xref: /freebsd/sys/contrib/openzfs/module/zfs/spa_stats.c (revision 7a7741af18d6c8a804cc643cb7ecda9d730c6aa6)
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 https://opensource.org/licenses/CDDL-1.0.
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 #include <sys/zfs_context.h>
23 #include <sys/spa_impl.h>
24 #include <sys/vdev_impl.h>
25 #include <sys/spa.h>
26 #include <zfs_comutil.h>
27 
28 /*
29  * Keeps stats on last N reads per spa_t, disabled by default.
30  */
31 static uint_t zfs_read_history = B_FALSE;
32 
33 /*
34  * Include cache hits in history, disabled by default.
35  */
36 static int zfs_read_history_hits = B_FALSE;
37 
38 /*
39  * Keeps stats on the last 100 txgs by default.
40  */
41 static uint_t zfs_txg_history = 100;
42 
43 /*
44  * Keeps stats on the last N MMP updates, disabled by default.
45  */
46 static uint_t zfs_multihost_history = B_FALSE;
47 
48 /*
49  * ==========================================================================
50  * SPA Read History Routines
51  * ==========================================================================
52  */
53 
54 /*
55  * Read statistics - Information exported regarding each arc_read call
56  */
57 typedef struct spa_read_history {
58 	hrtime_t	start;		/* time read completed */
59 	uint64_t	objset;		/* read from this objset */
60 	uint64_t	object;		/* read of this object number */
61 	uint64_t	level;		/* block's indirection level */
62 	uint64_t	blkid;		/* read of this block id */
63 	char		origin[24];	/* read originated from here */
64 	uint32_t	aflags;		/* ARC flags (cached, prefetch, etc.) */
65 	pid_t		pid;		/* PID of task doing read */
66 	char		comm[16];	/* process name of task doing read */
67 	procfs_list_node_t	srh_node;
68 } spa_read_history_t;
69 
70 static int
spa_read_history_show_header(struct seq_file * f)71 spa_read_history_show_header(struct seq_file *f)
72 {
73 	seq_printf(f, "%-8s %-16s %-8s %-8s %-8s %-8s %-8s "
74 	    "%-24s %-8s %-16s\n", "UID", "start", "objset", "object",
75 	    "level", "blkid", "aflags", "origin", "pid", "process");
76 
77 	return (0);
78 }
79 
80 static int
spa_read_history_show(struct seq_file * f,void * data)81 spa_read_history_show(struct seq_file *f, void *data)
82 {
83 	spa_read_history_t *srh = (spa_read_history_t *)data;
84 
85 	seq_printf(f, "%-8llu %-16llu 0x%-6llx "
86 	    "%-8lli %-8lli %-8lli 0x%-6x %-24s %-8i %-16s\n",
87 	    (u_longlong_t)srh->srh_node.pln_id, srh->start,
88 	    (longlong_t)srh->objset, (longlong_t)srh->object,
89 	    (longlong_t)srh->level, (longlong_t)srh->blkid,
90 	    srh->aflags, srh->origin, srh->pid, srh->comm);
91 
92 	return (0);
93 }
94 
95 /* Remove oldest elements from list until there are no more than 'size' left */
96 static void
spa_read_history_truncate(spa_history_list_t * shl,unsigned int size)97 spa_read_history_truncate(spa_history_list_t *shl, unsigned int size)
98 {
99 	spa_read_history_t *srh;
100 	while (shl->size > size) {
101 		srh = list_remove_head(&shl->procfs_list.pl_list);
102 		ASSERT3P(srh, !=, NULL);
103 		kmem_free(srh, sizeof (spa_read_history_t));
104 		shl->size--;
105 	}
106 
107 	if (size == 0)
108 		ASSERT(list_is_empty(&shl->procfs_list.pl_list));
109 }
110 
111 static int
spa_read_history_clear(procfs_list_t * procfs_list)112 spa_read_history_clear(procfs_list_t *procfs_list)
113 {
114 	spa_history_list_t *shl = procfs_list->pl_private;
115 	mutex_enter(&procfs_list->pl_lock);
116 	spa_read_history_truncate(shl, 0);
117 	mutex_exit(&procfs_list->pl_lock);
118 	return (0);
119 }
120 
121 static void
spa_read_history_init(spa_t * spa)122 spa_read_history_init(spa_t *spa)
123 {
124 	spa_history_list_t *shl = &spa->spa_stats.read_history;
125 
126 	shl->size = 0;
127 	shl->procfs_list.pl_private = shl;
128 	procfs_list_install("zfs",
129 	    spa_name(spa),
130 	    "reads",
131 	    0600,
132 	    &shl->procfs_list,
133 	    spa_read_history_show,
134 	    spa_read_history_show_header,
135 	    spa_read_history_clear,
136 	    offsetof(spa_read_history_t, srh_node));
137 }
138 
139 static void
spa_read_history_destroy(spa_t * spa)140 spa_read_history_destroy(spa_t *spa)
141 {
142 	spa_history_list_t *shl = &spa->spa_stats.read_history;
143 	procfs_list_uninstall(&shl->procfs_list);
144 	spa_read_history_truncate(shl, 0);
145 	procfs_list_destroy(&shl->procfs_list);
146 }
147 
148 void
spa_read_history_add(spa_t * spa,const zbookmark_phys_t * zb,uint32_t aflags)149 spa_read_history_add(spa_t *spa, const zbookmark_phys_t *zb, uint32_t aflags)
150 {
151 	spa_history_list_t *shl = &spa->spa_stats.read_history;
152 	spa_read_history_t *srh;
153 
154 	ASSERT3P(spa, !=, NULL);
155 	ASSERT3P(zb,  !=, NULL);
156 
157 	if (zfs_read_history == 0 && shl->size == 0)
158 		return;
159 
160 	if (zfs_read_history_hits == 0 && (aflags & ARC_FLAG_CACHED))
161 		return;
162 
163 	srh = kmem_zalloc(sizeof (spa_read_history_t), KM_SLEEP);
164 	strlcpy(srh->comm, getcomm(), sizeof (srh->comm));
165 	srh->start  = gethrtime();
166 	srh->objset = zb->zb_objset;
