Repository navigation
Expand file tree
/
Copy pathperf_trampoline.c
More file actions
639 lines (565 loc) · 21 KB
/
Copy pathperf_trampoline.c
File metadata and controls
639 lines (565 loc) · 21 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
/*
Perf trampoline instrumentation
===============================
This file contains instrumentation to allow to associate
calls to the CPython eval loop back to the names of the Python
functions and filename being executed.
Many native performance profilers like the Linux perf tools are
only available to 'see' the C stack when sampling from the profiled
process. This means that if we have the following python code:
import time
def foo(n):
# Some CPU intensive code
def bar(n):
foo(n)
def baz(n):
bar(n)
baz(10000000)
A performance profiler that is only able to see native frames will
produce the following backtrace when sampling from foo():
_PyEval_EvalFrameDefault -----> Evaluation frame of foo()
_PyEval_Vector
_PyFunction_Vectorcall
PyObject_Vectorcall
call_function
_PyEval_EvalFrameDefault ------> Evaluation frame of bar()
_PyEval_EvalFrame
_PyEval_Vector
_PyFunction_Vectorcall
PyObject_Vectorcall
call_function
_PyEval_EvalFrameDefault -------> Evaluation frame of baz()
_PyEval_EvalFrame
_PyEval_Vector
_PyFunction_Vectorcall
PyObject_Vectorcall
call_function
...
Py_RunMain
Because the profiler is only able to see the native frames and the native
function that runs the evaluation loop is the same (_PyEval_EvalFrameDefault)
then the profiler and any reporter generated by it will not be able to
associate the names of the Python functions and the filenames associated with
those calls, rendering the results useless in the Python world.
To fix this problem, we introduce the concept of a trampoline frame. A
trampoline frame is a piece of code that is unique per Python code object that
is executed before entering the CPython eval loop. This piece of code just
calls the original Python evaluation function (_PyEval_EvalFrameDefault) and
forwards all the arguments received. In this way, when a profiler samples
frames from the previous example it will see;
_PyEval_EvalFrameDefault -----> Evaluation frame of foo()
[Jit compiled code 3]
_PyEval_Vector
_PyFunction_Vectorcall
PyObject_Vectorcall
call_function
_PyEval_EvalFrameDefault ------> Evaluation frame of bar()
[Jit compiled code 2]
_PyEval_EvalFrame
_PyEval_Vector
_PyFunction_Vectorcall
PyObject_Vectorcall
call_function
_PyEval_EvalFrameDefault -------> Evaluation frame of baz()
[Jit compiled code 1]
_PyEval_EvalFrame
_PyEval_Vector
_PyFunction_Vectorcall
PyObject_Vectorcall
call_function
...
Py_RunMain
When we generate every unique copy of the trampoline (what here we called "[Jit
compiled code N]") we write the relationship between the compiled code and the
Python function that is associated with it. Every profiler requires this
information in a different format. For example, the Linux "perf" profiler
requires a file in "/tmp/perf-PID.map" (name and location not configurable)
with the following format:
<compiled code address> <compiled code size> <name of the compiled code>
If this file is available when "perf" generates reports, it will automatically
associate every trampoline with the Python function that it is associated with
allowing it to generate reports that include Python information. These reports
then can also be filtered in a way that *only* Python information appears.
Notice that for this to work, there must be a unique copied of the trampoline
per Python code object even if the code in the trampoline is the same. To
achieve this we have a assembly template in Objects/asm_trampiline.S that is
compiled into the Python executable/shared library. This template generates a
symbol that maps the start of the assembly code and another that marks the end
of the assembly code for the trampoline. Then, every time we need a unique
trampoline for a Python code object, we copy the assembly code into a mmaped
area that has executable permissions and we return the start of that area as
our trampoline function.
Asking for a mmap-ed memory area for trampoline is very wasteful so we
allocate big arenas of memory in a single mmap call, we populate the entire
arena with copies of the trampoline (this allows us to now have to invalidate
the icache for the instructions in the page) and then we return the next
available chunk every time someone asks for a new trampoline. We keep a linked
list of arenas in case the current memory arena is exhausted and another one is
needed.
For the best results, Python should be compiled with
CFLAGS="-fno-omit-frame-pointer -mno-omit-leaf-frame-pointer" as this allows
profilers to unwind using only the frame pointer and not on DWARF debug
information (note that as trampilines are dynamically generated there won't be
any DWARF information available for them).
