1 /*-
2 * SPDX-License-Identifier: BSD-4-Clause
3 *
4 * Copyright (c) 1990 University of Utah.
5 * Copyright (c) 1991 The Regents of the University of California.
6 * All rights reserved.
7 * Copyright (c) 1993, 1994 John S. Dyson
8 * Copyright (c) 1995, David Greenman
9 *
10 * This code is derived from software contributed to Berkeley by
11 * the Systems Programming Group of the University of Utah Computer
12 * Science Department.
13 *
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
16 * are met:
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.
29 *
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
40 * SUCH DAMAGE.
41 */
42
43 /*
44 * Page to/from files (vnodes).
45 */
46
47 /*
48 * TODO:
49 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
50 * greatly re-simplify the vnode_pager.
51 */
52
53 #include <sys/cdefs.h>
54 #include "opt_vm.h"
55
56 #include <sys/param.h>
57 #include <sys/kernel.h>
58 #include <sys/systm.h>
59 #include <sys/sysctl.h>
60 #include <sys/proc.h>
61 #include <sys/vnode.h>
62 #include <sys/mount.h>
63 #include <sys/bio.h>
64 #include <sys/buf.h>
65 #include <sys/vmmeter.h>
66 #include <sys/ktr.h>
67 #include <sys/limits.h>
68 #include <sys/conf.h>
69 #include <sys/refcount.h>
70 #include <sys/rwlock.h>
71 #include <sys/sf_buf.h>
72 #include <sys/domainset.h>
73 #include <sys/user.h>
74
75 #include <machine/atomic.h>
76
77 #include <vm/vm.h>
78 #include <vm/vm_param.h>
79 #include <vm/vm_object.h>
80 #include <vm/vm_page.h>
81 #include <vm/vm_pager.h>
82 #include <vm/vm_map.h>
83 #include <vm/vnode_pager.h>
84 #include <vm/vm_extern.h>
85 #include <vm/uma.h>
86
87 static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
88 daddr_t *rtaddress, int *run);
89 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
90 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
91 static void vnode_pager_dealloc(vm_object_t);
92 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *);
93 static int vnode_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
94 int *, vop_getpages_iodone_t, void *);
95 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *);
96 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
97 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t,
98 vm_ooffset_t, struct ucred *cred);
99 static int vnode_pager_generic_getpages_done(struct buf *);
100 static void vnode_pager_generic_getpages_done_async(struct buf *);
101 static void vnode_pager_update_writecount(vm_object_t, vm_offset_t,
102 vm_offset_t);
103 static void vnode_pager_release_writecount(vm_object_t, vm_offset_t,
104 vm_offset_t);
105 static void vnode_pager_getvp(vm_object_t, struct vnode **, bool *);
106
107 const struct pagerops vnodepagerops = {
108 .pgo_kvme_type = KVME_TYPE_VNODE,
109 .pgo_alloc = vnode_pager_alloc,
110 .pgo_dealloc = vnode_pager_dealloc,
111 .pgo_getpages = vnode_pager_getpages,
112 .pgo_getpages_async = vnode_pager_getpages_async,
113 .pgo_putpages = vnode_pager_putpages,
114 .pgo_haspage = vnode_pager_haspage,
115 .pgo_update_writecount = vnode_pager_update_writecount,
116 .pgo_release_writecount = vnode_pager_release_writecount,
117 .pgo_set_writeable_dirty = vm_object_set_writeable_dirty_,
118 .pgo_mightbedirty = vm_object_mightbedirty_,
119 .pgo_getvp = vnode_pager_getvp,
120 };
121
122 static struct domainset *vnode_domainset = NULL;
123
124 SYSCTL_PROC(_debug, OID_AUTO, vnode_domainset,
125 CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_RW, &vnode_domainset, 0,
126 sysctl_handle_domainset, "A", "Default vnode NUMA policy");
127
128 static int nvnpbufs;
129 SYSCTL_INT(_vm, OID_AUTO, vnode_pbufs, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
130 &nvnpbufs, 0, "number of physical buffers allocated for vnode pager");
131
132 static uma_zone_t vnode_pbuf_zone;
133
134 static void
vnode_pager_init(void * dummy)135 vnode_pager_init(void *dummy)
136 {
137
138 #ifdef __LP64__
139 nvnpbufs = nswbuf * 2;
140 #else
141 nvnpbufs = nswbuf / 2;
142 #endif
143 TUNABLE_INT_FETCH("vm.vnode_pbufs", &nvnpbufs);
144 vnode_pbuf_zone = pbuf_zsecond_create("vnpbuf", nvnpbufs);
145 }
146 SYSINIT(vnode_pager, SI_SUB_CPU, SI_ORDER_ANY, vnode_pager_init, NULL);
147
148 /* Create the VM system backing object for this vnode */
149 static int
vnode_create_vobject_any(struct vnode * vp,off_t isize,struct thread * td)150 vnode_create_vobject_any(struct vnode *vp, off_t isize, struct thread *td)
151 {
152 vm_object_t object;
153 vm_ooffset_t size;
154 bool last;
155
156 object = vp->v_object;
157 if (object != NULL)
158 return (0);
159
160 if (isize == VNODE_NO_SIZE) {
161 if (vn_getsize_locked(vp, &size, td->td_ucred) != 0)
162 return (0);
163 } else {
164 size = isize;
165 }
166
167 object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred);
168 /*
169 * Dereference the reference we just created. This assumes
170 * that the object is associated with the vp. We still have
171 * to serialize with vnode_pager_dealloc() for the last
172 * potential reference.
173 */
174 VM_OBJECT_RLOCK(object);
175 last = refcount_release(&object->ref_count);
176 VM_OBJECT_RUNLOCK(object);
177 if (last)
178 vrele(vp);
179
180 VNASSERT(vp->v_object != NULL, vp, ("%s: NULL object", __func__));
181
182 return (0);
183 }
184
185 int
vnode_create_vobject(struct vnode * vp,off_t isize,struct thread * td)186 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
187 {
188 VNASSERT(!vn_isdisk(vp), vp, ("%s: disk vnode", __func__));
189 VNASSERT(isize == VNODE_NO_SIZE || isize >= 0, vp,
190 ("%s: invalid size (%jd)", __func__, (intmax_t)isize));
191
192 if (!vn_canvmio(vp))
193 return (0);
194
195 return (vnode_create_vobject_any(vp, isize, td));
196 }
197
198 int
vnode_create_disk_vobject(struct vnode * vp,off_t isize,struct thread * td)199 vnode_create_disk_vobject(struct vnode *vp, off_t isize, struct thread *td)
200 {
201 VNASSERT(isize > 0, vp, ("%s: invalid size (%jd)", __func__,
202 (intmax_t)isize));
203
204 return (vnode_create_vobject_any(vp, isize, td));
205 }
206
207 void
vnode_destroy_vobject(struct vnode * vp)208 vnode_destroy_vobject(struct vnode *vp)
209 {
210 struct vm_object *obj;
211
212 obj = vp->v_object;
213 if (obj == NULL || obj->handle != vp)
214 return;
215 ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
216 VM_OBJECT_WLOCK(obj);
217 MPASS(obj->type == OBJT_VNODE);
218 umtx_shm_object_terminated(obj);
219 if (obj->ref_count == 0) {
220 KASSERT((obj->flags & OBJ_DEAD) == 0,
221 ("vnode_destroy_vobject: Terminating dead object"));
222 vm_object_set_flag(obj, OBJ_DEAD);
223
224 /*
225 * Clean pages and flush buffers.
226 */
227 vm_object_page_clean(obj, 0, 0, OBJPC_SYNC);
228 VM_OBJECT_WUNLOCK(obj);
229
230 vinvalbuf(vp, V_SAVE, 0, 0);
231
232 BO_LOCK(&vp->v_bufobj);
233 vp->v_bufobj.bo_flag |= BO_DEAD;
234 BO_UNLOCK(&vp->v_bufobj);
235
236 VM_OBJECT_WLOCK(obj);
237 vm_object_terminate(obj);
238 } else {
239 /*
240 * Woe to the process that tries to page now :-).
