null_vnops.c revision 143630
1193323Sed/*- 2193323Sed * Copyright (c) 1992, 1993 3193323Sed * The Regents of the University of California. All rights reserved. 4193323Sed * 5193323Sed * This code is derived from software contributed to Berkeley by 6193323Sed * John Heidemann of the UCLA Ficus project. 7193323Sed * 8193323Sed * Redistribution and use in source and binary forms, with or without 9193323Sed * modification, are permitted provided that the following conditions 10193323Sed * are met: 11193323Sed * 1. Redistributions of source code must retain the above copyright 12193323Sed * notice, this list of conditions and the following disclaimer. 13193323Sed * 2. Redistributions in binary form must reproduce the above copyright 14193323Sed * notice, this list of conditions and the following disclaimer in the 15193323Sed * documentation and/or other materials provided with the distribution. 16193323Sed * 4. Neither the name of the University nor the names of its contributors 17193323Sed * may be used to endorse or promote products derived from this software 18193323Sed * without specific prior written permission. 19193323Sed * 20193323Sed * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21198090Srdivacky * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22193323Sed * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23193323Sed * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24193323Sed * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25193323Sed * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26193323Sed * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27193323Sed * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28193323Sed * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29193323Sed * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30193323Sed * SUCH DAMAGE. 31234353Sdim * 32193323Sed * @(#)null_vnops.c 8.6 (Berkeley) 5/27/95 33193323Sed * 34193323Sed * Ancestors: 35193323Sed * @(#)lofs_vnops.c 1.2 (Berkeley) 6/18/92 36193323Sed * ...and... 37193323Sed * @(#)null_vnodeops.c 1.20 92/07/07 UCLA Ficus project 38193323Sed * 39193323Sed * $FreeBSD: head/sys/fs/nullfs/null_vnops.c 143630 2005-03-15 11:28:45Z jeff $ 40193323Sed */ 41193323Sed 42193323Sed/* 43193323Sed * Null Layer 44193323Sed * 45193323Sed * (See mount_nullfs(8) for more information.) 46193323Sed * 47193323Sed * The null layer duplicates a portion of the filesystem 48205218Srdivacky * name space under a new name. In this respect, it is 49205218Srdivacky * similar to the loopback filesystem. It differs from 50193323Sed * the loopback fs in two respects: it is implemented using 51218885Sdim * a stackable layers techniques, and its "null-node"s stack above 52218885Sdim * all lower-layer vnodes, not just over directory vnodes. 53218885Sdim * 54193323Sed * The null layer has two purposes. First, it serves as a demonstration 55193323Sed * of layering by proving a layer which does nothing. (It actually 56193323Sed * does everything the loopback filesystem does, which is slightly 57193323Sed * more than nothing.) Second, the null layer can serve as a prototype 58193323Sed * layer. Since it provides all necessary layer framework, 59193323Sed * new filesystem layers can be created very easily be starting 60193323Sed * with a null layer. 61193323Sed * 62198090Srdivacky * The remainder of this man page examines the null layer as a basis 63198090Srdivacky * for constructing new layers. 64198090Srdivacky * 65198090Srdivacky * 66198090Srdivacky * INSTANTIATING NEW NULL LAYERS 67198090Srdivacky * 68193323Sed * New null layers are created with mount_nullfs(8). 