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/limits.h> 65 #include <sys/conf.h> 66 #include <sys/sf_buf.h> 67 68 #include <machine/atomic.h> 69 70 #include <vm/vm.h> 71 #include <vm/vm_object.h> 72 #include <vm/vm_page.h> 73 #include <vm/vm_pager.h> 74 #include <vm/vm_map.h> 75 #include <vm/vnode_pager.h> 76 #include <vm/vm_extern.h> 77 78 static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, 79 daddr_t *rtaddress, int *run); 80 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m); 81 static int vnode_pager_input_old(vm_object_t object, vm_page_t m); 82 static void vnode_pager_dealloc(vm_object_t); 83 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int); 84 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 85 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *); 86 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t, 87 vm_ooffset_t, struct ucred *cred); 88 89 struct pagerops vnodepagerops = { 90 .pgo_alloc = vnode_pager_alloc, 91 .pgo_dealloc = vnode_pager_dealloc, 92 .pgo_getpages = vnode_pager_getpages, 93 .pgo_putpages = vnode_pager_putpages, 94 .pgo_haspage = vnode_pager_haspage, 95 }; 96 97 int vnode_pbuf_freecnt; 98 99 /* Create the VM system backing object for this vnode */ 100 int 101 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td) 102 { 103 vm_object_t object; 104 vm_ooffset_t size = isize; 105 struct vattr va; 106 107 if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE) 108 return (0); 109 110 while ((object = vp->v_object) != NULL) { 111 VM_OBJECT_LOCK(object); 112 if (!(object->flags & OBJ_DEAD)) { 113 VM_OBJECT_UNLOCK(object); 114 return (0); 115 } 116 VOP_UNLOCK(vp, 0); 117 vm_object_set_flag(object, OBJ_DISCONNECTWNT); 118 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vodead", 0); 119 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 120 } 121 122 if (size == 0) { 123 if (vn_isdisk(vp, NULL)) { 124 size = IDX_TO_OFF(INT_MAX); 125 } else { 126 if (VOP_GETATTR(vp, &va, td->td_ucred)) 127 return (0); 128 size = va.va_size; 129 } 130 } 131 132 object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred); 133 /* 134 * Dereference the reference we just created. This assumes 135 * that the object is associated with the vp. 136 */ 137 VM_OBJECT_LOCK(object); 138 object->ref_count--; 139 VM_OBJECT_UNLOCK(object); 140 vrele(vp); 141 142 KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object")); 143 144 return (0); 145 } 146 147 void 148 vnode_destroy_vobject(struct vnode *vp) 149 { 150 struct vm_object *obj; 151 152 obj = vp->v_object; 153 if (obj == NULL) 154 return; 155 ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject"); 156 VM_OBJECT_LOCK(obj); 157 if (obj->ref_count == 0) { 158 /* 159 * vclean() may be called twice. The first time 160 * removes the primary reference to the object, 161 * the second time goes one further and is a 162 * special-case to terminate the object. 163 * 164 * don't double-terminate the object 165 */ 166 if ((obj->flags & OBJ_DEAD) == 0) 167 vm_object_terminate(obj); 168 else 169 VM_OBJECT_UNLOCK(obj); 170 } else { 171 /* 172 * Woe to the process that tries to page now :-). 173 */ 174 vm_pager_deallocate(obj); 175 VM_OBJECT_UNLOCK(obj); 176 } 177 vp->v_object = NULL; 178 } 179 180 181 /* 182 * Allocate (or lookup) pager for a vnode. 183 * Handle is a vnode pointer. 