xref: /freebsd/sys/vm/vnode_pager.c (revision 9fd69f37d28cfd7438cac3eeb45fe9dd46b4d7dd)
1 /*-
2  * Copyright (c) 1990 University of Utah.
3  * Copyright (c) 1991 The Regents of the University of California.
4  * All rights reserved.
5  * Copyright (c) 1993, 1994 John S. Dyson
6  * Copyright (c) 1995, David Greenman
7  *
8  * This code is derived from software contributed to Berkeley by
9  * the Systems Programming Group of the University of Utah Computer
10  * Science Department.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by the University of
23  *	California, Berkeley and its contributors.
24  * 4. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *	from: @(#)vnode_pager.c	7.5 (Berkeley) 4/20/91
41  */
42 
43 /*
44  * Page to/from files (vnodes).
45  */
46 
47 /*
48  * TODO:
49  *	Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
50  *	greatly re-simplify the vnode_pager.
51  */
52 
53 #include <sys/cdefs.h>
54 __FBSDID("$FreeBSD$");
55 
56 #include <sys/param.h>
57 #include <sys/systm.h>
58 #include <sys/proc.h>
59 #include <sys/vnode.h>
60 #include <sys/mount.h>
61 #include <sys/bio.h>
62 #include <sys/buf.h>
63 #include <sys/vmmeter.h>
64 #include <sys/limits.h>
65 #include <sys/conf.h>
66 #include <sys/sf_buf.h>
67 
68 #include <machine/atomic.h>
69 
70 #include <vm/vm.h>
71 #include <vm/vm_object.h>
72 #include <vm/vm_page.h>
73 #include <vm/vm_pager.h>
74 #include <vm/vm_map.h>
75 #include <vm/vnode_pager.h>
76 #include <vm/vm_extern.h>
77 
78 static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
79     daddr_t *rtaddress, int *run);
80 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
81 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
82 static void vnode_pager_dealloc(vm_object_t);
83 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int);
84 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
85 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
86 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t,
87     vm_ooffset_t, struct ucred *cred);
88 
89 struct pagerops vnodepagerops = {
90 	.pgo_alloc =	vnode_pager_alloc,
91 	.pgo_dealloc =	vnode_pager_dealloc,
92 	.pgo_getpages =	vnode_pager_getpages,
93 	.pgo_putpages =	vnode_pager_putpages,
94 	.pgo_haspage =	vnode_pager_haspage,
95 };
96 
97 int vnode_pbuf_freecnt;
98 
99 /* Create the VM system backing object for this vnode */
100 int
101 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
102 {
103 	vm_object_t object;
104 	vm_ooffset_t size = isize;
105 	struct vattr va;
106 
107 	if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE)
108 		return (0);
109 
110 	while ((object = vp->v_object) != NULL) {
111 		VM_OBJECT_LOCK(object);
112 		if (!(object->flags & OBJ_DEAD)) {
113 			VM_OBJECT_UNLOCK(object);
114 			return (0);
115 		}
116 		VOP_UNLOCK(vp, 0);
117 		vm_object_set_flag(object, OBJ_DISCONNECTWNT);
118 		msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vodead", 0);
119 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
120 	}
121 
122 	if (size == 0) {
123 		if (vn_isdisk(vp, NULL)) {
124 			size = IDX_TO_OFF(INT_MAX);
125 		} else {
126 			if (VOP_GETATTR(vp, &va, td->td_ucred))
127 				return (0);
128 			size = va.va_size;
129 		}
130 	}
131 
132 	object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred);
133 	/*
134 	 * Dereference the reference we just created.  This assumes
135 	 * that the object is associated with the vp.
136 	 */
137 	VM_OBJECT_LOCK(object);
138 	object->ref_count--;
139 	VM_OBJECT_UNLOCK(object);
140 	vrele(vp);
141 
142 	KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
143 
144 	return (0);
145 }
146 
147 void
148 vnode_destroy_vobject(struct vnode *vp)
149 {
150 	struct vm_object *obj;
151 
152 	obj = vp->v_object;
153 	if (obj == NULL)
154 		return;
155 	ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
156 	VM_OBJECT_LOCK(obj);
157 	if (obj->ref_count == 0) {
158 		/*
159 		 * vclean() may be called twice. The first time
160 		 * removes the primary reference to the object,
161 		 * the second time goes one further and is a
162 		 * special-case to terminate the object.
163 		 *
164 		 * don't double-terminate the object
165 		 */
166 		if ((obj->flags & OBJ_DEAD) == 0)
167 			vm_object_terminate(obj);
168 		else
169 			VM_OBJECT_UNLOCK(obj);
170 	} else {
171 		/*
172 		 * Woe to the process that tries to page now :-).
173 		 */
174 		vm_pager_deallocate(obj);
175 		VM_OBJECT_UNLOCK(obj);
176 	}
177 	vp->v_object = NULL;
178 }
179 
180 
181 /*
182  * Allocate (or lookup) pager for a vnode.