69193323Sed * Mount_nullfs(8) takes two arguments, the pathname 70193323Sed * of the lower vfs (target-pn) and the pathname where the null 71193323Sed * layer will appear in the namespace (alias-pn). After 72198090Srdivacky * the null layer is put into place, the contents 73198090Srdivacky * of target-pn subtree will be aliased under alias-pn. 74193323Sed * 75193323Sed * 76193323Sed * OPERATION OF A NULL LAYER 77193323Sed * 78193323Sed * The null layer is the minimum filesystem layer, 79193323Sed * simply bypassing all possible operations to the lower layer 80193323Sed * for processing there. The majority of its activity centers 81193323Sed * on the bypass routine, through which nearly all vnode operations 82193323Sed * pass. 83193323Sed * 84193323Sed * The bypass routine accepts arbitrary vnode operations for 85193323Sed * handling by the lower layer. It begins by examing vnode 86193323Sed * operation arguments and replacing any null-nodes by their 87193323Sed * lower-layer equivlants. It then invokes the operation 88193323Sed * on the lower layer. Finally, it replaces the null-nodes 89193323Sed * in the arguments and, if a vnode is return by the operation, 90218885Sdim * stacks a null-node on top of the returned vnode. 91193323Sed * 92193323Sed * Although bypass handles most operations, vop_getattr, vop_lock, 93193323Sed * vop_unlock, vop_inactive, vop_reclaim, and vop_print are not 94218885Sdim * bypassed. Vop_getattr must change the fsid being returned. 95218885Sdim * Vop_lock and vop_unlock must handle any locking for the 96218885Sdim * current vnode as well as pass the lock request down. 97218885Sdim * Vop_inactive and vop_reclaim are not bypassed so that 98218885Sdim * they can handle freeing null-layer specific data. Vop_print 99193323Sed * is not bypassed to avoid excessive debugging information. 100193323Sed * Also, certain vnode operations change the locking state within 101193323Sed * the operation (create, mknod, remove, link, rename, mkdir, rmdir, 102193323Sed * and symlink). Ideally these operations should not change the 103193323Sed * lock state, but should be changed to let the caller of the 104193323Sed * function unlock them. Otherwise all intermediate vnode layers 105193323Sed * (such as union, umapfs, etc) must catch these functions to do 106193323Sed * the necessary locking at their layer. 107193323Sed * 108193323Sed * 109193323Sed * INSTANTIATING VNODE STACKS 110198090Srdivacky * 111198090Srdivacky * Mounting associates the null layer with a lower layer, 112198090Srdivacky * effect stacking two VFSes. Vnode stacks are instead 113198090Srdivacky * created on demand as files are accessed. 114193323Sed * 115193323Sed * The initial mount creates a single vnode stack for the 116193323Sed * root of the new null layer. All other vnode stacks 117193323Sed * are created as a result of vnode operations on 118193323Sed * this or other null vnode stacks. 119193323Sed * 120193323Sed * New vnode stacks come into existance as a result of 121193323Sed * an operation which returns a vnode. 122193323Sed * The bypass routine stacks a null-node above the new 123193323Sed * vnode before returning it to the caller. 124193323Sed * 125193323Sed * For example, imagine mounting a null layer with 126208599Srdivacky * "mount_nullfs /usr/include /dev/layer/null". 127218885Sdim * Changing directory to /dev/layer/null will assign 128208599Srdivacky * the root null-node (which was created when the null layer was mounted). 129208599Srdivacky * Now consider opening "sys". A vop_lookup would be 130193323Sed * done on the root null-node. This operation would bypass through 131193323Sed * to the lower layer which would return a vnode representing 132193323Sed * the UFS "sys". Null_bypass then builds a null-node 133193323Sed * aliasing the UFS "sys" and returns this to the caller. 