xref: /netbsd-src/sys/miscfs/genfs/genfs_io.c (revision f89f6560d453f5e37386cc7938c072d2f528b9fa)
1 /*	$NetBSD: genfs_io.c,v 1.60 2015/04/12 14:44:06 skrll 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 1.60 2015/04/12 14:44:06 skrll 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/vnode.h>
42 #include <sys/kmem.h>
43 #include <sys/kauth.h>
44 #include <sys/fstrans.h>
45 #include <sys/buf.h>
46 
47 #include <miscfs/genfs/genfs.h>
48 #include <miscfs/genfs/genfs_node.h>
49 #include <miscfs/specfs/specdev.h>
50 #include <miscfs/syncfs/syncfs.h>
51 
52 #include <uvm/uvm.h>
53 #include <uvm/uvm_pager.h>
54 
55 static int genfs_do_directio(struct vmspace *, vaddr_t, size_t, struct vnode *,
56     off_t, enum uio_rw);
57 static void genfs_dio_iodone(struct buf *);
58 
59 static int genfs_getpages_read(struct vnode *, struct vm_page **, int, off_t,
60     off_t, bool, bool, bool, bool);
61 static int genfs_do_io(struct vnode *, off_t, vaddr_t, size_t, int, enum uio_rw,
62     void (*)(struct buf *));
63 static void genfs_rel_pages(struct vm_page **, unsigned int);
64 static void genfs_markdirty(struct vnode *);
65 
66 int genfs_maxdio = MAXPHYS;
67 
68 static void
69 genfs_rel_pages(struct vm_page **pgs, unsigned int npages)
70 {
71 	unsigned int i;
72 
73 	for (i = 0; i < npages; i++) {
74 		struct vm_page *pg = pgs[i];
75 
76 		if (pg == NULL || pg == PGO_DONTCARE)
77 			continue;
78 		KASSERT(uvm_page_locked_p(pg));
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 static void
89 genfs_markdirty(struct vnode *vp)
90 {
91 	struct genfs_node * const gp = VTOG(vp);
92 
93 	KASSERT(mutex_owned(vp->v_interlock));
94 	gp->g_dirtygen++;
95 	if ((vp->v_iflag & VI_ONWORKLST) == 0) {
96 		vn_syncer_add_to_worklist(vp, filedelay);
97 	}
98 	if ((vp->v_iflag & (VI_WRMAP|VI_WRMAPDIRTY)) == VI_WRMAP) {
99 		vp->v_iflag |= VI_WRMAPDIRTY;
100 	}
101 }
102 
103 /*
104  * generic VM getpages routine.
105  * Return PG_BUSY pages for the given range,
106  * reading from backing store if necessary.
107  */
108 
109 int
110 genfs_getpages(void *v)
111 {
112 	struct vop_getpages_args /* {
113 		struct vnode *a_vp;
114 		voff_t a_offset;
115 		struct vm_page **a_m;
116 		int *a_count;
117 		int a_centeridx;
118 		vm_prot_t a_access_type;
119 		int a_advice;
120 		int a_flags;
121 	} */ * const ap = v;
122 
123 	off_t diskeof, memeof;
124 	int i, error, npages;
125 	const int flags = ap->a_flags;
126 	struct vnode * const vp = ap->a_vp;
127 	struct uvm_object * const uobj = &vp->v_uobj;
128 	const bool async = (flags & PGO_SYNCIO) == 0;
129 	const bool memwrite = (ap->a_access_type & VM_PROT_WRITE) != 0;
130 	const bool overwrite = (flags & PGO_OVERWRITE) != 0;
131 	const bool blockalloc = memwrite && (flags & PGO_NOBLOCKALLOC) == 0;
132 	const bool glocked = (flags & PGO_GLOCKHELD) != 0;
133 	const bool need_wapbl = blockalloc && vp->v_mount->mnt_wapbl;
134 	bool has_trans_wapbl = false;
135 	UVMHIST_FUNC("genfs_getpages"); UVMHIST_CALLED(ubchist);
136 
137 	UVMHIST_LOG(ubchist, "vp %p off 0x%x/%x count %d",
138 	    vp, ap->a_offset >> 32, ap->a_offset, *ap->a_count);
139 
140 	KASSERT(vp->v_type == VREG || vp->v_type == VDIR ||
141 	    vp->v_type == VLNK || vp->v_type == VBLK);
142 
143 startover:
144 	error = 0;
145 	const voff_t origvsize = vp->v_size;
146 	const off_t origoffset = ap->a_offset;
147 	const int orignpages = *ap->a_count;
148 
149 	GOP_SIZE(vp, origvsize, &diskeof, 0);
150 	if (flags & PGO_PASTEOF) {
151 		off_t newsize;
152 #if defined(DIAGNOSTIC)
153 		off_t writeeof;
154 #endif /* defined(DIAGNOSTIC) */
155 
156 		newsize = MAX(origvsize,
157 		    origoffset + (orignpages << PAGE_SHIFT));
158 		GOP_SIZE(vp, newsize, &memeof, GOP_SIZE_MEM);
159 #if defined(DIAGNOSTIC)
160 		GOP_SIZE(vp, vp->v_writesize, &writeeof, GOP_SIZE_MEM);
161 		if (newsize > round_page(writeeof)) {
162 			panic("%s: past eof: %" PRId64 " vs. %" PRId64,
163 			    __func__, newsize, round_page(writeeof));
164 		}
165 #endif /* defined(DIAGNOSTIC) */
166 	} else {
167 		GOP_SIZE(vp, origvsize, &memeof, GOP_SIZE_MEM);
168 	}
169 	KASSERT(ap->a_centeridx >= 0 || ap->a_centeridx <= orignpages);
170 	KASSERT((origoffset & (PAGE_SIZE - 1)) == 0 && origoffset >= 0);
171 	KASSERT(orignpages > 0);
172 
173 	/*
174 	 * Bounds-check the request.
175 	 */
176 
177 	if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= memeof) {
178 		if ((flags & PGO_LOCKED) == 0) {
179 			mutex_exit(uobj->vmobjlock);
180 		}
181 		UVMHIST_LOG(ubchist, "off 0x%x count %d goes past EOF 0x%x",
182 		    origoffset, *ap->a_count, memeof,0);
183 		error = EINVAL;
184 		goto out_err;
185 	}
186 
187 	/* uobj is locked */
188 
189 	if ((flags & PGO_NOTIMESTAMP) == 0 &&
190 	    (vp->v_type != VBLK ||
191 	    (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) {
192 		int updflags = 0;
193 
194 		if ((vp->v_mount->mnt_flag & MNT_NOATIME) == 0) {
195 			updflags = GOP_UPDATE_ACCESSED;
196 		}
197 		if (memwrite) {
198 			updflags |= GOP_UPDATE_MODIFIED;
199 		}
200 		if (updflags != 0) {
201 			GOP_MARKUPDATE(vp, updflags);
202 		}
203 	}
204 
205 	/*
206 	 * For PGO_LOCKED requests, just return whatever's in memory.
207 	 */
208 
209 	if (flags & PGO_LOCKED) {
210 		int nfound;
211 		struct vm_page *pg;
212 
213 		KASSERT(!glocked);
214 		npages = *ap->a_count;
215 #if defined(DEBUG)
216 		for (i = 0; i < npages; i++) {
217 			pg = ap->a_m[i];
218 			KASSERT(pg == NULL || pg == PGO_DONTCARE);
219 		}
220 #endif /* defined(DEBUG) */
221 		nfound = uvn_findpages(uobj, origoffset, &npages,
222 		    ap->a_m, UFP_NOWAIT|UFP_NOALLOC|(memwrite ? UFP_NORDONLY : 0));
223 		KASSERT(npages == *ap->a_count);
224 		if (nfound == 0) {
225 			error = EBUSY;
226 			goto out_err;
227 		}
228 		if (!genfs_node_rdtrylock(vp)) {
229 			genfs_rel_pages(ap->a_m, npages);
230 
231 			/*
232 			 * restore the array.
233 			 */
234 
235 			for (i = 0; i < npages; i++) {
236 				pg = ap->a_m[i];
237 
238 				if (pg != NULL && pg != PGO_DONTCARE) {
239 					ap->a_m[i] = NULL;
240 				}
241 				KASSERT(ap->a_m[i] == NULL ||
242 				    ap->a_m[i] == PGO_DONTCARE);
243 			}
244 		} else {
245 			genfs_node_unlock(vp);
246 		}
247 		error = (ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0);
248 		if (error == 0 && memwrite) {
249 			genfs_markdirty(vp);
250 		}
251 		goto out_err;
252 	}
253 	mutex_exit(uobj->vmobjlock);
254 
255 	/*
256 	 * find the requested pages and make some simple checks.
257 	 * leave space in the page array for a whole block.
