xref: /netbsd-src/sys/miscfs/genfs/genfs_io.c (revision da9817918ec7e88db2912a2882967c7570a83f47)
1 /*	$NetBSD: genfs_io.c,v 1.20 2009/04/18 15:40:33 pooka 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 1.20 2009/04/18 15:40:33 pooka 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 inline void genfs_rel_pages(struct vm_page **, int);
66 
67 int genfs_maxdio = MAXPHYS;
68 
69 static inline void
70 genfs_rel_pages(struct vm_page **pgs, int npages)
71 {
72 	int i;
73 
74 	for (i = 0; i < npages; i++) {
75 		struct vm_page *pg = pgs[i];
76 
77 		if (pg == NULL || pg == PGO_DONTCARE)
78 			continue;
79 		if (pg->flags & PG_FAKE) {
80 			pg->flags |= PG_RELEASED;
81 		}
82 	}
83 	mutex_enter(&uvm_pageqlock);
84 	uvm_page_unbusy(pgs, npages);
85 	mutex_exit(&uvm_pageqlock);
86 }
87 
88 /*
89  * generic VM getpages routine.
90  * Return PG_BUSY pages for the given range,
91  * reading from backing store if necessary.
92  */
93 
94 int
95 genfs_getpages(void *v)
96 {
97 	struct vop_getpages_args /* {
98 		struct vnode *a_vp;
99 		voff_t a_offset;
100 		struct vm_page **a_m;
101 		int *a_count;
102 		int a_centeridx;
103 		vm_prot_t a_access_type;
104 		int a_advice;
105 		int a_flags;
106 	} */ *ap = v;
107 
108 	off_t newsize, diskeof, memeof;
109 	off_t offset, origoffset, startoffset, endoffset;
110 	daddr_t lbn, blkno;
111 	int i, error, npages, orignpages, npgs, run, ridx, pidx, pcount;
112 	int fs_bshift, fs_bsize, dev_bshift;
113 	const int flags = ap->a_flags;
114 	size_t bytes, iobytes, tailstart, tailbytes, totalbytes, skipbytes;
115 	vaddr_t kva;
116 	struct buf *bp, *mbp;
117 	struct vnode *vp = ap->a_vp;
118 	struct vnode *devvp;
119 	struct genfs_node *gp = VTOG(vp);
120 	struct uvm_object *uobj = &vp->v_uobj;
121 	struct vm_page *pg, **pgs, *pgs_onstack[UBC_MAX_PAGES];
122 	int pgs_size;
123 	kauth_cred_t cred = curlwp->l_cred;		/* XXXUBC curlwp */
124 	const bool async = (flags & PGO_SYNCIO) == 0;
125 	const bool write = (ap->a_access_type & VM_PROT_WRITE) != 0;
126 	bool sawhole = false;
127 	bool has_trans = false;
128 	const bool overwrite = (flags & PGO_OVERWRITE) != 0;
129 	const bool blockalloc = write && (flags & PGO_NOBLOCKALLOC) == 0;
130 	voff_t origvsize;
131 	UVMHIST_FUNC("genfs_getpages"); UVMHIST_CALLED(ubchist);
132 
133 	UVMHIST_LOG(ubchist, "vp %p off 0x%x/%x count %d",
134 	    vp, ap->a_offset >> 32, ap->a_offset, *ap->a_count);
135 
136 	KASSERT(vp->v_type == VREG || vp->v_type == VDIR ||
137 	    vp->v_type == VLNK || vp->v_type == VBLK);
138 
139 	pgs = NULL;
140 	pgs_size = 0;
141 
142 startover:
143 	error = 0;
144 	origvsize = vp->v_size;
145 	origoffset = ap->a_offset;
146 	orignpages = *ap->a_count;
147 	GOP_SIZE(vp, origvsize, &diskeof, 0);
148 	if (flags & PGO_PASTEOF) {
149 #if defined(DIAGNOSTIC)
150 		off_t writeeof;
151 #endif /* defined(DIAGNOSTIC) */
152 
153 		newsize = MAX(origvsize,
154 		    origoffset + (orignpages << PAGE_SHIFT));
155 		GOP_SIZE(vp, newsize, &memeof, GOP_SIZE_MEM);
156 #if defined(DIAGNOSTIC)
157 		GOP_SIZE(vp, vp->v_writesize, &writeeof, GOP_SIZE_MEM);
158 		if (newsize > round_page(writeeof)) {
159 			panic("%s: past eof", __func__);
160 		}
161 #endif /* defined(DIAGNOSTIC) */
162 	} else {
163 		GOP_SIZE(vp, origvsize, &memeof, GOP_SIZE_MEM);
164 	}
165 	KASSERT(ap->a_centeridx >= 0 || ap->a_centeridx <= orignpages);
166 	KASSERT((origoffset & (PAGE_SIZE - 1)) == 0 && origoffset >= 0);
167 	KASSERT(orignpages > 0);
168 
169 	/*
170 	 * Bounds-check the request.
171 	 */
172 
173 	if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= memeof) {
174 		if ((flags & PGO_LOCKED) == 0) {
175 			mutex_exit(&uobj->vmobjlock);
176 		}
177 		UVMHIST_LOG(ubchist, "off 0x%x count %d goes past EOF 0x%x",
178 		    origoffset, *ap->a_count, memeof,0);
179 		error = EINVAL;
180 		goto out_err;
181 	}
182 
183 	/* uobj is locked */
184 
185 	if ((flags & PGO_NOTIMESTAMP) == 0 &&
186 	    (vp->v_type != VBLK ||
187 	    (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) {
188 		int updflags = 0;
189 
190 		if ((vp->v_mount->mnt_flag & MNT_NOATIME) == 0) {
191 			updflags = GOP_UPDATE_ACCESSED;
192 		}
193 		if (write) {
194 			updflags |= GOP_UPDATE_MODIFIED;
195 		}
196 		if (updflags != 0) {
197 			GOP_MARKUPDATE(vp, updflags);
198 		}
199 	}
200 
201 	if (write) {
202 		gp->g_dirtygen++;
203 		if ((vp->v_iflag & VI_ONWORKLST) == 0) {
204 			vn_syncer_add_to_worklist(vp, filedelay);
205 		}
206 		if ((vp->v_iflag & (VI_WRMAP|VI_WRMAPDIRTY)) == VI_WRMAP) {
207 			vp->v_iflag |= VI_WRMAPDIRTY;
208 		}
209 	}
210 
211 	/*
212 	 * For PGO_LOCKED requests, just return whatever's in memory.
213 	 */
214 
215 	if (flags & PGO_LOCKED) {
216 		int nfound;
217 
218 		npages = *ap->a_count;
219 #if defined(DEBUG)
220 		for (i = 0; i < npages; i++) {
221 			pg = ap->a_m[i];
222 			KASSERT(pg == NULL || pg == PGO_DONTCARE);
223 		}
224 #endif /* defined(DEBUG) */
225 		nfound = uvn_findpages(uobj, origoffset, &npages,
226 		    ap->a_m, UFP_NOWAIT|UFP_NOALLOC|(write ? UFP_NORDONLY : 0));
227 		KASSERT(npages == *ap->a_count);
228 		if (nfound == 0) {
229 			error = EBUSY;
230 			goto out_err;
231 		}
232 		if (!rw_tryenter(&gp->g_glock, RW_READER)) {
233 			genfs_rel_pages(ap->a_m, npages);
234 
235 			/*
236 			 * restore the array.