184 * 185 * MPSAFE 186 */ 187 vm_object_t 188 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 189 vm_ooffset_t offset, struct ucred *cred) 190 { 191 vm_object_t object; 192 struct vnode *vp; 193 194 /* 195 * Pageout to vnode, no can do yet. 196 */ 197 if (handle == NULL) 198 return (NULL); 199 200 vp = (struct vnode *) handle; 201 202 /* 203 * If the object is being terminated, wait for it to 204 * go away. 205 */ 206 retry: 207 while ((object = vp->v_object) != NULL) { 208 VM_OBJECT_LOCK(object); 209 if ((object->flags & OBJ_DEAD) == 0) 210 break; 211 vm_object_set_flag(object, OBJ_DISCONNECTWNT); 212 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vadead", 0); 213 } 214 215 if (vp->v_usecount == 0) 216 panic("vnode_pager_alloc: no vnode reference"); 217 218 if (object == NULL) { 219 /* 220 * Add an object of the appropriate size 221 */ 222 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); 223 224 object->un_pager.vnp.vnp_size = size; 225 226 object->handle = handle; 227 VI_LOCK(vp); 228 if (vp->v_object != NULL) { 229 /* 230 * Object has been created while we were sleeping 231 */ 232 VI_UNLOCK(vp); 233 vm_object_destroy(object); 234 goto retry; 235 } 236 vp->v_object = object; 237 VI_UNLOCK(vp); 238 } else { 239 object->ref_count++; 240 VM_OBJECT_UNLOCK(object); 241 } 242 vref(vp); 243 return (object); 244 } 245 246 /* 247 * The object must be locked. 248 */ 249 static void 250 vnode_pager_dealloc(object) 251 vm_object_t object; 252 { 253 struct vnode *vp; 254 int refs; 255 256 vp = object->handle; 257 if (vp == NULL) 258 panic("vnode_pager_dealloc: pager already dealloced"); 259 260 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 261 vm_object_pip_wait(object, "vnpdea"); 262 refs = object->ref_count; 263 264 object->handle = NULL; 265 object->type = OBJT_DEAD; 266 if (object->flags & OBJ_DISCONNECTWNT) { 267 vm_object_clear_flag(object, OBJ_DISCONNECTWNT); 268 wakeup(object); 269 } 270 ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc"); 271 vp->v_object = NULL; 272 vp->v_vflag &= ~VV_TEXT; 273 while (refs-- > 0) 274 vunref(vp); 275 } 276 277 static boolean_t 278 vnode_pager_haspage(object, pindex, before, after) 279 vm_object_t object; 280 vm_pindex_t pindex; 281 int *before; 282 int *after; 283 { 284 struct vnode *vp = object->handle; 285 daddr_t bn; 286 int err; 287 daddr_t reqblock; 288 int poff; 289 int bsize; 290 int pagesperblock, blocksperpage; 291 int vfslocked; 292 293 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 294 /* 295 * If no vp or vp is doomed or marked transparent to VM, we do not 296 * have the page. 297 */ 298 if (vp == NULL || vp->v_iflag & VI_DOOMED) 299 return FALSE; 300 /* 301 * If the offset is beyond end of file we do 302 * not have the page. 303 */ 304 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size) 305 return FALSE; 306 307 bsize = vp->v_mount->mnt_stat.f_iosize; 308 pagesperblock = bsize / PAGE_SIZE; 309 blocksperpage = 0; 310 if (pagesperblock > 0) { 311 reqblock = pindex / pagesperblock; 312 } else { 313 blocksperpage = (PAGE_SIZE / bsize); 314 reqblock = pindex * blocksperpage; 315 } 316 VM_OBJECT_UNLOCK(object); 317 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 318 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before); 319 VFS_UNLOCK_GIANT(vfslocked); 320 VM_OBJECT_LOCK(object); 321 if (err) 322 return TRUE; 323 if (bn == -1) 324 return FALSE; 325 if (pagesperblock > 0) { 326 poff = pindex - (reqblock * pagesperblock); 327 if (before) { 328 *before *= pagesperblock; 329 *before += poff; 330 } 331 if (after) { 332 int numafter; 333 *after *= pagesperblock; 334 numafter = pagesperblock - (poff + 1); 335 if (IDX_TO_OFF(pindex + numafter) > 336 object->un_pager.vnp.vnp_size) { 337 numafter = 338 OFF_TO_IDX(object->un_pager.vnp.vnp_size) - 339 pindex; 340 } 341 *after += numafter; 342 } 343 } else { 344 if (before) { 345 *before /= blocksperpage; 346 } 347 348 if (after) { 349 *after /= blocksperpage; 350 } 351 } 352 return TRUE; 353 } 354 355 /* 356 * Lets the VM system know about a change in size for a file. 357 * We adjust our own internal size and flush any cached pages in 358 * the associated object that are affected by the size change. 359 * 360 * Note: this routine may be invoked as a result of a pager put 361 * operation (possibly at object termination time), so we must be careful. 