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 */ 42 43/* 44 * Page to/from files (vnodes). 45 */ 46 47/* 48 * TODO: 49 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will 50 * greatly re-simplify the vnode_pager. 51 */ 52 53#include <sys/cdefs.h> 54__FBSDID("$FreeBSD: releng/10.3/sys/vm/vnode_pager.c 291454 2015-11-29 14:44:40Z kib $"); 55 56#include <sys/param.h> 57#include <sys/systm.h> 58#include <sys/proc.h> 59#include <sys/vnode.h> 60#include <sys/mount.h> 61#include <sys/bio.h> 62#include <sys/buf.h> 63#include <sys/vmmeter.h> 64#include <sys/limits.h> 65#include <sys/conf.h> 66#include <sys/rwlock.h> 67#include <sys/sf_buf.h> 68 69#include <machine/atomic.h> 70 71#include <vm/vm.h> 72#include <vm/vm_param.h> 73#include <vm/vm_object.h> 74#include <vm/vm_page.h> 75#include <vm/vm_pager.h> 76#include <vm/vm_map.h> 77#include <vm/vnode_pager.h> 78#include <vm/vm_extern.h> 79 80static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, 81 daddr_t *rtaddress, int *run); 82static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m); 83static int vnode_pager_input_old(vm_object_t object, vm_page_t m); 84static void vnode_pager_dealloc(vm_object_t); 85static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int); 86static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *); 87static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *); 88static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t, 89 vm_ooffset_t, struct ucred *cred); 90 91struct pagerops vnodepagerops = { 92 .pgo_alloc = vnode_pager_alloc, 93 .pgo_dealloc = vnode_pager_dealloc, 94 .pgo_getpages = vnode_pager_getpages, 95 .pgo_putpages = vnode_pager_putpages, 96 .pgo_haspage = vnode_pager_haspage, 97}; 98 99int vnode_pbuf_freecnt; 100 101/* Create the VM system backing object for this vnode */ 102int 103vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td) 104{ 105 vm_object_t object; 106 vm_ooffset_t size = isize; 107 struct vattr va; 108 109 if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE) 110 return (0); 111 112 while ((object = vp->v_object) != NULL) { 113 VM_OBJECT_WLOCK(object); 114 if (!(object->flags & OBJ_DEAD)) { 115 VM_OBJECT_WUNLOCK(object); 116 return (0); 117 } 118 VOP_UNLOCK(vp, 0); 119 vm_object_set_flag(object, OBJ_DISCONNECTWNT); 120 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vodead", 0); 121 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 122 } 123 124 if (size == 0) { 125 if (vn_isdisk(vp, NULL)) { 126 size = IDX_TO_OFF(INT_MAX); 127 } else { 128 if (VOP_GETATTR(vp, &va, td->td_ucred)) 129 return (0); 130 size = va.va_size; 131 } 132 } 133 134 object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred); 135 /* 136 * Dereference the reference we just created. This assumes 137 * that the object is associated with the vp. 138 */ 139 VM_OBJECT_WLOCK(object); 140 object->ref_count--; 141 VM_OBJECT_WUNLOCK(object); 142 vrele(vp); 143 144 KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object")); 145 146 return (0); 147} 148 149void 150vnode_destroy_vobject(struct vnode *vp) 151{ 152 struct vm_object *obj; 153 154 obj = vp->v_object; 155 if (obj == NULL) 156 return; 157 ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject"); 158 VM_OBJECT_WLOCK(obj); 159 if (obj->ref_count == 0) { 160 /* 161 * don't double-terminate the object 162 */ 163 if ((obj->flags & OBJ_DEAD) == 0) 164 vm_object_terminate(obj); 165 else 166 VM_OBJECT_WUNLOCK(obj); 167 } else { 168 /* 169 * Woe to the process that tries to page now :-). 170 */ 171 vm_pager_deallocate(obj); 172 VM_OBJECT_WUNLOCK(obj); 173 } 174 vp->v_object = NULL; 175} 176 177 178/* 179 * Allocate (or lookup) pager for a vnode. 180 * Handle is a vnode pointer. 