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 #include "opt_vm.h"
  57
  58 #include <sys/param.h>
  59 #include <sys/kernel.h>
  60 #include <sys/systm.h>
  61 #include <sys/sysctl.h>
  62 #include <sys/proc.h>
  63 #include <sys/vnode.h>
  64 #include <sys/mount.h>
  65 #include <sys/bio.h>
  66 #include <sys/buf.h>
  67 #include <sys/vmmeter.h>
  68 #include <sys/ktr.h>
  69 #include <sys/limits.h>
  70 #include <sys/conf.h>
  71 #include <sys/refcount.h>
  72 #include <sys/rwlock.h>
  73 #include <sys/sf_buf.h>
  74 #include <sys/domainset.h>
  75 #include <sys/user.h>
  76
  77 #include <machine/atomic.h>
  78
  79 #include <vm/vm.h>
  80 #include <vm/vm_param.h>
  81 #include <vm/vm_object.h>
  82 #include <vm/vm_page.h>
  83 #include <vm/vm_pager.h>
  84 #include <vm/vm_map.h>
  85 #include <vm/vnode_pager.h>
  86 #include <vm/vm_extern.h>
  87 #include <vm/uma.h>
  88
  89 static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
  90     daddr_t *rtaddress, int *run);
  91 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
  92 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
  93 static void vnode_pager_dealloc(vm_object_t);
  94 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *);
  95 static int vnode_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
  96     int *, vop_getpages_iodone_t, void *);
  97 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, int, int *);
  98 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
  99 static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t,
 100     vm_ooffset_t, struct ucred *cred);
 101 static int vnode_pager_generic_getpages_done(struct buf *);
 102 static void vnode_pager_generic_getpages_done_async(struct buf *);
 103 static void vnode_pager_update_writecount(vm_object_t, vm_offset_t,
 104     vm_offset_t);
 105 static void vnode_pager_release_writecount(vm_object_t, vm_offset_t,
 106     vm_offset_t);
 107 static void vnode_pager_getvp(vm_object_t, struct vnode **, bool *);
 108
 109 const struct pagerops vnodepagerops = {
 110         .pgo_kvme_type = KVME_TYPE_VNODE,
 111         .pgo_alloc =    vnode_pager_alloc,
 112         .pgo_dealloc =  vnode_pager_dealloc,
 113         .pgo_getpages = vnode_pager_getpages,
 114         .pgo_getpages_async = vnode_pager_getpages_async,
 115         .pgo_putpages = vnode_pager_putpages,
 116         .pgo_haspage =  vnode_pager_haspage,
 117         .pgo_update_writecount = vnode_pager_update_writecount,
 118         .pgo_release_writecount = vnode_pager_release_writecount,
 119         .pgo_set_writeable_dirty = vm_object_set_writeable_dirty_,
 120         .pgo_mightbedirty = vm_object_mightbedirty_,
 121         .pgo_getvp = vnode_pager_getvp,
 122 };
 123
 124 static struct domainset *vnode_domainset = NULL;
 125
 126 SYSCTL_PROC(_debug, OID_AUTO, vnode_domainset,
 127     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_RW, &vnode_domainset, 0,
 128     sysctl_handle_domainset, "A", "Default vnode NUMA policy");
 129
 130 static int nvnpbufs;
 131 SYSCTL_INT(_vm, OID_AUTO, vnode_pbufs, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
 132     &nvnpbufs, 0, "number of physical buffers allocated for vnode pager");
 133
 134 static uma_zone_t vnode_pbuf_zone;
 135
 136 static void
 137 vnode_pager_init(void *dummy)
 138 {
 139
 140 #ifdef __LP64__
 141         nvnpbufs = nswbuf * 2;
 142 #else
 143         nvnpbufs = nswbuf / 2;
 144 #endif
 145         TUNABLE_INT_FETCH("vm.vnode_pbufs", &nvnpbufs);
 146         vnode_pbuf_zone = pbuf_zsecond_create("vnpbuf", nvnpbufs);
 147 }
 148 SYSINIT(vnode_pager, SI_SUB_CPU, SI_ORDER_ANY, vnode_pager_init, NULL);
 149
 150 /* Create the VM system backing object for this vnode */
 151 int
 152 vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td)
 153 {
 154         vm_object_t object;
 155         vm_ooffset_t size = isize;
 156         struct vattr va;
 157         bool last;
 158
 159         if (!vn_isdisk(vp) && vn_canvmio(vp) == FALSE)
 160                 return (0);
 161
 162         object = vp->v_object;
 163         if (object != NULL)
 164                 return (0);
 165
 166         if (size == 0) {
 167                 if (vn_isdisk(vp)) {
 168                         size = IDX_TO_OFF(INT_MAX);
 169                 } else {
 170                         if (VOP_GETATTR(vp, &va, td->td_ucred))
 171                                 return (0);
 172                         size = va.va_size;
 173                 }
 174         }
 175
 176         object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred);
 177         /*
 178          * Dereference the reference we just created.  This assumes
 179          * that the object is associated with the vp.  We still have
 180          * to serialize with vnode_pager_dealloc() for the last
 181          * potential reference.
 182          */
 183         VM_OBJECT_RLOCK(object);
 184         last = refcount_release(&object->ref_count);
 185         VM_OBJECT_RUNLOCK(object);
 186         if (last)
 187                 vrele(vp);
 188
 189         KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object"));
 190
 191         return (0);
 192 }
 193
 194 void
 195 vnode_destroy_vobject(struct vnode *vp)
 196 {
 197         struct vm_object *obj;
 198
 199         obj = vp->v_object;
 200         if (obj == NULL || obj->handle != vp)
 201                 return;
 202         ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject");
 203         VM_OBJECT_WLOCK(obj);
 204         MPASS(obj->type == OBJT_VNODE);
 205         umtx_shm_object_terminated(obj);
 206         if (obj->ref_count == 0) {
 207                 KASSERT((obj->flags & OBJ_DEAD) == 0,
 208                    ("vnode_destroy_vobject: Terminating dead object"));
 209                 vm_object_set_flag(obj, OBJ_DEAD);
 210
 211                 /*
 212                  * Clean pages and flush buffers.
 213                  */
 214                 vm_object_page_clean(obj, 0, 0, OBJPC_SYNC);
 215                 VM_OBJECT_WUNLOCK(obj);
 216
 217                 vinvalbuf(vp, V_SAVE, 0, 0);
 218
 219                 BO_LOCK(&vp->v_bufobj);
 220                 vp->v_bufobj.bo_flag |= BO_DEAD;
 221                 BO_UNLOCK(&vp->v_bufobj);
 222
 223                 VM_OBJECT_WLOCK(obj);
 224                 vm_object_terminate(obj);
 225         } else {
 226                 /*
 227                  * Woe to the process that tries to page now :-).
 228                  */
 229                 vm_pager_deallocate(obj);
 230                 VM_OBJECT_WUNLOCK(obj);
 231         }
 232         KASSERT(vp->v_object == NULL, ("vp %p obj %p", vp, vp->v_object));
 233 }
 234
 235 /*
 236  * Allocate (or lookup) pager for a vnode.
 237  * Handle is a vnode pointer.
 238  */
 239 vm_object_t
 240 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
 241     vm_ooffset_t offset, struct ucred *cred)
 242 {
 243         vm_object_t object;
 244         struct vnode *vp;
 245
 246         /*
 247          * Pageout to vnode, no can do yet.
