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/5.3/sys/vm/vnode_pager.c 128992 2004-05-06 05:03:23Z alc $");
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 #include <sys/sf_buf.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/vm_extern.h>
74
75 static void vnode_pager_init(void);
76 static vm_offset_t vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
77 int *run);
78 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
79 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
80 static void vnode_pager_dealloc(vm_object_t);
81 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int);
82 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
83 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
84
85 struct pagerops vnodepagerops = {
86 .pgo_init = vnode_pager_init,
87 .pgo_alloc = vnode_pager_alloc,
88 .pgo_dealloc = vnode_pager_dealloc,
89 .pgo_getpages = vnode_pager_getpages,
90 .pgo_putpages = vnode_pager_putpages,
91 .pgo_haspage = vnode_pager_haspage,
92 };
93
94 int vnode_pbuf_freecnt;
95
96 static void
97 vnode_pager_init(void)
98 {
99
100 vnode_pbuf_freecnt = nswbuf / 2 + 1;
101 }
102
103 /*
104 * Allocate (or lookup) pager for a vnode.
105 * Handle is a vnode pointer.
106 *
107 * MPSAFE
108 */
109 vm_object_t
110 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
111 vm_ooffset_t offset)
112 {
113 vm_object_t object;
114 struct vnode *vp;
115
116 /*
117 * Pageout to vnode, no can do yet.
118 */
119 if (handle == NULL)
120 return (NULL);
121
122 vp = (struct vnode *) handle;
123
124 ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc");
125
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 return (object);
175 }
176
177 /*
178 * The object must be locked.
179 */
180 static void
181 vnode_pager_dealloc(object)
182 vm_object_t object;
183 {
184 struct vnode *vp = object->handle;
185
186 if (vp == NULL)
187 panic("vnode_pager_dealloc: pager already dealloced");
188
189 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
190 vm_object_pip_wait(object, "vnpdea");
191
192 object->handle = NULL;
193 object->type = OBJT_DEAD;
194 ASSERT_VOP_LOCKED(vp, "vnode_pager_dealloc");
195 vp->v_object = NULL;
196 vp->v_vflag &= ~(VV_TEXT | VV_OBJBUF);
197 }
198
199 static boolean_t
200 vnode_pager_haspage(object, pindex, before, after)
201 vm_object_t object;
202 vm_pindex_t pindex;
203 int *before;
204 int *after;
205 {
206 struct vnode *vp = object->handle;
207 daddr_t bn;
208 int err;
209 daddr_t reqblock;
210 int poff;
211 int bsize;
212 int pagesperblock, blocksperpage;
213
214 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
215 /*
216 * If no vp or vp is doomed or marked transparent to VM, we do not
217 * have the page.
218 */
219 if (vp == NULL)
220 return FALSE;
221
222 VI_LOCK(vp);
223 if (vp->v_iflag & VI_DOOMED) {
224 VI_UNLOCK(vp);
225 return FALSE;
226 }
227 VI_UNLOCK(vp);
228 /*
229 * If filesystem no longer mounted or offset beyond end of file we do
230 * not have the page.
231 */
232 if ((vp->v_mount == NULL) ||
233 (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size))
234 return FALSE;
235
236 bsize = vp->v_mount->mnt_stat.f_iosize;
237 pagesperblock = bsize / PAGE_SIZE;
238 blocksperpage = 0;
239 if (pagesperblock > 0) {
240 reqblock = pindex / pagesperblock;
241 } else {
242 blocksperpage = (PAGE_SIZE / bsize);
243 reqblock = pindex * blocksperpage;
244 }
245 VM_OBJECT_UNLOCK(object);
246 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
247 VM_OBJECT_LOCK(object);
248 if (err)
249 return TRUE;
250 if (bn == -1)
251 return FALSE;
252 if (pagesperblock > 0) {
253 poff = pindex - (reqblock * pagesperblock);
254 if (before) {
255 *before *= pagesperblock;
256 *before += poff;
257 }
258 if (after) {
259 int numafter;
260 *after *= pagesperblock;
261 numafter = pagesperblock - (poff + 1);
262 if (IDX_TO_OFF(pindex + numafter) >
263 object->un_pager.vnp.vnp_size) {
264 numafter =
265 OFF_TO_IDX(object->un_pager.vnp.vnp_size) -
266 pindex;
267 }
268 *after += numafter;
269 }
270 } else {
271 if (before) {
272 *before /= blocksperpage;
273 }
274
275 if (after) {
276 *after /= blocksperpage;
277 }
278 }
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 void
291 vnode_pager_setsize(vp, nsize)
292 struct vnode *vp;
293 vm_ooffset_t nsize;
294 {
295 vm_object_t object;
296 vm_page_t m;
297 vm_pindex_t nobjsize;
298
299 if ((object = vp->v_object) == NULL)
300 return;
301 VM_OBJECT_LOCK(object);
302 if (nsize == object->un_pager.vnp.vnp_size) {
303 /*
304 * Hasn't changed size
305 */
306 VM_OBJECT_UNLOCK(object);
307 return;
308 }
309 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
310 if (nsize < object->un_pager.vnp.vnp_size) {
311 /*
312 * File has shrunk. Toss any cached pages beyond the new EOF.
