fs/ceph/metric.c

   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 bool ceph_mdsc_send_metrics(struct ceph_mds_client *mdsc,
  12                                    struct ceph_mds_session *s)
  13 {
  14         struct ceph_metric_head *head;
  15         struct ceph_metric_cap *cap;
  16         struct ceph_metric_read_latency *read;
  17         struct ceph_metric_write_latency *write;
  18         struct ceph_metric_metadata_latency *meta;
  19         struct ceph_metric_dlease *dlease;
  20         struct ceph_opened_files *files;
  21         struct ceph_pinned_icaps *icaps;
  22         struct ceph_opened_inodes *inodes;
  23         struct ceph_read_io_size *rsize;
  24         struct ceph_write_io_size *wsize;
  25         struct ceph_client_metric *m = &mdsc->metric;
  26         u64 nr_caps = atomic64_read(&m->total_caps);
  27         u32 header_len = sizeof(struct ceph_metric_header);
  28         struct ceph_msg *msg;
  29         struct timespec64 ts;
  30         s64 sum;
  31         s32 items = 0;
  32         s32 len;
  33
  34         len = sizeof(*head) + sizeof(*cap) + sizeof(*read) + sizeof(*write)
  35               + sizeof(*meta) + sizeof(*dlease) + sizeof(*files)
  36               + sizeof(*icaps) + sizeof(*inodes) + sizeof(*rsize)
  37               + sizeof(*wsize);
  38
  39         msg = ceph_msg_new(CEPH_MSG_CLIENT_METRICS, len, GFP_NOFS, true);
  40         if (!msg) {
  41                 pr_err("send metrics to mds%d, failed to allocate message\n",
  42                        s->s_mds);
  43                 return false;
  44         }
  45
  46         head = msg->front.iov_base;
  47
  48         /* encode the cap metric */
  49         cap = (struct ceph_metric_cap *)(head + 1);
  50         cap->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_CAP_INFO);
  51         cap->header.ver = 1;
  52         cap->header.compat = 1;
  53         cap->header.data_len = cpu_to_le32(sizeof(*cap) - header_len);
  54         cap->hit = cpu_to_le64(percpu_counter_sum(&m->i_caps_hit));
  55         cap->mis = cpu_to_le64(percpu_counter_sum(&m->i_caps_mis));
  56         cap->total = cpu_to_le64(nr_caps);
  57         items++;
  58
  59         /* encode the read latency metric */
  60         read = (struct ceph_metric_read_latency *)(cap + 1);
  61         read->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_READ_LATENCY);
  62         read->header.ver = 1;
  63         read->header.compat = 1;
  64         read->header.data_len = cpu_to_le32(sizeof(*read) - header_len);
  65         sum = m->read_latency_sum;
  66         jiffies_to_timespec64(sum, &ts);
  67         read->sec = cpu_to_le32(ts.tv_sec);
  68         read->nsec = cpu_to_le32(ts.tv_nsec);
  69         items++;
  70
  71         /* encode the write latency metric */
  72         write = (struct ceph_metric_write_latency *)(read + 1);
  73         write->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_WRITE_LATENCY);
  74         write->header.ver = 1;
  75         write->header.compat = 1;
  76         write->header.data_len = cpu_to_le32(sizeof(*write) - header_len);
  77         sum = m->write_latency_sum;
  78         jiffies_to_timespec64(sum, &ts);
  79         write->sec = cpu_to_le32(ts.tv_sec);
  80         write->nsec = cpu_to_le32(ts.tv_nsec);
  81         items++;
  82
  83         /* encode the metadata latency metric */
  84         meta = (struct ceph_metric_metadata_latency *)(write + 1);
  85         meta->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_METADATA_LATENCY);
  86         meta->header.ver = 1;
  87         meta->header.compat = 1;
  88         meta->header.data_len = cpu_to_le32(sizeof(*meta) - header_len);
  89         sum = m->metadata_latency_sum;
  90         jiffies_to_timespec64(sum, &ts);
  91         meta->sec = cpu_to_le32(ts.tv_sec);
  92         meta->nsec = cpu_to_le32(ts.tv_nsec);
  93         items++;
  94
  95         /* encode the dentry lease metric */
  96         dlease = (struct ceph_metric_dlease *)(meta + 1);
  97         dlease->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_DENTRY_LEASE);
