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);
176 * Waiters were already handled during object
177 * termination. The exclusive vnode lock hopefully
178 * prevented new waiters from referencing the dying
181 KASSERT((obj->flags & OBJ_DISCONNECTWNT) == 0,
182 ("OBJ_DISCONNECTWNT set obj %p flags %x",
185 VM_OBJECT_WUNLOCK(obj);
189 * Woe to the process that tries to page now :-).
191 vm_pager_deallocate(obj);
192 VM_OBJECT_WUNLOCK(obj);
194 KASSERT(vp->v_object == NULL, ("vp %p obj %p", vp, vp->v_object));
199 * Allocate (or lookup) pager for a vnode.
200 * Handle is a vnode pointer.
205 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
206 vm_ooffset_t offset, struct ucred *cred)
212 * Pageout to vnode, no can do yet.
217 vp = (struct vnode *) handle;
220 * If the object is being terminated, wait for it to
224 while ((object = vp->v_object) != NULL) {
225 VM_OBJECT_WLOCK(object);
226 if ((object->flags & OBJ_DEAD) == 0)
228 vm_object_set_flag(object, OBJ_DISCONNECTWNT);
229 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vadead", 0);
232 KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference"));
234 if (object == NULL) {
236 * Add an object of the appropriate size
238 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
240 object->un_pager.vnp.vnp_size = size;
241 object->un_pager.vnp.writemappings = 0;
243 object->handle = handle;
245 if (vp->v_object != NULL) {
247 * Object has been created while we were sleeping
250 VM_OBJECT_WLOCK(object);
251 KASSERT(object->ref_count == 1,
252 ("leaked ref %p %d", object, object->ref_count));
253 object->type = OBJT_DEAD;
254 object->ref_count = 0;
255 VM_OBJECT_WUNLOCK(object);
256 vm_object_destroy(object);
259 vp->v_object = object;
263 #if VM_NRESERVLEVEL > 0
264 vm_object_color(object, 0);
266 VM_OBJECT_WUNLOCK(object);
273 * The object must be locked.
276 vnode_pager_dealloc(vm_object_t object)
283 panic("vnode_pager_dealloc: pager already dealloced");
285 VM_OBJECT_ASSERT_WLOCKED(object);
286 vm_object_pip_wait(object, "vnpdea");
287 refs = object->ref_count;
289 object->handle = NULL;
290 object->type = OBJT_DEAD;
291 if (object->flags & OBJ_DISCONNECTWNT) {
292 vm_object_clear_flag(object, OBJ_DISCONNECTWNT);
295 ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
296 if (object->un_pager.vnp.writemappings > 0) {
297 object->un_pager.vnp.writemappings = 0;
298 VOP_ADD_WRITECOUNT(vp, -1);
299 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
300 __func__, vp, vp->v_writecount);
304 VM_OBJECT_WUNLOCK(object);
307 VM_OBJECT_WLOCK(object);
311 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
314 struct vnode *vp = object->handle;
320 int pagesperblock, blocksperpage;
322 VM_OBJECT_ASSERT_WLOCKED(object);
324 * If no vp or vp is doomed or marked transparent to VM, we do not
327 if (vp == NULL || vp->v_iflag & VI_DOOMED)
330 * If the offset is beyond end of file we do
333 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
336 bsize = vp->v_mount->mnt_stat.f_iosize;
337 pagesperblock = bsize / PAGE_SIZE;
339 if (pagesperblock > 0) {
340 reqblock = pindex / pagesperblock;
342 blocksperpage = (PAGE_SIZE / bsize);
343 reqblock = pindex * blocksperpage;
345 VM_OBJECT_WUNLOCK(object);
346 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
347 VM_OBJECT_WLOCK(object);
352 if (pagesperblock > 0) {
353 poff = pindex - (reqblock * pagesperblock);
355 *before *= pagesperblock;
360 * The BMAP vop can report a partial block in the
361 * 'after', but must not report blocks after EOF.
362 * Assert the latter, and truncate 'after' in case
365 KASSERT((reqblock + *after) * pagesperblock <
366 roundup2(object->size, pagesperblock),
367 ("%s: reqblock %jd after %d size %ju", __func__,
368 (intmax_t )reqblock, *after,
369 (uintmax_t )object->size));
370 *after *= pagesperblock;
371 *after += pagesperblock - (poff + 1);
372 if (pindex + *after >= object->size)
373 *after = object->size - 1 - pindex;
377 *before /= blocksperpage;
381 *after /= blocksperpage;
388 * Lets the VM system know about a change in size for a file.
