| 1 | /* |
| 2 | * Copyright (c) 2004 The DragonFly Project. All rights reserved. |
| 3 | * |
| 4 | * This code is derived from software contributed to The DragonFly Project |
| 5 | * by Matthew Dillon <dillon@backplane.com> |
| 6 | * |
| 7 | * Redistribution and use in source and binary forms, with or without |
| 8 | * modification, are permitted provided that the following conditions |
| 9 | * are met: |
| 10 | * |
| 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 |
| 15 | * the documentation and/or other materials provided with the |
| 16 | * distribution. |
| 17 | * 3. Neither the name of The DragonFly Project nor the names of its |
| 18 | * contributors may be used to endorse or promote products derived |
| 19 | * from this software without specific, prior written permission. |
| 20 | * |
| 21 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
| 22 | * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
| 23 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS |
| 24 | * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE |
| 25 | * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, |
| 26 | * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, |
| 27 | * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; |
| 28 | * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED |
| 29 | * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, |
| 30 | * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT |
| 31 | * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
| 32 | * SUCH DAMAGE. |
| 33 | * |
| 34 | * $DragonFly: src/sys/kern/vfs_lock.c,v 1.30 2008/06/30 03:57:41 dillon Exp $ |
| 35 | */ |
| 36 | |
| 37 | /* |
| 38 | * External virtual filesystem routines |
| 39 | */ |
| 40 | #include "opt_ddb.h" |
| 41 | |
| 42 | #include <sys/param.h> |
| 43 | #include <sys/systm.h> |
| 44 | #include <sys/kernel.h> |
| 45 | #include <sys/malloc.h> |
| 46 | #include <sys/mount.h> |
| 47 | #include <sys/proc.h> |
| 48 | #include <sys/vnode.h> |
| 49 | #include <sys/buf.h> |
| 50 | #include <sys/sysctl.h> |
| 51 | |
| 52 | #include <machine/limits.h> |
| 53 | |
| 54 | #include <vm/vm.h> |
| 55 | #include <vm/vm_object.h> |
| 56 | |
| 57 | #include <sys/buf2.h> |
| 58 | #include <sys/thread2.h> |
| 59 | #include <sys/sysref2.h> |
| 60 | |
| 61 | static void vnode_terminate(struct vnode *vp); |
| 62 | static boolean_t vnode_ctor(void *obj, void *private, int ocflags); |
| 63 | static void vnode_dtor(void *obj, void *private); |
| 64 | |
| 65 | static MALLOC_DEFINE(M_VNODE, "vnodes", "vnode structures"); |
| 66 | static struct sysref_class vnode_sysref_class = { |
| 67 | .name = "vnode", |
| 68 | .mtype = M_VNODE, |
| 69 | .proto = SYSREF_PROTO_VNODE, |
| 70 | .offset = offsetof(struct vnode, v_sysref), |
| 71 | .objsize = sizeof(struct vnode), |
| 72 | .mag_capacity = 256, |
| 73 | .flags = SRC_MANAGEDINIT, |
| 74 | .ctor = vnode_ctor, |
| 75 | .dtor = vnode_dtor, |
| 76 | .ops = { |
| 77 | .terminate = (sysref_terminate_func_t)vnode_terminate |
| 78 | } |
| 79 | }; |
| 80 | |
| 81 | /* |
| 82 | * The vnode free list hold inactive vnodes. Aged inactive vnodes |
| 83 | * are inserted prior to the mid point, and otherwise inserted |
| 84 | * at the tail. |
| 85 | */ |
| 86 | static TAILQ_HEAD(freelst, vnode) vnode_free_list; |
