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/systm.h>
63 #include <sys/vnode.h>
64 #include <sys/mount.h>
67 #include <sys/vmmeter.h>
68 #include <sys/limits.h>
70 #include <sys/rwlock.h>
71 #include <sys/sf_buf.h>
73 #include <machine/atomic.h>
76 #include <vm/vm_param.h>
77 #include <vm/vm_object.h>
78 #include <vm/vm_page.h>
79 #include <vm/vm_pager.h>
80 #include <vm/vm_map.h>
81 #include <vm/vnode_pager.h>
82 #include <vm/vm_extern.h>
84 static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
85 daddr_t *rtaddress, int *run);
86 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
87 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
88 static void vnode_pager_dealloc(vm_object_t);
89 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *);
90 static int vnode_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
91 int *, vop_getpages_iodone_t, void *);
92 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *);
93 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
94 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t,
95 vm_ooffset_t, struct ucred *cred);
96 static int vnode_pager_generic_getpages_done(struct buf *);
97 static void vnode_pager_generic_getpages_done_async(struct buf *);
99 struct pagerops vnodepagerops = {
100 .pgo_alloc = vnode_pager_alloc,
101 .pgo_dealloc = vnode_pager_dealloc,
102 .pgo_getpages = vnode_pager_getpages,
103 .pgo_getpages_async = vnode_pager_getpages_async,
104 .pgo_putpages = vnode_pager_putpages,
105 .pgo_haspage = vnode_pager_haspage,
108 int vnode_pbuf_freecnt;
109 int vnode_async_pbuf_freecnt;
111 /* Create the VM system backing object for this vnode */
113 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
116 vm_ooffset_t size = isize;
119 if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE)
122 while ((object = vp->v_object) != NULL) {
123 VM_OBJECT_WLOCK(object);
124 if (!(object->flags & OBJ_DEAD)) {
125 VM_OBJECT_WUNLOCK(object);
129 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
130 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vodead", 0);
131 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
135 if (vn_isdisk(vp, NULL)) {
136 size = IDX_TO_OFF(INT_MAX);
138 if (VOP_GETATTR(vp, &va, td->td_ucred))
144 object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred);
146 * Dereference the reference we just created. This assumes
147 * that the object is associated with the vp.
149 VM_OBJECT_WLOCK(object);
151 VM_OBJECT_WUNLOCK(object);
154 KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
160 vnode_destroy_vobject(struct vnode *vp)
162 struct vm_object *obj;
167 ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
168 VM_OBJECT_WLOCK(obj);
169 umtx_shm_object_terminated(obj);
170 if (obj->ref_count == 0) {
172 * don't double-terminate the object
174 if ((obj->flags & OBJ_DEAD) == 0) {
175 vm_object_terminate(obj);
178 * Waiters were already handled during object
179 * termination. The exclusive vnode lock hopefully
180 * prevented new waiters from referencing the dying
183 KASSERT((obj->flags & OBJ_DISCONNECTWNT) == 0,
184 ("OBJ_DISCONNECTWNT set obj %p flags %x",
187 VM_OBJECT_WUNLOCK(obj);
191 * Woe to the process that tries to page now :-).
193 vm_pager_deallocate(obj);
194 VM_OBJECT_WUNLOCK(obj);
196 KASSERT(vp->v_object == NULL, ("vp %p obj %p", vp, vp->v_object));
201 * Allocate (or lookup) pager for a vnode.
202 * Handle is a vnode pointer.
207 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
208 vm_ooffset_t offset, struct ucred *cred)
214 * Pageout to vnode, no can do yet.
219 vp = (struct vnode *) handle;
222 * If the object is being terminated, wait for it to
226 while ((object = vp->v_object) != NULL) {
227 VM_OBJECT_WLOCK(object);
228 if ((object->flags & OBJ_DEAD) == 0)
230 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
231 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vadead", 0);
234 KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference"));
236 if (object == NULL) {
238 * Add an object of the appropriate size
240 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
242 object->un_pager.vnp.vnp_size = size;
243 object->un_pager.vnp.writemappings = 0;
245 object->handle = handle;
247 if (vp->v_object != NULL) {
249 * Object has been created while we were sleeping
252 VM_OBJECT_WLOCK(object);
253 KASSERT(object->ref_count == 1,
254 ("leaked ref %p %d", object, object->ref_count));
255 object->type = OBJT_DEAD;
256 object->ref_count = 0;
257 VM_OBJECT_WUNLOCK(object);
258 vm_object_destroy(object);
261 vp->v_object = object;
265 #if VM_NRESERVLEVEL > 0
266 vm_object_color(object, 0);
268 VM_OBJECT_WUNLOCK(object);
275 * The object must be locked.