167 	srh->object = zb->zb_object;
168 	srh->level  = zb->zb_level;
169 	srh->blkid  = zb->zb_blkid;
170 	srh->aflags = aflags;
171 	srh->pid    = getpid();
172 
173 	mutex_enter(&shl->procfs_list.pl_lock);
174 
175 	procfs_list_add(&shl->procfs_list, srh);
176 	shl->size++;
177 
178 	spa_read_history_truncate(shl, zfs_read_history);
179 
180 	mutex_exit(&shl->procfs_list.pl_lock);
181 }
182 
183 /*
184  * ==========================================================================
185  * SPA TXG History Routines
186  * ==========================================================================
187  */
188 
189 /*
190  * Txg statistics - Information exported regarding each txg sync
191  */
192 
193 typedef struct spa_txg_history {
194 	uint64_t	txg;		/* txg id */
195 	txg_state_t	state;		/* active txg state */
196 	uint64_t	nread;		/* number of bytes read */
197 	uint64_t	nwritten;	/* number of bytes written */
198 	uint64_t	reads;		/* number of read operations */
199 	uint64_t	writes;		/* number of write operations */
200 	uint64_t	ndirty;		/* number of dirty bytes */
201 	hrtime_t	times[TXG_STATE_COMMITTED]; /* completion times */
202 	procfs_list_node_t	sth_node;
203 } spa_txg_history_t;
204 
205 static int
spa_txg_history_show_header(struct seq_file * f)206 spa_txg_history_show_header(struct seq_file *f)
207 {
208 	seq_printf(f, "%-8s %-16s %-5s %-12s %-12s %-12s "
209 	    "%-8s %-8s %-12s %-12s %-12s %-12s\n", "txg", "birth", "state",
210 	    "ndirty", "nread", "nwritten", "reads", "writes",
211 	    "otime", "qtime", "wtime", "stime");
212 	return (0);
213 }
214 
215 static int
spa_txg_history_show(struct seq_file * f,void * data)216 spa_txg_history_show(struct seq_file *f, void *data)
217 {
218 	spa_txg_history_t *sth = (spa_txg_history_t *)data;
219 	uint64_t open = 0, quiesce = 0, wait = 0, sync = 0;
220 	char state;
221 
222 	switch (sth->state) {
223 		case TXG_STATE_BIRTH:		state = 'B';	break;
224 		case TXG_STATE_OPEN:		state = 'O';	break;
225 		case TXG_STATE_QUIESCED:	state = 'Q';	break;
226 		case TXG_STATE_WAIT_FOR_SYNC:	state = 'W';	break;
227 		case TXG_STATE_SYNCED:		state = 'S';	break;
228 		case TXG_STATE_COMMITTED:	state = 'C';	break;
229 		default:			state = '?';	break;
230 	}
231 
232 	if (sth->times[TXG_STATE_OPEN])
233 		open = sth->times[TXG_STATE_OPEN] -
234 		    sth->times[TXG_STATE_BIRTH];
235 
236 	if (sth->times[TXG_STATE_QUIESCED])
237 		quiesce = sth->times[TXG_STATE_QUIESCED] -
238 		    sth->times[TXG_STATE_OPEN];
239 
240 	if (sth->times[TXG_STATE_WAIT_FOR_SYNC])
241 		wait = sth->times[TXG_STATE_WAIT_FOR_SYNC] -
242 		    sth->times[TXG_STATE_QUIESCED];
243 
244 	if (sth->times[TXG_STATE_SYNCED])
245 		sync = sth->times[TXG_STATE_SYNCED] -
246 		    sth->times[TXG_STATE_WAIT_FOR_SYNC];
247 
248 	seq_printf(f, "%-8llu %-16llu %-5c %-12llu "
249 	    "%-12llu %-12llu %-8llu %-8llu %-12llu %-12llu %-12llu %-12llu\n",
250 	    (longlong_t)sth->txg, sth->times[TXG_STATE_BIRTH], state,
251 	    (u_longlong_t)sth->ndirty,
252 	    (u_longlong_t)sth->nread, (u_longlong_t)sth->nwritten,
253 	    (u_longlong_t)sth->reads, (u_longlong_t)sth->writes,
254 	    (u_longlong_t)open, (u_longlong_t)quiesce, (u_longlong_t)wait,
255 	    (u_longlong_t)sync);
256 
257 	return (0);
258 }
259 
260 /* Remove oldest elements from list until there are no more than 'size' left */
261 static void
spa_txg_history_truncate(spa_history_list_t * shl,unsigned int size)262 spa_txg_history_truncate(spa_history_list_t *shl, unsigned int size)
263 {
264 	spa_txg_history_t *sth;
265 	while (shl->size > size) {
266 		sth = list_remove_head(&shl->procfs_list.pl_list);
267 		ASSERT3P(sth, !=, NULL);
268 		kmem_free(sth, sizeof (spa_txg_history_t));
269 		shl->size--;
270 	}
271 
272 	if (size == 0)
273 		ASSERT(list_is_empty(&shl->procfs_list.pl_list));
274 
275 }
276 
277 static int
spa_txg_history_clear(procfs_list_t * procfs_list)278 spa_txg_history_clear(procfs_list_t *procfs_list)
279 {
280 	spa_history_list_t *shl = procfs_list->pl_private;
281 	mutex_enter(&procfs_list->pl_lock);
282 	spa_txg_history_truncate(shl, 0);
283 	mutex_exit(&procfs_list->pl_lock);
284 	return (0);
285 }
286 
287 static void
spa_txg_history_init(spa_t * spa)288 spa_txg_history_init(spa_t *spa)
289 {
290 	spa_history_list_t *shl = &spa->spa_stats.txg_history;
291 
292 	shl->size = 0;
293 	shl->procfs_list.pl_private = shl;
294 	procfs_list_install("zfs",
295 	    spa_name(spa),
296 	    "txgs",
297 	    0644,
298 	    &shl->procfs_list,
299 	    spa_txg_history_show,
300 	    spa_txg_history_show_header,
301 	    spa_txg_history_clear,
302 	    offsetof(spa_txg_history_t, sth_node));
303 }
304 
305 static void
spa_txg_history_destroy(spa_t * spa)306 spa_txg_history_destroy(spa_t *spa)
307 {
308 	spa_history_list_t *shl = &spa->spa_stats.txg_history;
309 	procfs_list_uninstall(&shl->procfs_list);
310 	spa_txg_history_truncate(shl, 0);
311 	procfs_list_destroy(&shl->procfs_list);
312 }
313 
314 /*
315  * Add a new txg to historical record.