241 */
242 vm_pager_deallocate(obj);
243 VM_OBJECT_WUNLOCK(obj);
244 }
245 KASSERT(vp->v_object == NULL, ("vp %p obj %p", vp, vp->v_object));
246 }
247
248 /*
249 * Allocate (or lookup) pager for a vnode.
250 * Handle is a vnode pointer.
251 */
252 vm_object_t
vnode_pager_alloc(void * handle,vm_ooffset_t size,vm_prot_t prot,vm_ooffset_t offset,struct ucred * cred)253 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
254 vm_ooffset_t offset, struct ucred *cred)
255 {
256 vm_object_t object;
257 struct vnode *vp;
258
259 /*
260 * Pageout to vnode, no can do yet.
261 */
262 if (handle == NULL)
263 return (NULL);
264
265 vp = (struct vnode *)handle;
266 ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc");
267 VNPASS(vp->v_usecount > 0, vp);
268 retry:
269 object = vp->v_object;
270
271 if (object == NULL) {
272 /*
273 * Add an object of the appropriate size
274 */
275 object = vm_object_allocate(OBJT_VNODE,
276 OFF_TO_IDX(round_page(size)));
277
278 object->un_pager.vnp.vnp_size = size;
279 object->un_pager.vnp.writemappings = 0;
280 object->domain.dr_policy = vnode_domainset;
281 object->handle = handle;
282 if ((vp->v_vflag & VV_VMSIZEVNLOCK) != 0) {
283 VM_OBJECT_WLOCK(object);
284 vm_object_set_flag(object, OBJ_SIZEVNLOCK);
285 VM_OBJECT_WUNLOCK(object);
286 }
287 VI_LOCK(vp);
288 if (vp->v_object != NULL) {
289 /*
290 * Object has been created while we were allocating.
291 */
292 VI_UNLOCK(vp);
293 VM_OBJECT_WLOCK(object);
294 KASSERT(object->ref_count == 1,
295 ("leaked ref %p %d", object, object->ref_count));
296 object->type = OBJT_DEAD;
297 refcount_init(&object->ref_count, 0);
298 VM_OBJECT_WUNLOCK(object);
299 vm_object_destroy(object);
300 goto retry;
301 }
302 vp->v_object = object;
303 VI_UNLOCK(vp);
304 vrefact(vp);
305 } else {
306 vm_object_reference(object);
307 #if VM_NRESERVLEVEL > 0
308 if ((object->flags & OBJ_COLORED) == 0) {
309 VM_OBJECT_WLOCK(object);
310 vm_object_color(object, 0);
311 VM_OBJECT_WUNLOCK(object);
312 }
313 #endif
314 }
315 return (object);
316 }
317
318 /*
319 * The object must be locked.
320 */
321 static void
vnode_pager_dealloc(vm_object_t object)322 vnode_pager_dealloc(vm_object_t object)
323 {
324 struct vnode *vp;
325 int refs;
326
327 vp = object->handle;
328 if (vp == NULL)
329 panic("vnode_pager_dealloc: pager already dealloced");
330
331 VM_OBJECT_ASSERT_WLOCKED(object);
332 vm_object_pip_wait(object, "vnpdea");
333 refs = object->ref_count;
334
335 object->handle = NULL;
336 object->type = OBJT_DEAD;
337 ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
338 if (object->un_pager.vnp.writemappings > 0) {
339 object->un_pager.vnp.writemappings = 0;
340 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
341 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
342 __func__, vp, vp->v_writecount);
343 }
344 vp->v_object = NULL;
345 VI_LOCK(vp);
346
347 /*
348 * vm_map_entry_set_vnode_text() cannot reach this vnode by
349 * following object->handle. Clear all text references now.
350 * This also clears the transient references from
351 * kern_execve(), which is fine because dead_vnodeops uses nop
352 * for VOP_UNSET_TEXT().
353 */
354 if (vp->v_writecount < 0)
355 vp->v_writecount = 0;
356 VI_UNLOCK(vp);
357 VM_OBJECT_WUNLOCK(object);
358 if (refs > 0)
359 vunref(vp);
360 VM_OBJECT_WLOCK(object);
361 }
362
363 static boolean_t
vnode_pager_haspage(vm_object_t object,vm_pindex_t pindex,int * before,int * after)364 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
365 int *after)
366 {
367 struct vnode *vp = object->handle;
368 daddr_t bn;
369 uintptr_t lockstate;
370 int err;
371 daddr_t reqblock;
372 int poff;
373 int bsize;
374 int pagesperblock, blocksperpage;
375
376 VM_OBJECT_ASSERT_LOCKED(object);
377 /*
378 * If no vp or vp is doomed or marked transparent to VM, we do not
379 * have the page.
380 */
381 if (vp == NULL || VN_IS_DOOMED(vp))
382 return FALSE;
383 /*
384 * If the offset is beyond end of file we do
385 * not have the page.
386 */
387 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
388 return FALSE;
389
390 bsize = vp->v_mount->mnt_stat.f_iosize;
391 pagesperblock = bsize / PAGE_SIZE;
392 blocksperpage = 0;
393 if (pagesperblock > 0) {
394 reqblock = pindex / pagesperblock;
395 } else {
396 blocksperpage = (PAGE_SIZE / bsize);
397 reqblock = pindex * blocksperpage;
398 }
399 lockstate = VM_OBJECT_DROP(object);
400 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
401 VM_OBJECT_PICKUP(object, lockstate);
402 if (err)
403 return TRUE;
404 if (bn == -1)
405 return FALSE;
406 if (pagesperblock > 0) {
407 poff = pindex - (reqblock * pagesperblock);
408 if (before) {
409 *before *= pagesperblock;
410 *before += poff;
411 }
412 if (after) {
413 /*
414 * The BMAP vop can report a partial block in the
415 * 'after', but must not report blocks after EOF.
416 * Assert the latter, and truncate 'after' in case
417 * of the former.
418 */
419 KASSERT((reqblock + *after) * pagesperblock <
420 roundup2(object->size, pagesperblock),
421 ("%s: reqblock %jd after %d size %ju", __func__,
422 (intmax_t )reqblock, *after,
423 (uintmax_t )object->size));
424 *after *= pagesperblock;
425 *after += pagesperblock - (poff + 1);
426 if (pindex + *after >= object->size)
427 *after = object->size - 1 - pindex;
428 }
429 } else {
430 if (before) {
431 *before /= blocksperpage;
432 }
433
434 if (after) {
435 *after /= blocksperpage;
436 }
437 }
438 return TRUE;
439 }
440
441 /*
442 * Internal routine clearing partial-page content
443 */
444 static void
vnode_pager_subpage_purge(struct vm_page * m,int base,int end)445 vnode_pager_subpage_purge(struct vm_page *m, int base, int end)
446 {
447 int size;
448
449 KASSERT(end > base && end <= PAGE_SIZE,
450 ("%s: start %d end %d", __func__, base, end));
451 size = end - base;
452
453 /*
454 * Clear out partial-page garbage in case
455 * the page has been mapped.
456 */
457 pmap_zero_page_area(m, base, size);
458
459 /*
460 * Update the valid bits to reflect the blocks
461 * that have been zeroed. Some of these valid
462 * bits may have already been set.
463 */
464 vm_page_set_valid_range(m, base, size);
465
466 /*
467 * Round up "base" to the next block boundary so
468 * that the dirty bit for a partially zeroed
469 * block is not cleared.
470 */
471 base = roundup2(base, DEV_BSIZE);
472 end = rounddown2(end, DEV_BSIZE);
473
474 if (end > base) {
475 /*
476 * Clear out partial-page dirty bits.
