2 * SPDX-License-Identifier: BSD-4-Clause
4 * Copyright (c) 1990 University of Utah.
5 * Copyright (c) 1991 The Regents of the University of California.
7 * Copyright (c) 1993, 1994 John S. Dyson
8 * Copyright (c) 1995, David Greenman
10 * This code is derived from software contributed to Berkeley by
11 * the Systems Programming Group of the University of Utah Computer
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
17 * 1. Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 * 2. Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in the
21 * documentation and/or other materials provided with the distribution.
22 * 3. All advertising materials mentioning features or use of this software
23 * must display the following acknowledgement:
24 * This product includes software developed by the University of
25 * California, Berkeley and its contributors.
26 * 4. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91
46 * Page to/from files (vnodes).
51 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
52 * greatly re-simplify the vnode_pager.
55 #include <sys/cdefs.h>
56 __FBSDID("$FreeBSD$");
60 #include <sys/param.h>
61 #include <sys/kernel.h>
62 #include <sys/systm.h>
63 #include <sys/sysctl.h>
65 #include <sys/vnode.h>
66 #include <sys/mount.h>
69 #include <sys/vmmeter.h>
70 #include <sys/limits.h>
72 #include <sys/rwlock.h>
73 #include <sys/sf_buf.h>
74 #include <sys/domainset.h>
76 #include <machine/atomic.h>
79 #include <vm/vm_param.h>
80 #include <vm/vm_object.h>
81 #include <vm/vm_page.h>
82 #include <vm/vm_pager.h>
83 #include <vm/vm_map.h>
84 #include <vm/vnode_pager.h>
85 #include <vm/vm_extern.h>
88 static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
89 daddr_t *rtaddress, int *run);
90 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
91 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
92 static void vnode_pager_dealloc(vm_object_t);
93 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *);
94 static int vnode_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
95 int *, vop_getpages_iodone_t, void *);
96 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *);
97 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
98 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t,
99 vm_ooffset_t, struct ucred *cred);
100 static int vnode_pager_generic_getpages_done(struct buf *);
101 static void vnode_pager_generic_getpages_done_async(struct buf *);
103 struct pagerops vnodepagerops = {
104 .pgo_alloc = vnode_pager_alloc,
105 .pgo_dealloc = vnode_pager_dealloc,
106 .pgo_getpages = vnode_pager_getpages,
107 .pgo_getpages_async = vnode_pager_getpages_async,
108 .pgo_putpages = vnode_pager_putpages,
109 .pgo_haspage = vnode_pager_haspage,
112 static struct domainset *vnode_domainset = NULL;
114 SYSCTL_PROC(_debug, OID_AUTO, vnode_domainset, CTLTYPE_STRING | CTLFLAG_RW,
115 &vnode_domainset, 0, sysctl_handle_domainset, "A",
116 "Default vnode NUMA policy");
118 static uma_zone_t vnode_pbuf_zone;
121 vnode_pager_init(void *dummy)
124 vnode_pbuf_zone = pbuf_zsecond_create("vnpbuf", nswbuf * 8);
126 SYSINIT(vnode_pager, SI_SUB_CPU, SI_ORDER_ANY, vnode_pager_init, NULL);
128 /* Create the VM system backing object for this vnode */
130 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
133 vm_ooffset_t size = isize;
136 if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE)
139 while ((object = vp->v_object) != NULL) {
140 VM_OBJECT_WLOCK(object);
141 if (!(object->flags & OBJ_DEAD)) {
142 VM_OBJECT_WUNLOCK(object);
146 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
147 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vodead", 0);
148 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
152 if (vn_isdisk(vp, NULL)) {
153 size = IDX_TO_OFF(INT_MAX);
155 if (VOP_GETATTR(vp, &va, td->td_ucred))
161 object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred);
163 * Dereference the reference we just created. This assumes
164 * that the object is associated with the vp.
166 VM_OBJECT_WLOCK(object);
168 VM_OBJECT_WUNLOCK(object);
171 KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
177 vnode_destroy_vobject(struct vnode *vp)
179 struct vm_object *obj;
184 ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
185 VM_OBJECT_WLOCK(obj);
186 umtx_shm_object_terminated(obj);
187 if (obj->ref_count == 0) {
189 * don't double-terminate the object
191 if ((obj->flags & OBJ_DEAD) == 0) {
192 vm_object_terminate(obj);
195 * Waiters were already handled during object
196 * termination. The exclusive vnode lock hopefully
197 * prevented new waiters from referencing the dying
200 KASSERT((obj->flags & OBJ_DISCONNECTWNT) == 0,
201 ("OBJ_DISCONNECTWNT set obj %p flags %x",
204 VM_OBJECT_WUNLOCK(obj);
208 * Woe to the process that tries to page now :-).
210 vm_pager_deallocate(obj);
211 VM_OBJECT_WUNLOCK(obj);
213 KASSERT(vp->v_object == NULL, ("vp %p obj %p", vp, vp->v_object));
218 * Allocate (or lookup) pager for a vnode.
219 * Handle is a vnode pointer.
