xref: /openbmc/linux/fs/ceph/metric.c (revision 078b39c9)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #include <linux/ceph/ceph_debug.h>
3 
4 #include <linux/types.h>
5 #include <linux/percpu_counter.h>
6 #include <linux/math64.h>
7 
8 #include "metric.h"
9 #include "mds_client.h"
10 
11 static void ktime_to_ceph_timespec(struct ceph_timespec *ts, ktime_t val)
12 {
13 	struct timespec64 t = ktime_to_timespec64(val);
14 	ceph_encode_timespec64(ts, &t);
15 }
16 
17 static bool ceph_mdsc_send_metrics(struct ceph_mds_client *mdsc,
18 				   struct ceph_mds_session *s)
19 {
20 	struct ceph_metric_head *head;
21 	struct ceph_metric_cap *cap;
22 	struct ceph_metric_read_latency *read;
23 	struct ceph_metric_write_latency *write;
24 	struct ceph_metric_metadata_latency *meta;
25 	struct ceph_metric_dlease *dlease;
26 	struct ceph_opened_files *files;
27 	struct ceph_pinned_icaps *icaps;
28 	struct ceph_opened_inodes *inodes;
29 	struct ceph_read_io_size *rsize;
30 	struct ceph_write_io_size *wsize;
31 	struct ceph_client_metric *m = &mdsc->metric;
32 	u64 nr_caps = atomic64_read(&m->total_caps);
33 	u32 header_len = sizeof(struct ceph_metric_header);
34 	struct ceph_msg *msg;
35 	s64 sum;
36 	s32 items = 0;
37 	s32 len;
38 
39 	/* Do not send the metrics until the MDS rank is ready */
40 	mutex_lock(&mdsc->mutex);
41 	if (ceph_mdsmap_get_state(mdsc->mdsmap, s->s_mds) != CEPH_MDS_STATE_ACTIVE) {
42 		mutex_unlock(&mdsc->mutex);
43 		return false;
44 	}
45 	mutex_unlock(&mdsc->mutex);
46 
47 	len = sizeof(*head) + sizeof(*cap) + sizeof(*read) + sizeof(*write)
48 	      + sizeof(*meta) + sizeof(*dlease) + sizeof(*files)
49 	      + sizeof(*icaps) + sizeof(*inodes) + sizeof(*rsize)
50 	      + sizeof(*wsize);
51 
52 	msg = ceph_msg_new(CEPH_MSG_CLIENT_METRICS, len, GFP_NOFS, true);
53 	if (!msg) {
54 		pr_err("send metrics to mds%d, failed to allocate message\n",
55 		       s->s_mds);
56 		return false;
57 	}
58 
59 	head = msg->front.iov_base;
60 
61 	/* encode the cap metric */
62 	cap = (struct ceph_metric_cap *)(head + 1);
63 	cap->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_CAP_INFO);
64 	cap->header.ver = 1;
65 	cap->header.compat = 1;
66 	cap->header.data_len = cpu_to_le32(sizeof(*cap) - header_len);
67 	cap->hit = cpu_to_le64(percpu_counter_sum(&m->i_caps_hit));
68 	cap->mis = cpu_to_le64(percpu_counter_sum(&m->i_caps_mis));
69 	cap->total = cpu_to_le64(nr_caps);
70 	items++;
71 
72 	/* encode the read latency metric */
73 	read = (struct ceph_metric_read_latency *)(cap + 1);
74 	read->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_READ_LATENCY);
75 	read->header.ver = 2;
76 	read->header.compat = 1;
77 	read->header.data_len = cpu_to_le32(sizeof(*read) - header_len);
78 	sum = m->metric[METRIC_READ].latency_sum;
79 	ktime_to_ceph_timespec(&read->lat, sum);
80 	ktime_to_ceph_timespec(&read->avg, m->metric[METRIC_READ].latency_avg);
81 	read->sq_sum = cpu_to_le64(m->metric[METRIC_READ].latency_sq_sum);
82 	read->count = cpu_to_le64(m->metric[METRIC_READ].total);
83 	items++;
84 
85 	/* encode the write latency metric */
