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, vm_ooffset_t); 87 88 struct pagerops vnodepagerops = { 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 96 int vnode_pbuf_freecnt; 97 98 /* Create the VM system backing object for this vnode */ 99 int 100 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td) 101 { 102 vm_object_t object; 103 vm_ooffset_t size = isize; 104 struct vattr va; 105 106 if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE) 107 return (0); 108 109 while ((object = vp->v_object) != NULL) { 110 VM_OBJECT_LOCK(object); 111 if (!(object->flags & OBJ_DEAD)) { 112 VM_OBJECT_UNLOCK(object); 113 return (0); 114 } 115 VOP_UNLOCK(vp, 0, td); 116 vm_object_set_flag(object, OBJ_DISCONNECTWNT); 117 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vodead", 0); 118 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 119 } 120 121 if (size == 0) { 122 if (vn_isdisk(vp, NULL)) { 123 size = IDX_TO_OFF(INT_MAX); 124 } else { 125 if (VOP_GETATTR(vp, &va, td->td_ucred, td) != 0) 126 return (0); 127 size = va.va_size; 128 } 129 } 130 131 object = vnode_pager_alloc(vp, size, 0, 0); 132 /* 133 * Dereference the reference we just created. This assumes 134 * that the object is associated with the vp. 135 */ 136 VM_OBJECT_LOCK(object); 137 object->ref_count--; 138 VM_OBJECT_UNLOCK(object); 139 vrele(vp); 140 141 KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object")); 142 143 return (0); 144 } 145 146 void 147 vnode_destroy_vobject(struct vnode *vp) 148 { 149 struct vm_object *obj; 150 151 obj = vp->v_object; 152 if (obj == NULL) 153 return; 154 ASSERT_VOP_LOCKED(vp, "vnode_destroy_vobject"); 155 VM_OBJECT_LOCK(obj); 156 if (obj->ref_count == 0) { 157 /* 158 * vclean() may be called twice. The first time 159 * removes the primary reference to the object, 160 * the second time goes one further and is a 161 * special-case to terminate the object. 162 * 163 * don't double-terminate the object 164 */ 165 if ((obj->flags & OBJ_DEAD) == 0) 166 vm_object_terminate(obj); 167 else 168 VM_OBJECT_UNLOCK(obj); 169 } else { 170 /* 171 * Woe to the process that tries to page now :-). 172 */ 173 vm_pager_deallocate(obj); 174 VM_OBJECT_UNLOCK(obj); 175 } 176 vp->v_object = NULL; 177 } 178 179 180 /* 181 * Allocate (or lookup) pager for a vnode. 182 * Handle is a vnode pointer. 183 * 184 * MPSAFE 185 */ 186 vm_object_t 187 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 188 vm_ooffset_t offset) 189 { 190 vm_object_t object; 191 struct vnode *vp; 192 193 /* 194 * Pageout to vnode, no can do yet. 195 */ 196 if (handle == NULL) 197 return (NULL); 198 199 vp = (struct vnode *) handle; 200 201 ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc"); 202 203 /* 204 * If the object is being terminated, wait for it to 205 * go away. 206 */ 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 * And 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 if (VFS_NEEDSGIANT(vp->v_mount)) 228 vm_object_set_flag(object, OBJ_NEEDGIANT); 229 vp->v_object = object; 230 } else { 231 object->ref_count++; 232 VM_OBJECT_UNLOCK(object); 233 } 234 vref(vp); 235 return (object); 236 } 237 238 /* 239 * The object must be locked. 