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 umtx_shm_object_terminated(obj);
168 if (obj->ref_count == 0) {
170 * don't double-terminate the object
172 if ((obj->flags & OBJ_DEAD) == 0)
173 vm_object_terminate(obj);
175 VM_OBJECT_WUNLOCK(obj);
178 * Woe to the process that tries to page now :-).
180 vm_pager_deallocate(obj);
181 VM_OBJECT_WUNLOCK(obj);
188 * Allocate (or lookup) pager for a vnode.
189 * Handle is a vnode pointer.
194 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
195 vm_ooffset_t offset, struct ucred *cred)
201 * Pageout to vnode, no can do yet.
206 vp = (struct vnode *) handle;
209 * If the object is being terminated, wait for it to
213 while ((object = vp->v_object) != NULL) {
214 VM_OBJECT_WLOCK(object);
215 if ((object->flags & OBJ_DEAD) == 0)
217 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
218 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vadead", 0);
221 KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference"));
223 if (object == NULL) {
225 * Add an object of the appropriate size
227 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
229 object->un_pager.vnp.vnp_size = size;
230 object->un_pager.vnp.writemappings = 0;
232 object->handle = handle;
234 if (vp->v_object != NULL) {
236 * Object has been created while we were sleeping
239 VM_OBJECT_WLOCK(object);
240 KASSERT(object->ref_count == 1,
241 ("leaked ref %p %d", object, object->ref_count));
242 object->type = OBJT_DEAD;
243 object->ref_count = 0;
244 VM_OBJECT_WUNLOCK(object);
245 vm_object_destroy(object);
248 vp->v_object = object;
252 #if VM_NRESERVLEVEL > 0
253 vm_object_color(object, 0);
255 VM_OBJECT_WUNLOCK(object);
262 * The object must be locked.
265 vnode_pager_dealloc(vm_object_t object)
272 panic("vnode_pager_dealloc: pager already dealloced");
274 VM_OBJECT_ASSERT_WLOCKED(object);
275 vm_object_pip_wait(object, "vnpdea");
276 refs = object->ref_count;
278 object->handle = NULL;
279 object->type = OBJT_DEAD;
280 if (object->flags & OBJ_DISCONNECTWNT) {
281 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
284 ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
285 if (object->un_pager.vnp.writemappings > 0) {
286 object->un_pager.vnp.writemappings = 0;
287 VOP_ADD_WRITECOUNT(vp, -1);
288 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
289 __func__, vp, vp->v_writecount);
293 VM_OBJECT_WUNLOCK(object);
296 VM_OBJECT_WLOCK(object);
300 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
303 struct vnode *vp = object->handle;
309 int pagesperblock, blocksperpage;
311 VM_OBJECT_ASSERT_WLOCKED(object);
313 * If no vp or vp is doomed or marked transparent to VM, we do not
316 if (vp == NULL || vp->v_iflag & VI_DOOMED)
319 * If the offset is beyond end of file we do
322 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
325 bsize = vp->v_mount->mnt_stat.f_iosize;
326 pagesperblock = bsize / PAGE_SIZE;
328 if (pagesperblock > 0) {
329 reqblock = pindex / pagesperblock;
331 blocksperpage = (PAGE_SIZE / bsize);
332 reqblock = pindex * blocksperpage;
334 VM_OBJECT_WUNLOCK(object);
335 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
336 VM_OBJECT_WLOCK(object);
341 if (pagesperblock > 0) {
342 poff = pindex - (reqblock * pagesperblock);
344 *before *= pagesperblock;
349 * The BMAP vop can report a partial block in the
350 * 'after', but must not report blocks after EOF.
351 * Assert the latter, and truncate 'after' in case
354 KASSERT((reqblock + *after) * pagesperblock <
355 roundup2(object->size, pagesperblock),
356 ("%s: reqblock %jd after %d size %ju", __func__,
357 (intmax_t )reqblock, *after,
358 (uintmax_t )object->size));
359 *after *= pagesperblock;
360 *after += pagesperblock - (poff + 1);
361 if (pindex + *after >= object->size)
362 *after = object->size - 1 - pindex;
366 *before /= blocksperpage;
370 *after /= blocksperpage;
377 * Lets the VM system know about a change in size for a file.
