kernel/vm: Remove some unused variables.
[dragonfly.git] / sys / vm / vnode_pager.c
CommitLineData
984263bc 1/*
0ac0a48a
MD
2 * (MPSAFE)
3 *
984263bc
MD
4 * Copyright (c) 1990 University of Utah.
5 * Copyright (c) 1991 The Regents of the University of California.
6 * All rights reserved.
7 * Copyright (c) 1993, 1994 John S. Dyson
8 * Copyright (c) 1995, David Greenman
9 *
10 * This code is derived from software contributed to Berkeley by
11 * the Systems Programming Group of the University of Utah Computer
12 * Science Department.
13 *
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
16 * are met:
17 * 1. Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 * 2. Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in the
21 * documentation and/or other materials provided with the distribution.
22 * 3. All advertising materials mentioning features or use of this software
23 * must display the following acknowledgement:
24 * This product includes software developed by the University of
25 * California, Berkeley and its contributors.
26 * 4. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
29 *
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40 * SUCH DAMAGE.
41 *
42 * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91
43 * $FreeBSD: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $
44 */
45
46/*
47 * Page to/from files (vnodes).
48 */
49
50/*
51 * TODO:
52 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
53 * greatly re-simplify the vnode_pager.
54 */
55
56#include <sys/param.h>
57#include <sys/systm.h>
5fd012e0 58#include <sys/kernel.h>
984263bc
MD
59#include <sys/proc.h>
60#include <sys/vnode.h>
61#include <sys/mount.h>
62#include <sys/buf.h>
63#include <sys/vmmeter.h>
64#include <sys/conf.h>
5c5185ae
SG
65
66#include <cpu/lwbuf.h>
984263bc
MD
67
68#include <vm/vm.h>
69#include <vm/vm_object.h>
70#include <vm/vm_page.h>
71#include <vm/vm_pager.h>
72#include <vm/vm_map.h>
73#include <vm/vnode_pager.h>
8d292090 74#include <vm/swap_pager.h>
984263bc
MD
75#include <vm/vm_extern.h>
76
1a54183b
MD
77#include <sys/thread2.h>
78#include <vm/vm_page2.h>
79
1388df65 80static void vnode_pager_dealloc (vm_object_t);
1b9d3514 81static int vnode_pager_getpage (vm_object_t, vm_page_t *, int);
1388df65 82static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *);
1b9d3514 83static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t);
984263bc
MD
84
85struct pagerops vnodepagerops = {
984263bc 86 vnode_pager_dealloc,
1b9d3514 87 vnode_pager_getpage,
984263bc 88 vnode_pager_putpages,
107e9bcc 89 vnode_pager_haspage
984263bc
MD
90};
91
35f59bfa
MD
92static struct krate vbadrate = { 1 };
93static struct krate vresrate = { 1 };
94
984263bc
MD
95int vnode_pbuf_freecnt = -1; /* start out unlimited */
96
97/*
b0d18f7d
MD
98 * Allocate a VM object for a vnode, typically a regular file vnode.
99 *
100 * Some additional information is required to generate a properly sized
101 * object which covers the entire buffer cache buffer straddling the file
102 * EOF. Userland does not see the extra pages as the VM fault code tests
103 * against v_filesize.
984263bc
MD
104 */
105vm_object_t
b0d18f7d
MD
106vnode_pager_alloc(void *handle, off_t length, vm_prot_t prot, off_t offset,
107 int blksize, int boff)
984263bc
MD
108{
109 vm_object_t object;
110 struct vnode *vp;
b0d18f7d
MD
111 off_t loffset;
112 vm_pindex_t lsize;
984263bc
MD
113
114 /*
115 * Pageout to vnode, no can do yet.
116 */
117 if (handle == NULL)
118 return (NULL);
119
120 /*
121 * XXX hack - This initialization should be put somewhere else.
122 */
123 if (vnode_pbuf_freecnt < 0) {
124 vnode_pbuf_freecnt = nswbuf / 2 + 1;
125 }
126
2de4f77e
MD
127 /*
128 * Serialize potential vnode/object teardowns and interlocks
129 */
130 vp = (struct vnode *)handle;
b12defdc 131 lwkt_gettoken(&vp->v_token);
984263bc
MD
132
133 /*
134 * Prevent race condition when allocating the object. This
135 * can happen with NFS vnodes since the nfsnode isn't locked.
