDATE_MODIFIED;
200 		}
201 		if (updflags != 0) {
202 			GOP_MARKUPDATE(vp, updflags);
203 		}
204 	}
205 
206 	/*
207 	 * For PGO_LOCKED requests, just return whatever's in memory.
208 	 */
209 
210 	if (flags & PGO_LOCKED) {
211 		int nfound;
212 		struct vm_page *pg;
213 
214 		KASSERT(!glocked);
215 		npages = *ap->a_count;
216 #if defined(DEBUG)
217 		for (i = 0; i < npages; i++) {
218 			pg = ap->a_m[i];
219 			KASSERT(pg == NULL || pg == PGO_DONTCARE);
220 		}
221 #endif /* defined(DEBUG) */
222 		nfound = uvn_findpages(uobj, origoffset, &npages,
223 		    ap->a_m, UFP_NOWAIT|UFP_NOALLOC|(memwrite ? UFP_NORDONLY : 0));
224 		KASSERT(npages == *ap->a_count);
225 		if (nfound == 0) {
226 			error = EBUSY;
227 			goto out_err;
228 		}
229 		if (!genfs_node_rdtrylock(vp)) {
230 			genfs_rel_pages(ap->a_m, npages);
231 
232 			/*
233 			 * restore the array.
234 			 */
235 
236 			for (i = 0; i < npages; i++) {
237 				pg = ap->a_m[i];
238 
239 				if (pg != NULL || pg != PGO_DONTCARE) {
240 					ap->a_m[i] = NULL;
241 				}
242 			}
243 		} else {
244 			genfs_node_unlock(vp);
245 		}
246 		error = (ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0);
247 		if (error == 0 && memwrite) {
248 			genfs_markdirty(vp);
249 		}
250 		goto out_err;
251 	}
252 	mutex_exit(&uobj->vmobjlock);
253 
254 	/*
255 	 * find the requested pages and make some simple checks.
256 	 * leave space in the page array for a whole block.
257 	 */
258 
259 	const int fs_bshift = (vp->v_type != VBLK) ?
260 	    vp->v_mount->mnt_fs_bshift : DEV_BSHIFT;
261 	const int dev_bshift = (vp->v_type != VBLK) ?
262 	    vp->v_mount->mnt_dev_bshift : DEV_BSHIFT;
263 	const int fs_bsize = 1 << fs_bshift;
264 #define	blk_mask	(fs_bsize - 1)
265 #define	trunc_blk(x)	((x) & ~blk_mask)
266 #define	round_blk(x)	(((x) + blk_mask) & ~blk_mask)
267 
268 	const int orignmempages = MIN(orignpages,
269 	    round_page(memeof - origoffset) >> PAGE_SHIFT);
270 	npages = orignmempages;
271 	const off_t startoffset = trunc_blk(origoffset);
272 	const off_t endoffset = MIN(
273 	    round_page(round_blk(origoffset + (npages << PAGE_SHIFT))),
274 	    round_page(memeof));
275 	const int ridx = (origoffset - startoffset) >> PAGE_SHIFT;
276 
277 	const int pgs_size = sizeof(struct vm_page *) *
278 	    ((endoffset - startoffset) >> PAGE_SHIFT);
279 	struct vm_page **pgs, *pgs_onstack[UBC_MAX_PAGES];
280 
281 	if (pgs_size > sizeof(pgs_onstack)) {
282 		pgs = kmem_zalloc(pgs_size, async ? KM_NOSLEEP : KM_SLEEP);
283 		if (pgs == NULL) {
284 			pgs = pgs_onstack;
285 			error = ENOMEM;
286 			goto out_err;
287 		}
288 	} else {
289 		pgs = pgs_onstack;
290 		(void)memset(pgs, 0, pgs_size);
291 	}
292 
293 	UVMHIST_LOG(ubchist, "ridx %d npages %d startoff %ld endoff %ld",
294 	    ridx, npages, startoffset, endoffset);
295 
296 	if (!has_trans) {
297 		fstrans_start(vp->v_mount, FSTRANS_SHARED);
298 		has_trans = true;
299 	}
300 
301 	/*
302 	 * hold g_glock to prevent a race with truncate.
303 	 *
304 	 * check if our idea of v_size is still valid.
305 	 */
306 
307 	KASSERT(!glocked || genfs_node_wrlocked(vp));
308 	if (!glocked) {
309 		if (blockalloc) {
310 			genfs_node_wrlock(vp);
311 		} else {
312 			genfs_node_rdlock(vp);
313 		}
314 	}
315 	mutex_enter(&uobj->vmobjlock);
316 	if (vp->v_size < origvsize) {
317 		if (!glocked) {
318 			genfs_node_unlock(vp);
319 		}
320 		if (pgs != pgs_onstack)
321 			kmem_free(pgs, pgs_size);
322 		goto startover;
323 	}
324 
325 	if (uvn_findpages(uobj, origoffset, &npages, &pgs[ridx],
326 	    async ? UFP_NOWAIT : UFP_ALL) != orignmempages) {
327 		if (!glocked) {
328 			genfs_node_unlock(vp);
329 		}
330 		KASSERT(async != 0);
331 		genfs_rel_pages(&pgs[ridx], orignmempages);
332 		mutex_exit(&uobj->vmobjlock);
333 		error = EBUSY;
334 		goto out_err_free;
335 	}
336 
337 	/*
338 	 * if the pages are already resident, just return them.
339 	 */
340 
341 	for (i = 0; i < npages; i++) {
342 		struct vm_page *pg = pgs[ridx + i];
343 
344 		if ((pg->flags & PG_FAKE) ||
345 		    (blockalloc && (pg->flags & PG_RDONLY))) {
346 			break;
347 		}
348 	}
349 	if (i == npages) {
350 		if (!glocked) {
351 			genfs_node_unlock(vp);
352 		}
353 		UVMHIST_LOG(ubchist, "returning cached pages", 0,0,0,0);
354 		npages += ridx;
355 		goto out;
356 	}
357 
358 	/*
359 	 * if PGO_OVERWRITE is set, don't bother reading the pages.
360 	 */
361 
362 	if (overwrite) {
363 		if (!glocked) {
364 			genfs_node_unlock(vp);
365 		}
366 		UVMHIST_LOG(ubchist, "PGO_OVERWRITE",0,0,0,0);
367 
368 		for (i = 0; i < npages; i++) {
369 			struct vm_page *pg = pgs[ridx + i];
370 
371 			pg->flags &= ~(PG_RDONLY|PG_CLEAN);
372 		}
373 		npages += ridx;
374 		goto out;
375 	}
376 
377 	/*
378 	 * the page wasn't resident and we're not overwriting,
379 	 * so we're going to have to do some i/o.
380 	 * find any additional pages needed to cover the expanded range.
381 	 */
382 
383 	npages = (endoffset - startoffset) >> PAGE_SHIFT;
384 	if (startoffset != origoffset || npages != orignmempages) {
385 		int npgs;
386 
387 		/*
388 		 * we need to avoid deadlocks caused by locking
389 		 * additional pages at lower offsets than pages we
390 		 * already have locked.  unlock them all and start over.
391 		 */
392 
393 		genfs_rel_pages(&pgs[ridx], orignmempages);
394 		memset(pgs, 0, pgs_size);
395 
396 		UVMHIST_LOG(ubchist, "reset npages start 0x%x end 0x%x",
397 		    startoffset, endoffset, 0,0);
398 		npgs = npages;
399 		if (uvn_findpages(uobj, startoffset, &npgs, pgs,
400 		    async ? UFP_NOWAIT : UFP_ALL) != npages) {
401 			if (!glocked) {
402 				genfs_node_unlock(vp);
403 			}
404 			KASSERT(async != 0);
405 			genfs_rel_pages(pgs, npages);
406 			mutex_exit(&uobj->vmobjlock);
407 			error = EBUSY;
408 			goto out_err_free;
409 		}
410 	}
411 
412 	mutex_exit(&uobj->vmobjlock);
413 
414     {
415 	size_t bytes, iobytes, tailstart, tailbytes, totalbytes, skipbytes;
416 	vaddr_t kva;
417 	struct buf *bp, *mbp;
418 	bool sawhole = false;
419 
420 	/*
421 	 * read the desired page(s).
