xref: /illumos-gate/usr/src/uts/common/fs/zfs/spa_history.c (revision 4b22b9337f359bfd063322244f5336cc7c6ffcfa)
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.h>
30 #include <sys/spa_impl.h>
31 #include <sys/zap.h>
32 #include <sys/dsl_synctask.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/utsname.h>
36 #include <sys/cmn_err.h>
37 #include <sys/sunddi.h>
38 #ifdef _KERNEL
39 #include <sys/zone.h>
40 #endif
41 
42 /*
43  * Routines to manage the on-disk history log.
44  *
45  * The history log is stored as a dmu object containing
46  * <packed record length, record nvlist> tuples.
47  *
48  * Where "record nvlist" is a nvlist containing uint64_ts and strings, and
49  * "packed record length" is the packed length of the "record nvlist" stored
50  * as a little endian uint64_t.
51  *
52  * The log is implemented as a ring buffer, though the original creation
53  * of the pool ('zpool create') is never overwritten.
54  *
55  * The history log is tracked as object 'spa_t::spa_history'.  The bonus buffer
56  * of 'spa_history' stores the offsets for logging/retrieving history as
57  * 'spa_history_phys_t'.  'sh_pool_create_len' is the ending offset in bytes of
58  * where the 'zpool create' record is stored.  This allows us to never
59  * overwrite the original creation of the pool.  'sh_phys_max_off' is the
60  * physical ending offset in bytes of the log.  This tells you the length of
61  * the buffer. 'sh_eof' is the logical EOF (in bytes).  Whenever a record
62  * is added, 'sh_eof' is incremented by the the size of the record.
63  * 'sh_eof' is never decremented.  'sh_bof' is the logical BOF (in bytes).
64  * This is where the consumer should start reading from after reading in
65  * the 'zpool create' portion of the log.
66  *
67  * 'sh_records_lost' keeps track of how many records have been overwritten
68  * and permanently lost.
69  */
70 
71 /* convert a logical offset to physical */
72 static uint64_t
73 spa_history_log_to_phys(uint64_t log_off, spa_history_phys_t *shpp)
74 {
75 	uint64_t phys_len;
76 
77 	phys_len = shpp->sh_phys_max_off - shpp->sh_pool_create_len;
78 	return ((log_off - shpp->sh_pool_create_len) % phys_len
79 	    + shpp->sh_pool_create_len);
80 }
81 
82 void
83 spa_history_create_obj(spa_t *spa, dmu_tx_t *tx)
84 {
85 	dmu_buf_t *dbp;
86 	spa_history_phys_t *shpp;
87 	objset_t *mos = spa->spa_meta_objset;
88 
89 	ASSERT(spa->spa_history == 0);
90 	spa->spa_history = dmu_object_alloc(mos, DMU_OT_SPA_HISTORY,
91 	    SPA_MAXBLOCKSIZE, DMU_OT_SPA_HISTORY_OFFSETS,
92 	    sizeof (spa_history_phys_t), tx);
93 
94 	VERIFY(zap_add(mos, DMU_POOL_DIRECTORY_OBJECT,
95 	    DMU_POOL_HISTORY, sizeof (uint64_t), 1,
96 	    &spa->spa_history, tx) == 0);
97 
98 	VERIFY(0 == dmu_bonus_hold(mos, spa->spa_history, FTAG, &dbp));
99 	ASSERT(dbp->db_size >= sizeof (spa_history_phys_t));
100 
101 	shpp = dbp->db_data;
102 	dmu_buf_will_dirty(dbp, tx);
103 
104 	/*
105 	 * Figure out maximum size of history log.  We set it at
106 	 * 1% of pool size, with a max of 32MB and min of 128KB.
107 	 */
108 	shpp->sh_phys_max_off = spa_get_dspace(spa) / 100;
109 	shpp->sh_phys_max_off = MIN(shpp->sh_phys_max_off, 32<<20);
110 	shpp->sh_phys_max_off = MAX(shpp->sh_phys_max_off, 128<<10);
111 
112 	dmu_buf_rele(dbp, FTAG);
113 }
114 
115 /*
116  * Change 'sh_bof' to the beginning of the next record.
