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 1.14 2004/05/13 17:40:19 dillon 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 #include <sys/sfbuf.h> 64 65 #include <vm/vm.h> 66 #include <vm/vm_object.h> 67 #include <vm/vm_page.h> 68 #include <vm/vm_pager.h> 69 #include <vm/vm_map.h> 70 #include <vm/vnode_pager.h> 71 #include <vm/vm_extern.h> 72 73 static vm_offset_t vnode_pager_addr (struct vnode *vp, vm_ooffset_t address, 74 int *run); 75 static void vnode_pager_iodone (struct buf *bp); 76 static int vnode_pager_input_smlfs (vm_object_t object, vm_page_t m); 77 static int vnode_pager_input_old (vm_object_t object, vm_page_t m); 78 static void vnode_pager_dealloc (vm_object_t); 79 static int vnode_pager_getpages (vm_object_t, vm_page_t *, int, int); 80 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *); 81 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t, int *, int *); 82 83 struct pagerops vnodepagerops = { 84 NULL, 85 vnode_pager_alloc, 86 vnode_pager_dealloc, 87 vnode_pager_getpages, 88 vnode_pager_putpages, 89 vnode_pager_haspage, 90 NULL 91 }; 92 93 int vnode_pbuf_freecnt = -1; /* start out unlimited */ 94 95 /* 96 * Allocate (or lookup) pager for a vnode. 97 * Handle is a vnode pointer. 98 */ 99 vm_object_t 100 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 101 vm_ooffset_t offset) 102 { 103 vm_object_t object; 104 struct vnode *vp; 105 106 /* 107 * Pageout to vnode, no can do yet. 108 */ 109 if (handle == NULL) 110 return (NULL); 111 112 /* 113 * XXX hack - This initialization should be put somewhere else. 114 */ 115 if (vnode_pbuf_freecnt < 0) { 116 vnode_pbuf_freecnt = nswbuf / 2 + 1; 117 } 118 119 vp = (struct vnode *) handle; 120 121 /* 122 * Prevent race condition when allocating the object. This 123 * can happen with NFS vnodes since the nfsnode isn't locked. 124 */ 125 while (vp->v_flag & VOLOCK) { 126 vp->v_flag |= VOWANT; 127 tsleep(vp, 0, "vnpobj", 0); 128 } 129 vp->v_flag |= VOLOCK; 130 131 /* 132 * If the object is being terminated, wait for it to 133 * go away. 134 */ 135 while (((object = vp->v_object) != NULL) && 136 (object->flags & OBJ_DEAD)) { 137 tsleep(object, 0, "vadead", 0); 138 } 139 140 if (vp->v_usecount == 0) 141 panic("vnode_pager_alloc: no vnode reference"); 142 143 if (object == NULL) { 144 /* 145 * And an object of the appropriate size 146 */ 147 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); 148 object->flags = 0; 149 150 object->un_pager.vnp.vnp_size = size; 151 152 object->handle = handle; 153 vp->v_object = object; 154 vp->v_usecount++; 155 } else { 156 object->ref_count++; 157 vp->v_usecount++; 158 } 159 160 vp->v_flag &= ~VOLOCK; 161 if (vp->v_flag & VOWANT) { 162 vp->v_flag &= ~VOWANT; 163 wakeup(vp); 164 } 165 return (object); 166 } 167 168 static void 169 vnode_pager_dealloc(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(vm_object_t object, vm_pindex_t pindex, int *before, 186 int *after) 187 { 188 struct vnode *vp = object->handle; 189 daddr_t bn; 190 int err; 191 daddr_t reqblock; 192 int poff; 193 int bsize; 194 int pagesperblock, blocksperpage; 195 196 /* 197 * If no vp or vp is doomed or marked transparent to VM, we do not 198 * have the page. 199 */ 200 if ((vp == NULL) || (vp->v_flag & VDOOMED)) 201 return FALSE; 202 203 /* 204 * If filesystem no longer mounted or offset beyond end of file we do 205 * not have the page. 206 */ 207 if ((vp->v_mount == NULL) || 208 (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)) 209 return FALSE; 210 211 bsize = vp->v_mount->mnt_stat.f_iosize; 212 pagesperblock = bsize / PAGE_SIZE; 213 blocksperpage = 0; 214 if (pagesperblock > 0) { 215 reqblock = pindex / pagesperblock; 216 } else { 217 blocksperpage = (PAGE_SIZE / bsize); 218 