134193323Sed * Later operations on the null-node "sys" will repeat this 135193323Sed * process when constructing other vnode stacks. 136193323Sed * 137193323Sed * 138193323Sed * CREATING OTHER FILE SYSTEM LAYERS 139193323Sed * 140193323Sed * One of the easiest ways to construct new filesystem layers is to make 141193323Sed * a copy of the null layer, rename all files and variables, and 142193323Sed * then begin modifing the copy. Sed can be used to easily rename 143193323Sed * all variables. 144198090Srdivacky * 145193323Sed * The umap layer is an example of a layer descended from the 146193323Sed * null layer. 147193323Sed * 148193323Sed * 149193323Sed * INVOKING OPERATIONS ON LOWER LAYERS 150193323Sed * 151193323Sed * There are two techniques to invoke operations on a lower layer 152193323Sed * when the operation cannot be completely bypassed. Each method 153193323Sed * is appropriate in different situations. In both cases, 154193323Sed * it is the responsibility of the aliasing layer to make 155193323Sed * the operation arguments "correct" for the lower layer 156193323Sed * by mapping a vnode arguments to the lower layer. 157193323Sed * 158193323Sed * The first approach is to call the aliasing layer's bypass routine. 159193323Sed * This method is most suitable when you wish to invoke the operation 160193323Sed * currently being handled on the lower layer. It has the advantage 161218885Sdim * that the bypass routine already must do argument mapping. 162208599Srdivacky * An example of this is null_getattrs in the null layer. 163193323Sed * 164193323Sed * A second approach is to directly invoke vnode operations on 165193323Sed * the lower layer with the VOP_OPERATIONNAME interface. 166193323Sed * The advantage of this method is that it is easy to invoke 167193323Sed * arbitrary operations on the lower layer. The disadvantage 168193323Sed * is that vnode arguments must be manualy mapped. 169193323Sed * 170193323Sed */ 171193323Sed 172193323Sed#include <sys/param.h> 173193323Sed#include <sys/systm.h> 174193323Sed#include <sys/conf.h> 175193323Sed#include <sys/kernel.h> 176193323Sed#include <sys/lock.h> 177193323Sed#include <sys/malloc.h> 178193323Sed#include <sys/mount.h> 179205218Srdivacky#include <sys/mutex.h> 180193323Sed#include <sys/namei.h> 181193323Sed#include <sys/sysctl.h> 182218885Sdim#include <sys/vnode.h> 183208599Srdivacky 184205218Srdivacky#include <fs/nullfs/null.h> 185193323Sed 186193323Sed#include <vm/vm.h> 187193323Sed#include <vm/vm_extern.h> 188193323Sed#include <vm/vm_object.h> 189193323Sed#include <vm/vnode_pager.h> 190218885Sdim 191218885Sdimstatic int null_bug_bypass = 0; /* for debugging: enables bypass printf'ing */ 192218885SdimSYSCTL_INT(_debug, OID_AUTO, nullfs_bug_bypass, CTLFLAG_RW, 193218885Sdim &null_bug_bypass, 0, ""); 194193323Sed 195218885Sdim/* 196218885Sdim * This is the 10-Apr-92 bypass routine. 197193323Sed * This version has been optimized for speed, throwing away some 198193323Sed * safety checks. It should still always work, but it's not as 199193323Sed * robust to programmer errors. 200198090Srdivacky * 201193323Sed * In general, we map all vnodes going down and unmap them on the way back. 202198090Srdivacky * As an exception to this, vnodes can be marked "unmapped" by setting 203193323Sed * the Nth bit in operation's vdesc_flags. 204193323Sed * 205193323Sed * Also, some BSD vnode operations have the side effect of vrele'ing 206193323Sed * their arguments. With stacking, the reference counts are held 207193323Sed * by the upper node, not the lower one, so we must handle these 208193323Sed * side-effects here. This is not of concern in Sun-derived systems 209193323Sed * since there are no such side-effects. 