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
  13 /* Top 8 bits reserved for kernel tracer use. */
  14 #if SIDE_BITS_PER_LONG == 64
  15 # define SIDE_EVENT_ENABLED_KERNEL_MASK                 0xFF00000000000000ULL
  16 # define SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK      0x8000000000000000ULL
  17
  18 /* Allow 2^56 tracer references on an event. */
  19 # define SIDE_EVENT_ENABLED_USER_MASK                   0x00FFFFFFFFFFFFFFULL
  20 #else
  21 # define SIDE_EVENT_ENABLED_KERNEL_MASK                 0xFF000000UL
  22 # define SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK      0x80000000UL
  23
  24 /* Allow 2^24 tracer references on an event. */
  25 # define SIDE_EVENT_ENABLED_USER_MASK                   0x00FFFFFFUL
  26 #endif
  27
  28 struct side_events_register_handle {
  29         struct side_list_node node;
  30         struct side_event_description **events;
  31         uint32_t nr_events;
  32 };
  33
  34 struct side_tracer_handle {
  35         struct side_list_node node;
  36         void (*cb)(enum side_tracer_notification notif,
  37                 struct side_event_description **events, uint32_t nr_events, void *priv);
  38         void *priv;
  39 };
  40
  41 struct side_callback {
  42         union {
  43                 void (*call)(const struct side_event_description *desc,
  44                         const struct side_arg_vec *side_arg_vec,
  45                         void *priv);
  46                 void (*call_variadic)(const struct side_event_description *desc,
  47                         const struct side_arg_vec *side_arg_vec,
  48                         const struct side_arg_dynamic_struct *var_struct,
  49                         void *priv);
  50         } u;
  51         void *priv;
  52 };
  53
  54 static struct side_rcu_gp_state rcu_gp;
  55
  56 /*
  57  * Lazy initialization for early use within library constructors.
  58  */
  59 static bool initialized;
  60 /*
  61  * Do not register/unregister any more events after destructor.
  62  */
  63 static bool finalized;
  64
  65 /*
  66  * Recursive mutex to allow tracer callbacks to use the side API.
  67  */
  68 static pthread_mutex_t side_lock = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
  69
  70 static DEFINE_SIDE_LIST_HEAD(side_events_list);
  71 static DEFINE_SIDE_LIST_HEAD(side_tracer_list);
  72
  73 /*
  74  * The empty callback has a NULL function callback pointer, which stops
  75  * iteration on the array of callbacks immediately.
  76  */
  77 const char side_empty_callback[sizeof(struct side_callback)];
  78
  79 void side_call(const struct side_event_state *event_state, const struct side_arg_vec *side_arg_vec)
  80 {
  81         struct side_rcu_read_state rcu_read_state;
  82         const struct side_event_state_0 *es0;
  83         const struct side_callback *side_cb;
  84         uintptr_t enabled;
  85
  86         if (side_unlikely(finalized))
  87                 return;
  88         if (side_unlikely(!initialized))
  89                 side_init();
  90         if (side_unlikely(event_state->version != 0))
  91                 abort();
  92         es0 = side_container_of(event_state, const struct side_event_state_0, parent);
  93         assert(!(es0->desc->flags & SIDE_EVENT_FLAG_VARIADIC));
  94         enabled = __atomic_load_n(&es0->enabled, __ATOMIC_RELAXED);
  95         if (side_unlikely(enabled & SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK)) {
  96                 // TODO: call kernel write.
  97         }
  98         side_rcu_read_begin(&rcu_gp, &rcu_read_state);
  99         for (side_cb = side_rcu_dereference(es0->callbacks); side_cb->u.call != NULL; side_cb++)
 100                 side_cb->u.call(es0->desc, side_arg_vec, side_cb->priv);
 101         side_rcu_read_end(&rcu_gp, &rcu_read_state);
 102 }
 103
 104 void side_call_variadic(const struct side_event_state *event_state,
 105         const struct side_arg_vec *side_arg_vec,
 106         const struct side_arg_dynamic_struct *var_struct)
 107 {
 108         struct side_rcu_read_state rcu_read_state;
 109         const struct side_event_state_0 *es0;
 110         const struct side_callback *side_cb;
 111         uintptr_t enabled;
 112
 113         if (side_unlikely(finalized))
 114                 return;
 115         if (side_unlikely(!initialized))
 116                 side_init();
 117         if (side_unlikely(event_state->version != 0))
 118                 abort();
 119         es0 = side_container_of(event_state, const struct side_event_state_0, parent);
 120         assert(es0->desc->flags & SIDE_EVENT_FLAG_VARIADIC);
 121         enabled = __atomic_load_n(&es0->enabled, __ATOMIC_RELAXED);
 122         if (side_unlikely(enabled & SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK)) {
 123                 // TODO: call kernel write.
