xref: /netbsd-src/sys/miscfs/genfs/genfs_io.c (revision 46f5119e40af2e51998f686b2fdcc76b5488f7f3)
1 /*	$NetBSD: genfs_io.c,v 1.48 2011/04/21 06:27:17 matt 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 1.48 2011/04/21 06:27:17 matt 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 #include <miscfs/syncfs/syncfs.h>
56 
57 #include <uvm/uvm.h>
58 #include <uvm/uvm_pager.h>
59 
60 static int genfs_do_directio(struct vmspace *, vaddr_t, size_t, struct vnode *,
61     off_t, enum uio_rw);
62 static void genfs_dio_iodone(struct buf *);
63 
64 static int genfs_do_io(struct vnode *, off_t, vaddr_t, size_t, int, enum uio_rw,
65     void (*)(struct buf *));
66 static void genfs_rel_pages(struct vm_page **, int);
67 static void genfs_markdirty(struct vnode *);
68 
69 int genfs_maxdio = MAXPHYS;
70 
71 static void
72 genfs_rel_pages(struct vm_page **pgs, int npages)
73 {
74 	int i;
75 
76 	for (i = 0; i < npages; i++) {
77 		struct vm_page *pg = pgs[i];
78 
79 		if (pg == NULL || pg == PGO_DONTCARE)
80 			continue;
81 		if (pg->flags & PG_FAKE) {
82 			pg->flags |= PG_RELEASED;
83 		}
84 	}
85 	mutex_enter(&uvm_pageqlock);
86 	uvm_page_unbusy(pgs, npages);
87 	mutex_exit(&uvm_pageqlock);
88 }
89 
90 static void
91 genfs_markdirty(struct vnode *vp)
92 {
93 	struct genfs_node * const gp = VTOG(vp);
94 
95 	KASSERT(mutex_owned(&vp->v_interlock));
96 	gp->g_dirtygen++;
97 	if ((vp->v_iflag & VI_ONWORKLST) == 0) {
98 		vn_syncer_add_to_worklist(vp, filedelay);
99 	}
100 	if ((vp->v_iflag & (VI_WRMAP|VI_WRMAPDIRTY)) == VI_WRMAP) {
101 		vp->v_iflag |= VI_WRMAPDIRTY;
102 	}
103 }
104 
105 /*
106  * generic VM getpages routine.
107  * Return PG_BUSY pages for the given range,
108  * reading from backing store if necessary.
109  */
110 
111 int
112 genfs_getpages(void *v)
113 {
114 	struct vop_getpages_args /* {
115 		struct vnode *a_vp;
116 		voff_t a_offset;
117 		struct vm_page **a_m;
118 		int *a_count;
119 		int a_centeridx;
120 		vm_prot_t a_access_type;
121 		int a_advice;
122 		int a_flags;
123 	} */ * const ap = v;
124 
125 	off_t diskeof, memeof;
126 	int i, error, npages;
127 	const int flags = ap->a_flags;
128 	struct vnode * const vp = ap->a_vp;
129 	struct uvm_object * const uobj = &vp->v_uobj;
130 	kauth_cred_t const cred = curlwp->l_cred;		/* XXXUBC curlwp */
131 	const bool async = (flags & PGO_SYNCIO) == 0;
132 	const bool memwrite = (ap->a_access_type & VM_PROT_WRITE) != 0;
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 	const bool need_wapbl = blockalloc && vp->v_mount->mnt_wapbl;
137 	bool has_trans_wapbl = false;
138 	UVMHIST_FUNC("genfs_getpages"); UVMHIST_CALLED(ubchist);
139 
140 	UVMHIST_LOG(ubchist, "vp %p off 0x%x/%x count %d",
141 	    vp, ap->a_offset >> 32, ap->a_offset, *ap->a_count);
142 
143 	KASSERT(vp->v_type == VREG || vp->v_type == VDIR ||
144 	    vp->v_type == VLNK || vp->v_type == VBLK);
145 
146 startover:
147 	error = 0;
148 	const voff_t origvsize = vp->v_size;
149 	const off_t origoffset = ap->a_offset;
150 	const int orignpages = *ap->a_count;
151 
152 	GOP_SIZE(vp, origvsize, &diskeof, 0);
153 	if (flags & PGO_PASTEOF) {
154 		off_t newsize;
155 #if defined(DIAGNOSTIC)
156 		off_t writeeof;
157 #endif /* defined(DIAGNOSTIC) */
158 
159 		newsize = MAX(origvsize,
160 		    origoffset + (orignpages << PAGE_SHIFT));
161 		GOP_SIZE(vp, newsize, &memeof, GOP_SIZE_MEM);
162 #if defined(DIAGNOSTIC)
163 		GOP_SIZE(vp, vp->v_writesize, &writeeof, GOP_SIZE_MEM);
164 		if (newsize > round_page(writeeof)) {
165 			panic("%s: past eof: %" PRId64 " vs. %" PRId64,
166 			    __func__, newsize, round_page(writeeof));
167 		}
168 #endif /* defined(DIAGNOSTIC) */
169 	} else {
170 		GOP_SIZE(vp, origvsize, &memeof, GOP_SIZE_MEM);
171 	}
172 	KASSERT(ap->a_centeridx >= 0 || ap->a_centeridx <= orignpages);
173 	KASSERT((origoffset & (PAGE_SIZE - 1)) == 0 && origoffset >= 0);
174 	KASSERT(orignpages > 0);
175 
176 	/*
177 	 * Bounds-check the request.
178 	 */
179 
180 	if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= memeof) {
181 		if ((flags & PGO_LOCKED) == 0) {
182 			mutex_exit(&uobj->vmobjlock);
183 		}
184 		UVMHIST_LOG(ubchist, "off 0x%x count %d goes past EOF 0x%x",
185 		    origoffset, *ap->a_count, memeof,0);
186 		error = EINVAL;
187 		goto out_err;
188 	}
189 
190 	/* uobj is locked */
191 
192 	if ((flags & PGO_NOTIMESTAMP) == 0 &&
193 	    (vp->v_type != VBLK ||
194 	    (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) {
195 		int updflags = 0;
196 
197 		if ((vp->v_mount->mnt_flag & MNT_NOATIME) == 0) {
198 			updflags = GOP_UPDATE_ACCESSED;
199 		}
200 		if (memwrite) {
201 			updflags |= GOP_UPDATE_MODIFIED;
202 		}
203 		if (updflags != 0) {
204 			GOP_MARKUPDATE(vp, updflags);
205 		}
206 	}
207 
208 	/*
209 	 * For PGO_LOCKED requests, just return whatever's in memory.
210 	 */
211 
212 	if (flags & PGO_LOCKED) {
213 		int nfound;
214 		struct vm_page *pg;
215 
216 		KASSERT(!glocked);
217 		npages = *ap->a_count;
218 #if defined(DEBUG)
219 		for (i = 0; i < npages; i++) {
220 			pg = ap->a_m[i];
221 			KASSERT(pg == NULL || pg == PGO_DONTCARE);
222 		}
223 #endif /* defined(DEBUG) */
224 		nfound = uvn_findpages(uobj, origoffset, &npages,
225 		    ap->a_m, UFP_NOWAIT|UFP_NOALLOC|(memwrite ? UFP_NORDONLY : 0));
226 		KASSERT(npages == *ap->a_count);
227 		if (nfound == 0) {
228 			error = EBUSY;
229 			goto out_err;
230 		}
231 		if (!genfs_node_rdtrylock(vp)) {
232 			genfs_rel_pages(ap->a_m, npages);
233 
234 			/*
235 			 * restore the array.
236 			 */
237 
238 			for (i = 0; i < npages; i++) {
239 				pg = ap->a_m[i];
240 
241 				if (pg != NULL && pg != PGO_DONTCARE) {
242 					ap->a_m[i] = NULL;
243 				}
244 				KASSERT(ap->a_m[i] == NULL ||
245 				    ap->a_m[i] == PGO_DONTCARE);
246 			}
247 		} else {
248 			genfs_node_unlock(vp);
249 		}
250 		error = (ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0);
251 		if (error == 0 && memwrite) {
252 			genfs_markdirty(vp);
253 		}
254 		goto out_err;
255 	}
256 	mutex_exit(&uobj->vmobjlock);
257 
258 	/*
259 	 * find the requested pages and make some simple checks.
260 	 * leave space in the page array for a whole block.
261 	 */
262 
263 	const int fs_bshift = (vp->v_type != VBLK) ?
264 	    vp->v_mount->mnt_fs_bshift : DEV_BSHIFT;
265 	const int dev_bshift = (vp->v_type != VBLK) ?
