src/bin/lttng-sessiond/timer.cpp

   1 /*
   2  * Copyright (C) 2017 Julien Desfossez <jdesfossez@efficios.com>
   3  * Copyright (C) 2018 Jérémie Galarneau <jeremie.galarneau@efficios.com>
   4  *
   5  * SPDX-License-Identifier: GPL-2.0-only
   6  *
   7  */
   8
   9 #define _LGPL_SOURCE
  10 #include <inttypes.h>
  11 #include <signal.h>
  12
  13 #include "timer.h"
  14 #include "health-sessiond.h"
  15 #include "rotation-thread.h"
  16 #include "thread.h"
  17
  18 #define LTTNG_SESSIOND_SIG_QS                           SIGRTMIN + 10
  19 #define LTTNG_SESSIOND_SIG_EXIT                         SIGRTMIN + 11
  20 #define LTTNG_SESSIOND_SIG_PENDING_ROTATION_CHECK       SIGRTMIN + 12
  21 #define LTTNG_SESSIOND_SIG_SCHEDULED_ROTATION           SIGRTMIN + 13
  22
  23 #define UINT_TO_PTR(value)                              \
  24         ({                                              \
  25                 LTTNG_ASSERT(value <= UINTPTR_MAX);             \
  26                 (void *) (uintptr_t) value;             \
  27         })
  28 #define PTR_TO_UINT(ptr) ((uintptr_t) ptr)
  29
  30 /*
  31  * Handle timer teardown race wrt memory free of private data by sessiond
  32  * signals are handled by a single thread, which permits a synchronization
  33  * point between handling of each signal. Internal lock ensures mutual
  34  * exclusion.
  35  */
  36 static
  37 struct timer_signal_data {
  38         /* Thread managing signals. */
  39         pthread_t tid;
  40         int qs_done;
  41         pthread_mutex_t lock;
  42 } timer_signal = {
  43         .tid = 0,
  44         .qs_done = 0,
  45         .lock = PTHREAD_MUTEX_INITIALIZER,
  46 };
  47
  48 /*
  49  * Set custom signal mask to current thread.
  50  */
  51 static
  52 void setmask(sigset_t *mask)
  53 {
  54         int ret;
  55
  56         ret = sigemptyset(mask);
  57         if (ret) {
  58                 PERROR("sigemptyset");
  59         }
  60         ret = sigaddset(mask, LTTNG_SESSIOND_SIG_QS);
  61         if (ret) {
  62                 PERROR("sigaddset teardown");
  63         }
  64         ret = sigaddset(mask, LTTNG_SESSIOND_SIG_EXIT);
  65         if (ret) {
  66                 PERROR("sigaddset exit");
  67         }
  68         ret = sigaddset(mask, LTTNG_SESSIOND_SIG_PENDING_ROTATION_CHECK);
  69         if (ret) {
  70                 PERROR("sigaddset pending rotation check");
  71         }
  72         ret = sigaddset(mask, LTTNG_SESSIOND_SIG_SCHEDULED_ROTATION);
  73         if (ret) {
  74                 PERROR("sigaddset scheduled rotation");
  75         }
  76 }
  77
  78 /*
  79  * This is the same function as timer_signal_thread_qs, when it
  80  * returns, it means that no timer signr is currently pending or being handled
  81  * by the timer thread. This cannot be called from the timer thread.
  82  */
  83 static
  84 void timer_signal_thread_qs(unsigned int signr)
  85 {
  86         sigset_t pending_set;
  87         int ret;
  88
  89         /*
  90          * We need to be the only thread interacting with the thread
  91          * that manages signals for teardown synchronization.