389 * We adjust our own internal size and flush any cached pages in
390 * the associated object that are affected by the size change.
392 * Note: this routine may be invoked as a result of a pager put
393 * operation (possibly at object termination time), so we must be careful.
396 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
400 vm_pindex_t nobjsize;
402 if ((object = vp->v_object) == NULL)
404 /* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */
405 VM_OBJECT_WLOCK(object);
406 if (object->type == OBJT_DEAD) {
407 VM_OBJECT_WUNLOCK(object);
410 KASSERT(object->type == OBJT_VNODE,
411 ("not vnode-backed object %p", object));
412 if (nsize == object->un_pager.vnp.vnp_size) {
414 * Hasn't changed size
416 VM_OBJECT_WUNLOCK(object);
419 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
420 if (nsize < object->un_pager.vnp.vnp_size) {
422 * File has shrunk. Toss any cached pages beyond the new EOF.
424 if (nobjsize < object->size)
425 vm_object_page_remove(object, nobjsize, object->size,
428 * this gets rid of garbage at the end of a page that is now
429 * only partially backed by the vnode.
431 * XXX for some reason (I don't know yet), if we take a
432 * completely invalid page and mark it partially valid
433 * it can screw up NFS reads, so we don't allow the case.
435 if ((nsize & PAGE_MASK) &&
436 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
438 int base = (int)nsize & PAGE_MASK;
439 int size = PAGE_SIZE - base;
442 * Clear out partial-page garbage in case
443 * the page has been mapped.
445 pmap_zero_page_area(m, base, size);
448 * Update the valid bits to reflect the blocks that
449 * have been zeroed. Some of these valid bits may
450 * have already been set.
452 vm_page_set_valid_range(m, base, size);
455 * Round "base" to the next block boundary so that the
456 * dirty bit for a partially zeroed block is not
459 base = roundup2(base, DEV_BSIZE);
462 * Clear out partial-page dirty bits.
464 * note that we do not clear out the valid
465 * bits. This would prevent bogus_page
466 * replacement from working properly.
468 vm_page_clear_dirty(m, base, PAGE_SIZE - base);
469 } else if ((nsize & PAGE_MASK) &&
470 vm_page_is_cached(object, OFF_TO_IDX(nsize))) {
471 vm_page_cache_free(object, OFF_TO_IDX(nsize),
475 object->un_pager.vnp.vnp_size = nsize;
476 object->size = nobjsize;
477 VM_OBJECT_WUNLOCK(object);
481 * calculate the linear (byte) disk address of specified virtual
485 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
496 if (vp->v_iflag & VI_DOOMED)
499 bsize = vp->v_mount->mnt_stat.f_iosize;
500 vblock = address / bsize;
501 voffset = address % bsize;
503 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
505 if (*rtaddress != -1)
506 *rtaddress += voffset / DEV_BSIZE;
509 *run *= bsize/PAGE_SIZE;
510 *run -= voffset/PAGE_SIZE;
518 * small block filesystem vnode pager input
521 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m)
534 if (vp->v_iflag & VI_DOOMED)
537 bsize = vp->v_mount->mnt_stat.f_iosize;
539 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
541 sf = sf_buf_alloc(m, 0);
543 for (i = 0; i < PAGE_SIZE / bsize; i++) {
544 vm_ooffset_t address;
546 bits = vm_page_bits(i * bsize, bsize);
550 address = IDX_TO_OFF(m->pindex) + i * bsize;
551 if (address >= object->un_pager.vnp.vnp_size) {
554 error = vnode_pager_addr(vp, address, &fileaddr, NULL);
558 if (fileaddr != -1) {
559 bp = getpbuf(&vnode_pbuf_freecnt);
561 /* build a minimal buffer header */
562 bp->b_iocmd = BIO_READ;
563 bp->b_iodone = bdone;
564 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
565 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
566 bp->b_rcred = crhold(curthread->td_ucred);
567 bp->b_wcred = crhold(curthread->td_ucred);
568 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
569 bp->b_blkno = fileaddr;
572 bp->b_bcount = bsize;
573 bp->b_bufsize = bsize;
574 bp->b_runningbufspace = bp->b_bufsize;
575 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
578 bp->b_iooffset = dbtob(bp->b_blkno);
581 bwait(bp, PVM, "vnsrd");
583 if ((bp->b_ioflags & BIO_ERROR) != 0)
587 * free the buffer header back to the swap buffer pool
591 relpbuf(bp, &vnode_pbuf_freecnt);
595 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
596 KASSERT((m->dirty & bits) == 0,
597 ("vnode_pager_input_smlfs: page %p is dirty", m));
598 VM_OBJECT_WLOCK(object);
600 VM_OBJECT_WUNLOCK(object);
604 return VM_PAGER_ERROR;
610 * old style vnode pager input routine
613 vnode_pager_input_old(vm_object_t object, vm_page_t m)
622 VM_OBJECT_ASSERT_WLOCKED(object);
626 * Return failure if beyond current EOF
628 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
632 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
633 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
635 VM_OBJECT_WUNLOCK(object);
638 * Allocate a kernel virtual address and initialize so that
639 * we can use VOP_READ/WRITE routines.
