xref: /illumos-gate/usr/src/uts/common/sys/zone.h (revision 81d43577)
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 /*
22  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright 2014 Nexenta Systems, Inc. All rights reserved.
24  */
25 
26 #ifndef _SYS_ZONE_H
27 #define	_SYS_ZONE_H
28 
29 #include <sys/types.h>
30 #include <sys/mutex.h>
31 #include <sys/param.h>
32 #include <sys/rctl.h>
33 #include <sys/ipc_rctl.h>
34 #include <sys/pset.h>
35 #include <sys/tsol/label.h>
36 #include <sys/cred.h>
37 #include <sys/netstack.h>
38 #include <sys/uadmin.h>
39 #include <sys/ksynch.h>
40 #include <sys/socket_impl.h>
41 #include <netinet/in.h>
42 
43 #ifdef	__cplusplus
44 extern "C" {
45 #endif
46 
47 /*
48  * NOTE
49  *
50  * The contents of this file are private to the implementation of
51  * Solaris and are subject to change at any time without notice.
52  * Applications and drivers using these interfaces may fail to
53  * run on future releases.
54  */
55 
56 /* Available both in kernel and for user space */
57 
58 /* zone id restrictions and special ids */
59 #define	MAX_ZONEID	9999
60 #define	MIN_USERZONEID	1	/* lowest user-creatable zone ID */
61 #define	MIN_ZONEID	0	/* minimum zone ID on system */
62 #define	GLOBAL_ZONEID	0
63 #define	ZONEID_WIDTH	4	/* for printf */
64 
65 /*
66  * Special zoneid_t token to refer to all zones.
67  */
68 #define	ALL_ZONES	(-1)
69 
70 /* system call subcodes */
71 #define	ZONE_CREATE		0
72 #define	ZONE_DESTROY		1
73 #define	ZONE_GETATTR		2
74 #define	ZONE_ENTER		3
75 #define	ZONE_LIST		4
76 #define	ZONE_SHUTDOWN		5
77 #define	ZONE_LOOKUP		6
78 #define	ZONE_BOOT		7
79 #define	ZONE_VERSION		8
80 #define	ZONE_SETATTR		9
81 #define	ZONE_ADD_DATALINK	10
82 #define	ZONE_DEL_DATALINK	11
83 #define	ZONE_CHECK_DATALINK	12
84 #define	ZONE_LIST_DATALINK	13
85 
86 /* zone attributes */
87 #define	ZONE_ATTR_ROOT		1
88 #define	ZONE_ATTR_NAME		2
89 #define	ZONE_ATTR_STATUS	3
90 #define	ZONE_ATTR_PRIVSET	4
91 #define	ZONE_ATTR_UNIQID	5
92 #define	ZONE_ATTR_POOLID	6
93 #define	ZONE_ATTR_INITPID	7
94 #define	ZONE_ATTR_SLBL		8
95 #define	ZONE_ATTR_INITNAME	9
96 #define	ZONE_ATTR_BOOTARGS	10
97 #define	ZONE_ATTR_BRAND		11
98 #define	ZONE_ATTR_PHYS_MCAP	12
99 #define	ZONE_ATTR_SCHED_CLASS	13
100 #define	ZONE_ATTR_FLAGS		14
101 #define	ZONE_ATTR_HOSTID	15
102 #define	ZONE_ATTR_FS_ALLOWED	16
103 #define	ZONE_ATTR_NETWORK	17
104 
105 /* Start of the brand-specific attribute namespace */
106 #define	ZONE_ATTR_BRAND_ATTRS	32768
107 
108 #define	ZONE_FS_ALLOWED_MAX	1024
109 
110 #define	ZONE_EVENT_CHANNEL	"com.sun:zones:status"
111 #define	ZONE_EVENT_STATUS_CLASS	"status"
112 #define	ZONE_EVENT_STATUS_SUBCLASS	"change"
113 
114 #define	ZONE_EVENT_UNINITIALIZED	"uninitialized"
115 #define	ZONE_EVENT_INITIALIZED		"initialized"
116 #define	ZONE_EVENT_READY		"ready"
117 #define	ZONE_EVENT_RUNNING		"running"
118 #define	ZONE_EVENT_SHUTTING_DOWN	"shutting_down"
119 
120 #define	ZONE_CB_NAME		"zonename"
121 #define	ZONE_CB_NEWSTATE	"newstate"
122 #define	ZONE_CB_OLDSTATE	"oldstate"
123 #define	ZONE_CB_TIMESTAMP	"when"
124 #define	ZONE_CB_ZONEID		"zoneid"
125 
126 /*
127  * Exit values that may be returned by scripts or programs invoked by various
128  * zone commands.
