vnode_pager.c revision 116279
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: head/sys/vm/vnode_pager.c 116279 2003-06-13 03:02:28Z alc $"); 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/conf.h> 65 66#include <vm/vm.h> 67#include <vm/vm_object.h> 68#include <vm/vm_page.h> 69#include <vm/vm_pager.h> 70#include <vm/vm_map.h> 71#include <vm/vnode_pager.h> 72#include <vm/vm_extern.h> 73 74static void vnode_pager_init(void); 75static vm_offset_t vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, 76 int *run); 77static void vnode_pager_iodone(struct buf *bp); 78static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m); 79static int vnode_pager_input_old(vm_object_t object, vm_page_t m); 80static void vnode_pager_dealloc(vm_object_t); 81static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int); 82static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 83static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *); 84 85struct pagerops vnodepagerops = { 86 vnode_pager_init, 87 vnode_pager_alloc, 88 vnode_pager_dealloc, 89 vnode_pager_getpages, 90 vnode_pager_putpages, 91 vnode_pager_haspage, 92 NULL 93}; 94 95int vnode_pbuf_freecnt; 96 97static void 98vnode_pager_init(void) 99{ 100 101 vnode_pbuf_freecnt = nswbuf / 2 + 1; 102} 103 104/* 105 * Allocate (or lookup) pager for a vnode. 106 * Handle is a vnode pointer. 107 * 108 * MPSAFE 109 */ 110vm_object_t 111vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 112 vm_ooffset_t offset) 113{ 114 vm_object_t object; 115 struct vnode *vp; 116 117 /* 118 * Pageout to vnode, no can do yet. 119 */ 120 if (handle == NULL) 121 return (NULL); 122 123 vp = (struct vnode *) handle; 124 125 ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc"); 126 127 mtx_lock(&Giant); 128 /* 129 * Prevent race condition when allocating the object. This 130 * can happen with NFS vnodes since the nfsnode isn't locked. 131 */ 132 VI_LOCK(vp); 133 while (vp->v_iflag & VI_OLOCK) { 134 vp->v_iflag |= VI_OWANT; 135 msleep(vp, VI_MTX(vp), PVM, "vnpobj", 0); 136 } 137 vp->v_iflag |= VI_OLOCK; 138 VI_UNLOCK(vp); 139 140 /* 141 * If the object is being terminated, wait for it to 142 * go away. 143 */ 144 while ((object = vp->v_object) != NULL) { 145 VM_OBJECT_LOCK(object); 146 if ((object->flags & OBJ_DEAD) == 0) 147 break; 148 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vadead", 0); 149 } 150 151 if (vp->v_usecount == 0) 152 panic("vnode_pager_alloc: no vnode reference"); 153 154 if (object == NULL) { 155 /* 156 * And an object of the appropriate size 157 */ 158 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); 159 160 object->un_pager.vnp.vnp_size = size; 161 162 object->handle = handle; 163 vp->v_object = object; 164 } else { 165 object->ref_count++; 166 VM_OBJECT_UNLOCK(object); 167 } 168 VI_LOCK(vp); 169 vp->v_usecount++; 170 vp->v_iflag &= ~VI_OLOCK; 171 if (vp->v_iflag & VI_OWANT) { 172 vp->v_iflag &= ~VI_OWANT; 173 wakeup(vp); 174 } 175 VI_UNLOCK(vp); 176 mtx_unlock(&Giant); 177 return (object); 178} 179 180/* 181 * The object must be locked. 