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>
55 #include <sys/kernel.h>
57 #include <sys/vnode.h>
58 #include <sys/mount.h>
60 #include <sys/vmmeter.h>
63 #include <cpu/lwbuf.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_pager.h>
69 #include <vm/vm_map.h>
70 #include <vm/vnode_pager.h>
71 #include <vm/swap_pager.h>
72 #include <vm/vm_extern.h>
74 #include <vm/vm_page2.h>
76 static pgo_dealloc_t vnode_pager_dealloc;
77 static pgo_getpage_t vnode_pager_getpage;
78 static pgo_putpages_t vnode_pager_putpages;
79 static pgo_haspage_t vnode_pager_haspage;
81 struct pagerops vnodepagerops = {
82 .pgo_dealloc = vnode_pager_dealloc,
83 .pgo_getpage = vnode_pager_getpage,
84 .pgo_putpages = vnode_pager_putpages,
85 .pgo_haspage = vnode_pager_haspage
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_kva / 2 + 1;
124 * Serialize potential vnode/object teardowns and interlocks
126 vp = (struct vnode *)handle;
127 lwkt_gettoken(&vp->v_token);
130 * If the object is being terminated, wait for it to
133 object = vp->v_object;
135 vm_object_hold(object);
136 KKASSERT((object->flags & OBJ_DEAD) == 0);
139 if (VREFCNT(vp) <= 0)
140 panic("vnode_pager_alloc: no vnode reference");
143 * Round up to the *next* block, then destroy the buffers in question.
144 * Since we are only removing some of the buffers we must rely on the
145 * scan count to determine whether a loop is necessary.
147 * Destroy any pages beyond the last buffer.
150 boff = (int)(length % blksize);
152 loffset = length + (blksize - boff);
155 lsize = OFF_TO_IDX(round_page64(loffset));
157 if (object == NULL) {
159 * And an object of the appropriate size
161 object = vm_object_allocate_hold(OBJT_VNODE, lsize);
162 object->handle = handle;
163 vp->v_object = object;
164 vp->v_filesize = length;
165 if (vp->v_mount && (vp->v_mount->mnt_kern_flag & MNTK_NOMSYNC))
166 vm_object_set_flag(object, OBJ_NOMSYNC);
169 vm_object_reference_quick(object); /* also vref's */
170 if (object->size != lsize) {
171 kprintf("vnode_pager_alloc: Warning, objsize "
172 "mismatch %jd/%jd vp=%p obj=%p\n",
173 (intmax_t)object->size,
177 if (vp->v_filesize != length) {
178 kprintf("vnode_pager_alloc: Warning, filesize "
179 "mismatch %jd/%jd vp=%p obj=%p\n",
180 (intmax_t)vp->v_filesize,
185 vm_object_drop(object);
186 lwkt_reltoken(&vp->v_token);
192 * Add a ref to a vnode's existing VM object, return the object or
193 * NULL if the vnode did not have one. This does not create the
194 * object (we can't since we don't know what the proper blocksize/boff
195 * is to match the VFS's use of the buffer cache).
197 * The vnode must be referenced and is typically open. The object should
198 * be stable in this situation.
200 * Returns the object with an additional reference but not locked.
203 vnode_pager_reference(struct vnode *vp)
207 if ((object = vp->v_object) != NULL)
208 vm_object_reference_quick(object); /* also vref's vnode */
213 vnode_pager_dealloc(vm_object_t object)
215 struct vnode *vp = object->handle;
218 panic("vnode_pager_dealloc: pager already dealloced");
220 vm_object_pip_wait(object, "vnpdea");
222 object->handle = NULL;
223 object->type = OBJT_DEAD;
225 vp->v_filesize = NOOFFSET;
226 vclrflags(vp, VTEXT | VOBJBUF);
227 swap_pager_freespace_all(object);
231 * Return whether the vnode pager has the requested page.
234 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex)
236 struct vnode *vp = object->handle;
244 * If no vp or vp is doomed or marked transparent to VM, we do not
247 if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
251 * If filesystem no longer mounted or offset beyond end of file we do
254 loffset = IDX_TO_OFF(pindex);
256 if (vp->v_mount == NULL || loffset >= vp->v_filesize)
259 bsize = vp->v_mount->mnt_stat.f_iosize;
260 voff = loffset % bsize;
263 * XXX (obsolete - before and after pointers are now NULL)
265 * BMAP returns byte counts before and after, where after
266 * is inclusive of the base page. haspage must return page
267 * counts before and after where after does not include the
270 * BMAP is allowed to return a *after of 0 for backwards
271 * compatibility. The base page is still considered valid if
272 * no error is returned.
