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 87static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, 88 daddr_t *rtaddress, int *run); 89static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m); 90static int vnode_pager_input_old(vm_object_t object, vm_page_t m); 91static void vnode_pager_dealloc(vm_object_t); 92static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *); 93static int vnode_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 94 int *, vop_getpages_iodone_t, void *); 95static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *); 96static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *); 97static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t, 98 vm_ooffset_t, struct ucred *cred); 99static int vnode_pager_generic_getpages_done(struct buf *); 100static void vnode_pager_generic_getpages_done_async(struct buf *); 101static void vnode_pager_update_writecount(vm_object_t, vm_offset_t, 102 vm_offset_t); 103static void vnode_pager_release_writecount(vm_object_t, vm_offset_t, 104 vm_offset_t); 105static void vnode_pager_getvp(vm_object_t, struct vnode **, bool *); 106 107const 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 122static struct domainset *vnode_domainset = NULL; 123 124SYSCTL_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 128static int nvnpbufs; 129SYSCTL_INT(_vm, OID_AUTO, vnode_pbufs, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 130 &nvnpbufs, 0, "number of physical buffers allocated for vnode pager"); 131 132static uma_zone_t vnode_pbuf_zone; 133 134static void 135vnode_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} 146SYSINIT(vnode_pager, SI_SUB_CPU, SI_ORDER_ANY, vnode_pager_init, NULL); 147 148/* Create the VM system backing object for this vnode */ 149static int 150vnode_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 185int 186vnode_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 198int 199vnode_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 207void 208vnode_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 */ 252vm_object_t 253vnode_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); 268retry: 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 */ 321static void 322vnode_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 363static boolean_t 364vnode_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 */ 444static void 445vnode_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 */ 496void 497vnode_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 } 555out: 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 */ 571void 572vnode_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 } 618out: 619 VM_OBJECT_WUNLOCK(object); 620} 621 622/* 623 * calculate the linear (byte) disk address of specified virtual 624 * file address 625 */ 626static int 627vnode_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 656static void 657vnode_pager_input_bdone(struct buf *bp) 658{ 659 runningbufwakeup(bp); 660 bdone(bp); 661} 662 663/* 664 * small block filesystem vnode pager input 665 */ 666static int 667vnode_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 bp->b_runningbufspace = bp->b_bufsize; 721 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 722 723 /* do the input */ 724 bp->b_iooffset = dbtob(bp->b_blkno); 725 bstrategy(bp); 726 727 bwait(bp, PVM, "vnsrd"); 728 729 if ((bp->b_ioflags & BIO_ERROR) != 0) { 730 KASSERT(bp->b_error != 0, 731 ("%s: buf error but b_error == 0\n", __func__)); 732 error = bp->b_error; 733 } 734 735 /* 736 * free the buffer header back to the swap buffer pool 737 */ 738 bp->b_vp = NULL; 739 pbrelbo(bp); 740 uma_zfree(vnode_pbuf_zone, bp); 741 if (error) 742 break; 743 } else 744 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize); 745 KASSERT((m->dirty & bits) == 0, 746 ("vnode_pager_input_smlfs: page %p is dirty", m)); 747 vm_page_bits_set(m, &m->valid, bits); 748 } 749 sf_buf_free(sf); 750 if (error) { 751 return