xref: /netbsd-src/sys/miscfs/genfs/genfs_io.c (revision 8b0f9554ff8762542c4defc4f70e1eb76fb508fa)
1 /*	$NetBSD: genfs_io.c,v 1.1 2007/10/17 16:45:00 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 1.1 2007/10/17 16:45:00 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 
51 #include <miscfs/genfs/genfs.h>
52 #include <miscfs/genfs/genfs_node.h>
53 #include <miscfs/specfs/specdev.h>
54 
55 #include <uvm/uvm.h>
56 #include <uvm/uvm_pager.h>
57 
58 static int genfs_do_directio(struct vmspace *, vaddr_t, size_t, struct vnode *,
59     off_t, enum uio_rw);
60 static void genfs_dio_iodone(struct buf *);
61 
62 static int genfs_do_io(struct vnode *, off_t, vaddr_t, size_t, int, enum uio_rw,
63     void (*)(struct buf *));
64 static inline void genfs_rel_pages(struct vm_page **, int);
65 
66 #define MAX_READ_PAGES	16 	/* XXXUBC 16 */
67 
68 int genfs_maxdio = MAXPHYS;
69 
70 static inline void
71 genfs_rel_pages(struct vm_page **pgs, int npages)
72 {
73 	int i;
74 
75 	for (i = 0; i < npages; i++) {
76 		struct vm_page *pg = pgs[i];
77 
78 		if (pg == NULL || pg == PGO_DONTCARE)
79 			continue;
80 		if (pg->flags & PG_FAKE) {
81 			pg->flags |= PG_RELEASED;
82 		}
83 	}
84 	uvm_lock_pageq();
85 	uvm_page_unbusy(pgs, npages);
86 	uvm_unlock_pageq();
87 }
88 
89 /*
90  * generic VM getpages routine.
91  * Return PG_BUSY pages for the given range,
92  * reading from backing store if necessary.
93  */
94 
95 int
96 genfs_getpages(void *v)
97 {
98 	struct vop_getpages_args /* {
99 		struct vnode *a_vp;
100 		voff_t a_offset;
101 		struct vm_page **a_m;
102 		int *a_count;
103 		int a_centeridx;
104 		vm_prot_t a_access_type;
105 		int a_advice;
106 		int a_flags;
107 	} */ *ap = v;
108 
109 	off_t newsize, diskeof, memeof;
110 	off_t offset, origoffset, startoffset, endoffset;
111 	daddr_t lbn, blkno;
112 	int i, error, npages, orignpages, npgs, run, ridx, pidx, pcount;
113 	int fs_bshift, fs_bsize, dev_bshift;
114 	int flags = ap->a_flags;
115 	size_t bytes, iobytes, tailstart, tailbytes, totalbytes, skipbytes;
116 	vaddr_t kva;
117 	struct buf *bp, *mbp;
118 	struct vnode *vp = ap->a_vp;
119 	struct vnode *devvp;
120 	struct genfs_node *gp = VTOG(vp);
121 	struct uvm_object *uobj = &vp->v_uobj;
122 	struct vm_page *pg, **pgs, *pgs_onstack[MAX_READ_PAGES];
123 	int pgs_size;
124 	kauth_cred_t cred = curlwp->l_cred;		/* XXXUBC curlwp */
125 	bool async = (flags & PGO_SYNCIO) == 0;
126 	bool write = (ap->a_access_type & VM_PROT_WRITE) != 0;
127 	bool sawhole = false;
128 	bool has_trans = false;
129 	bool overwrite = (flags & PGO_OVERWRITE) != 0;
130 	bool blockalloc = write && (flags & PGO_NOBLOCKALLOC) == 0;
131 	voff_t origvsize;
132 	UVMHIST_FUNC("genfs_getpages"); UVMHIST_CALLED(ubchist);
133 
134 	UVMHIST_LOG(ubchist, "vp %p off 0x%x/%x count %d",
135 	    vp, ap->a_offset >> 32, ap->a_offset, *ap->a_count);
136 
137 	KASSERT(vp->v_type == VREG || vp->v_type == VDIR ||
138 	    vp->v_type == VLNK || vp->v_type == VBLK);
139 
140 	/* XXXUBC temp limit */
141 	if (*ap->a_count > MAX_READ_PAGES) {
142 		panic("genfs_getpages: too many pages");
143 	}
144 
145 	pgs = pgs_onstack;
146 	pgs_size = sizeof(pgs_onstack);
147 
148 startover:
149 	error = 0;
150 	origvsize = vp->v_size;
151 	origoffset = ap->a_offset;
152 	orignpages = *ap->a_count;
153 	GOP_SIZE(vp, origvsize, &diskeof, 0);
154 	if (flags & PGO_PASTEOF) {
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", __func__);
166 		}
167 #endif /* defined(DIAGNOSTIC) */
168 	} else {
169 		GOP_SIZE(vp, origvsize, &memeof, GOP_SIZE_MEM);
170 	}
171 	KASSERT(ap->a_centeridx >= 0 || ap->a_centeridx <= orignpages);
172 	KASSERT((origoffset & (PAGE_SIZE - 1)) == 0 && origoffset >= 0);
173 	KASSERT(orignpages > 0);
174 
175 	/*
176 	 * Bounds-check the request.
177 	 */
178 
179 	if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= memeof) {
180 		if ((flags & PGO_LOCKED) == 0) {
181 			simple_unlock(&uobj->vmobjlock);
182 		}
183 		UVMHIST_LOG(ubchist, "off 0x%x count %d goes past EOF 0x%x",
184 		    origoffset, *ap->a_count, memeof,0);
185 		error = EINVAL;
186 		goto out_err;
187 	}
188 
189 	/* uobj is locked */
190 
191 	if ((flags & PGO_NOTIMESTAMP) == 0 &&
192 	    (vp->v_type != VBLK ||
193 	    (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) {
194 		int updflags = 0;
195 
196 		if ((vp->v_mount->mnt_flag & MNT_NOATIME) == 0) {
197 			updflags = GOP_UPDATE_ACCESSED;
198 		}
199 		if (write) {
200 			updflags |= GOP_UPDATE_MODIFIED;
201 		}
202 		if (updflags != 0) {
203 			GOP_MARKUPDATE(vp, updflags);
204 		}
205 	}
206 
207 	if (write) {
208 		gp->g_dirtygen++;
209 		if ((vp->v_iflag & VI_ONWORKLST) == 0) {
210 			vn_syncer_add_to_worklist(vp, filedelay);
211 		}
212 		if ((vp->v_iflag & (VI_WRMAP|VI_WRMAPDIRTY)) == VI_WRMAP) {
213 			vp->v_iflag |= VI_WRMAPDIRTY;
214 		}
215 	}
216 
217 	/*
218 	 * For PGO_LOCKED requests, just return whatever's in memory.
219 	 */
220 
221 	if (flags & PGO_LOCKED) {
222 		int nfound;
223 
224 		npages = *ap->a_count;
225 #if defined(DEBUG)
226 		for (i = 0; i < npages; i++) {
227 			pg = ap->a_m[i];
228 			KASSERT(pg == NULL || pg == PGO_DONTCARE);
229 		}
230 #endif /* defined(DEBUG) */
231 		nfound = uvn_findpages(uobj, origoffset, &npages,
232 		    ap->a_m, UFP_NOWAIT|UFP_NOALLOC|(write ? UFP_NORDONLY : 0));
233 		KASSERT(npages == *ap->a_count);
234 		if (nfound == 0) {
235 			error = EBUSY;
236 			goto out_err;
237 		}
238 		if (!rw_tryenter(&gp->g_glock, RW_READER)) {
239 			genfs_rel_pages(ap->a_m, npages);
240 
241 			/*
242 			 * restore the array.