reqblock = pindex * blocksperpage; 219 } 220 err = VOP_BMAP(vp, reqblock, (struct vnode **) 0, &bn, 221 after, before); 222 if (err) 223 return TRUE; 224 if ( bn == -1) 225 return FALSE; 226 if (pagesperblock > 0) { 227 poff = pindex - (reqblock * pagesperblock); 228 if (before) { 229 *before *= pagesperblock; 230 *before += poff; 231 } 232 if (after) { 233 int numafter; 234 *after *= pagesperblock; 235 numafter = pagesperblock - (poff + 1); 236 if (IDX_TO_OFF(pindex + numafter) > object->un_pager.vnp.vnp_size) { 237 numafter = OFF_TO_IDX((object->un_pager.vnp.vnp_size - IDX_TO_OFF(pindex))); 238 } 239 *after += numafter; 240 } 241 } else { 242 if (before) { 243 *before /= blocksperpage; 244 } 245 246 if (after) { 247 *after /= blocksperpage; 248 } 249 } 250 return TRUE; 251 } 252 253 /* 254 * Lets the VM system know about a change in size for a file. 255 * We adjust our own internal size and flush any cached pages in 256 * the associated object that are affected by the size change. 257 * 258 * Note: this routine may be invoked as a result of a pager put 259 * operation (possibly at object termination time), so we must be careful. 260 */ 261 void 262 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize) 263 { 264 vm_pindex_t nobjsize; 265 vm_object_t object = vp->v_object; 266 267 if (object == NULL) 268 return; 269 270 /* 271 * Hasn't changed size 272 */ 273 if (nsize == object->un_pager.vnp.vnp_size) 274 return; 275 276 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 277 278 /* 279 * File has shrunk. Toss any cached pages beyond the new EOF. 280 */ 281 if (nsize < object->un_pager.vnp.vnp_size) { 282 vm_freeze_copyopts(object, OFF_TO_IDX(nsize), object->size); 283 if (nobjsize < object->size) { 284 vm_object_page_remove(object, nobjsize, object->size, 285 FALSE); 286 } 287 /* 288 * This gets rid of garbage at the end of a page that is now 289 * only partially backed by the vnode. Since we are setting 290 * the entire page valid & clean after we are done we have 291 * to be sure that the portion of the page within the file 292 * bounds is already valid. If it isn't then making it 293 * valid would create a corrupt block. 294 */ 295 if (nsize & PAGE_MASK) { 296 vm_offset_t kva; 297 vm_page_t m; 298 299 m = vm_page_lookup(object, OFF_TO_IDX(nsize)); 300 if (m && m->valid) { 301 int base = (int)nsize & PAGE_MASK; 302 int size = PAGE_SIZE - base; 303 struct sf_buf *sf; 304 305 /* 306 * Clear out partial-page garbage in case 307 * the page has been mapped. 308 */ 309 sf = sf_buf_alloc(m, SFBA_QUICK); 310 kva = sf_buf_kva(sf); 311 bzero((caddr_t)kva + base, size); 312 sf_buf_free(sf); 313 314 /* 315 * XXX work around SMP data integrity race 316 * by unmapping the page from user processes. 317 * The garbage we just cleared may be mapped 318 * to a user process running on another cpu 319 * and this code is not running through normal 320 * I/O channels which handle SMP issues for 321 * us, so unmap page to synchronize all cpus. 322 * 323 * XXX should vm_pager_unmap_page() have 324 * dealt with this? 325 */ 326 vm_page_protect(m, VM_PROT_NONE); 327 328 /* 329 * Clear out partial-page dirty bits. This 330 * has the side effect of setting the valid 331 * bits, but that is ok. There are a bunch 332 * of places in the VM system where we expected 333 * m->dirty == VM_PAGE_BITS_ALL. The file EOF 334 * case is one of them. If the page is still 335 * partially dirty, make it fully dirty. 336 * 337 * note that we do not clear out the valid 338 * bits. This would prevent bogus_page 339 * replacement from working properly. 