210193323Sed * 211193323Sed * This makes the following assumptions: 212193323Sed * - only one returned vpp 213193323Sed * - no INOUT vpp's (Sun's vop_open has one of these) 214193323Sed * - the vnode operation vector of the first vnode should be used 215198090Srdivacky * to determine what implementation of the op should be invoked 216218885Sdim * - all mapped vnodes are of our vnode-type (NEEDSWORK: 217193323Sed * problems on rmdir'ing mount points and renaming?) 218193323Sed */ 219193323Sedint 220193323Sednull_bypass(struct vop_generic_args *ap) 221193323Sed{ 222193323Sed struct vnode **this_vp_p; 223193323Sed int error; 224193323Sed struct vnode *old_vps[VDESC_MAX_VPS]; 225193323Sed struct vnode **vps_p[VDESC_MAX_VPS]; 226198090Srdivacky struct vnode ***vppp; 227218885Sdim struct vnodeop_desc *descp = ap->a_desc; 228193323Sed int reles, i; 229193323Sed 230193323Sed if (null_bug_bypass) 231193323Sed printf ("null_bypass: %s\n", descp->vdesc_name); 232193323Sed 233249423Sdim#ifdef DIAGNOSTIC 234193323Sed /* 235193323Sed * We require at least one vp. 236193323Sed */ 237205218Srdivacky if (descp->vdesc_vp_offsets == NULL || 238193323Sed descp->vdesc_vp_offsets[0] == VDESC_NO_OFFSET) 239198090Srdivacky panic ("null_bypass: no vp's in map"); 240208599Srdivacky#endif 241205218Srdivacky 242205218Srdivacky /* 243193323Sed * Map the vnodes going in. 244193323Sed * Later, we'll invoke the operation based on 245218885Sdim * the first mapped vnode's operation vector. 246218885Sdim */ 247193323Sed reles = descp->vdesc_flags; 248193323Sed for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) { 249193323Sed if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET) 250193323Sed break; /* bail out at end of list */ 251193323Sed vps_p[i] = this_vp_p = 252193323Sed VOPARG_OFFSETTO(struct vnode**,descp->vdesc_vp_offsets[i],ap); 253198090Srdivacky /* 254218885Sdim * We're not guaranteed that any but the first vnode 255208599Srdivacky * are of our type. Check for and don't map any 256198090Srdivacky * that aren't. (We must always map first vp or vclean fails.) 257193323Sed */ 258193323Sed if (i && (*this_vp_p == NULLVP || 259193323Sed (*this_vp_p)->v_op != &null_vnodeops)) { 260193323Sed old_vps[i] = NULLVP; 261193323Sed } else { 262207618Srdivacky old_vps[i] = *this_vp_p; 263207618Srdivacky *(vps_p[i]) = NULLVPTOLOWERVP(*this_vp_p); 264207618Srdivacky /* 265193323Sed * XXX - Several operations have the side effect 266212793Sdim * of vrele'ing their vp's. We must account for 267193323Sed * that. (This should go away in the future.) 268198090Srdivacky */ 269193323Sed if (reles & VDESC_VP0_WILLRELE) 270207618Srdivacky VREF(*this_vp_p); 271218885Sdim } 272198090Srdivacky 273193323Sed } 274193323Sed 275198090Srdivacky /* 276198090Srdivacky * Call the operation on the lower layer 277198090Srdivacky * with the modified argument structure. 278193323Sed */ 279193323Sed if (vps_p[0] && *vps_p[0]) 280193323Sed error = VCALL(ap); 281207618Srdivacky else { 282221337Sdim printf("null_bypass: no map for %s\n", descp->vdesc_name); 283193323Sed error = EINVAL; 284193323Sed } 285208599Srdivacky 286193323Sed /* 287193323Sed * Maintain the illusion of call-by-value 288193323Sed * by restoring vnodes in the argument structure 289193323Sed * to their original value. 