258 	 */
259 
260 	const int fs_bshift = (vp->v_type != VBLK) ?
261 	    vp->v_mount->mnt_fs_bshift : DEV_BSHIFT;
262 	const int fs_bsize = 1 << fs_bshift;
263 #define	blk_mask	(fs_bsize - 1)
264 #define	trunc_blk(x)	((x) & ~blk_mask)
265 #define	round_blk(x)	(((x) + blk_mask) & ~blk_mask)
266 
267 	const int orignmempages = MIN(orignpages,
268 	    round_page(memeof - origoffset) >> PAGE_SHIFT);
269 	npages = orignmempages;
270 	const off_t startoffset = trunc_blk(origoffset);
271 	const off_t endoffset = MIN(
272 	    round_page(round_blk(origoffset + (npages << PAGE_SHIFT))),
273 	    round_page(memeof));
274 	const int ridx = (origoffset - startoffset) >> PAGE_SHIFT;
275 
276 	const int pgs_size = sizeof(struct vm_page *) *
277 	    ((endoffset - startoffset) >> PAGE_SHIFT);
278 	struct vm_page **pgs, *pgs_onstack[UBC_MAX_PAGES];
279 
280 	if (pgs_size > sizeof(pgs_onstack)) {
281 		pgs = kmem_zalloc(pgs_size, async ? KM_NOSLEEP : KM_SLEEP);
282 		if (pgs == NULL) {
283 			pgs = pgs_onstack;
284 			error = ENOMEM;
285 			goto out_err;
286 		}
287 	} else {
288 		pgs = pgs_onstack;
289 		(void)memset(pgs, 0, pgs_size);
290 	}
291 
292 	UVMHIST_LOG(ubchist, "ridx %d npages %d startoff %ld endoff %ld",
293 	    ridx, npages, startoffset, endoffset);
294 
295 	if (!has_trans_wapbl) {
296 		fstrans_start(vp->v_mount, FSTRANS_SHARED);
297 		/*
298 		 * XXX: This assumes that we come here only via
299 		 * the mmio path
300 		 */
301 		if (need_wapbl) {
302 			error = WAPBL_BEGIN(vp->v_mount);
303 			if (error) {
304 				fstrans_done(vp->v_mount);
305 				goto out_err_free;
306 			}
307 		}
308 		has_trans_wapbl = true;
309 	}
310 
311 	/*
312 	 * hold g_glock to prevent a race with truncate.
313 	 *
314 	 * check if our idea of v_size is still valid.
315 	 */
316 
317 	KASSERT(!glocked || genfs_node_wrlocked(vp));
318 	if (!glocked) {
319 		if (blockalloc) {
320 			genfs_node_wrlock(vp);
321 		} else {
322 			genfs_node_rdlock(vp);
323 		}
324 	}
325 	mutex_enter(uobj->vmobjlock);
326 	if (vp->v_size < origvsize) {
327 		if (!glocked) {
328 			genfs_node_unlock(vp);
329 		}
330 		if (pgs != pgs_onstack)
331 			kmem_free(pgs, pgs_size);
332 		goto startover;
333 	}
334 
335 	if (uvn_findpages(uobj, origoffset, &npages, &pgs[ridx],
336 	    async ? UFP_NOWAIT : UFP_ALL) != orignmempages) {
337 		if (!glocked) {
338 			genfs_node_unlock(vp);
339 		}
340 		KASSERT(async != 0);
341 		genfs_rel_pages(&pgs[ridx], orignmempages);
342 		mutex_exit(uobj->vmobjlock);
343 		error = EBUSY;
344 		goto out_err_free;
345 	}
346 
347 	/*
348 	 * if the pages are already resident, just return them.
349 	 */
350 
351 	for (i = 0; i < npages; i++) {
352 		struct vm_page *pg = pgs[ridx + i];
353 
354 		if ((pg->flags & PG_FAKE) ||
355 		    (blockalloc && (pg->flags & PG_RDONLY))) {
356 			break;
357 		}
358 	}
359 	if (i == npages) {
360 		if (!glocked) {
361 			genfs_node_unlock(vp);
362 		}
363 		UVMHIST_LOG(ubchist, "returning cached pages", 0,0,0,0);
364 		npages += ridx;
365 		goto out;
366 	}
367 
368 	/*
369 	 * if PGO_OVERWRITE is set, don't bother reading the pages.
370 	 */
371 
372 	if (overwrite) {
373 		if (!glocked) {
374 			genfs_node_unlock(vp);
375 		}
376 		UVMHIST_LOG(ubchist, "PGO_OVERWRITE",0,0,0,0);
377 
378 		for (i = 0; i < npages; i++) {
379 			struct vm_page *pg = pgs[ridx + i];
380 
381 			pg->flags &= ~(PG_RDONLY|PG_CLEAN);
382 		}
383 		npages += ridx;
384 		goto out;
385 	}
386 
387 	/*
388 	 * the page wasn't resident and we're not overwriting,
389 	 * so we're going to have to do some i/o.
390 	 * find any additional pages needed to cover the expanded range.
391 	 */
392 
393 	npages = (endoffset - startoffset) >> PAGE_SHIFT;
394 	if (startoffset != origoffset || npages != orignmempages) {
395 		int npgs;
396 
397 		/*
398 		 * we need to avoid deadlocks caused by locking
399 		 * additional pages at lower offsets than pages we
400 		 * already have locked.  unlock them all and start over.
401 		 */
402 
403 		genfs_rel_pages(&pgs[ridx], orignmempages);
404 		memset(pgs, 0, pgs_size);
405 
406 		UVMHIST_LOG(ubchist, "reset npages start 0x%x end 0x%x",
407 		    startoffset, endoffset, 0,0);
408 		npgs = npages;
409 		if (uvn_findpages(uobj, startoffset, &npgs, pgs,
410 		    async ? UFP_NOWAIT : UFP_ALL) != npages) {
411 			if (!glocked) {
412 				genfs_node_unlock(vp);
413 			}
414 			KASSERT(async != 0);
415 			genfs_rel_pages(pgs, npages);
416 			mutex_exit(uobj->vmobjlock);
417 			error = EBUSY;
418 			goto out_err_free;
419 		}
420 	}
421 
422 	mutex_exit(uobj->vmobjlock);
423 	error = genfs_getpages_read(vp, pgs, npages, startoffset, diskeof,
424 	    async, memwrite, blockalloc, glocked);
425 	if (error == 0 && async)
426 		goto out_err_free;
427 	if (!glocked) {
428 		genfs_node_unlock(vp);
429 	}
430 	mutex_enter(uobj->vmobjlock);
431 
432 	/*
433 	 * we're almost done!  release the pages...
434 	 * for errors, we free the pages.
435 	 * otherwise we activate them and mark them as valid and clean.
436 	 * also, unbusy pages that were not actually requested.
437 	 */
438 
439 	if (error) {
440 		genfs_rel_pages(pgs, npages);
441 		mutex_exit(uobj->vmobjlock);
442 		UVMHIST_LOG(ubchist, "returning error %d", error,0,0,0);
443 		goto out_err_free;
444 	}
445 
446 out:
447 	UVMHIST_LOG(ubchist, "succeeding, npages %d", npages,0,0,0);
448 	error = 0;
449 	mutex_enter(&uvm_pageqlock);
450 	for (i = 0; i < npages; i++) {
451 		struct vm_page *pg = pgs[i];
452 		if (pg == NULL) {
453 			continue;
454 		}
455 		UVMHIST_LOG(ubchist, "examining pg %p flags 0x%x",
456 		    pg, pg->flags, 0,0);
457 		if (pg->flags & PG_FAKE && !overwrite) {
458 			pg->flags &= ~(PG_FAKE);
459 			pmap_clear_modify(pgs[i]);
460 		}
461 		KASSERT(!memwrite || !blockalloc || (pg->flags & PG_RDONLY) == 0);
462 		if (i < ridx || i >= ridx + orignmempages || async) {
463 			UVMHIST_LOG(ubchist, "unbusy pg %p offset 0x%x",
464 			    pg, pg->offset,0,0);
465 			if (pg->flags & PG_WANTED) {
466 				wakeup(pg);
467 			}
468 			if (pg->flags & PG_FAKE) {
469 				KASSERT(overwrite);
470 				uvm_pagezero(pg);
471 			}
472 			if (pg->flags & PG_RELEASED) {
473 				uvm_pagefree(pg);
474 				continue;
475 			}
476 			uvm_pageenqueue(pg);
477 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
478 			UVM_PAGE_OWN(pg, NULL);
479 		}
480 	}
481 	mutex_exit(&uvm_pageqlock);
482 	if (memwrite) {
483 		genfs_markdirty(vp);
484 	}
485 	mutex_exit(uobj->vmobjlock);
486 	if (ap->a_m != NULL) {
487 		memcpy(ap->a_m, &pgs[ridx],
488 		    orignmempages * sizeof(struct vm_page *));
489 	}
490 
491 out_err_free:
492 	if (pgs != NULL && pgs != pgs_onstack)
493 		kmem_free(pgs, pgs_size);
494 out_err:
495 	if (has_trans_wapbl) {
496 		if (need_wapbl)
497 			WAPBL_END(vp->v_mount);
498 		fstrans_done(vp->v_mount);
499 	}
500 	return error;
501 }
502 
503 /*
504  * genfs_getpages_read: Read the pages in with VOP_BMAP/VOP_STRATEGY.
