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/9.1/sys/vm/vnode_pager.c 234466 2012-04-19 18:03:24Z mckusick $");
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 int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
79 daddr_t *rtaddress, 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,
87 vm_ooffset_t, struct ucred *cred);
88
89 struct pagerops vnodepagerops = {
90 .pgo_alloc = vnode_pager_alloc,
91 .pgo_dealloc = vnode_pager_dealloc,
92 .pgo_getpages = vnode_pager_getpages,
93 .pgo_putpages = vnode_pager_putpages,
94 .pgo_haspage = vnode_pager_haspage,
95 };
96
97 int vnode_pbuf_freecnt;
98
99 /* Create the VM system backing object for this vnode */
100 int
101 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
102 {
103 vm_object_t object;
104 vm_ooffset_t size = isize;
105 struct vattr va;
106
107 if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE)
108 return (0);
109
110 while ((object = vp->v_object) != NULL) {
111 VM_OBJECT_LOCK(object);
112 if (!(object->flags & OBJ_DEAD)) {
113 VM_OBJECT_UNLOCK(object);
114 return (0);
115 }
116 VOP_UNLOCK(vp, 0);
117 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
118 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vodead", 0);
119 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
120 }
121
122 if (size == 0) {
123 if (vn_isdisk(vp, NULL)) {
124 size = IDX_TO_OFF(INT_MAX);
125 } else {
126 if (VOP_GETATTR(vp, &va, td->td_ucred))
127 return (0);
128 size = va.va_size;
129 }
130 }
131
132 object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred);
133 /*
134 * Dereference the reference we just created. This assumes
135 * that the object is associated with the vp.
136 */
137 VM_OBJECT_LOCK(object);
138 object->ref_count--;
139 VM_OBJECT_UNLOCK(object);
140 vrele(vp);
141
142 KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
143
144 return (0);
145 }
146
147 void
148 vnode_destroy_vobject(struct vnode *vp)
149 {
150 struct vm_object *obj;
151
152 obj = vp->v_object;
153 if (obj == NULL)
154 return;
155 ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
156 VM_OBJECT_LOCK(obj);
157 if (obj->ref_count == 0) {
158 /*
159 * vclean() may be called twice. The first time
160 * removes the primary reference to the object,
161 * the second time goes one further and is a
162 * special-case to terminate the object.
163 *
164 * don't double-terminate the object
165 */
166 if ((obj->flags & OBJ_DEAD) == 0)
167 vm_object_terminate(obj);
168 else
169 VM_OBJECT_UNLOCK(obj);
170 } else {
171 /*
172 * Woe to the process that tries to page now :-).
173 */
174 vm_pager_deallocate(obj);
175 VM_OBJECT_UNLOCK(obj);
176 }
177 vp->v_object = NULL;
178 }
179
180
181 /*
182 * Allocate (or lookup) pager for a vnode.
183 * Handle is a vnode pointer.
184 *
185 * MPSAFE
186 */
187 vm_object_t
188 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
189 vm_ooffset_t offset, struct ucred *cred)
190 {
191 vm_object_t object;
192 struct vnode *vp;
193
194 /*
195 * Pageout to vnode, no can do yet.
196 */
197 if (handle == NULL)
198 return (NULL);
199
200 vp = (struct vnode *) handle;
201
202 /*
203 * If the object is being terminated, wait for it to
204 * go away.
205 */
206 retry:
207 while ((object = vp->v_object) != NULL) {
208 VM_OBJECT_LOCK(object);
209 if ((object->flags & OBJ_DEAD) == 0)
210 break;
211 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
212 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vadead", 0);
213 }
214
215 if (vp->v_usecount == 0)
216 panic("vnode_pager_alloc: no vnode reference");
217
218 if (object == NULL) {
219 /*
220 * Add an object of the appropriate size
221 */
222 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
223
224 object->un_pager.vnp.vnp_size = size;
225 object->un_pager.vnp.writemappings = 0;
226
227 object->handle = handle;
228 VI_LOCK(vp);
229 if (vp->v_object != NULL) {
230 /*
231 * Object has been created while we were sleeping
232 */
233 VI_UNLOCK(vp);
234 vm_object_destroy(object);
235 goto retry;
236 }
237 vp->v_object = object;
238 VI_UNLOCK(vp);
239 } else {
240 object->ref_count++;
241 VM_OBJECT_UNLOCK(object);
242 }
243 vref(vp);
244 return (object);
245 }
246
247 /*
248 * The object must be locked.
