X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/743b15655a24ee3fe9f458f383003e011db0558f..3903760236c30e3b5ace7a4eefac3a269d68957c:/bsd/vfs/vfs_cache.c?ds=sidebyside diff --git a/bsd/vfs/vfs_cache.c b/bsd/vfs/vfs_cache.c index fe7fe5269..b7876475d 100644 --- a/bsd/vfs/vfs_cache.c +++ b/bsd/vfs/vfs_cache.c @@ -1,23 +1,29 @@ /* - * Copyright (c) 2000-2003 Apple Computer, Inc. All rights reserved. + * Copyright (c) 2000-2015 Apple Inc. All rights reserved. * - * @APPLE_LICENSE_HEADER_START@ + * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ * - * The contents of this file constitute Original Code as defined in and - * are subject to the Apple Public Source License Version 1.1 (the - * "License"). You may not use this file except in compliance with the - * License. Please obtain a copy of the License at - * http://www.apple.com/publicsource and read it before using this file. + * This file contains Original Code and/or Modifications of Original Code + * as defined in and that are subject to the Apple Public Source License + * Version 2.0 (the 'License'). You may not use this file except in + * compliance with the License. The rights granted to you under the License + * may not be used to create, or enable the creation or redistribution of, + * unlawful or unlicensed copies of an Apple operating system, or to + * circumvent, violate, or enable the circumvention or violation of, any + * terms of an Apple operating system software license agreement. * - * This Original Code and all software distributed under the License are - * distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY KIND, EITHER + * Please obtain a copy of the License at + * http://www.opensource.apple.com/apsl/ and read it before using this file. + * + * The Original Code and all software distributed under the License are + * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES, * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE OR NON-INFRINGEMENT. Please see the - * License for the specific language governing rights and limitations - * under the License. + * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT. + * Please see the License for the specific language governing rights and + * limitations under the License. * - * @APPLE_LICENSE_HEADER_END@ + * @APPLE_OSREFERENCE_LICENSE_HEADER_END@ */ /* Copyright (c) 1995 NeXT Computer, Inc. All Rights Reserved */ /* @@ -58,16 +64,28 @@ * * @(#)vfs_cache.c 8.5 (Berkeley) 3/22/95 */ +/* + * NOTICE: This file was modified by SPARTA, Inc. in 2005 to introduce + * support for mandatory and extensible security protections. This notice + * is included in support of clause 2.2 (b) of the Apple Public License, + * Version 2.0. + */ #include #include #include #include #include +#include #include #include #include #include #include +#include + +#if CONFIG_MACF +#include +#endif /* * Name caching works as follows: @@ -96,22 +114,62 @@ u_long nchash; /* size of hash table - 1 */ long numcache; /* number of cache entries allocated */ int desiredNodes; int desiredNegNodes; +int ncs_negtotal; +int nc_disabled = 0; TAILQ_HEAD(, namecache) nchead; /* chain of all name cache entries */ TAILQ_HEAD(, namecache) neghead; /* chain of only negative cache entries */ + + +#if COLLECT_STATS + struct nchstats nchstats; /* cache effectiveness statistics */ +#define NCHSTAT(v) { \ + nchstats.v++; \ +} +#define NAME_CACHE_LOCK() name_cache_lock() +#define NAME_CACHE_UNLOCK() name_cache_unlock() +#define NAME_CACHE_LOCK_SHARED() name_cache_lock() + +#else + +#define NCHSTAT(v) +#define NAME_CACHE_LOCK() name_cache_lock() +#define NAME_CACHE_UNLOCK() name_cache_unlock() +#define NAME_CACHE_LOCK_SHARED() name_cache_lock_shared() + +#endif + + /* vars for name cache list lock */ lck_grp_t * namecache_lck_grp; lck_grp_attr_t * namecache_lck_grp_attr; lck_attr_t * namecache_lck_attr; -lck_mtx_t * namecache_mtx_lock; + +lck_grp_t * strcache_lck_grp; +lck_grp_attr_t * strcache_lck_grp_attr; +lck_attr_t * strcache_lck_attr; + +lck_rw_t * namecache_rw_lock; +lck_rw_t * strtable_rw_lock; + +#define NUM_STRCACHE_LOCKS 1024 + +lck_mtx_t strcache_mtx_locks[NUM_STRCACHE_LOCKS]; + static vnode_t cache_lookup_locked(vnode_t dvp, struct componentname *cnp); -static int remove_name_locked(const char *); -static char *add_name_locked(const char *, size_t, u_int, u_int); +static const char *add_name_internal(const char *, uint32_t, u_int, boolean_t, u_int); static void init_string_table(void); static void cache_delete(struct namecache *, int); -static void dump_string_table(void); +static void cache_enter_locked(vnode_t dvp, vnode_t vp, struct componentname *cnp, const char *strname); + +#ifdef DUMP_STRING_TABLE +/* + * Internal dump function used for debugging + */ +void dump_string_table(void); +#endif /* DUMP_STRING_TABLE */ static void init_crc32(void); static unsigned int crc32tab[256]; @@ -120,114 +178,571 @@ static unsigned int crc32tab[256]; #define NCHHASH(dvp, hash_val) \ (&nchashtbl[(dvp->v_id ^ (hash_val)) & nchashmask]) +/* + * This function tries to check if a directory vp is a subdirectory of dvp + * only from valid v_parent pointers. It is called with the name cache lock + * held and does not drop the lock anytime inside the function. + * + * It returns a boolean that indicates whether or not it was able to + * successfully infer the parent/descendent relationship via the v_parent + * pointers, or if it could not infer such relationship and that the decision + * must be delegated to the owning filesystem. + * + * If it does not defer the decision, i.e. it was successfuly able to determine + * the parent/descendent relationship, *is_subdir tells the caller if vp is a + * subdirectory of dvp. + * + * If the decision is deferred, *next_vp is where it stopped i.e. *next_vp + * is the vnode whose parent is to be determined from the filesystem. + * *is_subdir, in this case, is not indicative of anything and should be + * ignored. + * + * The return value and output args should be used as follows : + * + * defer = cache_check_vnode_issubdir(vp, dvp, is_subdir, next_vp); + * if (!defer) { + * if (*is_subdir) + * vp is subdirectory; + * else + * vp is not a subdirectory; + * } else { + * if (*next_vp) + * check this vnode's parent from the filesystem + * else + * error (likely because of forced unmount). + * } + * + */ +static boolean_t +cache_check_vnode_issubdir(vnode_t vp, vnode_t dvp, boolean_t *is_subdir, + vnode_t *next_vp) +{ + vnode_t tvp = vp; + int defer = FALSE; + *is_subdir = FALSE; + *next_vp = NULLVP; + while (1) { + mount_t tmp; -// -// This function builds the path to a filename in "buff". The -// length of the buffer *INCLUDING* the trailing zero byte is -// returned in outlen. NOTE: the length includes the trailing -// zero byte and thus the length is one greater than what strlen -// would return. This is important and lots of code elsewhere -// in the kernel assumes this behavior. -// + if (tvp == dvp) { + *is_subdir = TRUE; + break; + } else if (tvp == rootvnode) { + /* *is_subdir = FALSE */ + break; + } + + tmp = tvp->v_mount; + while ((tvp->v_flag & VROOT) && tmp && tmp->mnt_vnodecovered && + tvp != dvp && tvp != rootvnode) { + tvp = tmp->mnt_vnodecovered; + tmp = tvp->v_mount; + } + + /* + * If dvp is not at the top of a mount "stack" then + * vp is not a subdirectory of dvp either. + */ + if (tvp == dvp || tvp == rootvnode) { + /* *is_subdir = FALSE */ + break; + } + + if (!tmp) { + defer = TRUE; + *next_vp = NULLVP; + break; + } + + if ((tvp->v_flag & VISHARDLINK) || !(tvp->v_parent)) { + defer = TRUE; + *next_vp = tvp; + break; + } + + tvp = tvp->v_parent; + } + + return (defer); +} + +/* maximum times retry from potentially transient errors in vnode_issubdir */ +#define MAX_ERROR_RETRY 3 + +/* + * This function checks if a given directory (vp) is a subdirectory of dvp. + * It walks backwards from vp and if it hits dvp in its parent chain, + * it is a subdirectory. If it encounters the root directory, it is not + * a subdirectory. + * + * This function returns an error if it is unsuccessful and 0 on success. + * + * On entry (and exit) vp has an iocount and if this function has to take + * any iocounts on other vnodes in the parent chain traversal, it releases them. + */ int -build_path(vnode_t first_vp, char *buff, int buflen, int *outlen) +vnode_issubdir(vnode_t vp, vnode_t dvp, int *is_subdir, vfs_context_t ctx) { - vnode_t vp = first_vp; - char *end, *str; - int len, ret=0, counter=0; + vnode_t start_vp, tvp; + vnode_t vp_with_iocount; + int error = 0; + char dotdotbuf[] = ".."; + int error_retry_count = 0; /* retry count for potentially transient + errors */ + + *is_subdir = FALSE; + tvp = start_vp = vp; + /* + * Anytime we acquire an iocount in this function, we save the vnode + * in this variable and release it before exiting. + */ + vp_with_iocount = NULLVP; + + while (1) { + boolean_t defer; + vnode_t pvp; + uint32_t vid; + struct componentname cn; + boolean_t is_subdir_locked = FALSE; + + if (tvp == dvp) { + *is_subdir = TRUE; + break; + } else if (tvp == rootvnode) { + /* *is_subdir = FALSE */ + break; + } + + NAME_CACHE_LOCK_SHARED(); + + defer = cache_check_vnode_issubdir(tvp, dvp, &is_subdir_locked, + &tvp); + + if (defer && tvp) + vid = vnode_vid(tvp); + + NAME_CACHE_UNLOCK(); + + if (!defer) { + *is_subdir = is_subdir_locked; + break; + } + + if (!tvp) { + if (error_retry_count++ < MAX_ERROR_RETRY) { + tvp = vp; + continue; + } + error = ENOENT; + break; + } + + if (tvp != start_vp) { + if (vp_with_iocount) { + vnode_put(vp_with_iocount); + vp_with_iocount = NULLVP; + } + + error = vnode_getwithvid(tvp, vid); + if (error) { + if (error_retry_count++ < MAX_ERROR_RETRY) { + tvp = vp; + error = 0; + continue; + } + break; + } + + vp_with_iocount = tvp; + } + + bzero(&cn, sizeof(cn)); + cn.cn_nameiop = LOOKUP; + cn.cn_flags = ISLASTCN | ISDOTDOT; + cn.cn_context = ctx; + cn.cn_pnbuf = &dotdotbuf[0]; + cn.cn_pnlen = sizeof(dotdotbuf); + cn.cn_nameptr = cn.cn_pnbuf; + cn.cn_namelen = 2; + + pvp = NULLVP; + if ((error = VNOP_LOOKUP(tvp, &pvp, &cn, ctx))) + break; + + if (!