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