266 	    vp->v_mount->mnt_dev_bshift : DEV_BSHIFT;
267 	const int fs_bsize = 1 << fs_bshift;
268 #define	blk_mask	(fs_bsize - 1)
269 #define	trunc_blk(x)	((x) & ~blk_mask)
270 #define	round_blk(x)	(((x) + blk_mask) & ~blk_mask)
271 
272 	const int orignmempages = MIN(orignpages,
273 	    round_page(memeof - origoffset) >> PAGE_SHIFT);
274 	npages = orignmempages;
275 	const off_t startoffset = trunc_blk(origoffset);
276 	const off_t endoffset = MIN(
277 	    round_page(round_blk(origoffset + (npages << PAGE_SHIFT))),
278 	    round_page(memeof));
279 	const int ridx = (origoffset - startoffset) >> PAGE_SHIFT;
280 
281 	const int pgs_size = sizeof(struct vm_page *) *
282 	    ((endoffset - startoffset) >> PAGE_SHIFT);
283 	struct vm_page **pgs, *pgs_onstack[UBC_MAX_PAGES];
284 
285 	if (pgs_size > sizeof(pgs_onstack)) {
286 		pgs = kmem_zalloc(pgs_size, async ? KM_NOSLEEP : KM_SLEEP);
287 		if (pgs == NULL) {
288 			pgs = pgs_onstack;
289 			error = ENOMEM;
290 			goto out_err;
291 		}
292 	} else {
293 		pgs = pgs_onstack;
294 		(void)memset(pgs, 0, pgs_size);
295 	}
296 
297 	UVMHIST_LOG(ubchist, "ridx %d npages %d startoff %ld endoff %ld",
298 	    ridx, npages, startoffset, endoffset);
299 
300 	if (!has_trans_wapbl) {
301 		fstrans_start(vp->v_mount, FSTRANS_SHARED);
302 		/*
303 		 * XXX: This assumes that we come here only via
304 		 * the mmio path
305 		 */
306 		if (need_wapbl) {
307 			error = WAPBL_BEGIN(vp->v_mount);
308 			if (error) {
309 				fstrans_done(vp->v_mount);
310 				goto out_err_free;
311 			}
312 		}
313 		has_trans_wapbl = true;
314 	}
315 
316 	/*
317 	 * hold g_glock to prevent a race with truncate.
318 	 *
319 	 * check if our idea of v_size is still valid.
320 	 */
321 
322 	KASSERT(!glocked || genfs_node_wrlocked(vp));
323 	if (!glocked) {
324 		if (blockalloc) {
325 			genfs_node_wrlock(vp);
326 		} else {
327 			genfs_node_rdlock(vp);
328 		}
329 	}
330 	mutex_enter(&uobj->vmobjlock);
331 	if (vp->v_size < origvsize) {
332 		if (!glocked) {
333 			genfs_node_unlock(vp);
334 		}
335 		if (pgs != pgs_onstack)
336 			kmem_free(pgs, pgs_size);
337 		goto startover;
338 	}
339 
340 	if (uvn_findpages(uobj, origoffset, &npages, &pgs[ridx],
341 	    async ? UFP_NOWAIT : UFP_ALL) != orignmempages) {
342 		if (!glocked) {
343 			genfs_node_unlock(vp);
344 		}
345 		KASSERT(async != 0);
346 		genfs_rel_pages(&pgs[ridx], orignmempages);
347 		mutex_exit(&uobj->vmobjlock);
348 		error = EBUSY;
349 		goto out_err_free;
350 	}
351 
352 	/*
353 	 * if the pages are already resident, just return them.
354 	 */
355 
356 	for (i = 0; i < npages; i++) {
357 		struct vm_page *pg = pgs[ridx + i];
358 
359 		if ((pg->flags & PG_FAKE) ||
360 		    (blockalloc && (pg->flags & PG_RDONLY))) {
361 			break;
362 		}
363 	}
364 	if (i == npages) {
365 		if (!glocked) {
366 			genfs_node_unlock(vp);
367 		}
368 		UVMHIST_LOG(ubchist, "returning cached pages", 0,0,0,0);
369 		npages += ridx;
370 		goto out;
371 	}
372 
373 	/*
374 	 * if PGO_OVERWRITE is set, don't bother reading the pages.
375 	 */
376 
377 	if (overwrite) {
378 		if (!glocked) {
379 			genfs_node_unlock(vp);
380 		}
381 		UVMHIST_LOG(ubchist, "PGO_OVERWRITE",0,0,0,0);
382 
383 		for (i = 0; i < npages; i++) {
384 			struct vm_page *pg = pgs[ridx + i];
385 
386 			pg->flags &= ~(PG_RDONLY|PG_CLEAN);
387 		}
388 		npages += ridx;
389 		goto out;
390 	}
391 
392 	/*
393 	 * the page wasn't resident and we're not overwriting,
394 	 * so we're going to have to do some i/o.
395 	 * find any additional pages needed to cover the expanded range.
396 	 */
397 
398 	npages = (endoffset - startoffset) >> PAGE_SHIFT;
399 	if (startoffset != origoffset || npages != orignmempages) {
400 		int npgs;
401 
402 		/*
403 		 * we need to avoid deadlocks caused by locking
404 		 * additional pages at lower offsets than pages we
405 		 * already have locked.  unlock them all and start over.
406 		 */
407 
408 		genfs_rel_pages(&pgs[ridx], orignmempages);
409 		memset(pgs, 0, pgs_size);
410 
411 		UVMHIST_LOG(ubchist, "reset npages start 0x%x end 0x%x",
412 		    startoffset, endoffset, 0,0);
413 		npgs = npages;
414 		if (uvn_findpages(uobj, startoffset, &npgs, pgs,
415 		    async ? UFP_NOWAIT : UFP_ALL) != npages) {
416 			if (!glocked) {
417 				genfs_node_unlock(vp);
418 			}
419 			KASSERT(async != 0);
420 			genfs_rel_pages(pgs, npages);
421 			mutex_exit(&uobj->vmobjlock);
422 			error = EBUSY;
423 			goto out_err_free;
424 		}
425 	}
426 
427 	mutex_exit(&uobj->vmobjlock);
428 
429     {
430 	size_t bytes, iobytes, tailstart, tailbytes, totalbytes, skipbytes;
431 	vaddr_t kva;
432 	struct buf *bp, *mbp;
433 	bool sawhole = false;
434 
435 	/*
436 	 * read the desired page(s).
437 	 */
438 
439 	totalbytes = npages << PAGE_SHIFT;
440 	bytes = MIN(totalbytes, MAX(diskeof - startoffset, 0));
441 	tailbytes = totalbytes - bytes;
442 	skipbytes = 0;
443 
444 	kva = uvm_pagermapin(pgs, npages,
445 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
446 
447 	mbp = getiobuf(vp, true);
448 	mbp->b_bufsize = totalbytes;
449 	mbp->b_data = (void *)kva;
450 	mbp->b_resid = mbp->b_bcount = bytes;
451 	mbp->b_cflags = BC_BUSY;
452 	if (async) {
453 		mbp->b_flags = B_READ | B_ASYNC;
454 		mbp->b_iodone = uvm_aio_biodone;
455 	} else {
456 		mbp->b_flags = B_READ;
457 		mbp->b_iodone = NULL;
458 	}
459 	if (async)
460 		BIO_SETPRIO(mbp, BPRIO_TIMELIMITED);
461 	else
462 		BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL);
463 
464 	/*
465 	 * if EOF is in the middle of the range, zero the part past EOF.
466 	 * skip over pages which are not PG_FAKE since in that case they have
467 	 * valid data that we need to preserve.
468 	 */
469 
470 	tailstart = bytes;
471 	while (tailbytes > 0) {
472 		const int len = PAGE_SIZE - (tailstart & PAGE_MASK);
473 
474 		KASSERT(len <= tailbytes);
475 		if ((pgs[tailstart >> PAGE_SHIFT]->flags & PG_FAKE) != 0) {
476 			memset((void *)(kva + tailstart), 0, len);
477 			UVMHIST_LOG(ubchist, "tailbytes %p 0x%x 0x%x",
478 			    kva, tailstart, len, 0);
479 		}
480 		tailstart += len;
481 		tailbytes -= len;
482 	}
483 
484 	/*
485 	 * now loop over the pages, reading as needed.
486 	 */
487 
488 	bp = NULL;
489 	off_t offset;
490 	for (offset = startoffset;
491 	    bytes > 0;
492 	    offset += iobytes, bytes -= iobytes) {
493 		int run;
494 		daddr_t lbn, blkno;
495 		int pidx;
496 		struct vnode *devvp;
497 
498 		/*
499 		 * skip pages which don't need to be read.
