src/side.c

   1 // SPDX-License-Identifier: MIT
   2 /*
   3  * Copyright 2022 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   4  */
   5
   6 #include <side/trace.h>
   7 #include <string.h>
   8 #include <assert.h>
   9 #include <pthread.h>
  10 #include <unistd.h>
  11
  12 #include "rcu.h"
  13 #include "list.h"
  14 #include "rculist.h"
  15
  16 /* Top 8 bits reserved for shared tracer use. */
  17 #if SIDE_BITS_PER_LONG == 64
  18 # define SIDE_EVENT_ENABLED_SHARED_MASK                 0xFF00000000000000ULL
  19 # define SIDE_EVENT_ENABLED_SHARED_USER_EVENT_MASK      0x8000000000000000ULL
  20 # define SIDE_EVENT_ENABLED_SHARED_PTRACE_MASK          0x4000000000000000ULL
  21
  22 /* Allow 2^56 private tracer references on an event. */
  23 # define SIDE_EVENT_ENABLED_PRIVATE_MASK                0x00FFFFFFFFFFFFFFULL
  24 #else
  25 # define SIDE_EVENT_ENABLED_SHARED_MASK                 0xFF000000UL
  26 # define SIDE_EVENT_ENABLED_SHARED_USER_EVENT_MASK      0x80000000UL
  27 # define SIDE_EVENT_ENABLED_SHARED_PTRACE_MASK          0x40000000UL
  28
  29 /* Allow 2^24 private tracer references on an event. */
  30 # define SIDE_EVENT_ENABLED_PRIVATE_MASK                0x00FFFFFFUL
  31 #endif
  32
  33 #define SIDE_KEY_RESERVED_RANGE_END                     0x8
  34
  35 /* Key 0x0 is reserved to match all. */
  36 #define SIDE_KEY_MATCH_ALL                              0x0
  37 /* Key 0x1 is reserved for user event. */
  38 #define SIDE_KEY_USER_EVENT                             0x1
  39 /* Key 0x2 is reserved for ptrace. */
  40 #define SIDE_KEY_PTRACE                                 0x2
  41
  42 struct side_events_register_handle {
  43         struct side_list_node node;
  44         struct side_event_description **events;
  45         uint32_t nr_events;
  46 };
  47
  48 struct side_tracer_handle {
  49         struct side_list_node node;
  50         void (*cb)(enum side_tracer_notification notif,
  51                 struct side_event_description **events, uint32_t nr_events, void *priv);
  52         void *priv;
  53 };
  54
  55 struct side_statedump_notification {
  56         struct side_list_node node;
  57         uint64_t key;
  58 };
  59
  60 struct side_statedump_request_handle {
  61         struct side_list_node node;                     /* Statedump request RCU list node. */
  62         struct side_list_head notification_queue;       /* Queue of struct side_statedump_notification */
  63         void (*cb)(void);
  64         char *name;
  65         enum side_statedump_mode mode;
  66 };
  67
  68 struct side_callback {
  69         union {
  70                 void (*call)(const struct side_event_description *desc,
  71                         const struct side_arg_vec *side_arg_vec,
  72                         void *priv);
  73                 void (*call_variadic)(const struct side_event_description *desc,
  74                         const struct side_arg_vec *side_arg_vec,
  75                         const struct side_arg_dynamic_struct *var_struct,
  76                         void *priv);
  77         } u;
  78         void *priv;
  79         uint64_t key;
  80 };
  81
  82 enum agent_thread_state {
  83         AGENT_THREAD_STATE_BLOCKED = 0,
  84         AGENT_THREAD_STATE_HANDLE_REQUEST = (1 << 0),
  85         AGENT_THREAD_STATE_EXIT = (1 << 1),
  86 };
  87
  88 struct statedump_agent_thread {
  89         long ref;
  90         pthread_t id;
  91         enum agent_thread_state state;
  92 };
  93
  94 static struct side_rcu_gp_state event_rcu_gp, statedump_rcu_gp;
  95
  96 /*
  97  * Lazy initialization for early use within library constructors.
  98  */
  99 static bool initialized;
 100 /*
 101  * Do not register/unregister any more events after destructor.
 102  */
 103 static bool finalized;
 104
 105 /*
 106  * Recursive mutex to allow tracer callbacks to use the side API.