(tvp->v_flag & VISHARDLINK) && tvp->v_parent != pvp) { + (void)vnode_update_identity(tvp, pvp, NULL, 0, 0, + VNODE_UPDATE_PARENT); + } + + if (vp_with_iocount) + vnode_put(vp_with_iocount); + + vp_with_iocount = tvp = pvp; + } + + if (vp_with_iocount) + vnode_put(vp_with_iocount); + + return (error); +} + +/* + * This function builds the path to a filename in "buff". The + * length of the buffer *INCLUDING* the trailing zero byte is + * returned in outlen. NOTE: the length includes the trailing + * zero byte and thus the length is one greater than what strlen + * would return. This is important and lots of code elsewhere + * in the kernel assumes this behavior. + * + * This function can call vnop in file system if the parent vnode + * does not exist or when called for hardlinks via volfs path. + * If BUILDPATH_NO_FS_ENTER is set in flags, it only uses values present + * in the name cache and does not enter the file system. + * + * If BUILDPATH_CHECK_MOVED is set in flags, we return EAGAIN when + * we encounter ENOENT during path reconstruction. ENOENT means that + * one of the parents moved while we were building the path. The + * caller can special handle this case by calling build_path again. + * + * If BUILDPATH_VOLUME_RELATIVE is set in flags, we return path + * that is relative to the nearest mount point, i.e. do not + * cross over mount points during building the path. + * + * passed in vp must have a valid io_count reference + */ +int +build_path(vnode_t first_vp, char *buff, int buflen, int *outlen, int flags, vfs_context_t ctx) +{ + vnode_t vp, tvp; + vnode_t vp_with_iocount; + vnode_t proc_root_dir_vp; + char *end; + const char *str; + int len; + int ret = 0; + int fixhardlink; + + if (first_vp == NULLVP) + return (EINVAL); + + if (buflen <= 1) + return (ENOSPC); + + /* + * Grab the process fd so we can evaluate fd_rdir. + */ + if (vfs_context_proc(ctx)->p_fd) + proc_root_dir_vp = vfs_context_proc(ctx)->p_fd->fd_rdir; + else + proc_root_dir_vp = NULL; + + vp_with_iocount = NULLVP; +again: + vp = first_vp; end = &buff[buflen-1]; *end = '\0'; /* - * if this is the root dir of a file system... + * holding the NAME_CACHE_LOCK in shared mode is + * sufficient to stabilize both the vp->v_parent chain + * and the 'vp->v_mount->mnt_vnodecovered' chain + * + * if we need to drop this lock, we must first grab the v_id + * from the vnode we're currently working with... if that + * vnode doesn't already have an io_count reference (the vp + * passed in comes with one), we must grab a reference + * after we drop the NAME_CACHE_LOCK via vnode_getwithvid... + * deadlocks may result if you call vnode_get while holding + * the NAME_CACHE_LOCK... we lazily release the reference + * we pick up the next time we encounter a need to drop + * the NAME_CACHE_LOCK or before we return from this routine */ - if (vp && (vp->v_flag & VROOT) && vp->v_mount) { - /* - * then if it's the root fs, just put in a '/' and get out of here - */ - if (vp->v_mount->mnt_flag & MNT_ROOTFS) { + NAME_CACHE_LOCK_SHARED(); + + /* + * Check if this is the root of a file system. + */ + while (vp && vp->v_flag & VROOT) { + if (vp->v_mount == NULL) { + ret = EINVAL; + goto out_unlock; + } + if ((vp->v_mount->mnt_flag & MNT_ROOTFS) || (vp == proc_root_dir_vp)) { + /* + * It's the root of the root file system, so it's + * just "/". + */ *--end = '/'; - goto out; + + goto out_unlock; } else { - /* - * else just use the covered vnode to get the mount path + /* + * This the root of the volume and the caller does not + * want to cross mount points. Therefore just return + * '/' as the relative path. */ - vp = vp->v_mount->mnt_vnodecovered; + if (flags & BUILDPATH_VOLUME_RELATIVE) { + *--end = '/'; + goto out_unlock; + } else { + vp = vp->v_mount->mnt_vnodecovered; + } } } - name_cache_lock(); - while (vp && vp->v_parent != vp) { - /* - * the maximum depth of a file system hierarchy is MAXPATHLEN/2 - * (with single-char names separated by slashes). we panic if - * we've ever looped more than that. - */ - if (counter++ > MAXPATHLEN/2) { - panic("build_path: vnode parent chain is too long! vp 0x%x\n", vp); - } - str = vp->v_name; + while ((vp != NULLVP) && (vp->v_parent != vp)) { + int vid; - if (str == NULL) { - if (vp->v_parent != NULL) { - ret = EINVAL; + /* + * For hardlinks the v_name may be stale, so if its OK + * to enter a file system, ask the file system for the + * name and parent (below). + */ + fixhardlink = (vp->v_flag & VISHARDLINK) && + (vp->v_mount->mnt_kern_flag & MNTK_PATH_FROM_ID) && + !(flags & BUILDPATH_NO_FS_ENTER); + + if (!fixhardlink) { + str = vp->v_name; + + if (str == NULL || *str == '\0') { + if (vp->v_parent != NULL) + ret = EINVAL; + else + ret = ENOENT; + goto out_unlock; } - break; + len = strlen(str); + /* + * Check that there's enough space (including space for the '/') + */ + if ((end - buff) < (len + 1)) { + ret = ENOSPC; + goto out_unlock; + } + /* + * Copy the name backwards. + */ + str += len; + + for (; len > 0; len--) + *--end = *--str; + /* + * Add a path separator. + */ + *--end = '/'; } - len = strlen(str); /* - * check that there's enough space (make sure to include space for the '/') + * Walk up the parent chain. */ - if ((end - buff) < (len + 1)) { - ret = ENOSPC; - break; + if (((vp->v_parent != NULLVP) && !fixhardlink) || + (flags & BUILDPATH_NO_FS_ENTER)) { + + /* + * In this if () block we are not allowed to enter the filesystem + * to conclusively get the most accurate parent identifier. + * As a result, if 'vp' does not identify '/' and it + * does not have a valid v_parent, then error out + * and disallow further path construction + */ + if ((vp->v_parent == NULLVP) && (rootvnode != vp)) { + /* + * Only '/' is allowed to have a NULL parent + * pointer. Upper level callers should ideally + * re-drive name lookup on receiving a ENOENT. + */ + ret = ENOENT; + + /* The code below will exit early if 'tvp = vp' == NULL */ + } + vp = vp->v_parent; + + /* + * if the vnode we have in hand isn't a directory and it + * has a v_parent, then we started with the resource fork + * so skip up to avoid getting a duplicate copy of the + * file name in the path. + */ + if (vp && !vnode_isdir(vp) && vp->v_parent) { + vp = vp->v_parent; + } + } else { + /* + * No parent, go get it if supported. + */ + struct vnode_attr va; + vnode_t dvp; + + /* + * Make sure file system supports obtaining a path from id. + */ + if (!