xref: /dragonfly/sys/vm/vnode_pager.c (revision 2d396014)
1 /*
2  * (MPSAFE)
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. Neither the name of the University nor the names of its contributors
23  *    may be used to endorse or promote products derived from this software
24  *    without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	from: @(#)vnode_pager.c	7.5 (Berkeley) 4/20/91
39  * $FreeBSD: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $
40  */
41 
42 /*
43  * Page to/from files (vnodes).
44  */
45 
46 /*
47  * TODO:
48  *	Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
49  *	greatly re-simplify the vnode_pager.
50  */
51 
52 #include <sys/param.h>
53 #include <sys/systm.h>
54 #include <sys/uio.h>
55 #include <sys/kernel.h>
56 #include <sys/proc.h>
57 #include <sys/vnode.h>
58 #include <sys/mount.h>
59 #include <sys/buf.h>
60 #include <sys/vmmeter.h>
61 #include <sys/conf.h>
62 
63 #include <cpu/lwbuf.h>
64 
65 #include <vm/vm.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_pager.h>
69 #include <vm/vm_map.h>
70 #include <vm/vnode_pager.h>
71 #include <vm/swap_pager.h>
72 #include <vm/vm_extern.h>
73 
74 #include <vm/vm_page2.h>
75 
76 static	pgo_dealloc_t		vnode_pager_dealloc;
77 static	pgo_getpage_t		vnode_pager_getpage;
78 static	pgo_putpages_t		vnode_pager_putpages;
79 static	pgo_haspage_t		vnode_pager_haspage;
80 
81 struct pagerops vnodepagerops = {
82 	.pgo_dealloc =		vnode_pager_dealloc,
83 	.pgo_getpage =		vnode_pager_getpage,
84 	.pgo_putpages =		vnode_pager_putpages,
85 	.pgo_haspage =		vnode_pager_haspage
86 };
87 
88 static struct krate vbadrate = { 1 };
89 static struct krate vresrate = { 1 };
90 
91 long vnode_pbuf_freecnt = -1;	/* start out unlimited */
92 
93 /*
94  * Allocate a VM object for a vnode, typically a regular file vnode.
95  *
96  * Some additional information is required to generate a properly sized
97  * object which covers the entire buffer cache buffer straddling the file
98  * EOF.  Userland does not see the extra pages as the VM fault code tests
99  * against v_filesize.
100  */
101 vm_object_t
vnode_pager_alloc(void * handle,off_t length,vm_prot_t prot,off_t offset,int blksize,int boff)102 vnode_pager_alloc(void *handle, off_t length, vm_prot_t prot, off_t offset,
103 		  int blksize, int boff)
104 {
105 	vm_object_t object;
106 	struct vnode *vp;
107 	off_t loffset;
108 	vm_pindex_t lsize;
109 
110 	/*
111 	 * Pageout to vnode, no can do yet.
112 	 */
113 	if (handle == NULL)
114 		return (NULL);
115 
116 	/*
117 	 * XXX hack - This initialization should be put somewhere else.
118 	 */
119 	if (vnode_pbuf_freecnt < 0) {
120 	    vnode_pbuf_freecnt = nswbuf_kva / 2 + 1;
121 	}
122 
123 	/*
124 	 * Serialize potential vnode/object teardowns and interlocks
125 	 */
126 	vp = (struct vnode *)handle;
127 	lwkt_gettoken(&vp->v_token);
128 
129 	/*
130 	 * If the object is being terminated, wait for it to
131 	 * go away.
132 	 */
133 	object = vp->v_object;
134 	if (object) {
135 		vm_object_hold(object);
136 		KKASSERT((object->flags & OBJ_DEAD) == 0);
137 	}
138 
139 	if (VREFCNT(vp) <= 0)
140 		panic("vnode_pager_alloc: no vnode reference");
141 
142 	/*
143 	 * Round up to the *next* block, then destroy the buffers in question.
144 	 * Since we are only removing some of the buffers we must rely on the
145 	 * scan count to determine whether a loop is necessary.
146 	 *
147 	 * Destroy any pages beyond the last buffer.