422 	 */
423 
424 	totalbytes = npages << PAGE_SHIFT;
425 	bytes = MIN(totalbytes, MAX(diskeof - startoffset, 0));
426 	tailbytes = totalbytes - bytes;
427 	skipbytes = 0;
428 
429 	kva = uvm_pagermapin(pgs, npages,
430 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
431 
432 	mbp = getiobuf(vp, true);
433 	mbp->b_bufsize = totalbytes;
434 	mbp->b_data = (void *)kva;
435 	mbp->b_resid = mbp->b_bcount = bytes;
436 	mbp->b_cflags = BC_BUSY;
437 	if (async) {
438 		mbp->b_flags = B_READ | B_ASYNC;
439 		mbp->b_iodone = uvm_aio_biodone;
440 	} else {
441 		mbp->b_flags = B_READ;
442 		mbp->b_iodone = NULL;
443 	}
444 	if (async)
445 		BIO_SETPRIO(mbp, BPRIO_TIMELIMITED);
446 	else
447 		BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL);
448 
449 	/*
450 	 * if EOF is in the middle of the range, zero the part past EOF.
451 	 * skip over pages which are not PG_FAKE since in that case they have
452 	 * valid data that we need to preserve.
453 	 */
454 
455 	tailstart = bytes;
456 	while (tailbytes > 0) {
457 		const int len = PAGE_SIZE - (tailstart & PAGE_MASK);
458 
459 		KASSERT(len <= tailbytes);
460 		if ((pgs[tailstart >> PAGE_SHIFT]->flags & PG_FAKE) != 0) {
461 			memset((void *)(kva + tailstart), 0, len);
462 			UVMHIST_LOG(ubchist, "tailbytes %p 0x%x 0x%x",
463 			    kva, tailstart, len, 0);
464 		}
465 		tailstart += len;
466 		tailbytes -= len;
467 	}
468 
469 	/*
470 	 * now loop over the pages, reading as needed.
471 	 */
472 
473 	bp = NULL;
474 	off_t offset;
475 	for (offset = startoffset;
476 	    bytes > 0;
477 	    offset += iobytes, bytes -= iobytes) {
478 		int run;
479 		daddr_t lbn, blkno;
480 		int pidx;
481 		struct vnode *devvp;
482 
483 		/*
484 		 * skip pages which don't need to be read.
485 		 */
486 
487 		pidx = (offset - startoffset) >> PAGE_SHIFT;
488 		while ((pgs[pidx]->flags & PG_FAKE) == 0) {
489 			size_t b;
490 
491 			KASSERT((offset & (PAGE_SIZE - 1)) == 0);
492 			if ((pgs[pidx]->flags & PG_RDONLY)) {
493 				sawhole = true;
494 			}
495 			b = MIN(PAGE_SIZE, bytes);
496 			offset += b;
497 			bytes -= b;
498 			skipbytes += b;
499 			pidx++;
500 			UVMHIST_LOG(ubchist, "skipping, new offset 0x%x",
501 			    offset, 0,0,0);
502 			if (bytes == 0) {
503 				goto loopdone;
504 			}
505 		}
506 
507 		/*
508 		 * bmap the file to find out the blkno to read from and
509 		 * how much we can read in one i/o.  if bmap returns an error,
510 		 * skip the rest of the top-level i/o.
511 		 */
512 
513 		lbn = offset >> fs_bshift;
514 		error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run);
515 		if (error) {
516 			UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%x -> %d\n",
517 			    lbn,error,0,0);
518 			skipbytes += bytes;
519 			bytes = 0;
520 			goto loopdone;
521 		}
522 
523 		/*
524 		 * see how many pages can be read with this i/o.
525 		 * reduce the i/o size if necessary to avoid
526 		 * overwriting pages with valid data.
527 		 */
528 
529 		iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset,
530 		    bytes);
531 		if (offset + iobytes > round_page(offset)) {
532 			int pcount;
533 
534 			pcount = 1;
535 			while (pidx + pcount < npages &&
536 			    pgs[pidx + pcount]->flags & PG_FAKE) {
537 				pcount++;
538 			}
539 			iobytes = MIN(iobytes, (pcount << PAGE_SHIFT) -
540 			    (offset - trunc_page(offset)));
541 		}
542 
543 		/*
544 		 * if this block isn't allocated, zero it instead of
545 		 * reading it.  unless we are going to allocate blocks,
546 		 * mark the pages we zeroed PG_RDONLY.
547 		 */
548 
549 		if (blkno == (daddr_t)-1) {
550 			int holepages = (round_page(offset + iobytes) -
551 			    trunc_page(offset)) >> PAGE_SHIFT;
552 			UVMHIST_LOG(ubchist, "lbn 0x%x -> HOLE", lbn,0,0,0);
553 
554 			sawhole = true;
555 			memset((char *)kva + (offset - startoffset), 0,
556 			    iobytes);
557 			skipbytes += iobytes;
558 
559 			for (i = 0; i < holepages; i++) {
560 				if (memwrite) {
561 					pgs[pidx + i]->flags &= ~PG_CLEAN;
562 				}
563 				if (!blockalloc) {
564 					pgs[pidx + i]->flags |= PG_RDONLY;
565 				}
566 			}
567 			continue;
568 		}
569 
570 		/*
571 		 * allocate a sub-buf for this piece of the i/o
572 		 * (or just use mbp if there's only 1 piece),
573 		 * and start it going.
574 		 */
575 
576 		if (offset == startoffset && iobytes == bytes) {
577 			bp = mbp;
578 		} else {
579 			UVMHIST_LOG(ubchist, "vp %p bp %p num now %d",
580 			    vp, bp, vp->v_numoutput, 0);
581 			bp = getiobuf(vp, true);
582 			nestiobuf_setup(mbp, bp, offset - startoffset, iobytes);
583 		}
584 		bp->b_lblkno = 0;
585 
586 		/* adjust physical blkno for partial blocks */
587 		bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >>
588 		    dev_bshift);
589 
590 		UVMHIST_LOG(ubchist,
591 		    "bp %p offset 0x%x bcount 0x%x blkno 0x%x",
592 		    bp, offset, bp->b_bcount, bp->b_blkno);
593 
594 		VOP_STRATEGY(devvp, bp);
595 	}
596 
597 loopdone:
598 	nestiobuf_done(mbp, skipbytes, error);
599 	if (async) {
600 		UVMHIST_LOG(ubchist, "returning 0 (async)",0,0,0,0);
601 		if (!glocked) {
602 			genfs_node_unlock(vp);
603 		}
604 		error = 0;
605 		goto out_err_free;
606 	}
607 	if (bp != NULL) {
608 		error = biowait(mbp);
609 	}
610 
611 	/* Remove the mapping (make KVA available as soon as possible) */
612 	uvm_pagermapout(kva, npages);
613 
614 	/*
615 	 * if this we encountered a hole then we have to do a little more work.
616 	 * for read faults, we marked the page PG_RDONLY so that future
617 	 * write accesses to the page will fault again.
618 	 * for write faults, we must make sure that the backing store for
619 	 * the page is completely allocated while the pages are locked.
