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
44 * Page to/from files (vnodes).
49 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
50 * greatly re-simplify the vnode_pager.
53 #include <sys/cdefs.h>
54 __FBSDID("$FreeBSD$");
58 #include <sys/param.h>
59 #include <sys/systm.h>
61 #include <sys/vnode.h>
62 #include <sys/mount.h>
65 #include <sys/vmmeter.h>
66 #include <sys/limits.h>
68 #include <sys/rwlock.h>
69 #include <sys/sf_buf.h>
71 #include <machine/atomic.h>
74 #include <vm/vm_param.h>
75 #include <vm/vm_object.h>
76 #include <vm/vm_page.h>
77 #include <vm/vm_pager.h>
78 #include <vm/vm_map.h>
79 #include <vm/vnode_pager.h>
80 #include <vm/vm_extern.h>
82 static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
83 daddr_t *rtaddress, int *run);
84 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
85 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
86 static void vnode_pager_dealloc(vm_object_t);
87 static int vnode_pager_local_getpages0(struct vnode *, vm_page_t *, int, int,
88 vop_getpages_iodone_t, void *);
89 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int);
90 static int vnode_pager_getpages_async(vm_object_t, vm_page_t *, int, int,
91 vop_getpages_iodone_t, void *);
92 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *);
93 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
94 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t,
95 vm_ooffset_t, struct ucred *cred);
96 static int vnode_pager_generic_getpages_done(struct buf *);
97 static void vnode_pager_generic_getpages_done_async(struct buf *);
99 struct pagerops vnodepagerops = {
100 .pgo_alloc = vnode_pager_alloc,
101 .pgo_dealloc = vnode_pager_dealloc,
102 .pgo_getpages = vnode_pager_getpages,
103 .pgo_getpages_async = vnode_pager_getpages_async,
104 .pgo_putpages = vnode_pager_putpages,
105 .pgo_haspage = vnode_pager_haspage,
108 int vnode_pbuf_freecnt;
109 int vnode_async_pbuf_freecnt;
111 /* Create the VM system backing object for this vnode */
113 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
116 vm_ooffset_t size = isize;
119 if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE)
122 while ((object = vp->v_object) != NULL) {
123 VM_OBJECT_WLOCK(object);
124 if (!(object->flags & OBJ_DEAD)) {
125 VM_OBJECT_WUNLOCK(object);
129 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
130 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vodead", 0);
131 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
135 if (vn_isdisk(vp, NULL)) {
136 size = IDX_TO_OFF(INT_MAX);
138 if (VOP_GETATTR(vp, &va, td->td_ucred))
144 object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred);
146 * Dereference the reference we just created. This assumes
147 * that the object is associated with the vp.
149 VM_OBJECT_WLOCK(object);
151 VM_OBJECT_WUNLOCK(object);
154 KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
160 vnode_destroy_vobject(struct vnode *vp)
162 struct vm_object *obj;
167 ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
168 VM_OBJECT_WLOCK(obj);
169 if (obj->ref_count == 0) {
171 * don't double-terminate the object
173 if ((obj->flags & OBJ_DEAD) == 0)
174 vm_object_terminate(obj);
176 VM_OBJECT_WUNLOCK(obj);
179 * Woe to the process that tries to page now :-).
181 vm_pager_deallocate(obj);
182 VM_OBJECT_WUNLOCK(obj);
189 * Allocate (or lookup) pager for a vnode.
190 * Handle is a vnode pointer.
195 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
196 vm_ooffset_t offset, struct ucred *cred)
202 * Pageout to vnode, no can do yet.