224 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
225 vm_ooffset_t offset, struct ucred *cred)
231 * Pageout to vnode, no can do yet.
236 vp = (struct vnode *) handle;
239 * If the object is being terminated, wait for it to
243 while ((object = vp->v_object) != NULL) {
244 VM_OBJECT_WLOCK(object);
245 if ((object->flags & OBJ_DEAD) == 0)
247 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
248 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vadead", 0);
251 KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference"));
253 if (object == NULL) {
255 * Add an object of the appropriate size
257 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
259 object->un_pager.vnp.vnp_size = size;
260 object->un_pager.vnp.writemappings = 0;
261 object->domain.dr_policy = vnode_domainset;
263 object->handle = handle;
265 if (vp->v_object != NULL) {
267 * Object has been created while we were sleeping
270 VM_OBJECT_WLOCK(object);
271 KASSERT(object->ref_count == 1,
272 ("leaked ref %p %d", object, object->ref_count));
273 object->type = OBJT_DEAD;
274 object->ref_count = 0;
275 VM_OBJECT_WUNLOCK(object);
276 vm_object_destroy(object);
279 vp->v_object = object;
283 #if VM_NRESERVLEVEL > 0
284 vm_object_color(object, 0);
286 VM_OBJECT_WUNLOCK(object);
293 * The object must be locked.
296 vnode_pager_dealloc(vm_object_t object)
303 panic("vnode_pager_dealloc: pager already dealloced");
305 VM_OBJECT_ASSERT_WLOCKED(object);
306 vm_object_pip_wait(object, "vnpdea");
307 refs = object->ref_count;
309 object->handle = NULL;
310 object->type = OBJT_DEAD;
311 if (object->flags & OBJ_DISCONNECTWNT) {
312 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
315 ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
316 if (object->un_pager.vnp.writemappings > 0) {
317 object->un_pager.vnp.writemappings = 0;
318 VOP_ADD_WRITECOUNT(vp, -1);
319 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
320 __func__, vp, vp->v_writecount);
324 VM_OBJECT_WUNLOCK(object);
327 VM_OBJECT_WLOCK(object);
331 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
334 struct vnode *vp = object->handle;
340 int pagesperblock, blocksperpage;
342 VM_OBJECT_ASSERT_WLOCKED(object);
344 * If no vp or vp is doomed or marked transparent to VM, we do not
347 if (vp == NULL || vp->v_iflag & VI_DOOMED)
350 * If the offset is beyond end of file we do
353 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
356 bsize = vp->v_mount->mnt_stat.f_iosize;
357 pagesperblock = bsize / PAGE_SIZE;
359 if (pagesperblock > 0) {
360 reqblock = pindex / pagesperblock;
362 blocksperpage = (PAGE_SIZE / bsize);
363 reqblock = pindex * blocksperpage;
365 VM_OBJECT_WUNLOCK(object);
366 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
367 VM_OBJECT_WLOCK(object);
372 if (pagesperblock > 0) {
373 poff = pindex - (reqblock * pagesperblock);
375 *before *= pagesperblock;
380 * The BMAP vop can report a partial block in the
381 * 'after', but must not report blocks after EOF.
382 * Assert the latter, and truncate 'after' in case
385 KASSERT((reqblock + *after) * pagesperblock <
386 roundup2(object->size, pagesperblock),
387 ("%s: reqblock %jd after %d size %ju", __func__,
388 (intmax_t )reqblock, *after,
389 (uintmax_t )object->size));
390 *after *= pagesperblock;
391 *after += pagesperblock - (poff + 1);
392 if (pindex + *after >= object->size)
393 *after = object->size - 1 - pindex;
397 *before /= blocksperpage;
401 *after /= blocksperpage;
408 * Lets the VM system know about a change in size for a file.
409 * We adjust our own internal size and flush any cached pages in
410 * the associated object that are affected by the size change.
412 * Note: this routine may be invoked as a result of a pager put
413 * operation (possibly at object termination time), so we must be careful.
416 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
420 vm_pindex_t nobjsize;
422 if ((object = vp->v_object) == NULL)
424 /* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */
425 VM_OBJECT_WLOCK(object);
426 if (object->type == OBJT_DEAD) {
427 VM_OBJECT_WUNLOCK(object);
430 KASSERT(object->type == OBJT_VNODE,
431 ("not vnode-backed object %p", object));
432 if (nsize == object->un_pager.vnp.vnp_size) {
434 * Hasn't changed size
436 VM_OBJECT_WUNLOCK(object);
439 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
440 if (nsize < object->un_pager.vnp.vnp_size) {
442 * File has shrunk. Toss any cached pages beyond the new EOF.
444 if (nobjsize < object->size)
445 vm_object_page_remove(object, nobjsize, object->size,
448 * this gets rid of garbage at the end of a page that is now
449 * only partially backed by the vnode.
451 * XXX for some reason (I don't know yet), if we take a
452 * completely invalid page and mark it partially valid
453 * it can screw up NFS reads, so we don't allow the case.
455 if ((nsize & PAGE_MASK) &&
456 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
458 int base = (int)nsize & PAGE_MASK;
459 int size = PAGE_SIZE - base;
462 * Clear out partial-page garbage in case
463 * the page has been mapped.
