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 * $FreeBSD: releng/5.1/sys/vm/vnode_pager.c 114774 2003-05-06 02:45:28Z alc $
42 */
43
44 /*
45 * Page to/from files (vnodes).
46 */
47
48 /*
49 * TODO:
50 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
51 * greatly re-simplify the vnode_pager.
52 */
53
54 #include <sys/param.h>
55 #include <sys/systm.h>
56 #include <sys/proc.h>
57 #include <sys/vnode.h>
58 #include <sys/mount.h>
59 #include <sys/bio.h>
60 #include <sys/buf.h>
61 #include <sys/vmmeter.h>
62 #include <sys/conf.h>
63
64 #include <vm/vm.h>
65 #include <vm/vm_object.h>
66 #include <vm/vm_page.h>
67 #include <vm/vm_pager.h>
68 #include <vm/vm_map.h>
69 #include <vm/vnode_pager.h>
70 #include <vm/vm_extern.h>
71
72 static void vnode_pager_init(void);
73 static vm_offset_t vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
74 int *run);
75 static void vnode_pager_iodone(struct buf *bp);
76 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
77 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
78 static void vnode_pager_dealloc(vm_object_t);
79 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int);
80 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
81 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
82
83 struct pagerops vnodepagerops = {
84 vnode_pager_init,
85 vnode_pager_alloc,
86 vnode_pager_dealloc,
87 vnode_pager_getpages,
88 vnode_pager_putpages,
89 vnode_pager_haspage,
90 NULL
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 GIANT_REQUIRED;
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 err = VOP_BMAP(vp, reqblock, (struct vnode **) 0, &bn,
247 after, before);
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_pindex_t nobjsize;
296 vm_object_t object = vp->v_object;
297
298 GIANT_REQUIRED;
299
300 if (object == NULL)
301 return;
302
303 /*
304 * Hasn't changed size
305 */
306 if (nsize == object->un_pager.vnp.vnp_size)
307 return;
308
309 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
310
311 /*
312 * File has shrunk. Toss any cached pages beyond the new EOF.
313 */
314 if (nsize < object->un_pager.vnp.vnp_size) {
315 if (nobjsize < object->size) {
316 VM_OBJECT_LOCK(object);
317 vm_object_page_remove(object, nobjsize, object->size,
318 FALSE);
319 VM_OBJECT_UNLOCK(object);
320 }
321 /*
322 * this gets rid of garbage at the end of a page that is now
323 * only partially backed by the vnode.
324 *
325 * XXX for some reason (I don't know yet), if we take a
326 * completely invalid page and mark it partially valid
327 * it can screw up NFS reads, so we don't allow the case.
328 */
329 if (nsize & PAGE_MASK) {
330 vm_page_t m;
331
332 m = vm_page_lookup(object, OFF_TO_IDX(nsize));
333 if (m && m->valid) {
334 int base = (int)nsize & PAGE_MASK;
335 int size = PAGE_SIZE - base;
336
337 /*
338 * Clear out partial-page garbage in case
339 * the page has been mapped.
340 */
341 pmap_zero_page_area(m, base, size);
342
343 vm_page_lock_queues();
344 /*
345 * XXX work around SMP data integrity race
346 * by unmapping the page from user processes.
347 * The garbage we just cleared may be mapped
348 * to a user process running on another cpu
349 * and this code is not running through normal
350 * I/O channels which handle SMP issues for
351 * us, so unmap page to synchronize all cpus.
352 *
353 * XXX should vm_pager_unmap_page() have
354 * dealt with this?
355 */
356 pmap_remove_all(m);
357
358 /*
359 * Clear out partial-page dirty bits. This
360 * has the side effect of setting the valid
361 * bits, but that is ok. There are a bunch
362 * of places in the VM system where we expected
363 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
364 * case is one of them. If the page is still
365 * partially dirty, make it fully dirty.
366 *
367 * note that we do not clear out the valid
368 * bits. This would prevent bogus_page
369 * replacement from working properly.
