xref: /freebsd/sys/vm/vnode_pager.c (revision bdafb02fcb88389fd1ab684cfe734cb429d35618)
1 /*-
2  * SPDX-License-Identifier: BSD-4-Clause
3  *
4  * Copyright (c) 1990 University of Utah.
5  * Copyright (c) 1991 The Regents of the University of California.
6  * All rights reserved.
7  * Copyright (c) 1993, 1994 John S. Dyson
8  * Copyright (c) 1995, David Greenman
9  *
10  * This code is derived from software contributed to Berkeley by
11  * the Systems Programming Group of the University of Utah Computer
12  * Science Department.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  * 3. All advertising materials mentioning features or use of this software
23  *    must display the following acknowledgement:
24  *	This product includes software developed by the University of
25  *	California, Berkeley and its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	from: @(#)vnode_pager.c	7.5 (Berkeley) 4/20/91
43  */
44 
45 /*
46  * Page to/from files (vnodes).
47  */
48 
49 /*
50  * TODO:
51  *	Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
52  *	greatly re-simplify the vnode_pager.
53  */
54 
55 #include <sys/cdefs.h>
56 __FBSDID("$FreeBSD$");
57 
58 #include "opt_vm.h"
59 
60 #include <sys/param.h>
61 #include <sys/systm.h>
62 #include <sys/sysctl.h>
63 #include <sys/proc.h>
64 #include <sys/vnode.h>
65 #include <sys/mount.h>
66 #include <sys/bio.h>
67 #include <sys/buf.h>
68 #include <sys/vmmeter.h>
69 #include <sys/limits.h>
70 #include <sys/conf.h>
71 #include <sys/rwlock.h>
72 #include <sys/sf_buf.h>
73 #include <sys/domainset.h>
74 
75 #include <machine/atomic.h>
76 
77 #include <vm/vm.h>
78 #include <vm/vm_param.h>
79 #include <vm/vm_object.h>
80 #include <vm/vm_page.h>
81 #include <vm/vm_pager.h>
82 #include <vm/vm_map.h>
83 #include <vm/vnode_pager.h>
84 #include <vm/vm_extern.h>
85 
86 static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
87     daddr_t *rtaddress, int *run);
88 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
89 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
90 static void vnode_pager_dealloc(vm_object_t);
91 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *);
92 static int vnode_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
93     int *, vop_getpages_iodone_t, void *);
94 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *);
95 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
96 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t,
97     vm_ooffset_t, struct ucred *cred);
98 static int vnode_pager_generic_getpages_done(struct buf *);
99 static void vnode_pager_generic_getpages_done_async(struct buf *);
100 
101 struct pagerops vnodepagerops = {
102 	.pgo_alloc =	vnode_pager_alloc,
103 	.pgo_dealloc =	vnode_pager_dealloc,
104 	.pgo_getpages =	vnode_pager_getpages,
105 	.pgo_getpages_async = vnode_pager_getpages_async,
106 	.pgo_putpages =	vnode_pager_putpages,
107 	.pgo_haspage =	vnode_pager_haspage,
108 };
109 
110 int vnode_pbuf_freecnt;
111 int vnode_async_pbuf_freecnt;
112 
113 static struct domainset *vnode_domainset = NULL;
114 
115 SYSCTL_PROC(_debug, OID_AUTO, vnode_domainset, CTLTYPE_STRING | CTLFLAG_RW,
116     &vnode_domainset, 0, sysctl_handle_domainset, "A",
117     "Default vnode NUMA policy");
118 
119 /* Create the VM system backing object for this vnode */
120 int
121 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
122 {
123 	vm_object_t object;
124 	vm_ooffset_t size = isize;
125 	struct vattr va;
126 
127 	if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE)
128 		return (0);
129 
130 	while ((object = vp->v_object) != NULL) {
131 		VM_OBJECT_WLOCK(object);
132 		if (!(object->flags & OBJ_DEAD)) {
133 			VM_OBJECT_WUNLOCK(object);
134 			return (0);
135 		}
136 		VOP_UNLOCK(vp, 0);
137 		vm_object_set_flag(object, OBJ_DISCONNECTWNT);
138 		VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vodead", 0);
139 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
140 	}
141 
142 	if (size == 0) {
143 		if (vn_isdisk(vp, NULL)) {
144 			size = IDX_TO_OFF(INT_MAX);
145 		} else {
146 			if (VOP_GETATTR(vp, &va, td->td_ucred))
147 				return (0);
148 			size = va.va_size;
149 		}
150 	}
151 
152 	object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred);
153 	/*
154 	 * Dereference the reference we just created.  This assumes
155 	 * that the object is associated with the vp.
156 	 */
157 	VM_OBJECT_WLOCK(object);
158 	object->ref_count--;
159 	VM_OBJECT_WUNLOCK(object);
160 	vrele(vp);
161 
162 	KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
163 
164 	return (0);
165 }
166 
167 void
168 vnode_destroy_vobject(struct vnode *vp)
169 {
170 	struct vm_object *obj;
171 
172 	obj = vp->v_object;
173 	if (obj == NULL)
174 		return;
175 	ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
176 	VM_OBJECT_WLOCK(obj);
177 	umtx_shm_object_terminated(obj);
178 	if (obj->ref_count == 0) {
179 		/*
180 		 * don't double-terminate the object
181 		 */
182 		if ((obj->flags & OBJ_DEAD) == 0) {
183 			vm_object_terminate(obj);
184 		} else {
185 			/*
186 			 * Waiters were already handled during object
187 			 * termination.  The exclusive vnode lock hopefully
188 			 * prevented new waiters from referencing the dying
189 			 * object.
190 			 */
191 			KASSERT((obj->flags & OBJ_DISCONNECTWNT) == 0,
192 			    ("OBJ_DISCONNECTWNT set obj %p flags %x",
193 			    obj, obj->flags));
194 			vp->v_object = NULL;
195 			VM_OBJECT_WUNLOCK(obj);
196 		}
197 	} else {
198 		/*
199 		 * Woe to the process that tries to page now :-).
200 		 */
201 		vm_pager_deallocate(obj);
202 		VM_OBJECT_WUNLOCK(obj);
203 	}
204 	KASSERT(vp->v_object == NULL, ("vp %p obj %p", vp, vp->v_object));
205 }
206 
207 
208 /*
209  * Allocate (or lookup) pager for a vnode.
210  * Handle is a vnode pointer.
211  *
212  * MPSAFE
213  */
214 vm_object_t
215 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
216     vm_ooffset_t offset, struct ucred *cred)
217 {
218 	vm_object_t object;
219 	struct vnode *vp;
220 
221 	/*
222 	 * Pageout to vnode, no can do yet.
223 	 */
224 	if (handle == NULL)
225 		return (NULL);
226 
227 	vp = (struct vnode *) handle;
228 
229 	/*
230 	 * If the object is being terminated, wait for it to
231 	 * go away.
