vnode_pager.c revision 146340
1/*- 2 * Copyright (c) 1990 University of Utah. 3 * Copyright (c) 1991 The Regents of the University of California. 4 * All rights reserved. 5 * Copyright (c) 1993, 1994 John S. Dyson 6 * Copyright (c) 1995, David Greenman 7 * 8 * This code is derived from software contributed to Berkeley by 9 * the Systems Programming Group of the University of Utah Computer 10 * Science Department. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. All advertising materials mentioning features or use of this software 21 * must display the following acknowledgement: 22 * This product includes software developed by the University of 23 * California, Berkeley and its contributors. 24 * 4. Neither the name of the University nor the names of its contributors 25 * may be used to endorse or promote products derived from this software 26 * without specific prior written permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 38 * SUCH DAMAGE. 39 * 40 * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91 41 */ 42 43/* 44 * Page to/from files (vnodes). 45 */ 46 47/* 48 * TODO: 49 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will 50 * greatly re-simplify the vnode_pager. 51 */ 52 53#include <sys/cdefs.h> 54__FBSDID("$FreeBSD: head/sys/vm/vnode_pager.c 146340 2005-05-18 08:57:31Z bz $"); 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/limits.h> 65#include <sys/conf.h> 66#include <sys/sf_buf.h> 67 68#include <vm/vm.h> 69#include <vm/vm_object.h> 70#include <vm/vm_page.h> 71#include <vm/vm_pager.h> 72#include <vm/vm_map.h> 73#include <vm/vnode_pager.h> 74#include <vm/vm_extern.h> 75 76static void vnode_pager_init(void); 77static daddr_t vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, 78 int *run); 79static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m); 80static int vnode_pager_input_old(vm_object_t object, vm_page_t m); 81static void vnode_pager_dealloc(vm_object_t); 82static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int); 83static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 84static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *); 85static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t, vm_ooffset_t); 86 87struct pagerops vnodepagerops = { 88 .pgo_init = vnode_pager_init, 89 .pgo_alloc = vnode_pager_alloc, 90 .pgo_dealloc = vnode_pager_dealloc, 91 .pgo_getpages = vnode_pager_getpages, 92 .pgo_putpages = vnode_pager_putpages, 93 .pgo_haspage = vnode_pager_haspage, 94}; 95 96int vnode_pbuf_freecnt; 97 98static void 99vnode_pager_init(void) 100{ 101 102 vnode_pbuf_freecnt = nswbuf / 2 + 1; 103} 104 105/* Create the VM system backing object for this vnode */ 106int 107vnode_create_vobject(struct vnode *vp, size_t isize, struct thread *td) 108{ 109 vm_object_t object; 110 vm_ooffset_t size = isize; 111 struct vattr va; 112 113 if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE) 114 return (0); 115 116 while ((object = vp->v_object) != NULL) { 117 VM_OBJECT_LOCK(object); 118 if (!(object->flags & OBJ_DEAD)) { 119 VM_OBJECT_UNLOCK(object); 120 return (0); 121 } 122 VOP_UNLOCK(vp, 0, td); 123 vm_object_set_flag(object, OBJ_DISCONNECTWNT); 124 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vodead", 0); 125 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 126 } 127 128 if (size == 0) { 129 if (vn_isdisk(vp, NULL)) { 130 size = IDX_TO_OFF(INT_MAX); 131 } else { 132 if (VOP_GETATTR(vp, &va, td->td_ucred, td) != 0) 133 return (0); 134 size = va.va_size; 135 } 136 } 137 138 object = vnode_pager_alloc(vp, size, 0, 0); 139 /* 140 * Dereference the reference we just created. This assumes 141 * that the object is associated with the vp. 142 */ 143 VM_OBJECT_LOCK(object); 144 object->ref_count--; 145 VM_OBJECT_UNLOCK(object); 146 vrele(vp); 147 148 KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object")); 149 150 return (0); 151} 152 153void 154vnode_destroy_vobject(struct vnode *vp) 155{ 156 struct vm_object *obj; 157 158 obj = vp->v_object; 159 if (obj == NULL) 160 return; 161 ASSERT_VOP_LOCKED(vp, "vnode_destroy_vobject"); 162 VM_OBJECT_LOCK(obj); 163 if (obj->ref_count == 0) { 164 /* 165 * vclean() may be called twice. The first time 166 * removes the primary reference to the object, 167 * the second time goes one further and is a 168 * special-case to terminate the object. 