136 */
137 while (vp->v_flag & VOLOCK) {
2247fe02 138 vsetflags(vp, VOWANT);
377d4740 139 tsleep(vp, 0, "vnpobj", 0);
984263bc 140 }
2247fe02 141 vsetflags(vp, VOLOCK);
b12defdc 142 lwkt_reltoken(&vp->v_token);
984263bc
MD
143
144 /*
145 * If the object is being terminated, wait for it to
146 * go away.
147 */
b12defdc
MD
148 while ((object = vp->v_object) != NULL) {
149 vm_object_hold(object);
150 if ((object->flags & OBJ_DEAD) == 0)
151 break;
9e12ff11 152 vm_object_dead_sleep(object, "vadead");
b12defdc 153 vm_object_drop(object);
984263bc
MD
154 }
155
3c37c940 156 if (vp->v_sysref.refcnt <= 0)
984263bc
MD
157 panic("vnode_pager_alloc: no vnode reference");
158
b0d18f7d
MD
159 /*
160 * Round up to the *next* block, then destroy the buffers in question.
161 * Since we are only removing some of the buffers we must rely on the
162 * scan count to determine whether a loop is necessary.
163 *
164 * Destroy any pages beyond the last buffer.
165 */
166 if (boff < 0)
167 boff = (int)(length % blksize);
168 if (boff)
169 loffset = length + (blksize - boff);
170 else
171 loffset = length;
172 lsize = OFF_TO_IDX(round_page64(loffset));
173
984263bc
MD
174 if (object == NULL) {
175 /*
176 * And an object of the appropriate size
177 */
2b9050f5 178 object = vm_object_allocate_hold(OBJT_VNODE, lsize);
984263bc
MD
179 object->handle = handle;
180 vp->v_object = object;
b0d18f7d 181 vp->v_filesize = length;
2bc7505b 182 if (vp->v_mount && (vp->v_mount->mnt_kern_flag & MNTK_NOMSYNC))
2a418930 183 vm_object_set_flag(object, OBJ_NOMSYNC);
984263bc 184 } else {
b12defdc 185 object->ref_count++;
b0d18f7d
MD
186 if (object->size != lsize) {
187 kprintf("vnode_pager_alloc: Warning, objsize "
188 "mismatch %jd/%jd vp=%p obj=%p\n",
189 (intmax_t)object->size,
190 (intmax_t)lsize,
191 vp, object);
192 }
193 if (vp->v_filesize != length) {
973c11b9 194 kprintf("vnode_pager_alloc: Warning, filesize "
b0d18f7d
MD
195 "mismatch %jd/%jd vp=%p obj=%p\n",
196 (intmax_t)vp->v_filesize,
197 (intmax_t)length,
198 vp, object);
973c11b9 199 }
984263bc
MD
200 }
201
2de4f77e 202 vref(vp);
b12defdc 203 lwkt_gettoken(&vp->v_token);
2247fe02 204 vclrflags(vp, VOLOCK);
984263bc 205 if (vp->v_flag & VOWANT) {
5a648714
MD
206 vclrflags(vp, VOWANT);
207 wakeup(vp);
208 }
b12defdc
MD
209 lwkt_reltoken(&vp->v_token);
210
211 vm_object_drop(object);
2de4f77e 212
5a648714
MD
213 return (object);
214}
215
216/*
217 * Add a ref to a vnode's existing VM object, return the object or
218 * NULL if the vnode did not have one. This does not create the
219 * object (we can't since we don't know what the proper blocksize/boff
220 * is to match the VFS's use of the buffer cache).
221 */
222vm_object_t
223vnode_pager_reference(struct vnode *vp)
224{
225 vm_object_t object;
226
227 /*
228 * Prevent race condition when allocating the object. This
229 * can happen with NFS vnodes since the nfsnode isn't locked.
b12defdc
MD
230 *
231 * Serialize potential vnode/object teardowns and interlocks
5a648714 232 */
b12defdc 233 lwkt_gettoken(&vp->v_token);
5a648714
MD
234 while (vp->v_flag & VOLOCK) {
235 vsetflags(vp, VOWANT);
236 tsleep(vp, 0, "vnpobj", 0);
237 }
238 vsetflags(vp, VOLOCK);
b12defdc 239 lwkt_reltoken(&vp->v_token);
5a648714
MD
240
241 /*
242 * Prevent race conditions against deallocation of the VM
243 * object.
