Use RdSeed when available, and reduce RdRand load

This introduces support for autodetecting and using the RdSeed instruction.

In addition:
* In SeedFast, only 64 bits of entropy are generated through RdRand (256 was relatively slow).
* In SeedStartup, 256 bits of entropy are generated, using RdSeed (preferably) or RdRand (otherwise).

src/random.cpp

@@ -78,25 +78,119 @@ static inline int64_t GetPerformanceCounter() noexcept
 }
 
 #if defined(__x86_64__) || defined(__amd64__) || defined(__i386__)
-static bool rdrand_supported = false;
+static bool g_rdrand_supported = false;
+static bool g_rdseed_supported = false;
 static constexpr uint32_t CPUID_F1_ECX_RDRAND = 0x40000000;
+static constexpr uint32_t CPUID_F7_EBX_RDSEED = 0x00040000;
+#ifdef bit_RDRND
+static_assert(CPUID_F1_ECX_RDRAND == bit_RDRND, "Unexpected value for bit_RDRND");
+#endif
+#ifdef bit_RDSEED
+static_assert(CPUID_F7_EBX_RDSEED == bit_RDSEED, "Unexpected value for bit_RDSEED");
+#endif
+static void inline GetCPUID(uint32_t leaf, uint32_t subleaf, uint32_t& a, uint32_t& b, uint32_t& c, uint32_t& d)
+{
+    // We can't use __get_cpuid as it doesn't support subleafs.
+#ifdef __GNUC__
+    __cpuid_count(leaf, subleaf, a, b, c, d);
+#else
+    __asm__ ("cpuid" : "=a"(a), "=b"(b), "=c"(c), "=d"(d) : "0"(leaf), "2"(subleaf));
+#endif
+}
+
 static void InitHardwareRand()
 {
     uint32_t eax, ebx, ecx, edx;
-    if (__get_cpuid(1, &eax, &ebx, &ecx, &edx) && (ecx & CPUID_F1_ECX_RDRAND)) {
+    GetCPUID(1, 0, eax, ebx, ecx, edx);
-        rdrand_supported = true;
+    if (ecx & CPUID_F1_ECX_RDRAND) {
+        g_rdrand_supported = true;
+    }
+    GetCPUID(7, 0, eax, ebx, ecx, edx);
+    if (ebx & CPUID_F7_EBX_RDSEED) {
+        g_rdseed_supported = true;
     }
 }
 
 static void ReportHardwareRand()
 {
-    if (rdrand_supported) {
     // This must be done in a separate function, as HWRandInit() may be indirectly called
     // from global constructors, before logging is initialized.
+    if (g_rdseed_supported) {
+        LogPrintf("Using RdSeed as additional entropy source\n");
+    }
+    if (g_rdrand_supported) {
         LogPrintf("Using RdRand as an additional entropy source\n");
     }
 }
 
+/** Read 64 bits of entropy using rdrand.
+ *
+ * Must only be called when RdRand is supported.
+ */
+static uint64_t GetRdRand() noexcept
+{
+    // RdRand may very rarely fail. Invoke it up to 10 times in a loop to reduce this risk.
+#ifdef __i386__
+    uint8_t ok;
+    uint32_t r1, r2;
+    for (int i = 0; i < 10; ++i) {
+        __asm__ volatile (".byte 0x0f, 0xc7, 0xf0; setc %1" : "=a"(r1), "=q"(ok) :: "cc"); // rdrand %eax
+        if (ok) break;
+    }
+    for (int i = 0; i < 10; ++i) {
+        __asm__ volatile (".byte 0x0f, 0xc7, 0xf0; setc %1" : "=a"(r2), "=q"(ok) :: "cc"); // rdrand %eax
+        if (ok) break;
+    }
+    return (((uint64_t)r2) << 32) | r1;
+#elif defined(__x86_64__) || defined(__amd64__)
+    uint8_t ok;
+    uint64_t r1;
+    for (int i = 0; i < 10; ++i) {
+        __asm__ volatile (".byte 0x48, 0x0f, 0xc7, 0xf0; setc %1" : "=a"(r1), "=q"(ok) :: "cc"); // rdrand %rax
+        if (ok) break;
+    }
+    return r1;
+#else
+#error "RdRand is only supported on x86 and x86_64"
+#endif
+}
+
+/** Read 64 bits of entropy using rdseed.
+ *
+ * Must only be called when RdSeed is supported.
+ */
+static uint64_t GetRdSeed() noexcept
+{
+    // RdSeed may fail when the HW RNG is overloaded. Loop indefinitely until enough entropy is gathered,
+    // but pause after every failure.
+#ifdef __i386__
+    uint8_t ok;
+    uint32_t r1, r2;
+    do {
+        __asm__ volatile (".byte 0x0f, 0xc7, 0xf8; setc %1" : "=a"(r1), "=q"(ok) :: "cc"); // rdseed %eax
+        if (ok) break;
+        __asm__ volatile ("pause");
+    } while(true);
+    do {
+        __asm__ volatile (".byte 0x0f, 0xc7, 0xf8; setc %1" : "=a"(r2), "=q"(ok) :: "cc"); // rdseed %eax
+        if (ok) break;
+        __asm__ volatile ("pause");
+    } while(true);
+    return (((uint64_t)r2) << 32) | r1;
+#elif defined(__x86_64__) || defined(__amd64__)
+    uint8_t ok;
+    uint64_t r1;
+    do {
+        __asm__ volatile (".byte 0x48, 0x0f, 0xc7, 0xf8; setc %1" : "=a"(r1), "=q"(ok) :: "cc"); // rdseed %rax
+        if (ok) break;
+        __asm__ volatile ("pause");
+    } while(true);
+    return r1;
+#else
+#error "RdSeed is only supported on x86 and x86_64"
+#endif
+}
+
 #else
 /* Access to other hardware random number generators could be added here later,
  * assuming it is sufficiently fast (in the order of a few hundred CPU cycles).
@@ -107,40 +201,40 @@ static void InitHardwareRand() {}
 static void ReportHardwareRand() {}
 #endif
 
