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 * $FreeBSD: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $ 42 * $DragonFly: src/sys/vm/vnode_pager.c,v 1.33 2007/05/06 19:23:36 dillon Exp $ 43 */ 44 45 /* 46 * Page to/from files (vnodes). 47 */ 48 49 /* 50 * TODO: 51 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will 52 * greatly re-simplify the vnode_pager. 53 */ 54 55 #include <sys/param.h> 56 #include <sys/systm.h> 57 #include <sys/kernel.h> 58 #include <sys/proc.h> 59 #include <sys/vnode.h> 60 #include <sys/mount.h> 61 #include <sys/buf.h> 62 #include <sys/vmmeter.h> 63 #include <sys/conf.h> 64 #include <sys/sfbuf.h> 65 #include <sys/thread2.h> 66 67 #include <vm/vm.h> 68 #include <vm/vm_object.h> 69 #include <vm/vm_page.h> 70 #include <vm/vm_pager.h> 71 #include <vm/vm_map.h> 72 #include <vm/vnode_pager.h> 73 #include <vm/vm_extern.h> 74 75 static off_t vnode_pager_addr (struct vnode *vp, off_t loffset, int *run); 76 static void vnode_pager_iodone (struct bio *bio); 77 static int vnode_pager_input_smlfs (vm_object_t object, vm_page_t m); 78 static int vnode_pager_input_old (vm_object_t object, vm_page_t m); 79 static void vnode_pager_dealloc (vm_object_t); 80 static int vnode_pager_getpages (vm_object_t, vm_page_t *, int, int); 81 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *); 82 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t, int *, int *); 83 84 struct pagerops vnodepagerops = { 85 NULL, 86 vnode_pager_alloc, 87 vnode_pager_dealloc, 88 vnode_pager_getpages, 89 vnode_pager_putpages, 90 vnode_pager_haspage, 91 NULL 92 }; 93 94 int vnode_pbuf_freecnt = -1; /* start out unlimited */ 95 96 /* 97 * Allocate (or lookup) pager for a vnode. 98 * Handle is a vnode pointer. 99 */ 100 vm_object_t 101 vnode_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset) 102 { 103 vm_object_t object; 104 struct vnode *vp; 105 106 /* 107 * Pageout to vnode, no can do yet. 108 */ 109 if (handle == NULL) 110 return (NULL); 111 112 /* 113 * XXX hack - This initialization should be put somewhere else. 114 */ 115 if (vnode_pbuf_freecnt < 0) { 116 vnode_pbuf_freecnt = nswbuf / 2 + 1; 117 } 118 119 vp = (struct vnode *) handle; 120 121 /* 122 * Prevent race condition when allocating the object. This 123 * can happen with NFS vnodes since the nfsnode isn't locked. 124 */ 125 while (vp->v_flag & VOLOCK) { 126 vp->v_flag |= VOWANT; 127 tsleep(vp, 0, "vnpobj", 0); 128 } 129 vp->v_flag |= VOLOCK; 130 131 /* 132 * If the object is being terminated, wait for it to 133 * go away. 134 */ 135 while (((object = vp->v_object) != NULL) && 136 (object->flags & OBJ_DEAD)) { 137 tsleep(object, 0, "vadead", 0); 138 } 139 140 if (vp->v_sysref.refcnt <= 0) 141 panic("vnode_pager_alloc: no vnode reference"); 142 143 if (object == NULL) { 144 /* 145 * And an object of the appropriate size 146 */ 147 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); 148 object->flags = 0; 149 object->handle = handle; 150 vp->v_object = object; 151 vp->v_filesize = size; 152 } else { 153 object->ref_count++; 154 if (vp->v_filesize != size) 155 kprintf("vnode_pager_alloc: Warning, filesize mismatch %lld/%lld\n", vp->v_filesize, size); 156 } 157 vref(vp); 158 159 vp->v_flag &= ~VOLOCK; 160 if (vp->v_flag & VOWANT) { 161 vp->v_flag &= ~VOWANT; 162 wakeup(vp); 163 } 164 return (object); 165 } 166 167 static void 168 vnode_pager_dealloc(vm_object_t object) 169 { 170 struct vnode *vp = object->handle; 