1/*- 2 * Copyright (c) 1990 University of Utah. 3 * Copyright (c) 1991 The Regents of the University of California. 4 * All rights reserved. 5 * Copyright (c) 1993, 1994 John S. Dyson 6 * Copyright (c) 1995, David Greenman 7 * 8 * This code is derived from software contributed to Berkeley by 9 * the Systems Programming Group of the University of Utah Computer 10 * Science Department. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. All advertising materials mentioning features or use of this software 21 * must display the following acknowledgement: 22 * This product includes software developed by the University of 23 * California, Berkeley and its contributors. 24 * 4. Neither the name of the University nor the names of its contributors 25 * may be used to endorse or promote products derived from this software 26 * without specific prior written permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 38 * SUCH DAMAGE. 39 * 40 * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91 41 */ 42 43/* 44 * Page to/from files (vnodes). 45 */ 46 47/* 48 * TODO: 49 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will 50 * greatly re-simplify the vnode_pager. 51 */ 52 53#include <sys/cdefs.h> 54__FBSDID("$FreeBSD$"); 55 56#include <sys/param.h> 57#include <sys/systm.h> 58#include <sys/proc.h> 59#include <sys/vnode.h> 60#include <sys/mount.h> 61#include <sys/bio.h> 62#include <sys/buf.h> 63#include <sys/vmmeter.h> 64#include <sys/limits.h> 65#include <sys/conf.h> 66#include <sys/sf_buf.h> 67 68#include <machine/atomic.h> 69 70#include <vm/vm.h> 71#include <vm/vm_param.h> 72#include <vm/vm_object.h> 73#include <vm/vm_page.h> 74#include <vm/vm_pager.h> 75#include <vm/vm_map.h> 76#include <vm/vnode_pager.h> 77#include <vm/vm_extern.h> 78 79static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, 80 daddr_t *rtaddress, int *run); 81static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m); 82static int vnode_pager_input_old(vm_object_t object, vm_page_t m); 83static void vnode_pager_dealloc(vm_object_t); 84static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int); 85static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 86static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *); 87static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t, 88 vm_ooffset_t, struct ucred *cred); 89 90struct pagerops vnodepagerops = { 91 .pgo_alloc = vnode_pager_alloc, 92 .pgo_dealloc = vnode_pager_dealloc, 93 .pgo_getpages = vnode_pager_getpages, 94 .pgo_putpages = vnode_pager_putpages, 95 .pgo_haspage = vnode_pager_haspage, 96}; 97 98int vnode_pbuf_freecnt; 99 100/* Create the VM system backing object for this vnode */ 101int 102vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td) 103{ 104 vm_object_t object; 105 vm_ooffset_t size = isize; 106 struct vattr va; 107 108 if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE) 109 return (0); 110 111 while ((object = vp->v_object) != NULL) { 112 VM_OBJECT_LOCK(object); 113 if (!(object->flags & OBJ_DEAD)) { 114 VM_OBJECT_UNLOCK(object); 115 return (0); 116 } 117 VOP_UNLOCK(vp, 0); 118 vm_object_set_flag(object, OBJ_DISCONNECTWNT); 119 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vodead", 0); 120 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 121 } 122 123 if (size == 0) { 124 if (vn_isdisk(vp, NULL)) { 125 size = IDX_TO_OFF(INT_MAX); 126 } else { 127 if (VOP_GETATTR(vp, &va, td->td_ucred)) 128 return (0); 129 size = va.va_size; 130 } 131 } 132 133 object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred); 134 /* 135 * Dereference the reference we just created. This assumes 136 * that the object is associated with the vp. 137 */ 138 VM_OBJECT_LOCK(object); 139 object->ref_count--; 140 VM_OBJECT_UNLOCK(object); 141 vrele(vp); 142 143 KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object")); 144 145 return (0); 146} 147 148void 149vnode_destroy_vobject(struct vnode *vp) 150{ 151 struct vm_object *obj; 152 153 obj = vp->v_object; 154 if (obj == NULL) 155 return; 156 ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject"); 157 VM_OBJECT_LOCK(obj); 158 if (obj->ref_count == 0) { 159 /* 160 * vclean() may be called twice. The first time 161 * removes the primary reference to the object, 162 * the second time goes one further and is a 163 * special-case to terminate the object. 