182 */ 183static void 184vnode_pager_dealloc(object) 185 vm_object_t object; 186{ 187 struct vnode *vp = object->handle; 188 189 if (vp == NULL) 190 panic("vnode_pager_dealloc: pager already dealloced"); 191 192 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 193 vm_object_pip_wait(object, "vnpdea"); 194 195 object->handle = NULL; 196 object->type = OBJT_DEAD; 197 ASSERT_VOP_LOCKED(vp, "vnode_pager_dealloc"); 198 vp->v_object = NULL; 199 vp->v_vflag &= ~(VV_TEXT | VV_OBJBUF); 200} 201 202static boolean_t 203vnode_pager_haspage(object, pindex, before, after) 204 vm_object_t object; 205 vm_pindex_t pindex; 206 int *before; 207 int *after; 208{ 209 struct vnode *vp = object->handle; 210 daddr_t bn; 211 int err; 212 daddr_t reqblock; 213 int poff; 214 int bsize; 215 int pagesperblock, blocksperpage; 216 217 GIANT_REQUIRED; 218 /* 219 * If no vp or vp is doomed or marked transparent to VM, we do not 220 * have the page. 221 */ 222 if (vp == NULL) 223 return FALSE; 224 225 VI_LOCK(vp); 226 if (vp->v_iflag & VI_DOOMED) { 227 VI_UNLOCK(vp); 228 return FALSE; 229 } 230 VI_UNLOCK(vp); 231 /* 232 * If filesystem no longer mounted or offset beyond end of file we do 233 * not have the page. 234 */ 235 if ((vp->v_mount == NULL) || 236 (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)) 237 return FALSE; 238 239 bsize = vp->v_mount->mnt_stat.f_iosize; 240 pagesperblock = bsize / PAGE_SIZE; 241 blocksperpage = 0; 242 if (pagesperblock > 0) { 243 reqblock = pindex / pagesperblock; 244 } else { 245 blocksperpage = (PAGE_SIZE / bsize); 246 reqblock = pindex * blocksperpage; 247 } 248 err = VOP_BMAP(vp, reqblock, (struct vnode **) 0, &bn, 249 after, before); 250 if (err) 251 return TRUE; 252 if (bn == -1) 253 return FALSE; 254 if (pagesperblock > 0) { 255 poff = pindex - (reqblock * pagesperblock); 256 if (before) { 257 *before *= pagesperblock; 258 *before += poff; 259 } 260 if (after) { 261 int numafter; 262 *after *= pagesperblock; 263 numafter = pagesperblock - (poff + 1); 264 if (IDX_TO_OFF(pindex + numafter) > 265 object->un_pager.vnp.vnp_size) { 266 numafter = 267 OFF_TO_IDX(object->un_pager.vnp.vnp_size) - 268 pindex; 269 } 270 *after += numafter; 271 } 272 } else { 273 if (before) { 274 *before /= blocksperpage; 275 } 276 277 if (after) { 278 *after /= blocksperpage; 279 } 280 } 281 return TRUE; 282} 283 284/* 285 * Lets the VM system know about a change in size for a file. 286 * We adjust our own internal size and flush any cached pages in 287 * the associated object that are affected by the size change. 288 * 289 * Note: this routine may be invoked as a result of a pager put 290 * operation (possibly at object termination time), so we must be careful. 291 */ 292void 293vnode_pager_setsize(vp, nsize) 294 struct vnode *vp; 295 vm_ooffset_t nsize; 296{ 297 vm_object_t object; 298 vm_page_t m; 299 vm_pindex_t nobjsize; 300 301 if ((object = vp->v_object) == NULL) 302 return; 303 VM_OBJECT_LOCK(object); 304 if (nsize == object->un_pager.vnp.vnp_size) { 305 /* 306 * Hasn't changed size 307 */ 308 VM_OBJECT_UNLOCK(object); 309 return; 310 } 311 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 312 if (nsize < object->un_pager.vnp.vnp_size) { 313 /* 314 * File has shrunk. Toss any cached pages beyond the new EOF. 315 */ 316 if (nobjsize < object->size) 317 vm_object_page_remove(object, nobjsize, object->size, 318 FALSE); 319 /* 320 * this gets rid of garbage at the end of a page that is now 321 * only partially backed by the vnode. 322 * 323 * XXX for some reason (I don't know yet), if we take a 324 * completely invalid page and mark it partially valid 325 * it can screw up NFS reads, so we don't allow the case. 326 */ 327 if ((nsize & PAGE_MASK) && 328 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL) { 329 vm_page_lock_queues(); 330 if (m->valid) { 331 int base = (int)nsize & PAGE_MASK; 332 int size = PAGE_SIZE - base; 333 334 /* 335 * Clear out partial-page garbage in case 336 * the page has been mapped. 