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