xref: /titanic_50/usr/src/cmd/fm/fmd/common/fmd_dr.c (revision f3324781c875e2f9865c291e43f86ee710b0c145)
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 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 /*
29  * FMD Dynamic Reconfiguration (DR) Event Handling
30  *
31  * Fault manager scheme plug-ins must track characteristics of individual
32  * pieces of hardware.  As these components can be added or removed by a DR
33  * operation, we need to provide a means by which plug-ins can determine when
34  * they need to re-examine the current configuration.  We provide a simple
35  * mechanism whereby this task can be implemented using lazy evaluation: a
36  * simple 64-bit generation counter is maintained and incremented on *any* DR.
37  * Schemes can store the generation number in scheme-specific data structures,
38  * and then revalidate their contents if the current generation number has
39  * changed since the resource information was cached.  This method saves time,
40  * avoids the complexity of direct participation in DR, avoids the need for
41  * resource-specific processing of DR events, and is relatively easy to port
42  * to other systems that support dynamic reconfiguration.
43  *
44  * The dr generation is only incremented in response to hardware changes.  Since
45  * ASRUs can be in any scheme, including the device scheme, we must also be
46  * aware of software configuration changes which may affect the resource cache.
47  * In addition, we take a snapshot of the topology whenever a reconfiguration
48  * event occurs and notify any modules of the change.
49  */
50 
51 #include <sys/types.h>
52 #include <sys/sunddi.h>
53 #include <sys/sysevent/dr.h>
54 #include <sys/sysevent/eventdefs.h>
55 
56 #include <stdio.h>
57 #include <string.h>
58 #include <unistd.h>
59 #include <libsysevent.h>
60 
61 #undef MUTEX_HELD
62 #undef RW_READ_HELD
63 #undef RW_WRITE_HELD
64 
65 #include <fmd_asru.h>
66 #include <fmd_error.h>
67 #include <fmd_event.h>
68 #include <fmd_fmri.h>
69 #include <fmd_module.h>
70 #include <fmd_subr.h>
71 #include <fmd_topo.h>
72 #include <fmd.h>
73 
74 static void
75 fmd_dr_event(sysevent_t *sep)
76 {
77 	uint64_t gen;
78 	fmd_event_t *e;
79 	const char *class = sysevent_get_class_name(sep);
80 	hrtime_t evtime;
81 	fmd_topo_t *ftp, *prev;
82 	boolean_t update_topo = B_FALSE;
83 
84 	/*
85 	 * The dr generation is only changed in response to DR events.
86 	 */
87 	if (strcmp(class, EC_DR) == 0) {
88 		update_topo = B_TRUE;
89 
90 		(void) pthread_mutex_lock(&fmd.d_stats_lock);
91 		gen = fmd.d_stats->ds_dr_gen.fmds_value.ui64++;
92 		(void) pthread_mutex_unlock(&fmd.d_stats_lock);
93 
94 		TRACE((FMD_DBG_XPRT, "dr event %p, gen=%llu",
95 		    (void *)sep, gen));
96 	}
97 
98 	/*
99 	 * Take a topo snapshot and notify modules of the change.  Picking an
100 	 * accurate time here is difficult.  On one hand, we have the timestamp
101 	 * of the underlying sysevent, indicating when the reconfiguration event
102 	 * occurred.  On the other hand, we are taking the topo snapshot
103 	 * asynchronously, and hence the timestamp of the snapshot is the
104 	 * current time.  Pretending this topo snapshot was valid at the time
105 	 * the sysevent was posted seems wrong, so we instead opt for the
106 	 * current time as an upper bound on the snapshot validity.
107 	 *
108 	 * Along these lines, we keep track of the last time we dispatched a
109 	 * topo snapshot.  If the sysevent occurred before the last topo
110 	 * snapshot, then don't bother dispatching another topo change event.
111 	 * We've already indicated (to the best of our ability) the change in
112 	 * topology.  This prevents endless topo snapshots in response to a
113 	 * flurry of sysevents.
114 	 */
115 	sysevent_get_time(sep, &evtime);
116 	prev = fmd_topo_hold();
117 	if (evtime <= prev->ft_time &&
118 	    fmd.d_clockops == &fmd_timeops_native) {
119 		fmd_topo_rele(prev);
120 		return;
121 	}
122 	fmd_topo_rele(prev);
123 
124 	if (update_topo)
125 		fmd_topo_update();
126 
127 	ftp = fmd_topo_hold();
128 	e = fmd_event_create(FMD_EVT_TOPO, ftp->ft_time, NULL, ftp);
129 	fmd_modhash_dispatch(fmd.d_mod_hash, e);
130 }
131 
132 void
133 fmd_dr_init(void)
134 {
135 	const char *dr_subclasses[] = {
136 	    ESC_DR_AP_STATE_CHANGE
137 	};
138 	const char *zfs_subclasses[] = {
139 	    ESC_ZFS_VDEV_CLEAR,
140 	    ESC_ZFS_VDEV_REMOVE,
141 	    ESC_ZFS_POOL_DESTROY
142 	};
143 	const char *dev_subclasses[] = {
144 	    EC_SUB_ALL
145 	};
146 
147 	if (geteuid() != 0)
148 		return; /* legacy sysevent mechanism is still root-only */
149 
150 	if ((fmd.d_dr_hdl = sysevent_bind_handle(fmd_dr_event)) == NULL)
151 		fmd_error(EFMD_EXIT, "failed to bind handle for DR sysevent");
152 
153 	if (sysevent_subscribe_event(fmd.d_dr_hdl, EC_DR,
154 	    dr_subclasses, sizeof (dr_subclasses) / sizeof (char *)) == -1)
155 		fmd_error(EFMD_EXIT, "failed to subscribe to DR sysevents");
156 
157 	if (sysevent_subscribe_event(fmd.d_dr_hdl, EC_DEVFS,
158 	    dev_subclasses, sizeof (dev_subclasses) / sizeof (char *)) == -1)
159 		fmd_error(EFMD_EXIT, "failed to subscribe to devfs sysevents");
160 
161 	if (sysevent_subscribe_event(fmd.d_dr_hdl, EC_ZFS,
162 	    zfs_subclasses, sizeof (zfs_subclasses) / sizeof (char *)) == -1)
163 		fmd_error(EFMD_EXIT, "failed to subscribe to ZFS sysevents");
164 }
165 
166 void
167 fmd_dr_fini(void)
168 {
169 	if (fmd.d_dr_hdl != NULL) {
170 		sysevent_unsubscribe_event(fmd.d_dr_hdl, EC_DR);
171 		sysevent_unbind_handle(fmd.d_dr_hdl);
172 	}
173 }
174