244 */
b12defdc
MD
245 while ((object = vp->v_object) != NULL) {
246 vm_object_hold(object);
247 if ((object->flags & OBJ_DEAD) == 0)
248 break;
5a648714 249 vm_object_dead_sleep(object, "vadead");
b12defdc 250 vm_object_drop(object);
5a648714
MD
251 }
252
253 /*
254 * The object is expected to exist, the caller will handle
255 * NULL returns if it does not.
256 */
257 if (object) {
b12defdc 258 object->ref_count++;
5a648714
MD
259 vref(vp);
260 }
261
b12defdc 262 lwkt_gettoken(&vp->v_token);
5a648714
MD
263 vclrflags(vp, VOLOCK);
264 if (vp->v_flag & VOWANT) {
2247fe02 265 vclrflags(vp, VOWANT);
984263bc
MD
266 wakeup(vp);
267 }
b12defdc
MD
268 lwkt_reltoken(&vp->v_token);
269 if (object)
270 vm_object_drop(object);
2de4f77e 271
984263bc
MD
272 return (object);
273}
274
275static void
57e43348 276vnode_pager_dealloc(vm_object_t object)
984263bc 277{
5f910b2f 278 struct vnode *vp = object->handle;
984263bc
MD
279
280 if (vp == NULL)
281 panic("vnode_pager_dealloc: pager already dealloced");
282
283 vm_object_pip_wait(object, "vnpdea");
284
285 object->handle = NULL;
286 object->type = OBJT_DEAD;
287 vp->v_object = NULL;
57f7b636 288 vp->v_filesize = NOOFFSET;
2247fe02 289 vclrflags(vp, VTEXT | VOBJBUF);
8d292090 290 swap_pager_freespace_all(object);
984263bc
MD
291}
292
54078292
MD
293/*
294 * Return whether the vnode pager has the requested page. Return the
295 * number of disk-contiguous pages before and after the requested page,
296 * not including the requested page.
297 */
984263bc 298static boolean_t
1b9d3514 299vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex)
984263bc
MD
300{
301 struct vnode *vp = object->handle;
54078292
MD
302 off_t loffset;
303 off_t doffset;
304 int voff;
984263bc 305 int bsize;
54078292 306 int error;
984263bc
MD
307
308 /*
309 * If no vp or vp is doomed or marked transparent to VM, we do not
310 * have the page.
311 */
5fd012e0 312 if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
984263bc
MD
313 return FALSE;
314
315 /*
316 * If filesystem no longer mounted or offset beyond end of file we do
317 * not have the page.
318 */
54078292
MD
319 loffset = IDX_TO_OFF(pindex);
320
57f7b636 321 if (vp->v_mount == NULL || loffset >= vp->v_filesize)
984263bc
MD
322 return FALSE;
323
324 bsize = vp->v_mount->mnt_stat.f_iosize;
54078292
MD
325 voff = loffset % bsize;
326
bc823b32 327 /*
1b9d3514
MD
328 * XXX
329 *
bc823b32
MD
330 * BMAP returns byte counts before and after, where after
331 * is inclusive of the base page. haspage must return page
332 * counts before and after where after does not include the
333 * base page.
334 *
335 * BMAP is allowed to return a *after of 0 for backwards
336 * compatibility. The base page is still considered valid if
337 * no error is returned.
338 */
1b9d3514
MD
339 error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0);
340 if (error)
984263bc 341 return TRUE;
54078292 342 if (doffset == NOOFFSET)
984263bc 343 return FALSE;
984263bc
MD
344 return TRUE;
345}
346
347/*
348 * Lets the VM system know about a change in size for a file.
349 * We adjust our own internal size and flush any cached pages in
350 * the associated object that are affected by the size change.
