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_getpages(vm_object_t, vm_page_t *, int, int *, int *);
88 static int vnode_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
89 int *, vop_getpages_iodone_t, void *);
90 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *);
91 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
92 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t,
93 vm_ooffset_t, struct ucred *cred);
94 static int vnode_pager_generic_getpages_done(struct buf *);
95 static void vnode_pager_generic_getpages_done_async(struct buf *);
97 struct pagerops vnodepagerops = {
98 .pgo_alloc = vnode_pager_alloc,
99 .pgo_dealloc = vnode_pager_dealloc,
100 .pgo_getpages = vnode_pager_getpages,
101 .pgo_getpages_async = vnode_pager_getpages_async,
102 .pgo_putpages = vnode_pager_putpages,
103 .pgo_haspage = vnode_pager_haspage,
106 int vnode_pbuf_freecnt;
107 int vnode_async_pbuf_freecnt;
109 /* Create the VM system backing object for this vnode */
111 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
114 vm_ooffset_t size = isize;
117 if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE)
120 while ((object = vp->v_object) != NULL) {
121 VM_OBJECT_WLOCK(object);
122 if (!(object->flags & OBJ_DEAD)) {
123 VM_OBJECT_WUNLOCK(object);
127 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
128 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vodead", 0);
129 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
133 if (vn_isdisk(vp, NULL)) {
134 size = IDX_TO_OFF(INT_MAX);
136 if (VOP_GETATTR(vp, &va, td->td_ucred))
142 object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred);
144 * Dereference the reference we just created. This assumes
145 * that the object is associated with the vp.
147 VM_OBJECT_WLOCK(object);
149 VM_OBJECT_WUNLOCK(object);
152 KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
158 vnode_destroy_vobject(struct vnode *vp)
160 struct vm_object *obj;
165 ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
166 VM_OBJECT_WLOCK(obj);
167 if (obj->ref_count == 0) {
169 * don't double-terminate the object
171 if ((obj->flags & OBJ_DEAD) == 0)
172 vm_object_terminate(obj);
174 VM_OBJECT_WUNLOCK(obj);
177 * Woe to the process that tries to page now :-).
179 vm_pager_deallocate(obj);
180 VM_OBJECT_WUNLOCK(obj);
187 * Allocate (or lookup) pager for a vnode.
188 * Handle is a vnode pointer.
193 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
194 vm_ooffset_t offset, struct ucred *cred)
200 * Pageout to vnode, no can do yet.
205 vp = (struct vnode *) handle;
208 * If the object is being terminated, wait for it to
212 while ((object = vp->v_object) != NULL) {
213 VM_OBJECT_WLOCK(object);
214 if ((object->flags & OBJ_DEAD) == 0)
216 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
217 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vadead", 0);
220 KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference"));
222 if (object == NULL) {
224 * Add an object of the appropriate size
226 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
228 object->un_pager.vnp.vnp_size = size;
229 object->un_pager.vnp.writemappings = 0;
231 object->handle = handle;
233 if (vp->v_object != NULL) {
235 * Object has been created while we were sleeping
238 VM_OBJECT_WLOCK(object);
239 KASSERT(object->ref_count == 1,
240 ("leaked ref %p %d", object, object->ref_count));
241 object->type = OBJT_DEAD;
242 object->ref_count = 0;
243 VM_OBJECT_WUNLOCK(object);
244 vm_object_destroy(object);
247 vp->v_object = object;
251 #if VM_NRESERVLEVEL > 0
252 vm_object_color(object, 0);
254 VM_OBJECT_WUNLOCK(object);
261 * The object must be locked.
