xref: /illumos-gate/usr/src/uts/common/fs/zfs/spa_history.c (revision ef8846857fcf954444cdc77e72249afef48377d2)
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 /*
23  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/spa_impl.h>
30 #include <sys/zap.h>
31 #include <sys/dsl_synctask.h>
32 
33 /*
34  * Routines to manage the on-disk history log.
35  *
36  * The history log is stored as a dmu object containing
37  * <packed record length, record nvlist> tuples.
38  *
39  * Where "record nvlist" is a nvlist containing uint64_ts and strings, and
40  * "packed record length" is the packed length of the "record nvlist" stored
41  * as a little endian uint64_t.
42  *
43  * The log is implemented as a ring buffer, though the original creation
44  * of the pool ('zpool create') is never overwritten.
45  *
46  * The history log is tracked as object 'spa_t::spa_history'.  The bonus buffer
47  * of 'spa_history' stores the offsets for logging/retrieving history as
48  * 'spa_history_phys_t'.  'sh_pool_create_len' is the ending offset in bytes of
49  * where the 'zpool create' record is stored.  This allows us to never
50  * overwrite the original creation of the pool.  'sh_phys_max_off' is the
51  * physical ending offset in bytes of the log.  This tells you the length of
52  * the buffer. 'sh_eof' is the logical EOF (in bytes).  Whenever a record
53  * is added, 'sh_eof' is incremented by the the size of the record.
54  * 'sh_eof' is never decremented.  'sh_bof' is the logical BOF (in bytes).
55  * This is where the consumer should start reading from after reading in
56  * the 'zpool create' portion of the log.
57  *
58  * 'sh_records_lost' keeps track of how many records have been overwritten
59  * and permanently lost.
60  */
61 
62 typedef enum history_log_type {
63 	LOG_CMD_CREATE,
64 	LOG_CMD_NO_CREATE
65 } history_log_type_t;
66 
67 typedef struct history_arg {
68 	const char *ha_history_str;
69 	history_log_type_t ha_log_type;
70 } history_arg_t;
71 
72 /* convert a logical offset to physical */
73 static uint64_t
74 spa_history_log_to_phys(uint64_t log_off, spa_history_phys_t *shpp)
75 {
76 	uint64_t phys_len;
77 
78 	phys_len = shpp->sh_phys_max_off - shpp->sh_pool_create_len;
79 	return ((log_off - shpp->sh_pool_create_len) % phys_len
80 	    + shpp->sh_pool_create_len);
81 }
82 
83 void
84 spa_history_create_obj(spa_t *spa, dmu_tx_t *tx)
85 {
86 	dmu_buf_t *dbp;
87 	spa_history_phys_t *shpp;
88 	objset_t *mos = spa->spa_meta_objset;
89 
90 	ASSERT(spa->spa_history == 0);
91 	spa->spa_history = dmu_object_alloc(mos, DMU_OT_SPA_HISTORY,
92 	    SPA_MAXBLOCKSIZE, DMU_OT_SPA_HISTORY_OFFSETS,
93 	    sizeof (spa_history_phys_t), tx);
94 
95 	VERIFY(zap_add(mos, DMU_POOL_DIRECTORY_OBJECT,
96 	    DMU_POOL_HISTORY, sizeof (uint64_t), 1,
97 	    &spa->spa_history, tx) == 0);
98 
99 	VERIFY(0 == dmu_bonus_hold(mos, spa->spa_history, FTAG, &dbp));
100 	ASSERT(dbp->db_size >= sizeof (spa_history_phys_t));
101 
102 	shpp = dbp->db_data;
103 	dmu_buf_will_dirty(dbp, tx);
104 
105 	/*
106 	 * Figure out maximum size of history log.  We set it at
107 	 * 1% of pool size, with a max of 32MB and min of 128KB.
108 	 */
109 	shpp->sh_phys_max_off = spa_get_dspace(spa) / 100;
110 	shpp->sh_phys_max_off = MIN(shpp->sh_phys_max_off, 32<<20);
111 	shpp->sh_phys_max_off = MAX(shpp->sh_phys_max_off, 128<<10);
112 
113 	dmu_buf_rele(dbp, FTAG);
114 }
115 
116 /*
117  * Change 'sh_bof' to the beginning of the next record.
