1 /*- 2 * SPDX-License-Identifier: BSD-4-Clause 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 */ 44 45 /* 46 * Page to/from files (vnodes). 47 */ 48 49 /* 50 * TODO: 51 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will 52 * greatly re-simplify the vnode_pager. 53 */ 54 55 #include <sys/cdefs.h> 56 __FBSDID("$FreeBSD$"); 57 58 #include "opt_vm.h" 59 60 #include <sys/param.h> 61 #include <sys/kernel.h> 62 #include <sys/systm.h> 63 #include <sys/sysctl.h> 64 #include <sys/proc.h> 65 #include <sys/vnode.h> 66 #include <sys/mount.h> 67 #include <sys/bio.h> 68 #include <sys/buf.h> 69 #include <sys/vmmeter.h> 70 #include <sys/ktr.h> 71 #include <sys/limits.h> 72 #include <sys/conf.h> 73 #include <sys/refcount.h> 74 #include <sys/rwlock.h> 75 #include <sys/sf_buf.h> 76 #include <sys/domainset.h> 77 78 #include <machine/atomic.h> 79 80 #include <vm/vm.h> 81 #include <vm/vm_param.h> 82 #include <vm/vm_object.h> 83 #include <vm/vm_page.h> 84 #include <vm/vm_pager.h> 85 #include <vm/vm_map.h> 86 #include <vm/vnode_pager.h> 87 #include <vm/vm_extern.h> 88 #include <vm/uma.h> 89 90 static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, 91 daddr_t *rtaddress, int *run); 92 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m); 93 static int vnode_pager_input_old(vm_object_t object, vm_page_t m); 94 static void vnode_pager_dealloc(vm_object_t); 95 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *); 96 static int vnode_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 97 int *, vop_getpages_iodone_t, void *); 98 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *); 99 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *); 100 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t, 101 vm_ooffset_t, struct ucred *cred); 102 static int vnode_pager_generic_getpages_done(struct buf *); 103 static void vnode_pager_generic_getpages_done_async(struct buf *); 104 static void vnode_pager_update_writecount(vm_object_t, vm_offset_t, 105 vm_offset_t); 106 static void vnode_pager_release_writecount(vm_object_t, vm_offset_t, 107 vm_offset_t); 108 109 struct pagerops vnodepagerops = { 110 .pgo_alloc = vnode_pager_alloc, 111 .pgo_dealloc = vnode_pager_dealloc, 112 .pgo_getpages = vnode_pager_getpages, 113 .pgo_getpages_async = vnode_pager_getpages_async, 114 .pgo_putpages = vnode_pager_putpages, 115 .pgo_haspage = vnode_pager_haspage, 116 .pgo_update_writecount = vnode_pager_update_writecount, 117 .pgo_release_writecount = vnode_pager_release_writecount, 118 }; 119 120 static struct domainset *vnode_domainset = NULL; 121 122 SYSCTL_PROC(_debug, OID_AUTO, vnode_domainset, 123 CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_RW, &vnode_domainset, 0, 124 sysctl_handle_domainset, "A", "Default vnode NUMA policy"); 125 126 static int nvnpbufs; 127 SYSCTL_INT(_vm, OID_AUTO, vnode_pbufs, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 128 &nvnpbufs, 0, "number of physical buffers allocated for vnode pager"); 129 130 static uma_zone_t vnode_pbuf_zone; 131 132 static void 133 vnode_pager_init(void *dummy) 134 { 135 136 #ifdef __LP64__ 137 nvnpbufs = nswbuf * 2; 138 #else 139 nvnpbufs = nswbuf / 2; 140 #endif 141 TUNABLE_INT_FETCH("vm.vnode_pbufs", &nvnpbufs); 142 vnode_pbuf_zone = pbuf_zsecond_create("vnpbuf", nvnpbufs); 143 } 144 SYSINIT(vnode_pager, SI_SUB_CPU, SI_ORDER_ANY, vnode_pager_init, NULL); 145 146 /* Create the VM system backing object for this vnode */ 147 int 148 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td) 149 { 150 vm_object_t object; 151 vm_ooffset_t size = isize; 152 struct vattr va; 153 bool last; 154 155 if (!vn_isdisk(vp) && vn_canvmio(vp) == FALSE) 156 return (0); 157 158 object = vp->v_object; 159 if (object != NULL) 160 return (0); 161 162 if (size == 0) { 163 if (vn_isdisk(vp)) { 164 size = IDX_TO_OFF(INT_MAX); 165 } else { 166 if (VOP_GETATTR(vp, &va, td->td_ucred)) 167 return (0); 168 size = va.va_size; 169 } 170 } 171 172 object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred); 173 /* 174 * Dereference the reference we just created. This assumes 175 * that the object is associated with the vp. We still have 176 * to serialize with vnode_pager_dealloc() for the last 177 * potential reference. 178 */ 179 VM_OBJECT_RLOCK(object); 180 last = refcount_release(&object->ref_count); 181 VM_OBJECT_RUNLOCK(object); 182 if (last) 183 vrele(vp); 184 185 KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object")); 186 187 return (0); 188 } 189 190 void 191 vnode_destroy_vobject(struct vnode *vp) 192 { 193 struct vm_object *obj; 194 195 obj = vp->v_object; 196 if (obj == NULL || obj->handle != vp) 197 return; 198 ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject"); 199 VM_OBJECT_WLOCK(obj); 200 MPASS(obj->type == OBJT_VNODE); 201 umtx_shm_object_terminated(obj); 202 if (obj->ref_count == 0) { 203 KASSERT((obj->flags & OBJ_DEAD) == 0, 204 ("vnode_destroy_vobject: Terminating dead object")); 205 vm_object_set_flag(obj, OBJ_DEAD); 206 207 /* 208 * Clean pages and flush buffers. 209 */ 210 vm_object_page_clean(obj, 0, 0, OBJPC_SYNC); 211 VM_OBJECT_WUNLOCK(obj); 212 213 vinvalbuf(vp, V_SAVE, 0, 0); 214 215 BO_LOCK(&vp->v_bufobj); 216 vp->v_bufobj.bo_flag |= BO_DEAD; 217 BO_UNLOCK(&vp->v_bufobj); 218 219 VM_OBJECT_WLOCK(obj); 220 vm_object_terminate(obj); 221 } else { 222 /* 223 * Woe to the process that tries to page now :-). 224 */ 225 vm_pager_deallocate(obj); 226 VM_OBJECT_WUNLOCK(obj); 227 } 228 KASSERT(vp->v_object == NULL, ("vp %p obj %p", vp, vp->v_object)); 229 } 230 231 232 /* 233 * Allocate (or lookup) pager for a vnode. 234 * Handle is a vnode pointer. 