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 * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91 43 */ 44 45 /* 46 * Page to/from files (vnodes). 47 */ 48 49 /* 50 * TODO: 51 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will 52 * greatly re-simplify the vnode_pager. 53 */ 54 55 #include <sys/cdefs.h> 56 __FBSDID("$FreeBSD$"); 57 58 #include "opt_vm.h" 59 60 #include <sys/param.h> 61 #include <sys/kernel.h> 62 #include <sys/systm.h> 63 #include <sys/sysctl.h> 64 #include <sys/proc.h> 65 #include <sys/vnode.h> 66 #include <sys/mount.h> 67 #include <sys/bio.h> 68 #include <sys/buf.h> 69 #include <sys/vmmeter.h> 70 #include <sys/ktr.h> 71 #include <sys/limits.h> 72 #include <sys/conf.h> 73 #include <sys/rwlock.h> 74 #include <sys/sf_buf.h> 75 #include <sys/domainset.h> 76 77 #include <machine/atomic.h> 78 79 #include <vm/vm.h> 80 #include <vm/vm_param.h> 81 #include <vm/vm_object.h> 82 #include <vm/vm_page.h> 83 #include <vm/vm_pager.h> 84 #include <vm/vm_map.h> 85 #include <vm/vnode_pager.h> 86 #include <vm/vm_extern.h> 87 #include <vm/uma.h> 88 89 static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, 90 daddr_t *rtaddress, int *run); 91 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m); 92 static int vnode_pager_input_old(vm_object_t object, vm_page_t m); 93 static void vnode_pager_dealloc(vm_object_t); 94 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *); 95 static int vnode_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 96 int *, vop_getpages_iodone_t, void *); 97 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *); 98 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *); 99 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t, 100 vm_ooffset_t, struct ucred *cred); 101 static int vnode_pager_generic_getpages_done(struct buf *); 102 static void vnode_pager_generic_getpages_done_async(struct buf *); 103 104 struct pagerops vnodepagerops = { 105 .pgo_alloc = vnode_pager_alloc, 106 .pgo_dealloc = vnode_pager_dealloc, 107 .pgo_getpages = vnode_pager_getpages, 108 .pgo_getpages_async = vnode_pager_getpages_async, 109 .pgo_putpages = vnode_pager_putpages, 110 .pgo_haspage = vnode_pager_haspage, 111 }; 112 113 static struct domainset *vnode_domainset = NULL; 114 115 SYSCTL_PROC(_debug, OID_AUTO, vnode_domainset, CTLTYPE_STRING | CTLFLAG_RW, 116 &vnode_domainset, 0, sysctl_handle_domainset, "A", 117 "Default vnode NUMA policy"); 118 119 static int nvnpbufs; 120 SYSCTL_INT(_vm, OID_AUTO, vnode_pbufs, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 121 &nvnpbufs, 0, "number of physical buffers allocated for vnode pager"); 122 123 static uma_zone_t vnode_pbuf_zone; 124 125 static void 126 vnode_pager_init(void *dummy) 127 { 128 129 #ifdef __LP64__ 130 nvnpbufs = nswbuf * 2; 131 #else 132 nvnpbufs = nswbuf / 2; 133 #endif 134 TUNABLE_INT_FETCH("vm.vnode_pbufs", &nvnpbufs); 135 vnode_pbuf_zone = pbuf_zsecond_create("vnpbuf", nvnpbufs); 136 } 137 SYSINIT(vnode_pager, SI_SUB_CPU, SI_ORDER_ANY, vnode_pager_init, NULL); 138 139 /* Create the VM system backing object for this vnode */ 140 int 141 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td) 142 { 143 vm_object_t object; 144 vm_ooffset_t size = isize; 145 struct vattr va; 146 147 if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE) 148 return (0); 149 150 while ((object = vp->v_object) != NULL) { 151 VM_OBJECT_WLOCK(object); 152 if (!(object->flags & OBJ_DEAD)) { 153 VM_OBJECT_WUNLOCK(object); 154 return (0); 155 } 156 VOP_UNLOCK(vp, 0); 157 vm_object_set_flag(object, OBJ_DISCONNECTWNT); 158 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vodead", 0); 159 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 160 } 161 162 if (size == 0) { 163 if (vn_isdisk(vp, NULL)) { 164 size = IDX_TO_OFF(INT_MAX); 165 } else { 166 if (VOP_GETATTR(vp, &va, td->td_ucred)) 167 return (0); 168 size = va.va_size; 169 } 170 } 171 172 object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred); 173 /* 174 * Dereference the reference we just created. This assumes 175 * that the object is associated with the vp. 176 */ 177 VM_OBJECT_WLOCK(object); 178 object->ref_count--; 179 VM_OBJECT_WUNLOCK(object); 180 vrele(vp); 181 182 KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object")); 183 184 return (0); 185 } 186 187 void 188 vnode_destroy_vobject(struct vnode *vp) 189 { 190 struct vm_object *obj; 191 192 obj = vp->v_object; 193 if (obj == NULL) 194 return; 195 ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject"); 196 VM_OBJECT_WLOCK(obj); 197 umtx_shm_object_terminated(obj); 198 if (obj->ref_count == 0) { 199 /* 200 * don't double-terminate the object 201 */ 202 if ((obj->flags & OBJ_DEAD) == 0) { 203 vm_object_set_flag(obj, OBJ_DEAD); 204 205 /* 206 * Clean pages and flush buffers. 207 */ 208 vm_object_page_clean(obj, 0, 0, OBJPC_SYNC); 209 VM_OBJECT_WUNLOCK(obj); 210 211 vinvalbuf(vp, V_SAVE, 0, 0); 212 213 BO_LOCK(&vp->v_bufobj); 214 vp->v_bufobj.bo_flag |= BO_DEAD; 215 BO_UNLOCK(&vp->v_bufobj); 216 217 VM_OBJECT_WLOCK(obj); 218 vm_object_terminate(obj); 219 } else { 220 /* 221 * Waiters were already handled during object 222 * termination. The exclusive vnode lock hopefully 223 * prevented new waiters from referencing the dying 224 * object. 225 */ 226 KASSERT((obj->flags & OBJ_DISCONNECTWNT) == 0, 227 ("OBJ_DISCONNECTWNT set obj %p flags %x", 228 obj, obj->flags)); 229 vp->v_object = NULL; 230 VM_OBJECT_WUNLOCK(obj); 231 } 232 } else { 233 /* 234 * Woe to the process that tries to page now :-). 235 */ 236 vm_pager_deallocate(obj); 237 VM_OBJECT_WUNLOCK(obj); 238 } 239 KASSERT(vp->v_object == NULL, ("vp %p obj %p", vp, vp->v_object)); 240 } 241 242 243 /* 244 * Allocate (or lookup) pager for a vnode. 