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. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91 43 * $FreeBSD: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $ 44 */ 45 46 /* 47 * Page to/from files (vnodes). 48 */ 49 50 /* 51 * TODO: 52 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will 53 * greatly re-simplify the vnode_pager. 54 */ 55 56 #include <sys/param.h> 57 #include <sys/systm.h> 58 #include <sys/kernel.h> 59 #include <sys/proc.h> 60 #include <sys/vnode.h> 61 #include <sys/mount.h> 62 #include <sys/buf.h> 63 #include <sys/vmmeter.h> 64 #include <sys/conf.h> 65 66 #include <cpu/lwbuf.h> 67 68 #include <vm/vm.h> 69 #include <vm/vm_object.h> 70 #include <vm/vm_page.h> 71 #include <vm/vm_pager.h> 72 #include <vm/vm_map.h> 73 #include <vm/vnode_pager.h> 74 #include <vm/swap_pager.h> 75 #include <vm/vm_extern.h> 76 77 #include <sys/thread2.h> 78 #include <vm/vm_page2.h> 79 80 static void vnode_pager_dealloc (vm_object_t); 81 static int vnode_pager_getpage (vm_object_t, vm_page_t *, int); 82 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *); 83 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t); 84 85 struct pagerops vnodepagerops = { 86 vnode_pager_dealloc, 87 vnode_pager_getpage, 88 vnode_pager_putpages, 89 vnode_pager_haspage 90 }; 91 92 static struct krate vbadrate = { 1 }; 93 static struct krate vresrate = { 1 }; 94 95 int vnode_pbuf_freecnt = -1; /* start out unlimited */ 96 97 /* 98 * Allocate a VM object for a vnode, typically a regular file vnode. 99 * 100 * Some additional information is required to generate a properly sized 101 * object which covers the entire buffer cache buffer straddling the file 102 * EOF. Userland does not see the extra pages as the VM fault code tests 103 * against v_filesize. 104 */ 105 vm_object_t 106 vnode_pager_alloc(void *handle, off_t length, vm_prot_t prot, off_t offset, 107 int blksize, int boff) 108 { 109 vm_object_t object; 110 struct vnode *vp; 111 off_t loffset; 112 vm_pindex_t lsize; 113 114 /* 115 * Pageout to vnode, no can do yet. 116 */ 117 if (handle == NULL) 118 return (NULL); 119 120 /* 121 * XXX hack - This initialization should be put somewhere else. 122 */ 123 if (vnode_pbuf_freecnt < 0) { 124 vnode_pbuf_freecnt = nswbuf / 2 + 1; 125 } 126 127 /* 128 * Serialize potential vnode/object teardowns and interlocks 129 */ 130 vp = (struct vnode *)handle; 131 lwkt_gettoken(&vp->v_token); 132 133 /* 134 * Prevent race condition when allocating the object. This 135 * can happen with NFS vnodes since the nfsnode isn't locked. 136 */ 137 while (vp->v_flag & VOLOCK) { 138 vsetflags(vp, VOWANT); 139 tsleep(vp, 0, "vnpobj", 0); 140 } 141 vsetflags(vp, VOLOCK); 142 lwkt_reltoken(&vp->v_token); 143 144 /* 145 * If the object is being terminated, wait for it to 146 * go away. 147 */ 148 while ((object = vp->v_object) != NULL) { 149 vm_object_hold(object); 150 if ((object->flags & OBJ_DEAD) == 0) 151 break; 152 vm_object_dead_sleep(object, "vadead"); 153 vm_object_drop(object); 154 } 155 156 if (vp->v_sysref.refcnt <= 0) 157 panic("vnode_pager_alloc: no vnode reference"); 158 159 /* 160 * Round up to the *next* block, then destroy the buffers in question. 161 * Since we are only removing some of the buffers we must rely on the 162 * scan count to determine whether a loop is necessary. 163 * 164 * Destroy any pages beyond the last buffer. 