505  */
506 static int
507 genfs_getpages_read(struct vnode *vp, struct vm_page **pgs, int npages,
508     off_t startoffset, off_t diskeof,
509     bool async, bool memwrite, bool blockalloc, bool glocked)
510 {
511 	struct uvm_object * const uobj = &vp->v_uobj;
512 	const int fs_bshift = (vp->v_type != VBLK) ?
513 	    vp->v_mount->mnt_fs_bshift : DEV_BSHIFT;
514 	const int dev_bshift = (vp->v_type != VBLK) ?
515 	    vp->v_mount->mnt_dev_bshift : DEV_BSHIFT;
516 	kauth_cred_t const cred = curlwp->l_cred;		/* XXXUBC curlwp */
517 	size_t bytes, iobytes, tailstart, tailbytes, totalbytes, skipbytes;
518 	vaddr_t kva;
519 	struct buf *bp, *mbp;
520 	bool sawhole = false;
521 	int i;
522 	int error = 0;
523 
524 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
525 
526 	/*
527 	 * read the desired page(s).
528 	 */
529 
530 	totalbytes = npages << PAGE_SHIFT;
531 	bytes = MIN(totalbytes, MAX(diskeof - startoffset, 0));
532 	tailbytes = totalbytes - bytes;
533 	skipbytes = 0;
534 
535 	kva = uvm_pagermapin(pgs, npages,
536 	    UVMPAGER_MAPIN_READ | (async ? 0 : UVMPAGER_MAPIN_WAITOK));
537 	if (kva == 0)
538 		return EBUSY;
539 
540 	mbp = getiobuf(vp, true);
541 	mbp->b_bufsize = totalbytes;
542 	mbp->b_data = (void *)kva;
543 	mbp->b_resid = mbp->b_bcount = bytes;
544 	mbp->b_cflags = BC_BUSY;
545 	if (async) {
546 		mbp->b_flags = B_READ | B_ASYNC;
547 		mbp->b_iodone = uvm_aio_biodone;
548 	} else {
549 		mbp->b_flags = B_READ;
550 		mbp->b_iodone = NULL;
551 	}
552 	if (async)
553 		BIO_SETPRIO(mbp, BPRIO_TIMELIMITED);
554 	else
555 		BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL);
556 
557 	/*
558 	 * if EOF is in the middle of the range, zero the part past EOF.
559 	 * skip over pages which are not PG_FAKE since in that case they have
560 	 * valid data that we need to preserve.
561 	 */
562 
563 	tailstart = bytes;
564 	while (tailbytes > 0) {
565 		const int len = PAGE_SIZE - (tailstart & PAGE_MASK);
566 
567 		KASSERT(len <= tailbytes);
568 		if ((pgs[tailstart >> PAGE_SHIFT]->flags & PG_FAKE) != 0) {
569 			memset((void *)(kva + tailstart), 0, len);
570 			UVMHIST_LOG(ubchist, "tailbytes %p 0x%x 0x%x",
571 			    kva, tailstart, len, 0);
572 		}
573 		tailstart += len;
574 		tailbytes -= len;
575 	}
576 
577 	/*
578 	 * now loop over the pages, reading as needed.
579 	 */
580 
581 	bp = NULL;
582 	off_t offset;
583 	for (offset = startoffset;
584 	    bytes > 0;
585 	    offset += iobytes, bytes -= iobytes) {
586 		int run;
587 		daddr_t lbn, blkno;
588 		int pidx;
589 		struct vnode *devvp;
590 
591 		/*
592 		 * skip pages which don't need to be read.
593 		 */
594 
595 		pidx = (offset - startoffset) >> PAGE_SHIFT;
596 		while ((pgs[pidx]->flags & PG_FAKE) == 0) {
597 			size_t b;
598 
599 			KASSERT((offset & (PAGE_SIZE - 1)) == 0);
600 			if ((pgs[pidx]->flags & PG_RDONLY)) {
601 				sawhole = true;
602 			}
603 			b = MIN(PAGE_SIZE, bytes);
604 			offset += b;
605 			bytes -= b;
606 			skipbytes += b;
607 			pidx++;
608 			UVMHIST_LOG(ubchist, "skipping, new offset 0x%x",
609 			    offset, 0,0,0);
610 			if (bytes == 0) {
611 				goto loopdone;
612 			}
613 		}
614 
615 		/*
616 		 * bmap the file to find out the blkno to read from and
617 		 * how much we can read in one i/o.  if bmap returns an error,
618 		 * skip the rest of the top-level i/o.
619 		 */
620 
621 		lbn = offset >> fs_bshift;
622 		error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run);
623 		if (error) {
624 			UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%x -> %d\n",
625 			    lbn,error,0,0);
626 			skipbytes += bytes;
627 			bytes = 0;
628 			goto loopdone;
629 		}
630 
631 		/*
632 		 * see how many pages can be read with this i/o.
633 		 * reduce the i/o size if necessary to avoid
634 		 * overwriting pages with valid data.
635 		 */
636 
637 		iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset,
638 		    bytes);
639 		if (offset + iobytes > round_page(offset)) {
640 			int pcount;
641 
642 			pcount = 1;
643 			while (pidx + pcount < npages &&
644 			    pgs[pidx + pcount]->flags & PG_FAKE) {
645 				pcount++;
646 			}
647 			iobytes = MIN(iobytes, (pcount << PAGE_SHIFT) -
648 			    (offset - trunc_page(offset)));
649 		}
650 
651 		/*
652 		 * if this block isn't allocated, zero it instead of
653 		 * reading it.  unless we are going to allocate blocks,
654 		 * mark the pages we zeroed PG_RDONLY.
655 		 */
656 
657 		if (blkno == (daddr_t)-1) {
658 			int holepages = (round_page(offset + iobytes) -
659 			    trunc_page(offset)) >> PAGE_SHIFT;
660 			UVMHIST_LOG(ubchist, "lbn 0x%x -> HOLE", lbn,0,0,0);
661 
662 			sawhole = true;
663 			memset((char *)kva + (offset - startoffset), 0,
664 			    iobytes);
665 			skipbytes += iobytes;
666 
667 			mutex_enter(uobj->vmobjlock);
668 			for (i = 0; i < holepages; i++) {
669 				if (memwrite) {
670 					pgs[pidx + i]->flags &= ~PG_CLEAN;
671 				}
672 				if (!blockalloc) {
673 					pgs[pidx + i]->flags |= PG_RDONLY;
674 				}
675 			}
676 			mutex_exit(uobj->vmobjlock);
677 			continue;
678 		}
679 
680 		/*
681 		 * allocate a sub-buf for this piece of the i/o
682 		 * (or just use mbp if there's only 1 piece),
683 		 * and start it going.
684 		 */
685 
686 		if (offset == startoffset && iobytes == bytes) {
687 			bp = mbp;
688 		} else {
689 			UVMHIST_LOG(ubchist, "vp %p bp %p num now %d",
690 			    vp, bp, vp->v_numoutput, 0);
691 			bp = getiobuf(vp, true);
692 			nestiobuf_setup(mbp, bp, offset - startoffset, iobytes);
693 		}
694 		bp->b_lblkno = 0;
695 
696 		/* adjust physical blkno for partial blocks */
697 		bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >>
698 		    dev_bshift);
699 
700 		UVMHIST_LOG(ubchist,
701 		    "bp %p offset 0x%x bcount 0x%x blkno 0x%x",
702 		    bp, offset, bp->b_bcount, bp->b_blkno);
703 
704 		VOP_STRATEGY(devvp, bp);
705 	}
706 
707 loopdone:
708 	nestiobuf_done(mbp, skipbytes, error);
709 	if (async) {
710 		UVMHIST_LOG(ubchist, "returning 0 (async)",0,0,0,0);
711 		if (!glocked) {
712 			genfs_node_unlock(vp);
713 		}
714 		return 0;
715 	}
716 	if (bp != NULL) {
717 		error = biowait(mbp);
718 	}
719 
720 	/* Remove the mapping (make KVA available as soon as possible) */
721 	uvm_pagermapout(kva, npages);
722 
723 	/*
724 	 * if this we encountered a hole then we have to do a little more work.