118  */
119 static int
120 spa_history_advance_bof(spa_t *spa, spa_history_phys_t *shpp)
121 {
122 	objset_t *mos = spa->spa_meta_objset;
123 	uint64_t firstread, reclen, phys_bof;
124 	char buf[sizeof (reclen)];
125 	int err;
126 
127 	phys_bof = spa_history_log_to_phys(shpp->sh_bof, shpp);
128 	firstread = MIN(sizeof (reclen), shpp->sh_phys_max_off - phys_bof);
129 
130 	if ((err = dmu_read(mos, spa->spa_history, phys_bof, firstread,
131 	    buf)) != 0)
132 		return (err);
133 	if (firstread != sizeof (reclen)) {
134 		if ((err = dmu_read(mos, spa->spa_history,
135 		    shpp->sh_pool_create_len, sizeof (reclen) - firstread,
136 		    buf + firstread)) != 0)
137 			return (err);
138 	}
139 
140 	reclen = LE_64(*((uint64_t *)buf));
141 	shpp->sh_bof += reclen + sizeof (reclen);
142 	shpp->sh_records_lost++;
143 	return (0);
144 }
145 
146 static int
147 spa_history_write(spa_t *spa, void *buf, uint64_t len, spa_history_phys_t *shpp,
148     dmu_tx_t *tx)
149 {
150 	uint64_t firstwrite, phys_eof;
151 	objset_t *mos = spa->spa_meta_objset;
152 	int err;
153 
154 	ASSERT(MUTEX_HELD(&spa->spa_history_lock));
155 
156 	/* see if we need to reset logical BOF */
157 	while (shpp->sh_phys_max_off - shpp->sh_pool_create_len -
158 	    (shpp->sh_eof - shpp->sh_bof) <= len) {
159 		if ((err = spa_history_advance_bof(spa, shpp)) != 0)
160 			return (err);
161 	}
162 
163 	phys_eof = spa_history_log_to_phys(shpp->sh_eof, shpp);
164 	firstwrite = MIN(len, shpp->sh_phys_max_off - phys_eof);
165 	shpp->sh_eof += len;
166 	dmu_write(mos, spa->spa_history, phys_eof, firstwrite, buf, tx);
167 
168 	len -= firstwrite;
169 	if (len > 0) {
170 		/* write out the rest at the beginning of physical file */
171 		dmu_write(mos, spa->spa_history, shpp->sh_pool_create_len,
172 		    len, (char *)buf + firstwrite, tx);
173 	}
174 
175 	return (0);
176 }
177 
178 /*
179  * Write out a history event.
180  */
181 void
182 spa_history_log_sync(void *arg1, void *arg2, dmu_tx_t *tx)
183 {
184 	spa_t		*spa = arg1;
185 	history_arg_t	*hap = arg2;
186 	const char	*history_str = hap->ha_history_str;
187 	objset_t	*mos = spa->spa_meta_objset;
188 	dmu_buf_t	*dbp;
189 	spa_history_phys_t *shpp;
190 	size_t		reclen;
191 	uint64_t	le_len;
192 	nvlist_t	*nvrecord;
193 	char		*record_packed = NULL;
194 	int		ret;
195 
196 	if (history_str == NULL)
197 		return;
198 
199 	/*
200 	 * If we have an older pool that doesn't have a command
201 	 * history object, create it now.
202 	 */
203 	mutex_enter(&spa->spa_history_lock);
204 	if (!spa->spa_history)
205 		spa_history_create_obj(spa, tx);
206 	mutex_exit(&spa->spa_history_lock);
207 
208 	/*
209 	 * Get the offset of where we need to write via the bonus buffer.
210 	 * Update the offset when the write completes.
211 	 */
212 	VERIFY(0 == dmu_bonus_hold(mos, spa->spa_history, FTAG, &dbp));
213 	shpp = dbp->db_data;
214 
215 	dmu_buf_will_dirty(dbp, tx);
216 
217 #ifdef ZFS_DEBUG
218 	{
219 		dmu_object_info_t doi;
220 		dmu_object_info_from_db(dbp, &doi);
221 		ASSERT3U(doi.doi_bonus_type, ==, DMU_OT_SPA_HISTORY_OFFSETS);
222 	}
223 #endif
224 
225 	/* construct a nvlist of the current time and cmd string */
226 	VERIFY(nvlist_alloc(&nvrecord, NV_UNIQUE_NAME, KM_SLEEP) == 0);
227 	VERIFY(nvlist_add_uint64(nvrecord, ZPOOL_HIST_TIME,
228 	    gethrestime_sec()) == 0);
229 	VERIFY(nvlist_add_string(nvrecord, ZPOOL_HIST_CMD, history_str) == 0);
230 	VERIFY(nvlist_pack(nvrecord, &record_packed, &reclen,
231 	    NV_ENCODE_XDR, KM_SLEEP) == 0);
232 
233 	mutex_enter(&spa->spa_history_lock);
234 	if (hap->ha_log_type == LOG_CMD_CREATE)
235 		VERIFY(shpp->sh_eof == shpp->sh_pool_create_len);
236 
237 	/* write out the packed length as little endian */
238 	le_len = LE_64((uint64_t)reclen);
239 	ret = spa_history_write(spa, &le_len, sizeof (le_len), shpp, tx);
240 	if (!ret)
241 		ret = spa_history_write(spa, record_packed, reclen, shpp, tx);
242 
243 	if (!ret && hap->ha_log_type == LOG_CMD_CREATE) {
244 		shpp->sh_pool_create_len += sizeof (le_len) + reclen;
245 		shpp->sh_bof = shpp->sh_pool_create_len;
246 	}
247 
248 	mutex_exit(&spa->spa_history_lock);
249 	nvlist_free(nvrecord);
250 	kmem_free(record_packed, reclen);
251 	dmu_buf_rele(dbp, FTAG);
252 }
253 
254 /*
255  * Write out a history event.
