Re: [PATCH v2 0/2] Lock and Pointer guards

[Kees Cook <keescook@chromium.org>]

On Tue, Jun 06, 2023 at 07:50:47AM -0700, Linus Torvalds wrote:
> So you could have something like
> 
>   #define RAII(type, var, init, exit) \
>         __RAII(type, var, init, exit, __UNIQUE_ID(fn)
> 
>   #define __RAII(type, var, init, exit, exitname) \
>         void exitname(type *p) { exit } \
>         type var __attribute__((__cleanup__(exitname))) = (init)
> 
> and do all of the above with
> 
>     RAII(struct fd, fd, fdget(f), fdput(fd));

"fdput(fd)" needs to be "fdput(*p)", since otherwise "fdput(fd)" is
referencing "fd" before it has been declared.

But regardless, yes, Clang is angry about the nested function. Also,
while my toy[1] example doesn't show it, GCC may also generate code
that requires an executable stack for some instances (or at least it
did historically) that need trampolines.

[1] https://godbolt.org/z/WTjx6Gs7x

Also, more nits on naming: isn't this more accurately called Scope-based
Resource Management (SBRM) not RAII? (RAII is technically object lifetime,
and SBRM is scope entry/exit.)

-- 
Kees Cook