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 2015 Joyent, Inc. All rights reserved.
24 * Copyright 2014 Nexenta Systems, Inc. All rights reserved.
25 * Copyright 2014 Igor Kozhukhov <ikozhukhov@gmail.com>.
26 * Copyright 2016, Joyent, Inc.
27 */
28
29 #ifndef _SYS_ZONE_H
30 #define _SYS_ZONE_H
31
32 #include <sys/types.h>
33 #include <sys/mutex.h>
34 #include <sys/param.h>
35 #include <sys/rctl.h>
36 #include <sys/ipc_rctl.h>
37 #include <sys/pset.h>
38 #include <sys/tsol/label.h>
39 #include <sys/cred.h>
40 #include <sys/netstack.h>
41 #include <sys/uadmin.h>
42 #include <sys/ksynch.h>
43 #include <sys/socket_impl.h>
44 #include <netinet/in.h>
45
46 #ifdef __cplusplus
47 extern "C" {
48 #endif
49
50 /*
51 * NOTE
52 *
53 * The contents of this file are private to the implementation of
54 * Solaris and are subject to change at any time without notice.
55 * Applications and drivers using these interfaces may fail to
56 * run on future releases.
57 */
58
59 /* Available both in kernel and for user space */
60
61 /* zone id restrictions and special ids */
62 #define MAX_ZONEID 9999
63 #define MIN_USERZONEID 1 /* lowest user-creatable zone ID */
64 #define MIN_ZONEID 0 /* minimum zone ID on system */
65 #define GLOBAL_ZONEID 0
66 #define ZONEID_WIDTH 4 /* for printf */
67
68 /*
69 * Special zoneid_t token to refer to all zones.
70 */
71 #define ALL_ZONES (-1)
72
73 /* system call subcodes */
74 #define ZONE_CREATE 0
75 #define ZONE_DESTROY 1
76 #define ZONE_GETATTR 2
77 #define ZONE_ENTER 3
78 #define ZONE_LIST 4
79 #define ZONE_SHUTDOWN 5
80 #define ZONE_LOOKUP 6
81 #define ZONE_BOOT 7
82 #define ZONE_VERSION 8
83 #define ZONE_SETATTR 9
84 #define ZONE_ADD_DATALINK 10
85 #define ZONE_DEL_DATALINK 11
86 #define ZONE_CHECK_DATALINK 12
87 #define ZONE_LIST_DATALINK 13
88
89 /* zone attributes */
90 #define ZONE_ATTR_ROOT 1
91 #define ZONE_ATTR_NAME 2
92 #define ZONE_ATTR_STATUS 3
93 #define ZONE_ATTR_PRIVSET 4
94 #define ZONE_ATTR_UNIQID 5
95 #define ZONE_ATTR_POOLID 6
96 #define ZONE_ATTR_INITPID 7
97 #define ZONE_ATTR_SLBL 8
98 #define ZONE_ATTR_INITNAME 9
99 #define ZONE_ATTR_BOOTARGS 10
100 #define ZONE_ATTR_BRAND 11
101 #define ZONE_ATTR_PMCAP_NOVER 12
102 #define ZONE_ATTR_SCHED_CLASS 13
103 #define ZONE_ATTR_FLAGS 14
104 #define ZONE_ATTR_HOSTID 15
105 #define ZONE_ATTR_FS_ALLOWED 16
106 #define ZONE_ATTR_NETWORK 17
107 #define ZONE_ATTR_DID 18
108 #define ZONE_ATTR_PMCAP_PAGEOUT 19
109 #define ZONE_ATTR_INITNORESTART 20
110 #define ZONE_ATTR_PG_FLT_DELAY 21
111 #define ZONE_ATTR_RSS 22
112 #define ZONE_ATTR_APP_SVC_CT 23
113 #define ZONE_ATTR_SCHED_FIXEDHI 24
114
115 /* Start of the brand-specific attribute namespace */
116 #define ZONE_ATTR_BRAND_ATTRS 32768
117
118 #define ZONE_FS_ALLOWED_MAX 1024
119
120 #define ZONE_EVENT_CHANNEL "com.sun:zones:status"
121 #define ZONE_EVENT_STATUS_CLASS "status"
122 #define ZONE_EVENT_STATUS_SUBCLASS "change"
123
124 #define ZONE_EVENT_UNINITIALIZED "uninitialized"
125 #define ZONE_EVENT_INITIALIZED "initialized"
126 #define ZONE_EVENT_READY "ready"
127 #define ZONE_EVENT_RUNNING "running"
128 #define ZONE_EVENT_SHUTTING_DOWN "shutting_down"
129
130 #define ZONE_CB_NAME "zonename"
131 #define ZONE_CB_NEWSTATE "newstate"
132 #define ZONE_CB_OLDSTATE "oldstate"
133 #define ZONE_CB_TIMESTAMP "when"
134 #define ZONE_CB_ZONEID "zoneid"
135
136 /*
137 * Exit values that may be returned by scripts or programs invoked by various
138 * zone commands.
139 *
140 * These are defined as:
141 *
142 * ZONE_SUBPROC_OK
143 * ===============
144 * The subprocess completed successfully.