183  * Handle is a vnode pointer.
184  *
185  * MPSAFE
186  */
187 vm_object_t
188 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
189     vm_ooffset_t offset, struct ucred *cred)
190 {
191 	vm_object_t object;
192 	struct vnode *vp;
193 
194 	/*
195 	 * Pageout to vnode, no can do yet.
196 	 */
197 	if (handle == NULL)
198 		return (NULL);
199 
200 	vp = (struct vnode *) handle;
201 
202 	/*
203 	 * If the object is being terminated, wait for it to
204 	 * go away.
205 	 */
206 retry:
207 	while ((object = vp->v_object) != NULL) {
208 		VM_OBJECT_LOCK(object);
209 		if ((object->flags & OBJ_DEAD) == 0)
210 			break;
211 		vm_object_set_flag(object, OBJ_DISCONNECTWNT);
212 		msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vadead", 0);
213 	}
214 
215 	if (vp->v_usecount == 0)
216 		panic("vnode_pager_alloc: no vnode reference");
217 
218 	if (object == NULL) {
219 		/*
220 		 * Add an object of the appropriate size
221 		 */
222 		object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
223 
224 		object->un_pager.vnp.vnp_size = size;
225 
226 		object->handle = handle;
227 		VI_LOCK(vp);
228 		if (vp->v_object != NULL) {
229 			/*
230 			 * Object has been created while we were sleeping
231 			 */
232 			VI_UNLOCK(vp);
233 			vm_object_destroy(object);
234 			goto retry;
235 		}
236 		vp->v_object = object;
237 		VI_UNLOCK(vp);
238 	} else {
239 		object->ref_count++;
240 		VM_OBJECT_UNLOCK(object);
241 	}
242 	vref(vp);
243 	return (object);
244 }
245 
246 /*
247  *	The object must be locked.
248  */
249 static void
250 vnode_pager_dealloc(object)
251 	vm_object_t object;
252 {
253 	struct vnode *vp;
254 	int refs;
255 
256 	vp = object->handle;
257 	if (vp == NULL)
258 		panic("vnode_pager_dealloc: pager already dealloced");
259 
260 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
261 	vm_object_pip_wait(object, "vnpdea");
262 	refs = object->ref_count;
263 
264 	object->handle = NULL;
265 	object->type = OBJT_DEAD;
266 	if (object->flags & OBJ_DISCONNECTWNT) {
267 		vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
268 		wakeup(object);
269 	}
270 	ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
271 	vp->v_object = NULL;
272 	vp->v_vflag &= ~VV_TEXT;
273 	while (refs-- > 0)
274 		vunref(vp);
275 }
276 
277 static boolean_t
278 vnode_pager_haspage(object, pindex, before, after)
279 	vm_object_t object;
280 	vm_pindex_t pindex;
281 	int *before;
282 	int *after;
283 {
284 	struct vnode *vp = object->handle;
285 	daddr_t bn;
286 	int err;
287 	daddr_t reqblock;
288 	int poff;
289 	int bsize;
290 	int pagesperblock, blocksperpage;
291 	int vfslocked;
292 
293 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
294 	/*
295 	 * If no vp or vp is doomed or marked transparent to VM, we do not
296 	 * have the page.
297 	 */
298 	if (vp == NULL || vp->v_iflag & VI_DOOMED)
299 		return FALSE;
300 	/*
301 	 * If the offset is beyond end of file we do
302 	 * not have the page.
303 	 */
304 	if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
305 		return FALSE;
306 
307 	bsize = vp->v_mount->mnt_stat.f_iosize;
308 	pagesperblock = bsize / PAGE_SIZE;
309 	blocksperpage = 0;
310 	if (pagesperblock > 0) {
311 		reqblock = pindex / pagesperblock;
312 	} else {
313 		blocksperpage = (PAGE_SIZE / bsize);
314 		reqblock = pindex * blocksperpage;
315 	}
316 	VM_OBJECT_UNLOCK(object);
317 	vfslocked = VFS_LOCK_GIANT(vp->v_mount);
318 	err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
319 	VFS_UNLOCK_GIANT(vfslocked);
320 	VM_OBJECT_LOCK(object);
321 	if (err)
322 		return TRUE;
323 	if (bn == -1)
324 		return FALSE;
325 	if (pagesperblock > 0) {
326 		poff = pindex - (reqblock * pagesperblock);
327 		if (before) {
328 			*before *= pagesperblock;
329 			*before += poff;
330 		}
331 		if (after) {
332 			int numafter;
333 			*after *= pagesperblock;
334 			numafter = pagesperblock - (poff + 1);
335 			if (IDX_TO_OFF(pindex + numafter) >
336 			    object->un_pager.vnp.vnp_size) {
337 				numafter =
338 		    		    OFF_TO_IDX(object->un_pager.vnp.vnp_size) -
339 				    pindex;
340 			}
341 			*after += numafter;
342 		}
343 	} else {
344 		if (before) {
345 			*before /= blocksperpage;
346 		}
347 
348 		if (after) {
349 			*after /= blocksperpage;
350 		}
351 	}
352 	return TRUE;
353 }
354 
355 /*
356  * Lets the VM system know about a change in size for a file.