477 *
478 * note that we do not clear out the
479 * valid bits. This would prevent
480 * bogus_page replacement from working
481 * properly.
482 */
483 vm_page_clear_dirty(m, base, end - base);
484 }
485
486 }
487
488 /*
489 * Lets the VM system know about a change in size for a file.
490 * We adjust our own internal size and flush any cached pages in
491 * the associated object that are affected by the size change.
492 *
493 * Note: this routine may be invoked as a result of a pager put
494 * operation (possibly at object termination time), so we must be careful.
495 */
496 void
vnode_pager_setsize(struct vnode * vp,vm_ooffset_t nsize)497 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
498 {
499 vm_object_t object;
500 vm_page_t m;
501 vm_pindex_t nobjsize;
502
503 if ((object = vp->v_object) == NULL)
504 return;
505 #ifdef DEBUG_VFS_LOCKS
506 {
507 struct mount *mp;
508
509 mp = vp->v_mount;
510 if (mp != NULL && (mp->mnt_kern_flag & MNTK_VMSETSIZE_BUG) == 0)
511 assert_vop_elocked(vp,
512 "vnode_pager_setsize and not locked vnode");
513 }
514 #endif
515 VM_OBJECT_WLOCK(object);
516 if (object->type == OBJT_DEAD) {
517 VM_OBJECT_WUNLOCK(object);
518 return;
519 }
520 KASSERT(object->type == OBJT_VNODE,
521 ("not vnode-backed object %p", object));
522 if (nsize == object->un_pager.vnp.vnp_size) {
523 /*
524 * Hasn't changed size
525 */
526 VM_OBJECT_WUNLOCK(object);
527 return;
528 }
529 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
530 if (nsize < object->un_pager.vnp.vnp_size) {
531 /*
532 * File has shrunk. Toss any cached pages beyond the new EOF.
533 */
534 if (nobjsize < object->size)
535 vm_object_page_remove(object, nobjsize, object->size,
536 0);
537 /*
538 * this gets rid of garbage at the end of a page that is now
539 * only partially backed by the vnode.
540 *
541 * XXX for some reason (I don't know yet), if we take a
542 * completely invalid page and mark it partially valid
543 * it can screw up NFS reads, so we don't allow the case.
544 */
545 if (!(nsize & PAGE_MASK))
546 goto out;
547 m = vm_page_grab(object, OFF_TO_IDX(nsize), VM_ALLOC_NOCREAT);
548 if (m == NULL)
549 goto out;
550 if (!vm_page_none_valid(m))
551 vnode_pager_subpage_purge(m, (int)nsize & PAGE_MASK,
552 PAGE_SIZE);
553 vm_page_xunbusy(m);
554 }
555 out:
556 #if defined(__powerpc__) && !defined(__powerpc64__)
557 object->un_pager.vnp.vnp_size = nsize;
558 #else
559 atomic_store_64(&object->un_pager.vnp.vnp_size, nsize);
560 #endif
561 object->size = nobjsize;
562 VM_OBJECT_WUNLOCK(object);
563 }
564
565 /*
566 * Lets the VM system know about the purged range for a file. We toss away any
567 * cached pages in the associated object that are affected by the purge
568 * operation. Partial-page area not aligned to page boundaries will be zeroed
569 * and the dirty blocks in DEV_BSIZE unit within a page will not be flushed.
570 */
571 void
vnode_pager_purge_range(struct vnode * vp,vm_ooffset_t start,vm_ooffset_t end)572 vnode_pager_purge_range(struct vnode *vp, vm_ooffset_t start, vm_ooffset_t end)
573 {
574 struct vm_page *m;
575 struct vm_object *object;
576 vm_pindex_t pi, pistart, piend;
577 bool same_page;
578 int base, pend;
579
580 ASSERT_VOP_LOCKED(vp, "vnode_pager_purge_range");
581
582 object = vp->v_object;
583 pi = start + PAGE_MASK < start ? OBJ_MAX_SIZE :
584 OFF_TO_IDX(start + PAGE_MASK);
585 pistart = OFF_TO_IDX(start);
586 piend = end == 0 ? OBJ_MAX_SIZE : OFF_TO_IDX(end);
587 same_page = pistart == piend;
588 if ((end != 0 && end <= start) || object == NULL)
589 return;
590
591 VM_OBJECT_WLOCK(object);
592
593 if (pi < piend)
594 vm_object_page_remove(object, pi, piend, 0);
595
596 if ((start & PAGE_MASK) != 0) {
597 base = (int)start & PAGE_MASK;
598 pend = same_page ? (int)end & PAGE_MASK : PAGE_SIZE;
599 m = vm_page_grab(object, pistart, VM_ALLOC_NOCREAT);
600 if (m != NULL) {
601 if (!vm_page_none_valid(m))
602 vnode_pager_subpage_purge(m, base, pend);
603 vm_page_xunbusy(m);
604 }
605 if (same_page)
606 goto out;
607 }
608 if ((end & PAGE_MASK) != 0) {
609 base = same_page ? (int)start & PAGE_MASK : 0 ;
610 pend = (int)end & PAGE_MASK;
611 m = vm_page_grab(object, piend, VM_ALLOC_NOCREAT);
612 if (m != NULL) {
613 if (!vm_page_none_valid(m))
614 vnode_pager_subpage_purge(m, base, pend);
615 vm_page_xunbusy(m);
616 }
617 }
618 out:
619 VM_OBJECT_WUNLOCK(object);
620 }
621
622 /*
623 * calculate the linear (byte) disk address of specified virtual
624 * file address
625 */
626 static int
vnode_pager_addr(struct vnode * vp,vm_ooffset_t address,daddr_t * rtaddress,int * run)627 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
628 int *run)
629 {
630 int bsize;
631 int err;
632 daddr_t vblock;
633 daddr_t voffset;
634
635 if (VN_IS_DOOMED(vp))
636 return -1;
637
638 bsize = vp->v_mount->mnt_stat.f_iosize;
639 vblock = address / bsize;
640 voffset = address % bsize;
641
642 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL);
643 if (err == 0) {
644 if (*rtaddress != -1)
645 *rtaddress += voffset / DEV_BSIZE;
646 if (run) {
647 *run += 1;
648 *run *= bsize / PAGE_SIZE;
649 *run -= voffset / PAGE_SIZE;
650 }
651 }
652
653 return (err);
654 }
655
656 static void
vnode_pager_input_bdone(struct buf * bp)657 vnode_pager_input_bdone(struct buf *bp)
658 {
659 runningbufwakeup(bp);
660 bdone(bp);
661 }
662
663 /*
664 * small block filesystem vnode pager input
665 */
666 static int
vnode_pager_input_smlfs(vm_object_t object,vm_page_t m)667 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m)
668 {
669 struct vnode *vp;
670 struct bufobj *bo;
671 struct buf *bp;
672 struct sf_buf *sf;
673 daddr_t fileaddr;
674 vm_offset_t bsize;
675 vm_page_bits_t bits;
676 int error, i;
677
678 error = 0;
679 vp = object->handle;
680 if (VN_IS_DOOMED(vp))
681 return VM_PAGER_BAD;
682
683 bsize = vp->v_mount->mnt_stat.f_iosize;
684
685 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
686
687 sf = sf_buf_alloc(m, 0);
688
689 for (i = 0; i < PAGE_SIZE / bsize; i++) {
690 vm_ooffset_t address;
691
692 bits = vm_page_bits(i * bsize, bsize);
693 if (m->valid & bits)
694 continue;
695
696 address = IDX_TO_OFF(m->pindex) + i * bsize;
697 if (address >= object->un_pager.vnp.vnp_size) {
698 fileaddr = -1;
699 } else {
700 error = vnode_pager_addr(vp, address, &fileaddr, NULL);
701 if (error)
702 break;
703 }
704 if (fileaddr != -1) {
705 bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK);
706
707 /* build a minimal buffer header */
708 bp->b_iocmd = BIO_READ;
709 bp->b_iodone = vnode_pager_input_bdone;
710 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
711 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
712 bp->b_rcred = crhold(curthread->td_ucred);
713 bp->b_wcred = crhold(curthread->td_ucred);
714 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
715 bp->b_blkno = fileaddr;
716 pbgetbo(bo, bp);
717 bp->b_vp = vp;
718 bp->b_bcount = bsize;
719 bp->b_bufsize = bsize;
720 (void)runningbufclaim(bp, bp->b_bufsize);
721
722 /* do the input */
723 bp->b_iooffset = dbtob(bp->b_blkno);
