Repository navigation
Remove copy from bytearray to bytes in rawiobase_read() implementation #60107
Description
Activity
sbt commented
on Sep 10, 2012 sbtmannequinMannequinAuthorCurrently rawiobase_read() reads to a bytearray object and then copies the data to a bytes object.
There is a TODO comment saying that the bytes object should be created directly. The attached patch does that.
It works as along as the bytes object cannot leak to Python code, (imagine a custom readinto() method which plays with gc.get_referrers, then calls hash(b)...)
This is OK with this patch.sbt commented
on Sep 17, 2012 sbtmannequinMannequinAuthorNew patch which checks the refcount of the memoryview and bytes object after calling readinto().
If either refcount is larger than the expected value of 1, then the data is copied rather than resized.
If either refcount is larger than the expected value of 1, then the
data is copied rather than resized.I think that's a useless precaution. The bytes object cannot "leak" since you are using PyMemoryView_FromMemory(), which doesn't know about the original object.
Out of curiousity, have you done any benchmarks?
sbt commented
on Sep 17, 2012 sbtmannequinMannequinAuthorI think that's a useless precaution. The bytes object cannot "leak"
since you are using PyMemoryView_FromMemory(), which doesn't know about
the original object.The bytes object cannot "leak" so, as you say, checking that refcount is pointless. But the view might "leak", and since it does not own a reference to the base object we have a problem: we can't deallocate the bytes object for fear of breaking the view.
It looks like objects returned by PyMemoryView_FromMemory() must never be allowed to "leak", so I am not sure there are many circumstances in which PyMemoryView_FromMemory() is safe to use.
Perhaps using PyBuffer_FillInfo() and PyMemory_FromBuffer() would keep alive the bytes object while the view is alive, without letting the bytes object "leak".
Out of curiousity, have you done any benchmarks?
No.
The bytes object cannot "leak" so, as you say, checking that refcount
is pointless. But the view might "leak", and since it does not own a
reference to the base object we have a problem: we can't deallocate the
bytes object for fear of breaking the view.Indeed, that's a problem (but your patch does deallocate the bytes object).
It's quite fishy, I'm not sure how to solve the issue cleanly. Stefan, do you have an idea?skrah commented
on Sep 18, 2012 skrahmannequinMannequinSo the problem is that readinto(view) might result in several references
to view? I don't think that can be solved on the memoryview side.One could do:
view = PyMemoryView_FromObject(b); // Lie about writability ((PyMemoryViewObject *)view)->view.readonly = 0;
[...]
Then the view owns a reference to the bytes object. But that does not
solve the problem that writable memoryviews based on a readonly object
might be hanging around.sbt commented
on Sep 18, 2012 sbtmannequinMannequinAuthorThen the view owns a reference to the bytes object. But that does not
solve the problem that writable memoryviews based on a readonly object
might be hanging around.How about doing
PyObject_GetBuffer(b, &buf, PyBUF_WRITABLE); view = PyMemoryView_FromBuffer(&buf); // readinto view PyBuffer_Release(&buf);
Would attempts to access a "leaked" reference to view now result in ValueError("operation forbidden on released memoryview object")? If so then I think this would be safe.
sbt commented
on Sep 18, 2012 sbtmannequinMannequinAuthorThe current non-test uses of PyMemoryView_FromBuffer() are in _io.BufferedReader.read(), _io.BufferedWriter.write(), PyUnicode_Decode().
