On 1/15/2018 11:02 AM, Shahaf Shuler wrote:

Monday, January 15, 2018 12:21 PM, Ferruh Yigit:

quoted

quoted

I think all of the 3 should be in a single patch.
The reason is that the convert patch should maintain the same offloads

configuration needed for the application.

Perhaps I am missing some details about "mbuf fast free" offload, can you
please give more details about it, what does having or not having it mean?
Currently no PMD seems implemented it.

Sure,

FAST_FREE offload is the logical AND between the old txqflags of:
ETH_TXQ_FLAGS_NOREFCOUNT
ETH_TXQ_FLAGS_NOMULTMEM

The offload is just a performance optimization. As specified in the documentation [1] it enables the PMDs to further optimize the data path given the guarantees from the application. 
Not having it means possible performance degradation for some PMD which rely on it.

There is no PMD which implement it yet since not all PMDs moved to the new offloads API. However this flag is tested and translated into txqflags as part of rte_eth_convert_txq_offloads function.
Relevant PMDs for this offload will be: sfc, thunderx and i40e.

Thank you for clarification, I am OK to have it.

But since currently no PMD provide "DEV_TX_OFFLOAD_MBUF_FAST_FREE" capability,
and default txq_flags is overwritten, some PMDs lost this optimization until
they implement new capability, right?

[1]
/**< Device supports optimization for fast release of mbufs.                 
  *   When set application must guarantee that per-queue all mbufs comes from 
  *   the same mempool and has refcnt = 1.                                    
  */                                                                           

quoted

quoted

Before the convert patch the examples were using the default

configuration set by the PMD. In there the txq flags were set to ignore ref
count and to declare all mbufs are from the same pool.

quoted

The fast free Tx offload was added in order to keep this old offloads

configuration.

quoted

quoted

Wouldn't be better to enable new offloadings in a separate patch,
other than convert one? And I don't know if we want to enable that
specific offload for all samples.

As you can see, not all the examples has the FAST_FREE offloads, only the

entitled ones (i.e. single mempool and no ref count).

quoted

For example, ipv4_multicast doesn't set this offload flag.