171 172 if (vp == NULL) 173 panic("vnode_pager_dealloc: pager already dealloced"); 174 175 vm_object_pip_wait(object, "vnpdea"); 176 177 object->handle = NULL; 178 object->type = OBJT_DEAD; 179 vp->v_object = NULL; 180 vp->v_filesize = NOOFFSET; 181 vp->v_flag &= ~(VTEXT | VOBJBUF); 182 } 183 184 /* 185 * Return whether the vnode pager has the requested page. Return the 186 * number of disk-contiguous pages before and after the requested page, 187 * not including the requested page. 188 */ 189 static boolean_t 190 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 191 int *after) 192 { 193 struct vnode *vp = object->handle; 194 off_t loffset; 195 off_t doffset; 196 int voff; 197 int bsize; 198 int error; 199 200 /* 201 * If no vp or vp is doomed or marked transparent to VM, we do not 202 * have the page. 203 */ 204 if ((vp == NULL) || (vp->v_flag & VRECLAIMED)) 205 return FALSE; 206 207 /* 208 * If filesystem no longer mounted or offset beyond end of file we do 209 * not have the page. 210 */ 211 loffset = IDX_TO_OFF(pindex); 212 213 if (vp->v_mount == NULL || loffset >= vp->v_filesize) 214 return FALSE; 215 216 bsize = vp->v_mount->mnt_stat.f_iosize; 217 voff = loffset % bsize; 218 219 error = VOP_BMAP(vp, loffset - voff, NULL, &doffset, after, before); 220 if (error) 221 return TRUE; 222 if (doffset == NOOFFSET) 223 return FALSE; 224 225 if (before) { 226 *before = (*before + voff) >> PAGE_SHIFT; 227 } 228 if (after) { 229 *after -= voff; 230 if (loffset + *after > vp->v_filesize) 231 *after = vp->v_filesize - loffset; 232 *after >>= PAGE_SHIFT; 233 if (*after < 0) 234 *after = 0; 235 } 236 return TRUE; 237 } 238 239 /* 240 * Lets the VM system know about a change in size for a file. 241 * We adjust our own internal size and flush any cached pages in 242 * the associated object that are affected by the size change. 243 * 244 * NOTE: This routine may be invoked as a result of a pager put 245 * operation (possibly at object termination time), so we must be careful. 246 * 247 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that 248 * we do not blow up on the case. nsize will always be >= 0, however. 249 */ 250 void 251 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize) 252 { 253 vm_pindex_t nobjsize; 254 vm_pindex_t oobjsize; 255 vm_object_t object = vp->v_object; 256 257 if (object == NULL) 258 return; 259 260 /* 261 * Hasn't changed size 262 */ 263 if (nsize == vp->v_filesize) 264 return; 265 266 /* 267 * Has changed size. Adjust the VM object's size and v_filesize 268 * before we start scanning pages to prevent new pages from being 269 * allocated during the scan. 270 */ 271 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 272 oobjsize = object->size; 273 object->size = nobjsize; 274 275 /* 276 * File has shrunk. Toss any cached pages beyond the new EOF. 277 */ 278 if (nsize < vp->v_filesize) { 279 vp->v_filesize = nsize; 280 if (nobjsize < oobjsize) { 281 vm_object_page_remove(object, nobjsize, oobjsize, 282 FALSE); 283 } 284 /* 285 * This gets rid of garbage at the end of a page that is now 286 * only partially backed by the vnode. Since we are setting 287 * the entire page valid & clean after we are done we have 288 * to be sure that the portion of the page within the file 289 * bounds is already valid. If it isn't then making it 290 * valid would create a corrupt block. 