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 1.21 2006/02/17 19:18:08 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/kernel.h> 58 #include <sys/proc.h> 59 #include <sys/vnode.h> 60 #include <sys/mount.h> 61 #include <sys/buf.h> 62 #include <sys/vmmeter.h> 63 #include <sys/conf.h> 64 #include <sys/sfbuf.h> 65 #include <sys/thread2.h> 66 67 #include <vm/vm.h> 68 #include <vm/vm_object.h> 69 #include <vm/vm_page.h> 70 #include <vm/vm_pager.h> 71 #include <vm/vm_map.h> 72 #include <vm/vnode_pager.h> 73 #include <vm/vm_extern.h> 74 75 static vm_offset_t vnode_pager_addr (struct vnode *vp, vm_ooffset_t address, 76 int *run); 77 static void vnode_pager_iodone (struct bio *bio); 78 static int vnode_pager_input_smlfs (vm_object_t object, vm_page_t m); 79 static int vnode_pager_input_old (vm_object_t object, vm_page_t m); 80 static void vnode_pager_dealloc (vm_object_t); 81 static int vnode_pager_getpages (vm_object_t, vm_page_t *, int, int); 82 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *); 83 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t, int *, int *); 84 85 struct pagerops vnodepagerops = { 86 NULL, 87 vnode_pager_alloc, 88 vnode_pager_dealloc, 89 vnode_pager_getpages, 90 vnode_pager_putpages, 91 vnode_pager_haspage, 92 NULL 93 }; 94 95 int vnode_pbuf_freecnt = -1; /* start out unlimited */ 96 97 /* 98 * Allocate (or lookup) pager for a vnode. 99 * Handle is a vnode pointer. 100 */ 101 vm_object_t 102 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 103 vm_ooffset_t offset) 104 { 105 vm_object_t object; 106 struct vnode *vp; 107 108 /* 109 * Pageout to vnode, no can do yet. 110 */ 111 if (handle == NULL) 112 return (NULL); 113 114 /* 115 * XXX hack - This initialization should be put somewhere else. 116 */ 117 if (vnode_pbuf_freecnt < 0) { 118 vnode_pbuf_freecnt = nswbuf / 2 + 1; 119 } 120 121 vp = (struct vnode *) handle; 122 123 /* 124 * Prevent race condition when allocating the object. This 125 * can happen with NFS vnodes since the nfsnode isn't locked. 126 */ 127 while (vp->v_flag & VOLOCK) { 128 vp->v_flag |= VOWANT; 129 tsleep(vp, 0, "vnpobj", 0); 130 } 131 vp->v_flag |= VOLOCK; 132 133 /* 134 * If the object is being terminated, wait for it to 135 * go away. 136 */ 137 while (((object = vp->v_object) != NULL) && 138 (object->flags & OBJ_DEAD)) { 139 tsleep(object, 0, "vadead", 0); 140 } 141 142 if (vp->v_usecount == 0) 143 panic("vnode_pager_alloc: no vnode reference"); 144 145 if (object == NULL) { 146 /* 147 * And an object of the appropriate size 148 */ 149 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); 150 object->flags = 0; 151 152 object->un_pager.vnp.vnp_size = size; 153 154 object->handle = handle; 155 vp->v_object = object; 156 vp->v_usecount++; 157 } else { 158 object->ref_count++; 159 vp->v_usecount++; 160 } 161 162 vp->v_flag &= ~VOLOCK; 163 if (vp->v_flag & VOWANT) { 164 vp->v_flag &= ~VOWANT; 165 wakeup(vp); 166 } 167 return (object); 168 } 169 170 static void 171 vnode_pager_dealloc(vm_object_t object) 172 { 173 struct vnode *vp = object->handle; 174 175 if (vp == NULL) 176 panic("vnode_pager_dealloc: pager already dealloced"); 177 178 vm_object_pip_wait(object, "vnpdea"); 179 180 object->handle = NULL; 181 object->type = OBJT_DEAD; 182 vp->v_object = NULL; 183 vp->v_flag &= ~(VTEXT | VOBJBUF); 184 } 185 186 static boolean_t 187 