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_terminate(obj); 204 } else { 205 /* 206 * Waiters were already handled during object 207 * termination. The exclusive vnode lock hopefully 208 * prevented new waiters from referencing the dying 209 * object. 210 */ 211 KASSERT((obj->flags & OBJ_DISCONNECTWNT) == 0, 212 ("OBJ_DISCONNECTWNT set obj %p flags %x", 213 obj, obj->flags)); 214 vp->v_object = NULL; 215 VM_OBJECT_WUNLOCK(obj); 216 } 217 } else { 218 /* 219 * Woe to the process that tries to page now :-). 220 */ 221 vm_pager_deallocate(obj); 222 VM_OBJECT_WUNLOCK(obj); 223 } 224 KASSERT(vp->v_object == NULL, ("vp %p obj %p", vp, vp->v_object)); 225 } 226 227 228 /* 229 * Allocate (or lookup) pager for a vnode. 230 * Handle is a vnode pointer. 231 * 232 * MPSAFE 233 */ 234 vm_object_t 235 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 236 vm_ooffset_t offset, struct ucred *cred) 237 { 238 vm_object_t object; 239 struct vnode *vp; 240 241 /* 242 * Pageout to vnode, no can do yet. 243 */ 244 if (handle == NULL) 245 return (NULL); 246 247 vp = (struct vnode *) handle; 248 249 /* 250 * If the object is being terminated, wait for it to 251 * go away. 252 */ 253 retry: 254 while ((object = vp->v_object) != NULL) { 255 VM_OBJECT_WLOCK(object); 256 if ((object->flags & OBJ_DEAD) == 0) 257 break; 258 vm_object_set_flag(object, OBJ_DISCONNECTWNT); 259 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vadead", 0); 260 } 261 262 KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference")); 263 264 if (object == NULL) { 265 /* 266 * Add an object of the appropriate size 267 */ 268 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); 269 270 object->un_pager.vnp.vnp_size = size; 271 object->un_pager.vnp.writemappings = 0; 272 object->domain.dr_policy = vnode_domainset; 273 274 object->handle = handle; 275 VI_LOCK(vp); 276 if (vp->v_object != NULL) { 277 /* 278 * Object has been created while we were sleeping 279 */ 280 VI_UNLOCK(vp); 281 VM_OBJECT_WLOCK(object); 282 KASSERT(object->ref_count == 1, 283 ("leaked ref %p %d", object, object->ref_count)); 284 object->type = OBJT_DEAD; 285 object->ref_count = 0; 286 VM_OBJECT_WUNLOCK(object); 287 vm_object_destroy(object); 288 goto retry; 289 } 290 vp->v_object = object; 291 VI_UNLOCK(vp); 292 } else { 293 object->ref_count++; 294 #if VM_NRESERVLEVEL > 0 295 vm_object_color(object, 0); 296 #endif 297 VM_OBJECT_WUNLOCK(object); 298 } 299 vrefact(vp); 300 return (object); 301 } 302 303 /* 304 * The object must be locked. 305 */ 306 static void 307 vnode_pager_dealloc(vm_object_t object) 308 { 309 struct vnode *vp; 310 int refs; 311 312 vp = object->handle; 313 if (vp == NULL) 314 panic("vnode_pager_dealloc: pager already dealloced"); 315 316 VM_OBJECT_ASSERT_WLOCKED(object); 317 vm_object_pip_wait(object, "vnpdea"); 318 refs = object->ref_count; 319 320 object->handle = NULL; 321 object->type = OBJT_DEAD; 322 if (object->flags & OBJ_DISCONNECTWNT) { 323 vm_object_clear_flag(object, OBJ_DISCONNECTWNT); 324 wakeup(object); 325 } 326 ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc"); 327 if (object->un_pager.vnp.writemappings > 0) { 328 object->un_pager.vnp.writemappings = 0; 329 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 330 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 331 __func__, vp, vp->v_writecount); 332 } 333 vp->v_object = NULL; 334 VI_LOCK(vp); 335 336 /* 337 * vm_map_entry_set_vnode_text() cannot reach this vnode by 338 * following object->handle. Clear all text references now. 339 * This also clears the transient references from 340 * kern_execve(), which is fine because dead_vnodeops uses nop 341 * for VOP_UNSET_TEXT(). 342 */ 343 if (vp->v_writecount < 0) 344 vp->v_writecount = 0; 345 VI_UNLOCK(vp); 346 VM_OBJECT_WUNLOCK(object); 347 while (refs-- > 0) 348 vunref(vp); 349 VM_OBJECT_WLOCK(object); 350 } 351 352 static boolean_t 353 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 354 int *after) 355 { 356 struct vnode *vp = object->handle; 357 daddr_t bn; 358 uintptr_t lockstate; 359 int err; 360 daddr_t reqblock; 361 int poff; 362 int bsize; 363 int pagesperblock, blocksperpage; 364 365 VM_OBJECT_ASSERT_LOCKED(object); 366 /* 367 * If no vp or vp is doomed or marked transparent to VM, we do not 368 * have the page. 369 */ 370 if (vp == NULL || vp->v_iflag & VI_DOOMED) 371 return FALSE; 372 /* 373 * If the offset is beyond end of file we do 374 * not have the page. 375 */ 376 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size) 377 return FALSE; 378 379 bsize = vp->v_mount->mnt_stat.f_iosize; 380 pagesperblock = bsize / PAGE_SIZE; 381 blocksperpage = 0; 382 if (pagesperblock > 0) { 383 reqblock = pindex / pagesperblock; 384 } else { 385 blocksperpage = (PAGE_SIZE / bsize); 386 reqblock = pindex * blocksperpage; 387 } 388 lockstate = VM_OBJECT_DROP(object); 389 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before); 390 VM_OBJECT_PICKUP(object, lockstate); 391 if (err) 392 return TRUE; 393 if (bn == -1) 394 return FALSE; 395 if (pagesperblock > 0) { 396 poff = pindex - (reqblock * pagesperblock); 397 if (before) { 398 *before *= pagesperblock; 399 *before += poff; 400 } 401 if (after) { 402 /* 403 * The BMAP vop can report a partial block in the 404 * 'after', but must not report blocks after EOF. 405 * Assert the latter, and truncate 'after' in case 406 * of the former. 