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 / 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 || m->wire_count || (m->flags & PG_NEED_COMMIT)) { 421 vm_page_activate(m); 422 vm_page_wakeup(m); 423 } else { 424 vm_page_free(m); 425 } 426 } 427 428 /* 429 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 430 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to 431 * vnode_pager_generic_getpages() to implement the previous behaviour. 432 * 433 * All other FS's should use the bypass to get to the local media 434 * backing vp's VOP_GETPAGES. 435 */ 436 static int 437 vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess) 438 { 439 int rtval; 440 struct vnode *vp; 441 442 vp = object->handle; 443 rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess); 444 if (rtval == EOPNOTSUPP) 445 panic("vnode_pager: vfs's must implement vop_getpages"); 446 return rtval; 447 } 448 449 /* 450 * This is now called from local media FS's to operate against their 451 * own vnodes if they fail to implement VOP_GETPAGES. 452 * 453 * With all the caching local media devices do these days there is really 454 * very little point to attempting to restrict the I/O size to contiguous 455 * blocks on-disk, especially if our caller thinks we need all the specified 456 * pages. Just construct and issue a READ. 457 */ 458 int 459 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount, 460 int reqpage, int seqaccess) 461 { 462 struct iovec aiov; 463 struct uio auio; 464 off_t foff; 465 int error; 466 int count; 467 int i; 468 int ioflags; 469 470 /* 471 * Do not do anything if the vnode is bad. 472 */ 473 if (vp->v_mount == NULL) 474 return VM_PAGER_BAD; 475 476 /* 477 * Calculate the number of pages. Since we are paging in whole 478 * pages, adjust bytecount to be an integral multiple of the page 479 * size. It will be clipped to the file EOF later on. 480 */ 481 bytecount = round_page(bytecount); 482 count = bytecount / PAGE_SIZE; 483 484 /* 485 * We could check m[reqpage]->valid here and shortcut the operation, 486 * but doing so breaks read-ahead. Instead assume that the VM 487 * system has already done at least the check, don't worry about 488 * any races, and issue the VOP_READ to allow read-ahead to function. 489 * 490 * This keeps the pipeline full for I/O bound sequentially scanned 491 * mmap()'s 492 */ 493 /* don't shortcut */ 494 495 /* 496 * Discard pages past the file EOF. If the requested page is past 497 * the file EOF we just leave its valid bits set to 0, the caller 498 * expects to maintain ownership of the requested page. If the 499 * entire range is past file EOF discard everything and generate 500 * a pagein error. 501 */ 502 foff = IDX_TO_OFF(mpp[0]->pindex); 503 if (foff >= vp->v_filesize) { 504 for (i = 0; i < count; i++) { 505 if (i != reqpage) 506 vnode_pager_freepage(mpp[i]); 507 } 508 return VM_PAGER_ERROR; 509 } 510 511 if (foff + bytecount > vp->v_filesize) { 512 bytecount = vp->v_filesize - foff; 513 i = round_page(bytecount) / PAGE_SIZE; 514 while (count > i) { 515 --count; 516 if (count != reqpage) 517 vnode_pager_freepage(mpp[count]); 518 } 519 } 520 521 /* 522 * The size of the transfer is bytecount. bytecount will be an 523 * integral multiple of the page size unless it has been clipped 524 * to the file EOF. The transfer cannot exceed the file EOF. 525 * 526 * When dealing with real devices we must round-up to the device 527 * sector size. 528 */ 529 if (vp->v_type == VBLK || vp->v_type == VCHR) { 530 int secmask = vp->v_rdev->si_bsize_phys - 1; 531 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large", secmask + 1)); 532 bytecount = (bytecount + secmask) & ~secmask; 533 } 534 535 /* 536 * Severe hack to avoid deadlocks with the buffer cache 537 */ 538 for (i = 0; i < count; ++i) { 539 vm_page_t mt = mpp[i]; 540 541 vm_page_io_start(mt); 542 vm_page_wakeup(mt); 543 } 544 545 /* 546 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE 547 */ 548 ioflags = IO_VMIO; 549 if (seqaccess) 550 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 551 552 aiov.iov_base = NULL; 553 