620 	 */
621 
622 	if (!error && sawhole && blockalloc) {
623 		/*
624 		 * XXX: This assumes that we come here only via
625 		 * the mmio path
626 		 */
627 		if (vp->v_mount->mnt_wapbl) {
628 			error = WAPBL_BEGIN(vp->v_mount);
629 		}
630 
631 		if (!error) {
632 			error = GOP_ALLOC(vp, startoffset,
633 			    npages << PAGE_SHIFT, 0, cred);
634 			if (vp->v_mount->mnt_wapbl) {
635 				WAPBL_END(vp->v_mount);
636 			}
637 		}
638 
639 		UVMHIST_LOG(ubchist, "gop_alloc off 0x%x/0x%x -> %d",
640 		    startoffset, npages << PAGE_SHIFT, error,0);
641 		if (!error) {
642 			for (i = 0; i < npages; i++) {
643 				struct vm_page *pg = pgs[i];
644 
645 				if (pg == NULL) {
646 					continue;
647 				}
648 				pg->flags &= ~(PG_CLEAN|PG_RDONLY);
649 				UVMHIST_LOG(ubchist, "mark dirty pg %p",
650 				    pg,0,0,0);
651 			}
652 		}
653 	}
654 	if (!glocked) {
655 		genfs_node_unlock(vp);
656 	}
657 
658 	putiobuf(mbp);
659     }
660 
661 	mutex_enter(&uobj->vmobjlock);
662 
663 	/*
664 	 * we're almost done!  release the pages...
665 	 * for errors, we free the pages.
666 	 * otherwise we activate them and mark them as valid and clean.
667 	 * also, unbusy pages that were not actually requested.
668 	 */
669 
670 	if (error) {
671 		for (i = 0; i < npages; i++) {
672 			struct vm_page *pg = pgs[i];
673 
674 			if (pg == NULL) {
675 				continue;
676 			}
677 			UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x",
678 			    pg, pg->flags, 0,0);
679 			if (pg->flags & PG_FAKE) {
680 				pg->flags |= PG_RELEASED;
681 			}
682 		}
683 		mutex_enter(&uvm_pageqlock);
684 		uvm_page_unbusy(pgs, npages);
685 		mutex_exit(&uvm_pageqlock);
686 		mutex_exit(&uobj->vmobjlock);
687 		UVMHIST_LOG(ubchist, "returning error %d", error,0,0,0);
688 		goto out_err_free;
689 	}
690 
691 out:
692 	UVMHIST_LOG(ubchist, "succeeding, npages %d", npages,0,0,0);
693 	error = 0;
694 	mutex_enter(&uvm_pageqlock);
695 	for (i = 0; i < npages; i++) {
696 		struct vm_page *pg = pgs[i];
697 		if (pg == NULL) {
698 			continue;
699 		}
700 		UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x",
701 		    pg, pg->flags, 0,0);
702 		if (pg->flags & PG_FAKE && !overwrite) {
703 			pg->flags &= ~(PG_FAKE);
704 			pmap_clear_modify(pgs[i]);
705 		}
706 		KASSERT(!memwrite || !blockalloc || (pg->flags & PG_RDONLY) == 0);
707 		if (i < ridx || i >= ridx + orignmempages || async) {
708 			UVMHIST_LOG(ubchist, "unbusy pg %p offset 0x%x",
709 			    pg, pg->offset,0,0);
710 			if (pg->flags & PG_WANTED) {
711 				wakeup(pg);
712 			}
713 			if (pg->flags & PG_FAKE) {
714 				KASSERT(overwrite);
715 				uvm_pagezero(pg);
716 			}
717 			if (pg->flags & PG_RELEASED) {
718 				uvm_pagefree(pg);
719 				continue;
720 			}
721 			uvm_pageenqueue(pg);
722 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
723 			UVM_PAGE_OWN(pg, NULL);
724 		}
725 	}
726 	mutex_exit(&uvm_pageqlock);
727 	if (memwrite) {
728 		genfs_markdirty(vp);
729 	}
730 	mutex_exit(&uobj->vmobjlock);
731 	if (ap->a_m != NULL) {
732 		memcpy(ap->a_m, &pgs[ridx],
733 		    orignmempages * sizeof(struct vm_page *));
734 	}
735 
736 out_err_free:
737 	if (pgs != NULL && pgs != pgs_onstack)
738 		kmem_free(pgs, pgs_size);
739 out_err:
740 	if (has_trans)
741 		fstrans_done(vp->v_mount);
742 	return error;
743 }
744 
745 /*
746  * generic VM putpages routine.
747  * Write the given range of pages to backing store.
748  *
749  * => "offhi == 0" means flush all pages at or after "offlo".
750  * => object should be locked by caller.  we return with the
751  *      object unlocked.
752  * => if PGO_CLEANIT or PGO_SYNCIO is set, we may block (due to I/O).
753  *	thus, a caller might want to unlock higher level resources
754  *	(e.g. vm_map) before calling flush.
755  * => if neither PGO_CLEANIT nor PGO_SYNCIO is set, we will not block
756  * => if PGO_ALLPAGES is set, then all pages in the object will be processed.
757  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
758  *	that new pages are inserted on the tail end of the list.   thus,
759  *	we can make a complete pass through the object in one go by starting
760  *	at the head and working towards the tail (new pages are put in
761  *	front of us).
762  * => NOTE: we are allowed to lock the page queues, so the caller
763  *	must not be holding the page queue lock.
764  *
765  * note on "cleaning" object and PG_BUSY pages:
766  *	this routine is holding the lock on the object.   the only time
767  *	that it can run into a PG_BUSY page that it does not own is if
768  *	some other process has started I/O on the page (e.g. either
769  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
770  *	in, then it can not be dirty (!PG_CLEAN) because no one has
771  *	had a chance to modify it yet.    if the PG_BUSY page is being
772  *	paged out then it means that someone else has already started
773  *	cleaning the page for us (how nice!).    in this case, if we
774  *	have syncio specified, then after we make our pass through the
775  *	object we need to wait for the other PG_BUSY pages to clear
776  *	off (i.e. we need to do an iosync).   also note that once a
777  *	page is PG_BUSY it must stay in its object until it is un-busyed.
778  *
779  * note on page traversal:
780  *	we can traverse the pages in an object either by going down the
781  *	linked list in "uobj->memq", or we can go over the address range
782  *	by page doing hash table lookups for each address.    depending
783  *	on how many pages are in the object it may be cheaper to do one
784  *	or the other.   we set "by_list" to true if we are using memq.
785  *	if the cost of a hash lookup was equal to the cost of the list
786  *	traversal we could compare the number of pages in the start->stop
787  *	range to the total number of pages in the object.   however, it
788  *	seems that a hash table lookup is more expensive than the linked
789  *	list traversal, so we multiply the number of pages in the
790  *	range by an estimate of the relatively higher cost of the hash lookup.