129  *
130  * These are defined as:
131  *
132  *	ZONE_SUBPROC_OK
133  *	===============
134  *	The subprocess completed successfully.
135  *
136  *	ZONE_SUBPROC_USAGE
137  *	==================
138  *	The subprocess failed with a usage message, or a usage message should
139  *	be output in its behalf.
140  *
141  *	ZONE_SUBPROC_NOTCOMPLETE
142  *	========================
143  *	The subprocess did not complete, but the actions performed by the
144  *	subprocess require no recovery actions by the user.
145  *
146  *	For example, if the subprocess were called by "zoneadm install," the
147  *	installation of the zone did not succeed but the user need not perform
148  *	a "zoneadm uninstall" before attempting another install.
149  *
150  *	ZONE_SUBPROC_FATAL
151  *	==================
152  *	The subprocess failed in a fatal manner, usually one that will require
153  *	some type of recovery action by the user.
154  *
155  *	For example, if the subprocess were called by "zoneadm install," the
156  *	installation of the zone did not succeed and the user will need to
157  *	perform a "zoneadm uninstall" before another install attempt is
158  *	possible.
159  *
160  *	The non-success exit values are large to avoid accidental collision
161  *	with values used internally by some commands (e.g. "Z_ERR" and
162  *	"Z_USAGE" as used by zoneadm.)
163  */
164 #define	ZONE_SUBPROC_OK			0
165 #define	ZONE_SUBPROC_USAGE		253
166 #define	ZONE_SUBPROC_NOTCOMPLETE	254
167 #define	ZONE_SUBPROC_FATAL		255
168 
169 #ifdef _SYSCALL32
170 typedef struct {
171 	caddr32_t zone_name;
172 	caddr32_t zone_root;
173 	caddr32_t zone_privs;
174 	size32_t zone_privssz;
175 	caddr32_t rctlbuf;
176 	size32_t rctlbufsz;
177 	caddr32_t extended_error;
178 	caddr32_t zfsbuf;
179 	size32_t  zfsbufsz;
180 	int match;			/* match level */
181 	uint32_t doi;			/* DOI for label */
182 	caddr32_t label;		/* label associated with zone */
183 	int flags;
184 } zone_def32;
185 #endif
186 typedef struct {
187 	const char *zone_name;
188 	const char *zone_root;
189 	const struct priv_set *zone_privs;
190 	size_t zone_privssz;
191 	const char *rctlbuf;
192 	size_t rctlbufsz;
193 	int *extended_error;
194 	const char *zfsbuf;
195 	size_t zfsbufsz;
196 	int match;			/* match level */
197 	uint32_t doi;			/* DOI for label */
198 	const bslabel_t *label;		/* label associated with zone */
199 	int flags;
200 } zone_def;
201 
202 /* extended error information */
203 #define	ZE_UNKNOWN	0	/* No extended error info */
204 #define	ZE_CHROOTED	1	/* tried to zone_create from chroot */
205 #define	ZE_AREMOUNTS	2	/* there are mounts within the zone */
206 #define	ZE_LABELINUSE	3	/* label is already in use by some other zone */
207 
208 /*
209  * zone_status values
210  *
211  * You must modify zone_status_names in mdb(1M)'s genunix module
212  * (genunix/zone.c) when you modify this enum.
213  */
214 typedef enum {
215 	ZONE_IS_UNINITIALIZED = 0,
216 	ZONE_IS_INITIALIZED,
217 	ZONE_IS_READY,
218 	ZONE_IS_BOOTING,
219 	ZONE_IS_RUNNING,
220 	ZONE_IS_SHUTTING_DOWN,
221 	ZONE_IS_EMPTY,
222 	ZONE_IS_DOWN,
223 	ZONE_IS_DYING,
224 	ZONE_IS_DEAD
225 } zone_status_t;
226 #define	ZONE_MIN_STATE		ZONE_IS_UNINITIALIZED
227 #define	ZONE_MAX_STATE		ZONE_IS_DEAD
228 
229 /*
230  * Valid commands which may be issued by zoneadm to zoneadmd.  The kernel also
231  * communicates with zoneadmd, but only uses Z_REBOOT and Z_HALT.
