vnode_pager.c revision 189595
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 189595 2009-03-09 19:35:20Z jhb $"); 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 78static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, 79 daddr_t *rtaddress, int *run); 80static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m); 81static int vnode_pager_input_old(vm_object_t object, vm_page_t m); 82static void vnode_pager_dealloc(vm_object_t); 83static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int); 84static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 85static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *); 86static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t, vm_ooffset_t); 87 88struct 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 96int vnode_pbuf_freecnt; 97 98/* Create the VM system backing object for this vnode */ 99int 100vnode_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); 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); 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)) 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 146void 147vnode_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_ELOCKED(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 */ 186vm_object_t 187vnode_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 /* 202 * If the object is being terminated, wait for it to 203 * go away. 204 */ 205retry: 206 while ((object = vp->v_object) != NULL) { 207 VM_OBJECT_LOCK(object); 208 if ((object->flags & OBJ_DEAD) == 0) 209 break; 210 vm_object_set_flag(object, OBJ_DISCONNECTWNT); 211 msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vadead", 0); 212 } 213 214 if (vp->v_usecount == 0) 215 panic("vnode_pager_alloc: no vnode reference"); 216 217 if (object == NULL) { 218 /* 219 * Add an object of the appropriate size 220 */ 221 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); 222 223 object->un_pager.vnp.vnp_size = size; 224 225 object->handle = handle; 226 VI_LOCK(vp); 227 if (vp->v_object != NULL) { 228 /* 229 * Object has been created while we were sleeping 230 */ 231 VI_UNLOCK(vp); 232 vm_object_destroy(object); 233 goto retry; 234 } 235 vp->v_object = object; 236 VI_UNLOCK(vp); 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_ELOCKED(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 || vp->v_iflag & VI_DOOMED) 293 return FALSE; 294 /* 295 * If the offset is beyond end of file we do 296 * not have the page. 297 */ 298 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size) 299 return FALSE; 300 301 bsize = vp->v_mount->mnt_stat.f_iosize; 302 pagesperblock = bsize / PAGE_SIZE; 303 blocksperpage = 0; 304 if (pagesperblock > 0) { 305 reqblock = pindex / pagesperblock; 306 } else { 307 blocksperpage = (PAGE_SIZE / bsize); 308 reqblock = pindex * blocksperpage; 309 } 310 VM_OBJECT_UNLOCK(object); 311 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 312 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before); 313 VFS_UNLOCK_GIANT(vfslocked); 314 VM_OBJECT_LOCK(object); 315 if (err) 316 return TRUE; 317 if (bn == -1) 318 return FALSE; 319 if (pagesperblock > 0) { 320 poff = pindex - (reqblock * pagesperblock); 321 if (before) { 322 *before *= pagesperblock; 323 *before += poff; 324 } 325 if (after) { 326 int numafter; 327 *after *= pagesperblock; 328 numafter = pagesperblock - (poff + 1); 329 if (IDX_TO_OFF(pindex + numafter) > 330 object->un_pager.vnp.vnp_size) { 331 numafter = 332 OFF_TO_IDX(object->un_pager.vnp.vnp_size) - 333 pindex; 334 } 335 *after += numafter; 336 } 337 } else { 338 if (before) { 339 *before /= blocksperpage; 340 } 341 342 if (after) { 343 *after /= blocksperpage; 344 } 345 } 346 return TRUE; 347} 348 349/* 350 * Lets the VM system know about a change in size for a file. 351 * We adjust our own internal size and flush any cached pages in 352 * the associated object that are affected by the size change. 