181 * 182 * MPSAFE 183 */ 184vm_object_t 185vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 186 vm_ooffset_t offset, struct ucred *cred) 187{ 188 vm_object_t object; 189 struct vnode *vp; 190 191 /* 192 * Pageout to vnode, no can do yet. 193 */ 194 if (handle == NULL) 195 return (NULL); 196 197 vp = (struct vnode *) handle; 198 199 /* 200 * If the object is being terminated, wait for it to 201 * go away. 202 */ 203retry: 204 while ((object = vp->v_object) != NULL) { 205 VM_OBJECT_WLOCK(object); 206 if ((object->flags & OBJ_DEAD) == 0) 207 break; 208 vm_object_set_flag(object, OBJ_DISCONNECTWNT); 209 VM_OBJECT_SLEEP(object, object, PDROP | PVM, "vadead", 0); 210 } 211 212 KASSERT(vp->v_usecount != 0, ("vnode_pager_alloc: no vnode reference")); 213 214 if (object == NULL) { 215 /* 216 * Add an object of the appropriate size 217 */ 218 object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); 219 220 object->un_pager.vnp.vnp_size = size; 221 object->un_pager.vnp.writemappings = 0; 222 223 object->handle = handle; 224 VI_LOCK(vp); 225 if (vp->v_object != NULL) { 226 /* 227 * Object has been created while we were sleeping 228 */ 229 VI_UNLOCK(vp); 230 VM_OBJECT_WLOCK(object); 231 KASSERT(object->ref_count == 1, 232 ("leaked ref %p %d", object, object->ref_count)); 233 object->type = OBJT_DEAD; 234 object->ref_count = 0; 235 VM_OBJECT_WUNLOCK(object); 236 vm_object_destroy(object); 237 goto retry; 238 } 239 vp->v_object = object; 240 VI_UNLOCK(vp); 241 } else { 242 object->ref_count++; 243 VM_OBJECT_WUNLOCK(object); 244 } 245 vref(vp); 246 return (object); 247} 248 249/* 250 * The object must be locked. 251 */ 252static void 253vnode_pager_dealloc(object) 254 vm_object_t object; 255{ 256 struct vnode *vp; 257 int refs; 258 259 vp = object->handle; 260 if (vp == NULL) 261 panic("vnode_pager_dealloc: pager already dealloced"); 262 263 VM_OBJECT_ASSERT_WLOCKED(object); 264 vm_object_pip_wait(object, "vnpdea"); 265 refs = object->ref_count; 266 267 object->handle = NULL; 268 object->type = OBJT_DEAD; 269 if (object->flags & OBJ_DISCONNECTWNT) { 270 vm_object_clear_flag(object, OBJ_DISCONNECTWNT); 271 wakeup(object); 272 } 273 ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc"); 274 if (object->un_pager.vnp.writemappings > 0) { 275 object->un_pager.vnp.writemappings = 0; 276 VOP_ADD_WRITECOUNT(vp, -1); 277 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 278 __func__, vp, vp->v_writecount); 279 } 280 vp->v_object = NULL; 281 VOP_UNSET_TEXT(vp); 282 VM_OBJECT_WUNLOCK(object); 283 while (refs-- > 0) 284 vunref(vp); 285 VM_OBJECT_WLOCK(object); 286} 287 288static boolean_t 289vnode_pager_haspage(object, pindex, before, after) 290 vm_object_t object; 291 vm_pindex_t pindex; 292 int *before; 293 int *after; 294{ 295 struct vnode *vp = object->handle; 296 daddr_t bn; 297 int err; 298 daddr_t reqblock; 299 int poff; 300 int bsize; 301 int pagesperblock, blocksperpage; 302 303 VM_OBJECT_ASSERT_WLOCKED(object); 304 /* 305 * If no vp or vp is doomed or marked transparent to VM, we do not 306 * have the page. 307 */ 308 if (vp == NULL || vp->v_iflag & VI_DOOMED) 309 return FALSE; 310 /* 311 * If the offset is beyond end of file we do 312 * not have the page. 313 */ 314 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size) 315 return FALSE; 316 317 bsize = vp->v_mount->mnt_stat.f_iosize; 318 pagesperblock = bsize / PAGE_SIZE; 319 blocksperpage = 0; 320 if (pagesperblock > 0) { 321 reqblock = pindex / pagesperblock; 322 } else { 323 blocksperpage = (PAGE_SIZE / bsize); 324 reqblock = pindex * blocksperpage; 325 } 326 VM_OBJECT_WUNLOCK(object); 327 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before); 328 VM_OBJECT_WLOCK(object); 329 if (err) 330 return TRUE; 331 if (bn == -1) 332 return FALSE; 333 if (pagesperblock > 0) { 334 poff = pindex - (reqblock * pagesperblock); 335 if (before) { 336 *before *= pagesperblock; 337 *before += poff; 338 } 339 if (after) { 340 int numafter; 341 *after *= pagesperblock; 342 numafter = pagesperblock - (poff + 1); 343 if (IDX_TO_OFF(pindex + numafter) > 344 object->un_pager.vnp.vnp_size) { 345 numafter = 346 OFF_TO_IDX(object->un_pager.vnp.vnp_size) - 347 pindex; 348 } 349 *after += numafter; 350 } 351 } else { 352 if (before) { 353 *before /= blocksperpage; 354 } 355 356 if (after) { 357 *after /= blocksperpage; 358 } 359 } 360 return TRUE; 361} 362 363/* 364 * Lets the VM system know about a change in size for a file. 365 * We adjust our own internal size and flush any cached pages in 366 * the associated object that are affected by the size change. 