xref: /netbsd-src/sys/miscfs/genfs/genfs_io.c (revision 3816d47b2c42fcd6e549e3407f842a5b1a1d23ad)
1 /*	$NetBSD: genfs_io.c,v 1.21 2009/10/21 21:12:06 rmind 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 1.21 2009/10/21 21:12:06 rmind 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 	} else {
889 		pg = uvm_pagelookup(uobj, off);
890 	}
891 	nextpg = NULL;
892 	while (by_list || off < endoff) {
893 
894 		/*
895 		 * if the current page is not interesting, move on to the next.
896 		 */
897 
898 		KASSERT(pg == NULL || pg->uobject == uobj);
899 		KASSERT(pg == NULL ||
900 		    (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 ||
901 		    (pg->flags & PG_BUSY) != 0);
902 		if (by_list) {
903 			if (pg == &endmp) {
904 				break;
905 			}
906 			if (pg->offset < startoff || pg->offset >= endoff ||
907 			    pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
908 				if (pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
909 					wasclean = false;
910 				}
911 				pg = TAILQ_NEXT(pg, listq.queue);
912 				continue;
913 			}
914 			off = pg->offset;
915 		} else if (pg == NULL || pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
916 			if (pg != NULL) {
917 				wasclean = false;
918 			}
919 			off += PAGE_SIZE;
920 			if (off < endoff) {
921 				pg = uvm_pagelookup(uobj, off);
922 			}
923 			continue;
924 		}
925 
926 		/*
927 		 * if the current page needs to be cleaned and it's busy,
928 		 * wait for it to become unbusy.
929 		 */
930 
931 		yld = (l->l_cpu->ci_schedstate.spc_flags &
932 		    SPCF_SHOULDYIELD) && !pagedaemon;
933 		if (pg->flags & PG_BUSY || yld) {
934 			UVMHIST_LOG(ubchist, "busy %p", pg,0,0,0);
935 			if (flags & PGO_BUSYFAIL && pg->flags & PG_BUSY) {
936 				UVMHIST_LOG(ubchist, "busyfail %p", pg, 0,0,0);
937 				error = EDEADLK;
938 				if (busypg != NULL)
939 					*busypg = pg;
940 				break;
941 			}
942 			if (pagedaemon) {
943 				/*
944 				 * someone has taken the page while we
945 				 * dropped the lock for fstrans_start.
946 				 */
947 				break;
948 			}
949 			if (by_list) {
950 				TAILQ_INSERT_BEFORE(pg, &curmp, listq.queue);
951 				UVMHIST_LOG(ubchist, "curmp next %p",
952 				    TAILQ_NEXT(&curmp, listq.queue), 0,0,0);
953 			}
954 			if (yld) {
955 				mutex_exit(slock);
956 				preempt();
957 				mutex_enter(slock);
958 			} else {
959 				pg->flags |= PG_WANTED;
960 				UVM_UNLOCK_AND_WAIT(pg, slock, 0, "genput", 0);
961 				mutex_enter(slock);
962 			}
963 			if (by_list) {
964 				UVMHIST_LOG(ubchist, "after next %p",
965 				    TAILQ_NEXT(&curmp, listq.queue), 0,0,0);
966 				pg = TAILQ_NEXT(&curmp, listq.queue);
967 				TAILQ_REMOVE(&uobj->memq, &curmp, listq.queue);
968 			} else {
969 				pg = uvm_pagelookup(uobj, off);
970 			}
971 			continue;
972 		}
973 
974 		/*
975 		 * if we're freeing, remove all mappings of the page now.
976 		 * if we're cleaning, check if the page is needs to be cleaned.
977 		 */
978 
979 		if (flags & PGO_FREE) {
980 			pmap_page_protect(pg, VM_PROT_NONE);
981 		} else if (flags & PGO_CLEANIT) {
982 
983 			/*
984 			 * if we still have some hope to pull this vnode off
985 			 * from the syncer queue, write-protect the page.
986 			 */
987 
988 			if (cleanall && wasclean &&
989 			    gp->g_dirtygen == dirtygen) {
990 
991 				/*
992 				 * uobj pages get wired only by uvm_fault
993 				 * where uobj is locked.