240 */ 241 static void 242 vnode_pager_dealloc(object) 243 vm_object_t object; 244 { 245 struct vnode *vp = object->handle; 246 247 if (vp == NULL) 248 panic("vnode_pager_dealloc: pager already dealloced"); 249 250 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 251 vm_object_pip_wait(object, "vnpdea"); 252 253 object->handle = NULL; 254 object->type = OBJT_DEAD; 255 if (object->flags & OBJ_DISCONNECTWNT) { 256 vm_object_clear_flag(object, OBJ_DISCONNECTWNT); 257 wakeup(object); 258 } 259 ASSERT_VOP_LOCKED(vp, "vnode_pager_dealloc"); 260 vp->v_object = NULL; 261 vp->v_vflag &= ~VV_TEXT; 262 } 263 264 static boolean_t 265 vnode_pager_haspage(object, pindex, before, after) 266 vm_object_t object; 267 vm_pindex_t pindex; 268 int *before; 269 int *after; 270 { 271 struct vnode *vp = object->handle; 272 daddr_t bn; 273 int err; 274 daddr_t reqblock; 275 int poff; 276 int bsize; 277 int pagesperblock, blocksperpage; 278 int vfslocked; 279 280 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 281 /* 282 * If no vp or vp is doomed or marked transparent to VM, we do not 283 * have the page. 284 */ 285 if (vp == NULL || vp->v_iflag & VI_DOOMED) 286 return FALSE; 287 /* 288 * If the offset is beyond end of file we do 289 * not have the page. 290 */ 291 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size) 292 return FALSE; 293 294 bsize = vp->v_mount->mnt_stat.f_iosize; 295 pagesperblock = bsize / PAGE_SIZE; 296 blocksperpage = 0; 297 if (pagesperblock > 0) { 298 reqblock = pindex / pagesperblock; 299 } else { 300 blocksperpage = (PAGE_SIZE / bsize); 301 reqblock = pindex * blocksperpage; 302 } 303 VM_OBJECT_UNLOCK(object); 304 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 305 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before); 306 VFS_UNLOCK_GIANT(vfslocked); 307 VM_OBJECT_LOCK(object); 308 if (err) 309 return TRUE; 310 if (bn == -1) 311 return FALSE; 312 if (pagesperblock > 0) { 313 poff = pindex - (reqblock * pagesperblock); 314 if (before) { 315 *before *= pagesperblock; 316 *before += poff; 317 } 318 if (after) { 319 int numafter; 320 *after *= pagesperblock; 321 numafter = pagesperblock - (poff + 1); 322 if (IDX_TO_OFF(pindex + numafter) > 323 object->un_pager.vnp.vnp_size) { 324 numafter = 325 OFF_TO_IDX(object->un_pager.vnp.vnp_size) - 326 pindex; 327 } 328 *after += numafter; 329 } 330 } else { 331 if (before) { 332 *before /= blocksperpage; 333 } 334 335 if (after) { 336 *after /= blocksperpage; 337 } 338 } 339 return TRUE; 340 } 341 342 /* 343 * Lets the VM system know about a change in size for a file. 344 * We adjust our own internal size and flush any cached pages in 345 * the associated object that are affected by the size change. 346 * 347 * Note: this routine may be invoked as a result of a pager put 348 * operation (possibly at object termination time), so we must be careful. 349 */ 350 void 351 vnode_pager_setsize(vp, nsize) 352 struct vnode *vp; 353 vm_ooffset_t nsize; 354 { 355 vm_object_t object; 356 vm_page_t m; 357 vm_pindex_t nobjsize; 358 359 if ((object = vp->v_object) == NULL) 360 return; 361 VM_OBJECT_LOCK(object); 362 if (nsize == object->un_pager.vnp.vnp_size) { 363 /* 364 * Hasn't changed size 365 */ 366 VM_OBJECT_UNLOCK(object); 367 return; 368 } 369 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 370 if (nsize < object->un_pager.vnp.vnp_size) { 371 /* 372 * File has shrunk. Toss any cached pages beyond the new EOF. 373 */ 374 if (nobjsize < object->size) 375 vm_object_page_remove(object, nobjsize, object->size, 376 FALSE); 377 /* 378 * this gets rid of garbage at the end of a page that is now 379 * only partially backed by the vnode. 380 * 381 * XXX for some reason (I don't know yet), if we take a 382 * completely invalid page and mark it partially valid 383 * it can screw up NFS reads, so we don't allow the case. 384 */ 385 if ((nsize & PAGE_MASK) && 386 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL && 387 m->valid != 0) { 388 int base = (int)nsize & PAGE_MASK; 389 int size = PAGE_SIZE - base; 390 391 /* 392 * Clear out partial-page garbage in case 393 * the page has been mapped. 394 */ 395 pmap_zero_page_area(m, base, size); 396 397 /* 398 * XXX work around SMP data integrity race 399 * by unmapping the page from user processes. 