[foreign-abi] RFR: 8240173: Confusing overflow error when trying to dereference a nothing segment

[Maurizio Cimadamore]

On Thu, 27 Feb 2020 15:26:51 GMT, Henry Jen <henryjen at openjdk.org> wrote:

>> This simple patch attempts to generate a more explicit error message when trying to dereference an address which is based on the Nothing segment. Note that the first problem here is caused by the fact that the Nothing segment is considered "small" - which means most of the real address values will fall outside its range. While we could simply fix that (which will improve the error message - from reporting an overflow during offset computation to report an access outside the bounds of the segment), I thought it was better to mark the Nothing segment as non-accessible, and generate an explicit error message.
>
> Looks good. I think making NO_ACCESS explicit rather than depending on size checking is the right thing and glad to see the unused mask argument gone before we actually need it.

> This started to make me wonder whether this ?nothing? segment is really an ?untrusted? segment, which could even have a bound. It does represent something, but that something cannot be fully trusted and as such its contents cannot from be accessed within Java code.

There is an overlapping between Untrusted and Nothing. In the past we considered having distinction between trusted (or managed) and not trusted. Having the notion of the Nothing segment allowed us to speak about non-trusted addresses in a way which didn't need any other special concepts (and to reuse the 'rebase' operation to trust an untrusted address). In other words, there is a certain appeal in not adding a new kind of segment (untrusted) and simply piggy back on the concepts we already have. At the time, the only way to make a segment inaccessible was to give it zero length - but I'm planning a revamp of the access modes, to expose, instead of a single asReadOnly method a set of orthogonal access modes READ | WRITE | CLOSE - at which an untrusted segment will simply be a segment whose READ | WRITE (and maybe CLOSE) bits are unset.

-------------

PR: https://git.openjdk.java.net/panama-foreign/pull/34