kernel - Optimize the x86-64 lwbuf API
[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 $
e92ca23a 44 * $DragonFly: src/sys/vm/vnode_pager.c,v 1.43 2008/06/19 23:27:39 dillon Exp $
984263bc
MD
45 */
46
47/*
48 * Page to/from files (vnodes).
49 */
50
51/*
52 * TODO:
53 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
54 * greatly re-simplify the vnode_pager.
55 */
56
57#include <sys/param.h>
58#include <sys/systm.h>
5fd012e0 59#include <sys/kernel.h>
984263bc
MD
60#include <sys/proc.h>
61#include <sys/vnode.h>
62#include <sys/mount.h>
63#include <sys/buf.h>
64#include <sys/vmmeter.h>
65#include <sys/conf.h>
5c5185ae
SG
66
67#include <cpu/lwbuf.h>
984263bc
MD
68
69#include <vm/vm.h>
70#include <vm/vm_object.h>
71#include <vm/vm_page.h>
72#include <vm/vm_pager.h>
73#include <vm/vm_map.h>
74#include <vm/vnode_pager.h>
8d292090 75#include <vm/swap_pager.h>
984263bc
MD
76#include <vm/vm_extern.h>
77
1a54183b
MD
78#include <sys/thread2.h>
79#include <vm/vm_page2.h>
80
1388df65 81static void vnode_pager_dealloc (vm_object_t);
1b9d3514 82static int vnode_pager_getpage (vm_object_t, vm_page_t *, int);
1388df65 83static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *);
1b9d3514 84static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t);
984263bc
MD
85
86struct pagerops vnodepagerops = {
984263bc 87 vnode_pager_dealloc,
1b9d3514 88 vnode_pager_getpage,
984263bc 89 vnode_pager_putpages,
107e9bcc 90 vnode_pager_haspage
984263bc
MD
91};
92
35f59bfa
MD
93static struct krate vbadrate = { 1 };
94static struct krate vresrate = { 1 };
95
984263bc
MD
96int vnode_pbuf_freecnt = -1; /* start out unlimited */
97
98/*
b0d18f7d
MD
99 * Allocate a VM object for a vnode, typically a regular file vnode.
100 *
101 * Some additional information is required to generate a properly sized
102 * object which covers the entire buffer cache buffer straddling the file
103 * EOF. Userland does not see the extra pages as the VM fault code tests
104 * against v_filesize.
984263bc
MD
105 */
106vm_object_t
b0d18f7d
MD
107vnode_pager_alloc(void *handle, off_t length, vm_prot_t prot, off_t offset,
108 int blksize, int boff)
984263bc
MD
109{
110 vm_object_t object;
111 struct vnode *vp;
b0d18f7d
MD
112 off_t loffset;
113 vm_pindex_t lsize;
984263bc
MD
114
115 /*
116 * Pageout to vnode, no can do yet.
117 */
118 if (handle == NULL)
119 return (NULL);
120
121 /*
122 * XXX hack - This initialization should be put somewhere else.
123 */
124 if (vnode_pbuf_freecnt < 0) {
125 vnode_pbuf_freecnt = nswbuf / 2 + 1;
126 }
127
2de4f77e
MD
128 /*
129 * Serialize potential vnode/object teardowns and interlocks
130 */
131 vp = (struct vnode *)handle;
132 lwkt_gettoken(&vmobj_token);
984263bc
MD
133
134 /*
135 * Prevent race condition when allocating the object. This
136 * can happen with NFS vnodes since the nfsnode isn't locked.
137 */
138 while (vp->v_flag & VOLOCK) {
2247fe02 139 vsetflags(vp, VOWANT);
377d4740 140 tsleep(vp, 0, "vnpobj", 0);
984263bc 141 }
2247fe02 142 vsetflags(vp, VOLOCK);
984263bc
MD
143
144 /*
145 * If the object is being terminated, wait for it to
146 * go away.
147 */
148 while (((object = vp->v_object) != NULL) &&
149 (object->flags & OBJ_DEAD)) {
9e12ff11 150 vm_object_dead_sleep(object, "vadead");
984263bc
MD
151 }
152
3c37c940 153 if (vp->v_sysref.refcnt <= 0)
984263bc
MD
154 panic("vnode_pager_alloc: no vnode reference");
155
b0d18f7d
MD
156 /*
157 * Round up to the *next* block, then destroy the buffers in question.