VM_PAGER_ERROR; 752 } 753 return VM_PAGER_OK; 754} 755 756/* 757 * old style vnode pager input routine 758 */ 759static int 760vnode_pager_input_old(vm_object_t object, vm_page_t m) 761{ 762 struct uio auio; 763 struct iovec aiov; 764 int error; 765 int size; 766 struct sf_buf *sf; 767 struct vnode *vp; 768 769 VM_OBJECT_ASSERT_WLOCKED(object); 770 error = 0; 771 772 /* 773 * Return failure if beyond current EOF 774 */ 775 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 776 return VM_PAGER_BAD; 777 } else { 778 size = PAGE_SIZE; 779 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 780 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 781 vp = object->handle; 782 VM_OBJECT_WUNLOCK(object); 783 784 /* 785 * Allocate a kernel virtual address and initialize so that 786 * we can use VOP_READ/WRITE routines. 787 */ 788 sf = sf_buf_alloc(m, 0); 789 790 aiov.iov_base = (caddr_t)sf_buf_kva(sf); 791 aiov.iov_len = size; 792 auio.uio_iov = &aiov; 793 auio.uio_iovcnt = 1; 794 auio.uio_offset = IDX_TO_OFF(m->pindex); 795 auio.uio_segflg = UIO_SYSSPACE; 796 auio.uio_rw = UIO_READ; 797 auio.uio_resid = size; 798 auio.uio_td = curthread; 799 800 error = VOP_READ(vp, &auio, 0, curthread->td_ucred); 801 if (!error) { 802 int count = size - auio.uio_resid; 803 804 if (count == 0) 805 error = EINVAL; 806 else if (count != PAGE_SIZE) 807 bzero((caddr_t)sf_buf_kva(sf) + count, 808 PAGE_SIZE - count); 809 } 810 sf_buf_free(sf); 811 812 VM_OBJECT_WLOCK(object); 813 } 814 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m)); 815 if (!error) 816 vm_page_valid(m); 817 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 818} 819 820/* 821 * generic vnode pager input routine 822 */ 823 824/* 825 * Local media VFS's that do not implement their own VOP_GETPAGES 826 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages() 827 * to implement the previous behaviour. 828 * 829 * All other FS's should use the bypass to get to the local media 830 * backing vp's VOP_GETPAGES. 831 */ 832static int 833vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind, 834 int *rahead) 835{ 836 struct vnode *vp; 837 int rtval; 838 839 /* Handle is stable with paging in progress. */ 840 vp = object->handle; 841 rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead); 842 KASSERT(rtval != EOPNOTSUPP, 843 ("vnode_pager: FS getpages not implemented\n")); 844 return rtval; 845} 846 847static int 848vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count, 849 int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg) 850{ 851 struct vnode *vp; 852 int rtval; 853 854 vp = object->handle; 855 rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg); 856 KASSERT(rtval != EOPNOTSUPP, 857 ("vnode_pager: FS getpages_async not implemented\n")); 858 return (rtval); 859} 860 861/* 862 * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for 863 * local filesystems, where partially valid pages can only occur at 864 * the end of file. 865 */ 866int 867vnode_pager_local_getpages(struct vop_getpages_args *ap) 868{ 869 870 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count, 871 ap->a_rbehind, ap->a_rahead, NULL, NULL)); 872} 873 874int 875vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap) 876{ 877 int error; 878 879 error = vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count, 880 ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg); 881 if (error != 0 && ap->a_iodone != NULL) 882 ap->a_iodone(ap->a_arg, ap->a_m, ap->a_count, error); 883 return (error); 884} 885 886/* 887 * This is now called from local media FS's to operate against their 888 * own vnodes if they fail to implement VOP_GETPAGES. 