367 * 368 * Note: this routine may be invoked as a result of a pager put 369 * operation (possibly at object termination time), so we must be careful. 370 */ 371void 372vnode_pager_setsize(vp, nsize) 373 struct vnode *vp; 374 vm_ooffset_t nsize; 375{ 376 vm_object_t object; 377 vm_page_t m; 378 vm_pindex_t nobjsize; 379 380 if ((object = vp->v_object) == NULL) 381 return; 382/* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */ 383 VM_OBJECT_WLOCK(object); 384 if (object->type == OBJT_DEAD) { 385 VM_OBJECT_WUNLOCK(object); 386 return; 387 } 388 KASSERT(object->type == OBJT_VNODE, 389 ("not vnode-backed object %p", object)); 390 if (nsize == object->un_pager.vnp.vnp_size) { 391 /* 392 * Hasn't changed size 393 */ 394 VM_OBJECT_WUNLOCK(object); 395 return; 396 } 397 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 398 if (nsize < object->un_pager.vnp.vnp_size) { 399 /* 400 * File has shrunk. Toss any cached pages beyond the new EOF. 401 */ 402 if (nobjsize < object->size) 403 vm_object_page_remove(object, nobjsize, object->size, 404 0); 405 /* 406 * this gets rid of garbage at the end of a page that is now 407 * only partially backed by the vnode. 408 * 409 * XXX for some reason (I don't know yet), if we take a 410 * completely invalid page and mark it partially valid 411 * it can screw up NFS reads, so we don't allow the case. 412 */ 413 if ((nsize & PAGE_MASK) && 414 (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL && 415 m->valid != 0) { 416 int base = (int)nsize & PAGE_MASK; 417 int size = PAGE_SIZE - base; 418 419 /* 420 * Clear out partial-page garbage in case 421 * the page has been mapped. 422 */ 423 pmap_zero_page_area(m, base, size); 424 425 /* 426 * Update the valid bits to reflect the blocks that 427 * have been zeroed. Some of these valid bits may 428 * have already been set. 429 */ 430 vm_page_set_valid_range(m, base, size); 431 432 /* 433 * Round "base" to the next block boundary so that the 434 * dirty bit for a partially zeroed block is not 435 * cleared. 436 */ 437 base = roundup2(base, DEV_BSIZE); 438 439 /* 440 * Clear out partial-page dirty bits. 441 * 442 * note that we do not clear out the valid 443 * bits. This would prevent bogus_page 444 * replacement from working properly. 445 */ 446 vm_page_clear_dirty(m, base, PAGE_SIZE - base); 447 } else if ((nsize & PAGE_MASK) && 448 vm_page_is_cached(object, OFF_TO_IDX(nsize))) { 449 vm_page_cache_free(object, OFF_TO_IDX(nsize), 450 nobjsize); 451 } 452 } 453 object->un_pager.vnp.vnp_size = nsize; 454 object->size = nobjsize; 455 VM_OBJECT_WUNLOCK(object); 456} 457 458/* 459 * calculate the linear (byte) disk address of specified virtual 460 * file address 461 */ 462static int 463vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress, 464 int *run) 465{ 466 int bsize; 467 int err; 468 daddr_t vblock; 469 daddr_t voffset; 470 471 if (address < 0) 472 return -1; 473 474 if (vp->v_iflag & VI_DOOMED) 475 return -1; 476 477 bsize = vp->v_mount->mnt_stat.f_iosize; 478 vblock = address / bsize; 479 voffset = address % bsize; 480 481 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL); 482 if (err == 0) { 483 if (*rtaddress != -1) 484 *rtaddress += voffset / DEV_BSIZE; 485 if (run) { 486 *run += 1; 487 *run *= bsize/PAGE_SIZE; 488 *run -= voffset/PAGE_SIZE; 489 } 490 } 491 492 return (err); 493} 494 495/* 496 * small block filesystem vnode pager input 497 */ 498static int 499vnode_pager_input_smlfs(object, m) 500 vm_object_t object; 501 vm_page_t m; 502{ 503 struct vnode *vp; 504 struct bufobj *bo; 505 struct buf *bp; 506 struct sf_buf *sf; 507 daddr_t fileaddr; 508 vm_offset_t bsize; 509 vm_page_bits_t bits; 510 int error, i; 511 512 error = 0; 513 vp = object->handle; 514 if (vp->v_iflag & VI_DOOMED) 515 return VM_PAGER_BAD; 516 517 bsize = vp->v_mount->mnt_stat.f_iosize; 518 519 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL); 520 521 sf = sf_buf_alloc(m, 0); 522 523 for (i = 0; i < PAGE_SIZE / bsize; i++) { 524 vm_ooffset_t address; 525 526 bits = vm_page_bits(i * bsize, bsize); 527 if (m->valid & bits) 528 continue; 529 530 address = IDX_TO_OFF(m->pindex) + i * bsize; 531 if (address >= object->un_pager.vnp.vnp_size) { 532 fileaddr = -1; 533 } else { 534 error = vnode_pager_addr(vp, address, &fileaddr, NULL); 535 if (error) 536 break; 537 } 538 if (fileaddr != -1) { 539 bp = getpbuf(&vnode_pbuf_freecnt); 540 541 /* build a minimal buffer header */ 542 bp->b_iocmd = BIO_READ; 543 bp->b_iodone = bdone; 544 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 545 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 546 bp->b_rcred = crhold(curthread->td_ucred); 547 bp->b_wcred = crhold(curthread->td_ucred); 548 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize; 549 bp->b_blkno = fileaddr; 550 pbgetbo(bo, bp); 551 bp->b_vp = vp; 552 bp->b_bcount = bsize; 553 bp->b_bufsize = bsize; 554 bp->b_runningbufspace = bp->b_bufsize; 555 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 556 557 /* do the input */ 558 bp->b_iooffset = dbtob(bp->b_blkno); 559 bstrategy(bp); 560 561 bwait(bp, PVM, "vnsrd"); 562 563 if ((bp->b_ioflags & BIO_ERROR) != 0) 564 error = EIO; 565 566 /* 567 * free the buffer header back to the swap buffer pool 568 */ 569 bp->b_vp = NULL; 570 pbrelbo(bp); 571 relpbuf(bp, &vnode_pbuf_freecnt); 572 if (error) 573 break; 574 } else 575 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize); 576 KASSERT((m->dirty & bits) == 0, 577 ("vnode_pager_input_smlfs: page %p is dirty", m)); 578 VM_OBJECT_WLOCK(object); 579 m->valid |= bits; 580 VM_OBJECT_WUNLOCK(object); 581 } 582 sf_buf_free(sf); 583 if (error) { 584 return VM_PAGER_ERROR; 585 } 586 return VM_PAGER_OK; 587} 588 589/* 590 * old style vnode pager input routine 591 */ 592static int 593vnode_pager_input_old(object, m) 594 vm_object_t object; 595 vm_page_t m; 596{ 597 struct uio auio; 598 struct iovec aiov; 599 int error; 600 int size; 601 struct sf_buf *sf; 602 struct vnode *vp; 603 604 VM_OBJECT_ASSERT_WLOCKED(object); 605 error = 0; 606 607 /* 608 * Return failure if beyond current EOF 609 */ 610 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 611 return VM_PAGER_BAD; 612 } else { 613 size = PAGE_SIZE; 614 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 615 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 616 vp = object->handle; 617 VM_OBJECT_WUNLOCK(object); 618 619 /* 620 * Allocate a kernel virtual address and initialize so that 621 * we can use VOP_READ/WRITE routines. 622 */ 623 sf = sf_buf_alloc(m, 0); 624 625 aiov.iov_base = (caddr_t)sf_buf_kva(sf); 626 aiov.iov_len = size; 627 auio.uio_iov = &aiov; 628 auio.uio_iovcnt = 1; 629 auio.uio_offset = IDX_TO_OFF(m->pindex); 630 auio.uio_segflg = UIO_SYSSPACE; 631 auio.uio_rw = UIO_READ; 632 auio.uio_resid = size; 633 auio.uio_td = curthread; 634 635 error = VOP_READ(vp, &auio, 0, curthread->td_ucred); 636 if (!error) { 637 int count = size - auio.uio_resid; 638 639 if (count == 0) 640 error = EINVAL; 641 else if (count != PAGE_SIZE) 642 bzero((caddr_t)sf_buf_kva(sf) + count, 643 PAGE_SIZE - count); 644 } 645 sf_buf_free(sf); 646 647 VM_OBJECT_WLOCK(object); 648 } 649 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m)); 650 if (!error) 651 m->valid = VM_PAGE_BITS_ALL; 652 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 653} 654 655/* 656 * generic vnode pager input routine 657 */ 658 659/* 660 * Local media VFS's that do not implement their own VOP_GETPAGES 661 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages() 662 * to implement the previous behaviour. 663 * 664 * All other FS's should use the bypass to get to the local media 665 * backing vp's VOP_GETPAGES. 666 */ 667static int 668vnode_pager_getpages(object, m, count, reqpage) 669 vm_object_t object; 670 vm_page_t *m; 671 int count; 672 int reqpage; 673{ 674 int rtval; 675 struct vnode *vp; 676 int bytes = count * PAGE_SIZE; 677 678 vp = object->handle; 679 VM_OBJECT_WUNLOCK(object); 680 rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0); 681 KASSERT(rtval != EOPNOTSUPP, 682 ("vnode_pager: FS getpages not implemented\n")); 683 VM_OBJECT_WLOCK(object); 684 return rtval; 685} 686 687/* 688 * This is now called from local media FS's to operate against their 689 * own vnodes if they fail to implement VOP_GETPAGES. 