500 		 */
501 
502 		pidx = (offset - startoffset) >> PAGE_SHIFT;
503 		while ((pgs[pidx]->flags & PG_FAKE) == 0) {
504 			size_t b;
505 
506 			KASSERT((offset & (PAGE_SIZE - 1)) == 0);
507 			if ((pgs[pidx]->flags & PG_RDONLY)) {
508 				sawhole = true;
509 			}
510 			b = MIN(PAGE_SIZE, bytes);
511 			offset += b;
512 			bytes -= b;
513 			skipbytes += b;
514 			pidx++;
515 			UVMHIST_LOG(ubchist, "skipping, new offset 0x%x",
516 			    offset, 0,0,0);
517 			if (bytes == 0) {
518 				goto loopdone;
519 			}
520 		}
521 
522 		/*
523 		 * bmap the file to find out the blkno to read from and
524 		 * how much we can read in one i/o.  if bmap returns an error,
525 		 * skip the rest of the top-level i/o.
526 		 */
527 
528 		lbn = offset >> fs_bshift;
529 		error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run);
530 		if (error) {
531 			UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%x -> %d\n",
532 			    lbn,error,0,0);
533 			skipbytes += bytes;
534 			bytes = 0;
535 			goto loopdone;
536 		}
537 
538 		/*
539 		 * see how many pages can be read with this i/o.
540 		 * reduce the i/o size if necessary to avoid
541 		 * overwriting pages with valid data.
542 		 */
543 
544 		iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset,
545 		    bytes);
546 		if (offset + iobytes > round_page(offset)) {
547 			int pcount;
548 
549 			pcount = 1;
550 			while (pidx + pcount < npages &&
551 			    pgs[pidx + pcount]->flags & PG_FAKE) {
552 				pcount++;
553 			}
554 			iobytes = MIN(iobytes, (pcount << PAGE_SHIFT) -
555 			    (offset - trunc_page(offset)));
556 		}
557 
558 		/*
559 		 * if this block isn't allocated, zero it instead of
560 		 * reading it.  unless we are going to allocate blocks,
561 		 * mark the pages we zeroed PG_RDONLY.
562 		 */
563 
564 		if (blkno == (daddr_t)-1) {
565 			int holepages = (round_page(offset + iobytes) -
566 			    trunc_page(offset)) >> PAGE_SHIFT;
567 			UVMHIST_LOG(ubchist, "lbn 0x%x -> HOLE", lbn,0,0,0);
568 
569 			sawhole = true;
570 			memset((char *)kva + (offset - startoffset), 0,
571 			    iobytes);
572 			skipbytes += iobytes;
573 
574 			mutex_enter(&uobj->vmobjlock);
575 			for (i = 0; i < holepages; i++) {
576 				if (memwrite) {
577 					pgs[pidx + i]->flags &= ~PG_CLEAN;
578 				}
579 				if (!blockalloc) {
580 					pgs[pidx + i]->flags |= PG_RDONLY;
581 				}
582 			}
583 			mutex_exit(&uobj->vmobjlock);
584 			continue;
585 		}
586 
587 		/*
588 		 * allocate a sub-buf for this piece of the i/o
589 		 * (or just use mbp if there's only 1 piece),
590 		 * and start it going.
591 		 */
592 
593 		if (offset == startoffset && iobytes == bytes) {
594 			bp = mbp;
595 		} else {
596 			UVMHIST_LOG(ubchist, "vp %p bp %p num now %d",
597 			    vp, bp, vp->v_numoutput, 0);
598 			bp = getiobuf(vp, true);
599 			nestiobuf_setup(mbp, bp, offset - startoffset, iobytes);
600 		}
601 		bp->b_lblkno = 0;
602 
603 		/* adjust physical blkno for partial blocks */
604 		bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >>
605 		    dev_bshift);
606 
607 		UVMHIST_LOG(ubchist,
608 		    "bp %p offset 0x%x bcount 0x%x blkno 0x%x",
609 		    bp, offset, bp->b_bcount, bp->b_blkno);
610 
611 		VOP_STRATEGY(devvp, bp);
612 	}
613 
614 loopdone:
615 	nestiobuf_done(mbp, skipbytes, error);
616 	if (async) {
617 		UVMHIST_LOG(ubchist, "returning 0 (async)",0,0,0,0);
618 		if (!glocked) {
619 			genfs_node_unlock(vp);
620 		}
621 		error = 0;
622 		goto out_err_free;
623 	}
624 	if (bp != NULL) {
625 		error = biowait(mbp);
626 	}
627 
628 	/* Remove the mapping (make KVA available as soon as possible) */
629 	uvm_pagermapout(kva, npages);
630 
631 	/*
632 	 * if this we encountered a hole then we have to do a little more work.
633 	 * for read faults, we marked the page PG_RDONLY so that future
634 	 * write accesses to the page will fault again.
635 	 * for write faults, we must make sure that the backing store for
636 	 * the page is completely allocated while the pages are locked.
637 	 */
638 
639 	if (!error && sawhole && blockalloc) {
640 		error = GOP_ALLOC(vp, startoffset,
641 		    npages << PAGE_SHIFT, 0, cred);
642 		UVMHIST_LOG(ubchist, "gop_alloc off 0x%x/0x%x -> %d",
643 		    startoffset, npages << PAGE_SHIFT, error,0);
644 		if (!error) {
645 			mutex_enter(&uobj->vmobjlock);
646 			for (i = 0; i < npages; i++) {
647 				struct vm_page *pg = pgs[i];
648 
649 				if (pg == NULL) {
650 					continue;
651 				}
652 				pg->flags &= ~(PG_CLEAN|PG_RDONLY);
653 				UVMHIST_LOG(ubchist, "mark dirty pg %p",
654 				    pg,0,0,0);
655 			}
656 			mutex_exit(&uobj->vmobjlock);
657 		}
658 	}
659 	if (!glocked) {
660 		genfs_node_unlock(vp);
661 	}
662 
663 	putiobuf(mbp);
664     }
665 
666 	mutex_enter(&uobj->vmobjlock);
667 
668 	/*
669 	 * we're almost done!  release the pages...
670 	 * for errors, we free the pages.
671 	 * otherwise we activate them and mark them as valid and clean.
672 	 * also, unbusy pages that were not actually requested.
673 	 */
674 
675 	if (error) {
676 		for (i = 0; i < npages; i++) {
677 			struct vm_page *pg = pgs[i];
678 
679 			if (pg == NULL) {
680 				continue;
681 			}
682 			UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x",
683 			    pg, pg->flags, 0,0);
684 			if (pg->flags & PG_FAKE) {
685 				pg->flags |= PG_RELEASED;
686 			}
687 		}
688 		mutex_enter(&uvm_pageqlock);
689 		uvm_page_unbusy(pgs, npages);
690 		mutex_exit(&uvm_pageqlock);
691 		mutex_exit(&uobj->vmobjlock);
692 		UVMHIST_LOG(ubchist, "returning error %d", error,0,0,0);
693 		goto out_err_free;
694 	}
695 
696 out:
697 	UVMHIST_LOG(ubchist, "succeeding, npages %d", npages,0,0,0);
698 	error = 0;
699 	mutex_enter(&uvm_pageqlock);
700 	for (i = 0; i < npages; i++) {
701 		struct vm_page *pg = pgs[i];
702 		if (pg == NULL) {
703 			continue;
704 		}
705 		UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x",
706 		    pg, pg->flags, 0,0);
707 		if (pg->flags & PG_FAKE && !overwrite) {
708 			pg->flags &= ~(PG_FAKE);
709 			pmap_clear_modify(pgs[i]);
710 		}
711 		KASSERT(!memwrite || !blockalloc || (pg->flags & PG_RDONLY) == 0);
712 		if (i < ridx || i >= ridx + orignmempages || async) {
713 			UVMHIST_LOG(ubchist, "unbusy pg %p offset 0x%x",
714 			    pg, pg->offset,0,0);
715 			if (pg->flags & PG_WANTED) {
716 				wakeup(pg);
717 			}
718 			if (pg->flags & PG_FAKE) {
719 				KASSERT(overwrite);
720 				uvm_pagezero(pg);
721 			}
722 			if (pg->flags & PG_RELEASED) {
723 				uvm_pagefree(pg);
724 				continue;
725 			}
726 			uvm_pageenqueue(pg);
727 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
728 			UVM_PAGE_OWN(pg, NULL);
729 		}
730 	}
731 	mutex_exit(&uvm_pageqlock);
732 	if (memwrite) {
733 		genfs_markdirty(vp);
734 	}
735 	mutex_exit(&uobj->vmobjlock);
736 	if (ap->a_m != NULL) {
737 		memcpy(ap->a_m, &pgs[ridx],
738 		    orignmempages * sizeof(struct vm_page *));
739 	}
740 
741 out_err_free:
742 	if (pgs != NULL && pgs != pgs_onstack)
743 		kmem_free(pgs, pgs_size);
744 out_err:
745 	if (has_trans_wapbl) {
746 		if (need_wapbl)
747 			WAPBL_END(vp->v_mount);
748 		fstrans_done(vp->v_mount);
749 	}
750 	return error;
751 }
752 
753 /*
754  * generic VM putpages routine.
