FreeBSD/Linux Kernel Cross Reference
sys/vm/vnode_pager.c
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: releng/5.2/sys/vm/vnode_pager.c 122748 2003-11-15 09:54:11Z tjr $");
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/conf.h>
65
66 #include <vm/vm.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_pager.h>
70 #include <vm/vm_map.h>
71 #include <vm/vnode_pager.h>
72 #include <vm/vm_extern.h>
73
74 static void vnode_pager_init(void);
75 static vm_offset_t vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
76 int *run);
77 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
78 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
79 static void vnode_pager_dealloc(vm_object_t);
80 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, 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, int *, int *);
83
84 struct pagerops vnodepagerops = {
85 .pgo_init = vnode_pager_init,
86 .pgo_alloc = vnode_pager_alloc,
87 .pgo_dealloc = vnode_pager_dealloc,
88 .pgo_getpages = vnode_pager_getpages,
89 .pgo_putpages = vnode_pager_putpages,
90 .pgo_haspage = vnode_pager_haspage,
91 };
92
93 int vnode_pbuf_freecnt;
94
95 static void
96 vnode_pager_init(void)
97 {
98
99 vnode_pbuf_freecnt = nswbuf / 2 + 1;
100 }
101
102 /*
103 * Allocate (or lookup) pager for a vnode.
104 * Handle is a vnode pointer.
105 *
106 * MPSAFE
107 */
108 vm_object_t
109 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
110 vm_ooffset_t offset)
111 {
112 vm_object_t object;
113 struct vnode *vp;
114
115 /*
116 * Pageout to vnode, no can do yet.
117 */
118 if (handle == NULL)
119 return (NULL);
120
121 vp = (struct vnode *) handle;
122
123 ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc");
124
125 mtx_lock(&Giant);
126 /*
127 * Prevent race condition when allocating the object. This
128 * can happen with NFS vnodes since the nfsnode isn't locked.
129 */
130 VI_LOCK(vp);
131 while (vp->v_iflag & VI_OLOCK) {
132 vp->v_iflag |= VI_OWANT;
133 msleep(vp, VI_MTX(vp), PVM, "vnpobj", 0);
134 }
135 vp->v_iflag |= VI_OLOCK;
136 VI_UNLOCK(vp);
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 VM_OBJECT_LOCK(object);
144 if ((object->flags & OBJ_DEAD) == 0)
145 break;
146 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vadead", 0);
147 }
148
149 if (vp->v_usecount == 0)
150 panic("vnode_pager_alloc: no vnode reference");
151
152 if (object == NULL) {
153 /*
154 * And an object of the appropriate size
155 */
156 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
157
158 object->un_pager.vnp.vnp_size = size;
159
160 object->handle = handle;
161 vp->v_object = object;
162 } else {
163 object->ref_count++;
164 VM_OBJECT_UNLOCK(object);
165 }
166 VI_LOCK(vp);
167 vp->v_usecount++;
168 vp->v_iflag &= ~VI_OLOCK;
169 if (vp->v_iflag & VI_OWANT) {
170 vp->v_iflag &= ~VI_OWANT;
171 wakeup(vp);
172 }
173 VI_UNLOCK(vp);
174 mtx_unlock(&Giant);
175 return (object);
176 }
177
178 /*
179 * The object must be locked.
180 */
181 static void
182 vnode_pager_dealloc(object)
183 vm_object_t object;
184 {
185 struct vnode *vp = object->handle;
186
187 if (vp == NULL)
188 panic("vnode_pager_dealloc: pager already dealloced");
189
190 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
191 vm_object_pip_wait(object, "vnpdea");
192
193 object->handle = NULL;
194 object->type = OBJT_DEAD;
195 ASSERT_VOP_LOCKED(vp, "vnode_pager_dealloc");
196 vp->v_object = NULL;
197 vp->v_vflag &= ~(VV_TEXT | VV_OBJBUF);
198 }
199
200 static boolean_t
201 vnode_pager_haspage(object, pindex, before, after)
202 vm_object_t object;
203 vm_pindex_t pindex;
204 int *before;
205 int *after;
206 {
207 struct vnode *vp = object->handle;
208 daddr_t bn;
209 int err;
210 daddr_t reqblock;
211 int poff;
212 int bsize;
213 int pagesperblock, blocksperpage;
214
215 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
216 /*
217 * If no vp or vp is doomed or marked transparent to VM, we do not
218 * have the page.
