From 0cd455d17f8da5cf407d1b21a4dfbc7b04be3f00 Mon Sep 17 00:00:00 2001
From: Vitaliy Filippov <vitalif@yourcmc.ru>
Date: Tue, 5 Mar 2024 00:08:20 +0300
Subject: [PATCH] First just recheck version without actually re-reading block
 in vitastor-kv

---
 src/cluster_client.cpp |  2 +-
 src/kv_db.cpp          | 28 +++++++++++++++++++++++-----
 2 files changed, 24 insertions(+), 6 deletions(-)

diff --git a/src/cluster_client.cpp b/src/cluster_client.cpp
index 9050b6bd..70920452 100644
--- a/src/cluster_client.cpp
+++ b/src/cluster_client.cpp
@@ -673,7 +673,7 @@ bool cluster_client_t::check_rw(cluster_op_t *op)
         return false;
     }
     // Check alignment
-    if (!op->len && (op->opcode == OSD_OP_READ || op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP || op->opcode == OSD_OP_WRITE) ||
+    if (!op->len && (op->opcode == OSD_OP_READ_BITMAP || op->opcode == OSD_OP_READ_CHAIN_BITMAP || op->opcode == OSD_OP_WRITE) ||
         op->offset % pool_it->second.bitmap_granularity || op->len % pool_it->second.bitmap_granularity)
     {
         op->retval = -EINVAL;
diff --git a/src/kv_db.cpp b/src/kv_db.cpp
index 35a2d007..2bb26fb3 100644
--- a/src/kv_db.cpp
+++ b/src/kv_db.cpp
@@ -888,14 +888,23 @@ static void get_block(kv_db_t *db, uint64_t offset, int cur_level, int recheck_p
     op->opcode = OSD_OP_READ;
     op->inode = db->inode_id;
     op->offset = offset;
-    op->len = db->kv_block_size;
-    op->iov.push_back(malloc_or_die(op->len), op->len);
+    if (b_it != db->block_cache.end() && !b_it->second.invalidated && !b_it->second.updating)
+    {
+        // just recheck version - it's cheaper than re-reading the block
+        op->len = 0;
+    }
+    else
+    {
+        op->len = db->kv_block_size;
+        op->iov.push_back(malloc_or_die(op->len), op->len);
+    }
     op->callback = [=](cluster_op_t *op)
     {
         if (op->retval != op->len)
         {
             // error
-            free(op->iov.buf[0].iov_base);
+            if (op->len)
+                free(op->iov.buf[0].iov_base);
             cb(op->retval >= 0 ? -EIO : op->retval, BLK_NOCHANGE);
             delete op;
             return;
@@ -910,7 +919,8 @@ static void get_block(kv_db_t *db, uint64_t offset, int cur_level, int recheck_p
             if (blk->updating > 0 && recheck_policy == KV_RECHECK_WAIT)
             {
                 // Wait until block update stops
-                free(op->iov.buf[0].iov_base);
+                if (op->len)
+                    free(op->iov.buf[0].iov_base);
                 delete op;
                 db->continue_update.emplace(blk->offset, [=, blk_offset = blk->offset]()
                 {
@@ -924,6 +934,13 @@ static void get_block(kv_db_t *db, uint64_t offset, int cur_level, int recheck_p
         }
         else
         {
+            if (!op->len)
+            {
+                // Version check failed, re-read block
+                delete op;
+                get_block(db, offset, cur_level, recheck_policy, cb);
+                return;
+            }
             auto blk = &db->block_cache[op->offset];
             if (blk_it != db->block_cache.end())
             {
@@ -945,7 +962,8 @@ static void get_block(kv_db_t *db, uint64_t offset, int cur_level, int recheck_p
             }
             try_evict(db);
         }
-        free(op->iov.buf[0].iov_base);
+        if (op->len)
+            free(op->iov.buf[0].iov_base);
         delete op;
     };
     db->cli->execute(op);