290193323Sed */ 291193323Sed reles = descp->vdesc_flags; 292193323Sed for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) { 293212793Sdim if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET) 294212793Sdim break; /* bail out at end of list */ 295193323Sed if (old_vps[i]) { 296193323Sed *(vps_p[i]) = old_vps[i]; 297193323Sed#if 0 298212793Sdim if (reles & VDESC_VP0_WILLUNLOCK) 299193323Sed VOP_UNLOCK(*(vps_p[i]), LK_THISLAYER, curthread); 300193323Sed#endif 301193323Sed if (reles & VDESC_VP0_WILLRELE) 302193323Sed vrele(*(vps_p[i])); 303193323Sed } 304193323Sed } 305198090Srdivacky 306193323Sed /* 307198090Srdivacky * Map the possible out-going vpp 308198090Srdivacky * (Assumes that the lower layer always returns 309193323Sed * a VREF'ed vpp unless it gets an error.) 310193323Sed */ 311198090Srdivacky if (descp->vdesc_vpp_offset != VDESC_NO_OFFSET && 312193323Sed !(descp->vdesc_flags & VDESC_NOMAP_VPP) && 313193323Sed !error) { 314198090Srdivacky /* 315198090Srdivacky * XXX - even though some ops have vpp returned vp's, 316193323Sed * several ops actually vrele this before returning. 317193323Sed * We must avoid these ops. 318193323Sed * (This should go away when these ops are regularized.) 319193323Sed */ 320193323Sed if (descp->vdesc_flags & VDESC_VPP_WILLRELE) 321193323Sed goto out; 322207618Srdivacky vppp = VOPARG_OFFSETTO(struct vnode***, 323221337Sdim descp->vdesc_vpp_offset,ap); 324221337Sdim if (*vppp) 325193323Sed error = null_nodeget(old_vps[0]->v_mount, **vppp, *vppp); 326198090Srdivacky } 327193323Sed 328193323Sed out: 329193323Sed return (error); 330193323Sed} 331198090Srdivacky 332198090Srdivacky/* 333193323Sed * We have to carry on the locking protocol on the null layer vnodes 334193323Sed * as we progress through the tree. We also have to enforce read-only 335198090Srdivacky * if this layer is mounted read-only. 336193323Sed */ 337193323Sedstatic int 338193323Sednull_lookup(struct vop_lookup_args *ap) 339193323Sed{ 340193323Sed struct componentname *cnp = ap->a_cnp; 341193323Sed struct vnode *dvp = ap->a_dvp; 342207618Srdivacky struct thread *td = cnp->cn_thread; 343207618Srdivacky int flags = cnp->cn_flags; 344193323Sed struct vnode *vp, *ldvp, *lvp; 345207618Srdivacky int error; 346207618Srdivacky 347193323Sed if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) && 348193323Sed (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) 349198090Srdivacky return (EROFS); 350193323Sed /* 351193323Sed * Although it is possible to call null_bypass(), we'll do 352198090Srdivacky * a direct call to reduce overhead 353193323Sed */ 354193323Sed ldvp = NULLVPTOLOWERVP(dvp); 355193323Sed vp = lvp = NULL; 356193323Sed error = VOP_LOOKUP(ldvp, &lvp, cnp); 357193323Sed if (error == EJUSTRETURN && (flags & ISLASTCN) && 358193323Sed (dvp->v_mount->mnt_flag & MNT_RDONLY) && 359193323Sed (cnp->cn_nameiop == CREATE || cnp->cn_nameiop == RENAME)) 360193323Sed error = EROFS; 361193323Sed 362193323Sed /* 363193323Sed * Rely only on the PDIRUNLOCK flag which should be carefully 364193323Sed * tracked by underlying filesystem. 365193323Sed */ 366193323Sed if ((cnp->cn_flags & PDIRUNLOCK) && dvp->v_vnlock != ldvp->v_vnlock) 367193323Sed VOP_UNLOCK(dvp, LK_THISLAYER, td); 368193323Sed if ((error == 0 || error == EJUSTRETURN) && lvp != NULL) { 369193323Sed if (ldvp == lvp) { 370193323Sed *ap->a_vpp = dvp; 371193323Sed VREF(dvp); 372193323Sed vrele(lvp); 373193323Sed } else { 374212793Sdim error = null_nodeget(dvp->v_mount, lvp, &vp); 375212793Sdim if (error) { 376212793Sdim /* XXX Cleanup needed... */ 377212793Sdim panic("null_nodeget failed"); 378193323Sed } 379193323Sed *ap->a_vpp = vp; 380207618Srdivacky } 381207618Srdivacky } 382193323Sed return (error); 383193323Sed} 384193323Sed 385193323Sedstatic int 386207618Srdivackynull_open(struct vop_open_args *ap) 387207618Srdivacky{ 388207618Srdivacky int retval; 389207618Srdivacky struct vnode *vp, *ldvp; 390193323Sed 