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 * $DragonFly: src/sys/vm/vnode_pager.c,v 1.43 2008/06/19 23:27:39 dillon Exp $ 45 */ 46 47 /* 48 * Page to/from files (vnodes). 49 */ 50 51 /* 52 * TODO: 53 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will 54 * greatly re-simplify the vnode_pager. 55 */ 56 57 #include <sys/param.h> 58 #include <sys/systm.h> 59 #include <sys/kernel.h> 60 #include <sys/proc.h> 61 #include <sys/vnode.h> 62 #include <sys/mount.h> 63 #include <sys/buf.h> 64 #include <sys/vmmeter.h> 65 #include <sys/conf.h> 66 67 #include <cpu/lwbuf.h> 68 69 #include <vm/vm.h> 70 #include <vm/vm_object.h> 71 #include <vm/vm_page.h> 72 #include <vm/vm_pager.h> 73 #include <vm/vm_map.h> 74 #include <vm/vnode_pager.h> 75 #include <vm/swap_pager.h> 76 #include <vm/vm_extern.h> 77 78 #include <sys/thread2.h> 79 #include <vm/vm_page2.h> 80 81 static void vnode_pager_dealloc (vm_object_t); 82 static int vnode_pager_getpage (vm_object_t, vm_page_t *, int); 83 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *); 84 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t); 85 86 struct pagerops vnodepagerops = { 87 vnode_pager_dealloc, 88 vnode_pager_getpage, 89 vnode_pager_putpages, 90 vnode_pager_haspage 91 }; 92 93 static struct krate vbadrate = { 1 }; 94 static struct krate vresrate = { 1 }; 95 96 int vnode_pbuf_freecnt = -1; /* start out unlimited */ 97 98 /* 99 * Allocate a VM object for a vnode, typically a regular file vnode. 100 * 101 * Some additional information is required to generate a properly sized 102 * object which covers the entire buffer cache buffer straddling the file 103 * EOF. Userland does not see the extra pages as the VM fault code tests 104 * against v_filesize. 105 */ 106 vm_object_t 107 vnode_pager_alloc(void *handle, off_t length, vm_prot_t prot, off_t offset, 108 int blksize, int boff) 109 { 110 vm_object_t object; 111 struct vnode *vp; 112 off_t loffset; 113 vm_pindex_t lsize; 114 115 /* 116 * Pageout to vnode, no can do yet. 117 */ 118 if (handle == NULL) 119 return (NULL); 120 121 /* 122 * XXX hack - This initialization should be put somewhere else. 123 */ 124 if (vnode_pbuf_freecnt < 0) { 125 vnode_pbuf_freecnt = nswbuf / 2 + 1; 126 } 127 128 /* 129 * Serialize potential vnode/object teardowns and interlocks 130 */ 131 vp = (struct vnode *)handle; 132 lwkt_gettoken(&vp->v_token); 133 134 /* 135 * Prevent race condition when allocating the object. This 136 * can happen with NFS vnodes since the nfsnode isn't locked. 137 */ 138 while (vp->v_flag & VOLOCK) { 139 vsetflags(vp, VOWANT); 140 tsleep(vp, 0, "vnpobj", 0); 141 } 142 vsetflags(vp, VOLOCK); 143 lwkt_reltoken(&vp->v_token); 144 145 /* 146 * If the object is being terminated, wait for it to 147 * go away. 148 */ 149 while ((object = vp->v_object) != NULL) { 150 vm_object_hold(object); 151 if ((object->flags & OBJ_DEAD) == 0) 152 break; 153 vm_object_dead_sleep(object, "vadead"); 154 vm_object_drop(object); 155 } 156 157 if (vp->v_sysref.refcnt <= 0) 158 panic("vnode_pager_alloc: no vnode reference"); 159 160 /* 161 * Round up to the *next* block, then destroy the buffers in question. 162 * Since we are only removing some of the buffers we must rely on the 163 * scan count to determine whether a loop is necessary. 164 * 165 * Destroy any pages beyond the last buffer. 