4 * Copyright (c) 1990 University of Utah.
5 * Copyright (c) 1991 The Regents of the University of California.
7 * Copyright (c) 1993, 1994 John S. Dyson
8 * Copyright (c) 1995, David Greenman
10 * This code is derived from software contributed to Berkeley by
11 * the Systems Programming Group of the University of Utah Computer
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
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. Neither the name of the University nor the names of its contributors
23 * may be used to endorse or promote products derived from this software
24 * without specific prior written permission.
26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91
39 * $FreeBSD: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $
43 * Page to/from files (vnodes).
48 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
49 * greatly re-simplify the vnode_pager.
52 #include <sys/param.h>
53 #include <sys/systm.h>
54 #include <sys/kernel.h>
56 #include <sys/vnode.h>
57 #include <sys/mount.h>
59 #include <sys/vmmeter.h>
62 #include <cpu/lwbuf.h>
65 #include <vm/vm_object.h>
66 #include <vm/vm_page.h>
67 #include <vm/vm_pager.h>
68 #include <vm/vm_map.h>
69 #include <vm/vnode_pager.h>
70 #include <vm/swap_pager.h>
71 #include <vm/vm_extern.h>
73 #include <sys/thread2.h>
74 #include <vm/vm_page2.h>
76 static void vnode_pager_dealloc (vm_object_t);
77 static int vnode_pager_getpage (vm_object_t, vm_page_t *, int);
78 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *);
79 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t);
81 struct pagerops vnodepagerops = {
88 static struct krate vbadrate = { 1 };
89 static struct krate vresrate = { 1 };
91 long vnode_pbuf_freecnt = -1; /* start out unlimited */
94 * Allocate a VM object for a vnode, typically a regular file vnode.
96 * Some additional information is required to generate a properly sized
97 * object which covers the entire buffer cache buffer straddling the file
98 * EOF. Userland does not see the extra pages as the VM fault code tests
102 vnode_pager_alloc(void *handle, off_t length, vm_prot_t prot, off_t offset,
103 int blksize, int boff)
111 * Pageout to vnode, no can do yet.
117 * XXX hack - This initialization should be put somewhere else.
119 if (vnode_pbuf_freecnt < 0) {
120 vnode_pbuf_freecnt = nswbuf / 2 + 1;
124 * Serialize potential vnode/object teardowns and interlocks
126 vp = (struct vnode *)handle;
127 lwkt_gettoken(&vp->v_token);
130 * Prevent race condition when allocating the object. This
131 * can happen with NFS vnodes since the nfsnode isn't locked.
133 while (vp->v_flag & VOLOCK) {
134 vsetflags(vp, VOWANT);
135 tsleep(vp, 0, "vnpobj", 0);
137 vsetflags(vp, VOLOCK);
138 lwkt_reltoken(&vp->v_token);
141 * If the object is being terminated, wait for it to
144 while ((object = vp->v_object) != NULL) {
145 vm_object_hold(object);
146 if ((object->flags & OBJ_DEAD) == 0)
148 vm_object_dead_sleep(object, "vadead");
149 vm_object_drop(object);
152 if (vp->v_sysref.refcnt <= 0)
153 panic("vnode_pager_alloc: no vnode reference");
156 * Round up to the *next* block, then destroy the buffers in question.
157 * Since we are only removing some of the buffers we must rely on the
158 * scan count to determine whether a loop is necessary.
160 * Destroy any pages beyond the last buffer.
