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 (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                                 KASSERT(bp->b_error != 0,
 633                                     ("%s: buf error but b_error == 0\n", __func__));
 634                                 error = bp->b_error;
 635                         }
 636
 637                         /*
 638                          * free the buffer header back to the swap buffer pool
 639                          */
 640                         bp->b_vp = NULL;
 641                         pbrelbo(bp);
 642                         uma_zfree(vnode_pbuf_zone, bp);
 643                         if (error)
 644                                 break;
 645                 } else
 646                         bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize);
 647                 KASSERT((m->dirty & bits) == 0,
 648                     ("vnode_pager_input_smlfs: page %p is dirty", m));
 649                 vm_page_bits_set(m, &m->valid, bits);
 650         }
 651         sf_buf_free(sf);
 652         if (error) {
 653                 return VM_PAGER_ERROR;
 654         }
 655         return VM_PAGER_OK;
 656 }
 657
 658 /*
 659  * old style vnode pager input routine
 660  */
 661 static int
 662 vnode_pager_input_old(vm_object_t object, vm_page_t m)
 663 {
 664         struct uio auio;
 665         struct iovec aiov;
 666         int error;
 667         int size;
 668         struct sf_buf *sf;
 669         struct vnode *vp;
 670
 671         VM_OBJECT_ASSERT_WLOCKED(object);
 672         error = 0;
 673
 674         /*
 675          * Return failure if beyond current EOF
 676          */
 677         if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
 678                 return VM_PAGER_BAD;
 679         } else {
 680                 size = PAGE_SIZE;
 681                 if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
 682                         size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
 683                 vp = object->handle;
 684                 VM_OBJECT_WUNLOCK(object);
 685
 686                 /*
 687                  * Allocate a kernel virtual address and initialize so that
 688                  * we can use VOP_READ/WRITE routines.
 689                  */
 690                 sf = sf_buf_alloc(m, 0);
 691
 692                 aiov.iov_base = (caddr_t)sf_buf_kva(sf);
 693                 aiov.iov_len = size;
 694                 auio.uio_iov = &aiov;
 695                 auio.uio_iovcnt = 1;
 696                 auio.uio_offset = IDX_TO_OFF(m->pindex);
 697                 auio.uio_segflg = UIO_SYSSPACE;
 698                 auio.uio_rw = UIO_READ;
 699                 auio.uio_resid = size;
 700                 auio.uio_td = curthread;
 701
 702                 error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
 703                 if (!error) {
 704                         int count = size - auio.uio_resid;
 705
 706                         if (count == 0)
 707                                 error = EINVAL;
 708                         else if (count != PAGE_SIZE)
 709                                 bzero((caddr_t)sf_buf_kva(sf) + count,
 710                                     PAGE_SIZE - count);
 711                 }
 712                 sf_buf_free(sf);
 713
 714                 VM_OBJECT_WLOCK(object);
 715         }
 716         KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m));
 717         if (!error)
 718                 vm_page_valid(m);
 719         return error ? VM_PAGER_ERROR : VM_PAGER_OK;
 720 }
 721
 722 /*
 723  * generic vnode pager input routine
 724  */
 725
 726 /*
 727  * Local media VFS's that do not implement their own VOP_GETPAGES
 728  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
 729  * to implement the previous behaviour.
 730  *
 731  * All other FS's should use the bypass to get to the local media
 732  * backing vp's VOP_GETPAGES.
 733  */
 734 static int
 735 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
 736     int *rahead)
 737 {
 738         struct vnode *vp;
 739         int rtval;
 740
 741         /* Handle is stable with paging in progress. */
 742         vp = object->handle;
 743         rtval = VOP_GETPAGES(vp, m, count, rbehind, rahead);
 744         KASSERT(rtval != EOPNOTSUPP,
 745             ("vnode_pager: FS getpages not implemented\n"));
 746         return rtval;
 747 }
 748
 749 static int
 750 vnode_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
 751     int *rbehind, int *rahead, vop_getpages_iodone_t iodone, void *arg)
 752 {
 753         struct vnode *vp;
 754         int rtval;
 755
 756         vp = object->handle;
 757         rtval = VOP_GETPAGES_ASYNC(vp, m, count, rbehind, rahead, iodone, arg);
 758         KASSERT(rtval != EOPNOTSUPP,
 759             ("vnode_pager: FS getpages_async not implemented\n"));
 760         return (rtval);
 761 }
 762
 763 /*
 764  * The implementation of VOP_GETPAGES() and VOP_GETPAGES_ASYNC() for
 765  * local filesystems, where partially valid pages can only occur at
 766  * the end of file.