166 */ 167 if (boff < 0) 168 boff = (int)(length % blksize); 169 if (boff) 170 loffset = length + (blksize - boff); 171 else 172 loffset = length; 173 lsize = OFF_TO_IDX(round_page64(loffset)); 174 175 if (object == NULL) { 176 /* 177 * And an object of the appropriate size 178 */ 179 object = vm_object_allocate(OBJT_VNODE, lsize); 180 vm_object_hold(object); 181 object->flags = 0; 182 object->handle = handle; 183 vp->v_object = object; 184 vp->v_filesize = length; 185 if (vp->v_mount && (vp->v_mount->mnt_kern_flag & MNTK_NOMSYNC)) 186 vm_object_set_flag(object, OBJ_NOMSYNC); 187 } else { 188 object->ref_count++; 189 if (object->size != lsize) { 190 kprintf("vnode_pager_alloc: Warning, objsize " 191 "mismatch %jd/%jd vp=%p obj=%p\n", 192 (intmax_t)object->size, 193 (intmax_t)lsize, 194 vp, object); 195 } 196 if (vp->v_filesize != length) { 197 kprintf("vnode_pager_alloc: Warning, filesize " 198 "mismatch %jd/%jd vp=%p obj=%p\n", 199 (intmax_t)vp->v_filesize, 200 (intmax_t)length, 201 vp, object); 202 } 203 } 204 205 vref(vp); 206 lwkt_gettoken(&vp->v_token); 207 vclrflags(vp, VOLOCK); 208 if (vp->v_flag & VOWANT) { 209 vclrflags(vp, VOWANT); 210 wakeup(vp); 211 } 212 lwkt_reltoken(&vp->v_token); 213 214 vm_object_drop(object); 215 216 return (object); 217 } 218 219 /* 220 * Add a ref to a vnode's existing VM object, return the object or 221 * NULL if the vnode did not have one. This does not create the 222 * object (we can't since we don't know what the proper blocksize/boff 223 * is to match the VFS's use of the buffer cache). 224 */ 225 vm_object_t 226 vnode_pager_reference(struct vnode *vp) 227 { 228 vm_object_t object; 229 230 /* 231 * Prevent race condition when allocating the object. This 232 * can happen with NFS vnodes since the nfsnode isn't locked. 233 * 234 * Serialize potential vnode/object teardowns and interlocks 235 */ 236 lwkt_gettoken(&vp->v_token); 237 while (vp->v_flag & VOLOCK) { 238 vsetflags(vp, VOWANT); 239 tsleep(vp, 0, "vnpobj", 0); 240 } 241 vsetflags(vp, VOLOCK); 242 lwkt_reltoken(&vp->v_token); 243 244 /* 245 * Prevent race conditions against deallocation of the VM 246 * object. 247 */ 248 while ((object = vp->v_object) != NULL) { 249 vm_object_hold(object); 250 if ((object->flags & OBJ_DEAD) == 0) 251 break; 252 vm_object_dead_sleep(object, "vadead"); 253 vm_object_drop(object); 254 } 255 256 /* 257 * The object is expected to exist, the caller will handle 258 * NULL returns if it does not. 259 */ 260 if (object) { 261 object->ref_count++; 262 vref(vp); 263 } 264 265 lwkt_gettoken(&vp->v_token); 266 vclrflags(vp, VOLOCK); 267 if (vp->v_flag & VOWANT) { 268 vclrflags(vp, VOWANT); 269 wakeup(vp); 270 } 271 lwkt_reltoken(&vp->v_token); 272 if (object) 273 vm_object_drop(object); 274 275 return (object); 276 } 277 278 static void 279 vnode_pager_dealloc(vm_object_t object) 280 { 281 struct vnode *vp = object->handle; 282 283 if (vp == NULL) 284 panic("vnode_pager_dealloc: pager already dealloced"); 285 286 vm_object_pip_wait(object, "vnpdea"); 287 288 object->handle = NULL; 289 object->type = OBJT_DEAD; 290 vp->v_object = NULL; 291 vp->v_filesize = NOOFFSET; 292 vclrflags(vp, VTEXT | VOBJBUF); 293 swap_pager_freespace_all(object); 294 } 295 296 /* 297 * Return whether the vnode pager has the requested page. Return the 298 * number of disk-contiguous pages before and after the requested page, 299 * not including the requested page. 300 */ 301 static boolean_t 302 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex) 303 { 304 struct vnode *vp = object->handle; 305 off_t loffset; 306 off_t doffset; 307 int voff; 308 int bsize; 