791  */
792 
793 int
794 genfs_putpages(void *v)
795 {
796 	struct vop_putpages_args /* {
797 		struct vnode *a_vp;
798 		voff_t a_offlo;
799 		voff_t a_offhi;
800 		int a_flags;
801 	} */ * const ap = v;
802 
803 	return genfs_do_putpages(ap->a_vp, ap->a_offlo, ap->a_offhi,
804 	    ap->a_flags, NULL);
805 }
806 
807 int
808 genfs_do_putpages(struct vnode *vp, off_t startoff, off_t endoff,
809     int origflags, struct vm_page **busypg)
810 {
811 	struct uvm_object * const uobj = &vp->v_uobj;
812 	kmutex_t * const slock = &uobj->vmobjlock;
813 	off_t off;
814 	/* Even for strange MAXPHYS, the shift rounds down to a page */
815 #define maxpages (MAXPHYS >> PAGE_SHIFT)
816 	int i, error, npages, nback;
817 	int freeflag;
818 	struct vm_page *pgs[maxpages], *pg, *nextpg, *tpg, curmp, endmp;
819 	bool wasclean, by_list, needs_clean, yld;
820 	bool async = (origflags & PGO_SYNCIO) == 0;
821 	bool pagedaemon = curlwp == uvm.pagedaemon_lwp;
822 	struct lwp * const l = curlwp ? curlwp : &lwp0;
823 	struct genfs_node * const gp = VTOG(vp);
824 	int flags;
825 	int dirtygen;
826 	bool modified;
827 	bool need_wapbl;
828 	bool has_trans;
829 	bool cleanall;
830 	bool onworklst;
831 
832 	UVMHIST_FUNC("genfs_putpages"); UVMHIST_CALLED(ubchist);
833 
834 	KASSERT(origflags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE));
835 	KASSERT((startoff & PAGE_MASK) == 0 && (endoff & PAGE_MASK) == 0);
836 	KASSERT(startoff < endoff || endoff == 0);
837 
838 	UVMHIST_LOG(ubchist, "vp %p pages %d off 0x%x len 0x%x",
839 	    vp, uobj->uo_npages, startoff, endoff - startoff);
840 
841 	has_trans = false;
842 	need_wapbl = (!pagedaemon && vp->v_mount && vp->v_mount->mnt_wapbl &&
843 	    (origflags & PGO_JOURNALLOCKED) == 0);
844 
845 retry:
846 	modified = false;
847 	flags = origflags;
848 	KASSERT((vp->v_iflag & VI_ONWORKLST) != 0 ||
849 	    (vp->v_iflag & VI_WRMAPDIRTY) == 0);
850 	if (uobj->uo_npages == 0) {
851 		if (vp->v_iflag & VI_ONWORKLST) {
852 			vp->v_iflag &= ~VI_WRMAPDIRTY;
853 			if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL)
854 				vn_syncer_remove_from_worklist(vp);
855 		}
856 		if (has_trans) {
857 			if (need_wapbl)
858 				WAPBL_END(vp->v_mount);
859 			fstrans_done(vp->v_mount);
860 		}
861 		mutex_exit(slock);
862 		return (0);
863 	}
864 
865 	/*
866 	 * the vnode has pages, set up to process the request.
867 	 */
868 
869 	if (!has_trans && (flags & PGO_CLEANIT) != 0) {
870 		mutex_exit(slock);
871 		if (pagedaemon) {
872 			error = fstrans_start_nowait(vp->v_mount, FSTRANS_LAZY);
873 			if (error)
874 				return error;
875 		} else
876 			fstrans_start(vp->v_mount, FSTRANS_LAZY);
877 		if (need_wapbl) {
878 			error = WAPBL_BEGIN(vp->v_mount);
879 			if (error) {
880 				fstrans_done(vp->v_mount);
881 				return error;
882 			}
883 		}
884 		has_trans = true;
885 		mutex_enter(slock);
886 		goto retry;
887 	}
888 
889 	error = 0;
890 	wasclean = (vp->v_numoutput == 0);
891 	off = startoff;
892 	if (endoff == 0 || flags & PGO_ALLPAGES) {
893 		endoff = trunc_page(LLONG_MAX);
894 	}
895 	by_list = (uobj->uo_npages <=
896 	    ((endoff - startoff) >> PAGE_SHIFT) * UVM_PAGE_TREE_PENALTY);
897 
898 #if !defined(DEBUG)
899 	/*
900 	 * if this vnode is known not to have dirty pages,
901 	 * don't bother to clean it out.
902 	 */
903 
904 	if ((vp->v_iflag & VI_ONWORKLST) == 0) {
905 		if ((flags & (PGO_FREE|PGO_DEACTIVATE)) == 0) {
906 			goto skip_scan;
907 		}
908 		flags &= ~PGO_CLEANIT;
909 	}
910 #endif /* !defined(DEBUG) */
911 
912 	/*
913 	 * start the loop.  when scanning by list, hold the last page
914 	 * in the list before we start.  pages allocated after we start
915 	 * will be added to the end of the list, so we can stop at the
916 	 * current last page.
917 	 */
918 
919 	cleanall = (flags & PGO_CLEANIT) != 0 && wasclean &&
920 	    startoff == 0 && endoff == trunc_page(LLONG_MAX) &&
921 	    (vp->v_iflag & VI_ONWORKLST) != 0;
922 	dirtygen = gp->g_dirtygen;
923 	freeflag = pagedaemon ? PG_PAGEOUT : PG_RELEASED;
924 	if (by_list) {
925 		curmp.flags = PG_MARKER;
926 		endmp.flags = PG_MARKER;
927 		pg = TAILQ_FIRST(&uobj->memq);
928 		TAILQ_INSERT_TAIL(&uobj->memq, &endmp, listq.queue);
929 	} else {
930 		pg = uvm_pagelookup(uobj, off);
931 	}
932 	nextpg = NULL;
933 	while (by_list || off < endoff) {
934 
935 		/*
936 		 * if the current page is not interesting, move on to the next.
937 		 */
938 
939 		KASSERT(pg == NULL || pg->uobject == uobj ||
940 		    (pg->flags & PG_MARKER) != 0);
941 		KASSERT(pg == NULL ||
942 		    (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 ||
943 		    (pg->flags & (PG_BUSY|PG_MARKER)) != 0);
944 		if (by_list) {
945 			if (pg == &endmp) {
946 				break;
947 			}
948 			if (pg->flags & PG_MARKER) {
949 				pg = TAILQ_NEXT(pg, listq.queue);
950 				continue;
951 			}
952 			if (pg->offset < startoff || pg->offset >= endoff ||
953 			    pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
954 				if (pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
955 					wasclean = false;
956 				}
957 				pg = TAILQ_NEXT(pg, listq.queue);
958 				continue;
959 			}
960 			off = pg->offset;
961 		} else if (pg == NULL || pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
962 			if (pg != NULL) {
963 				wasclean = false;
964 			}
965 			off += PAGE_SIZE;
966 			if (off < endoff) {
967 				pg = uvm_pagelookup(uobj, off);
968 			}
969 			continue;
970 		}
971 
972 		/*
973 		 * if the current page needs to be cleaned and it's busy,
974 		 * wait for it to become unbusy.
975 		 */
976 
977 		yld = (l->l_cpu->ci_schedstate.spc_flags &
978 		    SPCF_SHOULDYIELD) && !pagedaemon;
979 		if (pg->flags & PG_BUSY || yld) {
980 			UVMHIST_LOG(ubchist, "busy %p", pg,0,0,0);
981 			if (flags & PGO_BUSYFAIL && pg->flags & PG_BUSY) {
982 				UVMHIST_LOG(ubchist, "busyfail %p", pg, 0,0,0);
983 				error = EDEADLK;
984 				if (busypg != NULL)
985 					*busypg = pg;
986 				break;
987 			}
988 			if (pagedaemon) {
989 				/*
990 				 * someone has taken the page while we
991 				 * dropped the lock for fstrans_start.
992 				 */
993 				break;
994 			}
995 			if (by_list) {
996 				TAILQ_INSERT_BEFORE(pg, &curmp, listq.queue);
997 				UVMHIST_LOG(ubchist, "curmp next %p",
998 				    TAILQ_NEXT(&curmp, listq.queue), 0,0,0);
999 			}
1000 			if (yld) {
1001 				mutex_exit(slock);
1002 				preempt();
1003 				mutex_enter(slock);
1004 			} else {
1005 				pg->flags |= PG_WANTED;
1006 				UVM_UNLOCK_AND_WAIT(pg, slock, 0, "genput", 0);
1007 				mutex_enter(slock);
1008 			}
1009 			if (by_list) {
1010 				UVMHIST_LOG(ubchist, "after next %p",
1011 				    TAILQ_NEXT(&curmp, listq.queue), 0,0,0);
1012 				pg = TAILQ_NEXT(&curmp, listq.queue);
1013 				TAILQ_REMOVE(&uobj->memq, &curmp, listq.queue);
1014 			} else {
1015 				pg = uvm_pagelookup(uobj, off);
1016 			}
1017 			continue;
1018 		}
1019 
1020 		/*
1021 		 * if we're freeing, remove all mappings of the page now.