232  */
233 typedef enum zone_cmd {
234 	Z_READY, Z_BOOT, Z_FORCEBOOT, Z_REBOOT, Z_HALT, Z_NOTE_UNINSTALLING,
235 	Z_MOUNT, Z_FORCEMOUNT, Z_UNMOUNT, Z_SHUTDOWN
236 } zone_cmd_t;
237 
238 /*
239  * The structure of a request to zoneadmd.
240  */
241 typedef struct zone_cmd_arg {
242 	uint64_t	uniqid;		/* unique "generation number" */
243 	zone_cmd_t	cmd;		/* requested action */
244 	uint32_t	_pad;		/* need consistent 32/64 bit alignmt */
245 	char locale[MAXPATHLEN];	/* locale in which to render messages */
246 	char bootbuf[BOOTARGS_MAX];	/* arguments passed to zone_boot() */
247 } zone_cmd_arg_t;
248 
249 /*
250  * Structure of zoneadmd's response to a request.  A NULL return value means
251  * the caller should attempt to restart zoneadmd and retry.
252  */
253 typedef struct zone_cmd_rval {
254 	int rval;			/* return value of request */
255 	char errbuf[1];	/* variable-sized buffer containing error messages */
256 } zone_cmd_rval_t;
257 
258 /*
259  * The zone support infrastructure uses the zone name as a component
260  * of unix domain (AF_UNIX) sockets, which are limited to 108 characters
261  * in length, so ZONENAME_MAX is limited by that.
262  */
263 #define	ZONENAME_MAX		64
264 
265 #define	GLOBAL_ZONENAME		"global"
266 
267 /*
268  * Extended Regular expression (see regex(5)) which matches all valid zone
269  * names.
270  */
271 #define	ZONENAME_REGEXP		"[a-zA-Z0-9][-_.a-zA-Z0-9]{0,62}"
272 
273 /*
274  * Where the zones support infrastructure places temporary files.
275  */
276 #define	ZONES_TMPDIR		"/var/run/zones"
277 
278 /*
279  * The path to the door used by clients to communicate with zoneadmd.
280  */
281 #define	ZONE_DOOR_PATH		ZONES_TMPDIR "/%s.zoneadmd_door"
282 
283 
284 /* zone_flags */
285 /*
286  * Threads that read or write the following flag must hold zone_lock.
287  */
288 #define	ZF_REFCOUNTS_LOGGED	0x1	/* a thread logged the zone's refs */
289 
290 /*
291  * The following threads are set when the zone is created and never changed.
292  * Threads that test for these flags don't have to hold zone_lock.
293  */
294 #define	ZF_HASHED_LABEL		0x2	/* zone has a unique label */
295 #define	ZF_IS_SCRATCH		0x4	/* scratch zone */
296 #define	ZF_NET_EXCL		0x8	/* Zone has an exclusive IP stack */
297 
298 
299 /* zone_create flags */
300 #define	ZCF_NET_EXCL		0x1	/* Create a zone with exclusive IP */
301 
302 /* zone network properties */
303 #define	ZONE_NETWORK_ADDRESS	1
304 #define	ZONE_NETWORK_DEFROUTER	2
305 
306 #define	ZONE_NET_ADDRNAME	"address"
307 #define	ZONE_NET_RTRNAME	"route"
308 
309 typedef struct zone_net_data {
310 	int zn_type;
311 	int zn_len;
312 	datalink_id_t zn_linkid;
313 	uint8_t zn_val[1];
314 } zone_net_data_t;
315 
316 
317 #ifdef _KERNEL
318 
319 /*
320  * We need to protect the definition of 'list_t' from userland applications and
321  * libraries which may be defining ther own versions.
322  */
323 #include <sys/list.h>
324 #include <sys/cpuvar.h>
325 
326 #define	GLOBAL_ZONEUNIQID	0	/* uniqid of the global zone */
327 
328 struct pool;
329 struct brand;
330 
331 /*
332  * Each of these constants identifies a kernel subsystem that acquires and
333  * releases zone references.  Each subsystem that invokes
334  * zone_hold_ref() and zone_rele_ref() should specify the
335  * zone_ref_subsys_t constant associated with the subsystem.  Tracked holds
336  * help users and developers quickly identify subsystems that stall zone
337  * shutdowns indefinitely.
338  *
339  * NOTE: You must modify zone_ref_subsys_names in usr/src/uts/common/os/zone.c
340  * when you modify this enumeration.