353 * 354 * Note: this routine may be invoked as a result of a pager put 355 * operation (possibly at object termination time), so we must be careful. 356 */ 357void 358vnode_pager_setsize(vp, nsize) 359 struct vnode *vp; 360 vm_ooffset_t nsize; 361{ 362 vm_object_t object; 363 vm_page_t m; 364 vm_pindex_t nobjsize; 365 366 if ((object = vp->v_object) == NULL) 367 return; 368/* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */ 369 VM_OBJECT_LOCK(object); 370 if (nsize == object->un_pager.vnp.vnp_size) { 371 /* 372 * Hasn't changed size 373 */ 374 VM_OBJECT_UNLOCK(object); 375 return; 376 } 377 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 378 if (nsize < object->un_pager.vnp.vnp_size) { 379 /* 380 * File has shrunk. Toss any cached pages beyond the new EOF. 381 */ 382 if (nobjsize < object->size) 383 vm_object_page_remove(object, nobjsize, object->size, 384 FALSE); 385 /* 386 * this gets rid of garbage at the end of a page that is now 387 * only partially backed by the vnode. 388 * 389 * XXX for some reason (I don't know yet), if we take a 390 * completely invalid page and mark it partially valid 391 * it can screw up NFS reads, so we don't allow the case. 392 */ 393 if ((nsize & PAGE_MASK) && 394 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL && 395 m->valid != 0) { 396 int base = (int)nsize & PAGE_MASK; 397 int size = PAGE_SIZE - base; 398 399 /* 400 * Clear out partial-page garbage in case 401 * the page has been mapped. 402 */ 403 pmap_zero_page_area(m, base, size); 404 405 /* 406 * Clear out partial-page dirty bits. This 407 * has the side effect of setting the valid 408 * bits, but that is ok. There are a bunch 409 * of places in the VM system where we expected 410 * m->dirty == VM_PAGE_BITS_ALL. The file EOF 411 * case is one of them. If the page is still 412 * partially dirty, make it fully dirty. 413 * 414 * note that we do not clear out the valid 415 * bits. This would prevent bogus_page 416 * replacement from working properly. 417 */ 418 vm_page_lock_queues(); 419 vm_page_set_validclean(m, base, size); 420 if (m->dirty != 0) 421 m->dirty = VM_PAGE_BITS_ALL; 422 vm_page_unlock_queues(); 423 } else if ((nsize & PAGE_MASK) && 424 __predict_false(object->cache != NULL)) { 425 vm_page_cache_free(object, OFF_TO_IDX(nsize), 426 nobjsize); 427 } 428 } 429 object->un_pager.vnp.vnp_size = nsize; 430 object->size = nobjsize; 431 VM_OBJECT_UNLOCK(object); 432} 433 434/* 435 * calculate the linear (byte) disk address of specified virtual 436 * file address 437 */ 438static int 439vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress, 440 int *run) 441{ 442 int bsize; 443 int err; 444 daddr_t vblock; 445 daddr_t voffset; 446 447 if (address < 0) 448 return -1; 449 450 if (vp->v_iflag & VI_DOOMED) 451 return -1; 452 453 bsize = vp->v_mount->mnt_stat.f_iosize; 454 vblock = address / bsize; 455 voffset = address % bsize; 456 457 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL); 458 if (err == 0) { 459 if (*rtaddress != -1) 460 *rtaddress += voffset / DEV_BSIZE; 461 if (run) { 462 *run += 1; 463 *run *= bsize/PAGE_SIZE; 464 *run -= voffset/PAGE_SIZE; 465 } 466 } 467 468 return (err); 469} 470 471/* 472 * small block filesystem vnode pager input 473 */ 474static int 475vnode_pager_input_smlfs(object, m) 476 vm_object_t object; 477 vm_page_t m; 478{ 479 int i; 480 struct vnode *vp; 481 struct bufobj *bo; 482 struct buf *bp; 483 struct sf_buf *sf; 484 daddr_t fileaddr; 485 vm_offset_t bsize; 486 int error = 0; 487 488 vp = object->handle; 489 if (vp->v_iflag & VI_DOOMED) 490 return VM_PAGER_BAD; 