| 1 | /* $NetBSD: vfs_cache.c,v 1.110 2016/07/07 06:55:43 msaitoh Exp $ */ |
| 2 | |
| 3 | /*- |
| 4 | * Copyright (c) 2008 The NetBSD Foundation, Inc. |
| 5 | * All rights reserved. |
| 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 | * 1. Redistributions of source code must retain the above copyright |
| 11 | * notice, this list of conditions and the following disclaimer. |
| 12 | * 2. Redistributions in binary form must reproduce the above copyright |
| 13 | * notice, this list of conditions and the following disclaimer in the |
| 14 | * documentation and/or other materials provided with the distribution. |
| 15 | * |
| 16 | * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS |
| 17 | * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED |
| 18 | * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR |
| 19 | * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS |
| 20 | * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR |
| 21 | * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF |
| 22 | * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS |
| 23 | * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN |
| 24 | * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) |
| 25 | * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE |
| 26 | * POSSIBILITY OF SUCH DAMAGE. |
| 27 | */ |
| 28 | |
| 29 | /* |
| 30 | * Copyright (c) 1989, 1993 |
| 31 | * The Regents of the University of California. All rights reserved. |
| 32 | * |
| 33 | * Redistribution and use in source and binary forms, with or without |
| 34 | * modification, are permitted provided that the following conditions |
| 35 | * are met: |
| 36 | * 1. Redistributions of source code must retain the above copyright |
| 37 | * notice, this list of conditions and the following disclaimer. |
| 38 | * 2. Redistributions in binary form must reproduce the above copyright |
| 39 | * notice, this list of conditions and the following disclaimer in the |
| 40 | * documentation and/or other materials provided with the distribution. |
| 41 | * 3. Neither the name of the University nor the names of its contributors |
| 42 | * may be used to endorse or promote products derived from this software |
| 43 | * without specific prior written permission. |
| 44 | * |
| 45 | * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND |
| 46 | * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 47 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 48 | * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE |
| 49 | * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
| 50 | * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS |
| 51 | * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) |
| 52 | * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
| 53 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY |
| 54 | * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
| 55 | * SUCH DAMAGE. |
| 56 | * |
| 57 | * @(#)vfs_cache.c 8.3 (Berkeley) 8/22/94 |
| 58 | */ |
| 59 | |
| 60 | #include <sys/cdefs.h> |
| 61 | __KERNEL_RCSID(0, "$NetBSD: vfs_cache.c,v 1.110 2016/07/07 06:55:43 msaitoh Exp $" ); |
| 62 | |
| 63 | #ifdef _KERNEL_OPT |
| 64 | #include "opt_ddb.h" |
| 65 | #include "opt_revcache.h" |
| 66 | #include "opt_dtrace.h" |
| 67 | #endif |
| 68 | |
| 69 | #include <sys/param.h> |
| 70 | #include <sys/systm.h> |
| 71 | #include <sys/sysctl.h> |
| 72 | #include <sys/time.h> |
| 73 | #include <sys/mount.h> |
| 74 | #include <sys/vnode.h> |
| 75 | #include <sys/namei.h> |
| 76 | #include <sys/errno.h> |
| 77 | #include <sys/pool.h> |
| 78 | #include <sys/mutex.h> |
| 79 | #include <sys/atomic.h> |
| 80 | #include <sys/kthread.h> |
| 81 | #include <sys/kernel.h> |
| 82 | #include <sys/cpu.h> |
| 83 | #include <sys/evcnt.h> |
| 84 | #include <sys/sdt.h> |
| 85 | |
| 86 | #define NAMECACHE_ENTER_REVERSE |
| 87 | /* |
| 88 | * Name caching works as follows: |
| 89 | * |
| 90 | * Names found by directory scans are retained in a cache |
| 91 | * for future reference. It is managed LRU, so frequently |
| 92 | * used names will hang around. Cache is indexed by hash value |
| 93 | * obtained from (dvp, name) where dvp refers to the directory |
| 94 | * containing name. |
| 95 | * |
| 96 | * For simplicity (and economy of storage), names longer than |
| 97 | * a maximum length of NCHNAMLEN are not cached; they occur |
| 98 | * infrequently in any case, and are almost never of interest. |
| 99 | * |
| 100 | * Upon reaching the last segment of a path, if the reference |
| 101 | * is for DELETE, or NOCACHE is set (rewrite), and the |
| 102 | * name is located in the cache, it will be dropped. |
| 103 | * The entry is dropped also when it was not possible to lock |
| 104 | * the cached vnode, either because vget() failed or the generation |
| 105 | * number has changed while waiting for the lock. |
| 106 | */ |
| 107 | |
| 108 | /* |
| 109 | * The locking in this subsystem works as follows: |
| 110 | * |
| 111 | * When an entry is added to the cache, via cache_enter(), |
| 112 | * namecache_lock is taken to exclude other writers. The new |
| 113 | * entry is added to the hash list in a way which permits |
| 114 | * concurrent lookups and invalidations in the cache done on |
| 115 | * other CPUs to continue in parallel. |
| 116 | * |
| 117 | * When a lookup is done in the cache, via cache_lookup() or |
| 118 | * cache_lookup_raw(), the per-cpu lock below is taken. This |
| 119 | * protects calls to cache_lookup_entry() and cache_invalidate() |
| 120 | * against cache_reclaim() but allows lookups to continue in |
| 121 | * parallel with cache_enter(). |
| 122 | * |
| 123 | * cache_revlookup() takes namecache_lock to exclude cache_enter() |
| 124 | * and cache_reclaim() since the list it operates on is not |
| 125 | * maintained to allow concurrent reads. |
| 126 | * |
| 127 | * When cache_reclaim() is called namecache_lock is held to hold |
| 128 | * off calls to cache_enter()/cache_revlookup() and each of the |
| 129 | * per-cpu locks is taken to hold off lookups. Holding all these |
