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/6.2/sys/vm/vnode_pager.c 163340 2006-10-14 06:04:32Z 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/limits.h>
65 #include <sys/conf.h>
66 #include <sys/sf_buf.h>
67
68 #include <machine/atomic.h>
69
70 #include <vm/vm.h>
71 #include <vm/vm_object.h>
72 #include <vm/vm_page.h>
73 #include <vm/vm_pager.h>
74 #include <vm/vm_map.h>
75 #include <vm/vnode_pager.h>
76 #include <vm/vm_extern.h>
77
78 static daddr_t vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
79 int *run);
80 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
81 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
82 static void vnode_pager_dealloc(vm_object_t);
83 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int);
84 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
85 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
86 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t, vm_ooffset_t);
87
88 struct pagerops vnodepagerops = {
89 .pgo_alloc = vnode_pager_alloc,
90 .pgo_dealloc = vnode_pager_dealloc,
91 .pgo_getpages = vnode_pager_getpages,
92 .pgo_putpages = vnode_pager_putpages,
93 .pgo_haspage = vnode_pager_haspage,
94 };
95
96 int vnode_pbuf_freecnt;
97
98 /*
99 * Compatibility function for RELENG_6, in which vnode_create_vobject()
100 * takes file size as size_t due to an oversight. The type may not just
101 * change to off_t because the ABI to 3rd party modules must be preserved
102 * for RELENG_6 lifetime.
103 */
104 int
105 vnode_create_vobject(struct vnode *vp, size_t isize __unused, struct thread *td)
106 {
107
108 /*
109 * Size of 0 will indicate to vnode_create_vobject_off()
110 * VOP_GETATTR() is to be called to get the actual size.
111 */
112 return (vnode_create_vobject_off(vp, 0, td));
113 }
114
115 /*
116 * Create the VM system backing object for this vnode -- for RELENG_6 only.
117 * In HEAD, vnode_create_vobject() has been fixed to take file size as off_t
118 * and so it can be used as is.
119 */
120 int
121 vnode_create_vobject_off(struct vnode *vp, off_t isize, struct thread *td)
122 {
123 vm_object_t object;
124 vm_ooffset_t size = isize;
125 struct vattr va;
126
127 if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE)
128 return (0);
129
130 while ((object = vp->v_object) != NULL) {
131 VM_OBJECT_LOCK(object);
132 if (!(object->flags & OBJ_DEAD)) {
133 VM_OBJECT_UNLOCK(object);
134 return (0);
135 }
136 VOP_UNLOCK(vp, 0, td);
137 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
138 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vodead", 0);
139 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
140 }
141
142 if (size == 0) {
143 if (vn_isdisk(vp, NULL)) {
144 size = IDX_TO_OFF(INT_MAX);
145 } else {
146 if (VOP_GETATTR(vp, &va, td->td_ucred, td) != 0)
147 return (0);
148 size = va.va_size;
149 }
150 }
151
152 object = vnode_pager_alloc(vp, size, 0, 0);
153 /*
154 * Dereference the reference we just created. This assumes
155 * that the object is associated with the vp.
156 */
157 VM_OBJECT_LOCK(object);
158 object->ref_count--;
159 VM_OBJECT_UNLOCK(object);
160 vrele(vp);
161
162 KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
163
164 return (0);
165 }
166
167 void
168 vnode_destroy_vobject(struct vnode *vp)
169 {
170 struct vm_object *obj;
171
172 obj = vp->v_object;
173 if (obj == NULL)
174 return;
175 ASSERT_VOP_LOCKED(vp, "vnode_destroy_vobject");
176 VM_OBJECT_LOCK(obj);
177 if (obj->ref_count == 0) {
178 /*
179 * vclean() may be called twice. The first time
180 * removes the primary reference to the object,
181 * the second time goes one further and is a
182 * special-case to terminate the object.
183 *
184 * don't double-terminate the object
185 */
186 if ((obj->flags & OBJ_DEAD) == 0)
187 vm_object_terminate(obj);
188 else
189 VM_OBJECT_UNLOCK(obj);
190 } else {
191 /*
192 * Woe to the process that tries to page now :-).
193 */
194 vm_pager_deallocate(obj);
195 VM_OBJECT_UNLOCK(obj);
196 }
197 vp->v_object = NULL;
198 }
199
200
201 /*
202 * Allocate (or lookup) pager for a vnode.
203 * Handle is a vnode pointer.
204 *
205 * MPSAFE
206 */
207 vm_object_t
208 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
209 vm_ooffset_t offset)
210 {
211 vm_object_t object;
212 struct vnode *vp;
213
214 /*
215 * Pageout to vnode, no can do yet.