341  */
342 typedef enum zone_ref_subsys {
343 	ZONE_REF_NFS,			/* NFS */
344 	ZONE_REF_NFSV4,			/* NFSv4 */
345 	ZONE_REF_SMBFS,			/* SMBFS */
346 	ZONE_REF_MNTFS,			/* MNTFS */
347 	ZONE_REF_LOFI,			/* LOFI devices */
348 	ZONE_REF_VFS,			/* VFS infrastructure */
349 	ZONE_REF_IPC,			/* IPC infrastructure */
350 	ZONE_REF_NUM_SUBSYS		/* This must be the last entry. */
351 } zone_ref_subsys_t;
352 
353 /*
354  * zone_ref represents a general-purpose references to a zone.  Each zone's
355  * references are linked into the zone's zone_t::zone_ref_list.  This allows
356  * debuggers to walk zones' references.
357  */
358 typedef struct zone_ref {
359 	struct zone	*zref_zone; /* the zone to which the reference refers */
360 	list_node_t	zref_linkage; /* linkage for zone_t::zone_ref_list */
361 } zone_ref_t;
362 
363 /*
364  * Structure to record list of ZFS datasets exported to a zone.
365  */
366 typedef struct zone_dataset {
367 	char		*zd_dataset;
368 	list_node_t	zd_linkage;
369 } zone_dataset_t;
370 
371 /*
372  * structure for zone kstats
373  */
374 typedef struct zone_kstat {
375 	kstat_named_t zk_zonename;
376 	kstat_named_t zk_usage;
377 	kstat_named_t zk_value;
378 } zone_kstat_t;
379 
380 struct cpucap;
381 
382 typedef struct {
383 	kstat_named_t	zm_zonename;	/* full name, kstat truncates name */
384 	kstat_named_t	zm_utime;
385 	kstat_named_t	zm_stime;
386 	kstat_named_t	zm_wtime;
387 	kstat_named_t	zm_avenrun1;
388 	kstat_named_t	zm_avenrun5;
389 	kstat_named_t	zm_avenrun15;
390 } zone_misc_kstat_t;
391 
392 typedef struct zone {
393 	/*
394 	 * zone_name is never modified once set.
395 	 */
396 	char		*zone_name;	/* zone's configuration name */
397 	/*
398 	 * zone_nodename and zone_domain are never freed once allocated.
399 	 */
400 	char		*zone_nodename;	/* utsname.nodename equivalent */
401 	char		*zone_domain;	/* srpc_domain equivalent */
402 	/*
403 	 * zone_hostid is used for per-zone hostid emulation.
404 	 * Currently it isn't modified after it's set (so no locks protect
405 	 * accesses), but that might have to change when we allow
406 	 * administrators to change running zones' properties.
407 	 *
408 	 * The global zone's zone_hostid must always be HW_INVALID_HOSTID so
409 	 * that zone_get_hostid() will function correctly.
410 	 */
411 	uint32_t	zone_hostid;	/* zone's hostid, HW_INVALID_HOSTID */
412 					/* if not emulated */
413 	/*
414 	 * zone_lock protects the following fields of a zone_t:
415 	 * 	zone_ref
416 	 * 	zone_cred_ref
417 	 * 	zone_subsys_ref
418 	 * 	zone_ref_list
419 	 * 	zone_ntasks
420 	 * 	zone_flags
421 	 * 	zone_zsd
422 	 *	zone_pfexecd
423 	 */
424 	kmutex_t	zone_lock;
425 	/*
426 	 * zone_linkage is the zone's linkage into the active or
427 	 * death-row list.  The field is protected by zonehash_lock.
428 	 */
429 	list_node_t	zone_linkage;
430 	zoneid_t	zone_id;	/* ID of zone */
431 	uint_t		zone_ref;	/* count of zone_hold()s on zone */
432 	uint_t		zone_cred_ref;	/* count of zone_hold_cred()s on zone */
433 	/*
434 	 * Fixed-sized array of subsystem-specific reference counts
435 	 * The sum of all of the counts must be less than or equal to zone_ref.
436 	 * The array is indexed by the counts' subsystems' zone_ref_subsys_t
437 	 * constants.
438 	 */
439 	uint_t		zone_subsys_ref[ZONE_REF_NUM_SUBSYS];
440 	list_t		zone_ref_list;	/* list of zone_ref_t structs */
441 	/*
442 	 * zone_rootvp and zone_rootpath can never be modified once set.