148 	 */
149 	if (boff < 0)
150 		boff = (int)(length % blksize);
151 	if (boff)
152 		loffset = length + (blksize - boff);
153 	else
154 		loffset = length;
155 	lsize = OFF_TO_IDX(round_page64(loffset));
156 
157 	if (object == NULL) {
158 		/*
159 		 * And an object of the appropriate size
160 		 */
161 		object = vm_object_allocate_hold(OBJT_VNODE, lsize);
162 		object->handle = handle;
163 		vp->v_object = object;
164 		vp->v_filesize = length;
165 		if (vp->v_mount && (vp->v_mount->mnt_kern_flag & MNTK_NOMSYNC))
166 			vm_object_set_flag(object, OBJ_NOMSYNC);
167 		vref(vp);
168 	} else {
169 		vm_object_reference_quick(object);	/* also vref's */
170 		if (object->size != lsize) {
171 			kprintf("vnode_pager_alloc: Warning, objsize "
172 				"mismatch %jd/%jd vp=%p obj=%p\n",
173 				(intmax_t)object->size,
174 				(intmax_t)lsize,
175 				vp, object);
176 		}
177 		if (vp->v_filesize != length) {
178 			kprintf("vnode_pager_alloc: Warning, filesize "
179 				"mismatch %jd/%jd vp=%p obj=%p\n",
180 				(intmax_t)vp->v_filesize,
181 				(intmax_t)length,
182 				vp, object);
183 		}
184 	}
185 	vm_object_drop(object);
186 	lwkt_reltoken(&vp->v_token);
187 
188 	return (object);
189 }
190 
191 /*
192  * Add a ref to a vnode's existing VM object, return the object or
193  * NULL if the vnode did not have one.  This does not create the
194  * object (we can't since we don't know what the proper blocksize/boff
195  * is to match the VFS's use of the buffer cache).
196  *
197  * The vnode must be referenced and is typically open.  The object should
198  * be stable in this situation.
199  *
200  * Returns the object with an additional reference but not locked.
201  */
202 vm_object_t
vnode_pager_reference(struct vnode * vp)203 vnode_pager_reference(struct vnode *vp)
204 {
205 	vm_object_t object;
206 
207 	if ((object = vp->v_object) != NULL)
208 		vm_object_reference_quick(object); /* also vref's vnode */
209 	return (object);
210 }
211 
212 static void
vnode_pager_dealloc(vm_object_t object)213 vnode_pager_dealloc(vm_object_t object)
214 {
215 	struct vnode *vp = object->handle;
216 
217 	if (vp == NULL)
218 		panic("vnode_pager_dealloc: pager already dealloced");
219 
220 	vm_object_pip_wait(object, "vnpdea");
221 
222 	object->handle = NULL;
223 	object->type = OBJT_DEAD;
224 	vp->v_object = NULL;
225 	vp->v_filesize = NOOFFSET;
226 	vclrflags(vp, VTEXT | VOBJBUF);
227 	swap_pager_freespace_all(object);
228 }
229 
230 /*
231  * Return whether the vnode pager has the requested page.
232  */
233 static boolean_t
vnode_pager_haspage(vm_object_t object,vm_pindex_t pindex)234 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex)
235 {
236 	struct vnode *vp = object->handle;
237 	off_t loffset;
238 	off_t doffset;
239 	int voff;
240 	int bsize;
241 	int error;
242 
243 	/*
244 	 * If no vp or vp is doomed or marked transparent to VM, we do not
245 	 * have the page.
246 	 */
247 	if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
248 		return FALSE;
249 
250 	/*
251 	 * If filesystem no longer mounted or offset beyond end of file we do
252 	 * not have the page.
253 	 */
254 	loffset = IDX_TO_OFF(pindex);
255 
256 	if (vp->v_mount == NULL || loffset >= vp->v_filesize)
257 		return FALSE;
258 
259 	bsize = vp->v_mount->mnt_stat.f_iosize;
260 	voff = loffset % bsize;
261 
262 	/*
263 	 * XXX (obsolete - before and after pointers are now NULL)
264 	 *
265 	 * BMAP returns byte counts before and after, where after
266 	 * is inclusive of the base page.  haspage must return page
267 	 * counts before and after where after does not include the
268 	 * base page.