(vp->v_mount->mnt_kern_flag & MNTK_PATH_FROM_ID)) { + ret = ENOENT; + goto out_unlock; + } + vid = vp->v_id; + + NAME_CACHE_UNLOCK(); + + if (vp != first_vp && vp != vp_with_iocount) { + if (vp_with_iocount) { + vnode_put(vp_with_iocount); + vp_with_iocount = NULLVP; + } + if (vnode_getwithvid(vp, vid)) + goto again; + vp_with_iocount = vp; + } + VATTR_INIT(&va); + VATTR_WANTED(&va, va_parentid); + + if (fixhardlink) { + VATTR_WANTED(&va, va_name); + MALLOC_ZONE(va.va_name, caddr_t, MAXPATHLEN, M_NAMEI, M_WAITOK); + } else { + va.va_name = NULL; + } + /* + * Ask the file system for its parent id and for its name (optional). + */ + ret = vnode_getattr(vp, &va, ctx); + + if (fixhardlink) { + if ((ret == 0) && (VATTR_IS_SUPPORTED(&va, va_name))) { + str = va.va_name; + vnode_update_identity(vp, NULL, str, strlen(str), 0, VNODE_UPDATE_NAME); + } else if (vp->v_name) { + str = vp->v_name; + ret = 0; + } else { + ret = ENOENT; + goto bad_news; + } + len = strlen(str); + + /* + * Check that there's enough space. + */ + if ((end - buff) < (len + 1)) { + ret = ENOSPC; + } else { + /* Copy the name backwards. */ + str += len; + + for (; len > 0; len--) { + *--end = *--str; + } + /* + * Add a path separator. + */ + *--end = '/'; + } +bad_news: + FREE_ZONE(va.va_name, MAXPATHLEN, M_NAMEI); + } + if (ret || !VATTR_IS_SUPPORTED(&va, va_parentid)) { + ret = ENOENT; + goto out; + } + /* + * Ask the file system for the parent vnode. + */ + if ((ret = VFS_VGET(vp->v_mount, (ino64_t)va.va_parentid, &dvp, ctx))) + goto out; + + if (!fixhardlink && (vp->v_parent != dvp)) + vnode_update_identity(vp, dvp, NULL, 0, 0, VNODE_UPDATE_PARENT); + + if (vp_with_iocount) + vnode_put(vp_with_iocount); + vp = dvp; + vp_with_iocount = vp; + + NAME_CACHE_LOCK_SHARED(); + + /* + * if the vnode we have in hand isn't a directory and it + * has a v_parent, then we started with the resource fork + * so skip up to avoid getting a duplicate copy of the + * file name in the path. + */ + if (vp && !vnode_isdir(vp) && vp->v_parent) + vp = vp->v_parent; } - /* - * copy it backwards - */ - str += len; - for (; len > 0; len--) { - *--end = *--str; + if (vp && (flags & BUILDPATH_CHECKACCESS)) { + vid = vp->v_id; + + NAME_CACHE_UNLOCK(); + + if (vp != first_vp && vp != vp_with_iocount) { + if (vp_with_iocount) { + vnode_put(vp_with_iocount); + vp_with_iocount = NULLVP; + } + if (vnode_getwithvid(vp, vid)) + goto again; + vp_with_iocount = vp; + } + if ((ret = vnode_authorize(vp, NULL, KAUTH_VNODE_SEARCH, ctx))) + goto out; /* no peeking */ + + NAME_CACHE_LOCK_SHARED(); } - /* - * put in the path separator - */ - *--end = '/'; /* - * walk up the chain (as long as we're not the root) + * When a mount point is crossed switch the vp. + * Continue until we find the root or we find + * a vnode that's not the root of a mounted + * file system. */ - if (vp == first_vp && (vp->v_flag & VROOT)) { - if (vp->v_mount && vp->v_mount->mnt_vnodecovered) { - vp = vp->v_mount->mnt_vnodecovered->v_parent; + tvp = vp; + + while (tvp) { + if (tvp == proc_root_dir_vp) + goto out_unlock; /* encountered the root */ + + if (!(tvp->v_flag & VROOT) || !tvp->v_mount) + break; /* not the root of a mounted FS */ + + if (flags & BUILDPATH_VOLUME_RELATIVE) { + /* Do not cross over mount points */ + tvp = NULL; } else { - vp = NULLVP; + tvp = tvp->v_mount->mnt_vnodecovered; } - } else { - vp = vp->v_parent; - } - /* - * check if we're crossing a mount point and - * switch the vp if we are. - */ - if (vp && (vp->v_flag & VROOT) && vp->v_mount) { - vp = vp->v_mount->mnt_vnodecovered; } + if (tvp == NULLVP) + goto out_unlock; + vp = tvp; } - name_cache_unlock(); +out_unlock: + NAME_CACHE_UNLOCK(); out: + if (vp_with_iocount) + vnode_put(vp_with_iocount); /* - * slide it down to the beginning of the buffer + * Slide the name down to the beginning of the buffer. */ memmove(buff, end, &buff[buflen] - end); - - *outlen = &buff[buflen] - end; // length includes the trailing zero byte + + /* + * length includes the trailing zero byte + */ + *outlen = &buff[buflen] - end; - return ret; + /* One of the parents was moved during path reconstruction. + * The caller is interested in knowing whether any of the + * parents moved via BUILDPATH_CHECK_MOVED, so return EAGAIN. + */ + if ((ret == ENOENT) && (flags & BUILDPATH_CHECK_MOVED)) { + ret = EAGAIN; + } + + return (ret); } @@ -242,7 +757,7 @@ vnode_getparent(vnode_t vp) vnode_t pvp = NULLVP; int pvid; - name_cache_lock(); + NAME_CACHE_LOCK_SHARED(); /* * v_parent is stable behind the name_cache lock * however, the only thing we can really guarantee @@ -253,40 +768,80 @@ vnode_getparent(vnode_t vp) if ( (pvp = vp->v_parent) != NULLVP ) { pvid = pvp->v_id; - name_cache_unlock(); + NAME_CACHE_UNLOCK(); if (vnode_getwithvid(pvp, pvid) != 0) pvp = NULL; } else - name_cache_unlock(); - + NAME_CACHE_UNLOCK(); return (pvp); } -char * +const char * vnode_getname(vnode_t vp) { - char *name = NULL; - - name_cache_lock(); + const char *name = NULL; + + NAME_CACHE_LOCK_SHARED(); if (vp->v_name) - name = add_name_locked(vp->v_name, strlen(vp->v_name), 0, 0); - name_cache_unlock(); + name = vfs_addname(vp->v_name, strlen(vp->v_name), 0, 0); + NAME_CACHE_UNLOCK(); return (name); } void -vnode_putname(char *name) +vnode_putname(const char *name) { - name_cache_lock(); + vfs_removename(name); +} - remove_name_locked(name); +static const char unknown_vnodename[] = "(unknown vnode name)"; - name_cache_unlock(); +const char * +vnode_getname_printable(vnode_t vp) +{ + const char *name = vnode_getname(vp); + if (name != NULL) + return name; + + switch (vp->v_type) { + case VCHR: + case VBLK: + { + /* + * Create an artificial dev name from + * major and minor device number + */ + char dev_name[64]; + (void) snprintf(dev_name, sizeof(dev_name), + "%c(%u, %u)", VCHR == vp->v_type ? 'c':'b', + major(vp->v_rdev), minor(vp->v_rdev)); + /* + * Add the newly created dev name to the name + * cache to allow easier cleanup. Also, + * vfs_addname allocates memory for the new name + * and returns it. + */ + NAME_CACHE_LOCK_SHARED(); + name = vfs_addname(dev_name, strlen(dev_name), 0, 0); + NAME_CACHE_UNLOCK(); + return name; + } + default: + return unknown_vnodename; + } } +void +vnode_putname_printable(const char *name) +{ + if (name == unknown_vnodename) + return; + vnode_putname(name); +} + /* * if VNODE_UPDATE_PARENT, and we can take @@ -302,52 +857,115 @@ vnode_putname(char *name) * if VNODE_UPDATE_CACHE, flush the name cache entries associated with vp */ void -vnode_update_identity(vnode_t vp, vnode_t dvp, char *name, int name_len, int name_hashval, int flags) +vnode_update_identity(vnode_t vp, vnode_t dvp, const char *name, int name_len, uint32_t name_hashval, int flags) { struct namecache *ncp; vnode_t old_parentvp = NULLVP; - +#if NAMEDSTREAMS + int isstream = (vp->v_flag & VISNAMEDSTREAM); + int kusecountbumped = 0; +#endif + kauth_cred_t tcred = NULL; + const char *vname = NULL; + const char *tname = NULL; if (flags & VNODE_UPDATE_PARENT) { - if (dvp && vnode_ref(dvp) != 0) - dvp = NULLVP; - } else + if (dvp && vnode_ref(dvp) != 0) { + dvp = NULLVP; + } +#if NAMEDSTREAMS + /* Don't count a stream's parent ref during unmounts */ + if (isstream && dvp && (dvp != vp) && (dvp != vp->v_parent) && (dvp->v_type == VREG)) { + vnode_lock_spin(dvp); + ++dvp->v_kusecount; + kusecountbumped = 1; + vnode_unlock(dvp); + } +#endif + } else { dvp = NULLVP; - name_cache_lock(); + } + if ( (flags & VNODE_UPDATE_NAME) ) { + if (name != vp->v_name) { + if (name && *name) { + if (name_len == 0) + name_len = strlen(name); + tname = vfs_addname(name, name_len, name_hashval, 0); + } + } else + flags &= ~VNODE_UPDATE_NAME; + } + if ( (flags & (VNODE_UPDATE_PURGE | VNODE_UPDATE_PARENT | VNODE_UPDATE_CACHE | VNODE_UPDATE_NAME)) ) { + + NAME_CACHE_LOCK(); + + if ( (flags & VNODE_UPDATE_PURGE) ) { + + if (vp->v_parent) + vp->v_parent->v_nc_generation++; - if ( (flags & VNODE_UPDATE_NAME) && (name != vp->v_name) ) { - if (vp->v_name != NULL) { - remove_name_locked(vp->v_name); - vp->v_name = NULL; + while ( (ncp = LIST_FIRST(&vp->v_nclinks)) ) + cache_delete(ncp, 1); + + while ( (ncp = TAILQ_FIRST(&vp->v_ncchildren)) ) + cache_delete(ncp, 1); + + /* + * Use a temp variable to avoid kauth_cred_unref() while NAME_CACHE_LOCK is held + */ + tcred = vp->v_cred; + vp->v_cred = NOCRED; + vp->v_authorized_actions = 0; } - if (name && *name) { - if (name_len == 0) - name_len = strlen(name); - vp->v_name = add_name_locked(name, name_len, name_hashval, 0); + if ( (flags & VNODE_UPDATE_NAME) ) { + vname = vp->v_name; + vp->v_name = tname; } - } - if (flags & VNODE_UPDATE_PARENT) { - if (dvp != vp && dvp != vp->v_parent) { - old_parentvp = vp->v_parent; - vp->v_parent = dvp; - dvp = NULLVP; + if (flags & VNODE_UPDATE_PARENT) { + if (dvp != vp && dvp != vp->v_parent) { + old_parentvp = vp->v_parent; + vp->v_parent = dvp; + dvp = NULLVP; - if (old_parentvp) - flags |= VNODE_UPDATE_CACHE; + if (old_parentvp) + flags |= VNODE_UPDATE_CACHE; + } } - } - if (flags & VNODE_UPDATE_CACHE) { - while ( (ncp = LIST_FIRST(&vp->v_nclinks)) ) - cache_delete(ncp, 1); - } - name_cache_unlock(); + if (flags & VNODE_UPDATE_CACHE) { + while ( (ncp = LIST_FIRST(&vp->v_nclinks)) ) + cache_delete(ncp, 1); + } + NAME_CACHE_UNLOCK(); - if (dvp != NULLVP) + if (vname != NULL) + vfs_removename(vname); + + if (IS_VALID_CRED(tcred)) + kauth_cred_unref(&tcred); + } + if (dvp != NULLVP) { +#if NAMEDSTREAMS + /* Back-out the ref we took if we lost a race for vp->v_parent. */ + if (kusecountbumped) { + vnode_lock_spin(dvp); + if (dvp->v_kusecount > 0) + --dvp->v_kusecount; + vnode_unlock(dvp); + } +#endif vnode_rele(dvp); - + } if (old_parentvp) { struct uthread *ut; +#if NAMEDSTREAMS + if (isstream) { + vnode_lock_spin(old_parentvp); + if ((old_parentvp->v_type != VDIR) && (old_parentvp->v_kusecount > 0)) + --old_parentvp->v_kusecount; + vnode_unlock(old_parentvp); + } +#endif ut = get_bsdthread_info(current_thread()); /* @@ -362,8 +980,7 @@ vnode_update_identity(vnode_t vp, vnode_t dvp, char *name, int name_len, int nam while ( (vp = old_parentvp) != NULLVP ) { - vnode_lock(vp); - + vnode_lock_spin(vp); vnode_rele_internal(vp, 