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 * $FreeBSD: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $ 42 * $DragonFly: src/sys/vm/vnode_pager.c,v 1.10 2003/08/20 08:03:01 rob Exp $ 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/param.h> 56 #include <sys/systm.h> 57 #include <sys/proc.h> 58 #include <sys/vnode.h> 59 #include <sys/mount.h> 60 #include <sys/buf.h> 61 #include <sys/vmmeter.h> 62 #include <sys/conf.h> 63 64 #include <vm/vm.h> 65 #include <vm/vm_object.h> 66 #include <vm/vm_page.h> 67 #include <vm/vm_pager.h> 68 #include <vm/vm_map.h> 69 #include <vm/vnode_pager.h> 70 #include <vm/vm_extern.h> 71 72 static vm_offset_t vnode_pager_addr (struct vnode *vp, vm_ooffset_t address, 73 int *run); 74 static void vnode_pager_iodone (struct buf *bp); 75 static int vnode_pager_input_smlfs (vm_object_t object, vm_page_t m); 76 static int vnode_pager_input_old (vm_object_t object, vm_page_t m); 77 static void vnode_pager_dealloc (vm_object_t); 78 static int vnode_pager_getpages (vm_object_t, vm_page_t *, int, int); 79 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *); 80 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t, int *, int *); 81 82 struct pagerops vnodepagerops = { 83 NULL, 84 vnode_pager_alloc, 85 vnode_pager_dealloc, 86 vnode_pager_getpages, 87 vnode_pager_putpages, 88 vnode_pager_haspage, 89 NULL 90 }; 91 92 int vnode_pbuf_freecnt = -1; /* start out unlimited */ 93 94 /* 95 * Allocate (or lookup) pager for a vnode. 96 * Handle is a vnode pointer. 97 */ 98 vm_object_t 99 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 100 vm_ooffset_t offset) 101 { 102 vm_object_t object; 103 struct vnode *vp; 104 105 /* 106 * Pageout to vnode, no can do yet. 107 */ 108 if (handle == NULL) 109 return (NULL); 110 111 /* 112 * XXX hack - This initialization should be put somewhere else. 113 */ 114 if (vnode_pbuf_freecnt < 0) { 115 vnode_pbuf_freecnt = nswbuf / 2 + 1; 116 } 117 118 vp = (struct vnode *) handle; 119 120 /* 121 * Prevent race condition when allocating the object. This 122 * can happen with NFS vnodes since the nfsnode isn't locked. 123 */ 124 while (vp->v_flag & VOLOCK) { 125 vp->v_flag |= VOWANT; 126 tsleep(vp, 0, "vnpobj", 0); 127 } 128 vp->v_flag |= VOLOCK; 129 130 /* 131 * If the object is being terminated, wait for it to 132 * go away. 133 */ 134 while (((object = vp->v_object) != NULL) && 135 (object->flags & OBJ_DEAD)) { 136 tsleep(object, 0, "vadead", 0); 137 } 138 139 if (vp->v_usecount == 0) 140 panic("vnode_pager_alloc: no vnode reference"); 141 142 if (object == NULL) { 143 /* 144 * And an object of the appropriate size 145 */ 146 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); 147 object->flags = 0; 148 149 object->un_pager.vnp.vnp_size = size; 150 151 object->handle = handle; 152 vp->v_object = object; 153 vp->v_usecount++; 154 } else { 155 object->ref_count++; 156 vp->v_usecount++; 157 } 158 159 vp->v_flag &= ~VOLOCK; 160 if (vp->v_flag & VOWANT) { 161 vp->v_flag &= ~VOWANT; 162 wakeup(vp); 163 } 164 return (object); 165 } 166 167 static void 168 vnode_pager_dealloc(object) 169 vm_object_t object; 170 { 171 struct vnode *vp = object->handle; 172 173 if (vp == NULL) 174 panic("vnode_pager_dealloc: pager already dealloced"); 175 176 vm_object_pip_wait(object, "vnpdea"); 177 178 object->handle = NULL; 179 