xref: /illumos-gate/usr/src/lib/libdevinfo/devinfo_devlink.c (revision bbf21555)

/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License (the "License").
 * You may not use this file except in compliance with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */
/*
 * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 * Copyright (c) 2016 by Delphix. All rights reserved.
 */

#include "libdevinfo.h"
#include "devinfo_devlink.h"
#include "device_info.h"

#undef	DEBUG
#ifndef	DEBUG
#define	NDEBUG 1
#else
#undef	NDEBUG
#endif

#include <assert.h>

static mutex_t update_mutex = DEFAULTMUTEX; /* Protects update record lock */
static mutex_t temp_file_mutex = DEFAULTMUTEX; /* for file creation tests */

static const size_t elem_sizes[DB_TYPES] = {
	sizeof (struct db_node),
	sizeof (struct db_minor),
	sizeof (struct db_link),
	sizeof (char)
};

/*
 * List of directories/files skipped while physically walking /dev
 * Paths are relative to "<root>/dev/"
 */
static const char *skip_dirs[] = {"fd"};
static const char *skip_files[] = {
	"stdout",
	"stdin",
	"stderr"
};

#define	N_SKIP_DIRS	(sizeof (skip_dirs) / sizeof (skip_dirs[0]))
#define	N_SKIP_FILES	(sizeof (skip_files) / sizeof (skip_files[0]))

#define	DI_TEST_DB	ETCDEV "di_test_db"

/*
 *
 * This file contains two sets of interfaces which operate on the reverse
 * links database. One set (which includes di_devlink_open()/_close())
 * allows link generators like devfsadm(8) and ucblinks(1B) (writers) to
 * populate the database with /devices -> /dev mappings. Another set
 * of interfaces (which includes di_devlink_init()/_fini()) allows
 * applications (readers) to lookup the database for /dev links corresponding
 * to a given minor.
 *
 * Writers operate on a cached version of the database. The cache is created
 * when di_devlink_open() is called. As links in /dev are created and removed,
 * the cache is updated to keep it in synch with /dev. When the /dev updates
 * are complete, the link generator calls di_devlink_close() which writes
 * out the cache to the database.
 *
 * Applications which need to lookup the database, call di_devlink_init().
 * di_devlink_init() checks the database file (if one exists). If the
 * database is valid, it is mapped into the address space of the
 * application. The database file consists of several segments. Each
 * segment can be mapped in independently and is mapped on demand.
 *
 *		   Database Layout
 *
 *		---------------------
 *		|	Magic #     |
 *		| ----------------- |
 *		|       Version	    |	HEADER
 *		| ----------------- |
 *		|        ...        |
 *		---------------------
 *		|		    |
 *		|		    |	NODES
 *		|	            |
 *		|		    |
 *		---------------------
 *		|		    |
 *		|		    |	MINORS
 *		|	            |
 *		|		    |
 *		---------------------
 *		|		    |
 *		|		    |   LINKS
 *		|	            |
 *		|		    |
 *		---------------------
 *		|		    |
 *		|		    |	STRINGS
 *		|	            |
 *		|		    |
 *		---------------------
 *
 * Readers can lookup /dev links for a specific minor or
 * lookup all /dev links. In the latter case, the node
 * and minor segments are not mapped in and the reader
 * walks through every link in the link segment.
 *
 */
di_devlink_handle_t
di_devlink_open(const char *root_dir, uint_t flags)
{
	int err;
	char path[PATH_MAX];
	struct di_devlink_handle *hdp;
	int retried = 0;

retry:
	/*
	 * Allocate a read-write handle but open the DB in readonly
	 * mode. We do writes only to a temporary copy of the database.
	 */
	if ((hdp = handle_alloc(root_dir, OPEN_RDWR)) == NULL) {
		return (NULL);
	}

	err = open_db(hdp, OPEN_RDONLY);

	/*
	 * We don't want to unlink the db at this point - if we did we
	 * would be creating a window where consumers would take a slow
	 * code path (and those consumers might also trigger requests for
	 * db creation, which we are already in the process of doing).
	 * When we are done with our update, we use rename to install the
	 * latest version of the db file.
	 */
	get_db_path(hdp, DB_FILE, path, sizeof (path));

