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. All advertising materials mentioning features or use of this software
23 * must display the following acknowledgement:
24 * This product includes software developed by the University of
25 * California, Berkeley and its contributors.
26 * 4. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91
43 * $FreeBSD: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $
44 * $DragonFly: src/sys/vm/vnode_pager.c,v 1.43 2008/06/19 23:27:39 dillon Exp $
48 * Page to/from files (vnodes).
53 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
54 * greatly re-simplify the vnode_pager.
57 #include <sys/param.h>
58 #include <sys/systm.h>
59 #include <sys/kernel.h>
61 #include <sys/vnode.h>
62 #include <sys/mount.h>
64 #include <sys/vmmeter.h>
67 #include <cpu/lwbuf.h>
70 #include <vm/vm_object.h>
71 #include <vm/vm_page.h>
72 #include <vm/vm_pager.h>
73 #include <vm/vm_map.h>
74 #include <vm/vnode_pager.h>
75 #include <vm/swap_pager.h>
76 #include <vm/vm_extern.h>
78 #include <sys/thread2.h>
79 #include <vm/vm_page2.h>
81 static void vnode_pager_dealloc (vm_object_t);
82 static int vnode_pager_getpage (vm_object_t, vm_page_t *, int);
83 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *);
84 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t);
86 struct pagerops vnodepagerops = {
93 static struct krate vbadrate = { 1 };
94 static struct krate vresrate = { 1 };
96 int vnode_pbuf_freecnt = -1; /* start out unlimited */
99 * Allocate a VM object for a vnode, typically a regular file vnode.
101 * Some additional information is required to generate a properly sized
102 * object which covers the entire buffer cache buffer straddling the file
103 * EOF. Userland does not see the extra pages as the VM fault code tests
104 * against v_filesize.
107 vnode_pager_alloc(void *handle, off_t length, vm_prot_t prot, off_t offset,
108 int blksize, int boff)
116 * Pageout to vnode, no can do yet.
122 * XXX hack - This initialization should be put somewhere else.
124 if (vnode_pbuf_freecnt < 0) {
125 vnode_pbuf_freecnt = nswbuf / 2 + 1;
129 * Serialize potential vnode/object teardowns and interlocks
131 vp = (struct vnode *)handle;
132 lwkt_gettoken(&vmobj_token);
135 * Prevent race condition when allocating the object. This
136 * can happen with NFS vnodes since the nfsnode isn't locked.
138 while (vp->v_flag & VOLOCK) {
139 vsetflags(vp, VOWANT);
140 tsleep(vp, 0, "vnpobj", 0);
142 vsetflags(vp, VOLOCK);
145 * If the object is being terminated, wait for it to
148 while (((object = vp->v_object) != NULL) &&
149 (object->flags & OBJ_DEAD)) {
150 vm_object_dead_sleep(object, "vadead");
153 if (vp->v_sysref.refcnt <= 0)
154 panic("vnode_pager_alloc: no vnode reference");
157 * Round up to the *next* block, then destroy the buffers in question.
158 * Since we are only removing some of the buffers we must rely on the
159 * scan count to determine whether a loop is necessary.
161 * Destroy any pages beyond the last buffer.
164 boff = (int)(length % blksize);
166 loffset = length + (blksize - boff);
169 lsize = OFF_TO_IDX(round_page64(loffset));
171 if (object == NULL) {
173 * And an object of the appropriate size
175 object = vm_object_allocate(OBJT_VNODE, lsize);
177 object->handle = handle;
178 vp->v_object = object;
179 vp->v_filesize = length;
180 if (vp->v_mount && (vp->v_mount->mnt_kern_flag & MNTK_NOMSYNC))
181 vm_object_set_flag(object, OBJ_NOMSYNC);
183 object->ref_count++; /* protected by vmobj_token */
184 if (object->size != lsize) {
185 kprintf("vnode_pager_alloc: Warning, objsize "
186 "mismatch %jd/%jd vp=%p obj=%p\n",
187 (intmax_t)object->size,
191 if (vp->v_filesize != length) {
192 kprintf("vnode_pager_alloc: Warning, filesize "
193 "mismatch %jd/%jd vp=%p obj=%p\n",
194 (intmax_t)vp->v_filesize,
201 vclrflags(vp, VOLOCK);
202 if (vp->v_flag & VOWANT) {
203 vclrflags(vp, VOWANT);
206 lwkt_reltoken(&vmobj_token);
212 * Add a ref to a vnode's existing VM object, return the object or
213 * NULL if the vnode did not have one. This does not create the
214 * object (we can't since we don't know what the proper blocksize/boff
215 * is to match the VFS's use of the buffer cache).