351 *
57f7b636 352 * NOTE: This routine may be invoked as a result of a pager put
984263bc 353 * operation (possibly at object termination time), so we must be careful.
57f7b636
MD
354 *
355 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
356 * we do not blow up on the case. nsize will always be >= 0, however.
984263bc
MD
357 */
358void
57e43348 359vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
984263bc
MD
360{
361 vm_pindex_t nobjsize;
c4b46cb4 362 vm_pindex_t oobjsize;
b12defdc 363 vm_object_t object;
984263bc 364
b12defdc
MD
365 while ((object = vp->v_object) != NULL) {
366 vm_object_hold(object);
367 if (vp->v_object == object)
368 break;
369 vm_object_drop(object);
370 }
984263bc
MD
371 if (object == NULL)
372 return;
373
374 /*
375 * Hasn't changed size
376 */
b12defdc
MD
377 if (nsize == vp->v_filesize) {
378 vm_object_drop(object);
984263bc 379 return;
b12defdc 380 }
573fb415 381
c4b46cb4
MD
382 /*
383 * Has changed size. Adjust the VM object's size and v_filesize
384 * before we start scanning pages to prevent new pages from being
385 * allocated during the scan.
386 */
984263bc 387 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
c4b46cb4
MD
388 oobjsize = object->size;
389 object->size = nobjsize;
984263bc
MD
390
391 /*
392 * File has shrunk. Toss any cached pages beyond the new EOF.
393 */
57f7b636 394 if (nsize < vp->v_filesize) {
135bd6a8 395 vp->v_filesize = nsize;
c4b46cb4
MD
396 if (nobjsize < oobjsize) {
397 vm_object_page_remove(object, nobjsize, oobjsize,
398 FALSE);
984263bc
MD
399 }
400 /*
8d429613
MD
401 * This gets rid of garbage at the end of a page that is now
402 * only partially backed by the vnode. Since we are setting
403 * the entire page valid & clean after we are done we have
404 * to be sure that the portion of the page within the file
405 * bounds is already valid. If it isn't then making it
406 * valid would create a corrupt block.
984263bc
MD
407 */
408 if (nsize & PAGE_MASK) {
409 vm_offset_t kva;
410 vm_page_t m;
411
b12defdc
MD
412 m = vm_page_lookup_busy_wait(object, OFF_TO_IDX(nsize),
413 TRUE, "vsetsz");
17cde63e 414
984263bc
MD
415 if (m && m->valid) {
416 int base = (int)nsize & PAGE_MASK;
417 int size = PAGE_SIZE - base;
5c5185ae 418 struct lwbuf *lwb;
7a683a24 419 struct lwbuf lwb_cache;
984263bc
MD
420
421 /*
422 * Clear out partial-page garbage in case
423 * the page has been mapped.
c7841cbe
MD
424 *
425 * This is byte aligned.
984263bc 426 */
7a683a24 427 lwb = lwbuf_alloc(m, &lwb_cache);
5c5185ae 428 kva = lwbuf_kva(lwb);
984263bc 429 bzero((caddr_t)kva + base, size);
5c5185ae 430 lwbuf_free(lwb);
984263bc
MD
431
432 /*
433 * XXX work around SMP data integrity race
434 * by unmapping the page from user processes.
435 * The garbage we just cleared may be mapped
436 * to a user process running on another cpu
437 * and this code is not running through normal
438 * I/O channels which handle SMP issues for
439 * us, so unmap page to synchronize all cpus.
440 *
441 * XXX should vm_pager_unmap_page() have
442 * dealt with this?
443 */
444 vm_page_protect(m, VM_PROT_NONE);
445
446 /*
447 * Clear out partial-page dirty bits. This
448 * has the side effect of setting the valid
449 * bits, but that is ok. There are a bunch
450 * of places in the VM system where we expected
451 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
452 * case is one of them. If the page is still
453 * partially dirty, make it fully dirty.
454 *
c7841cbe 455 * NOTE: We do not clear out the valid
984263bc
MD
456 * bits. This would prevent bogus_page
457 * replacement from working properly.
c7841cbe
MD
458 *
459 * NOTE: We do not want to clear the dirty
460 * bit for a partial DEV_BSIZE'd truncation!
461 * This is DEV_BSIZE aligned!