269 	 *
270 	 * BMAP is allowed to return a *after of 0 for backwards
271 	 * compatibility.  The base page is still considered valid if
272 	 * no error is returned.
273 	 */
274 	error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0);
275 	if (error)
276 		return TRUE;
277 	if (doffset == NOOFFSET)
278 		return FALSE;
279 	return TRUE;
280 }
281 
282 /*
283  * Lets the VM system know about a change in size for a file.
284  * We adjust our own internal size and flush any cached pages in
285  * the associated object that are affected by the size change.
286  *
287  * NOTE: This routine may be invoked as a result of a pager put
288  * operation (possibly at object termination time), so we must be careful.
289  *
290  * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
291  * we do not blow up on the case.  nsize will always be >= 0, however.
292  */
293 void
vnode_pager_setsize(struct vnode * vp,vm_ooffset_t nsize)294 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
295 {
296 	vm_pindex_t nobjsize;
297 	vm_pindex_t oobjsize;
298 	vm_object_t object;
299 
300 	object = vp->v_object;
301 	if (object == NULL)
302 		return;
303 	vm_object_hold(object);
304 	KKASSERT(vp->v_object == object);
305 
306 	/*
307 	 * Hasn't changed size
308 	 */
309 	if (nsize == vp->v_filesize) {
310 		vm_object_drop(object);
311 		return;
312 	}
313 
314 	/*
315 	 * Has changed size.  Adjust the VM object's size and v_filesize
316 	 * before we start scanning pages to prevent new pages from being
317 	 * allocated during the scan.
318 	 */
319 	nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
320 	oobjsize = object->size;
321 	object->size = nobjsize;
322 
323 	/*
324 	 * File has shrunk. Toss any cached pages beyond the new EOF.
325 	 */
326 	if (nsize < vp->v_filesize) {
327 		vp->v_filesize = nsize;
328 		if (nobjsize < oobjsize) {
329 			vm_object_page_remove(object, nobjsize, oobjsize,
330 					      FALSE);
331 		}
332 		/*
333 		 * This gets rid of garbage at the end of a page that is now
334 		 * only partially backed by the vnode.  Since we are setting
335 		 * the entire page valid & clean after we are done we have
336 		 * to be sure that the portion of the page within the file
337 		 * bounds is already valid.  If it isn't then making it
338 		 * valid would create a corrupt block.
339 		 */
340 		if (nsize & PAGE_MASK) {
341 			vm_offset_t kva;
342 			vm_page_t m;
343 
344 			m = vm_page_lookup_busy_wait(object, OFF_TO_IDX(nsize),
345 						     TRUE, "vsetsz");
346 
347 			if (m && m->valid) {
348 				int base = (int)nsize & PAGE_MASK;
349 				int size = PAGE_SIZE - base;
350 				struct lwbuf *lwb;
351 				struct lwbuf lwb_cache;
352 
353 				/*
354 				 * Clear out partial-page garbage in case
355 				 * the page has been mapped.
356 				 *
357 				 * This is byte aligned.
358 				 */
359 				lwb = lwbuf_alloc(m, &lwb_cache);
360 				kva = lwbuf_kva(lwb);
361 				bzero((caddr_t)kva + base, size);
362 				lwbuf_free(lwb);
363 
364 				/*
365 				 * XXX work around SMP data integrity race
366 				 * by unmapping the page from user processes.
367 				 * The garbage we just cleared may be mapped
368 				 * to a user process running on another cpu
369 				 * and this code is not running through normal
370 				 * I/O channels which handle SMP issues for
371 				 * us, so unmap page to synchronize all cpus.
372 				 *
373 				 * XXX should vm_pager_unmap_page() have
374 				 * dealt with this?
375 				 */
376 				vm_page_protect(m, VM_PROT_NONE);
377 
378 				/*
379 				 * Clear out partial-page dirty bits.  This
380 				 * has the side effect of setting the valid
381 				 * bits, but that is ok.  There are a bunch
382 				 * of places in the VM system where we expected
383 				 * m->dirty == VM_PAGE_BITS_ALL.  The file EOF
384 				 * case is one of them.  If the page is still
385 				 * partially dirty, make it fully dirty.