219 */
220 if (vp == NULL)
221 return FALSE;
222
223 VI_LOCK(vp);
224 if (vp->v_iflag & VI_DOOMED) {
225 VI_UNLOCK(vp);
226 return FALSE;
227 }
228 VI_UNLOCK(vp);
229 /*
230 * If filesystem no longer mounted or offset beyond end of file we do
231 * not have the page.
232 */
233 if ((vp->v_mount == NULL) ||
234 (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size))
235 return FALSE;
236
237 bsize = vp->v_mount->mnt_stat.f_iosize;
238 pagesperblock = bsize / PAGE_SIZE;
239 blocksperpage = 0;
240 if (pagesperblock > 0) {
241 reqblock = pindex / pagesperblock;
242 } else {
243 blocksperpage = (PAGE_SIZE / bsize);
244 reqblock = pindex * blocksperpage;
245 }
246 VM_OBJECT_UNLOCK(object);
247 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
248 VM_OBJECT_LOCK(object);
249 if (err)
250 return TRUE;
251 if (bn == -1)
252 return FALSE;
253 if (pagesperblock > 0) {
254 poff = pindex - (reqblock * pagesperblock);
255 if (before) {
256 *before *= pagesperblock;
257 *before += poff;
258 }
259 if (after) {
260 int numafter;
261 *after *= pagesperblock;
262 numafter = pagesperblock - (poff + 1);
263 if (IDX_TO_OFF(pindex + numafter) >
264 object->un_pager.vnp.vnp_size) {
265 numafter =
266 OFF_TO_IDX(object->un_pager.vnp.vnp_size) -
267 pindex;
268 }
269 *after += numafter;
270 }
271 } else {
272 if (before) {
273 *before /= blocksperpage;
274 }
275
276 if (after) {
277 *after /= blocksperpage;
278 }
279 }
280 return TRUE;
281 }
282
283 /*
284 * Lets the VM system know about a change in size for a file.
285 * We adjust our own internal size and flush any cached pages in
286 * the associated object that are affected by the size change.
287 *
288 * Note: this routine may be invoked as a result of a pager put
289 * operation (possibly at object termination time), so we must be careful.
290 */
291 void
292 vnode_pager_setsize(vp, nsize)
293 struct vnode *vp;
294 vm_ooffset_t nsize;
295 {
296 vm_object_t object;
297 vm_page_t m;
298 vm_pindex_t nobjsize;
299
300 if ((object = vp->v_object) == NULL)
301 return;
302 VM_OBJECT_LOCK(object);
303 if (nsize == object->un_pager.vnp.vnp_size) {
304 /*
305 * Hasn't changed size
306 */
307 VM_OBJECT_UNLOCK(object);
308 return;
309 }
310 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
311 if (nsize < object->un_pager.vnp.vnp_size) {
312 /*
313 * File has shrunk. Toss any cached pages beyond the new EOF.
314 */
315 if (nobjsize < object->size)
316 vm_object_page_remove(object, nobjsize, object->size,
317 FALSE);
318 /*
319 * this gets rid of garbage at the end of a page that is now
320 * only partially backed by the vnode.
321 *
322 * XXX for some reason (I don't know yet), if we take a
323 * completely invalid page and mark it partially valid
324 * it can screw up NFS reads, so we don't allow the case.
325 */
326 if ((nsize & PAGE_MASK) &&
327 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
328 m->valid != 0) {
329 int base = (int)nsize & PAGE_MASK;
330 int size = PAGE_SIZE - base;
331
332 /*
333 * Clear out partial-page garbage in case
334 * the page has been mapped.
335 */
336 pmap_zero_page_area(m, base, size);
337
338 /*
339 * XXX work around SMP data integrity race
340 * by unmapping the page from user processes.
341 * The garbage we just cleared may be mapped
342 * to a user process running on another cpu
343 * and this code is not running through normal
344 * I/O channels which handle SMP issues for
345 * us, so unmap page to synchronize all cpus.
346 *
347 * XXX should vm_pager_unmap_page() have
348 * dealt with this?
