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>
65 #include <sys/thread2.h>
68 #include <vm/vm_object.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_pager.h>
71 #include <vm/vm_map.h>
72 #include <vm/vnode_pager.h>
73 #include <vm/vm_extern.h>
75 static void vnode_pager_dealloc (vm_object_t);
76 static int vnode_pager_getpages (vm_object_t, vm_page_t *, int, int);
77 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *);
78 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t, int *, int *);
80 struct pagerops vnodepagerops = {
90 static struct krate vbadrate = { 1 };
91 static struct krate vresrate = { 1 };
93 int vnode_pbuf_freecnt = -1; /* start out unlimited */
96 * Allocate (or lookup) pager for a vnode.
97 * Handle is a vnode pointer.
100 vnode_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
106 * Pageout to vnode, no can do yet.
112 * XXX hack - This initialization should be put somewhere else.
114 if (vnode_pbuf_freecnt < 0) {
115 vnode_pbuf_freecnt = nswbuf / 2 + 1;
118 vp = (struct vnode *) handle;
121 * Prevent race condition when allocating the object. This
122 * can happen with NFS vnodes since the nfsnode isn't locked.
124 while (vp->v_flag & VOLOCK) {
125 vp->v_flag |= VOWANT;
126 tsleep(vp, 0, "vnpobj", 0);
128 vp->v_flag |= VOLOCK;
131 * If the object is being terminated, wait for it to
134 while (((object = vp->v_object) != NULL) &&
135 (object->flags & OBJ_DEAD)) {
136 vm_object_dead_sleep(object, "vadead");
139 if (vp->v_sysref.refcnt <= 0)
140 panic("vnode_pager_alloc: no vnode reference");
142 if (object == NULL) {
144 * And an object of the appropriate size
146 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
148 object->handle = handle;
149 vp->v_object = object;
150 vp->v_filesize = size;
153 if (vp->v_filesize != size)
154 kprintf("vnode_pager_alloc: Warning, filesize mismatch %lld/%lld\n", vp->v_filesize, size);
158 vp->v_flag &= ~VOLOCK;
159 if (vp->v_flag & VOWANT) {
160 vp->v_flag &= ~VOWANT;
167 vnode_pager_dealloc(vm_object_t object)
169 struct vnode *vp = object->handle;
172 panic("vnode_pager_dealloc: pager already dealloced");
174 vm_object_pip_wait(object, "vnpdea");
176 object->handle = NULL;
177 object->type = OBJT_DEAD;
179 vp->v_filesize = NOOFFSET;
180 vp->v_flag &= ~(VTEXT | VOBJBUF);
184 * Return whether the vnode pager has the requested page. Return the
185 * number of disk-contiguous pages before and after the requested page,
186 * not including the requested page.
189 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
192 struct vnode *vp = object->handle;
200 * If no vp or vp is doomed or marked transparent to VM, we do not
203 if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
207 * If filesystem no longer mounted or offset beyond end of file we do
210 loffset = IDX_TO_OFF(pindex);
212 if (vp->v_mount == NULL || loffset >= vp->v_filesize)
215 bsize = vp->v_mount->mnt_stat.f_iosize;
216 voff = loffset % bsize;
219 * BMAP returns byte counts before and after, where after
220 * is inclusive of the base page. haspage must return page
221 * counts before and after where after does not include the
224 * BMAP is allowed to return a *after of 0 for backwards
225 * compatibility. The base page is still considered valid if
226 * no error is returned.
228 error = VOP_BMAP(vp, loffset - voff, &doffset, after, before, 0);
236 if (doffset == NOOFFSET)
240 *before = (*before + voff) >> PAGE_SHIFT;
244 if (loffset + *after > vp->v_filesize)
245 *after = vp->v_filesize - loffset;
246 *after >>= PAGE_SHIFT;
254 * Lets the VM system know about a change in size for a file.
255 * We adjust our own internal size and flush any cached pages in
256 * the associated object that are affected by the size change.
258 * NOTE: This routine may be invoked as a result of a pager put
259 * operation (possibly at object termination time), so we must be careful.