362 */ 363 void 364 vnode_pager_setsize(vp, nsize) 365 struct vnode *vp; 366 vm_ooffset_t nsize; 367 { 368 vm_object_t object; 369 vm_page_t m; 370 vm_pindex_t nobjsize; 371 372 if ((object = vp->v_object) == NULL) 373 return; 374 /* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */ 375 VM_OBJECT_LOCK(object); 376 if (nsize == object->un_pager.vnp.vnp_size) { 377 /* 378 * Hasn't changed size 379 */ 380 VM_OBJECT_UNLOCK(object); 381 return; 382 } 383 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 384 if (nsize < object->un_pager.vnp.vnp_size) { 385 /* 386 * File has shrunk. Toss any cached pages beyond the new EOF. 387 */ 388 if (nobjsize < object->size) 389 vm_object_page_remove(object, nobjsize, object->size, 390 FALSE); 391 /* 392 * this gets rid of garbage at the end of a page that is now 393 * only partially backed by the vnode. 394 * 395 * XXX for some reason (I don't know yet), if we take a 396 * completely invalid page and mark it partially valid 397 * it can screw up NFS reads, so we don't allow the case. 398 */ 399 if ((nsize & PAGE_MASK) && 400 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL && 401 m->valid != 0) { 402 int base = (int)nsize & PAGE_MASK; 403 int size = PAGE_SIZE - base; 404 405 /* 406 * Clear out partial-page garbage in case 407 * the page has been mapped. 408 */ 409 pmap_zero_page_area(m, base, size); 410 411 /* 412 * Update the valid bits to reflect the blocks that 413 * have been zeroed. Some of these valid bits may 414 * have already been set. 415 */ 416 vm_page_set_valid(m, base, size); 417 418 /* 419 * Round "base" to the next block boundary so that the 420 * dirty bit for a partially zeroed block is not 421 * cleared. 422 */ 423 base = roundup2(base, DEV_BSIZE); 424 425 /* 426 * Clear out partial-page dirty bits. 427 * 428 * note that we do not clear out the valid 429 * bits. This would prevent bogus_page 430 * replacement from working properly. 431 */ 432 vm_page_clear_dirty(m, base, PAGE_SIZE - base); 433 } else if ((nsize & PAGE_MASK) && 434 __predict_false(object->cache != NULL)) { 435 vm_page_cache_free(object, OFF_TO_IDX(nsize), 436 nobjsize); 437 } 438 } 439 object->un_pager.vnp.vnp_size = nsize; 440 object->size = nobjsize; 441 VM_OBJECT_UNLOCK(object); 442 } 443 444 /* 445 * calculate the linear (byte) disk address of specified virtual 446 * file address 447 */ 448 static int 449 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress, 450 int *run) 451 { 452 int bsize; 453 int err; 454 daddr_t vblock; 455 daddr_t voffset; 456 457 if (address < 0) 458 return -1; 459 460 if (vp->v_iflag & VI_DOOMED) 461 return -1; 462 463 bsize = vp->v_mount->mnt_stat.f_iosize; 464 vblock = address / bsize; 465 voffset = address % bsize; 466 467 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL); 468 if (err == 0) { 469 if (*rtaddress != -1) 470 *rtaddress += voffset / DEV_BSIZE; 471 if (run) { 472 *run += 1; 473 *run *= bsize/PAGE_SIZE; 474 *run -= voffset/PAGE_SIZE; 475 } 476 } 477 478 return (err); 479 } 480 481 /* 482 * small block filesystem vnode pager input 483 */ 484 static int 485 vnode_pager_input_smlfs(object, m) 486 vm_object_t object; 487 vm_page_t m; 488 { 489 int bits, i; 490 struct vnode *vp; 491 struct bufobj *bo; 492 struct buf *bp; 493 struct sf_buf *sf; 494 daddr_t fileaddr; 495 vm_offset_t bsize; 496 int error = 0; 497 498 vp = object->handle; 499 if (vp->v_iflag & VI_DOOMED) 500 return VM_PAGER_BAD; 501 502 bsize = vp->v_mount->mnt_stat.f_iosize; 503 504 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL); 505 506 sf = sf_buf_alloc(m, 0); 507 508 for (i = 0; i < PAGE_SIZE / bsize; i++) { 509 vm_ooffset_t address; 510 511 bits = vm_page_bits(i * bsize, bsize); 512 if (m->valid & bits) 513 continue; 514 515 address = IDX_TO_OFF(m->pindex) + i * bsize; 516 if (address >= object->un_pager.vnp.vnp_size) { 517 fileaddr = -1; 518 } else { 519 error = vnode_pager_addr(vp, address, &fileaddr, NULL); 520 if (error) 521 break; 522 } 523 if (fileaddr != -1) { 524 bp = getpbuf(&vnode_pbuf_freecnt); 525 526 /* build a minimal buffer header */ 527 bp->b_iocmd = BIO_READ; 528 bp->b_iodone = bdone; 529 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 530 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 531 bp->b_rcred = crhold(curthread->td_ucred); 532 bp->b_wcred = crhold(curthread->td_ucred); 533 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize; 534 bp->b_blkno = fileaddr; 535 pbgetbo(bo, bp); 536 bp->b_bcount = bsize; 537 bp->b_bufsize = bsize; 538 bp->b_runningbufspace = bp->b_bufsize; 539 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 540 541 /* do the input */ 542 bp->b_iooffset = dbtob(bp->b_blkno); 543 bstrategy(bp); 544 545 bwait(bp, PVM, "vnsrd"); 546 547 if ((bp->b_ioflags & BIO_ERROR) != 0) 548 error = EIO; 549 550 /* 551 * free the buffer header back to the swap buffer pool 552 */ 553 pbrelbo(bp); 554 relpbuf(bp, &vnode_pbuf_freecnt); 555 if (error) 556 break; 557 } else 558 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize); 559 KASSERT((m->dirty & bits) == 0, 560 ("vnode_pager_input_smlfs: page %p is dirty", m)); 561 VM_OBJECT_LOCK(object); 562 m->valid |= bits; 563 VM_OBJECT_UNLOCK(object); 564 } 565 sf_buf_free(sf); 566 if (error) { 567 return VM_PAGER_ERROR; 568 } 569 return VM_PAGER_OK; 570 } 571 572 /* 573 * old style vnode pager input routine 574 */ 575 static int 576 vnode_pager_input_old(object, m) 577 vm_object_t object; 578 vm_page_t m; 579 { 580 struct uio auio; 581 struct iovec aiov; 582 int error; 583 int size; 584 struct sf_buf *sf; 585 struct vnode *vp; 586 587 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 588 error = 0; 589 590 /* 591 * Return failure if beyond current EOF 592 */ 593 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 594 return VM_PAGER_BAD; 595 } else { 596 size = PAGE_SIZE; 597 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 598 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 599 vp = object->handle; 600 VM_OBJECT_UNLOCK(object); 601 602 /* 603 * Allocate a kernel virtual address and initialize so that 604 * we can use VOP_READ/WRITE routines. 605 */ 606 sf = sf_buf_alloc(m, 0); 607 608 aiov.iov_base = (caddr_t)sf_buf_kva(sf); 609 aiov.iov_len = size; 610 auio.uio_iov = &aiov; 611 auio.uio_iovcnt = 1; 612 auio.uio_offset = IDX_TO_OFF(m->pindex); 613 auio.uio_segflg = UIO_SYSSPACE; 614 auio.uio_rw = UIO_READ; 615 auio.uio_resid = size; 616 auio.uio_td = curthread; 617 618 error = VOP_READ(vp, &auio, 0, curthread->td_ucred); 619 if (!error) { 620 int count = size - auio.uio_resid; 621 622 if (count == 0) 623 error = EINVAL; 624 else if (count != PAGE_SIZE) 625 bzero((caddr_t)sf_buf_kva(sf) + count, 626 PAGE_SIZE - count); 627 } 628 sf_buf_free(sf); 629 630 VM_OBJECT_LOCK(object); 631 } 632 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m)); 633 if (!error) 634 m->valid = VM_PAGE_BITS_ALL; 635 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 636 } 637 638 /* 639 * generic vnode pager input routine 640 */ 641 642 /* 643 * Local media VFS's that do not implement their own VOP_GETPAGES 644 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages() 645 * to implement the previous behaviour. 646 * 647 * All other FS's should use the bypass to get to the local media 648 * backing vp's VOP_GETPAGES. 649 */ 650 static int 651 vnode_pager_getpages(object, m, count, reqpage) 652 vm_object_t object; 653 vm_page_t *m; 654 int count; 655 int reqpage; 656 { 657 int rtval; 658 struct vnode *vp; 659 int bytes = count * PAGE_SIZE; 660 int vfslocked; 661 662 vp = object->handle; 663 VM_OBJECT_UNLOCK(object); 664 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 665 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0); 666 KASSERT(rtval != EOPNOTSUPP, 667 ("vnode_pager: FS getpages not implemented\n")); 668 VFS_UNLOCK_GIANT(vfslocked); 669 VM_OBJECT_LOCK(object); 670 return rtval; 671 } 672 673 /* 674 * This is now called from local media FS's to operate against their 675 * own vnodes if they fail to implement VOP_GETPAGES. 