725 	 * for read faults, we marked the page PG_RDONLY so that future
726 	 * write accesses to the page will fault again.
727 	 * for write faults, we must make sure that the backing store for
728 	 * the page is completely allocated while the pages are locked.
729 	 */
730 
731 	if (!error && sawhole && blockalloc) {
732 		error = GOP_ALLOC(vp, startoffset,
733 		    npages << PAGE_SHIFT, 0, cred);
734 		UVMHIST_LOG(ubchist, "gop_alloc off 0x%x/0x%x -> %d",
735 		    startoffset, npages << PAGE_SHIFT, error,0);
736 		if (!error) {
737 			mutex_enter(uobj->vmobjlock);
738 			for (i = 0; i < npages; i++) {
739 				struct vm_page *pg = pgs[i];
740 
741 				if (pg == NULL) {
742 					continue;
743 				}
744 				pg->flags &= ~(PG_CLEAN|PG_RDONLY);
745 				UVMHIST_LOG(ubchist, "mark dirty pg %p",
746 				    pg,0,0,0);
747 			}
748 			mutex_exit(uobj->vmobjlock);
749 		}
750 	}
751 
752 	putiobuf(mbp);
753 	return error;
754 }
755 
756 /*
757  * generic VM putpages routine.
758  * Write the given range of pages to backing store.
759  *
760  * => "offhi == 0" means flush all pages at or after "offlo".
761  * => object should be locked by caller.  we return with the
762  *      object unlocked.
763  * => if PGO_CLEANIT or PGO_SYNCIO is set, we may block (due to I/O).
764  *	thus, a caller might want to unlock higher level resources
765  *	(e.g. vm_map) before calling flush.
766  * => if neither PGO_CLEANIT nor PGO_SYNCIO is set, we will not block
767  * => if PGO_ALLPAGES is set, then all pages in the object will be processed.
768  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
769  *	that new pages are inserted on the tail end of the list.   thus,
770  *	we can make a complete pass through the object in one go by starting
771  *	at the head and working towards the tail (new pages are put in
772  *	front of us).
773  * => NOTE: we are allowed to lock the page queues, so the caller
774  *	must not be holding the page queue lock.
775  *
776  * note on "cleaning" object and PG_BUSY pages:
777  *	this routine is holding the lock on the object.   the only time
778  *	that it can run into a PG_BUSY page that it does not own is if
779  *	some other process has started I/O on the page (e.g. either
780  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
781  *	in, then it can not be dirty (!PG_CLEAN) because no one has
782  *	had a chance to modify it yet.    if the PG_BUSY page is being
783  *	paged out then it means that someone else has already started
784  *	cleaning the page for us (how nice!).    in this case, if we
785  *	have syncio specified, then after we make our pass through the
786  *	object we need to wait for the other PG_BUSY pages to clear
787  *	off (i.e. we need to do an iosync).   also note that once a
788  *	page is PG_BUSY it must stay in its object until it is un-busyed.
789  *
790  * note on page traversal:
791  *	we can traverse the pages in an object either by going down the
792  *	linked list in "uobj->memq", or we can go over the address range
793  *	by page doing hash table lookups for each address.    depending
794  *	on how many pages are in the object it may be cheaper to do one
795  *	or the other.   we set "by_list" to true if we are using memq.
796  *	if the cost of a hash lookup was equal to the cost of the list
797  *	traversal we could compare the number of pages in the start->stop
798  *	range to the total number of pages in the object.   however, it
799  *	seems that a hash table lookup is more expensive than the linked
800  *	list traversal, so we multiply the number of pages in the
801  *	range by an estimate of the relatively higher cost of the hash lookup.
802  */
803 
804 int
805 genfs_putpages(void *v)
806 {
807 	struct vop_putpages_args /* {
808 		struct vnode *a_vp;
809 		voff_t a_offlo;
810 		voff_t a_offhi;
811 		int a_flags;
812 	} */ * const ap = v;
813 
814 	return genfs_do_putpages(ap->a_vp, ap->a_offlo, ap->a_offhi,
815 	    ap->a_flags, NULL);
816 }
817 
818 int
819 genfs_do_putpages(struct vnode *vp, off_t startoff, off_t endoff,
820     int origflags, struct vm_page **busypg)
821 {
822 	struct uvm_object * const uobj = &vp->v_uobj;
823 	kmutex_t * const slock = uobj->vmobjlock;
824 	off_t off;
825 	/* Even for strange MAXPHYS, the shift rounds down to a page */
826 #define maxpages (MAXPHYS >> PAGE_SHIFT)
827 	int i, error, npages, nback;
828 	int freeflag;
829 	struct vm_page *pgs[maxpages], *pg, *nextpg, *tpg, curmp, endmp;
830 	bool wasclean, by_list, needs_clean, yld;
831 	bool async = (origflags & PGO_SYNCIO) == 0;
832 	bool pagedaemon = curlwp == uvm.pagedaemon_lwp;
833 	struct lwp * const l = curlwp ? curlwp : &lwp0;
834 	struct genfs_node * const gp = VTOG(vp);
835 	int flags;
836 	int dirtygen;
837 	bool modified;
838 	bool need_wapbl;
839 	bool has_trans;
840 	bool cleanall;
841 	bool onworklst;
842 
843 	UVMHIST_FUNC("genfs_putpages"); UVMHIST_CALLED(ubchist);
844 
845 	KASSERT(origflags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE));
846 	KASSERT((startoff & PAGE_MASK) == 0 && (endoff & PAGE_MASK) == 0);
847 	KASSERT(startoff < endoff || endoff == 0);
848 
849 	UVMHIST_LOG(ubchist, "vp %p pages %d off 0x%x len 0x%x",
850 	    vp, uobj->uo_npages, startoff, endoff - startoff);
851 
852 	has_trans = false;
853 	need_wapbl = (!pagedaemon && vp->v_mount && vp->v_mount->mnt_wapbl &&
854 	    (origflags & PGO_JOURNALLOCKED) == 0);
855 
856 retry:
857 	modified = false;
858 	flags = origflags;
859 	KASSERT((vp->v_iflag & VI_ONWORKLST) != 0 ||
860 	    (vp->v_iflag & VI_WRMAPDIRTY) == 0);
861 	if (uobj->uo_npages == 0) {
862 		if (vp->v_iflag & VI_ONWORKLST) {
863 			vp->v_iflag &= ~VI_WRMAPDIRTY;
864 			if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL)
865 				vn_syncer_remove_from_worklist(vp);
866 		}
867 		if (has_trans) {
868 			if (need_wapbl)
869 				WAPBL_END(vp->v_mount);
870 			fstrans_done(vp->v_mount);
871 		}
872 		mutex_exit(slock);
873 		return (0);
874 	}
875 
876 	/*
877 	 * the vnode has pages, set up to process the request.
878 	 */
879 
880 	if (!has_trans && (flags & PGO_CLEANIT) != 0) {
881 		mutex_exit(slock);
882 		if (pagedaemon) {
883 			error = fstrans_start_nowait(vp->v_mount, FSTRANS_LAZY);
884 			if (error)
885 				return error;
886 		} else
887 			fstrans_start(vp->v_mount, FSTRANS_LAZY);
888 		if (need_wapbl) {
889 			error = WAPBL_BEGIN(vp->v_mount);
890 			if (error) {
891 				fstrans_done(vp->v_mount);
892 				return error;
893 			}
894 		}
895 		has_trans = true;
896 		mutex_enter(slock);
897 		goto retry;
898 	}
899 
900 	error = 0;
901 	wasclean = (vp->v_numoutput == 0);
902 	off = startoff;
903 	if (endoff == 0 || flags & PGO_ALLPAGES) {
904 		endoff = trunc_page(LLONG_MAX);
905 	}
906 	by_list = (uobj->uo_npages <=
907 	    ((endoff - startoff) >> PAGE_SHIFT) * UVM_PAGE_TREE_PENALTY);
908 
909 	/*
910 	 * if this vnode is known not to have dirty pages,
911 	 * don't bother to clean it out.
912 	 */
913 
914 	if ((vp->v_iflag & VI_ONWORKLST) == 0) {
915 #if !defined(DEBUG)
916 		if ((flags & (PGO_FREE|PGO_DEACTIVATE)) == 0) {
917 			goto skip_scan;
918 		}
919 #endif /* !defined(DEBUG) */
920 		flags &= ~PGO_CLEANIT;
921 	}
922 
923 	/*
924 	 * start the loop.  when scanning by list, hold the last page
925 	 * in the list before we start.  pages allocated after we start
926 	 * will be added to the end of the list, so we can stop at the
927 	 * current last page.
