xref: /dflybsd-src/sys/vm/vnode_pager.c (revision c9e3d8f96688a159959b1af2d4fef14b744173e3)
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 	/*
350 	 * Has changed size.  Adjust the VM object's size and v_filesize
351 	 * before we start scanning pages to prevent new pages from being
352 	 * allocated during the scan.
353 	 */
354 	nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
355 	oobjsize = object->size;
356 	object->size = nobjsize;
357 
358 	/*
359 	 * File has shrunk. Toss any cached pages beyond the new EOF.
360 	 */
361 	if (nsize < vp->v_filesize) {
362 		vp->v_filesize = nsize;
363 		if (nobjsize < oobjsize) {
364 			vm_object_page_remove(object, nobjsize, oobjsize,
365 					      FALSE);
366 		}
367 		/*
368 		 * This gets rid of garbage at the end of a page that is now
369 		 * only partially backed by the vnode.  Since we are setting
370 		 * the entire page valid & clean after we are done we have
371 		 * to be sure that the portion of the page within the file
372 		 * bounds is already valid.  If it isn't then making it
373 		 * valid would create a corrupt block.
374 		 */
375 		if (nsize & PAGE_MASK) {
376 			vm_offset_t kva;
377 			vm_page_t m;
378 
379 			do {
380 				m = vm_page_lookup(object, OFF_TO_IDX(nsize));
381 			} while (m && vm_page_sleep_busy(m, TRUE, "vsetsz"));
382 
383 			if (m && m->valid) {
384 				int base = (int)nsize & PAGE_MASK;
385 				int size = PAGE_SIZE - base;
386 				struct lwbuf *lwb;
387 
388 				/*
389 				 * Clear out partial-page garbage in case
390 				 * the page has been mapped.
391 				 *
392 				 * This is byte aligned.
393 				 */
394 				vm_page_busy(m);
395 				lwb = lwbuf_alloc(m);
396 				kva = lwbuf_kva(lwb);
397 				bzero((caddr_t)kva + base, size);
398 				lwbuf_free(lwb);
399 
400 				/*
401 				 * XXX work around SMP data integrity race
402 				 * by unmapping the page from user processes.
403 				 * The garbage we just cleared may be mapped
404 				 * to a user process running on another cpu
405 				 * and this code is not running through normal
406 				 * I/O channels which handle SMP issues for
407 				 * us, so unmap page to synchronize all cpus.
408 				 *
409 				 * XXX should vm_pager_unmap_page() have
410 				 * dealt with this?
411 				 */
412 				vm_page_protect(m, VM_PROT_NONE);
413 
414 				/*
415 				 * Clear out partial-page dirty bits.  This
416 				 * has the side effect of setting the valid
417 				 * bits, but that is ok.  There are a bunch
418 				 * of places in the VM system where we expected
419 				 * m->dirty == VM_PAGE_BITS_ALL.  The file EOF
420 				 * case is one of them.  If the page is still
421 				 * partially dirty, make it fully dirty.
422 				 *
423 				 * NOTE: We do not clear out the valid
424 				 * bits.  This would prevent bogus_page
425 				 * replacement from working properly.
426 				 *
427 				 * NOTE: We do not want to clear the dirty
428 				 * bit for a partial DEV_BSIZE'd truncation!
429 				 * This is DEV_BSIZE aligned!
430 				 */
431 				vm_page_clear_dirty_beg_nonincl(m, base, size);
432 				if (m->dirty != 0)
433 					m->dirty = VM_PAGE_BITS_ALL;
434 				vm_page_wakeup(m);
435 			}
436 		}
437 	} else {
438 		vp->v_filesize = nsize;
439 	}
440 }
441 
442 /*
443  * Release a page busied for a getpages operation.  The page may have become
444  * wired (typically due to being used by the buffer cache) or otherwise been
445  * soft-busied and cannot be freed in that case.  A held page can still be
446  * freed.
447  */
448 void
449 vnode_pager_freepage(vm_page_t m)
450 {
451 	if (m->busy || m->wire_count) {
452 		vm_page_activate(m);
453 		vm_page_wakeup(m);
454 	} else {
455 		vm_page_free(m);
456 	}
457 }
458 
459 /*
460  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
461  * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
462  * vnode_pager_generic_getpages() to implement the previous behaviour.