264 vnode_pager_dealloc(vm_object_t object)
271 panic("vnode_pager_dealloc: pager already dealloced");
273 VM_OBJECT_ASSERT_WLOCKED(object);
274 vm_object_pip_wait(object, "vnpdea");
275 refs = object->ref_count;
277 object->handle = NULL;
278 object->type = OBJT_DEAD;
279 if (object->flags & OBJ_DISCONNECTWNT) {
280 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
283 ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
284 if (object->un_pager.vnp.writemappings > 0) {
285 object->un_pager.vnp.writemappings = 0;
286 VOP_ADD_WRITECOUNT(vp, -1);
287 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
288 __func__, vp, vp->v_writecount);
292 VM_OBJECT_WUNLOCK(object);
295 VM_OBJECT_WLOCK(object);
299 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
302 struct vnode *vp = object->handle;
308 int pagesperblock, blocksperpage;
310 VM_OBJECT_ASSERT_WLOCKED(object);
312 * If no vp or vp is doomed or marked transparent to VM, we do not
315 if (vp == NULL || vp->v_iflag & VI_DOOMED)
318 * If the offset is beyond end of file we do
321 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
324 bsize = vp->v_mount->mnt_stat.f_iosize;
325 pagesperblock = bsize / PAGE_SIZE;
327 if (pagesperblock > 0) {
328 reqblock = pindex / pagesperblock;
330 blocksperpage = (PAGE_SIZE / bsize);
331 reqblock = pindex * blocksperpage;
333 VM_OBJECT_WUNLOCK(object);
334 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
335 VM_OBJECT_WLOCK(object);
340 if (pagesperblock > 0) {
341 poff = pindex - (reqblock * pagesperblock);
343 *before *= pagesperblock;
348 * The BMAP vop can report a partial block in the
349 * 'after', but must not report blocks after EOF.
350 * Assert the latter, and truncate 'after' in case
353 KASSERT((reqblock + *after) * pagesperblock <
354 roundup2(object->size, pagesperblock),
355 ("%s: reqblock %jd after %d size %ju", __func__,
356 (intmax_t )reqblock, *after,
357 (uintmax_t )object->size));
358 *after *= pagesperblock;
359 *after += pagesperblock - (poff + 1);
360 if (pindex + *after >= object->size)
361 *after = object->size - 1 - pindex;
365 *before /= blocksperpage;
369 *after /= blocksperpage;
376 * Lets the VM system know about a change in size for a file.
377 * We adjust our own internal size and flush any cached pages in
378 * the associated object that are affected by the size change.
380 * Note: this routine may be invoked as a result of a pager put
381 * operation (possibly at object termination time), so we must be careful.
384 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
388 vm_pindex_t nobjsize;
390 if ((object = vp->v_object) == NULL)
392 /* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */
393 VM_OBJECT_WLOCK(object);
394 if (object->type == OBJT_DEAD) {
395 VM_OBJECT_WUNLOCK(object);
398 KASSERT(object->type == OBJT_VNODE,
399 ("not vnode-backed object %p", object));
400 if (nsize == object->un_pager.vnp.vnp_size) {
402 * Hasn't changed size
404 VM_OBJECT_WUNLOCK(object);
407 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
408 if (nsize < object->un_pager.vnp.vnp_size) {
410 * File has shrunk. Toss any cached pages beyond the new EOF.
412 if (nobjsize < object->size)
413 vm_object_page_remove(object, nobjsize, object->size,
416 * this gets rid of garbage at the end of a page that is now
417 * only partially backed by the vnode.
419 * XXX for some reason (I don't know yet), if we take a
420 * completely invalid page and mark it partially valid
421 * it can screw up NFS reads, so we don't allow the case.
423 if ((nsize & PAGE_MASK) &&
424 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
426 int base = (int)nsize & PAGE_MASK;
427 int size = PAGE_SIZE - base;
430 * Clear out partial-page garbage in case
431 * the page has been mapped.
433 pmap_zero_page_area(m, base, size);
436 * Update the valid bits to reflect the blocks that
437 * have been zeroed. Some of these valid bits may
438 * have already been set.