278 vnode_pager_dealloc(vm_object_t object)
285 panic("vnode_pager_dealloc: pager already dealloced");
287 VM_OBJECT_ASSERT_WLOCKED(object);
288 vm_object_pip_wait(object, "vnpdea");
289 refs = object->ref_count;
291 object->handle = NULL;
292 object->type = OBJT_DEAD;
293 if (object->flags & OBJ_DISCONNECTWNT) {
294 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
297 ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
298 if (object->un_pager.vnp.writemappings > 0) {
299 object->un_pager.vnp.writemappings = 0;
300 VOP_ADD_WRITECOUNT(vp, -1);
301 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
302 __func__, vp, vp->v_writecount);
306 VM_OBJECT_WUNLOCK(object);
309 VM_OBJECT_WLOCK(object);
313 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
316 struct vnode *vp = object->handle;
322 int pagesperblock, blocksperpage;
324 VM_OBJECT_ASSERT_WLOCKED(object);
326 * If no vp or vp is doomed or marked transparent to VM, we do not
329 if (vp == NULL || vp->v_iflag & VI_DOOMED)
332 * If the offset is beyond end of file we do
335 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
338 bsize = vp->v_mount->mnt_stat.f_iosize;
339 pagesperblock = bsize / PAGE_SIZE;
341 if (pagesperblock > 0) {
342 reqblock = pindex / pagesperblock;
344 blocksperpage = (PAGE_SIZE / bsize);
345 reqblock = pindex * blocksperpage;
347 VM_OBJECT_WUNLOCK(object);
348 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
349 VM_OBJECT_WLOCK(object);
354 if (pagesperblock > 0) {
355 poff = pindex - (reqblock * pagesperblock);
357 *before *= pagesperblock;
362 * The BMAP vop can report a partial block in the
363 * 'after', but must not report blocks after EOF.
364 * Assert the latter, and truncate 'after' in case
367 KASSERT((reqblock + *after) * pagesperblock <
368 roundup2(object->size, pagesperblock),
369 ("%s: reqblock %jd after %d size %ju", __func__,
370 (intmax_t )reqblock, *after,
371 (uintmax_t )object->size));
372 *after *= pagesperblock;
373 *after += pagesperblock - (poff + 1);
374 if (pindex + *after >= object->size)
375 *after = object->size - 1 - pindex;
379 *before /= blocksperpage;
383 *after /= blocksperpage;
390 * Lets the VM system know about a change in size for a file.
391 * We adjust our own internal size and flush any cached pages in
392 * the associated object that are affected by the size change.
394 * Note: this routine may be invoked as a result of a pager put
395 * operation (possibly at object termination time), so we must be careful.
398 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
402 vm_pindex_t nobjsize;
404 if ((object = vp->v_object) == NULL)
406 /* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */
407 VM_OBJECT_WLOCK(object);
408 if (object->type == OBJT_DEAD) {
409 VM_OBJECT_WUNLOCK(object);
412 KASSERT(object->type == OBJT_VNODE,
413 ("not vnode-backed object %p", object));
414 if (nsize == object->un_pager.vnp.vnp_size) {
416 * Hasn't changed size
418 VM_OBJECT_WUNLOCK(object);
421 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
422 if (nsize < object->un_pager.vnp.vnp_size) {
424 * File has shrunk. Toss any cached pages beyond the new EOF.
426 if (nobjsize < object->size)
427 vm_object_page_remove(object, nobjsize, object->size,
430 * this gets rid of garbage at the end of a page that is now
431 * only partially backed by the vnode.
433 * XXX for some reason (I don't know yet), if we take a
434 * completely invalid page and mark it partially valid
435 * it can screw up NFS reads, so we don't allow the case.
437 if ((nsize & PAGE_MASK) &&
438 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
440 int base = (int)nsize & PAGE_MASK;
441 int size = PAGE_SIZE - base;
444 * Clear out partial-page garbage in case
445 * the page has been mapped.
447 pmap_zero_page_area(m, base, size);
450 * Update the valid bits to reflect the blocks that
451 * have been zeroed. Some of these valid bits may
452 * have already been set.
454 vm_page_set_valid_range(m, base, size);
457 * Round "base" to the next block boundary so that the
458 * dirty bit for a partially zeroed block is not
461 base = roundup2(base, DEV_BSIZE);
464 * Clear out partial-page dirty bits.
466 * note that we do not clear out the valid
467 * bits. This would prevent bogus_page
468 * replacement from working properly.