86 	write = (struct ceph_metric_write_latency *)(read + 1);
87 	write->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_WRITE_LATENCY);
88 	write->header.ver = 2;
89 	write->header.compat = 1;
90 	write->header.data_len = cpu_to_le32(sizeof(*write) - header_len);
91 	sum = m->metric[METRIC_WRITE].latency_sum;
92 	ktime_to_ceph_timespec(&write->lat, sum);
93 	ktime_to_ceph_timespec(&write->avg, m->metric[METRIC_WRITE].latency_avg);
94 	write->sq_sum = cpu_to_le64(m->metric[METRIC_WRITE].latency_sq_sum);
95 	write->count = cpu_to_le64(m->metric[METRIC_WRITE].total);
96 	items++;
97 
98 	/* encode the metadata latency metric */
99 	meta = (struct ceph_metric_metadata_latency *)(write + 1);
100 	meta->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_METADATA_LATENCY);
101 	meta->header.ver = 2;
102 	meta->header.compat = 1;
103 	meta->header.data_len = cpu_to_le32(sizeof(*meta) - header_len);
104 	sum = m->metric[METRIC_METADATA].latency_sum;
105 	ktime_to_ceph_timespec(&meta->lat, sum);
106 	ktime_to_ceph_timespec(&meta->avg, m->metric[METRIC_METADATA].latency_avg);
107 	meta->sq_sum = cpu_to_le64(m->metric[METRIC_METADATA].latency_sq_sum);
108 	meta->count = cpu_to_le64(m->metric[METRIC_METADATA].total);
109 	items++;
110 
111 	/* encode the dentry lease metric */
112 	dlease = (struct ceph_metric_dlease *)(meta + 1);
113 	dlease->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_DENTRY_LEASE);
114 	dlease->header.ver = 1;
115 	dlease->header.compat = 1;
116 	dlease->header.data_len = cpu_to_le32(sizeof(*dlease) - header_len);
117 	dlease->hit = cpu_to_le64(percpu_counter_sum(&m->d_lease_hit));
118 	dlease->mis = cpu_to_le64(percpu_counter_sum(&m->d_lease_mis));
119 	dlease->total = cpu_to_le64(atomic64_read(&m->total_dentries));
120 	items++;
121 
122 	sum = percpu_counter_sum(&m->total_inodes);
123 
124 	/* encode the opened files metric */
125 	files = (struct ceph_opened_files *)(dlease + 1);
126 	files->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_OPENED_FILES);
127 	files->header.ver = 1;
128 	files->header.compat = 1;
129 	files->header.data_len = cpu_to_le32(sizeof(*files) - header_len);
130 	files->opened_files = cpu_to_le64(atomic64_read(&m->opened_files));
131 	files->total = cpu_to_le64(sum);
132 	items++;
133 
134 	/* encode the pinned icaps metric */
135 	icaps = (struct ceph_pinned_icaps *)(files + 1);
136 	icaps->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_PINNED_ICAPS);
137 	icaps->header.ver = 1;
138 	icaps->header.compat = 1;
139 	icaps->header.data_len = cpu_to_le32(sizeof(*icaps) - header_len);
140 	icaps->pinned_icaps = cpu_to_le64(nr_caps);
141 	icaps->total = cpu_to_le64(sum);
142 	items++;
143 
144 	/* encode the opened inodes metric */
145 	inodes = (struct ceph_opened_inodes *)(icaps + 1);
146 	inodes->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_OPENED_INODES);
147 	inodes->header.ver = 1;
148 	inodes->header.compat = 1;
149 	inodes->header.data_len = cpu_to_le32(sizeof(*inodes) - header_len);
150 	inodes->opened_inodes = cpu_to_le64(percpu_counter_sum(&m->opened_inodes));