  92          */
  93         pthread_mutex_lock(&timer_signal.lock);
  94
  95         /* Ensure we don't have any signal queued for this session. */
  96         for (;;) {
  97                 ret = sigemptyset(&pending_set);
  98                 if (ret == -1) {
  99                         PERROR("sigemptyset");
 100                 }
 101                 ret = sigpending(&pending_set);
 102                 if (ret == -1) {
 103                         PERROR("sigpending");
 104                 }
 105                 if (!sigismember(&pending_set, signr)) {
 106                         break;
 107                 }
 108                 caa_cpu_relax();
 109         }
 110
 111         /*
 112          * From this point, no new signal handler will be fired that would try to
 113          * access "session". However, we still need to wait for any currently
 114          * executing handler to complete.
 115          */
 116         cmm_smp_mb();
 117         CMM_STORE_SHARED(timer_signal.qs_done, 0);
 118         cmm_smp_mb();
 119
 120         /*
 121          * Kill with LTTNG_SESSIOND_SIG_QS, so signal management thread
 122          * wakes up.
 123          */
 124         kill(getpid(), LTTNG_SESSIOND_SIG_QS);
 125
 126         while (!CMM_LOAD_SHARED(timer_signal.qs_done)) {
 127                 caa_cpu_relax();
 128         }
 129         cmm_smp_mb();
 130
 131         pthread_mutex_unlock(&timer_signal.lock);
 132 }
 133
 134 /*
 135  * Start a timer on a session that will fire at a given interval
 136  * (timer_interval_us) and fire a given signal (signal).
 137  *
 138  * Returns a negative value on error, 0 if a timer was created, and
 139  * a positive value if no timer was created (not an error).
 140  */
 141 static
 142 int timer_start(timer_t *timer_id, struct ltt_session *session,
 143                 unsigned int timer_interval_us, int signal, bool one_shot)
 144 {
 145         int ret = 0, delete_ret;
 146         struct sigevent sev = {};
 147         struct itimerspec its;
 148
 149         sev.sigev_notify = SIGEV_SIGNAL;
 150         sev.sigev_signo = signal;
 151         sev.sigev_value.sival_ptr = session;
 152         ret = timer_create(CLOCK_MONOTONIC, &sev, timer_id);
 153         if (ret == -1) {
 154                 PERROR("timer_create");
 155                 goto end;
 156         }
 157
 158         its.it_value.tv_sec = timer_interval_us / 1000000;
 159         its.it_value.tv_nsec = (timer_interval_us % 1000000) * 1000;
 160         if (one_shot) {
 161                 its.it_interval.tv_sec = 0;
 162                 its.it_interval.tv_nsec = 0;
 163         } else {
 164                 its.it_interval.tv_sec = its.it_value.tv_sec;
 165                 its.it_interval.tv_nsec = its.it_value.tv_nsec;
 166         }
 167
 168         ret = timer_settime(*timer_id, 0, &its, NULL);
 169         if (ret == -1) {
 170                 PERROR("timer_settime");
 171                 goto error_destroy_timer;
 172         }
 173         goto end;
 174
 175 error_destroy_timer:
 176         delete_ret = timer_delete(*timer_id);
 177         if (delete_ret == -1) {
 178                 PERROR("timer_delete");
 179         }
 180
 181 end:
 182         return ret;
 183 }
 184
 185 static
 186 int timer_stop(timer_t *timer_id, int signal)
 187 {
 188         int ret = 0;
 189
 190         ret = timer_delete(*timer_id);
 191         if (ret == -1) {
 192                 PERROR("timer_delete");
 193                 goto end;
 194         }
 195
 196         timer_signal_thread_qs(signal);
 197         *timer_id = 0;
 198 end:
 199         return ret;
 200 }
 201
 202 int timer_session_rotation_pending_check_start(struct ltt_session *session,
 203                 unsigned int interval_us)
 204 {
 205         int ret;
 206
 207         if (!session_get(session)) {
 208                 ret = -1;
 209                 goto end;
 210         }
 211         DBG("Enabling session rotation pending check timer on session %" PRIu64,
 212                         session->id);
 213         /*
 214          * We arm this timer in a one-shot mode so we don't have to disable it
 215          * explicitly (which could deadlock if the timer thread is blocked
 216          * writing in the rotation_timer_pipe).