232 	 */
233 retry:
234 	while ((object = vp->v_object) != NULL) {
235 		VM_OBJECT_WLOCK(object);
236 		if ((object->flags & OBJ_DEAD) == 0)
237 			break;
238 		vm_object_set_flag(object, OBJ_DISCONNECTWNT);
239 		VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vadead", 0);
240 	}
241 
242 	KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference"));
243 
244 	if (object == NULL) {
245 		/*
246 		 * Add an object of the appropriate size
247 		 */
248 		object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
249 
250 		object->un_pager.vnp.vnp_size = size;
251 		object->un_pager.vnp.writemappings = 0;
252 		object->domain.dr_policy = vnode_domainset;
253 
254 		object->handle = handle;
255 		VI_LOCK(vp);
256 		if (vp->v_object != NULL) {
257 			/*
258 			 * Object has been created while we were sleeping
259 			 */
260 			VI_UNLOCK(vp);
261 			VM_OBJECT_WLOCK(object);
262 			KASSERT(object->ref_count == 1,
263 			    ("leaked ref %p %d", object, object->ref_count));
264 			object->type = OBJT_DEAD;
265 			object->ref_count = 0;
266 			VM_OBJECT_WUNLOCK(object);
267 			vm_object_destroy(object);
268 			goto retry;
269 		}
270 		vp->v_object = object;
271 		VI_UNLOCK(vp);
272 	} else {
273 		object->ref_count++;
274 #if VM_NRESERVLEVEL > 0
275 		vm_object_color(object, 0);
276 #endif
277 		VM_OBJECT_WUNLOCK(object);
278 	}
279 	vrefact(vp);
280 	return (object);
281 }
282 
283 /*
284  *	The object must be locked.
285  */
286 static void
287 vnode_pager_dealloc(vm_object_t object)
288 {
289 	struct vnode *vp;
290 	int refs;
291 
292 	vp = object->handle;
293 	if (vp == NULL)
294 		panic("vnode_pager_dealloc: pager already dealloced");
295 
296 	VM_OBJECT_ASSERT_WLOCKED(object);
297 	vm_object_pip_wait(object, "vnpdea");
298 	refs = object->ref_count;
299 
300 	object->handle = NULL;
301 	object->type = OBJT_DEAD;
302 	if (object->flags & OBJ_DISCONNECTWNT) {
303 		vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
304 		wakeup(object);
305 	}
306 	ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
307 	if (object->un_pager.vnp.writemappings > 0) {
308 		object->un_pager.vnp.writemappings = 0;
309 		VOP_ADD_WRITECOUNT(vp, -1);
310 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
311 		    __func__, vp, vp->v_writecount);
312 	}
313 	vp->v_object = NULL;
314 	VOP_UNSET_TEXT(vp);
315 	VM_OBJECT_WUNLOCK(object);
316 	while (refs-- > 0)
317 		vunref(vp);
318 	VM_OBJECT_WLOCK(object);
319 }
320 
321 static boolean_t
322 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
323     int *after)
324 {
325 	struct vnode *vp = object->handle;
326 	daddr_t bn;
327 	int err;
328 	daddr_t reqblock;
329 	int poff;
330 	int bsize;
331 	int pagesperblock, blocksperpage;
332 
333 	VM_OBJECT_ASSERT_WLOCKED(object);
334 	/*
335 	 * If no vp or vp is doomed or marked transparent to VM, we do not
336 	 * have the page.
337 	 */
338 	if (vp == NULL || vp->v_iflag & VI_DOOMED)
339 		return FALSE;
340 	/*
341 	 * If the offset is beyond end of file we do
342 	 * not have the page.
343 	 */
344 	if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
345 		return FALSE;
346 
347 	bsize = vp->v_mount->mnt_stat.f_iosize;
348 	pagesperblock = bsize / PAGE_SIZE;
349 	blocksperpage = 0;
350 	if (pagesperblock > 0) {
351 		reqblock = pindex / pagesperblock;
352 	} else {
353 		blocksperpage = (PAGE_SIZE / bsize);
354 		reqblock = pindex * blocksperpage;
355 	}
356 	VM_OBJECT_WUNLOCK(object);
357 	err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
358 	VM_OBJECT_WLOCK(object);
359 	if (err)
360 		return TRUE;
361 	if (bn == -1)
362 		return FALSE;
363 	if (pagesperblock > 0) {
364 		poff = pindex - (reqblock * pagesperblock);
365 		if (before) {
366 			*before *= pagesperblock;
367 			*before += poff;
368 		}
369 		if (after) {
370 			/*
371 			 * The BMAP vop can report a partial block in the
372 			 * 'after', but must not report blocks after EOF.
373 			 * Assert the latter, and truncate 'after' in case
374 			 * of the former.
375 			 */
376 			KASSERT((reqblock + *after) * pagesperblock <
377 			    roundup2(object->size, pagesperblock),
378 			    ("%s: reqblock %jd after %d size %ju", __func__,
379 			    (intmax_t )reqblock, *after,
380 			    (uintmax_t )object->size));
381 			*after *= pagesperblock;
382 			*after += pagesperblock - (poff + 1);
383 			if (pindex + *after >= object->size)
384 				*after = object->size - 1 - pindex;
385 		}
386 	} else {
387 		if (before) {
388 			*before /= blocksperpage;
389 		}
390 
391 		if (after) {
392 			*after /= blocksperpage;
393 		}
394 	}
395 	return TRUE;
396 }
397 
398 /*
399  * Lets the VM system know about a change in size for a file.
400  * We adjust our own internal size and flush any cached pages in
401  * the associated object that are affected by the size change.
402  *
403  * Note: this routine may be invoked as a result of a pager put
404  * operation (possibly at object termination time), so we must be careful.
405  */
406 void
407 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
408 {
409 	vm_object_t object;
410 	vm_page_t m;
411 	vm_pindex_t nobjsize;
412 
413 	if ((object = vp->v_object) == NULL)
414 		return;
415 /* 	ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */
416 	VM_OBJECT_WLOCK(object);
417 	if (object->type == OBJT_DEAD) {
418 		VM_OBJECT_WUNLOCK(object);
419 		return;
420 	}
421 	KASSERT(object->type == OBJT_VNODE,
422 	    ("not vnode-backed object %p", object));
423 	if (nsize == object->un_pager.vnp.vnp_size) {
424 		/*
425 		 * Hasn't changed size
426 		 */
427 		VM_OBJECT_WUNLOCK(object);
428 		return;
429 	}
430 	nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
431 	if (nsize < object->un_pager.vnp.vnp_size) {
432 		/*
433 		 * File has shrunk. Toss any cached pages beyond the new EOF.
434 		 */
435 		if (nobjsize < object->size)
436 			vm_object_page_remove(object, nobjsize, object->size,
437 			    0);
438 		/*
439 		 * this gets rid of garbage at the end of a page that is now
440 		 * only partially backed by the vnode.
441 		 *
442 		 * XXX for some reason (I don't know yet), if we take a
443 		 * completely invalid page and mark it partially valid
444 		 * it can screw up NFS reads, so we don't allow the case.
445 		 */
446 		if ((nsize & PAGE_MASK) &&
447 		    (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
448 		    m->valid != 0) {
449 			int base = (int)nsize & PAGE_MASK;
450 			int size = PAGE_SIZE - base;
451 
452 			/*
453 			 * Clear out partial-page garbage in case
454 			 * the page has been mapped.