386 				 *
387 				 * NOTE: We do not clear out the valid
388 				 * bits.  This would prevent bogus_page
389 				 * replacement from working properly.
390 				 *
391 				 * NOTE: We do not want to clear the dirty
392 				 * bit for a partial DEV_BSIZE'd truncation!
393 				 * This is DEV_BSIZE aligned!
394 				 */
395 				vm_page_clear_dirty_beg_nonincl(m, base, size);
396 				if (m->dirty != 0)
397 					m->dirty = VM_PAGE_BITS_ALL;
398 				vm_page_wakeup(m);
399 			} else if (m) {
400 				vm_page_wakeup(m);
401 			}
402 		}
403 	} else {
404 		vp->v_filesize = nsize;
405 	}
406 	vm_object_drop(object);
407 }
408 
409 /*
410  * Release a page busied for a getpages operation.  The page may have become
411  * wired (typically due to being used by the buffer cache) or otherwise been
412  * soft-busied and cannot be freed in that case.  A held page can still be
413  * freed.
414  */
415 void
vnode_pager_freepage(vm_page_t m)416 vnode_pager_freepage(vm_page_t m)
417 {
418 	if ((m->busy_count & PBUSY_MASK) ||
419 	    m->wire_count ||
420 	    (m->flags & PG_NEED_COMMIT)) {
421 		vm_page_activate(m);
422 		vm_page_wakeup(m);
423 	} else {
424 		vm_page_free(m);
425 	}
426 }
427 
428 /*
429  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
430  * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
431  * vnode_pager_generic_getpages() to implement the previous behaviour.
432  *
433  * All other FS's should use the bypass to get to the local media
434  * backing vp's VOP_GETPAGES.
435  */
436 static int
vnode_pager_getpage(vm_object_t object,vm_pindex_t pindex,vm_page_t * mpp,int seqaccess)437 vnode_pager_getpage(vm_object_t object, vm_pindex_t pindex,
438 		    vm_page_t *mpp, int seqaccess)
439 {
440 	int rtval;
441 	struct vnode *vp;
442 
443 	vp = object->handle;
444 	rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess);
445 	if (rtval == EOPNOTSUPP)
446 		panic("vnode_pager: vfs's must implement vop_getpages");
447 	return rtval;
448 }
449 
450 /*
451  * This is now called from local media FS's to operate against their
452  * own vnodes if they fail to implement VOP_GETPAGES.
453  *
454  * With all the caching local media devices do these days there is really
455  * very little point to attempting to restrict the I/O size to contiguous
456  * blocks on-disk, especially if our caller thinks we need all the specified
457  * pages.  Just construct and issue a READ.
458  */
459 int
vnode_pager_generic_getpages(struct vnode * vp,vm_page_t * mpp,int bytecount,int reqpage,int seqaccess)460 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount,
461 			     int reqpage, int seqaccess)
462 {
463 	struct iovec aiov;
464 	struct uio auio;
465 	off_t foff;
466 	int error;
467 	int count;
468 	int i;
469 	int ioflags;
470 	int obytecount;
471 
472 	/*
473 	 * Do not do anything if the vnode is bad.
474 	 */
475 	if (vp->v_mount == NULL)
476 		return VM_PAGER_BAD;
477 
478 	/*
479 	 * Calculate the number of pages.  Since we are paging in whole
480 	 * pages, adjust bytecount to be an integral multiple of the page
481 	 * size.  It will be clipped to the file EOF later on.
482 	 */
483 	bytecount = round_page(bytecount);
484 	count = bytecount / PAGE_SIZE;
485 
486 	/*
487 	 * We could check m[reqpage]->valid here and shortcut the operation,
488 	 * but doing so breaks read-ahead.  Instead assume that the VM
489 	 * system has already done at least the check, don't worry about
490 	 * any races, and issue the VOP_READ to allow read-ahead to function.
