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>
65 #include <cpu/lwbuf.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/swap_pager.h>
74 #include <vm/vm_extern.h>
76 #include <sys/thread2.h>
77 #include <vm/vm_page2.h>
79 static void vnode_pager_dealloc (vm_object_t);
80 static int vnode_pager_getpage (vm_object_t, vm_page_t *, int);
81 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *);
82 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t);
84 struct pagerops vnodepagerops = {
91 static struct krate vbadrate = { 1 };
92 static struct krate vresrate = { 1 };
94 int vnode_pbuf_freecnt = -1; /* start out unlimited */
97 * Allocate a VM object for a vnode, typically a regular file vnode.
99 * Some additional information is required to generate a properly sized
100 * object which covers the entire buffer cache buffer straddling the file
101 * EOF. Userland does not see the extra pages as the VM fault code tests
102 * against v_filesize.
105 vnode_pager_alloc(void *handle, off_t length, vm_prot_t prot, off_t offset,
106 int blksize, int boff)
114 * Pageout to vnode, no can do yet.
120 * XXX hack - This initialization should be put somewhere else.
122 if (vnode_pbuf_freecnt < 0) {
123 vnode_pbuf_freecnt = nswbuf / 2 + 1;
126 vp = (struct vnode *) handle;
129 * Prevent race condition when allocating the object. This
130 * can happen with NFS vnodes since the nfsnode isn't locked.
132 while (vp->v_flag & VOLOCK) {
133 vsetflags(vp, VOWANT);
134 tsleep(vp, 0, "vnpobj", 0);
136 vsetflags(vp, VOLOCK);
139 * If the object is being terminated, wait for it to
142 while (((object = vp->v_object) != NULL) &&
143 (object->flags & OBJ_DEAD)) {
144 vm_object_dead_sleep(object, "vadead");
147 if (vp->v_sysref.refcnt <= 0)
148 panic("vnode_pager_alloc: no vnode reference");
151 * Round up to the *next* block, then destroy the buffers in question.
152 * Since we are only removing some of the buffers we must rely on the
153 * scan count to determine whether a loop is necessary.
155 * Destroy any pages beyond the last buffer.
158 boff = (int)(length % blksize);
160 loffset = length + (blksize - boff);
163 lsize = OFF_TO_IDX(round_page64(loffset));
165 if (object == NULL) {
167 * And an object of the appropriate size
169 object = vm_object_allocate(OBJT_VNODE, lsize);
171 object->handle = handle;
172 vp->v_object = object;
173 vp->v_filesize = length;
174 if (vp->v_mount && (vp->v_mount->mnt_kern_flag & MNTK_NOMSYNC))
175 object->flags |= OBJ_NOMSYNC;
178 if (object->size != lsize) {
179 kprintf("vnode_pager_alloc: Warning, objsize "
180 "mismatch %jd/%jd vp=%p obj=%p\n",
181 (intmax_t)object->size,
185 if (vp->v_filesize != length) {
186 kprintf("vnode_pager_alloc: Warning, filesize "
187 "mismatch %jd/%jd vp=%p obj=%p\n",
188 (intmax_t)vp->v_filesize,
195 vclrflags(vp, VOLOCK);
196 if (vp->v_flag & VOWANT) {
197 vclrflags(vp, VOWANT);
204 * Add a ref to a vnode's existing VM object, return the object or
205 * NULL if the vnode did not have one. This does not create the
206 * object (we can't since we don't know what the proper blocksize/boff
207 * is to match the VFS's use of the buffer cache).
210 vnode_pager_reference(struct vnode *vp)
215 * Prevent race condition when allocating the object. This
216 * can happen with NFS vnodes since the nfsnode isn't locked.
218 while (vp->v_flag & VOLOCK) {
219 vsetflags(vp, VOWANT);
220 tsleep(vp, 0, "vnpobj", 0);
222 vsetflags(vp, VOLOCK);
225 * Prevent race conditions against deallocation of the VM
228 while (((object = vp->v_object) != NULL) &&
229 (object->flags & OBJ_DEAD)) {
230 vm_object_dead_sleep(object, "vadead");
234 * The object is expected to exist, the caller will handle
235 * NULL returns if it does not.
