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 1.43 2008/06/19 23:27:39 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 66 #include <vm/vm.h> 67 #include <vm/vm_object.h> 68 #include <vm/vm_page.h> 69 #include <vm/vm_pager.h> 70 #include <vm/vm_map.h> 71 #include <vm/vnode_pager.h> 72 #include <vm/swap_pager.h> 73 #include <vm/vm_extern.h> 74 75 #include <sys/thread2.h> 76 #include <vm/vm_page2.h> 77 78 static void vnode_pager_dealloc (vm_object_t); 79 static int vnode_pager_getpage (vm_object_t, vm_page_t *, int); 80 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *); 81 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t); 82 83 struct pagerops vnodepagerops = { 84 vnode_pager_dealloc, 85 vnode_pager_getpage, 86 vnode_pager_putpages, 87 vnode_pager_haspage 88 }; 89 90 static struct krate vbadrate = { 1 }; 91 static struct krate vresrate = { 1 }; 92 93 int vnode_pbuf_freecnt = -1; /* start out unlimited */ 94 95 /* 96 * Allocate a VM object for a vnode, typically a regular file vnode. 97 * 98 * Some additional information is required to generate a properly sized 99 * object which covers the entire buffer cache buffer straddling the file 100 * EOF. Userland does not see the extra pages as the VM fault code tests 101 * against v_filesize. 102 */ 103 vm_object_t 104 vnode_pager_alloc(void *handle, off_t length, vm_prot_t prot, off_t offset, 105 int blksize, int boff) 106 { 107 vm_object_t object; 108 struct vnode *vp; 109 off_t loffset; 110 vm_pindex_t lsize; 111 112 /* 113 * Pageout to vnode, no can do yet. 114 */ 115 if (handle == NULL) 116 return (NULL); 117 118 /* 119 * XXX hack - This initialization should be put somewhere else. 120 */ 121 if (vnode_pbuf_freecnt < 0) { 122 vnode_pbuf_freecnt = nswbuf / 2 + 1; 123 } 124 125 vp = (struct vnode *) handle; 126 127 /* 128 * Prevent race condition when allocating the object. This 129 * can happen with NFS vnodes since the nfsnode isn't locked. 130 */ 131 while (vp->v_flag & VOLOCK) { 132 vsetflags(vp, VOWANT); 133 tsleep(vp, 0, "vnpobj", 0); 134 } 135 vsetflags(vp, VOLOCK); 136 137 /* 138 * If the object is being terminated, wait for it to 139 * go away. 140 */ 141 while (((object = vp->v_object) != NULL) && 142 (object->flags & OBJ_DEAD)) { 143 vm_object_dead_sleep(object, "vadead"); 144 } 145 146 if (vp->v_sysref.refcnt <= 0) 147 panic("vnode_pager_alloc: no vnode reference"); 148 149 /* 150 * Round up to the *next* block, then destroy the buffers in question. 151 * Since we are only removing some of the buffers we must rely on the 152 * scan count to determine whether a loop is necessary. 153 * 154 * Destroy any pages beyond the last buffer. 155 */ 156 if (boff < 0) 157 boff = (int)(length % blksize); 158 if (boff) 159 loffset = length + (blksize - boff); 160 else 161 loffset = length; 162 lsize = OFF_TO_IDX(round_page64(loffset)); 163 164 if (object == NULL) { 165 /* 166 * And an object of the appropriate size 167 */ 168 object = vm_object_allocate(OBJT_VNODE, lsize); 169 object->flags = 0; 170 object->handle = handle; 171 vp->v_object = object; 172 vp->v_filesize = length; 173 if (vp->v_mount && (vp->v_mount->mnt_kern_flag & MNTK_NOMSYNC)) 174 object->flags |= OBJ_NOMSYNC; 175 } else { 176 object->ref_count++; 177 if (object->size != lsize) { 178 kprintf("vnode_pager_alloc: Warning, objsize " 179 "mismatch %jd/%jd vp=%p obj=%p\n", 180 (intmax_t)object->size, 181 (intmax_t)lsize, 182 vp, object); 183 } 184 if (vp->v_filesize != length) { 185 kprintf("vnode_pager_alloc: Warning, filesize " 186 "mismatch %jd/%jd vp=%p obj=%p\n", 187 (intmax_t)vp->v_filesize, 188 (intmax_t)length, 189 vp, object); 190 } 191 } 192 vref(vp); 193 194 vclrflags(vp, VOLOCK); 195 if (vp->v_flag & VOWANT) { 196 vclrflags(vp, VOWANT); 197 wakeup(vp); 198 } 199 return (object); 200 } 201 202 /* 203 * Add a ref to a vnode's existing VM object, return the object or 204 * NULL if the vnode did not have one. This does not create the 205 * object (we can't since we don't know what the proper blocksize/boff 206 * is to match the VFS's use of the buffer cache). 