641 sf = sf_buf_alloc(m, 0);
643 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
645 auio.uio_iov = &aiov;
647 auio.uio_offset = IDX_TO_OFF(m->pindex);
648 auio.uio_segflg = UIO_SYSSPACE;
649 auio.uio_rw = UIO_READ;
650 auio.uio_resid = size;
651 auio.uio_td = curthread;
653 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
655 int count = size - auio.uio_resid;
659 else if (count != PAGE_SIZE)
660 bzero((caddr_t)sf_buf_kva(sf) + count,
665 VM_OBJECT_WLOCK(object);
667 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
669 m->valid = VM_PAGE_BITS_ALL;
670 return error ? VM_PAGER_ERROR : VM_PAGER_OK;
674 * generic vnode pager input routine
678 * Local media VFS's that do not implement their own VOP_GETPAGES
679 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
680 * to implement the previous behaviour.
682 * All other FS's should use the bypass to get to the local media
683 * backing vp's VOP_GETPAGES.
686 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
693 VM_OBJECT_WUNLOCK(object);
694 rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead);
695 KASSERT(rtval != EOPNOTSUPP,
696 ("vnode_pager: FS getpages not implemented\n"));
697 VM_OBJECT_WLOCK(object);
702 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
703 int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg)
709 VM_OBJECT_WUNLOCK(object);
710 rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg);
711 KASSERT(rtval != EOPNOTSUPP,
712 ("vnode_pager: FS getpages_async not implemented\n"));
713 VM_OBJECT_WLOCK(object);
718 * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for
719 * local filesystems, where partially valid pages can only occur at
723 vnode_pager_local_getpages(struct vop_getpages_args *ap)
726 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
727 ap->a_rbehind, ap->a_rahead, NULL, NULL));
731 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap)
734 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
735 ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg));
739 * This is now called from local media FS's to operate against their
740 * own vnodes if they fail to implement VOP_GETPAGES.
743 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count,
744 int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg)
750 int bsize, pagesperblock, *freecnt;
751 int error, before, after, rbehind, rahead, poff, i;
752 int bytecount, secmask;
754 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
755 ("%s does not support devices", __func__));
757 if (vp->v_iflag & VI_DOOMED)
758 return (VM_PAGER_BAD);
760 object = vp->v_object;
761 foff = IDX_TO_OFF(m[0]->pindex);
762 bsize = vp->v_mount->mnt_stat.f_iosize;
763 pagesperblock = bsize / PAGE_SIZE;
765 KASSERT(foff < object->un_pager.vnp.vnp_size,
766 ("%s: page %p offset beyond vp %p size", __func__, m[0], vp));
767 KASSERT(count <= sizeof(bp->b_pages),
768 ("%s: requested %d pages", __func__, count));
771 * The last page has valid blocks. Invalid part can only
772 * exist at the end of file, and the page is made fully valid
773 * by zeroing in vm_pager_get_pages().
775 if (m[count - 1]->valid != 0 && --count == 0) {
777 iodone(arg, m, 1, 0);
778 return (VM_PAGER_OK);
782 * Synchronous and asynchronous paging operations use different
783 * free pbuf counters. This is done to avoid asynchronous requests
784 * to consume all pbufs.
