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