2 * Copyright (c) 1990 University of Utah.
3 * Copyright (c) 1991 The Regents of the University of California.
5 * Copyright (c) 1993, 1994 John S. Dyson
6 * Copyright (c) 1995, David Greenman
8 * This code is derived from software contributed to Berkeley by
9 * the Systems Programming Group of the University of Utah Computer
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
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.
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
40 * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91
41 * $FreeBSD: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $
42 * $DragonFly: src/sys/vm/vnode_pager.c,v 1.43 2008/06/19 23:27:39 dillon Exp $
46 * Page to/from files (vnodes).
51 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
52 * greatly re-simplify the vnode_pager.
55 #include <sys/param.h>
56 #include <sys/systm.h>
57 #include <sys/kernel.h>
59 #include <sys/vnode.h>
60 #include <sys/mount.h>
62 #include <sys/vmmeter.h>
64 #include <sys/sfbuf.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_pager.h>
70 #include <vm/vm_map.h>
71 #include <vm/vnode_pager.h>
72 #include <vm/vm_extern.h>
74 #include <sys/thread2.h>
75 #include <vm/vm_page2.h>
77 static void vnode_pager_dealloc (vm_object_t);
78 static int vnode_pager_getpage (vm_object_t, vm_page_t *, int);
79 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *);
80 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t);
82 struct pagerops vnodepagerops = {
92 static struct krate vbadrate = { 1 };
93 static struct krate vresrate = { 1 };
95 int vnode_pbuf_freecnt = -1; /* start out unlimited */
98 * Allocate (or lookup) pager for a vnode.
99 * Handle is a vnode pointer.
102 vnode_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
108 * Pageout to vnode, no can do yet.
114 * XXX hack - This initialization should be put somewhere else.
116 if (vnode_pbuf_freecnt < 0) {
117 vnode_pbuf_freecnt = nswbuf / 2 + 1;
120 vp = (struct vnode *) handle;
123 * Prevent race condition when allocating the object. This
124 * can happen with NFS vnodes since the nfsnode isn't locked.
126 while (vp->v_flag & VOLOCK) {
127 vsetflags(vp, VOWANT);
128 tsleep(vp, 0, "vnpobj", 0);
130 vsetflags(vp, VOLOCK);
133 * If the object is being terminated, wait for it to
136 while (((object = vp->v_object) != NULL) &&
137 (object->flags & OBJ_DEAD)) {
138 vm_object_dead_sleep(object, "vadead");
141 if (vp->v_sysref.refcnt <= 0)
142 panic("vnode_pager_alloc: no vnode reference");
144 if (object == NULL) {
146 * And an object of the appropriate size
148 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
150 object->handle = handle;
151 vp->v_object = object;
152 vp->v_filesize = size;
155 if (vp->v_filesize != size) {
156 kprintf("vnode_pager_alloc: Warning, filesize "
157 "mismatch %lld/%lld\n",
158 (long long)vp->v_filesize,
164 vclrflags(vp, VOLOCK);
165 if (vp->v_flag & VOWANT) {
166 vclrflags(vp, VOWANT);
173 vnode_pager_dealloc(vm_object_t object)
175 struct vnode *vp = object->handle;
178 panic("vnode_pager_dealloc: pager already dealloced");
180 vm_object_pip_wait(object, "vnpdea");
182 object->handle = NULL;
183 object->type = OBJT_DEAD;
185 vp->v_filesize = NOOFFSET;
186 vclrflags(vp, VTEXT | VOBJBUF);
190 * Return whether the vnode pager has the requested page. Return the
191 * number of disk-contiguous pages before and after the requested page,
192 * not including the requested page.
195 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex)
197 struct vnode *vp = object->handle;
205 * If no vp or vp is doomed or marked transparent to VM, we do not
208 if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
212 * If filesystem no longer mounted or offset beyond end of file we do
215 loffset = IDX_TO_OFF(pindex);
217 if (vp->v_mount == NULL || loffset >= vp->v_filesize)
220 bsize = vp->v_mount->mnt_stat.f_iosize;
221 voff = loffset % bsize;
226 * BMAP returns byte counts before and after, where after
227 * is inclusive of the base page. haspage must return page
228 * counts before and after where after does not include the
231 * BMAP is allowed to return a *after of 0 for backwards
232 * compatibility. The base page is still considered valid if
233 * no error is returned.