0, 0, 1); /* @@ -377,7 +994,7 @@ vnode_update_identity(vnode_t vp, vnode_t dvp, char *name, int name_len, int nam * we'll sit in this loop until we run into * a parent in this chain that is not in this state * - * make our check and the node_rele atomic + * make our check and the vnode_rele atomic * with respect to the current vnode we're working on * by holding the vnode lock * if vnode_rele deferred the vnode_reclaim and has put @@ -392,11 +1009,17 @@ vnode_update_identity(vnode_t vp, vnode_t dvp, char *name, int name_len, int nam * pull the parent pointer now so that when we do the * vnode_reclaim for each of the vnodes in the uu_vreclaims * list, we won't recurse back through here + * + * need to do a convert here in case vnode_rele_internal + * returns with the lock held in the spin mode... it + * can drop and retake the lock under certain circumstances */ - name_cache_lock(); + vnode_lock_convert(vp); + + NAME_CACHE_LOCK(); old_parentvp = vp->v_parent; vp->v_parent = NULLVP; - name_cache_unlock(); + NAME_CACHE_UNLOCK(); } else { /* * we're done... we ran into a vnode that isn't @@ -430,9 +1053,9 @@ vnode_update_identity(vnode_t vp, vnode_t dvp, char *name, int name_len, int nam * so that HFS can post-process the lookup. Also, volfs will call * VNOP_GETATTR2 to determine the parent, instead of using v_parent. */ -void vnode_set_hard_link(vnode_t vp) +void vnode_setmultipath(vnode_t vp) { - vnode_lock(vp); + vnode_lock_spin(vp); /* * In theory, we're changing the vnode's identity as far as the @@ -451,141 +1074,216 @@ void vnode_set_hard_link(vnode_t vp) } + +/* + * backwards compatibility + */ void vnode_uncache_credentials(vnode_t vp) { - kauth_cred_t ucred = NULL; + vnode_uncache_authorized_action(vp, KAUTH_INVALIDATE_CACHED_RIGHTS); +} - if (vp->v_cred) { - vnode_lock(vp); - ucred = vp->v_cred; - vp->v_cred = NULL; +/* + * use the exclusive form of NAME_CACHE_LOCK to protect the update of the + * following fields in the vnode: v_cred_timestamp, v_cred, v_authorized_actions + * we use this lock so that we can look at the v_cred and v_authorized_actions + * atomically while behind the NAME_CACHE_LOCK in shared mode in 'cache_lookup_path', + * which is the super-hot path... if we are updating the authorized actions for this + * vnode, we are already in the super-slow and far less frequented path so its not + * that bad that we take the lock exclusive for this case... of course we strive + * to hold it for the minimum amount of time possible + */ - vnode_unlock(vp); +void vnode_uncache_authorized_action(vnode_t vp, kauth_action_t action) +{ + kauth_cred_t tcred = NOCRED; + + NAME_CACHE_LOCK(); + + vp->v_authorized_actions &= ~action; - if (ucred) - kauth_cred_rele(ucred); + if (action == KAUTH_INVALIDATE_CACHED_RIGHTS && + IS_VALID_CRED(vp->v_cred)) { + /* + * Use a temp variable to avoid kauth_cred_unref() while NAME_CACHE_LOCK is held + */ + tcred = vp->v_cred; + vp->v_cred = NOCRED; } + NAME_CACHE_UNLOCK(); + + if (tcred != NOCRED) + kauth_cred_unref(&tcred); } -void vnode_cache_credentials(vnode_t vp, vfs_context_t context) +extern int bootarg_vnode_cache_defeat; /* default = 0, from bsd_init.c */ + +boolean_t +vnode_cache_is_authorized(vnode_t vp, vfs_context_t ctx, kauth_action_t action) { - kauth_cred_t ucred; - kauth_cred_t tcred = NOCRED; - struct timeval tv; + kauth_cred_t ucred; + boolean_t retval = FALSE; - ucred = vfs_context_ucred(context); + /* Boot argument to defeat rights caching */ + if (bootarg_vnode_cache_defeat) + return FALSE; - if (vp->v_cred != ucred || (vp->v_mount->mnt_kern_flag & MNTK_AUTH_OPAQUE)) { - vnode_lock(vp); + if ( (vp->v_mount->mnt_kern_flag & (MNTK_AUTH_OPAQUE | MNTK_AUTH_CACHE_TTL)) ) { + /* + * a TTL is enabled on the rights cache... handle it here + * a TTL of 0 indicates that no rights should be cached + */ + if (vp->v_mount->mnt_authcache_ttl) { + if ( !(vp->v_mount->mnt_kern_flag & MNTK_AUTH_CACHE_TTL) ) { + /* + * For filesystems marked only MNTK_AUTH_OPAQUE (generally network ones), + * we will only allow a SEARCH right on a directory to be cached... + * that cached right always has a default TTL associated with it + */ + if (action != KAUTH_VNODE_SEARCH || vp->v_type != VDIR) + vp = NULLVP; + } + if (vp != NULLVP && vnode_cache_is_stale(vp) == TRUE) { + vnode_uncache_authorized_action(vp, vp->v_authorized_actions); + vp = NULLVP; + } + } else + vp = NULLVP; + } + if (vp != NULLVP) { + ucred = vfs_context_ucred(ctx); - microuptime(&tv); - vp->v_cred_timestamp = tv.tv_sec; + NAME_CACHE_LOCK_SHARED(); - if (vp->v_cred != ucred) { - kauth_cred_ref(ucred); - - tcred = vp->v_cred; - vp->v_cred = ucred; - } - vnode_unlock(vp); - - if (tcred) - kauth_cred_rele(tcred); + if (vp->v_cred == ucred && (vp->v_authorized_actions & action) == action) + retval = TRUE; + + NAME_CACHE_UNLOCK(); } + return retval; } -/* reverse_lookup - lookup by walking back up the parent chain while leveraging - * use of the name cache lock in order to protect our starting vnode. - * NOTE - assumes you already have search access to starting point. - * returns 0 when we have reached the root, current working dir, or chroot root - * - */ -int -reverse_lookup(vnode_t start_vp, vnode_t *lookup_vpp, struct filedesc *fdp, vfs_context_t context, int *dp_authorized) + +void vnode_cache_authorized_action(vnode_t vp, vfs_context_t ctx, kauth_action_t action) { - int vid, done = 0; - int auth_opaque = 0; - vnode_t dp = start_vp; - vnode_t vp = NULLVP; - kauth_cred_t ucred; - struct timeval tv; + kauth_cred_t tcred = NOCRED; + kauth_cred_t ucred; + struct timeval tv; + boolean_t ttl_active = FALSE; + + ucred = vfs_context_ucred(ctx); - ucred = vfs_context_ucred(context); - *lookup_vpp = start_vp; + if (!IS_VALID_CRED(ucred) || action == 0) + return; + + if ( (vp->v_mount->mnt_kern_flag & (MNTK_AUTH_OPAQUE | MNTK_AUTH_CACHE_TTL)) ) { + /* + * a TTL is enabled on the rights cache... handle it here + * a TTL of 0 indicates that no rights should be cached + */ + if (vp->v_mount->mnt_authcache_ttl == 0) + return; - name_cache_lock(); + if ( !(vp->v_mount->mnt_kern_flag & MNTK_AUTH_CACHE_TTL) ) { + /* + * only cache SEARCH action for filesystems marked + * MNTK_AUTH_OPAQUE on VDIRs... + * the lookup_path code will time these out + */ + if ( (action & ~KAUTH_VNODE_SEARCH) || vp->v_type != VDIR ) + return; + } + ttl_active = TRUE; - if ( dp->v_mount && (dp->v_mount->mnt_kern_flag & MNTK_AUTH_OPAQUE) ) { - auth_opaque = 1; microuptime(&tv); } - for (;;) { - *dp_authorized = 0; + NAME_CACHE_LOCK(); - if (auth_opaque && ((tv.tv_sec - dp->v_cred_timestamp) > VCRED_EXPIRED)) - break; - if (dp->v_cred != ucred) - break; - /* - * indicate that we're allowed to traverse this directory... - * even if we bail for some reason, this information is valid and is used - * to avoid doing a vnode_authorize + if (vp->v_cred != ucred) { + kauth_cred_ref(ucred); + /* + * Use a temp variable to avoid kauth_cred_unref() while NAME_CACHE_LOCK is held */ - *dp_authorized = 1; + tcred = vp->v_cred; + vp->v_cred = ucred; + vp->v_authorized_actions = 0; + } + if (ttl_active == TRUE && vp->v_authorized_actions == 0) { + /* + * only reset the timestamnp on the + * first authorization cached after the previous + * timer has expired or we're switching creds... + * 'vnode_cache_is_authorized' will clear the + * authorized actions if the TTL is active and + * it has expired + */ + vp->v_cred_timestamp = tv.tv_sec; + } + vp->v_authorized_actions |= action; + + NAME_CACHE_UNLOCK(); + + if (IS_VALID_CRED(tcred)) + kauth_cred_unref(&tcred); +} - if ((dp->v_flag & VROOT) != 0 || /* Hit "/" */ - (dp == fdp->fd_cdir) || /* Hit process's working directory */ - (dp == fdp->fd_rdir)) { /* Hit process chroot()-ed root */ - done = 1; - break; - } - if ( (vp = dp->v_parent) == NULLVP) - break; +boolean_t vnode_cache_is_stale(vnode_t vp) +{ + struct timeval tv; + boolean_t retval; - dp = vp; - *lookup_vpp = dp; - } /* for (;;) */ + microuptime(&tv); - vid = dp->v_id; - - name_cache_unlock(); - - if (done == 0 && dp != start_vp) { - if (vnode_getwithvid(dp, vid) != 0) { - *lookup_vpp = start_vp; - } - } + if ((tv.tv_sec - vp->v_cred_timestamp) > vp->v_mount->mnt_authcache_ttl) + retval = TRUE; + else + retval = FALSE; - return((done == 1) ? 