676 */ 677 int 678 vnode_pager_generic_getpages(vp, m, bytecount, reqpage) 679 struct vnode *vp; 680 vm_page_t *m; 681 int bytecount; 682 int reqpage; 683 { 684 vm_object_t object; 685 vm_offset_t kva; 686 off_t foff, tfoff, nextoff; 687 int i, j, size, bsize, first; 688 daddr_t firstaddr, reqblock; 689 struct bufobj *bo; 690 int runpg; 691 int runend; 692 struct buf *bp; 693 int count; 694 int error; 695 696 object = vp->v_object; 697 count = bytecount / PAGE_SIZE; 698 699 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 700 ("vnode_pager_generic_getpages does not support devices")); 701 if (vp->v_iflag & VI_DOOMED) 702 return VM_PAGER_BAD; 703 704 bsize = vp->v_mount->mnt_stat.f_iosize; 705 706 /* get the UNDERLYING device for the file with VOP_BMAP() */ 707 708 /* 709 * originally, we did not check for an error return value -- assuming 710 * an fs always has a bmap entry point -- that assumption is wrong!!! 711 */ 712 foff = IDX_TO_OFF(m[reqpage]->pindex); 713 714 /* 715 * if we can't bmap, use old VOP code 716 */ 717 error = VOP_BMAP(vp, foff / bsize, &bo, &reqblock, NULL, NULL); 718 if (error == EOPNOTSUPP) { 719 VM_OBJECT_LOCK(object); 720 721 for (i = 0; i < count; i++) 722 if (i != reqpage) { 723 vm_page_lock(m[i]); 724 vm_page_free(m[i]); 725 vm_page_unlock(m[i]); 726 } 727 PCPU_INC(cnt.v_vnodein); 728 PCPU_INC(cnt.v_vnodepgsin); 729 error = vnode_pager_input_old(object, m[reqpage]); 730 VM_OBJECT_UNLOCK(object); 731 return (error); 732 } else if (error != 0) { 733 VM_OBJECT_LOCK(object); 734 for (i = 0; i < count; i++) 735 if (i != reqpage) { 736 vm_page_lock(m[i]); 737 vm_page_free(m[i]); 738 vm_page_unlock(m[i]); 739 } 740 VM_OBJECT_UNLOCK(object); 741 return (VM_PAGER_ERROR); 742 743 /* 744 * if the blocksize is smaller than a page size, then use 745 * special small filesystem code. NFS sometimes has a small 746 * blocksize, but it can handle large reads itself. 747 */ 748 } else if ((PAGE_SIZE / bsize) > 1 && 749 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) { 750 VM_OBJECT_LOCK(object); 751 for (i = 0; i < count; i++) 752 if (i != reqpage) { 753 vm_page_lock(m[i]); 754 vm_page_free(m[i]); 755 vm_page_unlock(m[i]); 756 } 757 VM_OBJECT_UNLOCK(object); 758 PCPU_INC(cnt.v_vnodein); 759 PCPU_INC(cnt.v_vnodepgsin); 760 return vnode_pager_input_smlfs(object, m[reqpage]); 761 } 762 763 /* 764 * If we have a completely valid page available to us, we can 765 * clean up and return. Otherwise we have to re-read the 766 * media. 767 */ 768 VM_OBJECT_LOCK(object); 769 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) { 770 for (i = 0; i < count; i++) 771 if (i != reqpage) { 772 vm_page_lock(m[i]); 773 vm_page_free(m[i]); 774 vm_page_unlock(m[i]); 775 } 776 VM_OBJECT_UNLOCK(object); 777 return VM_PAGER_OK; 778 } else if (reqblock == -1) { 779 pmap_zero_page(m[reqpage]); 780 KASSERT(m[reqpage]->dirty == 0, 781 ("vnode_pager_generic_getpages: page %p is dirty", m)); 782 m[reqpage]->valid = VM_PAGE_BITS_ALL; 783 for (i = 0; i < count; i++) 784 if (i != reqpage) { 785 vm_page_lock(m[i]); 786 vm_page_free(m[i]); 787 vm_page_unlock(m[i]); 788 } 789 VM_OBJECT_UNLOCK(object); 790 return (VM_PAGER_OK); 791 } 792 m[reqpage]->valid = 0; 793 VM_OBJECT_UNLOCK(object); 794 795 /* 796 * here on direct device I/O 797 */ 798 firstaddr = -1; 799 800 /* 801 * calculate the run that includes the required page 802 */ 803 for (first = 0, i = 0; i < count; i = runend) { 804 if (vnode_pager_addr(vp, IDX_TO_OFF(m[i]->pindex), &firstaddr, 805 &runpg) != 0) { 806 VM_OBJECT_LOCK(object); 807 for (; i < count; i++) 808 