337 */ 338 pmap_zero_page_area(m, base, size); 339 340 /* 341 * XXX work around SMP data integrity race 342 * by unmapping the page from user processes. 343 * The garbage we just cleared may be mapped 344 * to a user process running on another cpu 345 * and this code is not running through normal 346 * I/O channels which handle SMP issues for 347 * us, so unmap page to synchronize all cpus. 348 * 349 * XXX should vm_pager_unmap_page() have 350 * dealt with this? 351 */ 352 pmap_remove_all(m); 353 354 /* 355 * Clear out partial-page dirty bits. This 356 * has the side effect of setting the valid 357 * bits, but that is ok. There are a bunch 358 * of places in the VM system where we expected 359 * m->dirty == VM_PAGE_BITS_ALL. The file EOF 360 * case is one of them. If the page is still 361 * partially dirty, make it fully dirty. 362 * 363 * note that we do not clear out the valid 364 * bits. This would prevent bogus_page 365 * replacement from working properly. 366 */ 367 vm_page_set_validclean(m, base, size); 368 if (m->dirty != 0) 369 m->dirty = VM_PAGE_BITS_ALL; 370 } 371 vm_page_unlock_queues(); 372 } 373 } 374 object->un_pager.vnp.vnp_size = nsize; 375 object->size = nobjsize; 376 VM_OBJECT_UNLOCK(object); 377} 378 379/* 380 * calculate the linear (byte) disk address of specified virtual 381 * file address 382 */ 383static vm_offset_t 384vnode_pager_addr(vp, address, run) 385 struct vnode *vp; 386 vm_ooffset_t address; 387 int *run; 388{ 389 int rtaddress; 390 int bsize; 391 daddr_t block; 392 struct vnode *rtvp; 393 int err; 394 daddr_t vblock; 395 int voffset; 396 397 GIANT_REQUIRED; 398 if ((int) address < 0) 399 return -1; 400 401 if (vp->v_mount == NULL) 402 return -1; 403 404 bsize = vp->v_mount->mnt_stat.f_iosize; 405 vblock = address / bsize; 406 voffset = address % bsize; 407 408 err = VOP_BMAP(vp, vblock, &rtvp, &block, run, NULL); 409 410 if (err || (block == -1)) 411 rtaddress = -1; 412 else { 413 rtaddress = block + voffset / DEV_BSIZE; 414 if (run) { 415 *run += 1; 416 *run *= bsize/PAGE_SIZE; 417 *run -= voffset/PAGE_SIZE; 418 } 419 } 420 421 return rtaddress; 422} 423 424/* 425 * interrupt routine for I/O completion 426 */ 427static void 428vnode_pager_iodone(bp) 429 struct buf *bp; 430{ 431 bp->b_flags |= B_DONE; 432 wakeup(bp); 433} 434 435/* 436 * small block filesystem vnode pager input 437 */ 438static int 439vnode_pager_input_smlfs(object, m) 440 vm_object_t object; 441 vm_page_t m; 442{ 443 int i; 444 int s; 445 struct vnode *dp, *vp; 446 struct buf *bp; 447 vm_offset_t kva; 448 int fileaddr; 449 vm_offset_t bsize; 450 int error = 0; 451 452 GIANT_REQUIRED; 453 454 vp = object->handle; 455 if (vp->v_mount == NULL) 456 return VM_PAGER_BAD; 457 458 bsize = vp->v_mount->mnt_stat.f_iosize; 459 460 VOP_BMAP(vp, 0, &dp, 0, NULL, NULL); 461 462 kva = vm_pager_map_page(m); 463 464 for (i = 0; i < PAGE_SIZE / bsize; i++) { 465 vm_ooffset_t address; 466 467 if (vm_page_bits(i * bsize, bsize) & m->valid) 468 continue; 469 470 address = IDX_TO_OFF(m->pindex) + i * bsize; 471 if (address >= object->un_pager.vnp.vnp_size) { 472 fileaddr = -1; 473 } else { 474 fileaddr = vnode_pager_addr(vp, address, NULL); 475 } 476 if (fileaddr != -1) { 477 bp = getpbuf(&vnode_pbuf_freecnt); 478 479 /* build a minimal buffer header */ 480 bp->b_iocmd = BIO_READ; 481 bp->b_iodone = vnode_pager_iodone; 482 