984263bc 462 */
1a54183b 463 vm_page_clear_dirty_beg_nonincl(m, base, size);
984263bc
MD
464 if (m->dirty != 0)
465 m->dirty = VM_PAGE_BITS_ALL;
17cde63e 466 vm_page_wakeup(m);
b12defdc
MD
467 } else if (m) {
468 vm_page_wakeup(m);
984263bc
MD
469 }
470 }
135bd6a8
MD
471 } else {
472 vp->v_filesize = nsize;
984263bc 473 }
b12defdc 474 vm_object_drop(object);
984263bc
MD
475}
476
a55afca2
MD
477/*
478 * Release a page busied for a getpages operation. The page may have become
479 * wired (typically due to being used by the buffer cache) or otherwise been
480 * soft-busied and cannot be freed in that case. A held page can still be
481 * freed.
482 */
984263bc 483void
57e43348 484vnode_pager_freepage(vm_page_t m)
984263bc 485{
9bf025db 486 if (m->busy || m->wire_count || (m->flags & PG_NEED_COMMIT)) {
a55afca2
MD
487 vm_page_activate(m);
488 vm_page_wakeup(m);
489 } else {
490 vm_page_free(m);
491 }
984263bc
MD
492}
493
494/*
984263bc
MD
495 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
496 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
497 * vnode_pager_generic_getpages() to implement the previous behaviour.
498 *
499 * All other FS's should use the bypass to get to the local media
500 * backing vp's VOP_GETPAGES.
501 */
502static int
1b9d3514 503vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
984263bc
MD
504{
505 int rtval;
506 struct vnode *vp;
984263bc
MD
507
508 vp = object->handle;
1c48c952 509 rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess);
a05863a7 510 if (rtval == EOPNOTSUPP)
ed20d0e3 511 panic("vnode_pager: vfs's must implement vop_getpages");
984263bc
MD
512 return rtval;
513}
514
984263bc
MD
515/*
516 * This is now called from local media FS's to operate against their
517 * own vnodes if they fail to implement VOP_GETPAGES.
a05863a7
MD
518 *
519 * With all the caching local media devices do these days there is really
520 * very little point to attempting to restrict the I/O size to contiguous
521 * blocks on-disk, especially if our caller thinks we need all the specified
522 * pages. Just construct and issue a READ.
984263bc
MD
523 */
524int
1b9d3514 525vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount,
1c48c952 526 int reqpage, int seqaccess)
984263bc 527{
a05863a7
MD
528 struct iovec aiov;
529 struct uio auio;
530 off_t foff;
531 int error;
984263bc 532 int count;
a05863a7
MD
533 int i;
534 int ioflags;
984263bc
MD
535
536 /*
a05863a7 537 * Do not do anything if the vnode is bad.
984263bc 538 */
a05863a7
MD
539 if (vp->v_mount == NULL)
540 return VM_PAGER_BAD;
984263bc
MD
541
542 /*
a05863a7
MD
543 * Calculate the number of pages. Since we are paging in whole
544 * pages, adjust bytecount to be an integral multiple of the page
545 * size. It will be clipped to the file EOF later on.
984263bc 546 */
a05863a7
MD
547 bytecount = round_page(bytecount);
548 count = bytecount / PAGE_SIZE;
984263bc
MD
549
550 /*
cf1bb2a8
MD
551 * We could check m[reqpage]->valid here and shortcut the operation,
552 * but doing so breaks read-ahead. Instead assume that the VM
553 * system has already done at least the check, don't worry about
554 * any races, and issue the VOP_READ to allow read-ahead to function.
8d429613 555 *
cf1bb2a8
MD
556 * This keeps the pipeline full for I/O bound sequentially scanned
557 * mmap()'s
984263bc 558 */
cf1bb2a8 559 /* don't shortcut */
984263bc
MD
560
561 /*
a05863a7
MD
562 * Discard pages past the file EOF. If the requested page is past
563 * the file EOF we just leave its valid bits set to 0, the caller
564 * expects to maintain ownership of the requested page. If the
565 * entire range is past file EOF discard everything and generate
566 * a pagein error.
