Files
pycapnp/test/test_get_data_view.py
leoherz.liu 8b56bb8140 Fix: get_data_as_view retaion the backing reader/builder even after the view is released, leading to memory leakage (#406)
* fix: return empty memoryview for uninitialized DATA fields

Use a module-level sentinel when Cap'n Proto reports a NULL pointer with
zero size so PyBuffer_FillInfo receives a valid address for unset fields.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix: release buffer info if memoryview construction fails

PyBuffer_FillInfo pins `self` via buf.obj; call PyBuffer_Release on failure
so that reference is not leaked. This is safe for sentinel-backed empty views:
PyBuffer_Release only decrements buf.obj and does not free buf.buf.

Co-authored-by: Cursor <cursoragent@cursor.com>

* docs: clarify lifetime rules for zero-copy buffer views

Document borrowing semantics, mutation hazards, and empty DATA field
behavior for get_data_as_view and to_segment_views.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix: pin DATA field views via shared buffer exporter

Replace PyMemoryView_FromBuffer with a _BorrowedBufferView holder and
PyMemoryView_FromObject so get_data_as_view() correctly pins the struct
reader/builder for the memoryview lifetime. Generalize the same exporter
for to_segment_views() and add regression tests for packed payload release.

Co-authored-by: Cursor <cursoragent@cursor.com>

---------

Co-authored-by: bigtailfox <leoherz.liu@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-07-10 21:06:36 -07:00

