xref: /titanic_44/usr/src/uts/common/fs/zfs/zfs_fm.c (revision 68c92b9f236192fc4b1af4355a207acf7469bd59)
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 2007 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.
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 
103 	/*
104 	 * If we are doing a spa_tryimport(), ignore errors.
105 	 */
106 	if (spa->spa_load_state == SPA_LOAD_TRYIMPORT)
107 		return;
108 
109 	/*
110 	 * If we are in the middle of opening a pool, and the previous attempt
111 	 * failed, don't bother logging any new ereports - we're just going to
112 	 * get the same diagnosis anyway.
113 	 */
114 	if (spa->spa_load_state != SPA_LOAD_NONE &&
115 	    spa->spa_last_open_failed)
116 		return;
117 
118 	/*
119 	 * Ignore any errors from I/Os that we are going to retry anyway - we
120 	 * only generate errors from the final failure.  Checksum errors are
121 	 * generated after the pipeline stage responsible for retrying the I/O
122 	 * (VDEV_IO_ASSESS), so this only applies to standard I/O errors.
123 	 */
124 	if (zio && zio_should_retry(zio) && zio->io_error != ECKSUM)
125 		return;
126 
127 	/*
128 	 * If this is not a read or write zio, ignore the error.  This can occur
129 	 * if the DKIOCFLUSHWRITECACHE ioctl fails.
130 	 */
131 	if (zio && zio->io_type != ZIO_TYPE_READ &&
132 	    zio->io_type != ZIO_TYPE_WRITE)
133 		return;
134 
135 	if ((ereport = fm_nvlist_create(NULL)) == NULL)
136 		return;
137 
138 	if ((detector = fm_nvlist_create(NULL)) == NULL) {
139 		fm_nvlist_destroy(ereport, FM_NVA_FREE);
140 		return;
141 	}
142 
143 	/*
144 	 * Serialize ereport generation
145 	 */
146 	mutex_enter(&spa->spa_errlist_lock);
147 
148 	/*
149 	 * Determine the ENA to use for this event.  If we are in a loading
150 	 * state, use a SPA-wide ENA.  Otherwise, if we are in an I/O state, use
151 	 * a root zio-wide ENA.  Otherwise, simply use a unique ENA.
152 	 */
153 	if (spa->spa_load_state != SPA_LOAD_NONE) {
154 		if (spa->spa_ena == 0)
155 			spa->spa_ena = fm_ena_generate(0, FM_ENA_FMT1);
156 		ena = spa->spa_ena;
157 	} else if (zio != NULL && zio->io_logical != NULL) {
158 		if (zio->io_logical->io_ena == 0)
159 			zio->io_logical->io_ena =
160 			    fm_ena_generate(0, FM_ENA_FMT1);
161 		ena = zio->io_logical->io_ena;
162 	} else {
163 		ena = fm_ena_generate(0, FM_ENA_FMT1);
164 	}
165 
166 	/*
167 	 * Construct the full class, detector, and other standard FMA fields.
168 	 */
169 	(void) snprintf(class, sizeof (class), "%s.%s",
170 	    ZFS_ERROR_CLASS, subclass);
171 
172 	fm_fmri_zfs_set(detector, FM_ZFS_SCHEME_VERSION, spa_guid(spa),
173 	    vd != NULL ? vd->vdev_guid : 0);
174 
175 	fm_ereport_set(ereport, FM_EREPORT_VERSION, class, ena, detector, NULL);
176 
177 	/*
178 	 * Construct the per-ereport payload, depending on which parameters are
179 	 * passed in.
180 	 */
181 
182 	/*
183 	 * Generic payload members common to all ereports.
184 	 *
185 	 * The direct reference to spa_name is used rather than spa_name()
186 	 * because of the asynchronous nature of the zio pipeline.  spa_name()
187 	 * asserts that the config lock is held in some form.  This is always
188 	 * the case in I/O context, but because the check for RW_WRITER compares
189 	 * against 'curthread', we may be in an asynchronous context and blow
190 	 * this assert.  Rather than loosen this assert, we acknowledge that all
191 	 * contexts in which this function is called (pool open, I/O) are safe,
192 	 * and dereference the name directly.
193 	 */
194 	fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_POOL,
195 	    DATA_TYPE_STRING, spa->spa_name, FM_EREPORT_PAYLOAD_ZFS_POOL_GUID,
196 	    DATA_TYPE_UINT64, spa_guid(spa),
197 	    FM_EREPORT_PAYLOAD_ZFS_POOL_CONTEXT, DATA_TYPE_INT32,
198 	    spa->spa_load_state, NULL);
199 
200 	if (vd != NULL) {
201 		vdev_t *pvd = vd->vdev_parent;
202 
203 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID,
204 		    DATA_TYPE_UINT64, vd->vdev_guid,
205 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_TYPE,
206 		    DATA_TYPE_STRING, vd->vdev_ops->vdev_op_type, NULL);
207 		if (vd->vdev_path)
208 			fm_payload_set(ereport,
209 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_PATH,
210 			    DATA_TYPE_STRING, vd->vdev_path, NULL);
211 		if (vd->vdev_devid)
212 			fm_payload_set(ereport,
213 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_DEVID,
214 			    DATA_TYPE_STRING, vd->vdev_devid, NULL);
215 
216 		if (pvd != NULL) {
217 			fm_payload_set(ereport,
218 			    FM_EREPORT_PAYLOAD_ZFS_PARENT_GUID,
219 			    DATA_TYPE_UINT64, pvd->vdev_guid,
220 			    FM_EREPORT_PAYLOAD_ZFS_PARENT_TYPE,
221 			    DATA_TYPE_STRING, pvd->vdev_ops->vdev_op_type,
222 			    NULL);
223 			if (pvd->vdev_path)
224 				fm_payload_set(ereport,
225 				    FM_EREPORT_PAYLOAD_ZFS_PARENT_PATH,
226 				    DATA_TYPE_STRING, vd->vdev_path, NULL);
227 			if (pvd->vdev_devid)
228 				fm_payload_set(ereport,
229 				    FM_EREPORT_PAYLOAD_ZFS_PARENT_DEVID,
230 				    DATA_TYPE_STRING, pvd->vdev_devid, NULL);
231 		}
232 	}
233 
234 	if (zio != NULL) {
235 		/*
236 		 * Payload common to all I/Os.