*/
#include "Python.h"
#include "pycore_ceval.h" // _PyPerf_Callbacks
#include "pycore_interpframe.h" // _PyFrame_GetCode()
#include "pycore_mmap.h" // _PyAnnotateMemoryMap()
#include "pycore_runtime.h" // _PyRuntime
#ifdef PY_HAVE_PERF_TRAMPOLINE
#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>
#include <sys/mman.h> // mmap()
#include <sys/types.h>
#include <unistd.h> // sysconf()
#include <sys/time.h> // gettimeofday()
#if defined(__arm__) || defined(__arm64__) || defined(__aarch64__)
#define PY_HAVE_INVALIDATE_ICACHE
#if defined(__clang__) || defined(__GNUC__)
extern void __clear_cache(void *, void*);
#endif
static void invalidate_icache(char* begin, char*end) {
#if defined(__clang__) || defined(__GNUC__)
return __clear_cache(begin, end);
#else
return;
#endif
}
#endif
/* The function pointer is passed as last argument. The other three arguments
* are passed in the same order as the function requires. This results in
* shorter, more efficient ASM code for trampoline.
*/
typedef PyObject *(*py_evaluator)(PyThreadState *, _PyInterpreterFrame *,
int throwflag);
typedef PyObject *(*py_trampoline)(PyThreadState *, _PyInterpreterFrame *, int,
py_evaluator);
extern void *_Py_trampoline_func_start; // Start of the template of the
// assembly trampoline
extern void *
_Py_trampoline_func_end; // End of the template of the assembly trampoline
struct code_arena_st {
char *start_addr; // Start of the memory arena
char *current_addr; // Address of the current trampoline within the arena
size_t size; // Size of the memory arena
size_t size_left; // Remaining size of the memory arena
size_t code_size; // Size of the code of every trampoline in the arena
struct code_arena_st
*prev; // Pointer to the arena or NULL if this is the first arena.
};
typedef struct code_arena_st code_arena_t;
typedef struct trampoline_api_st trampoline_api_t;
enum perf_trampoline_type {
PERF_TRAMPOLINE_UNSET = 0,
PERF_TRAMPOLINE_TYPE_MAP = 1,
PERF_TRAMPOLINE_TYPE_JITDUMP = 2,
};
#define perf_status _PyRuntime.ceval.perf.status
#define extra_code_index _PyRuntime.ceval.perf.extra_code_index
#define perf_code_arena _PyRuntime.ceval.perf.code_arena
#define trampoline_api _PyRuntime.ceval.perf.trampoline_api
#define perf_map_file _PyRuntime.ceval.perf.map_file
#define persist_after_fork _PyRuntime.ceval.perf.persist_after_fork
#define perf_trampoline_type _PyRuntime.ceval.perf.perf_trampoline_type
#define prev_eval_frame _PyRuntime.ceval.perf.prev_eval_frame
#define trampoline_refcount _PyRuntime.ceval.perf.trampoline_refcount
#define code_watcher_id _PyRuntime.ceval.perf.code_watcher_id
static void free_code_arenas(void);
static void
perf_trampoline_clear_code_watcher(void)
{
if (code_watcher_id >= 0) {
PyCode_ClearWatcher(code_watcher_id);
code_watcher_id = -1;
}
extra_code_index = -1;
}
static void
perf_trampoline_reset_state(void)
{
free_code_arenas();
perf_trampoline_clear_code_watcher();
}
static int
perf_trampoline_code_watcher(PyCodeEvent event, PyCodeObject *co)
{
if (event != PY_CODE_EVENT_DESTROY) {
return 0;
}
if (extra_code_index == -1) {
return 0;
}
py_trampoline f = NULL;
int ret = _PyCode_GetExtra((PyObject *)co, extra_code_index, (void **)&f);
if (ret != 0 || f == NULL) {
return 0;
}
trampoline_refcount--;
if (trampoline_refcount == 0) {
perf_trampoline_reset_state();
}
return 0;
}
static void
perf_map_write_entry(void *state, const void *code_addr,
size_t code_size, PyCodeObject *co)
