xref: /illumos-gate/usr/src/lib/libdevinfo/devinfo_devlink.h (revision eff1e82143649b9c02592b224b2d4f2dba1274c5)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  *
21  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
22  * Use is subject to license terms.
23  * Copyright 2026 Oxide Computer Company
24  */
25 
26 
27 #ifndef _DEVINFO_DEVLINK_H
28 #define	_DEVINFO_DEVLINK_H
29 
30 #ifdef	__cplusplus
31 extern "C" {
32 #endif
33 
34 #define	_POSIX_PTHREAD_SEMANTICS	/* For readdir_r */
35 
36 #include <stdio.h>
37 #include <unistd.h>
38 #include <fcntl.h>
39 #include <string.h>
40 #include <thread.h>
41 #include <synch.h>
42 #include <libdevinfo.h>
43 #include <limits.h>
44 #include <stdlib.h>
45 #include <dirent.h>
46 #include <regex.h>
47 #include <errno.h>
48 #include <stdarg.h>
49 #include <sys/uio.h>
50 #include <sys/types.h>
51 #include <sys/stat.h>
52 #include <sys/time.h>
53 #include <sys/mman.h>
54 #include <sys/wait.h>
55 #include <door.h>
56 #include <signal.h>
57 #include <sys/statvfs.h>
58 #include <sys/ccompile.h>
59 #include <stdbool.h>
60 #include <time.h>
61 #include <sys/crc32.h>
62 
63 struct db_link {
64 	uint32_t attr;		/* primary or secondary */
65 	uint32_t path;		/* link path */
66 	uint32_t content;	/* link content */
67 	uint32_t sib;		/* next link for same minor */
68 };
69 
70 struct db_minor {
71 	uint32_t name;		/* minor name */
72 	uint32_t nodetype;	/* minor node type */
73 	uint32_t sib;		/* next minor for same node */
74 	uint32_t link;		/* next minor for same node */
75 };
76 
77 struct db_node {
78 	uint32_t path;		/* node path */
79 	uint32_t sib;		/* node's sibling */
80 	uint32_t child;		/* first child for this node */
81 	uint32_t minor;		/* first minor for node */
82 };
83 
84 typedef enum db_seg {
85 	DB_NODE = 0,
86 	DB_MINOR,
87 	DB_LINK,
88 	DB_STR,
89 	DB_TYPES,	/* Number of non-header segments */
90 	DB_HEADER
91 } db_seg_t;
92 
93 struct db_hdr {
94 	uint32_t magic;			/* Magic number	*/
95 	uint32_t vers;			/* database format version */
96 	uint32_t root_idx;		/* index for root node */
97 	uint32_t dngl_idx;		/* head of DB dangling links */
98 	uint32_t page_sz;		/* page size for mmap alignment	*/
99 	uint32_t update_count;		/* updates since last /dev synch up */
100 	uint32_t nelems[DB_TYPES];	/* Number of elements of each type */
101 	/*
102 	 * The following fields allow database corruption to be detected.
103 	 * The CRCs cover each segment's data and are stored by the writer.
104 	 * The writer identity fields allow a reader which detects
105 	 * corruption to report which process produced the database, and
106 	 * when.
