1 | /* |
2 | * Implementation of the Global Interpreter Lock (GIL). |
3 | */ |
4 | |
5 | #include <stdlib.h> |
6 | #include <errno.h> |
7 | |
8 | #include "pycore_atomic.h" |
9 | |
10 | |
11 | /* |
12 | Notes about the implementation: |
13 | |
14 | - The GIL is just a boolean variable (locked) whose access is protected |
15 | by a mutex (gil_mutex), and whose changes are signalled by a condition |
16 | variable (gil_cond). gil_mutex is taken for short periods of time, |
17 | and therefore mostly uncontended. |
18 | |
19 | - In the GIL-holding thread, the main loop (PyEval_EvalFrameEx) must be |
20 | able to release the GIL on demand by another thread. A volatile boolean |
21 | variable (gil_drop_request) is used for that purpose, which is checked |
22 | at every turn of the eval loop. That variable is set after a wait of |
23 | `interval` microseconds on `gil_cond` has timed out. |
24 | |
25 | [Actually, another volatile boolean variable (eval_breaker) is used |
26 | which ORs several conditions into one. Volatile booleans are |
27 | sufficient as inter-thread signalling means since Python is run |
28 | on cache-coherent architectures only.] |
29 | |
30 | - A thread wanting to take the GIL will first let pass a given amount of |
31 | time (`interval` microseconds) before setting gil_drop_request. This |
32 | encourages a defined switching period, but doesn't enforce it since |
33 | opcodes can take an arbitrary time to execute. |
34 | |
35 | The `interval` value is available for the user to read and modify |
36 | using the Python API `sys.{get,set}switchinterval()`. |
37 | |
38 | - When a thread releases the GIL and gil_drop_request is set, that thread |
39 | ensures that another GIL-awaiting thread gets scheduled. |
40 | It does so by waiting on a condition variable (switch_cond) until |
41 | the value of last_holder is changed to something else than its |
42 | own thread state pointer, indicating that another thread was able to |
43 | take the GIL. |
44 | |
45 | This is meant to prohibit the latency-adverse behaviour on multi-core |
46 | machines where one thread would speculatively release the GIL, but still |
47 | run and end up being the first to re-acquire it, making the "timeslices" |
48 | much longer than expected. |
49 | (Note: this mechanism is enabled with FORCE_SWITCHING above) |
50 | */ |
51 | |
52 | #include "condvar.h" |
53 | |
54 | #define MUTEX_INIT(mut) \ |
55 | if (PyMUTEX_INIT(&(mut))) { \ |
56 | Py_FatalError("PyMUTEX_INIT(" #mut ") failed"); }; |
57 | #define MUTEX_FINI(mut) \ |
58 | if (PyMUTEX_FINI(&(mut))) { \ |
59 | Py_FatalError("PyMUTEX_FINI(" #mut ") failed"); }; |
60 | #define MUTEX_LOCK(mut) \ |
61 | if (PyMUTEX_LOCK(&(mut))) { \ |
62 | Py_FatalError("PyMUTEX_LOCK(" #mut ") failed"); }; |
63 | #define MUTEX_UNLOCK(mut) \ |
64 | if (PyMUTEX_UNLOCK(&(mut))) { \ |
65 | Py_FatalError("PyMUTEX_UNLOCK(" #mut ") failed"); }; |
66 | |
67 | #define COND_INIT(cond) \ |
68 | if (PyCOND_INIT(&(cond))) { \ |
69 | Py_FatalError("PyCOND_INIT(" #cond ") failed"); }; |
70 | #define COND_FINI(cond) \ |
71 | if (PyCOND_FINI(&(cond))) { \ |
72 | Py_FatalError("PyCOND_FINI(" #cond ") failed"); }; |
73 | #define COND_SIGNAL(cond) \ |
74 | if (PyCOND_SIGNAL(&(cond))) { \ |
75 | Py_FatalError("PyCOND_SIGNAL(" #cond ") failed"); }; |