145 *
146 * ZONE_SUBPROC_USAGE
147 * ==================
148 * The subprocess failed with a usage message, or a usage message should
149 * be output in its behalf.
150 *
151 * ZONE_SUBPROC_NOTCOMPLETE
152 * ========================
153 * The subprocess did not complete, but the actions performed by the
154 * subprocess require no recovery actions by the user.
155 *
156 * For example, if the subprocess were called by "zoneadm install," the
157 * installation of the zone did not succeed but the user need not perform
158 * a "zoneadm uninstall" before attempting another install.
159 *
160 * ZONE_SUBPROC_FATAL
161 * ==================
162 * The subprocess failed in a fatal manner, usually one that will require
163 * some type of recovery action by the user.
164 *
165 * For example, if the subprocess were called by "zoneadm install," the
166 * installation of the zone did not succeed and the user will need to
167 * perform a "zoneadm uninstall" before another install attempt is
168 * possible.
169 *
170 * The non-success exit values are large to avoid accidental collision
171 * with values used internally by some commands (e.g. "Z_ERR" and
172 * "Z_USAGE" as used by zoneadm.)
173 */
174 #define ZONE_SUBPROC_OK 0
175 #define ZONE_SUBPROC_USAGE 253
176 #define ZONE_SUBPROC_NOTCOMPLETE 254
177 #define ZONE_SUBPROC_FATAL 255
178
179 #ifdef _SYSCALL32
180 typedef struct {
181 caddr32_t zone_name;
182 caddr32_t zone_root;
183 caddr32_t zone_privs;
184 size32_t zone_privssz;
185 caddr32_t rctlbuf;
186 size32_t rctlbufsz;
187 caddr32_t extended_error;
188 caddr32_t zfsbuf;
189 size32_t zfsbufsz;
190 int match; /* match level */
191 uint32_t doi; /* DOI for label */
192 caddr32_t label; /* label associated with zone */
193 int flags;
194 } zone_def32;
195 #endif
196 typedef struct {
197 const char *zone_name;
198 const char *zone_root;
199 const struct priv_set *zone_privs;
200 size_t zone_privssz;
201 const char *rctlbuf;
202 size_t rctlbufsz;
203 int *extended_error;
204 const char *zfsbuf;
205 size_t zfsbufsz;
206 int match; /* match level */
207 uint32_t doi; /* DOI for label */
208 const bslabel_t *label; /* label associated with zone */
209 int flags;
210 } zone_def;
211
212 /* extended error information */
213 #define ZE_UNKNOWN 0 /* No extended error info */
214 #define ZE_CHROOTED 1 /* tried to zone_create from chroot */
215 #define ZE_AREMOUNTS 2 /* there are mounts within the zone */
216 #define ZE_LABELINUSE 3 /* label is already in use by some other zone */
217
218 /*
219 * zone_status values
220 *
221 * You must modify zone_status_names in mdb(1M)'s genunix module
222 * (genunix/zone.c) when you modify this enum.
223 */
224 typedef enum {
225 ZONE_IS_UNINITIALIZED = 0,
226 ZONE_IS_INITIALIZED,
227 ZONE_IS_READY,
228 ZONE_IS_BOOTING,
229 ZONE_IS_RUNNING,
230 ZONE_IS_SHUTTING_DOWN,
231 ZONE_IS_EMPTY,
232 ZONE_IS_DOWN,
233 ZONE_IS_DYING,
234 ZONE_IS_DEAD
235 } zone_status_t;
236 #define ZONE_MIN_STATE ZONE_IS_UNINITIALIZED
237 #define ZONE_MAX_STATE ZONE_IS_DEAD
238
239 /*
240 * Valid commands which may be issued by zoneadm to zoneadmd. The kernel also
241 * communicates with zoneadmd, but only uses Z_REBOOT and Z_HALT.
242 */
243 typedef enum zone_cmd {
244 Z_READY, Z_BOOT, Z_FORCEBOOT, Z_REBOOT, Z_HALT, Z_NOTE_UNINSTALLING,
245 Z_MOUNT, Z_FORCEMOUNT, Z_UNMOUNT, Z_SHUTDOWN
246 } zone_cmd_t;
247
248 /*
249 * The structure of a request to zoneadmd.
250 */
251 typedef struct zone_cmd_arg {
252 uint64_t uniqid; /* unique "generation number" */
253 zone_cmd_t cmd; /* requested action */
254 int status; /* init status on shutdown */
255 uint32_t debug; /* enable brand hook debug */
256 char locale[MAXPATHLEN]; /* locale in which to render messages */
257 char bootbuf[BOOTARGS_MAX]; /* arguments passed to zone_boot() */
258 /* Needed for 32/64 zoneadm -> zoneadmd door arg size check. */
259 int pad;
260 } zone_cmd_arg_t;
261
262 /*
263 * Structure of zoneadmd's response to a request. A NULL return value means
264 * the caller should attempt to restart zoneadmd and retry.
