diff --git a/src/rdrand.rs b/src/rdrand.rs
index 910b5350..69f6a5d1 100644
--- a/src/rdrand.rs
+++ b/src/rdrand.rs
@@ -5,10 +5,11 @@
 // <LICENSE-MIT or https://opensource.org/licenses/MIT>, at your
 // option. This file may not be copied, modified, or distributed
 // except according to those terms.
-
-//! Implementation for SGX using RDRAND instruction
-use crate::{util::slice_as_uninit, Error};
-use core::mem::{self, MaybeUninit};
+use crate::{
+    util::{slice_as_uninit, LazyBool},
+    Error,
+};
+use core::mem::{size_of, MaybeUninit};
 
 cfg_if! {
     if #[cfg(target_arch = "x86_64")] {
@@ -24,74 +25,106 @@ cfg_if! {
 // Implementation Guide" - Section 5.2.1 and "Intel® 64 and IA-32 Architectures
 // Software Developer’s Manual" - Volume 1 - Section 7.3.17.1.
 const RETRY_LIMIT: usize = 10;
-const WORD_SIZE: usize = mem::size_of::<usize>();
 
 #[target_feature(enable = "rdrand")]
-unsafe fn rdrand() -> Result<[u8; WORD_SIZE], Error> {
+unsafe fn rdrand() -> Option<usize> {
     for _ in 0..RETRY_LIMIT {
-        let mut el = mem::zeroed();
-        if rdrand_step(&mut el) == 1 {
-            // AMD CPUs from families 14h to 16h (pre Ryzen) sometimes fail to
-            // set CF on bogus random data, so we check these values explicitly.
-            // See https://github.com/systemd/systemd/issues/11810#issuecomment-489727505
-            // We perform this check regardless of target to guard against
-            // any implementation that incorrectly fails to set CF.
-            if el != 0 && el != !0 {
-                return Ok(el.to_ne_bytes());
-            }
-            // Keep looping in case this was a false positive.
+        let mut val = 0;
+        if rdrand_step(&mut val) == 1 {
+            return Some(val as usize);
         }
     }
-    Err(Error::FAILED_RDRAND)
+    None
 }
 
-// "rdrand" target feature requires "+rdrnd" flag, see https://github.com/rust-lang/rust/issues/49653.
+// "rdrand" target feature requires "+rdrand" flag, see https://github.com/rust-lang/rust/issues/49653.
 #[cfg(all(target_env = "sgx", not(target_feature = "rdrand")))]
 compile_error!(
-    "SGX targets require 'rdrand' target feature. Enable by using -C target-feature=+rdrnd."
+    "SGX targets require 'rdrand' target feature. Enable by using -C target-feature=+rdrand."
 );
 
-#[cfg(target_feature = "rdrand")]
-fn is_rdrand_supported() -> bool {
-    true
+// Run a small self-test to make sure we aren't repeating values
+// Adapted from Linux's test in arch/x86/kernel/cpu/rdrand.c
+// Fails with probability < 2^(-90) on 32-bit systems
+#[target_feature(enable = "rdrand")]
+unsafe fn self_test() -> bool {
+    // On AMD, RDRAND returns 0xFF...FF on failure, count it as a collision.
+    let mut prev = !0; // TODO(MSRV 1.43): Move to usize::MAX
+    let mut fails = 0;
+    for _ in 0..8 {
+        match rdrand() {
+            Some(val) if val == prev => fails += 1,
+            Some(val) => prev = val,
+            None => return false,
+        };
+    }
+    fails <= 2
 }
 
-// TODO use is_x86_feature_detected!("rdrand") when that works in core. See:
-// https://github.com/rust-lang-nursery/stdsimd/issues/464
-#[cfg(not(target_feature = "rdrand"))]
-fn is_rdrand_supported() -> bool {
-    use crate::util::LazyBool;
+fn is_rdrand_good() -> bool {
+    #[cfg(not(target_feature = "rdrand"))]
+    {
+        // SAFETY: All Rust x86 targets are new enough to have CPUID, and we
+        // check that leaf 1 is supported before using it.
+        let cpuid0 = unsafe { arch::__cpuid(0) };
+        if cpuid0.eax < 1 {
+            return false;
+        }
+        let cpuid1 = unsafe { arch::__cpuid(1) };
 
-    // SAFETY: All Rust x86 targets are new enough to have CPUID, and if CPUID
-    // is supported, CPUID leaf 1 is always supported.
-    const FLAG: u32 = 1 << 30;
-    static HAS_RDRAND: LazyBool = LazyBool::new();
-    HAS_RDRAND.unsync_init(|| unsafe { (arch::__cpuid(1).ecx & FLAG) != 0 })
+        let vendor_id = [
+            cpuid0.ebx.to_le_bytes(),
+            cpuid0.edx.to_le_bytes(),
+            cpuid0.ecx.to_le_bytes(),
+        ];
+        if vendor_id == [*b"Auth", *b"enti", *b"cAMD"] {
+            let mut family = (cpuid1.eax >> 8) & 0xF;
+            if family == 0xF {
+                family += (cpuid1.eax >> 20) & 0xFF;
+            }
+            // AMD CPUs families before 17h (Zen) sometimes fail to set CF when
+            // RDRAND fails after suspend. Don't use RDRAND on those families.
+            // See https://bugzilla.redhat.com/show_bug.cgi?id=1150286
+            if family < 0x17 {
+                return false;
+            }
+        }
+
+        const RDRAND_FLAG: u32 = 1 << 30;
+        if cpuid1.ecx & RDRAND_FLAG == 0 {
+            return false;
+        }
+    }
+
+    // SAFETY: We have already checked that rdrand is available.
+    unsafe { self_test() }
 }
 
 pub fn getrandom_inner(dest: &mut [MaybeUninit<u8>]) -> Result<(), Error> {
-    if !is_rdrand_supported() {
+    static RDRAND_GOOD: LazyBool = LazyBool::new();
+    if !RDRAND_GOOD.unsync_init(is_rdrand_good) {
         return Err(Error::NO_RDRAND);
     }
-
-    // SAFETY: After this point, rdrand is supported, so calling the rdrand
-    // functions is not undefined behavior.
-    unsafe { rdrand_exact(dest) }
+    // SAFETY: After this point, we know rdrand is supported.
+    unsafe { rdrand_exact(dest) }.ok_or(Error::FAILED_RDRAND)
 }
 
+// TODO: make this function safe when we have feature(target_feature_11)
 #[target_feature(enable = "rdrand")]
-unsafe fn rdrand_exact(dest: &mut [MaybeUninit<u8>]) -> Result<(), Error> {
+unsafe fn rdrand_exact(dest: &mut [MaybeUninit<u8>]) -> Option<()> {
     // We use chunks_exact_mut instead of chunks_mut as it allows almost all
     // calls to memcpy to be elided by the compiler.
-    let mut chunks = dest.chunks_exact_mut(WORD_SIZE);
+    let mut chunks = dest.chunks_exact_mut(size_of::<usize>());
     for chunk in chunks.by_ref() {
-        chunk.copy_from_slice(slice_as_uninit(&rdrand()?));
+        let src = rdrand()?.to_ne_bytes();
+        chunk.copy_from_slice(slice_as_uninit(&src));
     }
 
     let tail = chunks.into_remainder();
     let n = tail.len();
     if n > 0 {
-        tail.copy_from_slice(slice_as_uninit(&rdrand()?[..n]));
+        let src = rdrand()?.to_ne_bytes();
+        tail.copy_from_slice(slice_as_uninit(&src[..n]));
     }
-    Ok(())
+    Some(())
 }