 107  */
 108 static pthread_mutex_t side_event_lock = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
 109 static pthread_mutex_t side_statedump_lock = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
 110 static pthread_mutex_t side_key_lock = PTHREAD_MUTEX_INITIALIZER;
 111
 112 /* Dynamic tracer key allocation. */
 113 static uint64_t side_key_next = SIDE_KEY_RESERVED_RANGE_END;
 114
 115 static struct statedump_agent_thread statedump_agent_thread;
 116
 117 static DEFINE_SIDE_LIST_HEAD(side_events_list);
 118 static DEFINE_SIDE_LIST_HEAD(side_tracer_list);
 119
 120 /*
 121  * The statedump request list is a RCU list to allow the agent thread to
 122  * iterate over this list with a RCU read-side lock.
 123  */
 124 static DEFINE_SIDE_LIST_HEAD(side_statedump_list);
 125
 126 /*
 127  * Callback filter key for state dump.
 128  */
 129 static __thread uint64_t filter_key = SIDE_KEY_MATCH_ALL;
 130
 131 /*
 132  * The empty callback has a NULL function callback pointer, which stops
 133  * iteration on the array of callbacks immediately.
 134  */
 135 const char side_empty_callback[sizeof(struct side_callback)];
 136
 137 side_static_event(side_statedump_begin, "side", "statedump_begin",
 138         SIDE_LOGLEVEL_INFO, side_field_list(side_field_string("name")));
 139 side_static_event(side_statedump_end, "side", "statedump_end",
 140         SIDE_LOGLEVEL_INFO, side_field_list(side_field_string("name")));
 141
 142 /*
 143  * side_ptrace_hook is a place holder for a debugger breakpoint.
 144  * var_struct is NULL if not variadic.
 145  */
 146 void side_ptrace_hook(const struct side_event_state *event_state __attribute__((unused)),
 147         const struct side_arg_vec *side_arg_vec __attribute__((unused)),
 148         const struct side_arg_dynamic_struct *var_struct __attribute__((unused)))
 149         __attribute__((noinline));
 150 void side_ptrace_hook(const struct side_event_state *event_state __attribute__((unused)),
 151         const struct side_arg_vec *side_arg_vec __attribute__((unused)),
 152         const struct side_arg_dynamic_struct *var_struct __attribute__((unused)))
 153 {
 154 }
 155
 156 static
 157 void _side_call(const struct side_event_state *event_state, const struct side_arg_vec *side_arg_vec, uint64_t key)
 158 {
 159         struct side_rcu_read_state rcu_read_state;
 160         const struct side_event_state_0 *es0;
 161         const struct side_callback *side_cb;
 162         uintptr_t enabled;
 163
 164         if (side_unlikely(finalized))
 165                 return;
 166         if (side_unlikely(!initialized))
 167                 side_init();
 168         if (side_unlikely(event_state->version != 0))
 169                 abort();
 170         es0 = side_container_of(event_state, const struct side_event_state_0, parent);
 171         assert(!(es0->desc->flags & SIDE_EVENT_FLAG_VARIADIC));
 172         enabled = __atomic_load_n(&es0->enabled, __ATOMIC_RELAXED);
 173         if (side_unlikely(enabled & SIDE_EVENT_ENABLED_SHARED_MASK)) {
 174                 if ((enabled & SIDE_EVENT_ENABLED_SHARED_USER_EVENT_MASK) &&
 175                     (key == SIDE_KEY_MATCH_ALL || key == SIDE_KEY_USER_EVENT)) {
 176                         // TODO: call kernel write.
 177                 }
 178                 if ((enabled & SIDE_EVENT_ENABLED_SHARED_PTRACE_MASK) &&
 179                     (key == SIDE_KEY_MATCH_ALL || key == SIDE_KEY_PTRACE))
 180                         side_ptrace_hook(event_state, side_arg_vec, NULL);
 181         }
 182         side_rcu_read_begin(&event_rcu_gp, &rcu_read_state);
 183         for (side_cb = side_rcu_dereference(es0->callbacks); side_cb->u.call != NULL; side_cb++) {
 184                 if (key != SIDE_KEY_MATCH_ALL && side_cb->key != SIDE_KEY_MATCH_ALL && side_cb->key != key)
 185                         continue;
 186                 side_cb->u.call(es0->desc, side_arg_vec, side_cb->priv);
 187         }
 188         side_rcu_read_end(&event_rcu_gp, &rcu_read_state);
 189 }
 190
 191 void side_call(const struct side_event_state *event_state, const struct side_arg_vec *side_arg_vec)
 192 {
 193         _side_call(event_state, side_arg_vec, SIDE_KEY_MATCH_ALL);
 194 }
 195
 196 void side_statedump_call(const struct side_event_state *event_state, const struct side_arg_vec *side_arg_vec)
 197 {
 198         _side_call(event_state, side_arg_vec, filter_key);
 199 }
 200
 201 static
 202 void _side_call_variadic(const struct side_event_state *event_state,
 203         const struct side_arg_vec *side_arg_vec,
 204         const struct side_arg_dynamic_struct *var_struct,
 205         uint64_t key)
 206 {
 207         struct side_rcu_read_state rcu_read_state;
 208         const struct side_event_state_0 *es0;
 209         const struct side_callback *side_cb;
 210         uintptr_t enabled;
 211
 212         if (side_unlikely(finalized))
 213                 return;
 214         if (side_unlikely(!initialized))
 215                 side_init();
 216         if (side_unlikely(event_state->version != 0))
 217                 abort();
 218         es0 = side_container_of(event_state, const struct side_event_state_0, parent);
 219         assert(es0->desc->flags & SIDE_EVENT_FLAG_VARIADIC);
 220         enabled = __atomic_load_n(&es0->enabled, __ATOMIC_RELAXED);
 221         if (side_unlikely(enabled & SIDE_EVENT_ENABLED_SHARED_MASK)) {
 222                 if ((enabled & SIDE_EVENT_ENABLED_SHARED_USER_EVENT_MASK) &&
 223                     (key == SIDE_KEY_MATCH_ALL || key == SIDE_KEY_USER_EVENT)) {
 224                         // TODO: call kernel write.