984263bc 567 */
1b9d3514 568 foff = IDX_TO_OFF(mpp[0]->pindex);
a05863a7
MD
569 if (foff >= vp->v_filesize) {
570 for (i = 0; i < count; i++) {
571 if (i != reqpage)
1b9d3514 572 vnode_pager_freepage(mpp[i]);
984263bc 573 }
a05863a7 574 return VM_PAGER_ERROR;
984263bc
MD
575 }
576
a05863a7
MD
577 if (foff + bytecount > vp->v_filesize) {
578 bytecount = vp->v_filesize - foff;
579 i = round_page(bytecount) / PAGE_SIZE;
580 while (count > i) {
581 --count;
582 if (count != reqpage)
1b9d3514 583 vnode_pager_freepage(mpp[count]);
984263bc 584 }
984263bc
MD
585 }
586
587 /*
a05863a7
MD
588 * The size of the transfer is bytecount. bytecount will be an
589 * integral multiple of the page size unless it has been clipped
590 * to the file EOF. The transfer cannot exceed the file EOF.
591 *
592 * When dealing with real devices we must round-up to the device
593 * sector size.
984263bc 594 */
a05863a7
MD
595 if (vp->v_type == VBLK || vp->v_type == VCHR) {
596 int secmask = vp->v_rdev->si_bsize_phys - 1;
ed20d0e3 597 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large", secmask + 1));
a05863a7 598 bytecount = (bytecount + secmask) & ~secmask;
984263bc
MD
599 }
600
984263bc 601 /*
ca28958c
MD
602 * Severe hack to avoid deadlocks with the buffer cache
603 */
604 for (i = 0; i < count; ++i) {
1b9d3514 605 vm_page_t mt = mpp[i];
ca28958c
MD
606
607 vm_page_io_start(mt);
608 vm_page_wakeup(mt);
609 }
610
611 /*
cf1bb2a8 612 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE
984263bc 613 */
a05863a7 614 ioflags = IO_VMIO;
1c48c952 615 if (seqaccess)
cf1bb2a8 616 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
a05863a7 617
1b9d3514 618 aiov.iov_base = NULL;
a05863a7
MD
619 aiov.iov_len = bytecount;
620 auio.uio_iov = &aiov;
621 auio.uio_iovcnt = 1;
622 auio.uio_offset = foff;
623 auio.uio_segflg = UIO_NOCOPY;
624 auio.uio_rw = UIO_READ;
625 auio.uio_resid = bytecount;
626 auio.uio_td = NULL;
12e4aaff
MD
627 mycpu->gd_cnt.v_vnodein++;
628 mycpu->gd_cnt.v_vnodepgsin += count;
984263bc 629
a05863a7 630 error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
984263bc
MD
631
632 /*
ca28958c
MD
633 * Severe hack to avoid deadlocks with the buffer cache
634 */
635 for (i = 0; i < count; ++i) {
b12defdc
MD
636 vm_page_busy_wait(mpp[i], FALSE, "getpgs");
637 vm_page_io_finish(mpp[i]);
ca28958c
MD
638 }
639
640 /*
a05863a7
MD
641 * Calculate the actual number of bytes read and clean up the
642 * page list.
984263bc 643 */
a05863a7 644 bytecount -= auio.uio_resid;
984263bc 645
a05863a7 646 for (i = 0; i < count; ++i) {
1b9d3514 647 vm_page_t mt = mpp[i];
a55afca2 648
a05863a7 649 if (i != reqpage) {
a55afca2 650 if (error == 0 && mt->valid) {
b12defdc 651 if (mt->flags & PG_REFERENCED)
984263bc
MD
652 vm_page_activate(mt);
653 else
654 vm_page_deactivate(mt);
655 vm_page_wakeup(mt);
656 } else {
657 vnode_pager_freepage(mt);
658 }
a55afca2
MD
659 } else if (mt->valid == 0) {
660 if (error == 0) {
fc9ed34d
MD
661 kprintf("page failed but no I/O error page "
662 "%p object %p pindex %d\n",
663 mt, mt->object, (int) mt->pindex);
a55afca2
MD
664 /* whoops, something happened */
665 error = EINVAL;
666 }
667 } else if (mt->valid != VM_PAGE_BITS_ALL) {
668 /*
669 * Zero-extend the requested page if necessary (if
670 * the filesystem is using a small block size).