165 */ 166 if (boff < 0) 167 boff = (int)(length % blksize); 168 if (boff) 169 loffset = length + (blksize - boff); 170 else 171 loffset = length; 172 lsize = OFF_TO_IDX(round_page64(loffset)); 173 174 if (object == NULL) { 175 /* 176 * And an object of the appropriate size 177 */ 178 object = vm_object_allocate_hold(OBJT_VNODE, lsize); 179 object->handle = handle; 180 vp->v_object = object; 181 vp->v_filesize = length; 182 if (vp->v_mount && (vp->v_mount->mnt_kern_flag & MNTK_NOMSYNC)) 183 vm_object_set_flag(object, OBJ_NOMSYNC); 184 } else { 185 object->ref_count++; 186 if (object->size != lsize) { 187 kprintf("vnode_pager_alloc: Warning, objsize " 188 "mismatch %jd/%jd vp=%p obj=%p\n", 189 (intmax_t)object->size, 190 (intmax_t)lsize, 191 vp, object); 192 } 193 if (vp->v_filesize != length) { 194 kprintf("vnode_pager_alloc: Warning, filesize " 195 "mismatch %jd/%jd vp=%p obj=%p\n", 196 (intmax_t)vp->v_filesize, 197 (intmax_t)length, 198 vp, object); 199 } 200 } 201 202 vref(vp); 203 lwkt_gettoken(&vp->v_token); 204 vclrflags(vp, VOLOCK); 205 if (vp->v_flag & VOWANT) { 206 vclrflags(vp, VOWANT); 207 wakeup(vp); 208 } 209 lwkt_reltoken(&vp->v_token); 210 211 vm_object_drop(object); 212 213 return (object); 214 } 215 216 /* 217 * Add a ref to a vnode's existing VM object, return the object or 218 * NULL if the vnode did not have one. This does not create the 219 * object (we can't since we don't know what the proper blocksize/boff 220 * is to match the VFS's use of the buffer cache). 221 */ 222 vm_object_t 223 vnode_pager_reference(struct vnode *vp) 224 { 225 vm_object_t object; 226 227 /* 228 * Prevent race condition when allocating the object. This 229 * can happen with NFS vnodes since the nfsnode isn't locked. 230 * 231 * Serialize potential vnode/object teardowns and interlocks 232 */ 233 lwkt_gettoken(&vp->v_token); 234 while (vp->v_flag & VOLOCK) { 235 vsetflags(vp, VOWANT); 236 tsleep(vp, 0, "vnpobj", 0); 237 } 238 vsetflags(vp, VOLOCK); 239 lwkt_reltoken(&vp->v_token); 240 241 /* 242 * Prevent race conditions against deallocation of the VM 243 * object. 244 */ 245 while ((object = vp->v_object) != NULL) { 246 vm_object_hold(object); 247 if ((object->flags & OBJ_DEAD) == 0) 248 break; 249 vm_object_dead_sleep(object, "vadead"); 250 vm_object_drop(object); 251 } 252 253 /* 254 * The object is expected to exist, the caller will handle 255 * NULL returns if it does not. 256 */ 257 if (object) { 258 object->ref_count++; 259 vref(vp); 260 } 261 262 lwkt_gettoken(&vp->v_token); 263 vclrflags(vp, VOLOCK); 264 if (vp->v_flag & VOWANT) { 265 vclrflags(vp, VOWANT); 266 wakeup(vp); 267 } 268 lwkt_reltoken(&vp->v_token); 269 if (object) 270 vm_object_drop(object); 271 272 return (object); 273 } 274 275 static void 276 vnode_pager_dealloc(vm_object_t object) 277 { 278 struct vnode *vp = object->handle; 279 280 if (vp == NULL) 281 panic("vnode_pager_dealloc: pager already dealloced"); 282 283 vm_object_pip_wait(object, "vnpdea"); 284 285 object->handle = NULL; 286 object->type = OBJT_DEAD; 287 vp->v_object = NULL; 288 vp->v_filesize = NOOFFSET; 289 vclrflags(vp, VTEXT | VOBJBUF); 290 swap_pager_freespace_all(object); 291 } 292 293 /* 294 * Return whether the vnode pager has the requested page. Return the 295 * number of disk-contiguous pages before and after the requested page, 296 * not including the requested page. 297 */ 298 static boolean_t 299 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex) 300 { 301 struct vnode *vp = object->handle; 302 off_t loffset; 303 off_t doffset; 304 int voff; 305 int bsize; 306 int