[ovs-dev] [PATCH v7] datapath: Add Stateless TCP Tunneling protocol.

[Jesse Gross]

On Wed, Apr 22, 2015 at 8:55 PM, Pravin Shelar <pshelar at nicira.com> wrote:
> On Wed, Apr 22, 2015 at 8:50 PM, Jesse Gross <jesse at nicira.com> wrote:
>> On Wed, Apr 22, 2015 at 8:37 PM, Pravin Shelar <pshelar at nicira.com> wrote:
>>> On Wed, Apr 22, 2015 at 8:29 PM, Jesse Gross <jesse at nicira.com> wrote:
>>>> On Wed, Apr 22, 2015 at 8:22 PM, Pravin Shelar <pshelar at nicira.com> wrote:
>>>>> On Wed, Apr 22, 2015 at 7:48 PM, Jesse Gross <jesse at nicira.com> wrote:
>>>>>> On Wed, Apr 22, 2015 at 3:38 PM, Pravin Shelar <pshelar at nicira.com> wrote:
>>>>>>> On Wed, Apr 22, 2015 at 3:24 PM, Jesse Gross <jesse at nicira.com> wrote:
>>>>>>>> On Wed, Apr 22, 2015 at 3:19 PM, Pravin Shelar <pshelar at nicira.com> wrote:
>>>>>>>>> On Wed, Apr 22, 2015 at 11:31 AM, Jesse Gross <jesse at nicira.com> wrote:
>>>>>>>>>> On Wed, Apr 22, 2015 at 11:09 AM, Pravin Shelar <pshelar at nicira.com> wrote:
>>>>>>>>>>> On Wed, Apr 22, 2015 at 10:52 AM, Jesse Gross <jesse at nicira.com> wrote:
>>>>>>>>>>>> On Wed, Apr 22, 2015 at 10:03 AM, Pravin Shelar <pshelar at nicira.com> wrote:
>>>>>>>>>>>>> On Mon, Apr 20, 2015 at 9:54 PM, Jesse Gross <jesse at nicira.com> wrote:
>>>>>>>>>>>>>> On Mon, Apr 20, 2015 at 7:33 PM, Pravin Shelar <pshelar at nicira.com> wrote:
>>>>>>>>>>>>>>> On Mon, Apr 20, 2015 at 5:56 PM, Jesse Gross <jesse at nicira.com> wrote:
>>>>>>>>>>>>>>>> On Mon, Apr 20, 2015 at 3:54 PM, Pravin Shelar <pshelar at nicira.com> wrote:
>>>>>>>>>>>>>>>>> On Mon, Apr 20, 2015 at 3:36 PM, Jesse Gross <jesse at nicira.com> wrote:
>>>>>>>>>>>>>>>>>> On Mon, Apr 20, 2015 at 1:29 PM, Pravin B Shelar <pshelar at nicira.com> wrote:
>>>>>>>>>>>>>>>>>>> diff --git a/datapath/linux/compat/stt.c b/datapath/linux/compat/stt.c
>>>>>>>>>>>>>>>>>>> new file mode 100644
>>>>>>>>>>>>>>>>>>> index 0000000..209bf1a
>>>>>>>>>>>>>>>>>>> --- /dev/null
>>>>>>>>>>>>>>>>>>> +++ b/datapath/linux/compat/stt.c
>>>>>>>>>>>>>>>>>>> +static void update_headers(struct sk_buff *skb, bool head,
>>>>>>>>>>>>>>>>>>> +                              unsigned int l4_offset, unsigned int hdr_len,
>>>>>>>>>>>>>>>>>>> +                              bool ipv4, u32 tcp_seq)
>>>>>>>>>>>>>>>>>> [...]
>>>>>>>>>>>>>>>>>>> +       skb->truesize = SKB_TRUESIZE(skb_end_offset(skb)) + skb->data_len;
>>>>>>>>>>>>>>>>>>> +}
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>> I wonder if there are any possible edge cases with resetting truesize
>>>>>>>>>>>>>>>>>> where the packet is still in someone's transmit queue (such as if we
>>>>>>>>>>>>>>>>>> are looping back packet). Do we need to orphan it first?
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> ok, I will orphan it in update_headers.
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> Just to clarify - I was mostly just thinking aloud on orphaning it.
>>>>>>>>>>>>>>>> I'm not totally sure if that is the right thing to do or if this is
>>>>>>>>>>>>>>>> the right place to do it. I'm not sure what the conceptual
>>>>>>>>>>>>>>>> justification would be for it and it could potentially result in the
>>>>>>>>>>>>>>>> sender's buffers not being properly limited. Perhaps not resetting the
>>>>>>>>>>>>>>>> truesize is the right thing too...
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> I have seen warning msg if we do no keep truesize update along with
>>>>>>>>>>>>>>> changes to skb.
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Hmm, interesting, what is the warning? I don't think that I have seen
>>>>>>>>>>>>>> that before.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Actually skb_try_coalesce() is updating it correctly. so there no need
>>>>>>>>>>>>> to change truesize anymore. I will update patch accordingly.
>>>>>>>>>>>>
>>>>>>>>>>>> That's much nicer. I also checked and other receive side code (like
>>>>>>>>>>>> TCP input) doesn't worry about the case where a local sender may still
>>>>>>>>>>>> be accounting for the packet since any type of loopback device does
>>>>>>>>>>>> call skb_orphan() in some form.
>>>>>>>>>>>>
>>>>>>>>>>>> I hate to bring this up but what about on transmit? In cases where we
>>>>>>>>>>>> merge or split skbs (skb_try_coalesce() and normalize_frag_list()
>>>>>>>>>>>> respectively) we do track the truesize for correctly for the result
