xref: /freebsd/sys/vm/vnode_pager.c (revision d0b2dbfa0ecf2bbc9709efc5e20baf8e4b44bbbf)
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 #include "opt_vm.h"
57 
58 #include <sys/param.h>
59 #include <sys/kernel.h>
60 #include <sys/systm.h>
61 #include <sys/sysctl.h>
62 #include <sys/proc.h>
63 #include <sys/vnode.h>
64 #include <sys/mount.h>
65 #include <sys/bio.h>
66 #include <sys/buf.h>
67 #include <sys/vmmeter.h>
68 #include <sys/ktr.h>
69 #include <sys/limits.h>
70 #include <sys/conf.h>
71 #include <sys/refcount.h>
72 #include <sys/rwlock.h>
73 #include <sys/sf_buf.h>
74 #include <sys/domainset.h>
75 #include <sys/user.h>
76 
77 #include <machine/atomic.h>
78 
79 #include <vm/vm.h>
80 #include <vm/vm_param.h>
81 #include <vm/vm_object.h>
82 #include <vm/vm_page.h>
83 #include <vm/vm_pager.h>
84 #include <vm/vm_map.h>
85 #include <vm/vnode_pager.h>
86 #include <vm/vm_extern.h>
87 #include <vm/uma.h>
88 
89 static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
90     daddr_t *rtaddress, int *run);
91 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
92 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
93 static void vnode_pager_dealloc(vm_object_t);
94 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *);
95 static int vnode_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
96     int *, vop_getpages_iodone_t, void *);
97 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *);
98 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
99 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t,
100     vm_ooffset_t, struct ucred *cred);
101 static int vnode_pager_generic_getpages_done(struct buf *);
102 static void vnode_pager_generic_getpages_done_async(struct buf *);
103 static void vnode_pager_update_writecount(vm_object_t, vm_offset_t,
104     vm_offset_t);
105 static void vnode_pager_release_writecount(vm_object_t, vm_offset_t,
106     vm_offset_t);
107 static void vnode_pager_getvp(vm_object_t, struct vnode **, bool *);
108 
109 const struct pagerops vnodepagerops = {
110 	.pgo_kvme_type = KVME_TYPE_VNODE,
111 	.pgo_alloc =	vnode_pager_alloc,
112 	.pgo_dealloc =	vnode_pager_dealloc,
113 	.pgo_getpages =	vnode_pager_getpages,
114 	.pgo_getpages_async = vnode_pager_getpages_async,
115 	.pgo_putpages =	vnode_pager_putpages,
116 	.pgo_haspage =	vnode_pager_haspage,
117 	.pgo_update_writecount = vnode_pager_update_writecount,
118 	.pgo_release_writecount = vnode_pager_release_writecount,
119 	.pgo_set_writeable_dirty = vm_object_set_writeable_dirty_,
120 	.pgo_mightbedirty = vm_object_mightbedirty_,
121 	.pgo_getvp = vnode_pager_getvp,
122 };
123 
124 static struct domainset *vnode_domainset = NULL;
125 
126 SYSCTL_PROC(_debug, OID_AUTO, vnode_domainset,
127     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_RW, &vnode_domainset, 0,
128     sysctl_handle_domainset, "A", "Default vnode NUMA policy");
129 
130 static int nvnpbufs;
131 SYSCTL_INT(_vm, OID_AUTO, vnode_pbufs, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
132     &nvnpbufs, 0, "number of physical buffers allocated for vnode pager");
133 
134 static uma_zone_t vnode_pbuf_zone;
135 
136 static void
137 vnode_pager_init(void *dummy)
138 {
139 
140 #ifdef __LP64__
141 	nvnpbufs = nswbuf * 2;
142 #else
143 	nvnpbufs = nswbuf / 2;
144 #endif
145 	TUNABLE_INT_FETCH("vm.vnode_pbufs", &nvnpbufs);
146 	vnode_pbuf_zone = pbuf_zsecond_create("vnpbuf", nvnpbufs);
147 }
148 SYSINIT(vnode_pager, SI_SUB_CPU, SI_ORDER_ANY, vnode_pager_init, NULL);
149 
150 /* Create the VM system backing object for this vnode */
151 int
152 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
153 {
154 	vm_object_t object;
155 	vm_ooffset_t size = isize;
156 	bool last;
157 
158 	if (!vn_isdisk(vp) && vn_canvmio(vp) == FALSE)
159 		return (0);
160 
161 	object = vp->v_object;
162 	if (object != NULL)
163 		return (0);
164 
165 	if (size == 0) {
166 		if (vn_isdisk(vp)) {
167 			size = IDX_TO_OFF(INT_MAX);
168 		} else {
169 			if (vn_getsize_locked(vp, &size, td->td_ucred) != 0)
170 				return (0);
171 		}
172 	}
173 
174 	object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred);
175 	/*
176 	 * Dereference the reference we just created.  This assumes
177 	 * that the object is associated with the vp.  We still have
178 	 * to serialize with vnode_pager_dealloc() for the last
179 	 * potential reference.
180 	 */
181 	VM_OBJECT_RLOCK(object);
182 	last = refcount_release(&object->ref_count);
183 	VM_OBJECT_RUNLOCK(object);
184 	if (last)
185 		vrele(vp);
186 
187 	KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
188 
189 	return (0);
190 }
191 
192 void
193 vnode_destroy_vobject(struct vnode *vp)
194 {
195 	struct vm_object *obj;
196 
197 	obj = vp->v_object;
198 	if (obj == NULL || obj->handle != vp)
199 		return;
200 	ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
201 	VM_OBJECT_WLOCK(obj);
202 	MPASS(obj->type == OBJT_VNODE);
203 	umtx_shm_object_terminated(obj);
204 	if (obj->ref_count == 0) {
205 		KASSERT((obj->flags & OBJ_DEAD) == 0,
206 		   ("vnode_destroy_vobject: Terminating dead object"));
207 		vm_object_set_flag(obj, OBJ_DEAD);
208 
209 		/*
210 		 * Clean pages and flush buffers.
211 		 */
212 		vm_object_page_clean(obj, 0, 0, OBJPC_SYNC);
213 		VM_OBJECT_WUNLOCK(obj);
214 
215 		vinvalbuf(vp, V_SAVE, 0, 0);
216 
217 		BO_LOCK(&vp->v_bufobj);
218 		vp->v_bufobj.bo_flag |= BO_DEAD;
219 		BO_UNLOCK(&vp->v_bufobj);
220 
221 		VM_OBJECT_WLOCK(obj);
222 		vm_object_terminate(obj);
223 	} else {
224 		/*
225 		 * Woe to the process that tries to page now :-).
226 		 */
227 		vm_pager_deallocate(obj);
228 		VM_OBJECT_WUNLOCK(obj);
229 	}
230 	KASSERT(vp->v_object == NULL, ("vp %p obj %p", vp, vp->v_object));
231 }
232 
233 /*
234  * Allocate (or lookup) pager for a vnode.
235  * Handle is a vnode pointer.
236  */
237 vm_object_t
238 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
239     vm_ooffset_t offset, struct ucred *cred)
240 {
241 	vm_object_t object;
242 	struct vnode *vp;
243 
244 	/*
245 	 * Pageout to vnode, no can do yet.