407 */ 408 KASSERT((reqblock + *after) * pagesperblock < 409 roundup2(object->size, pagesperblock), 410 ("%s: reqblock %jd after %d size %ju", __func__, 411 (intmax_t )reqblock, *after, 412 (uintmax_t )object->size)); 413 *after *= pagesperblock; 414 *after += pagesperblock - (poff + 1); 415 if (pindex + *after >= object->size) 416 *after = object->size - 1 - pindex; 417 } 418 } else { 419 if (before) { 420 *before /= blocksperpage; 421 } 422 423 if (after) { 424 *after /= blocksperpage; 425 } 426 } 427 return TRUE; 428 } 429 430 /* 431 * Lets the VM system know about a change in size for a file. 432 * We adjust our own internal size and flush any cached pages in 433 * the associated object that are affected by the size change. 434 * 435 * Note: this routine may be invoked as a result of a pager put 436 * operation (possibly at object termination time), so we must be careful. 437 */ 438 void 439 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize) 440 { 441 vm_object_t object; 442 vm_page_t m; 443 vm_pindex_t nobjsize; 444 445 if ((object = vp->v_object) == NULL) 446 return; 447 /* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */ 448 VM_OBJECT_WLOCK(object); 449 if (object->type == OBJT_DEAD) { 450 VM_OBJECT_WUNLOCK(object); 451 return; 452 } 453 KASSERT(object->type == OBJT_VNODE, 454 ("not vnode-backed object %p", object)); 455 if (nsize == object->un_pager.vnp.vnp_size) { 456 /* 457 * Hasn't changed size 458 */ 459 VM_OBJECT_WUNLOCK(object); 460 return; 461 } 462 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 463 if (nsize < object->un_pager.vnp.vnp_size) { 464 /* 465 * File has shrunk. Toss any cached pages beyond the new EOF. 466 */ 467 if (nobjsize < object->size) 468 vm_object_page_remove(object, nobjsize, object->size, 469 0); 470 /* 471 * this gets rid of garbage at the end of a page that is now 472 * only partially backed by the vnode. 473 * 474 * XXX for some reason (I don't know yet), if we take a 475 * completely invalid page and mark it partially valid 476 * it can screw up NFS reads, so we don't allow the case. 477 */ 478 if ((nsize & PAGE_MASK) && 479 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL && 480 m->valid != 0) { 481 int base = (int)nsize & PAGE_MASK; 482 int size = PAGE_SIZE - base; 483 484 /* 485 * Clear out partial-page garbage in case 486 * the page has been mapped. 487 */ 488 pmap_zero_page_area(m, base, size); 489 490 /* 491 * Update the valid bits to reflect the blocks that 492 * have been zeroed. Some of these valid bits may 493 * have already been set. 494 */ 495 vm_page_set_valid_range(m, base, size); 496 497 /* 498 * Round "base" to the next block boundary so that the 499 * dirty bit for a partially zeroed block is not 500 * cleared. 501 */ 502 base = roundup2(base, DEV_BSIZE); 503 504 /* 505 * Clear out partial-page dirty bits. 506 * 507 * note that we do not clear out the valid 508 * bits. This would prevent bogus_page 509 * replacement from working properly. 510 */ 511 vm_page_clear_dirty(m, base, PAGE_SIZE - base); 512 } 513 } 514 object->un_pager.vnp.vnp_size = nsize; 515 object->size = nobjsize; 516 VM_OBJECT_WUNLOCK(object); 517 } 518 519 /* 520 * calculate the linear (byte) disk address of specified virtual 521 * file address 522 */ 523 static int 524 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress, 525 int *run) 526 { 527 int bsize; 528 int err; 529 daddr_t vblock; 530 daddr_t voffset; 531 532 if (address < 0) 533 return -1; 534 535 if (vp->v_iflag & VI_DOOMED) 536 return -1; 537 538 bsize = vp->v_mount->mnt_stat.f_iosize; 539 vblock = address / bsize; 540 voffset = address % bsize; 541 542 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL); 543 if (err == 0) { 544 if (*rtaddress != -1) 545 *rtaddress += voffset / DEV_BSIZE; 546 if (run) { 547 *run += 1; 548 *run *= bsize / PAGE_SIZE; 549 *run -= voffset / PAGE_SIZE; 550 } 551 } 552 553 return (err); 554 } 555 556 /* 557 * small block filesystem vnode pager input 558 */ 559 static int 560 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m) 561 { 562 struct vnode *vp; 563 struct bufobj *bo; 564 struct buf *bp; 565 struct sf_buf *sf; 566 daddr_t fileaddr; 567 vm_offset_t bsize; 568 vm_page_bits_t bits; 569 int error, i; 570 571 error = 0; 572 vp = object->handle; 573 if (vp->v_iflag & VI_DOOMED) 574 return VM_PAGER_BAD; 575 576 bsize = vp->v_mount->mnt_stat.f_iosize; 577 578 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL); 579 580 sf = sf_buf_alloc(m, 0); 581 582 for (i = 0; i < PAGE_SIZE / bsize; i++) { 583 vm_ooffset_t address; 584 585 bits = vm_page_bits(i * bsize, bsize); 586 if (m->valid & bits) 587 continue; 588 589 address = IDX_TO_OFF(m->pindex) + i * bsize; 590 if (address >= object->un_pager.vnp.vnp_size) { 591 fileaddr = -1; 592 } else { 593 error = vnode_pager_addr(vp, address, &fileaddr, NULL); 594 if (error) 595 break; 596 } 597 if (fileaddr != -1) { 598 bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK); 599 600 /* build a minimal buffer header */ 601 bp->b_iocmd = BIO_READ; 602 bp->b_iodone = bdone; 603 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 604 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 605 bp->b_rcred = crhold(curthread->td_ucred); 606 bp->b_wcred = crhold(curthread->td_ucred); 607 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize; 608 bp->b_blkno = fileaddr; 609 pbgetbo(bo, bp); 610 bp->b_vp = vp; 611 bp->b_bcount = bsize; 612 bp->b_bufsize = bsize; 613 bp->b_runningbufspace = bp->b_bufsize; 614 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 615 616 /* do the input */ 617 bp->b_iooffset = dbtob(bp->b_blkno); 618 bstrategy(bp); 619 620 bwait(bp, PVM, "vnsrd"); 621 622 if ((bp->b_ioflags & BIO_ERROR) != 0) 623 error = EIO; 624 625 /* 626 * free the buffer header back to the swap buffer pool 627 */ 628 bp->b_vp = NULL; 629 pbrelbo(bp); 630 uma_zfree(vnode_pbuf_zone, bp); 631 if (error) 632 break; 633 } else 634 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize); 635 KASSERT((m->dirty & bits) == 0, 636 ("vnode_pager_input_smlfs: page %p is dirty", m)); 637 VM_OBJECT_WLOCK(object); 638 m->valid |= bits; 639 VM_OBJECT_WUNLOCK(object); 640 } 641 sf_buf_free(sf); 642 if (error) { 643 return VM_PAGER_ERROR; 644 } 645 return VM_PAGER_OK; 646 } 647 648 /* 649 * old style vnode pager input routine 650 */ 651 static int 652 vnode_pager_input_old(vm_object_t object, vm_page_t m) 653 { 654 struct uio auio; 655 struct iovec aiov; 656 int error; 657 int size; 658 struct sf_buf *sf; 659 struct vnode *vp; 660 661 VM_OBJECT_ASSERT_WLOCKED(object); 662 error = 0; 663 664 /* 665 * Return failure if beyond current EOF 666 */ 667 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 668 return VM_PAGER_BAD; 669 } else { 670 size = PAGE_SIZE; 671 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 672 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 673 vp = object->handle; 674 VM_OBJECT_WUNLOCK(object); 675 676 /* 677 * Allocate a kernel virtual address and initialize so that 678 * we can use VOP_READ/WRITE routines. 679 */ 680 sf = sf_buf_alloc(m, 0); 681 682 aiov.iov_base = (caddr_t)sf_buf_kva(sf); 683 aiov.iov_len = size; 684 auio.uio_iov = &aiov; 685 auio.uio_iovcnt = 1; 686 auio.uio_offset = IDX_TO_OFF(m->pindex); 687 auio.uio_segflg = UIO_SYSSPACE; 688 auio.uio_rw = UIO_READ; 689 auio.uio_resid = size; 690 auio.uio_td = curthread; 691 692 error = VOP_READ(vp, &auio, 0, curthread->td_ucred); 693 if (!error) { 694 int count = size - auio.uio_resid; 695 696 if (count == 0) 697 error = EINVAL; 698 else if (count != PAGE_SIZE) 699 bzero((caddr_t)sf_buf_kva(sf) + count, 700 PAGE_SIZE - count); 701 } 702 sf_buf_free(sf); 703 704 VM_OBJECT_WLOCK(object); 705 } 706 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m)); 707 if (!error) 708 m->valid = VM_PAGE_BITS_ALL; 709 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 710 } 711 712 /* 713 * generic vnode pager input routine 714 */ 715 716 /* 717 * Local media VFS's that do not implement their own VOP_GETPAGES 718 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages() 719 * to implement the previous behaviour. 720 * 721 * All other FS's should use the bypass to get to the local media 722 * backing vp's VOP_GETPAGES. 723 */ 724 static int 725 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind, 726 int *rahead) 727 { 728 struct vnode *vp; 729 int rtval; 730 731 vp = object->handle; 732 VM_OBJECT_WUNLOCK(object); 733 rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead); 734 KASSERT(rtval != EOPNOTSUPP, 735 ("vnode_pager: FS getpages not implemented\n")); 736 VM_OBJECT_WLOCK(object); 737 return rtval; 738 } 739 740 static int 741 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count, 742 int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg) 743 { 744 struct vnode *vp; 745 int rtval; 746 747 vp = object->handle; 748 VM_OBJECT_WUNLOCK(object); 749 rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg); 750 KASSERT(rtval != EOPNOTSUPP, 751 ("vnode_pager: FS getpages_async not implemented\n")); 752 VM_OBJECT_WLOCK(object); 753 return (rtval); 754 } 755 756 /* 757 * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for 758 * local filesystems, where partially valid pages can only occur at 759 * the end of file. 760 */ 761 int 762 vnode_pager_local_getpages(struct vop_getpages_args *ap) 763 { 764 765 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count, 766 ap->a_rbehind, ap->a_rahead, NULL, NULL)); 767 } 768 769 int 770 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap) 771 { 772 773 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count, 774 ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg)); 775 } 776 777 /* 778 * This is now called from local media FS's to operate against their 779 * own vnodes if they fail to implement VOP_GETPAGES. 