928 	 */
929 
930 	cleanall = (flags & PGO_CLEANIT) != 0 && wasclean &&
931 	    startoff == 0 && endoff == trunc_page(LLONG_MAX) &&
932 	    (vp->v_iflag & VI_ONWORKLST) != 0;
933 	dirtygen = gp->g_dirtygen;
934 	freeflag = pagedaemon ? PG_PAGEOUT : PG_RELEASED;
935 	if (by_list) {
936 		curmp.flags = PG_MARKER;
937 		endmp.flags = PG_MARKER;
938 		pg = TAILQ_FIRST(&uobj->memq);
939 		TAILQ_INSERT_TAIL(&uobj->memq, &endmp, listq.queue);
940 	} else {
941 		pg = uvm_pagelookup(uobj, off);
942 	}
943 	nextpg = NULL;
944 	while (by_list || off < endoff) {
945 
946 		/*
947 		 * if the current page is not interesting, move on to the next.
948 		 */
949 
950 		KASSERT(pg == NULL || pg->uobject == uobj ||
951 		    (pg->flags & PG_MARKER) != 0);
952 		KASSERT(pg == NULL ||
953 		    (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 ||
954 		    (pg->flags & (PG_BUSY|PG_MARKER)) != 0);
955 		if (by_list) {
956 			if (pg == &endmp) {
957 				break;
958 			}
959 			if (pg->flags & PG_MARKER) {
960 				pg = TAILQ_NEXT(pg, listq.queue);
961 				continue;
962 			}
963 			if (pg->offset < startoff || pg->offset >= endoff ||
964 			    pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
965 				if (pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
966 					wasclean = false;
967 				}
968 				pg = TAILQ_NEXT(pg, listq.queue);
969 				continue;
970 			}
971 			off = pg->offset;
972 		} else if (pg == NULL || pg->flags & (PG_RELEASED|PG_PAGEOUT)) {
973 			if (pg != NULL) {
974 				wasclean = false;
975 			}
976 			off += PAGE_SIZE;
977 			if (off < endoff) {
978 				pg = uvm_pagelookup(uobj, off);
979 			}
980 			continue;
981 		}
982 
983 		/*
984 		 * if the current page needs to be cleaned and it's busy,
985 		 * wait for it to become unbusy.
986 		 */
987 
988 		yld = (l->l_cpu->ci_schedstate.spc_flags &
989 		    SPCF_SHOULDYIELD) && !pagedaemon;
990 		if (pg->flags & PG_BUSY || yld) {
991 			UVMHIST_LOG(ubchist, "busy %p", pg,0,0,0);
992 			if (flags & PGO_BUSYFAIL && pg->flags & PG_BUSY) {
993 				UVMHIST_LOG(ubchist, "busyfail %p", pg, 0,0,0);
994 				error = EDEADLK;
995 				if (busypg != NULL)
996 					*busypg = pg;
997 				break;
998 			}
999 			if (pagedaemon) {
1000 				/*
1001 				 * someone has taken the page while we
1002 				 * dropped the lock for fstrans_start.
1003 				 */
1004 				break;
1005 			}
1006 			if (by_list) {
1007 				TAILQ_INSERT_BEFORE(pg, &curmp, listq.queue);
1008 				UVMHIST_LOG(ubchist, "curmp next %p",
1009 				    TAILQ_NEXT(&curmp, listq.queue), 0,0,0);
1010 			}
1011 			if (yld) {
1012 				mutex_exit(slock);
1013 				preempt();
1014 				mutex_enter(slock);
1015 			} else {
1016 				pg->flags |= PG_WANTED;
1017 				UVM_UNLOCK_AND_WAIT(pg, slock, 0, "genput", 0);
1018 				mutex_enter(slock);
1019 			}
1020 			if (by_list) {
1021 				UVMHIST_LOG(ubchist, "after next %p",
1022 				    TAILQ_NEXT(&curmp, listq.queue), 0,0,0);
1023 				pg = TAILQ_NEXT(&curmp, listq.queue);
1024 				TAILQ_REMOVE(&uobj->memq, &curmp, listq.queue);
1025 			} else {
1026 				pg = uvm_pagelookup(uobj, off);
1027 			}
1028 			continue;
1029 		}
1030 
1031 		/*
1032 		 * if we're freeing, remove all mappings of the page now.
1033 		 * if we're cleaning, check if the page is needs to be cleaned.
1034 		 */
1035 
1036 		if (flags & PGO_FREE) {
1037 			pmap_page_protect(pg, VM_PROT_NONE);
1038 		} else if (flags & PGO_CLEANIT) {
1039 
1040 			/*
1041 			 * if we still have some hope to pull this vnode off
1042 			 * from the syncer queue, write-protect the page.
1043 			 */
1044 
1045 			if (cleanall && wasclean &&
1046 			    gp->g_dirtygen == dirtygen) {
1047 
1048 				/*
1049 				 * uobj pages get wired only by uvm_fault
1050 				 * where uobj is locked.
1051 				 */
1052 
1053 				if (pg->wire_count == 0) {
1054 					pmap_page_protect(pg,
1055 					    VM_PROT_READ|VM_PROT_EXECUTE);
1056 				} else {
1057 					cleanall = false;
1058 				}
1059 			}
1060 		}
1061 
1062 		if (flags & PGO_CLEANIT) {
1063 			needs_clean = pmap_clear_modify(pg) ||
1064 			    (pg->flags & PG_CLEAN) == 0;
1065 			pg->flags |= PG_CLEAN;
1066 		} else {
1067 			needs_clean = false;
1068 		}
1069 
1070 		/*
1071 		 * if we're cleaning, build a cluster.
1072 		 * the cluster will consist of pages which are currently dirty,
1073 		 * but they will be returned to us marked clean.
1074 		 * if not cleaning, just operate on the one page.
1075 		 */
1076 
1077 		if (needs_clean) {
1078 			KDASSERT((vp->v_iflag & VI_ONWORKLST));
1079 			wasclean = false;
1080 			memset(pgs, 0, sizeof(pgs));
1081 			pg->flags |= PG_BUSY;
1082 			UVM_PAGE_OWN(pg, "genfs_putpages");
1083 
1084 			/*
1085 			 * first look backward.
1086 			 */
1087 
1088 			npages = MIN(maxpages >> 1, off >> PAGE_SHIFT);
1089 			nback = npages;
1090 			uvn_findpages(uobj, off - PAGE_SIZE, &nback, &pgs[0],
1091 			    UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY|UFP_BACKWARD);
1092 			if (nback) {
1093 				memmove(&pgs[0], &pgs[npages - nback],
1094 				    nback * sizeof(pgs[0]));
1095 				if (npages - nback < nback)
1096 					memset(&pgs[nback], 0,
1097 					    (npages - nback) * sizeof(pgs[0]));
1098 				else
1099 					memset(&pgs[npages - nback], 0,
1100 					    nback * sizeof(pgs[0]));
1101 			}
1102 
1103 			/*
1104 			 * then plug in our page of interest.
1105 			 */
1106 
1107 			pgs[nback] = pg;
1108 
1109 			/*
1110 			 * then look forward to fill in the remaining space in
1111 			 * the array of pages.
1112 			 */
1113 
1114 			npages = maxpages - nback - 1;
1115 			uvn_findpages(uobj, off + PAGE_SIZE, &npages,
1116 			    &pgs[nback + 1],
1117 			    UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY);
1118 			npages += nback + 1;
1119 		} else {
1120 			pgs[0] = pg;
1121 			npages = 1;
1122 			nback = 0;
1123 		}
1124 
1125 		/*
1126 		 * apply FREE or DEACTIVATE options if requested.
1127 		 */
1128 
1129 		if (flags & (PGO_DEACTIVATE|PGO_FREE)) {
1130 			mutex_enter(&uvm_pageqlock);
1131 		}
1132 		for (i = 0; i < npages; i++) {
1133 			tpg = pgs[i];
1134 			KASSERT(tpg->uobject == uobj);
1135 			if (by_list && tpg == TAILQ_NEXT(pg, listq.queue))
1136 				pg = tpg;
1137 			if (tpg->offset < startoff || tpg->offset >= endoff)
1138 				continue;
1139 			if (flags & PGO_DEACTIVATE && tpg->wire_count == 0) {
1140 				uvm_pagedeactivate(tpg);
1141 			} else if (flags & PGO_FREE) {
1142 				pmap_page_protect(tpg, VM_PROT_NONE);
1143 				if (tpg->flags & PG_BUSY) {
1144 					tpg->flags |= freeflag;
1145 					if (pagedaemon) {
1146 						uvm_pageout_start(1);
1147 						uvm_pagedequeue(tpg);
1148 					}
1149 				} else {
1150 
1151 					/*
1152 					 * ``page is not busy''
1153 					 * implies that npages is 1
1154 					 * and needs_clean is false.
