sys/vm/vnode_pager.c

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