235 error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0);
238 if (doffset == NOOFFSET)
244 * Lets the VM system know about a change in size for a file.
245 * We adjust our own internal size and flush any cached pages in
246 * the associated object that are affected by the size change.
248 * NOTE: This routine may be invoked as a result of a pager put
249 * operation (possibly at object termination time), so we must be careful.
251 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
252 * we do not blow up on the case. nsize will always be >= 0, however.
255 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
257 vm_pindex_t nobjsize;
258 vm_pindex_t oobjsize;
259 vm_object_t object = vp->v_object;
265 * Hasn't changed size
267 if (nsize == vp->v_filesize)
271 * Has changed size. Adjust the VM object's size and v_filesize
272 * before we start scanning pages to prevent new pages from being
273 * allocated during the scan.
275 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
276 oobjsize = object->size;
277 object->size = nobjsize;
280 * File has shrunk. Toss any cached pages beyond the new EOF.
282 if (nsize < vp->v_filesize) {
283 vp->v_filesize = nsize;
284 if (nobjsize < oobjsize) {
285 vm_object_page_remove(object, nobjsize, oobjsize,
289 * This gets rid of garbage at the end of a page that is now
290 * only partially backed by the vnode. Since we are setting
291 * the entire page valid & clean after we are done we have
292 * to be sure that the portion of the page within the file
293 * bounds is already valid. If it isn't then making it
294 * valid would create a corrupt block.
296 if (nsize & PAGE_MASK) {
301 m = vm_page_lookup(object, OFF_TO_IDX(nsize));
302 } while (m && vm_page_sleep_busy(m, TRUE, "vsetsz"));
305 int base = (int)nsize & PAGE_MASK;
306 int size = PAGE_SIZE - base;
310 * Clear out partial-page garbage in case
311 * the page has been mapped.
313 * This is byte aligned.
316 sf = sf_buf_alloc(m, SFB_CPUPRIVATE);
317 kva = sf_buf_kva(sf);
318 bzero((caddr_t)kva + base, size);
322 * XXX work around SMP data integrity race
323 * by unmapping the page from user processes.
324 * The garbage we just cleared may be mapped
325 * to a user process running on another cpu
326 * and this code is not running through normal
327 * I/O channels which handle SMP issues for
328 * us, so unmap page to synchronize all cpus.
330 * XXX should vm_pager_unmap_page() have
333 vm_page_protect(m, VM_PROT_NONE);
336 * Clear out partial-page dirty bits. This
337 * has the side effect of setting the valid
338 * bits, but that is ok. There are a bunch
339 * of places in the VM system where we expected
340 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
341 * case is one of them. If the page is still
342 * partially dirty, make it fully dirty.
344 * NOTE: We do not clear out the valid
345 * bits. This would prevent bogus_page
346 * replacement from working properly.
348 * NOTE: We do not want to clear the dirty
349 * bit for a partial DEV_BSIZE'd truncation!
350 * This is DEV_BSIZE aligned!
352 vm_page_clear_dirty_beg_nonincl(m, base, size);
354 m->dirty = VM_PAGE_BITS_ALL;
359 vp->v_filesize = nsize;
364 * Release a page busied for a getpages operation. The page may have become
365 * wired (typically due to being used by the buffer cache) or otherwise been
366 * soft-busied and cannot be freed in that case. A held page can still be
370 vnode_pager_freepage(vm_page_t m)
372 if (m->busy || m->wire_count) {
381 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
382 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
383 * vnode_pager_generic_getpages() to implement the previous behaviour.
385 * All other FS's should use the bypass to get to the local media
386 * backing vp's VOP_GETPAGES.
389 vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
395 rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess);
396 if (rtval == EOPNOTSUPP)
397 panic("vnode_pager: vfs's must implement vop_getpages\n");
402 * This is now called from local media FS's to operate against their
403 * own vnodes if they fail to implement VOP_GETPAGES.
405 * With all the caching local media devices do these days there is really
406 * very little point to attempting to restrict the I/O size to contiguous
407 * blocks on-disk, especially if our caller thinks we need all the specified
408 * pages. Just construct and issue a READ.