216 */
217 if (handle == NULL)
218 return (NULL);
219
220 vp = (struct vnode *) handle;
221
222 ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc");
223
224 /*
225 * If the object is being terminated, wait for it to
226 * go away.
227 */
228 while ((object = vp->v_object) != NULL) {
229 VM_OBJECT_LOCK(object);
230 if ((object->flags & OBJ_DEAD) == 0)
231 break;
232 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
233 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vadead", 0);
234 }
235
236 if (vp->v_usecount == 0)
237 panic("vnode_pager_alloc: no vnode reference");
238
239 if (object == NULL) {
240 /*
241 * And an object of the appropriate size
242 */
243 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
244
245 object->un_pager.vnp.vnp_size = size;
246
247 object->handle = handle;
248 if (VFS_NEEDSGIANT(vp->v_mount))
249 vm_object_set_flag(object, OBJ_NEEDGIANT);
250 vp->v_object = object;
251 } else {
252 object->ref_count++;
253 VM_OBJECT_UNLOCK(object);
254 }
255 vref(vp);
256 return (object);
257 }
258
259 /*
260 * The object must be locked.
261 */
262 static void
263 vnode_pager_dealloc(object)
264 vm_object_t object;
265 {
266 struct vnode *vp = object->handle;
267
268 if (vp == NULL)
269 panic("vnode_pager_dealloc: pager already dealloced");
270
271 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
272 vm_object_pip_wait(object, "vnpdea");
273
274 object->handle = NULL;
275 object->type = OBJT_DEAD;
276 if (object->flags & OBJ_DISCONNECTWNT) {
277 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
278 wakeup(object);
279 }
280 ASSERT_VOP_LOCKED(vp, "vnode_pager_dealloc");
281 vp->v_object = NULL;
282 vp->v_vflag &= ~VV_TEXT;
283 }
284
285 static boolean_t
286 vnode_pager_haspage(object, pindex, before, after)
287 vm_object_t object;
288 vm_pindex_t pindex;
289 int *before;
290 int *after;
291 {
292 struct vnode *vp = object->handle;
293 daddr_t bn;
294 int err;
295 daddr_t reqblock;
296 int poff;
297 int bsize;
298 int pagesperblock, blocksperpage;
299 int vfslocked;
300
301 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
302 /*
303 * If no vp or vp is doomed or marked transparent to VM, we do not
304 * have the page.
305 */
306 if (vp == NULL || vp->v_iflag & VI_DOOMED)
307 return FALSE;
308 /*
309 * If the offset is beyond end of file we do
310 * not have the page.
311 */
312 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
313 return FALSE;
314
315 bsize = vp->v_mount->mnt_stat.f_iosize;
316 pagesperblock = bsize / PAGE_SIZE;
317 blocksperpage = 0;
318 if (pagesperblock > 0) {
319 reqblock = pindex / pagesperblock;
320 } else {
321 blocksperpage = (PAGE_SIZE / bsize);
322 reqblock = pindex * blocksperpage;
323 }
324 VM_OBJECT_UNLOCK(object);
325 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
326 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
327 VFS_UNLOCK_GIANT(vfslocked);
328 VM_OBJECT_LOCK(object);
329 if (err)
330 return TRUE;
331 if (bn == -1)
332 return FALSE;
333 if (pagesperblock > 0) {
334 poff = pindex - (reqblock * pagesperblock);
335 if (before) {
336 *before *= pagesperblock;
337 *before += poff;
338 }
339 if (after) {
340 int numafter;
341 *after *= pagesperblock;
342 numafter = pagesperblock - (poff + 1);
343 if (IDX_TO_OFF(pindex + numafter) >
344 object->un_pager.vnp.vnp_size) {
345 numafter =
346 OFF_TO_IDX(object->un_pager.vnp.vnp_size) -
347 pindex;
348 }
349 *after += numafter;
350 }
351 } else {
352 if (before) {
353 *before /= blocksperpage;
354 }
355
356 if (after) {
357 *after /= blocksperpage;
358 }
359 }
360 return TRUE;
361 }
362
363 /*
364 * Lets the VM system know about a change in size for a file.
365 * We adjust our own internal size and flush any cached pages in
366 * the associated object that are affected by the size change.
367 *
368 * Note: this routine may be invoked as a result of a pager put
369 * operation (possibly at object termination time), so we must be careful.