455 			 */
456 			pmap_zero_page_area(m, base, size);
457 
458 			/*
459 			 * Update the valid bits to reflect the blocks that
460 			 * have been zeroed.  Some of these valid bits may
461 			 * have already been set.
462 			 */
463 			vm_page_set_valid_range(m, base, size);
464 
465 			/*
466 			 * Round "base" to the next block boundary so that the
467 			 * dirty bit for a partially zeroed block is not
468 			 * cleared.
469 			 */
470 			base = roundup2(base, DEV_BSIZE);
471 
472 			/*
473 			 * Clear out partial-page dirty bits.
474 			 *
475 			 * note that we do not clear out the valid
476 			 * bits.  This would prevent bogus_page
477 			 * replacement from working properly.
478 			 */
479 			vm_page_clear_dirty(m, base, PAGE_SIZE - base);
480 		}
481 	}
482 	object->un_pager.vnp.vnp_size = nsize;
483 	object->size = nobjsize;
484 	VM_OBJECT_WUNLOCK(object);
485 }
486 
487 /*
488  * calculate the linear (byte) disk address of specified virtual
489  * file address
490  */
491 static int
492 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
493     int *run)
494 {
495 	int bsize;
496 	int err;
497 	daddr_t vblock;
498 	daddr_t voffset;
499 
500 	if (address < 0)
501 		return -1;
502 
503 	if (vp->v_iflag & VI_DOOMED)
504 		return -1;
505 
506 	bsize = vp->v_mount->mnt_stat.f_iosize;
507 	vblock = address / bsize;
508 	voffset = address % bsize;
509 
510 	err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
511 	if (err == 0) {
512 		if (*rtaddress != -1)
513 			*rtaddress += voffset / DEV_BSIZE;
514 		if (run) {
515 			*run += 1;
516 			*run *= bsize/PAGE_SIZE;
517 			*run -= voffset/PAGE_SIZE;
518 		}
519 	}
520 
521 	return (err);
522 }
523 
524 /*
525  * small block filesystem vnode pager input
526  */
527 static int
528 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m)
529 {
530 	struct vnode *vp;
531 	struct bufobj *bo;
532 	struct buf *bp;
533 	struct sf_buf *sf;
534 	daddr_t fileaddr;
535 	vm_offset_t bsize;
536 	vm_page_bits_t bits;
537 	int error, i;
538 
539 	error = 0;
540 	vp = object->handle;
541 	if (vp->v_iflag & VI_DOOMED)
542 		return VM_PAGER_BAD;
543 
544 	bsize = vp->v_mount->mnt_stat.f_iosize;
545 
546 	VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
547 
548 	sf = sf_buf_alloc(m, 0);
549 
550 	for (i = 0; i < PAGE_SIZE / bsize; i++) {
551 		vm_ooffset_t address;
552 
553 		bits = vm_page_bits(i * bsize, bsize);
554 		if (m->valid & bits)
555 			continue;
556 
557 		address = IDX_TO_OFF(m->pindex) + i * bsize;
558 		if (address >= object->un_pager.vnp.vnp_size) {
559 			fileaddr = -1;
560 		} else {
561 			error = vnode_pager_addr(vp, address, &fileaddr, NULL);
562 			if (error)
563 				break;
564 		}
565 		if (fileaddr != -1) {
566 			bp = getpbuf(&vnode_pbuf_freecnt);
567 
568 			/* build a minimal buffer header */
569 			bp->b_iocmd = BIO_READ;
570 			bp->b_iodone = bdone;
571 			KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
572 			KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
573 			bp->b_rcred = crhold(curthread->td_ucred);
574 			bp->b_wcred = crhold(curthread->td_ucred);
575 			bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
576 			bp->b_blkno = fileaddr;
577 			pbgetbo(bo, bp);
578 			bp->b_vp = vp;
579 			bp->b_bcount = bsize;
580 			bp->b_bufsize = bsize;
581 			bp->b_runningbufspace = bp->b_bufsize;
582 			atomic_add_long(&runningbufspace, bp->b_runningbufspace);
583 
584 			/* do the input */
585 			bp->b_iooffset = dbtob(bp->b_blkno);
586 			bstrategy(bp);
587 
588 			bwait(bp, PVM, "vnsrd");
589 
590 			if ((bp->b_ioflags & BIO_ERROR) != 0)
591 				error = EIO;
592 
593 			/*
594 			 * free the buffer header back to the swap buffer pool
595 			 */
596 			bp->b_vp = NULL;
597 			pbrelbo(bp);
598 			relpbuf(bp, &vnode_pbuf_freecnt);
599 			if (error)
600 				break;
601 		} else
602 			bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
603 		KASSERT((m->dirty & bits) == 0,
604 		    ("vnode_pager_input_smlfs: page %p is dirty", m));
605 		VM_OBJECT_WLOCK(object);
606 		m->valid |= bits;
607 		VM_OBJECT_WUNLOCK(object);
608 	}
609 	sf_buf_free(sf);
610 	if (error) {
611 		return VM_PAGER_ERROR;
612 	}
613 	return VM_PAGER_OK;
614 }
615 
616 /*
617  * old style vnode pager input routine
618  */
619 static int
620 vnode_pager_input_old(vm_object_t object, vm_page_t m)
621 {
622 	struct uio auio;
623 	struct iovec aiov;
624 	int error;
625 	int size;
626 	struct sf_buf *sf;
627 	struct vnode *vp;
628 
629 	VM_OBJECT_ASSERT_WLOCKED(object);
630 	error = 0;
631 
632 	/*
633 	 * Return failure if beyond current EOF
634 	 */
635 	if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
636 		return VM_PAGER_BAD;
637 	} else {
638 		size = PAGE_SIZE;
639 		if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
640 			size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
641 		vp = object->handle;
642 		VM_OBJECT_WUNLOCK(object);
643 
644 		/*
645 		 * Allocate a kernel virtual address and initialize so that
646 		 * we can use VOP_READ/WRITE routines.
647 		 */
648 		sf = sf_buf_alloc(m, 0);
649 
650 		aiov.iov_base = (caddr_t)sf_buf_kva(sf);
651 		aiov.iov_len = size;
652 		auio.uio_iov = &aiov;
653 		auio.uio_iovcnt = 1;
654 		auio.uio_offset = IDX_TO_OFF(m->pindex);
655 		auio.uio_segflg = UIO_SYSSPACE;
656 		auio.uio_rw = UIO_READ;
657 		auio.uio_resid = size;
658 		auio.uio_td = curthread;
659 
660 		error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
661 		if (!error) {
662 			int count = size - auio.uio_resid;
663 
664 			if (count == 0)
665 				error = EINVAL;
666 			else if (count != PAGE_SIZE)
667 				bzero((caddr_t)sf_buf_kva(sf) + count,
668 				    PAGE_SIZE - count);
669 		}
670 		sf_buf_free(sf);
671 
672 		VM_OBJECT_WLOCK(object);
673 	}
674 	KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
675 	if (!error)
676 		m->valid = VM_PAGE_BITS_ALL;
677 	return error ? VM_PAGER_ERROR : VM_PAGER_OK;
678 }
679 
680 /*
681  * generic vnode pager input routine
682  */
683 
684 /*
685  * Local media VFS's that do not implement their own VOP_GETPAGES
686  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
687  * to implement the previous behaviour.
