null_vnops.c revision 116469
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 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the University of 19 * California, Berkeley and its contributors. 20 * 4. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * @(#)null_vnops.c 8.6 (Berkeley) 5/27/95 37 * 38 * Ancestors: 39 * @(#)lofs_vnops.c 1.2 (Berkeley) 6/18/92 40 * ...and... 41 * @(#)null_vnodeops.c 1.20 92/07/07 UCLA Ficus project 42 * 43 * $FreeBSD: head/sys/fs/nullfs/null_vnops.c 116469 2003-06-17 08:52:45Z tjr $ 44 */ 45 46/* 47 * Null Layer 48 * 49 * (See mount_nullfs(8) for more information.) 50 * 51 * The null layer duplicates a portion of the filesystem 52 * name space under a new name. In this respect, it is 53 * similar to the loopback filesystem. It differs from 54 * the loopback fs in two respects: it is implemented using 55 * a stackable layers techniques, and its "null-node"s stack above 56 * all lower-layer vnodes, not just over directory vnodes. 57 * 58 * The null layer has two purposes. First, it serves as a demonstration 59 * of layering by proving a layer which does nothing. (It actually 60 * does everything the loopback filesystem does, which is slightly 61 * more than nothing.) Second, the null layer can serve as a prototype 62 * layer. Since it provides all necessary layer framework, 63 * new filesystem layers can be created very easily be starting 64 * with a null layer. 65 * 66 * The remainder of this man page examines the null layer as a basis 67 * for constructing new layers. 68 * 69 * 70 * INSTANTIATING NEW NULL LAYERS 71 * 72 * New null layers are created with mount_nullfs(8). 73 * Mount_nullfs(8) takes two arguments, the pathname 74 * of the lower vfs (target-pn) and the pathname where the null 75 * layer will appear in the namespace (alias-pn). After 76 * the null layer is put into place, the contents 77 * of target-pn subtree will be aliased under alias-pn. 78 * 79 * 80 * OPERATION OF A NULL LAYER 81 * 82 * The null layer is the minimum filesystem layer, 83 * simply bypassing all possible operations to the lower layer 84 * for processing there. The majority of its activity centers 85 * on the bypass routine, through which nearly all vnode operations 86 * pass. 87 * 88 * The bypass routine accepts arbitrary vnode operations for 89 * handling by the lower layer. It begins by examing vnode 90 * operation arguments and replacing any null-nodes by their 91 * lower-layer equivlants. It then invokes the operation 92 * on the lower layer. Finally, it replaces the null-nodes 93 * in the arguments and, if a vnode is return by the operation, 94 * stacks a null-node on top of the returned vnode. 95 * 96 * Although bypass handles most operations, vop_getattr, vop_lock, 97 * vop_unlock, vop_inactive, vop_reclaim, and vop_print are not 98 * bypassed. Vop_getattr must change the fsid being returned. 99 * Vop_lock and vop_unlock must handle any locking for the 100 * current vnode as well as pass the lock request down. 101 * Vop_inactive and vop_reclaim are not bypassed so that 102 * they can handle freeing null-layer specific data. Vop_print 103 * is not bypassed to avoid excessive debugging information. 104 * Also, certain vnode operations change the locking state within 105 * the operation (create, mknod, remove, link, rename, mkdir, rmdir, 106 * and symlink). Ideally these operations should not change the 107 * lock state, but should be changed to let the caller of the 108 * function unlock them. Otherwise all intermediate vnode layers 109 * (such as union, umapfs, etc) must catch these functions to do 110 * the necessary locking at their layer. 