null_vnops.c revision 140165
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 * 4. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * @(#)null_vnops.c 8.6 (Berkeley) 5/27/95 33 * 34 * Ancestors: 35 * @(#)lofs_vnops.c 1.2 (Berkeley) 6/18/92 36 * ...and... 37 * @(#)null_vnodeops.c 1.20 92/07/07 UCLA Ficus project 38 * 39 * $FreeBSD: head/sys/fs/nullfs/null_vnops.c 140165 2005-01-13 07:53:01Z phk $ 40 */ 41 42/* 43 * Null Layer 44 * 45 * (See mount_nullfs(8) for more information.) 46 * 47 * The null layer duplicates a portion of the filesystem 48 * name space under a new name. In this respect, it is 49 * similar to the loopback filesystem. It differs from 50 * the loopback fs in two respects: it is implemented using 51 * a stackable layers techniques, and its "null-node"s stack above 52 * all lower-layer vnodes, not just over directory vnodes. 53 * 54 * The null layer has two purposes. First, it serves as a demonstration 55 * of layering by proving a layer which does nothing. (It actually 56 * does everything the loopback filesystem does, which is slightly 57 * more than nothing.) Second, the null layer can serve as a prototype 58 * layer. Since it provides all necessary layer framework, 59 * new filesystem layers can be created very easily be starting 60 * with a null layer. 61 * 62 * The remainder of this man page examines the null layer as a basis 63 * for constructing new layers. 64 * 65 * 66 * INSTANTIATING NEW NULL LAYERS 67 * 68 * New null layers are created with mount_nullfs(8). 69 * Mount_nullfs(8) takes two arguments, the pathname 70 * of the lower vfs (target-pn) and the pathname where the null 71 * layer will appear in the namespace (alias-pn). After 72 * the null layer is put into place, the contents 73 * of target-pn subtree will be aliased under alias-pn. 74 * 75 * 76 * OPERATION OF A NULL LAYER 77 * 78 * The null layer is the minimum filesystem layer, 79 * simply bypassing all possible operations to the lower layer 80 * for processing there. The majority of its activity centers 81 * on the bypass routine, through which nearly all vnode operations 82 * pass. 83 * 84 * The bypass routine accepts arbitrary vnode operations for 85 * handling by the lower layer. It begins by examing vnode 86 * operation arguments and replacing any null-nodes by their 87 * lower-layer equivlants. It then invokes the operation 88 * on the lower layer. Finally, it replaces the null-nodes 89 * in the arguments and, if a vnode is return by the operation, 90 * stacks a null-node on top of the returned vnode. 91 * 92 * Although bypass handles most operations, vop_getattr, vop_lock, 93 * vop_unlock, vop_inactive, vop_reclaim, and vop_print are not 94 * bypassed. Vop_getattr must change the fsid being returned. 95 * Vop_lock and vop_unlock must handle any locking for the 96 * current vnode as well as pass the lock request down. 97 * Vop_inactive and vop_reclaim are not bypassed so that 98 * they can handle freeing null-layer specific data. Vop_print 99 * is not bypassed to avoid excessive debugging information. 100 * Also, certain vnode operations change the locking state within 101 * the operation (create, mknod, remove, link, rename, mkdir, rmdir, 102 * and symlink). Ideally these operations should not change the 103 * lock state, but should be changed to let the caller of the 104 * function unlock them. Otherwise all intermediate vnode layers 105 * (such as union, umapfs, etc) must catch these functions to do 106 * the necessary locking at their layer. 107 * 108 * 109 * INSTANTIATING VNODE STACKS 110 * 111 * Mounting associates the null layer with a lower layer, 112 * effect stacking two VFSes. Vnode stacks are instead 113 * created on demand as files are accessed. 