249 */
250 static void
251 vnode_pager_dealloc(object)
252 vm_object_t object;
253 {
254 struct vnode *vp;
255 int refs;
256
257 vp = object->handle;
258 if (vp == NULL)
259 panic("vnode_pager_dealloc: pager already dealloced");
260
261 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
262 vm_object_pip_wait(object, "vnpdea");
263 refs = object->ref_count;
264
265 object->handle = NULL;
266 object->type = OBJT_DEAD;
267 if (object->flags & OBJ_DISCONNECTWNT) {
268 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
269 wakeup(object);
270 }
271 ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
272 if (object->un_pager.vnp.writemappings > 0) {
273 object->un_pager.vnp.writemappings = 0;
274 vp->v_writecount--;
275 }
276 vp->v_object = NULL;
277 vp->v_vflag &= ~VV_TEXT;
278 VM_OBJECT_UNLOCK(object);
279 while (refs-- > 0)
280 vunref(vp);
281 VM_OBJECT_LOCK(object);
282 }
283
284 static boolean_t
285 vnode_pager_haspage(object, pindex, before, after)
286 vm_object_t object;
287 vm_pindex_t pindex;
288 int *before;
289 int *after;
290 {
291 struct vnode *vp = object->handle;
292 daddr_t bn;
293 int err;
294 daddr_t reqblock;
295 int poff;
296 int bsize;
297 int pagesperblock, blocksperpage;
298 int vfslocked;
299
300 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
301 /*
302 * If no vp or vp is doomed or marked transparent to VM, we do not
303 * have the page.
304 */
305 if (vp == NULL || vp->v_iflag & VI_DOOMED)
306 return FALSE;
307 /*
308 * If the offset is beyond end of file we do
309 * not have the page.
310 */
311 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
312 return FALSE;
313
314 bsize = vp->v_mount->mnt_stat.f_iosize;
315 pagesperblock = bsize / PAGE_SIZE;
316 blocksperpage = 0;
317 if (pagesperblock > 0) {
318 reqblock = pindex / pagesperblock;
319 } else {
320 blocksperpage = (PAGE_SIZE / bsize);
321 reqblock = pindex * blocksperpage;
322 }
323 VM_OBJECT_UNLOCK(object);
324 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
325 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
326 VFS_UNLOCK_GIANT(vfslocked);
327 VM_OBJECT_LOCK(object);
328 if (err)
329 return TRUE;
330 if (bn == -1)
331 return FALSE;
332 if (pagesperblock > 0) {
333 poff = pindex - (reqblock * pagesperblock);
334 if (before) {
335 *before *= pagesperblock;
336 *before += poff;
337 }
338 if (after) {
339 int numafter;
340 *after *= pagesperblock;
341 numafter = pagesperblock - (poff + 1);
342 if (IDX_TO_OFF(pindex + numafter) >
343 object->un_pager.vnp.vnp_size) {
344 numafter =
345 OFF_TO_IDX(object->un_pager.vnp.vnp_size) -
346 pindex;
347 }
348 *after += numafter;
349 }
350 } else {
351 if (before) {
352 *before /= blocksperpage;
353 }
354
355 if (after) {
356 *after /= blocksperpage;
357 }
358 }
359 return TRUE;
360 }
361
362 /*
363 * Lets the VM system know about a change in size for a file.
364 * We adjust our own internal size and flush any cached pages in
365 * the associated object that are affected by the size change.
366 *
367 * Note: this routine may be invoked as a result of a pager put
368 * operation (possibly at object termination time), so we must be careful.
369 */
370 void
371 vnode_pager_setsize(vp, nsize)
372 struct vnode *vp;
373 vm_ooffset_t nsize;
374 {
375 vm_object_t object;
376 vm_page_t m;
377 vm_pindex_t nobjsize;
378
379 if ((object = vp->v_object) == NULL)
380 return;
381 /* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */
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 0);
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 * Update the valid bits to reflect the blocks that
420 * have been zeroed. Some of these valid bits may
421 * have already been set.
422 */
423 vm_page_set_valid(m, base, size);
424
425 /*
426 * Round "base" to the next block boundary so that the
427 * dirty bit for a partially zeroed block is not
428 * cleared.
429 */
430 base = roundup2(base, DEV_BSIZE);
431
432 /*
433 * Clear out partial-page dirty bits.
434 *
435 * note that we do not clear out the valid
436 * bits. This would prevent bogus_page
437 * replacement from working properly.