169 * 170 * don't double-terminate the object 171 */ 172 if ((obj->flags & OBJ_DEAD) == 0) 173 vm_object_terminate(obj); 174 else 175 VM_OBJECT_UNLOCK(obj); 176 } else { 177 /* 178 * Woe to the process that tries to page now :-). 179 */ 180 vm_pager_deallocate(obj); 181 VM_OBJECT_UNLOCK(obj); 182 } 183 vp->v_object = NULL; 184} 185 186 187/* 188 * Allocate (or lookup) pager for a vnode. 189 * Handle is a vnode pointer. 190 * 191 * MPSAFE 192 */ 193vm_object_t 194vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 195 vm_ooffset_t offset) 196{ 197 vm_object_t object; 198 struct vnode *vp; 199 200 /* 201 * Pageout to vnode, no can do yet. 202 */ 203 if (handle == NULL) 204 return (NULL); 205 206 vp = (struct vnode *) handle; 207 208 ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc"); 209 210 /* 211 * If the object is being terminated, wait for it to 212 * go away. 213 */ 214 while ((object = vp->v_object) != NULL) { 215 VM_OBJECT_LOCK(object); 216 if ((object->flags & OBJ_DEAD) == 0) 217 break; 218 vm_object_set_flag(object, OBJ_DISCONNECTWNT); 219 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vadead", 0); 220 } 221 222 if (vp->v_usecount == 0) 223 panic("vnode_pager_alloc: no vnode reference"); 224 225 if (object == NULL) { 226 /* 227 * And an object of the appropriate size 228 */ 229 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); 230 231 object->un_pager.vnp.vnp_size = size; 232 233 object->handle = handle; 234 if (VFS_NEEDSGIANT(vp->v_mount)) 235 vm_object_set_flag(object, OBJ_NEEDGIANT); 236 vp->v_object = object; 237 } else { 238 object->ref_count++; 239 VM_OBJECT_UNLOCK(object); 240 } 241 vref(vp); 242 return (object); 243} 244 245/* 246 * The object must be locked. 247 */ 248static void 249vnode_pager_dealloc(object) 250 vm_object_t object; 251{ 252 struct vnode *vp = object->handle; 253 254 if (vp == NULL) 255 panic("vnode_pager_dealloc: pager already dealloced"); 256 257 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 258 vm_object_pip_wait(object, "vnpdea"); 259 260 object->handle = NULL; 261 object->type = OBJT_DEAD; 262 if (object->flags & OBJ_DISCONNECTWNT) { 263 vm_object_clear_flag(object, OBJ_DISCONNECTWNT); 264 wakeup(object); 265 } 266 ASSERT_VOP_LOCKED(vp, "vnode_pager_dealloc"); 267 vp->v_object = NULL; 268 vp->v_vflag &= ~VV_TEXT; 269} 270 271static boolean_t 272vnode_pager_haspage(object, pindex, before, after) 273 vm_object_t object; 274 vm_pindex_t pindex; 275 int *before; 276 int *after; 277{ 278 struct vnode *vp = object->handle; 279 daddr_t bn; 280 int err; 281 daddr_t reqblock; 282 int poff; 283 int bsize; 284 int pagesperblock, blocksperpage; 285 int vfslocked; 286 287 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 288 /* 289 * If no vp or vp is doomed or marked transparent to VM, we do not 290 * have the page. 291 */ 292 if (vp == NULL) 293 return FALSE; 294 295 VI_LOCK(vp); 296 if (vp->v_iflag & VI_DOOMED) { 297 VI_UNLOCK(vp); 298 return FALSE; 299 } 300 VI_UNLOCK(vp); 301 /* 302 * If filesystem no longer mounted or offset beyond end of file we do 303 * not have the page. 304 */ 305 if ((vp->v_mount == NULL) || 306 (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)) 307 return FALSE; 308 309 bsize = vp->v_mount->mnt_stat.f_iosize; 310 pagesperblock = bsize / PAGE_SIZE; 311 blocksperpage = 0; 312 if (pagesperblock > 0) { 313 reqblock = pindex / pagesperblock; 314 } else { 315 blocksperpage = (PAGE_SIZE / bsize); 316 reqblock = pindex * blocksperpage; 317 } 318 VM_OBJECT_UNLOCK(object); 319 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 320 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before); 321 VFS_UNLOCK_GIANT(vfslocked); 322 VM_OBJECT_LOCK(object); 323 