291 */ 292 if (nsize & PAGE_MASK) { 293 vm_offset_t kva; 294 vm_page_t m; 295 296 m = vm_page_lookup(object, OFF_TO_IDX(nsize)); 297 if (m && m->valid) { 298 int base = (int)nsize & PAGE_MASK; 299 int size = PAGE_SIZE - base; 300 struct sf_buf *sf; 301 302 /* 303 * Clear out partial-page garbage in case 304 * the page has been mapped. 305 */ 306 sf = sf_buf_alloc(m, SFB_CPUPRIVATE); 307 kva = sf_buf_kva(sf); 308 bzero((caddr_t)kva + base, size); 309 sf_buf_free(sf); 310 311 /* 312 * XXX work around SMP data integrity race 313 * by unmapping the page from user processes. 314 * The garbage we just cleared may be mapped 315 * to a user process running on another cpu 316 * and this code is not running through normal 317 * I/O channels which handle SMP issues for 318 * us, so unmap page to synchronize all cpus. 319 * 320 * XXX should vm_pager_unmap_page() have 321 * dealt with this? 322 */ 323 vm_page_protect(m, VM_PROT_NONE); 324 325 /* 326 * Clear out partial-page dirty bits. This 327 * has the side effect of setting the valid 328 * bits, but that is ok. There are a bunch 329 * of places in the VM system where we expected 330 * m->dirty == VM_PAGE_BITS_ALL. The file EOF 331 * case is one of them. If the page is still 332 * partially dirty, make it fully dirty. 333 * 334 * note that we do not clear out the valid 335 * bits. This would prevent bogus_page 336 * replacement from working properly. 337 */ 338 vm_page_set_validclean(m, base, size); 339 if (m->dirty != 0) 340 m->dirty = VM_PAGE_BITS_ALL; 341 } 342 } 343 } else { 344 vp->v_filesize = nsize; 345 } 346 } 347 348 void 349 vnode_pager_freepage(vm_page_t m) 350 { 351 vm_page_free(m); 352 } 353 354 /* 355 * calculate the disk byte address of specified logical byte offset. The 356 * logical offset will be block-aligned. Return the number of contiguous 357 * pages that may be read from the underlying block device in *run. If 358 * *run is non-NULL, it will be set to a value of at least 1. 359 */ 360 static off_t 361 vnode_pager_addr(struct vnode *vp, off_t loffset, int *run) 362 { 363 struct vnode *rtvp; 364 off_t doffset; 365 int bsize; 366 int error; 367 int voff; 368 369 if (loffset < 0) 370 return -1; 371 372 if (vp->v_mount == NULL) 373 return -1; 374 375 /* 376 * Align loffset to a block boundary for the BMAP, then adjust the 377 * returned disk address appropriately. 378 */ 379 bsize = vp->v_mount->mnt_stat.f_iosize; 380 voff = loffset % bsize; 381 382 /* 383 * Map the block, adjust the disk offset so it represents the 384 * passed loffset rather then the block containing loffset. 385 */ 386 error = VOP_BMAP(vp, loffset - voff, &rtvp, &doffset, run, NULL); 387 if (error || doffset == NOOFFSET) { 388 doffset = NOOFFSET; 389 } else { 390 doffset += voff; 391 392 /* 393 * When calculating *run, which is the number of pages 394 * worth of data which can be read linearly from disk, 395 * the minimum return value is 1 page. 396 */ 397 if (run) { 398 *run = (*run - voff) >> PAGE_SHIFT; 399 if (*run < 1) 400 *run = 1; 401 } 402 403 } 404 return (doffset); 405 } 406 407 /* 408 * interrupt routine for I/O completion 409 */ 410 static void 411 vnode_pager_iodone(struct bio *bio) 412 { 413 struct buf *bp = bio->bio_buf; 414 415 bp->b_cmd = BUF_CMD_DONE; 416 wakeup(bp); 417 } 418 419 /* 420 * small block file system vnode pager input 421 */ 422 static int 423 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m) 424 { 425 int i; 426 struct vnode *dp, *vp; 427 struct buf *bp; 428 vm_offset_t kva; 429 