[
https://issues.apache.org/jira/browse/ZOOKEEPER-2841?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16079343#comment-16079343
]
ASF GitHub Bot commented on ZOOKEEPER-2841:
-------------------------------------------
Github user hanm commented on the issue:
https://github.com/apache/zookeeper/pull/306
Should we put the CMake files in a dedicated folder called "windows" or
something like that? Otherwise we will have the CMakeList file, and the
existing makefile and config files (which are not touched by this pull request)
coexisting under the C client folder. This might create confusions for user
because makefile can be generated from the CMakeList file, so the question of
which makefile is the source of truth and so on...
Ultimately we might drop existing makefiles in favor of the CMake so build
files across every platform are generated by a single source of truth
(cmakelist), but that's more scope of work (the ant build file has to be
modified and the system requirements of ZK has to be updated to make cmake a
dependency for source installations etc to avoid break existing workflows), so
for now I think it's good to scope cmake so it only impact windows C client
builds.
> ZooKeeper public include files leak porting changes
> ---------------------------------------------------
>
> Key: ZOOKEEPER-2841
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-2841
> Project: ZooKeeper
> Issue Type: Bug
> Components: c client
> Environment: Windows 10 with Visual Studio 2017
> Reporter: Andrew Schwartzmeyer
> Assignee: Andrew Schwartzmeyer
> Labels: windows
>
> The fundamental problem is that the port of the C client to Windows is now
> close to six years old, with very few updates. This port leaks a lot of
> changes that should be internal to ZooKeeper, and many of those changes are
> simply no longer relevant. The correct thing to do is attempt to refactor the
> Windows port for new versions of ZooKeeper, removing dead/unneeded porting
> code, and moving dangerous porting code to C files instead of public headers.
> Two primary examples of this problem are
> [ZOOKEEPER-2491|https://issues.apache.org/jira/browse/ZOOKEEPER-2491] and
> [MESOS-7541|https://issues.apache.org/jira/browse/MESOS-7541].
> The first issue stems from this ancient porting code:
> {noformat}
> #define snprintf _snprintf
> {noformat}
> in
> [winconfig.h|https://github.com/apache/zookeeper/blob/ddf0364903bf7ac7cd25b2e1927f0d9d3c7203c4/src/c/include/winconfig.h#L179].
> Newer versions of Windows C libraries define {{snprintf}} as a function, and
> so it cannot be redefined.
> The second issue comes from this undocumented change:
> {noformat}
> #undef AF_INET6
> {noformat}
> again in
> [winconfig.h|https://github.com/apache/zookeeper/blob/ddf0364903bf7ac7cd25b2e1927f0d9d3c7203c4/src/c/include/winconfig.h#L169]
> which breaks any library that uses IPv6 and {{winsock2.h}}.
> Furthermore, the inclusion of the following defines and headers causes
> terrible problems for consuming libraries, as they leak into ZooKeeper's
> public headers:
> {noformat}
> #define _CRT_SECURE_NO_WARNINGS
> #define WIN32_LEAN_AND_MEAN
> #include <Windows.h>
> #include <Winsock2.h>
> #include <winstdint.h>
> #include <process.h>
> #include <ws2tcpip.h>
> {noformat}
> Depending on the order that a project includes or compiles files, this may or
> may not cause {{WIN32_LEAN_AND_MEAN}} to become unexpectedly defined, and
> {{windows.h}} to be unexpectedly included. This problem is exacberated by the
> fact that the {{winsock2.h}} and {{windows.h}} headers are order-dependent
> (if you read up on this, you'll see that defining {{WIN32_LEAN_AND_MEAN}} was
> meant to work-around this).
> Going forward, porting changes should live next to where they are used,
> preferably in source files, not header files, so they remain contained.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