158 * Since we are only removing some of the buffers we must rely on the
159 * scan count to determine whether a loop is necessary.
160 *
161 * Destroy any pages beyond the last buffer.
162 */
163 if (boff < 0)
164 boff = (int)(length % blksize);
165 if (boff)
166 loffset = length + (blksize - boff);
167 else
168 loffset = length;
169 lsize = OFF_TO_IDX(round_page64(loffset));
170
984263bc
MD
171 if (object == NULL) {
172 /*
173 * And an object of the appropriate size
174 */
b0d18f7d 175 object = vm_object_allocate(OBJT_VNODE, lsize);
984263bc 176 object->flags = 0;
984263bc
MD
177 object->handle = handle;
178 vp->v_object = object;
b0d18f7d 179 vp->v_filesize = length;
2bc7505b 180 if (vp->v_mount && (vp->v_mount->mnt_kern_flag & MNTK_NOMSYNC))
2a418930 181 vm_object_set_flag(object, OBJ_NOMSYNC);
984263bc 182 } else {
2de4f77e 183 object->ref_count++; /* protected by vmobj_token */
b0d18f7d
MD
184 if (object->size != lsize) {
185 kprintf("vnode_pager_alloc: Warning, objsize "
186 "mismatch %jd/%jd vp=%p obj=%p\n",
187 (intmax_t)object->size,
188 (intmax_t)lsize,
189 vp, object);
190 }
191 if (vp->v_filesize != length) {
973c11b9 192 kprintf("vnode_pager_alloc: Warning, filesize "
b0d18f7d
MD
193 "mismatch %jd/%jd vp=%p obj=%p\n",
194 (intmax_t)vp->v_filesize,
195 (intmax_t)length,
196 vp, object);
973c11b9 197 }
984263bc
MD
198 }
199
2de4f77e 200 vref(vp);
2247fe02 201 vclrflags(vp, VOLOCK);
984263bc 202 if (vp->v_flag & VOWANT) {
5a648714
MD
203 vclrflags(vp, VOWANT);
204 wakeup(vp);
205 }
2de4f77e
MD
206 lwkt_reltoken(&vmobj_token);
207
5a648714
MD
208 return (object);
209}
210
211/*
212 * Add a ref to a vnode's existing VM object, return the object or
213 * NULL if the vnode did not have one. This does not create the
214 * object (we can't since we don't know what the proper blocksize/boff
215 * is to match the VFS's use of the buffer cache).
216 */
217vm_object_t
218vnode_pager_reference(struct vnode *vp)
219{
220 vm_object_t object;
221
222 /*
2de4f77e
MD
223 * Serialize potential vnode/object teardowns and interlocks
224 */
225 lwkt_gettoken(&vmobj_token);
226
227 /*
5a648714
MD
228 * Prevent race condition when allocating the object. This
229 * can happen with NFS vnodes since the nfsnode isn't locked.
230 */
231 while (vp->v_flag & VOLOCK) {
232 vsetflags(vp, VOWANT);
233 tsleep(vp, 0, "vnpobj", 0);
234 }
235 vsetflags(vp, VOLOCK);
236
237 /*
238 * Prevent race conditions against deallocation of the VM
239 * object.
240 */
241 while (((object = vp->v_object) != NULL) &&
242 (object->flags & OBJ_DEAD)) {
243 vm_object_dead_sleep(object, "vadead");
244 }
245
246 /*
247 * The object is expected to exist, the caller will handle
248 * NULL returns if it does not.
249 */
250 if (object) {
2de4f77e 251 object->ref_count++; /* protected by vmobj_token */
5a648714
MD
252 vref(vp);
253 }
254
255 vclrflags(vp, VOLOCK);
256 if (vp->v_flag & VOWANT) {
2247fe02 257 vclrflags(vp, VOWANT);
984263bc
MD
258 wakeup(vp);
259 }
2de4f77e
MD
260
261 lwkt_reltoken(&vmobj_token);
984263bc
MD
262 return (object);
263}
264
265static void
57e43348 266vnode_pager_dealloc(vm_object_t object)
984263bc 267{
5f910b2f 268 struct vnode *vp = object->handle;
984263bc
MD
269
270 if (vp == NULL)
271 panic("vnode_pager_dealloc: pager already dealloced");
272
273 vm_object_pip_wait(object, "vnpdea");
274
275 object->handle = NULL;
276 object->type = OBJT_DEAD;
277 vp->v_object = NULL;
57f7b636 278 vp->v_filesize = NOOFFSET;
2247fe02 279 vclrflags(vp, VTEXT | VOBJBUF);
8d292090 280 swap_pager_freespace_all(object);
984263bc
MD
281}
282
54078292
MD
283/*
284 * Return whether the vnode pager has the requested page. Return the
285 * number of disk-contiguous pages before and after the requested page,
286 * not including the requested page.