  98         dlease->header.ver = 1;
  99         dlease->header.compat = 1;
 100         dlease->header.data_len = cpu_to_le32(sizeof(*dlease) - header_len);
 101         dlease->hit = cpu_to_le64(percpu_counter_sum(&m->d_lease_hit));
 102         dlease->mis = cpu_to_le64(percpu_counter_sum(&m->d_lease_mis));
 103         dlease->total = cpu_to_le64(atomic64_read(&m->total_dentries));
 104         items++;
 105
 106         sum = percpu_counter_sum(&m->total_inodes);
 107
 108         /* encode the opened files metric */
 109         files = (struct ceph_opened_files *)(dlease + 1);
 110         files->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_OPENED_FILES);
 111         files->header.ver = 1;
 112         files->header.compat = 1;
 113         files->header.data_len = cpu_to_le32(sizeof(*files) - header_len);
 114         files->opened_files = cpu_to_le64(atomic64_read(&m->opened_files));
 115         files->total = cpu_to_le64(sum);
 116         items++;
 117
 118         /* encode the pinned icaps metric */
 119         icaps = (struct ceph_pinned_icaps *)(files + 1);
 120         icaps->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_PINNED_ICAPS);
 121         icaps->header.ver = 1;
 122         icaps->header.compat = 1;
 123         icaps->header.data_len = cpu_to_le32(sizeof(*icaps) - header_len);
 124         icaps->pinned_icaps = cpu_to_le64(nr_caps);
 125         icaps->total = cpu_to_le64(sum);
 126         items++;
 127
 128         /* encode the opened inodes metric */
 129         inodes = (struct ceph_opened_inodes *)(icaps + 1);
 130         inodes->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_OPENED_INODES);
 131         inodes->header.ver = 1;
 132         inodes->header.compat = 1;
 133         inodes->header.data_len = cpu_to_le32(sizeof(*inodes) - header_len);
 134         inodes->opened_inodes = cpu_to_le64(percpu_counter_sum(&m->opened_inodes));
 135         inodes->total = cpu_to_le64(sum);
 136         items++;
 137
 138         /* encode the read io size metric */
 139         rsize = (struct ceph_read_io_size *)(inodes + 1);
 140         rsize->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_READ_IO_SIZES);
 141         rsize->header.ver = 1;
 142         rsize->header.compat = 1;
 143         rsize->header.data_len = cpu_to_le32(sizeof(*rsize) - header_len);
 144         rsize->total_ops = cpu_to_le64(m->total_reads);
 145         rsize->total_size = cpu_to_le64(m->read_size_sum);
 146         items++;
 147
 148         /* encode the write io size metric */
 149         wsize = (struct ceph_write_io_size *)(rsize + 1);
 150         wsize->header.type = cpu_to_le32(CLIENT_METRIC_TYPE_WRITE_IO_SIZES);
 151         wsize->header.ver = 1;
 152         wsize->header.compat = 1;
 153         wsize->header.data_len = cpu_to_le32(sizeof(*wsize) - header_len);
 154         wsize->total_ops = cpu_to_le64(m->total_writes);
 155         wsize->total_size = cpu_to_le64(m->write_size_sum);
 156         items++;
 157
 158         put_unaligned_le32(items, &head->num);
 159         msg->front.iov_len = len;
 160         msg->hdr.version = cpu_to_le16(1);
 161         msg->hdr.compat_version = cpu_to_le16(1);
 162         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
 163         dout("client%llu send metrics to mds%d\n",
 164              ceph_client_gid(mdsc->fsc->client), s->s_mds);
 165         ceph_con_send(&s->s_con, msg);
 166
 167         return true;
 168 }
 169
 170
 171 static void metric_get_session(struct ceph_mds_client *mdsc)
 172 {
 173         struct ceph_mds_session *s;
 174         int i;
 175
 176         mutex_lock(&mdsc->mutex);
 177         for (i = 0; i < mdsc->max_sessions; i++) {
 178                 s = __ceph_lookup_mds_session(mdsc, i);
 179                 if (!s)
 180                         continue;
 181
 182                 /*
 183                  * Skip it if MDS doesn't support the metric collection,
 184                  * or the MDS will close the session's socket connection
 185                  * directly when it get this message.
