Subject: Re: [PATCH v3 net-next] bpf/verifier: track liveness for pruning
To: Daniel Borkmann <daniel@iogearbox.net>,
        Alexei Starovoitov <ast@fb.com>, <davem@davemloft.net>,
        Alexei Starovoitov <alexei.starovoitov@gmail.com>
References: <f40d3d54-c88a-7348-99ca-66db8075a8d5@solarflare.com>
 <89ff34f7-84ee-0e0a-3766-5b4d046189bf@fb.com>
 <b8a46388-74ea-81f9-df08-3b6b88042229@solarflare.com>
 <5d4e12aa-6861-a176-a8cf-a766bbca0a7a@fb.com>
 <b39b4a6f-907e-c684-b9e4-285ff9c7ca9a@solarflare.com>
 <599B4231.3080405@iogearbox.net>
CC: <netdev@vger.kernel.org>, <linux-kernel@vger.kernel.org>,
        iovisor-dev <iovisor-dev@lists.iovisor.org>
From: Edward Cree <ecree@solarflare.com>
Message-ID: <128a84cd-234d-f505-95e2-7561db974981@solarflare.com>
Date: Mon, 21 Aug 2017 21:44:40 +0100
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101
 Thunderbird/45.1.0
MIME-Version: 1.0
In-Reply-To: <599B4231.3080405@iogearbox.net>
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2011
Lines: 36

On 21/08/17 21:27, Daniel Borkmann wrote:
> On 08/21/2017 08:36 PM, Edward Cree wrote:
>> On 19/08/17 00:37, Alexei Starovoitov wrote:
> [...]
>> I'm tempted to just rip out env->varlen_map_value_access and always check
>>   the whole thing, because honestly I don't know what it was meant to do
>>   originally or how it can ever do any useful pruning.  While drastic, it
>>   does cause your test case to pass.
>
> Original intention from 484611357c19 ("bpf: allow access into map
> value arrays") was that it wouldn't potentially make pruning worse
> if PTR_TO_MAP_VALUE_ADJ was not used, meaning that we wouldn't need
> to take reg state's min_value and max_value into account for state
> checking; this was basically due to min_value / max_value is being
> adjusted/tracked on every alu/jmp ops for involved regs (e.g.
> adjust_reg_min_max_vals() and others that mangle them) even if we
> have the case that no actual dynamic map access is used throughout
> the program. To give an example on net tree, the bpf_lxc.o prog's
> section increases from 36,386 to 68,226 when env->varlen_map_value_access
> is always true, so it does have an effect. Did you do some checks
> on this on net-next?
I tested with the cilium progs and saw no change in insn count.  I
 suspect that for the normal case I already killed this optimisation
 when I did my unification patch, it was previously about ignoring
 min/max values on all regs (including scalars), whereas on net-next
 it only ignores them on map_value pointers; in practice this is
 useless because we tend to still have the offset scalar sitting in
 a register somewhere.  (Come to think of it, this may have been
 behind a large chunk of the #insn increase that my patches caused.)
Since we use umax_value in find_good_pkt_pointers() now (to check
 against MAX_PACKET_OFF and ensure our reg->range is really ok), we
 can't just stop caring about all min/max values just because we
 haven't done any variable map accesses.
I don't see a way around this.

-Ed