690 */ 691int 692vnode_pager_generic_getpages(vp, m, bytecount, reqpage) 693 struct vnode *vp; 694 vm_page_t *m; 695 int bytecount; 696 int reqpage; 697{ 698 vm_object_t object; 699 struct bufobj *bo; 700 struct buf *bp; 701 struct mount *mp; 702 vm_offset_t kva; 703 daddr_t firstaddr, reqblock; 704 off_t foff, nextoff, tfoff, pib; 705 int pbefore, pafter, i, size, bsize, first, last; 706 int count, error, before, after, secmask; 707 708 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 709 ("vnode_pager_generic_getpages does not support devices")); 710 if (vp->v_iflag & VI_DOOMED) 711 return (VM_PAGER_BAD); 712 713 object = vp->v_object; 714 count = bytecount / PAGE_SIZE; 715 bsize = vp->v_mount->mnt_stat.f_iosize; 716 717 /* get the UNDERLYING device for the file with VOP_BMAP() */ 718 719 /* 720 * originally, we did not check for an error return value -- assuming 721 * an fs always has a bmap entry point -- that assumption is wrong!!! 722 */ 723 foff = IDX_TO_OFF(m[reqpage]->pindex); 724 725 /* 726 * if we can't bmap, use old VOP code 727 */ 728 error = VOP_BMAP(vp, IDX_TO_OFF(m[reqpage]->pindex) / bsize, &bo, 729 &reqblock, &after, &before); 730 if (error == EOPNOTSUPP) { 731 VM_OBJECT_WLOCK(object); 732 733 for (i = 0; i < count; i++) 734 if (i != reqpage) { 735 vm_page_lock(m[i]); 736 vm_page_free(m[i]); 737 vm_page_unlock(m[i]); 738 } 739 PCPU_INC(cnt.v_vnodein); 740 PCPU_INC(cnt.v_vnodepgsin); 741 error = vnode_pager_input_old(object, m[reqpage]); 742 VM_OBJECT_WUNLOCK(object); 743 return (error); 744 } else if (error != 0) { 745 VM_OBJECT_WLOCK(object); 746 for (i = 0; i < count; i++) 747 if (i != reqpage) { 748 vm_page_lock(m[i]); 749 vm_page_free(m[i]); 750 vm_page_unlock(m[i]); 751 } 752 VM_OBJECT_WUNLOCK(object); 753 return (VM_PAGER_ERROR); 754 755 /* 756 * if the blocksize is smaller than a page size, then use 757 * special small filesystem code. NFS sometimes has a small 758 * blocksize, but it can handle large reads itself. 759 */ 760 } else if ((PAGE_SIZE / bsize) > 1 && 761 (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) { 762 VM_OBJECT_WLOCK(object); 763 for (i = 0; i < count; i++) 764 if (i != reqpage) { 765 vm_page_lock(m[i]); 766 vm_page_free(m[i]); 767 vm_page_unlock(m[i]); 768 } 769 VM_OBJECT_WUNLOCK(object); 770 PCPU_INC(cnt.v_vnodein); 771 PCPU_INC(cnt.v_vnodepgsin); 772 return (vnode_pager_input_smlfs(object, m[reqpage])); 773 } 774 775 /* 776 * If we have a completely valid page available to us, we can 777 * clean up and return. Otherwise we have to re-read the 778 * media. 779 */ 780 VM_OBJECT_WLOCK(object); 781 if (m[reqpage]->valid == VM_PAGE_BITS_ALL) { 782 for (i = 0; i < count; i++) 783 if (i != reqpage) { 784 vm_page_lock(m[i]); 785 vm_page_free(m[i]); 786 vm_page_unlock(m[i]); 787 } 788 VM_OBJECT_WUNLOCK(object); 789 return VM_PAGER_OK; 790 } else if (reqblock == -1) { 791 pmap_zero_page(m[reqpage]); 792 KASSERT(m[reqpage]->dirty == 0, 793 ("vnode_pager_generic_getpages: page %p is dirty", m)); 794 m[reqpage]->valid = VM_PAGE_BITS_ALL; 795 for (i = 0; i < count; i++) 796 if (i != reqpage) { 797 vm_page_lock(m[i]); 798 vm_page_free(m[i]); 799 vm_page_unlock(m[i]); 800 } 801 VM_OBJECT_WUNLOCK(object); 802 return (VM_PAGER_OK); 803 } 804 m[reqpage]->valid = 0; 805 VM_OBJECT_WUNLOCK(object); 806 807 pib = IDX_TO_OFF(m[reqpage]->pindex) % bsize; 808 pbefore = ((daddr_t)before * bsize + pib) / PAGE_SIZE; 809 pafter = ((daddr_t)(after + 1) * bsize - pib) / PAGE_SIZE - 1; 810 first = reqpage < pbefore ? 0 : reqpage - pbefore; 811 last = reqpage + pafter >= count ? count - 1 : reqpage + pafter; 812 if (first > 0 || last + 1 < count) { 813 VM_OBJECT_WLOCK(object); 814 for (i = 0; i < first; i++) { 815 vm_page_lock(m[i]); 816 vm_page_free(m[i]); 817 vm_page_unlock(m[i]); 818 } 819 for (i = last + 1; i < count; i++) { 820 vm_page_lock(m[i]); 821 vm_page_free(m[i]); 822 vm_page_unlock(m[i]); 823 } 824 VM_OBJECT_WUNLOCK(object); 825 } 826 827 /* 828 * here on direct device I/O 829 */ 830 firstaddr = reqblock; 831 firstaddr += pib / DEV_BSIZE; 832 firstaddr -= IDX_TO_OFF(reqpage - first) / DEV_BSIZE; 833 834 /* 835 * The first and last page have been calculated now, move 836 * input pages to be zero based, and adjust the count. 