671 */
672 vm_page_zero_invalid(mt, TRUE);
984263bc
MD
673 }
674 }
675 if (error) {
1b9d3514 676 kprintf("vnode_pager_getpage: I/O read error\n");
984263bc
MD
677 }
678 return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
679}
680
681/*
682 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
683 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
684 * vnode_pager_generic_putpages() to implement the previous behaviour.
685 *
cb1cf930
MD
686 * Caller has already cleared the pmap modified bits, if any.
687 *
984263bc
MD
688 * All other FS's should use the bypass to get to the local media
689 * backing vp's VOP_PUTPAGES.
690 */
691static void
57e43348 692vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
cb1cf930 693 boolean_t sync, int *rtvals)
984263bc
MD
694{
695 int rtval;
696 struct vnode *vp;
697 int bytes = count * PAGE_SIZE;
698
699 /*
700 * Force synchronous operation if we are extremely low on memory
701 * to prevent a low-memory deadlock. VOP operations often need to
702 * allocate more memory to initiate the I/O ( i.e. do a BMAP
703 * operation ). The swapper handles the case by limiting the amount
704 * of asynchronous I/O, but that sort of solution doesn't scale well
705 * for the vnode pager without a lot of work.
706 *
707 * Also, the backing vnode's iodone routine may not wake the pageout
708 * daemon up. This should be probably be addressed XXX.
709 */
710
0a4d4828
MD
711 if ((vmstats.v_free_count + vmstats.v_cache_count) <
712 vmstats.v_pageout_free_min) {
984263bc 713 sync |= OBJPC_SYNC;
0a4d4828 714 }
984263bc
MD
715
716 /*
717 * Call device-specific putpages function
718 */
984263bc
MD
719 vp = object->handle;
720 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
721 if (rtval == EOPNOTSUPP) {
086c1d7e 722 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
984263bc
MD
723 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
724 }
725}
726
727
728/*
729 * This is now called from local media FS's to operate against their
730 * own vnodes if they fail to implement VOP_PUTPAGES.
731 *
732 * This is typically called indirectly via the pageout daemon and
733 * clustering has already typically occured, so in general we ask the
734 * underlying filesystem to write the data out asynchronously rather
735 * then delayed.
736 */
737int
57e43348 738vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
a05863a7 739 int flags, int *rtvals)
984263bc
MD
740{
741 int i;
cb1cf930 742 int maxsize, ncount, count;
984263bc
MD
743 vm_ooffset_t poffset;
744 struct uio auio;
745 struct iovec aiov;
746 int error;
747 int ioflags;
748
984263bc
MD
749 count = bytecount / PAGE_SIZE;
750
751 for (i = 0; i < count; i++)
752 rtvals[i] = VM_PAGER_AGAIN;
753
754 if ((int) m[0]->pindex < 0) {
086c1d7e 755 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
984263bc
MD
756 (long)m[0]->pindex, m[0]->dirty);
757 rtvals[0] = VM_PAGER_BAD;
758 return VM_PAGER_BAD;
759 }
760
761 maxsize = count * PAGE_SIZE;
762 ncount = count;
763
764 poffset = IDX_TO_OFF(m[0]->pindex);
765
766 /*
767 * If the page-aligned write is larger then the actual file we
cb1cf930
MD
768 * have to invalidate pages occuring beyond the file EOF.
769 *
770 * If the file EOF resides in the middle of a page we still clear
771 * all of that page's dirty bits later on. If we didn't it would
772 * endlessly re-write.
984263bc
MD
773 *
774 * We do not under any circumstances truncate the valid bits, as
775 * this will screw up bogus page replacement.
93afe6be
MD
776 *
777 * The caller has already read-protected the pages. The VFS must
778 * use the buffer cache to wrap the pages. The pages might not
779 * be immediately flushed by the buffer cache but once under its
780 * control the pages themselves can wind up being marked clean
781 * and their covering buffer cache buffer can be marked dirty.