440 vm_page_set_valid_range(m, base, size);
443 * Round "base" to the next block boundary so that the
444 * dirty bit for a partially zeroed block is not
447 base = roundup2(base, DEV_BSIZE);
450 * Clear out partial-page dirty bits.
452 * note that we do not clear out the valid
453 * bits. This would prevent bogus_page
454 * replacement from working properly.
456 vm_page_clear_dirty(m, base, PAGE_SIZE - base);
457 } else if ((nsize & PAGE_MASK) &&
458 vm_page_is_cached(object, OFF_TO_IDX(nsize))) {
459 vm_page_cache_free(object, OFF_TO_IDX(nsize),
463 object->un_pager.vnp.vnp_size = nsize;
464 object->size = nobjsize;
465 VM_OBJECT_WUNLOCK(object);
469 * calculate the linear (byte) disk address of specified virtual
473 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
484 if (vp->v_iflag & VI_DOOMED)
487 bsize = vp->v_mount->mnt_stat.f_iosize;
488 vblock = address / bsize;
489 voffset = address % bsize;
491 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
493 if (*rtaddress != -1)
494 *rtaddress += voffset / DEV_BSIZE;
497 *run *= bsize/PAGE_SIZE;
498 *run -= voffset/PAGE_SIZE;
506 * small block filesystem vnode pager input
509 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m)
522 if (vp->v_iflag & VI_DOOMED)
525 bsize = vp->v_mount->mnt_stat.f_iosize;
527 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
529 sf = sf_buf_alloc(m, 0);
531 for (i = 0; i < PAGE_SIZE / bsize; i++) {
532 vm_ooffset_t address;
534 bits = vm_page_bits(i * bsize, bsize);
538 address = IDX_TO_OFF(m->pindex) + i * bsize;
539 if (address >= object->un_pager.vnp.vnp_size) {
542 error = vnode_pager_addr(vp, address, &fileaddr, NULL);
546 if (fileaddr != -1) {
547 bp = getpbuf(&vnode_pbuf_freecnt);
549 /* build a minimal buffer header */
550 bp->b_iocmd = BIO_READ;
551 bp->b_iodone = bdone;
552 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
553 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
554 bp->b_rcred = crhold(curthread->td_ucred);
555 bp->b_wcred = crhold(curthread->td_ucred);
556 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
557 bp->b_blkno = fileaddr;
560 bp->b_bcount = bsize;
561 bp->b_bufsize = bsize;
562 bp->b_runningbufspace = bp->b_bufsize;
563 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
566 bp->b_iooffset = dbtob(bp->b_blkno);
569 bwait(bp, PVM, "vnsrd");
571 if ((bp->b_ioflags & BIO_ERROR) != 0)
575 * free the buffer header back to the swap buffer pool
579 relpbuf(bp, &vnode_pbuf_freecnt);
583 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
584 KASSERT((m->dirty & bits) == 0,
585 ("vnode_pager_input_smlfs: page %p is dirty", m));
586 VM_OBJECT_WLOCK(object);
588 VM_OBJECT_WUNLOCK(object);
592 return VM_PAGER_ERROR;
598 * old style vnode pager input routine
601 vnode_pager_input_old(vm_object_t object, vm_page_t m)
610 VM_OBJECT_ASSERT_WLOCKED(object);
614 * Return failure if beyond current EOF
616 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
620 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
621 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
623 VM_OBJECT_WUNLOCK(object);
626 * Allocate a kernel virtual address and initialize so that
627 * we can use VOP_READ/WRITE routines.
629 sf = sf_buf_alloc(m, 0);
631 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
633 auio.uio_iov = &aiov;
635 auio.uio_offset = IDX_TO_OFF(m->pindex);
636 auio.uio_segflg = UIO_SYSSPACE;
637 auio.uio_rw = UIO_READ;
638 auio.uio_resid = size;
639 auio.uio_td = curthread;
641 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
643 int count = size - auio.uio_resid;
647 else if (count != PAGE_SIZE)
648 bzero((caddr_t)sf_buf_kva(sf) + count,
653 VM_OBJECT_WLOCK(object);
655 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
657 m->valid = VM_PAGE_BITS_ALL;
658 return error ? VM_PAGER_ERROR : VM_PAGER_OK;
662 * generic vnode pager input routine
666 * Local media VFS's that do not implement their own VOP_GETPAGES
667 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
668 * to implement the previous behaviour.