235 */ 236 vm_object_t 237 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 238 vm_ooffset_t offset, struct ucred *cred) 239 { 240 vm_object_t object; 241 struct vnode *vp; 242 243 /* 244 * Pageout to vnode, no can do yet. 245 */ 246 if (handle == NULL) 247 return (NULL); 248 249 vp = (struct vnode *)handle; 250 ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc"); 251 VNPASS(vp->v_usecount > 0, vp); 252 retry: 253 object = vp->v_object; 254 255 if (object == NULL) { 256 /* 257 * Add an object of the appropriate size 258 */ 259 object = vm_object_allocate(OBJT_VNODE, 260 OFF_TO_IDX(round_page(size))); 261 262 object->un_pager.vnp.vnp_size = size; 263 object->un_pager.vnp.writemappings = 0; 264 object->domain.dr_policy = vnode_domainset; 265 object->handle = handle; 266 if ((vp->v_vflag & VV_VMSIZEVNLOCK) != 0) { 267 VM_OBJECT_WLOCK(object); 268 vm_object_set_flag(object, OBJ_SIZEVNLOCK); 269 VM_OBJECT_WUNLOCK(object); 270 } 271 VI_LOCK(vp); 272 if (vp->v_object != NULL) { 273 /* 274 * Object has been created while we were allocating. 275 */ 276 VI_UNLOCK(vp); 277 VM_OBJECT_WLOCK(object); 278 KASSERT(object->ref_count == 1, 279 ("leaked ref %p %d", object, object->ref_count)); 280 object->type = OBJT_DEAD; 281 refcount_init(&object->ref_count, 0); 282 VM_OBJECT_WUNLOCK(object); 283 vm_object_destroy(object); 284 goto retry; 285 } 286 vp->v_object = object; 287 VI_UNLOCK(vp); 288 vrefact(vp); 289 } else { 290 vm_object_reference(object); 291 #if VM_NRESERVLEVEL > 0 292 if ((object->flags & OBJ_COLORED) == 0) { 293 VM_OBJECT_WLOCK(object); 294 vm_object_color(object, 0); 295 VM_OBJECT_WUNLOCK(object); 296 } 297 #endif 298 } 299 return (object); 300 } 301 302 /* 303 * The object must be locked. 304 */ 305 static void 306 vnode_pager_dealloc(vm_object_t object) 307 { 308 struct vnode *vp; 309 int refs; 310 311 vp = object->handle; 312 if (vp == NULL) 313 panic("vnode_pager_dealloc: pager already dealloced"); 314 315 VM_OBJECT_ASSERT_WLOCKED(object); 316 vm_object_pip_wait(object, "vnpdea"); 317 refs = object->ref_count; 318 319 object->handle = NULL; 320 object->type = OBJT_DEAD; 321 ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc"); 322 if (object->un_pager.vnp.writemappings > 0) { 323 object->un_pager.vnp.writemappings = 0; 324 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 325 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 326 __func__, vp, vp->v_writecount); 327 } 328 vp->v_object = NULL; 329 VI_LOCK(vp); 330 331 /* 332 * vm_map_entry_set_vnode_text() cannot reach this vnode by 333 * following object->handle. Clear all text references now. 334 * This also clears the transient references from 335 * kern_execve(), which is fine because dead_vnodeops uses nop 336 * for VOP_UNSET_TEXT(). 337 */ 338 if (vp->v_writecount < 0) 339 vp->v_writecount = 0; 340 VI_UNLOCK(vp); 341 VM_OBJECT_WUNLOCK(object); 342 if (refs > 0) 343 vunref(vp); 344 VM_OBJECT_WLOCK(object); 345 } 346 347 static boolean_t 348 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 349 int *after) 350 { 351 struct vnode *vp = object->handle; 352 daddr_t bn; 353 uintptr_t lockstate; 354 int err; 355 daddr_t reqblock; 356 int poff; 357 int bsize; 358 int pagesperblock, blocksperpage; 359 360 VM_OBJECT_ASSERT_LOCKED(object); 361 /* 362 * If no vp or vp is doomed or marked transparent to VM, we do not 363 * have the page. 364 */ 365 if (vp == NULL || VN_IS_DOOMED(vp)) 366 return FALSE; 367 /* 368 * If the offset is beyond end of file we do 369 * not have the page. 370 */ 371 if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size) 372 return FALSE; 373 374 bsize = vp->v_mount->mnt_stat.f_iosize; 375 pagesperblock = bsize / PAGE_SIZE; 376 blocksperpage = 0; 377 if (pagesperblock > 0) { 378 reqblock = pindex / pagesperblock; 379 } else { 380 blocksperpage = (PAGE_SIZE / bsize); 381 reqblock = pindex * blocksperpage; 382 } 383 lockstate = VM_OBJECT_DROP(object); 384 err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before); 385 VM_OBJECT_PICKUP(object, lockstate); 386 if (err) 387 return TRUE; 388 if (bn == -1) 389 return FALSE; 390 if (pagesperblock > 0) { 391 poff = pindex - (reqblock * pagesperblock); 392 if (before) { 393 *before *= pagesperblock; 394 *before += poff; 395 } 396 if (after) { 397 /* 398 * The BMAP vop can report a partial block in the 399 * 'after', but must not report blocks after EOF. 400 * Assert the latter, and truncate 'after' in case 401 * of the former. 402 */ 403 KASSERT((reqblock + *after) * pagesperblock < 404 roundup2(object->size, pagesperblock), 405 ("%s: reqblock %jd after %d size %ju", __func__, 406 (intmax_t )reqblock, *after, 407 (uintmax_t )object->size)); 408 *after *= pagesperblock; 409 *after += pagesperblock - (poff + 1); 410 if (pindex + *after >= object->size) 411 *after = object->size - 1 - pindex; 412 } 413 } else { 414 if (before) { 415 *before /= blocksperpage; 416 } 417 418 if (after) { 419 *after /= blocksperpage; 420 } 421 } 422 return TRUE; 423 } 424 425 /* 426 * Lets the VM system know about a change in size for a file. 427 * We adjust our own internal size and flush any cached pages in 428 * the associated object that are affected by the size change. 429 * 430 * Note: this routine may be invoked as a result of a pager put 431 * operation (possibly at object termination time), so we must be careful. 