724 bstrategy(bp);
725
726 bwait(bp, PVM, "vnsrd");
727
728 if ((bp->b_ioflags & BIO_ERROR) != 0) {
729 KASSERT(bp->b_error != 0,
730 ("%s: buf error but b_error == 0\n", __func__));
731 error = bp->b_error;
732 }
733
734 /*
735 * free the buffer header back to the swap buffer pool
736 */
737 bp->b_vp = NULL;
738 pbrelbo(bp);
739 uma_zfree(vnode_pbuf_zone, bp);
740 if (error)
741 break;
742 } else
743 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
744 KASSERT((m->dirty & bits) == 0,
745 ("vnode_pager_input_smlfs: page %p is dirty", m));
746 vm_page_bits_set(m, &m->valid, bits);
747 }
748 sf_buf_free(sf);
749 if (error) {
750 return VM_PAGER_ERROR;
751 }
752 return VM_PAGER_OK;
753 }
754
755 /*
756 * old style vnode pager input routine
757 */
758 static int
vnode_pager_input_old(vm_object_t object,vm_page_t m)759 vnode_pager_input_old(vm_object_t object, vm_page_t m)
760 {
761 struct uio auio;
762 struct iovec aiov;
763 int error;
764 int size;
765 struct sf_buf *sf;
766 struct vnode *vp;
767
768 VM_OBJECT_ASSERT_WLOCKED(object);
769 error = 0;
770
771 /*
772 * Return failure if beyond current EOF
773 */
774 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
775 return VM_PAGER_BAD;
776 } else {
777 size = PAGE_SIZE;
778 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
779 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
780 vp = object->handle;
781 VM_OBJECT_WUNLOCK(object);
782
783 /*
784 * Allocate a kernel virtual address and initialize so that
785 * we can use VOP_READ/WRITE routines.
786 */
787 sf = sf_buf_alloc(m, 0);
788
789 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
790 aiov.iov_len = size;
791 auio.uio_iov = &aiov;
792 auio.uio_iovcnt = 1;
793 auio.uio_offset = IDX_TO_OFF(m->pindex);
794 auio.uio_segflg = UIO_SYSSPACE;
795 auio.uio_rw = UIO_READ;
796 auio.uio_resid = size;
797 auio.uio_td = curthread;
798
799 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
800 if (!error) {
801 int count = size - auio.uio_resid;
802
803 if (count == 0)
804 error = EINVAL;
805 else if (count != PAGE_SIZE)
806 bzero((caddr_t)sf_buf_kva(sf) + count,
807 PAGE_SIZE - count);
808 }
809 sf_buf_free(sf);
810
811 VM_OBJECT_WLOCK(object);
812 }
813 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
814 if (!error)
815 vm_page_valid(m);
816 return error ? VM_PAGER_ERROR : VM_PAGER_OK;
817 }
818
819 /*
820 * generic vnode pager input routine
821 */
822
823 /*
824 * Local media VFS's that do not implement their own VOP_GETPAGES
825 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
826 * to implement the previous behaviour.
827 *
828 * All other FS's should use the bypass to get to the local media
829 * backing vp's VOP_GETPAGES.
830 */
831 static int
vnode_pager_getpages(vm_object_t object,vm_page_t * m,int count,int * rbehind,int * rahead)832 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
833 int *rahead)
834 {
835 struct vnode *vp;
836 int rtval;
837
838 /* Handle is stable with paging in progress. */
839 vp = object->handle;
840 rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead);
841 KASSERT(rtval != EOPNOTSUPP,
842 ("vnode_pager: FS getpages not implemented\n"));
843 return rtval;
844 }
845
846 static int
vnode_pager_getpages_async(vm_object_t object,vm_page_t * m,int count,int * rbehind,int * rahead,vop_getpages_iodone_t iodone,void * arg)847 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
848 int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg)
849 {
850 struct vnode *vp;
851 int rtval;
852
853 vp = object->handle;
854 rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg);
855 KASSERT(rtval != EOPNOTSUPP,
856 ("vnode_pager: FS getpages_async not implemented\n"));
857 return (rtval);
858 }
859
860 /*
861 * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for
862 * local filesystems, where partially valid pages can only occur at
863 * the end of file.
864 */
865 int
vnode_pager_local_getpages(struct vop_getpages_args * ap)866 vnode_pager_local_getpages(struct vop_getpages_args *ap)
867 {
868
869 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
870 ap->a_rbehind, ap->a_rahead, NULL, NULL));
871 }
872
873 int
vnode_pager_local_getpages_async(struct vop_getpages_async_args * ap)874 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap)
875 {
876 int error;
877
878 error = vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
879 ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg);
880 if (error != 0 && ap->a_iodone != NULL)
881 ap->a_iodone(ap->a_arg, ap->a_m, ap->a_count, error);
882 return (error);
883 }
884
885 /*
886 * This is now called from local media FS's to operate against their
887 * own vnodes if they fail to implement VOP_GETPAGES.
888 */
889 int
vnode_pager_generic_getpages(struct vnode * vp,vm_page_t * m,int count,int * a_rbehind,int * a_rahead,vop_getpages_iodone_t iodone,void * arg)890 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count,
891 int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg)
892 {
893 vm_object_t object;
894 struct bufobj *bo;
895 struct buf *bp;
896 off_t foff;
897 #ifdef INVARIANTS
898 off_t blkno0;
899 #endif
900 int bsize, pagesperblock;
901 int error, before, after, rbehind, rahead, poff, i;
902 int bytecount, secmask;
903
904 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
905 ("%s does not support devices", __func__));
906
907 if (VN_IS_DOOMED(vp))
908 return (VM_PAGER_BAD);
909
910 object = vp->v_object;
911 foff = IDX_TO_OFF(m[0]->pindex);
912 bsize = vp->v_mount->mnt_stat.f_iosize;
913 pagesperblock = bsize / PAGE_SIZE;
914
915 KASSERT(foff < object->un_pager.vnp.vnp_size,
916 ("%s: page %p offset beyond vp %p size", __func__, m[0], vp));
917 KASSERT(count <= atop(maxphys),
918 ("%s: requested %d pages", __func__, count));
919
920 /*
921 * The last page has valid blocks. Invalid part can only
922 * exist at the end of file, and the page is made fully valid
923 * by zeroing in vm_pager_get_pages().
924 */
925 if (!vm_page_none_valid(m[count - 1]) && --count == 0) {
926 if (iodone != NULL)
927 iodone(arg, m, 1, 0);
928 return (VM_PAGER_OK);
929 }
930
931 bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK);
932 MPASS((bp->b_flags & B_MAXPHYS) != 0);
933
934 /*
935 * Get the underlying device blocks for the file with VOP_BMAP().
936 * If the file system doesn't support VOP_BMAP, use old way of
937 * getting pages via VOP_READ.
938 */
939 error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before);
940 if (error == EOPNOTSUPP) {
941 uma_zfree(vnode_pbuf_zone, bp);
942 VM_OBJECT_WLOCK(object);
943 for (i = 0; i < count; i++) {
944 VM_CNT_INC(v_vnodein);
945 VM_CNT_INC(v_vnodepgsin);
946 error = vnode_pager_input_old(object, m[i]);
947 if (error)
948 break;
949 }
950 VM_OBJECT_WUNLOCK(object);
951 return (error);
952 } else if (error != 0) {
953 uma_zfree(vnode_pbuf_zone, bp);
954 return (VM_PAGER_ERROR);
955 }
956
957 /*
958 * If the file system supports BMAP, but blocksize is smaller
959 * than a page size, then use special small filesystem code.