	/*
	 * The flags argument is reserved for future use.
	 */
	if (flags != 0) {
		handle_free(&hdp); /* also closes the DB */
		errno = EINVAL;
		return (NULL);
	}

	if (cache_alloc(hdp) != 0) {
		handle_free(&hdp);
		return (NULL);
	}

	if (err) {
		/*
		 * Failed to open DB.
		 * The most likely cause is that DB file did not exist.
		 * Call di_devlink_close() to recreate the DB file and
		 * retry di_devlink_open().
		 */
		if (retried == 0) {
			(void) di_devlink_close(&hdp, 0);
			retried = 1;
			goto retry;
		}

		/*
		 * DB cannot be opened, just return the
		 * handle. We will recreate the DB later.
		 */
		return (hdp);
	}

	/* Read the database into the cache */
	CACHE(hdp)->update_count = DB_HDR(hdp)->update_count;
	(void) read_nodes(hdp, NULL, DB_HDR(hdp)->root_idx);
	(void) read_links(hdp, NULL, DB_HDR(hdp)->dngl_idx);

	(void) close_db(hdp);

	return (hdp);
}

static void
get_db_path(
	struct di_devlink_handle *hdp,
	const char *fname,
	char *buf,
	size_t blen)
{
	char *dir = NULL;

#ifdef	DEBUG
	if (dir = getenv(ALT_DB_DIR)) {
		(void) dprintf(DBG_INFO, "get_db_path: alternate db dir: %s\n",
		    dir);
	}
#endif
	if (dir == NULL) {
		dir = hdp->db_dir;
	}

	(void) snprintf(buf, blen, "%s/%s", dir, fname);
}

static int
open_db(struct di_devlink_handle *hdp, int flags)
{
	size_t sz;
	long page_sz;
	int fd, rv, flg;
	struct stat sbuf;
	uint32_t count[DB_TYPES] = {0};
	char path[PATH_MAX];
	void *cp;

	assert(!DB_OPEN(hdp));

#ifdef	DEBUG
	if (getenv(SKIP_DB)) {
		(void) dprintf(DBG_INFO, "open_db: skipping database\n");
		return (-1);
	}
#endif
	if ((page_sz = sysconf(_SC_PAGE_SIZE)) == -1) {
		return (-1);
	}

	/*
	 * Use O_TRUNC flag for write access, so that the subsequent ftruncate()
	 * call will zero-fill the entire file
	 */
	if (IS_RDONLY(flags)) {
		flg = O_RDONLY;
		get_db_path(hdp, DB_FILE, path, sizeof (path));
	} else {
		flg = O_RDWR|O_CREAT|O_TRUNC;
		get_db_path(hdp, DB_TMP, path, sizeof (path));
	}

	/*
	 * Avoid triggering /dev reconfigure for read when not present
	 */
	if (IS_RDONLY(flags) &&
	    (strncmp(path, "/dev/", 5) == 0) && !device_exists(path)) {
		return (-1);
	}

	if ((fd = open(path, flg, DB_PERMS)) == -1) {
		return (-1);
	}

	if (IS_RDONLY(flags)) {
		flg = PROT_READ;
		rv = fstat(fd, &sbuf);
		sz = sbuf.st_size;
	} else {
		flg = PROT_READ | PROT_WRITE;
		sz = size_db(hdp, page_sz, count);
		rv = ftruncate(fd, sz);
	}

	if (rv == -1 || sz < HDR_LEN) {
		if (rv != -1)
			errno = EINVAL;
		(void) close(fd);
		return (-1);
	}

	cp = mmap(0, HDR_LEN, flg, MAP_SHARED, fd, 0);
	if (cp == MAP_FAILED) {
		(void) close(fd);
		return (-1);
	}
	DB(hdp)->hdr = (struct db_hdr *)cp;
	DB(hdp)->db_fd = fd;
	DB(hdp)->flags = flags;

	if (IS_RDONLY(flags)) {
		rv = invalid_db(hdp, sz, page_sz);
	} else {
		rv = init_hdr(hdp, page_sz, count);
	}

	if (rv) {
		(void) dprintf(DBG_ERR, "open_db: invalid DB(%s)\n", path);
		(void) close_db(hdp);
		return (-1);
	} else {
		(void) dprintf(DBG_STEP, "open_db: DB(%s): opened\n", path);
		return (0);
	}
}