1155 					 */
1156 
1157 					nextpg = TAILQ_NEXT(tpg, listq.queue);
1158 					uvm_pagefree(tpg);
1159 					if (pagedaemon)
1160 						uvmexp.pdfreed++;
1161 				}
1162 			}
1163 		}
1164 		if (flags & (PGO_DEACTIVATE|PGO_FREE)) {
1165 			mutex_exit(&uvm_pageqlock);
1166 		}
1167 		if (needs_clean) {
1168 			modified = true;
1169 
1170 			/*
1171 			 * start the i/o.  if we're traversing by list,
1172 			 * keep our place in the list with a marker page.
1173 			 */
1174 
1175 			if (by_list) {
1176 				TAILQ_INSERT_AFTER(&uobj->memq, pg, &curmp,
1177 				    listq.queue);
1178 			}
1179 			mutex_exit(slock);
1180 			error = GOP_WRITE(vp, pgs, npages, flags);
1181 			mutex_enter(slock);
1182 			if (by_list) {
1183 				pg = TAILQ_NEXT(&curmp, listq.queue);
1184 				TAILQ_REMOVE(&uobj->memq, &curmp, listq.queue);
1185 			}
1186 			if (error) {
1187 				break;
1188 			}
1189 			if (by_list) {
1190 				continue;
1191 			}
1192 		}
1193 
1194 		/*
1195 		 * find the next page and continue if there was no error.
1196 		 */
1197 
1198 		if (by_list) {
1199 			if (nextpg) {
1200 				pg = nextpg;
1201 				nextpg = NULL;
1202 			} else {
1203 				pg = TAILQ_NEXT(pg, listq.queue);
1204 			}
1205 		} else {
1206 			off += (npages - nback) << PAGE_SHIFT;
1207 			if (off < endoff) {
1208 				pg = uvm_pagelookup(uobj, off);
1209 			}
1210 		}
1211 	}
1212 	if (by_list) {
1213 		TAILQ_REMOVE(&uobj->memq, &endmp, listq.queue);
1214 	}
1215 
1216 	if (modified && (vp->v_iflag & VI_WRMAPDIRTY) != 0 &&
1217 	    (vp->v_type != VBLK ||
1218 	    (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) {
1219 		GOP_MARKUPDATE(vp, GOP_UPDATE_MODIFIED);
1220 	}
1221 
1222 	/*
1223 	 * if we're cleaning and there was nothing to clean,
1224 	 * take us off the syncer list.  if we started any i/o
1225 	 * and we're doing sync i/o, wait for all writes to finish.
1226 	 */
1227 
1228 	if (cleanall && wasclean && gp->g_dirtygen == dirtygen &&
1229 	    (vp->v_iflag & VI_ONWORKLST) != 0) {
1230 #if defined(DEBUG)
1231 		TAILQ_FOREACH(pg, &uobj->memq, listq.queue) {
1232 			if ((pg->flags & (PG_FAKE | PG_MARKER)) != 0) {
1233 				continue;
1234 			}
1235 			if ((pg->flags & PG_CLEAN) == 0) {
1236 				printf("%s: %p: !CLEAN\n", __func__, pg);
1237 			}
1238 			if (pmap_is_modified(pg)) {
1239 				printf("%s: %p: modified\n", __func__, pg);
1240 			}
1241 		}
1242 #endif /* defined(DEBUG) */
1243 		vp->v_iflag &= ~VI_WRMAPDIRTY;
1244 		if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL)
1245 			vn_syncer_remove_from_worklist(vp);
1246 	}
1247 
1248 #if !defined(DEBUG)
1249 skip_scan:
1250 #endif /* !defined(DEBUG) */
1251 
1252 	/* Wait for output to complete. */
1253 	if (!wasclean && !async && vp->v_numoutput != 0) {
1254 		while (vp->v_numoutput != 0)
1255 			cv_wait(&vp->v_cv, slock);
1256 	}
1257 	onworklst = (vp->v_iflag & VI_ONWORKLST) != 0;
1258 	mutex_exit(slock);
1259 
1260 	if ((flags & PGO_RECLAIM) != 0 && onworklst) {
1261 		/*
1262 		 * in the case of PGO_RECLAIM, ensure to make the vnode clean.
1263 		 * retrying is not a big deal because, in many cases,
1264 		 * uobj->uo_npages is already 0 here.
1265 		 */
1266 		mutex_enter(slock);
1267 		goto retry;
1268 	}
1269 
1270 	if (has_trans) {
1271 		if (need_wapbl)
1272 			WAPBL_END(vp->v_mount);
1273 		fstrans_done(vp->v_mount);
1274 	}
1275 
1276 	return (error);
1277 }
1278 
1279 int
1280 genfs_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, int flags)
1281 {
1282 	off_t off;
1283 	vaddr_t kva;
1284 	size_t len;
1285 	int error;
1286 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1287 
1288 	UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x",
1289 	    vp, pgs, npages, flags);
1290 
1291 	off = pgs[0]->offset;
1292 	kva = uvm_pagermapin(pgs, npages,
1293 	    UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK);
1294 	len = npages << PAGE_SHIFT;
1295 
1296 	error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE,
1297 			    uvm_aio_biodone);
1298 
1299 	return error;
1300 }
1301 
1302 int
1303 genfs_gop_write_rwmap(struct vnode *vp, struct vm_page **pgs, int npages, int flags)
1304 {
1305 	off_t off;
1306 	vaddr_t kva;
1307 	size_t len;
1308 	int error;
1309 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1310 
1311 	UVMHIST_LOG(ubchist, "vp %p pgs %p npages %d flags 0x%x",
1312 	    vp, pgs, npages, flags);
1313 
1314 	off = pgs[0]->offset;
1315 	kva = uvm_pagermapin(pgs, npages,
1316 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
1317 	len = npages << PAGE_SHIFT;
1318 
1319 	error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE,
1320 			    uvm_aio_biodone);
1321 
1322 	return error;
1323 }
1324 
1325 /*
1326  * Backend routine for doing I/O to vnode pages.  Pages are already locked
1327  * and mapped into kernel memory.  Here we just look up the underlying
1328  * device block addresses and call the strategy routine.
1329  */
1330 
1331 static int
1332 genfs_do_io(struct vnode *vp, off_t off, vaddr_t kva, size_t len, int flags,
1333     enum uio_rw rw, void (*iodone)(struct buf *))
1334 {
1335 	int s, error;
1336 	int fs_bshift, dev_bshift;
1337 	off_t eof, offset, startoffset;
1338 	size_t bytes, iobytes, skipbytes;
1339 	struct buf *mbp, *bp;
1340 	const bool async = (flags & PGO_SYNCIO) == 0;
1341 	const bool lazy = (flags & PGO_LAZY) == 0;
1342 	const bool iowrite = rw == UIO_WRITE;
1343 	const int brw = iowrite ? B_WRITE : B_READ;
1344 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist);
1345 
1346 	UVMHIST_LOG(ubchist, "vp %p kva %p len 0x%x flags 0x%x",
1347 	    vp, kva, len, flags);
1348 
1349 	KASSERT(vp->v_size <= vp->v_writesize);
1350 	GOP_SIZE(vp, vp->v_writesize, &eof, 0);
1351 	if (vp->v_type != VBLK) {
1352 		fs_bshift = vp->v_mount->mnt_fs_bshift;
1353 		dev_bshift = vp->v_mount->mnt_dev_bshift;
1354 	} else {
1355 		fs_bshift = DEV_BSHIFT;
1356 		dev_bshift = DEV_BSHIFT;
1357 	}
1358 	error = 0;
1359 	startoffset = off;
1360 	bytes = MIN(len, eof - startoffset);
1361 	skipbytes = 0;
1362 	KASSERT(bytes != 0);
1363 
1364 	if (iowrite) {
1365 		mutex_enter(vp->v_interlock);
1366 		vp->v_numoutput += 2;
1367 		mutex_exit(vp->v_interlock);
1368 	}
1369 	mbp = getiobuf(vp, true);
1370 	UVMHIST_LOG(ubchist, "vp %p mbp %p num now %d bytes 0x%x",
1371 	    vp, mbp, vp->v_numoutput, bytes);
1372 	mbp->b_bufsize = len;
1373 	mbp->b_data = (void *)kva;
1374 	mbp->b_resid = mbp->b_bcount = bytes;
1375 	mbp->b_cflags = BC_BUSY | BC_AGE;
1376 	if (async) {
1377 		mbp->b_flags = brw | B_ASYNC;
1378 		mbp->b_iodone = iodone;
1379 	} else {
1380 		mbp->b_flags = brw;
1381 		mbp->b_iodone = NULL;
1382 	}
1383 	if (curlwp == uvm.pagedaemon_lwp)
1384 		BIO_SETPRIO(mbp, BPRIO_TIMELIMITED);
1385 	else if (async || lazy)
1386 		BIO_SETPRIO(mbp, BPRIO_TIMENONCRITICAL);
1387 	else
1388 		BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL);
1389 
1390 	bp = NULL;
1391 	for (offset = startoffset;
1392 	    bytes > 0;
1393 	    offset += iobytes, bytes -= iobytes) {
1394 		int run;
1395 		daddr_t lbn, blkno;
1396 		struct vnode *devvp;
1397 
1398 		/*
1399 		 * bmap the file to find out the blkno to read from and
1400 		 * how much we can read in one i/o.  if bmap returns an error,
1401 		 * skip the rest of the top-level i/o.