 767  */
 768 int
 769 vnode_pager_local_getpages(struct vop_getpages_args *ap)
 770 {
 771
 772         return (vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
 773             ap->a_rbehind, ap->a_rahead, NULL, NULL));
 774 }
 775
 776 int
 777 vnode_pager_local_getpages_async(struct vop_getpages_async_args *ap)
 778 {
 779         int error;
 780
 781         error = vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count,
 782             ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg);
 783         if (error != 0 && ap->a_iodone != NULL)
 784                 ap->a_iodone(ap->a_arg, ap->a_m, ap->a_count, error);
 785         return (error);
 786 }
 787
 788 /*
 789  * This is now called from local media FS's to operate against their
 790  * own vnodes if they fail to implement VOP_GETPAGES.
 791  */
 792 int
 793 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count,
 794     int *a_rbehind, int *a_rahead, vop_getpages_iodone_t iodone, void *arg)
 795 {
 796         vm_object_t object;
 797         struct bufobj *bo;
 798         struct buf *bp;
 799         off_t foff;
 800 #ifdef INVARIANTS
 801         off_t blkno0;
 802 #endif
 803         int bsize, pagesperblock;
 804         int error, before, after, rbehind, rahead, poff, i;
 805         int bytecount, secmask;
 806
 807         KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
 808             ("%s does not support devices", __func__));
 809
 810         if (VN_IS_DOOMED(vp))
 811                 return (VM_PAGER_BAD);
 812
 813         object = vp->v_object;
 814         foff = IDX_TO_OFF(m[0]->pindex);
 815         bsize = vp->v_mount->mnt_stat.f_iosize;
 816         pagesperblock = bsize / PAGE_SIZE;
 817
 818         KASSERT(foff < object->un_pager.vnp.vnp_size,
 819             ("%s: page %p offset beyond vp %p size", __func__, m[0], vp));
 820         KASSERT(count <= atop(maxphys),
 821             ("%s: requested %d pages", __func__, count));
 822
 823         /*
 824          * The last page has valid blocks.  Invalid part can only
 825          * exist at the end of file, and the page is made fully valid
 826          * by zeroing in vm_pager_get_pages().
 827          */
 828         if (!vm_page_none_valid(m[count - 1]) && --count == 0) {
 829                 if (iodone != NULL)
 830                         iodone(arg, m, 1, 0);
 831                 return (VM_PAGER_OK);
 832         }
 833
 834         bp = uma_zalloc(vnode_pbuf_zone, M_WAITOK);
 835         MPASS((bp->b_flags & B_MAXPHYS) != 0);
 836
 837         /*
 838          * Get the underlying device blocks for the file with VOP_BMAP().
 839          * If the file system doesn't support VOP_BMAP, use old way of
 840          * getting pages via VOP_READ.
 841          */
 842         error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before);
 843         if (error == EOPNOTSUPP) {
 844                 uma_zfree(vnode_pbuf_zone, bp);
 845                 VM_OBJECT_WLOCK(object);
 846                 for (i = 0; i < count; i++) {
 847                         VM_CNT_INC(v_vnodein);
 848                         VM_CNT_INC(v_vnodepgsin);
 849                         error = vnode_pager_input_old(object, m[i]);
 850                         if (error)
 851                                 break;
 852                 }
 853                 VM_OBJECT_WUNLOCK(object);
 854                 return (error);
 855         } else if (error != 0) {
 856                 uma_zfree(vnode_pbuf_zone, bp);
 857                 return (VM_PAGER_ERROR);
 858         }
 859
 860         /*
 861          * If the file system supports BMAP, but blocksize is smaller
 862          * than a page size, then use special small filesystem code.
 863          */
 864         if (pagesperblock == 0) {
 865                 uma_zfree(vnode_pbuf_zone, bp);
 866                 for (i = 0; i < count; i++) {
 867                         VM_CNT_INC(v_vnodein);
 868                         VM_CNT_INC(v_vnodepgsin);
 869                         error = vnode_pager_input_smlfs(object, m[i]);
 870                         if (error)
 871                                 break;
 872                 }
 873                 return (error);
 874         }
 875
 876         /*
 877          * A sparse file can be encountered only for a single page request,
 878          * which may not be preceded by call to vm_pager_haspage().
 879          */
 880         if (bp->b_blkno == -1) {
 881                 KASSERT(count == 1,
 882                     ("%s: array[%d] request to a sparse file %p", __func__,
 883                     count, vp));
 884                 uma_zfree(vnode_pbuf_zone, bp);
 885                 pmap_zero_page(m[0]);
 886                 KASSERT(m[0]->dirty == 0, ("%s: page %p is dirty",
 887                     __func__, m[0]));
 888                 vm_page_valid(m[0]);
 889                 return (VM_PAGER_OK);
 890         }
 891
 892 #ifdef INVARIANTS
 893         blkno0 = bp->b_blkno;
 894 #endif
 895         bp->b_blkno += (foff % bsize) / DEV_BSIZE;
 896
 897         /* Recalculate blocks available after/before to pages. */
 898         poff = (foff % bsize) / PAGE_SIZE;
 899         before *= pagesperblock;
 900         before += poff;
 901         after *= pagesperblock;
 902         after += pagesperblock - (poff + 1);
 903         if (m[0]->pindex + after >= object->size)
 904                 after = object->size - 1 - m[0]->pindex;
 905         KASSERT(count <= after + 1, ("%s: %d pages asked, can do only %d",
 906             __func__, count, after + 1));
 907         after -= count - 1;
 908
 909         /* Trim requested rbehind/rahead to possible values. */
 910         rbehind = a_rbehind ? *a_rbehind : 0;
 911         rahead = a_rahead ? *a_rahead : 0;
 912         rbehind = min(rbehind, before);
 913         rbehind = min(rbehind, m[0]->pindex);
 914         rahead = min(rahead, after);
 915         rahead = min(rahead, object->size - m[count - 1]->pindex);
 916         /*
 917          * Check that total amount of pages fit into buf.  Trim rbehind and
 918          * rahead evenly if not.