349 */
350 vm_page_lock_queues();
351 pmap_remove_all(m);
352
353 /*
354 * Clear out partial-page dirty bits. This
355 * has the side effect of setting the valid
356 * bits, but that is ok. There are a bunch
357 * of places in the VM system where we expected
358 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
359 * case is one of them. If the page is still
360 * partially dirty, make it fully dirty.
361 *
362 * note that we do not clear out the valid
363 * bits. This would prevent bogus_page
364 * replacement from working properly.
365 */
366 vm_page_set_validclean(m, base, size);
367 if (m->dirty != 0)
368 m->dirty = VM_PAGE_BITS_ALL;
369 vm_page_unlock_queues();
370 }
371 }
372 object->un_pager.vnp.vnp_size = nsize;
373 object->size = nobjsize;
374 VM_OBJECT_UNLOCK(object);
375 }
376
377 /*
378 * calculate the linear (byte) disk address of specified virtual
379 * file address
380 */
381 static vm_offset_t
382 vnode_pager_addr(vp, address, run)
383 struct vnode *vp;
384 vm_ooffset_t address;
385 int *run;
386 {
387 int rtaddress;
388 int bsize;
389 daddr_t block;
390 int err;
391 daddr_t vblock;
392 int voffset;
393
394 GIANT_REQUIRED;
395 if ((int) address < 0)
396 return -1;
397
398 if (vp->v_mount == NULL)
399 return -1;
400
401 bsize = vp->v_mount->mnt_stat.f_iosize;
402 vblock = address / bsize;
403 voffset = address % bsize;
404
405 err = VOP_BMAP(vp, vblock, NULL, &block, run, NULL);
406
407 if (err || (block == -1))
408 rtaddress = -1;
409 else {
410 rtaddress = block + voffset / DEV_BSIZE;
411 if (run) {
412 *run += 1;
413 *run *= bsize/PAGE_SIZE;
414 *run -= voffset/PAGE_SIZE;
415 }
416 }
417
418 return rtaddress;
419 }
420
421 /*
422 * small block filesystem vnode pager input
423 */
424 static int
425 vnode_pager_input_smlfs(object, m)
426 vm_object_t object;
427 vm_page_t m;
428 {
429 int i;
430 struct vnode *dp, *vp;
431 struct buf *bp;
432 vm_offset_t kva;
433 int fileaddr;
434 vm_offset_t bsize;
435 int error = 0;
436
437 GIANT_REQUIRED;
438
439 vp = object->handle;
440 if (vp->v_mount == NULL)
441 return VM_PAGER_BAD;
442
443 bsize = vp->v_mount->mnt_stat.f_iosize;
444
445 VOP_BMAP(vp, 0, &dp, 0, NULL, NULL);
446
447 kva = vm_pager_map_page(m);
448
449 for (i = 0; i < PAGE_SIZE / bsize; i++) {
450 vm_ooffset_t address;
451
452 if (vm_page_bits(i * bsize, bsize) & m->valid)
453 continue;
454
455 address = IDX_TO_OFF(m->pindex) + i * bsize;
456 if (address >= object->un_pager.vnp.vnp_size) {
457 fileaddr = -1;
458 } else {
459 fileaddr = vnode_pager_addr(vp, address, NULL);
460 }
461 if (fileaddr != -1) {
462 bp = getpbuf(&vnode_pbuf_freecnt);
463
464 /* build a minimal buffer header */
465 bp->b_iocmd = BIO_READ;
466 bp->b_iodone = bdone;
467 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
468 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
469 bp->b_rcred = crhold(curthread->td_ucred);
470 bp->b_wcred = crhold(curthread->td_ucred);
471 bp->b_data = (caddr_t) kva + i * bsize;
472 bp->b_blkno = fileaddr;
473 pbgetvp(dp, bp);
474 bp->b_bcount = bsize;
475 bp->b_bufsize = bsize;
476 bp->b_runningbufspace = bp->b_bufsize;
477 runningbufspace += bp->b_runningbufspace;
478
479 /* do the input */
480 bp->b_iooffset = dbtob(bp->b_blkno);
481 if (dp->v_type == VCHR)
482 VOP_SPECSTRATEGY(bp->b_vp, bp);
483 else
484 VOP_STRATEGY(bp->b_vp, bp);
485
486 /* we definitely need to be at splvm here */
487
488 bwait(bp, PVM, "vnsrd");
489
490 if ((bp->b_ioflags & BIO_ERROR) != 0)
491 error = EIO;
