From: lu.shasha <lu.shasha@aishu.cn>
Date: Thu, 18 Aug 2022 07:07:42 +0000 (+0800)
Subject: rgw: fix data corruption due to network jitter
X-Git-Tag: v16.2.11~243^2
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=b1ef8f95eb533cb63695e54f11ce49d8b5e2a3cf;p=ceph.git

rgw: fix data corruption due to network jitter

In function RGWRados::Object::Write::_do_write_meta rgw_rados_operate may return -ETIMEDOUT due to network jitter.
But head obj may eventually be written successfully even if return -ETIMEDOUT. Due to r < 0, meta.canceled be set true.
In such scenario, all the tail objects will removed in the destructor of RadosWriter, then data corruptted.
To avoid data corruption, when return -ETIMEDOUT, should clear all objects in RadosWriter.

Signed-off-by: Shasha Lu <lu.shasha@aishu.cn>
(cherry picked from commit 04282ae3c1d19bf185ee1df148eb84deb244630f)
---

diff --git a/src/rgw/rgw_putobj_processor.cc b/src/rgw/rgw_putobj_processor.cc
index 68730173edfb..a4baa2308a37 100644
--- a/src/rgw/rgw_putobj_processor.cc
+++ b/src/rgw/rgw_putobj_processor.cc
@@ -310,6 +310,11 @@ int AtomicObjectProcessor::complete(size_t accounted_size,
 
   r = obj_op->prepare(y);
   if (r < 0) {
+    if (r == -ETIMEDOUT) {
+      // The head object write may eventually succeed, clear the set of objects for deletion. if it
+      // doesn't ever succeed, we'll orphan any tail objects as if we'd crashed before that write
+      writer.clear_written();
+    }
     return r;
   }