 225                 }
 226                 if ((enabled & SIDE_EVENT_ENABLED_SHARED_PTRACE_MASK) &&
 227                     (key == SIDE_KEY_MATCH_ALL || key == SIDE_KEY_PTRACE))
 228                         side_ptrace_hook(event_state, side_arg_vec, var_struct);
 229         }
 230         side_rcu_read_begin(&event_rcu_gp, &rcu_read_state);
 231         for (side_cb = side_rcu_dereference(es0->callbacks); side_cb->u.call_variadic != NULL; side_cb++) {
 232                 if (key != SIDE_KEY_MATCH_ALL && side_cb->key != SIDE_KEY_MATCH_ALL && side_cb->key != key)
 233                         continue;
 234                 side_cb->u.call_variadic(es0->desc, side_arg_vec, var_struct, side_cb->priv);
 235         }
 236         side_rcu_read_end(&event_rcu_gp, &rcu_read_state);
 237 }
 238
 239 void side_call_variadic(const struct side_event_state *event_state,
 240         const struct side_arg_vec *side_arg_vec,
 241         const struct side_arg_dynamic_struct *var_struct)
 242 {
 243         _side_call_variadic(event_state, side_arg_vec, var_struct, SIDE_KEY_MATCH_ALL);
 244 }
 245
 246 void side_statedump_call_variadic(const struct side_event_state *event_state,
 247         const struct side_arg_vec *side_arg_vec,
 248         const struct side_arg_dynamic_struct *var_struct)
 249 {
 250         _side_call_variadic(event_state, side_arg_vec, var_struct, filter_key);
 251 }
 252
 253 static
 254 const struct side_callback *side_tracer_callback_lookup(
 255                 const struct side_event_description *desc,
 256                 void *call, void *priv, uint64_t key)
 257 {
 258         struct side_event_state *event_state = side_ptr_get(desc->state);
 259         const struct side_event_state_0 *es0;
 260         const struct side_callback *cb;
 261
 262         if (side_unlikely(event_state->version != 0))
 263                 abort();
 264         es0 = side_container_of(event_state, const struct side_event_state_0, parent);
 265         for (cb = es0->callbacks; cb->u.call != NULL; cb++) {
 266                 if ((void *) cb->u.call == call && cb->priv == priv && cb->key == key)
 267                         return cb;
 268         }
 269         return NULL;
 270 }
 271
 272 static
 273 int _side_tracer_callback_register(struct side_event_description *desc,
 274                 void *call, void *priv, uint64_t key)
 275 {
 276         struct side_event_state *event_state;
 277         struct side_callback *old_cb, *new_cb;
 278         struct side_event_state_0 *es0;
 279         int ret = SIDE_ERROR_OK;
 280         uint32_t old_nr_cb;
 281
 282         if (!call)
 283                 return SIDE_ERROR_INVAL;
 284         if (finalized)
 285                 return SIDE_ERROR_EXITING;
 286         if (!initialized)
 287                 side_init();
 288         pthread_mutex_lock(&side_event_lock);
 289         event_state = side_ptr_get(desc->state);
 290         if (side_unlikely(event_state->version != 0))
 291                 abort();
 292         es0 = side_container_of(event_state, struct side_event_state_0, parent);
 293         old_nr_cb = es0->nr_callbacks;
 294         if (old_nr_cb == UINT32_MAX) {
 295                 ret = SIDE_ERROR_INVAL;
 296                 goto unlock;
 297         }
 298         /* Reject duplicate (call, priv) tuples. */
 299         if (side_tracer_callback_lookup(desc, call, priv, key)) {
 300                 ret = SIDE_ERROR_EXIST;
 301                 goto unlock;
 302         }
 303         old_cb = (struct side_callback *) es0->callbacks;
 304         /* old_nr_cb + 1 (new cb) + 1 (NULL) */
 305         new_cb = (struct side_callback *) calloc(old_nr_cb + 2, sizeof(struct side_callback));
 306         if (!new_cb) {
 307                 ret = SIDE_ERROR_NOMEM;
 308                 goto unlock;
 309         }