207 vp = (struct vnode *) handle;
210 * If the object is being terminated, wait for it to
214 while ((object = vp->v_object) != NULL) {
215 VM_OBJECT_WLOCK(object);
216 if ((object->flags & OBJ_DEAD) == 0)
218 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
219 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vadead", 0);
222 KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference"));
224 if (object == NULL) {
226 * Add an object of the appropriate size
228 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
230 object->un_pager.vnp.vnp_size = size;
231 object->un_pager.vnp.writemappings = 0;
233 object->handle = handle;
235 if (vp->v_object != NULL) {
237 * Object has been created while we were sleeping
240 VM_OBJECT_WLOCK(object);
241 KASSERT(object->ref_count == 1,
242 ("leaked ref %p %d", object, object->ref_count));
243 object->type = OBJT_DEAD;
244 object->ref_count = 0;
245 VM_OBJECT_WUNLOCK(object);
246 vm_object_destroy(object);
249 vp->v_object = object;
253 #if VM_NRESERVLEVEL > 0
254 vm_object_color(object, 0);
256 VM_OBJECT_WUNLOCK(object);
263 * The object must be locked.
266 vnode_pager_dealloc(vm_object_t object)
273 panic("vnode_pager_dealloc: pager already dealloced");
275 VM_OBJECT_ASSERT_WLOCKED(object);
276 vm_object_pip_wait(object, "vnpdea");
277 refs = object->ref_count;
279 object->handle = NULL;
280 object->type = OBJT_DEAD;
281 if (object->flags & OBJ_DISCONNECTWNT) {
282 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
285 ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
286 if (object->un_pager.vnp.writemappings > 0) {
287 object->un_pager.vnp.writemappings = 0;
288 VOP_ADD_WRITECOUNT(vp, -1);
289 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
290 __func__, vp, vp->v_writecount);
294 VM_OBJECT_WUNLOCK(object);
297 VM_OBJECT_WLOCK(object);
301 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
304 struct vnode *vp = object->handle;
310 int pagesperblock, blocksperpage;
312 VM_OBJECT_ASSERT_WLOCKED(object);
314 * If no vp or vp is doomed or marked transparent to VM, we do not
317 if (vp == NULL || vp->v_iflag & VI_DOOMED)
320 * If the offset is beyond end of file we do
323 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
326 bsize = vp->v_mount->mnt_stat.f_iosize;
327 pagesperblock = bsize / PAGE_SIZE;
329 if (pagesperblock > 0) {
330 reqblock = pindex / pagesperblock;
332 blocksperpage = (PAGE_SIZE / bsize);
333 reqblock = pindex * blocksperpage;
335 VM_OBJECT_WUNLOCK(object);
336 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
337 VM_OBJECT_WLOCK(object);
342 if (pagesperblock > 0) {
343 poff = pindex - (reqblock * pagesperblock);
345 *before *= pagesperblock;
350 * The BMAP vop can report a partial block in the
351 * 'after', but must not report blocks after EOF.
352 * Assert the latter, and truncate 'after' in case
355 KASSERT((reqblock + *after) * pagesperblock <
356 roundup2(object->size, pagesperblock),
357 ("%s: reqblock %jd after %d size %ju", __func__,
358 (intmax_t )reqblock, *after,
359 (uintmax_t )object->size));
360 *after *= pagesperblock;
361 *after += pagesperblock - (poff + 1);
362 if (pindex + *after >= object->size)
363 *after = object->size - 1 - pindex;
367 *before /= blocksperpage;
371 *after /= blocksperpage;
378 * Lets the VM system know about a change in size for a file.
379 * We adjust our own internal size and flush any cached pages in
380 * the associated object that are affected by the size change.
382 * Note: this routine may be invoked as a result of a pager put
383 * operation (possibly at object termination time), so we must be careful.
386 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
390 vm_pindex_t nobjsize;
392 if ((object = vp->v_object) == NULL)
394 /* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */
395 VM_OBJECT_WLOCK(object);
396 if (object->type == OBJT_DEAD) {
397 VM_OBJECT_WUNLOCK(object);
400 KASSERT(object->type == OBJT_VNODE,
401 ("not vnode-backed object %p", object));
402 if (nsize == object->un_pager.vnp.vnp_size) {
404 * Hasn't changed size
406 VM_OBJECT_WUNLOCK(object);
409 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
410 if (nsize < object->un_pager.vnp.vnp_size) {
412 * File has shrunk. Toss any cached pages beyond the new EOF.
