proc->thread stage 3.5: Add an IO_CORE flag so coda doesn't have to dig
[dragonfly.git] / sys / vm / vnode_pager.c
CommitLineData
984263bc
MD
1/*
2 * Copyright (c) 1990 University of Utah.
3 * Copyright (c) 1991 The Regents of the University of California.
4 * All rights reserved.
5 * Copyright (c) 1993, 1994 John S. Dyson
6 * Copyright (c) 1995, David Greenman
7 *
8 * This code is derived from software contributed to Berkeley by
9 * the Systems Programming Group of the University of Utah Computer
10 * Science Department.
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by the University of
23 * California, Berkeley and its contributors.
24 * 4. Neither the name of the University nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
27 *
28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * SUCH DAMAGE.
39 *
40 * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91
41 * $FreeBSD: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $
1de703da 42 * $DragonFly: src/sys/vm/vnode_pager.c,v 1.2 2003/06/17 04:29:00 dillon Exp $
984263bc
MD
43 */
44
45/*
46 * Page to/from files (vnodes).
47 */
48
49/*
50 * TODO:
51 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
52 * greatly re-simplify the vnode_pager.
53 */
54
55#include <sys/param.h>
56#include <sys/systm.h>
57#include <sys/proc.h>
58#include <sys/vnode.h>
59#include <sys/mount.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
72static vm_offset_t vnode_pager_addr __P((struct vnode *vp, vm_ooffset_t address,
73 int *run));
74static void vnode_pager_iodone __P((struct buf *bp));
75static int vnode_pager_input_smlfs __P((vm_object_t object, vm_page_t m));
76static int vnode_pager_input_old __P((vm_object_t object, vm_page_t m));
77static void vnode_pager_dealloc __P((vm_object_t));
78static int vnode_pager_getpages __P((vm_object_t, vm_page_t *, int, int));
79static void vnode_pager_putpages __P((vm_object_t, vm_page_t *, int, boolean_t, int *));
80static boolean_t vnode_pager_haspage __P((vm_object_t, vm_pindex_t, int *, int *));
81
82struct pagerops vnodepagerops = {
83 NULL,
84 vnode_pager_alloc,
85 vnode_pager_dealloc,
86 vnode_pager_getpages,
87 vnode_pager_putpages,
88 vnode_pager_haspage,
89 NULL
90};
91
92int vnode_pbuf_freecnt = -1; /* start out unlimited */
93
94/*
95 * Allocate (or lookup) pager for a vnode.
96 * Handle is a vnode pointer.
97 */
98vm_object_t
99vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
100 vm_ooffset_t offset)
101{
102 vm_object_t object;
103 struct vnode *vp;
104
105 /*
106 * Pageout to vnode, no can do yet.
107 */
108 if (handle == NULL)
109 return (NULL);
110
111 /*
112 * XXX hack - This initialization should be put somewhere else.
113 */
114 if (vnode_pbuf_freecnt < 0) {
115 vnode_pbuf_freecnt = nswbuf / 2 + 1;
116 }
117
118 vp = (struct vnode *) handle;
119
120 /*
121 * Prevent race condition when allocating the object. This
122 * can happen with NFS vnodes since the nfsnode isn't locked.
123 */
124 while (vp->v_flag & VOLOCK) {
125 vp->v_flag |= VOWANT;
126 tsleep(vp, PVM, "vnpobj", 0);
127 }
128 vp->v_flag |= VOLOCK;
129
130 /*
131 * If the object is being terminated, wait for it to
132 * go away.