400 * The garbage we just cleared may be mapped 401 * to a user process running on another cpu 402 * and this code is not running through normal 403 * I/O channels which handle SMP issues for 404 * us, so unmap page to synchronize all cpus. 405 * 406 * XXX should vm_pager_unmap_page() have 407 * dealt with this? 408 */ 409 vm_page_lock_queues(); 410 pmap_remove_all(m); 411 412 /* 413 * Clear out partial-page dirty bits. This 414 * has the side effect of setting the valid 415 * bits, but that is ok. There are a bunch 416 * of places in the VM system where we expected 417 * m->dirty == VM_PAGE_BITS_ALL. The file EOF 418 * case is one of them. If the page is still 419 * partially dirty, make it fully dirty. 420 * 421 * note that we do not clear out the valid 422 * bits. This would prevent bogus_page 423 * replacement from working properly. 424 */ 425 vm_page_set_validclean(m, base, size); 426 if (m->dirty != 0) 427 m->dirty = VM_PAGE_BITS_ALL; 428 vm_page_unlock_queues(); 429 } 430 } 431 object->un_pager.vnp.vnp_size = nsize; 432 object->size = nobjsize; 433 VM_OBJECT_UNLOCK(object); 434 } 435 436 /* 437 * calculate the linear (byte) disk address of specified virtual 438 * file address 439 */ 440 static int 441 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress, 442 int *run) 443 { 444 int bsize; 445 int err; 446 daddr_t vblock; 447 daddr_t voffset; 448 449 if (address < 0) 450 return -1; 451 452 if (vp->v_iflag & VI_DOOMED) 453 return -1; 454 455 bsize = vp->v_mount->mnt_stat.f_iosize; 456 vblock = address / bsize; 457 voffset = address % bsize; 458 459 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL); 460 if (err == 0) { 461 if (*rtaddress != -1) 462 *rtaddress += voffset / DEV_BSIZE; 463 if (run) { 464 *run += 1; 465 *run *= bsize/PAGE_SIZE; 466 *run -= voffset/PAGE_SIZE; 467 } 468 } 469 470 return (err); 471 } 472 473 /* 474 * small block filesystem vnode pager input 475 */ 476 static int 477 vnode_pager_input_smlfs(object, m) 478 vm_object_t object; 479 vm_page_t m; 480 { 481 int i; 482 struct vnode *vp; 483 struct bufobj *bo; 484 struct buf *bp; 485 struct sf_buf *sf; 486 daddr_t fileaddr; 487 vm_offset_t bsize; 488 int error = 0; 489 490 vp = object->handle; 491 if (vp->v_iflag & VI_DOOMED) 492 return VM_PAGER_BAD; 493 494 bsize = vp->v_mount->mnt_stat.f_iosize; 495 496 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL); 497 498 sf = sf_buf_alloc(m, 0); 499 500 for (i = 0; i < PAGE_SIZE / bsize; i++) { 501 vm_ooffset_t address; 502 503 if (vm_page_bits(i * bsize, bsize) & m->valid) 504 continue; 505 506 address = IDX_TO_OFF(m->pindex) + i * bsize; 507 if (address >= object->un_pager.vnp.vnp_size) { 508 fileaddr = -1; 509 } else { 510 error = vnode_pager_addr(vp, address, &fileaddr, NULL); 511 if (error) 512 break; 513 } 514 if (fileaddr != -1) { 515 bp = getpbuf(&vnode_pbuf_freecnt); 516 517 /* build a minimal buffer header */ 518 bp->b_iocmd = BIO_READ; 519 bp->b_iodone = bdone; 520 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 521 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 522 bp->b_rcred = crhold(curthread->td_ucred); 523 bp->b_wcred = crhold(curthread->td_ucred); 524 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize; 525 bp->b_blkno = fileaddr; 526 pbgetbo(bo, bp); 527 bp->b_bcount = bsize; 528 bp->b_bufsize = bsize; 529 bp->b_runningbufspace = bp->b_bufsize; 530 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 531 532 /* do the input */ 533 bp->b_iooffset = dbtob(bp->b_blkno); 534 bstrategy(bp); 535 536 bwait(bp, PVM, "vnsrd"); 537 538 if ((bp->b_ioflags & BIO_ERROR) != 0) 539 error = EIO; 540 541 /* 542 * free the buffer header back to the swap buffer pool 