411 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount,
412 int reqpage, int seqaccess)
423 * Do not do anything if the vnode is bad.
425 if (vp->v_mount == NULL)
429 * Calculate the number of pages. Since we are paging in whole
430 * pages, adjust bytecount to be an integral multiple of the page
431 * size. It will be clipped to the file EOF later on.
433 bytecount = round_page(bytecount);
434 count = bytecount / PAGE_SIZE;
437 * We could check m[reqpage]->valid here and shortcut the operation,
438 * but doing so breaks read-ahead. Instead assume that the VM
439 * system has already done at least the check, don't worry about
440 * any races, and issue the VOP_READ to allow read-ahead to function.
442 * This keeps the pipeline full for I/O bound sequentially scanned
448 * Discard pages past the file EOF. If the requested page is past
449 * the file EOF we just leave its valid bits set to 0, the caller
450 * expects to maintain ownership of the requested page. If the
451 * entire range is past file EOF discard everything and generate
454 foff = IDX_TO_OFF(mpp[0]->pindex);
455 if (foff >= vp->v_filesize) {
456 for (i = 0; i < count; i++) {
458 vnode_pager_freepage(mpp[i]);
460 return VM_PAGER_ERROR;
463 if (foff + bytecount > vp->v_filesize) {
464 bytecount = vp->v_filesize - foff;
465 i = round_page(bytecount) / PAGE_SIZE;
468 if (count != reqpage)
469 vnode_pager_freepage(mpp[count]);
474 * The size of the transfer is bytecount. bytecount will be an
475 * integral multiple of the page size unless it has been clipped
476 * to the file EOF. The transfer cannot exceed the file EOF.
478 * When dealing with real devices we must round-up to the device
481 if (vp->v_type == VBLK || vp->v_type == VCHR) {
482 int secmask = vp->v_rdev->si_bsize_phys - 1;
483 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
484 bytecount = (bytecount + secmask) & ~secmask;
488 * Severe hack to avoid deadlocks with the buffer cache
490 for (i = 0; i < count; ++i) {
491 vm_page_t mt = mpp[i];
493 vm_page_io_start(mt);
498 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE
502 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
504 aiov.iov_base = NULL;
505 aiov.iov_len = bytecount;
506 auio.uio_iov = &aiov;
508 auio.uio_offset = foff;
509 auio.uio_segflg = UIO_NOCOPY;
510 auio.uio_rw = UIO_READ;
511 auio.uio_resid = bytecount;
513 mycpu->gd_cnt.v_vnodein++;
514 mycpu->gd_cnt.v_vnodepgsin += count;
516 error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
519 * Severe hack to avoid deadlocks with the buffer cache
521 for (i = 0; i < count; ++i) {
522 vm_page_t mt = mpp[i];
524 while (vm_page_sleep_busy(mt, FALSE, "getpgs"))
527 vm_page_io_finish(mt);
531 * Calculate the actual number of bytes read and clean up the
534 bytecount -= auio.uio_resid;
536 for (i = 0; i < count; ++i) {
537 vm_page_t mt = mpp[i];
540 if (error == 0 && mt->valid) {
541 if (mt->flags & PG_WANTED)
542 vm_page_activate(mt);
544 vm_page_deactivate(mt);
547 vnode_pager_freepage(mt);
549 } else if (mt->valid == 0) {
551 kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt, mt->object, (int) mt->pindex);
552 /* whoops, something happened */
555 } else if (mt->valid != VM_PAGE_BITS_ALL) {
557 * Zero-extend the requested page if necessary (if
558 * the filesystem is using a small block size).
560 vm_page_zero_invalid(mt, TRUE);
564 kprintf("vnode_pager_getpage: I/O read error\n");
566 return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
570 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
571 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
572 * vnode_pager_generic_putpages() to implement the previous behaviour.
574 * Caller has already cleared the pmap modified bits, if any.
576 * All other FS's should use the bypass to get to the local media
577 * backing vp's VOP_PUTPAGES.