237 			 */
238 
239 			for (i = 0; i < npages; i++) {
240 				pg = ap->a_m[i];
241 
242 				if (pg != NULL || pg != PGO_DONTCARE) {
243 					ap->a_m[i] = NULL;
244 				}
245 			}
246 		} else {
247 			rw_exit(&gp->g_glock);
248 		}
249 		error = (ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0);
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 	if (vp->v_type != VBLK) {
260 		fs_bshift = vp->v_mount->mnt_fs_bshift;
261 		dev_bshift = vp->v_mount->mnt_dev_bshift;
262 	} else {
263 		fs_bshift = DEV_BSHIFT;
264 		dev_bshift = DEV_BSHIFT;
265 	}
266 	fs_bsize = 1 << fs_bshift;
267 
268 	orignpages = MIN(orignpages,
269 	    round_page(memeof - origoffset) >> PAGE_SHIFT);
270 	npages = orignpages;
271 	startoffset = origoffset & ~(fs_bsize - 1);
272 	endoffset = round_page((origoffset + (npages << PAGE_SHIFT) +
273 	    fs_bsize - 1) & ~(fs_bsize - 1));
274 	endoffset = MIN(endoffset, round_page(memeof));
275 	ridx = (origoffset - startoffset) >> PAGE_SHIFT;
276 
277 	pgs_size = sizeof(struct vm_page *) *
278 	    ((endoffset - startoffset) >> PAGE_SHIFT);
279 	if (pgs_size > sizeof(pgs_onstack)) {
280 		pgs = kmem_zalloc(pgs_size, async ? KM_NOSLEEP : KM_SLEEP);
281 		if (pgs == NULL) {
282 			pgs = pgs_onstack;
283 			error = ENOMEM;
284 			goto out_err;
285 		}
286 	} else {
287 		pgs = pgs_onstack;
288 		(void)memset(pgs, 0, pgs_size);
289 	}
290 
291 
292 	UVMHIST_LOG(ubchist, "ridx %d npages %d startoff %ld endoff %ld",
293 	    ridx, npages, startoffset, endoffset);
294 
295 	if (!has_trans) {
296 		fstrans_start(vp->v_mount, FSTRANS_SHARED);
297 		has_trans = true;
298 	}
299 
300 	/*
301 	 * hold g_glock to prevent a race with truncate.
302 	 *
303 	 * check if our idea of v_size is still valid.
304 	 */
305 
306 	if (blockalloc) {
307 		rw_enter(&gp->g_glock, RW_WRITER);
308 	} else {
309 		rw_enter(&gp->g_glock, RW_READER);
310 	}
311 	mutex_enter(&uobj->vmobjlock);
312 	if (vp->v_size < origvsize) {
313 		rw_exit(&gp->g_glock);
314 		if (pgs != pgs_onstack)
315 			kmem_free(pgs, pgs_size);
316 		goto startover;
317 	}
318 
319 	if (uvn_findpages(uobj, origoffset, &npages, &pgs[ridx],
320 	    async ? UFP_NOWAIT : UFP_ALL) != orignpages) {
321 		rw_exit(&gp->g_glock);
322 		KASSERT(async != 0);
323 		genfs_rel_pages(&pgs[ridx], orignpages);
324 		mutex_exit(&uobj->vmobjlock);
325 		error = EBUSY;
326 		goto out_err;
327 	}
328 
329 	/*
330 	 * if the pages are already resident, just return them.
331 	 */
332 
333 	for (i = 0; i < npages; i++) {
334 		struct vm_page *pg1 = pgs[ridx + i];
335 
336 		if ((pg1->flags & PG_FAKE) ||
337 		    (blockalloc && (pg1->flags & PG_RDONLY))) {
338 			break;
339 		}
340 	}
341 	if (i == npages) {
342 		rw_exit(&gp->g_glock);
343 		UVMHIST_LOG(ubchist, "returning cached pages", 0,0,0,0);
344 		npages += ridx;
345 		goto out;
346 	}
347 
348 	/*
349 	 * if PGO_OVERWRITE is set, don't bother reading the pages.
350 	 */
351 
352 	if (overwrite) {
353 		rw_exit(&gp->g_glock);
354 		UVMHIST_LOG(ubchist, "PGO_OVERWRITE",0,0,0,0);
355 
356 		for (i = 0; i < npages; i++) {
357 			struct vm_page *pg1 = pgs[ridx + i];
358 
359 			pg1->flags &= ~(PG_RDONLY|PG_CLEAN);
360 		}
361 		npages += ridx;
362 		goto out;
363 	}
364 
365 	/*
366 	 * the page wasn't resident and we're not overwriting,
367 	 * so we're going to have to do some i/o.
368 	 * find any additional pages needed to cover the expanded range.
369 	 */
370 
371 	npages = (endoffset - startoffset) >> PAGE_SHIFT;
372 	if (startoffset != origoffset || npages != orignpages) {
373 
374 		/*
375 		 * we need to avoid deadlocks caused by locking
376 		 * additional pages at lower offsets than pages we
377 		 * already have locked.  unlock them all and start over.
378 		 */
379 
380 		genfs_rel_pages(&pgs[ridx], orignpages);
381 		memset(pgs, 0, pgs_size);
382 
383 		UVMHIST_LOG(ubchist, "reset npages start 0x%x end 0x%x",
384 		    startoffset, endoffset, 0,0);
385 		npgs = npages;
386 		if (uvn_findpages(uobj, startoffset, &npgs, pgs,
387 		    async ? UFP_NOWAIT : UFP_ALL) != npages) {
388 			rw_exit(&gp->g_glock);
389 			KASSERT(async != 0);
390 			genfs_rel_pages(pgs, npages);
391 			mutex_exit(&uobj->vmobjlock);
392 			error = EBUSY;
393 			goto out_err;
394 		}
395 	}
396 	mutex_exit(&uobj->vmobjlock);
397 
398 	/*
399 	 * read the desired page(s).
400 	 */
401 
402 	totalbytes = npages << PAGE_SHIFT;
403 	bytes = MIN(totalbytes, MAX(diskeof - startoffset, 0));
404 	tailbytes = totalbytes - bytes;
405 	skipbytes = 0;
406 
407 	kva = uvm_pagermapin(pgs, npages,
408 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
409 
410 	mbp = getiobuf(vp, true);
411 	mbp->b_bufsize = totalbytes;
412 	mbp->b_data = (void *)kva;
413 	mbp->b_resid = mbp->b_bcount = bytes;
414 	mbp->b_cflags = BC_BUSY;
415 	if (async) {
416 		mbp->b_flags = B_READ | B_ASYNC;
417 		mbp->b_iodone = uvm_aio_biodone;
418 	} else {
419 		mbp->b_flags = B_READ;
420 		mbp->b_iodone = NULL;
421 	}
422 	if (async)
423 		BIO_SETPRIO(mbp, BPRIO_TIMELIMITED);
424 	else
425 		BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL);
426 
427 	/*
428 	 * if EOF is in the middle of the range, zero the part past EOF.
429 	 * skip over pages which are not PG_FAKE since in that case they have
430 	 * valid data that we need to preserve.
431 	 */
432 
433 	tailstart = bytes;
434 	while (tailbytes > 0) {
435 		const int len = PAGE_SIZE - (tailstart & PAGE_MASK);
436 
437 		KASSERT(len <= tailbytes);
438 		if ((pgs[tailstart >> PAGE_SHIFT]->flags & PG_FAKE) != 0) {
439 			memset((void *)(kva + tailstart), 0, len);
440 			UVMHIST_LOG(ubchist, "tailbytes %p 0x%x 0x%x",
441 			    kva, tailstart, len, 0);
442 		}
443 		tailstart += len;
444 		tailbytes -= len;
445 	}
446 
447 	/*
448 	 * now loop over the pages, reading as needed.
449 	 */
450 
451 	bp = NULL;
452 	for (offset = startoffset;
453 	    bytes > 0;
454 	    offset += iobytes, bytes -= iobytes) {
455 
456 		/*
457 		 * skip pages which don't need to be read.
458 		 */
459 
460 		pidx = (offset - startoffset) >> PAGE_SHIFT;
461 		while ((pgs[pidx]->flags & PG_FAKE) == 0) {
462 			size_t b;
463 
464 			KASSERT((offset & (PAGE_SIZE - 1)) == 0);
465 			if ((pgs[pidx]->flags & PG_RDONLY)) {
466 				sawhole = true;
467 			}
468 			b = MIN(PAGE_SIZE, bytes);
469 			offset += b;
470 			bytes -= b;
471 			skipbytes += b;
472 			pidx++;
473 			UVMHIST_LOG(ubchist, "skipping, new offset 0x%x",
474 			    offset, 0,0,0);
475 			if (bytes == 0) {
476 				goto loopdone;
477 			}
478 		}
479 
480 		/*
481 		 * bmap the file to find out the blkno to read from and
482 		 * how much we can read in one i/o.  if bmap returns an error,
483 		 * skip the rest of the top-level i/o.