414 if (nobjsize < object->size)
415 vm_object_page_remove(object, nobjsize, object->size,
418 * this gets rid of garbage at the end of a page that is now
419 * only partially backed by the vnode.
421 * XXX for some reason (I don't know yet), if we take a
422 * completely invalid page and mark it partially valid
423 * it can screw up NFS reads, so we don't allow the case.
425 if ((nsize & PAGE_MASK) &&
426 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
428 int base = (int)nsize & PAGE_MASK;
429 int size = PAGE_SIZE - base;
432 * Clear out partial-page garbage in case
433 * the page has been mapped.
435 pmap_zero_page_area(m, base, size);
438 * Update the valid bits to reflect the blocks that
439 * have been zeroed. Some of these valid bits may
440 * have already been set.
442 vm_page_set_valid_range(m, base, size);
445 * Round "base" to the next block boundary so that the
446 * dirty bit for a partially zeroed block is not
449 base = roundup2(base, DEV_BSIZE);
452 * Clear out partial-page dirty bits.
454 * note that we do not clear out the valid
455 * bits. This would prevent bogus_page
456 * replacement from working properly.
458 vm_page_clear_dirty(m, base, PAGE_SIZE - base);
459 } else if ((nsize & PAGE_MASK) &&
460 vm_page_is_cached(object, OFF_TO_IDX(nsize))) {
461 vm_page_cache_free(object, OFF_TO_IDX(nsize),
465 object->un_pager.vnp.vnp_size = nsize;
466 object->size = nobjsize;
467 VM_OBJECT_WUNLOCK(object);
471 * calculate the linear (byte) disk address of specified virtual
475 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
486 if (vp->v_iflag & VI_DOOMED)
489 bsize = vp->v_mount->mnt_stat.f_iosize;
490 vblock = address / bsize;
491 voffset = address % bsize;
493 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
495 if (*rtaddress != -1)
496 *rtaddress += voffset / DEV_BSIZE;
499 *run *= bsize/PAGE_SIZE;
500 *run -= voffset/PAGE_SIZE;
508 * small block filesystem vnode pager input
511 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m)
524 if (vp->v_iflag & VI_DOOMED)
527 bsize = vp->v_mount->mnt_stat.f_iosize;
529 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
531 sf = sf_buf_alloc(m, 0);
533 for (i = 0; i < PAGE_SIZE / bsize; i++) {
534 vm_ooffset_t address;
536 bits = vm_page_bits(i * bsize, bsize);
540 address = IDX_TO_OFF(m->pindex) + i * bsize;
541 if (address >= object->un_pager.vnp.vnp_size) {
544 error = vnode_pager_addr(vp, address, &fileaddr, NULL);
548 if (fileaddr != -1) {
549 bp = getpbuf(&vnode_pbuf_freecnt);
551 /* build a minimal buffer header */
552 bp->b_iocmd = BIO_READ;
553 bp->b_iodone = bdone;
554 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
555 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
556 bp->b_rcred = crhold(curthread->td_ucred);
557 bp->b_wcred = crhold(curthread->td_ucred);
558 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
559 bp->b_blkno = fileaddr;
562 bp->b_bcount = bsize;
563 bp->b_bufsize = bsize;
564 bp->b_runningbufspace = bp->b_bufsize;
565 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
568 bp->b_iooffset = dbtob(bp->b_blkno);
571 bwait(bp, PVM, "vnsrd");
573 if ((bp->b_ioflags & BIO_ERROR) != 0)
577 * free the buffer header back to the swap buffer pool
581 relpbuf(bp, &vnode_pbuf_freecnt);
585 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
586 KASSERT((m->dirty & bits) == 0,
587 ("vnode_pager_input_smlfs: page %p is dirty", m));
588 VM_OBJECT_WLOCK(object);
590 VM_OBJECT_WUNLOCK(object);
594 return VM_PAGER_ERROR;
600 * old style vnode pager input routine
603 vnode_pager_input_old(vm_object_t object, vm_page_t m)
612 VM_OBJECT_ASSERT_WLOCKED(object);
616 * Return failure if beyond current EOF
618 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
622 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
623 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
625 VM_OBJECT_WUNLOCK(object);
628 * Allocate a kernel virtual address and initialize so that
629 * we can use VOP_READ/WRITE routines.