443 	 */
444 	struct vnode	*zone_rootvp;	/* zone's root vnode */
445 	char		*zone_rootpath;	/* Path to zone's root + '/' */
446 	ushort_t	zone_flags;	/* misc flags */
447 	zone_status_t	zone_status;	/* protected by zone_status_lock */
448 	uint_t		zone_ntasks;	/* number of tasks executing in zone */
449 	kmutex_t	zone_nlwps_lock; /* protects zone_nlwps, and *_nlwps */
450 					/* counters in projects and tasks */
451 					/* that are within the zone */
452 	rctl_qty_t	zone_nlwps;	/* number of lwps in zone */
453 	rctl_qty_t	zone_nlwps_ctl; /* protected by zone_rctls->rcs_lock */
454 	rctl_qty_t	zone_shmmax;	/* System V shared memory usage */
455 	ipc_rqty_t	zone_ipc;	/* System V IPC id resource usage */
456 
457 	uint_t		zone_rootpathlen; /* strlen(zone_rootpath) + 1 */
458 	uint32_t	zone_shares;	/* FSS shares allocated to zone */
459 	rctl_set_t	*zone_rctls;	/* zone-wide (zone.*) rctls */
460 	kmutex_t	zone_mem_lock;	/* protects zone_locked_mem and */
461 					/* kpd_locked_mem for all */
462 					/* projects in zone. */
463 					/* Also protects zone_max_swap */
464 					/* grab after p_lock, before rcs_lock */
465 	rctl_qty_t	zone_locked_mem;	/* bytes of locked memory in */
466 						/* zone */
467 	rctl_qty_t	zone_locked_mem_ctl;	/* Current locked memory */
468 						/* limit.  Protected by */
469 						/* zone_rctls->rcs_lock */
470 	rctl_qty_t	zone_max_swap; /* bytes of swap reserved by zone */
471 	rctl_qty_t	zone_max_swap_ctl;	/* current swap limit. */
472 						/* Protected by */
473 						/* zone_rctls->rcs_lock */
474 	kmutex_t	zone_rctl_lock;	/* protects zone_max_lofi */
475 	rctl_qty_t	zone_max_lofi; /* lofi devs for zone */
476 	rctl_qty_t	zone_max_lofi_ctl;	/* current lofi limit. */
477 						/* Protected by */
478 						/* zone_rctls->rcs_lock */
479 	list_t		zone_zsd;	/* list of Zone-Specific Data values */
480 	kcondvar_t	zone_cv;	/* used to signal state changes */
481 	struct proc	*zone_zsched;	/* Dummy kernel "zsched" process */
482 	pid_t		zone_proc_initpid; /* pid of "init" for this zone */
483 	char		*zone_initname;	/* fs path to 'init' */
484 	int		zone_boot_err;  /* for zone_boot() if boot fails */
485 	char		*zone_bootargs;	/* arguments passed via zone_boot() */
486 	uint64_t	zone_phys_mcap;	/* physical memory cap */
487 	/*
488 	 * zone_kthreads is protected by zone_status_lock.
489 	 */
490 	kthread_t	*zone_kthreads;	/* kernel threads in zone */
491 	struct priv_set	*zone_privset;	/* limit set for zone */
492 	/*
493 	 * zone_vfslist is protected by vfs_list_lock().
494 	 */
495 	struct vfs	*zone_vfslist;	/* list of FS's mounted in zone */
496 	uint64_t	zone_uniqid;	/* unique zone generation number */
497 	struct cred	*zone_kcred;	/* kcred-like, zone-limited cred */
498 	/*
499 	 * zone_pool is protected by pool_lock().
500 	 */
501 	struct pool	*zone_pool;	/* pool the zone is bound to */
502 	hrtime_t	zone_pool_mod;	/* last pool bind modification time */
503 	/* zone_psetid is protected by cpu_lock */
504 	psetid_t	zone_psetid;	/* pset the zone is bound to */
505 
506 	time_t		zone_boot_time;	/* Similar to boot_time */
507 
508 	/*
509 	 * The following two can be read without holding any locks.  They are
510 	 * updated under cpu_lock.
511 	 */
512 	int		zone_ncpus;  /* zone's idea of ncpus */
513 	int		zone_ncpus_online; /* zone's idea of ncpus_online */
514 	/*
515 	 * List of ZFS datasets exported to this zone.