484 		 */
485 
486 		lbn = offset >> fs_bshift;
487 		error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run);
488 		if (error) {
489 			UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%x -> %d\n",
490 			    lbn, error,0,0);
491 			skipbytes += bytes;
492 			goto loopdone;
493 		}
494 
495 		/*
496 		 * see how many pages can be read with this i/o.
497 		 * reduce the i/o size if necessary to avoid
498 		 * overwriting pages with valid data.
499 		 */
500 
501 		iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset,
502 		    bytes);
503 		if (offset + iobytes > round_page(offset)) {
504 			pcount = 1;
505 			while (pidx + pcount < npages &&
506 			    pgs[pidx + pcount]->flags & PG_FAKE) {
507 				pcount++;
508 			}
509 			iobytes = MIN(iobytes, (pcount << PAGE_SHIFT) -
510 			    (offset - trunc_page(offset)));
511 		}
512 
513 		/*
514 		 * if this block isn't allocated, zero it instead of
515 		 * reading it.  unless we are going to allocate blocks,
516 		 * mark the pages we zeroed PG_RDONLY.
517 		 */
518 
519 		if (blkno < 0) {
520 			int holepages = (round_page(offset + iobytes) -
521 			    trunc_page(offset)) >> PAGE_SHIFT;
522 			UVMHIST_LOG(ubchist, "lbn 0x%x -> HOLE", lbn,0,0,0);
523 
524 			sawhole = true;
525 			memset((char *)kva + (offset - startoffset), 0,
526 			    iobytes);
527 			skipbytes += iobytes;
528 
529 			for (i = 0; i < holepages; i++) {
530 				if (write) {
531 					pgs[pidx + i]->flags &= ~PG_CLEAN;
532 				}
533 				if (!blockalloc) {
534 					pgs[pidx + i]->flags |= PG_RDONLY;
535 				}
536 			}
537 			continue;
538 		}
539 
540 		/*
541 		 * allocate a sub-buf for this piece of the i/o
542 		 * (or just use mbp if there's only 1 piece),
543 		 * and start it going.
544 		 */
545 
546 		if (offset == startoffset && iobytes == bytes) {
547 			bp = mbp;
548 		} else {
549 			bp = getiobuf(vp, true);
550 			nestiobuf_setup(mbp, bp, offset - startoffset, iobytes);
551 		}
552 		bp->b_lblkno = 0;
553 
554 		/* adjust physical blkno for partial blocks */
555 		bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >>
556 		    dev_bshift);
557 
558 		UVMHIST_LOG(ubchist,
559 		    "bp %p offset 0x%x bcount 0x%x blkno 0x%x",
560 		    bp, offset, iobytes, bp->b_blkno);
561 
562 		VOP_STRATEGY(devvp, bp);
563 	}
564 
565 loopdone:
566 	nestiobuf_done(mbp, skipbytes, error);
567 	if (async) {
568 		UVMHIST_LOG(ubchist, "returning 0 (async)",0,0,0,0);
569 		rw_exit(&gp->g_glock);
570 		error = 0;
571 		goto out_err;
572 	}
573 	if (bp != NULL) {
574 		error = biowait(mbp);
575 	}
576 
577 	/* Remove the mapping (make KVA available as soon as possible) */
578 	uvm_pagermapout(kva, npages);
579 
580 	/*
581 	 * if this we encountered a hole then we have to do a little more work.
582 	 * for read faults, we marked the page PG_RDONLY so that future
583 	 * write accesses to the page will fault again.
584 	 * for write faults, we must make sure that the backing store for
585 	 * the page is completely allocated while the pages are locked.
586 	 */
587 
588 	if (!error && sawhole && blockalloc) {
589 		/*
590 		 * XXX: This assumes that we come here only via
591 		 * the mmio path
592 		 */
593 		if (vp->v_mount->mnt_wapbl) {
594 			error = WAPBL_BEGIN(vp->v_mount);
595 		}
596 
597 		if (!error) {
598 			error = GOP_ALLOC(vp, startoffset,
599 			    npages << PAGE_SHIFT, 0, cred);
600 			if (vp->v_mount->mnt_wapbl) {
601 				WAPBL_END(vp->v_mount);
602 			}
603 		}
604 
605 		UVMHIST_LOG(ubchist, "gop_alloc off 0x%x/0x%x -> %d",
606 		    startoffset, npages << PAGE_SHIFT, error,0);
607 		if (!error) {
608 			for (i = 0; i < npages; i++) {
609 				if (pgs[i] == NULL) {
610 					continue;
611 				}
612 				pgs[i]->flags &= ~(PG_CLEAN|PG_RDONLY);
613 				UVMHIST_LOG(ubchist, "mark dirty pg %p",
614 				    pgs[i],0,0,0);
615 			}
616 		}
617 	}
618 	rw_exit(&gp->g_glock);
619 
620 	putiobuf(mbp);
621 
622 	mutex_enter(&uobj->vmobjlock);
623 
624 	/*
625 	 * we're almost done!  release the pages...
626 	 * for errors, we free the pages.
627 	 * otherwise we activate them and mark them as valid and clean.
628 	 * also, unbusy pages that were not actually requested.
629 	 */
630 
631 	if (error) {
632 		for (i = 0; i < npages; i++) {
633 			if (pgs[i] == NULL) {
634 				continue;
635 			}
636 			UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x",
637 			    pgs[i], pgs[i]->flags, 0,0);
638 			if (pgs[i]->flags & PG_FAKE) {
639 				pgs[i]->flags |= PG_RELEASED;
640 			}
641 		}
642 		mutex_enter(&uvm_pageqlock);
643 		uvm_page_unbusy(pgs, npages);
644 		mutex_exit(&uvm_pageqlock);
645 		mutex_exit(&uobj->vmobjlock);
646 		UVMHIST_LOG(ubchist, "returning error %d", error,0,0,0);
647 		goto out_err;
648 	}
649 
650 out:
651 	UVMHIST_LOG(ubchist, "succeeding, npages %d", npages,0,0,0);
652 	error = 0;
653 	mutex_enter(&uvm_pageqlock);
654 	for (i = 0; i < npages; i++) {
655 		pg = pgs[i];
656 		if (pg == NULL) {
657 			continue;
658 		}
659 		UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x",
660 		    pg, pg->flags, 0,0);
661 		if (pg->flags & PG_FAKE && !overwrite) {
662 			pg->flags &= ~(PG_FAKE);
663 			pmap_clear_modify(pgs[i]);
664 		}
665 		KASSERT(!write || !blockalloc || (pg->flags & PG_RDONLY) == 0);
666 		if (i < ridx || i >= ridx + orignpages || async) {
667 			UVMHIST_LOG(ubchist, "unbusy pg %p offset 0x%x",
668 			    pg, pg->offset,0,0);
669 			if (pg->flags & PG_WANTED) {
670 				wakeup(pg);
671 			}
672 			if (pg->flags & PG_FAKE) {
673 				KASSERT(overwrite);
674 				uvm_pagezero(pg);
675 			}
676 			if (pg->flags & PG_RELEASED) {
677 				uvm_pagefree(pg);
678 				continue;
679 			}
680 			uvm_pageenqueue(pg);
681 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
682 			UVM_PAGE_OWN(pg, NULL);
683 		}
684 	}
685 	mutex_exit(&uvm_pageqlock);
686 	mutex_exit(&uobj->vmobjlock);
687 	if (ap->a_m != NULL) {
688 		memcpy(ap->a_m, &pgs[ridx],
689 		    orignpages * sizeof(struct vm_page *));
690 	}
691 
692 out_err:
693 	if (pgs != NULL && pgs != pgs_onstack)
694 		kmem_free(pgs, pgs_size);
695 	if (has_trans)
696 		fstrans_done(vp->v_mount);
697 	return (error);
698 }
699 
700 /*
701  * generic VM putpages routine.
702  * Write the given range of pages to backing store.
703  *
704  * => "offhi == 0" means flush all pages at or after "offlo".
705  * => object should be locked by caller.  we return with the
706  *      object unlocked.
707  * => if PGO_CLEANIT or PGO_SYNCIO is set, we may block (due to I/O).