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 483 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 484 bp->b_rcred = crhold(curthread->td_ucred); 485 bp->b_wcred = crhold(curthread->td_ucred); 486 bp->b_data = (caddr_t) kva + i * bsize; 487 bp->b_blkno = fileaddr; 488 pbgetvp(dp, bp); 489 bp->b_bcount = bsize; 490 bp->b_bufsize = bsize; 491 bp->b_runningbufspace = bp->b_bufsize; 492 runningbufspace += bp->b_runningbufspace; 493 494 /* do the input */ 495 VOP_SPECSTRATEGY(bp->b_vp, bp); 496 497 /* we definitely need to be at splvm here */ 498 499 s = splvm(); 500 while ((bp->b_flags & B_DONE) == 0) { 501 tsleep(bp, PVM, "vnsrd", 0); 502 } 503 splx(s); 504 if ((bp->b_ioflags & BIO_ERROR) != 0) 505 error = EIO; 506 507 /* 508 * free the buffer header back to the swap buffer pool 509 */ 510 relpbuf(bp, &vnode_pbuf_freecnt); 511 if (error) 512 break; 513 514 vm_page_lock_queues(); 515 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 516 vm_page_unlock_queues(); 517 } else { 518 vm_page_lock_queues(); 519 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 520 vm_page_unlock_queues(); 521 bzero((caddr_t) kva + i * bsize, bsize); 522 } 523 } 524 vm_pager_unmap_page(kva); 525 vm_page_lock_queues(); 526 pmap_clear_modify(m); 527 vm_page_flag_clear(m, PG_ZERO); 528 vm_page_unlock_queues(); 529 if (error) { 530 return VM_PAGER_ERROR; 531 } 532 return VM_PAGER_OK; 533 534} 535 536 537/* 538 * old style vnode pager output routine 539 */ 540static int 541vnode_pager_input_old(object, m) 542 vm_object_t object; 543 vm_page_t m; 544{ 545 struct uio auio; 546 struct iovec aiov; 547 int error; 548 int size; 549 vm_offset_t kva; 550 struct vnode *vp; 551 552 GIANT_REQUIRED; 553 error = 0; 554 555 /* 556 * Return failure if beyond current EOF 557 */ 558 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 559 return VM_PAGER_BAD; 560 } else { 561 size = PAGE_SIZE; 562 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 563 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 564 565 /* 566 * Allocate a kernel virtual address and initialize so that 567 * we can use VOP_READ/WRITE routines. 568 */ 569 kva = vm_pager_map_page(m); 570 571 vp = object->handle; 572 aiov.iov_base = (caddr_t) kva; 573 aiov.iov_len = size; 574 auio.uio_iov = &aiov; 575 auio.uio_iovcnt = 1; 576 auio.uio_offset = IDX_TO_OFF(m->pindex); 577 auio.uio_segflg = UIO_SYSSPACE; 578 auio.uio_rw = UIO_READ; 579 auio.uio_resid = size; 580 auio.uio_td = curthread; 581 582 error = VOP_READ(vp, &auio, 0, curthread->td_ucred); 583 if (!error) { 584 int count = size - auio.uio_resid; 585 586 if (count == 0) 587 error = EINVAL; 588 else if (count != PAGE_SIZE) 589 bzero((caddr_t) kva + count, PAGE_SIZE - count); 590 } 591 vm_pager_unmap_page(kva); 592 } 593 vm_page_lock_queues(); 594 pmap_clear_modify(m); 595 vm_page_undirty(m); 596 vm_page_flag_clear(m, PG_ZERO); 597 if (!error) 598 m->valid = VM_PAGE_BITS_ALL; 599 vm_page_unlock_queues(); 600 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 601} 602 603/* 604 * generic vnode pager input routine 605 */ 606 607/* 608 * Local media VFS's that do not implement their own VOP_GETPAGES 609 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages() 610 * to implement the previous behaviour. 611 * 612 * All other FS's should use the bypass to get to the local media 613 * backing vp's VOP_GETPAGES. 