 248          */
 249         if (handle == NULL)
 250                 return (NULL);
 251
 252         vp = (struct vnode *)handle;
 253         ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc");
 254         VNPASS(vp->v_usecount > 0, vp);
 255 retry:
 256         object = vp->v_object;
 257
 258         if (object == NULL) {
 259                 /*
 260                  * Add an object of the appropriate size
 261                  */
 262                 object = vm_object_allocate(OBJT_VNODE,
 263                     OFF_TO_IDX(round_page(size)));
 264
 265                 object->un_pager.vnp.vnp_size = size;
 266                 object->un_pager.vnp.writemappings = 0;
 267                 object->domain.dr_policy = vnode_domainset;
 268                 object->handle = handle;
 269                 if ((vp->v_vflag & VV_VMSIZEVNLOCK) != 0) {
 270                         VM_OBJECT_WLOCK(object);
 271                         vm_object_set_flag(object, OBJ_SIZEVNLOCK);
 272                         VM_OBJECT_WUNLOCK(object);
 273                 }
 274                 VI_LOCK(vp);
 275                 if (vp->v_object != NULL) {
 276                         /*
 277                          * Object has been created while we were allocating.
 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                         refcount_init(&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                 vrefact(vp);
 292         } else {
 293                 vm_object_reference(object);
 294 #if VM_NRESERVLEVEL > 0
 295                 if ((object->flags & OBJ_COLORED) == 0) {
 296                         VM_OBJECT_WLOCK(object);
 297                         vm_object_color(object, 0);
 298                         VM_OBJECT_WUNLOCK(object);
 299                 }
 300 #endif
 301         }
 302         return (object);
 303 }
 304
 305 /*
 306  *      The object must be locked.
 307  */
 308 static void
 309 vnode_pager_dealloc(vm_object_t object)
 310 {
 311         struct vnode *vp;
 312         int refs;
 313
 314         vp = object->handle;
 315         if (vp == NULL)
 316                 panic("vnode_pager_dealloc: pager already dealloced");
 317
 318         VM_OBJECT_ASSERT_WLOCKED(object);
 319         vm_object_pip_wait(object, "vnpdea");
 320         refs = object->ref_count;
 321
 322         object->handle = NULL;
 323         object->type = OBJT_DEAD;
 324         ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc");
 325         if (object->un_pager.vnp.writemappings > 0) {
 326                 object->un_pager.vnp.writemappings = 0;
 327                 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
 328                 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
 329                     __func__, vp, vp->v_writecount);
 330         }
 331         vp->v_object = NULL;
 332         VI_LOCK(vp);
 333
 334         /*
 335          * vm_map_entry_set_vnode_text() cannot reach this vnode by
 336          * following object->handle.  Clear all text references now.
 337          * This also clears the transient references from
 338          * kern_execve(), which is fine because dead_vnodeops uses nop
 339          * for VOP_UNSET_TEXT().
 340          */
 341         if (vp->v_writecount < 0)
 342                 vp->v_writecount = 0;
 343         VI_UNLOCK(vp);
 344         VM_OBJECT_WUNLOCK(object);
 345         if (refs > 0)
 346                 vunref(vp);
 347         VM_OBJECT_WLOCK(object);
 348 }
 349
 350 static boolean_t
 351 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
 352     int *after)
 353 {
 354         struct vnode *vp = object->handle;
 355         daddr_t bn;
 356         uintptr_t lockstate;
 357         int err;
 358         daddr_t reqblock;
 359         int poff;
 360         int bsize;
 361         int pagesperblock, blocksperpage;
 362
 363         VM_OBJECT_ASSERT_LOCKED(object);
 364         /*
 365          * If no vp or vp is doomed or marked transparent to VM, we do not
 366          * have the page.
 367          */
 368         if (vp == NULL || VN_IS_DOOMED(vp))
 369                 return FALSE;
 370         /*
 371          * If the offset is beyond end of file we do
 372          * not have the page.
 373          */
 374         if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size)
 375                 return FALSE;
 376
 377         bsize = vp->v_mount->mnt_stat.f_iosize;
 378         pagesperblock = bsize / PAGE_SIZE;
 379         blocksperpage = 0;
 380         if (pagesperblock > 0) {
 381                 reqblock = pindex / pagesperblock;
 382         } else {
 383                 blocksperpage = (PAGE_SIZE / bsize);
 384                 reqblock = pindex * blocksperpage;
 385         }
 386         lockstate = VM_OBJECT_DROP(object);
 387         err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
 388         VM_OBJECT_PICKUP(object, lockstate);
 389         if (err)
 390                 return TRUE;
 391         if (bn == -1)
 392                 return FALSE;
 393         if (pagesperblock > 0) {
 394                 poff = pindex - (reqblock * pagesperblock);
 395                 if (before) {
 396                         *before *= pagesperblock;
 397                         *before += poff;
 398                 }
 399                 if (after) {
 400                         /*
 401                          * The BMAP vop can report a partial block in the
 402                          * 'after', but must not report blocks after EOF.
 403                          * Assert the latter, and truncate 'after' in case
 404                          * of the former.
 405                          */
 406                         KASSERT((reqblock + *after) * pagesperblock <
 407                             roundup2(object->size, pagesperblock),
 408                             ("%s: reqblock %jd after %d size %ju", __func__,
 409                             (intmax_t )reqblock, *after,
 410                             (uintmax_t )object->size));
 411                         *after *= pagesperblock;
 412                         *after += pagesperblock - (poff + 1);
 413                         if (pindex + *after >= object->size)
 414                                 *after = object->size - 1 - pindex;
 415                 }
 416         } else {
 417                 if (before) {
 418                         *before /= blocksperpage;
 419                 }
 420
 421                 if (after) {
 422                         *after /= blocksperpage;
 423                 }
 424         }
 425         return TRUE;
 426 }
 427
 428 /*
 429  * Lets the VM system know about a change in size for a file.
 430  * We adjust our own internal size and flush any cached pages in
 431  * the associated object that are affected by the size change.
 432  *
 433  * Note: this routine may be invoked as a result of a pager put
 434  * operation (possibly at object termination time), so we must be careful.
 435  */
 436 void
 437 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
 438 {
 439         vm_object_t object;
 440         vm_page_t m;
 441         vm_pindex_t nobjsize;
 442
 443         if ((object = vp->v_object) == NULL)
 444                 return;
 445 #ifdef DEBUG_VFS_LOCKS
 446         {
 447                 struct mount *mp;
 448
 449                 mp = vp->v_mount;
 450                 if (mp != NULL && (mp->mnt_kern_flag & MNTK_VMSETSIZE_BUG) == 0)
 451                         assert_vop_elocked(vp,
 452                             "vnode_pager_setsize and not locked vnode");
 453         }
 454 #endif
 455         VM_OBJECT_WLOCK(object);
 456         if (object->type == OBJT_DEAD) {
 457                 VM_OBJECT_WUNLOCK(object);
 458                 return;
 459         }
 460         KASSERT(object->type == OBJT_VNODE,
 461             ("not vnode-backed object %p", object));
 462         if (nsize == object->un_pager.vnp.vnp_size) {
 463                 /*
 464                  * Hasn't changed size
 465                  */
 466                 VM_OBJECT_WUNLOCK(object);
 467                 return;
 468         }
 469         nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
 470         if (nsize < object->un_pager.vnp.vnp_size) {
 471                 /*
 472                  * File has shrunk. Toss any cached pages beyond the new EOF.