163 boff = (int)(length % blksize);
165 loffset = length + (blksize - boff);
168 lsize = OFF_TO_IDX(round_page64(loffset));
170 if (object == NULL) {
172 * And an object of the appropriate size
174 object = vm_object_allocate_hold(OBJT_VNODE, lsize);
175 object->handle = handle;
176 vp->v_object = object;
177 vp->v_filesize = length;
178 if (vp->v_mount && (vp->v_mount->mnt_kern_flag & MNTK_NOMSYNC))
179 vm_object_set_flag(object, OBJ_NOMSYNC);
182 if (object->size != lsize) {
183 kprintf("vnode_pager_alloc: Warning, objsize "
184 "mismatch %jd/%jd vp=%p obj=%p\n",
185 (intmax_t)object->size,
189 if (vp->v_filesize != length) {
190 kprintf("vnode_pager_alloc: Warning, filesize "
191 "mismatch %jd/%jd vp=%p obj=%p\n",
192 (intmax_t)vp->v_filesize,
199 lwkt_gettoken(&vp->v_token);
200 vclrflags(vp, VOLOCK);
201 if (vp->v_flag & VOWANT) {
202 vclrflags(vp, VOWANT);
205 lwkt_reltoken(&vp->v_token);
207 vm_object_drop(object);
213 * Add a ref to a vnode's existing VM object, return the object or
214 * NULL if the vnode did not have one. This does not create the
215 * object (we can't since we don't know what the proper blocksize/boff
216 * is to match the VFS's use of the buffer cache).
219 vnode_pager_reference(struct vnode *vp)
224 * Prevent race condition when allocating the object. This
225 * can happen with NFS vnodes since the nfsnode isn't locked.
227 * Serialize potential vnode/object teardowns and interlocks
229 lwkt_gettoken(&vp->v_token);
230 while (vp->v_flag & VOLOCK) {
231 vsetflags(vp, VOWANT);
232 tsleep(vp, 0, "vnpobj", 0);
234 vsetflags(vp, VOLOCK);
235 lwkt_reltoken(&vp->v_token);
238 * Prevent race conditions against deallocation of the VM
241 while ((object = vp->v_object) != NULL) {
242 vm_object_hold(object);
243 if ((object->flags & OBJ_DEAD) == 0)
245 vm_object_dead_sleep(object, "vadead");
246 vm_object_drop(object);
250 * The object is expected to exist, the caller will handle
251 * NULL returns if it does not.
258 lwkt_gettoken(&vp->v_token);
259 vclrflags(vp, VOLOCK);
260 if (vp->v_flag & VOWANT) {
261 vclrflags(vp, VOWANT);
264 lwkt_reltoken(&vp->v_token);
266 vm_object_drop(object);
272 vnode_pager_dealloc(vm_object_t object)
274 struct vnode *vp = object->handle;
277 panic("vnode_pager_dealloc: pager already dealloced");
279 vm_object_pip_wait(object, "vnpdea");
281 object->handle = NULL;
282 object->type = OBJT_DEAD;
284 vp->v_filesize = NOOFFSET;
285 vclrflags(vp, VTEXT | VOBJBUF);
286 swap_pager_freespace_all(object);
290 * Return whether the vnode pager has the requested page. Return the
291 * number of disk-contiguous pages before and after the requested page,
292 * not including the requested page.
295 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex)
297 struct vnode *vp = object->handle;
305 * If no vp or vp is doomed or marked transparent to VM, we do not
308 if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
312 * If filesystem no longer mounted or offset beyond end of file we do
315 loffset = IDX_TO_OFF(pindex);
317 if (vp->v_mount == NULL || loffset >= vp->v_filesize)
320 bsize = vp->v_mount->mnt_stat.f_iosize;
321 voff = loffset % bsize;
326 * BMAP returns byte counts before and after, where after
327 * is inclusive of the base page. haspage must return page
328 * counts before and after where after does not include the
331 * BMAP is allowed to return a *after of 0 for backwards
332 * compatibility. The base page is still considered valid if
333 * no error is returned.
335 error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0);
338 if (doffset == NOOFFSET)
344 * Lets the VM system know about a change in size for a file.
345 * We adjust our own internal size and flush any cached pages in
346 * the associated object that are affected by the size change.
348 * NOTE: This routine may be invoked as a result of a pager put
349 * operation (possibly at object termination time), so we must be careful.
351 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
352 * we do not blow up on the case. nsize will always be >= 0, however.
355 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
357 vm_pindex_t nobjsize;
358 vm_pindex_t oobjsize;
361 while ((object = vp->v_object) != NULL) {
362 vm_object_hold(object);
363 if (vp->v_object == object)
365 vm_object_drop(object);
371 * Hasn't changed size
373 if (nsize == vp->v_filesize) {
374 vm_object_drop(object);
379 * Has changed size. Adjust the VM object's size and v_filesize
380 * before we start scanning pages to prevent new pages from being
381 * allocated during the scan.