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 188 int *after) 189 { 190 struct vnode *vp = object->handle; 191 daddr_t bn; 192 int err; 193 daddr_t reqblock; 194 int poff; 195 int bsize; 196 int pagesperblock, blocksperpage; 197 198 /* 199 * If no vp or vp is doomed or marked transparent to VM, we do not 200 * have the page. 201 */ 202 if ((vp == NULL) || (vp->v_flag & VRECLAIMED)) 203 return FALSE; 204 205 /* 206 * If filesystem no longer mounted or offset beyond end of file we do 207 * not have the page. 208 */ 209 if ((vp->v_mount == NULL) || 210 (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)) 211 return FALSE; 212 213 bsize = vp->v_mount->mnt_stat.f_iosize; 214 pagesperblock = bsize / PAGE_SIZE; 215 blocksperpage = 0; 216 if (pagesperblock > 0) { 217 reqblock = pindex / pagesperblock; 218 } else { 219 blocksperpage = (PAGE_SIZE / bsize); 220 reqblock = pindex * blocksperpage; 221 } 222 err = VOP_BMAP(vp, reqblock, (struct vnode **) 0, &bn, 223 after, before); 224 if (err) 225 return TRUE; 226 if ( bn == -1) 227 return FALSE; 228 if (pagesperblock > 0) { 229 poff = pindex - (reqblock * pagesperblock); 230 if (before) { 231 *before *= pagesperblock; 232 *before += poff; 233 } 234 if (after) { 235 int numafter; 236 *after *= pagesperblock; 237 numafter = pagesperblock - (poff + 1); 238 if (IDX_TO_OFF(pindex + numafter) > object->un_pager.vnp.vnp_size) { 239 numafter = OFF_TO_IDX((object->un_pager.vnp.vnp_size - IDX_TO_OFF(pindex))); 240 } 241 *after += numafter; 242 } 243 } else { 244 if (before) { 245 *before /= blocksperpage; 246 } 247 248 if (after) { 249 *after /= blocksperpage; 250 } 251 } 252 return TRUE; 253 } 254 255 /* 256 * Lets the VM system know about a change in size for a file. 257 * We adjust our own internal size and flush any cached pages in 258 * the associated object that are affected by the size change. 259 * 260 * Note: this routine may be invoked as a result of a pager put 261 * operation (possibly at object termination time), so we must be careful. 262 */ 263 void 264 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize) 265 { 266 vm_pindex_t nobjsize; 267 vm_object_t object = vp->v_object; 268 269 if (object == NULL) 270 return; 271 272 /* 273 * Hasn't changed size 274 */ 275 if (nsize == object->un_pager.vnp.vnp_size) 276 return; 277 278 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 279 280 /* 281 * File has shrunk. Toss any cached pages beyond the new EOF. 282 */ 283 if (nsize < object->un_pager.vnp.vnp_size) { 284 vm_freeze_copyopts(object, OFF_TO_IDX(nsize), object->size); 285 if (nobjsize < object->size) { 286 vm_object_page_remove(object, nobjsize, object->size, 287 FALSE); 288 } 289 /* 290 * This gets rid of garbage at the end of a page that is now 291 * only partially backed by the vnode. Since we are setting 292 * the entire page valid & clean after we are done we have 293 * to be sure that the portion of the page within the file 294 * bounds is already valid. If it isn't then making it 295 * valid would create a corrupt block. 296 */ 297 if (nsize & PAGE_MASK) { 298 vm_offset_t kva; 299 vm_page_t m; 300 301 m = vm_page_lookup(object, OFF_TO_IDX(nsize)); 302 if (m && m->valid) { 303 int base = (int)nsize & PAGE_MASK; 304 int size = PAGE_SIZE - base; 305 struct sf_buf *sf; 306 307 /* 308 * Clear out partial-page garbage in case 309 * the page has been mapped. 