Re: [PATCH bpf-next] bpf: move rcu lock management out of BPF_PROG_RUN routines

[sdf@google.com]

On 04/13, Martin KaFai Lau wrote:
> On Wed, Apr 13, 2022 at 12:52:53PM -0700, Andrii Nakryiko wrote:
> > On Wed, Apr 13, 2022 at 12:39 PM <sdf@google.com> wrote:
> > >
> > > On 04/13, Andrii Nakryiko wrote:
> > > > On Wed, Apr 13, 2022 at 11:33 AM Stanislav Fomichev <sdf@google.com>
> > > > wrote:
> > > > >
> > > > > Commit 7d08c2c91171 ("bpf: Refactor BPF_PROG_RUN_ARRAY family of  
> macros
> > > > > into functions") switched a bunch of BPF_PROG_RUN macros to inline
> > > > > routines. This changed the semantic a bit. Due to arguments  
> expansion
> > > > > of macros, it used to be:
> > > > >
> > > > >         rcu_read_lock();
> > > > >         array = rcu_dereference(cgrp->bpf.effective[atype]);
> > > > >         ...
> > > > >
> > > > > Now, with with inline routines, we have:
> > > > >         array_rcu = rcu_dereference(cgrp->bpf.effective[atype]);
> > > > >         /* array_rcu can be kfree'd here */
> > > > >         rcu_read_lock();
> > > > >         array = rcu_dereference(array_rcu);
> > > > >
> > >
> > > > So subtle difference, wow...
> > >
> > > > But this open-coding of rcu_read_lock() seems very unfortunate as
> > > > well. Would making BPF_PROG_RUN_ARRAY back to a macro which only  
> does
> > > > rcu lock/unlock and grabs effective array and then calls static  
> inline
> > > > function be a viable solution?
> > >
> > > > #define BPF_PROG_RUN_ARRAY_CG_FLAGS(array_rcu, ctx, run_prog,  
> ret_flags) \
> > > >    ({
> > > >        int ret;
> > >
> > > >        rcu_read_lock();
> > > >        ret =
> > > > __BPF_PROG_RUN_ARRAY_CG_FLAGS(rcu_dereference(array_rcu), ....);
> > > >        rcu_read_unlock();
> > > >        ret;
> > > >    })
> > >
> > >
> > > > where __BPF_PROG_RUN_ARRAY_CG_FLAGS is what
> > > > BPF_PROG_RUN_ARRAY_CG_FLAGS is today but with __rcu annotation  
> dropped
> > > > (and no internal rcu stuff)?
> > >
> > > Yeah, that should work. But why do you think it's better to hide them?
> > > I find those automatic rcu locks deep in the call stack a bit obscure
> > > (when reasoning about sleepable vs non-sleepable contexts/bpf).
> > >
> > > I, as the caller, know that the effective array is rcu-managed (it
> > > has __rcu annotation) and it seems natural for me to grab rcu lock
> > > while work with it; I might grab it for some other things like cgroup
> > > anyway.
> >
> > If you think that having this more explicitly is better, I'm fine with
> > that as well. I thought a simpler invocation pattern would be good,
> > given we call bpf_prog_run_array variants in quite a lot of places. So
> > count me indifferent. I'm curious what others think.

> Would it work if the bpf_prog_run_array_cg() directly takes the
> 'struct cgroup *cgrp' argument instead of the array ?
> bpf_prog_run_array_cg() should know what protection is needed
> to get member from the cgrp ptr.  The sk call path should be able
> to provide a cgrp ptr.  For current cgrp, pass NULL as the cgrp
> pointer and then current will be used in bpf_prog_run_array_cg().
> A rcu_read_lock() is needed anyway to get the current's cgrp
> and can be done together in bpf_prog_run_array_cg().

> That there are only two remaining bpf_prog_run_array() usages
> from lirc and bpf_trace which are not too bad to have them
> directly do rcu_read_lock on their own struct ?

 From Andrii's original commit message:

     I think BPF_PROG_RUN_ARRAY_CG would benefit from further refactoring to  
accept
     struct cgroup and enum bpf_attach_type instead of bpf_prog_array,  
fetching
     cgrp->bpf.effective[type] and RCU-dereferencing it internally. But that
     required including include/linux/cgroup-defs.h, which I wasn't sure is  
ok with
     everyone.

I guess including cgroup-defs.h/bpf-cgroup-defs.h into bpf.h might still
be somewhat problematic?

But even if we pass the cgroup pointer, I'm assuming that this cgroup  
pointer
is still rcu-managed, right? So the callers still have to rcu-lock.
However, in most places we don't care and do "cgrp =  
sock_cgroup_ptr(&sk->sk_cgrp_data);"
but seems like it depends on the fact that sockets can't (yet?)
change their cgroup association and it's fine to not rcu-lock that
cgroup. Seems fragile, but ok. It always stumbles me when I see:

cgrp = sock_cgroup_ptr(&sk->sk_cgrp_data);
bpf_prog_run_array_cg_flags(cgrp.bpf->effective[atype], ...)

But then, with current, it becomes:

rcu_read_lock();
cgrp = task_dfl_cgroup(current);
bpf_prog_run_array_cg_flags(cgrp.bpf->effective[atype], ...)
rcu_read_unlock();

Idk, I might be overthinking it. I'll try to see if including
bpf-cgroup-defs.h and passing cgroup_bpf is workable.