if (err) 324 return TRUE; 325 if (bn == -1) 326 return FALSE; 327 if (pagesperblock > 0) { 328 poff = pindex - (reqblock * pagesperblock); 329 if (before) { 330 *before *= pagesperblock; 331 *before += poff; 332 } 333 if (after) { 334 int numafter; 335 *after *= pagesperblock; 336 numafter = pagesperblock - (poff + 1); 337 if (IDX_TO_OFF(pindex + numafter) > 338 object->un_pager.vnp.vnp_size) { 339 numafter = 340 OFF_TO_IDX(object->un_pager.vnp.vnp_size) - 341 pindex; 342 } 343 *after += numafter; 344 } 345 } else { 346 if (before) { 347 *before /= blocksperpage; 348 } 349 350 if (after) { 351 *after /= blocksperpage; 352 } 353 } 354 return TRUE; 355} 356 357/* 358 * Lets the VM system know about a change in size for a file. 359 * We adjust our own internal size and flush any cached pages in 360 * the associated object that are affected by the size change. 361 * 362 * Note: this routine may be invoked as a result of a pager put 363 * operation (possibly at object termination time), so we must be careful. 364 */ 365void 366vnode_pager_setsize(vp, nsize) 367 struct vnode *vp; 368 vm_ooffset_t nsize; 369{ 370 vm_object_t object; 371 vm_page_t m; 372 vm_pindex_t nobjsize; 373 374 if ((object = vp->v_object) == NULL) 375 return; 376 VM_OBJECT_LOCK(object); 377 if (nsize == object->un_pager.vnp.vnp_size) { 378 /* 379 * Hasn't changed size 380 */ 381 VM_OBJECT_UNLOCK(object); 382 return; 383 } 384 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 385 if (nsize < object->un_pager.vnp.vnp_size) { 386 /* 387 * File has shrunk. Toss any cached pages beyond the new EOF. 388 */ 389 if (nobjsize < object->size) 390 vm_object_page_remove(object, nobjsize, object->size, 391 FALSE); 392 /* 393 * this gets rid of garbage at the end of a page that is now 394 * only partially backed by the vnode. 395 * 396 * XXX for some reason (I don't know yet), if we take a 397 * completely invalid page and mark it partially valid 398 * it can screw up NFS reads, so we don't allow the case. 399 */ 400 if ((nsize & PAGE_MASK) && 401 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL && 402 m->valid != 0) { 403 int base = (int)nsize & PAGE_MASK; 404 int size = PAGE_SIZE - base; 405 406 /* 407 * Clear out partial-page garbage in case 408 * the page has been mapped. 409 */ 410 pmap_zero_page_area(m, base, size); 411 412 /* 413 * XXX work around SMP data integrity race 414 * by unmapping the page from user processes. 415 * The garbage we just cleared may be mapped 416 * to a user process running on another cpu 417 * and this code is not running through normal 418 * I/O channels which handle SMP issues for 419 * us, so unmap page to synchronize all cpus. 420 * 421 * XXX should vm_pager_unmap_page() have 422 * dealt with this? 423 */ 424 vm_page_lock_queues(); 425 pmap_remove_all(m); 426 427 /* 428 * Clear out partial-page dirty bits. This 429 * has the side effect of setting the valid 430 * bits, but that is ok. There are a bunch 431 * of places in the VM system where we expected 432 * m->dirty == VM_PAGE_BITS_ALL. The file EOF 433 * case is one of them. If the page is still 434 * partially dirty, make it fully dirty. 435 * 436 * note that we do not clear out the valid 437 * bits. This would prevent bogus_page 438 * replacement from working properly. 439 */ 440 vm_page_set_validclean(m, base, size); 441 if (m->dirty != 0) 442 m->dirty = VM_PAGE_BITS_ALL; 443 vm_page_unlock_queues(); 444 } 445 } 446 object->un_pager.vnp.vnp_size = nsize; 447 object->size = nobjsize; 448 VM_OBJECT_UNLOCK(object); 449} 450 451/* 452 * calculate the linear (byte) disk address of specified virtual 453 * file address 454 */ 455static daddr_t 456vnode_pager_addr(vp, address, run) 457 struct vnode *vp; 458 vm_ooffset_t address; 459 int *run; 460{ 461 daddr_t rtaddress; 462 int bsize; 463 daddr_t block; 464 int err; 465 daddr_t vblock; 466 daddr_t voffset; 467 468 if (address < 0) 469 return -1; 470 471 if (vp->v_mount == NULL) 472 return -1; 473 474 bsize = vp->v_mount->mnt_stat.f_iosize; 475 vblock = address / bsize; 476 voffset = address % bsize; 477 478 err = VOP_BMAP(vp, vblock, NULL, &block, run, NULL); 479 480 if (err || (block == -1)) 481 rtaddress = -1; 482 else { 483 rtaddress = block + voffset / DEV_BSIZE; 484 if (run) { 485 *run += 1; 486 *run *= bsize/PAGE_SIZE; 487 *run -= voffset/PAGE_SIZE; 488 } 489 } 490 491 return rtaddress; 492} 493 494/* 495 * small block filesystem vnode pager input 496 */ 497static int 498vnode_pager_input_smlfs(object, m) 499 vm_object_t object; 500 vm_page_t m; 501{ 502 int i; 503 struct vnode *vp; 504 struct bufobj *bo; 505 struct buf *bp; 506 struct sf_buf *sf; 507 daddr_t fileaddr; 508 vm_offset_t bsize; 509 int error = 0; 510 511 vp = object->handle; 512 if (vp->v_mount == NULL) 513 return VM_PAGER_BAD; 514 515 bsize = vp->v_mount->mnt_stat.f_iosize; 516 517 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL); 518 519 sf = sf_buf_alloc(m, 0); 520 521 for (i = 0; i < PAGE_SIZE / bsize; i++) { 522 vm_ooffset_t address; 523 524 if (vm_page_bits(i * bsize, bsize) & m->valid) 525 continue; 526 527 address = IDX_TO_OFF(m->pindex) + i * bsize; 528 if (address >= object->un_pager.vnp.vnp_size) { 529 fileaddr = -1; 530 } else { 531 fileaddr = vnode_pager_addr(vp, address, NULL); 532 } 533 if (fileaddr != -1) { 534 bp = getpbuf(&vnode_pbuf_freecnt); 535 536 /* build a minimal buffer header */ 537 bp->b_iocmd = BIO_READ; 538 bp->b_iodone = bdone; 539 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 540 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 541 bp->b_rcred = crhold(curthread->td_ucred); 542 bp->b_wcred = crhold(curthread->td_ucred); 543 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize; 544 bp->b_blkno = fileaddr; 545 pbgetbo(bo, bp); 546 bp->b_bcount = bsize; 547 bp->b_bufsize = bsize; 548 bp->b_runningbufspace = bp->b_bufsize; 549 runningbufspace += bp->b_runningbufspace; 550 551 /* do the input */ 552 bp->b_iooffset = dbtob(bp->b_blkno); 553 bstrategy(bp); 554 555 /* we definitely need to be at splvm here */ 556 557 bwait(bp, PVM, "vnsrd"); 558 559 if ((bp->b_ioflags & BIO_ERROR) != 0) 560 error = EIO; 561 562 /* 563 * free the buffer header back to the swap buffer pool 564 */ 565 pbrelbo(bp); 566 relpbuf(bp, &vnode_pbuf_freecnt); 567 if (error) 568 break; 569 570 VM_OBJECT_LOCK(object); 571 vm_page_lock_queues(); 572 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 573 vm_page_unlock_queues(); 574 VM_OBJECT_UNLOCK(object); 575 } else { 576 VM_OBJECT_LOCK(object); 577 vm_page_lock_queues(); 578 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 579 vm_page_unlock_queues(); 580 VM_OBJECT_UNLOCK(object); 581 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize); 582 } 583 } 584 sf_buf_free(sf); 585 vm_page_lock_queues(); 586 pmap_clear_modify(m); 587 vm_page_unlock_queues(); 588 if (error) { 589 return VM_PAGER_ERROR; 590 } 591 return VM_PAGER_OK; 592 593} 594 595 596/* 597 * old style vnode pager input routine 598 */ 599static int 600vnode_pager_input_old(object, m) 601 vm_object_t object; 602 vm_page_t m; 603{ 604 struct uio auio; 605 struct iovec aiov; 606 int error; 607 int size; 608 struct sf_buf *sf; 609 struct vnode *vp; 610 611 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 612 error = 0; 613 614 /* 615 * Return failure if beyond current EOF 616 */ 617 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 618 return VM_PAGER_BAD; 619 } else { 620 size = PAGE_SIZE; 621 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 622 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 623 vp = object->handle; 624 VM_OBJECT_UNLOCK(object); 625 626 /* 627 * Allocate a kernel virtual address and initialize so that 628 * we can use VOP_READ/WRITE routines. 