470 vm_page_clear_dirty(m, base, PAGE_SIZE - base);
473 object->un_pager.vnp.vnp_size = nsize;
474 object->size = nobjsize;
475 VM_OBJECT_WUNLOCK(object);
479 * calculate the linear (byte) disk address of specified virtual
483 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
494 if (vp->v_iflag & VI_DOOMED)
497 bsize = vp->v_mount->mnt_stat.f_iosize;
498 vblock = address / bsize;
499 voffset = address % bsize;
501 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
503 if (*rtaddress != -1)
504 *rtaddress += voffset / DEV_BSIZE;
507 *run *= bsize/PAGE_SIZE;
508 *run -= voffset/PAGE_SIZE;
516 * small block filesystem vnode pager input
519 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m)
532 if (vp->v_iflag & VI_DOOMED)
535 bsize = vp->v_mount->mnt_stat.f_iosize;
537 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
539 sf = sf_buf_alloc(m, 0);
541 for (i = 0; i < PAGE_SIZE / bsize; i++) {
542 vm_ooffset_t address;
544 bits = vm_page_bits(i * bsize, bsize);
548 address = IDX_TO_OFF(m->pindex) + i * bsize;
549 if (address >= object->un_pager.vnp.vnp_size) {
552 error = vnode_pager_addr(vp, address, &fileaddr, NULL);
556 if (fileaddr != -1) {
557 bp = getpbuf(&vnode_pbuf_freecnt);
559 /* build a minimal buffer header */
560 bp->b_iocmd = BIO_READ;
561 bp->b_iodone = bdone;
562 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
563 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
564 bp->b_rcred = crhold(curthread->td_ucred);
565 bp->b_wcred = crhold(curthread->td_ucred);
566 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
567 bp->b_blkno = fileaddr;
570 bp->b_bcount = bsize;
571 bp->b_bufsize = bsize;
572 bp->b_runningbufspace = bp->b_bufsize;
573 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
576 bp->b_iooffset = dbtob(bp->b_blkno);
579 bwait(bp, PVM, "vnsrd");
581 if ((bp->b_ioflags & BIO_ERROR) != 0)
585 * free the buffer header back to the swap buffer pool
589 relpbuf(bp, &vnode_pbuf_freecnt);
593 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
594 KASSERT((m->dirty & bits) == 0,
595 ("vnode_pager_input_smlfs: page %p is dirty", m));
596 VM_OBJECT_WLOCK(object);
598 VM_OBJECT_WUNLOCK(object);
602 return VM_PAGER_ERROR;
608 * old style vnode pager input routine
611 vnode_pager_input_old(vm_object_t object, vm_page_t m)
620 VM_OBJECT_ASSERT_WLOCKED(object);
624 * Return failure if beyond current EOF
626 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
630 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
631 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
633 VM_OBJECT_WUNLOCK(object);
636 * Allocate a kernel virtual address and initialize so that
637 * we can use VOP_READ/WRITE routines.
639 sf = sf_buf_alloc(m, 0);
641 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
643 auio.uio_iov = &aiov;
645 auio.uio_offset = IDX_TO_OFF(m->pindex);
646 auio.uio_segflg = UIO_SYSSPACE;
647 auio.uio_rw = UIO_READ;
648 auio.uio_resid = size;
649 auio.uio_td = curthread;
651 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
653 int count = size - auio.uio_resid;
657 else if (count != PAGE_SIZE)
658 bzero((caddr_t)sf_buf_kva(sf) + count,
663 VM_OBJECT_WLOCK(object);
665 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
667 m->valid = VM_PAGE_BITS_ALL;
668 return error ? VM_PAGER_ERROR : VM_PAGER_OK;
672 * generic vnode pager input routine
676 * Local media VFS's that do not implement their own VOP_GETPAGES
677 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
678 * to implement the previous behaviour.
680 * All other FS's should use the bypass to get to the local media
681 * backing vp's VOP_GETPAGES.
684 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
691 VM_OBJECT_WUNLOCK(object);
692 rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead);
693 KASSERT(rtval != EOPNOTSUPP,
694 ("vnode_pager: FS getpages not implemented\n"));
695 VM_OBJECT_WLOCK(object);
700 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
701 int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg)
707 VM_OBJECT_WUNLOCK(object);
708 rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg);
709 KASSERT(rtval != EOPNOTSUPP,
710 ("vnode_pager: FS getpages_async not implemented\n"));
711 VM_OBJECT_WLOCK(object);
716 * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for
717 * local filesystems, where partially valid pages can only occur at
721 vnode_pager_local_getpages(struct vop_getpages_args *ap)
724 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
725 ap->a_rbehind, ap->a_rahead, NULL, NULL));
729 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap)
732 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
733 ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg));
737 * This is now called from local media FS's to operate against their
738 * own vnodes if they fail to implement VOP_GETPAGES.
741 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count,
742 int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg)
751 int bsize, pagesperblock, *freecnt;
752 int error, before, after, rbehind, rahead, poff, i;
753 int bytecount, secmask;
755 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
756 ("%s does not support devices", __func__));
758 if (vp->v_iflag & VI_DOOMED)
759 return (VM_PAGER_BAD);
761 object = vp->v_object;
762 foff = IDX_TO_OFF(m[0]->pindex);
763 bsize = vp->v_mount->mnt_stat.f_iosize;
764 pagesperblock = bsize / PAGE_SIZE;
766 KASSERT(foff < object->un_pager.vnp.vnp_size,
767 ("%s: page %p offset beyond vp %p size", __func__, m[0], vp));
768 KASSERT(count <= sizeof(bp->b_pages),
769 ("%s: requested %d pages", __func__, count));
772 * The last page has valid blocks. Invalid part can only
773 * exist at the end of file, and the page is made fully valid
774 * by zeroing in vm_pager_get_pages().