370 */
371 vm_page_set_validclean(m, base, size);
372 if (m->dirty != 0)
373 m->dirty = VM_PAGE_BITS_ALL;
374 vm_page_unlock_queues();
375 }
376 }
377 }
378 object->un_pager.vnp.vnp_size = nsize;
379 object->size = nobjsize;
380 }
381
382 /*
383 * calculate the linear (byte) disk address of specified virtual
384 * file address
385 */
386 static vm_offset_t
387 vnode_pager_addr(vp, address, run)
388 struct vnode *vp;
389 vm_ooffset_t address;
390 int *run;
391 {
392 int rtaddress;
393 int bsize;
394 daddr_t block;
395 struct vnode *rtvp;
396 int err;
397 daddr_t vblock;
398 int voffset;
399
400 GIANT_REQUIRED;
401 if ((int) address < 0)
402 return -1;
403
404 if (vp->v_mount == NULL)
405 return -1;
406
407 bsize = vp->v_mount->mnt_stat.f_iosize;
408 vblock = address / bsize;
409 voffset = address % bsize;
410
411 err = VOP_BMAP(vp, vblock, &rtvp, &block, run, NULL);
412
413 if (err || (block == -1))
414 rtaddress = -1;
415 else {
416 rtaddress = block + voffset / DEV_BSIZE;
417 if (run) {
418 *run += 1;
419 *run *= bsize/PAGE_SIZE;
420 *run -= voffset/PAGE_SIZE;
421 }
422 }
423
424 return rtaddress;
425 }
426
427 /*
428 * interrupt routine for I/O completion
429 */
430 static void
431 vnode_pager_iodone(bp)
432 struct buf *bp;
433 {
434 bp->b_flags |= B_DONE;
435 wakeup(bp);
436 }
437
438 /*
439 * small block filesystem vnode pager input
440 */
441 static int
442 vnode_pager_input_smlfs(object, m)
443 vm_object_t object;
444 vm_page_t m;
445 {
446 int i;
447 int s;
448 struct vnode *dp, *vp;
449 struct buf *bp;
450 vm_offset_t kva;
451 int fileaddr;
452 vm_offset_t bsize;
453 int error = 0;
454
455 GIANT_REQUIRED;
456
457 vp = object->handle;
458 if (vp->v_mount == NULL)
459 return VM_PAGER_BAD;
460
461 bsize = vp->v_mount->mnt_stat.f_iosize;
462
463 VOP_BMAP(vp, 0, &dp, 0, NULL, NULL);
464
465 kva = vm_pager_map_page(m);
466
467 for (i = 0; i < PAGE_SIZE / bsize; i++) {
468 vm_ooffset_t address;
469
470 if (vm_page_bits(i * bsize, bsize) & m->valid)
471 continue;
472
473 address = IDX_TO_OFF(m->pindex) + i * bsize;
474 if (address >= object->un_pager.vnp.vnp_size) {
475 fileaddr = -1;
476 } else {
477 fileaddr = vnode_pager_addr(vp, address, NULL);
478 }
479 if (fileaddr != -1) {
480 bp = getpbuf(&vnode_pbuf_freecnt);
481
482 /* build a minimal buffer header */
483 bp->b_iocmd = BIO_READ;
484 bp->b_iodone = vnode_pager_iodone;
485 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
486 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
487 bp->b_rcred = crhold(curthread->td_ucred);
488 bp->b_wcred = crhold(curthread->td_ucred);
489 bp->b_data = (caddr_t) kva + i * bsize;
490 bp->b_blkno = fileaddr;
491 pbgetvp(dp, bp);
492 bp->b_bcount = bsize;
493 bp->b_bufsize = bsize;
494 bp->b_runningbufspace = bp->b_bufsize;
495 runningbufspace += bp->b_runningbufspace;
496
497 /* do the input */
498 VOP_SPECSTRATEGY(bp->b_vp, bp);
499
500 /* we definitely need to be at splvm here */
501
502 s = splvm();
503 while ((bp->b_flags & B_DONE) == 0) {
504 tsleep(bp, PVM, "vnsrd", 0);
505 }
506 splx(s);
507 if ((bp->b_ioflags & BIO_ERROR) != 0)
508 error = EIO;
509
510 /*
511 * free the buffer header back to the swap buffer pool
