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         pthread_cond_t worker_cond;
  93         pthread_cond_t waiter_cond;
  94 };
  95
  96 static struct side_rcu_gp_state event_rcu_gp, statedump_rcu_gp;
  97
  98 /*
  99  * Lazy initialization for early use within library constructors.
 100  */
 101 static bool initialized;
 102 /*
 103  * Do not register/unregister any more events after destructor.
 104  */
 105 static bool finalized;
 106
 107 /*
 108  * Recursive mutex to allow tracer callbacks to use the side API.
 109  */
 110 static pthread_mutex_t side_event_lock = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
 111 static pthread_mutex_t side_statedump_lock = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
 112 static pthread_mutex_t side_key_lock = PTHREAD_MUTEX_INITIALIZER;
 113 /*
 114  * The side_agent_thread_lock protects the life-time of the agent
 115  * thread: reference counting, creation, join. It is not taken by
 116  * the agent thread per se so it does not have a circular dependency
 117  * with pthread join.
 118  * The side_statedump_lock nests inside the side_agent_thread_lock.
 119  */
 120 static pthread_mutex_t side_agent_thread_lock = PTHREAD_MUTEX_INITIALIZER;
 121
 122 /* Dynamic tracer key allocation. */
 123 static uint64_t side_key_next = SIDE_KEY_RESERVED_RANGE_END;
 124
 125 static struct statedump_agent_thread statedump_agent_thread;
 126
 127 static DEFINE_SIDE_LIST_HEAD(side_events_list);
 128 static DEFINE_SIDE_LIST_HEAD(side_tracer_list);
 129
 130 /*
 131  * The statedump request list is a RCU list to allow the agent thread to
 132  * iterate over this list with a RCU read-side lock.
 133  */
 134 static DEFINE_SIDE_LIST_HEAD(side_statedump_list);
 135
 136 /*
 137  * Callback filter key for state dump.
 138  */
 139 static __thread uint64_t filter_key = SIDE_KEY_MATCH_ALL;
 140
 141 /*
 142  * The empty callback has a NULL function callback pointer, which stops
 143  * iteration on the array of callbacks immediately.
 144  */
 145 const char side_empty_callback[sizeof(struct side_callback)];
 146
 147 side_static_event(side_statedump_begin, "side", "statedump_begin",
 148         SIDE_LOGLEVEL_INFO, side_field_list(side_field_string("name")));
 149 side_static_event(side_statedump_end, "side", "statedump_end",
 150         SIDE_LOGLEVEL_INFO, side_field_list(side_field_string("name")));
 151
 152 /*
 153  * side_ptrace_hook is a place holder for a debugger breakpoint.
 154  * var_struct is NULL if not variadic.
 155  */
 156 void side_ptrace_hook(const struct side_event_state *event_state __attribute__((unused)),
 157         const struct side_arg_vec *side_arg_vec __attribute__((unused)),
 158         const struct side_arg_dynamic_struct *var_struct __attribute__((unused)))
 159         __attribute__((noinline));
 160 void side_ptrace_hook(const struct side_event_state *event_state __attribute__((unused)),
 161         const struct side_arg_vec *side_arg_vec __attribute__((unused)),
 162         const struct side_arg_dynamic_struct *var_struct __attribute__((unused)))
 163 {
 164 }
 165
 166 static
 167 void _side_call(const struct side_event_state *event_state, const struct side_arg_vec *side_arg_vec, uint64_t key)
 168 {
 169         struct side_rcu_read_state rcu_read_state;
 170         const struct side_event_state_0 *es0;
 171         const struct side_callback *side_cb;
 172         uintptr_t enabled;
 173
 174         if (side_unlikely(finalized))
 175                 return;
 176         if (side_unlikely(!initialized))
 177                 side_init();
 178         if (side_unlikely(event_state->version != 0))
 179                 abort();
 180         es0 = side_container_of(event_state, const struct side_event_state_0, parent);
 181         assert(!(es0->desc->flags & SIDE_EVENT_FLAG_VARIADIC));
 182         enabled = __atomic_load_n(&es0->enabled, __ATOMIC_RELAXED);
 183         if (side_unlikely(enabled & SIDE_EVENT_ENABLED_SHARED_MASK)) {
 184                 if ((enabled & SIDE_EVENT_ENABLED_SHARED_USER_EVENT_MASK) &&
 185                     (key == SIDE_KEY_MATCH_ALL || key == SIDE_KEY_USER_EVENT)) {
 186                         // TODO: call kernel write.