708  *	thus, a caller might want to unlock higher level resources
709  *	(e.g. vm_map) before calling flush.
710  * => if neither PGO_CLEANIT nor PGO_SYNCIO is set, we will not block
711  * => if PGO_ALLPAGES is set, then all pages in the object will be processed.
712  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
713  *	that new pages are inserted on the tail end of the list.   thus,
714  *	we can make a complete pass through the object in one go by starting
715  *	at the head and working towards the tail (new pages are put in
716  *	front of us).
717  * => NOTE: we are allowed to lock the page queues, so the caller
718  *	must not be holding the page queue lock.
719  *
720  * note on "cleaning" object and PG_BUSY pages:
721  *	this routine is holding the lock on the object.   the only time
722  *	that it can run into a PG_BUSY page that it does not own is if
723  *	some other process has started I/O on the page (e.g. either
724  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
725  *	in, then it can not be dirty (!PG_CLEAN) because no one has
726  *	had a chance to modify it yet.    if the PG_BUSY page is being
727  *	paged out then it means that someone else has already started
728  *	cleaning the page for us (how nice!).    in this case, if we
729  *	have syncio specified, then after we make our pass through the
730  *	object we need to wait for the other PG_BUSY pages to clear
731  *	off (i.e. we need to do an iosync).   also note that once a
732  *	page is PG_BUSY it must stay in its object until it is un-busyed.
733  *
734  * note on page traversal:
735  *	we can traverse the pages in an object either by going down the
736  *	linked list in "uobj->memq", or we can go over the address range
737  *	by page doing hash table lookups for each address.    depending
738  *	on how many pages are in the object it may be cheaper to do one
739  *	or the other.   we set "by_list" to true if we are using memq.
740  *	if the cost of a hash lookup was equal to the cost of the list
741  *	traversal we could compare the number of pages in the start->stop
742  *	range to the total number of pages in the object.   however, it
743  *	seems that a hash table lookup is more expensive than the linked
744  *	list traversal, so we multiply the number of pages in the
745  *	range by an estimate of the relatively higher cost of the hash lookup.
746  */
747 
748 int
749 genfs_putpages(void *v)
750 {
751 	struct vop_putpages_args /* {
752 		struct vnode *a_vp;
753 		voff_t a_offlo;
754 		voff_t a_offhi;
755 		int a_flags;
756 	} */ *ap = v;
757 
758 	return genfs_do_putpages(ap->a_vp, ap->a_offlo, ap->a_offhi,
759 	    ap->a_flags, NULL);
760 }
761 
762 int
763 genfs_do_putpages(struct vnode *vp, off_t startoff, off_t endoff,
764     int origflags, struct vm_page **busypg)
765 {
766 	struct uvm_object *uobj = &vp->v_uobj;
767 	kmutex_t *slock = &uobj->vmobjlock;
768 	off_t off;
769 	/* Even for strange MAXPHYS, the shift rounds down to a page */
770 #define maxpages (MAXPHYS >> PAGE_SHIFT)
771 	int i, error, npages, nback;
772 	int freeflag;
773 	struct vm_page *pgs[maxpages], *pg, *nextpg, *tpg, curmp, endmp;
774 	bool wasclean, by_list, needs_clean, yld;
775 	bool async = (origflags & PGO_SYNCIO) == 0;
776 	bool pagedaemon = curlwp == uvm.pagedaemon_lwp;
777 	struct lwp *l = curlwp ? curlwp : &lwp0;
778 	struct genfs_node *gp = VTOG(vp);
779 	int flags;
780 	int dirtygen;
781 	bool modified;
782 	bool need_wapbl;
783 	bool has_trans;
784 	bool cleanall;
785 	bool onworklst;
786 
787 	UVMHIST_FUNC("genfs_putpages"); UVMHIST_CALLED(ubchist);
788 
789 	KASSERT(origflags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE));
790 	KASSERT((startoff & PAGE_MASK) == 0 && (endoff & PAGE_MASK) == 0);
791 	KASSERT(startoff < endoff || endoff == 0);
792 
793 	UVMHIST_LOG(ubchist, "vp %p pages %d off 0x%x len 0x%x",
794 	    vp, uobj->uo_npages, startoff, endoff - startoff);
795 
796 	has_trans = false;
797 	need_wapbl = (!pagedaemon && vp->v_mount && vp->v_mount->mnt_wapbl &&
798 	    (origflags & PGO_JOURNALLOCKED) == 0);
799 
800 retry:
801 	modified = false;
802 	flags = origflags;
803 	KASSERT((vp->v_iflag & VI_ONWORKLST) != 0 ||
804 	    (vp->v_iflag & VI_WRMAPDIRTY) == 0);
805 	if (uobj->uo_npages == 0) {
806 		if (vp->v_iflag & VI_ONWORKLST) {
807 			vp->v_iflag &= ~VI_WRMAPDIRTY;
808 			if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL)
809 				vn_syncer_remove_from_worklist(vp);
810 		}
811 		if (has_trans) {
812 			if (need_wapbl)
813 				WAPBL_END(vp->v_mount);
814 			fstrans_done(vp->v_mount);
815 		}
816 		mutex_exit(slock);
817 		return (0);
818 	}
819 
820 	/*
821 	 * the vnode has pages, set up to process the request.
822 	 */
823 
824 	if (!has_trans && (flags & PGO_CLEANIT) != 0) {
825 		mutex_exit(slock);
826 		if (pagedaemon) {
827 			error = fstrans_start_nowait(vp->v_mount, FSTRANS_LAZY);
828 			if (error)
829 				return error;
830 		} else
831 			fstrans_start(vp->v_mount, FSTRANS_LAZY);
832 		if (need_wapbl) {
833 			error = WAPBL_BEGIN(vp->v_mount);
834 			if (error) {
835 				fstrans_done(vp->v_mount);
836 				return error;
837 			}
838 		}
839 		has_trans = true;
840 		mutex_enter(slock);
841 		goto retry;
842 	}
843 
844 	error = 0;
845 	wasclean = (vp->v_numoutput == 0);
846 	off = startoff;
847 	if (endoff == 0 || flags & PGO_ALLPAGES) {
848 		endoff = trunc_page(LLONG_MAX);
849 	}
850 	by_list = (uobj->uo_npages <=
851 	    ((endoff - startoff) >> PAGE_SHIFT) * UVM_PAGE_TREE_PENALTY);
852 
853 #if !defined(DEBUG)
854 	/*
855 	 * if this vnode is known not to have dirty pages,
856 	 * don't bother to clean it out.
857 	 */
858 
859 	if ((vp->v_iflag & VI_ONWORKLST) == 0) {
860 		if ((flags & (PGO_FREE|PGO_DEACTIVATE)) == 0) {
861 			goto skip_scan;
862 		}
863 		flags &= ~PGO_CLEANIT;
864 	}
865 #endif /* !defined(DEBUG) */
866 
867 	/*
868 	 * start the loop.  when scanning by list, hold the last page
869 	 * in the list before we start.  pages allocated after we start
870 	 * will be added to the end of the list, so we can stop at the
871 	 * current last page.
872 	 */
873 
874 	cleanall = (flags & PGO_CLEANIT) != 0 && wasclean &&
875 	    startoff == 0 && endoff == trunc_page(LLONG_MAX) &&
876 	    (vp->v_iflag & VI_ONWORKLST) != 0;
877 	dirtygen = gp->g_dirtygen;
878 	freeflag = pagedaemon ? PG_PAGEOUT : PG_RELEASED;
879 	if (by_list) {
880 		curmp.uobject = uobj;
881 		curmp.offset = (voff_t)-1;
882 		curmp.flags = PG_BUSY;
883 		endmp.uobject = uobj;
884 		endmp.offset = (voff_t)-1;
885 		endmp.flags = PG_BUSY;
886 		pg = TAILQ_FIRST(&uobj->memq);
887 		TAILQ_INSERT_TAIL(&uobj->memq, &endmp, listq.queue);
888 		uvm_lwp_hold(l);
889 	} else {
890 		pg = uvm_pagelookup(uobj, off);
891 	}
892 	nextpg = NULL;
893 	while (by_list || off < endoff) {
894 
895 		/*
896 		 * if the current page is not interesting, move on to the next.
