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> | |
8 | ||
9 | #include "rcu.h" | |
10 | #include "list.h" | |
11 | ||
12 | /* Top 8 bits reserved for kernel tracer use. */ | |
13 | #if SIDE_BITS_PER_LONG == 64 | |
14 | # define SIDE_EVENT_ENABLED_KERNEL_MASK 0xFF00000000000000ULL | |
15 | # define SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK 0x8000000000000000ULL | |
16 | ||
17 | /* Allow 2^56 tracer references on an event. */ | |
18 | # define SIDE_EVENT_ENABLED_USER_MASK 0x00FFFFFFFFFFFFFFULL | |
19 | #else | |
20 | # define SIDE_EVENT_ENABLED_KERNEL_MASK 0xFF000000UL | |
21 | # define SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK 0x80000000UL | |
22 | ||
23 | /* Allow 2^24 tracer references on an event. */ | |
24 | # define SIDE_EVENT_ENABLED_USER_MASK 0x00FFFFFFUL | |
25 | #endif | |
26 | ||
27 | struct side_events_register_handle { | |
28 | struct side_list_node node; | |
29 | struct side_event_description **events; | |
30 | uint32_t nr_events; | |
31 | }; | |
32 | ||
33 | struct side_tracer_handle { | |
34 | struct side_list_node node; | |
35 | void (*cb)(enum side_tracer_notification notif, | |
36 | struct side_event_description **events, uint32_t nr_events, void *priv); | |
37 | void *priv; | |
38 | }; | |
39 | ||
40 | static struct side_rcu_gp_state rcu_gp; | |
41 | ||
42 | /* | |
43 | * Lazy initialization for early use within library constructors. | |
44 | */ | |
45 | static bool initialized; | |
46 | /* | |
47 | * Do not register/unregister any more events after destructor. | |
48 | */ | |
49 | static bool finalized; | |
50 | ||
51 | /* | |
52 | * Recursive mutex to allow tracer callbacks to use the side API. | |
53 | */ | |
54 | static pthread_mutex_t side_lock = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP; | |
55 | ||
56 | static DEFINE_SIDE_LIST_HEAD(side_events_list); | |
57 | static DEFINE_SIDE_LIST_HEAD(side_tracer_list); | |
58 | ||
59 | /* | |
60 | * The empty callback has a NULL function callback pointer, which stops | |
61 | * iteration on the array of callbacks immediately. | |
62 | */ | |
63 | const struct side_callback side_empty_callback = { }; | |
64 | ||
65 | void side_call(const struct side_event_description *desc, const struct side_arg_vec *side_arg_vec) | |
66 | { | |
0b9e59d6 | 67 | struct side_event_state *event_state; |
67337c4a MD |
68 | struct side_rcu_read_state rcu_read_state; |
69 | const struct side_callback *side_cb; | |
70 | uintptr_t enabled; | |
71 | ||
72 | if (side_unlikely(finalized)) | |
73 | return; | |
74 | if (side_unlikely(!initialized)) | |
75 | side_init(); | |
76 | if (side_unlikely(desc->flags & SIDE_EVENT_FLAG_VARIADIC)) { | |
77 | printf("ERROR: unexpected variadic event description\n"); | |
78 | abort(); | |
79 | } | |
0b9e59d6 MD |
80 | event_state = side_ptr_get(desc->state); |
81 | enabled = __atomic_load_n(&event_state->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); | |
0b9e59d6 | 86 | for (side_cb = side_rcu_dereference(event_state->callbacks); side_cb->u.call != NULL; side_cb++) |
67337c4a MD |
87 | side_cb->u.call(desc, side_arg_vec, side_cb->priv); |
88 | side_rcu_read_end(&rcu_gp, &rcu_read_state); | |
89 | } | |
90 | ||
91 | void side_call_variadic(const struct side_event_description *desc, | |
92 | const struct side_arg_vec *side_arg_vec, | |