491 492 bsize = vp->v_mount->mnt_stat.f_iosize; 493 494 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL); 495 496 sf = sf_buf_alloc(m, 0); 497 498 for (i = 0; i < PAGE_SIZE / bsize; i++) { 499 vm_ooffset_t address; 500 501 if (vm_page_bits(i * bsize, bsize) & m->valid) 502 continue; 503 504 address = IDX_TO_OFF(m->pindex) + i * bsize; 505 if (address >= object->un_pager.vnp.vnp_size) { 506 fileaddr = -1; 507 } else { 508 error = vnode_pager_addr(vp, address, &fileaddr, NULL); 509 if (error) 510 break; 511 } 512 if (fileaddr != -1) { 513 bp = getpbuf(&vnode_pbuf_freecnt); 514 515 /* build a minimal buffer header */ 516 bp->b_iocmd = BIO_READ; 517 bp->b_iodone = bdone; 518 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 519 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 520 bp->b_rcred = crhold(curthread->td_ucred); 521 bp->b_wcred = crhold(curthread->td_ucred); 522 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize; 523 bp->b_blkno = fileaddr; 524 pbgetbo(bo, bp); 525 bp->b_bcount = bsize; 526 bp->b_bufsize = bsize; 527 bp->b_runningbufspace = bp->b_bufsize; 528 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 529 530 /* do the input */ 531 bp->b_iooffset = dbtob(bp->b_blkno); 532 bstrategy(bp); 533 534 bwait(bp, PVM, "vnsrd"); 535 536 if ((bp->b_ioflags & BIO_ERROR) != 0) 537 error = EIO; 538 539 /* 540 * free the buffer header back to the swap buffer pool 541 */ 542 pbrelbo(bp); 543 relpbuf(bp, &vnode_pbuf_freecnt); 544 if (error) 545 break; 546 547 VM_OBJECT_LOCK(object); 548 vm_page_lock_queues(); 549 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 550 vm_page_unlock_queues(); 551 VM_OBJECT_UNLOCK(object); 552 } else { 553 VM_OBJECT_LOCK(object); 554 vm_page_lock_queues(); 555 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 556 vm_page_unlock_queues(); 557 VM_OBJECT_UNLOCK(object); 558 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize); 559 } 560 } 561 sf_buf_free(sf); 562 vm_page_lock_queues(); 563 pmap_clear_modify(m); 564 vm_page_unlock_queues(); 565 if (error) { 566 return VM_PAGER_ERROR; 567 } 568 return VM_PAGER_OK; 569 570} 571 572 573/* 574 * old style vnode pager input routine 575 */ 576static int 577vnode_pager_input_old(object, m) 578 vm_object_t object; 579 vm_page_t m; 580{ 581 struct uio auio; 582 struct iovec aiov; 583 int error; 584 int size; 585 struct sf_buf *sf; 586 struct vnode *vp; 587 588 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 589 error = 0; 590 591 /* 592 * Return failure if beyond current EOF 593 */ 594 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 595 return VM_PAGER_BAD; 596 } else { 597 size = PAGE_SIZE; 598 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 599 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 600 vp = object->handle; 601 VM_OBJECT_UNLOCK(object); 602 603 /* 604 * Allocate a kernel virtual address and initialize so that 605 * we can use VOP_READ/WRITE routines. 606 */ 607 sf = sf_buf_alloc(m, 0); 608 609 aiov.iov_base = (caddr_t)sf_buf_kva(sf); 610 aiov.iov_len = size; 611 auio.uio_iov = &aiov; 612 auio.uio_iovcnt = 1; 613 auio.uio_offset = IDX_TO_OFF(m->pindex); 614 auio.uio_segflg = UIO_SYSSPACE; 615 auio.uio_rw = UIO_READ; 616 auio.uio_resid = size; 617 auio.uio_td = curthread; 618 619 error = VOP_READ(vp, &auio, 0, curthread->td_ucred); 620 if (!error) { 621 int count = size - auio.uio_resid; 622 623 if (count == 0) 624 error = EINVAL; 625 else if (count != PAGE_SIZE) 626 bzero((caddr_t)sf_buf_kva(sf) + count, 627 PAGE_SIZE - count); 628 } 629 sf_buf_free(sf); 630 631 VM_OBJECT_LOCK(object); 632 } 633 vm_page_lock_queues(); 634 pmap_clear_modify(m); 635 vm_page_undirty(m); 636 vm_page_unlock_queues(); 637 if (!error) 638 m->valid = VM_PAGE_BITS_ALL; 639 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 640} 641 642/* 643 * generic vnode pager input routine 644 */ 645 646/* 647 * Local media VFS's that do not implement their own VOP_GETPAGES 648 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages() 649 * to implement the previous behaviour. 