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 <vm/vm_page2.h>
75 static void vnode_pager_dealloc (vm_object_t);
76 static int vnode_pager_getpage (vm_object_t, vm_page_t *, int);
77 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, int, int *);
78 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t);
80 struct pagerops vnodepagerops = {
87 static struct krate vbadrate = { 1 };
88 static struct krate vresrate = { 1 };
90 long vnode_pbuf_freecnt = -1; /* start out unlimited */
93 * Allocate a VM object for a vnode, typically a regular file vnode.
95 * Some additional information is required to generate a properly sized
96 * object which covers the entire buffer cache buffer straddling the file
97 * EOF. Userland does not see the extra pages as the VM fault code tests
101 vnode_pager_alloc(void *handle, off_t length, vm_prot_t prot, off_t offset,
102 int blksize, int boff)
110 * Pageout to vnode, no can do yet.
116 * XXX hack - This initialization should be put somewhere else.
118 if (vnode_pbuf_freecnt < 0) {
119 vnode_pbuf_freecnt = nswbuf_kva / 2 + 1;
123 * Serialize potential vnode/object teardowns and interlocks
125 vp = (struct vnode *)handle;
126 lwkt_gettoken(&vp->v_token);
129 * If the object is being terminated, wait for it to
132 object = vp->v_object;
134 vm_object_hold(object);
135 KKASSERT((object->flags & OBJ_DEAD) == 0);
138 if (VREFCNT(vp) <= 0)
139 panic("vnode_pager_alloc: no vnode reference");
142 * Round up to the *next* block, then destroy the buffers in question.
143 * Since we are only removing some of the buffers we must rely on the
144 * scan count to determine whether a loop is necessary.
146 * Destroy any pages beyond the last buffer.
149 boff = (int)(length % blksize);
151 loffset = length + (blksize - boff);
154 lsize = OFF_TO_IDX(round_page64(loffset));
156 if (object == NULL) {
158 * And an object of the appropriate size
160 object = vm_object_allocate_hold(OBJT_VNODE, lsize);
161 object->handle = handle;
162 vp->v_object = object;
163 vp->v_filesize = length;
164 if (vp->v_mount && (vp->v_mount->mnt_kern_flag & MNTK_NOMSYNC))
165 vm_object_set_flag(object, OBJ_NOMSYNC);
168 vm_object_reference_quick(object); /* also vref's */
169 if (object->size != lsize) {
170 kprintf("vnode_pager_alloc: Warning, objsize "
171 "mismatch %jd/%jd vp=%p obj=%p\n",
172 (intmax_t)object->size,
176 if (vp->v_filesize != length) {
177 kprintf("vnode_pager_alloc: Warning, filesize "
178 "mismatch %jd/%jd vp=%p obj=%p\n",
179 (intmax_t)vp->v_filesize,
184 vm_object_drop(object);
185 lwkt_reltoken(&vp->v_token);
191 * Add a ref to a vnode's existing VM object, return the object or
192 * NULL if the vnode did not have one. This does not create the
193 * object (we can't since we don't know what the proper blocksize/boff
194 * is to match the VFS's use of the buffer cache).
196 * The vnode must be referenced and is typically open. The object should
197 * be stable in this situation.
199 * Returns the object with an additional reference but not locked.
202 vnode_pager_reference(struct vnode *vp)
206 if ((object = vp->v_object) != NULL)
207 vm_object_reference_quick(object); /* also vref's vnode */
212 vnode_pager_dealloc(vm_object_t object)
214 struct vnode *vp = object->handle;
217 panic("vnode_pager_dealloc: pager already dealloced");
219 vm_object_pip_wait(object, "vnpdea");
221 object->handle = NULL;
222 object->type = OBJT_DEAD;
224 vp->v_filesize = NOOFFSET;
225 vclrflags(vp, VTEXT | VOBJBUF);
226 swap_pager_freespace_all(object);
230 * Return whether the vnode pager has the requested page.
233 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex)
235 struct vnode *vp = object->handle;
243 * If no vp or vp is doomed or marked transparent to VM, we do not
246 if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
250 * If filesystem no longer mounted or offset beyond end of file we do
253 loffset = IDX_TO_OFF(pindex);
255 if (vp->v_mount == NULL || loffset >= vp->v_filesize)
258 bsize = vp->v_mount->mnt_stat.f_iosize;
259 voff = loffset % bsize;
262 * XXX (obsolete - before and after pointers are now NULL)
264 * BMAP returns byte counts before and after, where after
265 * is inclusive of the base page. haspage must return page
266 * counts before and after where after does not include the
269 * BMAP is allowed to return a *after of 0 for backwards
270 * compatibility. The base page is still considered valid if
271 * no error is returned.