107 	 */
108 	uint32_t crc[DB_TYPES];		/* CRC32 of each segment */
109 	uint32_t writer_pid;		/* process which wrote the DB */
110 	uint32_t hdr_pad;		/* explicit padding, always zero */
111 	uint64_t writer_time;		/* write time, seconds since epoch */
112 	char writer_exec[24];		/* basename of writer's executable */
113 };
114 
115 
116 typedef	struct cache_link {
117 	char   *path;			/* link path */
118 	char   *content;		/* link content	*/
119 	uint_t attr;			/* link attributes */
120 	struct cache_link *hash;	/* next link on same hash chain */
121 	struct cache_link *sib;		/* next link for same minor */
122 	struct cache_minor *minor;	/* minor for this link */
123 } cache_link_t;
124 
125 typedef	struct cache_minor {
126 	char *name;			/* minor name */
127 	char *nodetype;			/* minor nodetype */
128 	struct cache_node *node;	/* node for this minor */
129 	struct cache_minor *sib;	/* next minor for same node */
130 	struct cache_link *link;	/* first link pointing to minor */
131 } cache_minor_t;
132 
133 typedef struct cache_node {
134 	char	*path;			/* path	*/
135 	struct cache_node *parent;	/* node's parent */
136 	struct cache_node *sib;		/* node's sibling */
137 	struct cache_node *child;	/* first child for this node */
138 	struct cache_minor *minor;	/* first minor for node */
139 } cache_node_t;
140 
141 struct cache {
142 	uint_t	flags;			/* cache state */
143 	uint_t	update_count;		/* updates since /dev synchronization */
144 	uint_t	hash_sz;		/* number of hash chains */
145 	cache_link_t **hash;		/* hash table */
146 	cache_node_t *root;		/* root of cache tree */
147 	cache_link_t *dngl;		/* list of dangling links */
148 	cache_minor_t *last_minor;	/* last minor looked up	*/
149 };
150 
151 struct db {
152 	int db_fd;			/* database file */
153 	uint_t flags;			/* database open mode */
154 	struct db_hdr *hdr;		/* DB header */
155 	int  seg_prot[DB_TYPES];	/* protection for  segments */
156 	caddr_t seg_base[DB_TYPES];	/* base address for segments */
157 	uint32_t str_crc;		/* running CRC of written strings */
158 };
159 
160 struct di_devlink_handle {
161 	char *dev_dir;			/* <root-dir>/dev */
162 	char *db_dir;			/* <root-dir>/etc/dev */
163 	uint_t	flags;			/* handle flags	*/
164 	uint_t  error;			/* records errors encountered */
165 	int lock_fd;			/* lock file for updates */
166 	struct cache cache;
167 	struct db db;
168 };
169 
170 typedef struct link_desc {
171 	regex_t *regp;
172 	const char *minor_path;
173 	uint_t flags;
174 	void *arg;
175 	int (*fcn)(di_devlink_t, void *);
176 	int retval;
177 } link_desc_t;
178 
179 struct tnode {
180 	void *node;
181 	int flags;
182 	struct di_devlink_handle *handle;
183 };
184 
185 struct di_devlink {
186 	char *rel_path;
187 	char *abs_path;
188 	char *content;
189 	int type;
190 };
191 
192 typedef struct recurse {
193 	void *data;
194 	int (*fcn)(struct di_devlink_handle *, void *, const char *);
195 } recurse_t;
196 
197 /*
198  * Debug levels currently defined.
199  */
200 typedef enum {
201 	DBG_ERR = 1,
202 	DBG_LCK,
203 	DBG_INFO,
204 	DBG_STEP,
205 	DBG_ALL
206 } debug_level_t;
207 
208 
209 #define	DB_MAGIC	0xBAC2ACAB
210 #define	DB_FILE		".devlink_db"
211 #define	DB_TMP		".devlink_db_tmp"