378 * We adjust our own internal size and flush any cached pages in
379 * the associated object that are affected by the size change.
381 * Note: this routine may be invoked as a result of a pager put
382 * operation (possibly at object termination time), so we must be careful.
385 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
389 vm_pindex_t nobjsize;
391 if ((object = vp->v_object) == NULL)
393 /* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */
394 VM_OBJECT_WLOCK(object);
395 if (object->type == OBJT_DEAD) {
396 VM_OBJECT_WUNLOCK(object);
399 KASSERT(object->type == OBJT_VNODE,
400 ("not vnode-backed object %p", object));
401 if (nsize == object->un_pager.vnp.vnp_size) {
403 * Hasn't changed size
405 VM_OBJECT_WUNLOCK(object);
408 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
409 if (nsize < object->un_pager.vnp.vnp_size) {
411 * File has shrunk. Toss any cached pages beyond the new EOF.
413 if (nobjsize < object->size)
414 vm_object_page_remove(object, nobjsize, object->size,
417 * this gets rid of garbage at the end of a page that is now
418 * only partially backed by the vnode.
420 * XXX for some reason (I don't know yet), if we take a
421 * completely invalid page and mark it partially valid
422 * it can screw up NFS reads, so we don't allow the case.
424 if ((nsize & PAGE_MASK) &&
425 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
427 int base = (int)nsize & PAGE_MASK;
428 int size = PAGE_SIZE - base;
431 * Clear out partial-page garbage in case
432 * the page has been mapped.
434 pmap_zero_page_area(m, base, size);
437 * Update the valid bits to reflect the blocks that
438 * have been zeroed. Some of these valid bits may
439 * have already been set.
441 vm_page_set_valid_range(m, base, size);
444 * Round "base" to the next block boundary so that the
445 * dirty bit for a partially zeroed block is not
448 base = roundup2(base, DEV_BSIZE);
451 * Clear out partial-page dirty bits.
453 * note that we do not clear out the valid
454 * bits. This would prevent bogus_page
455 * replacement from working properly.
457 vm_page_clear_dirty(m, base, PAGE_SIZE - base);
458 } else if ((nsize & PAGE_MASK) &&
459 vm_page_is_cached(object, OFF_TO_IDX(nsize))) {
460 vm_page_cache_free(object, OFF_TO_IDX(nsize),
464 object->un_pager.vnp.vnp_size = nsize;
465 object->size = nobjsize;
466 VM_OBJECT_WUNLOCK(object);
470 * calculate the linear (byte) disk address of specified virtual
474 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
485 if (vp->v_iflag & VI_DOOMED)
488 bsize = vp->v_mount->mnt_stat.f_iosize;
489 vblock = address / bsize;
490 voffset = address % bsize;
492 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
494 if (*rtaddress != -1)
495 *rtaddress += voffset / DEV_BSIZE;
498 *run *= bsize/PAGE_SIZE;
499 *run -= voffset/PAGE_SIZE;
507 * small block filesystem vnode pager input
510 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m)
523 if (vp->v_iflag & VI_DOOMED)
526 bsize = vp->v_mount->mnt_stat.f_iosize;
528 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
530 sf = sf_buf_alloc(m, 0);
532 for (i = 0; i < PAGE_SIZE / bsize; i++) {
533 vm_ooffset_t address;
535 bits = vm_page_bits(i * bsize, bsize);
539 address = IDX_TO_OFF(m->pindex) + i * bsize;
540 if (address >= object->un_pager.vnp.vnp_size) {
543 error = vnode_pager_addr(vp, address, &fileaddr, NULL);
547 if (fileaddr != -1) {
548 bp = getpbuf(&vnode_pbuf_freecnt);
550 /* build a minimal buffer header */
551 bp->b_iocmd = BIO_READ;
552 bp->b_iodone = bdone;
553 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
554 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
555 bp->b_rcred = crhold(curthread->td_ucred);
556 bp->b_wcred = crhold(curthread->td_ucred);
557 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
558 bp->b_blkno = fileaddr;
561 bp->b_bcount = bsize;
562 bp->b_bufsize = bsize;
563 bp->b_runningbufspace = bp->b_bufsize;
564 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
567 bp->b_iooffset = dbtob(bp->b_blkno);
570 bwait(bp, PVM, "vnsrd");
572 if ((bp->b_ioflags & BIO_ERROR) != 0)
576 * free the buffer header back to the swap buffer pool
580 relpbuf(bp, &vnode_pbuf_freecnt);
584 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
585 KASSERT((m->dirty & bits) == 0,
586 ("vnode_pager_input_smlfs: page %p is dirty", m));
587 VM_OBJECT_WLOCK(object);
589 VM_OBJECT_WUNLOCK(object);
593 return VM_PAGER_ERROR;
599 * old style vnode pager input routine
602 vnode_pager_input_old(vm_object_t object, vm_page_t m)
611 VM_OBJECT_ASSERT_WLOCKED(object);
615 * Return failure if beyond current EOF
617 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
621 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
622 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
624 VM_OBJECT_WUNLOCK(object);
627 * Allocate a kernel virtual address and initialize so that
628 * we can use VOP_READ/WRITE routines.
