xref: /illumos-gate/usr/src/uts/common/fs/zfs/zfs_fm.c (revision abdf5d9abf528d6c318fd8533e09bc3cac1f228b)
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 2009 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 
100 	/*
101 	 * If we are doing a spa_tryimport(), ignore errors.
102 	 */
103 	if (spa->spa_load_state == SPA_LOAD_TRYIMPORT)
104 		return;
105 
106 	/*
107 	 * If we are in the middle of opening a pool, and the previous attempt
108 	 * failed, don't bother logging any new ereports - we're just going to
109 	 * get the same diagnosis anyway.
110 	 */
111 	if (spa->spa_load_state != SPA_LOAD_NONE &&
112 	    spa->spa_last_open_failed)
113 		return;
114 
115 	if (zio != NULL) {
116 		/*
117 		 * If this is not a read or write zio, ignore the error.  This
118 		 * can occur if the DKIOCFLUSHWRITECACHE ioctl fails.
119 		 */
120 		if (zio->io_type != ZIO_TYPE_READ &&
121 		    zio->io_type != ZIO_TYPE_WRITE)
122 			return;
123 
124 		/*
125 		 * Ignore any errors from speculative I/Os, as failure is an
126 		 * expected result.
127 		 */
128 		if (zio->io_flags & ZIO_FLAG_SPECULATIVE)
129 			return;
130 
131 		/*
132 		 * If this I/O is not a retry I/O, don't post an ereport.
133 		 * Otherwise, we risk making bad diagnoses based on B_FAILFAST
134 		 * I/Os.
135 		 */
136 		if (zio->io_error == EIO &&
137 		    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
138 			return;
139 
140 		if (vd != NULL) {
141 			/*
142 			 * If the vdev has already been marked as failing due
143 			 * to a failed probe, then ignore any subsequent I/O
144 			 * errors, as the DE will automatically fault the vdev
145 			 * on the first such failure.  This also catches cases
146 			 * where vdev_remove_wanted is set and the device has
147 			 * not yet been asynchronously placed into the REMOVED
148 			 * state.
149 			 */
150 			if (zio->io_vd == vd && !vdev_accessible(vd, zio))
151 				return;
152 
153 			/*
154 			 * Ignore checksum errors for reads from DTL regions of
155 			 * leaf vdevs.
156 			 */
157 			if (zio->io_type == ZIO_TYPE_READ &&
158 			    zio->io_error == ECKSUM &&
159 			    vd->vdev_ops->vdev_op_leaf &&
160 			    vdev_dtl_contains(vd, DTL_MISSING, zio->io_txg, 1))
161 				return;
162 		}
163 	}
164 
165 	/*
166 	 * For probe failure, we want to avoid posting ereports if we've
167 	 * already removed the device in the meantime.
168 	 */
169 	if (vd != NULL &&
170 	    strcmp(subclass, FM_EREPORT_ZFS_PROBE_FAILURE) == 0 &&
171 	    (vd->vdev_remove_wanted || vd->vdev_state == VDEV_STATE_REMOVED))
172 		return;
173 
174 	if ((ereport = fm_nvlist_create(NULL)) == NULL)
175 		return;
176 
177 	if ((detector = fm_nvlist_create(NULL)) == NULL) {
178 		fm_nvlist_destroy(ereport, FM_NVA_FREE);
179 		return;
180 	}
181 
182 	/*
183 	 * Serialize ereport generation
184 	 */
185 	mutex_enter(&spa->spa_errlist_lock);
186 
187 	/*
188 	 * Determine the ENA to use for this event.  If we are in a loading
189 	 * state, use a SPA-wide ENA.  Otherwise, if we are in an I/O state, use
190 	 * a root zio-wide ENA.  Otherwise, simply use a unique ENA.
191 	 */
192 	if (spa->spa_load_state != SPA_LOAD_NONE) {
193 		if (spa->spa_ena == 0)
194 			spa->spa_ena = fm_ena_generate(0, FM_ENA_FMT1);
195 		ena = spa->spa_ena;
196 	} else if (zio != NULL && zio->io_logical != NULL) {
197 		if (zio->io_logical->io_ena == 0)
198 			zio->io_logical->io_ena =
199 			    fm_ena_generate(0, FM_ENA_FMT1);
200 		ena = zio->io_logical->io_ena;
201 	} else {
202 		ena = fm_ena_generate(0, FM_ENA_FMT1);
203 	}
204 
205 	/*
206 	 * Construct the full class, detector, and other standard FMA fields.