256  */
257 int
258 spa_history_log(spa_t *spa, const char *history_str, uint64_t pool_create)
259 {
260 	history_arg_t ha;
261 
262 	ha.ha_history_str = history_str;
263 	ha.ha_log_type = pool_create ? LOG_CMD_CREATE : LOG_CMD_NO_CREATE;
264 	return (dsl_sync_task_do(spa_get_dsl(spa), NULL, spa_history_log_sync,
265 	    spa, &ha, 0));
266 }
267 
268 /*
269  * Read out the command history.
270  */
271 int
272 spa_history_get(spa_t *spa, uint64_t *offp, uint64_t *len, char *buf)
273 {
274 	objset_t *mos = spa->spa_meta_objset;
275 	dmu_buf_t *dbp;
276 	uint64_t read_len, phys_read_off, phys_eof;
277 	uint64_t leftover = 0;
278 	spa_history_phys_t *shpp;
279 	int err;
280 
281 	/*
282 	 * If the command history  doesn't exist (older pool),
283 	 * that's ok, just return ENOENT.
284 	 */
285 	if (!spa->spa_history)
286 		return (ENOENT);
287 
288 	if ((err = dmu_bonus_hold(mos, spa->spa_history, FTAG, &dbp)) != 0)
289 		return (err);
290 	shpp = dbp->db_data;
291 
292 #ifdef ZFS_DEBUG
293 	{
294 		dmu_object_info_t doi;
295 		dmu_object_info_from_db(dbp, &doi);
296 		ASSERT3U(doi.doi_bonus_type, ==, DMU_OT_SPA_HISTORY_OFFSETS);
297 	}
298 #endif
299 
300 	mutex_enter(&spa->spa_history_lock);
301 	phys_eof = spa_history_log_to_phys(shpp->sh_eof, shpp);
302 
303 	if (*offp < shpp->sh_pool_create_len) {
304 		/* read in just the zpool create history */
305 		phys_read_off = *offp;
306 		read_len = MIN(*len, shpp->sh_pool_create_len -
307 		    phys_read_off);
308 	} else {
309 		/*
310 		 * Need to reset passed in offset to BOF if the passed in
311 		 * offset has since been overwritten.
312 		 */
313 		*offp = MAX(*offp, shpp->sh_bof);
314 		phys_read_off = spa_history_log_to_phys(*offp, shpp);
315 
316 		/*
317 		 * Read up to the minimum of what the user passed down or
318 		 * the EOF (physical or logical).  If we hit physical EOF,
319 		 * use 'leftover' to read from the physical BOF.
320 		 */
321 		if (phys_read_off <= phys_eof) {
322 			read_len = MIN(*len, phys_eof - phys_read_off);
323 		} else {
324 			read_len = MIN(*len,
325 			    shpp->sh_phys_max_off - phys_read_off);
326 			if (phys_read_off + *len > shpp->sh_phys_max_off) {
327 				leftover = MIN(*len - read_len,
328 				    phys_eof - shpp->sh_pool_create_len);
329 			}
330 		}
331 	}
332 
333 	/* offset for consumer to use next */
334 	*offp += read_len + leftover;
335 
336 	/* tell the consumer how much you actually read */
337 	*len = read_len + leftover;
338 
339 	if (read_len == 0) {
340 		mutex_exit(&spa->spa_history_lock);
341 		dmu_buf_rele(dbp, FTAG);
342 		return (0);
343 	}
344 
345 	err = dmu_read(mos, spa->spa_history, phys_read_off, read_len, buf);
346 	if (leftover && err == 0) {
347 		err = dmu_read(mos, spa->spa_history, shpp->sh_pool_create_len,
348 		    leftover, buf + read_len);
349 	}
350 	mutex_exit(&spa->spa_history_lock);
351 
352 	dmu_buf_rele(dbp, FTAG);
353 	return (err);
354 }
355