 473                  */
 474                 if (nobjsize < object->size)
 475                         vm_object_page_remove(object, nobjsize, object->size,
 476                             0);
 477                 /*
 478                  * this gets rid of garbage at the end of a page that is now
 479                  * only partially backed by the vnode.
 480                  *
 481                  * XXX for some reason (I don't know yet), if we take a
 482                  * completely invalid page and mark it partially valid
 483                  * it can screw up NFS reads, so we don't allow the case.
 484                  */
 485                 if (!(nsize & PAGE_MASK))
 486                         goto out;
 487                 m = vm_page_grab(object, OFF_TO_IDX(nsize), VM_ALLOC_NOCREAT);
 488                 if (m == NULL)
 489                         goto out;
 490                 if (!vm_page_none_valid(m)) {
 491                         int base = (int)nsize & PAGE_MASK;
 492                         int size = PAGE_SIZE - base;
 493
 494                         /*
 495                          * Clear out partial-page garbage in case
 496                          * the page has been mapped.
 497                          */
 498                         pmap_zero_page_area(m, base, size);
 499
 500                         /*
 501                          * Update the valid bits to reflect the blocks that
 502                          * have been zeroed.  Some of these valid bits may
 503                          * have already been set.
 504                          */
 505                         vm_page_set_valid_range(m, base, size);
 506
 507                         /*
 508                          * Round "base" to the next block boundary so that the
 509                          * dirty bit for a partially zeroed block is not
 510                          * cleared.
 511                          */
 512                         base = roundup2(base, DEV_BSIZE);
 513
 514                         /*
 515                          * Clear out partial-page dirty bits.
 516                          *
 517                          * note that we do not clear out the valid
 518                          * bits.  This would prevent bogus_page
 519                          * replacement from working properly.
 520                          */
 521                         vm_page_clear_dirty(m, base, PAGE_SIZE - base);
 522                 }
 523                 vm_page_xunbusy(m);
 524         }
 525 out:
 526 #if defined(__powerpc__) && !defined(__powerpc64__)
 527         object->un_pager.vnp.vnp_size = nsize;
 528 #else
 529         atomic_store_64(&object->un_pager.vnp.vnp_size, nsize);
 530 #endif
 531         object->size = nobjsize;
 532         VM_OBJECT_WUNLOCK(object);
 533 }
 534
 535 /*
 536  * calculate the linear (byte) disk address of specified virtual
 537  * file address
 538  */
 539 static int
 540 vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress,
 541     int *run)
 542 {
 543         int bsize;
 544         int err;
 545         daddr_t vblock;
 546         daddr_t voffset;
 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 static void
 570 vnode_pager_input_bdone(struct buf *bp)
 571 {
 572         runningbufwakeup(bp);
 573         bdone(bp);
 574 }
 575
 576 /*
 577  * small block filesystem vnode pager input
 578  */
 579 static int
 580 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m)
 581 {
 582         struct vnode *vp;
 583         struct bufobj *bo;
 584         struct buf *bp;
 585         struct sf_buf *sf;
 586         daddr_t fileaddr;
 587         vm_offset_t bsize;
 588         vm_page_bits_t bits;
 589         int error, i;
 590
 591         error = 0;
 592         vp = object->handle;
 593         if (VN_IS_DOOMED(vp))
 594                 return VM_PAGER_BAD;
 595
 596         bsize = vp->v_mount->mnt_stat.f_iosize;
 597
 598         VOP_BMAP(vp, 0, &bo, 0, NULL, NULL);
 599
 600         sf = sf_buf_alloc(m, 0);
 601
 602         for (i = 0; i < PAGE_SIZE / bsize; i++) {
 603                 vm_ooffset_t address;
 604
 605                 bits = vm_page_bits(i * bsize, bsize);
 606                 if (m->valid & bits)
 607                         continue;
 608
 609                 address = IDX_TO_OFF(m->pindex) + i * bsize;
 610                 if (address >= object->un_pager.vnp.vnp_size) {
 611                         fileaddr = -1;
 612                 } else {
 613                         error = vnode_pager_addr(vp, address, &fileaddr, NULL);
 614                         if (error)
 615                                 break;
 616                 }
 617                 if (fileaddr != -1) {
 618                         bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK);
 619
 620                         /* build a minimal buffer header */
 621                         bp->b_iocmd = BIO_READ;
 622                         bp->b_iodone = vnode_pager_input_bdone;
 623                         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
 624                         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
 625                         bp->b_rcred = crhold(curthread->td_ucred);
 626                         bp->b_wcred = crhold(curthread->td_ucred);
 627                         bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize;
 628                         bp->b_blkno = fileaddr;
 629                         pbgetbo(bo, bp);
 630                         bp->b_vp = vp;
 631                         bp->b_bcount = bsize;
 632                         bp->b_bufsize = bsize;
 633                         bp->b_runningbufspace = bp->b_bufsize;
 634                         atomic_add_long(&runningbufspace, bp->b_runningbufspace);
 635
 636                         /* do the input */
 637                         bp->b_iooffset = dbtob(bp->b_blkno);
 638                         bstrategy(bp);
 639
 640                         bwait(bp, PVM, "vnsrd");
 641
 642                         if ((bp->b_ioflags & BIO_ERROR) != 0) {
 643                                 KASSERT(bp->b_error != 0,
 644                                     ("%s: buf error but b_error == 0\n", __func__));
 645                                 error = bp->b_error;
 646                         }
 647
 648                         /*
 649                          * free the buffer header back to the swap buffer pool
 650                          */
 651                         bp->b_vp = NULL;
 652                         pbrelbo(bp);
 653                         uma_zfree(vnode_pbuf_zone, bp);
 654                         if (error)
 655                                 break;
 656                 } else
 657                         bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
 658                 KASSERT((m->dirty & bits) == 0,
 659                     ("vnode_pager_input_smlfs: page %p is dirty", m));
 660                 vm_page_bits_set(m, &m->valid, bits);
 661         }
 662         sf_buf_free(sf);
 663         if (error) {
 664                 return VM_PAGER_ERROR;
 665         }
 666         return VM_PAGER_OK;
 667 }
 668
 669 /*
 670  * old style vnode pager input routine
 671  */
 672 static int
 673 vnode_pager_input_old(vm_object_t object, vm_page_t m)
 674 {
 675         struct uio auio;
 676         struct iovec aiov;
 677         int error;
 678         int size;
 679         struct sf_buf *sf;
 680         struct vnode *vp;
 681
 682         VM_OBJECT_ASSERT_WLOCKED(object);
 683         error = 0;
 684
 685         /*
 686          * Return failure if beyond current EOF
 687          */
 688         if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
 689                 return VM_PAGER_BAD;
 690         } else {
 691                 size = PAGE_SIZE;
 692                 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
 693                         size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
 694                 vp = object->handle;
 695                 VM_OBJECT_WUNLOCK(object);
 696
 697                 /*
 698                  * Allocate a kernel virtual address and initialize so that
 699                  * we can use VOP_READ/WRITE routines.