207 */ 208 vm_object_t 209 vnode_pager_reference(struct vnode *vp) 210 { 211 vm_object_t object; 212 213 /* 214 * Prevent race condition when allocating the object. This 215 * can happen with NFS vnodes since the nfsnode isn't locked. 216 */ 217 while (vp->v_flag & VOLOCK) { 218 vsetflags(vp, VOWANT); 219 tsleep(vp, 0, "vnpobj", 0); 220 } 221 vsetflags(vp, VOLOCK); 222 223 /* 224 * Prevent race conditions against deallocation of the VM 225 * object. 226 */ 227 while (((object = vp->v_object) != NULL) && 228 (object->flags & OBJ_DEAD)) { 229 vm_object_dead_sleep(object, "vadead"); 230 } 231 232 /* 233 * The object is expected to exist, the caller will handle 234 * NULL returns if it does not. 235 */ 236 if (object) { 237 object->ref_count++; 238 vref(vp); 239 } 240 241 vclrflags(vp, VOLOCK); 242 if (vp->v_flag & VOWANT) { 243 vclrflags(vp, VOWANT); 244 wakeup(vp); 245 } 246 return (object); 247 } 248 249 static void 250 vnode_pager_dealloc(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_pip_wait(object, "vnpdea"); 258 259 object->handle = NULL; 260 object->type = OBJT_DEAD; 261 vp->v_object = NULL; 262 vp->v_filesize = NOOFFSET; 263 vclrflags(vp, VTEXT | VOBJBUF); 264 swap_pager_freespace_all(object); 265 } 266 267 /* 268 * Return whether the vnode pager has the requested page. Return the 269 * number of disk-contiguous pages before and after the requested page, 270 * not including the requested page. 271 */ 272 static boolean_t 273 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex) 274 { 275 struct vnode *vp = object->handle; 276 off_t loffset; 277 off_t doffset; 278 int voff; 279 int bsize; 280 int error; 281 282 /* 283 * If no vp or vp is doomed or marked transparent to VM, we do not 284 * have the page. 285 */ 286 if ((vp == NULL) || (vp->v_flag & VRECLAIMED)) 287 return FALSE; 288 289 /* 290 * If filesystem no longer mounted or offset beyond end of file we do 291 * not have the page. 292 */ 293 loffset = IDX_TO_OFF(pindex); 294 295 if (vp->v_mount == NULL || loffset >= vp->v_filesize) 296 return FALSE; 297 298 bsize = vp->v_mount->mnt_stat.f_iosize; 299 voff = loffset % bsize; 300 301 /* 302 * XXX 303 * 304 * BMAP returns byte counts before and after, where after 305 * is inclusive of the base page. haspage must return page 306 * counts before and after where after does not include the 307 * base page. 308 * 309 * BMAP is allowed to return a *after of 0 for backwards 310 * compatibility. The base page is still considered valid if 311 * no error is returned. 312 */ 313 error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0); 314 if (error) 315 return TRUE; 316 if (doffset == NOOFFSET) 317 return FALSE; 318 return TRUE; 319 } 320 321 /* 322 * Lets the VM system know about a change in size for a file. 323 * We adjust our own internal size and flush any cached pages in 324 * the associated object that are affected by the size change. 325 * 326 * NOTE: This routine may be invoked as a result of a pager put 327 * operation (possibly at object termination time), so we must be careful. 328 * 329 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that 330 * we do not blow up on the case. nsize will always be >= 0, however. 