xref: /dflybsd-src/sys/vm/vnode_pager.c (revision 2e7bf158f373428dba2c765c927f14d9e94f00a4)
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  * $FreeBSD: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $
42  * $DragonFly: src/sys/vm/vnode_pager.c,v 1.43 2008/06/19 23:27:39 dillon Exp $
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/param.h>
56 #include <sys/systm.h>
57 #include <sys/kernel.h>
58 #include <sys/proc.h>
59 #include <sys/vnode.h>
60 #include <sys/mount.h>
61 #include <sys/buf.h>
62 #include <sys/vmmeter.h>
63 #include <sys/conf.h>
64 
65 #include <cpu/lwbuf.h>
66 
67 #include <vm/vm.h>
68 #include <vm/vm_object.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_pager.h>
71 #include <vm/vm_map.h>
72 #include <vm/vnode_pager.h>
73 #include <vm/swap_pager.h>
74 #include <vm/vm_extern.h>
75 
76 #include <sys/thread2.h>
77 #include <vm/vm_page2.h>
78 
79 static void vnode_pager_dealloc (vm_object_t);
80 static int vnode_pager_getpage (vm_object_t, vm_page_t *, int);
81 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *);
82 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t);
83 
84 struct pagerops vnodepagerops = {
85 	vnode_pager_dealloc,
86 	vnode_pager_getpage,
87 	vnode_pager_putpages,
88 	vnode_pager_haspage
89 };
90 
91 static struct krate vbadrate = { 1 };
92 static struct krate vresrate = { 1 };
93 
94 int vnode_pbuf_freecnt = -1;	/* start out unlimited */
95 
96 /*
97  * Allocate a VM object for a vnode, typically a regular file vnode.
98  *
99  * Some additional information is required to generate a properly sized
100  * object which covers the entire buffer cache buffer straddling the file
101  * EOF.  Userland does not see the extra pages as the VM fault code tests
102  * against v_filesize.
103  */
104 vm_object_t
105 vnode_pager_alloc(void *handle, off_t length, vm_prot_t prot, off_t offset,
106 		  int blksize, int boff)
107 {
108 	vm_object_t object;
109 	struct vnode *vp;
110 	off_t loffset;
111 	vm_pindex_t lsize;
112 
113 	/*
114 	 * Pageout to vnode, no can do yet.
115 	 */
116 	if (handle == NULL)
117 		return (NULL);
118 
119 	/*
120 	 * XXX hack - This initialization should be put somewhere else.
121 	 */
122 	if (vnode_pbuf_freecnt < 0) {
123 	    vnode_pbuf_freecnt = nswbuf / 2 + 1;
124 	}
125 
126 	vp = (struct vnode *) handle;
127 
128 	/*
129 	 * Prevent race condition when allocating the object. This
130 	 * can happen with NFS vnodes since the nfsnode isn't locked.
131 	 */
132 	while (vp->v_flag & VOLOCK) {
133 		vsetflags(vp, VOWANT);
134 		tsleep(vp, 0, "vnpobj", 0);
135 	}
136 	vsetflags(vp, VOLOCK);
137 
138 	/*
139 	 * If the object is being terminated, wait for it to
140 	 * go away.
141 	 */
142 	while (((object = vp->v_object) != NULL) &&
143 		(object->flags & OBJ_DEAD)) {
144 		vm_object_dead_sleep(object, "vadead");
145 	}
146 
147 	if (vp->v_sysref.refcnt <= 0)
148 		panic("vnode_pager_alloc: no vnode reference");
149 
150 	/*
151 	 * Round up to the *next* block, then destroy the buffers in question.
152 	 * Since we are only removing some of the buffers we must rely on the
153 	 * scan count to determine whether a loop is necessary.
154 	 *
155 	 * Destroy any pages beyond the last buffer.