133 */
134 while (((object = vp->v_object) != NULL) &&
135 (object->flags & OBJ_DEAD)) {
136 tsleep(object, PVM, "vadead", 0);
137 }
138
139 if (vp->v_usecount == 0)
140 panic("vnode_pager_alloc: no vnode reference");
141
142 if (object == NULL) {
143 /*
144 * And an object of the appropriate size
145 */
146 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
147 object->flags = 0;
148
149 object->un_pager.vnp.vnp_size = size;
150
151 object->handle = handle;
152 vp->v_object = object;
153 vp->v_usecount++;
154 } else {
155 object->ref_count++;
156 vp->v_usecount++;
157 }
158
159 vp->v_flag &= ~VOLOCK;
160 if (vp->v_flag & VOWANT) {
161 vp->v_flag &= ~VOWANT;
162 wakeup(vp);
163 }
164 return (object);
165}
166
167static void
168vnode_pager_dealloc(object)
169 vm_object_t object;
170{
171 register struct vnode *vp = object->handle;
172
173 if (vp == NULL)
174 panic("vnode_pager_dealloc: pager already dealloced");
175
176 vm_object_pip_wait(object, "vnpdea");
177
178 object->handle = NULL;
179 object->type = OBJT_DEAD;
180 vp->v_object = NULL;
181 vp->v_flag &= ~(VTEXT | VOBJBUF);
182}
183
184static boolean_t
185vnode_pager_haspage(object, pindex, before, after)
186 vm_object_t object;
187 vm_pindex_t pindex;
188 int *before;
189 int *after;
190{
191 struct vnode *vp = object->handle;
192 daddr_t bn;
193 int err;
194 daddr_t reqblock;
195 int poff;
196 int bsize;
197 int pagesperblock, blocksperpage;
198
199 /*
200 * If no vp or vp is doomed or marked transparent to VM, we do not
201 * have the page.
202 */
203 if ((vp == NULL) || (vp->v_flag & VDOOMED))
204 return FALSE;
205
206 /*
207 * If filesystem no longer mounted or offset beyond end of file we do
208 * not have the page.
209 */
210 if ((vp->v_mount == NULL) ||
211 (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size))
212 return FALSE;
213
214 bsize = vp->v_mount->mnt_stat.f_iosize;
215 pagesperblock = bsize / PAGE_SIZE;
216 blocksperpage = 0;
217 if (pagesperblock > 0) {
218 reqblock = pindex / pagesperblock;
219 } else {
220 blocksperpage = (PAGE_SIZE / bsize);
221 reqblock = pindex * blocksperpage;
222 }
223 err = VOP_BMAP(vp, reqblock, (struct vnode **) 0, &bn,
224 after, before);
225 if (err)
226 return TRUE;
227 if ( bn == -1)
228 return FALSE;
229 if (pagesperblock > 0) {
230 poff = pindex - (reqblock * pagesperblock);
231 if (before) {
232 *before *= pagesperblock;
233 *before += poff;
234 }
235 if (after) {
236 int numafter;
237 *after *= pagesperblock;
238 numafter = pagesperblock - (poff + 1);
239 if (IDX_TO_OFF(pindex + numafter) > object->un_pager.vnp.vnp_size) {
240 numafter = OFF_TO_IDX((object->un_pager.vnp.vnp_size - IDX_TO_OFF(pindex)));
241 }
242 *after += numafter;
243 }
244 } else {
245 if (before) {
246 *before /= blocksperpage;
247 }
248
249 if (after) {
250 *after /= blocksperpage;
251 }
252 }
253 return TRUE;
254}
255
256/*
257 * Lets the VM system know about a change in size for a file.
258 * We adjust our own internal size and flush any cached pages in
259 * the associated object that are affected by the size change.
260 *
261 * Note: this routine may be invoked as a result of a pager put
262 * operation (possibly at object termination time), so we must be careful.
263 */
264void
265vnode_pager_setsize(vp, nsize)
266 struct vnode *vp;
267 vm_ooffset_t nsize;
268{
269 vm_pindex_t nobjsize;
270 vm_object_t object = vp->v_object;
271
272 if (object == NULL)
273 return;
274
275 /*
276 * Hasn't changed size
277 */
278 if (nsize == object->un_pager.vnp.vnp_size)
279 return;
280
281 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
282
283 /*
284 * File has shrunk. Toss any cached pages beyond the new EOF.