243 			 */
244 
245 			for (i = 0; i < npages; i++) {
246 				pg = ap->a_m[i];
247 
248 				if (pg != NULL || pg != PGO_DONTCARE) {
249 					ap->a_m[i] = NULL;
250 				}
251 			}
252 		} else {
253 			rw_exit(&gp->g_glock);
254 		}
255 		error = (ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0);
256 		goto out_err;
257 	}
258 	simple_unlock(&uobj->vmobjlock);
259 
260 	/*
261 	 * find the requested pages and make some simple checks.
262 	 * leave space in the page array for a whole block.
263 	 */
264 
265 	if (vp->v_type != VBLK) {
266 		fs_bshift = vp->v_mount->mnt_fs_bshift;
267 		dev_bshift = vp->v_mount->mnt_dev_bshift;
268 	} else {
269 		fs_bshift = DEV_BSHIFT;
270 		dev_bshift = DEV_BSHIFT;
271 	}
272 	fs_bsize = 1 << fs_bshift;
273 
274 	orignpages = MIN(orignpages,
275 	    round_page(memeof - origoffset) >> PAGE_SHIFT);
276 	npages = orignpages;
277 	startoffset = origoffset & ~(fs_bsize - 1);
278 	endoffset = round_page((origoffset + (npages << PAGE_SHIFT) +
279 	    fs_bsize - 1) & ~(fs_bsize - 1));
280 	endoffset = MIN(endoffset, round_page(memeof));
281 	ridx = (origoffset - startoffset) >> PAGE_SHIFT;
282 
283 	pgs_size = sizeof(struct vm_page *) *
284 	    ((endoffset - startoffset) >> PAGE_SHIFT);
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 		memset(pgs, 0, pgs_size);
295 	}
296 	UVMHIST_LOG(ubchist, "ridx %d npages %d startoff %ld endoff %ld",
297 	    ridx, npages, startoffset, endoffset);
298 
299 	if (!has_trans) {
300 		fstrans_start(vp->v_mount, FSTRANS_SHARED);
301 		has_trans = true;
302 	}
303 
304 	/*
305 	 * hold g_glock to prevent a race with truncate.
306 	 *
307 	 * check if our idea of v_size is still valid.
308 	 */
309 
310 	if (blockalloc) {
311 		rw_enter(&gp->g_glock, RW_WRITER);
312 	} else {
313 		rw_enter(&gp->g_glock, RW_READER);
314 	}
315 	simple_lock(&uobj->vmobjlock);
316 	if (vp->v_size < origvsize) {
317 		rw_exit(&gp->g_glock);
318 		if (pgs != pgs_onstack)
319 			kmem_free(pgs, pgs_size);
320 		goto startover;
321 	}
322 
323 	if (uvn_findpages(uobj, origoffset, &npages, &pgs[ridx],
324 	    async ? UFP_NOWAIT : UFP_ALL) != orignpages) {
325 		rw_exit(&gp->g_glock);
326 		KASSERT(async != 0);
327 		genfs_rel_pages(&pgs[ridx], orignpages);
328 		simple_unlock(&uobj->vmobjlock);
329 		error = EBUSY;
330 		goto out_err;
331 	}
332 
333 	/*
334 	 * if the pages are already resident, just return them.
335 	 */
336 
337 	for (i = 0; i < npages; i++) {
338 		struct vm_page *pg1 = pgs[ridx + i];
339 
340 		if ((pg1->flags & PG_FAKE) ||
341 		    (blockalloc && (pg1->flags & PG_RDONLY))) {
342 			break;
343 		}
344 	}
345 	if (i == npages) {
346 		rw_exit(&gp->g_glock);
347 		UVMHIST_LOG(ubchist, "returning cached pages", 0,0,0,0);
348 		npages += ridx;
349 		goto out;
350 	}
351 
352 	/*
353 	 * if PGO_OVERWRITE is set, don't bother reading the pages.
354 	 */
355 
356 	if (overwrite) {
357 		rw_exit(&gp->g_glock);
358 		UVMHIST_LOG(ubchist, "PGO_OVERWRITE",0,0,0,0);
359 
360 		for (i = 0; i < npages; i++) {
361 			struct vm_page *pg1 = pgs[ridx + i];
362 
363 			pg1->flags &= ~(PG_RDONLY|PG_CLEAN);
364 		}
365 		npages += ridx;
366 		goto out;
367 	}
368 
369 	/*
370 	 * the page wasn't resident and we're not overwriting,
371 	 * so we're going to have to do some i/o.
372 	 * find any additional pages needed to cover the expanded range.
373 	 */
374 
375 	npages = (endoffset - startoffset) >> PAGE_SHIFT;
376 	if (startoffset != origoffset || npages != orignpages) {
377 
378 		/*
379 		 * we need to avoid deadlocks caused by locking
380 		 * additional pages at lower offsets than pages we
381 		 * already have locked.  unlock them all and start over.
382 		 */
383 
384 		genfs_rel_pages(&pgs[ridx], orignpages);
385 		memset(pgs, 0, pgs_size);
386 
387 		UVMHIST_LOG(ubchist, "reset npages start 0x%x end 0x%x",
388 		    startoffset, endoffset, 0,0);
389 		npgs = npages;
390 		if (uvn_findpages(uobj, startoffset, &npgs, pgs,
391 		    async ? UFP_NOWAIT : UFP_ALL) != npages) {
392 			rw_exit(&gp->g_glock);
393 			KASSERT(async != 0);
394 			genfs_rel_pages(pgs, npages);
395 			simple_unlock(&uobj->vmobjlock);
396 			error = EBUSY;
397 			goto out_err;
398 		}
399 	}
400 	simple_unlock(&uobj->vmobjlock);
401 
402 	/*
403 	 * read the desired page(s).
404 	 */
405 
406 	totalbytes = npages << PAGE_SHIFT;
407 	bytes = MIN(totalbytes, MAX(diskeof - startoffset, 0));
408 	tailbytes = totalbytes - bytes;
409 	skipbytes = 0;
410 
411 	kva = uvm_pagermapin(pgs, npages,
412 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
413 
414 	mbp = getiobuf();
415 	mbp->b_bufsize = totalbytes;
416 	mbp->b_data = (void *)kva;
417 	mbp->b_resid = mbp->b_bcount = bytes;
418 	mbp->b_flags = B_BUSY|B_READ| (async ? B_CALL|B_ASYNC : 0);
419 	mbp->b_iodone = (async ? uvm_aio_biodone : 0);
420 	mbp->b_vp = vp;
421 	if (async)
422 		BIO_SETPRIO(mbp, BPRIO_TIMELIMITED);
423 	else
424 		BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL);
425 
426 	/*
427 	 * if EOF is in the middle of the range, zero the part past EOF.
428 	 * skip over pages which are not PG_FAKE since in that case they have
429 	 * valid data that we need to preserve.
430 	 */
431 
432 	tailstart = bytes;
433 	while (tailbytes > 0) {
434 		const int len = PAGE_SIZE - (tailstart & PAGE_MASK);
435 
436 		KASSERT(len <= tailbytes);
437 		if ((pgs[tailstart >> PAGE_SHIFT]->flags & PG_FAKE) != 0) {
438 			memset((void *)(kva + tailstart), 0, len);
439 			UVMHIST_LOG(ubchist, "tailbytes %p 0x%x 0x%x",
440 			    kva, tailstart, len, 0);
441 		}
442 		tailstart += len;
443 		tailbytes -= len;
444 	}
445 
446 	/*
447 	 * now loop over the pages, reading as needed.
448 	 */
449 
450 	bp = NULL;
451 	for (offset = startoffset;
452 	    bytes > 0;
453 	    offset += iobytes, bytes -= iobytes) {
454 
455 		/*
456 		 * skip pages which don't need to be read.