/*
 * A handle can be allocated for read-only or read-write access
 */
static struct di_devlink_handle *
handle_alloc(const char *root_dir, uint_t flags)
{
	char dev_dir[PATH_MAX], path[PATH_MAX], db_dir[PATH_MAX];
	struct di_devlink_handle *hdp, proto = {0};
	int install = 0;
	int isroot = 0;
	struct stat sb;
	char can_path[PATH_MAX];

	assert(flags == OPEN_RDWR || flags == OPEN_RDONLY);

	dev_dir[0] = '\0';
	db_dir[0] = '\0';

	/*
	 * NULL and the empty string are equivalent to "/"
	 */
	if (root_dir && root_dir[0] != '\0') {

		if (root_dir[0] != '/') {
			errno = EINVAL;
			return (NULL);
		}

#ifdef	DEBUG
		/*LINTED*/
		assert(sizeof (dev_dir) >= PATH_MAX);
#endif
		if ((realpath(root_dir, dev_dir) == NULL) ||
		    (realpath(root_dir, db_dir) == NULL)) {
			return (NULL);
		}
	} else {
		/*
		 * The dev dir is at /dev i.e. we are not doing a -r /altroot
		 */
		isroot = 1;
	}

	if (strcmp(dev_dir, "/") == 0) {
		dev_dir[0] = 0;
		db_dir[0] = 0;
	} else {
		(void) strlcpy(db_dir, dev_dir, sizeof (db_dir));
	}

	(void) strlcat(dev_dir, DEV, sizeof (dev_dir));
	(void) strlcat(db_dir, ETCDEV, sizeof (db_dir));

	/*
	 * The following code is for install. Readers and writers need
	 * to be redirected to /tmp/etc/dev for the database file.
	 * Note that we test for readonly /etc by actually creating a
	 * file since statvfs is not a reliable method for determining
	 * readonly filesystems.
	 */
	install = 0;
	(void) snprintf(can_path, sizeof (can_path), "%s/%s", ETCDEV, DB_FILE);
	if (flags == OPEN_RDWR && isroot) {
		char di_test_db[PATH_MAX];
		int fd;
		(void) mutex_lock(&temp_file_mutex);
		(void) snprintf(di_test_db, sizeof (di_test_db), "%s.%d",
		    DI_TEST_DB, getpid());
		fd = open(di_test_db, O_CREAT|O_RDWR|O_EXCL, 0644);
		if (fd == -1 && errno == EROFS && stat(can_path, &sb) == -1)
			install = 1;
		if (fd != -1) {
			(void) close(fd);
			(void) unlink(di_test_db);
		}
		(void) mutex_unlock(&temp_file_mutex);
	} else if (isroot) {
		/*
		 * Readers can be non-privileged so we cannot test by creating
		 * a file in /etc/dev. Instead we check if the database
		 * file is missing in /etc/dev and is present in /tmp/etc/dev
		 * and is owned by root.
		 */
		char install_path[PATH_MAX];

		(void) snprintf(install_path, sizeof (install_path),
		    "/tmp%s/%s", ETCDEV, DB_FILE);
		if (stat(can_path, &sb) == -1 && stat(install_path, &sb)
		    != -1 && sb.st_uid == 0) {
			install = 1;
		}
	}

	/*
	 * Check if we are in install. If we are, the database will be in
	 * /tmp/etc/dev
	 */
	if (install)
		(void) snprintf(db_dir, sizeof (db_dir), "/tmp%s", ETCDEV);

	proto.dev_dir = dev_dir;
	proto.db_dir = db_dir;
	proto.flags = flags;
	proto.lock_fd = -1;

	/*
	 * Lock database if a read-write handle is being allocated.
	 * Locks are needed to protect against multiple writers.
	 * Readers don't need locks.
	 */
	if (HDL_RDWR(&proto)) {
		if (enter_db_lock(&proto, root_dir) != 1) {
			return (NULL);
		}
	}