463  *
464  * All other FS's should use the bypass to get to the local media
465  * backing vp's VOP_GETPAGES.
466  */
467 static int
468 vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
469 {
470 	int rtval;
471 	struct vnode *vp;
472 
473 	vp = object->handle;
474 	rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess);
475 	if (rtval == EOPNOTSUPP)
476 		panic("vnode_pager: vfs's must implement vop_getpages\n");
477 	return rtval;
478 }
479 
480 /*
481  * This is now called from local media FS's to operate against their
482  * own vnodes if they fail to implement VOP_GETPAGES.
483  *
484  * With all the caching local media devices do these days there is really
485  * very little point to attempting to restrict the I/O size to contiguous
486  * blocks on-disk, especially if our caller thinks we need all the specified
487  * pages.  Just construct and issue a READ.
488  */
489 int
490 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount,
491 			     int reqpage, int seqaccess)
492 {
493 	struct iovec aiov;
494 	struct uio auio;
495 	off_t foff;
496 	int error;
497 	int count;
498 	int i;
499 	int ioflags;
500 
501 	/*
502 	 * Do not do anything if the vnode is bad.
503 	 */
504 	if (vp->v_mount == NULL)
505 		return VM_PAGER_BAD;
506 
507 	/*
508 	 * Calculate the number of pages.  Since we are paging in whole
509 	 * pages, adjust bytecount to be an integral multiple of the page
510 	 * size.  It will be clipped to the file EOF later on.
511 	 */
512 	bytecount = round_page(bytecount);
513 	count = bytecount / PAGE_SIZE;
514 
515 	/*
516 	 * We could check m[reqpage]->valid here and shortcut the operation,
517 	 * but doing so breaks read-ahead.  Instead assume that the VM
518 	 * system has already done at least the check, don't worry about
519 	 * any races, and issue the VOP_READ to allow read-ahead to function.
520 	 *
521 	 * This keeps the pipeline full for I/O bound sequentially scanned
522 	 * mmap()'s
523 	 */
524 	/* don't shortcut */
525 
526 	/*
527 	 * Discard pages past the file EOF.  If the requested page is past
528 	 * the file EOF we just leave its valid bits set to 0, the caller
529 	 * expects to maintain ownership of the requested page.  If the
530 	 * entire range is past file EOF discard everything and generate
531 	 * a pagein error.
532 	 */
533 	foff = IDX_TO_OFF(mpp[0]->pindex);
534 	if (foff >= vp->v_filesize) {
535 		for (i = 0; i < count; i++) {
536 			if (i != reqpage)
537 				vnode_pager_freepage(mpp[i]);
538 		}
539 		return VM_PAGER_ERROR;
540 	}
541 
542 	if (foff + bytecount > vp->v_filesize) {
543 		bytecount = vp->v_filesize - foff;
544 		i = round_page(bytecount) / PAGE_SIZE;
545 		while (count > i) {
546 			--count;
547 			if (count != reqpage)
548 				vnode_pager_freepage(mpp[count]);
549 		}
550 	}
551 
552 	/*
553 	 * The size of the transfer is bytecount.  bytecount will be an
554 	 * integral multiple of the page size unless it has been clipped
555 	 * to the file EOF.  The transfer cannot exceed the file EOF.
556 	 *
557 	 * When dealing with real devices we must round-up to the device
558 	 * sector size.
559 	 */
560 	if (vp->v_type == VBLK || vp->v_type == VCHR) {
561 		int secmask = vp->v_rdev->si_bsize_phys - 1;
562 		KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
563 		bytecount = (bytecount + secmask) & ~secmask;
564 	}
565 
566 	/*
567 	 * Severe hack to avoid deadlocks with the buffer cache
568 	 */
569 	for (i = 0; i < count; ++i) {
570 		vm_page_t mt = mpp[i];
571 
572 		vm_page_io_start(mt);
573 		vm_page_wakeup(mt);
574 	}
575 
576 	/*
577 	 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE
578 	 */
579 	ioflags = IO_VMIO;
580 	if (seqaccess)
581 		ioflags |= IO_SEQMAX << IO_SEQSHIFT;
582 
583 	aiov.iov_base = NULL;
584 	aiov.iov_len = bytecount;
585 	auio.uio_iov = &aiov;
586 	auio.uio_iovcnt = 1;
587 	auio.uio_offset = foff;
588 	auio.uio_segflg = UIO_NOCOPY;
589 	auio.uio_rw = UIO_READ;
590 	auio.uio_resid = bytecount;
591 	auio.uio_td = NULL;
592 	mycpu->gd_cnt.v_vnodein++;
593 	mycpu->gd_cnt.v_vnodepgsin += count;
594 
595 	error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
596 
597 	/*
598 	 * Severe hack to avoid deadlocks with the buffer cache
599 	 */
600 	for (i = 0; i < count; ++i) {
601 		vm_page_t mt = mpp[i];
602 
603 		while (vm_page_sleep_busy(mt, FALSE, "getpgs"))
604 			;
605 		vm_page_busy(mt);
606 		vm_page_io_finish(mt);
607 	}
608 
609 	/*
610 	 * Calculate the actual number of bytes read and clean up the
611 	 * page list.