614 */ 615static int 616vnode_pager_getpages(object, m, count, reqpage) 617 vm_object_t object; 618 vm_page_t *m; 619 int count; 620 int reqpage; 621{ 622 int rtval; 623 struct vnode *vp; 624 int bytes = count * PAGE_SIZE; 625 626 vp = object->handle; 627 VM_OBJECT_UNLOCK(object); 628 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0); 629 KASSERT(rtval != EOPNOTSUPP, 630 ("vnode_pager: FS getpages not implemented\n")); 631 VM_OBJECT_LOCK(object); 632 return rtval; 633} 634 635/* 636 * This is now called from local media FS's to operate against their 637 * own vnodes if they fail to implement VOP_GETPAGES. 638 */ 639int 640vnode_pager_generic_getpages(vp, m, bytecount, reqpage) 641 struct vnode *vp; 642 vm_page_t *m; 643 int bytecount; 644 int reqpage; 645{ 646 vm_object_t object; 647 vm_offset_t kva; 648 off_t foff, tfoff, nextoff; 649 int i, j, size, bsize, first, firstaddr; 650 struct vnode *dp; 651 int runpg; 652 int runend; 653 struct buf *bp; 654 int s; 655 int count; 656 int error = 0; 657 658 GIANT_REQUIRED; 659 object = vp->v_object; 660 count = bytecount / PAGE_SIZE; 661 662 if (vp->v_mount == NULL) 663 return VM_PAGER_BAD; 664 665 bsize = vp->v_mount->mnt_stat.f_iosize; 666 667 /* get the UNDERLYING device for the file with VOP_BMAP() */ 668 669 /* 670 * originally, we did not check for an error return value -- assuming 671 * an fs always has a bmap entry point -- that assumption is wrong!!! 672 */ 673 foff = IDX_TO_OFF(m[reqpage]->pindex); 674 675 /* 676 * if we can't bmap, use old VOP code 677 */ 678 if (VOP_BMAP(vp, 0, &dp, 0, NULL, NULL)) { 679 vm_page_lock_queues(); 680 for (i = 0; i < count; i++) 681 if (i != reqpage) 682 vm_page_free(m[i]); 683 vm_page_unlock_queues(); 684 cnt.v_vnodein++; 685 cnt.v_vnodepgsin++; 686 return vnode_pager_input_old(object, m[reqpage]); 687 688 /* 689 * if the blocksize is smaller than a page size, then use 690 * special small filesystem code. NFS sometimes has a small 691 * blocksize, but it can handle large reads itself. 692 */ 693 } else if ((PAGE_SIZE / bsize) > 1 && 694 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) { 695 vm_page_lock_queues(); 696 for (i = 0; i < count; i++) 697 if (i != reqpage) 698 vm_page_free(m[i]); 699 vm_page_unlock_queues(); 700 cnt.v_vnodein++; 701 cnt.v_vnodepgsin++; 702 return vnode_pager_input_smlfs(object, m[reqpage]); 703 } 704 705 /* 706 * If we have a completely valid page available to us, we can 707 * clean up and return. Otherwise we have to re-read the 708 * media. 709 */ 710 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) { 711 vm_page_lock_queues(); 712 for (i = 0; i < count; i++) 713 if (i != reqpage) 714 vm_page_free(m[i]); 715 vm_page_unlock_queues(); 716 return VM_PAGER_OK; 717 } 718 m[reqpage]->valid = 0; 719 720 /* 721 * here on direct device I/O 722 */ 723 firstaddr = -1; 724 725 /* 726 * calculate the run that includes the required page 727 */ 728 for (first = 0, i = 0; i < count; i = runend) { 729 firstaddr = vnode_pager_addr(vp, 730 IDX_TO_OFF(m[i]->pindex), &runpg); 731 if (firstaddr == -1) { 732 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) { 733 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %d, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx", 734 firstaddr, (uintmax_t)(foff >> 32), 735 (uintmax_t)foff, 736 (uintmax_t) 737 (object->un_pager.vnp.vnp_size >> 32), 738 (uintmax_t)object->un_pager.vnp.vnp_size); 739 } 740 vm_page_lock_queues(); 741 vm_page_free(m[i]); 742 vm_page_unlock_queues(); 743 runend = i + 1; 744 first = runend; 745 continue; 746 } 747 runend = i + runpg; 748 if (runend <= reqpage) { 749 vm_page_lock_queues(); 750 for (j = i; j < runend; j++) 751 vm_page_free(m[j]); 752 vm_page_unlock_queues(); 753 } else { 754 if (runpg < (count - first)) { 755 vm_page_lock_queues(); 756 for (i = first + runpg; i < count; i++) 757 vm_page_free(m[i]); 758 vm_page_unlock_queues(); 759 count = first + runpg; 760 } 761 break; 762 } 763 first = runend; 764 } 765 766 /* 767 * the first and last page have been calculated now, move input pages 768 * to be zero based... 