Inconsistent behavior in dealing with partial data

[Leont]

It seems like the implementations on Linux have different behaviors regarding a partial read.

The read implementation will implicitly retry when interrupted until the desired length is reached.

The sysread implementations on the other hand will not retry, so if it would be interrupted while only partially having filled it will throw an error, even though the situation is entirely recoverable by issuing a new read.

AFAICT the XS implementation can exhibit either behavior depending on which backend is selected.

[david-dick]

Would you think a retry on receiving a EINTR when reading /dev/urandom or calling getrandom to be acceptable?