897 		 */
898 
899 		KASSERT(pg == NULL || pg->uobject == uobj);
900 		KASSERT(pg == NULL ||
901 		    (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 ||
902 		    (pg->flags & PG_BUSY) != 0);
903 		if (by_list) {
904 			if (pg == &endmp) {
905 				break;
906 			}
907 			if (pg->offset < startoff || pg->offset >= endoff ||
908 			    pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
909 				if (pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
910 					wasclean = false;
911 				}
912 				pg = TAILQ_NEXT(pg, listq.queue);
913 				continue;
914 			}
915 			off = pg->offset;
916 		} else if (pg == NULL || pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
917 			if (pg != NULL) {
918 				wasclean = false;
919 			}
920 			off += PAGE_SIZE;
921 			if (off < endoff) {
922 				pg = uvm_pagelookup(uobj, off);
923 			}
924 			continue;
925 		}
926 
927 		/*
928 		 * if the current page needs to be cleaned and it's busy,
929 		 * wait for it to become unbusy.
930 		 */
931 
932 		yld = (l->l_cpu->ci_schedstate.spc_flags &
933 		    SPCF_SHOULDYIELD) && !pagedaemon;
934 		if (pg->flags & PG_BUSY || yld) {
935 			UVMHIST_LOG(ubchist, "busy %p", pg,0,0,0);
936 			if (flags & PGO_BUSYFAIL && pg->flags & PG_BUSY) {
937 				UVMHIST_LOG(ubchist, "busyfail %p", pg, 0,0,0);
938 				error = EDEADLK;
939 				if (busypg != NULL)
940 					*busypg = pg;
941 				break;
942 			}
943 			if (pagedaemon) {
944 				/*
945 				 * someone has taken the page while we
946 				 * dropped the lock for fstrans_start.
947 				 */
948 				break;
949 			}
950 			if (by_list) {
951 				TAILQ_INSERT_BEFORE(pg, &curmp, listq.queue);
952 				UVMHIST_LOG(ubchist, "curmp next %p",
953 				    TAILQ_NEXT(&curmp, listq.queue), 0,0,0);
954 			}
955 			if (yld) {
956 				mutex_exit(slock);
957 				preempt();
958 				mutex_enter(slock);
959 			} else {
960 				pg->flags |= PG_WANTED;
961 				UVM_UNLOCK_AND_WAIT(pg, slock, 0, "genput", 0);
962 				mutex_enter(slock);
963 			}
964 			if (by_list) {
965 				UVMHIST_LOG(ubchist, "after next %p",
966 				    TAILQ_NEXT(&curmp, listq.queue), 0,0,0);
967 				pg = TAILQ_NEXT(&curmp, listq.queue);
968 				TAILQ_REMOVE(&uobj->memq, &curmp, listq.queue);
969 			} else {
970 				pg = uvm_pagelookup(uobj, off);
971 			}
972 			continue;
973 		}
974 
975 		/*
976 		 * if we're freeing, remove all mappings of the page now.
977 		 * if we're cleaning, check if the page is needs to be cleaned.
978 		 */
979 
980 		if (flags & PGO_FREE) {
981 			pmap_page_protect(pg, VM_PROT_NONE);
982 		} else if (flags & PGO_CLEANIT) {
983 
984 			/*
985 			 * if we still have some hope to pull this vnode off
986 			 * from the syncer queue, write-protect the page.
987 			 */
988 
989 			if (cleanall && wasclean &&
990 			    gp->g_dirtygen == dirtygen) {
991 
992 				/*
993 				 * uobj pages get wired only by uvm_fault
994 				 * where uobj is locked.
995 				 */
996 
997 				if (pg->wire_count == 0) {
998 					pmap_page_protect(pg,
999 					    VM_PROT_READ|VM_PROT_EXECUTE);
1000 				} else {
1001 					cleanall = false;
1002 				}
1003 			}
1004 		}
1005 
1006 		if (flags & PGO_CLEANIT) {
1007 			needs_clean = pmap_clear_modify(pg) ||
1008 			    (pg->flags & PG_CLEAN) == 0;
1009 			pg->flags |= PG_CLEAN;
1010 		} else {
1011 			needs_clean = false;
1012 		}
1013 
1014 		/*
1015 		 * if we're cleaning, build a cluster.
1016 		 * the cluster will consist of pages which are currently dirty,
1017 		 * but they will be returned to us marked clean.
1018 		 * if not cleaning, just operate on the one page.
1019 		 */
1020 
1021 		if (needs_clean) {
1022 			KDASSERT((vp->v_iflag & VI_ONWORKLST));
1023 			wasclean = false;
1024 			memset(pgs, 0, sizeof(pgs));
1025 			pg->flags |= PG_BUSY;
1026 			UVM_PAGE_OWN(pg, "genfs_putpages");
1027 
1028 			/*
1029 			 * first look backward.
1030 			 */
1031 
1032 			npages = MIN(maxpages >> 1, off >> PAGE_SHIFT);
1033 			nback = npages;
1034 			uvn_findpages(uobj, off - PAGE_SIZE, &nback, &pgs[0],
1035 			    UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY|UFP_BACKWARD);
1036 			if (nback) {
1037 				memmove(&pgs[0], &pgs[npages - nback],
1038 				    nback * sizeof(pgs[0]));
1039 				if (npages - nback < nback)
1040 					memset(&pgs[nback], 0,
1041 					    (npages - nback) * sizeof(pgs[0]));
1042 				else
1043 					memset(&pgs[npages - nback], 0,
1044 					    nback * sizeof(pgs[0]));
1045 			}
1046 
1047 			/*
1048 			 * then plug in our page of interest.
1049 			 */
1050 
1051 			pgs[nback] = pg;
1052 
1053 			/*
1054 			 * then look forward to fill in the remaining space in
1055 			 * the array of pages.
1056 			 */
1057 
1058 			npages = maxpages - nback - 1;
1059 			uvn_findpages(uobj, off + PAGE_SIZE, &npages,
1060 			    &pgs[nback + 1],
1061 			    UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY);
1062 			npages += nback + 1;
1063 		} else {
1064 			pgs[0] = pg;
1065 			npages = 1;
1066 			nback = 0;
1067 		}
1068 
1069 		/*
1070 		 * apply FREE or DEACTIVATE options if requested.
1071 		 */
1072 
1073 		if (flags & (PGO_DEACTIVATE|PGO_FREE)) {
1074 			mutex_enter(&uvm_pageqlock);
1075 		}
1076 		for (i = 0; i < npages; i++) {
1077 			tpg = pgs[i];
1078 			KASSERT(tpg->uobject == uobj);
1079 			if (by_list && tpg == TAILQ_NEXT(pg, listq.queue))
1080 				pg = tpg;
1081 			if (tpg->offset < startoff || tpg->offset >= endoff)
1082 				continue;
1083 			if (flags & PGO_DEACTIVATE && tpg->wire_count == 0) {
1084 				uvm_pagedeactivate(tpg);
1085 			} else if (flags & PGO_FREE) {
1086 				pmap_page_protect(tpg, VM_PROT_NONE);
1087 				if (tpg->flags & PG_BUSY) {
1088 					tpg->flags |= freeflag;
1089 					if (pagedaemon) {
1090 						uvm_pageout_start(1);
1091 						uvm_pagedequeue(tpg);
1092 					}
1093 				} else {
1094 
1095 					/*
1096 					 * ``page is not busy''
1097 					 * implies that npages is 1
1098 					 * and needs_clean is false.
1099 					 */
1100 
1101 					nextpg = TAILQ_NEXT(tpg, listq.queue);
1102 					uvm_pagefree(tpg);
1103 					if (pagedaemon)
1104 						uvmexp.pdfreed++;
1105 				}
1106 			}
1107 		}
1108 		if (flags & (PGO_DEACTIVATE|PGO_FREE)) {
1109 			mutex_exit(&uvm_pageqlock);
1110 		}
1111 		if (needs_clean) {
1112 			modified = true;
1113 
1114 			/*
1115 			 * start the i/o.  if we're traversing by list,
1116 			 * keep our place in the list with a marker page.