650 * 651 * All other FS's should use the bypass to get to the local media 652 * backing vp's VOP_GETPAGES. 653 */ 654static int 655vnode_pager_getpages(object, m, count, reqpage) 656 vm_object_t object; 657 vm_page_t *m; 658 int count; 659 int reqpage; 660{ 661 int rtval; 662 struct vnode *vp; 663 int bytes = count * PAGE_SIZE; 664 int vfslocked; 665 666 vp = object->handle; 667 VM_OBJECT_UNLOCK(object); 668 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 669 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0); 670 KASSERT(rtval != EOPNOTSUPP, 671 ("vnode_pager: FS getpages not implemented\n")); 672 VFS_UNLOCK_GIANT(vfslocked); 673 VM_OBJECT_LOCK(object); 674 return rtval; 675} 676 677/* 678 * This is now called from local media FS's to operate against their 679 * own vnodes if they fail to implement VOP_GETPAGES. 680 */ 681int 682vnode_pager_generic_getpages(vp, m, bytecount, reqpage) 683 struct vnode *vp; 684 vm_page_t *m; 685 int bytecount; 686 int reqpage; 687{ 688 vm_object_t object; 689 vm_offset_t kva; 690 off_t foff, tfoff, nextoff; 691 int i, j, size, bsize, first; 692 daddr_t firstaddr, reqblock; 693 struct bufobj *bo; 694 int runpg; 695 int runend; 696 struct buf *bp; 697 int count; 698 int error; 699 700 object = vp->v_object; 701 count = bytecount / PAGE_SIZE; 702 703 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 704 ("vnode_pager_generic_getpages does not support devices")); 705 if (vp->v_iflag & VI_DOOMED) 706 return VM_PAGER_BAD; 707 708 bsize = vp->v_mount->mnt_stat.f_iosize; 709 710 /* get the UNDERLYING device for the file with VOP_BMAP() */ 711 712 /* 713 * originally, we did not check for an error return value -- assuming 714 * an fs always has a bmap entry point -- that assumption is wrong!!! 715 */ 716 foff = IDX_TO_OFF(m[reqpage]->pindex); 717 718 /* 719 * if we can't bmap, use old VOP code 720 */ 721 error = VOP_BMAP(vp, foff / bsize, &bo, &reqblock, NULL, NULL); 722 if (error == EOPNOTSUPP) { 723 VM_OBJECT_LOCK(object); 724 vm_page_lock_queues(); 725 for (i = 0; i < count; i++) 726 if (i != reqpage) 727 vm_page_free(m[i]); 728 vm_page_unlock_queues(); 729 PCPU_INC(cnt.v_vnodein); 730 PCPU_INC(cnt.v_vnodepgsin); 731 error = vnode_pager_input_old(object, m[reqpage]); 732 VM_OBJECT_UNLOCK(object); 733 return (error); 734 } else if (error != 0) { 735 VM_OBJECT_LOCK(object); 736 vm_page_lock_queues(); 737 for (i = 0; i < count; i++) 738 if (i != reqpage) 739 vm_page_free(m[i]); 740 vm_page_unlock_queues(); 741 VM_OBJECT_UNLOCK(object); 742 return (VM_PAGER_ERROR); 743 744 /* 745 * if the blocksize is smaller than a page size, then use 746 * special small filesystem code. NFS sometimes has a small 747 * blocksize, but it can handle large reads itself. 748 */ 749 } else if ((PAGE_SIZE / bsize) > 1 && 750 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) { 751 VM_OBJECT_LOCK(object); 752 vm_page_lock_queues(); 753 for (i = 0; i < count; i++) 754 if (i != reqpage) 755 vm_page_free(m[i]); 756 vm_page_unlock_queues(); 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 vm_page_lock_queues(); 771 for (i = 0; i < count; i++) 772 if (i != reqpage) 773 vm_page_free(m[i]); 774 vm_page_unlock_queues(); 775 VM_OBJECT_UNLOCK(object); 776 return VM_PAGER_OK; 777 } else if (reqblock == -1) { 778 pmap_zero_page(m[reqpage]); 779 vm_page_undirty(m[reqpage]); 780 m[reqpage]->valid = VM_PAGE_BITS_ALL; 781 vm_page_lock_queues(); 782 for (i = 0; i < count; i++) 