114 * 115 * The initial mount creates a single vnode stack for the 116 * root of the new null layer. All other vnode stacks 117 * are created as a result of vnode operations on 118 * this or other null vnode stacks. 119 * 120 * New vnode stacks come into existance as a result of 121 * an operation which returns a vnode. 122 * The bypass routine stacks a null-node above the new 123 * vnode before returning it to the caller. 124 * 125 * For example, imagine mounting a null layer with 126 * "mount_nullfs /usr/include /dev/layer/null". 127 * Changing directory to /dev/layer/null will assign 128 * the root null-node (which was created when the null layer was mounted). 129 * Now consider opening "sys". A vop_lookup would be 130 * done on the root null-node. This operation would bypass through 131 * to the lower layer which would return a vnode representing 132 * the UFS "sys". Null_bypass then builds a null-node 133 * aliasing the UFS "sys" and returns this to the caller. 134 * Later operations on the null-node "sys" will repeat this 135 * process when constructing other vnode stacks. 136 * 137 * 138 * CREATING OTHER FILE SYSTEM LAYERS 139 * 140 * One of the easiest ways to construct new filesystem layers is to make 141 * a copy of the null layer, rename all files and variables, and 142 * then begin modifing the copy. Sed can be used to easily rename 143 * all variables. 144 * 145 * The umap layer is an example of a layer descended from the 146 * null layer. 147 * 148 * 149 * INVOKING OPERATIONS ON LOWER LAYERS 150 * 151 * There are two techniques to invoke operations on a lower layer 152 * when the operation cannot be completely bypassed. Each method 153 * is appropriate in different situations. In both cases, 154 * it is the responsibility of the aliasing layer to make 155 * the operation arguments "correct" for the lower layer 156 * by mapping a vnode arguments to the lower layer. 157 * 158 * The first approach is to call the aliasing layer's bypass routine. 159 * This method is most suitable when you wish to invoke the operation 160 * currently being handled on the lower layer. It has the advantage 161 * that the bypass routine already must do argument mapping. 162 * An example of this is null_getattrs in the null layer. 163 * 164 * A second approach is to directly invoke vnode operations on 165 * the lower layer with the VOP_OPERATIONNAME interface. 166 * The advantage of this method is that it is easy to invoke 167 * arbitrary operations on the lower layer. The disadvantage 168 * is that vnode arguments must be manualy mapped. 169 * 170 */ 171 172#include <sys/param.h> 173#include <sys/systm.h> 174#include <sys/conf.h> 175#include <sys/kernel.h> 176#include <sys/lock.h> 177#include <sys/malloc.h> 178#include <sys/mount.h> 179#include <sys/mutex.h> 180#include <sys/namei.h> 181#include <sys/sysctl.h> 182#include <sys/vnode.h> 183 184#include <fs/nullfs/null.h> 185 186#include <vm/vm.h> 187#include <vm/vm_extern.h> 188#include <vm/vm_object.h> 189#include <vm/vnode_pager.h> 190 191static int null_bug_bypass = 0; /* for debugging: enables bypass printf'ing */ 192SYSCTL_INT(_debug, OID_AUTO, nullfs_bug_bypass, CTLFLAG_RW, 193 &null_bug_bypass, 0, ""); 194 195static vop_access_t null_access; 196static vop_createvobject_t null_createvobject; 197static vop_destroyvobject_t null_destroyvobject; 198static vop_getattr_t null_getattr; 199static vop_getvobject_t null_getvobject; 200static vop_inactive_t null_inactive; 201static vop_islocked_t null_islocked; 202static vop_lock_t null_lock; 203static vop_lookup_t null_lookup; 204static vop_print_t null_print; 205static vop_reclaim_t null_reclaim; 206static vop_rename_t null_rename; 207static vop_setattr_t null_setattr; 208static vop_unlock_t null_unlock; 209 210/* 211 * This is the 10-Apr-92 bypass routine. 