273 error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0);
276 if (doffset == NOOFFSET)
282 * Lets the VM system know about a change in size for a file.
283 * We adjust our own internal size and flush any cached pages in
284 * the associated object that are affected by the size change.
286 * NOTE: This routine may be invoked as a result of a pager put
287 * operation (possibly at object termination time), so we must be careful.
289 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
290 * we do not blow up on the case. nsize will always be >= 0, however.
293 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
295 vm_pindex_t nobjsize;
296 vm_pindex_t oobjsize;
299 object = vp->v_object;
302 vm_object_hold(object);
303 KKASSERT(vp->v_object == object);
306 * Hasn't changed size
308 if (nsize == vp->v_filesize) {
309 vm_object_drop(object);
314 * Has changed size. Adjust the VM object's size and v_filesize
315 * before we start scanning pages to prevent new pages from being
316 * allocated during the scan.
318 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
319 oobjsize = object->size;
320 object->size = nobjsize;
323 * File has shrunk. Toss any cached pages beyond the new EOF.
325 if (nsize < vp->v_filesize) {
326 vp->v_filesize = nsize;
327 if (nobjsize < oobjsize) {
328 vm_object_page_remove(object, nobjsize, oobjsize,
332 * This gets rid of garbage at the end of a page that is now
333 * only partially backed by the vnode. Since we are setting
334 * the entire page valid & clean after we are done we have
335 * to be sure that the portion of the page within the file
336 * bounds is already valid. If it isn't then making it
337 * valid would create a corrupt block.
339 if (nsize & PAGE_MASK) {
343 m = vm_page_lookup_busy_wait(object, OFF_TO_IDX(nsize),
347 int base = (int)nsize & PAGE_MASK;
348 int size = PAGE_SIZE - base;
350 struct lwbuf lwb_cache;
353 * Clear out partial-page garbage in case
354 * the page has been mapped.
356 * This is byte aligned.
358 lwb = lwbuf_alloc(m, &lwb_cache);
359 kva = lwbuf_kva(lwb);
360 bzero((caddr_t)kva + base, size);
364 * XXX work around SMP data integrity race
365 * by unmapping the page from user processes.
366 * The garbage we just cleared may be mapped
367 * to a user process running on another cpu
368 * and this code is not running through normal
369 * I/O channels which handle SMP issues for
370 * us, so unmap page to synchronize all cpus.
372 * XXX should vm_pager_unmap_page() have
375 vm_page_protect(m, VM_PROT_NONE);
378 * Clear out partial-page dirty bits. This
379 * has the side effect of setting the valid
380 * bits, but that is ok. There are a bunch
381 * of places in the VM system where we expected
382 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
383 * case is one of them. If the page is still
384 * partially dirty, make it fully dirty.
386 * NOTE: We do not clear out the valid
387 * bits. This would prevent bogus_page
388 * replacement from working properly.
390 * NOTE: We do not want to clear the dirty
391 * bit for a partial DEV_BSIZE'd truncation!
392 * This is DEV_BSIZE aligned!
394 vm_page_clear_dirty_beg_nonincl(m, base, size);
396 m->dirty = VM_PAGE_BITS_ALL;
403 vp->v_filesize = nsize;
405 vm_object_drop(object);
409 * Release a page busied for a getpages operation. The page may have become
410 * wired (typically due to being used by the buffer cache) or otherwise been
411 * soft-busied and cannot be freed in that case. A held page can still be
415 vnode_pager_freepage(vm_page_t m)
417 if ((m->busy_count & PBUSY_MASK) ||
419 (m->flags & PG_NEED_COMMIT)) {
428 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
429 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
430 * vnode_pager_generic_getpages() to implement the previous behaviour.
432 * All other FS's should use the bypass to get to the local media
433 * backing vp's VOP_GETPAGES.
436 vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
442 rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess);
443 if (rtval == EOPNOTSUPP)
444 panic("vnode_pager: vfs's must implement vop_getpages");
449 * This is now called from local media FS's to operate against their
450 * own vnodes if they fail to implement VOP_GETPAGES.
452 * With all the caching local media devices do these days there is really
453 * very little point to attempting to restrict the I/O size to contiguous
454 * blocks on-disk, especially if our caller thinks we need all the specified
455 * pages. Just construct and issue a READ.
