Add rbd_skip_partial_discard flag

If we want to discard a range of an object, we will zero(use fallocate to punch a hole) the range now. In general this introduce some overhead(extra writes). If the filesystem ontop of RBD holding lots of small files, this behavior will bring big performance penalty. Adding a flag that allow user to control if they want to zero the range. Signed-off-by: Xiaoxi Chen <[email protected]>
zmc · Apr 1, 2015 · e7812b8 · e7812b8
1 parent 497737a
commit e7812b8
Show file tree

Hide file tree

Showing 2 changed files with 8 additions and 3 deletions.
diff --git a/src/common/config_opts.h b/src/common/config_opts.h
@@ -892,6 +892,7 @@ OPTION(rbd_clone_copy_on_read, OPT_BOOL, false)
 OPTION(rbd_blacklist_on_break_lock, OPT_BOOL, true) // whether to blacklist clients whose lock was broken
 OPTION(rbd_blacklist_expire_seconds, OPT_INT, 0) // number of seconds to blacklist - set to 0 for OSD default
 OPTION(rbd_request_timed_out_seconds, OPT_INT, 30) // number of seconds before maint request times out
+OPTION(rbd_skip_partial_discard, OPT_BOOL, false) // when trying to discard a range inside an object, set to true to skip zeroing the range.
 
 /*
  * The following options change the behavior for librbd's image creation methods that

diff --git a/src/librbd/internal.cc b/src/librbd/internal.cc
@@ -3576,9 +3576,13 @@ namespace librbd {
 	req = new AioTruncate(ictx, p->oid.name, p->objectno, p->offset, objectx, object_overlap,
 			      snapc, snap_id, req_comp);
       } else {
-	req = new AioZero(ictx, p->oid.name, p->objectno, p->offset, p->length,
-			  objectx, object_overlap,
-			  snapc, snap_id, req_comp);
+	if(ictx->cct->_conf->rbd_skip_partial_discard) {
+	  continue;
+	} else {
+	  req = new AioZero(ictx, p->oid.name, p->objectno, p->offset, p->length,
+			    objectx, object_overlap,
+			    snapc, snap_id, req_comp);
+	}
       }
 
       r = req->send();