543 */ 544 pbrelbo(bp); 545 relpbuf(bp, &vnode_pbuf_freecnt); 546 if (error) 547 break; 548 549 VM_OBJECT_LOCK(object); 550 vm_page_lock_queues(); 551 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 552 vm_page_unlock_queues(); 553 VM_OBJECT_UNLOCK(object); 554 } else { 555 VM_OBJECT_LOCK(object); 556 vm_page_lock_queues(); 557 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 558 vm_page_unlock_queues(); 559 VM_OBJECT_UNLOCK(object); 560 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize); 561 } 562 } 563 sf_buf_free(sf); 564 vm_page_lock_queues(); 565 pmap_clear_modify(m); 566 vm_page_unlock_queues(); 567 if (error) { 568 return VM_PAGER_ERROR; 569 } 570 return VM_PAGER_OK; 571 572 } 573 574 575 /* 576 * old style vnode pager input routine 577 */ 578 static int 579 vnode_pager_input_old(object, m) 580 vm_object_t object; 581 vm_page_t m; 582 { 583 struct uio auio; 584 struct iovec aiov; 585 int error; 586 int size; 587 struct sf_buf *sf; 588 struct vnode *vp; 589 590 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 591 error = 0; 592 593 /* 594 * Return failure if beyond current EOF 595 */ 596 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 597 return VM_PAGER_BAD; 598 } else { 599 size = PAGE_SIZE; 600 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 601 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 602 vp = object->handle; 603 VM_OBJECT_UNLOCK(object); 604 605 /* 606 * Allocate a kernel virtual address and initialize so that 607 * we can use VOP_READ/WRITE routines. 608 */ 609 sf = sf_buf_alloc(m, 0); 610 611 aiov.iov_base = (caddr_t)sf_buf_kva(sf); 612 aiov.iov_len = size; 613 auio.uio_iov = &aiov; 614 auio.uio_iovcnt = 1; 615 auio.uio_offset = IDX_TO_OFF(m->pindex); 616 auio.uio_segflg = UIO_SYSSPACE; 617 auio.uio_rw = UIO_READ; 618 auio.uio_resid = size; 619 auio.uio_td = curthread; 620 621 error = VOP_READ(vp, &auio, 0, curthread->td_ucred); 622 if (!error) { 623 int count = size - auio.uio_resid; 624 625 if (count == 0) 626 error = EINVAL; 627 else if (count != PAGE_SIZE) 628 bzero((caddr_t)sf_buf_kva(sf) + count, 629 PAGE_SIZE - count); 630 } 631 sf_buf_free(sf); 632 633 VM_OBJECT_LOCK(object); 634 } 635 vm_page_lock_queues(); 636 pmap_clear_modify(m); 637 vm_page_undirty(m); 638 vm_page_unlock_queues(); 639 if (!error) 640 m->valid = VM_PAGE_BITS_ALL; 641 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 642 } 643 644 /* 645 * generic vnode pager input routine 646 */ 647 648 /* 649 * Local media VFS's that do not implement their own VOP_GETPAGES 650 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages() 651 * to implement the previous behaviour. 652 * 653 * All other FS's should use the bypass to get to the local media 654 * backing vp's VOP_GETPAGES. 655 */ 656 static int 657 vnode_pager_getpages(object, m, count, reqpage) 658 vm_object_t object; 659 vm_page_t *m; 660 int count; 661 int reqpage; 662 { 663 int rtval; 664 struct vnode *vp; 665 int bytes = count * PAGE_SIZE; 666 int vfslocked; 667 668 vp = object->handle; 669 VM_OBJECT_UNLOCK(object); 670 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 671 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0); 672 KASSERT(rtval != EOPNOTSUPP, 673 ("vnode_pager: FS getpages not implemented\n")); 674 VFS_UNLOCK_GIANT(vfslocked); 675 VM_OBJECT_LOCK(object); 676 return rtval; 677 } 678 679 /* 680 * This is now called from local media FS's to operate against their 681 * own vnodes if they fail to implement VOP_GETPAGES. 