331 */ 332 void 333 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize) 334 { 335 vm_pindex_t nobjsize; 336 vm_pindex_t oobjsize; 337 vm_object_t object = vp->v_object; 338 339 if (object == NULL) 340 return; 341 342 /* 343 * Hasn't changed size 344 */ 345 if (nsize == vp->v_filesize) 346 return; 347 348 /* 349 * Has changed size. Adjust the VM object's size and v_filesize 350 * before we start scanning pages to prevent new pages from being 351 * allocated during the scan. 352 */ 353 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 354 oobjsize = object->size; 355 object->size = nobjsize; 356 357 /* 358 * File has shrunk. Toss any cached pages beyond the new EOF. 359 */ 360 if (nsize < vp->v_filesize) { 361 vp->v_filesize = nsize; 362 if (nobjsize < oobjsize) { 363 vm_object_page_remove(object, nobjsize, oobjsize, 364 FALSE); 365 } 366 /* 367 * This gets rid of garbage at the end of a page that is now 368 * only partially backed by the vnode. Since we are setting 369 * the entire page valid & clean after we are done we have 370 * to be sure that the portion of the page within the file 371 * bounds is already valid. If it isn't then making it 372 * valid would create a corrupt block. 373 */ 374 if (nsize & PAGE_MASK) { 375 vm_offset_t kva; 376 vm_page_t m; 377 378 do { 379 m = vm_page_lookup(object, OFF_TO_IDX(nsize)); 380 } while (m && vm_page_sleep_busy(m, TRUE, "vsetsz")); 381 382 if (m && m->valid) { 383 int base = (int)nsize & PAGE_MASK; 384 int size = PAGE_SIZE - base; 385 struct sf_buf *sf; 386 387 /* 388 * Clear out partial-page garbage in case 389 * the page has been mapped. 390 * 391 * This is byte aligned. 392 */ 393 vm_page_busy(m); 394 sf = sf_buf_alloc(m, SFB_CPUPRIVATE); 395 kva = sf_buf_kva(sf); 396 bzero((caddr_t)kva + base, size); 397 sf_buf_free(sf); 398 399 /* 400 * XXX work around SMP data integrity race 401 * by unmapping the page from user processes. 402 * The garbage we just cleared may be mapped 403 * to a user process running on another cpu 404 * and this code is not running through normal 405 * I/O channels which handle SMP issues for 406 * us, so unmap page to synchronize all cpus. 407 * 408 * XXX should vm_pager_unmap_page() have 409 * dealt with this? 410 */ 411 vm_page_protect(m, VM_PROT_NONE); 412 413 /* 414 * Clear out partial-page dirty bits. This 415 * has the side effect of setting the valid 416 * bits, but that is ok. There are a bunch 417 * of places in the VM system where we expected 418 * m->dirty == VM_PAGE_BITS_ALL. The file EOF 419 * case is one of them. If the page is still 420 * partially dirty, make it fully dirty. 421 * 422 * NOTE: We do not clear out the valid 423 * bits. This would prevent bogus_page 424 * replacement from working properly. 425 * 426 * NOTE: We do not want to clear the dirty 427 * bit for a partial DEV_BSIZE'd truncation! 428 * This is DEV_BSIZE aligned! 429 */ 430 vm_page_clear_dirty_beg_nonincl(m, base, size); 431 if (m->dirty != 0) 432 m->dirty = VM_PAGE_BITS_ALL; 433 vm_page_wakeup(m); 434 } 435 } 436 } else { 437 vp->v_filesize = nsize; 438 } 439 } 440 441 /* 442 * Release a page busied for a getpages operation. The page may have become 443 * wired (typically due to being used by the buffer cache) or otherwise been 444 * soft-busied and cannot be freed in that case. A held page can still be 445 * freed. 446 */ 447 void 448 vnode_pager_freepage(vm_page_t m) 449 { 450 if (m->busy || m->wire_count) { 451 vm_page_activate(m); 452 vm_page_wakeup(m); 453 } else { 454 vm_page_free(m); 455 } 456 } 457 458 /* 459 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 460 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to 461 * vnode_pager_generic_getpages() to implement the previous behaviour. 462 * 463 * All other FS's should use the bypass to get to the local media 464 * backing vp's VOP_GETPAGES. 