| 87 | static struct vnode vnode_free_mid; |
| 88 | static struct spinlock vfs_spin = SPINLOCK_INITIALIZER(vfs_spin); |
| 89 | |
| 90 | int freevnodes = 0; |
| 91 | SYSCTL_INT(_debug, OID_AUTO, freevnodes, CTLFLAG_RD, |
| 92 | &freevnodes, 0, ""); |
| 93 | static int wantfreevnodes = 25; |
| 94 | SYSCTL_INT(_debug, OID_AUTO, wantfreevnodes, CTLFLAG_RW, |
| 95 | &wantfreevnodes, 0, ""); |
| 96 | #ifdef TRACKVNODE |
| 97 | static ulong trackvnode; |
| 98 | SYSCTL_ULONG(_debug, OID_AUTO, trackvnode, CTLFLAG_RW, |
| 99 | &trackvnode, 0, ""); |
| 100 | #endif |
| 101 | |
| 102 | /* |
| 103 | * Called from vfsinit() |
| 104 | */ |
| 105 | void |
| 106 | vfs_lock_init(void) |
| 107 | { |
| 108 | TAILQ_INIT(&vnode_free_list); |
| 109 | TAILQ_INSERT_HEAD(&vnode_free_list, &vnode_free_mid, v_freelist); |
| 110 | spin_init(&vfs_spin); |
| 111 | } |
| 112 | |
| 113 | /* |
| 114 | * Misc functions |
| 115 | */ |
| 116 | static __inline |
| 117 | void |
| 118 | _vsetflags(struct vnode *vp, int flags) |
| 119 | { |
| 120 | atomic_set_int(&vp->v_flag, flags); |
| 121 | } |
| 122 | |
| 123 | static __inline |
| 124 | void |
| 125 | _vclrflags(struct vnode *vp, int flags) |
| 126 | { |
| 127 | atomic_clear_int(&vp->v_flag, flags); |
| 128 | } |
| 129 | |
| 130 | void |
| 131 | vsetflags(struct vnode *vp, int flags) |
| 132 | { |
| 133 | _vsetflags(vp, flags); |
| 134 | } |
| 135 | |
| 136 | void |
| 137 | vclrflags(struct vnode *vp, int flags) |
| 138 | { |
| 139 | _vclrflags(vp, flags); |
| 140 | } |
| 141 | |
| 142 | /* |
| 143 | * Inline helper functions. vbusy() and vfree() must be called while |
| 144 | * vp->v_spinlock is held. |
| 145 | * |
| 146 | * WARNING! This functions is typically called with v_spinlock held. |
| 147 | * |
| 148 | * MPSAFE |
| 149 | */ |
| 150 | static __inline |
| 151 | void |
| 152 | __vbusy(struct vnode *vp) |
| 153 | { |
| 154 | #ifdef TRACKVNODE |
| 155 | if ((ulong)vp == trackvnode) |
| 156 | kprintf("__vbusy %p %08x\n", vp, vp->v_flag); |
| 157 | #endif |
| 158 | spin_lock_wr(&vfs_spin); |
| 159 | TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); |
| 160 | freevnodes--; |
| 161 | _vclrflags(vp, VFREE); |
| 162 | spin_unlock_wr(&vfs_spin); |
| 163 | } |
| 164 | |
| 165 | /* |
| 166 | * WARNING! This functions is typically called with v_spinlock held. |
| 167 | * |
| 168 | * MPSAFE |
| 169 | */ |
| 170 | static __inline |
| 171 | void |
| 172 | __vfree(struct vnode *vp) |
| 173 | { |
| 174 | #ifdef TRACKVNODE |
| 175 | if ((ulong)vp == trackvnode) { |
| 176 | kprintf("__vfree %p %08x\n", vp, vp->v_flag); |
| 177 | print_backtrace(); |
| 178 | } |
| 179 | #endif |
| 180 | spin_lock_wr(&vfs_spin); |
| 181 | if (vp->v_flag & VRECLAIMED) |
| 182 | TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist); |
| 183 | else if (vp->v_flag & (VAGE0 | VAGE1)) |
| 184 | TAILQ_INSERT_BEFORE(&vnode_free_mid, vp, v_freelist); |
| 185 | else |
| 186 | TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); |
| 187 | freevnodes++; |
| 188 | _vsetflags(vp, VFREE); |
| 189 | spin_unlock_wr(&vfs_spin); |
| 190 | } |
| 191 | |
| 192 | /* |
| 193 | * WARNING! This functions is typically called with v_spinlock held. |
| 194 | * |
| 195 | * MPSAFE |
| 196 | */ |
| 197 | static __inline |
| 198 | void |
| 199 | __vfreetail(struct vnode *vp) |
| 200 | { |