889 */ 890int 891vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count, 892 int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg) 893{ 894 vm_object_t object; 895 struct bufobj *bo; 896 struct buf *bp; 897 off_t foff; 898#ifdef INVARIANTS 899 off_t blkno0; 900#endif 901 int bsize, pagesperblock; 902 int error, before, after, rbehind, rahead, poff, i; 903 int bytecount, secmask; 904 905 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 906 ("%s does not support devices", __func__)); 907 908 if (VN_IS_DOOMED(vp)) 909 return (VM_PAGER_BAD); 910 911 object = vp->v_object; 912 foff = IDX_TO_OFF(m[0]->pindex); 913 bsize = vp->v_mount->mnt_stat.f_iosize; 914 pagesperblock = bsize / PAGE_SIZE; 915 916 KASSERT(foff < object->un_pager.vnp.vnp_size, 917 ("%s: page %p offset beyond vp %p size", __func__, m[0], vp)); 918 KASSERT(count <= atop(maxphys), 919 ("%s: requested %d pages", __func__, count)); 920 921 /* 922 * The last page has valid blocks. Invalid part can only 923 * exist at the end of file, and the page is made fully valid 924 * by zeroing in vm_pager_get_pages(). 925 */ 926 if (!vm_page_none_valid(m[count - 1]) && --count == 0) { 927 if (iodone != NULL) 928 iodone(arg, m, 1, 0); 929 return (VM_PAGER_OK); 930 } 931 932 bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK); 933 MPASS((bp->b_flags & B_MAXPHYS) != 0); 934 935 /* 936 * Get the underlying device blocks for the file with VOP_BMAP(). 937 * If the file system doesn't support VOP_BMAP, use old way of 938 * getting pages via VOP_READ. 939 */ 940 error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before); 941 if (error == EOPNOTSUPP) { 942 uma_zfree(vnode_pbuf_zone, bp); 943 VM_OBJECT_WLOCK(object); 944 for (i = 0; i < count; i++) { 945 VM_CNT_INC(v_vnodein); 946 VM_CNT_INC(v_vnodepgsin); 947 error = vnode_pager_input_old(object, m[i]); 948 if (error) 949 break; 950 } 951 VM_OBJECT_WUNLOCK(object); 952 return (error); 953 } else if (error != 0) { 954 uma_zfree(vnode_pbuf_zone, bp); 955 return (VM_PAGER_ERROR); 956 } 957 958 /* 959 * If the file system supports BMAP, but blocksize is smaller 960 * than a page size, then use special small filesystem code. 961 */ 962 if (pagesperblock == 0) { 963 uma_zfree(vnode_pbuf_zone, bp); 964 for (i = 0; i < count; i++) { 965 VM_CNT_INC(v_vnodein); 966 VM_CNT_INC(v_vnodepgsin); 967 error = vnode_pager_input_smlfs(object, m[i]); 968 if (error) 969 break; 970 } 971 return (error); 972 } 973 974 /* 975 * A sparse file can be encountered only for a single page request, 976 * which may not be preceded by call to vm_pager_haspage(). 977 */ 978 if (bp->b_blkno == -1) { 979 KASSERT(count == 1, 980 ("%s: array[%d] request to a sparse file %p", __func__, 981 count, vp)); 982 uma_zfree(vnode_pbuf_zone, bp); 983 pmap_zero_page(m[0]); 984 KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty", 985 __func__, m[0])); 986 vm_page_valid(m[0]); 987 return (VM_PAGER_OK); 988 } 989 990#ifdef INVARIANTS 991 blkno0 = bp->b_blkno; 992#endif 993 bp->b_blkno += (foff % bsize) / DEV_BSIZE; 994 995 /* Recalculate blocks available after/before to pages. */ 996 poff = (foff % bsize) / PAGE_SIZE; 997 before *= pagesperblock; 998 before += poff; 999 after *= pagesperblock; 1000 after += pagesperblock - (poff + 1); 1001 if (m[0]->pindex + after >= object->size) 1002 after = object->size - 1 - m[0]->pindex; 1003 KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d", 1004 __func__, count, after + 1)); 1005 after -= count - 1; 1006 1007 /* Trim requested rbehind/rahead to possible values. */ 1008 rbehind = a_rbehind ? *a_rbehind : 0; 1009 rahead = a_rahead ? *a_rahead : 0; 1010 rbehind = min(rbehind, before); 1011 rbehind = min(rbehind, m[0]->pindex); 1012 rahead = min(rahead, after); 1013 rahead = min(rahead, object->size - m[count - 1]->pindex); 1014 /* 1015 * Check that total amount of pages fit into buf. Trim rbehind and 1016 * rahead evenly if not. 1017 */ 1018 if (rbehind + rahead + count > atop(maxphys)) { 1019 int trim, sum; 1020 1021 trim = rbehind + rahead + count - atop(maxphys) + 1; 1022 sum = rbehind + rahead; 1023 if (rbehind == before) { 1024 /* Roundup rbehind trim to block size. */ 1025 rbehind -= roundup(trim * rbehind / sum, pagesperblock); 1026 if (rbehind < 0) 1027 rbehind = 0; 