aiov.iov_len = bytecount; 554 auio.uio_iov = &aiov; 555 auio.uio_iovcnt = 1; 556 auio.uio_offset = foff; 557 auio.uio_segflg = UIO_NOCOPY; 558 auio.uio_rw = UIO_READ; 559 auio.uio_resid = bytecount; 560 auio.uio_td = NULL; 561 mycpu->gd_cnt.v_vnodein++; 562 mycpu->gd_cnt.v_vnodepgsin += count; 563 564 error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred); 565 566 /* 567 * Severe hack to avoid deadlocks with the buffer cache 568 */ 569 for (i = 0; i < count; ++i) { 570 vm_page_busy_wait(mpp[i], FALSE, "getpgs"); 571 vm_page_io_finish(mpp[i]); 572 } 573 574 /* 575 * Calculate the actual number of bytes read and clean up the 576 * page list. 577 */ 578 bytecount -= auio.uio_resid; 579 580 for (i = 0; i < count; ++i) { 581 vm_page_t mt = mpp[i]; 582 583 if (i != reqpage) { 584 if (error == 0 && mt->valid) { 585 if (mt->flags & PG_REFERENCED) 586 vm_page_activate(mt); 587 else 588 vm_page_deactivate(mt); 589 vm_page_wakeup(mt); 590 } else { 591 vnode_pager_freepage(mt); 592 } 593 } else if (mt->valid == 0) { 594 if (error == 0) { 595 kprintf("page failed but no I/O error page " 596 "%p object %p pindex %d\n", 597 mt, mt->object, (int) mt->pindex); 598 /* whoops, something happened */ 599 error = EINVAL; 600 } 601 } else if (mt->valid != VM_PAGE_BITS_ALL) { 602 /* 603 * Zero-extend the requested page if necessary (if 604 * the filesystem is using a small block size). 605 */ 606 vm_page_zero_invalid(mt, TRUE); 607 } 608 } 609 if (error) { 610 kprintf("vnode_pager_getpage: I/O read error\n"); 611 } 612 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 613 } 614 615 /* 616 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 617 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 618 * vnode_pager_generic_putpages() to implement the previous behaviour. 619 * 620 * Caller has already cleared the pmap modified bits, if any. 621 * 622 * All other FS's should use the bypass to get to the local media 623 * backing vp's VOP_PUTPAGES. 624 */ 625 static void 626 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count, 627 int sync, int *rtvals) 628 { 629 int rtval; 630 struct vnode *vp; 631 int bytes = count * PAGE_SIZE; 632 633 /* 634 * Force synchronous operation if we are extremely low on memory 635 * to prevent a low-memory deadlock. VOP operations often need to 636 * allocate more memory to initiate the I/O ( i.e. do a BMAP 637 * operation ). The swapper handles the case by limiting the amount 638 * of asynchronous I/O, but that sort of solution doesn't scale well 639 * for the vnode pager without a lot of work. 640 * 641 * Also, the backing vnode's iodone routine may not wake the pageout 642 * daemon up. This should be probably be addressed XXX. 643 */ 644 645 if ((vmstats.v_free_count + vmstats.v_cache_count) < 646 vmstats.v_pageout_free_min) { 647 sync |= OBJPC_SYNC; 648 } 649 650 /* 651 * Call device-specific putpages function 652 */ 653 vp = object->handle; 654 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 655 if (rtval == EOPNOTSUPP) { 656 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n"); 657 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals); 658 } 659 } 660 661 662 /* 663 * This is now called from local media FS's to operate against their 664 * own vnodes if they fail to implement VOP_PUTPAGES. 665 * 666 * This is typically called indirectly via the pageout daemon and 667 * clustering has already typically occured, so in general we ask the 668 * underlying filesystem to write the data out asynchronously rather 669 * then delayed. 670 */ 671 int 672 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount, 673 int flags, int *rtvals) 674 { 675 int i; 676 int maxsize, ncount, count; 677 vm_ooffset_t poffset; 678 struct uio auio; 679 struct iovec aiov; 680 int error; 681 int ioflags; 682 683 count = bytecount / PAGE_SIZE; 684 685 for (i = 0; i < count; i++) 686 rtvals[i] = VM_PAGER_AGAIN; 687 688 if ((int) m[0]->pindex < 0) { 689 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n", 690 (long)m[0]->pindex, m[0]->dirty); 691 rtvals[0] = VM_PAGER_BAD; 692 return VM_PAGER_BAD; 693 } 694 695 maxsize = count * PAGE_SIZE; 696 ncount = count; 697 698 poffset = IDX_TO_OFF(m[0]->pindex); 699 700 /* 701 * If the page-aligned write is larger then the actual file we 702 * have to invalidate pages occuring beyond the file EOF. 703 * 704 * If the file EOF resides in the middle of a page we still clear 705 * all of that page's dirty bits later on. If we didn't it would 706 * endlessly re-write. 