245 * Handle is a vnode pointer. 246 * 247 * MPSAFE 248 */ 249 vm_object_t 250 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 251 vm_ooffset_t offset, struct ucred *cred) 252 { 253 vm_object_t object; 254 struct vnode *vp; 255 256 /* 257 * Pageout to vnode, no can do yet. 258 */ 259 if (handle == NULL) 260 return (NULL); 261 262 vp = (struct vnode *) handle; 263 264 /* 265 * If the object is being terminated, wait for it to 266 * go away. 267 */ 268 retry: 269 while ((object = vp->v_object) != NULL) { 270 VM_OBJECT_WLOCK(object); 271 if ((object->flags & OBJ_DEAD) == 0) 272 break; 273 vm_object_set_flag(object, OBJ_DISCONNECTWNT); 274 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vadead", 0); 275 } 276 277 KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference")); 278 279 if (object == NULL) { 280 /* 281 * Add an object of the appropriate size 282 */ 283 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); 284 285 object->un_pager.vnp.vnp_size = size; 286 object->un_pager.vnp.writemappings = 0; 287 object->domain.dr_policy = vnode_domainset; 288 289 object->handle = handle; 290 VI_LOCK(vp); 291 if (vp->v_object != NULL) { 292 /* 293 * Object has been created while we were sleeping 294 */ 295 VI_UNLOCK(vp); 296 VM_OBJECT_WLOCK(object); 297 KASSERT(object->ref_count == 1, 298 ("leaked ref %p %d", object, object->ref_count)); 299 object->type = OBJT_DEAD; 300 object->ref_count = 0; 301 VM_OBJECT_WUNLOCK(object); 302 vm_object_destroy(object); 303 goto retry; 304 } 305 vp->v_object = object; 306 VI_UNLOCK(vp); 307 } else { 308 object->ref_count++; 309 #if VM_NRESERVLEVEL > 0 310 vm_object_color(object, 0); 311 #endif 312 VM_OBJECT_WUNLOCK(object); 313 } 314 vrefact(vp); 315 return (object); 316 } 317 318 /* 319 * The object must be locked. 320 */ 321 static void 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 if (object->flags & OBJ_DISCONNECTWNT) { 338 vm_object_clear_flag(object, OBJ_DISCONNECTWNT); 339 wakeup(object); 340 } 341 ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc"); 342 if (object->un_pager.vnp.writemappings > 0) { 343 object->un_pager.vnp.writemappings = 0; 344 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 345 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 346 __func__, vp, vp->v_writecount); 347 } 348 vp->v_object = NULL; 349 VI_LOCK(vp); 350 351 /* 352 * vm_map_entry_set_vnode_text() cannot reach this vnode by 353 * following object->handle. Clear all text references now. 354 * This also clears the transient references from 355 * kern_execve(), which is fine because dead_vnodeops uses nop 356 * for VOP_UNSET_TEXT(). 357 */ 358 if (vp->v_writecount < 0) 359 vp->v_writecount = 0; 360 VI_UNLOCK(vp); 361 VM_OBJECT_WUNLOCK(object); 362 while (refs-- > 0) 363 vunref(vp); 364 VM_OBJECT_WLOCK(object); 365 } 366 367 static boolean_t 368 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 369 int *after) 370 { 371 struct vnode *vp = object->handle; 372 daddr_t bn; 373 uintptr_t lockstate; 374 int err; 375 daddr_t reqblock; 376 int poff; 377 int bsize; 378 int pagesperblock, blocksperpage; 379 380 VM_OBJECT_ASSERT_LOCKED(object); 381 /* 382 * If no vp or vp is doomed or marked transparent to VM, we do not 383 * have the page. 384 */ 385 if (vp == NULL || vp->v_iflag & VI_DOOMED) 386 return FALSE; 387 /* 388 * If the offset is beyond end of file we do 389 * not have the page. 390 */ 391 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size) 392 return FALSE; 393 394 bsize = vp->v_mount->mnt_stat.f_iosize; 395 pagesperblock = bsize / PAGE_SIZE; 396 blocksperpage = 0; 397 if (pagesperblock > 0) { 398 reqblock = pindex / pagesperblock; 399 } else { 400 blocksperpage = (PAGE_SIZE / bsize); 401 reqblock = pindex * blocksperpage; 402 } 403 lockstate = VM_OBJECT_DROP(object); 404 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before); 405 VM_OBJECT_PICKUP(object, lockstate); 406 if (err) 407 return TRUE; 408 if (bn == -1) 409 return FALSE; 410 if (pagesperblock > 0) { 411 poff = pindex - (reqblock * pagesperblock); 412 if (before) { 413 *before *= pagesperblock; 414 *before += poff; 415 } 416 if (after) { 417 /* 418 * The BMAP vop can report a partial block in the 419 * 'after', but must not report blocks after EOF. 420 * Assert the latter, and truncate 'after' in case 421 * of the former. 422 */ 423 KASSERT((reqblock + *after) * pagesperblock < 424 roundup2(object->size, pagesperblock), 425 ("%s: reqblock %jd after %d size %ju", __func__, 426 (intmax_t )reqblock, *after, 427 (uintmax_t )object->size)); 428 *after *= pagesperblock; 429 *after += pagesperblock - (poff + 1); 430 if (pindex + *after >= object->size) 431 *after = object->size - 1 - pindex; 432 } 433 } else { 434 if (before) { 435 *before /= blocksperpage; 436 } 437 438 if (after) { 439 *after /= blocksperpage; 440 } 441 } 442 return TRUE; 443 } 444 445 /* 446 * Lets the VM system know about a change in size for a file. 447 * We adjust our own internal size and flush any cached pages in 448 * the associated object that are affected by the size change. 449 * 450 * Note: this routine may be invoked as a result of a pager put 451 * operation (possibly at object termination time), so we must be careful. 