512 */
513 relpbuf(bp, &vnode_pbuf_freecnt);
514 if (error)
515 break;
516
517 vm_page_lock_queues();
518 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
519 vm_page_unlock_queues();
520 } else {
521 vm_page_lock_queues();
522 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
523 vm_page_unlock_queues();
524 bzero((caddr_t) kva + i * bsize, bsize);
525 }
526 }
527 vm_pager_unmap_page(kva);
528 vm_page_lock_queues();
529 pmap_clear_modify(m);
530 vm_page_flag_clear(m, PG_ZERO);
531 vm_page_unlock_queues();
532 if (error) {
533 return VM_PAGER_ERROR;
534 }
535 return VM_PAGER_OK;
536
537 }
538
539
540 /*
541 * old style vnode pager output routine
542 */
543 static int
544 vnode_pager_input_old(object, m)
545 vm_object_t object;
546 vm_page_t m;
547 {
548 struct uio auio;
549 struct iovec aiov;
550 int error;
551 int size;
552 vm_offset_t kva;
553 struct vnode *vp;
554
555 GIANT_REQUIRED;
556 error = 0;
557
558 /*
559 * Return failure if beyond current EOF
560 */
561 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
562 return VM_PAGER_BAD;
563 } else {
564 size = PAGE_SIZE;
565 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
566 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
567
568 /*
569 * Allocate a kernel virtual address and initialize so that
570 * we can use VOP_READ/WRITE routines.
571 */
572 kva = vm_pager_map_page(m);
573
574 vp = object->handle;
575 aiov.iov_base = (caddr_t) kva;
576 aiov.iov_len = size;
577 auio.uio_iov = &aiov;
578 auio.uio_iovcnt = 1;
579 auio.uio_offset = IDX_TO_OFF(m->pindex);
580 auio.uio_segflg = UIO_SYSSPACE;
581 auio.uio_rw = UIO_READ;
582 auio.uio_resid = size;
583 auio.uio_td = curthread;
584
585 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
586 if (!error) {
587 int count = size - auio.uio_resid;
588
589 if (count == 0)
590 error = EINVAL;
591 else if (count != PAGE_SIZE)
592 bzero((caddr_t) kva + count, PAGE_SIZE - count);
593 }
594 vm_pager_unmap_page(kva);
595 }
596 vm_page_lock_queues();
597 pmap_clear_modify(m);
598 vm_page_undirty(m);
599 vm_page_flag_clear(m, PG_ZERO);
600 if (!error)
601 m->valid = VM_PAGE_BITS_ALL;
602 vm_page_unlock_queues();
603 return error ? VM_PAGER_ERROR : VM_PAGER_OK;
604 }
605
606 /*
607 * generic vnode pager input routine
608 */
609
610 /*
611 * Local media VFS's that do not implement their own VOP_GETPAGES
612 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
613 * to implement the previous behaviour.
614 *
615 * All other FS's should use the bypass to get to the local media
616 * backing vp's VOP_GETPAGES.
617 */
618 static int
619 vnode_pager_getpages(object, m, count, reqpage)
620 vm_object_t object;
621 vm_page_t *m;
622 int count;
623 int reqpage;
624 {
625 int rtval;
626 struct vnode *vp;
627 int bytes = count * PAGE_SIZE;
628
629 GIANT_REQUIRED;
630 vp = object->handle;
631 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
632 KASSERT(rtval != EOPNOTSUPP,
633 ("vnode_pager: FS getpages not implemented\n"));
634 return rtval;
635 }
636
637 /*
638 * This is now called from local media FS's to operate against their
639 * own vnodes if they fail to implement VOP_GETPAGES.
