1 /* 2 * Copyright (c) 1992, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * John Heidemann of the UCLA Ficus project. 7 * 8 * %sccs.include.redist.c% 9 * 10 * @(#)null_vnops.c 8.5 (Berkeley) 05/22/95 11 * 12 * Ancestors: 13 * @(#)lofs_vnops.c 1.2 (Berkeley) 6/18/92 14 * $Id: lofs_vnops.c,v 1.11 1992/05/30 10:05:43 jsp Exp jsp $ 15 * ...and... 16 * @(#)null_vnodeops.c 1.20 92/07/07 UCLA Ficus project 17 */ 18 19 /* 20 * Null Layer 21 * 22 * (See mount_null(8) for more information.) 23 * 24 * The null layer duplicates a portion of the file system 25 * name space under a new name. In this respect, it is 26 * similar to the loopback file system. It differs from 27 * the loopback fs in two respects: it is implemented using 28 * a stackable layers techniques, and it's "null-node"s stack above 29 * all lower-layer vnodes, not just over directory vnodes. 30 * 31 * The null layer has two purposes. First, it serves as a demonstration 32 * of layering by proving a layer which does nothing. (It actually 33 * does everything the loopback file system does, which is slightly 34 * more than nothing.) Second, the null layer can serve as a prototype 35 * layer. Since it provides all necessary layer framework, 36 * new file system layers can be created very easily be starting 37 * with a null layer. 38 * 39 * The remainder of this man page examines the null layer as a basis 40 * for constructing new layers. 41 * 42 * 43 * INSTANTIATING NEW NULL LAYERS 44 * 45 * New null layers are created with mount_null(8). 46 * Mount_null(8) takes two arguments, the pathname 47 * of the lower vfs (target-pn) and the pathname where the null 48 * layer will appear in the namespace (alias-pn). After 49 * the null layer is put into place, the contents 50 * of target-pn subtree will be aliased under alias-pn. 51 * 52 * 53 * OPERATION OF A NULL LAYER 54 * 55 * The null layer is the minimum file system layer, 56 * simply bypassing all possible operations to the lower layer 57 * for processing there. The majority of its activity centers 58 * on the bypass routine, though which nearly all vnode operations 59 * pass. 60 * 61 * The bypass routine accepts arbitrary vnode operations for 62 * handling by the lower layer. It begins by examing vnode 63 * operation arguments and replacing any null-nodes by their 64 * lower-layer equivlants. It then invokes the operation 65 * on the lower layer. Finally, it replaces the null-nodes 66 * in the arguments and, if a vnode is return by the operation, 67 * stacks a null-node on top of the returned vnode. 68 * 69 * Although bypass handles most operations, vop_getattr, vop_lock, 70 * vop_unlock, vop_inactive, vop_reclaim, and vop_print are not 71 * bypassed. Vop_getattr must change the fsid being returned. 72 * Vop_lock and vop_unlock must handle any locking for the 73 * current vnode as well as pass the lock request down. 74 * Vop_inactive and vop_reclaim are not bypassed so that 75 * they can handle freeing null-layer specific data. Vop_print 76 * is not bypassed to avoid excessive debugging information. 77 * Also, certain vnode operations change the locking state within 78 * the operation (create, mknod, remove, link, rename, mkdir, rmdir, 79 * and symlink). Ideally these operations should not change the 80 * lock state, but should be changed to let the caller of the 81 * function unlock them. Otherwise all intermediate vnode layers 82 * (such as union, umapfs, etc) must catch these functions to do 83 * the necessary locking at their layer. 84 * 85 * 86 * INSTANTIATING VNODE STACKS 87 * 88 * Mounting associates the null layer with a lower layer, 89 * effect stacking two VFSes. Vnode stacks are instead 90 * created on demand as files are accessed. 91 * 92 * The initial mount creates a single vnode stack for the 93 * root of the new null layer. All other vnode stacks 94 * are created as a result of vnode operations on 95 * this or other null vnode stacks. 96 * 97 * New vnode stacks come into existance as a result of 98 * an operation which returns a vnode. 99 * The bypass routine stacks a null-node above the new 100 * vnode before returning it to the caller. 101 * 102 * For example, imagine mounting a null layer with 103 * "mount_null /usr/include /dev/layer/null". 