xref: /titanic_41/usr/src/uts/common/fs/zfs/zfs_fm.c (revision 5dfd244acc8f144280c5bc8f69ed941185fc3ccc)
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 		if (vd != NULL) {
132 			/*
133 			 * If the vdev has already been marked as failing due
134 			 * to a failed probe, then ignore any subsequent I/O
135 			 * errors, as the DE will automatically fault the vdev
136 			 * on the first such failure.  This also catches cases
137 			 * where vdev_remove_wanted is set and the device has
138 			 * not yet been asynchronously placed into the REMOVED
139 			 * state.
140 			 */
141 			if (zio->io_vd == vd &&
142 			    !vdev_accessible(vd, zio) &&
143 			    strcmp(subclass, FM_EREPORT_ZFS_PROBE_FAILURE) != 0)
144 				return;
145 
146 			/*
147 			 * Ignore checksum errors for reads from DTL regions of
148 			 * leaf vdevs.
149 			 */
150 			if (zio->io_type == ZIO_TYPE_READ &&
151 			    zio->io_error == ECKSUM &&
152 			    vd->vdev_ops->vdev_op_leaf &&
153 			    vdev_dtl_contains(vd, DTL_MISSING, zio->io_txg, 1))
154 				return;
155 		}
156 	}
157 
158 	if ((ereport = fm_nvlist_create(NULL)) == NULL)
159 		return;
160 
161 	if ((detector = fm_nvlist_create(NULL)) == NULL) {
162 		fm_nvlist_destroy(ereport, FM_NVA_FREE);
163 		return;
164 	}
165 
166 	/*
167 	 * Serialize ereport generation
168 	 */
169 	mutex_enter(&spa->spa_errlist_lock);
170 
171 	/*
172 	 * Determine the ENA to use for this event.  If we are in a loading
173 	 * state, use a SPA-wide ENA.  Otherwise, if we are in an I/O state, use
174 	 * a root zio-wide ENA.  Otherwise, simply use a unique ENA.
175 	 */
176 	if (spa->spa_load_state != SPA_LOAD_NONE) {
177 		if (spa->spa_ena == 0)
178 			spa->spa_ena = fm_ena_generate(0, FM_ENA_FMT1);
179 		ena = spa->spa_ena;
180 	} else if (zio != NULL && zio->io_logical != NULL) {
181 		if (zio->io_logical->io_ena == 0)
182 			zio->io_logical->io_ena =
183 			    fm_ena_generate(0, FM_ENA_FMT1);
184 		ena = zio->io_logical->io_ena;
185 	} else {
186 		ena = fm_ena_generate(0, FM_ENA_FMT1);
187 	}
188 
189 	/*
190 	 * Construct the full class, detector, and other standard FMA fields.
191 	 */
192 	(void) snprintf(class, sizeof (class), "%s.%s",
193 	    ZFS_ERROR_CLASS, subclass);
194 
195 	fm_fmri_zfs_set(detector, FM_ZFS_SCHEME_VERSION, spa_guid(spa),
196 	    vd != NULL ? vd->vdev_guid : 0);
197 
198 	fm_ereport_set(ereport, FM_EREPORT_VERSION, class, ena, detector, NULL);
199 
200 	/*
201 	 * Construct the per-ereport payload, depending on which parameters are
202 	 * passed in.
203 	 */
204 
205 	/*
206 	 * Generic payload members common to all ereports.
207 	 */
208 	fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_POOL,
209 	    DATA_TYPE_STRING, spa_name(spa), FM_EREPORT_PAYLOAD_ZFS_POOL_GUID,
210 	    DATA_TYPE_UINT64, spa_guid(spa),
211 	    FM_EREPORT_PAYLOAD_ZFS_POOL_CONTEXT, DATA_TYPE_INT32,
212 	    spa->spa_load_state, NULL);
213 
214 	if (spa != NULL) {
215 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_POOL_FAILMODE,
216 		    DATA_TYPE_STRING,
217 		    spa_get_failmode(spa) == ZIO_FAILURE_MODE_WAIT ?
218 		    FM_EREPORT_FAILMODE_WAIT :
219 		    spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE ?
220 		    FM_EREPORT_FAILMODE_CONTINUE : FM_EREPORT_FAILMODE_PANIC,
221 		    NULL);
222 	}
223 
224 	if (vd != NULL) {
225 		vdev_t *pvd = vd->vdev_parent;
226 
227 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID,
228 		    DATA_TYPE_UINT64, vd->vdev_guid,
229 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_TYPE,
230 		    DATA_TYPE_STRING, vd->vdev_ops->vdev_op_type, NULL);
231 		if (vd->vdev_path != NULL)
232 			fm_payload_set(ereport,
233 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_PATH,
234 			    DATA_TYPE_STRING, vd->vdev_path, NULL);
235 		if (vd->vdev_devid != NULL)
236 			fm_payload_set(ereport,
237 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_DEVID,
238 			    DATA_TYPE_STRING, vd->vdev_devid, NULL);
239 		if (vd->vdev_fru != NULL)
240 			fm_payload_set(ereport,
241 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_FRU,
242 			    DATA_TYPE_STRING, vd->vdev_fru, NULL);
243 
244 		if (pvd != NULL) {
245 			fm_payload_set(ereport,
246 			    FM_EREPORT_PAYLOAD_ZFS_PARENT_GUID,
247 			    DATA_TYPE_UINT64, pvd->vdev_guid,
248 			    FM_EREPORT_PAYLOAD_ZFS_PARENT_TYPE,
249 			    DATA_TYPE_STRING, pvd->vdev_ops->vdev_op_type,
250 			    NULL);
251 			if (pvd->vdev_path)
252 				fm_payload_set(ereport,
253 				    FM_EREPORT_PAYLOAD_ZFS_PARENT_PATH,
254 				    DATA_TYPE_STRING, pvd->vdev_path, NULL);
255 			if (pvd->vdev_devid)
256 				fm_payload_set(ereport,
257 				    FM_EREPORT_PAYLOAD_ZFS_PARENT_DEVID,
258 				    DATA_TYPE_STRING, pvd->vdev_devid, NULL);
259 		}
260 	}
261 
262 	if (zio != NULL) {
263 		/*
264 		 * Payload common to all I/Os.