438 */
439 vm_page_clear_dirty(m, base, PAGE_SIZE - base);
440 } else if ((nsize & PAGE_MASK) &&
441 __predict_false(object->cache != NULL)) {
442 vm_page_cache_free(object, OFF_TO_IDX(nsize),
443 nobjsize);
444 }
445 }
446 object->un_pager.vnp.vnp_size = nsize;
447 object->size = nobjsize;
448 VM_OBJECT_UNLOCK(object);
449 }
450
451 /*
452 * calculate the linear (byte) disk address of specified virtual
453 * file address
454 */
455 static int
456 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
457 int *run)
458 {
459 int bsize;
460 int err;
461 daddr_t vblock;
462 daddr_t voffset;
463
464 if (address < 0)
465 return -1;
466
467 if (vp->v_iflag & VI_DOOMED)
468 return -1;
469
470 bsize = vp->v_mount->mnt_stat.f_iosize;
471 vblock = address / bsize;
472 voffset = address % bsize;
473
474 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
475 if (err == 0) {
476 if (*rtaddress != -1)
477 *rtaddress += voffset / DEV_BSIZE;
478 if (run) {
479 *run += 1;
480 *run *= bsize/PAGE_SIZE;
481 *run -= voffset/PAGE_SIZE;
482 }
483 }
484
485 return (err);
486 }
487
488 /*
489 * small block filesystem vnode pager input
490 */
491 static int
492 vnode_pager_input_smlfs(object, m)
493 vm_object_t object;
494 vm_page_t m;
495 {
496 struct vnode *vp;
497 struct bufobj *bo;
498 struct buf *bp;
499 struct sf_buf *sf;
500 daddr_t fileaddr;
501 vm_offset_t bsize;
502 vm_page_bits_t bits;
503 int error, i;
504
505 error = 0;
506 vp = object->handle;
507 if (vp->v_iflag & VI_DOOMED)
508 return VM_PAGER_BAD;
509
510 bsize = vp->v_mount->mnt_stat.f_iosize;
511
512 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
513
514 sf = sf_buf_alloc(m, 0);
515
516 for (i = 0; i < PAGE_SIZE / bsize; i++) {
517 vm_ooffset_t address;
518
519 bits = vm_page_bits(i * bsize, bsize);
520 if (m->valid & bits)
521 continue;
522
523 address = IDX_TO_OFF(m->pindex) + i * bsize;
524 if (address >= object->un_pager.vnp.vnp_size) {
525 fileaddr = -1;
526 } else {
527 error = vnode_pager_addr(vp, address, &fileaddr, NULL);
528 if (error)
529 break;
530 }
531 if (fileaddr != -1) {
532 bp = getpbuf(&vnode_pbuf_freecnt);
533
534 /* build a minimal buffer header */
535 bp->b_iocmd = BIO_READ;
536 bp->b_iodone = bdone;
537 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
538 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
539 bp->b_rcred = crhold(curthread->td_ucred);
540 bp->b_wcred = crhold(curthread->td_ucred);
541 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
542 bp->b_blkno = fileaddr;
543 pbgetbo(bo, bp);
544 bp->b_vp = vp;
545 bp->b_bcount = bsize;
546 bp->b_bufsize = bsize;
547 bp->b_runningbufspace = bp->b_bufsize;
548 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
549
550 /* do the input */
551 bp->b_iooffset = dbtob(bp->b_blkno);
552 bstrategy(bp);
553
554 bwait(bp, PVM, "vnsrd");
555
556 if ((bp->b_ioflags & BIO_ERROR) != 0)
557 error = EIO;
558
559 /*
560 * free the buffer header back to the swap buffer pool
561 */
562 bp->b_vp = NULL;
563 pbrelbo(bp);
564 relpbuf(bp, &vnode_pbuf_freecnt);
565 if (error)
566 break;
567 } else
568 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
569 KASSERT((m->dirty & bits) == 0,
570 ("vnode_pager_input_smlfs: page %p is dirty", m));
571 VM_OBJECT_LOCK(object);
572 m->valid |= bits;
573 VM_OBJECT_UNLOCK(object);
574 }
575 sf_buf_free(sf);
576 if (error) {
577 return VM_PAGER_ERROR;
578 }
579 return VM_PAGER_OK;
580 }
581
582 /*
583 * old style vnode pager input routine
584 */
585 static int
586 vnode_pager_input_old(object, m)
587 vm_object_t object;
588 vm_page_t m;
589 {
590 struct uio auio;
591 struct iovec aiov;
592 int error;
593 int size;
594 struct sf_buf *sf;
595 struct vnode *vp;
596
597 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
598 error = 0;
599
600 /*
601 * Return failure if beyond current EOF
602 */
603 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
604 return VM_PAGER_BAD;
605 } else {
606 size = PAGE_SIZE;
607 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
608 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
609 vp = object->handle;
610 VM_OBJECT_UNLOCK(object);
611
612 /*
613 * Allocate a kernel virtual address and initialize so that
614 * we can use VOP_READ/WRITE routines.