432 */ 433 void 434 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize) 435 { 436 vm_object_t object; 437 vm_page_t m; 438 vm_pindex_t nobjsize; 439 440 if ((object = vp->v_object) == NULL) 441 return; 442 #ifdef DEBUG_VFS_LOCKS 443 { 444 struct mount *mp; 445 446 mp = vp->v_mount; 447 if (mp != NULL && (mp->mnt_kern_flag & MNTK_VMSETSIZE_BUG) == 0) 448 assert_vop_elocked(vp, 449 "vnode_pager_setsize and not locked vnode"); 450 } 451 #endif 452 VM_OBJECT_WLOCK(object); 453 if (object->type == OBJT_DEAD) { 454 VM_OBJECT_WUNLOCK(object); 455 return; 456 } 457 KASSERT(object->type == OBJT_VNODE, 458 ("not vnode-backed object %p", object)); 459 if (nsize == object->un_pager.vnp.vnp_size) { 460 /* 461 * Hasn't changed size 462 */ 463 VM_OBJECT_WUNLOCK(object); 464 return; 465 } 466 nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); 467 if (nsize < object->un_pager.vnp.vnp_size) { 468 /* 469 * File has shrunk. Toss any cached pages beyond the new EOF. 470 */ 471 if (nobjsize < object->size) 472 vm_object_page_remove(object, nobjsize, object->size, 473 0); 474 /* 475 * this gets rid of garbage at the end of a page that is now 476 * only partially backed by the vnode. 477 * 478 * XXX for some reason (I don't know yet), if we take a 479 * completely invalid page and mark it partially valid 480 * it can screw up NFS reads, so we don't allow the case. 481 */ 482 if (!(nsize & PAGE_MASK)) 483 goto out; 484 m = vm_page_grab(object, OFF_TO_IDX(nsize), VM_ALLOC_NOCREAT); 485 if (m == NULL) 486 goto out; 487 if (!vm_page_none_valid(m)) { 488 int base = (int)nsize & PAGE_MASK; 489 int size = PAGE_SIZE - base; 490 491 /* 492 * Clear out partial-page garbage in case 493 * the page has been mapped. 494 */ 495 pmap_zero_page_area(m, base, size); 496 497 /* 498 * Update the valid bits to reflect the blocks that 499 * have been zeroed. Some of these valid bits may 500 * have already been set. 501 */ 502 vm_page_set_valid_range(m, base, size); 503 504 /* 505 * Round "base" to the next block boundary so that the 506 * dirty bit for a partially zeroed block is not 507 * cleared. 508 */ 509 base = roundup2(base, DEV_BSIZE); 510 511 /* 512 * Clear out partial-page dirty bits. 513 * 514 * note that we do not clear out the valid 515 * bits. This would prevent bogus_page 516 * replacement from working properly. 517 */ 518 vm_page_clear_dirty(m, base, PAGE_SIZE - base); 519 } 520 vm_page_xunbusy(m); 521 } 522 out: 523 #if defined(__powerpc__) && !defined(__powerpc64__) 524 object->un_pager.vnp.vnp_size = nsize; 525 #else 526 atomic_store_64(&object->un_pager.vnp.vnp_size, nsize); 527 #endif 528 object->size = nobjsize; 529 VM_OBJECT_WUNLOCK(object); 530 } 531 532 /* 533 * calculate the linear (byte) disk address of specified virtual 534 * file address 535 */ 536 static int 537 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress, 538 int *run) 539 { 540 int bsize; 541 int err; 542 daddr_t vblock; 543 daddr_t voffset; 544 545 if (address < 0) 546 return -1; 547 548 if (VN_IS_DOOMED(vp)) 549 return -1; 550 551 bsize = vp->v_mount->mnt_stat.f_iosize; 552 vblock = address / bsize; 553 voffset = address % bsize; 554 555 err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL); 556 if (err == 0) { 557 if (*rtaddress != -1) 558 *rtaddress += voffset / DEV_BSIZE; 559 if (run) { 560 *run += 1; 561 *run *= bsize / PAGE_SIZE; 562 *run -= voffset / PAGE_SIZE; 563 } 564 } 565 566 return (err); 567 } 568 569 /* 570 * small block filesystem vnode pager input 571 */ 572 static int 573 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m) 574 { 575 struct vnode *vp; 576 struct bufobj *bo; 577 struct buf *bp; 578 struct sf_buf *sf; 579 daddr_t fileaddr; 580 vm_offset_t bsize; 581 vm_page_bits_t bits; 582 int error, i; 583 584 error = 0; 585 vp = object->handle; 586 if (VN_IS_DOOMED(vp)) 587 return VM_PAGER_BAD; 588 589 bsize = vp->v_mount->mnt_stat.f_iosize; 590 591 VOP_BMAP(vp, 0, &bo, 0, NULL, NULL); 592 593 sf = sf_buf_alloc(m, 0); 594 595 for (i = 0; i < PAGE_SIZE / bsize; i++) { 596 vm_ooffset_t address; 597 598 bits = vm_page_bits(i * bsize, bsize); 599 if (m->valid & bits) 600 continue; 601 602 address = IDX_TO_OFF(m->pindex) + i * bsize; 603 if (address >= object->un_pager.vnp.vnp_size) { 604 fileaddr = -1; 605 } else { 606 error = vnode_pager_addr(vp, address, &fileaddr, NULL); 607 if (error) 608 break; 609 } 610 if (fileaddr != -1) { 611 bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK); 612 613 /* build a minimal buffer header */ 614 bp->b_iocmd = BIO_READ; 615 bp->b_iodone = bdone; 616 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 617 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 618 bp->b_rcred = crhold(curthread->td_ucred); 619 bp->b_wcred = crhold(curthread->td_ucred); 620 bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize; 621 bp->b_blkno = fileaddr; 622 pbgetbo(bo, bp); 623 bp->b_vp = vp; 624 bp->b_bcount = bsize; 625 bp->b_bufsize = bsize; 626 bp->b_runningbufspace = bp->b_bufsize; 627 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 628 629 /* do the input */ 630 bp->b_iooffset = dbtob(bp->b_blkno); 631 bstrategy(bp); 632 633 bwait(bp, PVM, "vnsrd"); 634 635 if ((bp->b_ioflags & BIO_ERROR) != 0) { 636 KASSERT(bp->b_error != 0, 637 ("%s: buf error but b_error == 0\n", __func__)); 638 error = bp->b_error; 639 } 640 641 /* 642 * free the buffer header back to the swap buffer pool 643 */ 644 bp->b_vp = NULL; 645 pbrelbo(bp); 646 uma_zfree(vnode_pbuf_zone, bp); 647 if (error) 648 break; 649 } else 650 bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize); 651 KASSERT((m->dirty & bits) == 0, 652 ("vnode_pager_input_smlfs: page %p is dirty", m)); 653 vm_page_bits_set(m, &m->valid, bits); 654 } 655 sf_buf_free(sf); 656 if (error) { 657 return VM_PAGER_ERROR; 658 } 659 return VM_PAGER_OK; 660 } 661 662 /* 663 * old style vnode pager input routine 664 */ 665 static int 666 vnode_pager_input_old(vm_object_t object, vm_page_t m) 667 { 668 struct uio auio; 669 struct iovec aiov; 670 int error; 671 int size; 672 struct sf_buf *sf; 673 struct vnode *vp; 674 675 VM_OBJECT_ASSERT_WLOCKED(object); 676 error = 0; 677 678 /* 679 * Return failure if beyond current EOF 680 */ 681 if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { 682 return VM_PAGER_BAD; 683 } else { 684 size = PAGE_SIZE; 685 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) 686 size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); 687 vp = object->handle; 688 VM_OBJECT_WUNLOCK(object); 689 690 /* 691 * Allocate a kernel virtual address and initialize so that 692 * we can use VOP_READ/WRITE routines. 