265 		 */
266 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_ZIO_ERR,
267 		    DATA_TYPE_INT32, zio->io_error, NULL);
268 
269 		/*
270 		 * If the 'size' parameter is non-zero, it indicates this is a
271 		 * RAID-Z or other I/O where the physical offset and length are
272 		 * provided for us, instead of within the zio_t.
273 		 */
274 		if (vd != NULL) {
275 			if (size)
276 				fm_payload_set(ereport,
277 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_OFFSET,
278 				    DATA_TYPE_UINT64, stateoroffset,
279 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_SIZE,
280 				    DATA_TYPE_UINT64, size, NULL);
281 			else
282 				fm_payload_set(ereport,
283 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_OFFSET,
284 				    DATA_TYPE_UINT64, zio->io_offset,
285 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_SIZE,
286 				    DATA_TYPE_UINT64, zio->io_size, NULL);
287 		}
288 
289 		/*
290 		 * Payload for I/Os with corresponding logical information.
291 		 */
292 		if (zio->io_logical != NULL)
293 			fm_payload_set(ereport,
294 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_OBJSET,
295 			    DATA_TYPE_UINT64,
296 			    zio->io_logical->io_bookmark.zb_objset,
297 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_OBJECT,
298 			    DATA_TYPE_UINT64,
299 			    zio->io_logical->io_bookmark.zb_object,
300 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_LEVEL,
301 			    DATA_TYPE_INT64,
302 			    zio->io_logical->io_bookmark.zb_level,
303 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_BLKID,
304 			    DATA_TYPE_UINT64,
305 			    zio->io_logical->io_bookmark.zb_blkid, NULL);
306 	} else if (vd != NULL) {
307 		/*
308 		 * If we have a vdev but no zio, this is a device fault, and the
309 		 * 'stateoroffset' parameter indicates the previous state of the
310 		 * vdev.
311 		 */
312 		fm_payload_set(ereport,
313 		    FM_EREPORT_PAYLOAD_ZFS_PREV_STATE,
314 		    DATA_TYPE_UINT64, stateoroffset, NULL);
315 	}
316 	mutex_exit(&spa->spa_errlist_lock);
317 
318 	fm_ereport_post(ereport, EVCH_SLEEP);
319 
320 	fm_nvlist_destroy(ereport, FM_NVA_FREE);
321 	fm_nvlist_destroy(detector, FM_NVA_FREE);
322 #endif
323 }
324 
325 static void
326 zfs_post_common(spa_t *spa, vdev_t *vd, const char *name)
327 {
328 #ifdef _KERNEL
329 	nvlist_t *resource;
330 	char class[64];
331 
332 	if ((resource = fm_nvlist_create(NULL)) == NULL)
333 		return;
334 
335 	(void) snprintf(class, sizeof (class), "%s.%s.%s", FM_RSRC_RESOURCE,
336 	    ZFS_ERROR_CLASS, name);
337 	VERIFY(nvlist_add_uint8(resource, FM_VERSION, FM_RSRC_VERSION) == 0);
338 	VERIFY(nvlist_add_string(resource, FM_CLASS, class) == 0);
339 	VERIFY(nvlist_add_uint64(resource,
340 	    FM_EREPORT_PAYLOAD_ZFS_POOL_GUID, spa_guid(spa)) == 0);
341 	if (vd)
342 		VERIFY(nvlist_add_uint64(resource,
343 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID, vd->vdev_guid) == 0);
344 
345 	fm_ereport_post(resource, EVCH_SLEEP);
346 
347 	fm_nvlist_destroy(resource, FM_NVA_FREE);
348 #endif
349 }
350 
351 /*
352  * The 'resource.fs.zfs.removed' event is an internal signal that the given vdev
353  * has been removed from the system.  This will cause the DE to ignore any
354  * recent I/O errors, inferring that they are due to the asynchronous device
355  * removal.
356  */
357 void
358 zfs_post_remove(spa_t *spa, vdev_t *vd)
359 {
360 	zfs_post_common(spa, vd, FM_RESOURCE_REMOVED);
361 }
362 
363 /*
364  * The 'resource.fs.zfs.autoreplace' event is an internal signal that the pool
365  * has the 'autoreplace' property set, and therefore any broken vdevs will be
366  * handled by higher level logic, and no vdev fault should be generated.
367  */
368 void
369 zfs_post_autoreplace(spa_t *spa, vdev_t *vd)
370 {
371 	zfs_post_common(spa, vd, FM_RESOURCE_AUTOREPLACE);
372 }
373