Commit | Line | Data |
---|---|---|
67337c4a MD |
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> | |
b1bf768c | 8 | #include <assert.h> |
67337c4a MD |
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 | 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 | ||
0d747f98 | 66 | void side_call(const struct side_event_state *event_state, const struct side_arg_vec *side_arg_vec) |
67337c4a MD |
67 | { |
68 | struct side_rcu_read_state rcu_read_state; | |
b2a84b9f | 69 | const struct side_event_state_0 *es0; |
67337c4a MD |
70 | const struct side_callback *side_cb; |
71 | uintptr_t enabled; | |
72 | ||
73 | if (side_unlikely(finalized)) | |
74 | return; | |
75 | if (side_unlikely(!initialized)) | |
76 | side_init(); | |
b2a84b9f MD |
77 | if (side_unlikely(event_state->version != 0)) |
78 | abort(); | |
79 | es0 = side_container_of(event_state, const struct side_event_state_0, p); | |
7269a8a3 | 80 | assert(!(es0->desc->flags & SIDE_EVENT_FLAG_VARIADIC)); |
b2a84b9f | 81 | enabled = __atomic_load_n(&es0->enabled, __ATOMIC_RELAXED); |
67337c4a MD |
82 | if (side_unlikely(enabled & SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK)) { |
83 | // TODO: call kernel write. | |
84 | } | |
85 | side_rcu_read_begin(&rcu_gp, &rcu_read_state); | |
7269a8a3 MD |
86 | for (side_cb = side_rcu_dereference(es0->callbacks); side_cb->u.call != NULL; side_cb++) |
87 | side_cb->u.call(es0->desc, side_arg_vec, side_cb->priv); | |
67337c4a MD |
88 | side_rcu_read_end(&rcu_gp, &rcu_read_state); |
89 | } | |
90 | ||
0d747f98 | 91 | void side_call_variadic(const struct side_event_state *event_state, |
67337c4a MD |
92 | const struct side_arg_vec *side_arg_vec, |
93 | const struct side_arg_dynamic_struct *var_struct) | |
94 | { | |
95 | struct side_rcu_read_state rcu_read_state; | |
b2a84b9f | 96 | const struct side_event_state_0 *es0; |
67337c4a MD |
97 | const struct side_callback *side_cb; |
98 | uintptr_t enabled; | |
99 | ||
100 | if (side_unlikely(finalized)) | |
101 | return; | |
102 | if (side_unlikely(!initialized)) | |
103 | side_init(); | |
b2a84b9f MD |
104 | if (side_unlikely(event_state->version != 0)) |
105 | abort(); | |
106 | es0 = side_container_of(event_state, const struct side_event_state_0, p); | |
7269a8a3 | 107 | assert(es0->desc->flags & SIDE_EVENT_FLAG_VARIADIC); |
b2a84b9f | 108 | enabled = __atomic_load_n(&es0->enabled, __ATOMIC_RELAXED); |
67337c4a MD |
109 | if (side_unlikely(enabled & SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK)) { |
110 | // TODO: call kernel write. | |
111 | } | |
112 | side_rcu_read_begin(&rcu_gp, &rcu_read_state); | |
7269a8a3 MD |
113 | for (side_cb = side_rcu_dereference(es0->callbacks); side_cb->u.call_variadic != NULL; side_cb++) |
114 | side_cb->u.call_variadic(es0->desc, side_arg_vec, var_struct, side_cb->priv); | |
67337c4a MD |
115 | side_rcu_read_end(&rcu_gp, &rcu_read_state); |
116 | } | |
117 | ||
118 | static | |
119 | const struct side_callback *side_tracer_callback_lookup( | |
120 | const struct side_event_description *desc, | |
121 | void *call, void *priv) | |
122 | { | |
0b9e59d6 | 123 | struct side_event_state *event_state = side_ptr_get(desc->state); |
b2a84b9f | 124 | const struct side_event_state_0 *es0; |