370 */
371 void
372 vnode_pager_setsize(vp, nsize)
373 struct vnode *vp;
374 vm_ooffset_t nsize;
375 {
376 vm_object_t object;
377 vm_page_t m;
378 vm_pindex_t nobjsize;
379
380 if ((object = vp->v_object) == NULL)
381 return;
382 VM_OBJECT_LOCK(object);
383 if (nsize == object->un_pager.vnp.vnp_size) {
384 /*
385 * Hasn't changed size
386 */
387 VM_OBJECT_UNLOCK(object);
388 return;
389 }
390 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
391 if (nsize < object->un_pager.vnp.vnp_size) {
392 /*
393 * File has shrunk. Toss any cached pages beyond the new EOF.
394 */
395 if (nobjsize < object->size)
396 vm_object_page_remove(object, nobjsize, object->size,
397 FALSE);
398 /*
399 * this gets rid of garbage at the end of a page that is now
400 * only partially backed by the vnode.
401 *
402 * XXX for some reason (I don't know yet), if we take a
403 * completely invalid page and mark it partially valid
404 * it can screw up NFS reads, so we don't allow the case.
405 */
406 if ((nsize & PAGE_MASK) &&
407 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
408 m->valid != 0) {
409 int base = (int)nsize & PAGE_MASK;
410 int size = PAGE_SIZE - base;
411
412 /*
413 * Clear out partial-page garbage in case
414 * the page has been mapped.
415 */
416 pmap_zero_page_area(m, base, size);
417
418 /*
419 * XXX work around SMP data integrity race
420 * by unmapping the page from user processes.
421 * The garbage we just cleared may be mapped
422 * to a user process running on another cpu
423 * and this code is not running through normal
424 * I/O channels which handle SMP issues for
425 * us, so unmap page to synchronize all cpus.
426 *
427 * XXX should vm_pager_unmap_page() have
428 * dealt with this?
429 */
430 vm_page_lock_queues();
431 pmap_remove_all(m);
432
433 /*
434 * Clear out partial-page dirty bits. This
435 * has the side effect of setting the valid
436 * bits, but that is ok. There are a bunch
437 * of places in the VM system where we expected
438 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
439 * case is one of them. If the page is still
440 * partially dirty, make it fully dirty.
441 *
442 * note that we do not clear out the valid
443 * bits. This would prevent bogus_page
444 * replacement from working properly.
445 */
446 vm_page_set_validclean(m, base, size);
447 if (m->dirty != 0)
448 m->dirty = VM_PAGE_BITS_ALL;
449 vm_page_unlock_queues();
450 }
451 }
452 object->un_pager.vnp.vnp_size = nsize;
453 object->size = nobjsize;
454 VM_OBJECT_UNLOCK(object);
455 }
456
457 /*
458 * calculate the linear (byte) disk address of specified virtual
459 * file address
460 */
461 static daddr_t
462 vnode_pager_addr(vp, address, run)
463 struct vnode *vp;
464 vm_ooffset_t address;
465 int *run;
466 {
467 daddr_t rtaddress;
468 int bsize;
469 daddr_t block;
470 int err;
471 daddr_t vblock;
472 daddr_t voffset;
473
474 if (address < 0)
475 return -1;
476
477 if (vp->v_iflag & VI_DOOMED)
478 return -1;
479
480 bsize = vp->v_mount->mnt_stat.f_iosize;
481 vblock = address / bsize;
482 voffset = address % bsize;
483
484 err = VOP_BMAP(vp, vblock, NULL, &block, run, NULL);
485
486 if (err || (block == -1))
487 rtaddress = -1;
488 else {
489 rtaddress = block + voffset / DEV_BSIZE;
490 if (run) {
491 *run += 1;
492 *run *= bsize/PAGE_SIZE;
493 *run -= voffset/PAGE_SIZE;
494 }
495 }
496
497 return rtaddress;
498 }
499
500 /*
501 * small block filesystem vnode pager input
502 */
503 static int
504 vnode_pager_input_smlfs(object, m)
505 vm_object_t object;
506 vm_page_t m;
507 {
508 int i;
509 struct vnode *vp;
510 struct bufobj *bo;
511 struct buf *bp;
512 struct sf_buf *sf;
513 daddr_t fileaddr;
514 vm_offset_t bsize;
515 int error = 0;
516
517 vp = object->handle;
518 if (vp->v_iflag & VI_DOOMED)
519 return VM_PAGER_BAD;
520
521 bsize = vp->v_mount->mnt_stat.f_iosize;
522
523 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
524
525 sf = sf_buf_alloc(m, 0);
526
