On Fri, Jan 31, 2025 at 11:29:05AM -0800, Linus Torvalds wrote:
> On Fri, 31 Jan 2025 at 00:32, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> >
> > This commit lifts definition(s) of QSTR() into linux/dcache.h,
> > converts it to compound literal (all bcachefs users are fine
> > with that) and converts assorted open-coded instances to using
> > that.
>
> Looks fine to me. I do wonder if the cast to 'struct qstr' should be
> part of QSTR_INIT, so that you can use that the same way when you
> already know the length.
>
> We have code like this in fs/overlayfs/namei.c:
>
> *name = (struct qstr) QSTR_INIT(n, s - n);
>
> in bcachefs has
>
> return (struct qstr) QSTR_INIT(d.v->d_name, bch2_dirent_name_bytes(d));
>
> So both of them would seem to want to have that cast as part of the
> QSTR_INIT() thing.
>
> Or maybe we could just make QSTR() itself more powerful, and do
> something like this:
>
> #define QSTR_INIT(n, l, ...) { { { .len = l } }, .name = n }
> #define QSTR(a, ...) (struct qstr) QSTR_INIT(a , ## __VA_ARGS__, strlen(a))
>
> which allows you to write either "QSTR(str)" or "QSTR(str, len)", and
> defaults to using 'strlen()' when no length is given.
>
> Because if we have heper macros, let's make them _helpful_.
>
> Side note: you missed a few places in the core VFS code that could
> also use this new cleanup:
>
> struct qstr this = QSTR_INIT("pts", 3);
> ...
> child = d_hash_and_lookup(parent, &this);
>
> can now be just
>
> child = d_hash_and_lookup(parent, &QSTR("pts"));
>
> but sadly a few more are static initializers and can't use 'strlen()'.
There's also this in fs/fuse/inode.c
const struct qstr name = QSTR_INIT(".", 1);
and
struct qstr null_name = QSTR_INIT(NULL_FILE_NAME,
sizeof(NULL_FILE_NAME)-1);
in selinuxfs - I wasn't trying to get them all, just the infrastructure
and enough examples to demonstrate the usefulness. Once the definition
is merged, those could be switched at leisure.