struct sf_buf *sf; 430 off_t doffset; 431 vm_offset_t bsize; 432 int error = 0; 433 434 vp = object->handle; 435 if (vp->v_mount == NULL) 436 return VM_PAGER_BAD; 437 438 bsize = vp->v_mount->mnt_stat.f_iosize; 439 440 441 VOP_BMAP(vp, (off_t)0, &dp, NULL, NULL, NULL); 442 443 sf = sf_buf_alloc(m, 0); 444 kva = sf_buf_kva(sf); 445 446 for (i = 0; i < PAGE_SIZE / bsize; i++) { 447 off_t loffset; 448 449 if (vm_page_bits(i * bsize, bsize) & m->valid) 450 continue; 451 452 loffset = IDX_TO_OFF(m->pindex) + i * bsize; 453 if (loffset >= vp->v_filesize) { 454 doffset = NOOFFSET; 455 } else { 456 doffset = vnode_pager_addr(vp, loffset, NULL); 457 } 458 if (doffset != NOOFFSET) { 459 bp = getpbuf(&vnode_pbuf_freecnt); 460 461 /* build a minimal buffer header */ 462 bp->b_data = (caddr_t) kva + i * bsize; 463 bp->b_bio1.bio_done = vnode_pager_iodone; 464 bp->b_bio1.bio_offset = doffset; 465 bp->b_bcount = bsize; 466 bp->b_runningbufspace = bsize; 467 runningbufspace += bp->b_runningbufspace; 468 bp->b_cmd = BUF_CMD_READ; 469 470 /* do the input */ 471 vn_strategy(dp, &bp->b_bio1); 472 473 /* we definitely need to be at splvm here */ 474 475 crit_enter(); 476 while (bp->b_cmd != BUF_CMD_DONE) 477 tsleep(bp, 0, "vnsrd", 0); 478 crit_exit(); 479 if ((bp->b_flags & B_ERROR) != 0) 480 error = EIO; 481 482 /* 483 * free the buffer header back to the swap buffer pool 484 */ 485 relpbuf(bp, &vnode_pbuf_freecnt); 486 if (error) 487 break; 488 489 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 490 } else { 491 vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize); 492 bzero((caddr_t) kva + i * bsize, bsize); 493 } 494 } 495 sf_buf_free(sf); 496 pmap_clear_modify(m); 497 vm_page_flag_clear(m, PG_ZERO); 498 if (error) { 499 return VM_PAGER_ERROR; 500 } 501 return VM_PAGER_OK; 502 503 } 504 505 506 /* 507 * old style vnode pager output routine 508 */ 509 static int 510 vnode_pager_input_old(vm_object_t object, vm_page_t m) 511 { 512 struct uio auio; 513 struct iovec aiov; 514 int error; 515 int size; 516 vm_offset_t kva; 517 struct sf_buf *sf; 518 struct vnode *vp; 519 520 error = 0; 521 vp = object->handle; 522 523 /* 524 * Return failure if beyond current EOF 525 */ 526 if (IDX_TO_OFF(m->pindex) >= vp->v_filesize) { 527 return VM_PAGER_BAD; 528 } else { 529 size = PAGE_SIZE; 530 if (IDX_TO_OFF(m->pindex) + size > vp->v_filesize) 531 size = vp->v_filesize - IDX_TO_OFF(m->pindex); 532 533 /* 534 * Allocate a kernel virtual address and initialize so that 535 * we can use VOP_READ/WRITE routines. 536 */ 537 sf = sf_buf_alloc(m, 0); 538 kva = sf_buf_kva(sf); 539 540 aiov.iov_base = (caddr_t) kva; 541 aiov.iov_len = size; 542 auio.uio_iov = &aiov; 543 auio.uio_iovcnt = 1; 544 auio.uio_offset = IDX_TO_OFF(m->pindex); 545 auio.uio_segflg = UIO_SYSSPACE; 546 auio.uio_rw = UIO_READ; 547 auio.uio_resid = size; 548 auio.uio_td = curthread; 549 550 error = VOP_READ(((struct vnode *)object->handle), 551 &auio, 0, proc0.p_ucred); 552 if (!error) { 553 int count = size - auio.uio_resid; 554 555 if (count == 0) 556 error = EINVAL; 557 else if (count != PAGE_SIZE) 558 bzero((caddr_t) kva + count, PAGE_SIZE - count); 559 } 560 sf_buf_free(sf); 561 } 562 pmap_clear_modify(m); 563 vm_page_undirty(m); 564 vm_page_flag_clear(m, PG_ZERO); 565 if (!error) 566 m->valid = VM_PAGE_BITS_ALL; 567 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 568 } 569 570 /* 571 * generic vnode pager input routine 572 */ 573 574 /* 575 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 576 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to 577 * vnode_pager_generic_getpages() to implement the previous behaviour. 