994 				 */
995 
996 				if (pg->wire_count == 0) {
997 					pmap_page_protect(pg,
998 					    VM_PROT_READ|VM_PROT_EXECUTE);
999 				} else {
1000 					cleanall = false;
1001 				}
1002 			}
1003 		}
1004 
1005 		if (flags & PGO_CLEANIT) {
1006 			needs_clean = pmap_clear_modify(pg) ||
1007 			    (pg->flags & PG_CLEAN) == 0;
1008 			pg->flags |= PG_CLEAN;
1009 		} else {
1010 			needs_clean = false;
1011 		}
1012 
1013 		/*
1014 		 * if we're cleaning, build a cluster.
1015 		 * the cluster will consist of pages which are currently dirty,
1016 		 * but they will be returned to us marked clean.
1017 		 * if not cleaning, just operate on the one page.
1018 		 */
1019 
1020 		if (needs_clean) {
1021 			KDASSERT((vp->v_iflag & VI_ONWORKLST));
1022 			wasclean = false;
1023 			memset(pgs, 0, sizeof(pgs));
1024 			pg->flags |= PG_BUSY;
1025 			UVM_PAGE_OWN(pg, "genfs_putpages");
1026 
1027 			/*
1028 			 * first look backward.
1029 			 */
1030 
1031 			npages = MIN(maxpages >> 1, off >> PAGE_SHIFT);
1032 			nback = npages;
1033 			uvn_findpages(uobj, off - PAGE_SIZE, &nback, &pgs[0],
1034 			    UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY|UFP_BACKWARD);
1035 			if (nback) {
1036 				memmove(&pgs[0], &pgs[npages - nback],
1037 				    nback * sizeof(pgs[0]));
1038 				if (npages - nback < nback)
1039 					memset(&pgs[nback], 0,
1040 					    (npages - nback) * sizeof(pgs[0]));
1041 				else
1042 					memset(&pgs[npages - nback], 0,
1043 					    nback * sizeof(pgs[0]));
1044 			}
1045 
1046 			/*
1047 			 * then plug in our page of interest.
1048 			 */
1049 
1050 			pgs[nback] = pg;
1051 
1052 			/*
1053 			 * then look forward to fill in the remaining space in
1054 			 * the array of pages.
1055 			 */
1056 
1057 			npages = maxpages - nback - 1;
1058 			uvn_findpages(uobj, off + PAGE_SIZE, &npages,
1059 			    &pgs[nback + 1],
1060 			    UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY);
1061 			npages += nback + 1;
1062 		} else {
1063 			pgs[0] = pg;
1064 			npages = 1;
1065 			nback = 0;
1066 		}
1067 
1068 		/*
1069 		 * apply FREE or DEACTIVATE options if requested.
1070 		 */
1071 
1072 		if (flags & (PGO_DEACTIVATE|PGO_FREE)) {
1073 			mutex_enter(&uvm_pageqlock);
1074 		}
1075 		for (i = 0; i < npages; i++) {
1076 			tpg = pgs[i];
1077 			KASSERT(tpg->uobject == uobj);
1078 			if (by_list && tpg == TAILQ_NEXT(pg, listq.queue))
1079 				pg = tpg;
1080 			if (tpg->offset < startoff || tpg->offset >= endoff)
1081 				continue;
1082 			if (flags & PGO_DEACTIVATE && tpg->wire_count == 0) {
1083 				uvm_pagedeactivate(tpg);
1084 			} else if (flags & PGO_FREE) {
1085 				pmap_page_protect(tpg, VM_PROT_NONE);
1086 				if (tpg->flags & PG_BUSY) {
1087 					tpg->flags |= freeflag;
1088 					if (pagedaemon) {
1089 						uvm_pageout_start(1);
1090 						uvm_pagedequeue(tpg);
1091 					}
1092 				} else {
1093 
1094 					/*
1095 					 * ``page is not busy''
1096 					 * implies that npages is 1
1097 					 * and needs_clean is false.