492
493 /*
494 * free the buffer header back to the swap buffer pool
495 */
496 relpbuf(bp, &vnode_pbuf_freecnt);
497 if (error)
498 break;
499
500 VM_OBJECT_LOCK(object);
501 vm_page_lock_queues();
502 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
503 vm_page_unlock_queues();
504 VM_OBJECT_UNLOCK(object);
505 } else {
506 VM_OBJECT_LOCK(object);
507 vm_page_lock_queues();
508 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
509 vm_page_unlock_queues();
510 VM_OBJECT_UNLOCK(object);
511 bzero((caddr_t) kva + i * bsize, bsize);
512 }
513 }
514 vm_pager_unmap_page(kva);
515 vm_page_lock_queues();
516 pmap_clear_modify(m);
517 vm_page_flag_clear(m, PG_ZERO);
518 vm_page_unlock_queues();
519 if (error) {
520 return VM_PAGER_ERROR;
521 }
522 return VM_PAGER_OK;
523
524 }
525
526
527 /*
528 * old style vnode pager output routine
529 */
530 static int
531 vnode_pager_input_old(object, m)
532 vm_object_t object;
533 vm_page_t m;
534 {
535 struct uio auio;
536 struct iovec aiov;
537 int error;
538 int size;
539 vm_offset_t kva;
540 struct vnode *vp;
541
542 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
543 error = 0;
544
545 /*
546 * Return failure if beyond current EOF
547 */
548 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
549 return VM_PAGER_BAD;
550 } else {
551 size = PAGE_SIZE;
552 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
553 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
554 vp = object->handle;
555 VM_OBJECT_UNLOCK(object);
556
557 /*
558 * Allocate a kernel virtual address and initialize so that
559 * we can use VOP_READ/WRITE routines.
560 */
561 kva = vm_pager_map_page(m);
562
563 aiov.iov_base = (caddr_t) kva;
564 aiov.iov_len = size;
565 auio.uio_iov = &aiov;
566 auio.uio_iovcnt = 1;
567 auio.uio_offset = IDX_TO_OFF(m->pindex);
568 auio.uio_segflg = UIO_SYSSPACE;
569 auio.uio_rw = UIO_READ;
570 auio.uio_resid = size;
571 auio.uio_td = curthread;
572
573 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
574 if (!error) {
575 int count = size - auio.uio_resid;
576
577 if (count == 0)
578 error = EINVAL;
579 else if (count != PAGE_SIZE)
580 bzero((caddr_t) kva + count, PAGE_SIZE - count);
581 }
582 vm_pager_unmap_page(kva);
583
584 VM_OBJECT_LOCK(object);
585 }
586 vm_page_lock_queues();
587 pmap_clear_modify(m);
588 vm_page_undirty(m);
589 vm_page_flag_clear(m, PG_ZERO);
590 vm_page_unlock_queues();
591 if (!error)
592 m->valid = VM_PAGE_BITS_ALL;
593 return error ? VM_PAGER_ERROR : VM_PAGER_OK;
594 }
595
596 /*
597 * generic vnode pager input routine
598 */
599
600 /*
601 * Local media VFS's that do not implement their own VOP_GETPAGES
602 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
603 * to implement the previous behaviour.
604 *
605 * All other FS's should use the bypass to get to the local media
606 * backing vp's VOP_GETPAGES.
607 */
608 static int
609 vnode_pager_getpages(object, m, count, reqpage)
610 vm_object_t object;
611 vm_page_t *m;
612 int count;
613 int reqpage;
614 {
615 int rtval;
616 struct vnode *vp;
617 int bytes = count * PAGE_SIZE;
618
619 vp = object->handle;
620 VM_OBJECT_UNLOCK(object);
621 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
622 KASSERT(rtval != EOPNOTSUPP,
623 ("vnode_pager: FS getpages not implemented\n"));
624 VM_OBJECT_LOCK(object);
625 return rtval;
626 }
627
628 /*
629 * This is now called from local media FS's to operate against their
630 * own vnodes if they fail to implement VOP_GETPAGES.