631 sf = sf_buf_alloc(m, 0);
633 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
635 auio.uio_iov = &aiov;
637 auio.uio_offset = IDX_TO_OFF(m->pindex);
638 auio.uio_segflg = UIO_SYSSPACE;
639 auio.uio_rw = UIO_READ;
640 auio.uio_resid = size;
641 auio.uio_td = curthread;
643 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
645 int count = size - auio.uio_resid;
649 else if (count != PAGE_SIZE)
650 bzero((caddr_t)sf_buf_kva(sf) + count,
655 VM_OBJECT_WLOCK(object);
657 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
659 m->valid = VM_PAGE_BITS_ALL;
660 return error ? VM_PAGER_ERROR : VM_PAGER_OK;
664 * generic vnode pager input routine
668 * Local media VFS's that do not implement their own VOP_GETPAGES
669 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
670 * to implement the previous behaviour.
672 * All other FS's should use the bypass to get to the local media
673 * backing vp's VOP_GETPAGES.
676 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
680 int bytes = count * PAGE_SIZE;
683 VM_OBJECT_WUNLOCK(object);
684 rtval = VOP_GETPAGES(vp, m, bytes, reqpage);
685 KASSERT(rtval != EOPNOTSUPP,
686 ("vnode_pager: FS getpages not implemented\n"));
687 VM_OBJECT_WLOCK(object);
692 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
693 int reqpage, vop_getpages_iodone_t iodone, void *arg)
699 VM_OBJECT_WUNLOCK(object);
700 rtval = VOP_GETPAGES_ASYNC(vp, m, count * PAGE_SIZE, reqpage,
702 KASSERT(rtval != EOPNOTSUPP,
703 ("vnode_pager: FS getpages_async not implemented\n"));
704 VM_OBJECT_WLOCK(object);
709 * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for
710 * local filesystems, where partially valid pages can only occur at
714 vnode_pager_local_getpages(struct vop_getpages_args *ap)
717 return (vnode_pager_local_getpages0(ap->a_vp, ap->a_m, ap->a_count,
718 ap->a_reqpage, NULL, NULL));
722 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap)
725 return (vnode_pager_local_getpages0(ap->a_vp, ap->a_m, ap->a_count,
726 ap->a_reqpage, ap->a_iodone, ap->a_arg));
730 vnode_pager_local_getpages0(struct vnode *vp, vm_page_t *m, int bytecount,
731 int reqpage, vop_getpages_iodone_t iodone, void *arg)
738 * Since the caller has busied the requested page, that page's valid
739 * field will not be changed by other threads.
741 vm_page_assert_xbusied(mreq);
744 * The requested page has valid blocks. Invalid part can only
745 * exist at the end of file, and the page is made fully valid
746 * by zeroing in vm_pager_get_pages(). Free non-requested
747 * pages, since no i/o is done to read its content.
749 if (mreq->valid != 0) {
750 vm_pager_free_nonreq(mreq->object, m, reqpage,
751 round_page(bytecount) / PAGE_SIZE, FALSE);
753 iodone(arg, m, reqpage, 0);
754 return (VM_PAGER_OK);
757 return (vnode_pager_generic_getpages(vp, m, bytecount, reqpage,
762 * This is now called from local media FS's to operate against their
763 * own vnodes if they fail to implement VOP_GETPAGES.
766 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int bytecount,
767 int reqpage, vop_getpages_iodone_t iodone, void *arg)
772 daddr_t firstaddr, reqblock;
774 int pbefore, pafter, i, size, bsize, first, last, *freecnt;
775 int count, error, before, after, secmask;
777 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
778 ("vnode_pager_generic_getpages does not support devices"));
779 if (vp->v_iflag & VI_DOOMED)
780 return (VM_PAGER_BAD);
782 object = vp->v_object;
783 count = bytecount / PAGE_SIZE;
784 bsize = vp->v_mount->mnt_stat.f_iosize;
787 * Synchronous and asynchronous paging operations use different
788 * free pbuf counters. This is done to avoid asynchronous requests
789 * to consume all pbufs.
