[Gluster-devel] add-brick

[Emmanuel Dreyfus]

> I understand this code checks that the layout cover the whole space,
> is that right? Then it must be upset that layout->list[0] does not cover
> anything. Since the error is transcient, I susepct a race condition:
> the layout would be filled after that check. Is it possible? Where is
> the layout crafted?

I improved my test by completely deleting and re-creating the 
volume before adding a brick. Here is what happens when I add a brick:

1-vndfs-client-1: Connected to 192.0.2.103:24027, attached to remote 
  volume '/export/vnd1a'.
1-vndfs-client-1: Server and Client lk-version numbers are not same, 
  reopening the fds
0-fuse: switched to graph 1
1-vndfs-client-1: Server lk version = 1
1-vndfs-dht: missing disk layout on vndfs-client-0. err = -1
1-dht_layout_merge: ==> layout[0] 0 - 0 err -1
1-dht_layout_merge: ==> layout[1] 0 - 0 err 0
1-vndfs-dht: missing disk layout on vndfs-client-1. err = -1
1-dht_layout_merge: ==> layout[0] 0 - 0 err -1
1-dht_layout_merge: ==> layout[1] 0 - 0 err -1

I am not sure it is expected behavior. The broken layout does not
raise EINVAL to process using the filesystem, but later similar
treatment will.

After playing a bit I tested the race condition with this patch:

--- a/xlators/cluster/dht/src/dht-common.c
+++ b/xlators/cluster/dht/src/dht-common.c
@@ -477,6 +477,11 @@ unlock:
                         ret = dht_layout_normalize (this, &local->loc, layout);
 
                         if (ret != 0) {
+                                if (strcmp(local->loc.path, "/") == 0) {
+                                        gf_log (this->name, GF_LOG_WARNING,
+                                           "wAit 2s for DHT to settle...");
+                                        sleep(2);
+                                }
                                 gf_log (this->name, GF_LOG_DEBUG,
                                         "fixing assignment on %s",
                                         local->loc.path);

Here is the kind og log it procudes. I do not always see the EINVAL 
in the log, but it is never seen by processes using the filesystem.
At least during the tests I did.

[2012-08-19 06:04:06.288131] I [fuse-bridge.c:4195:fuse_graph_setup] 
  0-fuse: switched to graph 1
[2012-08-19 06:04:06.289052] I [client-handshake.c:453:
  client_set_lk_version_cbk] 1-vndfs-client-1: Server lk version = 1
[2012-08-19 06:04:06.294234] W [dht-common.c:482:dht_lookup_dir_cbk] 
  1-vndfs-dht: wait 2s for DHT to settle...
[2012-08-19 06:04:08.306937] I [client.c:2151:notify] 0-vndfs-client-0: 
  current graph is no longer active, destroying rpc_client 
[2012-08-19 06:04:08.308114] I [client.c:2090:client_rpc_notify] 
  0-vndfs-client-0: disconnected
[2012-08-19 06:04:08.309833] W [fuse-resolve.c:151:fuse_resolve_gfid_cbk] 
  0-fuse: 4e4b4110-a585-4aae-b919-b2416355f5d1: failed to resolve 
  (Invalid argument)
[2012-08-19 06:04:08.310275] E [fuse-bridge.c:353:fuse_lookup_resume] 
  0-fuse: failed to resolve path (null)

But this probably does not really fix the problem. I got an unreproductible
ENOENT for a directory while copying a hierarchy for instance.

-- 
Emmanuel Dreyfus
manu at netbsd.org