246 	 */
247 	if (handle == NULL)
248 		return (NULL);
249 
250 	vp = (struct vnode *)handle;
251 	ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc");
252 	VNPASS(vp->v_usecount > 0, vp);
253 retry:
254 	object = vp->v_object;
255 
256 	if (object == NULL) {
257 		/*
258 		 * Add an object of the appropriate size
259 		 */
260 		object = vm_object_allocate(OBJT_VNODE,
261 		    OFF_TO_IDX(round_page(size)));
262 
263 		object->un_pager.vnp.vnp_size = size;
264 		object->un_pager.vnp.writemappings = 0;
265 		object->domain.dr_policy = vnode_domainset;
266 		object->handle = handle;
267 		if ((vp->v_vflag & VV_VMSIZEVNLOCK) != 0) {
268 			VM_OBJECT_WLOCK(object);
269 			vm_object_set_flag(object, OBJ_SIZEVNLOCK);
270 			VM_OBJECT_WUNLOCK(object);
271 		}
272 		VI_LOCK(vp);
273 		if (vp->v_object != NULL) {
274 			/*
275 			 * Object has been created while we were allocating.
276 			 */
277 			VI_UNLOCK(vp);
278 			VM_OBJECT_WLOCK(object);
279 			KASSERT(object->ref_count == 1,
280 			    ("leaked ref %p %d", object, object->ref_count));
281 			object->type = OBJT_DEAD;
282 			refcount_init(&object->ref_count, 0);
283 			VM_OBJECT_WUNLOCK(object);
284 			vm_object_destroy(object);
285 			goto retry;
286 		}
287 		vp->v_object = object;
288 		VI_UNLOCK(vp);
289 		vrefact(vp);
290 	} else {
291 		vm_object_reference(object);
292 #if VM_NRESERVLEVEL > 0
293 		if ((object->flags & OBJ_COLORED) == 0) {
294 			VM_OBJECT_WLOCK(object);
295 			vm_object_color(object, 0);
296 			VM_OBJECT_WUNLOCK(object);
297 		}
298 #endif
299 	}
300 	return (object);
301 }
302 
303 /*
304  *	The object must be locked.
305  */
306 static void
307 vnode_pager_dealloc(vm_object_t object)
308 {
309 	struct vnode *vp;
310 	int refs;
311 
312 	vp = object->handle;
313 	if (vp == NULL)
314 		panic("vnode_pager_dealloc: pager already dealloced");
315 
316 	VM_OBJECT_ASSERT_WLOCKED(object);
317 	vm_object_pip_wait(object, "vnpdea");
318 	refs = object->ref_count;
319 
320 	object->handle = NULL;
321 	object->type = OBJT_DEAD;
322 	ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
323 	if (object->un_pager.vnp.writemappings > 0) {
324 		object->un_pager.vnp.writemappings = 0;
325 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
326 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
327 		    __func__, vp, vp->v_writecount);
328 	}
329 	vp->v_object = NULL;
330 	VI_LOCK(vp);
331 
332 	/*
333 	 * vm_map_entry_set_vnode_text() cannot reach this vnode by
334 	 * following object->handle.  Clear all text references now.
335 	 * This also clears the transient references from
336 	 * kern_execve(), which is fine because dead_vnodeops uses nop
337 	 * for VOP_UNSET_TEXT().
338 	 */
339 	if (vp->v_writecount < 0)
340 		vp->v_writecount = 0;
341 	VI_UNLOCK(vp);
342 	VM_OBJECT_WUNLOCK(object);
343 	if (refs > 0)
344 		vunref(vp);
345 	VM_OBJECT_WLOCK(object);
346 }
347 
348 static boolean_t
349 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
350     int *after)
351 {
352 	struct vnode *vp = object->handle;
353 	daddr_t bn;
354 	uintptr_t lockstate;
355 	int err;
356 	daddr_t reqblock;
357 	int poff;
358 	int bsize;
359 	int pagesperblock, blocksperpage;
360 
361 	VM_OBJECT_ASSERT_LOCKED(object);
362 	/*
363 	 * If no vp or vp is doomed or marked transparent to VM, we do not
364 	 * have the page.
365 	 */
366 	if (vp == NULL || VN_IS_DOOMED(vp))
367 		return FALSE;
368 	/*
369 	 * If the offset is beyond end of file we do
370 	 * not have the page.
371 	 */
372 	if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
373 		return FALSE;
374 
375 	bsize = vp->v_mount->mnt_stat.f_iosize;
376 	pagesperblock = bsize / PAGE_SIZE;
377 	blocksperpage = 0;
378 	if (pagesperblock > 0) {
379 		reqblock = pindex / pagesperblock;
380 	} else {
381 		blocksperpage = (PAGE_SIZE / bsize);
382 		reqblock = pindex * blocksperpage;
383 	}
384 	lockstate = VM_OBJECT_DROP(object);
385 	err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
386 	VM_OBJECT_PICKUP(object, lockstate);
387 	if (err)
388 		return TRUE;
389 	if (bn == -1)
390 		return FALSE;
391 	if (pagesperblock > 0) {
392 		poff = pindex - (reqblock * pagesperblock);
393 		if (before) {
394 			*before *= pagesperblock;
395 			*before += poff;
396 		}
397 		if (after) {
398 			/*
399 			 * The BMAP vop can report a partial block in the
400 			 * 'after', but must not report blocks after EOF.
401 			 * Assert the latter, and truncate 'after' in case
402 			 * of the former.
403 			 */
404 			KASSERT((reqblock + *after) * pagesperblock <
405 			    roundup2(object->size, pagesperblock),
406 			    ("%s: reqblock %jd after %d size %ju", __func__,
407 			    (intmax_t )reqblock, *after,
408 			    (uintmax_t )object->size));
409 			*after *= pagesperblock;
410 			*after += pagesperblock - (poff + 1);
411 			if (pindex + *after >= object->size)
412 				*after = object->size - 1 - pindex;
413 		}
414 	} else {
415 		if (before) {
416 			*before /= blocksperpage;
417 		}
418 
419 		if (after) {
420 			*after /= blocksperpage;
421 		}
422 	}
423 	return TRUE;
424 }
425 
426 /*
427  * Internal routine clearing partial-page content
428  */
429 static void
430 vnode_pager_subpage_purge(struct vm_page *m, int base, int end)
431 {
432 	int size;
433 
434 	KASSERT(end > base && end <= PAGE_SIZE,
435 	    ("%s: start %d end %d", __func__, base, end));
436 	size = end - base;
437 
438 	/*
439 	 * Clear out partial-page garbage in case
440 	 * the page has been mapped.
441 	 */
442 	pmap_zero_page_area(m, base, size);
443 
444 	/*
445 	 * Update the valid bits to reflect the blocks
446 	 * that have been zeroed.  Some of these valid
447 	 * bits may have already been set.
448 	 */
449 	vm_page_set_valid_range(m, base, size);
450 
451 	/*
452 	 * Round up "base" to the next block boundary so
453 	 * that the dirty bit for a partially zeroed
454 	 * block is not cleared.
455 	 */
456 	base = roundup2(base, DEV_BSIZE);
457 	end = rounddown2(end, DEV_BSIZE);
458 
459 	if (end > base) {
460 		/*
461 		 * Clear out partial-page dirty bits.
462 		 *
463 		 * note that we do not clear out the
464 		 * valid bits.  This would prevent
465 		 * bogus_page replacement from working
466 		 * properly.
467 		 */
468 		vm_page_clear_dirty(m, base, end - base);
469 	}
470 
471 }
472 
473 /*
474  * Lets the VM system know about a change in size for a file.
475  * We adjust our own internal size and flush any cached pages in
476  * the associated object that are affected by the size change.
477  *
478  * Note: this routine may be invoked as a result of a pager put
479  * operation (possibly at object termination time), so we must be careful.
