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.
828 } else if ((PAGE_SIZE / bsize) > 1) {
829 relpbuf(bp, freecnt);
830 vm_pager_free_nonreq(object, m, reqpage, count, FALSE);
831 PCPU_INC(cnt.v_vnodein);
832 PCPU_INC(cnt.v_vnodepgsin);
833 return (vnode_pager_input_smlfs(object, m[reqpage]));
837 * Since the caller has busied the requested page, that page's valid
838 * field will not be changed by other threads.
840 vm_page_assert_xbusied(m[reqpage]);
843 * If we have a completely valid page available to us, we can
844 * clean up and return. Otherwise we have to re-read the
847 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
848 relpbuf(bp, freecnt);
849 vm_pager_free_nonreq(object, m, reqpage, count, FALSE);
850 return (VM_PAGER_OK);
851 } else if (reqblock == -1) {
852 relpbuf(bp, freecnt);
853 pmap_zero_page(m[reqpage]);
854 KASSERT(m[reqpage]->dirty == 0,
855 ("vnode_pager_generic_getpages: page %p is dirty", m));
856 VM_OBJECT_WLOCK(object);
857 m[reqpage]->valid = VM_PAGE_BITS_ALL;
858 vm_pager_free_nonreq(object, m, reqpage, count, TRUE);
859 VM_OBJECT_WUNLOCK(object);
860 return (VM_PAGER_OK);
861 } else if (m[reqpage]->valid != 0) {
862 VM_OBJECT_WLOCK(object);
863 m[reqpage]->valid = 0;
864 VM_OBJECT_WUNLOCK(object);
867 pib = IDX_TO_OFF(m[reqpage]->pindex) % bsize;
868 pbefore = ((daddr_t)before * bsize + pib) / PAGE_SIZE;
869 pafter = ((daddr_t)(after + 1) * bsize - pib) / PAGE_SIZE - 1;
870 first = reqpage < pbefore ? 0 : reqpage - pbefore;
871 last = reqpage + pafter >= count ? count - 1 : reqpage + pafter;
872 if (first > 0 || last + 1 < count) {
873 VM_OBJECT_WLOCK(object);
874 for (i = 0; i < first; i++) {
877 vm_page_unlock(m[i]);
879 for (i = last + 1; i < count; i++) {
882 vm_page_unlock(m[i]);
884 VM_OBJECT_WUNLOCK(object);
888 * here on direct device I/O
890 firstaddr = reqblock;
891 firstaddr += pib / DEV_BSIZE;
892 firstaddr -= IDX_TO_OFF(reqpage - first) / DEV_BSIZE;
895 * The first and last page have been calculated now, move
896 * input pages to be zero based, and adjust the count.
900 count = last - first + 1;
903 * calculate the file virtual address for the transfer
905 foff = IDX_TO_OFF(m[0]->pindex);
908 * calculate the size of the transfer
910 size = count * PAGE_SIZE;
911 KASSERT(count > 0, ("zero count"));
912 if ((foff + size) > object->un_pager.vnp.vnp_size)
913 size = object->un_pager.vnp.vnp_size - foff;
914 KASSERT(size > 0, ("zero size"));
917 * round up physical size for real devices.
919 secmask = bo->bo_bsize - 1;
920 KASSERT(secmask < PAGE_SIZE && secmask > 0,
921 ("vnode_pager_generic_getpages: sector size %d too large",
923 size = (size + secmask) & ~secmask;
926 * and map the pages to be read into the kva, if the filesystem
927 * requires mapped buffers.