682 */ 683 int 684 vnode_pager_generic_getpages(vp, m, bytecount, reqpage) 685 struct vnode *vp; 686 vm_page_t *m; 687 int bytecount; 688 int reqpage; 689 { 690 vm_object_t object; 691 vm_offset_t kva; 692 off_t foff, tfoff, nextoff; 693 int i, j, size, bsize, first; 694 daddr_t firstaddr, reqblock; 695 struct bufobj *bo; 696 int runpg; 697 int runend; 698 struct buf *bp; 699 int count; 700 int error; 701 702 object = vp->v_object; 703 count = bytecount / PAGE_SIZE; 704 705 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 706 ("vnode_pager_generic_getpages does not support devices")); 707 if (vp->v_iflag & VI_DOOMED) 708 return VM_PAGER_BAD; 709 710 bsize = vp->v_mount->mnt_stat.f_iosize; 711 712 /* get the UNDERLYING device for the file with VOP_BMAP() */ 713 714 /* 715 * originally, we did not check for an error return value -- assuming 716 * an fs always has a bmap entry point -- that assumption is wrong!!! 717 */ 718 foff = IDX_TO_OFF(m[reqpage]->pindex); 719 720 /* 721 * if we can't bmap, use old VOP code 722 */ 723 error = VOP_BMAP(vp, foff / bsize, &bo, &reqblock, NULL, NULL); 724 if (error == EOPNOTSUPP) { 725 VM_OBJECT_LOCK(object); 726 vm_page_lock_queues(); 727 for (i = 0; i < count; i++) 728 if (i != reqpage) 729 vm_page_free(m[i]); 730 vm_page_unlock_queues(); 731 PCPU_INC(cnt.v_vnodein); 732 PCPU_INC(cnt.v_vnodepgsin); 733 error = vnode_pager_input_old(object, m[reqpage]); 734 VM_OBJECT_UNLOCK(object); 735 return (error); 736 } else if (error != 0) { 737 VM_OBJECT_LOCK(object); 738 vm_page_lock_queues(); 739 for (i = 0; i < count; i++) 740 if (i != reqpage) 741 vm_page_free(m[i]); 742 vm_page_unlock_queues(); 743 VM_OBJECT_UNLOCK(object); 744 return (VM_PAGER_ERROR); 745 746 /* 747 * if the blocksize is smaller than a page size, then use 748 * special small filesystem code. NFS sometimes has a small 749 * blocksize, but it can handle large reads itself. 750 */ 751 } else if ((PAGE_SIZE / bsize) > 1 && 752 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) { 753 VM_OBJECT_LOCK(object); 754 vm_page_lock_queues(); 755 for (i = 0; i < count; i++) 756 if (i != reqpage) 757 vm_page_free(m[i]); 758 vm_page_unlock_queues(); 759 VM_OBJECT_UNLOCK(object); 760 PCPU_INC(cnt.v_vnodein); 761 PCPU_INC(cnt.v_vnodepgsin); 762 return vnode_pager_input_smlfs(object, m[reqpage]); 763 } 764 765 /* 766 * If we have a completely valid page available to us, we can 767 * clean up and return. Otherwise we have to re-read the 768 * media. 769 */ 770 VM_OBJECT_LOCK(object); 771 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) { 772 vm_page_lock_queues(); 773 for (i = 0; i < count; i++) 774 if (i != reqpage) 775 vm_page_free(m[i]); 776 vm_page_unlock_queues(); 777 VM_OBJECT_UNLOCK(object); 778 return VM_PAGER_OK; 779 } else if (reqblock == -1) { 780 pmap_zero_page(m[reqpage]); 781 vm_page_undirty(m[reqpage]); 782 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 if (vnode_pager_addr(vp, IDX_TO_OFF(m[i]->pindex), &firstaddr, 804 &runpg) != 0) { 805 VM_OBJECT_LOCK(object); 806 vm_page_lock_queues(); 807 for (; i < count; i++) 808 if (i != reqpage) 809 vm_page_free(m[i]); 810 vm_page_unlock_queues(); 811 VM_OBJECT_UNLOCK(object); 812 return (VM_PAGER_ERROR); 813 } 814 if (firstaddr == -1) { 815 VM_OBJECT_LOCK(object); 816 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) { 817 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx", 818 (intmax_t)firstaddr, (uintmax_t)(foff >> 32), 819 (uintmax_t)foff, 820 (uintmax_t) 821 (object->un_pager.vnp.vnp_size >> 32), 822 (uintmax_t)object->un_pager.vnp.vnp_size); 823 } 824 vm_page_lock_queues(); 825 vm_page_free(m[i]); 826 vm_page_unlock_queues(); 827 VM_OBJECT_UNLOCK(object); 828 runend = i + 1; 829 first = runend; 830 continue; 831 } 832 runend = i + runpg; 833 if (runend <= reqpage) { 834 VM_OBJECT_LOCK(object); 835 vm_page_lock_queues(); 836 for (j = i; j < runend; j++) 837 vm_page_free(m[j]); 838 vm_page_unlock_queues(); 839 VM_OBJECT_UNLOCK(object); 840 } else { 841 if (runpg < (count - first)) { 842 VM_OBJECT_LOCK(object); 843 vm_page_lock_queues(); 844 for (i = first + runpg; i < count; i++) 845 vm_page_free(m[i]); 846 vm_page_unlock_queues(); 847 VM_OBJECT_UNLOCK(object); 848 count = first + runpg; 849 } 850 break; 851 } 852 first = runend; 853 } 854 855 /* 856 * the first and last page have been calculated now, move input pages 857 * to be zero based... 