776 if (m[count - 1]->valid != 0 && --count == 0) {
778 iodone(arg, m, 1, 0);
779 return (VM_PAGER_OK);
783 * Synchronous and asynchronous paging operations use different
784 * free pbuf counters. This is done to avoid asynchronous requests
785 * to consume all pbufs.
786 * Allocate the pbuf at the very beginning of the function, so that
787 * if we are low on certain kind of pbufs don't even proceed to BMAP,
790 freecnt = iodone != NULL ?
791 &vnode_async_pbuf_freecnt : &vnode_pbuf_freecnt;
792 bp = getpbuf(freecnt);
795 * Get the underlying device blocks for the file with VOP_BMAP().
796 * If the file system doesn't support VOP_BMAP, use old way of
797 * getting pages via VOP_READ.
799 error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before);
800 if (error == EOPNOTSUPP) {
801 relpbuf(bp, freecnt);
802 VM_OBJECT_WLOCK(object);
803 for (i = 0; i < count; i++) {
804 VM_CNT_INC(v_vnodein);
805 VM_CNT_INC(v_vnodepgsin);
806 error = vnode_pager_input_old(object, m[i]);
810 VM_OBJECT_WUNLOCK(object);
812 } else if (error != 0) {
813 relpbuf(bp, freecnt);
814 return (VM_PAGER_ERROR);
818 * If the file system supports BMAP, but blocksize is smaller
819 * than a page size, then use special small filesystem code.
821 if (pagesperblock == 0) {
822 relpbuf(bp, freecnt);
823 for (i = 0; i < count; i++) {
824 VM_CNT_INC(v_vnodein);
825 VM_CNT_INC(v_vnodepgsin);
826 error = vnode_pager_input_smlfs(object, m[i]);
834 * A sparse file can be encountered only for a single page request,
835 * which may not be preceded by call to vm_pager_haspage().
837 if (bp->b_blkno == -1) {
839 ("%s: array[%d] request to a sparse file %p", __func__,
841 relpbuf(bp, freecnt);
842 pmap_zero_page(m[0]);
843 KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty",
845 VM_OBJECT_WLOCK(object);
846 m[0]->valid = VM_PAGE_BITS_ALL;
847 VM_OBJECT_WUNLOCK(object);
848 return (VM_PAGER_OK);
852 blkno0 = bp->b_blkno;
854 bp->b_blkno += (foff % bsize) / DEV_BSIZE;
856 /* Recalculate blocks available after/before to pages. */
857 poff = (foff % bsize) / PAGE_SIZE;
858 before *= pagesperblock;
860 after *= pagesperblock;
861 after += pagesperblock - (poff + 1);
862 if (m[0]->pindex + after >= object->size)
863 after = object->size - 1 - m[0]->pindex;
864 KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d",
865 __func__, count, after + 1));
868 /* Trim requested rbehind/rahead to possible values. */
869 rbehind = a_rbehind ? *a_rbehind : 0;
870 rahead = a_rahead ? *a_rahead : 0;
871 rbehind = min(rbehind, before);
872 rbehind = min(rbehind, m[0]->pindex);
873 rahead = min(rahead, after);
874 rahead = min(rahead, object->size - m[count - 1]->pindex);
876 * Check that total amount of pages fit into buf. Trim rbehind and
877 * rahead evenly if not.
879 if (rbehind + rahead + count > nitems(bp->b_pages)) {
882 trim = rbehind + rahead + count - nitems(bp->b_pages) + 1;
883 sum = rbehind + rahead;
884 if (rbehind == before) {
885 /* Roundup rbehind trim to block size. */
886 rbehind -= roundup(trim * rbehind / sum, pagesperblock);
890 rbehind -= trim * rbehind / sum;
891 rahead -= trim * rahead / sum;
893 KASSERT(rbehind + rahead + count <= nitems(bp->b_pages),
894 ("%s: behind %d ahead %d count %d", __func__,
895 rbehind, rahead, count));
898 * Fill in the bp->b_pages[] array with requested and optional
899 * read behind or read ahead pages. Read behind pages are looked
900 * up in a backward direction, down to a first cached page. Same
901 * for read ahead pages, but there is no need to shift the array
902 * in case of encountering a cached page.