{
const char *entry = "";
if (co->co_qualname != NULL) {
entry = PyUnicode_AsUTF8(co->co_qualname);
}
const char *filename = "";
if (co->co_filename != NULL) {
filename = PyUnicode_AsUTF8(co->co_filename);
}
size_t perf_map_entry_size = snprintf(NULL, 0, "py::%s:%s", entry, filename) + 1;
char* perf_map_entry = (char*) PyMem_RawMalloc(perf_map_entry_size);
if (perf_map_entry == NULL) {
return;
}
snprintf(perf_map_entry, perf_map_entry_size, "py::%s:%s", entry, filename);
PyUnstable_WritePerfMapEntry(code_addr, code_size, perf_map_entry);
PyMem_RawFree(perf_map_entry);
}
static void*
perf_map_init_state(void)
{
PyUnstable_PerfMapState_Init();
trampoline_api.code_padding = 0;
trampoline_api.code_alignment = 32;
perf_trampoline_type = PERF_TRAMPOLINE_TYPE_MAP;
return NULL;
}
static int
perf_map_free_state(void *state)
{
PyUnstable_PerfMapState_Fini();
return 0;
}
_PyPerf_Callbacks _Py_perfmap_callbacks = {
&perf_map_init_state,
&perf_map_write_entry,
&perf_map_free_state,
};
static size_t round_up(int64_t value, int64_t multiple) {
if (multiple == 0) {
// Avoid division by zero
return value;
}
int64_t remainder = value % multiple;
if (remainder == 0) {
// Value is already a multiple of 'multiple'
return value;
}
// Calculate the difference to the next multiple
int64_t difference = multiple - remainder;
// Add the difference to the value
int64_t rounded_up_value = value + difference;
return rounded_up_value;
}
// TRAMPOLINE MANAGEMENT API
static int
new_code_arena(void)
{
// non-trivial programs typically need 64 to 256 kiB.
size_t mem_size = 4096 * 16;
assert(mem_size % sysconf(_SC_PAGESIZE) == 0);
char *memory =
mmap(NULL, // address
mem_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS,
-1, // fd (not used here)
0); // offset (not used here)
if (memory == MAP_FAILED) {
PyErr_SetFromErrno(PyExc_OSError);
PyErr_FormatUnraisable("Failed to create new mmap for perf trampoline");
perf_status = PERF_STATUS_FAILED;
return -1;
}
(void)_PyAnnotateMemoryMap(memory, mem_size, "cpython:perf_trampoline");
void *start = &_Py_trampoline_func_start;
void *end = &_Py_trampoline_func_end;
size_t code_size = end - start;
size_t unaligned_size = code_size + trampoline_api.code_padding;
size_t chunk_size = round_up(unaligned_size, trampoline_api.code_alignment);
assert(chunk_size % trampoline_api.code_alignment == 0);
// TODO: Check the effect of alignment of the code chunks. Initial investigation
// showed that this has no effect on performance in x86-64 or aarch64 and the current
// version has the advantage that the unwinder in GDB can unwind across JIT-ed code.
//
// We should check the values in the future and see if there is a
// measurable performance improvement by rounding trampolines up to 32-bit
// or 64-bit alignment.
size_t n_copies = mem_size / chunk_size;
for (size_t i = 0; i < n_copies; i++) {
memcpy(memory + i * chunk_size, start, code_size * sizeof(char));
}
// Some systems may prevent us from creating executable code on the fly.
int res = mprotect(memory, mem_size, PROT_READ | PROT_EXEC);
if (res == -1) {
PyErr_SetFromErrno(PyExc_OSError);
munmap(memory, mem_size);
PyErr_FormatUnraisable("Failed to set mmap for perf trampoline to "
"PROT_READ | PROT_EXEC");
return -1;
}
#ifdef PY_HAVE_INVALIDATE_ICACHE
// Before the JIT can run a block of code that has been emitted it must invalidate
// the instruction cache on some platforms like arm and aarch64.