1028 } else 1029 rbehind -= trim * rbehind / sum; 1030 rahead -= trim * rahead / sum; 1031 } 1032 KASSERT(rbehind + rahead + count <= atop(maxphys), 1033 ("%s: behind %d ahead %d count %d maxphys %lu", __func__, 1034 rbehind, rahead, count, maxphys)); 1035 1036 /* 1037 * Fill in the bp->b_pages[] array with requested and optional 1038 * read behind or read ahead pages. Read behind pages are looked 1039 * up in a backward direction, down to a first cached page. Same 1040 * for read ahead pages, but there is no need to shift the array 1041 * in case of encountering a cached page. 1042 */ 1043 i = bp->b_npages = 0; 1044 if (rbehind) { 1045 vm_pindex_t startpindex, tpindex; 1046 vm_page_t p; 1047 1048 VM_OBJECT_WLOCK(object); 1049 startpindex = m[0]->pindex - rbehind; 1050 if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL && 1051 p->pindex >= startpindex) 1052 startpindex = p->pindex + 1; 1053 1054 /* tpindex is unsigned; beware of numeric underflow. */ 1055 for (tpindex = m[0]->pindex - 1; 1056 tpindex >= startpindex && tpindex < m[0]->pindex; 1057 tpindex--, i++) { 1058 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1059 if (p == NULL) { 1060 /* Shift the array. */ 1061 for (int j = 0; j < i; j++) 1062 bp->b_pages[j] = bp->b_pages[j + 1063 tpindex + 1 - startpindex]; 1064 break; 1065 } 1066 bp->b_pages[tpindex - startpindex] = p; 1067 } 1068 1069 bp->b_pgbefore = i; 1070 bp->b_npages += i; 1071 bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE; 1072 } else 1073 bp->b_pgbefore = 0; 1074 1075 /* Requested pages. */ 1076 for (int j = 0; j < count; j++, i++) 1077 bp->b_pages[i] = m[j]; 1078 bp->b_npages += count; 1079 1080 if (rahead) { 1081 vm_pindex_t endpindex, tpindex; 1082 vm_page_t p; 1083 1084 if (!VM_OBJECT_WOWNED(object)) 1085 VM_OBJECT_WLOCK(object); 1086 endpindex = m[count - 1]->pindex + rahead + 1; 1087 if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL && 1088 p->pindex < endpindex) 1089 endpindex = p->pindex; 1090 if (endpindex > object->size) 1091 endpindex = object->size; 1092 1093 for (tpindex = m[count - 1]->pindex + 1; 1094 tpindex < endpindex; i++, tpindex++) { 1095 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 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 = bp->b_runningbufspace = 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 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 1174 VM_CNT_INC(v_vnodein); 1175 VM_CNT_ADD(v_vnodepgsin, bp->b_npages); 1176 1177 if (iodone != NULL) { /* async */ 1178 bp->b_pgiodone = iodone; 1179 bp->b_caller1 = arg; 1180 bp->b_iodone = vnode_pager_generic_getpages_done_async; 1181 bp->b_flags |= B_ASYNC; 1182 BUF_KERNPROC(bp); 1183 bstrategy(bp); 1184 return (VM_PAGER_OK); 1185 } else { 1186 bp->b_iodone = bdone; 1187 bstrategy(bp); 1188 bwait(bp, PVM, "vnread"); 1189 error = vnode_pager_generic_getpages_done(bp); 1190 for (i = 0; i < bp->b_npages; i++) 1191 bp->b_pages[i] = NULL; 1192 bp->b_vp = NULL; 1193 pbrelbo(bp); 1194 uma_zfree(vnode_pbuf_zone, bp); 1195 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); 1196 } 1197} 1198 1199static void 1200vnode_pager_generic_getpages_done_async(struct buf *bp) 1201{ 1202 int error; 1203 1204 error = vnode_pager_generic_getpages_done(bp); 1205 /* Run the iodone upon the requested range. */ 1206 bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore, 1207 bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error); 1208 for (int i = 0; i < bp->b_npages; i++) 1209 bp->b_pages[i] = NULL; 1210 bp->b_vp = NULL; 1211 pbrelbo(bp); 1212 uma_zfree(vnode_pbuf_zone, bp); 1213} 1214 1215static int 1216vnode_pager_generic_getpages_done(struct buf *bp) 1217{ 1218 vm_object_t object; 1219 off_t tfoff, nextoff; 1220 int i, error; 1221 1222 KASSERT((bp->b_ioflags & BIO_ERROR) == 0 || bp->b_error != 0, 1223 ("%s: buf error but b_error == 0\n", __func__)); 1224 error = (bp->b_ioflags & BIO_ERROR) != 0 ? bp->b_error : 0; 1225 object = bp->b_vp->v_object; 1226 