164 * 165 * don't double-terminate the object 166 */ 167 if ((obj->flags & OBJ_DEAD) == 0) 168 vm_object_terminate(obj); 169 else 170 VM_OBJECT_UNLOCK(obj); 171 } else { 172 /* 173 * Woe to the process that tries to page now :-). 174 */ 175 vm_pager_deallocate(obj); 176 VM_OBJECT_UNLOCK(obj); 177 } 178 vp->v_object = NULL; 179} 180 181 182/* 183 * Allocate (or lookup) pager for a vnode. 184 * Handle is a vnode pointer. 185 * 186 * MPSAFE 187 */ 188vm_object_t 189vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 190 vm_ooffset_t offset, struct ucred *cred) 191{ 192 vm_object_t object; 193 struct vnode *vp; 194 195 /* 196 * Pageout to vnode, no can do yet. 197 */ 198 if (handle == NULL) 199 return (NULL); 200 201 vp = (struct vnode *) handle; 202 203 /* 204 * If the object is being terminated, wait for it to 205 * go away. 206 */ 207retry: 208 while ((object = vp->v_object) != NULL) { 209 VM_OBJECT_LOCK(object); 210 if ((object->flags & OBJ_DEAD) == 0) 211 break; 212 vm_object_set_flag(object, OBJ_DISCONNECTWNT); 213 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vadead", 0); 214 } 215 216 KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference")); 217 218 if (object == NULL) { 219 /* 220 * Add an object of the appropriate size 221 */ 222 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); 223 224 object->un_pager.vnp.vnp_size = size; 225 object->un_pager.vnp.writemappings = 0; 226 227 object->handle = handle; 228 VI_LOCK(vp); 229 if (vp->v_object != NULL) { 230 /* 231 * Object has been created while we were sleeping 232 */ 233 VI_UNLOCK(vp); 234 vm_object_destroy(object); 235 goto retry; 236 } 237 vp->v_object = object; 238 VI_UNLOCK(vp); 239 } else { 240 object->ref_count++; 241 VM_OBJECT_UNLOCK(object); 242 } 243 vref(vp); 244 return (object); 245} 246 247/* 248 * The object must be locked. 249 */ 250static void 251vnode_pager_dealloc(object) 252 vm_object_t object; 253{ 254 struct vnode *vp; 255 int refs; 256 257 vp = object->handle; 258 if (vp == NULL) 259 panic("vnode_pager_dealloc: pager already dealloced"); 260 261 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 262 vm_object_pip_wait(object, "vnpdea"); 263 refs = object->ref_count; 264 265 object->handle = NULL; 266 object->type = OBJT_DEAD; 267 if (object->flags & OBJ_DISCONNECTWNT) { 268 vm_object_clear_flag(object, OBJ_DISCONNECTWNT); 269 wakeup(object); 270 } 271 ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc"); 272 if (object->un_pager.vnp.writemappings > 0) { 273 object->un_pager.vnp.writemappings = 0; 274 VOP_ADD_WRITECOUNT(vp, -1); 275 } 276 vp->v_object = NULL; 277 VOP_UNSET_TEXT(vp); 278 VM_OBJECT_UNLOCK(object); 279 while (refs-- > 0) 280 vunref(vp); 281 VM_OBJECT_LOCK(object); 282} 283 284static boolean_t 285vnode_pager_haspage(object, pindex, before, after) 286 vm_object_t object; 287 vm_pindex_t pindex; 288 int *before; 289 int *after; 290{ 291 struct vnode *vp = object->handle; 292 daddr_t bn; 293 int err; 294 daddr_t reqblock; 295 int poff; 296 int bsize; 297 int pagesperblock, blocksperpage; 298 int vfslocked; 299 300 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 301 /* 302 * If no vp or vp is doomed or marked transparent to VM, we do not 303 * have the page. 304 */ 305 if (vp == NULL || vp->v_iflag & VI_DOOMED) 306 return FALSE; 307 /* 308 * If the offset is beyond end of file we do 309 * not have the page. 310 */ 311 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size) 312 return FALSE; 313 314 bsize = vp->v_mount->mnt_stat.f_iosize; 315 pagesperblock = bsize / PAGE_SIZE; 316 blocksperpage = 0; 317 if (pagesperblock > 0) { 318 reqblock = pindex / pagesperblock; 319 } else { 320 blocksperpage = (PAGE_SIZE / bsize); 321 reqblock = pindex * blocksperpage; 322 } 323 VM_OBJECT_UNLOCK(object); 324 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 325 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before); 326 VFS_UNLOCK_GIANT(vfslocked); 327 VM_OBJECT_LOCK(object); 328 if (err) 329 return TRUE; 330 if (bn == -1) 331 return FALSE; 332 if (pagesperblock > 0) { 333 poff = pindex - (reqblock * pagesperblock); 334 if (before) { 335 *before *= pagesperblock; 336 *before += poff; 337 } 338 if (after) { 339 int numafter; 340 *after *= pagesperblock; 341 numafter = pagesperblock - (poff + 1); 342 if (IDX_TO_OFF(pindex + numafter) > 343 object->un_pager.vnp.vnp_size) { 344 numafter = 345 OFF_TO_IDX(object->un_pager.vnp.vnp_size) - 346 pindex; 347 } 348 *after += numafter; 349 } 350 } else { 351 if (before) { 352 *before /= blocksperpage; 353 } 354 355 if (after) { 356 *after /= blocksperpage; 357 } 358 } 359 return TRUE; 360} 361 362/* 363 * Lets the VM system know about a change in size for a file. 364 * We adjust our own internal size and flush any cached pages in 365 * the associated object that are affected by the size change. 