 186                  */
 187                 if (check_session_state(s) &&
 188                     test_bit(CEPHFS_FEATURE_METRIC_COLLECT, &s->s_features)) {
 189                         mdsc->metric.session = s;
 190                         break;
 191                 }
 192
 193                 ceph_put_mds_session(s);
 194         }
 195         mutex_unlock(&mdsc->mutex);
 196 }
 197
 198 static void metric_delayed_work(struct work_struct *work)
 199 {
 200         struct ceph_client_metric *m =
 201                 container_of(work, struct ceph_client_metric, delayed_work.work);
 202         struct ceph_mds_client *mdsc =
 203                 container_of(m, struct ceph_mds_client, metric);
 204
 205         if (mdsc->stopping)
 206                 return;
 207
 208         if (!m->session || !check_session_state(m->session)) {
 209                 if (m->session) {
 210                         ceph_put_mds_session(m->session);
 211                         m->session = NULL;
 212                 }
 213                 metric_get_session(mdsc);
 214         }
 215         if (m->session) {
 216                 ceph_mdsc_send_metrics(mdsc, m->session);
 217                 metric_schedule_delayed(m);
 218         }
 219 }
 220
 221 int ceph_metric_init(struct ceph_client_metric *m)
 222 {
 223         int ret;
 224
 225         if (!m)
 226                 return -EINVAL;
 227
 228         atomic64_set(&m->total_dentries, 0);
 229         ret = percpu_counter_init(&m->d_lease_hit, 0, GFP_KERNEL);
 230         if (ret)
 231                 return ret;
 232
 233         ret = percpu_counter_init(&m->d_lease_mis, 0, GFP_KERNEL);
 234         if (ret)
 235                 goto err_d_lease_mis;
 236
 237         atomic64_set(&m->total_caps, 0);
 238         ret = percpu_counter_init(&m->i_caps_hit, 0, GFP_KERNEL);
 239         if (ret)
 240                 goto err_i_caps_hit;
 241
 242         ret = percpu_counter_init(&m->i_caps_mis, 0, GFP_KERNEL);
 243         if (ret)
 244                 goto err_i_caps_mis;
 245
 246         spin_lock_init(&m->read_metric_lock);
 247         m->read_latency_sq_sum = 0;
 248         m->read_latency_min = KTIME_MAX;
 249         m->read_latency_max = 0;
 250         m->total_reads = 0;
 251         m->read_latency_sum = 0;
 252         m->read_size_min = U64_MAX;
 253         m->read_size_max = 0;
 254         m->read_size_sum = 0;
 255
 256         spin_lock_init(&m->write_metric_lock);
 257         m->write_latency_sq_sum = 0;
 258         m->write_latency_min = KTIME_MAX;
 259         m->write_latency_max = 0;
 260         m->total_writes = 0;
 261         m->write_latency_sum = 0;
 262         m->write_size_min = U64_MAX;
 263         m->write_size_max = 0;
 264         m->write_size_sum = 0;
 265
 266         spin_lock_init(&m->metadata_metric_lock);
 267         m->metadata_latency_sq_sum = 0;
 268         m->metadata_latency_min = KTIME_MAX;
 269         m->metadata_latency_max = 0;
 270         m->total_metadatas = 0;
 271         m->metadata_latency_sum = 0;
 272
 273         atomic64_set(&m->opened_files, 0);
 274         ret = percpu_counter_init(&m->opened_inodes, 0, GFP_KERNEL);
 275         if (ret)
 276                 goto err_opened_inodes;
 277         ret = percpu_counter_init(&m->total_inodes, 0, GFP_KERNEL);
 278         if (ret)
 279                 goto err_total_inodes;
 280
 281         m->session = NULL;
 282         INIT_DELAYED_WORK(&m->delayed_work, metric_delayed_work);
 283
 284         return 0;
 285
 286 err_total_inodes:
 287         percpu_counter_destroy(&m->opened_inodes);
 288 err_opened_inodes:
 289         percpu_counter_destroy(&m->i_caps_mis);
 290 err_i_caps_mis:
 291         percpu_counter_destroy(&m->i_caps_hit);
 292 err_i_caps_hit:
 293         percpu_counter_destroy(&m->d_lease_mis);
 294 err_d_lease_mis:
 295         percpu_counter_destroy(&m->d_lease_hit);
 296
 297         return ret;
 298 }
 299
 300 void ceph_metric_destroy(struct ceph_client_metric *m)
 301 {
 302         if (!m)
 303                 return;
 304
 305         percpu_counter_destroy(&m->total_inodes);
 306         percpu_counter_destroy(&m->opened_inodes);
