1 /* 2 * Copyright (c) 1990 University of Utah. 3 * Copyright (c) 1991, 1993 4 * The Regents of the University of California. All rights reserved. 5 * 6 * This code is derived from software contributed to Berkeley by 7 * the Systems Programming Group of the University of Utah Computer 8 * Science Department. 9 * 10 * %sccs.include.redist.c% 11 * 12 * @(#)vnode_pager.c 8.8 (Berkeley) 02/13/94 13 */ 14 15 /* 16 * Page to/from files (vnodes). 17 * 18 * TODO: 19 * pageouts 20 * fix credential use (uses current process credentials now) 21 */ 22 23 #include <sys/param.h> 24 #include <sys/systm.h> 25 #include <sys/proc.h> 26 #include <sys/malloc.h> 27 #include <sys/vnode.h> 28 #include <sys/uio.h> 29 #include <sys/mount.h> 30 31 #include <vm/vm.h> 32 #include <vm/vm_page.h> 33 #include <vm/vnode_pager.h> 34 35 struct pagerlst vnode_pager_list; /* list of managed vnodes */ 36 37 #ifdef DEBUG 38 int vpagerdebug = 0x00; 39 #define VDB_FOLLOW 0x01 40 #define VDB_INIT 0x02 41 #define VDB_IO 0x04 42 #define VDB_FAIL 0x08 43 #define VDB_ALLOC 0x10 44 #define VDB_SIZE 0x20 45 #endif 46 47 static vm_pager_t vnode_pager_alloc 48 __P((caddr_t, vm_size_t, vm_prot_t, vm_offset_t)); 49 static void vnode_pager_cluster 50 __P((vm_pager_t, vm_offset_t, 51 vm_offset_t *, vm_offset_t *)); 52 static void vnode_pager_dealloc __P((vm_pager_t)); 53 static int vnode_pager_getpage 54 __P((vm_pager_t, vm_page_t *, int, boolean_t)); 55 static boolean_t vnode_pager_haspage __P((vm_pager_t, vm_offset_t)); 56 static void vnode_pager_init __P((void)); 57 static int vnode_pager_io 58 __P((vn_pager_t, vm_page_t *, int, 59 boolean_t, enum uio_rw)); 60 static boolean_t vnode_pager_putpage 61 __P((vm_pager_t, vm_page_t *, int, boolean_t)); 62 63 struct pagerops vnodepagerops = { 64 vnode_pager_init, 65 vnode_pager_alloc, 66 vnode_pager_dealloc, 67 vnode_pager_getpage, 68 vnode_pager_putpage, 69 vnode_pager_haspage, 70 vnode_pager_cluster 71 }; 72 73 static void 74 vnode_pager_init() 75 { 76 #ifdef DEBUG 77 if (vpagerdebug & VDB_FOLLOW) 78 printf("vnode_pager_init()\n"); 79 #endif 80 TAILQ_INIT(&vnode_pager_list); 81 } 82 83 /* 84 * Allocate (or lookup) pager for a vnode. 85 * Handle is a vnode pointer. 86 */ 87 static vm_pager_t 88 vnode_pager_alloc(handle, size, prot, foff) 89 caddr_t handle; 90 vm_size_t size; 91 vm_prot_t prot; 92 vm_offset_t foff; 93 { 94 register vm_pager_t pager; 95 register vn_pager_t vnp; 96 vm_object_t object; 97 struct vattr vattr; 98 struct vnode *vp; 99 struct proc *p = curproc; /* XXX */ 100 101 #ifdef DEBUG 102 if (vpagerdebug & (VDB_FOLLOW|VDB_ALLOC)) 103 printf("vnode_pager_alloc(%x, %x, %x)\n", handle, size, prot); 104 #endif 105 /* 106 * Pageout to vnode, no can do yet. 107 */ 108 if (handle == NULL) 109 return(NULL); 110 111 /* 112 * Vnodes keep a pointer to any associated pager so no need to 113 * lookup with vm_pager_lookup. 