785 * Allocate the pbuf at the very beginning of the function, so that
786 * if we are low on certain kind of pbufs don't even proceed to BMAP,
789 freecnt = iodone != NULL ?
790 &vnode_async_pbuf_freecnt : &vnode_pbuf_freecnt;
791 bp = getpbuf(freecnt);
794 * Get the underlying device blocks for the file with VOP_BMAP().
795 * If the file system doesn't support VOP_BMAP, use old way of
796 * getting pages via VOP_READ.
798 error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before);
799 if (error == EOPNOTSUPP) {
800 relpbuf(bp, freecnt);
801 VM_OBJECT_WLOCK(object);
802 for (i = 0; i < count; i++) {
803 PCPU_INC(cnt.v_vnodein);
804 PCPU_INC(cnt.v_vnodepgsin);
805 error = vnode_pager_input_old(object, m[i]);
809 VM_OBJECT_WUNLOCK(object);
811 } else if (error != 0) {
812 relpbuf(bp, freecnt);
813 return (VM_PAGER_ERROR);
817 * If the file system supports BMAP, but blocksize is smaller
818 * than a page size, then use special small filesystem code.
820 if (pagesperblock == 0) {
821 relpbuf(bp, freecnt);
822 for (i = 0; i < count; i++) {
823 PCPU_INC(cnt.v_vnodein);
824 PCPU_INC(cnt.v_vnodepgsin);
825 error = vnode_pager_input_smlfs(object, m[i]);
833 * A sparse file can be encountered only for a single page request,
834 * which may not be preceded by call to vm_pager_haspage().
836 if (bp->b_blkno == -1) {
838 ("%s: array[%d] request to a sparse file %p", __func__,
840 relpbuf(bp, freecnt);
841 pmap_zero_page(m[0]);
842 KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty",
844 VM_OBJECT_WLOCK(object);
845 m[0]->valid = VM_PAGE_BITS_ALL;
846 VM_OBJECT_WUNLOCK(object);
847 return (VM_PAGER_OK);
850 bp->b_blkno += (foff % bsize) / DEV_BSIZE;
852 /* Recalculate blocks available after/before to pages. */
853 poff = (foff % bsize) / PAGE_SIZE;
854 before *= pagesperblock;
856 after *= pagesperblock;
857 after += pagesperblock - (poff + 1);
858 if (m[0]->pindex + after >= object->size)
859 after = object->size - 1 - m[0]->pindex;
860 KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d",
861 __func__, count, after + 1));
864 /* Trim requested rbehind/rahead to possible values. */
865 rbehind = a_rbehind ? *a_rbehind : 0;
866 rahead = a_rahead ? *a_rahead : 0;
867 rbehind = min(rbehind, before);
868 rbehind = min(rbehind, m[0]->pindex);
869 rahead = min(rahead, after);
870 rahead = min(rahead, object->size - m[count - 1]->pindex);
871 KASSERT(rbehind + rahead + count <= sizeof(bp->b_pages),
872 ("%s: behind %d ahead %d count %d", __func__,
873 rbehind, rahead, count));
876 * Fill in the bp->b_pages[] array with requested and optional
877 * read behind or read ahead pages. Read behind pages are looked
878 * up in a backward direction, down to a first cached page. Same
879 * for read ahead pages, but there is no need to shift the array
880 * in case of encountering a cached page.