629 */ 630 sf = sf_buf_alloc(m, 0); 631 632 aiov.iov_base = (caddr_t)sf_buf_kva(sf); 633 aiov.iov_len = size; 634 auio.uio_iov = &aiov; 635 auio.uio_iovcnt = 1; 636 auio.uio_offset = IDX_TO_OFF(m->pindex); 637 auio.uio_segflg = UIO_SYSSPACE; 638 auio.uio_rw = UIO_READ; 639 auio.uio_resid = size; 640 auio.uio_td = curthread; 641 642 error = VOP_READ(vp, &auio, 0, curthread->td_ucred); 643 if (!error) { 644 int count = size - auio.uio_resid; 645 646 if (count == 0) 647 error = EINVAL; 648 else if (count != PAGE_SIZE) 649 bzero((caddr_t)sf_buf_kva(sf) + count, 650 PAGE_SIZE - count); 651 } 652 sf_buf_free(sf); 653 654 VM_OBJECT_LOCK(object); 655 } 656 vm_page_lock_queues(); 657 pmap_clear_modify(m); 658 vm_page_undirty(m); 659 vm_page_unlock_queues(); 660 if (!error) 661 m->valid = VM_PAGE_BITS_ALL; 662 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 663} 664 665/* 666 * generic vnode pager input routine 667 */ 668 669/* 670 * Local media VFS's that do not implement their own VOP_GETPAGES 671 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages() 672 * to implement the previous behaviour. 673 * 674 * All other FS's should use the bypass to get to the local media 675 * backing vp's VOP_GETPAGES. 676 */ 677static int 678vnode_pager_getpages(object, m, count, reqpage) 679 vm_object_t object; 680 vm_page_t *m; 681 int count; 682 int reqpage; 683{ 684 int rtval; 685 struct vnode *vp; 686 int bytes = count * PAGE_SIZE; 687 int vfslocked; 688 689 vp = object->handle; 690 VM_OBJECT_UNLOCK(object); 691 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 692 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0); 693 KASSERT(rtval != EOPNOTSUPP, 694 ("vnode_pager: FS getpages not implemented\n")); 695 VFS_UNLOCK_GIANT(vfslocked); 696 VM_OBJECT_LOCK(object); 697 return rtval; 698} 699 700/* 701 * This is now called from local media FS's to operate against their 702 * own vnodes if they fail to implement VOP_GETPAGES. 703 */ 704int 705vnode_pager_generic_getpages(vp, m, bytecount, reqpage) 706 struct vnode *vp; 707 vm_page_t *m; 708 int bytecount; 709 int reqpage; 710{ 711 vm_object_t object; 712 vm_offset_t kva; 713 off_t foff, tfoff, nextoff; 714 int i, j, size, bsize, first; 715 daddr_t firstaddr; 716 struct bufobj *bo; 717 int runpg; 718 int runend; 719 struct buf *bp; 720 int count; 721 int error = 0; 722 723 object = vp->v_object; 724 count = bytecount / PAGE_SIZE; 725 726 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 727 ("vnode_pager_generic_getpages does not support devices")); 728 if (vp->v_mount == NULL) 729 return VM_PAGER_BAD; 730 731 bsize = vp->v_mount->mnt_stat.f_iosize; 732 733 /* get the UNDERLYING device for the file with VOP_BMAP() */ 734 735 /* 736 * originally, we did not check for an error return value -- assuming 737 * an fs always has a bmap entry point -- that assumption is wrong!!! 738 */ 739 foff = IDX_TO_OFF(m[reqpage]->pindex); 740 741 /* 742 * if we can't bmap, use old VOP code 743 */ 744 if (VOP_BMAP(vp, 0, &bo, 0, NULL, NULL)) { 745 VM_OBJECT_LOCK(object); 746 vm_page_lock_queues(); 747 for (i = 0; i < count; i++) 748 if (i != reqpage) 749 vm_page_free(m[i]); 750 vm_page_unlock_queues(); 751 cnt.v_vnodein++; 752 cnt.v_vnodepgsin++; 753 error = vnode_pager_input_old(object, m[reqpage]); 754 VM_OBJECT_UNLOCK(object); 755 return (error); 756 757 /* 758 * if the blocksize is smaller than a page size, then use 759 * special small filesystem code. NFS sometimes has a small 760 * blocksize, but it can handle large reads itself. 761 */ 762 } else if ((PAGE_SIZE / bsize) > 1 && 763 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) { 764 VM_OBJECT_LOCK(object); 765 vm_page_lock_queues(); 766 for (i = 0; i < count; i++) 767 if (i != reqpage) 768 vm_page_free(m[i]); 769 vm_page_unlock_queues(); 770 VM_OBJECT_UNLOCK(object); 771 cnt.v_vnodein++; 772 cnt.v_vnodepgsin++; 773 return vnode_pager_input_smlfs(object, m[reqpage]); 774 } 775 776 /* 777 * If we have a completely valid page available to us, we can 778 * clean up and return. Otherwise we have to re-read the 779 * media. 