755  * Write the given range of pages to backing store.
756  *
757  * => "offhi == 0" means flush all pages at or after "offlo".
758  * => object should be locked by caller.  we return with the
759  *      object unlocked.
760  * => if PGO_CLEANIT or PGO_SYNCIO is set, we may block (due to I/O).
761  *	thus, a caller might want to unlock higher level resources
762  *	(e.g. vm_map) before calling flush.
763  * => if neither PGO_CLEANIT nor PGO_SYNCIO is set, we will not block
764  * => if PGO_ALLPAGES is set, then all pages in the object will be processed.
765  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
766  *	that new pages are inserted on the tail end of the list.   thus,
767  *	we can make a complete pass through the object in one go by starting
768  *	at the head and working towards the tail (new pages are put in
769  *	front of us).
770  * => NOTE: we are allowed to lock the page queues, so the caller
771  *	must not be holding the page queue lock.
772  *
773  * note on "cleaning" object and PG_BUSY pages:
774  *	this routine is holding the lock on the object.   the only time
775  *	that it can run into a PG_BUSY page that it does not own is if
776  *	some other process has started I/O on the page (e.g. either
777  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
778  *	in, then it can not be dirty (!PG_CLEAN) because no one has
779  *	had a chance to modify it yet.    if the PG_BUSY page is being
780  *	paged out then it means that someone else has already started
781  *	cleaning the page for us (how nice!).    in this case, if we
782  *	have syncio specified, then after we make our pass through the
783  *	object we need to wait for the other PG_BUSY pages to clear
784  *	off (i.e. we need to do an iosync).   also note that once a
785  *	page is PG_BUSY it must stay in its object until it is un-busyed.
786  *
787  * note on page traversal:
788  *	we can traverse the pages in an object either by going down the
789  *	linked list in "uobj->memq", or we can go over the address range
790  *	by page doing hash table lookups for each address.    depending
791  *	on how many pages are in the object it may be cheaper to do one
792  *	or the other.   we set "by_list" to true if we are using memq.
793  *	if the cost of a hash lookup was equal to the cost of the list
794  *	traversal we could compare the number of pages in the start->stop
795  *	range to the total number of pages in the object.   however, it
796  *	seems that a hash table lookup is more expensive than the linked
797  *	list traversal, so we multiply the number of pages in the
798  *	range by an estimate of the relatively higher cost of the hash lookup.
799  */
800 
801 int
802 genfs_putpages(void *v)
803 {
804 	struct vop_putpages_args /* {
805 		struct vnode *a_vp;
806 		voff_t a_offlo;
807 		voff_t a_offhi;
808 		int a_flags;
809 	} */ * const ap = v;
810 
811 	return genfs_do_putpages(ap->a_vp, ap->a_offlo, ap->a_offhi,
812 	    ap->a_flags, NULL);
813 }
814 
815 int
816 genfs_do_putpages(struct vnode *vp, off_t startoff, off_t endoff,
817     int origflags, struct vm_page **busypg)
818 {
819 	struct uvm_object * const uobj = &vp->v_uobj;
820 	kmutex_t * const slock = &uobj->vmobjlock;
821 	off_t off;
822 	/* Even for strange MAXPHYS, the shift rounds down to a page */
823 #define maxpages (MAXPHYS >> PAGE_SHIFT)
824 	int i, error, npages, nback;
825 	int freeflag;
826 	struct vm_page *pgs[maxpages], *pg, *nextpg, *tpg, curmp, endmp;
827 	bool wasclean, by_list, needs_clean, yld;
828 	bool async = (origflags & PGO_SYNCIO) == 0;
829 	bool pagedaemon = curlwp == uvm.pagedaemon_lwp;
830 	struct lwp * const l = curlwp ? curlwp : &lwp0;
831 	struct genfs_node * const gp = VTOG(vp);
832 	int flags;
833 	int dirtygen;
834 	bool modified;
835 	bool need_wapbl;
836 	bool has_trans;
837 	bool cleanall;
838 	bool onworklst;
839 
840 	UVMHIST_FUNC("genfs_putpages"); UVMHIST_CALLED(ubchist);
841 
842 	KASSERT(origflags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE));
843 	KASSERT((startoff & PAGE_MASK) == 0 && (endoff & PAGE_MASK) == 0);
844 	KASSERT(startoff < endoff || endoff == 0);
845 
846 	UVMHIST_LOG(ubchist, "vp %p pages %d off 0x%x len 0x%x",
847 	    vp, uobj->uo_npages, startoff, endoff - startoff);
848 
849 	has_trans = false;
850 	need_wapbl = (!pagedaemon && vp->v_mount && vp->v_mount->mnt_wapbl &&
851 	    (origflags & PGO_JOURNALLOCKED) == 0);
852 
853 retry:
854 	modified = false;
855 	flags = origflags;
856 	KASSERT((vp->v_iflag & VI_ONWORKLST) != 0 ||
857 	    (vp->v_iflag & VI_WRMAPDIRTY) == 0);
858 	if (uobj->uo_npages == 0) {
859 		if (vp->v_iflag & VI_ONWORKLST) {
860 			vp->v_iflag &= ~VI_WRMAPDIRTY;
861 			if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL)
862 				vn_syncer_remove_from_worklist(vp);
863 		}
864 		if (has_trans) {
865 			if (need_wapbl)
866 				WAPBL_END(vp->v_mount);
867 			fstrans_done(vp->v_mount);
868 		}
869 		mutex_exit(slock);
870 		return (0);
871 	}
872 
873 	/*
874 	 * the vnode has pages, set up to process the request.
875 	 */
876 
877 	if (!has_trans && (flags & PGO_CLEANIT) != 0) {
878 		mutex_exit(slock);
879 		if (pagedaemon) {
880 			error = fstrans_start_nowait(vp->v_mount, FSTRANS_LAZY);
881 			if (error)
882 				return error;
883 		} else
884 			fstrans_start(vp->v_mount, FSTRANS_LAZY);
885 		if (need_wapbl) {
886 			error = WAPBL_BEGIN(vp->v_mount);
887 			if (error) {
888 				fstrans_done(vp->v_mount);
889 				return error;
890 			}
891 		}
892 		has_trans = true;
893 		mutex_enter(slock);
894 		goto retry;
895 	}
896 
897 	error = 0;
898 	wasclean = (vp->v_numoutput == 0);
899 	off = startoff;
900 	if (endoff == 0 || flags & PGO_ALLPAGES) {
901 		endoff = trunc_page(LLONG_MAX);
902 	}
903 	by_list = (uobj->uo_npages <=
904 	    ((endoff - startoff) >> PAGE_SHIFT) * UVM_PAGE_TREE_PENALTY);
905 
906 	/*
907 	 * if this vnode is known not to have dirty pages,
908 	 * don't bother to clean it out.
909 	 */
910 
911 	if ((vp->v_iflag & VI_ONWORKLST) == 0) {
912 #if !defined(DEBUG)
913 		if ((flags & (PGO_FREE|PGO_DEACTIVATE)) == 0) {
914 			goto skip_scan;
915 		}
916 #endif /* !defined(DEBUG) */
917 		flags &= ~PGO_CLEANIT;
918 	}
919 
920 	/*
921 	 * start the loop.  when scanning by list, hold the last page
922 	 * in the list before we start.  pages allocated after we start
923 	 * will be added to the end of the list, so we can stop at the
924 	 * current last page.
925 	 */
926 
927 	cleanall = (flags & PGO_CLEANIT) != 0 && wasclean &&
928 	    startoff == 0 && endoff == trunc_page(LLONG_MAX) &&
929 	    (vp->v_iflag & VI_ONWORKLST) != 0;
930 	dirtygen = gp->g_dirtygen;
931 	freeflag = pagedaemon ? PG_PAGEOUT : PG_RELEASED;
932 	if (by_list) {
933 		curmp.flags = PG_MARKER;
934 		endmp.flags = PG_MARKER;
935 		pg = TAILQ_FIRST(&uobj->memq);
936 		TAILQ_INSERT_TAIL(&uobj->memq, &endmp, listq.queue);
937 	} else {
938 		pg = uvm_pagelookup(uobj, off);
939 	}
940 	nextpg = NULL;
941 	while (by_list || off < endoff) {
942 
943 		/*
944 		 * if the current page is not interesting, move on to the next.