 124         }
 125         side_rcu_read_begin(&rcu_gp, &rcu_read_state);
 126         for (side_cb = side_rcu_dereference(es0->callbacks); side_cb->u.call_variadic != NULL; side_cb++)
 127                 side_cb->u.call_variadic(es0->desc, side_arg_vec, var_struct, side_cb->priv);
 128         side_rcu_read_end(&rcu_gp, &rcu_read_state);
 129 }
 130
 131 static
 132 const struct side_callback *side_tracer_callback_lookup(
 133                 const struct side_event_description *desc,
 134                 void *call, void *priv)
 135 {
 136         struct side_event_state *event_state = side_ptr_get(desc->state);
 137         const struct side_event_state_0 *es0;
 138         const struct side_callback *cb;
 139
 140         if (side_unlikely(event_state->version != 0))
 141                 abort();
 142         es0 = side_container_of(event_state, const struct side_event_state_0, parent);
 143         for (cb = es0->callbacks; cb->u.call != NULL; cb++) {
 144                 if ((void *) cb->u.call == call && cb->priv == priv)
 145                         return cb;
 146         }
 147         return NULL;
 148 }
 149
 150 static
 151 int _side_tracer_callback_register(struct side_event_description *desc,
 152                 void *call, void *priv)
 153 {
 154         struct side_event_state *event_state;
 155         struct side_callback *old_cb, *new_cb;
 156         struct side_event_state_0 *es0;
 157         int ret = SIDE_ERROR_OK;
 158         uint32_t old_nr_cb;
 159
 160         if (!call)
 161                 return SIDE_ERROR_INVAL;
 162         if (finalized)
 163                 return SIDE_ERROR_EXITING;
 164         if (!initialized)
 165                 side_init();
 166         pthread_mutex_lock(&side_lock);
 167         event_state = side_ptr_get(desc->state);
 168         if (side_unlikely(event_state->version != 0))
 169                 abort();
 170         es0 = side_container_of(event_state, struct side_event_state_0, parent);
 171         old_nr_cb = es0->nr_callbacks;
 172         if (old_nr_cb == UINT32_MAX) {
 173                 ret = SIDE_ERROR_INVAL;
 174                 goto unlock;
 175         }
 176         /* Reject duplicate (call, priv) tuples. */
 177         if (side_tracer_callback_lookup(desc, call, priv)) {
 178                 ret = SIDE_ERROR_EXIST;
 179                 goto unlock;
 180         }
 181         old_cb = (struct side_callback *) es0->callbacks;
 182         /* old_nr_cb + 1 (new cb) + 1 (NULL) */
 183         new_cb = (struct side_callback *) calloc(old_nr_cb + 2, sizeof(struct side_callback));
 184         if (!new_cb) {
 185                 ret = SIDE_ERROR_NOMEM;
 186                 goto unlock;
 187         }
 188         memcpy(new_cb, old_cb, old_nr_cb);
 189         if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
 190                 new_cb[old_nr_cb].u.call_variadic =
 191                         (side_tracer_callback_variadic_func) call;
 192         else
 193                 new_cb[old_nr_cb].u.call =
 194                         (side_tracer_callback_func) call;
 195         new_cb[old_nr_cb].priv = priv;
 196         /* High order bits are already zeroed. */
 197         side_rcu_assign_pointer(es0->callbacks, new_cb);
 198         side_rcu_wait_grace_period(&rcu_gp);
 199         if (old_nr_cb)
 200                 free(old_cb);
 201         es0->nr_callbacks++;
 202         /* Increment concurrently with kernel setting the top bits. */
 203         if (!old_nr_cb)
 204                 (void) __atomic_add_fetch(&es0->enabled, 1, __ATOMIC_RELAXED);
 205 unlock:
 206         pthread_mutex_unlock(&side_lock);
 207         return ret;
 208 }
 209
 210 int side_tracer_callback_register(struct side_event_description *desc,
 211                 side_tracer_callback_func call,
 212                 void *priv)
 213 {
 214         if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
 215                 return SIDE_ERROR_INVAL;
 216         return _side_tracer_callback_register(desc, (void *) call, priv);
 217 }
 218
 219 int side_tracer_callback_variadic_register(struct side_event_description *desc,
 220                 side_tracer_callback_variadic_func call_variadic,
 221                 void *priv)
 222 {
 223         if (!(desc->flags & SIDE_EVENT_FLAG_VARIADIC))
 224                 return SIDE_ERROR_INVAL;
 225         return _side_tracer_callback_register(desc, (void *) call_variadic, priv);
 226 }
 227
 228 static int _side_tracer_callback_unregister(struct side_event_description *desc,
 229                 void *call, void *priv)
 230 {
 231         struct side_event_state *event_state;
 232         struct side_callback *old_cb, *new_cb;
 233         const struct side_callback *cb_pos;
 234         struct side_event_state_0 *es0;
 235         uint32_t pos_idx;
 236         int ret = SIDE_ERROR_OK;
 237         uint32_t old_nr_cb;
 238
