mmap read is racy

[totaam]

It probably does not cause real problems because the mmap area is big enough that we don't overlap reads and writes.
09408c4 changed from array.from_buffer to memoryview but the result is the same: we have been using zero-copy since at least 2013: 2c3f04a

The fast-path using single-chunk encoding is preferred by the server side.
The client wraps this memory in a memoryview.

The problem is that the shared data offsets are updated synchronously:

xpra/xpra/net/mmap.py

Line 280 in 38fce45

data_start.value = offset + length

Instead, we should wait until the memoryview is freed before allowing writes to this area.
The lazy fix would be to call tobytes().
A better fix would be to obtain the memory pointer and use makebuf with a custom free function.