693 */ 694 sf = sf_buf_alloc(m, 0); 695 696 aiov.iov_base = (caddr_t)sf_buf_kva(sf); 697 aiov.iov_len = size; 698 auio.uio_iov = &aiov; 699 auio.uio_iovcnt = 1; 700 auio.uio_offset = IDX_TO_OFF(m->pindex); 701 auio.uio_segflg = UIO_SYSSPACE; 702 auio.uio_rw = UIO_READ; 703 auio.uio_resid = size; 704 auio.uio_td = curthread; 705 706 error = VOP_READ(vp, &auio, 0, curthread->td_ucred); 707 if (!error) { 708 int count = size - auio.uio_resid; 709 710 if (count == 0) 711 error = EINVAL; 712 else if (count != PAGE_SIZE) 713 bzero((caddr_t)sf_buf_kva(sf) + count, 714 PAGE_SIZE - count); 715 } 716 sf_buf_free(sf); 717 718 VM_OBJECT_WLOCK(object); 719 } 720 KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m)); 721 if (!error) 722 vm_page_valid(m); 723 return error ? VM_PAGER_ERROR : VM_PAGER_OK; 724 } 725 726 /* 727 * generic vnode pager input routine 728 */ 729 730 /* 731 * Local media VFS's that do not implement their own VOP_GETPAGES 732 * should have their VOP_GETPAGES call to vnode_pager_generic_getpages() 733 * to implement the previous behaviour. 734 * 735 * All other FS's should use the bypass to get to the local media 736 * backing vp's VOP_GETPAGES. 737 */ 738 static int 739 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind, 740 int *rahead) 741 { 742 struct vnode *vp; 743 int rtval; 744 745 /* Handle is stable with paging in progress. */ 746 vp = object->handle; 747 rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead); 748 KASSERT(rtval != EOPNOTSUPP, 749 ("vnode_pager: FS getpages not implemented\n")); 750 return rtval; 751 } 752 753 static int 754 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count, 755 int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg) 756 { 757 struct vnode *vp; 758 int rtval; 759 760 vp = object->handle; 761 rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg); 762 KASSERT(rtval != EOPNOTSUPP, 763 ("vnode_pager: FS getpages_async not implemented\n")); 764 return (rtval); 765 } 766 767 /* 768 * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for 769 * local filesystems, where partially valid pages can only occur at 770 * the end of file. 771 */ 772 int 773 vnode_pager_local_getpages(struct vop_getpages_args *ap) 774 { 775 776 return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count, 777 ap->a_rbehind, ap->a_rahead, NULL, NULL)); 778 } 779 780 int 781 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap) 782 { 783 int error; 784 785 error = vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count, 786 ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg); 787 if (error != 0 && ap->a_iodone != NULL) 788 ap->a_iodone(ap->a_arg, ap->a_m, ap->a_count, error); 789 return (error); 790 } 791 792 /* 793 * This is now called from local media FS's to operate against their 794 * own vnodes if they fail to implement VOP_GETPAGES. 795 */ 796 int 797 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count, 798 int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg) 799 { 800 vm_object_t object; 801 struct bufobj *bo; 802 struct buf *bp; 803 off_t foff; 804 #ifdef INVARIANTS 805 off_t blkno0; 806 #endif 807 int bsize, pagesperblock; 808 int error, before, after, rbehind, rahead, poff, i; 809 int bytecount, secmask; 810 811 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 812 ("%s does not support devices", __func__)); 813 814 if (VN_IS_DOOMED(vp)) 815 return (VM_PAGER_BAD); 816 817 object = vp->v_object; 818 foff = IDX_TO_OFF(m[0]->pindex); 819 bsize = vp->v_mount->mnt_stat.f_iosize; 820 pagesperblock = bsize / PAGE_SIZE; 821 822 KASSERT(foff < object->un_pager.vnp.vnp_size, 823 ("%s: page %p offset beyond vp %p size", __func__, m[0], vp)); 824 KASSERT(count <= nitems(bp->b_pages), 825 ("%s: requested %d pages", __func__, count)); 826 827 /* 828 * The last page has valid blocks. Invalid part can only 829 * exist at the end of file, and the page is made fully valid 830 * by zeroing in vm_pager_get_pages(). 831 */ 832 if (!vm_page_none_valid(m[count - 1]) && --count == 0) { 833 if (iodone != NULL) 834 iodone(arg, m, 1, 0); 835 return (VM_PAGER_OK); 836 } 837 838 bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK); 839 840 /* 841 * Get the underlying device blocks for the file with VOP_BMAP(). 842 * If the file system doesn't support VOP_BMAP, use old way of 843 * getting pages via VOP_READ. 844 */ 845 error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before); 846 if (error == EOPNOTSUPP) { 847 uma_zfree(vnode_pbuf_zone, bp); 848 VM_OBJECT_WLOCK(object); 849 for (i = 0; i < count; i++) { 850 VM_CNT_INC(v_vnodein); 851 VM_CNT_INC(v_vnodepgsin); 852 error = vnode_pager_input_old(object, m[i]); 853 if (error) 854 break; 855 } 856 VM_OBJECT_WUNLOCK(object); 857 return (error); 858 } else if (error != 0) { 859 uma_zfree(vnode_pbuf_zone, bp); 860 return (VM_PAGER_ERROR); 861 } 862 863 /* 864 * If the file system supports BMAP, but blocksize is smaller 865 * than a page size, then use special small filesystem code. 866 */ 867 if (pagesperblock == 0) { 868 uma_zfree(vnode_pbuf_zone, bp); 869 for (i = 0; i < count; i++) { 870 VM_CNT_INC(v_vnodein); 871 VM_CNT_INC(v_vnodepgsin); 872 error = vnode_pager_input_smlfs(object, m[i]); 873 if (error) 874 break; 875 } 876 return (error); 877 } 878 879 /* 880 * A sparse file can be encountered only for a single page request, 881 * which may not be preceded by call to vm_pager_haspage(). 