117  */
118 static int
119 spa_history_advance_bof(spa_t *spa, spa_history_phys_t *shpp)
120 {
121 	objset_t *mos = spa->spa_meta_objset;
122 	uint64_t firstread, reclen, phys_bof;
123 	char buf[sizeof (reclen)];
124 	int err;
125 
126 	phys_bof = spa_history_log_to_phys(shpp->sh_bof, shpp);
127 	firstread = MIN(sizeof (reclen), shpp->sh_phys_max_off - phys_bof);
128 
129 	if ((err = dmu_read(mos, spa->spa_history, phys_bof, firstread,
130 	    buf)) != 0)
131 		return (err);
132 	if (firstread != sizeof (reclen)) {
133 		if ((err = dmu_read(mos, spa->spa_history,
134 		    shpp->sh_pool_create_len, sizeof (reclen) - firstread,
135 		    buf + firstread)) != 0)
136 			return (err);
137 	}
138 
139 	reclen = LE_64(*((uint64_t *)buf));
140 	shpp->sh_bof += reclen + sizeof (reclen);
141 	shpp->sh_records_lost++;
142 	return (0);
143 }
144 
145 static int
146 spa_history_write(spa_t *spa, void *buf, uint64_t len, spa_history_phys_t *shpp,
147     dmu_tx_t *tx)
148 {
149 	uint64_t firstwrite, phys_eof;
150 	objset_t *mos = spa->spa_meta_objset;
151 	int err;
152 
153 	ASSERT(MUTEX_HELD(&spa->spa_history_lock));
154 
155 	/* see if we need to reset logical BOF */
156 	while (shpp->sh_phys_max_off - shpp->sh_pool_create_len -
157 	    (shpp->sh_eof - shpp->sh_bof) <= len) {
158 		if ((err = spa_history_advance_bof(spa, shpp)) != 0) {
159 			return (err);
160 		}
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 static char *
179 spa_history_zone()
180 {
181 #ifdef _KERNEL
182 	return (curproc->p_zone->zone_name);
183 #else
184 	return ("global");
185 #endif
186 }
187 
188 /*
189  * Write out a history event.
190  */
191 static void
192 spa_history_log_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
193 {
194 	spa_t		*spa = arg1;
195 	history_arg_t	*hap = arg2;
196 	const char	*history_str = hap->ha_history_str;
197 	objset_t	*mos = spa->spa_meta_objset;
198 	dmu_buf_t	*dbp;
199 	spa_history_phys_t *shpp;
200 	size_t		reclen;
201 	uint64_t	le_len;
202 	nvlist_t	*nvrecord;
203 	char		*record_packed = NULL;
204 	int		ret;
205 
206 	/*
207 	 * If we have an older pool that doesn't have a command
208 	 * history object, create it now.
209 	 */
210 	mutex_enter(&spa->spa_history_lock);
211 	if (!spa->spa_history)
212 		spa_history_create_obj(spa, tx);
213 	mutex_exit(&spa->spa_history_lock);
214 
215 	/*
216 	 * Get the offset of where we need to write via the bonus buffer.
217 	 * Update the offset when the write completes.