1022 		 * if we're cleaning, check if the page is needs to be cleaned.
1023 		 */
1024 
1025 		if (flags & PGO_FREE) {
1026 			pmap_page_protect(pg, VM_PROT_NONE);
1027 		} else if (flags & PGO_CLEANIT) {
1028 
1029 			/*
1030 			 * if we still have some hope to pull this vnode off
1031 			 * from the syncer queue, write-protect the page.
1032 			 */
1033 
1034 			if (cleanall && wasclean &&
1035 			    gp->g_dirtygen == dirtygen) {
1036 
1037 				/*
1038 				 * uobj pages get wired only by uvm_fault
1039 				 * where uobj is locked.
1040 				 */
1041 
1042 				if (pg->wire_count == 0) {
1043 					pmap_page_protect(pg,
1044 					    VM_PROT_READ|VM_PROT_EXECUTE);
1045 				} else {
1046 					cleanall = false;
1047 				}
1048 			}
1049 		}
1050 
1051 		if (flags & PGO_CLEANIT) {
1052 			needs_clean = pmap_clear_modify(pg) ||
1053 			    (pg->flags & PG_CLEAN) == 0;
1054 			pg->flags |= PG_CLEAN;
1055 		} else {
1056 			needs_clean = false;
1057 		}
1058 
1059 		/*
1060 		 * if we're cleaning, build a cluster.
1061 		 * the cluster will consist of pages which are currently dirty,
1062 		 * but they will be returned to us marked clean.
1063 		 * if not cleaning, just operate on the one page.
1064 		 */
1065 
1066 		if (needs_clean) {
1067 			KDASSERT((vp->v_iflag & VI_ONWORKLST));
1068 			wasclean = false;
1069 			memset(pgs, 0, sizeof(pgs));
1070 			pg->flags |= PG_BUSY;
1071 			UVM_PAGE_OWN(pg, "genfs_putpages");
1072 
1073 			/*
1074 			 * first look backward.
1075 			 */
1076 
1077 			npages = MIN(maxpages >> 1, off >> PAGE_SHIFT);
1078 			nback = npages;
1079 			uvn_findpages(uobj, off - PAGE_SIZE, &nback, &pgs[0],
1080 			    UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY|UFP_BACKWARD);
1081 			if (nback) {
1082 				memmove(&pgs[0], &pgs[npages - nback],
1083 				    nback * sizeof(pgs[0]));
1084 				if (npages - nback < nback)
1085 					memset(&pgs[nback], 0,
1086 					    (npages - nback) * sizeof(pgs[0]));
1087 				else
1088 					memset(&pgs[npages - nback], 0,
1089 					    nback * sizeof(pgs[0]));
1090 			}
1091 
1092 			/*
1093 			 * then plug in our page of interest.
1094 			 */
1095 
1096 			pgs[nback] = pg;
1097 
1098 			/*
1099 			 * then look forward to fill in the remaining space in
1100 			 * the array of pages.
1101 			 */
1102 
1103 			npages = maxpages - nback - 1;
1104 			uvn_findpages(uobj, off + PAGE_SIZE, &npages,
1105 			    &pgs[nback + 1],
1106 			    UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY);
1107 			npages += nback + 1;
1108 		} else {
1109 			pgs[0] = pg;
1110 			npages = 1;
1111 			nback = 0;
1112 		}
1113 
1114 		/*
1115 		 * apply FREE or DEACTIVATE options if requested.
1116 		 */
1117 
1118 		if (flags & (PGO_DEACTIVATE|PGO_FREE)) {
1119 			mutex_enter(&uvm_pageqlock);
1120 		}
1121 		for (i = 0; i < npages; i++) {
1122 			tpg = pgs[i];
1123 			KASSERT(tpg->uobject == uobj);
1124 			if (by_list && tpg == TAILQ_NEXT(pg, listq.queue))
1125 				pg = tpg;
1126 			if (tpg->offset < startoff || tpg->offset >= endoff)
1127 				continue;
1128 			if (flags & PGO_DEACTIVATE && tpg->wire_count == 0) {
1129 				uvm_pagedeactivate(tpg);
1130 			} else if (flags & PGO_FREE) {
1131 				pmap_page_protect(tpg, VM_PROT_NONE);
1132 				if (tpg->flags & PG_BUSY) {
1133 					tpg->flags |= freeflag;
1134 					if (pagedaemon) {
1135 						uvm_pageout_start(1);
1136 						uvm_pagedequeue(tpg);
1137 					}
1138 				} else {
1139 
1140 					/*
1141 					 * ``page is not busy''
1142 					 * implies that npages is 1
1143 					 * and needs_clean is false.
1144 					 */
1145 
1146 					nextpg = TAILQ_NEXT(tpg, listq.queue);
1147 					uvm_pagefree(tpg);
1148 					if (pagedaemon)
1149 						uvmexp.pdfreed++;
1150 				}
1151 			}
1152 		}
1153 		if (flags & (PGO_DEACTIVATE|PGO_FREE)) {
1154 			mutex_exit(&uvm_pageqlock);
1155 		}
1156 		if (needs_clean) {
1157 			modified = true;
1158 
1159 			/*
1160 			 * start the i/o.  if we're traversing by list,
1161 			 * keep our place in the list with a marker page.
1162 			 */
1163 
1164 			if (by_list) {
1165 				TAILQ_INSERT_AFTER(&uobj->memq, pg, &curmp,
1166 				    listq.queue);
1167 			}
1168 			mutex_exit(slock);
1169 			error = GOP_WRITE(vp, pgs, npages, flags);
1170 			mutex_enter(slock);
1171 			if (by_list) {
1172 				pg = TAILQ_NEXT(&curmp, listq.queue);
1173 				TAILQ_REMOVE(&uobj->memq, &curmp, listq.queue);
1174 			}
1175 			if (error) {
1176 				break;
1177 			}
1178 			if (by_list) {
1179 				continue;
1180 			}
1181 		}
1182 
1183 		/*
1184 		 * find the next page and continue if there was no error.
1185 		 */
1186 
1187 		if (by_list) {
1188 			if (nextpg) {
1189 				pg = nextpg;
1190 				nextpg = NULL;
1191 			} else {
1192 				pg = TAILQ_NEXT(pg, listq.queue);
1193 			}
1194 		} else {
1195 			off += (npages - nback) << PAGE_SHIFT;
1196 			if (off < endoff) {
1197 				pg = uvm_pagelookup(uobj, off);
1198 			}
1199 		}
1200 	}
1201 	if (by_list) {
1202 		TAILQ_REMOVE(&uobj->memq, &endmp, listq.queue);
1203 	}
1204 
1205 	if (modified && (vp->v_iflag & VI_WRMAPDIRTY) != 0 &&
1206 	    (vp->v_type != VBLK ||
1207 	    (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) {
1208 		GOP_MARKUPDATE(vp, GOP_UPDATE_MODIFIED);
1209 	}
1210 
1211 	/*
1212 	 * if we're cleaning and there was nothing to clean,
1213 	 * take us off the syncer list.  if we started any i/o
1214 	 * and we're doing sync i/o, wait for all writes to finish.