882 */ 883 if (bp->b_blkno == -1) { 884 KASSERT(count == 1, 885 ("%s: array[%d] request to a sparse file %p", __func__, 886 count, vp)); 887 uma_zfree(vnode_pbuf_zone, bp); 888 pmap_zero_page(m[0]); 889 KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty", 890 __func__, m[0])); 891 vm_page_valid(m[0]); 892 return (VM_PAGER_OK); 893 } 894 895 #ifdef INVARIANTS 896 blkno0 = bp->b_blkno; 897 #endif 898 bp->b_blkno += (foff % bsize) / DEV_BSIZE; 899 900 /* Recalculate blocks available after/before to pages. */ 901 poff = (foff % bsize) / PAGE_SIZE; 902 before *= pagesperblock; 903 before += poff; 904 after *= pagesperblock; 905 after += pagesperblock - (poff + 1); 906 if (m[0]->pindex + after >= object->size) 907 after = object->size - 1 - m[0]->pindex; 908 KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d", 909 __func__, count, after + 1)); 910 after -= count - 1; 911 912 /* Trim requested rbehind/rahead to possible values. */ 913 rbehind = a_rbehind ? *a_rbehind : 0; 914 rahead = a_rahead ? *a_rahead : 0; 915 rbehind = min(rbehind, before); 916 rbehind = min(rbehind, m[0]->pindex); 917 rahead = min(rahead, after); 918 rahead = min(rahead, object->size - m[count - 1]->pindex); 919 /* 920 * Check that total amount of pages fit into buf. Trim rbehind and 921 * rahead evenly if not. 922 */ 923 if (rbehind + rahead + count > nitems(bp->b_pages)) { 924 int trim, sum; 925 926 trim = rbehind + rahead + count - nitems(bp->b_pages) + 1; 927 sum = rbehind + rahead; 928 if (rbehind == before) { 929 /* Roundup rbehind trim to block size. */ 930 rbehind -= roundup(trim * rbehind / sum, pagesperblock); 931 if (rbehind < 0) 932 rbehind = 0; 933 } else 934 rbehind -= trim * rbehind / sum; 935 rahead -= trim * rahead / sum; 936 } 937 KASSERT(rbehind + rahead + count <= nitems(bp->b_pages), 938 ("%s: behind %d ahead %d count %d", __func__, 939 rbehind, rahead, count)); 940 941 /* 942 * Fill in the bp->b_pages[] array with requested and optional 943 * read behind or read ahead pages. Read behind pages are looked 944 * up in a backward direction, down to a first cached page. Same 945 * for read ahead pages, but there is no need to shift the array 946 * in case of encountering a cached page. 947 */ 948 i = bp->b_npages = 0; 949 if (rbehind) { 950 vm_pindex_t startpindex, tpindex; 951 vm_page_t p; 952 953 VM_OBJECT_WLOCK(object); 954 startpindex = m[0]->pindex - rbehind; 955 if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL && 956 p->pindex >= startpindex) 957 startpindex = p->pindex + 1; 958 959 /* tpindex is unsigned; beware of numeric underflow. */ 960 for (tpindex = m[0]->pindex - 1; 961 tpindex >= startpindex && tpindex < m[0]->pindex; 962 tpindex--, i++) { 963 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 964 if (p == NULL) { 965 /* Shift the array. */ 966 for (int j = 0; j < i; j++) 967 bp->b_pages[j] = bp->b_pages[j + 968 tpindex + 1 - startpindex]; 969 break; 970 } 971 bp->b_pages[tpindex - startpindex] = p; 972 } 973 974 bp->b_pgbefore = i; 975 bp->b_npages += i; 976 bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE; 977 } else 978 bp->b_pgbefore = 0; 979 980 /* Requested pages. */ 981 for (int j = 0; j < count; j++, i++) 982 bp->b_pages[i] = m[j]; 983 bp->b_npages += count; 984 985 if (rahead) { 986 vm_pindex_t endpindex, tpindex; 987 vm_page_t p; 988 989 if (!VM_OBJECT_WOWNED(object)) 990 VM_OBJECT_WLOCK(object); 991 endpindex = m[count - 1]->pindex + rahead + 1; 992 if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL && 993 p->pindex < endpindex) 994 endpindex = p->pindex; 995 if (endpindex > object->size) 996 endpindex = object->size; 997 998 for (tpindex = m[count - 1]->pindex + 1; 999 tpindex < endpindex; i++, tpindex++) { 1000 p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1001 if (p == NULL) 1002 break; 1003 bp->b_pages[i] = p; 1004 } 1005 1006 bp->b_pgafter = i - bp->b_npages; 1007 bp->b_npages = i; 1008 } else 1009 bp->b_pgafter = 0; 1010 1011 if (VM_OBJECT_WOWNED(object)) 1012 VM_OBJECT_WUNLOCK(object); 1013 1014 /* Report back actual behind/ahead read. */ 1015 if (a_rbehind) 1016 *a_rbehind = bp->b_pgbefore; 1017 if (a_rahead) 1018 *a_rahead = bp->b_pgafter; 1019 1020 #ifdef INVARIANTS 1021 KASSERT(bp->b_npages <= nitems(bp->b_pages), 1022 ("%s: buf %p overflowed", __func__, bp)); 1023 for (int j = 1, prev = 0; j < bp->b_npages; j++) { 1024 if (bp->b_pages[j] == bogus_page) 1025 continue; 1026 KASSERT(bp->b_pages[j]->pindex - bp->b_pages[prev]->pindex == 1027 j - prev, ("%s: pages array not consecutive, bp %p", 1028 __func__, bp)); 1029 prev = j; 1030 } 1031 #endif 1032 1033 /* 1034 * Recalculate first offset and bytecount with regards to read behind. 1035 * Truncate bytecount to vnode real size and round up physical size 1036 * for real devices. 1037 */ 1038 foff = IDX_TO_OFF(bp->b_pages[0]->pindex); 1039 bytecount = bp->b_npages << PAGE_SHIFT; 1040 if ((foff + bytecount) > object->un_pager.vnp.vnp_size) 1041 bytecount = object->un_pager.vnp.vnp_size - foff; 1042 secmask = bo->bo_bsize - 1; 1043 KASSERT(secmask < PAGE_SIZE && secmask > 0, 1044 ("%s: sector size %d too large", __func__, secmask + 1)); 1045 bytecount = (bytecount + secmask) & ~secmask; 1046 1047 /* 1048 * And map the pages to be read into the kva, if the filesystem 1049 * requires mapped buffers. 