960 */
961 if (pagesperblock == 0) {
962 uma_zfree(vnode_pbuf_zone, bp);
963 for (i = 0; i < count; i++) {
964 VM_CNT_INC(v_vnodein);
965 VM_CNT_INC(v_vnodepgsin);
966 error = vnode_pager_input_smlfs(object, m[i]);
967 if (error)
968 break;
969 }
970 return (error);
971 }
972
973 /*
974 * A sparse file can be encountered only for a single page request,
975 * which may not be preceded by call to vm_pager_haspage().
976 */
977 if (bp->b_blkno == -1) {
978 KASSERT(count == 1,
979 ("%s: array[%d] request to a sparse file %p", __func__,
980 count, vp));
981 uma_zfree(vnode_pbuf_zone, bp);
982 pmap_zero_page(m[0]);
983 KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty",
984 __func__, m[0]));
985 vm_page_valid(m[0]);
986 return (VM_PAGER_OK);
987 }
988
989 #ifdef INVARIANTS
990 blkno0 = bp->b_blkno;
991 #endif
992 bp->b_blkno += (foff % bsize) / DEV_BSIZE;
993
994 /* Recalculate blocks available after/before to pages. */
995 poff = (foff % bsize) / PAGE_SIZE;
996 before *= pagesperblock;
997 before += poff;
998 after *= pagesperblock;
999 after += pagesperblock - (poff + 1);
1000 if (m[0]->pindex + after >= object->size)
1001 after = object->size - 1 - m[0]->pindex;
1002 KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d",
1003 __func__, count, after + 1));
1004 after -= count - 1;
1005
1006 /* Trim requested rbehind/rahead to possible values. */
1007 rbehind = a_rbehind ? *a_rbehind : 0;
1008 rahead = a_rahead ? *a_rahead : 0;
1009 rbehind = min(rbehind, before);
1010 rbehind = min(rbehind, m[0]->pindex);
1011 rahead = min(rahead, after);
1012 rahead = min(rahead, object->size - m[count - 1]->pindex);
1013 /*
1014 * Check that total amount of pages fit into buf. Trim rbehind and
1015 * rahead evenly if not.
1016 */
1017 if (rbehind + rahead + count > atop(maxphys)) {
1018 int trim, sum;
1019
1020 trim = rbehind + rahead + count - atop(maxphys) + 1;
1021 sum = rbehind + rahead;
1022 if (rbehind == before) {
1023 /* Roundup rbehind trim to block size. */
1024 rbehind -= roundup(trim * rbehind / sum, pagesperblock);
1025 if (rbehind < 0)
1026 rbehind = 0;
1027 } else
1028 rbehind -= trim * rbehind / sum;
1029 rahead -= trim * rahead / sum;
1030 }
1031 KASSERT(rbehind + rahead + count <= atop(maxphys),
1032 ("%s: behind %d ahead %d count %d maxphys %lu", __func__,
1033 rbehind, rahead, count, maxphys));
1034
1035 /*
1036 * Fill in the bp->b_pages[] array with requested and optional
1037 * read behind or read ahead pages. Read behind pages are looked
1038 * up in a backward direction, down to a first cached page. Same
1039 * for read ahead pages, but there is no need to shift the array
1040 * in case of encountering a cached page.
1041 */
1042 i = bp->b_npages = 0;
1043 if (rbehind) {
1044 vm_pindex_t startpindex, tpindex;
1045 vm_page_t mpred, p;
1046
1047 VM_OBJECT_WLOCK(object);
1048 tpindex = m[0]->pindex;
1049 startpindex = MAX(tpindex, rbehind) - rbehind;
1050 if ((mpred = TAILQ_PREV(m[0], pglist, listq)) != NULL)
1051 startpindex = MAX(startpindex, mpred->pindex + 1);
1052
1053 /* Stepping backward from pindex, mpred doesn't change. */
1054 for (; tpindex-- > startpindex; i++) {
1055 p = vm_page_alloc_after(object, tpindex,
1056 VM_ALLOC_NORMAL, mpred);
1057 if (p == NULL) {
1058 /* Shift the array. */
1059 for (int j = 0; j < i; j++)
1060 bp->b_pages[j] = bp->b_pages[j +
1061 tpindex + 1 - startpindex];
1062 break;
1063 }
1064 bp->b_pages[tpindex - startpindex] = p;
1065 }
1066
1067 bp->b_pgbefore = i;
1068 bp->b_npages += i;
1069 bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE;
1070 } else
1071 bp->b_pgbefore = 0;
1072
1073 /* Requested pages. */
1074 for (int j = 0; j < count; j++, i++)
1075 bp->b_pages[i] = m[j];
1076 bp->b_npages += count;
1077
1078 if (rahead) {
1079 vm_pindex_t endpindex, tpindex;
1080 vm_page_t p;
1081
1082 if (!VM_OBJECT_WOWNED(object))
1083 VM_OBJECT_WLOCK(object);
1084 endpindex = m[count - 1]->pindex + rahead + 1;
1085 if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL &&
1086 p->pindex < endpindex)
1087 endpindex = p->pindex;
1088 if (endpindex > object->size)
1089 endpindex = object->size;
1090
1091 p = m[count - 1];
1092 for (tpindex = p->pindex + 1;
1093 tpindex < endpindex; i++, tpindex++) {
1094 p = vm_page_alloc_after(object, tpindex,
1095 VM_ALLOC_NORMAL, p);
1096 if (p == NULL)
1097 break;
1098 bp->b_pages[i] = p;
1099 }
1100
1101 bp->b_pgafter = i - bp->b_npages;
1102 bp->b_npages = i;
1103 } else
1104 bp->b_pgafter = 0;
1105
1106 if (VM_OBJECT_WOWNED(object))
1107 VM_OBJECT_WUNLOCK(object);
1108
1109 /* Report back actual behind/ahead read. */
1110 if (a_rbehind)
1111 *a_rbehind = bp->b_pgbefore;
1112 if (a_rahead)
1113 *a_rahead = bp->b_pgafter;
1114
1115 #ifdef INVARIANTS
1116 KASSERT(bp->b_npages <= atop(maxphys),
1117 ("%s: buf %p overflowed", __func__, bp));
1118 for (int j = 1, prev = 0; j < bp->b_npages; j++) {
1119 if (bp->b_pages[j] == bogus_page)
1120 continue;
1121 KASSERT(bp->b_pages[j]->pindex - bp->b_pages[prev]->pindex ==
1122 j - prev, ("%s: pages array not consecutive, bp %p",
1123 __func__, bp));
1124 prev = j;
1125 }
1126 #endif
1127
1128 /*
1129 * Recalculate first offset and bytecount with regards to read behind.
1130 * Truncate bytecount to vnode real size and round up physical size
1131 * for real devices.
1132 */
1133 foff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1134 bytecount = bp->b_npages << PAGE_SHIFT;
1135 if ((foff + bytecount) > object->un_pager.vnp.vnp_size)
1136 bytecount = object->un_pager.vnp.vnp_size - foff;
1137 secmask = bo->bo_bsize - 1;
1138 KASSERT(secmask < PAGE_SIZE && secmask > 0,
1139 ("%s: sector size %d too large", __func__, secmask + 1));
1140 bytecount = (bytecount + secmask) & ~secmask;
1141
1142 /*
1143 * And map the pages to be read into the kva, if the filesystem
1144 * requires mapped buffers.
