Fix RocksDB SIGILL error on Raspberry PI 4 (#7233)

Summary:
Issue：https://github.com/facebook/rocksdb/issues/7042

No PMULL runtime check will lead to SIGILL on a Raspberry pi 4.

Leverage 'getauxval' to get Hardware-Cap to detect whether target
platform does support PMULL or not in runtime.

Consider the condition that the target platform does support crc32 but not support PMULL.
In this condition, the code should leverage the crc32 instruction
rather than skip all hardware crc32 instruction.

Pull Request resolved: https://github.com/facebook/rocksdb/pull/7233

Reviewed By: jay-zhuang

Differential Revision: D23790116

fbshipit-source-id: a3ebd821fbd4a38dd2f59064adbb7c3013ee8140

util/crc32c.cc

@@ -41,6 +41,10 @@
 
 #endif
 
+#if defined(__linux__) && defined(HAVE_ARM64_CRC)
+bool pmull_runtime_flag = false;
+#endif
+
 namespace ROCKSDB_NAMESPACE {
 namespace crc32c {
 
@@ -494,6 +498,7 @@ std::string IsFastCrc32Supported() {
   if (crc32c_runtime_check()) {
     has_fast_crc = true;
     arch = "Arm64";
+    pmull_runtime_flag = crc32c_pmull_runtime_check();
   } else {
     has_fast_crc = false;
     arch = "Arm64";
@@ -1224,6 +1229,7 @@ static inline Function Choose_Extend() {
   return isAltiVec() ? ExtendPPCImpl : ExtendImpl<Slow_CRC32>;
 #elif defined(__linux__) && defined(HAVE_ARM64_CRC)
   if(crc32c_runtime_check()) {
+    pmull_runtime_flag = crc32c_pmull_runtime_check();
     return ExtendARMImpl;
   } else {
     return ExtendImpl<Slow_CRC32>;

util/crc32c_arm64.cc

@@ -14,6 +14,9 @@
 #ifndef HWCAP_CRC32
 #define HWCAP_CRC32 (1 << 7)
 #endif
+#ifndef HWCAP_PMULL
+#define HWCAP_PMULL (1 << 4)
+#endif
 
 #ifdef HAVE_ARM64_CRYPTO
 /* unfolding to compute 8 * 3 = 24 bytes parallelly */
@@ -35,6 +38,8 @@
   } while (0)
 #endif
 
+extern bool pmull_runtime_flag;
+
 uint32_t crc32c_runtime_check(void) {
 #ifdef ROCKSDB_AUXV_GETAUXVAL_PRESENT
   uint64_t auxv = getauxval(AT_HWCAP);
@@ -44,6 +49,15 @@ uint32_t crc32c_runtime_check(void) {
 #endif
 }
 
+bool crc32c_pmull_runtime_check(void) {
+#ifdef ROCKSDB_AUXV_GETAUXVAL_PRESENT
+  uint64_t auxv = getauxval(AT_HWCAP);
+  return (auxv & HWCAP_PMULL) != 0;
+#else
+  return false;
+#endif
+}
+
 #ifdef ROCKSDB_UBSAN_RUN
 #if defined(__clang__)
 __attribute__((__no_sanitize__("alignment")))
@@ -58,6 +72,13 @@ uint32_t crc32c_arm64(uint32_t crc, unsigned char const *data,
   int length = (int)len;
   crc ^= 0xffffffff;
 