265 */
266 typedef struct zone_cmd_rval {
267 int rval; /* return value of request */
268 char errbuf[1]; /* variable-sized buffer containing error messages */
269 } zone_cmd_rval_t;
270
271 /*
272 * The zone support infrastructure uses the zone name as a component
273 * of unix domain (AF_UNIX) sockets, which are limited to 108 characters
274 * in length, so ZONENAME_MAX is limited by that.
275 */
276 #define ZONENAME_MAX 64
277
278 #define GLOBAL_ZONENAME "global"
279
280 /*
281 * Extended Regular expression (see regex(5)) which matches all valid zone
282 * names.
283 */
284 #define ZONENAME_REGEXP "[a-zA-Z0-9][-_.a-zA-Z0-9]{0,62}"
285
286 /*
287 * Where the zones support infrastructure places temporary files.
288 */
289 #define ZONES_TMPDIR "/var/run/zones"
290
291 /*
292 * The path to the door used by clients to communicate with zoneadmd.
293 */
294 #define ZONE_DOOR_PATH ZONES_TMPDIR "/%s.zoneadmd_door"
295
296
297 /* zone_flags */
298 /*
299 * Threads that read or write the following flag must hold zone_lock.
300 */
301 #define ZF_REFCOUNTS_LOGGED 0x1 /* a thread logged the zone's refs */
302
303 /*
304 * The following threads are set when the zone is created and never changed.
305 * Threads that test for these flags don't have to hold zone_lock.
306 */
307 #define ZF_HASHED_LABEL 0x2 /* zone has a unique label */
308 #define ZF_IS_SCRATCH 0x4 /* scratch zone */
309 #define ZF_NET_EXCL 0x8 /* Zone has an exclusive IP stack */
310
311
312 /* zone_create flags */
313 #define ZCF_NET_EXCL 0x1 /* Create a zone with exclusive IP */
314
315 /* zone network properties */
316 #define ZONE_NETWORK_ADDRESS 1
317 #define ZONE_NETWORK_DEFROUTER 2
318
319 #define ZONE_NET_ADDRNAME "address"
320 #define ZONE_NET_RTRNAME "route"
321
322 typedef struct zone_net_data {
323 int zn_type;
324 int zn_len;
325 datalink_id_t zn_linkid;
326 uint8_t zn_val[1];
327 } zone_net_data_t;
328
329
330 #ifdef _KERNEL
331
332 /*
333 * We need to protect the definition of 'list_t' from userland applications and
334 * libraries which may be defining ther own versions.
335 */
336 #include <sys/list.h>
337 #include <sys/loadavg.h>
338
339 #define GLOBAL_ZONEUNIQID 0 /* uniqid of the global zone */
340
341 struct pool;
342 struct brand;
343
344 /*
345 * Each of these constants identifies a kernel subsystem that acquires and
346 * releases zone references. Each subsystem that invokes
347 * zone_hold_ref() and zone_rele_ref() should specify the
348 * zone_ref_subsys_t constant associated with the subsystem. Tracked holds
349 * help users and developers quickly identify subsystems that stall zone
350 * shutdowns indefinitely.
351 *
352 * NOTE: You must modify zone_ref_subsys_names in usr/src/uts/common/os/zone.c
353 * when you modify this enumeration.
354 */
355 typedef enum zone_ref_subsys {
356 ZONE_REF_NFS, /* NFS */
357 ZONE_REF_NFSV4, /* NFSv4 */
358 ZONE_REF_SMBFS, /* SMBFS */
359 ZONE_REF_MNTFS, /* MNTFS */
360 ZONE_REF_LOFI, /* LOFI devices */
361 ZONE_REF_VFS, /* VFS infrastructure */
362 ZONE_REF_IPC, /* IPC infrastructure */
363 ZONE_REF_NUM_SUBSYS /* This must be the last entry. */
364 } zone_ref_subsys_t;
365
366 /*
367 * zone_ref represents a general-purpose references to a zone. Each zone's
368 * references are linked into the zone's zone_t::zone_ref_list. This allows
369 * debuggers to walk zones' references.
370 */
371 typedef struct zone_ref {
372 struct zone *zref_zone; /* the zone to which the reference refers */
373 list_node_t zref_linkage; /* linkage for zone_t::zone_ref_list */
374 } zone_ref_t;
375
376 /*
377 * Structure to record list of ZFS datasets exported to a zone.
