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