465 */ 466 static int 467 vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess) 468 { 469 int rtval; 470 struct vnode *vp; 471 472 vp = object->handle; 473 rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess); 474 if (rtval == EOPNOTSUPP) 475 panic("vnode_pager: vfs's must implement vop_getpages\n"); 476 return rtval; 477 } 478 479 /* 480 * This is now called from local media FS's to operate against their 481 * own vnodes if they fail to implement VOP_GETPAGES. 482 * 483 * With all the caching local media devices do these days there is really 484 * very little point to attempting to restrict the I/O size to contiguous 485 * blocks on-disk, especially if our caller thinks we need all the specified 486 * pages. Just construct and issue a READ. 487 */ 488 int 489 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount, 490 int reqpage, int seqaccess) 491 { 492 struct iovec aiov; 493 struct uio auio; 494 off_t foff; 495 int error; 496 int count; 497 int i; 498 int ioflags; 499 500 /* 501 * Do not do anything if the vnode is bad. 502 */ 503 if (vp->v_mount == NULL) 504 return VM_PAGER_BAD; 505 506 /* 507 * Calculate the number of pages. Since we are paging in whole 508 * pages, adjust bytecount to be an integral multiple of the page 509 * size. It will be clipped to the file EOF later on. 510 */ 511 bytecount = round_page(bytecount); 512 count = bytecount / PAGE_SIZE; 513 514 /* 515 * We could check m[reqpage]->valid here and shortcut the operation, 516 * but doing so breaks read-ahead. Instead assume that the VM 517 * system has already done at least the check, don't worry about 518 * any races, and issue the VOP_READ to allow read-ahead to function. 519 * 520 * This keeps the pipeline full for I/O bound sequentially scanned 521 * mmap()'s 522 */ 523 /* don't shortcut */ 524 525 /* 526 * Discard pages past the file EOF. If the requested page is past 527 * the file EOF we just leave its valid bits set to 0, the caller 528 * expects to maintain ownership of the requested page. If the 529 * entire range is past file EOF discard everything and generate 530 * a pagein error. 531 */ 532 foff = IDX_TO_OFF(mpp[0]->pindex); 533 if (foff >= vp->v_filesize) { 534 for (i = 0; i < count; i++) { 535 if (i != reqpage) 536 vnode_pager_freepage(mpp[i]); 537 } 538 return VM_PAGER_ERROR; 539 } 540 541 if (foff + bytecount > vp->v_filesize) { 542 bytecount = vp->v_filesize - foff; 543 i = round_page(bytecount) / PAGE_SIZE; 544 while (count > i) { 545 --count; 546 if (count != reqpage) 547 vnode_pager_freepage(mpp[count]); 548 } 549 } 550 551 /* 552 * The size of the transfer is bytecount. bytecount will be an 553 * integral multiple of the page size unless it has been clipped 554 * to the file EOF. The transfer cannot exceed the file EOF. 555 * 556 * When dealing with real devices we must round-up to the device 557 * sector size. 558 */ 559 if (vp->v_type == VBLK || vp->v_type == VCHR) { 560 int secmask = vp->v_rdev->si_bsize_phys - 1; 561 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1)); 562 bytecount = (bytecount + secmask) & ~secmask; 563 } 564 565 /* 566 * Severe hack to avoid deadlocks with the buffer cache 567 */ 568 for (i = 0; i < count; ++i) { 569 vm_page_t mt = mpp[i]; 570 571 vm_page_io_start(mt); 572 vm_page_wakeup(mt); 573 } 574 575 /* 576 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE 577 */ 578 ioflags = IO_VMIO; 579 if (seqaccess) 580 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 581 582 aiov.iov_base = NULL; 583 aiov.iov_len = bytecount; 584 auio.uio_iov = &aiov; 585 auio.uio_iovcnt = 1; 586 auio.uio_offset = foff; 587 auio.uio_segflg = UIO_NOCOPY; 588 auio.uio_rw = UIO_READ; 589 auio.uio_resid = bytecount; 590 auio.uio_td = NULL; 591 mycpu->gd_cnt.v_vnodein++; 592 mycpu->gd_cnt.v_vnodepgsin += count; 593 594 error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred); 595 596 /* 597 * Severe hack to avoid deadlocks with the buffer cache 598 */ 599 for (i = 0; i < count; ++i) { 600 vm_page_t mt = mpp[i]; 601 602 while (vm_page_sleep_busy(mt, FALSE, "getpgs")) 603 ; 604 vm_page_busy(mt); 605 vm_page_io_finish(mt); 606 } 607 608 /* 609 * Calculate the actual number of bytes read and clean up the 610 * page list. 611 */ 612 bytecount -= auio.uio_resid; 613 614 for (i = 0; i < count; ++i) { 615 vm_page_t mt = mpp[i]; 616 617 if (i != reqpage) { 618 if (error == 0 && mt->valid) { 619 if (mt->flags & PG_WANTED) 620 vm_page_activate(mt); 621 else 622 vm_page_deactivate(mt); 623 vm_page_wakeup(mt); 624 } else { 625 vnode_pager_freepage(mt); 626 } 627 } else if (mt->valid == 0) { 628 if (error == 0) { 629 kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt, mt->object, (int) mt->pindex); 630 /* whoops, something happened */ 631 error = EINVAL; 632 } 633 } else if (mt->valid != VM_PAGE_BITS_ALL) { 634 /* 635 * Zero-extend the requested page if necessary (if 636 * the filesystem is using a small block size). 637 */ 638 vm_page_zero_invalid(mt, TRUE); 639 } 640 } 641 if (error) { 642 kprintf("vnode_pager_getpage: I/O read error\n"); 643 } 644 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 645 } 646 647 /* 648 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 649 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 650 * vnode_pager_generic_putpages() to implement the previous behaviour. 651 * 652 * Caller has already cleared the pmap modified bits, if any. 653 * 654 * All other FS's should use the bypass to get to the local media 655 * backing vp's VOP_PUTPAGES. 656 */ 657 static void 658 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count, 659 boolean_t sync, int *rtvals) 660 { 661 int rtval; 662 struct vnode *vp; 663 int bytes = count * PAGE_SIZE; 664 665 /* 666 * Force synchronous operation if we are extremely low on memory 667 * to prevent a low-memory deadlock. VOP operations often need to 668 * allocate more memory to initiate the I/O ( i.e. do a BMAP 669 * operation ). The swapper handles the case by limiting the amount 670 * of asynchronous I/O, but that sort of solution doesn't scale well 671 * for the vnode pager without a lot of work. 672 * 673 * Also, the backing vnode's iodone routine may not wake the pageout 674 * daemon up. This should be probably be addressed XXX. 675 */ 676 677 if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min) 678 sync |= OBJPC_SYNC; 679 680 /* 681 * Call device-specific putpages function 682 */ 683 vp = object->handle; 684 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 685 if (rtval == EOPNOTSUPP) { 686 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n"); 687 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals); 688 } 689 } 690 691 692 /* 693 * This is now called from local media FS's to operate against their 694 * own vnodes if they fail to implement VOP_PUTPAGES. 695 * 696 * This is typically called indirectly via the pageout daemon and 697 * clustering has already typically occured, so in general we ask the 698 * underlying filesystem to write the data out asynchronously rather 699 * then delayed. 