93 | const struct side_arg_dynamic_struct *var_struct) | |
94 | { | |
0b9e59d6 | 95 | struct side_event_state *event_state; |
67337c4a MD |
96 | struct side_rcu_read_state rcu_read_state; |
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(); | |
104 | if (side_unlikely(!(desc->flags & SIDE_EVENT_FLAG_VARIADIC))) { | |
105 | printf("ERROR: unexpected non-variadic event description\n"); | |
106 | abort(); | |
107 | } | |
0b9e59d6 MD |
108 | event_state = side_ptr_get(desc->state); |
109 | enabled = __atomic_load_n(&event_state->enabled, __ATOMIC_RELAXED); | |
67337c4a MD |
110 | if (side_unlikely(enabled & SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK)) { |
111 | // TODO: call kernel write. | |
112 | } | |
113 | side_rcu_read_begin(&rcu_gp, &rcu_read_state); | |
0b9e59d6 | 114 | for (side_cb = side_rcu_dereference(event_state->callbacks); side_cb->u.call_variadic != NULL; side_cb++) |
67337c4a MD |
115 | side_cb->u.call_variadic(desc, side_arg_vec, var_struct, side_cb->priv); |
116 | side_rcu_read_end(&rcu_gp, &rcu_read_state); | |
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 | { | |
0b9e59d6 | 124 | struct side_event_state *event_state = side_ptr_get(desc->state); |
67337c4a MD |
125 | const struct side_callback *cb; |
126 | ||
0b9e59d6 | 127 | for (cb = event_state->callbacks; cb->u.call != NULL; cb++) { |
67337c4a MD |
128 | if ((void *) cb->u.call == call && cb->priv == priv) |
129 | return cb; | |
130 | } | |
131 | return NULL; | |
132 | } | |
133 | ||
134 | static | |
135 | int _side_tracer_callback_register(struct side_event_description *desc, | |
136 | void *call, void *priv) | |
137 | { | |
0b9e59d6 | 138 | struct side_event_state *event_state; |
67337c4a MD |
139 | struct side_callback *old_cb, *new_cb; |
140 | int ret = SIDE_ERROR_OK; | |
141 | uint32_t old_nr_cb; | |
142 | ||
143 | if (!call) | |
144 | return SIDE_ERROR_INVAL; | |
145 | if (finalized) | |
146 | return SIDE_ERROR_EXITING; | |
147 | if (!initialized) | |
148 | side_init(); | |
149 | pthread_mutex_lock(&side_lock); | |
0b9e59d6 MD |
150 | event_state = side_ptr_get(desc->state); |
151 | old_nr_cb = event_state->nr_callbacks; | |
67337c4a MD |
152 | if (old_nr_cb == UINT32_MAX) { |
153 | ret = SIDE_ERROR_INVAL; | |
154 | goto unlock; | |
155 | } | |
156 | /* Reject duplicate (call, priv) tuples. */ | |
157 | if (side_tracer_callback_lookup(desc, call, priv)) { | |
158 | ret = SIDE_ERROR_EXIST; | |
159 | goto unlock; | |
160 | } | |
0b9e59d6 | 161 | old_cb = (struct side_callback *) event_state->callbacks; |
67337c4a MD |
162 | /* old_nr_cb + 1 (new cb) + 1 (NULL) */ |
163 | new_cb = (struct side_callback *) calloc(old_nr_cb + 2, sizeof(struct side_callback)); | |
164 | if (!new_cb) { | |
165 | ret = SIDE_ERROR_NOMEM; | |
166 | goto unlock; | |
167 | } | |
168 | memcpy(new_cb, old_cb, old_nr_cb); | |
169 | if (desc->flags & SIDE_EVENT_FLAG_VARIADIC) | |
170 | new_cb[old_nr_cb].u.call_variadic = | |
171 | (side_tracer_callback_variadic_func) call; | |
172 | else | |
173 | new_cb[old_nr_cb].u.call = | |
174 | (side_tracer_callback_func) call; | |
175 | new_cb[old_nr_cb].priv = priv; | |
0b9e59d6 | 176 | side_rcu_assign_pointer(event_state->callbacks, new_cb); |
67337c4a MD |
177 | side_rcu_wait_grace_period(&rcu_gp); |
178 | if (old_nr_cb) | |
179 | free(old_cb); | |