904 i = bp->b_npages = 0;
906 vm_pindex_t startpindex, tpindex;
909 VM_OBJECT_WLOCK(object);
910 startpindex = m[0]->pindex - rbehind;
911 if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL &&
912 p->pindex >= startpindex)
913 startpindex = p->pindex + 1;
915 /* tpindex is unsigned; beware of numeric underflow. */
916 for (tpindex = m[0]->pindex - 1;
917 tpindex >= startpindex && tpindex < m[0]->pindex;
919 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
921 /* Shift the array. */
922 for (int j = 0; j < i; j++)
923 bp->b_pages[j] = bp->b_pages[j +
924 tpindex + 1 - startpindex];
927 bp->b_pages[tpindex - startpindex] = p;
932 bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE;
936 /* Requested pages. */
937 for (int j = 0; j < count; j++, i++)
938 bp->b_pages[i] = m[j];
939 bp->b_npages += count;
942 vm_pindex_t endpindex, tpindex;
945 if (!VM_OBJECT_WOWNED(object))
946 VM_OBJECT_WLOCK(object);
947 endpindex = m[count - 1]->pindex + rahead + 1;
948 if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL &&
949 p->pindex < endpindex)
950 endpindex = p->pindex;
951 if (endpindex > object->size)
952 endpindex = object->size;
954 for (tpindex = m[count - 1]->pindex + 1;
955 tpindex < endpindex; i++, tpindex++) {
956 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
962 bp->b_pgafter = i - bp->b_npages;
967 if (VM_OBJECT_WOWNED(object))
968 VM_OBJECT_WUNLOCK(object);
970 /* Report back actual behind/ahead read. */
972 *a_rbehind = bp->b_pgbefore;
974 *a_rahead = bp->b_pgafter;
977 KASSERT(bp->b_npages <= nitems(bp->b_pages),
978 ("%s: buf %p overflowed", __func__, bp));
979 for (int j = 1, prev = 0; j < bp->b_npages; j++) {
980 if (bp->b_pages[j] == bogus_page)
982 KASSERT(bp->b_pages[j]->pindex - bp->b_pages[prev]->pindex ==
983 j - prev, ("%s: pages array not consecutive, bp %p",
990 * Recalculate first offset and bytecount with regards to read behind.
991 * Truncate bytecount to vnode real size and round up physical size
994 foff = IDX_TO_OFF(bp->b_pages[0]->pindex);
995 bytecount = bp->b_npages << PAGE_SHIFT;
996 if ((foff + bytecount) > object->un_pager.vnp.vnp_size)
997 bytecount = object->un_pager.vnp.vnp_size - foff;
998 secmask = bo->bo_bsize - 1;
999 KASSERT(secmask < PAGE_SIZE && secmask > 0,
1000 ("%s: sector size %d too large", __func__, secmask + 1));
1001 bytecount = (bytecount + secmask) & ~secmask;
1004 * And map the pages to be read into the kva, if the filesystem
1005 * requires mapped buffers.
1007 if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
1008 unmapped_buf_allowed) {
1009 bp->b_data = unmapped_buf;
1012 bp->b_data = bp->b_kvabase;
1013 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1016 /* Build a minimal buffer header. */
1017 bp->b_iocmd = BIO_READ;
1018 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
1019 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
1020 bp->b_rcred = crhold(curthread->td_ucred);
1021 bp->b_wcred = crhold(curthread->td_ucred);
1024 bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount;
1025 bp->b_iooffset = dbtob(bp->b_blkno);
1026 KASSERT(IDX_TO_OFF(m[0]->pindex - bp->b_pages[0]->pindex) ==
1027 (blkno0 - bp->b_blkno) * DEV_BSIZE +
1028 IDX_TO_OFF(m[0]->pindex) % bsize,
1029 ("wrong offsets bsize %d m[0] %ju b_pages[0] %ju "
1030 "blkno0 %ju b_blkno %ju", bsize,
1031 (uintmax_t)m[0]->pindex, (uintmax_t)bp->b_pages[0]->pindex,
1032 (uintmax_t)blkno0, (uintmax_t)bp->b_blkno));
1034 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
1035 VM_CNT_INC(v_vnodein);
1036 VM_CNT_ADD(v_vnodepgsin, bp->b_npages);
1038 if (iodone != NULL) { /* async */
1039 bp->b_pgiodone = iodone;
1040 bp->b_caller1 = arg;
1041 bp->b_iodone = vnode_pager_generic_getpages_done_async;
1042 bp->b_flags |= B_ASYNC;
1045 return (VM_PAGER_OK);
1047 bp->b_iodone = bdone;
1049 bwait(bp, PVM, "vnread");
1050 error = vnode_pager_generic_getpages_done(bp);
1051 for (i = 0; i < bp->b_npages; i++)
1052 bp->b_pages[i] = NULL;
1055 relpbuf(bp, &vnode_pbuf_freecnt);
1056 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
1061 vnode_pager_generic_getpages_done_async(struct buf *bp)
1065 error = vnode_pager_generic_getpages_done(bp);
1066 /* Run the iodone upon the requested range. */
1067 bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore,
1068 bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error);
1069 for (int i = 0; i < bp->b_npages; i++)
1070 bp->b_pages[i] = NULL;
1073 relpbuf(bp, &vnode_async_pbuf_freecnt);
1077 vnode_pager_generic_getpages_done(struct buf *bp)
1080 off_t tfoff, nextoff;
1083 error = (bp->b_ioflags & BIO_ERROR) != 0 ? EIO : 0;
1084 object = bp->b_vp->v_object;
1086 if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1087 if (!buf_mapped(bp)) {
1088 bp->b_data = bp->b_kvabase;
1089 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages,
1092 bzero(bp->b_data + bp->b_bcount,
1093 PAGE_SIZE * bp->b_npages - bp->b_bcount);
1095 if (buf_mapped(bp)) {
1096 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1097 bp->b_data = unmapped_buf;
1100 VM_OBJECT_WLOCK(object);
1101 for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1102 i < bp->b_npages; i++, tfoff = nextoff) {
1105 nextoff = tfoff + PAGE_SIZE;
1106 mt = bp->b_pages[i];
1108 if (nextoff <= object->un_pager.vnp.vnp_size) {
1110 * Read filled up entire page.