491 	 *
492 	 * This keeps the pipeline full for I/O bound sequentially scanned
493 	 * mmap()'s
494 	 */
495 	/* don't shortcut */
496 
497 	/*
498 	 * Discard pages past the file EOF.  If the requested page is past
499 	 * the file EOF we just leave its valid bits set to 0, the caller
500 	 * expects to maintain ownership of the requested page.  If the
501 	 * entire range is past file EOF discard everything and generate
502 	 * a pagein error.
503 	 */
504 	foff = IDX_TO_OFF(mpp[0]->pindex);
505 	if (foff >= vp->v_filesize) {
506 		for (i = 0; i < count; i++) {
507 			if (i != reqpage)
508 				vnode_pager_freepage(mpp[i]);
509 		}
510 		return VM_PAGER_ERROR;
511 	}
512 
513 	if (foff + bytecount > vp->v_filesize) {
514 		bytecount = vp->v_filesize - foff;
515 		i = round_page(bytecount) / PAGE_SIZE;
516 		while (count > i) {
517 			--count;
518 			if (count != reqpage)
519 				vnode_pager_freepage(mpp[count]);
520 		}
521 	}
522 
523 	/*
524 	 * The size of the transfer is bytecount.  bytecount will be an
525 	 * integral multiple of the page size unless it has been clipped
526 	 * to the file EOF.  The transfer cannot exceed the file EOF.
527 	 *
528 	 * When dealing with real devices we must round-up to the device
529 	 * sector size.
530 	 */
531 	if (vp->v_type == VBLK || vp->v_type == VCHR) {
532 		int secmask = vp->v_rdev->si_bsize_phys - 1;
533 		KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large", secmask + 1));
534 		bytecount = (bytecount + secmask) & ~secmask;
535 	}
536 	obytecount = bytecount;
537 
538 	/*
539 	 * Severe hack to avoid deadlocks with the buffer cache
540 	 */
541 	for (i = 0; i < count; ++i) {
542 		vm_page_t mt = mpp[i];
543 
544 		vm_page_io_start(mt);
545 		vm_page_wakeup(mt);
546 	}
547 
548 	/*
549 	 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE
550 	 */
551 	ioflags = IO_VMIO;
552 	if (seqaccess)
553 		ioflags |= IO_SEQMAX << IO_SEQSHIFT;
554 
555 	aiov.iov_base = NULL;
556 	aiov.iov_len = bytecount;
557 	auio.uio_iov = &aiov;
558 	auio.uio_iovcnt = 1;
559 	auio.uio_offset = foff;
560 	auio.uio_segflg = UIO_NOCOPY;
561 	auio.uio_rw = UIO_READ;
562 	auio.uio_resid = bytecount;
563 	auio.uio_td = NULL;
564 	mycpu->gd_cnt.v_vnodein++;
565 	mycpu->gd_cnt.v_vnodepgsin += count;
566 
567 	error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
568 
569 	/*
570 	 * Severe hack to avoid deadlocks with the buffer cache
571 	 */
572 	for (i = 0; i < count; ++i) {
573 		vm_page_busy_wait(mpp[i], FALSE, "getpgs");
574 		vm_page_io_finish(mpp[i]);
575 	}
576 
577 	/*
578 	 * Calculate the actual number of bytes read and clean up the
579 	 * page list.
580 	 */
581 	bytecount -= auio.uio_resid;
582 
583 	for (i = 0; i < count; ++i) {
584 		vm_page_t mt = mpp[i];
585 
586 		if (i != reqpage) {
587 			if (error == 0 && mt->valid) {
588 				if (mt->flags & PG_REFERENCED)
589 					vm_page_activate(mt);
590 				else
591 					vm_page_deactivate(mt);
592 				vm_page_wakeup(mt);
593 			} else {
594 				vnode_pager_freepage(mt);
595 			}
596 		} else if (mt->valid == 0) {
597 			if (error == 0) {
598 				kprintf("page failed but no I/O error page "
599 					"%p object %p pindex %d\n",
600 					mt, mt->object, (int) mt->pindex);
601 				kprintf("i=%d foff=%016lx bytecount=%d/%d "
602 					"uioresid=%zd\n",
603 					i, foff, obytecount, bytecount,
604 					auio.uio_resid);
605 				/* whoops, something happened */
606 				error = EINVAL;
607 			}
608 		} else if (mt->valid != VM_PAGE_BITS_ALL) {
609 			/*
610 			 * Zero-extend the requested page if necessary (if
611 			 * the filesystem is using a small block size).
