xref: /illumos-gate/usr/src/uts/common/fs/zfs/zfs_fm.c (revision a38ddfee9c8c6b6c5a2947ff52fd2338362a4444)
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 #include <sys/spa.h>
27 #include <sys/spa_impl.h>
28 #include <sys/vdev.h>
29 #include <sys/vdev_impl.h>
30 #include <sys/zio.h>
31 
32 #include <sys/fm/fs/zfs.h>
33 #include <sys/fm/protocol.h>
34 #include <sys/fm/util.h>
35 #include <sys/sysevent.h>
36 
37 /*
38  * This general routine is responsible for generating all the different ZFS
39  * ereports.  The payload is dependent on the class, and which arguments are
40  * supplied to the function:
41  *
42  * 	EREPORT			POOL	VDEV	IO
43  * 	block			X	X	X
44  * 	data			X		X
45  * 	device			X	X
46  * 	pool			X
47  *
48  * If we are in a loading state, all errors are chained together by the same
49  * SPA-wide ENA (Error Numeric Association).
50  *
51  * For isolated I/O requests, we get the ENA from the zio_t. The propagation
52  * gets very complicated due to RAID-Z, gang blocks, and vdev caching.  We want
53  * to chain together all ereports associated with a logical piece of data.  For
54  * read I/Os, there  are basically three 'types' of I/O, which form a roughly
55  * layered diagram:
56  *
57  *      +---------------+
58  * 	| Aggregate I/O |	No associated logical data or device
59  * 	+---------------+
60  *              |
61  *              V
62  * 	+---------------+	Reads associated with a piece of logical data.
63  * 	|   Read I/O    |	This includes reads on behalf of RAID-Z,
64  * 	+---------------+       mirrors, gang blocks, retries, etc.
65  *              |
66  *              V
67  * 	+---------------+	Reads associated with a particular device, but
68  * 	| Physical I/O  |	no logical data.  Issued as part of vdev caching
69  * 	+---------------+	and I/O aggregation.
70  *
71  * Note that 'physical I/O' here is not the same terminology as used in the rest
72  * of ZIO.  Typically, 'physical I/O' simply means that there is no attached
73  * blockpointer.  But I/O with no associated block pointer can still be related
74  * to a logical piece of data (i.e. RAID-Z requests).
75  *
76  * Purely physical I/O always have unique ENAs.  They are not related to a
77  * particular piece of logical data, and therefore cannot be chained together.
78  * We still generate an ereport, but the DE doesn't correlate it with any
79  * logical piece of data.  When such an I/O fails, the delegated I/O requests
80  * will issue a retry, which will trigger the 'real' ereport with the correct
81  * ENA.
82  *
83  * We keep track of the ENA for a ZIO chain through the 'io_logical' member.
84  * When a new logical I/O is issued, we set this to point to itself.  Child I/Os
85  * then inherit this pointer, so that when it is first set subsequent failures
86  * will use the same ENA.  For vdev cache fill and queue aggregation I/O,
87  * this pointer is set to NULL, and no ereport will be generated (since it
88  * doesn't actually correspond to any particular device or piece of data,
89  * and the caller will always retry without caching or queueing anyway).
90  */
91 void
92 zfs_ereport_post(const char *subclass, spa_t *spa, vdev_t *vd, zio_t *zio,
93     uint64_t stateoroffset, uint64_t size)
94 {
95 #ifdef _KERNEL
96 	nvlist_t *ereport, *detector;
97 	uint64_t ena;
98 	char class[64];
99 	int state;
100 
101 	/*
102 	 * If we are doing a spa_tryimport(), ignore errors.
103 	 */
104 	if (spa->spa_load_state == SPA_LOAD_TRYIMPORT)
105 		return;
106 
107 	/*
108 	 * If we are in the middle of opening a pool, and the previous attempt
109 	 * failed, don't bother logging any new ereports - we're just going to
110 	 * get the same diagnosis anyway.
111 	 */
112 	if (spa->spa_load_state != SPA_LOAD_NONE &&
113 	    spa->spa_last_open_failed)
114 		return;
115 
116 	if (zio != NULL) {
117 		/*
118 		 * If this is not a read or write zio, ignore the error.  This
119 		 * can occur if the DKIOCFLUSHWRITECACHE ioctl fails.
120 		 */
121 		if (zio->io_type != ZIO_TYPE_READ &&
122 		    zio->io_type != ZIO_TYPE_WRITE)
123 			return;
124 
125 		/*
126 		 * Ignore any errors from speculative I/Os, as failure is an
127 		 * expected result.
128 		 */
129 		if (zio->io_flags & ZIO_FLAG_SPECULATIVE)
130 			return;
131 
132 		/*
133 		 * If the vdev has already been marked as failing due to a
134 		 * failed probe, then ignore any subsequent I/O errors, as the
135 		 * DE will automatically fault the vdev on the first such
136 		 * failure.