631 */
632 int
633 vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
634 struct vnode *vp;
635 vm_page_t *m;
636 int bytecount;
637 int reqpage;
638 {
639 vm_object_t object;
640 vm_offset_t kva;
641 off_t foff, tfoff, nextoff;
642 int i, j, size, bsize, first, firstaddr;
643 struct vnode *dp;
644 int runpg;
645 int runend;
646 struct buf *bp;
647 int count;
648 int error = 0;
649
650 GIANT_REQUIRED;
651 object = vp->v_object;
652 count = bytecount / PAGE_SIZE;
653
654 if (vp->v_mount == NULL)
655 return VM_PAGER_BAD;
656
657 bsize = vp->v_mount->mnt_stat.f_iosize;
658
659 /* get the UNDERLYING device for the file with VOP_BMAP() */
660
661 /*
662 * originally, we did not check for an error return value -- assuming
663 * an fs always has a bmap entry point -- that assumption is wrong!!!
664 */
665 foff = IDX_TO_OFF(m[reqpage]->pindex);
666
667 /*
668 * if we can't bmap, use old VOP code
669 */
670 if (VOP_BMAP(vp, 0, &dp, 0, NULL, NULL)) {
671 VM_OBJECT_LOCK(object);
672 vm_page_lock_queues();
673 for (i = 0; i < count; i++)
674 if (i != reqpage)
675 vm_page_free(m[i]);
676 vm_page_unlock_queues();
677 cnt.v_vnodein++;
678 cnt.v_vnodepgsin++;
679 error = vnode_pager_input_old(object, m[reqpage]);
680 VM_OBJECT_UNLOCK(object);
681 return (error);
682
683 /*
684 * if the blocksize is smaller than a page size, then use
685 * special small filesystem code. NFS sometimes has a small
686 * blocksize, but it can handle large reads itself.
687 */
688 } else if ((PAGE_SIZE / bsize) > 1 &&
689 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
690 VM_OBJECT_LOCK(object);
691 vm_page_lock_queues();
692 for (i = 0; i < count; i++)
693 if (i != reqpage)
694 vm_page_free(m[i]);
695 vm_page_unlock_queues();
696 VM_OBJECT_UNLOCK(object);
697 cnt.v_vnodein++;
698 cnt.v_vnodepgsin++;
699 return vnode_pager_input_smlfs(object, m[reqpage]);
700 }
701
702 /*
703 * If we have a completely valid page available to us, we can
704 * clean up and return. Otherwise we have to re-read the
705 * media.
706 */
707 VM_OBJECT_LOCK(object);
708 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
709 vm_page_lock_queues();
710 for (i = 0; i < count; i++)
711 if (i != reqpage)
712 vm_page_free(m[i]);
713 vm_page_unlock_queues();
714 VM_OBJECT_UNLOCK(object);
715 return VM_PAGER_OK;
716 }
717 m[reqpage]->valid = 0;
718 VM_OBJECT_UNLOCK(object);
719
720 /*
721 * here on direct device I/O
722 */
723 firstaddr = -1;
724
725 /*
726 * calculate the run that includes the required page
727 */
728 for (first = 0, i = 0; i < count; i = runend) {
729 firstaddr = vnode_pager_addr(vp,
730 IDX_TO_OFF(m[i]->pindex), &runpg);
731 if (firstaddr == -1) {
732 VM_OBJECT_LOCK(object);
733 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
734 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %d, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
735 firstaddr, (uintmax_t)(foff >> 32),
736 (uintmax_t)foff,
737 (uintmax_t)
738 (object->un_pager.vnp.vnp_size >> 32),
739 (uintmax_t)object->un_pager.vnp.vnp_size);
740 }
741 vm_page_lock_queues();
742 vm_page_free(m[i]);
743 vm_page_unlock_queues();
744 VM_OBJECT_UNLOCK(object);
745 runend = i + 1;
746 first = runend;
747 continue;
748 }
749 runend = i + runpg;
750 if (runend <= reqpage) {
751 VM_OBJECT_LOCK(object);
752 vm_page_lock_queues();
753 for (j = i; j < runend; j++)
754 vm_page_free(m[j]);
755 vm_page_unlock_queues();
756 VM_OBJECT_UNLOCK(object);
757 } else {
758 if (runpg < (count - first)) {
759 VM_OBJECT_LOCK(object);
760 vm_page_lock_queues();
761 for (i = first + runpg; i < count; i++)
762 vm_page_free(m[i]);
763 vm_page_unlock_queues();
764 VM_OBJECT_UNLOCK(object);
765 count = first + runpg;
766 }
767 break;
768 }
769 first = runend;
770 }
771
772 /*
773 * the first and last page have been calculated now, move input pages
774 * to be zero based...