 187                 }
 188                 if ((enabled & SIDE_EVENT_ENABLED_SHARED_PTRACE_MASK) &&
 189                     (key == SIDE_KEY_MATCH_ALL || key == SIDE_KEY_PTRACE))
 190                         side_ptrace_hook(event_state, side_arg_vec, NULL);
 191         }
 192         side_rcu_read_begin(&event_rcu_gp, &rcu_read_state);
 193         for (side_cb = side_rcu_dereference(es0->callbacks); side_cb->u.call != NULL; side_cb++) {
 194                 if (key != SIDE_KEY_MATCH_ALL && side_cb->key != SIDE_KEY_MATCH_ALL && side_cb->key != key)
 195                         continue;
 196                 side_cb->u.call(es0->desc, side_arg_vec, side_cb->priv);
 197         }
 198         side_rcu_read_end(&event_rcu_gp, &rcu_read_state);
 199 }
 200
 201 void side_call(const struct side_event_state *event_state, const struct side_arg_vec *side_arg_vec)
 202 {
 203         _side_call(event_state, side_arg_vec, SIDE_KEY_MATCH_ALL);
 204 }
 205
 206 void side_statedump_call(const struct side_event_state *event_state, const struct side_arg_vec *side_arg_vec)
 207 {
 208         _side_call(event_state, side_arg_vec, filter_key);
 209 }
 210
 211 static
 212 void _side_call_variadic(const struct side_event_state *event_state,
 213         const struct side_arg_vec *side_arg_vec,
 214         const struct side_arg_dynamic_struct *var_struct,
 215         uint64_t key)
 216 {
 217         struct side_rcu_read_state rcu_read_state;
 218         const struct side_event_state_0 *es0;
 219         const struct side_callback *side_cb;
 220         uintptr_t enabled;
 221
 222         if (side_unlikely(finalized))
 223                 return;
 224         if (side_unlikely(!initialized))
 225                 side_init();
 226         if (side_unlikely(event_state->version != 0))
 227                 abort();
 228         es0 = side_container_of(event_state, const struct side_event_state_0, parent);
 229         assert(es0->desc->flags & SIDE_EVENT_FLAG_VARIADIC);
 230         enabled = __atomic_load_n(&es0->enabled, __ATOMIC_RELAXED);
 231         if (side_unlikely(enabled & SIDE_EVENT_ENABLED_SHARED_MASK)) {
 232                 if ((enabled & SIDE_EVENT_ENABLED_SHARED_USER_EVENT_MASK) &&
 233                     (key == SIDE_KEY_MATCH_ALL || key == SIDE_KEY_USER_EVENT)) {
 234                         // TODO: call kernel write.
 235                 }
 236                 if ((enabled & SIDE_EVENT_ENABLED_SHARED_PTRACE_MASK) &&
 237                     (key == SIDE_KEY_MATCH_ALL || key == SIDE_KEY_PTRACE))
 238                         side_ptrace_hook(event_state, side_arg_vec, var_struct);
 239         }
 240         side_rcu_read_begin(&event_rcu_gp, &rcu_read_state);
 241         for (side_cb = side_rcu_dereference(es0->callbacks); side_cb->u.call_variadic != NULL; side_cb++) {
 242                 if (key != SIDE_KEY_MATCH_ALL && side_cb->key != SIDE_KEY_MATCH_ALL && side_cb->key != key)
 243                         continue;
 244                 side_cb->u.call_variadic(es0->desc, side_arg_vec, var_struct, side_cb->priv);
 245         }
 246         side_rcu_read_end(&event_rcu_gp, &rcu_read_state);
 247 }
 248
 249 void side_call_variadic(const struct side_event_state *event_state,
 250         const struct side_arg_vec *side_arg_vec,
 251         const struct side_arg_dynamic_struct *var_struct)
 252 {
 253         _side_call_variadic(event_state, side_arg_vec, var_struct, SIDE_KEY_MATCH_ALL);
 254 }
 255
 256 void side_statedump_call_variadic(const struct side_event_state *event_state,
 257         const struct side_arg_vec *side_arg_vec,
 258         const struct side_arg_dynamic_struct *var_struct)
 259 {
 260         _side_call_variadic(event_state, side_arg_vec, var_struct, filter_key);
 261 }
 262
 263 static
 264 const struct side_callback *side_tracer_callback_lookup(
 265                 const struct side_event_description *desc,
 266                 void *call, void *priv, uint64_t key)
 267 {
 268         struct side_event_state *event_state = side_ptr_get(desc->state);
 269         const struct side_event_state_0 *es0;
 270         const struct side_callback *cb;
 271
 272         if (side_unlikely(event_state->version != 0))
 273                 abort();
 274         es0 = side_container_of(event_state, const struct side_event_state_0, parent);
 275         for (cb = es0->callbacks; cb->u.call != NULL; cb++) {
 276                 if ((void *) cb->u.call == call && cb->priv == priv && cb->key == key)