1117 			 */
1118 
1119 			if (by_list) {
1120 				TAILQ_INSERT_AFTER(&uobj->memq, pg, &curmp,
1121 				    listq.queue);
1122 			}
1123 			mutex_exit(slock);
1124 			error = GOP_WRITE(vp, pgs, npages, flags);
1125 			mutex_enter(slock);
1126 			if (by_list) {
1127 				pg = TAILQ_NEXT(&curmp, listq.queue);
1128 				TAILQ_REMOVE(&uobj->memq, &curmp, listq.queue);
1129 			}
1130 			if (error) {
1131 				break;
1132 			}
1133 			if (by_list) {
1134 				continue;
1135 			}
1136 		}
1137 
1138 		/*
1139 		 * find the next page and continue if there was no error.
1140 		 */
1141 
1142 		if (by_list) {
1143 			if (nextpg) {
1144 				pg = nextpg;
1145 				nextpg = NULL;
1146 			} else {
1147 				pg = TAILQ_NEXT(pg, listq.queue);
1148 			}
1149 		} else {
1150 			off += (npages - nback) << PAGE_SHIFT;
1151 			if (off < endoff) {
1152 				pg = uvm_pagelookup(uobj, off);
1153 			}
1154 		}
1155 	}
1156 	if (by_list) {
1157 		TAILQ_REMOVE(&uobj->memq, &endmp, listq.queue);
1158 		uvm_lwp_rele(l);
1159 	}
1160 
1161 	if (modified && (vp->v_iflag & VI_WRMAPDIRTY) != 0 &&
1162 	    (vp->v_type != VBLK ||
1163 	    (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) {
1164 		GOP_MARKUPDATE(vp, GOP_UPDATE_MODIFIED);
1165 	}
1166 
1167 	/*
1168 	 * if we're cleaning and there was nothing to clean,
1169 	 * take us off the syncer list.  if we started any i/o
1170 	 * and we're doing sync i/o, wait for all writes to finish.
1171 	 */
1172 
1173 	if (cleanall && wasclean && gp->g_dirtygen == dirtygen &&
1174 	    (vp->v_iflag & VI_ONWORKLST) != 0) {
1175 #if defined(DEBUG)
1176 		TAILQ_FOREACH(pg, &uobj->memq, listq.queue) {
1177 			if ((pg->flags & PG_CLEAN) == 0) {
1178 				printf("%s: %p: !CLEAN\n", __func__, pg);
1179 			}
1180 			if (pmap_is_modified(pg)) {
1181 				printf("%s: %p: modified\n", __func__, pg);
1182 			}
1183 		}
1184 #endif /* defined(DEBUG) */
1185 		vp->v_iflag &= ~VI_WRMAPDIRTY;
1186 		if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL)
1187 			vn_syncer_remove_from_worklist(vp);
1188 	}
1189 
1190 #if !defined(DEBUG)
1191 skip_scan:
1192 #endif /* !defined(DEBUG) */
1193 
1194 	/* Wait for output to complete. */
1195 	if (!wasclean && !async && vp->v_numoutput != 0) {
1196 		while (vp->v_numoutput != 0)
1197 			cv_wait(&vp->v_cv, slock);
1198 	}
1199 	onworklst = (vp->v_iflag & VI_ONWORKLST) != 0;
1200 	mutex_exit(slock);
1201 
1202 	if ((flags & PGO_RECLAIM) != 0 && onworklst) {
1203 		/*
1204 		 * in the case of PGO_RECLAIM, ensure to make the vnode clean.
1205 		 * retrying is not a big deal because, in many cases,
1206 		 * uobj->uo_npages is already 0 here.
1207 		 */
1208 		mutex_enter(slock);
1209 		goto retry;
1210 	}
1211 
1212 	if (has_trans) {
1213 		if (need_wapbl)
1214 			WAPBL_END(vp->v_mount);
1215 		fstrans_done(vp->v_mount);
1216 	}
1217 
1218 	return (error);
1219 }
1220 
1221 int
1222 genfs_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, int flags)
1223 {
1224 	off_t off;
1225 	vaddr_t kva;
1226 	size_t len;
1227 	int error;
1228 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1229 
1230 	UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x",
1231 	    vp, pgs, npages, flags);
1232 
1233 	off = pgs[0]->offset;
1234 	kva = uvm_pagermapin(pgs, npages,
1235 	    UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK);
1236 	len = npages << PAGE_SHIFT;
1237 
1238 	error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE,
1239 			    uvm_aio_biodone);
1240 
1241 	return error;
1242 }
1243 
1244 int
1245 genfs_gop_write_rwmap(struct vnode *vp, struct vm_page **pgs, int npages, int flags)
1246 {
1247 	off_t off;
1248 	vaddr_t kva;
1249 	size_t len;
1250 	int error;
1251 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1252 
1253 	UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x",
1254 	    vp, pgs, npages, flags);
1255 
1256 	off = pgs[0]->offset;
1257 	kva = uvm_pagermapin(pgs, npages,
1258 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
1259 	len = npages << PAGE_SHIFT;
1260 
1261 	error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE,
1262 			    uvm_aio_biodone);
1263 
1264 	return error;
1265 }
1266 
1267 /*
1268  * Backend routine for doing I/O to vnode pages.  Pages are already locked
1269  * and mapped into kernel memory.  Here we just look up the underlying
1270  * device block addresses and call the strategy routine.
1271  */
1272 
1273 static int
1274 genfs_do_io(struct vnode *vp, off_t off, vaddr_t kva, size_t len, int flags,
1275     enum uio_rw rw, void (*iodone)(struct buf *))
1276 {
1277 	int s, error, run;
1278 	int fs_bshift, dev_bshift;
1279 	off_t eof, offset, startoffset;
1280 	size_t bytes, iobytes, skipbytes;
1281 	daddr_t lbn, blkno;
1282 	struct buf *mbp, *bp;
1283 	struct vnode *devvp;
1284 	bool async = (flags & PGO_SYNCIO) == 0;
1285 	bool write = rw == UIO_WRITE;
1286 	int brw = write ? B_WRITE : B_READ;
1287 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1288 
1289 	UVMHIST_LOG(ubchist, "vp %p kva %p len 0x%x flags 0x%x",
1290 	    vp, kva, len, flags);
1291 
1292 	KASSERT(vp->v_size <= vp->v_writesize);
1293 	GOP_SIZE(vp, vp->v_writesize, &eof, 0);
1294 	if (vp->v_type != VBLK) {
1295 		fs_bshift = vp->v_mount->mnt_fs_bshift;
1296 		dev_bshift = vp->v_mount->mnt_dev_bshift;
1297 	} else {
1298 		fs_bshift = DEV_BSHIFT;
1299 		dev_bshift = DEV_BSHIFT;
1300 	}
1301 	error = 0;
1302 	startoffset = off;
1303 	bytes = MIN(len, eof - startoffset);
1304 	skipbytes = 0;
1305 	KASSERT(bytes != 0);
1306 
1307 	if (write) {
1308 		mutex_enter(&vp->v_interlock);
1309 		vp->v_numoutput += 2;
1310 		mutex_exit(&vp->v_interlock);
1311 	}
1312 	mbp = getiobuf(vp, true);
1313 	UVMHIST_LOG(ubchist, "vp %p mbp %p num now %d bytes 0x%x",
1314 	    vp, mbp, vp->v_numoutput, bytes);
1315 	mbp->b_bufsize = len;
1316 	mbp->b_data = (void *)kva;
1317 	mbp->b_resid = mbp->b_bcount = bytes;
1318 	mbp->b_cflags = BC_BUSY | BC_AGE;
1319 	if (async) {
1320 		mbp->b_flags = brw | B_ASYNC;
1321 		mbp->b_iodone = iodone;
1322 	} else {
1323 		mbp->b_flags = brw;
1324 		mbp->b_iodone = NULL;
1325 	}
1326 	if (curlwp == uvm.pagedaemon_lwp)
1327 		BIO_SETPRIO(mbp, BPRIO_TIMELIMITED);
