Re: [net-next 11/13] igb: Make Tx budget for NAPI user adjustable

[Alexander Duyck <alexander.h.duyck@intel.com>]

On 09/19/2011 09:05 AM, Ben Hutchings wrote:
> On Mon, 2011-09-19 at 08:48 -0700, Alexander Duyck wrote:
>> On 09/17/2011 10:04 AM, Ben Hutchings wrote:
>>> On Sat, 2011-09-17 at 01:04 -0700, Jeff Kirsher wrote:
>>>> From: Alexander Duyck<alexander.h.duyck@intel.com>
>>>>
>>>> This change is meant to make the NAPI budget limits for transmit
>>>> adjustable.  By doing this it is possible to tune the value for optimal
>>>> performance with applications such as routing.
>>> [...]
>>>> --- a/drivers/net/ethernet/intel/igb/igb_ethtool.c
>>>> +++ b/drivers/net/ethernet/intel/igb/igb_ethtool.c
>>>> @@ -1989,6 +1989,9 @@ static int igb_set_coalesce(struct net_device *netdev,
>>>>    	if ((adapter->flags&   IGB_FLAG_QUEUE_PAIRS)&&   ec->tx_coalesce_usecs)
>>>>    		return -EINVAL;
>>>>
>>>> +	if (ec->tx_max_coalesced_frames_irq)
>>>> +		adapter->tx_work_limit = ec->tx_max_coalesced_frames_irq;
>>>> +
>>> [...]
>>>
>>> I don't think it really makes sense to conflate NAPI and interrupt
>>> moderation parameters.  This really ought to be added to NAPI itself.
>>>
>>> (NAPI contexts really ought to be exposed through sysfs somehow.  I
>>> think we've discussed this before, and it's tricky due to the lack of a
>>> consistent mapping between those contexts and net devices.)
>>>
>>> Ben.
>> All NAPI does is move things from a hard interrupt to a soft interrupt
>> in the case of TX cleanup.  If it wasn't for NAPI we would be calling
>> ixgbe_clean_tx_irq directly from the interrupt handler and would still
>> be using the same limiting value.  This is why placing it here makes sense.
> But tx_max_coalesced_frames_irq is not supposed to be a work limit (and
> such a work limit doesn't seem useful in the absence of NAPI).  As I
> understand it, it is supposed to be an alternate moderation value for
> the hardware to use if a frame is sent while the IRQ handler is running.
>
> Ben.
The fact is ixgbe has been using this parameter this way for over 2 
years now and the main goal of this patch was just to synchronize how 
things work on igb and ixgbe.

Our hardware doesn't have a mechanism for firing an interrupt after X 
number of frames so instead we simply have modified things so that we 
will only process X number of frames and then fire another 
interrupt/poll if needed.  As such we aren't that far out of compliance 
with the meaning of how this parameter is supposed to be used.

Thanks,

Alex