 277                         return cb;
 278         }
 279         return NULL;
 280 }
 281
 282 static
 283 int _side_tracer_callback_register(struct side_event_description *desc,
 284                 void *call, void *priv, uint64_t key)
 285 {
 286         struct side_event_state *event_state;
 287         struct side_callback *old_cb, *new_cb;
 288         struct side_event_state_0 *es0;
 289         int ret = SIDE_ERROR_OK;
 290         uint32_t old_nr_cb;
 291
 292         if (!call)
 293                 return SIDE_ERROR_INVAL;
 294         if (finalized)
 295                 return SIDE_ERROR_EXITING;
 296         if (!initialized)
 297                 side_init();
 298         pthread_mutex_lock(&side_event_lock);
 299         event_state = side_ptr_get(desc->state);
 300         if (side_unlikely(event_state->version != 0))
 301                 abort();
 302         es0 = side_container_of(event_state, struct side_event_state_0, parent);
 303         old_nr_cb = es0->nr_callbacks;
 304         if (old_nr_cb == UINT32_MAX) {
 305                 ret = SIDE_ERROR_INVAL;
 306                 goto unlock;
 307         }
 308         /* Reject duplicate (call, priv) tuples. */
 309         if (side_tracer_callback_lookup(desc, call, priv, key)) {
 310                 ret = SIDE_ERROR_EXIST;
 311                 goto unlock;
 312         }
 313         old_cb = (struct side_callback *) es0->callbacks;
 314         /* old_nr_cb + 1 (new cb) + 1 (NULL) */
 315         new_cb = (struct side_callback *) calloc(old_nr_cb + 2, sizeof(struct side_callback));
 316         if (!new_cb) {
 317                 ret = SIDE_ERROR_NOMEM;
 318                 goto unlock;
 319         }
 320         memcpy(new_cb, old_cb, old_nr_cb);
 321         if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
 322                 new_cb[old_nr_cb].u.call_variadic =
 323                         (side_tracer_callback_variadic_func) call;
 324         else
 325                 new_cb[old_nr_cb].u.call =
 326                         (side_tracer_callback_func) call;
 327         new_cb[old_nr_cb].priv = priv;
 328         new_cb[old_nr_cb].key = key;
 329         /* High order bits are already zeroed. */
 330         side_rcu_assign_pointer(es0->callbacks, new_cb);
 331         side_rcu_wait_grace_period(&event_rcu_gp);
 332         if (old_nr_cb)
 333                 free(old_cb);
 334         es0->nr_callbacks++;
 335         /* Increment concurrently with kernel setting the top bits. */
 336         if (!old_nr_cb)
 337                 (void) __atomic_add_fetch(&es0->enabled, 1, __ATOMIC_RELAXED);
 338 unlock:
 339         pthread_mutex_unlock(&side_event_lock);
 340         return ret;
 341 }
 342
 343 int side_tracer_callback_register(struct side_event_description *desc,
 344                 side_tracer_callback_func call,
 345                 void *priv, uint64_t key)
 346 {
 347         if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
 348                 return SIDE_ERROR_INVAL;
 349         return _side_tracer_callback_register(desc, (void *) call, priv, key);
 350 }
 351
 352 int side_tracer_callback_variadic_register(struct side_event_description *desc,
 353                 side_tracer_callback_variadic_func call_variadic,
 354                 void *priv, uint64_t key)
 355 {
 356         if (!(desc->flags & SIDE_EVENT_FLAG_VARIADIC))
 357                 return SIDE_ERROR_INVAL;
 358         return _side_tracer_callback_register(desc, (void *) call_variadic, priv, key);
 359 }
 360
 361 static int _side_tracer_callback_unregister(struct side_event_description *desc,
 362                 void *call, void *priv, uint64_t key)
 363 {
 364         struct side_event_state *event_state;
 365         struct side_callback *old_cb, *new_cb;
 366         const struct side_callback *cb_pos;
 367         struct side_event_state_0 *es0;
 368         uint32_t pos_idx;
 369         int ret = SIDE_ERROR_OK;
 370         uint32_t old_nr_cb;
 371
 372         if (!call)
 373                 return SIDE_ERROR_INVAL;
 374         if (finalized)
 375                 return SIDE_ERROR_EXITING;
 376         if (!initialized)
 377                 side_init();
 378         pthread_mutex_lock(&side_event_lock);
 379         event_state = side_ptr_get(desc->state);
 380         if (side_unlikely(event_state->version != 0))
 381                 abort();
 382         es0 = side_container_of(event_state, struct side_event_state_0, parent);