783 if (i != reqpage) 784 vm_page_free(m[i]); 785 vm_page_unlock_queues(); 786 VM_OBJECT_UNLOCK(object); 787 return (VM_PAGER_OK); 788 } 789 m[reqpage]->valid = 0; 790 VM_OBJECT_UNLOCK(object); 791 792 /* 793 * here on direct device I/O 794 */ 795 firstaddr = -1; 796 797 /* 798 * calculate the run that includes the required page 799 */ 800 for (first = 0, i = 0; i < count; i = runend) { 801 if (vnode_pager_addr(vp, IDX_TO_OFF(m[i]->pindex), &firstaddr, 802 &runpg) != 0) { 803 VM_OBJECT_LOCK(object); 804 vm_page_lock_queues(); 805 for (; i < count; i++) 806 if (i != reqpage) 807 vm_page_free(m[i]); 808 vm_page_unlock_queues(); 809 VM_OBJECT_UNLOCK(object); 810 return (VM_PAGER_ERROR); 811 } 812 if (firstaddr == -1) { 813 VM_OBJECT_LOCK(object); 814 if (i == reqpage && foff < object->un_pager.vnp.vnp_size) { 815 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx", 816 (intmax_t)firstaddr, (uintmax_t)(foff >> 32), 817 (uintmax_t)foff, 818 (uintmax_t) 819 (object->un_pager.vnp.vnp_size >> 32), 820 (uintmax_t)object->un_pager.vnp.vnp_size); 821 } 822 vm_page_lock_queues(); 823 vm_page_free(m[i]); 824 vm_page_unlock_queues(); 825 VM_OBJECT_UNLOCK(object); 826 runend = i + 1; 827 first = runend; 828 continue; 829 } 830 runend = i + runpg; 831 if (runend <= reqpage) { 832 VM_OBJECT_LOCK(object); 833 vm_page_lock_queues(); 834 for (j = i; j < runend; j++) 835 vm_page_free(m[j]); 836 vm_page_unlock_queues(); 837 VM_OBJECT_UNLOCK(object); 838 } else { 839 if (runpg < (count - first)) { 840 VM_OBJECT_LOCK(object); 841 vm_page_lock_queues(); 842 for (i = first + runpg; i < count; i++) 843 vm_page_free(m[i]); 844 vm_page_unlock_queues(); 845 VM_OBJECT_UNLOCK(object); 846 count = first + runpg; 847 } 848 break; 849 } 850 first = runend; 851 } 852 853 /* 854 * the first and last page have been calculated now, move input pages 855 * to be zero based... 856 */ 857 if (first != 0) { 858 m += first; 859 count -= first; 860 reqpage -= first; 861 } 862 863 /* 864 * calculate the file virtual address for the transfer 865 */ 866 foff = IDX_TO_OFF(m[0]->pindex); 867 868 /* 869 * calculate the size of the transfer 870 */ 871 size = count * PAGE_SIZE; 872 KASSERT(count > 0, ("zero count")); 873 if ((foff + size) > object->un_pager.vnp.vnp_size) 874 size = object->un_pager.vnp.vnp_size - foff; 875 KASSERT(size > 0, ("zero size")); 876 877 /* 878 * round up physical size for real devices. 879 */ 880 if (1) { 881 int secmask = bo->bo_bsize - 1; 882 KASSERT(secmask < PAGE_SIZE && secmask > 0, 883 ("vnode_pager_generic_getpages: sector size %d too large", 884 secmask + 1)); 885 size = (size + secmask) & ~secmask; 886 } 887 888 bp = getpbuf(&vnode_pbuf_freecnt); 889 kva = (vm_offset_t) bp->b_data; 890 891 /* 892 * and map the pages to be read into the kva 893 */ 894 pmap_qenter(kva, m, count); 895 896 /* build a minimal buffer header */ 897 bp->b_iocmd = BIO_READ; 898 bp->b_iodone = bdone; 899 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 900 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 901 bp->b_rcred = crhold(curthread->td_ucred); 902 bp->b_wcred = crhold(curthread->td_ucred); 903 bp->b_blkno = firstaddr; 904 pbgetbo(bo, bp); 905 bp->b_bcount = size; 906 bp->b_bufsize = size; 907 bp->b_runningbufspace = bp->b_bufsize; 908 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 909 910 PCPU_INC(cnt.v_vnodein); 911 PCPU_ADD(cnt.v_vnodepgsin, count); 912 913 /* do the input */ 914 bp->b_iooffset = dbtob(bp->b_blkno); 915 bstrategy(bp); 916 917 bwait(bp, PVM, "vnread"); 918 919 if ((bp->b_ioflags & BIO_ERROR) != 0) 920 error = EIO; 921 922 if (!error) { 923 if (size != count * PAGE_SIZE) 924 bzero((caddr_t) kva + size, PAGE_SIZE * count - size); 925 } 926 pmap_qremove(kva, count); 927 928 /* 929 * free the buffer header back to the swap buffer pool 930 */ 931 pbrelbo(bp); 932 relpbuf(bp, &vnode_pbuf_freecnt); 933 934 VM_OBJECT_LOCK(object); 935 vm_page_lock_queues(); 936 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) { 937 vm_page_t mt; 938 939 nextoff = tfoff + PAGE_SIZE; 940 mt = m[i]; 941 942 if (nextoff <= object->un_pager.vnp.vnp_size) { 943 /* 944 * Read filled up entire page. 945 */ 946 mt->valid = VM_PAGE_BITS_ALL; 947 vm_page_undirty(mt); /* should be an assert? XXX */ 948 pmap_clear_modify(mt); 949 } else { 950 /* 951 * Read did not fill up entire page. Since this 952 * is getpages, the page may be mapped, so we have 953 * to zero the invalid portions of the page even 954 * though we aren't setting them valid. 955 * 956 * Currently we do not set the entire page valid, 957 * we just try to clear the piece that we couldn't 958 * read. 959 */ 960 vm_page_set_validclean(mt, 0, 961 object->un_pager.vnp.vnp_size - tfoff); 962 /* handled by vm_fault now */ 963 /* vm_page_zero_invalid(mt, FALSE); */ 964 } 965 966 if (i != reqpage) { 967 968 /* 969 * whether or not to leave the page activated is up in 970 * the air, but we should put the page on a page queue 971 * somewhere. (it already is in the object). Result: 972 * It appears that empirical results show that 973 * deactivating pages is best. 974 */ 975 976 /* 977 * just in case someone was asking for this page we 978 * now tell them that it is ok to use 979 */ 980 if (!error) { 981 if (mt->oflags & VPO_WANTED) 982 vm_page_activate(mt); 983 else 984 vm_page_deactivate(mt); 985 vm_page_wakeup(mt); 986 } else { 987 vm_page_free(mt); 988 } 989 } 990 } 991 vm_page_unlock_queues(); 992 VM_OBJECT_UNLOCK(object); 993 if (error) { 994 printf("vnode_pager_getpages: I/O read error\n"); 995 } 996 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 997} 998 999/* 1000 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 1001 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 1002 * vnode_pager_generic_putpages() to implement the previous behaviour. 1003 * 1004 * All other FS's should use the bypass to get to the local media 1005 * backing vp's VOP_PUTPAGES. 1006 */ 1007static void 1008vnode_pager_putpages(object, m, count, sync, rtvals) 1009 vm_object_t object; 1010 vm_page_t *m; 1011 int count; 1012 boolean_t sync; 1013 int *rtvals; 1014{ 1015 int rtval; 1016 struct vnode *vp; 1017 struct mount *mp; 1018 int bytes = count * PAGE_SIZE; 1019 1020 /* 1021 * Force synchronous operation if we are extremely low on memory 1022 * to prevent a low-memory deadlock. VOP operations often need to 1023 * allocate more memory to initiate the I/O ( i.e. do a BMAP 1024 * operation ). The swapper handles the case by limiting the amount 1025 * of asynchronous I/O, but that sort of solution doesn't scale well 1026 * for the vnode pager without a lot of work. 1027 * 1028 * Also, the backing vnode's iodone routine may not wake the pageout 1029 * daemon up. This should be probably be addressed XXX. 1030 */ 1031 1032 if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min) 1033 sync |= OBJPC_SYNC; 1034 1035 /* 1036 * Call device-specific putpages function 1037 */ 1038 vp = object->handle; 1039 VM_OBJECT_UNLOCK(object); 1040 if (vp->v_type != VREG) 1041 mp = NULL; 1042 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 1043 KASSERT(rtval != EOPNOTSUPP, 1044 ("vnode_pager: stale FS putpages\n")); 1045 VM_OBJECT_LOCK(object); 1046} 1047 1048 1049/* 1050 * This is now called from local media FS's to operate against their 1051 * own vnodes if they fail to implement VOP_PUTPAGES. 1052 * 1053 * This is typically called indirectly via the pageout daemon and 1054 * clustering has already typically occured, so in general we ask the 1055 * underlying filesystem to write the data out asynchronously rather 1056 * then delayed. 