780 */ 781 VM_OBJECT_LOCK(object); 782 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) { 783 vm_page_lock_queues(); 784 for (i = 0; i < count; i++) 785 if (i != reqpage) 786 vm_page_free(m[i]); 787 vm_page_unlock_queues(); 788 VM_OBJECT_UNLOCK(object); 789 return VM_PAGER_OK; 790 } 791 m[reqpage]->valid = 0; 792 VM_OBJECT_UNLOCK(object); 793 794 /* 795 * here on direct device I/O 796 */ 797 firstaddr = -1; 798 799 /* 800 * calculate the run that includes the required page 801 */ 802 for (first = 0, i = 0; i < count; i = runend) { 803 firstaddr = vnode_pager_addr(vp, 804 IDX_TO_OFF(m[i]->pindex), &runpg); 805 if (firstaddr == -1) { 806 VM_OBJECT_LOCK(object); 807 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) { 808 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx", 809 (intmax_t)firstaddr, (uintmax_t)(foff >> 32), 810 (uintmax_t)foff, 811 (uintmax_t) 812 (object->un_pager.vnp.vnp_size >> 32), 813 (uintmax_t)object->un_pager.vnp.vnp_size); 814 } 815 vm_page_lock_queues(); 816 vm_page_free(m[i]); 817 vm_page_unlock_queues(); 818 VM_OBJECT_UNLOCK(object); 819 runend = i + 1; 820 first = runend; 821 continue; 822 } 823 runend = i + runpg; 824 if (runend <= reqpage) { 825 VM_OBJECT_LOCK(object); 826 vm_page_lock_queues(); 827 for (j = i; j < runend; j++) 828 vm_page_free(m[j]); 829 vm_page_unlock_queues(); 830 VM_OBJECT_UNLOCK(object); 831 } else { 832 if (runpg < (count - first)) { 833 VM_OBJECT_LOCK(object); 834 vm_page_lock_queues(); 835 for (i = first + runpg; i < count; i++) 836 vm_page_free(m[i]); 837 vm_page_unlock_queues(); 838 VM_OBJECT_UNLOCK(object); 839 count = first + runpg; 840 } 841 break; 842 } 843 first = runend; 844 } 845 846 /* 847 * the first and last page have been calculated now, move input pages 848 * to be zero based... 849 */ 850 if (first != 0) { 851 for (i = first; i < count; i++) { 852 m[i - first] = m[i]; 853 } 854 count -= first; 855 reqpage -= first; 856 } 857 858 /* 859 * calculate the file virtual address for the transfer 860 */ 861 foff = IDX_TO_OFF(m[0]->pindex); 862 863 /* 864 * calculate the size of the transfer 865 */ 866 size = count * PAGE_SIZE; 867 KASSERT(count > 0, ("zero count")); 868 if ((foff + size) > object->un_pager.vnp.vnp_size) 869 size = object->un_pager.vnp.vnp_size - foff; 870 KASSERT(size > 0, ("zero size")); 871 872 /* 873 * round up physical size for real devices. 874 */ 875 if (1) { 876 int secmask = bo->bo_bsize - 1; 877 KASSERT(secmask < PAGE_SIZE && secmask > 0, 878 ("vnode_pager_generic_getpages: sector size %d too large", 879 secmask + 1)); 880 size = (size + secmask) & ~secmask; 881 } 882 883 bp = getpbuf(&vnode_pbuf_freecnt); 884 kva = (vm_offset_t) bp->b_data; 885 886 /* 887 * and map the pages to be read into the kva 888 */ 889 pmap_qenter(kva, m, count); 890 891 /* build a minimal buffer header */ 892 bp->b_iocmd = BIO_READ; 893 bp->b_iodone = bdone; 894 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 895 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 896 bp->b_rcred = crhold(curthread->td_ucred); 897 bp->b_wcred = crhold(curthread->td_ucred); 898 bp->b_blkno = firstaddr; 899 pbgetbo(bo, bp); 900 bp->b_bcount = size; 901 bp->b_bufsize = size; 902 bp->b_runningbufspace = bp->b_bufsize; 903 runningbufspace += bp->b_runningbufspace; 904 905 cnt.v_vnodein++; 906 cnt.v_vnodepgsin += count; 907 908 /* do the input */ 909 bp->b_iooffset = dbtob(bp->b_blkno); 910 bstrategy(bp); 911 912 bwait(bp, PVM, "vnread"); 913 914 if ((bp->b_ioflags & BIO_ERROR) != 0) 915 error = EIO; 916 917 if (!error) { 918 if (size != count * PAGE_SIZE) 919 bzero((caddr_t) kva + size, PAGE_SIZE * count - size); 920 } 921 pmap_qremove(kva, count); 922 923 /* 924 * free the buffer header back to the swap buffer pool 925 */ 926 pbrelbo(bp); 927 relpbuf(bp, &vnode_pbuf_freecnt); 928 929 VM_OBJECT_LOCK(object); 930 vm_page_lock_queues(); 931 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) { 932 vm_page_t mt; 933 934 nextoff = tfoff + PAGE_SIZE; 935 mt = m[i]; 936 937 if (nextoff <= object->un_pager.vnp.vnp_size) { 938 /* 939 * Read filled up entire page. 940 */ 941 mt->valid = VM_PAGE_BITS_ALL; 942 vm_page_undirty(mt); /* should be an assert? XXX */ 943 pmap_clear_modify(mt); 944 } else { 945 /* 946 * Read did not fill up entire page. Since this 947 * is getpages, the page may be mapped, so we have 948 * to zero the invalid portions of the page even 949 * though we aren't setting them valid. 