Re: [PATCH 2/2] vfs: support statx(..., NULL, AT_EMPTY_PATH, ...)

[Christian Brauner <brauner@kernel.org>]

On Wed, Jul 03, 2024 at 08:14:15PM GMT, Christian Brauner wrote:
> On Wed, Jul 03, 2024 at 10:54:53AM GMT, Linus Torvalds wrote:
> > On Wed, 3 Jul 2024 at 10:40, Linus Torvalds
> > <torvalds@linux-foundation.org> wrote:
> > >
> > > Oh wow. Shows just *how* long ago that was - and how long ago I looked
> > > at 32-bit code. Because clearly, I was wrong.
> > 
> > Ok, so clearly any *new* 32-bit architecture should use 'struct statx'
> > as 'struct stat', and at least avoid the conversion pain.
> > 
> > Of course, if using <asm-generic/stat.h> like loongarch does, that is
> > very much not what happens. You get those old models with just 'long'.
> > 
> > So any architecture that didn't do that 'stat == statx' and has
> > binaries with old stat models should just continue to have them.
> > 
> > It's not like we can get rid of the kernel side code for that all _anyway_.
> 
> Fwiw, the original motivation for that whole "let's do NULL with
> AT_EMPTY_PATH" (somewhat independent from the generic use of it) that
> somehow morphed into this discussion was that the Chrome Sandbox has
> rewrites fstatat() system calls to fstat() via SECCOMP_RET_TRAP:
> 
>   if (args.nr == __NR_fstatat_default) {
>     if (*reinterpret_cast<const char*>(args.args[1]) == '\0' &&
>         args.args[3] == static_cast<uint64_t>(AT_EMPTY_PATH)) {
>       return syscall(__NR_fstat_default, static_cast<int>(args.args[0]),
>                      reinterpret_cast<default_stat_struct*>(args.args[2]));
>     }
> 
> while also disabling statx() completely because they can't (easily)
> rewrite it and don't want to allow it unless we have NULL for
> AT_EMPTY_PATH (which we'll have soon ofc).
> 
> In any case in [1] I proposed they add back fstat()/fstatat64() which
> should get that problem solved because they can rewrite that thing.
> 
> In any case, which one of these does a new architecture have to add for
> reasonable backward compatibility:

Going by riscv added in 2017 it would be:

newstat()
newlstat()
newfstat()
newfstatat()
statx()

> 
> fstat()
> fstat64()
> fstatat64()
> 
> lstat()
> lstat64()
> 
> stat()
> stat64()
> statx()
> 
> newstat()
> newlstat()
> newfstat()
> newfstatat()
> 
> Because really that's a complete mess and we have all sorts of overflow
> issues and odd failures in the varioius variants. And the userspace
> ifdefery in libcs is just as bad if not very much worse.
> 
> [1]: https://lore.kernel.org/lkml/20240226-altmodisch-gedeutet-91c5ba2f6071@brauner