790 * Allocate the pbuf at the very beginning of the function, so that
791 * if we are low on certain kind of pbufs don't even proceed to BMAP,
794 freecnt = iodone != NULL ?
795 &vnode_async_pbuf_freecnt : &vnode_pbuf_freecnt;
796 bp = getpbuf(freecnt);
799 * Get the underlying device blocks for the file with VOP_BMAP().
800 * If the file system doesn't support VOP_BMAP, use old way of
801 * getting pages via VOP_READ.
803 error = VOP_BMAP(vp, IDX_TO_OFF(m[reqpage]->pindex) / bsize, &bo,
804 &reqblock, &after, &before);
805 if (error == EOPNOTSUPP) {
806 relpbuf(bp, freecnt);
807 VM_OBJECT_WLOCK(object);
808 for (i = 0; i < count; i++)
812 vm_page_unlock(m[i]);
814 PCPU_INC(cnt.v_vnodein);
815 PCPU_INC(cnt.v_vnodepgsin);
816 error = vnode_pager_input_old(object, m[reqpage]);
817 VM_OBJECT_WUNLOCK(object);
819 } else if (error != 0) {
820 relpbuf(bp, freecnt);
821 vm_pager_free_nonreq(object, m, reqpage, count, FALSE);
822 return (VM_PAGER_ERROR);
825 * if the blocksize is smaller than a page size, then use
826 * special small filesystem code. NFS sometimes has a small
827 * blocksize, but it can handle large reads itself.
829 } else if ((PAGE_SIZE / bsize) > 1 &&
830 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
831 relpbuf(bp, freecnt);
832 vm_pager_free_nonreq(object, m, reqpage, count, FALSE);
833 PCPU_INC(cnt.v_vnodein);
834 PCPU_INC(cnt.v_vnodepgsin);
835 return (vnode_pager_input_smlfs(object, m[reqpage]));
839 * Since the caller has busied the requested page, that page's valid
840 * field will not be changed by other threads.
842 vm_page_assert_xbusied(m[reqpage]);
845 * If we have a completely valid page available to us, we can
846 * clean up and return. Otherwise we have to re-read the
849 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
850 relpbuf(bp, freecnt);
851 vm_pager_free_nonreq(object, m, reqpage, count, FALSE);
852 return (VM_PAGER_OK);
853 } else if (reqblock == -1) {
854 relpbuf(bp, freecnt);
855 pmap_zero_page(m[reqpage]);
856 KASSERT(m[reqpage]->dirty == 0,
857 ("vnode_pager_generic_getpages: page %p is dirty", m));
858 VM_OBJECT_WLOCK(object);
859 m[reqpage]->valid = VM_PAGE_BITS_ALL;
860 vm_pager_free_nonreq(object, m, reqpage, count, TRUE);
861 VM_OBJECT_WUNLOCK(object);
862 return (VM_PAGER_OK);
863 } else if (m[reqpage]->valid != 0) {
864 VM_OBJECT_WLOCK(object);
865 m[reqpage]->valid = 0;
866 VM_OBJECT_WUNLOCK(object);
869 pbefore = (daddr_t)before * bsize / PAGE_SIZE;
870 pafter = (daddr_t)after * bsize / PAGE_SIZE;
871 first = reqpage < pbefore ? 0 : reqpage - pbefore;
872 last = reqpage + pafter >= count ? count - 1 : reqpage + pafter;
873 if (first > 0 || last + 1 < count) {
874 VM_OBJECT_WLOCK(object);
875 for (i = 0; i < first; i++) {
878 vm_page_unlock(m[i]);
880 for (i = last + 1; i < count; i++) {
883 vm_page_unlock(m[i]);
885 VM_OBJECT_WUNLOCK(object);
889 * here on direct device I/O
891 firstaddr = reqblock;
892 firstaddr += (IDX_TO_OFF(m[reqpage]->pindex) % bsize) / DEV_BSIZE;
893 firstaddr -= IDX_TO_OFF(reqpage - first) / DEV_BSIZE;
896 * The first and last page have been calculated now, move
897 * input pages to be zero based, and adjust the count.