| 130 | * locks essentially idles the subsystem, ensuring there are no |
| 131 | * concurrent references to the cache entries being freed. |
| 132 | * |
| 133 | * 32 bit per-cpu statistic counters (struct nchstats_percpu) are |
| 134 | * incremented when the operations they count are performed while |
| 135 | * running on the corresponding CPU. Frequently individual counters |
| 136 | * are incremented while holding a lock (either a per-cpu lock or |
| 137 | * namecache_lock) sufficient to preclude concurrent increments |
| 138 | * being done to the same counter, so non-atomic increments are |
| 139 | * done using the COUNT() macro. Counters which are incremented |
| 140 | * when one of these locks is not held use the COUNT_UNL() macro |
| 141 | * instead. COUNT_UNL() could be defined to do atomic increments |
| 142 | * but currently just does what COUNT() does, on the theory that |
| 143 | * it is unlikely the non-atomic increment will be interrupted |
| 144 | * by something on the same CPU that increments the same counter, |
| 145 | * but even if it does happen the consequences aren't serious. |
| 146 | * |
| 147 | * N.B.: Attempting to protect COUNT_UNL() increments by taking |
| 148 | * a per-cpu lock in the namecache_count_*() functions causes |
| 149 | * a deadlock. Don't do that, use atomic increments instead if |
| 150 | * the imperfections here bug you. |
| 151 | * |
| 152 | * The 64 bit system-wide statistic counts (struct nchstats) are |
| 153 | * maintained by sampling the per-cpu counters periodically, adding |
| 154 | * in the deltas since the last samples and recording the current |
| 155 | * samples to use to compute the next delta. The sampling is done |
| 156 | * as a side effect of cache_reclaim() which is run periodically, |
| 157 | * for its own purposes, often enough to avoid overflow of the 32 |
| 158 | * bit counters. While sampling in this fashion requires no locking |
| 159 | * it is never-the-less done only after all locks have been taken by |
| 160 | * cache_reclaim() to allow cache_stat_sysctl() to hold off |
| 161 | * cache_reclaim() with minimal locking. |
| 162 | * |
| 163 | * cache_stat_sysctl() takes its CPU's per-cpu lock to hold off |
| 164 | * cache_reclaim() so that it can copy the subsystem total stats |
| 165 | * without them being concurrently modified. If CACHE_STATS_CURRENT |
| 166 | * is defined it also harvests the per-cpu increments into the total, |
| 167 | * which again requires cache_reclaim() to be held off. |
| 168 | * |
| 169 | * The per-cpu data (a lock and the per-cpu stats structures) |
| 170 | * are defined next. |
| 171 | */ |
| 172 | struct nchstats_percpu _NAMEI_CACHE_STATS(uint32_t); |
| 173 | |
| 174 | struct nchcpu { |
| 175 | kmutex_t cpu_lock; |
| 176 | struct nchstats_percpu cpu_stats; |
| 177 | /* XXX maybe __cacheline_aligned would improve this? */ |
| 178 | struct nchstats_percpu cpu_stats_last; /* from last sample */ |
| 179 | }; |
| 180 | |
| 181 | /* |
| 182 | * The type for the hash code. While the hash function generates a |
| 183 | * u32, the hash code has historically been passed around as a u_long, |
| 184 | * and the value is modified by xor'ing a uintptr_t, so it's not |
| 185 | * entirely clear what the best type is. For now I'll leave it |
| 186 | * unchanged as u_long. |
| 187 | */ |
| 188 | |
| 189 | typedef u_long nchash_t; |
| 190 | |
| 191 | /* |
| 192 | * Structures associated with name cacheing. |
| 193 | */ |
| 194 | |
| 195 | static kmutex_t *namecache_lock __read_mostly; |
| 196 | static pool_cache_t namecache_cache __read_mostly; |
| 197 | static TAILQ_HEAD(, namecache) nclruhead __cacheline_aligned; |
| 198 | |
| 199 | static LIST_HEAD(nchashhead, namecache) *nchashtbl __read_mostly; |
| 200 | static u_long nchash __read_mostly; |
| 201 | |
| 202 | #define NCHASH2(hash, dvp) \ |
| 203 | (((hash) ^ ((uintptr_t)(dvp) >> 3)) & nchash) |
| 204 | |
| 205 | static LIST_HEAD(ncvhashhead, namecache) *ncvhashtbl __read_mostly; |
| 206 | static u_long ncvhash __read_mostly; |
| 207 | |
| 208 | #define NCVHASH(vp) (((uintptr_t)(vp) >> 3) & ncvhash) |
| 209 | |
| 210 | /* Number of cache entries allocated. */ |
| 211 | static long numcache __cacheline_aligned; |
| 212 | |
| 213 | /* Garbage collection queue and number of entries pending in it. */ |
| 214 | static void *cache_gcqueue; |
| 215 | static u_int cache_gcpend; |
| 216 | |
| 217 | /* Cache effectiveness statistics. This holds total from per-cpu stats */ |
| 218 | struct nchstats nchstats __cacheline_aligned; |
| 219 | |
| 220 | /* |
| 221 | * Macros to count an event, update the central stats with per-cpu |
| 222 | * values and add current per-cpu increments to the subsystem total |
| 223 | * last collected by cache_reclaim(). |
| 224 | */ |
| 225 | #define CACHE_STATS_CURRENT /* nothing */ |
| 226 | |
| 227 | #define COUNT(cpup, f) ((cpup)->cpu_stats.f++) |
| 228 | |
| 229 | #define UPDATE(cpup, f) do { \ |
| 230 | struct nchcpu *Xcpup = (cpup); \ |
| 231 | uint32_t Xcnt = (volatile uint32_t) Xcpup->cpu_stats.f; \ |
| 232 | nchstats.f += Xcnt - Xcpup->cpu_stats_last.f; \ |
| 233 | Xcpup->cpu_stats_last.f = Xcnt; \ |
| 234 | } while (/* CONSTCOND */ 0) |
| 235 | |
| 236 | #define ADD(stats, cpup, f) do { \ |
| 237 | struct nchcpu *Xcpup = (cpup); \ |
| 238 | stats.f += Xcpup->cpu_stats.f - Xcpup->cpu_stats_last.f; \ |
| 239 | } while (/* CONSTCOND */ 0) |
| 240 | |
| 241 | /* Do unlocked stats the same way. Use a different name to allow mind changes */ |
| 242 | #define COUNT_UNL(cpup, f) COUNT((cpup), f) |
| 243 | |
| 244 | static const int cache_lowat = 95; |
| 245 | static const int cache_hiwat = 98; |
| 246 | static const int cache_hottime = 5; /* number of seconds */ |
| 247 | static int doingcache = 1; /* 1 => enable the cache */ |
| 248 | |
| 249 | static struct evcnt cache_ev_scan; |
| 250 | static struct evcnt cache_ev_gc; |
| 251 | static struct evcnt cache_ev_over; |
| 252 | static struct evcnt cache_ev_under; |
| 253 | static struct evcnt cache_ev_forced; |
| 254 | |
| 255 | static void cache_invalidate(struct namecache *); |