357  * We adjust our own internal size and flush any cached pages in
358  * the associated object that are affected by the size change.
359  *
360  * Note: this routine may be invoked as a result of a pager put
361  * operation (possibly at object termination time), so we must be careful.
362  */
363 void
364 vnode_pager_setsize(vp, nsize)
365 	struct vnode *vp;
366 	vm_ooffset_t nsize;
367 {
368 	vm_object_t object;
369 	vm_page_t m;
370 	vm_pindex_t nobjsize;
371 
372 	if ((object = vp->v_object) == NULL)
373 		return;
374 /* 	ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */
375 	VM_OBJECT_LOCK(object);
376 	if (nsize == object->un_pager.vnp.vnp_size) {
377 		/*
378 		 * Hasn't changed size
379 		 */
380 		VM_OBJECT_UNLOCK(object);
381 		return;
382 	}
383 	nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
384 	if (nsize < object->un_pager.vnp.vnp_size) {
385 		/*
386 		 * File has shrunk. Toss any cached pages beyond the new EOF.
387 		 */
388 		if (nobjsize < object->size)
389 			vm_object_page_remove(object, nobjsize, object->size,
390 			    FALSE);
391 		/*
392 		 * this gets rid of garbage at the end of a page that is now
393 		 * only partially backed by the vnode.
394 		 *
395 		 * XXX for some reason (I don't know yet), if we take a
396 		 * completely invalid page and mark it partially valid
397 		 * it can screw up NFS reads, so we don't allow the case.
398 		 */
399 		if ((nsize & PAGE_MASK) &&
400 		    (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
401 		    m->valid != 0) {
402 			int base = (int)nsize & PAGE_MASK;
403 			int size = PAGE_SIZE - base;
404 
405 			/*
406 			 * Clear out partial-page garbage in case
407 			 * the page has been mapped.
408 			 */
409 			pmap_zero_page_area(m, base, size);
410 
411 			/*
412 			 * Update the valid bits to reflect the blocks that
413 			 * have been zeroed.  Some of these valid bits may
414 			 * have already been set.
415 			 */
416 			vm_page_set_valid(m, base, size);
417 
418 			/*
419 			 * Round "base" to the next block boundary so that the
420 			 * dirty bit for a partially zeroed block is not
421 			 * cleared.
422 			 */
423 			base = roundup2(base, DEV_BSIZE);
424 
425 			/*
426 			 * Clear out partial-page dirty bits.
427 			 *
428 			 * note that we do not clear out the valid
429 			 * bits.  This would prevent bogus_page
430 			 * replacement from working properly.
431 			 */
432 			vm_page_lock_queues();
433 			vm_page_clear_dirty(m, base, PAGE_SIZE - base);
434 			vm_page_unlock_queues();
435 		} else if ((nsize & PAGE_MASK) &&
436 		    __predict_false(object->cache != NULL)) {
437 			vm_page_cache_free(object, OFF_TO_IDX(nsize),
438 			    nobjsize);
439 		}
440 	}
441 	object->un_pager.vnp.vnp_size = nsize;
442 	object->size = nobjsize;
443 	VM_OBJECT_UNLOCK(object);
444 }
445 
446 /*
447  * calculate the linear (byte) disk address of specified virtual
448  * file address
449  */
450 static int
451 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
452     int *run)
453 {
454 	int bsize;
455 	int err;
456 	daddr_t vblock;
457 	daddr_t voffset;
458 
459 	if (address < 0)
460 		return -1;
461 
462 	if (vp->v_iflag & VI_DOOMED)
463 		return -1;
464 
465 	bsize = vp->v_mount->mnt_stat.f_iosize;
466 	vblock = address / bsize;
467 	voffset = address % bsize;
468 
469 	err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
470 	if (err == 0) {
471 		if (*rtaddress != -1)
472 			*rtaddress += voffset / DEV_BSIZE;
473 		if (run) {
474 			*run += 1;
475 			*run *= bsize/PAGE_SIZE;
476 			*run -= voffset/PAGE_SIZE;
477 		}
478 	}
479 
480 	return (err);
481 }
482 
483 /*
484  * small block filesystem vnode pager input
485  */
486 static int
487 vnode_pager_input_smlfs(object, m)
488 	vm_object_t object;
489 	vm_page_t m;
490 {
491 	int bits, i;
492 	struct vnode *vp;
493 	struct bufobj *bo;
494 	struct buf *bp;
495 	struct sf_buf *sf;
496 	daddr_t fileaddr;
497 	vm_offset_t bsize;
498 	int error = 0;
499 
500 	vp = object->handle;
501 	if (vp->v_iflag & VI_DOOMED)
502 		return VM_PAGER_BAD;