465 pmap_zero_page_area(m, base, size);
468 * Update the valid bits to reflect the blocks that
469 * have been zeroed. Some of these valid bits may
470 * have already been set.
472 vm_page_set_valid_range(m, base, size);
475 * Round "base" to the next block boundary so that the
476 * dirty bit for a partially zeroed block is not
479 base = roundup2(base, DEV_BSIZE);
482 * Clear out partial-page dirty bits.
484 * note that we do not clear out the valid
485 * bits. This would prevent bogus_page
486 * replacement from working properly.
488 vm_page_clear_dirty(m, base, PAGE_SIZE - base);
491 object->un_pager.vnp.vnp_size = nsize;
492 object->size = nobjsize;
493 VM_OBJECT_WUNLOCK(object);
497 * calculate the linear (byte) disk address of specified virtual
501 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
512 if (vp->v_iflag & VI_DOOMED)
515 bsize = vp->v_mount->mnt_stat.f_iosize;
516 vblock = address / bsize;
517 voffset = address % bsize;
519 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
521 if (*rtaddress != -1)
522 *rtaddress += voffset / DEV_BSIZE;
525 *run *= bsize/PAGE_SIZE;
526 *run -= voffset/PAGE_SIZE;
534 * small block filesystem vnode pager input
537 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m)
550 if (vp->v_iflag & VI_DOOMED)
553 bsize = vp->v_mount->mnt_stat.f_iosize;
555 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
557 sf = sf_buf_alloc(m, 0);
559 for (i = 0; i < PAGE_SIZE / bsize; i++) {
560 vm_ooffset_t address;
562 bits = vm_page_bits(i * bsize, bsize);
566 address = IDX_TO_OFF(m->pindex) + i * bsize;
567 if (address >= object->un_pager.vnp.vnp_size) {
570 error = vnode_pager_addr(vp, address, &fileaddr, NULL);
574 if (fileaddr != -1) {
575 bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK);
577 /* build a minimal buffer header */
578 bp->b_iocmd = BIO_READ;
579 bp->b_iodone = bdone;
580 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
581 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
582 bp->b_rcred = crhold(curthread->td_ucred);
583 bp->b_wcred = crhold(curthread->td_ucred);
584 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
585 bp->b_blkno = fileaddr;
588 bp->b_bcount = bsize;
589 bp->b_bufsize = bsize;
590 bp->b_runningbufspace = bp->b_bufsize;
591 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
594 bp->b_iooffset = dbtob(bp->b_blkno);
597 bwait(bp, PVM, "vnsrd");
599 if ((bp->b_ioflags & BIO_ERROR) != 0)
603 * free the buffer header back to the swap buffer pool
607 uma_zfree(vnode_pbuf_zone, bp);
611 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
612 KASSERT((m->dirty & bits) == 0,
613 ("vnode_pager_input_smlfs: page %p is dirty", m));
614 VM_OBJECT_WLOCK(object);
616 VM_OBJECT_WUNLOCK(object);
620 return VM_PAGER_ERROR;
626 * old style vnode pager input routine
629 vnode_pager_input_old(vm_object_t object, vm_page_t m)
638 VM_OBJECT_ASSERT_WLOCKED(object);
642 * Return failure if beyond current EOF
644 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
648 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
649 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
651 VM_OBJECT_WUNLOCK(object);
654 * Allocate a kernel virtual address and initialize so that
655 * we can use VOP_READ/WRITE routines.
657 sf = sf_buf_alloc(m, 0);
659 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
661 auio.uio_iov = &aiov;
663 auio.uio_offset = IDX_TO_OFF(m->pindex);
664 auio.uio_segflg = UIO_SYSSPACE;
665 auio.uio_rw = UIO_READ;
666 auio.uio_resid = size;
667 auio.uio_td = curthread;
669 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
671 int count = size - auio.uio_resid;
675 else if (count != PAGE_SIZE)
676 bzero((caddr_t)sf_buf_kva(sf) + count,
681 VM_OBJECT_WLOCK(object);
683 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
685 m->valid = VM_PAGE_BITS_ALL;
686 return error ? VM_PAGER_ERROR : VM_PAGER_OK;
690 * generic vnode pager input routine
694 * Local media VFS's that do not implement their own VOP_GETPAGES
695 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
696 * to implement the previous behaviour.
698 * All other FS's should use the bypass to get to the local media
699 * backing vp's VOP_GETPAGES.
702 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
709 VM_OBJECT_WUNLOCK(object);
710 rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead);
711 KASSERT(rtval != EOPNOTSUPP,
712 ("vnode_pager: FS getpages not implemented\n"));
713 VM_OBJECT_WLOCK(object);
718 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
719 int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg)
725 VM_OBJECT_WUNLOCK(object);
726 rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg);
727 KASSERT(rtval != EOPNOTSUPP,
728 ("vnode_pager: FS getpages_async not implemented\n"));
729 VM_OBJECT_WLOCK(object);
734 * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for
735 * local filesystems, where partially valid pages can only occur at
739 vnode_pager_local_getpages(struct vop_getpages_args *ap)
742 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
743 ap->a_rbehind, ap->a_rahead, NULL, NULL));
747 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap)
750 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
751 ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg));
755 * This is now called from local media FS's to operate against their
756 * own vnodes if they fail to implement VOP_GETPAGES.