480  */
481 void
482 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
483 {
484 	vm_object_t object;
485 	vm_page_t m;
486 	vm_pindex_t nobjsize;
487 
488 	if ((object = vp->v_object) == NULL)
489 		return;
490 #ifdef DEBUG_VFS_LOCKS
491 	{
492 		struct mount *mp;
493 
494 		mp = vp->v_mount;
495 		if (mp != NULL && (mp->mnt_kern_flag & MNTK_VMSETSIZE_BUG) == 0)
496 			assert_vop_elocked(vp,
497 			    "vnode_pager_setsize and not locked vnode");
498 	}
499 #endif
500 	VM_OBJECT_WLOCK(object);
501 	if (object->type == OBJT_DEAD) {
502 		VM_OBJECT_WUNLOCK(object);
503 		return;
504 	}
505 	KASSERT(object->type == OBJT_VNODE,
506 	    ("not vnode-backed object %p", object));
507 	if (nsize == object->un_pager.vnp.vnp_size) {
508 		/*
509 		 * Hasn't changed size
510 		 */
511 		VM_OBJECT_WUNLOCK(object);
512 		return;
513 	}
514 	nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
515 	if (nsize < object->un_pager.vnp.vnp_size) {
516 		/*
517 		 * File has shrunk. Toss any cached pages beyond the new EOF.
518 		 */
519 		if (nobjsize < object->size)
520 			vm_object_page_remove(object, nobjsize, object->size,
521 			    0);
522 		/*
523 		 * this gets rid of garbage at the end of a page that is now
524 		 * only partially backed by the vnode.
525 		 *
526 		 * XXX for some reason (I don't know yet), if we take a
527 		 * completely invalid page and mark it partially valid
528 		 * it can screw up NFS reads, so we don't allow the case.
529 		 */
530 		if (!(nsize & PAGE_MASK))
531 			goto out;
532 		m = vm_page_grab(object, OFF_TO_IDX(nsize), VM_ALLOC_NOCREAT);
533 		if (m == NULL)
534 			goto out;
535 		if (!vm_page_none_valid(m))
536 			vnode_pager_subpage_purge(m, (int)nsize & PAGE_MASK,
537 			    PAGE_SIZE);
538 		vm_page_xunbusy(m);
539 	}
540 out:
541 #if defined(__powerpc__) && !defined(__powerpc64__)
542 	object->un_pager.vnp.vnp_size = nsize;
543 #else
544 	atomic_store_64(&object->un_pager.vnp.vnp_size, nsize);
545 #endif
546 	object->size = nobjsize;
547 	VM_OBJECT_WUNLOCK(object);
548 }
549 
550 /*
551  * Lets the VM system know about the purged range for a file. We toss away any
552  * cached pages in the associated object that are affected by the purge
553  * operation. Partial-page area not aligned to page boundaries will be zeroed
554  * and the dirty blocks in DEV_BSIZE unit within a page will not be flushed.
555  */
556 void
557 vnode_pager_purge_range(struct vnode *vp, vm_ooffset_t start, vm_ooffset_t end)
558 {
559 	struct vm_page *m;
560 	struct vm_object *object;
561 	vm_pindex_t pi, pistart, piend;
562 	bool same_page;
563 	int base, pend;
564 
565 	ASSERT_VOP_LOCKED(vp, "vnode_pager_purge_range");
566 
567 	object = vp->v_object;
568 	pi = start + PAGE_MASK < start ? OBJ_MAX_SIZE :
569 	    OFF_TO_IDX(start + PAGE_MASK);
570 	pistart = OFF_TO_IDX(start);
571 	piend = end == 0 ? OBJ_MAX_SIZE : OFF_TO_IDX(end);
572 	same_page = pistart == piend;
573 	if ((end != 0 && end <= start) || object == NULL)
574 		return;
575 
576 	VM_OBJECT_WLOCK(object);
577 
578 	if (pi < piend)
579 		vm_object_page_remove(object, pi, piend, 0);
580 
581 	if ((start & PAGE_MASK) != 0) {
582 		base = (int)start & PAGE_MASK;
583 		pend = same_page ? (int)end & PAGE_MASK : PAGE_SIZE;
584 		m = vm_page_grab(object, pistart, VM_ALLOC_NOCREAT);
585 		if (m != NULL) {
586 			if (!vm_page_none_valid(m))
587 				vnode_pager_subpage_purge(m, base, pend);
588 			vm_page_xunbusy(m);
589 		}
590 		if (same_page)
591 			goto out;
592 	}
593 	if ((end & PAGE_MASK) != 0) {
594 		base = same_page ? (int)start & PAGE_MASK : 0 ;
595 		pend = (int)end & PAGE_MASK;
596 		m = vm_page_grab(object, piend, VM_ALLOC_NOCREAT);
597 		if (m != NULL) {
598 			if (!vm_page_none_valid(m))
599 				vnode_pager_subpage_purge(m, base, pend);
600 			vm_page_xunbusy(m);
601 		}
602 	}
603 out:
604 	VM_OBJECT_WUNLOCK(object);
605 }
606 
607 /*
608  * calculate the linear (byte) disk address of specified virtual
609  * file address
610  */
611 static int
612 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
613     int *run)
614 {
615 	int bsize;
616 	int err;
617 	daddr_t vblock;
618 	daddr_t voffset;
619 
620 	if (VN_IS_DOOMED(vp))
621 		return -1;
622 
623 	bsize = vp->v_mount->mnt_stat.f_iosize;
624 	vblock = address / bsize;
625 	voffset = address % bsize;
626 
627 	err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
628 	if (err == 0) {
629 		if (*rtaddress != -1)
630 			*rtaddress += voffset / DEV_BSIZE;
631 		if (run) {
632 			*run += 1;
633 			*run *= bsize / PAGE_SIZE;
634 			*run -= voffset / PAGE_SIZE;
635 		}
636 	}
637 
638 	return (err);
639 }
640 
641 static void
642 vnode_pager_input_bdone(struct buf *bp)
643 {
644 	runningbufwakeup(bp);
645 	bdone(bp);
646 }
647 
648 /*
649  * small block filesystem vnode pager input
650  */
651 static int
652 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m)
653 {
654 	struct vnode *vp;
655 	struct bufobj *bo;
656 	struct buf *bp;
657 	struct sf_buf *sf;
658 	daddr_t fileaddr;
659 	vm_offset_t bsize;
660 	vm_page_bits_t bits;
661 	int error, i;
662 
663 	error = 0;
664 	vp = object->handle;
665 	if (VN_IS_DOOMED(vp))
666 		return VM_PAGER_BAD;
667 
668 	bsize = vp->v_mount->mnt_stat.f_iosize;
669 
670 	VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
671 
672 	sf = sf_buf_alloc(m, 0);
673 
674 	for (i = 0; i < PAGE_SIZE / bsize; i++) {
675 		vm_ooffset_t address;
676 
677 		bits = vm_page_bits(i * bsize, bsize);
678 		if (m->valid & bits)
679 			continue;
680 
681 		address = IDX_TO_OFF(m->pindex) + i * bsize;
682 		if (address >= object->un_pager.vnp.vnp_size) {
683 			fileaddr = -1;
684 		} else {
685 			error = vnode_pager_addr(vp, address, &fileaddr, NULL);
686 			if (error)
687 				break;
688 		}
689 		if (fileaddr != -1) {
690 			bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK);
691 
692 			/* build a minimal buffer header */
693 			bp->b_iocmd = BIO_READ;
694 			bp->b_iodone = vnode_pager_input_bdone;
695 			KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
696 			KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
697 			bp->b_rcred = crhold(curthread->td_ucred);
698 			bp->b_wcred = crhold(curthread->td_ucred);
699 			bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
700 			bp->b_blkno = fileaddr;
701 			pbgetbo(bo, bp);
702 			bp->b_vp = vp;
703 			bp->b_bcount = bsize;
704 			bp->b_bufsize = bsize;
705 			bp->b_runningbufspace = bp->b_bufsize;
706 			atomic_add_long(&runningbufspace, bp->b_runningbufspace);
707 
708 			/* do the input */
709 			bp->b_iooffset = dbtob(bp->b_blkno);
710 			bstrategy(bp);
711 
712 			bwait(bp, PVM, "vnsrd");
713 
714 			if ((bp->b_ioflags & BIO_ERROR) != 0) {
715 				KASSERT(bp->b_error != 0,
716 				    ("%s: buf error but b_error == 0\n", __func__));
717 				error = bp->b_error;
718 			}
719 
720 			/*
721 			 * free the buffer header back to the swap buffer pool
722 			 */
723 			bp->b_vp = NULL;
724 			pbrelbo(bp);
725 			uma_zfree(vnode_pbuf_zone, bp);
726 			if (error)
727 				break;
728 		} else
729 			bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
730 		KASSERT((m->dirty & bits) == 0,
731 		    ("vnode_pager_input_smlfs: page %p is dirty", m));
732 		vm_page_bits_set(m, &m->valid, bits);
733 	}
734 	sf_buf_free(sf);
735 	if (error) {
736 		return VM_PAGER_ERROR;
737 	}
738 	return VM_PAGER_OK;
739 }
740 
741 /*
742  * old style vnode pager input routine
743  */
744 static int
745 vnode_pager_input_old(vm_object_t object, vm_page_t m)
746 {
747 	struct uio auio;
748 	struct iovec aiov;
749 	int error;
750 	int size;
751 	struct sf_buf *sf;
752 	struct vnode *vp;
753 
754 	VM_OBJECT_ASSERT_WLOCKED(object);
755 	error = 0;
756 
757 	/*
758 	 * Return failure if beyond current EOF
759 	 */
760 	if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
761 		return VM_PAGER_BAD;
762 	} else {
763 		size = PAGE_SIZE;
764 		if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
765 			size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
766 		vp = object->handle;
767 		VM_OBJECT_WUNLOCK(object);
768 
769 		/*
770 		 * Allocate a kernel virtual address and initialize so that
771 		 * we can use VOP_READ/WRITE routines.