It looks like they can each be made to leak a memoryview that references a deallocated buffer. (Maybe the answer is Don't Do That.)
import codecs, sys def decode(buf): global view view = buf return codecs.latin_1_decode(buf) def getregentry(): return codecs.CodecInfo(name='foobar', decode=decode, encode=codecs.latin_1_encode) @codecs.register def search_function(encoding): if encoding == 'foobar': return codecs.CodecInfo(*getregentry()) b = b'hello'.upper() b.decode('foobar') print(view.tobytes()) # => b'HELLO' del b x = b'dump'.upper() print(view.tobytes()) # => b'DUMP\x00' import io, sys class File(io.RawIOBase): def readinto(self, buf): global view view = buf n = len(buf) buf[:] = b'x'*n return n def readable(self): return True f = io.BufferedReader(File()) f.read(1) print(view[:5].tobytes()) # => b'xxxxx' del f print(view[:5].tobytes()) # => b'\xdd\xdd\xdd\xdd\xdd'
sbt commented
on Sep 18, 2012 sbtmannequinMannequinAuthorI am rather confused about the ownership semantics when one uses PyMemoryView_FromBuffer().
It looks as though PyMemoryView_FromBuffer() "steals" ownership of the buffer since, when the associated _PyManagedBufferObject is garbage collected, PyBuffer_Release() is called on its copy of the buffer info. However, the _PyManagedBufferObject does not own a reference of the base object, so one still needs to decref the base object (at some time when it is safe to do so).
So am I right in thinking that
PyObject_GetBuffer(obj, &buf, ...); view = PyMemoryView_FromBuffer(&buf); // view->master owns the buffer, but view->master->obj == NULL ... Py_DECREF(view); // releases buffer (assuming no other exports) Py_XDECREF(buf.obj);
has balanced refcounting and is more or less equivalent to
view = PyMemoryView_FromObject(obj); ... Py_DECREF(view);
The documentation is not very helpful. It just says that calls to PyObject_GetBuffer() must be matched with calls to PyBuffer_Release().
skrah commented
on Sep 18, 2012 skrahmannequinMannequinRichard Oudkerk <report@bugs.python.org> wrote:
PyObject_GetBuffer(b, &buf, PyBUF_WRITABLE); view = PyMemoryView_FromBuffer(&buf); // readinto view PyBuffer_Release(&buf);Would attempts to access a "leaked" reference to view now result in ValueError("operation forbidden on released memoryview object")? If so then I think this would be safe.
You would need to call memory_release(). Perhaps we can just expose it on the
C-API level as PyMemoryView_Release().IMO the use of PyObject_GetBuffer() should be discouraged. The semantics
aren't clear (see bpo-15821). I'd suggest using:-
A buffer provider is involved (the case here):
PyMemoryView_From Object()
-
A piece of memory needs to be wrapped temporarily and no references
to the memoryview are "leaked" on the Python level:
PyMemoryView_FromMemory()-
A piece of memory needs to be packaged as a memoryview with automatic
cleanup in mbuf_dealloc():PyMemoryView_FromBufferWithCleanup() (proposed in msg169613)
So I think the combination of PyMemoryView_FromObject() with a call to
PyMemoryView_Release() should indeed work here.-
skrah commented
on Sep 18, 2012 skrahmannequinMannequinRichard Oudkerk <report@bugs.python.org> wrote:
The documentation is not very helpful. It just says that calls
to PyObject_GetBuffer() must be matched with calls to PyBuffer_Release().Yes, we need to sort that out, see bpo-15821.
7 remaining items
BreamoreBoy commented
on Jul 16, 2014 BreamoreBoymannequinMannequinPlease note this is also referred to from bpo-15994.
Another approach here is to make it so
bytearraycan be converted to bytes without a copy. Looking to see if can use_io.BytesIOin this case / performance is reasonable with that which already exposes the "mutable buffer" ->byteswithout copying.- added a commit that references this issue
on Nov 13, 2025 - changed the title
[-]Make rawiobase_read() read directly to bytes object[/-][+]Remove copy from `bytearray` to `bytes` in rawiobase_read() implementation[/+]on Nov 13, 2025 - added a commit that references this issue
on Nov 20, 2025
Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.
Show more details
GitHub fields:
bugs.python.org fields:
Linked PRs
io.RawIOBase.readfrom reading into a temporary bytearray #138616