 383         cb_pos = side_tracer_callback_lookup(desc, call, priv, key);
 384         if (!cb_pos) {
 385                 ret = SIDE_ERROR_NOENT;
 386                 goto unlock;
 387         }
 388         old_nr_cb = es0->nr_callbacks;
 389         old_cb = (struct side_callback *) es0->callbacks;
 390         if (old_nr_cb == 1) {
 391                 new_cb = (struct side_callback *) &side_empty_callback;
 392         } else {
 393                 pos_idx = cb_pos - es0->callbacks;
 394                 /* Remove entry at pos_idx. */
 395                 /* old_nr_cb - 1 (removed cb) + 1 (NULL) */
 396                 new_cb = (struct side_callback *) calloc(old_nr_cb, sizeof(struct side_callback));
 397                 if (!new_cb) {
 398                         ret = SIDE_ERROR_NOMEM;
 399                         goto unlock;
 400                 }
 401                 memcpy(new_cb, old_cb, pos_idx);
 402                 memcpy(&new_cb[pos_idx], &old_cb[pos_idx + 1], old_nr_cb - pos_idx - 1);
 403         }
 404         /* High order bits are already zeroed. */
 405         side_rcu_assign_pointer(es0->callbacks, new_cb);
 406         side_rcu_wait_grace_period(&event_rcu_gp);
 407         free(old_cb);
 408         es0->nr_callbacks--;
 409         /* Decrement concurrently with kernel setting the top bits. */
 410         if (old_nr_cb == 1)
 411                 (void) __atomic_add_fetch(&es0->enabled, -1, __ATOMIC_RELAXED);
 412 unlock:
 413         pthread_mutex_unlock(&side_event_lock);
 414         return ret;
 415 }
 416
 417 int side_tracer_callback_unregister(struct side_event_description *desc,
 418                 side_tracer_callback_func call,
 419                 void *priv, uint64_t key)
 420 {
 421         if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
 422                 return SIDE_ERROR_INVAL;
 423         return _side_tracer_callback_unregister(desc, (void *) call, priv, key);
 424 }
 425
 426 int side_tracer_callback_variadic_unregister(struct side_event_description *desc,
 427                 side_tracer_callback_variadic_func call_variadic,
 428                 void *priv, uint64_t key)
 429 {
 430         if (!(desc->flags & SIDE_EVENT_FLAG_VARIADIC))
 431                 return SIDE_ERROR_INVAL;
 432         return _side_tracer_callback_unregister(desc, (void *) call_variadic, priv, key);
 433 }
 434
 435 struct side_events_register_handle *side_events_register(struct side_event_description **events, uint32_t nr_events)
 436 {
 437         struct side_events_register_handle *events_handle = NULL;
 438         struct side_tracer_handle *tracer_handle;
 439
 440         if (finalized)
 441                 return NULL;
 442         if (!initialized)
 443                 side_init();
 444         events_handle = (struct side_events_register_handle *)
 445                         calloc(1, sizeof(struct side_events_register_handle));
 446         if (!events_handle)
 447                 return NULL;
 448         events_handle->events = events;
 449         events_handle->nr_events = nr_events;
 450
 451         pthread_mutex_lock(&side_event_lock);
 452         side_list_insert_node_tail(&side_events_list, &events_handle->node);
 453         side_list_for_each_entry(tracer_handle, &side_tracer_list, node) {
 454                 tracer_handle->cb(SIDE_TRACER_NOTIFICATION_INSERT_EVENTS,
 455                         events, nr_events, tracer_handle->priv);
 456         }
 457         pthread_mutex_unlock(&side_event_lock);
 458         //TODO: call event batch register ioctl
 459         return events_handle;
 460 }
 461
 462 static
 463 void side_event_remove_callbacks(struct side_event_description *desc)
 464 {
 465         struct side_event_state *event_state = side_ptr_get(desc->state);
 466         struct side_event_state_0 *es0;
 467         struct side_callback *old_cb;
 468         uint32_t nr_cb;
 469
 470         if (side_unlikely(event_state->version != 0))
 471                 abort();
 472         es0 = side_container_of(event_state, struct side_event_state_0, parent);
 473         nr_cb = es0->nr_callbacks;
 474         if (!nr_cb)
 475                 return;
 476         old_cb = (struct side_callback *) es0->callbacks;
 477         (void) __atomic_add_fetch(&es0->enabled, -1, __ATOMIC_RELAXED);
 478         /*
 479          * Setting the state back to 0 cb and empty callbacks out of
 480          * caution. This should not matter because instrumentation is
 481          * unreachable.
