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