527 for (i = 0; i < PAGE_SIZE / bsize; i++) {
528 vm_ooffset_t address;
529
530 if (vm_page_bits(i * bsize, bsize) & m->valid)
531 continue;
532
533 address = IDX_TO_OFF(m->pindex) + i * bsize;
534 if (address >= object->un_pager.vnp.vnp_size) {
535 fileaddr = -1;
536 } else {
537 fileaddr = vnode_pager_addr(vp, address, NULL);
538 }
539 if (fileaddr != -1) {
540 bp = getpbuf(&vnode_pbuf_freecnt);
541
542 /* build a minimal buffer header */
543 bp->b_iocmd = BIO_READ;
544 bp->b_iodone = bdone;
545 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
546 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
547 bp->b_rcred = crhold(curthread->td_ucred);
548 bp->b_wcred = crhold(curthread->td_ucred);
549 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
550 bp->b_blkno = fileaddr;
551 pbgetbo(bo, bp);
552 bp->b_bcount = bsize;
553 bp->b_bufsize = bsize;
554 bp->b_runningbufspace = bp->b_bufsize;
555 atomic_add_int(&runningbufspace, bp->b_runningbufspace);
556
557 /* do the input */
558 bp->b_iooffset = dbtob(bp->b_blkno);
559 bstrategy(bp);
560
561 bwait(bp, PVM, "vnsrd");
562
563 if ((bp->b_ioflags & BIO_ERROR) != 0)
564 error = EIO;
565
566 /*
567 * free the buffer header back to the swap buffer pool
568 */
569 pbrelbo(bp);
570 relpbuf(bp, &vnode_pbuf_freecnt);
571 if (error)
572 break;
573
574 VM_OBJECT_LOCK(object);
575 vm_page_lock_queues();
576 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
577 vm_page_unlock_queues();
578 VM_OBJECT_UNLOCK(object);
579 } else {
580 VM_OBJECT_LOCK(object);
581 vm_page_lock_queues();
582 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
583 vm_page_unlock_queues();
584 VM_OBJECT_UNLOCK(object);
585 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
586 }
587 }
588 sf_buf_free(sf);
589 vm_page_lock_queues();
590 pmap_clear_modify(m);
591 vm_page_unlock_queues();
592 if (error) {
593 return VM_PAGER_ERROR;
594 }
595 return VM_PAGER_OK;
596
597 }
598
599
600 /*
601 * old style vnode pager input routine
602 */
603 static int
604 vnode_pager_input_old(object, m)
605 vm_object_t object;
606 vm_page_t m;
607 {
608 struct uio auio;
609 struct iovec aiov;
610 int error;
611 int size;
612 struct sf_buf *sf;
613 struct vnode *vp;
614
615 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
616 error = 0;
617
618 /*
619 * Return failure if beyond current EOF
620 */
621 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
622 return VM_PAGER_BAD;
623 } else {
624 size = PAGE_SIZE;
625 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
626 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
627 vp = object->handle;
628 VM_OBJECT_UNLOCK(object);
629
630 /*
631 * Allocate a kernel virtual address and initialize so that
632 * we can use VOP_READ/WRITE routines.
633 */
634 sf = sf_buf_alloc(m, 0);
635
636 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
637 aiov.iov_len = size;
638 auio.uio_iov = &aiov;
639 auio.uio_iovcnt = 1;
640 auio.uio_offset = IDX_TO_OFF(m->pindex);
641 auio.uio_segflg = UIO_SYSSPACE;
642 auio.uio_rw = UIO_READ;
643 auio.uio_resid = size;
644 auio.uio_td = curthread;
645
646 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
647 if (!error) {
648 int count = size - auio.uio_resid;
649
650 if (count == 0)
651 error = EINVAL;
652 else if (count != PAGE_SIZE)
653 bzero((caddr_t)sf_buf_kva(sf) + count,
654 PAGE_SIZE - count);
655 }
656 sf_buf_free(sf);
657
658 VM_OBJECT_LOCK(object);
659 }
660 vm_page_lock_queues();
661 pmap_clear_modify(m);
662 vm_page_undirty(m);
663 vm_page_unlock_queues();
664 if (!error)
665 m->valid = VM_PAGE_BITS_ALL;
666 return error ? VM_PAGER_ERROR : VM_PAGER_OK;
667 }
668
669 /*
670 * generic vnode pager input routine
671 */
672
673 /*
674 * Local media VFS's that do not implement their own VOP_GETPAGES
675 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
676 * to implement the previous behaviour.
677 *
678 * All other FS's should use the bypass to get to the local media
679 * backing vp's VOP_GETPAGES.