293 lines
7.9 KiB
Python

import os
import tempfile
import weakref
from pathlib import Path
import pytest
import capnp
import sys
import gc
@pytest.fixture(scope="module")
def all_types():
"""Load the standard all_types.capnp schema."""
directory = os.path.dirname(__file__)
return capnp.load(os.path.join(directory, "all_types.capnp"))
def test_set_bytes_get_bytes(all_types):
"""
Scenario 1: Set Byte -> Get Byte
Verify standard behavior: writing bytes results in reading bytes.
"""
msg = all_types.TestAllTypes.new_message()
input_data = b"hello_world"
# Set
msg.dataField = input_data
# Get
output_data = msg.dataField
# Verify
assert isinstance(output_data, bytes)
assert output_data == input_data
def test_set_view_get_bytes(all_types):
"""
Scenario 2: Set View -> Get Byte
Verify compatibility: Passing a memoryview sets the data,
but standard attribute access returns a bytes copy.
"""
msg = all_types.TestAllTypes.new_message()
# Create a memoryview source
raw_source = bytearray(b"view_source")
view = memoryview(raw_source)
# Set via memoryview
msg.dataField = view
# Get via standard attribute
output_data = msg.dataField
# Verify
assert isinstance(output_data, bytes)
assert output_data == b"view_source"
def test_set_bytes_get_view_and_modify(all_types):
"""
Scenario 3: Set Byte -> Get View
Verify the high-performance API get_data_as_view.
The view must be writable and modifications must reflect in the message.
"""
msg = all_types.TestAllTypes.new_message()
# Initial write
msg.dataField = b"ABCDE"
# Get view via new API
view = msg.get_data_as_view("dataField")
# Verify view properties
assert isinstance(view, memoryview)
assert view.readonly is False
assert view.tobytes() == b"ABCDE"
# Verify in-place modification
view[0] = ord("Z") # Change 'A' to 'Z'
# Verify modification is reflected in standard access
assert msg.dataField == b"ZBCDE"
def test_reader_vs_builder_view(all_types):
"""
Verify that Builder views are writable, but Reader views are read-only.
"""
# 1. Builder phase
builder = all_types.TestAllTypes.new_message()
builder.dataField = b"test_rw"
builder_view = builder.get_data_as_view("dataField")
assert builder_view.readonly is False
builder_view[0] = ord("T") # Modification allowed
# 2. Reader phase
reader = builder.as_reader()
# Standard Get
assert reader.dataField == b"Test_rw"
# Reader get_data_as_view
reader_view = reader.get_data_as_view("dataField")
assert isinstance(reader_view, memoryview)
assert reader_view.readonly is True
# Attempting to modify Reader view should raise TypeError
with pytest.raises(TypeError):
reader_view[0] = ord("X")
def test_nested_struct_data(all_types):
"""
Verify that get_data_as_view works correctly on nested structs.
"""
msg = all_types.TestAllTypes.new_message()
# Initialize nested struct
inner = msg.init("structField")
inner.int32Field = 100
inner.dataField = b"nested_data"
# 1. Verify standard access
assert msg.structField.dataField == b"nested_data"
# 2. Verify nested get_data_as_view
view = msg.structField.get_data_as_view("dataField")
assert isinstance(view, memoryview)
assert view.tobytes() == b"nested_data"
# Modify nested data
view[0] = ord("N")
assert msg.structField.dataField == b"Nested_data"
def test_corner_cases_values(all_types):
"""
Test edge cases: Empty bytes and binary data with nulls.
"""
msg = all_types.TestAllTypes.new_message()
# Case A: Empty Bytes
msg.dataField = b""
assert msg.dataField == b""
view = msg.get_data_as_view("dataField")
assert len(view) == 0
# Case B: Binary data containing null bytes
binary_data = b"\x00\xff\x00\x01"
msg.dataField = binary_data
assert msg.dataField == binary_data
assert msg.get_data_as_view("dataField").tobytes() == binary_data
def test_uninitialized_data_get_view(all_types):
"""
Default DATA fields should expose an empty memoryview instead of failing on a NULL buffer pointer.
"""
builder = all_types.TestAllTypes.new_message()
builder_view = builder.get_data_as_view("dataField")
assert isinstance(builder_view, memoryview)
assert builder_view.readonly is False
assert len(builder_view) == 0
assert builder_view.tobytes() == b""
reader = all_types.TestAllTypes.new_message().as_reader()
reader_view = reader.get_data_as_view("dataField")
assert isinstance(reader_view, memoryview)
assert reader_view.readonly is True
assert len(reader_view) == 0
assert reader_view.tobytes() == b""
with pytest.raises(IndexError):
builder_view[0] = 0xFF
with pytest.raises(ValueError):
builder_view[0:1] = b"\xff"
def test_error_wrong_type(all_types):
"""
Test error handling: Calling get_data_as_view on non-Data fields.
"""
msg = all_types.TestAllTypes.new_message()
msg.int32Field = 123
msg.textField = "I am text"
# Attempt on Int field
with pytest.raises(TypeError) as excinfo:
msg.get_data_as_view("int32Field")
assert "not a DATA field" in str(excinfo.value)
# Attempt on Text field
with pytest.raises(TypeError) as excinfo:
msg.get_data_as_view("textField")
assert "not a DATA field" in str(excinfo.value)
def test_error_missing_field(all_types):
"""
Test error handling: Accessing a non-existent field name.
"""
msg = all_types.TestAllTypes.new_message()
# Accessing a missing field should raise AttributeError (standard Python behavior)
with pytest.raises(AttributeError) as excinfo:
msg.get_data_as_view("non_existent_field")
# Optional: Verify the error message contains the field name
assert "non_existent_field" in str(excinfo.value)
def test_view_keeps_message_alive(all_types):
"""
Verify that a View keeps messages alive.
"""
msg = all_types.TestAllTypes.new_message()
expected_data = b"persistence_check"
msg.dataField = expected_data
initial_ref_count = sys.getrefcount(msg)
view = msg.get_data_as_view("dataField")
new_ref_count = sys.getrefcount(msg)
assert new_ref_count > initial_ref_count, (
f"View failed to hold reference to Message! (Old: {initial_ref_count}, New: {new_ref_count})"
)
print(f"\n[Ref Check] Success: Ref count increased from {initial_ref_count} to {new_ref_count}")
del msg
gc.collect()
assert view.tobytes() == expected_data
def test_data_view_exports_through_buffer_exporter(all_types):
"""Returned memoryviews should pin an internal exporter, not bare pointers."""
msg = all_types.TestAllTypes.new_message()
msg.dataField = b"exporter_check"
view = msg.get_data_as_view("dataField")
assert isinstance(view, memoryview)
assert view.obj is not None
assert len(view.obj) == len(view)
def test_data_view_survives_del_builder(all_types):
msg = all_types.TestAllTypes.new_message()
msg.dataField = b"persistence_check"
view = msg.get_data_as_view("dataField")
del msg
gc.collect()
assert view.tobytes() == b"persistence_check"
def test_data_view_releases_packed_payload():
schema_text = """
@0x9d7d4f087df9b6e1;
struct BlobMsg {
data @0 :Data;
}
"""
class Payload(bytearray):
pass
td = tempfile.TemporaryDirectory()
path = Path(td.name) / "blob.capnp"
path.write_text(schema_text)
schema = capnp.load(str(path))
try:
payload = Payload(schema.BlobMsg.new_message(data=b"x" * 4096).to_bytes_packed())
payload_ref = weakref.ref(payload)
reader = schema.BlobMsg.from_bytes_packed(payload)
view = reader.get_data_as_view("data")
view.release()
del view, reader, payload
gc.collect()
assert payload_ref() is None
finally:
td.cleanup()