950 * 951 * Currently we do not set the entire page valid, 952 * we just try to clear the piece that we couldn't 953 * read. 954 */ 955 vm_page_set_validclean(mt, 0, 956 object->un_pager.vnp.vnp_size - tfoff); 957 /* handled by vm_fault now */ 958 /* vm_page_zero_invalid(mt, FALSE); */ 959 } 960 961 if (i != reqpage) { 962 963 /* 964 * whether or not to leave the page activated is up in 965 * the air, but we should put the page on a page queue 966 * somewhere. (it already is in the object). Result: 967 * It appears that empirical results show that 968 * deactivating pages is best. 969 */ 970 971 /* 972 * just in case someone was asking for this page we 973 * now tell them that it is ok to use 974 */ 975 if (!error) { 976 if (mt->flags & PG_WANTED) 977 vm_page_activate(mt); 978 else 979 vm_page_deactivate(mt); 980 vm_page_wakeup(mt); 981 } else { 982 vm_page_free(mt); 983 } 984 } 985 } 986 vm_page_unlock_queues(); 987 VM_OBJECT_UNLOCK(object); 988 if (error) { 989 printf("vnode_pager_getpages: I/O read error\n"); 990 } 991 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 992} 993 994/* 995 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 996 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 997 * vnode_pager_generic_putpages() to implement the previous behaviour. 998 * 999 * All other FS's should use the bypass to get to the local media 1000 * backing vp's VOP_PUTPAGES. 1001 */ 1002static void 1003vnode_pager_putpages(object, m, count, sync, rtvals) 1004 vm_object_t object; 1005 vm_page_t *m; 1006 int count; 1007 boolean_t sync; 1008 int *rtvals; 1009{ 1010 int rtval; 1011 struct vnode *vp; 1012 struct mount *mp; 1013 int bytes = count * PAGE_SIZE; 1014 1015 /* 1016 * Force synchronous operation if we are extremely low on memory 1017 * to prevent a low-memory deadlock. VOP operations often need to 1018 * allocate more memory to initiate the I/O ( i.e. do a BMAP 1019 * operation ). The swapper handles the case by limiting the amount 1020 * of asynchronous I/O, but that sort of solution doesn't scale well 1021 * for the vnode pager without a lot of work. 1022 * 1023 * Also, the backing vnode's iodone routine may not wake the pageout 1024 * daemon up. This should be probably be addressed XXX. 1025 */ 1026 1027 if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min) 1028 sync |= OBJPC_SYNC; 1029 1030 /* 1031 * Call device-specific putpages function 1032 */ 1033 vp = object->handle; 1034 VM_OBJECT_UNLOCK(object); 1035 if (vp->v_type != VREG) 1036 mp = NULL; 1037 (void)vn_start_write(vp, &mp, V_WAIT); 1038 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 1039 KASSERT(rtval != EOPNOTSUPP, 1040 ("vnode_pager: stale FS putpages\n")); 1041 vn_finished_write(mp); 1042 VM_OBJECT_LOCK(object); 1043} 1044 1045 1046/* 1047 * This is now called from local media FS's to operate against their 1048 * own vnodes if they fail to implement VOP_PUTPAGES. 1049 * 1050 * This is typically called indirectly via the pageout daemon and 1051 * clustering has already typically occured, so in general we ask the 1052 * underlying filesystem to write the data out asynchronously rather 1053 * then delayed. 1054 */ 1055int 1056vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals) 1057 struct vnode *vp; 1058 vm_page_t *m; 1059 int bytecount; 1060 int flags; 1061 int *rtvals; 1062{ 1063 int i; 1064 vm_object_t object; 1065 int count; 1066 1067 int maxsize, ncount; 1068 vm_ooffset_t poffset; 1069 struct uio auio; 1070 struct iovec aiov; 1071 int error; 1072 int ioflags; 1073 1074 object = vp->v_object; 1075 count = bytecount / PAGE_SIZE; 1076 1077 for (i = 0; i < count; i++) 1078 rtvals[i] = VM_PAGER_AGAIN; 1079 1080 if ((int64_t)m[0]->pindex < 0) { 1081 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n", 1082 (long)m[0]->pindex, (u_long)m[0]->dirty); 1083 rtvals[0] = VM_PAGER_BAD; 1084 return