207 	 */
208 	(void) snprintf(class, sizeof (class), "%s.%s",
209 	    ZFS_ERROR_CLASS, subclass);
210 
211 	fm_fmri_zfs_set(detector, FM_ZFS_SCHEME_VERSION, spa_guid(spa),
212 	    vd != NULL ? vd->vdev_guid : 0);
213 
214 	fm_ereport_set(ereport, FM_EREPORT_VERSION, class, ena, detector, NULL);
215 
216 	/*
217 	 * Construct the per-ereport payload, depending on which parameters are
218 	 * passed in.
219 	 */
220 
221 	/*
222 	 * Generic payload members common to all ereports.
223 	 */
224 	fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_POOL,
225 	    DATA_TYPE_STRING, spa_name(spa), FM_EREPORT_PAYLOAD_ZFS_POOL_GUID,
226 	    DATA_TYPE_UINT64, spa_guid(spa),
227 	    FM_EREPORT_PAYLOAD_ZFS_POOL_CONTEXT, DATA_TYPE_INT32,
228 	    spa->spa_load_state, NULL);
229 
230 	if (spa != NULL) {
231 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_POOL_FAILMODE,
232 		    DATA_TYPE_STRING,
233 		    spa_get_failmode(spa) == ZIO_FAILURE_MODE_WAIT ?
234 		    FM_EREPORT_FAILMODE_WAIT :
235 		    spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE ?
236 		    FM_EREPORT_FAILMODE_CONTINUE : FM_EREPORT_FAILMODE_PANIC,
237 		    NULL);
238 	}
239 
240 	if (vd != NULL) {
241 		vdev_t *pvd = vd->vdev_parent;
242 
243 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID,
244 		    DATA_TYPE_UINT64, vd->vdev_guid,
245 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_TYPE,
246 		    DATA_TYPE_STRING, vd->vdev_ops->vdev_op_type, NULL);
247 		if (vd->vdev_path != NULL)
248 			fm_payload_set(ereport,
249 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_PATH,
250 			    DATA_TYPE_STRING, vd->vdev_path, NULL);
251 		if (vd->vdev_devid != NULL)
252 			fm_payload_set(ereport,
253 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_DEVID,
254 			    DATA_TYPE_STRING, vd->vdev_devid, NULL);
255 		if (vd->vdev_fru != NULL)
256 			fm_payload_set(ereport,
257 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_FRU,
258 			    DATA_TYPE_STRING, vd->vdev_fru, NULL);
259 
260 		if (pvd != NULL) {
261 			fm_payload_set(ereport,
262 			    FM_EREPORT_PAYLOAD_ZFS_PARENT_GUID,
263 			    DATA_TYPE_UINT64, pvd->vdev_guid,
264 			    FM_EREPORT_PAYLOAD_ZFS_PARENT_TYPE,
265 			    DATA_TYPE_STRING, pvd->vdev_ops->vdev_op_type,
266 			    NULL);
267 			if (pvd->vdev_path)
268 				fm_payload_set(ereport,
269 				    FM_EREPORT_PAYLOAD_ZFS_PARENT_PATH,
270 				    DATA_TYPE_STRING, pvd->vdev_path, NULL);
271 			if (pvd->vdev_devid)
272 				fm_payload_set(ereport,
273 				    FM_EREPORT_PAYLOAD_ZFS_PARENT_DEVID,
274 				    DATA_TYPE_STRING, pvd->vdev_devid, NULL);
275 		}
276 	}
277 
278 	if (zio != NULL) {
279 		/*
280 		 * Payload common to all I/Os.
281 		 */
282 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_ZIO_ERR,
283 		    DATA_TYPE_INT32, zio->io_error, NULL);
284 
285 		/*
286 		 * If the 'size' parameter is non-zero, it indicates this is a
287 		 * RAID-Z or other I/O where the physical offset and length are
288 		 * provided for us, instead of within the zio_t.