772 		 */
773 		sf = sf_buf_alloc(m, 0);
774 
775 		aiov.iov_base = (caddr_t)sf_buf_kva(sf);
776 		aiov.iov_len = size;
777 		auio.uio_iov = &aiov;
778 		auio.uio_iovcnt = 1;
779 		auio.uio_offset = IDX_TO_OFF(m->pindex);
780 		auio.uio_segflg = UIO_SYSSPACE;
781 		auio.uio_rw = UIO_READ;
782 		auio.uio_resid = size;
783 		auio.uio_td = curthread;
784 
785 		error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
786 		if (!error) {
787 			int count = size - auio.uio_resid;
788 
789 			if (count == 0)
790 				error = EINVAL;
791 			else if (count != PAGE_SIZE)
792 				bzero((caddr_t)sf_buf_kva(sf) + count,
793 				    PAGE_SIZE - count);
794 		}
795 		sf_buf_free(sf);
796 
797 		VM_OBJECT_WLOCK(object);
798 	}
799 	KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
800 	if (!error)
801 		vm_page_valid(m);
802 	return error ? VM_PAGER_ERROR : VM_PAGER_OK;
803 }
804 
805 /*
806  * generic vnode pager input routine
807  */
808 
809 /*
810  * Local media VFS's that do not implement their own VOP_GETPAGES
811  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
812  * to implement the previous behaviour.
813  *
814  * All other FS's should use the bypass to get to the local media
815  * backing vp's VOP_GETPAGES.
816  */
817 static int
818 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
819     int *rahead)
820 {
821 	struct vnode *vp;
822 	int rtval;
823 
824 	/* Handle is stable with paging in progress. */
825 	vp = object->handle;
826 	rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead);
827 	KASSERT(rtval != EOPNOTSUPP,
828 	    ("vnode_pager: FS getpages not implemented\n"));
829 	return rtval;
830 }
831 
832 static int
833 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
834     int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg)
835 {
836 	struct vnode *vp;
837 	int rtval;
838 
839 	vp = object->handle;
840 	rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg);
841 	KASSERT(rtval != EOPNOTSUPP,
842 	    ("vnode_pager: FS getpages_async not implemented\n"));
843 	return (rtval);
844 }
845 
846 /*
847  * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for
848  * local filesystems, where partially valid pages can only occur at
849  * the end of file.
850  */
851 int
852 vnode_pager_local_getpages(struct vop_getpages_args *ap)
853 {
854 
855 	return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
856 	    ap->a_rbehind, ap->a_rahead, NULL, NULL));
857 }
858 
859 int
860 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap)
861 {
862 	int error;
863 
864 	error = vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
865 	    ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg);
866 	if (error != 0 && ap->a_iodone != NULL)
867 		ap->a_iodone(ap->a_arg, ap->a_m, ap->a_count, error);
868 	return (error);
869 }
870 
871 /*
872  * This is now called from local media FS's to operate against their
873  * own vnodes if they fail to implement VOP_GETPAGES.
874  */
875 int
876 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count,
877     int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg)
878 {
879 	vm_object_t object;
880 	struct bufobj *bo;
881 	struct buf *bp;
882 	off_t foff;
883 #ifdef INVARIANTS
884 	off_t blkno0;
885 #endif
886 	int bsize, pagesperblock;
887 	int error, before, after, rbehind, rahead, poff, i;
888 	int bytecount, secmask;
889 
890 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
891 	    ("%s does not support devices", __func__));
892 
893 	if (VN_IS_DOOMED(vp))
894 		return (VM_PAGER_BAD);
895 
896 	object = vp->v_object;
897 	foff = IDX_TO_OFF(m[0]->pindex);
898 	bsize = vp->v_mount->mnt_stat.f_iosize;
899 	pagesperblock = bsize / PAGE_SIZE;
900 
901 	KASSERT(foff < object->un_pager.vnp.vnp_size,
902 	    ("%s: page %p offset beyond vp %p size", __func__, m[0], vp));
903 	KASSERT(count <= atop(maxphys),
904 	    ("%s: requested %d pages", __func__, count));
905 
906 	/*
907 	 * The last page has valid blocks.  Invalid part can only
908 	 * exist at the end of file, and the page is made fully valid
909 	 * by zeroing in vm_pager_get_pages().
910 	 */
911 	if (!vm_page_none_valid(m[count - 1]) && --count == 0) {
912 		if (iodone != NULL)
913 			iodone(arg, m, 1, 0);
914 		return (VM_PAGER_OK);
915 	}
916 
917 	bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK);
918 	MPASS((bp->b_flags & B_MAXPHYS) != 0);
919 
920 	/*
921 	 * Get the underlying device blocks for the file with VOP_BMAP().
922 	 * If the file system doesn't support VOP_BMAP, use old way of
923 	 * getting pages via VOP_READ.
924 	 */
925 	error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before);
926 	if (error == EOPNOTSUPP) {
927 		uma_zfree(vnode_pbuf_zone, bp);
928 		VM_OBJECT_WLOCK(object);
929 		for (i = 0; i < count; i++) {
930 			VM_CNT_INC(v_vnodein);
931 			VM_CNT_INC(v_vnodepgsin);
932 			error = vnode_pager_input_old(object, m[i]);
933 			if (error)
934 				break;
935 		}
936 		VM_OBJECT_WUNLOCK(object);
937 		return (error);
938 	} else if (error != 0) {
939 		uma_zfree(vnode_pbuf_zone, bp);
940 		return (VM_PAGER_ERROR);
941 	}
942 
943 	/*
944 	 * If the file system supports BMAP, but blocksize is smaller
945 	 * than a page size, then use special small filesystem code.