0b9e59d6 | 180 | event_state->nr_callbacks++; |
67337c4a MD |
181 | /* Increment concurrently with kernel setting the top bits. */ |
182 | if (!old_nr_cb) | |
0b9e59d6 | 183 | (void) __atomic_add_fetch(&event_state->enabled, 1, __ATOMIC_RELAXED); |
67337c4a MD |
184 | unlock: |
185 | pthread_mutex_unlock(&side_lock); | |
186 | return ret; | |
187 | } | |
188 | ||
189 | int side_tracer_callback_register(struct side_event_description *desc, | |
190 | side_tracer_callback_func call, | |
191 | void *priv) | |
192 | { | |
193 | if (desc->flags & SIDE_EVENT_FLAG_VARIADIC) | |
194 | return SIDE_ERROR_INVAL; | |
195 | return _side_tracer_callback_register(desc, (void *) call, priv); | |
196 | } | |
197 | ||
198 | int side_tracer_callback_variadic_register(struct side_event_description *desc, | |
199 | side_tracer_callback_variadic_func call_variadic, | |
200 | void *priv) | |
201 | { | |
202 | if (!(desc->flags & SIDE_EVENT_FLAG_VARIADIC)) | |
203 | return SIDE_ERROR_INVAL; | |
204 | return _side_tracer_callback_register(desc, (void *) call_variadic, priv); | |
205 | } | |
206 | ||
207 | static int _side_tracer_callback_unregister(struct side_event_description *desc, | |
208 | void *call, void *priv) | |
209 | { | |
0b9e59d6 | 210 | struct side_event_state *event_state; |
67337c4a MD |
211 | struct side_callback *old_cb, *new_cb; |
212 | const struct side_callback *cb_pos; | |
213 | uint32_t pos_idx; | |
214 | int ret = SIDE_ERROR_OK; | |
215 | uint32_t old_nr_cb; | |
216 | ||
217 | if (!call) | |
218 | return SIDE_ERROR_INVAL; | |
219 | if (finalized) | |
220 | return SIDE_ERROR_EXITING; | |
221 | if (!initialized) | |
222 | side_init(); | |
223 | pthread_mutex_lock(&side_lock); | |
0b9e59d6 | 224 | event_state = side_ptr_get(desc->state); |
67337c4a MD |
225 | cb_pos = side_tracer_callback_lookup(desc, call, priv); |
226 | if (!cb_pos) { | |
227 | ret = SIDE_ERROR_NOENT; | |
228 | goto unlock; | |
229 | } | |
0b9e59d6 MD |
230 | old_nr_cb = event_state->nr_callbacks; |
231 | old_cb = (struct side_callback *) event_state->callbacks; | |
67337c4a MD |
232 | if (old_nr_cb == 1) { |
233 | new_cb = (struct side_callback *) &side_empty_callback; | |
234 | } else { | |
0b9e59d6 | 235 | pos_idx = cb_pos - event_state->callbacks; |
67337c4a MD |
236 | /* Remove entry at pos_idx. */ |
237 | /* old_nr_cb - 1 (removed cb) + 1 (NULL) */ | |
238 | new_cb = (struct side_callback *) calloc(old_nr_cb, sizeof(struct side_callback)); | |
239 | if (!new_cb) { | |
240 | ret = SIDE_ERROR_NOMEM; | |
241 | goto unlock; | |
242 | } | |
243 | memcpy(new_cb, old_cb, pos_idx); | |
244 | memcpy(&new_cb[pos_idx], &old_cb[pos_idx + 1], old_nr_cb - pos_idx - 1); | |
245 | } | |
0b9e59d6 | 246 | side_rcu_assign_pointer(event_state->callbacks, new_cb); |
67337c4a MD |
247 | side_rcu_wait_grace_period(&rcu_gp); |
248 | free(old_cb); | |
0b9e59d6 | 249 | event_state->nr_callbacks--; |
67337c4a MD |
250 | /* Decrement concurrently with kernel setting the top bits. */ |
251 | if (old_nr_cb == 1) | |
0b9e59d6 | 252 | (void) __atomic_add_fetch(&event_state->enabled, -1, __ATOMIC_RELAXED); |
67337c4a MD |
253 | unlock: |
254 | pthread_mutex_unlock(&side_lock); | |
255 | return ret; | |
256 | } | |
257 | ||
258 | int side_tracer_callback_unregister(struct side_event_description *desc, | |
259 | side_tracer_callback_func call, | |
