Re: RFC: i40e xmit path HW limitation

On 07/30/15 20:01, Stephen Hemminger wrote:

On Thu, 30 Jul 2015 19:50:27 +0300
Vlad Zolotarov [off-list ref] wrote:

quoted

On 07/30/15 19:20, Avi Kivity wrote:

quoted

On 07/30/2015 07:17 PM, Stephen Hemminger wrote:

quoted

On Thu, 30 Jul 2015 17:57:33 +0300
Vlad Zolotarov [off-list ref] wrote:

quoted

Hi, Konstantin, Helin,
there is a documented limitation of xl710 controllers (i40e driver)
which is not handled in any way by a DPDK driver.
   From the datasheet chapter 8.4.1:

"• A single transmit packet may span up to 8 buffers (up to 8 data
descriptors per packet including
both the header and payload buffers).
• The total number of data descriptors for the whole TSO (explained
later on in this chapter) is
unlimited as long as each segment within the TSO obeys the previous
rule (up to 8 data descriptors
per segment for both the TSO header and the segment payload buffers)."

This means that, for instance, long cluster with small fragments has to
be linearized before it may be placed on the HW ring.
In more standard environments like Linux or FreeBSD drivers the
solution
is straight forward - call skb_linearize()/m_collapse() corresponding.
In the non-conformist environment like DPDK life is not that easy -
there is no easy way to collapse the cluster into a linear buffer from
inside the device driver
since device driver doesn't allocate memory in a fast path and utilizes
the user allocated pools only.

Here are two proposals for a solution:

   1. We may provide a callback that would return a user TRUE if a give
      cluster has to be linearized and it should always be called before
      rte_eth_tx_burst(). Alternatively it may be called from inside the
      rte_eth_tx_burst() and rte_eth_tx_burst() is changed to return
some
      error code for a case when one of the clusters it's given has
to be
      linearized.
   2. Another option is to allocate a mempool in the driver with the
      elements consuming a single page each (standard 2KB buffers would
      do). Number of elements in the pool should be as Tx ring length
      multiplied by "64KB/(linear data length of the buffer in the pool
      above)". Here I use 64KB as a maximum packet length and not taking
      into an account esoteric things like "Giant" TSO mentioned in the
      spec above. Then we may actually go and linearize the cluster if
      needed on top of the buffers from the pool above, post the buffer
      from the mempool above on the HW ring, link the original
cluster to
      that new cluster (using the private data) and release it when the
      send is done.

Or just silently drop heavily scattered packets (and increment oerrors)
with a PMD_TX_LOG debug message.

I think a DPDK driver doesn't have to accept all possible mbufs and do
extra work. It seems reasonable to expect caller to be well behaved
in this restricted ecosystem.

How can the caller know what's well behaved?  It's device dependent.

+1

Stephen, how do you imagine this well-behaved application? Having switch
case by an underlying device type and then "well-behaving" correspondingly?
Not to mention that to "well-behave" the application writer has to read
HW specs and understand them, which would limit the amount of DPDK
developers to a very small amount of people... ;) Not to mention that
the mentioned above switch-case would be a super ugly thing to be found
in an application that would raise a big question about the
justification of a DPDK existence as as SDK providing device drivers
interface. ;)

Either have a RTE_MAX_MBUF_SEGMENTS

And what would it be in our care? 8? This would limit the maximum TSO packet to 16KB for 2KB buffers.

that is global or
a mbuf_linearize function?  Driver already can stash the
mbuf pool used for Rx and reuse it for the transient Tx buffers.

First of all who can guaranty that that pool would meet our needs - namely have large enough buffers?
Secondly, using user's Rx mempool for that would be really not nice (read - dirty) towards the user that may had allocated the specific amount of buffers in it according to some calculations that didn't include the usage from the Tx flow.

And lastly and most importantly, this would require using the atomic operations during access to Rx mempool, that would both require a specific mempool initialization and would significantly hit the performance.