378 */
379 typedef struct zone_dataset {
380 char *zd_dataset;
381 list_node_t zd_linkage;
382 } zone_dataset_t;
383
384 /*
385 * structure for zone kstats
386 */
387 typedef struct zone_kstat {
388 kstat_named_t zk_zonename;
389 kstat_named_t zk_usage;
390 kstat_named_t zk_value;
391 } zone_kstat_t;
392
393 struct cpucap;
394
395 typedef struct {
396 hrtime_t cycle_start;
397 uint_t cycle_cnt;
398 hrtime_t zone_avg_cnt;
399 } sys_zio_cntr_t;
400
401 typedef struct {
402 kstat_named_t zv_zonename;
403 kstat_named_t zv_nread;
404 kstat_named_t zv_reads;
405 kstat_named_t zv_rtime;
406 kstat_named_t zv_rlentime;
407 kstat_named_t zv_rcnt;
408 kstat_named_t zv_nwritten;
409 kstat_named_t zv_writes;
410 kstat_named_t zv_wtime;
411 kstat_named_t zv_wlentime;
412 kstat_named_t zv_wcnt;
413 kstat_named_t zv_10ms_ops;
414 kstat_named_t zv_100ms_ops;
415 kstat_named_t zv_1s_ops;
416 kstat_named_t zv_10s_ops;
417 kstat_named_t zv_delay_cnt;
418 kstat_named_t zv_delay_time;
419 } zone_vfs_kstat_t;
420
421 typedef struct {
422 kstat_named_t zz_zonename;
423 kstat_named_t zz_nread;
424 kstat_named_t zz_reads;
425 kstat_named_t zz_rtime;
426 kstat_named_t zz_rlentime;
427 kstat_named_t zz_nwritten;
428 kstat_named_t zz_writes;
429 kstat_named_t zz_waittime;
430 } zone_zfs_kstat_t;
431
432 typedef struct {
433 kstat_named_t zm_zonename;
434 kstat_named_t zm_rss;
435 kstat_named_t zm_phys_cap;
436 kstat_named_t zm_swap;
437 kstat_named_t zm_swap_cap;
438 kstat_named_t zm_nover;
439 kstat_named_t zm_pagedout;
440 kstat_named_t zm_pgpgin;
441 kstat_named_t zm_anonpgin;
442 kstat_named_t zm_execpgin;
443 kstat_named_t zm_fspgin;
444 kstat_named_t zm_anon_alloc_fail;
445 kstat_named_t zm_pf_throttle;
446 kstat_named_t zm_pf_throttle_usec;
447 } zone_mcap_kstat_t;
448
449 typedef struct {
450 kstat_named_t zm_zonename; /* full name, kstat truncates name */
451 kstat_named_t zm_utime;
452 kstat_named_t zm_stime;
453 kstat_named_t zm_wtime;
454 kstat_named_t zm_avenrun1;
455 kstat_named_t zm_avenrun5;
456 kstat_named_t zm_avenrun15;
457 kstat_named_t zm_ffcap;
458 kstat_named_t zm_ffnoproc;
459 kstat_named_t zm_ffnomem;
460 kstat_named_t zm_ffmisc;
461 kstat_named_t zm_nested_intp;
462 kstat_named_t zm_init_pid;
463 kstat_named_t zm_boot_time;
464 } zone_misc_kstat_t;
465
466 typedef struct zone {
467 /*
468 * zone_name is never modified once set.
469 */
470 char *zone_name; /* zone's configuration name */
471 /*
472 * zone_nodename and zone_domain are never freed once allocated.
473 */
474 char *zone_nodename; /* utsname.nodename equivalent */
475 char *zone_domain; /* srpc_domain equivalent */
476 /*
477 * zone_hostid is used for per-zone hostid emulation.
478 * Currently it isn't modified after it's set (so no locks protect
479 * accesses), but that might have to change when we allow
480 * administrators to change running zones' properties.
481 *
482 * The global zone's zone_hostid must always be HW_INVALID_HOSTID so
483 * that zone_get_hostid() will function correctly.
484 */
485 uint32_t zone_hostid; /* zone's hostid, HW_INVALID_HOSTID */
486 /* if not emulated */
487 /*
488 * zone_lock protects the following fields of a zone_t:
489 * zone_ref
490 * zone_cred_ref
491 * zone_subsys_ref
492 * zone_ref_list
493 * zone_ntasks
494 * zone_flags
495 * zone_zsd
496 * zone_pfexecd
497 */
498 kmutex_t zone_lock;
499 /*
500 * zone_linkage is the zone's linkage into the active or
501 * death-row list. The field is protected by zonehash_lock.
502 */
503 list_node_t zone_linkage;
504 zoneid_t zone_id; /* ID of zone */
505 uint_t zone_ref; /* count of zone_hold()s on zone */
506 uint_t zone_cred_ref; /* count of zone_hold_cred()s on zone */
507 /*
508 * Fixed-sized array of subsystem-specific reference counts
509 * The sum of all of the counts must be less than or equal to zone_ref.
510 * The array is indexed by the counts' subsystems' zone_ref_subsys_t
511 * constants.
512 */
513 uint_t zone_subsys_ref[ZONE_REF_NUM_SUBSYS];
514 list_t zone_ref_list; /* list of zone_ref_t structs */
515 /*
516 * zone_rootvp and zone_rootpath can never be modified once set.