837 */ 838 m += first; 839 reqpage -= first; 840 count = last - first + 1; 841 842 /* 843 * calculate the file virtual address for the transfer 844 */ 845 foff = IDX_TO_OFF(m[0]->pindex); 846 847 /* 848 * calculate the size of the transfer 849 */ 850 size = count * PAGE_SIZE; 851 KASSERT(count > 0, ("zero count")); 852 if ((foff + size) > object->un_pager.vnp.vnp_size) 853 size = object->un_pager.vnp.vnp_size - foff; 854 KASSERT(size > 0, ("zero size")); 855 856 /* 857 * round up physical size for real devices. 858 */ 859 secmask = bo->bo_bsize - 1; 860 KASSERT(secmask < PAGE_SIZE && secmask > 0, 861 ("vnode_pager_generic_getpages: sector size %d too large", 862 secmask + 1)); 863 size = (size + secmask) & ~secmask; 864 865 bp = getpbuf(&vnode_pbuf_freecnt); 866 kva = (vm_offset_t)bp->b_data; 867 868 /* 869 * and map the pages to be read into the kva, if the filesystem 870 * requires mapped buffers. 871 */ 872 mp = vp->v_mount; 873 if (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 && 874 unmapped_buf_allowed) { 875 bp->b_data = unmapped_buf; 876 bp->b_kvabase = unmapped_buf; 877 bp->b_offset = 0; 878 bp->b_flags |= B_UNMAPPED; 879 bp->b_npages = count; 880 for (i = 0; i < count; i++) 881 bp->b_pages[i] = m[i]; 882 } else 883 pmap_qenter(kva, m, count); 884 885 /* build a minimal buffer header */ 886 bp->b_iocmd = BIO_READ; 887 bp->b_iodone = bdone; 888 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 889 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 890 bp->b_rcred = crhold(curthread->td_ucred); 891 bp->b_wcred = crhold(curthread->td_ucred); 892 bp->b_blkno = firstaddr; 893 pbgetbo(bo, bp); 894 bp->b_vp = vp; 895 bp->b_bcount = size; 896 bp->b_bufsize = size; 897 bp->b_runningbufspace = bp->b_bufsize; 898 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 899 900 PCPU_INC(cnt.v_vnodein); 901 PCPU_ADD(cnt.v_vnodepgsin, count); 902 903 /* do the input */ 904 bp->b_iooffset = dbtob(bp->b_blkno); 905 bstrategy(bp); 906 907 bwait(bp, PVM, "vnread"); 908 909 if ((bp->b_ioflags & BIO_ERROR) != 0) 910 error = EIO; 911 912 if (error == 0 && size != count * PAGE_SIZE) { 913 if ((bp->b_flags & B_UNMAPPED) != 0) { 914 bp->b_flags &= ~B_UNMAPPED; 915 pmap_qenter(kva, m, count); 916 } 917 bzero((caddr_t)kva + size, PAGE_SIZE * count - size); 918 } 919 if ((bp->b_flags & B_UNMAPPED) == 0) 920 pmap_qremove(kva, count); 921 if (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0) { 922 bp->b_data = (caddr_t)kva; 923 bp->b_kvabase = (caddr_t)kva; 924 bp->b_flags &= ~B_UNMAPPED; 925 for (i = 0; i < count; i++) 926 bp->b_pages[i] = NULL; 927 } 928 929 /* 930 * free the buffer header back to the swap buffer pool 931 */ 932 bp->b_vp = NULL; 933 pbrelbo(bp); 934 relpbuf(bp, &vnode_pbuf_freecnt); 935 936 VM_OBJECT_WLOCK(object); 937 for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) { 938 vm_page_t mt; 939 940 nextoff = tfoff + PAGE_SIZE; 941 mt = m[i]; 942 943 if (nextoff <= object->un_pager.vnp.vnp_size) { 944 /* 945 * Read filled up entire page. 946 */ 947 mt->valid = VM_PAGE_BITS_ALL; 948 KASSERT(mt->dirty == 0, 949 ("vnode_pager_generic_getpages: page %p is dirty", 950 mt)); 951 KASSERT(!pmap_page_is_mapped(mt), 952 ("vnode_pager_generic_getpages: page %p is mapped", 953 mt)); 954 } else { 955 /* 956 * Read did not fill up entire page. 957 * 958 * Currently we do not set the entire page valid, 959 * we just try to clear the piece that we couldn't 960 * read. 961 */ 962 vm_page_set_valid_range(mt, 0, 963 object->un_pager.vnp.vnp_size - tfoff); 964 KASSERT((mt->dirty & vm_page_bits(0, 965 object->un_pager.vnp.vnp_size - tfoff)) == 0, 966 ("vnode_pager_generic_getpages: page %p is dirty", 967 mt)); 968 } 969 970 if (i != reqpage) 971 vm_page_readahead_finish(mt); 972 } 973 VM_OBJECT_WUNLOCK(object); 974 if (error) { 975 printf("vnode_pager_getpages: I/O read error\n"); 976 } 977 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 978} 979 980/* 981 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 982 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 983 * vnode_pager_generic_putpages() to implement the previous behaviour. 984 * 985 * All other FS's should use the bypass to get to the local media 986 * backing vp's VOP_PUTPAGES. 