[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Wed, 12 Dec 2018 14:00:31 -0800
From: Alexei Starovoitov <alexei.starovoitov@...il.com>
To: Edward Cree <ecree@...arflare.com>
Cc: Alexei Starovoitov <ast@...nel.org>,
"David S . Miller" <davem@...emloft.net>, daniel@...earbox.net,
jakub.kicinski@...ronome.com, jiong.wang@...ronome.com,
netdev@...r.kernel.org, kernel-team@...com
Subject: Re: [PATCH bpf-next 4/4] bpf: add self-check logic to liveness
analysis
On Wed, Dec 12, 2018 at 08:58:33PM +0000, Edward Cree wrote:
> On 12/12/18 05:28, Alexei Starovoitov wrote:
> > introduce REG_LIVE_DONE to check the liveness propagation
> > and prepare the states for merging
> > See algorithm description in clean_live_states().
> > No difference in tests.
> >
> > Signed-off-by: Alexei Starovoitov <ast@...nel.org>
> This feels a bit magic to me, relying as it does on seeing the state
> again in is_state_visited() after the last continuation is explored
> (rather than the marking happening as soon as the last exit is
> reached).
what is "last exit" ?
Around 'process_bpf_exit' bit of code the verifier doesn't know which
state lists are not going to be changed.
> Also, why do you clean_live_states() in or after the states_equal()
> loop, rather than doing it (in just one place) before it?
true. can move the call to the beginning of is_state_visited().
> AIUI, being
> in is_state_visited() already implies that all explored states are
> DONE, whether any of them match cur_state or not.
Unfortunately not.
That's exactly the issue with liveness that I want to address
with this additional safety check.
For subprograms states_equal() checks callsite equivalence.
That's what is saving the existing liveness from producing incorrect
results.
The states in state lists of subprogs are still going to be changed.
> A different way I previously thought of was to have a refcount in
> verifier states (at the time we had a single parent rather than per-
> register parents) counting live children, that falls to 0 when all
> continuations have been walked. That was something I did in my
> bounded loops RFC.
> With something like that, we could check refcount != 0 in mark_reg_read
> and check refcount == 0 in explored states in is_state_visited. Seems
> to me like that gets you the same thing and also adds the guarantee
> that our explored_states really are fully explored.
refcnt was my initial approach, but it needs to walk parentage chain.
Also push/pop_stack needs full walk of all chains too.
That is too expensive.
What kind of refcnt you had in mind?
> Rest of series looks good, have my Ack for patches 1-3.
> (Though, maybe use a few more capital letters in your commit messages?)
Meaning capitalize first letter of the sentences?
Powered by blists - more mailing lists