Re: Network Stack Budgeting

["Tomas Pilar (tpilar)" <tpilar@solarflare.com>]

On 29/01/2019 03:20, Fu, Siyuan wrote:
> Hi, Tomas
>
>> -----Original Message-----
>> From: Tomas Pilar (tpilar) [mailto:tpilar@solarflare.com]
>> Sent: Monday, January 28, 2019 7:24 PM
>> To: Fu, Siyuan <siyuan.fu@intel.com>; Wu, Jiaxin <jiaxin.wu@intel.com>; Laszlo
>> Ersek <lersek@redhat.com>; edk2-devel@lists.01.org
>> Cc: Ye, Ting <ting.ye@intel.com>
>> Subject: Re: [edk2] Network Stack Budgeting
>>
>> Hi Siyuan,
>>
>> I had to do some tweaks due to the fact that people in the wild want to PXE
>> boot whole WDS installations which are upwards of 300MB.
>>
>> I have an internal set of queues where the NIC receives into. Snp.Receive()
>> will get a packet from this store. I had to implement a separate queue for
>> unicast/multicast/broadcast traffic so that I could reprioritise and deliver
>> unicast traffic while under ARP spam. I tied internal poll to receive and
>> transmit, so Snp.Receive() will poll the NIC and then dequeue a packet while
>> Snp.Transmit() will transmit the packet and then poll the NIC to get a
>> completion. I also implemented a busy poll for 100us or until the next packet
>> comes in when the NIC is polled.
> I see some problems here. Both Snp.Receive() and Snp.Transmit() should be non-blocking
> interface in current design and UEFI spec, but it's not true in your driver.
> For Receive(), it could poll the NIC only once and return either a received packet
> or an error, it should not do any blocking poll to wait for a packet.
On some platforms (tftp clients), if the Snp.Receive() returns EFI_NOT_READY they do not poll the NIC until the next MnpSystemPoll(). Hence the delay before returning EFI_NOT_READY in case the packet is still incoming.
> For Transmit(), it just need to put the packet into the TX queue and return directly,
> without checking if the TX is completed. 
> Instead of that, Snp.GetStatus() should be
> used to check if the packet has really been sent out.
I've seen quite a few platforms where GetStatus was only called by the MnpSystemPoll() and the application would not call it. This means our hardware rings are stuffed with completions and we stop being able to transmit. Yes, you could say that we should solve the application, but unfortunately I can't do that.
>
> May I know how did you control the 100us timeout of the busy poll (by which API?),
> and what's the TPL of this busy poll?
>
Create a timeout timer with value of 1000. Then busy poll in a loop until packet is received or a timer is triggered.
This runs at TPL_CALLBACK while the EfxPoll function itself runs at TPL_NOTIFY.

-- 
  /* Set the poll moderation timeout to 100us */
  gBS->SetTimer(NetDev->PollModeratorTimer, TimerRelative, 1000);

  while (!InterruptStatus) {
    EfxPoll(NULL, NetDev->Efx);
    InterruptStatus = EfxStatus(NetDev->Efx);

    if (gBS->CheckEvent(NetDev->PollModeratorTimer))
      return InterruptStatus;
  }

  return InterruptStatus;
--

>> Otherwise I have seen a pathological case where the TFTP client would pull one
>> packet per second but the TFTP_RETRANSMIT interval on the server was also one
>> second and everything was awful.
> Why TFTP client just pull one packet per second? I think it’s not correct and it
> could use the poll() function to accelerate the receive. Why you trying to solve a
> TFTP layer problem in SNP layer? It breaks the design principle of network layer model.
Yes, I appreciate the principle. However, in practice we don't get to sell adapters that do not PXE boot and it's pointless to argue with customers that they have a rubbish TFTP client implementation. So we put in workarounds into the driver.
>> In theory the busy poll moderator might cause an issue in a pathological case,
>> if it stalls the machine for 100us for each packet - with 100 packets per
>> second this would eat only 10ms.
> Not only 10ms, it's at least 10ms, and in theory. If you have multiple NICs in the
> system, this time will double. If someone tries to poll the NIC, the time will increase
> significantly. It also depends on who you measure the 100us timeout, I did see some
> platform which has a minimum timer interval limitation, that you thought you are using
> a 100us timeout but it’s actually 1ms or even 10ms.
>
> So my suggestion is do NOT add any busy wait in SNP layer API, keep it as unblocking,
> and the try to solve the application layer problem in application layer.
I'll try dropping the busy poll and see whether that helps.

