xref: /titanic_50/usr/src/uts/common/fs/zfs/zfs_fm.c (revision c7158ae983f5a04c4a998f468ecefba6d23ba721)
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 #include <sys/spa.h>
29 #include <sys/spa_impl.h>
30 #include <sys/vdev.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/zio.h>
33 
34 #include <sys/fm/fs/zfs.h>
35 #include <sys/fm/protocol.h>
36 #include <sys/fm/util.h>
37 #include <sys/sysevent.h>
38 
39 /*
40  * This general routine is responsible for generating all the different ZFS
41  * ereports.  The payload is dependent on the class, and which arguments are
42  * supplied to the function:
43  *
44  * 	EREPORT			POOL	VDEV	IO
45  * 	block			X	X	X
46  * 	data			X		X
47  * 	device			X	X
48  * 	pool			X
49  *
50  * If we are in a loading state, all errors are chained together by the same
51  * SPA-wide ENA (Error Numeric Association).
52  *
53  * For isolated I/O requests, we get the ENA from the zio_t. The propagation
54  * gets very complicated due to RAID-Z, gang blocks, and vdev caching.  We want
55  * to chain together all ereports associated with a logical piece of data.  For
56  * read I/Os, there  are basically three 'types' of I/O, which form a roughly
57  * layered diagram:
58  *
59  *      +---------------+
60  * 	| Aggregate I/O |	No associated logical data or device
61  * 	+---------------+
62  *              |
63  *              V
64  * 	+---------------+	Reads associated with a piece of logical data.
65  * 	|   Read I/O    |	This includes reads on behalf of RAID-Z,
66  * 	+---------------+       mirrors, gang blocks, retries, etc.
67  *              |
68  *              V
69  * 	+---------------+	Reads associated with a particular device, but
70  * 	| Physical I/O  |	no logical data.  Issued as part of vdev caching
71  * 	+---------------+	and I/O aggregation.
72  *
73  * Note that 'physical I/O' here is not the same terminology as used in the rest
74  * of ZIO.  Typically, 'physical I/O' simply means that there is no attached
75  * blockpointer.  But I/O with no associated block pointer can still be related
76  * to a logical piece of data (i.e. RAID-Z requests).
77  *
78  * Purely physical I/O always have unique ENAs.  They are not related to a
79  * particular piece of logical data, and therefore cannot be chained together.
80  * We still generate an ereport, but the DE doesn't correlate it with any
81  * logical piece of data.  When such an I/O fails, the delegated I/O requests
82  * will issue a retry, which will trigger the 'real' ereport with the correct
83  * ENA.
84  *
85  * We keep track of the ENA for a ZIO chain through the 'io_logical' member.
86  * When a new logical I/O is issued, we set this to point to itself.  Child I/Os
87  * then inherit this pointer, so that when it is first set subsequent failures
88  * will use the same ENA.  If a physical I/O is issued (by passing the
89  * ZIO_FLAG_NOBOOKMARK flag), then this pointer is reset, guaranteeing that a
90  * unique ENA will be generated.  For an aggregate I/O, this pointer is set to
91  * NULL, and no ereport will be generated (since it doesn't actually correspond
92  * to any particular device or piece of data).
93  */
94 void
95 zfs_ereport_post(const char *subclass, spa_t *spa, vdev_t *vd, zio_t *zio,
96     uint64_t stateoroffset, uint64_t size)
97 {
98 #ifdef _KERNEL
99 	nvlist_t *ereport, *detector;
100 	uint64_t ena;
101 	char class[64];
102 	int state;
103 
104 	/*
105 	 * If we are doing a spa_tryimport(), ignore errors.
106 	 */
107 	if (spa->spa_load_state == SPA_LOAD_TRYIMPORT)
108 		return;
109 
110 	/*
111 	 * If we are in the middle of opening a pool, and the previous attempt
112 	 * failed, don't bother logging any new ereports - we're just going to
113 	 * get the same diagnosis anyway.
114 	 */
115 	if (spa->spa_load_state != SPA_LOAD_NONE &&
116 	    spa->spa_last_open_failed)
117 		return;
118 
119 	/*
120 	 * Ignore any errors from I/Os that we are going to retry anyway - we
121 	 * only generate errors from the final failure.  Checksum errors are
122 	 * generated after the pipeline stage responsible for retrying the I/O
123 	 * (VDEV_IO_ASSESS), so this only applies to standard I/O errors.
124 	 */
125 	if (zio && zio_should_retry(zio) && zio->io_error != ECKSUM)
126 		return;
127 
128 	/*
129 	 * If this is not a read or write zio, ignore the error.  This can occur
130 	 * if the DKIOCFLUSHWRITECACHE ioctl fails.