| 201 | #ifdef TRACKVNODE |
| 202 | if ((ulong)vp == trackvnode) |
| 203 | kprintf("__vfreetail %p %08x\n", vp, vp->v_flag); |
| 204 | #endif |
| 205 | spin_lock_wr(&vfs_spin); |
| 206 | TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist); |
| 207 | freevnodes++; |
| 208 | _vsetflags(vp, VFREE); |
| 209 | spin_unlock_wr(&vfs_spin); |
| 210 | } |
| 211 | |
| 212 | /* |
| 213 | * Return a C boolean if we should put the vnode on the freelist (VFREE), |
| 214 | * or leave it / mark it as VCACHED. |
| 215 | * |
| 216 | * This routine is only valid if the vnode is already either VFREE or |
| 217 | * VCACHED, or if it can become VFREE or VCACHED via vnode_terminate(). |
| 218 | * |
| 219 | * WARNING! This functions is typically called with v_spinlock held. |
| 220 | * |
| 221 | * MPSAFE |
| 222 | */ |
| 223 | static __inline boolean_t |
| 224 | vshouldfree(struct vnode *vp) |
| 225 | { |
| 226 | return (vp->v_auxrefs == 0 && |
| 227 | (vp->v_object == NULL || vp->v_object->resident_page_count == 0)); |
| 228 | } |
| 229 | |
| 230 | /* |
| 231 | * Add a ref to an active vnode. This function should never be called |
| 232 | * with an inactive vnode (use vget() instead). |
| 233 | * |
| 234 | * MPSAFE |
| 235 | */ |
| 236 | void |
| 237 | vref(struct vnode *vp) |
| 238 | { |
| 239 | KKASSERT(vp->v_sysref.refcnt > 0 && |
| 240 | (vp->v_flag & (VFREE|VINACTIVE)) == 0); |
| 241 | sysref_get(&vp->v_sysref); |
| 242 | } |
| 243 | |
| 244 | /* |
| 245 | * Release a ref on an active or inactive vnode. The sysref termination |
| 246 | * function will be called when the active last active reference is released, |
| 247 | * and the vnode is returned to the objcache when the last inactive |
| 248 | * reference is released. |
| 249 | */ |
| 250 | void |
| 251 | vrele(struct vnode *vp) |
| 252 | { |
| 253 | sysref_put(&vp->v_sysref); |
| 254 | } |
| 255 | |
| 256 | /* |
| 257 | * Add an auxiliary data structure reference to the vnode. Auxiliary |
| 258 | * references do not change the state of the vnode or prevent them |
| 259 | * from being deactivated, reclaimed, or placed on the free list. |
| 260 | * |
| 261 | * An auxiliary reference DOES prevent the vnode from being destroyed, |
| 262 | * allowing you to vx_lock() it, test state, etc. |
| 263 | * |
| 264 | * An auxiliary reference DOES NOT move a vnode out of the VFREE state |
| 265 | * once it has entered it. |
| 266 | * |
| 267 | * MPSAFE |
| 268 | */ |
| 269 | void |
| 270 | vhold(struct vnode *vp) |
| 271 | { |
| 272 | KKASSERT(vp->v_sysref.refcnt != 0); |
| 273 | atomic_add_int(&vp->v_auxrefs, 1); |
| 274 | } |
| 275 | |
| 276 | /* |
| 277 | * Remove an auxiliary reference from the vnode. |
| 278 | * |
| 279 | * vdrop needs to check for a VCACHE->VFREE transition to catch cases |
| 280 | * where a vnode is held past its reclamation. |
| 281 | * |
| 282 | * MPSAFE |
| 283 | */ |
| 284 | void |
| 285 | vdrop(struct vnode *vp) |
| 286 | { |
| 287 | KKASSERT(vp->v_sysref.refcnt != 0 && vp->v_auxrefs > 0); |
| 288 | spin_lock_wr(&vp->v_spinlock); |
| 289 | atomic_subtract_int(&vp->v_auxrefs, 1); |
| 290 | if ((vp->v_flag & VCACHED) && vshouldfree(vp)) { |
| 291 | _vclrflags(vp, VCACHED); |
| 292 | __vfree(vp); |
| 293 | } |
| 294 | spin_unlock_wr(&vp->v_spinlock); |
| 295 | } |
| 296 | |
| 297 | /* |