1227 runningbufwakeup(bp); 1228 1229 if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) { 1230 if (!buf_mapped(bp)) { 1231 bp->b_data = bp->b_kvabase; 1232 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, 1233 bp->b_npages); 1234 } 1235 bzero(bp->b_data + bp->b_bcount, 1236 PAGE_SIZE * bp->b_npages - bp->b_bcount); 1237 } 1238 if (buf_mapped(bp)) { 1239 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1240 bp->b_data = unmapped_buf; 1241 } 1242 1243 /* 1244 * If the read failed, we must free any read ahead/behind pages here. 1245 * The requested pages are freed by the caller (for sync requests) 1246 * or by the bp->b_pgiodone callback (for async requests). 1247 */ 1248 if (error != 0) { 1249 VM_OBJECT_WLOCK(object); 1250 for (i = 0; i < bp->b_pgbefore; i++) 1251 vm_page_free_invalid(bp->b_pages[i]); 1252 for (i = bp->b_npages - bp->b_pgafter; i < bp->b_npages; i++) 1253 vm_page_free_invalid(bp->b_pages[i]); 1254 VM_OBJECT_WUNLOCK(object); 1255 return (error); 1256 } 1257 1258 /* Read lock to protect size. */ 1259 VM_OBJECT_RLOCK(object); 1260 for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex); 1261 i < bp->b_npages; i++, tfoff = nextoff) { 1262 vm_page_t mt; 1263 1264 nextoff = tfoff + PAGE_SIZE; 1265 mt = bp->b_pages[i]; 1266 if (mt == bogus_page) 1267 continue; 1268 1269 if (nextoff <= object->un_pager.vnp.vnp_size) { 1270 /* 1271 * Read filled up entire page. 1272 */ 1273 vm_page_valid(mt); 1274 KASSERT(mt->dirty == 0, 1275 ("%s: page %p is dirty", __func__, mt)); 1276 KASSERT(!pmap_page_is_mapped(mt), 1277 ("%s: page %p is mapped", __func__, mt)); 1278 } else { 1279 /* 1280 * Read did not fill up entire page. 1281 * 1282 * Currently we do not set the entire page valid, 1283 * we just try to clear the piece that we couldn't 1284 * read. 1285 */ 1286 vm_page_set_valid_range(mt, 0, 1287 object->un_pager.vnp.vnp_size - tfoff); 1288 KASSERT((mt->dirty & vm_page_bits(0, 1289 object->un_pager.vnp.vnp_size - tfoff)) == 0, 1290 ("%s: page %p is dirty", __func__, mt)); 1291 } 1292 1293 if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter) 1294 vm_page_readahead_finish(mt); 1295 } 1296 VM_OBJECT_RUNLOCK(object); 1297 1298 return (error); 1299} 1300 1301/* 1302 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 1303 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 1304 * vnode_pager_generic_putpages() to implement the previous behaviour. 1305 * 1306 * All other FS's should use the bypass to get to the local media 1307 * backing vp's VOP_PUTPAGES. 1308 */ 1309static void 1310vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1311 int flags, int *rtvals) 1312{ 1313 int rtval __diagused; 1314 struct vnode *vp; 1315 int bytes = count * PAGE_SIZE; 1316 1317 /* 1318 * Force synchronous operation if we are extremely low on memory 1319 * to prevent a low-memory deadlock. VOP operations often need to 1320 * allocate more memory to initiate the I/O ( i.e. do a BMAP 1321 * operation ). The swapper handles the case by limiting the amount 1322 * of asynchronous I/O, but that sort of solution doesn't scale well 1323 * for the vnode pager without a lot of work. 1324 * 1325 * Also, the backing vnode's iodone routine may not wake the pageout 1326 * daemon up. This should be probably be addressed XXX. 1327 */ 1328 1329 if (vm_page_count_min()) 1330 flags |= VM_PAGER_PUT_SYNC; 1331 1332 /* 1333 * Call device-specific putpages function 1334 */ 1335 vp = object->handle; 1336 VM_OBJECT_WUNLOCK(object); 1337 rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals); 1338 KASSERT(rtval != EOPNOTSUPP, 1339 ("vnode_pager: stale FS putpages\n")); 1340 VM_OBJECT_WLOCK(object); 1341} 1342 1343static int 1344vn_off2bidx(vm_ooffset_t offset) 1345{ 1346 1347 return ((offset & PAGE_MASK) / DEV_BSIZE); 1348} 1349 1350static bool 1351vn_dirty_blk(vm_page_t m, vm_ooffset_t offset) 1352{ 1353 1354 KASSERT(IDX_TO_OFF(m->pindex) <= offset && 1355 offset < IDX_TO_OFF(m->pindex + 1), 1356 ("page %p pidx %ju offset %ju", m, (uintmax_t)m->pindex, 1357 (uintmax_t)offset)); 1358 return ((m->dirty & ((vm_page_bits_t)1 << vn_off2bidx(offset))) != 0); 1359} 1360 1361/* 1362 * This is now called from local media FS's to operate against their 1363 * own vnodes if they fail to implement VOP_PUTPAGES. 