383 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
384 oobjsize = object->size;
385 object->size = nobjsize;
388 * File has shrunk. Toss any cached pages beyond the new EOF.
390 if (nsize < vp->v_filesize) {
391 vp->v_filesize = nsize;
392 if (nobjsize < oobjsize) {
393 vm_object_page_remove(object, nobjsize, oobjsize,
397 * This gets rid of garbage at the end of a page that is now
398 * only partially backed by the vnode. Since we are setting
399 * the entire page valid & clean after we are done we have
400 * to be sure that the portion of the page within the file
401 * bounds is already valid. If it isn't then making it
402 * valid would create a corrupt block.
404 if (nsize & PAGE_MASK) {
408 m = vm_page_lookup_busy_wait(object, OFF_TO_IDX(nsize),
412 int base = (int)nsize & PAGE_MASK;
413 int size = PAGE_SIZE - base;
415 struct lwbuf lwb_cache;
418 * Clear out partial-page garbage in case
419 * the page has been mapped.
421 * This is byte aligned.
423 lwb = lwbuf_alloc(m, &lwb_cache);
424 kva = lwbuf_kva(lwb);
425 bzero((caddr_t)kva + base, size);
429 * XXX work around SMP data integrity race
430 * by unmapping the page from user processes.
431 * The garbage we just cleared may be mapped
432 * to a user process running on another cpu
433 * and this code is not running through normal
434 * I/O channels which handle SMP issues for
435 * us, so unmap page to synchronize all cpus.
437 * XXX should vm_pager_unmap_page() have
440 vm_page_protect(m, VM_PROT_NONE);
443 * Clear out partial-page dirty bits. This
444 * has the side effect of setting the valid
445 * bits, but that is ok. There are a bunch
446 * of places in the VM system where we expected
447 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
448 * case is one of them. If the page is still
449 * partially dirty, make it fully dirty.
451 * NOTE: We do not clear out the valid
452 * bits. This would prevent bogus_page
453 * replacement from working properly.
455 * NOTE: We do not want to clear the dirty
456 * bit for a partial DEV_BSIZE'd truncation!
457 * This is DEV_BSIZE aligned!
459 vm_page_clear_dirty_beg_nonincl(m, base, size);
461 m->dirty = VM_PAGE_BITS_ALL;
468 vp->v_filesize = nsize;
470 vm_object_drop(object);
474 * Release a page busied for a getpages operation. The page may have become
475 * wired (typically due to being used by the buffer cache) or otherwise been
476 * soft-busied and cannot be freed in that case. A held page can still be
480 vnode_pager_freepage(vm_page_t m)
482 if (m->busy || m->wire_count || (m->flags & PG_NEED_COMMIT)) {
491 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
492 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
493 * vnode_pager_generic_getpages() to implement the previous behaviour.
495 * All other FS's should use the bypass to get to the local media
496 * backing vp's VOP_GETPAGES.
499 vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
505 rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess);
506 if (rtval == EOPNOTSUPP)
507 panic("vnode_pager: vfs's must implement vop_getpages");
512 * This is now called from local media FS's to operate against their
513 * own vnodes if they fail to implement VOP_GETPAGES.
515 * With all the caching local media devices do these days there is really
516 * very little point to attempting to restrict the I/O size to contiguous
517 * blocks on-disk, especially if our caller thinks we need all the specified
518 * pages. Just construct and issue a READ.
521 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount,
522 int reqpage, int seqaccess)
533 * Do not do anything if the vnode is bad.
535 if (vp->v_mount == NULL)
539 * Calculate the number of pages. Since we are paging in whole
540 * pages, adjust bytecount to be an integral multiple of the page
541 * size. It will be clipped to the file EOF later on.
543 bytecount = round_page(bytecount);
544 count = bytecount / PAGE_SIZE;
547 * We could check m[reqpage]->valid here and shortcut the operation,
548 * but doing so breaks read-ahead. Instead assume that the VM
549 * system has already done at least the check, don't worry about
550 * any races, and issue the VOP_READ to allow read-ahead to function.