1402 		 */
1403 
1404 		lbn = offset >> fs_bshift;
1405 		error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run);
1406 		if (error) {
1407 			UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%x -> %d\n",
1408 			    lbn,error,0,0);
1409 			skipbytes += bytes;
1410 			bytes = 0;
1411 			goto loopdone;
1412 		}
1413 
1414 		/*
1415 		 * see how many pages can be read with this i/o.
1416 		 * reduce the i/o size if necessary to avoid
1417 		 * overwriting pages with valid data.
1418 		 */
1419 
1420 		iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset,
1421 		    bytes);
1422 
1423 		/*
1424 		 * if this block isn't allocated, zero it instead of
1425 		 * reading it.  unless we are going to allocate blocks,
1426 		 * mark the pages we zeroed PG_RDONLY.
1427 		 */
1428 
1429 		if (blkno == (daddr_t)-1) {
1430 			if (!iowrite) {
1431 				memset((char *)kva + (offset - startoffset), 0,
1432 				    iobytes);
1433 			}
1434 			skipbytes += iobytes;
1435 			continue;
1436 		}
1437 
1438 		/*
1439 		 * allocate a sub-buf for this piece of the i/o
1440 		 * (or just use mbp if there's only 1 piece),
1441 		 * and start it going.
1442 		 */
1443 
1444 		if (offset == startoffset && iobytes == bytes) {
1445 			bp = mbp;
1446 		} else {
1447 			UVMHIST_LOG(ubchist, "vp %p bp %p num now %d",
1448 			    vp, bp, vp->v_numoutput, 0);
1449 			bp = getiobuf(vp, true);
1450 			nestiobuf_setup(mbp, bp, offset - startoffset, iobytes);
1451 		}
1452 		bp->b_lblkno = 0;
1453 
1454 		/* adjust physical blkno for partial blocks */
1455 		bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >>
1456 		    dev_bshift);
1457 
1458 		UVMHIST_LOG(ubchist,
1459 		    "bp %p offset 0x%x bcount 0x%x blkno 0x%x",
1460 		    bp, offset, bp->b_bcount, bp->b_blkno);
1461 
1462 		VOP_STRATEGY(devvp, bp);
1463 	}
1464 
1465 loopdone:
1466 	if (skipbytes) {
1467 		UVMHIST_LOG(ubchist, "skipbytes %d", skipbytes, 0,0,0);
1468 	}
1469 	nestiobuf_done(mbp, skipbytes, error);
1470 	if (async) {
1471 		UVMHIST_LOG(ubchist, "returning 0 (async)", 0,0,0,0);
1472 		return (0);
1473 	}
1474 	UVMHIST_LOG(ubchist, "waiting for mbp %p", mbp,0,0,0);
1475 	error = biowait(mbp);
1476 	s = splbio();
1477 	(*iodone)(mbp);
1478 	splx(s);
1479 	UVMHIST_LOG(ubchist, "returning, error %d", error,0,0,0);
1480 	return (error);
1481 }
1482 
1483 int
1484 genfs_compat_getpages(void *v)
1485 {
1486 	struct vop_getpages_args /* {
1487 		struct vnode *a_vp;
1488 		voff_t a_offset;
1489 		struct vm_page **a_m;
1490 		int *a_count;
1491 		int a_centeridx;
1492 		vm_prot_t a_access_type;
1493 		int a_advice;
1494 		int a_flags;
1495 	} */ *ap = v;
1496 
1497 	off_t origoffset;
1498 	struct vnode *vp = ap->a_vp;
1499 	struct uvm_object *uobj = &vp->v_uobj;
1500 	struct vm_page *pg, **pgs;
1501 	vaddr_t kva;
1502 	int i, error, orignpages, npages;
1503 	struct iovec iov;
1504 	struct uio uio;
1505 	kauth_cred_t cred = curlwp->l_cred;
1506 	const bool memwrite = (ap->a_access_type & VM_PROT_WRITE) != 0;
1507 
1508 	error = 0;
1509 	origoffset = ap->a_offset;
1510 	orignpages = *ap->a_count;
1511 	pgs = ap->a_m;
1512 
1513 	if (ap->a_flags & PGO_LOCKED) {
1514 		uvn_findpages(uobj, origoffset, ap->a_count, ap->a_m,
1515 		    UFP_NOWAIT|UFP_NOALLOC| (memwrite ? UFP_NORDONLY : 0));
1516 
1517 		error = ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0;
1518 		if (error == 0 && memwrite) {
1519 			genfs_markdirty(vp);
1520 		}
1521 		return error;
1522 	}
1523 	if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= vp->v_size) {
1524 		mutex_exit(uobj->vmobjlock);
1525 		return EINVAL;
1526 	}
1527 	if ((ap->a_flags & PGO_SYNCIO) == 0) {
1528 		mutex_exit(uobj->vmobjlock);
1529 		return 0;
1530 	}
1531 	npages = orignpages;
1532 	uvn_findpages(uobj, origoffset, &npages, pgs, UFP_ALL);
1533 	mutex_exit(uobj->vmobjlock);
1534 	kva = uvm_pagermapin(pgs, npages,
1535 	    UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK);
1536 	for (i = 0; i < npages; i++) {
1537 		pg = pgs[i];
1538 		if ((pg->flags & PG_FAKE) == 0) {
1539 			continue;
1540 		}
1541 		iov.iov_base = (char *)kva + (i << PAGE_SHIFT);
1542 		iov.iov_len = PAGE_SIZE;
1543 		uio.uio_iov = &iov;
1544 		uio.uio_iovcnt = 1;
1545 		uio.uio_offset = origoffset + (i << PAGE_SHIFT);
1546 		uio.uio_rw = UIO_READ;
1547 		uio.uio_resid = PAGE_SIZE;
1548 		UIO_SETUP_SYSSPACE(&uio);
1549 		/* XXX vn_lock */
1550 		error = VOP_READ(vp, &uio, 0, cred);
1551 		if (error) {
1552 			break;
1553 		}
1554 		if (uio.uio_resid) {
1555 			memset(iov.iov_base, 0, uio.uio_resid);
1556 		}
1557 	}
1558 	uvm_pagermapout(kva, npages);
1559 	mutex_enter(uobj->vmobjlock);
1560 	mutex_enter(&uvm_pageqlock);
1561 	for (i = 0; i < npages; i++) {
1562 		pg = pgs[i];
1563 		if (error && (pg->flags & PG_FAKE) != 0) {
1564 			pg->flags |= PG_RELEASED;
1565 		} else {
1566 			pmap_clear_modify(pg);
1567 			uvm_pageactivate(pg);
1568 		}
1569 	}
1570 	if (error) {
1571 		uvm_page_unbusy(pgs, npages);
1572 	}
1573 	mutex_exit(&uvm_pageqlock);
1574 	if (error == 0 && memwrite) {
1575 		genfs_markdirty(vp);
1576 	}
1577 	mutex_exit(uobj->vmobjlock);
1578 	return error;
1579 }
1580 
1581 int
1582 genfs_compat_gop_write(struct vnode *vp, struct vm_page **pgs, int npages,
1583     int flags)
1584 {
1585 	off_t offset;
1586 	struct iovec iov;
1587 	struct uio uio;
1588 	kauth_cred_t cred = curlwp->l_cred;
1589 	struct buf *bp;
1590 	vaddr_t kva;
1591 	int error;
1592 
1593 	offset = pgs[0]->offset;
1594 	kva = uvm_pagermapin(pgs, npages,
1595 	    UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK);
1596 
1597 	iov.iov_base = (void *)kva;
1598 	iov.iov_len = npages << PAGE_SHIFT;
1599 	uio.uio_iov = &iov;
1600 	uio.uio_iovcnt = 1;
1601 	uio.uio_offset = offset;
1602 	uio.uio_rw = UIO_WRITE;
1603 	uio.uio_resid = npages << PAGE_SHIFT;
1604 	UIO_SETUP_SYSSPACE(&uio);
1605 	/* XXX vn_lock */
1606 	error = VOP_WRITE(vp, &uio, 0, cred);
1607 
1608 	mutex_enter(vp->v_interlock);
1609 	vp->v_numoutput++;
1610 	mutex_exit(vp->v_interlock);
1611 
1612 	bp = getiobuf(vp, true);
1613 	bp->b_cflags = BC_BUSY | BC_AGE;
1614 	bp->b_lblkno = offset >> vp->v_mount->mnt_fs_bshift;
1615 	bp->b_data = (char *)kva;
1616 	bp->b_bcount = npages << PAGE_SHIFT;
1617 	bp->b_bufsize = npages << PAGE_SHIFT;
1618 	bp->b_resid = 0;
1619 	bp->b_error = error;
1620 	uvm_aio_aiodone(bp);
1621 	return (error);
1622 }
1623 
1624 /*
1625  * Process a uio using direct I/O.  If we reach a part of the request
1626  * which cannot be processed in this fashion for some reason, just return.
