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