invalidate_icache(memory, memory + mem_size);
#endif
code_arena_t *new_arena = PyMem_RawCalloc(1, sizeof(code_arena_t));
if (new_arena == NULL) {
PyErr_NoMemory();
munmap(memory, mem_size);
PyErr_FormatUnraisable("Failed to allocate new code arena struct for perf trampoline");
return -1;
}
new_arena->start_addr = memory;
new_arena->current_addr = memory;
new_arena->size = mem_size;
new_arena->size_left = mem_size;
new_arena->code_size = code_size;
new_arena->prev = perf_code_arena;
perf_code_arena = new_arena;
return 0;
}
static void
free_code_arenas(void)
{
code_arena_t *cur = perf_code_arena;
code_arena_t *prev;
perf_code_arena = NULL; // invalid static pointer
while (cur) {
munmap(cur->start_addr, cur->size);
prev = cur->prev;
PyMem_RawFree(cur);
cur = prev;
}
}
static inline py_trampoline
code_arena_new_code(code_arena_t *code_arena)
{
py_trampoline trampoline = (py_trampoline)code_arena->current_addr;
size_t total_code_size = round_up(code_arena->code_size + trampoline_api.code_padding,
trampoline_api.code_alignment);
assert(total_code_size % trampoline_api.code_alignment == 0);
code_arena->size_left -= total_code_size;
code_arena->current_addr += total_code_size;
return trampoline;
}
static inline py_trampoline
compile_trampoline(void)
{
size_t total_code_size = round_up(perf_code_arena->code_size + trampoline_api.code_padding, 16);
if ((perf_code_arena == NULL) ||
(perf_code_arena->size_left <= total_code_size)) {
if (new_code_arena() < 0) {
return NULL;
}
}
assert(perf_code_arena->size_left <= perf_code_arena->size);
return code_arena_new_code(perf_code_arena);
}
static PyObject *
py_trampoline_evaluator(PyThreadState *ts, _PyInterpreterFrame *frame,
int throw)
{
if (perf_status == PERF_STATUS_FAILED ||
perf_status == PERF_STATUS_NO_INIT) {
goto default_eval;
}
PyCodeObject *co = _PyFrame_GetCode(frame);
py_trampoline f = NULL;
assert(extra_code_index != -1);
int ret = _PyCode_GetExtra((PyObject *)co, extra_code_index, (void **)&f);
if (ret != 0 || f == NULL) {
// This is the first time we see this code object so we need
// to compile a trampoline for it.
py_trampoline new_trampoline = compile_trampoline();
if (new_trampoline == NULL) {
goto default_eval;
}
trampoline_api.write_state(trampoline_api.state, new_trampoline,
perf_code_arena->code_size, co);
_PyCode_SetExtra((PyObject *)co, extra_code_index,
(void *)new_trampoline);
trampoline_refcount++;
f = new_trampoline;
}
assert(f != NULL);
return f(ts, frame, throw, prev_eval_frame != NULL ? prev_eval_frame : _PyEval_EvalFrameDefault);
default_eval:
// Something failed, fall back to the default evaluator.
if (prev_eval_frame) {
return prev_eval_frame(ts, frame, throw);
}
return _PyEval_EvalFrameDefault(ts, frame, throw);
}
#endif // PY_HAVE_PERF_TRAMPOLINE
int PyUnstable_PerfTrampoline_CompileCode(PyCodeObject *co)
{
#ifdef PY_HAVE_PERF_TRAMPOLINE
py_trampoline f = NULL;
assert(extra_code_index != -1);
int ret = _PyCode_GetExtra((PyObject *)co, extra_code_index, (void **)&f);
if (ret != 0 || f == NULL) {
py_trampoline new_trampoline = compile_trampoline();
if (new_trampoline == NULL) {
return 0;
}
trampoline_api.write_state(trampoline_api.state, new_trampoline,
perf_code_arena->code_size, co);
trampoline_refcount++;
return _PyCode_SetExtra((PyObject *)co, extra_code_index,
(void *)new_trampoline);
}
#endif // PY_HAVE_PERF_TRAMPOLINE
return 0;
}
int
_PyIsPerfTrampolineActive(void)
{
#ifdef PY_HAVE_PERF_TRAMPOLINE
PyThreadState *tstate = _PyThreadState_GET();
return tstate->interp->eval_frame == py_trampoline_evaluator;
#endif
return 0;
}
void
_PyPerfTrampoline_GetCallbacks(_PyPerf_Callbacks *callbacks)
{
if (callbacks == NULL) {
return;
}
#ifdef PY_HAVE_PERF_TRAMPOLINE
callbacks->init_state = trampoline_api.init_state;