| 256 | static struct namecache *cache_lookup_entry( |
| 257 | const struct vnode *, const char *, size_t); |
| 258 | static void cache_thread(void *); |
| 259 | static void cache_invalidate(struct namecache *); |
| 260 | static void cache_disassociate(struct namecache *); |
| 261 | static void cache_reclaim(void); |
| 262 | static int cache_ctor(void *, void *, int); |
| 263 | static void cache_dtor(void *, void *); |
| 264 | |
| 265 | static struct sysctllog *sysctllog; |
| 266 | static void sysctl_cache_stat_setup(void); |
| 267 | |
| 268 | SDT_PROVIDER_DEFINE(vfs); |
| 269 | |
| 270 | SDT_PROBE_DEFINE1(vfs, namecache, invalidate, done, "struct vnode *" ); |
| 271 | SDT_PROBE_DEFINE1(vfs, namecache, purge, parents, "struct vnode *" ); |
| 272 | SDT_PROBE_DEFINE1(vfs, namecache, purge, children, "struct vnode *" ); |
| 273 | SDT_PROBE_DEFINE2(vfs, namecache, purge, name, "char *" , "size_t" ); |
| 274 | SDT_PROBE_DEFINE1(vfs, namecache, purge, vfs, "struct mount *" ); |
| 275 | SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *" , |
| 276 | "char *" , "size_t" ); |
| 277 | SDT_PROBE_DEFINE3(vfs, namecache, lookup, miss, "struct vnode *" , |
| 278 | "char *" , "size_t" ); |
| 279 | SDT_PROBE_DEFINE3(vfs, namecache, lookup, toolong, "struct vnode *" , |
| 280 | "char *" , "size_t" ); |
| 281 | SDT_PROBE_DEFINE2(vfs, namecache, revlookup, success, "struct vnode *" , |
| 282 | "struct vnode *" ); |
| 283 | SDT_PROBE_DEFINE2(vfs, namecache, revlookup, fail, "struct vnode *" , |
| 284 | "int" ); |
| 285 | SDT_PROBE_DEFINE2(vfs, namecache, prune, done, "int" , "int" ); |
| 286 | SDT_PROBE_DEFINE3(vfs, namecache, enter, toolong, "struct vnode *" , |
| 287 | "char *" , "size_t" ); |
| 288 | SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *" , |
| 289 | "char *" , "size_t" ); |
| 290 | |
| 291 | /* |
| 292 | * Compute the hash for an entry. |
| 293 | * |
| 294 | * (This is for now a wrapper around namei_hash, whose interface is |
| 295 | * for the time being slightly inconvenient.) |
| 296 | */ |
| 297 | static nchash_t |
| 298 | cache_hash(const char *name, size_t namelen) |
| 299 | { |
| 300 | const char *endptr; |
| 301 | |
| 302 | endptr = name + namelen; |
| 303 | return namei_hash(name, &endptr); |
| 304 | } |
| 305 | |
| 306 | /* |
| 307 | * Invalidate a cache entry and enqueue it for garbage collection. |
| 308 | * The caller needs to hold namecache_lock or a per-cpu lock to hold |
| 309 | * off cache_reclaim(). |
| 310 | */ |
| 311 | static void |
| 312 | cache_invalidate(struct namecache *ncp) |
| 313 | { |
| 314 | void *head; |
| 315 | |
| 316 | KASSERT(mutex_owned(&ncp->nc_lock)); |
| 317 | |
| 318 | if (ncp->nc_dvp != NULL) { |
| 319 | SDT_PROBE(vfs, namecache, invalidate, done, ncp->nc_dvp, |
| 320 | 0, 0, 0, 0); |
| 321 | |
| 322 | ncp->nc_vp = NULL; |
| 323 | ncp->nc_dvp = NULL; |
| 324 | do { |
| 325 | head = cache_gcqueue; |
| 326 | ncp->nc_gcqueue = head; |
| 327 | } while (atomic_cas_ptr(&cache_gcqueue, head, ncp) != head); |
| 328 | atomic_inc_uint(&cache_gcpend); |
| 329 | } |
| 330 | } |
| 331 | |
| 332 | /* |
| 333 | * Disassociate a namecache entry from any vnodes it is attached to, |
| 334 | * and remove from the global LRU list. |
| 335 | */ |
| 336 | static void |
| 337 | cache_disassociate(struct namecache *ncp) |
| 338 | { |
| 339 | |
| 340 | KASSERT(mutex_owned(namecache_lock)); |
| 341 | KASSERT(ncp->nc_dvp == NULL); |
| 342 | |
| 343 | if (ncp->nc_lru.tqe_prev != NULL) { |
| 344 | TAILQ_REMOVE(&nclruhead, ncp, nc_lru); |
| 345 | ncp->nc_lru.tqe_prev = NULL; |
| 346 | } |
| 347 | if (ncp->nc_vhash.le_prev != NULL) { |
| 348 | LIST_REMOVE(ncp, nc_vhash); |
| 349 | ncp->nc_vhash.le_prev = NULL; |
| 350 | } |
| 351 | if (ncp->nc_vlist.le_prev != NULL) { |
| 352 | LIST_REMOVE(ncp, nc_vlist); |
| 353 | ncp->nc_vlist.le_prev = NULL; |
| 354 | } |
| 355 | if (ncp->nc_dvlist.le_prev != NULL) { |
| 356 | LIST_REMOVE(ncp, nc_dvlist); |
| 357 | ncp->nc_dvlist.le_prev = NULL; |
| 358 | } |
| 359 | } |
| 360 | |
| 361 | /* |
| 362 | * Lock all CPUs to prevent any cache lookup activity. Conceptually, |
| 363 | * this locks out all "readers". |
| 364 | */ |
| 365 | static void |
| 366 | cache_lock_cpus(void) |
| 367 | { |
| 368 | CPU_INFO_ITERATOR cii; |
| 369 | struct cpu_info *ci; |
| 370 | struct nchcpu *cpup; |
| 371 | |
| 372 | /* |
| 373 | * Lock out all CPUs first, then harvest per-cpu stats. This |
| 374 | * is probably not quite as cache-efficient as doing the lock |
| 375 | * and harvest at the same time, but allows cache_stat_sysctl() |
| 376 | * to make do with a per-cpu lock. |
| 377 | */ |
| 378 | for (CPU_INFO_FOREACH(cii, ci)) { |
| 379 | cpup = ci->ci_data.cpu_nch; |
| 380 | mutex_enter(&cpup->cpu_lock); |
| 381 | } |
| 382 | for (CPU_INFO_FOREACH(cii, ci)) { |
| 383 | cpup = ci->ci_data.cpu_nch; |
| 384 | UPDATE(cpup, ncs_goodhits); |
| 385 | UPDATE(cpup, ncs_neghits); |
| 386 | UPDATE(cpup, ncs_badhits); |
| 387 | UPDATE(cpup, ncs_falsehits); |
| 388 | UPDATE(cpup, ncs_miss); |
| 389 | UPDATE(cpup, ncs_long); |
| 390 | UPDATE(cpup, ncs_pass2); |
| 391 | UPDATE(cpup, ncs_2passes); |
| 392 | UPDATE(cpup, ncs_revhits); |
| 393 | UPDATE(cpup, ncs_revmiss); |
| 394 | } |
| 395 | } |
| 396 | |
| 397 | /* |
| 398 | * Release all CPU locks. |
| 399 | */ |
| 400 | static void |
| 401 | cache_unlock_cpus(void) |
| 402 | { |
| 403 | CPU_INFO_ITERATOR cii; |
| 404 | struct cpu_info *ci; |
| 405 | struct nchcpu *cpup; |
| 406 | |
| 407 | for (CPU_INFO_FOREACH(cii, ci)) { |
| 408 | cpup = ci->ci_data.cpu_nch; |
| 409 | mutex_exit(&cpup->cpu_lock); |
| 410 | } |
| 411 | } |
| 412 | |
| 413 | /* |
| 414 | * Find a single cache entry and return it locked. |
| 415 | * The caller needs to hold namecache_lock or a per-cpu lock to hold |
| 416 | * off cache_reclaim(). |
| 417 | */ |
| 418 | static struct namecache * |
| 419 | cache_lookup_entry(const struct vnode *dvp, const char *name, size_t namelen) |
| 420 | { |
| 421 | struct nchashhead *ncpp; |
| 422 | struct namecache *ncp; |
| 423 | nchash_t hash; |
| 424 | |
| 425 | KASSERT(dvp != NULL); |