552 * This keeps the pipeline full for I/O bound sequentially scanned
558 * Discard pages past the file EOF. If the requested page is past
559 * the file EOF we just leave its valid bits set to 0, the caller
560 * expects to maintain ownership of the requested page. If the
561 * entire range is past file EOF discard everything and generate
564 foff = IDX_TO_OFF(mpp[0]->pindex);
565 if (foff >= vp->v_filesize) {
566 for (i = 0; i < count; i++) {
568 vnode_pager_freepage(mpp[i]);
570 return VM_PAGER_ERROR;
573 if (foff + bytecount > vp->v_filesize) {
574 bytecount = vp->v_filesize - foff;
575 i = round_page(bytecount) / PAGE_SIZE;
578 if (count != reqpage)
579 vnode_pager_freepage(mpp[count]);
584 * The size of the transfer is bytecount. bytecount will be an
585 * integral multiple of the page size unless it has been clipped
586 * to the file EOF. The transfer cannot exceed the file EOF.
588 * When dealing with real devices we must round-up to the device
591 if (vp->v_type == VBLK || vp->v_type == VCHR) {
592 int secmask = vp->v_rdev->si_bsize_phys - 1;
593 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large", secmask + 1));
594 bytecount = (bytecount + secmask) & ~secmask;
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 vm_page_io_start(mt);
608 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE
612 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
614 aiov.iov_base = NULL;
615 aiov.iov_len = bytecount;
616 auio.uio_iov = &aiov;
618 auio.uio_offset = foff;
619 auio.uio_segflg = UIO_NOCOPY;
620 auio.uio_rw = UIO_READ;
621 auio.uio_resid = bytecount;
623 mycpu->gd_cnt.v_vnodein++;
624 mycpu->gd_cnt.v_vnodepgsin += count;
626 error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
629 * Severe hack to avoid deadlocks with the buffer cache
631 for (i = 0; i < count; ++i) {
632 vm_page_busy_wait(mpp[i], FALSE, "getpgs");
633 vm_page_io_finish(mpp[i]);
637 * Calculate the actual number of bytes read and clean up the
640 bytecount -= auio.uio_resid;
642 for (i = 0; i < count; ++i) {
643 vm_page_t mt = mpp[i];
646 if (error == 0 && mt->valid) {
647 if (mt->flags & PG_REFERENCED)
648 vm_page_activate(mt);
650 vm_page_deactivate(mt);
653 vnode_pager_freepage(mt);
655 } else if (mt->valid == 0) {
657 kprintf("page failed but no I/O error page "
658 "%p object %p pindex %d\n",
659 mt, mt->object, (int) mt->pindex);
660 /* whoops, something happened */
663 } else if (mt->valid != VM_PAGE_BITS_ALL) {
665 * Zero-extend the requested page if necessary (if
666 * the filesystem is using a small block size).
668 vm_page_zero_invalid(mt, TRUE);
672 kprintf("vnode_pager_getpage: I/O read error\n");
674 return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
678 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
679 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
680 * vnode_pager_generic_putpages() to implement the previous behaviour.
682 * Caller has already cleared the pmap modified bits, if any.
684 * All other FS's should use the bypass to get to the local media
685 * backing vp's VOP_PUTPAGES.
688 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
689 boolean_t sync, int *rtvals)
693 int bytes = count * PAGE_SIZE;
696 * Force synchronous operation if we are extremely low on memory
697 * to prevent a low-memory deadlock. VOP operations often need to
698 * allocate more memory to initiate the I/O ( i.e. do a BMAP
699 * operation ). The swapper handles the case by limiting the amount
700 * of asynchronous I/O, but that sort of solution doesn't scale well
701 * for the vnode pager without a lot of work.
703 * Also, the backing vnode's iodone routine may not wake the pageout
704 * daemon up. This should be probably be addressed XXX.
707 if ((vmstats.v_free_count + vmstats.v_cache_count) <
708 vmstats.v_pageout_free_min) {
713 * Call device-specific putpages function
716 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
717 if (rtval == EOPNOTSUPP) {
718 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
719 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
725 * This is now called from local media FS's to operate against their
726 * own vnodes if they fail to implement VOP_PUTPAGES.
