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: src/sys/vm/vnode_pager.c,v 1.196.2.8 2005/08/16 15:09:44 sobomax Exp $");
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 <machine/atomic.h>
68
69 #include <vm/vm.h>
70 #include <vm/vm_object.h>
71 #include <vm/vm_page.h>
72 #include <vm/vm_pager.h>
73 #include <vm/vm_map.h>
74 #include <vm/vnode_pager.h>
75 #include <vm/vm_extern.h>
76
77 static daddr_t vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
78 int *run);
79 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
80 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
81 static void vnode_pager_dealloc(vm_object_t);
82 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int);
83 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
84 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
85
86 struct pagerops vnodepagerops = {
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 /*
97 * Allocate (or lookup) pager for a vnode.
98 * Handle is a vnode pointer.
99 *
100 * MPSAFE
101 */
102 vm_object_t
103 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
104 vm_ooffset_t offset)
105 {
106 vm_object_t object;
107 struct vnode *vp;
108
109 /*
110 * Pageout to vnode, no can do yet.
111 */
112 if (handle == NULL)
113 return (NULL);
114
115 vp = (struct vnode *) handle;
116
117 ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc");
118
119 /*
120 * Prevent race condition when allocating the object. This
121 * can happen with NFS vnodes since the nfsnode isn't locked.
122 */
123 VI_LOCK(vp);
124 while (vp->v_iflag & VI_OLOCK) {
125 vp->v_iflag |= VI_OWANT;
126 msleep(vp, VI_MTX(vp), PVM, "vnpobj", 0);
127 }
128 vp->v_iflag |= VI_OLOCK;
129 VI_UNLOCK(vp);
130
131 /*
132 * If the object is being terminated, wait for it to
133 * go away.
134 */
135 while ((object = vp->v_object) != NULL) {
136 VM_OBJECT_LOCK(object);
137 if ((object->flags & OBJ_DEAD) == 0)
138 break;
139 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
140 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vadead", 0);
141 }
142
143 if (vp->v_usecount == 0)
144 panic("vnode_pager_alloc: no vnode reference");
145
146 if (object == NULL) {
147 /*
148 * And an object of the appropriate size
149 */
150 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
151
152 object->un_pager.vnp.vnp_size = size;
153
154 object->handle = handle;
155 vp->v_object = object;
156 } else {
157 object->ref_count++;
158 VM_OBJECT_UNLOCK(object);
159 }
160 VI_LOCK(vp);
161 vp->v_usecount++;
162 vp->v_iflag &= ~VI_OLOCK;
163 if (vp->v_iflag & VI_OWANT) {
164 vp->v_iflag &= ~VI_OWANT;
165 wakeup(vp);
166 }
167 VI_UNLOCK(vp);
168 return (object);
169 }
170
171 /*
172 * The object must be locked.
173 */
174 static void
175 vnode_pager_dealloc(object)
176 vm_object_t object;
177 {
178 struct vnode *vp = object->handle;
179
180 if (vp == NULL)
181 panic("vnode_pager_dealloc: pager already dealloced");
182
183 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
184 vm_object_pip_wait(object, "vnpdea");
185
186 object->handle = NULL;
187 object->type = OBJT_DEAD;
188 if (object->flags & OBJ_DISCONNECTWNT) {
189 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
190 wakeup(object);
191 }
192 ASSERT_VOP_LOCKED(vp, "vnode_pager_dealloc");
193 vp->v_object = NULL;
194 vp->v_vflag &= ~(VV_TEXT | VV_OBJBUF);
195 }
196
197 static boolean_t
198 vnode_pager_haspage(object, pindex, before, after)
199 vm_object_t object;
200 vm_pindex_t pindex;
201 int *before;
202 int *after;
203 {
204 struct vnode *vp = object->handle;
205 daddr_t bn;
206 int err;
207 daddr_t reqblock;
208 int poff;
209 int bsize;
210 int pagesperblock, blocksperpage;
211
212 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
213 /*
214 * If no vp or vp is doomed or marked transparent to VM, we do not
215 * have the page.
216 */
217 if (vp == NULL)
218 return FALSE;
219
220 VI_LOCK(vp);
221 if (vp->v_iflag & VI_DOOMED) {
222 VI_UNLOCK(vp);
223 return FALSE;
224 }
225 VI_UNLOCK(vp);
226 /*
227 * If filesystem no longer mounted or offset beyond end of file we do
228 * not have the page.