 919          */
 920         if (rbehind + rahead + count > atop(maxphys)) {
 921                 int trim, sum;
 922
 923                 trim = rbehind + rahead + count - atop(maxphys) + 1;
 924                 sum = rbehind + rahead;
 925                 if (rbehind == before) {
 926                         /* Roundup rbehind trim to block size. */
 927                         rbehind -= roundup(trim * rbehind / sum, pagesperblock);
 928                         if (rbehind < 0)
 929                                 rbehind = 0;
 930                 } else
 931                         rbehind -= trim * rbehind / sum;
 932                 rahead -= trim * rahead / sum;
 933         }
 934         KASSERT(rbehind + rahead + count <= atop(maxphys),
 935             ("%s: behind %d ahead %d count %d maxphys %lu", __func__,
 936             rbehind, rahead, count, maxphys));
 937
 938         /*
 939          * Fill in the bp->b_pages[] array with requested and optional
 940          * read behind or read ahead pages.  Read behind pages are looked
 941          * up in a backward direction, down to a first cached page.  Same
 942          * for read ahead pages, but there is no need to shift the array
 943          * in case of encountering a cached page.
 944          */
 945         i = bp->b_npages = 0;
 946         if (rbehind) {
 947                 vm_pindex_t startpindex, tpindex;
 948                 vm_page_t p;
 949
 950                 VM_OBJECT_WLOCK(object);
 951                 startpindex = m[0]->pindex - rbehind;
 952                 if ((p = TAILQ_PREV(m[0], pglist, listq)) != NULL &&
 953                     p->pindex >= startpindex)
 954                         startpindex = p->pindex + 1;
 955
 956                 /* tpindex is unsigned; beware of numeric underflow. */
 957                 for (tpindex = m[0]->pindex - 1;
 958                     tpindex >= startpindex && tpindex < m[0]->pindex;
 959                     tpindex--, i++) {
 960                         p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
 961                         if (p == NULL) {
 962                                 /* Shift the array. */
 963                                 for (int j = 0; j < i; j++)
 964                                         bp->b_pages[j] = bp->b_pages[j +
 965                                             tpindex + 1 - startpindex];
 966                                 break;
 967                         }
 968                         bp->b_pages[tpindex - startpindex] = p;
 969                 }
 970
 971                 bp->b_pgbefore = i;
 972                 bp->b_npages += i;
 973                 bp->b_blkno -= IDX_TO_OFF(i) / DEV_BSIZE;
 974         } else
 975                 bp->b_pgbefore = 0;
 976
 977         /* Requested pages. */
 978         for (int j = 0; j < count; j++, i++)
 979                 bp->b_pages[i] = m[j];
 980         bp->b_npages += count;
 981
 982         if (rahead) {
 983                 vm_pindex_t endpindex, tpindex;
 984                 vm_page_t p;
 985
 986                 if (!VM_OBJECT_WOWNED(object))
 987                         VM_OBJECT_WLOCK(object);
 988                 endpindex = m[count - 1]->pindex + rahead + 1;
 989                 if ((p = TAILQ_NEXT(m[count - 1], listq)) != NULL &&
 990                     p->pindex < endpindex)
 991                         endpindex = p->pindex;
 992                 if (endpindex > object->size)
 993                         endpindex = object->size;
 994
 995                 for (tpindex = m[count - 1]->pindex + 1;
 996                     tpindex < endpindex; i++, tpindex++) {
 997                         p = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
 998                         if (p == NULL)
 999                                 break;
1000                         bp->b_pages[i] = p;
1001                 }
1002
1003                 bp->b_pgafter = i - bp->b_npages;
1004                 bp->b_npages = i;
1005         } else
1006                 bp->b_pgafter = 0;
1007
1008         if (VM_OBJECT_WOWNED(object))
1009                 VM_OBJECT_WUNLOCK(object);
1010
1011         /* Report back actual behind/ahead read. */
1012         if (a_rbehind)
1013                 *a_rbehind = bp->b_pgbefore;
1014         if (a_rahead)
1015                 *a_rahead = bp->b_pgafter;
1016
1017 #ifdef INVARIANTS
1018         KASSERT(bp->b_npages <= atop(maxphys),
1019             ("%s: buf %p overflowed", __func__, bp));
1020         for (int j = 1, prev = 0; j < bp->b_npages; j++) {
1021                 if (bp->b_pages[j] == bogus_page)
1022                         continue;
1023                 KASSERT(bp->b_pages[j]->pindex - bp->b_pages[prev]->pindex ==
1024                     j - prev, ("%s: pages array not consecutive, bp %p",
1025                      __func__, bp));
1026                 prev = j;
1027         }
1028 #endif
1029
1030         /*
1031          * Recalculate first offset and bytecount with regards to read behind.
