On Sun, Sep 29, 2024 at 7:08 PM Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
>
> On Sun, Sep 29, 2024 at 10:04:00AM -0700, Alexei Starovoitov wrote:
> > On Fri, Sep 27, 2024 at 11:41 AM Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
> > > +static void *kmem_cache_iter_seq_start(struct seq_file *seq, loff_t *pos)
> > > +{
> > > + loff_t cnt = 0;
> > > + struct kmem_cache *s = NULL;
> > > +
> > > + mutex_lock(&slab_mutex);
> >
> > It would be better to find a way to iterate slabs without holding
> > the mutex for the duration of the loop.
> > Maybe use refcnt to hold the kmem_cache while bpf prog is looking at it?
>
> Do you mean that you want to not hold slab_mutex while BPF program is
> running?
yes.
> Maybe we can allocates an arary of pointers to the slab cahe
> (with refcounts) at the beginning and iterate them instead. And call
> kmem_cache_destroy() for each entry at the end. Is it ok to you?
That doesn't sound efficient.
Just grab a refcnt on kmem_cache before running the prog ?
Drop refcnt, and grab a mutex again to do a next step.
kmem_cache_iter_seq_next() will be running with mutex held, of course.