67337c4a MD |
125 | const struct side_callback *cb; |
126 | ||
b2a84b9f MD |
127 | if (side_unlikely(event_state->version != 0)) |
128 | abort(); | |
129 | es0 = side_container_of(event_state, const struct side_event_state_0, p); | |
7269a8a3 | 130 | for (cb = es0->callbacks; cb->u.call != NULL; cb++) { |
67337c4a MD |
131 | if ((void *) cb->u.call == call && cb->priv == priv) |
132 | return cb; | |
133 | } | |
134 | return NULL; | |
135 | } | |
136 | ||
137 | static | |
138 | int _side_tracer_callback_register(struct side_event_description *desc, | |
139 | void *call, void *priv) | |
140 | { | |
0b9e59d6 | 141 | struct side_event_state *event_state; |
67337c4a | 142 | struct side_callback *old_cb, *new_cb; |
b2a84b9f | 143 | struct side_event_state_0 *es0; |
67337c4a MD |
144 | int ret = SIDE_ERROR_OK; |
145 | uint32_t old_nr_cb; | |
146 | ||
147 | if (!call) | |
148 | return SIDE_ERROR_INVAL; | |
149 | if (finalized) | |
150 | return SIDE_ERROR_EXITING; | |
151 | if (!initialized) | |
152 | side_init(); | |
153 | pthread_mutex_lock(&side_lock); | |
0b9e59d6 | 154 | event_state = side_ptr_get(desc->state); |
b2a84b9f MD |
155 | if (side_unlikely(event_state->version != 0)) |
156 | abort(); | |
157 | es0 = side_container_of(event_state, struct side_event_state_0, p); | |
3cac1780 | 158 | old_nr_cb = es0->nr_callbacks; |
67337c4a MD |
159 | if (old_nr_cb == UINT32_MAX) { |
160 | ret = SIDE_ERROR_INVAL; | |
161 | goto unlock; | |
162 | } | |
163 | /* Reject duplicate (call, priv) tuples. */ | |
164 | if (side_tracer_callback_lookup(desc, call, priv)) { | |
165 | ret = SIDE_ERROR_EXIST; | |
166 | goto unlock; | |
167 | } | |
7269a8a3 | 168 | old_cb = (struct side_callback *) es0->callbacks; |
67337c4a MD |
169 | /* old_nr_cb + 1 (new cb) + 1 (NULL) */ |
170 | new_cb = (struct side_callback *) calloc(old_nr_cb + 2, sizeof(struct side_callback)); | |
171 | if (!new_cb) { | |
172 | ret = SIDE_ERROR_NOMEM; | |
173 | goto unlock; | |
174 | } | |
175 | memcpy(new_cb, old_cb, old_nr_cb); | |
176 | if (desc->flags & SIDE_EVENT_FLAG_VARIADIC) | |
177 | new_cb[old_nr_cb].u.call_variadic = | |
178 | (side_tracer_callback_variadic_func) call; | |
179 | else | |
180 | new_cb[old_nr_cb].u.call = | |
181 | (side_tracer_callback_func) call; | |
182 | new_cb[old_nr_cb].priv = priv; | |
f60d8121 | 183 | /* High order bits are already zeroed. */ |
7269a8a3 | 184 | side_rcu_assign_pointer(es0->callbacks, new_cb); |
67337c4a MD |
185 | side_rcu_wait_grace_period(&rcu_gp); |
186 | if (old_nr_cb) | |
187 | free(old_cb); | |
3cac1780 | 188 | es0->nr_callbacks++; |
67337c4a MD |
189 | /* Increment concurrently with kernel setting the top bits. */ |
190 | if (!old_nr_cb) | |
b2a84b9f | 191 | (void) __atomic_add_fetch(&es0->enabled, 1, __ATOMIC_RELAXED); |
67337c4a MD |
192 | unlock: |
193 | pthread_mutex_unlock(&side_lock); | |
194 | return ret; | |
195 | } | |
196 | ||
197 | int side_tracer_callback_register(struct side_event_description *desc, | |
198 | side_tracer_callback_func call, | |
199 | void *priv) | |
200 | { | |
201 | if (desc->flags & SIDE_EVENT_FLAG_VARIADIC) | |
202 | return SIDE_ERROR_INVAL; | |
203 | return _side_tracer_callback_register(desc, (void *) call, priv); | |
204 | } | |
205 | ||