| 426 | hash = cache_hash(name, namelen); |
| 427 | ncpp = &nchashtbl[NCHASH2(hash, dvp)]; |
| 428 | |
| 429 | LIST_FOREACH(ncp, ncpp, nc_hash) { |
| 430 | membar_datadep_consumer(); /* for Alpha... */ |
| 431 | if (ncp->nc_dvp != dvp || |
| 432 | ncp->nc_nlen != namelen || |
| 433 | memcmp(ncp->nc_name, name, (u_int)ncp->nc_nlen)) |
| 434 | continue; |
| 435 | mutex_enter(&ncp->nc_lock); |
| 436 | if (__predict_true(ncp->nc_dvp == dvp)) { |
| 437 | ncp->nc_hittime = hardclock_ticks; |
| 438 | SDT_PROBE(vfs, namecache, lookup, hit, dvp, |
| 439 | name, namelen, 0, 0); |
| 440 | return ncp; |
| 441 | } |
| 442 | /* Raced: entry has been nullified. */ |
| 443 | mutex_exit(&ncp->nc_lock); |
| 444 | } |
| 445 | |
| 446 | SDT_PROBE(vfs, namecache, lookup, miss, dvp, |
| 447 | name, namelen, 0, 0); |
| 448 | return NULL; |
| 449 | } |
| 450 | |
| 451 | /* |
| 452 | * Look for a the name in the cache. We don't do this |
| 453 | * if the segment name is long, simply so the cache can avoid |
| 454 | * holding long names (which would either waste space, or |
| 455 | * add greatly to the complexity). |
| 456 | * |
| 457 | * Lookup is called with DVP pointing to the directory to search, |
| 458 | * and CNP providing the name of the entry being sought: cn_nameptr |
| 459 | * is the name, cn_namelen is its length, and cn_flags is the flags |
| 460 | * word from the namei operation. |
| 461 | * |
| 462 | * DVP must be locked. |
| 463 | * |
| 464 | * There are three possible non-error return states: |
| 465 | * 1. Nothing was found in the cache. Nothing is known about |
| 466 | * the requested name. |
| 467 | * 2. A negative entry was found in the cache, meaning that the |
| 468 | * requested name definitely does not exist. |
| 469 | * 3. A positive entry was found in the cache, meaning that the |
| 470 | * requested name does exist and that we are providing the |
| 471 | * vnode. |
| 472 | * In these cases the results are: |
| 473 | * 1. 0 returned; VN is set to NULL. |
| 474 | * 2. 1 returned; VN is set to NULL. |
| 475 | * 3. 1 returned; VN is set to the vnode found. |
| 476 | * |
| 477 | * The additional result argument ISWHT is set to zero, unless a |
| 478 | * negative entry is found that was entered as a whiteout, in which |
| 479 | * case ISWHT is set to one. |
| 480 | * |
| 481 | * The ISWHT_RET argument pointer may be null. In this case an |
| 482 | * assertion is made that the whiteout flag is not set. File systems |
| 483 | * that do not support whiteouts can/should do this. |
| 484 | * |
| 485 | * Filesystems that do support whiteouts should add ISWHITEOUT to |
| 486 | * cnp->cn_flags if ISWHT comes back nonzero. |
| 487 | * |
| 488 | * When a vnode is returned, it is locked, as per the vnode lookup |
| 489 | * locking protocol. |
| 490 | * |
| 491 | * There is no way for this function to fail, in the sense of |
| 492 | * generating an error that requires aborting the namei operation. |
| 493 | * |
| 494 | * (Prior to October 2012, this function returned an integer status, |
| 495 | * and a vnode, and mucked with the flags word in CNP for whiteouts. |
| 496 | * The integer status was -1 for "nothing found", ENOENT for "a |
| 497 | * negative entry found", 0 for "a positive entry found", and possibly |
| 498 | * other errors, and the value of VN might or might not have been set |
| 499 | * depending on what error occurred.) |
| 500 | */ |
| 501 | int |
| 502 | cache_lookup(struct vnode *dvp, const char *name, size_t namelen, |
| 503 | uint32_t nameiop, uint32_t cnflags, |
| 504 | int *iswht_ret, struct vnode **vn_ret) |
| 505 | { |
| 506 | struct namecache *ncp; |
| 507 | struct vnode *vp; |
| 508 | struct nchcpu *cpup; |
| 509 | int error, ret_value; |
| 510 | |
| 511 | |
| 512 | /* Establish default result values */ |
| 513 | if (iswht_ret != NULL) { |
| 514 | *iswht_ret = 0; |
| 515 | } |
| 516 | *vn_ret = NULL; |
| 517 | |
| 518 | if (__predict_false(!doingcache)) { |
| 519 | return 0; |
| 520 | } |
| 521 | |
| 522 | cpup = curcpu()->ci_data.cpu_nch; |
| 523 | mutex_enter(&cpup->cpu_lock); |
| 524 | if (__predict_false(namelen > NCHNAMLEN)) { |
| 525 | SDT_PROBE(vfs, namecache, lookup, toolong, dvp, |
| 526 | name, namelen, 0, 0); |
| 527 | COUNT(cpup, ncs_long); |
| 528 | mutex_exit(&cpup->cpu_lock); |
| 529 | /* found nothing */ |
| 530 | return 0; |
| 531 | } |
| 532 | |
| 533 | ncp = cache_lookup_entry(dvp, name, namelen); |
| 534 | if (__predict_false(ncp == NULL)) { |
| 535 | COUNT(cpup, ncs_miss); |
| 536 | mutex_exit(&cpup->cpu_lock); |
| 537 | /* found nothing */ |
| 538 | return 0; |
| 539 | } |
| 540 | if ((cnflags & MAKEENTRY) == 0) { |
| 541 | COUNT(cpup, ncs_badhits); |
| 542 | /* |
| 543 | * Last component and we are renaming or deleting, |
| 544 | * the cache entry is invalid, or otherwise don't |
| 545 | * want cache entry to exist. |
| 546 | */ |
| 547 | cache_invalidate(ncp); |
| 548 | mutex_exit(&ncp->nc_lock); |
| 549 | mutex_exit(&cpup->cpu_lock); |
| 550 | /* found nothing */ |
| 551 | return 0; |
| 552 | } |
| 553 | if (ncp->nc_vp == NULL) { |
| 554 | if (iswht_ret != NULL) { |
| 555 | /* |
| 556 | * Restore the ISWHITEOUT flag saved earlier. |
| 557 | */ |
| 558 | KASSERT((ncp->nc_flags & ~ISWHITEOUT) == 0); |
| 559 | *iswht_ret = (ncp->nc_flags & ISWHITEOUT) != 0; |
| 560 | } else { |
| 561 | KASSERT(ncp->nc_flags == 0); |
| 562 | } |
| 563 | |
| 564 | if (__predict_true(nameiop != CREATE || |
| 565 | (cnflags & ISLASTCN) == 0)) { |
| 566 | COUNT(cpup, ncs_neghits); |
| 567 | /* found neg entry; vn is already null from above */ |
| 568 | ret_value = 1; |
| 569 | } else { |
| 570 | COUNT(cpup, ncs_badhits); |
| 571 | /* |
| 572 | * Last component and we are preparing to create |
| 573 | * the named object, so flush the negative cache |
| 574 | * entry. |
| 575 | */ |
| 576 | cache_invalidate(ncp); |
| 577 | /* found nothing */ |
| 578 | ret_value = 0; |
| 579 | } |
| 580 | mutex_exit(&ncp->nc_lock); |
| 581 | mutex_exit(&cpup->cpu_lock); |
| 582 | return ret_value; |
| 583 | } |
| 584 | |
| 585 | vp = ncp->nc_vp; |
| 586 | mutex_enter(vp->v_interlock); |
