On Mon, Oct 7, 2024 at 4:50 PM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
>
> On Mon, Oct 07, 2024 at 10:52:37AM +0200, Uros Bizjak wrote:
> > Use atomic64_inc_return(&ref) instead of atomic64_add_return(1, &ref)
> > to use optimized implementation and ease register pressure around
> > the primitive for targets that implement optimized variant.
> >
> > Signed-off-by: Uros Bizjak <ubizjak@xxxxxxxxx>
> > Cc: Alexander Viro <viro@xxxxxxxxxxxxxxxxxx>
> > Cc: Christian Brauner <brauner@xxxxxxxxxx>
> > Cc: Jan Kara <jack@xxxxxxx>
> > ---
> > fs/namespace.c | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/fs/namespace.c b/fs/namespace.c
> > index 93c377816d75..9a3c251d033d 100644
> > --- a/fs/namespace.c
> > +++ b/fs/namespace.c
> > @@ -3901,7 +3901,7 @@ static struct mnt_namespace *alloc_mnt_ns(struct user_namespace *user_ns, bool a
> > }
> > new_ns->ns.ops = &mntns_operations;
> > if (!anon)
> > - new_ns->seq = atomic64_add_return(1, &mnt_ns_seq);
> > + new_ns->seq = atomic64_inc_return(&mnt_ns_seq);
>
> On which load do you see that path hot enough for the change to
> make any difference???
It is not performance, but code size improvement, as stated in the
commit message.
The difference on x86_32 (that implements atomic64_inc_return()) is:
eeb: b8 01 00 00 00 mov $0x1,%eax
ef0: 31 d2 xor %edx,%edx
ef2: b9 20 00 00 00 mov $0x20,%ecx
ef3: R_386_32 .data
ef7: e8 fc ff ff ff call ef8 <alloc_mnt_ns+0xd0>
ef8: R_386_PC32 atomic64_add_return_cx8
efc: 89 46 20 mov %eax,0x20(%esi)
eff: 89 56 24 mov %edx,0x24(%esi)
vs:
eeb: be 20 00 00 00 mov $0x20,%esi
eec: R_386_32 .data
ef0: e8 fc ff ff ff call ef1 <alloc_mnt_ns+0xc9>
ef1: R_386_PC32 atomic64_inc_return_cx8
ef5: 89 43 20 mov %eax,0x20(%ebx)
ef8: 89 53 24 mov %edx,0x24(%ebx)
Uros.