1032          * Truncate bytecount to vnode real size and round up physical size
1033          * for real devices.
1034          */
1035         foff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1036         bytecount = bp->b_npages << PAGE_SHIFT;
1037         if ((foff + bytecount) > object->un_pager.vnp.vnp_size)
1038                 bytecount = object->un_pager.vnp.vnp_size - foff;
1039         secmask = bo->bo_bsize - 1;
1040         KASSERT(secmask < PAGE_SIZE && secmask > 0,
1041             ("%s: sector size %d too large", __func__, secmask + 1));
1042         bytecount = (bytecount + secmask) & ~secmask;
1043
1044         /*
1045          * And map the pages to be read into the kva, if the filesystem
1046          * requires mapped buffers.
1047          */
1048         if ((vp->v_mount->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0 &&
1049             unmapped_buf_allowed) {
1050                 bp->b_data = unmapped_buf;
1051                 bp->b_offset = 0;
1052         } else {
1053                 bp->b_data = bp->b_kvabase;
1054                 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1055         }
1056
1057         /* Build a minimal buffer header. */
1058         bp->b_iocmd = BIO_READ;
1059         KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
1060         KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
1061         bp->b_rcred = crhold(curthread->td_ucred);
1062         bp->b_wcred = crhold(curthread->td_ucred);
1063         pbgetbo(bo, bp);
1064         bp->b_vp = vp;
1065         bp->b_bcount = bp->b_bufsize = bp->b_runningbufspace = bytecount;
1066         bp->b_iooffset = dbtob(bp->b_blkno);
1067         KASSERT(IDX_TO_OFF(m[0]->pindex - bp->b_pages[0]->pindex) ==
1068             (blkno0 - bp->b_blkno) * DEV_BSIZE +
1069             IDX_TO_OFF(m[0]->pindex) % bsize,
1070             ("wrong offsets bsize %d m[0] %ju b_pages[0] %ju "
1071             "blkno0 %ju b_blkno %ju", bsize,
1072             (uintmax_t)m[0]->pindex, (uintmax_t)bp->b_pages[0]->pindex,
1073             (uintmax_t)blkno0, (uintmax_t)bp->b_blkno));
1074
1075         atomic_add_long(&runningbufspace, bp->b_runningbufspace);
1076         VM_CNT_INC(v_vnodein);
1077         VM_CNT_ADD(v_vnodepgsin, bp->b_npages);
1078
1079         if (iodone != NULL) { /* async */
1080                 bp->b_pgiodone = iodone;
1081                 bp->b_caller1 = arg;
1082                 bp->b_iodone = vnode_pager_generic_getpages_done_async;
1083                 bp->b_flags |= B_ASYNC;
1084                 BUF_KERNPROC(bp);
1085                 bstrategy(bp);
1086                 return (VM_PAGER_OK);
1087         } else {
1088                 bp->b_iodone = bdone;
1089                 bstrategy(bp);
1090                 bwait(bp, PVM, "vnread");
1091                 error = vnode_pager_generic_getpages_done(bp);
1092                 for (i = 0; i < bp->b_npages; i++)
1093                         bp->b_pages[i] = NULL;
1094                 bp->b_vp = NULL;
1095                 pbrelbo(bp);
1096                 uma_zfree(vnode_pbuf_zone, bp);
1097                 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
1098         }
1099 }
1100
1101 static void
1102 vnode_pager_generic_getpages_done_async(struct buf *bp)
1103 {
1104         int error;
1105
1106         error = vnode_pager_generic_getpages_done(bp);
1107         /* Run the iodone upon the requested range. */
1108         bp->b_pgiodone(bp->b_caller1, bp->b_pages + bp->b_pgbefore,
1109             bp->b_npages - bp->b_pgbefore - bp->b_pgafter, error);
1110         for (int i = 0; i < bp->b_npages; i++)
1111                 bp->b_pages[i] = NULL;
1112         bp->b_vp = NULL;
1113         pbrelbo(bp);
1114         uma_zfree(vnode_pbuf_zone, bp);
1115 }
1116
1117 static int
1118 vnode_pager_generic_getpages_done(struct buf *bp)
1119 {
1120         vm_object_t object;
1121         off_t tfoff, nextoff;
1122         int i, error;
1123
1124         KASSERT((bp->b_ioflags & BIO_ERROR) == 0 || bp->b_error != 0,
1125             ("%s: buf error but b_error == 0\n", __func__));
1126         error = (bp->b_ioflags & BIO_ERROR) != 0 ? bp->b_error : 0;
1127         object = bp->b_vp->v_object;
1128
1129         if (error == 0 && bp->b_bcount != bp->b_npages * PAGE_SIZE) {
1130                 if (!buf_mapped(bp)) {
1131                         bp->b_data = bp->b_kvabase;
1132                         pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages,
1133                             bp->b_npages);
1134                 }
1135                 bzero(bp->b_data + bp->b_bcount,
1136                     PAGE_SIZE * bp->b_npages - bp->b_bcount);
1137         }
1138         if (buf_mapped(bp)) {
1139                 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1140                 bp->b_data = unmapped_buf;
1141         }
1142
1143         /*
1144          * If the read failed, we must free any read ahead/behind pages here.