1050 */ 1051 if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 && 1052 unmapped_buf_allowed) { 1053 bp->b_data = unmapped_buf; 1054 bp->b_offset = 0; 1055 } else { 1056 bp->b_data = bp->b_kvabase; 1057 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 1058 } 1059 1060 /* Build a minimal buffer header. */ 1061 bp->b_iocmd = BIO_READ; 1062 KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); 1063 KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); 1064 bp->b_rcred = crhold(curthread->td_ucred); 1065 bp->b_wcred = crhold(curthread->td_ucred); 1066 pbgetbo(bo, bp); 1067 bp->b_vp = vp; 1068 bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount; 1069 bp->b_iooffset = dbtob(bp->b_blkno); 1070 KASSERT(IDX_TO_OFF(m[0]->pindex - bp->b_pages[0]->pindex) == 1071 (blkno0 - bp->b_blkno) * DEV_BSIZE + 1072 IDX_TO_OFF(m[0]->pindex) % bsize, 1073 ("wrong offsets bsize %d m[0] %ju b_pages[0] %ju " 1074 "blkno0 %ju b_blkno %ju", bsize, 1075 (uintmax_t)m[0]->pindex, (uintmax_t)bp->b_pages[0]->pindex, 1076 (uintmax_t)blkno0, (uintmax_t)bp->b_blkno)); 1077 1078 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 1079 VM_CNT_INC(v_vnodein); 1080 VM_CNT_ADD(v_vnodepgsin, bp->b_npages); 1081 1082 if (iodone != NULL) { /* async */ 1083 bp->b_pgiodone = iodone; 1084 bp->b_caller1 = arg; 1085 bp->b_iodone = vnode_pager_generic_getpages_done_async; 1086 bp->b_flags |= B_ASYNC; 1087 BUF_KERNPROC(bp); 1088 bstrategy(bp); 1089 return (VM_PAGER_OK); 1090 } else { 1091 bp->b_iodone = bdone; 1092 bstrategy(bp); 1093 bwait(bp, PVM, "vnread"); 1094 error = vnode_pager_generic_getpages_done(bp); 1095 for (i = 0; i < bp->b_npages; i++) 1096 bp->b_pages[i] = NULL; 1097 bp->b_vp = NULL; 1098 pbrelbo(bp); 1099 uma_zfree(vnode_pbuf_zone, bp); 1100 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); 1101 } 1102 } 1103 1104 static void 1105 vnode_pager_generic_getpages_done_async(struct buf *bp) 1106 { 1107 int error; 1108 1109 error = vnode_pager_generic_getpages_done(bp); 1110 /* Run the iodone upon the requested range. */ 1111 bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore, 1112 bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error); 1113 for (int i = 0; i < bp->b_npages; i++) 1114 bp->b_pages[i] = NULL; 1115 bp->b_vp = NULL; 1116 pbrelbo(bp); 1117 uma_zfree(vnode_pbuf_zone, bp); 1118 } 1119 1120 static int 1121 vnode_pager_generic_getpages_done(struct buf *bp) 1122 { 1123 vm_object_t object; 1124 off_t tfoff, nextoff; 1125 int i, error; 1126 1127 KASSERT((bp->b_ioflags & BIO_ERROR) == 0 || bp->b_error != 0, 1128 ("%s: buf error but b_error == 0\n", __func__)); 1129 error = (bp->b_ioflags & BIO_ERROR) != 0 ? bp->b_error : 0; 1130 object = bp->b_vp->v_object; 1131 1132 if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) { 1133 if (!buf_mapped(bp)) { 1134 bp->b_data = bp->b_kvabase; 1135 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, 1136 bp->b_npages); 1137 } 1138 bzero(bp->b_data + bp->b_bcount, 1139 PAGE_SIZE * bp->b_npages - bp->b_bcount); 1140 } 1141 if (buf_mapped(bp)) { 1142 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1143 bp->b_data = unmapped_buf; 1144 } 1145 1146 /* 1147 * If the read failed, we must free any read ahead/behind pages here. 1148 * The requested pages are freed by the caller (for sync requests) 1149 * or by the bp->b_pgiodone callback (for async requests). 1150 */ 1151 if (error != 0) { 1152 VM_OBJECT_WLOCK(object); 1153 for (i = 0; i < bp->b_pgbefore; i++) 1154 vm_page_free_invalid(bp->b_pages[i]); 1155 for (i = bp->b_npages - bp->b_pgafter; i < bp->b_npages; i++) 1156 vm_page_free_invalid(bp->b_pages[i]); 1157 VM_OBJECT_WUNLOCK(object); 1158 return (error); 1159 } 1160 1161 /* Read lock to protect size. */ 1162 VM_OBJECT_RLOCK(object); 1163 for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex); 1164 i < bp->b_npages; i++, tfoff = nextoff) { 1165 vm_page_t mt; 1166 1167 nextoff = tfoff + PAGE_SIZE; 1168 mt = bp->b_pages[i]; 1169 if (mt == bogus_page) 1170 continue; 1171 1172 if (nextoff <= object->un_pager.vnp.vnp_size) { 1173 /* 1174 * Read filled up entire page. 1175 */ 1176 vm_page_valid(mt); 1177 KASSERT(mt->dirty == 0, 1178 ("%s: page %p is dirty", __func__, mt)); 1179 KASSERT(!pmap_page_is_mapped(mt), 1180 ("%s: page %p is mapped", __func__, mt)); 1181 } else { 1182 /* 1183 * Read did not fill up entire page. 1184 * 1185 * Currently we do not set the entire page valid, 1186 * we just try to clear the piece that we couldn't 1187 * read. 1188 */ 1189 vm_page_set_valid_range(mt, 0, 1190 object->un_pager.vnp.vnp_size - tfoff); 1191 KASSERT((mt->dirty & vm_page_bits(0, 1192 object->un_pager.vnp.vnp_size - tfoff)) == 0, 1193 ("%s: page %p is dirty", __func__, mt)); 1194 } 1195 1196 if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter) 1197 vm_page_readahead_finish(mt); 1198 } 1199 VM_OBJECT_RUNLOCK(object); 1200 1201 return (error); 1202 } 1203 1204 /* 1205 * EOPNOTSUPP is no longer legal. For local media VFS's that do not 1206 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to 1207 * vnode_pager_generic_putpages() to implement the previous behaviour. 1208 * 1209 * All other FS's should use the bypass to get to the local media 1210 * backing vp's VOP_PUTPAGES. 1211 */ 1212 static void 1213 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1214 int flags, int *rtvals) 1215 { 1216 int rtval; 1217 struct vnode *vp; 1218 int bytes = count * PAGE_SIZE; 1219 1220 /* 1221 * Force synchronous operation if we are extremely low on memory 1222 * to prevent a low-memory deadlock. VOP operations often need to 1223 * allocate more memory to initiate the I/O ( i.e. do a BMAP 1224 * operation ). The swapper handles the case by limiting the amount 1225 * of asynchronous I/O, but that sort of solution doesn't scale well 1226 * for the vnode pager without a lot of work. 1227 * 1228 * Also, the backing vnode's iodone routine may not wake the pageout 1229 * daemon up. This should be probably be addressed XXX. 1230 */ 1231 1232 if (vm_page_count_min()) 1233 flags |= VM_PAGER_PUT_SYNC; 1234 1235 /* 1236 * Call device-specific putpages function 1237 */ 1238 vp = object->handle; 1239 VM_OBJECT_WUNLOCK(object); 1240 rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals); 1241 KASSERT(rtval != EOPNOTSUPP, 1242 ("vnode_pager: stale FS putpages\n")); 1243 VM_OBJECT_WLOCK(object); 1244 } 1245 1246 static int 1247 vn_off2bidx(vm_ooffset_t offset) 1248 { 1249 1250 return ((offset & PAGE_MASK) / DEV_BSIZE); 1251 } 1252 1253 static bool 1254 vn_dirty_blk(vm_page_t m, vm_ooffset_t offset) 1255 { 1256 1257 KASSERT(IDX_TO_OFF(m->pindex) <= offset && 1258 offset < IDX_TO_OFF(m->pindex + 1), 1259 ("page %p pidx %ju offset %ju", m, (uintmax_t)m->pindex, 1260 (uintmax_t)offset)); 1261 return ((m->dirty & ((vm_page_bits_t)1 << vn_off2bidx(offset))) != 0); 1262 } 1263 1264 /* 1265 * This is now called from local media FS's to operate against their 1266 * own vnodes if they fail to implement VOP_PUTPAGES. 