From ba238245618e41ab37e48ab1f26d99c59f2a0f4e Mon Sep 17 00:00:00 2001
From: Vitaliy Filippov <vitalif@yourcmc.ru>
Date: Mon, 6 Jan 2020 21:01:11 +0300
Subject: [PATCH] Allow to disable zero fill

---
 blockstore_impl.h    | 5 +++++
 blockstore_open.cpp  | 4 ++++
 blockstore_write.cpp | 9 +++++----
 3 files changed, 14 insertions(+), 4 deletions(-)

diff --git a/blockstore_impl.h b/blockstore_impl.h
index d3559e2b..62e87611 100644
--- a/blockstore_impl.h
+++ b/blockstore_impl.h
@@ -195,7 +195,12 @@ class blockstore_impl_t
 
     bool readonly = false;
     // FIXME: separate flags for data, metadata and journal
+    // It is safe to disable fsync() if drive write cache is writethrough
     bool disable_fsync = false;
+    // It is safe to disable zero fill if drive is zeroed before formatting.
+    // For example, with TRIM and Deterministic Read Zeroes after TRIM.
+    // FIXME: OP_DELETE should trim/zero out the block.
+    bool zerofill_enabled = false;
     bool inmemory_meta = false;
     void *metadata_buffer = NULL;
 
diff --git a/blockstore_open.cpp b/blockstore_open.cpp
index 52002171..4741ec8f 100644
--- a/blockstore_open.cpp
+++ b/blockstore_open.cpp
@@ -10,6 +10,10 @@ void blockstore_impl_t::calc_lengths(blockstore_config_t & config)
     {
         disable_fsync = true;
     }
+    if (config["zerofill"] == "true" || config["zerofill"] == "1" || config["zerofill"] == "yes")
+    {
+        zerofill_enabled = true;
+    }
     // data
     data_len = data_size - data_offset;
     if (data_fd == meta_fd && data_offset < meta_offset)
diff --git a/blockstore_write.cpp b/blockstore_write.cpp
index 3a85b556..7089519e 100644
--- a/blockstore_write.cpp
+++ b/blockstore_write.cpp
@@ -102,10 +102,10 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
 #endif
         data_alloc->set(loc, true);
         int vcnt = 0;
-        if (op->version == 1 && op->len != block_size)
+        uint64_t stripe_offset = 0;
+        if (op->len != block_size && zerofill_enabled)
         {
-            // Zero fill newly allocated object. First write is always a big write
-            // FIXME: Add "no-zero-fill" mode which will just leave random garbage (insecure, but may be useful)
+            // Zero fill newly allocated object if required
             if (op->offset > 0)
                 PRIV(op)->iov_zerofill[vcnt++] = (struct iovec){ zero_object, op->offset };
             PRIV(op)->iov_zerofill[vcnt++] = (struct iovec){ op->buf, op->len };
@@ -118,10 +118,11 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
             vcnt = 1;
             PRIV(op)->iov_zerofill[0] = (struct iovec){ op->buf, op->len };
             data->iov.iov_len = op->len; // to check it in the callback
+            stripe_offset = op->offset;
         }
         data->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
         my_uring_prep_writev(
-            sqe, data_fd, PRIV(op)->iov_zerofill, vcnt, data_offset + (loc << block_order)
+            sqe, data_fd, PRIV(op)->iov_zerofill, vcnt, data_offset + (loc << block_order) + stripe_offset
         );
         PRIV(op)->pending_ops = 1;
         PRIV(op)->min_used_journal_sector = PRIV(op)->max_used_journal_sector = 0;