261 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
262 * we do not blow up on the case. nsize will always be >= 0, however.
265 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
267 vm_pindex_t nobjsize;
268 vm_pindex_t oobjsize;
269 vm_object_t object = vp->v_object;
275 * Hasn't changed size
277 if (nsize == vp->v_filesize)
281 * Has changed size. Adjust the VM object's size and v_filesize
282 * before we start scanning pages to prevent new pages from being
283 * allocated during the scan.
285 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
286 oobjsize = object->size;
287 object->size = nobjsize;
290 * File has shrunk. Toss any cached pages beyond the new EOF.
292 if (nsize < vp->v_filesize) {
293 vp->v_filesize = nsize;
294 if (nobjsize < oobjsize) {
295 vm_object_page_remove(object, nobjsize, oobjsize,
299 * This gets rid of garbage at the end of a page that is now
300 * only partially backed by the vnode. Since we are setting
301 * the entire page valid & clean after we are done we have
302 * to be sure that the portion of the page within the file
303 * bounds is already valid. If it isn't then making it
304 * valid would create a corrupt block.
306 if (nsize & PAGE_MASK) {
311 m = vm_page_lookup(object, OFF_TO_IDX(nsize));
312 } while (m && vm_page_sleep_busy(m, TRUE, "vsetsz"));
315 int base = (int)nsize & PAGE_MASK;
316 int size = PAGE_SIZE - base;
320 * Clear out partial-page garbage in case
321 * the page has been mapped.
324 sf = sf_buf_alloc(m, SFB_CPUPRIVATE);
325 kva = sf_buf_kva(sf);
326 bzero((caddr_t)kva + base, size);
330 * XXX work around SMP data integrity race
331 * by unmapping the page from user processes.
332 * The garbage we just cleared may be mapped
333 * to a user process running on another cpu
334 * and this code is not running through normal
335 * I/O channels which handle SMP issues for
336 * us, so unmap page to synchronize all cpus.
338 * XXX should vm_pager_unmap_page() have
341 vm_page_protect(m, VM_PROT_NONE);
344 * Clear out partial-page dirty bits. This
345 * has the side effect of setting the valid
346 * bits, but that is ok. There are a bunch
347 * of places in the VM system where we expected
348 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
349 * case is one of them. If the page is still
350 * partially dirty, make it fully dirty.
352 * note that we do not clear out the valid
353 * bits. This would prevent bogus_page
354 * replacement from working properly.
356 vm_page_set_validclean(m, base, size);
358 m->dirty = VM_PAGE_BITS_ALL;
363 vp->v_filesize = nsize;
368 * Release a page busied for a getpages operation. The page may have become
369 * wired (typically due to being used by the buffer cache) or otherwise been
370 * soft-busied and cannot be freed in that case. A held page can still be
374 vnode_pager_freepage(vm_page_t m)
376 if (m->busy || m->wire_count) {
385 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
386 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
387 * vnode_pager_generic_getpages() to implement the previous behaviour.
389 * All other FS's should use the bypass to get to the local media
390 * backing vp's VOP_GETPAGES.
393 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
397 int bytes = count * PAGE_SIZE;
400 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
401 if (rtval == EOPNOTSUPP)
402 panic("vnode_pager: vfs's must implement vop_getpages\n");
407 * This is now called from local media FS's to operate against their
408 * own vnodes if they fail to implement VOP_GETPAGES.
410 * With all the caching local media devices do these days there is really
411 * very little point to attempting to restrict the I/O size to contiguous
412 * blocks on-disk, especially if our caller thinks we need all the specified
413 * pages. Just construct and issue a READ.
416 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int bytecount,
428 * Do not do anything if the vnode is bad.
430 if (vp->v_mount == NULL)
434 * Calculate the number of pages. Since we are paging in whole
435 * pages, adjust bytecount to be an integral multiple of the page
436 * size. It will be clipped to the file EOF later on.