1112 mt->valid = VM_PAGE_BITS_ALL;
1113 KASSERT(mt->dirty == 0,
1114 ("%s: page %p is dirty", __func__, mt));
1115 KASSERT(!pmap_page_is_mapped(mt),
1116 ("%s: page %p is mapped", __func__, mt));
1119 * Read did not fill up entire page.
1121 * Currently we do not set the entire page valid,
1122 * we just try to clear the piece that we couldn't
1125 vm_page_set_valid_range(mt, 0,
1126 object->un_pager.vnp.vnp_size - tfoff);
1127 KASSERT((mt->dirty & vm_page_bits(0,
1128 object->un_pager.vnp.vnp_size - tfoff)) == 0,
1129 ("%s: page %p is dirty", __func__, mt));
1132 if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter)
1133 vm_page_readahead_finish(mt);
1135 VM_OBJECT_WUNLOCK(object);
1137 printf("%s: I/O read error %d\n", __func__, error);
1143 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
1144 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1145 * vnode_pager_generic_putpages() to implement the previous behaviour.
1147 * All other FS's should use the bypass to get to the local media
1148 * backing vp's VOP_PUTPAGES.
1151 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1152 int flags, int *rtvals)
1156 int bytes = count * PAGE_SIZE;
1159 * Force synchronous operation if we are extremely low on memory
1160 * to prevent a low-memory deadlock. VOP operations often need to
1161 * allocate more memory to initiate the I/O ( i.e. do a BMAP
1162 * operation ). The swapper handles the case by limiting the amount
1163 * of asynchronous I/O, but that sort of solution doesn't scale well
1164 * for the vnode pager without a lot of work.
1166 * Also, the backing vnode's iodone routine may not wake the pageout
1167 * daemon up. This should be probably be addressed XXX.
1170 if (vm_page_count_min())
1171 flags |= VM_PAGER_PUT_SYNC;
1174 * Call device-specific putpages function
1176 vp = object->handle;
1177 VM_OBJECT_WUNLOCK(object);
1178 rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1179 KASSERT(rtval != EOPNOTSUPP,
1180 ("vnode_pager: stale FS putpages\n"));
1181 VM_OBJECT_WLOCK(object);
1185 vn_off2bidx(vm_ooffset_t offset)
1188 return ((offset & PAGE_MASK) / DEV_BSIZE);
1192 vn_dirty_blk(vm_page_t m, vm_ooffset_t offset)
1195 KASSERT(IDX_TO_OFF(m->pindex) <= offset &&
1196 offset < IDX_TO_OFF(m->pindex + 1),
1197 ("page %p pidx %ju offset %ju", m, (uintmax_t)m->pindex,
1198 (uintmax_t)offset));
1199 return ((m->dirty & ((vm_page_bits_t)1 << vn_off2bidx(offset))) != 0);
1203 * This is now called from local media FS's to operate against their
1204 * own vnodes if they fail to implement VOP_PUTPAGES.
1206 * This is typically called indirectly via the pageout daemon and
1207 * clustering has already typically occurred, so in general we ask the
1208 * underlying filesystem to write the data out asynchronously rather
1212 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1213 int flags, int *rtvals)
1217 vm_ooffset_t maxblksz, next_offset, poffset, prev_offset;
1220 off_t prev_resid, wrsz;
1221 int count, error, i, maxsize, ncount, pgoff, ppscheck;
1223 static struct timeval lastfail;
1226 object = vp->v_object;
1227 count = bytecount / PAGE_SIZE;
1229 for (i = 0; i < count; i++)
1230 rtvals[i] = VM_PAGER_ERROR;
1232 if ((int64_t)ma[0]->pindex < 0) {
1233 printf("vnode_pager_generic_putpages: "
1234 "attempt to write meta-data 0x%jx(%lx)\n",
1235 (uintmax_t)ma[0]->pindex, (u_long)ma[0]->dirty);
1236 rtvals[0] = VM_PAGER_BAD;
1237 return (VM_PAGER_BAD);
1240 maxsize = count * PAGE_SIZE;
1243 poffset = IDX_TO_OFF(ma[0]->pindex);
1246 * If the page-aligned write is larger then the actual file we
1247 * have to invalidate pages occurring beyond the file EOF. However,
1248 * there is an edge case where a file may not be page-aligned where
1249 * the last page is partially invalid. In this case the filesystem
1250 * may not properly clear the dirty bits for the entire page (which
1251 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1252 * With the page locked we are free to fix-up the dirty bits here.