640 */
641 int
642 vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
643 struct vnode *vp;
644 vm_page_t *m;
645 int bytecount;
646 int reqpage;
647 {
648 vm_object_t object;
649 vm_offset_t kva;
650 off_t foff, tfoff, nextoff;
651 int i, j, size, bsize, first, firstaddr;
652 struct vnode *dp;
653 int runpg;
654 int runend;
655 struct buf *bp;
656 int s;
657 int count;
658 int error = 0;
659
660 GIANT_REQUIRED;
661 object = vp->v_object;
662 count = bytecount / PAGE_SIZE;
663
664 if (vp->v_mount == NULL)
665 return VM_PAGER_BAD;
666
667 bsize = vp->v_mount->mnt_stat.f_iosize;
668
669 /* get the UNDERLYING device for the file with VOP_BMAP() */
670
671 /*
672 * originally, we did not check for an error return value -- assuming
673 * an fs always has a bmap entry point -- that assumption is wrong!!!
674 */
675 foff = IDX_TO_OFF(m[reqpage]->pindex);
676
677 /*
678 * if we can't bmap, use old VOP code
679 */
680 if (VOP_BMAP(vp, 0, &dp, 0, NULL, NULL)) {
681 vm_page_lock_queues();
682 for (i = 0; i < count; i++)
683 if (i != reqpage)
684 vm_page_free(m[i]);
685 vm_page_unlock_queues();
686 cnt.v_vnodein++;
687 cnt.v_vnodepgsin++;
688 return vnode_pager_input_old(object, m[reqpage]);
689
690 /*
691 * if the blocksize is smaller than a page size, then use
692 * special small filesystem code. NFS sometimes has a small
693 * blocksize, but it can handle large reads itself.
694 */
695 } else if ((PAGE_SIZE / bsize) > 1 &&
696 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
697 vm_page_lock_queues();
698 for (i = 0; i < count; i++)
699 if (i != reqpage)
700 vm_page_free(m[i]);
701 vm_page_unlock_queues();
702 cnt.v_vnodein++;
703 cnt.v_vnodepgsin++;
704 return vnode_pager_input_smlfs(object, m[reqpage]);
705 }
706
707 /*
708 * If we have a completely valid page available to us, we can
709 * clean up and return. Otherwise we have to re-read the
710 * media.
711 */
712 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
713 vm_page_lock_queues();
714 for (i = 0; i < count; i++)
715 if (i != reqpage)
716 vm_page_free(m[i]);
717 vm_page_unlock_queues();
718 return VM_PAGER_OK;
719 }
720 m[reqpage]->valid = 0;
721
722 /*
723 * here on direct device I/O
724 */
725 firstaddr = -1;
726
727 /*
728 * calculate the run that includes the required page
729 */
730 for (first = 0, i = 0; i < count; i = runend) {
731 firstaddr = vnode_pager_addr(vp,
732 IDX_TO_OFF(m[i]->pindex), &runpg);
733 if (firstaddr == -1) {
734 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
735 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %d, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
736 firstaddr, (uintmax_t)(foff >> 32),
737 (uintmax_t)foff,
738 (uintmax_t)
739 (object->un_pager.vnp.vnp_size >> 32),
740 (uintmax_t)object->un_pager.vnp.vnp_size);
741 }
742 vm_page_lock_queues();
743 vm_page_free(m[i]);
744 vm_page_unlock_queues();
745 runend = i + 1;
746 first = runend;
747 continue;
748 }
749 runend = i + runpg;
750 if (runend <= reqpage) {
751 vm_page_lock_queues();
752 for (j = i; j < runend; j++)
753 vm_page_free(m[j]);
754 vm_page_unlock_queues();
755 } else {
756 if (runpg < (count - first)) {
757 vm_page_lock_queues();
758 for (i = first + runpg; i < count; i++)
759 vm_page_free(m[i]);
760 vm_page_unlock_queues();
761 count = first + runpg;
762 }
763 break;
764 }
765 first = runend;
766 }
767
768 /*
769 * the first and last page have been calculated now, move input pages
770 * to be zero based...
