[
https://issues.apache.org/jira/browse/MESOS-3335?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15245869#comment-15245869
]
Benjamin Bannier commented on MESOS-3335:
-----------------------------------------
This appears a bigger problem than just possibly mutilating some help string.
Above call to {{add}} uses a pointer to memory not necessarily associated to
the {{Flags}} object for both reading and writing, and like you show even in
situations where the scope holding on to that memory has already been cleaned
up. We do use this {{add}} overload (there are actually two of them) to add a
large portion of existing {{Flag}} values.
This being undefined behavior we cannot rely on this error just messing up
things locally; in fact e.g., an optimized test built built with clang-3.9.0
will immediately segfault because of this.
I believe a possible fix would be to enforce that we only pass pointers to
member variables and add the usage sites get the actual values via the current
{{Flags}} object; this is already used partially to implement {{add}} support
of {{Flag}} values from derived {{Flags}}.
> FlagsBase copy-ctor leads to dangling pointer
> ---------------------------------------------
>
> Key: MESOS-3335
> URL: https://issues.apache.org/jira/browse/MESOS-3335
> Project: Mesos
> Issue Type: Bug
> Reporter: Neil Conway
> Assignee: Benjamin Bannier
> Priority: Minor
> Attachments: lambda_capture_bug.cpp
>
>
> Per [#3328], ubsan detects the following problem:
> [ RUN ] FaultToleranceTest.ReregisterCompletedFrameworks
> /mesos/3rdparty/libprocess/3rdparty/stout/include/stout/flags/flags.hpp:303:25:
> runtime error: load of value 33, which is not a valid value for type 'bool'
> I believe what is going on here is the following:
> * The test calls StartMaster(), which does MesosTest::CreateMasterFlags()
> * MesosTest::CreateMasterFlags() allocates a new master::Flags on the stack,
> which is subsequently copy-constructed back to StartMaster()
> * The FlagsBase constructor is:
> bq. {{FlagsBase() { add(&help, "help", "...", false); }}}
> where "help" is a member variable -- i.e., it is allocated on the stack in
> this case.
> * {{FlagsBase()::add}} captures {{&help}}, e.g.:
> {noformat}
> flag.stringify = [t1](const FlagsBase&) -> Option<std::string> {
> return stringify(*t1);
> };}}
> {noformat}
> * The implicit copy constructor for FlagsBase is just going to copy the
> lambda above, i.e., the result of the copy constructor will have a lambda
> that points into MesosTest::CreateMasterFlags()'s stack frame, which is bad
> news.
> Not sure the right fix -- comments welcome. You could define a copy-ctor for
> FlagsBase that does something gross (basically remove the old help flag and
> define a new one that points into the target of the copy), but that seems,
> well, gross.
> Probably not a pressing-problem to fix -- AFAICS worst symptom is that we end
> up reading one byte from some random stack location when serving
> {{state.json}}, for example.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