517 */
518 struct vnode *zone_rootvp; /* zone's root vnode */
519 char *zone_rootpath; /* Path to zone's root + '/' */
520 ushort_t zone_flags; /* misc flags */
521 zone_status_t zone_status; /* protected by zone_status_lock */
522 uint_t zone_ntasks; /* number of tasks executing in zone */
523 kmutex_t zone_nlwps_lock; /* protects zone_nlwps, and *_nlwps */
524 /* counters in projects and tasks */
525 /* that are within the zone */
526 rctl_qty_t zone_nlwps; /* number of lwps in zone */
527 rctl_qty_t zone_nlwps_ctl; /* protected by zone_rctls->rcs_lock */
528 rctl_qty_t zone_shmmax; /* System V shared memory usage */
529 ipc_rqty_t zone_ipc; /* System V IPC id resource usage */
530
531 uint_t zone_rootpathlen; /* strlen(zone_rootpath) + 1 */
532 uint32_t zone_shares; /* FSS shares allocated to zone */
533 rctl_set_t *zone_rctls; /* zone-wide (zone.*) rctls */
534 kmutex_t zone_mem_lock; /* protects zone_locked_mem and */
535 /* kpd_locked_mem for all */
536 /* projects in zone. */
537 /* Also protects zone_max_swap */
538 /* grab after p_lock, before rcs_lock */
539 rctl_qty_t zone_locked_mem; /* bytes of locked memory in */
540 /* zone */
541 rctl_qty_t zone_locked_mem_ctl; /* Current locked memory */
542 /* limit. Protected by */
543 /* zone_rctls->rcs_lock */
544 rctl_qty_t zone_max_swap; /* bytes of swap reserved by zone */
545 rctl_qty_t zone_max_swap_ctl; /* current swap limit. */
546 /* Protected by */
547 /* zone_rctls->rcs_lock */
548 kmutex_t zone_rctl_lock; /* protects zone_max_lofi */
549 rctl_qty_t zone_max_lofi; /* lofi devs for zone */
550 rctl_qty_t zone_max_lofi_ctl; /* current lofi limit. */
551 /* Protected by */
552 /* zone_rctls->rcs_lock */
553 list_t zone_zsd; /* list of Zone-Specific Data values */
554 kcondvar_t zone_cv; /* used to signal state changes */
555 struct proc *zone_zsched; /* Dummy kernel "zsched" process */
556 pid_t zone_proc_initpid; /* pid of "init" for this zone */
557 char *zone_initname; /* fs path to 'init' */
558 int zone_init_status; /* init's exit status */
559 int zone_boot_err; /* for zone_boot() if boot fails */
560 char *zone_bootargs; /* arguments passed via zone_boot() */
561 rctl_qty_t zone_phys_mem_ctl; /* current phys. memory limit */
562 /*
563 * zone_kthreads is protected by zone_status_lock.
564 */
565 kthread_t *zone_kthreads; /* kernel threads in zone */
566 struct priv_set *zone_privset; /* limit set for zone */
567 /*
568 * zone_vfslist is protected by vfs_list_lock().
569 */
570 struct vfs *zone_vfslist; /* list of FS's mounted in zone */
571 uint64_t zone_uniqid; /* unique zone generation number */
572 struct cred *zone_kcred; /* kcred-like, zone-limited cred */
573 /*
574 * zone_pool is protected by pool_lock().
575 */
576 struct pool *zone_pool; /* pool the zone is bound to */
577 hrtime_t zone_pool_mod; /* last pool bind modification time */
578 /* zone_psetid is protected by cpu_lock */
579 psetid_t zone_psetid; /* pset the zone is bound to */
580
581 time_t zone_boot_time; /* Similar to boot_time */
582
583 /*
584 * The following two can be read without holding any locks. They are
585 * updated under cpu_lock.
586 */
587 int zone_ncpus; /* zone's idea of ncpus */
588 int zone_ncpus_online; /* zone's idea of ncpus_online */
589 /*
590 * List of ZFS datasets exported to this zone.
591 */
592 list_t zone_datasets; /* list of datasets */
593
594 ts_label_t *zone_slabel; /* zone sensitivity label */
595 int zone_match; /* require label match for packets */
596 tsol_mlp_list_t zone_mlps; /* MLPs on zone-private addresses */
597
598 boolean_t zone_restart_init; /* Restart init if it dies? */
599 boolean_t zone_reboot_on_init_exit; /* Reboot if init dies? */
600 boolean_t zone_setup_app_contract; /* setup contract? */
601 struct brand *zone_brand; /* zone's brand */
602 void *zone_brand_data; /* store brand specific data */
603 id_t zone_defaultcid; /* dflt scheduling class id */
604 boolean_t zone_fixed_hipri; /* fixed sched. hi prio */
605 kstat_t *zone_swapresv_kstat;
606 kstat_t *zone_lockedmem_kstat;
607 /*
608 * zone_dl_list is protected by zone_lock
609 */
610 list_t zone_dl_list;
611 netstack_t *zone_netstack;
612 struct cpucap *zone_cpucap; /* CPU caps data */
613
614 /*
615 * Data and counters used for ZFS fair-share disk IO.
616 */
617 rctl_qty_t zone_zfs_io_pri; /* ZFS IO priority */
618 uint_t zone_zfs_queued[2]; /* sync I/O enqueued count */
619 uint64_t zone_zfs_weight; /* used to prevent starvation */
620 uint64_t zone_io_util; /* IO utilization metric */
621 boolean_t zone_io_util_above_avg; /* IO util percent > avg. */
622 uint16_t zone_io_delay; /* IO delay on logical r/w */
623 kmutex_t zone_stg_io_lock; /* protects IO window data */
624 sys_zio_cntr_t zone_rd_ops; /* Counters for ZFS reads, */
625 sys_zio_cntr_t zone_wr_ops; /* writes and */
626 sys_zio_cntr_t zone_lwr_ops; /* logical writes. */
627
628 /*
629 * kstats and counters for I/O ops and bytes.