1057 */ 1058int 1059vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals) 1060 struct vnode *vp; 1061 vm_page_t *m; 1062 int bytecount; 1063 int flags; 1064 int *rtvals; 1065{ 1066 int i; 1067 vm_object_t object; 1068 int count; 1069 1070 int maxsize, ncount; 1071 vm_ooffset_t poffset; 1072 struct uio auio; 1073 struct iovec aiov; 1074 int error; 1075 int ioflags; 1076 int ppscheck = 0; 1077 static struct timeval lastfail; 1078 static int curfail; 1079 1080 object = vp->v_object; 1081 count = bytecount / PAGE_SIZE; 1082 1083 for (i = 0; i < count; i++) 1084 rtvals[i] = VM_PAGER_AGAIN; 1085 1086 if ((int64_t)m[0]->pindex < 0) { 1087 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n", 1088 (long)m[0]->pindex, (u_long)m[0]->dirty); 1089 rtvals[0] = VM_PAGER_BAD; 1090 return VM_PAGER_BAD; 1091 } 1092 1093 maxsize = count * PAGE_SIZE; 1094 ncount = count; 1095 1096 poffset = IDX_TO_OFF(m[0]->pindex); 1097 1098 /* 1099 * If the page-aligned write is larger then the actual file we 1100 * have to invalidate pages occuring beyond the file EOF. However, 1101 * there is an edge case where a file may not be page-aligned where 1102 * the last page is partially invalid. In this case the filesystem 1103 * may not properly clear the dirty bits for the entire page (which 1104 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 1105 * With the page locked we are free to fix-up the dirty bits here. 1106 * 1107 * We do not under any circumstances truncate the valid bits, as 1108 * this will screw up bogus page replacement. 1109 */ 1110 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 1111 if (object->un_pager.vnp.vnp_size > poffset) { 1112 int pgoff; 1113 1114 maxsize = object->un_pager.vnp.vnp_size - poffset; 1115 ncount = btoc(maxsize); 1116 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 1117 vm_page_lock_queues(); 1118 vm_page_clear_dirty(m[ncount - 1], pgoff, 1119 PAGE_SIZE - pgoff); 1120 vm_page_unlock_queues(); 1121 } 1122 } else { 1123 maxsize = 0; 1124 ncount = 0; 1125 } 1126 if (ncount < count) { 1127 for (i = ncount; i < count; i++) { 1128 rtvals[i] = VM_PAGER_BAD; 1129 } 1130 } 1131 } 1132 1133 /* 1134 * pageouts are already clustered, use IO_ASYNC t o force a bawrite() 1135 * rather then a bdwrite() to prevent paging I/O from saturating 1136 * the buffer cache. Dummy-up the sequential heuristic to cause 1137 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 1138 * the system decides how to cluster. 1139 */ 1140 ioflags = IO_VMIO; 1141 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 1142 ioflags |= IO_SYNC; 1143 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 1144 ioflags |= IO_ASYNC; 1145 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 1146 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 1147 1148 aiov.iov_base = (caddr_t) 0; 1149 aiov.iov_len = maxsize; 1150 auio.uio_iov = &aiov; 1151 auio.uio_iovcnt = 1; 1152 auio.uio_offset = poffset; 1153 auio.uio_segflg = UIO_NOCOPY; 1154 auio.uio_rw = UIO_WRITE; 1155 auio.uio_resid = maxsize; 1156 auio.uio_td = (struct thread *) 0; 1157 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred); 1158 PCPU_INC(cnt.v_vnodeout); 1159 PCPU_ADD(cnt.v_vnodepgsout, ncount); 1160 1161 if (error) { 1162 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1))) 1163 printf("vnode_pager_putpages: I/O error %d\n", error); 1164 } 1165 if (auio.uio_resid) { 1166 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1)) 1167 printf("vnode_pager_putpages: residual I/O %d at %lu\n", 1168 auio.uio_resid, (u_long)m[0]->pindex); 1169 } 1170 for (i = 0; i < ncount; i++) { 1171 rtvals[i] = VM_PAGER_OK; 1172 } 1173 return rtvals[0]; 1174} 1175