gh-60107: avoid io.RawIOBase.read from reading into a temporary bytearray

[taskevich]

implements: https://github.com/python/cpython/issues/60107

• Issue: Remove copy from bytearray to bytes in rawiobase_read() implementation #60107

[python-cla-bot]

All commit authors signed the Contributor License Agreement.

[picnixz]

Can you indicate how efficient this actually becomes by using pyperf? there was not clear consensus on the issue considering what should be done.

[picnixz]

While I think this addition is fine, I'm worried that it could change performances a bit. So I would appreciate if we had numbers indicating whether this changed or not the performance (you can look at https://github.com/python/pyperformance for inspiration of how to write benchmarks).

By the way, avoid merging main into your branch unless it's meant to fix CI tests. And remove the NEWS entry for now. We'll add it later if needed. If possible, it would be good if we also had tests where we have a class inheriting from RawIOBase and which tries to mess with the buffer passed to readinto

[auvipy]

should we also consider adding some extra unit tests for this change?

[picnixz]

This was already mentioned:

If possible, it would be good if we also had tests where we have a class inheriting from RawIOBase and which tries to mess with the buffer passed to readinto

[picnixz]

Maybe I wasn't precise enough. RawIOBase.read() is implemented in terms of readinto. So, what we want to do is provide a readinto method that does something bad as it's given the internal buffer we constructed in C. I don't know if it's possible to cause a segfault on main with that approach though.

Stated otherwise, we want some class:

classEvilReadInto(MockRawIOWithoutRead): defreadinto(self, buf): # do something bad with 'buf'

And then

r=EvilReadInto(something_here) r.read() # on main, this call should crash, but not with this patch

[picnixz]

I'm going to convert it into a draft until a regression can be found between main and this PR. IOW, we want to be able to do something evil with the buffer we are given in readinto when using main, but we want that evil thing to be impossible with this patch.

[cmaloney]

Thanks for the pull request @taskevich ! A slightly different fix for this has landed in #141532 so going to close this one.