 700                  */
 701                 sf = sf_buf_alloc(m, 0);
 702
 703                 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
 704                 aiov.iov_len = size;
 705                 auio.uio_iov = &aiov;
 706                 auio.uio_iovcnt = 1;
 707                 auio.uio_offset = IDX_TO_OFF(m->pindex);
 708                 auio.uio_segflg = UIO_SYSSPACE;
 709                 auio.uio_rw = UIO_READ;
 710                 auio.uio_resid = size;
 711                 auio.uio_td = curthread;
 712
 713                 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
 714                 if (!error) {
 715                         int count = size - auio.uio_resid;
 716
 717                         if (count == 0)
 718                                 error = EINVAL;
 719                         else if (count != PAGE_SIZE)
 720                                 bzero((caddr_t)sf_buf_kva(sf) + count,
 721                                     PAGE_SIZE - count);
 722                 }
 723                 sf_buf_free(sf);
 724
 725                 VM_OBJECT_WLOCK(object);
 726         }
 727         KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
 728         if (!error)
 729                 vm_page_valid(m);
 730         return error ? VM_PAGER_ERROR : VM_PAGER_OK;
 731 }
 732
 733 /*
 734  * generic vnode pager input routine
 735  */
 736
 737 /*
 738  * Local media VFS's that do not implement their own VOP_GETPAGES
 739  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
 740  * to implement the previous behaviour.
 741  *
 742  * All other FS's should use the bypass to get to the local media
 743  * backing vp's VOP_GETPAGES.
 744  */
 745 static int
 746 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
 747     int *rahead)
 748 {
 749         struct vnode *vp;
 750         int rtval;
 751
 752         /* Handle is stable with paging in progress. */
 753         vp = object->handle;
 754         rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead);
 755         KASSERT(rtval != EOPNOTSUPP,
 756             ("vnode_pager: FS getpages not implemented\n"));
 757         return rtval;
 758 }
 759
 760 static int
 761 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
 762     int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg)
 763 {
 764         struct vnode *vp;
 765         int rtval;
 766
 767         vp = object->handle;
 768         rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg);
 769         KASSERT(rtval != EOPNOTSUPP,
 770             ("vnode_pager: FS getpages_async not implemented\n"));
 771         return (rtval);
 772 }
 773
 774 /*
 775  * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for
 776  * local filesystems, where partially valid pages can only occur at
 777  * the end of file.
 778  */
 779 int
 780 vnode_pager_local_getpages(struct vop_getpages_args *ap)
 781 {
 782
 783         return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
 784             ap->a_rbehind, ap->a_rahead, NULL, NULL));
 785 }
 786
 787 int
 788 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap)
 789 {
 790         int error;
 791
 792         error = vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
 793             ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg);
 794         if (error != 0 && ap->a_iodone != NULL)
 795                 ap->a_iodone(ap->a_arg, ap->a_m, ap->a_count, error);
 796         return (error);
 797 }
 798
 799 /*
 800  * This is now called from local media FS's to operate against their
 801  * own vnodes if they fail to implement VOP_GETPAGES.
 802  */
 803 int
 804 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count,
 805     int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg)
 806 {
 807         vm_object_t object;
 808         struct bufobj *bo;
 809         struct buf *bp;
 810         off_t foff;
 811 #ifdef INVARIANTS
 812         off_t blkno0;
 813 #endif
 814         int bsize, pagesperblock;
 815         int error, before, after, rbehind, rahead, poff, i;
 816         int bytecount, secmask;
 817
 818         KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
 819             ("%s does not support devices", __func__));
 820
 821         if (VN_IS_DOOMED(vp))
 822                 return (VM_PAGER_BAD);
 823
 824         object = vp->v_object;
 825         foff = IDX_TO_OFF(m[0]->pindex);
 826         bsize = vp->v_mount->mnt_stat.f_iosize;
 827         pagesperblock = bsize / PAGE_SIZE;
 828
 829         KASSERT(foff < object->un_pager.vnp.vnp_size,
 830             ("%s: page %p offset beyond vp %p size", __func__, m[0], vp));
 831         KASSERT(count <= atop(maxphys),
 832             ("%s: requested %d pages", __func__, count));
 833
 834         /*
 835          * The last page has valid blocks.  Invalid part can only
 836          * exist at the end of file, and the page is made fully valid
 837          * by zeroing in vm_pager_get_pages().
 838          */
 839         if (!vm_page_none_valid(m[count - 1]) && --count == 0) {
 840                 if (iodone != NULL)
 841                         iodone(arg, m, 1, 0);
 842                 return (VM_PAGER_OK);
 843         }
 844
 845         bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK);
 846         MPASS((bp->b_flags & B_MAXPHYS) != 0);
 847
 848         /*
 849          * Get the underlying device blocks for the file with VOP_BMAP().
 850          * If the file system doesn't support VOP_BMAP, use old way of
 851          * getting pages via VOP_READ.
 852          */
 853         error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before);
 854         if (error == EOPNOTSUPP) {
 855                 uma_zfree(vnode_pbuf_zone, bp);
 856                 VM_OBJECT_WLOCK(object);
 857                 for (i = 0; i < count; i++) {
 858                         VM_CNT_INC(v_vnodein);
 859                         VM_CNT_INC(v_vnodepgsin);
 860                         error = vnode_pager_input_old(object, m[i]);
 861                         if (error)
 862                                 break;
 863                 }
 864                 VM_OBJECT_WUNLOCK(object);
 865                 return (error);
 866         } else if (error != 0) {
 867                 uma_zfree(vnode_pbuf_zone, bp);
 868                 return (VM_PAGER_ERROR);
 869         }
 870
 871         /*
 872          * If the file system supports BMAP, but blocksize is smaller
 873          * than a page size, then use special small filesystem code.
 874          */
 875         if (pagesperblock == 0) {
 876                 uma_zfree(vnode_pbuf_zone, bp);
 877                 for (i = 0; i < count; i++) {
 878                         VM_CNT_INC(v_vnodein);
 879                         VM_CNT_INC(v_vnodepgsin);
 880                         error = vnode_pager_input_smlfs(object, m[i]);
 881                         if (error)
 882                                 break;
 883                 }
 884                 return (error);
 885         }
 886
 887         /*
 888          * A sparse file can be encountered only for a single page request,
 889          * which may not be preceded by call to vm_pager_haspage().
 890          */
 891         if (bp->b_blkno == -1) {
 892                 KASSERT(count == 1,
 893                     ("%s: array[%d] request to a sparse file %p", __func__,
 894                     count, vp));
 895                 uma_zfree(vnode_pbuf_zone, bp);
 896                 pmap_zero_page(m[0]);
 897                 KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty",
 898                     __func__, m[0]));
 899                 vm_page_valid(m[0]);
 900                 return (VM_PAGER_OK);
 901         }
 902
 903 #ifdef INVARIANTS
 904         blkno0 = bp->b_blkno;
 905 #endif
 906         bp->b_blkno += (foff % bsize) / DEV_BSIZE;
 907
 908         /* Recalculate blocks available after/before to pages. */
 909         poff = (foff % bsize) / PAGE_SIZE;
 910         before *= pagesperblock;
 911         before += poff;
 912         after *= pagesperblock;
 913         after += pagesperblock - (poff + 1);
 914         if (m[0]->pindex + after >= object->size)
 915                 after = object->size - 1 - m[0]->pindex;
 916         KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d",
 917             __func__, count, after + 1));
 918         after -= count - 1;
 919
 920         /* Trim requested rbehind/rahead to possible values. */
 921         rbehind = a_rbehind ? *a_rbehind : 0;
 922         rahead = a_rahead ? *a_rahead : 0;
 923         rbehind = min(rbehind, before);
 924         rbehind = min(rbehind, m[0]->pindex);
 925         rahead = min(rahead, after);
 926         rahead = min(rahead, object->size - m[count - 1]->pindex);
 927         /*
 928          * Check that total amount of pages fit into buf.  Trim rbehind and
 929          * rahead evenly if not.