630 */
631 kmutex_t zone_io_lock; /* protects I/O statistics */
632 kstat_t *zone_io_ksp;
633 kstat_io_t *zone_io_kiop;
634
635 /*
636 * kstats and counters for VFS ops and bytes.
637 */
638 kmutex_t zone_vfs_lock; /* protects VFS statistics */
639 kstat_t *zone_vfs_ksp;
640 kstat_io_t zone_vfs_rwstats;
641 zone_vfs_kstat_t *zone_vfs_stats;
642
643 /*
644 * kstats for ZFS I/O ops and bytes.
645 */
646 kmutex_t zone_zfs_lock; /* protects ZFS statistics */
647 kstat_io_t zone_zfs_rwstats;
648 zone_zfs_kstat_t *zone_zfs_stats;
649
650 /*
651 * Solaris Auditing per-zone audit context
652 */
653 struct au_kcontext *zone_audit_kctxt;
654 /*
655 * For private use by mntfs.
656 */
657 struct mntelem *zone_mntfs_db;
658 krwlock_t zone_mntfs_db_lock;
659
660 struct klpd_reg *zone_pfexecd;
661
662 char *zone_fs_allowed;
663 rctl_qty_t zone_nprocs; /* number of processes in the zone */
664 rctl_qty_t zone_nprocs_ctl; /* current limit protected by */
665 /* zone_rctls->rcs_lock */
666 kstat_t *zone_nprocs_kstat;
667
668 /*
669 * kstats and counters for physical memory capping.
670 */
671 rctl_qty_t zone_phys_mem; /* current bytes of phys. mem. (RSS) */
672 kstat_t *zone_physmem_kstat;
673 uint64_t zone_mcap_nover; /* # of times over phys. cap */
674 uint64_t zone_mcap_pagedout; /* bytes of mem. paged out */
675 kmutex_t zone_mcap_lock; /* protects mcap statistics */
676 kstat_t *zone_mcap_ksp;
677 zone_mcap_kstat_t *zone_mcap_stats;
678 uint64_t zone_pgpgin; /* pages paged in */
679 uint64_t zone_anonpgin; /* anon pages paged in */
680 uint64_t zone_execpgin; /* exec pages paged in */
681 uint64_t zone_fspgin; /* fs pages paged in */
682 uint64_t zone_anon_alloc_fail; /* cnt of anon alloc fails */
683 uint64_t zone_pf_throttle; /* cnt of page flt throttles */
684 uint64_t zone_pf_throttle_usec; /* time of page flt throttles */
685
686 /* Num usecs to throttle page fault when zone is over phys. mem cap */
687 uint32_t zone_pg_flt_delay;
688
689 /*
690 * Misc. kstats and counters for zone cpu-usage aggregation.
691 * The zone_Xtime values are the sum of the micro-state accounting
692 * values for all threads that are running or have run in the zone.
693 * This is tracked in msacct.c as threads change state.
694 * The zone_stime is the sum of the LMS_SYSTEM times.
695 * The zone_utime is the sum of the LMS_USER times.
696 * The zone_wtime is the sum of the LMS_WAIT_CPU times.
697 * As with per-thread micro-state accounting values, these values are
698 * not scaled to nanosecs. The scaling is done by the
699 * zone_misc_kstat_update function when kstats are requested.
700 */
701 kmutex_t zone_misc_lock; /* protects misc statistics */
702 kstat_t *zone_misc_ksp;
703 zone_misc_kstat_t *zone_misc_stats;
704 uint64_t zone_stime; /* total system time */
705 uint64_t zone_utime; /* total user time */
706 uint64_t zone_wtime; /* total time waiting in runq */
707 /* fork-fail kstat tracking */
708 uint32_t zone_ffcap; /* hit an rctl cap */
709 uint32_t zone_ffnoproc; /* get proc/lwp error */
710 uint32_t zone_ffnomem; /* as_dup/memory error */
711 uint32_t zone_ffmisc; /* misc. other error */
712
713 uint32_t zone_nested_intp; /* nested interp. kstat */
714
715 struct loadavg_s zone_loadavg; /* loadavg for this zone */
716 uint64_t zone_hp_avenrun[3]; /* high-precision avenrun */
717 int zone_avenrun[3]; /* FSCALED avg. run queue len */
718
719 /*
720 * FSS stats updated once per second by fss_decay_usage.
721 */
722 uint32_t zone_fss_gen; /* FSS generation cntr */
723 uint64_t zone_run_ticks; /* tot # of ticks running */
724
725 /*
726 * DTrace-private per-zone state
727 */
728 int zone_dtrace_getf; /* # of unprivileged getf()s */
729
730 /*
731 * Synchronization primitives used to synchronize between mounts and
732 * zone creation/destruction.
