1 /* 2 * Copyright (c) 1990 University of Utah. 3 * Copyright (c) 1991 The Regents of the University of California. 4 * All rights reserved. 5 * Copyright (c) 1993, 1994 John S. Dyson 6 * Copyright (c) 1995, David Greenman 7 * 8 * This code is derived from software contributed to Berkeley by 9 * the Systems Programming Group of the University of Utah Computer 10 * Science Department. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. All advertising materials mentioning features or use of this software 21 * must display the following acknowledgement: 22 * This product includes software developed by the University of 23 * California, Berkeley and its contributors. 24 * 4. Neither the name of the University nor the names of its contributors 25 * may be used to endorse or promote products derived from this software 26 * without specific prior written permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 38 * SUCH DAMAGE. 39 * 40 * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91 41 */ 42 43 /* 44 * Page to/from files (vnodes). 45 */ 46 47 /* 48 * TODO: 49 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will 50 * greatly re-simplify the vnode_pager. 51 */ 52 53 #include <sys/cdefs.h> 54 __FBSDID("$FreeBSD$"); 55 56 #include <sys/param.h> 57 #include <sys/systm.h> 58 #include <sys/proc.h> 59 #include <sys/vnode.h> 60 #include <sys/mount.h> 61 #include <sys/bio.h> 62 #include <sys/buf.h> 63 #include <sys/vmmeter.h> 64 #include <sys/conf.h> 65 #include <sys/sf_buf.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 void vnode_pager_init(void); 76 static vm_offset_t vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, 77 int *run); 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 .pgo_init = vnode_pager_init, 87 .pgo_alloc = vnode_pager_alloc, 88 .pgo_dealloc = vnode_pager_dealloc, 89 .pgo_getpages = vnode_pager_getpages, 90 .pgo_putpages = vnode_pager_putpages, 91 .pgo_haspage = vnode_pager_haspage, 92 }; 93 94 int vnode_pbuf_freecnt; 95 96 static void 97 vnode_pager_init(void) 98 { 99 100 vnode_pbuf_freecnt = nswbuf / 2 + 1; 101 } 102 103 /* 104 * Allocate (or lookup) pager for a vnode. 105 * Handle is a vnode pointer. 106 * 107 * MPSAFE 108 */ 109 vm_object_t 110 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 111 vm_ooffset_t offset) 112 { 113 vm_object_t object; 114 struct vnode *vp; 115 116 /* 117 * Pageout to vnode, no can do yet. 118 */ 119 if (handle == NULL) 120 return (NULL); 121 122 vp = (struct vnode *) handle; 123 124 ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc"); 125 126 /* 127 * Prevent race condition when allocating the object. This 128 * can happen with NFS vnodes since the nfsnode isn't locked. 129 */ 130 VI_LOCK(vp); 131 while (vp->v_iflag & VI_OLOCK) { 132 vp->v_iflag |= VI_OWANT; 133 msleep(vp, VI_MTX(vp), PVM, "vnpobj", 0); 134 } 135 vp->v_iflag |= VI_OLOCK; 136 VI_UNLOCK(vp); 137 138 /* 139 * If the object is being terminated, wait for it to 140 * go away. 141 */ 142 while ((object = vp->v_object) != NULL) { 143 VM_OBJECT_LOCK(object); 144 if ((object->flags & OBJ_DEAD) == 0) 145 break; 146 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vadead", 0); 147 } 148 149 if (vp->v_usecount == 0) 150 panic("vnode_pager_alloc: no vnode reference"); 151 152 if (object == NULL) { 153 /* 154 * And an object of the appropriate size 155 */ 156 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); 157 158 object->un_pager.vnp.vnp_size = size; 159 160 object->handle = handle; 161 vp->v_object = object; 162 } else { 163 object->ref_count++; 164 VM_OBJECT_UNLOCK(object); 165 } 166 VI_LOCK(vp); 167 vp->v_usecount++; 168 vp->v_iflag &= ~VI_OLOCK; 169 if (vp->v_iflag & VI_OWANT) { 170 vp->v_iflag &= ~VI_OWANT; 171 wakeup(vp); 172 } 173 VI_UNLOCK(vp); 174 return (object); 175 } 176 177 /* 178 * The object must be locked. 