260 | void *priv) | |
261 | { | |
262 | if (desc->flags & SIDE_EVENT_FLAG_VARIADIC) | |
263 | return SIDE_ERROR_INVAL; | |
264 | return _side_tracer_callback_unregister(desc, (void *) call, priv); | |
265 | } | |
266 | ||
267 | int side_tracer_callback_variadic_unregister(struct side_event_description *desc, | |
268 | side_tracer_callback_variadic_func call_variadic, | |
269 | void *priv) | |
270 | { | |
271 | if (!(desc->flags & SIDE_EVENT_FLAG_VARIADIC)) | |
272 | return SIDE_ERROR_INVAL; | |
273 | return _side_tracer_callback_unregister(desc, (void *) call_variadic, priv); | |
274 | } | |
275 | ||
276 | struct side_events_register_handle *side_events_register(struct side_event_description **events, uint32_t nr_events) | |
277 | { | |
278 | struct side_events_register_handle *events_handle = NULL; | |
279 | struct side_tracer_handle *tracer_handle; | |
280 | ||
281 | if (finalized) | |
282 | return NULL; | |
283 | if (!initialized) | |
284 | side_init(); | |
285 | events_handle = (struct side_events_register_handle *) | |
286 | calloc(1, sizeof(struct side_events_register_handle)); | |
287 | if (!events_handle) | |
288 | return NULL; | |
289 | events_handle->events = events; | |
290 | events_handle->nr_events = nr_events; | |
291 | ||
292 | pthread_mutex_lock(&side_lock); | |
293 | side_list_insert_node_tail(&side_events_list, &events_handle->node); | |
294 | side_list_for_each_entry(tracer_handle, &side_tracer_list, node) { | |
295 | tracer_handle->cb(SIDE_TRACER_NOTIFICATION_INSERT_EVENTS, | |
296 | events, nr_events, tracer_handle->priv); | |
297 | } | |
298 | pthread_mutex_unlock(&side_lock); | |
299 | //TODO: call event batch register ioctl | |
300 | return events_handle; | |
301 | } | |
302 | ||
303 | static | |
304 | void side_event_remove_callbacks(struct side_event_description *desc) | |
305 | { | |
0b9e59d6 MD |
306 | struct side_event_state *event_state = side_ptr_get(desc->state); |
307 | uint32_t nr_cb = event_state->nr_callbacks; | |
67337c4a MD |
308 | struct side_callback *old_cb; |
309 | ||
310 | if (!nr_cb) | |
311 | return; | |
0b9e59d6 MD |
312 | old_cb = (struct side_callback *) event_state->callbacks; |
313 | (void) __atomic_add_fetch(&event_state->enabled, -1, __ATOMIC_RELAXED); | |
67337c4a MD |
314 | /* |
315 | * Setting the state back to 0 cb and empty callbacks out of | |
316 | * caution. This should not matter because instrumentation is | |
317 | * unreachable. | |
318 | */ | |
0b9e59d6 MD |
319 | event_state->nr_callbacks = 0; |
320 | side_rcu_assign_pointer(event_state->callbacks, &side_empty_callback); | |
67337c4a MD |
321 | /* |
322 | * No need to wait for grace period because instrumentation is | |
323 | * unreachable. | |
324 | */ | |
325 | free(old_cb); | |
326 | } | |
327 | ||
328 | /* | |
329 | * Unregister event handle. At this point, all side events in that | |
330 | * handle should be unreachable. | |
331 | */ | |
332 | void side_events_unregister(struct side_events_register_handle *events_handle) | |
333 | { | |
334 | struct side_tracer_handle *tracer_handle; | |
335 | uint32_t i; | |
336 | ||
337 | if (!events_handle) | |
338 | return; | |
339 | if (finalized) | |
340 | return; | |
341 | if (!initialized) | |
342 | side_init(); | |
343 | pthread_mutex_lock(&side_lock); | |
344 | side_list_remove_node(&events_handle->node); | |