218 	 */
219 	VERIFY(0 == dmu_bonus_hold(mos, spa->spa_history, FTAG, &dbp));
220 	shpp = dbp->db_data;
221 
222 	dmu_buf_will_dirty(dbp, tx);
223 
224 #ifdef ZFS_DEBUG
225 	{
226 		dmu_object_info_t doi;
227 		dmu_object_info_from_db(dbp, &doi);
228 		ASSERT3U(doi.doi_bonus_type, ==, DMU_OT_SPA_HISTORY_OFFSETS);
229 	}
230 #endif
231 
232 	VERIFY(nvlist_alloc(&nvrecord, NV_UNIQUE_NAME, KM_SLEEP) == 0);
233 	VERIFY(nvlist_add_uint64(nvrecord, ZPOOL_HIST_TIME,
234 	    gethrestime_sec()) == 0);
235 	VERIFY(nvlist_add_uint64(nvrecord, ZPOOL_HIST_WHO,
236 	    (uint64_t)crgetuid(cr)) == 0);
237 	if (hap->ha_zone[0] != '\0')
238 		VERIFY(nvlist_add_string(nvrecord, ZPOOL_HIST_ZONE,
239 		    hap->ha_zone) == 0);
240 #ifdef _KERNEL
241 	VERIFY(nvlist_add_string(nvrecord, ZPOOL_HIST_HOST,
242 	    utsname.nodename) == 0);
243 #endif
244 	if (hap->ha_log_type == LOG_CMD_POOL_CREATE ||
245 	    hap->ha_log_type == LOG_CMD_NORMAL) {
246 		VERIFY(nvlist_add_string(nvrecord, ZPOOL_HIST_CMD,
247 		    history_str) == 0);
248 	} else {
249 		VERIFY(nvlist_add_uint64(nvrecord, ZPOOL_HIST_INT_EVENT,
250 		    hap->ha_event) == 0);
251 		VERIFY(nvlist_add_uint64(nvrecord, ZPOOL_HIST_TXG,
252 		    tx->tx_txg) == 0);
253 		VERIFY(nvlist_add_string(nvrecord, ZPOOL_HIST_INT_STR,
254 		    history_str) == 0);
255 	}
256 
257 	VERIFY(nvlist_size(nvrecord, &reclen, NV_ENCODE_XDR) == 0);
258 	record_packed = kmem_alloc(reclen, KM_SLEEP);
259 
260 	VERIFY(nvlist_pack(nvrecord, &record_packed, &reclen,
261 	    NV_ENCODE_XDR, KM_SLEEP) == 0);
262 
263 	mutex_enter(&spa->spa_history_lock);
264 	if (hap->ha_log_type == LOG_CMD_POOL_CREATE)
265 		VERIFY(shpp->sh_eof == shpp->sh_pool_create_len);
266 
267 	/* write out the packed length as little endian */
268 	le_len = LE_64((uint64_t)reclen);
269 	ret = spa_history_write(spa, &le_len, sizeof (le_len), shpp, tx);
270 	if (!ret)
271 		ret = spa_history_write(spa, record_packed, reclen, shpp, tx);
272 
273 	if (!ret && hap->ha_log_type == LOG_CMD_POOL_CREATE) {
274 		shpp->sh_pool_create_len += sizeof (le_len) + reclen;
275 		shpp->sh_bof = shpp->sh_pool_create_len;
276 	}
277 
278 	mutex_exit(&spa->spa_history_lock);
279 	nvlist_free(nvrecord);
280 	kmem_free(record_packed, reclen);
281 	dmu_buf_rele(dbp, FTAG);
282 
283 	if (hap->ha_log_type == LOG_INTERNAL) {
284 		kmem_free((void*)hap->ha_history_str, HIS_MAX_RECORD_LEN);
285 		kmem_free(hap, sizeof (history_arg_t));
286 	}
287 }
288 
289 /*
290  * Write out a history event.
291  */
292 int
293 spa_history_log(spa_t *spa, const char *history_str, history_log_type_t what)
294 {
295 	history_arg_t ha;
296 
297 	ASSERT(what != LOG_INTERNAL);
298 
299 	ha.ha_history_str = history_str;
300 	ha.ha_log_type = what;
301 	(void) strlcpy(ha.ha_zone, spa_history_zone(), sizeof (ha.ha_zone));
302 	return (dsl_sync_task_do(spa_get_dsl(spa), NULL, spa_history_log_sync,
303 	    spa, &ha, 0));
304 }
305 
306 /*
307  * Read out the command history.
308  */
309 int
310 spa_history_get(spa_t *spa, uint64_t *offp, uint64_t *len, char *buf)
311 {
312 	objset_t *mos = spa->spa_meta_objset;
313 	dmu_buf_t *dbp;
314 	uint64_t read_len, phys_read_off, phys_eof;
315 	uint64_t leftover = 0;
316 	spa_history_phys_t *shpp;
317 	int err;
318 
319 	/*
320 	 * If the command history  doesn't exist (older pool),
321 	 * that's ok, just return ENOENT.