316  */
317 void
spa_txg_history_add(spa_t * spa,uint64_t txg,hrtime_t birth_time)318 spa_txg_history_add(spa_t *spa, uint64_t txg, hrtime_t birth_time)
319 {
320 	spa_history_list_t *shl = &spa->spa_stats.txg_history;
321 	spa_txg_history_t *sth;
322 
323 	if (zfs_txg_history == 0 && shl->size == 0)
324 		return;
325 
326 	sth = kmem_zalloc(sizeof (spa_txg_history_t), KM_SLEEP);
327 	sth->txg = txg;
328 	sth->state = TXG_STATE_OPEN;
329 	sth->times[TXG_STATE_BIRTH] = birth_time;
330 
331 	mutex_enter(&shl->procfs_list.pl_lock);
332 	procfs_list_add(&shl->procfs_list, sth);
333 	shl->size++;
334 	spa_txg_history_truncate(shl, zfs_txg_history);
335 	mutex_exit(&shl->procfs_list.pl_lock);
336 }
337 
338 /*
339  * Set txg state completion time and increment current state.
340  */
341 int
spa_txg_history_set(spa_t * spa,uint64_t txg,txg_state_t completed_state,hrtime_t completed_time)342 spa_txg_history_set(spa_t *spa, uint64_t txg, txg_state_t completed_state,
343     hrtime_t completed_time)
344 {
345 	spa_history_list_t *shl = &spa->spa_stats.txg_history;
346 	spa_txg_history_t *sth;
347 	int error = ENOENT;
348 
349 	if (zfs_txg_history == 0)
350 		return (0);
351 
352 	mutex_enter(&shl->procfs_list.pl_lock);
353 	for (sth = list_tail(&shl->procfs_list.pl_list); sth != NULL;
354 	    sth = list_prev(&shl->procfs_list.pl_list, sth)) {
355 		if (sth->txg == txg) {
356 			sth->times[completed_state] = completed_time;
357 			sth->state++;
358 			error = 0;
359 			break;
360 		}
361 	}
362 	mutex_exit(&shl->procfs_list.pl_lock);
363 
364 	return (error);
365 }
366 
367 /*
368  * Set txg IO stats.
369  */
370 static int
spa_txg_history_set_io(spa_t * spa,uint64_t txg,uint64_t nread,uint64_t nwritten,uint64_t reads,uint64_t writes,uint64_t ndirty)371 spa_txg_history_set_io(spa_t *spa, uint64_t txg, uint64_t nread,
372     uint64_t nwritten, uint64_t reads, uint64_t writes, uint64_t ndirty)
373 {
374 	spa_history_list_t *shl = &spa->spa_stats.txg_history;
375 	spa_txg_history_t *sth;
376 	int error = ENOENT;
377 
378 	if (zfs_txg_history == 0)
379 		return (0);
380 
381 	mutex_enter(&shl->procfs_list.pl_lock);
382 	for (sth = list_tail(&shl->procfs_list.pl_list); sth != NULL;
383 	    sth = list_prev(&shl->procfs_list.pl_list, sth)) {
384 		if (sth->txg == txg) {
385 			sth->nread = nread;
386 			sth->nwritten = nwritten;
387 			sth->reads = reads;
388 			sth->writes = writes;
389 			sth->ndirty = ndirty;
390 			error = 0;
391 			break;
392 		}
393 	}
394 	mutex_exit(&shl->procfs_list.pl_lock);
395 
396 	return (error);
397 }
398 
399 txg_stat_t *
spa_txg_history_init_io(spa_t * spa,uint64_t txg,dsl_pool_t * dp)400 spa_txg_history_init_io(spa_t *spa, uint64_t txg, dsl_pool_t *dp)
401 {
402 	txg_stat_t *ts;
403 
404 	if (zfs_txg_history == 0)
405 		return (NULL);
406 
407 	ts = kmem_alloc(sizeof (txg_stat_t), KM_SLEEP);
408 
409 	spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
410 	vdev_get_stats(spa->spa_root_vdev, &ts->vs1);
411 	spa_config_exit(spa, SCL_CONFIG, FTAG);
412 
413 	ts->txg = txg;
414 	ts->ndirty = dp->dp_dirty_pertxg[txg & TXG_MASK];
415 
416 	spa_txg_history_set(spa, txg, TXG_STATE_WAIT_FOR_SYNC, gethrtime());
417 
418 	return (ts);
419 }
420 
421 void
spa_txg_history_fini_io(spa_t * spa,txg_stat_t * ts)422 spa_txg_history_fini_io(spa_t *spa, txg_stat_t *ts)
423 {
424 	if (ts == NULL)
425 		return;
426 
427 	if (zfs_txg_history == 0) {
428 		kmem_free(ts, sizeof (txg_stat_t));
429 		return;
430 	}
431 
432 	spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
433 	vdev_get_stats(spa->spa_root_vdev, &ts->vs2);
434 	spa_config_exit(spa, SCL_CONFIG, FTAG);
435 
436 	spa_txg_history_set(spa, ts->txg, TXG_STATE_SYNCED, gethrtime());
437 	spa_txg_history_set_io(spa, ts->txg,
438 	    ts->vs2.vs_bytes[ZIO_TYPE_READ] - ts->vs1.vs_bytes[ZIO_TYPE_READ],
439 	    ts->vs2.vs_bytes[ZIO_TYPE_WRITE] - ts->vs1.vs_bytes[ZIO_TYPE_WRITE],
440 	    ts->vs2.vs_ops[ZIO_TYPE_READ] - ts->vs1.vs_ops[ZIO_TYPE_READ],
441 	    ts->vs2.vs_ops[ZIO_TYPE_WRITE] - ts->vs1.vs_ops[ZIO_TYPE_WRITE],
442 	    ts->ndirty);
443 
444 	kmem_free(ts, sizeof (txg_stat_t));
445 }
446 
447 /*
448  * ==========================================================================
449  * SPA TX Assign Histogram Routines
450  * ==========================================================================
451  */
452 
453 /*
454  * Tx statistics - Information exported regarding dmu_tx_assign time.
