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