452 */ 453 void 454 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize) 455 { 456 vm_object_t object; 457 vm_page_t m; 458 vm_pindex_t nobjsize; 459 460 if ((object = vp->v_object) == NULL) 461 return; 462 /* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */ 463 VM_OBJECT_WLOCK(object); 464 if (object->type == OBJT_DEAD) { 465 VM_OBJECT_WUNLOCK(object); 466 return; 467 } 468 KASSERT(object->type == OBJT_VNODE, 469 ("not vnode-backed object %p", object)); 470 if (nsize == object->un_pager.vnp.vnp_size) { 471 /* 472 * Hasn't changed size 473 */ 474 VM_OBJECT_WUNLOCK(object); 475 return; 476 } 477 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 478 if (nsize < object->un_pager.vnp.vnp_size) { 479 /* 480 * File has shrunk. Toss any cached pages beyond the new EOF. 481 */ 482 if (nobjsize < object->size) 483 vm_object_page_remove(object, nobjsize, object->size, 484 0); 485 /* 486 * this gets rid of garbage at the end of a page that is now 487 * only partially backed by the vnode. 488 * 489 * XXX for some reason (I don't know yet), if we take a 490 * completely invalid page and mark it partially valid 491 * it can screw up NFS reads, so we don't allow the case. 492 */ 493 if ((nsize & PAGE_MASK) && 494 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL && 495 m->valid != 0) { 496 int base = (int)nsize & PAGE_MASK; 497 int size = PAGE_SIZE - base; 498 499 /* 500 * Clear out partial-page garbage in case 501 * the page has been mapped. 502 */ 503 pmap_zero_page_area(m, base, size); 504 505 /* 506 * Update the valid bits to reflect the blocks that 507 * have been zeroed. Some of these valid bits may 508 * have already been set. 509 */ 510 vm_page_set_valid_range(m, base, size); 511 512 /* 513 * Round "base" to the next block boundary so that the 514 * dirty bit for a partially zeroed block is not 515 * cleared. 516 */ 517 base = roundup2(base, DEV_BSIZE); 518 519 /* 520 * Clear out partial-page dirty bits. 521 * 522 * note that we do not clear out the valid 523 * bits. This would prevent bogus_page 524 * replacement from working properly. 525 */ 526 vm_page_clear_dirty(m, base, PAGE_SIZE - base); 527 } 528 } 529 object->un_pager.vnp.vnp_size = nsize; 530 object->size = nobjsize; 531 VM_OBJECT_WUNLOCK(object); 532 } 533 534 /* 535 * calculate the linear (byte) disk address of specified virtual 536 * file address 537 */ 538 static int 539 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress, 540 int *run) 541 { 542 int bsize; 543 int err; 544 daddr_t vblock; 545 daddr_t voffset; 546 547 if (address < 0) 548 return -1; 549 550 if (vp->v_iflag & VI_DOOMED) 551 return -1; 552 553 bsize = vp->v_mount->mnt_stat.f_iosize; 554 vblock = address / bsize; 555 voffset = address % bsize; 556 557 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL); 558 if (err == 0) { 559 if (*rtaddress != -1) 560 *rtaddress += voffset / DEV_BSIZE; 561 if (run) { 562 *run += 1; 563 *run *= bsize / PAGE_SIZE; 564 *run -= voffset / PAGE_SIZE; 565 } 566 } 567 568 return (err); 569 } 570 571 /* 572 * small block filesystem vnode pager input 573 */ 574 static int 575 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m) 576 { 577 struct vnode *vp; 578 struct bufobj *bo; 579 struct buf *bp; 580 struct sf_buf *sf; 581 daddr_t fileaddr; 582 vm_offset_t bsize; 583 vm_page_bits_t bits; 584 int error, i; 585 586 error = 0; 587 vp = object->handle; 588 if (vp->v_iflag & VI_DOOMED) 589 return VM_PAGER_BAD; 590 591 bsize = vp->v_mount->mnt_stat.f_iosize; 592 593 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL); 594 595 sf = sf_buf_alloc(m, 0); 596 597 for (i = 0; i < PAGE_SIZE / bsize; i++) { 598 vm_ooffset_t address; 599 600 bits = vm_page_bits(i * bsize, bsize); 601 if (m->valid & bits) 602 continue; 603 604 address = IDX_TO_OFF(m->pindex) + i * bsize; 605 if (address >= object->un_pager.vnp.vnp_size) { 606 fileaddr = -1; 607 } else { 608 error = vnode_pager_addr(vp, address, &fileaddr, NULL); 609 if (error) 610 break; 611 } 612 if (fileaddr != -1) { 613 bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK); 614 615 /* build a minimal buffer header */ 616 bp->b_iocmd = BIO_READ; 617 bp->b_iodone = bdone; 618 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 619 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 620 bp->b_rcred = crhold(curthread->td_ucred); 621 bp->b_wcred = crhold(curthread->td_ucred); 622 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize; 623 bp->b_blkno = fileaddr; 624 pbgetbo(bo, bp); 625 bp->b_vp = vp; 626 bp->b_bcount = bsize; 627 bp->b_bufsize = bsize; 628 bp->b_runningbufspace = bp->b_bufsize; 629 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 630 631 /* do the input */ 632 bp->b_iooffset = dbtob(bp->b_blkno); 633 bstrategy(bp); 634 635 bwait(bp, PVM, "vnsrd"); 636 637 if ((bp->b_ioflags & BIO_ERROR) != 0) 638 error = EIO; 639 640 /* 641 * free the buffer header back to the swap buffer pool 642 */ 643 bp->b_vp = NULL; 644 pbrelbo(bp); 645 uma_zfree(vnode_pbuf_zone, bp); 646 if (error) 647 break; 648 } else 649 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize); 650 KASSERT((m->dirty & bits) == 0, 651 ("vnode_pager_input_smlfs: page %p is dirty", m)); 652 VM_OBJECT_WLOCK(object); 653 m->valid |= bits; 654 VM_OBJECT_WUNLOCK(object); 655 } 656 sf_buf_free(sf); 657 if (error) { 658 return VM_PAGER_ERROR; 659 } 660 return VM_PAGER_OK; 661 } 662 663 /* 664 * old style vnode pager input routine 665 */ 666 static int 667 vnode_pager_input_old(vm_object_t object, vm_page_t m) 668 { 669 struct uio auio; 670 struct iovec aiov; 671 int error; 672 int size; 673 struct sf_buf *sf; 674 struct vnode *vp; 675 676 VM_OBJECT_ASSERT_WLOCKED(object); 677 error = 0; 678 679 /* 680 * Return failure if beyond current EOF 681 */ 682 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 683 return VM_PAGER_BAD; 684 } else { 685 size = PAGE_SIZE; 686 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 687 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 688 vp = object->handle; 689 VM_OBJECT_WUNLOCK(object); 690 691 /* 692 * Allocate a kernel virtual address and initialize so that 693 * we can use VOP_READ/WRITE routines. 