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