1627  * The caller must handle some additional part of the request using
1628  * buffered I/O before trying direct I/O again.
1629  */
1630 
1631 void
1632 genfs_directio(struct vnode *vp, struct uio *uio, int ioflag)
1633 {
1634 	struct vmspace *vs;
1635 	struct iovec *iov;
1636 	vaddr_t va;
1637 	size_t len;
1638 	const int mask = DEV_BSIZE - 1;
1639 	int error;
1640 	bool need_wapbl = (vp->v_mount && vp->v_mount->mnt_wapbl &&
1641 	    (ioflag & IO_JOURNALLOCKED) == 0);
1642 
1643 	/*
1644 	 * We only support direct I/O to user space for now.
1645 	 */
1646 
1647 	if (VMSPACE_IS_KERNEL_P(uio->uio_vmspace)) {
1648 		return;
1649 	}
1650 
1651 	/*
1652 	 * If the vnode is mapped, we would need to get the getpages lock
1653 	 * to stabilize the bmap, but then we would get into trouble while
1654 	 * locking the pages if the pages belong to this same vnode (or a
1655 	 * multi-vnode cascade to the same effect).  Just fall back to
1656 	 * buffered I/O if the vnode is mapped to avoid this mess.
1657 	 */
1658 
1659 	if (vp->v_vflag & VV_MAPPED) {
1660 		return;
1661 	}
1662 
1663 	if (need_wapbl) {
1664 		error = WAPBL_BEGIN(vp->v_mount);
1665 		if (error)
1666 			return;
1667 	}
1668 
1669 	/*
1670 	 * Do as much of the uio as possible with direct I/O.
1671 	 */
1672 
1673 	vs = uio->uio_vmspace;
1674 	while (uio->uio_resid) {
1675 		iov = uio->uio_iov;
1676 		if (iov->iov_len == 0) {
1677 			uio->uio_iov++;
1678 			uio->uio_iovcnt--;
1679 			continue;
1680 		}
1681 		va = (vaddr_t)iov->iov_base;
1682 		len = MIN(iov->iov_len, genfs_maxdio);
1683 		len &= ~mask;
1684 
1685 		/*
1686 		 * If the next chunk is smaller than DEV_BSIZE or extends past
1687 		 * the current EOF, then fall back to buffered I/O.
1688 		 */
1689 
1690 		if (len == 0 || uio->uio_offset + len > vp->v_size) {
1691 			break;
1692 		}
1693 
1694 		/*
1695 		 * Check alignment.  The file offset must be at least
1696 		 * sector-aligned.  The exact constraint on memory alignment
1697 		 * is very hardware-dependent, but requiring sector-aligned
1698 		 * addresses there too is safe.
1699 		 */
1700 
1701 		if (uio->uio_offset & mask || va & mask) {
1702 			break;
1703 		}
1704 		error = genfs_do_directio(vs, va, len, vp, uio->uio_offset,
1705 					  uio->uio_rw);
1706 		if (error) {
1707 			break;
1708 		}
1709 		iov->iov_base = (char *)iov->iov_base + len;
1710 		iov->iov_len -= len;
1711 		uio->uio_offset += len;
1712 		uio->uio_resid -= len;
1713 	}
1714 
1715 	if (need_wapbl)
1716 		WAPBL_END(vp->v_mount);
1717 }
1718 
1719 /*
1720  * Iodone routine for direct I/O.  We don't do much here since the request is
1721  * always synchronous, so the caller will do most of the work after biowait().
1722  */
1723 
1724 static void
1725 genfs_dio_iodone(struct buf *bp)
1726 {
1727 
1728 	KASSERT((bp->b_flags & B_ASYNC) == 0);
1729 	if ((bp->b_flags & B_READ) == 0 && (bp->b_cflags & BC_AGE) != 0) {
1730 		mutex_enter(bp->b_objlock);
1731 		vwakeup(bp);
1732 		mutex_exit(bp->b_objlock);
1733 	}
1734 	putiobuf(bp);
1735 }
1736 
1737 /*
1738  * Process one chunk of a direct I/O request.
1739  */
1740 
1741 static int
1742 genfs_do_directio(struct vmspace *vs, vaddr_t uva, size_t len, struct vnode *vp,
1743     off_t off, enum uio_rw rw)
1744 {
1745 	struct vm_map *map;
1746 	struct pmap *upm, *kpm __unused;
1747 	size_t klen = round_page(uva + len) - trunc_page(uva);
1748 	off_t spoff, epoff;
1749 	vaddr_t kva, puva;
1750 	paddr_t pa;
1751 	vm_prot_t prot;
1752 	int error, rv __diagused, poff, koff;
1753 	const int pgoflags = PGO_CLEANIT | PGO_SYNCIO | PGO_JOURNALLOCKED |
1754 		(rw == UIO_WRITE ? PGO_FREE : 0);
1755 
1756 	/*
1757 	 * For writes, verify that this range of the file already has fully
1758 	 * allocated backing store.  If there are any holes, just punt and
1759 	 * make the caller take the buffered write path.
1760 	 */
1761 
1762 	if (rw == UIO_WRITE) {
1763 		daddr_t lbn, elbn, blkno;
1764 		int bsize, bshift, run;
1765 
1766 		bshift = vp->v_mount->mnt_fs_bshift;
1767 		bsize = 1 << bshift;
1768 		lbn = off >> bshift;
1769 		elbn = (off + len + bsize - 1) >> bshift;
1770 		while (lbn < elbn) {
1771 			error = VOP_BMAP(vp, lbn, NULL, &blkno, &run);
1772 			if (error) {
1773 				return error;
1774 			}
1775 			if (blkno == (daddr_t)-1) {
1776 				return ENOSPC;
1777 			}
1778 			lbn += 1 + run;
1779 		}
1780 	}
1781 
1782 	/*
1783 	 * Flush any cached pages for parts of the file that we're about to
1784 	 * access.  If we're writing, invalidate pages as well.
1785 	 */
1786 
1787 	spoff = trunc_page(off);
1788 	epoff = round_page(off + len);
1789 	mutex_enter(vp->v_interlock);
1790 	error = VOP_PUTPAGES(vp, spoff, epoff, pgoflags);
1791 	if (error) {
1792 		return error;
1793 	}
1794 
1795 	/*
1796 	 * Wire the user pages and remap them into kernel memory.
1797 	 */
1798 
1799 	prot = rw == UIO_READ ? VM_PROT_READ | VM_PROT_WRITE : VM_PROT_READ;
1800 	error = uvm_vslock(vs, (void *)uva, len, prot);
1801 	if (error) {
1802 		return error;
1803 	}
1804 
1805 	map = &vs->vm_map;
1806 	upm = vm_map_pmap(map);
1807 	kpm = vm_map_pmap(kernel_map);
1808 	puva = trunc_page(uva);
1809 	kva = uvm_km_alloc(kernel_map, klen, atop(puva) & uvmexp.colormask,
1810 	    UVM_KMF_VAONLY | UVM_KMF_WAITVA | UVM_KMF_COLORMATCH);
1811 	for (poff = 0; poff < klen; poff += PAGE_SIZE) {
1812 		rv = pmap_extract(upm, puva + poff, &pa);
1813 		KASSERT(rv);
1814 		pmap_kenter_pa(kva + poff, pa, prot, PMAP_WIRED);
1815 	}
1816 	pmap_update(kpm);
1817 
1818 	/*
1819 	 * Do the I/O.
1820 	 */
1821 
1822 	koff = uva - trunc_page(uva);
1823 	error = genfs_do_io(vp, off, kva + koff, len, PGO_SYNCIO, rw,
1824 			    genfs_dio_iodone);
1825 
1826 	/*
1827 	 * Tear down the kernel mapping.
1828 	 */
1829 
1830 	pmap_kremove(kva, klen);
1831 	pmap_update(kpm);
1832 	uvm_km_free(kernel_map, kva, klen, UVM_KMF_VAONLY);
1833 
1834 	/*
1835 	 * Unwire the user pages.
1836 	 */
1837 
1838 	uvm_vsunlock(vs, (void *)uva, len);
1839 	return error;
1840 }
1841