137 		 */
138 		if (vd != NULL &&
139 		    (!vdev_readable(vd) || !vdev_writeable(vd)) &&
140 		    strcmp(subclass, FM_EREPORT_ZFS_PROBE_FAILURE) != 0)
141 			return;
142 	}
143 
144 	if ((ereport = fm_nvlist_create(NULL)) == NULL)
145 		return;
146 
147 	if ((detector = fm_nvlist_create(NULL)) == NULL) {
148 		fm_nvlist_destroy(ereport, FM_NVA_FREE);
149 		return;
150 	}
151 
152 	/*
153 	 * Serialize ereport generation
154 	 */
155 	mutex_enter(&spa->spa_errlist_lock);
156 
157 	/*
158 	 * Determine the ENA to use for this event.  If we are in a loading
159 	 * state, use a SPA-wide ENA.  Otherwise, if we are in an I/O state, use
160 	 * a root zio-wide ENA.  Otherwise, simply use a unique ENA.
161 	 */
162 	if (spa->spa_load_state != SPA_LOAD_NONE) {
163 		if (spa->spa_ena == 0)
164 			spa->spa_ena = fm_ena_generate(0, FM_ENA_FMT1);
165 		ena = spa->spa_ena;
166 	} else if (zio != NULL && zio->io_logical != NULL) {
167 		if (zio->io_logical->io_ena == 0)
168 			zio->io_logical->io_ena =
169 			    fm_ena_generate(0, FM_ENA_FMT1);
170 		ena = zio->io_logical->io_ena;
171 	} else {
172 		ena = fm_ena_generate(0, FM_ENA_FMT1);
173 	}
174 
175 	/*
176 	 * Construct the full class, detector, and other standard FMA fields.
177 	 */
178 	(void) snprintf(class, sizeof (class), "%s.%s",
179 	    ZFS_ERROR_CLASS, subclass);
180 
181 	fm_fmri_zfs_set(detector, FM_ZFS_SCHEME_VERSION, spa_guid(spa),
182 	    vd != NULL ? vd->vdev_guid : 0);
183 
184 	fm_ereport_set(ereport, FM_EREPORT_VERSION, class, ena, detector, NULL);
185 
186 	/*
187 	 * Construct the per-ereport payload, depending on which parameters are
188 	 * passed in.
189 	 */
190 
191 	/*
192 	 * If we are importing a faulted pool, then we treat it like an open,
193 	 * not an import.  Otherwise, the DE will ignore all faults during
194 	 * import, since the default behavior is to mark the devices as
195 	 * persistently unavailable, not leave them in the faulted state.
196 	 */
197 	state = spa->spa_import_faulted ? SPA_LOAD_OPEN : spa->spa_load_state;
198 
199 	/*
200 	 * Generic payload members common to all ereports.
201 	 */
202 	fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_POOL,
203 	    DATA_TYPE_STRING, spa_name(spa), FM_EREPORT_PAYLOAD_ZFS_POOL_GUID,
204 	    DATA_TYPE_UINT64, spa_guid(spa),
205 	    FM_EREPORT_PAYLOAD_ZFS_POOL_CONTEXT, DATA_TYPE_INT32,
206 	    state, NULL);
207 
208 	if (spa != NULL) {
209 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_POOL_FAILMODE,
210 		    DATA_TYPE_STRING,
211 		    spa_get_failmode(spa) == ZIO_FAILURE_MODE_WAIT ?
212 		    FM_EREPORT_FAILMODE_WAIT :
213 		    spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE ?
214 		    FM_EREPORT_FAILMODE_CONTINUE : FM_EREPORT_FAILMODE_PANIC,
215 		    NULL);
216 	}
217 
218 	if (vd != NULL) {
219 		vdev_t *pvd = vd->vdev_parent;
220 
221 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID,
222 		    DATA_TYPE_UINT64, vd->vdev_guid,
223 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_TYPE,
224 		    DATA_TYPE_STRING, vd->vdev_ops->vdev_op_type, NULL);
225 		if (vd->vdev_path)
226 			fm_payload_set(ereport,
227 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_PATH,
228 			    DATA_TYPE_STRING, vd->vdev_path, NULL);
229 		if (vd->vdev_devid)
230 			fm_payload_set(ereport,
231 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_DEVID,
232 			    DATA_TYPE_STRING, vd->vdev_devid, NULL);
233 
234 		if (pvd != NULL) {
235 			fm_payload_set(ereport,
236 			    FM_EREPORT_PAYLOAD_ZFS_PARENT_GUID,
237 			    DATA_TYPE_UINT64, pvd->vdev_guid,
238 			    FM_EREPORT_PAYLOAD_ZFS_PARENT_TYPE,
239 			    DATA_TYPE_STRING, pvd->vdev_ops->vdev_op_type,
240 			    NULL);
241 			if (pvd->vdev_path)
242 				fm_payload_set(ereport,
243 				    FM_EREPORT_PAYLOAD_ZFS_PARENT_PATH,
244 				    DATA_TYPE_STRING, pvd->vdev_path, NULL);
245 			if (pvd->vdev_devid)
246 				fm_payload_set(ereport,
247 				    FM_EREPORT_PAYLOAD_ZFS_PARENT_DEVID,
248 				    DATA_TYPE_STRING, pvd->vdev_devid, NULL);
249 		}
250 	}
251 
252 	if (zio != NULL) {
253 		/*
254 		 * Payload common to all I/Os.