503 
504 	bsize = vp->v_mount->mnt_stat.f_iosize;
505 
506 	VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
507 
508 	sf = sf_buf_alloc(m, 0);
509 
510 	for (i = 0; i < PAGE_SIZE / bsize; i++) {
511 		vm_ooffset_t address;
512 
513 		bits = vm_page_bits(i * bsize, bsize);
514 		if (m->valid & bits)
515 			continue;
516 
517 		address = IDX_TO_OFF(m->pindex) + i * bsize;
518 		if (address >= object->un_pager.vnp.vnp_size) {
519 			fileaddr = -1;
520 		} else {
521 			error = vnode_pager_addr(vp, address, &fileaddr, NULL);
522 			if (error)
523 				break;
524 		}
525 		if (fileaddr != -1) {
526 			bp = getpbuf(&vnode_pbuf_freecnt);
527 
528 			/* build a minimal buffer header */
529 			bp->b_iocmd = BIO_READ;
530 			bp->b_iodone = bdone;
531 			KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
532 			KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
533 			bp->b_rcred = crhold(curthread->td_ucred);
534 			bp->b_wcred = crhold(curthread->td_ucred);
535 			bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
536 			bp->b_blkno = fileaddr;
537 			pbgetbo(bo, bp);
538 			bp->b_bcount = bsize;
539 			bp->b_bufsize = bsize;
540 			bp->b_runningbufspace = bp->b_bufsize;
541 			atomic_add_long(&runningbufspace, bp->b_runningbufspace);
542 
543 			/* do the input */
544 			bp->b_iooffset = dbtob(bp->b_blkno);
545 			bstrategy(bp);
546 
547 			bwait(bp, PVM, "vnsrd");
548 
549 			if ((bp->b_ioflags & BIO_ERROR) != 0)
550 				error = EIO;
551 
552 			/*
553 			 * free the buffer header back to the swap buffer pool
554 			 */
555 			pbrelbo(bp);
556 			relpbuf(bp, &vnode_pbuf_freecnt);
557 			if (error)
558 				break;
559 		} else
560 			bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
561 		KASSERT((m->dirty & bits) == 0,
562 		    ("vnode_pager_input_smlfs: page %p is dirty", m));
563 		VM_OBJECT_LOCK(object);
564 		m->valid |= bits;
565 		VM_OBJECT_UNLOCK(object);
566 	}
567 	sf_buf_free(sf);
568 	if (error) {
569 		return VM_PAGER_ERROR;
570 	}
571 	return VM_PAGER_OK;
572 }
573 
574 /*
575  * old style vnode pager input routine
576  */
577 static int
578 vnode_pager_input_old(object, m)
579 	vm_object_t object;
580 	vm_page_t m;
581 {
582 	struct uio auio;
583 	struct iovec aiov;
584 	int error;
585 	int size;
586 	struct sf_buf *sf;
587 	struct vnode *vp;
588 
589 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
590 	error = 0;
591 
592 	/*
593 	 * Return failure if beyond current EOF
594 	 */
595 	if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
596 		return VM_PAGER_BAD;
597 	} else {
598 		size = PAGE_SIZE;
599 		if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
600 			size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
601 		vp = object->handle;
602 		VM_OBJECT_UNLOCK(object);
603 
604 		/*
605 		 * Allocate a kernel virtual address and initialize so that
606 		 * we can use VOP_READ/WRITE routines.
607 		 */
608 		sf = sf_buf_alloc(m, 0);
609 
610 		aiov.iov_base = (caddr_t)sf_buf_kva(sf);
611 		aiov.iov_len = size;
612 		auio.uio_iov = &aiov;
613 		auio.uio_iovcnt = 1;
614 		auio.uio_offset = IDX_TO_OFF(m->pindex);
615 		auio.uio_segflg = UIO_SYSSPACE;
616 		auio.uio_rw = UIO_READ;
617 		auio.uio_resid = size;
618 		auio.uio_td = curthread;
619 
620 		error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
621 		if (!error) {
622 			int count = size - auio.uio_resid;
623 
624 			if (count == 0)
625 				error = EINVAL;
626 			else if (count != PAGE_SIZE)
627 				bzero((caddr_t)sf_buf_kva(sf) + count,
628 				    PAGE_SIZE - count);
629 		}
630 		sf_buf_free(sf);
631 
632 		VM_OBJECT_LOCK(object);
633 	}
634 	KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
635 	if (!error)
636 		m->valid = VM_PAGE_BITS_ALL;
637 	return error ? VM_PAGER_ERROR : VM_PAGER_OK;
638 }
639 
640 /*
641  * generic vnode pager input routine
642  */
643 
644 /*
645  * Local media VFS's that do not implement their own VOP_GETPAGES
646  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
647  * to implement the previous behaviour.