 482          */
 483         es0->nr_callbacks = 0;
 484         side_rcu_assign_pointer(es0->callbacks, &side_empty_callback);
 485         /*
 486          * No need to wait for grace period because instrumentation is
 487          * unreachable.
 488          */
 489         free(old_cb);
 490 }
 491
 492 /*
 493  * Unregister event handle. At this point, all side events in that
 494  * handle should be unreachable.
 495  */
 496 void side_events_unregister(struct side_events_register_handle *events_handle)
 497 {
 498         struct side_tracer_handle *tracer_handle;
 499         uint32_t i;
 500
 501         if (!events_handle)
 502                 return;
 503         if (finalized)
 504                 return;
 505         if (!initialized)
 506                 side_init();
 507         pthread_mutex_lock(&side_event_lock);
 508         side_list_remove_node(&events_handle->node);
 509         side_list_for_each_entry(tracer_handle, &side_tracer_list, node) {
 510                 tracer_handle->cb(SIDE_TRACER_NOTIFICATION_REMOVE_EVENTS,
 511                         events_handle->events, events_handle->nr_events,
 512                         tracer_handle->priv);
 513         }
 514         for (i = 0; i < events_handle->nr_events; i++) {
 515                 struct side_event_description *event = events_handle->events[i];
 516
 517                 /* Skip NULL pointers */
 518                 if (!event)
 519                         continue;
 520                 side_event_remove_callbacks(event);
 521         }
 522         pthread_mutex_unlock(&side_event_lock);
 523         //TODO: call event batch unregister ioctl
 524         free(events_handle);
 525 }
 526
 527 struct side_tracer_handle *side_tracer_event_notification_register(
 528                 void (*cb)(enum side_tracer_notification notif,
 529                         struct side_event_description **events, uint32_t nr_events, void *priv),
 530                 void *priv)
 531 {
 532         struct side_tracer_handle *tracer_handle;
 533         struct side_events_register_handle *events_handle;
 534
 535         if (finalized)
 536                 return NULL;
 537         if (!initialized)
 538                 side_init();
 539         tracer_handle = (struct side_tracer_handle *)
 540                                 calloc(1, sizeof(struct side_tracer_handle));
 541         if (!tracer_handle)
 542                 return NULL;
 543         pthread_mutex_lock(&side_event_lock);
 544         tracer_handle->cb = cb;
 545         tracer_handle->priv = priv;
 546         side_list_insert_node_tail(&side_tracer_list, &tracer_handle->node);
 547         side_list_for_each_entry(events_handle, &side_events_list, node) {
 548                 cb(SIDE_TRACER_NOTIFICATION_INSERT_EVENTS,
 549                         events_handle->events, events_handle->nr_events, priv);
 550         }
 551         pthread_mutex_unlock(&side_event_lock);
 552         return tracer_handle;
 553 }
 554
 555 void side_tracer_event_notification_unregister(struct side_tracer_handle *tracer_handle)
 556 {
 557         struct side_events_register_handle *events_handle;
 558
 559         if (finalized)
 560                 return;
 561         if (!initialized)
 562                 side_init();
 563         pthread_mutex_lock(&side_event_lock);
 564         side_list_for_each_entry(events_handle, &side_events_list, node) {
 565                 tracer_handle->cb(SIDE_TRACER_NOTIFICATION_REMOVE_EVENTS,
 566                         events_handle->events, events_handle->nr_events,
 567                         tracer_handle->priv);
 568         }
 569         side_list_remove_node(&tracer_handle->node);
 570         pthread_mutex_unlock(&side_event_lock);
 571         free(tracer_handle);
 572 }
 573
 574 /* Called with side_statedump_lock held. */
 575 static
 576 void queue_statedump_pending(struct side_statedump_request_handle *handle, uint64_t key)
 577 {
 578         struct side_statedump_notification *notif;
 579
 580         notif = (struct side_statedump_notification *) calloc(1, sizeof(struct side_statedump_notification));
 581         if (!notif)
 582                 abort();
 583         notif->key = key;
 584         side_list_insert_node_tail(&handle->notification_queue, &notif->node);
 585         if (handle->mode == SIDE_STATEDUMP_MODE_AGENT_THREAD) {
 586                 (void)__atomic_or_fetch(&statedump_agent_thread.state, AGENT_THREAD_STATE_HANDLE_REQUEST, __ATOMIC_SEQ_CST);