156 	 */
157 	if (boff < 0)
158 		boff = (int)(length % blksize);
159 	if (boff)
160 		loffset = length + (blksize - boff);
161 	else
162 		loffset = length;
163 	lsize = OFF_TO_IDX(round_page64(loffset));
164 
165 	if (object == NULL) {
166 		/*
167 		 * And an object of the appropriate size
168 		 */
169 		object = vm_object_allocate(OBJT_VNODE, lsize);
170 		object->flags = 0;
171 		object->handle = handle;
172 		vp->v_object = object;
173 		vp->v_filesize = length;
174 		if (vp->v_mount && (vp->v_mount->mnt_kern_flag & MNTK_NOMSYNC))
175 			object->flags |= OBJ_NOMSYNC;
176 	} else {
177 		object->ref_count++;
178 		if (object->size != lsize) {
179 			kprintf("vnode_pager_alloc: Warning, objsize "
180 				"mismatch %jd/%jd vp=%p obj=%p\n",
181 				(intmax_t)object->size,
182 				(intmax_t)lsize,
183 				vp, object);
184 		}
185 		if (vp->v_filesize != length) {
186 			kprintf("vnode_pager_alloc: Warning, filesize "
187 				"mismatch %jd/%jd vp=%p obj=%p\n",
188 				(intmax_t)vp->v_filesize,
189 				(intmax_t)length,
190 				vp, object);
191 		}
192 	}
193 	vref(vp);
194 
195 	vclrflags(vp, VOLOCK);
196 	if (vp->v_flag & VOWANT) {
197 		vclrflags(vp, VOWANT);
198 		wakeup(vp);
199 	}
200 	return (object);
201 }
202 
203 /*
204  * Add a ref to a vnode's existing VM object, return the object or
205  * NULL if the vnode did not have one.  This does not create the
206  * object (we can't since we don't know what the proper blocksize/boff
207  * is to match the VFS's use of the buffer cache).
208  */
209 vm_object_t
210 vnode_pager_reference(struct vnode *vp)
211 {
212 	vm_object_t object;
213 
214 	/*
215 	 * Prevent race condition when allocating the object. This
216 	 * can happen with NFS vnodes since the nfsnode isn't locked.
217 	 */
218 	while (vp->v_flag & VOLOCK) {
219 		vsetflags(vp, VOWANT);
220 		tsleep(vp, 0, "vnpobj", 0);
221 	}
222 	vsetflags(vp, VOLOCK);
223 
224 	/*
225 	 * Prevent race conditions against deallocation of the VM
226 	 * object.
227 	 */
228 	while (((object = vp->v_object) != NULL) &&
229 		(object->flags & OBJ_DEAD)) {
230 		vm_object_dead_sleep(object, "vadead");
231 	}
232 
233 	/*
234 	 * The object is expected to exist, the caller will handle
235 	 * NULL returns if it does not.
236 	 */
237 	if (object) {
238 		object->ref_count++;
239 		vref(vp);
240 	}
241 
242 	vclrflags(vp, VOLOCK);
243 	if (vp->v_flag & VOWANT) {
244 		vclrflags(vp, VOWANT);
245 		wakeup(vp);
246 	}
247 	return (object);
248 }
249 
250 static void
251 vnode_pager_dealloc(vm_object_t object)
252 {
253 	struct vnode *vp = object->handle;
254 
255 	if (vp == NULL)
256 		panic("vnode_pager_dealloc: pager already dealloced");
257 
258 	vm_object_pip_wait(object, "vnpdea");
259 
260 	object->handle = NULL;
261 	object->type = OBJT_DEAD;
262 	vp->v_object = NULL;
263 	vp->v_filesize = NOOFFSET;
264 	vclrflags(vp, VTEXT | VOBJBUF);
265 	swap_pager_freespace_all(object);
266 }
267 
268 /*
269  * Return whether the vnode pager has the requested page.  Return the
270  * number of disk-contiguous pages before and after the requested page,
271  * not including the requested page.