179 */ 180 static void 181 vnode_pager_dealloc(object) 182 vm_object_t object; 183 { 184 struct vnode *vp = object->handle; 185 186 if (vp == NULL) 187 panic("vnode_pager_dealloc: pager already dealloced"); 188 189 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 190 vm_object_pip_wait(object, "vnpdea"); 191 192 object->handle = NULL; 193 object->type = OBJT_DEAD; 194 ASSERT_VOP_LOCKED(vp, "vnode_pager_dealloc"); 195 vp->v_object = NULL; 196 vp->v_vflag &= ~(VV_TEXT | VV_OBJBUF); 197 } 198 199 static boolean_t 200 vnode_pager_haspage(object, pindex, before, after) 201 vm_object_t object; 202 vm_pindex_t pindex; 203 int *before; 204 int *after; 205 { 206 struct vnode *vp = object->handle; 207 daddr_t bn; 208 int err; 209 daddr_t reqblock; 210 int poff; 211 int bsize; 212 int pagesperblock, blocksperpage; 213 214 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 215 /* 216 * If no vp or vp is doomed or marked transparent to VM, we do not 217 * have the page. 218 */ 219 if (vp == NULL) 220 return FALSE; 221 222 VI_LOCK(vp); 223 if (vp->v_iflag & VI_DOOMED) { 224 VI_UNLOCK(vp); 225 return FALSE; 226 } 227 VI_UNLOCK(vp); 228 /* 229 * If filesystem no longer mounted or offset beyond end of file we do 230 * not have the page. 231 */ 232 if ((vp->v_mount == NULL) || 233 (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)) 234 return FALSE; 235 236 bsize = vp->v_mount->mnt_stat.f_iosize; 237 pagesperblock = bsize / PAGE_SIZE; 238 blocksperpage = 0; 239 if (pagesperblock > 0) { 240 reqblock = pindex / pagesperblock; 241 } else { 242 blocksperpage = (PAGE_SIZE / bsize); 243 reqblock = pindex * blocksperpage; 244 } 245 VM_OBJECT_UNLOCK(object); 246 mtx_lock(&Giant); 247 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before); 248 mtx_unlock(&Giant); 249 VM_OBJECT_LOCK(object); 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 */ 292 void 293 vnode_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 m->valid != 0) { 330 int base = (int)nsize & PAGE_MASK; 331 int size = PAGE_SIZE - base; 332 333 /* 334 * Clear out partial-page garbage in case 335 * the page has been mapped. 336 */ 337 pmap_zero_page_area(m, base, size); 338 339 /* 340 * XXX work around SMP data integrity race 341 * by unmapping the page from user processes. 342 * The garbage we just cleared may be mapped 343 * to a user process running on another cpu 344 * and this code is not running through normal 345 * I/O channels which handle SMP issues for 346 * us, so unmap page to synchronize all cpus. 347 * 348 * XXX should vm_pager_unmap_page() have 349 * dealt with this? 350 */ 351 vm_page_lock_queues(); 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 vm_page_unlock_queues(); 371 } 372 } 373 object->un_pager.vnp.vnp_size = nsize; 374 object->size = nobjsize; 375 VM_OBJECT_UNLOCK(object); 376 } 377 378 /* 379 * calculate the linear (byte) disk address of specified virtual 380 * file