366 * 367 * Note: this routine may be invoked as a result of a pager put 368 * operation (possibly at object termination time), so we must be careful. 369 */ 370void 371vnode_pager_setsize(vp, nsize) 372 struct vnode *vp; 373 vm_ooffset_t nsize; 374{ 375 vm_object_t object; 376 vm_page_t m; 377 vm_pindex_t nobjsize; 378 379 if ((object = vp->v_object) == NULL) 380 return; 381/* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */ 382 VM_OBJECT_LOCK(object); 383 if (object->type == OBJT_DEAD) { 384 VM_OBJECT_UNLOCK(object); 385 return; 386 } 387 KASSERT(object->type == OBJT_VNODE, 388 ("not vnode-backed object %p", object)); 389 if (nsize == object->un_pager.vnp.vnp_size) { 390 /* 391 * Hasn't changed size 392 */ 393 VM_OBJECT_UNLOCK(object); 394 return; 395 } 396 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 397 if (nsize < object->un_pager.vnp.vnp_size) { 398 /* 399 * File has shrunk. Toss any cached pages beyond the new EOF. 400 */ 401 if (nobjsize < object->size) 402 vm_object_page_remove(object, nobjsize, object->size, 403 0); 404 /* 405 * this gets rid of garbage at the end of a page that is now 406 * only partially backed by the vnode. 407 * 408 * XXX for some reason (I don't know yet), if we take a 409 * completely invalid page and mark it partially valid 410 * it can screw up NFS reads, so we don't allow the case. 411 */ 412 if ((nsize & PAGE_MASK) && 413 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL && 414 m->valid != 0) { 415 int base = (int)nsize & PAGE_MASK; 416 int size = PAGE_SIZE - base; 417 418 /* 419 * Clear out partial-page garbage in case 420 * the page has been mapped. 421 */ 422 pmap_zero_page_area(m, base, size); 423 424 /* 425 * Update the valid bits to reflect the blocks that 426 * have been zeroed. Some of these valid bits may 427 * have already been set. 428 */ 429 vm_page_set_valid(m, base, size); 430 431 /* 432 * Round "base" to the next block boundary so that the 433 * dirty bit for a partially zeroed block is not 434 * cleared. 435 */ 436 base = roundup2(base, DEV_BSIZE); 437 438 /* 439 * Clear out partial-page dirty bits. 440 * 441 * note that we do not clear out the valid 442 * bits. This would prevent bogus_page 443 * replacement from working properly. 444 */ 445 vm_page_clear_dirty(m, base, PAGE_SIZE - base); 446 } else if ((nsize & PAGE_MASK) && 447 __predict_false(object->cache != NULL)) { 448 vm_page_cache_free(object, OFF_TO_IDX(nsize), 449 nobjsize); 450 } 451 } 452 object->un_pager.vnp.vnp_size = nsize; 453 object->size = nobjsize; 454 VM_OBJECT_UNLOCK(object); 455} 456 457/* 458 * calculate the linear (byte) disk address of specified virtual 459 * file address 460 */ 461static int 462vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress, 463 int *run) 464{ 465 int bsize; 466 int err; 467 daddr_t vblock; 468 daddr_t voffset; 469 470 if (address < 0) 471 return -1; 472 473 if (vp->v_iflag & VI_DOOMED) 474 return -1; 475 476 bsize = vp->v_mount->mnt_stat.f_iosize; 477 vblock = address / bsize; 478 voffset = address % bsize; 479 480 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL); 481 if (err == 0) { 482 if (*rtaddress != -1) 483 *rtaddress += voffset / DEV_BSIZE; 484 if (run) { 485 *run += 1; 486 *run *= bsize/PAGE_SIZE; 487 *run -= voffset/PAGE_SIZE; 488 } 489 } 490 491 return (err); 492} 493 494/* 495 * small block filesystem vnode pager input 496 */ 497static int 498vnode_pager_input_smlfs(object, m) 499 vm_object_t object; 500 vm_page_t m; 501{ 502 struct vnode *vp; 503 struct bufobj *bo; 504 struct buf *bp; 505 struct sf_buf *sf; 506 daddr_t fileaddr; 507 vm_offset_t bsize; 508 vm_page_bits_t bits; 509 int error, i; 510 511 error = 0; 512 vp = object->handle; 513 if (vp->v_iflag & VI_DOOMED) 514 return VM_PAGER_BAD; 515 516 bsize = vp->v_mount->mnt_stat.f_iosize; 517 518 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL); 519 520 sf = sf_buf_alloc(m, 0); 521 522 for (i = 0; i < PAGE_SIZE / bsize; i++) { 523 vm_ooffset_t address; 524 525 bits = vm_page_bits(i * bsize, bsize); 526 if (m->valid & bits) 527 continue; 528 529 address = IDX_TO_OFF(m->pindex) + i * bsize; 530 if (address >= object->un_pager.vnp.vnp_size) { 531 fileaddr = -1; 532 } else { 533 error = vnode_pager_addr(vp, address, &fileaddr, NULL); 534 if (error) 535 break; 536 } 537 if (fileaddr != -1) { 538 bp = getpbuf(&vnode_pbuf_freecnt); 539 540 /* build a minimal buffer header */ 541 bp->b_iocmd = BIO_READ; 542 bp->b_iodone = bdone; 543 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 544 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 545 bp->b_rcred = crhold(curthread->td_ucred); 546 bp->b_wcred = crhold(curthread->td_ucred); 547 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize; 548 bp->b_blkno = fileaddr; 549 pbgetbo(bo, bp); 550 bp->b_vp = vp; 551 bp->b_bcount = bsize; 552 bp->b_bufsize = bsize; 553 bp->b_runningbufspace = bp->b_bufsize; 554 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 555 556 /* do the input */ 557 bp->b_iooffset = dbtob(bp->b_blkno); 558 bstrategy(bp); 559 560 bwait(bp, PVM, "vnsrd"); 561 562 if ((bp->b_ioflags & BIO_ERROR) != 0) 563 error = EIO; 564 565 /* 566 * free the buffer header back to the swap buffer pool 567 */ 568 bp->b_vp = NULL; 569 pbrelbo(bp); 570 relpbuf(bp, &vnode_pbuf_freecnt); 571 if (error) 572 break; 573 } else 574 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize); 575 KASSERT((m->dirty & bits) == 0, 576 ("vnode_pager_input_smlfs: page %p is dirty", m)); 577 VM_OBJECT_LOCK(object); 578 m->valid |= bits; 579 VM_OBJECT_UNLOCK(object); 580 } 581 sf_buf_free(sf); 582 if (error) { 583 return VM_PAGER_ERROR; 584 } 585 return VM_PAGER_OK; 586} 587 588/* 589 * old style vnode pager input routine 590 */ 591static int 592vnode_pager_input_old(object, m) 593 vm_object_t object; 594 vm_page_t m; 595{ 596 struct uio auio; 597 struct iovec aiov; 598 int error; 599 int size; 600 struct sf_buf *sf; 601 struct vnode *vp; 602 603 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 604 error = 0; 605 606 /* 607 * Return failure if beyond current EOF 608 */ 609 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 610 return VM_PAGER_BAD; 611 } else { 612 size = PAGE_SIZE; 613 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 614 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 615 vp = object->handle; 616 VM_OBJECT_UNLOCK(object); 617 618 /* 619 * Allocate a kernel virtual address and initialize so that 620 * we can use VOP_READ/WRITE routines. 621 */ 622 sf = sf_buf_alloc(m, 0); 623 624 aiov.iov_base = (caddr_t)sf_buf_kva(sf); 625 aiov.iov_len = size; 626 auio.uio_iov = &aiov; 627 auio.uio_iovcnt = 1; 628 auio.uio_offset = IDX_TO_OFF(m->pindex); 629 auio.uio_segflg = UIO_SYSSPACE; 630 auio.uio_rw = UIO_READ; 631 auio.uio_resid = size; 632 auio.uio_td = curthread; 633 634 error = VOP_READ(vp, &auio, 0, curthread->td_ucred); 635 if (!error) { 636 int count = size - auio.uio_resid; 637 638 if (count == 0) 639 error = EINVAL; 640 else if (count != PAGE_SIZE) 641 bzero((caddr_t)sf_buf_kva(sf) + count, 642 PAGE_SIZE - count); 643 } 644 sf_buf_free(sf); 645 646 VM_OBJECT_LOCK(object); 647 } 648 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m)); 649 if (!error) 650 m->valid = VM_PAGE_BITS_ALL; 651 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 652} 653 654/* 655 * generic vnode pager input routine 656 */ 657 658/* 659 * Local media VFS's that do not implement their own VOP_GETPAGES 660 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages() 661 * to implement the previous behaviour. 662 * 663 * All other FS's should use the bypass to get to the local media 664 * backing vp's VOP_GETPAGES. 665 */ 666static int 667vnode_pager_getpages(object, m, count, reqpage) 668 vm_object_t object; 669 vm_page_t *m; 670 int count; 671 int reqpage; 672{ 673 int rtval; 674 struct vnode *vp; 675 int bytes = count * PAGE_SIZE; 676 int vfslocked; 677 678 vp = object->handle; 679 VM_OBJECT_UNLOCK(object); 680 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 681 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0); 682 KASSERT(rtval != EOPNOTSUPP, 683 ("vnode_pager: FS getpages not implemented\n")); 684 VFS_UNLOCK_GIANT(vfslocked); 685 VM_OBJECT_LOCK(object); 686 return rtval; 687} 688 689/* 690 * This is now called from local media FS's to operate against their 691 * own vnodes if they fail to implement VOP_GETPAGES. 