212 #define	DB_LOCK		".devlink_db_lock"
213 #define	DB_CORRUPT	".devlink_db_corrupt"
214 #define	DB_PERMS	(S_IRUSR | S_IRGRP | S_IROTH | S_IWUSR)
215 #define	DB_LOCK_PERMS	DB_PERMS
216 #define	DB_VERSION	2	/* version 2 added CRCs and writer identity */
217 
218 #define	DB_NIL		0
219 
220 #define	DEV		"/dev"
221 #define	ETCDEV		"/etc/dev"
222 #define	DEVICES_SUFFIX	"ices"
223 
224 #define	HDR_LEN			sizeof (struct db_hdr)
225 
226 #define	AVG_CHAIN_SIZE		20   /* Average number of links per chain */
227 #define	MIN_HASH_SIZE		1024 /* Min number of chains in hash table */
228 #define	MAX_UPDATE_INTERVAL	5 /* Max DB writes before synching with /dev */
229 #define	MAX_LOCK_RETRY		5 /* Max attempts at locking the update lock */
230 
231 /*
232  * Various flags private to the implementation
233  */
234 #define	A_PRIMARY		0x0001U
235 #define	A_SECONDARY		0x0002U
236 #define	A_LINK_TYPES		0x0003U	/* Mask */
237 #define	A_VALID			0x0004U
238 
239 #define	TYPE_DB			0x0008U
240 #define	TYPE_CACHE		0x0010U
241 #define	CREATE_FLAG		0x0020U
242 
243 #define	INSERT_HEAD		0x0040U
244 #define	INSERT_TAIL		0x0080U
245 #define	OPEN_RDWR		0x0100U
246 #define	OPEN_RDONLY		0x0200U
247 #define	OPEN_FLAGS		0x0300U	/* Mask */
248 #define	UNLINK_FROM_HASH	0x0400U
249 
250 #define	SET_VALID_ATTR(a)	((a) |= A_VALID)
251 #define	CLR_VALID_ATTR(a)	((a) &= ~A_VALID)
252 #define	GET_VALID_ATTR(a)	((a) & A_VALID)
253 
254 #define	SET_DB_ERR(h)	((h)->error = 1)
255 #define	DB_ERR(h)	((h)->error)
256 
257 #define	LOOKUP_DB(f)	((f) & TYPE_DB)
258 #define	LOOKUP_CACHE(f)	((f) & TYPE_CACHE)
259 #define	CREATE_ELEM(f)	((f) & CREATE_FLAG)
260 
261 #define	IS_RDWR(f)	(((f) & OPEN_FLAGS) == OPEN_RDWR)
262 #define	IS_RDONLY(f)	(((f) & OPEN_FLAGS) == OPEN_RDONLY)
263 
264 #define	HDL_RDWR(h)	(((h)->flags & OPEN_FLAGS) == OPEN_RDWR)
265 #define	HDL_RDONLY(h)	(((h)->flags & OPEN_FLAGS) == OPEN_RDONLY)
266 
267 #define	CACHE(h)		(&(h)->cache)
268 #define	CACHE_ROOT(h)		(CACHE(h)->root)
269 #define	CACHE_HASH(h, i)	(CACHE(h)->hash[i])
270 #define	CACHE_LAST(h)	(CACHE(h)->last_minor)
271 #define	CACHE_EMPTY(h)	(CACHE(h)->root == NULL && CACHE(h)->dngl == NULL)
272 
273 #define	DB(h)			(&(h)->db)
274 #define	DB_HDR(h)		(DB(h)->hdr)
275 #define	DB_NUM(h, t)		(DB_HDR(h)->nelems[t])
276 #define	DB_SEG(h, t)		(DB(h)->seg_base[t])
277 #define	DB_SEG_PROT(h, t)	(DB(h)->seg_prot[t])
278 
279 #define	DB_OPEN(h)	(DB_HDR(h) != NULL)
280 #define	DB_RDWR(h)	((DB(h)->flags & OPEN_FLAGS) == OPEN_RDWR)
281 #define	DB_RDONLY(h)	((DB(h)->flags & OPEN_FLAGS) == OPEN_RDONLY)
282 
283 #define	DB_EMPTY(h)	(DB_HDR(h)->root_idx == DB_NIL && \
284 			    DB_HDR(h)->dngl_idx == DB_NIL)
285 
286 #define	TYPE_NONE(f)	(((f) & DI_LINK_TYPES) == 0)
287 #define	TYPE_PRI(f)	(((f) & DI_LINK_TYPES) == DI_PRIMARY_LINK)
288 #define	TYPE_SEC(f)	(((f) & DI_LINK_TYPES) == DI_SECONDARY_LINK)
289 #define	LINK_TYPE(f)	((f) & DI_LINK_TYPES)
290 #define	VALID_TYPE(f)	(TYPE_NONE(f) || TYPE_PRI(f) || TYPE_SEC(f))
291 
292 #define	VALID_STR(h, i, s)   ((i) + strlen(s) + 1 <= DB_HDR(h)->nelems[DB_STR])
293 #define	VALID_INDEX(h, t, i) ((i) < DB_HDR(h)->nelems[t])
294 
295 /*
296  * Environment variables used by DEBUG version of code.