858 */ 859 if (first != 0) { 860 m += first; 861 count -= first; 862 reqpage -= first; 863 } 864 865 /* 866 * calculate the file virtual address for the transfer 867 */ 868 foff = IDX_TO_OFF(m[0]->pindex); 869 870 /* 871 * calculate the size of the transfer 872 */ 873 size = count * PAGE_SIZE; 874 KASSERT(count > 0, ("zero count")); 875 if ((foff + size) > object->un_pager.vnp.vnp_size) 876 size = object->un_pager.vnp.vnp_size - foff; 877 KASSERT(size > 0, ("zero size")); 878 879 /* 880 * round up physical size for real devices. 881 */ 882 if (1) { 883 int secmask = bo->bo_bsize - 1; 884 KASSERT(secmask < PAGE_SIZE && secmask > 0, 885 ("vnode_pager_generic_getpages: sector size %d too large", 886 secmask + 1)); 887 size = (size + secmask) & ~secmask; 888 } 889 890 bp = getpbuf(&vnode_pbuf_freecnt); 891 kva = (vm_offset_t) bp->b_data; 892 893 /* 894 * and map the pages to be read into the kva 895 */ 896 pmap_qenter(kva, m, count); 897 898 /* build a minimal buffer header */ 899 bp->b_iocmd = BIO_READ; 900 bp->b_iodone = bdone; 901 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 902 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 903 bp->b_rcred = crhold(curthread->td_ucred); 904 bp->b_wcred = crhold(curthread->td_ucred); 905 bp->b_blkno = firstaddr; 906 pbgetbo(bo, bp); 907 bp->b_bcount = size; 908 bp->b_bufsize = size; 909 bp->b_runningbufspace = bp->b_bufsize; 910 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 911 912 PCPU_INC(cnt.v_vnodein); 913 PCPU_ADD(cnt.v_vnodepgsin, count); 914 915 /* do the input */ 916 bp->b_iooffset = dbtob(bp->b_blkno); 917 bstrategy(bp); 918 919 bwait(bp, PVM, "vnread"); 920 921 if ((bp->b_ioflags & BIO_ERROR) != 0) 922 error = EIO; 923 924 if (!error) { 925 if (size != count * PAGE_SIZE) 926 bzero((caddr_t) kva + size, PAGE_SIZE * count - size); 927 } 928 pmap_qremove(kva, count); 929 930 /* 931 * free the buffer header back to the swap buffer pool 932 */ 933 pbrelbo(bp); 934 relpbuf(bp, &vnode_pbuf_freecnt); 935 936 VM_OBJECT_LOCK(object); 937 vm_page_lock_queues(); 938 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) { 939 vm_page_t mt; 940 941 nextoff = tfoff + PAGE_SIZE; 942 mt = m[i]; 943 944 if (nextoff <= object->un_pager.vnp.vnp_size) { 945 /* 946 * Read filled up entire page. 947 */ 948 mt->valid = VM_PAGE_BITS_ALL; 949 vm_page_undirty(mt); /* should be an assert? XXX */ 950 pmap_clear_modify(mt); 951 } else { 952 /* 953 * Read did not fill up entire page. Since this 954 * is getpages, the page may be mapped, so we have 955 * to zero the invalid portions of the page even 956 * though we aren't setting them valid. 957 * 958 * Currently we do not set the entire page valid, 959 * we just try to clear the piece that we couldn't 960 * read. 961 */ 962 vm_page_set_validclean(mt, 0, 963 object->un_pager.vnp.vnp_size - tfoff); 964 /* handled by vm_fault now */ 965 /* vm_page_zero_invalid(mt, FALSE); */ 966 } 967 968 if (i != reqpage) { 969 970 /* 971 * whether or not to leave the page activated is up in 972 * the air, but we should put the page on a page queue 973 * somewhere. (it already is in the object). Result: 974 * It appears that empirical results show that 975 * deactivating pages is best. 976 */ 977 978 /* 979 * just in case someone