1145          * The requested pages are freed by the caller (for sync requests)
1146          * or by the bp->b_pgiodone callback (for async requests).
1147          */
1148         if (error != 0) {
1149                 VM_OBJECT_WLOCK(object);
1150                 for (i = 0; i < bp->b_pgbefore; i++)
1151                         vm_page_free_invalid(bp->b_pages[i]);
1152                 for (i = bp->b_npages - bp->b_pgafter; i < bp->b_npages; i++)
1153                         vm_page_free_invalid(bp->b_pages[i]);
1154                 VM_OBJECT_WUNLOCK(object);
1155                 return (error);
1156         }
1157
1158         /* Read lock to protect size. */
1159         VM_OBJECT_RLOCK(object);
1160         for (i = 0, tfoff = IDX_TO_OFF(bp->b_pages[0]->pindex);
1161             i < bp->b_npages; i++, tfoff = nextoff) {
1162                 vm_page_t mt;
1163
1164                 nextoff = tfoff + PAGE_SIZE;
1165                 mt = bp->b_pages[i];
1166                 if (mt == bogus_page)
1167                         continue;
1168
1169                 if (nextoff <= object->un_pager.vnp.vnp_size) {
1170                         /*
1171                          * Read filled up entire page.
1172                          */
1173                         vm_page_valid(mt);
1174                         KASSERT(mt->dirty == 0,
1175                             ("%s: page %p is dirty", __func__, mt));
1176                         KASSERT(!pmap_page_is_mapped(mt),
1177                             ("%s: page %p is mapped", __func__, mt));
1178                 } else {
1179                         /*
1180                          * Read did not fill up entire page.
1181                          *
1182                          * Currently we do not set the entire page valid,
1183                          * we just try to clear the piece that we couldn't
1184                          * read.
1185                          */
1186                         vm_page_set_valid_range(mt, 0,
1187                             object->un_pager.vnp.vnp_size - tfoff);
1188                         KASSERT((mt->dirty & vm_page_bits(0,
1189                             object->un_pager.vnp.vnp_size - tfoff)) == 0,
1190                             ("%s: page %p is dirty", __func__, mt));
1191                 }
1192
1193                 if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter)
1194                         vm_page_readahead_finish(mt);
1195         }
1196         VM_OBJECT_RUNLOCK(object);
1197
1198         return (error);
1199 }
1200
1201 /*
1202  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
1203  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
1204  * vnode_pager_generic_putpages() to implement the previous behaviour.
1205  *
1206  * All other FS's should use the bypass to get to the local media
1207  * backing vp's VOP_PUTPAGES.
1208  */
1209 static void
1210 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1211     int flags, int *rtvals)
1212 {
1213         int rtval;
1214         struct vnode *vp;
1215         int bytes = count * PAGE_SIZE;
1216
1217         /*
1218          * Force synchronous operation if we are extremely low on memory
1219          * to prevent a low-memory deadlock.  VOP operations often need to
1220          * allocate more memory to initiate the I/O ( i.e. do a BMAP
1221          * operation ).  The swapper handles the case by limiting the amount
1222          * of asynchronous I/O, but that sort of solution doesn't scale well
1223          * for the vnode pager without a lot of work.
1224          *
1225          * Also, the backing vnode's iodone routine may not wake the pageout
1226          * daemon up.  This should be probably be addressed XXX.
1227          */
1228
1229         if (vm_page_count_min())
1230                 flags |= VM_PAGER_PUT_SYNC;
1231
1232         /*
1233          * Call device-specific putpages function
1234          */
1235         vp = object->handle;
1236         VM_OBJECT_WUNLOCK(object);
1237         rtval = VOP_PUTPAGES(vp, m, bytes, flags, rtvals);
1238         KASSERT(rtval != EOPNOTSUPP,
1239             ("vnode_pager: stale FS putpages\n"));
1240         VM_OBJECT_WLOCK(object);
1241 }
1242
1243 static int
1244 vn_off2bidx(vm_ooffset_t offset)
1245 {
1246
1247         return ((offset & PAGE_MASK) / DEV_BSIZE);
1248 }
1249
1250 static bool
1251 vn_dirty_blk(vm_page_t m, vm_ooffset_t offset)
1252 {
1253
1254         KASSERT(IDX_TO_OFF(m->pindex) <= offset &&
1255             offset < IDX_TO_OFF(m->pindex + 1),
1256             ("page %p pidx %ju offset %ju", m, (uintmax_t)m->pindex,
1257             (uintmax_t)offset));
1258         return ((m->dirty & ((vm_page_bits_t)1 << vn_off2bidx(offset))) != 0);
1259 }
1260
1261 /*
1262  * This is now called from local media FS's to operate against their
1263  * own vnodes if they fail to implement VOP_PUTPAGES.