0 : -1); + return retval; } + + +/* + * Returns: 0 Success + * ERECYCLE vnode was recycled from underneath us. Force lookup to be re-driven from namei. + * This errno value should not be seen by anyone outside of the kernel. + */ int -cache_lookup_path(struct nameidata *ndp, struct componentname *cnp, vnode_t dp, vfs_context_t context, int *trailing_slash, int *dp_authorized) +cache_lookup_path(struct nameidata *ndp, struct componentname *cnp, vnode_t dp, + vfs_context_t ctx, int *dp_authorized, vnode_t last_dp) { char *cp; /* pointer into pathname argument */ - int vid, vvid; - int auth_opaque = 0; + int vid; + int vvid = 0; /* protected by vp != NULLVP */ vnode_t vp = NULLVP; vnode_t tdp = NULLVP; kauth_cred_t ucred; - struct timeval tv; + boolean_t ttl_enabled = FALSE; + struct timeval tv; + mount_t mp; unsigned int hash; + int error = 0; + boolean_t dotdotchecked = FALSE; - ucred = vfs_context_ucred(context); - *trailing_slash = 0; +#if CONFIG_TRIGGERS + vnode_t trigger_vp; +#endif /* CONFIG_TRIGGERS */ - name_cache_lock(); + ucred = vfs_context_ucred(ctx); + ndp->ni_flag &= ~(NAMEI_TRAILINGSLASH); + NAME_CACHE_LOCK_SHARED(); - if ( dp->v_mount && (dp->v_mount->mnt_kern_flag & MNTK_AUTH_OPAQUE) ) { - auth_opaque = 1; + if ( dp->v_mount && (dp->v_mount->mnt_kern_flag & (MNTK_AUTH_OPAQUE | MNTK_AUTH_CACHE_TTL)) ) { + ttl_enabled = TRUE; microuptime(&tv); } for (;;) { - /* + /* * Search a directory. * * The cn_hash value is for use by cache_lookup @@ -596,7 +1294,7 @@ cache_lookup_path(struct nameidata *ndp, struct componentname *cnp, vnode_t dp, cp = cnp->cn_nameptr; while (*cp && (*cp != '/')) { - hash ^= crc32tab[((hash >> 24) ^ (unsigned char)*cp++)]; + hash = crc32tab[((hash >> 24) ^ (unsigned char)*cp++)] ^ hash << 8; } /* * the crc generator can legitimately generate @@ -623,7 +1321,7 @@ cache_lookup_path(struct nameidata *ndp, struct componentname *cnp, vnode_t dp, ndp->ni_pathlen--; if (*cp == '\0') { - *trailing_slash = 1; + ndp->ni_flag |= NAMEI_TRAILINGSLASH; *ndp->ni_next = '\0'; } } @@ -638,12 +1336,62 @@ cache_lookup_path(struct nameidata *ndp, struct componentname *cnp, vnode_t dp, cnp->cn_flags |= ISDOTDOT; *dp_authorized = 0; +#if NAMEDRSRCFORK + /* + * Process a request for a file's resource fork. + * + * Consume the _PATH_RSRCFORKSPEC suffix and tag the path. + */ + if ((ndp->ni_pathlen == sizeof(_PATH_RSRCFORKSPEC)) && + (cp[1] == '.' && cp[2] == '.') && + bcmp(cp, _PATH_RSRCFORKSPEC, sizeof(_PATH_RSRCFORKSPEC)) == 0) { + /* Skip volfs file systems that don't support native streams. */ + if ((dp->v_mount != NULL) && + (dp->v_mount->mnt_flag & MNT_DOVOLFS) && + (dp->v_mount->mnt_kern_flag & MNTK_NAMED_STREAMS) == 0) { + goto skiprsrcfork; + } + cnp->cn_flags |= CN_WANTSRSRCFORK; + cnp->cn_flags |= ISLASTCN; + ndp->ni_next[0] = '\0'; + ndp->ni_pathlen = 1; + } +skiprsrcfork: +#endif + +#if CONFIG_MACF - if (auth_opaque && ((tv.tv_sec - dp->v_cred_timestamp) > VCRED_EXPIRED)) + /* + * Name cache provides authorization caching (see below) + * that will short circuit MAC checks in lookup(). + * We must perform MAC check here. On denial + * dp_authorized will remain 0 and second check will + * be perfomed in lookup(). + */ + if (!(cnp->cn_flags & DONOTAUTH)) { + error = mac_vnode_check_lookup(ctx, dp, cnp); + if (error) { + NAME_CACHE_UNLOCK(); + goto errorout; + } + } +#endif /* MAC */ + if (ttl_enabled && ((tv.tv_sec - dp->v_cred_timestamp) > dp->v_mount->mnt_authcache_ttl)) break; - if (dp->v_cred != ucred) + /* + * NAME_CACHE_LOCK holds these fields stable + * + * We can't cache KAUTH_VNODE_SEARCHBYANYONE for root correctly + * so we make an ugly check for root here. root is always + * allowed and breaking out of here only to find out that is + * authorized by virtue of being root is very very expensive. + */ + if ((dp->v_cred != ucred || !(dp->v_authorized_actions & KAUTH_VNODE_SEARCH)) && + !(dp->v_authorized_actions & KAUTH_VNODE_SEARCHBYANYONE) && + !vfs_context_issuser(ctx)) break; + /* * indicate that we're allowed to traverse this directory... * even if we fail the cache lookup or decide to bail for @@ -653,36 +1401,105 @@ cache_lookup_path(struct nameidata *ndp, struct componentname *cnp, vnode_t dp, *dp_authorized = 1; if ( (cnp->cn_flags & (ISLASTCN | ISDOTDOT)) ) { - if (cnp->cn_nameiop != LOOKUP) - break; - if (cnp->cn_flags & (LOCKPARENT | NOCACHE)) - break; + if (cnp->cn_nameiop != LOOKUP) + break; + if (cnp->cn_flags & LOCKPARENT) + break; + if (cnp->cn_flags & NOCACHE) + break; if (cnp->cn_flags & ISDOTDOT) { + /* + * Force directory hardlinks to go to + * file system for ".." requests. + */ + if (dp && (dp->v_flag & VISHARDLINK)) { + break; + } /* * Quit here only if we can't use * the parent directory pointer or * don't have one. Otherwise, we'll * use it below. */ - if ((dp->v_flag & VROOT) || + if ((dp->v_flag & VROOT) || + dp == ndp->ni_rootdir || dp->v_parent == NULLVP) break; } } + if ((cnp->cn_flags & CN_SKIPNAMECACHE)) { + /* + * Force lookup to go to the filesystem with + * all cnp fields set up. + */ + break; + } + /* * "." and ".." aren't supposed to be cached, so check * for them before checking the cache. */ if (cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.') vp = dp; - else if (cnp->cn_flags & ISDOTDOT) - vp = dp->v_parent; - else { + else if ( (cnp->cn_flags & ISDOTDOT) ) { + /* + * If this is a chrooted process, we need to check if + * the process is trying to break out of its chrooted + * jail. We do that by trying to determine if dp is + * a subdirectory of ndp->ni_rootdir. If we aren't + * able to determine that by the v_parent pointers, we + * will leave the fast path. + * + * Since this function may see dotdot components + * many times and it has the name cache lock held for + * the entire duration, we optimise this by doing this + * check only once per cache_lookup_path call. + * If dotdotchecked is set, it means we've done this + * check once already and don't need to do it again. + */ + if (!dotdotchecked && (ndp->ni_rootdir != rootvnode)) { + vnode_t tvp = dp; + boolean_t defer = FALSE; + boolean_t is_subdir = FALSE; + + defer = cache_check_vnode_issubdir(tvp, + ndp->ni_rootdir, &is_subdir, &tvp); + + if (defer) { + /* defer to Filesystem */ + break; + } else if (!is_subdir) { + /* + * This process is trying to break out + * of its chrooted jail, so all its + * dotdot accesses will be translated to + * its root directory. + */ + vp = ndp->ni_rootdir; + } else { + /* + * All good, let this dotdot access + * proceed normally + */ + vp = dp->v_parent; + } + dotdotchecked = TRUE; + } else { + vp = dp->v_parent; + } + } else { if ( (vp = cache_lookup_locked(dp, cnp)) == NULLVP) break; - } + if ( (vp->v_flag & VISHARDLINK) ) { + /* + * The file system wants a VNOP_LOOKUP on this vnode + */ + vp = NULL; + break; + } + } if ( (cnp->cn_flags & ISLASTCN) ) break; @@ -691,8 +1508,25 @@ cache_lookup_path(struct nameidata *ndp, struct componentname *cnp, vnode_t dp, vp = NULL; break; } - if (vp->v_mountedhere && ((cnp->cn_flags & NOCROSSMOUNT) == 0)) - break; + + if ( (mp = vp->v_mountedhere) && ((cnp->cn_flags & NOCROSSMOUNT) == 0)) { + vnode_t tmp_vp = mp->mnt_realrootvp; + if (tmp_vp == NULLVP || mp->mnt_generation != mount_generation || + mp->mnt_realrootvp_vid != tmp_vp->v_id) + break; + vp = tmp_vp; + } + +#if CONFIG_TRIGGERS + /* + * After traversing all mountpoints stacked here, if we have a + * trigger in hand, resolve it. Note that we don't need to + * leave the fast path if the mount has already happened. + */ + if (vp->v_resolve) + break; +#endif /* CONFIG_TRIGGERS */ + dp = vp; vp = NULLVP; @@ -708,8 +1542,7 @@ cache_lookup_path(struct nameidata *ndp, struct componentname *cnp, vnode_t dp, vvid = vp->v_id; vid = dp->v_id; - name_cache_unlock(); - + NAME_CACHE_UNLOCK(); if ((vp != NULLVP) && (vp->v_type != VLNK) && ((cnp->cn_flags & (ISLASTCN | LOCKPARENT | WANTPARENT | SAVESTART)) == ISLASTCN)) { @@ -726,34 +1559,54 @@ cache_lookup_path(struct nameidata *ndp, struct componentname *cnp, vnode_t dp, dp = NULLVP; } else { need_dp: - /* + /* * return the last directory we looked at - * with an io reference held + * with an io reference held. If it was the one passed + * in as a result of the last iteration of VNOP_LOOKUP, + * it should already hold an io ref. No need to increase ref. */ - if (dp == ndp->ni_usedvp) { - /* - * if this vnode matches the one passed in via USEDVP - * than this context already holds an io_count... just - * use vnode_get to get an extra ref for lookup to play - * with... can't use the getwithvid variant here because - * it will block behind a vnode_drain which would result - * in a deadlock (since we already own an io_count that the - * vnode_drain is waiting on)... vnode_get grabs the io_count - * immediately w/o waiting... it always succeeds - */ - vnode_get(dp); - } else if ( (vnode_getwithvid(dp, vid)) ) { - /* - * failure indicates the vnode - * changed identity or is being - * TERMINATED... in either case - * punt this lookup - */ - return (ENOENT); + if (last_dp != dp){ + + if (dp == ndp->ni_usedvp) { + /* + * if this vnode matches the one passed in via USEDVP + * than this context already holds an io_count... just + * use vnode_get to get an extra ref for lookup to play + * with... can't use the getwithvid variant here because + * it will block behind a vnode_drain which would result + * in a deadlock (since we already own an io_count that the + * vnode_drain is waiting on)... vnode_get grabs the io_count + * immediately w/o waiting... it