692 */ 693int 694vnode_pager_generic_getpages(vp, m, bytecount, reqpage) 695 struct vnode *vp; 696 vm_page_t *m; 697 int bytecount; 698 int reqpage; 699{ 700 vm_object_t object; 701 vm_offset_t kva; 702 off_t foff, tfoff, nextoff; 703 int i, j, size, bsize, first; 704 daddr_t firstaddr, reqblock; 705 struct bufobj *bo; 706 int runpg; 707 int runend; 708 struct buf *bp; 709 struct mount *mp; 710 int count; 711 int error; 712 713 object = vp->v_object; 714 count = bytecount / PAGE_SIZE; 715 716 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 717 ("vnode_pager_generic_getpages does not support devices")); 718 if (vp->v_iflag & VI_DOOMED) 719 return VM_PAGER_BAD; 720 721 bsize = vp->v_mount->mnt_stat.f_iosize; 722 723 /* get the UNDERLYING device for the file with VOP_BMAP() */ 724 725 /* 726 * originally, we did not check for an error return value -- assuming 727 * an fs always has a bmap entry point -- that assumption is wrong!!! 728 */ 729 foff = IDX_TO_OFF(m[reqpage]->pindex); 730 731 /* 732 * if we can't bmap, use old VOP code 733 */ 734 error = VOP_BMAP(vp, foff / bsize, &bo, &reqblock, NULL, NULL); 735 if (error == EOPNOTSUPP) { 736 VM_OBJECT_LOCK(object); 737 738 for (i = 0; i < count; i++) 739 if (i != reqpage) { 740 vm_page_lock(m[i]); 741 vm_page_free(m[i]); 742 vm_page_unlock(m[i]); 743 } 744 PCPU_INC(cnt.v_vnodein); 745 PCPU_INC(cnt.v_vnodepgsin); 746 error = vnode_pager_input_old(object, m[reqpage]); 747 VM_OBJECT_UNLOCK(object); 748 return (error); 749 } else if (error != 0) { 750 VM_OBJECT_LOCK(object); 751 for (i = 0; i < count; i++) 752 if (i != reqpage) { 753 vm_page_lock(m[i]); 754 vm_page_free(m[i]); 755 vm_page_unlock(m[i]); 756 } 757 VM_OBJECT_UNLOCK(object); 758 return (VM_PAGER_ERROR); 759 760 /* 761 * if the blocksize is smaller than a page size, then use 762 * special small filesystem code. NFS sometimes has a small 763 * blocksize, but it can handle large reads itself. 764 */ 765 } else if ((PAGE_SIZE / bsize) > 1 && 766 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) { 767 VM_OBJECT_LOCK(object); 768 for (i = 0; i < count; i++) 769 if (i != reqpage) { 770 vm_page_lock(m[i]); 771 vm_page_free(m[i]); 772 vm_page_unlock(m[i]); 773 } 774 VM_OBJECT_UNLOCK(object); 775 PCPU_INC(cnt.v_vnodein); 776 PCPU_INC(cnt.v_vnodepgsin); 777 return vnode_pager_input_smlfs(object, m[reqpage]); 778 } 779 780 /* 781 * If we have a completely valid page available to us, we can 782 * clean up and return. Otherwise we have to re-read the 783 * media. 784 */ 785 VM_OBJECT_LOCK(object); 786 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) { 787 for (i = 0; i < count; i++) 788 if (i != reqpage) { 789 vm_page_lock(m[i]); 790 vm_page_free(m[i]); 791 vm_page_unlock(m[i]); 792 } 793 VM_OBJECT_UNLOCK(object); 794 return VM_PAGER_OK; 795 } else if (reqblock == -1) { 796 pmap_zero_page(m[reqpage]); 797 KASSERT(m[reqpage]->dirty == 0, 798 ("vnode_pager_generic_getpages: page %p is dirty", m)); 799 m[reqpage]->valid = VM_PAGE_BITS_ALL; 800 for (i = 0; i < count; i++) 801 if (i != reqpage) { 802 vm_page_lock(m[i]); 803 vm_page_free(m[i]); 804 vm_page_unlock(m[i]); 805 } 806 VM_OBJECT_UNLOCK(object); 807 return (VM_PAGER_OK); 808 } 809 m[reqpage]->valid = 0; 810 VM_OBJECT_UNLOCK(object); 811 812 /* 813 * here on direct device I/O 814 */ 815 firstaddr = -1; 816 817 /* 818 * calculate the run that includes the required page 819 */ 820 for (first = 0, i = 0; i < count; i = runend) { 821 if (vnode_pager_addr(vp, IDX_TO_OFF(m[i]->pindex), &firstaddr, 822 &runpg) != 0) { 823 VM_OBJECT_LOCK(object); 824 for (; i < count; i++) 825 if (i != reqpage) { 826 vm_page_lock(m[i]); 827 vm_page_free(m[i]); 828 vm_page_unlock(m[i]); 829 } 830 VM_OBJECT_UNLOCK(object); 831 return (VM_PAGER_ERROR); 832 } 833 if (firstaddr == -1) { 834 VM_OBJECT_LOCK(object); 835 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) { 836 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx", 837 (intmax_t)firstaddr, (uintmax_t)(foff >> 32), 838 (uintmax_t)foff, 839 (uintmax_t) 840 (object->un_pager.vnp.vnp_size >> 32), 841 (uintmax_t)object->un_pager.vnp.vnp_size); 842 } 843 vm_page_lock(m[i]); 844 vm_page_free(m[i]); 845 vm_page_unlock(m[i]); 846 VM_OBJECT_UNLOCK(object); 847 runend = i + 1; 848 first = runend; 849 continue; 850 } 851 runend = i + runpg; 852 if (runend <= reqpage) { 853 VM_OBJECT_LOCK(object); 854 for (j = i; j < runend; j++) { 855 vm_page_lock(m[j]); 856 vm_page_free(m[j]); 857 vm_page_unlock(m[j]); 858 } 859 VM_OBJECT_UNLOCK(object); 860 } else { 861 if (runpg < (count - first)) { 862 VM_OBJECT_LOCK(object); 863 for (i = first + runpg; i < count; i++) { 864 vm_page_lock(m[i]); 865 vm_page_free(m[i]); 866 vm_page_unlock(m[i]); 867 } 868 VM_OBJECT_UNLOCK(object); 869 count = first + runpg; 870 } 871 break; 872 } 873 first = runend; 874 } 875 876 /* 877 * the first and last page have been calculated now, move input pages 878 * to be zero based... 