945 		 */
946 
947 		KASSERT(pg == NULL || pg->uobject == uobj ||
948 		    (pg->flags & PG_MARKER) != 0);
949 		KASSERT(pg == NULL ||
950 		    (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 ||
951 		    (pg->flags & (PG_BUSY|PG_MARKER)) != 0);
952 		if (by_list) {
953 			if (pg == &endmp) {
954 				break;
955 			}
956 			if (pg->flags & PG_MARKER) {
957 				pg = TAILQ_NEXT(pg, listq.queue);
958 				continue;
959 			}
960 			if (pg->offset < startoff || pg->offset >= endoff ||
961 			    pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
962 				if (pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
963 					wasclean = false;
964 				}
965 				pg = TAILQ_NEXT(pg, listq.queue);
966 				continue;
967 			}
968 			off = pg->offset;
969 		} else if (pg == NULL || pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
970 			if (pg != NULL) {
971 				wasclean = false;
972 			}
973 			off += PAGE_SIZE;
974 			if (off < endoff) {
975 				pg = uvm_pagelookup(uobj, off);
976 			}
977 			continue;
978 		}
979 
980 		/*
981 		 * if the current page needs to be cleaned and it's busy,
982 		 * wait for it to become unbusy.
983 		 */
984 
985 		yld = (l->l_cpu->ci_schedstate.spc_flags &
986 		    SPCF_SHOULDYIELD) && !pagedaemon;
987 		if (pg->flags & PG_BUSY || yld) {
988 			UVMHIST_LOG(ubchist, "busy %p", pg,0,0,0);
989 			if (flags & PGO_BUSYFAIL && pg->flags & PG_BUSY) {
990 				UVMHIST_LOG(ubchist, "busyfail %p", pg, 0,0,0);
991 				error = EDEADLK;
992 				if (busypg != NULL)
993 					*busypg = pg;
994 				break;
995 			}
996 			if (pagedaemon) {
997 				/*
998 				 * someone has taken the page while we
999 				 * dropped the lock for fstrans_start.
1000 				 */
1001 				break;
1002 			}
1003 			if (by_list) {
1004 				TAILQ_INSERT_BEFORE(pg, &curmp, listq.queue);
1005 				UVMHIST_LOG(ubchist, "curmp next %p",
1006 				    TAILQ_NEXT(&curmp, listq.queue), 0,0,0);
1007 			}
1008 			if (yld) {
1009 				mutex_exit(slock);
1010 				preempt();
1011 				mutex_enter(slock);
1012 			} else {
1013 				pg->flags |= PG_WANTED;
1014 				UVM_UNLOCK_AND_WAIT(pg, slock, 0, "genput", 0);
1015 				mutex_enter(slock);
1016 			}
1017 			if (by_list) {
1018 				UVMHIST_LOG(ubchist, "after next %p",
1019 				    TAILQ_NEXT(&curmp, listq.queue), 0,0,0);
1020 				pg = TAILQ_NEXT(&curmp, listq.queue);
1021 				TAILQ_REMOVE(&uobj->memq, &curmp, listq.queue);
1022 			} else {
1023 				pg = uvm_pagelookup(uobj, off);
1024 			}
1025 			continue;
1026 		}
1027 
1028 		/*
1029 		 * if we're freeing, remove all mappings of the page now.
1030 		 * if we're cleaning, check if the page is needs to be cleaned.
1031 		 */
1032 
1033 		if (flags & PGO_FREE) {
1034 			pmap_page_protect(pg, VM_PROT_NONE);
1035 		} else if (flags & PGO_CLEANIT) {
1036 
1037 			/*
1038 			 * if we still have some hope to pull this vnode off
1039 			 * from the syncer queue, write-protect the page.
1040 			 */
1041 
1042 			if (cleanall && wasclean &&
1043 			    gp->g_dirtygen == dirtygen) {
1044 
1045 				/*
1046 				 * uobj pages get wired only by uvm_fault
1047 				 * where uobj is locked.
1048 				 */
1049 
1050 				if (pg->wire_count == 0) {
1051 					pmap_page_protect(pg,
1052 					    VM_PROT_READ|VM_PROT_EXECUTE);
1053 				} else {
1054 					cleanall = false;
1055 				}
1056 			}
1057 		}
1058 
1059 		if (flags & PGO_CLEANIT) {
1060 			needs_clean = pmap_clear_modify(pg) ||
1061 			    (pg->flags & PG_CLEAN) == 0;
1062 			pg->flags |= PG_CLEAN;
1063 		} else {
1064 			needs_clean = false;
1065 		}
1066 
1067 		/*
1068 		 * if we're cleaning, build a cluster.
1069 		 * the cluster will consist of pages which are currently dirty,
1070 		 * but they will be returned to us marked clean.
1071 		 * if not cleaning, just operate on the one page.
1072 		 */
1073 
1074 		if (needs_clean) {
1075 			KDASSERT((vp->v_iflag & VI_ONWORKLST));
1076 			wasclean = false;
1077 			memset(pgs, 0, sizeof(pgs));
1078 			pg->flags |= PG_BUSY;
1079 			UVM_PAGE_OWN(pg, "genfs_putpages");
1080 
1081 			/*
1082 			 * first look backward.
1083 			 */
1084 
1085 			npages = MIN(maxpages >> 1, off >> PAGE_SHIFT);
1086 			nback = npages;
1087 			uvn_findpages(uobj, off - PAGE_SIZE, &nback, &pgs[0],
1088 			    UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY|UFP_BACKWARD);
1089 			if (nback) {
1090 				memmove(&pgs[0], &pgs[npages - nback],
1091 				    nback * sizeof(pgs[0]));
1092 				if (npages - nback < nback)
1093 					memset(&pgs[nback], 0,
1094 					    (npages - nback) * sizeof(pgs[0]));
1095 				else
1096 					memset(&pgs[npages - nback], 0,
1097 					    nback * sizeof(pgs[0]));
1098 			}
1099 
1100 			/*
1101 			 * then plug in our page of interest.
1102 			 */
1103 
1104 			pgs[nback] = pg;
1105 
1106 			/*
1107 			 * then look forward to fill in the remaining space in
1108 			 * the array of pages.
1109 			 */
1110 
1111 			npages = maxpages - nback - 1;
1112 			uvn_findpages(uobj, off + PAGE_SIZE, &npages,
1113 			    &pgs[nback + 1],
1114 			    UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY);
1115 			npages += nback + 1;
1116 		} else {
1117 			pgs[0] = pg;
1118 			npages = 1;
1119 			nback = 0;
1120 		}
1121 
1122 		/*
1123 		 * apply FREE or DEACTIVATE options if requested.
1124 		 */
1125 
1126 		if (flags & (PGO_DEACTIVATE|PGO_FREE)) {
1127 			mutex_enter(&uvm_pageqlock);
1128 		}
1129 		for (i = 0; i < npages; i++) {
1130 			tpg = pgs[i];
1131 			KASSERT(tpg->uobject == uobj);
1132 			if (by_list && tpg == TAILQ_NEXT(pg, listq.queue))
1133 				pg = tpg;
1134 			if (tpg->offset < startoff || tpg->offset >= endoff)
1135 				continue;
1136 			if (flags & PGO_DEACTIVATE && tpg->wire_count == 0) {
1137 				uvm_pagedeactivate(tpg);
1138 			} else if (flags & PGO_FREE) {
1139 				pmap_page_protect(tpg, VM_PROT_NONE);
1140 				if (tpg->flags & PG_BUSY) {
1141 					tpg->flags |= freeflag;
1142 					if (pagedaemon) {
1143 						uvm_pageout_start(1);
1144 						uvm_pagedequeue(tpg);
1145 					}
1146 				} else {
1147 
1148 					/*
1149 					 * ``page is not busy''
1150 					 * implies that npages is 1
1151 					 * and needs_clean is false.
1152 					 */
1153 
1154 					nextpg = TAILQ_NEXT(tpg, listq.queue);
1155 					uvm_pagefree(tpg);
1156 					if (pagedaemon)
1157 						uvmexp.pdfreed++;
1158 				}
1159 			}
1160 		}
1161 		if (flags & (PGO_DEACTIVATE|PGO_FREE)) {
1162 			mutex_exit(&uvm_pageqlock);
1163 		}
1164 		if (needs_clean) {
1165 			modified = true;
1166 
1167 			/*
1168 			 * start the i/o.  if we're traversing by list,
1169 			 * keep our place in the list with a marker page.