946 	 */
947 	if (pagesperblock == 0) {
948 		uma_zfree(vnode_pbuf_zone, bp);
949 		for (i = 0; i < count; i++) {
950 			VM_CNT_INC(v_vnodein);
951 			VM_CNT_INC(v_vnodepgsin);
952 			error = vnode_pager_input_smlfs(object, m[i]);
953 			if (error)
954 				break;
955 		}
956 		return (error);
957 	}
958 
959 	/*
960 	 * A sparse file can be encountered only for a single page request,
961 	 * which may not be preceded by call to vm_pager_haspage().
962 	 */
963 	if (bp->b_blkno == -1) {
964 		KASSERT(count == 1,
965 		    ("%s: array[%d] request to a sparse file %p", __func__,
966 		    count, vp));
967 		uma_zfree(vnode_pbuf_zone, bp);
968 		pmap_zero_page(m[0]);
969 		KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty",
970 		    __func__, m[0]));
971 		vm_page_valid(m[0]);
972 		return (VM_PAGER_OK);
973 	}
974 
975 #ifdef INVARIANTS
976 	blkno0 = bp->b_blkno;
977 #endif
978 	bp->b_blkno += (foff % bsize) / DEV_BSIZE;
979 
980 	/* Recalculate blocks available after/before to pages. */
981 	poff = (foff % bsize) / PAGE_SIZE;
982 	before *= pagesperblock;
983 	before += poff;
984 	after *= pagesperblock;
985 	after += pagesperblock - (poff + 1);
986 	if (m[0]->pindex + after >= object->size)
987 		after = object->size - 1 - m[0]->pindex;
988 	KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d",
989 	    __func__, count, after + 1));
990 	after -= count - 1;
991 
992 	/* Trim requested rbehind/rahead to possible values. */
993 	rbehind = a_rbehind ? *a_rbehind : 0;
994 	rahead = a_rahead ? *a_rahead : 0;
995 	rbehind = min(rbehind, before);
996 	rbehind = min(rbehind, m[0]->pindex);
997 	rahead = min(rahead, after);
998 	rahead = min(rahead, object->size - m[count - 1]->pindex);
999 	/*
1000 	 * Check that total amount of pages fit into buf.  Trim rbehind and
1001 	 * rahead evenly if not.
1002 	 */
1003 	if (rbehind + rahead + count > atop(maxphys)) {
1004 		int trim, sum;
1005 
1006 		trim = rbehind + rahead + count - atop(maxphys) + 1;
1007 		sum = rbehind + rahead;
1008 		if (rbehind == before) {
1009 			/* Roundup rbehind trim to block size. */
1010 			rbehind -= roundup(trim * rbehind / sum, pagesperblock);
1011 			if (rbehind < 0)
1012 				rbehind = 0;
1013 		} else
1014 			rbehind -= trim * rbehind / sum;
1015 		rahead -= trim * rahead / sum;
1016 	}
1017 	KASSERT(rbehind + rahead + count <= atop(maxphys),
1018 	    ("%s: behind %d ahead %d count %d maxphys %lu", __func__,
1019 	    rbehind, rahead, count, maxphys));
1020 
1021 	/*
1022 	 * Fill in the bp->b_pages[] array with requested and optional
1023 	 * read behind or read ahead pages.  Read behind pages are looked
1024 	 * up in a backward direction, down to a first cached page.  Same
1025 	 * for read ahead pages, but there is no need to shift the array
1026 	 * in case of encountering a cached page.
1027 	 */
1028 	i = bp->b_npages = 0;
1029 	if (rbehind) {
1030 		vm_pindex_t startpindex, tpindex;
1031 		vm_page_t p;
1032 
1033 		VM_OBJECT_WLOCK(object);
1034 		startpindex = m[0]->pindex - rbehind;
1035 		if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL &&
1036 		    p->pindex >= startpindex)
1037 			startpindex = p->pindex + 1;
1038 
1039 		/* tpindex is unsigned; beware of numeric underflow. */
1040 		for (tpindex = m[0]->pindex - 1;
1041 		    tpindex >= startpindex && tpindex < m[0]->pindex;
1042 		    tpindex--, i++) {
1043 			p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1044 			if (p == NULL) {
1045 				/* Shift the array. */
1046 				for (int j = 0; j < i; j++)
1047 					bp->b_pages[j] = bp->b_pages[j +
1048 					    tpindex + 1 - startpindex];
1049 				break;
1050 			}
1051 			bp->b_pages[tpindex - startpindex] = p;
1052 		}
1053 
1054 		bp->b_pgbefore = i;
1055 		bp->b_npages += i;
1056 		bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE;
1057 	} else
1058 		bp->b_pgbefore = 0;
1059 
1060 	/* Requested pages. */
1061 	for (int j = 0; j < count; j++, i++)
1062 		bp->b_pages[i] = m[j];
1063 	bp->b_npages += count;
1064 
1065 	if (rahead) {
1066 		vm_pindex_t endpindex, tpindex;
1067 		vm_page_t p;
1068 
1069 		if (!VM_OBJECT_WOWNED(object))
1070 			VM_OBJECT_WLOCK(object);
1071 		endpindex = m[count - 1]->pindex + rahead + 1;
1072 		if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL &&
1073 		    p->pindex < endpindex)
1074 			endpindex = p->pindex;
1075 		if (endpindex > object->size)
1076 			endpindex = object->size;
1077 
1078 		for (tpindex = m[count - 1]->pindex + 1;
1079 		    tpindex < endpindex; i++, tpindex++) {
1080 			p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1081 			if (p == NULL)
1082 				break;
1083 			bp->b_pages[i] = p;
1084 		}
1085 
1086 		bp->b_pgafter = i - bp->b_npages;
1087 		bp->b_npages = i;
1088 	} else
1089 		bp->b_pgafter = 0;
1090 
1091 	if (VM_OBJECT_WOWNED(object))
1092 		VM_OBJECT_WUNLOCK(object);
1093 
1094 	/* Report back actual behind/ahead read. */
1095 	if (a_rbehind)
1096 		*a_rbehind = bp->b_pgbefore;
1097 	if (a_rahead)
1098 		*a_rahead = bp->b_pgafter;
1099 
1100 #ifdef INVARIANTS
1101 	KASSERT(bp->b_npages <= atop(maxphys),
1102 	    ("%s: buf %p overflowed", __func__, bp));
1103 	for (int j = 1, prev = 0; j < bp->b_npages; j++) {
1104 		if (bp->b_pages[j] == bogus_page)
1105 			continue;
1106 		KASSERT(bp->b_pages[j]->pindex - bp->b_pages[prev]->pindex ==
1107 		    j - prev, ("%s: pages array not consecutive, bp %p",
1108 		     __func__, bp));
1109 		prev = j;
1110 	}
1111 #endif
1112 
1113 	/*
1114 	 * Recalculate first offset and bytecount with regards to read behind.
1115 	 * Truncate bytecount to vnode real size and round up physical size
1116 	 * for real devices.
1117 	 */
1118 	foff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1119 	bytecount = bp->b_npages << PAGE_SHIFT;
1120 	if ((foff + bytecount) > object->un_pager.vnp.vnp_size)
1121 		bytecount = object->un_pager.vnp.vnp_size - foff;
1122 	secmask = bo->bo_bsize - 1;
1123 	KASSERT(secmask < PAGE_SIZE && secmask > 0,
1124 	    ("%s: sector size %d too large", __func__, secmask + 1));
1125 	bytecount = (bytecount + secmask) & ~secmask;
1126 
1127 	/*
1128 	 * And map the pages to be read into the kva, if the filesystem
1129 	 * requires mapped buffers.