| 298 | * This function is called when the last active reference on the vnode |
| 299 | * is released, typically via vrele(). SYSREF will give the vnode a |
| 300 | * negative ref count, indicating that it is undergoing termination or |
| 301 | * is being set aside for the cache, and one final sysref_put() is |
| 302 | * required to actually return it to the memory subsystem. |
| 303 | * |
| 304 | * However, because vnodes may have auxiliary structural references via |
| 305 | * v_auxrefs, we must interlock auxiliary references against termination |
| 306 | * via the VX lock mechanism. It is possible for a vnode to be reactivated |
| 307 | * while we were blocked on the lock. |
| 308 | * |
| 309 | * MPSAFE |
| 310 | */ |
| 311 | void |
| 312 | vnode_terminate(struct vnode *vp) |
| 313 | { |
| 314 | vx_lock(vp); |
| 315 | if (sysref_isinactive(&vp->v_sysref)) { |
| 316 | /* |
| 317 | * Deactivate the vnode by marking it VFREE or VCACHED. |
| 318 | * The vnode can be reactivated from either state until |
| 319 | * reclaimed. These states inherit the 'last' sysref on the |
| 320 | * vnode. |
| 321 | * |
| 322 | * NOTE: There may be additional inactive references from |
| 323 | * other entities blocking on the VX lock while we hold it, |
| 324 | * but this does not prevent us from changing the vnode's |
| 325 | * state. |
| 326 | * |
| 327 | * NOTE: The vnode could already be marked inactive. XXX |
| 328 | * how? |
| 329 | * |
| 330 | * NOTE: v_mount may be NULL due to assignment to |
| 331 | * dead_vnode_vops |
| 332 | * |
| 333 | * NOTE: The vnode may be marked inactive with dirty buffers |
| 334 | * or dirty pages in its cached VM object still present. |
| 335 | */ |
| 336 | if ((vp->v_flag & VINACTIVE) == 0) { |
| 337 | _vsetflags(vp, VINACTIVE); |
| 338 | if (vp->v_mount) |
| 339 | VOP_INACTIVE(vp); |
| 340 | } |
| 341 | spin_lock_wr(&vp->v_spinlock); |
| 342 | KKASSERT((vp->v_flag & (VFREE|VCACHED)) == 0); |
| 343 | if (vshouldfree(vp)) |
| 344 | __vfree(vp); |
| 345 | else |
| 346 | _vsetflags(vp, VCACHED); /* inactive but not yet free*/ |
| 347 | spin_unlock_wr(&vp->v_spinlock); |
| 348 | vx_unlock(vp); |
| 349 | } else { |
| 350 | /* |
| 351 | * Someone reactivated the vnode while were blocked on the |
| 352 | * VX lock. Release the VX lock and release the (now active) |
| 353 | * last reference which is no longer last. |
| 354 | */ |
| 355 | vx_unlock(vp); |
| 356 | vrele(vp); |
| 357 | } |
| 358 | } |
| 359 | |
| 360 | /* |
| 361 | * Physical vnode constructor / destructor. These are only executed on |
| 362 | * the backend of the objcache. They are NOT executed on every vnode |
| 363 | * allocation or deallocation. |
| 364 | * |
| 365 | * MPSAFE |
| 366 | */ |
| 367 | boolean_t |
| 368 | vnode_ctor(void *obj, void *private, int ocflags) |
| 369 | { |
| 370 | struct vnode *vp = obj; |
| 371 | |
| 372 | lwkt_token_init(&vp->v_token); |
| 373 | lockinit(&vp->v_lock, "vnode", 0, 0); |
| 374 | ccms_dataspace_init(&vp->v_ccms); |
| 375 | TAILQ_INIT(&vp->v_namecache); |
| 376 | RB_INIT(&vp->v_rbclean_tree); |
| 377 | RB_INIT(&vp->v_rbdirty_tree); |
| 378 | RB_INIT(&vp->v_rbhash_tree); |
| 379 | return(TRUE); |
| 380 | } |
| 381 | |
| 382 | /* |
| 383 | * MPSAFE |
| 384 | */ |
| 385 | void |
| 386 | vnode_dtor(void *obj, void *private) |
| 387 | { |
| 388 | struct vnode *vp = obj; |