1170 			 */
1171 
1172 			if (by_list) {
1173 				TAILQ_INSERT_AFTER(&uobj->memq, pg, &curmp,
1174 				    listq.queue);
1175 			}
1176 			mutex_exit(slock);
1177 			error = GOP_WRITE(vp, pgs, npages, flags);
1178 			mutex_enter(slock);
1179 			if (by_list) {
1180 				pg = TAILQ_NEXT(&curmp, listq.queue);
1181 				TAILQ_REMOVE(&uobj->memq, &curmp, listq.queue);
1182 			}
1183 			if (error) {
1184 				break;
1185 			}
1186 			if (by_list) {
1187 				continue;
1188 			}
1189 		}
1190 
1191 		/*
1192 		 * find the next page and continue if there was no error.
1193 		 */
1194 
1195 		if (by_list) {
1196 			if (nextpg) {
1197 				pg = nextpg;
1198 				nextpg = NULL;
1199 			} else {
1200 				pg = TAILQ_NEXT(pg, listq.queue);
1201 			}
1202 		} else {
1203 			off += (npages - nback) << PAGE_SHIFT;
1204 			if (off < endoff) {
1205 				pg = uvm_pagelookup(uobj, off);
1206 			}
1207 		}
1208 	}
1209 	if (by_list) {
1210 		TAILQ_REMOVE(&uobj->memq, &endmp, listq.queue);
1211 	}
1212 
1213 	if (modified && (vp->v_iflag & VI_WRMAPDIRTY) != 0 &&
1214 	    (vp->v_type != VBLK ||
1215 	    (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) {
1216 		GOP_MARKUPDATE(vp, GOP_UPDATE_MODIFIED);
1217 	}
1218 
1219 	/*
1220 	 * if we're cleaning and there was nothing to clean,
1221 	 * take us off the syncer list.  if we started any i/o
1222 	 * and we're doing sync i/o, wait for all writes to finish.
1223 	 */
1224 
1225 	if (cleanall && wasclean && gp->g_dirtygen == dirtygen &&
1226 	    (vp->v_iflag & VI_ONWORKLST) != 0) {
1227 #if defined(DEBUG)
1228 		TAILQ_FOREACH(pg, &uobj->memq, listq.queue) {
1229 			if ((pg->flags & (PG_FAKE | PG_MARKER)) != 0) {
1230 				continue;
1231 			}
1232 			if ((pg->flags & PG_CLEAN) == 0) {
1233 				printf("%s: %p: !CLEAN\n", __func__, pg);
1234 			}
1235 			if (pmap_is_modified(pg)) {
1236 				printf("%s: %p: modified\n", __func__, pg);
1237 			}
1238 		}
1239 #endif /* defined(DEBUG) */
1240 		vp->v_iflag &= ~VI_WRMAPDIRTY;
1241 		if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL)
1242 			vn_syncer_remove_from_worklist(vp);
1243 	}
1244 
1245 #if !defined(DEBUG)
1246 skip_scan:
1247 #endif /* !defined(DEBUG) */
1248 
1249 	/* Wait for output to complete. */
1250 	if (!wasclean && !async && vp->v_numoutput != 0) {
1251 		while (vp->v_numoutput != 0)
1252 			cv_wait(&vp->v_cv, slock);
1253 	}
1254 	onworklst = (vp->v_iflag & VI_ONWORKLST) != 0;
1255 	mutex_exit(slock);
1256 
1257 	if ((flags & PGO_RECLAIM) != 0 && onworklst) {
1258 		/*
1259 		 * in the case of PGO_RECLAIM, ensure to make the vnode clean.
1260 		 * retrying is not a big deal because, in many cases,
1261 		 * uobj->uo_npages is already 0 here.
1262 		 */
1263 		mutex_enter(slock);
1264 		goto retry;
1265 	}
1266 
1267 	if (has_trans) {
1268 		if (need_wapbl)
1269 			WAPBL_END(vp->v_mount);
1270 		fstrans_done(vp->v_mount);
1271 	}
1272 
1273 	return (error);
1274 }
1275 
1276 int
1277 genfs_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, int flags)
1278 {
1279 	off_t off;
1280 	vaddr_t kva;
1281 	size_t len;
1282 	int error;
1283 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1284 
1285 	UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x",
1286 	    vp, pgs, npages, flags);
1287 
1288 	off = pgs[0]->offset;
1289 	kva = uvm_pagermapin(pgs, npages,
1290 	    UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK);
1291 	len = npages << PAGE_SHIFT;
1292 
1293 	error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE,
1294 			    uvm_aio_biodone);
1295 
1296 	return error;
1297 }
1298 
1299 int
1300 genfs_gop_write_rwmap(struct vnode *vp, struct vm_page **pgs, int npages, int flags)
1301 {
1302 	off_t off;
1303 	vaddr_t kva;
1304 	size_t len;
1305 	int error;
1306 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1307 
1308 	UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x",
1309 	    vp, pgs, npages, flags);
1310 
1311 	off = pgs[0]->offset;
1312 	kva = uvm_pagermapin(pgs, npages,
1313 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
1314 	len = npages << PAGE_SHIFT;
1315 
1316 	error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE,
1317 			    uvm_aio_biodone);
1318 
1319 	return error;
1320 }
1321 
1322 /*
1323  * Backend routine for doing I/O to vnode pages.  Pages are already locked
1324  * and mapped into kernel memory.  Here we just look up the underlying
1325  * device block addresses and call the strategy routine.
1326  */
1327 
1328 static int
1329 genfs_do_io(struct vnode *vp, off_t off, vaddr_t kva, size_t len, int flags,
1330     enum uio_rw rw, void (*iodone)(struct buf *))
1331 {
1332 	int s, error;
1333 	int fs_bshift, dev_bshift;
1334 	off_t eof, offset, startoffset;
1335 	size_t bytes, iobytes, skipbytes;
1336 	struct buf *mbp, *bp;
1337 	const bool async = (flags & PGO_SYNCIO) == 0;
1338 	const bool iowrite = rw == UIO_WRITE;
1339 	const int brw = iowrite ? B_WRITE : B_READ;
1340 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1341 
1342 	UVMHIST_LOG(ubchist, "vp %p kva %p len 0x%x flags 0x%x",
1343 	    vp, kva, len, flags);
1344 
1345 	KASSERT(vp->v_size <= vp->v_writesize);
1346 	GOP_SIZE(vp, vp->v_writesize, &eof, 0);
1347 	if (vp->v_type != VBLK) {
1348 		fs_bshift = vp->v_mount->mnt_fs_bshift;
1349 		dev_bshift = vp->v_mount->mnt_dev_bshift;
1350 	} else {
1351 		fs_bshift = DEV_BSHIFT;
1352 		dev_bshift = DEV_BSHIFT;
1353 	}
1354 	error = 0;
1355 	startoffset = off;
1356 	bytes = MIN(len, eof - startoffset);
1357 	skipbytes = 0;
1358 	KASSERT(bytes != 0);
1359 
1360 	if (iowrite) {
1361 		mutex_enter(&vp->v_interlock);
1362 		vp->v_numoutput += 2;
1363 		mutex_exit(&vp->v_interlock);
1364 	}
1365 	mbp = getiobuf(vp, true);
1366 	UVMHIST_LOG(ubchist, "vp %p mbp %p num now %d bytes 0x%x",
1367 	    vp, mbp, vp->v_numoutput, bytes);
1368 	mbp->b_bufsize = len;
1369 	mbp->b_data = (void *)kva;
1370 	mbp->b_resid = mbp->b_bcount = bytes;
1371 	mbp->b_cflags = BC_BUSY | BC_AGE;
1372 	if (async) {
1373 		mbp->b_flags = brw | B_ASYNC;
1374 		mbp->b_iodone = iodone;
1375 	} else {
1376 		mbp->b_flags = brw;
1377 		mbp->b_iodone = NULL;
1378 	}
1379 	if (curlwp == uvm.pagedaemon_lwp)
1380 		BIO_SETPRIO(mbp, BPRIO_TIMELIMITED);
1381 	else if (async)
1382 		BIO_SETPRIO(mbp, BPRIO_TIMENONCRITICAL);
1383 	else
1384 		BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL);
1385 
1386 	bp = NULL;
1387 	for (offset = startoffset;
1388 	    bytes > 0;
1389 	    offset += iobytes, bytes -= iobytes) {
1390 		int run;
1391 		daddr_t lbn, blkno;
1392 		struct vnode *devvp;
1393 
1394 		/*
1395 		 * bmap the file to find out the blkno to read from and
1396 		 * how much we can read in one i/o.  if bmap returns an error,
1397 		 * skip the rest of the top-level i/o.