455  */
456 
457 /*
458  * When the kstat is written zero all buckets.  When the kstat is read
459  * count the number of trailing buckets set to zero and update ks_ndata
460  * such that they are not output.
461  */
462 static int
spa_tx_assign_update(kstat_t * ksp,int rw)463 spa_tx_assign_update(kstat_t *ksp, int rw)
464 {
465 	spa_t *spa = ksp->ks_private;
466 	spa_history_kstat_t *shk = &spa->spa_stats.tx_assign_histogram;
467 	int i;
468 
469 	if (rw == KSTAT_WRITE) {
470 		for (i = 0; i < shk->count; i++)
471 			((kstat_named_t *)shk->priv)[i].value.ui64 = 0;
472 	}
473 
474 	for (i = shk->count; i > 0; i--)
475 		if (((kstat_named_t *)shk->priv)[i-1].value.ui64 != 0)
476 			break;
477 
478 	ksp->ks_ndata = i;
479 	ksp->ks_data_size = i * sizeof (kstat_named_t);
480 
481 	return (0);
482 }
483 
484 static void
spa_tx_assign_init(spa_t * spa)485 spa_tx_assign_init(spa_t *spa)
486 {
487 	spa_history_kstat_t *shk = &spa->spa_stats.tx_assign_histogram;
488 	char *name;
489 	kstat_named_t *ks;
490 	kstat_t *ksp;
491 	int i;
492 
493 	mutex_init(&shk->lock, NULL, MUTEX_DEFAULT, NULL);
494 
495 	shk->count = 42; /* power of two buckets for 1ns to 2,199s */
496 	shk->size = shk->count * sizeof (kstat_named_t);
497 	shk->priv = kmem_alloc(shk->size, KM_SLEEP);
498 
499 	name = kmem_asprintf("zfs/%s", spa_name(spa));
500 
501 	for (i = 0; i < shk->count; i++) {
502 		ks = &((kstat_named_t *)shk->priv)[i];
503 		ks->data_type = KSTAT_DATA_UINT64;
504 		ks->value.ui64 = 0;
505 		(void) snprintf(ks->name, KSTAT_STRLEN, "%llu ns",
506 		    (u_longlong_t)1 << i);
507 	}
508 
509 	ksp = kstat_create(name, 0, "dmu_tx_assign", "misc",
510 	    KSTAT_TYPE_NAMED, 0, KSTAT_FLAG_VIRTUAL);
511 	shk->kstat = ksp;
512 
513 	if (ksp) {
514 		ksp->ks_lock = &shk->lock;
515 		ksp->ks_data = shk->priv;
516 		ksp->ks_ndata = shk->count;
517 		ksp->ks_data_size = shk->size;
518 		ksp->ks_private = spa;
519 		ksp->ks_update = spa_tx_assign_update;
520 		kstat_install(ksp);
521 	}
522 	kmem_strfree(name);
523 }
524 
525 static void
spa_tx_assign_destroy(spa_t * spa)526 spa_tx_assign_destroy(spa_t *spa)
527 {
528 	spa_history_kstat_t *shk = &spa->spa_stats.tx_assign_histogram;
529 	kstat_t *ksp;
530 
531 	ksp = shk->kstat;
532 	if (ksp)
533 		kstat_delete(ksp);
534 
535 	kmem_free(shk->priv, shk->size);
536 	mutex_destroy(&shk->lock);
537 }
538 
539 void
spa_tx_assign_add_nsecs(spa_t * spa,uint64_t nsecs)540 spa_tx_assign_add_nsecs(spa_t *spa, uint64_t nsecs)
541 {
542 	spa_history_kstat_t *shk = &spa->spa_stats.tx_assign_histogram;
543 	uint64_t idx = 0;
544 
545 	while (((1ULL << idx) < nsecs) && (idx < shk->size - 1))
546 		idx++;
547 
548 	atomic_inc_64(&((kstat_named_t *)shk->priv)[idx].value.ui64);
549 }
550 
551 /*
552  * ==========================================================================
553  * SPA MMP History Routines
554  * ==========================================================================
555  */
556 
557 /*
558  * MMP statistics - Information exported regarding attempted MMP writes
559  *   For MMP writes issued, fields used as per comments below.
560  *   For MMP writes skipped, an entry represents a span of time when
561  *      writes were skipped for same reason (error from mmp_random_leaf).