670 * All other FS's should use the bypass to get to the local media
671 * backing vp's VOP_GETPAGES.
674 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
681 VM_OBJECT_WUNLOCK(object);
682 rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead);
683 KASSERT(rtval != EOPNOTSUPP,
684 ("vnode_pager: FS getpages not implemented\n"));
685 VM_OBJECT_WLOCK(object);
690 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
691 int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg)
697 VM_OBJECT_WUNLOCK(object);
698 rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg);
699 KASSERT(rtval != EOPNOTSUPP,
700 ("vnode_pager: FS getpages_async not implemented\n"));
701 VM_OBJECT_WLOCK(object);
706 * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for
707 * local filesystems, where partially valid pages can only occur at
711 vnode_pager_local_getpages(struct vop_getpages_args *ap)
714 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
715 ap->a_rbehind, ap->a_rahead, NULL, NULL));
719 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap)
722 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
723 ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg));
727 * This is now called from local media FS's to operate against their
728 * own vnodes if they fail to implement VOP_GETPAGES.
731 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count,
732 int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg)
738 int bsize, pagesperblock, *freecnt;
739 int error, before, after, rbehind, rahead, poff, i;
740 int bytecount, secmask;
742 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
743 ("%s does not support devices", __func__));
745 if (vp->v_iflag & VI_DOOMED)
746 return (VM_PAGER_BAD);
748 object = vp->v_object;
749 foff = IDX_TO_OFF(m[0]->pindex);
750 bsize = vp->v_mount->mnt_stat.f_iosize;
751 pagesperblock = bsize / PAGE_SIZE;
753 KASSERT(foff < object->un_pager.vnp.vnp_size,
754 ("%s: page %p offset beyond vp %p size", __func__, m[0], vp));
755 KASSERT(count <= sizeof(bp->b_pages),
756 ("%s: requested %d pages", __func__, count));
759 * The last page has valid blocks. Invalid part can only
760 * exist at the end of file, and the page is made fully valid
761 * by zeroing in vm_pager_get_pages().
763 if (m[count - 1]->valid != 0 && --count == 0) {
765 iodone(arg, m, 1, 0);
766 return (VM_PAGER_OK);
770 * Synchronous and asynchronous paging operations use different
771 * free pbuf counters. This is done to avoid asynchronous requests
772 * to consume all pbufs.
773 * Allocate the pbuf at the very beginning of the function, so that
774 * if we are low on certain kind of pbufs don't even proceed to BMAP,
777 freecnt = iodone != NULL ?
778 &vnode_async_pbuf_freecnt : &vnode_pbuf_freecnt;
779 bp = getpbuf(freecnt);
782 * Get the underlying device blocks for the file with VOP_BMAP().
783 * If the file system doesn't support VOP_BMAP, use old way of
784 * getting pages via VOP_READ.
786 error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before);
787 if (error == EOPNOTSUPP) {
788 relpbuf(bp, freecnt);
789 VM_OBJECT_WLOCK(object);
790 for (i = 0; i < count; i++) {
791 PCPU_INC(cnt.v_vnodein);
792 PCPU_INC(cnt.v_vnodepgsin);
793 error = vnode_pager_input_old(object, m[i]);
797 VM_OBJECT_WUNLOCK(object);
799 } else if (error != 0) {
800 relpbuf(bp, freecnt);
801 return (VM_PAGER_ERROR);
805 * If the file system supports BMAP, but blocksize is smaller
806 * than a page size, then use special small filesystem code.