457 		 */
458 
459 		pidx = (offset - startoffset) >> PAGE_SHIFT;
460 		while ((pgs[pidx]->flags & PG_FAKE) == 0) {
461 			size_t b;
462 
463 			KASSERT((offset & (PAGE_SIZE - 1)) == 0);
464 			if ((pgs[pidx]->flags & PG_RDONLY)) {
465 				sawhole = true;
466 			}
467 			b = MIN(PAGE_SIZE, bytes);
468 			offset += b;
469 			bytes -= b;
470 			skipbytes += b;
471 			pidx++;
472 			UVMHIST_LOG(ubchist, "skipping, new offset 0x%x",
473 			    offset, 0,0,0);
474 			if (bytes == 0) {
475 				goto loopdone;
476 			}
477 		}
478 
479 		/*
480 		 * bmap the file to find out the blkno to read from and
481 		 * how much we can read in one i/o.  if bmap returns an error,
482 		 * skip the rest of the top-level i/o.
483 		 */
484 
485 		lbn = offset >> fs_bshift;
486 		error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run);
487 		if (error) {
488 			UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%x -> %d\n",
489 			    lbn, error,0,0);
490 			skipbytes += bytes;
491 			goto loopdone;
492 		}
493 
494 		/*
495 		 * see how many pages can be read with this i/o.
496 		 * reduce the i/o size if necessary to avoid
497 		 * overwriting pages with valid data.
498 		 */
499 
500 		iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset,
501 		    bytes);
502 		if (offset + iobytes > round_page(offset)) {
503 			pcount = 1;
504 			while (pidx + pcount < npages &&
505 			    pgs[pidx + pcount]->flags & PG_FAKE) {
506 				pcount++;
507 			}
508 			iobytes = MIN(iobytes, (pcount << PAGE_SHIFT) -
509 			    (offset - trunc_page(offset)));
510 		}
511 
512 		/*
513 		 * if this block isn't allocated, zero it instead of
514 		 * reading it.  unless we are going to allocate blocks,
515 		 * mark the pages we zeroed PG_RDONLY.
516 		 */
517 
518 		if (blkno < 0) {
519 			int holepages = (round_page(offset + iobytes) -
520 			    trunc_page(offset)) >> PAGE_SHIFT;
521 			UVMHIST_LOG(ubchist, "lbn 0x%x -> HOLE", lbn,0,0,0);
522 
523 			sawhole = true;
524 			memset((char *)kva + (offset - startoffset), 0,
525 			    iobytes);
526 			skipbytes += iobytes;
527 
528 			for (i = 0; i < holepages; i++) {
529 				if (write) {
530 					pgs[pidx + i]->flags &= ~PG_CLEAN;
531 				}
532 				if (!blockalloc) {
533 					pgs[pidx + i]->flags |= PG_RDONLY;
534 				}
535 			}
536 			continue;
537 		}
538 
539 		/*
540 		 * allocate a sub-buf for this piece of the i/o
541 		 * (or just use mbp if there's only 1 piece),
542 		 * and start it going.
543 		 */
544 
545 		if (offset == startoffset && iobytes == bytes) {
546 			bp = mbp;
547 		} else {
548 			bp = getiobuf();
549 			nestiobuf_setup(mbp, bp, offset - startoffset, iobytes);
550 		}
551 		bp->b_lblkno = 0;
552 
553 		/* adjust physical blkno for partial blocks */
554 		bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >>
555 		    dev_bshift);
556 
557 		UVMHIST_LOG(ubchist,
558 		    "bp %p offset 0x%x bcount 0x%x blkno 0x%x",
559 		    bp, offset, iobytes, bp->b_blkno);
560 
561 		VOP_STRATEGY(devvp, bp);
562 	}
563 
564 loopdone:
565 	nestiobuf_done(mbp, skipbytes, error);
566 	if (async) {
567 		UVMHIST_LOG(ubchist, "returning 0 (async)",0,0,0,0);
568 		rw_exit(&gp->g_glock);
569 		error = 0;
570 		goto out_err;
571 	}
572 	if (bp != NULL) {
573 		error = biowait(mbp);
574 	}
575 	putiobuf(mbp);
576 	uvm_pagermapout(kva, npages);
577 
578 	/*
579 	 * if this we encountered a hole then we have to do a little more work.
580 	 * for read faults, we marked the page PG_RDONLY so that future
581 	 * write accesses to the page will fault again.
582 	 * for write faults, we must make sure that the backing store for
583 	 * the page is completely allocated while the pages are locked.
584 	 */
585 
586 	if (!error && sawhole && blockalloc) {
587 		error = GOP_ALLOC(vp, startoffset, npages << PAGE_SHIFT, 0,
588 		    cred);
589 		UVMHIST_LOG(ubchist, "gop_alloc off 0x%x/0x%x -> %d",
590 		    startoffset, npages << PAGE_SHIFT, error,0);
591 		if (!error) {
592 			for (i = 0; i < npages; i++) {
593 				if (pgs[i] == NULL) {
594 					continue;
595 				}
596 				pgs[i]->flags &= ~(PG_CLEAN|PG_RDONLY);
597 				UVMHIST_LOG(ubchist, "mark dirty pg %p",
598 				    pgs[i],0,0,0);
599 			}
600 		}
601 	}
602 	rw_exit(&gp->g_glock);
603 	simple_lock(&uobj->vmobjlock);
604 
605 	/*
606 	 * we're almost done!  release the pages...
607 	 * for errors, we free the pages.
608 	 * otherwise we activate them and mark them as valid and clean.
609 	 * also, unbusy pages that were not actually requested.
610 	 */
611 
612 	if (error) {
613 		for (i = 0; i < npages; i++) {
614 			if (pgs[i] == NULL) {
615 				continue;
616 			}
617 			UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x",
618 			    pgs[i], pgs[i]->flags, 0,0);
619 			if (pgs[i]->flags & PG_FAKE) {
620 				pgs[i]->flags |= PG_RELEASED;
621 			}
622 		}
623 		uvm_lock_pageq();
624 		uvm_page_unbusy(pgs, npages);
625 		uvm_unlock_pageq();
626 		simple_unlock(&uobj->vmobjlock);
627 		UVMHIST_LOG(ubchist, "returning error %d", error,0,0,0);
628 		goto out_err;
629 	}
630 
631 out:
632 	UVMHIST_LOG(ubchist, "succeeding, npages %d", npages,0,0,0);
633 	error = 0;
634 	uvm_lock_pageq();
635 	for (i = 0; i < npages; i++) {
636 		pg = pgs[i];
637 		if (pg == NULL) {
638 			continue;
639 		}
640 		UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x",
641 		    pg, pg->flags, 0,0);
642 		if (pg->flags & PG_FAKE && !overwrite) {
643 			pg->flags &= ~(PG_FAKE);
644 			pmap_clear_modify(pgs[i]);
645 		}
646 		KASSERT(!write || !blockalloc || (pg->flags & PG_RDONLY) == 0);
647 		if (i < ridx || i >= ridx + orignpages || async) {
648 			UVMHIST_LOG(ubchist, "unbusy pg %p offset 0x%x",
649 			    pg, pg->offset,0,0);
650 			if (pg->flags & PG_WANTED) {
651 				wakeup(pg);
652 			}
653 			if (pg->flags & PG_FAKE) {
654 				KASSERT(overwrite);
655 				uvm_pagezero(pg);
656 			}
657 			if (pg->flags & PG_RELEASED) {
658 				uvm_pagefree(pg);
659 				continue;
660 			}
661 			uvm_pageenqueue(pg);
662 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
663 			UVM_PAGE_OWN(pg, NULL);
664 		}
665 	}
666 	uvm_unlock_pageq();
667 	simple_unlock(&uobj->vmobjlock);
668 	if (ap->a_m != NULL) {
669 		memcpy(ap->a_m, &pgs[ridx],
670 		    orignpages * sizeof(struct vm_page *));
671 	}
672 
673 out_err:
674 	if (pgs != pgs_onstack)
675 		kmem_free(pgs, pgs_size);
676 	if (has_trans)
677 		fstrans_done(vp->v_mount);
678 	return (error);
679 }
680 
681 /*
682  * generic VM putpages routine.