error; 307 308 /* 309 * If no vp or vp is doomed or marked transparent to VM, we do not 310 * have the page. 311 */ 312 if ((vp == NULL) || (vp->v_flag & VRECLAIMED)) 313 return FALSE; 314 315 /* 316 * If filesystem no longer mounted or offset beyond end of file we do 317 * not have the page. 318 */ 319 loffset = IDX_TO_OFF(pindex); 320 321 if (vp->v_mount == NULL || loffset >= vp->v_filesize) 322 return FALSE; 323 324 bsize = vp->v_mount->mnt_stat.f_iosize; 325 voff = loffset % bsize; 326 327 /* 328 * XXX 329 * 330 * BMAP returns byte counts before and after, where after 331 * is inclusive of the base page. haspage must return page 332 * counts before and after where after does not include the 333 * base page. 334 * 335 * BMAP is allowed to return a *after of 0 for backwards 336 * compatibility. The base page is still considered valid if 337 * no error is returned. 338 */ 339 error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0); 340 if (error) 341 return TRUE; 342 if (doffset == NOOFFSET) 343 return FALSE; 344 return TRUE; 345 } 346 347 /* 348 * Lets the VM system know about a change in size for a file. 349 * We adjust our own internal size and flush any cached pages in 350 * the associated object that are affected by the size change. 351 * 352 * NOTE: This routine may be invoked as a result of a pager put 353 * operation (possibly at object termination time), so we must be careful. 354 * 355 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that 356 * we do not blow up on the case. nsize will always be >= 0, however. 357 */ 358 void 359 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize) 360 { 361 vm_pindex_t nobjsize; 362 vm_pindex_t oobjsize; 363 vm_object_t object; 364 365 while ((object = vp->v_object) != NULL) { 366 vm_object_hold(object); 367 if (vp->v_object == object) 368 break; 369 vm_object_drop(object); 370 } 371 if (object == NULL) 372 return; 373 374 /* 375 * Hasn't changed size 376 */ 377 if (nsize == vp->v_filesize) { 378 vm_object_drop(object); 379 return; 380 } 381 382 /* 383 * Has changed size. Adjust the VM object's size and v_filesize 384 * before we start scanning pages to prevent new pages from being 385 * allocated during the scan. 386 */ 387 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 388 oobjsize = object->size; 389 object->size = nobjsize; 390 391 /* 392 * File has shrunk. Toss any cached pages beyond the new EOF. 393 */ 394 if (nsize < vp->v_filesize) { 395 vp->v_filesize = nsize; 396 if (nobjsize < oobjsize) { 397 vm_object_page_remove(object, nobjsize, oobjsize, 398 FALSE); 399 } 400 /* 401 * This gets rid of garbage at the end of a page that is now 402 * only partially backed by the vnode. Since we are setting 403 * the entire page valid & clean after we are done we have 404 * to be sure that the portion of the page within the file 405 * bounds is already valid. If it isn't then making it 406 * valid would create a corrupt block. 407 */ 408 if (nsize & PAGE_MASK) { 409 vm_offset_t kva; 410 vm_page_t m; 411 412 m = vm_page_lookup_busy_wait(object, OFF_TO_IDX(nsize), 413 TRUE, "vsetsz"); 414 415 if (m && m->valid) { 416 int base = (int)nsize & PAGE_MASK; 417 int size = PAGE_SIZE - base; 418 struct lwbuf *lwb; 419 struct lwbuf lwb_cache; 420 421 /* 422 * Clear out partial-page garbage in case 423 * the page has been mapped. 424 * 425 * This is byte aligned. 