272  */
273 static boolean_t
274 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex)
275 {
276 	struct vnode *vp = object->handle;
277 	off_t loffset;
278 	off_t doffset;
279 	int voff;
280 	int bsize;
281 	int error;
282 
283 	/*
284 	 * If no vp or vp is doomed or marked transparent to VM, we do not
285 	 * have the page.
286 	 */
287 	if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
288 		return FALSE;
289 
290 	/*
291 	 * If filesystem no longer mounted or offset beyond end of file we do
292 	 * not have the page.
293 	 */
294 	loffset = IDX_TO_OFF(pindex);
295 
296 	if (vp->v_mount == NULL || loffset >= vp->v_filesize)
297 		return FALSE;
298 
299 	bsize = vp->v_mount->mnt_stat.f_iosize;
300 	voff = loffset % bsize;
301 
302 	/*
303 	 * XXX
304 	 *
305 	 * BMAP returns byte counts before and after, where after
306 	 * is inclusive of the base page.  haspage must return page
307 	 * counts before and after where after does not include the
308 	 * base page.
309 	 *
310 	 * BMAP is allowed to return a *after of 0 for backwards
311 	 * compatibility.  The base page is still considered valid if
312 	 * no error is returned.
313 	 */
314 	error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0);
315 	if (error)
316 		return TRUE;
317 	if (doffset == NOOFFSET)
318 		return FALSE;
319 	return TRUE;
320 }
321 
322 /*
323  * Lets the VM system know about a change in size for a file.
324  * We adjust our own internal size and flush any cached pages in
325  * the associated object that are affected by the size change.
326  *
327  * NOTE: This routine may be invoked as a result of a pager put
328  * operation (possibly at object termination time), so we must be careful.
329  *
330  * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
331  * we do not blow up on the case.  nsize will always be >= 0, however.
332  */
333 void
334 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
335 {
336 	vm_pindex_t nobjsize;
337 	vm_pindex_t oobjsize;
338 	vm_object_t object = vp->v_object;
339 
340 	if (object == NULL)
341 		return;
342 
343 	/*
344 	 * Hasn't changed size
345 	 */
346 	if (nsize == vp->v_filesize)
347 		return;
348 
349 	lwkt_gettoken(&vm_token);
350 
351 	/*
352 	 * Has changed size.  Adjust the VM object's size and v_filesize
353 	 * before we start scanning pages to prevent new pages from being
354 	 * allocated during the scan.
355 	 */
356 	nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
357 	oobjsize = object->size;
358 	object->size = nobjsize;
359 
360 	/*
361 	 * File has shrunk. Toss any cached pages beyond the new EOF.
362 	 */
363 	if (nsize < vp->v_filesize) {
364 		vp->v_filesize = nsize;
365 		if (nobjsize < oobjsize) {
366 			vm_object_page_remove(object, nobjsize, oobjsize,
367 					      FALSE);
368 		}
369 		/*
370 		 * This gets rid of garbage at the end of a page that is now
371 		 * only partially backed by the vnode.  Since we are setting
372 		 * the entire page valid & clean after we are done we have
373 		 * to be sure that the portion of the page within the file
374 		 * bounds is already valid.  If it isn't then making it
375 		 * valid would create a corrupt block.
376 		 */
377 		if (nsize & PAGE_MASK) {
378 			vm_offset_t kva;
379 			vm_page_t m;
380 
381 			do {
382 				m = vm_page_lookup(object, OFF_TO_IDX(nsize));
383 			} while (m && vm_page_sleep_busy(m, TRUE, "vsetsz"));
384 
385 			if (m && m->valid) {
386 				int base = (int)nsize & PAGE_MASK;
387 				int size = PAGE_SIZE - base;
388 				struct lwbuf *lwb;
389 
390 				/*
391 				 * Clear out partial-page garbage in case
392 				 * the page has been mapped.
393 				 *
394 				 * This is byte aligned.