391193323Sed vp = ap->a_vp; 392207618Srdivacky ldvp = NULLVPTOLOWERVP(vp); 393207618Srdivacky retval = null_bypass(&ap->a_gen); 394207618Srdivacky if (retval == 0) 395207618Srdivacky vp->v_object = ldvp->v_object; 396207618Srdivacky return (retval); 397193323Sed} 398193323Sed 399193323Sed/* 400193323Sed * Setattr call. Disallow write attempts if the layer is mounted read-only. 401193323Sed */ 402193323Sedstatic int 403193323Sednull_setattr(struct vop_setattr_args *ap) 404193323Sed{ 405193323Sed struct vnode *vp = ap->a_vp; 406218885Sdim struct vattr *vap = ap->a_vap; 407193323Sed 408193323Sed if ((vap->va_flags != VNOVAL || vap->va_uid != (uid_t)VNOVAL || 409193323Sed vap->va_gid != (gid_t)VNOVAL || vap->va_atime.tv_sec != VNOVAL || 410193323Sed vap->va_mtime.tv_sec != VNOVAL || vap->va_mode != (mode_t)VNOVAL) && 411207618Srdivacky (vp->v_mount->mnt_flag & MNT_RDONLY)) 412207618Srdivacky return (EROFS); 413207618Srdivacky if (vap->va_size != VNOVAL) { 414193323Sed switch (vp->v_type) { 415207618Srdivacky case VDIR: 416212793Sdim return (EISDIR); 417207618Srdivacky case VCHR: 418207618Srdivacky case VBLK: 419193323Sed case VSOCK: 420198090Srdivacky case VFIFO: 421234353Sdim if (vap->va_flags != VNOVAL) 422234353Sdim return (EOPNOTSUPP); 423198090Srdivacky return (0); 424193323Sed case VREG: 425193323Sed case VLNK: 426193323Sed default: 427207618Srdivacky /* 428193323Sed * Disallow write attempts if the filesystem is 429193323Sed * mounted read-only. 430193323Sed */ 431193323Sed if (vp->v_mount->mnt_flag & MNT_RDONLY) 432193323Sed return (EROFS); 433207618Srdivacky } 434223013Sdim } 435193323Sed 436212793Sdim return (null_bypass((struct vop_generic_args *)ap)); 437212793Sdim} 438193323Sed 439207618Srdivacky/* 440212793Sdim * We handle getattr only to change the fsid. 441193323Sed */ 442212793Sdimstatic int 443212793Sdimnull_getattr(struct vop_getattr_args *ap) 444207618Srdivacky{ 445207618Srdivacky int error; 446193323Sed 447193323Sed if ((error = null_bypass((struct vop_generic_args *)ap)) != 0) 448193323Sed return (error); 449193323Sed 450198090Srdivacky ap->a_vap->va_fsid = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0]; 451193323Sed return (0); 452193323Sed} 453198090Srdivacky 454193323Sed/* 455193323Sed * Handle to disallow write access if mounted read-only. 456193323Sed */ 457193323Sedstatic int 458198090Srdivackynull_access(struct vop_access_args *ap) 459198090Srdivacky{ 460198090Srdivacky struct vnode *vp = ap->a_vp; 461198090Srdivacky mode_t mode = ap->a_mode; 462193323Sed 463193323Sed /* 464193323Sed * Disallow write attempts on read-only layers; 465193323Sed * unless the file is a socket, fifo, or a block or 466193323Sed * character device resident on the filesystem. 467193323Sed */ 468193323Sed if (mode & VWRITE) { 469193323Sed switch (vp->v_type) { 470193323Sed case VDIR: 471193323Sed case VLNK: 472193323Sed case VREG: 473 if (vp->v_mount->mnt_flag & MNT_RDONLY) 474 return (EROFS); 475 break; 476 default: 477 break; 478 } 479 } 480 return (null_bypass((struct vop_generic_args *)ap)); 481} 482 483/* 484 * We handle this to eliminate null FS to lower FS 485 * file moving. Don't know why we don't allow this, 486 * possibly we should. 487 */ 488static int 489null_rename(struct vop_rename_args *ap) 490{ 491 struct vnode *tdvp = ap->a_tdvp; 492 struct vnode *fvp = ap->a_fvp; 493 struct vnode *fdvp = ap->a_fdvp; 494 struct vnode *tvp = ap->a_tvp; 495 496 /* Check for cross-device rename. */ 497 if ((fvp->v_mount != tdvp->v_mount) || 498 (tvp && (fvp->v_mount != tvp->v_mount))) { 499 if (tdvp == tvp) 500 vrele(tdvp); 501 else 502 vput(tdvp); 503 if (tvp) 504 vput(tvp); 505 vrele(fdvp); 506 vrele(fvp); 507 return (EXDEV); 508 } 509 510 return (null_bypass((struct vop_generic_args *)ap)); 511} 512 513/* 514 * We need to process our own vnode lock and then clear the 515 * interlock flag as it applies only to our vnode, not the 516 * vnodes below us on the stack. 