615 */
616 sf = sf_buf_alloc(m, 0);
617
618 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
619 aiov.iov_len = size;
620 auio.uio_iov = &aiov;
621 auio.uio_iovcnt = 1;
622 auio.uio_offset = IDX_TO_OFF(m->pindex);
623 auio.uio_segflg = UIO_SYSSPACE;
624 auio.uio_rw = UIO_READ;
625 auio.uio_resid = size;
626 auio.uio_td = curthread;
627
628 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
629 if (!error) {
630 int count = size - auio.uio_resid;
631
632 if (count == 0)
633 error = EINVAL;
634 else if (count != PAGE_SIZE)
635 bzero((caddr_t)sf_buf_kva(sf) + count,
636 PAGE_SIZE - count);
637 }
638 sf_buf_free(sf);
639
640 VM_OBJECT_LOCK(object);
641 }
642 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
643 if (!error)
644 m->valid = VM_PAGE_BITS_ALL;
645 return error ? VM_PAGER_ERROR : VM_PAGER_OK;
646 }
647
648 /*
649 * generic vnode pager input routine
650 */
651
652 /*
653 * Local media VFS's that do not implement their own VOP_GETPAGES
654 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
655 * to implement the previous behaviour.
656 *
657 * All other FS's should use the bypass to get to the local media
658 * backing vp's VOP_GETPAGES.
659 */
660 static int
661 vnode_pager_getpages(object, m, count, reqpage)
662 vm_object_t object;
663 vm_page_t *m;
664 int count;
665 int reqpage;
666 {
667 int rtval;
668 struct vnode *vp;
669 int bytes = count * PAGE_SIZE;
670 int vfslocked;
671
672 vp = object->handle;
673 VM_OBJECT_UNLOCK(object);
674 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
675 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
676 KASSERT(rtval != EOPNOTSUPP,
677 ("vnode_pager: FS getpages not implemented\n"));
678 VFS_UNLOCK_GIANT(vfslocked);
679 VM_OBJECT_LOCK(object);
680 return rtval;
681 }
682
683 /*
684 * This is now called from local media FS's to operate against their
685 * own vnodes if they fail to implement VOP_GETPAGES.
686 */
687 int
688 vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
689 struct vnode *vp;
690 vm_page_t *m;
691 int bytecount;
692 int reqpage;
693 {
694 vm_object_t object;
695 vm_offset_t kva;
696 off_t foff, tfoff, nextoff;
697 int i, j, size, bsize, first;
698 daddr_t firstaddr, reqblock;
699 struct bufobj *bo;
700 int runpg;
701 int runend;
702 struct buf *bp;
703 int count;
704 int error;
705
706 object = vp->v_object;
707 count = bytecount / PAGE_SIZE;
708
709 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
710 ("vnode_pager_generic_getpages does not support devices"));
711 if (vp->v_iflag & VI_DOOMED)
712 return VM_PAGER_BAD;
713
714 bsize = vp->v_mount->mnt_stat.f_iosize;
715
716 /* get the UNDERLYING device for the file with VOP_BMAP() */
717
718 /*
719 * originally, we did not check for an error return value -- assuming
720 * an fs always has a bmap entry point -- that assumption is wrong!!!
721 */
722 foff = IDX_TO_OFF(m[reqpage]->pindex);
723
724 /*
725 * if we can't bmap, use old VOP code
726 */
727 error = VOP_BMAP(vp, foff / bsize, &bo, &reqblock, NULL, NULL);
728 if (error == EOPNOTSUPP) {
729 VM_OBJECT_LOCK(object);
730
731 for (i = 0; i < count; i++)
732 if (i != reqpage) {
733 vm_page_lock(m[i]);
734 vm_page_free(m[i]);
735 vm_page_unlock(m[i]);
736 }
737 PCPU_INC(cnt.v_vnodein);
738 PCPU_INC(cnt.v_vnodepgsin);
739 error = vnode_pager_input_old(object, m[reqpage]);
740 VM_OBJECT_UNLOCK(object);
741 return (error);
742 } else if (error != 0) {
743 VM_OBJECT_LOCK(object);
744 for (i = 0; i < count; i++)
745 if (i != reqpage) {
746 vm_page_lock(m[i]);
747 vm_page_free(m[i]);
748 vm_page_unlock(m[i]);
749 }
750 VM_OBJECT_UNLOCK(object);
751 return (VM_PAGER_ERROR);
752
753 /*
754 * if the blocksize is smaller than a page size, then use
755 * special small filesystem code. NFS sometimes has a small
756 * blocksize, but it can handle large reads itself.