 310         memcpy(new_cb, old_cb, old_nr_cb);
 311         if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
 312                 new_cb[old_nr_cb].u.call_variadic =
 313                         (side_tracer_callback_variadic_func) call;
 314         else
 315                 new_cb[old_nr_cb].u.call =
 316                         (side_tracer_callback_func) call;
 317         new_cb[old_nr_cb].priv = priv;
 318         new_cb[old_nr_cb].key = key;
 319         /* High order bits are already zeroed. */
 320         side_rcu_assign_pointer(es0->callbacks, new_cb);
 321         side_rcu_wait_grace_period(&event_rcu_gp);
 322         if (old_nr_cb)
 323                 free(old_cb);
 324         es0->nr_callbacks++;
 325         /* Increment concurrently with kernel setting the top bits. */
 326         if (!old_nr_cb)
 327                 (void) __atomic_add_fetch(&es0->enabled, 1, __ATOMIC_RELAXED);
 328 unlock:
 329         pthread_mutex_unlock(&side_event_lock);
 330         return ret;
 331 }
 332
 333 int side_tracer_callback_register(struct side_event_description *desc,
 334                 side_tracer_callback_func call,
 335                 void *priv, uint64_t key)
 336 {
 337         if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
 338                 return SIDE_ERROR_INVAL;
 339         return _side_tracer_callback_register(desc, (void *) call, priv, key);
 340 }
 341
 342 int side_tracer_callback_variadic_register(struct side_event_description *desc,
 343                 side_tracer_callback_variadic_func call_variadic,
 344                 void *priv, uint64_t key)
 345 {
 346         if (!(desc->flags & SIDE_EVENT_FLAG_VARIADIC))
 347                 return SIDE_ERROR_INVAL;
 348         return _side_tracer_callback_register(desc, (void *) call_variadic, priv, key);
 349 }
 350
 351 static int _side_tracer_callback_unregister(struct side_event_description *desc,
 352                 void *call, void *priv, uint64_t key)
 353 {
 354         struct side_event_state *event_state;
 355         struct side_callback *old_cb, *new_cb;
 356         const struct side_callback *cb_pos;
 357         struct side_event_state_0 *es0;
 358         uint32_t pos_idx;
 359         int ret = SIDE_ERROR_OK;
 360         uint32_t old_nr_cb;
 361
 362         if (!call)
 363                 return SIDE_ERROR_INVAL;
 364         if (finalized)
 365                 return SIDE_ERROR_EXITING;
 366         if (!initialized)
 367                 side_init();
 368         pthread_mutex_lock(&side_event_lock);
 369         event_state = side_ptr_get(desc->state);
 370         if (side_unlikely(event_state->version != 0))
 371                 abort();
 372         es0 = side_container_of(event_state, struct side_event_state_0, parent);
 373         cb_pos = side_tracer_callback_lookup(desc, call, priv, key);
 374         if (!cb_pos) {
 375                 ret = SIDE_ERROR_NOENT;
 376                 goto unlock;
 377         }
 378         old_nr_cb = es0->nr_callbacks;
 379         old_cb = (struct side_callback *) es0->callbacks;
 380         if (old_nr_cb == 1) {
 381                 new_cb = (struct side_callback *) &side_empty_callback;
 382         } else {
 383                 pos_idx = cb_pos - es0->callbacks;
 384                 /* Remove entry at pos_idx. */
 385                 /* old_nr_cb - 1 (removed cb) + 1 (NULL) */
 386                 new_cb = (struct side_callback *) calloc(old_nr_cb, sizeof(struct side_callback));
 387                 if (!new_cb) {
 388                         ret = SIDE_ERROR_NOMEM;
 389                         goto unlock;
 390                 }
 391                 memcpy(new_cb, old_cb, pos_idx);
 392                 memcpy(&new_cb[pos_idx], &old_cb[pos_idx + 1], old_nr_cb - pos_idx - 1);
 393         }
 394         /* High order bits are already zeroed. */