680 */
681 static int
682 vnode_pager_getpages(object, m, count, reqpage)
683 vm_object_t object;
684 vm_page_t *m;
685 int count;
686 int reqpage;
687 {
688 int rtval;
689 struct vnode *vp;
690 int bytes = count * PAGE_SIZE;
691 int vfslocked;
692
693 vp = object->handle;
694 VM_OBJECT_UNLOCK(object);
695 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
696 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
697 KASSERT(rtval != EOPNOTSUPP,
698 ("vnode_pager: FS getpages not implemented\n"));
699 VFS_UNLOCK_GIANT(vfslocked);
700 VM_OBJECT_LOCK(object);
701 return rtval;
702 }
703
704 /*
705 * This is now called from local media FS's to operate against their
706 * own vnodes if they fail to implement VOP_GETPAGES.
707 */
708 int
709 vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
710 struct vnode *vp;
711 vm_page_t *m;
712 int bytecount;
713 int reqpage;
714 {
715 vm_object_t object;
716 vm_offset_t kva;
717 off_t foff, tfoff, nextoff;
718 int i, j, size, bsize, first;
719 daddr_t firstaddr, reqblock;
720 struct bufobj *bo;
721 int runpg;
722 int runend;
723 struct buf *bp;
724 int count;
725 int error = 0;
726
727 object = vp->v_object;
728 count = bytecount / PAGE_SIZE;
729
730 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
731 ("vnode_pager_generic_getpages does not support devices"));
732 if (vp->v_iflag & VI_DOOMED)
733 return VM_PAGER_BAD;
734
735 bsize = vp->v_mount->mnt_stat.f_iosize;
736
737 /* get the UNDERLYING device for the file with VOP_BMAP() */
738
739 /*
740 * originally, we did not check for an error return value -- assuming
741 * an fs always has a bmap entry point -- that assumption is wrong!!!
742 */
743 foff = IDX_TO_OFF(m[reqpage]->pindex);
744
745 /*
746 * if we can't bmap, use old VOP code
747 */
748 if (VOP_BMAP(vp, foff / bsize, &bo, &reqblock, NULL, NULL)) {
749 VM_OBJECT_LOCK(object);
750 vm_page_lock_queues();
751 for (i = 0; i < count; i++)
752 if (i != reqpage)
753 vm_page_free(m[i]);
754 vm_page_unlock_queues();
755 cnt.v_vnodein++;
756 cnt.v_vnodepgsin++;
757 error = vnode_pager_input_old(object, m[reqpage]);
758 VM_OBJECT_UNLOCK(object);
759 return (error);
760
761 /*
762 * if the blocksize is smaller than a page size, then use
763 * special small filesystem code. NFS sometimes has a small
764 * blocksize, but it can handle large reads itself.
765 */
766 } else if ((PAGE_SIZE / bsize) > 1 &&
767 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
768 VM_OBJECT_LOCK(object);
769 vm_page_lock_queues();
770 for (i = 0; i < count; i++)
771 if (i != reqpage)
772 vm_page_free(m[i]);
773 vm_page_unlock_queues();
774 VM_OBJECT_UNLOCK(object);
775 cnt.v_vnodein++;
776 cnt.v_vnodepgsin++;
777 return vnode_pager_input_smlfs(object, m[reqpage]);
778 }
779
780 /*
781 * If we have a completely valid page available to us, we can
782 * clean up and return. Otherwise we have to re-read the
783 * media.