630 sf = sf_buf_alloc(m, 0);
632 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
634 auio.uio_iov = &aiov;
636 auio.uio_offset = IDX_TO_OFF(m->pindex);
637 auio.uio_segflg = UIO_SYSSPACE;
638 auio.uio_rw = UIO_READ;
639 auio.uio_resid = size;
640 auio.uio_td = curthread;
642 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
644 int count = size - auio.uio_resid;
648 else if (count != PAGE_SIZE)
649 bzero((caddr_t)sf_buf_kva(sf) + count,
654 VM_OBJECT_WLOCK(object);
656 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
658 m->valid = VM_PAGE_BITS_ALL;
659 return error ? VM_PAGER_ERROR : VM_PAGER_OK;
663 * generic vnode pager input routine
667 * Local media VFS's that do not implement their own VOP_GETPAGES
668 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
669 * to implement the previous behaviour.
671 * All other FS's should use the bypass to get to the local media
672 * backing vp's VOP_GETPAGES.
675 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
682 VM_OBJECT_WUNLOCK(object);
683 rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead);
684 KASSERT(rtval != EOPNOTSUPP,
685 ("vnode_pager: FS getpages not implemented\n"));
686 VM_OBJECT_WLOCK(object);
691 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
692 int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg)
698 VM_OBJECT_WUNLOCK(object);
699 rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg);
700 KASSERT(rtval != EOPNOTSUPP,
701 ("vnode_pager: FS getpages_async not implemented\n"));
702 VM_OBJECT_WLOCK(object);
707 * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for
708 * local filesystems, where partially valid pages can only occur at
712 vnode_pager_local_getpages(struct vop_getpages_args *ap)
715 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
716 ap->a_rbehind, ap->a_rahead, NULL, NULL));
720 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap)
723 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
724 ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg));
728 * This is now called from local media FS's to operate against their
729 * own vnodes if they fail to implement VOP_GETPAGES.
732 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count,
733 int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg)
739 int bsize, pagesperblock, *freecnt;
740 int error, before, after, rbehind, rahead, poff, i;
741 int bytecount, secmask;
743 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
744 ("%s does not support devices", __func__));
746 if (vp->v_iflag & VI_DOOMED)
747 return (VM_PAGER_BAD);
749 object = vp->v_object;
750 foff = IDX_TO_OFF(m[0]->pindex);
751 bsize = vp->v_mount->mnt_stat.f_iosize;
752 pagesperblock = bsize / PAGE_SIZE;
754 KASSERT(foff < object->un_pager.vnp.vnp_size,
755 ("%s: page %p offset beyond vp %p size", __func__, m[0], vp));
756 KASSERT(count <= sizeof(bp->b_pages),
757 ("%s: requested %d pages", __func__, count));
760 * The last page has valid blocks. Invalid part can only
761 * exist at the end of file, and the page is made fully valid
762 * by zeroing in vm_pager_get_pages().
764 if (m[count - 1]->valid != 0 && --count == 0) {
766 iodone(arg, m, 1, 0);
767 return (VM_PAGER_OK);
771 * Synchronous and asynchronous paging operations use different
772 * free pbuf counters. This is done to avoid asynchronous requests
773 * to consume all pbufs.