address 381 */ 382 static vm_offset_t 383 vnode_pager_addr(vp, address, run) 384 struct vnode *vp; 385 vm_ooffset_t address; 386 int *run; 387 { 388 int rtaddress; 389 int bsize; 390 daddr_t block; 391 int err; 392 daddr_t vblock; 393 int voffset; 394 395 GIANT_REQUIRED; 396 if ((int) address < 0) 397 return -1; 398 399 if (vp->v_mount == NULL) 400 return -1; 401 402 bsize = vp->v_mount->mnt_stat.f_iosize; 403 vblock = address / bsize; 404 voffset = address % bsize; 405 406 err = VOP_BMAP(vp, vblock, NULL, &block, run, NULL); 407 408 if (err || (block == -1)) 409 rtaddress = -1; 410 else { 411 rtaddress = block + voffset / DEV_BSIZE; 412 if (run) { 413 *run += 1; 414 *run *= bsize/PAGE_SIZE; 415 *run -= voffset/PAGE_SIZE; 416 } 417 } 418 419 return rtaddress; 420 } 421 422 /* 423 * small block filesystem vnode pager input 424 */ 425 static int 426 vnode_pager_input_smlfs(object, m) 427 vm_object_t object; 428 vm_page_t m; 429 { 430 int i; 431 struct vnode *dp, *vp; 432 struct buf *bp; 433 struct sf_buf *sf; 434 int fileaddr; 435 vm_offset_t bsize; 436 int error = 0; 437 438 GIANT_REQUIRED; 439 440 vp = object->handle; 441 if (vp->v_mount == NULL) 442 return VM_PAGER_BAD; 443 444 bsize = vp->v_mount->mnt_stat.f_iosize; 445 446 VOP_BMAP(vp, 0, &dp, 0, NULL, NULL); 447 448 sf = sf_buf_alloc(m, 0); 449 450 for (i = 0; i < PAGE_SIZE / bsize; i++) { 451 vm_ooffset_t address; 452 453 if (vm_page_bits(i * bsize, bsize) & m->valid) 454 continue; 455 456 address = IDX_TO_OFF(m->pindex) + i * bsize; 457 if (address >= object->un_pager.vnp.vnp_size) { 458 fileaddr = -1; 459 } else { 460 fileaddr = vnode_pager_addr(vp, address, NULL); 461 } 462 if (fileaddr != -1) { 463 bp = getpbuf(&vnode_pbuf_freecnt); 464 465 /* build a minimal buffer header */ 466 bp->b_iocmd = BIO_READ; 467 bp->b_iodone = bdone; 468 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 469 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 470 bp->b_rcred = crhold(curthread->td_ucred); 471 bp->b_wcred = crhold(curthread->td_ucred); 472 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize; 473 bp->b_blkno = fileaddr; 474 pbgetvp(dp, bp); 475 bp->b_bcount = bsize; 476 bp->b_bufsize = bsize; 477 bp->b_runningbufspace = bp->b_bufsize; 478 runningbufspace += bp->b_runningbufspace; 479 480 /* do the input */ 481 bp->b_iooffset = dbtob(bp->b_blkno); 482 if (dp->v_type == VCHR) 483 VOP_SPECSTRATEGY(bp->b_vp, bp); 484 else 485 VOP_STRATEGY(bp->b_vp, bp); 486 487 /* we definitely need to be at splvm here */ 488 489 bwait(bp, PVM, "vnsrd"); 490 491 if ((bp->b_ioflags & BIO_ERROR) != 0) 492 error = EIO; 493 494 /* 495 * free the buffer header back to the swap buffer pool 496 */ 497 relpbuf(bp, &vnode_pbuf_freecnt); 498 if (error) 499 break; 500 501 VM_OBJECT_LOCK(object); 502 vm_page_lock_queues(); 503 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 504 vm_page_unlock_queues(); 505 VM_OBJECT_UNLOCK(object); 506 } else { 507 VM_OBJECT_LOCK(object); 508 vm_page_lock_queues(); 509 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 510 vm_page_unlock_queues(); 511 VM_OBJECT_UNLOCK(object); 512 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize); 513 } 514 } 515 sf_buf_free(sf); 516 vm_page_lock_queues(); 517 pmap_clear_modify(m); 518 vm_page_unlock_queues(); 519 if (error) { 520 return VM_PAGER_ERROR; 521 } 522 return VM_PAGER_OK; 523 524 } 525 526 527 /* 528 * old style vnode pager output routine 529 */ 530 static int 531 vnode_pager_input_old(object, m) 532 vm_object_t object; 533 vm_page_t m; 534 { 535 struct uio auio; 536 struct iovec aiov; 537 int error; 538 int size; 539 struct sf_buf *sf; 540 struct vnode *vp; 541 542 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 543 error = 0; 544 545 /* 546 * Return failure if beyond current EOF 547 */ 548 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 549 return VM_PAGER_BAD; 550 } else { 551 size = PAGE_SIZE; 552 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 553 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 554 vp = object->handle; 555 VM_OBJECT_UNLOCK(object); 556 557 /* 558 * Allocate a kernel virtual address and initialize so that 559 * we can use VOP_READ/WRITE routines. 560 */ 561 sf = sf_buf_alloc(m, 0); 562 563 aiov.iov_base = (caddr_t)sf_buf_kva(sf); 564 aiov.iov_len = size; 565 auio.uio_iov = &aiov; 566 auio.uio_iovcnt = 1; 567 auio.uio_offset = IDX_TO_OFF(m->pindex); 568 auio.uio_segflg = UIO_SYSSPACE; 569 auio.uio_rw = UIO_READ; 570 auio.uio_resid = size; 571 auio.uio_td = curthread; 572 573 error = VOP_READ(vp, &auio, 0, curthread->td_ucred); 574 if (!error) { 575 int count = size - auio.uio_resid; 576 577 if (count == 0) 578 error = EINVAL; 579 else if (count != PAGE_SIZE) 580 bzero((caddr_t)sf_buf_kva(sf) + count, 581 PAGE_SIZE - count); 582 } 583 sf_buf_free(sf); 584 585 VM_OBJECT_LOCK(object); 586 } 587 vm_page_lock_queues(); 588 pmap_clear_modify(m); 589 vm_page_undirty(m); 590 vm_page_unlock_queues(); 591 if (!error) 592 m->valid = VM_PAGE_BITS_ALL; 593 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 594 } 595 596 /* 597 * generic vnode pager input routine 598 */ 599 600 /* 601 * Local media VFS's that do not implement their own VOP_GETPAGES 602 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages() 603 * to implement the previous behaviour. 604 * 605 * All other FS's should use the bypass to get to the local media 606 * backing vp's VOP_GETPAGES. 607 */ 608 static int 609 vnode_pager_getpages(object, m, count, reqpage) 610 vm_object_t object; 611 vm_page_t *m; 612 int count; 613 int reqpage; 614 { 615 int rtval; 616 struct vnode *vp; 617 int bytes = count * PAGE_SIZE; 618 619 vp = object->handle; 620 VM_OBJECT_UNLOCK(object); 621 mtx_lock(&Giant); 622 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0); 623 KASSERT(rtval != EOPNOTSUPP, 624 ("vnode_pager: FS getpages not implemented\n")); 625 mtx_unlock(&Giant); 626 VM_OBJECT_LOCK(object); 627 return rtval; 628 } 629 630 /* 631 * This is now called from local media FS's to operate against their 632 * own vnodes if they fail to implement VOP_GETPAGES. 633 */ 634 int 635 vnode_pager_generic_getpages(vp, m, bytecount, reqpage) 636 struct vnode *vp; 637 vm_page_t *m; 638 int bytecount; 639 int reqpage; 640 { 641 vm_object_t object; 642 vm_offset_t kva; 643 off_t foff, tfoff, nextoff; 644 int i, j, size, bsize, first, firstaddr; 645 struct vnode *dp; 646 int runpg; 647 int runend; 648 struct buf *bp; 649 int count; 650 int error = 0; 651 652 GIANT_REQUIRED; 653 object = vp->v_object; 654 count = bytecount / PAGE_SIZE; 655 656 if (vp->v_mount == NULL) 657 return VM_PAGER_BAD; 658 659 bsize = vp->v_mount->mnt_stat.f_iosize; 660 661 /* get the UNDERLYING device for the file with VOP_BMAP() */ 662 663 /* 664 * originally, we did not check for an error return value -- assuming 665 * an fs always has a bmap entry point -- that assumption is wrong!!! 