733 */
734 int zone_mounts_in_progress;
735 kcondvar_t zone_mount_cv;
736 kmutex_t zone_mount_lock;
737 } zone_t;
738
739 /*
740 * Special value of zone_psetid to indicate that pools are disabled.
741 */
742 #define ZONE_PS_INVAL PS_MYID
743
744
745 extern zone_t zone0;
746 extern zone_t *global_zone;
747 extern uint_t maxzones;
748 extern rctl_hndl_t rc_zone_nlwps;
749 extern rctl_hndl_t rc_zone_nprocs;
750
751 extern long zone(int, void *, void *, void *, void *);
752 extern void zone_zsd_init(void);
753 extern void zone_init(void);
754 extern void zone_hold(zone_t *);
755 extern void zone_rele(zone_t *);
756 extern void zone_init_ref(zone_ref_t *);
757 extern void zone_hold_ref(zone_t *, zone_ref_t *, zone_ref_subsys_t);
758 extern void zone_rele_ref(zone_ref_t *, zone_ref_subsys_t);
759 extern void zone_cred_hold(zone_t *);
760 extern void zone_cred_rele(zone_t *);
761 extern void zone_task_hold(zone_t *);
762 extern void zone_task_rele(zone_t *);
763 extern zone_t *zone_find_by_id(zoneid_t);
764 extern zone_t *zone_find_by_label(const ts_label_t *);
765 extern zone_t *zone_find_by_name(char *);
766 extern zone_t *zone_find_by_any_path(const char *, boolean_t);
767 extern zone_t *zone_find_by_path(const char *);
768 extern zoneid_t getzoneid(void);
769 extern zone_t *zone_find_by_id_nolock(zoneid_t);
770 extern int zone_datalink_walk(zoneid_t, int (*)(datalink_id_t, void *), void *);
771 extern int zone_check_datalink(zoneid_t *, datalink_id_t);
772 extern void zone_loadavg_update();
773
774 /*
775 * Zone-specific data (ZSD) APIs
776 */
777 /*
778 * The following is what code should be initializing its zone_key_t to if it
779 * calls zone_getspecific() without necessarily knowing that zone_key_create()
780 * has been called on the key.
781 */
782 #define ZONE_KEY_UNINITIALIZED 0
783
784 typedef uint_t zone_key_t;
785
786 extern void zone_key_create(zone_key_t *, void *(*)(zoneid_t),
787 void (*)(zoneid_t, void *), void (*)(zoneid_t, void *));
788 extern int zone_key_delete(zone_key_t);
789 extern void *zone_getspecific(zone_key_t, zone_t *);
790 extern int zone_setspecific(zone_key_t, zone_t *, const void *);
791
792 /*
793 * The definition of a zsd_entry is truly private to zone.c and is only
794 * placed here so it can be shared with mdb.
795 *
796 * State maintained for each zone times each registered key, which tracks
797 * the state of the create, shutdown and destroy callbacks.
798 *
799 * zsd_flags is used to keep track of pending actions to avoid holding locks
800 * when calling the create/shutdown/destroy callbacks, since doing so
801 * could lead to deadlocks.
802 */
803 struct zsd_entry {
804 zone_key_t zsd_key; /* Key used to lookup value */
805 void *zsd_data; /* Caller-managed value */
806 /*
807 * Callbacks to be executed when a zone is created, shutdown, and
808 * destroyed, respectively.
809 */
810 void *(*zsd_create)(zoneid_t);
811 void (*zsd_shutdown)(zoneid_t, void *);
812 void (*zsd_destroy)(zoneid_t, void *);
813 list_node_t zsd_linkage;
814 uint16_t zsd_flags; /* See below */
815 kcondvar_t zsd_cv;
816 };
817
818 /*
819 * zsd_flags
820 */
821 #define ZSD_CREATE_NEEDED 0x0001
822 #define ZSD_CREATE_INPROGRESS 0x0002
823 #define ZSD_CREATE_COMPLETED 0x0004
824 #define ZSD_SHUTDOWN_NEEDED 0x0010
825 #define ZSD_SHUTDOWN_INPROGRESS 0x0020
826 #define ZSD_SHUTDOWN_COMPLETED 0x0040
827 #define ZSD_DESTROY_NEEDED 0x0100
828 #define ZSD_DESTROY_INPROGRESS 0x0200
829 #define ZSD_DESTROY_COMPLETED 0x0400
830
831 #define ZSD_CREATE_ALL \
832 (ZSD_CREATE_NEEDED|ZSD_CREATE_INPROGRESS|ZSD_CREATE_COMPLETED)
833 #define ZSD_SHUTDOWN_ALL \
834 (ZSD_SHUTDOWN_NEEDED|ZSD_SHUTDOWN_INPROGRESS|ZSD_SHUTDOWN_COMPLETED)
835 #define ZSD_DESTROY_ALL \
836 (ZSD_DESTROY_NEEDED|ZSD_DESTROY_INPROGRESS|ZSD_DESTROY_COMPLETED)
837
838 #define ZSD_ALL_INPROGRESS \
839 (ZSD_CREATE_INPROGRESS|ZSD_SHUTDOWN_INPROGRESS|ZSD_DESTROY_INPROGRESS)
840
841 /*
842 * Macros to help with zone visibility restrictions.
