On Thu, Jun 09, 2022 at 03:34:19PM +0100, Richard W.M. Jones wrote:
On Thu, Jun 09, 2022 at 08:34:47AM -0500, Eric Blake wrote:
The Py_BuildValue "y#" format copies data; this is because Python assumes our C memory can go out of scope, while the user's python callback can stash off whatever bytearray object it got. But this copying is inefficient. Now that we already have a Python buffer-like object in scope for the duration of all structured reads (pread_structured since commit f8c76c01, aio_pread_structured since commit 4f15d0e9), we know that the C memory parameter to our chunk_callback is already (a portion of) the underlying contiguous memory of an original Python object. If we pass the Python callback a memoryview object over the correct slice of that memory, then Python will correctly handle the intricacies of what happens if the callback saves our memoryview object for later use, all without any data copying overhead, making use of the pread_structured form much faster.
Acked-by: Richard W.M. Jones <rjones@redhat.com>
Series is now pushed as 67dd054..8bb89d6 -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3266 Virtualization: qemu.org | libvirt.org