 930          */
 931         if (rbehind + rahead + count > atop(maxphys)) {
 932                 int trim, sum;
 933
 934                 trim = rbehind + rahead + count - atop(maxphys) + 1;
 935                 sum = rbehind + rahead;
 936                 if (rbehind == before) {
 937                         /* Roundup rbehind trim to block size. */
 938                         rbehind -= roundup(trim * rbehind / sum, pagesperblock);
 939                         if (rbehind < 0)
 940                                 rbehind = 0;
 941                 } else
 942                         rbehind -= trim * rbehind / sum;
 943                 rahead -= trim * rahead / sum;
 944         }
 945         KASSERT(rbehind + rahead + count <= atop(maxphys),
 946             ("%s: behind %d ahead %d count %d maxphys %lu", __func__,
 947             rbehind, rahead, count, maxphys));
 948
 949         /*
 950          * Fill in the bp->b_pages[] array with requested and optional
 951          * read behind or read ahead pages.  Read behind pages are looked
 952          * up in a backward direction, down to a first cached page.  Same
 953          * for read ahead pages, but there is no need to shift the array
 954          * in case of encountering a cached page.
 955          */
 956         i = bp->b_npages = 0;
 957         if (rbehind) {
 958                 vm_pindex_t startpindex, tpindex;
 959                 vm_page_t p;
 960
 961                 VM_OBJECT_WLOCK(object);
 962                 startpindex = m[0]->pindex - rbehind;
 963                 if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL &&
 964                     p->pindex >= startpindex)
 965                         startpindex = p->pindex + 1;
 966
 967                 /* tpindex is unsigned; beware of numeric underflow. */
 968                 for (tpindex = m[0]->pindex - 1;
 969                     tpindex >= startpindex && tpindex < m[0]->pindex;
 970                     tpindex--, i++) {
 971                         p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
 972                         if (p == NULL) {
 973                                 /* Shift the array. */
 974                                 for (int j = 0; j < i; j++)
 975                                         bp->b_pages[j] = bp->b_pages[j +
 976                                             tpindex + 1 - startpindex];
 977                                 break;
 978                         }
 979                         bp->b_pages[tpindex - startpindex] = p;
 980                 }
 981
 982                 bp->b_pgbefore = i;
 983                 bp->b_npages += i;
 984                 bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE;
 985         } else
 986                 bp->b_pgbefore = 0;
 987
 988         /* Requested pages. */
 989         for (int j = 0; j < count; j++, i++)
 990                 bp->b_pages[i] = m[j];
 991         bp->b_npages += count;
 992
 993         if (rahead) {
 994                 vm_pindex_t endpindex, tpindex;
 995                 vm_page_t p;
 996
 997                 if (!VM_OBJECT_WOWNED(object))
 998                         VM_OBJECT_WLOCK(object);
 999                 endpindex = m[count - 1]->pindex + rahead + 1;
1000                 if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL &&
1001                     p->pindex < endpindex)
1002                         endpindex = p->pindex;
1003                 if (endpindex > object->size)
1004                         endpindex = object->size;
1005
1006                 for (tpindex = m[count - 1]->pindex + 1;
1007                     tpindex < endpindex; i++, tpindex++) {
1008                         p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1009                         if (p == NULL)
1010                                 break;
1011                         bp->b_pages[i] = p;
1012                 }
1013
1014                 bp->b_pgafter = i - bp->b_npages;
1015                 bp->b_npages = i;
1016         } else
1017                 bp->b_pgafter = 0;
1018
1019         if (VM_OBJECT_WOWNED(object))
1020                 VM_OBJECT_WUNLOCK(object);
1021
1022         /* Report back actual behind/ahead read. */
1023         if (a_rbehind)
1024                 *a_rbehind = bp->b_pgbefore;
1025         if (a_rahead)
1026                 *a_rahead = bp->b_pgafter;
1027
1028 #ifdef INVARIANTS
1029         KASSERT(bp->b_npages <= atop(maxphys),
1030             ("%s: buf %p overflowed", __func__, bp));
1031         for (int j = 1, prev = 0; j < bp->b_npages; j++) {
1032                 if (bp->b_pages[j] == bogus_page)
1033                         continue;
1034                 KASSERT(bp->b_pages[j]->pindex - bp->b_pages[prev]->pindex ==
1035                     j - prev, ("%s: pages array not consecutive, bp %p",
1036                      __func__, bp));
1037                 prev = j;
1038         }
1039 #endif
1040
1041         /*
1042          * Recalculate first offset and bytecount with regards to read behind.
1043          * Truncate bytecount to vnode real size and round up physical size
1044          * for real devices.
1045          */
1046         foff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1047         bytecount = bp->b_npages << PAGE_SHIFT;
1048         if ((foff + bytecount) > object->un_pager.vnp.vnp_size)
1049                 bytecount = object->un_pager.vnp.vnp_size - foff;
1050         secmask = bo->bo_bsize - 1;
1051         KASSERT(secmask < PAGE_SIZE && secmask > 0,
1052             ("%s: sector size %d too large", __func__, secmask + 1));
1053         bytecount = (bytecount + secmask) & ~secmask;
1054
1055         /*
1056          * And map the pages to be read into the kva, if the filesystem
1057          * requires mapped buffers.