808 if (pagesperblock == 0) {
809 relpbuf(bp, freecnt);
810 for (i = 0; i < count; i++) {
811 PCPU_INC(cnt.v_vnodein);
812 PCPU_INC(cnt.v_vnodepgsin);
813 error = vnode_pager_input_smlfs(object, m[i]);
821 * A sparse file can be encountered only for a single page request,
822 * which may not be preceeded by call to vm_pager_haspage().
824 if (bp->b_blkno == -1) {
826 ("%s: array[%d] request to a sparse file %p", __func__,
828 relpbuf(bp, freecnt);
829 pmap_zero_page(m[0]);
830 KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty",
832 VM_OBJECT_WLOCK(object);
833 m[0]->valid = VM_PAGE_BITS_ALL;
834 VM_OBJECT_WUNLOCK(object);
835 return (VM_PAGER_OK);
838 bp->b_blkno += (foff % bsize) / DEV_BSIZE;
840 /* Recalculate blocks available after/before to pages. */
841 poff = (foff % bsize) / PAGE_SIZE;
842 before *= pagesperblock;
844 after *= pagesperblock;
845 after += pagesperblock - (poff + 1);
846 if (m[0]->pindex + after >= object->size)
847 after = object->size - 1 - m[0]->pindex;
848 KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d",
849 __func__, count, after + 1));
852 /* Trim requested rbehind/rahead to possible values. */
853 rbehind = a_rbehind ? *a_rbehind : 0;
854 rahead = a_rahead ? *a_rahead : 0;
855 rbehind = min(rbehind, before);
856 rbehind = min(rbehind, m[0]->pindex);
857 rahead = min(rahead, after);
858 rahead = min(rahead, object->size - m[count - 1]->pindex);
859 KASSERT(rbehind + rahead + count <= sizeof(bp->b_pages),
860 ("%s: behind %d ahead %d count %d", __func__,
861 rbehind, rahead, count));
864 * Fill in the bp->b_pages[] array with requested and optional
865 * read behind or read ahead pages. Read behind pages are looked
866 * up in a backward direction, down to a first cached page. Same
867 * for read ahead pages, but there is no need to shift the array
868 * in case of encountering a cached page.
870 i = bp->b_npages = 0;
872 vm_pindex_t startpindex, tpindex;
875 VM_OBJECT_WLOCK(object);
876 startpindex = m[0]->pindex - rbehind;
877 if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL &&
878 p->pindex >= startpindex)
879 startpindex = p->pindex + 1;
881 /* tpindex is unsigned; beware of numeric underflow. */
882 for (tpindex = m[0]->pindex - 1;
883 tpindex >= startpindex && tpindex < m[0]->pindex;
885 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
886 VM_ALLOC_IFNOTCACHED);
888 /* Shift the array. */
889 for (int j = 0; j < i; j++)
890 bp->b_pages[j] = bp->b_pages[j +
891 tpindex + 1 - startpindex];
894 bp->b_pages[tpindex - startpindex] = p;
899 bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE;
903 /* Requested pages. */
904 for (int j = 0; j < count; j++, i++)
905 bp->b_pages[i] = m[j];
906 bp->b_npages += count;
909 vm_pindex_t endpindex, tpindex;
912 if (!VM_OBJECT_WOWNED(object))
913 VM_OBJECT_WLOCK(object);
914 endpindex = m[count - 1]->pindex + rahead + 1;
915 if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL &&
916 p->pindex < endpindex)
917 endpindex = p->pindex;
918 if (endpindex > object->size)
919 endpindex = object->size;
921 for (tpindex = m[count - 1]->pindex + 1;
922 tpindex < endpindex; i++, tpindex++) {
923 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
924 VM_ALLOC_IFNOTCACHED);
930 bp->b_pgafter = i - bp->b_npages;
935 if (VM_OBJECT_WOWNED(object))
936 VM_OBJECT_WUNLOCK(object);
938 /* Report back actual behind/ahead read. */
940 *a_rbehind = bp->b_pgbefore;
942 *a_rahead = bp->b_pgafter;
944 KASSERT(bp->b_npages <= sizeof(bp->b_pages),
945 ("%s: buf %p overflowed", __func__, bp));
948 * Recalculate first offset and bytecount with regards to read behind.