666 */ 667 foff = IDX_TO_OFF(m[reqpage]->pindex); 668 669 /* 670 * if we can't bmap, use old VOP code 671 */ 672 if (VOP_BMAP(vp, 0, &dp, 0, NULL, NULL)) { 673 VM_OBJECT_LOCK(object); 674 vm_page_lock_queues(); 675 for (i = 0; i < count; i++) 676 if (i != reqpage) 677 vm_page_free(m[i]); 678 vm_page_unlock_queues(); 679 cnt.v_vnodein++; 680 cnt.v_vnodepgsin++; 681 error = vnode_pager_input_old(object, m[reqpage]); 682 VM_OBJECT_UNLOCK(object); 683 return (error); 684 685 /* 686 * if the blocksize is smaller than a page size, then use 687 * special small filesystem code. NFS sometimes has a small 688 * blocksize, but it can handle large reads itself. 689 */ 690 } else if ((PAGE_SIZE / bsize) > 1 && 691 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) { 692 VM_OBJECT_LOCK(object); 693 vm_page_lock_queues(); 694 for (i = 0; i < count; i++) 695 if (i != reqpage) 696 vm_page_free(m[i]); 697 vm_page_unlock_queues(); 698 VM_OBJECT_UNLOCK(object); 699 cnt.v_vnodein++; 700 cnt.v_vnodepgsin++; 701 return vnode_pager_input_smlfs(object, m[reqpage]); 702 } 703 704 /* 705 * If we have a completely valid page available to us, we can 706 * clean up and return. Otherwise we have to re-read the 707 * media. 708 */ 709 VM_OBJECT_LOCK(object); 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 VM_OBJECT_UNLOCK(object); 717 return VM_PAGER_OK; 718 } 719 m[reqpage]->valid = 0; 720 VM_OBJECT_UNLOCK(object); 721 722 /* 723 * here on direct device I/O 724 */ 725 firstaddr = -1; 726 727 /* 728 * calculate the run that includes the required page 729 */ 730 for (first = 0, i = 0; i < count; i = runend) { 731 firstaddr = vnode_pager_addr(vp, 732 IDX_TO_OFF(m[i]->pindex), &runpg); 733 if (firstaddr == -1) { 734 VM_OBJECT_LOCK(object); 735 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) { 736 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %d, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx", 737 firstaddr, (uintmax_t)(foff >> 32), 738 (uintmax_t)foff, 739 (uintmax_t) 740 (object->un_pager.vnp.vnp_size >> 32), 741 (uintmax_t)object->un_pager.vnp.vnp_size); 742 } 743 vm_page_lock_queues(); 744 vm_page_free(m[i]); 745 vm_page_unlock_queues(); 746 VM_OBJECT_UNLOCK(object); 747 runend = i + 1; 748 first = runend; 749 continue; 750 } 751 runend = i + runpg; 752 if (runend <= reqpage) { 753 VM_OBJECT_LOCK(object); 754 vm_page_lock_queues(); 755 for (j = i; j < runend; j++) 756 vm_page_free(m[j]); 757 vm_page_unlock_queues(); 758 VM_OBJECT_UNLOCK(object); 759 } else { 760 if (runpg < (count - first)) { 761 VM_OBJECT_LOCK(object); 762 vm_page_lock_queues(); 763 for (i = first + runpg; i < count; i++) 764 vm_page_free(m[i]); 765 vm_page_unlock_queues(); 766 VM_OBJECT_UNLOCK(object); 767 count = first + runpg; 768 } 769 break; 770 } 771 first = runend; 772 } 773 774 /* 775 * the first and last page have been calculated now, move input pages 776 * to be zero based... 777 */ 778 if (first != 0) { 779 for (i = first; i < count; i++) { 780 m[i - first] = m[i]; 781 } 782 count -= first; 783 reqpage -= first; 784 } 785 786 /* 787 * calculate the file virtual address for the transfer 788 */ 789 foff = IDX_TO_OFF(m[0]->pindex); 790 791 /* 792 * calculate the size of the transfer 793 */ 794 size = count * PAGE_SIZE; 795 if ((foff + size) > object->un_pager.vnp.vnp_size) 796 size = object->un_pager.vnp.vnp_size - foff; 797 798 /* 799 * round up physical size for real devices. 