 395         side_rcu_assign_pointer(es0->callbacks, new_cb);
 396         side_rcu_wait_grace_period(&event_rcu_gp);
 397         free(old_cb);
 398         es0->nr_callbacks--;
 399         /* Decrement concurrently with kernel setting the top bits. */
 400         if (old_nr_cb == 1)
 401                 (void) __atomic_add_fetch(&es0->enabled, -1, __ATOMIC_RELAXED);
 402 unlock:
 403         pthread_mutex_unlock(&side_event_lock);
 404         return ret;
 405 }
 406
 407 int side_tracer_callback_unregister(struct side_event_description *desc,
 408                 side_tracer_callback_func call,
 409                 void *priv, uint64_t key)
 410 {
 411         if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
 412                 return SIDE_ERROR_INVAL;
 413         return _side_tracer_callback_unregister(desc, (void *) call, priv, key);
 414 }
 415
 416 int side_tracer_callback_variadic_unregister(struct side_event_description *desc,
 417                 side_tracer_callback_variadic_func call_variadic,
 418                 void *priv, uint64_t key)
 419 {
 420         if (!(desc->flags & SIDE_EVENT_FLAG_VARIADIC))
 421                 return SIDE_ERROR_INVAL;
 422         return _side_tracer_callback_unregister(desc, (void *) call_variadic, priv, key);
 423 }
 424
 425 struct side_events_register_handle *side_events_register(struct side_event_description **events, uint32_t nr_events)
 426 {
 427         struct side_events_register_handle *events_handle = NULL;
 428         struct side_tracer_handle *tracer_handle;
 429
 430         if (finalized)
 431                 return NULL;
 432         if (!initialized)
 433                 side_init();
 434         events_handle = (struct side_events_register_handle *)
 435                         calloc(1, sizeof(struct side_events_register_handle));
 436         if (!events_handle)
 437                 return NULL;
 438         events_handle->events = events;
 439         events_handle->nr_events = nr_events;
 440
 441         pthread_mutex_lock(&side_event_lock);
 442         side_list_insert_node_tail(&side_events_list, &events_handle->node);
 443         side_list_for_each_entry(tracer_handle, &side_tracer_list, node) {
 444                 tracer_handle->cb(SIDE_TRACER_NOTIFICATION_INSERT_EVENTS,
 445                         events, nr_events, tracer_handle->priv);
 446         }
 447         pthread_mutex_unlock(&side_event_lock);
 448         //TODO: call event batch register ioctl
 449         return events_handle;
 450 }
 451
 452 static
 453 void side_event_remove_callbacks(struct side_event_description *desc)
 454 {
 455         struct side_event_state *event_state = side_ptr_get(desc->state);
 456         struct side_event_state_0 *es0;
 457         struct side_callback *old_cb;
 458         uint32_t nr_cb;
 459
 460         if (side_unlikely(event_state->version != 0))
 461                 abort();
 462         es0 = side_container_of(event_state, struct side_event_state_0, parent);
 463         nr_cb = es0->nr_callbacks;
 464         if (!nr_cb)
 465                 return;
 466         old_cb = (struct side_callback *) es0->callbacks;
 467         (void) __atomic_add_fetch(&es0->enabled, -1, __ATOMIC_RELAXED);
 468         /*
 469          * Setting the state back to 0 cb and empty callbacks out of
 470          * caution. This should not matter because instrumentation is
 471          * unreachable.
 472          */
 473         es0->nr_callbacks = 0;
 474         side_rcu_assign_pointer(es0->callbacks, &side_empty_callback);
 475         /*
 476          * No need to wait for grace period because instrumentation is
 477          * unreachable.
 478          */
 479         free(old_cb);
 480 }
 481
 482 /*
 483  * Unregister event handle. At this point, all side events in that
 484  * handle should be unreachable.
