Movatterモバイル変換

[0]ホーム
URL:

Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

 Sign up
Appearance settings

Commitb24c916

Browse files
authored
gh-112529: Make the GC scheduling thread-safe (#114880)
The GC keeps track of the number of allocations (less deallocations)since the last GC. This buffers the count in thread-local state and usesatomic operations to modify the per-interpreter count. The thread-localbuffering avoids contention on shared state.A consequence is that the GC scheduling is not as precise, so"test_sneaky_frame_object" is skipped because it requires that the GC berun exactly after allocating a frame object.
1 parentf92857a commitb24c916

File tree

7 files changed

+71
-16
lines changed

7 files changed

+71
-16
lines changed

‎Include/internal/pycore_gc.h‎

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -260,6 +260,13 @@ struct _gc_runtime_state {
260260
Py_ssize_tlong_lived_pending;
261261
};
262262

263+
#ifdefPy_GIL_DISABLED
264+
struct_gc_thread_state {
265+
/* Thread-local allocation count. */
266+
Py_ssize_talloc_count;
267+
};
268+
#endif
269+
263270

264271
externvoid_PyGC_InitState(struct_gc_runtime_state*);
265272

‎Include/internal/pycore_tstate.h‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -28,6 +28,7 @@ typedef struct _PyThreadStateImpl {
2828
PyThreadStatebase;
2929

3030
#ifdefPy_GIL_DISABLED
31+
struct_gc_thread_stategc;
3132
struct_mimalloc_thread_statemimalloc;
3233
struct_Py_object_freelistsfreelists;
3334
struct_brc_thread_statebrc;

‎Lib/test/test_frame.py‎

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -13,7 +13,7 @@
1313
_testcapi=None
1414

1515
fromtestimportsupport
16-
fromtest.supportimportthreading_helper
16+
fromtest.supportimportthreading_helper,Py_GIL_DISABLED
1717
fromtest.support.script_helperimportassert_python_ok
1818

1919

@@ -294,6 +294,7 @@ def gen():
294294
assert_python_ok("-c",code)
295295

296296
@support.cpython_only
297+
@unittest.skipIf(Py_GIL_DISABLED,"test requires precise GC scheduling")
297298
deftest_sneaky_frame_object(self):
298299

299300
deftrace(frame,event,arg):

‎Lib/test/test_gc.py‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -363,6 +363,7 @@ def __del__(self):
363363
# To minimize variations, though, we first store the get_count() results
364364
# and check them at the end.
365365
@refcount_test
366+
@unittest.skipIf(Py_GIL_DISABLED,'needs precise allocation counts')
366367
deftest_get_count(self):
367368
gc.collect()
368369
a,b,c=gc.get_count()

‎Modules/gcmodule.c‎

Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -201,6 +201,16 @@ gc_get_count_impl(PyObject *module)
201201
/*[clinic end generated code: output=354012e67b16398f input=a392794a08251751]*/
202202
{
203203
GCState*gcstate=get_gc_state();
204+
205+
#ifdefPy_GIL_DISABLED
206+
_PyThreadStateImpl*tstate= (_PyThreadStateImpl*)_PyThreadState_GET();
207+
struct_gc_thread_state*gc=&tstate->gc;
208+
209+
// Flush the local allocation count to the global count
210+
_Py_atomic_add_int(&gcstate->generations[0].count, (int)gc->alloc_count);
211+
gc->alloc_count=0;
212+
#endif
213+
204214
returnPy_BuildValue("(iii)",
205215
gcstate->generations[0].count,
206216
gcstate->generations[1].count,

‎Objects/typeobject.c‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1835,6 +1835,8 @@ _PyType_AllocNoTrack(PyTypeObject *type, Py_ssize_t nitems)
18351835
if (presize) {
18361836
((PyObject**)alloc)[0]=NULL;
18371837
((PyObject**)alloc)[1]=NULL;
1838+
}
1839+
if (PyType_IS_GC(type)) {
18381840
_PyObject_GC_Link(obj);
18391841
}
18401842
memset(obj,'\0',size);

‎Python/gc_free_threading.c‎

Lines changed: 48 additions & 15 deletions
Original file line numberDiff line numberDiff line change
@@ -23,6 +23,11 @@ typedef struct _gc_runtime_state GCState;
2323
# defineGC_DEBUG
2424
#endif
2525

26+
// Each thread buffers the count of allocated objects in a thread-local
27+
// variable up to +/- this amount to reduce the overhead of updating
28+
// the global count.
29+
#defineLOCAL_ALLOC_COUNT_THRESHOLD 512
30+
2631
// Automatically choose the generation that needs collecting.
2732
#defineGENERATION_AUTO (-1)
2833