648  *
649  * All other FS's should use the bypass to get to the local media
650  * backing vp's VOP_GETPAGES.
651  */
652 static int
653 vnode_pager_getpages(object, m, count, reqpage)
654 	vm_object_t object;
655 	vm_page_t *m;
656 	int count;
657 	int reqpage;
658 {
659 	int rtval;
660 	struct vnode *vp;
661 	int bytes = count * PAGE_SIZE;
662 	int vfslocked;
663 
664 	vp = object->handle;
665 	VM_OBJECT_UNLOCK(object);
666 	vfslocked = VFS_LOCK_GIANT(vp->v_mount);
667 	rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
668 	KASSERT(rtval != EOPNOTSUPP,
669 	    ("vnode_pager: FS getpages not implemented\n"));
670 	VFS_UNLOCK_GIANT(vfslocked);
671 	VM_OBJECT_LOCK(object);
672 	return rtval;
673 }
674 
675 /*
676  * This is now called from local media FS's to operate against their
677  * own vnodes if they fail to implement VOP_GETPAGES.
678  */
679 int
680 vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
681 	struct vnode *vp;
682 	vm_page_t *m;
683 	int bytecount;
684 	int reqpage;
685 {
686 	vm_object_t object;
687 	vm_offset_t kva;
688 	off_t foff, tfoff, nextoff;
689 	int i, j, size, bsize, first;
690 	daddr_t firstaddr, reqblock;
691 	struct bufobj *bo;
692 	int runpg;
693 	int runend;
694 	struct buf *bp;
695 	int count;
696 	int error;
697 
698 	object = vp->v_object;
699 	count = bytecount / PAGE_SIZE;
700 
701 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
702 	    ("vnode_pager_generic_getpages does not support devices"));
703 	if (vp->v_iflag & VI_DOOMED)
704 		return VM_PAGER_BAD;
705 
706 	bsize = vp->v_mount->mnt_stat.f_iosize;
707 
708 	/* get the UNDERLYING device for the file with VOP_BMAP() */
709 
710 	/*
711 	 * originally, we did not check for an error return value -- assuming
712 	 * an fs always has a bmap entry point -- that assumption is wrong!!!
713 	 */
714 	foff = IDX_TO_OFF(m[reqpage]->pindex);
715 
716 	/*
717 	 * if we can't bmap, use old VOP code
718 	 */
719 	error = VOP_BMAP(vp, foff / bsize, &bo, &reqblock, NULL, NULL);
720 	if (error == EOPNOTSUPP) {
721 		VM_OBJECT_LOCK(object);
722 		vm_page_lock_queues();
723 		for (i = 0; i < count; i++)
724 			if (i != reqpage)
725 				vm_page_free(m[i]);
726 		vm_page_unlock_queues();
727 		PCPU_INC(cnt.v_vnodein);
728 		PCPU_INC(cnt.v_vnodepgsin);
729 		error = vnode_pager_input_old(object, m[reqpage]);
730 		VM_OBJECT_UNLOCK(object);
731 		return (error);
732 	} else if (error != 0) {
733 		VM_OBJECT_LOCK(object);
734 		vm_page_lock_queues();
735 		for (i = 0; i < count; i++)
736 			if (i != reqpage)
737 				vm_page_free(m[i]);
738 		vm_page_unlock_queues();
739 		VM_OBJECT_UNLOCK(object);
740 		return (VM_PAGER_ERROR);
741 
742 		/*
743 		 * if the blocksize is smaller than a page size, then use
744 		 * special small filesystem code.  NFS sometimes has a small
745 		 * blocksize, but it can handle large reads itself.
746 		 */
747 	} else if ((PAGE_SIZE / bsize) > 1 &&
748 	    (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
749 		VM_OBJECT_LOCK(object);
750 		vm_page_lock_queues();
751 		for (i = 0; i < count; i++)
752 			if (i != reqpage)
753 				vm_page_free(m[i]);
754 		vm_page_unlock_queues();
755 		VM_OBJECT_UNLOCK(object);
756 		PCPU_INC(cnt.v_vnodein);
757 		PCPU_INC(cnt.v_vnodepgsin);
758 		return vnode_pager_input_smlfs(object, m[reqpage]);
759 	}
760 
761 	/*
762 	 * If we have a completely valid page available to us, we can
763 	 * clean up and return.  Otherwise we have to re-read the
764 	 * media.