object->type = OBJT_DEAD; 180 vp->v_object = NULL; 181 vp->v_flag &= ~(VTEXT | VOBJBUF); 182 } 183 184 static boolean_t 185 vnode_pager_haspage(object, pindex, before, after) 186 vm_object_t object; 187 vm_pindex_t pindex; 188 int *before; 189 int *after; 190 { 191 struct vnode *vp = object->handle; 192 daddr_t bn; 193 int err; 194 daddr_t reqblock; 195 int poff; 196 int bsize; 197 int pagesperblock, blocksperpage; 198 199 /* 200 * If no vp or vp is doomed or marked transparent to VM, we do not 201 * have the page. 202 */ 203 if ((vp == NULL) || (vp->v_flag & VDOOMED)) 204 return FALSE; 205 206 /* 207 * If filesystem no longer mounted or offset beyond end of file we do 208 * not have the page. 209 */ 210 if ((vp->v_mount == NULL) || 211 (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)) 212 return FALSE; 213 214 bsize = vp->v_mount->mnt_stat.f_iosize; 215 pagesperblock = bsize / PAGE_SIZE; 216 blocksperpage = 0; 217 if (pagesperblock > 0) { 218 reqblock = pindex / pagesperblock; 219 } else { 220 blocksperpage = (PAGE_SIZE / bsize); 221 reqblock = pindex * blocksperpage; 222 } 223 err = VOP_BMAP(vp, reqblock, (struct vnode **) 0, &bn, 224 after, before); 225 if (err) 226 return TRUE; 227 if ( bn == -1) 228 return FALSE; 229 if (pagesperblock > 0) { 230 poff = pindex - (reqblock * pagesperblock); 231 if (before) { 232 *before *= pagesperblock; 233 *before += poff; 234 } 235 if (after) { 236 int numafter; 237 *after *= pagesperblock; 238 numafter = pagesperblock - (poff + 1); 239 if (IDX_TO_OFF(pindex + numafter) > object->un_pager.vnp.vnp_size) { 240 numafter = OFF_TO_IDX((object->un_pager.vnp.vnp_size - IDX_TO_OFF(pindex))); 241 } 242 *after += numafter; 243 } 244 } else { 245 if (before) { 246 *before /= blocksperpage; 247 } 248 249 if (after) { 250 *after /= blocksperpage; 251 } 252 } 253 return TRUE; 254 } 255 256 /* 257 * Lets the VM system know about a change in size for a file. 258 * We adjust our own internal size and flush any cached pages in 259 * the associated object that are affected by the size change. 260 * 261 * Note: this routine may be invoked as a result of a pager put 262 * operation (possibly at object termination time), so we must be careful. 263 */ 264 void 265 vnode_pager_setsize(vp, nsize) 266 struct vnode *vp; 267 vm_ooffset_t nsize; 268 { 269 vm_pindex_t nobjsize; 270 vm_object_t object = vp->v_object; 271 272 if (object == NULL) 273 return; 274 275 /* 276 * Hasn't changed size 277 */ 278 if (nsize == object->un_pager.vnp.vnp_size) 279 return; 280 281 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 282 283 /* 284 * File has shrunk. Toss any cached pages beyond the new EOF. 285 */ 286 if (nsize < object->un_pager.vnp.vnp_size) { 287 vm_freeze_copyopts(object, OFF_TO_IDX(nsize), object->size); 288 if (nobjsize < object->size) { 289 vm_object_page_remove(object, nobjsize, object->size, 290 FALSE); 291 } 292 /* 293 * this gets rid of garbage at the end of a page that is now 294 * only partially backed by the vnode. 295 * 296 * XXX for some reason (I don't know yet), if we take a 297 * completely invalid page and mark it partially valid 298 * it can screw up NFS reads, so we don't allow the case. 299 */ 300 if (nsize & PAGE_MASK) { 301 vm_offset_t kva; 302 vm_page_t m; 303 304 m = vm_page_lookup(object, OFF_TO_IDX(nsize)); 305 if (m && m->valid) { 306 int base = (int)nsize & PAGE_MASK; 307 int size = PAGE_SIZE - base; 308 309 /* 310 * Clear out partial-page garbage in case 311 * the page has been mapped. 