882 i = bp->b_npages = 0;
884 vm_pindex_t startpindex, tpindex;
887 VM_OBJECT_WLOCK(object);
888 startpindex = m[0]->pindex - rbehind;
889 if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL &&
890 p->pindex >= startpindex)
891 startpindex = p->pindex + 1;
893 /* tpindex is unsigned; beware of numeric underflow. */
894 for (tpindex = m[0]->pindex - 1;
895 tpindex >= startpindex && tpindex < m[0]->pindex;
897 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
898 VM_ALLOC_IFNOTCACHED);
900 /* Shift the array. */
901 for (int j = 0; j < i; j++)
902 bp->b_pages[j] = bp->b_pages[j +
903 tpindex + 1 - startpindex];
906 bp->b_pages[tpindex - startpindex] = p;
911 bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE;
915 /* Requested pages. */
916 for (int j = 0; j < count; j++, i++)
917 bp->b_pages[i] = m[j];
918 bp->b_npages += count;
921 vm_pindex_t endpindex, tpindex;
924 if (!VM_OBJECT_WOWNED(object))
925 VM_OBJECT_WLOCK(object);
926 endpindex = m[count - 1]->pindex + rahead + 1;
927 if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL &&
928 p->pindex < endpindex)
929 endpindex = p->pindex;
930 if (endpindex > object->size)
931 endpindex = object->size;
933 for (tpindex = m[count - 1]->pindex + 1;
934 tpindex < endpindex; i++, tpindex++) {
935 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
936 VM_ALLOC_IFNOTCACHED);
942 bp->b_pgafter = i - bp->b_npages;
947 if (VM_OBJECT_WOWNED(object))
948 VM_OBJECT_WUNLOCK(object);
950 /* Report back actual behind/ahead read. */
952 *a_rbehind = bp->b_pgbefore;
954 *a_rahead = bp->b_pgafter;
956 KASSERT(bp->b_npages <= sizeof(bp->b_pages),
957 ("%s: buf %p overflowed", __func__, bp));
960 * Recalculate first offset and bytecount with regards to read behind.
961 * Truncate bytecount to vnode real size and round up physical size
964 foff = IDX_TO_OFF(bp->b_pages[0]->pindex);
965 bytecount = bp->b_npages << PAGE_SHIFT;
966 if ((foff + bytecount) > object->un_pager.vnp.vnp_size)
967 bytecount = object->un_pager.vnp.vnp_size - foff;
968 secmask = bo->bo_bsize - 1;
969 KASSERT(secmask < PAGE_SIZE && secmask > 0,
970 ("%s: sector size %d too large", __func__, secmask + 1));
971 bytecount = (bytecount + secmask) & ~secmask;
974 * And map the pages to be read into the kva, if the filesystem
975 * requires mapped buffers.
977 if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
978 unmapped_buf_allowed) {
979 bp->b_data = unmapped_buf;
982 bp->b_data = bp->b_kvabase;
983 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
986 /* Build a minimal buffer header. */
987 bp->b_iocmd = BIO_READ;
988 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
989 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
990 bp->b_rcred = crhold(curthread->td_ucred);
991 bp->b_wcred = crhold(curthread->td_ucred);
994 bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount;
995 bp->b_iooffset = dbtob(bp->b_blkno);
997 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
998 PCPU_INC(cnt.v_vnodein);
999 PCPU_ADD(cnt.v_vnodepgsin, bp->b_npages);
1001 if (iodone != NULL) { /* async */
1002 bp->b_pgiodone = iodone;
1003 bp->b_caller1 = arg;
1004 bp->b_iodone = vnode_pager_generic_getpages_done_async;
1005 bp->b_flags |= B_ASYNC;
1008 return (VM_PAGER_OK);
1010 bp->b_iodone = bdone;
1012 bwait(bp, PVM, "vnread");
1013 error = vnode_pager_generic_getpages_done(bp);
1014 for (i = 0; i < bp->b_npages; i++)
1015 bp->b_pages[i] = NULL;
1018 relpbuf(bp, &vnode_pbuf_freecnt);
1019 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
1024 vnode_pager_generic_getpages_done_async(struct buf *bp)
1028 error = vnode_pager_generic_getpages_done(bp);
1029 /* Run the iodone upon the requested range. */
1030 bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore,
1031 bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error);
1032 for (int i = 0; i < bp->b_npages; i++)
1033 bp->b_pages[i] = NULL;
1036 relpbuf(bp, &vnode_async_pbuf_freecnt);
1040 vnode_pager_generic_getpages_done(struct buf *bp)
1043 off_t tfoff, nextoff;
1046 error = (bp->b_ioflags & BIO_ERROR) != 0 ? EIO : 0;
1047 object = bp->b_vp->v_object;
1049 if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1050 if (!buf_mapped(bp)) {
1051 bp->b_data = bp->b_kvabase;
1052 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages,
1055 bzero(bp->b_data + bp->b_bcount,
1056 PAGE_SIZE * bp->b_npages - bp->b_bcount);
1058 if (buf_mapped(bp)) {
1059 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1060 bp->b_data = unmapped_buf;
1063 VM_OBJECT_WLOCK(object);
1064 for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1065 i < bp->b_npages; i++, tfoff = nextoff) {
1068 nextoff = tfoff + PAGE_SIZE;
1069 mt = bp->b_pages[i];
1071 if (nextoff <= object->un_pager.vnp.vnp_size) {
1073 * Read filled up entire page.