775 */
776 if (first != 0) {
777 for (i = first; i < count; i++) {
778 m[i - first] = m[i];
779 }
780 count -= first;
781 reqpage -= first;
782 }
783
784 /*
785 * calculate the file virtual address for the transfer
786 */
787 foff = IDX_TO_OFF(m[0]->pindex);
788
789 /*
790 * calculate the size of the transfer
791 */
792 size = count * PAGE_SIZE;
793 if ((foff + size) > object->un_pager.vnp.vnp_size)
794 size = object->un_pager.vnp.vnp_size - foff;
795
796 /*
797 * round up physical size for real devices.
798 */
799 if (dp->v_type == VBLK || dp->v_type == VCHR) {
800 int secmask = dp->v_rdev->si_bsize_phys - 1;
801 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
802 size = (size + secmask) & ~secmask;
803 }
804
805 bp = getpbuf(&vnode_pbuf_freecnt);
806 kva = (vm_offset_t) bp->b_data;
807
808 /*
809 * and map the pages to be read into the kva
810 */
811 pmap_qenter(kva, m, count);
812
813 /* build a minimal buffer header */
814 bp->b_iocmd = BIO_READ;
815 bp->b_iodone = bdone;
816 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
817 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
818 bp->b_rcred = crhold(curthread->td_ucred);
819 bp->b_wcred = crhold(curthread->td_ucred);
820 bp->b_blkno = firstaddr;
821 pbgetvp(dp, bp);
822 bp->b_bcount = size;
823 bp->b_bufsize = size;
824 bp->b_runningbufspace = bp->b_bufsize;
825 runningbufspace += bp->b_runningbufspace;
826
827 cnt.v_vnodein++;
828 cnt.v_vnodepgsin += count;
829
830 /* do the input */
831 bp->b_iooffset = dbtob(bp->b_blkno);
832 if (dp->v_type == VCHR)
833 VOP_SPECSTRATEGY(bp->b_vp, bp);
834 else
835 VOP_STRATEGY(bp->b_vp, bp);
836
837 bwait(bp, PVM, "vnread");
838
839 if ((bp->b_ioflags & BIO_ERROR) != 0)
840 error = EIO;
841
842 if (!error) {
843 if (size != count * PAGE_SIZE)
844 bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
845 }
846 pmap_qremove(kva, count);
847
848 /*
849 * free the buffer header back to the swap buffer pool
850 */
851 relpbuf(bp, &vnode_pbuf_freecnt);
852
853 VM_OBJECT_LOCK(object);
854 vm_page_lock_queues();
855 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
856 vm_page_t mt;
857
858 nextoff = tfoff + PAGE_SIZE;
859 mt = m[i];
860
861 if (nextoff <= object->un_pager.vnp.vnp_size) {
862 /*
863 * Read filled up entire page.
864 */
865 mt->valid = VM_PAGE_BITS_ALL;
866 vm_page_undirty(mt); /* should be an assert? XXX */
867 pmap_clear_modify(mt);
868 } else {
869 /*
870 * Read did not fill up entire page. Since this
871 * is getpages, the page may be mapped, so we have
872 * to zero the invalid portions of the page even
873 * though we aren't setting them valid.
874 *
875 * Currently we do not set the entire page valid,
876 * we just try to clear the piece that we couldn't
877 * read.
878 */
879 vm_page_set_validclean(mt, 0,
880 object->un_pager.vnp.vnp_size - tfoff);
881 /* handled by vm_fault now */
882 /* vm_page_zero_invalid(mt, FALSE); */
883 }
884
885 vm_page_flag_clear(mt, PG_ZERO);
886 if (i != reqpage) {
887
888 /*
889 * whether or not to leave the page activated is up in
890 * the air, but we should put the page on a page queue
891 * somewhere. (it already is in the object). Result:
892 * It appears that empirical results show that
893 * deactivating pages is best.