218 vnode_pager_reference(struct vnode *vp)
223 * Serialize potential vnode/object teardowns and interlocks
225 lwkt_gettoken(&vmobj_token);
228 * Prevent race condition when allocating the object. This
229 * can happen with NFS vnodes since the nfsnode isn't locked.
231 while (vp->v_flag & VOLOCK) {
232 vsetflags(vp, VOWANT);
233 tsleep(vp, 0, "vnpobj", 0);
235 vsetflags(vp, VOLOCK);
238 * Prevent race conditions against deallocation of the VM
241 while (((object = vp->v_object) != NULL) &&
242 (object->flags & OBJ_DEAD)) {
243 vm_object_dead_sleep(object, "vadead");
247 * The object is expected to exist, the caller will handle
248 * NULL returns if it does not.
251 object->ref_count++; /* protected by vmobj_token */
255 vclrflags(vp, VOLOCK);
256 if (vp->v_flag & VOWANT) {
257 vclrflags(vp, VOWANT);
261 lwkt_reltoken(&vmobj_token);
266 vnode_pager_dealloc(vm_object_t object)
268 struct vnode *vp = object->handle;
271 panic("vnode_pager_dealloc: pager already dealloced");
273 vm_object_pip_wait(object, "vnpdea");
275 object->handle = NULL;
276 object->type = OBJT_DEAD;
278 vp->v_filesize = NOOFFSET;
279 vclrflags(vp, VTEXT | VOBJBUF);
280 swap_pager_freespace_all(object);
284 * Return whether the vnode pager has the requested page. Return the
285 * number of disk-contiguous pages before and after the requested page,
286 * not including the requested page.
289 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex)
291 struct vnode *vp = object->handle;
299 * If no vp or vp is doomed or marked transparent to VM, we do not
302 if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
306 * If filesystem no longer mounted or offset beyond end of file we do
309 loffset = IDX_TO_OFF(pindex);
311 if (vp->v_mount == NULL || loffset >= vp->v_filesize)
314 bsize = vp->v_mount->mnt_stat.f_iosize;
315 voff = loffset % bsize;
320 * BMAP returns byte counts before and after, where after
321 * is inclusive of the base page. haspage must return page
322 * counts before and after where after does not include the
325 * BMAP is allowed to return a *after of 0 for backwards
326 * compatibility. The base page is still considered valid if
327 * no error is returned.
329 error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0);
332 if (doffset == NOOFFSET)
338 * Lets the VM system know about a change in size for a file.
339 * We adjust our own internal size and flush any cached pages in
340 * the associated object that are affected by the size change.
342 * NOTE: This routine may be invoked as a result of a pager put
343 * operation (possibly at object termination time), so we must be careful.
345 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
346 * we do not blow up on the case. nsize will always be >= 0, however.
349 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
351 vm_pindex_t nobjsize;
352 vm_pindex_t oobjsize;
353 vm_object_t object = vp->v_object;
359 * Hasn't changed size
361 if (nsize == vp->v_filesize)
364 lwkt_gettoken(&vm_token);
367 * Has changed size. Adjust the VM object's size and v_filesize
368 * before we start scanning pages to prevent new pages from being
369 * allocated during the scan.
