1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1990 University of Utah. 5 * Copyright (c) 1991 The Regents of the University of California. 6 * All rights reserved. 7 * Copyright (c) 1993, 1994 John S. Dyson 8 * Copyright (c) 1995, David Greenman 9 * 10 * This code is derived from software contributed to Berkeley by 11 * the Systems Programming Group of the University of Utah Computer 12 * Science Department. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91 39 * $FreeBSD: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $ 40 */ 41 42 /* 43 * Page to/from files (vnodes). 44 */ 45 46 /* 47 * TODO: 48 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will 49 * greatly re-simplify the vnode_pager. 50 */ 51 52 #include <sys/param.h> 53 #include <sys/systm.h> 54 #include <sys/kernel.h> 55 #include <sys/proc.h> 56 #include <sys/vnode.h> 57 #include <sys/mount.h> 58 #include <sys/buf.h> 59 #include <sys/vmmeter.h> 60 #include <sys/conf.h> 61 62 #include <cpu/lwbuf.h> 63 64 #include <vm/vm.h> 65 #include <vm/vm_object.h> 66 #include <vm/vm_page.h> 67 #include <vm/vm_pager.h> 68 #include <vm/vm_map.h> 69 #include <vm/vnode_pager.h> 70 #include <vm/swap_pager.h> 71 #include <vm/vm_extern.h> 72 73 #include <sys/thread2.h> 74 #include <vm/vm_page2.h> 75 76 static void vnode_pager_dealloc (vm_object_t); 77 static int vnode_pager_getpage (vm_object_t, vm_page_t *, int); 78 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, int, int *); 79 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t); 80 81 struct pagerops vnodepagerops = { 82 vnode_pager_dealloc, 83 vnode_pager_getpage, 84 vnode_pager_putpages, 85 vnode_pager_haspage 86 }; 87 88 static struct krate vbadrate = { 1 }; 89 static struct krate vresrate = { 1 }; 90 91 long vnode_pbuf_freecnt = -1; /* start out unlimited */ 92 93 /* 94 * Allocate a VM object for a vnode, typically a regular file vnode. 95 * 96 * Some additional information is required to generate a properly sized 97 * object which covers the entire buffer cache buffer straddling the file 98 * EOF. Userland does not see the extra pages as the VM fault code tests 99 * against v_filesize. 100 */ 101 vm_object_t 102 vnode_pager_alloc(void *handle, off_t length, vm_prot_t prot, off_t offset, 103 int blksize, int boff) 104 { 105 vm_object_t object; 106 struct vnode *vp; 107 off_t loffset; 108 vm_pindex_t lsize; 109 110 /* 111 * Pageout to vnode, no can do yet. 112 */ 113 if (handle == NULL) 114 return (NULL); 115 116 /* 117 * XXX hack - This initialization should be put somewhere else. 118 */ 119 if (vnode_pbuf_freecnt < 0) { 120 vnode_pbuf_freecnt = nswbuf_kva / 2 + 1; 121 } 122 123 /* 124 * Serialize potential vnode/object teardowns and interlocks 125 */ 126 vp = (struct vnode *)handle; 127 lwkt_gettoken(&vp->v_token); 128 129 /* 130 * If the object is being terminated, wait for it to 131 * go away. 132 */ 133 object = vp->v_object; 134 if (object) { 135 vm_object_hold(object); 136 KKASSERT((object->flags & OBJ_DEAD) == 0); 137 } 138 139 if (VREFCNT(vp) <= 0) 140 panic("vnode_pager_alloc: no vnode reference"); 141 142 /* 143 * Round up to the *next* block, then destroy the buffers in question. 144 * Since we are only removing some of the buffers we must rely on the 145 * scan count to determine whether a loop is necessary. 146 * 147 * Destroy any pages beyond the last buffer. 