Cheers,
Tom
>
> Best Regards,
> Siyuan
>
>> Cheers,
>> Tom
>>
>> On 27/01/2019 14:28, Fu, Siyuan wrote:
>>> Hi, Tom
>>>
>>>> -----Original Message-----
>>>> From: edk2-devel [mailto:edk2-devel-bounces@lists.01.org] On Behalf Of
>> Tomas
>>>> Pilar (tpilar)
>>>> Sent: Friday, January 25, 2019 8:09 PM
>>>> To: Wu, Jiaxin <jiaxin.wu@intel.com>; Laszlo Ersek <lersek@redhat.com>;
>> edk2-
>>>> devel@lists.01.org
>>>> Cc: Ye, Ting <ting.ye@intel.com>
>>>> Subject: Re: [edk2] Network Stack Budgeting
>>>>
>>>> Yeah, that makes sense I suppose. The end result is however that the
>> network
>>>> device is 'opened' as soon as ConnectController() is called rather than
>> when
>>>> someone wants to do networking. This seems wrong and directly leads to a
>> DoS
>>>> of a platform in case of heavy network load (unless we implement budgeting
>> in
>>>> the network driver, which is a really not what it should be doing).
>>> It's true that a configured MNP will start a background system poll to
>> receive
>>> network packets, but normally it won't lead to a DoS, because the receive
>> frequency
>>> is actually very low. The MNP driver only tries to receive 1 packet from SNP
>> for every
>>> 10 milliseconds (see MNP_SYS_POLL_INTERVAL, MnpSystemPoll() and related
>> timer event).
>>> So no matter how heavy the network is, MNP driver (and upper layer stack)
>> will
>>> only process at most 100 packets per second, all other packets should be
>> dropped
>>> by UNDI driver directly.
>>>
>>> So if there is nobody doing busy network receiving, the network stack cost
>> will be
>>> at most 100 packets per NIC device and per second.
>>>
>>> In your first email you mentioned "some performance improvements to my
>> network driver",
>>> may I know what's the improvement method you are using?
>>>
>>> Best Regards,
>>> Siyuan
>>>
>>>> How do you suggest we solve this problem?
>>>>
>>>> Cheers,
>>>> Tom
>>>>
>>>> On 25/01/2019 08:44, Wu, Jiaxin wrote:
>>>>> Hi Tom,
>>>>>
>>>>> One thing I want to highlight is that our design of network stack is not
>>>> only for the PXE/HTTP boot trigged in BootManager, but also to make sure
>> it's
>>>> workable once there is any MNP instance configured by upper drivers
>>>> (ARP/IPv4/IPv6).
>>>>> Take ARP/IP as an example, once ARP/IP are started, we need a heartbeat to
>>>> process any ARP requests, which is required by ARP functionality. In such a
>>>> case, SNP must be started to make ARP/IP drivers works well.
>>>>> Thanks,
>>>>> Jiaxin
>>>>>
>>>>>> -----Original Message-----
>>>>>> From: edk2-devel [mailto:edk2-devel-bounces@lists.01.org] On Behalf Of
>>>>>> Tomas Pilar (tpilar)
>>>>>> Sent: Friday, January 25, 2019 1:43 AM
>>>>>> To: Laszlo Ersek <lersek@redhat.com>; edk2-devel@lists.01.org
>>>>>> Subject: Re: [edk2] Network Stack Budgeting
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 24/01/2019 16:49, Laszlo Ersek wrote:
>>>>>>> On 01/24/19 14:25, Tomas Pilar (tpilar) wrote:
>>>>>>>> Hmm,
>>>>>>>>
>>>>>>>> Mnp->Configure() will eventually call MnpStartSnp().
>>>>>>>>
>>>>>>>> A grep for Mnp->Configure shows that:
>>>>>>>> * ArpDxe performs this on DriverBinding.Start()
>>>>>>>> * Ip6Dxe performs this on DriverBinding.Start()
>>>>>>>>
>>>>>>>> Ipv4 and DnsDhcp do this as a part of their Configure() they expose in
>>>> the
>>>>>> API.
>>>>>>> Yes, that makes sense. All of the above drivers are UEFI drivers that
>>>>>>> follow the UEFI driver model, AIUI. As long as the controller is not
>>>>>>> connected from BDS, no BindingStart() function should be called in these.
>>>>>> Ah, but I would expect the BDS to call ConnectController() on the NIC,
>> but
>>>> I
>>>>>> would not expect the network stack to be started unless the device is
>>>>>> actually chosen for PXE booting. In other words, the above protocols
>> should
>>>>>> follow the example of EFI_DNS4_PROTOCOL that binds against the device
>>>>>> during BDS but .Configure() is not automatically called by
>>>>>> DriverBinding.Start().
>>>>>>
>>>>>> .Configure() should be called by the BootManager if networking is
>> actually
>>>> to
>>>>>> be done. That in turn will configure Mnp and start Snp.
>>>>>>
>>>>>> Cheers,
>>>>>> Tom
>>>>>>
>>>>>> _______________________________________________
>>>>>> edk2-devel mailing list
>>>>>> edk2-devel@lists.01.org
>>>>>> https://lists.01.org/mailman/listinfo/edk2-devel
>>>> _______________________________________________
>>>> edk2-devel mailing list
>>>> edk2-devel@lists.01.org
>>>> https://lists.01.org/mailman/listinfo/edk2-devel