1145 */
1146 if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
1147 unmapped_buf_allowed) {
1148 bp->b_data = unmapped_buf;
1149 bp->b_offset = 0;
1150 } else {
1151 bp->b_data = bp->b_kvabase;
1152 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1153 }
1154
1155 /* Build a minimal buffer header. */
1156 bp->b_iocmd = BIO_READ;
1157 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
1158 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
1159 bp->b_rcred = crhold(curthread->td_ucred);
1160 bp->b_wcred = crhold(curthread->td_ucred);
1161 pbgetbo(bo, bp);
1162 bp->b_vp = vp;
1163 bp->b_bcount = bp->b_bufsize = bytecount;
1164 bp->b_iooffset = dbtob(bp->b_blkno);
1165 KASSERT(IDX_TO_OFF(m[0]->pindex - bp->b_pages[0]->pindex) ==
1166 (blkno0 - bp->b_blkno) * DEV_BSIZE +
1167 IDX_TO_OFF(m[0]->pindex) % bsize,
1168 ("wrong offsets bsize %d m[0] %ju b_pages[0] %ju "
1169 "blkno0 %ju b_blkno %ju", bsize,
1170 (uintmax_t)m[0]->pindex, (uintmax_t)bp->b_pages[0]->pindex,
1171 (uintmax_t)blkno0, (uintmax_t)bp->b_blkno));
1172
1173 (void)runningbufclaim(bp, bp->b_bufsize);
1174
1175 VM_CNT_INC(v_vnodein);
1176 VM_CNT_ADD(v_vnodepgsin, bp->b_npages);
1177
1178 if (iodone != NULL) { /* async */
1179 bp->b_pgiodone = iodone;
1180 bp->b_caller1 = arg;
1181 bp->b_iodone = vnode_pager_generic_getpages_done_async;
1182 bp->b_flags |= B_ASYNC;
1183 BUF_KERNPROC(bp);
1184 bstrategy(bp);
1185 return (VM_PAGER_OK);
1186 } else {
1187 bp->b_iodone = bdone;
1188 bstrategy(bp);
1189 bwait(bp, PVM, "vnread");
1190 error = vnode_pager_generic_getpages_done(bp);
1191 for (i = 0; i < bp->b_npages; i++)
1192 bp->b_pages[i] = NULL;
1193 bp->b_vp = NULL;
1194 pbrelbo(bp);
1195 uma_zfree(vnode_pbuf_zone, bp);
1196 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
1197 }
1198 }
1199
1200 static void
vnode_pager_generic_getpages_done_async(struct buf * bp)1201 vnode_pager_generic_getpages_done_async(struct buf *bp)
1202 {
1203 int error;
1204
1205 error = vnode_pager_generic_getpages_done(bp);
1206 /* Run the iodone upon the requested range. */
1207 bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore,
1208 bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error);
1209 for (int i = 0; i < bp->b_npages; i++)
1210 bp->b_pages[i] = NULL;
1211 bp->b_vp = NULL;
1212 pbrelbo(bp);
1213 uma_zfree(vnode_pbuf_zone, bp);
1214 }
1215
1216 static int
vnode_pager_generic_getpages_done(struct buf * bp)1217 vnode_pager_generic_getpages_done(struct buf *bp)
1218 {
1219 vm_object_t object;
1220 off_t tfoff, nextoff;
1221 int i, error;
1222
1223 KASSERT((bp->b_ioflags & BIO_ERROR) == 0 || bp->b_error != 0,
1224 ("%s: buf error but b_error == 0\n", __func__));
1225 error = (bp->b_ioflags & BIO_ERROR) != 0 ? bp->b_error : 0;
1226 object = bp->b_vp->v_object;
1227
1228 runningbufwakeup(bp);
1229
1230 if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1231 if (!buf_mapped(bp)) {
1232 bp->b_data = bp->b_kvabase;
1233 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages,
1234 bp->b_npages);
1235 }
1236 bzero(bp->b_data + bp->b_bcount,
1237 PAGE_SIZE * bp->b_npages - bp->b_bcount);
1238 }
1239 if (buf_mapped(bp)) {
1240 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1241 bp->b_data = unmapped_buf;
1242 }
1243
1244 /*
1245 * If the read failed, we must free any read ahead/behind pages here.
1246 * The requested pages are freed by the caller (for sync requests)
1247 * or by the bp->b_pgiodone callback (for async requests).
1248 */
1249 if (error != 0) {
1250 VM_OBJECT_WLOCK(object);
1251 for (i = 0; i < bp->b_pgbefore; i++)
1252 vm_page_free_invalid(bp->b_pages[i]);
1253 for (i = bp->b_npages - bp->b_pgafter; i < bp->b_npages; i++)
1254 vm_page_free_invalid(bp->b_pages[i]);
1255 VM_OBJECT_WUNLOCK(object);
1256 return (error);
1257 }
1258
1259 /* Read lock to protect size. */
1260 VM_OBJECT_RLOCK(object);
1261 for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1262 i < bp->b_npages; i++, tfoff = nextoff) {
1263 vm_page_t mt;
1264
1265 nextoff = tfoff + PAGE_SIZE;
1266 mt = bp->b_pages[i];
1267 if (mt == bogus_page)
1268 continue;
1269
1270 if (nextoff <= object->un_pager.vnp.vnp_size) {
1271 /*
1272 * Read filled up entire page.
1273 */
1274 vm_page_valid(mt);
1275 KASSERT(mt->dirty == 0,
1276 ("%s: page %p is dirty", __func__, mt));
1277 KASSERT(!pmap_page_is_mapped(mt),
1278 ("%s: page %p is mapped", __func__, mt));
1279 } else {
1280 /*
1281 * Read did not fill up entire page.
1282 *
1283 * Currently we do not set the entire page valid,
1284 * we just try to clear the piece that we couldn't
1285 * read.
1286 */
1287 vm_page_set_valid_range(mt, 0,
1288 object->un_pager.vnp.vnp_size - tfoff);
1289 KASSERT((mt->dirty & vm_page_bits(0,
1290 object->un_pager.vnp.vnp_size - tfoff)) == 0,
1291 ("%s: page %p is dirty", __func__, mt));
1292 }
1293
1294 if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter)
1295 vm_page_readahead_finish(mt);
1296 }
1297 VM_OBJECT_RUNLOCK(object);
1298
1299 return (error);
1300 }
1301
1302 /*
1303 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
1304 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1305 * vnode_pager_generic_putpages() to implement the previous behaviour.
1306 *
1307 * All other FS's should use the bypass to get to the local media
1308 * backing vp's VOP_PUTPAGES.
1309 */
1310 static void
vnode_pager_putpages(vm_object_t object,vm_page_t * m,int count,int flags,int * rtvals)1311 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1312 int flags, int *rtvals)
1313 {
1314 int rtval __diagused;
1315 struct vnode *vp;
1316 int bytes = count * PAGE_SIZE;
1317
1318 /*
1319 * Force synchronous operation if we are extremely low on memory
1320 * to prevent a low-memory deadlock. VOP operations often need to
1321 * allocate more memory to initiate the I/O ( i.e. do a BMAP
1322 * operation ). The swapper handles the case by limiting the amount
1323 * of asynchronous I/O, but that sort of solution doesn't scale well
1324 * for the vnode pager without a lot of work.
1325 *
1326 * Also, the backing vnode's iodone routine may not wake the pageout
1327 * daemon up. This should be probably be addressed XXX.