285 */
286 if (nsize < object->un_pager.vnp.vnp_size) {
287 vm_freeze_copyopts(object, OFF_TO_IDX(nsize), object->size);
288 if (nobjsize < object->size) {
289 vm_object_page_remove(object, nobjsize, object->size,
290 FALSE);
291 }
292 /*
293 * this gets rid of garbage at the end of a page that is now
294 * only partially backed by the vnode.
295 *
296 * XXX for some reason (I don't know yet), if we take a
297 * completely invalid page and mark it partially valid
298 * it can screw up NFS reads, so we don't allow the case.
299 */
300 if (nsize & PAGE_MASK) {
301 vm_offset_t kva;
302 vm_page_t m;
303
304 m = vm_page_lookup(object, OFF_TO_IDX(nsize));
305 if (m && m->valid) {
306 int base = (int)nsize & PAGE_MASK;
307 int size = PAGE_SIZE - base;
308
309 /*
310 * Clear out partial-page garbage in case
311 * the page has been mapped.
312 */
313 kva = vm_pager_map_page(m);
314 bzero((caddr_t)kva + base, size);
315 vm_pager_unmap_page(kva);
316
317 /*
318 * XXX work around SMP data integrity race
319 * by unmapping the page from user processes.
320 * The garbage we just cleared may be mapped
321 * to a user process running on another cpu
322 * and this code is not running through normal
323 * I/O channels which handle SMP issues for
324 * us, so unmap page to synchronize all cpus.
325 *
326 * XXX should vm_pager_unmap_page() have
327 * dealt with this?
328 */
329 vm_page_protect(m, VM_PROT_NONE);
330
331 /*
332 * Clear out partial-page dirty bits. This
333 * has the side effect of setting the valid
334 * bits, but that is ok. There are a bunch
335 * of places in the VM system where we expected
336 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
337 * case is one of them. If the page is still
338 * partially dirty, make it fully dirty.
339 *
340 * note that we do not clear out the valid
341 * bits. This would prevent bogus_page
342 * replacement from working properly.
343 */
344 vm_page_set_validclean(m, base, size);
345 if (m->dirty != 0)
346 m->dirty = VM_PAGE_BITS_ALL;
347 }
348 }
349 }
350 object->un_pager.vnp.vnp_size = nsize;
351 object->size = nobjsize;
352}
353
354void
355vnode_pager_freepage(m)
356 vm_page_t m;
357{
358 vm_page_free(m);
359}
360
361/*
362 * calculate the linear (byte) disk address of specified virtual
363 * file address
364 */
365static vm_offset_t
366vnode_pager_addr(vp, address, run)
367 struct vnode *vp;
368 vm_ooffset_t address;
369 int *run;
370{
371 int rtaddress;
372 int bsize;
373 daddr_t block;
374 struct vnode *rtvp;
375 int err;
376 daddr_t vblock;
377 int voffset;
378
379 if ((int) address < 0)
380 return -1;
381
382 if (vp->v_mount == NULL)
383 return -1;
384
385 bsize = vp->v_mount->mnt_stat.f_iosize;
386 vblock = address / bsize;
387 voffset = address % bsize;
388
389 err = VOP_BMAP(vp, vblock, &rtvp, &block, run, NULL);
390
391 if (err || (block == -1))
392 rtaddress = -1;
393 else {
394 rtaddress = block + voffset / DEV_BSIZE;
395 if( run) {
396 *run += 1;
397 *run *= bsize/PAGE_SIZE;
398 *run -= voffset/PAGE_SIZE;
399 }
400 }
401
402 return rtaddress;
403}
404
405/*
406 * interrupt routine for I/O completion
407 */
408static void
409vnode_pager_iodone(bp)
410 struct buf *bp;
411{
412 bp->b_flags |= B_DONE;
413 wakeup(bp);
414}
415
416/*
417 * small block file system vnode pager input
418 */
419static int
420vnode_pager_input_smlfs(object, m)
421 vm_object_t object;
422 vm_page_t m;
423{
424 int i;
425 int s;
426 struct vnode *dp, *vp;
427 struct buf *bp;
428 vm_offset_t kva;
429 int fileaddr;
430 vm_offset_t bsize;
431 int error = 0;
432
433 vp = object->handle;
434 if (vp->v_mount == NULL)