| 587 | mutex_exit(&ncp->nc_lock); |
| 588 | mutex_exit(&cpup->cpu_lock); |
| 589 | |
| 590 | /* |
| 591 | * Unlocked except for the vnode interlock. Call vget(). |
| 592 | */ |
| 593 | error = vget(vp, LK_NOWAIT, false /* !wait */); |
| 594 | if (error) { |
| 595 | KASSERT(error == EBUSY); |
| 596 | /* |
| 597 | * This vnode is being cleaned out. |
| 598 | * XXX badhits? |
| 599 | */ |
| 600 | COUNT_UNL(cpup, ncs_falsehits); |
| 601 | /* found nothing */ |
| 602 | return 0; |
| 603 | } |
| 604 | |
| 605 | COUNT_UNL(cpup, ncs_goodhits); |
| 606 | /* found it */ |
| 607 | *vn_ret = vp; |
| 608 | return 1; |
| 609 | } |
| 610 | |
| 611 | |
| 612 | /* |
| 613 | * Cut-'n-pasted version of the above without the nameiop argument. |
| 614 | */ |
| 615 | int |
| 616 | cache_lookup_raw(struct vnode *dvp, const char *name, size_t namelen, |
| 617 | uint32_t cnflags, |
| 618 | int *iswht_ret, struct vnode **vn_ret) |
| 619 | { |
| 620 | struct namecache *ncp; |
| 621 | struct vnode *vp; |
| 622 | struct nchcpu *cpup; |
| 623 | int error; |
| 624 | |
| 625 | /* Establish default results. */ |
| 626 | if (iswht_ret != NULL) { |
| 627 | *iswht_ret = 0; |
| 628 | } |
| 629 | *vn_ret = NULL; |
| 630 | |
| 631 | if (__predict_false(!doingcache)) { |
| 632 | /* found nothing */ |
| 633 | return 0; |
| 634 | } |
| 635 | |
| 636 | cpup = curcpu()->ci_data.cpu_nch; |
| 637 | mutex_enter(&cpup->cpu_lock); |
| 638 | if (__predict_false(namelen > NCHNAMLEN)) { |
| 639 | COUNT(cpup, ncs_long); |
| 640 | mutex_exit(&cpup->cpu_lock); |
| 641 | /* found nothing */ |
| 642 | return 0; |
| 643 | } |
| 644 | ncp = cache_lookup_entry(dvp, name, namelen); |
| 645 | if (__predict_false(ncp == NULL)) { |
| 646 | COUNT(cpup, ncs_miss); |
| 647 | mutex_exit(&cpup->cpu_lock); |
| 648 | /* found nothing */ |
| 649 | return 0; |
| 650 | } |
| 651 | vp = ncp->nc_vp; |
| 652 | if (vp == NULL) { |
| 653 | /* |
| 654 | * Restore the ISWHITEOUT flag saved earlier. |
| 655 | */ |
| 656 | if (iswht_ret != NULL) { |
| 657 | KASSERT((ncp->nc_flags & ~ISWHITEOUT) == 0); |
| 658 | /*cnp->cn_flags |= ncp->nc_flags;*/ |
| 659 | *iswht_ret = (ncp->nc_flags & ISWHITEOUT) != 0; |
| 660 | } |
| 661 | COUNT(cpup, ncs_neghits); |
| 662 | mutex_exit(&ncp->nc_lock); |
| 663 | mutex_exit(&cpup->cpu_lock); |
| 664 | /* found negative entry; vn is already null from above */ |
| 665 | return 1; |
| 666 | } |
| 667 | mutex_enter(vp->v_interlock); |
| 668 | mutex_exit(&ncp->nc_lock); |
| 669 | mutex_exit(&cpup->cpu_lock); |
| 670 | |
| 671 | /* |
| 672 | * Unlocked except for the vnode interlock. Call vget(). |
| 673 | */ |
| 674 | error = vget(vp, LK_NOWAIT, false /* !wait */); |
| 675 | if (error) { |
| 676 | KASSERT(error == EBUSY); |
| 677 | /* |
| 678 | * This vnode is being cleaned out. |
| 679 | * XXX badhits? |
| 680 | */ |
| 681 | COUNT_UNL(cpup, ncs_falsehits); |
| 682 | /* found nothing */ |
| 683 | return 0; |
| 684 | } |
| 685 | |
| 686 | COUNT_UNL(cpup, ncs_goodhits); /* XXX can be "badhits" */ |
| 687 | /* found it */ |
| 688 | *vn_ret = vp; |
| 689 | return 1; |
| 690 | } |
| 691 | |
| 692 | /* |
| 693 | * Scan cache looking for name of directory entry pointing at vp. |
| 694 | * |
| 695 | * If the lookup succeeds the vnode is referenced and stored in dvpp. |
| 696 | * |
| 697 | * If bufp is non-NULL, also place the name in the buffer which starts |
| 698 | * at bufp, immediately before *bpp, and move bpp backwards to point |
| 699 | * at the start of it. (Yes, this is a little baroque, but it's done |
| 700 | * this way to cater to the whims of getcwd). |
| 701 | * |
| 702 | * Returns 0 on success, -1 on cache miss, positive errno on failure. |
| 703 | */ |
| 704 | int |
| 705 | cache_revlookup(struct vnode *vp, struct vnode **dvpp, char **bpp, char *bufp) |
| 706 | { |
| 707 | struct namecache *ncp; |
| 708 | struct vnode *dvp; |
| 709 | struct ncvhashhead *nvcpp; |
| 710 | struct nchcpu *cpup; |
| 711 | char *bp; |
| 712 | int error, nlen; |
| 713 | |
| 714 | if (!doingcache) |
| 715 | goto out; |
| 716 | |
| 717 | nvcpp = &ncvhashtbl[NCVHASH(vp)]; |
| 718 | |
| 719 | /* |
| 720 | * We increment counters in the local CPU's per-cpu stats. |
| 721 | * We don't take the per-cpu lock, however, since this function |
| 722 | * is the only place these counters are incremented so no one |
| 723 | * will be racing with us to increment them. |
| 724 | */ |
| 725 | cpup = curcpu()->ci_data.cpu_nch; |
| 726 | mutex_enter(namecache_lock); |
| 727 | LIST_FOREACH(ncp, nvcpp, nc_vhash) { |
| 728 | mutex_enter(&ncp->nc_lock); |
| 729 | if (ncp->nc_vp == vp && |
| 730 | (dvp = ncp->nc_dvp) != NULL && |
| 731 | dvp != vp) { /* avoid pesky . entries.. */ |
| 732 | |
| 733 | #ifdef DIAGNOSTIC |
| 734 | if (ncp->nc_nlen == 1 && |
| 735 | ncp->nc_name[0] == '.') |
| 736 | panic("cache_revlookup: found entry for ." ); |
| 737 | |
| 738 | if (ncp->nc_nlen == 2 && |
| 739 | ncp->nc_name[0] == '.' && |
| 740 | ncp->nc_name[1] == '.') |
| 741 | panic("cache_revlookup: found entry for .." ); |
| 742 | #endif |
| 743 | COUNT(cpup, ncs_revhits); |
| 744 | nlen = ncp->nc_nlen; |
| 745 | |
| 746 | if (bufp) { |
| 747 | bp = *bpp; |
| 748 | bp -= nlen; |
| 749 | if (bp <= bufp) { |
| 750 | *dvpp = NULL; |
| 751 | mutex_exit(&ncp->nc_lock); |
| 752 | mutex_exit(namecache_lock); |
| 753 | SDT_PROBE(vfs, namecache, revlookup, |
| 754 | fail, vp, ERANGE, 0, 0, 0); |
| 755 | return (ERANGE); |
| 756 | } |
| 757 | memcpy(bp, ncp->nc_name, nlen); |
| 758 | *bpp = bp; |
| 759 | } |
| 760 | |
| 761 | mutex_enter(dvp->v_interlock); |
| 762 | mutex_exit(&ncp->nc_lock); |
| 763 | mutex_exit(namecache_lock); |
| 764 | error = vget(dvp, LK_NOWAIT, false /* !wait */); |
| 765 | if (error) { |
| 766 | KASSERT(error == EBUSY); |
| 767 | if (bufp) |
| 768 | (*bpp) += nlen; |
| 769 | *dvpp = NULL; |
| 770 | SDT_PROBE(vfs, namecache, revlookup, fail, vp, |
| 771 | error, 0, 0, 0); |
| 772 | return -1; |
| 773 | } |
| 774 | *dvpp = dvp; |
| 775 | SDT_PROBE(vfs, namecache, revlookup, success, vp, dvp, |
| 776 | 0, 0, 0); |
| 777 | return (0); |
| 778 | } |
| 779 | mutex_exit(&ncp->nc_lock); |