800 */ 801 if (dp->v_type == VBLK || dp->v_type == VCHR) { 802 int secmask = dp->v_rdev->si_bsize_phys - 1; 803 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1)); 804 size = (size + secmask) & ~secmask; 805 } 806 807 bp = getpbuf(&vnode_pbuf_freecnt); 808 kva = (vm_offset_t) bp->b_data; 809 810 /* 811 * and map the pages to be read into the kva 812 */ 813 pmap_qenter(kva, m, count); 814 815 /* build a minimal buffer header */ 816 bp->b_iocmd = BIO_READ; 817 bp->b_iodone = bdone; 818 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 819 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 820 bp->b_rcred = crhold(curthread->td_ucred); 821 bp->b_wcred = crhold(curthread->td_ucred); 822 bp->b_blkno = firstaddr; 823 pbgetvp(dp, bp); 824 bp->b_bcount = size; 825 bp->b_bufsize = size; 826 bp->b_runningbufspace = bp->b_bufsize; 827 runningbufspace += bp->b_runningbufspace; 828 829 cnt.v_vnodein++; 830 cnt.v_vnodepgsin += count; 831 832 /* do the input */ 833 bp->b_iooffset = dbtob(bp->b_blkno); 834 if (dp->v_type == VCHR) 835 VOP_SPECSTRATEGY(bp->b_vp, bp); 836 else 837 VOP_STRATEGY(bp->b_vp, bp); 838 839 bwait(bp, PVM, "vnread"); 840 841 if ((bp->b_ioflags & BIO_ERROR) != 0) 842 error = EIO; 843 844 if (!error) { 845 if (size != count * PAGE_SIZE) 846 bzero((caddr_t) kva + size, PAGE_SIZE * count - size); 847 } 848 pmap_qremove(kva, count); 849 850 /* 851 * free the buffer header back to the swap buffer pool 852 */ 853 relpbuf(bp, &vnode_pbuf_freecnt); 854 855 VM_OBJECT_LOCK(object); 856 vm_page_lock_queues(); 857 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) { 858 vm_page_t mt; 859 860 nextoff = tfoff + PAGE_SIZE; 861 mt = m[i]; 862 863 if (nextoff <= object->un_pager.vnp.vnp_size) { 864 /* 865 * Read filled up entire page. 866 */ 867 mt->valid = VM_PAGE_BITS_ALL; 868 vm_page_undirty(mt); /* should be an assert? XXX */ 869 pmap_clear_modify(mt); 870 } else { 871 /* 872 * Read did not fill up entire page. Since this 873 * is getpages, the page may be mapped, so we have 874 * to zero the invalid portions of the page even 875 * though we aren't setting them valid. 876 * 877 * Currently we do not set the entire page valid, 878 * we just try to clear the piece that we couldn't 879 * read. 880 */ 881 vm_page_set_validclean(mt, 0, 882 object->un_pager.vnp.vnp_size - tfoff); 883 /* handled by vm_fault now */ 884 /* vm_page_zero_invalid(mt, FALSE); */ 885 } 886 887 if (i != reqpage) { 888 889 /* 890 * whether or not to leave the page activated is up in 891 * the air, but we should put the page on a page queue 892 * somewhere. (it already is in the object). Result: 893 * It appears that empirical results show that 894 * deactivating pages is best. 895 */ 896 897 /* 898 * just in case someone was asking for this page we 899 * now tell them that it is ok to use 900 */ 901 if (!error) { 902 if (mt->flags & PG_WANTED) 903 vm_page_activate(mt); 904 else 905 vm_page_deactivate(mt); 906 vm_page_wakeup(mt); 907 } else { 908 vm_page_free(mt); 909 } 910 } 911 } 912 vm_page_unlock_queues(); 913 VM_OBJECT_UNLOCK(object); 914 if (error) { 915 printf("vnode_pager_getpages: I/O read error\n"); 916 } 917 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 918 } 919 920 /* 921 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 922 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 923 * vnode_pager_generic_putpages() to implement the previous behaviour. 