780 */ 781 int 782 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count, 783 int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg) 784 { 785 vm_object_t object; 786 struct bufobj *bo; 787 struct buf *bp; 788 off_t foff; 789 #ifdef INVARIANTS 790 off_t blkno0; 791 #endif 792 int bsize, pagesperblock; 793 int error, before, after, rbehind, rahead, poff, i; 794 int bytecount, secmask; 795 796 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 797 ("%s does not support devices", __func__)); 798 799 if (vp->v_iflag & VI_DOOMED) 800 return (VM_PAGER_BAD); 801 802 object = vp->v_object; 803 foff = IDX_TO_OFF(m[0]->pindex); 804 bsize = vp->v_mount->mnt_stat.f_iosize; 805 pagesperblock = bsize / PAGE_SIZE; 806 807 KASSERT(foff < object->un_pager.vnp.vnp_size, 808 ("%s: page %p offset beyond vp %p size", __func__, m[0], vp)); 809 KASSERT(count <= nitems(bp->b_pages), 810 ("%s: requested %d pages", __func__, count)); 811 812 /* 813 * The last page has valid blocks. Invalid part can only 814 * exist at the end of file, and the page is made fully valid 815 * by zeroing in vm_pager_get_pages(). 816 */ 817 if (m[count - 1]->valid != 0 && --count == 0) { 818 if (iodone != NULL) 819 iodone(arg, m, 1, 0); 820 return (VM_PAGER_OK); 821 } 822 823 bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK); 824 825 /* 826 * Get the underlying device blocks for the file with VOP_BMAP(). 827 * If the file system doesn't support VOP_BMAP, use old way of 828 * getting pages via VOP_READ. 829 */ 830 error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before); 831 if (error == EOPNOTSUPP) { 832 uma_zfree(vnode_pbuf_zone, bp); 833 VM_OBJECT_WLOCK(object); 834 for (i = 0; i < count; i++) { 835 VM_CNT_INC(v_vnodein); 836 VM_CNT_INC(v_vnodepgsin); 837 error = vnode_pager_input_old(object, m[i]); 838 if (error) 839 break; 840 } 841 VM_OBJECT_WUNLOCK(object); 842 return (error); 843 } else if (error != 0) { 844 uma_zfree(vnode_pbuf_zone, bp); 845 return (VM_PAGER_ERROR); 846 } 847 848 /* 849 * If the file system supports BMAP, but blocksize is smaller 850 * than a page size, then use special small filesystem code. 851 */ 852 if (pagesperblock == 0) { 853 uma_zfree(vnode_pbuf_zone, bp); 854 for (i = 0; i < count; i++) { 855 VM_CNT_INC(v_vnodein); 856 VM_CNT_INC(v_vnodepgsin); 857 error = vnode_pager_input_smlfs(object, m[i]); 858 if (error) 859 break; 860 } 861 return (error); 862 } 863 864 /* 865 * A sparse file can be encountered only for a single page request, 866 * which may not be preceded by call to vm_pager_haspage(). 867 */ 868 if (bp->b_blkno == -1) { 869 KASSERT(count == 1, 870 ("%s: array[%d] request to a sparse file %p", __func__, 871 count, vp)); 872 uma_zfree(vnode_pbuf_zone, bp); 873 pmap_zero_page(m[0]); 874 KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty", 875 __func__, m[0])); 876 VM_OBJECT_WLOCK(object); 877 m[0]->valid = VM_PAGE_BITS_ALL; 878 VM_OBJECT_WUNLOCK(object); 879 return (VM_PAGER_OK); 880 } 881 882 #ifdef INVARIANTS 883 blkno0 = bp->b_blkno; 884 #endif 885 bp->b_blkno += (foff % bsize) / DEV_BSIZE; 886 887 /* Recalculate blocks available after/before to pages. */ 888 poff = (foff % bsize) / PAGE_SIZE; 889 before *= pagesperblock; 890 before += poff; 891 after *= pagesperblock; 892 after += pagesperblock - (poff + 1); 893 if (m[0]->pindex + after >= object->size) 894 after = object->size - 1 - m[0]->pindex; 895 KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d", 896 __func__, count, after + 1)); 897 after -= count - 1; 898 899 /* Trim requested rbehind/rahead to possible values. */ 900 rbehind = a_rbehind ? *a_rbehind : 0; 901 rahead = a_rahead ? *a_rahead : 0; 902 rbehind = min(rbehind, before); 903 rbehind = min(rbehind, m[0]->pindex); 904 rahead = min(rahead, after); 905 rahead = min(rahead, object->size - m[count - 1]->pindex); 906 /* 907 * Check that total amount of pages fit into buf. Trim rbehind and 908 * rahead evenly if not. 909 */ 910 if (rbehind + rahead + count > nitems(bp->b_pages)) { 911 int trim, sum; 912 913 trim = rbehind + rahead + count - nitems(bp->b_pages) + 1; 914 sum = rbehind + rahead; 915 if (rbehind == before) { 916 /* Roundup rbehind trim to block size. */ 917 rbehind -= roundup(trim * rbehind / sum, pagesperblock); 918 if (rbehind < 0) 919 rbehind = 0; 920 } else 921 rbehind -= trim * rbehind / sum; 922 rahead -= trim * rahead / sum; 923 } 924 KASSERT(rbehind + rahead + count <= nitems(bp->b_pages), 925 ("%s: behind %d ahead %d count %d", __func__, 926 rbehind, rahead, count)); 927 928 /* 929 * Fill in the bp->b_pages[] array with requested and optional 930 * read behind or read ahead pages. Read behind pages are looked 931 * up in a backward direction, down to a first cached page. Same 932 * for read ahead pages, but there is no need to shift the array 933 * in case of encountering a cached page. 934 */ 935 i = bp->b_npages = 0; 936 if (rbehind) { 937 vm_pindex_t startpindex, tpindex; 938 vm_page_t p; 939 940 VM_OBJECT_WLOCK(object); 941 startpindex = m[0]->pindex - rbehind; 942 if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL && 943 p->pindex >= startpindex) 944 startpindex = p->pindex + 1; 945 946 /* tpindex is unsigned; beware of numeric underflow. */ 947 for (tpindex = m[0]->pindex - 1; 948 tpindex >= startpindex && tpindex < m[0]->pindex; 949 tpindex--, i++) { 950 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 951 if (p == NULL) { 952 /* Shift the array. */ 953 for (int j = 0; j < i; j++) 954 bp->b_pages[j] = bp->b_pages[j + 955 tpindex + 1 - startpindex]; 956 break; 957 } 958 bp->b_pages[tpindex - startpindex] = p; 959 } 960 961 bp->b_pgbefore = i; 962 bp->b_npages += i; 963 bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE; 964 } else 965 bp->b_pgbefore = 0; 966 967 /* Requested pages. */ 968 for (int j = 0; j < count; j++, i++) 969 bp->b_pages[i] = m[j]; 970 bp->b_npages += count; 971 972 if (rahead) { 973 vm_pindex_t endpindex, tpindex; 974 vm_page_t p; 975 976 if (!VM_OBJECT_WOWNED(object)) 977 VM_OBJECT_WLOCK(object); 978 endpindex = m[count - 1]->pindex + rahead + 1; 979 if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL && 980 p->pindex < endpindex) 981 endpindex = p->pindex; 982 if (endpindex > object->size) 983 endpindex = object->size; 984 985 for (tpindex = m[count - 1]->pindex + 1; 986 tpindex < endpindex; i++, tpindex++) { 987 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 988 if (p == NULL) 989 break; 990 bp->b_pages[i] = p; 991 } 992 993 bp->b_pgafter = i - bp->b_npages; 994 bp->b_npages = i; 995 } else 996 bp->b_pgafter = 0; 997 998 if (VM_OBJECT_WOWNED(object)) 999 VM_OBJECT_WUNLOCK(object); 1000 1001 /* Report back actual behind/ahead read. */ 1002 if (a_rbehind) 1003 *a_rbehind = bp->b_pgbefore; 1004 if (a_rahead) 1005 *a_rahead = bp->b_pgafter; 1006 1007 #ifdef INVARIANTS 1008 KASSERT(bp->b_npages <= nitems(bp->b_pages), 1009 ("%s: buf %p overflowed", __func__, bp)); 1010 for (int j = 1, prev = 0; j < bp->b_npages; j++) { 1011 if (bp->b_pages[j] == bogus_page) 1012 continue; 1013 KASSERT(bp->b_pages[j]->pindex - bp->b_pages[prev]->pindex == 1014 j - prev, ("%s: pages array not consecutive, bp %p", 1015 __func__, bp)); 1016 prev = j; 1017 } 1018 #endif 1019 1020 /* 1021 * Recalculate first offset and bytecount with regards to read behind. 1022 * Truncate bytecount to vnode real size and round up physical size 1023 * for real devices. 1024 */ 1025 foff = IDX_TO_OFF(bp->b_pages[0]->pindex); 1026 bytecount = bp->b_npages << PAGE_SHIFT; 1027 if ((foff + bytecount) > object->un_pager.vnp.vnp_size) 1028 bytecount = object->un_pager.vnp.vnp_size - foff; 1029 secmask = bo->bo_bsize - 1; 1030 KASSERT(secmask < PAGE_SIZE && secmask > 0, 1031 ("%s: sector size %d too large", __func__, secmask + 1)); 1032 bytecount = (bytecount + secmask) & ~secmask; 1033 1034 /* 1035 * And map the pages to be read into the kva, if the filesystem 1036 * requires mapped buffers. 1037 */ 1038 if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 && 1039 unmapped_buf_allowed) { 1040 bp->b_data = unmapped_buf; 1041 bp->b_offset = 0; 1042 } else { 1043 bp->b_data = bp->b_kvabase; 1044 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 1045 } 1046 1047 /* Build a minimal buffer header. */ 1048 bp->b_iocmd = BIO_READ; 1049 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 1050 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 1051 bp->b_rcred = crhold(curthread->td_ucred); 1052 bp->b_wcred = crhold(curthread->td_ucred); 1053 pbgetbo(bo, bp); 1054 bp->b_vp = vp; 1055 bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount; 1056 bp->b_iooffset = dbtob(bp->b_blkno); 1057 KASSERT(IDX_TO_OFF(m[0]->pindex - bp->b_pages[0]->pindex) == 1058 (blkno0 - bp->b_blkno) * DEV_BSIZE + 1059 IDX_TO_OFF(m[0]->pindex) % bsize, 1060 ("wrong offsets bsize %d m[0] %ju b_pages[0] %ju " 1061 "blkno0 %ju b_blkno %ju", bsize, 1062 (uintmax_t)m[0]->pindex, (uintmax_t)bp->b_pages[0]->pindex, 1063 (uintmax_t)blkno0, (uintmax_t)bp->b_blkno)); 1064 1065 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 1066 VM_CNT_INC(v_vnodein); 1067 VM_CNT_ADD(v_vnodepgsin, bp->b_npages); 1068 1069 if (iodone != NULL) { /* async */ 1070 bp->b_pgiodone = iodone; 1071 bp->b_caller1 = arg; 1072 bp->b_iodone = vnode_pager_generic_getpages_done_async; 1073 bp->b_flags |= B_ASYNC; 1074 BUF_KERNPROC(bp); 1075 bstrategy(bp); 1076 return (VM_PAGER_OK); 1077 } else { 1078 bp->b_iodone = bdone; 1079 bstrategy(bp); 1080 bwait(bp, PVM, "vnread"); 1081 error = vnode_pager_generic_getpages_done(bp); 1082 for (i = 0; i < bp->b_npages; i++) 1083 bp->b_pages[i] = NULL; 1084 bp->b_vp = NULL; 1085 pbrelbo(bp); 1086 uma_zfree(vnode_pbuf_zone, bp); 1087 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); 1088 } 1089 } 1090 1091 static void 1092 vnode_pager_generic_getpages_done_async(struct buf *bp) 1093 { 1094 int error; 1095 1096 error = vnode_pager_generic_getpages_done(bp); 1097 /* Run the iodone upon the requested range. */ 1098 bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore, 1099 bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error); 1100 for (int i = 0; i < bp->b_npages; i++) 1101 bp->b_pages[i] = NULL; 1102 bp->b_vp = NULL; 1103 pbrelbo(bp); 1104 uma_zfree(vnode_pbuf_zone, bp); 