always succeeds + */ + vnode_get(dp); + } else if ((error = vnode_getwithvid_drainok(dp, vid))) { + /* + * failure indicates the vnode + * changed identity or is being + * TERMINATED... in either case + * punt this lookup. + * + * don't necessarily return ENOENT, though, because + * we really want to go back to disk and make sure it's + * there or not if someone else is changing this + * vnode. That being said, the one case where we do want + * to return ENOENT is when the vnode's mount point is + * in the process of unmounting and we might cause a deadlock + * in our attempt to take an iocount. An ENODEV error return + * is from vnode_get* is an indication this but we change that + * ENOENT for upper layers. + */ + if (error == ENODEV) { + error = ENOENT; + } else { + error = ERECYCLE; + } + goto errorout; + } } } if (vp != NULLVP) { - if ( (vnode_getwithvid(vp, vvid)) ) { + if ( (vnode_getwithvid_drainok(vp, vvid)) ) { vp = NULLVP; /* @@ -770,64 +1623,92 @@ need_dp: } } } + ndp->ni_dvp = dp; ndp->ni_vp = vp; - return (0); +#if CONFIG_TRIGGERS + trigger_vp = vp ? vp : dp; + if ((error == 0) && (trigger_vp != NULLVP) && vnode_isdir(trigger_vp)) { + error = vnode_trigger_resolve(trigger_vp, ndp, ctx); + if (error) { + if (vp) + vnode_put(vp); + if (dp) + vnode_put(dp); + goto errorout; + } + } +#endif /* CONFIG_TRIGGERS */ + +errorout: + /* + * If we came into cache_lookup_path after an iteration of the lookup loop that + * resulted in a call to VNOP_LOOKUP, then VNOP_LOOKUP returned a vnode with a io ref + * on it. It is now the job of cache_lookup_path to drop the ref on this vnode + * when it is no longer needed. If we get to this point, and last_dp is not NULL + * and it is ALSO not the dvp we want to return to caller of this function, it MUST be + * the case that we got to a subsequent path component and this previous vnode is + * no longer needed. We can then drop the io ref on it. + */ + if ((last_dp != NULLVP) && (last_dp != ndp->ni_dvp)){ + vnode_put(last_dp); + } + + //initialized to 0, should be the same if no error cases occurred. + return error; } static vnode_t cache_lookup_locked(vnode_t dvp, struct componentname *cnp) { - register struct namecache *ncp; - register struct nchashhead *ncpp; - register long namelen = cnp->cn_namelen; - char *nameptr = cnp->cn_nameptr; - unsigned int hashval = (cnp->cn_hash & NCHASHMASK); - vnode_t vp; + struct namecache *ncp; + struct nchashhead *ncpp; + long namelen = cnp->cn_namelen; + unsigned int hashval = cnp->cn_hash; + if (nc_disabled) { + return NULL; + } + ncpp = NCHHASH(dvp, cnp->cn_hash); LIST_FOREACH(ncp, ncpp, nc_hash) { if ((ncp->nc_dvp == dvp) && (ncp->nc_hashval == hashval)) { - if (memcmp(ncp->nc_name, nameptr, namelen) == 0 && ncp->nc_name[namelen] == 0) + if (memcmp(ncp->nc_name, cnp->cn_nameptr, namelen) == 0 && ncp->nc_name[namelen] == 0) break; } } - if (ncp == 0) + if (ncp == 0) { /* * We failed to find an entry */ + NCHSTAT(ncs_miss); return (NULL); - - vp = ncp->nc_vp; - if (vp && (vp->v_flag & VISHARDLINK)) { - /* - * The file system wants a VNOP_LOOKUP on this vnode - */ - vp = NULL; } - - return (vp); + NCHSTAT(ncs_goodhits); + + return (ncp->nc_vp); } +unsigned int hash_string(const char *cp, int len); // // Have to take a len argument because we may only need to // hash part of a componentname. // -static unsigned int +unsigned int hash_string(const char *cp, int len) { unsigned hash = 0; if (len) { while (len--) { - hash ^= crc32tab[((hash >> 24) ^ (unsigned char)*cp++)]; + hash = crc32tab[((hash >> 24) ^ (unsigned char)*cp++)] ^ hash << 8; } } else { while (*cp != '\0') { - hash ^= crc32tab[((hash >> 24) ^ (unsigned char)*cp++)]; + hash = crc32tab[((hash >> 24) ^ (unsigned char)*cp++)] ^ hash << 8; } } /* @@ -857,55 +1738,69 @@ hash_string(const char *cp, int len) */ int -cache_lookup(dvp, vpp, cnp) - struct vnode *dvp; - struct vnode **vpp; - struct componentname *cnp; +cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp) { - register struct namecache *ncp; - register struct nchashhead *ncpp; - register long namelen = cnp->cn_namelen; - char *nameptr = cnp->cn_nameptr; - unsigned int hashval = (cnp->cn_hash & NCHASHMASK); + struct namecache *ncp; + struct nchashhead *ncpp; + long namelen = cnp->cn_namelen; + unsigned int hashval; + boolean_t have_exclusive = FALSE; uint32_t vid; vnode_t vp; - name_cache_lock(); + if (cnp->cn_hash == 0) + cnp->cn_hash = hash_string(cnp->cn_nameptr, cnp->cn_namelen); + hashval = cnp->cn_hash; + + if (nc_disabled) { + return 0; + } + + NAME_CACHE_LOCK_SHARED(); +relook: ncpp = NCHHASH(dvp, cnp->cn_hash); LIST_FOREACH(ncp, ncpp, nc_hash) { if ((ncp->nc_dvp == dvp) && (ncp->nc_hashval == hashval)) { - if (memcmp(ncp->nc_name, nameptr, namelen) == 0 && ncp->nc_name[namelen] == 0) + if (memcmp(ncp->nc_name, cnp->cn_nameptr, namelen) == 0 && ncp->nc_name[namelen] == 0) break; } } /* We failed to find an entry */ if (ncp == 0) { - nchstats.ncs_miss++; - name_cache_unlock(); + NCHSTAT(ncs_miss); + NAME_CACHE_UNLOCK(); return (0); } /* We don't want to have an entry, so dump it */ if ((cnp->cn_flags & MAKEENTRY) == 0) { - nchstats.ncs_badhits++; - cache_delete(ncp, 1); - name_cache_unlock(); - return (0); + if (have_exclusive == TRUE) { + NCHSTAT(ncs_badhits); + cache_delete(ncp, 1); + NAME_CACHE_UNLOCK(); + return (0); + } + NAME_CACHE_UNLOCK(); + NAME_CACHE_LOCK(); + have_exclusive = TRUE; + goto relook; } vp = ncp->nc_vp; /* We found a "positive" match, return the vnode */ if (vp) { - nchstats.ncs_goodhits++; + NCHSTAT(ncs_goodhits); vid = vp->v_id; - name_cache_unlock(); + NAME_CACHE_UNLOCK(); if (vnode_getwithvid(vp, vid)) { - name_cache_lock(); - nchstats.ncs_badvid++; - name_cache_unlock(); +#if COLLECT_STATS + NAME_CACHE_LOCK(); + NCHSTAT(ncs_badvid); + NAME_CACHE_UNLOCK(); +#endif return (0); } *vpp = vp; @@ -914,49 +1809,121 @@ cache_lookup(dvp, vpp, cnp) /* We found a negative match, and want to create it, so purge */ if (cnp->cn_nameiop == CREATE || cnp->cn_nameiop == RENAME) { - nchstats.ncs_badhits++; - cache_delete(ncp, 1); - name_cache_unlock(); - return (0); + if (have_exclusive == TRUE) { + NCHSTAT(ncs_badhits); + cache_delete(ncp, 1); + NAME_CACHE_UNLOCK(); + return (0); + } + NAME_CACHE_UNLOCK(); + NAME_CACHE_LOCK(); + have_exclusive = TRUE; + goto relook; } /* * We found a "negative" match, ENOENT notifies client of this match. - * The nc_whiteout field records whether this is a whiteout. */ - nchstats.ncs_neghits++; + NCHSTAT(ncs_neghits); - if (ncp->nc_whiteout) - cnp->cn_flags |= ISWHITEOUT; - name_cache_unlock(); + NAME_CACHE_UNLOCK(); return (ENOENT); } +const char * +cache_enter_create(vnode_t dvp, vnode_t vp, struct componentname *cnp) +{ + const char *strname; + + if (cnp->cn_hash == 0) + cnp->cn_hash = hash_string(cnp->cn_nameptr, cnp->cn_namelen); + + /* + * grab 2 references on the string entered + * one for the cache_enter_locked to consume + * and the second to be consumed by v_name (vnode_create call point) + */ + strname = add_name_internal(cnp->cn_nameptr, cnp->cn_namelen, cnp->cn_hash, TRUE, 0); + + NAME_CACHE_LOCK(); + + cache_enter_locked(dvp, vp, cnp, strname); + + NAME_CACHE_UNLOCK(); + + return (strname); +} + + +/* + * Add an entry to the cache... + * but first check to see if the directory + * that this entry is to be associated with has + * had any cache_purges applied since we took + * our identity snapshot... this check needs to + * be done behind the name cache lock + */ +void +cache_enter_with_gen(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, int gen) +{ + + if (cnp->cn_hash == 0) + cnp->cn_hash = hash_string(cnp->cn_nameptr, cnp->cn_namelen); + + NAME_CACHE_LOCK(); + + if (dvp->v_nc_generation == gen) + (void)cache_enter_locked(dvp, vp, cnp, NULL); + + NAME_CACHE_UNLOCK(); +} + + /* * Add an entry to the cache. */ void -cache_enter(dvp, vp, cnp) - struct vnode *dvp; - struct vnode *vp; - struct componentname *cnp; +cache_enter(struct vnode *dvp, struct vnode *vp, struct componentname *cnp) { - register struct namecache *ncp, *negp; - register struct nchashhead *ncpp; + const char *strname; if (cnp->cn_hash == 0) cnp->cn_hash = hash_string(cnp->cn_nameptr, cnp->cn_namelen); - name_cache_lock(); + /* + * grab 1 reference on the string entered + * for the cache_enter_locked to consume + */ + strname = add_name_internal(cnp->cn_nameptr, cnp->cn_namelen, cnp->cn_hash, FALSE, 0); + + NAME_CACHE_LOCK(); + + cache_enter_locked(dvp, vp, cnp, strname); + + NAME_CACHE_UNLOCK(); +} + - /* if the entry is for -ve caching vp is null */ +static void +cache_enter_locked(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, const char *strname) +{ + struct namecache *ncp, *negp; + struct nchashhead *ncpp; + + if (nc_disabled) + return; + + /* + * if the entry is for -ve caching vp is null + */ if ((vp != NULLVP) && (LIST_FIRST(&vp->v_nclinks))) { /* * someone beat us to