309 int error; 310 311 /* 312 * If no vp or vp is doomed or marked transparent to VM, we do not 313 * have the page. 314 */ 315 if ((vp == NULL) || (vp->v_flag & VRECLAIMED)) 316 return FALSE; 317 318 /* 319 * If filesystem no longer mounted or offset beyond end of file we do 320 * not have the page. 321 */ 322 loffset = IDX_TO_OFF(pindex); 323 324 if (vp->v_mount == NULL || loffset >= vp->v_filesize) 325 return FALSE; 326 327 bsize = vp->v_mount->mnt_stat.f_iosize; 328 voff = loffset % bsize; 329 330 /* 331 * XXX 332 * 333 * BMAP returns byte counts before and after, where after 334 * is inclusive of the base page. haspage must return page 335 * counts before and after where after does not include the 336 * base page. 337 * 338 * BMAP is allowed to return a *after of 0 for backwards 339 * compatibility. The base page is still considered valid if 340 * no error is returned. 341 */ 342 error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0); 343 if (error) 344 return TRUE; 345 if (doffset == NOOFFSET) 346 return FALSE; 347 return TRUE; 348 } 349 350 /* 351 * Lets the VM system know about a change in size for a file. 352 * We adjust our own internal size and flush any cached pages in 353 * the associated object that are affected by the size change. 354 * 355 * NOTE: This routine may be invoked as a result of a pager put 356 * operation (possibly at object termination time), so we must be careful. 357 * 358 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that 359 * we do not blow up on the case. nsize will always be >= 0, however. 360 */ 361 void 362 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize) 363 { 364 vm_pindex_t nobjsize; 365 vm_pindex_t oobjsize; 366 vm_object_t object; 367 368 while ((object = vp->v_object) != NULL) { 369 vm_object_hold(object); 370 if (vp->v_object == object) 371 break; 372 vm_object_drop(object); 373 } 374 if (object == NULL) 375 return; 376 377 /* 378 * Hasn't changed size 379 */ 380 if (nsize == vp->v_filesize) { 381 vm_object_drop(object); 382 return; 383 } 384 385 /* 386 * Has changed size. Adjust the VM object's size and v_filesize 387 * before we start scanning pages to prevent new pages from being 388 * allocated during the scan. 389 */ 390 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 391 oobjsize = object->size; 392 object->size = nobjsize; 393 394 /* 395 * File has shrunk. Toss any cached pages beyond the new EOF. 396 */ 397 if (nsize < vp->v_filesize) { 398 vp->v_filesize = nsize; 399 if (nobjsize < oobjsize) { 400 vm_object_page_remove(object, nobjsize, oobjsize, 401 FALSE); 402 } 403 /* 404 * This gets rid of garbage at the end of a page that is now 405 * only partially backed by the vnode. Since we are setting 406 * the entire page valid & clean after we are done we have 407 * to be sure that the portion of the page within the file 408 * bounds is already valid. If it isn't then making it 409 * valid would create a corrupt block. 410 */ 411 if (nsize & PAGE_MASK) { 412 vm_offset_t kva; 413 vm_page_t m; 414 415 m = vm_page_lookup_busy_wait(object, OFF_TO_IDX(nsize), 416 TRUE, "vsetsz"); 417 418 if (m && m->valid) { 419 int base = (int)nsize & PAGE_MASK; 420 int size = PAGE_SIZE - base; 421 struct lwbuf *lwb; 422 struct lwbuf lwb_cache; 423 424 /* 425 * Clear out partial-page garbage in case 426 * the page has been mapped. 427 * 428 * This is byte aligned. 