212 * This version has been optimized for speed, throwing away some 213 * safety checks. It should still always work, but it's not as 214 * robust to programmer errors. 215 * 216 * In general, we map all vnodes going down and unmap them on the way back. 217 * As an exception to this, vnodes can be marked "unmapped" by setting 218 * the Nth bit in operation's vdesc_flags. 219 * 220 * Also, some BSD vnode operations have the side effect of vrele'ing 221 * their arguments. With stacking, the reference counts are held 222 * by the upper node, not the lower one, so we must handle these 223 * side-effects here. This is not of concern in Sun-derived systems 224 * since there are no such side-effects. 225 * 226 * This makes the following assumptions: 227 * - only one returned vpp 228 * - no INOUT vpp's (Sun's vop_open has one of these) 229 * - the vnode operation vector of the first vnode should be used 230 * to determine what implementation of the op should be invoked 231 * - all mapped vnodes are of our vnode-type (NEEDSWORK: 232 * problems on rmdir'ing mount points and renaming?) 233 */ 234int 235null_bypass(ap) 236 struct vop_generic_args /* { 237 struct vnodeop_desc *a_desc; 238 <other random data follows, presumably> 239 } */ *ap; 240{ 241 register struct vnode **this_vp_p; 242 int error; 243 struct vnode *old_vps[VDESC_MAX_VPS]; 244 struct vnode **vps_p[VDESC_MAX_VPS]; 245 struct vnode ***vppp; 246 struct vnodeop_desc *descp = ap->a_desc; 247 int reles, i; 248 249 if (null_bug_bypass) 250 printf ("null_bypass: %s\n", descp->vdesc_name); 251 252#ifdef DIAGNOSTIC 253 /* 254 * We require at least one vp. 255 */ 256 if (descp->vdesc_vp_offsets == NULL || 257 descp->vdesc_vp_offsets[0] == VDESC_NO_OFFSET) 258 panic ("null_bypass: no vp's in map"); 259#endif 260 261 /* 262 * Map the vnodes going in. 263 * Later, we'll invoke the operation based on 264 * the first mapped vnode's operation vector. 265 */ 266 reles = descp->vdesc_flags; 267 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) { 268 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET) 269 break; /* bail out at end of list */ 270 vps_p[i] = this_vp_p = 271 VOPARG_OFFSETTO(struct vnode**,descp->vdesc_vp_offsets[i],ap); 272 /* 273 * We're not guaranteed that any but the first vnode 274 * are of our type. Check for and don't map any 275 * that aren't. (We must always map first vp or vclean fails.) 276 */ 277 if (i && (*this_vp_p == NULLVP || 278 (*this_vp_p)->v_op != &null_vnodeops)) { 279 old_vps[i] = NULLVP; 280 } else { 281 old_vps[i] = *this_vp_p; 282 *(vps_p[i]) = NULLVPTOLOWERVP(*this_vp_p); 283 /* 284 * XXX - Several operations have the side effect 285 * of vrele'ing their vp's. We must account for 286 * that. (This should go away in the future.) 287 */ 288 if (reles & VDESC_VP0_WILLRELE) 289 VREF(*this_vp_p); 290 } 291 292 } 293 294 /* 295 * Call the operation on the lower layer 296 * with the modified argument structure. 297 */ 298 if (vps_p[0] && *vps_p[0]) 299 error = VCALL(ap); 300 else { 301 printf("null_bypass: no map for %s\n", descp->vdesc_name); 302 error = EINVAL; 303 } 304 305 /* 306 * Maintain the illusion of call-by-value 307 * by restoring vnodes in the argument structure 308 * to their original value. 309 */ 310 reles = descp->vdesc_flags; 311 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) { 312 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET) 313 break; /* bail out at end of list */ 314 if (old_vps[i]) { 315 *(vps_p[i]) = old_vps[i]; 316#if 0 317 if (reles & VDESC_VP0_WILLUNLOCK) 318 VOP_UNLOCK(*(vps_p[i]), LK_THISLAYER, curthread); 319#endif 320 if (reles & VDESC_VP0_WILLRELE) 321 vrele(*(vps_p[i])); 322 } 323 } 324 325 /* 326 * Map the possible out-going vpp 327 * (Assumes that the lower layer always returns 328 * a VREF'ed vpp unless it gets an error.) 