if (i != reqpage) { 809 vm_page_lock(m[i]); 810 vm_page_free(m[i]); 811 vm_page_unlock(m[i]); 812 } 813 VM_OBJECT_UNLOCK(object); 814 return (VM_PAGER_ERROR); 815 } 816 if (firstaddr == -1) { 817 VM_OBJECT_LOCK(object); 818 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) { 819 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx", 820 (intmax_t)firstaddr, (uintmax_t)(foff >> 32), 821 (uintmax_t)foff, 822 (uintmax_t) 823 (object->un_pager.vnp.vnp_size >> 32), 824 (uintmax_t)object->un_pager.vnp.vnp_size); 825 } 826 vm_page_lock(m[i]); 827 vm_page_free(m[i]); 828 vm_page_unlock(m[i]); 829 VM_OBJECT_UNLOCK(object); 830 runend = i + 1; 831 first = runend; 832 continue; 833 } 834 runend = i + runpg; 835 if (runend <= reqpage) { 836 VM_OBJECT_LOCK(object); 837 for (j = i; j < runend; j++) { 838 vm_page_lock(m[j]); 839 vm_page_free(m[j]); 840 vm_page_unlock(m[j]); 841 } 842 VM_OBJECT_UNLOCK(object); 843 } else { 844 if (runpg < (count - first)) { 845 VM_OBJECT_LOCK(object); 846 for (i = first + runpg; i < count; i++) { 847 vm_page_lock(m[i]); 848 vm_page_free(m[i]); 849 vm_page_unlock(m[i]); 850 } 851 VM_OBJECT_UNLOCK(object); 852 count = first + runpg; 853 } 854 break; 855 } 856 first = runend; 857 } 858 859 /* 860 * the first and last page have been calculated now, move input pages 861 * to be zero based... 862 */ 863 if (first != 0) { 864 m += first; 865 count -= first; 866 reqpage -= first; 867 } 868 869 /* 870 * calculate the file virtual address for the transfer 871 */ 872 foff = IDX_TO_OFF(m[0]->pindex); 873 874 /* 875 * calculate the size of the transfer 876 */ 877 size = count * PAGE_SIZE; 878 KASSERT(count > 0, ("zero count")); 879 if ((foff + size) > object->un_pager.vnp.vnp_size) 880 size = object->un_pager.vnp.vnp_size - foff; 881 KASSERT(size > 0, ("zero size")); 882 883 /* 884 * round up physical size for real devices. 885 */ 886 if (1) { 887 int secmask = bo->bo_bsize - 1; 888 KASSERT(secmask < PAGE_SIZE && secmask > 0, 889 ("vnode_pager_generic_getpages: sector size %d too large", 890 secmask + 1)); 891 size = (size + secmask) & ~secmask; 892 } 893 894 bp = getpbuf(&vnode_pbuf_freecnt); 895 kva = (vm_offset_t) bp->b_data; 896 897 /* 898 * and map the pages to be read into the kva 899 */ 900 pmap_qenter(kva, m, count); 901 902 /* build a minimal buffer header */ 903 bp->b_iocmd = BIO_READ; 904 bp->b_iodone = bdone; 905 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 906 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 907 bp->b_rcred = crhold(curthread->td_ucred); 908 bp->b_wcred = crhold(curthread->td_ucred); 909 bp->b_blkno = firstaddr; 910 pbgetbo(bo, bp); 911 bp->b_bcount = size; 912 bp->b_bufsize = size; 913 bp->b_runningbufspace = bp->b_bufsize; 914 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 915 916 PCPU_INC(cnt.v_vnodein); 917 PCPU_ADD(cnt.v_vnodepgsin, count); 918 919 /* do the input */ 920 bp->b_iooffset = dbtob(bp->b_blkno); 921 bstrategy(bp); 922 923 bwait(bp, PVM, "vnread"); 924 925 if ((bp->b_ioflags & BIO_ERROR) != 0) 926 error = EIO; 927 928 if (!error) { 929 if (size != count * PAGE_SIZE) 930 bzero((caddr_t) kva + size, PAGE_SIZE * count - size); 931 } 932 pmap_qremove(kva, count); 933 934 /* 935 * free the buffer header back to the swap buffer pool 936 */ 937 pbrelbo(bp); 938 relpbuf(bp, &vnode_pbuf_freecnt); 939 940 VM_OBJECT_LOCK(object); 941 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) { 942 vm_page_t mt; 943 944 nextoff = tfoff + PAGE_SIZE; 945 mt = m[i]; 946 947 if (nextoff <= object->un_pager.vnp.vnp_size) { 948 /* 949 * Read filled up entire page. 950 */ 951 mt->valid = VM_PAGE_BITS_ALL; 952 KASSERT(mt->dirty == 0, 953 ("vnode_pager_generic_getpages: page %p is dirty", 954 mt)); 955 KASSERT(!pmap_page_is_mapped(mt), 956 ("vnode_pager_generic_getpages: page %p is mapped", 957 mt)); 958 } else { 959 /* 960 * Read did not