Abstract

This PEP proposes an extension to the buffer interface called the ‘locked buffer interface’.

The locked buffer interface avoids the flaws of the ‘old’ buffer interface [1] as defined in Python versions up to and including 2.2, and has the following semantics:

• The lifetime of the retrieved pointer is clearly defined and controlled by the client.
• The buffer size is returned as a ‘size_t’ data type, which allows access to large buffers on platforms where sizeof(int) != sizeof(void *).

(Guido comments: This second sounds like a change we could also make to the “old” buffer interface, if we introduce another flag bit that’s not part of the default flags.)

Specification

The locked buffer interface exposes new functions which return the size and the pointer to the internal memory block of any python object which chooses to implement this interface.

Retrieving a buffer from an object puts this object in a locked state during which the buffer may not be freed, resized, or reallocated.

The object must be unlocked again by releasing the buffer if it’s no longer used by calling another function in the locked buffer interface. If the object never resizes or reallocates the buffer during its lifetime, this function may be NULL. Failure to call this function (if it is != NULL) is a programming error and may have unexpected results.

The locked buffer interface omits the memory segment model which is present in the old buffer interface - only a single memory block can be exposed.

The memory blocks can be accessed without holding the global interpreter lock.

Implementation

Define a new flag in Include/object.h:

/* PyBufferProcs contains bf_acquirelockedreadbuffer,
   bf_acquirelockedwritebuffer, and bf_releaselockedbuffer */
#define Py_TPFLAGS_HAVE_LOCKEDBUFFER (1L<<15)

This flag would be included in Py_TPFLAGS_DEFAULT:

#define Py_TPFLAGS_DEFAULT  ( \
                    ....
                    Py_TPFLAGS_HAVE_LOCKEDBUFFER | \
                    ....
                    0)

Extend the PyBufferProcs structure by new fields in Include/object.h:

typedef size_t (*acquirelockedreadbufferproc)(PyObject *,
                                              const void **);
typedef size_t (*acquirelockedwritebufferproc)(PyObject *,
                                               void **);
typedef void (*releaselockedbufferproc)(PyObject *);

typedef struct {
    getreadbufferproc bf_getreadbuffer;
    getwritebufferproc bf_getwritebuffer;
    getsegcountproc bf_getsegcount;
    getcharbufferproc bf_getcharbuffer;
    /* locked buffer interface functions */
    acquirelockedreadbufferproc bf_acquirelockedreadbuffer;
    acquirelockedwritebufferproc bf_acquirelockedwritebuffer;
    releaselockedbufferproc bf_releaselockedbuffer;
} PyBufferProcs;

The new fields are present if the Py_TPFLAGS_HAVE_LOCKEDBUFFER flag is set in the object’s type.

The Py_TPFLAGS_HAVE_LOCKEDBUFFER flag implies the Py_TPFLAGS_HAVE_GETCHARBUFFER flag.

The acquirelockedreadbufferproc and acquirelockedwritebufferproc functions return the size in bytes of the memory block on success, and fill in the passed void * pointer on success. If these functions fail - either because an error occurs or no memory block is exposed - they must set the void * pointer to NULL and raise an exception. The return value is undefined in these cases and should not be used.

If calls to these functions succeed, eventually the buffer must be released by a call to the releaselockedbufferproc, supplying the original object as argument. The releaselockedbufferproc cannot fail. For objects that actually maintain an internal lock count it would be a fatal error if the releaselockedbufferproc function would be called too often, leading to a negative lock count.

Similar to the ‘old’ buffer interface, any of these functions may be set to NULL, but it is strongly recommended to implement the releaselockedbufferproc function (even if it does nothing) if any of the acquireread/writelockedbufferproc functions are implemented, to discourage extension writers from checking for a NULL value and not calling it.

These functions aren’t supposed to be called directly, they are called through convenience functions declared in Include/abstract.h:

int PyObject_AcquireLockedReadBuffer(PyObject *obj,
                                    const void **buffer,
                                    size_t *buffer_len);

int PyObject_AcquireLockedWriteBuffer(PyObject *obj,
                                      void **buffer,
                                      size_t *buffer_len);

void PyObject_ReleaseLockedBuffer(PyObject *obj);

The former two functions return 0 on success, set buffer to the memory location and buffer_len to the length of the memory block in bytes. On failure, or if the locked buffer interface is not implemented by obj, they return -1 and set an exception.

The latter function doesn’t return anything, and cannot fail.

Backward Compatibility

The size of the PyBufferProcs structure changes if this proposal is implemented, but the type’s tp_flags slot can be used to determine if the additional fields are present.

Reference Implementation

An implementation has been uploaded to the SourceForge patch manager as https://bugs.python.org/issue652857.

Additional Notes/Comments

Python strings, unicode strings, mmap objects, and array objects would expose the locked buffer interface.

mmap and array objects would actually enter a locked state while the buffer is active, this is not needed for strings and unicode objects. Resizing locked array objects is not allowed and will raise an exception. Whether closing a locked mmap object is an error or will only be deferred until the lock count reaches zero is an implementation detail.

Guido recommends

But I’m still very concerned that if most built-in types (e.g. strings, bytes) don’t implement the release functionality, it’s too easy for an extension to seem to work while forgetting to release the buffer.

I recommend that at least some built-in types implement the acquire/release functionality with a counter, and assert that the counter is zero when the object is deleted – if the assert fails, someone DECREF’ed their reference to the object without releasing it. (The rule should be that you must own a reference to the object while you’ve acquired the object.)

For strings that might be impractical because the string object would have to grow 4 bytes to hold the counter; but the new bytes object (PEP 296) could easily implement the counter, and the array object too – that way there will be plenty of opportunity to test proper use of the protocol.

Community Feedback

Greg Ewing doubts the locked buffer interface is needed at all, he thinks the normal buffer interface could be used if the pointer is (re)fetched each time it’s used. This seems to be dangerous, because even innocent looking calls to the Python API like Py_DECREF() may trigger execution of arbitrary Python code.

The first version of this proposal didn’t have the release function, but it turned out that this would have been too restrictive: mmap and array objects wouldn’t have been able to implement it, because mmap objects can be closed anytime if not locked, and array objects could resize or reallocate the buffer.

This PEP will probably be rejected because nobody except the author needs it.

References

Copyright

This document has been placed in the public domain.