1254 * We do not under any circumstances truncate the valid bits, as
1255 * this will screw up bogus page replacement.
1257 VM_OBJECT_RLOCK(object);
1258 if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1259 if (!VM_OBJECT_TRYUPGRADE(object)) {
1260 VM_OBJECT_RUNLOCK(object);
1261 VM_OBJECT_WLOCK(object);
1262 if (maxsize + poffset <= object->un_pager.vnp.vnp_size)
1265 if (object->un_pager.vnp.vnp_size > poffset) {
1266 maxsize = object->un_pager.vnp.vnp_size - poffset;
1267 ncount = btoc(maxsize);
1268 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1269 pgoff = roundup2(pgoff, DEV_BSIZE);
1272 * If the object is locked and the following
1273 * conditions hold, then the page's dirty
1274 * field cannot be concurrently changed by a
1278 vm_page_assert_sbusied(m);
1279 KASSERT(!pmap_page_is_write_mapped(m),
1280 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1281 MPASS(m->dirty != 0);
1282 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1289 for (i = ncount; i < count; i++)
1290 rtvals[i] = VM_PAGER_BAD;
1292 VM_OBJECT_LOCK_DOWNGRADE(object);
1295 auio.uio_iov = &aiov;
1296 auio.uio_segflg = UIO_NOCOPY;
1297 auio.uio_rw = UIO_WRITE;
1299 maxblksz = roundup2(poffset + maxsize, DEV_BSIZE);
1301 for (prev_offset = poffset; prev_offset < maxblksz;) {
1302 /* Skip clean blocks. */
1303 for (in_hole = true; in_hole && prev_offset < maxblksz;) {
1304 m = ma[OFF_TO_IDX(prev_offset - poffset)];
1305 for (i = vn_off2bidx(prev_offset);
1306 i < sizeof(vm_page_bits_t) * NBBY &&
1307 prev_offset < maxblksz; i++) {
1308 if (vn_dirty_blk(m, prev_offset)) {
1312 prev_offset += DEV_BSIZE;
1318 /* Find longest run of dirty blocks. */
1319 for (next_offset = prev_offset; next_offset < maxblksz;) {
1320 m = ma[OFF_TO_IDX(next_offset - poffset)];
1321 for (i = vn_off2bidx(next_offset);
1322 i < sizeof(vm_page_bits_t) * NBBY &&
1323 next_offset < maxblksz; i++) {
1324 if (!vn_dirty_blk(m, next_offset))
1326 next_offset += DEV_BSIZE;
1330 if (next_offset > poffset + maxsize)
1331 next_offset = poffset + maxsize;
1334 * Getting here requires finding a dirty block in the
1335 * 'skip clean blocks' loop.
1337 MPASS(prev_offset < next_offset);
1339 VM_OBJECT_RUNLOCK(object);
1340 aiov.iov_base = NULL;
1341 auio.uio_iovcnt = 1;
1342 auio.uio_offset = prev_offset;
1343 prev_resid = auio.uio_resid = aiov.iov_len = next_offset -
1345 error = VOP_WRITE(vp, &auio,
1346 vnode_pager_putpages_ioflags(flags), curthread->td_ucred);
1348 wrsz = prev_resid - auio.uio_resid;
1350 if (ppsratecheck(&lastfail, &curfail, 1) != 0) {
1351 vn_printf(vp, "vnode_pager_putpages: "
1352 "zero-length write at %ju resid %zd\n",
1353 auio.uio_offset, auio.uio_resid);
1355 VM_OBJECT_RLOCK(object);
1359 /* Adjust the starting offset for next iteration. */
1360 prev_offset += wrsz;
1361 MPASS(auio.uio_offset == prev_offset);
1364 if (error != 0 && (ppscheck = ppsratecheck(&lastfail,
1366 vn_printf(vp, "vnode_pager_putpages: I/O error %d\n",
1368 if (auio.uio_resid != 0 && (ppscheck != 0 ||
1369 ppsratecheck(&lastfail, &curfail, 1) != 0))
1370 vn_printf(vp, "vnode_pager_putpages: residual I/O %zd "
1371 "at %ju\n", auio.uio_resid,
1372 (uintmax_t)ma[0]->pindex);
1373 VM_OBJECT_RLOCK(object);
1374 if (error != 0 || auio.uio_resid != 0)
1378 /* Mark completely processed pages. */
1379 for (i = 0; i < OFF_TO_IDX(prev_offset - poffset); i++)
1380 rtvals[i] = VM_PAGER_OK;
1381 /* Mark partial EOF page. */
1382 if (prev_offset == poffset + maxsize && (prev_offset & PAGE_MASK) != 0)
1383 rtvals[i++] = VM_PAGER_OK;
1384 /* Unwritten pages in range, free bonus if the page is clean. */
1385 for (; i < ncount; i++)
1386 rtvals[i] = ma[i]->dirty == 0 ? VM_PAGER_OK : VM_PAGER_ERROR;
1387 VM_OBJECT_RUNLOCK(object);
1388 VM_CNT_ADD(v_vnodepgsout, i);
1389 VM_CNT_INC(v_vnodeout);
1394 vnode_pager_putpages_ioflags(int pager_flags)
1399 * Pageouts are already clustered, use IO_ASYNC to force a
1400 * bawrite() rather then a bdwrite() to prevent paging I/O
1401 * from saturating the buffer cache. Dummy-up the sequential
1402 * heuristic to cause large ranges to cluster. If neither
1403 * IO_SYNC or IO_ASYNC is set, the system decides how to
1407 if ((pager_flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) != 0)
1409 else if ((pager_flags & VM_PAGER_CLUSTER_OK) == 0)
1410 ioflags |= IO_ASYNC;
1411 ioflags |= (pager_flags & VM_PAGER_PUT_INVAL) != 0 ? IO_INVAL: 0;
1412 ioflags |= (pager_flags & VM_PAGER_PUT_NOREUSE) != 0 ? IO_NOREUSE : 0;
1413 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1418 * vnode_pager_undirty_pages().
