This way it can be used to fast cancel/undo the last command.
Also make the tip message a litter bit nicer..
Signed-off-by: xie xingguo <xie.xingguo@zte.com.cn>
There is a lot of good stuff going on here, but nobody is investing in xio
and it is not expected to be the path forward for RDMA. If that ever
changes, we can resurrect the code. Until then, let's clean up the tree
and reduce friction for changes going forward.
Signed-off-by: Sage Weil <sage@redhat.com>
The crush/builder.c crush_add_bucket method resizes the max_buckets array
but a power of 2 when it has to expand, but the code in CrushWrapper was
assuming that if the array grew the pos for the new bucket would be the
last position in the new array. This led to a situation where the
crush_choose_arg_map args array size didn't match max_buckets, and
eventually caused a crash.
Fixes: http://tracker.ceph.com/issues/38664
Signed-off-by: Sage Weil <sage@redhat.com>
This avoids a huge pg merge from 100s to 4, which takes a long time and
makes the teuthology scrub cleanup time out.
Signed-off-by: Sage Weil <sage@redhat.com>
Henceforth, we'll require explicit `allow` caps for commands, or for the
config-key service. Blanket caps are no longer allowed for the
config-key service, except for 'allow *'.
Signed-off-by: Joao Eduardo Luis <joao@suse.de>
in b38b8e980c, we changed the upper
limit of size of `config key` 's value to 64k, so we need to update
the test accordingly.
Fixes: http://tracker.ceph.com/issues/36260
Signed-off-by: Kefu Chai <kchai@redhat.com>
5s -> 5m to gives us more leeway for when the mons are thrashing.
Also, *only* set this timeout when we expect a timeout. If we don't,
wait forever.
Signed-off-by: Sage Weil <sage@redhat.com>
If the bucket is empty or does not have weight-set weights yet, avoid
crashing when populating the parent bucket.
Fixes: http://tracker.ceph.com/issues/23386
Signed-off-by: Sage Weil <sage@redhat.com>
because the mgrclient will be waiting for the mgrmap if the mgrmap
is not available, when the client is about to send a mgr command.
and monitor will drop the subscription requests if the client does not
have enough cap for reading mon, so unlike mon commands, the client
won't get an EACCES return code in this case.
in this change, a timeout machinary is introduced. and the client
will wait for "rados-mon-op-timeout" before it gives up. if this
setting is 0, it will wait forever.
Fixes: https://tracker.ceph.com/issues/23627
Signed-off-by: Kefu Chai <kchai@redhat.com>
Keep a standalone wrapper for the workunit, so we can test it locally,
leveraging the ceph-helpers to do the setup. Keep a workunit to be
exercised by teuthology.
Signed-off-by: Joao Eduardo Luis <joao@suse.de>
So we can combine "crush add-bucket" with "crush move" command,
and hence avoid making two separate changes to the osdmap,
and hence slow down map-epoch generation.
Signed-off-by: xie xingguo <xie.xingguo@zte.com.cn>
This command returns all crush rules that are currently
referencing the device class specified by user.
Signed-off-by: xie xingguo <xie.xingguo@zte.com.cn>
User may specify a rule with the same name of the pool that it serves.
Since a pool can be renamed, so does the rule.
Signed-off-by: xie xingguo <xie.xingguo@zte.com.cn>
we cannot overwrite existing dev class, and "osd_class_update_on_start"
is true by default (see 0c885d6). so we should remove all device classes before
setting them.
Signed-off-by: Kefu Chai <kchai@redhat.com>
/bin/bash is a Linuxism. Other operating systems install bash to
different paths. Use /usr/bin/env in shebangs to find bash.
Signed-off-by: Alan Somers <asomers@gmail.com>
The CRUSH rule creation is busted (rules and buckets out of order), but
after I fix that it doesn't seem to run right anyway. Remove it.
We get the mon thrasher coverage from rados/monthrash already; I don't
think this is adding meaningful coverage for the amount of effort it takes
to maintain.
Signed-off-by: Sage Weil <sage@redhat.com>