 307         percpu_counter_destroy(&m->i_caps_mis);
 308         percpu_counter_destroy(&m->i_caps_hit);
 309         percpu_counter_destroy(&m->d_lease_mis);
 310         percpu_counter_destroy(&m->d_lease_hit);
 311
 312         cancel_delayed_work_sync(&m->delayed_work);
 313
 314         ceph_put_mds_session(m->session);
 315 }
 316
 317 #define METRIC_UPDATE_MIN_MAX(min, max, new)    \
 318 {                                               \
 319         if (unlikely(new < min))                \
 320                 min = new;                      \
 321         if (unlikely(new > max))                \
 322                 max = new;                      \
 323 }
 324
 325 static inline void __update_stdev(ktime_t total, ktime_t lsum,
 326                                   ktime_t *sq_sump, ktime_t lat)
 327 {
 328         ktime_t avg, sq;
 329
 330         if (unlikely(total == 1))
 331                 return;
 332
 333         /* the sq is (lat - old_avg) * (lat - new_avg) */
 334         avg = DIV64_U64_ROUND_CLOSEST((lsum - lat), (total - 1));
 335         sq = lat - avg;
 336         avg = DIV64_U64_ROUND_CLOSEST(lsum, total);
 337         sq = sq * (lat - avg);
 338         *sq_sump += sq;
 339 }
 340
 341 void ceph_update_read_metrics(struct ceph_client_metric *m,
 342                               ktime_t r_start, ktime_t r_end,
 343                               unsigned int size, int rc)
 344 {
 345         ktime_t lat = ktime_sub(r_end, r_start);
 346         ktime_t total;
 347
 348         if (unlikely(rc < 0 && rc != -ENOENT && rc != -ETIMEDOUT))
 349                 return;
 350
 351         spin_lock(&m->read_metric_lock);
 352         total = ++m->total_reads;
 353         m->read_size_sum += size;
 354         m->read_latency_sum += lat;
 355         METRIC_UPDATE_MIN_MAX(m->read_size_min,
 356                               m->read_size_max,
 357                               size);
 358         METRIC_UPDATE_MIN_MAX(m->read_latency_min,
 359                               m->read_latency_max,
 360                               lat);
 361         __update_stdev(total, m->read_latency_sum,
 362                        &m->read_latency_sq_sum, lat);
 363         spin_unlock(&m->read_metric_lock);
 364 }
 365
 366 void ceph_update_write_metrics(struct ceph_client_metric *m,
 367                                ktime_t r_start, ktime_t r_end,
 368                                unsigned int size, int rc)
 369 {
 370         ktime_t lat = ktime_sub(r_end, r_start);
 371         ktime_t total;
 372
 373         if (unlikely(rc && rc != -ETIMEDOUT))
 374                 return;
 375
 376         spin_lock(&m->write_metric_lock);
 377         total = ++m->total_writes;
 378         m->write_size_sum += size;
 379         m->write_latency_sum += lat;
 380         METRIC_UPDATE_MIN_MAX(m->write_size_min,
 381                               m->write_size_max,
 382                               size);
 383         METRIC_UPDATE_MIN_MAX(m->write_latency_min,
 384                               m->write_latency_max,
 385                               lat);
 386         __update_stdev(total, m->write_latency_sum,
 387                        &m->write_latency_sq_sum, lat);
 388         spin_unlock(&m->write_metric_lock);
 389 }
 390
 391 void ceph_update_metadata_metrics(struct ceph_client_metric *m,
 392                                   ktime_t r_start, ktime_t r_end,
 393                                   int rc)
 394 {
 395         ktime_t lat = ktime_sub(r_end, r_start);
 396         ktime_t total;
 397
 398         if (unlikely(rc && rc != -ENOENT))
 399                 return;
 400
 401         spin_lock(&m->metadata_metric_lock);
 402         total = ++m->total_metadatas;
 403         m->metadata_latency_sum += lat;
 404         METRIC_UPDATE_MIN_MAX(m->metadata_latency_min,
 405                               m->metadata_latency_max,
 406                               lat);
 407         __update_stdev(total, m->metadata_latency_sum,
 408                        &m->metadata_latency_sq_sum, lat);
 409         spin_unlock(&m->metadata_metric_lock);
 410 }