297  */
298 #define	SKIP_DB		"DEBUG_SKIP_DB"
299 #define	SKIP_LAST_CACHE	"DEBUG_SKIP_LAST_CACHE"
300 #define	ALT_DB_DIR	"DEBUG_ALT_DB_DIR"
301 
302 /*
303  * Function prototypes
304  */
305 static struct di_devlink_handle *handle_alloc(const char *dev_dir,
306     uint_t flags);
307 static int cache_alloc(struct di_devlink_handle *hdp);
308 static int open_db(struct di_devlink_handle *hdp, int flags);
309 static int invalid_db(struct di_devlink_handle *hdp, size_t fsize, long pg_sz);
310 static int read_nodes(struct di_devlink_handle *hdp, cache_node_t *pcnp,
311     uint32_t nidx);
312 static int read_minors(struct di_devlink_handle *hdp, cache_node_t *pcnp,
313     uint32_t nidx);
314 static int read_links(struct di_devlink_handle *hdp, cache_minor_t *pcmp,
315     uint32_t nidx);
316 static int init_hdr(struct di_devlink_handle *hdp, long page_sz,
317     uint32_t *count);
318 static size_t size_db(struct di_devlink_handle *hdp, long page_sz,
319     uint32_t *count);
320 static size_t seg_size(struct di_devlink_handle *hdp, int seg);
321 
322 static cache_node_t *node_insert(struct di_devlink_handle *hdp,
323     cache_node_t *pcnp, const char *path, int insert);
324 static cache_minor_t *minor_insert(struct di_devlink_handle *hdp,
325     cache_node_t *pcnp, const char *name, const char *nodetype,
326     cache_minor_t **prev);
327 static cache_link_t *link_insert(struct di_devlink_handle *hdp,
328     cache_minor_t *mnp, const char *path, const char *content, uint32_t attr);
329 
330 static void minor_delete(di_devlink_handle_t hdp, cache_minor_t *cmnp);
331 static void link_delete(di_devlink_handle_t hdp, cache_link_t *clp);
332 
333 static int write_nodes(struct di_devlink_handle *hdp, struct db_node *pdnp,
334     cache_node_t *cnp, uint32_t *next);
335 static int write_minors(struct di_devlink_handle *hdp, struct db_node *pdnp,
336     cache_minor_t *cmnp, uint32_t *next);
337 static int write_links(struct di_devlink_handle *hdp, struct db_minor *pdmp,
338     cache_link_t *clp, uint32_t *next);
339 static void rm_link_from_hash(struct di_devlink_handle *hdp, cache_link_t *clp);
340 static uint32_t write_string(struct di_devlink_handle *hdp, const char *str,
341     uint32_t *next);
342 static int close_db(struct di_devlink_handle *hdp);
343 static uint32_t devlink_crc32(uint32_t crc, const void *buf, size_t len);
344 static uint32_t segment_crc(struct di_devlink_handle *hdp, db_seg_t seg,
345     int prot);
346 static bool verify_db_crc(struct di_devlink_handle *hdp);
347 static void seal_db(struct di_devlink_handle *hdp, uint32_t *next);
348 static void devlink_db_fault(struct di_devlink_handle *hdp,
349     const char *fmt, ...) __PRINTFLIKE(2) __NORETURN;
350 static void cache_free(struct di_devlink_handle *hdp);
351 static void handle_free(struct di_devlink_handle **pp);
352 static void resolve_dangling_links(struct di_devlink_handle *hdp);
353 static void subtree_free(struct di_devlink_handle *hdp, cache_node_t **pp);
354 static void node_free(cache_node_t **pp);
355 static void minor_free(struct di_devlink_handle *hdp, cache_minor_t **pp);
356 static void link_free(cache_link_t **pp);
357 static void count_node(cache_node_t *cnp, uint32_t *count);
358 static void count_minor(cache_minor_t *mnp, uint32_t *count);
359 static void count_link(cache_link_t *clp, uint32_t *count);
360 static void count_string(const char *str, uint32_t *count);
361 static int visit_node(const char *path, void *arg);
362 static int walk_tree(char *cur, void *arg,
363     int (*node_callback)(const char *path, void *arg));
364 static void *lookup_node(struct di_devlink_handle *hdp, char *path,
365     const int flags);
366 static cache_link_t *add_link(struct di_devlink_handle *hdp, const char *link,
367     const char *content, int primary);
368 
369 static void *lookup_minor(struct di_devlink_handle *hdp, const char *minor_path,
370     const char *nodetype, const int flags);
371 static cache_link_t *link_hash(di_devlink_handle_t hdp, const char *link,
372     uint_t flags);
373 