774 * Allocate the pbuf at the very beginning of the function, so that
775 * if we are low on certain kind of pbufs don't even proceed to BMAP,
778 freecnt = iodone != NULL ?
779 &vnode_async_pbuf_freecnt : &vnode_pbuf_freecnt;
780 bp = getpbuf(freecnt);
783 * Get the underlying device blocks for the file with VOP_BMAP().
784 * If the file system doesn't support VOP_BMAP, use old way of
785 * getting pages via VOP_READ.
787 error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before);
788 if (error == EOPNOTSUPP) {
789 relpbuf(bp, freecnt);
790 VM_OBJECT_WLOCK(object);
791 for (i = 0; i < count; i++) {
792 PCPU_INC(cnt.v_vnodein);
793 PCPU_INC(cnt.v_vnodepgsin);
794 error = vnode_pager_input_old(object, m[i]);
798 VM_OBJECT_WUNLOCK(object);
800 } else if (error != 0) {
801 relpbuf(bp, freecnt);
802 return (VM_PAGER_ERROR);
806 * If the file system supports BMAP, but blocksize is smaller
807 * than a page size, then use special small filesystem code.
809 if (pagesperblock == 0) {
810 relpbuf(bp, freecnt);
811 for (i = 0; i < count; i++) {
812 PCPU_INC(cnt.v_vnodein);
813 PCPU_INC(cnt.v_vnodepgsin);
814 error = vnode_pager_input_smlfs(object, m[i]);
822 * A sparse file can be encountered only for a single page request,
823 * which may not be preceded by call to vm_pager_haspage().
825 if (bp->b_blkno == -1) {
827 ("%s: array[%d] request to a sparse file %p", __func__,
829 relpbuf(bp, freecnt);
830 pmap_zero_page(m[0]);
831 KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty",
833 VM_OBJECT_WLOCK(object);
834 m[0]->valid = VM_PAGE_BITS_ALL;
835 VM_OBJECT_WUNLOCK(object);
836 return (VM_PAGER_OK);
839 bp->b_blkno += (foff % bsize) / DEV_BSIZE;
841 /* Recalculate blocks available after/before to pages. */
842 poff = (foff % bsize) / PAGE_SIZE;
843 before *= pagesperblock;
845 after *= pagesperblock;
846 after += pagesperblock - (poff + 1);
847 if (m[0]->pindex + after >= object->size)
848 after = object->size - 1 - m[0]->pindex;
849 KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d",
850 __func__, count, after + 1));
853 /* Trim requested rbehind/rahead to possible values. */
854 rbehind = a_rbehind ? *a_rbehind : 0;
855 rahead = a_rahead ? *a_rahead : 0;
856 rbehind = min(rbehind, before);
857 rbehind = min(rbehind, m[0]->pindex);
858 rahead = min(rahead, after);
859 rahead = min(rahead, object->size - m[count - 1]->pindex);
860 KASSERT(rbehind + rahead + count <= sizeof(bp->b_pages),
861 ("%s: behind %d ahead %d count %d", __func__,
862 rbehind, rahead, count));
865 * Fill in the bp->b_pages[] array with requested and optional
866 * read behind or read ahead pages. Read behind pages are looked
867 * up in a backward direction, down to a first cached page. Same
868 * for read ahead pages, but there is no need to shift the array
869 * in case of encountering a cached page.