148 */ 149 if (boff < 0) 150 boff = (int)(length % blksize); 151 if (boff) 152 loffset = length + (blksize - boff); 153 else 154 loffset = length; 155 lsize = OFF_TO_IDX(round_page64(loffset)); 156 157 if (object == NULL) { 158 /* 159 * And an object of the appropriate size 160 */ 161 object = vm_object_allocate_hold(OBJT_VNODE, lsize); 162 object->handle = handle; 163 vp->v_object = object; 164 vp->v_filesize = length; 165 if (vp->v_mount && (vp->v_mount->mnt_kern_flag & MNTK_NOMSYNC)) 166 vm_object_set_flag(object, OBJ_NOMSYNC); 167 vref(vp); 168 } else { 169 vm_object_reference_quick(object); /* also vref's */ 170 if (object->size != lsize) { 171 kprintf("vnode_pager_alloc: Warning, objsize " 172 "mismatch %jd/%jd vp=%p obj=%p\n", 173 (intmax_t)object->size, 174 (intmax_t)lsize, 175 vp, object); 176 } 177 if (vp->v_filesize != length) { 178 kprintf("vnode_pager_alloc: Warning, filesize " 179 "mismatch %jd/%jd vp=%p obj=%p\n", 180 (intmax_t)vp->v_filesize, 181 (intmax_t)length, 182 vp, object); 183 } 184 } 185 vm_object_drop(object); 186 lwkt_reltoken(&vp->v_token); 187 188 return (object); 189 } 190 191 /* 192 * Add a ref to a vnode's existing VM object, return the object or 193 * NULL if the vnode did not have one. This does not create the 194 * object (we can't since we don't know what the proper blocksize/boff 195 * is to match the VFS's use of the buffer cache). 196 * 197 * The vnode must be referenced and is typically open. The object should 198 * be stable in this situation. 199 * 200 * Returns the object with an additional reference but not locked. 201 */ 202 vm_object_t 203 vnode_pager_reference(struct vnode *vp) 204 { 205 vm_object_t object; 206 207 if ((object = vp->v_object) != NULL) 208 vm_object_reference_quick(object); /* also vref's vnode */ 209 return (object); 210 } 211 212 static void 213 vnode_pager_dealloc(vm_object_t object) 214 { 215 struct vnode *vp = object->handle; 216 217 if (vp == NULL) 218 panic("vnode_pager_dealloc: pager already dealloced"); 219 220 vm_object_pip_wait(object, "vnpdea"); 221 222 object->handle = NULL; 223 object->type = OBJT_DEAD; 224 vp->v_object = NULL; 225 vp->v_filesize = NOOFFSET; 226 vclrflags(vp, VTEXT | VOBJBUF); 227 swap_pager_freespace_all(object); 228 } 229 230 /* 231 * Return whether the vnode pager has the requested page. Return the 232 * number of disk-contiguous pages before and after the requested page, 233 * not including the requested page. 234 */ 235 static boolean_t 236 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex) 237 { 238 struct vnode *vp = object->handle; 239 off_t loffset; 240 off_t doffset; 241 int voff; 242 int bsize; 243 int error; 244 245 /* 246 * If no vp or vp is doomed or marked transparent to VM, we do not 247 * have the page. 248 */ 249 if ((vp == NULL) || (vp->v_flag & VRECLAIMED)) 250 return FALSE; 251 252 /* 253 * If filesystem no longer mounted or offset beyond end of file we do 254 * not have the page. 255 */ 256 loffset = IDX_TO_OFF(pindex); 257 258 if (vp->v_mount == NULL || loffset >= vp->v_filesize) 259 return FALSE; 260 261 bsize = vp->v_mount->mnt_stat.f_iosize; 262 voff = loffset % bsize; 263 264 /* 265 * XXX 266 * 267 * BMAP returns byte counts before and after, where after 268 * is inclusive of the base page. haspage must return page 269 * counts before and after where after does not include the 270 * base page. 271 * 272 * BMAP is allowed to return a *after of 0 for backwards 273 * compatibility. The base page is still considered valid if 274 * no error is returned. 275 */ 276 error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0); 277 if (error) 278 return TRUE; 279 if (doffset == NOOFFSET) 280 return FALSE; 281 return TRUE; 282 } 283 284 /* 285 * Lets the VM system know about a change in size for a file. 286 * We adjust our own internal size and flush any cached pages in 287 * the associated object that are affected by the size change. 288 * 289 * NOTE: This routine may be invoked as a result of a pager put 290 * operation (possibly at object termination time), so we must be careful. 291 * 292 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that 293 * we do not blow up on the case. nsize will always be >= 0, however. 