517 */ 518static int 519null_lock(struct vop_lock_args *ap) 520{ 521 struct vnode *vp = ap->a_vp; 522 int flags = ap->a_flags; 523 struct thread *td = ap->a_td; 524 struct vnode *lvp; 525 int error; 526 struct null_node *nn; 527 528 if (flags & LK_THISLAYER) { 529 if (vp->v_vnlock != NULL) { 530 /* lock is shared across layers */ 531 if (flags & LK_INTERLOCK) 532 mtx_unlock(&vp->v_interlock); 533 return 0; 534 } 535 error = lockmgr(&vp->v_lock, flags & ~LK_THISLAYER, 536 &vp->v_interlock, td); 537 return (error); 538 } 539 540 if ((flags & LK_INTERLOCK) == 0) { 541 VI_LOCK(vp); 542 flags |= LK_INTERLOCK; 543 } 544 if (vp->v_vnlock != NULL) { 545 /* 546 * The lower level has exported a struct lock to us. Use 547 * it so that all vnodes in the stack lock and unlock 548 * simultaneously. Note: we don't DRAIN the lock as DRAIN 549 * decommissions the lock - just because our vnode is 550 * going away doesn't mean the struct lock below us is. 551 * LK_EXCLUSIVE is fine. 552 */ 553 nn = VTONULL(vp); 554 if ((flags & LK_TYPE_MASK) == LK_DRAIN) { 555 NULLFSDEBUG("null_lock: avoiding LK_DRAIN\n"); 556 /* 557 * Emulate lock draining by waiting for all other 558 * pending locks to complete. Afterwards the 559 * lockmgr call might block, but no other threads 560 * will attempt to use this nullfs vnode due to the 561 * VI_DOOMED flag. 562 */ 563 while (nn->null_pending_locks > 0) { 564 nn->null_drain_wakeup = 1; 565 msleep(&nn->null_pending_locks, 566 VI_MTX(vp), 567 PVFS, 568 "nuldr", 0); 569 } 570 error = lockmgr(vp->v_vnlock, 571 (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE, 572 VI_MTX(vp), td); 573 return error; 574 } 575 nn->null_pending_locks++; 576 error = lockmgr(vp->v_vnlock, flags, &vp->v_interlock, td); 577 VI_LOCK(vp); 578 /* 579 * If we're called from vrele then v_usecount can have been 0 580 * and another process might have initiated a recycle 581 * operation. When that happens, just back out. 582 */ 583 if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0 && 584 td != vp->v_vxthread) { 585 lockmgr(vp->v_vnlock, 586 (flags & ~LK_TYPE_MASK) | LK_RELEASE, 587 VI_MTX(vp), td); 588 VI_LOCK(vp); 589 error = ENOENT; 590 } 591 nn->null_pending_locks--; 592 /* 593 * Wakeup the process draining the vnode after all 594 * pending lock attempts has been failed. 595 */ 596 if (nn->null_pending_locks == 0 && 597 nn->null_drain_wakeup != 0) { 598 nn->null_drain_wakeup = 0; 599 wakeup(&nn->null_pending_locks); 600 } 601 VI_UNLOCK(vp); 602 return error; 603 } else { 604 /* 605 * To prevent race conditions involving doing a lookup 606 * on "..", we have to lock the lower node, then lock our 607 * node. Most of the time it won't matter that we lock our 608 * node (as any locking would need the lower one locked 609 * first). But we can LK_DRAIN the upper lock as a step 610 * towards decomissioning it. 611 */ 612 lvp = NULLVPTOLOWERVP(vp); 613 if (lvp == NULL) 614 return (lockmgr(&vp->v_lock, flags, &vp->v_interlock, td)); 615 if (flags & LK_INTERLOCK) { 616 mtx_unlock(&vp->v_interlock); 617 flags &= ~LK_INTERLOCK; 618 } 619 if ((flags & LK_TYPE_MASK) == LK_DRAIN) { 620 error = VOP_LOCK(lvp, 621 (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE, td); 622 } else 623 error = VOP_LOCK(lvp, flags, td); 624 if (error) 625 return (error); 626 error = lockmgr(&vp->v_lock, flags, &vp->v_interlock, td); 627 if (error) 628 VOP_UNLOCK(lvp, 0, td); 629 return (error); 630 } 631} 632 633/* 634 * We need to process our own vnode unlock and then clear the 635 * interlock flag as it applies only to our vnode, not the 636 * vnodes below us on the stack. 