371 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
372 oobjsize = object->size;
373 object->size = nobjsize;
376 * File has shrunk. Toss any cached pages beyond the new EOF.
378 if (nsize < vp->v_filesize) {
379 vp->v_filesize = nsize;
380 if (nobjsize < oobjsize) {
381 vm_object_page_remove(object, nobjsize, oobjsize,
385 * This gets rid of garbage at the end of a page that is now
386 * only partially backed by the vnode. Since we are setting
387 * the entire page valid & clean after we are done we have
388 * to be sure that the portion of the page within the file
389 * bounds is already valid. If it isn't then making it
390 * valid would create a corrupt block.
392 if (nsize & PAGE_MASK) {
397 m = vm_page_lookup(object, OFF_TO_IDX(nsize));
398 } while (m && vm_page_sleep_busy(m, TRUE, "vsetsz"));
401 int base = (int)nsize & PAGE_MASK;
402 int size = PAGE_SIZE - base;
404 struct lwbuf lwb_cache;
407 * Clear out partial-page garbage in case
408 * the page has been mapped.
410 * This is byte aligned.
413 lwb = lwbuf_alloc(m, &lwb_cache);
414 kva = lwbuf_kva(lwb);
415 bzero((caddr_t)kva + base, size);
419 * XXX work around SMP data integrity race
420 * by unmapping the page from user processes.
421 * The garbage we just cleared may be mapped
422 * to a user process running on another cpu
423 * and this code is not running through normal
424 * I/O channels which handle SMP issues for
425 * us, so unmap page to synchronize all cpus.
427 * XXX should vm_pager_unmap_page() have
430 vm_page_protect(m, VM_PROT_NONE);
433 * Clear out partial-page dirty bits. This
434 * has the side effect of setting the valid
435 * bits, but that is ok. There are a bunch
436 * of places in the VM system where we expected
437 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
438 * case is one of them. If the page is still
439 * partially dirty, make it fully dirty.
441 * NOTE: We do not clear out the valid
442 * bits. This would prevent bogus_page
443 * replacement from working properly.
445 * NOTE: We do not want to clear the dirty
446 * bit for a partial DEV_BSIZE'd truncation!
447 * This is DEV_BSIZE aligned!
449 vm_page_clear_dirty_beg_nonincl(m, base, size);
451 m->dirty = VM_PAGE_BITS_ALL;
456 vp->v_filesize = nsize;
458 lwkt_reltoken(&vm_token);
462 * Release a page busied for a getpages operation. The page may have become
463 * wired (typically due to being used by the buffer cache) or otherwise been
464 * soft-busied and cannot be freed in that case. A held page can still be
468 vnode_pager_freepage(vm_page_t m)
470 if (m->busy || m->wire_count) {
479 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
480 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
481 * vnode_pager_generic_getpages() to implement the previous behaviour.
483 * All other FS's should use the bypass to get to the local media
484 * backing vp's VOP_GETPAGES.
487 vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
493 rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess);
494 if (rtval == EOPNOTSUPP)
495 panic("vnode_pager: vfs's must implement vop_getpages\n");
500 * This is now called from local media FS's to operate against their
501 * own vnodes if they fail to implement VOP_GETPAGES.
503 * With all the caching local media devices do these days there is really
504 * very little point to attempting to restrict the I/O size to contiguous
505 * blocks on-disk, especially if our caller thinks we need all the specified
506 * pages. Just construct and issue a READ.
509 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount,
510 int reqpage, int seqaccess)
521 * Do not do anything if the vnode is bad.
523 if (vp->v_mount == NULL)
527 * Calculate the number of pages. Since we are paging in whole
528 * pages, adjust bytecount to be an integral multiple of the page
529 * size. It will be clipped to the file EOF later on.
531 bytecount = round_page(bytecount);
532 count = bytecount / PAGE_SIZE;
535 * We could check m[reqpage]->valid here and shortcut the operation,
536 * but doing so breaks read-ahead. Instead assume that the VM
537 * system has already done at least the check, don't worry about
538 * any races, and issue the VOP_READ to allow read-ahead to function.