294 */ 295 void 296 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize) 297 { 298 vm_pindex_t nobjsize; 299 vm_pindex_t oobjsize; 300 vm_object_t object; 301 302 object = vp->v_object; 303 if (object == NULL) 304 return; 305 vm_object_hold(object); 306 KKASSERT(vp->v_object == object); 307 308 /* 309 * Hasn't changed size 310 */ 311 if (nsize == vp->v_filesize) { 312 vm_object_drop(object); 313 return; 314 } 315 316 /* 317 * Has changed size. Adjust the VM object's size and v_filesize 318 * before we start scanning pages to prevent new pages from being 319 * allocated during the scan. 320 */ 321 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 322 oobjsize = object->size; 323 object->size = nobjsize; 324 325 /* 326 * File has shrunk. Toss any cached pages beyond the new EOF. 327 */ 328 if (nsize < vp->v_filesize) { 329 vp->v_filesize = nsize; 330 if (nobjsize < oobjsize) { 331 vm_object_page_remove(object, nobjsize, oobjsize, 332 FALSE); 333 } 334 /* 335 * This gets rid of garbage at the end of a page that is now 336 * only partially backed by the vnode. Since we are setting 337 * the entire page valid & clean after we are done we have 338 * to be sure that the portion of the page within the file 339 * bounds is already valid. If it isn't then making it 340 * valid would create a corrupt block. 341 */ 342 if (nsize & PAGE_MASK) { 343 vm_offset_t kva; 344 vm_page_t m; 345 346 m = vm_page_lookup_busy_wait(object, OFF_TO_IDX(nsize), 347 TRUE, "vsetsz"); 348 349 if (m && m->valid) { 350 int base = (int)nsize & PAGE_MASK; 351 int size = PAGE_SIZE - base; 352 struct lwbuf *lwb; 353 struct lwbuf lwb_cache; 354 355 /* 356 * Clear out partial-page garbage in case 357 * the page has been mapped. 358 * 359 * This is byte aligned. 360 */ 361 lwb = lwbuf_alloc(m, &lwb_cache); 362 kva = lwbuf_kva(lwb); 363 bzero((caddr_t)kva + base, size); 364 lwbuf_free(lwb); 365 366 /* 367 * XXX work around SMP data integrity race 368 * by unmapping the page from user processes. 369 * The garbage we just cleared may be mapped 370 * to a user process running on another cpu 371 * and this code is not running through normal 372 * I/O channels which handle SMP issues for 373 * us, so unmap page to synchronize all cpus. 374 * 375 * XXX should vm_pager_unmap_page() have 376 * dealt with this? 377 */ 378 vm_page_protect(m, VM_PROT_NONE); 379 380 /* 381 * Clear out partial-page dirty bits. This 382 * has the side effect of setting the valid 383 * bits, but that is ok. There are a bunch 384 * of places in the VM system where we expected 385 * m->dirty == VM_PAGE_BITS_ALL. The file EOF 386 * case is one of them. If the page is still 387 * partially dirty, make it fully dirty. 388 * 389 * NOTE: We do not clear out the valid 390 * bits. This would prevent bogus_page 391 * replacement from working properly. 392 * 393 * NOTE: We do not want to clear the dirty 394 * bit for a partial DEV_BSIZE'd truncation! 395 * This is DEV_BSIZE aligned! 396 */ 397 vm_page_clear_dirty_beg_nonincl(m, base, size); 398 if (m->dirty != 0) 399 m->dirty = VM_PAGE_BITS_ALL; 400 vm_page_wakeup(m); 401 } else if (m) { 402 vm_page_wakeup(m); 403 } 404 } 405 } else { 406 vp->v_filesize = nsize; 407 } 408 vm_object_drop(object); 409 } 410 411 /* 412 * Release a page busied for a getpages operation. The page may have become 413 * wired (typically due to being used by the buffer cache) or otherwise been 414 * soft-busied and cannot be freed in that case. A held page can still be 415 * freed. 416 */ 417 void 418 vnode_pager_freepage(vm_page_t m) 419 { 420 if ((m->busy_count & PBUSY_MASK) || 421 m->wire_count || 422 (m->flags & PG_NEED_COMMIT)) { 423 vm_page_activate(m); 424 vm_page_wakeup(m); 425 } else { 426 vm_page_free(m); 427 } 428 } 429 430 /* 431 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 432 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to 433 * vnode_pager_generic_getpages() to implement the previous behaviour. 434 * 435 * All other FS's should use the bypass to get to the local media 436 * backing vp's VOP_GETPAGES. 