374 static void hash_insert(struct di_devlink_handle *hdp, cache_link_t *clp);
375 static uint_t hashfn(struct di_devlink_handle *hdp, const char *str);
376 static void get_db_path(struct di_devlink_handle *hdp, const char *fname,
377     char *buf, size_t blen);
378 
379 static struct db_node *get_node(struct di_devlink_handle *hdp, uint32_t idx);
380 static struct db_node *set_node(struct di_devlink_handle *hdp, uint32_t idx);
381 
382 static struct db_minor *get_minor(struct di_devlink_handle *hdp, uint32_t idx);
383 static struct db_minor *set_minor(struct di_devlink_handle *hdp, uint32_t idx);
384 
385 static struct db_link *get_link(struct di_devlink_handle *hdp, uint32_t idx);
386 static struct db_link *set_link(struct di_devlink_handle *hdp, uint32_t idx);
387 
388 static char *get_string(struct di_devlink_handle *hdp, uint32_t idx);
389 static char *set_string(struct di_devlink_handle *hdp, uint32_t idx);
390 
391 static void *map_seg(struct di_devlink_handle *hdp, uint32_t idx, int prot,
392     db_seg_t seg);
393 
394 static int walk_db(struct di_devlink_handle *hdp, link_desc_t *linkp);
395 static int walk_all_links(struct di_devlink_handle *hdp, link_desc_t *linkp);
396 static int walk_matching_links(struct di_devlink_handle *hdp,
397     link_desc_t *linkp);
398 static int visit_link(struct di_devlink_handle *hdp, link_desc_t *linkp,
399     struct di_devlink *vlp);
400 
401 static void walk_cache_minor(di_devlink_handle_t hdp, const char *mpath,
402     link_desc_t *linkp);
403 static int walk_cache_links(di_devlink_handle_t hdp, cache_link_t *clp,
404     link_desc_t *linkp);
405 static void walk_all_cache(di_devlink_handle_t hdp, link_desc_t *linkp);
406 static int cache_dev_link(struct di_devlink_handle *hdp, void *data,
407     const char *link_path);
408 
409 static int walk_dev(struct di_devlink_handle *hdp, link_desc_t *linkp);
410 static int recurse_dev(struct di_devlink_handle *hdp, recurse_t *rp);
411 static int do_recurse(const char *dir, struct di_devlink_handle *hdp,
412     recurse_t *rp, int *retp);
413 
414 static int check_attr(uint32_t attr);
415 static int attr2type(uint32_t attr);
416 
417 static int check_args(link_desc_t *linkp);
418 
419 static void *get_last_node(struct di_devlink_handle *hdp, const char *path,
420     int flags);
421 static void *get_last_minor(struct di_devlink_handle *hdp,
422     const char *devfs_path, const char *minor_name, int flags);
423 static void set_last_minor(struct di_devlink_handle *hdp, cache_minor_t *cmnp,
424     int flags);
425 
426 static int enter_db_lock(struct di_devlink_handle *hdp, const char *root_dir);
427 static void exit_db_lock(struct di_devlink_handle *hdp);
428 
429 static char *minor_colon(const char *path);
430 static const char *rel_path(struct di_devlink_handle *hdp, const char *path);
431 static int link_flag(uint_t flags);
432 static int s_readlink(const char *link, char *buf, size_t blen);
433 static cache_minor_t *link2minor(struct di_devlink_handle *hdp,
434     cache_link_t *clp);
435 static int link_cmp(cache_link_t *clp, const char *content, int type);
436 static void delete_unused_nodes(di_devlink_handle_t hdp, cache_node_t *cnp);
437 static void delete_unused_minor(di_devlink_handle_t hdp, cache_minor_t *cmnp);
438 static int synchronize_db(di_devlink_handle_t hdp);
439 static void devlink_dprintf(debug_level_t msglevel, const char *fmt, ...);
440 static di_devlink_handle_t devlink_snapshot(const char *root_dir);
441 static int devlink_create(const char *root, const char *name, int dca_flags);
442 static int dca_init(const char *name, struct dca_off *dcp, int dca_flags);
443 static void exec_cmd(const char *root, struct dca_off *dcp);
444 static int do_exec(const char *path, char *const argv[]);
445 static int start_daemon(const char *root, int install);
446 static int daemon_call(const char *root, struct dca_off *dcp);
447 
448 int is_minor_node(const char *contents, const char **mn_root);
449 char *s_realpath(const char *path, char *resolved_path);
450 
451 #ifdef	__cplusplus
452 }
453 #endif
454 
455 #endif /* _DEVINFO_DEVLINK_H */
456