| 780 | } |
| 781 | COUNT(cpup, ncs_revmiss); |
| 782 | mutex_exit(namecache_lock); |
| 783 | out: |
| 784 | *dvpp = NULL; |
| 785 | return (-1); |
| 786 | } |
| 787 | |
| 788 | /* |
| 789 | * Add an entry to the cache |
| 790 | */ |
| 791 | void |
| 792 | cache_enter(struct vnode *dvp, struct vnode *vp, |
| 793 | const char *name, size_t namelen, uint32_t cnflags) |
| 794 | { |
| 795 | struct namecache *ncp; |
| 796 | struct namecache *oncp; |
| 797 | struct nchashhead *ncpp; |
| 798 | struct ncvhashhead *nvcpp; |
| 799 | nchash_t hash; |
| 800 | |
| 801 | /* First, check whether we can/should add a cache entry. */ |
| 802 | if ((cnflags & MAKEENTRY) == 0 || |
| 803 | __predict_false(namelen > NCHNAMLEN || !doingcache)) { |
| 804 | SDT_PROBE(vfs, namecache, enter, toolong, vp, name, namelen, |
| 805 | 0, 0); |
| 806 | return; |
| 807 | } |
| 808 | |
| 809 | SDT_PROBE(vfs, namecache, enter, done, vp, name, namelen, 0, 0); |
| 810 | if (numcache > desiredvnodes) { |
| 811 | mutex_enter(namecache_lock); |
| 812 | cache_ev_forced.ev_count++; |
| 813 | cache_reclaim(); |
| 814 | mutex_exit(namecache_lock); |
| 815 | } |
| 816 | |
| 817 | ncp = pool_cache_get(namecache_cache, PR_WAITOK); |
| 818 | mutex_enter(namecache_lock); |
| 819 | numcache++; |
| 820 | |
| 821 | /* |
| 822 | * Concurrent lookups in the same directory may race for a |
| 823 | * cache entry. if there's a duplicated entry, free it. |
| 824 | */ |
| 825 | oncp = cache_lookup_entry(dvp, name, namelen); |
| 826 | if (oncp) { |
| 827 | cache_invalidate(oncp); |
| 828 | mutex_exit(&oncp->nc_lock); |
| 829 | } |
| 830 | |
| 831 | /* Grab the vnode we just found. */ |
| 832 | mutex_enter(&ncp->nc_lock); |
| 833 | ncp->nc_vp = vp; |
| 834 | ncp->nc_flags = 0; |
| 835 | ncp->nc_hittime = 0; |
| 836 | ncp->nc_gcqueue = NULL; |
| 837 | if (vp == NULL) { |
| 838 | /* |
| 839 | * For negative hits, save the ISWHITEOUT flag so we can |
| 840 | * restore it later when the cache entry is used again. |
| 841 | */ |
| 842 | ncp->nc_flags = cnflags & ISWHITEOUT; |
| 843 | } |
| 844 | |
| 845 | /* Fill in cache info. */ |
| 846 | ncp->nc_dvp = dvp; |
| 847 | LIST_INSERT_HEAD(&dvp->v_dnclist, ncp, nc_dvlist); |
| 848 | if (vp) |
| 849 | LIST_INSERT_HEAD(&vp->v_nclist, ncp, nc_vlist); |
| 850 | else { |
| 851 | ncp->nc_vlist.le_prev = NULL; |
| 852 | ncp->nc_vlist.le_next = NULL; |
| 853 | } |
| 854 | KASSERT(namelen <= NCHNAMLEN); |
| 855 | ncp->nc_nlen = namelen; |
| 856 | memcpy(ncp->nc_name, name, (unsigned)ncp->nc_nlen); |
| 857 | TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru); |
| 858 | hash = cache_hash(name, namelen); |
| 859 | ncpp = &nchashtbl[NCHASH2(hash, dvp)]; |
| 860 | |
| 861 | /* |
| 862 | * Flush updates before making visible in table. No need for a |
| 863 | * memory barrier on the other side: to see modifications the |
| 864 | * list must be followed, meaning a dependent pointer load. |
| 865 | * The below is LIST_INSERT_HEAD() inlined, with the memory |
| 866 | * barrier included in the correct place. |
| 867 | */ |
| 868 | if ((ncp->nc_hash.le_next = ncpp->lh_first) != NULL) |
| 869 | ncpp->lh_first->nc_hash.le_prev = &ncp->nc_hash.le_next; |
| 870 | ncp->nc_hash.le_prev = &ncpp->lh_first; |
| 871 | membar_producer(); |
| 872 | ncpp->lh_first = ncp; |
| 873 | |
| 874 | ncp->nc_vhash.le_prev = NULL; |
| 875 | ncp->nc_vhash.le_next = NULL; |
| 876 | |
| 877 | /* |
| 878 | * Create reverse-cache entries (used in getcwd) for directories. |
| 879 | * (and in linux procfs exe node) |
| 880 | */ |
| 881 | if (vp != NULL && |
| 882 | vp != dvp && |
| 883 | #ifndef NAMECACHE_ENTER_REVERSE |
| 884 | vp->v_type == VDIR && |
| 885 | #endif |
| 886 | (ncp->nc_nlen > 2 || |
| 887 | (ncp->nc_nlen > 1 && ncp->nc_name[1] != '.') || |
| 888 | (/* ncp->nc_nlen > 0 && */ ncp->nc_name[0] != '.'))) { |
| 889 | nvcpp = &ncvhashtbl[NCVHASH(vp)]; |
| 890 | LIST_INSERT_HEAD(nvcpp, ncp, nc_vhash); |
| 891 | } |
| 892 | mutex_exit(&ncp->nc_lock); |
| 893 | mutex_exit(namecache_lock); |
| 894 | } |
| 895 | |
| 896 | /* |
| 897 | * Name cache initialization, from vfs_init() when we are booting |
| 898 | */ |
| 899 | void |
| 900 | nchinit(void) |
| 901 | { |
| 902 | int error; |
| 903 | |
| 904 | TAILQ_INIT(&nclruhead); |
| 905 | namecache_cache = pool_cache_init(sizeof(struct namecache), |
| 906 | coherency_unit, 0, 0, "ncache" , NULL, IPL_NONE, cache_ctor, |
| 907 | cache_dtor, NULL); |
| 908 | KASSERT(namecache_cache != NULL); |
| 909 | |
| 910 | namecache_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_NONE); |
| 911 | |
| 912 | nchashtbl = hashinit(desiredvnodes, HASH_LIST, true, &nchash); |
| 913 | ncvhashtbl = |
| 914 | #ifdef NAMECACHE_ENTER_REVERSE |
| 915 | hashinit(desiredvnodes, HASH_LIST, true, &ncvhash); |
| 916 | #else |
| 917 | hashinit(desiredvnodes/8, HASH_LIST, true, &ncvhash); |
| 918 | #endif |
| 919 | |
| 920 | error = kthread_create(PRI_VM, KTHREAD_MPSAFE, NULL, cache_thread, |
| 921 | NULL, NULL, "cachegc" ); |
| 922 | if (error != 0) |
| 923 | panic("nchinit %d" , error); |
| 924 | |
| 925 | evcnt_attach_dynamic(&cache_ev_scan, EVCNT_TYPE_MISC, NULL, |
| 926 | "namecache" , "entries scanned" ); |
| 927 | evcnt_attach_dynamic(&cache_ev_gc, EVCNT_TYPE_MISC, NULL, |
| 928 | "namecache" , "entries collected" ); |
| 929 | evcnt_attach_dynamic(&cache_ev_over, EVCNT_TYPE_MISC, NULL, |
| 930 | "namecache" , "over scan target" ); |
| 931 | evcnt_attach_dynamic(&cache_ev_under, EVCNT_TYPE_MISC, NULL, |
| 932 | "namecache" , "under scan target" ); |
| 933 | evcnt_attach_dynamic(&cache_ev_forced, EVCNT_TYPE_MISC, NULL, |
| 934 | "namecache" , "forced reclaims" ); |
| 935 | |
| 936 | sysctl_cache_stat_setup(); |
| 937 | } |
| 938 | |
| 939 | static int |
| 940 | cache_ctor(void *arg, void *obj, int flag) |
| 941 | { |
| 942 | struct namecache *ncp; |
| 943 | |
| 944 | ncp = obj; |
| 945 | mutex_init(&ncp->nc_lock, MUTEX_DEFAULT, IPL_NONE); |
| 946 | |
| 947 | return 0; |
| 948 | } |
| 949 | |
| 950 | static void |
| 951 | cache_dtor(void *arg, void *obj) |
| 952 | { |
| 953 | struct namecache *ncp; |