VM_PAGER_BAD; 1085 } 1086 1087 maxsize = count * PAGE_SIZE; 1088 ncount = count; 1089 1090 poffset = IDX_TO_OFF(m[0]->pindex); 1091 1092 /* 1093 * If the page-aligned write is larger then the actual file we 1094 * have to invalidate pages occuring beyond the file EOF. However, 1095 * there is an edge case where a file may not be page-aligned where 1096 * the last page is partially invalid. In this case the filesystem 1097 * may not properly clear the dirty bits for the entire page (which 1098 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 1099 * With the page locked we are free to fix-up the dirty bits here. 1100 * 1101 * We do not under any circumstances truncate the valid bits, as 1102 * this will screw up bogus page replacement. 1103 */ 1104 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 1105 if (object->un_pager.vnp.vnp_size > poffset) { 1106 int pgoff; 1107 1108 maxsize = object->un_pager.vnp.vnp_size - poffset; 1109 ncount = btoc(maxsize); 1110 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 1111 vm_page_lock_queues(); 1112 vm_page_clear_dirty(m[ncount - 1], pgoff, 1113 PAGE_SIZE - pgoff); 1114 vm_page_unlock_queues(); 1115 } 1116 } else { 1117 maxsize = 0; 1118 ncount = 0; 1119 } 1120 if (ncount < count) { 1121 for (i = ncount; i < count; i++) { 1122 rtvals[i] = VM_PAGER_BAD; 1123 } 1124 } 1125 } 1126 1127 /* 1128 * pageouts are already clustered, use IO_ASYNC t o force a bawrite() 1129 * rather then a bdwrite() to prevent paging I/O from saturating 1130 * the buffer cache. Dummy-up the sequential heuristic to cause 1131 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 1132 * the system decides how to cluster. 1133 */ 1134 ioflags = IO_VMIO; 1135 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 1136 ioflags |= IO_SYNC; 1137 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 1138 ioflags |= IO_ASYNC; 1139 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 1140 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 1141 1142 aiov.iov_base = (caddr_t) 0; 1143 aiov.iov_len = maxsize; 1144 auio.uio_iov = &aiov; 1145 auio.uio_iovcnt = 1; 1146 auio.uio_offset = poffset; 1147 auio.uio_segflg = UIO_NOCOPY; 1148 auio.uio_rw = UIO_WRITE; 1149 auio.uio_resid = maxsize; 1150 auio.uio_td = (struct thread *) 0; 1151 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred); 1152 cnt.v_vnodeout++; 1153 cnt.v_vnodepgsout += ncount; 1154 1155 if (error) { 1156 printf("vnode_pager_putpages: I/O error %d\n", error); 1157 } 1158 if (auio.uio_resid) { 1159 printf("vnode_pager_putpages: residual I/O %d at %lu\n", 1160 auio.uio_resid, (u_long)m[0]->pindex); 1161 } 1162 for (i = 0; i < ncount; i++) { 1163 rtvals[i] = VM_PAGER_OK; 1164 } 1165 return rtvals[0]; 1166} 1167 1168struct vnode * 1169vnode_pager_lock(vm_object_t first_object) 1170{ 1171 struct vnode *vp; 1172 vm_object_t backing_object, object; 1173 1174 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 1175 for (object = first_object; object != NULL; object = backing_object) { 1176 if (object->type != OBJT_VNODE) { 1177 if ((backing_object = object->backing_object) != NULL) 1178 VM_OBJECT_LOCK(backing_object); 1179 if (object != first_object) 1180 VM_OBJECT_UNLOCK(object); 1181 continue; 1182 } 1183 retry: 1184 if (object->flags & OBJ_DEAD) { 1185 if (object != first_object) 1186 VM_OBJECT_UNLOCK(object); 1187 return NULL; 1188 } 1189 vp = object->handle; 1190 VI_LOCK(vp); 1191 VM_OBJECT_UNLOCK(object); 1192 if (first_object != object) 1193 VM_OBJECT_UNLOCK(first_object); 1194 VFS_ASSERT_GIANT(vp->v_mount); 1195 if (vget(vp, LK_CANRECURSE | LK_INTERLOCK | 1196 LK_RETRY | LK_SHARED, curthread)) { 1197 VM_OBJECT_LOCK(first_object); 1198 if (object != first_object) 1199 VM_OBJECT_LOCK(object); 1200 if (object->type != OBJT_VNODE) { 1201 if (object != first_object) 1202 VM_OBJECT_UNLOCK(object); 1203 return NULL; 1204 } 1205 printf("vnode_pager_lock: retrying\n"); 1206 goto retry; 1207 } 1208 VM_OBJECT_LOCK(first_object); 1209 return (vp); 1210 } 1211 return NULL; 1212} 1213