 217          *
 218          * Instead, we re-arm it if needed after the rotation_pending check as
 219          * returned. Also, this timer is usually only needed once, so there is
 220          * no need to go through the whole signal teardown scheme everytime.
 221          */
 222         ret = timer_start(&session->rotation_pending_check_timer,
 223                         session, interval_us,
 224                         LTTNG_SESSIOND_SIG_PENDING_ROTATION_CHECK,
 225                         /* one-shot */ true);
 226         if (ret == 0) {
 227                 session->rotation_pending_check_timer_enabled = true;
 228         }
 229 end:
 230         return ret;
 231 }
 232
 233 /*
 234  * Call with session and session_list locks held.
 235  */
 236 int timer_session_rotation_pending_check_stop(struct ltt_session *session)
 237 {
 238         int ret;
 239
 240         LTTNG_ASSERT(session);
 241         LTTNG_ASSERT(session->rotation_pending_check_timer_enabled);
 242
 243         DBG("Disabling session rotation pending check timer on session %" PRIu64,
 244                         session->id);
 245         ret = timer_stop(&session->rotation_pending_check_timer,
 246                         LTTNG_SESSIOND_SIG_PENDING_ROTATION_CHECK);
 247         if (ret == -1) {
 248                 ERR("Failed to stop rotate_pending_check timer");
 249         } else {
 250                 session->rotation_pending_check_timer_enabled = false;
 251                 /*
 252                  * The timer's reference to the session can be released safely.
 253                  */
 254                 session_put(session);
 255         }
 256         return ret;
 257 }
 258
 259 /*
 260  * Call with session and session_list locks held.
 261  */
 262 int timer_session_rotation_schedule_timer_start(struct ltt_session *session,
 263                 unsigned int interval_us)
 264 {
 265         int ret;
 266
 267         if (!session_get(session)) {
 268                 ret = -1;
 269                 goto end;
 270         }
 271         DBG("Enabling scheduled rotation timer on session \"%s\" (%ui %s)", session->name,
 272                         interval_us, USEC_UNIT);
 273         ret = timer_start(&session->rotation_schedule_timer, session,
 274                         interval_us, LTTNG_SESSIOND_SIG_SCHEDULED_ROTATION,
 275                         /* one-shot */ false);
 276         if (ret < 0) {
 277                 goto end;
 278         }
 279         session->rotation_schedule_timer_enabled = true;
 280 end:
 281         return ret;
 282 }
 283
 284 /*
 285  * Call with session and session_list locks held.
 286  */
 287 int timer_session_rotation_schedule_timer_stop(struct ltt_session *session)
 288 {
 289         int ret = 0;
 290
 291         LTTNG_ASSERT(session);
 292
 293         if (!session->rotation_schedule_timer_enabled) {
 294                 goto end;
 295         }
 296
 297         DBG("Disabling scheduled rotation timer on session %s", session->name);
 298         ret = timer_stop(&session->rotation_schedule_timer,
 299                         LTTNG_SESSIOND_SIG_SCHEDULED_ROTATION);
 300         if (ret < 0) {
 301                 ERR("Failed to stop scheduled rotation timer of session \"%s\"",
 302                                 session->name);
 303                 goto end;
 304         }
 305
 306         session->rotation_schedule_timer_enabled = false;
 307         /* The timer's reference to the session can be released safely. */
 308         session_put(session);
 309         ret = 0;
 310 end:
 311         return ret;
 312 }
 313
 314 /*
 315  * Block the RT signals for the entire process. It must be called from the
 316  * sessiond main before creating the threads
 317  */
 318 int timer_signal_init(void)
 319 {
 320         int ret;
 321         sigset_t mask;
 322
 323         /* Block signal for entire process, so only our thread processes it. */
 324         setmask(&mask);
 325         ret = pthread_sigmask(SIG_BLOCK, &mask, NULL);
 326         if (ret) {
 327                 errno = ret;
 328                 PERROR("pthread_sigmask");
 329                 return -1;
 330         }
 331         return 0;
 332 }
 333
 334 /*
 335  * This thread is the sighandler for the timer signals.