949 * Truncate bytecount to vnode real size and round up physical size
952 foff = IDX_TO_OFF(bp->b_pages[0]->pindex);
953 bytecount = bp->b_npages << PAGE_SHIFT;
954 if ((foff + bytecount) > object->un_pager.vnp.vnp_size)
955 bytecount = object->un_pager.vnp.vnp_size - foff;
956 secmask = bo->bo_bsize - 1;
957 KASSERT(secmask < PAGE_SIZE && secmask > 0,
958 ("%s: sector size %d too large", __func__, secmask + 1));
959 bytecount = (bytecount + secmask) & ~secmask;
962 * And map the pages to be read into the kva, if the filesystem
963 * requires mapped buffers.
965 if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
966 unmapped_buf_allowed) {
967 bp->b_data = unmapped_buf;
970 bp->b_data = bp->b_kvabase;
971 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
974 /* Build a minimal buffer header. */
975 bp->b_iocmd = BIO_READ;
976 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
977 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
978 bp->b_rcred = crhold(curthread->td_ucred);
979 bp->b_wcred = crhold(curthread->td_ucred);
982 bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount;
983 bp->b_iooffset = dbtob(bp->b_blkno);
985 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
986 PCPU_INC(cnt.v_vnodein);
987 PCPU_ADD(cnt.v_vnodepgsin, bp->b_npages);
989 if (iodone != NULL) { /* async */
990 bp->b_pgiodone = iodone;
992 bp->b_iodone = vnode_pager_generic_getpages_done_async;
993 bp->b_flags |= B_ASYNC;
996 return (VM_PAGER_OK);
998 bp->b_iodone = bdone;
1000 bwait(bp, PVM, "vnread");
1001 error = vnode_pager_generic_getpages_done(bp);
1002 for (i = 0; i < bp->b_npages; i++)
1003 bp->b_pages[i] = NULL;
1006 relpbuf(bp, &vnode_pbuf_freecnt);
1007 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
1012 vnode_pager_generic_getpages_done_async(struct buf *bp)
1016 error = vnode_pager_generic_getpages_done(bp);
1017 /* Run the iodone upon the requested range. */
1018 bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore,
1019 bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error);
1020 for (int i = 0; i < bp->b_npages; i++)
1021 bp->b_pages[i] = NULL;
1024 relpbuf(bp, &vnode_async_pbuf_freecnt);
1028 vnode_pager_generic_getpages_done(struct buf *bp)
1031 off_t tfoff, nextoff;
1034 error = (bp->b_ioflags & BIO_ERROR) != 0 ? EIO : 0;
1035 object = bp->b_vp->v_object;
1037 if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1038 if (!buf_mapped(bp)) {
1039 bp->b_data = bp->b_kvabase;
1040 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages,
1043 bzero(bp->b_data + bp->b_bcount,
1044 PAGE_SIZE * bp->b_npages - bp->b_bcount);
1046 if (buf_mapped(bp)) {
1047 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1048 bp->b_data = unmapped_buf;
1051 VM_OBJECT_WLOCK(object);
1052 for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1053 i < bp->b_npages; i++, tfoff = nextoff) {
1056 nextoff = tfoff + PAGE_SIZE;
1057 mt = bp->b_pages[i];
1059 if (nextoff <= object->un_pager.vnp.vnp_size) {
1061 * Read filled up entire page.
1063 mt->valid = VM_PAGE_BITS_ALL;
1064 KASSERT(mt->dirty == 0,
1065 ("%s: page %p is dirty", __func__, mt));
1066 KASSERT(!pmap_page_is_mapped(mt),
1067 ("%s: page %p is mapped", __func__, mt));
1070 * Read did not fill up entire page.