1264  *
1265  * This is typically called indirectly via the pageout daemon and
1266  * clustering has already typically occurred, so in general we ask the
1267  * underlying filesystem to write the data out asynchronously rather
1268  * then delayed.
1269  */
1270 int
1271 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount,
1272     int flags, int *rtvals)
1273 {
1274         vm_object_t object;
1275         vm_page_t m;
1276         vm_ooffset_t maxblksz, next_offset, poffset, prev_offset;
1277         struct uio auio;
1278         struct iovec aiov;
1279         off_t prev_resid, wrsz;
1280         int count, error, i, maxsize, ncount, pgoff, ppscheck;
1281         bool in_hole;
1282         static struct timeval lastfail;
1283         static int curfail;
1284
1285         object = vp->v_object;
1286         count = bytecount / PAGE_SIZE;
1287
1288         for (i = 0; i < count; i++)
1289                 rtvals[i] = VM_PAGER_ERROR;
1290
1291         if ((int64_t)ma[0]->pindex < 0) {
1292                 printf("vnode_pager_generic_putpages: "
1293                     "attempt to write meta-data 0x%jx(%lx)\n",
1294                     (uintmax_t)ma[0]->pindex, (u_long)ma[0]->dirty);
1295                 rtvals[0] = VM_PAGER_BAD;
1296                 return (VM_PAGER_BAD);
1297         }
1298
1299         maxsize = count * PAGE_SIZE;
1300         ncount = count;
1301
1302         poffset = IDX_TO_OFF(ma[0]->pindex);
1303
1304         /*
1305          * If the page-aligned write is larger then the actual file we
1306          * have to invalidate pages occurring beyond the file EOF.  However,
1307          * there is an edge case where a file may not be page-aligned where
1308          * the last page is partially invalid.  In this case the filesystem
1309          * may not properly clear the dirty bits for the entire page (which
1310          * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1311          * With the page busied we are free to fix up the dirty bits here.
1312          *
1313          * We do not under any circumstances truncate the valid bits, as
1314          * this will screw up bogus page replacement.
1315          */
1316         VM_OBJECT_RLOCK(object);
1317         if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1318                 if (object->un_pager.vnp.vnp_size > poffset) {
1319                         maxsize = object->un_pager.vnp.vnp_size - poffset;
1320                         ncount = btoc(maxsize);
1321                         if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1322                                 pgoff = roundup2(pgoff, DEV_BSIZE);
1323
1324                                 /*
1325                                  * If the page is busy and the following
1326                                  * conditions hold, then the page's dirty
1327                                  * field cannot be concurrently changed by a
1328                                  * pmap operation.
1329                                  */
1330                                 m = ma[ncount - 1];
1331                                 vm_page_assert_sbusied(m);
1332                                 KASSERT(!pmap_page_is_write_mapped(m),
1333                 ("vnode_pager_generic_putpages: page %p is not read-only", m));
1334                                 MPASS(m->dirty != 0);
1335                                 vm_page_clear_dirty(m, pgoff, PAGE_SIZE -
1336                                     pgoff);
1337                         }
1338                 } else {
1339                         maxsize = 0;
1340                         ncount = 0;
1341                 }
1342                 for (i = ncount; i < count; i++)
1343                         rtvals[i] = VM_PAGER_BAD;
1344         }
1345         VM_OBJECT_RUNLOCK(object);
1346
1347         auio.uio_iov = &aiov;
1348         auio.uio_segflg = UIO_NOCOPY;
1349         auio.uio_rw = UIO_WRITE;
1350         auio.uio_td = NULL;
1351         maxblksz = roundup2(poffset + maxsize, DEV_BSIZE);
1352
1353         for (prev_offset = poffset; prev_offset < maxblksz;) {
1354                 /* Skip clean blocks. */
1355                 for (in_hole = true; in_hole && prev_offset < maxblksz;) {
1356                         m = ma[OFF_TO_IDX(prev_offset - poffset)];
1357                         for (i = vn_off2bidx(prev_offset);
1358                             i < sizeof(vm_page_bits_t) * NBBY &&
1359                             prev_offset < maxblksz; i++) {
1360                                 if (vn_dirty_blk(m, prev_offset)) {
1361                                         in_hole = false;
1362                                         break;
1363                                 }
1364                                 prev_offset += DEV_BSIZE;
1365                         }
1366                 }
1367                 if (in_hole)
1368                         goto write_done;
1369
1370                 /* Find longest run of dirty blocks. */
1371                 for (next_offset = prev_offset; next_offset < maxblksz;) {
1372                         m = ma[OFF_TO_IDX(next_offset - poffset)];
1373                         for (i = vn_off2bidx(next_offset);
1374                             i < sizeof(vm_page_bits_t) * NBBY &&
1375                             next_offset < maxblksz; i++) {
1376                                 if (!vn_dirty_blk(m, next_offset))
1377                                         goto start_write;
1378                                 next_offset += DEV_BSIZE;
1379                         }
1380                 }
1381 start_write:
1382                 if (next_offset > poffset + maxsize)
1383                         next_offset = poffset + maxsize;
1384
1385                 /*
1386                  * Getting here requires finding a dirty block in the
1387                  * 'skip clean blocks' loop.