612 	 */
613 	bytecount -= auio.uio_resid;
614 
615 	for (i = 0; i < count; ++i) {
616 		vm_page_t mt = mpp[i];
617 
618 		if (i != reqpage) {
619 			if (error == 0 && mt->valid) {
620 				if (mt->flags & PG_WANTED)
621 					vm_page_activate(mt);
622 				else
623 					vm_page_deactivate(mt);
624 				vm_page_wakeup(mt);
625 			} else {
626 				vnode_pager_freepage(mt);
627 			}
628 		} else if (mt->valid == 0) {
629 			if (error == 0) {
630 				kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt, mt->object, (int) mt->pindex);
631 				/* whoops, something happened */
632 				error = EINVAL;
633 			}
634 		} else if (mt->valid != VM_PAGE_BITS_ALL) {
635 			/*
636 			 * Zero-extend the requested page if necessary (if
637 			 * the filesystem is using a small block size).
638 			 */
639 			vm_page_zero_invalid(mt, TRUE);
640 		}
641 	}
642 	if (error) {
643 		kprintf("vnode_pager_getpage: I/O read error\n");
644 	}
645 	return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
646 }
647 
648 /*
649  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
650  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
651  * vnode_pager_generic_putpages() to implement the previous behaviour.
652  *
653  * Caller has already cleared the pmap modified bits, if any.
654  *
655  * All other FS's should use the bypass to get to the local media
656  * backing vp's VOP_PUTPAGES.
657  */
658 static void
659 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
660 		     boolean_t sync, int *rtvals)
661 {
662 	int rtval;
663 	struct vnode *vp;
664 	int bytes = count * PAGE_SIZE;
665 
666 	/*
667 	 * Force synchronous operation if we are extremely low on memory
668 	 * to prevent a low-memory deadlock.  VOP operations often need to
669 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
670 	 * operation ).  The swapper handles the case by limiting the amount
671 	 * of asynchronous I/O, but that sort of solution doesn't scale well
672 	 * for the vnode pager without a lot of work.
673 	 *
674 	 * Also, the backing vnode's iodone routine may not wake the pageout
675 	 * daemon up.  This should be probably be addressed XXX.
676 	 */
677 
678 	if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min)
679 		sync |= OBJPC_SYNC;
680 
681 	/*
682 	 * Call device-specific putpages function
683 	 */
684 	vp = object->handle;
685 	rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
686 	if (rtval == EOPNOTSUPP) {
687 	    kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
688 	    rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
689 	}
690 }
691 
692 
693 /*
694  * This is now called from local media FS's to operate against their
695  * own vnodes if they fail to implement VOP_PUTPAGES.
696  *
697  * This is typically called indirectly via the pageout daemon and
698  * clustering has already typically occured, so in general we ask the
699  * underlying filesystem to write the data out asynchronously rather
700  * then delayed.
701  */
702 int
703 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
704 			     int flags, int *rtvals)
705 {
706 	int i;
707 	vm_object_t object;
708 	int maxsize, ncount, count;
709 	vm_ooffset_t poffset;
710 	struct uio auio;
711 	struct iovec aiov;
712 	int error;
713 	int ioflags;
714 
715 	object = vp->v_object;
716 	count = bytecount / PAGE_SIZE;
717 
718 	for (i = 0; i < count; i++)
719 		rtvals[i] = VM_PAGER_AGAIN;
720 
721 	if ((int) m[0]->pindex < 0) {
722 		kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
723 			(long)m[0]->pindex, m[0]->dirty);
724 		rtvals[0] = VM_PAGER_BAD;
725 		return VM_PAGER_BAD;
726 	}
727 
728 	maxsize = count * PAGE_SIZE;
729 	ncount = count;
730 
731 	poffset = IDX_TO_OFF(m[0]->pindex);
732 
733 	/*
734 	 * If the page-aligned write is larger then the actual file we
735 	 * have to invalidate pages occuring beyond the file EOF.