1328 	else if (async)
1329 		BIO_SETPRIO(mbp, BPRIO_TIMENONCRITICAL);
1330 	else
1331 		BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL);
1332 
1333 	bp = NULL;
1334 	for (offset = startoffset;
1335 	    bytes > 0;
1336 	    offset += iobytes, bytes -= iobytes) {
1337 		lbn = offset >> fs_bshift;
1338 		error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run);
1339 		if (error) {
1340 			UVMHIST_LOG(ubchist, "VOP_BMAP() -> %d", error,0,0,0);
1341 			skipbytes += bytes;
1342 			bytes = 0;
1343 			break;
1344 		}
1345 
1346 		iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset,
1347 		    bytes);
1348 		if (blkno == (daddr_t)-1) {
1349 			if (!write) {
1350 				memset((char *)kva + (offset - startoffset), 0,
1351 				   iobytes);
1352 			}
1353 			skipbytes += iobytes;
1354 			continue;
1355 		}
1356 
1357 		/* if it's really one i/o, don't make a second buf */
1358 		if (offset == startoffset && iobytes == bytes) {
1359 			bp = mbp;
1360 		} else {
1361 			UVMHIST_LOG(ubchist, "vp %p bp %p num now %d",
1362 			    vp, bp, vp->v_numoutput, 0);
1363 			bp = getiobuf(vp, true);
1364 			nestiobuf_setup(mbp, bp, offset - startoffset, iobytes);
1365 		}
1366 		bp->b_lblkno = 0;
1367 
1368 		/* adjust physical blkno for partial blocks */
1369 		bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >>
1370 		    dev_bshift);
1371 		UVMHIST_LOG(ubchist,
1372 		    "vp %p offset 0x%x bcount 0x%x blkno 0x%x",
1373 		    vp, offset, bp->b_bcount, bp->b_blkno);
1374 
1375 		VOP_STRATEGY(devvp, bp);
1376 	}
1377 	if (skipbytes) {
1378 		UVMHIST_LOG(ubchist, "skipbytes %d", skipbytes, 0,0,0);
1379 	}
1380 	nestiobuf_done(mbp, skipbytes, error);
1381 	if (async) {
1382 		UVMHIST_LOG(ubchist, "returning 0 (async)", 0,0,0,0);
1383 		return (0);
1384 	}
1385 	UVMHIST_LOG(ubchist, "waiting for mbp %p", mbp,0,0,0);
1386 	error = biowait(mbp);
1387 	s = splbio();
1388 	(*iodone)(mbp);
1389 	splx(s);
1390 	UVMHIST_LOG(ubchist, "returning, error %d", error,0,0,0);
1391 	return (error);
1392 }
1393 
1394 int
1395 genfs_compat_getpages(void *v)
1396 {
1397 	struct vop_getpages_args /* {
1398 		struct vnode *a_vp;
1399 		voff_t a_offset;
1400 		struct vm_page **a_m;
1401 		int *a_count;
1402 		int a_centeridx;
1403 		vm_prot_t a_access_type;
1404 		int a_advice;
1405 		int a_flags;
1406 	} */ *ap = v;
1407 
1408 	off_t origoffset;
1409 	struct vnode *vp = ap->a_vp;
1410 	struct uvm_object *uobj = &vp->v_uobj;
1411 	struct vm_page *pg, **pgs;
1412 	vaddr_t kva;
1413 	int i, error, orignpages, npages;
1414 	struct iovec iov;
1415 	struct uio uio;
1416 	kauth_cred_t cred = curlwp->l_cred;
1417 	bool write = (ap->a_access_type & VM_PROT_WRITE) != 0;
1418 
1419 	error = 0;
1420 	origoffset = ap->a_offset;
1421 	orignpages = *ap->a_count;
1422 	pgs = ap->a_m;
1423 
1424 	if (write && (vp->v_iflag & VI_ONWORKLST) == 0) {
1425 		vn_syncer_add_to_worklist(vp, filedelay);
1426 	}
1427 	if (ap->a_flags & PGO_LOCKED) {
1428 		uvn_findpages(uobj, origoffset, ap->a_count, ap->a_m,
1429 		    UFP_NOWAIT|UFP_NOALLOC| (write ? UFP_NORDONLY : 0));
1430 
1431 		return (ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0);
1432 	}
1433 	if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= vp->v_size) {
1434 		mutex_exit(&uobj->vmobjlock);
1435 		return (EINVAL);
1436 	}
1437 	if ((ap->a_flags & PGO_SYNCIO) == 0) {
1438 		mutex_exit(&uobj->vmobjlock);
1439 		return 0;
1440 	}
1441 	npages = orignpages;
1442 	uvn_findpages(uobj, origoffset, &npages, pgs, UFP_ALL);
1443 	mutex_exit(&uobj->vmobjlock);
1444 	kva = uvm_pagermapin(pgs, npages,
1445 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
1446 	for (i = 0; i < npages; i++) {
1447 		pg = pgs[i];
1448 		if ((pg->flags & PG_FAKE) == 0) {
1449 			continue;
1450 		}
1451 		iov.iov_base = (char *)kva + (i << PAGE_SHIFT);
1452 		iov.iov_len = PAGE_SIZE;
1453 		uio.uio_iov = &iov;
1454 		uio.uio_iovcnt = 1;
1455 		uio.uio_offset = origoffset + (i << PAGE_SHIFT);
1456 		uio.uio_rw = UIO_READ;
1457 		uio.uio_resid = PAGE_SIZE;
1458 		UIO_SETUP_SYSSPACE(&uio);
1459 		/* XXX vn_lock */
1460 		error = VOP_READ(vp, &uio, 0, cred);
1461 		if (error) {
1462 			break;
1463 		}
1464 		if (uio.uio_resid) {
1465 			memset(iov.iov_base, 0, uio.uio_resid);
1466 		}
1467 	}
1468 	uvm_pagermapout(kva, npages);
1469 	mutex_enter(&uobj->vmobjlock);
1470 	mutex_enter(&uvm_pageqlock);
1471 	for (i = 0; i < npages; i++) {
1472 		pg = pgs[i];
1473 		if (error && (pg->flags & PG_FAKE) != 0) {
1474 			pg->flags |= PG_RELEASED;
1475 		} else {
1476 			pmap_clear_modify(pg);
1477 			uvm_pageactivate(pg);
1478 		}
1479 	}
1480 	if (error) {
1481 		uvm_page_unbusy(pgs, npages);
1482 	}
1483 	mutex_exit(&uvm_pageqlock);
1484 	mutex_exit(&uobj->vmobjlock);
1485 	return (error);
1486 }
1487 
1488 int
1489 genfs_compat_gop_write(struct vnode *vp, struct vm_page **pgs, int npages,
1490     int flags)
1491 {
1492 	off_t offset;
1493 	struct iovec iov;
1494 	struct uio uio;
1495 	kauth_cred_t cred = curlwp->l_cred;
1496 	struct buf *bp;
1497 	vaddr_t kva;
1498 	int error;
1499 
1500 	offset = pgs[0]->offset;
1501 	kva = uvm_pagermapin(pgs, npages,
1502 	    UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK);
1503 
1504 	iov.iov_base = (void *)kva;
1505 	iov.iov_len = npages << PAGE_SHIFT;
1506 	uio.uio_iov = &iov;
1507 	uio.uio_iovcnt = 1;
1508 	uio.uio_offset = offset;
1509 	uio.uio_rw = UIO_WRITE;
1510 	uio.uio_resid = npages << PAGE_SHIFT;
1511 	UIO_SETUP_SYSSPACE(&uio);
1512 	/* XXX vn_lock */
1513 	error = VOP_WRITE(vp, &uio, 0, cred);
1514 
1515 	mutex_enter(&vp->v_interlock);
1516 	vp->v_numoutput++;
1517 	mutex_exit(&vp->v_interlock);
1518 
1519 	bp = getiobuf(vp, true);
1520 	bp->b_cflags = BC_BUSY | BC_AGE;
1521 	bp->b_lblkno = offset >> vp->v_mount->mnt_fs_bshift;
1522 	bp->b_data = (char *)kva;
1523 	bp->b_bcount = npages << PAGE_SHIFT;
1524 	bp->b_bufsize = npages << PAGE_SHIFT;
1525 	bp->b_resid = 0;
1526 	bp->b_error = error;
1527 	uvm_aio_aiodone(bp);
1528 	return (error);
1529 }
1530 
1531 /*
1532  * Process a uio using direct I/O.  If we reach a part of the request
1533  * which cannot be processed in this fashion for some reason, just return.