688  *
689  * All other FS's should use the bypass to get to the local media
690  * backing vp's VOP_GETPAGES.
691  */
692 static int
693 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
694     int *rahead)
695 {
696 	struct vnode *vp;
697 	int rtval;
698 
699 	vp = object->handle;
700 	VM_OBJECT_WUNLOCK(object);
701 	rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead);
702 	KASSERT(rtval != EOPNOTSUPP,
703 	    ("vnode_pager: FS getpages not implemented\n"));
704 	VM_OBJECT_WLOCK(object);
705 	return rtval;
706 }
707 
708 static int
709 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
710     int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg)
711 {
712 	struct vnode *vp;
713 	int rtval;
714 
715 	vp = object->handle;
716 	VM_OBJECT_WUNLOCK(object);
717 	rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg);
718 	KASSERT(rtval != EOPNOTSUPP,
719 	    ("vnode_pager: FS getpages_async not implemented\n"));
720 	VM_OBJECT_WLOCK(object);
721 	return (rtval);
722 }
723 
724 /*
725  * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for
726  * local filesystems, where partially valid pages can only occur at
727  * the end of file.
728  */
729 int
730 vnode_pager_local_getpages(struct vop_getpages_args *ap)
731 {
732 
733 	return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
734 	    ap->a_rbehind, ap->a_rahead, NULL, NULL));
735 }
736 
737 int
738 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap)
739 {
740 
741 	return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
742 	    ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg));
743 }
744 
745 /*
746  * This is now called from local media FS's to operate against their
747  * own vnodes if they fail to implement VOP_GETPAGES.
748  */
749 int
750 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count,
751     int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg)
752 {
753 	vm_object_t object;
754 	struct bufobj *bo;
755 	struct buf *bp;
756 	off_t foff;
757 #ifdef INVARIANTS
758 	off_t blkno0;
759 #endif
760 	int bsize, pagesperblock, *freecnt;
761 	int error, before, after, rbehind, rahead, poff, i;
762 	int bytecount, secmask;
763 
764 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
765 	    ("%s does not support devices", __func__));
766 
767 	if (vp->v_iflag & VI_DOOMED)
768 		return (VM_PAGER_BAD);
769 
770 	object = vp->v_object;
771 	foff = IDX_TO_OFF(m[0]->pindex);
772 	bsize = vp->v_mount->mnt_stat.f_iosize;
773 	pagesperblock = bsize / PAGE_SIZE;
774 
775 	KASSERT(foff < object->un_pager.vnp.vnp_size,
776 	    ("%s: page %p offset beyond vp %p size", __func__, m[0], vp));
777 	KASSERT(count <= sizeof(bp->b_pages),
778 	    ("%s: requested %d pages", __func__, count));
779 
780 	/*
781 	 * The last page has valid blocks.  Invalid part can only
782 	 * exist at the end of file, and the page is made fully valid
783 	 * by zeroing in vm_pager_get_pages().
784 	 */
785 	if (m[count - 1]->valid != 0 && --count == 0) {
786 		if (iodone != NULL)
787 			iodone(arg, m, 1, 0);
788 		return (VM_PAGER_OK);
789 	}
790 
791 	/*
792 	 * Synchronous and asynchronous paging operations use different
793 	 * free pbuf counters.  This is done to avoid asynchronous requests
794 	 * to consume all pbufs.
795 	 * Allocate the pbuf at the very beginning of the function, so that
796 	 * if we are low on certain kind of pbufs don't even proceed to BMAP,
797 	 * but sleep.
798 	 */
799 	freecnt = iodone != NULL ?
800 	    &vnode_async_pbuf_freecnt : &vnode_pbuf_freecnt;
801 	bp = getpbuf(freecnt);
802 
803 	/*
804 	 * Get the underlying device blocks for the file with VOP_BMAP().
805 	 * If the file system doesn't support VOP_BMAP, use old way of
806 	 * getting pages via VOP_READ.
807 	 */
808 	error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before);
809 	if (error == EOPNOTSUPP) {
810 		relpbuf(bp, freecnt);
811 		VM_OBJECT_WLOCK(object);
812 		for (i = 0; i < count; i++) {
813 			VM_CNT_INC(v_vnodein);
814 			VM_CNT_INC(v_vnodepgsin);
815 			error = vnode_pager_input_old(object, m[i]);
816 			if (error)
817 				break;
818 		}
819 		VM_OBJECT_WUNLOCK(object);
820 		return (error);
821 	} else if (error != 0) {
822 		relpbuf(bp, freecnt);
823 		return (VM_PAGER_ERROR);
824 	}
825 
826 	/*
827 	 * If the file system supports BMAP, but blocksize is smaller
828 	 * than a page size, then use special small filesystem code.
829 	 */
830 	if (pagesperblock == 0) {
831 		relpbuf(bp, freecnt);
832 		for (i = 0; i < count; i++) {
833 			VM_CNT_INC(v_vnodein);
834 			VM_CNT_INC(v_vnodepgsin);
835 			error = vnode_pager_input_smlfs(object, m[i]);
836 			if (error)
837 				break;
838 		}
839 		return (error);
840 	}
841 
842 	/*
843 	 * A sparse file can be encountered only for a single page request,
844 	 * which may not be preceded by call to vm_pager_haspage().
845 	 */
846 	if (bp->b_blkno == -1) {
847 		KASSERT(count == 1,
848 		    ("%s: array[%d] request to a sparse file %p", __func__,
849 		    count, vp));
850 		relpbuf(bp, freecnt);
851 		pmap_zero_page(m[0]);
852 		KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty",
853 		    __func__, m[0]));
854 		VM_OBJECT_WLOCK(object);
855 		m[0]->valid = VM_PAGE_BITS_ALL;
856 		VM_OBJECT_WUNLOCK(object);
857 		return (VM_PAGER_OK);
858 	}
859 
860 #ifdef INVARIANTS
861 	blkno0 = bp->b_blkno;
862 #endif
863 	bp->b_blkno += (foff % bsize) / DEV_BSIZE;
864 
865 	/* Recalculate blocks available after/before to pages. */
866 	poff = (foff % bsize) / PAGE_SIZE;
867 	before *= pagesperblock;
868 	before += poff;
869 	after *= pagesperblock;
870 	after += pagesperblock - (poff + 1);
871 	if (m[0]->pindex + after >= object->size)
872 		after = object->size - 1 - m[0]->pindex;
873 	KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d",
874 	    __func__, count, after + 1));
875 	after -= count - 1;
876 
877 	/* Trim requested rbehind/rahead to possible values. */
878 	rbehind = a_rbehind ? *a_rbehind : 0;
879 	rahead = a_rahead ? *a_rahead : 0;
880 	rbehind = min(rbehind, before);
881 	rbehind = min(rbehind, m[0]->pindex);
882 	rahead = min(rahead, after);
883 	rahead = min(rahead, object->size - m[count - 1]->pindex);
884 	/*
885 	 * Check that total amount of pages fit into buf.  Trim rbehind and
886 	 * rahead evenly if not.