1072 * Currently we do not set the entire page valid,
1073 * we just try to clear the piece that we couldn't
1076 vm_page_set_valid_range(mt, 0,
1077 object->un_pager.vnp.vnp_size - tfoff);
1078 KASSERT((mt->dirty & vm_page_bits(0,
1079 object->un_pager.vnp.vnp_size - tfoff)) == 0,
1080 ("%s: page %p is dirty", __func__, mt));
1083 if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter)
1084 vm_page_readahead_finish(mt);
1086 VM_OBJECT_WUNLOCK(object);
1088 printf("%s: I/O read error %d\n", __func__, error);
1094 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
1095 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1096 * vnode_pager_generic_putpages() to implement the previous behaviour.
1098 * All other FS's should use the bypass to get to the local media
1099 * backing vp's VOP_PUTPAGES.
1102 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1103 int flags, int *rtvals)
1107 int bytes = count * PAGE_SIZE;
1110 * Force synchronous operation if we are extremely low on memory
1111 * to prevent a low-memory deadlock. VOP operations often need to
1112 * allocate more memory to initiate the I/O ( i.e. do a BMAP
1113 * operation ). The swapper handles the case by limiting the amount
1114 * of asynchronous I/O, but that sort of solution doesn't scale well
1115 * for the vnode pager without a lot of work.
1117 * Also, the backing vnode's iodone routine may not wake the pageout
1118 * daemon up. This should be probably be addressed XXX.
1121 if (vm_cnt.v_free_count + vm_cnt.v_cache_count <
1122 vm_cnt.v_pageout_free_min)
1123 flags |= VM_PAGER_PUT_SYNC;
1126 * Call device-specific putpages function
1128 vp = object->handle;
1129 VM_OBJECT_WUNLOCK(object);
1130 rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1131 KASSERT(rtval != EOPNOTSUPP,
1132 ("vnode_pager: stale FS putpages\n"));
1133 VM_OBJECT_WLOCK(object);
1138 * This is now called from local media FS's to operate against their
1139 * own vnodes if they fail to implement VOP_PUTPAGES.
1141 * This is typically called indirectly via the pageout daemon and
1142 * clustering has already typically occured, so in general we ask the
1143 * underlying filesystem to write the data out asynchronously rather
1147 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1148 int flags, int *rtvals)
1155 int maxsize, ncount;
1156 vm_ooffset_t poffset;
1162 static struct timeval lastfail;
1165 object = vp->v_object;
1166 count = bytecount / PAGE_SIZE;
1168 for (i = 0; i < count; i++)
1169 rtvals[i] = VM_PAGER_ERROR;
1171 if ((int64_t)ma[0]->pindex < 0) {
1172 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1173 (long)ma[0]->pindex, (u_long)ma[0]->dirty);
1174 rtvals[0] = VM_PAGER_BAD;
1175 return VM_PAGER_BAD;
1178 maxsize = count * PAGE_SIZE;
1181 poffset = IDX_TO_OFF(ma[0]->pindex);
1184 * If the page-aligned write is larger then the actual file we
1185 * have to invalidate pages occuring beyond the file EOF. However,
1186 * there is an edge case where a file may not be page-aligned where
1187 * the last page is partially invalid. In this case the filesystem
1188 * may not properly clear the dirty bits for the entire page (which
1189 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1190 * With the page locked we are free to fix-up the dirty bits here.
1192 * We do not under any circumstances truncate the valid bits, as
1193 * this will screw up bogus page replacement.