395 				 */
396 				vm_page_busy(m);
397 				lwb = lwbuf_alloc(m);
398 				kva = lwbuf_kva(lwb);
399 				bzero((caddr_t)kva + base, size);
400 				lwbuf_free(lwb);
401 
402 				/*
403 				 * XXX work around SMP data integrity race
404 				 * by unmapping the page from user processes.
405 				 * The garbage we just cleared may be mapped
406 				 * to a user process running on another cpu
407 				 * and this code is not running through normal
408 				 * I/O channels which handle SMP issues for
409 				 * us, so unmap page to synchronize all cpus.
410 				 *
411 				 * XXX should vm_pager_unmap_page() have
412 				 * dealt with this?
413 				 */
414 				vm_page_protect(m, VM_PROT_NONE);
415 
416 				/*
417 				 * Clear out partial-page dirty bits.  This
418 				 * has the side effect of setting the valid
419 				 * bits, but that is ok.  There are a bunch
420 				 * of places in the VM system where we expected
421 				 * m->dirty == VM_PAGE_BITS_ALL.  The file EOF
422 				 * case is one of them.  If the page is still
423 				 * partially dirty, make it fully dirty.
424 				 *
425 				 * NOTE: We do not clear out the valid
426 				 * bits.  This would prevent bogus_page
427 				 * replacement from working properly.
428 				 *
429 				 * NOTE: We do not want to clear the dirty
430 				 * bit for a partial DEV_BSIZE'd truncation!
431 				 * This is DEV_BSIZE aligned!
432 				 */
433 				vm_page_clear_dirty_beg_nonincl(m, base, size);
434 				if (m->dirty != 0)
435 					m->dirty = VM_PAGE_BITS_ALL;
436 				vm_page_wakeup(m);
437 			}
438 		}
439 	} else {
440 		vp->v_filesize = nsize;
441 	}
442 	lwkt_reltoken(&vm_token);
443 }
444 
445 /*
446  * Release a page busied for a getpages operation.  The page may have become
447  * wired (typically due to being used by the buffer cache) or otherwise been
448  * soft-busied and cannot be freed in that case.  A held page can still be
449  * freed.
450  */
451 void
452 vnode_pager_freepage(vm_page_t m)
453 {
454 	if (m->busy || m->wire_count) {
455 		vm_page_activate(m);
456 		vm_page_wakeup(m);
457 	} else {
458 		vm_page_free(m);
459 	}
460 }
461 
462 /*
463  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
464  * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
465  * vnode_pager_generic_getpages() to implement the previous behaviour.
466  *
467  * All other FS's should use the bypass to get to the local media
468  * backing vp's VOP_GETPAGES.
469  */
470 static int
471 vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
472 {
473 	int rtval;
474 	struct vnode *vp;
475 
476 	vp = object->handle;
477 	rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess);
478 	if (rtval == EOPNOTSUPP)
479 		panic("vnode_pager: vfs's must implement vop_getpages\n");
480 	return rtval;
481 }
482 
483 /*
484  * This is now called from local media FS's to operate against their
485  * own vnodes if they fail to implement VOP_GETPAGES.
486  *
487  * With all the caching local media devices do these days there is really
488  * very little point to attempting to restrict the I/O size to contiguous
489  * blocks on-disk, especially if our caller thinks we need all the specified
490  * pages.  Just construct and issue a READ.
491  */
492 int
493 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount,
494 			     int reqpage, int seqaccess)
495 {
496 	struct iovec aiov;
497 	struct uio auio;
498 	off_t foff;
499 	int error;
500 	int count;
501 	int i;
502 	int ioflags;
503 
504 	/*
505 	 * Do not do anything if the vnode is bad.
506 	 */
507 	if (vp->v_mount == NULL)
508 		return VM_PAGER_BAD;
509 
510 	/*
511 	 * Calculate the number of pages.  Since we are paging in whole
512 	 * pages, adjust bytecount to be an integral multiple of the page
513 	 * size.  It will be clipped to the file EOF later on.