1388                  */
1389                 MPASS(prev_offset < next_offset);
1390
1391                 aiov.iov_base = NULL;
1392                 auio.uio_iovcnt = 1;
1393                 auio.uio_offset = prev_offset;
1394                 prev_resid = auio.uio_resid = aiov.iov_len = next_offset -
1395                     prev_offset;
1396                 error = VOP_WRITE(vp, &auio,
1397                     vnode_pager_putpages_ioflags(flags), curthread->td_ucred);
1398
1399                 wrsz = prev_resid - auio.uio_resid;
1400                 if (wrsz == 0) {
1401                         if (ppsratecheck(&lastfail, &curfail, 1) != 0) {
1402                                 vn_printf(vp, "vnode_pager_putpages: "
1403                                     "zero-length write at %ju resid %zd\n",
1404                                     auio.uio_offset, auio.uio_resid);
1405                         }
1406                         break;
1407                 }
1408
1409                 /* Adjust the starting offset for next iteration. */
1410                 prev_offset += wrsz;
1411                 MPASS(auio.uio_offset == prev_offset);
1412
1413                 ppscheck = 0;
1414                 if (error != 0 && (ppscheck = ppsratecheck(&lastfail,
1415                     &curfail, 1)) != 0)
1416                         vn_printf(vp, "vnode_pager_putpages: I/O error %d\n",
1417                             error);
1418                 if (auio.uio_resid != 0 && (ppscheck != 0 ||
1419                     ppsratecheck(&lastfail, &curfail, 1) != 0))
1420                         vn_printf(vp, "vnode_pager_putpages: residual I/O %zd "
1421                             "at %ju\n", auio.uio_resid,
1422                             (uintmax_t)ma[0]->pindex);
1423                 if (error != 0 || auio.uio_resid != 0)
1424                         break;
1425         }
1426 write_done:
1427         /* Mark completely processed pages. */
1428         for (i = 0; i < OFF_TO_IDX(prev_offset - poffset); i++)
1429                 rtvals[i] = VM_PAGER_OK;
1430         /* Mark partial EOF page. */
1431         if (prev_offset == poffset + maxsize && (prev_offset & PAGE_MASK) != 0)
1432                 rtvals[i++] = VM_PAGER_OK;
1433         /* Unwritten pages in range, free bonus if the page is clean. */
1434         for (; i < ncount; i++)
1435                 rtvals[i] = ma[i]->dirty == 0 ? VM_PAGER_OK : VM_PAGER_ERROR;
1436         VM_CNT_ADD(v_vnodepgsout, i);
1437         VM_CNT_INC(v_vnodeout);
1438         return (rtvals[0]);
1439 }
1440
1441 int
1442 vnode_pager_putpages_ioflags(int pager_flags)
1443 {
1444         int ioflags;
1445
1446         /*
1447          * Pageouts are already clustered, use IO_ASYNC to force a
1448          * bawrite() rather then a bdwrite() to prevent paging I/O
1449          * from saturating the buffer cache.  Dummy-up the sequential
1450          * heuristic to cause large ranges to cluster.  If neither
1451          * IO_SYNC or IO_ASYNC is set, the system decides how to
1452          * cluster.
1453          */
1454         ioflags = IO_VMIO;
1455         if ((pager_flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) != 0)
1456                 ioflags |= IO_SYNC;
1457         else if ((pager_flags & VM_PAGER_CLUSTER_OK) == 0)
1458                 ioflags |= IO_ASYNC;
1459         ioflags |= (pager_flags & VM_PAGER_PUT_INVAL) != 0 ? IO_INVAL: 0;
1460         ioflags |= (pager_flags & VM_PAGER_PUT_NOREUSE) != 0 ? IO_NOREUSE : 0;
1461         ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1462         return (ioflags);
1463 }
1464
1465 /*
1466  * vnode_pager_undirty_pages().
1467  *
1468  * A helper to mark pages as clean after pageout that was possibly
1469  * done with a short write.  The lpos argument specifies the page run
1470  * length in bytes, and the written argument specifies how many bytes
1471  * were actually written.  eof is the offset past the last valid byte
1472  * in the vnode using the absolute file position of the first byte in
1473  * the run as the base from which it is computed.