312 */ 313 kva = vm_pager_map_page(m); 314 bzero((caddr_t)kva + base, size); 315 vm_pager_unmap_page(kva); 316 317 /* 318 * XXX work around SMP data integrity race 319 * by unmapping the page from user processes. 320 * The garbage we just cleared may be mapped 321 * to a user process running on another cpu 322 * and this code is not running through normal 323 * I/O channels which handle SMP issues for 324 * us, so unmap page to synchronize all cpus. 325 * 326 * XXX should vm_pager_unmap_page() have 327 * dealt with this? 328 */ 329 vm_page_protect(m, VM_PROT_NONE); 330 331 /* 332 * Clear out partial-page dirty bits. This 333 * has the side effect of setting the valid 334 * bits, but that is ok. There are a bunch 335 * of places in the VM system where we expected 336 * m->dirty == VM_PAGE_BITS_ALL. The file EOF 337 * case is one of them. If the page is still 338 * partially dirty, make it fully dirty. 339 * 340 * note that we do not clear out the valid 341 * bits. This would prevent bogus_page 342 * replacement from working properly. 343 */ 344 vm_page_set_validclean(m, base, size); 345 if (m->dirty != 0) 346 m->dirty = VM_PAGE_BITS_ALL; 347 } 348 } 349 } 350 object->un_pager.vnp.vnp_size = nsize; 351 object->size = nobjsize; 352 } 353 354 void 355 vnode_pager_freepage(m) 356 vm_page_t m; 357 { 358 vm_page_free(m); 359 } 360 361 /* 362 * calculate the linear (byte) disk address of specified virtual 363 * file address 364 */ 365 static vm_offset_t 366 vnode_pager_addr(vp, address, run) 367 struct vnode *vp; 368 vm_ooffset_t address; 369 int *run; 370 { 371 int rtaddress; 372 int bsize; 373 daddr_t block; 374 struct vnode *rtvp; 375 int err; 376 daddr_t vblock; 377 int voffset; 378 379 if ((int) address < 0) 380 return -1; 381 382 if (vp->v_mount == NULL) 383 return -1; 384 385 bsize = vp->v_mount->mnt_stat.f_iosize; 386 vblock = address / bsize; 387 voffset = address % bsize; 388 389 err = VOP_BMAP(vp, vblock, &rtvp, &block, run, NULL); 390 391 if (err || (block == -1)) 392 rtaddress = -1; 393 else { 394 rtaddress = block + voffset / DEV_BSIZE; 395 if( run) { 396 *run += 1; 397 *run *= bsize/PAGE_SIZE; 398 *run -= voffset/PAGE_SIZE; 399 } 400 } 401 402 return rtaddress; 403 } 404 405 /* 406 * interrupt routine for I/O completion 407 */ 408 static void 409 vnode_pager_iodone(bp) 410 struct buf *bp; 411 { 412 bp->b_flags |= B_DONE; 413 wakeup(bp); 414 } 415 416 /* 417 * small block file system vnode pager input 418 */ 419 static int 420 vnode_pager_input_smlfs(object, m) 421 vm_object_t object; 422 vm_page_t m; 423 { 424 int i; 425 int s; 426 struct vnode *dp, *vp; 427 struct buf *bp; 428 vm_offset_t kva; 429 int fileaddr; 430 vm_offset_t bsize; 431 int error = 0; 432 433 vp = object->handle; 434 if (vp->v_mount == NULL) 435 return VM_PAGER_BAD; 436 437 bsize = vp->v_mount->mnt_stat.f_iosize; 438 439 440 VOP_BMAP(vp, 0, &dp, 0, NULL, NULL); 441 442 kva = vm_pager_map_page(m); 443 444 for (i = 0; i < PAGE_SIZE / bsize; i++) { 445 vm_ooffset_t address; 446 447 if (vm_page_bits(i * bsize, bsize) & m->valid) 448 continue; 449 450 address = IDX_TO_OFF(m->pindex) + i * bsize; 451 if (address >= object->un_pager.vnp.vnp_size) { 