	DB(&proto)->db_fd = -1;

	hdp = calloc(1, sizeof (struct di_devlink_handle));
	if (hdp == NULL) {
		goto error;
	}

	*hdp = proto;

	/*
	 * The handle hdp now contains a pointer to local storage
	 * in the dev_dir field (obtained from the proto handle).
	 * In the following line, a dynamically allocated version
	 * is substituted.
	 */

	if ((hdp->dev_dir = strdup(proto.dev_dir)) == NULL) {
		free(hdp);
		goto error;
	}

	if ((hdp->db_dir = strdup(proto.db_dir)) == NULL) {
		free(hdp->dev_dir);
		free(hdp);
		goto error;
	}

	return (hdp);

error:
	if (HDL_RDWR(&proto)) {
		/* Unlink DB file on error */
		get_db_path(&proto, DB_FILE, path, sizeof (path));
		(void) unlink(path);
		exit_db_lock(&proto);
	}
	return (NULL);
}


static int
cache_alloc(struct di_devlink_handle *hdp)
{
	size_t hash_sz = 0;

	assert(HDL_RDWR(hdp));

	if (DB_OPEN(hdp)) {
		hash_sz = DB_NUM(hdp, DB_LINK) / AVG_CHAIN_SIZE;
	}
	hash_sz = (hash_sz >= MIN_HASH_SIZE) ? hash_sz : MIN_HASH_SIZE;

	CACHE(hdp)->hash = calloc(hash_sz, sizeof (cache_link_t *));
	if (CACHE(hdp)->hash == NULL) {
		return (-1);
	}
	CACHE(hdp)->hash_sz = hash_sz;

	return (0);
}


static int
invalid_db(struct di_devlink_handle *hdp, size_t fsize, long page_sz)
{
	int i;
	char *cp;
	size_t sz;

	if (DB_HDR(hdp)->magic != DB_MAGIC || DB_HDR(hdp)->vers != DB_VERSION) {
		return (1);
	}

	if (DB_HDR(hdp)->page_sz == 0 || DB_HDR(hdp)->page_sz != page_sz) {
		return (1);
	}

	sz = seg_size(hdp, DB_HEADER);
	for (i = 0; i < DB_TYPES; i++) {
		(void) dprintf(DBG_INFO, "N[%u] = %u\n", i, DB_NUM(hdp, i));
		/* There must be at least 1 element of each type */
		if (DB_NUM(hdp, i) < 1) {
			return (1);
		}
		sz += seg_size(hdp, i);
		assert(sz % page_sz == 0);
	}

	if (sz != fsize) {
		return (1);
	}

	if (!VALID_INDEX(hdp, DB_NODE, DB_HDR(hdp)->root_idx)) {
		return (1);
	}

	if (!VALID_INDEX(hdp, DB_LINK, DB_HDR(hdp)->dngl_idx)) {
		return (1);
	}

	if (DB_EMPTY(hdp)) {
		return (1);
	}

	/*
	 * The last character in the string segment must be a NUL char.
	 */
	cp = get_string(hdp, DB_NUM(hdp, DB_STR) - 1);
	if (cp == NULL || *cp != '\0') {
		return (1);
	}

	return (0);
}

static int
read_nodes(struct di_devlink_handle *hdp, cache_node_t *pcnp, uint32_t nidx)
{
	char *path;
	cache_node_t *cnp;
	struct db_node *dnp;
	const char *fcn = "read_nodes";

	assert(HDL_RDWR(hdp));

	/*
	 * parent node should be NULL only for the root node
	 */
	if ((pcnp == NULL) ^ (nidx == DB_HDR(hdp)->root_idx)) {
		(void) dprintf(DBG_ERR, "%s: invalid parent or index(%u)\n",
		    fcn, nidx);
		SET_DB_ERR(hdp);
		return (-1);
	}

	for (; dnp = get_node(hdp, nidx); nidx = dnp->sib) {

		path = get_string(hdp, dnp->path);