871 i = bp->b_npages = 0;
873 vm_pindex_t startpindex, tpindex;
876 VM_OBJECT_WLOCK(object);
877 startpindex = m[0]->pindex - rbehind;
878 if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL &&
879 p->pindex >= startpindex)
880 startpindex = p->pindex + 1;
882 /* tpindex is unsigned; beware of numeric underflow. */
883 for (tpindex = m[0]->pindex - 1;
884 tpindex >= startpindex && tpindex < m[0]->pindex;
886 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
887 VM_ALLOC_IFNOTCACHED);
889 /* Shift the array. */
890 for (int j = 0; j < i; j++)
891 bp->b_pages[j] = bp->b_pages[j +
892 tpindex + 1 - startpindex];
895 bp->b_pages[tpindex - startpindex] = p;
900 bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE;
904 /* Requested pages. */
905 for (int j = 0; j < count; j++, i++)
906 bp->b_pages[i] = m[j];
907 bp->b_npages += count;
910 vm_pindex_t endpindex, tpindex;
913 if (!VM_OBJECT_WOWNED(object))
914 VM_OBJECT_WLOCK(object);
915 endpindex = m[count - 1]->pindex + rahead + 1;
916 if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL &&
917 p->pindex < endpindex)
918 endpindex = p->pindex;
919 if (endpindex > object->size)
920 endpindex = object->size;
922 for (tpindex = m[count - 1]->pindex + 1;
923 tpindex < endpindex; i++, tpindex++) {
924 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
925 VM_ALLOC_IFNOTCACHED);
931 bp->b_pgafter = i - bp->b_npages;
936 if (VM_OBJECT_WOWNED(object))
937 VM_OBJECT_WUNLOCK(object);
939 /* Report back actual behind/ahead read. */
941 *a_rbehind = bp->b_pgbefore;
943 *a_rahead = bp->b_pgafter;
945 KASSERT(bp->b_npages <= sizeof(bp->b_pages),
946 ("%s: buf %p overflowed", __func__, bp));
949 * Recalculate first offset and bytecount with regards to read behind.
950 * Truncate bytecount to vnode real size and round up physical size
953 foff = IDX_TO_OFF(bp->b_pages[0]->pindex);
954 bytecount = bp->b_npages << PAGE_SHIFT;
955 if ((foff + bytecount) > object->un_pager.vnp.vnp_size)
956 bytecount = object->un_pager.vnp.vnp_size - foff;
957 secmask = bo->bo_bsize - 1;
958 KASSERT(secmask < PAGE_SIZE && secmask > 0,
959 ("%s: sector size %d too large", __func__, secmask + 1));
960 bytecount = (bytecount + secmask) & ~secmask;
963 * And map the pages to be read into the kva, if the filesystem
964 * requires mapped buffers.
966 if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
967 unmapped_buf_allowed) {
968 bp->b_data = unmapped_buf;
971 bp->b_data = bp->b_kvabase;
972 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
975 /* Build a minimal buffer header. */
976 bp->b_iocmd = BIO_READ;
977 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
978 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
979 bp->b_rcred = crhold(curthread->td_ucred);
980 bp->b_wcred = crhold(curthread->td_ucred);
983 bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount;
984 bp->b_iooffset = dbtob(bp->b_blkno);
986 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
987 PCPU_INC(cnt.v_vnodein);
988 PCPU_ADD(cnt.v_vnodepgsin, bp->b_npages);
990 if (iodone != NULL) { /* async */
991 bp->b_pgiodone = iodone;
993 bp->b_iodone = vnode_pager_generic_getpages_done_async;
994 bp->b_flags |= B_ASYNC;
997 return (VM_PAGER_OK);
999 bp->b_iodone = bdone;
1001 bwait(bp, PVM, "vnread");
1002 error = vnode_pager_generic_getpages_done(bp);
1003 for (i = 0; i < bp->b_npages; i++)
1004 bp->b_pages[i] = NULL;
1007 relpbuf(bp, &vnode_pbuf_freecnt);
1008 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
1013 vnode_pager_generic_getpages_done_async(struct buf *bp)
1017 error = vnode_pager_generic_getpages_done(bp);
1018 /* Run the iodone upon the requested range. */
1019 bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore,
1020 bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error);
1021 for (int i = 0; i < bp->b_npages; i++)
1022 bp->b_pages[i] = NULL;
1025 relpbuf(bp, &vnode_async_pbuf_freecnt);
1029 vnode_pager_generic_getpages_done(struct buf *bp)
1032 off_t tfoff, nextoff;
1035 error = (bp->b_ioflags & BIO_ERROR) != 0 ? EIO : 0;
1036 object = bp->b_vp->v_object;
1038 if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1039 if (!buf_mapped(bp)) {
1040 bp->b_data = bp->b_kvabase;
1041 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages,
1044 bzero(bp->b_data + bp->b_bcount,
1045 PAGE_SIZE * bp->b_npages - bp->b_bcount);
1047 if (buf_mapped(bp)) {
1048 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1049 bp->b_data = unmapped_buf;
1052 VM_OBJECT_WLOCK(object);
1053 for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1054 i < bp->b_npages; i++, tfoff = nextoff) {
1057 nextoff = tfoff + PAGE_SIZE;
1058 mt = bp->b_pages[i];
1060 if (nextoff <= object->un_pager.vnp.vnp_size) {
1062 * Read filled up entire page.