1058          */
1059         if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
1060             unmapped_buf_allowed) {
1061                 bp->b_data = unmapped_buf;
1062                 bp->b_offset = 0;
1063         } else {
1064                 bp->b_data = bp->b_kvabase;
1065                 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1066         }
1067
1068         /* Build a minimal buffer header. */
1069         bp->b_iocmd = BIO_READ;
1070         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
1071         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
1072         bp->b_rcred = crhold(curthread->td_ucred);
1073         bp->b_wcred = crhold(curthread->td_ucred);
1074         pbgetbo(bo, bp);
1075         bp->b_vp = vp;
1076         bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount;
1077         bp->b_iooffset = dbtob(bp->b_blkno);
1078         KASSERT(IDX_TO_OFF(m[0]->pindex - bp->b_pages[0]->pindex) ==
1079             (blkno0 - bp->b_blkno) * DEV_BSIZE +
1080             IDX_TO_OFF(m[0]->pindex) % bsize,
1081             ("wrong offsets bsize %d m[0] %ju b_pages[0] %ju "
1082             "blkno0 %ju b_blkno %ju", bsize,
1083             (uintmax_t)m[0]->pindex, (uintmax_t)bp->b_pages[0]->pindex,
1084             (uintmax_t)blkno0, (uintmax_t)bp->b_blkno));
1085
1086         atomic_add_long(&runningbufspace, bp->b_runningbufspace);
1087         VM_CNT_INC(v_vnodein);
1088         VM_CNT_ADD(v_vnodepgsin, bp->b_npages);
1089
1090         if (iodone != NULL) { /* async */
1091                 bp->b_pgiodone = iodone;
1092                 bp->b_caller1 = arg;
1093                 bp->b_iodone = vnode_pager_generic_getpages_done_async;
1094                 bp->b_flags |= B_ASYNC;
1095                 BUF_KERNPROC(bp);
1096                 bstrategy(bp);
1097                 return (VM_PAGER_OK);
1098         } else {
1099                 bp->b_iodone = bdone;
1100                 bstrategy(bp);
1101                 bwait(bp, PVM, "vnread");
1102                 error = vnode_pager_generic_getpages_done(bp);
1103                 for (i = 0; i < bp->b_npages; i++)
1104                         bp->b_pages[i] = NULL;
1105                 bp->b_vp = NULL;
1106                 pbrelbo(bp);
1107                 uma_zfree(vnode_pbuf_zone, bp);
1108                 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
1109         }
1110 }
1111
1112 static void
1113 vnode_pager_generic_getpages_done_async(struct buf *bp)
1114 {
1115         int error;
1116
1117         error = vnode_pager_generic_getpages_done(bp);
1118         /* Run the iodone upon the requested range. */
1119         bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore,
1120             bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error);
1121         for (int i = 0; i < bp->b_npages; i++)
1122                 bp->b_pages[i] = NULL;
1123         bp->b_vp = NULL;
1124         pbrelbo(bp);
1125         uma_zfree(vnode_pbuf_zone, bp);
1126 }
1127
1128 static int
1129 vnode_pager_generic_getpages_done(struct buf *bp)
1130 {
1131         vm_object_t object;
1132         off_t tfoff, nextoff;
1133         int i, error;
1134
1135         KASSERT((bp->b_ioflags & BIO_ERROR) == 0 || bp->b_error != 0,
1136             ("%s: buf error but b_error == 0\n", __func__));
1137         error = (bp->b_ioflags & BIO_ERROR) != 0 ? bp->b_error : 0;
1138         object = bp->b_vp->v_object;
1139
1140         runningbufwakeup(bp);
1141
1142         if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1143                 if (!buf_mapped(bp)) {
1144                         bp->b_data = bp->b_kvabase;
1145                         pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages,
1146                             bp->b_npages);
1147                 }
1148                 bzero(bp->b_data + bp->b_bcount,
1149                     PAGE_SIZE * bp->b_npages - bp->b_bcount);
1150         }
1151         if (buf_mapped(bp)) {
1152                 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1153                 bp->b_data = unmapped_buf;
1154         }
1155
1156         /*
1157          * If the read failed, we must free any read ahead/behind pages here.
1158          * The requested pages are freed by the caller (for sync requests)
1159          * or by the bp->b_pgiodone callback (for async requests).
1160          */
1161         if (error != 0) {
1162                 VM_OBJECT_WLOCK(object);
1163                 for (i = 0; i < bp->b_pgbefore; i++)
1164                         vm_page_free_invalid(bp->b_pages[i]);
1165                 for (i = bp->b_npages - bp->b_pgafter; i < bp->b_npages; i++)
1166                         vm_page_free_invalid(bp->b_pages[i]);
1167                 VM_OBJECT_WUNLOCK(object);
1168                 return (error);
1169         }
1170
1171         /* Read lock to protect size. */
1172         VM_OBJECT_RLOCK(object);
1173         for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1174             i < bp->b_npages; i++, tfoff = nextoff) {
1175                 vm_page_t mt;
1176
1177                 nextoff = tfoff + PAGE_SIZE;
1178                 mt = bp->b_pages[i];
1179                 if (mt == bogus_page)
1180                         continue;
1181
1182                 if (nextoff <= object->un_pager.vnp.vnp_size) {
1183                         /*
1184                          * Read filled up entire page.
1185                          */
1186                         vm_page_valid(mt);
1187                         KASSERT(mt->dirty == 0,
1188                             ("%s: page %p is dirty", __func__, mt));
1189                         KASSERT(!pmap_page_is_mapped(mt),
1190                             ("%s: page %p is mapped", __func__, mt));
1191                 } else {
1192                         /*
1193                          * Read did not fill up entire page.
1194                          *
1195                          * Currently we do not set the entire page valid,
1196                          * we just try to clear the piece that we couldn't
1197                          * read.
1198                          */
1199                         vm_page_set_valid_range(mt, 0,
1200                             object->un_pager.vnp.vnp_size - tfoff);
1201                         KASSERT((mt->dirty & vm_page_bits(0,
1202                             object->un_pager.vnp.vnp_size - tfoff)) == 0,
1203                             ("%s: page %p is dirty", __func__, mt));
1204                 }
1205
1206                 if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter)
1207                         vm_page_readahead_finish(mt);
1208         }
1209         VM_OBJECT_RUNLOCK(object);
1210
1211         return (error);
1212 }
1213
1214 /*
1215  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
1216  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1217  * vnode_pager_generic_putpages() to implement the previous behaviour.
1218  *
1219  * All other FS's should use the bypass to get to the local media
1220  * backing vp's VOP_PUTPAGES.
1221  */
1222 static void
1223 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1224     int flags, int *rtvals)
1225 {
1226         int rtval;
1227         struct vnode *vp;
1228         int bytes = count * PAGE_SIZE;
1229
1230         /*
1231          * Force synchronous operation if we are extremely low on memory
1232          * to prevent a low-memory deadlock.  VOP operations often need to
1233          * allocate more memory to initiate the I/O ( i.e. do a BMAP
1234          * operation ).  The swapper handles the case by limiting the amount
1235          * of asynchronous I/O, but that sort of solution doesn't scale well
1236          * for the vnode pager without a lot of work.
1237          *
1238          * Also, the backing vnode's iodone routine may not wake the pageout
1239          * daemon up.  This should be probably be addressed XXX.
1240          */
1241
1242         if (vm_page_count_min())
1243                 flags |= VM_PAGER_PUT_SYNC;
1244
1245         /*
1246          * Call device-specific putpages function
1247          */
1248         vp = object->handle;
1249         VM_OBJECT_WUNLOCK(object);
1250         rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1251         KASSERT(rtval != EOPNOTSUPP,
1252             ("vnode_pager: stale FS putpages\n"));
1253         VM_OBJECT_WLOCK(object);
1254 }
1255
1256 static int
1257 vn_off2bidx(vm_ooffset_t offset)
1258 {
1259
1260         return ((offset & PAGE_MASK) / DEV_BSIZE);
1261 }
1262
1263 static bool
1264 vn_dirty_blk(vm_page_t m, vm_ooffset_t offset)
1265 {
1266
1267         KASSERT(IDX_TO_OFF(m->pindex) <= offset &&
1268             offset < IDX_TO_OFF(m->pindex + 1),
1269             ("page %p pidx %ju offset %ju", m, (uintmax_t)m->pindex,
1270             (uintmax_t)offset));
1271         return ((m->dirty & ((vm_page_bits_t)1 << vn_off2bidx(offset))) != 0);
1272 }
1273
1274 /*
1275  * This is now called from local media FS's to operate against their
1276  * own vnodes if they fail to implement VOP_PUTPAGES.