562  *      Differences are:
563  *      timestamp	time first write skipped, if >1 skipped in a row
564  *      mmp_delay	delay value at timestamp
565  *      vdev_guid	number of writes skipped
566  *      io_error	one of enum mmp_error
567  *      duration	time span (ns) of skipped writes
568  */
569 
570 typedef struct spa_mmp_history {
571 	uint64_t	mmp_node_id;	/* unique # for updates */
572 	uint64_t	txg;		/* txg of last sync */
573 	uint64_t	timestamp;	/* UTC time MMP write issued */
574 	uint64_t	mmp_delay;	/* mmp_thread.mmp_delay at timestamp */
575 	uint64_t	vdev_guid;	/* unique ID of leaf vdev */
576 	char		*vdev_path;
577 	int		vdev_label;	/* vdev label */
578 	int		io_error;	/* error status of MMP write */
579 	hrtime_t	error_start;	/* hrtime of start of error period */
580 	hrtime_t	duration;	/* time from submission to completion */
581 	procfs_list_node_t	smh_node;
582 } spa_mmp_history_t;
583 
584 static int
spa_mmp_history_show_header(struct seq_file * f)585 spa_mmp_history_show_header(struct seq_file *f)
586 {
587 	seq_printf(f, "%-10s %-10s %-10s %-6s %-10s %-12s %-24s "
588 	    "%-10s %s\n", "id", "txg", "timestamp", "error", "duration",
589 	    "mmp_delay", "vdev_guid", "vdev_label", "vdev_path");
590 	return (0);
591 }
592 
593 static int
spa_mmp_history_show(struct seq_file * f,void * data)594 spa_mmp_history_show(struct seq_file *f, void *data)
595 {
596 	spa_mmp_history_t *smh = (spa_mmp_history_t *)data;
597 	char skip_fmt[] = "%-10llu %-10llu %10llu %#6llx %10lld %12llu %-24llu "
598 	    "%-10lld %s\n";
599 	char write_fmt[] = "%-10llu %-10llu %10llu %6lld %10lld %12llu %-24llu "
600 	    "%-10lld %s\n";
601 
602 	seq_printf(f, (smh->error_start ? skip_fmt : write_fmt),
603 	    (u_longlong_t)smh->mmp_node_id, (u_longlong_t)smh->txg,
604 	    (u_longlong_t)smh->timestamp, (longlong_t)smh->io_error,
605 	    (longlong_t)smh->duration, (u_longlong_t)smh->mmp_delay,
606 	    (u_longlong_t)smh->vdev_guid, (u_longlong_t)smh->vdev_label,
607 	    (smh->vdev_path ? smh->vdev_path : "-"));
608 
609 	return (0);
610 }
611 
612 /* Remove oldest elements from list until there are no more than 'size' left */
613 static void
spa_mmp_history_truncate(spa_history_list_t * shl,unsigned int size)614 spa_mmp_history_truncate(spa_history_list_t *shl, unsigned int size)
615 {
616 	spa_mmp_history_t *smh;
617 	while (shl->size > size) {
618 		smh = list_remove_head(&shl->procfs_list.pl_list);
619 		if (smh->vdev_path)
620 			kmem_strfree(smh->vdev_path);
621 		kmem_free(smh, sizeof (spa_mmp_history_t));
622 		shl->size--;
623 	}
624 
625 	if (size == 0)
626 		ASSERT(list_is_empty(&shl->procfs_list.pl_list));
627 
628 }
629 
630 static int
spa_mmp_history_clear(procfs_list_t * procfs_list)631 spa_mmp_history_clear(procfs_list_t *procfs_list)
632 {
633 	spa_history_list_t *shl = procfs_list->pl_private;
634 	mutex_enter(&procfs_list->pl_lock);
635 	spa_mmp_history_truncate(shl, 0);
636 	mutex_exit(&procfs_list->pl_lock);
637 	return (0);
638 }
639 
640 static void
spa_mmp_history_init(spa_t * spa)641 spa_mmp_history_init(spa_t *spa)
642 {
643 	spa_history_list_t *shl = &spa->spa_stats.mmp_history;
644 
645 	shl->size = 0;
646 
647 	shl->procfs_list.pl_private = shl;
648 	procfs_list_install("zfs",
649 	    spa_name(spa),
650 	    "multihost",
651 	    0644,
652 	    &shl->procfs_list,
653 	    spa_mmp_history_show,
654 	    spa_mmp_history_show_header,
655 	    spa_mmp_history_clear,
656 	    offsetof(spa_mmp_history_t, smh_node));
657 }
658 
659 static void
spa_mmp_history_destroy(spa_t * spa)660 spa_mmp_history_destroy(spa_t *spa)
661 {
662 	spa_history_list_t *shl = &spa->spa_stats.mmp_history;
663 	procfs_list_uninstall(&shl->procfs_list);
664 	spa_mmp_history_truncate(shl, 0);
665 	procfs_list_destroy(&shl->procfs_list);
666 }
667 
668 /*
669  * Set duration in existing "skip" record to how long we have waited for a leaf
670  * vdev to become available.
671  *
672  * Important that we start search at the tail of the list where new
673  * records are inserted, so this is normally an O(1) operation.
674  */
675 int
spa_mmp_history_set_skip(spa_t * spa,uint64_t mmp_node_id)676 spa_mmp_history_set_skip(spa_t *spa, uint64_t mmp_node_id)
677 {
678 	spa_history_list_t *shl = &spa->spa_stats.mmp_history;
679 	spa_mmp_history_t *smh;
680 	int error = ENOENT;
681 
682 	if (zfs_multihost_history == 0 && shl->size == 0)
683 		return (0);
684 
685 	mutex_enter(&shl->procfs_list.pl_lock);
686 	for (smh = list_tail(&shl->procfs_list.pl_list); smh != NULL;
687 	    smh = list_prev(&shl->procfs_list.pl_list, smh)) {
688 		if (smh->mmp_node_id == mmp_node_id) {
689 			ASSERT3U(smh->io_error, !=, 0);
690 			smh->duration = gethrtime() - smh->error_start;
691 			smh->vdev_guid++;
692 			error = 0;
693 			break;
694 		}
695 	}
696 	mutex_exit(&shl->procfs_list.pl_lock);
697 
698 	return (error);
699 }
700 
701 /*
702  * Set MMP write duration and error status in existing record.
703  * See comment re: search order above spa_mmp_history_set_skip().