| 389 | |
| 390 | ccms_dataspace_destroy(&vp->v_ccms); |
| 391 | } |
| 392 | |
| 393 | /**************************************************************** |
| 394 | * VX LOCKING FUNCTIONS * |
| 395 | **************************************************************** |
| 396 | * |
| 397 | * These functions lock vnodes for reclamation and deactivation related |
| 398 | * activities. The caller must already be holding some sort of reference |
| 399 | * on the vnode. |
| 400 | * |
| 401 | * MPSAFE |
| 402 | */ |
| 403 | void |
| 404 | vx_lock(struct vnode *vp) |
| 405 | { |
| 406 | lockmgr(&vp->v_lock, LK_EXCLUSIVE); |
| 407 | } |
| 408 | |
| 409 | static int |
| 410 | vx_lock_nonblock(struct vnode *vp) |
| 411 | { |
| 412 | return(lockmgr(&vp->v_lock, LK_EXCLUSIVE | LK_NOWAIT)); |
| 413 | } |
| 414 | |
| 415 | void |
| 416 | vx_unlock(struct vnode *vp) |
| 417 | { |
| 418 | lockmgr(&vp->v_lock, LK_RELEASE); |
| 419 | } |
| 420 | |
| 421 | /**************************************************************** |
| 422 | * VNODE ACQUISITION FUNCTIONS * |
| 423 | **************************************************************** |
| 424 | * |
| 425 | * These functions must be used when accessing a vnode via an auxiliary |
| 426 | * reference such as the namecache or free list, or when you wish to |
| 427 | * do a combo ref+lock sequence. |
| 428 | * |
| 429 | * These functions are MANDATORY for any code chain accessing a vnode |
| 430 | * whos activation state is not known. |
| 431 | * |
| 432 | * vget() can be called with LK_NOWAIT and will return EBUSY if the |
| 433 | * lock cannot be immediately acquired. |
| 434 | * |
| 435 | * vget()/vput() are used when reactivation is desired. |
| 436 | * |
| 437 | * vx_get() and vx_put() are used when reactivation is not desired. |
| 438 | */ |
| 439 | int |
| 440 | vget(struct vnode *vp, int flags) |
| 441 | { |
| 442 | int error; |
| 443 | |
| 444 | /* |
| 445 | * A lock type must be passed |
| 446 | */ |
| 447 | if ((flags & LK_TYPE_MASK) == 0) { |
| 448 | panic("vget() called with no lock specified!"); |
| 449 | /* NOT REACHED */ |
| 450 | } |
| 451 | |
| 452 | /* |
| 453 | * Reference the structure and then acquire the lock. 0->1 |
| 454 | * transitions and refs during termination are allowed here so |
| 455 | * call sysref directly. |
| 456 | */ |
| 457 | sysref_get(&vp->v_sysref); |
| 458 | if ((error = vn_lock(vp, flags)) != 0) { |
| 459 | /* |
| 460 | * The lock failed, undo and return an error. |
| 461 | */ |
| 462 | sysref_put(&vp->v_sysref); |
| 463 | } else if (vp->v_flag & VRECLAIMED) { |
| 464 | /* |
| 465 | * The node is being reclaimed and cannot be reactivated |
| 466 | * any more, undo and return ENOENT. |
| 467 | */ |
| 468 | vn_unlock(vp); |
| 469 | vrele(vp); |
| 470 | error = ENOENT; |
| 471 | } else { |
| 472 | /* |
| 473 | * If the vnode is marked VFREE or VCACHED it needs to be |
| 474 | * reactivated, otherwise it had better already be active. |
| 475 | * VINACTIVE must also be cleared. |
| 476 | * |
| 477 | * In the VFREE/VCACHED case we have to throw away the |
| 478 | * sysref that was earmarking those cases and preventing |
| 479 | * the vnode from being destroyed. Our sysref is still held. |
| 480 | */ |
| 481 | spin_lock_wr(&vp->v_spinlock); |
| 482 | if (vp->v_flag & VFREE) { |