426 */ 427 lwb = lwbuf_alloc(m, &lwb_cache); 428 kva = lwbuf_kva(lwb); 429 bzero((caddr_t)kva + base, size); 430 lwbuf_free(lwb); 431 432 /* 433 * XXX work around SMP data integrity race 434 * by unmapping the page from user processes. 435 * The garbage we just cleared may be mapped 436 * to a user process running on another cpu 437 * and this code is not running through normal 438 * I/O channels which handle SMP issues for 439 * us, so unmap page to synchronize all cpus. 440 * 441 * XXX should vm_pager_unmap_page() have 442 * dealt with this? 443 */ 444 vm_page_protect(m, VM_PROT_NONE); 445 446 /* 447 * Clear out partial-page dirty bits. This 448 * has the side effect of setting the valid 449 * bits, but that is ok. There are a bunch 450 * of places in the VM system where we expected 451 * m->dirty == VM_PAGE_BITS_ALL. The file EOF 452 * case is one of them. If the page is still 453 * partially dirty, make it fully dirty. 454 * 455 * NOTE: We do not clear out the valid 456 * bits. This would prevent bogus_page 457 * replacement from working properly. 458 * 459 * NOTE: We do not want to clear the dirty 460 * bit for a partial DEV_BSIZE'd truncation! 461 * This is DEV_BSIZE aligned! 462 */ 463 vm_page_clear_dirty_beg_nonincl(m, base, size); 464 if (m->dirty != 0) 465 m->dirty = VM_PAGE_BITS_ALL; 466 vm_page_wakeup(m); 467 } else if (m) { 468 vm_page_wakeup(m); 469 } 470 } 471 } else { 472 vp->v_filesize = nsize; 473 } 474 vm_object_drop(object); 475 } 476 477 /* 478 * Release a page busied for a getpages operation. The page may have become 479 * wired (typically due to being used by the buffer cache) or otherwise been 480 * soft-busied and cannot be freed in that case. A held page can still be 481 * freed. 482 */ 483 void 484 vnode_pager_freepage(vm_page_t m) 485 { 486 if (m->busy || m->wire_count || (m->flags & PG_NEED_COMMIT)) { 487 vm_page_activate(m); 488 vm_page_wakeup(m); 489 } else { 490 vm_page_free(m); 491 } 492 } 493 494 /* 495 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 496 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to 497 * vnode_pager_generic_getpages() to implement the previous behaviour. 498 * 499 * All other FS's should use the bypass to get to the local media 500 * backing vp's VOP_GETPAGES. 501 */ 502 static int 503 vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess) 504 { 505 int rtval; 506 struct vnode *vp; 507 508 vp = object->handle; 509 rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess); 510 if (rtval == EOPNOTSUPP) 511 panic("vnode_pager: vfs's must implement vop_getpages"); 512 return rtval; 513 } 514 515 /* 516 * This is now called from local media FS's to operate against their 517 * own vnodes if they fail to implement VOP_GETPAGES. 518 * 519 * With all the caching local media devices do these days there is really 520 * very little point to attempting to restrict the I/O size to contiguous 521 * blocks on-disk, especially if our caller thinks we need all the specified 522 * pages. Just construct and issue a READ. 523 */ 524 int 525 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount, 526 int reqpage, int seqaccess) 527 { 528 struct iovec aiov; 529 struct uio auio; 530 off_t foff; 531 int error; 532 int count; 533 int i; 534 int ioflags; 535 536 /* 537 * Do not do anything if the vnode is bad. 538 */ 539 if (vp->v_mount == NULL) 540 return VM_PAGER_BAD; 541 542 /* 543 * Calculate the number of pages. Since we are paging in whole 544 * pages, adjust bytecount to be an integral multiple of the page 545 * size. It will be clipped to the file EOF later on. 546 */ 547 bytecount = round_page(bytecount); 548 count = bytecount / PAGE_SIZE; 549 550 /* 551 * We could check m[reqpage]->valid here and shortcut the operation, 552 * but doing so breaks read-ahead. Instead assume that the VM 553 * system has already done at least the check, don't worry about 554 * any races, and issue the VOP_READ to allow read-ahead to function. 