1364 * 1365 * This is typically called indirectly via the pageout daemon and 1366 * clustering has already typically occurred, so in general we ask the 1367 * underlying filesystem to write the data out asynchronously rather 1368 * then delayed. 1369 */ 1370int 1371vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount, 1372 int flags, int *rtvals) 1373{ 1374 vm_object_t object; 1375 vm_page_t m; 1376 vm_ooffset_t max_offset, next_offset, poffset, prev_offset; 1377 struct uio auio; 1378 struct iovec aiov; 1379 off_t prev_resid, wrsz; 1380 int count, error, i, maxsize, ncount, pgoff, ppscheck; 1381 bool in_hole; 1382 static struct timeval lastfail; 1383 static int curfail; 1384 1385 object = vp->v_object; 1386 count = bytecount / PAGE_SIZE; 1387 1388 for (i = 0; i < count; i++) 1389 rtvals[i] = VM_PAGER_ERROR; 1390 1391 if ((int64_t)ma[0]->pindex < 0) { 1392 printf("vnode_pager_generic_putpages: " 1393 "attempt to write meta-data 0x%jx(%lx)\n", 1394 (uintmax_t)ma[0]->pindex, (u_long)ma[0]->dirty); 1395 rtvals[0] = VM_PAGER_BAD; 1396 return (VM_PAGER_BAD); 1397 } 1398 1399 maxsize = count * PAGE_SIZE; 1400 ncount = count; 1401 1402 poffset = IDX_TO_OFF(ma[0]->pindex); 1403 1404 /* 1405 * If the page-aligned write is larger then the actual file we 1406 * have to invalidate pages occurring beyond the file EOF. However, 1407 * there is an edge case where a file may not be page-aligned where 1408 * the last page is partially invalid. In this case the filesystem 1409 * may not properly clear the dirty bits for the entire page (which 1410 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 1411 * With the page busied we are free to fix up the dirty bits here. 1412 * 1413 * We do not under any circumstances truncate the valid bits, as 1414 * this will screw up bogus page replacement. 1415 */ 1416 VM_OBJECT_RLOCK(object); 1417 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 1418 if (object->un_pager.vnp.vnp_size > poffset) { 1419 maxsize = object->un_pager.vnp.vnp_size - poffset; 1420 ncount = btoc(maxsize); 1421 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 1422 pgoff = roundup2(pgoff, DEV_BSIZE); 1423 1424 /* 1425 * If the page is busy and the following 1426 * conditions hold, then the page's dirty 1427 * field cannot be concurrently changed by a 1428 * pmap operation. 1429 */ 1430 m = ma[ncount - 1]; 1431 vm_page_assert_sbusied(m); 1432 KASSERT(!pmap_page_is_write_mapped(m), 1433 ("vnode_pager_generic_putpages: page %p is not read-only", m)); 1434 MPASS(m->dirty != 0); 1435 vm_page_clear_dirty(m, pgoff, PAGE_SIZE - 1436 pgoff); 1437 } 1438 } else { 1439 maxsize = 0; 1440 ncount = 0; 1441 } 1442 for (i = ncount; i < count; i++) 1443 rtvals[i] = VM_PAGER_BAD; 1444 } 1445 VM_OBJECT_RUNLOCK(object); 1446 1447 auio.uio_iov = &aiov; 1448 auio.uio_segflg = UIO_NOCOPY; 1449 auio.uio_rw = UIO_WRITE; 1450 auio.uio_td = NULL; 1451 max_offset = roundup2(poffset + maxsize, DEV_BSIZE); 1452 1453 for (prev_offset = poffset; prev_offset < max_offset;) { 1454 /* Skip clean blocks. */ 1455 for (in_hole = true; in_hole && prev_offset < max_offset;) { 1456 m = ma[OFF_TO_IDX(prev_offset - poffset)]; 1457 for (i = vn_off2bidx(prev_offset); 1458 i < sizeof(vm_page_bits_t) * NBBY && 1459 prev_offset < max_offset; i++) { 1460 if (vn_dirty_blk(m, prev_offset)) { 1461 in_hole = false; 1462 break; 1463 } 1464 prev_offset += DEV_BSIZE; 1465 } 1466 } 1467 if (in_hole) 1468 goto write_done; 1469 1470 /* Find longest run of dirty blocks. */ 1471 for (next_offset = prev_offset; next_offset < max_offset;) { 1472 m = ma[OFF_TO_IDX(next_offset - poffset)]; 1473 for (i = vn_off2bidx(next_offset); 1474 i < sizeof(vm_page_bits_t) * NBBY && 1475 next_offset < max_offset; i++) { 1476 if (!vn_dirty_blk(m, next_offset)) 1477 goto start_write; 1478 next_offset += DEV_BSIZE; 1479 } 1480 } 1481start_write: 1482 if (next_offset > poffset + maxsize) 1483 next_offset = poffset + maxsize; 1484 if (prev_offset == next_offset) 1485 goto write_done; 1486 1487 /* 1488 * Getting here requires finding a dirty block in the 1489 * 'skip clean blocks' loop. 1490 */ 1491 1492 aiov.iov_base = NULL; 1493 auio.uio_iovcnt = 1; 1494 auio.uio_offset = prev_offset; 1495 prev_resid = auio.uio_resid = aiov.iov_len = next_offset - 1496 prev_offset; 1497 error = VOP_WRITE(vp, &auio, 1498 vnode_pager_putpages_ioflags(flags), curthread->td_ucred); 1499 1500 wrsz = prev_resid - auio.uio_resid; 1501 if (wrsz == 0) { 1502 if (ppsratecheck(&lastfail, &curfail, 1) != 0) { 1503 vn_printf(vp, "vnode_pager_putpages: " 1504 "zero-length write at %ju resid %zd\n", 1505 auio.uio_offset, auio.uio_resid); 1506 } 1507 break; 1508 } 1509 1510 /* Adjust the starting offset for next iteration. */ 1511 prev_offset += wrsz; 1512 MPASS(auio.uio_offset == prev_offset); 1513 1514 ppscheck = 0; 1515 if (error != 0 && (ppscheck = ppsratecheck(&lastfail, 1516 &curfail, 1)) != 0) 1517 vn_printf(vp, "vnode_pager_putpages: I/O error %d\n", 1518 error); 1519 if (auio.uio_resid != 0 && (ppscheck != 0 || 1520 ppsratecheck(&lastfail, &curfail, 1) != 0)) 1521 vn_printf(vp, "vnode_pager_putpages: residual I/O %zd " 1522 "at %ju\n", auio.uio_resid, 1523 (uintmax_t)ma[0]->pindex); 1524 if (error != 0 || auio.uio_resid != 0) 1525 break; 1526 } 1527write_done: 1528 /* Mark completely processed pages. */ 1529 for (i = 0; i < OFF_TO_IDX(prev_offset - poffset); i++) 1530 rtvals[i] = VM_PAGER_OK; 1531 /* Mark partial EOF page. */ 1532 if (prev_offset == poffset + maxsize && (prev_offset & PAGE_MASK) != 0) 1533 rtvals[i++] = VM_PAGER_OK; 1534 /* Unwritten pages in range, free bonus if the page is clean. */ 1535 for (; i < ncount; i++) 1536 rtvals[i] = ma[i]->dirty == 0 ? VM_PAGER_OK : VM_PAGER_ERROR; 1537 VM_CNT_ADD(v_vnodepgsout, i); 1538 VM_CNT_INC(v_vnodeout); 1539 return (rtvals[0]); 1540} 1541 1542int 1543vnode_pager_putpages_ioflags(int pager_flags) 1544{ 1545 int ioflags; 1546 1547 /* 1548 * Pageouts are already clustered, use IO_ASYNC to force a 1549 * bawrite() rather then a bdwrite() to prevent paging I/O 1550 * from saturating the buffer cache. Dummy-up the sequential 1551 * heuristic to cause large ranges to cluster. If neither 1552 * IO_SYNC or IO_ASYNC is set, the system decides how to 1553 * cluster. 1554 */ 1555 ioflags = IO_VMIO; 1556 if ((pager_flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) != 0) 1557 ioflags |= IO_SYNC; 1558 else if ((pager_flags & VM_PAGER_CLUSTER_OK) == 0) 1559 ioflags |= IO_ASYNC; 1560 ioflags |= (pager_flags & VM_PAGER_PUT_INVAL) != 0 ? IO_INVAL: 0; 1561 ioflags |= (pager_flags & VM_PAGER_PUT_NOREUSE) != 0 ? IO_NOREUSE : 0; 1562 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 1563 return (ioflags); 1564} 1565 1566/* 1567 * vnode_pager_undirty_pages(). 1568 * 1569 * A helper to mark pages as clean after pageout that was possibly 1570 * done with a short write. The lpos argument specifies the page run 1571 * length in bytes, and the written argument specifies how many bytes 1572 * were actually written. eof is the offset past the last valid byte 1573 * in the vnode using the absolute file position of the first byte in 1574 * the run as the base from which it is computed. 1575 */ 1576void 1577vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written, off_t eof, 1578 int lpos) 1579{ 1580 int i, pos, pos_devb; 1581 1582 if (written == 0 && eof >= lpos) 1583 return; 1584 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) { 1585 if (pos < trunc_page(written)) { 1586 rtvals[i] = VM_PAGER_OK; 1587 vm_page_undirty(ma[i]); 1588 } else { 1589 /* Partially written page. */ 1590 rtvals[i] = VM_PAGER_AGAIN; 1591 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK); 1592 } 1593 } 1594 if (eof >= lpos) /* avoid truncation */ 1595 return; 1596 for (pos = eof, i = OFF_TO_IDX(trunc_page(pos)); pos < lpos; i++) { 1597 if (pos != trunc_page(pos)) { 1598 /* 1599 * The page contains the last valid byte in 1600 * the vnode, mark the rest of the page as 1601 * clean, potentially making the whole page 1602 * clean. 