929 if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
930 unmapped_buf_allowed) {
931 bp->b_data = unmapped_buf;
934 bp->b_data = bp->b_kvabase;
935 pmap_qenter((vm_offset_t)bp->b_data, m, count);
938 /* build a minimal buffer header */
939 bp->b_iocmd = BIO_READ;
940 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
941 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
942 bp->b_rcred = crhold(curthread->td_ucred);
943 bp->b_wcred = crhold(curthread->td_ucred);
944 bp->b_blkno = firstaddr;
948 bp->b_bufsize = size;
949 bp->b_runningbufspace = bp->b_bufsize;
950 for (i = 0; i < count; i++)
951 bp->b_pages[i] = m[i];
952 bp->b_npages = count;
953 bp->b_pager.pg_reqpage = reqpage;
954 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
956 PCPU_INC(cnt.v_vnodein);
957 PCPU_ADD(cnt.v_vnodepgsin, count);
960 bp->b_iooffset = dbtob(bp->b_blkno);
962 if (iodone != NULL) { /* async */
963 bp->b_pager.pg_iodone = iodone;
965 bp->b_iodone = vnode_pager_generic_getpages_done_async;
966 bp->b_flags |= B_ASYNC;
971 bp->b_iodone = bdone;
973 bwait(bp, PVM, "vnread");
974 error = vnode_pager_generic_getpages_done(bp);
975 for (i = 0; i < bp->b_npages; i++)
976 bp->b_pages[i] = NULL;
979 relpbuf(bp, &vnode_pbuf_freecnt);
982 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
986 vnode_pager_generic_getpages_done_async(struct buf *bp)
990 error = vnode_pager_generic_getpages_done(bp);
991 bp->b_pager.pg_iodone(bp->b_caller1, bp->b_pages,
992 bp->b_pager.pg_reqpage, error);
993 for (int i = 0; i < bp->b_npages; i++)
994 bp->b_pages[i] = NULL;
997 relpbuf(bp, &vnode_async_pbuf_freecnt);
1001 vnode_pager_generic_getpages_done(struct buf *bp)
1004 off_t tfoff, nextoff;
1007 error = (bp->b_ioflags & BIO_ERROR) != 0 ? EIO : 0;
1008 object = bp->b_vp->v_object;
1010 if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1011 if (!buf_mapped(bp)) {
1012 bp->b_data = bp->b_kvabase;
1013 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages,
1016 bzero(bp->b_data + bp->b_bcount,
1017 PAGE_SIZE * bp->b_npages - bp->b_bcount);
1019 if (buf_mapped(bp)) {
1020 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1021 bp->b_data = unmapped_buf;
1024 VM_OBJECT_WLOCK(object);
1025 for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1026 i < bp->b_npages; i++, tfoff = nextoff) {
1029 nextoff = tfoff + PAGE_SIZE;
1030 mt = bp->b_pages[i];
1032 if (nextoff <= object->un_pager.vnp.vnp_size) {
1034 * Read filled up entire page.
1036 mt->valid = VM_PAGE_BITS_ALL;
1037 KASSERT(mt->dirty == 0,
1038 ("%s: page %p is dirty", __func__, mt));
1039 KASSERT(!pmap_page_is_mapped(mt),
1040 ("%s: page %p is mapped", __func__, mt));
1043 * Read did not fill up entire page.
1045 * Currently we do not set the entire page valid,
1046 * we just try to clear the piece that we couldn't
1049 vm_page_set_valid_range(mt, 0,
1050 object->un_pager.vnp.vnp_size - tfoff);
1051 KASSERT((mt->dirty & vm_page_bits(0,
1052 object->un_pager.vnp.vnp_size - tfoff)) == 0,
1053 ("%s: page %p is dirty", __func__, mt));
1056 if (i != bp->b_pager.pg_reqpage)
1057 vm_page_readahead_finish(mt);
1059 VM_OBJECT_WUNLOCK(object);
1061 printf("%s: I/O read error %d\n", __func__, error);
1067 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
1068 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1069 * vnode_pager_generic_putpages() to implement the previous behaviour.
1071 * All other FS's should use the bypass to get to the local media
1072 * backing vp's VOP_PUTPAGES.
1075 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1076 int flags, int *rtvals)
1080 int bytes = count * PAGE_SIZE;
1083 * Force synchronous operation if we are extremely low on memory
1084 * to prevent a low-memory deadlock. VOP operations often need to
1085 * allocate more memory to initiate the I/O ( i.e. do a BMAP
1086 * operation ). The swapper handles the case by limiting the amount
1087 * of asynchronous I/O, but that sort of solution doesn't scale well
1088 * for the vnode pager without a lot of work.
1090 * Also, the backing vnode's iodone routine may not wake the pageout
1091 * daemon up. This should be probably be addressed XXX.
1094 if (vm_cnt.v_free_count + vm_cnt.v_cache_count <
1095 vm_cnt.v_pageout_free_min)
1096 flags |= VM_PAGER_PUT_SYNC;
1099 * Call device-specific putpages function
1101 vp = object->handle;
1102 VM_OBJECT_WUNLOCK(object);
1103 rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1104 KASSERT(rtval != EOPNOTSUPP,
1105 ("vnode_pager: stale FS putpages\n"));
1106 VM_OBJECT_WLOCK(object);
1111 * This is now called from local media FS's to operate against their
1112 * own vnodes if they fail to implement VOP_PUTPAGES.
1114 * This is typically called indirectly via the pageout daemon and
1115 * clustering has already typically occured, so in general we ask the
1116 * underlying filesystem to write the data out asynchronously rather
1120 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1121 int flags, int *rtvals)
1128 int maxsize, ncount;
1129 vm_ooffset_t poffset;
1135 static struct timeval lastfail;
1138 object = vp->v_object;
1139 count = bytecount / PAGE_SIZE;
1141 for (i = 0; i < count; i++)
1142 rtvals[i] = VM_PAGER_ERROR;
1144 if ((int64_t)ma[0]->pindex < 0) {
1145 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1146 (long)ma[0]->pindex, (u_long)ma[0]->dirty);
1147 rtvals[0] = VM_PAGER_BAD;
1148 return VM_PAGER_BAD;
1151 maxsize = count * PAGE_SIZE;
1154 poffset = IDX_TO_OFF(ma[0]->pindex);
1157 * If the page-aligned write is larger then the actual file we
1158 * have to invalidate pages occuring beyond the file EOF. However,
1159 * there is an edge case where a file may not be page-aligned where
1160 * the last page is partially invalid. In this case the filesystem
1161 * may not properly clear the dirty bits for the entire page (which
1162 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1163 * With the page locked we are free to fix-up the dirty bits here.