329 */ 330 if (descp->vdesc_vpp_offset != VDESC_NO_OFFSET && 331 !(descp->vdesc_flags & VDESC_NOMAP_VPP) && 332 !error) { 333 /* 334 * XXX - even though some ops have vpp returned vp's, 335 * several ops actually vrele this before returning. 336 * We must avoid these ops. 337 * (This should go away when these ops are regularized.) 338 */ 339 if (descp->vdesc_flags & VDESC_VPP_WILLRELE) 340 goto out; 341 vppp = VOPARG_OFFSETTO(struct vnode***, 342 descp->vdesc_vpp_offset,ap); 343 if (*vppp) 344 error = null_nodeget(old_vps[0]->v_mount, **vppp, *vppp); 345 } 346 347 out: 348 return (error); 349} 350 351/* 352 * We have to carry on the locking protocol on the null layer vnodes 353 * as we progress through the tree. We also have to enforce read-only 354 * if this layer is mounted read-only. 355 */ 356static int 357null_lookup(ap) 358 struct vop_lookup_args /* { 359 struct vnode * a_dvp; 360 struct vnode ** a_vpp; 361 struct componentname * a_cnp; 362 } */ *ap; 363{ 364 struct componentname *cnp = ap->a_cnp; 365 struct vnode *dvp = ap->a_dvp; 366 struct thread *td = cnp->cn_thread; 367 int flags = cnp->cn_flags; 368 struct vnode *vp, *ldvp, *lvp; 369 int error; 370 371 if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) && 372 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) 373 return (EROFS); 374 /* 375 * Although it is possible to call null_bypass(), we'll do 376 * a direct call to reduce overhead 377 */ 378 ldvp = NULLVPTOLOWERVP(dvp); 379 vp = lvp = NULL; 380 error = VOP_LOOKUP(ldvp, &lvp, cnp); 381 if (error == EJUSTRETURN && (flags & ISLASTCN) && 382 (dvp->v_mount->mnt_flag & MNT_RDONLY) && 383 (cnp->cn_nameiop == CREATE || cnp->cn_nameiop == RENAME)) 384 error = EROFS; 385 386 /* 387 * Rely only on the PDIRUNLOCK flag which should be carefully 388 * tracked by underlying filesystem. 389 */ 390 if ((cnp->cn_flags & PDIRUNLOCK) && dvp->v_vnlock != ldvp->v_vnlock) 391 VOP_UNLOCK(dvp, LK_THISLAYER, td); 392 if ((error == 0 || error == EJUSTRETURN) && lvp != NULL) { 393 if (ldvp == lvp) { 394 *ap->a_vpp = dvp; 395 VREF(dvp); 396 vrele(lvp); 397 } else { 398 error = null_nodeget(dvp->v_mount, lvp, &vp); 399 if (error) { 400 /* XXX Cleanup needed... */ 401 panic("null_nodeget failed"); 402 } 403 *ap->a_vpp = vp; 404 } 405 } 406 return (error); 407} 408 409/* 410 * Setattr call. Disallow write attempts if the layer is mounted read-only. 411 */ 412static int 413null_setattr(ap) 414 struct vop_setattr_args /* { 415 struct vnodeop_desc *a_desc; 416 struct vnode *a_vp; 417 struct vattr *a_vap; 418 struct ucred *a_cred; 419 struct thread *a_td; 420 } */ *ap; 421{ 422 struct vnode *vp = ap->a_vp; 423 struct vattr *vap = ap->a_vap; 424 425 if ((vap->va_flags != VNOVAL || vap->va_uid != (uid_t)VNOVAL || 426 vap->va_gid != (gid_t)VNOVAL || vap->va_atime.tv_sec != VNOVAL || 427 vap->va_mtime.tv_sec != VNOVAL || vap->va_mode != (mode_t)VNOVAL) && 428 (vp->v_mount->mnt_flag & MNT_RDONLY)) 429 return (EROFS); 430 if (vap->va_size != VNOVAL) { 431 switch (vp->v_type) { 432 case VDIR: 433 return (EISDIR); 434 case VCHR: 435 case VBLK: 436 case VSOCK: 437 case VFIFO: 438 if (vap->va_flags != VNOVAL) 439 return (EOPNOTSUPP); 440 return (0); 441 case VREG: 442 case VLNK: 443 default: 444 /* 445 * Disallow write attempts if the filesystem is 446 * mounted read-only. 447 */ 448 if (vp->v_mount->mnt_flag & MNT_RDONLY) 449 return (EROFS); 450 } 451 } 452 453 return (null_bypass((struct vop_generic_args *)ap)); 454} 455 456/* 457 * We handle getattr only to change the fsid. 458 */ 459static int 460null_getattr(ap) 461 struct vop_getattr_args /* { 462 struct vnode *a_vp; 463 struct vattr *a_vap; 464 struct ucred *a_cred; 465 struct thread *a_td; 466 } */ *ap; 467{ 468 int error; 469 470 if ((error = null_bypass((struct vop_generic_args *)ap)) != 0) 471 return (error); 472 473 ap->a_vap->va_fsid = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0]; 474 return (0); 475} 476 477/* 478 * Handle to disallow write access if mounted read-only. 479 */ 480static int 481null_access(ap) 482 struct vop_access_args /* { 483 struct vnode *a_vp; 484 int a_mode; 485 struct ucred *a_cred; 486 struct thread *a_td; 487 } */ *ap; 488{ 489 struct vnode *vp = ap->a_vp; 490 mode_t mode = ap->a_mode; 491 492 /* 493 * Disallow write attempts on read-only layers; 494 * unless the file is a socket, fifo, or a block or 495 * character device resident on the filesystem. 496 */ 497 if (mode & VWRITE) { 498 switch (vp->v_type) { 499 case VDIR: 500 case VLNK: 501 case VREG: 502 if (vp->v_mount->mnt_flag & MNT_RDONLY) 503 return (EROFS); 504 break; 505 default: 506 break; 507 } 508 } 509 return (null_bypass((struct vop_generic_args *)ap)); 510} 511 512/* 513 * We handle this to eliminate null FS to lower FS 514 * file moving. Don't know why we don't allow this, 515 * possibly we should. 516 */ 517static int 518null_rename(ap) 519 struct vop_rename_args /* { 520 struct vnode *a_fdvp; 521 struct vnode *a_fvp; 522 struct componentname *a_fcnp; 523 struct vnode *a_tdvp; 524 struct vnode *a_tvp; 525 struct componentname *a_tcnp; 526 } */ *ap; 527{ 528 struct vnode *tdvp = ap->a_tdvp; 529 struct vnode *fvp = ap->a_fvp; 530 struct vnode *fdvp = ap->a_fdvp; 531 struct vnode *tvp = ap->a_tvp; 532 533 /* Check for cross-device rename. */ 534 if ((fvp->v_mount != tdvp->v_mount) || 535 (tvp && (fvp->v_mount != tvp->v_mount))) { 536 if (tdvp == tvp) 537 vrele(tdvp); 538 else 539 vput(tdvp); 540 if (tvp) 541 vput(tvp); 542 vrele(fdvp); 543 vrele(fvp); 544 return (EXDEV); 545 } 546 547 return (null_bypass((struct vop_generic_args *)ap)); 548} 549 550/* 551 * We need to process our own vnode lock and then clear the 552 * interlock flag as it applies only to our vnode, not the 553 * vnodes below us on the stack. 554 */ 555static int 556null_lock(ap) 557 struct vop_lock_args /* { 558 struct vnode *a_vp; 559 int a_flags; 560 struct thread *a_td; 561 } */ *ap; 562{ 563 struct vnode *vp = ap->a_vp; 564 int flags = ap->a_flags; 565 struct thread *td = ap->a_td; 566 struct vnode *lvp; 567 int error; 568 struct null_node *nn; 569 570 if (flags & LK_THISLAYER) { 571 if (vp->v_vnlock != NULL) { 572 /* lock is shared across layers */ 573 if (flags & LK_INTERLOCK) 574 mtx_unlock(&vp->v_interlock); 575 return 0; 576 } 577 error = lockmgr(&vp->v_lock, flags & ~LK_THISLAYER, 578 &vp->v_interlock, td); 579 return (error); 580 } 581 582 if (vp->v_vnlock != NULL) { 583 /* 584 * The lower level has exported a struct lock to us. Use 585 * it so that all vnodes in the stack lock and unlock 586 * simultaneously. Note: we don't DRAIN the lock as DRAIN 587 * decommissions the lock - just because our vnode is 588 * going away doesn't mean the struct lock below us is. 589 * LK_EXCLUSIVE is fine. 590 */ 591 if ((flags & LK_INTERLOCK) == 0) { 592 VI_LOCK(vp); 593 flags |= LK_INTERLOCK; 594 } 595 nn = VTONULL(vp); 596 if ((flags & LK_TYPE_MASK) == LK_DRAIN) { 597 NULLFSDEBUG("null_lock: avoiding LK_DRAIN\n"); 598 /* 599 * Emulate lock draining by waiting for all other 600 * pending locks to complete. Afterwards the 601 * lockmgr call might block, but no other threads 