340 */ 341 vm_page_set_validclean(m, base, size); 342 if (m->dirty != 0) 343 m->dirty = VM_PAGE_BITS_ALL; 344 } 345 } 346 } 347 object->un_pager.vnp.vnp_size = nsize; 348 object->size = nobjsize; 349 } 350 351 void 352 vnode_pager_freepage(vm_page_t m) 353 { 354 vm_page_free(m); 355 } 356 357 /* 358 * calculate the linear (byte) disk address of specified virtual 359 * file address 360 */ 361 static vm_offset_t 362 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, int *run) 363 { 364 int rtaddress; 365 int bsize; 366 daddr_t block; 367 struct vnode *rtvp; 368 int err; 369 daddr_t vblock; 370 int voffset; 371 372 if ((int) address < 0) 373 return -1; 374 375 if (vp->v_mount == NULL) 376 return -1; 377 378 bsize = vp->v_mount->mnt_stat.f_iosize; 379 vblock = address / bsize; 380 voffset = address % bsize; 381 382 err = VOP_BMAP(vp, vblock, &rtvp, &block, run, NULL); 383 384 if (err || (block == -1)) 385 rtaddress = -1; 386 else { 387 rtaddress = block + voffset / DEV_BSIZE; 388 if( run) { 389 *run += 1; 390 *run *= bsize/PAGE_SIZE; 391 *run -= voffset/PAGE_SIZE; 392 } 393 } 394 395 return rtaddress; 396 } 397 398 /* 399 * interrupt routine for I/O completion 400 */ 401 static void 402 vnode_pager_iodone(struct buf *bp) 403 { 404 bp->b_flags |= B_DONE; 405 wakeup(bp); 406 } 407 408 /* 409 * small block file system vnode pager input 410 */ 411 static int 412 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m) 413 { 414 int i; 415 int s; 416 struct vnode *dp, *vp; 417 struct buf *bp; 418 vm_offset_t kva; 419 struct sf_buf *sf; 420 int fileaddr; 421 vm_offset_t bsize; 422 int error = 0; 423 424 vp = object->handle; 425 if (vp->v_mount == NULL) 426 return VM_PAGER_BAD; 427 428 bsize = vp->v_mount->mnt_stat.f_iosize; 429 430 431 VOP_BMAP(vp, 0, &dp, 0, NULL, NULL); 432 433 sf = sf_buf_alloc(m, 0); 434 kva = sf_buf_kva(sf); 435 436 for (i = 0; i < PAGE_SIZE / bsize; i++) { 437 vm_ooffset_t address; 438 439 if (vm_page_bits(i * bsize, bsize) & m->valid) 440 continue; 441 442 address = IDX_TO_OFF(m->pindex) + i * bsize; 443 if (address >= object->un_pager.vnp.vnp_size) { 444 fileaddr = -1; 445 } else { 446 fileaddr = vnode_pager_addr(vp, address, NULL); 447 } 448 if (fileaddr != -1) { 449 bp = getpbuf(&vnode_pbuf_freecnt); 450 451 /* build a minimal buffer header */ 452 bp->b_flags = B_READ | B_CALL; 453 bp->b_iodone = vnode_pager_iodone; 454 bp->b_data = (caddr_t) kva + i * bsize; 455 bp->b_blkno = fileaddr; 456 pbgetvp(dp, bp); 457 bp->b_bcount = bsize; 458 bp->b_bufsize = bsize; 459 bp->b_runningbufspace = bp->b_bufsize; 460 runningbufspace += bp->b_runningbufspace; 461 462 /* do the input */ 463 VOP_STRATEGY(bp->b_vp, bp); 464 465 /* we definitely need to be at splvm here */ 466 467 s = splvm(); 468 while ((bp->b_flags & B_DONE) == 0) { 469 tsleep(bp, 0, "vnsrd", 0); 470 } 471 splx(s); 472 if ((bp->b_flags & B_ERROR) != 0) 473 error = EIO; 474 475 /* 476 * free the buffer header back to the swap buffer pool 477 */ 478 relpbuf(bp, &vnode_pbuf_freecnt); 479 if (error) 480 break; 481 482 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 483 } else { 484 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 485 bzero((caddr_t) kva + i * bsize, bsize); 486 } 487 } 488 sf_buf_free(sf); 489 pmap_clear_modify(m); 490 vm_page_flag_clear(m, PG_ZERO); 491 if (error) { 492 return VM_PAGER_ERROR; 493 } 494 return VM_PAGER_OK; 495 496 } 497 498 499 /* 500 * old style vnode pager output routine 501 */ 502 static int 503 vnode_pager_input_old(vm_object_t object, vm_page_t m) 504 { 505 struct uio auio; 506 struct iovec aiov; 507 int error; 508 int size; 509 vm_offset_t kva; 510 struct sf_buf *sf; 511 512 error = 0; 513 514 /* 515 * Return failure if beyond current EOF 516 */ 517 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 518 return VM_PAGER_BAD; 519 } else { 520 size = PAGE_SIZE; 521 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 522 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 523 524 /* 525 * Allocate a kernel virtual address and initialize so that 526 * we can use VOP_READ/WRITE routines. 