1398 		 */
1399 
1400 		lbn = offset >> fs_bshift;
1401 		error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run);
1402 		if (error) {
1403 			UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%x -> %d\n",
1404 			    lbn,error,0,0);
1405 			skipbytes += bytes;
1406 			bytes = 0;
1407 			goto loopdone;
1408 		}
1409 
1410 		/*
1411 		 * see how many pages can be read with this i/o.
1412 		 * reduce the i/o size if necessary to avoid
1413 		 * overwriting pages with valid data.
1414 		 */
1415 
1416 		iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset,
1417 		    bytes);
1418 
1419 		/*
1420 		 * if this block isn't allocated, zero it instead of
1421 		 * reading it.  unless we are going to allocate blocks,
1422 		 * mark the pages we zeroed PG_RDONLY.
1423 		 */
1424 
1425 		if (blkno == (daddr_t)-1) {
1426 			if (!iowrite) {
1427 				memset((char *)kva + (offset - startoffset), 0,
1428 				    iobytes);
1429 			}
1430 			skipbytes += iobytes;
1431 			continue;
1432 		}
1433 
1434 		/*
1435 		 * allocate a sub-buf for this piece of the i/o
1436 		 * (or just use mbp if there's only 1 piece),
1437 		 * and start it going.
1438 		 */
1439 
1440 		if (offset == startoffset && iobytes == bytes) {
1441 			bp = mbp;
1442 		} else {
1443 			UVMHIST_LOG(ubchist, "vp %p bp %p num now %d",
1444 			    vp, bp, vp->v_numoutput, 0);
1445 			bp = getiobuf(vp, true);
1446 			nestiobuf_setup(mbp, bp, offset - startoffset, iobytes);
1447 		}
1448 		bp->b_lblkno = 0;
1449 
1450 		/* adjust physical blkno for partial blocks */
1451 		bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >>
1452 		    dev_bshift);
1453 
1454 		UVMHIST_LOG(ubchist,
1455 		    "bp %p offset 0x%x bcount 0x%x blkno 0x%x",
1456 		    bp, offset, bp->b_bcount, bp->b_blkno);
1457 
1458 		VOP_STRATEGY(devvp, bp);
1459 	}
1460 
1461 loopdone:
1462 	if (skipbytes) {
1463 		UVMHIST_LOG(ubchist, "skipbytes %d", skipbytes, 0,0,0);
1464 	}
1465 	nestiobuf_done(mbp, skipbytes, error);
1466 	if (async) {
1467 		UVMHIST_LOG(ubchist, "returning 0 (async)", 0,0,0,0);
1468 		return (0);
1469 	}
1470 	UVMHIST_LOG(ubchist, "waiting for mbp %p", mbp,0,0,0);
1471 	error = biowait(mbp);
1472 	s = splbio();
1473 	(*iodone)(mbp);
1474 	splx(s);
1475 	UVMHIST_LOG(ubchist, "returning, error %d", error,0,0,0);
1476 	return (error);
1477 }
1478 
1479 int
1480 genfs_compat_getpages(void *v)
1481 {
1482 	struct vop_getpages_args /* {
1483 		struct vnode *a_vp;
1484 		voff_t a_offset;
1485 		struct vm_page **a_m;
1486 		int *a_count;
1487 		int a_centeridx;
1488 		vm_prot_t a_access_type;
1489 		int a_advice;
1490 		int a_flags;
1491 	} */ *ap = v;
1492 
1493 	off_t origoffset;
1494 	struct vnode *vp = ap->a_vp;
1495 	struct uvm_object *uobj = &vp->v_uobj;
1496 	struct vm_page *pg, **pgs;
1497 	vaddr_t kva;
1498 	int i, error, orignpages, npages;
1499 	struct iovec iov;
1500 	struct uio uio;
1501 	kauth_cred_t cred = curlwp->l_cred;
1502 	const bool memwrite = (ap->a_access_type & VM_PROT_WRITE) != 0;
1503 
1504 	error = 0;
1505 	origoffset = ap->a_offset;
1506 	orignpages = *ap->a_count;
1507 	pgs = ap->a_m;
1508 
1509 	if (ap->a_flags & PGO_LOCKED) {
1510 		uvn_findpages(uobj, origoffset, ap->a_count, ap->a_m,
1511 		    UFP_NOWAIT|UFP_NOALLOC| (memwrite ? UFP_NORDONLY : 0));
1512 
1513 		error = ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0;
1514 		if (error == 0 && memwrite) {
1515 			genfs_markdirty(vp);
1516 		}
1517 		return error;
1518 	}
1519 	if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= vp->v_size) {
1520 		mutex_exit(&uobj->vmobjlock);
1521 		return EINVAL;
1522 	}
1523 	if ((ap->a_flags & PGO_SYNCIO) == 0) {
1524 		mutex_exit(&uobj->vmobjlock);
1525 		return 0;
1526 	}
1527 	npages = orignpages;
1528 	uvn_findpages(uobj, origoffset, &npages, pgs, UFP_ALL);
1529 	mutex_exit(&uobj->vmobjlock);
1530 	kva = uvm_pagermapin(pgs, npages,
1531 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
1532 	for (i = 0; i < npages; i++) {
1533 		pg = pgs[i];
1534 		if ((pg->flags & PG_FAKE) == 0) {
1535 			continue;
1536 		}
1537 		iov.iov_base = (char *)kva + (i << PAGE_SHIFT);
1538 		iov.iov_len = PAGE_SIZE;
1539 		uio.uio_iov = &iov;
1540 		uio.uio_iovcnt = 1;
1541 		uio.uio_offset = origoffset + (i << PAGE_SHIFT);
1542 		uio.uio_rw = UIO_READ;
1543 		uio.uio_resid = PAGE_SIZE;
1544 		UIO_SETUP_SYSSPACE(&uio);
1545 		/* XXX vn_lock */
1546 		error = VOP_READ(vp, &uio, 0, cred);
1547 		if (error) {
1548 			break;
1549 		}
1550 		if (uio.uio_resid) {
1551 			memset(iov.iov_base, 0, uio.uio_resid);
1552 		}
1553 	}
1554 	uvm_pagermapout(kva, npages);
1555 	mutex_enter(&uobj->vmobjlock);
1556 	mutex_enter(&uvm_pageqlock);
1557 	for (i = 0; i < npages; i++) {
1558 		pg = pgs[i];
1559 		if (error && (pg->flags & PG_FAKE) != 0) {
1560 			pg->flags |= PG_RELEASED;
1561 		} else {
1562 			pmap_clear_modify(pg);
1563 			uvm_pageactivate(pg);
1564 		}
1565 	}
1566 	if (error) {
1567 		uvm_page_unbusy(pgs, npages);
1568 	}
1569 	mutex_exit(&uvm_pageqlock);
1570 	if (error == 0 && memwrite) {
1571 		genfs_markdirty(vp);
1572 	}
1573 	mutex_exit(&uobj->vmobjlock);
1574 	return error;
1575 }
1576 
1577 int
1578 genfs_compat_gop_write(struct vnode *vp, struct vm_page **pgs, int npages,
1579     int flags)
1580 {
1581 	off_t offset;
1582 	struct iovec iov;
1583 	struct uio uio;
1584 	kauth_cred_t cred = curlwp->l_cred;
1585 	struct buf *bp;
1586 	vaddr_t kva;
1587 	int error;
1588 
1589 	offset = pgs[0]->offset;
1590 	kva = uvm_pagermapin(pgs, npages,
1591 	    UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK);
1592 
1593 	iov.iov_base = (void *)kva;
1594 	iov.iov_len = npages << PAGE_SHIFT;
1595 	uio.uio_iov = &iov;
1596 	uio.uio_iovcnt = 1;
1597 	uio.uio_offset = offset;
1598 	uio.uio_rw = UIO_WRITE;
1599 	uio.uio_resid = npages << PAGE_SHIFT;
1600 	UIO_SETUP_SYSSPACE(&uio);
1601 	/* XXX vn_lock */
1602 	error = VOP_WRITE(vp, &uio, 0, cred);
1603 
1604 	mutex_enter(&vp->v_interlock);
1605 	vp->v_numoutput++;
1606 	mutex_exit(&vp->v_interlock);
1607 
1608 	bp = getiobuf(vp, true);
1609 	bp->b_cflags = BC_BUSY | BC_AGE;
1610 	bp->b_lblkno = offset >> vp->v_mount->mnt_fs_bshift;
1611 	bp->b_data = (char *)kva;
1612 	bp->b_bcount = npages << PAGE_SHIFT;
1613 	bp->b_bufsize = npages << PAGE_SHIFT;
1614 	bp->b_resid = 0;
1615 	bp->b_error = error;
1616 	uvm_aio_aiodone(bp);
1617 	return (error);
1618 }
1619 
1620 /*
1621  * Process a uio using direct I/O.  If we reach a part of the request
1622  * which cannot be processed in this fashion for some reason, just return.