555 * 556 * This keeps the pipeline full for I/O bound sequentially scanned 557 * mmap()'s 558 */ 559 /* don't shortcut */ 560 561 /* 562 * Discard pages past the file EOF. If the requested page is past 563 * the file EOF we just leave its valid bits set to 0, the caller 564 * expects to maintain ownership of the requested page. If the 565 * entire range is past file EOF discard everything and generate 566 * a pagein error. 567 */ 568 foff = IDX_TO_OFF(mpp[0]->pindex); 569 if (foff >= vp->v_filesize) { 570 for (i = 0; i < count; i++) { 571 if (i != reqpage) 572 vnode_pager_freepage(mpp[i]); 573 } 574 return VM_PAGER_ERROR; 575 } 576 577 if (foff + bytecount > vp->v_filesize) { 578 bytecount = vp->v_filesize - foff; 579 i = round_page(bytecount) / PAGE_SIZE; 580 while (count > i) { 581 --count; 582 if (count != reqpage) 583 vnode_pager_freepage(mpp[count]); 584 } 585 } 586 587 /* 588 * The size of the transfer is bytecount. bytecount will be an 589 * integral multiple of the page size unless it has been clipped 590 * to the file EOF. The transfer cannot exceed the file EOF. 591 * 592 * When dealing with real devices we must round-up to the device 593 * sector size. 594 */ 595 if (vp->v_type == VBLK || vp->v_type == VCHR) { 596 int secmask = vp->v_rdev->si_bsize_phys - 1; 597 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large", secmask + 1)); 598 bytecount = (bytecount + secmask) & ~secmask; 599 } 600 601 /* 602 * Severe hack to avoid deadlocks with the buffer cache 603 */ 604 for (i = 0; i < count; ++i) { 605 vm_page_t mt = mpp[i]; 606 607 vm_page_io_start(mt); 608 vm_page_wakeup(mt); 609 } 610 611 /* 612 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE 613 */ 614 ioflags = IO_VMIO; 615 if (seqaccess) 616 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 617 618 aiov.iov_base = NULL; 619 aiov.iov_len = bytecount; 620 auio.uio_iov = &aiov; 621 auio.uio_iovcnt = 1; 622 auio.uio_offset = foff; 623 auio.uio_segflg = UIO_NOCOPY; 624 auio.uio_rw = UIO_READ; 625 auio.uio_resid = bytecount; 626 auio.uio_td = NULL; 627 mycpu->gd_cnt.v_vnodein++; 628 mycpu->gd_cnt.v_vnodepgsin += count; 629 630 error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred); 631 632 /* 633 * Severe hack to avoid deadlocks with the buffer cache 634 */ 635 for (i = 0; i < count; ++i) { 636 vm_page_busy_wait(mpp[i], FALSE, "getpgs"); 637 vm_page_io_finish(mpp[i]); 638 } 639 640 /* 641 * Calculate the actual number of bytes read and clean up the 642 * page list. 643 */ 644 bytecount -= auio.uio_resid; 645 646 for (i = 0; i < count; ++i) { 647 vm_page_t mt = mpp[i]; 648 649 if (i != reqpage) { 650 if (error == 0 && mt->valid) { 651 if (mt->flags & PG_REFERENCED) 652 vm_page_activate(mt); 653 else 654 vm_page_deactivate(mt); 655 vm_page_wakeup(mt); 656 } else { 657 vnode_pager_freepage(mt); 658 } 659 } else if (mt->valid == 0) { 660 if (error == 0) { 661 kprintf("page failed but no I/O error page " 662 "%p object %p pindex %d\n", 663 mt, mt->object, (int) mt->pindex); 664 /* whoops, something happened */ 665 error = EINVAL; 666 } 667 } else if (mt->valid != VM_PAGE_BITS_ALL) { 668 /* 669 * Zero-extend the requested page if necessary (if 670 * the filesystem is using a small block size). 