1215 	 */
1216 
1217 	if (cleanall && wasclean && gp->g_dirtygen == dirtygen &&
1218 	    (vp->v_iflag & VI_ONWORKLST) != 0) {
1219 #if defined(DEBUG)
1220 		TAILQ_FOREACH(pg, &uobj->memq, listq.queue) {
1221 			if ((pg->flags & PG_MARKER) != 0) {
1222 				continue;
1223 			}
1224 			if ((pg->flags & PG_CLEAN) == 0) {
1225 				printf("%s: %p: !CLEAN\n", __func__, pg);
1226 			}
1227 			if (pmap_is_modified(pg)) {
1228 				printf("%s: %p: modified\n", __func__, pg);
1229 			}
1230 		}
1231 #endif /* defined(DEBUG) */
1232 		vp->v_iflag &= ~VI_WRMAPDIRTY;
1233 		if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL)
1234 			vn_syncer_remove_from_worklist(vp);
1235 	}
1236 
1237 #if !defined(DEBUG)
1238 skip_scan:
1239 #endif /* !defined(DEBUG) */
1240 
1241 	/* Wait for output to complete. */
1242 	if (!wasclean && !async && vp->v_numoutput != 0) {
1243 		while (vp->v_numoutput != 0)
1244 			cv_wait(&vp->v_cv, slock);
1245 	}
1246 	onworklst = (vp->v_iflag & VI_ONWORKLST) != 0;
1247 	mutex_exit(slock);
1248 
1249 	if ((flags & PGO_RECLAIM) != 0 && onworklst) {
1250 		/*
1251 		 * in the case of PGO_RECLAIM, ensure to make the vnode clean.
1252 		 * retrying is not a big deal because, in many cases,
1253 		 * uobj->uo_npages is already 0 here.
1254 		 */
1255 		mutex_enter(slock);
1256 		goto retry;
1257 	}
1258 
1259 	if (has_trans) {
1260 		if (need_wapbl)
1261 			WAPBL_END(vp->v_mount);
1262 		fstrans_done(vp->v_mount);
1263 	}
1264 
1265 	return (error);
1266 }
1267 
1268 int
1269 genfs_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, int flags)
1270 {
1271 	off_t off;
1272 	vaddr_t kva;
1273 	size_t len;
1274 	int error;
1275 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1276 
1277 	UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x",
1278 	    vp, pgs, npages, flags);
1279 
1280 	off = pgs[0]->offset;
1281 	kva = uvm_pagermapin(pgs, npages,
1282 	    UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK);
1283 	len = npages << PAGE_SHIFT;
1284 
1285 	error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE,
1286 			    uvm_aio_biodone);
1287 
1288 	return error;
1289 }
1290 
1291 int
1292 genfs_gop_write_rwmap(struct vnode *vp, struct vm_page **pgs, int npages, int flags)
1293 {
1294 	off_t off;
1295 	vaddr_t kva;
1296 	size_t len;
1297 	int error;
1298 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1299 
1300 	UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x",
1301 	    vp, pgs, npages, flags);
1302 
1303 	off = pgs[0]->offset;
1304 	kva = uvm_pagermapin(pgs, npages,
1305 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
1306 	len = npages << PAGE_SHIFT;
1307 
1308 	error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE,
1309 			    uvm_aio_biodone);
1310 
1311 	return error;
1312 }
1313 
1314 /*
1315  * Backend routine for doing I/O to vnode pages.  Pages are already locked
1316  * and mapped into kernel memory.  Here we just look up the underlying
1317  * device block addresses and call the strategy routine.
1318  */
1319 
1320 static int
1321 genfs_do_io(struct vnode *vp, off_t off, vaddr_t kva, size_t len, int flags,
1322     enum uio_rw rw, void (*iodone)(struct buf *))
1323 {
1324 	int s, error;
1325 	int fs_bshift, dev_bshift;
1326 	off_t eof, offset, startoffset;
1327 	size_t bytes, iobytes, skipbytes;
1328 	struct buf *mbp, *bp;
1329 	const bool async = (flags & PGO_SYNCIO) == 0;
1330 	const bool iowrite = rw == UIO_WRITE;
1331 	const int brw = iowrite ? B_WRITE : B_READ;
1332 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1333 
1334 	UVMHIST_LOG(ubchist, "vp %p kva %p len 0x%x flags 0x%x",
1335 	    vp, kva, len, flags);
1336 
1337 	KASSERT(vp->v_size <= vp->v_writesize);
1338 	GOP_SIZE(vp, vp->v_writesize, &eof, 0);
1339 	if (vp->v_type != VBLK) {
1340 		fs_bshift = vp->v_mount->mnt_fs_bshift;
1341 		dev_bshift = vp->v_mount->mnt_dev_bshift;
1342 	} else {
1343 		fs_bshift = DEV_BSHIFT;
1344 		dev_bshift = DEV_BSHIFT;
1345 	}
1346 	error = 0;
1347 	startoffset = off;
1348 	bytes = MIN(len, eof - startoffset);
1349 	skipbytes = 0;
1350 	KASSERT(bytes != 0);
1351 
1352 	if (iowrite) {
1353 		mutex_enter(&vp->v_interlock);
1354 		vp->v_numoutput += 2;
1355 		mutex_exit(&vp->v_interlock);
1356 	}
1357 	mbp = getiobuf(vp, true);
1358 	UVMHIST_LOG(ubchist, "vp %p mbp %p num now %d bytes 0x%x",
1359 	    vp, mbp, vp->v_numoutput, bytes);
1360 	mbp->b_bufsize = len;
1361 	mbp->b_data = (void *)kva;
1362 	mbp->b_resid = mbp->b_bcount = bytes;
1363 	mbp->b_cflags = BC_BUSY | BC_AGE;
1364 	if (async) {
1365 		mbp->b_flags = brw | B_ASYNC;
1366 		mbp->b_iodone = iodone;
1367 	} else {
1368 		mbp->b_flags = brw;
1369 		mbp->b_iodone = NULL;
1370 	}
1371 	if (curlwp == uvm.pagedaemon_lwp)
1372 		BIO_SETPRIO(mbp, BPRIO_TIMELIMITED);
1373 	else if (async)
1374 		BIO_SETPRIO(mbp, BPRIO_TIMENONCRITICAL);
1375 	else
1376 		BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL);
1377 
1378 	bp = NULL;
1379 	for (offset = startoffset;
1380 	    bytes > 0;
1381 	    offset += iobytes, bytes -= iobytes) {
1382 		int run;
1383 		daddr_t lbn, blkno;
1384 		struct vnode *devvp;
1385 
1386 		/*
1387 		 * bmap the file to find out the blkno to read from and
1388 		 * how much we can read in one i/o.  if bmap returns an error,
1389 		 * skip the rest of the top-level i/o.
1390 		 */
1391 
1392 		lbn = offset >> fs_bshift;
1393 		error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run);
1394 		if (error) {
1395 			UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%x -> %d\n",
1396 			    lbn,error,0,0);
1397 			skipbytes += bytes;
1398 			bytes = 0;
1399 			goto loopdone;
1400 		}
1401 
1402 		/*
1403 		 * see how many pages can be read with this i/o.
1404 		 * reduce the i/o size if necessary to avoid
1405 		 * overwriting pages with valid data.
1406 		 */
1407 
1408 		iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset,
1409 		    bytes);
1410 
1411 		/*
1412 		 * if this block isn't allocated, zero it instead of
1413 		 * reading it.  unless we are going to allocate blocks,
1414 		 * mark the pages we zeroed PG_RDONLY.
1415 		 */
1416 
1417 		if (blkno == (daddr_t)-1) {
1418 			if (!iowrite) {
1419 				memset((char *)kva + (offset - startoffset), 0,
1420 				    iobytes);
1421 			}
1422 			skipbytes += iobytes;
1423 			continue;
1424 		}
1425 
1426 		/*
1427 		 * allocate a sub-buf for this piece of the i/o
1428 		 * (or just use mbp if there's only 1 piece),
1429 		 * and start it going.