612 			 */
613 			vm_page_zero_invalid(mt, TRUE);
614 		}
615 	}
616 	if (error) {
617 		kprintf("vnode_pager_getpage: I/O read error\n");
618 	}
619 	return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
620 }
621 
622 /*
623  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
624  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
625  * vnode_pager_generic_putpages() to implement the previous behaviour.
626  *
627  * Caller has already cleared the pmap modified bits, if any.
628  *
629  * All other FS's should use the bypass to get to the local media
630  * backing vp's VOP_PUTPAGES.
631  */
632 static void
vnode_pager_putpages(vm_object_t object,vm_page_t * m,int count,int flags,int * rtvals)633 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
634 		     int flags, int *rtvals)
635 {
636 	int rtval;
637 	struct vnode *vp;
638 	int bytes = count * PAGE_SIZE;
639 
640 	/*
641 	 * Force synchronous operation if we are extremely low on memory
642 	 * to prevent a low-memory deadlock.  VOP operations often need to
643 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
644 	 * operation ).  The swapper handles the case by limiting the amount
645 	 * of asynchronous I/O, but that sort of solution doesn't scale well
646 	 * for the vnode pager without a lot of work.
647 	 *
648 	 * Also, the backing vnode's iodone routine may not wake the pageout
649 	 * daemon up.  This should be probably be addressed XXX.
650 	 */
651 
652 	if ((vmstats.v_free_count + vmstats.v_cache_count) <
653 	    vmstats.v_pageout_free_min) {
654 		flags |= OBJPC_SYNC;
655 	}
656 
657 	/*
658 	 * Call device-specific putpages function
659 	 */
660 	vp = object->handle;
661 	rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals, 0);
662 	if (rtval == EOPNOTSUPP) {
663 	    kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
664 	    rtval = vnode_pager_generic_putpages( vp, m, bytes, flags, rtvals);
665 	}
666 }
667 
668 
669 /*
670  * This is now called from local media FS's to operate against their
671  * own vnodes if they fail to implement VOP_PUTPAGES.
672  *
673  * This is typically called indirectly via the pageout daemon and
674  * clustering has already typically occured, so in general we ask the
675  * underlying filesystem to write the data out asynchronously rather
676  * then delayed.
677  */
678 int
vnode_pager_generic_putpages(struct vnode * vp,vm_page_t * m,int bytecount,int flags,int * rtvals)679 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
680 			     int flags, int *rtvals)
681 {
682 	int i;
683 	int maxsize, ncount, count;
684 	vm_ooffset_t poffset;
685 	struct uio auio;
686 	struct iovec aiov;
687 	int error;
688 	int ioflags;
689 
690 	count = bytecount / PAGE_SIZE;
691 
692 	for (i = 0; i < count; i++)
693 		rtvals[i] = VM_PAGER_AGAIN;
694 
695 	if ((int) m[0]->pindex < 0) {
696 		kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
697 			(long)m[0]->pindex, m[0]->dirty);
698 		rtvals[0] = VM_PAGER_BAD;
699 		return VM_PAGER_BAD;
700 	}
701 
702 	maxsize = count * PAGE_SIZE;
703 	ncount = count;
704 
705 	poffset = IDX_TO_OFF(m[0]->pindex);
706 
707 	/*
708 	 * If the page-aligned write is larger then the actual file we
709 	 * have to invalidate pages occuring beyond the file EOF.
710 	 *
711 	 * If the file EOF resides in the middle of a page we still clear
712 	 * all of that page's dirty bits later on.  If we didn't it would
713 	 * endlessly re-write.
714 	 *
715 	 * We do not under any circumstances truncate the valid bits, as
716 	 * this will screw up bogus page replacement.
717 	 *
718 	 * The caller has already read-protected the pages.  The VFS must
719 	 * use the buffer cache to wrap the pages.  The pages might not
720 	 * be immediately flushed by the buffer cache but once under its
721 	 * control the pages themselves can wind up being marked clean
722 	 * and their covering buffer cache buffer can be marked dirty.