527 */ 528 sf = sf_buf_alloc(m, 0); 529 kva = sf_buf_kva(sf); 530 531 aiov.iov_base = (caddr_t) kva; 532 aiov.iov_len = size; 533 auio.uio_iov = &aiov; 534 auio.uio_iovcnt = 1; 535 auio.uio_offset = IDX_TO_OFF(m->pindex); 536 auio.uio_segflg = UIO_SYSSPACE; 537 auio.uio_rw = UIO_READ; 538 auio.uio_resid = size; 539 auio.uio_td = curthread; 540 541 error = VOP_READ(object->handle, &auio, 0, proc0.p_ucred); 542 if (!error) { 543 int count = size - auio.uio_resid; 544 545 if (count == 0) 546 error = EINVAL; 547 else if (count != PAGE_SIZE) 548 bzero((caddr_t) kva + count, PAGE_SIZE - count); 549 } 550 sf_buf_free(sf); 551 } 552 pmap_clear_modify(m); 553 vm_page_undirty(m); 554 vm_page_flag_clear(m, PG_ZERO); 555 if (!error) 556 m->valid = VM_PAGE_BITS_ALL; 557 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 558 } 559 560 /* 561 * generic vnode pager input routine 562 */ 563 564 /* 565 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 566 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to 567 * vnode_pager_generic_getpages() to implement the previous behaviour. 568 * 569 * All other FS's should use the bypass to get to the local media 570 * backing vp's VOP_GETPAGES. 571 */ 572 static int 573 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage) 574 { 575 int rtval; 576 struct vnode *vp; 577 int bytes = count * PAGE_SIZE; 578 579 vp = object->handle; 580 /* 581 * XXX temporary diagnostic message to help track stale FS code, 582 * Returning EOPNOTSUPP from here may make things unhappy. 583 */ 584 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0); 585 if (rtval == EOPNOTSUPP) { 586 printf("vnode_pager: *** WARNING *** stale FS getpages\n"); 587 rtval = vnode_pager_generic_getpages( vp, m, bytes, reqpage); 588 } 589 return rtval; 590 } 591 592 593 /* 594 * This is now called from local media FS's to operate against their 595 * own vnodes if they fail to implement VOP_GETPAGES. 596 */ 597 int 598 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int bytecount, 599 int reqpage) 600 { 601 vm_object_t object; 602 vm_offset_t kva; 603 off_t foff, tfoff, nextoff; 604 int i, size, bsize, first, firstaddr; 605 struct vnode *dp; 606 int runpg; 607 int runend; 608 struct buf *bp; 609 int s; 610 int count; 611 int error = 0; 612 613 object = vp->v_object; 614 count = bytecount / PAGE_SIZE; 615 616 if (vp->v_mount == NULL) 617 return VM_PAGER_BAD; 618 619 bsize = vp->v_mount->mnt_stat.f_iosize; 620 621 /* get the UNDERLYING device for the file with VOP_BMAP() */ 622 623 /* 624 * originally, we did not check for an error return value -- assuming 625 * an fs always has a bmap entry point -- that assumption is wrong!!! 626 */ 627 foff = IDX_TO_OFF(m[reqpage]->pindex); 628 629 /* 630 * if we can't bmap, use old VOP code 631 */ 632 if (VOP_BMAP(vp, 0, &dp, 0, NULL, NULL)) { 633 for (i = 0; i < count; i++) { 634 if (i != reqpage) { 635 vnode_pager_freepage(m[i]); 636 } 637 } 638 mycpu->gd_cnt.v_vnodein++; 639 mycpu->gd_cnt.v_vnodepgsin++; 640 return vnode_pager_input_old(object, m[reqpage]); 641 642 /* 643 * if the blocksize is smaller than a page size, then use 644 * special small filesystem code. NFS sometimes has a small 645 * blocksize, but it can handle large reads itself. 646 */ 647 } else if ((PAGE_SIZE / bsize) > 1 && 648 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) { 649 for (i = 0; i < count; i++) { 650 if (i != reqpage) { 651 vnode_pager_freepage(m[i]); 652 } 653 } 654 mycpu->gd_cnt.v_vnodein++; 655 mycpu->gd_cnt.v_vnodepgsin++; 656 return vnode_pager_input_smlfs(object, m[reqpage]); 657 } 658 659 /* 660 * If we have a completely valid page available to us, we can 661 * clean up and return. Otherwise we have to re-read the 662 * media. 