Re: Dynamic modules: MODULE_HANDLE_SIGNALS etc.

[Daniel Colascione <dancol@dancol.org>]

[-- Attachment #1: Type: text/plain, Size: 4445 bytes --]

On 12/21/2015 10:15 AM, Paul Eggert wrote:
> Eli Zaretskii wrote:
> 
>> Are you
>> now saying something different from what you said back then, i.e. that
>> we cannot rely on any function/macro from lisp.h to be signal-safe?
> 
> Yes and no. As I understood it, that old conversation was about
> functions that explicitly signal or throw, and it's safe to assume that
> EQ, NILP, etc. won't do that. The new conversation is about running out
> of memory, which is a different form of non-local exit. 

IMHO, we should treat OOM exactly like other sorts of error. It's
dangerous to make some functions infallible.

> There may be
> other forms, such as operating-system signals (I don't recall exactly).

OS signals should go through the usual Emacs event loop, right?

>> If so, we should add the necessary protection, in the form of calls to
>> MODULE_FUNCTION_BEGIN, to emacs-module.c functions that until now
>> relied on those lisp.h functions/macros to be safe.
> 
> This wouldn't suffice for these other non-local exits, I think; at
> least, not as currently constructed.
> 
>> AFAIK, proper C++ exception handling
>> requires non-trivial amounts of stack space that is not available when
>> there's stack overflow, where you have at most a single guard page to
>> work with.
> 
> There should be workarounds for that. Surely the C++ community has run
> into this problem and has solutions. If we want to support C++ modules,
> we need to employ them.

The standard requires runtimes reserve enough memory to throw
std::bad_alloc. All Emacs has to do is make sure control flow reaches
the C++ level.

>> I think there is some misunderstanding here, or some confusion,
>> perhaps mine: emacs-module.c is not supposed to deal with any C++
>> exceptions.  C++ exceptions are supposed to be caught at the C++
>> level, below emacs-module.c, and handled there.  An exception that
>> isn't caught will be recorded and will cause all the subsequent calls
>> to Lisp or to emacs-module.c function to fail,
> 
> Why bother? If C++ exceptions are supposed to be caught by the C++
> module in question, why does Emacs need to worry about C++ exceptions
> that are not caught?

IMHO, it should be the module's job to make sure C++ exceptions don't
propagate through Emacs stack frames. Emacs shouldn't know or care about
C++ exceptions in any way.

You previously wrote that,

> If emacs-module.c or the Emacs exception-handling mechanism really
> needs to be rewritten in C++ in order to catch C++ exceptions nicely,
> then somebody with C++ expertise should do that.

I think there's a fundamental misunderstanding here. Emacs signals and
C++ exceptions are completely separate mechanisms. There's no reason
Emacs has to care about C++ at all. When we say that we want C++
exceptions to work, the Emacs-relevant meaning is that Emacs should
return always use conventional local returns, not longjmp, so that the
stack unwinding facilities of other languages (e.g., C++) work properly.

>> What emacs-module.c does with non-local exits of _any_ kind is record
>> the first occurrence of such an exit, and silently return to the
>> caller, thus allowing the C++ objects on the stack to be destroyed
>> normally.  IOW, it defers the exit until internal--module-call is
>> about to return.  What problems do you see with that which cause you
>> to think it's error-prone, let alone dysfunctional?
> 
> It uses a different model at the C level from what one sees in Elisp, or
> from what one normally sees in C for that matter.  I don't feel that I
> will really understand the model unless I see some actual modules that
> do function calls and exception handling; but it's hard to believe that
> a model that does silent returns and that defers returns until later and
> that records some returns but not others will be problem-free. Wouldn't
> it be simpler to have a module invoke analogs of 'condition-case' and/or
> 'catch', and to dispense with the funcall_exit stuff?
> 

Both the Python and Java extension APIs implement high-level exceptions
with low-level state exactly the way we're talking about here, and the
result has been generally usable.

Of course it's possible for module authors (or Emacs developers) to
introduce bugs with either model, but using an explicit error indication
results in bugs that are easier to notice and easier to fix.


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 819 bytes --]