-static bool GetHardwareRand(unsigned char* ent32) noexcept {
+/** Add 64 bits of entropy gathered from hardware to hasher. Do nothing if not supported. */
+static void SeedHardwareFast(CSHA512& hasher) noexcept {
 #if defined(__x86_64__) || defined(__amd64__) || defined(__i386__)
-    if (rdrand_supported) {
+    if (g_rdrand_supported) {
-        uint8_t ok;
+        uint64_t out = GetRdRand();
-        // Not all assemblers support the rdrand instruction, write it in hex.
+        hasher.Write((const unsigned char*)&out, sizeof(out));
-#ifdef __i386__
+        return;
-        for (int iter = 0; iter < 4; ++iter) {
+    }
-            uint32_t r1, r2;
+#endif
-            __asm__ volatile (".byte 0x0f, 0xc7, 0xf0;" // rdrand %eax
+}
-                              ".byte 0x0f, 0xc7, 0xf2;" // rdrand %edx
+
-                              "setc %2" :
+/** Add 256 bits of entropy gathered from hardware to hasher. Do nothing if not supported. */
-                              "=a"(r1), "=d"(r2), "=q"(ok) :: "cc");
+static void SeedHardwareSlow(CSHA512& hasher) noexcept {
-            if (!ok) return false;
+#if defined(__x86_64__) || defined(__amd64__) || defined(__i386__)
-            WriteLE32(ent32 + 8 * iter, r1);
+    // When we want 256 bits of entropy, prefer RdSeed over RdRand, as it's
-            WriteLE32(ent32 + 8 * iter + 4, r2);
+    // guaranteed to produce independent randomness on every call.
-        }
+    if (g_rdseed_supported) {
-#else
+        for (int i = 0; i < 4; ++i) {
-        uint64_t r1, r2, r3, r4;
+            uint64_t out = GetRdSeed();
-        __asm__ volatile (".byte 0x48, 0x0f, 0xc7, 0xf0, " // rdrand %rax
+            hasher.Write((const unsigned char*)&out, sizeof(out));
-                                "0x48, 0x0f, 0xc7, 0xf3, " // rdrand %rbx
+        }
-                                "0x48, 0x0f, 0xc7, 0xf1, " // rdrand %rcx
+        return;
-                                "0x48, 0x0f, 0xc7, 0xf2; " // rdrand %rdx
+    }
-                          "setc %4" :
+    // When falling back to RdRand, XOR the result of 1024 results.
-                          "=a"(r1), "=b"(r2), "=c"(r3), "=d"(r4), "=q"(ok) :: "cc");
+    // This guarantees a reseeding occurs between each.
-        if (!ok) return false;
+    if (g_rdrand_supported) {
-        WriteLE64(ent32, r1);
+        for (int i = 0; i < 4; ++i) {
-        WriteLE64(ent32 + 8, r2);
+            uint64_t out = 0;
-        WriteLE64(ent32 + 16, r3);
+            for (int j = 0; j < 1024; ++j) out ^= GetRdRand();
-        WriteLE64(ent32 + 24, r4);
+            hasher.Write((const unsigned char*)&out, sizeof(out));
-#endif
+        }
-        return true;
+        return;
     }
 #endif
-    return false;
 }
 
 static void RandAddSeedPerfmon(CSHA512& hasher)
@@ -407,8 +501,7 @@ static void SeedFast(CSHA512& hasher) noexcept
     hasher.Write((const unsigned char*)&ptr, sizeof(ptr));
 
     // Hardware randomness is very fast when available; use it always.
-    bool have_hw_rand = GetHardwareRand(buffer);
+    SeedHardwareFast(hasher);
-    if (have_hw_rand) hasher.Write(buffer, sizeof(buffer));
 
     // High-precision timestamp
     SeedTimestamp(hasher);
@@ -460,6 +553,9 @@ static void SeedStartup(CSHA512& hasher) noexcept
     RAND_screen();
 #endif
 
+    // Gather 256 bits of hardware randomness, if available
+    SeedHardwareSlow(hasher);
+
     // Everything that the 'slow' seeder includes.
     SeedSlow(hasher);
 

src/random.h

@@ -24,7 +24,7 @@
  *   perform 'fast' seeding, consisting of mixing in:
  *   - A stack pointer (indirectly committing to calling thread and call stack)
  *   - A high-precision timestamp (rdtsc when available, c++ high_resolution_clock otherwise)
- *   - Hardware RNG (rdrand) when available.
+ *   - 64 bits from the hardware RNG (rdrand) when available.
  *   These entropy sources are very fast, and only designed to protect against situations
  *   where a VM state restore/copy results in multiple systems with the same randomness.
  *   FastRandomContext on the other hand does not protect against this once created, but
@@ -48,6 +48,7 @@
  *
  * On first use of the RNG (regardless of what function is called first), all entropy
  * sources used in the 'slow' seeder are included, but also:
+ * - 256 bits from the hardware RNG (rdseed or rdrand) when available.
  * - (On Windows) Performance monitoring data from the OS.
  * - (On Windows) Through OpenSSL, the screen contents.
  *