671 */ 672 vm_page_zero_invalid(mt, TRUE); 673 } 674 } 675 if (error) { 676 kprintf("vnode_pager_getpage: I/O read error\n"); 677 } 678 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 679 } 680 681 /* 682 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 683 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 684 * vnode_pager_generic_putpages() to implement the previous behaviour. 685 * 686 * Caller has already cleared the pmap modified bits, if any. 687 * 688 * All other FS's should use the bypass to get to the local media 689 * backing vp's VOP_PUTPAGES. 690 */ 691 static void 692 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count, 693 boolean_t sync, int *rtvals) 694 { 695 int rtval; 696 struct vnode *vp; 697 int bytes = count * PAGE_SIZE; 698 699 /* 700 * Force synchronous operation if we are extremely low on memory 701 * to prevent a low-memory deadlock. VOP operations often need to 702 * allocate more memory to initiate the I/O ( i.e. do a BMAP 703 * operation ). The swapper handles the case by limiting the amount 704 * of asynchronous I/O, but that sort of solution doesn't scale well 705 * for the vnode pager without a lot of work. 706 * 707 * Also, the backing vnode's iodone routine may not wake the pageout 708 * daemon up. This should be probably be addressed XXX. 709 */ 710 711 if ((vmstats.v_free_count + vmstats.v_cache_count) < 712 vmstats.v_pageout_free_min) { 713 sync |= OBJPC_SYNC; 714 } 715 716 /* 717 * Call device-specific putpages function 718 */ 719 vp = object->handle; 720 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 721 if (rtval == EOPNOTSUPP) { 722 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n"); 723 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals); 724 } 725 } 726 727 728 /* 729 * This is now called from local media FS's to operate against their 730 * own vnodes if they fail to implement VOP_PUTPAGES. 731 * 732 * This is typically called indirectly via the pageout daemon and 733 * clustering has already typically occured, so in general we ask the 734 * underlying filesystem to write the data out asynchronously rather 735 * then delayed. 736 */ 737 int 738 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount, 739 int flags, int *rtvals) 740 { 741 int i; 742 vm_object_t object; 743 int maxsize, ncount, count; 744 vm_ooffset_t poffset; 745 struct uio auio; 746 struct iovec aiov; 747 int error; 748 int ioflags; 749 750 object = vp->v_object; 751 count = bytecount / PAGE_SIZE; 752 753 for (i = 0; i < count; i++) 754 rtvals[i] = VM_PAGER_AGAIN; 755 756 if ((int) m[0]->pindex < 0) { 757 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n", 758 (long)m[0]->pindex, m[0]->dirty); 759 rtvals[0] = VM_PAGER_BAD; 760 return VM_PAGER_BAD; 761 } 762 763 maxsize = count * PAGE_SIZE; 764 ncount = count; 765 766 poffset = IDX_TO_OFF(m[0]->pindex); 767 768 /* 769 * If the page-aligned write is larger then the actual file we 770 * have to invalidate pages occuring beyond the file EOF. 771 * 772 * If the file EOF resides in the middle of a page we still clear 773 * all of that page's dirty bits later on. If we didn't it would 774 * endlessly re-write. 775 * 776 * We do not under any circumstances truncate the valid bits, as 777 * this will screw up bogus page replacement. 