757 */
758 } else if ((PAGE_SIZE / bsize) > 1 &&
759 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
760 VM_OBJECT_LOCK(object);
761 for (i = 0; i < count; i++)
762 if (i != reqpage) {
763 vm_page_lock(m[i]);
764 vm_page_free(m[i]);
765 vm_page_unlock(m[i]);
766 }
767 VM_OBJECT_UNLOCK(object);
768 PCPU_INC(cnt.v_vnodein);
769 PCPU_INC(cnt.v_vnodepgsin);
770 return vnode_pager_input_smlfs(object, m[reqpage]);
771 }
772
773 /*
774 * If we have a completely valid page available to us, we can
775 * clean up and return. Otherwise we have to re-read the
776 * media.
777 */
778 VM_OBJECT_LOCK(object);
779 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
780 for (i = 0; i < count; i++)
781 if (i != reqpage) {
782 vm_page_lock(m[i]);
783 vm_page_free(m[i]);
784 vm_page_unlock(m[i]);
785 }
786 VM_OBJECT_UNLOCK(object);
787 return VM_PAGER_OK;
788 } else if (reqblock == -1) {
789 pmap_zero_page(m[reqpage]);
790 KASSERT(m[reqpage]->dirty == 0,
791 ("vnode_pager_generic_getpages: page %p is dirty", m));
792 m[reqpage]->valid = VM_PAGE_BITS_ALL;
793 for (i = 0; i < count; i++)
794 if (i != reqpage) {
795 vm_page_lock(m[i]);
796 vm_page_free(m[i]);
797 vm_page_unlock(m[i]);
798 }
799 VM_OBJECT_UNLOCK(object);
800 return (VM_PAGER_OK);
801 }
802 m[reqpage]->valid = 0;
803 VM_OBJECT_UNLOCK(object);
804
805 /*
806 * here on direct device I/O
807 */
808 firstaddr = -1;
809
810 /*
811 * calculate the run that includes the required page
812 */
813 for (first = 0, i = 0; i < count; i = runend) {
814 if (vnode_pager_addr(vp, IDX_TO_OFF(m[i]->pindex), &firstaddr,
815 &runpg) != 0) {
816 VM_OBJECT_LOCK(object);
817 for (; i < count; i++)
818 if (i != reqpage) {
819 vm_page_lock(m[i]);
820 vm_page_free(m[i]);
821 vm_page_unlock(m[i]);
822 }
823 VM_OBJECT_UNLOCK(object);
824 return (VM_PAGER_ERROR);
825 }
826 if (firstaddr == -1) {
827 VM_OBJECT_LOCK(object);
828 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
829 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
830 (intmax_t)firstaddr, (uintmax_t)(foff >> 32),
831 (uintmax_t)foff,
832 (uintmax_t)
833 (object->un_pager.vnp.vnp_size >> 32),
834 (uintmax_t)object->un_pager.vnp.vnp_size);
835 }
836 vm_page_lock(m[i]);
837 vm_page_free(m[i]);
838 vm_page_unlock(m[i]);
839 VM_OBJECT_UNLOCK(object);
840 runend = i + 1;
841 first = runend;
842 continue;
843 }
844 runend = i + runpg;
845 if (runend <= reqpage) {
846 VM_OBJECT_LOCK(object);
847 for (j = i; j < runend; j++) {
848 vm_page_lock(m[j]);
849 vm_page_free(m[j]);
850 vm_page_unlock(m[j]);
851 }
852 VM_OBJECT_UNLOCK(object);
853 } else {
854 if (runpg < (count - first)) {
855 VM_OBJECT_LOCK(object);
856 for (i = first + runpg; i < count; i++) {
857 vm_page_lock(m[i]);
858 vm_page_free(m[i]);
859 vm_page_unlock(m[i]);
860 }
861 VM_OBJECT_UNLOCK(object);
862 count = first + runpg;
863 }
864 break;
865 }
866 first = runend;
867 }
868
869 /*
870 * the first and last page have been calculated now, move input pages
871 * to be zero based...
872 */
873 if (first != 0) {
874 m += first;
875 count -= first;
876 reqpage -= first;
877 }
878
879 /*
880 * calculate the file virtual address for the transfer
881 */
882 foff = IDX_TO_OFF(m[0]->pindex);
883
884 /*
885 * calculate the size of the transfer
886 */
887 size = count * PAGE_SIZE;
888 KASSERT(count > 0, ("zero count"));
889 if ((foff + size) > object->un_pager.vnp.vnp_size)
890 size = object->un_pager.vnp.vnp_size - foff;
891 KASSERT(size > 0, ("zero size"));
892
893 /*
894 * round up physical size for real devices.