274 error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0);
277 if (doffset == NOOFFSET)
283 * Lets the VM system know about a change in size for a file.
284 * We adjust our own internal size and flush any cached pages in
285 * the associated object that are affected by the size change.
287 * NOTE: This routine may be invoked as a result of a pager put
288 * operation (possibly at object termination time), so we must be careful.
290 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
291 * we do not blow up on the case. nsize will always be >= 0, however.
294 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
296 vm_pindex_t nobjsize;
297 vm_pindex_t oobjsize;
300 object = vp->v_object;
303 vm_object_hold(object);
304 KKASSERT(vp->v_object == object);
307 * Hasn't changed size
309 if (nsize == vp->v_filesize) {
310 vm_object_drop(object);
315 * Has changed size. Adjust the VM object's size and v_filesize
316 * before we start scanning pages to prevent new pages from being
317 * allocated during the scan.
319 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
320 oobjsize = object->size;
321 object->size = nobjsize;
324 * File has shrunk. Toss any cached pages beyond the new EOF.
326 if (nsize < vp->v_filesize) {
327 vp->v_filesize = nsize;
328 if (nobjsize < oobjsize) {
329 vm_object_page_remove(object, nobjsize, oobjsize,
333 * This gets rid of garbage at the end of a page that is now
334 * only partially backed by the vnode. Since we are setting
335 * the entire page valid & clean after we are done we have
336 * to be sure that the portion of the page within the file
337 * bounds is already valid. If it isn't then making it
338 * valid would create a corrupt block.
340 if (nsize & PAGE_MASK) {
344 m = vm_page_lookup_busy_wait(object, OFF_TO_IDX(nsize),
348 int base = (int)nsize & PAGE_MASK;
349 int size = PAGE_SIZE - base;
351 struct lwbuf lwb_cache;
354 * Clear out partial-page garbage in case
355 * the page has been mapped.
357 * This is byte aligned.
359 lwb = lwbuf_alloc(m, &lwb_cache);
360 kva = lwbuf_kva(lwb);
361 bzero((caddr_t)kva + base, size);
365 * XXX work around SMP data integrity race
366 * by unmapping the page from user processes.
367 * The garbage we just cleared may be mapped
368 * to a user process running on another cpu
369 * and this code is not running through normal
370 * I/O channels which handle SMP issues for
371 * us, so unmap page to synchronize all cpus.
373 * XXX should vm_pager_unmap_page() have
376 vm_page_protect(m, VM_PROT_NONE);
379 * Clear out partial-page dirty bits. This
380 * has the side effect of setting the valid
381 * bits, but that is ok. There are a bunch
382 * of places in the VM system where we expected
383 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
384 * case is one of them. If the page is still
385 * partially dirty, make it fully dirty.
387 * NOTE: We do not clear out the valid
388 * bits. This would prevent bogus_page
389 * replacement from working properly.
391 * NOTE: We do not want to clear the dirty
392 * bit for a partial DEV_BSIZE'd truncation!
393 * This is DEV_BSIZE aligned!
395 vm_page_clear_dirty_beg_nonincl(m, base, size);
397 m->dirty = VM_PAGE_BITS_ALL;
404 vp->v_filesize = nsize;
406 vm_object_drop(object);
410 * Release a page busied for a getpages operation. The page may have become
411 * wired (typically due to being used by the buffer cache) or otherwise been
412 * soft-busied and cannot be freed in that case. A held page can still be
416 vnode_pager_freepage(vm_page_t m)
418 if ((m->busy_count & PBUSY_MASK) ||
420 (m->flags & PG_NEED_COMMIT)) {
429 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
430 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
431 * vnode_pager_generic_getpages() to implement the previous behaviour.
433 * All other FS's should use the bypass to get to the local media
434 * backing vp's VOP_GETPAGES.
437 vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
443 rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess);
444 if (rtval == EOPNOTSUPP)
445 panic("vnode_pager: vfs's must implement vop_getpages");
450 * This is now called from local media FS's to operate against their
451 * own vnodes if they fail to implement VOP_GETPAGES.
453 * With all the caching local media devices do these days there is really
454 * very little point to attempting to restrict the I/O size to contiguous
455 * blocks on-disk, especially if our caller thinks we need all the specified
456 * pages. Just construct and issue a READ.
459 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount,
460 int reqpage, int seqaccess)
472 * Do not do anything if the vnode is bad.