887 	 */
888 	if (rbehind + rahead + count > nitems(bp->b_pages)) {
889 		int trim, sum;
890 
891 		trim = rbehind + rahead + count - nitems(bp->b_pages) + 1;
892 		sum = rbehind + rahead;
893 		if (rbehind == before) {
894 			/* Roundup rbehind trim to block size. */
895 			rbehind -= roundup(trim * rbehind / sum, pagesperblock);
896 			if (rbehind < 0)
897 				rbehind = 0;
898 		} else
899 			rbehind -= trim * rbehind / sum;
900 		rahead -= trim * rahead / sum;
901 	}
902 	KASSERT(rbehind + rahead + count <= nitems(bp->b_pages),
903 	    ("%s: behind %d ahead %d count %d", __func__,
904 	    rbehind, rahead, count));
905 
906 	/*
907 	 * Fill in the bp->b_pages[] array with requested and optional
908 	 * read behind or read ahead pages.  Read behind pages are looked
909 	 * up in a backward direction, down to a first cached page.  Same
910 	 * for read ahead pages, but there is no need to shift the array
911 	 * in case of encountering a cached page.
912 	 */
913 	i = bp->b_npages = 0;
914 	if (rbehind) {
915 		vm_pindex_t startpindex, tpindex;
916 		vm_page_t p;
917 
918 		VM_OBJECT_WLOCK(object);
919 		startpindex = m[0]->pindex - rbehind;
920 		if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL &&
921 		    p->pindex >= startpindex)
922 			startpindex = p->pindex + 1;
923 
924 		/* tpindex is unsigned; beware of numeric underflow. */
925 		for (tpindex = m[0]->pindex - 1;
926 		    tpindex >= startpindex && tpindex < m[0]->pindex;
927 		    tpindex--, i++) {
928 			p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
929 			if (p == NULL) {
930 				/* Shift the array. */
931 				for (int j = 0; j < i; j++)
932 					bp->b_pages[j] = bp->b_pages[j +
933 					    tpindex + 1 - startpindex];
934 				break;
935 			}
936 			bp->b_pages[tpindex - startpindex] = p;
937 		}
938 
939 		bp->b_pgbefore = i;
940 		bp->b_npages += i;
941 		bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE;
942 	} else
943 		bp->b_pgbefore = 0;
944 
945 	/* Requested pages. */
946 	for (int j = 0; j < count; j++, i++)
947 		bp->b_pages[i] = m[j];
948 	bp->b_npages += count;
949 
950 	if (rahead) {
951 		vm_pindex_t endpindex, tpindex;
952 		vm_page_t p;
953 
954 		if (!VM_OBJECT_WOWNED(object))
955 			VM_OBJECT_WLOCK(object);
956 		endpindex = m[count - 1]->pindex + rahead + 1;
957 		if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL &&
958 		    p->pindex < endpindex)
959 			endpindex = p->pindex;
960 		if (endpindex > object->size)
961 			endpindex = object->size;
962 
963 		for (tpindex = m[count - 1]->pindex + 1;
964 		    tpindex < endpindex; i++, tpindex++) {
965 			p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
966 			if (p == NULL)
967 				break;
968 			bp->b_pages[i] = p;
969 		}
970 
971 		bp->b_pgafter = i - bp->b_npages;
972 		bp->b_npages = i;
973 	} else
974 		bp->b_pgafter = 0;
975 
976 	if (VM_OBJECT_WOWNED(object))
977 		VM_OBJECT_WUNLOCK(object);
978 
979 	/* Report back actual behind/ahead read. */
980 	if (a_rbehind)
981 		*a_rbehind = bp->b_pgbefore;
982 	if (a_rahead)
983 		*a_rahead = bp->b_pgafter;
984 
985 #ifdef INVARIANTS
986 	KASSERT(bp->b_npages <= nitems(bp->b_pages),
987 	    ("%s: buf %p overflowed", __func__, bp));
988 	for (int j = 1, prev = 0; j < bp->b_npages; j++) {
989 		if (bp->b_pages[j] == bogus_page)
990 			continue;
991 		KASSERT(bp->b_pages[j]->pindex - bp->b_pages[prev]->pindex ==
992 		    j - prev, ("%s: pages array not consecutive, bp %p",
993 		     __func__, bp));
994 		prev = j;
995 	}
996 #endif
997 
998 	/*
999 	 * Recalculate first offset and bytecount with regards to read behind.
1000 	 * Truncate bytecount to vnode real size and round up physical size
1001 	 * for real devices.
1002 	 */
1003 	foff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1004 	bytecount = bp->b_npages << PAGE_SHIFT;
1005 	if ((foff + bytecount) > object->un_pager.vnp.vnp_size)
1006 		bytecount = object->un_pager.vnp.vnp_size - foff;
1007 	secmask = bo->bo_bsize - 1;
1008 	KASSERT(secmask < PAGE_SIZE && secmask > 0,
1009 	    ("%s: sector size %d too large", __func__, secmask + 1));
1010 	bytecount = (bytecount + secmask) & ~secmask;
1011 
1012 	/*
1013 	 * And map the pages to be read into the kva, if the filesystem
1014 	 * requires mapped buffers.