895 */
896 if (1) {
897 int secmask = bo->bo_bsize - 1;
898 KASSERT(secmask < PAGE_SIZE && secmask > 0,
899 ("vnode_pager_generic_getpages: sector size %d too large",
900 secmask + 1));
901 size = (size + secmask) & ~secmask;
902 }
903
904 bp = getpbuf(&vnode_pbuf_freecnt);
905 kva = (vm_offset_t) bp->b_data;
906
907 /*
908 * and map the pages to be read into the kva
909 */
910 pmap_qenter(kva, m, count);
911
912 /* build a minimal buffer header */
913 bp->b_iocmd = BIO_READ;
914 bp->b_iodone = bdone;
915 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
916 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
917 bp->b_rcred = crhold(curthread->td_ucred);
918 bp->b_wcred = crhold(curthread->td_ucred);
919 bp->b_blkno = firstaddr;
920 pbgetbo(bo, bp);
921 bp->b_vp = vp;
922 bp->b_bcount = size;
923 bp->b_bufsize = size;
924 bp->b_runningbufspace = bp->b_bufsize;
925 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
926
927 PCPU_INC(cnt.v_vnodein);
928 PCPU_ADD(cnt.v_vnodepgsin, count);
929
930 /* do the input */
931 bp->b_iooffset = dbtob(bp->b_blkno);
932 bstrategy(bp);
933
934 bwait(bp, PVM, "vnread");
935
936 if ((bp->b_ioflags & BIO_ERROR) != 0)
937 error = EIO;
938
939 if (!error) {
940 if (size != count * PAGE_SIZE)
941 bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
942 }
943 pmap_qremove(kva, count);
944
945 /*
946 * free the buffer header back to the swap buffer pool
947 */
948 bp->b_vp = NULL;
949 pbrelbo(bp);
950 relpbuf(bp, &vnode_pbuf_freecnt);
951
952 VM_OBJECT_LOCK(object);
953 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
954 vm_page_t mt;
955
956 nextoff = tfoff + PAGE_SIZE;
957 mt = m[i];
958
959 if (nextoff <= object->un_pager.vnp.vnp_size) {
960 /*
961 * Read filled up entire page.
962 */
963 mt->valid = VM_PAGE_BITS_ALL;
964 KASSERT(mt->dirty == 0,
965 ("vnode_pager_generic_getpages: page %p is dirty",
966 mt));
967 KASSERT(!pmap_page_is_mapped(mt),
968 ("vnode_pager_generic_getpages: page %p is mapped",
969 mt));
970 } else {
971 /*
972 * Read did not fill up entire page.
973 *
974 * Currently we do not set the entire page valid,
975 * we just try to clear the piece that we couldn't
976 * read.
977 */
978 vm_page_set_valid(mt, 0,
979 object->un_pager.vnp.vnp_size - tfoff);
980 KASSERT((mt->dirty & vm_page_bits(0,
981 object->un_pager.vnp.vnp_size - tfoff)) == 0,
982 ("vnode_pager_generic_getpages: page %p is dirty",
983 mt));
984 }
985
986 if (i != reqpage) {
987
988 /*
989 * whether or not to leave the page activated is up in
990 * the air, but we should put the page on a page queue
991 * somewhere. (it already is in the object). Result:
992 * It appears that empirical results show that
993 * deactivating pages is best.
994 */
995
996 /*
997 * just in case someone was asking for this page we
998 * now tell them that it is ok to use
999 */
1000 if (!error) {
1001 if (mt->oflags & VPO_WANTED) {
1002 vm_page_lock(mt);
1003 vm_page_activate(mt);
1004 vm_page_unlock(mt);
1005 } else {
1006 vm_page_lock(mt);
1007 vm_page_deactivate(mt);
1008 vm_page_unlock(mt);
1009 }
1010 vm_page_wakeup(mt);
1011 } else {
1012 vm_page_lock(mt);
1013 vm_page_free(mt);
1014 vm_page_unlock(mt);
1015 }
1016 }
1017 }
1018 VM_OBJECT_UNLOCK(object);
1019 if (error) {
1020 printf("vnode_pager_getpages: I/O read error\n");
1021 }
1022 return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
1023 }
1024
1025 /*
1026 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
1027 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1028 * vnode_pager_generic_putpages() to implement the previous behaviour.
1029 *
1030 * All other FS's should use the bypass to get to the local media
1031 * backing vp's VOP_PUTPAGES.