438 bytecount = round_page(bytecount);
439 count = bytecount / PAGE_SIZE;
442 * If we have a completely valid page available to us, we can
443 * clean up and return. Otherwise we have to re-read the
446 * Note that this does not work with NFS, so NFS has its own
447 * getpages routine. The problem is that NFS can have partially
448 * valid pages associated with the buffer cache due to the piecemeal
449 * write support. If we were to fall through and re-read the media
450 * as we do here, dirty data could be lost.
452 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
453 for (i = 0; i < count; i++) {
455 vnode_pager_freepage(m[i]);
461 * Discard pages past the file EOF. If the requested page is past
462 * the file EOF we just leave its valid bits set to 0, the caller
463 * expects to maintain ownership of the requested page. If the
464 * entire range is past file EOF discard everything and generate
467 foff = IDX_TO_OFF(m[0]->pindex);
468 if (foff >= vp->v_filesize) {
469 for (i = 0; i < count; i++) {
471 vnode_pager_freepage(m[i]);
473 return VM_PAGER_ERROR;
476 if (foff + bytecount > vp->v_filesize) {
477 bytecount = vp->v_filesize - foff;
478 i = round_page(bytecount) / PAGE_SIZE;
481 if (count != reqpage)
482 vnode_pager_freepage(m[count]);
487 * The size of the transfer is bytecount. bytecount will be an
488 * integral multiple of the page size unless it has been clipped
489 * to the file EOF. The transfer cannot exceed the file EOF.
491 * When dealing with real devices we must round-up to the device
494 if (vp->v_type == VBLK || vp->v_type == VCHR) {
495 int secmask = vp->v_rdev->si_bsize_phys - 1;
496 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
497 bytecount = (bytecount + secmask) & ~secmask;
501 * Severe hack to avoid deadlocks with the buffer cache
503 for (i = 0; i < count; ++i) {
506 vm_page_io_start(mt);
511 * Issue the I/O without any read-ahead
514 /*ioflags |= IO_SEQMAX << IO_SEQSHIFT;*/
516 aiov.iov_base = (caddr_t) 0;
517 aiov.iov_len = bytecount;
518 auio.uio_iov = &aiov;
520 auio.uio_offset = foff;
521 auio.uio_segflg = UIO_NOCOPY;
522 auio.uio_rw = UIO_READ;
523 auio.uio_resid = bytecount;
525 mycpu->gd_cnt.v_vnodein++;
526 mycpu->gd_cnt.v_vnodepgsin += count;
528 error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
531 * Severe hack to avoid deadlocks with the buffer cache
533 for (i = 0; i < count; ++i) {
536 while (vm_page_sleep_busy(mt, FALSE, "getpgs"))
539 vm_page_io_finish(mt);
543 * Calculate the actual number of bytes read and clean up the
546 bytecount -= auio.uio_resid;
548 for (i = 0; i < count; ++i) {
552 if (error == 0 && mt->valid) {
553 if (mt->flags & PG_WANTED)
554 vm_page_activate(mt);
556 vm_page_deactivate(mt);
559 vnode_pager_freepage(mt);
561 } else if (mt->valid == 0) {
563 kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt, mt->object, (int) mt->pindex);
564 /* whoops, something happened */
567 } else if (mt->valid != VM_PAGE_BITS_ALL) {
569 * Zero-extend the requested page if necessary (if
570 * the filesystem is using a small block size).
572 vm_page_zero_invalid(mt, TRUE);
576 kprintf("vnode_pager_getpages: I/O read error\n");
578 return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
582 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
583 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
584 * vnode_pager_generic_putpages() to implement the previous behaviour.
586 * All other FS's should use the bypass to get to the local media
587 * backing vp's VOP_PUTPAGES.
590 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
591 boolean_t sync, int *rtvals)
595 int bytes = count * PAGE_SIZE;
598 * Force synchronous operation if we are extremely low on memory
599 * to prevent a low-memory deadlock. VOP operations often need to
600 * allocate more memory to initiate the I/O ( i.e. do a BMAP
601 * operation ). The swapper handles the case by limiting the amount
602 * of asynchronous I/O, but that sort of solution doesn't scale well
603 * for the vnode pager without a lot of work.