578 * 579 * All other FS's should use the bypass to get to the local media 580 * backing vp's VOP_GETPAGES. 581 */ 582 static int 583 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage) 584 { 585 int rtval; 586 struct vnode *vp; 587 int bytes = count * PAGE_SIZE; 588 589 vp = object->handle; 590 /* 591 * XXX temporary diagnostic message to help track stale FS code, 592 * Returning EOPNOTSUPP from here may make things unhappy. 593 */ 594 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0); 595 if (rtval == EOPNOTSUPP) { 596 kprintf("vnode_pager: *** WARNING *** stale FS getpages\n"); 597 rtval = vnode_pager_generic_getpages( vp, m, bytes, reqpage); 598 } 599 return rtval; 600 } 601 602 603 /* 604 * This is now called from local media FS's to operate against their 605 * own vnodes if they fail to implement VOP_GETPAGES. 606 */ 607 int 608 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int bytecount, 609 int reqpage) 610 { 611 vm_object_t object; 612 vm_offset_t kva; 613 off_t foff, tfoff, nextoff; 614 int i, size, bsize, first; 615 off_t firstaddr; 616 struct vnode *dp; 617 int runpg; 618 int runend; 619 struct buf *bp; 620 int count; 621 int error = 0; 622 623 object = vp->v_object; 624 count = bytecount / PAGE_SIZE; 625 626 if (vp->v_mount == NULL) 627 return VM_PAGER_BAD; 628 629 bsize = vp->v_mount->mnt_stat.f_iosize; 630 631 /* get the UNDERLYING device for the file with VOP_BMAP() */ 632 633 /* 634 * originally, we did not check for an error return value -- assuming 635 * an fs always has a bmap entry point -- that assumption is wrong!!! 636 */ 637 foff = IDX_TO_OFF(m[reqpage]->pindex); 638 639 /* 640 * if we can't bmap, use old VOP code 641 */ 642 if (VOP_BMAP(vp, (off_t)0, &dp, NULL, NULL, NULL)) { 643 for (i = 0; i < count; i++) { 644 if (i != reqpage) { 645 vnode_pager_freepage(m[i]); 646 } 647 } 648 mycpu->gd_cnt.v_vnodein++; 649 mycpu->gd_cnt.v_vnodepgsin++; 650 return vnode_pager_input_old(object, m[reqpage]); 651 652 /* 653 * if the blocksize is smaller than a page size, then use 654 * special small filesystem code. NFS sometimes has a small 655 * blocksize, but it can handle large reads itself. 656 */ 657 } else if ((PAGE_SIZE / bsize) > 1 && 658 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) { 659 for (i = 0; i < count; i++) { 660 if (i != reqpage) { 661 vnode_pager_freepage(m[i]); 662 } 663 } 664 mycpu->gd_cnt.v_vnodein++; 665 mycpu->gd_cnt.v_vnodepgsin++; 666 return vnode_pager_input_smlfs(object, m[reqpage]); 667 } 668 669 /* 670 * If we have a completely valid page available to us, we can 671 * clean up and return. Otherwise we have to re-read the 672 * media. 673 * 674 * Note that this does not work with NFS, so NFS has its own 675 * getpages routine. The problem is that NFS can have partially 676 * valid pages associated with the buffer cache due to the piecemeal 677 * write support. If we were to fall through and re-read the media 678 * as we do here, dirty data could be lost. 679 */ 680 681 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) { 682 for (i = 0; i < count; i++) { 683 if (i != reqpage) 684 vnode_pager_freepage(m[i]); 685 } 686 return VM_PAGER_OK; 687 } 688 m[reqpage]->valid = 0; 689 690 /* 691 * here on direct