was asking for this page we 980 * now tell them that it is ok to use 981 */ 982 if (!error) { 983 if (mt->oflags & VPO_WANTED) 984 vm_page_activate(mt); 985 else 986 vm_page_deactivate(mt); 987 vm_page_wakeup(mt); 988 } else { 989 vm_page_free(mt); 990 } 991 } 992 } 993 vm_page_unlock_queues(); 994 VM_OBJECT_UNLOCK(object); 995 if (error) { 996 printf("vnode_pager_getpages: I/O read error\n"); 997 } 998 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 999 } 1000 1001 /* 1002 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 1003 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 1004 * vnode_pager_generic_putpages() to implement the previous behaviour. 1005 * 1006 * All other FS's should use the bypass to get to the local media 1007 * backing vp's VOP_PUTPAGES. 1008 */ 1009 static void 1010 vnode_pager_putpages(object, m, count, sync, rtvals) 1011 vm_object_t object; 1012 vm_page_t *m; 1013 int count; 1014 boolean_t sync; 1015 int *rtvals; 1016 { 1017 int rtval; 1018 struct vnode *vp; 1019 struct mount *mp; 1020 int bytes = count * PAGE_SIZE; 1021 1022 /* 1023 * Force synchronous operation if we are extremely low on memory 1024 * to prevent a low-memory deadlock. VOP operations often need to 1025 * allocate more memory to initiate the I/O ( i.e. do a BMAP 1026 * operation ). The swapper handles the case by limiting the amount 1027 * of asynchronous I/O, but that sort of solution doesn't scale well 1028 * for the vnode pager without a lot of work. 1029 * 1030 * Also, the backing vnode's iodone routine may not wake the pageout 1031 * daemon up. This should be probably be addressed XXX. 1032 */ 1033 1034 if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min) 1035 sync |= OBJPC_SYNC; 1036 1037 /* 1038 * Call device-specific putpages function 1039 */ 1040 vp = object->handle; 1041 VM_OBJECT_UNLOCK(object); 1042 if (vp->v_type != VREG) 1043 mp = NULL; 1044 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 1045 KASSERT(rtval != EOPNOTSUPP, 1046 ("vnode_pager: stale FS putpages\n")); 1047 VM_OBJECT_LOCK(object); 1048 } 1049 1050 1051 /* 1052 * This is now called from local media FS's to operate against their 1053 * own vnodes if they fail to implement VOP_PUTPAGES. 1054 * 1055 * This is typically called indirectly via the pageout daemon and 1056 * clustering has already typically occured, so in general we ask the 1057 * underlying filesystem to write the data out asynchronously rather 1058 * then delayed. 1059 */ 1060 int 1061 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals) 1062 struct vnode *vp; 1063 vm_page_t *m; 1064 int bytecount; 1065 int flags; 1066 int *rtvals; 1067 { 1068 int i; 1069 vm_object_t object; 1070 int count; 1071 1072 int maxsize, ncount; 1073 vm_ooffset_t poffset; 1074 struct uio auio; 1075 struct iovec aiov; 1076 int error; 1077 int ioflags; 1078 int ppscheck = 0; 1079 static struct timeval lastfail; 1080 static int curfail; 1081 1082 object = vp->v_object; 1083 count = bytecount / PAGE_SIZE; 1084 1085 for (i = 0; i < count; i++) 1086 rtvals[i] = VM_PAGER_AGAIN; 1087 1088 if ((int64_t)m[0]->pindex < 0) { 1089 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n", 1090 (long)m[0]->pindex, (u_long)m[0]->dirty); 1091 rtvals[0] = VM_PAGER_BAD; 1092 return VM_PAGER_BAD; 1093 } 1094 1095 maxsize = count * PAGE_SIZE; 1096 ncount = count; 1097 1098 poffset = IDX_TO_OFF(m[0]->pindex); 1099 1100 /* 1101 * If the page-aligned write is larger then the actual file we 1102 * have to invalidate pages occuring beyond the file EOF. However, 1103 * there is an edge case where a file may not be page-aligned where 1104 * the last page is partially invalid. In this case the filesystem 1105 * may not properly clear the dirty bits for the entire page (which 1106 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 1107 * With the page locked we are free to fix-up the dirty bits here. 1108 * 1109 * We do not under any circumstances truncate the valid bits, as 1110 * this will screw up bogus page replacement. 