310 */ 311 sf = sf_buf_alloc(m, SFB_CPUPRIVATE); 312 kva = sf_buf_kva(sf); 313 bzero((caddr_t)kva + base, size); 314 sf_buf_free(sf); 315 316 /* 317 * XXX work around SMP data integrity race 318 * by unmapping the page from user processes. 319 * The garbage we just cleared may be mapped 320 * to a user process running on another cpu 321 * and this code is not running through normal 322 * I/O channels which handle SMP issues for 323 * us, so unmap page to synchronize all cpus. 324 * 325 * XXX should vm_pager_unmap_page() have 326 * dealt with this? 327 */ 328 vm_page_protect(m, VM_PROT_NONE); 329 330 /* 331 * Clear out partial-page dirty bits. This 332 * has the side effect of setting the valid 333 * bits, but that is ok. There are a bunch 334 * of places in the VM system where we expected 335 * m->dirty == VM_PAGE_BITS_ALL. The file EOF 336 * case is one of them. If the page is still 337 * partially dirty, make it fully dirty. 338 * 339 * note that we do not clear out the valid 340 * bits. This would prevent bogus_page 341 * replacement from working properly. 342 */ 343 vm_page_set_validclean(m, base, size); 344 if (m->dirty != 0) 345 m->dirty = VM_PAGE_BITS_ALL; 346 } 347 } 348 } 349 object->un_pager.vnp.vnp_size = nsize; 350 object->size = nobjsize; 351 } 352 353 void 354 vnode_pager_freepage(vm_page_t m) 355 { 356 vm_page_free(m); 357 } 358 359 /* 360 * calculate the linear (byte) disk address of specified virtual 361 * file address 362 */ 363 static vm_offset_t 364 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, int *run) 365 { 366 int rtaddress; 367 int bsize; 368 daddr_t block; 369 struct vnode *rtvp; 370 int err; 371 daddr_t vblock; 372 int voffset; 373 374 if (address < 0) 375 return -1; 376 377 if (vp->v_mount == NULL) 378 return -1; 379 380 bsize = vp->v_mount->mnt_stat.f_iosize; 381 vblock = address / bsize; 382 voffset = address % bsize; 383 384 err = VOP_BMAP(vp, vblock, &rtvp, &block, run, NULL); 385 386 if (err || (block == -1)) 387 rtaddress = -1; 388 else { 389 rtaddress = block + voffset / DEV_BSIZE; 390 if( run) { 391 *run += 1; 392 *run *= bsize/PAGE_SIZE; 393 *run -= voffset/PAGE_SIZE; 394 } 395 } 396 397 return rtaddress; 398 } 399 400 /* 401 * interrupt routine for I/O completion 402 */ 403 static void 404 vnode_pager_iodone(struct bio *bio) 405 { 406 struct buf *bp = bio->bio_buf; 407 408 bp->b_flags |= B_DONE; 409 wakeup(bp); 410 } 411 412 /* 413 * small block file system vnode pager input 414 */ 415 static int 416 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m) 417 { 418 int i; 419 struct vnode *dp, *vp; 420 struct buf *bp; 421 vm_offset_t kva; 422 struct sf_buf *sf; 423 int fileaddr; 424 vm_offset_t bsize; 425 int error = 0; 426 427 vp = object->handle; 428 if (vp->v_mount == NULL) 429 return VM_PAGER_BAD; 430 431 bsize = vp->v_mount->mnt_stat.f_iosize; 432 433 434 VOP_BMAP(vp, 0, &dp, 0, NULL, NULL); 435 436 sf = sf_buf_alloc(m, 0); 437 kva = sf_buf_kva(sf); 438 439 for (i = 0; i < PAGE_SIZE / bsize; i++) { 440 vm_ooffset_t address; 441 442 if (vm_page_bits(i * bsize, bsize) & m->valid) 443 continue; 444 445 address = IDX_TO_OFF(m->pindex) + i * bsize; 446 if (address >= object->un_pager.vnp.vnp_size) { 447 fileaddr = -1; 448 } else { 449 fileaddr = vnode_pager_addr(vp, address, NULL); 450 } 451 if (fileaddr != -1) { 452 bp = getpbuf(&vnode_pbuf_freecnt); 