784 */
785 VM_OBJECT_LOCK(object);
786 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
787 vm_page_lock_queues();
788 for (i = 0; i < count; i++)
789 if (i != reqpage)
790 vm_page_free(m[i]);
791 vm_page_unlock_queues();
792 VM_OBJECT_UNLOCK(object);
793 return VM_PAGER_OK;
794 } else if (reqblock == -1) {
795 pmap_zero_page(m[reqpage]);
796 vm_page_undirty(m[reqpage]);
797 m[reqpage]->valid = VM_PAGE_BITS_ALL;
798 vm_page_lock_queues();
799 for (i = 0; i < count; i++)
800 if (i != reqpage)
801 vm_page_free(m[i]);
802 vm_page_unlock_queues();
803 VM_OBJECT_UNLOCK(object);
804 return (VM_PAGER_OK);
805 }
806 m[reqpage]->valid = 0;
807 VM_OBJECT_UNLOCK(object);
808
809 /*
810 * here on direct device I/O
811 */
812 firstaddr = -1;
813
814 /*
815 * calculate the run that includes the required page
816 */
817 for (first = 0, i = 0; i < count; i = runend) {
818 firstaddr = vnode_pager_addr(vp,
819 IDX_TO_OFF(m[i]->pindex), &runpg);
820 if (firstaddr == -1) {
821 VM_OBJECT_LOCK(object);
822 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
823 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
824 (intmax_t)firstaddr, (uintmax_t)(foff >> 32),
825 (uintmax_t)foff,
826 (uintmax_t)
827 (object->un_pager.vnp.vnp_size >> 32),
828 (uintmax_t)object->un_pager.vnp.vnp_size);
829 }
830 vm_page_lock_queues();
831 vm_page_free(m[i]);
832 vm_page_unlock_queues();
833 VM_OBJECT_UNLOCK(object);
834 runend = i + 1;
835 first = runend;
836 continue;
837 }
838 runend = i + runpg;
839 if (runend <= reqpage) {
840 VM_OBJECT_LOCK(object);
841 vm_page_lock_queues();
842 for (j = i; j < runend; j++)
843 vm_page_free(m[j]);
844 vm_page_unlock_queues();
845 VM_OBJECT_UNLOCK(object);
846 } else {
847 if (runpg < (count - first)) {
848 VM_OBJECT_LOCK(object);
849 vm_page_lock_queues();
850 for (i = first + runpg; i < count; i++)
851 vm_page_free(m[i]);
852 vm_page_unlock_queues();
853 VM_OBJECT_UNLOCK(object);
854 count = first + runpg;
855 }
856 break;
857 }
858 first = runend;
859 }
860
861 /*
862 * the first and last page have been calculated now, move input pages
863 * to be zero based...
864 */
865 if (first != 0) {
866 for (i = first; i < count; i++) {
867 m[i - first] = m[i];
868 }
869 count -= first;
870 reqpage -= first;
871 }
872
873 /*
874 * calculate the file virtual address for the transfer
875 */
876 foff = IDX_TO_OFF(m[0]->pindex);
877
878 /*
879 * calculate the size of the transfer
880 */
881 size = count * PAGE_SIZE;
882 KASSERT(count > 0, ("zero count"));
883 if ((foff + size) > object->un_pager.vnp.vnp_size)
884 size = object->un_pager.vnp.vnp_size - foff;
885 KASSERT(size > 0, ("zero size"));
886
887 /*
888 * round up physical size for real devices.
889 */
890 if (1) {
891 int secmask = bo->bo_bsize - 1;
892 KASSERT(secmask < PAGE_SIZE && secmask > 0,
893 ("vnode_pager_generic_getpages: sector size %d too large",
894 secmask + 1));
895 size = (size + secmask) & ~secmask;
896 }
897
898 bp = getpbuf(&vnode_pbuf_freecnt);
899 kva = (vm_offset_t) bp->b_data;
900
901 /*
902 * and map the pages to be read into the kva
903 */
904 pmap_qenter(kva, m, count);
905
906 /* build a minimal buffer header */
907 bp->b_iocmd = BIO_READ;
908 bp->b_iodone = bdone;
909 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
910 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
911 bp->b_rcred = crhold(curthread->td_ucred);
912 bp->b_wcred = crhold(curthread->td_ucred);
913 bp->b_blkno = firstaddr;
914 pbgetbo(bo, bp);
915 bp->b_bcount = size;
916 bp->b_bufsize = size;
917 bp->b_runningbufspace = bp->b_bufsize;
918 atomic_add_int(&runningbufspace, bp->b_runningbufspace);
919
920 cnt.v_vnodein++;
921 cnt.v_vnodepgsin += count;
922
923 /* do the input */
924 bp->b_iooffset = dbtob(bp->b_blkno);
925 bstrategy(bp);
926
927 bwait(bp, PVM, "vnread");
928
929 if ((bp->b_ioflags & BIO_ERROR) != 0)
930 error = EIO;
931
932 if (!error) {
933 if (size != count * PAGE_SIZE)
934 bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
935 }
936 pmap_qremove(kva, count);
937
938 /*
939 * free the buffer header back to the swap buffer pool
940 */
941 pbrelbo(bp);
942 relpbuf(bp, &vnode_pbuf_freecnt);
943
944 VM_OBJECT_LOCK(object);
945 vm_page_lock_queues();
946 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
947 vm_page_t mt;
948
949 nextoff = tfoff + PAGE_SIZE;
950 mt = m[i];
951
952 if (nextoff <= object->un_pager.vnp.vnp_size) {
953 /*
954 * Read filled up entire page.
955 */
956 mt->valid = VM_PAGE_BITS_ALL;
957 vm_page_undirty(mt); /* should be an assert? XXX */
958 pmap_clear_modify(mt);
959 } else {
960 /*
961 * Read did not fill up entire page. Since this
962 * is getpages, the page may be mapped, so we have
963 * to zero the invalid portions of the page even
964 * though we aren't setting them valid.