894 */
895
896 /*
897 * just in case someone was asking for this page we
898 * now tell them that it is ok to use
899 */
900 if (!error) {
901 if (mt->flags & PG_WANTED)
902 vm_page_activate(mt);
903 else
904 vm_page_deactivate(mt);
905 vm_page_wakeup(mt);
906 } else {
907 vm_page_free(mt);
908 }
909 }
910 }
911 vm_page_unlock_queues();
912 VM_OBJECT_UNLOCK(object);
913 if (error) {
914 printf("vnode_pager_getpages: I/O read error\n");
915 }
916 return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
917 }
918
919 /*
920 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
921 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
922 * vnode_pager_generic_putpages() to implement the previous behaviour.
923 *
924 * All other FS's should use the bypass to get to the local media
925 * backing vp's VOP_PUTPAGES.
926 */
927 static void
928 vnode_pager_putpages(object, m, count, sync, rtvals)
929 vm_object_t object;
930 vm_page_t *m;
931 int count;
932 boolean_t sync;
933 int *rtvals;
934 {
935 int rtval;
936 struct vnode *vp;
937 struct mount *mp;
938 int bytes = count * PAGE_SIZE;
939
940 GIANT_REQUIRED;
941 /*
942 * Force synchronous operation if we are extremely low on memory
943 * to prevent a low-memory deadlock. VOP operations often need to
944 * allocate more memory to initiate the I/O ( i.e. do a BMAP
945 * operation ). The swapper handles the case by limiting the amount
946 * of asynchronous I/O, but that sort of solution doesn't scale well
947 * for the vnode pager without a lot of work.
948 *
949 * Also, the backing vnode's iodone routine may not wake the pageout
950 * daemon up. This should be probably be addressed XXX.
951 */
952
953 if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
954 sync |= OBJPC_SYNC;
955
956 /*
957 * Call device-specific putpages function
958 */
959 vp = object->handle;
960 VM_OBJECT_UNLOCK(object);
961 if (vp->v_type != VREG)
962 mp = NULL;
963 (void)vn_start_write(vp, &mp, V_WAIT);
964 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
965 KASSERT(rtval != EOPNOTSUPP,
966 ("vnode_pager: stale FS putpages\n"));
967 vn_finished_write(mp);
968 VM_OBJECT_LOCK(object);
969 }
970
971
972 /*
973 * This is now called from local media FS's to operate against their
974 * own vnodes if they fail to implement VOP_PUTPAGES.
975 *
976 * This is typically called indirectly via the pageout daemon and
977 * clustering has already typically occured, so in general we ask the
978 * underlying filesystem to write the data out asynchronously rather
979 * then delayed.
980 */
981 int
982 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals)
983 struct vnode *vp;
984 vm_page_t *m;
985 int bytecount;
986 int flags;
987 int *rtvals;
988 {
989 int i;
990 vm_object_t object;
991 int count;
992
993 int maxsize, ncount;
994 vm_ooffset_t poffset;
995 struct uio auio;
996 struct iovec aiov;
997 int error;
998 int ioflags;
999
1000 GIANT_REQUIRED;
1001 object = vp->v_object;
1002 count = bytecount / PAGE_SIZE;
1003
1004 for (i = 0; i < count; i++)
1005 rtvals[i] = VM_PAGER_AGAIN;
1006
1007 if ((int) m[0]->pindex < 0) {
1008 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1009 (long)m[0]->pindex, (u_long)m[0]->dirty);
1010 rtvals[0] = VM_PAGER_BAD;
1011 return VM_PAGER_BAD;
1012 }
1013
1014 maxsize = count * PAGE_SIZE;
1015 ncount = count;
1016
1017 poffset = IDX_TO_OFF(m[0]->pindex);
1018
1019 /*
1020 * If the page-aligned write is larger then the actual file we
1021 * have to invalidate pages occuring beyond the file EOF. However,
1022 * there is an edge case where a file may not be page-aligned where
1023 * the last page is partially invalid. In this case the filesystem
1024 * may not properly clear the dirty bits for the entire page (which
1025 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1026 * With the page locked we are free to fix-up the dirty bits here.