984263bc 782 */
fb69d1c8
MD
783 if (poffset + maxsize > vp->v_filesize) {
784 if (poffset < vp->v_filesize) {
57f7b636 785 maxsize = vp->v_filesize - poffset;
984263bc 786 ncount = btoc(maxsize);
984263bc
MD
787 } else {
788 maxsize = 0;
789 ncount = 0;
790 }
791 if (ncount < count) {
792 for (i = ncount; i < count; i++) {
793 rtvals[i] = VM_PAGER_BAD;
794 }
795 }
796 }
797
798 /*
799 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
800 * rather then a bdwrite() to prevent paging I/O from saturating
801 * the buffer cache. Dummy-up the sequential heuristic to cause
802 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
803 * the system decides how to cluster.
804 */
805 ioflags = IO_VMIO;
806 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
807 ioflags |= IO_SYNC;
808 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
809 ioflags |= IO_ASYNC;
810 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
811 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
812
813 aiov.iov_base = (caddr_t) 0;
814 aiov.iov_len = maxsize;
815 auio.uio_iov = &aiov;
816 auio.uio_iovcnt = 1;
817 auio.uio_offset = poffset;
818 auio.uio_segflg = UIO_NOCOPY;
819 auio.uio_rw = UIO_WRITE;
820 auio.uio_resid = maxsize;
dadab5e9 821 auio.uio_td = NULL;
8a8d5d85 822 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
12e4aaff
MD
823 mycpu->gd_cnt.v_vnodeout++;
824 mycpu->gd_cnt.v_vnodepgsout += ncount;
984263bc
MD
825
826 if (error) {
35f59bfa
MD
827 krateprintf(&vbadrate,
828 "vnode_pager_putpages: I/O error %d\n", error);
984263bc
MD
829 }
830 if (auio.uio_resid) {
35f59bfa 831 krateprintf(&vresrate,
bfc09ba0 832 "vnode_pager_putpages: residual I/O %zd at %lu\n",
35f59bfa 833 auio.uio_resid, (u_long)m[0]->pindex);
984263bc 834 }
cb1cf930
MD
835 if (error == 0) {
836 for (i = 0; i < ncount; i++) {
837 rtvals[i] = VM_PAGER_OK;
838 vm_page_undirty(m[i]);
839 }
840 }
984263bc
MD
841 return rtvals[0];
842}
843
b12defdc
MD
844/*
845 * Run the chain and if the bottom-most object is a vnode-type lock the
846 * underlying vnode. A locked vnode or NULL is returned.
847 */
984263bc 848struct vnode *
dadab5e9 849vnode_pager_lock(vm_object_t object)
984263bc 850{
b12defdc
MD
851 struct vnode *vp = NULL;
852 vm_object_t lobject;
853 vm_object_t tobject;
5fd012e0 854 int error;
984263bc 855
b12defdc
MD
856 if (object == NULL)
857 return(NULL);
984263bc 858
b12defdc
MD
859 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
860 lobject = object;
861
862 while (lobject->type != OBJT_VNODE) {
863 if (lobject->flags & OBJ_DEAD)
864 break;
865 tobject = lobject->backing_object;
866 if (tobject == NULL)
867 break;
868 vm_object_hold(tobject);
869 if (tobject == lobject->backing_object) {
870 if (lobject != object) {
871 vm_object_lock_swap();
872 vm_object_drop(lobject);
dadab5e9 873 }
b12defdc
MD
874 lobject = tobject;
875 } else {
876 vm_object_drop(tobject);
877 }
878 }
879 while (lobject->type == OBJT_VNODE &&
880 (lobject->flags & OBJ_DEAD) == 0) {
881 /*
882 * Extract the vp
883 */
884 vp = lobject->handle;
885 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
886 if (error == 0) {
887 if (lobject->handle == vp)
888 break;
889 vput(vp);
890 } else {
891 kprintf("vnode_pager_lock: vp %p error %d "
892 "lockstatus %d, retrying\n",
893 vp, error,
894 lockstatus(&vp->v_lock, curthread));
5fd012e0 895 tsleep(object->handle, 0, "vnpgrl", hz);
984263bc 896 }
b12defdc 897 vp = NULL;
984263bc 898 }
b12defdc
MD
899 if (lobject != object)
900 vm_object_drop(lobject);
901 return (vp);
984263bc 902}