fill up entire page. 961 * 962 * Currently we do not set the entire page valid, 963 * we just try to clear the piece that we couldn't 964 * read. 965 */ 966 vm_page_set_valid(mt, 0, 967 object->un_pager.vnp.vnp_size - tfoff); 968 KASSERT((mt->dirty & vm_page_bits(0, 969 object->un_pager.vnp.vnp_size - tfoff)) == 0, 970 ("vnode_pager_generic_getpages: page %p is dirty", 971 mt)); 972 } 973 974 if (i != reqpage) { 975 976 /* 977 * whether or not to leave the page activated is up in 978 * the air, but we should put the page on a page queue 979 * somewhere. (it already is in the object). Result: 980 * It appears that empirical results show that 981 * deactivating pages is best. 982 */ 983 984 /* 985 * just in case someone was asking for this page we 986 * now tell them that it is ok to use 987 */ 988 if (!error) { 989 if (mt->oflags & VPO_WANTED) { 990 vm_page_lock(mt); 991 vm_page_activate(mt); 992 vm_page_unlock(mt); 993 } else { 994 vm_page_lock(mt); 995 vm_page_deactivate(mt); 996 vm_page_unlock(mt); 997 } 998 vm_page_wakeup(mt); 999 } else { 1000 vm_page_lock(mt); 1001 vm_page_free(mt); 1002 vm_page_unlock(mt); 1003 } 1004 } 1005 } 1006 VM_OBJECT_UNLOCK(object); 1007 if (error) { 1008 printf("vnode_pager_getpages: I/O read error\n"); 1009 } 1010 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 1011 } 1012 1013 /* 1014 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 1015 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 1016 * vnode_pager_generic_putpages() to implement the previous behaviour. 1017 * 1018 * All other FS's should use the bypass to get to the local media 1019 * backing vp's VOP_PUTPAGES. 1020 */ 1021 static void 1022 vnode_pager_putpages(object, m, count, sync, rtvals) 1023 vm_object_t object; 1024 vm_page_t *m; 1025 int count; 1026 boolean_t sync; 1027 int *rtvals; 1028 { 1029 int rtval; 1030 struct vnode *vp; 1031 int bytes = count * PAGE_SIZE; 1032 1033 /* 1034 * Force synchronous operation if we are extremely low on memory 1035 * to prevent a low-memory deadlock. VOP operations often need to 1036 * allocate more memory to initiate the I/O ( i.e. do a BMAP 1037 * operation ). The swapper handles the case by limiting the amount 1038 * of asynchronous I/O, but that sort of solution doesn't scale well 1039 * for the vnode pager without a lot of work. 1040 * 1041 * Also, the backing vnode's iodone routine may not wake the pageout 1042 * daemon up. This should be probably be addressed XXX. 1043 */ 1044 1045 if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min) 1046 sync |= OBJPC_SYNC; 1047 1048 /* 1049 * Call device-specific putpages function 1050 */ 1051 vp = object->handle; 1052 VM_OBJECT_UNLOCK(object); 1053 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 1054 KASSERT(rtval != EOPNOTSUPP, 1055 ("vnode_pager: stale FS putpages\n")); 1056 VM_OBJECT_LOCK(object); 1057 } 1058 1059 1060 /* 1061 * This is now called from local media FS's to operate against their 1062 * own vnodes if they fail to implement VOP_PUTPAGES. 1063 * 1064 * This is typically called indirectly via the pageout daemon and 1065 * clustering has already typically occured, so in general we ask the 1066 * underlying filesystem to write the data out asynchronously rather 1067 * then delayed. 1068 */ 1069 int 1070 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount, 1071 int flags, int *rtvals) 1072 { 1073 int i; 1074 vm_object_t object; 1075 vm_page_t m; 1076 int count; 1077 1078 int maxsize, ncount; 1079 vm_ooffset_t poffset; 1080 struct uio auio; 1081 struct iovec aiov; 1082 int error; 1083 int ioflags; 1084 int ppscheck = 0; 1085 static struct timeval lastfail; 1086 static int curfail; 1087 1088 object = vp->v_object; 1089 count = bytecount / PAGE_SIZE; 1090 1091 for (i = 0; i < count; i++) 1092 rtvals[i] = VM_PAGER_ERROR; 1093 1094 if ((int64_t)ma[0]->pindex < 0) { 1095 