429 */ 430 lwb = lwbuf_alloc(m, &lwb_cache); 431 kva = lwbuf_kva(lwb); 432 bzero((caddr_t)kva + base, size); 433 lwbuf_free(lwb); 434 435 /* 436 * XXX work around SMP data integrity race 437 * by unmapping the page from user processes. 438 * The garbage we just cleared may be mapped 439 * to a user process running on another cpu 440 * and this code is not running through normal 441 * I/O channels which handle SMP issues for 442 * us, so unmap page to synchronize all cpus. 443 * 444 * XXX should vm_pager_unmap_page() have 445 * dealt with this? 446 */ 447 vm_page_protect(m, VM_PROT_NONE); 448 449 /* 450 * Clear out partial-page dirty bits. This 451 * has the side effect of setting the valid 452 * bits, but that is ok. There are a bunch 453 * of places in the VM system where we expected 454 * m->dirty == VM_PAGE_BITS_ALL. The file EOF 455 * case is one of them. If the page is still 456 * partially dirty, make it fully dirty. 457 * 458 * NOTE: We do not clear out the valid 459 * bits. This would prevent bogus_page 460 * replacement from working properly. 461 * 462 * NOTE: We do not want to clear the dirty 463 * bit for a partial DEV_BSIZE'd truncation! 464 * This is DEV_BSIZE aligned! 465 */ 466 vm_page_clear_dirty_beg_nonincl(m, base, size); 467 if (m->dirty != 0) 468 m->dirty = VM_PAGE_BITS_ALL; 469 vm_page_wakeup(m); 470 } else if (m) { 471 vm_page_wakeup(m); 472 } 473 } 474 } else { 475 vp->v_filesize = nsize; 476 } 477 vm_object_drop(object); 478 } 479 480 /* 481 * Release a page busied for a getpages operation. The page may have become 482 * wired (typically due to being used by the buffer cache) or otherwise been 483 * soft-busied and cannot be freed in that case. A held page can still be 484 * freed. 485 */ 486 void 487 vnode_pager_freepage(vm_page_t m) 488 { 489 if (m->busy || m->wire_count) { 490 vm_page_activate(m); 491 vm_page_wakeup(m); 492 } else { 493 vm_page_free(m); 494 } 495 } 496 497 /* 498 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 499 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to 500 * vnode_pager_generic_getpages() to implement the previous behaviour. 501 * 502 * All other FS's should use the bypass to get to the local media 503 * backing vp's VOP_GETPAGES. 504 */ 505 static int 506 vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess) 507 { 508 int rtval; 509 struct vnode *vp; 510 511 vp = object->handle; 512 rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess); 513 if (rtval == EOPNOTSUPP) 514 panic("vnode_pager: vfs's must implement vop_getpages\n"); 515 return rtval; 516 } 517 518 /* 519 * This is now called from local media FS's to operate against their 520 * own vnodes if they fail to implement VOP_GETPAGES. 521 * 522 * With all the caching local media devices do these days there is really 523 * very little point to attempting to restrict the I/O size to contiguous 524 * blocks on-disk, especially if our caller thinks we need all the specified 525 * pages. Just construct and issue a READ. 526 */ 527 int 528 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount, 529 int reqpage, int seqaccess) 530 { 531 struct iovec aiov; 532 struct uio auio; 533 off_t foff; 534 int error; 535 int count; 536 int i; 537 int ioflags; 538 539 /* 540 * Do not do anything if the vnode is bad. 541 */ 542 if (vp->v_mount == NULL) 543 return VM_PAGER_BAD; 544 545 /* 546 * Calculate the number of pages. Since we are paging in whole 547 * pages, adjust bytecount to be an integral multiple of the page 548 * size. It will be clipped to the file EOF later on. 549 */ 550 bytecount = round_page(bytecount); 551 count = bytecount / PAGE_SIZE; 552 553 /* 554 * We could check m[reqpage]->valid here and shortcut the operation, 555 * but doing so breaks read-ahead. Instead assume that the VM 556 * system has already done at least the check, don't worry about 557 * any races, and issue the VOP_READ to allow read-ahead to function. 