On Mon, Aug 12, 2024 at 07:44:17PM +0200, Mickaël Salaün wrote: No opinion about the core idea, I'll note though that this conflicts with a patch to move f_owner out of the struct: https://lore.kernel.org/linux-fsdevel/20240809-koriander-biobauer-6237cbc106f3@brauner/ Presumably nothing which can't get sorted out with some shoveling. I do have actionable remark concerning creds though: both get_cred and put_cred are slow. Sorting that out is on my todo list. In the meantime adding more calls can be avoided: > diff --git a/fs/file_table.c b/fs/file_table.c > index 4f03beed4737..d28b76aef4f3 100644 > --- a/fs/file_table.c > +++ b/fs/file_table.c > @@ -66,6 +66,7 @@ static inline void file_free(struct file *f) > if (likely(!(f->f_mode & FMODE_NOACCOUNT))) > percpu_counter_dec(&nr_files); > put_cred(f->f_cred); > + put_cred(f->f_owner.cred); if (likely(f->f_cred == f->f_owner.cred)) { put_cred_many(f->f_cred, 2); } else { put_cred(f->f_cred); put_cred(f->f_owner.cred); } > if (unlikely(f->f_mode & FMODE_BACKING)) { > path_put(backing_file_user_path(f)); > kfree(backing_file(f)); > @@ -149,9 +150,11 @@ static int init_file(struct file *f, int flags, const struct cred *cred) > int error; > > f->f_cred = get_cred(cred); > + f->f_owner.cred = get_cred(cred); f->f_cred = f->f_owner.cred = get_cred_many(cred, 2); > error = security_file_alloc(f); > if (unlikely(error)) { > put_cred(f->f_cred); > + put_cred(f->f_owner.cred); put_cred_many(cred, 2); > return error; > }