663 * 664 * Note that this does not work with NFS, so NFS has its own 665 * getpages routine. The problem is that NFS can have partially 666 * valid pages associated with the buffer cache due to the piecemeal 667 * write support. If we were to fall through and re-read the media 668 * as we do here, dirty data could be lost. 669 */ 670 671 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) { 672 for (i = 0; i < count; i++) { 673 if (i != reqpage) 674 vnode_pager_freepage(m[i]); 675 } 676 return VM_PAGER_OK; 677 } 678 m[reqpage]->valid = 0; 679 680 /* 681 * here on direct device I/O 682 */ 683 684 firstaddr = -1; 685 /* 686 * calculate the run that includes the required page 687 */ 688 for(first = 0, i = 0; i < count; i = runend) { 689 firstaddr = vnode_pager_addr(vp, 690 IDX_TO_OFF(m[i]->pindex), &runpg); 691 if (firstaddr == -1) { 692 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) { 693 /* XXX no %qd in kernel. */ 694 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %d, foff: 0x%lx%08lx, vnp_size: 0x%lx%08lx", 695 firstaddr, (u_long)(foff >> 32), 696 (u_long)(u_int32_t)foff, 697 (u_long)(u_int32_t) 698 (object->un_pager.vnp.vnp_size >> 32), 699 (u_long)(u_int32_t) 700 object->un_pager.vnp.vnp_size); 701 } 702 vnode_pager_freepage(m[i]); 703 runend = i + 1; 704 first = runend; 705 continue; 706 } 707 runend = i + runpg; 708 if (runend <= reqpage) { 709 int j; 710 for (j = i; j < runend; j++) { 711 vnode_pager_freepage(m[j]); 712 } 713 } else { 714 if (runpg < (count - first)) { 715 for (i = first + runpg; i < count; i++) 716 vnode_pager_freepage(m[i]); 717 count = first + runpg; 718 } 719 break; 720 } 721 first = runend; 722 } 723 724 /* 725 * the first and last page have been calculated now, move input pages 726 * to be zero based... 727 */ 728 if (first != 0) { 729 for (i = first; i < count; i++) { 730 m[i - first] = m[i]; 731 } 732 count -= first; 733 reqpage -= first; 734 } 735 736 /* 737 * calculate the file virtual address for the transfer 738 */ 739 foff = IDX_TO_OFF(m[0]->pindex); 740 741 /* 742 * calculate the size of the transfer 743 */ 744 size = count * PAGE_SIZE; 745 if ((foff + size) > object->un_pager.vnp.vnp_size) 746 size = object->un_pager.vnp.vnp_size - foff; 747 748 /* 749 * round up physical size for real devices. 750 */ 751 if (dp->v_type == VBLK || dp->v_type == VCHR) { 752 int secmask = dp->v_rdev->si_bsize_phys - 1; 753 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1)); 754 size = (size + secmask) & ~secmask; 755 } 756 757 bp = getpbuf(&vnode_pbuf_freecnt); 758 kva = (vm_offset_t) bp->b_data; 759 760 /* 761 * and map the pages to be read into the kva 762 */ 763 pmap_qenter(kva, m, count); 764 765 /* build a minimal buffer header */ 766 bp->b_flags = B_READ | B_CALL; 767 bp->b_iodone = vnode_pager_iodone; 768 /* B_PHYS is not set, but it is nice to fill this in */ 769 bp->b_blkno = firstaddr; 770 pbgetvp(dp, bp); 771 bp->b_bcount = size; 772 bp->b_bufsize = size; 773 bp->b_runningbufspace = bp->b_bufsize; 774 runningbufspace += bp->b_runningbufspace; 775 776 mycpu->gd_cnt.v_vnodein++; 777 mycpu->gd_cnt.v_vnodepgsin += count; 778 779 /* do the input */ 780 VOP_STRATEGY(bp->b_vp, bp); 781 782 s = splvm(); 783 /* we definitely need to be at splvm here */ 784 785 while ((bp->b_flags & B_DONE) == 0) { 786 tsleep(bp, 0, "vnread", 0); 787 } 788 splx(s); 789 if ((bp->b_flags & B_ERROR) != 0) 790 error = EIO; 791 792 if (!error) { 793 if (size != count * PAGE_SIZE) 794 bzero((caddr_t) kva + size, PAGE_SIZE * count - size); 795 } 796 pmap_qremove(kva, count); 