452 fileaddr = -1; 453 } else { 454 fileaddr = vnode_pager_addr(vp, address, NULL); 455 } 456 if (fileaddr != -1) { 457 bp = getpbuf(&vnode_pbuf_freecnt); 458 459 /* build a minimal buffer header */ 460 bp->b_flags = B_READ | B_CALL; 461 bp->b_iodone = vnode_pager_iodone; 462 bp->b_data = (caddr_t) kva + i * bsize; 463 bp->b_blkno = fileaddr; 464 pbgetvp(dp, bp); 465 bp->b_bcount = bsize; 466 bp->b_bufsize = bsize; 467 bp->b_runningbufspace = bp->b_bufsize; 468 runningbufspace += bp->b_runningbufspace; 469 470 /* do the input */ 471 VOP_STRATEGY(bp->b_vp, bp); 472 473 /* we definitely need to be at splvm here */ 474 475 s = splvm(); 476 while ((bp->b_flags & B_DONE) == 0) { 477 tsleep(bp, 0, "vnsrd", 0); 478 } 479 splx(s); 480 if ((bp->b_flags & B_ERROR) != 0) 481 error = EIO; 482 483 /* 484 * free the buffer header back to the swap buffer pool 485 */ 486 relpbuf(bp, &vnode_pbuf_freecnt); 487 if (error) 488 break; 489 490 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 491 } else { 492 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 493 bzero((caddr_t) kva + i * bsize, bsize); 494 } 495 } 496 vm_pager_unmap_page(kva); 497 pmap_clear_modify(m); 498 vm_page_flag_clear(m, PG_ZERO); 499 if (error) { 500 return VM_PAGER_ERROR; 501 } 502 return VM_PAGER_OK; 503 504 } 505 506 507 /* 508 * old style vnode pager output routine 509 */ 510 static int 511 vnode_pager_input_old(object, m) 512 vm_object_t object; 513 vm_page_t m; 514 { 515 struct uio auio; 516 struct iovec aiov; 517 int error; 518 int size; 519 vm_offset_t kva; 520 521 error = 0; 522 523 /* 524 * Return failure if beyond current EOF 525 */ 526 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 527 return VM_PAGER_BAD; 528 } else { 529 size = PAGE_SIZE; 530 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 531 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 532 533 /* 534 * Allocate a kernel virtual address and initialize so that 535 * we can use VOP_READ/WRITE routines. 536 */ 537 kva = vm_pager_map_page(m); 538 539 aiov.iov_base = (caddr_t) kva; 540 aiov.iov_len = size; 541 auio.uio_iov = &aiov; 542 auio.uio_iovcnt = 1; 543 auio.uio_offset = IDX_TO_OFF(m->pindex); 544 auio.uio_segflg = UIO_SYSSPACE; 545 auio.uio_rw = UIO_READ; 546 auio.uio_resid = size; 547 auio.uio_td = curthread; 548 549 error = VOP_READ(object->handle, &auio, 0, proc0.p_ucred); 550 if (!error) { 551 int count = size - auio.uio_resid; 552 553 if (count == 0) 554 error = EINVAL; 555 else if (count != PAGE_SIZE) 556 bzero((caddr_t) kva + count, PAGE_SIZE - count); 557 } 558 vm_pager_unmap_page(kva); 559 } 560 pmap_clear_modify(m); 561 vm_page_undirty(m); 562 vm_page_flag_clear(m, PG_ZERO); 563 if (!error) 564 m->valid = VM_PAGE_BITS_ALL; 565 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 566 } 567 568 /* 569 * generic vnode pager input routine 570 */ 571 572 /* 573 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 574 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to 575 * vnode_pager_generic_getpages() to implement the previous behaviour. 576 * 577 * All other FS's should use the bypass to get to the local media 578 * backing vp's VOP_GETPAGES. 