		/*
		 * Insert at head of list to recreate original order
		 */
		cnp = node_insert(hdp, pcnp, path, INSERT_HEAD);
		if (cnp == NULL) {
			SET_DB_ERR(hdp);
			break;
		}

		assert(strcmp(path, "/") ^ (nidx == DB_HDR(hdp)->root_idx));
		assert(strcmp(path, "/") != 0 || dnp->sib == DB_NIL);

		if (read_minors(hdp, cnp, dnp->minor) != 0 ||
		    read_nodes(hdp, cnp, dnp->child) != 0) {
			break;
		}

		(void) dprintf(DBG_STEP, "%s: node[%u]: %s\n", fcn, nidx,
		    cnp->path);
	}

	return (dnp ? -1 : 0);
}

static int
read_minors(struct di_devlink_handle *hdp, cache_node_t *pcnp, uint32_t nidx)
{
	cache_minor_t *cmnp;
	struct db_minor *dmp;
	char *name, *nodetype;
	const char *fcn = "read_minors";

	assert(HDL_RDWR(hdp));

	if (pcnp == NULL) {
		(void) dprintf(DBG_ERR, "%s: minor[%u]: orphan minor\n", fcn,
		    nidx);
		SET_DB_ERR(hdp);
		return (-1);
	}

	for (; dmp = get_minor(hdp, nidx); nidx = dmp->sib) {

		name = get_string(hdp, dmp->name);
		nodetype = get_string(hdp, dmp->nodetype);

		cmnp = minor_insert(hdp, pcnp, name, nodetype, NULL);
		if (cmnp == NULL) {
			SET_DB_ERR(hdp);
			break;
		}

		(void) dprintf(DBG_STEP, "%s: minor[%u]: %s\n", fcn, nidx,
		    cmnp->name);

		if (read_links(hdp, cmnp, dmp->link) != 0) {
			break;
		}
	}

	return (dmp ? -1 : 0);
}

/*
 * If the link is dangling the corresponding minor will be absent.
 */
static int
read_links(struct di_devlink_handle *hdp, cache_minor_t *pcmp, uint32_t nidx)
{
	cache_link_t *clp;
	struct db_link *dlp;
	char *path, *content;

	assert(HDL_RDWR(hdp));

	if (nidx != DB_NIL &&
	    ((pcmp == NULL) ^ (nidx == DB_HDR(hdp)->dngl_idx))) {
		(void) dprintf(DBG_ERR, "read_links: invalid minor or"
		    " index(%u)\n", nidx);
		SET_DB_ERR(hdp);
		return (-1);
	}

	for (; dlp = get_link(hdp, nidx); nidx = dlp->sib) {

		path = get_string(hdp, dlp->path);
		content = get_string(hdp, dlp->content);

		clp = link_insert(hdp, pcmp, path, content, dlp->attr);
		if (clp == NULL) {
			SET_DB_ERR(hdp);
			break;
		}

		(void) dprintf(DBG_STEP, "read_links: link[%u]: %s%s\n",
		    nidx, clp->path, pcmp == NULL ? "(DANGLING)" : "");
	}

	return (dlp ? -1 : 0);
}

int
di_devlink_close(di_devlink_handle_t *pp, int flag)
{
	int i, rv;
	char tmp[PATH_MAX];
	char file[PATH_MAX];
	uint32_t next[DB_TYPES] = {0};
	struct di_devlink_handle *hdp;

	if (pp == NULL || *pp == NULL || !HDL_RDWR(*pp)) {
		errno = EINVAL;
		return (-1);
	}

	hdp = *pp;
	*pp = NULL;

	/*
	 * The caller encountered some error in their processing.
	 * so handle isn't valid. Discard it and return success.
	 */
	if (flag == DI_LINK_ERROR) {
		handle_free(&hdp);
		return (0);
	}

	if (DB_ERR(hdp)) {
		handle_free(&hdp);
		errno = EINVAL;
		return (-1);
	}

	/*
	 * Extract the DB path before the handle is freed.
	 */
	get_db_path(hdp, DB_FILE, file, sizeof (file));
	get_db_path(hdp, DB_TMP, tmp, sizeof (tmp));

	/*
	 * update database with actual contents of /dev
	 */
	(void) dprintf(DBG_INFO, "di_devlink_close: update_count = %u\n",
	    CACHE(hdp)->update_count);