1267 * 1268 * This is typically called indirectly via the pageout daemon and 1269 * clustering has already typically occurred, so in general we ask the 1270 * underlying filesystem to write the data out asynchronously rather 1271 * then delayed. 1272 */ 1273 int 1274 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount, 1275 int flags, int *rtvals) 1276 { 1277 vm_object_t object; 1278 vm_page_t m; 1279 vm_ooffset_t maxblksz, next_offset, poffset, prev_offset; 1280 struct uio auio; 1281 struct iovec aiov; 1282 off_t prev_resid, wrsz; 1283 int count, error, i, maxsize, ncount, pgoff, ppscheck; 1284 bool in_hole; 1285 static struct timeval lastfail; 1286 static int curfail; 1287 1288 object = vp->v_object; 1289 count = bytecount / PAGE_SIZE; 1290 1291 for (i = 0; i < count; i++) 1292 rtvals[i] = VM_PAGER_ERROR; 1293 1294 if ((int64_t)ma[0]->pindex < 0) { 1295 printf("vnode_pager_generic_putpages: " 1296 "attempt to write meta-data 0x%jx(%lx)\n", 1297 (uintmax_t)ma[0]->pindex, (u_long)ma[0]->dirty); 1298 rtvals[0] = VM_PAGER_BAD; 1299 return (VM_PAGER_BAD); 1300 } 1301 1302 maxsize = count * PAGE_SIZE; 1303 ncount = count; 1304 1305 poffset = IDX_TO_OFF(ma[0]->pindex); 1306 1307 /* 1308 * If the page-aligned write is larger then the actual file we 1309 * have to invalidate pages occurring beyond the file EOF. However, 1310 * there is an edge case where a file may not be page-aligned where 1311 * the last page is partially invalid. In this case the filesystem 1312 * may not properly clear the dirty bits for the entire page (which 1313 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). 1314 * With the page busied we are free to fix up the dirty bits here. 1315 * 1316 * We do not under any circumstances truncate the valid bits, as 1317 * this will screw up bogus page replacement. 1318 */ 1319 VM_OBJECT_RLOCK(object); 1320 if (maxsize + poffset > object->un_pager.vnp.vnp_size) { 1321 if (object->un_pager.vnp.vnp_size > poffset) { 1322 maxsize = object->un_pager.vnp.vnp_size - poffset; 1323 ncount = btoc(maxsize); 1324 if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { 1325 pgoff = roundup2(pgoff, DEV_BSIZE); 1326 1327 /* 1328 * If the page is busy and the following 1329 * conditions hold, then the page's dirty 1330 * field cannot be concurrently changed by a 1331 * pmap operation. 1332 */ 1333 m = ma[ncount - 1]; 1334 vm_page_assert_sbusied(m); 1335 KASSERT(!pmap_page_is_write_mapped(m), 1336 ("vnode_pager_generic_putpages: page %p is not read-only", m)); 1337 MPASS(m->dirty != 0); 1338 vm_page_clear_dirty(m, pgoff, PAGE_SIZE - 1339 pgoff); 1340 } 1341 } else { 1342 maxsize = 0; 1343 ncount = 0; 1344 } 1345 for (i = ncount; i < count; i++) 1346 rtvals[i] = VM_PAGER_BAD; 1347 } 1348 VM_OBJECT_RUNLOCK(object); 1349 1350 auio.uio_iov = &aiov; 1351 auio.uio_segflg = UIO_NOCOPY; 1352 auio.uio_rw = UIO_WRITE; 1353 auio.uio_td = NULL; 1354 maxblksz = roundup2(poffset + maxsize, DEV_BSIZE); 1355 1356 for (prev_offset = poffset; prev_offset < maxblksz;) { 1357 /* Skip clean blocks. */ 1358 for (in_hole = true; in_hole && prev_offset < maxblksz;) { 1359 m = ma[OFF_TO_IDX(prev_offset - poffset)]; 1360 for (i = vn_off2bidx(prev_offset); 1361 i < sizeof(vm_page_bits_t) * NBBY && 1362 prev_offset < maxblksz; i++) { 1363 if (vn_dirty_blk(m, prev_offset)) { 1364 in_hole = false; 1365 break; 1366 } 1367 prev_offset += DEV_BSIZE; 1368 } 1369 } 1370 if (in_hole) 1371 goto write_done; 1372 1373 /* Find longest run of dirty blocks. */ 1374 for (next_offset = prev_offset; next_offset < maxblksz;) { 1375 m = ma[OFF_TO_IDX(next_offset - poffset)]; 1376 for (i = vn_off2bidx(next_offset); 1377 i < sizeof(vm_page_bits_t) * NBBY && 1378 next_offset < maxblksz; i++) { 1379 if (!vn_dirty_blk(m, next_offset)) 1380 goto start_write; 1381 next_offset += DEV_BSIZE; 1382 } 1383 } 1384 start_write: 1385 if (next_offset > poffset + maxsize) 1386 next_offset = poffset + maxsize; 1387 1388 /* 1389 * Getting here requires finding a dirty block in the 1390 * 'skip clean blocks' loop. 1391 */ 1392 MPASS(prev_offset < next_offset); 1393 1394 aiov.iov_base = NULL; 1395 auio.uio_iovcnt = 1; 1396 auio.uio_offset = prev_offset; 1397 prev_resid = auio.uio_resid = aiov.iov_len = next_offset - 1398 prev_offset; 1399 error = VOP_WRITE(vp, &auio, 1400 vnode_pager_putpages_ioflags(flags), curthread->td_ucred); 1401 1402 wrsz = prev_resid - auio.uio_resid; 1403 if (wrsz == 0) { 1404 if (ppsratecheck(&lastfail, &curfail, 1) != 0) { 1405 vn_printf(vp, "vnode_pager_putpages: " 1406 "zero-length write at %ju resid %zd\n", 1407 auio.uio_offset, auio.uio_resid); 1408 } 1409 break; 1410 } 1411 1412 /* Adjust the starting offset for next iteration. */ 1413 prev_offset += wrsz; 1414 MPASS(auio.uio_offset == prev_offset); 1415 1416 ppscheck = 0; 1417 if (error != 0 && (ppscheck = ppsratecheck(&lastfail, 1418 &curfail, 1)) != 0) 1419 vn_printf(vp, "vnode_pager_putpages: I/O error %d\n", 1420 error); 1421 if (auio.uio_resid != 0 && (ppscheck != 0 || 1422 ppsratecheck(&lastfail, &curfail, 1) != 0)) 1423 vn_printf(vp, "vnode_pager_putpages: residual I/O %zd " 1424 "at %ju\n", auio.uio_resid, 