1105 } 1106 1107 static int 1108 vnode_pager_generic_getpages_done(struct buf *bp) 1109 { 1110 vm_object_t object; 1111 off_t tfoff, nextoff; 1112 int i, error; 1113 1114 error = (bp->b_ioflags & BIO_ERROR) != 0 ? EIO : 0; 1115 object = bp->b_vp->v_object; 1116 1117 if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) { 1118 if (!buf_mapped(bp)) { 1119 bp->b_data = bp->b_kvabase; 1120 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, 1121 bp->b_npages); 1122 } 1123 bzero(bp->b_data + bp->b_bcount, 1124 PAGE_SIZE * bp->b_npages - bp->b_bcount); 1125 } 1126 if (buf_mapped(bp)) { 1127 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1128 bp->b_data = unmapped_buf; 1129 } 1130 1131 VM_OBJECT_WLOCK(object); 1132 for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex); 1133 i < bp->b_npages; i++, tfoff = nextoff) { 1134 vm_page_t mt; 1135 1136 nextoff = tfoff + PAGE_SIZE; 1137 mt = bp->b_pages[i]; 1138 1139 if (nextoff <= object->un_pager.vnp.vnp_size) { 1140 /* 1141 * Read filled up entire page. 1142 */ 1143 mt->valid = VM_PAGE_BITS_ALL; 1144 KASSERT(mt->dirty == 0, 1145 ("%s: page %p is dirty", __func__, mt)); 1146 KASSERT(!pmap_page_is_mapped(mt), 1147 ("%s: page %p is mapped", __func__, mt)); 1148 } else { 1149 /* 1150 * Read did not fill up entire page. 1151 * 1152 * Currently we do not set the entire page valid, 1153 * we just try to clear the piece that we couldn't 1154 * read. 1155 */ 1156 vm_page_set_valid_range(mt, 0, 1157 object->un_pager.vnp.vnp_size - tfoff); 1158 KASSERT((mt->dirty & vm_page_bits(0, 1159 object->un_pager.vnp.vnp_size - tfoff)) == 0, 1160 ("%s: page %p is dirty", __func__, mt)); 1161 } 1162 1163 if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter) 1164 vm_page_readahead_finish(mt); 1165 } 1166 VM_OBJECT_WUNLOCK(object); 1167 if (error != 0) 1168 printf("%s: I/O read error %d\n", __func__, error); 1169 1170 return (error); 1171 } 1172 1173 /* 1174 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 1175 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 1176 * vnode_pager_generic_putpages() to implement the previous behaviour. 1177 * 1178 * All other FS's should use the bypass to get to the local media 1179 * backing vp's VOP_PUTPAGES. 1180 */ 1181 static void 1182 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1183 int flags, int *rtvals) 1184 { 1185 int rtval; 1186 struct vnode *vp; 1187 int bytes = count * PAGE_SIZE; 1188 1189 /* 1190 * Force synchronous operation if we are extremely low on memory 1191 * to prevent a low-memory deadlock. VOP operations often need to 1192 * allocate more memory to initiate the I/O ( i.e. do a BMAP 1193 * operation ). The swapper handles the case by limiting the amount 1194 * of asynchronous I/O, but that sort of solution doesn't scale well 1195 * for the vnode pager without a lot of work. 1196 * 1197 * Also, the backing vnode's iodone routine may not wake the pageout 1198 * daemon up. This should be probably be addressed XXX. 1199 */ 1200 1201 if (vm_page_count_min()) 1202 flags |= VM_PAGER_PUT_SYNC; 1203 1204 /* 1205 * Call device-specific putpages function 1206 */ 1207 vp = object->handle; 1208 VM_OBJECT_WUNLOCK(object); 1209 rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals); 1210 KASSERT(rtval != EOPNOTSUPP, 1211 ("vnode_pager: stale FS putpages\n")); 1212 VM_OBJECT_WLOCK(object); 1213 } 1214 1215 static int 1216 vn_off2bidx(vm_ooffset_t offset) 1217 { 1218 1219 return ((offset & PAGE_MASK) / DEV_BSIZE); 1220 } 1221 1222 static bool 1223 vn_dirty_blk(vm_page_t m, vm_ooffset_t offset) 1224 { 1225 1226 KASSERT(IDX_TO_OFF(m->pindex) <= offset && 1227 offset < IDX_TO_OFF(m->pindex + 1), 1228 ("page %p pidx %ju offset %ju", m, (uintmax_t)m->pindex, 1229 (uintmax_t)offset)); 1230 return ((m->dirty & ((vm_page_bits_t)1 << vn_off2bidx(offset))) != 0); 1231 } 1232 1233 /* 1234 * This is now called from local media FS's to operate against their 1235 * own vnodes if they fail to implement VOP_PUTPAGES. 1236 * 1237 * This is typically called indirectly via the pageout daemon and 1238 * clustering has already typically occurred, so in general we ask the 1239 * underlying filesystem to write the data out asynchronously rather 1240 * then delayed. 1241 */ 1242 int 1243 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount, 1244 int flags, int *rtvals) 1245 { 1246 vm_object_t object; 1247 vm_page_t m; 1248 vm_ooffset_t maxblksz, next_offset, poffset, prev_offset; 1249 struct uio auio; 1250 struct iovec aiov; 1251 off_t prev_resid, wrsz; 1252 int count, error, i, maxsize, ncount, pgoff, ppscheck; 1253 bool in_hole; 1254 static struct timeval lastfail; 1255 static int curfail; 1256 1257 object = vp->v_object; 1258 count = bytecount / PAGE_SIZE; 1259 1260 for (i = 0; i < count; i++) 1261 rtvals[i] = VM_PAGER_ERROR; 1262 1263 if ((int64_t)ma[0]->pindex < 0) { 1264 printf("vnode_pager_generic_putpages: " 1265 "attempt to write meta-data 0x%jx(%lx)\n", 1266 (uintmax_t)ma[0]->pindex, (u_long)ma[0]->dirty); 1267 rtvals[0] = VM_PAGER_BAD; 1268 return (VM_PAGER_BAD); 1269 } 1270 1271 maxsize = count * PAGE_SIZE; 1272 ncount = count; 1273 1274 poffset = IDX_TO_OFF(ma[0]->pindex); 1275 1276 /* 1277 * If the page-aligned write is larger then the actual file we 1278 * have to invalidate pages occurring beyond the file EOF. However, 1279 * there is an edge case where a file may not be page-aligned where 1280 * the last page is partially invalid. In this case the filesystem 1281 * may not properly clear the dirty bits for the entire page (which 1282 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 1283 * With the page locked we are free to fix-up the dirty bits here. 1284 * 1285 * We do not under any circumstances truncate the valid bits, as 1286 * this will screw up bogus page replacement. 