558 * 559 * This keeps the pipeline full for I/O bound sequentially scanned 560 * mmap()'s 561 */ 562 /* don't shortcut */ 563 564 /* 565 * Discard pages past the file EOF. If the requested page is past 566 * the file EOF we just leave its valid bits set to 0, the caller 567 * expects to maintain ownership of the requested page. If the 568 * entire range is past file EOF discard everything and generate 569 * a pagein error. 570 */ 571 foff = IDX_TO_OFF(mpp[0]->pindex); 572 if (foff >= vp->v_filesize) { 573 for (i = 0; i < count; i++) { 574 if (i != reqpage) 575 vnode_pager_freepage(mpp[i]); 576 } 577 return VM_PAGER_ERROR; 578 } 579 580 if (foff + bytecount > vp->v_filesize) { 581 bytecount = vp->v_filesize - foff; 582 i = round_page(bytecount) / PAGE_SIZE; 583 while (count > i) { 584 --count; 585 if (count != reqpage) 586 vnode_pager_freepage(mpp[count]); 587 } 588 } 589 590 /* 591 * The size of the transfer is bytecount. bytecount will be an 592 * integral multiple of the page size unless it has been clipped 593 * to the file EOF. The transfer cannot exceed the file EOF. 594 * 595 * When dealing with real devices we must round-up to the device 596 * sector size. 597 */ 598 if (vp->v_type == VBLK || vp->v_type == VCHR) { 599 int secmask = vp->v_rdev->si_bsize_phys - 1; 600 KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1)); 601 bytecount = (bytecount + secmask) & ~secmask; 602 } 603 604 /* 605 * Severe hack to avoid deadlocks with the buffer cache 606 */ 607 for (i = 0; i < count; ++i) { 608 vm_page_t mt = mpp[i]; 609 610 vm_page_io_start(mt); 611 vm_page_wakeup(mt); 612 } 613 614 /* 615 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE 616 */ 617 ioflags = IO_VMIO; 618 if (seqaccess) 619 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 620 621 aiov.iov_base = NULL; 622 aiov.iov_len = bytecount; 623 auio.uio_iov = &aiov; 624 auio.uio_iovcnt = 1; 625 auio.uio_offset = foff; 626 auio.uio_segflg = UIO_NOCOPY; 627 auio.uio_rw = UIO_READ; 628 auio.uio_resid = bytecount; 629 auio.uio_td = NULL; 630 mycpu->gd_cnt.v_vnodein++; 631 mycpu->gd_cnt.v_vnodepgsin += count; 632 633 error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred); 634 635 /* 636 * Severe hack to avoid deadlocks with the buffer cache 637 */ 638 for (i = 0; i < count; ++i) { 639 vm_page_busy_wait(mpp[i], FALSE, "getpgs"); 640 vm_page_io_finish(mpp[i]); 641 } 642 643 /* 644 * Calculate the actual number of bytes read and clean up the 645 * page list. 646 */ 647 bytecount -= auio.uio_resid; 648 649 for (i = 0; i < count; ++i) { 650 vm_page_t mt = mpp[i]; 651 652 if (i != reqpage) { 653 if (error == 0 && mt->valid) { 654 if (mt->flags & PG_REFERENCED) 655 vm_page_activate(mt); 656 else 657 vm_page_deactivate(mt); 658 vm_page_wakeup(mt); 659 } else { 660 vnode_pager_freepage(mt); 661 } 662 } else if (mt->valid == 0) { 663 if (error == 0) { 664 kprintf("page failed but no I/O error page " 665 "%p object %p pindex %d\n", 666 mt, mt->object, (int) mt->pindex); 667 /* whoops, something happened */ 668 error = EINVAL; 669 } 670 } else if (mt->valid != VM_PAGE_BITS_ALL) { 671 /* 672 * Zero-extend the requested page if necessary (if 673 * the filesystem is using a small block size). 