694 */ 695 sf = sf_buf_alloc(m, 0); 696 697 aiov.iov_base = (caddr_t)sf_buf_kva(sf); 698 aiov.iov_len = size; 699 auio.uio_iov = &aiov; 700 auio.uio_iovcnt = 1; 701 auio.uio_offset = IDX_TO_OFF(m->pindex); 702 auio.uio_segflg = UIO_SYSSPACE; 703 auio.uio_rw = UIO_READ; 704 auio.uio_resid = size; 705 auio.uio_td = curthread; 706 707 error = VOP_READ(vp, &auio, 0, curthread->td_ucred); 708 if (!error) { 709 int count = size - auio.uio_resid; 710 711 if (count == 0) 712 error = EINVAL; 713 else if (count != PAGE_SIZE) 714 bzero((caddr_t)sf_buf_kva(sf) + count, 715 PAGE_SIZE - count); 716 } 717 sf_buf_free(sf); 718 719 VM_OBJECT_WLOCK(object); 720 } 721 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m)); 722 if (!error) 723 m->valid = VM_PAGE_BITS_ALL; 724 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 725 } 726 727 /* 728 * generic vnode pager input routine 729 */ 730 731 /* 732 * Local media VFS's that do not implement their own VOP_GETPAGES 733 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages() 734 * to implement the previous behaviour. 735 * 736 * All other FS's should use the bypass to get to the local media 737 * backing vp's VOP_GETPAGES. 738 */ 739 static int 740 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind, 741 int *rahead) 742 { 743 struct vnode *vp; 744 int rtval; 745 746 vp = object->handle; 747 VM_OBJECT_WUNLOCK(object); 748 rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead); 749 KASSERT(rtval != EOPNOTSUPP, 750 ("vnode_pager: FS getpages not implemented\n")); 751 VM_OBJECT_WLOCK(object); 752 return rtval; 753 } 754 755 static int 756 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count, 757 int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg) 758 { 759 struct vnode *vp; 760 int rtval; 761 762 vp = object->handle; 763 VM_OBJECT_WUNLOCK(object); 764 rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg); 765 KASSERT(rtval != EOPNOTSUPP, 766 ("vnode_pager: FS getpages_async not implemented\n")); 767 VM_OBJECT_WLOCK(object); 768 return (rtval); 769 } 770 771 /* 772 * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for 773 * local filesystems, where partially valid pages can only occur at 774 * the end of file. 775 */ 776 int 777 vnode_pager_local_getpages(struct vop_getpages_args *ap) 778 { 779 780 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count, 781 ap->a_rbehind, ap->a_rahead, NULL, NULL)); 782 } 783 784 int 785 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap) 786 { 787 788 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count, 789 ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg)); 790 } 791 792 /* 793 * This is now called from local media FS's to operate against their 794 * own vnodes if they fail to implement VOP_GETPAGES. 795 */ 796 int 797 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count, 798 int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg) 799 { 800 vm_object_t object; 801 struct bufobj *bo; 802 struct buf *bp; 803 off_t foff; 804 #ifdef INVARIANTS 805 off_t blkno0; 806 #endif 807 int bsize, pagesperblock; 808 int error, before, after, rbehind, rahead, poff, i; 809 int bytecount, secmask; 810 811 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 812 ("%s does not support devices", __func__)); 813 814 if (vp->v_iflag & VI_DOOMED) 815 return (VM_PAGER_BAD); 816 817 object = vp->v_object; 818 foff = IDX_TO_OFF(m[0]->pindex); 819 bsize = vp->v_mount->mnt_stat.f_iosize; 820 pagesperblock = bsize / PAGE_SIZE; 821 822 KASSERT(foff < object->un_pager.vnp.vnp_size, 823 ("%s: page %p offset beyond vp %p size", __func__, m[0], vp)); 824 KASSERT(count <= nitems(bp->b_pages), 825 ("%s: requested %d pages", __func__, count)); 826 827 /* 828 * The last page has valid blocks. Invalid part can only 829 * exist at the end of file, and the page is made fully valid 830 * by zeroing in vm_pager_get_pages(). 831 */ 832 if (m[count - 1]->valid != 0 && --count == 0) { 833 if (iodone != NULL) 834 iodone(arg, m, 1, 0); 835 return (VM_PAGER_OK); 836 } 837 838 bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK); 839 840 /* 841 * Get the underlying device blocks for the file with VOP_BMAP(). 842 * If the file system doesn't support VOP_BMAP, use old way of 843 * getting pages via VOP_READ. 844 */ 845 error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before); 846 if (error == EOPNOTSUPP) { 847 uma_zfree(vnode_pbuf_zone, bp); 848 VM_OBJECT_WLOCK(object); 849 for (i = 0; i < count; i++) { 850 VM_CNT_INC(v_vnodein); 851 VM_CNT_INC(v_vnodepgsin); 852 error = vnode_pager_input_old(object, m[i]); 853 if (error) 854 break; 855 } 856 VM_OBJECT_WUNLOCK(object); 857 return (error); 858 } else if (error != 0) { 859 uma_zfree(vnode_pbuf_zone, bp); 860 return (VM_PAGER_ERROR); 861 } 862 863 /* 864 * If the file system supports BMAP, but blocksize is smaller 865 * than a page size, then use special small filesystem code. 866 */ 867 if (pagesperblock == 0) { 868 uma_zfree(vnode_pbuf_zone, bp); 869 for (i = 0; i < count; i++) { 870 VM_CNT_INC(v_vnodein); 871 VM_CNT_INC(v_vnodepgsin); 872 error = vnode_pager_input_smlfs(object, m[i]); 873 if (error) 874 break; 875 } 876 return (error); 877 } 878 879 /* 880 * A sparse file can be encountered only for a single page request, 881 * which may not be preceded by call to vm_pager_haspage(). 