1098 					 */
1099 
1100 					nextpg = TAILQ_NEXT(tpg, listq.queue);
1101 					uvm_pagefree(tpg);
1102 					if (pagedaemon)
1103 						uvmexp.pdfreed++;
1104 				}
1105 			}
1106 		}
1107 		if (flags & (PGO_DEACTIVATE|PGO_FREE)) {
1108 			mutex_exit(&uvm_pageqlock);
1109 		}
1110 		if (needs_clean) {
1111 			modified = true;
1112 
1113 			/*
1114 			 * start the i/o.  if we're traversing by list,
1115 			 * keep our place in the list with a marker page.
1116 			 */
1117 
1118 			if (by_list) {
1119 				TAILQ_INSERT_AFTER(&uobj->memq, pg, &curmp,
1120 				    listq.queue);
1121 			}
1122 			mutex_exit(slock);
1123 			error = GOP_WRITE(vp, pgs, npages, flags);
1124 			mutex_enter(slock);
1125 			if (by_list) {
1126 				pg = TAILQ_NEXT(&curmp, listq.queue);
1127 				TAILQ_REMOVE(&uobj->memq, &curmp, listq.queue);
1128 			}
1129 			if (error) {
1130 				break;
1131 			}
1132 			if (by_list) {
1133 				continue;
1134 			}
1135 		}
1136 
1137 		/*
1138 		 * find the next page and continue if there was no error.
1139 		 */
1140 
1141 		if (by_list) {
1142 			if (nextpg) {
1143 				pg = nextpg;
1144 				nextpg = NULL;
1145 			} else {
1146 				pg = TAILQ_NEXT(pg, listq.queue);
1147 			}
1148 		} else {
1149 			off += (npages - nback) << PAGE_SHIFT;
1150 			if (off < endoff) {
1151 				pg = uvm_pagelookup(uobj, off);
1152 			}
1153 		}
1154 	}
1155 	if (by_list) {
1156 		TAILQ_REMOVE(&uobj->memq, &endmp, listq.queue);
1157 	}
1158 
1159 	if (modified && (vp->v_iflag & VI_WRMAPDIRTY) != 0 &&
1160 	    (vp->v_type != VBLK ||
1161 	    (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) {
1162 		GOP_MARKUPDATE(vp, GOP_UPDATE_MODIFIED);
1163 	}
1164 
1165 	/*
1166 	 * if we're cleaning and there was nothing to clean,
1167 	 * take us off the syncer list.  if we started any i/o
1168 	 * and we're doing sync i/o, wait for all writes to finish.
1169 	 */
1170 
1171 	if (cleanall && wasclean && gp->g_dirtygen == dirtygen &&
1172 	    (vp->v_iflag & VI_ONWORKLST) != 0) {
1173 #if defined(DEBUG)
1174 		TAILQ_FOREACH(pg, &uobj->memq, listq.queue) {
1175 			if ((pg->flags & PG_CLEAN) == 0) {
1176 				printf("%s: %p: !CLEAN\n", __func__, pg);
1177 			}
1178 			if (pmap_is_modified(pg)) {
1179 				printf("%s: %p: modified\n", __func__, pg);
1180 			}
1181 		}
1182 #endif /* defined(DEBUG) */
1183 		vp->v_iflag &= ~VI_WRMAPDIRTY;
1184 		if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL)
1185 			vn_syncer_remove_from_worklist(vp);
1186 	}
1187 
1188 #if !defined(DEBUG)
1189 skip_scan:
1190 #endif /* !defined(DEBUG) */
1191 
1192 	/* Wait for output to complete. */
1193 	if (!wasclean && !async && vp->v_numoutput != 0) {
1194 		while (vp->v_numoutput != 0)
1195 			cv_wait(&vp->v_cv, slock);
1196 	}
1197 	onworklst = (vp->v_iflag & VI_ONWORKLST) != 0;
1198 	mutex_exit(slock);
1199 
1200 	if ((flags & PGO_RECLAIM) != 0 && onworklst) {
1201 		/*
1202 		 * in the case of PGO_RECLAIM, ensure to make the vnode clean.
1203 		 * retrying is not a big deal because, in many cases,
1204 		 * uobj->uo_npages is already 0 here.