@@ -959,6 +964,41 @@ gc_should_collect(GCState *gcstate)
959964
gcstate->generations[1].threshold==0);
960965
}
961966

967+
staticvoid
968+
record_allocation(PyThreadState*tstate)
969+
{
970+
struct_gc_thread_state*gc=&((_PyThreadStateImpl*)tstate)->gc;
971+
972+
// We buffer the allocation count to avoid the overhead of atomic
973+
// operations for every allocation.
974+
gc->alloc_count++;
975+
if (gc->alloc_count >=LOCAL_ALLOC_COUNT_THRESHOLD) {
976+
// TODO: Use Py_ssize_t for the generation count.
977+
GCState*gcstate=&tstate->interp->gc;
978+
_Py_atomic_add_int(&gcstate->generations[0].count, (int)gc->alloc_count);
979+
gc->alloc_count=0;
980+
981+
if (gc_should_collect(gcstate)&&
982+
!_Py_atomic_load_int_relaxed(&gcstate->collecting))
983+
{
984+
_Py_ScheduleGC(tstate->interp);
985+
}
986+
}
987+
}
988+
989+
staticvoid
990+
record_deallocation(PyThreadState*tstate)
991+
{
992+
struct_gc_thread_state*gc=&((_PyThreadStateImpl*)tstate)->gc;
993+
994+
gc->alloc_count--;
995+
if (gc->alloc_count <=-LOCAL_ALLOC_COUNT_THRESHOLD) {
996+
GCState*gcstate=&tstate->interp->gc;
997+
_Py_atomic_add_int(&gcstate->generations[0].count, (int)gc->alloc_count);
998+
gc->alloc_count=0;
999+
}
1000+
}
1001+
9621002
staticvoid
9631003
gc_collect_internal(PyInterpreterState*interp,structcollection_state*state)
9641004
{
@@ -981,6 +1021,9 @@ gc_collect_internal(PyInterpreterState *interp, struct collection_state *state)
9811021
}
9821022
}
9831023

1024+
// Record the number of live GC objects
1025+
interp->gc.long_lived_total=state->long_lived_total;
1026+
9841027
// Clear weakrefs and enqueue callbacks (but do not call them).
9851028
clear_weakrefs(state);
9861029
_PyEval_StartTheWorld(interp);
@@ -1090,7 +1133,6 @@ gc_collect_main(PyThreadState *tstate, int generation, _PyGC_Reason reason)
10901133

10911134
m=state.collected;
10921135
n=state.uncollectable;
1093-
gcstate->long_lived_total=state.long_lived_total;
10941136

10951137
if (gcstate->debug&_PyGC_DEBUG_STATS) {
10961138
doubled=_PyTime_AsSecondsDouble(_PyTime_GetPerfCounter()-t1);
@@ -1530,15 +1572,7 @@ _Py_ScheduleGC(PyInterpreterState *interp)
15301572
void
15311573
_PyObject_GC_Link(PyObject*op)
15321574
{
1533-
PyThreadState*tstate=_PyThreadState_GET();
1534-
GCState*gcstate=&tstate->interp->gc;
1535-
gcstate->generations[0].count++;
1536-
1537-
if (gc_should_collect(gcstate)&&
1538-
!_Py_atomic_load_int_relaxed(&gcstate->collecting))
1539-
{
1540-
_Py_ScheduleGC(tstate->interp);
1541-
}
1575+
record_allocation(_PyThreadState_GET());
15421576
}
15431577

15441578
void
@@ -1564,7 +1598,7 @@ gc_alloc(PyTypeObject *tp, size_t basicsize, size_t presize)
15641598
((PyObject**)mem)[1]=NULL;
15651599
}
15661600
PyObject*op= (PyObject*)(mem+presize);
1567-
_PyObject_GC_Link(op);
1601+
record_allocation(tstate);
15681602
returnop;
15691603
}
15701604

@@ -1646,10 +1680,9 @@ PyObject_GC_Del(void *op)
16461680
PyErr_SetRaisedException(exc);
16471681
#endif
16481682
}
1649-
GCState*gcstate=get_gc_state();
1650-
if (gcstate->generations[0].count>0) {
1651-
gcstate->generations[0].count--;
1652-
}
1683+
1684+
record_deallocation(_PyThreadState_GET());
1685+
16531686
PyObject_Free(((char*)op)-presize);
16541687
}
16551688

0 commit comments

Comments
 (0)
[8]ページ先頭
©2009-2025 Movatter.jp