So, as pointed out by haypo, blksize_t is actually signed long on Linux. This means that my patch (as well as the current code) is not right.
Both with and without my patch, io_open function uses "int" to store blksize_t and it also passes it to one of PyBuffered{Random,Reader,Writer}_Type. These three however use Py_ssize_t, which is inconsistent with io_open, and it's not correct, too.
I'm unsure how to proceed here - should I fix buffer size types throughout the _io module to long and submit one big patch? It doesn't feel right to put two not-very-related changes into one patch (I'm afraid that changing buffer sizes to long everywhere will result in a rather large patch). Or should I first submit a patch that fixes the long usage and then rebase the patch attached to this issue on top of it?

Thanks a lot.