	/*
	 * For performance reasons, synchronization of the database
	 * with /dev is turned off by default. However, applications
	 * with appropriate permissions can request a "sync" by
	 * calling di_devlink_update().
	 */
	if (CACHE(hdp)->update_count == 0) {
		CACHE(hdp)->update_count = 1;
		(void) dprintf(DBG_INFO,
		    "di_devlink_close: synchronizing DB\n");
		(void) synchronize_db(hdp);
	}

	/*
	 * Resolve dangling links AFTER synchronizing DB with /dev as the
	 * synchronization process may create dangling links.
	 */
	resolve_dangling_links(hdp);

	/*
	 * All changes to the cache are complete. Write out the cache
	 * to the database only if it is not empty.
	 */
	if (CACHE_EMPTY(hdp)) {
		(void) dprintf(DBG_INFO, "di_devlink_close: skipping write\n");
		(void) unlink(file);
		handle_free(&hdp);
		return (0);
	}

	if (open_db(hdp, OPEN_RDWR) != 0) {
		handle_free(&hdp);
		return (-1);
	}

	/*
	 * Keep track of array assignments. There is at least
	 * 1 element (the "NIL" element) per type.
	 */
	for (i = 0; i < DB_TYPES; i++) {
		next[i] = 1;
	}

	(void) write_nodes(hdp, NULL, CACHE_ROOT(hdp), next);
	(void) write_links(hdp, NULL, CACHE(hdp)->dngl, next);
	DB_HDR(hdp)->update_count = CACHE(hdp)->update_count;

	rv = close_db(hdp);

	if (rv != 0 || DB_ERR(hdp) || rename(tmp, file) != 0) {
		(void) dprintf(DBG_ERR, "di_devlink_close: %s error: %s\n",
		    rv ? "close_db" : "DB or rename", strerror(errno));
		(void) unlink(tmp);
		(void) unlink(file);
		handle_free(&hdp);
		return (-1);
	}

	handle_free(&hdp);

	(void) dprintf(DBG_INFO, "di_devlink_close: wrote DB(%s)\n", file);

	return (0);
}

/*
 * Inits the database header.
 */
static int
init_hdr(struct di_devlink_handle *hdp, long page_sz, uint32_t *count)
{
	int i;

	DB_HDR(hdp)->magic = DB_MAGIC;
	DB_HDR(hdp)->vers = DB_VERSION;
	DB_HDR(hdp)->root_idx = DB_NIL;
	DB_HDR(hdp)->dngl_idx = DB_NIL;
	DB_HDR(hdp)->page_sz = (uint32_t)page_sz;

	for (i = 0; i < DB_TYPES; i++) {
		assert(count[i] >= 1);
		DB_NUM(hdp, i) = count[i];
	}

	return (0);
}

static int
write_nodes(
	struct di_devlink_handle *hdp,
	struct db_node *pdnp,
	cache_node_t *cnp,
	uint32_t *next)
{
	uint32_t idx;
	struct db_node *dnp;
	const char *fcn = "write_nodes";

	assert(HDL_RDWR(hdp));

	for (; cnp != NULL; cnp = cnp->sib) {

		assert(cnp->path != NULL);

		/* parent node should only be NULL for root node */
		if ((pdnp == NULL) ^ (cnp == CACHE_ROOT(hdp))) {
			(void) dprintf(DBG_ERR, "%s: invalid parent for: %s\n",
			    fcn, cnp->path);
			SET_DB_ERR(hdp);
			break;
		}

		assert((strcmp(cnp->path, "/") != 0) ^
		    (cnp == CACHE_ROOT(hdp)));

		idx = next[DB_NODE];
		if ((dnp = set_node(hdp, idx)) == NULL) {
			SET_DB_ERR(hdp);
			break;
		}

		dnp->path = write_string(hdp, cnp->path, next);
		if (dnp->path == DB_NIL) {
			SET_DB_ERR(hdp);
			break;
		}
		/* commit write for this node */
		next[DB_NODE]++;

		if (pdnp == NULL) {
			assert(DB_HDR(hdp)->root_idx == DB_NIL);
			DB_HDR(hdp)->root_idx = idx;
		} else {
			dnp->sib = pdnp->child;
			pdnp->child = idx;
		}