1130 	 */
1131 	if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
1132 	    unmapped_buf_allowed) {
1133 		bp->b_data = unmapped_buf;
1134 		bp->b_offset = 0;
1135 	} else {
1136 		bp->b_data = bp->b_kvabase;
1137 		pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1138 	}
1139 
1140 	/* Build a minimal buffer header. */
1141 	bp->b_iocmd = BIO_READ;
1142 	KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
1143 	KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
1144 	bp->b_rcred = crhold(curthread->td_ucred);
1145 	bp->b_wcred = crhold(curthread->td_ucred);
1146 	pbgetbo(bo, bp);
1147 	bp->b_vp = vp;
1148 	bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount;
1149 	bp->b_iooffset = dbtob(bp->b_blkno);
1150 	KASSERT(IDX_TO_OFF(m[0]->pindex - bp->b_pages[0]->pindex) ==
1151 	    (blkno0 - bp->b_blkno) * DEV_BSIZE +
1152 	    IDX_TO_OFF(m[0]->pindex) % bsize,
1153 	    ("wrong offsets bsize %d m[0] %ju b_pages[0] %ju "
1154 	    "blkno0 %ju b_blkno %ju", bsize,
1155 	    (uintmax_t)m[0]->pindex, (uintmax_t)bp->b_pages[0]->pindex,
1156 	    (uintmax_t)blkno0, (uintmax_t)bp->b_blkno));
1157 
1158 	atomic_add_long(&runningbufspace, bp->b_runningbufspace);
1159 	VM_CNT_INC(v_vnodein);
1160 	VM_CNT_ADD(v_vnodepgsin, bp->b_npages);
1161 
1162 	if (iodone != NULL) { /* async */
1163 		bp->b_pgiodone = iodone;
1164 		bp->b_caller1 = arg;
1165 		bp->b_iodone = vnode_pager_generic_getpages_done_async;
1166 		bp->b_flags |= B_ASYNC;
1167 		BUF_KERNPROC(bp);
1168 		bstrategy(bp);
1169 		return (VM_PAGER_OK);
1170 	} else {
1171 		bp->b_iodone = bdone;
1172 		bstrategy(bp);
1173 		bwait(bp, PVM, "vnread");
1174 		error = vnode_pager_generic_getpages_done(bp);
1175 		for (i = 0; i < bp->b_npages; i++)
1176 			bp->b_pages[i] = NULL;
1177 		bp->b_vp = NULL;
1178 		pbrelbo(bp);
1179 		uma_zfree(vnode_pbuf_zone, bp);
1180 		return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
1181 	}
1182 }
1183 
1184 static void
1185 vnode_pager_generic_getpages_done_async(struct buf *bp)
1186 {
1187 	int error;
1188 
1189 	error = vnode_pager_generic_getpages_done(bp);
1190 	/* Run the iodone upon the requested range. */
1191 	bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore,
1192 	    bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error);
1193 	for (int i = 0; i < bp->b_npages; i++)
1194 		bp->b_pages[i] = NULL;
1195 	bp->b_vp = NULL;
1196 	pbrelbo(bp);
1197 	uma_zfree(vnode_pbuf_zone, bp);
1198 }
1199 
1200 static int
1201 vnode_pager_generic_getpages_done(struct buf *bp)
1202 {
1203 	vm_object_t object;
1204 	off_t tfoff, nextoff;
1205 	int i, error;
1206 
1207 	KASSERT((bp->b_ioflags & BIO_ERROR) == 0 || bp->b_error != 0,
1208 	    ("%s: buf error but b_error == 0\n", __func__));
1209 	error = (bp->b_ioflags & BIO_ERROR) != 0 ? bp->b_error : 0;
1210 	object = bp->b_vp->v_object;
1211 
1212 	runningbufwakeup(bp);
1213 
1214 	if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1215 		if (!buf_mapped(bp)) {
1216 			bp->b_data = bp->b_kvabase;
1217 			pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages,
1218 			    bp->b_npages);
1219 		}
1220 		bzero(bp->b_data + bp->b_bcount,
1221 		    PAGE_SIZE * bp->b_npages - bp->b_bcount);
1222 	}
1223 	if (buf_mapped(bp)) {
1224 		pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1225 		bp->b_data = unmapped_buf;
1226 	}
1227 
1228 	/*
1229 	 * If the read failed, we must free any read ahead/behind pages here.
1230 	 * The requested pages are freed by the caller (for sync requests)
1231 	 * or by the bp->b_pgiodone callback (for async requests).
1232 	 */
1233 	if (error != 0) {
1234 		VM_OBJECT_WLOCK(object);
1235 		for (i = 0; i < bp->b_pgbefore; i++)
1236 			vm_page_free_invalid(bp->b_pages[i]);
1237 		for (i = bp->b_npages - bp->b_pgafter; i < bp->b_npages; i++)
1238 			vm_page_free_invalid(bp->b_pages[i]);
1239 		VM_OBJECT_WUNLOCK(object);
1240 		return (error);
1241 	}
1242 
1243 	/* Read lock to protect size. */
1244 	VM_OBJECT_RLOCK(object);
1245 	for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1246 	    i < bp->b_npages; i++, tfoff = nextoff) {
1247 		vm_page_t mt;
1248 
1249 		nextoff = tfoff + PAGE_SIZE;
1250 		mt = bp->b_pages[i];
1251 		if (mt == bogus_page)
1252 			continue;
1253 
1254 		if (nextoff <= object->un_pager.vnp.vnp_size) {
1255 			/*
1256 			 * Read filled up entire page.
1257 			 */
1258 			vm_page_valid(mt);
1259 			KASSERT(mt->dirty == 0,
1260 			    ("%s: page %p is dirty", __func__, mt));
1261 			KASSERT(!pmap_page_is_mapped(mt),
1262 			    ("%s: page %p is mapped", __func__, mt));
1263 		} else {
1264 			/*
1265 			 * Read did not fill up entire page.
1266 			 *
1267 			 * Currently we do not set the entire page valid,
1268 			 * we just try to clear the piece that we couldn't
1269 			 * read.
1270 			 */
1271 			vm_page_set_valid_range(mt, 0,
1272 			    object->un_pager.vnp.vnp_size - tfoff);
1273 			KASSERT((mt->dirty & vm_page_bits(0,
1274 			    object->un_pager.vnp.vnp_size - tfoff)) == 0,
1275 			    ("%s: page %p is dirty", __func__, mt));
1276 		}
1277 
1278 		if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter)
1279 			vm_page_readahead_finish(mt);
1280 	}
1281 	VM_OBJECT_RUNLOCK(object);
1282 
1283 	return (error);
1284 }
1285 
1286 /*
1287  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
1288  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1289  * vnode_pager_generic_putpages() to implement the previous behaviour.
1290  *
1291  * All other FS's should use the bypass to get to the local media
1292  * backing vp's VOP_PUTPAGES.
1293  */
1294 static void
1295 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1296     int flags, int *rtvals)
1297 {
1298 	int rtval __diagused;
1299 	struct vnode *vp;
1300 	int bytes = count * PAGE_SIZE;
1301 
1302 	/*
1303 	 * Force synchronous operation if we are extremely low on memory
1304 	 * to prevent a low-memory deadlock.  VOP operations often need to
1305 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
1306 	 * operation ).  The swapper handles the case by limiting the amount
1307 	 * of asynchronous I/O, but that sort of solution doesn't scale well
1308 	 * for the vnode pager without a lot of work.
1309 	 *
1310 	 * Also, the backing vnode's iodone routine may not wake the pageout
1311 	 * daemon up.  This should be probably be addressed XXX.