759 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count,
760 int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg)
769 int bsize, pagesperblock;
770 int error, before, after, rbehind, rahead, poff, i;
771 int bytecount, secmask;
773 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
774 ("%s does not support devices", __func__));
776 if (vp->v_iflag & VI_DOOMED)
777 return (VM_PAGER_BAD);
779 object = vp->v_object;
780 foff = IDX_TO_OFF(m[0]->pindex);
781 bsize = vp->v_mount->mnt_stat.f_iosize;
782 pagesperblock = bsize / PAGE_SIZE;
784 KASSERT(foff < object->un_pager.vnp.vnp_size,
785 ("%s: page %p offset beyond vp %p size", __func__, m[0], vp));
786 KASSERT(count <= sizeof(bp->b_pages),
787 ("%s: requested %d pages", __func__, count));
790 * The last page has valid blocks. Invalid part can only
791 * exist at the end of file, and the page is made fully valid
792 * by zeroing in vm_pager_get_pages().
794 if (m[count - 1]->valid != 0 && --count == 0) {
796 iodone(arg, m, 1, 0);
797 return (VM_PAGER_OK);
800 bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK);
803 * Get the underlying device blocks for the file with VOP_BMAP().
804 * If the file system doesn't support VOP_BMAP, use old way of
805 * getting pages via VOP_READ.
807 error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before);
808 if (error == EOPNOTSUPP) {
809 uma_zfree(vnode_pbuf_zone, bp);
810 VM_OBJECT_WLOCK(object);
811 for (i = 0; i < count; i++) {
812 VM_CNT_INC(v_vnodein);
813 VM_CNT_INC(v_vnodepgsin);
814 error = vnode_pager_input_old(object, m[i]);
818 VM_OBJECT_WUNLOCK(object);
820 } else if (error != 0) {
821 uma_zfree(vnode_pbuf_zone, bp);
822 return (VM_PAGER_ERROR);
826 * If the file system supports BMAP, but blocksize is smaller
827 * than a page size, then use special small filesystem code.
829 if (pagesperblock == 0) {
830 uma_zfree(vnode_pbuf_zone, bp);
831 for (i = 0; i < count; i++) {
832 VM_CNT_INC(v_vnodein);
833 VM_CNT_INC(v_vnodepgsin);
834 error = vnode_pager_input_smlfs(object, m[i]);
842 * A sparse file can be encountered only for a single page request,
843 * which may not be preceded by call to vm_pager_haspage().
845 if (bp->b_blkno == -1) {
847 ("%s: array[%d] request to a sparse file %p", __func__,
849 uma_zfree(vnode_pbuf_zone, bp);
850 pmap_zero_page(m[0]);
851 KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty",
853 VM_OBJECT_WLOCK(object);
854 m[0]->valid = VM_PAGE_BITS_ALL;
855 VM_OBJECT_WUNLOCK(object);
856 return (VM_PAGER_OK);
860 blkno0 = bp->b_blkno;
862 bp->b_blkno += (foff % bsize) / DEV_BSIZE;
864 /* Recalculate blocks available after/before to pages. */
865 poff = (foff % bsize) / PAGE_SIZE;
866 before *= pagesperblock;
868 after *= pagesperblock;
869 after += pagesperblock - (poff + 1);
870 if (m[0]->pindex + after >= object->size)
871 after = object->size - 1 - m[0]->pindex;
872 KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d",
873 __func__, count, after + 1));
876 /* Trim requested rbehind/rahead to possible values. */
877 rbehind = a_rbehind ? *a_rbehind : 0;
878 rahead = a_rahead ? *a_rahead : 0;
879 rbehind = min(rbehind, before);
880 rbehind = min(rbehind, m[0]->pindex);
881 rahead = min(rahead, after);
882 rahead = min(rahead, object->size - m[count - 1]->pindex);
884 * Check that total amount of pages fit into buf. Trim rbehind and
885 * rahead evenly if not.
887 if (rbehind + rahead + count > nitems(bp->b_pages)) {
890 trim = rbehind + rahead + count - nitems(bp->b_pages) + 1;
891 sum = rbehind + rahead;
892 if (rbehind == before) {
893 /* Roundup rbehind trim to block size. */
894 rbehind -= roundup(trim * rbehind / sum, pagesperblock);
898 rbehind -= trim * rbehind / sum;
899 rahead -= trim * rahead / sum;
901 KASSERT(rbehind + rahead + count <= nitems(bp->b_pages),
902 ("%s: behind %d ahead %d count %d", __func__,
903 rbehind, rahead, count));
906 * Fill in the bp->b_pages[] array with requested and optional
907 * read behind or read ahead pages. Read behind pages are looked
908 * up in a backward direction, down to a first cached page. Same
909 * for read ahead pages, but there is no need to shift the array
910 * in case of encountering a cached page.