882 */ 883 if (bp->b_blkno == -1) { 884 KASSERT(count == 1, 885 ("%s: array[%d] request to a sparse file %p", __func__, 886 count, vp)); 887 uma_zfree(vnode_pbuf_zone, bp); 888 pmap_zero_page(m[0]); 889 KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty", 890 __func__, m[0])); 891 VM_OBJECT_WLOCK(object); 892 m[0]->valid = VM_PAGE_BITS_ALL; 893 VM_OBJECT_WUNLOCK(object); 894 return (VM_PAGER_OK); 895 } 896 897 #ifdef INVARIANTS 898 blkno0 = bp->b_blkno; 899 #endif 900 bp->b_blkno += (foff % bsize) / DEV_BSIZE; 901 902 /* Recalculate blocks available after/before to pages. */ 903 poff = (foff % bsize) / PAGE_SIZE; 904 before *= pagesperblock; 905 before += poff; 906 after *= pagesperblock; 907 after += pagesperblock - (poff + 1); 908 if (m[0]->pindex + after >= object->size) 909 after = object->size - 1 - m[0]->pindex; 910 KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d", 911 __func__, count, after + 1)); 912 after -= count - 1; 913 914 /* Trim requested rbehind/rahead to possible values. */ 915 rbehind = a_rbehind ? *a_rbehind : 0; 916 rahead = a_rahead ? *a_rahead : 0; 917 rbehind = min(rbehind, before); 918 rbehind = min(rbehind, m[0]->pindex); 919 rahead = min(rahead, after); 920 rahead = min(rahead, object->size - m[count - 1]->pindex); 921 /* 922 * Check that total amount of pages fit into buf. Trim rbehind and 923 * rahead evenly if not. 924 */ 925 if (rbehind + rahead + count > nitems(bp->b_pages)) { 926 int trim, sum; 927 928 trim = rbehind + rahead + count - nitems(bp->b_pages) + 1; 929 sum = rbehind + rahead; 930 if (rbehind == before) { 931 /* Roundup rbehind trim to block size. */ 932 rbehind -= roundup(trim * rbehind / sum, pagesperblock); 933 if (rbehind < 0) 934 rbehind = 0; 935 } else 936 rbehind -= trim * rbehind / sum; 937 rahead -= trim * rahead / sum; 938 } 939 KASSERT(rbehind + rahead + count <= nitems(bp->b_pages), 940 ("%s: behind %d ahead %d count %d", __func__, 941 rbehind, rahead, count)); 942 943 /* 944 * Fill in the bp->b_pages[] array with requested and optional 945 * read behind or read ahead pages. Read behind pages are looked 946 * up in a backward direction, down to a first cached page. Same 947 * for read ahead pages, but there is no need to shift the array 948 * in case of encountering a cached page. 949 */ 950 i = bp->b_npages = 0; 951 if (rbehind) { 952 vm_pindex_t startpindex, tpindex; 953 vm_page_t p; 954 955 VM_OBJECT_WLOCK(object); 956 startpindex = m[0]->pindex - rbehind; 957 if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL && 958 p->pindex >= startpindex) 959 startpindex = p->pindex + 1; 960 961 /* tpindex is unsigned; beware of numeric underflow. */ 962 for (tpindex = m[0]->pindex - 1; 963 tpindex >= startpindex && tpindex < m[0]->pindex; 964 tpindex--, i++) { 965 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 966 if (p == NULL) { 967 /* Shift the array. */ 968 for (int j = 0; j < i; j++) 969 bp->b_pages[j] = bp->b_pages[j + 970 tpindex + 1 - startpindex]; 971 break; 972 } 973 bp->b_pages[tpindex - startpindex] = p; 974 } 975 976 bp->b_pgbefore = i; 977 bp->b_npages += i; 978 bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE; 979 } else 980 bp->b_pgbefore = 0; 981 982 /* Requested pages. */ 983 for (int j = 0; j < count; j++, i++) 984 bp->b_pages[i] = m[j]; 985 bp->b_npages += count; 986 987 if (rahead) { 988 vm_pindex_t endpindex, tpindex; 989 vm_page_t p; 990 991 if (!VM_OBJECT_WOWNED(object)) 992 VM_OBJECT_WLOCK(object); 993 endpindex = m[count - 1]->pindex + rahead + 1; 994 if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL && 995 p->pindex < endpindex) 996 endpindex = p->pindex; 997 if (endpindex > object->size) 998 endpindex = object->size; 999 1000 for (tpindex = m[count - 1]->pindex + 1; 1001 tpindex < endpindex; i++, tpindex++) { 1002 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1003 if (p == NULL) 1004 break; 1005 bp->b_pages[i] = p; 1006 } 1007 1008 bp->b_pgafter = i - bp->b_npages; 1009 bp->b_npages = i; 1010 } else 1011 bp->b_pgafter = 0; 1012 1013 if (VM_OBJECT_WOWNED(object)) 1014 VM_OBJECT_WUNLOCK(object); 1015 1016 /* Report back actual behind/ahead read. */ 1017 if (a_rbehind) 1018 *a_rbehind = bp->b_pgbefore; 1019 if (a_rahead) 1020 *a_rahead = bp->b_pgafter; 1021 1022 #ifdef INVARIANTS 1023 KASSERT(bp->b_npages <= nitems(bp->b_pages), 1024 ("%s: buf %p overflowed", __func__, bp)); 1025 for (int j = 1, prev = 0; j < bp->b_npages; j++) { 1026 if (bp->b_pages[j] == bogus_page) 1027 continue; 1028 KASSERT(bp->b_pages[j]->pindex - bp->b_pages[prev]->pindex == 1029 j - prev, ("%s: pages array not consecutive, bp %p", 1030 __func__, bp)); 1031 prev = j; 1032 } 1033 #endif 1034 1035 /* 1036 * Recalculate first offset and bytecount with regards to read behind. 1037 * Truncate bytecount to vnode real size and round up physical size 1038 * for real devices. 1039 */ 1040 foff = IDX_TO_OFF(bp->b_pages[0]->pindex); 1041 bytecount = bp->b_npages << PAGE_SHIFT; 1042 if ((foff + bytecount) > object->un_pager.vnp.vnp_size) 1043 bytecount = object->un_pager.vnp.vnp_size - foff; 1044 secmask = bo->bo_bsize - 1; 1045 KASSERT(secmask < PAGE_SIZE && secmask > 0, 1046 ("%s: sector size %d too large", __func__, secmask + 1)); 1047 bytecount = (bytecount + secmask) & ~secmask; 1048 1049 /* 1050 * And map the pages to be read into the kva, if the filesystem 1051 * requires mapped buffers. 