1075 mt->valid = VM_PAGE_BITS_ALL;
1076 KASSERT(mt->dirty == 0,
1077 ("%s: page %p is dirty", __func__, mt));
1078 KASSERT(!pmap_page_is_mapped(mt),
1079 ("%s: page %p is mapped", __func__, mt));
1082 * Read did not fill up entire page.
1084 * Currently we do not set the entire page valid,
1085 * we just try to clear the piece that we couldn't
1088 vm_page_set_valid_range(mt, 0,
1089 object->un_pager.vnp.vnp_size - tfoff);
1090 KASSERT((mt->dirty & vm_page_bits(0,
1091 object->un_pager.vnp.vnp_size - tfoff)) == 0,
1092 ("%s: page %p is dirty", __func__, mt));
1095 if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter)
1096 vm_page_readahead_finish(mt);
1098 VM_OBJECT_WUNLOCK(object);
1100 printf("%s: I/O read error %d\n", __func__, error);
1106 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
1107 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1108 * vnode_pager_generic_putpages() to implement the previous behaviour.
1110 * All other FS's should use the bypass to get to the local media
1111 * backing vp's VOP_PUTPAGES.
1114 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1115 int flags, int *rtvals)
1119 int bytes = count * PAGE_SIZE;
1122 * Force synchronous operation if we are extremely low on memory
1123 * to prevent a low-memory deadlock. VOP operations often need to
1124 * allocate more memory to initiate the I/O ( i.e. do a BMAP
1125 * operation ). The swapper handles the case by limiting the amount
1126 * of asynchronous I/O, but that sort of solution doesn't scale well
1127 * for the vnode pager without a lot of work.
1129 * Also, the backing vnode's iodone routine may not wake the pageout
1130 * daemon up. This should be probably be addressed XXX.
1133 if (vm_cnt.v_free_count + vm_cnt.v_cache_count <
1134 vm_cnt.v_pageout_free_min)
1135 flags |= VM_PAGER_PUT_SYNC;
1138 * Call device-specific putpages function
1140 vp = object->handle;
1141 VM_OBJECT_WUNLOCK(object);
1142 rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1143 KASSERT(rtval != EOPNOTSUPP,
1144 ("vnode_pager: stale FS putpages\n"));
1145 VM_OBJECT_WLOCK(object);
1150 * This is now called from local media FS's to operate against their
1151 * own vnodes if they fail to implement VOP_PUTPAGES.
1153 * This is typically called indirectly via the pageout daemon and
1154 * clustering has already typically occurred, so in general we ask the
1155 * underlying filesystem to write the data out asynchronously rather
1159 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1160 int flags, int *rtvals)
1167 int maxsize, ncount;
1168 vm_ooffset_t poffset;
1174 static struct timeval lastfail;
1177 object = vp->v_object;
1178 count = bytecount / PAGE_SIZE;
1180 for (i = 0; i < count; i++)
1181 rtvals[i] = VM_PAGER_ERROR;
1183 if ((int64_t)ma[0]->pindex < 0) {
1184 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1185 (long)ma[0]->pindex, (u_long)ma[0]->dirty);
1186 rtvals[0] = VM_PAGER_BAD;
1187 return VM_PAGER_BAD;
1190 maxsize = count * PAGE_SIZE;
1193 poffset = IDX_TO_OFF(ma[0]->pindex);
1196 * If the page-aligned write is larger then the actual file we
1197 * have to invalidate pages occurring beyond the file EOF. However,
1198 * there is an edge case where a file may not be page-aligned where
1199 * the last page is partially invalid. In this case the filesystem
1200 * may not properly clear the dirty bits for the entire page (which
1201 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1202 * With the page locked we are free to fix-up the dirty bits here.
1204 * We do not under any circumstances truncate the valid bits, as
1205 * this will screw up bogus page replacement.