| 954 | |
| 955 | ncp = obj; |
| 956 | mutex_destroy(&ncp->nc_lock); |
| 957 | } |
| 958 | |
| 959 | /* |
| 960 | * Called once for each CPU in the system as attached. |
| 961 | */ |
| 962 | void |
| 963 | cache_cpu_init(struct cpu_info *ci) |
| 964 | { |
| 965 | struct nchcpu *cpup; |
| 966 | size_t sz; |
| 967 | |
| 968 | sz = roundup2(sizeof(*cpup), coherency_unit) + coherency_unit; |
| 969 | cpup = kmem_zalloc(sz, KM_SLEEP); |
| 970 | cpup = (void *)roundup2((uintptr_t)cpup, coherency_unit); |
| 971 | mutex_init(&cpup->cpu_lock, MUTEX_DEFAULT, IPL_NONE); |
| 972 | ci->ci_data.cpu_nch = cpup; |
| 973 | } |
| 974 | |
| 975 | /* |
| 976 | * Name cache reinitialization, for when the maximum number of vnodes increases. |
| 977 | */ |
| 978 | void |
| 979 | nchreinit(void) |
| 980 | { |
| 981 | struct namecache *ncp; |
| 982 | struct nchashhead *oldhash1, *hash1; |
| 983 | struct ncvhashhead *oldhash2, *hash2; |
| 984 | u_long i, oldmask1, oldmask2, mask1, mask2; |
| 985 | |
| 986 | hash1 = hashinit(desiredvnodes, HASH_LIST, true, &mask1); |
| 987 | hash2 = |
| 988 | #ifdef NAMECACHE_ENTER_REVERSE |
| 989 | hashinit(desiredvnodes, HASH_LIST, true, &mask2); |
| 990 | #else |
| 991 | hashinit(desiredvnodes/8, HASH_LIST, true, &mask2); |
| 992 | #endif |
| 993 | mutex_enter(namecache_lock); |
| 994 | cache_lock_cpus(); |
| 995 | oldhash1 = nchashtbl; |
| 996 | oldmask1 = nchash; |
| 997 | nchashtbl = hash1; |
| 998 | nchash = mask1; |
| 999 | oldhash2 = ncvhashtbl; |
| 1000 | oldmask2 = ncvhash; |
| 1001 | ncvhashtbl = hash2; |
| 1002 | ncvhash = mask2; |
| 1003 | for (i = 0; i <= oldmask1; i++) { |
| 1004 | while ((ncp = LIST_FIRST(&oldhash1[i])) != NULL) { |
| 1005 | LIST_REMOVE(ncp, nc_hash); |
| 1006 | ncp->nc_hash.le_prev = NULL; |
| 1007 | } |
| 1008 | } |
| 1009 | for (i = 0; i <= oldmask2; i++) { |
| 1010 | while ((ncp = LIST_FIRST(&oldhash2[i])) != NULL) { |
| 1011 | LIST_REMOVE(ncp, nc_vhash); |
| 1012 | ncp->nc_vhash.le_prev = NULL; |
| 1013 | } |
| 1014 | } |
| 1015 | cache_unlock_cpus(); |
| 1016 | mutex_exit(namecache_lock); |
| 1017 | hashdone(oldhash1, HASH_LIST, oldmask1); |
| 1018 | hashdone(oldhash2, HASH_LIST, oldmask2); |
| 1019 | } |
| 1020 | |
| 1021 | /* |
| 1022 | * Cache flush, a particular vnode; called when a vnode is renamed to |
| 1023 | * hide entries that would now be invalid |
| 1024 | */ |
| 1025 | void |
| 1026 | cache_purge1(struct vnode *vp, const char *name, size_t namelen, int flags) |
| 1027 | { |
| 1028 | struct namecache *ncp, *ncnext; |
| 1029 | |
| 1030 | mutex_enter(namecache_lock); |
| 1031 | if (flags & PURGE_PARENTS) { |
| 1032 | SDT_PROBE(vfs, namecache, purge, parents, vp, 0, 0, 0, 0); |
| 1033 | |
| 1034 | for (ncp = LIST_FIRST(&vp->v_nclist); ncp != NULL; |
| 1035 | ncp = ncnext) { |
| 1036 | ncnext = LIST_NEXT(ncp, nc_vlist); |
| 1037 | mutex_enter(&ncp->nc_lock); |
| 1038 | cache_invalidate(ncp); |
| 1039 | mutex_exit(&ncp->nc_lock); |
| 1040 | cache_disassociate(ncp); |
| 1041 | } |
| 1042 | } |
| 1043 | if (flags & PURGE_CHILDREN) { |
| 1044 | SDT_PROBE(vfs, namecache, purge, children, vp, 0, 0, 0, 0); |
| 1045 | for (ncp = LIST_FIRST(&vp->v_dnclist); ncp != NULL; |
| 1046 | ncp = ncnext) { |
| 1047 | ncnext = LIST_NEXT(ncp, nc_dvlist); |
| 1048 | mutex_enter(&ncp->nc_lock); |
| 1049 | cache_invalidate(ncp); |
| 1050 | mutex_exit(&ncp->nc_lock); |
| 1051 | cache_disassociate(ncp); |
| 1052 | } |
| 1053 | } |
| 1054 | if (name != NULL) { |
| 1055 | SDT_PROBE(vfs, namecache, purge, name, name, namelen, 0, 0, 0); |
| 1056 | ncp = cache_lookup_entry(vp, name, namelen); |
| 1057 | if (ncp) { |
| 1058 | cache_invalidate(ncp); |
| 1059 | mutex_exit(&ncp->nc_lock); |
| 1060 | cache_disassociate(ncp); |
| 1061 | } |
| 1062 | } |
| 1063 | mutex_exit(namecache_lock); |
| 1064 | } |
| 1065 | |
| 1066 | /* |
| 1067 | * Cache flush, a whole filesystem; called when filesys is umounted to |
| 1068 | * remove entries that would now be invalid. |
| 1069 | */ |
| 1070 | void |
| 1071 | cache_purgevfs(struct mount *mp) |
| 1072 | { |
| 1073 | struct namecache *ncp, *nxtcp; |
| 1074 | |
| 1075 | SDT_PROBE(vfs, namecache, purge, vfs, mp, 0, 0, 0, 0); |
| 1076 | mutex_enter(namecache_lock); |
| 1077 | for (ncp = TAILQ_FIRST(&nclruhead); ncp != NULL; ncp = nxtcp) { |
| 1078 | nxtcp = TAILQ_NEXT(ncp, nc_lru); |
| 1079 | mutex_enter(&ncp->nc_lock); |
| 1080 | if (ncp->nc_dvp != NULL && ncp->nc_dvp->v_mount == mp) { |
| 1081 | /* Free the resources we had. */ |
| 1082 | cache_invalidate(ncp); |
| 1083 | cache_disassociate(ncp); |
| 1084 | } |
| 1085 | mutex_exit(&ncp->nc_lock); |
| 1086 | } |
| 1087 | cache_reclaim(); |
| 1088 | mutex_exit(namecache_lock); |
| 1089 | } |
| 1090 | |
| 1091 | /* |
| 1092 | * Scan global list invalidating entries until we meet a preset target. |
| 1093 | * Prefer to invalidate entries that have not scored a hit within |
| 1094 | * cache_hottime seconds. We sort the LRU list only for this routine's |
| 1095 | * benefit. |
| 1096 | */ |
| 1097 | static void |
| 1098 | cache_prune(int incache, int target) |
| 1099 | { |
| 1100 | struct namecache *ncp, *nxtcp, *sentinel; |
| 1101 | int items, recent, tryharder; |
| 1102 | |
| 1103 | KASSERT(mutex_owned(namecache_lock)); |
| 1104 | |
| 1105 | SDT_PROBE(vfs, namecache, prune, done, incache, target, 0, 0, 0); |
| 1106 | items = 0; |
| 1107 | tryharder = 0; |
| 1108 | recent = hardclock_ticks - hz * cache_hottime; |
| 1109 | sentinel = NULL; |
| 1110 | for (ncp = TAILQ_FIRST(&nclruhead); ncp != NULL; ncp = nxtcp) { |
| 1111 | if (incache <= target) |
| 1112 | break; |
| 1113 | items++; |
| 1114 | nxtcp = TAILQ_NEXT(ncp, nc_lru); |
| 1115 | if (ncp == sentinel) { |
| 1116 | /* |
| 1117 | * If we looped back on ourself, then ignore |
| 1118 | * recent entries and purge whatever we find. |
| 1119 | */ |
| 1120 | tryharder = 1; |
| 1121 | } |
| 1122 | if (ncp->nc_dvp == NULL) |
| 1123 | continue; |
| 1124 | if (!tryharder && (ncp->nc_hittime - recent) > 0) { |
| 1125 | if (sentinel == NULL) |
| 1126 | sentinel = ncp; |
| 1127 | TAILQ_REMOVE(&nclruhead, ncp, nc_lru); |