769 */ 770 if (first != 0) { 771 for (i = first; i < count; i++) { 772 m[i - first] = m[i]; 773 } 774 count -= first; 775 reqpage -= first; 776 } 777 778 /* 779 * calculate the file virtual address for the transfer 780 */ 781 foff = IDX_TO_OFF(m[0]->pindex); 782 783 /* 784 * calculate the size of the transfer 785 */ 786 size = count * PAGE_SIZE; 787 if ((foff + size) > object->un_pager.vnp.vnp_size) 788 size = object->un_pager.vnp.vnp_size - foff; 789 790 /* 791 * round up physical size for real devices. 792 */ 793 if (dp->v_type == VBLK || dp->v_type == VCHR) { 794 int secmask = dp->v_rdev->si_bsize_phys - 1; 795 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1)); 796 size = (size + secmask) & ~secmask; 797 } 798 799 bp = getpbuf(&vnode_pbuf_freecnt); 800 kva = (vm_offset_t) bp->b_data; 801 802 /* 803 * and map the pages to be read into the kva 804 */ 805 pmap_qenter(kva, m, count); 806 807 /* build a minimal buffer header */ 808 bp->b_iocmd = BIO_READ; 809 bp->b_iodone = vnode_pager_iodone; 810 /* B_PHYS is not set, but it is nice to fill this in */ 811 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 812 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 813 bp->b_rcred = crhold(curthread->td_ucred); 814 bp->b_wcred = crhold(curthread->td_ucred); 815 bp->b_blkno = firstaddr; 816 pbgetvp(dp, bp); 817 bp->b_bcount = size; 818 bp->b_bufsize = size; 819 bp->b_runningbufspace = bp->b_bufsize; 820 runningbufspace += bp->b_runningbufspace; 821 822 cnt.v_vnodein++; 823 cnt.v_vnodepgsin += count; 824 825 /* do the input */ 826 if (dp->v_type == VCHR) 827 VOP_SPECSTRATEGY(bp->b_vp, bp); 828 else 829 VOP_STRATEGY(bp->b_vp, bp); 830 831 s = splvm(); 832 /* we definitely need to be at splvm here */ 833 834 while ((bp->b_flags & B_DONE) == 0) { 835 tsleep(bp, PVM, "vnread", 0); 836 } 837 splx(s); 838 if ((bp->b_ioflags & BIO_ERROR) != 0) 839 error = EIO; 840 841 if (!error) { 842 if (size != count * PAGE_SIZE) 843 bzero((caddr_t) kva + size, PAGE_SIZE * count - size); 844 } 845 pmap_qremove(kva, count); 846 847 /* 848 * free the buffer header back to the swap buffer pool 849 */ 850 relpbuf(bp, &vnode_pbuf_freecnt); 851 852 vm_page_lock_queues(); 853 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) { 854 vm_page_t mt; 855 856 nextoff = tfoff + PAGE_SIZE; 857 mt = m[i]; 858 859 if (nextoff <= object->un_pager.vnp.vnp_size) { 860 /* 861 * Read filled up entire page. 862 */ 863 mt->valid = VM_PAGE_BITS_ALL; 864 vm_page_undirty(mt); /* should be an assert? XXX */ 865 pmap_clear_modify(mt); 866 } else { 867 /* 868 * Read did not fill up entire page. Since this 869 * is getpages, the page may be mapped, so we have 870 * to zero the invalid portions of the page even 871 * though we aren't setting them valid. 