diff options
Diffstat (limited to 'Objects')
| -rw-r--r-- | Objects/codeobject.c | 2 | ||||
| -rw-r--r-- | Objects/dictobject.c | 4 | ||||
| -rw-r--r-- | Objects/listobject.c | 3 | ||||
| -rw-r--r-- | Objects/namespaceobject.c | 7 | ||||
| -rw-r--r-- | Objects/object.c | 76 | ||||
| -rw-r--r-- | Objects/obmalloc.c | 96 |
6 files changed, 123 insertions, 65 deletions
diff --git a/Objects/codeobject.c b/Objects/codeobject.c index 3f53d4cfeb2..91772bc9d19 100644 --- a/Objects/codeobject.c +++ b/Objects/codeobject.c @@ -3369,7 +3369,7 @@ create_tlbc_lock_held(PyCodeObject *co, Py_ssize_t idx) } memcpy(new_tlbc->entries, tlbc->entries, tlbc->size * sizeof(void *)); _Py_atomic_store_ptr_release(&co->co_tlbc, new_tlbc); - _PyMem_FreeDelayed(tlbc); + _PyMem_FreeDelayed(tlbc, tlbc->size * sizeof(void *)); tlbc = new_tlbc; } char *bc = PyMem_Calloc(1, _PyCode_NBYTES(co)); diff --git a/Objects/dictobject.c b/Objects/dictobject.c index 72901377cbb..6b7b150f0e2 100644 --- a/Objects/dictobject.c +++ b/Objects/dictobject.c @@ -813,7 +813,7 @@ free_keys_object(PyDictKeysObject *keys, bool use_qsbr) { #ifdef Py_GIL_DISABLED if (use_qsbr) { - _PyMem_FreeDelayed(keys); + _PyMem_FreeDelayed(keys, _PyDict_KeysSize(keys)); return; } #endif @@ -858,7 +858,7 @@ free_values(PyDictValues *values, bool use_qsbr) assert(values->embedded == 0); #ifdef Py_GIL_DISABLED if (use_qsbr) { - _PyMem_FreeDelayed(values); + _PyMem_FreeDelayed(values, values_size_from_count(values->capacity)); return; } #endif diff --git a/Objects/listobject.c b/Objects/listobject.c index c5895645a2d..23d3472b6d4 100644 --- a/Objects/listobject.c +++ b/Objects/listobject.c @@ -61,7 +61,8 @@ free_list_items(PyObject** items, bool use_qsbr) #ifdef Py_GIL_DISABLED _PyListArray *array = _Py_CONTAINER_OF(items, _PyListArray, ob_item); if (use_qsbr) { - _PyMem_FreeDelayed(array); + size_t size = sizeof(_PyListArray) + array->allocated * sizeof(PyObject *); + _PyMem_FreeDelayed(array, size); } else { PyMem_Free(array); diff --git a/Objects/namespaceobject.c b/Objects/namespaceobject.c index caebe6bf543..0fc2bcea4cb 100644 --- a/Objects/namespaceobject.c +++ b/Objects/namespaceobject.c @@ -124,9 +124,10 @@ namespace_repr(PyObject *ns) if (PyUnicode_Check(key) && PyUnicode_GET_LENGTH(key) > 0) { PyObject *value, *item; - value = PyDict_GetItemWithError(d, key); - if (value != NULL) { + int has_key = PyDict_GetItemRef(d, key, &value); + if (has_key == 1) { item = PyUnicode_FromFormat("%U=%R", key, value); + Py_DECREF(value); if (item == NULL) { loop_error = 1; } @@ -135,7 +136,7 @@ namespace_repr(PyObject *ns) Py_DECREF(item); } } - else if (PyErr_Occurred()) { + else if (has_key < 0) { loop_error = 1; } } diff --git a/Objects/object.c b/Objects/object.c index 6a581a19604..4d60128b092 100644 --- a/Objects/object.c +++ b/Objects/object.c @@ -3034,57 +3034,28 @@ finally: /* Trashcan support. */ -#ifndef Py_GIL_DISABLED -/* We need to store a pointer in the refcount field of - * an object. It is important that we never store 0 (NULL). -* It is also important to not make the object appear immortal, -* or it might be untracked by the cycle GC. */ -static uintptr_t -pointer_to_safe_refcount(void *ptr) -{ - uintptr_t full = (uintptr_t)ptr; - assert((full & 3) == 0); -#if SIZEOF_VOID_P > 4 - uint32_t refcnt = (uint32_t)full; - if (refcnt >= (uint32_t)_Py_IMMORTAL_MINIMUM_REFCNT) { - full = full - ((uintptr_t)_Py_IMMORTAL_MINIMUM_REFCNT) + 1; - } - return full + 2; -#else - // Make the top two bits 0, so it appears mortal. - return (full >> 2) + 1; -#endif -} - -static void * -safe_refcount_to_pointer(uintptr_t refcnt) -{ -#if SIZEOF_VOID_P > 4 - if (refcnt & 1) { - refcnt += _Py_IMMORTAL_MINIMUM_REFCNT - 1; - } - return (void *)(refcnt - 2); -#else - return (void *)((refcnt -1) << 2); -#endif -} -#endif - /* Add op to the gcstate->trash_delete_later list. Called when the current - * call-stack depth gets large. op must be a currently untracked gc'ed - * object, with refcount 0. Py_DECREF must already have been called on it. + * call-stack depth gets large. op must be a gc'ed object, with refcount 0. + * Py_DECREF must already have been called on it. */ void _PyTrash_thread_deposit_object(PyThreadState *tstate, PyObject *op) { _PyObject_ASSERT(op, Py_REFCNT(op) == 0); + PyTypeObject *tp = Py_TYPE(op); + assert(tp->tp_flags & Py_TPFLAGS_HAVE_GC); + int tracked = 0; + if (tp->tp_is_gc == NULL || tp->tp_is_gc(op)) { + tracked = _PyObject_GC_IS_TRACKED(op); + if (tracked) { + _PyObject_GC_UNTRACK(op); + } + } + uintptr_t tagged_ptr = ((uintptr_t)tstate->delete_later) | tracked; #ifdef Py_GIL_DISABLED - op->ob_tid = (uintptr_t)tstate->delete_later; + op->ob_tid = tagged_ptr; #else - /* Store the delete_later pointer in the refcnt field. */ - uintptr_t refcnt = pointer_to_safe_refcount(tstate->delete_later); - *((uintptr_t*)op) = refcnt; - assert(!_Py_IsImmortal(op)); + _Py_AS_GC(op)->_gc_next = tagged_ptr; #endif tstate->delete_later = op; } @@ -3099,17 +3070,17 @@ _PyTrash_thread_destroy_chain(PyThreadState *tstate) destructor dealloc = Py_TYPE(op)->tp_dealloc; #ifdef Py_GIL_DISABLED - tstate->delete_later = (PyObject*) op->ob_tid; + uintptr_t tagged_ptr = op->ob_tid; op->ob_tid = 0; _Py_atomic_store_ssize_relaxed(&op->ob_ref_shared, _Py_REF_MERGED); #else - /* Get the delete_later pointer from the refcnt field. - * See _PyTrash_thread_deposit_object(). */ - uintptr_t refcnt = *((uintptr_t*)op); - tstate->delete_later = safe_refcount_to_pointer(refcnt); - op->ob_refcnt = 0; + uintptr_t tagged_ptr = _Py_AS_GC(op)->_gc_next; + _Py_AS_GC(op)->_gc_next = 0; #endif - + tstate->delete_later = (PyObject *)(tagged_ptr & ~1); + if (tagged_ptr & 1) { + _PyObject_GC_TRACK(op); + } /* Call the deallocator directly. This used to try to * fool Py_DECREF into calling it indirectly, but * Py_DECREF was already called on this object, and in @@ -3183,10 +3154,11 @@ void _Py_Dealloc(PyObject *op) { PyTypeObject *type = Py_TYPE(op); + unsigned long gc_flag = type->tp_flags & Py_TPFLAGS_HAVE_GC; destructor dealloc = type->tp_dealloc; PyThreadState *tstate = _PyThreadState_GET(); intptr_t margin = _Py_RecursionLimit_GetMargin(tstate); - if (margin < 2) { + if (margin < 2 && gc_flag) { _PyTrash_thread_deposit_object(tstate, (PyObject *)op); return; } @@ -3232,7 +3204,7 @@ _Py_Dealloc(PyObject *op) Py_XDECREF(old_exc); Py_DECREF(type); #endif - if (tstate->delete_later && margin >= 4) { + if (tstate->delete_later && margin >= 4 && gc_flag) { _PyTrash_thread_destroy_chain(tstate); } } diff --git a/Objects/obmalloc.c b/Objects/obmalloc.c index d4b8327cb73..deb7fd957e5 100644 --- a/Objects/obmalloc.c +++ b/Objects/obmalloc.c @@ -124,6 +124,33 @@ _PyMem_mi_page_is_safe_to_free(mi_page_t *page) } +#ifdef Py_GIL_DISABLED + +// If we are deferring collection of more than this amount of memory for +// mimalloc pages, advance the write sequence. Advancing allows these +// pages to be re-used in a different thread or for a different size class. +#define QSBR_PAGE_MEM_LIMIT 4096*20 + +// Return true if the global write sequence should be advanced for a mimalloc +// page that is deferred from collection. +static bool +should_advance_qsbr_for_page(struct _qsbr_thread_state *qsbr, mi_page_t *page) +{ + size_t bsize = mi_page_block_size(page); + size_t page_size = page->capacity*bsize; + if (page_size > QSBR_PAGE_MEM_LIMIT) { + qsbr->deferred_page_memory = 0; + return true; + } + qsbr->deferred_page_memory += page_size; + if (qsbr->deferred_page_memory > QSBR_PAGE_MEM_LIMIT) { + qsbr->deferred_page_memory = 0; + return true; + } + return false; +} +#endif + static bool _PyMem_mi_page_maybe_free(mi_page_t *page, mi_page_queue_t *pq, bool force) { @@ -139,7 +166,14 @@ _PyMem_mi_page_maybe_free(mi_page_t *page, mi_page_queue_t *pq, bool force) _PyMem_mi_page_clear_qsbr(page); page->retire_expire = 0; - page->qsbr_goal = _Py_qsbr_deferred_advance(tstate->qsbr); + + if (should_advance_qsbr_for_page(tstate->qsbr, page)) { + page->qsbr_goal = _Py_qsbr_advance(tstate->qsbr->shared); + } + else { + page->qsbr_goal = _Py_qsbr_shared_next(tstate->qsbr->shared); + } + llist_insert_tail(&tstate->mimalloc.page_list, &page->qsbr_node); return false; } @@ -1141,8 +1175,44 @@ free_work_item(uintptr_t ptr, delayed_dealloc_cb cb, void *state) } } + +#ifdef Py_GIL_DISABLED + +// For deferred advance on free: the number of deferred items before advancing +// the write sequence. This is based on WORK_ITEMS_PER_CHUNK. We ideally +// want to process a chunk before it overflows. +#define QSBR_DEFERRED_LIMIT 127 + +// If the deferred memory exceeds 1 MiB, advance the write sequence. This +// helps limit memory usage due to QSBR delaying frees too long. +#define QSBR_FREE_MEM_LIMIT 1024*1024 + +// Return true if the global write sequence should be advanced for a deferred +// memory free. +static bool +should_advance_qsbr_for_free(struct _qsbr_thread_state *qsbr, size_t size) +{ + if (size > QSBR_FREE_MEM_LIMIT) { + qsbr->deferred_count = 0; + qsbr->deferred_memory = 0; + qsbr->should_process = true; + return true; + } + qsbr->deferred_count++; + qsbr->deferred_memory += size; + if (qsbr->deferred_count > QSBR_DEFERRED_LIMIT || + qsbr->deferred_memory > QSBR_FREE_MEM_LIMIT) { + qsbr->deferred_count = 0; + qsbr->deferred_memory = 0; + qsbr->should_process = true; + return true; + } + return false; +} +#endif + static void -free_delayed(uintptr_t ptr) +free_delayed(uintptr_t ptr, size_t size) { #ifndef Py_GIL_DISABLED free_work_item(ptr, NULL, NULL); @@ -1200,23 +1270,32 @@ free_delayed(uintptr_t ptr) } assert(buf != NULL && buf->wr_idx < WORK_ITEMS_PER_CHUNK); - uint64_t seq = _Py_qsbr_deferred_advance(tstate->qsbr); + uint64_t seq; + if (should_advance_qsbr_for_free(tstate->qsbr, size)) { + seq = _Py_qsbr_advance(tstate->qsbr->shared); + } + else { + seq = _Py_qsbr_shared_next(tstate->qsbr->shared); + } buf->array[buf->wr_idx].ptr = ptr; buf->array[buf->wr_idx].qsbr_goal = seq; buf->wr_idx++; if (buf->wr_idx == WORK_ITEMS_PER_CHUNK) { + // Normally the processing of delayed items is done from the eval + // breaker. Processing here is a safety measure to ensure too much + // work does not accumulate. _PyMem_ProcessDelayed((PyThreadState *)tstate); } #endif } void -_PyMem_FreeDelayed(void *ptr) +_PyMem_FreeDelayed(void *ptr, size_t size) { assert(!((uintptr_t)ptr & 0x01)); if (ptr != NULL) { - free_delayed((uintptr_t)ptr); + free_delayed((uintptr_t)ptr, size); } } @@ -1226,7 +1305,10 @@ _PyObject_XDecRefDelayed(PyObject *ptr) { assert(!((uintptr_t)ptr & 0x01)); if (ptr != NULL) { - free_delayed(((uintptr_t)ptr)|0x01); + // We use 0 as the size since we don't have an easy way to know the + // actual size. If we are freeing many objects, the write sequence + // will be advanced due to QSBR_DEFERRED_LIMIT. + free_delayed(((uintptr_t)ptr)|0x01, 0); } } #endif @@ -1317,6 +1399,8 @@ _PyMem_ProcessDelayed(PyThreadState *tstate) PyInterpreterState *interp = tstate->interp; _PyThreadStateImpl *tstate_impl = (_PyThreadStateImpl *)tstate; + tstate_impl->qsbr->should_process = false; + // Process thread-local work process_queue(&tstate_impl->mem_free_queue, tstate_impl, true, NULL, NULL); |