516 	 */
517 	list_t		zone_datasets;	/* list of datasets */
518 
519 	ts_label_t	*zone_slabel;	/* zone sensitivity label */
520 	int		zone_match;	/* require label match for packets */
521 	tsol_mlp_list_t zone_mlps;	/* MLPs on zone-private addresses */
522 
523 	boolean_t	zone_restart_init;	/* Restart init if it dies? */
524 	struct brand	*zone_brand;		/* zone's brand */
525 	void 		*zone_brand_data;	/* store brand specific data */
526 	id_t		zone_defaultcid;	/* dflt scheduling class id */
527 	kstat_t		*zone_swapresv_kstat;
528 	kstat_t		*zone_lockedmem_kstat;
529 	/*
530 	 * zone_dl_list is protected by zone_lock
531 	 */
532 	list_t		zone_dl_list;
533 	netstack_t	*zone_netstack;
534 	struct cpucap	*zone_cpucap;	/* CPU caps data */
535 	/*
536 	 * Solaris Auditing per-zone audit context
537 	 */
538 	struct au_kcontext	*zone_audit_kctxt;
539 	/*
540 	 * For private use by mntfs.
541 	 */
542 	struct mntelem	*zone_mntfs_db;
543 	krwlock_t	zone_mntfs_db_lock;
544 
545 	struct klpd_reg		*zone_pfexecd;
546 
547 	char		*zone_fs_allowed;
548 	rctl_qty_t	zone_nprocs;	/* number of processes in the zone */
549 	rctl_qty_t	zone_nprocs_ctl;	/* current limit protected by */
550 						/* zone_rctls->rcs_lock */
551 	kstat_t		*zone_nprocs_kstat;
552 
553 	/*
554 	 * Misc. kstats and counters for zone cpu-usage aggregation.
555 	 * The zone_Xtime values are the sum of the micro-state accounting
556 	 * values for all threads that are running or have run in the zone.
557 	 * This is tracked in msacct.c as threads change state.
558 	 * The zone_stime is the sum of the LMS_SYSTEM times.
559 	 * The zone_utime is the sum of the LMS_USER times.
560 	 * The zone_wtime is the sum of the LMS_WAIT_CPU times.
561 	 * As with per-thread micro-state accounting values, these values are
562 	 * not scaled to nanosecs.  The scaling is done by the
563 	 * zone_misc_kstat_update function when kstats are requested.
564 	 */
565 	kmutex_t	zone_misc_lock;		/* protects misc statistics */
566 	kstat_t		*zone_misc_ksp;
567 	zone_misc_kstat_t *zone_misc_stats;
568 	uint64_t	zone_stime;		/* total system time */
569 	uint64_t	zone_utime;		/* total user time */
570 	uint64_t	zone_wtime;		/* total time waiting in runq */
571 
572 	struct loadavg_s zone_loadavg;		/* loadavg for this zone */
573 	uint64_t	zone_hp_avenrun[3];	/* high-precision avenrun */
574 	int		zone_avenrun[3];	/* FSCALED avg. run queue len */
575 
576 	/*
577 	 * DTrace-private per-zone state
578 	 */
579 	int		zone_dtrace_getf;	/* # of unprivileged getf()s */
580 } zone_t;
581 
582 /*
583  * Special value of zone_psetid to indicate that pools are disabled.
584  */
585 #define	ZONE_PS_INVAL	PS_MYID
586 
587 
588 extern zone_t zone0;
589 extern zone_t *global_zone;
590 extern uint_t maxzones;
591 extern rctl_hndl_t rc_zone_nlwps;
592 extern rctl_hndl_t rc_zone_nprocs;
593 
594 extern long zone(int, void *, void *, void *, void *);
595 extern void zone_zsd_init(void);
596 extern void zone_init(void);
597 extern void zone_hold(zone_t *);
598 extern void zone_rele(zone_t *);
599 extern void zone_init_ref(zone_ref_t *);
600 extern void zone_hold_ref(zone_t *, zone_ref_t *, zone_ref_subsys_t);
601 extern void zone_rele_ref(zone_ref_t *, zone_ref_subsys_t);
602 extern void zone_cred_hold(zone_t *);
603 extern void zone_cred_rele(zone_t *);
604 extern void zone_task_hold(zone_t *);
605 extern void zone_task_rele(zone_t *);
606 extern zone_t *zone_find_by_id(zoneid_t);
607 extern zone_t *zone_find_by_label(const ts_label_t *);
608 extern zone_t *zone_find_by_name(char *);
609 extern zone_t *zone_find_by_any_path(const char *, boolean_t);
610 extern zone_t *zone_find_by_path(const char *);
611 extern zoneid_t getzoneid(void);
612 extern zone_t *zone_find_by_id_nolock(zoneid_t);
613 extern int zone_datalink_walk(zoneid_t, int (*)(datalink_id_t, void *), void *);
614 extern int zone_check_datalink(zoneid_t *, datalink_id_t);
615 extern void zone_loadavg_update();
616 
617 /*
618  * Zone-specific data (ZSD) APIs
619  */
620 /*
621  * The following is what code should be initializing its zone_key_t to if it
622  * calls zone_getspecific() without necessarily knowing that zone_key_create()
623  * has been called on the key.
