Correctly initialize 'installed_handlers'

[fcharlie]

Because static variables are not initialized properly,
temporary files may not be deleted when receive-pack receives a signal.

[fcharlie]

/submit

[gitgitgadget]

Submitted as pull.630.git.1588921514146.gitgitgadget@gmail.com

[gitgitgadget]

On the Git mailing list, Carlo Marcelo Arenas Belón wrote (reply to this):

On Fri, May 08, 2020 at 07:05:13AM +0000, Force Charlie via GitGitGadget wrote:
> From: Force Charlie <charlieio@outlook.com>
> 
> Because static variables are not initialized properly,

what do you mean by "properly"?, all static variables are set to 0;
that is a warranty of the language (all the way to K&R) and any C compiler
should enforce that as part of the standard.

> temporary files may not be deleted when receive-pack receives a signal.

the way this is handled would seem to indicate otherwise

if (!installed_handlers) {
                atexit(remove_tmp_objdir);
                sigchain_push_common(remove_tmp_objdir_on_signal);
                installed_handlers++;
}

there is no explicit locking and so there might be a thread race
condition, but the code below wouldn't make a difference in that
case.

could you elaborate more on how to reproduce the problem?, I suspect
that if there was a problem then suppressing whatever signal that
was triggered before sigchain_push_common might help, but the window
is too short to be a likely issue.

Carlo

[fcharlie]

This problem should be caused by other reasons so close this pr

[gitgitgadget]

On the Git mailing list, Junio C Hamano wrote (reply to this):

Carlo Marcelo Arenas Belón <carenas@gmail.com> writes:

> the way this is handled would seem to indicate otherwise
>
> if (!installed_handlers) {
>                 atexit(remove_tmp_objdir);
>                 sigchain_push_common(remove_tmp_objdir_on_signal);
>                 installed_handlers++;
> }

It is a curious piece of code.  

The "prepare a file-scope static and do something and increment it
when it is 0" pattern expects the function to be called many times
and do the guarded thing only once.  However, there is this code:

	if (the_tmp_objdir)
		BUG(...);

before we do anything else, and then before that "arrange to clean
up, but do so just once" block, there is

	the_tmp_objdir = t;

where t is the pointer to a "struct tmp_objdir" instance.  So one
part of the function expects to be called at most once, while
another part is prepared to be called more than once.

Almost all of this function is attributed to 2564d994 (tmp-objdir:
introduce API for temporary object directories, 2016-10-03), so
let's see if Peff remembers anything about this curiosity.

Thanks.

[gitgitgadget]

On the Git mailing list, Jeff King wrote (reply to this):

On Fri, May 08, 2020 at 08:36:36AM -0700, Junio C Hamano wrote:

> Carlo Marcelo Arenas Belón <carenas@gmail.com> writes:
> 
> > the way this is handled would seem to indicate otherwise
> >
> > if (!installed_handlers) {
> >                 atexit(remove_tmp_objdir);
> >                 sigchain_push_common(remove_tmp_objdir_on_signal);
> >                 installed_handlers++;
> > }
> 
> It is a curious piece of code.  
> 
> The "prepare a file-scope static and do something and increment it
> when it is 0" pattern expects the function to be called many times
> and do the guarded thing only once.  However, there is this code:
> 
> 	if (the_tmp_objdir)
> 		BUG(...);
> 
> before we do anything else, and then before that "arrange to clean
> up, but do so just once" block, there is
> 
> 	the_tmp_objdir = t;
> 
> where t is the pointer to a "struct tmp_objdir" instance.  So one
> part of the function expects to be called at most once, while
> another part is prepared to be called more than once.
> 
> Almost all of this function is attributed to 2564d994 (tmp-objdir:
> introduce API for temporary object directories, 2016-10-03), so
> let's see if Peff remembers anything about this curiosity.

There is "only once per program" and "only one at a time". When the
tmp_objdir is destroyed (either directly or via tmp_objdir_migrate), we
set the_tmp_objdir back to NULL, and you are free to then create another
one.

In practice there's only one caller (receive-pack) and it only ever uses
one tmp_objdir per program, so it's mostly academic. But tmp-objdir.c
was written to be as reusable and least-surprising as possible. I would
have avoided the "one at a time" rule if I could, but the semantics are
unclear (if you have two active, which one should object-writes go to?).

The atexit and signal handlers could be removed when there's no
tmp_objdir active, but there's no easy way to remove them (there's
nothing portable at all for atexit, and for sigchain we don't know if
somebody else has pushed in the meantime).

-Peff