1015 	 */
1016 	if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
1017 	    unmapped_buf_allowed) {
1018 		bp->b_data = unmapped_buf;
1019 		bp->b_offset = 0;
1020 	} else {
1021 		bp->b_data = bp->b_kvabase;
1022 		pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1023 	}
1024 
1025 	/* Build a minimal buffer header. */
1026 	bp->b_iocmd = BIO_READ;
1027 	KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
1028 	KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
1029 	bp->b_rcred = crhold(curthread->td_ucred);
1030 	bp->b_wcred = crhold(curthread->td_ucred);
1031 	pbgetbo(bo, bp);
1032 	bp->b_vp = vp;
1033 	bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount;
1034 	bp->b_iooffset = dbtob(bp->b_blkno);
1035 	KASSERT(IDX_TO_OFF(m[0]->pindex - bp->b_pages[0]->pindex) ==
1036 	    (blkno0 - bp->b_blkno) * DEV_BSIZE +
1037 	    IDX_TO_OFF(m[0]->pindex) % bsize,
1038 	    ("wrong offsets bsize %d m[0] %ju b_pages[0] %ju "
1039 	    "blkno0 %ju b_blkno %ju", bsize,
1040 	    (uintmax_t)m[0]->pindex, (uintmax_t)bp->b_pages[0]->pindex,
1041 	    (uintmax_t)blkno0, (uintmax_t)bp->b_blkno));
1042 
1043 	atomic_add_long(&runningbufspace, bp->b_runningbufspace);
1044 	VM_CNT_INC(v_vnodein);
1045 	VM_CNT_ADD(v_vnodepgsin, bp->b_npages);
1046 
1047 	if (iodone != NULL) { /* async */
1048 		bp->b_pgiodone = iodone;
1049 		bp->b_caller1 = arg;
1050 		bp->b_iodone = vnode_pager_generic_getpages_done_async;
1051 		bp->b_flags |= B_ASYNC;
1052 		BUF_KERNPROC(bp);
1053 		bstrategy(bp);
1054 		return (VM_PAGER_OK);
1055 	} else {
1056 		bp->b_iodone = bdone;
1057 		bstrategy(bp);
1058 		bwait(bp, PVM, "vnread");
1059 		error = vnode_pager_generic_getpages_done(bp);
1060 		for (i = 0; i < bp->b_npages; i++)
1061 			bp->b_pages[i] = NULL;
1062 		bp->b_vp = NULL;
1063 		pbrelbo(bp);
1064 		relpbuf(bp, &vnode_pbuf_freecnt);
1065 		return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
1066 	}
1067 }
1068 
1069 static void
1070 vnode_pager_generic_getpages_done_async(struct buf *bp)
1071 {
1072 	int error;
1073 
1074 	error = vnode_pager_generic_getpages_done(bp);
1075 	/* Run the iodone upon the requested range. */
1076 	bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore,
1077 	    bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error);
1078 	for (int i = 0; i < bp->b_npages; i++)
1079 		bp->b_pages[i] = NULL;
1080 	bp->b_vp = NULL;
1081 	pbrelbo(bp);
1082 	relpbuf(bp, &vnode_async_pbuf_freecnt);
1083 }
1084 
1085 static int
1086 vnode_pager_generic_getpages_done(struct buf *bp)
1087 {
1088 	vm_object_t object;
1089 	off_t tfoff, nextoff;
1090 	int i, error;
1091 
1092 	error = (bp->b_ioflags & BIO_ERROR) != 0 ? EIO : 0;
1093 	object = bp->b_vp->v_object;
1094 
1095 	if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1096 		if (!buf_mapped(bp)) {
1097 			bp->b_data = bp->b_kvabase;
1098 			pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages,
1099 			    bp->b_npages);
1100 		}
1101 		bzero(bp->b_data + bp->b_bcount,
1102 		    PAGE_SIZE * bp->b_npages - bp->b_bcount);
1103 	}
1104 	if (buf_mapped(bp)) {
1105 		pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1106 		bp->b_data = unmapped_buf;
1107 	}
1108 
1109 	VM_OBJECT_WLOCK(object);
1110 	for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1111 	    i < bp->b_npages; i++, tfoff = nextoff) {
1112 		vm_page_t mt;
1113 
1114 		nextoff = tfoff + PAGE_SIZE;
1115 		mt = bp->b_pages[i];
1116 
1117 		if (nextoff <= object->un_pager.vnp.vnp_size) {
1118 			/*
1119 			 * Read filled up entire page.
1120 			 */
1121 			mt->valid = VM_PAGE_BITS_ALL;
1122 			KASSERT(mt->dirty == 0,
1123 			    ("%s: page %p is dirty", __func__, mt));
1124 			KASSERT(!pmap_page_is_mapped(mt),
1125 			    ("%s: page %p is mapped", __func__, mt));
1126 		} else {
1127 			/*
1128 			 * Read did not fill up entire page.
1129 			 *
1130 			 * Currently we do not set the entire page valid,
1131 			 * we just try to clear the piece that we couldn't
1132 			 * read.
1133 			 */
1134 			vm_page_set_valid_range(mt, 0,
1135 			    object->un_pager.vnp.vnp_size - tfoff);
1136 			KASSERT((mt->dirty & vm_page_bits(0,
1137 			    object->un_pager.vnp.vnp_size - tfoff)) == 0,
1138 			    ("%s: page %p is dirty", __func__, mt));
1139 		}
1140 
1141 		if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter)
1142 			vm_page_readahead_finish(mt);
1143 	}
1144 	VM_OBJECT_WUNLOCK(object);
1145 	if (error != 0)
1146 		printf("%s: I/O read error %d\n", __func__, error);
1147 
1148 	return (error);
1149 }
1150 
1151 /*
1152  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
1153  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1154  * vnode_pager_generic_putpages() to implement the previous behaviour.
1155  *
1156  * All other FS's should use the bypass to get to the local media
1157  * backing vp's VOP_PUTPAGES.
1158  */
1159 static void
1160 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1161     int flags, int *rtvals)
1162 {
1163 	int rtval;
1164 	struct vnode *vp;
1165 	int bytes = count * PAGE_SIZE;
1166 
1167 	/*
1168 	 * Force synchronous operation if we are extremely low on memory
1169 	 * to prevent a low-memory deadlock.  VOP operations often need to
1170 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
1171 	 * operation ).  The swapper handles the case by limiting the amount
1172 	 * of asynchronous I/O, but that sort of solution doesn't scale well
1173 	 * for the vnode pager without a lot of work.
1174 	 *
1175 	 * Also, the backing vnode's iodone routine may not wake the pageout
1176 	 * daemon up.  This should be probably be addressed XXX.
1177 	 */
1178 
1179 	if (vm_page_count_min())
1180 		flags |= VM_PAGER_PUT_SYNC;
1181 
1182 	/*
1183 	 * Call device-specific putpages function
1184 	 */
1185 	vp = object->handle;
1186 	VM_OBJECT_WUNLOCK(object);
1187 	rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1188 	KASSERT(rtval != EOPNOTSUPP,
1189 	    ("vnode_pager: stale FS putpages\n"));
1190 	VM_OBJECT_WLOCK(object);
1191 }
1192 
1193 static int
1194 vn_off2bidx(vm_ooffset_t offset)
1195 {
1196 
1197 	return ((offset & PAGE_MASK) / DEV_BSIZE);
1198 }
1199 
1200 static bool
1201 vn_dirty_blk(vm_page_t m, vm_ooffset_t offset)
1202 {
1203 
1204 	KASSERT(IDX_TO_OFF(m->pindex) <= offset &&
1205 	    offset < IDX_TO_OFF(m->pindex + 1),
1206 	    ("page %p pidx %ju offset %ju", m, (uintmax_t)m->pindex,
1207 	    (uintmax_t)offset));
1208 	return ((m->dirty & ((vm_page_bits_t)1 << vn_off2bidx(offset))) != 0);
1209 }
1210 
1211 /*
1212  * This is now called from local media FS's to operate against their
1213  * own vnodes if they fail to implement VOP_PUTPAGES.
1214  *
1215  * This is typically called indirectly via the pageout daemon and
1216  * clustering has already typically occurred, so in general we ask the
1217  * underlying filesystem to write the data out asynchronously rather
1218  * then delayed.