 587                 pthread_cond_broadcast(&statedump_agent_thread.worker_cond);
 588         }
 589 }
 590
 591 /* Called with side_statedump_lock held. */
 592 static
 593 void unqueue_statedump_pending(struct side_statedump_request_handle *handle, uint64_t key)
 594 {
 595         struct side_statedump_notification *notif, *tmp;
 596
 597         side_list_for_each_entry_safe(notif, tmp, &handle->notification_queue, node) {
 598                 if (key == SIDE_KEY_MATCH_ALL || key == notif->key) {
 599                         side_list_remove_node(&notif->node);
 600                         free(notif);
 601                 }
 602         }
 603 }
 604
 605 static
 606 void side_statedump_run(struct side_statedump_request_handle *handle,
 607                 struct side_statedump_notification *notif)
 608 {
 609         /* Invoke the state dump callback specifically for the tracer key. */
 610         filter_key = notif->key;
 611         side_statedump_event_call(side_statedump_begin,
 612                 side_arg_list(side_arg_string(handle->name)));
 613         handle->cb();
 614         side_statedump_event_call(side_statedump_end,
 615                 side_arg_list(side_arg_string(handle->name)));
 616         filter_key = SIDE_KEY_MATCH_ALL;
 617 }
 618
 619 static
 620 void _side_statedump_run_pending_requests(struct side_statedump_request_handle *handle)
 621 {
 622         struct side_statedump_notification *notif, *tmp;
 623         DEFINE_SIDE_LIST_HEAD(tmp_head);
 624
 625         pthread_mutex_lock(&side_statedump_lock);
 626         side_list_splice(&handle->notification_queue, &tmp_head);
 627         side_list_head_init(&handle->notification_queue);
 628         pthread_mutex_unlock(&side_statedump_lock);
 629
 630         /* We are now sole owner of the tmp_head list. */
 631         side_list_for_each_entry(notif, &tmp_head, node)
 632                 side_statedump_run(handle, notif);
 633         side_list_for_each_entry_safe(notif, tmp, &tmp_head, node)
 634                 free(notif);
 635
 636         if (handle->mode == SIDE_STATEDUMP_MODE_AGENT_THREAD) {
 637                 pthread_mutex_lock(&side_statedump_lock);
 638                 pthread_cond_broadcast(&statedump_agent_thread.waiter_cond);
 639                 pthread_mutex_unlock(&side_statedump_lock);
 640         }
 641 }
 642
 643 static
 644 void *statedump_agent_func(void *arg __attribute__((unused)))
 645 {
 646         for (;;) {
 647                 struct side_statedump_request_handle *handle;
 648                 struct side_rcu_read_state rcu_read_state;
 649                 enum agent_thread_state state;
 650
 651                 pthread_mutex_lock(&side_statedump_lock);
 652                 for (;;) {
 653                         state = __atomic_load_n(&statedump_agent_thread.state, __ATOMIC_SEQ_CST);
 654                         if (state == AGENT_THREAD_STATE_BLOCKED)
 655                                 pthread_cond_wait(&statedump_agent_thread.worker_cond, &side_statedump_lock);
 656                         else
 657                                 break;
 658                 }
 659                 pthread_mutex_unlock(&side_statedump_lock);
 660                 if (state & AGENT_THREAD_STATE_EXIT)
 661                         break;
 662                 (void)__atomic_and_fetch(&statedump_agent_thread.state, ~AGENT_THREAD_STATE_HANDLE_REQUEST, __ATOMIC_SEQ_CST);
 663                 side_rcu_read_begin(&statedump_rcu_gp, &rcu_read_state);
 664                 side_list_for_each_entry_rcu(handle, &side_statedump_list, node)
 665                         _side_statedump_run_pending_requests(handle);
 666                 side_rcu_read_end(&statedump_rcu_gp, &rcu_read_state);
 667         }
 668         return NULL;
 669 }
 670
 671 /* Called with side_agent_thread_lock and side_statedump_lock held. */
 672 static
 673 void statedump_agent_thread_get(void)
 674 {
 675         int ret;
 676
 677         if (statedump_agent_thread.ref++)
 678                 return;
 679         pthread_cond_init(&statedump_agent_thread.worker_cond, NULL);
 680         pthread_cond_init(&statedump_agent_thread.waiter_cond, NULL);
 681         statedump_agent_thread.state = AGENT_THREAD_STATE_BLOCKED;
 682         ret = pthread_create(&statedump_agent_thread.id, NULL,
 683                         statedump_agent_func, NULL);
 684         if (ret) {
 685                 abort();
 686         }
 687 }
 688
 689 /*
 690  * Called with side_agent_thread_lock and side_statedump_lock held.