255 		 */
256 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_ZIO_ERR,
257 		    DATA_TYPE_INT32, zio->io_error, NULL);
258 
259 		/*
260 		 * If the 'size' parameter is non-zero, it indicates this is a
261 		 * RAID-Z or other I/O where the physical offset and length are
262 		 * provided for us, instead of within the zio_t.
263 		 */
264 		if (vd != NULL) {
265 			if (size)
266 				fm_payload_set(ereport,
267 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_OFFSET,
268 				    DATA_TYPE_UINT64, stateoroffset,
269 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_SIZE,
270 				    DATA_TYPE_UINT64, size, NULL);
271 			else
272 				fm_payload_set(ereport,
273 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_OFFSET,
274 				    DATA_TYPE_UINT64, zio->io_offset,
275 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_SIZE,
276 				    DATA_TYPE_UINT64, zio->io_size, NULL);
277 		}
278 
279 		/*
280 		 * Payload for I/Os with corresponding logical information.
281 		 */
282 		if (zio->io_logical != NULL)
283 			fm_payload_set(ereport,
284 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_OBJSET,
285 			    DATA_TYPE_UINT64,
286 			    zio->io_logical->io_bookmark.zb_objset,
287 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_OBJECT,
288 			    DATA_TYPE_UINT64,
289 			    zio->io_logical->io_bookmark.zb_object,
290 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_LEVEL,
291 			    DATA_TYPE_INT64,
292 			    zio->io_logical->io_bookmark.zb_level,
293 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_BLKID,
294 			    DATA_TYPE_UINT64,
295 			    zio->io_logical->io_bookmark.zb_blkid, NULL);
296 	} else if (vd != NULL) {
297 		/*
298 		 * If we have a vdev but no zio, this is a device fault, and the
299 		 * 'stateoroffset' parameter indicates the previous state of the
300 		 * vdev.
301 		 */
302 		fm_payload_set(ereport,
303 		    FM_EREPORT_PAYLOAD_ZFS_PREV_STATE,
304 		    DATA_TYPE_UINT64, stateoroffset, NULL);
305 	}
306 	mutex_exit(&spa->spa_errlist_lock);
307 
308 	fm_ereport_post(ereport, EVCH_SLEEP);
309 
310 	fm_nvlist_destroy(ereport, FM_NVA_FREE);
311 	fm_nvlist_destroy(detector, FM_NVA_FREE);
312 #endif
313 }
314 
315 static void
316 zfs_post_common(spa_t *spa, vdev_t *vd, const char *name)
317 {
318 #ifdef _KERNEL
319 	nvlist_t *resource;
320 	char class[64];
321 
322 	if ((resource = fm_nvlist_create(NULL)) == NULL)
323 		return;
324 
325 	(void) snprintf(class, sizeof (class), "%s.%s.%s", FM_RSRC_RESOURCE,
326 	    ZFS_ERROR_CLASS, name);
327 	VERIFY(nvlist_add_uint8(resource, FM_VERSION, FM_RSRC_VERSION) == 0);
328 	VERIFY(nvlist_add_string(resource, FM_CLASS, class) == 0);
329 	VERIFY(nvlist_add_uint64(resource,
330 	    FM_EREPORT_PAYLOAD_ZFS_POOL_GUID, spa_guid(spa)) == 0);
331 	if (vd)
332 		VERIFY(nvlist_add_uint64(resource,
333 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID, vd->vdev_guid) == 0);
334 
335 	fm_ereport_post(resource, EVCH_SLEEP);
336 
337 	fm_nvlist_destroy(resource, FM_NVA_FREE);
338 #endif
339 }
340 
341 /*
342  * The 'resource.fs.zfs.removed' event is an internal signal that the given vdev
343  * has been removed from the system.  This will cause the DE to ignore any
344  * recent I/O errors, inferring that they are due to the asynchronous device
345  * removal.
346  */
347 void
348 zfs_post_remove(spa_t *spa, vdev_t *vd)
349 {
350 	zfs_post_common(spa, vd, FM_RESOURCE_REMOVED);
351 }
352 
353 /*
354  * The 'resource.fs.zfs.autoreplace' event is an internal signal that the pool
355  * has the 'autoreplace' property set, and therefore any broken vdevs will be
356  * handled by higher level logic, and no vdev fault should be generated.
357  */
358 void
359 zfs_post_autoreplace(spa_t *spa, vdev_t *vd)
360 {
361 	zfs_post_common(spa, vd, FM_RESOURCE_AUTOREPLACE);
362 }
363