1195 VM_OBJECT_WLOCK(object);
1196 if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1197 if (object->un_pager.vnp.vnp_size > poffset) {
1200 maxsize = object->un_pager.vnp.vnp_size - poffset;
1201 ncount = btoc(maxsize);
1202 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1204 * If the object is locked and the following
1205 * conditions hold, then the page's dirty
1206 * field cannot be concurrently changed by a
1210 vm_page_assert_sbusied(m);
1211 KASSERT(!pmap_page_is_write_mapped(m),
1212 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1213 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1220 if (ncount < count) {
1221 for (i = ncount; i < count; i++) {
1222 rtvals[i] = VM_PAGER_BAD;
1226 VM_OBJECT_WUNLOCK(object);
1229 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
1230 * rather then a bdwrite() to prevent paging I/O from saturating
1231 * the buffer cache. Dummy-up the sequential heuristic to cause
1232 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
1233 * the system decides how to cluster.
1236 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1238 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1239 ioflags |= IO_ASYNC;
1240 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1241 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1243 aiov.iov_base = (caddr_t) 0;
1244 aiov.iov_len = maxsize;
1245 auio.uio_iov = &aiov;
1246 auio.uio_iovcnt = 1;
1247 auio.uio_offset = poffset;
1248 auio.uio_segflg = UIO_NOCOPY;
1249 auio.uio_rw = UIO_WRITE;
1250 auio.uio_resid = maxsize;
1251 auio.uio_td = (struct thread *) 0;
1252 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1253 PCPU_INC(cnt.v_vnodeout);
1254 PCPU_ADD(cnt.v_vnodepgsout, ncount);
1257 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1)))
1258 printf("vnode_pager_putpages: I/O error %d\n", error);
1260 if (auio.uio_resid) {
1261 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1))
1262 printf("vnode_pager_putpages: residual I/O %zd at %lu\n",
1263 auio.uio_resid, (u_long)ma[0]->pindex);
1265 for (i = 0; i < ncount; i++) {
1266 rtvals[i] = VM_PAGER_OK;
1272 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written)
1279 obj = ma[0]->object;
1280 VM_OBJECT_WLOCK(obj);
1281 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1282 if (pos < trunc_page(written)) {
1283 rtvals[i] = VM_PAGER_OK;
1284 vm_page_undirty(ma[i]);
1286 /* Partially written page. */
1287 rtvals[i] = VM_PAGER_AGAIN;
1288 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1291 VM_OBJECT_WUNLOCK(obj);
1295 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1299 vm_ooffset_t old_wm;
1301 VM_OBJECT_WLOCK(object);
1302 if (object->type != OBJT_VNODE) {
1303 VM_OBJECT_WUNLOCK(object);
1306 old_wm = object->un_pager.vnp.writemappings;
1307 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1308 vp = object->handle;
1309 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1310 ASSERT_VOP_ELOCKED(vp, "v_writecount inc");
1311 VOP_ADD_WRITECOUNT(vp, 1);
1312 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1313 __func__, vp, vp->v_writecount);
1314 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1315 ASSERT_VOP_ELOCKED(vp, "v_writecount dec");
1316 VOP_ADD_WRITECOUNT(vp, -1);
1317 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1318 __func__, vp, vp->v_writecount);
1320 VM_OBJECT_WUNLOCK(object);
1324 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1331 VM_OBJECT_WLOCK(object);
1334 * First, recheck the object type to account for the race when
1335 * the vnode is reclaimed.
1337 if (object->type != OBJT_VNODE) {
1338 VM_OBJECT_WUNLOCK(object);
1343 * Optimize for the case when writemappings is not going to
1347 if (object->un_pager.vnp.writemappings != inc) {
1348 object->un_pager.vnp.writemappings -= inc;
1349 VM_OBJECT_WUNLOCK(object);
1353 vp = object->handle;
1355 VM_OBJECT_WUNLOCK(object);
1357 vn_start_write(vp, &mp, V_WAIT);
1358 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1361 * Decrement the object's writemappings, by swapping the start
1362 * and end arguments for vnode_pager_update_writecount(). If
1363 * there was not a race with vnode reclaimation, then the
1364 * vnode's v_writecount is decremented.
1366 vnode_pager_update_writecount(object, end, start);
1370 vn_finished_write(mp);