1534  * The caller must handle some additional part of the request using
1535  * buffered I/O before trying direct I/O again.
1536  */
1537 
1538 void
1539 genfs_directio(struct vnode *vp, struct uio *uio, int ioflag)
1540 {
1541 	struct vmspace *vs;
1542 	struct iovec *iov;
1543 	vaddr_t va;
1544 	size_t len;
1545 	const int mask = DEV_BSIZE - 1;
1546 	int error;
1547 	bool need_wapbl = (vp->v_mount && vp->v_mount->mnt_wapbl &&
1548 	    (ioflag & IO_JOURNALLOCKED) == 0);
1549 
1550 	/*
1551 	 * We only support direct I/O to user space for now.
1552 	 */
1553 
1554 	if (VMSPACE_IS_KERNEL_P(uio->uio_vmspace)) {
1555 		return;
1556 	}
1557 
1558 	/*
1559 	 * If the vnode is mapped, we would need to get the getpages lock
1560 	 * to stabilize the bmap, but then we would get into trouble whil e
1561 	 * locking the pages if the pages belong to this same vnode (or a
1562 	 * multi-vnode cascade to the same effect).  Just fall back to
1563 	 * buffered I/O if the vnode is mapped to avoid this mess.
1564 	 */
1565 
1566 	if (vp->v_vflag & VV_MAPPED) {
1567 		return;
1568 	}
1569 
1570 	if (need_wapbl) {
1571 		error = WAPBL_BEGIN(vp->v_mount);
1572 		if (error)
1573 			return;
1574 	}
1575 
1576 	/*
1577 	 * Do as much of the uio as possible with direct I/O.
1578 	 */
1579 
1580 	vs = uio->uio_vmspace;
1581 	while (uio->uio_resid) {
1582 		iov = uio->uio_iov;
1583 		if (iov->iov_len == 0) {
1584 			uio->uio_iov++;
1585 			uio->uio_iovcnt--;
1586 			continue;
1587 		}
1588 		va = (vaddr_t)iov->iov_base;
1589 		len = MIN(iov->iov_len, genfs_maxdio);
1590 		len &= ~mask;
1591 
1592 		/*
1593 		 * If the next chunk is smaller than DEV_BSIZE or extends past
1594 		 * the current EOF, then fall back to buffered I/O.
1595 		 */
1596 
1597 		if (len == 0 || uio->uio_offset + len > vp->v_size) {
1598 			break;
1599 		}
1600 
1601 		/*
1602 		 * Check alignment.  The file offset must be at least
1603 		 * sector-aligned.  The exact constraint on memory alignment
1604 		 * is very hardware-dependent, but requiring sector-aligned
1605 		 * addresses there too is safe.
1606 		 */
1607 
1608 		if (uio->uio_offset & mask || va & mask) {
1609 			break;
1610 		}
1611 		error = genfs_do_directio(vs, va, len, vp, uio->uio_offset,
1612 					  uio->uio_rw);
1613 		if (error) {
1614 			break;
1615 		}
1616 		iov->iov_base = (char *)iov->iov_base + len;
1617 		iov->iov_len -= len;
1618 		uio->uio_offset += len;
1619 		uio->uio_resid -= len;
1620 	}
1621 
1622 	if (need_wapbl)
1623 		WAPBL_END(vp->v_mount);
1624 }
1625 
1626 /*
1627  * Iodone routine for direct I/O.  We don't do much here since the request is
1628  * always synchronous, so the caller will do most of the work after biowait().
1629  */
1630 
1631 static void
1632 genfs_dio_iodone(struct buf *bp)
1633 {
1634 
1635 	KASSERT((bp->b_flags & B_ASYNC) == 0);
1636 	if ((bp->b_flags & B_READ) == 0 && (bp->b_cflags & BC_AGE) != 0) {
1637 		mutex_enter(bp->b_objlock);
1638 		vwakeup(bp);
1639 		mutex_exit(bp->b_objlock);
1640 	}
1641 	putiobuf(bp);
1642 }
1643 
1644 /*
1645  * Process one chunk of a direct I/O request.
1646  */
1647 
1648 static int
1649 genfs_do_directio(struct vmspace *vs, vaddr_t uva, size_t len, struct vnode *vp,
1650     off_t off, enum uio_rw rw)
1651 {
1652 	struct vm_map *map;
1653 	struct pmap *upm, *kpm;
1654 	size_t klen = round_page(uva + len) - trunc_page(uva);
1655 	off_t spoff, epoff;
1656 	vaddr_t kva, puva;
1657 	paddr_t pa;
1658 	vm_prot_t prot;
1659 	int error, rv, poff, koff;
1660 	const int pgoflags = PGO_CLEANIT | PGO_SYNCIO | PGO_JOURNALLOCKED |
1661 		(rw == UIO_WRITE ? PGO_FREE : 0);
1662 
1663 	/*
1664 	 * For writes, verify that this range of the file already has fully
1665 	 * allocated backing store.  If there are any holes, just punt and
1666 	 * make the caller take the buffered write path.
1667 	 */
1668 
1669 	if (rw == UIO_WRITE) {
1670 		daddr_t lbn, elbn, blkno;
1671 		int bsize, bshift, run;
1672 
1673 		bshift = vp->v_mount->mnt_fs_bshift;
1674 		bsize = 1 << bshift;
1675 		lbn = off >> bshift;
1676 		elbn = (off + len + bsize - 1) >> bshift;
1677 		while (lbn < elbn) {
1678 			error = VOP_BMAP(vp, lbn, NULL, &blkno, &run);
1679 			if (error) {
1680 				return error;
1681 			}
1682 			if (blkno == (daddr_t)-1) {
1683 				return ENOSPC;
1684 			}
1685 			lbn += 1 + run;
1686 		}
1687 	}
1688 
1689 	/*
1690 	 * Flush any cached pages for parts of the file that we're about to
1691 	 * access.  If we're writing, invalidate pages as well.
1692 	 */
1693 
1694 	spoff = trunc_page(off);
1695 	epoff = round_page(off + len);
1696 	mutex_enter(&vp->v_interlock);
1697 	error = VOP_PUTPAGES(vp, spoff, epoff, pgoflags);
1698 	if (error) {
1699 		return error;
1700 	}
1701 
1702 	/*
1703 	 * Wire the user pages and remap them into kernel memory.
1704 	 */
1705 
1706 	prot = rw == UIO_READ ? VM_PROT_READ | VM_PROT_WRITE : VM_PROT_READ;
1707 	error = uvm_vslock(vs, (void *)uva, len, prot);
1708 	if (error) {
1709 		return error;
1710 	}
1711 
1712 	map = &vs->vm_map;
1713 	upm = vm_map_pmap(map);
1714 	kpm = vm_map_pmap(kernel_map);
1715 	kva = uvm_km_alloc(kernel_map, klen, 0,
1716 			   UVM_KMF_VAONLY | UVM_KMF_WAITVA);
1717 	puva = trunc_page(uva);
1718 	for (poff = 0; poff < klen; poff += PAGE_SIZE) {
1719 		rv = pmap_extract(upm, puva + poff, &pa);
1720 		KASSERT(rv);
1721 		pmap_enter(kpm, kva + poff, pa, prot, prot | PMAP_WIRED);
1722 	}
1723 	pmap_update(kpm);
1724 
1725 	/*
1726 	 * Do the I/O.
1727 	 */
1728 
1729 	koff = uva - trunc_page(uva);
1730 	error = genfs_do_io(vp, off, kva + koff, len, PGO_SYNCIO, rw,
1731 			    genfs_dio_iodone);
1732 
1733 	/*
1734 	 * Tear down the kernel mapping.
1735 	 */
1736 
1737 	pmap_remove(kpm, kva, kva + klen);
1738 	pmap_update(kpm);
1739 	uvm_km_free(kernel_map, kva, klen, UVM_KMF_VAONLY);
1740 
1741 	/*
1742 	 * Unwire the user pages.
1743 	 */
1744 
1745 	uvm_vsunlock(vs, (void *)uva, len);
1746 	return error;
1747 }
1748 
1749