1165 * We do not under any circumstances truncate the valid bits, as
1166 * this will screw up bogus page replacement.
1168 VM_OBJECT_WLOCK(object);
1169 if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1170 if (object->un_pager.vnp.vnp_size > poffset) {
1173 maxsize = object->un_pager.vnp.vnp_size - poffset;
1174 ncount = btoc(maxsize);
1175 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1177 * If the object is locked and the following
1178 * conditions hold, then the page's dirty
1179 * field cannot be concurrently changed by a
1183 vm_page_assert_sbusied(m);
1184 KASSERT(!pmap_page_is_write_mapped(m),
1185 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1186 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1193 if (ncount < count) {
1194 for (i = ncount; i < count; i++) {
1195 rtvals[i] = VM_PAGER_BAD;
1199 VM_OBJECT_WUNLOCK(object);
1202 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
1203 * rather then a bdwrite() to prevent paging I/O from saturating
1204 * the buffer cache. Dummy-up the sequential heuristic to cause
1205 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
1206 * the system decides how to cluster.
1209 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1211 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1212 ioflags |= IO_ASYNC;
1213 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1214 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1216 aiov.iov_base = (caddr_t) 0;
1217 aiov.iov_len = maxsize;
1218 auio.uio_iov = &aiov;
1219 auio.uio_iovcnt = 1;
1220 auio.uio_offset = poffset;
1221 auio.uio_segflg = UIO_NOCOPY;
1222 auio.uio_rw = UIO_WRITE;
1223 auio.uio_resid = maxsize;
1224 auio.uio_td = (struct thread *) 0;
1225 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1226 PCPU_INC(cnt.v_vnodeout);
1227 PCPU_ADD(cnt.v_vnodepgsout, ncount);
1230 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1)))
1231 printf("vnode_pager_putpages: I/O error %d\n", error);
1233 if (auio.uio_resid) {
1234 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1))
1235 printf("vnode_pager_putpages: residual I/O %zd at %lu\n",
1236 auio.uio_resid, (u_long)ma[0]->pindex);
1238 for (i = 0; i < ncount; i++) {
1239 rtvals[i] = VM_PAGER_OK;
1245 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written)
1252 obj = ma[0]->object;
1253 VM_OBJECT_WLOCK(obj);
1254 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1255 if (pos < trunc_page(written)) {
1256 rtvals[i] = VM_PAGER_OK;
1257 vm_page_undirty(ma[i]);
1259 /* Partially written page. */
1260 rtvals[i] = VM_PAGER_AGAIN;
1261 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1264 VM_OBJECT_WUNLOCK(obj);
1268 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1272 vm_ooffset_t old_wm;
1274 VM_OBJECT_WLOCK(object);
1275 if (object->type != OBJT_VNODE) {
1276 VM_OBJECT_WUNLOCK(object);
1279 old_wm = object->un_pager.vnp.writemappings;
1280 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1281 vp = object->handle;
1282 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1283 ASSERT_VOP_ELOCKED(vp, "v_writecount inc");
1284 VOP_ADD_WRITECOUNT(vp, 1);
1285 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1286 __func__, vp, vp->v_writecount);
1287 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1288 ASSERT_VOP_ELOCKED(vp, "v_writecount dec");
1289 VOP_ADD_WRITECOUNT(vp, -1);
1290 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1291 __func__, vp, vp->v_writecount);
1293 VM_OBJECT_WUNLOCK(object);
1297 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1304 VM_OBJECT_WLOCK(object);
1307 * First, recheck the object type to account for the race when
1308 * the vnode is reclaimed.
1310 if (object->type != OBJT_VNODE) {
1311 VM_OBJECT_WUNLOCK(object);
1316 * Optimize for the case when writemappings is not going to
1320 if (object->un_pager.vnp.writemappings != inc) {
1321 object->un_pager.vnp.writemappings -= inc;
1322 VM_OBJECT_WUNLOCK(object);
1326 vp = object->handle;
1328 VM_OBJECT_WUNLOCK(object);
1330 vn_start_write(vp, &mp, V_WAIT);
1331 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1334 * Decrement the object's writemappings, by swapping the start
1335 * and end arguments for vnode_pager_update_writecount(). If
1336 * there was not a race with vnode reclaimation, then the
1337 * vnode's v_writecount is decremented.
1339 vnode_pager_update_writecount(object, end, start);
1343 vn_finished_write(mp);