540 * This keeps the pipeline full for I/O bound sequentially scanned
546 * Discard pages past the file EOF. If the requested page is past
547 * the file EOF we just leave its valid bits set to 0, the caller
548 * expects to maintain ownership of the requested page. If the
549 * entire range is past file EOF discard everything and generate
552 foff = IDX_TO_OFF(mpp[0]->pindex);
553 if (foff >= vp->v_filesize) {
554 for (i = 0; i < count; i++) {
556 vnode_pager_freepage(mpp[i]);
558 return VM_PAGER_ERROR;
561 if (foff + bytecount > vp->v_filesize) {
562 bytecount = vp->v_filesize - foff;
563 i = round_page(bytecount) / PAGE_SIZE;
566 if (count != reqpage)
567 vnode_pager_freepage(mpp[count]);
572 * The size of the transfer is bytecount. bytecount will be an
573 * integral multiple of the page size unless it has been clipped
574 * to the file EOF. The transfer cannot exceed the file EOF.
576 * When dealing with real devices we must round-up to the device
579 if (vp->v_type == VBLK || vp->v_type == VCHR) {
580 int secmask = vp->v_rdev->si_bsize_phys - 1;
581 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
582 bytecount = (bytecount + secmask) & ~secmask;
586 * Severe hack to avoid deadlocks with the buffer cache
588 for (i = 0; i < count; ++i) {
589 vm_page_t mt = mpp[i];
591 vm_page_io_start(mt);
596 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE
600 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
602 aiov.iov_base = NULL;
603 aiov.iov_len = bytecount;
604 auio.uio_iov = &aiov;
606 auio.uio_offset = foff;
607 auio.uio_segflg = UIO_NOCOPY;
608 auio.uio_rw = UIO_READ;
609 auio.uio_resid = bytecount;
611 mycpu->gd_cnt.v_vnodein++;
612 mycpu->gd_cnt.v_vnodepgsin += count;
614 error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
617 * Severe hack to avoid deadlocks with the buffer cache
619 lwkt_gettoken(&vm_token);
620 for (i = 0; i < count; ++i) {
621 vm_page_t mt = mpp[i];
623 while (vm_page_sleep_busy(mt, FALSE, "getpgs"))
626 vm_page_io_finish(mt);
628 lwkt_reltoken(&vm_token);
631 * Calculate the actual number of bytes read and clean up the
634 bytecount -= auio.uio_resid;
636 for (i = 0; i < count; ++i) {
637 vm_page_t mt = mpp[i];
640 if (error == 0 && mt->valid) {
641 if (mt->flags & PG_WANTED)
642 vm_page_activate(mt);
644 vm_page_deactivate(mt);
647 vnode_pager_freepage(mt);
649 } else if (mt->valid == 0) {
651 kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt, mt->object, (int) mt->pindex);
652 /* whoops, something happened */
655 } else if (mt->valid != VM_PAGE_BITS_ALL) {
657 * Zero-extend the requested page if necessary (if
658 * the filesystem is using a small block size).
660 vm_page_zero_invalid(mt, TRUE);
664 kprintf("vnode_pager_getpage: I/O read error\n");
666 return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
670 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
671 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
672 * vnode_pager_generic_putpages() to implement the previous behaviour.
674 * Caller has already cleared the pmap modified bits, if any.
676 * All other FS's should use the bypass to get to the local media
677 * backing vp's VOP_PUTPAGES.
680 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
681 boolean_t sync, int *rtvals)
685 int bytes = count * PAGE_SIZE;
688 * Force synchronous operation if we are extremely low on memory
689 * to prevent a low-memory deadlock. VOP operations often need to
690 * allocate more memory to initiate the I/O ( i.e. do a BMAP
691 * operation ). The swapper handles the case by limiting the amount
692 * of asynchronous I/O, but that sort of solution doesn't scale well
693 * for the vnode pager without a lot of work.