| 483 | __vbusy(vp); |
| 484 | spin_unlock_wr(&vp->v_spinlock); |
| 485 | sysref_put(&vp->v_sysref); |
| 486 | sysref_activate(&vp->v_sysref); |
| 487 | } else if (vp->v_flag & VCACHED) { |
| 488 | _vclrflags(vp, VCACHED); |
| 489 | spin_unlock_wr(&vp->v_spinlock); |
| 490 | sysref_put(&vp->v_sysref); |
| 491 | sysref_activate(&vp->v_sysref); |
| 492 | } else { |
| 493 | spin_unlock_wr(&vp->v_spinlock); |
| 494 | if (sysref_isinactive(&vp->v_sysref)) { |
| 495 | sysref_activate(&vp->v_sysref); |
| 496 | kprintf("Warning vp %p reactivation race\n", |
| 497 | vp); |
| 498 | } |
| 499 | } |
| 500 | _vclrflags(vp, VINACTIVE); |
| 501 | error = 0; |
| 502 | } |
| 503 | return(error); |
| 504 | } |
| 505 | |
| 506 | /* |
| 507 | * MPSAFE |
| 508 | */ |
| 509 | void |
| 510 | vput(struct vnode *vp) |
| 511 | { |
| 512 | vn_unlock(vp); |
| 513 | vrele(vp); |
| 514 | } |
| 515 | |
| 516 | /* |
| 517 | * XXX The vx_*() locks should use auxrefs, not the main reference counter. |
| 518 | * |
| 519 | * MPSAFE |
| 520 | */ |
| 521 | void |
| 522 | vx_get(struct vnode *vp) |
| 523 | { |
| 524 | sysref_get(&vp->v_sysref); |
| 525 | lockmgr(&vp->v_lock, LK_EXCLUSIVE); |
| 526 | } |
| 527 | |
| 528 | /* |
| 529 | * MPSAFE |
| 530 | */ |
| 531 | int |
| 532 | vx_get_nonblock(struct vnode *vp) |
| 533 | { |
| 534 | int error; |
| 535 | |
| 536 | sysref_get(&vp->v_sysref); |
| 537 | error = lockmgr(&vp->v_lock, LK_EXCLUSIVE | LK_NOWAIT); |
| 538 | if (error) |
| 539 | sysref_put(&vp->v_sysref); |
| 540 | return(error); |
| 541 | } |
| 542 | |
| 543 | /* |
| 544 | * Relase a VX lock that also held a ref on the vnode. |
| 545 | * |
| 546 | * vx_put needs to check for a VCACHE->VFREE transition to catch the |
| 547 | * case where e.g. vnlru issues a vgone*(). |
| 548 | * |
| 549 | * MPSAFE |
| 550 | */ |
| 551 | void |
| 552 | vx_put(struct vnode *vp) |
| 553 | { |
| 554 | spin_lock_wr(&vp->v_spinlock); |
| 555 | if ((vp->v_flag & VCACHED) && vshouldfree(vp)) { |
| 556 | _vclrflags(vp, VCACHED); |
| 557 | __vfree(vp); |
| 558 | } |
| 559 | spin_unlock_wr(&vp->v_spinlock); |
| 560 | lockmgr(&vp->v_lock, LK_RELEASE); |
| 561 | sysref_put(&vp->v_sysref); |
| 562 | } |
| 563 | |
| 564 | /* |
| 565 | * Try to reuse a vnode from the free list. NOTE: The returned vnode |
| 566 | * is not completely initialized. |
| 567 | * |
| 568 | * MPSAFE |
| 569 | */ |
| 570 | static |
| 571 | struct vnode * |
| 572 | allocfreevnode(void) |
| 573 | { |
| 574 | struct vnode *vp; |
| 575 | int count; |
| 576 | |
| 577 | for (count = 0; count < freevnodes; count++) { |
| 578 | /* |
| 579 | * Note that regardless of how we block in this loop, |
| 580 | * we only get here if freevnodes != 0 so there |
| 581 | * had better be something on the list. |
| 582 | * |
| 583 | * Try to lock the first vnode on the free list. |
| 584 | * Cycle if we can't. |
| 585 | * |
| 586 | * XXX NOT MP SAFE |
| 587 | */ |
| 588 | spin_lock_wr(&vfs_spin); |
| 589 | vp = TAILQ_FIRST(&vnode_free_list); |
| 590 | if (vp == &vnode_free_mid) |
| 591 | vp = TAILQ_NEXT(vp, v_freelist); |
| 592 | if (vx_lock_nonblock(vp)) { |
| 593 | KKASSERT(vp->v_flag & VFREE); |
| 594 | TAILQ_REMOVE(&vnode_free_list, vp, v_freelist); |
| 595 | TAILQ_INSERT_TAIL(&vnode_free_list, |
| 596 | vp, v_freelist); |
| 597 | spin_unlock_wr(&vfs_spin); |
| 598 | continue; |