872 * 873 * Currently we do not set the entire page valid, 874 * we just try to clear the piece that we couldn't 875 * read. 876 */ 877 vm_page_set_validclean(mt, 0, 878 object->un_pager.vnp.vnp_size - tfoff); 879 /* handled by vm_fault now */ 880 /* vm_page_zero_invalid(mt, FALSE); */ 881 } 882 883 vm_page_flag_clear(mt, PG_ZERO); 884 if (i != reqpage) { 885 886 /* 887 * whether or not to leave the page activated is up in 888 * the air, but we should put the page on a page queue 889 * somewhere. (it already is in the object). Result: 890 * It appears that empirical results show that 891 * deactivating pages is best. 892 */ 893 894 /* 895 * just in case someone was asking for this page we 896 * now tell them that it is ok to use 897 */ 898 if (!error) { 899 if (mt->flags & PG_WANTED) 900 vm_page_activate(mt); 901 else 902 vm_page_deactivate(mt); 903 vm_page_wakeup(mt); 904 } else { 905 vm_page_free(mt); 906 } 907 } 908 } 909 vm_page_unlock_queues(); 910 if (error) { 911 printf("vnode_pager_getpages: I/O read error\n"); 912 } 913 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 914} 915 916/* 917 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 918 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 919 * vnode_pager_generic_putpages() to implement the previous behaviour. 920 * 921 * All other FS's should use the bypass to get to the local media 922 * backing vp's VOP_PUTPAGES. 923 */ 924static void 925vnode_pager_putpages(object, m, count, sync, rtvals) 926 vm_object_t object; 927 vm_page_t *m; 928 int count; 929 boolean_t sync; 930 int *rtvals; 931{ 932 int rtval; 933 struct vnode *vp; 934 struct mount *mp; 935 int bytes = count * PAGE_SIZE; 936 937 GIANT_REQUIRED; 938 /* 939 * Force synchronous operation if we are extremely low on memory 940 * to prevent a low-memory deadlock. VOP operations often need to 941 * allocate more memory to initiate the I/O ( i.e. do a BMAP 942 * operation ). The swapper handles the case by limiting the amount 943 * of asynchronous I/O, but that sort of solution doesn't scale well 944 * for the vnode pager without a lot of work. 945 * 946 * Also, the backing vnode's iodone routine may not wake the pageout 947 * daemon up. This should be probably be addressed XXX. 948 */ 949 950 if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min) 951 sync |= OBJPC_SYNC; 952 953 /* 954 * Call device-specific putpages function 955 */ 956 vp = object->handle; 957 if (vp->v_type != VREG) 958 mp = NULL; 959 (void)vn_start_write(vp, &mp, V_WAIT); 960 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 961 KASSERT(rtval != EOPNOTSUPP, 962 ("vnode_pager: stale FS putpages\n")); 963 vn_finished_write(mp); 964} 965 966 967/* 968 * This is now called from local media FS's to operate against their 969 * own vnodes if they fail to implement VOP_PUTPAGES. 970 * 971 * This is typically called indirectly via the pageout daemon and 972 * clustering has already typically occured, so in general we ask the 973 * underlying filesystem to write the data out asynchronously rather 974 * then delayed. 975 */ 976int 977vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals) 978 struct vnode *vp; 979 vm_page_t *m; 980 int bytecount; 981 int flags; 982 int *rtvals; 983{ 984 int i; 985 vm_object_t object; 986 int count; 987 988 int maxsize, ncount; 989 vm_ooffset_t poffset; 990 struct uio auio; 991 struct iovec aiov; 992 int error; 993 int ioflags; 994 995 GIANT_REQUIRED; 996 object = vp->v_object; 997 count = bytecount / PAGE_SIZE; 998 999 for (i = 0; i < count; i++) 1000 rtvals[i] = VM_PAGER_AGAIN; 1001 1002 if ((int) m[0]->pindex < 0) { 1003 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n", 