912 i = bp->b_npages = 0;
914 vm_pindex_t startpindex, tpindex;
917 VM_OBJECT_WLOCK(object);
918 startpindex = m[0]->pindex - rbehind;
919 if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL &&
920 p->pindex >= startpindex)
921 startpindex = p->pindex + 1;
923 /* tpindex is unsigned; beware of numeric underflow. */
924 for (tpindex = m[0]->pindex - 1;
925 tpindex >= startpindex && tpindex < m[0]->pindex;
927 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
929 /* Shift the array. */
930 for (int j = 0; j < i; j++)
931 bp->b_pages[j] = bp->b_pages[j +
932 tpindex + 1 - startpindex];
935 bp->b_pages[tpindex - startpindex] = p;
940 bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE;
944 /* Requested pages. */
945 for (int j = 0; j < count; j++, i++)
946 bp->b_pages[i] = m[j];
947 bp->b_npages += count;
950 vm_pindex_t endpindex, tpindex;
953 if (!VM_OBJECT_WOWNED(object))
954 VM_OBJECT_WLOCK(object);
955 endpindex = m[count - 1]->pindex + rahead + 1;
956 if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL &&
957 p->pindex < endpindex)
958 endpindex = p->pindex;
959 if (endpindex > object->size)
960 endpindex = object->size;
962 for (tpindex = m[count - 1]->pindex + 1;
963 tpindex < endpindex; i++, tpindex++) {
964 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
970 bp->b_pgafter = i - bp->b_npages;
975 if (VM_OBJECT_WOWNED(object))
976 VM_OBJECT_WUNLOCK(object);
978 /* Report back actual behind/ahead read. */
980 *a_rbehind = bp->b_pgbefore;
982 *a_rahead = bp->b_pgafter;
985 KASSERT(bp->b_npages <= nitems(bp->b_pages),
986 ("%s: buf %p overflowed", __func__, bp));
987 for (int j = 1, prev = 0; j < bp->b_npages; j++) {
988 if (bp->b_pages[j] == bogus_page)
990 KASSERT(bp->b_pages[j]->pindex - bp->b_pages[prev]->pindex ==
991 j - prev, ("%s: pages array not consecutive, bp %p",
998 * Recalculate first offset and bytecount with regards to read behind.
999 * Truncate bytecount to vnode real size and round up physical size
1002 foff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1003 bytecount = bp->b_npages << PAGE_SHIFT;
1004 if ((foff + bytecount) > object->un_pager.vnp.vnp_size)
1005 bytecount = object->un_pager.vnp.vnp_size - foff;
1006 secmask = bo->bo_bsize - 1;
1007 KASSERT(secmask < PAGE_SIZE && secmask > 0,
1008 ("%s: sector size %d too large", __func__, secmask + 1));
1009 bytecount = (bytecount + secmask) & ~secmask;
1012 * And map the pages to be read into the kva, if the filesystem
1013 * requires mapped buffers.
1015 if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
1016 unmapped_buf_allowed) {
1017 bp->b_data = unmapped_buf;
1020 bp->b_data = bp->b_kvabase;
1021 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1024 /* Build a minimal buffer header. */
1025 bp->b_iocmd = BIO_READ;
1026 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
1027 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
1028 bp->b_rcred = crhold(curthread->td_ucred);
1029 bp->b_wcred = crhold(curthread->td_ucred);
1032 bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount;
1033 bp->b_iooffset = dbtob(bp->b_blkno);
1034 KASSERT(IDX_TO_OFF(m[0]->pindex - bp->b_pages[0]->pindex) ==
1035 (blkno0 - bp->b_blkno) * DEV_BSIZE +
1036 IDX_TO_OFF(m[0]->pindex) % bsize,
1037 ("wrong offsets bsize %d m[0] %ju b_pages[0] %ju "
1038 "blkno0 %ju b_blkno %ju", bsize,
1039 (uintmax_t)m[0]->pindex, (uintmax_t)bp->b_pages[0]->pindex,
1040 (uintmax_t)blkno0, (uintmax_t)bp->b_blkno));
1042 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
1043 VM_CNT_INC(v_vnodein);
1044 VM_CNT_ADD(v_vnodepgsin, bp->b_npages);
1046 if (iodone != NULL) { /* async */
1047 bp->b_pgiodone = iodone;
1048 bp->b_caller1 = arg;
1049 bp->b_iodone = vnode_pager_generic_getpages_done_async;
1050 bp->b_flags |= B_ASYNC;
1053 return (VM_PAGER_OK);
1055 bp->b_iodone = bdone;
1057 bwait(bp, PVM, "vnread");
1058 error = vnode_pager_generic_getpages_done(bp);
1059 for (i = 0; i < bp->b_npages; i++)
1060 bp->b_pages[i] = NULL;
1063 uma_zfree(vnode_pbuf_zone, bp);
1064 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
1069 vnode_pager_generic_getpages_done_async(struct buf *bp)
1073 error = vnode_pager_generic_getpages_done(bp);
1074 /* Run the iodone upon the requested range. */
1075 bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore,
1076 bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error);
1077 for (int i = 0; i < bp->b_npages; i++)
1078 bp->b_pages[i] = NULL;
1081 uma_zfree(vnode_pbuf_zone, bp);
1085 vnode_pager_generic_getpages_done(struct buf *bp)
1088 off_t tfoff, nextoff;
1091 error = (bp->b_ioflags & BIO_ERROR) != 0 ? EIO : 0;
1092 object = bp->b_vp->v_object;
1094 if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1095 if (!buf_mapped(bp)) {
1096 bp->b_data = bp->b_kvabase;
1097 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages,
1100 bzero(bp->b_data + bp->b_bcount,
1101 PAGE_SIZE * bp->b_npages - bp->b_bcount);
1103 if (buf_mapped(bp)) {
1104 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1105 bp->b_data = unmapped_buf;
1108 VM_OBJECT_WLOCK(object);
1109 for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1110 i < bp->b_npages; i++, tfoff = nextoff) {
1113 nextoff = tfoff + PAGE_SIZE;
1114 mt = bp->b_pages[i];
1116 if (nextoff <= object->un_pager.vnp.vnp_size) {
1118 * Read filled up entire page.