| 599 | } |
| 600 | spin_unlock_wr(&vfs_spin); |
| 601 | #ifdef TRACKVNODE |
| 602 | if ((ulong)vp == trackvnode) |
| 603 | kprintf("allocfreevnode %p %08x\n", vp, vp->v_flag); |
| 604 | #endif |
| 605 | |
| 606 | /* |
| 607 | * With the vnode locked we can safely remove it |
| 608 | * from the free list. We inherit the reference |
| 609 | * that was previously associated with the vnode |
| 610 | * being on the free list. |
| 611 | */ |
| 612 | KKASSERT((vp->v_flag & (VFREE|VINACTIVE)) == |
| 613 | (VFREE|VINACTIVE)); |
| 614 | KKASSERT(sysref_isinactive(&vp->v_sysref)); |
| 615 | __vbusy(vp); |
| 616 | |
| 617 | /* |
| 618 | * Holding the VX lock on an inactive vnode prevents it |
| 619 | * from being reactivated or reused. New namecache |
| 620 | * associations can only be made using active vnodes. |
| 621 | * |
| 622 | * Another thread may be blocked on our vnode lock while |
| 623 | * holding a namecache lock. We can only reuse this vnode |
| 624 | * if we can clear all namecache associations without |
| 625 | * blocking. |
| 626 | */ |
| 627 | if ((vp->v_flag & VRECLAIMED) == 0) { |
| 628 | if (cache_inval_vp_nonblock(vp)) { |
| 629 | __vfreetail(vp); |
| 630 | vx_unlock(vp); |
| 631 | continue; |
| 632 | } |
| 633 | vgone_vxlocked(vp); |
| 634 | /* vnode is still VX locked */ |
| 635 | } |
| 636 | |
| 637 | /* |
| 638 | * We can reuse the vnode if no primary or auxiliary |
| 639 | * references remain other then ours, else put it |
| 640 | * back on the free list and keep looking. |
| 641 | * |
| 642 | * Either the free list inherits the last reference |
| 643 | * or we fall through and sysref_activate() the last |
| 644 | * reference. |
| 645 | * |
| 646 | * Since the vnode is in a VRECLAIMED state, no new |
| 647 | * namecache associations could have been made. |
| 648 | */ |
| 649 | KKASSERT(TAILQ_EMPTY(&vp->v_namecache)); |
| 650 | if (vp->v_auxrefs || |
| 651 | !sysref_islastdeactivation(&vp->v_sysref)) { |
| 652 | __vfreetail(vp); |
| 653 | vx_unlock(vp); |
| 654 | continue; |
| 655 | } |
| 656 | |
| 657 | /* |
| 658 | * Return a VX locked vnode suitable for reuse. The caller |
| 659 | * inherits the sysref. |
| 660 | */ |
| 661 | return(vp); |
| 662 | } |
| 663 | return(NULL); |
| 664 | } |
| 665 | |
| 666 | /* |
| 667 | * Obtain a new vnode from the freelist, allocating more if necessary. |
| 668 | * The returned vnode is VX locked & refd. |
| 669 | * |
| 670 | * All new vnodes set the VAGE flags. An open() of the vnode will |
| 671 | * decrement the (2-bit) flags. Vnodes which are opened several times |
| 672 | * are thus retained in the cache over vnodes which are merely stat()d. |
| 673 | * |
| 674 | * MPSAFE |
| 675 | */ |
| 676 | struct vnode * |
| 677 | allocvnode(int lktimeout, int lkflags) |
| 678 | { |
| 679 | struct vnode *vp; |
| 680 | |
| 681 | /* |
| 682 | * Try to reuse vnodes if we hit the max. This situation only |
| 683 | * occurs in certain large-memory (2G+) situations. We cannot |
| 684 | * attempt to directly reclaim vnodes due to nasty recursion |
| 685 | * problems. |
| 686 | */ |
| 687 | while (numvnodes - freevnodes > desiredvnodes) |
| 688 | vnlru_proc_wait(); |
| 689 | |
| 690 | /* |
| 691 | * Try to build up as many vnodes as we can before reallocating |
| 692 | * from the free list. A vnode on the free list simply means |