114 */ 115 vp = (struct vnode *)handle; 116 pager = (vm_pager_t)vp->v_vmdata; 117 if (pager == NULL) { 118 /* 119 * Allocate pager structures 120 */ 121 pager = (vm_pager_t)malloc(sizeof *pager, M_VMPAGER, M_WAITOK); 122 if (pager == NULL) 123 return(NULL); 124 vnp = (vn_pager_t)malloc(sizeof *vnp, M_VMPGDATA, M_WAITOK); 125 if (vnp == NULL) { 126 free((caddr_t)pager, M_VMPAGER); 127 return(NULL); 128 } 129 /* 130 * And an object of the appropriate size 131 */ 132 if (VOP_GETATTR(vp, &vattr, p->p_ucred, p) == 0) { 133 object = vm_object_allocate(round_page(vattr.va_size)); 134 vm_object_enter(object, pager); 135 vm_object_setpager(object, pager, 0, TRUE); 136 } else { 137 free((caddr_t)vnp, M_VMPGDATA); 138 free((caddr_t)pager, M_VMPAGER); 139 return(NULL); 140 } 141 /* 142 * Hold a reference to the vnode and initialize pager data. 143 */ 144 VREF(vp); 145 vnp->vnp_flags = 0; 146 vnp->vnp_vp = vp; 147 vnp->vnp_size = vattr.va_size; 148 TAILQ_INSERT_TAIL(&vnode_pager_list, pager, pg_list); 149 pager->pg_handle = handle; 150 pager->pg_type = PG_VNODE; 151 pager->pg_flags = 0; 152 pager->pg_ops = &vnodepagerops; 153 pager->pg_data = vnp; 154 vp->v_vmdata = (caddr_t)pager; 155 } else { 156 /* 157 * vm_object_lookup() will remove the object from the 158 * cache if found and also gain a reference to the object. 159 */ 160 object = vm_object_lookup(pager); 161 #ifdef DEBUG 162 vnp = (vn_pager_t)pager->pg_data; 163 #endif 164 } 165 #ifdef DEBUG 166 if (vpagerdebug & VDB_ALLOC) 167 printf("vnode_pager_setup: vp %x sz %x pager %x object %x\n", 168 vp, vnp->vnp_size, pager, object); 169 #endif 170 return(pager); 171 } 172 173 static void 174 vnode_pager_dealloc(pager) 175 vm_pager_t pager; 176 { 177 register vn_pager_t vnp = (vn_pager_t)pager->pg_data; 178 register struct vnode *vp; 179 #ifdef NOTDEF 180 struct proc *p = curproc; /* XXX */ 181 #endif 182 183 #ifdef DEBUG 184 if (vpagerdebug & VDB_FOLLOW) 185 printf("vnode_pager_dealloc(%x)\n", pager); 186 #endif 187 if (vp = vnp->vnp_vp) { 188 vp->v_vmdata = NULL; 189 vp->v_flag &= ~VTEXT; 190 #if NOTDEF 191 /* can hang if done at reboot on NFS FS */ 192 (void) VOP_FSYNC(vp, p->p_ucred, p); 193 #endif 194 vrele(vp); 195 } 196 TAILQ_REMOVE(&vnode_pager_list, pager, pg_list); 197 free((caddr_t)vnp, M_VMPGDATA); 198 free((caddr_t)pager, M_VMPAGER); 199 } 200 201 static int 202 vnode_pager_getpage(pager, mlist, npages, sync) 203 vm_pager_t pager; 204 vm_page_t *mlist; 205 int npages; 206 boolean_t sync; 207 { 208 209 #ifdef DEBUG 210 if (vpagerdebug & VDB_FOLLOW) 211 printf("vnode_pager_getpage(%x, %x, %x, %x)\n", 212 pager, mlist, npages, sync); 213 #endif 214 return(vnode_pager_io((vn_pager_t)pager->pg_data, 215 mlist, npages, sync, UIO_READ)); 216 } 217 218 static boolean_t 219 vnode_pager_putpage(pager, mlist, npages, sync) 220 vm_pager_t pager; 221 vm_page_t *mlist; 222 int npages; 223 boolean_t sync; 224 { 225 int err; 226 227 #ifdef DEBUG 228 if (vpagerdebug & VDB_FOLLOW) 229 printf("vnode_pager_putpage(%x, %x, %x, %x)\n", 230 pager, mlist, npages, sync); 231 #endif 232 if (pager == NULL) 233 return (FALSE); /* ??? */ 234 err = vnode_pager_io((vn_pager_t)pager->pg_data, 235 mlist, npages, sync, UIO_WRITE); 236 /* 237 * If the operation was successful, mark the pages clean. 