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