1287 */ 1288 VM_OBJECT_RLOCK(object); 1289 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 1290 if (!VM_OBJECT_TRYUPGRADE(object)) { 1291 VM_OBJECT_RUNLOCK(object); 1292 VM_OBJECT_WLOCK(object); 1293 if (maxsize + poffset <= object->un_pager.vnp.vnp_size) 1294 goto downgrade; 1295 } 1296 if (object->un_pager.vnp.vnp_size > poffset) { 1297 maxsize = object->un_pager.vnp.vnp_size - poffset; 1298 ncount = btoc(maxsize); 1299 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 1300 pgoff = roundup2(pgoff, DEV_BSIZE); 1301 1302 /* 1303 * If the object is locked and the following 1304 * conditions hold, then the page's dirty 1305 * field cannot be concurrently changed by a 1306 * pmap operation. 1307 */ 1308 m = ma[ncount - 1]; 1309 vm_page_assert_sbusied(m); 1310 KASSERT(!pmap_page_is_write_mapped(m), 1311 ("vnode_pager_generic_putpages: page %p is not read-only", m)); 1312 MPASS(m->dirty != 0); 1313 vm_page_clear_dirty(m, pgoff, PAGE_SIZE - 1314 pgoff); 1315 } 1316 } else { 1317 maxsize = 0; 1318 ncount = 0; 1319 } 1320 for (i = ncount; i < count; i++) 1321 rtvals[i] = VM_PAGER_BAD; 1322 downgrade: 1323 VM_OBJECT_LOCK_DOWNGRADE(object); 1324 } 1325 1326 auio.uio_iov = &aiov; 1327 auio.uio_segflg = UIO_NOCOPY; 1328 auio.uio_rw = UIO_WRITE; 1329 auio.uio_td = NULL; 1330 maxblksz = roundup2(poffset + maxsize, DEV_BSIZE); 1331 1332 for (prev_offset = poffset; prev_offset < maxblksz;) { 1333 /* Skip clean blocks. */ 1334 for (in_hole = true; in_hole && prev_offset < maxblksz;) { 1335 m = ma[OFF_TO_IDX(prev_offset - poffset)]; 1336 for (i = vn_off2bidx(prev_offset); 1337 i < sizeof(vm_page_bits_t) * NBBY && 1338 prev_offset < maxblksz; i++) { 1339 if (vn_dirty_blk(m, prev_offset)) { 1340 in_hole = false; 1341 break; 1342 } 1343 prev_offset += DEV_BSIZE; 1344 } 1345 } 1346 if (in_hole) 1347 goto write_done; 1348 1349 /* Find longest run of dirty blocks. */ 1350 for (next_offset = prev_offset; next_offset < maxblksz;) { 1351 m = ma[OFF_TO_IDX(next_offset - poffset)]; 1352 for (i = vn_off2bidx(next_offset); 1353 i < sizeof(vm_page_bits_t) * NBBY && 1354 next_offset < maxblksz; i++) { 1355 if (!vn_dirty_blk(m, next_offset)) 1356 goto start_write; 1357 next_offset += DEV_BSIZE; 1358 } 1359 } 1360 start_write: 1361 if (next_offset > poffset + maxsize) 1362 next_offset = poffset + maxsize; 1363 1364 /* 1365 * Getting here requires finding a dirty block in the 1366 * 'skip clean blocks' loop. 1367 */ 1368 MPASS(prev_offset < next_offset); 1369 1370 VM_OBJECT_RUNLOCK(object); 1371 aiov.iov_base = NULL; 1372 auio.uio_iovcnt = 1; 1373 auio.uio_offset = prev_offset; 1374 prev_resid = auio.uio_resid = aiov.iov_len = next_offset - 1375 prev_offset; 1376 error = VOP_WRITE(vp, &auio, 1377 vnode_pager_putpages_ioflags(flags), curthread->td_ucred); 1378 1379 wrsz = prev_resid - auio.uio_resid; 1380 if (wrsz == 0) { 1381 if (ppsratecheck(&lastfail, &curfail, 1) != 0) { 1382 vn_printf(vp, "vnode_pager_putpages: " 1383 "zero-length write at %ju resid %zd\n", 1384 auio.uio_offset, auio.uio_resid); 1385 } 1386 VM_OBJECT_RLOCK(object); 1387 break; 1388 } 1389 1390 /* Adjust the starting offset for next iteration. */ 1391 prev_offset += wrsz; 1392 MPASS(auio.uio_offset == prev_offset); 1393 1394 ppscheck = 0; 1395 if (error != 0 && (ppscheck = ppsratecheck(&lastfail, 1396 &curfail, 1)) != 0) 1397 vn_printf(vp, "vnode_pager_putpages: I/O error %d\n", 1398 error); 1399 if (auio.uio_resid != 0 && (ppscheck != 0 || 1400 ppsratecheck(&lastfail, &curfail, 1) != 0)) 1401 vn_printf(vp, "vnode_pager_putpages: residual I/O %zd " 1402 "at %ju\n", auio.uio_resid, 1403 (uintmax_t)ma[0]->pindex); 1404 VM_OBJECT_RLOCK(object); 1405 if (error != 0 || auio.uio_resid != 0) 1406 break; 1407 } 1408 write_done: 1409 /* Mark completely processed pages. */ 1410 for (i = 0; i < OFF_TO_IDX(prev_offset - poffset); i++) 1411 rtvals[i] = VM_PAGER_OK; 1412 /* Mark partial EOF page. */ 1413 if (prev_offset == poffset + maxsize && (prev_offset & PAGE_MASK) != 0) 1414 rtvals[i++] = VM_PAGER_OK; 1415 /* Unwritten pages in range, free bonus if the page is clean. */ 1416 for (; i < ncount; i++) 1417 rtvals[i] = ma[i]->dirty == 0 ? VM_PAGER_OK : VM_PAGER_ERROR; 1418 VM_OBJECT_RUNLOCK(object); 1419 VM_CNT_ADD(v_vnodepgsout, i); 1420 VM_CNT_INC(v_vnodeout); 1421 return (rtvals[0]); 1422 } 1423 1424 int 1425 vnode_pager_putpages_ioflags(int pager_flags) 1426 { 1427 int ioflags; 1428 1429 /* 1430 * Pageouts are already clustered, use IO_ASYNC to force a 1431 * bawrite() rather then a bdwrite() to prevent paging I/O 1432 * from saturating the buffer cache. Dummy-up the sequential 1433 * heuristic to cause large ranges to cluster. If neither 1434 * IO_SYNC or IO_ASYNC is set, the system decides how to 1435 * cluster. 