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n", 1096 (long)ma[0]->pindex, (u_long)ma[0]->dirty); 1097 rtvals[0] = VM_PAGER_BAD; 1098 return VM_PAGER_BAD; 1099 } 1100 1101 maxsize = count * PAGE_SIZE; 1102 ncount = count; 1103 1104 poffset = IDX_TO_OFF(ma[0]->pindex); 1105 1106 /* 1107 * If the page-aligned write is larger then the actual file we 1108 * have to invalidate pages occuring beyond the file EOF. However, 1109 * there is an edge case where a file may not be page-aligned where 1110 * the last page is partially invalid. In this case the filesystem 1111 * may not properly clear the dirty bits for the entire page (which 1112 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 1113 * With the page locked we are free to fix-up the dirty bits here. 1114 * 1115 * We do not under any circumstances truncate the valid bits, as 1116 * this will screw up bogus page replacement. 1117 */ 1118 VM_OBJECT_LOCK(object); 1119 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 1120 if (object->un_pager.vnp.vnp_size > poffset) { 1121 int pgoff; 1122 1123 maxsize = object->un_pager.vnp.vnp_size - poffset; 1124 ncount = btoc(maxsize); 1125 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 1126 /* 1127 * If the object is locked and the following 1128 * conditions hold, then the page's dirty 1129 * field cannot be concurrently changed by a 1130 * pmap operation. 1131 */ 1132 m = ma[ncount - 1]; 1133 KASSERT(m->busy > 0, 1134 ("vnode_pager_generic_putpages: page %p is not busy", m)); 1135 KASSERT((m->flags & PG_WRITEABLE) == 0, 1136 ("vnode_pager_generic_putpages: page %p is not read-only", m)); 1137 vm_page_clear_dirty(m, pgoff, PAGE_SIZE - 1138 pgoff); 1139 } 1140 } else { 1141 maxsize = 0; 1142 ncount = 0; 1143 } 1144 if (ncount < count) { 1145 for (i = ncount; i < count; i++) { 1146 rtvals[i] = VM_PAGER_BAD; 1147 } 1148 } 1149 } 1150 VM_OBJECT_UNLOCK(object); 1151 1152 /* 1153 * pageouts are already clustered, use IO_ASYNC t o force a bawrite() 1154 * rather then a bdwrite() to prevent paging I/O from saturating 1155 * the buffer cache. Dummy-up the sequential heuristic to cause 1156 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 1157 * the system decides how to cluster. 1158 */ 1159 ioflags = IO_VMIO; 1160 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 1161 ioflags |= IO_SYNC; 1162 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 1163 ioflags |= IO_ASYNC; 1164 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 1165 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 1166 1167 aiov.iov_base = (caddr_t) 0; 1168 aiov.iov_len = maxsize; 1169 auio.uio_iov = &aiov; 1170 auio.uio_iovcnt = 1; 1171 auio.uio_offset = poffset; 1172 auio.uio_segflg = UIO_NOCOPY; 1173 auio.uio_rw = UIO_WRITE; 1174 auio.uio_resid = maxsize; 1175 auio.uio_td = (struct thread *) 0; 1176 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred); 1177 PCPU_INC(cnt.v_vnodeout); 1178 PCPU_ADD(cnt.v_vnodepgsout, ncount); 1179 1180 if (error) { 1181 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1))) 1182 printf("vnode_pager_putpages: I/O error %d\n", error); 1183 } 1184 if (auio.uio_resid) { 1185 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1)) 1186 printf("vnode_pager_putpages: residual I/O %zd at %lu\n", 1187 auio.uio_resid, (u_long)ma[0]->pindex); 1188 } 1189 for (i = 0; i < ncount; i++) { 1190 rtvals[i] = VM_PAGER_OK; 1191 } 1192 return rtvals[0]; 1193 } 1194 1195 void 1196 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written) 1197 { 1198 int i, pos; 1199 1200 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) { 1201 if (pos < trunc_page(written)) { 1202 rtvals[i] = VM_PAGER_OK; 1203 vm_page_undirty(ma[i]); 1204 } else { 1205 /* Partially written page. */ 1206 rtvals[i] = VM_PAGER_AGAIN; 1207 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK); 1208 } 1209 } 1210 } 1211