437 */ 438 static int 439 vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess) 440 { 441 int rtval; 442 struct vnode *vp; 443 444 vp = object->handle; 445 rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess); 446 if (rtval == EOPNOTSUPP) 447 panic("vnode_pager: vfs's must implement vop_getpages"); 448 return rtval; 449 } 450 451 /* 452 * This is now called from local media FS's to operate against their 453 * own vnodes if they fail to implement VOP_GETPAGES. 454 * 455 * With all the caching local media devices do these days there is really 456 * very little point to attempting to restrict the I/O size to contiguous 457 * blocks on-disk, especially if our caller thinks we need all the specified 458 * pages. Just construct and issue a READ. 459 */ 460 int 461 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount, 462 int reqpage, int seqaccess) 463 { 464 struct iovec aiov; 465 struct uio auio; 466 off_t foff; 467 int error; 468 int count; 469 int i; 470 int ioflags; 471 int obytecount; 472 473 /* 474 * Do not do anything if the vnode is bad. 475 */ 476 if (vp->v_mount == NULL) 477 return VM_PAGER_BAD; 478 479 /* 480 * Calculate the number of pages. Since we are paging in whole 481 * pages, adjust bytecount to be an integral multiple of the page 482 * size. It will be clipped to the file EOF later on. 483 */ 484 bytecount = round_page(bytecount); 485 count = bytecount / PAGE_SIZE; 486 487 /* 488 * We could check m[reqpage]->valid here and shortcut the operation, 489 * but doing so breaks read-ahead. Instead assume that the VM 490 * system has already done at least the check, don't worry about 491 * any races, and issue the VOP_READ to allow read-ahead to function. 492 * 493 * This keeps the pipeline full for I/O bound sequentially scanned 494 * mmap()'s 495 */ 496 /* don't shortcut */ 497 498 /* 499 * Discard pages past the file EOF. If the requested page is past 500 * the file EOF we just leave its valid bits set to 0, the caller 501 * expects to maintain ownership of the requested page. If the 502 * entire range is past file EOF discard everything and generate 503 * a pagein error. 504 */ 505 foff = IDX_TO_OFF(mpp[0]->pindex); 506 if (foff >= vp->v_filesize) { 507 for (i = 0; i < count; i++) { 508 if (i != reqpage) 509 vnode_pager_freepage(mpp[i]); 510 } 511 return VM_PAGER_ERROR; 512 } 513 514 if (foff + bytecount > vp->v_filesize) { 515 bytecount = vp->v_filesize - foff; 516 i = round_page(bytecount) / PAGE_SIZE; 517 while (count > i) { 518 --count; 519 if (count != reqpage) 520 vnode_pager_freepage(mpp[count]); 521 } 522 } 523 524 /* 525 * The size of the transfer is bytecount. bytecount will be an 526 * integral multiple of the page size unless it has been clipped 527 * to the file EOF. The transfer cannot exceed the file EOF. 528 * 529 * When dealing with real devices we must round-up to the device 530 * sector size. 531 */ 532 if (vp->v_type == VBLK || vp->v_type == VCHR) { 533 int secmask = vp->v_rdev->si_bsize_phys - 1; 534 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large", secmask + 1)); 535 bytecount = (bytecount + secmask) & ~secmask; 536 } 537 obytecount = bytecount; 538 539 /* 540 * Severe hack to avoid deadlocks with the buffer cache 541 */ 542 for (i = 0; i < count; ++i) { 543 vm_page_t mt = mpp[i]; 544 545 vm_page_io_start(mt); 546 vm_page_wakeup(mt); 547 } 548 549 /* 550 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE 551 */ 552 ioflags = IO_VMIO; 553 if (seqaccess) 554 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 555 556 aiov.iov_base = NULL; 557 aiov.iov_len = bytecount; 558 auio.uio_iov = &aiov; 559 auio.uio_iovcnt = 1; 560 auio.uio_offset = foff; 561 auio.uio_segflg = UIO_NOCOPY; 562 auio.uio_rw = UIO_READ; 563 auio.uio_resid = bytecount; 564 auio.uio_td = NULL; 565 mycpu->gd_cnt.v_vnodein++; 566 mycpu->gd_cnt.v_vnodepgsin += count; 567 568 error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred); 569 570 /* 571 * Severe hack to avoid deadlocks with the buffer cache 572 */ 573 for (i = 0; i < count; ++i) { 574 vm_page_busy_wait(mpp[i], FALSE, "getpgs"); 575 vm_page_io_finish(mpp[i]); 576 } 577 578 /* 579 * Calculate the actual number of bytes read and clean up the 580 * page list. 581 */ 582 bytecount -= auio.uio_resid; 583 584 for (i = 0; i < count; ++i) { 585 vm_page_t mt = mpp[i]; 586 587 if (i != reqpage) { 588 if (error == 0 && mt->valid) { 589 if (mt->flags & PG_REFERENCED) 590 vm_page_activate(mt); 591 else 592 vm_page_deactivate(mt); 593 vm_page_wakeup(mt); 594 } else { 595 vnode_pager_freepage(mt); 596 } 597 } else if (mt->valid == 0) { 598 if (error == 0) { 599 kprintf("page failed but no I/O error page " 600 "%p object %p pindex %d\n", 601 mt, mt->object, (int) mt->pindex); 602 kprintf("i=%d foff=%016lx bytecount=%d/%d" 603 "uioresid=%zd\n", 604 i, foff, obytecount, bytecount, 605 auio.uio_resid); 606 /* whoops, something happened */ 607 error = EINVAL; 608 } 609 } else if (mt->valid != VM_PAGE_BITS_ALL) { 610 /* 611 * Zero-extend the requested page if necessary (if 612 * the filesystem is using a small block size). 613 */ 614 vm_page_zero_invalid(mt, TRUE); 615 } 616 } 617 if (error) { 618 kprintf("vnode_pager_getpage: I/O read error\n"); 619 } 620 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 621 } 622 623 /* 624 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 625 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 626 * vnode_pager_generic_putpages() to implement the previous behaviour. 627 * 628 * Caller has already cleared the pmap modified bits, if any. 629 * 630 * All other FS's should use the bypass to get to the local media 631 * backing vp's VOP_PUTPAGES. 632 */ 633 static void 634 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count, 635 int sync, int *rtvals) 636 { 637 int rtval; 638 struct vnode *vp; 639 int bytes = count * PAGE_SIZE; 640 641 /* 642 * Force synchronous operation if we are extremely low on memory 643 * to prevent a low-memory deadlock. VOP operations often need to 644 * allocate more memory to initiate the I/O ( i.e. do a BMAP 645 * operation ). The swapper handles the case by limiting the amount 646 * of asynchronous I/O, but that sort of solution doesn't scale well 647 * for the vnode pager without a lot of work. 648 * 649 * Also, the backing vnode's iodone routine may not wake the pageout 650 * daemon up. This should be probably be addressed XXX. 651 */ 652 653 if ((vmstats.v_free_count + vmstats.v_cache_count) < 654 vmstats.v_pageout_free_min) { 655 sync |= OBJPC_SYNC; 656 } 657 658 /* 659 * Call device-specific putpages function 660 */ 661 vp = object->handle; 662 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 663 if (rtval == EOPNOTSUPP) { 664 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n"); 665 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals); 666 } 667 } 668 669 670 /* 671 * This is now called from local media FS's to operate against their 672 * own vnodes if they fail to implement VOP_PUTPAGES. 673 * 674 * This is typically called indirectly via the pageout daemon and 675 * clustering has already typically occured, so in general we ask the 676 * underlying filesystem to write the data out asynchronously rather 677 * then delayed. 678 */ 679 int 680 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount, 681 int flags, int *rtvals) 682 { 683 int i; 684 int maxsize, ncount, count; 685 vm_ooffset_t poffset; 686 struct uio auio; 687 struct iovec aiov; 688 int error; 689 int ioflags; 690 691 count = bytecount / PAGE_SIZE; 692 693 for (i = 0; i < count; i++) 694 rtvals[i] = VM_PAGER_AGAIN; 695 696 if ((int) m[0]->pindex < 0) { 697 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n", 698 (long)m[0]->pindex, m[0]->dirty); 699 rtvals[0] = VM_PAGER_BAD; 700 return VM_PAGER_BAD; 701 } 702 703 maxsize = count * PAGE_SIZE; 704 ncount = count; 705 706 poffset = IDX_TO_OFF(m[0]->pindex); 707 708 /* 709 * If the page-aligned write is larger then the actual file we 710 * have to invalidate pages occuring beyond the file EOF. 711 * 712 * If the file EOF resides in the middle of a page we still clear 713 * all of that page's dirty bits later on. If we didn't it would 714 * endlessly re-write. 