579 */ 580 static int 581 vnode_pager_getpages(object, m, count, reqpage) 582 vm_object_t object; 583 vm_page_t *m; 584 int count; 585 int reqpage; 586 { 587 int rtval; 588 struct vnode *vp; 589 int bytes = count * PAGE_SIZE; 590 591 vp = object->handle; 592 /* 593 * XXX temporary diagnostic message to help track stale FS code, 594 * Returning EOPNOTSUPP from here may make things unhappy. 595 */ 596 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0); 597 if (rtval == EOPNOTSUPP) { 598 printf("vnode_pager: *** WARNING *** stale FS getpages\n"); 599 rtval = vnode_pager_generic_getpages( vp, m, bytes, reqpage); 600 } 601 return rtval; 602 } 603 604 605 /* 606 * This is now called from local media FS's to operate against their 607 * own vnodes if they fail to implement VOP_GETPAGES. 608 */ 609 int 610 vnode_pager_generic_getpages(vp, m, bytecount, reqpage) 611 struct vnode *vp; 612 vm_page_t *m; 613 int bytecount; 614 int reqpage; 615 { 616 vm_object_t object; 617 vm_offset_t kva; 618 off_t foff, tfoff, nextoff; 619 int i, size, bsize, first, firstaddr; 620 struct vnode *dp; 621 int runpg; 622 int runend; 623 struct buf *bp; 624 int s; 625 int count; 626 int error = 0; 627 628 object = vp->v_object; 629 count = bytecount / PAGE_SIZE; 630 631 if (vp->v_mount == NULL) 632 return VM_PAGER_BAD; 633 634 bsize = vp->v_mount->mnt_stat.f_iosize; 635 636 /* get the UNDERLYING device for the file with VOP_BMAP() */ 637 638 /* 639 * originally, we did not check for an error return value -- assuming 640 * an fs always has a bmap entry point -- that assumption is wrong!!! 641 */ 642 foff = IDX_TO_OFF(m[reqpage]->pindex); 643 644 /* 645 * if we can't bmap, use old VOP code 646 */ 647 if (VOP_BMAP(vp, 0, &dp, 0, NULL, NULL)) { 648 for (i = 0; i < count; i++) { 649 if (i != reqpage) { 650 vnode_pager_freepage(m[i]); 651 } 652 } 653 mycpu->gd_cnt.v_vnodein++; 654 mycpu->gd_cnt.v_vnodepgsin++; 655 return vnode_pager_input_old(object, m[reqpage]); 656 657 /* 658 * if the blocksize is smaller than a page size, then use 659 * special small filesystem code. NFS sometimes has a small 660 * blocksize, but it can handle large reads itself. 661 */ 662 } else if ((PAGE_SIZE / bsize) > 1 && 663 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) { 664 for (i = 0; i < count; i++) { 665 if (i != reqpage) { 666 vnode_pager_freepage(m[i]); 667 } 668 } 669 mycpu->gd_cnt.v_vnodein++; 670 mycpu->gd_cnt.v_vnodepgsin++; 671 return vnode_pager_input_smlfs(object, m[reqpage]); 672 } 673 674 /* 675 * If we have a completely valid page available to us, we can 676 * clean up and return. Otherwise we have to re-read the 677 * media. 678 */ 679 680 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) { 681 for (i = 0; i < count; i++) { 682 if (i != reqpage) 683 vnode_pager_freepage(m[i]); 684 } 685 return VM_PAGER_OK; 686 } 687 m[reqpage]->valid = 0; 688 689 /* 690 * here on direct device I/O 691 */ 692 693 firstaddr = -1; 694 /* 695 * calculate the run that includes the required page 696 */ 697 for(first = 0, i = 0; i < count; i = runend) { 698 firstaddr = vnode_pager_addr(vp, 699 IDX_TO_OFF(m[i]->pindex), &runpg); 700 if (firstaddr == -1) { 701 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) { 702 /* XXX no %qd in kernel. */ 703 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %d, foff: 0x%lx%08lx, vnp_size: 0x%lx%08lx", 704 firstaddr, (u_long)(foff >> 32), 705 (u_long)(u_int32_t)foff, 706 (u_long)(u_int32_t) 707 (object->un_pager.vnp.vnp_size >> 32), 708 (u_long)(u_int32_t) 709 object->un_pager.vnp.vnp_size); 710 } 711 vnode_pager_freepage(m[i]); 712 runend = i + 1; 713 first = runend; 714 continue; 715 } 716 runend = i + runpg; 717 if (runend <= reqpage) { 718 int j; 719 for (j = i; j < runend; j++) { 720 vnode_pager_freepage(m[j]); 721 } 722 } else { 723 if (runpg < (count - first)) { 724 for (i = first + runpg; i < count; i++) 725 vnode_pager_freepage(m[i]); 726 count = first + runpg; 727 } 728 break; 729 } 730 first = runend; 731 } 732 733 /* 734 * the first and last page have been calculated now, move input pages 735 * to be zero based... 736 */ 737 if (first != 0) { 738 for (i = first; i < count; i++) { 739 m[i - first] = m[i]; 740 } 741 count -= first; 742 reqpage -= first; 743 } 744 745 /* 746 * calculate the file virtual address for the transfer 747 */ 748 foff = IDX_TO_OFF(m[0]->pindex); 749 750 /* 751 * calculate the size of the transfer 752 */ 753 size = count * PAGE_SIZE; 754 if ((foff + size) > object->un_pager.vnp.vnp_size) 755 size = object->un_pager.vnp.vnp_size - foff; 756 757 /* 758 * round up physical size for real devices. 759 */ 760 if (dp->v_type == VBLK || dp->v_type == VCHR) { 761 int secmask = dp->v_rdev->si_bsize_phys - 1; 762 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1)); 763 size = (size + secmask) & ~secmask; 764 } 765 766 bp = getpbuf(&vnode_pbuf_freecnt); 767 kva = (vm_offset_t) bp->b_data; 768 769 /* 770 * and map the pages to be read into the kva 771 */ 772 pmap_qenter(kva, m, count); 773 774 /* build a minimal buffer header */ 775 bp->b_flags = B_READ | B_CALL; 776 bp->b_iodone = vnode_pager_iodone; 777 /* B_PHYS is not set, but it is nice to fill this in */ 778 bp->b_blkno = firstaddr; 779 pbgetvp(dp, bp); 780 bp->b_bcount = size; 781 bp->b_bufsize = size; 782 bp->b_runningbufspace = bp->b_bufsize; 783 runningbufspace += bp->b_runningbufspace; 784 785 mycpu->gd_cnt.v_vnodein++; 786 mycpu->gd_cnt.v_vnodepgsin += count; 787 788 /* do the input */ 789 VOP_STRATEGY(bp->b_vp, bp); 790 791 s = splvm(); 792 /* we definitely need to be at splvm here */ 793 794 while ((bp->b_flags & B_DONE) == 0) { 795 tsleep(bp, 0, "vnread", 0); 796 } 797 splx(s); 798 if ((bp->b_flags & B_ERROR) != 0) 799 error = EIO; 800 801 if (!error) { 802 if (size != count * PAGE_SIZE) 803 bzero((caddr_t) kva + size, PAGE_SIZE * count - size); 804 } 805 pmap_qremove(kva, count); 806 807 /* 808 * free the buffer header back to the swap buffer pool 809 */ 810 relpbuf(bp, &vnode_pbuf_freecnt); 811 812 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) { 813 vm_page_t mt; 814 815 nextoff = tfoff + PAGE_SIZE; 816 mt = m[i]; 817 818 if (nextoff <= object->un_pager.vnp.vnp_size) { 819 /* 820 * Read filled up entire page. 821 */ 822 mt->valid = VM_PAGE_BITS_ALL; 823 vm_page_undirty(mt); /* should be an assert? XXX */ 824 pmap_clear_modify(mt); 825 } else { 826 /* 827 * Read did not fill up entire page. Since this 828 * is getpages, the page may be mapped, so we have 829 * to zero the invalid portions of the page even 830 * though we aren't setting them valid. 