238 */ 239 if (err == VM_PAGER_OK) { 240 while (npages--) { 241 (*mlist)->flags |= PG_CLEAN; 242 pmap_clear_modify(VM_PAGE_TO_PHYS(*mlist)); 243 mlist++; 244 } 245 } 246 return(err); 247 } 248 249 static boolean_t 250 vnode_pager_haspage(pager, offset) 251 vm_pager_t pager; 252 vm_offset_t offset; 253 { 254 register vn_pager_t vnp = (vn_pager_t)pager->pg_data; 255 daddr_t bn; 256 int err; 257 258 #ifdef DEBUG 259 if (vpagerdebug & VDB_FOLLOW) 260 printf("vnode_pager_haspage(%x, %x)\n", pager, offset); 261 #endif 262 263 /* 264 * Offset beyond end of file, do not have the page 265 * Lock the vnode first to make sure we have the most recent 266 * version of the size. 267 */ 268 VOP_LOCK(vnp->vnp_vp); 269 if (offset >= vnp->vnp_size) { 270 VOP_UNLOCK(vnp->vnp_vp); 271 #ifdef DEBUG 272 if (vpagerdebug & (VDB_FAIL|VDB_SIZE)) 273 printf("vnode_pager_haspage: pg %x, off %x, size %x\n", 274 pager, offset, vnp->vnp_size); 275 #endif 276 return(FALSE); 277 } 278 279 /* 280 * Read the index to find the disk block to read 281 * from. If there is no block, report that we don't 282 * have this data. 283 * 284 * Assumes that the vnode has whole page or nothing. 285 */ 286 err = VOP_BMAP(vnp->vnp_vp, 287 offset / vnp->vnp_vp->v_mount->mnt_stat.f_iosize, 288 (struct vnode **)0, &bn, NULL); 289 VOP_UNLOCK(vnp->vnp_vp); 290 if (err) { 291 #ifdef DEBUG 292 if (vpagerdebug & VDB_FAIL) 293 printf("vnode_pager_haspage: BMAP err %d, pg %x, off %x\n", 294 err, pager, offset); 295 #endif 296 return(TRUE); 297 } 298 return((long)bn < 0 ? FALSE : TRUE); 299 } 300 301 static void 302 vnode_pager_cluster(pager, offset, loffset, hoffset) 303 vm_pager_t pager; 304 vm_offset_t offset; 305 vm_offset_t *loffset; 306 vm_offset_t *hoffset; 307 { 308 vn_pager_t vnp = (vn_pager_t)pager->pg_data; 309 vm_offset_t loff, hoff; 310 311 #ifdef DEBUG 312 if (vpagerdebug & VDB_FOLLOW) 313 printf("vnode_pager_cluster(%x, %x) ", pager, offset); 314 #endif 315 loff = offset; 316 if (loff >= vnp->vnp_size) 317 panic("vnode_pager_cluster: bad offset"); 318 /* 319 * XXX could use VOP_BMAP to get maxcontig value 320 */ 321 hoff = loff + MAXBSIZE; 322 if (hoff > round_page(vnp->vnp_size)) 323 hoff = round_page(vnp->vnp_size); 324 325 *loffset = loff; 326 *hoffset = hoff; 327 #ifdef DEBUG 328 if (vpagerdebug & VDB_FOLLOW) 329 printf("returns [%x-%x]\n", loff, hoff); 330 #endif 331 } 332 333 /* 334 * (XXX) 335 * Lets the VM system know about a change in size for a file. 336 * If this vnode is mapped into some address space (i.e. we have a pager 337 * for it) we adjust our own internal size and flush any cached pages in 338 * the associated object that are affected by the size change. 