High - Memory allocated on stack by xblcoud_<put/delete/download> may
have been accessed by the callbacks after the function has
exited. Fixed by making xblcoud_<put/delete/download> to wait
for completion of all asynchronous requests.
- xbcloud/xbstream on centos 7 did not preserve alphabetical
order. Fixed by passing ls output through the sort command.
- Older libcurl used on Centos 6 freed internal structures of an
easy_handle when curl_multi_remove_handle was called. These
structures were than accessed in Http_connection::finalize. Fix
swaps two calls.
- Older libcurl version used in Centos 6 could fail to return
valid value for CURLINFO_EFFECTIVE_URL. Error code check was
added.
- Older libcurl version used in Centos 6 does not understand
Content-Length header passed manually and ended up sending two
Content-Length headers. Which broke S3 authentication.
- Older libcurl version used in Centos 6 can return
CURLM_CALL_MULTI_PERFORM in response to
curl_multi_socket_action. This error code was not handled by
mcode_or_die.
- Class Object_store did not have virtual destructor which caused
memory leaks of small amount of resources.
- Because of the bad timings, event loop could have started after we
pushed some work into the queue which caused missed async events and
timeouts. Fixed by adding a kickoff timer who's only goal is to
indicate that event loop has started.
- Even though libcurl claims to be thread-safe, this is not always true
when it comes to using it with OpenSSL. xbcloud makes all HTTP
requests using curl_multi in single thread. There is another thread
which can call curl_easy_escape. curl_easy handle was created for each
of such calls. Crash was observed on Debian 9 and the stack trace
included curl_easy_cleanup followed by OpenSSL cleanup code. It is
known that in order to use older versions of OpenSSL in multi-threaded
environment, one needed to provide locking callbacks. Debian 9,
however, comes with OpenSSL 1.1.0j, which has built-in support for
multi-threading which should just work. Fix (more a workaround) is to
use single curl_easy for all curl_easy_escape calls, guard the calls
with mutex (which is not really needed at the moment, because all url
escape calls come from one thread).