474 if (vp->v_mount == NULL)
478 * Calculate the number of pages. Since we are paging in whole
479 * pages, adjust bytecount to be an integral multiple of the page
480 * size. It will be clipped to the file EOF later on.
482 bytecount = round_page(bytecount);
483 count = bytecount / PAGE_SIZE;
486 * We could check m[reqpage]->valid here and shortcut the operation,
487 * but doing so breaks read-ahead. Instead assume that the VM
488 * system has already done at least the check, don't worry about
489 * any races, and issue the VOP_READ to allow read-ahead to function.
491 * This keeps the pipeline full for I/O bound sequentially scanned
497 * Discard pages past the file EOF. If the requested page is past
498 * the file EOF we just leave its valid bits set to 0, the caller
499 * expects to maintain ownership of the requested page. If the
500 * entire range is past file EOF discard everything and generate
503 foff = IDX_TO_OFF(mpp[0]->pindex);
504 if (foff >= vp->v_filesize) {
505 for (i = 0; i < count; i++) {
507 vnode_pager_freepage(mpp[i]);
509 return VM_PAGER_ERROR;
512 if (foff + bytecount > vp->v_filesize) {
513 bytecount = vp->v_filesize - foff;
514 i = round_page(bytecount) / PAGE_SIZE;
517 if (count != reqpage)
518 vnode_pager_freepage(mpp[count]);
523 * The size of the transfer is bytecount. bytecount will be an
524 * integral multiple of the page size unless it has been clipped
525 * to the file EOF. The transfer cannot exceed the file EOF.
527 * When dealing with real devices we must round-up to the device
530 if (vp->v_type == VBLK || vp->v_type == VCHR) {
531 int secmask = vp->v_rdev->si_bsize_phys - 1;
532 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large", secmask + 1));
533 bytecount = (bytecount + secmask) & ~secmask;
535 obytecount = bytecount;
538 * Severe hack to avoid deadlocks with the buffer cache
540 for (i = 0; i < count; ++i) {
541 vm_page_t mt = mpp[i];
543 vm_page_io_start(mt);
548 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE
552 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
554 aiov.iov_base = NULL;
555 aiov.iov_len = bytecount;
556 auio.uio_iov = &aiov;
558 auio.uio_offset = foff;
559 auio.uio_segflg = UIO_NOCOPY;
560 auio.uio_rw = UIO_READ;
561 auio.uio_resid = bytecount;
563 mycpu->gd_cnt.v_vnodein++;
564 mycpu->gd_cnt.v_vnodepgsin += count;
566 error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
569 * Severe hack to avoid deadlocks with the buffer cache
571 for (i = 0; i < count; ++i) {
572 vm_page_busy_wait(mpp[i], FALSE, "getpgs");
573 vm_page_io_finish(mpp[i]);
577 * Calculate the actual number of bytes read and clean up the
580 bytecount -= auio.uio_resid;
582 for (i = 0; i < count; ++i) {
583 vm_page_t mt = mpp[i];
586 if (error == 0 && mt->valid) {
587 if (mt->flags & PG_REFERENCED)
588 vm_page_activate(mt);
590 vm_page_deactivate(mt);
593 vnode_pager_freepage(mt);
595 } else if (mt->valid == 0) {
597 kprintf("page failed but no I/O error page "
598 "%p object %p pindex %d\n",
599 mt, mt->object, (int) mt->pindex);
600 kprintf("i=%d foff=%016lx bytecount=%d/%d "
602 i, foff, obytecount, bytecount,
604 /* whoops, something happened */
607 } else if (mt->valid != VM_PAGE_BITS_ALL) {
609 * Zero-extend the requested page if necessary (if
610 * the filesystem is using a small block size).
612 vm_page_zero_invalid(mt, TRUE);
616 kprintf("vnode_pager_getpage: I/O read error\n");
618 return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
622 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
623 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
624 * vnode_pager_generic_putpages() to implement the previous behaviour.
626 * Caller has already cleared the pmap modified bits, if any.
628 * All other FS's should use the bypass to get to the local media
629 * backing vp's VOP_PUTPAGES.
632 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
633 int flags, int *rtvals)
637 int bytes = count * PAGE_SIZE;
640 * Force synchronous operation if we are extremely low on memory
641 * to prevent a low-memory deadlock. VOP operations often need to
642 * allocate more memory to initiate the I/O ( i.e. do a BMAP
643 * operation ). The swapper handles the case by limiting the amount
644 * of asynchronous I/O, but that sort of solution doesn't scale well
645 * for the vnode pager without a lot of work.
647 * Also, the backing vnode's iodone routine may not wake the pageout
648 * daemon up. This should be probably be addressed XXX.