76 | #define COND_WAIT(cond, mut) \ |
77 | if (PyCOND_WAIT(&(cond), &(mut))) { \ |
78 | Py_FatalError("PyCOND_WAIT(" #cond ") failed"); }; |
79 | #define COND_TIMED_WAIT(cond, mut, microseconds, timeout_result) \ |
80 | { \ |
81 | int r = PyCOND_TIMEDWAIT(&(cond), &(mut), (microseconds)); \ |
82 | if (r < 0) \ |
83 | Py_FatalError("PyCOND_WAIT(" #cond ") failed"); \ |
84 | if (r) /* 1 == timeout, 2 == impl. can't say, so assume timeout */ \ |
85 | timeout_result = 1; \ |
86 | else \ |
87 | timeout_result = 0; \ |
88 | } \ |
89 | |
90 | |
91 | #define DEFAULT_INTERVAL 5000 |
92 | |
93 | static void _gil_initialize(struct _gil_runtime_state *gil) |
94 | { |
95 | _Py_atomic_int uninitialized = {-1}; |
96 | gil->locked = uninitialized; |
97 | gil->interval = DEFAULT_INTERVAL; |
98 | } |
99 | |
100 | static int gil_created(struct _gil_runtime_state *gil) |
101 | { |
102 | return (_Py_atomic_load_explicit(&gil->locked, _Py_memory_order_acquire) >= 0); |
103 | } |
104 | |
105 | static void create_gil(struct _gil_runtime_state *gil) |
106 | { |
107 | MUTEX_INIT(gil->mutex); |
108 | #ifdef FORCE_SWITCHING |
109 | MUTEX_INIT(gil->switch_mutex); |
110 | #endif |
111 | COND_INIT(gil->cond); |
112 | #ifdef FORCE_SWITCHING |
113 | COND_INIT(gil->switch_cond); |
114 | #endif |
115 | _Py_atomic_store_relaxed(&gil->last_holder, 0); |
116 | _Py_ANNOTATE_RWLOCK_CREATE(&gil->locked); |
117 | _Py_atomic_store_explicit(&gil->locked, 0, _Py_memory_order_release); |
118 | } |
119 | |
120 | static void destroy_gil(struct _gil_runtime_state *gil) |
121 | { |
122 | /* some pthread-like implementations tie the mutex to the cond |
123 | * and must have the cond destroyed first. |
124 | */ |
125 | COND_FINI(gil->cond); |
126 | MUTEX_FINI(gil->mutex); |
127 | #ifdef FORCE_SWITCHING |
128 | COND_FINI(gil->switch_cond); |
129 | MUTEX_FINI(gil->switch_mutex); |
130 | #endif |
131 | _Py_atomic_store_explicit(&gil->locked, -1, |
132 | _Py_memory_order_release); |
133 | _Py_ANNOTATE_RWLOCK_DESTROY(&gil->locked); |
134 | } |
135 | |
136 | static void recreate_gil(struct _gil_runtime_state *gil) |
137 | { |
138 | _Py_ANNOTATE_RWLOCK_DESTROY(&gil->locked); |
139 | /* XXX should we destroy the old OS resources here? */ |
140 | create_gil(gil); |
141 | } |
142 | |
143 | static void |
144 | drop_gil(struct _ceval_runtime_state *ceval, struct _ceval_state *ceval2, |
145 | PyThreadState *tstate) |
146 | { |
147 | #ifdef EXPERIMENTAL_ISOLATED_SUBINTERPRETERS |
148 | struct _gil_runtime_state *gil = &ceval2->gil; |
149 | #else |
150 | struct _gil_runtime_state *gil = &ceval->gil; |
151 | #endif |
152 | if (!_Py_atomic_load_relaxed(&gil->locked)) { |
153 | Py_FatalError("drop_gil: GIL is not locked" ); |
154 | } |
155 | |
156 | /* tstate is allowed to be NULL (early interpreter init) */ |
157 | if (tstate != NULL) { |
158 | /* Sub-interpreter support: threads might have been switched |
159 | under our feet using PyThreadState_Swap(). Fix the GIL last |
160 | holder variable so that our heuristics work. */ |
161 | _Py_atomic_store_relaxed(&gil->last_holder, (uintptr_t)tstate); |
162 | } |
163 | |
164 | MUTEX_LOCK(gil->mutex); |
165 | _Py_ANNOTATE_RWLOCK_RELEASED(&gil->locked, /*is_write=*/1); |
166 | _Py_atomic_store_relaxed(&gil->locked, 0); |
167 | COND_SIGNAL(gil->cond); |