624  */
625 #define	ZONE_KEY_UNINITIALIZED	0
626 
627 typedef uint_t zone_key_t;
628 
629 extern void	zone_key_create(zone_key_t *, void *(*)(zoneid_t),
630     void (*)(zoneid_t, void *), void (*)(zoneid_t, void *));
631 extern int 	zone_key_delete(zone_key_t);
632 extern void	*zone_getspecific(zone_key_t, zone_t *);
633 extern int	zone_setspecific(zone_key_t, zone_t *, const void *);
634 
635 /*
636  * The definition of a zsd_entry is truly private to zone.c and is only
637  * placed here so it can be shared with mdb.
638  *
639  * State maintained for each zone times each registered key, which tracks
640  * the state of the create, shutdown and destroy callbacks.
641  *
642  * zsd_flags is used to keep track of pending actions to avoid holding locks
643  * when calling the create/shutdown/destroy callbacks, since doing so
644  * could lead to deadlocks.
645  */
646 struct zsd_entry {
647 	zone_key_t		zsd_key;	/* Key used to lookup value */
648 	void			*zsd_data;	/* Caller-managed value */
649 	/*
650 	 * Callbacks to be executed when a zone is created, shutdown, and
651 	 * destroyed, respectively.
652 	 */
653 	void			*(*zsd_create)(zoneid_t);
654 	void			(*zsd_shutdown)(zoneid_t, void *);
655 	void			(*zsd_destroy)(zoneid_t, void *);
656 	list_node_t		zsd_linkage;
657 	uint16_t 		zsd_flags;	/* See below */
658 	kcondvar_t		zsd_cv;
659 };
660 
661 /*
662  * zsd_flags
663  */
664 #define	ZSD_CREATE_NEEDED	0x0001
665 #define	ZSD_CREATE_INPROGRESS	0x0002
666 #define	ZSD_CREATE_COMPLETED	0x0004
667 #define	ZSD_SHUTDOWN_NEEDED	0x0010
668 #define	ZSD_SHUTDOWN_INPROGRESS	0x0020
669 #define	ZSD_SHUTDOWN_COMPLETED	0x0040
670 #define	ZSD_DESTROY_NEEDED	0x0100
671 #define	ZSD_DESTROY_INPROGRESS	0x0200
672 #define	ZSD_DESTROY_COMPLETED	0x0400
673 
674 #define	ZSD_CREATE_ALL	\
675 	(ZSD_CREATE_NEEDED|ZSD_CREATE_INPROGRESS|ZSD_CREATE_COMPLETED)
676 #define	ZSD_SHUTDOWN_ALL	\
677 	(ZSD_SHUTDOWN_NEEDED|ZSD_SHUTDOWN_INPROGRESS|ZSD_SHUTDOWN_COMPLETED)
678 #define	ZSD_DESTROY_ALL	\
679 	(ZSD_DESTROY_NEEDED|ZSD_DESTROY_INPROGRESS|ZSD_DESTROY_COMPLETED)
680 
681 #define	ZSD_ALL_INPROGRESS \
682 	(ZSD_CREATE_INPROGRESS|ZSD_SHUTDOWN_INPROGRESS|ZSD_DESTROY_INPROGRESS)
683 
684 /*
685  * Macros to help with zone visibility restrictions.
686  */
687 
688 /*
689  * Is process in the global zone?
690  */
691 #define	INGLOBALZONE(p) \
692 	((p)->p_zone == global_zone)
693 
694 /*
695  * Can process view objects in given zone?
696  */
697 #define	HASZONEACCESS(p, zoneid) \
698 	((p)->p_zone->zone_id == (zoneid) || INGLOBALZONE(p))
699 
700 /*
701  * Convenience macro to see if a resolved path is visible from within a
702  * given zone.