707 * 708 * We do not under any circumstances truncate the valid bits, as 709 * this will screw up bogus page replacement. 710 * 711 * The caller has already read-protected the pages. The VFS must 712 * use the buffer cache to wrap the pages. The pages might not 713 * be immediately flushed by the buffer cache but once under its 714 * control the pages themselves can wind up being marked clean 715 * and their covering buffer cache buffer can be marked dirty. 716 */ 717 if (poffset + maxsize > vp->v_filesize) { 718 if (poffset < vp->v_filesize) { 719 maxsize = vp->v_filesize - poffset; 720 ncount = btoc(maxsize); 721 } else { 722 maxsize = 0; 723 ncount = 0; 724 } 725 if (ncount < count) { 726 for (i = ncount; i < count; i++) { 727 rtvals[i] = VM_PAGER_BAD; 728 } 729 } 730 } 731 732 /* 733 * pageouts are already clustered, use IO_ASYNC to force a bawrite() 734 * rather then a bdwrite() to prevent paging I/O from saturating 735 * the buffer cache. Dummy-up the sequential heuristic to cause 736 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 737 * the system decides how to cluster. 738 */ 739 ioflags = IO_VMIO; 740 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 741 ioflags |= IO_SYNC; 742 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 743 ioflags |= IO_ASYNC; 744 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 745 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 746 747 aiov.iov_base = (caddr_t) 0; 748 aiov.iov_len = maxsize; 749 auio.uio_iov = &aiov; 750 auio.uio_iovcnt = 1; 751 auio.uio_offset = poffset; 752 auio.uio_segflg = UIO_NOCOPY; 753 auio.uio_rw = UIO_WRITE; 754 auio.uio_resid = maxsize; 755 auio.uio_td = NULL; 756 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred); 757 mycpu->gd_cnt.v_vnodeout++; 758 mycpu->gd_cnt.v_vnodepgsout += ncount; 759 760 if (error) { 761 krateprintf(&vbadrate, 762 "vnode_pager_putpages: I/O error %d\n", error); 763 } 764 if (auio.uio_resid) { 765 krateprintf(&vresrate, 766 "vnode_pager_putpages: residual I/O %zd at %lu\n", 767 auio.uio_resid, (u_long)m[0]->pindex); 768 } 769 if (error == 0) { 770 for (i = 0; i < ncount; i++) { 771 rtvals[i] = VM_PAGER_OK; 772 vm_page_undirty(m[i]); 773 } 774 } 775 return rtvals[0]; 776 } 777 778 /* 779 * Run the chain and if the bottom-most object is a vnode-type lock the 780 * underlying vnode. A locked vnode or NULL is returned. 781 */ 782 struct vnode * 783 vnode_pager_lock(vm_object_t object) 784 { 785 struct vnode *vp = NULL; 786 vm_object_t lobject; 787 vm_object_t tobject; 788 int error; 789 790 if (object == NULL) 791 return(NULL); 792 793 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 794 lobject = object; 795 796 while (lobject->type != OBJT_VNODE) { 797 if (lobject->flags & OBJ_DEAD) 798 break; 799 tobject = lobject->backing_object; 800 if (tobject == NULL) 801 break; 802 vm_object_hold_shared(tobject); 803 if (tobject == lobject->backing_object) { 804 if (lobject != object) { 805 vm_object_lock_swap(); 806 vm_object_drop(lobject); 807 } 808 lobject = tobject; 809 } else { 810 vm_object_drop(tobject); 811 } 812 } 813 while (lobject->type == OBJT_VNODE && 814 (lobject->flags & OBJ_DEAD) == 0) { 815 /* 816 * Extract the vp 817 */ 818 vp = lobject->handle; 819 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE); 820 if (error == 0) { 821 if (lobject->handle == vp) 822 break; 823 vput(vp); 824 } else { 825 kprintf("vnode_pager_lock: vp %p error %d " 826 "lockstatus %d, retrying\n", 827 vp, error, 828 lockstatus(&vp->v_lock, curthread)); 829 tsleep(object->handle, 0, "vnpgrl", hz); 830 } 831 vp = NULL; 832 } 833 if (lobject != object) 834 vm_object_drop(lobject); 835 return (vp); 836 } 837