rgw: prevent another leftover bucket index olh entry scenario

If a call to bucket_index_link_olh or bucket_index_unlink_instance
fails, its associated pending xattr may have prevented the olh object
from being removed by another thread. We should do a best effort
cleanup attempt for this case by calling update_olh before returning
an error to the caller.

Signed-off-by: Cory Snyder <csnyder@1111systems.com>
(cherry picked from commit 570adec5bb8142f5baf1f05f0040e8afdb11ec05)

Conflicts:
	src/rgw/driver/rados/rgw_rados.cc

Cherry-pick notes:
- Conflicts due to rgw_rados file being moved into driver dir in later versions

diff --git a/src/rgw/rgw_rados.cc b/src/rgw/rgw_rados.cc
index 4004135747e4a77d68c28b08ed9c354bba1ee577..1e115f98b26fa7a5f9dfdd1c62e6b2566f1a5fd9 100644 (file)
--- a/src/rgw/rgw_rados.cc
+++ b/src/rgw/rgw_rados.cc
@@ -7520,6 +7520,13 @@ int RGWRados::set_olh(const DoutPrefixProvider *dpp, RGWObjectCtx& obj_ctx, cons
         }
         continue;
       }
+      // it's possible that the pending xattr from this op prevented the olh
+      // object from being cleaned by another thread that was deleting the last
+      // existing version. We invoke a best-effort update_olh here to handle this case.
+      int r = update_olh(dpp, obj_ctx, state, bucket_info, olh_obj);
+      if (r < 0 && r != -ECANCELED) {
+        ldpp_dout(dpp, 20) << "update_olh() target_obj=" << olh_obj << " returned " << r << dendl;
+      }
       return ret;
     }
     break;
@@ -7578,9 +7585,17 @@ int RGWRados::unlink_obj_instance(const DoutPrefixProvider *dpp, RGWObjectCtx& o
     ret = bucket_index_unlink_instance(dpp, bucket_info, target_obj, op_tag, olh_tag, olh_epoch, zones_trace);
     if (ret < 0) {
       ldpp_dout(dpp, 20) << "bucket_index_unlink_instance() target_obj=" << target_obj << " returned " << ret << dendl;
+      olh_cancel_modification(dpp, bucket_info, *state, olh_obj, op_tag, y);
       if (ret == -ECANCELED) {
         continue;
       }
+      // it's possible that the pending xattr from this op prevented the olh
+      // object from being cleaned by another thread that was deleting the last
+      // existing version. We invoke a best-effort update_olh here to handle this case.
+      int r = update_olh(dpp, obj_ctx, state, bucket_info, olh_obj, zones_trace);
+      if (r < 0 && r != -ECANCELED) {
+        ldpp_dout(dpp, 20) << "update_olh() target_obj=" << olh_obj << " returned " << r << dendl;
+      }
       return ret;
     }
     break;