On Sun, 9 Nov 2025 at 21:17, Al Viro <[email protected]> wrote:
>
> That's more about weird callers of getname(), but...
>
> #ifdef CONFIG_SYSFS_SYSCALL
> static int fs_index(const char __user * __name)
> {
> struct file_system_type * tmp;
> struct filename *name;
> int err, index;
>
> name = getname(__name);
Yeah, ok, this is certainly a somewhat unusual pattern in that "name"
here is not a pathname, but at the same time I can't fault this code
for using a convenient function for "allocate and copy a string from
user space".
> Yes, really - echo $((`sed -ne "/.\<$1$/=" </proc/filesystems` - 1))
> apparently does deserve a syscall. Multiplexor, as well (other
> subfunctions are about as hard to implement in userland)...
I think those are all "crazy legacy from back in the dark ages when we
thought iBCS2 was a goal".
I doubt anybody uses that 'sysfs()' system call, and it's behind the
SYSFS_SYSCALL config variable that was finally made "default n" this
year, but has actually had a help-message that called it obsolete
since at least 2014.
The code predates not just git, but the bitkeeper history too - and
we've long since removed all the actual iBCS2 code (see for example
commit 612a95b4e053: "x86: remove iBCS support", which removed some
binfmt left-overs - back in 2008).
> IMO things like "xfs" or "ceph" don't look like pathnames - if
> anything, we ought to use copy_mount_string() for consistency with
> mount(2)...
Oh, absolutely not.
But that code certainly could just do strndup_user(). That's the
normal thing for "get a string from user space" these days, but it
didn't historically exist..
That said, I think that code will just get removed, so it's not even
worth worrying about. I don't think anybody even *noticed* that we
made it "default n" after all these years.
Linus