131 	 */
132 	if (zio && zio->io_type != ZIO_TYPE_READ &&
133 	    zio->io_type != ZIO_TYPE_WRITE)
134 		return;
135 
136 	if ((ereport = fm_nvlist_create(NULL)) == NULL)
137 		return;
138 
139 	if ((detector = fm_nvlist_create(NULL)) == NULL) {
140 		fm_nvlist_destroy(ereport, FM_NVA_FREE);
141 		return;
142 	}
143 
144 	/*
145 	 * Serialize ereport generation
146 	 */
147 	mutex_enter(&spa->spa_errlist_lock);
148 
149 	/*
150 	 * Determine the ENA to use for this event.  If we are in a loading
151 	 * state, use a SPA-wide ENA.  Otherwise, if we are in an I/O state, use
152 	 * a root zio-wide ENA.  Otherwise, simply use a unique ENA.
153 	 */
154 	if (spa->spa_load_state != SPA_LOAD_NONE) {
155 		if (spa->spa_ena == 0)
156 			spa->spa_ena = fm_ena_generate(0, FM_ENA_FMT1);
157 		ena = spa->spa_ena;
158 	} else if (zio != NULL && zio->io_logical != NULL) {
159 		if (zio->io_logical->io_ena == 0)
160 			zio->io_logical->io_ena =
161 			    fm_ena_generate(0, FM_ENA_FMT1);
162 		ena = zio->io_logical->io_ena;
163 	} else {
164 		ena = fm_ena_generate(0, FM_ENA_FMT1);
165 	}
166 
167 	/*
168 	 * Construct the full class, detector, and other standard FMA fields.
169 	 */
170 	(void) snprintf(class, sizeof (class), "%s.%s",
171 	    ZFS_ERROR_CLASS, subclass);
172 
173 	fm_fmri_zfs_set(detector, FM_ZFS_SCHEME_VERSION, spa_guid(spa),
174 	    vd != NULL ? vd->vdev_guid : 0);
175 
176 	fm_ereport_set(ereport, FM_EREPORT_VERSION, class, ena, detector, NULL);
177 
178 	/*
179 	 * Construct the per-ereport payload, depending on which parameters are
180 	 * passed in.
181 	 */
182 
183 	/*
184 	 * If we are importing a faulted pool, then we treat it like an open,
185 	 * not an import.  Otherwise, the DE will ignore all faults during
186 	 * import, since the default behavior is to mark the devices as
187 	 * persistently unavailable, not leave them in the faulted state.
188 	 */
189 	state = spa->spa_import_faulted ? SPA_LOAD_OPEN : spa->spa_load_state;
190 
191 	/*
192 	 * Generic payload members common to all ereports.
193 	 *
194 	 * The direct reference to spa_name is used rather than spa_name()
195 	 * because of the asynchronous nature of the zio pipeline.  spa_name()
196 	 * asserts that the config lock is held in some form.  This is always
197 	 * the case in I/O context, but because the check for RW_WRITER compares
198 	 * against 'curthread', we may be in an asynchronous context and blow
199 	 * this assert.  Rather than loosen this assert, we acknowledge that all
200 	 * contexts in which this function is called (pool open, I/O) are safe,
201 	 * and dereference the name directly.
202 	 */
203 	fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_POOL,
204 	    DATA_TYPE_STRING, spa->spa_name, FM_EREPORT_PAYLOAD_ZFS_POOL_GUID,
205 	    DATA_TYPE_UINT64, spa_guid(spa),
206 	    FM_EREPORT_PAYLOAD_ZFS_POOL_CONTEXT, DATA_TYPE_INT32,
207 	    state, NULL);
208 
209 	if (spa != NULL) {
210 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_POOL_FAILMODE,
211 		    DATA_TYPE_STRING,
212 		    spa_get_failmode(spa) == ZIO_FAILURE_MODE_WAIT ?
213 		    FM_EREPORT_FAILMODE_WAIT :
214 		    spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE ?
215 		    FM_EREPORT_FAILMODE_CONTINUE : FM_EREPORT_FAILMODE_PANIC,
216 		    NULL);
217 	}
218 
219 	if (vd != NULL) {
220 		vdev_t *pvd = vd->vdev_parent;
221 
222 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID,
223 		    DATA_TYPE_UINT64, vd->vdev_guid,
224 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_TYPE,
225 		    DATA_TYPE_STRING, vd->vdev_ops->vdev_op_type, NULL);
226 		if (vd->vdev_path)
227 			fm_payload_set(ereport,
228 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_PATH,
229 			    DATA_TYPE_STRING, vd->vdev_path, NULL);
230 		if (vd->vdev_devid)
231 			fm_payload_set(ereport,
232 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_DEVID,
233 			    DATA_TYPE_STRING, vd->vdev_devid, NULL);
234 
235 		if (pvd != NULL) {
236 			fm_payload_set(ereport,
237 			    FM_EREPORT_PAYLOAD_ZFS_PARENT_GUID,
238 			    DATA_TYPE_UINT64, pvd->vdev_guid,
239 			    FM_EREPORT_PAYLOAD_ZFS_PARENT_TYPE,
240 			    DATA_TYPE_STRING, pvd->vdev_ops->vdev_op_type,
241 			    NULL);
242 			if (pvd->vdev_path)
243 				fm_payload_set(ereport,
244 				    FM_EREPORT_PAYLOAD_ZFS_PARENT_PATH,
245 				    DATA_TYPE_STRING, pvd->vdev_path, NULL);
246 			if (pvd->vdev_devid)
247 				fm_payload_set(ereport,
248 				    FM_EREPORT_PAYLOAD_ZFS_PARENT_DEVID,
249 				    DATA_TYPE_STRING, pvd->vdev_devid, NULL);
250 		}
251 	}
252 
253 	if (zio != NULL) {
254 		/*
255 		 * Payload common to all I/Os.