602 * will attempt to use this nullfs vnode due to the 603 * VI_XLOCK flag. 604 */ 605 while (nn->null_pending_locks > 0) { 606 nn->null_drain_wakeup = 1; 607 msleep(&nn->null_pending_locks, 608 VI_MTX(vp), 609 PVFS, 610 "nuldr", 0); 611 } 612 error = lockmgr(vp->v_vnlock, 613 (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE, 614 VI_MTX(vp), td); 615 return error; 616 } 617 nn->null_pending_locks++; 618 error = lockmgr(vp->v_vnlock, flags, &vp->v_interlock, td); 619 VI_LOCK(vp); 620 /* 621 * If we're called from vrele then v_usecount can have been 0 622 * and another process might have initiated a recycle 623 * operation. When that happens, just back out. 624 */ 625 if (error == 0 && (vp->v_iflag & VI_XLOCK) != 0 && 626 td != vp->v_vxthread) { 627 lockmgr(vp->v_vnlock, 628 (flags & ~LK_TYPE_MASK) | LK_RELEASE, 629 VI_MTX(vp), td); 630 VI_LOCK(vp); 631 error = ENOENT; 632 } 633 nn->null_pending_locks--; 634 /* 635 * Wakeup the process draining the vnode after all 636 * pending lock attempts has been failed. 637 */ 638 if (nn->null_pending_locks == 0 && 639 nn->null_drain_wakeup != 0) { 640 nn->null_drain_wakeup = 0; 641 wakeup(&nn->null_pending_locks); 642 } 643 if (error == ENOENT && (vp->v_iflag & VI_XLOCK) != 0 && 644 vp->v_vxthread != curthread) { 645 vp->v_iflag |= VI_XWANT; 646 msleep(vp, VI_MTX(vp), PINOD, "nulbo", 0); 647 } 648 VI_UNLOCK(vp); 649 return error; 650 } else { 651 /* 652 * To prevent race conditions involving doing a lookup 653 * on "..", we have to lock the lower node, then lock our 654 * node. Most of the time it won't matter that we lock our 655 * node (as any locking would need the lower one locked 656 * first). But we can LK_DRAIN the upper lock as a step 657 * towards decomissioning it. 658 */ 659 lvp = NULLVPTOLOWERVP(vp); 660 if (lvp == NULL) 661 return (lockmgr(&vp->v_lock, flags, &vp->v_interlock, td)); 662 if (flags & LK_INTERLOCK) { 663 mtx_unlock(&vp->v_interlock); 664 flags &= ~LK_INTERLOCK; 665 } 666 if ((flags & LK_TYPE_MASK) == LK_DRAIN) { 667 error = VOP_LOCK(lvp, 668 (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE, td); 669 } else 670 error = VOP_LOCK(lvp, flags, td); 671 if (error) 672 return (error); 673 error = lockmgr(&vp->v_lock, flags, &vp->v_interlock, td); 674 if (error) 675 VOP_UNLOCK(lvp, 0, td); 676 return (error); 677 } 678} 679 680/* 681 * We need to process our own vnode unlock and then clear the 682 * interlock flag as it applies only to our vnode, not the 683 * vnodes below us on the stack. 684 */ 685static int 686null_unlock(ap) 687 struct vop_unlock_args /* { 688 struct vnode *a_vp; 689 int a_flags; 690 struct thread *a_td; 691 } */ *ap; 692{ 693 struct vnode *vp = ap->a_vp; 694 int flags = ap->a_flags; 695 struct thread *td = ap->a_td; 696 struct vnode *lvp; 697 698 if (vp->v_vnlock != NULL) { 699 if (flags & LK_THISLAYER) 700 return 0; /* the lock is shared across layers */ 701 flags &= ~LK_THISLAYER; 702 return (lockmgr(vp->v_vnlock, flags | LK_RELEASE, 703 &vp->v_interlock, td)); 704 } 705 lvp = NULLVPTOLOWERVP(vp); 706 if (lvp == NULL) 707 return (lockmgr(&vp->v_lock, flags | LK_RELEASE, &vp->v_interlock, td)); 708 if ((flags & LK_THISLAYER) == 0) { 709 if (flags & LK_INTERLOCK) { 710 mtx_unlock(&vp->v_interlock); 711 flags &= ~LK_INTERLOCK; 712 } 713 VOP_UNLOCK(lvp, flags & ~LK_INTERLOCK, td); 714 } else 715 flags &= ~LK_THISLAYER; 716 return (lockmgr(&vp->v_lock, flags | LK_RELEASE, &vp->v_interlock, td)); 717} 718 719static int 720null_islocked(ap) 721 struct vop_islocked_args /* { 722 struct vnode *a_vp; 723 struct thread *a_td; 724 } */ *ap; 725{ 726 struct vnode *vp = ap->a_vp; 727 struct thread *td = ap->a_td; 728 729 if (vp->v_vnlock != NULL) 730 return (lockstatus(vp->v_vnlock, td)); 731 return (lockstatus(&vp->v_lock, td)); 732} 733 734/* 735 * There is no way to tell that someone issued remove/rmdir operation 736 * on the underlying filesystem. For now we just have to release lowevrp 737 * as soon as possible. 