From 01fb2a99bd21c22921ca960c012a4d8fc889e920 Mon Sep 17 00:00:00 2001
From: John Sully <john@csquare.ca>
Date: Fri, 26 Mar 2021 23:48:24 +0000
Subject: [PATCH] Prefetch keys even in pure RAM scenarios

Former-commit-id: d7219de186d60a5a437c1828ac97117eaad34819
---
 src/db.cpp         | 25 ++++++++++++++++++++++++-
 src/networking.cpp |  4 ++--
 2 files changed, 26 insertions(+), 3 deletions(-)

diff --git a/src/db.cpp b/src/db.cpp
index 524cdb66c..9978a8ca5 100644
--- a/src/db.cpp
+++ b/src/db.cpp
@@ -35,6 +35,11 @@
 #include <signal.h>
 #include <ctype.h>
 
+// Needed for prefetch
+#if defined(__x86_64__) || defined(__i386__)
+#include <xmmintrin.h>
+#endif
+
 /* Database backup. */
 struct dbBackup {
     const redisDbPersistentDataSnapshot **dbarray;
@@ -3002,8 +3007,26 @@ int dbnumFromDb(redisDb *db)
 
 void redisDbPersistentData::prefetchKeysAsync(client *c, parsed_command &command)
 {
-    if (m_spstorage == nullptr)
+    if (m_spstorage == nullptr) {
+#if defined(__x86_64__) || defined(__i386__)
+        // We do a quick 'n dirty check for set & get.  Anything else is too slow.
+        //  Should the user do something weird like remap them then the worst that will
+        //  happen is we don't prefetch or we prefetch wrong data.  A mild perf hit, but
+        //  not dangerous
+        const char *cmd = szFromObj(command.argv[0]);
+        if (!strcasecmp(cmd, "set") || !strcasecmp(cmd, "get")) {
+            auto h = dictSdsHash(szFromObj(command.argv[1]));
+            for (int iht = 0; iht < 2; ++iht) {
+                auto hT = h & c->db->m_pdict->ht[iht].sizemask;
+                if (c->db->m_pdict->ht[iht].table != nullptr)
+                    _mm_prefetch(c->db->m_pdict->ht[iht].table[hT], _MM_HINT_T1);
+                if (!dictIsRehashing(c->db->m_pdict))
+                    break;
+            }
+        }
+#endif
         return;
+    }
 
     AeLocker lock;
 
diff --git a/src/networking.cpp b/src/networking.cpp
index 15aa6f43a..2919750bc 100644
--- a/src/networking.cpp
+++ b/src/networking.cpp
@@ -2358,8 +2358,8 @@ void parseClientCommandBuffer(client *c) {
             serverAssert(c->vecqueuedcmd.back().reploff >= 0);
         }
 
-        /* Prefetch if we have a storage provider and we're not in the global lock */
-        if (cqueries < c->vecqueuedcmd.size() && g_pserver->m_pstorageFactory != nullptr && !GlobalLocksAcquired()) {
+        /* Prefetch outside the lock for better perf */
+        if (cqueries < c->vecqueuedcmd.size() && !GlobalLocksAcquired()) {
             auto &query = c->vecqueuedcmd.back();
             if (query.argc > 0 && query.argc == query.argcMax) {
                 c->db->prefetchKeysAsync(c, query);