229 */
230 if ((vp->v_mount == NULL) ||
231 (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size))
232 return FALSE;
233
234 bsize = vp->v_mount->mnt_stat.f_iosize;
235 pagesperblock = bsize / PAGE_SIZE;
236 blocksperpage = 0;
237 if (pagesperblock > 0) {
238 reqblock = pindex / pagesperblock;
239 } else {
240 blocksperpage = (PAGE_SIZE / bsize);
241 reqblock = pindex * blocksperpage;
242 }
243 VM_OBJECT_UNLOCK(object);
244 mtx_lock(&Giant);
245 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
246 mtx_unlock(&Giant);
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 daddr_t
381 vnode_pager_addr(vp, address, run)
382 struct vnode *vp;
383 vm_ooffset_t address;
384 int *run;
385 {
386 daddr_t rtaddress;
387 int bsize;
388 daddr_t block;
389 int err;
390 daddr_t vblock;
391 daddr_t voffset;
392
393 GIANT_REQUIRED;
394 if (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 daddr_t 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 atomic_add_int(&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 input 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;
643 daddr_t firstaddr;
644 struct vnode *dp;
645 int runpg;
646 int runend;
647 struct buf *bp;
648 int count;
649 int error = 0;
650
651 GIANT_REQUIRED;
652 object = vp->v_object;
653 count = bytecount / PAGE_SIZE;
654
655 if (vp->v_mount == NULL)
656 return VM_PAGER_BAD;
657
658 bsize = vp->v_mount->mnt_stat.f_iosize;
659
660 /* get the UNDERLYING device for the file with VOP_BMAP() */
661
662 /*
663 * originally, we did not check for an error return value -- assuming
664 * an fs always has a bmap entry point -- that assumption is wrong!!!
665 */
666 foff = IDX_TO_OFF(m[reqpage]->pindex);
667
668 /*
669 * if we can't bmap, use old VOP code
670 */
671 if (VOP_BMAP(vp, 0, &dp, 0, NULL, NULL)) {
672 VM_OBJECT_LOCK(object);
673 vm_page_lock_queues();
674 for (i = 0; i < count; i++)
675 if (i != reqpage)
676 vm_page_free(m[i]);
677 vm_page_unlock_queues();
678 cnt.v_vnodein++;
679 cnt.v_vnodepgsin++;
680 error = vnode_pager_input_old(object, m[reqpage]);
681 VM_OBJECT_UNLOCK(object);
682 return (error);
683
684 /*
685 * if the blocksize is smaller than a page size, then use
686 * special small filesystem code. NFS sometimes has a small
687 * blocksize, but it can handle large reads itself.
688 */
689 } else if ((PAGE_SIZE / bsize) > 1 &&
690 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
691 VM_OBJECT_LOCK(object);
692 vm_page_lock_queues();
693 for (i = 0; i < count; i++)
694 if (i != reqpage)
695 vm_page_free(m[i]);
696 vm_page_unlock_queues();
697 VM_OBJECT_UNLOCK(object);
698 cnt.v_vnodein++;
699 cnt.v_vnodepgsin++;
700 return vnode_pager_input_smlfs(object, m[reqpage]);
701 }
702
703 /*
704 * If we have a completely valid page available to us, we can
705 * clean up and return. Otherwise we have to re-read the
706 * media.