287 */
984263bc 288static boolean_t
1b9d3514 289vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex)
984263bc
MD
290{
291 struct vnode *vp = object->handle;
54078292
MD
292 off_t loffset;
293 off_t doffset;
294 int voff;
984263bc 295 int bsize;
54078292 296 int error;
984263bc
MD
297
298 /*
299 * If no vp or vp is doomed or marked transparent to VM, we do not
300 * have the page.
301 */
5fd012e0 302 if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
984263bc
MD
303 return FALSE;
304
305 /*
306 * If filesystem no longer mounted or offset beyond end of file we do
307 * not have the page.
308 */
54078292
MD
309 loffset = IDX_TO_OFF(pindex);
310
57f7b636 311 if (vp->v_mount == NULL || loffset >= vp->v_filesize)
984263bc
MD
312 return FALSE;
313
314 bsize = vp->v_mount->mnt_stat.f_iosize;
54078292
MD
315 voff = loffset % bsize;
316
bc823b32 317 /*
1b9d3514
MD
318 * XXX
319 *
bc823b32
MD
320 * BMAP returns byte counts before and after, where after
321 * is inclusive of the base page. haspage must return page
322 * counts before and after where after does not include the
323 * base page.
324 *
325 * BMAP is allowed to return a *after of 0 for backwards
326 * compatibility. The base page is still considered valid if
327 * no error is returned.
328 */
1b9d3514
MD
329 error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0);
330 if (error)
984263bc 331 return TRUE;
54078292 332 if (doffset == NOOFFSET)
984263bc 333 return FALSE;
984263bc
MD
334 return TRUE;
335}
336
337/*
338 * Lets the VM system know about a change in size for a file.
339 * We adjust our own internal size and flush any cached pages in
340 * the associated object that are affected by the size change.
341 *
57f7b636 342 * NOTE: This routine may be invoked as a result of a pager put
984263bc 343 * operation (possibly at object termination time), so we must be careful.
57f7b636
MD
344 *
345 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
346 * we do not blow up on the case. nsize will always be >= 0, however.
984263bc
MD
347 */
348void
57e43348 349vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
984263bc
MD
350{
351 vm_pindex_t nobjsize;
c4b46cb4 352 vm_pindex_t oobjsize;
984263bc
MD
353 vm_object_t object = vp->v_object;
354
355 if (object == NULL)
356 return;
357
358 /*
359 * Hasn't changed size
360 */
57f7b636 361 if (nsize == vp->v_filesize)
984263bc
MD
362 return;
363
573fb415
MD
364 lwkt_gettoken(&vm_token);
365
c4b46cb4
MD
366 /*
367 * Has changed size. Adjust the VM object's size and v_filesize
368 * before we start scanning pages to prevent new pages from being
369 * allocated during the scan.
370 */
984263bc 371 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
c4b46cb4
MD
372 oobjsize = object->size;
373 object->size = nobjsize;
984263bc
MD
374
375 /*
376 * File has shrunk. Toss any cached pages beyond the new EOF.
377 */
57f7b636 378 if (nsize < vp->v_filesize) {
135bd6a8 379 vp->v_filesize = nsize;
c4b46cb4
MD
380 if (nobjsize < oobjsize) {
381 vm_object_page_remove(object, nobjsize, oobjsize,
382 FALSE);
984263bc
MD
383 }
384 /*
8d429613
MD
385 * This gets rid of garbage at the end of a page that is now
386 * only partially backed by the vnode. Since we are setting
387 * the entire page valid & clean after we are done we have
388 * to be sure that the portion of the page within the file
389 * bounds is already valid. If it isn't then making it
390 * valid would create a corrupt block.