206 | int side_tracer_callback_variadic_register(struct side_event_description *desc, | |
207 | side_tracer_callback_variadic_func call_variadic, | |
208 | void *priv) | |
209 | { | |
210 | if (!(desc->flags & SIDE_EVENT_FLAG_VARIADIC)) | |
211 | return SIDE_ERROR_INVAL; | |
212 | return _side_tracer_callback_register(desc, (void *) call_variadic, priv); | |
213 | } | |
214 | ||
215 | static int _side_tracer_callback_unregister(struct side_event_description *desc, | |
216 | void *call, void *priv) | |
217 | { | |
0b9e59d6 | 218 | struct side_event_state *event_state; |
67337c4a MD |
219 | struct side_callback *old_cb, *new_cb; |
220 | const struct side_callback *cb_pos; | |
b2a84b9f | 221 | struct side_event_state_0 *es0; |
67337c4a MD |
222 | uint32_t pos_idx; |
223 | int ret = SIDE_ERROR_OK; | |
224 | uint32_t old_nr_cb; | |
225 | ||
226 | if (!call) | |
227 | return SIDE_ERROR_INVAL; | |
228 | if (finalized) | |
229 | return SIDE_ERROR_EXITING; | |
230 | if (!initialized) | |
231 | side_init(); | |
232 | pthread_mutex_lock(&side_lock); | |
0b9e59d6 | 233 | event_state = side_ptr_get(desc->state); |
b2a84b9f MD |
234 | if (side_unlikely(event_state->version != 0)) |
235 | abort(); | |
236 | es0 = side_container_of(event_state, struct side_event_state_0, p); | |
67337c4a MD |
237 | cb_pos = side_tracer_callback_lookup(desc, call, priv); |
238 | if (!cb_pos) { | |
239 | ret = SIDE_ERROR_NOENT; | |
240 | goto unlock; | |
241 | } | |
3cac1780 | 242 | old_nr_cb = es0->nr_callbacks; |
7269a8a3 | 243 | old_cb = (struct side_callback *) es0->callbacks; |
67337c4a MD |
244 | if (old_nr_cb == 1) { |
245 | new_cb = (struct side_callback *) &side_empty_callback; | |
246 | } else { | |
7269a8a3 | 247 | pos_idx = cb_pos - es0->callbacks; |
67337c4a MD |
248 | /* Remove entry at pos_idx. */ |
249 | /* old_nr_cb - 1 (removed cb) + 1 (NULL) */ | |
250 | new_cb = (struct side_callback *) calloc(old_nr_cb, sizeof(struct side_callback)); | |
251 | if (!new_cb) { | |
252 | ret = SIDE_ERROR_NOMEM; | |
253 | goto unlock; | |
254 | } | |
255 | memcpy(new_cb, old_cb, pos_idx); | |
256 | memcpy(&new_cb[pos_idx], &old_cb[pos_idx + 1], old_nr_cb - pos_idx - 1); | |
257 | } | |
f60d8121 | 258 | /* High order bits are already zeroed. */ |
7269a8a3 | 259 | side_rcu_assign_pointer(es0->callbacks, new_cb); |
67337c4a MD |
260 | side_rcu_wait_grace_period(&rcu_gp); |
261 | free(old_cb); | |
3cac1780 | 262 | es0->nr_callbacks--; |
67337c4a MD |
263 | /* Decrement concurrently with kernel setting the top bits. */ |
264 | if (old_nr_cb == 1) | |
b2a84b9f | 265 | (void) __atomic_add_fetch(&es0->enabled, -1, __ATOMIC_RELAXED); |
67337c4a MD |
266 | unlock: |
267 | pthread_mutex_unlock(&side_lock); | |
268 | return ret; | |
269 | } | |
270 | ||
271 | int side_tracer_callback_unregister(struct side_event_description *desc, | |
272 | side_tracer_callback_func call, | |
273 | void *priv) | |
274 | { | |
275 | if (desc->flags & SIDE_EVENT_FLAG_VARIADIC) | |
276 | return SIDE_ERROR_INVAL; | |
277 | return _side_tracer_callback_unregister(desc, (void *) call, priv); | |
278 | } | |
279 | ||
280 | int side_tracer_callback_variadic_unregister(struct side_event_description *desc, | |
281 | side_tracer_callback_variadic_func call_variadic, | |