| 1128 | TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru); |
| 1129 | continue; |
| 1130 | } |
| 1131 | mutex_enter(&ncp->nc_lock); |
| 1132 | if (ncp->nc_dvp != NULL) { |
| 1133 | cache_invalidate(ncp); |
| 1134 | cache_disassociate(ncp); |
| 1135 | incache--; |
| 1136 | } |
| 1137 | mutex_exit(&ncp->nc_lock); |
| 1138 | } |
| 1139 | cache_ev_scan.ev_count += items; |
| 1140 | } |
| 1141 | |
| 1142 | /* |
| 1143 | * Collect dead cache entries from all CPUs and garbage collect. |
| 1144 | */ |
| 1145 | static void |
| 1146 | cache_reclaim(void) |
| 1147 | { |
| 1148 | struct namecache *ncp, *next; |
| 1149 | int items; |
| 1150 | |
| 1151 | KASSERT(mutex_owned(namecache_lock)); |
| 1152 | |
| 1153 | /* |
| 1154 | * If the number of extant entries not awaiting garbage collection |
| 1155 | * exceeds the high water mark, then reclaim stale entries until we |
| 1156 | * reach our low water mark. |
| 1157 | */ |
| 1158 | items = numcache - cache_gcpend; |
| 1159 | if (items > (uint64_t)desiredvnodes * cache_hiwat / 100) { |
| 1160 | cache_prune(items, (int)((uint64_t)desiredvnodes * |
| 1161 | cache_lowat / 100)); |
| 1162 | cache_ev_over.ev_count++; |
| 1163 | } else |
| 1164 | cache_ev_under.ev_count++; |
| 1165 | |
| 1166 | /* |
| 1167 | * Stop forward lookup activity on all CPUs and garbage collect dead |
| 1168 | * entries. |
| 1169 | */ |
| 1170 | cache_lock_cpus(); |
| 1171 | ncp = cache_gcqueue; |
| 1172 | cache_gcqueue = NULL; |
| 1173 | items = cache_gcpend; |
| 1174 | cache_gcpend = 0; |
| 1175 | while (ncp != NULL) { |
| 1176 | next = ncp->nc_gcqueue; |
| 1177 | cache_disassociate(ncp); |
| 1178 | KASSERT(ncp->nc_dvp == NULL); |
| 1179 | if (ncp->nc_hash.le_prev != NULL) { |
| 1180 | LIST_REMOVE(ncp, nc_hash); |
| 1181 | ncp->nc_hash.le_prev = NULL; |
| 1182 | } |
| 1183 | pool_cache_put(namecache_cache, ncp); |
| 1184 | ncp = next; |
| 1185 | } |
| 1186 | cache_unlock_cpus(); |
| 1187 | numcache -= items; |
| 1188 | cache_ev_gc.ev_count += items; |
| 1189 | } |
| 1190 | |
| 1191 | /* |
| 1192 | * Cache maintainence thread, awakening once per second to: |
| 1193 | * |
| 1194 | * => keep number of entries below the high water mark |
| 1195 | * => sort pseudo-LRU list |
| 1196 | * => garbage collect dead entries |
| 1197 | */ |
| 1198 | static void |
| 1199 | cache_thread(void *arg) |
| 1200 | { |
| 1201 | |
| 1202 | mutex_enter(namecache_lock); |
| 1203 | for (;;) { |
| 1204 | cache_reclaim(); |
| 1205 | kpause("cachegc" , false, hz, namecache_lock); |
| 1206 | } |
| 1207 | } |
| 1208 | |
| 1209 | #ifdef DDB |
| 1210 | void |
| 1211 | namecache_print(struct vnode *vp, void (*pr)(const char *, ...)) |
| 1212 | { |
| 1213 | struct vnode *dvp = NULL; |
| 1214 | struct namecache *ncp; |
| 1215 | |
| 1216 | TAILQ_FOREACH(ncp, &nclruhead, nc_lru) { |
| 1217 | if (ncp->nc_vp == vp && ncp->nc_dvp != NULL) { |
| 1218 | (*pr)("name %.*s\n" , ncp->nc_nlen, ncp->nc_name); |
| 1219 | dvp = ncp->nc_dvp; |
| 1220 | } |
| 1221 | } |
| 1222 | if (dvp == NULL) { |
| 1223 | (*pr)("name not found\n" ); |
| 1224 | return; |
| 1225 | } |
| 1226 | vp = dvp; |
| 1227 | TAILQ_FOREACH(ncp, &nclruhead, nc_lru) { |
| 1228 | if (ncp->nc_vp == vp) { |
| 1229 | (*pr)("parent %.*s\n" , ncp->nc_nlen, ncp->nc_name); |
| 1230 | } |
| 1231 | } |
| 1232 | } |
| 1233 | #endif |
| 1234 | |
| 1235 | void |
| 1236 | namecache_count_pass2(void) |
| 1237 | { |
| 1238 | struct nchcpu *cpup = curcpu()->ci_data.cpu_nch; |
| 1239 | |
| 1240 | COUNT_UNL(cpup, ncs_pass2); |
| 1241 | } |
| 1242 | |
| 1243 | void |
| 1244 | namecache_count_2passes(void) |
| 1245 | { |
| 1246 | struct nchcpu *cpup = curcpu()->ci_data.cpu_nch; |
| 1247 | |
| 1248 | COUNT_UNL(cpup, ncs_2passes); |
| 1249 | } |
| 1250 | |
| 1251 | /* |
| 1252 | * Fetch the current values of the stats. We return the most |
| 1253 | * recent values harvested into nchstats by cache_reclaim(), which |
| 1254 | * will be less than a second old. |
| 1255 | */ |
| 1256 | static int |
| 1257 | cache_stat_sysctl(SYSCTLFN_ARGS) |
| 1258 | { |
| 1259 | struct nchstats stats; |
| 1260 | struct nchcpu *my_cpup; |
| 1261 | #ifdef CACHE_STATS_CURRENT |
| 1262 | CPU_INFO_ITERATOR cii; |
| 1263 | struct cpu_info *ci; |
| 1264 | #endif /* CACHE_STATS_CURRENT */ |
| 1265 | |
| 1266 | if (oldp == NULL) { |
| 1267 | *oldlenp = sizeof(stats); |
| 1268 | return 0; |
| 1269 | } |
| 1270 | |
| 1271 | if (*oldlenp < sizeof(stats)) { |
| 1272 | *oldlenp = 0; |
| 1273 | return 0; |
| 1274 | } |
| 1275 | |
| 1276 | /* |
| 1277 | * Take this CPU's per-cpu lock to hold off cache_reclaim() |
| 1278 | * from doing a stats update while doing minimal damage to |
| 1279 | * concurrent operations. |
| 1280 | */ |
| 1281 | sysctl_unlock(); |
| 1282 | my_cpup = curcpu()->ci_data.cpu_nch; |
| 1283 | mutex_enter(&my_cpup->cpu_lock); |
| 1284 | stats = nchstats; |
| 1285 | #ifdef CACHE_STATS_CURRENT |
| 1286 | for (CPU_INFO_FOREACH(cii, ci)) { |
| 1287 | struct nchcpu *cpup = ci->ci_data.cpu_nch; |
| 1288 | |
| 1289 | ADD(stats, cpup, ncs_goodhits); |
| 1290 | ADD(stats, cpup, ncs_neghits); |
| 1291 | ADD(stats, cpup, ncs_badhits); |
| 1292 | ADD(stats, cpup, ncs_falsehits); |
| 1293 | ADD(stats, cpup, ncs_miss); |
| 1294 | ADD(stats, cpup, ncs_long); |
| 1295 | ADD(stats, cpup, ncs_pass2); |
| 1296 | ADD(stats, cpup, ncs_2passes); |
| 1297 | ADD(stats, cpup, ncs_revhits); |
| 1298 | ADD(stats, cpup, ncs_revmiss); |
| 1299 | } |
| 1300 | #endif /* CACHE_STATS_CURRENT */ |
| 1301 | mutex_exit(&my_cpup->cpu_lock); |
| 1302 | sysctl_relock(); |
| 1303 | |
| 1304 | *oldlenp = sizeof(stats); |
| 1305 | return sysctl_copyout(l, &stats, oldp, sizeof(stats)); |
| 1306 | } |
| 1307 | |
| 1308 | static void |
| 1309 | sysctl_cache_stat_setup(void) |
| 1310 | { |
| 1311 | |
| 1312 | KASSERT(sysctllog == NULL); |
| 1313 | sysctl_createv(&sysctllog, 0, NULL, NULL, |
| 1314 | CTLFLAG_PERMANENT, |
| 1315 | CTLTYPE_STRUCT, "namecache_stats" , |
| 1316 | SYSCTL_DESCR("namecache statistics" ), |
| 1317 | cache_stat_sysctl, 0, NULL, 0, |
| 1318 | CTL_VFS, CTL_CREATE, CTL_EOL); |
| 1319 | } |
| 1320 | |