		(void) dprintf(DBG_STEP, "%s: node[%u]: %s\n", fcn, idx,
		    cnp->path);

		if (write_minors(hdp, dnp, cnp->minor, next) != 0 ||
		    write_nodes(hdp, dnp, cnp->child, next) != 0) {
			break;
		}
	}

	return (cnp ? -1 : 0);
}

static int
write_minors(
	struct di_devlink_handle *hdp,
	struct db_node *pdnp,
	cache_minor_t *cmnp,
	uint32_t *next)
{
	uint32_t idx;
	struct db_minor *dmp;
	const char *fcn = "write_minors";

	assert(HDL_RDWR(hdp));

	if (pdnp == NULL) {
		(void) dprintf(DBG_ERR, "%s: no node for minor: %s\n", fcn,
		    cmnp ? cmnp->name : "<NULL>");
		SET_DB_ERR(hdp);
		return (-1);
	}

	for (; cmnp != NULL; cmnp = cmnp->sib) {

		assert(cmnp->name != NULL);

		idx = next[DB_MINOR];
		if ((dmp = set_minor(hdp, idx)) == NULL) {
			SET_DB_ERR(hdp);
			break;
		}

		dmp->name = write_string(hdp, cmnp->name, next);
		dmp->nodetype = write_string(hdp, cmnp->nodetype, next);
		if (dmp->name == DB_NIL || dmp->nodetype == DB_NIL) {
			dmp->name = dmp->nodetype = DB_NIL;
			SET_DB_ERR(hdp);
			break;
		}

		/* Commit writes to this minor */
		next[DB_MINOR]++;

		dmp->sib = pdnp->minor;
		pdnp->minor = idx;

		(void) dprintf(DBG_STEP, "%s: minor[%u]: %s\n", fcn, idx,
		    cmnp->name);

		if (write_links(hdp, dmp, cmnp->link, next) != 0) {
			break;
		}
	}

	return (cmnp ? -1 : 0);
}

static int
write_links(
	struct di_devlink_handle *hdp,
	struct db_minor *pdmp,
	cache_link_t *clp,
	uint32_t *next)
{
	uint32_t idx;
	struct db_link *dlp;
	const char *fcn = "write_links";

	assert(HDL_RDWR(hdp));

	/* A NULL minor if and only if the links are dangling */
	if (clp != NULL && ((pdmp == NULL) ^ (clp == CACHE(hdp)->dngl))) {
		(void) dprintf(DBG_ERR, "%s: invalid minor for link\n", fcn);
		SET_DB_ERR(hdp);
		return (-1);
	}

	for (; clp != NULL; clp = clp->sib) {

		assert(clp->path != NULL);

		if ((pdmp == NULL) ^ (clp->minor == NULL)) {
			(void) dprintf(DBG_ERR, "%s: invalid minor for link"
			    "(%s)\n", fcn, clp->path);
			SET_DB_ERR(hdp);
			break;
		}

		idx = next[DB_LINK];
		if ((dlp = set_link(hdp, idx)) == NULL) {
			SET_DB_ERR(hdp);
			break;
		}

		dlp->path = write_string(hdp, clp->path, next);
		dlp->content = write_string(hdp, clp->content, next);
		if (dlp->path == DB_NIL || dlp->content == DB_NIL) {
			dlp->path = dlp->content = DB_NIL;
			SET_DB_ERR(hdp);
			break;
		}

		dlp->attr = clp->attr;

		/* Commit writes to this link */
		next[DB_LINK]++;

		if (pdmp != NULL) {
			dlp->sib = pdmp->link;
			pdmp->link = idx;
		} else {
			dlp->sib = DB_HDR(hdp)->dngl_idx;
			DB_HDR(hdp)->dngl_idx = idx;
		}