1312 	 */
1313 
1314 	if (vm_page_count_min())
1315 		flags |= VM_PAGER_PUT_SYNC;
1316 
1317 	/*
1318 	 * Call device-specific putpages function
1319 	 */
1320 	vp = object->handle;
1321 	VM_OBJECT_WUNLOCK(object);
1322 	rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1323 	KASSERT(rtval != EOPNOTSUPP,
1324 	    ("vnode_pager: stale FS putpages\n"));
1325 	VM_OBJECT_WLOCK(object);
1326 }
1327 
1328 static int
1329 vn_off2bidx(vm_ooffset_t offset)
1330 {
1331 
1332 	return ((offset & PAGE_MASK) / DEV_BSIZE);
1333 }
1334 
1335 static bool
1336 vn_dirty_blk(vm_page_t m, vm_ooffset_t offset)
1337 {
1338 
1339 	KASSERT(IDX_TO_OFF(m->pindex) <= offset &&
1340 	    offset < IDX_TO_OFF(m->pindex + 1),
1341 	    ("page %p pidx %ju offset %ju", m, (uintmax_t)m->pindex,
1342 	    (uintmax_t)offset));
1343 	return ((m->dirty & ((vm_page_bits_t)1 << vn_off2bidx(offset))) != 0);
1344 }
1345 
1346 /*
1347  * This is now called from local media FS's to operate against their
1348  * own vnodes if they fail to implement VOP_PUTPAGES.
1349  *
1350  * This is typically called indirectly via the pageout daemon and
1351  * clustering has already typically occurred, so in general we ask the
1352  * underlying filesystem to write the data out asynchronously rather
1353  * then delayed.
1354  */
1355 int
1356 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1357     int flags, int *rtvals)
1358 {
1359 	vm_object_t object;
1360 	vm_page_t m;
1361 	vm_ooffset_t maxblksz, next_offset, poffset, prev_offset;
1362 	struct uio auio;
1363 	struct iovec aiov;
1364 	off_t prev_resid, wrsz;
1365 	int count, error, i, maxsize, ncount, pgoff, ppscheck;
1366 	bool in_hole;
1367 	static struct timeval lastfail;
1368 	static int curfail;
1369 
1370 	object = vp->v_object;
1371 	count = bytecount / PAGE_SIZE;
1372 
1373 	for (i = 0; i < count; i++)
1374 		rtvals[i] = VM_PAGER_ERROR;
1375 
1376 	if ((int64_t)ma[0]->pindex < 0) {
1377 		printf("vnode_pager_generic_putpages: "
1378 		    "attempt to write meta-data 0x%jx(%lx)\n",
1379 		    (uintmax_t)ma[0]->pindex, (u_long)ma[0]->dirty);
1380 		rtvals[0] = VM_PAGER_BAD;
1381 		return (VM_PAGER_BAD);
1382 	}
1383 
1384 	maxsize = count * PAGE_SIZE;
1385 	ncount = count;
1386 
1387 	poffset = IDX_TO_OFF(ma[0]->pindex);
1388 
1389 	/*
1390 	 * If the page-aligned write is larger then the actual file we
1391 	 * have to invalidate pages occurring beyond the file EOF.  However,
1392 	 * there is an edge case where a file may not be page-aligned where
1393 	 * the last page is partially invalid.  In this case the filesystem
1394 	 * may not properly clear the dirty bits for the entire page (which
1395 	 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1396 	 * With the page busied we are free to fix up the dirty bits here.
1397 	 *
1398 	 * We do not under any circumstances truncate the valid bits, as
1399 	 * this will screw up bogus page replacement.
1400 	 */
1401 	VM_OBJECT_RLOCK(object);
1402 	if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1403 		if (object->un_pager.vnp.vnp_size > poffset) {
1404 			maxsize = object->un_pager.vnp.vnp_size - poffset;
1405 			ncount = btoc(maxsize);
1406 			if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1407 				pgoff = roundup2(pgoff, DEV_BSIZE);
1408 
1409 				/*
1410 				 * If the page is busy and the following
1411 				 * conditions hold, then the page's dirty
1412 				 * field cannot be concurrently changed by a
1413 				 * pmap operation.
1414 				 */
1415 				m = ma[ncount - 1];
1416 				vm_page_assert_sbusied(m);
1417 				KASSERT(!pmap_page_is_write_mapped(m),
1418 		("vnode_pager_generic_putpages: page %p is not read-only", m));
1419 				MPASS(m->dirty != 0);
1420 				vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1421 				    pgoff);
1422 			}
1423 		} else {
1424 			maxsize = 0;
1425 			ncount = 0;
1426 		}
1427 		for (i = ncount; i < count; i++)
1428 			rtvals[i] = VM_PAGER_BAD;
1429 	}
1430 	VM_OBJECT_RUNLOCK(object);
1431 
1432 	auio.uio_iov = &aiov;
1433 	auio.uio_segflg = UIO_NOCOPY;
1434 	auio.uio_rw = UIO_WRITE;
1435 	auio.uio_td = NULL;
1436 	maxblksz = roundup2(poffset + maxsize, DEV_BSIZE);
1437 
1438 	for (prev_offset = poffset; prev_offset < maxblksz;) {
1439 		/* Skip clean blocks. */
1440 		for (in_hole = true; in_hole && prev_offset < maxblksz;) {
1441 			m = ma[OFF_TO_IDX(prev_offset - poffset)];
1442 			for (i = vn_off2bidx(prev_offset);
1443 			    i < sizeof(vm_page_bits_t) * NBBY &&
1444 			    prev_offset < maxblksz; i++) {
1445 				if (vn_dirty_blk(m, prev_offset)) {
1446 					in_hole = false;
1447 					break;
1448 				}
1449 				prev_offset += DEV_BSIZE;
1450 			}
1451 		}
1452 		if (in_hole)
1453 			goto write_done;
1454 
1455 		/* Find longest run of dirty blocks. */
1456 		for (next_offset = prev_offset; next_offset < maxblksz;) {
1457 			m = ma[OFF_TO_IDX(next_offset - poffset)];
1458 			for (i = vn_off2bidx(next_offset);
1459 			    i < sizeof(vm_page_bits_t) * NBBY &&
1460 			    next_offset < maxblksz; i++) {
1461 				if (!vn_dirty_blk(m, next_offset))
1462 					goto start_write;
1463 				next_offset += DEV_BSIZE;
1464 			}
1465 		}
1466 start_write:
1467 		if (next_offset > poffset + maxsize)
1468 			next_offset = poffset + maxsize;
1469 
1470 		/*
1471 		 * Getting here requires finding a dirty block in the
1472 		 * 'skip clean blocks' loop.
