Re: [Intel-wired-lan] [PATCH net-next v4 0/3] ethtool: Dynamic RSS context indirection table resizing

Re: [Intel-wired-lan] [PATCH net-next v4 0/3] ethtool: Dynamic RSS context indirection table resizing

[Jakub Kicinski]

On Fri, 13 Mar 2026 08:13:12 +0100 Björn Töpel wrote:
> Some NICs (e.g. bnxt) change their RSS indirection table size based on
> the queue count, because the hardware table is a shared resource. The
> ethtool core locks ctx->indir_size at context creation, so drivers
> have to reject channel changes when RSS contexts exist.
> 
> This series adds resize helpers and wires them up in bnxt.

Sorry Bjorn, I was typing the explanation below and I realized that 
we may be violating user intent. We should already record the user_size
from rss_set_prep_indir() as part of the context (and presumably some
netdev state for the main context?) and don't allow shrinking the
context below that mark..


Now for the broader audience - my understanding is that the RSS table
is a precious resource for most if not all drivers.
The direction of this work is to allow user to explicitly specify what
RSS table size they want. The Netlink API for RSS already allows users
to send tables smaller than what the device reports (Netlink code just
"replicates" the table). So if the user asks for X entries the driver
should be able to allocate a table of any size as long as its a
multiple of X. This series only support "global" resizing but it
should be simple to add the state for the driver to mark in the context
what table size it actually allocated. The only reason we haven't added
the support for that is that it's quite hard to know how to implement
it without docs for the FW/HW.. fbnic has fixed table sizes so we can't
even prototype there :S

Please reply if you'd like to collaborate on adding support for the
user-defined RSS context sizing.

Re: [Intel-wired-lan] [PATCH net-next v4 0/3] ethtool: Dynamic RSS context indirection table resizing

[Björn Töpel]

Jakub Kicinski <kuba@kernel.org> writes:

> On Fri, 13 Mar 2026 08:13:12 +0100 Björn Töpel wrote:
>> Some NICs (e.g. bnxt) change their RSS indirection table size based on
>> the queue count, because the hardware table is a shared resource. The
>> ethtool core locks ctx->indir_size at context creation, so drivers
>> have to reject channel changes when RSS contexts exist.
>> 
>> This series adds resize helpers and wires them up in bnxt.
>
> Sorry Bjorn, I was typing the explanation below and I realized that 
> we may be violating user intent. We should already record the user_size
> from rss_set_prep_indir() as part of the context (and presumably some
> netdev state for the main context?) and don't allow shrinking the
> context below that mark..

Don't be sorry! Good catch, and glad you caught it now rather than after
it landed.

I'll respin with user_size tracked in the context (and netdev state for
context 0) as a lower bound, so resize never folds below the size the
user explicitly configured.

WDYT?


Björn

Re: [Intel-wired-lan] [PATCH net-next v4 0/3] ethtool: Dynamic RSS context indirection table resizing

[Jakub Kicinski]

On Sun, 15 Mar 2026 13:30:33 +0100 Björn Töpel wrote:
> Jakub Kicinski <kuba@kernel.org> writes:
> 
> > On Fri, 13 Mar 2026 08:13:12 +0100 Björn Töpel wrote:  
> >> Some NICs (e.g. bnxt) change their RSS indirection table size based on
> >> the queue count, because the hardware table is a shared resource. The
> >> ethtool core locks ctx->indir_size at context creation, so drivers
> >> have to reject channel changes when RSS contexts exist.
> >> 
> >> This series adds resize helpers and wires them up in bnxt.  
> >
> > Sorry Bjorn, I was typing the explanation below and I realized that 
> > we may be violating user intent. We should already record the user_size
> > from rss_set_prep_indir() as part of the context (and presumably some
> > netdev state for the main context?) and don't allow shrinking the
> > context below that mark..  
> 
> Don't be sorry! Good catch, and glad you caught it now rather than after
> it landed.
> 
> I'll respin with user_size tracked in the context (and netdev state for
> context 0) as a lower bound, so resize never folds below the size the
> user explicitly configured.
> 
> WDYT?

Yes, keep in mind that AFAIR context 0 is not currently tracked
in the xarray. So maybe add the size of the main / default context 
as a field in struct ethtool_netdev_state ?  Or we can start tracking
it (partially). IDK what's cleaner.