104 * Changing directory to /dev/layer/null will assign 105 * the root null-node (which was created when the null layer was mounted). 106 * Now consider opening "sys". A vop_lookup would be 107 * done on the root null-node. This operation would bypass through 108 * to the lower layer which would return a vnode representing 109 * the UFS "sys". Null_bypass then builds a null-node 110 * aliasing the UFS "sys" and returns this to the caller. 111 * Later operations on the null-node "sys" will repeat this 112 * process when constructing other vnode stacks. 113 * 114 * 115 * CREATING OTHER FILE SYSTEM LAYERS 116 * 117 * One of the easiest ways to construct new file system layers is to make 118 * a copy of the null layer, rename all files and variables, and 119 * then begin modifing the copy. Sed can be used to easily rename 120 * all variables. 121 * 122 * The umap layer is an example of a layer descended from the 123 * null layer. 124 * 125 * 126 * INVOKING OPERATIONS ON LOWER LAYERS 127 * 128 * There are two techniques to invoke operations on a lower layer 129 * when the operation cannot be completely bypassed. Each method 130 * is appropriate in different situations. In both cases, 131 * it is the responsibility of the aliasing layer to make 132 * the operation arguments "correct" for the lower layer 133 * by mapping an vnode arguments to the lower layer. 134 * 135 * The first approach is to call the aliasing layer's bypass routine. 136 * This method is most suitable when you wish to invoke the operation 137 * currently being hanldled on the lower layer. It has the advantage 138 * that the bypass routine already must do argument mapping. 139 * An example of this is null_getattrs in the null layer. 140 * 141 * A second approach is to directly invoked vnode operations on 142 * the lower layer with the VOP_OPERATIONNAME interface. 143 * The advantage of this method is that it is easy to invoke 144 * arbitrary operations on the lower layer. The disadvantage 145 * is that vnodes arguments must be manualy mapped. 146 * 147 */ 148 149 #include <sys/param.h> 150 #include <sys/systm.h> 151 #include <sys/proc.h> 152 #include <sys/time.h> 153 #include <sys/types.h> 154 #include <sys/vnode.h> 155 #include <sys/mount.h> 156 #include <sys/namei.h> 157 #include <sys/malloc.h> 158 #include <sys/buf.h> 159 #include <miscfs/nullfs/null.h> 160 161 162 int null_bug_bypass = 0; /* for debugging: enables bypass printf'ing */ 163 164 /* 165 * This is the 10-Apr-92 bypass routine. 166 * This version has been optimized for speed, throwing away some 167 * safety checks. It should still always work, but it's not as 168 * robust to programmer errors. 169 * Define SAFETY to include some error checking code. 170 * 171 * In general, we map all vnodes going down and unmap them on the way back. 172 * As an exception to this, vnodes can be marked "unmapped" by setting 173 * the Nth bit in operation's vdesc_flags. 174 * 175 * Also, some BSD vnode operations have the side effect of vrele'ing 176 * their arguments. With stacking, the reference counts are held 177 * by the upper node, not the lower one, so we must handle these 178 * side-effects here. This is not of concern in Sun-derived systems 179 * since there are no such side-effects. 180 * 181 * This makes the following assumptions: 182 * - only one returned vpp 183 * - no INOUT vpp's (Sun's vop_open has one of these) 184 * - the vnode operation vector of the first vnode should be used 185 * to determine what implementation of the op should be invoked 186 * - all mapped vnodes are of our vnode-type (NEEDSWORK: 187 * problems on rmdir'ing mount points and renaming?) 188 */ 189 int 190 null_bypass(ap) 191 struct vop_generic_args /* { 192 struct vnodeop_desc *a_desc; 193 <other random data follows, presumably> 194 } */ *ap; 195 { 196 extern int (**null_vnodeop_p)(); /* not extern, really "forward" */ 197 register struct vnode **this_vp_p; 198 int error; 199 struct vnode *old_vps[VDESC_MAX_VPS]; 200 struct vnode **vps_p[VDESC_MAX_VPS]; 201 struct vnode ***vppp; 202 struct vnodeop_desc *descp = ap->a_desc; 203 int reles, i; 204 205 if (null_bug_bypass) 206 printf ("null_bypass: %s\n", descp->vdesc_name); 207 208 #ifdef SAFETY 209 /* 210 * We require at least one vp. 211 */ 212 if (descp->vdesc_vp_offsets == NULL || 213 descp->vdesc_vp_offsets[0] == VDESC_NO_OFFSET) 214 panic ("null_bypass: no vp's in map.