1277  *
1278  * This is typically called indirectly via the pageout daemon and
1279  * clustering has already typically occurred, so in general we ask the
1280  * underlying filesystem to write the data out asynchronously rather
1281  * then delayed.
1282  */
1283 int
1284 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1285     int flags, int *rtvals)
1286 {
1287         vm_object_t object;
1288         vm_page_t m;
1289         vm_ooffset_t max_offset, next_offset, poffset, prev_offset;
1290         struct uio auio;
1291         struct iovec aiov;
1292         off_t prev_resid, wrsz;
1293         int count, error, i, maxsize, ncount, pgoff, ppscheck;
1294         bool in_hole;
1295         static struct timeval lastfail;
1296         static int curfail;
1297
1298         object = vp->v_object;
1299         count = bytecount / PAGE_SIZE;
1300
1301         for (i = 0; i < count; i++)
1302                 rtvals[i] = VM_PAGER_ERROR;
1303
1304         if ((int64_t)ma[0]->pindex < 0) {
1305                 printf("vnode_pager_generic_putpages: "
1306                     "attempt to write meta-data 0x%jx(%lx)\n",
1307                     (uintmax_t)ma[0]->pindex, (u_long)ma[0]->dirty);
1308                 rtvals[0] = VM_PAGER_BAD;
1309                 return (VM_PAGER_BAD);
1310         }
1311
1312         maxsize = count * PAGE_SIZE;
1313         ncount = count;
1314
1315         poffset = IDX_TO_OFF(ma[0]->pindex);
1316
1317         /*
1318          * If the page-aligned write is larger then the actual file we
1319          * have to invalidate pages occurring beyond the file EOF.  However,
1320          * there is an edge case where a file may not be page-aligned where
1321          * the last page is partially invalid.  In this case the filesystem
1322          * may not properly clear the dirty bits for the entire page (which
1323          * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1324          * With the page busied we are free to fix up the dirty bits here.
1325          *
1326          * We do not under any circumstances truncate the valid bits, as
1327          * this will screw up bogus page replacement.
1328          */
1329         VM_OBJECT_RLOCK(object);
1330         if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1331                 if (object->un_pager.vnp.vnp_size > poffset) {
1332                         maxsize = object->un_pager.vnp.vnp_size - poffset;
1333                         ncount = btoc(maxsize);
1334                         if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1335                                 pgoff = roundup2(pgoff, DEV_BSIZE);
1336
1337                                 /*
1338                                  * If the page is busy and the following
1339                                  * conditions hold, then the page's dirty
1340                                  * field cannot be concurrently changed by a
1341                                  * pmap operation.
1342                                  */
1343                                 m = ma[ncount - 1];
1344                                 vm_page_assert_sbusied(m);
1345                                 KASSERT(!pmap_page_is_write_mapped(m),
1346                 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1347                                 MPASS(m->dirty != 0);
1348                                 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1349                                     pgoff);
1350                         }
1351                 } else {
1352                         maxsize = 0;
1353                         ncount = 0;
1354                 }
1355                 for (i = ncount; i < count; i++)
1356                         rtvals[i] = VM_PAGER_BAD;
1357         }
1358         VM_OBJECT_RUNLOCK(object);
1359
1360         auio.uio_iov = &aiov;
1361         auio.uio_segflg = UIO_NOCOPY;
1362         auio.uio_rw = UIO_WRITE;
1363         auio.uio_td = NULL;
1364         max_offset = roundup2(poffset + maxsize, DEV_BSIZE);
1365
1366         for (prev_offset = poffset; prev_offset < max_offset;) {
1367                 /* Skip clean blocks. */
1368                 for (in_hole = true; in_hole && prev_offset < max_offset;) {
1369                         m = ma[OFF_TO_IDX(prev_offset - poffset)];
1370                         for (i = vn_off2bidx(prev_offset);
1371                             i < sizeof(vm_page_bits_t) * NBBY &&
1372                             prev_offset < max_offset; i++) {
1373                                 if (vn_dirty_blk(m, prev_offset)) {
1374                                         in_hole = false;
1375                                         break;
1376                                 }
1377                                 prev_offset += DEV_BSIZE;
1378                         }
1379                 }
1380                 if (in_hole)
1381                         goto write_done;
1382
1383                 /* Find longest run of dirty blocks. */
1384                 for (next_offset = prev_offset; next_offset < max_offset;) {
1385                         m = ma[OFF_TO_IDX(next_offset - poffset)];
1386                         for (i = vn_off2bidx(next_offset);
1387                             i < sizeof(vm_page_bits_t) * NBBY &&
1388                             next_offset < max_offset; i++) {
1389                                 if (!vn_dirty_blk(m, next_offset))
1390                                         goto start_write;
1391                                 next_offset += DEV_BSIZE;
1392                         }
1393                 }
1394 start_write:
1395                 if (next_offset > poffset + maxsize)
1396                         next_offset = poffset + maxsize;
1397                 if (prev_offset == next_offset)
1398                         goto write_done;
1399
1400                 /*
1401                  * Getting here requires finding a dirty block in the
1402                  * 'skip clean blocks' loop.
1403                  */
1404
1405                 aiov.iov_base = NULL;
1406                 auio.uio_iovcnt = 1;
1407                 auio.uio_offset = prev_offset;
1408                 prev_resid = auio.uio_resid = aiov.iov_len = next_offset -
1409                     prev_offset;
1410                 error = VOP_WRITE(vp, &auio,
1411                     vnode_pager_putpages_ioflags(flags), curthread->td_ucred);
1412
1413                 wrsz = prev_resid - auio.uio_resid;
1414                 if (wrsz == 0) {
1415                         if (ppsratecheck(&lastfail, &curfail, 1) != 0) {
1416                                 vn_printf(vp, "vnode_pager_putpages: "
1417                                     "zero-length write at %ju resid %zd\n",
1418                                     auio.uio_offset, auio.uio_resid);
1419                         }
1420                         break;
1421                 }
1422
1423                 /* Adjust the starting offset for next iteration. */
1424                 prev_offset += wrsz;
1425                 MPASS(auio.uio_offset == prev_offset);
1426
1427                 ppscheck = 0;
1428                 if (error != 0 && (ppscheck = ppsratecheck(&lastfail,
1429                     &curfail, 1)) != 0)
1430                         vn_printf(vp, "vnode_pager_putpages: I/O error %d\n",
1431                             error);
1432                 if (auio.uio_resid != 0 && (ppscheck != 0 ||
1433                     ppsratecheck(&lastfail, &curfail, 1) != 0))
1434                         vn_printf(vp, "vnode_pager_putpages: residual I/O %zd "
1435                             "at %ju\n", auio.uio_resid,
1436                             (uintmax_t)ma[0]->pindex);
1437                 if (error != 0 || auio.uio_resid != 0)
1438                         break;
1439         }
1440 write_done:
1441         /* Mark completely processed pages. */
1442         for (i = 0; i < OFF_TO_IDX(prev_offset - poffset); i++)
1443                 rtvals[i] = VM_PAGER_OK;
1444         /* Mark partial EOF page. */
1445         if (prev_offset == poffset + maxsize && (prev_offset & PAGE_MASK) != 0)
1446                 rtvals[i++] = VM_PAGER_OK;
1447         /* Unwritten pages in range, free bonus if the page is clean. */
1448         for (; i < ncount; i++)
1449                 rtvals[i] = ma[i]->dirty == 0 ? VM_PAGER_OK : VM_PAGER_ERROR;
1450         VM_CNT_ADD(v_vnodepgsout, i);
1451         VM_CNT_INC(v_vnodeout);
1452         return (rtvals[0]);
1453 }
1454
1455 int
1456 vnode_pager_putpages_ioflags(int pager_flags)
1457 {
1458         int ioflags;
1459
1460         /*
1461          * Pageouts are already clustered, use IO_ASYNC to force a
1462          * bawrite() rather then a bdwrite() to prevent paging I/O
1463          * from saturating the buffer cache.  Dummy-up the sequential
1464          * heuristic to cause large ranges to cluster.  If neither
1465          * IO_SYNC or IO_ASYNC is set, the system decides how to
1466          * cluster.