637 */ 638static int 639null_unlock(struct vop_unlock_args *ap) 640{ 641 struct vnode *vp = ap->a_vp; 642 int flags = ap->a_flags; 643 struct thread *td = ap->a_td; 644 struct vnode *lvp; 645 646 if (vp->v_vnlock != NULL) { 647 if (flags & LK_THISLAYER) 648 return 0; /* the lock is shared across layers */ 649 flags &= ~LK_THISLAYER; 650 return (lockmgr(vp->v_vnlock, flags | LK_RELEASE, 651 &vp->v_interlock, td)); 652 } 653 lvp = NULLVPTOLOWERVP(vp); 654 if (lvp == NULL) 655 return (lockmgr(&vp->v_lock, flags | LK_RELEASE, &vp->v_interlock, td)); 656 if ((flags & LK_THISLAYER) == 0) { 657 if (flags & LK_INTERLOCK) { 658 mtx_unlock(&vp->v_interlock); 659 flags &= ~LK_INTERLOCK; 660 } 661 VOP_UNLOCK(lvp, flags & ~LK_INTERLOCK, td); 662 } else 663 flags &= ~LK_THISLAYER; 664 return (lockmgr(&vp->v_lock, flags | LK_RELEASE, &vp->v_interlock, td)); 665} 666 667static int 668null_islocked(struct vop_islocked_args *ap) 669{ 670 struct vnode *vp = ap->a_vp; 671 struct thread *td = ap->a_td; 672 673 if (vp->v_vnlock != NULL) 674 return (lockstatus(vp->v_vnlock, td)); 675 return (lockstatus(&vp->v_lock, td)); 676} 677 678/* 679 * There is no way to tell that someone issued remove/rmdir operation 680 * on the underlying filesystem. For now we just have to release lowevrp 681 * as soon as possible. 682 * 683 * Note, we can't release any resources nor remove vnode from hash before 684 * appropriate VXLOCK stuff is is done because other process can find this 685 * vnode in hash during inactivation and may be sitting in vget() and waiting 686 * for null_inactive to unlock vnode. Thus we will do all those in VOP_RECLAIM. 687 */ 688static int 689null_inactive(struct vop_inactive_args *ap) 690{ 691 struct vnode *vp = ap->a_vp; 692 struct thread *td = ap->a_td; 693 694 vp->v_object = NULL; 695 696 /* 697 * If this is the last reference, then free up the vnode 698 * so as not to tie up the lower vnodes. 699 */ 700 vrecycle(vp, td); 701 702 return (0); 703} 704 705/* 706 * Now, the VXLOCK is in force and we're free to destroy the null vnode. 707 */ 708static int 709null_reclaim(struct vop_reclaim_args *ap) 710{ 711 struct vnode *vp = ap->a_vp; 712 struct null_node *xp = VTONULL(vp); 713 struct vnode *lowervp = xp->null_lowervp; 714 struct lock *vnlock; 715 716 if (lowervp) { 717 null_hashrem(xp); 718 719 vrele(lowervp); 720 vrele(lowervp); 721 } 722 723 vp->v_data = NULL; 724 vp->v_object = NULL; 725 vnlock = vp->v_vnlock; 726 lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL, curthread); 727 vp->v_vnlock = &vp->v_lock; 728 transferlockers(vnlock, vp->v_vnlock); 729 lockmgr(vnlock, LK_RELEASE, NULL, curthread); 730 FREE(xp, M_NULLFSNODE); 731 732 return (0); 733} 734 735static int 736null_print(struct vop_print_args *ap) 737{ 738 struct vnode *vp = ap->a_vp; 739 printf("\tvp=%p, lowervp=%p\n", vp, NULLVPTOLOWERVP(vp)); 740 return (0); 741} 742 743/* 744 * Global vfs data structures 745 */ 746struct vop_vector null_vnodeops = { 747 .vop_bypass = null_bypass, 748 749 .vop_access = null_access, 750 .vop_bmap = VOP_EOPNOTSUPP, 751 .vop_getattr = null_getattr, 752 .vop_getwritemount = vop_stdgetwritemount, 753 .vop_inactive = null_inactive, 754 .vop_islocked = null_islocked, 755 .vop_lock = null_lock, 756 .vop_lookup = null_lookup, 757 .vop_open = null_open, 758 .vop_print = null_print, 759 .vop_reclaim = null_reclaim, 760 .vop_rename = null_rename, 761 .vop_setattr = null_setattr, 762 .vop_strategy = VOP_EOPNOTSUPP, 763 .vop_unlock = null_unlock, 764}; 765