svn commit: r344817 - in head/sys: dev/e1000 net
Andrew Gallatin
gallatin at cs.duke.edu
Fri Mar 8 19:24:06 UTC 2019
On 3/5/19 4:06 PM, Matthew Macy wrote:
> This represents a misunderstanding of how defines are used. This left
> the option open to the user to enable the use of larger than page size
> buffers as it does enable better performance. Over the course of a
> long uptime memory can get too fragmented. However, this left it open
> to the end consumer.
>
> I'd like to see this reverted with perhaps a better name for the
> define and the addition of an explanatory comment.
I'd strongly prefer that it stay removed. Since it is not hooked to an
option, no user is ever going to find it. This really should have been
a tuneable (since it is done at ring init time, rather than rx buffer
alloc time), but nobody cared enough to make it actually usable.
From brief memories of performance tuning 10G adapters 14 years ago,
the differences between page-sized and 9k jumbos were minimal even back
then (1/3 as many mbuf alloc/free, smaller chains). So I'm not
convinced that it is worth bringing back in any form.
My general feeling is that the more of this code that we can remove, the
better. Iflib is tricky enough that it is already challenging to reason
about and maintain. Removing code which is for all intents and purposes
unreachable and never tested is Good Thing.
Drew
More information about the svn-src-all
mailing list