Re: Xen Console input very slow in recent UEFI

[Julien Grall <julien.grall@linaro.org>]

Hi Laszlo,

Thank you for your help.

On 26/10/17 16:20, Laszlo Ersek wrote:
> On 10/26/17 17:13, Laszlo Ersek wrote:
>> Hello Julien,
>>
>> On 10/26/17 13:05, Julien Grall wrote:
>>> Hi all,
>>>
>>> I was doing more testing of UEFI in Xen guests and noticed some slow
>>> down when using the shell. The characters are only echoed after a second
>>> or two that is a bit annoying.
>>>
>>> The change that introduced this issue is 4cf3f37c87 "MdeModulePkg
>>> SerialDxe: Process timeout consistently in SerialRead".
>>>
>>> The Serial Driver for Xen PV console is very simple (see
>>> OvmfPkg/Library/XenConsoleSerialPortLib). So I am not sure where the
>>> root cause is.
>>>
>>> Would anyone have any tips on it?
>>
>> The exact same issue has been encountered earlier under QEMU, please
>> refer to the following sub-thread (please read it to end):
>>
>> http://mid.mail-archive.com/b748580c-cb51-32c9-acf9-780841ef15da@redhat.com
>>
>> The fix was commit 5f0f5e90ae8c ("ArmVirtPkg/FdtPL011SerialPortLib: call
>> PL011UartLib in all SerialPortLib APIs", 2017-08-16).
>>
>> I think if you can implement the same for XenConsoleSerialPortLib, that
>> should return to working state as well.
> 
> Hmmm, wait, at a closer look, it looks like
> 
>    OvmfPkg/Library/XenConsoleSerialPortLib/XenConsoleSerialPortLib.c
> 
> already implements that suggestion? (I.e., it sets
> EFI_SERIAL_INPUT_BUFFER_EMPTY in *Control as necessary?)
> 
> Are we sure the SerialPortPoll() function works correctly? I don't see
> any MemoryFence() calls in SerialPortPoll(), around checking the fields
> in (*mXenConsoleInterface). Could that be the problem?

I am not entirely sure. But I added a couple of MemoryFence() in 
SerialPort just in case to clear that from potential cause:

XENCONS_RING_IDX  Consumer, Producer;

if (!mXenConsoleInterface) {
     return FALSE;
}

MemoryFence ();

Consumer = mXenConsoleInterface->in_cons;
Producer = mXenConsoleInterface->in_prod;

MemoryFence ();

return Consumer != Producer;

I also added some debug printk (using a different interface) to confirm 
the value of Consumer and Producer are valid.  I can see the Producer 
increasing every time a key is pressed and then soon followed by 
SerialPortRead incrementing Consumer.

I did more debugging and find out the following is happening in 
TerminalConInTimerHandler (MdeModulePkg/Universal/Console/TerminalDxe) 
when a character is received:

   1) GetControl will return EFI_SERIAL_INPUT_BUFFER_EMPTY unset
      	=> Entering in the loop to fetch character from the serial
   2) GetOneKeyFromSerial()
	=> Return directly with the character read
   3) Looping as the fifo is not full and no error
   4) GetOneKeyFromSerial() -> SerialRead()
   	=> No more character so SerialPortPoll() will return FALSE and loop 
until timeout
	=> Return EFI_TIMEOUT
   5) Exiting the loop from TerminalConInTimerHandler
   6) Characters are printed

So the step 4) will introduce the timeout seen and delay the echoing of 
the characters received.

I could see a couple of solutions to fix it:
	1) Remove the timeout from SerialPortRead and rely on either
		a) caller to handle timeout
		b) each UART driver
	2) TerminalConInTimerHandler to check at every iteration whether the 
buffer is empty.

Any other ideas?

Cheers,

-- 
Julien Grall