1219  */
1220 int
1221 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1222     int flags, int *rtvals)
1223 {
1224 	vm_object_t object;
1225 	vm_page_t m;
1226 	vm_ooffset_t maxblksz, next_offset, poffset, prev_offset;
1227 	struct uio auio;
1228 	struct iovec aiov;
1229 	off_t prev_resid, wrsz;
1230 	int count, error, i, maxsize, ncount, pgoff, ppscheck;
1231 	bool in_hole;
1232 	static struct timeval lastfail;
1233 	static int curfail;
1234 
1235 	object = vp->v_object;
1236 	count = bytecount / PAGE_SIZE;
1237 
1238 	for (i = 0; i < count; i++)
1239 		rtvals[i] = VM_PAGER_ERROR;
1240 
1241 	if ((int64_t)ma[0]->pindex < 0) {
1242 		printf("vnode_pager_generic_putpages: "
1243 		    "attempt to write meta-data 0x%jx(%lx)\n",
1244 		    (uintmax_t)ma[0]->pindex, (u_long)ma[0]->dirty);
1245 		rtvals[0] = VM_PAGER_BAD;
1246 		return (VM_PAGER_BAD);
1247 	}
1248 
1249 	maxsize = count * PAGE_SIZE;
1250 	ncount = count;
1251 
1252 	poffset = IDX_TO_OFF(ma[0]->pindex);
1253 
1254 	/*
1255 	 * If the page-aligned write is larger then the actual file we
1256 	 * have to invalidate pages occurring beyond the file EOF.  However,
1257 	 * there is an edge case where a file may not be page-aligned where
1258 	 * the last page is partially invalid.  In this case the filesystem
1259 	 * may not properly clear the dirty bits for the entire page (which
1260 	 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1261 	 * With the page locked we are free to fix-up the dirty bits here.
1262 	 *
1263 	 * We do not under any circumstances truncate the valid bits, as
1264 	 * this will screw up bogus page replacement.
1265 	 */
1266 	VM_OBJECT_RLOCK(object);
1267 	if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1268 		if (!VM_OBJECT_TRYUPGRADE(object)) {
1269 			VM_OBJECT_RUNLOCK(object);
1270 			VM_OBJECT_WLOCK(object);
1271 			if (maxsize + poffset <= object->un_pager.vnp.vnp_size)
1272 				goto downgrade;
1273 		}
1274 		if (object->un_pager.vnp.vnp_size > poffset) {
1275 			maxsize = object->un_pager.vnp.vnp_size - poffset;
1276 			ncount = btoc(maxsize);
1277 			if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1278 				pgoff = roundup2(pgoff, DEV_BSIZE);
1279 
1280 				/*
1281 				 * If the object is locked and the following
1282 				 * conditions hold, then the page's dirty
1283 				 * field cannot be concurrently changed by a
1284 				 * pmap operation.
1285 				 */
1286 				m = ma[ncount - 1];
1287 				vm_page_assert_sbusied(m);
1288 				KASSERT(!pmap_page_is_write_mapped(m),
1289 		("vnode_pager_generic_putpages: page %p is not read-only", m));
1290 				MPASS(m->dirty != 0);
1291 				vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1292 				    pgoff);
1293 			}
1294 		} else {
1295 			maxsize = 0;
1296 			ncount = 0;
1297 		}
1298 		for (i = ncount; i < count; i++)
1299 			rtvals[i] = VM_PAGER_BAD;
1300 downgrade:
1301 		VM_OBJECT_LOCK_DOWNGRADE(object);
1302 	}
1303 
1304 	auio.uio_iov = &aiov;
1305 	auio.uio_segflg = UIO_NOCOPY;
1306 	auio.uio_rw = UIO_WRITE;
1307 	auio.uio_td = NULL;
1308 	maxblksz = roundup2(poffset + maxsize, DEV_BSIZE);
1309 
1310 	for (prev_offset = poffset; prev_offset < maxblksz;) {
1311 		/* Skip clean blocks. */
1312 		for (in_hole = true; in_hole && prev_offset < maxblksz;) {
1313 			m = ma[OFF_TO_IDX(prev_offset - poffset)];
1314 			for (i = vn_off2bidx(prev_offset);
1315 			    i < sizeof(vm_page_bits_t) * NBBY &&
1316 			    prev_offset < maxblksz; i++) {
1317 				if (vn_dirty_blk(m, prev_offset)) {
1318 					in_hole = false;
1319 					break;
1320 				}
1321 				prev_offset += DEV_BSIZE;
1322 			}
1323 		}
1324 		if (in_hole)
1325 			goto write_done;
1326 
1327 		/* Find longest run of dirty blocks. */
1328 		for (next_offset = prev_offset; next_offset < maxblksz;) {
1329 			m = ma[OFF_TO_IDX(next_offset - poffset)];
1330 			for (i = vn_off2bidx(next_offset);
1331 			    i < sizeof(vm_page_bits_t) * NBBY &&
1332 			    next_offset < maxblksz; i++) {
1333 				if (!vn_dirty_blk(m, next_offset))
1334 					goto start_write;
1335 				next_offset += DEV_BSIZE;
1336 			}
1337 		}
1338 start_write:
1339 		if (next_offset > poffset + maxsize)
1340 			next_offset = poffset + maxsize;
1341 
1342 		/*
1343 		 * Getting here requires finding a dirty block in the
1344 		 * 'skip clean blocks' loop.
1345 		 */
1346 		MPASS(prev_offset < next_offset);
1347 
1348 		VM_OBJECT_RUNLOCK(object);
1349 		aiov.iov_base = NULL;
1350 		auio.uio_iovcnt = 1;
1351 		auio.uio_offset = prev_offset;
1352 		prev_resid = auio.uio_resid = aiov.iov_len = next_offset -
1353 		    prev_offset;
1354 		error = VOP_WRITE(vp, &auio,
1355 		    vnode_pager_putpages_ioflags(flags), curthread->td_ucred);
1356 
1357 		wrsz = prev_resid - auio.uio_resid;
1358 		if (wrsz == 0) {
1359 			if (ppsratecheck(&lastfail, &curfail, 1) != 0) {
1360 				vn_printf(vp, "vnode_pager_putpages: "
1361 				    "zero-length write at %ju resid %zd\n",
1362 				    auio.uio_offset, auio.uio_resid);
1363 			}
1364 			VM_OBJECT_RLOCK(object);
1365 			break;
1366 		}
1367 
1368 		/* Adjust the starting offset for next iteration. */
1369 		prev_offset += wrsz;
1370 		MPASS(auio.uio_offset == prev_offset);
1371 
1372 		ppscheck = 0;
1373 		if (error != 0 && (ppscheck = ppsratecheck(&lastfail,
1374 		    &curfail, 1)) != 0)
1375 			vn_printf(vp, "vnode_pager_putpages: I/O error %d\n",
1376 			    error);
1377 		if (auio.uio_resid != 0 && (ppscheck != 0 ||
1378 		    ppsratecheck(&lastfail, &curfail, 1) != 0))
1379 			vn_printf(vp, "vnode_pager_putpages: residual I/O %zd "
1380 			    "at %ju\n", auio.uio_resid,
1381 			    (uintmax_t)ma[0]->pindex);
1382 		VM_OBJECT_RLOCK(object);
1383 		if (error != 0 || auio.uio_resid != 0)
1384 			break;
1385 	}
1386 write_done:
1387 	/* Mark completely processed pages. */
1388 	for (i = 0; i < OFF_TO_IDX(prev_offset - poffset); i++)
1389 		rtvals[i] = VM_PAGER_OK;
1390 	/* Mark partial EOF page. */
1391 	if (prev_offset == poffset + maxsize && (prev_offset & PAGE_MASK) != 0)
1392 		rtvals[i++] = VM_PAGER_OK;
1393 	/* Unwritten pages in range, free bonus if the page is clean. */
1394 	for (; i < ncount; i++)
1395 		rtvals[i] = ma[i]->dirty == 0 ? VM_PAGER_OK : VM_PAGER_ERROR;
1396 	VM_OBJECT_RUNLOCK(object);
1397 	VM_CNT_ADD(v_vnodepgsout, i);
1398 	VM_CNT_INC(v_vnodeout);
1399 	return (rtvals[0]);
1400 }
1401 
1402 int
1403 vnode_pager_putpages_ioflags(int pager_flags)
1404 {
1405 	int ioflags;
1406 
1407 	/*
1408 	 * Pageouts are already clustered, use IO_ASYNC to force a
1409 	 * bawrite() rather then a bdwrite() to prevent paging I/O
1410 	 * from saturating the buffer cache.  Dummy-up the sequential
1411 	 * heuristic to cause large ranges to cluster.  If neither
1412 	 * IO_SYNC or IO_ASYNC is set, the system decides how to
1413 	 * cluster.