674 */ 675 vm_page_zero_invalid(mt, TRUE); 676 } 677 } 678 if (error) { 679 kprintf("vnode_pager_getpage: I/O read error\n"); 680 } 681 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 682 } 683 684 /* 685 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 686 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 687 * vnode_pager_generic_putpages() to implement the previous behaviour. 688 * 689 * Caller has already cleared the pmap modified bits, if any. 690 * 691 * All other FS's should use the bypass to get to the local media 692 * backing vp's VOP_PUTPAGES. 693 */ 694 static void 695 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count, 696 boolean_t sync, int *rtvals) 697 { 698 int rtval; 699 struct vnode *vp; 700 int bytes = count * PAGE_SIZE; 701 702 /* 703 * Force synchronous operation if we are extremely low on memory 704 * to prevent a low-memory deadlock. VOP operations often need to 705 * allocate more memory to initiate the I/O ( i.e. do a BMAP 706 * operation ). The swapper handles the case by limiting the amount 707 * of asynchronous I/O, but that sort of solution doesn't scale well 708 * for the vnode pager without a lot of work. 709 * 710 * Also, the backing vnode's iodone routine may not wake the pageout 711 * daemon up. This should be probably be addressed XXX. 712 */ 713 714 if ((vmstats.v_free_count + vmstats.v_cache_count) < 715 vmstats.v_pageout_free_min) { 716 sync |= OBJPC_SYNC; 717 } 718 719 /* 720 * Call device-specific putpages function 721 */ 722 vp = object->handle; 723 rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); 724 if (rtval == EOPNOTSUPP) { 725 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n"); 726 rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals); 727 } 728 } 729 730 731 /* 732 * This is now called from local media FS's to operate against their 733 * own vnodes if they fail to implement VOP_PUTPAGES. 734 * 735 * This is typically called indirectly via the pageout daemon and 736 * clustering has already typically occured, so in general we ask the 737 * underlying filesystem to write the data out asynchronously rather 738 * then delayed. 739 */ 740 int 741 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount, 742 int flags, int *rtvals) 743 { 744 int i; 745 vm_object_t object; 746 int maxsize, ncount, count; 747 vm_ooffset_t poffset; 748 struct uio auio; 749 struct iovec aiov; 750 int error; 751 int ioflags; 752 753 object = vp->v_object; 754 count = bytecount / PAGE_SIZE; 755 756 for (i = 0; i < count; i++) 757 rtvals[i] = VM_PAGER_AGAIN; 758 759 if ((int) m[0]->pindex < 0) { 760 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n", 761 (long)m[0]->pindex, m[0]->dirty); 762 rtvals[0] = VM_PAGER_BAD; 763 return VM_PAGER_BAD; 764 } 765 766 maxsize = count * PAGE_SIZE; 767 ncount = count; 768 769 poffset = IDX_TO_OFF(m[0]->pindex); 770 771 /* 772 * If the page-aligned write is larger then the actual file we 773 * have to invalidate pages occuring beyond the file EOF. 774 * 775 * If the file EOF resides in the middle of a page we still clear 776 * all of that page's dirty bits later on. If we didn't it would 777 * endlessly re-write. 778 * 779 * We do not under any circumstances truncate the valid bits, as 780 * this will screw up bogus page replacement. 