1467          */
1468         ioflags = IO_VMIO;
1469         if ((pager_flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) != 0)
1470                 ioflags |= IO_SYNC;
1471         else if ((pager_flags & VM_PAGER_CLUSTER_OK) == 0)
1472                 ioflags |= IO_ASYNC;
1473         ioflags |= (pager_flags & VM_PAGER_PUT_INVAL) != 0 ? IO_INVAL: 0;
1474         ioflags |= (pager_flags & VM_PAGER_PUT_NOREUSE) != 0 ? IO_NOREUSE : 0;
1475         ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1476         return (ioflags);
1477 }
1478
1479 /*
1480  * vnode_pager_undirty_pages().
1481  *
1482  * A helper to mark pages as clean after pageout that was possibly
1483  * done with a short write.  The lpos argument specifies the page run
1484  * length in bytes, and the written argument specifies how many bytes
1485  * were actually written.  eof is the offset past the last valid byte
1486  * in the vnode using the absolute file position of the first byte in
1487  * the run as the base from which it is computed.
1488  */
1489 void
1490 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written, off_t eof,
1491     int lpos)
1492 {
1493         vm_object_t obj;
1494         int i, pos, pos_devb;
1495
1496         if (written == 0 && eof >= lpos)
1497                 return;
1498         obj = ma[0]->object;
1499         for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1500                 if (pos < trunc_page(written)) {
1501                         rtvals[i] = VM_PAGER_OK;
1502                         vm_page_undirty(ma[i]);
1503                 } else {
1504                         /* Partially written page. */
1505                         rtvals[i] = VM_PAGER_AGAIN;
1506                         vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1507                 }
1508         }
1509         if (eof >= lpos) /* avoid truncation */
1510                 return;
1511         for (pos = eof, i = OFF_TO_IDX(trunc_page(pos)); pos < lpos; i++) {
1512                 if (pos != trunc_page(pos)) {
1513                         /*
1514                          * The page contains the last valid byte in
1515                          * the vnode, mark the rest of the page as
1516                          * clean, potentially making the whole page
1517                          * clean.
1518                          */
1519                         pos_devb = roundup2(pos & PAGE_MASK, DEV_BSIZE);
1520                         vm_page_clear_dirty(ma[i], pos_devb, PAGE_SIZE -
1521                             pos_devb);
1522
1523                         /*
1524                          * If the page was cleaned, report the pageout
1525                          * on it as successful.  msync() no longer
1526                          * needs to write out the page, endlessly
1527                          * creating write requests and dirty buffers.
1528                          */
1529                         if (ma[i]->dirty == 0)
1530                                 rtvals[i] = VM_PAGER_OK;
1531
1532                         pos = round_page(pos);
1533                 } else {
1534                         /* vm_pageout_flush() clears dirty */
1535                         rtvals[i] = VM_PAGER_BAD;
1536                         pos += PAGE_SIZE;
1537                 }
1538         }
1539 }
1540
1541 static void
1542 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1543     vm_offset_t end)
1544 {
1545         struct vnode *vp;
1546         vm_ooffset_t old_wm;
1547
1548         VM_OBJECT_WLOCK(object);
1549         if (object->type != OBJT_VNODE) {
1550                 VM_OBJECT_WUNLOCK(object);
1551                 return;
1552         }
1553         old_wm = object->un_pager.vnp.writemappings;
1554         object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1555         vp = object->handle;
1556         if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1557                 ASSERT_VOP_LOCKED(vp, "v_writecount inc");
1558                 VOP_ADD_WRITECOUNT_CHECKED(vp, 1);
1559                 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1560                     __func__, vp, vp->v_writecount);
1561         } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1562                 ASSERT_VOP_LOCKED(vp, "v_writecount dec");
1563                 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1564                 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1565                     __func__, vp, vp->v_writecount);
1566         }
1567         VM_OBJECT_WUNLOCK(object);
1568 }
1569
1570 static void
1571 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1572     vm_offset_t end)
1573 {
1574         struct vnode *vp;
1575         struct mount *mp;
1576         vm_offset_t inc;
1577
1578         VM_OBJECT_WLOCK(object);
1579
1580         /*
1581          * First, recheck the object type to account for the race when
1582          * the vnode is reclaimed.
1583          */
1584         if (object->type != OBJT_VNODE) {
1585                 VM_OBJECT_WUNLOCK(object);
1586                 return;
1587         }
1588
1589         /*
1590          * Optimize for the case when writemappings is not going to
1591          * zero.
1592          */
1593         inc = end - start;
1594         if (object->un_pager.vnp.writemappings != inc) {
1595                 object->un_pager.vnp.writemappings -= inc;
1596                 VM_OBJECT_WUNLOCK(object);
1597                 return;
1598         }
1599
1600         vp = object->handle;
1601         vhold(vp);
1602         VM_OBJECT_WUNLOCK(object);
1603         mp = NULL;
1604         vn_start_write(vp, &mp, V_WAIT);
1605         vn_lock(vp, LK_SHARED | LK_RETRY);
1606
1607         /*
1608          * Decrement the object's writemappings, by swapping the start
1609          * and end arguments for vnode_pager_update_writecount().  If
1610          * there was not a race with vnode reclaimation, then the
1611          * vnode's v_writecount is decremented.
1612          */
1613         vnode_pager_update_writecount(object, end, start);
1614         VOP_UNLOCK(vp);
1615         vdrop(vp);
1616         if (mp != NULL)
1617                 vn_finished_write(mp);
1618 }
1619
1620 static void
1621 vnode_pager_getvp(vm_object_t object, struct vnode **vpp, bool *vp_heldp)
1622 {
1623         *vpp = object->handle;
1624 }
1625
1626 static void
1627 vnode_pager_clean1(struct vnode *vp, int sync_flags)
1628 {
1629         struct vm_object *obj;
1630
1631         ASSERT_VOP_LOCKED(vp, "needs lock for writes");
1632         obj = vp->v_object;
1633         if (obj == NULL)
1634                 return;
1635
1636         VM_OBJECT_WLOCK(obj);
1637         vm_object_page_clean(obj, 0, 0, sync_flags);
1638         VM_OBJECT_WUNLOCK(obj);
1639 }
1640
1641 void
1642 vnode_pager_clean_sync(struct vnode *vp)
1643 {
1644         vnode_pager_clean1(vp, OBJPC_SYNC);
1645 }
1646
1647 void
1648 vnode_pager_clean_async(struct vnode *vp)
1649 {
1650         vnode_pager_clean1(vp, 0);
1651 }