1004 (long)m[0]->pindex, m[0]->dirty); 1005 rtvals[0] = VM_PAGER_BAD; 1006 return VM_PAGER_BAD; 1007 } 1008 1009 maxsize = count * PAGE_SIZE; 1010 ncount = count; 1011 1012 poffset = IDX_TO_OFF(m[0]->pindex); 1013 1014 /* 1015 * If the page-aligned write is larger then the actual file we 1016 * have to invalidate pages occuring beyond the file EOF. However, 1017 * there is an edge case where a file may not be page-aligned where 1018 * the last page is partially invalid. In this case the filesystem 1019 * may not properly clear the dirty bits for the entire page (which 1020 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 1021 * With the page locked we are free to fix-up the dirty bits here. 1022 * 1023 * We do not under any circumstances truncate the valid bits, as 1024 * this will screw up bogus page replacement. 1025 */ 1026 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 1027 if (object->un_pager.vnp.vnp_size > poffset) { 1028 int pgoff; 1029 1030 maxsize = object->un_pager.vnp.vnp_size - poffset; 1031 ncount = btoc(maxsize); 1032 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 1033 vm_page_clear_dirty(m[ncount - 1], pgoff, 1034 PAGE_SIZE - pgoff); 1035 } 1036 } else { 1037 maxsize = 0; 1038 ncount = 0; 1039 } 1040 if (ncount < count) { 1041 for (i = ncount; i < count; i++) { 1042 rtvals[i] = VM_PAGER_BAD; 1043 } 1044 } 1045 } 1046 1047 /* 1048 * pageouts are already clustered, use IO_ASYNC t o force a bawrite() 1049 * rather then a bdwrite() to prevent paging I/O from saturating 1050 * the buffer cache. Dummy-up the sequential heuristic to cause 1051 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 1052 * the system decides how to cluster. 1053 */ 1054 ioflags = IO_VMIO; 1055 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 1056 ioflags |= IO_SYNC; 1057 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 1058 ioflags |= IO_ASYNC; 1059 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 1060 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 1061 1062 aiov.iov_base = (caddr_t) 0; 1063 aiov.iov_len = maxsize; 1064 auio.uio_iov = &aiov; 1065 auio.uio_iovcnt = 1; 1066 auio.uio_offset = poffset; 1067 auio.uio_segflg = UIO_NOCOPY; 1068 auio.uio_rw = UIO_WRITE; 1069 auio.uio_resid = maxsize; 1070 auio.uio_td = (struct thread *) 0; 1071 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred); 1072 cnt.v_vnodeout++; 1073 cnt.v_vnodepgsout += ncount; 1074 1075 if (error) { 1076 printf("vnode_pager_putpages: I/O error %d\n", error); 1077 } 1078 if (auio.uio_resid) { 1079 printf("vnode_pager_putpages: residual I/O %d at %lu\n", 1080 auio.uio_resid, (u_long)m[0]->pindex); 1081 } 1082 for (i = 0; i < ncount; i++) { 1083 rtvals[i] = VM_PAGER_OK; 1084 } 1085 return rtvals[0]; 1086} 1087 1088struct vnode * 1089vnode_pager_lock(object) 1090 vm_object_t object; 1091{ 1092 struct thread *td = curthread; /* XXX */ 1093 1094 GIANT_REQUIRED; 1095 1096 for (; object != NULL; object = object->backing_object) { 1097 if (object->type != OBJT_VNODE) 1098 continue; 1099 if (object->flags & OBJ_DEAD) { 1100 return NULL; 1101 } 1102 1103 /* XXX; If object->handle can change, we need to cache it. */ 1104 while (vget(object->handle, 1105 LK_NOPAUSE | LK_SHARED | LK_RETRY | LK_CANRECURSE, td)){ 1106 if ((object->flags & OBJ_DEAD) || (object->type != OBJT_VNODE)) 1107 return NULL; 1108 printf("vnode_pager_lock: retrying\n"); 1109 } 1110 return object->handle; 1111 } 1112 return NULL; 1113} 1114