683  * Write the given range of pages to backing store.
684  *
685  * => "offhi == 0" means flush all pages at or after "offlo".
686  * => object should be locked by caller.  we return with the
687  *      object unlocked.
688  * => if PGO_CLEANIT or PGO_SYNCIO is set, we may block (due to I/O).
689  *	thus, a caller might want to unlock higher level resources
690  *	(e.g. vm_map) before calling flush.
691  * => if neither PGO_CLEANIT nor PGO_SYNCIO is set, we will not block
692  * => if PGO_ALLPAGES is set, then all pages in the object will be processed.
693  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
694  *	that new pages are inserted on the tail end of the list.   thus,
695  *	we can make a complete pass through the object in one go by starting
696  *	at the head and working towards the tail (new pages are put in
697  *	front of us).
698  * => NOTE: we are allowed to lock the page queues, so the caller
699  *	must not be holding the page queue lock.
700  *
701  * note on "cleaning" object and PG_BUSY pages:
702  *	this routine is holding the lock on the object.   the only time
703  *	that it can run into a PG_BUSY page that it does not own is if
704  *	some other process has started I/O on the page (e.g. either
705  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
706  *	in, then it can not be dirty (!PG_CLEAN) because no one has
707  *	had a chance to modify it yet.    if the PG_BUSY page is being
708  *	paged out then it means that someone else has already started
709  *	cleaning the page for us (how nice!).    in this case, if we
710  *	have syncio specified, then after we make our pass through the
711  *	object we need to wait for the other PG_BUSY pages to clear
712  *	off (i.e. we need to do an iosync).   also note that once a
713  *	page is PG_BUSY it must stay in its object until it is un-busyed.
714  *
715  * note on page traversal:
716  *	we can traverse the pages in an object either by going down the
717  *	linked list in "uobj->memq", or we can go over the address range
718  *	by page doing hash table lookups for each address.    depending
719  *	on how many pages are in the object it may be cheaper to do one
720  *	or the other.   we set "by_list" to true if we are using memq.
721  *	if the cost of a hash lookup was equal to the cost of the list
722  *	traversal we could compare the number of pages in the start->stop
723  *	range to the total number of pages in the object.   however, it
724  *	seems that a hash table lookup is more expensive than the linked
725  *	list traversal, so we multiply the number of pages in the
726  *	range by an estimate of the relatively higher cost of the hash lookup.
727  */
728 
729 int
730 genfs_putpages(void *v)
731 {
732 	struct vop_putpages_args /* {
733 		struct vnode *a_vp;
734 		voff_t a_offlo;
735 		voff_t a_offhi;
736 		int a_flags;
737 	} */ *ap = v;
738 
739 	return genfs_do_putpages(ap->a_vp, ap->a_offlo, ap->a_offhi,
740 	    ap->a_flags, NULL);
741 }
742 
743 int
744 genfs_do_putpages(struct vnode *vp, off_t startoff, off_t endoff, int flags,
745 	struct vm_page **busypg)
746 {
747 	struct uvm_object *uobj = &vp->v_uobj;
748 	struct simplelock *slock = &uobj->vmobjlock;
749 	off_t off;
750 	/* Even for strange MAXPHYS, the shift rounds down to a page */
751 #define maxpages (MAXPHYS >> PAGE_SHIFT)
752 	int i, s, error, npages, nback;
753 	int freeflag;
754 	struct vm_page *pgs[maxpages], *pg, *nextpg, *tpg, curmp, endmp;
755 	bool wasclean, by_list, needs_clean, yld;
756 	bool async = (flags & PGO_SYNCIO) == 0;
757 	bool pagedaemon = curlwp == uvm.pagedaemon_lwp;
758 	struct lwp *l = curlwp ? curlwp : &lwp0;
759 	struct genfs_node *gp = VTOG(vp);
760 	int dirtygen;
761 	bool modified = false;
762 	bool has_trans = false;
763 	bool cleanall;
764 
765 	UVMHIST_FUNC("genfs_putpages"); UVMHIST_CALLED(ubchist);
766 
767 	KASSERT(flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE));
768 	KASSERT((startoff & PAGE_MASK) == 0 && (endoff & PAGE_MASK) == 0);
769 	KASSERT(startoff < endoff || endoff == 0);
770 
771 	UVMHIST_LOG(ubchist, "vp %p pages %d off 0x%x len 0x%x",
772 	    vp, uobj->uo_npages, startoff, endoff - startoff);
773 
774 	KASSERT((vp->v_iflag & VI_ONWORKLST) != 0 ||
775 	    (vp->v_iflag & VI_WRMAPDIRTY) == 0);
776 	if (uobj->uo_npages == 0) {
777 		s = splbio();
778 		if (vp->v_iflag & VI_ONWORKLST) {
779 			vp->v_iflag &= ~VI_WRMAPDIRTY;
780 			if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL)
781 				vn_syncer_remove_from_worklist(vp);
782 		}
783 		splx(s);
784 		simple_unlock(slock);
785 		return (0);
786 	}
787 
788 	/*
789 	 * the vnode has pages, set up to process the request.
790 	 */
791 
792 	if ((flags & PGO_CLEANIT) != 0) {
793 		simple_unlock(slock);
794 		if (pagedaemon) {
795 			error = fstrans_start_nowait(vp->v_mount, FSTRANS_LAZY);
796 			if (error)
797 				return error;
798 		} else
799 			fstrans_start(vp->v_mount, FSTRANS_LAZY);
800 		has_trans = true;
801 		simple_lock(slock);
802 	}
803 
804 	error = 0;
805 	s = splbio();
806 	simple_lock(&global_v_numoutput_slock);
807 	wasclean = (vp->v_numoutput == 0);
808 	simple_unlock(&global_v_numoutput_slock);
809 	splx(s);
810 	off = startoff;
811 	if (endoff == 0 || flags & PGO_ALLPAGES) {
812 		endoff = trunc_page(LLONG_MAX);
813 	}
814 	by_list = (uobj->uo_npages <=
815 	    ((endoff - startoff) >> PAGE_SHIFT) * UVM_PAGE_HASH_PENALTY);
816 
817 #if !defined(DEBUG)
818 	/*
819 	 * if this vnode is known not to have dirty pages,
820 	 * don't bother to clean it out.
821 	 */
822 
823 	if ((vp->v_iflag & VI_ONWORKLST) == 0) {
824 		if ((flags & (PGO_FREE|PGO_DEACTIVATE)) == 0) {
825 			goto skip_scan;
826 		}
827 		flags &= ~PGO_CLEANIT;
828 	}
829 #endif /* !defined(DEBUG) */
830 
831 	/*
832 	 * start the loop.  when scanning by list, hold the last page
833 	 * in the list before we start.  pages allocated after we start
834 	 * will be added to the end of the list, so we can stop at the
835 	 * current last page.
836 	 */
837 
838 	cleanall = (flags & PGO_CLEANIT) != 0 && wasclean &&
839 	    startoff == 0 && endoff == trunc_page(LLONG_MAX) &&
840 	    (vp->v_iflag & VI_ONWORKLST) != 0;
841 	dirtygen = gp->g_dirtygen;
842 	freeflag = pagedaemon ? PG_PAGEOUT : PG_RELEASED;
843 	if (by_list) {
844 		curmp.uobject = uobj;
845 		curmp.offset = (voff_t)-1;
846 		curmp.flags = PG_BUSY;
847 		endmp.uobject = uobj;
848 		endmp.offset = (voff_t)-1;
849 		endmp.flags = PG_BUSY;
850 		pg = TAILQ_FIRST(&uobj->memq);
851 		TAILQ_INSERT_TAIL(&uobj->memq, &endmp, listq);
852 		uvm_lwp_hold(l);
853 	} else {
854 		pg = uvm_pagelookup(uobj, off);
855 	}
856 	nextpg = NULL;
857 	while (by_list || off < endoff) {
858 
859 		/*
860 		 * if the current page is not interesting, move on to the next.