984263bc
MD
391 */
392 if (nsize & PAGE_MASK) {
393 vm_offset_t kva;
394 vm_page_t m;
395
17cde63e
MD
396 do {
397 m = vm_page_lookup(object, OFF_TO_IDX(nsize));
398 } while (m && vm_page_sleep_busy(m, TRUE, "vsetsz"));
399
984263bc
MD
400 if (m && m->valid) {
401 int base = (int)nsize & PAGE_MASK;
402 int size = PAGE_SIZE - base;
5c5185ae 403 struct lwbuf *lwb;
7a683a24 404 struct lwbuf lwb_cache;
984263bc
MD
405
406 /*
407 * Clear out partial-page garbage in case
408 * the page has been mapped.
c7841cbe
MD
409 *
410 * This is byte aligned.
984263bc 411 */
17cde63e 412 vm_page_busy(m);
7a683a24 413 lwb = lwbuf_alloc(m, &lwb_cache);
5c5185ae 414 kva = lwbuf_kva(lwb);
984263bc 415 bzero((caddr_t)kva + base, size);
5c5185ae 416 lwbuf_free(lwb);
984263bc
MD
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_protect(m, VM_PROT_NONE);
431
432 /*
433 * Clear out partial-page dirty bits. This
434 * has the side effect of setting the valid
435 * bits, but that is ok. There are a bunch
436 * of places in the VM system where we expected
437 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
438 * case is one of them. If the page is still
439 * partially dirty, make it fully dirty.
440 *
c7841cbe 441 * NOTE: We do not clear out the valid
984263bc
MD
442 * bits. This would prevent bogus_page
443 * replacement from working properly.
c7841cbe
MD
444 *
445 * NOTE: We do not want to clear the dirty
446 * bit for a partial DEV_BSIZE'd truncation!
447 * This is DEV_BSIZE aligned!
984263bc 448 */
1a54183b 449 vm_page_clear_dirty_beg_nonincl(m, base, size);
984263bc
MD
450 if (m->dirty != 0)
451 m->dirty = VM_PAGE_BITS_ALL;
17cde63e 452 vm_page_wakeup(m);
984263bc
MD
453 }
454 }
135bd6a8
MD
455 } else {
456 vp->v_filesize = nsize;
984263bc 457 }
573fb415 458 lwkt_reltoken(&vm_token);
984263bc
MD
459}
460
a55afca2
MD
461/*
462 * Release a page busied for a getpages operation. The page may have become
463 * wired (typically due to being used by the buffer cache) or otherwise been
464 * soft-busied and cannot be freed in that case. A held page can still be
465 * freed.
466 */
984263bc 467void
57e43348 468vnode_pager_freepage(vm_page_t m)
984263bc 469{
a55afca2
MD
470 if (m->busy || m->wire_count) {
471 vm_page_activate(m);
472 vm_page_wakeup(m);
473 } else {
474 vm_page_free(m);
475 }
984263bc
MD
476}
477
478/*
984263bc
MD
479 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
480 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
481 * vnode_pager_generic_getpages() to implement the previous behaviour.
482 *
483 * All other FS's should use the bypass to get to the local media
484 * backing vp's VOP_GETPAGES.
485 */
486static int
1b9d3514 487vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
984263bc
MD
488{
489 int rtval;
490 struct vnode *vp;
984263bc
MD
491
492 vp = object->handle;
1c48c952 493 rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess);
a05863a7
MD
494 if (rtval == EOPNOTSUPP)
495 panic("vnode_pager: vfs's must implement vop_getpages\n");
984263bc
MD
496 return rtval;
497}
498
984263bc
MD
499/*
500 * This is now called from local media FS's to operate against their
501 * own vnodes if they fail to implement VOP_GETPAGES.
a05863a7
MD
502 *
503 * With all the caching local media devices do these days there is really
504 * very little point to attempting to restrict the I/O size to contiguous
505 * blocks on-disk, especially if our caller thinks we need all the specified
506 * pages. Just construct and issue a READ.
984263bc
MD
507 */
508int
1b9d3514 509vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount,
1c48c952 510 int reqpage, int seqaccess)
984263bc 511{
a05863a7
MD
512 struct iovec aiov;
513 struct uio auio;
514 off_t foff;
515 int error;
984263bc 516 int count;
a05863a7
MD
517 int i;
518 int ioflags;
984263bc
MD
519
520 /*
a05863a7 521 * Do not do anything if the vnode is bad.
