linux: make ptrace(2) return EIO when trying to peek invalid address

[trasz]

Summary

Previously we've returned the error from native ptrace(2), ENOMEM.
This confused Linux strace(2).

[trasz]

Btw, the ENOMEM return is not documented in the ptrace(2) man page. I wonder what's the better way: updating the man page, or changing the return?

[emaste]

Indeed to me EIO seems a better error for this case than ENOMEM, although I am not sure if anything depends on the current value. @markj @jhb ?

[markj]

In D29925#671482, @emaste wrote:

Indeed to me EIO seems a better error for this case than ENOMEM, although I am not sure if anything depends on the current value. @markj @jhb ?

I'm not sure. At some point PT_READ_I could return either EFAULT or ENOMEM depending on how the fault failed, but I squashed that to just ENOMEM in some refactoring years ago. EFAULT would have made more sense in retrospect.

I would be inclined to leave native ptrace alone and just do the translation here. Perhaps proactively squash EFAULT to EIO as well.

[jhb]

I agree with Mark. I think EFAULT is perhaps the better error code for the native case.

[emaste]

Mmmm, yeah EFAULT seems best.

[trasz]

Either way, it's not EIO, so perhaps let's get this one committed, and then I'll follow up with the native stuff.

[emaste]

Fine with me, but I agree with @markj that making it error == ENOMEM || error == EFAULT already seems like a decent idea.

[trasz]

TBH I'm not a fan of adding code "just in case", except for assertions; to me, it makes reasoning about the code harder.