453 454 /* build a minimal buffer header */ 455 bp->b_flags = B_READ; 456 bp->b_data = (caddr_t) kva + i * bsize; 457 bp->b_bio1.bio_done = vnode_pager_iodone; 458 bp->b_bio1.bio_blkno = fileaddr; 459 pbgetvp(dp, bp); 460 bp->b_bcount = bsize; 461 bp->b_bufsize = bsize; 462 bp->b_runningbufspace = bp->b_bufsize; 463 runningbufspace += bp->b_runningbufspace; 464 465 /* do the input */ 466 vn_strategy(dp, &bp->b_bio1); 467 468 /* we definitely need to be at splvm here */ 469 470 crit_enter(); 471 while ((bp->b_flags & B_DONE) == 0) { 472 tsleep(bp, 0, "vnsrd", 0); 473 } 474 crit_exit(); 475 if ((bp->b_flags & B_ERROR) != 0) 476 error = EIO; 477 478 /* 479 * free the buffer header back to the swap buffer pool 480 */ 481 relpbuf(bp, &vnode_pbuf_freecnt); 482 if (error) 483 break; 484 485 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 486 } else { 487 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 488 bzero((caddr_t) kva + i * bsize, bsize); 489 } 490 } 491 sf_buf_free(sf); 492 pmap_clear_modify(m); 493 vm_page_flag_clear(m, PG_ZERO); 494 if (error) { 495 return VM_PAGER_ERROR; 496 } 497 return VM_PAGER_OK; 498 499 } 500 501 502 /* 503 * old style vnode pager output routine 504 */ 505 static int 506 vnode_pager_input_old(vm_object_t object, vm_page_t m) 507 { 508 struct uio auio; 509 struct iovec aiov; 510 int error; 511 int size; 512 vm_offset_t kva; 513 struct sf_buf *sf; 514 515 error = 0; 516 517 /* 518 * Return failure if beyond current EOF 519 */ 520 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 521 return VM_PAGER_BAD; 522 } else { 523 size = PAGE_SIZE; 524 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 525 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 526 527 /* 528 * Allocate a kernel virtual address and initialize so that 529 * we can use VOP_READ/WRITE routines. 530 */ 531 sf = sf_buf_alloc(m, 0); 532 kva = sf_buf_kva(sf); 533 534 aiov.iov_base = (caddr_t) kva; 535 aiov.iov_len = size; 536 auio.uio_iov = &aiov; 537 auio.uio_iovcnt = 1; 538 auio.uio_offset = IDX_TO_OFF(m->pindex); 539 auio.uio_segflg = UIO_SYSSPACE; 540 auio.uio_rw = UIO_READ; 541 auio.uio_resid = size; 542 auio.uio_td = curthread; 543 544 error = VOP_READ(((struct vnode *)object->handle), 545 &auio, 0, proc0.p_ucred); 546 if (!error) { 547 int count = size - auio.uio_resid; 548 549 if (count == 0) 550 error = EINVAL; 551 else if (count != PAGE_SIZE) 552 bzero((caddr_t) kva + count, PAGE_SIZE - count); 553 } 554 sf_buf_free(sf); 555 } 556 pmap_clear_modify(m); 557 vm_page_undirty(m); 558 vm_page_flag_clear(m, PG_ZERO); 559 if (!error) 560 m->valid = VM_PAGE_BITS_ALL; 561 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 562 } 563 564 /* 565 * generic vnode pager input routine 566 */ 567 568 /* 569 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 570 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to 571 * vnode_pager_generic_getpages() to implement the previous behaviour. 572 * 573 * All other FS's should use the bypass to get to the local media 574 * backing vp's VOP_GETPAGES. 575 */ 576 static int 577 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage) 578 { 579 int rtval; 580 struct vnode *vp; 581 int bytes = count * PAGE_SIZE; 582 583 vp = object->handle; 584 /* 585 * XXX temporary diagnostic message to help track stale FS code, 586 * Returning EOPNOTSUPP from here may make things unhappy. 