435 return VM_PAGER_BAD;
436
437 bsize = vp->v_mount->mnt_stat.f_iosize;
438
439
440 VOP_BMAP(vp, 0, &dp, 0, NULL, NULL);
441
442 kva = vm_pager_map_page(m);
443
444 for (i = 0; i < PAGE_SIZE / bsize; i++) {
445 vm_ooffset_t address;
446
447 if (vm_page_bits(i * bsize, bsize) & m->valid)
448 continue;
449
450 address = IDX_TO_OFF(m->pindex) + i * bsize;
451 if (address >= object->un_pager.vnp.vnp_size) {
452 fileaddr = -1;
453 } else {
454 fileaddr = vnode_pager_addr(vp, address, NULL);
455 }
456 if (fileaddr != -1) {
457 bp = getpbuf(&vnode_pbuf_freecnt);
458
459 /* build a minimal buffer header */
460 bp->b_flags = B_READ | B_CALL;
461 bp->b_iodone = vnode_pager_iodone;
462 bp->b_rcred = bp->b_wcred = curproc->p_ucred;
463 if (bp->b_rcred != NOCRED)
464 crhold(bp->b_rcred);
465 if (bp->b_wcred != NOCRED)
466 crhold(bp->b_wcred);
467 bp->b_data = (caddr_t) kva + i * bsize;
468 bp->b_blkno = fileaddr;
469 pbgetvp(dp, bp);
470 bp->b_bcount = bsize;
471 bp->b_bufsize = bsize;
472 bp->b_runningbufspace = bp->b_bufsize;
473 runningbufspace += bp->b_runningbufspace;
474
475 /* do the input */
476 VOP_STRATEGY(bp->b_vp, bp);
477
478 /* we definitely need to be at splvm here */
479
480 s = splvm();
481 while ((bp->b_flags & B_DONE) == 0) {
482 tsleep(bp, PVM, "vnsrd", 0);
483 }
484 splx(s);
485 if ((bp->b_flags & B_ERROR) != 0)
486 error = EIO;
487
488 /*
489 * free the buffer header back to the swap buffer pool
490 */
491 relpbuf(bp, &vnode_pbuf_freecnt);
492 if (error)
493 break;
494
495 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
496 } else {
497 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
498 bzero((caddr_t) kva + i * bsize, bsize);
499 }
500 }
501 vm_pager_unmap_page(kva);
502 pmap_clear_modify(m);
503 vm_page_flag_clear(m, PG_ZERO);
504 if (error) {
505 return VM_PAGER_ERROR;
506 }
507 return VM_PAGER_OK;
508
509}
510
511
512/*
513 * old style vnode pager output routine
514 */
515static int
516vnode_pager_input_old(object, m)
517 vm_object_t object;
518 vm_page_t m;
519{
520 struct uio auio;
521 struct iovec aiov;
522 int error;
523 int size;
524 vm_offset_t kva;
525
526 error = 0;
527
528 /*
529 * Return failure if beyond current EOF
530 */
531 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
532 return VM_PAGER_BAD;
533 } else {
534 size = PAGE_SIZE;
535 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
536 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
537
538 /*
539 * Allocate a kernel virtual address and initialize so that
540 * we can use VOP_READ/WRITE routines.
541 */
542 kva = vm_pager_map_page(m);
543
544 aiov.iov_base = (caddr_t) kva;
545 aiov.iov_len = size;
546 auio.uio_iov = &aiov;
547 auio.uio_iovcnt = 1;
548 auio.uio_offset = IDX_TO_OFF(m->pindex);
549 auio.uio_segflg = UIO_SYSSPACE;
550 auio.uio_rw = UIO_READ;
551 auio.uio_resid = size;
552 auio.uio_procp = curproc;
553
554 error = VOP_READ(object->handle, &auio, 0, curproc->p_ucred);
555 if (!error) {
556 register int count = size - auio.uio_resid;
557
558 if (count == 0)
559 error = EINVAL;
560 else if (count != PAGE_SIZE)
561 bzero((caddr_t) kva + count, PAGE_SIZE - count);
562 }
563 vm_pager_unmap_page(kva);
564 }
565 pmap_clear_modify(m);
566 vm_page_undirty(m);
567 vm_page_flag_clear(m, PG_ZERO);
568 if (!error)
569 m->valid = VM_PAGE_BITS_ALL;
570 return error ? VM_PAGER_ERROR : VM_PAGER_OK;
571}
572
573/*
574 * generic vnode pager input routine
575 */
576
577/*
578 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
579 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
580 * vnode_pager_generic_getpages() to implement the previous behaviour.