237 		 */
238 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_ZIO_ERR,
239 		    DATA_TYPE_INT32, zio->io_error, NULL);
240 
241 		/*
242 		 * If the 'size' parameter is non-zero, it indicates this is a
243 		 * RAID-Z or other I/O where the physical offset and length are
244 		 * provided for us, instead of within the zio_t.
245 		 */
246 		if (vd != NULL) {
247 			if (size)
248 				fm_payload_set(ereport,
249 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_OFFSET,
250 				    DATA_TYPE_UINT64, stateoroffset,
251 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_SIZE,
252 				    DATA_TYPE_UINT64, size, NULL);
253 			else
254 				fm_payload_set(ereport,
255 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_OFFSET,
256 				    DATA_TYPE_UINT64, zio->io_offset,
257 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_SIZE,
258 				    DATA_TYPE_UINT64, zio->io_size, NULL);
259 		}
260 
261 		/*
262 		 * Payload for I/Os with corresponding logical information.
263 		 */
264 		if (zio->io_logical != NULL)
265 			fm_payload_set(ereport,
266 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_OBJSET,
267 			    DATA_TYPE_UINT64,
268 			    zio->io_logical->io_bookmark.zb_objset,
269 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_OBJECT,
270 			    DATA_TYPE_UINT64,
271 			    zio->io_logical->io_bookmark.zb_object,
272 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_LEVEL,
273 			    DATA_TYPE_INT32,
274 			    zio->io_logical->io_bookmark.zb_level,
275 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_BLKID,
276 			    DATA_TYPE_UINT64,
277 			    zio->io_logical->io_bookmark.zb_blkid, NULL);
278 	} else if (vd != NULL) {
279 		/*
280 		 * If we have a vdev but no zio, this is a device fault, and the
281 		 * 'stateoroffset' parameter indicates the previous state of the
282 		 * vdev.
283 		 */
284 		fm_payload_set(ereport,
285 		    FM_EREPORT_PAYLOAD_ZFS_PREV_STATE,
286 		    DATA_TYPE_UINT64, stateoroffset, NULL);
287 	}
288 	mutex_exit(&spa->spa_errlist_lock);
289 
290 	fm_ereport_post(ereport, EVCH_SLEEP);
291 
292 	fm_nvlist_destroy(ereport, FM_NVA_FREE);
293 	fm_nvlist_destroy(detector, FM_NVA_FREE);
294 #endif
295 }
296 
297 static void
298 zfs_post_common(spa_t *spa, vdev_t *vd, const char *name)
299 {
300 #ifdef _KERNEL
301 	nvlist_t *resource;
302 	char class[64];
303 
304 	if ((resource = fm_nvlist_create(NULL)) == NULL)
305 		return;
306 
307 	(void) snprintf(class, sizeof (class), "%s.%s.%s", FM_RSRC_RESOURCE,
308 	    ZFS_ERROR_CLASS, name);
309 	VERIFY(nvlist_add_uint8(resource, FM_VERSION, FM_RSRC_VERSION) == 0);
310 	VERIFY(nvlist_add_string(resource, FM_CLASS, class) == 0);
311 	VERIFY(nvlist_add_uint64(resource,
312 	    FM_EREPORT_PAYLOAD_ZFS_POOL_GUID, spa_guid(spa)) == 0);
313 	if (vd)
314 		VERIFY(nvlist_add_uint64(resource,
315 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID, vd->vdev_guid) == 0);
316 
317 	fm_ereport_post(resource, EVCH_SLEEP);
318 
319 	fm_nvlist_destroy(resource, FM_NVA_FREE);
320 #endif
321 }
322 
323 /*
324  * The 'resource.fs.zfs.ok' event is an internal signal that the associated
325  * resource (pool or disk) has been identified by ZFS as healthy.  This will
326  * then trigger the DE to close the associated case, if any.
327  */
328 void
329 zfs_post_ok(spa_t *spa, vdev_t *vd)
330 {
331 	zfs_post_common(spa, vd, FM_RESOURCE_OK);
332 }
333 
334 /*
335  * The 'resource.fs.zfs.removed' event is an internal signal that the given vdev
336  * has been removed from the system.  This will cause the DE to ignore any
337  * recent I/O errors, inferring that they are due to the asynchronous device
338  * removal.
339  */
340 void
341 zfs_post_remove(spa_t *spa, vdev_t *vd)
342 {
343 	zfs_post_common(spa, vd, FM_RESOURCE_REMOVED);
344 }
345 
346 /*
347  * The 'resource.fs.zfs.autoreplace' event is an internal signal that the pool
348  * has the 'autoreplace' property set, and therefore any broken vdevs will be
349  * handled by higher level logic, and no vdev fault should be generated.
350  */
351 void
352 zfs_post_autoreplace(spa_t *spa, vdev_t *vd)
353 {
354 	zfs_post_common(spa, vd, FM_RESOURCE_AUTOREPLACE);
355 }
356