111 * 112 * 113 * INSTANTIATING VNODE STACKS 114 * 115 * Mounting associates the null layer with a lower layer, 116 * effect stacking two VFSes. Vnode stacks are instead 117 * created on demand as files are accessed. 118 * 119 * The initial mount creates a single vnode stack for the 120 * root of the new null layer. All other vnode stacks 121 * are created as a result of vnode operations on 122 * this or other null vnode stacks. 123 * 124 * New vnode stacks come into existance as a result of 125 * an operation which returns a vnode. 126 * The bypass routine stacks a null-node above the new 127 * vnode before returning it to the caller. 128 * 129 * For example, imagine mounting a null layer with 130 * "mount_nullfs /usr/include /dev/layer/null". 131 * Changing directory to /dev/layer/null will assign 132 * the root null-node (which was created when the null layer was mounted). 133 * Now consider opening "sys". A vop_lookup would be 134 * done on the root null-node. This operation would bypass through 135 * to the lower layer which would return a vnode representing 136 * the UFS "sys". Null_bypass then builds a null-node 137 * aliasing the UFS "sys" and returns this to the caller. 138 * Later operations on the null-node "sys" will repeat this 139 * process when constructing other vnode stacks. 140 * 141 * 142 * CREATING OTHER FILE SYSTEM LAYERS 143 * 144 * One of the easiest ways to construct new filesystem layers is to make 145 * a copy of the null layer, rename all files and variables, and 146 * then begin modifing the copy. Sed can be used to easily rename 147 * all variables. 148 * 149 * The umap layer is an example of a layer descended from the 150 * null layer. 151 * 152 * 153 * INVOKING OPERATIONS ON LOWER LAYERS 154 * 155 * There are two techniques to invoke operations on a lower layer 156 * when the operation cannot be completely bypassed. Each method 157 * is appropriate in different situations. In both cases, 158 * it is the responsibility of the aliasing layer to make 159 * the operation arguments "correct" for the lower layer 160 * by mapping a vnode arguments to the lower layer. 161 * 162 * The first approach is to call the aliasing layer's bypass routine. 163 * This method is most suitable when you wish to invoke the operation 164 * currently being handled on the lower layer. It has the advantage 165 * that the bypass routine already must do argument mapping. 166 * An example of this is null_getattrs in the null layer. 167 * 168 * A second approach is to directly invoke vnode operations on 169 * the lower layer with the VOP_OPERATIONNAME interface. 170 * The advantage of this method is that it is easy to invoke 171 * arbitrary operations on the lower layer. The disadvantage 172 * is that vnode arguments must be manualy mapped. 