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