771 */
772 if (first != 0) {
773 for (i = first; i < count; i++) {
774 m[i - first] = m[i];
775 }
776 count -= first;
777 reqpage -= first;
778 }
779
780 /*
781 * calculate the file virtual address for the transfer
782 */
783 foff = IDX_TO_OFF(m[0]->pindex);
784
785 /*
786 * calculate the size of the transfer
787 */
788 size = count * PAGE_SIZE;
789 if ((foff + size) > object->un_pager.vnp.vnp_size)
790 size = object->un_pager.vnp.vnp_size - foff;
791
792 /*
793 * round up physical size for real devices.
794 */
795 if (dp->v_type == VBLK || dp->v_type == VCHR) {
796 int secmask = dp->v_rdev->si_bsize_phys - 1;
797 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
798 size = (size + secmask) & ~secmask;
799 }
800
801 bp = getpbuf(&vnode_pbuf_freecnt);
802 kva = (vm_offset_t) bp->b_data;
803
804 /*
805 * and map the pages to be read into the kva
806 */
807 pmap_qenter(kva, m, count);
808
809 /* build a minimal buffer header */
810 bp->b_iocmd = BIO_READ;
811 bp->b_iodone = vnode_pager_iodone;
812 /* B_PHYS is not set, but it is nice to fill this in */
813 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
814 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
815 bp->b_rcred = crhold(curthread->td_ucred);
816 bp->b_wcred = crhold(curthread->td_ucred);
817 bp->b_blkno = firstaddr;
818 pbgetvp(dp, bp);
819 bp->b_bcount = size;
820 bp->b_bufsize = size;
821 bp->b_runningbufspace = bp->b_bufsize;
822 runningbufspace += bp->b_runningbufspace;
823
824 cnt.v_vnodein++;
825 cnt.v_vnodepgsin += count;
826
827 /* do the input */
828 if (dp->v_type == VCHR)
829 VOP_SPECSTRATEGY(bp->b_vp, bp);
830 else
831 VOP_STRATEGY(bp->b_vp, bp);
832
833 s = splvm();
834 /* we definitely need to be at splvm here */
835
836 while ((bp->b_flags & B_DONE) == 0) {
837 tsleep(bp, PVM, "vnread", 0);
838 }
839 splx(s);
840 if ((bp->b_ioflags & BIO_ERROR) != 0)
841 error = EIO;
842
843 if (!error) {
844 if (size != count * PAGE_SIZE)
845 bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
846 }
847 pmap_qremove(kva, count);
848
849 /*
850 * free the buffer header back to the swap buffer pool
851 */
852 relpbuf(bp, &vnode_pbuf_freecnt);
853
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 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 if (vp->v_type != VREG)
960 mp = NULL;
961 (void)vn_start_write(vp, &mp, V_WAIT);
962 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
963 KASSERT(rtval != EOPNOTSUPP,
964 ("vnode_pager: stale FS putpages\n"));
965 vn_finished_write(mp);
966 }
967
968
969 /*
970 * This is now called from local media FS's to operate against their
971 * own vnodes if they fail to implement VOP_PUTPAGES.
972 *
973 * This is typically called indirectly via the pageout daemon and
974 * clustering has already typically occured, so in general we ask the
975 * underlying filesystem to write the data out asynchronously rather
976 * then delayed.
977 */
978 int
979 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals)
980 struct vnode *vp;
981 vm_page_t *m;
982 int bytecount;
983 int flags;
984 int *rtvals;
985 {
986 int i;
987 vm_object_t object;
988 int count;
989
990 int maxsize, ncount;
991 vm_ooffset_t poffset;
992 struct uio auio;
993 struct iovec aiov;
994 int error;
995 int ioflags;
996
997 GIANT_REQUIRED;
998 object = vp->v_object;
999 count = bytecount / PAGE_SIZE;
1000
1001 for (i = 0; i < count; i++)
1002 rtvals[i] = VM_PAGER_AGAIN;
1003
1004 if ((int) m[0]->pindex < 0) {
1005 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
1006 (long)m[0]->pindex, m[0]->dirty);
1007 rtvals[0] = VM_PAGER_BAD;
1008 return VM_PAGER_BAD;
1009 }
1010
1011 maxsize = count * PAGE_SIZE;
1012 ncount = count;
1013
1014 poffset = IDX_TO_OFF(m[0]->pindex);
1015
1016 /*
1017 * If the page-aligned write is larger then the actual file we
1018 * have to invalidate pages occuring beyond the file EOF. However,
1019 * there is an edge case where a file may not be page-aligned where
1020 * the last page is partially invalid. In this case the filesystem
1021 * may not properly clear the dirty bits for the entire page (which
1022 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1023 * With the page locked we are free to fix-up the dirty bits here.