282 | void *priv) | |
283 | { | |
284 | if (!(desc->flags & SIDE_EVENT_FLAG_VARIADIC)) | |
285 | return SIDE_ERROR_INVAL; | |
286 | return _side_tracer_callback_unregister(desc, (void *) call_variadic, priv); | |
287 | } | |
288 | ||
289 | struct side_events_register_handle *side_events_register(struct side_event_description **events, uint32_t nr_events) | |
290 | { | |
291 | struct side_events_register_handle *events_handle = NULL; | |
292 | struct side_tracer_handle *tracer_handle; | |
293 | ||
294 | if (finalized) | |
295 | return NULL; | |
296 | if (!initialized) | |
297 | side_init(); | |
298 | events_handle = (struct side_events_register_handle *) | |
299 | calloc(1, sizeof(struct side_events_register_handle)); | |
300 | if (!events_handle) | |
301 | return NULL; | |
302 | events_handle->events = events; | |
303 | events_handle->nr_events = nr_events; | |
304 | ||
305 | pthread_mutex_lock(&side_lock); | |
306 | side_list_insert_node_tail(&side_events_list, &events_handle->node); | |
307 | side_list_for_each_entry(tracer_handle, &side_tracer_list, node) { | |
308 | tracer_handle->cb(SIDE_TRACER_NOTIFICATION_INSERT_EVENTS, | |
309 | events, nr_events, tracer_handle->priv); | |
310 | } | |
311 | pthread_mutex_unlock(&side_lock); | |
312 | //TODO: call event batch register ioctl | |
313 | return events_handle; | |
314 | } | |
315 | ||
316 | static | |
317 | void side_event_remove_callbacks(struct side_event_description *desc) | |
318 | { | |
0b9e59d6 | 319 | struct side_event_state *event_state = side_ptr_get(desc->state); |
b2a84b9f | 320 | struct side_event_state_0 *es0; |
67337c4a | 321 | struct side_callback *old_cb; |
3cac1780 | 322 | uint32_t nr_cb; |
67337c4a | 323 | |
b2a84b9f MD |
324 | if (side_unlikely(event_state->version != 0)) |
325 | abort(); | |
326 | es0 = side_container_of(event_state, struct side_event_state_0, p); | |
3cac1780 MD |
327 | nr_cb = es0->nr_callbacks; |
328 | if (!nr_cb) | |
329 | return; | |
7269a8a3 | 330 | old_cb = (struct side_callback *) es0->callbacks; |
b2a84b9f | 331 | (void) __atomic_add_fetch(&es0->enabled, -1, __ATOMIC_RELAXED); |
67337c4a MD |
332 | /* |
333 | * Setting the state back to 0 cb and empty callbacks out of | |
334 | * caution. This should not matter because instrumentation is | |
335 | * unreachable. | |
336 | */ | |
3cac1780 | 337 | es0->nr_callbacks = 0; |
7269a8a3 | 338 | side_rcu_assign_pointer(es0->callbacks, &side_empty_callback); |
67337c4a MD |
339 | /* |
340 | * No need to wait for grace period because instrumentation is | |
341 | * unreachable. | |
342 | */ | |
343 | free(old_cb); | |
344 | } | |
345 | ||
346 | /* | |
347 | * Unregister event handle. At this point, all side events in that | |
348 | * handle should be unreachable. | |
349 | */ | |
350 | void side_events_unregister(struct side_events_register_handle *events_handle) | |
351 | { | |
352 | struct side_tracer_handle *tracer_handle; | |
353 | uint32_t i; | |
354 | ||
355 | if (!events_handle) | |
356 | return; | |
357 | if (finalized) | |
358 | return; | |
359 | if (!initialized) | |
360 | side_init(); | |
361 | pthread_mutex_lock(&side_lock); | |
362 | side_list_remove_node(&events_handle->node); | |
363 | side_list_for_each_entry(tracer_handle, &side_tracer_list, node) { | |
