Hello,
On Fri, Oct 22, 2021 at 03:47:33PM -0700, Martin KaFai Lau wrote:
...
> > + for_each_process_thread(g, p) {
> I am thinking if this loop can be done in bpf iter.
>
> If the bpf_local_storage_map is sleepable safe (not yet done but there is
> an earlier attempt [0]), bpf_local_storage_update() should be able to
> alloc without GFP_ATOMIC by sleepable bpf prog and this potentially
> will be useful in general for other sleepable use cases.
>
> For example, if a sleepable bpf iter prog can run in this loop (or the existing
> bpf task iter loop is as good?), the iter bpf prog can call
> bpf_task_storage_get(BPF_SK_STORAGE_GET_F_CREATE) on a sleepable
> bpf_local_storage_map.
Yeah, whatever that can walk all the existing tasks should do, and I think
the locked section can be shrunk too.
percpu_down_write(&threadgroup_rwsem);
list_add_tail(&smap->prealloc_node, &prealloc_smaps);
percpu_up_write(&threadgroup_rwsem);
// Here, it's guaranteed that all new tasks are guaranteed to
// prealloc on fork.
Iterate all tasks in whatever way and allocate if necessary;
Thanks.
--
tejun