 336  */
 337 static
 338 void *thread_timer(void *data)
 339 {
 340         int signr;
 341         sigset_t mask;
 342         siginfo_t info;
 343         struct timer_thread_parameters *ctx = (timer_thread_parameters *) data;
 344
 345         rcu_register_thread();
 346         rcu_thread_online();
 347
 348         health_register(the_health_sessiond, HEALTH_SESSIOND_TYPE_TIMER);
 349         health_code_update();
 350
 351         /* Only self thread will receive signal mask. */
 352         setmask(&mask);
 353         CMM_STORE_SHARED(timer_signal.tid, pthread_self());
 354
 355         while (1) {
 356                 health_code_update();
 357
 358                 health_poll_entry();
 359                 signr = sigwaitinfo(&mask, &info);
 360                 health_poll_exit();
 361
 362                 /*
 363                  * NOTE: cascading conditions are used instead of a switch case
 364                  * since the use of SIGRTMIN in the definition of the signals'
 365                  * values prevents the reduction to an integer constant.
 366                  */
 367                 if (signr == -1) {
 368                         if (errno != EINTR) {
 369                                 PERROR("sigwaitinfo");
 370                         }
 371                         continue;
 372                 } else if (signr == LTTNG_SESSIOND_SIG_QS) {
 373                         cmm_smp_mb();
 374                         CMM_STORE_SHARED(timer_signal.qs_done, 1);
 375                         cmm_smp_mb();
 376                 } else if (signr == LTTNG_SESSIOND_SIG_EXIT) {
 377                         goto end;
 378                 } else if (signr == LTTNG_SESSIOND_SIG_PENDING_ROTATION_CHECK) {
 379                         struct ltt_session *session =
 380                                         (struct ltt_session *) info.si_value.sival_ptr;
 381
 382                         rotation_thread_enqueue_job(ctx->rotation_thread_job_queue,
 383                                         ROTATION_THREAD_JOB_TYPE_CHECK_PENDING_ROTATION,
 384                                         session);
 385                 } else if (signr == LTTNG_SESSIOND_SIG_SCHEDULED_ROTATION) {
 386                         rotation_thread_enqueue_job(ctx->rotation_thread_job_queue,
 387                                         ROTATION_THREAD_JOB_TYPE_SCHEDULED_ROTATION,
 388                                         (struct ltt_session *) info.si_value.sival_ptr);
 389                         /*
 390                          * The scheduled periodic rotation timer is not in
 391                          * "one-shot" mode. The reference to the session is not
 392                          * released since the timer is still enabled and can
 393                          * still fire.
 394                          */
 395                 } else {
 396                         ERR("Unexpected signal %d", info.si_signo);
 397                 }
 398         }
 399
 400 end:
 401         DBG("Thread exit");
 402         health_unregister(the_health_sessiond);
 403         rcu_thread_offline();
 404         rcu_unregister_thread();
 405         return NULL;
 406 }
 407
 408 static
 409 bool shutdown_timer_thread(void *data)
 410 {
 411         return kill(getpid(), LTTNG_SESSIOND_SIG_EXIT) == 0;
 412 }
 413
 414 bool launch_timer_thread(
 415                 struct timer_thread_parameters *timer_thread_parameters)
 416 {
 417         struct lttng_thread *thread;
 418
 419         thread = lttng_thread_create("Timer",
 420                         thread_timer,
 421                         shutdown_timer_thread,
 422                         NULL,
 423                         timer_thread_parameters);
 424         if (!thread) {
 425                 goto error;
 426         }
 427         lttng_thread_put(thread);
 428         return true;
 429 error:
 430         return false;
 431 }