715 * 716 * We do not under any circumstances truncate the valid bits, as 717 * this will screw up bogus page replacement. 718 * 719 * The caller has already read-protected the pages. The VFS must 720 * use the buffer cache to wrap the pages. The pages might not 721 * be immediately flushed by the buffer cache but once under its 722 * control the pages themselves can wind up being marked clean 723 * and their covering buffer cache buffer can be marked dirty. 724 */ 725 if (poffset + maxsize > vp->v_filesize) { 726 if (poffset < vp->v_filesize) { 727 maxsize = vp->v_filesize - poffset; 728 ncount = btoc(maxsize); 729 } else { 730 maxsize = 0; 731 ncount = 0; 732 } 733 if (ncount < count) { 734 for (i = ncount; i < count; i++) { 735 rtvals[i] = VM_PAGER_BAD; 736 } 737 } 738 } 739 740 /* 741 * pageouts are already clustered, use IO_ASYNC to force a bawrite() 742 * rather then a bdwrite() to prevent paging I/O from saturating 743 * the buffer cache. Dummy-up the sequential heuristic to cause 744 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 745 * the system decides how to cluster. 746 */ 747 ioflags = IO_VMIO; 748 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 749 ioflags |= IO_SYNC; 750 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 751 ioflags |= IO_ASYNC; 752 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 753 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 754 755 aiov.iov_base = (caddr_t) 0; 756 aiov.iov_len = maxsize; 757 auio.uio_iov = &aiov; 758 auio.uio_iovcnt = 1; 759 auio.uio_offset = poffset; 760 auio.uio_segflg = UIO_NOCOPY; 761 auio.uio_rw = UIO_WRITE; 762 auio.uio_resid = maxsize; 763 auio.uio_td = NULL; 764 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred); 765 mycpu->gd_cnt.v_vnodeout++; 766 mycpu->gd_cnt.v_vnodepgsout += ncount; 767 768 if (error) { 769 krateprintf(&vbadrate, 770 "vnode_pager_putpages: I/O error %d\n", error); 771 } 772 if (auio.uio_resid) { 773 krateprintf(&vresrate, 774 "vnode_pager_putpages: residual I/O %zd at %lu\n", 775 auio.uio_resid, (u_long)m[0]->pindex); 776 } 777 if (error == 0) { 778 for (i = 0; i < ncount; i++) { 779 rtvals[i] = VM_PAGER_OK; 780 vm_page_undirty(m[i]); 781 } 782 } 783 return rtvals[0]; 784 } 785 786 /* 787 * Run the chain and if the bottom-most object is a vnode-type lock the 788 * underlying vnode. A locked vnode or NULL is returned. 789 */ 790 struct vnode * 791 vnode_pager_lock(vm_object_t object) 792 { 793 struct vnode *vp = NULL; 794 vm_object_t lobject; 795 vm_object_t tobject; 796 int error; 797 798 if (object == NULL) 799 return(NULL); 800 801 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 802 lobject = object; 803 804 while (lobject->type != OBJT_VNODE) { 805 if (lobject->flags & OBJ_DEAD) 806 break; 807 tobject = lobject->backing_object; 808 if (tobject == NULL) 809 break; 810 vm_object_hold_shared(tobject); 811 if (tobject == lobject->backing_object) { 812 if (lobject != object) { 813 vm_object_lock_swap(); 814 vm_object_drop(lobject); 815 } 816 lobject = tobject; 817 } else { 818 vm_object_drop(tobject); 819 } 820 } 821 while (lobject->type == OBJT_VNODE && 822 (lobject->flags & OBJ_DEAD) == 0) { 823 /* 824 * Extract the vp 825 */ 826 vp = lobject->handle; 827 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE); 828 if (error == 0) { 829 if (lobject->handle == vp) 830 break; 831 vput(vp); 832 } else { 833 kprintf("vnode_pager_lock: vp %p error %d " 834 "lockstatus %d, retrying\n", 835 vp, error, 836 lockstatus(&vp->v_lock, curthread)); 837 tsleep(object->handle, 0, "vnpgrl", hz); 838 } 839 vp = NULL; 840 } 841 if (lobject != object) 842 vm_object_drop(lobject); 843 return (vp); 844 } 845