879 */ 880 if (first != 0) { 881 m += first; 882 count -= first; 883 reqpage -= first; 884 } 885 886 /* 887 * calculate the file virtual address for the transfer 888 */ 889 foff = IDX_TO_OFF(m[0]->pindex); 890 891 /* 892 * calculate the size of the transfer 893 */ 894 size = count * PAGE_SIZE; 895 KASSERT(count > 0, ("zero count")); 896 if ((foff + size) > object->un_pager.vnp.vnp_size) 897 size = object->un_pager.vnp.vnp_size - foff; 898 KASSERT(size > 0, ("zero size")); 899 900 /* 901 * round up physical size for real devices. 902 */ 903 if (1) { 904 int secmask = bo->bo_bsize - 1; 905 KASSERT(secmask < PAGE_SIZE && secmask > 0, 906 ("vnode_pager_generic_getpages: sector size %d too large", 907 secmask + 1)); 908 size = (size + secmask) & ~secmask; 909 } 910 911 bp = getpbuf(&vnode_pbuf_freecnt); 912 kva = (vm_offset_t)bp->b_data; 913 914 /* 915 * and map the pages to be read into the kva, if the filesystem 916 * requires mapped buffers. 917 */ 918 mp = vp->v_mount; 919 if (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 && 920 unmapped_buf_allowed) { 921 bp->b_data = unmapped_buf; 922 bp->b_kvabase = unmapped_buf; 923 bp->b_offset = 0; 924 bp->b_flags |= B_UNMAPPED; 925 bp->b_npages = count; 926 for (i = 0; i < count; i++) 927 bp->b_pages[i] = m[i]; 928 } else 929 pmap_qenter(kva, m, count); 930 931 /* build a minimal buffer header */ 932 bp->b_iocmd = BIO_READ; 933 bp->b_iodone = bdone; 934 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 935 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 936 bp->b_rcred = crhold(curthread->td_ucred); 937 bp->b_wcred = crhold(curthread->td_ucred); 938 bp->b_blkno = firstaddr; 939 pbgetbo(bo, bp); 940 bp->b_vp = vp; 941 bp->b_bcount = size; 942 bp->b_bufsize = size; 943 bp->b_runningbufspace = bp->b_bufsize; 944 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 945 946 PCPU_INC(cnt.v_vnodein); 947 PCPU_ADD(cnt.v_vnodepgsin, count); 948 949 /* do the input */ 950 bp->b_iooffset = dbtob(bp->b_blkno); 951 bstrategy(bp); 952 953 bwait(bp, PVM, "vnread"); 954 955 if ((bp->b_ioflags & BIO_ERROR) != 0) 956 error = EIO; 957 958 if (error == 0 && size != count * PAGE_SIZE) { 959 if ((bp->b_flags & B_UNMAPPED) != 0) { 960 bp->b_flags &= ~B_UNMAPPED; 961 pmap_qenter(kva, m, count); 962 } 963 bzero((caddr_t)kva + size, PAGE_SIZE * count - size); 964 } 965 if ((bp->b_flags & B_UNMAPPED) == 0) 966 pmap_qremove(kva, count); 967 if (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0) { 968 bp->b_data = (caddr_t)kva; 969 bp->b_kvabase = (caddr_t)kva; 970 bp->b_flags &= ~B_UNMAPPED; 971 for (i = 0; i < count; i++) 972 bp->b_pages[i] = NULL; 973 } 974 975 /* 976 * free the buffer header back to the swap buffer pool 977 */ 978 bp->b_vp = NULL; 979 pbrelbo(bp); 980 relpbuf(bp, &vnode_pbuf_freecnt); 981 982 VM_OBJECT_LOCK(object); 983 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) { 984 vm_page_t mt; 985 986 nextoff = tfoff + PAGE_SIZE; 987 mt = m[i]; 988 989 if (nextoff <= object->un_pager.vnp.vnp_size) { 990 /* 991 * Read filled up entire page. 992 */ 993 mt->valid = VM_PAGE_BITS_ALL; 994 KASSERT(mt->dirty == 0, 995 ("vnode_pager_generic_getpages: page %p is dirty", 996 mt)); 997 KASSERT(!pmap_page_is_mapped(mt), 998 ("vnode_pager_generic_getpages: page %p is mapped", 999 mt)); 1000 } else { 1001 /* 1002 * Read did not fill up entire page. 1003 * 1004 * Currently we do not set the entire page valid, 1005 * we just try to clear the piece that we couldn't 1006 * read. 1007 */ 1008 vm_page_set_valid(mt, 0, 1009 object->un_pager.vnp.vnp_size - tfoff); 1010 KASSERT((mt->dirty & vm_page_bits(0, 1011 object->un_pager.vnp.vnp_size - tfoff)) == 0, 1012 ("vnode_pager_generic_getpages: page %p is dirty", 1013 mt)); 1014 } 1015 1016 if (i != reqpage) 1017 vm_page_readahead_finish(mt); 1018 } 1019 VM_OBJECT_UNLOCK(object); 1020 if (error) { 1021 printf("vnode_pager_getpages: I/O read error\n"); 1022 } 1023 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 1024} 1025 1026/* 1027 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 1028 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 1029 * vnode_pager_generic_putpages() to implement the previous behaviour. 