173 * 174 */ 175 176#include <sys/param.h> 177#include <sys/systm.h> 178#include <sys/conf.h> 179#include <sys/kernel.h> 180#include <sys/lock.h> 181#include <sys/malloc.h> 182#include <sys/mount.h> 183#include <sys/mutex.h> 184#include <sys/namei.h> 185#include <sys/sysctl.h> 186#include <sys/vnode.h> 187 188#include <fs/nullfs/null.h> 189 190#include <vm/vm.h> 191#include <vm/vm_extern.h> 192#include <vm/vm_object.h> 193#include <vm/vnode_pager.h> 194 195static int null_bug_bypass = 0; /* for debugging: enables bypass printf'ing */ 196SYSCTL_INT(_debug, OID_AUTO, nullfs_bug_bypass, CTLFLAG_RW, 197 &null_bug_bypass, 0, ""); 198 199static int null_access(struct vop_access_args *ap); 200static int null_createvobject(struct vop_createvobject_args *ap); 201static int null_destroyvobject(struct vop_destroyvobject_args *ap); 202static int null_getattr(struct vop_getattr_args *ap); 203static int null_getvobject(struct vop_getvobject_args *ap); 204static int null_inactive(struct vop_inactive_args *ap); 205static int null_islocked(struct vop_islocked_args *ap); 206static int null_lock(struct vop_lock_args *ap); 207static int null_lookup(struct vop_lookup_args *ap); 208static int null_open(struct vop_open_args *ap); 209static int null_print(struct vop_print_args *ap); 210static int null_reclaim(struct vop_reclaim_args *ap); 211static int null_rename(struct vop_rename_args *ap); 212static int null_setattr(struct vop_setattr_args *ap); 213static int null_unlock(struct vop_unlock_args *ap); 214 215/* 216 * This is the 10-Apr-92 bypass routine. 217 * This version has been optimized for speed, throwing away some 218 * safety checks. It should still always work, but it's not as 219 * robust to programmer errors. 220 * 221 * In general, we map all vnodes going down and unmap them on the way back. 222 * As an exception to this, vnodes can be marked "unmapped" by setting 223 * the Nth bit in operation's vdesc_flags. 224 * 225 * Also, some BSD vnode operations have the side effect of vrele'ing 226 * their arguments. With stacking, the reference counts are held 227 * by the upper node, not the lower one, so we must handle these 228 * side-effects here. This is not of concern in Sun-derived systems 229 * since there are no such side-effects. 230 * 231 * This makes the following assumptions: 232 * - only one returned vpp 233 * - no INOUT vpp's (Sun's vop_open has one of these) 234 * - the vnode operation vector of the first vnode should be used 235 * to determine what implementation of the op should be invoked 236 * - all mapped vnodes are of our vnode-type (NEEDSWORK: 237 * problems on rmdir'ing mount points and renaming?) 238 */ 239int 240null_bypass(ap) 241 struct vop_generic_args /* { 242 struct vnodeop_desc *a_desc; 243 <other random data follows, presumably> 244 } */ *ap; 245{ 246 register struct vnode **this_vp_p; 247 int error; 248 struct vnode *old_vps[VDESC_MAX_VPS]; 249 struct vnode **vps_p[VDESC_MAX_VPS]; 250 struct vnode ***vppp; 251 struct vnodeop_desc *descp = ap->a_desc; 252 int reles, i; 253 254 if (null_bug_bypass) 255 printf ("null_bypass: %s\n", descp->vdesc_name); 256 257#ifdef DIAGNOSTIC 258 /* 259 * We require at least one vp. 260 */ 261 if (descp->vdesc_vp_offsets == NULL || 262 descp->vdesc_vp_offsets[0] == VDESC_NO_OFFSET) 263 panic ("null_bypass: no vp's in map"); 264#endif 265 266 /* 267 * Map the vnodes going in. 268 * Later, we'll invoke the operation based on 269 * the first mapped vnode's operation vector. 270 */ 271 reles = descp->vdesc_flags; 272 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) { 273 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET) 274 break; /* bail out at end of list */ 275 vps_p[i] = this_vp_p = 276 VOPARG_OFFSETTO(struct vnode**,descp->vdesc_vp_offsets[i],ap); 277 /* 278 * We're not guaranteed that any but the first vnode 279 * are of our type. Check for and don't map any 280 * that aren't. (We must always map first vp or vclean fails.) 