1205 		 */
1206 		mutex_enter(slock);
1207 		goto retry;
1208 	}
1209 
1210 	if (has_trans) {
1211 		if (need_wapbl)
1212 			WAPBL_END(vp->v_mount);
1213 		fstrans_done(vp->v_mount);
1214 	}
1215 
1216 	return (error);
1217 }
1218 
1219 int
1220 genfs_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, int flags)
1221 {
1222 	off_t off;
1223 	vaddr_t kva;
1224 	size_t len;
1225 	int error;
1226 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1227 
1228 	UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x",
1229 	    vp, pgs, npages, flags);
1230 
1231 	off = pgs[0]->offset;
1232 	kva = uvm_pagermapin(pgs, npages,
1233 	    UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK);
1234 	len = npages << PAGE_SHIFT;
1235 
1236 	error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE,
1237 			    uvm_aio_biodone);
1238 
1239 	return error;
1240 }
1241 
1242 int
1243 genfs_gop_write_rwmap(struct vnode *vp, struct vm_page **pgs, int npages, int flags)
1244 {
1245 	off_t off;
1246 	vaddr_t kva;
1247 	size_t len;
1248 	int error;
1249 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1250 
1251 	UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x",
1252 	    vp, pgs, npages, flags);
1253 
1254 	off = pgs[0]->offset;
1255 	kva = uvm_pagermapin(pgs, npages,
1256 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
1257 	len = npages << PAGE_SHIFT;
1258 
1259 	error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE,
1260 			    uvm_aio_biodone);
1261 
1262 	return error;
1263 }
1264 
1265 /*
1266  * Backend routine for doing I/O to vnode pages.  Pages are already locked
1267  * and mapped into kernel memory.  Here we just look up the underlying
1268  * device block addresses and call the strategy routine.
1269  */
1270 
1271 static int
1272 genfs_do_io(struct vnode *vp, off_t off, vaddr_t kva, size_t len, int flags,
1273     enum uio_rw rw, void (*iodone)(struct buf *))
1274 {
1275 	int s, error, run;
1276 	int fs_bshift, dev_bshift;
1277 	off_t eof, offset, startoffset;
1278 	size_t bytes, iobytes, skipbytes;
1279 	daddr_t lbn, blkno;
1280 	struct buf *mbp, *bp;
1281 	struct vnode *devvp;
1282 	bool async = (flags & PGO_SYNCIO) == 0;
1283 	bool write = rw == UIO_WRITE;
1284 	int brw = write ? B_WRITE : B_READ;
1285 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1286 
1287 	UVMHIST_LOG(ubchist, "vp %p kva %p len 0x%x flags 0x%x",
1288 	    vp, kva, len, flags);
1289 
1290 	KASSERT(vp->v_size <= vp->v_writesize);
1291 	GOP_SIZE(vp, vp->v_writesize, &eof, 0);
1292 	if (vp->v_type != VBLK) {
1293 		fs_bshift = vp->v_mount->mnt_fs_bshift;
1294 		dev_bshift = vp->v_mount->mnt_dev_bshift;
1295 	} else {
1296 		fs_bshift = DEV_BSHIFT;
1297 		dev_bshift = DEV_BSHIFT;
1298 	}
1299 	error = 0;
1300 	startoffset = off;
1301 	bytes = MIN(len, eof - startoffset);
1302 	skipbytes = 0;
1303 	KASSERT(bytes != 0);
1304 
1305 	if (write) {
1306 		mutex_enter(&vp->v_interlock);
1307 		vp->v_numoutput += 2;
1308 		mutex_exit(&vp->v_interlock);
1309 	}
1310 	mbp = getiobuf(vp, true);
1311 	UVMHIST_LOG(ubchist, "vp %p mbp %p num now %d bytes 0x%x",
1312 	    vp, mbp, vp->v_numoutput, bytes);
1313 	mbp->b_bufsize = len;
1314 	mbp->b_data = (void *)kva;
1315 	mbp->b_resid = mbp->b_bcount = bytes;