651 if ((vmstats.v_free_count + vmstats.v_cache_count) <
652 vmstats.v_pageout_free_min) {
657 * Call device-specific putpages function
660 rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals, 0);
661 if (rtval == EOPNOTSUPP) {
662 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
663 rtval = vnode_pager_generic_putpages( vp, m, bytes, flags, rtvals);
669 * This is now called from local media FS's to operate against their
670 * own vnodes if they fail to implement VOP_PUTPAGES.
672 * This is typically called indirectly via the pageout daemon and
673 * clustering has already typically occured, so in general we ask the
674 * underlying filesystem to write the data out asynchronously rather
678 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
679 int flags, int *rtvals)
682 int maxsize, ncount, count;
683 vm_ooffset_t poffset;
689 count = bytecount / PAGE_SIZE;
691 for (i = 0; i < count; i++)
692 rtvals[i] = VM_PAGER_AGAIN;
694 if ((int) m[0]->pindex < 0) {
695 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
696 (long)m[0]->pindex, m[0]->dirty);
697 rtvals[0] = VM_PAGER_BAD;
701 maxsize = count * PAGE_SIZE;
704 poffset = IDX_TO_OFF(m[0]->pindex);
707 * If the page-aligned write is larger then the actual file we
708 * have to invalidate pages occuring beyond the file EOF.
710 * If the file EOF resides in the middle of a page we still clear
711 * all of that page's dirty bits later on. If we didn't it would
712 * endlessly re-write.
714 * We do not under any circumstances truncate the valid bits, as
715 * this will screw up bogus page replacement.
717 * The caller has already read-protected the pages. The VFS must
718 * use the buffer cache to wrap the pages. The pages might not
719 * be immediately flushed by the buffer cache but once under its
720 * control the pages themselves can wind up being marked clean
721 * and their covering buffer cache buffer can be marked dirty.
723 if (poffset + maxsize > vp->v_filesize) {
724 if (poffset < vp->v_filesize) {
725 maxsize = vp->v_filesize - poffset;
726 ncount = btoc(maxsize);
731 if (ncount < count) {
732 for (i = ncount; i < count; i++) {
733 rtvals[i] = VM_PAGER_BAD;
739 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
740 * rather then a bdwrite() to prevent paging I/O from saturating
741 * the buffer cache. Dummy-up the sequential heuristic to cause
742 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
743 * the system decides how to cluster.
746 if (flags & (OBJPC_SYNC | OBJPC_INVAL))
748 else if ((flags & OBJPC_CLUSTER_OK) == 0)
750 ioflags |= (flags & OBJPC_INVAL) ? IO_INVAL: 0;
751 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
753 aiov.iov_base = (caddr_t) 0;
754 aiov.iov_len = maxsize;
755 auio.uio_iov = &aiov;
757 auio.uio_offset = poffset;
758 auio.uio_segflg = UIO_NOCOPY;
759 auio.uio_rw = UIO_WRITE;
760 auio.uio_resid = maxsize;
762 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
763 mycpu->gd_cnt.v_vnodeout++;
764 mycpu->gd_cnt.v_vnodepgsout += ncount;
767 krateprintf(&vbadrate,
768 "vnode_pager_putpages: I/O error %d\n", error);
770 if (auio.uio_resid) {
771 krateprintf(&vresrate,
772 "vnode_pager_putpages: residual I/O %zd at %lu\n",
773 auio.uio_resid, (u_long)m[0]->pindex);
776 for (i = 0; i < ncount; i++) {
777 rtvals[i] = VM_PAGER_OK;
778 vm_page_undirty(m[i]);
785 * Run the chain and if the bottom-most object is a vnode-type lock the
786 * underlying vnode. A locked vnode or NULL is returned.
788 * Caller must hold the first object.
791 vnode_pager_lock(vm_map_backing_t ba)
793 vm_map_backing_t lba;
801 while (lba->backing_ba)
802 lba = lba->backing_ba;
803 if ((lobject = lba->object) == NULL)
806 vm_object_hold_shared(lobject);
808 while (lobject->type == OBJT_VNODE &&
809 (lobject->flags & OBJ_DEAD) == 0) {
813 vp = lobject->handle;
814 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
816 if (lobject->handle == vp)
820 kprintf("vnode_pager_lock: vp %p error %d "
821 "lockstatus %d, retrying\n",
823 lockstatus(&vp->v_lock, curthread));
824 tsleep(lobject->handle, 0, "vnpgrl", hz);
829 vm_object_drop(lobject);