[libcxx] random_device, for OpenBSD specify optimal entropy properties

[brad]

Summary

Submitting upstream from OpenBSD tree:

The random_device is an abstraction to the underlying OS PRNG. In OpenBSD's case, we know it has optimal entropy properties so specify what that is.

[ldionne]

libcxx/src/random.cpp

37

You can include this unconditionally. Please include it alongside "random" and "system_error" above.

179

Couldn't we do a better job and use ioctl on Linuxes? See cppreference: https://en.cppreference.com/w/cpp/numeric/random/random_device/entropy

[brad]

libcxx/src/random.cpp

37

You can include this unconditionally. Please include it alongside "random" and "system_error" above.

Ah. The original diff was like that. I figured you would want it guarded like this.

Immediately below them? In the same form? "limits" or as is?

[ldionne]

libcxx/src/random.cpp

37

Just #include "limits". It's a header provided by libc++, and compile-times are not really a concern for building the libc++ library, so I prefer to go for simplicity and removing as many #ifdefs as I can.

[brad]

Updated diff after moving header up to the top and removing the OpenBSD guard.

[brad]

Don't care about the removal of that one line.

[ldionne]

libcxx/src/random.cpp

179

Ping on this question. Is it a possibility to improve our implementation on all Linux platforms instead?

[ldionne]

@brad I suggest we abandon this in favour of D94953. Does it solve your problem?

[brad]

@ldionne I don't understand how a Linux specific diff would help with another operating system. I could update this after D94953 goes in and eliminate the first chunk.

[ldionne]

In D94571#2506880, @brad wrote:

@ldionne I don't understand how a Linux specific diff would help with another operating system. I could update this after D94953 goes in and eliminate the first chunk.

I thought OpenBSD and Linux might behave the same in that regard. It does look like OpenBSD has ioctl, but not RNDGETENTCNT: https://man.openbsd.org/ioctl.2. Can you confirm this?

[brad]

Rebased.

[curdeius]

libcxx/src/random.cpp

192

Just asking. I guess it cannot happen that the first branch of #if is true in OpenBSD, right?

193

Please use result_type here or changed to unsigned above for consistency.

[brad]

libcxx/src/random.cpp

192

Yes, this RNDGETENTCNT ioctl is Linux specific.

193

Ok.

[brad]

Updated.

libcxx/src/random.cpp

 }
 
 #else
 #error "Random device not implemented for this architecture"
 #endif
 
 double
 random_device::entropy() const _NOEXCEPT
 {
 #if defined(_LIBCPP_USING_DEV_RANDOM) && defined(RNDGETENTCNT)
   int ent;
   if (::ioctl(__f_, RNDGETENTCNT, &ent) < 0)
     return 0;
 
   if (ent < 0)
     return 0;
 
   if (ent > std::numeric_limits<result_type>::digits)
     return std::numeric_limits<result_type>::digits;
 
   return ent;
+#elif defined(__OpenBSD__)
+  return std::numeric_limits<unsigned int>::digits;
 #else
   return 0;
 #endif
 }
 
 _LIBCPP_END_NAMESPACE_STD