704  */
705 int
spa_mmp_history_set(spa_t * spa,uint64_t mmp_node_id,int io_error,hrtime_t duration)706 spa_mmp_history_set(spa_t *spa, uint64_t mmp_node_id, int io_error,
707     hrtime_t duration)
708 {
709 	spa_history_list_t *shl = &spa->spa_stats.mmp_history;
710 	spa_mmp_history_t *smh;
711 	int error = ENOENT;
712 
713 	if (zfs_multihost_history == 0 && shl->size == 0)
714 		return (0);
715 
716 	mutex_enter(&shl->procfs_list.pl_lock);
717 	for (smh = list_tail(&shl->procfs_list.pl_list); smh != NULL;
718 	    smh = list_prev(&shl->procfs_list.pl_list, smh)) {
719 		if (smh->mmp_node_id == mmp_node_id) {
720 			ASSERT(smh->io_error == 0);
721 			smh->io_error = io_error;
722 			smh->duration = duration;
723 			error = 0;
724 			break;
725 		}
726 	}
727 	mutex_exit(&shl->procfs_list.pl_lock);
728 
729 	return (error);
730 }
731 
732 /*
733  * Add a new MMP historical record.
734  * error == 0 : a write was issued.
735  * error != 0 : a write was not issued because no leaves were found.
736  */
737 void
spa_mmp_history_add(spa_t * spa,uint64_t txg,uint64_t timestamp,uint64_t mmp_delay,vdev_t * vd,int label,uint64_t mmp_node_id,int error)738 spa_mmp_history_add(spa_t *spa, uint64_t txg, uint64_t timestamp,
739     uint64_t mmp_delay, vdev_t *vd, int label, uint64_t mmp_node_id,
740     int error)
741 {
742 	spa_history_list_t *shl = &spa->spa_stats.mmp_history;
743 	spa_mmp_history_t *smh;
744 
745 	if (zfs_multihost_history == 0 && shl->size == 0)
746 		return;
747 
748 	smh = kmem_zalloc(sizeof (spa_mmp_history_t), KM_SLEEP);
749 	smh->txg = txg;
750 	smh->timestamp = timestamp;
751 	smh->mmp_delay = mmp_delay;
752 	if (vd) {
753 		smh->vdev_guid = vd->vdev_guid;
754 		if (vd->vdev_path)
755 			smh->vdev_path = kmem_strdup(vd->vdev_path);
756 	}
757 	smh->vdev_label = label;
758 	smh->mmp_node_id = mmp_node_id;
759 
760 	if (error) {
761 		smh->io_error = error;
762 		smh->error_start = gethrtime();
763 		smh->vdev_guid = 1;
764 	}
765 
766 	mutex_enter(&shl->procfs_list.pl_lock);
767 	procfs_list_add(&shl->procfs_list, smh);
768 	shl->size++;
769 	spa_mmp_history_truncate(shl, zfs_multihost_history);
770 	mutex_exit(&shl->procfs_list.pl_lock);
771 }
772 
773 static void *
spa_state_addr(kstat_t * ksp,loff_t n)774 spa_state_addr(kstat_t *ksp, loff_t n)
775 {
776 	if (n == 0)
777 		return (ksp->ks_private);	/* return the spa_t */
778 	return (NULL);
779 }
780 
781 static int
spa_state_data(char * buf,size_t size,void * data)782 spa_state_data(char *buf, size_t size, void *data)
783 {
784 	spa_t *spa = (spa_t *)data;
785 	(void) snprintf(buf, size, "%s\n", spa_state_to_name(spa));
786 	return (0);
787 }
788 
789 /*
790  * Return the state of the pool in /proc/spl/kstat/zfs/<pool>/state.
791  *
792  * This is a lock-less read of the pool's state (unlike using 'zpool', which
793  * can potentially block for seconds).  Because it doesn't block, it can useful
794  * as a pool heartbeat value.
795  */
796 static void
spa_state_init(spa_t * spa)797 spa_state_init(spa_t *spa)
798 {
799 	spa_history_kstat_t *shk = &spa->spa_stats.state;
800 	char *name;
801 	kstat_t *ksp;
802 
803 	mutex_init(&shk->lock, NULL, MUTEX_DEFAULT, NULL);
804 
805 	name = kmem_asprintf("zfs/%s", spa_name(spa));
806 	ksp = kstat_create(name, 0, "state", "misc",
807 	    KSTAT_TYPE_RAW, 0, KSTAT_FLAG_VIRTUAL);
808 
809 	shk->kstat = ksp;
810 	if (ksp) {
811 		ksp->ks_lock = &shk->lock;
812 		ksp->ks_data = NULL;
813 		ksp->ks_private = spa;
814 		ksp->ks_flags |= KSTAT_FLAG_NO_HEADERS;
815 		kstat_set_raw_ops(ksp, NULL, spa_state_data, spa_state_addr);
816 		kstat_install(ksp);
817 	}
818 
819 	kmem_strfree(name);
820 }
821 
822 static int
spa_guid_data(char * buf,size_t size,void * data)823 spa_guid_data(char *buf, size_t size, void *data)
824 {
825 	spa_t *spa = (spa_t *)data;
826 	(void) snprintf(buf, size, "%llu\n", (u_longlong_t)spa_guid(spa));
827 	return (0);
828 }
829 
830 static void
spa_guid_init(spa_t * spa)831 spa_guid_init(spa_t *spa)
832 {
833 	spa_history_kstat_t *shk = &spa->spa_stats.guid;
834 	char *name;
835 	kstat_t *ksp;
836 
837 	mutex_init(&shk->lock, NULL, MUTEX_DEFAULT, NULL);