1120 mt->valid = VM_PAGE_BITS_ALL;
1121 KASSERT(mt->dirty == 0,
1122 ("%s: page %p is dirty", __func__, mt));
1123 KASSERT(!pmap_page_is_mapped(mt),
1124 ("%s: page %p is mapped", __func__, mt));
1127 * Read did not fill up entire page.
1129 * Currently we do not set the entire page valid,
1130 * we just try to clear the piece that we couldn't
1133 vm_page_set_valid_range(mt, 0,
1134 object->un_pager.vnp.vnp_size - tfoff);
1135 KASSERT((mt->dirty & vm_page_bits(0,
1136 object->un_pager.vnp.vnp_size - tfoff)) == 0,
1137 ("%s: page %p is dirty", __func__, mt));
1140 if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter)
1141 vm_page_readahead_finish(mt);
1143 VM_OBJECT_WUNLOCK(object);
1145 printf("%s: I/O read error %d\n", __func__, error);
1151 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
1152 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1153 * vnode_pager_generic_putpages() to implement the previous behaviour.
1155 * All other FS's should use the bypass to get to the local media
1156 * backing vp's VOP_PUTPAGES.
1159 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1160 int flags, int *rtvals)
1164 int bytes = count * PAGE_SIZE;
1167 * Force synchronous operation if we are extremely low on memory
1168 * to prevent a low-memory deadlock. VOP operations often need to
1169 * allocate more memory to initiate the I/O ( i.e. do a BMAP
1170 * operation ). The swapper handles the case by limiting the amount
1171 * of asynchronous I/O, but that sort of solution doesn't scale well
1172 * for the vnode pager without a lot of work.
1174 * Also, the backing vnode's iodone routine may not wake the pageout
1175 * daemon up. This should be probably be addressed XXX.
1178 if (vm_page_count_min())
1179 flags |= VM_PAGER_PUT_SYNC;
1182 * Call device-specific putpages function
1184 vp = object->handle;
1185 VM_OBJECT_WUNLOCK(object);
1186 rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1187 KASSERT(rtval != EOPNOTSUPP,
1188 ("vnode_pager: stale FS putpages\n"));
1189 VM_OBJECT_WLOCK(object);
1193 vn_off2bidx(vm_ooffset_t offset)
1196 return ((offset & PAGE_MASK) / DEV_BSIZE);
1200 vn_dirty_blk(vm_page_t m, vm_ooffset_t offset)
1203 KASSERT(IDX_TO_OFF(m->pindex) <= offset &&
1204 offset < IDX_TO_OFF(m->pindex + 1),
1205 ("page %p pidx %ju offset %ju", m, (uintmax_t)m->pindex,
1206 (uintmax_t)offset));
1207 return ((m->dirty & ((vm_page_bits_t)1 << vn_off2bidx(offset))) != 0);
1211 * This is now called from local media FS's to operate against their
1212 * own vnodes if they fail to implement VOP_PUTPAGES.
1214 * This is typically called indirectly via the pageout daemon and
1215 * clustering has already typically occurred, so in general we ask the
1216 * underlying filesystem to write the data out asynchronously rather
1220 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1221 int flags, int *rtvals)
1225 vm_ooffset_t maxblksz, next_offset, poffset, prev_offset;
1228 off_t prev_resid, wrsz;
1229 int count, error, i, maxsize, ncount, pgoff, ppscheck;
1231 static struct timeval lastfail;
1234 object = vp->v_object;
1235 count = bytecount / PAGE_SIZE;
1237 for (i = 0; i < count; i++)
1238 rtvals[i] = VM_PAGER_ERROR;
1240 if ((int64_t)ma[0]->pindex < 0) {
1241 printf("vnode_pager_generic_putpages: "
1242 "attempt to write meta-data 0x%jx(%lx)\n",
1243 (uintmax_t)ma[0]->pindex, (u_long)ma[0]->dirty);
1244 rtvals[0] = VM_PAGER_BAD;
1245 return (VM_PAGER_BAD);
1248 maxsize = count * PAGE_SIZE;
1251 poffset = IDX_TO_OFF(ma[0]->pindex);
1254 * If the page-aligned write is larger then the actual file we
1255 * have to invalidate pages occurring beyond the file EOF. However,
1256 * there is an edge case where a file may not be page-aligned where
1257 * the last page is partially invalid. In this case the filesystem
1258 * may not properly clear the dirty bits for the entire page (which
1259 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1260 * With the page locked we are free to fix-up the dirty bits here.