781 * 782 * The caller has already read-protected the pages. The VFS must 783 * use the buffer cache to wrap the pages. The pages might not 784 * be immediately flushed by the buffer cache but once under its 785 * control the pages themselves can wind up being marked clean 786 * and their covering buffer cache buffer can be marked dirty. 787 */ 788 if (poffset + maxsize > vp->v_filesize) { 789 if (poffset < vp->v_filesize) { 790 maxsize = vp->v_filesize - poffset; 791 ncount = btoc(maxsize); 792 } else { 793 maxsize = 0; 794 ncount = 0; 795 } 796 if (ncount < count) { 797 for (i = ncount; i < count; i++) { 798 rtvals[i] = VM_PAGER_BAD; 799 } 800 } 801 } 802 803 /* 804 * pageouts are already clustered, use IO_ASYNC to force a bawrite() 805 * rather then a bdwrite() to prevent paging I/O from saturating 806 * the buffer cache. Dummy-up the sequential heuristic to cause 807 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 808 * the system decides how to cluster. 809 */ 810 ioflags = IO_VMIO; 811 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 812 ioflags |= IO_SYNC; 813 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 814 ioflags |= IO_ASYNC; 815 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 816 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 817 818 aiov.iov_base = (caddr_t) 0; 819 aiov.iov_len = maxsize; 820 auio.uio_iov = &aiov; 821 auio.uio_iovcnt = 1; 822 auio.uio_offset = poffset; 823 auio.uio_segflg = UIO_NOCOPY; 824 auio.uio_rw = UIO_WRITE; 825 auio.uio_resid = maxsize; 826 auio.uio_td = NULL; 827 error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred); 828 mycpu->gd_cnt.v_vnodeout++; 829 mycpu->gd_cnt.v_vnodepgsout += ncount; 830 831 if (error) { 832 krateprintf(&vbadrate, 833 "vnode_pager_putpages: I/O error %d\n", error); 834 } 835 if (auio.uio_resid) { 836 krateprintf(&vresrate, 837 "vnode_pager_putpages: residual I/O %zd at %lu\n", 838 auio.uio_resid, (u_long)m[0]->pindex); 839 } 840 if (error == 0) { 841 for (i = 0; i < ncount; i++) { 842 rtvals[i] = VM_PAGER_OK; 843 vm_page_undirty(m[i]); 844 } 845 } 846 return rtvals[0]; 847 } 848 849 /* 850 * Run the chain and if the bottom-most object is a vnode-type lock the 851 * underlying vnode. A locked vnode or NULL is returned. 852 */ 853 struct vnode * 854 vnode_pager_lock(vm_object_t object) 855 { 856 struct vnode *vp = NULL; 857 vm_object_t lobject; 858 vm_object_t tobject; 859 int error; 860 861 if (object == NULL) 862 return(NULL); 863 864 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 865 lobject = object; 866 867 while (lobject->type != OBJT_VNODE) { 868 if (lobject->flags & OBJ_DEAD) 869 break; 870 tobject = lobject->backing_object; 871 if (tobject == NULL) 872 break; 873 vm_object_hold(tobject); 874 if (tobject == lobject->backing_object) { 875 if (lobject != object) { 876 vm_object_lock_swap(); 877 vm_object_drop(lobject); 878 } 879 lobject = tobject; 880 } else { 881 vm_object_drop(tobject); 882 } 883 } 884 while (lobject->type == OBJT_VNODE && 885 (lobject->flags & OBJ_DEAD) == 0) { 886 /* 887 * Extract the vp 888 */ 889 vp = lobject->handle; 890 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE); 891 if (error == 0) { 892 if (lobject->handle == vp) 893 break; 894 vput(vp); 895 } else { 896 kprintf("vnode_pager_lock: vp %p error %d " 897 "lockstatus %d, retrying\n", 898 vp, error, 899 lockstatus(&vp->v_lock, curthread)); 900 tsleep(object->handle, 0, "vnpgrl", hz); 901 } 902 vp = NULL; 903 } 904 if (lobject != object) 905 vm_object_drop(lobject); 906 return (vp); 907 } 908