838 
839 	name = kmem_asprintf("zfs/%s", spa_name(spa));
840 
841 	ksp = kstat_create(name, 0, "guid", "misc",
842 	    KSTAT_TYPE_RAW, 0, KSTAT_FLAG_VIRTUAL);
843 
844 	shk->kstat = ksp;
845 	if (ksp) {
846 		ksp->ks_lock = &shk->lock;
847 		ksp->ks_data = NULL;
848 		ksp->ks_private = spa;
849 		ksp->ks_flags |= KSTAT_FLAG_NO_HEADERS;
850 		kstat_set_raw_ops(ksp, NULL, spa_guid_data, spa_state_addr);
851 		kstat_install(ksp);
852 	}
853 
854 	kmem_strfree(name);
855 }
856 
857 static void
spa_health_destroy(spa_t * spa)858 spa_health_destroy(spa_t *spa)
859 {
860 	spa_history_kstat_t *shk = &spa->spa_stats.state;
861 	kstat_t *ksp = shk->kstat;
862 	if (ksp)
863 		kstat_delete(ksp);
864 
865 	mutex_destroy(&shk->lock);
866 }
867 
868 static void
spa_guid_destroy(spa_t * spa)869 spa_guid_destroy(spa_t *spa)
870 {
871 	spa_history_kstat_t *shk = &spa->spa_stats.guid;
872 	kstat_t *ksp = shk->kstat;
873 	if (ksp)
874 		kstat_delete(ksp);
875 
876 	mutex_destroy(&shk->lock);
877 }
878 
879 static const spa_iostats_t spa_iostats_template = {
880 	{ "trim_extents_written",		KSTAT_DATA_UINT64 },
881 	{ "trim_bytes_written",			KSTAT_DATA_UINT64 },
882 	{ "trim_extents_skipped",		KSTAT_DATA_UINT64 },
883 	{ "trim_bytes_skipped",			KSTAT_DATA_UINT64 },
884 	{ "trim_extents_failed",		KSTAT_DATA_UINT64 },
885 	{ "trim_bytes_failed",			KSTAT_DATA_UINT64 },
886 	{ "autotrim_extents_written",		KSTAT_DATA_UINT64 },
887 	{ "autotrim_bytes_written",		KSTAT_DATA_UINT64 },
888 	{ "autotrim_extents_skipped",		KSTAT_DATA_UINT64 },
889 	{ "autotrim_bytes_skipped",		KSTAT_DATA_UINT64 },
890 	{ "autotrim_extents_failed",		KSTAT_DATA_UINT64 },
891 	{ "autotrim_bytes_failed",		KSTAT_DATA_UINT64 },
892 	{ "simple_trim_extents_written",	KSTAT_DATA_UINT64 },
893 	{ "simple_trim_bytes_written",		KSTAT_DATA_UINT64 },
894 	{ "simple_trim_extents_skipped",	KSTAT_DATA_UINT64 },
895 	{ "simple_trim_bytes_skipped",		KSTAT_DATA_UINT64 },
896 	{ "simple_trim_extents_failed",		KSTAT_DATA_UINT64 },
897 	{ "simple_trim_bytes_failed",		KSTAT_DATA_UINT64 },
898 	{ "arc_read_count",			KSTAT_DATA_UINT64 },
899 	{ "arc_read_bytes",			KSTAT_DATA_UINT64 },
900 	{ "arc_write_count",			KSTAT_DATA_UINT64 },
901 	{ "arc_write_bytes",			KSTAT_DATA_UINT64 },
902 	{ "direct_read_count",			KSTAT_DATA_UINT64 },
903 	{ "direct_read_bytes",			KSTAT_DATA_UINT64 },
904 	{ "direct_write_count",			KSTAT_DATA_UINT64 },
905 	{ "direct_write_bytes",			KSTAT_DATA_UINT64 },
906 };
907 
908 #define	SPA_IOSTATS_ADD(stat, val) \
909     atomic_add_64(&iostats->stat.value.ui64, (val));
910 
911 void
spa_iostats_trim_add(spa_t * spa,trim_type_t type,uint64_t extents_written,uint64_t bytes_written,uint64_t extents_skipped,uint64_t bytes_skipped,uint64_t extents_failed,uint64_t bytes_failed)912 spa_iostats_trim_add(spa_t *spa, trim_type_t type,
913     uint64_t extents_written, uint64_t bytes_written,
914     uint64_t extents_skipped, uint64_t bytes_skipped,
915     uint64_t extents_failed, uint64_t bytes_failed)
916 {
917 	spa_history_kstat_t *shk = &spa->spa_stats.iostats;
918 	kstat_t *ksp = shk->kstat;
919 	spa_iostats_t *iostats;
920 
921 	if (ksp == NULL)
922 		return;
923 
924 	iostats = ksp->ks_data;
925 	if (type == TRIM_TYPE_MANUAL) {
926 		SPA_IOSTATS_ADD(trim_extents_written, extents_written);
927 		SPA_IOSTATS_ADD(trim_bytes_written, bytes_written);
928 		SPA_IOSTATS_ADD(trim_extents_skipped, extents_skipped);
929 		SPA_IOSTATS_ADD(trim_bytes_skipped, bytes_skipped);
930 		SPA_IOSTATS_ADD(trim_extents_failed, extents_failed);
931 		SPA_IOSTATS_ADD(trim_bytes_failed, bytes_failed);
932 	} else if (type == TRIM_TYPE_AUTO) {
933 		SPA_IOSTATS_ADD(autotrim_extents_written, extents_written);
934 		SPA_IOSTATS_ADD(autotrim_bytes_written, bytes_written);
935 		SPA_IOSTATS_ADD(autotrim_extents_skipped, extents_skipped);
936 		SPA_IOSTATS_ADD(autotrim_bytes_skipped, bytes_skipped);
937 		SPA_IOSTATS_ADD(autotrim_extents_failed, extents_failed);
938 		SPA_IOSTATS_ADD(autotrim_bytes_failed, bytes_failed);
939 	} else {
940 		SPA_IOSTATS_ADD(simple_trim_extents_written, extents_written);