987 */ 988static void 989vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count, 990 int flags, int *rtvals) 991{ 992 int rtval; 993 struct vnode *vp; 994 int bytes = count * PAGE_SIZE; 995 996 /* 997 * Force synchronous operation if we are extremely low on memory 998 * to prevent a low-memory deadlock. VOP operations often need to 999 * allocate more memory to initiate the I/O ( i.e. do a BMAP 1000 * operation ). The swapper handles the case by limiting the amount 1001 * of asynchronous I/O, but that sort of solution doesn't scale well 1002 * for the vnode pager without a lot of work. 1003 * 1004 * Also, the backing vnode's iodone routine may not wake the pageout 1005 * daemon up. This should be probably be addressed XXX. 1006 */ 1007 1008 if (cnt.v_free_count + cnt.v_cache_count < cnt.v_pageout_free_min) 1009 flags |= VM_PAGER_PUT_SYNC; 1010 1011 /* 1012 * Call device-specific putpages function 1013 */ 1014 vp = object->handle; 1015 VM_OBJECT_WUNLOCK(object); 1016 rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals, 0); 1017 KASSERT(rtval != EOPNOTSUPP, 1018 ("vnode_pager: stale FS putpages\n")); 1019 VM_OBJECT_WLOCK(object); 1020} 1021 1022 1023/* 1024 * This is now called from local media FS's to operate against their 1025 * own vnodes if they fail to implement VOP_PUTPAGES. 1026 * 1027 * This is typically called indirectly via the pageout daemon and 1028 * clustering has already typically occured, so in general we ask the 1029 * underlying filesystem to write the data out asynchronously rather 1030 * then delayed. 1031 */ 1032int 1033vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount, 1034 int flags, int *rtvals) 1035{ 1036 int i; 1037 vm_object_t object; 1038 vm_page_t m; 1039 int count; 1040 1041 int maxsize, ncount; 1042 vm_ooffset_t poffset; 1043 struct uio auio; 1044 struct iovec aiov; 1045 int error; 1046 int ioflags; 1047 int ppscheck = 0; 1048 static struct timeval lastfail; 1049 static int curfail; 1050 1051 object = vp->v_object; 1052 count = bytecount / PAGE_SIZE; 1053 1054 for (i = 0; i < count; i++) 1055 rtvals[i] = VM_PAGER_ERROR; 1056 1057 if ((int64_t)ma[0]->pindex < 0) { 1058 printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n", 1059 (long)ma[0]->pindex, (u_long)ma[0]->dirty); 1060 rtvals[0] = VM_PAGER_BAD; 1061 return VM_PAGER_BAD; 1062 } 1063 1064 maxsize = count * PAGE_SIZE; 1065 ncount = count; 1066 1067 poffset = IDX_TO_OFF(ma[0]->pindex); 1068 1069 /* 1070 * If the page-aligned write is larger then the actual file we 1071 * have to invalidate pages occuring beyond the file EOF. However, 1072 * there is an edge case where a file may not be page-aligned where 1073 * the last page is partially invalid. In this case the filesystem 1074 * may not properly clear the dirty bits for the entire page (which 1075 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 1076 * With the page locked we are free to fix-up the dirty bits here. 1077 * 1078 * We do not under any circumstances truncate the valid bits, as 1079 * this will screw up bogus page replacement. 1080 */ 1081 VM_OBJECT_WLOCK(object); 1082 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 1083 if (object->un_pager.vnp.vnp_size > poffset) { 1084 int pgoff; 1085 1086 maxsize = object->un_pager.vnp.vnp_size - poffset; 1087 ncount = btoc(maxsize); 1088 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 1089 /* 1090 * If the object is locked and the following 1091 * conditions hold, then the page's dirty 1092 * field cannot be concurrently changed by a 1093 * pmap operation. 1094 */ 1095 m = ma[ncount - 1]; 1096 vm_page_assert_sbusied(m); 1097 KASSERT(!pmap_page_is_write_mapped(m), 1098 ("vnode_pager_generic_putpages: page %p is not read-only", m)); 1099 vm_page_clear_dirty(m, pgoff, PAGE_SIZE - 1100 pgoff); 1101 } 1102 } else { 1103 maxsize = 0; 1104 ncount = 0; 1105 } 1106 if (ncount < count) { 1107 for (i = ncount; i < count; i++) { 1108 rtvals[i] = VM_PAGER_BAD; 1109 } 1110 } 1111 } 1112 VM_OBJECT_WUNLOCK(object); 1113 1114 /* 1115 * pageouts are already clustered, use IO_ASYNC to force a bawrite() 1116 * rather then a bdwrite() to prevent paging I/O from saturating 1117 * the buffer cache. Dummy-up the sequential heuristic to cause 1118 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, 1119 * the system decides how to cluster. 