281 */ 282 if (i && (*this_vp_p == NULLVP || 283 (*this_vp_p)->v_op != null_vnodeop_p)) { 284 old_vps[i] = NULLVP; 285 } else { 286 old_vps[i] = *this_vp_p; 287 *(vps_p[i]) = NULLVPTOLOWERVP(*this_vp_p); 288 /* 289 * XXX - Several operations have the side effect 290 * of vrele'ing their vp's. We must account for 291 * that. (This should go away in the future.) 292 */ 293 if (reles & VDESC_VP0_WILLRELE) 294 VREF(*this_vp_p); 295 } 296 297 } 298 299 /* 300 * Call the operation on the lower layer 301 * with the modified argument structure. 302 */ 303 if (vps_p[0] && *vps_p[0]) 304 error = VCALL(*(vps_p[0]), descp->vdesc_offset, ap); 305 else { 306 printf("null_bypass: no map for %s\n", descp->vdesc_name); 307 error = EINVAL; 308 } 309 310 /* 311 * Maintain the illusion of call-by-value 312 * by restoring vnodes in the argument structure 313 * to their original value. 314 */ 315 reles = descp->vdesc_flags; 316 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) { 317 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET) 318 break; /* bail out at end of list */ 319 if (old_vps[i]) { 320 *(vps_p[i]) = old_vps[i]; 321#if 0 322 if (reles & VDESC_VP0_WILLUNLOCK) 323 VOP_UNLOCK(*(vps_p[i]), LK_THISLAYER, curthread); 324#endif 325 if (reles & VDESC_VP0_WILLRELE) 326 vrele(*(vps_p[i])); 327 } 328 } 329 330 /* 331 * Map the possible out-going vpp 332 * (Assumes that the lower layer always returns 333 * a VREF'ed vpp unless it gets an error.) 334 */ 335 if (descp->vdesc_vpp_offset != VDESC_NO_OFFSET && 336 !(descp->vdesc_flags & VDESC_NOMAP_VPP) && 337 !error) { 338 /* 339 * XXX - even though some ops have vpp returned vp's, 340 * several ops actually vrele this before returning. 341 * We must avoid these ops. 342 * (This should go away when these ops are regularized.) 343 */ 344 if (descp->vdesc_flags & VDESC_VPP_WILLRELE) 345 goto out; 346 vppp = VOPARG_OFFSETTO(struct vnode***, 347 descp->vdesc_vpp_offset,ap); 348 if (*vppp) 349 error = null_nodeget(old_vps[0]->v_mount, **vppp, *vppp); 350 } 351 352 out: 353 return (error); 354} 355 356/* 357 * We have to carry on the locking protocol on the null layer vnodes 358 * as we progress through the tree. We also have to enforce read-only 359 * if this layer is mounted read-only. 360 */ 361static int 362null_lookup(ap) 363 struct vop_lookup_args /* { 364 struct vnode * a_dvp; 365 struct vnode ** a_vpp; 366 struct componentname * a_cnp; 367 } */ *ap; 368{ 369 struct componentname *cnp = ap->a_cnp; 370 struct vnode *dvp = ap->a_dvp; 371 struct thread *td = cnp->cn_thread; 372 int flags = cnp->cn_flags; 373 struct vnode *vp, *ldvp, *lvp; 374 int error; 375 376 if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) && 377 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) 378 return (EROFS); 379 /* 380 * Although it is possible to call null_bypass(), we'll do 381 * a direct call to reduce overhead 382 */ 383 ldvp = NULLVPTOLOWERVP(dvp); 384 vp = lvp = NULL; 385 error = VOP_LOOKUP(ldvp, &lvp, cnp); 386 if (error == EJUSTRETURN && (flags & ISLASTCN) && 387 (dvp->v_mount->mnt_flag & MNT_RDONLY) && 388 (cnp->cn_nameiop == CREATE || cnp->cn_nameiop == RENAME)) 389 error = EROFS; 390 391 /* 392 * Rely only on the PDIRUNLOCK flag which should be carefully 393 * tracked by underlying filesystem. 