xie xingguo [Tue, 1 Mar 2016 11:03:00 +0000 (19:03 +0800)]
msg: async: start over after failing to bind a port in specified range
The original report associated with this fix can be found at
http://tracker.ceph.com/issues/13002. I reopen one in case it
needs to be backported or something alike.
Marcus Watts [Sat, 27 Feb 2016 07:34:14 +0000 (02:34 -0500)]
cmake fix: build civetweb with the right include path.
In an earlier commit, civetweb.c got moved into a separate cmake 'object'
library civetweb_common_objs so that it could be built just once for use
with several different targets. At the time, there was a separate global
"include_directory" to give the right include path for including "civetweb.h".
A later commit in master created an rgw_a library, and restricted civetweb.h's
include path for only building rgw_a objects. So, as the product of
these these two commits, the target_include_directories command for the
civetweb include path now needs to be applied to civetweb_common_objs
and not rgw_a.
Marcus Watts [Tue, 23 Feb 2016 23:35:21 +0000 (18:35 -0500)]
Use fixed version of civetweb to avoid warning.
The warning about SSLv23_client_method is because when using dynamic
linking, this should always be non-zero. That's inside of civetweb,
so this commit is to pull the submodule up to the fix.
Marcus Watts [Thu, 11 Feb 2016 00:14:06 +0000 (19:14 -0500)]
Use ld.so to link in ssl crypto (cmake changes).
Using dlopen() to load in ssl crypto makes me cringe.
Also it requires either installing the devel package,
or appending library versioning to the library name
civetweb tries to load, which is ugly. This is only
a license problem for people whose system distribution
does not include openssl.
vsjulu [Mon, 22 Feb 2016 19:40:41 +0000 (11:40 -0800)]
XIO: Changes to fix incorrect ip being assigned in case of multiple RDMA ports.remove rdma local and return error instead to avoid issues with using rdma_local incase of multiple ports.
This allows the http client to turn off ssl certificate peer checking,
which is turned on by default. This is useful in cases like when
Keystone is SSL terminated with a self signed certificate.
The option `rgw_keystone_verify_ssl` (default true) can be toggled if
self signed certs are used, so that swift and s3 apis using keystone
authentication can work.
Fixes: #14583 Reported-by: Karol Mroz <mroz.karol@gmail.com> Signed-off-by: Abhishek Lekshmanan <abhishek@suse.com>
Casey Bodley [Thu, 25 Feb 2016 18:07:54 +0000 (13:07 -0500)]
rgw: use pimpl pattern for RGWPeriodHistory
this removes the dependency on RGWPeriod from rgw_rados.h, which breaks
a circular dependency between rgw_rados.h -> rgw_metadata.h ->
rgw_period_history.h -> rgw_rados.h
Reported-by: Willem Jan Withagen <wjw@digiware.nl> Signed-off-by: Casey Bodley <cbodley@redhat.com>
Samuel Just [Mon, 16 Nov 2015 18:06:45 +0000 (10:06 -0800)]
src/osd: use unique_ptr for backend trasaction, move into submit_transaction
submit_transaction takes ownership of the transaction implicitely. Make
this implicit using an rvalue ref to force usage of std::move and move
constructor.
Samuel Just [Fri, 13 Nov 2015 18:37:45 +0000 (10:37 -0800)]
ReplicatedPG: move do_osd_op_effects out of eval_repop
This is better anyway, do_osd_op_effects on a non-client op was somewhat
silly without an op (even if you filled in the other fields like
watch_connects, it would just silently skip it). It also eliminates
another dependence of RepGather on OpContext.
Samuel Just [Fri, 12 Feb 2016 17:32:45 +0000 (09:32 -0800)]
ReplicatedPG: move client reply handling out of eval_repop
execute_ctx is the only path which actually has a client op,
let's clarify eval_repop by moving that logic inline there.
This also eliminates most of the repop->ctx users from
eval_repop.
Samuel Just [Wed, 11 Nov 2015 20:40:16 +0000 (12:40 -0800)]
osd/: replace simple_repop_.* with simple_opc_.*
This way, we don't expose the RepGather structure
up into the users (which pretty much exclusively
use the repop->ctx member anyway). This will pave
the way to removing RepGather::ctx.
Part of this involves generalizing the repop callback
members (queue_snap_trimmer and on_applied) to
on_applied, on_committed, on_success, and on_finish
on both OpContext and RepGather.
Erwan Velu [Thu, 25 Feb 2016 17:06:17 +0000 (18:06 +0100)]
test/ceph_objectstore_tool: Deleting btrfs subvolumes before exiting
When running the ceph_objectstore_tool test suite, we shall delete the btrfs
subvolumes that have been created before trying deleting the content unless that will generate a permission denied
issue.
While running a make check on a btrfs system, many subvolumes are let at the end
of the build. It's pretty common to have several hundreds of those.
btrfs is pretty sensible to the path when requesting a subvolume removal.
The current code was misleading the path and didn't deleted the remaining
volumes.
This patch list the current subvolumes, filter thoses created by the
test process and ajust the path because brtfs reports
erwan/chroot/ceph/src/testdir/test-7202/dev/osd1/snap_439
while regarding the current working directory we want to delete :
testdir/test-7202/dev/osd1/snap_439
Piotr Dałek [Thu, 25 Feb 2016 14:12:52 +0000 (15:12 +0100)]
config: increase default async op threads
Default value of two async threads is enough for light workloads,
but is insufficient for heavy ones, reducing maximum throughput on
slower CPUs that have plenty of logical cores. Three async threads
give noticeable perf jump without any noticeable downsides.
Signed-off-by: Piotr Dałek <piotr.dalek@ts.fujitsu.com>
Kefu Chai [Thu, 25 Feb 2016 13:06:04 +0000 (21:06 +0800)]
osd: always cleanup the scrub results
the destructor of Scrub::Store asserts that `this->results` should
always be empty when it is free'd. before this change, if we are
repairing, the results is not flushed or cleaned up, this crashes
osd daemon when a new PG internval comes. also we should not keep
the results around even we don't want to populate them to the omap
if we are repairing, it's just a waste.
Loic Dachary [Mon, 22 Feb 2016 04:38:46 +0000 (11:38 +0700)]
tests: test_pidfile.sh lingering processes
For the mon, using ps aux to find the process will only show part of the
command line and the $RUNID may be missing. For the OSD, the $RUNID is
passed using an environment variable and will not show at all.
There is no need to test both ceph-osd and ceph-mon since they both end
up in the same code path, pidfile_write. Add a log line that only shows
when the --pid-file is empty and grep for it in the simplest possible
way.
The TEST_pidfile test failed to kill a process because there was
no pidfile with its pid. This is fixed by backuping the pidfile before
it is modified.
Kefu Chai [Sat, 30 Jan 2016 07:22:49 +0000 (15:22 +0800)]
osd: persist inconsistent snapsets using omap
the inconsistent snapsets are identified in ReplicatedPG::_scrub()
after we compared the authorized objects with their replica/shards.
these inconsistent information are stored in the omap of objects
with prefix "SCRUB_SS_".
Kefu Chai [Sat, 30 Jan 2016 06:38:48 +0000 (14:38 +0800)]
osd: persist inconsistent objs using omap
persist inconsistent objects found when comparing the ScrubMaps
collected from replica/shards. the discrepancies between the auth
copy and the replica are identified as inconsistencies. and hence
encoded into the omap of an object of the temp coll of the PG in
question.
scrub_types.{h,cpp} are introduced to hide the details of how we
persist the scrub types from the librados client.
projects / ceph.git / log