 691  * Returns true if join for agent thread is needed.
 692  */
 693 static
 694 bool statedump_agent_thread_put(void)
 695 {
 696         if (--statedump_agent_thread.ref)
 697                 return false;
 698         (void)__atomic_or_fetch(&statedump_agent_thread.state, AGENT_THREAD_STATE_EXIT, __ATOMIC_SEQ_CST);
 699         pthread_cond_broadcast(&statedump_agent_thread.worker_cond);
 700         return true;
 701 }
 702
 703 /* Called with side_agent_thread_lock held. */
 704 static
 705 void statedump_agent_thread_join(void)
 706 {
 707         int ret;
 708         void *retval;
 709
 710         ret = pthread_join(statedump_agent_thread.id, &retval);
 711         if (ret) {
 712                 abort();
 713         }
 714         statedump_agent_thread.state = AGENT_THREAD_STATE_BLOCKED;
 715         pthread_cond_destroy(&statedump_agent_thread.worker_cond);
 716         pthread_cond_destroy(&statedump_agent_thread.waiter_cond);
 717 }
 718
 719 struct side_statedump_request_handle *
 720         side_statedump_request_notification_register(const char *state_name,
 721                 void (*statedump_cb)(void),
 722                 enum side_statedump_mode mode)
 723 {
 724         struct side_statedump_request_handle *handle;
 725         char *name;
 726
 727         if (finalized)
 728                 return NULL;
 729         if (!initialized)
 730                 side_init();
 731         /*
 732          * The statedump request notification should not be registered
 733          * from a notification callback.
 734          */
 735         assert(!filter_key);
 736         handle = (struct side_statedump_request_handle *)
 737                                 calloc(1, sizeof(struct side_statedump_request_handle));
 738         if (!handle)
 739                 return NULL;
 740         name = strdup(state_name);
 741         if (!name)
 742                 goto name_nomem;
 743         handle->cb = statedump_cb;
 744         handle->name = name;
 745         handle->mode = mode;
 746         side_list_head_init(&handle->notification_queue);
 747
 748         if (mode == SIDE_STATEDUMP_MODE_AGENT_THREAD)
 749                 pthread_mutex_lock(&side_agent_thread_lock);
 750         pthread_mutex_lock(&side_statedump_lock);
 751         if (mode == SIDE_STATEDUMP_MODE_AGENT_THREAD)
 752                 statedump_agent_thread_get();
 753         side_list_insert_node_tail_rcu(&side_statedump_list, &handle->node);
 754         /* Queue statedump pending for all tracers. */
 755         queue_statedump_pending(handle, SIDE_KEY_MATCH_ALL);
 756         pthread_mutex_unlock(&side_statedump_lock);
 757
 758         if (mode == SIDE_STATEDUMP_MODE_AGENT_THREAD) {
 759                 pthread_mutex_unlock(&side_agent_thread_lock);
 760
 761                 pthread_mutex_lock(&side_statedump_lock);
 762                 while (!side_list_empty(&handle->notification_queue))
 763                         pthread_cond_wait(&statedump_agent_thread.waiter_cond, &side_statedump_lock);
 764                 pthread_mutex_unlock(&side_statedump_lock);
 765         }
 766
 767         return handle;
 768
 769 name_nomem:
 770         free(handle);
 771         return NULL;
 772 }
 773
 774 void side_statedump_request_notification_unregister(struct side_statedump_request_handle *handle)
 775 {
 776         bool join = false;
 777
 778         if (finalized)
 779                 return;
 780         if (!initialized)
 781                 side_init();
 782         assert(!filter_key);
 783
 784         if (handle->mode == SIDE_STATEDUMP_MODE_AGENT_THREAD)
 785                 pthread_mutex_lock(&side_agent_thread_lock);
 786         pthread_mutex_lock(&side_statedump_lock);
 787         unqueue_statedump_pending(handle, SIDE_KEY_MATCH_ALL);
 788         side_list_remove_node_rcu(&handle->node);
 789         if (handle->mode == SIDE_STATEDUMP_MODE_AGENT_THREAD)
 790                 join = statedump_agent_thread_put();
 791         pthread_mutex_unlock(&side_statedump_lock);
 792         if (join)
 793                 statedump_agent_thread_join();
 794         if (handle->mode == SIDE_STATEDUMP_MODE_AGENT_THREAD)
 795                 pthread_mutex_unlock(&side_agent_thread_lock);
 796
 797         side_rcu_wait_grace_period(&statedump_rcu_gp);
 798         free(handle->name);
 799         free(handle);
 800 }
 801
 802 /* Returns true if the handle has pending statedump requests. */
 803 bool side_statedump_poll_pending_requests(struct side_statedump_request_handle *handle)
 804 {
 805         bool ret;
 806
 807         if (handle->mode != SIDE_STATEDUMP_MODE_POLLING)
 808                 return false;
 809         pthread_mutex_lock(&side_statedump_lock);
 810         ret = !side_list_empty(&handle->notification_queue);
 811         pthread_mutex_unlock(&side_statedump_lock);
 812         return ret;
 813 }
 814
 815 /*
 816  * Only polling mode state dump handles allow application to explicitly handle the
 817  * pending requests.
