Hi Michał,

On Thu, Jul 29, 2021 at 08:40:00AM +0200, Michał Krawczyk wrote:

śr., 28 lip 2021 o 16:27 Olivier Matz [off-list ref] napisał(a):

quoted

Hi Michał,

On Wed, Jul 14, 2021 at 12:02:32PM +0200, Michał Krawczyk wrote:

quoted

pon., 12 lip 2021 o 19:03 Ghalem Boudour [off-list ref] napisał(a):

quoted

The DPDK ENA driver does not provide multi-segment tx offload capability.
Let's add DEV_TX_OFFLOAD_MULTI_SEGS to ports offload capability by default.

Hi Ghalem,

This patch enables announcement of the DEV_TX_OFFLOAD_MULTI_SEGS
capability, but still the application may not request this offload.

As ENA PMD currently assumes all the mbufs may have multiple segments
(and we don't have fast-path for the other cases), I suggest
overwriting this flag in the ena_dev_configure(), similar to what
we're doing with the DEV_RX_OFFLOAD_RSS_HASH flag.

To give some more context, our application currently checks if the
driver supports multi-segments by checking its capabilities, and asks
for the feature if it is advertised.

When dealing with drivers that do not advertise this capability, our app
linearizes the segmented mbufs before sending them to the driver.

I think this is the proper way to use the API: if the driver supports to
handle multisegmented mbufs, it should advertise the capability.

Regards,
Olivier

Hi Olivier,

I agree we should advertise it. However, after advertising this
option, the application should ask PMD to use this feature if I
understand the offload API correctly.

Yes

I was thinking about the PMD behavior if this option won't be
requested by the application - it would be the same as for the
multisegment setup, that's why I suggested to also override this
feature at the configuration step.

With the RSS hash feature we're doing exactly the same - first we're
advertising that it's available, and then (if RSS was enabled), we're
setting this offload as enabled, even though the application didn't
request it explicitly.

I have no strong opinion here. If there is no behavior change when the
option is set or unset by the application, I think both options are
acceptable: either return what the application asked for, or return the
option always set (as you do for hash).

Anyway, I won't be holding push of this patch as we're close to rc3.
@Ghalem, please add the fixline and Cc the dpdk@stable.org in the
commit log, so we will have this commit backported.

Ghalem is currently off, I'll take care of it.

The multiseg Tx feature is there since the beginning, so the Fixes
line will be:
Fixes: 1173fca25af9 ("ena: add polling-mode driver")

I can also add the flag in ena_dev_configure(), if you feel it is more
consistent. Please let me know.

Thanks,
Olivier