151 	inodes->total = cpu_to_le64(sum);
152 	items++;
153 
154 	/* encode the read io size metric */
155 	rsize = (struct ceph_read_io_size *)(inodes + 1);
156 	rsize->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_READ_IO_SIZES);
157 	rsize->header.ver = 1;
158 	rsize->header.compat = 1;
159 	rsize->header.data_len = cpu_to_le32(sizeof(*rsize) - header_len);
160 	rsize->total_ops = cpu_to_le64(m->metric[METRIC_READ].total);
161 	rsize->total_size = cpu_to_le64(m->metric[METRIC_READ].size_sum);
162 	items++;
163 
164 	/* encode the write io size metric */
165 	wsize = (struct ceph_write_io_size *)(rsize + 1);
166 	wsize->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_WRITE_IO_SIZES);
167 	wsize->header.ver = 1;
168 	wsize->header.compat = 1;
169 	wsize->header.data_len = cpu_to_le32(sizeof(*wsize) - header_len);
170 	wsize->total_ops = cpu_to_le64(m->metric[METRIC_WRITE].total);
171 	wsize->total_size = cpu_to_le64(m->metric[METRIC_WRITE].size_sum);
172 	items++;
173 
174 	put_unaligned_le32(items, &head->num);
175 	msg->front.iov_len = len;
176 	msg->hdr.version = cpu_to_le16(1);
177 	msg->hdr.compat_version = cpu_to_le16(1);
178 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
179 	ceph_con_send(&s->s_con, msg);
180 
181 	return true;
182 }
183 
184 
185 static void metric_get_session(struct ceph_mds_client *mdsc)
186 {
187 	struct ceph_mds_session *s;
188 	int i;
189 
190 	mutex_lock(&mdsc->mutex);
191 	for (i = 0; i < mdsc->max_sessions; i++) {
192 		s = __ceph_lookup_mds_session(mdsc, i);
193 		if (!s)
194 			continue;
195 
196 		/*
197 		 * Skip it if MDS doesn't support the metric collection,
198 		 * or the MDS will close the session's socket connection
199 		 * directly when it get this message.
200 		 */
201 		if (check_session_state(s) &&
202 		    test_bit(CEPHFS_FEATURE_METRIC_COLLECT, &s->s_features)) {
203 			mdsc->metric.session = s;
204 			break;
205 		}
206 
207 		ceph_put_mds_session(s);
208 	}
209 	mutex_unlock(&mdsc->mutex);
210 }
211 
212 static void metric_delayed_work(struct work_struct *work)
213 {
214 	struct ceph_client_metric *m =
215 		container_of(work, struct ceph_client_metric, delayed_work.work);
216 	struct ceph_mds_client *mdsc =
217 		container_of(m, struct ceph_mds_client, metric);
218 
219 	if (mdsc->stopping)
220 		return;
221 
222 	if (!m->session || !check_session_state(m->session)) {
223 		if (m->session) {
224 			ceph_put_mds_session(m->session);
225 			m->session = NULL;
226 		}
227 		metric_get_session(mdsc);
228 	}
229 	if (m->session) {
230 		ceph_mdsc_send_metrics(mdsc, m->session);
231 		metric_schedule_delayed(m);
232 	}
233 }
234 
235 int ceph_metric_init(struct ceph_client_metric *m)
236 {
237 	struct ceph_metric *metric;
238 	int ret, i;
239 
240 	if (!m)
241 		return -EINVAL;
242 
243 	atomic64_set(&m->total_dentries, 0);
244 	ret = percpu_counter_init(&m->d_lease_hit, 0, GFP_KERNEL);
245 	if (ret)
246 		return ret;
247 
248 	ret = percpu_counter_init(&m->d_lease_mis, 0, GFP_KERNEL);
249 	if (ret)
250 		goto err_d_lease_mis;
251 
252 	atomic64_set(&m->total_caps, 0);