1024 *
1025 * We do not under any circumstances truncate the valid bits, as
1026 * this will screw up bogus page replacement.
1027 */
1028 if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1029 if (object->un_pager.vnp.vnp_size > poffset) {
1030 int pgoff;
1031
1032 maxsize = object->un_pager.vnp.vnp_size - poffset;
1033 ncount = btoc(maxsize);
1034 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1035 vm_page_clear_dirty(m[ncount - 1], pgoff,
1036 PAGE_SIZE - pgoff);
1037 }
1038 } else {
1039 maxsize = 0;
1040 ncount = 0;
1041 }
1042 if (ncount < count) {
1043 for (i = ncount; i < count; i++) {
1044 rtvals[i] = VM_PAGER_BAD;
1045 }
1046 }
1047 }
1048
1049 /*
1050 * pageouts are already clustered, use IO_ASYNC t o force a bawrite()
1051 * rather then a bdwrite() to prevent paging I/O from saturating
1052 * the buffer cache. Dummy-up the sequential heuristic to cause
1053 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
1054 * the system decides how to cluster.
1055 */
1056 ioflags = IO_VMIO;
1057 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1058 ioflags |= IO_SYNC;
1059 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1060 ioflags |= IO_ASYNC;
1061 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1062 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1063
1064 aiov.iov_base = (caddr_t) 0;
1065 aiov.iov_len = maxsize;
1066 auio.uio_iov = &aiov;
1067 auio.uio_iovcnt = 1;
1068 auio.uio_offset = poffset;
1069 auio.uio_segflg = UIO_NOCOPY;
1070 auio.uio_rw = UIO_WRITE;
1071 auio.uio_resid = maxsize;
1072 auio.uio_td = (struct thread *) 0;
1073 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1074 cnt.v_vnodeout++;
1075 cnt.v_vnodepgsout += ncount;
1076
1077 if (error) {
1078 printf("vnode_pager_putpages: I/O error %d\n", error);
1079 }
1080 if (auio.uio_resid) {
1081 printf("vnode_pager_putpages: residual I/O %d at %lu\n",
1082 auio.uio_resid, (u_long)m[0]->pindex);
1083 }
1084 for (i = 0; i < ncount; i++) {
1085 rtvals[i] = VM_PAGER_OK;
1086 }
1087 return rtvals[0];
1088 }
1089
1090 struct vnode *
1091 vnode_pager_lock(object)
1092 vm_object_t object;
1093 {
1094 struct thread *td = curthread; /* XXX */
1095
1096 GIANT_REQUIRED;
1097
1098 for (; object != NULL; object = object->backing_object) {
1099 if (object->type != OBJT_VNODE)
1100 continue;
1101 if (object->flags & OBJ_DEAD) {
1102 return NULL;
1103 }
1104
1105 /* XXX; If object->handle can change, we need to cache it. */
1106 while (vget(object->handle,
1107 LK_NOPAUSE | LK_SHARED | LK_RETRY | LK_CANRECURSE, td)){
1108 if ((object->flags & OBJ_DEAD) || (object->type != OBJT_VNODE))
1109 return NULL;
1110 printf("vnode_pager_lock: retrying\n");
1111 }
1112 return object->handle;
1113 }
1114 return NULL;
1115 }
Cache object: 4c34820c2f2e1434023ba4f3c1cb2b19
|