1316 	mbp->b_cflags = BC_BUSY | BC_AGE;
1317 	if (async) {
1318 		mbp->b_flags = brw | B_ASYNC;
1319 		mbp->b_iodone = iodone;
1320 	} else {
1321 		mbp->b_flags = brw;
1322 		mbp->b_iodone = NULL;
1323 	}
1324 	if (curlwp == uvm.pagedaemon_lwp)
1325 		BIO_SETPRIO(mbp, BPRIO_TIMELIMITED);
1326 	else if (async)
1327 		BIO_SETPRIO(mbp, BPRIO_TIMENONCRITICAL);
1328 	else
1329 		BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL);
1330 
1331 	bp = NULL;
1332 	for (offset = startoffset;
1333 	    bytes > 0;
1334 	    offset += iobytes, bytes -= iobytes) {
1335 		lbn = offset >> fs_bshift;
1336 		error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run);
1337 		if (error) {
1338 			UVMHIST_LOG(ubchist, "VOP_BMAP() -> %d", error,0,0,0);
1339 			skipbytes += bytes;
1340 			bytes = 0;
1341 			break;
1342 		}
1343 
1344 		iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset,
1345 		    bytes);
1346 		if (blkno == (daddr_t)-1) {
1347 			if (!write) {
1348 				memset((char *)kva + (offset - startoffset), 0,
1349 				   iobytes);
1350 			}
1351 			skipbytes += iobytes;
1352 			continue;
1353 		}
1354 
1355 		/* if it's really one i/o, don't make a second buf */
1356 		if (offset == startoffset && iobytes == bytes) {
1357 			bp = mbp;
1358 		} else {
1359 			UVMHIST_LOG(ubchist, "vp %p bp %p num now %d",
1360 			    vp, bp, vp->v_numoutput, 0);
1361 			bp = getiobuf(vp, true);
1362 			nestiobuf_setup(mbp, bp, offset - startoffset, iobytes);
1363 		}
1364 		bp->b_lblkno = 0;
1365 
1366 		/* adjust physical blkno for partial blocks */
1367 		bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >>
1368 		    dev_bshift);
1369 		UVMHIST_LOG(ubchist,
1370 		    "vp %p offset 0x%x bcount 0x%x blkno 0x%x",
1371 		    vp, offset, bp->b_bcount, bp->b_blkno);
1372 
1373 		VOP_STRATEGY(devvp, bp);
1374 	}
1375 	if (skipbytes) {
1376 		UVMHIST_LOG(ubchist, "skipbytes %d", skipbytes, 0,0,0);
1377 	}
1378 	nestiobuf_done(mbp, skipbytes, error);
1379 	if (async) {
1380 		UVMHIST_LOG(ubchist, "returning 0 (async)", 0,0,0,0);
1381 		return (0);
1382 	}
1383 	UVMHIST_LOG(ubchist, "waiting for mbp %p", mbp,0,0,0);
1384 	error = biowait(mbp);
1385 	s = splbio();
1386 	(*iodone)(mbp);
1387 	splx(s);
1388 	UVMHIST_LOG(ubchist, "returning, error %d", error,0,0,0);
1389 	return (error);
1390 }
1391 
1392 int
1393 genfs_compat_getpages(void *v)
1394 {
1395 	struct vop_getpages_args /* {
1396 		struct vnode *a_vp;
1397 		voff_t a_offset;
1398 		struct vm_page **a_m;
1399 		int *a_count;
1400 		int a_centeridx;
1401 		vm_prot_t a_access_type;
1402 		int a_advice;
1403 		int a_flags;
1404 	} */ *ap = v;
1405 
1406 	off_t origoffset;
1407 	struct vnode *vp = ap->a_vp;
1408 	struct uvm_object *uobj = &vp->v_uobj;
1409 	struct vm_page *pg, **pgs;
1410 	vaddr_t kva;
1411 	int i, error, orignpages, npages;
1412 	struct iovec iov;
1413 	struct uio uio;
1414 	kauth_cred_t cred = curlwp->l_cred;
1415 	bool write = (ap->a_access_type & VM_PROT_WRITE) != 0;
1416 
1417 	error = 0;
1418 	origoffset = ap->a_offset;
1419 	orignpages = *ap->a_count;
1420 	pgs = ap->a_m;
1421 
1422 	if (write && (vp->v_iflag & VI_ONWORKLST) == 0) {
1423 		vn_syncer_add_to_worklist(vp, filedelay);
1424 	}
1425 	if (ap->a_flags & PGO_LOCKED) {