581 *
582 * All other FS's should use the bypass to get to the local media
583 * backing vp's VOP_GETPAGES.
584 */
585static int
586vnode_pager_getpages(object, m, count, reqpage)
587 vm_object_t object;
588 vm_page_t *m;
589 int count;
590 int reqpage;
591{
592 int rtval;
593 struct vnode *vp;
594 int bytes = count * PAGE_SIZE;
595
596 vp = object->handle;
597 /*
598 * XXX temporary diagnostic message to help track stale FS code,
599 * Returning EOPNOTSUPP from here may make things unhappy.
600 */
601 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
602 if (rtval == EOPNOTSUPP) {
603 printf("vnode_pager: *** WARNING *** stale FS getpages\n");
604 rtval = vnode_pager_generic_getpages( vp, m, bytes, reqpage);
605 }
606 return rtval;
607}
608
609
610/*
611 * This is now called from local media FS's to operate against their
612 * own vnodes if they fail to implement VOP_GETPAGES.
613 */
614int
615vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
616 struct vnode *vp;
617 vm_page_t *m;
618 int bytecount;
619 int reqpage;
620{
621 vm_object_t object;
622 vm_offset_t kva;
623 off_t foff, tfoff, nextoff;
624 int i, size, bsize, first, firstaddr;
625 struct vnode *dp;
626 int runpg;
627 int runend;
628 struct buf *bp;
629 int s;
630 int count;
631 int error = 0;
632
633 object = vp->v_object;
634 count = bytecount / PAGE_SIZE;
635
636 if (vp->v_mount == NULL)
637 return VM_PAGER_BAD;
638
639 bsize = vp->v_mount->mnt_stat.f_iosize;
640
641 /* get the UNDERLYING device for the file with VOP_BMAP() */
642
643 /*
644 * originally, we did not check for an error return value -- assuming
645 * an fs always has a bmap entry point -- that assumption is wrong!!!
646 */
647 foff = IDX_TO_OFF(m[reqpage]->pindex);
648
649 /*
650 * if we can't bmap, use old VOP code
651 */
652 if (VOP_BMAP(vp, 0, &dp, 0, NULL, NULL)) {
653 for (i = 0; i < count; i++) {
654 if (i != reqpage) {
655 vnode_pager_freepage(m[i]);
656 }
657 }
658 cnt.v_vnodein++;
659 cnt.v_vnodepgsin++;
660 return vnode_pager_input_old(object, m[reqpage]);
661
662 /*
663 * if the blocksize is smaller than a page size, then use
664 * special small filesystem code. NFS sometimes has a small
665 * blocksize, but it can handle large reads itself.
666 */
667 } else if ((PAGE_SIZE / bsize) > 1 &&
668 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
669 for (i = 0; i < count; i++) {
670 if (i != reqpage) {
671 vnode_pager_freepage(m[i]);
672 }
673 }
674 cnt.v_vnodein++;
675 cnt.v_vnodepgsin++;
676 return vnode_pager_input_smlfs(object, m[reqpage]);
677 }
678
679 /*
680 * If we have a completely valid page available to us, we can
681 * clean up and return. Otherwise we have to re-read the
682 * media.