1262 * We do not under any circumstances truncate the valid bits, as
1263 * this will screw up bogus page replacement.
1265 VM_OBJECT_RLOCK(object);
1266 if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1267 if (!VM_OBJECT_TRYUPGRADE(object)) {
1268 VM_OBJECT_RUNLOCK(object);
1269 VM_OBJECT_WLOCK(object);
1270 if (maxsize + poffset <= object->un_pager.vnp.vnp_size)
1273 if (object->un_pager.vnp.vnp_size > poffset) {
1274 maxsize = object->un_pager.vnp.vnp_size - poffset;
1275 ncount = btoc(maxsize);
1276 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1277 pgoff = roundup2(pgoff, DEV_BSIZE);
1280 * If the object is locked and the following
1281 * conditions hold, then the page's dirty
1282 * field cannot be concurrently changed by a
1286 vm_page_assert_sbusied(m);
1287 KASSERT(!pmap_page_is_write_mapped(m),
1288 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1289 MPASS(m->dirty != 0);
1290 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1297 for (i = ncount; i < count; i++)
1298 rtvals[i] = VM_PAGER_BAD;
1300 VM_OBJECT_LOCK_DOWNGRADE(object);
1303 auio.uio_iov = &aiov;
1304 auio.uio_segflg = UIO_NOCOPY;
1305 auio.uio_rw = UIO_WRITE;
1307 maxblksz = roundup2(poffset + maxsize, DEV_BSIZE);
1309 for (prev_offset = poffset; prev_offset < maxblksz;) {
1310 /* Skip clean blocks. */
1311 for (in_hole = true; in_hole && prev_offset < maxblksz;) {
1312 m = ma[OFF_TO_IDX(prev_offset - poffset)];
1313 for (i = vn_off2bidx(prev_offset);
1314 i < sizeof(vm_page_bits_t) * NBBY &&
1315 prev_offset < maxblksz; i++) {
1316 if (vn_dirty_blk(m, prev_offset)) {
1320 prev_offset += DEV_BSIZE;
1326 /* Find longest run of dirty blocks. */
1327 for (next_offset = prev_offset; next_offset < maxblksz;) {
1328 m = ma[OFF_TO_IDX(next_offset - poffset)];
1329 for (i = vn_off2bidx(next_offset);
1330 i < sizeof(vm_page_bits_t) * NBBY &&
1331 next_offset < maxblksz; i++) {
1332 if (!vn_dirty_blk(m, next_offset))
1334 next_offset += DEV_BSIZE;
1338 if (next_offset > poffset + maxsize)
1339 next_offset = poffset + maxsize;
1342 * Getting here requires finding a dirty block in the
1343 * 'skip clean blocks' loop.
1345 MPASS(prev_offset < next_offset);
1347 VM_OBJECT_RUNLOCK(object);
1348 aiov.iov_base = NULL;
1349 auio.uio_iovcnt = 1;
1350 auio.uio_offset = prev_offset;
1351 prev_resid = auio.uio_resid = aiov.iov_len = next_offset -
1353 error = VOP_WRITE(vp, &auio,
1354 vnode_pager_putpages_ioflags(flags), curthread->td_ucred);
1356 wrsz = prev_resid - auio.uio_resid;
1358 if (ppsratecheck(&lastfail, &curfail, 1) != 0) {
1359 vn_printf(vp, "vnode_pager_putpages: "
1360 "zero-length write at %ju resid %zd\n",
1361 auio.uio_offset, auio.uio_resid);
1363 VM_OBJECT_RLOCK(object);
1367 /* Adjust the starting offset for next iteration. */
1368 prev_offset += wrsz;
1369 MPASS(auio.uio_offset == prev_offset);
1372 if (error != 0 && (ppscheck = ppsratecheck(&lastfail,
1374 vn_printf(vp, "vnode_pager_putpages: I/O error %d\n",
1376 if (auio.uio_resid != 0 && (ppscheck != 0 ||
1377 ppsratecheck(&lastfail, &curfail, 1) != 0))
1378 vn_printf(vp, "vnode_pager_putpages: residual I/O %zd "
1379 "at %ju\n", auio.uio_resid,
1380 (uintmax_t)ma[0]->pindex);
1381 VM_OBJECT_RLOCK(object);
1382 if (error != 0 || auio.uio_resid != 0)
1386 /* Mark completely processed pages. */
1387 for (i = 0; i < OFF_TO_IDX(prev_offset - poffset); i++)
1388 rtvals[i] = VM_PAGER_OK;
1389 /* Mark partial EOF page. */
1390 if (prev_offset == poffset + maxsize && (prev_offset & PAGE_MASK) != 0)
1391 rtvals[i++] = VM_PAGER_OK;
1392 /* Unwritten pages in range, free bonus if the page is clean. */
1393 for (; i < ncount; i++)
1394 rtvals[i] = ma[i]->dirty == 0 ? VM_PAGER_OK : VM_PAGER_ERROR;
1395 VM_OBJECT_RUNLOCK(object);
1396 VM_CNT_ADD(v_vnodepgsout, i);
1397 VM_CNT_INC(v_vnodeout);
1402 vnode_pager_putpages_ioflags(int pager_flags)
1407 * Pageouts are already clustered, use IO_ASYNC to force a
1408 * bawrite() rather then a bdwrite() to prevent paging I/O
1409 * from saturating the buffer cache. Dummy-up the sequential
1410 * heuristic to cause large ranges to cluster. If neither
1411 * IO_SYNC or IO_ASYNC is set, the system decides how to
1415 if ((pager_flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) != 0)
1417 else if ((pager_flags & VM_PAGER_CLUSTER_OK) == 0)
1418 ioflags |= IO_ASYNC;
1419 ioflags |= (pager_flags & VM_PAGER_PUT_INVAL) != 0 ? IO_INVAL: 0;
1420 ioflags |= (pager_flags & VM_PAGER_PUT_NOREUSE) != 0 ? IO_NOREUSE : 0;
1421 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1426 * vnode_pager_undirty_pages().