168 | MUTEX_UNLOCK(gil->mutex); |
169 | |
170 | #ifdef FORCE_SWITCHING |
171 | if (_Py_atomic_load_relaxed(&ceval2->gil_drop_request) && tstate != NULL) { |
172 | MUTEX_LOCK(gil->switch_mutex); |
173 | /* Not switched yet => wait */ |
174 | if (((PyThreadState*)_Py_atomic_load_relaxed(&gil->last_holder)) == tstate) |
175 | { |
176 | assert(is_tstate_valid(tstate)); |
177 | RESET_GIL_DROP_REQUEST(tstate->interp); |
178 | /* NOTE: if COND_WAIT does not atomically start waiting when |
179 | releasing the mutex, another thread can run through, take |
180 | the GIL and drop it again, and reset the condition |
181 | before we even had a chance to wait for it. */ |
182 | COND_WAIT(gil->switch_cond, gil->switch_mutex); |
183 | } |
184 | MUTEX_UNLOCK(gil->switch_mutex); |
185 | } |
186 | #endif |
187 | } |
188 | |
189 | |
190 | /* Check if a Python thread must exit immediately, rather than taking the GIL |
191 | if Py_Finalize() has been called. |
192 | |
193 | When this function is called by a daemon thread after Py_Finalize() has been |
194 | called, the GIL does no longer exist. |
195 | |
196 | tstate must be non-NULL. */ |
197 | static inline int |
198 | tstate_must_exit(PyThreadState *tstate) |
199 | { |
200 | /* bpo-39877: Access _PyRuntime directly rather than using |
201 | tstate->interp->runtime to support calls from Python daemon threads. |
202 | After Py_Finalize() has been called, tstate can be a dangling pointer: |
203 | point to PyThreadState freed memory. */ |
204 | PyThreadState *finalizing = _PyRuntimeState_GetFinalizing(&_PyRuntime); |
205 | return (finalizing != NULL && finalizing != tstate); |
206 | } |
207 | |
208 | |
209 | /* Take the GIL. |
210 | |
211 | The function saves errno at entry and restores its value at exit. |
212 | |
213 | tstate must be non-NULL. */ |
214 | static void |
215 | take_gil(PyThreadState *tstate) |
216 | { |
217 | int err = errno; |
218 | |
219 | assert(tstate != NULL); |
220 | |
221 | if (tstate_must_exit(tstate)) { |
222 | /* bpo-39877: If Py_Finalize() has been called and tstate is not the |
223 | thread which called Py_Finalize(), exit immediately the thread. |
224 | |
225 | This code path can be reached by a daemon thread after Py_Finalize() |
226 | completes. In this case, tstate is a dangling pointer: points to |
227 | PyThreadState freed memory. */ |
228 | PyThread_exit_thread(); |
229 | } |
230 | |
231 | assert(is_tstate_valid(tstate)); |
232 | PyInterpreterState *interp = tstate->interp; |
233 | struct _ceval_runtime_state *ceval = &interp->runtime->ceval; |
234 | struct _ceval_state *ceval2 = &interp->ceval; |
235 | #ifdef EXPERIMENTAL_ISOLATED_SUBINTERPRETERS |
236 | struct _gil_runtime_state *gil = &ceval2->gil; |
237 | #else |
238 | struct _gil_runtime_state *gil = &ceval->gil; |
239 | #endif |
240 | |
241 | /* Check that _PyEval_InitThreads() was called to create the lock */ |
242 | assert(gil_created(gil)); |
243 | |
244 | MUTEX_LOCK(gil->mutex); |
245 | |
246 | if (!_Py_atomic_load_relaxed(&gil->locked)) { |
247 | goto _ready; |
248 | } |
249 | |
250 | while (_Py_atomic_load_relaxed(&gil->locked)) { |
251 | unsigned long saved_switchnum = gil->switch_number; |
252 | |
253 | unsigned long interval = (gil->interval >= 1 ? gil->interval : 1); |
254 | int timed_out = 0; |