1328 */
1329
1330 if (vm_page_count_min())
1331 flags |= VM_PAGER_PUT_SYNC;
1332
1333 /*
1334 * Call device-specific putpages function
1335 */
1336 vp = object->handle;
1337 VM_OBJECT_WUNLOCK(object);
1338 rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1339 KASSERT(rtval != EOPNOTSUPP,
1340 ("vnode_pager: stale FS putpages\n"));
1341 VM_OBJECT_WLOCK(object);
1342 }
1343
1344 static int
vn_off2bidx(vm_ooffset_t offset)1345 vn_off2bidx(vm_ooffset_t offset)
1346 {
1347
1348 return ((offset & PAGE_MASK) / DEV_BSIZE);
1349 }
1350
1351 static bool
vn_dirty_blk(vm_page_t m,vm_ooffset_t offset)1352 vn_dirty_blk(vm_page_t m, vm_ooffset_t offset)
1353 {
1354
1355 KASSERT(IDX_TO_OFF(m->pindex) <= offset &&
1356 offset < IDX_TO_OFF(m->pindex + 1),
1357 ("page %p pidx %ju offset %ju", m, (uintmax_t)m->pindex,
1358 (uintmax_t)offset));
1359 return ((m->dirty & ((vm_page_bits_t)1 << vn_off2bidx(offset))) != 0);
1360 }
1361
1362 /*
1363 * This is now called from local media FS's to operate against their
1364 * own vnodes if they fail to implement VOP_PUTPAGES.
1365 *
1366 * This is typically called indirectly via the pageout daemon and
1367 * clustering has already typically occurred, so in general we ask the
1368 * underlying filesystem to write the data out asynchronously rather
1369 * then delayed.
1370 */
1371 int
vnode_pager_generic_putpages(struct vnode * vp,vm_page_t * ma,int bytecount,int flags,int * rtvals)1372 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1373 int flags, int *rtvals)
1374 {
1375 vm_object_t object;
1376 vm_page_t m;
1377 vm_ooffset_t max_offset, next_offset, poffset, prev_offset;
1378 struct uio auio;
1379 struct iovec aiov;
1380 off_t prev_resid, wrsz;
1381 int count, error, i, maxsize, ncount, pgoff, ppscheck;
1382 bool in_hole;
1383 static struct timeval lastfail;
1384 static int curfail;
1385
1386 object = vp->v_object;
1387 count = bytecount / PAGE_SIZE;
1388
1389 for (i = 0; i < count; i++)
1390 rtvals[i] = VM_PAGER_ERROR;
1391
1392 if ((int64_t)ma[0]->pindex < 0) {
1393 printf("vnode_pager_generic_putpages: "
1394 "attempt to write meta-data 0x%jx(%lx)\n",
1395 (uintmax_t)ma[0]->pindex, (u_long)ma[0]->dirty);
1396 rtvals[0] = VM_PAGER_BAD;
1397 return (VM_PAGER_BAD);
1398 }
1399
1400 maxsize = count * PAGE_SIZE;
1401 ncount = count;
1402
1403 poffset = IDX_TO_OFF(ma[0]->pindex);
1404
1405 /*
1406 * If the page-aligned write is larger then the actual file we
1407 * have to invalidate pages occurring beyond the file EOF. However,
1408 * there is an edge case where a file may not be page-aligned where
1409 * the last page is partially invalid. In this case the filesystem
1410 * may not properly clear the dirty bits for the entire page (which
1411 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1412 * With the page busied we are free to fix up the dirty bits here.
1413 *
1414 * We do not under any circumstances truncate the valid bits, as
1415 * this will screw up bogus page replacement.
1416 */
1417 VM_OBJECT_RLOCK(object);
1418 if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1419 if (object->un_pager.vnp.vnp_size > poffset) {
1420 maxsize = object->un_pager.vnp.vnp_size - poffset;
1421 ncount = btoc(maxsize);
1422 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1423 pgoff = roundup2(pgoff, DEV_BSIZE);
1424
1425 /*
1426 * If the page is busy and the following
1427 * conditions hold, then the page's dirty
1428 * field cannot be concurrently changed by a
1429 * pmap operation.
1430 */
1431 m = ma[ncount - 1];
1432 vm_page_assert_sbusied(m);
1433 KASSERT(!pmap_page_is_write_mapped(m),
1434 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1435 MPASS(m->dirty != 0);
1436 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1437 pgoff);
1438 }
1439 } else {
1440 maxsize = 0;
1441 ncount = 0;
1442 }
1443 for (i = ncount; i < count; i++)
1444 rtvals[i] = VM_PAGER_BAD;
1445 }
1446 VM_OBJECT_RUNLOCK(object);
1447
1448 auio.uio_iov = &aiov;
1449 auio.uio_segflg = UIO_NOCOPY;
1450 auio.uio_rw = UIO_WRITE;
1451 auio.uio_td = NULL;
1452 max_offset = roundup2(poffset + maxsize, DEV_BSIZE);
1453
1454 for (prev_offset = poffset; prev_offset < max_offset;) {
1455 /* Skip clean blocks. */
1456 for (in_hole = true; in_hole && prev_offset < max_offset;) {
1457 m = ma[OFF_TO_IDX(prev_offset - poffset)];
1458 for (i = vn_off2bidx(prev_offset);
1459 i < sizeof(vm_page_bits_t) * NBBY &&
1460 prev_offset < max_offset; i++) {
1461 if (vn_dirty_blk(m, prev_offset)) {
1462 in_hole = false;
1463 break;
1464 }
1465 prev_offset += DEV_BSIZE;
1466 }
1467 }
1468 if (in_hole)
1469 goto write_done;
1470
1471 /* Find longest run of dirty blocks. */
1472 for (next_offset = prev_offset; next_offset < max_offset;) {
1473 m = ma[OFF_TO_IDX(next_offset - poffset)];
1474 for (i = vn_off2bidx(next_offset);
1475 i < sizeof(vm_page_bits_t) * NBBY &&
1476 next_offset < max_offset; i++) {
1477 if (!vn_dirty_blk(m, next_offset))
1478 goto start_write;
1479 next_offset += DEV_BSIZE;
1480 }
1481 }
1482 start_write:
1483 if (next_offset > poffset + maxsize)
1484 next_offset = poffset + maxsize;
1485 if (prev_offset == next_offset)
1486 goto write_done;
1487
1488 /*
1489 * Getting here requires finding a dirty block in the
1490 * 'skip clean blocks' loop.
1491 */
1492
1493 aiov.iov_base = NULL;
1494 auio.uio_iovcnt = 1;
1495 auio.uio_offset = prev_offset;
1496 prev_resid = auio.uio_resid = aiov.iov_len = next_offset -
1497 prev_offset;
1498 error = VOP_WRITE(vp, &auio,
1499 vnode_pager_putpages_ioflags(flags), curthread->td_ucred);
1500
1501 wrsz = prev_resid - auio.uio_resid;
1502 if (wrsz == 0) {
1503 if (ppsratecheck(&lastfail, &curfail, 1) != 0) {
1504 vn_printf(vp, "vnode_pager_putpages: "
1505 "zero-length write at %ju resid %zd\n",
1506 auio.uio_offset, auio.uio_resid);
1507 }
1508 break;
1509 }
1510
1511 /* Adjust the starting offset for next iteration. */
1512 prev_offset += wrsz;
1513 MPASS(auio.uio_offset == prev_offset);
1514
1515 ppscheck = 0;
1516 if (error != 0 && (ppscheck = ppsratecheck(&lastfail,
1517 &curfail, 1)) != 0)
1518 vn_printf(vp, "vnode_pager_putpages: I/O error %d\n",
1519 error);
1520 if (auio.uio_resid != 0 && (ppscheck != 0 ||
1521 ppsratecheck(&lastfail, &curfail, 1) != 0))
1522 vn_printf(vp, "vnode_pager_putpages: residual I/O %zd "
1523 "at %ju\n", auio.uio_resid,
1524 (uintmax_t)ma[0]->pindex);
1525 if (error != 0 || auio.uio_resid != 0)
1526 break;
1527 }
1528 write_done:
1529 /* Mark completely processed pages. */
1530 for (i = 0; i < OFF_TO_IDX(prev_offset - poffset); i++)
1531 rtvals[i] = VM_PAGER_OK;
1532 /* Mark partial EOF page. */
1533 if (prev_offset == poffset + maxsize && (prev_offset & PAGE_MASK) != 0)
1534 rtvals[i++] = VM_PAGER_OK;
1535 /* Unwritten pages in range, free bonus if the page is clean. */
1536 for (; i < ncount; i++)
1537 rtvals[i] = ma[i]->dirty == 0 ? VM_PAGER_OK : VM_PAGER_ERROR;
1538 VM_CNT_ADD(v_vnodepgsout, i);
1539 VM_CNT_INC(v_vnodeout);
1540 return (rtvals[0]);
1541 }
1542
1543 int
vnode_pager_putpages_ioflags(int pager_flags)1544 vnode_pager_putpages_ioflags(int pager_flags)
1545 {
1546 int ioflags;
1547
1548 /*
1549 * Pageouts are already clustered, use IO_ASYNC to force a
1550 * bawrite() rather then a bdwrite() to prevent paging I/O
1551 * from saturating the buffer cache. Dummy-up the sequential
1552 * heuristic to cause large ranges to cluster. If neither
1553 * IO_SYNC or IO_ASYNC is set, the system decides how to
1554 * cluster.