861 		 */
862 
863 		KASSERT(pg == NULL || pg->uobject == uobj);
864 		KASSERT(pg == NULL ||
865 		    (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 ||
866 		    (pg->flags & PG_BUSY) != 0);
867 		if (by_list) {
868 			if (pg == &endmp) {
869 				break;
870 			}
871 			if (pg->offset < startoff || pg->offset >= endoff ||
872 			    pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
873 				if (pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
874 					wasclean = false;
875 				}
876 				pg = TAILQ_NEXT(pg, listq);
877 				continue;
878 			}
879 			off = pg->offset;
880 		} else if (pg == NULL || pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
881 			if (pg != NULL) {
882 				wasclean = false;
883 			}
884 			off += PAGE_SIZE;
885 			if (off < endoff) {
886 				pg = uvm_pagelookup(uobj, off);
887 			}
888 			continue;
889 		}
890 
891 		/*
892 		 * if the current page needs to be cleaned and it's busy,
893 		 * wait for it to become unbusy.
894 		 */
895 
896 		yld = (l->l_cpu->ci_schedstate.spc_flags &
897 		    SPCF_SHOULDYIELD) && !pagedaemon;
898 		if (pg->flags & PG_BUSY || yld) {
899 			UVMHIST_LOG(ubchist, "busy %p", pg,0,0,0);
900 			if (flags & PGO_BUSYFAIL && pg->flags & PG_BUSY) {
901 				UVMHIST_LOG(ubchist, "busyfail %p", pg, 0,0,0);
902 				error = EDEADLK;
903 				if (busypg != NULL)
904 					*busypg = pg;
905 				break;
906 			}
907 			if (pagedaemon) {
908 				/*
909 				 * someone has taken the page while we
910 				 * dropped the lock for fstrans_start.
911 				 */
912 				break;
913 			}
914 			if (by_list) {
915 				TAILQ_INSERT_BEFORE(pg, &curmp, listq);
916 				UVMHIST_LOG(ubchist, "curmp next %p",
917 				    TAILQ_NEXT(&curmp, listq), 0,0,0);
918 			}
919 			if (yld) {
920 				simple_unlock(slock);
921 				preempt();
922 				simple_lock(slock);
923 			} else {
924 				pg->flags |= PG_WANTED;
925 				UVM_UNLOCK_AND_WAIT(pg, slock, 0, "genput", 0);
926 				simple_lock(slock);
927 			}
928 			if (by_list) {
929 				UVMHIST_LOG(ubchist, "after next %p",
930 				    TAILQ_NEXT(&curmp, listq), 0,0,0);
931 				pg = TAILQ_NEXT(&curmp, listq);
932 				TAILQ_REMOVE(&uobj->memq, &curmp, listq);
933 			} else {
934 				pg = uvm_pagelookup(uobj, off);
935 			}
936 			continue;
937 		}
938 
939 		/*
940 		 * if we're freeing, remove all mappings of the page now.
941 		 * if we're cleaning, check if the page is needs to be cleaned.
942 		 */
943 
944 		if (flags & PGO_FREE) {
945 			pmap_page_protect(pg, VM_PROT_NONE);
946 		} else if (flags & PGO_CLEANIT) {
947 
948 			/*
949 			 * if we still have some hope to pull this vnode off
950 			 * from the syncer queue, write-protect the page.
951 			 */
952 
953 			if (cleanall && wasclean &&
954 			    gp->g_dirtygen == dirtygen) {
955 
956 				/*
957 				 * uobj pages get wired only by uvm_fault
958 				 * where uobj is locked.
959 				 */
960 
961 				if (pg->wire_count == 0) {
962 					pmap_page_protect(pg,
963 					    VM_PROT_READ|VM_PROT_EXECUTE);
964 				} else {
965 					cleanall = false;
966 				}
967 			}
968 		}
969 
970 		if (flags & PGO_CLEANIT) {
971 			needs_clean = pmap_clear_modify(pg) ||
972 			    (pg->flags & PG_CLEAN) == 0;
973 			pg->flags |= PG_CLEAN;
974 		} else {
975 			needs_clean = false;
976 		}
977 
978 		/*
979 		 * if we're cleaning, build a cluster.
980 		 * the cluster will consist of pages which are currently dirty,
981 		 * but they will be returned to us marked clean.
982 		 * if not cleaning, just operate on the one page.
983 		 */
984 
985 		if (needs_clean) {
986 			KDASSERT((vp->v_iflag & VI_ONWORKLST));
987 			wasclean = false;
988 			memset(pgs, 0, sizeof(pgs));
989 			pg->flags |= PG_BUSY;
990 			UVM_PAGE_OWN(pg, "genfs_putpages");
991 
992 			/*
993 			 * first look backward.
994 			 */
995 
996 			npages = MIN(maxpages >> 1, off >> PAGE_SHIFT);
997 			nback = npages;
998 			uvn_findpages(uobj, off - PAGE_SIZE, &nback, &pgs[0],
999 			    UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY|UFP_BACKWARD);
1000 			if (nback) {
1001 				memmove(&pgs[0], &pgs[npages - nback],
1002 				    nback * sizeof(pgs[0]));
1003 				if (npages - nback < nback)
1004 					memset(&pgs[nback], 0,
1005 					    (npages - nback) * sizeof(pgs[0]));
1006 				else
1007 					memset(&pgs[npages - nback], 0,
1008 					    nback * sizeof(pgs[0]));
1009 			}
1010 
1011 			/*
1012 			 * then plug in our page of interest.
1013 			 */
1014 
1015 			pgs[nback] = pg;
1016 
1017 			/*
1018 			 * then look forward to fill in the remaining space in
1019 			 * the array of pages.
1020 			 */
1021 
1022 			npages = maxpages - nback - 1;
1023 			uvn_findpages(uobj, off + PAGE_SIZE, &npages,
1024 			    &pgs[nback + 1],
1025 			    UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY);
1026 			npages += nback + 1;
1027 		} else {
1028 			pgs[0] = pg;
1029 			npages = 1;
1030 			nback = 0;
1031 		}
1032 
1033 		/*
1034 		 * apply FREE or DEACTIVATE options if requested.
1035 		 */
1036 
1037 		if (flags & (PGO_DEACTIVATE|PGO_FREE)) {
1038 			uvm_lock_pageq();
1039 		}
1040 		for (i = 0; i < npages; i++) {
1041 			tpg = pgs[i];
1042 			KASSERT(tpg->uobject == uobj);
1043 			if (by_list && tpg == TAILQ_NEXT(pg, listq))
1044 				pg = tpg;
1045 			if (tpg->offset < startoff || tpg->offset >= endoff)
1046 				continue;
1047 			if (flags & PGO_DEACTIVATE && tpg->wire_count == 0) {
1048 				(void) pmap_clear_reference(tpg);
1049 				uvm_pagedeactivate(tpg);
1050 			} else if (flags & PGO_FREE) {
1051 				pmap_page_protect(tpg, VM_PROT_NONE);
1052 				if (tpg->flags & PG_BUSY) {
1053 					tpg->flags |= freeflag;
1054 					if (pagedaemon) {
1055 						uvmexp.paging++;
1056 						uvm_pagedequeue(tpg);
1057 					}
1058 				} else {
1059 
1060 					/*
1061 					 * ``page is not busy''
1062 					 * implies that npages is 1
1063 					 * and needs_clean is false.
1064 					 */
1065 
1066 					nextpg = TAILQ_NEXT(tpg, listq);
1067 					uvm_pagefree(tpg);
1068 					if (pagedaemon)
1069 						uvmexp.pdfreed++;
1070 				}
1071 			}
1072 		}
1073 		if (flags & (PGO_DEACTIVATE|PGO_FREE)) {
1074 			uvm_unlock_pageq();
1075 		}
1076 		if (needs_clean) {
1077 			modified = true;
1078 
1079 			/*
1080 			 * start the i/o.  if we're traversing by list,
1081 			 * keep our place in the list with a marker page.