695 * Also, the backing vnode's iodone routine may not wake the pageout
696 * daemon up. This should be probably be addressed XXX.
699 if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min)
703 * Call device-specific putpages function
706 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
707 if (rtval == EOPNOTSUPP) {
708 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
709 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
715 * This is now called from local media FS's to operate against their
716 * own vnodes if they fail to implement VOP_PUTPAGES.
718 * This is typically called indirectly via the pageout daemon and
719 * clustering has already typically occured, so in general we ask the
720 * underlying filesystem to write the data out asynchronously rather
724 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
725 int flags, int *rtvals)
729 int maxsize, ncount, count;
730 vm_ooffset_t poffset;
736 object = vp->v_object;
737 count = bytecount / PAGE_SIZE;
739 for (i = 0; i < count; i++)
740 rtvals[i] = VM_PAGER_AGAIN;
742 if ((int) m[0]->pindex < 0) {
743 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
744 (long)m[0]->pindex, m[0]->dirty);
745 rtvals[0] = VM_PAGER_BAD;
749 maxsize = count * PAGE_SIZE;
752 poffset = IDX_TO_OFF(m[0]->pindex);
755 * If the page-aligned write is larger then the actual file we
756 * have to invalidate pages occuring beyond the file EOF.
758 * If the file EOF resides in the middle of a page we still clear
759 * all of that page's dirty bits later on. If we didn't it would
760 * endlessly re-write.
762 * We do not under any circumstances truncate the valid bits, as
763 * this will screw up bogus page replacement.
765 * The caller has already read-protected the pages. The VFS must
766 * use the buffer cache to wrap the pages. The pages might not
767 * be immediately flushed by the buffer cache but once under its
768 * control the pages themselves can wind up being marked clean
769 * and their covering buffer cache buffer can be marked dirty.
771 if (poffset + maxsize > vp->v_filesize) {
772 if (poffset < vp->v_filesize) {
773 maxsize = vp->v_filesize - poffset;
774 ncount = btoc(maxsize);
779 if (ncount < count) {
780 for (i = ncount; i < count; i++) {
781 rtvals[i] = VM_PAGER_BAD;
787 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
788 * rather then a bdwrite() to prevent paging I/O from saturating
789 * the buffer cache. Dummy-up the sequential heuristic to cause
790 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
791 * the system decides how to cluster.
794 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
796 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
798 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
799 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
801 aiov.iov_base = (caddr_t) 0;
802 aiov.iov_len = maxsize;
803 auio.uio_iov = &aiov;
805 auio.uio_offset = poffset;
806 auio.uio_segflg = UIO_NOCOPY;
807 auio.uio_rw = UIO_WRITE;
808 auio.uio_resid = maxsize;
810 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
811 mycpu->gd_cnt.v_vnodeout++;
812 mycpu->gd_cnt.v_vnodepgsout += ncount;
815 krateprintf(&vbadrate,
816 "vnode_pager_putpages: I/O error %d\n", error);
818 if (auio.uio_resid) {
819 krateprintf(&vresrate,
820 "vnode_pager_putpages: residual I/O %zd at %lu\n",
821 auio.uio_resid, (u_long)m[0]->pindex);
824 for (i = 0; i < ncount; i++) {
825 rtvals[i] = VM_PAGER_OK;
826 vm_page_undirty(m[i]);
833 vnode_pager_lock(vm_object_t object)
835 struct thread *td = curthread; /* XXX */
838 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
840 for (; object != NULL; object = object->backing_object) {
841 if (object->type != OBJT_VNODE)
843 if (object->flags & OBJ_DEAD)
847 struct vnode *vp = object->handle;
848 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
850 if (object->handle != vp) {
856 if ((object->flags & OBJ_DEAD) ||
857 (object->type != OBJT_VNODE)) {
860 kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp, error, lockstatus(&vp->v_lock, td));
861 tsleep(object->handle, 0, "vnpgrl", hz);