1436 */ 1437 ioflags = IO_VMIO; 1438 if ((pager_flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) != 0) 1439 ioflags |= IO_SYNC; 1440 else if ((pager_flags & VM_PAGER_CLUSTER_OK) == 0) 1441 ioflags |= IO_ASYNC; 1442 ioflags |= (pager_flags & VM_PAGER_PUT_INVAL) != 0 ? IO_INVAL: 0; 1443 ioflags |= (pager_flags & VM_PAGER_PUT_NOREUSE) != 0 ? IO_NOREUSE : 0; 1444 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 1445 return (ioflags); 1446 } 1447 1448 /* 1449 * vnode_pager_undirty_pages(). 1450 * 1451 * A helper to mark pages as clean after pageout that was possibly 1452 * done with a short write. The lpos argument specifies the page run 1453 * length in bytes, and the written argument specifies how many bytes 1454 * were actually written. eof is the offset past the last valid byte 1455 * in the vnode using the absolute file position of the first byte in 1456 * the run as the base from which it is computed. 1457 */ 1458 void 1459 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written, off_t eof, 1460 int lpos) 1461 { 1462 vm_object_t obj; 1463 int i, pos, pos_devb; 1464 1465 if (written == 0 && eof >= lpos) 1466 return; 1467 obj = ma[0]->object; 1468 VM_OBJECT_WLOCK(obj); 1469 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) { 1470 if (pos < trunc_page(written)) { 1471 rtvals[i] = VM_PAGER_OK; 1472 vm_page_undirty(ma[i]); 1473 } else { 1474 /* Partially written page. */ 1475 rtvals[i] = VM_PAGER_AGAIN; 1476 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK); 1477 } 1478 } 1479 if (eof >= lpos) /* avoid truncation */ 1480 goto done; 1481 for (pos = eof, i = OFF_TO_IDX(trunc_page(pos)); pos < lpos; i++) { 1482 if (pos != trunc_page(pos)) { 1483 /* 1484 * The page contains the last valid byte in 1485 * the vnode, mark the rest of the page as 1486 * clean, potentially making the whole page 1487 * clean. 1488 */ 1489 pos_devb = roundup2(pos & PAGE_MASK, DEV_BSIZE); 1490 vm_page_clear_dirty(ma[i], pos_devb, PAGE_SIZE - 1491 pos_devb); 1492 1493 /* 1494 * If the page was cleaned, report the pageout 1495 * on it as successful. msync() no longer 1496 * needs to write out the page, endlessly 1497 * creating write requests and dirty buffers. 1498 */ 1499 if (ma[i]->dirty == 0) 1500 rtvals[i] = VM_PAGER_OK; 1501 1502 pos = round_page(pos); 1503 } else { 1504 /* vm_pageout_flush() clears dirty */ 1505 rtvals[i] = VM_PAGER_BAD; 1506 pos += PAGE_SIZE; 1507 } 1508 } 1509 done: 1510 VM_OBJECT_WUNLOCK(obj); 1511 } 1512 1513 void 1514 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start, 1515 vm_offset_t end) 1516 { 1517 struct vnode *vp; 1518 vm_ooffset_t old_wm; 1519 1520 VM_OBJECT_WLOCK(object); 1521 if (object->type != OBJT_VNODE) { 1522 VM_OBJECT_WUNLOCK(object); 1523 return; 1524 } 1525 old_wm = object->un_pager.vnp.writemappings; 1526 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start; 1527 vp = object->handle; 1528 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) { 1529 ASSERT_VOP_LOCKED(vp, "v_writecount inc"); 1530 VOP_ADD_WRITECOUNT_CHECKED(vp, 1); 1531 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", 1532 __func__, vp, vp->v_writecount); 1533 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) { 1534 ASSERT_VOP_LOCKED(vp, "v_writecount dec"); 1535 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 1536 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 1537 __func__, vp, vp->v_writecount); 1538 } 1539 VM_OBJECT_WUNLOCK(object); 1540 } 1541 1542 void 1543 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start, 1544 vm_offset_t end) 1545 { 1546 struct vnode *vp; 1547 struct mount *mp; 1548 vm_offset_t inc; 1549 1550 VM_OBJECT_WLOCK(object); 1551 1552 /* 1553 * First, recheck the object type to account for the race when 1554 * the vnode is reclaimed. 1555 */ 1556 if (object->type != OBJT_VNODE) { 1557 VM_OBJECT_WUNLOCK(object); 1558 return; 1559 } 1560 1561 /* 1562 * Optimize for the case when writemappings is not going to 1563 * zero. 1564 */ 1565 inc = end - start; 1566 if (object->un_pager.vnp.writemappings != inc) { 1567 object->un_pager.vnp.writemappings -= inc; 1568 VM_OBJECT_WUNLOCK(object); 1569 return; 1570 } 1571 1572 vp = object->handle; 1573 vhold(vp); 1574 VM_OBJECT_WUNLOCK(object); 1575 mp = NULL; 1576 vn_start_write(vp, &mp, V_WAIT); 1577 vn_lock(vp, LK_SHARED | LK_RETRY); 1578 1579 /* 1580 * Decrement the object's writemappings, by swapping the start 1581 * and end arguments for vnode_pager_update_writecount(). If 1582 * there was not a race with vnode reclaimation, then the 1583 * vnode's v_writecount is decremented. 1584 */ 1585 vnode_pager_update_writecount(object, end, start); 1586 VOP_UNLOCK(vp, 0); 1587 vdrop(vp); 1588 if (mp != NULL) 1589 vn_finished_write(mp); 1590 } 1591