1082 			 */
1083 
1084 			if (by_list) {
1085 				TAILQ_INSERT_AFTER(&uobj->memq, pg, &curmp,
1086 				    listq);
1087 			}
1088 			simple_unlock(slock);
1089 			error = GOP_WRITE(vp, pgs, npages, flags);
1090 			simple_lock(slock);
1091 			if (by_list) {
1092 				pg = TAILQ_NEXT(&curmp, listq);
1093 				TAILQ_REMOVE(&uobj->memq, &curmp, listq);
1094 			}
1095 			if (error) {
1096 				break;
1097 			}
1098 			if (by_list) {
1099 				continue;
1100 			}
1101 		}
1102 
1103 		/*
1104 		 * find the next page and continue if there was no error.
1105 		 */
1106 
1107 		if (by_list) {
1108 			if (nextpg) {
1109 				pg = nextpg;
1110 				nextpg = NULL;
1111 			} else {
1112 				pg = TAILQ_NEXT(pg, listq);
1113 			}
1114 		} else {
1115 			off += (npages - nback) << PAGE_SHIFT;
1116 			if (off < endoff) {
1117 				pg = uvm_pagelookup(uobj, off);
1118 			}
1119 		}
1120 	}
1121 	if (by_list) {
1122 		TAILQ_REMOVE(&uobj->memq, &endmp, listq);
1123 		uvm_lwp_rele(l);
1124 	}
1125 
1126 	if (modified && (vp->v_iflag & VI_WRMAPDIRTY) != 0 &&
1127 	    (vp->v_type != VBLK ||
1128 	    (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) {
1129 		GOP_MARKUPDATE(vp, GOP_UPDATE_MODIFIED);
1130 	}
1131 
1132 	/*
1133 	 * if we're cleaning and there was nothing to clean,
1134 	 * take us off the syncer list.  if we started any i/o
1135 	 * and we're doing sync i/o, wait for all writes to finish.
1136 	 */
1137 
1138 	s = splbio();
1139 	if (cleanall && wasclean && gp->g_dirtygen == dirtygen &&
1140 	    (vp->v_iflag & VI_ONWORKLST) != 0) {
1141 		vp->v_iflag &= ~VI_WRMAPDIRTY;
1142 		if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL)
1143 			vn_syncer_remove_from_worklist(vp);
1144 	}
1145 	splx(s);
1146 
1147 #if !defined(DEBUG)
1148 skip_scan:
1149 #endif /* !defined(DEBUG) */
1150 	if (!wasclean && !async) {
1151 		s = splbio();
1152 		/*
1153 		 * XXX - we want simple_unlock(&global_v_numoutput_slock);
1154 		 *	 but the slot in ltsleep() is taken!
1155 		 * XXX - try to recover from missed wakeups with a timeout..
1156 		 *	 must think of something better.
1157 		 */
1158 		while (vp->v_numoutput != 0) {
1159 			vp->v_iflag |= VI_BWAIT;
1160 			UVM_UNLOCK_AND_WAIT(&vp->v_numoutput, slock, false,
1161 			    "genput2", hz);
1162 			simple_lock(slock);
1163 		}
1164 		splx(s);
1165 	}
1166 	simple_unlock(slock);
1167 
1168 	if (has_trans)
1169 		fstrans_done(vp->v_mount);
1170 
1171 	return (error);
1172 }
1173 
1174 int
1175 genfs_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, int flags)
1176 {
1177 	off_t off;
1178 	vaddr_t kva;
1179 	size_t len;
1180 	int error;
1181 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1182 
1183 	UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x",
1184 	    vp, pgs, npages, flags);
1185 
1186 	off = pgs[0]->offset;
1187 	kva = uvm_pagermapin(pgs, npages,
1188 	    UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK);
1189 	len = npages << PAGE_SHIFT;
1190 
1191 	error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE,
1192 			    uvm_aio_biodone);
1193 
1194 	return error;
1195 }
1196 
1197 /*
1198  * Backend routine for doing I/O to vnode pages.  Pages are already locked
1199  * and mapped into kernel memory.  Here we just look up the underlying
1200  * device block addresses and call the strategy routine.
1201  */
1202 
1203 static int
1204 genfs_do_io(struct vnode *vp, off_t off, vaddr_t kva, size_t len, int flags,
1205     enum uio_rw rw, void (*iodone)(struct buf *))
1206 {
1207 	int s, error, run;
1208 	int fs_bshift, dev_bshift;
1209 	off_t eof, offset, startoffset;
1210 	size_t bytes, iobytes, skipbytes;
1211 	daddr_t lbn, blkno;
1212 	struct buf *mbp, *bp;
1213 	struct vnode *devvp;
1214 	bool async = (flags & PGO_SYNCIO) == 0;
1215 	bool write = rw == UIO_WRITE;
1216 	int brw = write ? B_WRITE : B_READ;
1217 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1218 
1219 	UVMHIST_LOG(ubchist, "vp %p kva %p len 0x%x flags 0x%x",
1220 	    vp, kva, len, flags);
1221 
1222 	KASSERT(vp->v_size <= vp->v_writesize);
1223 	GOP_SIZE(vp, vp->v_writesize, &eof, 0);
1224 	if (vp->v_type != VBLK) {
1225 		fs_bshift = vp->v_mount->mnt_fs_bshift;
1226 		dev_bshift = vp->v_mount->mnt_dev_bshift;
1227 	} else {
1228 		fs_bshift = DEV_BSHIFT;
1229 		dev_bshift = DEV_BSHIFT;
1230 	}
1231 	error = 0;
1232 	startoffset = off;
1233 	bytes = MIN(len, eof - startoffset);
1234 	skipbytes = 0;
1235 	KASSERT(bytes != 0);
1236 
1237 	if (write) {
1238 		s = splbio();
1239 		simple_lock(&global_v_numoutput_slock);
1240 		vp->v_numoutput += 2;
1241 		simple_unlock(&global_v_numoutput_slock);
1242 		splx(s);
1243 	}
1244 	mbp = getiobuf();
1245 	UVMHIST_LOG(ubchist, "vp %p mbp %p num now %d bytes 0x%x",
1246 	    vp, mbp, vp->v_numoutput, bytes);
1247 	mbp->b_bufsize = len;
1248 	mbp->b_data = (void *)kva;
1249 	mbp->b_resid = mbp->b_bcount = bytes;
1250 	mbp->b_flags = B_BUSY | brw | B_AGE | (async ? (B_CALL | B_ASYNC) : 0);
1251 	mbp->b_iodone = iodone;
1252 	mbp->b_vp = vp;
1253 	if (curlwp == uvm.pagedaemon_lwp)
1254 		BIO_SETPRIO(mbp, BPRIO_TIMELIMITED);
1255 	else if (async)
1256 		BIO_SETPRIO(mbp, BPRIO_TIMENONCRITICAL);
1257 	else
1258 		BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL);
1259 
1260 	bp = NULL;
1261 	for (offset = startoffset;
1262 	    bytes > 0;
1263 	    offset += iobytes, bytes -= iobytes) {
1264 		lbn = offset >> fs_bshift;
1265 		error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run);
1266 		if (error) {
1267 			UVMHIST_LOG(ubchist, "VOP_BMAP() -> %d", error,0,0,0);
1268 			skipbytes += bytes;
1269 			bytes = 0;
1270 			break;
1271 		}
1272 
1273 		iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset,
1274 		    bytes);
1275 		if (blkno == (daddr_t)-1) {
1276 			if (!write) {
1277 				memset((char *)kva + (offset - startoffset), 0,
1278 				   iobytes);
1279 			}
1280 			skipbytes += iobytes;
1281 			continue;
1282 		}
1283 
1284 		/* if it's really one i/o, don't make a second buf */
1285 		if (offset == startoffset && iobytes == bytes) {
1286 			bp = mbp;
1287 		} else {
1288 			UVMHIST_LOG(ubchist, "vp %p bp %p num now %d",
1289 			    vp, bp, vp->v_numoutput, 0);
1290 			bp = getiobuf();
1291 			nestiobuf_setup(mbp, bp, offset - startoffset, iobytes);
1292 		}
1293 		bp->b_lblkno = 0;
1294 
1295 		/* adjust physical blkno for partial blocks */
1296 		bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >>
1297 		    dev_bshift);
1298 		UVMHIST_LOG(ubchist,
1299 		    "vp %p offset 0x%x bcount 0x%x blkno 0x%x",
1300 		    vp, offset, bp->b_bcount, bp->b_blkno);
1301 
1302 		VOP_STRATEGY(devvp, bp);
1303 	}
1304 	if (skipbytes) {
1305 		UVMHIST_LOG(ubchist, "skipbytes %d", skipbytes, 0,0,0);
1306 	}
1307 	nestiobuf_done(mbp, skipbytes, error);
1308 	if (async) {
1309 		UVMHIST_LOG(ubchist, "returning 0 (async)", 0,0,0,0);
1310 		return (0);
1311 	}
1312 	UVMHIST_LOG(ubchist, "waiting for mbp %p", mbp,0,0,0);
1313 	error = biowait(mbp);
1314 	s = splbio();
1315 	(*iodone)(mbp);
1316 	splx(s);
1317 	UVMHIST_LOG(ubchist, "returning, error %d", error,0,0,0);
1318 	return (error);
1319 }
1320 
1321 /*
1322  * VOP_PUTPAGES() for vnodes which never have pages.