458 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount,
459 int reqpage, int seqaccess)
471 * Do not do anything if the vnode is bad.
473 if (vp->v_mount == NULL)
477 * Calculate the number of pages. Since we are paging in whole
478 * pages, adjust bytecount to be an integral multiple of the page
479 * size. It will be clipped to the file EOF later on.
481 bytecount = round_page(bytecount);
482 count = bytecount / PAGE_SIZE;
485 * We could check m[reqpage]->valid here and shortcut the operation,
486 * but doing so breaks read-ahead. Instead assume that the VM
487 * system has already done at least the check, don't worry about
488 * any races, and issue the VOP_READ to allow read-ahead to function.
490 * This keeps the pipeline full for I/O bound sequentially scanned
496 * Discard pages past the file EOF. If the requested page is past
497 * the file EOF we just leave its valid bits set to 0, the caller
498 * expects to maintain ownership of the requested page. If the
499 * entire range is past file EOF discard everything and generate
502 foff = IDX_TO_OFF(mpp[0]->pindex);
503 if (foff >= vp->v_filesize) {
504 for (i = 0; i < count; i++) {
506 vnode_pager_freepage(mpp[i]);
508 return VM_PAGER_ERROR;
511 if (foff + bytecount > vp->v_filesize) {
512 bytecount = vp->v_filesize - foff;
513 i = round_page(bytecount) / PAGE_SIZE;
516 if (count != reqpage)
517 vnode_pager_freepage(mpp[count]);
522 * The size of the transfer is bytecount. bytecount will be an
523 * integral multiple of the page size unless it has been clipped
524 * to the file EOF. The transfer cannot exceed the file EOF.
526 * When dealing with real devices we must round-up to the device
529 if (vp->v_type == VBLK || vp->v_type == VCHR) {
530 int secmask = vp->v_rdev->si_bsize_phys - 1;
531 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large", secmask + 1));
532 bytecount = (bytecount + secmask) & ~secmask;
534 obytecount = bytecount;
537 * Severe hack to avoid deadlocks with the buffer cache
539 for (i = 0; i < count; ++i) {
540 vm_page_t mt = mpp[i];
542 vm_page_io_start(mt);
547 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE
551 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
553 aiov.iov_base = NULL;
554 aiov.iov_len = bytecount;
555 auio.uio_iov = &aiov;
557 auio.uio_offset = foff;
558 auio.uio_segflg = UIO_NOCOPY;
559 auio.uio_rw = UIO_READ;
560 auio.uio_resid = bytecount;
562 mycpu->gd_cnt.v_vnodein++;
563 mycpu->gd_cnt.v_vnodepgsin += count;
565 error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
568 * Severe hack to avoid deadlocks with the buffer cache
570 for (i = 0; i < count; ++i) {
571 vm_page_busy_wait(mpp[i], FALSE, "getpgs");
572 vm_page_io_finish(mpp[i]);
576 * Calculate the actual number of bytes read and clean up the
579 bytecount -= auio.uio_resid;
581 for (i = 0; i < count; ++i) {
582 vm_page_t mt = mpp[i];
585 if (error == 0 && mt->valid) {
586 if (mt->flags & PG_REFERENCED)
587 vm_page_activate(mt);
589 vm_page_deactivate(mt);
592 vnode_pager_freepage(mt);
594 } else if (mt->valid == 0) {
596 kprintf("page failed but no I/O error page "
597 "%p object %p pindex %d\n",
598 mt, mt->object, (int) mt->pindex);
599 kprintf("i=%d foff=%016lx bytecount=%d/%d "
601 i, foff, obytecount, bytecount,
603 /* whoops, something happened */
606 } else if (mt->valid != VM_PAGE_BITS_ALL) {
608 * Zero-extend the requested page if necessary (if
609 * the filesystem is using a small block size).
611 vm_page_zero_invalid(mt, TRUE);
615 kprintf("vnode_pager_getpage: I/O read error\n");
617 return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
621 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
622 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
623 * vnode_pager_generic_putpages() to implement the previous behaviour.
625 * Caller has already cleared the pmap modified bits, if any.
627 * All other FS's should use the bypass to get to the local media
628 * backing vp's VOP_PUTPAGES.
631 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
632 int sync, int *rtvals)
636 int bytes = count * PAGE_SIZE;
639 * Force synchronous operation if we are extremely low on memory
640 * to prevent a low-memory deadlock. VOP operations often need to
641 * allocate more memory to initiate the I/O ( i.e. do a BMAP
642 * operation ). The swapper handles the case by limiting the amount
643 * of asynchronous I/O, but that sort of solution doesn't scale well
644 * for the vnode pager without a lot of work.