738 * 739 * Note, we can't release any resources nor remove vnode from hash before 740 * appropriate VXLOCK stuff is is done because other process can find this 741 * vnode in hash during inactivation and may be sitting in vget() and waiting 742 * for null_inactive to unlock vnode. Thus we will do all those in VOP_RECLAIM. 743 */ 744static int 745null_inactive(ap) 746 struct vop_inactive_args /* { 747 struct vnode *a_vp; 748 struct thread *a_td; 749 } */ *ap; 750{ 751 struct vnode *vp = ap->a_vp; 752 struct thread *td = ap->a_td; 753 754 VOP_UNLOCK(vp, 0, td); 755 756 /* 757 * If this is the last reference, then free up the vnode 758 * so as not to tie up the lower vnodes. 759 */ 760 vrecycle(vp, NULL, td); 761 762 return (0); 763} 764 765/* 766 * Now, the VXLOCK is in force and we're free to destroy the null vnode. 767 */ 768static int 769null_reclaim(ap) 770 struct vop_reclaim_args /* { 771 struct vnode *a_vp; 772 struct thread *a_td; 773 } */ *ap; 774{ 775 struct vnode *vp = ap->a_vp; 776 struct null_node *xp = VTONULL(vp); 777 struct vnode *lowervp = xp->null_lowervp; 778 779 if (lowervp) { 780 null_hashrem(xp); 781 782 vrele(lowervp); 783 vrele(lowervp); 784 } 785 786 vp->v_data = NULL; 787 vp->v_vnlock = &vp->v_lock; 788 FREE(xp, M_NULLFSNODE); 789 790 return (0); 791} 792 793static int 794null_print(ap) 795 struct vop_print_args /* { 796 struct vnode *a_vp; 797 } */ *ap; 798{ 799 register struct vnode *vp = ap->a_vp; 800 printf("\tvp=%p, lowervp=%p\n", vp, NULLVPTOLOWERVP(vp)); 801 return (0); 802} 803 804/* 805 * Let an underlying filesystem do the work 806 */ 807static int 808null_createvobject(ap) 809 struct vop_createvobject_args /* { 810 struct vnode *vp; 811 struct ucred *cred; 812 struct thread *td; 813 } */ *ap; 814{ 815 struct vnode *vp = ap->a_vp; 816 struct vnode *lowervp = VTONULL(vp) ? NULLVPTOLOWERVP(vp) : NULL; 817 int error; 818 819 if (vp->v_type == VNON || lowervp == NULL) 820 return 0; 821 error = VOP_CREATEVOBJECT(lowervp, ap->a_cred, ap->a_td); 822 if (error) 823 return (error); 824 vp->v_vflag |= VV_OBJBUF; 825 return (0); 826} 827 828/* 829 * We have nothing to destroy and this operation shouldn't be bypassed. 830 */ 831static int 832null_destroyvobject(ap) 833 struct vop_destroyvobject_args /* { 834 struct vnode *vp; 835 } */ *ap; 836{ 837 struct vnode *vp = ap->a_vp; 838 839 vp->v_vflag &= ~VV_OBJBUF; 840 return (0); 841} 842 843static int 844null_getvobject(ap) 845 struct vop_getvobject_args /* { 846 struct vnode *vp; 847 struct vm_object **objpp; 848 } */ *ap; 849{ 850 struct vnode *lvp = NULLVPTOLOWERVP(ap->a_vp); 851 852 if (lvp == NULL) 853 return EINVAL; 854 return (VOP_GETVOBJECT(lvp, ap->a_objpp)); 855} 856 857/* 858 * Global vfs data structures 859 */ 860struct vop_vector null_vnodeops = { 861 .vop_bypass = null_bypass, 862 863 .vop_access = null_access, 864 .vop_bmap = VOP_EOPNOTSUPP, 865 .vop_createvobject = null_createvobject, 866 .vop_destroyvobject = null_destroyvobject, 867 .vop_getattr = null_getattr, 868 .vop_getvobject = null_getvobject, 869 .vop_getwritemount = vop_stdgetwritemount, 870 .vop_inactive = null_inactive, 871 .vop_islocked = null_islocked, 872 .vop_lock = null_lock, 873 .vop_lookup = null_lookup, 874 .vop_print = null_print, 875 .vop_reclaim = null_reclaim, 876 .vop_rename = null_rename, 877 .vop_setattr = null_setattr, 878 .vop_strategy = VOP_EOPNOTSUPP, 879 .vop_unlock = null_unlock, 880}; 881