 818  */
 819 int side_statedump_run_pending_requests(struct side_statedump_request_handle *handle)
 820 {
 821         if (handle->mode != SIDE_STATEDUMP_MODE_POLLING)
 822                 return SIDE_ERROR_INVAL;
 823         _side_statedump_run_pending_requests(handle);
 824         return SIDE_ERROR_OK;
 825 }
 826
 827 /*
 828  * Request a state dump for tracer callbacks identified with "key".
 829  */
 830 int side_tracer_statedump_request(uint64_t key)
 831 {
 832         struct side_statedump_request_handle *handle;
 833
 834         if (key == SIDE_KEY_MATCH_ALL)
 835                 return SIDE_ERROR_INVAL;
 836         pthread_mutex_lock(&side_statedump_lock);
 837         side_list_for_each_entry(handle, &side_statedump_list, node)
 838                 queue_statedump_pending(handle, key);
 839         pthread_mutex_lock(&side_statedump_lock);
 840         return SIDE_ERROR_OK;
 841 }
 842
 843 /*
 844  * Cancel a statedump request.
 845  */
 846 int side_tracer_statedump_request_cancel(uint64_t key)
 847 {
 848         struct side_statedump_request_handle *handle;
 849
 850         if (key == SIDE_KEY_MATCH_ALL)
 851                 return SIDE_ERROR_INVAL;
 852         pthread_mutex_lock(&side_statedump_lock);
 853         side_list_for_each_entry(handle, &side_statedump_list, node)
 854                 unqueue_statedump_pending(handle, key);
 855         pthread_mutex_lock(&side_statedump_lock);
 856         return SIDE_ERROR_OK;
 857 }
 858
 859 /*
 860  * Tracer keys are represented on 64-bit. Return SIDE_ERROR_NOMEM on
 861  * overflow (which should never happen in practice).
 862  */
 863 int side_tracer_request_key(uint64_t *key)
 864 {
 865         int ret = SIDE_ERROR_OK;
 866
 867         pthread_mutex_lock(&side_key_lock);
 868         if (side_key_next == 0) {
 869                 ret = SIDE_ERROR_NOMEM;
 870                 goto end;
 871         }
 872         *key = side_key_next++;
 873 end:
 874         pthread_mutex_unlock(&side_key_lock);
 875         return ret;
 876 }
 877
 878 void side_init(void)
 879 {
 880         if (initialized)
 881                 return;
 882         side_rcu_gp_init(&event_rcu_gp);
 883         side_rcu_gp_init(&statedump_rcu_gp);
 884         initialized = true;
 885 }
 886
 887 /*
 888  * side_exit() is executed from a library destructor. It can be called
 889  * explicitly at application exit as well. Concurrent side API use is
 890  * not expected at that point.
 891  */
 892 void side_exit(void)
 893 {
 894         struct side_events_register_handle *handle, *tmp;
 895
 896         if (finalized)
 897                 return;
 898         side_list_for_each_entry_safe(handle, tmp, &side_events_list, node)
 899                 side_events_unregister(handle);
 900         side_rcu_gp_exit(&event_rcu_gp);
 901         side_rcu_gp_exit(&statedump_rcu_gp);
 902         finalized = true;
 903 }