 485  */
 486 void side_events_unregister(struct side_events_register_handle *events_handle)
 487 {
 488         struct side_tracer_handle *tracer_handle;
 489         uint32_t i;
 490
 491         if (!events_handle)
 492                 return;
 493         if (finalized)
 494                 return;
 495         if (!initialized)
 496                 side_init();
 497         pthread_mutex_lock(&side_event_lock);
 498         side_list_remove_node(&events_handle->node);
 499         side_list_for_each_entry(tracer_handle, &side_tracer_list, node) {
 500                 tracer_handle->cb(SIDE_TRACER_NOTIFICATION_REMOVE_EVENTS,
 501                         events_handle->events, events_handle->nr_events,
 502                         tracer_handle->priv);
 503         }
 504         for (i = 0; i < events_handle->nr_events; i++) {
 505                 struct side_event_description *event = events_handle->events[i];
 506
 507                 /* Skip NULL pointers */
 508                 if (!event)
 509                         continue;
 510                 side_event_remove_callbacks(event);
 511         }
 512         pthread_mutex_unlock(&side_event_lock);
 513         //TODO: call event batch unregister ioctl
 514         free(events_handle);
 515 }
 516
 517 struct side_tracer_handle *side_tracer_event_notification_register(
 518                 void (*cb)(enum side_tracer_notification notif,
 519                         struct side_event_description **events, uint32_t nr_events, void *priv),
 520                 void *priv)
 521 {
 522         struct side_tracer_handle *tracer_handle;
 523         struct side_events_register_handle *events_handle;
 524
 525         if (finalized)
 526                 return NULL;
 527         if (!initialized)
 528                 side_init();
 529         tracer_handle = (struct side_tracer_handle *)
 530                                 calloc(1, sizeof(struct side_tracer_handle));
 531         if (!tracer_handle)
 532                 return NULL;
 533         pthread_mutex_lock(&side_event_lock);
 534         tracer_handle->cb = cb;
 535         tracer_handle->priv = priv;
 536         side_list_insert_node_tail(&side_tracer_list, &tracer_handle->node);
 537         side_list_for_each_entry(events_handle, &side_events_list, node) {
 538                 cb(SIDE_TRACER_NOTIFICATION_INSERT_EVENTS,
 539                         events_handle->events, events_handle->nr_events, priv);
 540         }
 541         pthread_mutex_unlock(&side_event_lock);
 542         return tracer_handle;
 543 }
 544
 545 void side_tracer_event_notification_unregister(struct side_tracer_handle *tracer_handle)
 546 {
 547         struct side_events_register_handle *events_handle;
 548
 549         if (finalized)
 550                 return;
 551         if (!initialized)
 552                 side_init();
 553         pthread_mutex_lock(&side_event_lock);
 554         side_list_for_each_entry(events_handle, &side_events_list, node) {
 555                 tracer_handle->cb(SIDE_TRACER_NOTIFICATION_REMOVE_EVENTS,
 556                         events_handle->events, events_handle->nr_events,
 557                         tracer_handle->priv);
 558         }
 559         side_list_remove_node(&tracer_handle->node);
 560         pthread_mutex_unlock(&side_event_lock);
 561         free(tracer_handle);
 562 }
 563
 564 /* Called with side_statedump_lock held. */
 565 static
 566 void queue_statedump_pending(struct side_statedump_request_handle *handle, uint64_t key)
 567 {
 568         struct side_statedump_notification *notif;
 569
 570         notif = (struct side_statedump_notification *) calloc(1, sizeof(struct side_statedump_notification));
 571         if (!notif)
 572                 abort();
 573         notif->key = key;
 574         side_list_insert_node_tail(&handle->notification_queue, &notif->node);
 575         if (handle->mode == SIDE_STATEDUMP_MODE_AGENT_THREAD)
 576                 (void)__atomic_or_fetch(&statedump_agent_thread.state, AGENT_THREAD_STATE_HANDLE_REQUEST, __ATOMIC_SEQ_CST);
 577 }
 578
 579 /* Called with side_statedump_lock held. */
 580 static
 581 void unqueue_statedump_pending(struct side_statedump_request_handle *handle, uint64_t key)
 582 {
 583         struct side_statedump_notification *notif, *tmp;
 584
 585         side_list_for_each_entry_safe(notif, tmp, &handle->notification_queue, node) {
 586                 if (key == SIDE_KEY_MATCH_ALL || key == notif->key) {
 587                         side_list_remove_node(&notif->node);
 588                         free(notif);
 589                 }
 590         }
 591 }
 592
 593 static
 594 void side_statedump_run(struct side_statedump_request_handle *handle,
 595                 struct side_statedump_notification *notif)
 596 {
 597         /* Invoke the state dump callback specifically for the tracer key. */
 598         filter_key = notif->key;