313 */
314 if (nobjsize < object->size)
315 vm_object_page_remove(object, nobjsize, object->size,
316 FALSE);
317 /*
318 * this gets rid of garbage at the end of a page that is now
319 * only partially backed by the vnode.
320 *
321 * XXX for some reason (I don't know yet), if we take a
322 * completely invalid page and mark it partially valid
323 * it can screw up NFS reads, so we don't allow the case.
324 */
325 if ((nsize & PAGE_MASK) &&
326 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
327 m->valid != 0) {
328 int base = (int)nsize & PAGE_MASK;
329 int size = PAGE_SIZE - base;
330
331 /*
332 * Clear out partial-page garbage in case
333 * the page has been mapped.
334 */
335 pmap_zero_page_area(m, base, size);
336
337 /*
338 * XXX work around SMP data integrity race
339 * by unmapping the page from user processes.
340 * The garbage we just cleared may be mapped
341 * to a user process running on another cpu
342 * and this code is not running through normal
343 * I/O channels which handle SMP issues for
344 * us, so unmap page to synchronize all cpus.
345 *
346 * XXX should vm_pager_unmap_page() have
347 * dealt with this?
348 */
349 vm_page_lock_queues();
350 pmap_remove_all(m);
351
352 /*
353 * Clear out partial-page dirty bits. This
354 * has the side effect of setting the valid
355 * bits, but that is ok. There are a bunch
356 * of places in the VM system where we expected
357 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
358 * case is one of them. If the page is still
359 * partially dirty, make it fully dirty.
360 *
361 * note that we do not clear out the valid
362 * bits. This would prevent bogus_page
363 * replacement from working properly.
364 */
365 vm_page_set_validclean(m, base, size);
366 if (m->dirty != 0)
367 m->dirty = VM_PAGE_BITS_ALL;
368 vm_page_unlock_queues();
369 }
370 }
371 object->un_pager.vnp.vnp_size = nsize;
372 object->size = nobjsize;
373 VM_OBJECT_UNLOCK(object);
374 }
375
376 /*
377 * calculate the linear (byte) disk address of specified virtual
378 * file address
379 */
380 static vm_offset_t
381 vnode_pager_addr(vp, address, run)
382 struct vnode *vp;
383 vm_ooffset_t address;
384 int *run;
385 {
386 int rtaddress;
387 int bsize;
388 daddr_t block;
389 int err;
390 daddr_t vblock;
391 int voffset;
392
393 GIANT_REQUIRED;
394 if ((int) address < 0)
395 return -1;
396
397 if (vp->v_mount == NULL)
398 return -1;
399
400 bsize = vp->v_mount->mnt_stat.f_iosize;
401 vblock = address / bsize;
402 voffset = address % bsize;
403
404 err = VOP_BMAP(vp, vblock, NULL, &block, run, NULL);
405
406 if (err || (block == -1))
407 rtaddress = -1;
408 else {
409 rtaddress = block + voffset / DEV_BSIZE;
410 if (run) {
411 *run += 1;
412 *run *= bsize/PAGE_SIZE;
413 *run -= voffset/PAGE_SIZE;
414 }
415 }
416
417 return rtaddress;
418 }
419
420 /*
421 * small block filesystem vnode pager input
422 */
423 static int
424 vnode_pager_input_smlfs(object, m)
425 vm_object_t object;
426 vm_page_t m;
427 {
428 int i;
429 struct vnode *dp, *vp;
430 struct buf *bp;
431 struct sf_buf *sf;
432 int fileaddr;
433 vm_offset_t bsize;
434 int error = 0;
435
436 GIANT_REQUIRED;
437
438 vp = object->handle;
439 if (vp->v_mount == NULL)
440 return VM_PAGER_BAD;
441
442 bsize = vp->v_mount->mnt_stat.f_iosize;
443
444 VOP_BMAP(vp, 0, &dp, 0, NULL, NULL);
445
446 sf = sf_buf_alloc(m, 0);
447
448 for (i = 0; i < PAGE_SIZE / bsize; i++) {
449 vm_ooffset_t address;
450
451 if (vm_page_bits(i * bsize, bsize) & m->valid)
452 continue;
453
454 address = IDX_TO_OFF(m->pindex) + i * bsize;
455 if (address >= object->un_pager.vnp.vnp_size) {
456 fileaddr = -1;
457 } else {
458 fileaddr = vnode_pager_addr(vp, address, NULL);