965 *
966 * Currently we do not set the entire page valid,
967 * we just try to clear the piece that we couldn't
968 * read.
969 */
970 vm_page_set_validclean(mt, 0,
971 object->un_pager.vnp.vnp_size - tfoff);
972 /* handled by vm_fault now */
973 /* vm_page_zero_invalid(mt, FALSE); */
974 }
975
976 if (i != reqpage) {
977
978 /*
979 * whether or not to leave the page activated is up in
980 * the air, but we should put the page on a page queue
981 * somewhere. (it already is in the object). Result:
982 * It appears that empirical results show that
983 * deactivating pages is best.
984 */
985
986 /*
987 * just in case someone was asking for this page we
988 * now tell them that it is ok to use
989 */
990 if (!error) {
991 if (mt->flags & PG_WANTED)
992 vm_page_activate(mt);
993 else
994 vm_page_deactivate(mt);
995 vm_page_wakeup(mt);
996 } else {
997 vm_page_free(mt);
998 }
999 }
1000 }
1001 vm_page_unlock_queues();
1002 VM_OBJECT_UNLOCK(object);
1003 if (error) {
1004 printf("vnode_pager_getpages: I/O read error\n");
1005 }
1006 return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
1007 }
1008
1009 /*
1010 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
1011 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1012 * vnode_pager_generic_putpages() to implement the previous behaviour.
1013 *
1014 * All other FS's should use the bypass to get to the local media
1015 * backing vp's VOP_PUTPAGES.
1016 */
1017 static void
1018 vnode_pager_putpages(object, m, count, sync, rtvals)
1019 vm_object_t object;
1020 vm_page_t *m;
1021 int count;
1022 boolean_t sync;
1023 int *rtvals;
1024 {
1025 int rtval;
1026 struct vnode *vp;
1027 struct mount *mp;
1028 int bytes = count * PAGE_SIZE;
1029
1030 /*
1031 * Force synchronous operation if we are extremely low on memory
1032 * to prevent a low-memory deadlock. VOP operations often need to
1033 * allocate more memory to initiate the I/O ( i.e. do a BMAP
1034 * operation ). The swapper handles the case by limiting the amount
1035 * of asynchronous I/O, but that sort of solution doesn't scale well
1036 * for the vnode pager without a lot of work.
1037 *
1038 * Also, the backing vnode's iodone routine may not wake the pageout
1039 * daemon up. This should be probably be addressed XXX.
1040 */
1041
1042 if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
1043 sync |= OBJPC_SYNC;
1044
1045 /*
1046 * Call device-specific putpages function
1047 */
1048 vp = object->handle;
1049 VM_OBJECT_UNLOCK(object);
1050 if (vp->v_type != VREG)
1051 mp = NULL;
1052 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
1053 KASSERT(rtval != EOPNOTSUPP,
1054 ("vnode_pager: stale FS putpages\n"));
1055 VM_OBJECT_LOCK(object);
1056 }
1057
1058
1059 /*
1060 * This is now called from local media FS's to operate against their
1061 * own vnodes if they fail to implement VOP_PUTPAGES.
1062 *
1063 * This is typically called indirectly via the pageout daemon and
1064 * clustering has already typically occured, so in general we ask the
1065 * underlying filesystem to write the data out asynchronously rather
1066 * then delayed.
1067 */
1068 int
1069 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals)
1070 struct vnode *vp;
1071 vm_page_t *m;
1072 int bytecount;
1073 int flags;
1074 int *rtvals;
1075 {
1076 int i;
1077 vm_object_t object;
1078 int count;
1079
1080 int maxsize, ncount;
1081 vm_ooffset_t poffset;
1082 struct uio auio;
1083 struct iovec aiov;
1084 int error;
1085 int ioflags;
1086 int ppscheck = 0;
1087 static struct timeval lastfail;
1088 static int curfail;
1089
1090 object = vp->v_object;
1091 count = bytecount / PAGE_SIZE;
1092
1093 for (i = 0; i < count; i++)
1094 rtvals[i] = VM_PAGER_AGAIN;
1095
1096 if ((int64_t)m[0]->pindex < 0) {
1097 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1098 (long)m[0]->pindex, (u_long)m[0]->dirty);
1099 rtvals[0] = VM_PAGER_BAD;
1100 return VM_PAGER_BAD;
1101 }
1102
1103 maxsize = count * PAGE_SIZE;
1104 ncount = count;
1105
1106 poffset = IDX_TO_OFF(m[0]->pindex);
1107
1108 /*
1109 * If the page-aligned write is larger then the actual file we
1110 * have to invalidate pages occuring beyond the file EOF. However,
1111 * there is an edge case where a file may not be page-aligned where
1112 * the last page is partially invalid. In this case the filesystem
1113 * may not properly clear the dirty bits for the entire page (which
1114 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1115 * With the page locked we are free to fix-up the dirty bits here.