322 	 */
323 	if (!spa->spa_history)
324 		return (ENOENT);
325 
326 	if ((err = dmu_bonus_hold(mos, spa->spa_history, FTAG, &dbp)) != 0)
327 		return (err);
328 	shpp = dbp->db_data;
329 
330 #ifdef ZFS_DEBUG
331 	{
332 		dmu_object_info_t doi;
333 		dmu_object_info_from_db(dbp, &doi);
334 		ASSERT3U(doi.doi_bonus_type, ==, DMU_OT_SPA_HISTORY_OFFSETS);
335 	}
336 #endif
337 
338 	mutex_enter(&spa->spa_history_lock);
339 	phys_eof = spa_history_log_to_phys(shpp->sh_eof, shpp);
340 
341 	if (*offp < shpp->sh_pool_create_len) {
342 		/* read in just the zpool create history */
343 		phys_read_off = *offp;
344 		read_len = MIN(*len, shpp->sh_pool_create_len -
345 		    phys_read_off);
346 	} else {
347 		/*
348 		 * Need to reset passed in offset to BOF if the passed in
349 		 * offset has since been overwritten.
350 		 */
351 		*offp = MAX(*offp, shpp->sh_bof);
352 		phys_read_off = spa_history_log_to_phys(*offp, shpp);
353 
354 		/*
355 		 * Read up to the minimum of what the user passed down or
356 		 * the EOF (physical or logical).  If we hit physical EOF,
357 		 * use 'leftover' to read from the physical BOF.
358 		 */
359 		if (phys_read_off <= phys_eof) {
360 			read_len = MIN(*len, phys_eof - phys_read_off);
361 		} else {
362 			read_len = MIN(*len,
363 			    shpp->sh_phys_max_off - phys_read_off);
364 			if (phys_read_off + *len > shpp->sh_phys_max_off) {
365 				leftover = MIN(*len - read_len,
366 				    phys_eof - shpp->sh_pool_create_len);
367 			}
368 		}
369 	}
370 
371 	/* offset for consumer to use next */
372 	*offp += read_len + leftover;
373 
374 	/* tell the consumer how much you actually read */
375 	*len = read_len + leftover;
376 
377 	if (read_len == 0) {
378 		mutex_exit(&spa->spa_history_lock);
379 		dmu_buf_rele(dbp, FTAG);
380 		return (0);
381 	}
382 
383 	err = dmu_read(mos, spa->spa_history, phys_read_off, read_len, buf);
384 	if (leftover && err == 0) {
385 		err = dmu_read(mos, spa->spa_history, shpp->sh_pool_create_len,
386 		    leftover, buf + read_len);
387 	}
388 	mutex_exit(&spa->spa_history_lock);
389 
390 	dmu_buf_rele(dbp, FTAG);
391 	return (err);
392 }
393 
394 void
395 spa_history_internal_log(history_internal_events_t event, spa_t *spa,
396     dmu_tx_t *tx, cred_t *cr, const char *fmt, ...)
397 {
398 	history_arg_t *hap;
399 	char *str;
400 	va_list adx;
401 
402 	hap = kmem_alloc(sizeof (history_arg_t), KM_SLEEP);
403 	str = kmem_alloc(HIS_MAX_RECORD_LEN, KM_SLEEP);
404 
405 	va_start(adx, fmt);
406 	(void) vsnprintf(str, HIS_MAX_RECORD_LEN, fmt, adx);
407 	va_end(adx);
408 
409 	hap->ha_log_type = LOG_INTERNAL;
410 	hap->ha_history_str = str;
411 	hap->ha_event = event;
412 	hap->ha_zone[0] = '\0';
413 
414 	if (dmu_tx_is_syncing(tx)) {
415 		spa_history_log_sync(spa, hap, cr, tx);
416 	} else {
417 		dsl_sync_task_do_nowait(spa_get_dsl(spa), NULL,
418 		    spa_history_log_sync, spa, hap, 0, tx);
419 	}
420 	/* spa_history_log_sync() will free hap and str */
421 }
422