 239         if (!call)
 240                 return SIDE_ERROR_INVAL;
 241         if (finalized)
 242                 return SIDE_ERROR_EXITING;
 243         if (!initialized)
 244                 side_init();
 245         pthread_mutex_lock(&side_lock);
 246         event_state = side_ptr_get(desc->state);
 247         if (side_unlikely(event_state->version != 0))
 248                 abort();
 249         es0 = side_container_of(event_state, struct side_event_state_0, parent);
 250         cb_pos = side_tracer_callback_lookup(desc, call, priv);
 251         if (!cb_pos) {
 252                 ret = SIDE_ERROR_NOENT;
 253                 goto unlock;
 254         }
 255         old_nr_cb = es0->nr_callbacks;
 256         old_cb = (struct side_callback *) es0->callbacks;
 257         if (old_nr_cb == 1) {
 258                 new_cb = (struct side_callback *) &side_empty_callback;
 259         } else {
 260                 pos_idx = cb_pos - es0->callbacks;
 261                 /* Remove entry at pos_idx. */
 262                 /* old_nr_cb - 1 (removed cb) + 1 (NULL) */
 263                 new_cb = (struct side_callback *) calloc(old_nr_cb, sizeof(struct side_callback));
 264                 if (!new_cb) {
 265                         ret = SIDE_ERROR_NOMEM;
 266                         goto unlock;
 267                 }
 268                 memcpy(new_cb, old_cb, pos_idx);
 269                 memcpy(&new_cb[pos_idx], &old_cb[pos_idx + 1], old_nr_cb - pos_idx - 1);
 270         }
 271         /* High order bits are already zeroed. */
 272         side_rcu_assign_pointer(es0->callbacks, new_cb);
 273         side_rcu_wait_grace_period(&rcu_gp);
 274         free(old_cb);
 275         es0->nr_callbacks--;
 276         /* Decrement concurrently with kernel setting the top bits. */
 277         if (old_nr_cb == 1)
 278                 (void) __atomic_add_fetch(&es0->enabled, -1, __ATOMIC_RELAXED);
 279 unlock:
 280         pthread_mutex_unlock(&side_lock);
 281         return ret;
 282 }
 283
 284 int side_tracer_callback_unregister(struct side_event_description *desc,
 285                 side_tracer_callback_func call,
 286                 void *priv)
 287 {
 288         if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
 289                 return SIDE_ERROR_INVAL;
 290         return _side_tracer_callback_unregister(desc, (void *) call, priv);
 291 }
 292
 293 int side_tracer_callback_variadic_unregister(struct side_event_description *desc,
 294                 side_tracer_callback_variadic_func call_variadic,
 295                 void *priv)
 296 {
 297         if (!(desc->flags & SIDE_EVENT_FLAG_VARIADIC))
 298                 return SIDE_ERROR_INVAL;
 299         return _side_tracer_callback_unregister(desc, (void *) call_variadic, priv);
 300 }
 301
 302 struct side_events_register_handle *side_events_register(struct side_event_description **events, uint32_t nr_events)
 303 {
 304         struct side_events_register_handle *events_handle = NULL;
 305         struct side_tracer_handle *tracer_handle;
 306
 307         if (finalized)
 308                 return NULL;
 309         if (!initialized)
 310                 side_init();
 311         events_handle = (struct side_events_register_handle *)
 312                         calloc(1, sizeof(struct side_events_register_handle));
 313         if (!events_handle)
 314                 return NULL;
 315         events_handle->events = events;
 316         events_handle->nr_events = nr_events;
 317
 318         pthread_mutex_lock(&side_lock);
 319         side_list_insert_node_tail(&side_events_list, &events_handle->node);
 320         side_list_for_each_entry(tracer_handle, &side_tracer_list, node) {
 321                 tracer_handle->cb(SIDE_TRACER_NOTIFICATION_INSERT_EVENTS,
 322                         events, nr_events, tracer_handle->priv);
 323         }
 324         pthread_mutex_unlock(&side_lock);
 325         //TODO: call event batch register ioctl
 326         return events_handle;
 327 }
 328
 329 static
 330 void side_event_remove_callbacks(struct side_event_description *desc)
 331 {
 332         struct side_event_state *event_state = side_ptr_get(desc->state);
 333         struct side_event_state_0 *es0;
 334         struct side_callback *old_cb;
 335         uint32_t nr_cb;
 336
 337         if (side_unlikely(event_state->version != 0))
 338                 abort();
 339         es0 = side_container_of(event_state, struct side_event_state_0, parent);
 340         nr_cb = es0->nr_callbacks;
 341         if (!nr_cb)
 342                 return;
 343         old_cb = (struct side_callback *) es0->callbacks;
 344         (void) __atomic_add_fetch(&es0->enabled, -1, __ATOMIC_RELAXED);
 345         /*
 346          * Setting the state back to 0 cb and empty callbacks out of
 347          * caution. This should not matter because instrumentation is
 348          * unreachable.
