rgw: RGWMetaSyncShardCR drops stack refs on destruction

[cbodley]

if the coroutine is canceled before collect_children() can clean up
all of its child stacks, those stack refs will leak. store these
stacks as boost::intrusive_ptr so the ref is dropped automatically on
destruction

Fixes: http://tracker.ceph.com/issues/18300

[yehudasa]

looks good, has it been tested?

[cbodley]

i added a dout to the RGWMetaSyncShardCR destructor and killed the gateway in the middle of test_multi.py:

2016-12-22 11:11:18.958001 7fb24c500700  1 meta sync: RGWMetaSyncShardCR canceled, releasing 2 stacks
2016-12-22 11:11:18.958077 7fb24c500700  1 meta sync: RGWMetaSyncShardCR canceled, releasing 3 stacks

and the valgrind came out clean

[liewegas]

rebased onto kraken and pushed wip-18300 to ceph-ci so we can run through rgw suite

[liewegas]

http://pulpito.ceph.com/sage-2016-12-26_18:32:29-rgw-wip-sage-testing---basic-smithi/

still has some valgrind failures, both leaks and the invalid read. see /a/sage-2016-12-26_18:32:29-rgw-wip-sage-testing---basic-smithi/668072

[cbodley]

thanks for running this @liewegas! looking through the results

[cbodley]

i identified 4 different valgrind failures in those results. one was the invalid read bug that Matt's been working on. i opened tracker issues for the other three:

• multisite: memory leak in RGWCloneMetaLogCoroutine::state_store_mdlog_entries()
• multisite: memory leak from RGWSimpleRadosLockCR::send_request()
• multisite: use after free in RGWCloneMetaLogCoroutine::state_read_shard_status() (exposed by the fix in this PR)

the original leak addressed by this PR did not reproduce

[cbodley]

added two commits to address the use-after-free issue that was uncovered by the initial fix. teuthology run pending..

[cbodley]

updates:

• marked [DNM] because the original fix uncovered a use-after-free
• changed RGWMetadataLog::get_info_async() interface to address use-after-free, but it caused a couple segfaults (http://pulpito.ceph.com/cbodley-2017-01-09_10:33:07-rgw-wip-cbodley-testing---basic-mira/)

[cbodley]

I didn't realize that there's a loop in RGWCloneMetaLogCoroutine::operate(), so state_read_shard_status() was being called multiple times with the same completion, and this was hitting an assertion in librados. This means the completion's lifetime has to be separate from RGWCloneMetaLogCoroutine. I pushed an updated version that allocates the completion separately, and cancels its callback in ~RGWCloneMetaLogCoroutine(). Will schedule another run of wip-cbodley-testing.

[cbodley]

latest run had some leaks of RGWMetadataLogInfoCompletion. i added a fix for the interaction between boost::intrusive_ptr and RefCountedObj. i updated the wip-cbodley-testing branch (and included Matt's fixes from #12767) and will reschedule as soon as it's built

[cbodley]

teuthology results were green, except for a single valgrind failure in the osd and an s3test failure due to timeouts on slow requests

[smithfarm]

@cbodley The jewel backport is non-trivial due to the use of boost::intrusive_ptr which is not available in jewel. If you want to do it, the backport tracker issue is http://tracker.ceph.com/issues/18563