242 vclrflags(vp, VOLOCK);
243 if (vp->v_flag & VOWANT) {
244 vclrflags(vp, VOWANT);
251 vnode_pager_dealloc(vm_object_t object)
253 struct vnode *vp = object->handle;
256 panic("vnode_pager_dealloc: pager already dealloced");
258 vm_object_pip_wait(object, "vnpdea");
260 object->handle = NULL;
261 object->type = OBJT_DEAD;
263 vp->v_filesize = NOOFFSET;
264 vclrflags(vp, VTEXT | VOBJBUF);
265 swap_pager_freespace_all(object);
269 * Return whether the vnode pager has the requested page. Return the
270 * number of disk-contiguous pages before and after the requested page,
271 * not including the requested page.
274 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex)
276 struct vnode *vp = object->handle;
284 * If no vp or vp is doomed or marked transparent to VM, we do not
287 if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
291 * If filesystem no longer mounted or offset beyond end of file we do
294 loffset = IDX_TO_OFF(pindex);
296 if (vp->v_mount == NULL || loffset >= vp->v_filesize)
299 bsize = vp->v_mount->mnt_stat.f_iosize;
300 voff = loffset % bsize;
305 * BMAP returns byte counts before and after, where after
306 * is inclusive of the base page. haspage must return page
307 * counts before and after where after does not include the
310 * BMAP is allowed to return a *after of 0 for backwards
311 * compatibility. The base page is still considered valid if
312 * no error is returned.
314 error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0);
317 if (doffset == NOOFFSET)
323 * Lets the VM system know about a change in size for a file.
324 * We adjust our own internal size and flush any cached pages in
325 * the associated object that are affected by the size change.
327 * NOTE: This routine may be invoked as a result of a pager put
328 * operation (possibly at object termination time), so we must be careful.
330 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
331 * we do not blow up on the case. nsize will always be >= 0, however.
334 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
336 vm_pindex_t nobjsize;
337 vm_pindex_t oobjsize;
338 vm_object_t object = vp->v_object;
344 * Hasn't changed size
346 if (nsize == vp->v_filesize)
350 * Has changed size. Adjust the VM object's size and v_filesize
351 * before we start scanning pages to prevent new pages from being
352 * allocated during the scan.
354 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
355 oobjsize = object->size;
356 object->size = nobjsize;
359 * File has shrunk. Toss any cached pages beyond the new EOF.
361 if (nsize < vp->v_filesize) {
362 vp->v_filesize = nsize;
363 if (nobjsize < oobjsize) {
364 vm_object_page_remove(object, nobjsize, oobjsize,
368 * This gets rid of garbage at the end of a page that is now
369 * only partially backed by the vnode. Since we are setting
370 * the entire page valid & clean after we are done we have
371 * to be sure that the portion of the page within the file
372 * bounds is already valid. If it isn't then making it
373 * valid would create a corrupt block.
375 if (nsize & PAGE_MASK) {
380 m = vm_page_lookup(object, OFF_TO_IDX(nsize));
381 } while (m && vm_page_sleep_busy(m, TRUE, "vsetsz"));
384 int base = (int)nsize & PAGE_MASK;
385 int size = PAGE_SIZE - base;
389 * Clear out partial-page garbage in case
390 * the page has been mapped.
392 * This is byte aligned.
395 lwb = lwbuf_alloc(m);
396 kva = lwbuf_kva(lwb);
397 bzero((caddr_t)kva + base, size);
401 * XXX work around SMP data integrity race
402 * by unmapping the page from user processes.
403 * The garbage we just cleared may be mapped
404 * to a user process running on another cpu
405 * and this code is not running through normal
406 * I/O channels which handle SMP issues for
407 * us, so unmap page to synchronize all cpus.
409 * XXX should vm_pager_unmap_page() have
412 vm_page_protect(m, VM_PROT_NONE);
415 * Clear out partial-page dirty bits. This
416 * has the side effect of setting the valid
417 * bits, but that is ok. There are a bunch
418 * of places in the VM system where we expected
419 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
420 * case is one of them. If the page is still
421 * partially dirty, make it fully dirty.
423 * NOTE: We do not clear out the valid
424 * bits. This would prevent bogus_page
425 * replacement from working properly.
427 * NOTE: We do not want to clear the dirty
428 * bit for a partial DEV_BSIZE'd truncation!
429 * This is DEV_BSIZE aligned!
431 vm_page_clear_dirty_beg_nonincl(m, base, size);
433 m->dirty = VM_PAGE_BITS_ALL;
438 vp->v_filesize = nsize;
443 * Release a page busied for a getpages operation. The page may have become
444 * wired (typically due to being used by the buffer cache) or otherwise been
445 * soft-busied and cannot be freed in that case. A held page can still be
449 vnode_pager_freepage(vm_page_t m)
451 if (m->busy || m->wire_count) {
460 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
461 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
462 * vnode_pager_generic_getpages() to implement the previous behaviour.