255 | COND_TIMED_WAIT(gil->cond, gil->mutex, interval, timed_out); |
256 | |
257 | /* If we timed out and no switch occurred in the meantime, it is time |
258 | to ask the GIL-holding thread to drop it. */ |
259 | if (timed_out && |
260 | _Py_atomic_load_relaxed(&gil->locked) && |
261 | gil->switch_number == saved_switchnum) |
262 | { |
263 | if (tstate_must_exit(tstate)) { |
264 | MUTEX_UNLOCK(gil->mutex); |
265 | PyThread_exit_thread(); |
266 | } |
267 | assert(is_tstate_valid(tstate)); |
268 | |
269 | SET_GIL_DROP_REQUEST(interp); |
270 | } |
271 | } |
272 | |
273 | _ready: |
274 | #ifdef FORCE_SWITCHING |
275 | /* This mutex must be taken before modifying gil->last_holder: |
276 | see drop_gil(). */ |
277 | MUTEX_LOCK(gil->switch_mutex); |
278 | #endif |
279 | /* We now hold the GIL */ |
280 | _Py_atomic_store_relaxed(&gil->locked, 1); |
281 | _Py_ANNOTATE_RWLOCK_ACQUIRED(&gil->locked, /*is_write=*/1); |
282 | |
283 | if (tstate != (PyThreadState*)_Py_atomic_load_relaxed(&gil->last_holder)) { |
284 | _Py_atomic_store_relaxed(&gil->last_holder, (uintptr_t)tstate); |
285 | ++gil->switch_number; |
286 | } |
287 | |
288 | #ifdef FORCE_SWITCHING |
289 | COND_SIGNAL(gil->switch_cond); |
290 | MUTEX_UNLOCK(gil->switch_mutex); |
291 | #endif |
292 | |
293 | if (tstate_must_exit(tstate)) { |
294 | /* bpo-36475: If Py_Finalize() has been called and tstate is not |
295 | the thread which called Py_Finalize(), exit immediately the |
296 | thread. |
297 | |
298 | This code path can be reached by a daemon thread which was waiting |
299 | in take_gil() while the main thread called |
300 | wait_for_thread_shutdown() from Py_Finalize(). */ |
301 | MUTEX_UNLOCK(gil->mutex); |
302 | drop_gil(ceval, ceval2, tstate); |
303 | PyThread_exit_thread(); |
304 | } |
305 | assert(is_tstate_valid(tstate)); |
306 | |
307 | if (_Py_atomic_load_relaxed(&ceval2->gil_drop_request)) { |
308 | RESET_GIL_DROP_REQUEST(interp); |
309 | } |
310 | else { |
311 | /* bpo-40010: eval_breaker should be recomputed to be set to 1 if there |
312 | is a pending signal: signal received by another thread which cannot |
313 | handle signals. |
314 | |
315 | Note: RESET_GIL_DROP_REQUEST() calls COMPUTE_EVAL_BREAKER(). */ |
316 | COMPUTE_EVAL_BREAKER(interp, ceval, ceval2); |
317 | } |
318 | |
319 | /* Don't access tstate if the thread must exit */ |
320 | if (tstate->async_exc != NULL) { |
321 | _PyEval_SignalAsyncExc(tstate->interp); |
322 | } |
323 | |
324 | MUTEX_UNLOCK(gil->mutex); |
325 | |
326 | errno = err; |
327 | } |
328 | |
329 | void _PyEval_SetSwitchInterval(unsigned long microseconds) |
330 | { |
331 | #ifdef EXPERIMENTAL_ISOLATED_SUBINTERPRETERS |
332 | PyInterpreterState *interp = PyInterpreterState_Get(); |
333 | struct _gil_runtime_state *gil = &interp->ceval.gil; |
334 | #else |
335 | struct _gil_runtime_state *gil = &_PyRuntime.ceval.gil; |
336 | #endif |
337 | gil->interval = microseconds; |
338 | } |
339 | |
340 | unsigned long _PyEval_GetSwitchInterval() |
341 | { |
342 | #ifdef EXPERIMENTAL_ISOLATED_SUBINTERPRETERS |
343 | PyInterpreterState *interp = PyInterpreterState_Get(); |
344 | struct _gil_runtime_state *gil = &interp->ceval.gil; |
345 | #else |
346 | struct _gil_runtime_state *gil = &_PyRuntime.ceval.gil; |
347 | #endif |
348 | return gil->interval; |
349 | } |
350 | |