394 */ 395 if (cnp->cn_flags & PDIRUNLOCK) 396 VOP_UNLOCK(dvp, LK_THISLAYER, td); 397 if ((error == 0 || error == EJUSTRETURN) && lvp != NULL) { 398 if (ldvp == lvp) { 399 *ap->a_vpp = dvp; 400 VREF(dvp); 401 vrele(lvp); 402 } else { 403 error = null_nodeget(dvp->v_mount, lvp, &vp); 404 if (error) { 405 /* XXX Cleanup needed... */ 406 panic("null_nodeget failed"); 407 } 408 *ap->a_vpp = vp; 409 } 410 } 411 return (error); 412} 413 414/* 415 * Setattr call. Disallow write attempts if the layer is mounted read-only. 416 */ 417static int 418null_setattr(ap) 419 struct vop_setattr_args /* { 420 struct vnodeop_desc *a_desc; 421 struct vnode *a_vp; 422 struct vattr *a_vap; 423 struct ucred *a_cred; 424 struct thread *a_td; 425 } */ *ap; 426{ 427 struct vnode *vp = ap->a_vp; 428 struct vattr *vap = ap->a_vap; 429 430 if ((vap->va_flags != VNOVAL || vap->va_uid != (uid_t)VNOVAL || 431 vap->va_gid != (gid_t)VNOVAL || vap->va_atime.tv_sec != VNOVAL || 432 vap->va_mtime.tv_sec != VNOVAL || vap->va_mode != (mode_t)VNOVAL) && 433 (vp->v_mount->mnt_flag & MNT_RDONLY)) 434 return (EROFS); 435 if (vap->va_size != VNOVAL) { 436 switch (vp->v_type) { 437 case VDIR: 438 return (EISDIR); 439 case VCHR: 440 case VBLK: 441 case VSOCK: 442 case VFIFO: 443 if (vap->va_flags != VNOVAL) 444 return (EOPNOTSUPP); 445 return (0); 446 case VREG: 447 case VLNK: 448 default: 449 /* 450 * Disallow write attempts if the filesystem is 451 * mounted read-only. 452 */ 453 if (vp->v_mount->mnt_flag & MNT_RDONLY) 454 return (EROFS); 455 } 456 } 457 458 return (null_bypass((struct vop_generic_args *)ap)); 459} 460 461/* 462 * We handle getattr only to change the fsid. 463 */ 464static int 465null_getattr(ap) 466 struct vop_getattr_args /* { 467 struct vnode *a_vp; 468 struct vattr *a_vap; 469 struct ucred *a_cred; 470 struct thread *a_td; 471 } */ *ap; 472{ 473 int error; 474 475 if ((error = null_bypass((struct vop_generic_args *)ap)) != 0) 476 return (error); 477 478 ap->a_vap->va_fsid = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0]; 479 return (0); 480} 481 482/* 483 * Handle to disallow write access if mounted read-only. 484 */ 485static int 486null_access(ap) 487 struct vop_access_args /* { 488 struct vnode *a_vp; 489 int a_mode; 490 struct ucred *a_cred; 491 struct thread *a_td; 492 } */ *ap; 493{ 494 struct vnode *vp = ap->a_vp; 495 mode_t mode = ap->a_mode; 496 497 /* 498 * Disallow write attempts on read-only layers; 499 * unless the file is a socket, fifo, or a block or 500 * character device resident on the filesystem. 501 */ 502 if (mode & VWRITE) { 503 switch (vp->v_type) { 504 case VDIR: 505 case VLNK: 506 case VREG: 507 if (vp->v_mount->mnt_flag & MNT_RDONLY) 508 return (EROFS); 509 break; 510 default: 511 break; 512 } 513 } 514 return (null_bypass((struct vop_generic_args *)ap)); 515} 516 517/* 518 * We must handle open to be able to catch MNT_NODEV and friends. 519 */ 520static int 521null_open(ap) 522 struct vop_open_args /* { 523 struct vnode *a_vp; 524 int a_mode; 525 struct ucred *a_cred; 526 struct thread *a_td; 527 } */ *ap; 528{ 529 struct vnode *vp = ap->a_vp; 530 struct vnode *lvp = NULLVPTOLOWERVP(ap->a_vp); 531 532 if ((vp->v_mount->mnt_flag & MNT_NODEV) && 533 (lvp->v_type == VBLK || lvp->v_type == VCHR)) 534 return ENXIO; 535 536 return (null_bypass((struct vop_generic_args *)ap)); 537} 538 539/* 540 * We handle this to eliminate null FS to lower FS 541 * file moving. Don't know why we don't allow this, 542 * possibly we should. 