984263bc 522 */
a05863a7
MD
523 if (vp->v_mount == NULL)
524 return VM_PAGER_BAD;
984263bc
MD
525
526 /*
a05863a7
MD
527 * Calculate the number of pages. Since we are paging in whole
528 * pages, adjust bytecount to be an integral multiple of the page
529 * size. It will be clipped to the file EOF later on.
984263bc 530 */
a05863a7
MD
531 bytecount = round_page(bytecount);
532 count = bytecount / PAGE_SIZE;
984263bc
MD
533
534 /*
cf1bb2a8
MD
535 * We could check m[reqpage]->valid here and shortcut the operation,
536 * but doing so breaks read-ahead. Instead assume that the VM
537 * system has already done at least the check, don't worry about
538 * any races, and issue the VOP_READ to allow read-ahead to function.
8d429613 539 *
cf1bb2a8
MD
540 * This keeps the pipeline full for I/O bound sequentially scanned
541 * mmap()'s
984263bc 542 */
cf1bb2a8 543 /* don't shortcut */
984263bc
MD
544
545 /*
a05863a7
MD
546 * Discard pages past the file EOF. If the requested page is past
547 * the file EOF we just leave its valid bits set to 0, the caller
548 * expects to maintain ownership of the requested page. If the
549 * entire range is past file EOF discard everything and generate
550 * a pagein error.
984263bc 551 */
1b9d3514 552 foff = IDX_TO_OFF(mpp[0]->pindex);
a05863a7
MD
553 if (foff >= vp->v_filesize) {
554 for (i = 0; i < count; i++) {
555 if (i != reqpage)
1b9d3514 556 vnode_pager_freepage(mpp[i]);
984263bc 557 }
a05863a7 558 return VM_PAGER_ERROR;
984263bc
MD
559 }
560
a05863a7
MD
561 if (foff + bytecount > vp->v_filesize) {
562 bytecount = vp->v_filesize - foff;
563 i = round_page(bytecount) / PAGE_SIZE;
564 while (count > i) {
565 --count;
566 if (count != reqpage)
1b9d3514 567 vnode_pager_freepage(mpp[count]);
984263bc 568 }
984263bc
MD
569 }
570
571 /*
a05863a7
MD
572 * The size of the transfer is bytecount. bytecount will be an
573 * integral multiple of the page size unless it has been clipped
574 * to the file EOF. The transfer cannot exceed the file EOF.
575 *
576 * When dealing with real devices we must round-up to the device
577 * sector size.
984263bc 578 */
a05863a7
MD
579 if (vp->v_type == VBLK || vp->v_type == VCHR) {
580 int secmask = vp->v_rdev->si_bsize_phys - 1;
984263bc 581 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
a05863a7 582 bytecount = (bytecount + secmask) & ~secmask;
984263bc
MD
583 }
584
984263bc 585 /*
ca28958c
MD
586 * Severe hack to avoid deadlocks with the buffer cache
587 */
588 for (i = 0; i < count; ++i) {
1b9d3514 589 vm_page_t mt = mpp[i];
ca28958c
MD
590
591 vm_page_io_start(mt);
592 vm_page_wakeup(mt);
593 }
594
595 /*
cf1bb2a8 596 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE
984263bc 597 */
a05863a7 598 ioflags = IO_VMIO;
1c48c952 599 if (seqaccess)
cf1bb2a8 600 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
a05863a7 601
1b9d3514 602 aiov.iov_base = NULL;
a05863a7
MD
603 aiov.iov_len = bytecount;
604 auio.uio_iov = &aiov;
605 auio.uio_iovcnt = 1;
606 auio.uio_offset = foff;
607 auio.uio_segflg = UIO_NOCOPY;
608 auio.uio_rw = UIO_READ;
609 auio.uio_resid = bytecount;
610 auio.uio_td = NULL;
12e4aaff
MD
611 mycpu->gd_cnt.v_vnodein++;
612 mycpu->gd_cnt.v_vnodepgsin += count;
984263bc 613
a05863a7 614 error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
984263bc
MD
615
616 /*
ca28958c
MD
617 * Severe hack to avoid deadlocks with the buffer cache
618 */
573fb415 619 lwkt_gettoken(&vm_token);
ca28958c 620 for (i = 0; i < count; ++i) {
1b9d3514 621 vm_page_t mt = mpp[i];
ca28958c
MD
622
623 while (vm_page_sleep_busy(mt, FALSE, "getpgs"))
624 ;
625 vm_page_busy(mt);
626 vm_page_io_finish(mt);
627 }
573fb415 628 lwkt_reltoken(&vm_token);
ca28958c
MD
629
630 /*
a05863a7
MD
631 * Calculate the actual number of bytes read and clean up the
632 * page list.