514 	 */
515 	bytecount = round_page(bytecount);
516 	count = bytecount / PAGE_SIZE;
517 
518 	/*
519 	 * We could check m[reqpage]->valid here and shortcut the operation,
520 	 * but doing so breaks read-ahead.  Instead assume that the VM
521 	 * system has already done at least the check, don't worry about
522 	 * any races, and issue the VOP_READ to allow read-ahead to function.
523 	 *
524 	 * This keeps the pipeline full for I/O bound sequentially scanned
525 	 * mmap()'s
526 	 */
527 	/* don't shortcut */
528 
529 	/*
530 	 * Discard pages past the file EOF.  If the requested page is past
531 	 * the file EOF we just leave its valid bits set to 0, the caller
532 	 * expects to maintain ownership of the requested page.  If the
533 	 * entire range is past file EOF discard everything and generate
534 	 * a pagein error.
535 	 */
536 	foff = IDX_TO_OFF(mpp[0]->pindex);
537 	if (foff >= vp->v_filesize) {
538 		for (i = 0; i < count; i++) {
539 			if (i != reqpage)
540 				vnode_pager_freepage(mpp[i]);
541 		}
542 		return VM_PAGER_ERROR;
543 	}
544 
545 	if (foff + bytecount > vp->v_filesize) {
546 		bytecount = vp->v_filesize - foff;
547 		i = round_page(bytecount) / PAGE_SIZE;
548 		while (count > i) {
549 			--count;
550 			if (count != reqpage)
551 				vnode_pager_freepage(mpp[count]);
552 		}
553 	}
554 
555 	/*
556 	 * The size of the transfer is bytecount.  bytecount will be an
557 	 * integral multiple of the page size unless it has been clipped
558 	 * to the file EOF.  The transfer cannot exceed the file EOF.
559 	 *
560 	 * When dealing with real devices we must round-up to the device
561 	 * sector size.
562 	 */
563 	if (vp->v_type == VBLK || vp->v_type == VCHR) {
564 		int secmask = vp->v_rdev->si_bsize_phys - 1;
565 		KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
566 		bytecount = (bytecount + secmask) & ~secmask;
567 	}
568 
569 	/*
570 	 * Severe hack to avoid deadlocks with the buffer cache
571 	 */
572 	for (i = 0; i < count; ++i) {
573 		vm_page_t mt = mpp[i];
574 
575 		vm_page_io_start(mt);
576 		vm_page_wakeup(mt);
577 	}
578 
579 	/*
580 	 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE
581 	 */
582 	ioflags = IO_VMIO;
583 	if (seqaccess)
584 		ioflags |= IO_SEQMAX << IO_SEQSHIFT;
585 
586 	aiov.iov_base = NULL;
587 	aiov.iov_len = bytecount;
588 	auio.uio_iov = &aiov;
589 	auio.uio_iovcnt = 1;
590 	auio.uio_offset = foff;
591 	auio.uio_segflg = UIO_NOCOPY;
592 	auio.uio_rw = UIO_READ;
593 	auio.uio_resid = bytecount;
594 	auio.uio_td = NULL;
595 	mycpu->gd_cnt.v_vnodein++;
596 	mycpu->gd_cnt.v_vnodepgsin += count;
597 
598 	error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
599 
600 	/*
601 	 * Severe hack to avoid deadlocks with the buffer cache
602 	 */
603 	lwkt_gettoken(&vm_token);
604 	for (i = 0; i < count; ++i) {
605 		vm_page_t mt = mpp[i];
606 
607 		while (vm_page_sleep_busy(mt, FALSE, "getpgs"))
608 			;
609 		vm_page_busy(mt);
610 		vm_page_io_finish(mt);
611 	}
612 	lwkt_reltoken(&vm_token);
613 
614 	/*
615 	 * Calculate the actual number of bytes read and clean up the
616 	 * page list.