1032 */
1033 static void
1034 vnode_pager_putpages(object, m, count, sync, rtvals)
1035 vm_object_t object;
1036 vm_page_t *m;
1037 int count;
1038 boolean_t sync;
1039 int *rtvals;
1040 {
1041 int rtval;
1042 struct vnode *vp;
1043 int bytes = count * PAGE_SIZE;
1044
1045 /*
1046 * Force synchronous operation if we are extremely low on memory
1047 * to prevent a low-memory deadlock. VOP operations often need to
1048 * allocate more memory to initiate the I/O ( i.e. do a BMAP
1049 * operation ). The swapper handles the case by limiting the amount
1050 * of asynchronous I/O, but that sort of solution doesn't scale well
1051 * for the vnode pager without a lot of work.
1052 *
1053 * Also, the backing vnode's iodone routine may not wake the pageout
1054 * daemon up. This should be probably be addressed XXX.
1055 */
1056
1057 if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
1058 sync |= OBJPC_SYNC;
1059
1060 /*
1061 * Call device-specific putpages function
1062 */
1063 vp = object->handle;
1064 VM_OBJECT_UNLOCK(object);
1065 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
1066 KASSERT(rtval != EOPNOTSUPP,
1067 ("vnode_pager: stale FS putpages\n"));
1068 VM_OBJECT_LOCK(object);
1069 }
1070
1071
1072 /*
1073 * This is now called from local media FS's to operate against their
1074 * own vnodes if they fail to implement VOP_PUTPAGES.
1075 *
1076 * This is typically called indirectly via the pageout daemon and
1077 * clustering has already typically occured, so in general we ask the
1078 * underlying filesystem to write the data out asynchronously rather
1079 * then delayed.
1080 */
1081 int
1082 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1083 int flags, int *rtvals)
1084 {
1085 int i;
1086 vm_object_t object;
1087 vm_page_t m;
1088 int count;
1089
1090 int maxsize, ncount;
1091 vm_ooffset_t poffset;
1092 struct uio auio;
1093 struct iovec aiov;
1094 int error;
1095 int ioflags;
1096 int ppscheck = 0;
1097 static struct timeval lastfail;
1098 static int curfail;
1099
1100 object = vp->v_object;
1101 count = bytecount / PAGE_SIZE;
1102
1103 for (i = 0; i < count; i++)
1104 rtvals[i] = VM_PAGER_ERROR;
1105
1106 if ((int64_t)ma[0]->pindex < 0) {
1107 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1108 (long)ma[0]->pindex, (u_long)ma[0]->dirty);
1109 rtvals[0] = VM_PAGER_BAD;
1110 return VM_PAGER_BAD;
1111 }
1112
1113 maxsize = count * PAGE_SIZE;
1114 ncount = count;
1115
1116 poffset = IDX_TO_OFF(ma[0]->pindex);
1117
1118 /*
1119 * If the page-aligned write is larger then the actual file we
1120 * have to invalidate pages occuring beyond the file EOF. However,
1121 * there is an edge case where a file may not be page-aligned where
1122 * the last page is partially invalid. In this case the filesystem
1123 * may not properly clear the dirty bits for the entire page (which
1124 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1125 * With the page locked we are free to fix-up the dirty bits here.
1126 *
1127 * We do not under any circumstances truncate the valid bits, as
1128 * this will screw up bogus page replacement.
1129 */
1130 VM_OBJECT_LOCK(object);
1131 if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1132 if (object->un_pager.vnp.vnp_size > poffset) {
1133 int pgoff;
1134
1135 maxsize = object->un_pager.vnp.vnp_size - poffset;
1136 ncount = btoc(maxsize);
1137 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1138 /*
1139 * If the object is locked and the following
1140 * conditions hold, then the page's dirty
1141 * field cannot be concurrently changed by a
1142 * pmap operation.
1143 */
1144 m = ma[ncount - 1];
1145 KASSERT(m->busy > 0,
1146 ("vnode_pager_generic_putpages: page %p is not busy", m));
1147 KASSERT((m->aflags & PGA_WRITEABLE) == 0,
1148 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1149 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1150 pgoff);
1151 }
1152 } else {
1153 maxsize = 0;
1154 ncount = 0;
1155 }
1156 if (ncount < count) {
1157 for (i = ncount; i < count; i++) {
1158 rtvals[i] = VM_PAGER_BAD;
1159 }
1160 }
1161 }
1162 VM_OBJECT_UNLOCK(object);
1163
1164 /*
1165 * pageouts are already clustered, use IO_ASYNC t o force a bawrite()
1166 * rather then a bdwrite() to prevent paging I/O from saturating
1167 * the buffer cache. Dummy-up the sequential heuristic to cause
1168 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
1169 * the system decides how to cluster.