464 * All other FS's should use the bypass to get to the local media
465 * backing vp's VOP_GETPAGES.
468 vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
474 rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess);
475 if (rtval == EOPNOTSUPP)
476 panic("vnode_pager: vfs's must implement vop_getpages\n");
481 * This is now called from local media FS's to operate against their
482 * own vnodes if they fail to implement VOP_GETPAGES.
484 * With all the caching local media devices do these days there is really
485 * very little point to attempting to restrict the I/O size to contiguous
486 * blocks on-disk, especially if our caller thinks we need all the specified
487 * pages. Just construct and issue a READ.
490 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount,
491 int reqpage, int seqaccess)
502 * Do not do anything if the vnode is bad.
504 if (vp->v_mount == NULL)
508 * Calculate the number of pages. Since we are paging in whole
509 * pages, adjust bytecount to be an integral multiple of the page
510 * size. It will be clipped to the file EOF later on.
512 bytecount = round_page(bytecount);
513 count = bytecount / PAGE_SIZE;
516 * We could check m[reqpage]->valid here and shortcut the operation,
517 * but doing so breaks read-ahead. Instead assume that the VM
518 * system has already done at least the check, don't worry about
519 * any races, and issue the VOP_READ to allow read-ahead to function.
521 * This keeps the pipeline full for I/O bound sequentially scanned
527 * Discard pages past the file EOF. If the requested page is past
528 * the file EOF we just leave its valid bits set to 0, the caller
529 * expects to maintain ownership of the requested page. If the
530 * entire range is past file EOF discard everything and generate
533 foff = IDX_TO_OFF(mpp[0]->pindex);
534 if (foff >= vp->v_filesize) {
535 for (i = 0; i < count; i++) {
537 vnode_pager_freepage(mpp[i]);
539 return VM_PAGER_ERROR;
542 if (foff + bytecount > vp->v_filesize) {
543 bytecount = vp->v_filesize - foff;
544 i = round_page(bytecount) / PAGE_SIZE;
547 if (count != reqpage)
548 vnode_pager_freepage(mpp[count]);
553 * The size of the transfer is bytecount. bytecount will be an
554 * integral multiple of the page size unless it has been clipped
555 * to the file EOF. The transfer cannot exceed the file EOF.
557 * When dealing with real devices we must round-up to the device
560 if (vp->v_type == VBLK || vp->v_type == VCHR) {
561 int secmask = vp->v_rdev->si_bsize_phys - 1;
562 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
563 bytecount = (bytecount + secmask) & ~secmask;
567 * Severe hack to avoid deadlocks with the buffer cache
569 for (i = 0; i < count; ++i) {
570 vm_page_t mt = mpp[i];
572 vm_page_io_start(mt);
577 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE
581 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
583 aiov.iov_base = NULL;
584 aiov.iov_len = bytecount;
585 auio.uio_iov = &aiov;
587 auio.uio_offset = foff;
588 auio.uio_segflg = UIO_NOCOPY;
589 auio.uio_rw = UIO_READ;
590 auio.uio_resid = bytecount;
592 mycpu->gd_cnt.v_vnodein++;
593 mycpu->gd_cnt.v_vnodepgsin += count;
595 error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
598 * Severe hack to avoid deadlocks with the buffer cache
600 for (i = 0; i < count; ++i) {
601 vm_page_t mt = mpp[i];
603 while (vm_page_sleep_busy(mt, FALSE, "getpgs"))
606 vm_page_io_finish(mt);
610 * Calculate the actual number of bytes read and clean up the
613 bytecount -= auio.uio_resid;
615 for (i = 0; i < count; ++i) {
616 vm_page_t mt = mpp[i];
619 if (error == 0 && mt->valid) {
620 if (mt->flags & PG_WANTED)
621 vm_page_activate(mt);
623 vm_page_deactivate(mt);
626 vnode_pager_freepage(mt);
628 } else if (mt->valid == 0) {
630 kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt, mt->object, (int) mt->pindex);
631 /* whoops, something happened */
634 } else if (mt->valid != VM_PAGE_BITS_ALL) {
636 * Zero-extend the requested page if necessary (if
637 * the filesystem is using a small block size).
639 vm_page_zero_invalid(mt, TRUE);
643 kprintf("vnode_pager_getpage: I/O read error\n");
645 return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
649 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
650 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
651 * vnode_pager_generic_putpages() to implement the previous behaviour.