1323  */
1324 
1325 int
1326 genfs_null_putpages(void *v)
1327 {
1328 	struct vop_putpages_args /* {
1329 		struct vnode *a_vp;
1330 		voff_t a_offlo;
1331 		voff_t a_offhi;
1332 		int a_flags;
1333 	} */ *ap = v;
1334 	struct vnode *vp = ap->a_vp;
1335 
1336 	KASSERT(vp->v_uobj.uo_npages == 0);
1337 	simple_unlock(&vp->v_interlock);
1338 	return (0);
1339 }
1340 
1341 int
1342 genfs_compat_getpages(void *v)
1343 {
1344 	struct vop_getpages_args /* {
1345 		struct vnode *a_vp;
1346 		voff_t a_offset;
1347 		struct vm_page **a_m;
1348 		int *a_count;
1349 		int a_centeridx;
1350 		vm_prot_t a_access_type;
1351 		int a_advice;
1352 		int a_flags;
1353 	} */ *ap = v;
1354 
1355 	off_t origoffset;
1356 	struct vnode *vp = ap->a_vp;
1357 	struct uvm_object *uobj = &vp->v_uobj;
1358 	struct vm_page *pg, **pgs;
1359 	vaddr_t kva;
1360 	int i, error, orignpages, npages;
1361 	struct iovec iov;
1362 	struct uio uio;
1363 	kauth_cred_t cred = curlwp->l_cred;
1364 	bool write = (ap->a_access_type & VM_PROT_WRITE) != 0;
1365 
1366 	error = 0;
1367 	origoffset = ap->a_offset;
1368 	orignpages = *ap->a_count;
1369 	pgs = ap->a_m;
1370 
1371 	if (write && (vp->v_iflag & VI_ONWORKLST) == 0) {
1372 		vn_syncer_add_to_worklist(vp, filedelay);
1373 	}
1374 	if (ap->a_flags & PGO_LOCKED) {
1375 		uvn_findpages(uobj, origoffset, ap->a_count, ap->a_m,
1376 		    UFP_NOWAIT|UFP_NOALLOC| (write ? UFP_NORDONLY : 0));
1377 
1378 		return (ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0);
1379 	}
1380 	if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= vp->v_size) {
1381 		simple_unlock(&uobj->vmobjlock);
1382 		return (EINVAL);
1383 	}
1384 	if ((ap->a_flags & PGO_SYNCIO) == 0) {
1385 		simple_unlock(&uobj->vmobjlock);
1386 		return 0;
1387 	}
1388 	npages = orignpages;
1389 	uvn_findpages(uobj, origoffset, &npages, pgs, UFP_ALL);
1390 	simple_unlock(&uobj->vmobjlock);
1391 	kva = uvm_pagermapin(pgs, npages,
1392 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
1393 	for (i = 0; i < npages; i++) {
1394 		pg = pgs[i];
1395 		if ((pg->flags & PG_FAKE) == 0) {
1396 			continue;
1397 		}
1398 		iov.iov_base = (char *)kva + (i << PAGE_SHIFT);
1399 		iov.iov_len = PAGE_SIZE;
1400 		uio.uio_iov = &iov;
1401 		uio.uio_iovcnt = 1;
1402 		uio.uio_offset = origoffset + (i << PAGE_SHIFT);
1403 		uio.uio_rw = UIO_READ;
1404 		uio.uio_resid = PAGE_SIZE;
1405 		UIO_SETUP_SYSSPACE(&uio);
1406 		/* XXX vn_lock */
1407 		error = VOP_READ(vp, &uio, 0, cred);
1408 		if (error) {
1409 			break;
1410 		}
1411 		if (uio.uio_resid) {
1412 			memset(iov.iov_base, 0, uio.uio_resid);
1413 		}
1414 	}
1415 	uvm_pagermapout(kva, npages);
1416 	simple_lock(&uobj->vmobjlock);
1417 	uvm_lock_pageq();
1418 	for (i = 0; i < npages; i++) {
1419 		pg = pgs[i];
1420 		if (error && (pg->flags & PG_FAKE) != 0) {
1421 			pg->flags |= PG_RELEASED;
1422 		} else {
1423 			pmap_clear_modify(pg);
1424 			uvm_pageactivate(pg);
1425 		}
1426 	}
1427 	if (error) {
1428 		uvm_page_unbusy(pgs, npages);
1429 	}
1430 	uvm_unlock_pageq();
1431 	simple_unlock(&uobj->vmobjlock);
1432 	return (error);
1433 }
1434 
1435 int
1436 genfs_compat_gop_write(struct vnode *vp, struct vm_page **pgs, int npages,
1437     int flags)
1438 {
1439 	off_t offset;
1440 	struct iovec iov;
1441 	struct uio uio;
1442 	kauth_cred_t cred = curlwp->l_cred;
1443 	struct buf *bp;
1444 	vaddr_t kva;
1445 	int s, error;
1446 
1447 	offset = pgs[0]->offset;
1448 	kva = uvm_pagermapin(pgs, npages,
1449 	    UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK);
1450 
1451 	iov.iov_base = (void *)kva;
1452 	iov.iov_len = npages << PAGE_SHIFT;
1453 	uio.uio_iov = &iov;
1454 	uio.uio_iovcnt = 1;
1455 	uio.uio_offset = offset;
1456 	uio.uio_rw = UIO_WRITE;
1457 	uio.uio_resid = npages << PAGE_SHIFT;
1458 	UIO_SETUP_SYSSPACE(&uio);
1459 	/* XXX vn_lock */
1460 	error = VOP_WRITE(vp, &uio, 0, cred);
1461 
1462 	s = splbio();
1463 	V_INCR_NUMOUTPUT(vp);
1464 	splx(s);
1465 
1466 	bp = getiobuf();
1467 	bp->b_flags = B_BUSY | B_WRITE | B_AGE;
1468 	bp->b_vp = vp;
1469 	bp->b_lblkno = offset >> vp->v_mount->mnt_fs_bshift;
1470 	bp->b_data = (char *)kva;
1471 	bp->b_bcount = npages << PAGE_SHIFT;
1472 	bp->b_bufsize = npages << PAGE_SHIFT;
1473 	bp->b_resid = 0;
1474 	bp->b_error = error;
1475 	uvm_aio_aiodone(bp);
1476 	return (error);
1477 }
1478 
1479 /*
1480  * Process a uio using direct I/O.  If we reach a part of the request
1481  * which cannot be processed in this fashion for some reason, just return.