778 * 779 * The caller has already read-protected the pages. The VFS must 780 * use the buffer cache to wrap the pages. The pages might not 781 * be immediately flushed by the buffer cache but once under its 782 * control the pages themselves can wind up being marked clean 783 * and their covering buffer cache buffer can be marked dirty. 784 */ 785 if (poffset + maxsize > vp->v_filesize) { 786 if (poffset < vp->v_filesize) { 787 maxsize = vp->v_filesize - poffset; 788 ncount = btoc(maxsize); 789 } else { 790 maxsize = 0; 791 ncount = 0; 792 } 793 if (ncount < count) { 794 for (i = ncount; i < count; i++) { 795 rtvals[i] = VM_PAGER_BAD; 796 } 797 } 798 } 799 800 /* 801 * pageouts are already clustered, use IO_ASYNC to force a bawrite() 802 * rather then a bdwrite() to prevent paging I/O from saturating 803 * the buffer cache. Dummy-up the sequential heuristic to cause 804 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 805 * the system decides how to cluster. 806 */ 807 ioflags = IO_VMIO; 808 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 809 ioflags |= IO_SYNC; 810 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 811 ioflags |= IO_ASYNC; 812 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 813 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 814 815 aiov.iov_base = (caddr_t) 0; 816 aiov.iov_len = maxsize; 817 auio.uio_iov = &aiov; 818 auio.uio_iovcnt = 1; 819 auio.uio_offset = poffset; 820 auio.uio_segflg = UIO_NOCOPY; 821 auio.uio_rw = UIO_WRITE; 822 auio.uio_resid = maxsize; 823 auio.uio_td = NULL; 824 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred); 825 mycpu->gd_cnt.v_vnodeout++; 826 mycpu->gd_cnt.v_vnodepgsout += ncount; 827 828 if (error) { 829 krateprintf(&vbadrate, 830 "vnode_pager_putpages: I/O error %d\n", error); 831 } 832 if (auio.uio_resid) { 833 krateprintf(&vresrate, 834 "vnode_pager_putpages: residual I/O %zd at %lu\n", 835 auio.uio_resid, (u_long)m[0]->pindex); 836 } 837 if (error == 0) { 838 for (i = 0; i < ncount; i++) { 839 rtvals[i] = VM_PAGER_OK; 840 vm_page_undirty(m[i]); 841 } 842 } 843 return rtvals[0]; 844 } 845 846 /* 847 * Run the chain and if the bottom-most object is a vnode-type lock the 848 * underlying vnode. A locked vnode or NULL is returned. 849 */ 850 struct vnode * 851 vnode_pager_lock(vm_object_t object) 852 { 853 struct vnode *vp = NULL; 854 vm_object_t lobject; 855 vm_object_t tobject; 856 int error; 857 858 if (object == NULL) 859 return(NULL); 860 861 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 862 lobject = object; 863 864 while (lobject->type != OBJT_VNODE) { 865 if (lobject->flags & OBJ_DEAD) 866 break; 867 tobject = lobject->backing_object; 868 if (tobject == NULL) 869 break; 870 vm_object_hold(tobject); 871 if (tobject == lobject->backing_object) { 872 if (lobject != object) { 873 vm_object_lock_swap(); 874 vm_object_drop(lobject); 875 } 876 lobject = tobject; 877 } else { 878 vm_object_drop(tobject); 879 } 880 } 881 while (lobject->type == OBJT_VNODE && 882 (lobject->flags & OBJ_DEAD) == 0) { 883 /* 884 * Extract the vp 885 */ 886 vp = lobject->handle; 887 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE); 888 if (error == 0) { 889 if (lobject->handle == vp) 890 break; 891 vput(vp); 892 } else { 893 kprintf("vnode_pager_lock: vp %p error %d " 894 "lockstatus %d, retrying\n", 895 vp, error, 896 lockstatus(&vp->v_lock, curthread)); 897 tsleep(object->handle, 0, "vnpgrl", hz); 898 } 899 vp = NULL; 900 } 901 if (lobject != object) 902 vm_object_drop(lobject); 903 return (vp); 904 } 905