\n"); 215 #endif 216 217 /* 218 * Map the vnodes going in. 219 * Later, we'll invoke the operation based on 220 * the first mapped vnode's operation vector. 221 */ 222 reles = descp->vdesc_flags; 223 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) { 224 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET) 225 break; /* bail out at end of list */ 226 vps_p[i] = this_vp_p = 227 VOPARG_OFFSETTO(struct vnode**,descp->vdesc_vp_offsets[i],ap); 228 /* 229 * We're not guaranteed that any but the first vnode 230 * are of our type. Check for and don't map any 231 * that aren't. (We must always map first vp or vclean fails.) 232 */ 233 if (i && (*this_vp_p == NULL || 234 (*this_vp_p)->v_op != null_vnodeop_p)) { 235 old_vps[i] = NULL; 236 } else { 237 old_vps[i] = *this_vp_p; 238 *(vps_p[i]) = NULLVPTOLOWERVP(*this_vp_p); 239 /* 240 * XXX - Several operations have the side effect 241 * of vrele'ing their vp's. We must account for 242 * that. (This should go away in the future.) 243 */ 244 if (reles & 1) 245 VREF(*this_vp_p); 246 } 247 248 } 249 250 /* 251 * Call the operation on the lower layer 252 * with the modified argument structure. 253 */ 254 error = VCALL(*(vps_p[0]), descp->vdesc_offset, ap); 255 256 /* 257 * Maintain the illusion of call-by-value 258 * by restoring vnodes in the argument structure 259 * to their original value. 260 */ 261 reles = descp->vdesc_flags; 262 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) { 263 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET) 264 break; /* bail out at end of list */ 265 if (old_vps[i]) { 266 *(vps_p[i]) = old_vps[i]; 267 if (reles & 1) 268 vrele(*(vps_p[i])); 269 } 270 } 271 272 /* 273 * Map the possible out-going vpp 274 * (Assumes that the lower layer always returns 275 * a VREF'ed vpp unless it gets an error.) 276 */ 277 if (descp->vdesc_vpp_offset != VDESC_NO_OFFSET && 278 !(descp->vdesc_flags & VDESC_NOMAP_VPP) && 279 !error) { 280 /* 281 * XXX - even though some ops have vpp returned vp's, 282 * several ops actually vrele this before returning. 283 * We must avoid these ops. 284 * (This should go away when these ops are regularized.) 285 */ 286 if (descp->vdesc_flags & VDESC_VPP_WILLRELE) 287 goto out; 288 vppp = VOPARG_OFFSETTO(struct vnode***, 289 descp->vdesc_vpp_offset,ap); 290 error = null_node_create(old_vps[0]->v_mount, **vppp, *vppp); 291 } 292 293 out: 294 return (error); 295 } 296 297 /* 298 * We have to carry on the locking protocol on the null layer vnodes 299 * as we progress through the tree. 300 */ 301 null_lookup(ap) 302 struct vop_lookup_args /* { 303 struct vnode * a_dvp; 304 struct vnode ** a_vpp; 305 struct componentname * a_cnp; 306 } */ *ap; 307 { 308 struct proc *p = ap->a_cnp->cn_proc; 309 struct vop_lock_args lockargs; 310 struct vop_unlock_args unlockargs; 311 struct vnode *dvp, *vp; 312 int error; 313 314 error = null_bypass(ap); 315 /* 316 * We must do the same locking and unlocking at this layer as 317 * is done in the layers below us. We could figure this out 318 * based on the error return and the LASTCN, LOCKPARENT, and 319 * LOCKLEAF flags. However, it is more expidient to just find 320 * out the state of the lower level vnodes and set ours to the 321 * same state. 322 */ 323 dvp = ap->a_dvp; 324 vp = *ap->a_vpp; 325 if (dvp == vp) 326 return (error); 327 if (!VOP_ISLOCKED(dvp)) { 328 unlockargs.a_vp = dvp; 329 unlockargs.a_flags = 0; 330 unlockargs.a_p = p; 331 vop_nounlock(&unlockargs); 332 } 333 if (vp != NULL && VOP_ISLOCKED(vp)) { 334 lockargs.a_vp = vp; 335 lockargs.a_flags = LK_SHARED; 336 lockargs.a_p = p; 337 vop_nolock(&lockargs); 338 } 339 return (error); 340 } 341 342 /* 343 * We handle getattr only to change the fsid. 344 */ 345 int 346 null_getattr(ap) 347 struct vop_getattr_args /* { 348 struct vnode *a_vp; 349 struct vattr *a_vap; 350 struct ucred *a_cred; 351 struct proc *a_p; 352 } */ *ap; 353 { 354 int error; 355 356 if (error = null_bypass(ap)) 357 return (error); 358 /* Requires that arguments be restored. */ 359 ap->a_vap->va_fsid = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0]; 360 return (0); 361 } 362 363 /* 364 * We need to process our own vnode lock and then clear the 365 * interlock flag as it applies only to our vnode, not the 366 * vnodes below us on the stack. 