707 */
708 VM_OBJECT_LOCK(object);
709 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
710 vm_page_lock_queues();
711 for (i = 0; i < count; i++)
712 if (i != reqpage)
713 vm_page_free(m[i]);
714 vm_page_unlock_queues();
715 VM_OBJECT_UNLOCK(object);
716 return VM_PAGER_OK;
717 }
718 m[reqpage]->valid = 0;
719 VM_OBJECT_UNLOCK(object);
720
721 /*
722 * here on direct device I/O
723 */
724 firstaddr = -1;
725
726 /*
727 * calculate the run that includes the required page
728 */
729 for (first = 0, i = 0; i < count; i = runend) {
730 firstaddr = vnode_pager_addr(vp,
731 IDX_TO_OFF(m[i]->pindex), &runpg);
732 if (firstaddr == -1) {
733 VM_OBJECT_LOCK(object);
734 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
735 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
736 (intmax_t)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 VM_OBJECT_UNLOCK(object);
746 runend = i + 1;
747 first = runend;
748 continue;
749 }
750 runend = i + runpg;
751 if (runend <= reqpage) {
752 VM_OBJECT_LOCK(object);
753 vm_page_lock_queues();
754 for (j = i; j < runend; j++)
755 vm_page_free(m[j]);
756 vm_page_unlock_queues();
757 VM_OBJECT_UNLOCK(object);
758 } else {
759 if (runpg < (count - first)) {
760 VM_OBJECT_LOCK(object);
761 vm_page_lock_queues();
762 for (i = first + runpg; i < count; i++)
763 vm_page_free(m[i]);
764 vm_page_unlock_queues();
765 VM_OBJECT_UNLOCK(object);
766 count = first + runpg;
767 }
768 break;
769 }
770 first = runend;
771 }
772
773 /*
774 * the first and last page have been calculated now, move input pages
775 * to be zero based...
776 */
777 if (first != 0) {
778 for (i = first; i < count; i++) {
779 m[i - first] = m[i];
780 }
781 count -= first;
782 reqpage -= first;
783 }
784
785 /*
786 * calculate the file virtual address for the transfer
787 */
788 foff = IDX_TO_OFF(m[0]->pindex);
789
790 /*
791 * calculate the size of the transfer
792 */
793 size = count * PAGE_SIZE;
794 if ((foff + size) > object->un_pager.vnp.vnp_size)
795 size = object->un_pager.vnp.vnp_size - foff;
796
797 /*
798 * round up physical size for real devices.
799 */
800 if (dp->v_type == VBLK || dp->v_type == VCHR) {
801 int secmask = dp->v_rdev->si_bsize_phys - 1;
802 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
803 size = (size + secmask) & ~secmask;
804 }
805
806 bp = getpbuf(&vnode_pbuf_freecnt);
807 kva = (vm_offset_t) bp->b_data;
808
809 /*
810 * and map the pages to be read into the kva
811 */
812 pmap_qenter(kva, m, count);
813
814 /* build a minimal buffer header */
815 bp->b_iocmd = BIO_READ;
816 bp->b_iodone = bdone;
817 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
818 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
819 bp->b_rcred = crhold(curthread->td_ucred);
820 bp->b_wcred = crhold(curthread->td_ucred);
821 bp->b_blkno = firstaddr;
822 pbgetvp(dp, bp);
823 bp->b_bcount = size;
824 bp->b_bufsize = size;
825 bp->b_runningbufspace = bp->b_bufsize;
826 atomic_add_int(&runningbufspace, bp->b_runningbufspace);
827
828 cnt.v_vnodein++;
829 cnt.v_vnodepgsin += count;
830
831 /* do the input */
832 bp->b_iooffset = dbtob(bp->b_blkno);
833 if (dp->v_type == VCHR)
834 VOP_SPECSTRATEGY(bp->b_vp, bp);
835 else
836 VOP_STRATEGY(bp->b_vp, bp);
837
838 bwait(bp, PVM, "vnread");
839
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_OBJECT_LOCK(object);
855 vm_page_lock_queues();
856 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
857 vm_page_t mt;
858
859 nextoff = tfoff + PAGE_SIZE;
860 mt = m[i];
861
862 if (nextoff <= object->un_pager.vnp.vnp_size) {
863 /*
864 * Read filled up entire page.
865 */
866 mt->valid = VM_PAGE_BITS_ALL;
867 vm_page_undirty(mt); /* should be an assert? XXX */
868 pmap_clear_modify(mt);
869 } else {
870 /*
871 * Read did not fill up entire page. Since this
872 * is getpages, the page may be mapped, so we have
873 * to zero the invalid portions of the page even
874 * though we aren't setting them valid.
875 *
876 * Currently we do not set the entire page valid,
877 * we just try to clear the piece that we couldn't
878 * read.
879 */
880 vm_page_set_validclean(mt, 0,
881 object->un_pager.vnp.vnp_size - tfoff);
882 /* handled by vm_fault now */
883 /* vm_page_zero_invalid(mt, FALSE); */
884 }
885
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 ((int64_t)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: 693d3b33e74c6fab997949bc4675c12f
|