Re: [PATCH] eventfs: Have inodes have unique inode numbers

[Linus Torvalds]

On Sat, 27 Jan 2024 at 07:27, David Laight <David.Laight@xxxxxxxxxx> wrote:
>
> Doesn't Linux support 64bit inode numbers?
> They solve the wrap problem...

Yes, but we've historically had issues with actually exposing them.

The legacy stat() code has things like this:

        tmp.st_ino = stat->ino;
        if (sizeof(tmp.st_ino) < sizeof(stat->ino) && tmp.st_ino != stat->ino)
                return -EOVERFLOW;

so if you have really old 32-bit user space, you generally do not
actually want to have 64-bit inode numbers.

This is why "get_next_ino()" returns a strictly 32-bit only inode
number. You don't want to error out on a 'fstat()' just because you're
on a big system that has been running for a long time.

Now, 'stat64' was introduced for this reason back in 2.3.34, so back
in 1999. So any half-way modern 32-bit environment doesn't have that
issue, and maybe it's time to just sunset all the old stat() calls.

Of course, the *really* old stat has a 16-bit inode number. Search for
__old_kernel_stat to still see that. That's more of a curiosity than
anything else.

          Linus