984263bc 633 */
a05863a7 634 bytecount -= auio.uio_resid;
984263bc 635
a05863a7 636 for (i = 0; i < count; ++i) {
1b9d3514 637 vm_page_t mt = mpp[i];
a55afca2 638
a05863a7 639 if (i != reqpage) {
a55afca2 640 if (error == 0 && mt->valid) {
984263bc
MD
641 if (mt->flags & PG_WANTED)
642 vm_page_activate(mt);
643 else
644 vm_page_deactivate(mt);
645 vm_page_wakeup(mt);
646 } else {
647 vnode_pager_freepage(mt);
648 }
a55afca2
MD
649 } else if (mt->valid == 0) {
650 if (error == 0) {
651 kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt, mt->object, (int) mt->pindex);
652 /* whoops, something happened */
653 error = EINVAL;
654 }
655 } else if (mt->valid != VM_PAGE_BITS_ALL) {
656 /*
657 * Zero-extend the requested page if necessary (if
658 * the filesystem is using a small block size).
659 */
660 vm_page_zero_invalid(mt, TRUE);
984263bc
MD
661 }
662 }
663 if (error) {
1b9d3514 664 kprintf("vnode_pager_getpage: I/O read error\n");
984263bc
MD
665 }
666 return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
667}
668
669/*
670 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
671 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
672 * vnode_pager_generic_putpages() to implement the previous behaviour.
673 *
cb1cf930
MD
674 * Caller has already cleared the pmap modified bits, if any.
675 *
984263bc
MD
676 * All other FS's should use the bypass to get to the local media
677 * backing vp's VOP_PUTPAGES.
678 */
679static void
57e43348 680vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
cb1cf930 681 boolean_t sync, int *rtvals)
984263bc
MD
682{
683 int rtval;
684 struct vnode *vp;
685 int bytes = count * PAGE_SIZE;
686
687 /*
688 * Force synchronous operation if we are extremely low on memory
689 * to prevent a low-memory deadlock. VOP operations often need to
690 * allocate more memory to initiate the I/O ( i.e. do a BMAP
691 * operation ). The swapper handles the case by limiting the amount
692 * of asynchronous I/O, but that sort of solution doesn't scale well
693 * for the vnode pager without a lot of work.
694 *
695 * Also, the backing vnode's iodone routine may not wake the pageout
696 * daemon up. This should be probably be addressed XXX.
697 */
698
12e4aaff 699 if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min)
984263bc
MD
700 sync |= OBJPC_SYNC;
701
702 /*
703 * Call device-specific putpages function
704 */
984263bc
MD
705 vp = object->handle;
706 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
707 if (rtval == EOPNOTSUPP) {
086c1d7e 708 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
984263bc
MD
709 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
710 }
711}
712
713
714/*
715 * This is now called from local media FS's to operate against their
716 * own vnodes if they fail to implement VOP_PUTPAGES.
717 *
718 * This is typically called indirectly via the pageout daemon and
719 * clustering has already typically occured, so in general we ask the
720 * underlying filesystem to write the data out asynchronously rather
721 * then delayed.
722 */
723int
57e43348 724vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
a05863a7 725 int flags, int *rtvals)
984263bc
MD
726{
727 int i;
728 vm_object_t object;
cb1cf930 729 int maxsize, ncount, count;
984263bc
MD
730 vm_ooffset_t poffset;
731 struct uio auio;
732 struct iovec aiov;
733 int error;
734 int ioflags;
735
736 object = vp->v_object;
737 count = bytecount / PAGE_SIZE;
738
739 for (i = 0; i < count; i++)
740 rtvals[i] = VM_PAGER_AGAIN;
741
742 if ((int) m[0]->pindex < 0) {
086c1d7e 743 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
984263bc
MD
744 (long)m[0]->pindex, m[0]->dirty);
745 rtvals[0] = VM_PAGER_BAD;
746 return VM_PAGER_BAD;
747 }
748
749 maxsize = count * PAGE_SIZE;
750 ncount = count;
751
752 poffset = IDX_TO_OFF(m[0]->pindex);
753
754 /*
755 * If the page-aligned write is larger then the actual file we
cb1cf930
MD
756 * have to invalidate pages occuring beyond the file EOF.