901 count = last - first + 1;
904 * calculate the file virtual address for the transfer
906 foff = IDX_TO_OFF(m[0]->pindex);
909 * calculate the size of the transfer
911 size = count * PAGE_SIZE;
912 KASSERT(count > 0, ("zero count"));
913 if ((foff + size) > object->un_pager.vnp.vnp_size)
914 size = object->un_pager.vnp.vnp_size - foff;
915 KASSERT(size > 0, ("zero size"));
918 * round up physical size for real devices.
920 secmask = bo->bo_bsize - 1;
921 KASSERT(secmask < PAGE_SIZE && secmask > 0,
922 ("vnode_pager_generic_getpages: sector size %d too large",
924 size = (size + secmask) & ~secmask;
927 * and map the pages to be read into the kva, if the filesystem
928 * requires mapped buffers.
930 if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
931 unmapped_buf_allowed) {
932 bp->b_data = unmapped_buf;
935 bp->b_data = bp->b_kvabase;
936 pmap_qenter((vm_offset_t)bp->b_data, m, count);
939 /* build a minimal buffer header */
940 bp->b_iocmd = BIO_READ;
941 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
942 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
943 bp->b_rcred = crhold(curthread->td_ucred);
944 bp->b_wcred = crhold(curthread->td_ucred);
945 bp->b_blkno = firstaddr;
949 bp->b_bufsize = size;
950 bp->b_runningbufspace = bp->b_bufsize;
951 for (i = 0; i < count; i++)
952 bp->b_pages[i] = m[i];
953 bp->b_npages = count;
954 bp->b_pager.pg_reqpage = reqpage;
955 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
957 PCPU_INC(cnt.v_vnodein);
958 PCPU_ADD(cnt.v_vnodepgsin, count);
961 bp->b_iooffset = dbtob(bp->b_blkno);
963 if (iodone != NULL) { /* async */
964 bp->b_pager.pg_iodone = iodone;
966 bp->b_iodone = vnode_pager_generic_getpages_done_async;
967 bp->b_flags |= B_ASYNC;
972 bp->b_iodone = bdone;
974 bwait(bp, PVM, "vnread");
975 error = vnode_pager_generic_getpages_done(bp);
976 for (i = 0; i < bp->b_npages; i++)
977 bp->b_pages[i] = NULL;
980 relpbuf(bp, &vnode_pbuf_freecnt);
983 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
987 vnode_pager_generic_getpages_done_async(struct buf *bp)
991 error = vnode_pager_generic_getpages_done(bp);
992 bp->b_pager.pg_iodone(bp->b_caller1, bp->b_pages,
993 bp->b_pager.pg_reqpage, error);
994 for (int i = 0; i < bp->b_npages; i++)
995 bp->b_pages[i] = NULL;
998 relpbuf(bp, &vnode_async_pbuf_freecnt);
1002 vnode_pager_generic_getpages_done(struct buf *bp)
1005 off_t tfoff, nextoff;
1008 error = (bp->b_ioflags & BIO_ERROR) != 0 ? EIO : 0;
1009 object = bp->b_vp->v_object;
1011 if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1012 if (!buf_mapped(bp)) {
1013 bp->b_data = bp->b_kvabase;
1014 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages,
1017 bzero(bp->b_data + bp->b_bcount,
1018 PAGE_SIZE * bp->b_npages - bp->b_bcount);
1020 if (buf_mapped(bp)) {
1021 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1022 bp->b_data = unmapped_buf;
1025 VM_OBJECT_WLOCK(object);
1026 for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1027 i < bp->b_npages; i++, tfoff = nextoff) {
1030 nextoff = tfoff + PAGE_SIZE;
1031 mt = bp->b_pages[i];
1033 if (nextoff <= object->un_pager.vnp.vnp_size) {
1035 * Read filled up entire page.