728 * This is typically called indirectly via the pageout daemon and
729 * clustering has already typically occured, so in general we ask the
730 * underlying filesystem to write the data out asynchronously rather
734 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
735 int flags, int *rtvals)
738 int maxsize, ncount, count;
739 vm_ooffset_t poffset;
745 count = bytecount / PAGE_SIZE;
747 for (i = 0; i < count; i++)
748 rtvals[i] = VM_PAGER_AGAIN;
750 if ((int) m[0]->pindex < 0) {
751 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
752 (long)m[0]->pindex, m[0]->dirty);
753 rtvals[0] = VM_PAGER_BAD;
757 maxsize = count * PAGE_SIZE;
760 poffset = IDX_TO_OFF(m[0]->pindex);
763 * If the page-aligned write is larger then the actual file we
764 * have to invalidate pages occuring beyond the file EOF.
766 * If the file EOF resides in the middle of a page we still clear
767 * all of that page's dirty bits later on. If we didn't it would
768 * endlessly re-write.
770 * We do not under any circumstances truncate the valid bits, as
771 * this will screw up bogus page replacement.
773 * The caller has already read-protected the pages. The VFS must
774 * use the buffer cache to wrap the pages. The pages might not
775 * be immediately flushed by the buffer cache but once under its
776 * control the pages themselves can wind up being marked clean
777 * and their covering buffer cache buffer can be marked dirty.
779 if (poffset + maxsize > vp->v_filesize) {
780 if (poffset < vp->v_filesize) {
781 maxsize = vp->v_filesize - poffset;
782 ncount = btoc(maxsize);
787 if (ncount < count) {
788 for (i = ncount; i < count; i++) {
789 rtvals[i] = VM_PAGER_BAD;
795 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
796 * rather then a bdwrite() to prevent paging I/O from saturating
797 * the buffer cache. Dummy-up the sequential heuristic to cause
798 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
799 * the system decides how to cluster.
802 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
804 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
806 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
807 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
809 aiov.iov_base = (caddr_t) 0;
810 aiov.iov_len = maxsize;
811 auio.uio_iov = &aiov;
813 auio.uio_offset = poffset;
814 auio.uio_segflg = UIO_NOCOPY;
815 auio.uio_rw = UIO_WRITE;
816 auio.uio_resid = maxsize;
818 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
819 mycpu->gd_cnt.v_vnodeout++;
820 mycpu->gd_cnt.v_vnodepgsout += ncount;
823 krateprintf(&vbadrate,
824 "vnode_pager_putpages: I/O error %d\n", error);
826 if (auio.uio_resid) {
827 krateprintf(&vresrate,
828 "vnode_pager_putpages: residual I/O %zd at %lu\n",
829 auio.uio_resid, (u_long)m[0]->pindex);
832 for (i = 0; i < ncount; i++) {
833 rtvals[i] = VM_PAGER_OK;
834 vm_page_undirty(m[i]);
841 * Run the chain and if the bottom-most object is a vnode-type lock the
842 * underlying vnode. A locked vnode or NULL is returned.
845 vnode_pager_lock(vm_object_t object)
847 struct vnode *vp = NULL;
855 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
858 while (lobject->type != OBJT_VNODE) {
859 if (lobject->flags & OBJ_DEAD)
861 tobject = lobject->backing_object;
864 vm_object_hold_shared(tobject);
865 if (tobject == lobject->backing_object) {
866 if (lobject != object) {
867 vm_object_lock_swap();
868 vm_object_drop(lobject);
872 vm_object_drop(tobject);
875 while (lobject->type == OBJT_VNODE &&
876 (lobject->flags & OBJ_DEAD) == 0) {
880 vp = lobject->handle;
881 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
883 if (lobject->handle == vp)
887 kprintf("vnode_pager_lock: vp %p error %d "
888 "lockstatus %d, retrying\n",
890 lockstatus(&vp->v_lock, curthread));
891 tsleep(object->handle, 0, "vnpgrl", hz);
895 if (lobject != object)
896 vm_object_drop(lobject);