1473 		 */
1474 		MPASS(prev_offset < next_offset);
1475 
1476 		aiov.iov_base = NULL;
1477 		auio.uio_iovcnt = 1;
1478 		auio.uio_offset = prev_offset;
1479 		prev_resid = auio.uio_resid = aiov.iov_len = next_offset -
1480 		    prev_offset;
1481 		error = VOP_WRITE(vp, &auio,
1482 		    vnode_pager_putpages_ioflags(flags), curthread->td_ucred);
1483 
1484 		wrsz = prev_resid - auio.uio_resid;
1485 		if (wrsz == 0) {
1486 			if (ppsratecheck(&lastfail, &curfail, 1) != 0) {
1487 				vn_printf(vp, "vnode_pager_putpages: "
1488 				    "zero-length write at %ju resid %zd\n",
1489 				    auio.uio_offset, auio.uio_resid);
1490 			}
1491 			break;
1492 		}
1493 
1494 		/* Adjust the starting offset for next iteration. */
1495 		prev_offset += wrsz;
1496 		MPASS(auio.uio_offset == prev_offset);
1497 
1498 		ppscheck = 0;
1499 		if (error != 0 && (ppscheck = ppsratecheck(&lastfail,
1500 		    &curfail, 1)) != 0)
1501 			vn_printf(vp, "vnode_pager_putpages: I/O error %d\n",
1502 			    error);
1503 		if (auio.uio_resid != 0 && (ppscheck != 0 ||
1504 		    ppsratecheck(&lastfail, &curfail, 1) != 0))
1505 			vn_printf(vp, "vnode_pager_putpages: residual I/O %zd "
1506 			    "at %ju\n", auio.uio_resid,
1507 			    (uintmax_t)ma[0]->pindex);
1508 		if (error != 0 || auio.uio_resid != 0)
1509 			break;
1510 	}
1511 write_done:
1512 	/* Mark completely processed pages. */
1513 	for (i = 0; i < OFF_TO_IDX(prev_offset - poffset); i++)
1514 		rtvals[i] = VM_PAGER_OK;
1515 	/* Mark partial EOF page. */
1516 	if (prev_offset == poffset + maxsize && (prev_offset & PAGE_MASK) != 0)
1517 		rtvals[i++] = VM_PAGER_OK;
1518 	/* Unwritten pages in range, free bonus if the page is clean. */
1519 	for (; i < ncount; i++)
1520 		rtvals[i] = ma[i]->dirty == 0 ? VM_PAGER_OK : VM_PAGER_ERROR;
1521 	VM_CNT_ADD(v_vnodepgsout, i);
1522 	VM_CNT_INC(v_vnodeout);
1523 	return (rtvals[0]);
1524 }
1525 
1526 int
1527 vnode_pager_putpages_ioflags(int pager_flags)
1528 {
1529 	int ioflags;
1530 
1531 	/*
1532 	 * Pageouts are already clustered, use IO_ASYNC to force a
1533 	 * bawrite() rather then a bdwrite() to prevent paging I/O
1534 	 * from saturating the buffer cache.  Dummy-up the sequential
1535 	 * heuristic to cause large ranges to cluster.  If neither
1536 	 * IO_SYNC or IO_ASYNC is set, the system decides how to
1537 	 * cluster.
1538 	 */
1539 	ioflags = IO_VMIO;
1540 	if ((pager_flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) != 0)
1541 		ioflags |= IO_SYNC;
1542 	else if ((pager_flags & VM_PAGER_CLUSTER_OK) == 0)
1543 		ioflags |= IO_ASYNC;
1544 	ioflags |= (pager_flags & VM_PAGER_PUT_INVAL) != 0 ? IO_INVAL: 0;
1545 	ioflags |= (pager_flags & VM_PAGER_PUT_NOREUSE) != 0 ? IO_NOREUSE : 0;
1546 	ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1547 	return (ioflags);
1548 }
1549 
1550 /*
1551  * vnode_pager_undirty_pages().
1552  *
1553  * A helper to mark pages as clean after pageout that was possibly
1554  * done with a short write.  The lpos argument specifies the page run
1555  * length in bytes, and the written argument specifies how many bytes
1556  * were actually written.  eof is the offset past the last valid byte
1557  * in the vnode using the absolute file position of the first byte in
1558  * the run as the base from which it is computed.
1559  */
1560 void
1561 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written, off_t eof,
1562     int lpos)
1563 {
1564 	int i, pos, pos_devb;
1565 
1566 	if (written == 0 && eof >= lpos)
1567 		return;
1568 	for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1569 		if (pos < trunc_page(written)) {
1570 			rtvals[i] = VM_PAGER_OK;
1571 			vm_page_undirty(ma[i]);
1572 		} else {
1573 			/* Partially written page. */
1574 			rtvals[i] = VM_PAGER_AGAIN;
1575 			vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1576 		}
1577 	}
1578 	if (eof >= lpos) /* avoid truncation */
1579 		return;
1580 	for (pos = eof, i = OFF_TO_IDX(trunc_page(pos)); pos < lpos; i++) {
1581 		if (pos != trunc_page(pos)) {
1582 			/*
1583 			 * The page contains the last valid byte in
1584 			 * the vnode, mark the rest of the page as
1585 			 * clean, potentially making the whole page
1586 			 * clean.
1587 			 */
1588 			pos_devb = roundup2(pos & PAGE_MASK, DEV_BSIZE);
1589 			vm_page_clear_dirty(ma[i], pos_devb, PAGE_SIZE -
1590 			    pos_devb);
1591 
1592 			/*
1593 			 * If the page was cleaned, report the pageout
1594 			 * on it as successful.  msync() no longer
1595 			 * needs to write out the page, endlessly
1596 			 * creating write requests and dirty buffers.
1597 			 */
1598 			if (ma[i]->dirty == 0)
1599 				rtvals[i] = VM_PAGER_OK;
1600 
1601 			pos = round_page(pos);
1602 		} else {
1603 			/* vm_pageout_flush() clears dirty */
1604 			rtvals[i] = VM_PAGER_BAD;
1605 			pos += PAGE_SIZE;
1606 		}
1607 	}
1608 }
1609 
1610 static void
1611 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1612     vm_offset_t end)
1613 {
1614 	struct vnode *vp;
1615 	vm_ooffset_t old_wm;
1616 
1617 	VM_OBJECT_WLOCK(object);
1618 	if (object->type != OBJT_VNODE) {
1619 		VM_OBJECT_WUNLOCK(object);
1620 		return;
1621 	}
1622 	old_wm = object->un_pager.vnp.writemappings;
1623 	object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1624 	vp = object->handle;
1625 	if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1626 		ASSERT_VOP_LOCKED(vp, "v_writecount inc");
1627 		VOP_ADD_WRITECOUNT_CHECKED(vp, 1);
1628 		CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1629 		    __func__, vp, vp->v_writecount);
1630 	} else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1631 		ASSERT_VOP_LOCKED(vp, "v_writecount dec");
1632 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1633 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1634 		    __func__, vp, vp->v_writecount);
1635 	}
1636 	VM_OBJECT_WUNLOCK(object);
1637 }
1638 
1639 static void
1640 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1641     vm_offset_t end)
1642 {
1643 	struct vnode *vp;
1644 	struct mount *mp;
1645 	vm_offset_t inc;
1646 
1647 	VM_OBJECT_WLOCK(object);
1648 
1649 	/*
1650 	 * First, recheck the object type to account for the race when
1651 	 * the vnode is reclaimed.
1652 	 */
1653 	if (object->type != OBJT_VNODE) {
1654 		VM_OBJECT_WUNLOCK(object);
1655 		return;
1656 	}
1657 
1658 	/*
1659 	 * Optimize for the case when writemappings is not going to
1660 	 * zero.
1661 	 */
1662 	inc = end - start;
1663 	if (object->un_pager.vnp.writemappings != inc) {
1664 		object->un_pager.vnp.writemappings -= inc;
1665 		VM_OBJECT_WUNLOCK(object);
1666 		return;
1667 	}
1668 
1669 	vp = object->handle;
1670 	vhold(vp);
1671 	VM_OBJECT_WUNLOCK(object);
1672 	mp = NULL;
1673 	vn_start_write(vp, &mp, V_WAIT);
1674 	vn_lock(vp, LK_SHARED | LK_RETRY);
1675 
1676 	/*
1677 	 * Decrement the object's writemappings, by swapping the start
1678 	 * and end arguments for vnode_pager_update_writecount().  If
1679 	 * there was not a race with vnode reclaimation, then the
1680 	 * vnode's v_writecount is decremented.
1681 	 */
1682 	vnode_pager_update_writecount(object, end, start);
1683 	VOP_UNLOCK(vp);
1684 	vdrop(vp);
1685 	if (mp != NULL)
1686 		vn_finished_write(mp);
1687 }
1688 
1689 static void
1690 vnode_pager_getvp(vm_object_t object, struct vnode **vpp, bool *vp_heldp)
1691 {
1692 	*vpp = object->handle;
1693 }
1694