367 */ 368 int 369 null_lock(ap) 370 struct vop_lock_args /* { 371 struct vnode *a_vp; 372 int a_flags; 373 struct proc *a_p; 374 } */ *ap; 375 { 376 377 vop_nolock(ap); 378 if ((ap->a_flags & LK_TYPE_MASK) == LK_DRAIN) 379 return (0); 380 ap->a_flags &= ~LK_INTERLOCK; 381 return (null_bypass(ap)); 382 } 383 384 /* 385 * We need to process our own vnode unlock and then clear the 386 * interlock flag as it applies only to our vnode, not the 387 * vnodes below us on the stack. 388 */ 389 int 390 null_unlock(ap) 391 struct vop_unlock_args /* { 392 struct vnode *a_vp; 393 int a_flags; 394 struct proc *a_p; 395 } */ *ap; 396 { 397 struct vnode *vp = ap->a_vp; 398 399 vop_nounlock(ap); 400 ap->a_flags &= ~LK_INTERLOCK; 401 return (null_bypass(ap)); 402 } 403 404 int 405 null_inactive(ap) 406 struct vop_inactive_args /* { 407 struct vnode *a_vp; 408 struct proc *a_p; 409 } */ *ap; 410 { 411 /* 412 * Do nothing (and _don't_ bypass). 413 * Wait to vrele lowervp until reclaim, 414 * so that until then our null_node is in the 415 * cache and reusable. 416 * 417 * NEEDSWORK: Someday, consider inactive'ing 418 * the lowervp and then trying to reactivate it 419 * with capabilities (v_id) 420 * like they do in the name lookup cache code. 421 * That's too much work for now. 422 */ 423 VOP_UNLOCK(ap->a_vp, 0, ap->a_p); 424 return (0); 425 } 426 427 int 428 null_reclaim(ap) 429 struct vop_reclaim_args /* { 430 struct vnode *a_vp; 431 struct proc *a_p; 432 } */ *ap; 433 { 434 struct vnode *vp = ap->a_vp; 435 struct null_node *xp = VTONULL(vp); 436 struct vnode *lowervp = xp->null_lowervp; 437 438 /* 439 * Note: in vop_reclaim, vp->v_op == dead_vnodeop_p, 440 * so we can't call VOPs on ourself. 441 */ 442 /* After this assignment, this node will not be re-used. */ 443 xp->null_lowervp = NULL; 444 LIST_REMOVE(xp, null_hash); 445 FREE(vp->v_data, M_TEMP); 446 vp->v_data = NULL; 447 vrele (lowervp); 448 return (0); 449 } 450 451 int 452 null_print(ap) 453 struct vop_print_args /* { 454 struct vnode *a_vp; 455 } */ *ap; 456 { 457 register struct vnode *vp = ap->a_vp; 458 printf ("\ttag VT_NULLFS, vp=%x, lowervp=%x\n", vp, NULLVPTOLOWERVP(vp)); 459 return (0); 460 } 461 462 /* 463 * XXX - vop_strategy must be hand coded because it has no 464 * vnode in its arguments. 465 * This goes away with a merged VM/buffer cache. 466 */ 467 int 468 null_strategy(ap) 469 struct vop_strategy_args /* { 470 struct buf *a_bp; 471 } */ *ap; 472 { 473 struct buf *bp = ap->a_bp; 474 int error; 475 struct vnode *savedvp; 476 477 savedvp = bp->b_vp; 478 bp->b_vp = NULLVPTOLOWERVP(bp->b_vp); 479 480 error = VOP_STRATEGY(bp); 481 482 bp->b_vp = savedvp; 483 484 return (error); 485 } 486 487 /* 488 * XXX - like vop_strategy, vop_bwrite must be hand coded because it has no 489 * vnode in its arguments. 490 * This goes away with a merged VM/buffer cache. 491 */ 492 int 493 null_bwrite(ap) 494 struct vop_bwrite_args /* { 495 struct buf *a_bp; 496 } */ *ap; 497 { 498 struct buf *bp = ap->a_bp; 499 int error; 500 struct vnode *savedvp; 501 502 savedvp = bp->b_vp; 503 bp->b_vp = NULLVPTOLOWERVP(bp->b_vp); 504 505 error = VOP_BWRITE(bp); 506 507 bp->b_vp = savedvp; 508 509 return (error); 510 } 511 512 /* 513 * Global vfs data structures 514 */ 515 int (**null_vnodeop_p)(); 516 struct vnodeopv_entry_desc null_vnodeop_entries[] = { 517 { &vop_default_desc, null_bypass }, 518 519 { &vop_lookup_desc, null_lookup }, 520 { &vop_getattr_desc, null_getattr }, 521 { &vop_lock_desc, null_lock }, 522 { &vop_unlock_desc, null_unlock }, 523 { &vop_inactive_desc, null_inactive }, 524 { &vop_reclaim_desc, null_reclaim }, 525 { &vop_print_desc, null_print }, 526 527 { &vop_strategy_desc, null_strategy }, 528 { &vop_bwrite_desc, null_bwrite }, 529 530 { (struct vnodeop_desc*)NULL, (int(*)())NULL } 531 }; 532 struct vnodeopv_desc null_vnodeop_opv_desc = 533 { &null_vnodeop_p, null_vnodeop_entries }; 534