1111 */ 1112 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 1113 if (object->un_pager.vnp.vnp_size > poffset) { 1114 int pgoff; 1115 1116 maxsize = object->un_pager.vnp.vnp_size - poffset; 1117 ncount = btoc(maxsize); 1118 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 1119 vm_page_lock_queues(); 1120 vm_page_clear_dirty(m[ncount - 1], pgoff, 1121 PAGE_SIZE - pgoff); 1122 vm_page_unlock_queues(); 1123 } 1124 } else { 1125 maxsize = 0; 1126 ncount = 0; 1127 } 1128 if (ncount < count) { 1129 for (i = ncount; i < count; i++) { 1130 rtvals[i] = VM_PAGER_BAD; 1131 } 1132 } 1133 } 1134 1135 /* 1136 * pageouts are already clustered, use IO_ASYNC t o force a bawrite() 1137 * rather then a bdwrite() to prevent paging I/O from saturating 1138 * the buffer cache. Dummy-up the sequential heuristic to cause 1139 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 1140 * the system decides how to cluster. 1141 */ 1142 ioflags = IO_VMIO; 1143 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 1144 ioflags |= IO_SYNC; 1145 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 1146 ioflags |= IO_ASYNC; 1147 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 1148 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 1149 1150 aiov.iov_base = (caddr_t) 0; 1151 aiov.iov_len = maxsize; 1152 auio.uio_iov = &aiov; 1153 auio.uio_iovcnt = 1; 1154 auio.uio_offset = poffset; 1155 auio.uio_segflg = UIO_NOCOPY; 1156 auio.uio_rw = UIO_WRITE; 1157 auio.uio_resid = maxsize; 1158 auio.uio_td = (struct thread *) 0; 1159 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred); 1160 PCPU_INC(cnt.v_vnodeout); 1161 PCPU_ADD(cnt.v_vnodepgsout, ncount); 1162 1163 if (error) { 1164 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1))) 1165 printf("vnode_pager_putpages: I/O error %d\n", error); 1166 } 1167 if (auio.uio_resid) { 1168 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1)) 1169 printf("vnode_pager_putpages: residual I/O %d at %lu\n", 1170 auio.uio_resid, (u_long)m[0]->pindex); 1171 } 1172 for (i = 0; i < ncount; i++) { 1173 rtvals[i] = VM_PAGER_OK; 1174 } 1175 return rtvals[0]; 1176 } 1177 1178 struct vnode * 1179 vnode_pager_lock(vm_object_t first_object) 1180 { 1181 struct vnode *vp; 1182 vm_object_t backing_object, object; 1183 1184 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 1185 for (object = first_object; object != NULL; object = backing_object) { 1186 if (object->type != OBJT_VNODE) { 1187 if ((backing_object = object->backing_object) != NULL) 1188 VM_OBJECT_LOCK(backing_object); 1189 if (object != first_object) 1190 VM_OBJECT_UNLOCK(object); 1191 continue; 1192 } 1193 retry: 1194 if (object->flags & OBJ_DEAD) { 1195 if (object != first_object) 1196 VM_OBJECT_UNLOCK(object); 1197 return NULL; 1198 } 1199 vp = object->handle; 1200 VI_LOCK(vp); 1201 VM_OBJECT_UNLOCK(object); 1202 if (first_object != object) 1203 VM_OBJECT_UNLOCK(first_object); 1204 VFS_ASSERT_GIANT(vp->v_mount); 1205 if (vget(vp, LK_CANRECURSE | LK_INTERLOCK | 1206 LK_RETRY | LK_SHARED, curthread)) { 1207 VM_OBJECT_LOCK(first_object); 1208 if (object != first_object) 1209 VM_OBJECT_LOCK(object); 1210 if (object->type != OBJT_VNODE) { 1211 if (object != first_object) 1212 VM_OBJECT_UNLOCK(object); 1213 return NULL; 1214 } 1215 printf("vnode_pager_lock: retrying\n"); 1216 goto retry; 1217 } 1218 VM_OBJECT_LOCK(first_object); 1219 return (vp); 1220 } 1221 return NULL; 1222 } 1223