543 */ 544static int 545null_rename(ap) 546 struct vop_rename_args /* { 547 struct vnode *a_fdvp; 548 struct vnode *a_fvp; 549 struct componentname *a_fcnp; 550 struct vnode *a_tdvp; 551 struct vnode *a_tvp; 552 struct componentname *a_tcnp; 553 } */ *ap; 554{ 555 struct vnode *tdvp = ap->a_tdvp; 556 struct vnode *fvp = ap->a_fvp; 557 struct vnode *fdvp = ap->a_fdvp; 558 struct vnode *tvp = ap->a_tvp; 559 560 /* Check for cross-device rename. */ 561 if ((fvp->v_mount != tdvp->v_mount) || 562 (tvp && (fvp->v_mount != tvp->v_mount))) { 563 if (tdvp == tvp) 564 vrele(tdvp); 565 else 566 vput(tdvp); 567 if (tvp) 568 vput(tvp); 569 vrele(fdvp); 570 vrele(fvp); 571 return (EXDEV); 572 } 573 574 return (null_bypass((struct vop_generic_args *)ap)); 575} 576 577/* 578 * We need to process our own vnode lock and then clear the 579 * interlock flag as it applies only to our vnode, not the 580 * vnodes below us on the stack. 581 */ 582static int 583null_lock(ap) 584 struct vop_lock_args /* { 585 struct vnode *a_vp; 586 int a_flags; 587 struct thread *a_td; 588 } */ *ap; 589{ 590 struct vnode *vp = ap->a_vp; 591 int flags = ap->a_flags; 592 struct thread *td = ap->a_td; 593 struct vnode *lvp; 594 int error; 595 struct null_node *nn; 596 597 if (flags & LK_THISLAYER) { 598 if (vp->v_vnlock != NULL) { 599 /* lock is shared across layers */ 600 if (flags & LK_INTERLOCK) 601 mtx_unlock(&vp->v_interlock); 602 return 0; 603 } 604 error = lockmgr(&vp->v_lock, flags & ~LK_THISLAYER, 605 &vp->v_interlock, td); 606 return (error); 607 } 608 609 if (vp->v_vnlock != NULL) { 610 /* 611 * The lower level has exported a struct lock to us. Use 612 * it so that all vnodes in the stack lock and unlock 613 * simultaneously. Note: we don't DRAIN the lock as DRAIN 614 * decommissions the lock - just because our vnode is 615 * going away doesn't mean the struct lock below us is. 616 * LK_EXCLUSIVE is fine. 617 */ 618 if ((flags & LK_INTERLOCK) == 0) { 619 VI_LOCK(vp); 620 flags |= LK_INTERLOCK; 621 } 622 nn = VTONULL(vp); 623 if ((flags & LK_TYPE_MASK) == LK_DRAIN) { 624 NULLFSDEBUG("null_lock: avoiding LK_DRAIN\n"); 625 /* 626 * Emulate lock draining by waiting for all other 627 * pending locks to complete. Afterwards the 628 * lockmgr call might block, but no other threads 629 * will attempt to use this nullfs vnode due to the 630 * VI_XLOCK flag. 631 */ 632 while (nn->null_pending_locks > 0) { 633 nn->null_drain_wakeup = 1; 634 msleep(&nn->null_pending_locks, 635 VI_MTX(vp), 636 PVFS, 637 "nuldr", 0); 638 } 639 error = lockmgr(vp->v_vnlock, 640 (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE, 641 VI_MTX(vp), td); 642 return error; 643 } 644 nn->null_pending_locks++; 645 error = lockmgr(vp->v_vnlock, flags, &vp->v_interlock, td); 646 VI_LOCK(vp); 647 /* 648 * If we're called from vrele then v_usecount can have been 0 649 * and another process might have initiated a recycle 650 * operation. When that happens, just back out. 651 */ 652 if (error == 0 && (vp->v_iflag & VI_XLOCK) != 0 && 653 td != vp->v_vxproc) { 654 lockmgr(vp->v_vnlock, 655 (flags & ~LK_TYPE_MASK) | LK_RELEASE, 656 VI_MTX(vp), td); 657 VI_LOCK(vp); 658 error = ENOENT; 659 } 660 nn->null_pending_locks--; 661 /* 662 * Wakeup the process draining the vnode after all 663 * pending lock attempts has been failed. 