253 	ret = percpu_counter_init(&m->i_caps_hit, 0, GFP_KERNEL);
254 	if (ret)
255 		goto err_i_caps_hit;
256 
257 	ret = percpu_counter_init(&m->i_caps_mis, 0, GFP_KERNEL);
258 	if (ret)
259 		goto err_i_caps_mis;
260 
261 	for (i = 0; i < METRIC_MAX; i++) {
262 		metric = &m->metric[i];
263 		spin_lock_init(&metric->lock);
264 		metric->size_sum = 0;
265 		metric->size_min = U64_MAX;
266 		metric->size_max = 0;
267 		metric->total = 0;
268 		metric->latency_sum = 0;
269 		metric->latency_avg = 0;
270 		metric->latency_sq_sum = 0;
271 		metric->latency_min = KTIME_MAX;
272 		metric->latency_max = 0;
273 	}
274 
275 	atomic64_set(&m->opened_files, 0);
276 	ret = percpu_counter_init(&m->opened_inodes, 0, GFP_KERNEL);
277 	if (ret)
278 		goto err_opened_inodes;
279 	ret = percpu_counter_init(&m->total_inodes, 0, GFP_KERNEL);
280 	if (ret)
281 		goto err_total_inodes;
282 
283 	m->session = NULL;
284 	INIT_DELAYED_WORK(&m->delayed_work, metric_delayed_work);
285 
286 	return 0;
287 
288 err_total_inodes:
289 	percpu_counter_destroy(&m->opened_inodes);
290 err_opened_inodes:
291 	percpu_counter_destroy(&m->i_caps_mis);
292 err_i_caps_mis:
293 	percpu_counter_destroy(&m->i_caps_hit);
294 err_i_caps_hit:
295 	percpu_counter_destroy(&m->d_lease_mis);
296 err_d_lease_mis:
297 	percpu_counter_destroy(&m->d_lease_hit);
298 
299 	return ret;
300 }
301 
302 void ceph_metric_destroy(struct ceph_client_metric *m)
303 {
304 	if (!m)
305 		return;
306 
307 	cancel_delayed_work_sync(&m->delayed_work);
308 
309 	percpu_counter_destroy(&m->total_inodes);
310 	percpu_counter_destroy(&m->opened_inodes);
311 	percpu_counter_destroy(&m->i_caps_mis);
312 	percpu_counter_destroy(&m->i_caps_hit);
313 	percpu_counter_destroy(&m->d_lease_mis);
314 	percpu_counter_destroy(&m->d_lease_hit);
315 
316 	ceph_put_mds_session(m->session);
317 }
318 
319 #define METRIC_UPDATE_MIN_MAX(min, max, new)	\
320 {						\
321 	if (unlikely(new < min))		\
322 		min = new;			\
323 	if (unlikely(new > max))		\
324 		max = new;			\
325 }
326 
327 static inline void __update_mean_and_stdev(ktime_t total, ktime_t *lavg,
328 					   ktime_t *sq_sump, ktime_t lat)
329 {
330 	ktime_t avg;
331 
332 	if (unlikely(total == 1)) {
333 		*lavg = lat;
334 	} else {
335 		/* the sq is (lat - old_avg) * (lat - new_avg) */
336 		avg = *lavg + div64_s64(lat - *lavg, total);
337 		*sq_sump += (lat - *lavg)*(lat - avg);
338 		*lavg = avg;
339 	}
340 }
341 
342 void ceph_update_metrics(struct ceph_metric *m,
343 			 ktime_t r_start, ktime_t r_end,
344 			 unsigned int size, int rc)
345 {
346 	ktime_t lat = ktime_sub(r_end, r_start);
347 	ktime_t total;
348 
349 	if (unlikely(rc < 0 && rc != -ENOENT && rc != -ETIMEDOUT))
350 		return;
351 
352 	spin_lock(&m->lock);
353 	total = ++m->total;
354 	m->size_sum += size;
355 	METRIC_UPDATE_MIN_MAX(m->size_min, m->size_max, size);
356 	m->latency_sum += lat;
357 	METRIC_UPDATE_MIN_MAX(m->latency_min, m->latency_max, lat);
358 	__update_mean_and_stdev(total, &m->latency_avg,	&m->latency_sq_sum,
359 				lat);
360 	spin_unlock(&m->lock);
361 }
362