459 }
460 if (fileaddr != -1) {
461 bp = getpbuf(&vnode_pbuf_freecnt);
462
463 /* build a minimal buffer header */
464 bp->b_iocmd = BIO_READ;
465 bp->b_iodone = bdone;
466 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
467 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
468 bp->b_rcred = crhold(curthread->td_ucred);
469 bp->b_wcred = crhold(curthread->td_ucred);
470 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
471 bp->b_blkno = fileaddr;
472 pbgetvp(dp, bp);
473 bp->b_bcount = bsize;
474 bp->b_bufsize = bsize;
475 bp->b_runningbufspace = bp->b_bufsize;
476 runningbufspace += bp->b_runningbufspace;
477
478 /* do the input */
479 bp->b_iooffset = dbtob(bp->b_blkno);
480 if (dp->v_type == VCHR)
481 VOP_SPECSTRATEGY(bp->b_vp, bp);
482 else
483 VOP_STRATEGY(bp->b_vp, bp);
484
485 /* we definitely need to be at splvm here */
486
487 bwait(bp, PVM, "vnsrd");
488
489 if ((bp->b_ioflags & BIO_ERROR) != 0)
490 error = EIO;
491
492 /*
493 * free the buffer header back to the swap buffer pool
494 */
495 relpbuf(bp, &vnode_pbuf_freecnt);
496 if (error)
497 break;
498
499 VM_OBJECT_LOCK(object);
500 vm_page_lock_queues();
501 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
502 vm_page_unlock_queues();
503 VM_OBJECT_UNLOCK(object);
504 } else {
505 VM_OBJECT_LOCK(object);
506 vm_page_lock_queues();
507 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
508 vm_page_unlock_queues();
509 VM_OBJECT_UNLOCK(object);
510 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
511 }
512 }
513 sf_buf_free(sf);
514 vm_page_lock_queues();
515 pmap_clear_modify(m);
516 vm_page_unlock_queues();
517 if (error) {
518 return VM_PAGER_ERROR;
519 }
520 return VM_PAGER_OK;
521
522 }
523
524
525 /*
526 * old style vnode pager output routine
527 */
528 static int
529 vnode_pager_input_old(object, m)
530 vm_object_t object;
531 vm_page_t m;
532 {
533 struct uio auio;
534 struct iovec aiov;
535 int error;
536 int size;
537 struct sf_buf *sf;
538 struct vnode *vp;
539
540 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
541 error = 0;
542
543 /*
544 * Return failure if beyond current EOF
545 */
546 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
547 return VM_PAGER_BAD;
548 } else {
549 size = PAGE_SIZE;
550 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
551 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
552 vp = object->handle;
553 VM_OBJECT_UNLOCK(object);
554
555 /*
556 * Allocate a kernel virtual address and initialize so that
557 * we can use VOP_READ/WRITE routines.
558 */
559 sf = sf_buf_alloc(m, 0);
560
561 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
562 aiov.iov_len = size;
563 auio.uio_iov = &aiov;
564 auio.uio_iovcnt = 1;
565 auio.uio_offset = IDX_TO_OFF(m->pindex);
566 auio.uio_segflg = UIO_SYSSPACE;
567 auio.uio_rw = UIO_READ;
568 auio.uio_resid = size;
569 auio.uio_td = curthread;
570
571 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
572 if (!error) {
573 int count = size - auio.uio_resid;
574
575 if (count == 0)
576 error = EINVAL;
577 else if (count != PAGE_SIZE)
578 bzero((caddr_t)sf_buf_kva(sf) + count,
579 PAGE_SIZE - count);
580 }
581 sf_buf_free(sf);
582
583 VM_OBJECT_LOCK(object);
584 }
585 vm_page_lock_queues();
586 pmap_clear_modify(m);
587 vm_page_undirty(m);
588 vm_page_unlock_queues();
589 if (!error)
590 m->valid = VM_PAGE_BITS_ALL;
591 return error ? VM_PAGER_ERROR : VM_PAGER_OK;
592 }
593
594 /*
595 * generic vnode pager input routine
596 */
597
598 /*
599 * Local media VFS's that do not implement their own VOP_GETPAGES
600 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
601 * to implement the previous behaviour.