757 *
758 * If the file EOF resides in the middle of a page we still clear
759 * all of that page's dirty bits later on. If we didn't it would
760 * endlessly re-write.
984263bc
MD
761 *
762 * We do not under any circumstances truncate the valid bits, as
763 * this will screw up bogus page replacement.
93afe6be
MD
764 *
765 * The caller has already read-protected the pages. The VFS must
766 * use the buffer cache to wrap the pages. The pages might not
767 * be immediately flushed by the buffer cache but once under its
768 * control the pages themselves can wind up being marked clean
769 * and their covering buffer cache buffer can be marked dirty.
984263bc 770 */
fb69d1c8
MD
771 if (poffset + maxsize > vp->v_filesize) {
772 if (poffset < vp->v_filesize) {
57f7b636 773 maxsize = vp->v_filesize - poffset;
984263bc 774 ncount = btoc(maxsize);
984263bc
MD
775 } else {
776 maxsize = 0;
777 ncount = 0;
778 }
779 if (ncount < count) {
780 for (i = ncount; i < count; i++) {
781 rtvals[i] = VM_PAGER_BAD;
782 }
783 }
784 }
785
786 /*
787 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
788 * rather then a bdwrite() to prevent paging I/O from saturating
789 * the buffer cache. Dummy-up the sequential heuristic to cause
790 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
791 * the system decides how to cluster.
792 */
793 ioflags = IO_VMIO;
794 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
795 ioflags |= IO_SYNC;
796 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
797 ioflags |= IO_ASYNC;
798 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
799 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
800
801 aiov.iov_base = (caddr_t) 0;
802 aiov.iov_len = maxsize;
803 auio.uio_iov = &aiov;
804 auio.uio_iovcnt = 1;
805 auio.uio_offset = poffset;
806 auio.uio_segflg = UIO_NOCOPY;
807 auio.uio_rw = UIO_WRITE;
808 auio.uio_resid = maxsize;
dadab5e9 809 auio.uio_td = NULL;
8a8d5d85 810 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
12e4aaff
MD
811 mycpu->gd_cnt.v_vnodeout++;
812 mycpu->gd_cnt.v_vnodepgsout += ncount;
984263bc
MD
813
814 if (error) {
35f59bfa
MD
815 krateprintf(&vbadrate,
816 "vnode_pager_putpages: I/O error %d\n", error);
984263bc
MD
817 }
818 if (auio.uio_resid) {
35f59bfa 819 krateprintf(&vresrate,
bfc09ba0 820 "vnode_pager_putpages: residual I/O %zd at %lu\n",
35f59bfa 821 auio.uio_resid, (u_long)m[0]->pindex);
984263bc 822 }
cb1cf930
MD
823 if (error == 0) {
824 for (i = 0; i < ncount; i++) {
825 rtvals[i] = VM_PAGER_OK;
826 vm_page_undirty(m[i]);
827 }
828 }
984263bc
MD
829 return rtvals[0];
830}
831
832struct vnode *
dadab5e9 833vnode_pager_lock(vm_object_t object)
984263bc 834{
dadab5e9 835 struct thread *td = curthread; /* XXX */
5fd012e0 836 int error;
984263bc 837
0ac0a48a
MD
838 ASSERT_LWKT_TOKEN_HELD(&vm_token);
839
984263bc
MD
840 for (; object != NULL; object = object->backing_object) {
841 if (object->type != OBJT_VNODE)
842 continue;
843 if (object->flags & OBJ_DEAD)
844 return NULL;
845
5fd012e0
MD
846 for (;;) {
847 struct vnode *vp = object->handle;
ab6f251b 848 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
5fd012e0
MD
849 if (error == 0) {
850 if (object->handle != vp) {
851 vput(vp);
852 continue;
853 }
854 return (vp);
855 }
dadab5e9
MD
856 if ((object->flags & OBJ_DEAD) ||
857 (object->type != OBJT_VNODE)) {
984263bc 858 return NULL;
dadab5e9 859 }
086c1d7e 860 kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp, error, lockstatus(&vp->v_lock, td));
5fd012e0 861 tsleep(object->handle, 0, "vnpgrl", hz);
984263bc 862 }
984263bc
MD
863 }
864 return NULL;
865}