653 * Caller has already cleared the pmap modified bits, if any.
655 * All other FS's should use the bypass to get to the local media
656 * backing vp's VOP_PUTPAGES.
659 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
660 boolean_t sync, int *rtvals)
664 int bytes = count * PAGE_SIZE;
667 * Force synchronous operation if we are extremely low on memory
668 * to prevent a low-memory deadlock. VOP operations often need to
669 * allocate more memory to initiate the I/O ( i.e. do a BMAP
670 * operation ). The swapper handles the case by limiting the amount
671 * of asynchronous I/O, but that sort of solution doesn't scale well
672 * for the vnode pager without a lot of work.
674 * Also, the backing vnode's iodone routine may not wake the pageout
675 * daemon up. This should be probably be addressed XXX.
678 if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min)
682 * Call device-specific putpages function
685 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
686 if (rtval == EOPNOTSUPP) {
687 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
688 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
694 * This is now called from local media FS's to operate against their
695 * own vnodes if they fail to implement VOP_PUTPAGES.
697 * This is typically called indirectly via the pageout daemon and
698 * clustering has already typically occured, so in general we ask the
699 * underlying filesystem to write the data out asynchronously rather
703 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
704 int flags, int *rtvals)
708 int maxsize, ncount, count;
709 vm_ooffset_t poffset;
715 object = vp->v_object;
716 count = bytecount / PAGE_SIZE;
718 for (i = 0; i < count; i++)
719 rtvals[i] = VM_PAGER_AGAIN;
721 if ((int) m[0]->pindex < 0) {
722 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
723 (long)m[0]->pindex, m[0]->dirty);
724 rtvals[0] = VM_PAGER_BAD;
728 maxsize = count * PAGE_SIZE;
731 poffset = IDX_TO_OFF(m[0]->pindex);
734 * If the page-aligned write is larger then the actual file we
735 * have to invalidate pages occuring beyond the file EOF.
737 * If the file EOF resides in the middle of a page we still clear
738 * all of that page's dirty bits later on. If we didn't it would
739 * endlessly re-write.
741 * We do not under any circumstances truncate the valid bits, as
742 * this will screw up bogus page replacement.
744 * The caller has already read-protected the pages. The VFS must
745 * use the buffer cache to wrap the pages. The pages might not
746 * be immediately flushed by the buffer cache but once under its
747 * control the pages themselves can wind up being marked clean
748 * and their covering buffer cache buffer can be marked dirty.
750 if (poffset + maxsize > vp->v_filesize) {
751 if (poffset < vp->v_filesize) {
752 maxsize = vp->v_filesize - poffset;
753 ncount = btoc(maxsize);
758 if (ncount < count) {
759 for (i = ncount; i < count; i++) {
760 rtvals[i] = VM_PAGER_BAD;
766 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
767 * rather then a bdwrite() to prevent paging I/O from saturating
768 * the buffer cache. Dummy-up the sequential heuristic to cause
769 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
770 * the system decides how to cluster.
773 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
775 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
777 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
778 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
780 aiov.iov_base = (caddr_t) 0;
781 aiov.iov_len = maxsize;
782 auio.uio_iov = &aiov;
784 auio.uio_offset = poffset;
785 auio.uio_segflg = UIO_NOCOPY;
786 auio.uio_rw = UIO_WRITE;
787 auio.uio_resid = maxsize;
789 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
790 mycpu->gd_cnt.v_vnodeout++;
791 mycpu->gd_cnt.v_vnodepgsout += ncount;
794 krateprintf(&vbadrate,
795 "vnode_pager_putpages: I/O error %d\n", error);
797 if (auio.uio_resid) {
798 krateprintf(&vresrate,
799 "vnode_pager_putpages: residual I/O %zd at %lu\n",
800 auio.uio_resid, (u_long)m[0]->pindex);
803 for (i = 0; i < ncount; i++) {
804 rtvals[i] = VM_PAGER_OK;
805 vm_page_undirty(m[i]);
812 vnode_pager_lock(vm_object_t object)
814 struct thread *td = curthread; /* XXX */
817 for (; object != NULL; object = object->backing_object) {
818 if (object->type != OBJT_VNODE)
820 if (object->flags & OBJ_DEAD)
824 struct vnode *vp = object->handle;
825 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
827 if (object->handle != vp) {
833 if ((object->flags & OBJ_DEAD) ||
834 (object->type != OBJT_VNODE)) {
837 kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp, error, lockstatus(&vp->v_lock, td));
838 tsleep(object->handle, 0, "vnpgrl", hz);