 599         side_statedump_event_call(side_statedump_begin,
 600                 side_arg_list(side_arg_string(handle->name)));
 601         handle->cb();
 602         side_statedump_event_call(side_statedump_end,
 603                 side_arg_list(side_arg_string(handle->name)));
 604         filter_key = SIDE_KEY_MATCH_ALL;
 605 }
 606
 607 static
 608 void _side_statedump_run_pending_requests(struct side_statedump_request_handle *handle)
 609 {
 610         struct side_statedump_notification *notif, *tmp;
 611         DEFINE_SIDE_LIST_HEAD(tmp_head);
 612
 613         pthread_mutex_lock(&side_statedump_lock);
 614         side_list_splice(&handle->notification_queue, &tmp_head);
 615         side_list_head_init(&handle->notification_queue);
 616         pthread_mutex_unlock(&side_statedump_lock);
 617
 618         /* We are now sole owner of the tmp_head list. */
 619         side_list_for_each_entry(notif, &tmp_head, node)
 620                 side_statedump_run(handle, notif);
 621         side_list_for_each_entry_safe(notif, tmp, &tmp_head, node)
 622                 free(notif);
 623 }
 624
 625 static
 626 void *statedump_agent_func(void *arg __attribute__((unused)))
 627 {
 628         for (;;) {
 629                 struct side_statedump_request_handle *handle;
 630                 struct side_rcu_read_state rcu_read_state;
 631                 enum agent_thread_state state;
 632
 633                 /* TODO: futex-based wakeup. */
 634                 state = __atomic_load_n(&statedump_agent_thread.state, __ATOMIC_SEQ_CST);
 635                 if (state == AGENT_THREAD_STATE_BLOCKED) {
 636                         sleep(1);
 637                         continue;
 638                 }
 639                 if (state & AGENT_THREAD_STATE_EXIT)
 640                         break;
 641                 (void)__atomic_and_fetch(&statedump_agent_thread.state, ~AGENT_THREAD_STATE_HANDLE_REQUEST, __ATOMIC_SEQ_CST);
 642                 side_rcu_read_begin(&statedump_rcu_gp, &rcu_read_state);
 643                 side_list_for_each_entry_rcu(handle, &side_statedump_list, node)
 644                         _side_statedump_run_pending_requests(handle);
 645                 side_rcu_read_end(&statedump_rcu_gp, &rcu_read_state);
 646         }
 647         return NULL;
 648 }
 649
 650 static
 651 void statedump_agent_thread_get(void)
 652 {
 653         int ret;
 654
 655         if (statedump_agent_thread.ref++)
 656                 return;
 657         statedump_agent_thread.state = AGENT_THREAD_STATE_BLOCKED;
 658         ret = pthread_create(&statedump_agent_thread.id, NULL,
 659                         statedump_agent_func, NULL);
 660         if (ret) {
 661                 abort();
 662         }
 663 }
 664
 665 static
 666 void statedump_agent_thread_put(void)
 667 {
 668         int ret;
 669         void *retval;
 670
 671         if (--statedump_agent_thread.ref)
 672                 return;
 673         (void)__atomic_or_fetch(&statedump_agent_thread.state, AGENT_THREAD_STATE_EXIT, __ATOMIC_SEQ_CST);
 674         ret = pthread_join(statedump_agent_thread.id, &retval);
 675         if (ret) {
 676                 abort();
 677         }
 678         statedump_agent_thread.state = AGENT_THREAD_STATE_BLOCKED;
 679 }
 680
 681 struct side_statedump_request_handle *
 682         side_statedump_request_notification_register(const char *state_name,
 683                 void (*statedump_cb)(void),
 684                 enum side_statedump_mode mode)
 685 {
 686         struct side_statedump_request_handle *handle;
 687         char *name;
 688
 689         if (finalized)
 690                 return NULL;
 691         if (!initialized)
 692                 side_init();
 693         /*
 694          * The statedump request notification should not be registered
 695          * from a notification callback.
 696          */
 697         assert(!filter_key);
 698         handle = (struct side_statedump_request_handle *)
 699                                 calloc(1, sizeof(struct side_statedump_request_handle));
 700         if (!handle)
 701                 return NULL;
 702         name = strdup(state_name);
 703         if (!name)
 704                 goto name_nomem;
 705         handle->cb = statedump_cb;
 706         handle->name = name;
 707         handle->mode = mode;
 708         side_list_head_init(&handle->notification_queue);
 709
 710         pthread_mutex_lock(&side_statedump_lock);
 711         if (mode == SIDE_STATEDUMP_MODE_AGENT_THREAD)
 712                 statedump_agent_thread_get();
 713         side_list_insert_node_tail_rcu(&side_statedump_list, &handle->node);
 714         /* Queue statedump pending for all tracers. */
 715         queue_statedump_pending(handle, SIDE_KEY_MATCH_ALL);
 716         pthread_mutex_unlock(&side_statedump_lock);
 717
 718         if (mode == SIDE_STATEDUMP_MODE_AGENT_THREAD) {
 719                 for (;;) {