797 798 /* 799 * free the buffer header back to the swap buffer pool 800 */ 801 relpbuf(bp, &vnode_pbuf_freecnt); 802 803 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) { 804 vm_page_t mt; 805 806 nextoff = tfoff + PAGE_SIZE; 807 mt = m[i]; 808 809 if (nextoff <= object->un_pager.vnp.vnp_size) { 810 /* 811 * Read filled up entire page. 812 */ 813 mt->valid = VM_PAGE_BITS_ALL; 814 vm_page_undirty(mt); /* should be an assert? XXX */ 815 pmap_clear_modify(mt); 816 } else { 817 /* 818 * Read did not fill up entire page. Since this 819 * is getpages, the page may be mapped, so we have 820 * to zero the invalid portions of the page even 821 * though we aren't setting them valid. 822 * 823 * Currently we do not set the entire page valid, 824 * we just try to clear the piece that we couldn't 825 * read. 826 */ 827 vm_page_set_validclean(mt, 0, 828 object->un_pager.vnp.vnp_size - tfoff); 829 /* handled by vm_fault now */ 830 /* vm_page_zero_invalid(mt, FALSE); */ 831 } 832 833 vm_page_flag_clear(mt, PG_ZERO); 834 if (i != reqpage) { 835 836 /* 837 * whether or not to leave the page activated is up in 838 * the air, but we should put the page on a page queue 839 * somewhere. (it already is in the object). Result: 840 * It appears that empirical results show that 841 * deactivating pages is best. 842 */ 843 844 /* 845 * just in case someone was asking for this page we 846 * now tell them that it is ok to use 847 */ 848 if (!error) { 849 if (mt->flags & PG_WANTED) 850 vm_page_activate(mt); 851 else 852 vm_page_deactivate(mt); 853 vm_page_wakeup(mt); 854 } else { 855 vnode_pager_freepage(mt); 856 } 857 } 858 } 859 if (error) { 860 printf("vnode_pager_getpages: I/O read error\n"); 861 } 862 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 863 } 864 865 /* 866 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 867 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 868 * vnode_pager_generic_putpages() to implement the previous behaviour. 869 * 870 * All other FS's should use the bypass to get to the local media 871 * backing vp's VOP_PUTPAGES. 872 */ 873 static void 874 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count, 875 boolean_t sync, int *rtvals) 876 { 877 int rtval; 878 struct vnode *vp; 879 int bytes = count * PAGE_SIZE; 880 881 /* 882 * Force synchronous operation if we are extremely low on memory 883 * to prevent a low-memory deadlock. VOP operations often need to 884 * allocate more memory to initiate the I/O ( i.e. do a BMAP 885 * operation ). The swapper handles the case by limiting the amount 886 * of asynchronous I/O, but that sort of solution doesn't scale well 887 * for the vnode pager without a lot of work. 888 * 889 * Also, the backing vnode's iodone routine may not wake the pageout 890 * daemon up. This should be probably be addressed XXX. 891 */ 892 893 if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min) 894 sync |= OBJPC_SYNC; 895 896 /* 897 * Call device-specific putpages function 898 */ 899 900 vp = object->handle; 901 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 902 if (rtval == EOPNOTSUPP) { 903 printf("vnode_pager: *** WARNING *** stale FS putpages\n"); 904 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals); 905 } 906 } 907 908 909 /* 910 * This is now called from local media FS's to operate against their 911 * own vnodes if they fail to implement VOP_PUTPAGES. 912 * 913 * This is typically called indirectly via the pageout daemon and 914 * clustering has already typically occured, so in general we ask the 915 * underlying filesystem to write the data out asynchronously rather 916 * then delayed. 