🐽 Fix memory leak in chunked streams handling code (#3631)

CHANGES/3631.bugfix

@@ -0,0 +1 @@
+Cleanup per-chunk data in generic data read. Memory leak fixed.

aiohttp/streams.py

@@ -460,6 +460,11 @@ def _read_nowait_chunk(self, n: int) -> bytes:
         self._size -= len(data)
         self._cursor += len(data)
 
+        chunk_splits = self._http_chunk_splits
+        # Prevent memory leak: drop useless chunk splits
+        while chunk_splits and chunk_splits[0] < self._cursor:
+            chunk_splits.pop(0)
+
         if self._size < self._low_water and self._protocol._reading_paused:
             self._protocol.resume_reading()
         return data

tests/test_streams.py

@@ -1,7 +1,12 @@
 """Tests for streams.py"""
 
+import abc
 import asyncio
+import gc
 import re
+import types
+from collections import defaultdict
+from itertools import groupby
 from unittest import mock
 
 import pytest
@@ -30,6 +35,36 @@ def protocol():
     return mock.Mock(_reading_paused=False)
 
 
+MEMLEAK_SKIP_TYPES = (
+    *(getattr(types, name) for name in types.__all__ if name.endswith('Type')),
+    mock.Mock,
+    abc.ABCMeta,
+)
+
+def get_memory_usage(obj):
+    objs=[obj]
+    # Memory leak may be caused by leaked links to same objects.
+    # Without link counting, [1,2,3] may be tracked the same as [1,2,3,3,3,3,3,3]
+    known = defaultdict(int)
+    known[id(obj)] += 1
+
+    while objs:
+        refs = gc.get_referents(*objs)
+        objs = []
+        for obj in refs:
+            if isinstance(obj, MEMLEAK_SKIP_TYPES):
+                continue
+            i = id(obj)
+            known[i] += 1
+            if known[i] == 1:
+                objs.append(obj)
+
+        # Make list of objects uniq
+        objs.sort(key=id)
+        objs = [next(g) for (i, g) in groupby(objs, id)]
+
+    return sum(cnt for cnt in known.values())
+
 class TestStreamReader:
 
     DATA = b'line1\nline2\nline3\n'
@@ -739,6 +774,9 @@ async def test_readchunk_with_other_read_calls(self) -> None:
         stream.begin_http_chunk_receiving()
         stream.feed_data(b'part2')
         stream.end_http_chunk_receiving()
+        stream.begin_http_chunk_receiving()
+        stream.feed_data(b'part3')
+        stream.end_http_chunk_receiving()
 
         data = await stream.read(7)
         assert b'part1pa' == data
@@ -747,11 +785,40 @@ async def test_readchunk_with_other_read_calls(self) -> None:
         assert b'rt2' == data
         assert end_of_chunk
 
+        # Corner case between read/readchunk
+        data = await stream.read(5)
+        assert b'part3' == data
+
+        data, end_of_chunk = await stream.readchunk()
+        assert b'' == data
+        assert end_of_chunk
+
         stream.feed_eof()
+
         data, end_of_chunk = await stream.readchunk()
         assert b'' == data
         assert not end_of_chunk
 
+    async def test_chunksplits_memory_leak(self) -> None:
+        """ Test for memoryleak on chunksplits """
+        stream = self._make_one()
+
+        # Warm + offset from standard numbers
+        stream.begin_http_chunk_receiving()
+        stream.feed_data(b'Y' * 500)
+        stream.end_http_chunk_receiving()
+        await stream.read(500)
+
+        before = get_memory_usage(stream)
+        for _ in range(300):
+            stream.begin_http_chunk_receiving()
+            stream.feed_data(b'X')
+            stream.end_http_chunk_receiving()
+        await stream.read(N)
+        after = get_memory_usage(stream)
+
+        assert abs(after - before) == 0
+
     async def test_read_empty_chunks(self) -> None:
         """Test that feeding empty chunks does not break stream"""
         stream = self._make_one()