1037 mt->valid = VM_PAGE_BITS_ALL;
1038 KASSERT(mt->dirty == 0,
1039 ("%s: page %p is dirty", __func__, mt));
1040 KASSERT(!pmap_page_is_mapped(mt),
1041 ("%s: page %p is mapped", __func__, mt));
1044 * Read did not fill up entire page.
1046 * Currently we do not set the entire page valid,
1047 * we just try to clear the piece that we couldn't
1050 vm_page_set_valid_range(mt, 0,
1051 object->un_pager.vnp.vnp_size - tfoff);
1052 KASSERT((mt->dirty & vm_page_bits(0,
1053 object->un_pager.vnp.vnp_size - tfoff)) == 0,
1054 ("%s: page %p is dirty", __func__, mt));
1057 if (i != bp->b_pager.pg_reqpage)
1058 vm_page_readahead_finish(mt);
1060 VM_OBJECT_WUNLOCK(object);
1062 printf("%s: I/O read error %d\n", __func__, error);
1068 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
1069 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1070 * vnode_pager_generic_putpages() to implement the previous behaviour.
1072 * All other FS's should use the bypass to get to the local media
1073 * backing vp's VOP_PUTPAGES.
1076 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1077 int flags, int *rtvals)
1081 int bytes = count * PAGE_SIZE;
1084 * Force synchronous operation if we are extremely low on memory
1085 * to prevent a low-memory deadlock. VOP operations often need to
1086 * allocate more memory to initiate the I/O ( i.e. do a BMAP
1087 * operation ). The swapper handles the case by limiting the amount
1088 * of asynchronous I/O, but that sort of solution doesn't scale well
1089 * for the vnode pager without a lot of work.
1091 * Also, the backing vnode's iodone routine may not wake the pageout
1092 * daemon up. This should be probably be addressed XXX.
1095 if (vm_cnt.v_free_count + vm_cnt.v_cache_count <
1096 vm_cnt.v_pageout_free_min)
1097 flags |= VM_PAGER_PUT_SYNC;
1100 * Call device-specific putpages function
1102 vp = object->handle;
1103 VM_OBJECT_WUNLOCK(object);
1104 rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1105 KASSERT(rtval != EOPNOTSUPP,
1106 ("vnode_pager: stale FS putpages\n"));
1107 VM_OBJECT_WLOCK(object);
1112 * This is now called from local media FS's to operate against their
1113 * own vnodes if they fail to implement VOP_PUTPAGES.
1115 * This is typically called indirectly via the pageout daemon and
1116 * clustering has already typically occured, so in general we ask the
1117 * underlying filesystem to write the data out asynchronously rather
1121 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1122 int flags, int *rtvals)
1129 int maxsize, ncount;
1130 vm_ooffset_t poffset;
1136 static struct timeval lastfail;
1139 object = vp->v_object;
1140 count = bytecount / PAGE_SIZE;
1142 for (i = 0; i < count; i++)
1143 rtvals[i] = VM_PAGER_ERROR;
1145 if ((int64_t)ma[0]->pindex < 0) {
1146 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1147 (long)ma[0]->pindex, (u_long)ma[0]->dirty);
1148 rtvals[0] = VM_PAGER_BAD;
1149 return VM_PAGER_BAD;
1152 maxsize = count * PAGE_SIZE;
1155 poffset = IDX_TO_OFF(ma[0]->pindex);
1158 * If the page-aligned write is larger then the actual file we
1159 * have to invalidate pages occuring beyond the file EOF. However,
1160 * there is an edge case where a file may not be page-aligned where
1161 * the last page is partially invalid. In this case the filesystem
1162 * may not properly clear the dirty bits for the entire page (which
1163 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1164 * With the page locked we are free to fix-up the dirty bits here.
1166 * We do not under any circumstances truncate the valid bits, as
1167 * this will screw up bogus page replacement.