1426 		uvn_findpages(uobj, origoffset, ap->a_count, ap->a_m,
1427 		    UFP_NOWAIT|UFP_NOALLOC| (write ? UFP_NORDONLY : 0));
1428 
1429 		return (ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0);
1430 	}
1431 	if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= vp->v_size) {
1432 		mutex_exit(&uobj->vmobjlock);
1433 		return (EINVAL);
1434 	}
1435 	if ((ap->a_flags & PGO_SYNCIO) == 0) {
1436 		mutex_exit(&uobj->vmobjlock);
1437 		return 0;
1438 	}
1439 	npages = orignpages;
1440 	uvn_findpages(uobj, origoffset, &npages, pgs, UFP_ALL);
1441 	mutex_exit(&uobj->vmobjlock);
1442 	kva = uvm_pagermapin(pgs, npages,
1443 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
1444 	for (i = 0; i < npages; i++) {
1445 		pg = pgs[i];
1446 		if ((pg->flags & PG_FAKE) == 0) {
1447 			continue;
1448 		}
1449 		iov.iov_base = (char *)kva + (i << PAGE_SHIFT);
1450 		iov.iov_len = PAGE_SIZE;
1451 		uio.uio_iov = &iov;
1452 		uio.uio_iovcnt = 1;
1453 		uio.uio_offset = origoffset + (i << PAGE_SHIFT);
1454 		uio.uio_rw = UIO_READ;
1455 		uio.uio_resid = PAGE_SIZE;
1456 		UIO_SETUP_SYSSPACE(&uio);
1457 		/* XXX vn_lock */
1458 		error = VOP_READ(vp, &uio, 0, cred);
1459 		if (error) {
1460 			break;
1461 		}
1462 		if (uio.uio_resid) {
1463 			memset(iov.iov_base, 0, uio.uio_resid);
1464 		}
1465 	}
1466 	uvm_pagermapout(kva, npages);
1467 	mutex_enter(&uobj->vmobjlock);
1468 	mutex_enter(&uvm_pageqlock);
1469 	for (i = 0; i < npages; i++) {
1470 		pg = pgs[i];
1471 		if (error && (pg->flags & PG_FAKE) != 0) {
1472 			pg->flags |= PG_RELEASED;
1473 		} else {
1474 			pmap_clear_modify(pg);
1475 			uvm_pageactivate(pg);
1476 		}
1477 	}
1478 	if (error) {
1479 		uvm_page_unbusy(pgs, npages);
1480 	}
1481 	mutex_exit(&uvm_pageqlock);
1482 	mutex_exit(&uobj->vmobjlock);
1483 	return (error);
1484 }
1485 
1486 int
1487 genfs_compat_gop_write(struct vnode *vp, struct vm_page **pgs, int npages,
1488     int flags)
1489 {
1490 	off_t offset;
1491 	struct iovec iov;
1492 	struct uio uio;
1493 	kauth_cred_t cred = curlwp->l_cred;
1494 	struct buf *bp;
1495 	vaddr_t kva;
1496 	int error;
1497 
1498 	offset = pgs[0]->offset;
1499 	kva = uvm_pagermapin(pgs, npages,
1500 	    UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK);
1501 
1502 	iov.iov_base = (void *)kva;
1503 	iov.iov_len = npages << PAGE_SHIFT;
1504 	uio.uio_iov = &iov;
1505 	uio.uio_iovcnt = 1;
1506 	uio.uio_offset = offset;
1507 	uio.uio_rw = UIO_WRITE;
1508 	uio.uio_resid = npages << PAGE_SHIFT;
1509 	UIO_SETUP_SYSSPACE(&uio);
1510 	/* XXX vn_lock */
1511 	error = VOP_WRITE(vp, &uio, 0, cred);
1512 
1513 	mutex_enter(&vp->v_interlock);
1514 	vp->v_numoutput++;
1515 	mutex_exit(&vp->v_interlock);
1516 
1517 	bp = getiobuf(vp, true);
1518 	bp->b_cflags = BC_BUSY | BC_AGE;
1519 	bp->b_lblkno = offset >> vp->v_mount->mnt_fs_bshift;
1520 	bp->b_data = (char *)kva;
1521 	bp->b_bcount = npages << PAGE_SHIFT;
1522 	bp->b_bufsize = npages << PAGE_SHIFT;
1523 	bp->b_resid = 0;
1524 	bp->b_error = error;
1525 	uvm_aio_aiodone(bp);
1526 	return (error);
1527 }
1528 
1529 /*
1530  * Process a uio using direct I/O.  If we reach a part of the request
1531  * which cannot be processed in this fashion for some reason, just return.