device I/O 692 */ 693 694 firstaddr = -1; 695 /* 696 * calculate the run that includes the required page 697 */ 698 for(first = 0, i = 0; i < count; i = runend) { 699 firstaddr = vnode_pager_addr(vp, IDX_TO_OFF(m[i]->pindex), 700 &runpg); 701 if (firstaddr == -1) { 702 if (i == reqpage && foff < vp->v_filesize) { 703 /* XXX no %qd in kernel. */ 704 panic("vnode_pager_getpages: unexpected missing page: firstaddr: %012llx, foff: 0x%012llx, v_filesize: 0x%012llx", 705 firstaddr, foff, vp->v_filesize); 706 } 707 vnode_pager_freepage(m[i]); 708 runend = i + 1; 709 first = runend; 710 continue; 711 } 712 runend = i + runpg; 713 if (runend <= reqpage) { 714 int j; 715 for (j = i; j < runend; j++) { 716 vnode_pager_freepage(m[j]); 717 } 718 } else { 719 if (runpg < (count - first)) { 720 for (i = first + runpg; i < count; i++) 721 vnode_pager_freepage(m[i]); 722 count = first + runpg; 723 } 724 break; 725 } 726 first = runend; 727 } 728 729 /* 730 * the first and last page have been calculated now, move input pages 731 * to be zero based... 732 */ 733 if (first != 0) { 734 for (i = first; i < count; i++) { 735 m[i - first] = m[i]; 736 } 737 count -= first; 738 reqpage -= first; 739 } 740 741 /* 742 * calculate the file virtual address for the transfer 743 */ 744 foff = IDX_TO_OFF(m[0]->pindex); 745 746 /* 747 * calculate the size of the transfer 748 */ 749 size = count * PAGE_SIZE; 750 if ((foff + size) > vp->v_filesize) 751 size = vp->v_filesize - foff; 752 753 /* 754 * round up physical size for real devices. 755 */ 756 if (dp->v_type == VBLK || dp->v_type == VCHR) { 757 int secmask = dp->v_rdev->si_bsize_phys - 1; 758 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1)); 759 size = (size + secmask) & ~secmask; 760 } 761 762 bp = getpbuf(&vnode_pbuf_freecnt); 763 kva = (vm_offset_t) bp->b_data; 764 765 /* 766 * and map the pages to be read into the kva 767 */ 768 pmap_qenter(kva, m, count); 769 770 /* build a minimal buffer header */ 771 bp->b_bio1.bio_done = vnode_pager_iodone; 772 bp->b_bio1.bio_offset = firstaddr; 773 bp->b_bcount = size; 774 bp->b_runningbufspace = size; 775 runningbufspace += bp->b_runningbufspace; 776 bp->b_cmd = BUF_CMD_READ; 777 778 mycpu->gd_cnt.v_vnodein++; 779 mycpu->gd_cnt.v_vnodepgsin += count; 780 781 /* do the input */ 782 vn_strategy(dp, &bp->b_bio1); 783 784 crit_enter(); 785 /* we definitely need to be at splvm here */ 786 787 while (bp->b_cmd != BUF_CMD_DONE) 788 tsleep(bp, 0, "vnread", 0); 789 crit_exit(); 790 if ((bp->b_flags & B_ERROR) != 0) 791 error = EIO; 792 793 if (!error) { 794 if (size != count * PAGE_SIZE) 795 bzero((caddr_t) kva + size, PAGE_SIZE * count - size); 796 } 797 pmap_qremove(kva, count); 798 799 /* 800 * free the buffer header back to the swap buffer pool 801 */ 802 relpbuf(bp, &vnode_pbuf_freecnt); 803 804 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) { 805 vm_page_t mt; 806 807 nextoff = tfoff + PAGE_SIZE; 808 mt = m[i]; 809 810 if (nextoff <= vp->v_filesize) { 811 /* 812 * Read filled up entire page. 813 */ 814 mt->valid = VM_PAGE_BITS_ALL; 815 vm_page_undirty(mt); /* should be an assert? XXX */ 816 pmap_clear_modify(mt); 817 } else { 818 /* 819 * Read did not fill up entire page. Since this 820 * is getpages, the page may be mapped, so we have 821 * to zero the invalid portions of the page even 822 * though we aren't setting them valid. 