1064 mt->valid = VM_PAGE_BITS_ALL;
1065 KASSERT(mt->dirty == 0,
1066 ("%s: page %p is dirty", __func__, mt));
1067 KASSERT(!pmap_page_is_mapped(mt),
1068 ("%s: page %p is mapped", __func__, mt));
1071 * Read did not fill up entire page.
1073 * Currently we do not set the entire page valid,
1074 * we just try to clear the piece that we couldn't
1077 vm_page_set_valid_range(mt, 0,
1078 object->un_pager.vnp.vnp_size - tfoff);
1079 KASSERT((mt->dirty & vm_page_bits(0,
1080 object->un_pager.vnp.vnp_size - tfoff)) == 0,
1081 ("%s: page %p is dirty", __func__, mt));
1084 if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter)
1085 vm_page_readahead_finish(mt);
1087 VM_OBJECT_WUNLOCK(object);
1089 printf("%s: I/O read error %d\n", __func__, error);
1095 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
1096 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1097 * vnode_pager_generic_putpages() to implement the previous behaviour.
1099 * All other FS's should use the bypass to get to the local media
1100 * backing vp's VOP_PUTPAGES.
1103 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1104 int flags, int *rtvals)
1108 int bytes = count * PAGE_SIZE;
1111 * Force synchronous operation if we are extremely low on memory
1112 * to prevent a low-memory deadlock. VOP operations often need to
1113 * allocate more memory to initiate the I/O ( i.e. do a BMAP
1114 * operation ). The swapper handles the case by limiting the amount
1115 * of asynchronous I/O, but that sort of solution doesn't scale well
1116 * for the vnode pager without a lot of work.
1118 * Also, the backing vnode's iodone routine may not wake the pageout
1119 * daemon up. This should be probably be addressed XXX.
1122 if (vm_cnt.v_free_count + vm_cnt.v_cache_count <
1123 vm_cnt.v_pageout_free_min)
1124 flags |= VM_PAGER_PUT_SYNC;
1127 * Call device-specific putpages function
1129 vp = object->handle;
1130 VM_OBJECT_WUNLOCK(object);
1131 rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1132 KASSERT(rtval != EOPNOTSUPP,
1133 ("vnode_pager: stale FS putpages\n"));
1134 VM_OBJECT_WLOCK(object);
1139 * This is now called from local media FS's to operate against their
1140 * own vnodes if they fail to implement VOP_PUTPAGES.
1142 * This is typically called indirectly via the pageout daemon and
1143 * clustering has already typically occurred, so in general we ask the
1144 * underlying filesystem to write the data out asynchronously rather
1148 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1149 int flags, int *rtvals)
1156 int maxsize, ncount;
1157 vm_ooffset_t poffset;
1163 static struct timeval lastfail;
1166 object = vp->v_object;
1167 count = bytecount / PAGE_SIZE;
1169 for (i = 0; i < count; i++)
1170 rtvals[i] = VM_PAGER_ERROR;
1172 if ((int64_t)ma[0]->pindex < 0) {
1173 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1174 (long)ma[0]->pindex, (u_long)ma[0]->dirty);
1175 rtvals[0] = VM_PAGER_BAD;
1176 return VM_PAGER_BAD;
1179 maxsize = count * PAGE_SIZE;
1182 poffset = IDX_TO_OFF(ma[0]->pindex);
1185 * If the page-aligned write is larger then the actual file we
1186 * have to invalidate pages occurring beyond the file EOF. However,
1187 * there is an edge case where a file may not be page-aligned where
1188 * the last page is partially invalid. In this case the filesystem
1189 * may not properly clear the dirty bits for the entire page (which
1190 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1191 * With the page locked we are free to fix-up the dirty bits here.
1193 * We do not under any circumstances truncate the valid bits, as
1194 * this will screw up bogus page replacement.