924 * 925 * All other FS's should use the bypass to get to the local media 926 * backing vp's VOP_PUTPAGES. 927 */ 928 static void 929 vnode_pager_putpages(object, m, count, sync, rtvals) 930 vm_object_t object; 931 vm_page_t *m; 932 int count; 933 boolean_t sync; 934 int *rtvals; 935 { 936 int rtval; 937 struct vnode *vp; 938 struct mount *mp; 939 int bytes = count * PAGE_SIZE; 940 941 GIANT_REQUIRED; 942 /* 943 * Force synchronous operation if we are extremely low on memory 944 * to prevent a low-memory deadlock. VOP operations often need to 945 * allocate more memory to initiate the I/O ( i.e. do a BMAP 946 * operation ). The swapper handles the case by limiting the amount 947 * of asynchronous I/O, but that sort of solution doesn't scale well 948 * for the vnode pager without a lot of work. 949 * 950 * Also, the backing vnode's iodone routine may not wake the pageout 951 * daemon up. This should be probably be addressed XXX. 952 */ 953 954 if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min) 955 sync |= OBJPC_SYNC; 956 957 /* 958 * Call device-specific putpages function 959 */ 960 vp = object->handle; 961 VM_OBJECT_UNLOCK(object); 962 if (vp->v_type != VREG) 963 mp = NULL; 964 (void)vn_start_write(vp, &mp, V_WAIT); 965 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 966 KASSERT(rtval != EOPNOTSUPP, 967 ("vnode_pager: stale FS putpages\n")); 968 vn_finished_write(mp); 969 VM_OBJECT_LOCK(object); 970 } 971 972 973 /* 974 * This is now called from local media FS's to operate against their 975 * own vnodes if they fail to implement VOP_PUTPAGES. 976 * 977 * This is typically called indirectly via the pageout daemon and 978 * clustering has already typically occured, so in general we ask the 979 * underlying filesystem to write the data out asynchronously rather 980 * then delayed. 981 */ 982 int 983 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals) 984 struct vnode *vp; 985 vm_page_t *m; 986 int bytecount; 987 int flags; 988 int *rtvals; 989 { 990 int i; 991 vm_object_t object; 992 int count; 993 994 int maxsize, ncount; 995 vm_ooffset_t poffset; 996 struct uio auio; 997 struct iovec aiov; 998 int error; 999 int ioflags; 1000 1001 GIANT_REQUIRED; 1002 object = vp->v_object; 1003 count = bytecount / PAGE_SIZE; 1004 1005 for (i = 0; i < count; i++) 1006 rtvals[i] = VM_PAGER_AGAIN; 1007 1008 if ((int) m[0]->pindex < 0) { 1009 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n", 1010 (long)m[0]->pindex, (u_long)m[0]->dirty); 1011 rtvals[0] = VM_PAGER_BAD; 1012 return VM_PAGER_BAD; 1013 } 1014 1015 maxsize = count * PAGE_SIZE; 1016 ncount = count; 1017 1018 poffset = IDX_TO_OFF(m[0]->pindex); 1019 1020 /* 1021 * If the page-aligned write is larger then the actual file we 1022 * have to invalidate pages occuring beyond the file EOF. However, 1023 * there is an edge case where a file may not be page-aligned where 1024 * the last page is partially invalid. In this case the filesystem 1025 * may not properly clear the dirty bits for the entire page (which 1026 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 1027 * With the page locked we are free to fix-up the dirty bits here. 1028 * 1029 * We do not under any circumstances truncate the valid bits, as 1030 * this will screw up bogus page replacement. 