1425 (uintmax_t)ma[0]->pindex); 1426 if (error != 0 || auio.uio_resid != 0) 1427 break; 1428 } 1429 write_done: 1430 /* Mark completely processed pages. */ 1431 for (i = 0; i < OFF_TO_IDX(prev_offset - poffset); i++) 1432 rtvals[i] = VM_PAGER_OK; 1433 /* Mark partial EOF page. */ 1434 if (prev_offset == poffset + maxsize && (prev_offset & PAGE_MASK) != 0) 1435 rtvals[i++] = VM_PAGER_OK; 1436 /* Unwritten pages in range, free bonus if the page is clean. */ 1437 for (; i < ncount; i++) 1438 rtvals[i] = ma[i]->dirty == 0 ? VM_PAGER_OK : VM_PAGER_ERROR; 1439 VM_CNT_ADD(v_vnodepgsout, i); 1440 VM_CNT_INC(v_vnodeout); 1441 return (rtvals[0]); 1442 } 1443 1444 int 1445 vnode_pager_putpages_ioflags(int pager_flags) 1446 { 1447 int ioflags; 1448 1449 /* 1450 * Pageouts are already clustered, use IO_ASYNC to force a 1451 * bawrite() rather then a bdwrite() to prevent paging I/O 1452 * from saturating the buffer cache. Dummy-up the sequential 1453 * heuristic to cause large ranges to cluster. If neither 1454 * IO_SYNC or IO_ASYNC is set, the system decides how to 1455 * cluster. 1456 */ 1457 ioflags = IO_VMIO; 1458 if ((pager_flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) != 0) 1459 ioflags |= IO_SYNC; 1460 else if ((pager_flags & VM_PAGER_CLUSTER_OK) == 0) 1461 ioflags |= IO_ASYNC; 1462 ioflags |= (pager_flags & VM_PAGER_PUT_INVAL) != 0 ? IO_INVAL: 0; 1463 ioflags |= (pager_flags & VM_PAGER_PUT_NOREUSE) != 0 ? IO_NOREUSE : 0; 1464 ioflags |= IO_SEQMAX << IO_SEQSHIFT; 1465 return (ioflags); 1466 } 1467 1468 /* 1469 * vnode_pager_undirty_pages(). 1470 * 1471 * A helper to mark pages as clean after pageout that was possibly 1472 * done with a short write. The lpos argument specifies the page run 1473 * length in bytes, and the written argument specifies how many bytes 1474 * were actually written. eof is the offset past the last valid byte 1475 * in the vnode using the absolute file position of the first byte in 1476 * the run as the base from which it is computed. 1477 */ 1478 void 1479 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written, off_t eof, 1480 int lpos) 1481 { 1482 vm_object_t obj; 1483 int i, pos, pos_devb; 1484 1485 if (written == 0 && eof >= lpos) 1486 return; 1487 obj = ma[0]->object; 1488 for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) { 1489 if (pos < trunc_page(written)) { 1490 rtvals[i] = VM_PAGER_OK; 1491 vm_page_undirty(ma[i]); 1492 } else { 1493 /* Partially written page. */ 1494 rtvals[i] = VM_PAGER_AGAIN; 1495 vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK); 1496 } 1497 } 1498 if (eof >= lpos) /* avoid truncation */ 1499 return; 1500 for (pos = eof, i = OFF_TO_IDX(trunc_page(pos)); pos < lpos; i++) { 1501 if (pos != trunc_page(pos)) { 1502 /* 1503 * The page contains the last valid byte in 1504 * the vnode, mark the rest of the page as 1505 * clean, potentially making the whole page 1506 * clean. 1507 */ 1508 pos_devb = roundup2(pos & PAGE_MASK, DEV_BSIZE); 1509 vm_page_clear_dirty(ma[i], pos_devb, PAGE_SIZE - 1510 pos_devb); 1511 1512 /* 1513 * If the page was cleaned, report the pageout 1514 * on it as successful. msync() no longer 1515 * needs to write out the page, endlessly 1516 * creating write requests and dirty buffers. 1517 */ 1518 if (ma[i]->dirty == 0) 1519 rtvals[i] = VM_PAGER_OK; 1520 1521 pos = round_page(pos); 1522 } else { 1523 /* vm_pageout_flush() clears dirty */ 1524 rtvals[i] = VM_PAGER_BAD; 1525 pos += PAGE_SIZE; 1526 } 1527 } 1528 } 1529 1530 static void 1531 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start, 1532 vm_offset_t end) 1533 { 1534 struct vnode *vp; 1535 vm_ooffset_t old_wm; 1536 1537 VM_OBJECT_WLOCK(object); 1538 if (object->type != OBJT_VNODE) { 1539 VM_OBJECT_WUNLOCK(object); 1540 return; 1541 } 1542 old_wm = object->un_pager.vnp.writemappings; 1543 object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start; 1544 vp = object->handle; 1545 if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) { 1546 ASSERT_VOP_LOCKED(vp, "v_writecount inc"); 1547 VOP_ADD_WRITECOUNT_CHECKED(vp, 1); 1548 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", 1549 __func__, vp, vp->v_writecount); 1550 } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) { 1551 ASSERT_VOP_LOCKED(vp, "v_writecount dec"); 1552 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 1553 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 1554 __func__, vp, vp->v_writecount); 1555 } 1556 VM_OBJECT_WUNLOCK(object); 1557 } 1558 1559 static void 1560 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start, 1561 vm_offset_t end) 1562 { 1563 struct vnode *vp; 1564 struct mount *mp; 1565 vm_offset_t inc; 1566 1567 VM_OBJECT_WLOCK(object); 1568 1569 /* 1570 * First, recheck the object type to account for the race when 1571 * the vnode is reclaimed. 1572 */ 1573 if (object->type != OBJT_VNODE) { 1574 VM_OBJECT_WUNLOCK(object); 1575 return; 1576 } 1577 1578 /* 1579 * Optimize for the case when writemappings is not going to 1580 * zero. 1581 */ 1582 inc = end - start; 1583 if (object->un_pager.vnp.writemappings != inc) { 1584 object->un_pager.vnp.writemappings -= inc; 1585 VM_OBJECT_WUNLOCK(object); 1586 return; 1587 } 1588 1589 vp = object->handle; 1590 vhold(vp); 1591 VM_OBJECT_WUNLOCK(object); 1592 mp = NULL; 1593 vn_start_write(vp, &mp, V_WAIT); 1594 vn_lock(vp, LK_SHARED | LK_RETRY); 1595 1596 /* 1597 * Decrement the object's writemappings, by swapping the start 1598 * and end arguments for vnode_pager_update_writecount(). If 1599 * there was not a race with vnode reclaimation, then the 1600 * vnode's v_writecount is decremented. 1601 */ 1602 vnode_pager_update_writecount(object, end, start); 1603 VOP_UNLOCK(vp); 1604 vdrop(vp); 1605 if (mp != NULL) 1606 vn_finished_write(mp); 1607 } 1608