683 */
684
685 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
686 for (i = 0; i < count; i++) {
687 if (i != reqpage)
688 vnode_pager_freepage(m[i]);
689 }
690 return VM_PAGER_OK;
691 }
692 m[reqpage]->valid = 0;
693
694 /*
695 * here on direct device I/O
696 */
697
698 firstaddr = -1;
699 /*
700 * calculate the run that includes the required page
701 */
702 for(first = 0, i = 0; i < count; i = runend) {
703 firstaddr = vnode_pager_addr(vp,
704 IDX_TO_OFF(m[i]->pindex), &runpg);
705 if (firstaddr == -1) {
706 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
707 /* XXX no %qd in kernel. */
708 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %d, foff: 0x%lx%08lx, vnp_size: 0x%lx%08lx",
709 firstaddr, (u_long)(foff >> 32),
710 (u_long)(u_int32_t)foff,
711 (u_long)(u_int32_t)
712 (object->un_pager.vnp.vnp_size >> 32),
713 (u_long)(u_int32_t)
714 object->un_pager.vnp.vnp_size);
715 }
716 vnode_pager_freepage(m[i]);
717 runend = i + 1;
718 first = runend;
719 continue;
720 }
721 runend = i + runpg;
722 if (runend <= reqpage) {
723 int j;
724 for (j = i; j < runend; j++) {
725 vnode_pager_freepage(m[j]);
726 }
727 } else {
728 if (runpg < (count - first)) {
729 for (i = first + runpg; i < count; i++)
730 vnode_pager_freepage(m[i]);
731 count = first + runpg;
732 }
733 break;
734 }
735 first = runend;
736 }
737
738 /*
739 * the first and last page have been calculated now, move input pages
740 * to be zero based...
741 */
742 if (first != 0) {
743 for (i = first; i < count; i++) {
744 m[i - first] = m[i];
745 }
746 count -= first;
747 reqpage -= first;
748 }
749
750 /*
751 * calculate the file virtual address for the transfer
752 */
753 foff = IDX_TO_OFF(m[0]->pindex);
754
755 /*
756 * calculate the size of the transfer
757 */
758 size = count * PAGE_SIZE;
759 if ((foff + size) > object->un_pager.vnp.vnp_size)
760 size = object->un_pager.vnp.vnp_size - foff;
761
762 /*
763 * round up physical size for real devices.
764 */
765 if (dp->v_type == VBLK || dp->v_type == VCHR) {
766 int secmask = dp->v_rdev->si_bsize_phys - 1;
767 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
768 size = (size + secmask) & ~secmask;
769 }
770
771 bp = getpbuf(&vnode_pbuf_freecnt);
772 kva = (vm_offset_t) bp->b_data;
773
774 /*
775 * and map the pages to be read into the kva
776 */
777 pmap_qenter(kva, m, count);
778
779 /* build a minimal buffer header */
780 bp->b_flags = B_READ | B_CALL;
781 bp->b_iodone = vnode_pager_iodone;
782 /* B_PHYS is not set, but it is nice to fill this in */
783 bp->b_rcred = bp->b_wcred = curproc->p_ucred;
784 if (bp->b_rcred != NOCRED)
785 crhold(bp->b_rcred);
786 if (bp->b_wcred != NOCRED)
787 crhold(bp->b_wcred);
788 bp->b_blkno = firstaddr;
789 pbgetvp(dp, bp);
790 bp->b_bcount = size;
791 bp->b_bufsize = size;
792 bp->b_runningbufspace = bp->b_bufsize;
793 runningbufspace += bp->b_runningbufspace;
794
795 cnt.v_vnodein++;
796 cnt.v_vnodepgsin += count;
797
798 /* do the input */
799 VOP_STRATEGY(bp->b_vp, bp);
800
801 s = splvm();
802 /* we definitely need to be at splvm here */
803
804 while ((bp->b_flags & B_DONE) == 0) {
805 tsleep(bp, PVM, "vnread", 0);
806 }
807 splx(s);
808 if ((bp->b_flags & B_ERROR) != 0)
809 error = EIO;
810
811 if (!error) {
812 if (size != count * PAGE_SIZE)
813 bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
814 }
815 pmap_qremove(kva, count);
816
817 /*
818 * free the buffer header back to the swap buffer pool
819 */
820 relpbuf(bp, &vnode_pbuf_freecnt);
821
822 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
823 vm_page_t mt;
824
825 nextoff = tfoff + PAGE_SIZE;
826 mt = m[i];
827
828 if (nextoff <= object->un_pager.vnp.vnp_size) {
829 /*
830 * Read filled up entire page.