364 | tracer_handle->cb(SIDE_TRACER_NOTIFICATION_REMOVE_EVENTS, | |
365 | events_handle->events, events_handle->nr_events, | |
366 | tracer_handle->priv); | |
367 | } | |
368 | for (i = 0; i < events_handle->nr_events; i++) { | |
369 | struct side_event_description *event = events_handle->events[i]; | |
370 | ||
371 | /* Skip NULL pointers */ | |
372 | if (!event) | |
373 | continue; | |
374 | side_event_remove_callbacks(event); | |
375 | } | |
376 | pthread_mutex_unlock(&side_lock); | |
377 | //TODO: call event batch unregister ioctl | |
378 | free(events_handle); | |
379 | } | |
380 | ||
381 | struct side_tracer_handle *side_tracer_event_notification_register( | |
382 | void (*cb)(enum side_tracer_notification notif, | |
383 | struct side_event_description **events, uint32_t nr_events, void *priv), | |
384 | void *priv) | |
385 | { | |
386 | struct side_tracer_handle *tracer_handle; | |
387 | struct side_events_register_handle *events_handle; | |
388 | ||
389 | if (finalized) | |
390 | return NULL; | |
391 | if (!initialized) | |
392 | side_init(); | |
393 | tracer_handle = (struct side_tracer_handle *) | |
394 | calloc(1, sizeof(struct side_tracer_handle)); | |
395 | if (!tracer_handle) | |
396 | return NULL; | |
397 | pthread_mutex_lock(&side_lock); | |
398 | tracer_handle->cb = cb; | |
399 | tracer_handle->priv = priv; | |
400 | side_list_insert_node_tail(&side_tracer_list, &tracer_handle->node); | |
401 | side_list_for_each_entry(events_handle, &side_events_list, node) { | |
402 | cb(SIDE_TRACER_NOTIFICATION_INSERT_EVENTS, | |
403 | events_handle->events, events_handle->nr_events, priv); | |
404 | } | |
405 | pthread_mutex_unlock(&side_lock); | |
406 | return tracer_handle; | |
407 | } | |
408 | ||
409 | void side_tracer_event_notification_unregister(struct side_tracer_handle *tracer_handle) | |
410 | { | |
411 | struct side_events_register_handle *events_handle; | |
412 | ||
413 | if (finalized) | |
414 | return; | |
415 | if (!initialized) | |
416 | side_init(); | |
417 | pthread_mutex_lock(&side_lock); | |
418 | side_list_for_each_entry(events_handle, &side_events_list, node) { | |
419 | tracer_handle->cb(SIDE_TRACER_NOTIFICATION_REMOVE_EVENTS, | |
420 | events_handle->events, events_handle->nr_events, | |
421 | tracer_handle->priv); | |
422 | } | |
423 | side_list_remove_node(&tracer_handle->node); | |
424 | pthread_mutex_unlock(&side_lock); | |
be787080 | 425 | free(tracer_handle); |
67337c4a MD |
426 | } |
427 | ||
428 | void side_init(void) | |
429 | { | |
430 | if (initialized) | |
431 | return; | |
432 | side_rcu_gp_init(&rcu_gp); | |
433 | initialized = true; | |
434 | } | |
435 | ||
436 | /* | |
437 | * side_exit() is executed from a library destructor. It can be called | |
438 | * explicitly at application exit as well. Concurrent side API use is | |
439 | * not expected at that point. | |
440 | */ | |
441 | void side_exit(void) | |
442 | { | |
443 | struct side_events_register_handle *handle, *tmp; | |
444 | ||
445 | if (finalized) | |
446 | return; | |
447 | side_list_for_each_entry_safe(handle, tmp, &side_events_list, node) | |
448 | side_events_unregister(handle); | |
449 | side_rcu_gp_exit(&rcu_gp); | |
450 | finalized = true; | |
451 | } |