1430 		 */
1431 
1432 		if (offset == startoffset && iobytes == bytes) {
1433 			bp = mbp;
1434 		} else {
1435 			UVMHIST_LOG(ubchist, "vp %p bp %p num now %d",
1436 			    vp, bp, vp->v_numoutput, 0);
1437 			bp = getiobuf(vp, true);
1438 			nestiobuf_setup(mbp, bp, offset - startoffset, iobytes);
1439 		}
1440 		bp->b_lblkno = 0;
1441 
1442 		/* adjust physical blkno for partial blocks */
1443 		bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >>
1444 		    dev_bshift);
1445 
1446 		UVMHIST_LOG(ubchist,
1447 		    "bp %p offset 0x%x bcount 0x%x blkno 0x%x",
1448 		    bp, offset, bp->b_bcount, bp->b_blkno);
1449 
1450 		VOP_STRATEGY(devvp, bp);
1451 	}
1452 
1453 loopdone:
1454 	if (skipbytes) {
1455 		UVMHIST_LOG(ubchist, "skipbytes %d", skipbytes, 0,0,0);
1456 	}
1457 	nestiobuf_done(mbp, skipbytes, error);
1458 	if (async) {
1459 		UVMHIST_LOG(ubchist, "returning 0 (async)", 0,0,0,0);
1460 		return (0);
1461 	}
1462 	UVMHIST_LOG(ubchist, "waiting for mbp %p", mbp,0,0,0);
1463 	error = biowait(mbp);
1464 	s = splbio();
1465 	(*iodone)(mbp);
1466 	splx(s);
1467 	UVMHIST_LOG(ubchist, "returning, error %d", error,0,0,0);
1468 	return (error);
1469 }
1470 
1471 int
1472 genfs_compat_getpages(void *v)
1473 {
1474 	struct vop_getpages_args /* {
1475 		struct vnode *a_vp;
1476 		voff_t a_offset;
1477 		struct vm_page **a_m;
1478 		int *a_count;
1479 		int a_centeridx;
1480 		vm_prot_t a_access_type;
1481 		int a_advice;
1482 		int a_flags;
1483 	} */ *ap = v;
1484 
1485 	off_t origoffset;
1486 	struct vnode *vp = ap->a_vp;
1487 	struct uvm_object *uobj = &vp->v_uobj;
1488 	struct vm_page *pg, **pgs;
1489 	vaddr_t kva;
1490 	int i, error, orignpages, npages;
1491 	struct iovec iov;
1492 	struct uio uio;
1493 	kauth_cred_t cred = curlwp->l_cred;
1494 	const bool memwrite = (ap->a_access_type & VM_PROT_WRITE) != 0;
1495 
1496 	error = 0;
1497 	origoffset = ap->a_offset;
1498 	orignpages = *ap->a_count;
1499 	pgs = ap->a_m;
1500 
1501 	if (ap->a_flags & PGO_LOCKED) {
1502 		uvn_findpages(uobj, origoffset, ap->a_count, ap->a_m,
1503 		    UFP_NOWAIT|UFP_NOALLOC| (memwrite ? UFP_NORDONLY : 0));
1504 
1505 		error = ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0;
1506 		if (error == 0 && memwrite) {
1507 			genfs_markdirty(vp);
1508 		}
1509 		return error;
1510 	}
1511 	if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= vp->v_size) {
1512 		mutex_exit(&uobj->vmobjlock);
1513 		return EINVAL;
1514 	}
1515 	if ((ap->a_flags & PGO_SYNCIO) == 0) {
1516 		mutex_exit(&uobj->vmobjlock);
1517 		return 0;
1518 	}
1519 	npages = orignpages;
1520 	uvn_findpages(uobj, origoffset, &npages, pgs, UFP_ALL);
1521 	mutex_exit(&uobj->vmobjlock);
1522 	kva = uvm_pagermapin(pgs, npages,
1523 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
1524 	for (i = 0; i < npages; i++) {
1525 		pg = pgs[i];
1526 		if ((pg->flags & PG_FAKE) == 0) {
1527 			continue;
1528 		}
1529 		iov.iov_base = (char *)kva + (i << PAGE_SHIFT);
1530 		iov.iov_len = PAGE_SIZE;
1531 		uio.uio_iov = &iov;
1532 		uio.uio_iovcnt = 1;
1533 		uio.uio_offset = origoffset + (i << PAGE_SHIFT);
1534 		uio.uio_rw = UIO_READ;
1535 		uio.uio_resid = PAGE_SIZE;
1536 		UIO_SETUP_SYSSPACE(&uio);
1537 		/* XXX vn_lock */
1538 		error = VOP_READ(vp, &uio, 0, cred);
1539 		if (error) {
1540 			break;
1541 		}
1542 		if (uio.uio_resid) {
1543 			memset(iov.iov_base, 0, uio.uio_resid);
1544 		}
1545 	}
1546 	uvm_pagermapout(kva, npages);
1547 	mutex_enter(&uobj->vmobjlock);
1548 	mutex_enter(&uvm_pageqlock);
1549 	for (i = 0; i < npages; i++) {
1550 		pg = pgs[i];
1551 		if (error && (pg->flags & PG_FAKE) != 0) {
1552 			pg->flags |= PG_RELEASED;
1553 		} else {
1554 			pmap_clear_modify(pg);
1555 			uvm_pageactivate(pg);
1556 		}
1557 	}
1558 	if (error) {
1559 		uvm_page_unbusy(pgs, npages);
1560 	}
1561 	mutex_exit(&uvm_pageqlock);
1562 	if (error == 0 && memwrite) {
1563 		genfs_markdirty(vp);
1564 	}
1565 	mutex_exit(&uobj->vmobjlock);
1566 	return error;
1567 }
1568 
1569 int
1570 genfs_compat_gop_write(struct vnode *vp, struct vm_page **pgs, int npages,
1571     int flags)
1572 {
1573 	off_t offset;
1574 	struct iovec iov;
1575 	struct uio uio;
1576 	kauth_cred_t cred = curlwp->l_cred;
1577 	struct buf *bp;
1578 	vaddr_t kva;
1579 	int error;
1580 
1581 	offset = pgs[0]->offset;
1582 	kva = uvm_pagermapin(pgs, npages,
1583 	    UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK);
1584 
1585 	iov.iov_base = (void *)kva;
1586 	iov.iov_len = npages << PAGE_SHIFT;
1587 	uio.uio_iov = &iov;
1588 	uio.uio_iovcnt = 1;
1589 	uio.uio_offset = offset;
1590 	uio.uio_rw = UIO_WRITE;
1591 	uio.uio_resid = npages << PAGE_SHIFT;
1592 	UIO_SETUP_SYSSPACE(&uio);
1593 	/* XXX vn_lock */
1594 	error = VOP_WRITE(vp, &uio, 0, cred);
1595 
1596 	mutex_enter(&vp->v_interlock);
1597 	vp->v_numoutput++;
1598 	mutex_exit(&vp->v_interlock);
1599 
1600 	bp = getiobuf(vp, true);
1601 	bp->b_cflags = BC_BUSY | BC_AGE;
1602 	bp->b_lblkno = offset >> vp->v_mount->mnt_fs_bshift;
1603 	bp->b_data = (char *)kva;
1604 	bp->b_bcount = npages << PAGE_SHIFT;
1605 	bp->b_bufsize = npages << PAGE_SHIFT;
1606 	bp->b_resid = 0;
1607 	bp->b_error = error;
1608 	uvm_aio_aiodone(bp);
1609 	return (error);
1610 }
1611 
1612 /*
1613  * Process a uio using direct I/O.  If we reach a part of the request
1614  * which cannot be processed in this fashion for some reason, just return.
1615  * The caller must handle some additional part of the request using
1616  * buffered I/O before trying direct I/O again.