602 *
603 * All other FS's should use the bypass to get to the local media
604 * backing vp's VOP_GETPAGES.
605 */
606 static int
607 vnode_pager_getpages(object, m, count, reqpage)
608 vm_object_t object;
609 vm_page_t *m;
610 int count;
611 int reqpage;
612 {
613 int rtval;
614 struct vnode *vp;
615 int bytes = count * PAGE_SIZE;
616
617 vp = object->handle;
618 VM_OBJECT_UNLOCK(object);
619 mtx_lock(&Giant);
620 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
621 KASSERT(rtval != EOPNOTSUPP,
622 ("vnode_pager: FS getpages not implemented\n"));
623 mtx_unlock(&Giant);
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 if (i != reqpage) {
886
887 /*
888 * whether or not to leave the page activated is up in
889 * the air, but we should put the page on a page queue
890 * somewhere. (it already is in the object). Result:
891 * It appears that empirical results show that
892 * deactivating pages is best.
893 */
894
895 /*
896 * just in case someone was asking for this page we
897 * now tell them that it is ok to use
898 */
899 if (!error) {
900 if (mt->flags & PG_WANTED)
901 vm_page_activate(mt);
902 else
903 vm_page_deactivate(mt);
904 vm_page_wakeup(mt);
905 } else {
906 vm_page_free(mt);
907 }
908 }
909 }
910 vm_page_unlock_queues();
911 VM_OBJECT_UNLOCK(object);
912 if (error) {
913 printf("vnode_pager_getpages: I/O read error\n");
914 }
915 return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
916 }
917
918 /*
919 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
920 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
921 * vnode_pager_generic_putpages() to implement the previous behaviour.
922 *
923 * All other FS's should use the bypass to get to the local media
924 * backing vp's VOP_PUTPAGES.
925 */
926 static void
927 vnode_pager_putpages(object, m, count, sync, rtvals)
928 vm_object_t object;
929 vm_page_t *m;
930 int count;
931 boolean_t sync;
932 int *rtvals;
933 {
934 int rtval;
935 struct vnode *vp;
936 struct mount *mp;
937 int bytes = count * PAGE_SIZE;
938
939 GIANT_REQUIRED;
940 /*
941 * Force synchronous operation if we are extremely low on memory
942 * to prevent a low-memory deadlock. VOP operations often need to
943 * allocate more memory to initiate the I/O ( i.e. do a BMAP
944 * operation ). The swapper handles the case by limiting the amount
945 * of asynchronous I/O, but that sort of solution doesn't scale well
946 * for the vnode pager without a lot of work.
947 *
948 * Also, the backing vnode's iodone routine may not wake the pageout
949 * daemon up. This should be probably be addressed XXX.
950 */
951
952 if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
953 sync |= OBJPC_SYNC;
954
955 /*
956 * Call device-specific putpages function
957 */
958 vp = object->handle;
959 VM_OBJECT_UNLOCK(object);
960 if (vp->v_type != VREG)
961 mp = NULL;
962 (void)vn_start_write(vp, &mp, V_WAIT);
963 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
964 KASSERT(rtval != EOPNOTSUPP,
965 ("vnode_pager: stale FS putpages\n"));
966 vn_finished_write(mp);
967 VM_OBJECT_LOCK(object);
968 }
969
970
971 /*
972 * This is now called from local media FS's to operate against their
973 * own vnodes if they fail to implement VOP_PUTPAGES.
974 *
975 * This is typically called indirectly via the pageout daemon and
976 * clustering has already typically occured, so in general we ask the
977 * underlying filesystem to write the data out asynchronously rather
978 * then delayed.
979 */
980 int
981 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals)
982 struct vnode *vp;
983 vm_page_t *m;
984 int bytecount;
985 int flags;
986 int *rtvals;
987 {
988 int i;
989 vm_object_t object;
990 int count;
991
992 int maxsize, ncount;
993 vm_ooffset_t poffset;
994 struct uio auio;
995 struct iovec aiov;
996 int error;
997 int ioflags;
998
999 GIANT_REQUIRED;
1000 object = vp->v_object;
1001 count = bytecount / PAGE_SIZE;
1002
1003 for (i = 0; i < count; i++)
1004 rtvals[i] = VM_PAGER_AGAIN;
1005
1006 if ((int) m[0]->pindex < 0) {
1007 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1008 (long)m[0]->pindex, (u_long)m[0]->dirty);
1009 rtvals[0] = VM_PAGER_BAD;
1010 return VM_PAGER_BAD;
1011 }
1012
1013 maxsize = count * PAGE_SIZE;
1014 ncount = count;
1015
1016 poffset = IDX_TO_OFF(m[0]->pindex);
1017
1018 /*
1019 * If the page-aligned write is larger then the actual file we
1020 * have to invalidate pages occuring beyond the file EOF. However,
1021 * there is an edge case where a file may not be page-aligned where
1022 * the last page is partially invalid. In this case the filesystem
1023 * may not properly clear the dirty bits for the entire page (which
1024 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1025 * With the page locked we are free to fix-up the dirty bits here.