345 | side_list_for_each_entry(tracer_handle, &side_tracer_list, node) { | |
346 | tracer_handle->cb(SIDE_TRACER_NOTIFICATION_REMOVE_EVENTS, | |
347 | events_handle->events, events_handle->nr_events, | |
348 | tracer_handle->priv); | |
349 | } | |
350 | for (i = 0; i < events_handle->nr_events; i++) { | |
351 | struct side_event_description *event = events_handle->events[i]; | |
352 | ||
353 | /* Skip NULL pointers */ | |
354 | if (!event) | |
355 | continue; | |
356 | side_event_remove_callbacks(event); | |
357 | } | |
358 | pthread_mutex_unlock(&side_lock); | |
359 | //TODO: call event batch unregister ioctl | |
360 | free(events_handle); | |
361 | } | |
362 | ||
363 | struct side_tracer_handle *side_tracer_event_notification_register( | |
364 | void (*cb)(enum side_tracer_notification notif, | |
365 | struct side_event_description **events, uint32_t nr_events, void *priv), | |
366 | void *priv) | |
367 | { | |
368 | struct side_tracer_handle *tracer_handle; | |
369 | struct side_events_register_handle *events_handle; | |
370 | ||
371 | if (finalized) | |
372 | return NULL; | |
373 | if (!initialized) | |
374 | side_init(); | |
375 | tracer_handle = (struct side_tracer_handle *) | |
376 | calloc(1, sizeof(struct side_tracer_handle)); | |
377 | if (!tracer_handle) | |
378 | return NULL; | |
379 | pthread_mutex_lock(&side_lock); | |
380 | tracer_handle->cb = cb; | |
381 | tracer_handle->priv = priv; | |
382 | side_list_insert_node_tail(&side_tracer_list, &tracer_handle->node); | |
383 | side_list_for_each_entry(events_handle, &side_events_list, node) { | |
384 | cb(SIDE_TRACER_NOTIFICATION_INSERT_EVENTS, | |
385 | events_handle->events, events_handle->nr_events, priv); | |
386 | } | |
387 | pthread_mutex_unlock(&side_lock); | |
388 | return tracer_handle; | |
389 | } | |
390 | ||
391 | void side_tracer_event_notification_unregister(struct side_tracer_handle *tracer_handle) | |
392 | { | |
393 | struct side_events_register_handle *events_handle; | |
394 | ||
395 | if (finalized) | |
396 | return; | |
397 | if (!initialized) | |
398 | side_init(); | |
399 | pthread_mutex_lock(&side_lock); | |
400 | side_list_for_each_entry(events_handle, &side_events_list, node) { | |
401 | tracer_handle->cb(SIDE_TRACER_NOTIFICATION_REMOVE_EVENTS, | |
402 | events_handle->events, events_handle->nr_events, | |
403 | tracer_handle->priv); | |
404 | } | |
405 | side_list_remove_node(&tracer_handle->node); | |
406 | pthread_mutex_unlock(&side_lock); | |
be787080 | 407 | free(tracer_handle); |
67337c4a MD |
408 | } |
409 | ||
410 | void side_init(void) | |
411 | { | |
412 | if (initialized) | |
413 | return; | |
414 | side_rcu_gp_init(&rcu_gp); | |
415 | initialized = true; | |
416 | } | |
417 | ||
418 | /* | |
419 | * side_exit() is executed from a library destructor. It can be called | |
420 | * explicitly at application exit as well. Concurrent side API use is | |
421 | * not expected at that point. | |
422 | */ | |
423 | void side_exit(void) | |
424 | { | |
425 | struct side_events_register_handle *handle, *tmp; | |
426 | ||
427 | if (finalized) | |
428 | return; | |
429 | side_list_for_each_entry_safe(handle, tmp, &side_events_list, node) | |
430 | side_events_unregister(handle); | |
431 | side_rcu_gp_exit(&rcu_gp); | |
432 | finalized = true; | |
433 | } |