587 */ 588 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0); 589 if (rtval == EOPNOTSUPP) { 590 printf("vnode_pager: *** WARNING *** stale FS getpages\n"); 591 rtval = vnode_pager_generic_getpages( vp, m, bytes, reqpage); 592 } 593 return rtval; 594 } 595 596 597 /* 598 * This is now called from local media FS's to operate against their 599 * own vnodes if they fail to implement VOP_GETPAGES. 600 */ 601 int 602 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int bytecount, 603 int reqpage) 604 { 605 vm_object_t object; 606 vm_offset_t kva; 607 off_t foff, tfoff, nextoff; 608 int i, size, bsize, first, firstaddr; 609 struct vnode *dp; 610 int runpg; 611 int runend; 612 struct buf *bp; 613 int count; 614 int error = 0; 615 616 object = vp->v_object; 617 count = bytecount / PAGE_SIZE; 618 619 if (vp->v_mount == NULL) 620 return VM_PAGER_BAD; 621 622 bsize = vp->v_mount->mnt_stat.f_iosize; 623 624 /* get the UNDERLYING device for the file with VOP_BMAP() */ 625 626 /* 627 * originally, we did not check for an error return value -- assuming 628 * an fs always has a bmap entry point -- that assumption is wrong!!! 629 */ 630 foff = IDX_TO_OFF(m[reqpage]->pindex); 631 632 /* 633 * if we can't bmap, use old VOP code 634 */ 635 if (VOP_BMAP(vp, 0, &dp, 0, NULL, NULL)) { 636 for (i = 0; i < count; i++) { 637 if (i != reqpage) { 638 vnode_pager_freepage(m[i]); 639 } 640 } 641 mycpu->gd_cnt.v_vnodein++; 642 mycpu->gd_cnt.v_vnodepgsin++; 643 return vnode_pager_input_old(object, m[reqpage]); 644 645 /* 646 * if the blocksize is smaller than a page size, then use 647 * special small filesystem code. NFS sometimes has a small 648 * blocksize, but it can handle large reads itself. 649 */ 650 } else if ((PAGE_SIZE / bsize) > 1 && 651 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) { 652 for (i = 0; i < count; i++) { 653 if (i != reqpage) { 654 vnode_pager_freepage(m[i]); 655 } 656 } 657 mycpu->gd_cnt.v_vnodein++; 658 mycpu->gd_cnt.v_vnodepgsin++; 659 return vnode_pager_input_smlfs(object, m[reqpage]); 660 } 661 662 /* 663 * If we have a completely valid page available to us, we can 664 * clean up and return. Otherwise we have to re-read the 665 * media. 666 * 667 * Note that this does not work with NFS, so NFS has its own 668 * getpages routine. The problem is that NFS can have partially 669 * valid pages associated with the buffer cache due to the piecemeal 670 * write support. If we were to fall through and re-read the media 671 * as we do here, dirty data could be lost. 672 */ 673 674 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) { 675 for (i = 0; i < count; i++) { 676 if (i != reqpage) 677 vnode_pager_freepage(m[i]); 678 } 679 return VM_PAGER_OK; 680 } 681 m[reqpage]->valid = 0; 682 683 /* 684 * here on direct device I/O 685 */ 686 687 firstaddr = -1; 688 /* 689 * calculate the run that includes the required page 690 */ 691 for(first = 0, i = 0; i < count; i = runend) { 692 firstaddr = vnode_pager_addr(vp, 693 IDX_TO_OFF(m[i]->pindex), &runpg); 694 if (firstaddr == -1) { 695 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) { 696 /* XXX no %qd in kernel. */ 697 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %d, foff: 0x%lx%08lx, vnp_size: 0x%lx%08lx", 698 firstaddr, (u_long)(foff >> 32), 699 (u_long)(u_int32_t)foff, 700 (u_long)(u_int32_t) 701 (object->un_pager.vnp.vnp_size >> 32), 702 (u_long)(u_int32_t) 703 object->un_pager.vnp.vnp_size); 704 } 705 vnode_pager_freepage(m[i]); 706 runend = i + 1; 707 first = runend; 708 continue; 709 } 710 runend = i + runpg; 711 if (runend <= reqpage) { 712 int j; 713 for (j = i; j < runend; j++) { 714 vnode_pager_freepage(m[j]); 715 } 716 } else { 717 if (runpg < (count - first)) { 718 for (i = first + runpg; i < count; i++) 719 vnode_pager_freepage(m[i]); 720 count = first + runpg; 721 } 722 break; 723 } 724 first = runend; 725 } 726 727 /* 728 * the first and last page have been calculated now, move input pages 729 * to be zero based... 730 */ 731 if (first != 0) { 732 for (i = first; i < count; i++) { 733 m[i - first] = m[i]; 734 } 735 count -= first; 736 reqpage -= first; 737 } 738 739 /* 740 * calculate the file virtual address for the transfer 741 */ 742 foff = IDX_TO_OFF(m[0]->pindex); 743 744 /* 745 * calculate the size of the transfer 746 */ 747 size = count * PAGE_SIZE; 748 if ((foff + size) > object->un_pager.vnp.vnp_size) 749 size = object->un_pager.vnp.vnp_size - foff; 750 751 /* 752 * round up physical size for real devices. 753 */ 754 if (dp->v_type == VBLK || dp->v_type == VCHR) { 755 int secmask = dp->v_rdev->si_bsize_phys - 1; 756 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1)); 757 size = (size + secmask) & ~secmask; 758 } 759 760 bp = getpbuf(&vnode_pbuf_freecnt); 761 kva = (vm_offset_t) bp->b_data; 762 763 /* 764 * and map the pages to be read into the kva 765 */ 766 pmap_qenter(kva, m, count); 767 768 /* build a minimal buffer header */ 769 bp->b_flags = B_READ; 770 bp->b_bio1.bio_done = vnode_pager_iodone; 771 bp->b_bio1.bio_blkno = firstaddr; 772 pbgetvp(dp, bp); 773 bp->b_bcount = size; 774 bp->b_bufsize = size; 775 bp->b_runningbufspace = bp->b_bufsize; 776 runningbufspace += bp->b_runningbufspace; 777 778 mycpu->gd_cnt.v_vnodein++; 779 mycpu->gd_cnt.v_vnodepgsin += count; 780 781 /* do the input */ 782 vn_strategy(dp, &bp->b_bio1); 783 784 crit_enter(); 785 /* we definitely need to be at splvm here */ 786 787 while ((bp->b_flags & B_DONE) == 0) { 788 tsleep(bp, 0, "vnread", 0); 789 } 790 crit_exit(); 791 if ((bp->b_flags & B_ERROR) != 0) 792 error = EIO; 793 794 if (!error) { 795 if (size != count * PAGE_SIZE) 796 bzero((caddr_t) kva + size, PAGE_SIZE * count - size); 797 } 798 pmap_qremove(kva, count); 799 800 /* 801 * free the buffer header back to the swap buffer pool 802 */ 803 relpbuf(bp, &vnode_pbuf_freecnt); 804 805 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) { 806 vm_page_t mt; 807 808 nextoff = tfoff + PAGE_SIZE; 809 mt = m[i]; 810 811 if (nextoff <= object->un_pager.vnp.vnp_size) { 812 /* 813 * Read filled up entire page. 814 */ 815 mt->valid = VM_PAGE_BITS_ALL; 816 vm_page_undirty(mt); /* should be an assert? XXX */ 817 pmap_clear_modify(mt); 818 } else { 819 /* 820 * Read did not fill up entire page. Since this 821 * is getpages, the page may be mapped, so we have 822 * to zero the invalid portions of the page even 823 * though we aren't setting them valid. 