1414 	 */
1415 	ioflags = IO_VMIO;
1416 	if ((pager_flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) != 0)
1417 		ioflags |= IO_SYNC;
1418 	else if ((pager_flags & VM_PAGER_CLUSTER_OK) == 0)
1419 		ioflags |= IO_ASYNC;
1420 	ioflags |= (pager_flags & VM_PAGER_PUT_INVAL) != 0 ? IO_INVAL: 0;
1421 	ioflags |= (pager_flags & VM_PAGER_PUT_NOREUSE) != 0 ? IO_NOREUSE : 0;
1422 	ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1423 	return (ioflags);
1424 }
1425 
1426 /*
1427  * vnode_pager_undirty_pages().
1428  *
1429  * A helper to mark pages as clean after pageout that was possibly
1430  * done with a short write.  The lpos argument specifies the page run
1431  * length in bytes, and the written argument specifies how many bytes
1432  * were actually written.  eof is the offset past the last valid byte
1433  * in the vnode using the absolute file position of the first byte in
1434  * the run as the base from which it is computed.
1435  */
1436 void
1437 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written, off_t eof,
1438     int lpos)
1439 {
1440 	vm_object_t obj;
1441 	int i, pos, pos_devb;
1442 
1443 	if (written == 0 && eof >= lpos)
1444 		return;
1445 	obj = ma[0]->object;
1446 	VM_OBJECT_WLOCK(obj);
1447 	for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1448 		if (pos < trunc_page(written)) {
1449 			rtvals[i] = VM_PAGER_OK;
1450 			vm_page_undirty(ma[i]);
1451 		} else {
1452 			/* Partially written page. */
1453 			rtvals[i] = VM_PAGER_AGAIN;
1454 			vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1455 		}
1456 	}
1457 	if (eof >= lpos) /* avoid truncation */
1458 		goto done;
1459 	for (pos = eof, i = OFF_TO_IDX(trunc_page(pos)); pos < lpos; i++) {
1460 		if (pos != trunc_page(pos)) {
1461 			/*
1462 			 * The page contains the last valid byte in
1463 			 * the vnode, mark the rest of the page as
1464 			 * clean, potentially making the whole page
1465 			 * clean.
1466 			 */
1467 			pos_devb = roundup2(pos & PAGE_MASK, DEV_BSIZE);
1468 			vm_page_clear_dirty(ma[i], pos_devb, PAGE_SIZE -
1469 			    pos_devb);
1470 
1471 			/*
1472 			 * If the page was cleaned, report the pageout
1473 			 * on it as successful.  msync() no longer
1474 			 * needs to write out the page, endlessly
1475 			 * creating write requests and dirty buffers.
1476 			 */
1477 			if (ma[i]->dirty == 0)
1478 				rtvals[i] = VM_PAGER_OK;
1479 
1480 			pos = round_page(pos);
1481 		} else {
1482 			/* vm_pageout_flush() clears dirty */
1483 			rtvals[i] = VM_PAGER_BAD;
1484 			pos += PAGE_SIZE;
1485 		}
1486 	}
1487 done:
1488 	VM_OBJECT_WUNLOCK(obj);
1489 }
1490 
1491 void
1492 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1493     vm_offset_t end)
1494 {
1495 	struct vnode *vp;
1496 	vm_ooffset_t old_wm;
1497 
1498 	VM_OBJECT_WLOCK(object);
1499 	if (object->type != OBJT_VNODE) {
1500 		VM_OBJECT_WUNLOCK(object);
1501 		return;
1502 	}
1503 	old_wm = object->un_pager.vnp.writemappings;
1504 	object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1505 	vp = object->handle;
1506 	if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1507 		ASSERT_VOP_ELOCKED(vp, "v_writecount inc");
1508 		VOP_ADD_WRITECOUNT(vp, 1);
1509 		CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1510 		    __func__, vp, vp->v_writecount);
1511 	} else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1512 		ASSERT_VOP_ELOCKED(vp, "v_writecount dec");
1513 		VOP_ADD_WRITECOUNT(vp, -1);
1514 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1515 		    __func__, vp, vp->v_writecount);
1516 	}
1517 	VM_OBJECT_WUNLOCK(object);
1518 }
1519 
1520 void
1521 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1522     vm_offset_t end)
1523 {
1524 	struct vnode *vp;
1525 	struct mount *mp;
1526 	vm_offset_t inc;
1527 
1528 	VM_OBJECT_WLOCK(object);
1529 
1530 	/*
1531 	 * First, recheck the object type to account for the race when
1532 	 * the vnode is reclaimed.
1533 	 */
1534 	if (object->type != OBJT_VNODE) {
1535 		VM_OBJECT_WUNLOCK(object);
1536 		return;
1537 	}
1538 
1539 	/*
1540 	 * Optimize for the case when writemappings is not going to
1541 	 * zero.
1542 	 */
1543 	inc = end - start;
1544 	if (object->un_pager.vnp.writemappings != inc) {
1545 		object->un_pager.vnp.writemappings -= inc;
1546 		VM_OBJECT_WUNLOCK(object);
1547 		return;
1548 	}
1549 
1550 	vp = object->handle;
1551 	vhold(vp);
1552 	VM_OBJECT_WUNLOCK(object);
1553 	mp = NULL;
1554 	vn_start_write(vp, &mp, V_WAIT);
1555 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1556 
1557 	/*
1558 	 * Decrement the object's writemappings, by swapping the start
1559 	 * and end arguments for vnode_pager_update_writecount().  If
1560 	 * there was not a race with vnode reclaimation, then the
1561 	 * vnode's v_writecount is decremented.
1562 	 */
1563 	vnode_pager_update_writecount(object, end, start);
1564 	VOP_UNLOCK(vp, 0);
1565 	vdrop(vp);
1566 	if (mp != NULL)
1567 		vn_finished_write(mp);
1568 }
1569