1052 */ 1053 if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 && 1054 unmapped_buf_allowed) { 1055 bp->b_data = unmapped_buf; 1056 bp->b_offset = 0; 1057 } else { 1058 bp->b_data = bp->b_kvabase; 1059 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 1060 } 1061 1062 /* Build a minimal buffer header. */ 1063 bp->b_iocmd = BIO_READ; 1064 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 1065 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 1066 bp->b_rcred = crhold(curthread->td_ucred); 1067 bp->b_wcred = crhold(curthread->td_ucred); 1068 pbgetbo(bo, bp); 1069 bp->b_vp = vp; 1070 bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount; 1071 bp->b_iooffset = dbtob(bp->b_blkno); 1072 KASSERT(IDX_TO_OFF(m[0]->pindex - bp->b_pages[0]->pindex) == 1073 (blkno0 - bp->b_blkno) * DEV_BSIZE + 1074 IDX_TO_OFF(m[0]->pindex) % bsize, 1075 ("wrong offsets bsize %d m[0] %ju b_pages[0] %ju " 1076 "blkno0 %ju b_blkno %ju", bsize, 1077 (uintmax_t)m[0]->pindex, (uintmax_t)bp->b_pages[0]->pindex, 1078 (uintmax_t)blkno0, (uintmax_t)bp->b_blkno)); 1079 1080 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 1081 VM_CNT_INC(v_vnodein); 1082 VM_CNT_ADD(v_vnodepgsin, bp->b_npages); 1083 1084 if (iodone != NULL) { /* async */ 1085 bp->b_pgiodone = iodone; 1086 bp->b_caller1 = arg; 1087 bp->b_iodone = vnode_pager_generic_getpages_done_async; 1088 bp->b_flags |= B_ASYNC; 1089 BUF_KERNPROC(bp); 1090 bstrategy(bp); 1091 return (VM_PAGER_OK); 1092 } else { 1093 bp->b_iodone = bdone; 1094 bstrategy(bp); 1095 bwait(bp, PVM, "vnread"); 1096 error = vnode_pager_generic_getpages_done(bp); 1097 for (i = 0; i < bp->b_npages; i++) 1098 bp->b_pages[i] = NULL; 1099 bp->b_vp = NULL; 1100 pbrelbo(bp); 1101 uma_zfree(vnode_pbuf_zone, bp); 1102 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); 1103 } 1104 } 1105 1106 static void 1107 vnode_pager_generic_getpages_done_async(struct buf *bp) 1108 { 1109 int error; 1110 1111 error = vnode_pager_generic_getpages_done(bp); 1112 /* Run the iodone upon the requested range. */ 1113 bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore, 1114 bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error); 1115 for (int i = 0; i < bp->b_npages; i++) 1116 bp->b_pages[i] = NULL; 1117 bp->b_vp = NULL; 1118 pbrelbo(bp); 1119 uma_zfree(vnode_pbuf_zone, bp); 1120 } 1121 1122 static int 1123 vnode_pager_generic_getpages_done(struct buf *bp) 1124 { 1125 vm_object_t object; 1126 off_t tfoff, nextoff; 1127 int i, error; 1128 1129 error = (bp->b_ioflags & BIO_ERROR) != 0 ? EIO : 0; 1130 object = bp->b_vp->v_object; 1131 1132 if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) { 1133 if (!buf_mapped(bp)) { 1134 bp->b_data = bp->b_kvabase; 1135 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, 1136 bp->b_npages); 1137 } 1138 bzero(bp->b_data + bp->b_bcount, 1139 PAGE_SIZE * bp->b_npages - bp->b_bcount); 1140 } 1141 if (buf_mapped(bp)) { 1142 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1143 bp->b_data = unmapped_buf; 1144 } 1145 1146 VM_OBJECT_WLOCK(object); 1147 for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex); 1148 i < bp->b_npages; i++, tfoff = nextoff) { 1149 vm_page_t mt; 1150 1151 nextoff = tfoff + PAGE_SIZE; 1152 mt = bp->b_pages[i]; 1153 1154 if (nextoff <= object->un_pager.vnp.vnp_size) { 1155 /* 1156 * Read filled up entire page. 1157 */ 1158 mt->valid = VM_PAGE_BITS_ALL; 1159 KASSERT(mt->dirty == 0, 1160 ("%s: page %p is dirty", __func__, mt)); 1161 KASSERT(!pmap_page_is_mapped(mt), 1162 ("%s: page %p is mapped", __func__, mt)); 1163 } else { 1164 /* 1165 * Read did not fill up entire page. 1166 * 1167 * Currently we do not set the entire page valid, 1168 * we just try to clear the piece that we couldn't 1169 * read. 1170 */ 1171 vm_page_set_valid_range(mt, 0, 1172 object->un_pager.vnp.vnp_size - tfoff); 1173 KASSERT((mt->dirty & vm_page_bits(0, 1174 object->un_pager.vnp.vnp_size - tfoff)) == 0, 1175 ("%s: page %p is dirty", __func__, mt)); 1176 } 1177 1178 if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter) 1179 vm_page_readahead_finish(mt); 1180 } 1181 VM_OBJECT_WUNLOCK(object); 1182 if (error != 0) 1183 printf("%s: I/O read error %d\n", __func__, error); 1184 1185 return (error); 1186 } 1187 1188 /* 1189 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 1190 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 1191 * vnode_pager_generic_putpages() to implement the previous behaviour. 1192 * 1193 * All other FS's should use the bypass to get to the local media 1194 * backing vp's VOP_PUTPAGES. 1195 */ 1196 static void 1197 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1198 int flags, int *rtvals) 1199 { 1200 int rtval; 1201 struct vnode *vp; 1202 int bytes = count * PAGE_SIZE; 1203 1204 /* 1205 * Force synchronous operation if we are extremely low on memory 1206 * to prevent a low-memory deadlock. VOP operations often need to 1207 * allocate more memory to initiate the I/O ( i.e. do a BMAP 1208 * operation ). The swapper handles the case by limiting the amount 1209 * of asynchronous I/O, but that sort of solution doesn't scale well 1210 * for the vnode pager without a lot of work. 1211 * 1212 * Also, the backing vnode's iodone routine may not wake the pageout 1213 * daemon up. This should be probably be addressed XXX. 1214 */ 1215 1216 if (vm_page_count_min()) 1217 flags |= VM_PAGER_PUT_SYNC; 1218 1219 /* 1220 * Call device-specific putpages function 1221 */ 1222 vp = object->handle; 1223 VM_OBJECT_WUNLOCK(object); 1224 rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals); 1225 KASSERT(rtval != EOPNOTSUPP, 1226 ("vnode_pager: stale FS putpages\n")); 1227 VM_OBJECT_WLOCK(object); 1228 } 1229 1230 static int 1231 vn_off2bidx(vm_ooffset_t offset) 1232 { 1233 1234 return ((offset & PAGE_MASK) / DEV_BSIZE); 1235 } 1236 1237 static bool 1238 vn_dirty_blk(vm_page_t m, vm_ooffset_t offset) 1239 { 1240 1241 KASSERT(IDX_TO_OFF(m->pindex) <= offset && 1242 offset < IDX_TO_OFF(m->pindex + 1), 1243 ("page %p pidx %ju offset %ju", m, (uintmax_t)m->pindex, 1244 (uintmax_t)offset)); 1245 return ((m->dirty & ((vm_page_bits_t)1 << vn_off2bidx(offset))) != 0); 1246 } 1247 1248 /* 1249 * This is now called from local media FS's to operate against their 1250 * own vnodes if they fail to implement VOP_PUTPAGES. 