[dpdk-dev] IXGBE error statistics

[dpdk-dev] IXGBE error statistics

[Igor Ryzhov]

Hello everyone.

Investigating IXGBE driver I found an mspdc counter (MAC Short Packet Discard). And I am wondering why this counter is not used in the calculation of total RX errors (ierrors field in rte_eth_stats structure). Is it already a part of another counter, for example, rlec (Receive Length Error)? Or is it a bug?

Another one question is about incompleteness of rte_eth_stats structure. IXGBE and other drivers have a lot of counters but only a part of them is represented in rte_eth_stats. Is there any valuable reasons for that or it's just not implemented?

Best regards,
Igor Ryzhov

Re: [dpdk-dev] IXGBE error statistics

[Van Haaren, Harry]

> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Igor Ryzhov
> Hello everyone.

Hi Igor,

> Investigating IXGBE driver I found an mspdc counter (MAC Short Packet
> Discard). And I am wondering why this counter is not used in the calculation
> of total RX errors (ierrors field in rte_eth_stats structure). Is it already a part
> of another counter, for example, rlec (Receive Length Error)? Or is it a bug?

There has been a discussion on list recently involving ixgbe stats, and certain
packets triggering multiple stats registers - the datasheet doesn't mention
this could be the case for the mspdc register, I will research this issue and
get back to you.

> Another one question is about incompleteness of rte_eth_stats structure.
> IXGBE and other drivers have a lot of counters but only a part of them is
> represented in rte_eth_stats. Is there any valuable reasons for that or it's
> just not implemented?

The rte_eth_stats struct presents the most general statistics that every NIC exposes.
In 2.1, and extended statistics API was added which allows NICs to expose stats
that are unique to that NIC. Currently ixgbe is the only driver that has the xstats API
implemented, I am working on patches to implement the functionality for the other
Intel drivers.

As part of testing the xstats implementation for each driver, I can test the exact
behavior of the mspdc counter, and if it is mis-counted this should become clear.

Cheers, -Harry

Re: [dpdk-dev] IXGBE error statistics

[Igor Ryzhov]

Hello, Harry.

Thank you, I'll wait for result of mspdc testing.

About rte_eth_stats - I found that not generic fields of the structure are all deprecated already. I will research xstats API, thank you.

Best regards,
Igor

> 18 сент. 2015 г., в 11:04, Van Haaren, Harry <harry.van.haaren@intel.com> написал(а):
> 
>> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Igor Ryzhov
>> Hello everyone.
> 
> Hi Igor,
> 
>> Investigating IXGBE driver I found an mspdc counter (MAC Short Packet
>> Discard). And I am wondering why this counter is not used in the calculation
>> of total RX errors (ierrors field in rte_eth_stats structure). Is it already a part
>> of another counter, for example, rlec (Receive Length Error)? Or is it a bug?
> 
> There has been a discussion on list recently involving ixgbe stats, and certain
> packets triggering multiple stats registers - the datasheet doesn't mention
> this could be the case for the mspdc register, I will research this issue and
> get back to you.
> 
>> Another one question is about incompleteness of rte_eth_stats structure.
>> IXGBE and other drivers have a lot of counters but only a part of them is
>> represented in rte_eth_stats. Is there any valuable reasons for that or it's
>> just not implemented?
> 
> The rte_eth_stats struct presents the most general statistics that every NIC exposes.
> In 2.1, and extended statistics API was added which allows NICs to expose stats
> that are unique to that NIC. Currently ixgbe is the only driver that has the xstats API
> implemented, I am working on patches to implement the functionality for the other
> Intel drivers.
> 
> As part of testing the xstats implementation for each driver, I can test the exact
> behavior of the mspdc counter, and if it is mis-counted this should become clear.
> 
> Cheers, -Harry

Re: [dpdk-dev] IXGBE error statistics

[Van Haaren, Harry]

> From: Igor Ryzhov [mailto:iryzhov@arccn.ru]
> Thank you, I'll wait for result of mspdc testing.

Hi Igor,

Regarding your original question:
The datasheet says that a packet with total size < 32 bytes is
discarded by MAC layer and counted in the mspdc register.

If possible, I would like to add a test for this type of packet when
testing the accuracy of xstats. I've tried using a packet generator
(48 bytes seems the smallest it is willing to generate for me), and
using another NIC won't work, as smaller packets are padded to
64 bytes on TX.
 
Do you have a method of reproducing this < 32 byte mspdc error packet? 
-Harry

Re: [dpdk-dev] IXGBE error statistics

[Igor Ryzhov]

Hi, Harry.

Sorry, but minimum size of packets we can generate is 64 bytes long.

Best regards,
Igor

> 21 сент. 2015 г., в 16:45, Van Haaren, Harry <harry.van.haaren@intel.com> написал(а):
> 
>> From: Igor Ryzhov [mailto:iryzhov@arccn.ru]
>> Thank you, I'll wait for result of mspdc testing.
> 
> Hi Igor,
> 
> Regarding your original question:
> The datasheet says that a packet with total size < 32 bytes is
> discarded by MAC layer and counted in the mspdc register.
> 
> If possible, I would like to add a test for this type of packet when
> testing the accuracy of xstats. I've tried using a packet generator
> (48 bytes seems the smallest it is willing to generate for me), and
> using another NIC won't work, as smaller packets are padded to
> 64 bytes on TX.
> 
> Do you have a method of reproducing this < 32 byte mspdc error packet? 
> -Harry

Re: [dpdk-dev] IXGBE error statistics

[Kyle Larose]

On Mon, Sep 21, 2015 at 9:45 AM, Van Haaren, Harry
<harry.van.haaren@intel.com> wrote:
>> From: Igor Ryzhov [mailto:iryzhov@arccn.ru]
>> Thank you, I'll wait for result of mspdc testing.
>
> Hi Igor,
>
> Regarding your original question:
> The datasheet says that a packet with total size < 32 bytes is
> discarded by MAC layer and counted in the mspdc register.
>
> If possible, I would like to add a test for this type of packet when
> testing the accuracy of xstats. I've tried using a packet generator
> (48 bytes seems the smallest it is willing to generate for me), and
> using another NIC won't work, as smaller packets are padded to
> 64 bytes on TX.
>
> Do you have a method of reproducing this < 32 byte mspdc error packet?
> -Harry

Can you disable the MAC padding on the NIC through a backdoor register access?

Re: [dpdk-dev] IXGBE error statistics

[Van Haaren, Harry]

> From: Kyle Larose [mailto:eomereadig@gmail.com]
> Can you disable the MAC padding on the NIC through a backdoor register
> access?

Good idea - yes it looks like that's possible, TXPADEN register.

I'll look into it, and post back later. Thanks, -Harry

Re: [dpdk-dev] IXGBE error statistics

[Van Haaren, Harry]

> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Van Haaren, Harry
> I'll look into it [mspdc error register stats], and post back later.

Hi guys,

Attempts to generate a 32 byte packet and make mspdc increment have not been successful - however from the datasheet and a theoretical point-of-view, it makes sense to add mspdc to rx_errors. Hence I've sent a patch that will include mspdc in the rx error count:

http://dpdk.org/dev/patchwork/patch/8064/

Thanks, -Harry