1532  * The caller must handle some additional part of the request using
1533  * buffered I/O before trying direct I/O again.
1534  */
1535 
1536 void
1537 genfs_directio(struct vnode *vp, struct uio *uio, int ioflag)
1538 {
1539 	struct vmspace *vs;
1540 	struct iovec *iov;
1541 	vaddr_t va;
1542 	size_t len;
1543 	const int mask = DEV_BSIZE - 1;
1544 	int error;
1545 	bool need_wapbl = (vp->v_mount && vp->v_mount->mnt_wapbl &&
1546 	    (ioflag & IO_JOURNALLOCKED) == 0);
1547 
1548 	/*
1549 	 * We only support direct I/O to user space for now.
1550 	 */
1551 
1552 	if (VMSPACE_IS_KERNEL_P(uio->uio_vmspace)) {
1553 		return;
1554 	}
1555 
1556 	/*
1557 	 * If the vnode is mapped, we would need to get the getpages lock
1558 	 * to stabilize the bmap, but then we would get into trouble whil e
1559 	 * locking the pages if the pages belong to this same vnode (or a
1560 	 * multi-vnode cascade to the same effect).  Just fall back to
1561 	 * buffered I/O if the vnode is mapped to avoid this mess.
1562 	 */
1563 
1564 	if (vp->v_vflag & VV_MAPPED) {
1565 		return;
1566 	}
1567 
1568 	if (need_wapbl) {
1569 		error = WAPBL_BEGIN(vp->v_mount);
1570 		if (error)
1571 			return;
1572 	}
1573 
1574 	/*
1575 	 * Do as much of the uio as possible with direct I/O.
1576 	 */
1577 
1578 	vs = uio->uio_vmspace;
1579 	while (uio->uio_resid) {
1580 		iov = uio->uio_iov;
1581 		if (iov->iov_len == 0) {
1582 			uio->uio_iov++;
1583 			uio->uio_iovcnt--;
1584 			continue;
1585 		}
1586 		va = (vaddr_t)iov->iov_base;
1587 		len = MIN(iov->iov_len, genfs_maxdio);
1588 		len &= ~mask;
1589 
1590 		/*
1591 		 * If the next chunk is smaller than DEV_BSIZE or extends past
1592 		 * the current EOF, then fall back to buffered I/O.
1593 		 */
1594 
1595 		if (len == 0 || uio->uio_offset + len > vp->v_size) {
1596 			break;
1597 		}
1598 
1599 		/*
1600 		 * Check alignment.  The file offset must be at least
1601 		 * sector-aligned.  The exact constraint on memory alignment
1602 		 * is very hardware-dependent, but requiring sector-aligned
1603 		 * addresses there too is safe.
1604 		 */
1605 
1606 		if (uio->uio_offset & mask || va & mask) {
1607 			break;
1608 		}
1609 		error = genfs_do_directio(vs, va, len, vp, uio->uio_offset,
1610 					  uio->uio_rw);
1611 		if (error) {
1612 			break;
1613 		}
1614 		iov->iov_base = (char *)iov->iov_base + len;
1615 		iov->iov_len -= len;
1616 		uio->uio_offset += len;
1617 		uio->uio_resid -= len;
1618 	}
1619 
1620 	if (need_wapbl)
1621 		WAPBL_END(vp->v_mount);
1622 }
1623 
1624 /*
1625  * Iodone routine for direct I/O.  We don't do much here since the request is
1626  * always synchronous, so the caller will do most of the work after biowait().
1627  */
1628 
1629 static void
1630 genfs_dio_iodone(struct buf *bp)
1631 {
1632 
1633 	KASSERT((bp->b_flags & B_ASYNC) == 0);
1634 	if ((bp->b_flags & B_READ) == 0 && (bp->b_cflags & BC_AGE) != 0) {
1635 		mutex_enter(bp->b_objlock);
1636 		vwakeup(bp);
1637 		mutex_exit(bp->b_objlock);
1638 	}
1639 	putiobuf(bp);
1640 }
1641 
1642 /*
1643  * Process one chunk of a direct I/O request.