617 	 */
618 	bytecount -= auio.uio_resid;
619 
620 	for (i = 0; i < count; ++i) {
621 		vm_page_t mt = mpp[i];
622 
623 		if (i != reqpage) {
624 			if (error == 0 && mt->valid) {
625 				if (mt->flags & PG_WANTED)
626 					vm_page_activate(mt);
627 				else
628 					vm_page_deactivate(mt);
629 				vm_page_wakeup(mt);
630 			} else {
631 				vnode_pager_freepage(mt);
632 			}
633 		} else if (mt->valid == 0) {
634 			if (error == 0) {
635 				kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt, mt->object, (int) mt->pindex);
636 				/* whoops, something happened */
637 				error = EINVAL;
638 			}
639 		} else if (mt->valid != VM_PAGE_BITS_ALL) {
640 			/*
641 			 * Zero-extend the requested page if necessary (if
642 			 * the filesystem is using a small block size).
643 			 */
644 			vm_page_zero_invalid(mt, TRUE);
645 		}
646 	}
647 	if (error) {
648 		kprintf("vnode_pager_getpage: I/O read error\n");
649 	}
650 	return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
651 }
652 
653 /*
654  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
655  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
656  * vnode_pager_generic_putpages() to implement the previous behaviour.
657  *
658  * Caller has already cleared the pmap modified bits, if any.
659  *
660  * All other FS's should use the bypass to get to the local media
661  * backing vp's VOP_PUTPAGES.
662  */
663 static void
664 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
665 		     boolean_t sync, int *rtvals)
666 {
667 	int rtval;
668 	struct vnode *vp;
669 	int bytes = count * PAGE_SIZE;
670 
671 	/*
672 	 * Force synchronous operation if we are extremely low on memory
673 	 * to prevent a low-memory deadlock.  VOP operations often need to
674 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
675 	 * operation ).  The swapper handles the case by limiting the amount
676 	 * of asynchronous I/O, but that sort of solution doesn't scale well
677 	 * for the vnode pager without a lot of work.
678 	 *
679 	 * Also, the backing vnode's iodone routine may not wake the pageout
680 	 * daemon up.  This should be probably be addressed XXX.
681 	 */
682 
683 	if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min)
684 		sync |= OBJPC_SYNC;
685 
686 	/*
687 	 * Call device-specific putpages function
688 	 */
689 	vp = object->handle;
690 	rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
691 	if (rtval == EOPNOTSUPP) {
692 	    kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
693 	    rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
694 	}
695 }
696 
697 
698 /*
699  * This is now called from local media FS's to operate against their
700  * own vnodes if they fail to implement VOP_PUTPAGES.
701  *
702  * This is typically called indirectly via the pageout daemon and
703  * clustering has already typically occured, so in general we ask the
704  * underlying filesystem to write the data out asynchronously rather
705  * then delayed.
706  */
707 int
708 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
709 			     int flags, int *rtvals)
710 {
711 	int i;
712 	vm_object_t object;
713 	int maxsize, ncount, count;
714 	vm_ooffset_t poffset;
715 	struct uio auio;
716 	struct iovec aiov;
717 	int error;
718 	int ioflags;
719 
720 	object = vp->v_object;
721 	count = bytecount / PAGE_SIZE;
722 
723 	for (i = 0; i < count; i++)
724 		rtvals[i] = VM_PAGER_AGAIN;
725 
726 	if ((int) m[0]->pindex < 0) {
727 		kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
728 			(long)m[0]->pindex, m[0]->dirty);
729 		rtvals[0] = VM_PAGER_BAD;
730 		return VM_PAGER_BAD;
731 	}
732 
733 	maxsize = count * PAGE_SIZE;
734 	ncount = count;
735 
736 	poffset = IDX_TO_OFF(m[0]->pindex);
737 
738 	/*
739 	 * If the page-aligned write is larger then the actual file we
740 	 * have to invalidate pages occuring beyond the file EOF.
