> On Jan. 18, 2017, 2:32 a.m., Joseph Wu wrote:
> > 3rdparty/stout/include/stout/os/windows/rmdir.hpp, lines 120-123
> > <https://reviews.apache.org/r/55327/diff/2/?file=1605538#file1605538line120>
> >
> >     Note: This is actually unreachable.

:| Embarrassing.


- Alex


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/55327/#review162019
-----------------------------------------------------------


On Jan. 15, 2017, 8:40 a.m., Alex Clemmer wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/55327/
> -----------------------------------------------------------
> 
> (Updated Jan. 15, 2017, 8:40 a.m.)
> 
> 
> Review request for mesos, Andrew Schwartzmeyer, Daniel Pravat, and Joseph Wu.
> 
> 
> Repository: mesos
> 
> 
> Description
> -------
> 
> The Windows implementation of `os::rmdir` will fail to delete "hanging"
> symlinks (i.e., symlinks whose targets do not exist). Note that on
> Windows this bug is specific to symlinks whose targets are _deleted_,
> since it is impossible to create a symlink whose target does not exist.
> 
> The primary issue that causes this problem is that it is very difficult
> to tell whether a symlink points at a directory or a file unless you
> resolve the symlink and determine whether the target is a directory or a
> file. In situations where the target does not exist, we can't use this
> information, and so `os::rmdir` occasionally mis-routes a symlink to
> (what was) a directory to a `::remove` call, which will fail with a
> cryptic error.
> 
> To fix this behavior, this commit will introduce code that simply tries
> to remove the reparse point with both `RemoveDirectory` and
> `DeleteFile`, and if either succeeds, we report success for the
> operation. This represents a "best effort"; in the case that the reparse
> point represents something more exotic than a symlink, we will still
> fail, but by choosing not to verify whether the target is a directory or
> a file, we simplify the code and still obtain the outcome of having
> deleted the directory.
> 
> This commit is the primary blocker for MESOS-6707, as deleting the Agent
> sandbox will sometimes cause us to delete the latest run directory for
> the executor before the symlinked `latest` directory itself. This causes
> the delete to fail, and then the GC tests to fail, since they tend to
> assert the directory does not exist.
> 
> 
> Diffs
> -----
> 
>   3rdparty/stout/include/stout/os/windows/rmdir.hpp 
> 4437484c068e9ef046e0be14683c97db447f2da1 
>   3rdparty/stout/tests/os/rmdir_tests.cpp 
> 988d41b7fdd11cc96ce005671a7c62d1b5a3615d 
> 
> Diff: https://reviews.apache.org/r/55327/diff/
> 
> 
> Testing
> -------
> 
> 
> Thanks,
> 
> Alex Clemmer
> 
>

Reply via email to