765 	 */
766 	VM_OBJECT_LOCK(object);
767 	if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
768 		vm_page_lock_queues();
769 		for (i = 0; i < count; i++)
770 			if (i != reqpage)
771 				vm_page_free(m[i]);
772 		vm_page_unlock_queues();
773 		VM_OBJECT_UNLOCK(object);
774 		return VM_PAGER_OK;
775 	} else if (reqblock == -1) {
776 		pmap_zero_page(m[reqpage]);
777 		KASSERT(m[reqpage]->dirty == 0,
778 		    ("vnode_pager_generic_getpages: page %p is dirty", m));
779 		m[reqpage]->valid = VM_PAGE_BITS_ALL;
780 		vm_page_lock_queues();
781 		for (i = 0; i < count; i++)
782 			if (i != reqpage)
783 				vm_page_free(m[i]);
784 		vm_page_unlock_queues();
785 		VM_OBJECT_UNLOCK(object);
786 		return (VM_PAGER_OK);
787 	}
788 	m[reqpage]->valid = 0;
789 	VM_OBJECT_UNLOCK(object);
790 
791 	/*
792 	 * here on direct device I/O
793 	 */
794 	firstaddr = -1;
795 
796 	/*
797 	 * calculate the run that includes the required page
798 	 */
799 	for (first = 0, i = 0; i < count; i = runend) {
800 		if (vnode_pager_addr(vp, IDX_TO_OFF(m[i]->pindex), &firstaddr,
801 		    &runpg) != 0) {
802 			VM_OBJECT_LOCK(object);
803 			vm_page_lock_queues();
804 			for (; i < count; i++)
805 				if (i != reqpage)
806 					vm_page_free(m[i]);
807 			vm_page_unlock_queues();
808 			VM_OBJECT_UNLOCK(object);
809 			return (VM_PAGER_ERROR);
810 		}
811 		if (firstaddr == -1) {
812 			VM_OBJECT_LOCK(object);
813 			if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
814 				panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
815 				    (intmax_t)firstaddr, (uintmax_t)(foff >> 32),
816 				    (uintmax_t)foff,
817 				    (uintmax_t)
818 				    (object->un_pager.vnp.vnp_size >> 32),
819 				    (uintmax_t)object->un_pager.vnp.vnp_size);
820 			}
821 			vm_page_lock_queues();
822 			vm_page_free(m[i]);
823 			vm_page_unlock_queues();
824 			VM_OBJECT_UNLOCK(object);
825 			runend = i + 1;
826 			first = runend;
827 			continue;
828 		}
829 		runend = i + runpg;
830 		if (runend <= reqpage) {
831 			VM_OBJECT_LOCK(object);
832 			vm_page_lock_queues();
833 			for (j = i; j < runend; j++)
834 				vm_page_free(m[j]);
835 			vm_page_unlock_queues();
836 			VM_OBJECT_UNLOCK(object);
837 		} else {
838 			if (runpg < (count - first)) {
839 				VM_OBJECT_LOCK(object);
840 				vm_page_lock_queues();
841 				for (i = first + runpg; i < count; i++)
842 					vm_page_free(m[i]);
843 				vm_page_unlock_queues();
844 				VM_OBJECT_UNLOCK(object);
845 				count = first + runpg;
846 			}
847 			break;
848 		}
849 		first = runend;
850 	}
851 
852 	/*
853 	 * the first and last page have been calculated now, move input pages
854 	 * to be zero based...
855 	 */
856 	if (first != 0) {
857 		m += first;
858 		count -= first;
859 		reqpage -= first;
860 	}
861 
862 	/*
863 	 * calculate the file virtual address for the transfer
864 	 */
865 	foff = IDX_TO_OFF(m[0]->pindex);
866 
867 	/*
868 	 * calculate the size of the transfer
869 	 */
870 	size = count * PAGE_SIZE;
871 	KASSERT(count > 0, ("zero count"));
872 	if ((foff + size) > object->un_pager.vnp.vnp_size)
873 		size = object->un_pager.vnp.vnp_size - foff;
874 	KASSERT(size > 0, ("zero size"));
875 
876 	/*
877 	 * round up physical size for real devices.