1026 *
1027 * We do not under any circumstances truncate the valid bits, as
1028 * this will screw up bogus page replacement.
1029 */
1030 if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1031 if (object->un_pager.vnp.vnp_size > poffset) {
1032 int pgoff;
1033
1034 maxsize = object->un_pager.vnp.vnp_size - poffset;
1035 ncount = btoc(maxsize);
1036 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1037 vm_page_lock_queues();
1038 vm_page_clear_dirty(m[ncount - 1], pgoff,
1039 PAGE_SIZE - pgoff);
1040 vm_page_unlock_queues();
1041 }
1042 } else {
1043 maxsize = 0;
1044 ncount = 0;
1045 }
1046 if (ncount < count) {
1047 for (i = ncount; i < count; i++) {
1048 rtvals[i] = VM_PAGER_BAD;
1049 }
1050 }
1051 }
1052
1053 /*
1054 * pageouts are already clustered, use IO_ASYNC t o force a bawrite()
1055 * rather then a bdwrite() to prevent paging I/O from saturating
1056 * the buffer cache. Dummy-up the sequential heuristic to cause
1057 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
1058 * the system decides how to cluster.
1059 */
1060 ioflags = IO_VMIO;
1061 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1062 ioflags |= IO_SYNC;
1063 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1064 ioflags |= IO_ASYNC;
1065 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1066 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1067
1068 aiov.iov_base = (caddr_t) 0;
1069 aiov.iov_len = maxsize;
1070 auio.uio_iov = &aiov;
1071 auio.uio_iovcnt = 1;
1072 auio.uio_offset = poffset;
1073 auio.uio_segflg = UIO_NOCOPY;
1074 auio.uio_rw = UIO_WRITE;
1075 auio.uio_resid = maxsize;
1076 auio.uio_td = (struct thread *) 0;
1077 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1078 cnt.v_vnodeout++;
1079 cnt.v_vnodepgsout += ncount;
1080
1081 if (error) {
1082 printf("vnode_pager_putpages: I/O error %d\n", error);
1083 }
1084 if (auio.uio_resid) {
1085 printf("vnode_pager_putpages: residual I/O %d at %lu\n",
1086 auio.uio_resid, (u_long)m[0]->pindex);
1087 }
1088 for (i = 0; i < ncount; i++) {
1089 rtvals[i] = VM_PAGER_OK;
1090 }
1091 return rtvals[0];
1092 }
1093
1094 struct vnode *
1095 vnode_pager_lock(vm_object_t first_object)
1096 {
1097 struct vnode *vp;
1098 vm_object_t backing_object, object;
1099
1100 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
1101 for (object = first_object; object != NULL; object = backing_object) {
1102 if (object->type != OBJT_VNODE) {
1103 if ((backing_object = object->backing_object) != NULL)
1104 VM_OBJECT_LOCK(backing_object);
1105 if (object != first_object)
1106 VM_OBJECT_UNLOCK(object);
1107 continue;
1108 }
1109 retry:
1110 if (object->flags & OBJ_DEAD) {
1111 if (object != first_object)
1112 VM_OBJECT_UNLOCK(object);
1113 return NULL;
1114 }
1115 vp = object->handle;
1116 VI_LOCK(vp);
1117 VM_OBJECT_UNLOCK(object);
1118 if (first_object != object)
1119 VM_OBJECT_UNLOCK(first_object);
1120 if (vget(vp, LK_CANRECURSE | LK_INTERLOCK | LK_NOPAUSE |
1121 LK_RETRY | LK_SHARED, curthread)) {
1122 VM_OBJECT_LOCK(first_object);
1123 if (object != first_object)
1124 VM_OBJECT_LOCK(object);
1125 if (object->type != OBJT_VNODE) {
1126 if (object != first_object)
1127 VM_OBJECT_UNLOCK(object);
1128 return NULL;
1129 }
1130 printf("vnode_pager_lock: retrying\n");
1131 goto retry;
1132 }
1133 VM_OBJECT_LOCK(first_object);
1134 return (vp);
1135 }
1136 return NULL;
1137 }
Cache object: 71323a682ad828c22491202a9c518244
|