1169 VM_OBJECT_WLOCK(object);
1170 if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1171 if (object->un_pager.vnp.vnp_size > poffset) {
1174 maxsize = object->un_pager.vnp.vnp_size - poffset;
1175 ncount = btoc(maxsize);
1176 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1178 * If the object is locked and the following
1179 * conditions hold, then the page's dirty
1180 * field cannot be concurrently changed by a
1184 vm_page_assert_sbusied(m);
1185 KASSERT(!pmap_page_is_write_mapped(m),
1186 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1187 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1194 if (ncount < count) {
1195 for (i = ncount; i < count; i++) {
1196 rtvals[i] = VM_PAGER_BAD;
1200 VM_OBJECT_WUNLOCK(object);
1203 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
1204 * rather then a bdwrite() to prevent paging I/O from saturating
1205 * the buffer cache. Dummy-up the sequential heuristic to cause
1206 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
1207 * the system decides how to cluster.
1210 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1212 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1213 ioflags |= IO_ASYNC;
1214 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1215 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1217 aiov.iov_base = (caddr_t) 0;
1218 aiov.iov_len = maxsize;
1219 auio.uio_iov = &aiov;
1220 auio.uio_iovcnt = 1;
1221 auio.uio_offset = poffset;
1222 auio.uio_segflg = UIO_NOCOPY;
1223 auio.uio_rw = UIO_WRITE;
1224 auio.uio_resid = maxsize;
1225 auio.uio_td = (struct thread *) 0;
1226 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1227 PCPU_INC(cnt.v_vnodeout);
1228 PCPU_ADD(cnt.v_vnodepgsout, ncount);
1231 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1)))
1232 printf("vnode_pager_putpages: I/O error %d\n", error);
1234 if (auio.uio_resid) {
1235 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1))
1236 printf("vnode_pager_putpages: residual I/O %zd at %lu\n",
1237 auio.uio_resid, (u_long)ma[0]->pindex);
1239 for (i = 0; i < ncount; i++) {
1240 rtvals[i] = VM_PAGER_OK;
1246 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written)
1253 obj = ma[0]->object;
1254 VM_OBJECT_WLOCK(obj);
1255 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1256 if (pos < trunc_page(written)) {
1257 rtvals[i] = VM_PAGER_OK;
1258 vm_page_undirty(ma[i]);
1260 /* Partially written page. */
1261 rtvals[i] = VM_PAGER_AGAIN;
1262 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1265 VM_OBJECT_WUNLOCK(obj);
1269 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1273 vm_ooffset_t old_wm;
1275 VM_OBJECT_WLOCK(object);
1276 if (object->type != OBJT_VNODE) {
1277 VM_OBJECT_WUNLOCK(object);
1280 old_wm = object->un_pager.vnp.writemappings;
1281 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1282 vp = object->handle;
1283 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1284 ASSERT_VOP_ELOCKED(vp, "v_writecount inc");
1285 VOP_ADD_WRITECOUNT(vp, 1);
1286 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1287 __func__, vp, vp->v_writecount);
1288 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1289 ASSERT_VOP_ELOCKED(vp, "v_writecount dec");
1290 VOP_ADD_WRITECOUNT(vp, -1);
1291 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1292 __func__, vp, vp->v_writecount);
1294 VM_OBJECT_WUNLOCK(object);
1298 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1305 VM_OBJECT_WLOCK(object);
1308 * First, recheck the object type to account for the race when
1309 * the vnode is reclaimed.
1311 if (object->type != OBJT_VNODE) {
1312 VM_OBJECT_WUNLOCK(object);
1317 * Optimize for the case when writemappings is not going to
1321 if (object->un_pager.vnp.writemappings != inc) {
1322 object->un_pager.vnp.writemappings -= inc;
1323 VM_OBJECT_WUNLOCK(object);
1327 vp = object->handle;
1329 VM_OBJECT_WUNLOCK(object);
1331 vn_start_write(vp, &mp, V_WAIT);
1332 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1335 * Decrement the object's writemappings, by swapping the start
1336 * and end arguments for vnode_pager_update_writecount(). If
1337 * there was not a race with vnode reclaimation, then the
1338 * vnode's v_writecount is decremented.
1340 vnode_pager_update_writecount(object, end, start);
1344 vn_finished_write(mp);