605 * Also, the backing vnode's iodone routine may not wake the pageout
606 * daemon up. This should be probably be addressed XXX.
609 if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min)
613 * Call device-specific putpages function
617 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
618 if (rtval == EOPNOTSUPP) {
619 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
620 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
626 * This is now called from local media FS's to operate against their
627 * own vnodes if they fail to implement VOP_PUTPAGES.
629 * This is typically called indirectly via the pageout daemon and
630 * clustering has already typically occured, so in general we ask the
631 * underlying filesystem to write the data out asynchronously rather
635 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
636 int flags, int *rtvals)
643 vm_ooffset_t poffset;
649 object = vp->v_object;
650 count = bytecount / PAGE_SIZE;
652 for (i = 0; i < count; i++)
653 rtvals[i] = VM_PAGER_AGAIN;
655 if ((int) m[0]->pindex < 0) {
656 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
657 (long)m[0]->pindex, m[0]->dirty);
658 rtvals[0] = VM_PAGER_BAD;
662 maxsize = count * PAGE_SIZE;
665 poffset = IDX_TO_OFF(m[0]->pindex);
668 * If the page-aligned write is larger then the actual file we
669 * have to invalidate pages occuring beyond the file EOF. However,
670 * there is an edge case where a file may not be page-aligned where
671 * the last page is partially invalid. In this case the filesystem
672 * may not properly clear the dirty bits for the entire page (which
673 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
674 * With the page locked we are free to fix-up the dirty bits here.
676 * We do not under any circumstances truncate the valid bits, as
677 * this will screw up bogus page replacement.
679 * The caller has already read-protected the pages. The VFS must
680 * use the buffer cache to wrap the pages. The pages might not
681 * be immediately flushed by the buffer cache but once under its
682 * control the pages themselves can wind up being marked clean
683 * and their covering buffer cache buffer can be marked dirty.
685 if (maxsize + poffset > vp->v_filesize) {
686 if (vp->v_filesize > poffset) {
689 maxsize = vp->v_filesize - poffset;
690 ncount = btoc(maxsize);
691 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
692 vm_page_clear_dirty(m[ncount - 1], pgoff,
699 if (ncount < count) {
700 for (i = ncount; i < count; i++) {
701 rtvals[i] = VM_PAGER_BAD;
707 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
708 * rather then a bdwrite() to prevent paging I/O from saturating
709 * the buffer cache. Dummy-up the sequential heuristic to cause
710 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
711 * the system decides how to cluster.
714 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
716 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
718 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
719 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
721 aiov.iov_base = (caddr_t) 0;
722 aiov.iov_len = maxsize;
723 auio.uio_iov = &aiov;
725 auio.uio_offset = poffset;
726 auio.uio_segflg = UIO_NOCOPY;
727 auio.uio_rw = UIO_WRITE;
728 auio.uio_resid = maxsize;
730 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
731 mycpu->gd_cnt.v_vnodeout++;
732 mycpu->gd_cnt.v_vnodepgsout += ncount;
735 krateprintf(&vbadrate,
736 "vnode_pager_putpages: I/O error %d\n", error);
738 if (auio.uio_resid) {
739 krateprintf(&vresrate,
740 "vnode_pager_putpages: residual I/O %d at %lu\n",
741 auio.uio_resid, (u_long)m[0]->pindex);
743 for (i = 0; i < ncount; i++)
744 rtvals[i] = VM_PAGER_OK;
749 vnode_pager_lock(vm_object_t object)
751 struct thread *td = curthread; /* XXX */
754 for (; object != NULL; object = object->backing_object) {
755 if (object->type != OBJT_VNODE)
757 if (object->flags & OBJ_DEAD)
761 struct vnode *vp = object->handle;
762 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
764 if (object->handle != vp) {
770 if ((object->flags & OBJ_DEAD) ||
771 (object->type != OBJT_VNODE)) {
774 kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp, error, lockstatus(&vp->v_lock, td));
775 tsleep(object->handle, 0, "vnpgrl", hz);