callbacks->write_state = trampoline_api.write_state;
callbacks->free_state = trampoline_api.free_state;
#endif
return;
}
int
_PyPerfTrampoline_SetCallbacks(_PyPerf_Callbacks *callbacks)
{
if (callbacks == NULL) {
return -1;
}
#ifdef PY_HAVE_PERF_TRAMPOLINE
if (trampoline_api.state) {
_PyPerfTrampoline_Fini();
}
trampoline_api.init_state = callbacks->init_state;
trampoline_api.write_state = callbacks->write_state;
trampoline_api.free_state = callbacks->free_state;
trampoline_api.state = NULL;
#endif
return 0;
}
int
_PyPerfTrampoline_Init(int activate)
{
#ifdef PY_HAVE_PERF_TRAMPOLINE
PyThreadState *tstate = _PyThreadState_GET();
if (code_watcher_id == 0) {
// Initialize to -1 since 0 is a valid watcher ID
code_watcher_id = -1;
}
if (!activate) {
_PyInterpreterState_SetEvalFrameFunc(tstate->interp, prev_eval_frame);
perf_status = PERF_STATUS_NO_INIT;
}
else if (tstate->interp->eval_frame != py_trampoline_evaluator) {
prev_eval_frame = _PyInterpreterState_GetEvalFrameFunc(tstate->interp);
_PyInterpreterState_SetEvalFrameFunc(tstate->interp, py_trampoline_evaluator);
extra_code_index = _PyEval_RequestCodeExtraIndex(NULL);
if (extra_code_index == -1) {
return -1;
}
if (trampoline_api.state == NULL && trampoline_api.init_state != NULL) {
trampoline_api.state = trampoline_api.init_state();
}
if (new_code_arena() < 0) {
return -1;
}
code_watcher_id = PyCode_AddWatcher(perf_trampoline_code_watcher);
if (code_watcher_id < 0) {
PyErr_FormatUnraisable("Failed to register code watcher for perf trampoline");
free_code_arenas();
return -1;
}
trampoline_refcount = 1; // Base refcount held by the system
perf_status = PERF_STATUS_OK;
}
#endif
return 0;
}
int
_PyPerfTrampoline_Fini(void)
{
#ifdef PY_HAVE_PERF_TRAMPOLINE
if (perf_status != PERF_STATUS_OK) {
return 0;
}
PyThreadState *tstate = _PyThreadState_GET();
if (tstate->interp->eval_frame == py_trampoline_evaluator) {
_PyInterpreterState_SetEvalFrameFunc(tstate->interp, NULL);
}
if (perf_status == PERF_STATUS_OK) {
trampoline_api.free_state(trampoline_api.state);
perf_trampoline_type = PERF_TRAMPOLINE_UNSET;
}
// Prevent new trampolines from being created
perf_status = PERF_STATUS_NO_INIT;
// Decrement base refcount. If refcount reaches 0, all code objects are already
// dead so clean up now. Otherwise, watcher remains active to clean up when last
// code object dies; extra_code_index stays valid so watcher can identify them.
trampoline_refcount--;
if (trampoline_refcount == 0) {
perf_trampoline_reset_state();
}
#endif
return 0;
}
int
PyUnstable_PerfTrampoline_SetPersistAfterFork(int enable){
#ifdef PY_HAVE_PERF_TRAMPOLINE
persist_after_fork = enable;
return persist_after_fork;
#endif
return 0;
}
PyStatus
_PyPerfTrampoline_AfterFork_Child(void)
{
#ifdef PY_HAVE_PERF_TRAMPOLINE
if (persist_after_fork) {
if (perf_trampoline_type != PERF_TRAMPOLINE_TYPE_MAP) {
return PyStatus_Error("Failed to copy perf map file as perf trampoline type is not type map.");
}
_PyPerfTrampoline_Fini();
char filename[256];
pid_t parent_pid = getppid();
snprintf(filename, sizeof(filename), "/tmp/perf-%d.map", parent_pid);
if (PyUnstable_CopyPerfMapFile(filename) != 0) {
return PyStatus_Error("Failed to copy perf map file.");
}
} else {
// Restart trampoline in file in child.
int was_active = _PyIsPerfTrampolineActive();
_PyPerfTrampoline_Fini();
if (was_active) {
// After fork, Fini may leave the old code watcher registered
// if trampolined code objects from the parent still exist
// (trampoline_refcount > 0). Clear it unconditionally before
// Init registers a new one, but keep the old arenas mapped: the
// child may still need to return through trampoline frames that
// were on the C stack at fork().
perf_trampoline_clear_code_watcher();
_PyPerfTrampoline_Init(1);
}
}
#endif
return PyStatus_Ok();
}