878 	 */
879 	if (1) {
880 		int secmask = bo->bo_bsize - 1;
881 		KASSERT(secmask < PAGE_SIZE && secmask > 0,
882 		    ("vnode_pager_generic_getpages: sector size %d too large",
883 		    secmask + 1));
884 		size = (size + secmask) & ~secmask;
885 	}
886 
887 	bp = getpbuf(&vnode_pbuf_freecnt);
888 	kva = (vm_offset_t) bp->b_data;
889 
890 	/*
891 	 * and map the pages to be read into the kva
892 	 */
893 	pmap_qenter(kva, m, count);
894 
895 	/* build a minimal buffer header */
896 	bp->b_iocmd = BIO_READ;
897 	bp->b_iodone = bdone;
898 	KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
899 	KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
900 	bp->b_rcred = crhold(curthread->td_ucred);
901 	bp->b_wcred = crhold(curthread->td_ucred);
902 	bp->b_blkno = firstaddr;
903 	pbgetbo(bo, bp);
904 	bp->b_bcount = size;
905 	bp->b_bufsize = size;
906 	bp->b_runningbufspace = bp->b_bufsize;
907 	atomic_add_long(&runningbufspace, bp->b_runningbufspace);
908 
909 	PCPU_INC(cnt.v_vnodein);
910 	PCPU_ADD(cnt.v_vnodepgsin, count);
911 
912 	/* do the input */
913 	bp->b_iooffset = dbtob(bp->b_blkno);
914 	bstrategy(bp);
915 
916 	bwait(bp, PVM, "vnread");
917 
918 	if ((bp->b_ioflags & BIO_ERROR) != 0)
919 		error = EIO;
920 
921 	if (!error) {
922 		if (size != count * PAGE_SIZE)
923 			bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
924 	}
925 	pmap_qremove(kva, count);
926 
927 	/*
928 	 * free the buffer header back to the swap buffer pool
929 	 */
930 	pbrelbo(bp);
931 	relpbuf(bp, &vnode_pbuf_freecnt);
932 
933 	VM_OBJECT_LOCK(object);
934 	vm_page_lock_queues();
935 	for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
936 		vm_page_t mt;
937 
938 		nextoff = tfoff + PAGE_SIZE;
939 		mt = m[i];
940 
941 		if (nextoff <= object->un_pager.vnp.vnp_size) {
942 			/*
943 			 * Read filled up entire page.
944 			 */
945 			mt->valid = VM_PAGE_BITS_ALL;
946 			KASSERT(mt->dirty == 0,
947 			    ("vnode_pager_generic_getpages: page %p is dirty",
948 			    mt));
949 			KASSERT(!pmap_page_is_mapped(mt),
950 			    ("vnode_pager_generic_getpages: page %p is mapped",
951 			    mt));
952 		} else {
953 			/*
954 			 * Read did not fill up entire page.
955 			 *
956 			 * Currently we do not set the entire page valid,
957 			 * we just try to clear the piece that we couldn't
958 			 * read.
959 			 */
960 			vm_page_set_valid(mt, 0,
961 			    object->un_pager.vnp.vnp_size - tfoff);
962 			KASSERT((mt->dirty & vm_page_bits(0,
963 			    object->un_pager.vnp.vnp_size - tfoff)) == 0,
964 			    ("vnode_pager_generic_getpages: page %p is dirty",
965 			    mt));
966 		}
967 
968 		if (i != reqpage) {
969 
970 			/*
971 			 * whether or not to leave the page activated is up in
972 			 * the air, but we should put the page on a page queue
973 			 * somewhere. (it already is in the object). Result:
974 			 * It appears that empirical results show that
975 			 * deactivating pages is best.
976 			 */
977 
978 			/*
979 			 * just in case someone was asking for this page we
980 			 * now tell them that it is ok to use
981 			 */
982 			if (!error) {
983 				if (mt->oflags & VPO_WANTED)
984 					vm_page_activate(mt);
985 				else
986 					vm_page_deactivate(mt);
987 				vm_page_wakeup(mt);
988 			} else {
989 				vm_page_free(mt);
990 			}
991 		}
992 	}
993 	vm_page_unlock_queues();
994 	VM_OBJECT_UNLOCK(object);
995 	if (error) {
996 		printf("vnode_pager_getpages: I/O read error\n");
997 	}
998 	return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
999 }
1000 
1001 /*
1002  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
1003  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1004  * vnode_pager_generic_putpages() to implement the previous behaviour.
1005  *
1006  * All other FS's should use the bypass to get to the local media
1007  * backing vp's VOP_PUTPAGES.
1008  */
1009 static void
1010 vnode_pager_putpages(object, m, count, sync, rtvals)
1011 	vm_object_t object;
1012 	vm_page_t *m;
1013 	int count;
1014 	boolean_t sync;
1015 	int *rtvals;
1016 {
1017 	int rtval;
1018 	struct vnode *vp;
1019 	int bytes = count * PAGE_SIZE;
1020 
1021 	/*
1022 	 * Force synchronous operation if we are extremely low on memory
1023 	 * to prevent a low-memory deadlock.  VOP operations often need to
1024 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
1025 	 * operation ).  The swapper handles the case by limiting the amount
1026 	 * of asynchronous I/O, but that sort of solution doesn't scale well
1027 	 * for the vnode pager without a lot of work.
1028 	 *
1029 	 * Also, the backing vnode's iodone routine may not wake the pageout
1030 	 * daemon up.  This should be probably be addressed XXX.
1031 	 */
1032 
1033 	if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
1034 		sync |= OBJPC_SYNC;
1035 
1036 	/*
1037 	 * Call device-specific putpages function
1038 	 */
1039 	vp = object->handle;
1040 	VM_OBJECT_UNLOCK(object);
1041 	rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
1042 	KASSERT(rtval != EOPNOTSUPP,
1043 	    ("vnode_pager: stale FS putpages\n"));
1044 	VM_OBJECT_LOCK(object);
1045 }
1046 
1047 
1048 /*
1049  * This is now called from local media FS's to operate against their
1050  * own vnodes if they fail to implement VOP_PUTPAGES.