580 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
581 boolean_t sync, int *rtvals)
585 int bytes = count * PAGE_SIZE;
588 * Force synchronous operation if we are extremely low on memory
589 * to prevent a low-memory deadlock. VOP operations often need to
590 * allocate more memory to initiate the I/O ( i.e. do a BMAP
591 * operation ). The swapper handles the case by limiting the amount
592 * of asynchronous I/O, but that sort of solution doesn't scale well
593 * for the vnode pager without a lot of work.
595 * Also, the backing vnode's iodone routine may not wake the pageout
596 * daemon up. This should be probably be addressed XXX.
599 if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min)
603 * Call device-specific putpages function
606 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
607 if (rtval == EOPNOTSUPP) {
608 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
609 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
615 * This is now called from local media FS's to operate against their
616 * own vnodes if they fail to implement VOP_PUTPAGES.
618 * This is typically called indirectly via the pageout daemon and
619 * clustering has already typically occured, so in general we ask the
620 * underlying filesystem to write the data out asynchronously rather
624 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
625 int flags, int *rtvals)
629 int maxsize, ncount, count;
630 vm_ooffset_t poffset;
636 object = vp->v_object;
637 count = bytecount / PAGE_SIZE;
639 for (i = 0; i < count; i++)
640 rtvals[i] = VM_PAGER_AGAIN;
642 if ((int) m[0]->pindex < 0) {
643 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
644 (long)m[0]->pindex, m[0]->dirty);
645 rtvals[0] = VM_PAGER_BAD;
649 maxsize = count * PAGE_SIZE;
652 poffset = IDX_TO_OFF(m[0]->pindex);
655 * If the page-aligned write is larger then the actual file we
656 * have to invalidate pages occuring beyond the file EOF.
658 * If the file EOF resides in the middle of a page we still clear
659 * all of that page's dirty bits later on. If we didn't it would
660 * endlessly re-write.
662 * We do not under any circumstances truncate the valid bits, as
663 * this will screw up bogus page replacement.
665 * The caller has already read-protected the pages. The VFS must
666 * use the buffer cache to wrap the pages. The pages might not
667 * be immediately flushed by the buffer cache but once under its
668 * control the pages themselves can wind up being marked clean
669 * and their covering buffer cache buffer can be marked dirty.
671 if (poffset + maxsize > vp->v_filesize) {
672 if (poffset < vp->v_filesize) {
673 maxsize = vp->v_filesize - poffset;
674 ncount = btoc(maxsize);
679 if (ncount < count) {
680 for (i = ncount; i < count; i++) {
681 rtvals[i] = VM_PAGER_BAD;
687 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
688 * rather then a bdwrite() to prevent paging I/O from saturating
689 * the buffer cache. Dummy-up the sequential heuristic to cause
690 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
691 * the system decides how to cluster.
694 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
696 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
698 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
699 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
701 aiov.iov_base = (caddr_t) 0;
702 aiov.iov_len = maxsize;
703 auio.uio_iov = &aiov;
705 auio.uio_offset = poffset;
706 auio.uio_segflg = UIO_NOCOPY;
707 auio.uio_rw = UIO_WRITE;
708 auio.uio_resid = maxsize;
710 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
711 mycpu->gd_cnt.v_vnodeout++;
712 mycpu->gd_cnt.v_vnodepgsout += ncount;
715 krateprintf(&vbadrate,
716 "vnode_pager_putpages: I/O error %d\n", error);
718 if (auio.uio_resid) {
719 krateprintf(&vresrate,
720 "vnode_pager_putpages: residual I/O %zd at %lu\n",
721 auio.uio_resid, (u_long)m[0]->pindex);
724 for (i = 0; i < ncount; i++) {
725 rtvals[i] = VM_PAGER_OK;
726 vm_page_undirty(m[i]);
733 vnode_pager_lock(vm_object_t object)
735 struct thread *td = curthread; /* XXX */
738 for (; object != NULL; object = object->backing_object) {
739 if (object->type != OBJT_VNODE)
741 if (object->flags & OBJ_DEAD)
745 struct vnode *vp = object->handle;
746 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
748 if (object->handle != vp) {
754 if ((object->flags & OBJ_DEAD) ||
755 (object->type != OBJT_VNODE)) {
758 kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp, error, lockstatus(&vp->v_lock, td));
759 tsleep(object->handle, 0, "vnpgrl", hz);