831 * 832 * Currently we do not set the entire page valid, 833 * we just try to clear the piece that we couldn't 834 * read. 835 */ 836 vm_page_set_validclean(mt, 0, 837 object->un_pager.vnp.vnp_size - tfoff); 838 /* handled by vm_fault now */ 839 /* vm_page_zero_invalid(mt, FALSE); */ 840 } 841 842 vm_page_flag_clear(mt, PG_ZERO); 843 if (i != reqpage) { 844 845 /* 846 * whether or not to leave the page activated is up in 847 * the air, but we should put the page on a page queue 848 * somewhere. (it already is in the object). Result: 849 * It appears that empirical results show that 850 * deactivating pages is best. 851 */ 852 853 /* 854 * just in case someone was asking for this page we 855 * now tell them that it is ok to use 856 */ 857 if (!error) { 858 if (mt->flags & PG_WANTED) 859 vm_page_activate(mt); 860 else 861 vm_page_deactivate(mt); 862 vm_page_wakeup(mt); 863 } else { 864 vnode_pager_freepage(mt); 865 } 866 } 867 } 868 if (error) { 869 printf("vnode_pager_getpages: I/O read error\n"); 870 } 871 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 872 } 873 874 /* 875 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 876 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 877 * vnode_pager_generic_putpages() to implement the previous behaviour. 878 * 879 * All other FS's should use the bypass to get to the local media 880 * backing vp's VOP_PUTPAGES. 881 */ 882 static void 883 vnode_pager_putpages(object, m, count, sync, rtvals) 884 vm_object_t object; 885 vm_page_t *m; 886 int count; 887 boolean_t sync; 888 int *rtvals; 889 { 890 int rtval; 891 struct vnode *vp; 892 int bytes = count * PAGE_SIZE; 893 894 /* 895 * Force synchronous operation if we are extremely low on memory 896 * to prevent a low-memory deadlock. VOP operations often need to 897 * allocate more memory to initiate the I/O ( i.e. do a BMAP 898 * operation ). The swapper handles the case by limiting the amount 899 * of asynchronous I/O, but that sort of solution doesn't scale well 900 * for the vnode pager without a lot of work. 901 * 902 * Also, the backing vnode's iodone routine may not wake the pageout 903 * daemon up. This should be probably be addressed XXX. 904 */ 905 906 if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min) 907 sync |= OBJPC_SYNC; 908 909 /* 910 * Call device-specific putpages function 911 */ 912 913 vp = object->handle; 914 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 915 if (rtval == EOPNOTSUPP) { 916 printf("vnode_pager: *** WARNING *** stale FS putpages\n"); 917 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals); 918 } 919 } 920 921 922 /* 923 * This is now called from local media FS's to operate against their 924 * own vnodes if they fail to implement VOP_PUTPAGES. 925 * 926 * This is typically called indirectly via the pageout daemon and 927 * clustering has already typically occured, so in general we ask the 928 * underlying filesystem to write the data out asynchronously rather 929 * then delayed. 930 */ 931 int 932 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals) 933 struct vnode *vp; 934 vm_page_t *m; 935 int bytecount; 936 int flags; 937 int *rtvals; 938 { 939 int i; 940 vm_object_t object; 941 int count; 942 943 int maxsize, ncount; 944 vm_ooffset_t poffset; 945 struct uio auio; 946 struct iovec aiov; 947 int error; 948 int ioflags; 949 950 object = vp->v_object; 951 count = bytecount / PAGE_SIZE; 952 953 for (i = 0; i < count; i++) 954 rtvals[i] = VM_PAGER_AGAIN; 955 956 if ((int) m[0]->pindex < 0) { 957 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n", 