xref: /titanic_52/usr/src/cmd/fm/fmd/common/fmd_dr.c (revision 193974072f41a843678abf5f61979c748687e66b)
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 2008 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 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 	const char *subclass = sysevent_get_subclass_name(sep);
81 	hrtime_t evtime;
82 	fmd_topo_t *ftp, *prev;
83 	boolean_t update_topo = B_FALSE;
84 
85 	if (strcmp(class, EC_DR) == 0) {
86 		if (strcmp(subclass, ESC_DR_AP_STATE_CHANGE) != 0 &&
87 		    strcmp(subclass, ESC_DR_TARGET_STATE_CHANGE) != 0)
88 			return;
89 
90 		/*
91 		 * The DR generation is only changed in response to DR events.
92 		 */
93 		update_topo = B_TRUE;
94 
95 		(void) pthread_mutex_lock(&fmd.d_stats_lock);
96 		gen = fmd.d_stats->ds_dr_gen.fmds_value.ui64++;
97 		(void) pthread_mutex_unlock(&fmd.d_stats_lock);
98 
99 		TRACE((FMD_DBG_XPRT, "dr event %p, gen=%llu",
100 		    (void *)sep, gen));
101 	} else if (strcmp(class, EC_DEVFS) == 0) {
102 		/*
103 		 * A devfs configuration event can change the topology,
104 		 * as disk nodes only exist when the device is configured.
105 		 */
106 		update_topo = B_TRUE;
107 	} else if (strcmp(class, EC_ZFS) == 0) {
108 		/*
109 		 * These events can change the resource cache.
110 		 */
111 		if (strcmp(subclass, ESC_ZFS_VDEV_CLEAR) != 0 &&
112 		    strcmp(subclass, ESC_ZFS_VDEV_REMOVE) != 0 &&
113 		    strcmp(subclass, ESC_ZFS_POOL_DESTROY) != 0)
114 			return;
115 	} else if (strcmp(class, EC_DEV_ADD) == 0 ||
116 	    strcmp(class, EC_DEV_REMOVE) == 0) {
117 		if (strcmp(subclass, ESC_DISK) != 0)
118 			return;
119 	}
120 
121 	/*
122 	 * Take a topo snapshot and notify modules of the change.  Picking an
123 	 * accurate time here is difficult.  On one hand, we have the timestamp
124 	 * of the underlying sysevent, indicating when the reconfiguration event
125 	 * occurred.  On the other hand, we are taking the topo snapshot
126 	 * asynchronously, and hence the timestamp of the snapshot is the
127 	 * current time.  Pretending this topo snapshot was valid at the time
128 	 * the sysevent was posted seems wrong, so we instead opt for the
129 	 * current time as an upper bound on the snapshot validity.
130 	 *
131 	 * Along these lines, we keep track of the last time we dispatched a
132 	 * topo snapshot.  If the sysevent occurred before the last topo
133 	 * snapshot, then don't bother dispatching another topo change event.
134 	 * We've already indicated (to the best of our ability) the change in
135 	 * topology.  This prevents endless topo snapshots in response to a
136 	 * flurry of sysevents.
137 	 */
138 	sysevent_get_time(sep, &evtime);
139 	prev = fmd_topo_hold();
140 	if (evtime <= prev->ft_time &&
141 	    fmd.d_clockops == &fmd_timeops_native) {
142 		fmd_topo_rele(prev);
143 		return;
144 	}
145 	fmd_topo_rele(prev);
146 
147 	if (update_topo)
148 		fmd_topo_update();
149 
150 	ftp = fmd_topo_hold();
151 	e = fmd_event_create(FMD_EVT_TOPO, ftp->ft_time, NULL, ftp);
152 	fmd_modhash_dispatch(fmd.d_mod_hash, e);
153 }
154