664 */ 665 if (nn->null_pending_locks == 0 && 666 nn->null_drain_wakeup != 0) { 667 nn->null_drain_wakeup = 0; 668 wakeup(&nn->null_pending_locks); 669 } 670 if (error == ENOENT && (vp->v_iflag & VI_XLOCK) != 0 && 671 vp->v_vxproc != curthread) { 672 vp->v_iflag |= VI_XWANT; 673 msleep(vp, VI_MTX(vp), PINOD, "nulbo", 0); 674 } 675 VI_UNLOCK(vp); 676 return error; 677 } else { 678 /* 679 * To prevent race conditions involving doing a lookup 680 * on "..", we have to lock the lower node, then lock our 681 * node. Most of the time it won't matter that we lock our 682 * node (as any locking would need the lower one locked 683 * first). But we can LK_DRAIN the upper lock as a step 684 * towards decomissioning it. 685 */ 686 lvp = NULLVPTOLOWERVP(vp); 687 if (lvp == NULL) 688 return (lockmgr(&vp->v_lock, flags, &vp->v_interlock, td)); 689 if (flags & LK_INTERLOCK) { 690 mtx_unlock(&vp->v_interlock); 691 flags &= ~LK_INTERLOCK; 692 } 693 if ((flags & LK_TYPE_MASK) == LK_DRAIN) { 694 error = VOP_LOCK(lvp, 695 (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE, td); 696 } else 697 error = VOP_LOCK(lvp, flags, td); 698 if (error) 699 return (error); 700 error = lockmgr(&vp->v_lock, flags, &vp->v_interlock, td); 701 if (error) 702 VOP_UNLOCK(lvp, 0, td); 703 return (error); 704 } 705} 706 707/* 708 * We need to process our own vnode unlock and then clear the 709 * interlock flag as it applies only to our vnode, not the 710 * vnodes below us on the stack. 711 */ 712static int 713null_unlock(ap) 714 struct vop_unlock_args /* { 715 struct vnode *a_vp; 716 int a_flags; 717 struct thread *a_td; 718 } */ *ap; 719{ 720 struct vnode *vp = ap->a_vp; 721 int flags = ap->a_flags; 722 struct thread *td = ap->a_td; 723 struct vnode *lvp; 724 725 if (vp->v_vnlock != NULL) { 726 if (flags & LK_THISLAYER) 727 return 0; /* the lock is shared across layers */ 728 flags &= ~LK_THISLAYER; 729 return (lockmgr(vp->v_vnlock, flags | LK_RELEASE, 730 &vp->v_interlock, td)); 731 } 732 lvp = NULLVPTOLOWERVP(vp); 733 if (lvp == NULL) 734 return (lockmgr(&vp->v_lock, flags | LK_RELEASE, &vp->v_interlock, td)); 735 if ((flags & LK_THISLAYER) == 0) { 736 if (flags & LK_INTERLOCK) { 737 mtx_unlock(&vp->v_interlock); 738 flags &= ~LK_INTERLOCK; 739 } 740 VOP_UNLOCK(lvp, flags & ~LK_INTERLOCK, td); 741 } else 742 flags &= ~LK_THISLAYER; 743 return (lockmgr(&vp->v_lock, flags | LK_RELEASE, &vp->v_interlock, td)); 744} 745 746static int 747null_islocked(ap) 748 struct vop_islocked_args /* { 749 struct vnode *a_vp; 750 struct thread *a_td; 751 } */ *ap; 752{ 753 struct vnode *vp = ap->a_vp; 754 struct thread *td = ap->a_td; 755 756 if (vp->v_vnlock != NULL) 757 return (lockstatus(vp->v_vnlock, td)); 758 return (lockstatus(&vp->v_lock, td)); 759} 760 761/* 762 * There is no way to tell that someone issued remove/rmdir operation 763 * on the underlying filesystem. For now we just have to release lowevrp 764 * as soon as possible. 765 * 766 * Note, we can't release any resources nor remove vnode from hash before 767 * appropriate VXLOCK stuff is is done because other process can find this 768 * vnode in hash during inactivation and may be sitting in vget() and waiting 769 * for null_inactive to unlock vnode. Thus we will do all those in VOP_RECLAIM. 770 */ 771static int 772null_inactive(ap) 773 struct vop_inactive_args /* { 774 struct vnode *a_vp; 775 struct thread *a_td; 776 } */ *ap; 777{ 778 struct vnode *vp = ap->a_vp; 779 struct thread *td = ap->a_td; 780 781 VOP_UNLOCK(vp, 0, td); 782 783 /* 784 * If this is the last reference, then free up the vnode 785 * so as not to tie up the lower vnodes. 