823 * 824 * Currently we do not set the entire page valid, 825 * we just try to clear the piece that we couldn't 826 * read. 827 */ 828 vm_page_set_validclean(mt, 0, vp->v_filesize - tfoff); 829 /* handled by vm_fault now */ 830 /* vm_page_zero_invalid(mt, FALSE); */ 831 } 832 833 vm_page_flag_clear(mt, PG_ZERO); 834 if (i != reqpage) { 835 836 /* 837 * whether or not to leave the page activated is up in 838 * the air, but we should put the page on a page queue 839 * somewhere. (it already is in the object). Result: 840 * It appears that empirical results show that 841 * deactivating pages is best. 842 */ 843 844 /* 845 * just in case someone was asking for this page we 846 * now tell them that it is ok to use 847 */ 848 if (!error) { 849 if (mt->flags & PG_WANTED) 850 vm_page_activate(mt); 851 else 852 vm_page_deactivate(mt); 853 vm_page_wakeup(mt); 854 } else { 855 vnode_pager_freepage(mt); 856 } 857 } 858 } 859 if (error) { 860 kprintf("vnode_pager_getpages: I/O read error\n"); 861 } 862 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 863 } 864 865 /* 866 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 867 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 868 * vnode_pager_generic_putpages() to implement the previous behaviour. 869 * 870 * All other FS's should use the bypass to get to the local media 871 * backing vp's VOP_PUTPAGES. 872 */ 873 static void 874 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count, 875 boolean_t sync, int *rtvals) 876 { 877 int rtval; 878 struct vnode *vp; 879 int bytes = count * PAGE_SIZE; 880 881 /* 882 * Force synchronous operation if we are extremely low on memory 883 * to prevent a low-memory deadlock. VOP operations often need to 884 * allocate more memory to initiate the I/O ( i.e. do a BMAP 885 * operation ). The swapper handles the case by limiting the amount 886 * of asynchronous I/O, but that sort of solution doesn't scale well 887 * for the vnode pager without a lot of work. 888 * 889 * Also, the backing vnode's iodone routine may not wake the pageout 890 * daemon up. This should be probably be addressed XXX. 891 */ 892 893 if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min) 894 sync |= OBJPC_SYNC; 895 896 /* 897 * Call device-specific putpages function 898 */ 899 900 vp = object->handle; 901 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 902 if (rtval == EOPNOTSUPP) { 903 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n"); 904 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals); 905 } 906 } 907 908 909 /* 910 * This is now called from local media FS's to operate against their 911 * own vnodes if they fail to implement VOP_PUTPAGES. 912 * 913 * This is typically called indirectly via the pageout daemon and 914 * clustering has already typically occured, so in general we ask the 915 * underlying filesystem to write the data out asynchronously rather 916 * then delayed. 917 */ 918 int 919 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount, 920 int flags, int *rtvals) 921 { 922 int i; 923 vm_object_t object; 924 int count; 925 926 int maxsize, ncount; 927 vm_ooffset_t poffset; 928 struct uio auio; 929 struct iovec aiov; 930 int error; 931 int ioflags; 932 933 object = vp->v_object; 934 count = bytecount / PAGE_SIZE; 935 936 for (i = 0; i < count; i++) 937 rtvals[i] = VM_PAGER_AGAIN; 938 939 if ((int) m[0]->pindex < 0) { 940 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n", 941 (long)m[0]->pindex, m[0]->dirty); 942 rtvals[0] = VM_PAGER_BAD; 