 349          */
 350         es0->nr_callbacks = 0;
 351         side_rcu_assign_pointer(es0->callbacks, &side_empty_callback);
 352         /*
 353          * No need to wait for grace period because instrumentation is
 354          * unreachable.
 355          */
 356         free(old_cb);
 357 }
 358
 359 /*
 360  * Unregister event handle. At this point, all side events in that
 361  * handle should be unreachable.
 362  */
 363 void side_events_unregister(struct side_events_register_handle *events_handle)
 364 {
 365         struct side_tracer_handle *tracer_handle;
 366         uint32_t i;
 367
 368         if (!events_handle)
 369                 return;
 370         if (finalized)
 371                 return;
 372         if (!initialized)
 373                 side_init();
 374         pthread_mutex_lock(&side_lock);
 375         side_list_remove_node(&events_handle->node);
 376         side_list_for_each_entry(tracer_handle, &side_tracer_list, node) {
 377                 tracer_handle->cb(SIDE_TRACER_NOTIFICATION_REMOVE_EVENTS,
 378                         events_handle->events, events_handle->nr_events,
 379                         tracer_handle->priv);
 380         }
 381         for (i = 0; i < events_handle->nr_events; i++) {
 382                 struct side_event_description *event = events_handle->events[i];
 383
 384                 /* Skip NULL pointers */
 385                 if (!event)
 386                         continue;
 387                 side_event_remove_callbacks(event);
 388         }
 389         pthread_mutex_unlock(&side_lock);
 390         //TODO: call event batch unregister ioctl
 391         free(events_handle);
 392 }
 393
 394 struct side_tracer_handle *side_tracer_event_notification_register(
 395                 void (*cb)(enum side_tracer_notification notif,
 396                         struct side_event_description **events, uint32_t nr_events, void *priv),
 397                 void *priv)
 398 {
 399         struct side_tracer_handle *tracer_handle;
 400         struct side_events_register_handle *events_handle;
 401
 402         if (finalized)
 403                 return NULL;
 404         if (!initialized)
 405                 side_init();
 406         tracer_handle = (struct side_tracer_handle *)
 407                                 calloc(1, sizeof(struct side_tracer_handle));
 408         if (!tracer_handle)
 409                 return NULL;
 410         pthread_mutex_lock(&side_lock);
 411         tracer_handle->cb = cb;
 412         tracer_handle->priv = priv;
 413         side_list_insert_node_tail(&side_tracer_list, &tracer_handle->node);
 414         side_list_for_each_entry(events_handle, &side_events_list, node) {
 415                 cb(SIDE_TRACER_NOTIFICATION_INSERT_EVENTS,
 416                         events_handle->events, events_handle->nr_events, priv);
 417         }
 418         pthread_mutex_unlock(&side_lock);
 419         return tracer_handle;
 420 }
 421
 422 void side_tracer_event_notification_unregister(struct side_tracer_handle *tracer_handle)
 423 {
 424         struct side_events_register_handle *events_handle;
 425
 426         if (finalized)
 427                 return;
 428         if (!initialized)
 429                 side_init();
 430         pthread_mutex_lock(&side_lock);
 431         side_list_for_each_entry(events_handle, &side_events_list, node) {
 432                 tracer_handle->cb(SIDE_TRACER_NOTIFICATION_REMOVE_EVENTS,
 433                         events_handle->events, events_handle->nr_events,
 434                         tracer_handle->priv);
 435         }
 436         side_list_remove_node(&tracer_handle->node);
 437         pthread_mutex_unlock(&side_lock);
 438         free(tracer_handle);
 439 }
 440
 441 void side_init(void)
 442 {
 443         if (initialized)
 444                 return;
 445         side_rcu_gp_init(&rcu_gp);
 446         initialized = true;
 447 }
 448
 449 /*
 450  * side_exit() is executed from a library destructor. It can be called
 451  * explicitly at application exit as well. Concurrent side API use is
 452  * not expected at that point.
 453  */
 454 void side_exit(void)
 455 {
 456         struct side_events_register_handle *handle, *tmp;
 457
 458         if (finalized)
 459                 return;
 460         side_list_for_each_entry_safe(handle, tmp, &side_events_list, node)
 461                 side_events_unregister(handle);
 462         side_rcu_gp_exit(&rcu_gp);
 463         finalized = true;
 464 }