Passing a memoryview we avoid unneeded copy of the original buffer. On the python side memoryview object can be used for slicing, writing to file, or sending to socket. This may break plugins assuming that the they get a bytearray, but good python code should not care about the type of the buffer, only about the behaviour. Testing with a plugin writing to /dev/null shows 2.7x speedup. Real plugin will probably show much smaller improvement.

Without patch: $ time qemu-img convert -p -f raw -O raw -n /var/tmp/disk.img nbd://localhost/ (100.00/100%) real 0m1.284s user 0m0.091s sys 0m0.576s With patch: $ time qemu-img convert -p -f raw -O raw -n /var/tmp/disk.img nbd://localhost/ (100.00/100%) real 0m0.477s user 0m0.078s sys 0m0.653s --- More info on how I tested this: # Creating test image $ dd if=/dev/zero bs=1M count=1024 | tr "\0" "U" > /var/tmp/disk.img $ cat zero.py import builtins def open(readonly): return builtins.open("/dev/zero", "r+b") def get_size(h): return 1024**3 def pwrite(h, buf, offset): h.write(buf) def pread(h, count, offset): raise NotImplementedError def close(h): h.close() # Running nbdkit $ ./nbdkit -f -v python zero.py plugins/python/python.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/plugins/python/python.c b/plugins/python/python.c index 214fffb..2b4361a 100644 --- a/plugins/python/python.c +++ b/plugins/python/python.c @@ -496,8 +496,8 @@ py_pwrite (void *handle, const void *buf, PyErr_Clear (); r = PyObject_CallFunction (fn, "ONL", obj, - PyByteArray_FromStringAndSize (buf, count), - offset, NULL); + PyMemoryView_FromMemory ((char *)buf, count, PyBUF_READ), + offset, NULL); Py_DECREF (fn); if (check_python_failure ("pwrite") == -1) return -1; -- 2.21.0

Passing a memoryview we avoid unneeded copy of the original buffer. On the python side memoryview object can be used for slicing, writing to file, or sending to socket. This may break plugins assuming that the they get a bytearray, but good python code should not care about the type of the buffer, only about the behaviour. Testing with a plugin writing to /dev/null shows 2.7x speedup. Real plugin will probably show much smaller improvement. Without patch: $ time qemu-img convert -p -f raw -O raw -n /var/tmp/disk.img nbd://localhost/ (100.00/100%) real 0m1.284s user 0m0.091s sys 0m0.576s With patch: $ time qemu-img convert -p -f raw -O raw -n /var/tmp/disk.img nbd://localhost/ (100.00/100%) real 0m0.477s user 0m0.078s sys 0m0.653s --- More info on how I tested this: # Creating test image $ dd if=/dev/zero bs=1M count=1024 | tr "\0" "U" > /var/tmp/disk.img $ cat zero.py import builtins def open(readonly):     return builtins.open("/dev/zero", "r+b") def get_size(h):     return 1024**3 def pwrite(h, buf, offset):     h.write(buf) def pread(h, count, offset):     raise NotImplementedError def close(h):     h.close() # Running nbdkit $ ./nbdkit -f -v python zero.py plugins/python/python.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/plugins/python/python.c b/plugins/python/python.c index 214fffb..2b4361a 100644 --- a/plugins/python/python.c +++ b/plugins/python/python.c @@ -496,8 +496,8 @@ py_pwrite (void *handle, const void *buf, PyErr_Clear (); r = PyObject_CallFunction (fn, "ONL", obj, - PyByteArray_FromStringAndSize (buf, count), - offset, NULL); + PyMemoryView_FromMemory ((char *)buf, count, PyBUF_READ), + offset, NULL); Py_DECREF (fn); if (check_python_failure ("pwrite") == -1) return -1;