700 */ 701 int 702 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount, 703 int flags, int *rtvals) 704 { 705 int i; 706 vm_object_t object; 707 int maxsize, ncount, count; 708 vm_ooffset_t poffset; 709 struct uio auio; 710 struct iovec aiov; 711 int error; 712 int ioflags; 713 714 object = vp->v_object; 715 count = bytecount / PAGE_SIZE; 716 717 for (i = 0; i < count; i++) 718 rtvals[i] = VM_PAGER_AGAIN; 719 720 if ((int) m[0]->pindex < 0) { 721 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n", 722 (long)m[0]->pindex, m[0]->dirty); 723 rtvals[0] = VM_PAGER_BAD; 724 return VM_PAGER_BAD; 725 } 726 727 maxsize = count * PAGE_SIZE; 728 ncount = count; 729 730 poffset = IDX_TO_OFF(m[0]->pindex); 731 732 /* 733 * If the page-aligned write is larger then the actual file we 734 * have to invalidate pages occuring beyond the file EOF. 735 * 736 * If the file EOF resides in the middle of a page we still clear 737 * all of that page's dirty bits later on. If we didn't it would 738 * endlessly re-write. 739 * 740 * We do not under any circumstances truncate the valid bits, as 741 * this will screw up bogus page replacement. 742 * 743 * The caller has already read-protected the pages. The VFS must 744 * use the buffer cache to wrap the pages. The pages might not 745 * be immediately flushed by the buffer cache but once under its 746 * control the pages themselves can wind up being marked clean 747 * and their covering buffer cache buffer can be marked dirty. 748 */ 749 if (poffset + maxsize > vp->v_filesize) { 750 if (poffset < vp->v_filesize) { 751 maxsize = vp->v_filesize - poffset; 752 ncount = btoc(maxsize); 753 } else { 754 maxsize = 0; 755 ncount = 0; 756 } 757 if (ncount < count) { 758 for (i = ncount; i < count; i++) { 759 rtvals[i] = VM_PAGER_BAD; 760 } 761 } 762 } 763 764 /* 765 * pageouts are already clustered, use IO_ASYNC to force a bawrite() 766 * rather then a bdwrite() to prevent paging I/O from saturating 767 * the buffer cache. Dummy-up the sequential heuristic to cause 768 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 769 * the system decides how to cluster. 770 */ 771 ioflags = IO_VMIO; 772 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 773 ioflags |= IO_SYNC; 774 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 775 ioflags |= IO_ASYNC; 776 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 777 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 778 779 aiov.iov_base = (caddr_t) 0; 780 aiov.iov_len = maxsize; 781 auio.uio_iov = &aiov; 782 auio.uio_iovcnt = 1; 783 auio.uio_offset = poffset; 784 auio.uio_segflg = UIO_NOCOPY; 785 auio.uio_rw = UIO_WRITE; 786 auio.uio_resid = maxsize; 787 auio.uio_td = NULL; 788 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred); 789 mycpu->gd_cnt.v_vnodeout++; 790 mycpu->gd_cnt.v_vnodepgsout += ncount; 791 792 if (error) { 793 krateprintf(&vbadrate, 794 "vnode_pager_putpages: I/O error %d\n", error); 795 } 796 if (auio.uio_resid) { 797 krateprintf(&vresrate, 798 "vnode_pager_putpages: residual I/O %zd at %lu\n", 799 auio.uio_resid, (u_long)m[0]->pindex); 800 } 801 if (error == 0) { 802 for (i = 0; i < ncount; i++) { 803 rtvals[i] = VM_PAGER_OK; 804 vm_page_undirty(m[i]); 805 } 806 } 807 return rtvals[0]; 808 } 809 810 struct vnode * 811 vnode_pager_lock(vm_object_t object) 812 { 813 struct thread *td = curthread; /* XXX */ 814 int error; 815 816 for (; object != NULL; object = object->backing_object) { 817 if (object->type != OBJT_VNODE) 818 continue; 819 if (object->flags & OBJ_DEAD) 820 return NULL; 821 822 for (;;) { 823 struct vnode *vp = object->handle; 824 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE); 825 if (error == 0) { 826 if (object->handle != vp) { 827 vput(vp); 828 continue; 829 } 830 return (vp); 831 } 832 if ((object->flags & OBJ_DEAD) || 833 (object->type != OBJT_VNODE)) { 834 return NULL; 835 } 836 kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp, error, lockstatus(&vp->v_lock, td)); 837 tsleep(object->handle, 0, "vnpgrl", hz); 838 } 839 } 840 return NULL; 841 } 842