339 * 340 * Note: this routine may be invoked as a result of a pager put 341 * operation (possibly at object termination time), so we must be careful. 342 */ 343 void 344 vnode_pager_setsize(vp, nsize) 345 struct vnode *vp; 346 u_long nsize; 347 { 348 register vn_pager_t vnp; 349 register vm_object_t object; 350 vm_pager_t pager; 351 352 /* 353 * Not a mapped vnode 354 */ 355 if (vp == NULL || vp->v_type != VREG || vp->v_vmdata == NULL) 356 return; 357 /* 358 * Hasn't changed size 359 */ 360 pager = (vm_pager_t)vp->v_vmdata; 361 vnp = (vn_pager_t)pager->pg_data; 362 if (nsize == vnp->vnp_size) 363 return; 364 /* 365 * No object. 366 * This can happen during object termination since 367 * vm_object_page_clean is called after the object 368 * has been removed from the hash table, and clean 369 * may cause vnode write operations which can wind 370 * up back here. 371 */ 372 object = vm_object_lookup(pager); 373 if (object == NULL) 374 return; 375 376 #ifdef DEBUG 377 if (vpagerdebug & (VDB_FOLLOW|VDB_SIZE)) 378 printf("vnode_pager_setsize: vp %x obj %x osz %d nsz %d\n", 379 vp, object, vnp->vnp_size, nsize); 380 #endif 381 /* 382 * File has shrunk. 383 * Toss any cached pages beyond the new EOF. 384 */ 385 if (nsize < vnp->vnp_size) { 386 vm_object_lock(object); 387 vm_object_page_remove(object, 388 (vm_offset_t)nsize, vnp->vnp_size); 389 vm_object_unlock(object); 390 } 391 vnp->vnp_size = (vm_offset_t)nsize; 392 vm_object_deallocate(object); 393 } 394 395 void 396 vnode_pager_umount(mp) 397 register struct mount *mp; 398 { 399 register vm_pager_t pager, npager; 400 struct vnode *vp; 401 402 for (pager = vnode_pager_list.tqh_first; pager != NULL; pager = npager){ 403 /* 404 * Save the next pointer now since uncaching may 405 * terminate the object and render pager invalid 406 */ 407 npager = pager->pg_list.tqe_next; 408 vp = ((vn_pager_t)pager->pg_data)->vnp_vp; 409 if (mp == (struct mount *)0 || vp->v_mount == mp) { 410 VOP_LOCK(vp); 411 (void) vnode_pager_uncache(vp); 412 VOP_UNLOCK(vp); 413 } 414 } 415 } 416 417 /* 418 * Remove vnode associated object from the object cache. 419 * 420 * XXX unlock the vnode if it is currently locked. 421 * We must do this since uncaching the object may result in its 422 * destruction which may initiate paging activity which may necessitate 423 * re-locking the vnode. 424 */ 425 boolean_t 426 vnode_pager_uncache(vp) 427 register struct vnode *vp; 428 { 429 register vm_object_t object; 430 boolean_t uncached; 431 vm_pager_t pager; 432 433 /* 434 * Not a mapped vnode 435 */ 436 pager = (vm_pager_t)vp->v_vmdata; 437 if (pager == NULL) 438 return (TRUE); 439 #ifdef DEBUG 440 if (!VOP_ISLOCKED(vp)) { 441 extern int (**nfsv2_vnodeop_p)(); 442 443 if (vp->v_op != nfsv2_vnodeop_p) 444 panic("vnode_pager_uncache: vnode not locked!"); 445 } 446 #endif 447 /* 448 * Must use vm_object_lookup() as it actually removes 449 * the object from the cache list. 450 */ 451 object = vm_object_lookup(pager); 452 if (object) { 453 uncached = (object->ref_count <= 1); 454 VOP_UNLOCK(vp); 455 pager_cache(object, FALSE); 456 VOP_LOCK(vp); 457 } else 458 uncached = TRUE; 459 return(uncached); 460 } 461 462 static int 463 vnode_pager_io(vnp, mlist, npages, sync, rw) 464 register vn_pager_t vnp; 465 vm_page_t *mlist; 466 int npages; 467 boolean_t sync; 468 enum uio_rw rw; 469 { 470 struct uio auio; 471 struct iovec aiov; 472 vm_offset_t kva, foff; 473 int error, size; 474 struct proc *p = curproc; /* XXX */ 475 476 /* XXX */ 477 vm_page_t m; 478 if (npages != 1) 479 panic("vnode_pager_io: cannot handle multiple pages"); 480 m = *mlist; 481 /* XXX */ 482 483 #ifdef DEBUG 484 if (vpagerdebug & VDB_FOLLOW) 485 printf("vnode_pager_io(%x, %x, %c): vnode %x\n", 486 vnp, m, rw == UIO_READ ? 'R' : 'W', vnp->vnp_vp); 487 #endif 488 foff = m->offset + m->object->paging_offset; 489 /* 490 * Allocate a kernel virtual address and initialize so that 491 * we can use VOP_READ/WRITE routines. 492 */ 493 kva = vm_pager_map_pages(mlist, npages, sync); 494 if (kva == NULL) 495 return(VM_PAGER_AGAIN); 496 /* 497 * After all of the potentially blocking operations have been 498 * performed, we can do the size checks: 499 * read beyond EOF (returns error) 500 * short read 501 */ 502 VOP_LOCK(vnp->vnp_vp); 503 if (foff >= vnp->vnp_size) { 504 VOP_UNLOCK(vnp->vnp_vp); 505 vm_pager_unmap_pages(kva, npages); 506 #ifdef DEBUG 507 if (vpagerdebug & VDB_SIZE) 508 printf("vnode_pager_io: vp %x, off %d size %d\n", 509 vnp->vnp_vp, foff, vnp->vnp_size); 510 #endif 511 return(VM_PAGER_BAD); 512 } 513 if (foff + PAGE_SIZE > vnp->vnp_size) 514 size = vnp->vnp_size - foff; 515 else 516 size = PAGE_SIZE; 517 aiov.iov_base = (caddr_t)kva; 518 aiov.iov_len = size; 519 auio.uio_iov = &aiov; 520 auio.uio_iovcnt = 1; 521 auio.uio_offset = foff; 522 auio.uio_segflg = UIO_SYSSPACE; 523 auio.uio_rw = rw; 524 auio.uio_resid = size; 525 auio.uio_procp = (struct proc *)0; 526 #ifdef DEBUG 527 if (vpagerdebug & VDB_IO) 528 printf("vnode_pager_io: vp %x kva %x foff %x size %x", 529 vnp->vnp_vp, kva, foff, size); 530 #endif 531 if (rw == UIO_READ) 532 error = VOP_READ(vnp->vnp_vp, &auio, 0, p->p_ucred); 533 else 534 error = VOP_WRITE(vnp->vnp_vp, &auio, 0, p->p_ucred); 535 VOP_UNLOCK(vnp->vnp_vp); 536 #ifdef DEBUG 537 if (vpagerdebug & VDB_IO) { 538 if (error || auio.uio_resid) 539 printf(" returns error %x, resid %x", 540 error, auio.uio_resid); 541 printf("\n"); 542 } 543 #endif 544 if (!error) { 545 register int count = size - auio.uio_resid; 546 547 if (count == 0) 548 error = EINVAL; 549 else if (count != PAGE_SIZE && rw == UIO_READ) 550 bzero((void *)(kva + count), PAGE_SIZE - count); 551 } 552 vm_pager_unmap_pages(kva, npages); 553 return (error ? VM_PAGER_ERROR : VM_PAGER_OK); 554 } 555