xref: /illumos-gate/usr/src/cmd/fm/fmd/common/fmd_ctl.c (revision c51cb4bc539e1650eb5bb4f805cc779bfce99c06)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 
23 /*
24  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
25  * Use is subject to license terms.
26  */
27 
28 #pragma ident	"%Z%%M%	%I%	%E% SMI"
29 
30 /*
31  * FMD Control Event Subsystem
32  *
33  * This file provides a simple and extensible subsystem for the processing of
34  * synchronous control events that can be received from the event transport
35  * and used to control the behavior of the fault manager itself.  At present
36  * this feature is used for the implementation of simulation controls such as
37  * advancing the simulated clock using events sent by the fminject utility.
38  * Control events are assigned a class of the form "resource.fm.fmd.*" and
39  * are assigned a callback function defined in the _fmd_ctls[] table below.
40  * As control events are received by the event transport, they are assigned a
41  * special event type (ev_type = FMD_EVT_CTL) and the ev_data member is used
42  * to refer to a fmd_ctl_t data structure, managed by the functions below.
43  *
44  * Control events are implemented so that they are synchronous with respect to
45  * the rest of the fault manager event stream, which is usually asynchronous
46  * (that is, the transport dispatch thread and the module receive threads all
47  * execute in parallel).  Synchronous processing is required for control events
48  * so that they can affect global state (e.g. the simulated clock) and ensure
49  * that the results of any state changes are seen by *all* subsequent events.
50  *
51  * To achieve synchronization, the event itself implements a thread barrier:
52  * the fmd_ctl_t maintains a reference count that mirrors the fmd_event_t
53  * reference count (which for ctls counts the number of modules the event
54  * was dispatched to).  As each module receive thread dequeues the event, it
55  * calls fmd_event_rele() to discard the event, which calls fmd_ctl_rele().
56  * fmd_ctl_rele() decrements the ctl's reference count but blocks there waiting
57  * for *all* other references to be released.  When all threads have reached
58  * the barrier, the final caller of fmd_ctl_rele() executes the control event
59  * callback function and then wakes everyone else up.  The transport dispatch
60  * thread, blocked in fmd_modhash_dispatch(), is typically this final caller.
61  */
62 
63 #include <strings.h>
64 #include <limits.h>
65 #include <signal.h>
66 
67 #include <fmd_protocol.h>
68 #include <fmd_alloc.h>
69 #include <fmd_error.h>
70 #include <fmd_subr.h>
71 #include <fmd_time.h>
72 #include <fmd_module.h>
73 #include <fmd_thread.h>
74 #include <fmd_ctl.h>
75 
76 #include <fmd.h>
77 
78 static void
79 fmd_ctl_addhrt(nvlist_t *nvl)
80 {
81 	int64_t delta = 0;
82 
83 	(void) nvlist_lookup_int64(nvl, FMD_CTL_ADDHRT_DELTA, &delta);
84 	fmd_time_addhrtime(delta);
85 
86 	/*
87 	 * If the non-adjustable clock has reached the apocalypse, fmd(1M)
88 	 * should exit gracefully: queue a SIGTERM for the main thread.
89 	 */
90 	if (fmd_time_gethrtime() == INT64_MAX)
91 		(void) pthread_kill(fmd.d_rmod->mod_thread->thr_tid, SIGTERM);
92 }
93 
94 static void
95 fmd_ctl_inval(nvlist_t *nvl)
96 {
97 	char *class = "<unknown>";
98 
99 	(void) nvlist_lookup_string(nvl, FM_CLASS, &class);
100 	fmd_error(EFMD_CTL_INVAL, "ignoring invalid control event %s\n", class);
101 }
102 
103 /*ARGSUSED*/
104 static void
105 fmd_ctl_pause(nvlist_t *nvl)
106 {
107 	fmd_dprintf(FMD_DBG_DISP, "unpausing modules from ctl barrier\n");
108 }
109 
110 static const fmd_ctl_desc_t _fmd_ctls[] = {
111 	{ FMD_CTL_ADDHRT, FMD_CTL_ADDHRT_VERS1, fmd_ctl_addhrt },
112 	{ NULL, UINT_MAX, fmd_ctl_inval }
113 };
114 
115 fmd_ctl_t *
116 fmd_ctl_init(nvlist_t *nvl)
117 {
118 	fmd_ctl_t *cp = fmd_alloc(sizeof (fmd_ctl_t), FMD_SLEEP);
119 
120 	const fmd_ctl_desc_t *dp;
121 	uint8_t vers;
122 	char *class;
123 
124 	(void) pthread_mutex_init(&cp->ctl_lock, NULL);
125 	(void) pthread_cond_init(&cp->ctl_cv, NULL);
126 
127 	cp->ctl_nvl = nvl;
128 	cp->ctl_refs = 0;
129 
130 	if (nvl == NULL) {
131 		cp->ctl_func = fmd_ctl_pause;
132 		return (cp);
133 	}
134 
135 	if (nvlist_lookup_string(nvl, FM_CLASS, &class) != 0 ||
136 	    nvlist_lookup_uint8(nvl, FM_VERSION, &vers) != 0)
137 		fmd_panic("ctl_init called with bad nvlist %p", (void *)nvl);
138 
139 	for (dp = _fmd_ctls; dp->cde_class != NULL; dp++) {
140 		if (strcmp(class, dp->cde_class) == 0)
141 			break;
142 	}
143 
144 	cp->ctl_func = vers > dp->cde_vers ? &fmd_ctl_inval : dp->cde_func;
145 	return (cp);
146 }
147 
148 void
149 fmd_ctl_fini(fmd_ctl_t *cp)
150 {
151 	fmd_free(cp, sizeof (fmd_ctl_t));
152 }
153 
154 /*
155  * Increment the ref count on the fmd_ctl_t to correspond to a reference to the
156  * fmd_event_t.  This count is used to implement a barrier in fmd_ctl_rele().
157  */
158 void
159 fmd_ctl_hold(fmd_ctl_t *cp)
160 {
161 	(void) pthread_mutex_lock(&cp->ctl_lock);
162 
163 	cp->ctl_refs++;
164 	ASSERT(cp->ctl_refs != 0);
165 
166 	(void) pthread_mutex_unlock(&cp->ctl_lock);
167 }
168 
169 /*
170  * Decrement the reference count on the fmd_ctl_t.  If this rele() is the last
171  * one, then execute the callback function and release all the other callers.
172  * Otherwise enter a loop waiting on ctl_cv for other threads to call rele().
173  */
174 void
175 fmd_ctl_rele(fmd_ctl_t *cp)
176 {
177 	(void) pthread_mutex_lock(&cp->ctl_lock);
178 
179 	ASSERT(cp->ctl_refs != 0);
180 	cp->ctl_refs--;
181 
182 	if (cp->ctl_refs == 0) {
183 		cp->ctl_func(cp->ctl_nvl);
184 		(void) pthread_cond_broadcast(&cp->ctl_cv);
185 	} else {
186 		while (cp->ctl_refs != 0)
187 			(void) pthread_cond_wait(&cp->ctl_cv, &cp->ctl_lock);
188 	}
189 
190 	(void) pthread_mutex_unlock(&cp->ctl_lock);
191 }
192