958 (long)m[0]->pindex, m[0]->dirty); 959 rtvals[0] = VM_PAGER_BAD; 960 return VM_PAGER_BAD; 961 } 962 963 maxsize = count * PAGE_SIZE; 964 ncount = count; 965 966 poffset = IDX_TO_OFF(m[0]->pindex); 967 968 /* 969 * If the page-aligned write is larger then the actual file we 970 * have to invalidate pages occuring beyond the file EOF. However, 971 * there is an edge case where a file may not be page-aligned where 972 * the last page is partially invalid. In this case the filesystem 973 * may not properly clear the dirty bits for the entire page (which 974 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 975 * With the page locked we are free to fix-up the dirty bits here. 976 * 977 * We do not under any circumstances truncate the valid bits, as 978 * this will screw up bogus page replacement. 979 */ 980 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 981 if (object->un_pager.vnp.vnp_size > poffset) { 982 int pgoff; 983 984 maxsize = object->un_pager.vnp.vnp_size - poffset; 985 ncount = btoc(maxsize); 986 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 987 vm_page_clear_dirty(m[ncount - 1], pgoff, 988 PAGE_SIZE - pgoff); 989 } 990 } else { 991 maxsize = 0; 992 ncount = 0; 993 } 994 if (ncount < count) { 995 for (i = ncount; i < count; i++) { 996 rtvals[i] = VM_PAGER_BAD; 997 } 998 } 999 } 1000 1001 /* 1002 * pageouts are already clustered, use IO_ASYNC to force a bawrite() 1003 * rather then a bdwrite() to prevent paging I/O from saturating 1004 * the buffer cache. Dummy-up the sequential heuristic to cause 1005 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 1006 * the system decides how to cluster. 1007 */ 1008 ioflags = IO_VMIO; 1009 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 1010 ioflags |= IO_SYNC; 1011 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 1012 ioflags |= IO_ASYNC; 1013 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 1014 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 1015 1016 aiov.iov_base = (caddr_t) 0; 1017 aiov.iov_len = maxsize; 1018 auio.uio_iov = &aiov; 1019 auio.uio_iovcnt = 1; 1020 auio.uio_offset = poffset; 1021 auio.uio_segflg = UIO_NOCOPY; 1022 auio.uio_rw = UIO_WRITE; 1023 auio.uio_resid = maxsize; 1024 auio.uio_td = NULL; 1025 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred); 1026 mycpu->gd_cnt.v_vnodeout++; 1027 mycpu->gd_cnt.v_vnodepgsout += ncount; 1028 1029 if (error) { 1030 printf("vnode_pager_putpages: I/O error %d\n", error); 1031 } 1032 if (auio.uio_resid) { 1033 printf("vnode_pager_putpages: residual I/O %d at %lu\n", 1034 auio.uio_resid, (u_long)m[0]->pindex); 1035 } 1036 for (i = 0; i < ncount; i++) { 1037 rtvals[i] = VM_PAGER_OK; 1038 } 1039 return rtvals[0]; 1040 } 1041 1042 struct vnode * 1043 vnode_pager_lock(vm_object_t object) 1044 { 1045 struct thread *td = curthread; /* XXX */ 1046 1047 for (; object != NULL; object = object->backing_object) { 1048 if (object->type != OBJT_VNODE) 1049 continue; 1050 if (object->flags & OBJ_DEAD) 1051 return NULL; 1052 1053 while (vget(object->handle, 1054 LK_NOPAUSE | LK_SHARED | LK_RETRY | LK_CANRECURSE, td)) { 1055 if ((object->flags & OBJ_DEAD) || 1056 (object->type != OBJT_VNODE)) { 1057 return NULL; 1058 } 1059 printf("vnode_pager_lock: retrying\n"); 1060 } 1061 return object->handle; 1062 } 1063 return NULL; 1064 } 1065