1170 */
1171 ioflags = IO_VMIO;
1172 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1173 ioflags |= IO_SYNC;
1174 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1175 ioflags |= IO_ASYNC;
1176 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1177 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1178
1179 aiov.iov_base = (caddr_t) 0;
1180 aiov.iov_len = maxsize;
1181 auio.uio_iov = &aiov;
1182 auio.uio_iovcnt = 1;
1183 auio.uio_offset = poffset;
1184 auio.uio_segflg = UIO_NOCOPY;
1185 auio.uio_rw = UIO_WRITE;
1186 auio.uio_resid = maxsize;
1187 auio.uio_td = (struct thread *) 0;
1188 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1189 PCPU_INC(cnt.v_vnodeout);
1190 PCPU_ADD(cnt.v_vnodepgsout, ncount);
1191
1192 if (error) {
1193 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1)))
1194 printf("vnode_pager_putpages: I/O error %d\n", error);
1195 }
1196 if (auio.uio_resid) {
1197 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1))
1198 printf("vnode_pager_putpages: residual I/O %zd at %lu\n",
1199 auio.uio_resid, (u_long)ma[0]->pindex);
1200 }
1201 for (i = 0; i < ncount; i++) {
1202 rtvals[i] = VM_PAGER_OK;
1203 }
1204 return rtvals[0];
1205 }
1206
1207 void
1208 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written)
1209 {
1210 vm_object_t obj;
1211 int i, pos;
1212
1213 if (written == 0)
1214 return;
1215 obj = ma[0]->object;
1216 VM_OBJECT_LOCK(obj);
1217 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1218 if (pos < trunc_page(written)) {
1219 rtvals[i] = VM_PAGER_OK;
1220 vm_page_undirty(ma[i]);
1221 } else {
1222 /* Partially written page. */
1223 rtvals[i] = VM_PAGER_AGAIN;
1224 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1225 }
1226 }
1227 VM_OBJECT_UNLOCK(obj);
1228 }
1229
1230 void
1231 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1232 vm_offset_t end)
1233 {
1234 struct vnode *vp;
1235 vm_ooffset_t old_wm;
1236
1237 VM_OBJECT_LOCK(object);
1238 if (object->type != OBJT_VNODE) {
1239 VM_OBJECT_UNLOCK(object);
1240 return;
1241 }
1242 old_wm = object->un_pager.vnp.writemappings;
1243 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1244 vp = object->handle;
1245 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1246 ASSERT_VOP_ELOCKED(vp, "v_writecount inc");
1247 vp->v_writecount++;
1248 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1249 ASSERT_VOP_ELOCKED(vp, "v_writecount dec");
1250 vp->v_writecount--;
1251 }
1252 VM_OBJECT_UNLOCK(object);
1253 }
1254
1255 void
1256 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1257 vm_offset_t end)
1258 {
1259 struct vnode *vp;
1260 struct mount *mp;
1261 vm_offset_t inc;
1262 int vfslocked;
1263
1264 VM_OBJECT_LOCK(object);
1265
1266 /*
1267 * First, recheck the object type to account for the race when
1268 * the vnode is reclaimed.
1269 */
1270 if (object->type != OBJT_VNODE) {
1271 VM_OBJECT_UNLOCK(object);
1272 return;
1273 }
1274
1275 /*
1276 * Optimize for the case when writemappings is not going to
1277 * zero.
1278 */
1279 inc = end - start;
1280 if (object->un_pager.vnp.writemappings != inc) {
1281 object->un_pager.vnp.writemappings -= inc;
1282 VM_OBJECT_UNLOCK(object);
1283 return;
1284 }
1285
1286 vp = object->handle;
1287 vhold(vp);
1288 VM_OBJECT_UNLOCK(object);
1289 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1290 mp = NULL;
1291 vn_start_write(vp, &mp, V_WAIT);
1292 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1293
1294 /*
1295 * Decrement the object's writemappings, by swapping the start
1296 * and end arguments for vnode_pager_update_writecount(). If
1297 * there was not a race with vnode reclaimation, then the
1298 * vnode's v_writecount is decremented.
1299 */
1300 vnode_pager_update_writecount(object, end, start);
1301 VOP_UNLOCK(vp, 0);
1302 vdrop(vp);
1303 if (mp != NULL)
1304 vn_finished_write(mp);
1305 VFS_UNLOCK_GIANT(vfslocked);
1306 }
Cache object: 48b5a0c5f417423e6131306d41d13b1a
|