| 693 | * that it is inactive with no resident pages. It may or may not |
| 694 | * have been reclaimed and could have valuable information associated |
| 695 | * with it that we shouldn't throw away unless we really need to. |
| 696 | * |
| 697 | * HAMMER NOTE: Re-establishing a vnode is a fairly expensive |
| 698 | * operation for HAMMER but this should benefit UFS as well. |
| 699 | */ |
| 700 | if (freevnodes >= wantfreevnodes && numvnodes >= desiredvnodes) |
| 701 | vp = allocfreevnode(); |
| 702 | else |
| 703 | vp = NULL; |
| 704 | if (vp == NULL) { |
| 705 | vp = sysref_alloc(&vnode_sysref_class); |
| 706 | lockmgr(&vp->v_lock, LK_EXCLUSIVE); |
| 707 | numvnodes++; |
| 708 | } |
| 709 | |
| 710 | /* |
| 711 | * We are using a managed sysref class, vnode fields are only |
| 712 | * zerod on initial allocation from the backing store, not |
| 713 | * on reallocation. Thus we have to clear these fields for both |
| 714 | * reallocation and reuse. |
| 715 | */ |
| 716 | #ifdef INVARIANTS |
| 717 | if (vp->v_data) |
| 718 | panic("cleaned vnode isn't"); |
| 719 | if (bio_track_active(&vp->v_track_read) || |
| 720 | bio_track_active(&vp->v_track_write)) { |
| 721 | panic("Clean vnode has pending I/O's"); |
| 722 | } |
| 723 | if (vp->v_flag & VONWORKLST) |
| 724 | panic("Clean vnode still pending on syncer worklist!"); |
| 725 | if (!RB_EMPTY(&vp->v_rbdirty_tree)) |
| 726 | panic("Clean vnode still has dirty buffers!"); |
| 727 | if (!RB_EMPTY(&vp->v_rbclean_tree)) |
| 728 | panic("Clean vnode still has clean buffers!"); |
| 729 | if (!RB_EMPTY(&vp->v_rbhash_tree)) |
| 730 | panic("Clean vnode still on hash tree!"); |
| 731 | KKASSERT(vp->v_mount == NULL); |
| 732 | #endif |
| 733 | vp->v_flag = VAGE0 | VAGE1; |
| 734 | vp->v_lastw = 0; |
| 735 | vp->v_lasta = 0; |
| 736 | vp->v_cstart = 0; |
| 737 | vp->v_clen = 0; |
| 738 | vp->v_socket = 0; |
| 739 | vp->v_opencount = 0; |
| 740 | vp->v_writecount = 0; /* XXX */ |
| 741 | |
| 742 | /* |
| 743 | * lktimeout only applies when LK_TIMELOCK is used, and only |
| 744 | * the pageout daemon uses it. The timeout may not be zero |
| 745 | * or the pageout daemon can deadlock in low-VM situations. |
| 746 | */ |
| 747 | if (lktimeout == 0) |
| 748 | lktimeout = hz / 10; |
| 749 | lockreinit(&vp->v_lock, "vnode", lktimeout, lkflags); |
| 750 | KKASSERT(TAILQ_EMPTY(&vp->v_namecache)); |
| 751 | /* exclusive lock still held */ |
| 752 | |
| 753 | /* |
| 754 | * Note: sysref needs to be activated to convert -0x40000000 to +1. |
| 755 | * The -0x40000000 comes from the last ref on reuse, and from |
| 756 | * sysref_init() on allocate. |
| 757 | */ |
| 758 | sysref_activate(&vp->v_sysref); |
| 759 | vp->v_filesize = NOOFFSET; |
| 760 | vp->v_type = VNON; |
| 761 | vp->v_tag = 0; |
| 762 | vp->v_ops = NULL; |
| 763 | vp->v_data = NULL; |
| 764 | KKASSERT(vp->v_mount == NULL); |
| 765 | |
| 766 | return (vp); |
| 767 | } |
| 768 | |
| 769 | /* |
| 770 | * MPSAFE |
| 771 | */ |
| 772 | int |
| 773 | freesomevnodes(int n) |
| 774 | { |
| 775 | struct vnode *vp; |
| 776 | int count = 0; |
| 777 | |
| 778 | while (n) { |
| 779 | --n; |
| 780 | if ((vp = allocfreevnode()) == NULL) |
| 781 | break; |
| 782 | vx_put(vp); |
| 783 | --numvnodes; |
| 784 | } |
| 785 | return(count); |
| 786 | } |
| 787 | |