+  /*
+   * Pmull runtime check here.
+   * Raspberry Pi supports crc32 but doesn't support pmull.
+   * Skip Crc32c Parallel computation if no crypto extension available.
+   */
+  if (pmull_runtime_flag) {
+/* Macro (HAVE_ARM64_CRYPTO) is used for compiling check  */
 #ifdef HAVE_ARM64_CRYPTO
 /* Crc32c Parallel computation
  *   Algorithm comes from Intel whitepaper:
@@ -68,51 +89,53 @@ uint32_t crc32c_arm64(uint32_t crc, unsigned char const *data,
  *   One Block: 42(BLK_LENGTH) * 8(step length: crc32c_u64) bytes
  */
 #define BLK_LENGTH 42
-  while (length >= 1024) {
-    uint64_t t0, t1;
-    uint32_t crc0 = 0, crc1 = 0, crc2 = 0;
-
-    /* Parallel Param:
-     *   k0 = CRC32(x ^ (42 * 8 * 8 * 2 - 1));
-     *   k1 = CRC32(x ^ (42 * 8 * 8 - 1));
-     */
-    uint32_t k0 = 0xe417f38a, k1 = 0x8f158014;
-
-    /* Prefetch data for following block to avoid cache miss */
-    PREF1KL1((uint8_t *)buf64, 1024);
-
-    /* First 8 byte for better pipelining */
-    crc0 = crc32c_u64(crc, *buf64++);
-
-    /* 3 blocks crc32c parallel computation
-     * Macro unfolding to compute parallelly
-     * 168 * 6 = 1008 (bytes)
-     */
-    CRC32C7X24BYTES(0);
-    CRC32C7X24BYTES(1);
-    CRC32C7X24BYTES(2);
-    CRC32C7X24BYTES(3);
-    CRC32C7X24BYTES(4);
-    CRC32C7X24BYTES(5);
-    buf64 += (BLK_LENGTH * 3);
-
-    /* Last 8 bytes */
-    crc = crc32c_u64(crc2, *buf64++);
-
-    t0 = (uint64_t)vmull_p64(crc0, k0);
-    t1 = (uint64_t)vmull_p64(crc1, k1);
-
-    /* Merge (crc0, crc1, crc2) -> crc */
-    crc1 = crc32c_u64(0, t1);
-    crc ^= crc1;
-    crc0 = crc32c_u64(0, t0);
-    crc ^= crc0;
-
-    length -= 1024;
-  }
-
-  if (length == 0) return crc ^ (0xffffffffU);
+    while (length >= 1024) {
+      uint64_t t0, t1;
+      uint32_t crc0 = 0, crc1 = 0, crc2 = 0;
+
+      /* Parallel Param:
+       *   k0 = CRC32(x ^ (42 * 8 * 8 * 2 - 1));
+       *   k1 = CRC32(x ^ (42 * 8 * 8 - 1));
+       */
+      uint32_t k0 = 0xe417f38a, k1 = 0x8f158014;
+
+      /* Prefetch data for following block to avoid cache miss */
+      PREF1KL1((uint8_t *)buf64, 1024);
+
+      /* First 8 byte for better pipelining */
+      crc0 = crc32c_u64(crc, *buf64++);
+
+      /* 3 blocks crc32c parallel computation
+       * Macro unfolding to compute parallelly
+       * 168 * 6 = 1008 (bytes)
+       */
+      CRC32C7X24BYTES(0);
+      CRC32C7X24BYTES(1);
+      CRC32C7X24BYTES(2);
+      CRC32C7X24BYTES(3);
+      CRC32C7X24BYTES(4);
+      CRC32C7X24BYTES(5);
+      buf64 += (BLK_LENGTH * 3);
+
+      /* Last 8 bytes */
+      crc = crc32c_u64(crc2, *buf64++);
+
+      t0 = (uint64_t)vmull_p64(crc0, k0);
+      t1 = (uint64_t)vmull_p64(crc1, k1);
+
+      /* Merge (crc0, crc1, crc2) -> crc */
+      crc1 = crc32c_u64(0, t1);
+      crc ^= crc1;
+      crc0 = crc32c_u64(0, t0);
+      crc ^= crc0;
+
+      length -= 1024;
+    }
+
+    if (length == 0) return crc ^ (0xffffffffU);
 #endif
+  }  // if Pmull runtime check here
+
   buf8 = (const uint8_t *)buf64;
   while (length >= 8) {
     crc = crc32c_u64(crc, *(const uint64_t *)buf8);

util/crc32c_arm64.h

@@ -35,6 +35,7 @@
 
 extern uint32_t crc32c_arm64(uint32_t crc, unsigned char const *data, unsigned len);
 extern uint32_t crc32c_runtime_check(void);
+extern bool crc32c_pmull_runtime_check(void);
 
 #ifdef __ARM_FEATURE_CRYPTO
 #define HAVE_ARM64_CRYPTO