256 		 */
257 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_ZIO_ERR,
258 		    DATA_TYPE_INT32, zio->io_error, NULL);
259 
260 		/*
261 		 * If the 'size' parameter is non-zero, it indicates this is a
262 		 * RAID-Z or other I/O where the physical offset and length are
263 		 * provided for us, instead of within the zio_t.
264 		 */
265 		if (vd != NULL) {
266 			if (size)
267 				fm_payload_set(ereport,
268 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_OFFSET,
269 				    DATA_TYPE_UINT64, stateoroffset,
270 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_SIZE,
271 				    DATA_TYPE_UINT64, size, NULL);
272 			else
273 				fm_payload_set(ereport,
274 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_OFFSET,
275 				    DATA_TYPE_UINT64, zio->io_offset,
276 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_SIZE,
277 				    DATA_TYPE_UINT64, zio->io_size, NULL);
278 		}
279 
280 		/*
281 		 * Payload for I/Os with corresponding logical information.
282 		 */
283 		if (zio->io_logical != NULL)
284 			fm_payload_set(ereport,
285 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_OBJSET,
286 			    DATA_TYPE_UINT64,
287 			    zio->io_logical->io_bookmark.zb_objset,
288 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_OBJECT,
289 			    DATA_TYPE_UINT64,
290 			    zio->io_logical->io_bookmark.zb_object,
291 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_LEVEL,
292 			    DATA_TYPE_INT64,
293 			    zio->io_logical->io_bookmark.zb_level,
294 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_BLKID,
295 			    DATA_TYPE_UINT64,
296 			    zio->io_logical->io_bookmark.zb_blkid, NULL);
297 	} else if (vd != NULL) {
298 		/*
299 		 * If we have a vdev but no zio, this is a device fault, and the
300 		 * 'stateoroffset' parameter indicates the previous state of the
301 		 * vdev.
302 		 */
303 		fm_payload_set(ereport,
304 		    FM_EREPORT_PAYLOAD_ZFS_PREV_STATE,
305 		    DATA_TYPE_UINT64, stateoroffset, NULL);
306 	}
307 	mutex_exit(&spa->spa_errlist_lock);
308 
309 	fm_ereport_post(ereport, EVCH_SLEEP);
310 
311 	fm_nvlist_destroy(ereport, FM_NVA_FREE);
312 	fm_nvlist_destroy(detector, FM_NVA_FREE);
313 #endif
314 }
315 
316 static void
317 zfs_post_common(spa_t *spa, vdev_t *vd, const char *name)
318 {
319 #ifdef _KERNEL
320 	nvlist_t *resource;
321 	char class[64];
322 
323 	if ((resource = fm_nvlist_create(NULL)) == NULL)
324 		return;
325 
326 	(void) snprintf(class, sizeof (class), "%s.%s.%s", FM_RSRC_RESOURCE,
327 	    ZFS_ERROR_CLASS, name);
328 	VERIFY(nvlist_add_uint8(resource, FM_VERSION, FM_RSRC_VERSION) == 0);
329 	VERIFY(nvlist_add_string(resource, FM_CLASS, class) == 0);
330 	VERIFY(nvlist_add_uint64(resource,
331 	    FM_EREPORT_PAYLOAD_ZFS_POOL_GUID, spa_guid(spa)) == 0);
332 	if (vd)
333 		VERIFY(nvlist_add_uint64(resource,
334 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID, vd->vdev_guid) == 0);
335 
336 	fm_ereport_post(resource, EVCH_SLEEP);
337 
338 	fm_nvlist_destroy(resource, FM_NVA_FREE);
339 #endif
340 }
341 
342 /*
343  * The 'resource.fs.zfs.removed' event is an internal signal that the given vdev
344  * has been removed from the system.  This will cause the DE to ignore any
345  * recent I/O errors, inferring that they are due to the asynchronous device
346  * removal.
347  */
348 void
349 zfs_post_remove(spa_t *spa, vdev_t *vd)
350 {
351 	zfs_post_common(spa, vd, FM_RESOURCE_REMOVED);
352 }
353 
354 /*
355  * The 'resource.fs.zfs.autoreplace' event is an internal signal that the pool
356  * has the 'autoreplace' property set, and therefore any broken vdevs will be
357  * handled by higher level logic, and no vdev fault should be generated.
358  */
359 void
360 zfs_post_autoreplace(spa_t *spa, vdev_t *vd)
361 {
362 	zfs_post_common(spa, vd, FM_RESOURCE_AUTOREPLACE);
363 }
364