786 */ 787 vrecycle(vp, NULL, td); 788 789 return (0); 790} 791 792/* 793 * Now, the VXLOCK is in force and we're free to destroy the null vnode. 794 */ 795static int 796null_reclaim(ap) 797 struct vop_reclaim_args /* { 798 struct vnode *a_vp; 799 struct thread *a_td; 800 } */ *ap; 801{ 802 struct vnode *vp = ap->a_vp; 803 struct null_node *xp = VTONULL(vp); 804 struct vnode *lowervp = xp->null_lowervp; 805 806 if (lowervp) { 807 null_hashrem(xp); 808 809 vrele(lowervp); 810 vrele(lowervp); 811 } 812 813 vp->v_data = NULL; 814 vp->v_vnlock = &vp->v_lock; 815 FREE(xp, M_NULLFSNODE); 816 817 return (0); 818} 819 820static int 821null_print(ap) 822 struct vop_print_args /* { 823 struct vnode *a_vp; 824 } */ *ap; 825{ 826 register struct vnode *vp = ap->a_vp; 827 printf("\tvp=%p, lowervp=%p\n", vp, NULLVPTOLOWERVP(vp)); 828 return (0); 829} 830 831/* 832 * Let an underlying filesystem do the work 833 */ 834static int 835null_createvobject(ap) 836 struct vop_createvobject_args /* { 837 struct vnode *vp; 838 struct ucred *cred; 839 struct thread *td; 840 } */ *ap; 841{ 842 struct vnode *vp = ap->a_vp; 843 struct vnode *lowervp = VTONULL(vp) ? NULLVPTOLOWERVP(vp) : NULL; 844 int error; 845 846 if (vp->v_type == VNON || lowervp == NULL) 847 return 0; 848 error = VOP_CREATEVOBJECT(lowervp, ap->a_cred, ap->a_td); 849 if (error) 850 return (error); 851 vp->v_vflag |= VV_OBJBUF; 852 return (0); 853} 854 855/* 856 * We have nothing to destroy and this operation shouldn't be bypassed. 857 */ 858static int 859null_destroyvobject(ap) 860 struct vop_destroyvobject_args /* { 861 struct vnode *vp; 862 } */ *ap; 863{ 864 struct vnode *vp = ap->a_vp; 865 866 vp->v_vflag &= ~VV_OBJBUF; 867 return (0); 868} 869 870static int 871null_getvobject(ap) 872 struct vop_getvobject_args /* { 873 struct vnode *vp; 874 struct vm_object **objpp; 875 } */ *ap; 876{ 877 struct vnode *lvp = NULLVPTOLOWERVP(ap->a_vp); 878 879 if (lvp == NULL) 880 return EINVAL; 881 return (VOP_GETVOBJECT(lvp, ap->a_objpp)); 882} 883 884/* 885 * Global vfs data structures 886 */ 887vop_t **null_vnodeop_p; 888static struct vnodeopv_entry_desc null_vnodeop_entries[] = { 889 { &vop_default_desc, (vop_t *) null_bypass }, 890 891 { &vop_access_desc, (vop_t *) null_access }, 892 { &vop_bmap_desc, (vop_t *) vop_eopnotsupp }, 893 { &vop_createvobject_desc, (vop_t *) null_createvobject }, 894 { &vop_destroyvobject_desc, (vop_t *) null_destroyvobject }, 895 { &vop_getattr_desc, (vop_t *) null_getattr }, 896 { &vop_getvobject_desc, (vop_t *) null_getvobject }, 897 { &vop_getwritemount_desc, (vop_t *) vop_stdgetwritemount}, 898 { &vop_inactive_desc, (vop_t *) null_inactive }, 899 { &vop_islocked_desc, (vop_t *) null_islocked }, 900 { &vop_lock_desc, (vop_t *) null_lock }, 901 { &vop_lookup_desc, (vop_t *) null_lookup }, 902 { &vop_open_desc, (vop_t *) null_open }, 903 { &vop_print_desc, (vop_t *) null_print }, 904 { &vop_reclaim_desc, (vop_t *) null_reclaim }, 905 { &vop_rename_desc, (vop_t *) null_rename }, 906 { &vop_setattr_desc, (vop_t *) null_setattr }, 907 { &vop_strategy_desc, (vop_t *) vop_eopnotsupp }, 908 { &vop_unlock_desc, (vop_t *) null_unlock }, 909 { NULL, NULL } 910}; 911static struct vnodeopv_desc null_vnodeop_opv_desc = 912 { &null_vnodeop_p, null_vnodeop_entries }; 913 914VNODEOP_SET(null_vnodeop_opv_desc); 915