1207 VM_OBJECT_WLOCK(object);
1208 if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1209 if (object->un_pager.vnp.vnp_size > poffset) {
1212 maxsize = object->un_pager.vnp.vnp_size - poffset;
1213 ncount = btoc(maxsize);
1214 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1216 * If the object is locked and the following
1217 * conditions hold, then the page's dirty
1218 * field cannot be concurrently changed by a
1222 vm_page_assert_sbusied(m);
1223 KASSERT(!pmap_page_is_write_mapped(m),
1224 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1225 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1232 if (ncount < count) {
1233 for (i = ncount; i < count; i++) {
1234 rtvals[i] = VM_PAGER_BAD;
1238 VM_OBJECT_WUNLOCK(object);
1241 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
1242 * rather then a bdwrite() to prevent paging I/O from saturating
1243 * the buffer cache. Dummy-up the sequential heuristic to cause
1244 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
1245 * the system decides how to cluster.
1248 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1250 else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1251 ioflags |= IO_ASYNC;
1252 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1253 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1255 aiov.iov_base = (caddr_t) 0;
1256 aiov.iov_len = maxsize;
1257 auio.uio_iov = &aiov;
1258 auio.uio_iovcnt = 1;
1259 auio.uio_offset = poffset;
1260 auio.uio_segflg = UIO_NOCOPY;
1261 auio.uio_rw = UIO_WRITE;
1262 auio.uio_resid = maxsize;
1263 auio.uio_td = (struct thread *) 0;
1264 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1265 PCPU_INC(cnt.v_vnodeout);
1266 PCPU_ADD(cnt.v_vnodepgsout, ncount);
1269 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1)))
1270 printf("vnode_pager_putpages: I/O error %d\n", error);
1272 if (auio.uio_resid) {
1273 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1))
1274 printf("vnode_pager_putpages: residual I/O %zd at %lu\n",
1275 auio.uio_resid, (u_long)ma[0]->pindex);
1277 for (i = 0; i < ncount; i++) {
1278 rtvals[i] = VM_PAGER_OK;
1284 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written)
1291 obj = ma[0]->object;
1292 VM_OBJECT_WLOCK(obj);
1293 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1294 if (pos < trunc_page(written)) {
1295 rtvals[i] = VM_PAGER_OK;
1296 vm_page_undirty(ma[i]);
1298 /* Partially written page. */
1299 rtvals[i] = VM_PAGER_AGAIN;
1300 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1303 VM_OBJECT_WUNLOCK(obj);
1307 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1311 vm_ooffset_t old_wm;
1313 VM_OBJECT_WLOCK(object);
1314 if (object->type != OBJT_VNODE) {
1315 VM_OBJECT_WUNLOCK(object);
1318 old_wm = object->un_pager.vnp.writemappings;
1319 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1320 vp = object->handle;
1321 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1322 ASSERT_VOP_ELOCKED(vp, "v_writecount inc");
1323 VOP_ADD_WRITECOUNT(vp, 1);
1324 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1325 __func__, vp, vp->v_writecount);
1326 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1327 ASSERT_VOP_ELOCKED(vp, "v_writecount dec");
1328 VOP_ADD_WRITECOUNT(vp, -1);
1329 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1330 __func__, vp, vp->v_writecount);
1332 VM_OBJECT_WUNLOCK(object);
1336 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1343 VM_OBJECT_WLOCK(object);
1346 * First, recheck the object type to account for the race when
1347 * the vnode is reclaimed.
1349 if (object->type != OBJT_VNODE) {
1350 VM_OBJECT_WUNLOCK(object);
1355 * Optimize for the case when writemappings is not going to
1359 if (object->un_pager.vnp.writemappings != inc) {
1360 object->un_pager.vnp.writemappings -= inc;
1361 VM_OBJECT_WUNLOCK(object);
1365 vp = object->handle;
1367 VM_OBJECT_WUNLOCK(object);
1369 vn_start_write(vp, &mp, V_WAIT);
1370 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1373 * Decrement the object's writemappings, by swapping the start
1374 * and end arguments for vnode_pager_update_writecount(). If
1375 * there was not a race with vnode reclaimation, then the
1376 * vnode's v_writecount is decremented.
1378 vnode_pager_update_writecount(object, end, start);
1382 vn_finished_write(mp);