1051  *
1052  * This is typically called indirectly via the pageout daemon and
1053  * clustering has already typically occured, so in general we ask the
1054  * underlying filesystem to write the data out asynchronously rather
1055  * then delayed.
1056  */
1057 int
1058 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals)
1059 	struct vnode *vp;
1060 	vm_page_t *m;
1061 	int bytecount;
1062 	int flags;
1063 	int *rtvals;
1064 {
1065 	int i;
1066 	vm_object_t object;
1067 	int count;
1068 
1069 	int maxsize, ncount;
1070 	vm_ooffset_t poffset;
1071 	struct uio auio;
1072 	struct iovec aiov;
1073 	int error;
1074 	int ioflags;
1075 	int ppscheck = 0;
1076 	static struct timeval lastfail;
1077 	static int curfail;
1078 
1079 	object = vp->v_object;
1080 	count = bytecount / PAGE_SIZE;
1081 
1082 	for (i = 0; i < count; i++)
1083 		rtvals[i] = VM_PAGER_AGAIN;
1084 
1085 	if ((int64_t)m[0]->pindex < 0) {
1086 		printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1087 			(long)m[0]->pindex, (u_long)m[0]->dirty);
1088 		rtvals[0] = VM_PAGER_BAD;
1089 		return VM_PAGER_BAD;
1090 	}
1091 
1092 	maxsize = count * PAGE_SIZE;
1093 	ncount = count;
1094 
1095 	poffset = IDX_TO_OFF(m[0]->pindex);
1096 
1097 	/*
1098 	 * If the page-aligned write is larger then the actual file we
1099 	 * have to invalidate pages occuring beyond the file EOF.  However,
1100 	 * there is an edge case where a file may not be page-aligned where
1101 	 * the last page is partially invalid.  In this case the filesystem
1102 	 * may not properly clear the dirty bits for the entire page (which
1103 	 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1104 	 * With the page locked we are free to fix-up the dirty bits here.
1105 	 *
1106 	 * We do not under any circumstances truncate the valid bits, as
1107 	 * this will screw up bogus page replacement.
1108 	 */
1109 	if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1110 		if (object->un_pager.vnp.vnp_size > poffset) {
1111 			int pgoff;
1112 
1113 			maxsize = object->un_pager.vnp.vnp_size - poffset;
1114 			ncount = btoc(maxsize);
1115 			if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1116 				vm_page_lock_queues();
1117 				vm_page_clear_dirty(m[ncount - 1], pgoff,
1118 					PAGE_SIZE - pgoff);
1119 				vm_page_unlock_queues();
1120 			}
1121 		} else {
1122 			maxsize = 0;
1123 			ncount = 0;
1124 		}
1125 		if (ncount < count) {
1126 			for (i = ncount; i < count; i++) {
1127 				rtvals[i] = VM_PAGER_BAD;
1128 			}
1129 		}
1130 	}
1131 
1132 	/*
1133 	 * pageouts are already clustered, use IO_ASYNC t o force a bawrite()
1134 	 * rather then a bdwrite() to prevent paging I/O from saturating
1135 	 * the buffer cache.  Dummy-up the sequential heuristic to cause
1136 	 * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
1137 	 * the system decides how to cluster.
1138 	 */
1139 	ioflags = IO_VMIO;
1140 	if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1141 		ioflags |= IO_SYNC;
1142 	else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1143 		ioflags |= IO_ASYNC;
1144 	ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1145 	ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1146 
1147 	aiov.iov_base = (caddr_t) 0;
1148 	aiov.iov_len = maxsize;
1149 	auio.uio_iov = &aiov;
1150 	auio.uio_iovcnt = 1;
1151 	auio.uio_offset = poffset;
1152 	auio.uio_segflg = UIO_NOCOPY;
1153 	auio.uio_rw = UIO_WRITE;
1154 	auio.uio_resid = maxsize;
1155 	auio.uio_td = (struct thread *) 0;
1156 	error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1157 	PCPU_INC(cnt.v_vnodeout);
1158 	PCPU_ADD(cnt.v_vnodepgsout, ncount);
1159 
1160 	if (error) {
1161 		if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1)))
1162 			printf("vnode_pager_putpages: I/O error %d\n", error);
1163 	}
1164 	if (auio.uio_resid) {
1165 		if (ppscheck || ppsratecheck(&lastfail, &curfail, 1))
1166 			printf("vnode_pager_putpages: residual I/O %zd at %lu\n",
1167 			    auio.uio_resid, (u_long)m[0]->pindex);
1168 	}
1169 	for (i = 0; i < ncount; i++) {
1170 		rtvals[i] = VM_PAGER_OK;
1171 	}
1172 	return rtvals[0];
1173 }
1174