1474  */
1475 void
1476 vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written, off_t eof,
1477     int lpos)
1478 {
1479         vm_object_t obj;
1480         int i, pos, pos_devb;
1481
1482         if (written == 0 && eof >= lpos)
1483                 return;
1484         obj = ma[0]->object;
1485         for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) {
1486                 if (pos < trunc_page(written)) {
1487                         rtvals[i] = VM_PAGER_OK;
1488                         vm_page_undirty(ma[i]);
1489                 } else {
1490                         /* Partially written page. */
1491                         rtvals[i] = VM_PAGER_AGAIN;
1492                         vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK);
1493                 }
1494         }
1495         if (eof >= lpos) /* avoid truncation */
1496                 return;
1497         for (pos = eof, i = OFF_TO_IDX(trunc_page(pos)); pos < lpos; i++) {
1498                 if (pos != trunc_page(pos)) {
1499                         /*
1500                          * The page contains the last valid byte in
1501                          * the vnode, mark the rest of the page as
1502                          * clean, potentially making the whole page
1503                          * clean.
1504                          */
1505                         pos_devb = roundup2(pos & PAGE_MASK, DEV_BSIZE);
1506                         vm_page_clear_dirty(ma[i], pos_devb, PAGE_SIZE -
1507                             pos_devb);
1508
1509                         /*
1510                          * If the page was cleaned, report the pageout
1511                          * on it as successful.  msync() no longer
1512                          * needs to write out the page, endlessly
1513                          * creating write requests and dirty buffers.
1514                          */
1515                         if (ma[i]->dirty == 0)
1516                                 rtvals[i] = VM_PAGER_OK;
1517
1518                         pos = round_page(pos);
1519                 } else {
1520                         /* vm_pageout_flush() clears dirty */
1521                         rtvals[i] = VM_PAGER_BAD;
1522                         pos += PAGE_SIZE;
1523                 }
1524         }
1525 }
1526
1527 static void
1528 vnode_pager_update_writecount(vm_object_t object, vm_offset_t start,
1529     vm_offset_t end)
1530 {
1531         struct vnode *vp;
1532         vm_ooffset_t old_wm;
1533
1534         VM_OBJECT_WLOCK(object);
1535         if (object->type != OBJT_VNODE) {
1536                 VM_OBJECT_WUNLOCK(object);
1537                 return;
1538         }
1539         old_wm = object->un_pager.vnp.writemappings;
1540         object->un_pager.vnp.writemappings += (vm_ooffset_t)end - start;
1541         vp = object->handle;
1542         if (old_wm == 0 && object->un_pager.vnp.writemappings != 0) {
1543                 ASSERT_VOP_LOCKED(vp, "v_writecount inc");
1544                 VOP_ADD_WRITECOUNT_CHECKED(vp, 1);
1545                 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
1546                     __func__, vp, vp->v_writecount);
1547         } else if (old_wm != 0 && object->un_pager.vnp.writemappings == 0) {
1548                 ASSERT_VOP_LOCKED(vp, "v_writecount dec");
1549                 VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1550                 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
1551                     __func__, vp, vp->v_writecount);
1552         }
1553         VM_OBJECT_WUNLOCK(object);
1554 }
1555
1556 static void
1557 vnode_pager_release_writecount(vm_object_t object, vm_offset_t start,
1558     vm_offset_t end)
1559 {
1560         struct vnode *vp;
1561         struct mount *mp;
1562         vm_offset_t inc;
1563
1564         VM_OBJECT_WLOCK(object);
1565
1566         /*
1567          * First, recheck the object type to account for the race when
1568          * the vnode is reclaimed.
1569          */
1570         if (object->type != OBJT_VNODE) {
1571                 VM_OBJECT_WUNLOCK(object);
1572                 return;
1573         }
1574
1575         /*
1576          * Optimize for the case when writemappings is not going to
1577          * zero.
1578          */
1579         inc = end - start;
1580         if (object->un_pager.vnp.writemappings != inc) {
1581                 object->un_pager.vnp.writemappings -= inc;
1582                 VM_OBJECT_WUNLOCK(object);
1583                 return;
1584         }
1585
1586         vp = object->handle;
1587         vhold(vp);
1588         VM_OBJECT_WUNLOCK(object);
1589         mp = NULL;
1590         vn_start_write(vp, &mp, V_WAIT);
1591         vn_lock(vp, LK_SHARED | LK_RETRY);
1592
1593         /*
1594          * Decrement the object's writemappings, by swapping the start
1595          * and end arguments for vnode_pager_update_writecount().  If
1596          * there was not a race with vnode reclaimation, then the
1597          * vnode's v_writecount is decremented.
1598          */
1599         vnode_pager_update_writecount(object, end, start);
1600         VOP_UNLOCK(vp);
1601         vdrop(vp);
1602         if (mp != NULL)
1603                 vn_finished_write(mp);
1604 }