1030 * 1031 * All other FS's should use the bypass to get to the local media 1032 * backing vp's VOP_PUTPAGES. 1033 */ 1034static void 1035vnode_pager_putpages(object, m, count, sync, rtvals) 1036 vm_object_t object; 1037 vm_page_t *m; 1038 int count; 1039 boolean_t sync; 1040 int *rtvals; 1041{ 1042 int rtval; 1043 struct vnode *vp; 1044 int bytes = count * PAGE_SIZE; 1045 1046 /* 1047 * Force synchronous operation if we are extremely low on memory 1048 * to prevent a low-memory deadlock. VOP operations often need to 1049 * allocate more memory to initiate the I/O ( i.e. do a BMAP 1050 * operation ). The swapper handles the case by limiting the amount 1051 * of asynchronous I/O, but that sort of solution doesn't scale well 1052 * for the vnode pager without a lot of work. 1053 * 1054 * Also, the backing vnode's iodone routine may not wake the pageout 1055 * daemon up. This should be probably be addressed XXX. 1056 */ 1057 1058 if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min) 1059 sync |= OBJPC_SYNC; 1060 1061 /* 1062 * Call device-specific putpages function 1063 */ 1064 vp = object->handle; 1065 VM_OBJECT_UNLOCK(object); 1066 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 1067 KASSERT(rtval != EOPNOTSUPP, 1068 ("vnode_pager: stale FS putpages\n")); 1069 VM_OBJECT_LOCK(object); 1070} 1071 1072 1073/* 1074 * This is now called from local media FS's to operate against their 1075 * own vnodes if they fail to implement VOP_PUTPAGES. 1076 * 1077 * This is typically called indirectly via the pageout daemon and 1078 * clustering has already typically occured, so in general we ask the 1079 * underlying filesystem to write the data out asynchronously rather 1080 * then delayed. 1081 */ 1082int 1083vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount, 1084 int flags, int *rtvals) 1085{ 1086 int i; 1087 vm_object_t object; 1088 vm_page_t m; 1089 int count; 1090 1091 int maxsize, ncount; 1092 vm_ooffset_t poffset; 1093 struct uio auio; 1094 struct iovec aiov; 1095 int error; 1096 int ioflags; 1097 int ppscheck = 0; 1098 static struct timeval lastfail; 1099 static int curfail; 1100 1101 object = vp->v_object; 1102 count = bytecount / PAGE_SIZE; 1103 1104 for (i = 0; i < count; i++) 1105 rtvals[i] = VM_PAGER_ERROR; 1106 1107 if ((int64_t)ma[0]->pindex < 0) { 1108 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n", 1109 (long)ma[0]->pindex, (u_long)ma[0]->dirty); 1110 rtvals[0] = VM_PAGER_BAD; 1111 return VM_PAGER_BAD; 1112 } 1113 1114 maxsize = count * PAGE_SIZE; 1115 ncount = count; 1116 1117 poffset = IDX_TO_OFF(ma[0]->pindex); 1118 1119 /* 1120 * If the page-aligned write is larger then the actual file we 1121 * have to invalidate pages occuring beyond the file EOF. However, 1122 * there is an edge case where a file may not be page-aligned where 1123 * the last page is partially invalid. In this case the filesystem 1124 * may not properly clear the dirty bits for the entire page (which 1125 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 1126 * With the page locked we are free to fix-up the dirty bits here. 1127 * 1128 * We do not under any circumstances truncate the valid bits, as 1129 * this will screw up bogus page replacement. 1130 */ 1131 VM_OBJECT_LOCK(object); 1132 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 1133 if (object->un_pager.vnp.vnp_size > poffset) { 1134 int pgoff; 1135 1136 maxsize = object->un_pager.vnp.vnp_size - poffset; 1137 ncount = btoc(maxsize); 1138 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 1139 /* 1140 * If the object is locked and the following 1141 * conditions hold, then the page's dirty 1142 * field cannot be concurrently changed by a 1143 * pmap operation. 