1482  * The caller must handle some additional part of the request using
1483  * buffered I/O before trying direct I/O again.
1484  */
1485 
1486 void
1487 genfs_directio(struct vnode *vp, struct uio *uio, int ioflag)
1488 {
1489 	struct vmspace *vs;
1490 	struct iovec *iov;
1491 	vaddr_t va;
1492 	size_t len;
1493 	const int mask = DEV_BSIZE - 1;
1494 	int error;
1495 
1496 	/*
1497 	 * We only support direct I/O to user space for now.
1498 	 */
1499 
1500 	if (VMSPACE_IS_KERNEL_P(uio->uio_vmspace)) {
1501 		return;
1502 	}
1503 
1504 	/*
1505 	 * If the vnode is mapped, we would need to get the getpages lock
1506 	 * to stabilize the bmap, but then we would get into trouble whil e
1507 	 * locking the pages if the pages belong to this same vnode (or a
1508 	 * multi-vnode cascade to the same effect).  Just fall back to
1509 	 * buffered I/O if the vnode is mapped to avoid this mess.
1510 	 */
1511 
1512 	if (vp->v_vflag & VV_MAPPED) {
1513 		return;
1514 	}
1515 
1516 	/*
1517 	 * Do as much of the uio as possible with direct I/O.
1518 	 */
1519 
1520 	vs = uio->uio_vmspace;
1521 	while (uio->uio_resid) {
1522 		iov = uio->uio_iov;
1523 		if (iov->iov_len == 0) {
1524 			uio->uio_iov++;
1525 			uio->uio_iovcnt--;
1526 			continue;
1527 		}
1528 		va = (vaddr_t)iov->iov_base;
1529 		len = MIN(iov->iov_len, genfs_maxdio);
1530 		len &= ~mask;
1531 
1532 		/*
1533 		 * If the next chunk is smaller than DEV_BSIZE or extends past
1534 		 * the current EOF, then fall back to buffered I/O.
1535 		 */
1536 
1537 		if (len == 0 || uio->uio_offset + len > vp->v_size) {
1538 			return;
1539 		}
1540 
1541 		/*
1542 		 * Check alignment.  The file offset must be at least
1543 		 * sector-aligned.  The exact constraint on memory alignment
1544 		 * is very hardware-dependent, but requiring sector-aligned
1545 		 * addresses there too is safe.
1546 		 */
1547 
1548 		if (uio->uio_offset & mask || va & mask) {
1549 			return;
1550 		}
1551 		error = genfs_do_directio(vs, va, len, vp, uio->uio_offset,
1552 					  uio->uio_rw);
1553 		if (error) {
1554 			break;
1555 		}
1556 		iov->iov_base = (char *)iov->iov_base + len;
1557 		iov->iov_len -= len;
1558 		uio->uio_offset += len;
1559 		uio->uio_resid -= len;
1560 	}
1561 }
1562 
1563 /*
1564  * Iodone routine for direct I/O.  We don't do much here since the request is
1565  * always synchronous, so the caller will do most of the work after biowait().
1566  */
1567 
1568 static void
1569 genfs_dio_iodone(struct buf *bp)
1570 {
1571 	int s;
1572 
1573 	KASSERT((bp->b_flags & B_ASYNC) == 0);
1574 	s = splbio();
1575 	if ((bp->b_flags & (B_READ | B_AGE)) == B_AGE) {
1576 		vwakeup(bp);
1577 	}
1578 	putiobuf(bp);
1579 	splx(s);
1580 }
1581 
1582 /*
1583  * Process one chunk of a direct I/O request.
1584  */
1585 
1586 static int
1587 genfs_do_directio(struct vmspace *vs, vaddr_t uva, size_t len, struct vnode *vp,
1588     off_t off, enum uio_rw rw)
1589 {
1590 	struct vm_map *map;
1591 	struct pmap *upm, *kpm;
1592 	size_t klen = round_page(uva + len) - trunc_page(uva);
1593 	off_t spoff, epoff;
1594 	vaddr_t kva, puva;
1595 	paddr_t pa;
1596 	vm_prot_t prot;
1597 	int error, rv, poff, koff;
1598 	const int pgoflags = PGO_CLEANIT | PGO_SYNCIO |
1599 		(rw == UIO_WRITE ? PGO_FREE : 0);
1600 
1601 	/*
1602 	 * For writes, verify that this range of the file already has fully
1603 	 * allocated backing store.  If there are any holes, just punt and
1604 	 * make the caller take the buffered write path.
1605 	 */
1606 
1607 	if (rw == UIO_WRITE) {
1608 		daddr_t lbn, elbn, blkno;
1609 		int bsize, bshift, run;
1610 
1611 		bshift = vp->v_mount->mnt_fs_bshift;
1612 		bsize = 1 << bshift;
1613 		lbn = off >> bshift;
1614 		elbn = (off + len + bsize - 1) >> bshift;
1615 		while (lbn < elbn) {
1616 			error = VOP_BMAP(vp, lbn, NULL, &blkno, &run);
1617 			if (error) {
1618 				return error;
1619 			}
1620 			if (blkno == (daddr_t)-1) {
1621 				return ENOSPC;
1622 			}
1623 			lbn += 1 + run;
1624 		}
1625 	}
1626 
1627 	/*
1628 	 * Flush any cached pages for parts of the file that we're about to
1629 	 * access.  If we're writing, invalidate pages as well.
1630 	 */
1631 
1632 	spoff = trunc_page(off);
1633 	epoff = round_page(off + len);
1634 	simple_lock(&vp->v_interlock);
1635 	error = VOP_PUTPAGES(vp, spoff, epoff, pgoflags);
1636 	if (error) {
1637 		return error;
1638 	}
1639 
1640 	/*
1641 	 * Wire the user pages and remap them into kernel memory.
1642 	 */
1643 
1644 	prot = rw == UIO_READ ? VM_PROT_READ | VM_PROT_WRITE : VM_PROT_READ;
1645 	error = uvm_vslock(vs, (void *)uva, len, prot);
1646 	if (error) {
1647 		return error;
1648 	}
1649 
1650 	map = &vs->vm_map;
1651 	upm = vm_map_pmap(map);
1652 	kpm = vm_map_pmap(kernel_map);
1653 	kva = uvm_km_alloc(kernel_map, klen, 0,
1654 			   UVM_KMF_VAONLY | UVM_KMF_WAITVA);
1655 	puva = trunc_page(uva);
1656 	for (poff = 0; poff < klen; poff += PAGE_SIZE) {
1657 		rv = pmap_extract(upm, puva + poff, &pa);
1658 		KASSERT(rv);
1659 		pmap_enter(kpm, kva + poff, pa, prot, prot | PMAP_WIRED);
1660 	}
1661 	pmap_update(kpm);
1662 
1663 	/*
1664 	 * Do the I/O.
1665 	 */
1666 
1667 	koff = uva - trunc_page(uva);
1668 	error = genfs_do_io(vp, off, kva + koff, len, PGO_SYNCIO, rw,
1669 			    genfs_dio_iodone);
1670 
1671 	/*
1672 	 * Tear down the kernel mapping.
1673 	 */
1674 
1675 	pmap_remove(kpm, kva, kva + klen);
1676 	pmap_update(kpm);
1677 	uvm_km_free(kernel_map, kva, klen, UVM_KMF_VAONLY);
1678 
1679 	/*
1680 	 * Unwire the user pages.
1681 	 */
1682 
1683 	uvm_vsunlock(vs, (void *)uva, len);
1684 	return error;
1685 }
1686