1617  */
1618 
1619 void
1620 genfs_directio(struct vnode *vp, struct uio *uio, int ioflag)
1621 {
1622 	struct vmspace *vs;
1623 	struct iovec *iov;
1624 	vaddr_t va;
1625 	size_t len;
1626 	const int mask = DEV_BSIZE - 1;
1627 	int error;
1628 	bool need_wapbl = (vp->v_mount && vp->v_mount->mnt_wapbl &&
1629 	    (ioflag & IO_JOURNALLOCKED) == 0);
1630 
1631 	/*
1632 	 * We only support direct I/O to user space for now.
1633 	 */
1634 
1635 	if (VMSPACE_IS_KERNEL_P(uio->uio_vmspace)) {
1636 		return;
1637 	}
1638 
1639 	/*
1640 	 * If the vnode is mapped, we would need to get the getpages lock
1641 	 * to stabilize the bmap, but then we would get into trouble whil e
1642 	 * locking the pages if the pages belong to this same vnode (or a
1643 	 * multi-vnode cascade to the same effect).  Just fall back to
1644 	 * buffered I/O if the vnode is mapped to avoid this mess.
1645 	 */
1646 
1647 	if (vp->v_vflag & VV_MAPPED) {
1648 		return;
1649 	}
1650 
1651 	if (need_wapbl) {
1652 		error = WAPBL_BEGIN(vp->v_mount);
1653 		if (error)
1654 			return;
1655 	}
1656 
1657 	/*
1658 	 * Do as much of the uio as possible with direct I/O.
1659 	 */
1660 
1661 	vs = uio->uio_vmspace;
1662 	while (uio->uio_resid) {
1663 		iov = uio->uio_iov;
1664 		if (iov->iov_len == 0) {
1665 			uio->uio_iov++;
1666 			uio->uio_iovcnt--;
1667 			continue;
1668 		}
1669 		va = (vaddr_t)iov->iov_base;
1670 		len = MIN(iov->iov_len, genfs_maxdio);
1671 		len &= ~mask;
1672 
1673 		/*
1674 		 * If the next chunk is smaller than DEV_BSIZE or extends past
1675 		 * the current EOF, then fall back to buffered I/O.
1676 		 */
1677 
1678 		if (len == 0 || uio->uio_offset + len > vp->v_size) {
1679 			break;
1680 		}
1681 
1682 		/*
1683 		 * Check alignment.  The file offset must be at least
1684 		 * sector-aligned.  The exact constraint on memory alignment
1685 		 * is very hardware-dependent, but requiring sector-aligned
1686 		 * addresses there too is safe.
1687 		 */
1688 
1689 		if (uio->uio_offset & mask || va & mask) {
1690 			break;
1691 		}
1692 		error = genfs_do_directio(vs, va, len, vp, uio->uio_offset,
1693 					  uio->uio_rw);
1694 		if (error) {
1695 			break;
1696 		}
1697 		iov->iov_base = (char *)iov->iov_base + len;
1698 		iov->iov_len -= len;
1699 		uio->uio_offset += len;
1700 		uio->uio_resid -= len;
1701 	}
1702 
1703 	if (need_wapbl)
1704 		WAPBL_END(vp->v_mount);
1705 }
1706 
1707 /*
1708  * Iodone routine for direct I/O.  We don't do much here since the request is
1709  * always synchronous, so the caller will do most of the work after biowait().
1710  */
1711 
1712 static void
1713 genfs_dio_iodone(struct buf *bp)
1714 {
1715 
1716 	KASSERT((bp->b_flags & B_ASYNC) == 0);
1717 	if ((bp->b_flags & B_READ) == 0 && (bp->b_cflags & BC_AGE) != 0) {
1718 		mutex_enter(bp->b_objlock);
1719 		vwakeup(bp);
1720 		mutex_exit(bp->b_objlock);
1721 	}
1722 	putiobuf(bp);
1723 }
1724 
1725 /*
1726  * Process one chunk of a direct I/O request.
1727  */
1728 
1729 static int
1730 genfs_do_directio(struct vmspace *vs, vaddr_t uva, size_t len, struct vnode *vp,
1731     off_t off, enum uio_rw rw)
1732 {
1733 	struct vm_map *map;
1734 	struct pmap *upm, *kpm;
1735 	size_t klen = round_page(uva + len) - trunc_page(uva);
1736 	off_t spoff, epoff;
1737 	vaddr_t kva, puva;
1738 	paddr_t pa;
1739 	vm_prot_t prot;
1740 	int error, rv, poff, koff;
1741 	const int pgoflags = PGO_CLEANIT | PGO_SYNCIO | PGO_JOURNALLOCKED |
1742 		(rw == UIO_WRITE ? PGO_FREE : 0);
1743 
1744 	/*
1745 	 * For writes, verify that this range of the file already has fully
1746 	 * allocated backing store.  If there are any holes, just punt and
1747 	 * make the caller take the buffered write path.
1748 	 */
1749 
1750 	if (rw == UIO_WRITE) {
1751 		daddr_t lbn, elbn, blkno;
1752 		int bsize, bshift, run;
1753 
1754 		bshift = vp->v_mount->mnt_fs_bshift;
1755 		bsize = 1 << bshift;
1756 		lbn = off >> bshift;
1757 		elbn = (off + len + bsize - 1) >> bshift;
1758 		while (lbn < elbn) {
1759 			error = VOP_BMAP(vp, lbn, NULL, &blkno, &run);
1760 			if (error) {
1761 				return error;
1762 			}
1763 			if (blkno == (daddr_t)-1) {
1764 				return ENOSPC;
1765 			}
1766 			lbn += 1 + run;
1767 		}
1768 	}
1769 
1770 	/*
1771 	 * Flush any cached pages for parts of the file that we're about to
1772 	 * access.  If we're writing, invalidate pages as well.
1773 	 */
1774 
1775 	spoff = trunc_page(off);
1776 	epoff = round_page(off + len);
1777 	mutex_enter(&vp->v_interlock);
1778 	error = VOP_PUTPAGES(vp, spoff, epoff, pgoflags);
1779 	if (error) {
1780 		return error;
1781 	}
1782 
1783 	/*
1784 	 * Wire the user pages and remap them into kernel memory.
1785 	 */
1786 
1787 	prot = rw == UIO_READ ? VM_PROT_READ | VM_PROT_WRITE : VM_PROT_READ;
1788 	error = uvm_vslock(vs, (void *)uva, len, prot);
1789 	if (error) {
1790 		return error;
1791 	}
1792 
1793 	map = &vs->vm_map;
1794 	upm = vm_map_pmap(map);
1795 	kpm = vm_map_pmap(kernel_map);
1796 	kva = uvm_km_alloc(kernel_map, klen, 0,
1797 			   UVM_KMF_VAONLY | UVM_KMF_WAITVA);
1798 	puva = trunc_page(uva);
1799 	for (poff = 0; poff < klen; poff += PAGE_SIZE) {
1800 		rv = pmap_extract(upm, puva + poff, &pa);
1801 		KASSERT(rv);
1802 		pmap_enter(kpm, kva + poff, pa, prot, prot | PMAP_WIRED);
1803 	}
1804 	pmap_update(kpm);
1805 
1806 	/*
1807 	 * Do the I/O.
1808 	 */
1809 
1810 	koff = uva - trunc_page(uva);
1811 	error = genfs_do_io(vp, off, kva + koff, len, PGO_SYNCIO, rw,
1812 			    genfs_dio_iodone);
1813 
1814 	/*
1815 	 * Tear down the kernel mapping.
1816 	 */
1817 
1818 	pmap_remove(kpm, kva, kva + klen);
1819 	pmap_update(kpm);
1820 	uvm_km_free(kernel_map, kva, klen, UVM_KMF_VAONLY);
1821 
1822 	/*
1823 	 * Unwire the user pages.
1824 	 */
1825 
1826 	uvm_vsunlock(vs, (void *)uva, len);
1827 	return error;
1828 }
1829 
1830