1144 */ 1145 m = ma[ncount - 1]; 1146 KASSERT(m->busy > 0, 1147 ("vnode_pager_generic_putpages: page %p is not busy", m)); 1148 KASSERT(!pmap_page_is_write_mapped(m), 1149 ("vnode_pager_generic_putpages: page %p is not read-only", m)); 1150 vm_page_clear_dirty(m, pgoff, PAGE_SIZE - 1151 pgoff); 1152 } 1153 } else { 1154 maxsize = 0; 1155 ncount = 0; 1156 } 1157 if (ncount < count) { 1158 for (i = ncount; i < count; i++) { 1159 rtvals[i] = VM_PAGER_BAD; 1160 } 1161 } 1162 } 1163 VM_OBJECT_UNLOCK(object); 1164 1165 /* 1166 * pageouts are already clustered, use IO_ASYNC t o force a bawrite() 1167 * rather then a bdwrite() to prevent paging I/O from saturating 1168 * the buffer cache. Dummy-up the sequential heuristic to cause 1169 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 1170 * the system decides how to cluster. 1171 */ 1172 ioflags = IO_VMIO; 1173 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 1174 ioflags |= IO_SYNC; 1175 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 1176 ioflags |= IO_ASYNC; 1177 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 1178 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 1179 1180 aiov.iov_base = (caddr_t) 0; 1181 aiov.iov_len = maxsize; 1182 auio.uio_iov = &aiov; 1183 auio.uio_iovcnt = 1; 1184 auio.uio_offset = poffset; 1185 auio.uio_segflg = UIO_NOCOPY; 1186 auio.uio_rw = UIO_WRITE; 1187 auio.uio_resid = maxsize; 1188 auio.uio_td = (struct thread *) 0; 1189 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred); 1190 PCPU_INC(cnt.v_vnodeout); 1191 PCPU_ADD(cnt.v_vnodepgsout, ncount); 1192 1193 if (error) { 1194 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1))) 1195 printf("vnode_pager_putpages: I/O error %d\n", error); 1196 } 1197 if (auio.uio_resid) { 1198 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1)) 1199 printf("vnode_pager_putpages: residual I/O %zd at %lu\n", 1200 auio.uio_resid, (u_long)ma[0]->pindex); 1201 } 1202 for (i = 0; i < ncount; i++) { 1203 rtvals[i] = VM_PAGER_OK; 1204 } 1205 return rtvals[0]; 1206} 1207 1208void 1209vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written) 1210{ 1211 vm_object_t obj; 1212 int i, pos; 1213 1214 if (written == 0) 1215 return; 1216 obj = ma[0]->object; 1217 VM_OBJECT_LOCK(obj); 1218 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) { 1219 if (pos < trunc_page(written)) { 1220 rtvals[i] = VM_PAGER_OK; 1221 vm_page_undirty(ma[i]); 1222 } else { 1223 /* Partially written page. */ 1224 rtvals[i] = VM_PAGER_AGAIN; 1225 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK); 1226 } 1227 } 1228 VM_OBJECT_UNLOCK(obj); 1229} 1230 1231void 1232vnode_pager_update_writecount(vm_object_t object, vm_offset_t start, 1233 vm_offset_t end) 1234{ 1235 struct vnode *vp; 1236 vm_ooffset_t old_wm; 1237 1238 VM_OBJECT_LOCK(object); 1239 if (object->type != OBJT_VNODE) { 1240 VM_OBJECT_UNLOCK(object); 1241 return; 1242 } 1243 old_wm = object->un_pager.vnp.writemappings; 1244 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start; 1245 vp = object->handle; 1246 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) { 1247 ASSERT_VOP_ELOCKED(vp, "v_writecount inc"); 1248 VOP_ADD_WRITECOUNT(vp, 1); 1249 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) { 1250 ASSERT_VOP_ELOCKED(vp, "v_writecount dec"); 1251 VOP_ADD_WRITECOUNT(vp, -1); 1252 } 1253 VM_OBJECT_UNLOCK(object); 1254} 1255 1256void 1257vnode_pager_release_writecount(vm_object_t object, vm_offset_t start, 1258 vm_offset_t end) 1259{ 1260 struct vnode *vp; 1261 struct mount *mp; 1262 vm_offset_t inc; 1263 int vfslocked; 1264 1265 VM_OBJECT_LOCK(object); 1266 1267 /* 1268 * First, recheck the object type to account for the race when 1269 * the vnode is reclaimed. 1270 */ 1271 if (object->type != OBJT_VNODE) { 1272 VM_OBJECT_UNLOCK(object); 1273 return; 1274 } 1275 1276 /* 1277 * Optimize for the case when writemappings is not going to 1278 * zero. 1279 */ 1280 inc = end - start; 1281 if (object->un_pager.vnp.writemappings != inc) { 1282 object->un_pager.vnp.writemappings -= inc; 1283 VM_OBJECT_UNLOCK(object); 1284 return; 1285 } 1286 1287 vp = object->handle; 1288 vhold(vp); 1289 VM_OBJECT_UNLOCK(object); 1290 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 1291 mp = NULL; 1292 vn_start_write(vp, &mp, V_WAIT); 1293 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1294 1295 /* 1296 * Decrement the object's writemappings, by swapping the start 1297 * and end arguments for vnode_pager_update_writecount(). If 1298 * there was not a race with vnode reclaimation, then the 1299 * vnode's v_writecount is decremented. 1300 */ 1301 vnode_pager_update_writecount(object, end, start); 1302 VOP_UNLOCK(vp, 0); 1303 vdrop(vp); 1304 if (mp != NULL) 1305 vn_finished_write(mp); 1306 VFS_UNLOCK_GIANT(vfslocked); 1307} 1308