1428 * A helper to mark pages as clean after pageout that was possibly
1429 * done with a short write. The lpos argument specifies the page run
1430 * length in bytes, and the written argument specifies how many bytes
1431 * were actually written. eof is the offset past the last valid byte
1432 * in the vnode using the absolute file position of the first byte in
1433 * the run as the base from which it is computed.
1436 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written, off_t eof,
1440 int i, pos, pos_devb;
1442 if (written == 0 && eof >= lpos)
1444 obj = ma[0]->object;
1445 VM_OBJECT_WLOCK(obj);
1446 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1447 if (pos < trunc_page(written)) {
1448 rtvals[i] = VM_PAGER_OK;
1449 vm_page_undirty(ma[i]);
1451 /* Partially written page. */
1452 rtvals[i] = VM_PAGER_AGAIN;
1453 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1456 if (eof >= lpos) /* avoid truncation */
1458 for (pos = eof, i = OFF_TO_IDX(trunc_page(pos)); pos < lpos; i++) {
1459 if (pos != trunc_page(pos)) {
1461 * The page contains the last valid byte in
1462 * the vnode, mark the rest of the page as
1463 * clean, potentially making the whole page
1466 pos_devb = roundup2(pos & PAGE_MASK, DEV_BSIZE);
1467 vm_page_clear_dirty(ma[i], pos_devb, PAGE_SIZE -
1471 * If the page was cleaned, report the pageout
1472 * on it as successful. msync() no longer
1473 * needs to write out the page, endlessly
1474 * creating write requests and dirty buffers.
1476 if (ma[i]->dirty == 0)
1477 rtvals[i] = VM_PAGER_OK;
1479 pos = round_page(pos);
1481 /* vm_pageout_flush() clears dirty */
1482 rtvals[i] = VM_PAGER_BAD;
1487 VM_OBJECT_WUNLOCK(obj);
1491 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1495 vm_ooffset_t old_wm;
1497 VM_OBJECT_WLOCK(object);
1498 if (object->type != OBJT_VNODE) {
1499 VM_OBJECT_WUNLOCK(object);
1502 old_wm = object->un_pager.vnp.writemappings;
1503 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1504 vp = object->handle;
1505 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1506 ASSERT_VOP_ELOCKED(vp, "v_writecount inc");
1507 VOP_ADD_WRITECOUNT(vp, 1);
1508 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1509 __func__, vp, vp->v_writecount);
1510 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1511 ASSERT_VOP_ELOCKED(vp, "v_writecount dec");
1512 VOP_ADD_WRITECOUNT(vp, -1);
1513 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1514 __func__, vp, vp->v_writecount);
1516 VM_OBJECT_WUNLOCK(object);
1520 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1527 VM_OBJECT_WLOCK(object);
1530 * First, recheck the object type to account for the race when
1531 * the vnode is reclaimed.
1533 if (object->type != OBJT_VNODE) {
1534 VM_OBJECT_WUNLOCK(object);
1539 * Optimize for the case when writemappings is not going to
1543 if (object->un_pager.vnp.writemappings != inc) {
1544 object->un_pager.vnp.writemappings -= inc;
1545 VM_OBJECT_WUNLOCK(object);
1549 vp = object->handle;
1551 VM_OBJECT_WUNLOCK(object);
1553 vn_start_write(vp, &mp, V_WAIT);
1554 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1557 * Decrement the object's writemappings, by swapping the start
1558 * and end arguments for vnode_pager_update_writecount(). If
1559 * there was not a race with vnode reclaimation, then the
1560 * vnode's v_writecount is decremented.
1562 vnode_pager_update_writecount(object, end, start);
1566 vn_finished_write(mp);