1420 * A helper to mark pages as clean after pageout that was possibly
1421 * done with a short write. The lpos argument specifies the page run
1422 * length in bytes, and the written argument specifies how many bytes
1423 * were actually written. eof is the offset past the last valid byte
1424 * in the vnode using the absolute file position of the first byte in
1425 * the run as the base from which it is computed.
1428 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written, off_t eof,
1432 int i, pos, pos_devb;
1434 if (written == 0 && eof >= lpos)
1436 obj = ma[0]->object;
1437 VM_OBJECT_WLOCK(obj);
1438 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1439 if (pos < trunc_page(written)) {
1440 rtvals[i] = VM_PAGER_OK;
1441 vm_page_undirty(ma[i]);
1443 /* Partially written page. */
1444 rtvals[i] = VM_PAGER_AGAIN;
1445 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1448 if (eof >= lpos) /* avoid truncation */
1450 for (pos = eof, i = OFF_TO_IDX(trunc_page(pos)); pos < lpos; i++) {
1451 if (pos != trunc_page(pos)) {
1453 * The page contains the last valid byte in
1454 * the vnode, mark the rest of the page as
1455 * clean, potentially making the whole page
1458 pos_devb = roundup2(pos & PAGE_MASK, DEV_BSIZE);
1459 vm_page_clear_dirty(ma[i], pos_devb, PAGE_SIZE -
1463 * If the page was cleaned, report the pageout
1464 * on it as successful. msync() no longer
1465 * needs to write out the page, endlessly
1466 * creating write requests and dirty buffers.
1468 if (ma[i]->dirty == 0)
1469 rtvals[i] = VM_PAGER_OK;
1471 pos = round_page(pos);
1473 /* vm_pageout_flush() clears dirty */
1474 rtvals[i] = VM_PAGER_BAD;
1479 VM_OBJECT_WUNLOCK(obj);
1483 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1487 vm_ooffset_t old_wm;
1489 VM_OBJECT_WLOCK(object);
1490 if (object->type != OBJT_VNODE) {
1491 VM_OBJECT_WUNLOCK(object);
1494 old_wm = object->un_pager.vnp.writemappings;
1495 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1496 vp = object->handle;
1497 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1498 ASSERT_VOP_ELOCKED(vp, "v_writecount inc");
1499 VOP_ADD_WRITECOUNT(vp, 1);
1500 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1501 __func__, vp, vp->v_writecount);
1502 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1503 ASSERT_VOP_ELOCKED(vp, "v_writecount dec");
1504 VOP_ADD_WRITECOUNT(vp, -1);
1505 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1506 __func__, vp, vp->v_writecount);
1508 VM_OBJECT_WUNLOCK(object);
1512 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1519 VM_OBJECT_WLOCK(object);
1522 * First, recheck the object type to account for the race when
1523 * the vnode is reclaimed.
1525 if (object->type != OBJT_VNODE) {
1526 VM_OBJECT_WUNLOCK(object);
1531 * Optimize for the case when writemappings is not going to
1535 if (object->un_pager.vnp.writemappings != inc) {
1536 object->un_pager.vnp.writemappings -= inc;
1537 VM_OBJECT_WUNLOCK(object);
1541 vp = object->handle;
1543 VM_OBJECT_WUNLOCK(object);
1545 vn_start_write(vp, &mp, V_WAIT);
1546 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1549 * Decrement the object's writemappings, by swapping the start
1550 * and end arguments for vnode_pager_update_writecount(). If
1551 * there was not a race with vnode reclaimation, then the
1552 * vnode's v_writecount is decremented.
1554 vnode_pager_update_writecount(object, end, start);
1558 vn_finished_write(mp);