		(void) dprintf(DBG_STEP, "%s: link[%u]: %s%s\n", fcn, idx,
		    clp->path, pdmp == NULL ? "(DANGLING)" : "");
	}

	return (clp ? -1 : 0);
}


static uint32_t
write_string(struct di_devlink_handle *hdp, const char *str, uint32_t *next)
{
	char *dstr;
	uint32_t idx;

	assert(HDL_RDWR(hdp));

	if (str == NULL) {
		(void) dprintf(DBG_ERR, "write_string: NULL argument\n");
		return (DB_NIL);
	}

	idx = next[DB_STR];
	if (!VALID_STR(hdp, idx, str)) {
		(void) dprintf(DBG_ERR, "write_string: invalid index[%u],"
		    " string(%s)\n", idx, str);
		return (DB_NIL);
	}

	if ((dstr = set_string(hdp, idx)) == NULL) {
		return (DB_NIL);
	}

	(void) strcpy(dstr, str);

	next[DB_STR] += strlen(dstr) + 1;

	return (idx);
}

static int
close_db(struct di_devlink_handle *hdp)
{
	int i, rv = 0;
	size_t sz;

	if (!DB_OPEN(hdp)) {
#ifdef	DEBUG
		assert(DB(hdp)->db_fd == -1);
		assert(DB(hdp)->flags == 0);
		for (i = 0; i < DB_TYPES; i++) {
			assert(DB_SEG(hdp, i) == NULL);
			assert(DB_SEG_PROT(hdp, i) == 0);
		}
#endif
		return (0);
	}

	/* Unmap header after unmapping all other mapped segments */
	for (i = 0; i < DB_TYPES; i++) {
		if (DB_SEG(hdp, i)) {
			sz = seg_size(hdp, i);
			if (DB_RDWR(hdp))
				rv += msync(DB_SEG(hdp, i), sz, MS_SYNC);
			(void) munmap(DB_SEG(hdp, i), sz);
			DB_SEG(hdp, i) = NULL;
			DB_SEG_PROT(hdp, i) = 0;
		}
	}

	if (DB_RDWR(hdp))
		rv += msync((caddr_t)DB_HDR(hdp), HDR_LEN, MS_SYNC);
	(void) munmap((caddr_t)DB_HDR(hdp), HDR_LEN);
	DB(hdp)->hdr = NULL;

	(void) close(DB(hdp)->db_fd);
	DB(hdp)->db_fd = -1;
	DB(hdp)->flags = 0;

	return (rv ? -1 : 0);
}


static void
cache_free(struct di_devlink_handle *hdp)
{
	cache_link_t *clp;

	subtree_free(hdp, &(CACHE_ROOT(hdp)));
	assert(CACHE_LAST(hdp) == NULL);

	/*
	 * Don't bother removing links from hash table chains,
	 * as we are freeing the hash table itself.
	 */
	while (CACHE(hdp)->dngl != NULL) {
		clp = CACHE(hdp)->dngl;
		CACHE(hdp)->dngl = clp->sib;
		assert(clp->minor == NULL);
		link_free(&clp);
	}

	assert((CACHE(hdp)->hash == NULL) ^ (CACHE(hdp)->hash_sz != 0));

	free(CACHE(hdp)->hash);
	CACHE(hdp)->hash = NULL;
	CACHE(hdp)->hash_sz = 0;
}

static void
handle_free(struct di_devlink_handle **pp)
{
	struct di_devlink_handle *hdp = *pp;

	*pp = NULL;

	if (hdp == NULL)
		return;

	(void) close_db(hdp);
	cache_free(hdp);

	if (HDL_RDWR(hdp))
		exit_db_lock(hdp);
	assert(hdp->lock_fd == -1);

	free(hdp->dev_dir);
	free(hdp->db_dir);
	free(hdp);
}

/*
 * Frees the tree rooted at a node. Siblings of the subtree root
 * have to be handled by the caller.
 */
static void
subtree_free(struct di_devlink_handle *hdp, cache_node_t **pp)
{
	cache_node_t *np;
	cache_link_t *clp;
	cache_minor_t *cmnp;

	if (pp == NULL || *pp == NULL)
		return;

	while ((*pp)->child != NULL) {
		np = (*pp)->child;
		(*pp)->child = np->sib;
		subtree_free(hdp, &np);
	}

	while ((*pp)->minor != NULL) {
		cmnp = (*pp)->minor;
		(*pp)->minor = cmnp->sib;

		while (cmnp->link != NULL) {
			clp = cmnp->link;
			cmnp->link = clp->sib;
			rm_link_from_hash(hdp, clp);
			link_free(&clp);