>>>>>>>>>>>> but the individual pieces might not have the right destructors or
>>>>>>>>>>>> might not have their truesize updated for the destructor they do have.
>>>>>>>>>>>
>>>>>>>>>>> How about about we update merged skb stats (len, data_len, truesize)
>>>>>>>>>>> according to *delta_truesize we get from skb_try_coalesce() and then
>>>>>>>>>>> free the skb?
>>>>>>>>>>
>>>>>>>>>> I think that would work for the skb_try_coalesce() case (although I
>>>>>>>>>> would only worry about truesize, not the lengths). For
>>>>>>>>>> normalize_frag_list() I think we would either have to add a destructor
>>>>>>>>>> or not update truesize. I'm not sure that the condition where we have
>>>>>>>>>> frag_lists and a destructor actually ever happens in practice though.
>>>>>>>>>
>>>>>>>>> I am not sure why do we need a destructor for normalize_frag_list
>>>>>>>>> changes. Even with the changes in the function truesize is consistent
>>>>>>>>> for given skb memory usage.
>>>>>>>>
>>>>>>>> If you have a packet with a frag_list, all of the memory for the
>>>>>>>> individual elements will be accounted for in the truesize in the top
>>>>>>>> level skb. This skb could also be accounted to some socket and have a
>>>>>>>> destructor. When we break apart the list, we remove the truesize from
>>>>>>>> the head because the memory is now part of individual packets.
>>>>>>>> However, the destructor is presumably only on the head and so only its
>>>>>>>> memory will be removed from the socket accounting when it is freed but
>>>>>>>> not each of the other skbs that came from it.
>>>>>>>
>>>>>>> ok. In that case we can not have our own destructor since there is one
>>>>>>> already (I am not sure if we can use skb->cb to restore original). not
>>>>>>> changing true size can complicate skb coalesce, since it does update
>>>>>>> truesize. Easy option would be orphan skb if we are going to coalesce
>>>>>>> its fragments.
>>>>>>
>>>>>> I'm not sure that we need our own destructor. What do you think about
>>>>>> just replicating the original destructor onto each of the newly
>>>>>> generated skbs?
>>>>>>
>>>>> In that case we assume there is no state associated with the skb, that
>>>>> might not be always true.
>>>>
>>>> What state do you mean? If you mean a destructor on the individual
>>>> skbs, I think that is already true because only the top level
>>>> destructor will get called when the original skb is freed.
>>>
>>> If destructor is replicated on the individual skbs then it will be
>>> called for each of those skbs when it is freed.
>>
>> Right; if we've correctly apportioned truesize among the individual
>> skbs isn't that the goal?
>
> It works for destructors which only care for skb-truesize. But
> destructor callback also takes argument, we also need to replicate
> that and we do not have enough information to do it.

I'm not quite sure what you mean by an argument (I don't see anything
really in either tcp_wfree or sock_wfree) but I agree that this could
be dangerous. There's no guarantee that total truesize on the head
packet accurately accounts for all of the individual frag_list members
(although for the main users like IP fragment reassembly this appears
to be the case).

I'm a little nervous about orphaning the packets in this case since it
can potentially affect socket accounting and from a high level,
there's no logical reason to do so (i.e. we're not crossing a boundary
or something like that).

I guess that in many cases (particularly on the transmit side), moving
the truesize around in normalize_frag_list() and skb_try_coalesce()
will effectively cancel each other out if we can merge things together
(although I guess that depends on the source of the frag_list, IP
fragment reassembly first tries to coalesce and then falls back on
creating frag_lists, in which case us trying to merge things is a
waste of time). It would be nice if we could relegate this to a corner
case and just linearize in those situations. Or better yet, more
deeply understand the conditions where we get the frag_list on
transmit and see if it's really worth trying to coalesce them at all.
grepping drivers/net I don't see a lot of frag_list creation. Even on
receive, I'm not totally sure that breaking up frag_lists makes sense
if skb_segment() can handle it now...