943 return VM_PAGER_BAD; 944 } 945 946 maxsize = count * PAGE_SIZE; 947 ncount = count; 948 949 poffset = IDX_TO_OFF(m[0]->pindex); 950 951 /* 952 * If the page-aligned write is larger then the actual file we 953 * have to invalidate pages occuring beyond the file EOF. However, 954 * there is an edge case where a file may not be page-aligned where 955 * the last page is partially invalid. In this case the filesystem 956 * may not properly clear the dirty bits for the entire page (which 957 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 958 * With the page locked we are free to fix-up the dirty bits here. 959 * 960 * We do not under any circumstances truncate the valid bits, as 961 * this will screw up bogus page replacement. 962 */ 963 if (maxsize + poffset > vp->v_filesize) { 964 if (vp->v_filesize > poffset) { 965 int pgoff; 966 967 maxsize = vp->v_filesize - poffset; 968 ncount = btoc(maxsize); 969 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 970 vm_page_clear_dirty(m[ncount - 1], pgoff, 971 PAGE_SIZE - pgoff); 972 } 973 } else { 974 maxsize = 0; 975 ncount = 0; 976 } 977 if (ncount < count) { 978 for (i = ncount; i < count; i++) { 979 rtvals[i] = VM_PAGER_BAD; 980 } 981 } 982 } 983 984 /* 985 * pageouts are already clustered, use IO_ASYNC to force a bawrite() 986 * rather then a bdwrite() to prevent paging I/O from saturating 987 * the buffer cache. Dummy-up the sequential heuristic to cause 988 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 989 * the system decides how to cluster. 990 */ 991 ioflags = IO_VMIO; 992 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 993 ioflags |= IO_SYNC; 994 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 995 ioflags |= IO_ASYNC; 996 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 997 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 998 999 aiov.iov_base = (caddr_t) 0; 1000 aiov.iov_len = maxsize; 1001 auio.uio_iov = &aiov; 1002 auio.uio_iovcnt = 1; 1003 auio.uio_offset = poffset; 1004 auio.uio_segflg = UIO_NOCOPY; 1005 auio.uio_rw = UIO_WRITE; 1006 auio.uio_resid = maxsize; 1007 auio.uio_td = NULL; 1008 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred); 1009 mycpu->gd_cnt.v_vnodeout++; 1010 mycpu->gd_cnt.v_vnodepgsout += ncount; 1011 1012 if (error) { 1013 kprintf("vnode_pager_putpages: I/O error %d\n", error); 1014 } 1015 if (auio.uio_resid) { 1016 kprintf("vnode_pager_putpages: residual I/O %d at %lu\n", 1017 auio.uio_resid, (u_long)m[0]->pindex); 1018 } 1019 for (i = 0; i < ncount; i++) { 1020 rtvals[i] = VM_PAGER_OK; 1021 } 1022 return rtvals[0]; 1023 } 1024 1025 struct vnode * 1026 vnode_pager_lock(vm_object_t object) 1027 { 1028 struct thread *td = curthread; /* XXX */ 1029 int error; 1030 1031 for (; object != NULL; object = object->backing_object) { 1032 if (object->type != OBJT_VNODE) 1033 continue; 1034 if (object->flags & OBJ_DEAD) 1035 return NULL; 1036 1037 for (;;) { 1038 struct vnode *vp = object->handle; 1039 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE); 1040 if (error == 0) { 1041 if (object->handle != vp) { 1042 vput(vp); 1043 continue; 1044 } 1045 return (vp); 1046 } 1047 if ((object->flags & OBJ_DEAD) || 1048 (object->type != OBJT_VNODE)) { 1049 return NULL; 1050 } 1051 kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp, error, lockstatus(&vp->v_lock, td)); 1052 tsleep(object->handle, 0, "vnpgrl", hz); 1053 } 1054 } 1055 return NULL; 1056 } 1057