1196 VM_OBJECT_WLOCK(object);
1197 if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1198 if (object->un_pager.vnp.vnp_size > poffset) {
1201 maxsize = object->un_pager.vnp.vnp_size - poffset;
1202 ncount = btoc(maxsize);
1203 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1205 * If the object is locked and the following
1206 * conditions hold, then the page's dirty
1207 * field cannot be concurrently changed by a
1211 vm_page_assert_sbusied(m);
1212 KASSERT(!pmap_page_is_write_mapped(m),
1213 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1214 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1221 if (ncount < count) {
1222 for (i = ncount; i < count; i++) {
1223 rtvals[i] = VM_PAGER_BAD;
1227 VM_OBJECT_WUNLOCK(object);
1230 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
1231 * rather then a bdwrite() to prevent paging I/O from saturating
1232 * the buffer cache. Dummy-up the sequential heuristic to cause
1233 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
1234 * the system decides how to cluster.
1237 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1239 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1240 ioflags |= IO_ASYNC;
1241 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1242 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1244 aiov.iov_base = (caddr_t) 0;
1245 aiov.iov_len = maxsize;
1246 auio.uio_iov = &aiov;
1247 auio.uio_iovcnt = 1;
1248 auio.uio_offset = poffset;
1249 auio.uio_segflg = UIO_NOCOPY;
1250 auio.uio_rw = UIO_WRITE;
1251 auio.uio_resid = maxsize;
1252 auio.uio_td = (struct thread *) 0;
1253 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1254 PCPU_INC(cnt.v_vnodeout);
1255 PCPU_ADD(cnt.v_vnodepgsout, ncount);
1258 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1)))
1259 printf("vnode_pager_putpages: I/O error %d\n", error);
1261 if (auio.uio_resid) {
1262 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1))
1263 printf("vnode_pager_putpages: residual I/O %zd at %lu\n",
1264 auio.uio_resid, (u_long)ma[0]->pindex);
1266 for (i = 0; i < ncount; i++) {
1267 rtvals[i] = VM_PAGER_OK;
1273 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written)
1280 obj = ma[0]->object;
1281 VM_OBJECT_WLOCK(obj);
1282 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1283 if (pos < trunc_page(written)) {
1284 rtvals[i] = VM_PAGER_OK;
1285 vm_page_undirty(ma[i]);
1287 /* Partially written page. */
1288 rtvals[i] = VM_PAGER_AGAIN;
1289 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1292 VM_OBJECT_WUNLOCK(obj);
1296 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1300 vm_ooffset_t old_wm;
1302 VM_OBJECT_WLOCK(object);
1303 if (object->type != OBJT_VNODE) {
1304 VM_OBJECT_WUNLOCK(object);
1307 old_wm = object->un_pager.vnp.writemappings;
1308 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1309 vp = object->handle;
1310 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1311 ASSERT_VOP_ELOCKED(vp, "v_writecount inc");
1312 VOP_ADD_WRITECOUNT(vp, 1);
1313 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1314 __func__, vp, vp->v_writecount);
1315 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1316 ASSERT_VOP_ELOCKED(vp, "v_writecount dec");
1317 VOP_ADD_WRITECOUNT(vp, -1);
1318 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1319 __func__, vp, vp->v_writecount);
1321 VM_OBJECT_WUNLOCK(object);
1325 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1332 VM_OBJECT_WLOCK(object);
1335 * First, recheck the object type to account for the race when
1336 * the vnode is reclaimed.
1338 if (object->type != OBJT_VNODE) {
1339 VM_OBJECT_WUNLOCK(object);
1344 * Optimize for the case when writemappings is not going to
1348 if (object->un_pager.vnp.writemappings != inc) {
1349 object->un_pager.vnp.writemappings -= inc;
1350 VM_OBJECT_WUNLOCK(object);
1354 vp = object->handle;
1356 VM_OBJECT_WUNLOCK(object);
1358 vn_start_write(vp, &mp, V_WAIT);
1359 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1362 * Decrement the object's writemappings, by swapping the start
1363 * and end arguments for vnode_pager_update_writecount(). If
1364 * there was not a race with vnode reclaimation, then the
1365 * vnode's v_writecount is decremented.
1367 vnode_pager_update_writecount(object, end, start);
1371 vn_finished_write(mp);