831 */
832 mt->valid = VM_PAGE_BITS_ALL;
833 vm_page_undirty(mt); /* should be an assert? XXX */
834 pmap_clear_modify(mt);
835 } else {
836 /*
837 * Read did not fill up entire page. Since this
838 * is getpages, the page may be mapped, so we have
839 * to zero the invalid portions of the page even
840 * though we aren't setting them valid.
841 *
842 * Currently we do not set the entire page valid,
843 * we just try to clear the piece that we couldn't
844 * read.
845 */
846 vm_page_set_validclean(mt, 0,
847 object->un_pager.vnp.vnp_size - tfoff);
848 /* handled by vm_fault now */
849 /* vm_page_zero_invalid(mt, FALSE); */
850 }
851
852 vm_page_flag_clear(mt, PG_ZERO);
853 if (i != reqpage) {
854
855 /*
856 * whether or not to leave the page activated is up in
857 * the air, but we should put the page on a page queue
858 * somewhere. (it already is in the object). Result:
859 * It appears that empirical results show that
860 * deactivating pages is best.
861 */
862
863 /*
864 * just in case someone was asking for this page we
865 * now tell them that it is ok to use
866 */
867 if (!error) {
868 if (mt->flags & PG_WANTED)
869 vm_page_activate(mt);
870 else
871 vm_page_deactivate(mt);
872 vm_page_wakeup(mt);
873 } else {
874 vnode_pager_freepage(mt);
875 }
876 }
877 }
878 if (error) {
879 printf("vnode_pager_getpages: I/O read error\n");
880 }
881 return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
882}
883
884/*
885 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
886 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
887 * vnode_pager_generic_putpages() to implement the previous behaviour.
888 *
889 * All other FS's should use the bypass to get to the local media
890 * backing vp's VOP_PUTPAGES.
891 */
892static void
893vnode_pager_putpages(object, m, count, sync, rtvals)
894 vm_object_t object;
895 vm_page_t *m;
896 int count;
897 boolean_t sync;
898 int *rtvals;
899{
900 int rtval;
901 struct vnode *vp;
902 int bytes = count * PAGE_SIZE;
903
904 /*
905 * Force synchronous operation if we are extremely low on memory
906 * to prevent a low-memory deadlock. VOP operations often need to
907 * allocate more memory to initiate the I/O ( i.e. do a BMAP
908 * operation ). The swapper handles the case by limiting the amount
909 * of asynchronous I/O, but that sort of solution doesn't scale well
910 * for the vnode pager without a lot of work.
911 *
912 * Also, the backing vnode's iodone routine may not wake the pageout
913 * daemon up. This should be probably be addressed XXX.
914 */
915
916 if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
917 sync |= OBJPC_SYNC;
918
919 /*
920 * Call device-specific putpages function
921 */
922
923 vp = object->handle;
924 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
925 if (rtval == EOPNOTSUPP) {
926 printf("vnode_pager: *** WARNING *** stale FS putpages\n");
927 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
928 }
929}
930
931
932/*
933 * This is now called from local media FS's to operate against their
934 * own vnodes if they fail to implement VOP_PUTPAGES.
935 *
936 * This is typically called indirectly via the pageout daemon and
937 * clustering has already typically occured, so in general we ask the
938 * underlying filesystem to write the data out asynchronously rather
939 * then delayed.