843 */
844
845 /*
846 * Is process in the global zone?
847 */
848 #define INGLOBALZONE(p) \
849 ((p)->p_zone == global_zone)
850
851 /*
852 * Can process view objects in given zone?
853 */
854 #define HASZONEACCESS(p, zoneid) \
855 ((p)->p_zone->zone_id == (zoneid) || INGLOBALZONE(p))
856
857 /*
858 * Convenience macro to see if a resolved path is visible from within a
859 * given zone.
860 *
861 * The basic idea is that the first (zone_rootpathlen - 1) bytes of the
862 * two strings must be equal. Since the rootpathlen has a trailing '/',
863 * we want to skip everything in the path up to (but not including) the
864 * trailing '/'.
865 */
866 #define ZONE_PATH_VISIBLE(path, zone) \
867 (strncmp((path), (zone)->zone_rootpath, \
868 (zone)->zone_rootpathlen - 1) == 0)
869
870 /*
871 * Convenience macro to go from the global view of a path to that seen
872 * from within said zone. It is the responsibility of the caller to
873 * ensure that the path is a resolved one (ie, no '..'s or '.'s), and is
874 * in fact visible from within the zone.
875 */
876 #define ZONE_PATH_TRANSLATE(path, zone) \
877 (ASSERT(ZONE_PATH_VISIBLE(path, zone)), \
878 (path) + (zone)->zone_rootpathlen - 2)
879
880 /*
881 * Special processes visible in all zones.
882 */
883 #define ZONE_SPECIALPID(x) ((x) == 0 || (x) == 1)
884
885 /*
886 * Zone-safe version of thread_create() to be used when the caller wants to
887 * create a kernel thread to run within the current zone's context.
888 */
889 extern kthread_t *zthread_create(caddr_t, size_t, void (*)(), void *, size_t,
890 pri_t);
891 extern void zthread_exit(void);
892
893 /*
894 * Functions for an external observer to register interest in a zone's status
895 * change. Observers will be woken up when the zone status equals the status
896 * argument passed in (in the case of zone_status_timedwait, the function may
897 * also return because of a timeout; zone_status_wait_sig may return early due
898 * to a signal being delivered; zone_status_timedwait_sig may return for any of
899 * the above reasons).
900 *
901 * Otherwise these behave identically to cv_timedwait(), cv_wait(), and
902 * cv_wait_sig() respectively.
903 */
904 extern clock_t zone_status_timedwait(zone_t *, clock_t, zone_status_t);
905 extern clock_t zone_status_timedwait_sig(zone_t *, clock_t, zone_status_t);
906 extern void zone_status_wait(zone_t *, zone_status_t);
907 extern int zone_status_wait_sig(zone_t *, zone_status_t);
908
909 /*
910 * Get the status of the zone (at the time it was called). The state may
911 * have progressed by the time it is returned.
912 */
913 extern zone_status_t zone_status_get(zone_t *);
914
915 /*
916 * Safely get the hostid of the specified zone (defaults to machine's hostid
917 * if the specified zone doesn't emulate a hostid). Passing NULL retrieves
918 * the global zone's (i.e., physical system's) hostid.
919 */
920 extern uint32_t zone_get_hostid(zone_t *);
921
922 /*
923 * Get the "kcred" credentials corresponding to the given zone.
924 */
925 extern struct cred *zone_get_kcred(zoneid_t);
926
927 /*
928 * Get/set the pool the zone is currently bound to.
929 */
930 extern struct pool *zone_pool_get(zone_t *);
931 extern void zone_pool_set(zone_t *, struct pool *);
932
933 /*
934 * Get/set the pset the zone is currently using.
935 */
936 extern psetid_t zone_pset_get(zone_t *);
937 extern void zone_pset_set(zone_t *, psetid_t);
938
939 /*
940 * Get the number of cpus/online-cpus visible from the given zone.
941 */
942 extern int zone_ncpus_get(zone_t *);
943 extern int zone_ncpus_online_get(zone_t *);
944
945 /*
946 * Returns true if the named pool/dataset is visible in the current zone.
947 */
948 extern int zone_dataset_visible(const char *, int *);
949 extern int zone_dataset_visible_inzone(zone_t *, const char *, int *);
950
951 /*
952 * zone version of kadmin()
953 */
954 extern int zone_kadmin(int, int, const char *, cred_t *);
955 extern void zone_shutdown_global(void);
956
957 extern void mount_in_progress(zone_t *);
958 extern void mount_completed(zone_t *);
959
960 extern int zone_walk(int (*)(zone_t *, void *), void *);
961
962 extern rctl_hndl_t rc_zone_locked_mem;
963 extern rctl_hndl_t rc_zone_max_swap;
964 extern rctl_hndl_t rc_zone_phys_mem;
965 extern rctl_hndl_t rc_zone_max_lofi;
966
967 #endif /* _KERNEL */
968
969 #ifdef __cplusplus
970 }
971 #endif
972
973 #endif /* _SYS_ZONE_H */