1027 *
1028 * We do not under any circumstances truncate the valid bits, as
1029 * this will screw up bogus page replacement.
1030 */
1031 if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1032 if (object->un_pager.vnp.vnp_size > poffset) {
1033 int pgoff;
1034
1035 maxsize = object->un_pager.vnp.vnp_size - poffset;
1036 ncount = btoc(maxsize);
1037 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1038 vm_page_lock_queues();
1039 vm_page_clear_dirty(m[ncount - 1], pgoff,
1040 PAGE_SIZE - pgoff);
1041 vm_page_unlock_queues();
1042 }
1043 } else {
1044 maxsize = 0;
1045 ncount = 0;
1046 }
1047 if (ncount < count) {
1048 for (i = ncount; i < count; i++) {
1049 rtvals[i] = VM_PAGER_BAD;
1050 }
1051 }
1052 }
1053
1054 /*
1055 * pageouts are already clustered, use IO_ASYNC t o force a bawrite()
1056 * rather then a bdwrite() to prevent paging I/O from saturating
1057 * the buffer cache. Dummy-up the sequential heuristic to cause
1058 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
1059 * the system decides how to cluster.
1060 */
1061 ioflags = IO_VMIO;
1062 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1063 ioflags |= IO_SYNC;
1064 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1065 ioflags |= IO_ASYNC;
1066 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1067 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1068
1069 aiov.iov_base = (caddr_t) 0;
1070 aiov.iov_len = maxsize;
1071 auio.uio_iov = &aiov;
1072 auio.uio_iovcnt = 1;
1073 auio.uio_offset = poffset;
1074 auio.uio_segflg = UIO_NOCOPY;
1075 auio.uio_rw = UIO_WRITE;
1076 auio.uio_resid = maxsize;
1077 auio.uio_td = (struct thread *) 0;
1078 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1079 cnt.v_vnodeout++;
1080 cnt.v_vnodepgsout += ncount;
1081
1082 if (error) {
1083 printf("vnode_pager_putpages: I/O error %d\n", error);
1084 }
1085 if (auio.uio_resid) {
1086 printf("vnode_pager_putpages: residual I/O %d at %lu\n",
1087 auio.uio_resid, (u_long)m[0]->pindex);
1088 }
1089 for (i = 0; i < ncount; i++) {
1090 rtvals[i] = VM_PAGER_OK;
1091 }
1092 return rtvals[0];
1093 }
1094
1095 struct vnode *
1096 vnode_pager_lock(vm_object_t first_object)
1097 {
1098 struct vnode *vp;
1099 vm_object_t backing_object, object;
1100
1101 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
1102 for (object = first_object; object != NULL; object = backing_object) {
1103 if (object->type != OBJT_VNODE) {
1104 if ((backing_object = object->backing_object) != NULL)
1105 VM_OBJECT_LOCK(backing_object);
1106 if (object != first_object)
1107 VM_OBJECT_UNLOCK(object);
1108 continue;
1109 }
1110 retry:
1111 if (object->flags & OBJ_DEAD) {
1112 if (object != first_object)
1113 VM_OBJECT_UNLOCK(object);
1114 return NULL;
1115 }
1116 vp = object->handle;
1117 VI_LOCK(vp);
1118 VM_OBJECT_UNLOCK(object);
1119 if (first_object != object)
1120 VM_OBJECT_UNLOCK(first_object);
1121 if (vget(vp, LK_CANRECURSE | LK_INTERLOCK | LK_NOPAUSE |
1122 LK_RETRY | LK_SHARED, curthread)) {
1123 VM_OBJECT_LOCK(first_object);
1124 if (object != first_object)
1125 VM_OBJECT_LOCK(object);
1126 if (object->type != OBJT_VNODE) {
1127 if (object != first_object)
1128 VM_OBJECT_UNLOCK(object);
1129 return NULL;
1130 }
1131 printf("vnode_pager_lock: retrying\n");
1132 goto retry;
1133 }
1134 VM_OBJECT_LOCK(first_object);
1135 return (vp);
1136 }
1137 return NULL;
1138 }
Cache object: efd749da866917406e876d28a408419d
|