703  *
704  * The basic idea is that the first (zone_rootpathlen - 1) bytes of the
705  * two strings must be equal.  Since the rootpathlen has a trailing '/',
706  * we want to skip everything in the path up to (but not including) the
707  * trailing '/'.
708  */
709 #define	ZONE_PATH_VISIBLE(path, zone) \
710 	(strncmp((path), (zone)->zone_rootpath,		\
711 	    (zone)->zone_rootpathlen - 1) == 0)
712 
713 /*
714  * Convenience macro to go from the global view of a path to that seen
715  * from within said zone.  It is the responsibility of the caller to
716  * ensure that the path is a resolved one (ie, no '..'s or '.'s), and is
717  * in fact visible from within the zone.
718  */
719 #define	ZONE_PATH_TRANSLATE(path, zone)	\
720 	(ASSERT(ZONE_PATH_VISIBLE(path, zone)),	\
721 	(path) + (zone)->zone_rootpathlen - 2)
722 
723 /*
724  * Special processes visible in all zones.
725  */
726 #define	ZONE_SPECIALPID(x)	 ((x) == 0 || (x) == 1)
727 
728 /*
729  * Zone-safe version of thread_create() to be used when the caller wants to
730  * create a kernel thread to run within the current zone's context.
731  */
732 extern kthread_t *zthread_create(caddr_t, size_t, void (*)(), void *, size_t,
733     pri_t);
734 extern void zthread_exit(void);
735 
736 /*
737  * Functions for an external observer to register interest in a zone's status
738  * change.  Observers will be woken up when the zone status equals the status
739  * argument passed in (in the case of zone_status_timedwait, the function may
740  * also return because of a timeout; zone_status_wait_sig may return early due
741  * to a signal being delivered; zone_status_timedwait_sig may return for any of
742  * the above reasons).
743  *
744  * Otherwise these behave identically to cv_timedwait(), cv_wait(), and
745  * cv_wait_sig() respectively.
746  */
747 extern clock_t zone_status_timedwait(zone_t *, clock_t, zone_status_t);
748 extern clock_t zone_status_timedwait_sig(zone_t *, clock_t, zone_status_t);
749 extern void zone_status_wait(zone_t *, zone_status_t);
750 extern int zone_status_wait_sig(zone_t *, zone_status_t);
751 
752 /*
753  * Get the status  of the zone (at the time it was called).  The state may
754  * have progressed by the time it is returned.
755  */
756 extern zone_status_t zone_status_get(zone_t *);
757 
758 /*
759  * Safely get the hostid of the specified zone (defaults to machine's hostid
760  * if the specified zone doesn't emulate a hostid).  Passing NULL retrieves
761  * the global zone's (i.e., physical system's) hostid.
762  */
763 extern uint32_t zone_get_hostid(zone_t *);
764 
765 /*
766  * Get the "kcred" credentials corresponding to the given zone.
767  */
768 extern struct cred *zone_get_kcred(zoneid_t);
769 
770 /*
771  * Get/set the pool the zone is currently bound to.
772  */
773 extern struct pool *zone_pool_get(zone_t *);
774 extern void zone_pool_set(zone_t *, struct pool *);
775 
776 /*
777  * Get/set the pset the zone is currently using.
778  */
779 extern psetid_t zone_pset_get(zone_t *);
780 extern void zone_pset_set(zone_t *, psetid_t);
781 
782 /*
783  * Get the number of cpus/online-cpus visible from the given zone.
784  */
785 extern int zone_ncpus_get(zone_t *);
786 extern int zone_ncpus_online_get(zone_t *);
787 
788 /*
789  * Returns true if the named pool/dataset is visible in the current zone.
790  */
791 extern int zone_dataset_visible(const char *, int *);
792 
793 /*
794  * zone version of kadmin()
795  */
796 extern int zone_kadmin(int, int, const char *, cred_t *);
797 extern void zone_shutdown_global(void);
798 
799 extern void mount_in_progress(void);
800 extern void mount_completed(void);
801 
802 extern int zone_walk(int (*)(zone_t *, void *), void *);
803 
804 extern rctl_hndl_t rc_zone_locked_mem;
805 extern rctl_hndl_t rc_zone_max_swap;
806 extern rctl_hndl_t rc_zone_max_lofi;
807 
808 #endif	/* _KERNEL */
809 
810 #ifdef	__cplusplus
811 }
812 #endif
813 
814 #endif	/* _SYS_ZONE_H */
815