917 */ 918 int 919 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount, 920 int flags, int *rtvals) 921 { 922 int i; 923 vm_object_t object; 924 int count; 925 926 int maxsize, ncount; 927 vm_ooffset_t poffset; 928 struct uio auio; 929 struct iovec aiov; 930 int error; 931 int ioflags; 932 933 object = vp->v_object; 934 count = bytecount / PAGE_SIZE; 935 936 for (i = 0; i < count; i++) 937 rtvals[i] = VM_PAGER_AGAIN; 938 939 if ((int) m[0]->pindex < 0) { 940 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n", 941 (long)m[0]->pindex, m[0]->dirty); 942 rtvals[0] = VM_PAGER_BAD; 943 return VM_PAGER_BAD; 944 } 945 946 maxsize = count * PAGE_SIZE; 947 ncount = count; 948 949 poffset = IDX_TO_OFF(m[0]->pindex); 950 951 /* 952 * If the page-aligned write is larger then the actual file we 953 * have to invalidate pages occuring beyond the file EOF. However, 954 * there is an edge case where a file may not be page-aligned where 955 * the last page is partially invalid. In this case the filesystem 956 * may not properly clear the dirty bits for the entire page (which 957 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 958 * With the page locked we are free to fix-up the dirty bits here. 959 * 960 * We do not under any circumstances truncate the valid bits, as 961 * this will screw up bogus page replacement. 962 */ 963 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 964 if (object->un_pager.vnp.vnp_size > poffset) { 965 int pgoff; 966 967 maxsize = object->un_pager.vnp.vnp_size - poffset; 968 ncount = btoc(maxsize); 969 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 970 vm_page_clear_dirty(m[ncount - 1], pgoff, 971 PAGE_SIZE - pgoff); 972 } 973 } else { 974 maxsize = 0; 975 ncount = 0; 976 } 977 if (ncount < count) { 978 for (i = ncount; i < count; i++) { 979 rtvals[i] = VM_PAGER_BAD; 980 } 981 } 982 } 983 984 /* 985 * pageouts are already clustered, use IO_ASYNC to force a bawrite() 986 * rather then a bdwrite() to prevent paging I/O from saturating 987 * the buffer cache. Dummy-up the sequential heuristic to cause 988 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 989 * the system decides how to cluster. 990 */ 991 ioflags = IO_VMIO; 992 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 993 ioflags |= IO_SYNC; 994 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 995 ioflags |= IO_ASYNC; 996 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 997 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 998 999 aiov.iov_base = (caddr_t) 0; 1000 aiov.iov_len = maxsize; 1001 auio.uio_iov = &aiov; 1002 auio.uio_iovcnt = 1; 1003 auio.uio_offset = poffset; 1004 auio.uio_segflg = UIO_NOCOPY; 1005 auio.uio_rw = UIO_WRITE; 1006 auio.uio_resid = maxsize; 1007 auio.uio_td = NULL; 1008 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred); 1009 mycpu->gd_cnt.v_vnodeout++; 1010 mycpu->gd_cnt.v_vnodepgsout += ncount; 1011 1012 if (error) { 1013 printf("vnode_pager_putpages: I/O error %d\n", error); 1014 } 1015 if (auio.uio_resid) { 1016 printf("vnode_pager_putpages: residual I/O %d at %lu\n", 1017 auio.uio_resid, (u_long)m[0]->pindex); 1018 } 1019 for (i = 0; i < ncount; i++) { 1020 rtvals[i] = VM_PAGER_OK; 1021 } 1022 return rtvals[0]; 1023 } 1024 1025 struct vnode * 1026 vnode_pager_lock(vm_object_t object) 1027 { 1028 struct thread *td = curthread; /* XXX */ 1029 1030 for (; object != NULL; object = object->backing_object) { 1031 if (object->type != OBJT_VNODE) 1032 continue; 1033 if (object->flags & OBJ_DEAD) 1034 return NULL; 1035 1036 while (vget(object->handle, NULL, 1037 LK_NOPAUSE | LK_SHARED | LK_RETRY | LK_CANRECURSE, td)) { 1038 if ((object->flags & OBJ_DEAD) || 1039 (object->type != OBJT_VNODE)) { 1040 return NULL; 1041 } 1042 printf("vnode_pager_lock: retrying\n"); 1043 } 1044 return object->handle; 1045 } 1046 return NULL; 1047 } 1048