646 * Also, the backing vnode's iodone routine may not wake the pageout
647 * daemon up. This should be probably be addressed XXX.
650 if ((vmstats.v_free_count + vmstats.v_cache_count) <
651 vmstats.v_pageout_free_min) {
656 * Call device-specific putpages function
659 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
660 if (rtval == EOPNOTSUPP) {
661 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
662 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
668 * This is now called from local media FS's to operate against their
669 * own vnodes if they fail to implement VOP_PUTPAGES.
671 * This is typically called indirectly via the pageout daemon and
672 * clustering has already typically occured, so in general we ask the
673 * underlying filesystem to write the data out asynchronously rather
677 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
678 int flags, int *rtvals)
681 int maxsize, ncount, count;
682 vm_ooffset_t poffset;
688 count = bytecount / PAGE_SIZE;
690 for (i = 0; i < count; i++)
691 rtvals[i] = VM_PAGER_AGAIN;
693 if ((int) m[0]->pindex < 0) {
694 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
695 (long)m[0]->pindex, m[0]->dirty);
696 rtvals[0] = VM_PAGER_BAD;
700 maxsize = count * PAGE_SIZE;
703 poffset = IDX_TO_OFF(m[0]->pindex);
706 * If the page-aligned write is larger then the actual file we
707 * have to invalidate pages occuring beyond the file EOF.
709 * If the file EOF resides in the middle of a page we still clear
710 * all of that page's dirty bits later on. If we didn't it would
711 * endlessly re-write.
713 * We do not under any circumstances truncate the valid bits, as
714 * this will screw up bogus page replacement.
716 * The caller has already read-protected the pages. The VFS must
717 * use the buffer cache to wrap the pages. The pages might not
718 * be immediately flushed by the buffer cache but once under its
719 * control the pages themselves can wind up being marked clean
720 * and their covering buffer cache buffer can be marked dirty.
722 if (poffset + maxsize > vp->v_filesize) {
723 if (poffset < vp->v_filesize) {
724 maxsize = vp->v_filesize - poffset;
725 ncount = btoc(maxsize);
730 if (ncount < count) {
731 for (i = ncount; i < count; i++) {
732 rtvals[i] = VM_PAGER_BAD;
738 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
739 * rather then a bdwrite() to prevent paging I/O from saturating
740 * the buffer cache. Dummy-up the sequential heuristic to cause
741 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
742 * the system decides how to cluster.
745 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
747 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
749 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
750 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
752 aiov.iov_base = (caddr_t) 0;
753 aiov.iov_len = maxsize;
754 auio.uio_iov = &aiov;
756 auio.uio_offset = poffset;
757 auio.uio_segflg = UIO_NOCOPY;
758 auio.uio_rw = UIO_WRITE;
759 auio.uio_resid = maxsize;
761 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
762 mycpu->gd_cnt.v_vnodeout++;
763 mycpu->gd_cnt.v_vnodepgsout += ncount;
766 krateprintf(&vbadrate,
767 "vnode_pager_putpages: I/O error %d\n", error);
769 if (auio.uio_resid) {
770 krateprintf(&vresrate,
771 "vnode_pager_putpages: residual I/O %zd at %lu\n",
772 auio.uio_resid, (u_long)m[0]->pindex);
775 for (i = 0; i < ncount; i++) {
776 rtvals[i] = VM_PAGER_OK;
777 vm_page_undirty(m[i]);
784 * Run the chain and if the bottom-most object is a vnode-type lock the
785 * underlying vnode. A locked vnode or NULL is returned.
787 * Caller must hold the first object.
790 vnode_pager_lock(vm_map_backing_t ba)
792 vm_map_backing_t lba;
800 while (lba->backing_ba)
801 lba = lba->backing_ba;
802 if ((lobject = lba->object) == NULL)
805 vm_object_hold_shared(lobject);
807 while (lobject->type == OBJT_VNODE &&
808 (lobject->flags & OBJ_DEAD) == 0) {
812 vp = lobject->handle;
813 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
815 if (lobject->handle == vp)
819 kprintf("vnode_pager_lock: vp %p error %d "
820 "lockstatus %d, retrying\n",
822 lockstatus(&vp->v_lock, curthread));
823 tsleep(lobject->handle, 0, "vnpgrl", hz);
828 vm_object_drop(lobject);