the punch.. * this vnode is already in the cache */ - name_cache_unlock(); - return; + if (strname != NULL) + vfs_removename(strname); + return; } /* * We allocate a new entry if we are less than the maximum @@ -969,7 +1936,7 @@ cache_enter(dvp, vp, cnp) /* * Allocate one more entry */ - ncp = (struct namecache *)_MALLOC_ZONE((u_long)sizeof *ncp, M_CACHE, M_WAITOK); + ncp = (struct namecache *)_MALLOC_ZONE(sizeof(*ncp), M_CACHE, M_WAITOK); numcache++; } else { /* @@ -983,11 +1950,11 @@ cache_enter(dvp, vp, cnp) * still in use... we need to * delete it before re-using it */ - nchstats.ncs_stolen++; + NCHSTAT(ncs_stolen); cache_delete(ncp, 0); } } - nchstats.ncs_enters++; + NCHSTAT(ncs_enters); /* * Fill in cache info, if vp is NULL this is a "negative" cache entry. @@ -995,8 +1962,29 @@ cache_enter(dvp, vp, cnp) ncp->nc_vp = vp; ncp->nc_dvp = dvp; ncp->nc_hashval = cnp->cn_hash; - ncp->nc_whiteout = FALSE; - ncp->nc_name = add_name_locked(cnp->cn_nameptr, cnp->cn_namelen, cnp->cn_hash, 0); + + if (strname == NULL) + ncp->nc_name = add_name_internal(cnp->cn_nameptr, cnp->cn_namelen, cnp->cn_hash, FALSE, 0); + else + ncp->nc_name = strname; + + // + // If the bytes of the name associated with the vnode differ, + // use the name associated with the vnode since the file system + // may have set that explicitly in the case of a lookup on a + // case-insensitive file system where the case of the looked up + // name differs from what is on disk. For more details, see: + // FSEvents doesn't always decompose diacritical unicode chars in the paths of the changed directories + // + const char *vn_name = vp ? vp->v_name : NULL; + unsigned int len = vn_name ? strlen(vn_name) : 0; + if (vn_name && ncp && ncp->nc_name && strncmp(ncp->nc_name, vn_name, len) != 0) { + unsigned int hash = hash_string(vn_name, len); + + vfs_removename(ncp->nc_name); + ncp->nc_name = add_name_internal(vn_name, len, hash, FALSE, 0); + ncp->nc_hashval = hash; + } /* * make us the newest entry in the cache @@ -1007,7 +1995,7 @@ cache_enter(dvp, vp, cnp) ncpp = NCHHASH(dvp, cnp->cn_hash); #if DIAGNOSTIC { - register struct namecache *p; + struct namecache *p; for (p = ncpp->lh_first; p != 0; p = p->nc_hash.le_next) if (p == ncp) @@ -1028,24 +2016,19 @@ cache_enter(dvp, vp, cnp) } else { /* * this is a negative cache entry (vp == NULL) - * stick it on the negative cache list - * and record the whiteout state + * stick it on the negative cache list. */ TAILQ_INSERT_TAIL(&neghead, ncp, nc_un.nc_negentry); - if (cnp->cn_flags & ISWHITEOUT) - ncp->nc_whiteout = TRUE; - nchstats.ncs_negtotal++; + ncs_negtotal++; - if (nchstats.ncs_negtotal > desiredNegNodes) { + if (ncs_negtotal > desiredNegNodes) { /* * if we've reached our desired limit * of negative cache entries, delete * the oldest */ negp = TAILQ_FIRST(&neghead); - TAILQ_REMOVE(&neghead, negp, nc_un.nc_negentry); - cache_delete(negp, 1); } } @@ -1053,9 +2036,10 @@ cache_enter(dvp, vp, cnp) * add us to the list of name cache entries that * are children of dvp */ - LIST_INSERT_HEAD(&dvp->v_ncchildren, ncp, nc_child); - - name_cache_unlock(); + if (vp) + TAILQ_INSERT_TAIL(&dvp->v_ncchildren, ncp, nc_child); + else + TAILQ_INSERT_HEAD(&dvp->v_ncchildren, ncp, nc_child); } @@ -1095,6 +2079,8 @@ static void init_crc32(void) void nchinit(void) { + int i; + desiredNegNodes = (desiredvnodes / 10); desiredNodes = desiredvnodes + desiredNegNodes; @@ -1103,39 +2089,55 @@ nchinit(void) init_crc32(); - nchashtbl = hashinit(MAX(4096, (2 *desiredNodes)), M_CACHE, &nchash); + nchashtbl = hashinit(MAX(CONFIG_NC_HASH, (2 *desiredNodes)), M_CACHE, &nchash); nchashmask = nchash; nchash++; init_string_table(); - /* Allocate mount list lock group attribute and group */ + /* Allocate name cache lock group attribute and group */ namecache_lck_grp_attr= lck_grp_attr_alloc_init(); - lck_grp_attr_setstat(namecache_lck_grp_attr); namecache_lck_grp = lck_grp_alloc_init("Name Cache", namecache_lck_grp_attr); - /* Allocate mount list lock attribute */ + /* Allocate name cache lock attribute */ namecache_lck_attr = lck_attr_alloc_init(); - //lck_attr_setdebug(namecache_lck_attr); - /* Allocate mount list lock */ - namecache_mtx_lock = lck_mtx_alloc_init(namecache_lck_grp, namecache_lck_attr); + /* Allocate name cache lock */ + namecache_rw_lock = lck_rw_alloc_init(namecache_lck_grp, namecache_lck_attr); + + + /* Allocate string cache lock group attribute and group */ + strcache_lck_grp_attr= lck_grp_attr_alloc_init(); + strcache_lck_grp = lck_grp_alloc_init("String Cache", strcache_lck_grp_attr); + + /* Allocate string cache lock attribute */ + strcache_lck_attr = lck_attr_alloc_init(); + + /* Allocate string cache lock */ + strtable_rw_lock = lck_rw_alloc_init(strcache_lck_grp, strcache_lck_attr); + + for (i = 0; i < NUM_STRCACHE_LOCKS; i++) + lck_mtx_init(&strcache_mtx_locks[i], strcache_lck_grp, strcache_lck_attr); +} +void +name_cache_lock_shared(void) +{ + lck_rw_lock_shared(namecache_rw_lock); } void name_cache_lock(void) { - lck_mtx_lock(namecache_mtx_lock); + lck_rw_lock_exclusive(namecache_rw_lock); } void name_cache_unlock(void) { - lck_mtx_unlock(namecache_mtx_lock); - + lck_rw_done(namecache_rw_lock); } @@ -1154,7 +2156,7 @@ resize_namecache(u_int newsize) dNodes = newsize + dNegNodes; // we don't support shrinking yet - if (dNodes < desiredNodes) { + if (dNodes <= desiredNodes) { return 0; } new_table = hashinit(2 * dNodes, M_CACHE, &nchashmask); @@ -1164,7 +2166,7 @@ resize_namecache(u_int newsize) return ENOMEM; } - name_cache_lock(); + NAME_CACHE_LOCK(); // do the switch! old_table = nchashtbl; nchashtbl = new_table; @@ -1192,7 +2194,7 @@ resize_namecache(u_int newsize) desiredNodes = dNodes; desiredNegNodes = dNegNodes; - name_cache_unlock(); + NAME_CACHE_UNLOCK(); FREE(old_table, M_CACHE); return 0; @@ -1201,15 +2203,15 @@ resize_namecache(u_int newsize) static void cache_delete(struct namecache *ncp, int age_entry) { - nchstats.ncs_deletes++; + NCHSTAT(ncs_deletes); if (ncp->nc_vp) { LIST_REMOVE(ncp, nc_un.nc_link); } else { TAILQ_REMOVE(&neghead, ncp, nc_un.nc_negentry); - nchstats.ncs_negtotal--; + ncs_negtotal--; } - LIST_REMOVE(ncp, nc_child); + TAILQ_REMOVE(&(ncp->nc_dvp->v_ncchildren), ncp, nc_child); LIST_REMOVE(ncp, nc_hash); /* @@ -1228,7 +2230,7 @@ cache_delete(struct namecache *ncp, int age_entry) TAILQ_REMOVE(&nchead, ncp, nc_entry); TAILQ_INSERT_HEAD(&nchead, ncp, nc_entry); } - remove_name_locked(ncp->nc_name); + vfs_removename(ncp->nc_name); ncp->nc_name = NULL; } @@ -1241,19 +2243,36 @@ void cache_purge(vnode_t vp) { struct namecache *ncp; + kauth_cred_t tcred = NULL; - if ((LIST_FIRST(&vp->v_nclinks) == NULL) && (LIST_FIRST(&vp->v_ncchildren) == NULL)) + if ((LIST_FIRST(&vp->v_nclinks) == NULL) && + (TAILQ_FIRST(&vp->v_ncchildren) == NULL) && + (vp->v_cred == NOCRED) && + (vp->v_parent == NULLVP)) return; - name_cache_lock(); + NAME_CACHE_LOCK(); + + if (vp->v_parent) + vp->v_parent->v_nc_generation++; while ( (ncp = LIST_FIRST(&vp->v_nclinks)) ) cache_delete(ncp, 1); - while ( (ncp = LIST_FIRST(&vp->v_ncchildren)) ) + while ( (ncp = TAILQ_FIRST(&vp->v_ncchildren)) ) cache_delete(ncp, 1); - name_cache_unlock(); + /* + * Use a temp variable to avoid kauth_cred_unref() while NAME_CACHE_LOCK is held + */ + tcred = vp->v_cred; + vp->v_cred = NOCRED; + vp->v_authorized_actions = 0; + + NAME_CACHE_UNLOCK(); + + if (IS_VALID_CRED(tcred)) + kauth_cred_unref(&tcred); } /* @@ -1266,15 +2285,18 @@ cache_purge(vnode_t vp) void cache_purge_negatives(vnode_t vp) { - struct namecache *ncp; + struct namecache *ncp, *next_ncp; + + NAME_CACHE_LOCK(); - name_cache_lock(); + TAILQ_FOREACH_SAFE(ncp, &vp->v_ncchildren, nc_child, next_ncp) { + if (ncp->nc_vp) + break; - LIST_FOREACH(ncp, &vp->v_ncchildren, nc_child) - if (ncp->nc_vp == NULL) - cache_delete(ncp , 1); + cache_delete(ncp, 1); + } - name_cache_unlock(); + NAME_CACHE_UNLOCK(); } /* @@ -1284,13 +2306,12 @@ cache_purge_negatives(vnode_t vp) * entries at the same time. */ void -cache_purgevfs(mp) - struct mount *mp; +cache_purgevfs(struct mount *mp) { struct nchashhead *ncpp; struct namecache *ncp; - name_cache_lock(); + NAME_CACHE_LOCK(); /* Scan hash tables for applicable entries */ for (ncpp = &nchashtbl[nchash - 1]; ncpp >= nchashtbl; ncpp--) { restart: @@ -1301,7 +2322,7 @@ restart: } } } - name_cache_unlock(); + NAME_CACHE_UNLOCK(); } @@ -1311,195 +2332,229 @@ restart: // static LIST_HEAD(stringhead, string_t) *string_ref_table; static u_long string_table_mask; -static uint32_t max_chain_len=0; -static struct stringhead *long_chain_head=NULL; static uint32_t filled_buckets=0; -static uint32_t num_dups=0; -static uint32_t nstrings=0; + typedef struct string_t { LIST_ENTRY(string_t) hash_chain; - unsigned char *str; + const char *str; uint32_t refcount; } string_t; - -static int +static void resize_string_ref_table(void) { - struct stringhead *new_table; - struct stringhead *old_table; - struct stringhead *old_head, *head; - string_t *entry, *next; - uint32_t i, hashval; - u_long new_mask, old_mask; - - new_table = hashinit((string_table_mask + 1) * 2, M_CACHE, &new_mask); - if (new_table == NULL) { - return ENOMEM; - } + struct stringhead *new_table; + struct stringhead *old_table; + struct stringhead *old_head, *head; + string_t *entry, *next; + uint32_t i, hashval; + u_long new_mask, old_mask; - // do the switch! - old_table = string_ref_table; - string_ref_table = new_table; - old_mask = string_table_mask; - string_table_mask = new_mask; + /* + * need to hold the table lock exclusively + * in order to grow the table... need to recheck + * the need to resize again after we've taken + * the lock exclusively in case some other thread + * beat us to the punch + */ + lck_rw_lock_exclusive(strtable_rw_lock); - printf("resize: max chain len %d, new table size %d\n", - max_chain_len, new_mask + 1); - max_chain_len = 0; - long_chain_head = NULL; - filled_buckets = 0; + if (4 * filled_buckets < ((string_table_mask + 1) * 3)) { + lck_rw_done(strtable_rw_lock); + return; + } + new_table = hashinit((string_table_mask + 1) * 2, M_CACHE, &new_mask); - // walk the old table and insert all the entries into - // the new table - // - for(i=0; i <= old_mask; i++) { - old_head = &old_table[i]; - for (entry=old_head->lh_first; entry != NULL; entry=next) { - hashval = hash_string(entry->str, 0); - head = &string_ref_table[hashval & string_table_mask]; - if (head->lh_first == NULL) { - filled_buckets++; - } + if (new_table == NULL) { + printf("failed to resize the hash table.\n"); + lck_rw_done(strtable_rw_lock); + return; + } - next = entry->hash_chain.le_next; - LIST_INSERT_HEAD(head, entry, hash_chain); + // do the switch! + old_table = string_ref_table; + string_ref_table = new_table; + old_mask = string_table_mask; + string_table_mask = new_mask; + filled_buckets = 0; + + // walk the old table and insert all the entries into + // the new table + // + for (i = 0; i <= old_mask; i++) { + old_head = &old_table[i]; + for (entry = old_head->lh_first; entry != NULL; entry = next) { + hashval = hash_string((const char *)entry->str, 0); + head = &string_ref_table[hashval & string_table_mask]; + if (head->lh_first == NULL) { + filled_buckets++; + } + next = entry->hash_chain.le_next; + LIST_INSERT_HEAD(head, entry, hash_chain); + } } - } - - FREE(old_table, M_CACHE); + lck_rw_done(strtable_rw_lock); - return 0; + FREE(old_table, M_CACHE); } static void init_string_table(void) { - string_ref_table = hashinit(4096, M_CACHE, &string_table_mask); + string_ref_table = hashinit(CONFIG_VFS_NAMES, M_CACHE, &string_table_mask); } -char * -vfs_addname(const char *name, size_t len, u_int hashval, u_int flags) +const char * +vfs_addname(const char *name, uint32_t len, u_int hashval, u_int flags) { - char * ptr; - - name_cache_lock(); - ptr = add_name_locked(name, len, hashval, flags); - name_cache_unlock(); - - return(ptr); + return (add_name_internal(name, len, hashval, FALSE, flags)); } -static char * -add_name_locked(const char *name, size_t len, u_int hashval, __unused u_int flags) + +static const char * +add_name_internal(const char *name, uint32_t len, u_int hashval, boolean_t need_extra_ref, __unused u_int flags) { - struct stringhead *head; - string_t *entry; - uint32_t chain_len = 0; + struct stringhead *head; + string_t *entry; + uint32_t chain_len = 0; + uint32_t hash_index; + uint32_t lock_index; + char *ptr; - // - // If the table gets more than 3/4 full, resize it - // - if (4*filled_buckets >= ((string_table_mask + 1) * 3)) { - if (resize_string_ref_table() != 0) { - printf("failed to resize the hash table.\n"); - } - } - if (hashval == 0) { - hashval = hash_string(name, 0); - } + if (len > MAXPATHLEN) + len = MAXPATHLEN; - head = &string_ref_table[hashval & string_table_mask]; - for (entry=head->lh_first; entry != NULL; chain_len++, entry=entry->hash_chain.le_next) { - if (memcmp(entry->str, name, len) == 0 && entry->str[len] == '\0') { - entry->refcount++; - num_dups++; - break; + /* + * if the length already accounts for the null-byte, then + * subtract one so later on we don't index past the end + * of the string. + */ + if (len > 0 && name[len-1] == '\0') { + len--; + } + if (hashval == 0) { + hashval = hash_string(name, len); } - } - if (entry == NULL) { - // it wasn't already there so add it. - MALLOC(entry, string_t *, sizeof(string_t) + len + 1, M_TEMP, M_WAITOK); + /* + * take this lock 'shared' to keep the hash stable + * if someone else decides to grow the pool they + * will take this lock exclusively + */ + lck_rw_lock_shared(strtable_rw_lock); - // have to get "head" again because we could have blocked - // in malloc and thus head could have changed. - // - head = &string_ref_table[hashval & string_table_mask]; - if (head->lh_first == NULL) { - filled_buckets++; + /* + * If the table gets more than 3/4 full, resize it + */ + if (4 * filled_buckets >= ((string_table_mask + 1) * 3)) { + lck_rw_done(strtable_rw_lock); + + resize_string_ref_table(); + + lck_rw_lock_shared(strtable_rw_lock); } + hash_index = hashval & string_table_mask; + lock_index = hash_index % NUM_STRCACHE_LOCKS; + + head = &string_ref_table[hash_index]; - entry->str = (char *)((char *)entry + sizeof(string_t)); - strncpy(entry->str, name, len); - entry->str[len] = '\0'; - entry->refcount = 1; - LIST_INSERT_HEAD(head, entry, hash_chain); + lck_mtx_lock_spin(&strcache_mtx_locks[lock_index]); - if (chain_len > max_chain_len) { - max_chain_len = chain_len; - long_chain_head = head; + for (entry = head->lh_first; entry != NULL; chain_len++, entry = entry->hash_chain.le_next) { + if (memcmp(entry->str, name, len) == 0 && entry->str[len] == 0) { + entry->refcount++; + break; + } } + if (entry == NULL) { + lck_mtx_convert_spin(&strcache_mtx_locks[lock_index]); + /* + * it wasn't already there so add it. + */ + MALLOC(entry, string_t *, sizeof(string_t) + len + 1, M_TEMP, M_WAITOK); - nstrings++; - } + if (head->lh_first == NULL) { + OSAddAtomic(1, &filled_buckets); + } + ptr = (char *)((char *)entry + sizeof(string_t)); + strncpy(ptr, name, len); + ptr[len] = '\0'; + entry->str = ptr; + entry->refcount = 1; + LIST_INSERT_HEAD(head, entry, hash_chain); + } + if (need_extra_ref == TRUE) + entry->refcount++; - return entry->str; + lck_mtx_unlock(&strcache_mtx_locks[lock_index]); + lck_rw_done(strtable_rw_lock); + + return (const char *)entry->str; } + int vfs_removename(const char *nameref) { - int i; + struct stringhead *head; + string_t *entry; + uint32_t hashval; + uint32_t hash_index; + uint32_t lock_index; + int retval = ENOENT; - name_cache_lock(); - i = remove_name_locked(nameref); - name_cache_unlock(); + hashval = hash_string(nameref, 0); - return(i); - -} + /* + * take this lock 'shared' to keep the hash stable + * if someone else decides to grow the pool they + * will take this lock exclusively + */ + lck_rw_lock_shared(strtable_rw_lock); + /* + * must compute the head behind the table lock + * since the size and location of the table + * can change on the fly + */ + hash_index = hashval & string_table_mask; + lock_index = hash_index % NUM_STRCACHE_LOCKS; + head = &string_ref_table[hash_index]; -static int -remove_name_locked(const char *nameref) -{ - struct stringhead *head; - string_t *entry; - uint32_t hashval; - char * ptr; - - hashval = hash_string(nameref, 0); - head = &string_ref_table[hashval & string_table_mask]; - for (entry=head->lh_first; entry != NULL; entry=entry->hash_chain.le_next) { - if (entry->str == (unsigned char *)nameref) { - entry->refcount--; - if (entry->refcount == 0) { - LIST_REMOVE(entry, hash_chain); - if (head->lh_first == NULL) { - filled_buckets--; - } - ptr = entry->str; - entry->str = NULL; - nstrings--; + lck_mtx_lock_spin(&strcache_mtx_locks[lock_index]); - FREE(entry, M_TEMP); - } else { - num_dups--; - } + for (entry = head->lh_first; entry != NULL; entry = entry->hash_chain.le_next) { + if (entry->str == nameref) { + entry->refcount--; + + if (entry->refcount == 0) { + LIST_REMOVE(entry, hash_chain); - return 0; + if (head->lh_first == NULL) { + OSAddAtomic(-1, &filled_buckets); + } + } else { + entry = NULL; + } + retval = 0; + break; + } } - } + lck_mtx_unlock(&strcache_mtx_locks[lock_index]); + lck_rw_done(strtable_rw_lock); + + if (entry != NULL) + FREE(entry, M_TEMP); - return ENOENT; + return retval; } +#ifdef DUMP_STRING_TABLE void dump_string_table(void) { @@ -1507,12 +2562,14 @@ dump_string_table(void) string_t *entry; u_long i; - name_cache_lock(); + lck_rw_lock_shared(strtable_rw_lock); + for (i = 0; i <= string_table_mask; i++) { head = &string_ref_table[i]; for (entry=head->lh_first; entry != NULL; entry=entry->hash_chain.le_next) { printf("%6d - %s\n", entry->refcount, entry->str); } } - name_cache_unlock(); + lck_rw_done(strtable_rw_lock); } +#endif /* DUMP_STRING_TABLE */