1623  * The caller must handle some additional part of the request using
1624  * buffered I/O before trying direct I/O again.
1625  */
1626 
1627 void
1628 genfs_directio(struct vnode *vp, struct uio *uio, int ioflag)
1629 {
1630 	struct vmspace *vs;
1631 	struct iovec *iov;
1632 	vaddr_t va;
1633 	size_t len;
1634 	const int mask = DEV_BSIZE - 1;
1635 	int error;
1636 	bool need_wapbl = (vp->v_mount && vp->v_mount->mnt_wapbl &&
1637 	    (ioflag & IO_JOURNALLOCKED) == 0);
1638 
1639 	/*
1640 	 * We only support direct I/O to user space for now.
1641 	 */
1642 
1643 	if (VMSPACE_IS_KERNEL_P(uio->uio_vmspace)) {
1644 		return;
1645 	}
1646 
1647 	/*
1648 	 * If the vnode is mapped, we would need to get the getpages lock
1649 	 * to stabilize the bmap, but then we would get into trouble whil e
1650 	 * locking the pages if the pages belong to this same vnode (or a
1651 	 * multi-vnode cascade to the same effect).  Just fall back to
1652 	 * buffered I/O if the vnode is mapped to avoid this mess.
1653 	 */
1654 
1655 	if (vp->v_vflag & VV_MAPPED) {
1656 		return;
1657 	}
1658 
1659 	if (need_wapbl) {
1660 		error = WAPBL_BEGIN(vp->v_mount);
1661 		if (error)
1662 			return;
1663 	}
1664 
1665 	/*
1666 	 * Do as much of the uio as possible with direct I/O.
1667 	 */
1668 
1669 	vs = uio->uio_vmspace;
1670 	while (uio->uio_resid) {
1671 		iov = uio->uio_iov;
1672 		if (iov->iov_len == 0) {
1673 			uio->uio_iov++;
1674 			uio->uio_iovcnt--;
1675 			continue;
1676 		}
1677 		va = (vaddr_t)iov->iov_base;
1678 		len = MIN(iov->iov_len, genfs_maxdio);
1679 		len &= ~mask;
1680 
1681 		/*
1682 		 * If the next chunk is smaller than DEV_BSIZE or extends past
1683 		 * the current EOF, then fall back to buffered I/O.
1684 		 */
1685 
1686 		if (len == 0 || uio->uio_offset + len > vp->v_size) {
1687 			break;
1688 		}
1689 
1690 		/*
1691 		 * Check alignment.  The file offset must be at least
1692 		 * sector-aligned.  The exact constraint on memory alignment
1693 		 * is very hardware-dependent, but requiring sector-aligned
1694 		 * addresses there too is safe.
1695 		 */
1696 
1697 		if (uio->uio_offset & mask || va & mask) {
1698 			break;
1699 		}
1700 		error = genfs_do_directio(vs, va, len, vp, uio->uio_offset,
1701 					  uio->uio_rw);
1702 		if (error) {
1703 			break;
1704 		}
1705 		iov->iov_base = (char *)iov->iov_base + len;
1706 		iov->iov_len -= len;
1707 		uio->uio_offset += len;
1708 		uio->uio_resid -= len;
1709 	}
1710 
1711 	if (need_wapbl)
1712 		WAPBL_END(vp->v_mount);
1713 }
1714 
1715 /*
1716  * Iodone routine for direct I/O.  We don't do much here since the request is
1717  * always synchronous, so the caller will do most of the work after biowait().
1718  */
1719 
1720 static void
1721 genfs_dio_iodone(struct buf *bp)
1722 {
1723 
1724 	KASSERT((bp->b_flags & B_ASYNC) == 0);
1725 	if ((bp->b_flags & B_READ) == 0 && (bp->b_cflags & BC_AGE) != 0) {
1726 		mutex_enter(bp->b_objlock);
1727 		vwakeup(bp);
1728 		mutex_exit(bp->b_objlock);
1729 	}
1730 	putiobuf(bp);
1731 }
1732 
1733 /*
1734  * Process one chunk of a direct I/O request.
1735  */
1736 
1737 static int
1738 genfs_do_directio(struct vmspace *vs, vaddr_t uva, size_t len, struct vnode *vp,
1739     off_t off, enum uio_rw rw)
1740 {
1741 	struct vm_map *map;
1742 	struct pmap *upm, *kpm;
1743 	size_t klen = round_page(uva + len) - trunc_page(uva);
1744 	off_t spoff, epoff;
1745 	vaddr_t kva, puva;
1746 	paddr_t pa;
1747 	vm_prot_t prot;
1748 	int error, rv, poff, koff;
1749 	const int pgoflags = PGO_CLEANIT | PGO_SYNCIO | PGO_JOURNALLOCKED |
1750 		(rw == UIO_WRITE ? PGO_FREE : 0);
1751 
1752 	/*
1753 	 * For writes, verify that this range of the file already has fully
1754 	 * allocated backing store.  If there are any holes, just punt and
1755 	 * make the caller take the buffered write path.
1756 	 */
1757 
1758 	if (rw == UIO_WRITE) {
1759 		daddr_t lbn, elbn, blkno;
1760 		int bsize, bshift, run;
1761 
1762 		bshift = vp->v_mount->mnt_fs_bshift;
1763 		bsize = 1 << bshift;
1764 		lbn = off >> bshift;
1765 		elbn = (off + len + bsize - 1) >> bshift;
1766 		while (lbn < elbn) {
1767 			error = VOP_BMAP(vp, lbn, NULL, &blkno, &run);
1768 			if (error) {
1769 				return error;
1770 			}
1771 			if (blkno == (daddr_t)-1) {
1772 				return ENOSPC;
1773 			}
1774 			lbn += 1 + run;
1775 		}
1776 	}
1777 
1778 	/*
1779 	 * Flush any cached pages for parts of the file that we're about to
1780 	 * access.  If we're writing, invalidate pages as well.
1781 	 */
1782 
1783 	spoff = trunc_page(off);
1784 	epoff = round_page(off + len);
1785 	mutex_enter(&vp->v_interlock);
1786 	error = VOP_PUTPAGES(vp, spoff, epoff, pgoflags);
1787 	if (error) {
1788 		return error;
1789 	}
1790 
1791 	/*
1792 	 * Wire the user pages and remap them into kernel memory.
1793 	 */
1794 
1795 	prot = rw == UIO_READ ? VM_PROT_READ | VM_PROT_WRITE : VM_PROT_READ;
1796 	error = uvm_vslock(vs, (void *)uva, len, prot);
1797 	if (error) {
1798 		return error;
1799 	}
1800 
1801 	map = &vs->vm_map;
1802 	upm = vm_map_pmap(map);
1803 	kpm = vm_map_pmap(kernel_map);
1804 	kva = uvm_km_alloc(kernel_map, klen, 0,
1805 			   UVM_KMF_VAONLY | UVM_KMF_WAITVA);
1806 	puva = trunc_page(uva);
1807 	for (poff = 0; poff < klen; poff += PAGE_SIZE) {
1808 		rv = pmap_extract(upm, puva + poff, &pa);
1809 		KASSERT(rv);
1810 		pmap_enter(kpm, kva + poff, pa, prot, prot | PMAP_WIRED);
1811 	}
1812 	pmap_update(kpm);
1813 
1814 	/*
1815 	 * Do the I/O.
1816 	 */
1817 
1818 	koff = uva - trunc_page(uva);
1819 	error = genfs_do_io(vp, off, kva + koff, len, PGO_SYNCIO, rw,
1820 			    genfs_dio_iodone);
1821 
1822 	/*
1823 	 * Tear down the kernel mapping.
1824 	 */
1825 
1826 	pmap_remove(kpm, kva, kva + klen);
1827 	pmap_update(kpm);
1828 	uvm_km_free(kernel_map, kva, klen, UVM_KMF_VAONLY);
1829 
1830 	/*
1831 	 * Unwire the user pages.
1832 	 */
1833 
1834 	uvm_vsunlock(vs, (void *)uva, len);
1835 	return error;
1836 }
1837 
1838