289 		 */
290 		if (vd != NULL) {
291 			if (size)
292 				fm_payload_set(ereport,
293 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_OFFSET,
294 				    DATA_TYPE_UINT64, stateoroffset,
295 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_SIZE,
296 				    DATA_TYPE_UINT64, size, NULL);
297 			else
298 				fm_payload_set(ereport,
299 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_OFFSET,
300 				    DATA_TYPE_UINT64, zio->io_offset,
301 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_SIZE,
302 				    DATA_TYPE_UINT64, zio->io_size, NULL);
303 		}
304 
305 		/*
306 		 * Payload for I/Os with corresponding logical information.
307 		 */
308 		if (zio->io_logical != NULL)
309 			fm_payload_set(ereport,
310 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_OBJSET,
311 			    DATA_TYPE_UINT64,
312 			    zio->io_logical->io_bookmark.zb_objset,
313 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_OBJECT,
314 			    DATA_TYPE_UINT64,
315 			    zio->io_logical->io_bookmark.zb_object,
316 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_LEVEL,
317 			    DATA_TYPE_INT64,
318 			    zio->io_logical->io_bookmark.zb_level,
319 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_BLKID,
320 			    DATA_TYPE_UINT64,
321 			    zio->io_logical->io_bookmark.zb_blkid, NULL);
322 	} else if (vd != NULL) {
323 		/*
324 		 * If we have a vdev but no zio, this is a device fault, and the
325 		 * 'stateoroffset' parameter indicates the previous state of the
326 		 * vdev.
327 		 */
328 		fm_payload_set(ereport,
329 		    FM_EREPORT_PAYLOAD_ZFS_PREV_STATE,
330 		    DATA_TYPE_UINT64, stateoroffset, NULL);
331 	}
332 	mutex_exit(&spa->spa_errlist_lock);
333 
334 	fm_ereport_post(ereport, EVCH_SLEEP);
335 
336 	fm_nvlist_destroy(ereport, FM_NVA_FREE);
337 	fm_nvlist_destroy(detector, FM_NVA_FREE);
338 #endif
339 }
340 
341 static void
342 zfs_post_common(spa_t *spa, vdev_t *vd, const char *name)
343 {
344 #ifdef _KERNEL
345 	nvlist_t *resource;
346 	char class[64];
347 
348 	if (spa->spa_load_state == SPA_LOAD_TRYIMPORT)
349 		return;
350 
351 	if ((resource = fm_nvlist_create(NULL)) == NULL)
352 		return;
353 
354 	(void) snprintf(class, sizeof (class), "%s.%s.%s", FM_RSRC_RESOURCE,
355 	    ZFS_ERROR_CLASS, name);
356 	VERIFY(nvlist_add_uint8(resource, FM_VERSION, FM_RSRC_VERSION) == 0);
357 	VERIFY(nvlist_add_string(resource, FM_CLASS, class) == 0);
358 	VERIFY(nvlist_add_uint64(resource,
359 	    FM_EREPORT_PAYLOAD_ZFS_POOL_GUID, spa_guid(spa)) == 0);
360 	if (vd)
361 		VERIFY(nvlist_add_uint64(resource,
362 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID, vd->vdev_guid) == 0);
363 
364 	fm_ereport_post(resource, EVCH_SLEEP);
365 
366 	fm_nvlist_destroy(resource, FM_NVA_FREE);
367 #endif
368 }
369 
370 /*
371  * The 'resource.fs.zfs.removed' event is an internal signal that the given vdev
372  * has been removed from the system.  This will cause the DE to ignore any
373  * recent I/O errors, inferring that they are due to the asynchronous device
374  * removal.
375  */
376 void
377 zfs_post_remove(spa_t *spa, vdev_t *vd)
378 {
379 	zfs_post_common(spa, vd, FM_RESOURCE_REMOVED);
380 }
381 
382 /*
383  * The 'resource.fs.zfs.autoreplace' event is an internal signal that the pool
384  * has the 'autoreplace' property set, and therefore any broken vdevs will be
385  * handled by higher level logic, and no vdev fault should be generated.
386  */
387 void
388 zfs_post_autoreplace(spa_t *spa, vdev_t *vd)
389 {
390 	zfs_post_common(spa, vd, FM_RESOURCE_AUTOREPLACE);
391 }
392