940 */
941int
942vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals)
943 struct vnode *vp;
944 vm_page_t *m;
945 int bytecount;
946 int flags;
947 int *rtvals;
948{
949 int i;
950 vm_object_t object;
951 int count;
952
953 int maxsize, ncount;
954 vm_ooffset_t poffset;
955 struct uio auio;
956 struct iovec aiov;
957 int error;
958 int ioflags;
959
960 object = vp->v_object;
961 count = bytecount / PAGE_SIZE;
962
963 for (i = 0; i < count; i++)
964 rtvals[i] = VM_PAGER_AGAIN;
965
966 if ((int) m[0]->pindex < 0) {
967 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
968 (long)m[0]->pindex, m[0]->dirty);
969 rtvals[0] = VM_PAGER_BAD;
970 return VM_PAGER_BAD;
971 }
972
973 maxsize = count * PAGE_SIZE;
974 ncount = count;
975
976 poffset = IDX_TO_OFF(m[0]->pindex);
977
978 /*
979 * If the page-aligned write is larger then the actual file we
980 * have to invalidate pages occuring beyond the file EOF. However,
981 * there is an edge case where a file may not be page-aligned where
982 * the last page is partially invalid. In this case the filesystem
983 * may not properly clear the dirty bits for the entire page (which
984 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
985 * With the page locked we are free to fix-up the dirty bits here.
986 *
987 * We do not under any circumstances truncate the valid bits, as
988 * this will screw up bogus page replacement.
989 */
990 if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
991 if (object->un_pager.vnp.vnp_size > poffset) {
992 int pgoff;
993
994 maxsize = object->un_pager.vnp.vnp_size - poffset;
995 ncount = btoc(maxsize);
996 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
997 vm_page_clear_dirty(m[ncount - 1], pgoff,
998 PAGE_SIZE - pgoff);
999 }
1000 } else {
1001 maxsize = 0;
1002 ncount = 0;
1003 }
1004 if (ncount < count) {
1005 for (i = ncount; i < count; i++) {
1006 rtvals[i] = VM_PAGER_BAD;
1007 }
1008 }
1009 }
1010
1011 /*
1012 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
1013 * rather then a bdwrite() to prevent paging I/O from saturating
1014 * the buffer cache. Dummy-up the sequential heuristic to cause
1015 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
1016 * the system decides how to cluster.
1017 */
1018 ioflags = IO_VMIO;
1019 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1020 ioflags |= IO_SYNC;
1021 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1022 ioflags |= IO_ASYNC;
1023 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1024 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1025
1026 aiov.iov_base = (caddr_t) 0;
1027 aiov.iov_len = maxsize;
1028 auio.uio_iov = &aiov;
1029 auio.uio_iovcnt = 1;
1030 auio.uio_offset = poffset;
1031 auio.uio_segflg = UIO_NOCOPY;
1032 auio.uio_rw = UIO_WRITE;
1033 auio.uio_resid = maxsize;
1034 auio.uio_procp = (struct proc *) 0;
1035 error = VOP_WRITE(vp, &auio, ioflags, curproc->p_ucred);
1036 cnt.v_vnodeout++;
1037 cnt.v_vnodepgsout += ncount;
1038
1039 if (error) {
1040 printf("vnode_pager_putpages: I/O error %d\n", error);
1041 }
1042 if (auio.uio_resid) {
1043 printf("vnode_pager_putpages: residual I/O %d at %lu\n",
1044 auio.uio_resid, (u_long)m[0]->pindex);
1045 }
1046 for (i = 0; i < ncount; i++) {
1047 rtvals[i] = VM_PAGER_OK;
1048 }
1049 return rtvals[0];
1050}
1051
1052struct vnode *
1053vnode_pager_lock(object)
1054 vm_object_t object;
1055{
1056 struct proc *p = curproc; /* XXX */
1057
1058 for (; object != NULL; object = object->backing_object) {
1059 if (object->type != OBJT_VNODE)
1060 continue;
1061 if (object->flags & OBJ_DEAD)
1062 return NULL;
1063
1064 while (vget(object->handle,
1065 LK_NOPAUSE | LK_SHARED | LK_RETRY | LK_CANRECURSE, p)) {
1066 if ((object->flags & OBJ_DEAD) || (object->type != OBJT_VNODE))
1067 return NULL;
1068 printf("vnode_pager_lock: retrying\n");
1069 }
1070 return object->handle;
1071 }
1072 return NULL;
1073}