824 * 825 * Currently we do not set the entire page valid, 826 * we just try to clear the piece that we couldn't 827 * read. 828 */ 829 vm_page_set_validclean(mt, 0, 830 object->un_pager.vnp.vnp_size - tfoff); 831 /* handled by vm_fault now */ 832 /* vm_page_zero_invalid(mt, FALSE); */ 833 } 834 835 vm_page_flag_clear(mt, PG_ZERO); 836 if (i != reqpage) { 837 838 /* 839 * whether or not to leave the page activated is up in 840 * the air, but we should put the page on a page queue 841 * somewhere. (it already is in the object). Result: 842 * It appears that empirical results show that 843 * deactivating pages is best. 844 */ 845 846 /* 847 * just in case someone was asking for this page we 848 * now tell them that it is ok to use 849 */ 850 if (!error) { 851 if (mt->flags & PG_WANTED) 852 vm_page_activate(mt); 853 else 854 vm_page_deactivate(mt); 855 vm_page_wakeup(mt); 856 } else { 857 vnode_pager_freepage(mt); 858 } 859 } 860 } 861 if (error) { 862 printf("vnode_pager_getpages: I/O read error\n"); 863 } 864 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 865 } 866 867 /* 868 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 869 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 870 * vnode_pager_generic_putpages() to implement the previous behaviour. 871 * 872 * All other FS's should use the bypass to get to the local media 873 * backing vp's VOP_PUTPAGES. 874 */ 875 static void 876 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count, 877 boolean_t sync, int *rtvals) 878 { 879 int rtval; 880 struct vnode *vp; 881 int bytes = count * PAGE_SIZE; 882 883 /* 884 * Force synchronous operation if we are extremely low on memory 885 * to prevent a low-memory deadlock. VOP operations often need to 886 * allocate more memory to initiate the I/O ( i.e. do a BMAP 887 * operation ). The swapper handles the case by limiting the amount 888 * of asynchronous I/O, but that sort of solution doesn't scale well 889 * for the vnode pager without a lot of work. 890 * 891 * Also, the backing vnode's iodone routine may not wake the pageout 892 * daemon up. This should be probably be addressed XXX. 893 */ 894 895 if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min) 896 sync |= OBJPC_SYNC; 897 898 /* 899 * Call device-specific putpages function 900 */ 901 902 vp = object->handle; 903 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 904 if (rtval == EOPNOTSUPP) { 905 printf("vnode_pager: *** WARNING *** stale FS putpages\n"); 906 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals); 907 } 908 } 909 910 911 /* 912 * This is now called from local media FS's to operate against their 913 * own vnodes if they fail to implement VOP_PUTPAGES. 914 * 915 * This is typically called indirectly via the pageout daemon and 916 * clustering has already typically occured, so in general we ask the 917 * underlying filesystem to write the data out asynchronously rather 918 * then delayed. 919 */ 920 int 921 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount, 922 int flags, int *rtvals) 923 { 924 int i; 925 vm_object_t object; 926 int count; 927 928 int maxsize, ncount; 929 vm_ooffset_t poffset; 930 struct uio auio; 931 struct iovec aiov; 932 int error; 933 int ioflags; 934 935 object = vp->v_object; 936 count = bytecount / PAGE_SIZE; 937 938 for (i = 0; i < count; i++) 939 rtvals[i] = VM_PAGER_AGAIN; 940 941 if ((int) m[0]->pindex < 0) { 942 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n", 943 (long)m[0]->pindex, m[0]->dirty); 944 rtvals[0] = VM_PAGER_BAD; 945 return VM_PAGER_BAD; 946 } 947 948 maxsize = count * PAGE_SIZE; 949 ncount = count; 950 951 poffset = IDX_TO_OFF(m[0]->pindex); 952 