1251 * 1252 * This is typically called indirectly via the pageout daemon and 1253 * clustering has already typically occurred, so in general we ask the 1254 * underlying filesystem to write the data out asynchronously rather 1255 * then delayed. 1256 */ 1257 int 1258 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount, 1259 int flags, int *rtvals) 1260 { 1261 vm_object_t object; 1262 vm_page_t m; 1263 vm_ooffset_t maxblksz, next_offset, poffset, prev_offset; 1264 struct uio auio; 1265 struct iovec aiov; 1266 off_t prev_resid, wrsz; 1267 int count, error, i, maxsize, ncount, pgoff, ppscheck; 1268 bool in_hole; 1269 static struct timeval lastfail; 1270 static int curfail; 1271 1272 object = vp->v_object; 1273 count = bytecount / PAGE_SIZE; 1274 1275 for (i = 0; i < count; i++) 1276 rtvals[i] = VM_PAGER_ERROR; 1277 1278 if ((int64_t)ma[0]->pindex < 0) { 1279 printf("vnode_pager_generic_putpages: " 1280 "attempt to write meta-data 0x%jx(%lx)\n", 1281 (uintmax_t)ma[0]->pindex, (u_long)ma[0]->dirty); 1282 rtvals[0] = VM_PAGER_BAD; 1283 return (VM_PAGER_BAD); 1284 } 1285 1286 maxsize = count * PAGE_SIZE; 1287 ncount = count; 1288 1289 poffset = IDX_TO_OFF(ma[0]->pindex); 1290 1291 /* 1292 * If the page-aligned write is larger then the actual file we 1293 * have to invalidate pages occurring beyond the file EOF. However, 1294 * there is an edge case where a file may not be page-aligned where 1295 * the last page is partially invalid. In this case the filesystem 1296 * may not properly clear the dirty bits for the entire page (which 1297 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 1298 * With the page locked we are free to fix-up the dirty bits here. 1299 * 1300 * We do not under any circumstances truncate the valid bits, as 1301 * this will screw up bogus page replacement. 1302 */ 1303 VM_OBJECT_RLOCK(object); 1304 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 1305 if (!VM_OBJECT_TRYUPGRADE(object)) { 1306 VM_OBJECT_RUNLOCK(object); 1307 VM_OBJECT_WLOCK(object); 1308 if (maxsize + poffset <= object->un_pager.vnp.vnp_size) 1309 goto downgrade; 1310 } 1311 if (object->un_pager.vnp.vnp_size > poffset) { 1312 maxsize = object->un_pager.vnp.vnp_size - poffset; 1313 ncount = btoc(maxsize); 1314 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 1315 pgoff = roundup2(pgoff, DEV_BSIZE); 1316 1317 /* 1318 * If the object is locked and the following 1319 * conditions hold, then the page's dirty 1320 * field cannot be concurrently changed by a 1321 * pmap operation. 1322 */ 1323 m = ma[ncount - 1]; 1324 vm_page_assert_sbusied(m); 1325 KASSERT(!pmap_page_is_write_mapped(m), 1326 ("vnode_pager_generic_putpages: page %p is not read-only", m)); 1327 MPASS(m->dirty != 0); 1328 vm_page_clear_dirty(m, pgoff, PAGE_SIZE - 1329 pgoff); 1330 } 1331 } else { 1332 maxsize = 0; 1333 ncount = 0; 1334 } 1335 for (i = ncount; i < count; i++) 1336 rtvals[i] = VM_PAGER_BAD; 1337 downgrade: 1338 VM_OBJECT_LOCK_DOWNGRADE(object); 1339 } 1340 1341 auio.uio_iov = &aiov; 1342 auio.uio_segflg = UIO_NOCOPY; 1343 auio.uio_rw = UIO_WRITE; 1344 auio.uio_td = NULL; 1345 maxblksz = roundup2(poffset + maxsize, DEV_BSIZE); 1346 1347 for (prev_offset = poffset; prev_offset < maxblksz;) { 1348 /* Skip clean blocks. */ 1349 for (in_hole = true; in_hole && prev_offset < maxblksz;) { 1350 m = ma[OFF_TO_IDX(prev_offset - poffset)]; 1351 for (i = vn_off2bidx(prev_offset); 1352 i < sizeof(vm_page_bits_t) * NBBY && 1353 prev_offset < maxblksz; i++) { 1354 if (vn_dirty_blk(m, prev_offset)) { 1355 in_hole = false; 1356 break; 1357 } 1358 prev_offset += DEV_BSIZE; 1359 } 1360 } 1361 if (in_hole) 1362 goto write_done; 1363 1364 /* Find longest run of dirty blocks. */ 1365 for (next_offset = prev_offset; next_offset < maxblksz;) { 1366 m = ma[OFF_TO_IDX(next_offset - poffset)]; 1367 for (i = vn_off2bidx(next_offset); 1368 i < sizeof(vm_page_bits_t) * NBBY && 1369 next_offset < maxblksz; i++) { 1370 if (!vn_dirty_blk(m, next_offset)) 1371 goto start_write; 1372 next_offset += DEV_BSIZE; 1373 } 1374 } 1375 start_write: 1376 if (next_offset > poffset + maxsize) 1377 next_offset = poffset + maxsize; 1378 1379 /* 1380 * Getting here requires finding a dirty block in the 1381 * 'skip clean blocks' loop. 1382 */ 1383 MPASS(prev_offset < next_offset); 1384 1385 VM_OBJECT_RUNLOCK(object); 1386 aiov.iov_base = NULL; 1387 auio.uio_iovcnt = 1; 1388 auio.uio_offset = prev_offset; 1389 prev_resid = auio.uio_resid = aiov.iov_len = next_offset - 1390 prev_offset; 1391 error = VOP_WRITE(vp, &auio, 1392 vnode_pager_putpages_ioflags(flags), curthread->td_ucred); 1393 1394 wrsz = prev_resid - auio.uio_resid; 1395 if (wrsz == 0) { 1396 if (ppsratecheck(&lastfail, &curfail, 1) != 0) { 1397 vn_printf(vp, "vnode_pager_putpages: " 1398 "zero-length write at %ju resid %zd\n", 1399 auio.uio_offset, auio.uio_resid); 1400 } 1401 VM_OBJECT_RLOCK(object); 1402 break; 1403 } 1404 1405 /* Adjust the starting offset for next iteration. */ 1406 prev_offset += wrsz; 1407 MPASS(auio.uio_offset == prev_offset); 1408 1409 ppscheck = 0; 1410 if (error != 0 && (ppscheck = ppsratecheck(&lastfail, 1411 &curfail, 1)) != 0) 1412 vn_printf(vp, "vnode_pager_putpages: I/O error %d\n", 1413 error); 1414 if (auio.uio_resid != 0 && (ppscheck != 0 || 1415 ppsratecheck(&lastfail, &curfail, 1) != 0)) 1416 vn_printf(vp, "vnode_pager_putpages: residual I/O %zd " 1417 "at %ju\n", auio.uio_resid, 1418 (uintmax_t)ma[0]->pindex); 1419 VM_OBJECT_RLOCK(object); 1420 if (error != 0 || auio.uio_resid != 0) 1421 break; 1422 } 1423 write_done: 1424 /* Mark completely processed pages. */ 1425 for (i = 0; i < OFF_TO_IDX(prev_offset - poffset); i++) 1426 rtvals[i] = VM_PAGER_OK; 1427 /* Mark partial EOF page. */ 1428 if (prev_offset == poffset + maxsize && (prev_offset & PAGE_MASK) != 0) 1429 rtvals[i++] = VM_PAGER_OK; 1430 /* Unwritten pages in range, free bonus if the page is clean. */ 1431 for (; i < ncount; i++) 1432 rtvals[i] = ma[i]->dirty == 0 ? VM_PAGER_OK : VM_PAGER_ERROR; 1433 VM_OBJECT_RUNLOCK(object); 1434 VM_CNT_ADD(v_vnodepgsout, i); 1435 VM_CNT_INC(v_vnodeout); 1436 return (rtvals[0]); 1437 } 1438 1439 int 1440 vnode_pager_putpages_ioflags(int pager_flags) 1441 { 1442 int ioflags; 1443 1444 /* 1445 * Pageouts are already clustered, use IO_ASYNC to force a 1446 * bawrite() rather then a bdwrite() to prevent paging I/O 1447 * from saturating the buffer cache. Dummy-up the sequential 1448 * heuristic to cause large ranges to cluster. If neither 1449 * IO_SYNC or IO_ASYNC is set, the system decides how to 1450 * cluster. 1451 */ 1452 ioflags = IO_VMIO; 1453 if ((pager_flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) != 0) 1454 ioflags |= IO_SYNC; 1455 else if ((pager_flags & VM_PAGER_CLUSTER_OK) == 0) 1456 ioflags |= IO_ASYNC; 1457 ioflags |= (pager_flags & VM_PAGER_PUT_INVAL) != 0 ? IO_INVAL: 0; 1458 ioflags |= (pager_flags & VM_PAGER_PUT_NOREUSE) != 0 ? IO_NOREUSE : 0; 1459 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 1460 return (ioflags); 1461 } 1462 1463 /* 1464 * vnode_pager_undirty_pages(). 1465 * 1466 * A helper to mark pages as clean after pageout that was possibly 1467 * done with a short write. The lpos argument specifies the page run 1468 * length in bytes, and the written argument specifies how many bytes 1469 * were actually written. eof is the offset past the last valid byte 1470 * in the vnode using the absolute file position of the first byte in 1471 * the run as the base from which it is computed. 1472 */ 1473 void 1474 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written, off_t eof, 1475 int lpos) 1476 { 1477 vm_object_t obj; 1478 int i, pos, pos_devb; 1479 1480 if (written == 0 && eof >= lpos) 1481 return; 1482 obj = ma[0]->object; 1483 VM_OBJECT_WLOCK(obj); 1484 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) { 1485 if (pos < trunc_page(written)) { 1486 rtvals[i] = VM_PAGER_OK; 1487 vm_page_undirty(ma[i]); 1488 } else { 1489 /* Partially written page. */ 1490 rtvals[i] = VM_PAGER_AGAIN; 1491 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK); 1492 } 1493 } 1494 if (eof >= lpos) /* avoid truncation */ 1495 goto done; 1496 for (pos = eof, i = OFF_TO_IDX(trunc_page(pos)); pos < lpos; i++) { 1497 if (pos != trunc_page(pos)) { 1498 /* 1499 * The page contains the last valid byte in 1500 * the vnode, mark the rest of the page as 1501 * clean, potentially making the whole page 1502 * clean. 1503 */ 1504 pos_devb = roundup2(pos & PAGE_MASK, DEV_BSIZE); 1505 vm_page_clear_dirty(ma[i], pos_devb, PAGE_SIZE - 1506 pos_devb); 1507 1508 /* 1509 * If the page was cleaned, report the pageout 1510 * on it as successful. msync() no longer 1511 * needs to write out the page, endlessly 1512 * creating write requests and dirty buffers. 1513 */ 1514 if (ma[i]->dirty == 0) 1515 rtvals[i] = VM_PAGER_OK; 1516 1517 pos = round_page(pos); 1518 } else { 1519 /* vm_pageout_flush() clears dirty */ 1520 rtvals[i] = VM_PAGER_BAD; 1521 pos += PAGE_SIZE; 1522 } 1523 } 1524 done: 1525 VM_OBJECT_WUNLOCK(obj); 1526 } 1527 1528 void 1529 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start, 1530 vm_offset_t end) 1531 { 1532 struct vnode *vp; 1533 vm_ooffset_t old_wm; 1534 1535 VM_OBJECT_WLOCK(object); 1536 if (object->type != OBJT_VNODE) { 1537 VM_OBJECT_WUNLOCK(object); 1538 return; 1539 } 1540 old_wm = object->un_pager.vnp.writemappings; 1541 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start; 1542 vp = object->handle; 1543 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) { 1544 ASSERT_VOP_LOCKED(vp, "v_writecount inc"); 1545 VOP_ADD_WRITECOUNT_CHECKED(vp, 1); 1546 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", 1547 __func__, vp, vp->v_writecount); 1548 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) { 1549 ASSERT_VOP_LOCKED(vp, "v_writecount dec"); 1550 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 1551 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 1552 __func__, vp, vp->v_writecount); 1553 } 1554 VM_OBJECT_WUNLOCK(object); 1555 } 1556 1557 void 1558 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start, 1559 vm_offset_t end) 1560 { 1561 struct vnode *vp; 1562 struct mount *mp; 1563 vm_offset_t inc; 1564 1565 VM_OBJECT_WLOCK(object); 1566 1567 /* 1568 * First, recheck the object type to account for the race when 1569 * the vnode is reclaimed. 1570 */ 1571 if (object->type != OBJT_VNODE) { 1572 VM_OBJECT_WUNLOCK(object); 1573 return; 1574 } 1575 1576 /* 1577 * Optimize for the case when writemappings is not going to 1578 * zero. 1579 */ 1580 inc = end - start; 1581 if (object->un_pager.vnp.writemappings != inc) { 1582 object->un_pager.vnp.writemappings -= inc; 1583 VM_OBJECT_WUNLOCK(object); 1584 return; 1585 } 1586 1587 vp = object->handle; 1588 vhold(vp); 1589 VM_OBJECT_WUNLOCK(object); 1590 mp = NULL; 1591 vn_start_write(vp, &mp, V_WAIT); 1592 vn_lock(vp, LK_SHARED | LK_RETRY); 1593 1594 /* 1595 * Decrement the object's writemappings, by swapping the start 1596 * and end arguments for vnode_pager_update_writecount(). If 1597 * there was not a race with vnode reclaimation, then the 1598 * vnode's v_writecount is decremented. 1599 */ 1600 vnode_pager_update_writecount(object, end, start); 1601 VOP_UNLOCK(vp, 0); 1602 vdrop(vp); 1603 if (mp != NULL) 1604 vn_finished_write(mp); 1605 } 1606