736 	 *
737 	 * If the file EOF resides in the middle of a page we still clear
738 	 * all of that page's dirty bits later on.  If we didn't it would
739 	 * endlessly re-write.
740 	 *
741 	 * We do not under any circumstances truncate the valid bits, as
742 	 * this will screw up bogus page replacement.
743 	 *
744 	 * The caller has already read-protected the pages.  The VFS must
745 	 * use the buffer cache to wrap the pages.  The pages might not
746 	 * be immediately flushed by the buffer cache but once under its
747 	 * control the pages themselves can wind up being marked clean
748 	 * and their covering buffer cache buffer can be marked dirty.
749 	 */
750 	if (poffset + maxsize > vp->v_filesize) {
751 		if (poffset < vp->v_filesize) {
752 			maxsize = vp->v_filesize - poffset;
753 			ncount = btoc(maxsize);
754 		} else {
755 			maxsize = 0;
756 			ncount = 0;
757 		}
758 		if (ncount < count) {
759 			for (i = ncount; i < count; i++) {
760 				rtvals[i] = VM_PAGER_BAD;
761 			}
762 		}
763 	}
764 
765 	/*
766 	 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
767 	 * rather then a bdwrite() to prevent paging I/O from saturating
768 	 * the buffer cache.  Dummy-up the sequential heuristic to cause
769 	 * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
770 	 * the system decides how to cluster.
771 	 */
772 	ioflags = IO_VMIO;
773 	if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
774 		ioflags |= IO_SYNC;
775 	else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
776 		ioflags |= IO_ASYNC;
777 	ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
778 	ioflags |= IO_SEQMAX << IO_SEQSHIFT;
779 
780 	aiov.iov_base = (caddr_t) 0;
781 	aiov.iov_len = maxsize;
782 	auio.uio_iov = &aiov;
783 	auio.uio_iovcnt = 1;
784 	auio.uio_offset = poffset;
785 	auio.uio_segflg = UIO_NOCOPY;
786 	auio.uio_rw = UIO_WRITE;
787 	auio.uio_resid = maxsize;
788 	auio.uio_td = NULL;
789 	error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
790 	mycpu->gd_cnt.v_vnodeout++;
791 	mycpu->gd_cnt.v_vnodepgsout += ncount;
792 
793 	if (error) {
794 		krateprintf(&vbadrate,
795 			    "vnode_pager_putpages: I/O error %d\n", error);
796 	}
797 	if (auio.uio_resid) {
798 		krateprintf(&vresrate,
799 			    "vnode_pager_putpages: residual I/O %zd at %lu\n",
800 			    auio.uio_resid, (u_long)m[0]->pindex);
801 	}
802 	if (error == 0) {
803 		for (i = 0; i < ncount; i++) {
804 			rtvals[i] = VM_PAGER_OK;
805 			vm_page_undirty(m[i]);
806 		}
807 	}
808 	return rtvals[0];
809 }
810 
811 struct vnode *
812 vnode_pager_lock(vm_object_t object)
813 {
814 	struct thread *td = curthread;	/* XXX */
815 	int error;
816 
817 	for (; object != NULL; object = object->backing_object) {
818 		if (object->type != OBJT_VNODE)
819 			continue;
820 		if (object->flags & OBJ_DEAD)
821 			return NULL;
822 
823 		for (;;) {
824 			struct vnode *vp = object->handle;
825 			error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
826 			if (error == 0) {
827 				if (object->handle != vp) {
828 					vput(vp);
829 					continue;
830 				}
831 				return (vp);
832 			}
833 			if ((object->flags & OBJ_DEAD) ||
834 			    (object->type != OBJT_VNODE)) {
835 				return NULL;
836 			}
837 			kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp, error, lockstatus(&vp->v_lock, td));
838 			tsleep(object->handle, 0, "vnpgrl", hz);
839 		}
840 	}
841 	return NULL;
842 }
843