953 /* 954 * If the page-aligned write is larger then the actual file we 955 * have to invalidate pages occuring beyond the file EOF. However, 956 * there is an edge case where a file may not be page-aligned where 957 * the last page is partially invalid. In this case the filesystem 958 * may not properly clear the dirty bits for the entire page (which 959 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 960 * With the page locked we are free to fix-up the dirty bits here. 961 * 962 * We do not under any circumstances truncate the valid bits, as 963 * this will screw up bogus page replacement. 964 */ 965 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 966 if (object->un_pager.vnp.vnp_size > poffset) { 967 int pgoff; 968 969 maxsize = object->un_pager.vnp.vnp_size - poffset; 970 ncount = btoc(maxsize); 971 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 972 vm_page_clear_dirty(m[ncount - 1], pgoff, 973 PAGE_SIZE - pgoff); 974 } 975 } else { 976 maxsize = 0; 977 ncount = 0; 978 } 979 if (ncount < count) { 980 for (i = ncount; i < count; i++) { 981 rtvals[i] = VM_PAGER_BAD; 982 } 983 } 984 } 985 986 /* 987 * pageouts are already clustered, use IO_ASYNC to force a bawrite() 988 * rather then a bdwrite() to prevent paging I/O from saturating 989 * the buffer cache. Dummy-up the sequential heuristic to cause 990 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 991 * the system decides how to cluster. 992 */ 993 ioflags = IO_VMIO; 994 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 995 ioflags |= IO_SYNC; 996 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 997 ioflags |= IO_ASYNC; 998 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 999 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 1000 1001 aiov.iov_base = (caddr_t) 0; 1002 aiov.iov_len = maxsize; 1003 auio.uio_iov = &aiov; 1004 auio.uio_iovcnt = 1; 1005 auio.uio_offset = poffset; 1006 auio.uio_segflg = UIO_NOCOPY; 1007 auio.uio_rw = UIO_WRITE; 1008 auio.uio_resid = maxsize; 1009 auio.uio_td = NULL; 1010 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred); 1011 mycpu->gd_cnt.v_vnodeout++; 1012 mycpu->gd_cnt.v_vnodepgsout += ncount; 1013 1014 if (error) { 1015 printf("vnode_pager_putpages: I/O error %d\n", error); 1016 } 1017 if (auio.uio_resid) { 1018 printf("vnode_pager_putpages: residual I/O %d at %lu\n", 1019 auio.uio_resid, (u_long)m[0]->pindex); 1020 } 1021 for (i = 0; i < ncount; i++) { 1022 rtvals[i] = VM_PAGER_OK; 1023 } 1024 return rtvals[0]; 1025 } 1026 1027 struct vnode * 1028 vnode_pager_lock(vm_object_t object) 1029 { 1030 struct thread *td = curthread; /* XXX */ 1031 int error; 1032 1033 for (; object != NULL; object = object->backing_object) { 1034 if (object->type != OBJT_VNODE) 1035 continue; 1036 if (object->flags & OBJ_DEAD) 1037 return NULL; 1038 1039 for (;;) { 1040 struct vnode *vp = object->handle; 1041 error = vget(vp, LK_NOPAUSE | LK_SHARED | 1042 LK_RETRY | LK_CANRECURSE, td); 1043 if (error == 0) { 1044 if (object->handle != vp) { 1045 vput(vp); 1046 continue; 1047 } 1048 return (vp); 1049 } 1050 if ((object->flags & OBJ_DEAD) || 1051 (object->type != OBJT_VNODE)) { 1052 return NULL; 1053 } 1054 printf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp, error, lockstatus(&vp->v_lock, td)); 1055 tsleep(object->handle, 0, "vnpgrl", hz); 1056 } 1057 } 1058 return NULL; 1059 } 1060