941 		SPA_IOSTATS_ADD(simple_trim_bytes_written, bytes_written);
942 		SPA_IOSTATS_ADD(simple_trim_extents_skipped, extents_skipped);
943 		SPA_IOSTATS_ADD(simple_trim_bytes_skipped, bytes_skipped);
944 		SPA_IOSTATS_ADD(simple_trim_extents_failed, extents_failed);
945 		SPA_IOSTATS_ADD(simple_trim_bytes_failed, bytes_failed);
946 	}
947 }
948 
949 void
spa_iostats_read_add(spa_t * spa,uint64_t size,uint64_t iops,uint32_t flags)950 spa_iostats_read_add(spa_t *spa, uint64_t size, uint64_t iops, uint32_t flags)
951 {
952 	spa_history_kstat_t *shk = &spa->spa_stats.iostats;
953 	kstat_t *ksp = shk->kstat;
954 
955 	if (ksp == NULL)
956 		return;
957 
958 	spa_iostats_t *iostats = ksp->ks_data;
959 	if (flags & DMU_DIRECTIO) {
960 		SPA_IOSTATS_ADD(direct_read_count, iops);
961 		SPA_IOSTATS_ADD(direct_read_bytes, size);
962 	} else {
963 		SPA_IOSTATS_ADD(arc_read_count, iops);
964 		SPA_IOSTATS_ADD(arc_read_bytes, size);
965 	}
966 }
967 
968 void
spa_iostats_write_add(spa_t * spa,uint64_t size,uint64_t iops,uint32_t flags)969 spa_iostats_write_add(spa_t *spa, uint64_t size, uint64_t iops, uint32_t flags)
970 {
971 	spa_history_kstat_t *shk = &spa->spa_stats.iostats;
972 	kstat_t *ksp = shk->kstat;
973 
974 	if (ksp == NULL)
975 		return;
976 
977 	spa_iostats_t *iostats = ksp->ks_data;
978 	if (flags & DMU_DIRECTIO) {
979 		SPA_IOSTATS_ADD(direct_write_count, iops);
980 		SPA_IOSTATS_ADD(direct_write_bytes, size);
981 	} else {
982 		SPA_IOSTATS_ADD(arc_write_count, iops);
983 		SPA_IOSTATS_ADD(arc_write_bytes, size);
984 	}
985 }
986 
987 static int
spa_iostats_update(kstat_t * ksp,int rw)988 spa_iostats_update(kstat_t *ksp, int rw)
989 {
990 	if (rw == KSTAT_WRITE) {
991 		memcpy(ksp->ks_data, &spa_iostats_template,
992 		    sizeof (spa_iostats_t));
993 	}
994 
995 	return (0);
996 }
997 
998 static void
spa_iostats_init(spa_t * spa)999 spa_iostats_init(spa_t *spa)
1000 {
1001 	spa_history_kstat_t *shk = &spa->spa_stats.iostats;
1002 
1003 	mutex_init(&shk->lock, NULL, MUTEX_DEFAULT, NULL);
1004 
1005 	char *name = kmem_asprintf("zfs/%s", spa_name(spa));
1006 	kstat_t *ksp = kstat_create(name, 0, "iostats", "misc",
1007 	    KSTAT_TYPE_NAMED, sizeof (spa_iostats_t) / sizeof (kstat_named_t),
1008 	    KSTAT_FLAG_VIRTUAL);
1009 
1010 	shk->kstat = ksp;
1011 	if (ksp) {
1012 		int size = sizeof (spa_iostats_t);
1013 		ksp->ks_lock = &shk->lock;
1014 		ksp->ks_private = spa;
1015 		ksp->ks_update = spa_iostats_update;
1016 		ksp->ks_data = kmem_alloc(size, KM_SLEEP);
1017 		memcpy(ksp->ks_data, &spa_iostats_template, size);
1018 		kstat_install(ksp);
1019 	}
1020 
1021 	kmem_strfree(name);
1022 }
1023 
1024 static void
spa_iostats_destroy(spa_t * spa)1025 spa_iostats_destroy(spa_t *spa)
1026 {
1027 	spa_history_kstat_t *shk = &spa->spa_stats.iostats;
1028 	kstat_t *ksp = shk->kstat;
1029 	if (ksp) {
1030 		kmem_free(ksp->ks_data, sizeof (spa_iostats_t));
1031 		kstat_delete(ksp);
1032 	}
1033 
1034 	mutex_destroy(&shk->lock);
1035 }
1036 
1037 void
spa_stats_init(spa_t * spa)1038 spa_stats_init(spa_t *spa)
1039 {
1040 	spa_read_history_init(spa);
1041 	spa_txg_history_init(spa);
1042 	spa_tx_assign_init(spa);
1043 	spa_mmp_history_init(spa);
1044 	spa_state_init(spa);
1045 	spa_guid_init(spa);
1046 	spa_iostats_init(spa);
1047 }
1048 
1049 void
spa_stats_destroy(spa_t * spa)1050 spa_stats_destroy(spa_t *spa)
1051 {
1052 	spa_iostats_destroy(spa);
1053 	spa_health_destroy(spa);
1054 	spa_tx_assign_destroy(spa);
1055 	spa_txg_history_destroy(spa);
1056 	spa_read_history_destroy(spa);
1057 	spa_mmp_history_destroy(spa);
1058 	spa_guid_destroy(spa);
1059 }
1060 
1061 ZFS_MODULE_PARAM(zfs, zfs_, read_history, UINT, ZMOD_RW,
1062 	"Historical statistics for the last N reads");
1063 
1064 ZFS_MODULE_PARAM(zfs, zfs_, read_history_hits, INT, ZMOD_RW,
1065 	"Include cache hits in read history");
1066 
1067 ZFS_MODULE_PARAM(zfs_txg, zfs_txg_, history, UINT, ZMOD_RW,
1068 	"Historical statistics for the last N txgs");
1069 
1070 ZFS_MODULE_PARAM(zfs_multihost, zfs_multihost_, history, UINT, ZMOD_RW,
1071 	"Historical statistics for last N multihost writes");
1072