From: Radoslaw Zarzynski <rzarzyns@redhat.com>
Date: Mon, 29 Jan 2018 16:27:09 +0000 (+0100)
Subject: os/bluestore: mark derivatives of AioContext as final.
X-Git-Tag: v13.0.2~296^2
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=aff36389e705735d5c0b89b27e015a00ce3444fd;p=ceph.git

os/bluestore: mark derivatives of AioContext as final.

Profiling `AioReadBatch`, a freshly introduced (at the moment in
a wip branch; see PR #19380) descendant of `AioContext` showed
that CPU cycles are burnt unnecessarily because of the dynamically
polymorphic destructor call that can't be inlined:

It's quite probable the same affects `TransContext` and `DeferredBatch`.
Marking the classes as final allows compiler to inline the dtors calls
in `::aio_finish` and, hopefully, optimize some of the extra overhead.

Signed-off-by: Radoslaw Zarzynski <rzarzyns@redhat.com>
---

diff --git a/src/os/bluestore/BlueStore.h b/src/os/bluestore/BlueStore.h
index 32f4cde3d9cd..5ab31791d109 100644
--- a/src/os/bluestore/BlueStore.h
+++ b/src/os/bluestore/BlueStore.h
@@ -1480,7 +1480,7 @@ public:
     }
   };
 
-  struct TransContext : public AioContext {
+  struct TransContext final : public AioContext {
     MEMPOOL_CLASS_HELPERS();
 
     typedef enum {
@@ -1619,7 +1619,7 @@ public:
       boost::intrusive::list_member_hook<>,
       &TransContext::deferred_queue_item> > deferred_queue_t;
 
-  struct DeferredBatch : public AioContext {
+  struct DeferredBatch final : public AioContext {
     OpSequencer *osr;
     struct deferred_io {
       bufferlist bl;    ///< data