723 	 */
724 	if (poffset + maxsize > vp->v_filesize) {
725 		if (poffset < vp->v_filesize) {
726 			maxsize = vp->v_filesize - poffset;
727 			ncount = btoc(maxsize);
728 		} else {
729 			maxsize = 0;
730 			ncount = 0;
731 		}
732 		if (ncount < count) {
733 			for (i = ncount; i < count; i++) {
734 				rtvals[i] = VM_PAGER_BAD;
735 			}
736 		}
737 	}
738 
739 	/*
740 	 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
741 	 * rather then a bdwrite() to prevent paging I/O from saturating
742 	 * the buffer cache.  Dummy-up the sequential heuristic to cause
743 	 * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
744 	 * the system decides how to cluster.
745 	 */
746 	ioflags = IO_VMIO;
747 	if (flags & (OBJPC_SYNC | OBJPC_INVAL))
748 		ioflags |= IO_SYNC;
749 	else if ((flags & OBJPC_CLUSTER_OK) == 0)
750 		ioflags |= IO_ASYNC;
751 	ioflags |= (flags & OBJPC_INVAL) ? IO_INVAL: 0;
752 	ioflags |= IO_SEQMAX << IO_SEQSHIFT;
753 
754 	aiov.iov_base = (caddr_t) 0;
755 	aiov.iov_len = maxsize;
756 	auio.uio_iov = &aiov;
757 	auio.uio_iovcnt = 1;
758 	auio.uio_offset = poffset;
759 	auio.uio_segflg = UIO_NOCOPY;
760 	auio.uio_rw = UIO_WRITE;
761 	auio.uio_resid = maxsize;
762 	auio.uio_td = NULL;
763 	error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
764 	mycpu->gd_cnt.v_vnodeout++;
765 	mycpu->gd_cnt.v_vnodepgsout += ncount;
766 
767 	if (error) {
768 		krateprintf(&vbadrate,
769 			    "vnode_pager_putpages: I/O error %d\n", error);
770 	}
771 	if (auio.uio_resid) {
772 		krateprintf(&vresrate,
773 			    "vnode_pager_putpages: residual I/O %zd at %lu\n",
774 			    auio.uio_resid, (u_long)m[0]->pindex);
775 	}
776 	if (error == 0) {
777 		for (i = 0; i < ncount; i++) {
778 			rtvals[i] = VM_PAGER_OK;
779 			vm_page_undirty(m[i]);
780 		}
781 	}
782 	return rtvals[0];
783 }
784 
785 /*
786  * Run the chain and if the bottom-most object is a vnode-type lock the
787  * underlying vnode.  A locked vnode or NULL is returned.
788  *
789  * Caller must hold the first object.
790  */
791 struct vnode *
vnode_pager_lock(vm_map_backing_t ba)792 vnode_pager_lock(vm_map_backing_t ba)
793 {
794 	vm_map_backing_t lba;
795 	struct vnode *vp = NULL;
796 	vm_object_t lobject;
797 	int error;
798 
799 	if (ba == NULL)
800 		return NULL;
801 	lba = ba;
802 	while (lba->backing_ba)
803 		lba = lba->backing_ba;
804 	if ((lobject = lba->object) == NULL)
805 		return NULL;
806 	if (lba != ba)
807 		vm_object_hold_shared(lobject);
808 
809 	while (lobject->type == OBJT_VNODE &&
810 	       (lobject->flags & OBJ_DEAD) == 0) {
811 		/*
812 		 * Extract the vp
813 		 */
814 		vp = lobject->handle;
815 		error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
816 		if (error == 0) {
817 			if (lobject->handle == vp)
818 				break;
819 			vput(vp);
820 		} else {
821 			kprintf("vnode_pager_lock: vp %p error %d "
822 				"lockstatus %d, retrying\n",
823 				vp, error,
824 				lockstatus(&vp->v_lock, curthread));
825 			tsleep(lobject->handle, 0, "vnpgrl", hz);
826 		}
827 		vp = NULL;
828 	}
829 	if (lba != ba)
830 		vm_object_drop(lobject);
831 	return (vp);
832 }
833