741 	 *
742 	 * If the file EOF resides in the middle of a page we still clear
743 	 * all of that page's dirty bits later on.  If we didn't it would
744 	 * endlessly re-write.
745 	 *
746 	 * We do not under any circumstances truncate the valid bits, as
747 	 * this will screw up bogus page replacement.
748 	 *
749 	 * The caller has already read-protected the pages.  The VFS must
750 	 * use the buffer cache to wrap the pages.  The pages might not
751 	 * be immediately flushed by the buffer cache but once under its
752 	 * control the pages themselves can wind up being marked clean
753 	 * and their covering buffer cache buffer can be marked dirty.
754 	 */
755 	if (poffset + maxsize > vp->v_filesize) {
756 		if (poffset < vp->v_filesize) {
757 			maxsize = vp->v_filesize - poffset;
758 			ncount = btoc(maxsize);
759 		} else {
760 			maxsize = 0;
761 			ncount = 0;
762 		}
763 		if (ncount < count) {
764 			for (i = ncount; i < count; i++) {
765 				rtvals[i] = VM_PAGER_BAD;
766 			}
767 		}
768 	}
769 
770 	/*
771 	 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
772 	 * rather then a bdwrite() to prevent paging I/O from saturating
773 	 * the buffer cache.  Dummy-up the sequential heuristic to cause
774 	 * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
775 	 * the system decides how to cluster.
776 	 */
777 	ioflags = IO_VMIO;
778 	if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
779 		ioflags |= IO_SYNC;
780 	else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
781 		ioflags |= IO_ASYNC;
782 	ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
783 	ioflags |= IO_SEQMAX << IO_SEQSHIFT;
784 
785 	aiov.iov_base = (caddr_t) 0;
786 	aiov.iov_len = maxsize;
787 	auio.uio_iov = &aiov;
788 	auio.uio_iovcnt = 1;
789 	auio.uio_offset = poffset;
790 	auio.uio_segflg = UIO_NOCOPY;
791 	auio.uio_rw = UIO_WRITE;
792 	auio.uio_resid = maxsize;
793 	auio.uio_td = NULL;
794 	error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
795 	mycpu->gd_cnt.v_vnodeout++;
796 	mycpu->gd_cnt.v_vnodepgsout += ncount;
797 
798 	if (error) {
799 		krateprintf(&vbadrate,
800 			    "vnode_pager_putpages: I/O error %d\n", error);
801 	}
802 	if (auio.uio_resid) {
803 		krateprintf(&vresrate,
804 			    "vnode_pager_putpages: residual I/O %zd at %lu\n",
805 			    auio.uio_resid, (u_long)m[0]->pindex);
806 	}
807 	if (error == 0) {
808 		for (i = 0; i < ncount; i++) {
809 			rtvals[i] = VM_PAGER_OK;
810 			vm_page_undirty(m[i]);
811 		}
812 	}
813 	return rtvals[0];
814 }
815 
816 struct vnode *
817 vnode_pager_lock(vm_object_t object)
818 {
819 	struct thread *td = curthread;	/* XXX */
820 	int error;
821 
822 	for (; object != NULL; object = object->backing_object) {
823 		if (object->type != OBJT_VNODE)
824 			continue;
825 		if (object->flags & OBJ_DEAD)
826 			return NULL;
827 
828 		for (;;) {
829 			struct vnode *vp = object->handle;
830 			error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
831 			if (error == 0) {
832 				if (object->handle != vp) {
833 					vput(vp);
834 					continue;
835 				}
836 				return (vp);
837 			}
838 			if ((object->flags & OBJ_DEAD) ||
839 			    (object->type != OBJT_VNODE)) {
840 				return NULL;
841 			}
842 			kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp, error, lockstatus(&vp->v_lock, td));
843 			tsleep(object->handle, 0, "vnpgrl", hz);
844 		}
845 	}
846 	return NULL;
847 }
848