 720                         bool is_empty;
 721
 722                         /* TODO futex based wakeup. */
 723                         pthread_mutex_lock(&side_statedump_lock);
 724                         is_empty = side_list_empty(&handle->notification_queue);
 725                         pthread_mutex_unlock(&side_statedump_lock);
 726                         if (is_empty)
 727                                 break;
 728                         sleep(1);
 729                 }
 730         }
 731
 732         return handle;
 733
 734 name_nomem:
 735         free(handle);
 736         return NULL;
 737 }
 738
 739 void side_statedump_request_notification_unregister(struct side_statedump_request_handle *handle)
 740 {
 741         if (finalized)
 742                 return;
 743         if (!initialized)
 744                 side_init();
 745         assert(!filter_key);
 746
 747         pthread_mutex_lock(&side_statedump_lock);
 748         unqueue_statedump_pending(handle, SIDE_KEY_MATCH_ALL);
 749         side_list_remove_node_rcu(&handle->node);
 750         if (handle->mode == SIDE_STATEDUMP_MODE_AGENT_THREAD)
 751                 statedump_agent_thread_put();
 752         pthread_mutex_unlock(&side_statedump_lock);
 753
 754         side_rcu_wait_grace_period(&statedump_rcu_gp);
 755         free(handle->name);
 756         free(handle);
 757 }
 758
 759 /* Returns true if the handle has pending statedump requests. */
 760 bool side_statedump_poll_pending_requests(struct side_statedump_request_handle *handle)
 761 {
 762         bool ret;
 763
 764         if (handle->mode != SIDE_STATEDUMP_MODE_POLLING)
 765                 return false;
 766         pthread_mutex_lock(&side_statedump_lock);
 767         ret = !side_list_empty(&handle->notification_queue);
 768         pthread_mutex_unlock(&side_statedump_lock);
 769         return ret;
 770 }
 771
 772 /*
 773  * Only polling mode state dump handles allow application to explicitly handle the
 774  * pending requests.
 775  */
 776 int side_statedump_run_pending_requests(struct side_statedump_request_handle *handle)
 777 {
 778         if (handle->mode != SIDE_STATEDUMP_MODE_POLLING)
 779                 return SIDE_ERROR_INVAL;
 780         _side_statedump_run_pending_requests(handle);
 781         return SIDE_ERROR_OK;
 782 }
 783
 784 /*
 785  * Request a state dump for tracer callbacks identified with "key".
 786  */
 787 int side_tracer_statedump_request(uint64_t key)
 788 {
 789         struct side_statedump_request_handle *handle;
 790
 791         if (key == SIDE_KEY_MATCH_ALL)
 792                 return SIDE_ERROR_INVAL;
 793         pthread_mutex_lock(&side_statedump_lock);
 794         side_list_for_each_entry(handle, &side_statedump_list, node)
 795                 queue_statedump_pending(handle, key);
 796         pthread_mutex_lock(&side_statedump_lock);
 797         return SIDE_ERROR_OK;
 798 }
 799
 800 /*
 801  * Cancel a statedump request.
 802  */
 803 int side_tracer_statedump_request_cancel(uint64_t key)
 804 {
 805         struct side_statedump_request_handle *handle;
 806
 807         if (key == SIDE_KEY_MATCH_ALL)
 808                 return SIDE_ERROR_INVAL;
 809         pthread_mutex_lock(&side_statedump_lock);
 810         side_list_for_each_entry(handle, &side_statedump_list, node)
 811                 unqueue_statedump_pending(handle, key);
 812         pthread_mutex_lock(&side_statedump_lock);
 813         return SIDE_ERROR_OK;
 814 }
 815
 816 /*
 817  * Tracer keys are represented on 64-bit. Return SIDE_ERROR_NOMEM on
 818  * overflow (which should never happen in practice).
 819  */
 820 int side_tracer_request_key(uint64_t *key)
 821 {
 822         int ret = SIDE_ERROR_OK;
 823
 824         pthread_mutex_lock(&side_key_lock);
 825         if (side_key_next == 0) {
 826                 ret = SIDE_ERROR_NOMEM;
 827                 goto end;
 828         }
 829         *key = side_key_next++;
 830 end:
 831         pthread_mutex_unlock(&side_key_lock);
 832         return ret;
 833 }
 834
 835 void side_init(void)
 836 {
 837         if (initialized)
 838                 return;
 839         side_rcu_gp_init(&event_rcu_gp);
 840         side_rcu_gp_init(&statedump_rcu_gp);
 841         initialized = true;
 842 }
 843
 844 /*
 845  * side_exit() is executed from a library destructor. It can be called
 846  * explicitly at application exit as well. Concurrent side API use is
 847  * not expected at that point.
 848  */
 849 void side_exit(void)
 850 {
 851         struct side_events_register_handle *handle, *tmp;
 852
 853         if (finalized)
 854                 return;
 855         side_list_for_each_entry_safe(handle, tmp, &side_events_list, node)
 856                 side_events_unregister(handle);
 857         side_rcu_gp_exit(&event_rcu_gp);
 858         side_rcu_gp_exit(&statedump_rcu_gp);
 859         finalized = true;
 860 }