1116 *
1117 * We do not under any circumstances truncate the valid bits, as
1118 * this will screw up bogus page replacement.
1119 */
1120 if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1121 if (object->un_pager.vnp.vnp_size > poffset) {
1122 int pgoff;
1123
1124 maxsize = object->un_pager.vnp.vnp_size - poffset;
1125 ncount = btoc(maxsize);
1126 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1127 vm_page_lock_queues();
1128 vm_page_clear_dirty(m[ncount - 1], pgoff,
1129 PAGE_SIZE - pgoff);
1130 vm_page_unlock_queues();
1131 }
1132 } else {
1133 maxsize = 0;
1134 ncount = 0;
1135 }
1136 if (ncount < count) {
1137 for (i = ncount; i < count; i++) {
1138 rtvals[i] = VM_PAGER_BAD;
1139 }
1140 }
1141 }
1142
1143 /*
1144 * pageouts are already clustered, use IO_ASYNC t o force a bawrite()
1145 * rather then a bdwrite() to prevent paging I/O from saturating
1146 * the buffer cache. Dummy-up the sequential heuristic to cause
1147 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
1148 * the system decides how to cluster.
1149 */
1150 ioflags = IO_VMIO;
1151 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1152 ioflags |= IO_SYNC;
1153 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1154 ioflags |= IO_ASYNC;
1155 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1156 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1157
1158 aiov.iov_base = (caddr_t) 0;
1159 aiov.iov_len = maxsize;
1160 auio.uio_iov = &aiov;
1161 auio.uio_iovcnt = 1;
1162 auio.uio_offset = poffset;
1163 auio.uio_segflg = UIO_NOCOPY;
1164 auio.uio_rw = UIO_WRITE;
1165 auio.uio_resid = maxsize;
1166 auio.uio_td = (struct thread *) 0;
1167 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1168 cnt.v_vnodeout++;
1169 cnt.v_vnodepgsout += ncount;
1170
1171 if (error) {
1172 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1)))
1173 printf("vnode_pager_putpages: I/O error %d\n", error);
1174 }
1175 if (auio.uio_resid) {
1176 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1))
1177 printf("vnode_pager_putpages: residual I/O %d at %lu\n",
1178 auio.uio_resid, (u_long)m[0]->pindex);
1179 }
1180 for (i = 0; i < ncount; i++) {
1181 rtvals[i] = VM_PAGER_OK;
1182 }
1183 return rtvals[0];
1184 }
1185
1186 struct vnode *
1187 vnode_pager_lock(vm_object_t first_object)
1188 {
1189 struct vnode *vp;
1190 vm_object_t backing_object, object;
1191
1192 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
1193 for (object = first_object; object != NULL; object = backing_object) {
1194 if (object->type != OBJT_VNODE) {
1195 if ((backing_object = object->backing_object) != NULL)
1196 VM_OBJECT_LOCK(backing_object);
1197 if (object != first_object)
1198 VM_OBJECT_UNLOCK(object);
1199 continue;
1200 }
1201 retry:
1202 if (object->flags & OBJ_DEAD) {
1203 if (object != first_object)
1204 VM_OBJECT_UNLOCK(object);
1205 return NULL;
1206 }
1207 vp = object->handle;
1208 VI_LOCK(vp);
1209 VM_OBJECT_UNLOCK(object);
1210 if (first_object != object)
1211 VM_OBJECT_UNLOCK(first_object);
1212 VFS_ASSERT_GIANT(vp->v_mount);
1213 if (vget(vp, LK_CANRECURSE | LK_INTERLOCK |
1214 LK_RETRY | LK_SHARED, curthread)) {
1215 VM_OBJECT_LOCK(first_object);
1216 if (object != first_object)
1217 VM_OBJECT_LOCK(object);
1218 if (object->type != OBJT_VNODE) {
1219 if (object != first_object)
1220 VM_OBJECT_UNLOCK(object);
1221 return NULL;
1222 }
1223 printf("vnode_pager_lock: retrying\n");
1224 goto retry;
1225 }
1226 VM_OBJECT_LOCK(first_object);
1227 return (vp);
1228 }
1229 return NULL;
1230 }
Cache object: fad02ae144fde8d2c41f743837643ee8
|