1603 */ 1604 pos_devb = roundup2(pos & PAGE_MASK, DEV_BSIZE); 1605 vm_page_clear_dirty(ma[i], pos_devb, PAGE_SIZE - 1606 pos_devb); 1607 1608 /* 1609 * If the page was cleaned, report the pageout 1610 * on it as successful. msync() no longer 1611 * needs to write out the page, endlessly 1612 * creating write requests and dirty buffers. 1613 */ 1614 if (ma[i]->dirty == 0) 1615 rtvals[i] = VM_PAGER_OK; 1616 1617 pos = round_page(pos); 1618 } else { 1619 /* vm_pageout_flush() clears dirty */ 1620 rtvals[i] = VM_PAGER_BAD; 1621 pos += PAGE_SIZE; 1622 } 1623 } 1624} 1625 1626static void 1627vnode_pager_update_writecount(vm_object_t object, vm_offset_t start, 1628 vm_offset_t end) 1629{ 1630 struct vnode *vp; 1631 vm_ooffset_t old_wm; 1632 1633 VM_OBJECT_WLOCK(object); 1634 if (object->type != OBJT_VNODE) { 1635 VM_OBJECT_WUNLOCK(object); 1636 return; 1637 } 1638 old_wm = object->un_pager.vnp.writemappings; 1639 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start; 1640 vp = object->handle; 1641 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) { 1642 ASSERT_VOP_LOCKED(vp, "v_writecount inc"); 1643 VOP_ADD_WRITECOUNT_CHECKED(vp, 1); 1644 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", 1645 __func__, vp, vp->v_writecount); 1646 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) { 1647 ASSERT_VOP_LOCKED(vp, "v_writecount dec"); 1648 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 1649 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 1650 __func__, vp, vp->v_writecount); 1651 } 1652 VM_OBJECT_WUNLOCK(object); 1653} 1654 1655static void 1656vnode_pager_release_writecount(vm_object_t object, vm_offset_t start, 1657 vm_offset_t end) 1658{ 1659 struct vnode *vp; 1660 struct mount *mp; 1661 vm_offset_t inc; 1662 1663 VM_OBJECT_WLOCK(object); 1664 1665 /* 1666 * First, recheck the object type to account for the race when 1667 * the vnode is reclaimed. 1668 */ 1669 if (object->type != OBJT_VNODE) { 1670 VM_OBJECT_WUNLOCK(object); 1671 return; 1672 } 1673 1674 /* 1675 * Optimize for the case when writemappings is not going to 1676 * zero. 1677 */ 1678 inc = end - start; 1679 if (object->un_pager.vnp.writemappings != inc) { 1680 object->un_pager.vnp.writemappings -= inc; 1681 VM_OBJECT_WUNLOCK(object); 1682 return; 1683 } 1684 1685 vp = object->handle; 1686 vhold(vp); 1687 VM_OBJECT_WUNLOCK(object); 1688 mp = NULL; 1689 vn_start_write(vp, &mp, V_WAIT); 1690 vn_lock(vp, LK_SHARED | LK_RETRY); 1691 1692 /* 1693 * Decrement the object's writemappings, by swapping the start 1694 * and end arguments for vnode_pager_update_writecount(). If 1695 * there was not a race with vnode reclaimation, then the 1696 * vnode's v_writecount is decremented. 1697 */ 1698 vnode_pager_update_writecount(object, end, start); 1699 VOP_UNLOCK(vp); 1700 vdrop(vp); 1701 if (mp != NULL) 1702 vn_finished_write(mp); 1703} 1704 1705static void 1706vnode_pager_getvp(vm_object_t object, struct vnode **vpp, bool *vp_heldp) 1707{ 1708 *vpp = object->handle; 1709} 1710 1711static void 1712vnode_pager_clean1(struct vnode *vp, int sync_flags) 1713{ 1714 struct vm_object *obj; 1715 1716 ASSERT_VOP_LOCKED(vp, "needs lock for writes"); 1717 obj = vp->v_object; 1718 if (obj == NULL) 1719 return; 1720 1721 VM_OBJECT_WLOCK(obj); 1722 vm_object_page_clean(obj, 0, 0, sync_flags); 1723 VM_OBJECT_WUNLOCK(obj); 1724} 1725 1726void 1727vnode_pager_clean_sync(struct vnode *vp) 1728{ 1729 vnode_pager_clean1(vp, OBJPC_SYNC); 1730} 1731 1732void 1733vnode_pager_clean_async(struct vnode *vp) 1734{ 1735 vnode_pager_clean1(vp, 0); 1736} 1737