Re: [bpf PATCH v2 2/6] bpf: sockmap only allow ESTABLISHED sock state

[John Fastabend <john.fastabend@gmail.com>]

On 06/18/2018 02:17 PM, Martin KaFai Lau wrote:
> On Mon, Jun 18, 2018 at 07:50:19AM -0700, John Fastabend wrote:
>> On 06/14/2018 05:18 PM, Martin KaFai Lau wrote:
>>> On Thu, Jun 14, 2018 at 09:44:52AM -0700, John Fastabend wrote:
>>>> Per the note in the TLS ULP (which is actually a generic statement
>>>> regarding ULPs)
>>>>
>>>>  /* The TLS ulp is currently supported only for TCP sockets
>>>>   * in ESTABLISHED state.
>>>>   * Supporting sockets in LISTEN state will require us
>>>>   * to modify the accept implementation to clone rather then
>>>>   * share the ulp context.
>>>>   */
>>> Can you explain how that apply to bpf_tcp ulp?
>>>
>>> My understanding is the "ulp context" referred in TLS ulp is
>>> the tls_context stored in icsk_ulp_data but I don't see bpf_tcp's
>>> ulp is using icsk_ulp_data.
>>>
>>> Others LGTM.
>>>
>>
>> So I think you are right we could probably allow it
>> here but I am thinking I'll leave the check for now
>> anyways for a couple reasons. First, we will shortly
>> add support to allow ULP types to coexist. At the moment
>> the two ULP types can not coexist. When this happens it
>> looks like we will need to restrict to only ESTABLISHED
>> types or somehow make all ULPs work in all states.
>>
>> Second, I don't have any use cases (nor can I think of
>> any) for the sock{map|hash} ULP to be running on a non
>> ESTABLISHED socket. Its not clear to me that having the
>> sendmsg/sendpage hooks for a LISTEN socket makes sense.
>> I would rather restrict it now and if we add something
>> later where it makes sense to run on non-ESTABLISHED
>> socks we can remove the check.
> Make sense if there is no use case.  It will be helpful if the commit log
> is updated accordingly.  Thanks!
> 
> Acked-by: Martin KaFai Lau <kafai@fb.com>
> 

The fall-out of this patch got a bit ugly. It doesn't
make much sense to me to allow transitioning into
ESTABLISH state (via tcp_disconnect) and not allow adding
the socks up front. But the fix via unhash callback, subsequent
patch, ended up causing a few issues. First to avoid racing
with transitions through update logic we had to use
sock_lock(sk) in the update handler. Which means we
can't use the normal ./kernel/bpf/syscall.c map update
logic and had to special case it so that preempt and
rcu were not used until after the lock was taken because
sock_lock can sleep. Then after running over night I noticed
a couple WARNINGS related to sk_forward_alloc not being
zero'd correctly on sock teardown. The issue is unhash
doesn't have sock_lock either and can be done while a
sendmsg/sendpage are running resulting in incorrectly
removing scatterlists. :(

All in all the "fix" got ugly so lets stay with the minimal
required set and allow non-established socks. It shouldn't
hurt anything even if from a use case perspective its a bit
useless. Then in bpf-next when we allow ULPs to coexist we
need to have some fix to handle this. But I would rather do
that in *next branches instead of bpf/net branches.

Thanks for all the useful feedback. I'm going to send a
set of three patches now to address the specific issues,
(a) IPV6 socks, (b) bucket lock missing and (c) uref release
missing.

Thanks,
John