1120 */ 1121 ioflags = IO_VMIO; 1122 if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) 1123 ioflags |= IO_SYNC; 1124 else if ((flags & VM_PAGER_CLUSTER_OK) == 0) 1125 ioflags |= IO_ASYNC; 1126 ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; 1127 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 1128 1129 aiov.iov_base = (caddr_t) 0; 1130 aiov.iov_len = maxsize; 1131 auio.uio_iov = &aiov; 1132 auio.uio_iovcnt = 1; 1133 auio.uio_offset = poffset; 1134 auio.uio_segflg = UIO_NOCOPY; 1135 auio.uio_rw = UIO_WRITE; 1136 auio.uio_resid = maxsize; 1137 auio.uio_td = (struct thread *) 0; 1138 error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred); 1139 PCPU_INC(cnt.v_vnodeout); 1140 PCPU_ADD(cnt.v_vnodepgsout, ncount); 1141 1142 if (error) { 1143 if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1))) 1144 printf("vnode_pager_putpages: I/O error %d\n", error); 1145 } 1146 if (auio.uio_resid) { 1147 if (ppscheck || ppsratecheck(&lastfail, &curfail, 1)) 1148 printf("vnode_pager_putpages: residual I/O %zd at %lu\n", 1149 auio.uio_resid, (u_long)ma[0]->pindex); 1150 } 1151 for (i = 0; i < ncount; i++) { 1152 rtvals[i] = VM_PAGER_OK; 1153 } 1154 return rtvals[0]; 1155} 1156 1157void 1158vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written) 1159{ 1160 vm_object_t obj; 1161 int i, pos; 1162 1163 if (written == 0) 1164 return; 1165 obj = ma[0]->object; 1166 VM_OBJECT_WLOCK(obj); 1167 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) { 1168 if (pos < trunc_page(written)) { 1169 rtvals[i] = VM_PAGER_OK; 1170 vm_page_undirty(ma[i]); 1171 } else { 1172 /* Partially written page. */ 1173 rtvals[i] = VM_PAGER_AGAIN; 1174 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK); 1175 } 1176 } 1177 VM_OBJECT_WUNLOCK(obj); 1178} 1179 1180void 1181vnode_pager_update_writecount(vm_object_t object, vm_offset_t start, 1182 vm_offset_t end) 1183{ 1184 struct vnode *vp; 1185 vm_ooffset_t old_wm; 1186 1187 VM_OBJECT_WLOCK(object); 1188 if (object->type != OBJT_VNODE) { 1189 VM_OBJECT_WUNLOCK(object); 1190 return; 1191 } 1192 old_wm = object->un_pager.vnp.writemappings; 1193 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start; 1194 vp = object->handle; 1195 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) { 1196 ASSERT_VOP_ELOCKED(vp, "v_writecount inc"); 1197 VOP_ADD_WRITECOUNT(vp, 1); 1198 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", 1199 __func__, vp, vp->v_writecount); 1200 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) { 1201 ASSERT_VOP_ELOCKED(vp, "v_writecount dec"); 1202 VOP_ADD_WRITECOUNT(vp, -1); 1203 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 1204 __func__, vp, vp->v_writecount); 1205 } 1206 VM_OBJECT_WUNLOCK(object); 1207} 1208 1209void 1210vnode_pager_release_writecount(vm_object_t object, vm_offset_t start, 1211 vm_offset_t end) 1212{ 1213 struct vnode *vp; 1214 struct mount *mp; 1215 vm_offset_t inc; 1216 1217 VM_OBJECT_WLOCK(object); 1218 1219 /* 1220 * First, recheck the object type to account for the race when 1221 * the vnode is reclaimed. 1222 */ 1223 if (object->type != OBJT_VNODE) { 1224 VM_OBJECT_WUNLOCK(object); 1225 return; 1226 } 1227 1228 /* 1229 * Optimize for the case when writemappings is not going to 1230 * zero. 1231 */ 1232 inc = end - start; 1233 if (object->un_pager.vnp.writemappings != inc) { 1234 object->un_pager.vnp.writemappings -= inc; 1235 VM_OBJECT_WUNLOCK(object); 1236 return; 1237 } 1238 1239 vp = object->handle; 1240 vhold(vp); 1241 VM_OBJECT_WUNLOCK(object); 1242 mp = NULL; 1243 vn_start_write(vp, &mp, V_WAIT); 1244 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1245 1246 /* 1247 * Decrement the object's writemappings, by swapping the start 1248 * and end arguments for vnode_pager_update_writecount(). If 1249 * there was not a race with vnode reclaimation, then the 1250 * vnode's v_writecount is decremented. 1251 */ 1252 vnode_pager_update_writecount(object, end, start); 1253 VOP_UNLOCK(vp, 0); 1254 vdrop(vp); 1255 if (mp != NULL) 1256 vn_finished_write(mp); 1257} 1258