1555 */
1556 ioflags = IO_VMIO;
1557 if ((pager_flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) != 0)
1558 ioflags |= IO_SYNC;
1559 else if ((pager_flags & VM_PAGER_CLUSTER_OK) == 0)
1560 ioflags |= IO_ASYNC;
1561 ioflags |= (pager_flags & VM_PAGER_PUT_INVAL) != 0 ? IO_INVAL: 0;
1562 ioflags |= (pager_flags & VM_PAGER_PUT_NOREUSE) != 0 ? IO_NOREUSE : 0;
1563 ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1564 return (ioflags);
1565 }
1566
1567 /*
1568 * vnode_pager_undirty_pages().
1569 *
1570 * A helper to mark pages as clean after pageout that was possibly
1571 * done with a short write. The lpos argument specifies the page run
1572 * length in bytes, and the written argument specifies how many bytes
1573 * were actually written. eof is the offset past the last valid byte
1574 * in the vnode using the absolute file position of the first byte in
1575 * the run as the base from which it is computed.
1576 */
1577 void
vnode_pager_undirty_pages(vm_page_t * ma,int * rtvals,int written,off_t eof,int lpos)1578 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written, off_t eof,
1579 int lpos)
1580 {
1581 int i, pos, pos_devb;
1582
1583 if (written == 0 && eof >= lpos)
1584 return;
1585 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1586 if (pos < trunc_page(written)) {
1587 rtvals[i] = VM_PAGER_OK;
1588 vm_page_undirty(ma[i]);
1589 } else {
1590 /* Partially written page. */
1591 rtvals[i] = VM_PAGER_AGAIN;
1592 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1593 }
1594 }
1595 if (eof >= lpos) /* avoid truncation */
1596 return;
1597 for (pos = eof, i = OFF_TO_IDX(trunc_page(pos)); pos < lpos; i++) {
1598 if (pos != trunc_page(pos)) {
1599 /*
1600 * The page contains the last valid byte in
1601 * the vnode, mark the rest of the page as
1602 * clean, potentially making the whole page
1603 * clean.
1604 */
1605 pos_devb = roundup2(pos & PAGE_MASK, DEV_BSIZE);
1606 vm_page_clear_dirty(ma[i], pos_devb, PAGE_SIZE -
1607 pos_devb);
1608
1609 /*
1610 * If the page was cleaned, report the pageout
1611 * on it as successful. msync() no longer
1612 * needs to write out the page, endlessly
1613 * creating write requests and dirty buffers.
1614 */
1615 if (ma[i]->dirty == 0)
1616 rtvals[i] = VM_PAGER_OK;
1617
1618 pos = round_page(pos);
1619 } else {
1620 /* vm_pageout_flush() clears dirty */
1621 rtvals[i] = VM_PAGER_BAD;
1622 pos += PAGE_SIZE;
1623 }
1624 }
1625 }
1626
1627 static void
vnode_pager_update_writecount(vm_object_t object,vm_offset_t start,vm_offset_t end)1628 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1629 vm_offset_t end)
1630 {
1631 struct vnode *vp;
1632 vm_ooffset_t old_wm;
1633
1634 VM_OBJECT_WLOCK(object);
1635 if (object->type != OBJT_VNODE) {
1636 VM_OBJECT_WUNLOCK(object);
1637 return;
1638 }
1639 old_wm = object->un_pager.vnp.writemappings;
1640 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1641 vp = object->handle;
1642 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1643 ASSERT_VOP_LOCKED(vp, "v_writecount inc");
1644 VOP_ADD_WRITECOUNT_CHECKED(vp, 1);
1645 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1646 __func__, vp, vp->v_writecount);
1647 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1648 ASSERT_VOP_LOCKED(vp, "v_writecount dec");
1649 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1650 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1651 __func__, vp, vp->v_writecount);
1652 }
1653 VM_OBJECT_WUNLOCK(object);
1654 }
1655
1656 static void
vnode_pager_release_writecount(vm_object_t object,vm_offset_t start,vm_offset_t end)1657 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1658 vm_offset_t end)
1659 {
1660 struct vnode *vp;
1661 struct mount *mp;
1662 vm_offset_t inc;
1663
1664 VM_OBJECT_WLOCK(object);
1665
1666 /*
1667 * First, recheck the object type to account for the race when
1668 * the vnode is reclaimed.
1669 */
1670 if (object->type != OBJT_VNODE) {
1671 VM_OBJECT_WUNLOCK(object);
1672 return;
1673 }
1674
1675 /*
1676 * Optimize for the case when writemappings is not going to
1677 * zero.
1678 */
1679 inc = end - start;
1680 if (object->un_pager.vnp.writemappings != inc) {
1681 object->un_pager.vnp.writemappings -= inc;
1682 VM_OBJECT_WUNLOCK(object);
1683 return;
1684 }
1685
1686 vp = object->handle;
1687 vhold(vp);
1688 VM_OBJECT_WUNLOCK(object);
1689 mp = NULL;
1690 vn_start_write(vp, &mp, V_WAIT);
1691 vn_lock(vp, LK_SHARED | LK_RETRY);
1692
1693 /*
1694 * Decrement the object's writemappings, by swapping the start
1695 * and end arguments for vnode_pager_update_writecount(). If
1696 * there was not a race with vnode reclaimation, then the
1697 * vnode's v_writecount is decremented.
1698 */
1699 vnode_pager_update_writecount(object, end, start);
1700 VOP_UNLOCK(vp);
1701 vdrop(vp);
1702 if (mp != NULL)
1703 vn_finished_write(mp);
1704 }
1705
1706 static void
vnode_pager_getvp(vm_object_t object,struct vnode ** vpp,bool * vp_heldp)1707 vnode_pager_getvp(vm_object_t object, struct vnode **vpp, bool *vp_heldp)
1708 {
1709 *vpp = object->handle;
1710 }
1711
1712 static void
vnode_pager_clean1(struct vnode * vp,int sync_flags)1713 vnode_pager_clean1(struct vnode *vp, int sync_flags)
1714 {
1715 struct vm_object *obj;
1716
1717 ASSERT_VOP_LOCKED(vp, "needs lock for writes");
1718 obj = vp->v_object;
1719 if (obj == NULL)
1720 return;
1721
1722 VM_OBJECT_WLOCK(obj);
1723 vm_object_page_clean(obj, 0, 0, sync_flags);
1724 VM_OBJECT_WUNLOCK(obj);
1725 }
1726
1727 void
vnode_pager_clean_sync(struct vnode * vp)1728 vnode_pager_clean_sync(struct vnode *vp)
1729 {
1730 vnode_pager_clean1(vp, OBJPC_SYNC);
1731 }
1732
1733 void
vnode_pager_clean_async(struct vnode * vp)1734 vnode_pager_clean_async(struct vnode *vp)
1735 {
1736 vnode_pager_clean1(vp, 0);
1737 }
1738