1031 */ 1032 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 1033 if (object->un_pager.vnp.vnp_size > poffset) { 1034 int pgoff; 1035 1036 maxsize = object->un_pager.vnp.vnp_size - poffset; 1037 ncount = btoc(maxsize); 1038 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 1039 vm_page_lock_queues(); 1040 vm_page_clear_dirty(m[ncount - 1], pgoff, 1041 PAGE_SIZE - pgoff); 1042 vm_page_unlock_queues(); 1043 } 1044 } else { 1045 maxsize = 0; 1046 ncount = 0; 1047 } 1048 if (ncount < count) { 1049 for (i = ncount; i < count; i++) { 1050 rtvals[i] = VM_PAGER_BAD; 1051 } 1052 } 1053 } 1054 1055 /* 1056 * pageouts are already clustered, use IO_ASYNC t o force a bawrite() 1057 * rather then a bdwrite() to prevent paging I/O from saturating 1058 * the buffer cache. Dummy-up the sequential heuristic to cause 1059 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 1060 * the system decides how to cluster. 1061 */ 1062 ioflags = IO_VMIO; 1063 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 1064 ioflags |= IO_SYNC; 1065 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 1066 ioflags |= IO_ASYNC; 1067 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 1068 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 1069 1070 aiov.iov_base = (caddr_t) 0; 1071 aiov.iov_len = maxsize; 1072 auio.uio_iov = &aiov; 1073 auio.uio_iovcnt = 1; 1074 auio.uio_offset = poffset; 1075 auio.uio_segflg = UIO_NOCOPY; 1076 auio.uio_rw = UIO_WRITE; 1077 auio.uio_resid = maxsize; 1078 auio.uio_td = (struct thread *) 0; 1079 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred); 1080 cnt.v_vnodeout++; 1081 cnt.v_vnodepgsout += ncount; 1082 1083 if (error) { 1084 printf("vnode_pager_putpages: I/O error %d\n", error); 1085 } 1086 if (auio.uio_resid) { 1087 printf("vnode_pager_putpages: residual I/O %d at %lu\n", 1088 auio.uio_resid, (u_long)m[0]->pindex); 1089 } 1090 for (i = 0; i < ncount; i++) { 1091 rtvals[i] = VM_PAGER_OK; 1092 } 1093 return rtvals[0]; 1094 } 1095 1096 struct vnode * 1097 vnode_pager_lock(vm_object_t first_object) 1098 { 1099 struct vnode *vp; 1100 vm_object_t backing_object, object; 1101 1102 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 1103 for (object = first_object; object != NULL; object = backing_object) { 1104 if (object->type != OBJT_VNODE) { 1105 if ((backing_object = object->backing_object) != NULL) 1106 VM_OBJECT_LOCK(backing_object); 1107 if (object != first_object) 1108 VM_OBJECT_UNLOCK(object); 1109 continue; 1110 } 1111 retry: 1112 if (object->flags & OBJ_DEAD) { 1113 if (object != first_object) 1114 VM_OBJECT_UNLOCK(object); 1115 return NULL; 1116 } 1117 vp = object->handle; 1118 VI_LOCK(vp); 1119 VM_OBJECT_UNLOCK(object); 1120 if (first_object != object) 1121 VM_OBJECT_UNLOCK(first_object); 1122 if (vget(vp, LK_CANRECURSE | LK_INTERLOCK | LK_NOPAUSE | 1123 LK_RETRY | LK_SHARED, curthread)) { 1124 VM_OBJECT_LOCK(first_object); 1125 if (object != first_object) 1126 VM_OBJECT_LOCK(object); 1127 if (object->type != OBJT_VNODE) { 1128 if (object != first_object) 1129 VM_OBJECT_UNLOCK(object); 1130 return NULL; 1131 } 1132 printf("vnode_pager_lock: retrying\n"); 1133 goto retry; 1134 } 1135 VM_OBJECT_LOCK(first_object); 1136 return (vp); 1137 } 1138 return NULL; 1139 } 1140