1644  */
1645 
1646 static int
1647 genfs_do_directio(struct vmspace *vs, vaddr_t uva, size_t len, struct vnode *vp,
1648     off_t off, enum uio_rw rw)
1649 {
1650 	struct vm_map *map;
1651 	struct pmap *upm, *kpm;
1652 	size_t klen = round_page(uva + len) - trunc_page(uva);
1653 	off_t spoff, epoff;
1654 	vaddr_t kva, puva;
1655 	paddr_t pa;
1656 	vm_prot_t prot;
1657 	int error, rv, poff, koff;
1658 	const int pgoflags = PGO_CLEANIT | PGO_SYNCIO | PGO_JOURNALLOCKED |
1659 		(rw == UIO_WRITE ? PGO_FREE : 0);
1660 
1661 	/*
1662 	 * For writes, verify that this range of the file already has fully
1663 	 * allocated backing store.  If there are any holes, just punt and
1664 	 * make the caller take the buffered write path.
1665 	 */
1666 
1667 	if (rw == UIO_WRITE) {
1668 		daddr_t lbn, elbn, blkno;
1669 		int bsize, bshift, run;
1670 
1671 		bshift = vp->v_mount->mnt_fs_bshift;
1672 		bsize = 1 << bshift;
1673 		lbn = off >> bshift;
1674 		elbn = (off + len + bsize - 1) >> bshift;
1675 		while (lbn < elbn) {
1676 			error = VOP_BMAP(vp, lbn, NULL, &blkno, &run);
1677 			if (error) {
1678 				return error;
1679 			}
1680 			if (blkno == (daddr_t)-1) {
1681 				return ENOSPC;
1682 			}
1683 			lbn += 1 + run;
1684 		}
1685 	}
1686 
1687 	/*
1688 	 * Flush any cached pages for parts of the file that we're about to
1689 	 * access.  If we're writing, invalidate pages as well.
1690 	 */
1691 
1692 	spoff = trunc_page(off);
1693 	epoff = round_page(off + len);
1694 	mutex_enter(&vp->v_interlock);
1695 	error = VOP_PUTPAGES(vp, spoff, epoff, pgoflags);
1696 	if (error) {
1697 		return error;
1698 	}
1699 
1700 	/*
1701 	 * Wire the user pages and remap them into kernel memory.
1702 	 */
1703 
1704 	prot = rw == UIO_READ ? VM_PROT_READ | VM_PROT_WRITE : VM_PROT_READ;
1705 	error = uvm_vslock(vs, (void *)uva, len, prot);
1706 	if (error) {
1707 		return error;
1708 	}
1709 
1710 	map = &vs->vm_map;
1711 	upm = vm_map_pmap(map);
1712 	kpm = vm_map_pmap(kernel_map);
1713 	kva = uvm_km_alloc(kernel_map, klen, 0,
1714 			   UVM_KMF_VAONLY | UVM_KMF_WAITVA);
1715 	puva = trunc_page(uva);
1716 	for (poff = 0; poff < klen; poff += PAGE_SIZE) {
1717 		rv = pmap_extract(upm, puva + poff, &pa);
1718 		KASSERT(rv);
1719 		pmap_enter(kpm, kva + poff, pa, prot, prot | PMAP_WIRED);
1720 	}
1721 	pmap_update(kpm);
1722 
1723 	/*
1724 	 * Do the I/O.
1725 	 */
1726 
1727 	koff = uva - trunc_page(uva);
1728 	error = genfs_do_io(vp, off, kva + koff, len, PGO_SYNCIO, rw,
1729 			    genfs_dio_iodone);
1730 
1731 	/*
1732 	 * Tear down the kernel mapping.
1733 	 */
1734 
1735 	pmap_remove(kpm, kva, kva + klen);
1736 	pmap_update(kpm);
1737 	uvm_km_free(kernel_map, kva, klen, UVM_KMF_VAONLY);
1738 
1739 	/*
1740 	 * Unwire the user pages.
1741 	 */
1742 
1743 	uvm_vsunlock(vs, (void *)uva, len);
1744 	return error;
1745 }
1746 
1747