Message-ID: <20180102100817.GB2051@nanopsycho.orion>
Date:   Tue, 2 Jan 2018 11:08:17 +0100
From:   Jiri Pirko <jiri@...nulli.us>
To:     Arkadi Sharshevsky <arkadis@...lanox.com>
Cc:     netdev@...r.kernel.org, dsa@...ulusnetworks.com,
        roopa@...ulusnetworks.com, davem@...emloft.net, mlxsw@...lanox.com,
        andrew@...n.ch, vivien.didelot@...oirfairelinux.com,
        f.fainelli@...il.com, michael.chan@...adcom.com,
        ganeshgr@...lsio.com, saeedm@...lanox.com, matanb@...lanox.com,
        leonro@...lanox.com, idosch@...lanox.com,
        jakub.kicinski@...ronome.com, ast@...nel.org, daniel@...earbox.net,
        simon.horman@...ronome.com, pieter.jansenvanvuuren@...ronome.com,
        john.hurley@...ronome.com, alexander.h.duyck@...el.com,
        linville@...driver.com, gospo@...adcom.com,
        steven.lin1@...adcom.com, yuvalm@...lanox.com,
        ogerlitz@...lanox.com
Subject: Re: [patch net-next v2 00/10] Add support for resource abstraction

Mon, Jan 01, 2018 at 03:58:33PM CET, arkadis@...lanox.com wrote:
>
>
>On 12/26/2017 01:23 PM, Jiri Pirko wrote:
>> From: Jiri Pirko <jiri@...lanox.com>
>> 
>> Many of the ASIC's internal resources are limited and are shared between
>> several hardware procedures. For example, unified hash-based memory can
>> be used for many lookup purposes, like FDB and LPM. In many cases the user
>> can provide a partitioning scheme for such a resource in order to perform
>> fine tuning for his application. In such cases performing driver reload is
>> needed for the changes to take place, thus this patchset also adds support
>> for hot reload.
>> 
>> Such an abstraction can be coupled with devlink's dpipe interface, which
>> models the ASIC's pipeline as a graph of match/action tables. By modeling
>> the hardware resource object, and by coupling it to several dpipe tables,
>> further visibility can be achieved in order to debug ASIC-wide issues.
>> 
>> The proposed interface will provide the user the ability to understand the
>> limitations of the hardware, and receive notification regarding its occupancy.
>> Furthermore, monitoring the resource occupancy can be done in real-time and
>> can be useful in many cases.
>> ---
>> Userspace part prototype can be found at https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Farkadis%2Fiproute2%2F&data=02%7C01%7Carkadis%40mellanox.com%7C1ae3d8b4854a454e21e008d54c5329e3%7Ca652971c7d2e4d9ba6a4d149256f461b%7C0%7C0%7C636498842440762657&sdata=7MC2BFQFxjnmHqy2sOOL9VEa4ZGq6e5Z2n2WvuNgyFk%3D&reserved=0
>> at resource_dev branch.
>> 
>> v1->v2
>> - Add resource size attribute.
>> - Fix split bug.
>> 
>
>Just to summarize the current fixes required:
>
>1. ERIF dpipe table size is reporting wrong size. More precisely the
>   ERIF table does not take rifs, so it should not be linked to the rif
>   bank resource (is not part of this patchset, future extension).
>2. Extended ACK user-space bug.
>3. ABI documentation- Not sure we agreed upon it, Jiri?

Question is where to put it. It is mlxsw-specific thing, moreover,
Spectrum-specific thing, same as dpipe tables etc. Not sure. Perhaps
Documentation/networking/mlxsw.txt ?


>
>If I missed something please respond. Nothing of the fixes mentioned
>above is relevant for this patchset actually.
>
>Couple of key-points:
>
>1. Constrains\trade off about setting the sizes - this can be obtained
>   trivially from the resource tree nested structure.
>2. Dpipe provides the mapping of hardware processes to resources.
>3. Units - each resource specifies his units, if dpipe table's size is
>   X and its related to some resource its size is normalized to that
>   resources basic unit.
>
>IMO this is the most hardware exact interaction, and this is the way it
>should be exported from the kernel, if something is not presented in
>'user' convenient way some utilities can be implemented in userspace
>to easily do it. Furthermore, some examples will be provided for the
>whole kvd tree partition for different cases (IPv6 heavy etc..).
>Advanced user will be able to tweak it as they like.
>
>Regarding the 'switchdev' layer I think that kernel's software tables
>like nexthops/neigh/routes should be mapped to dpipe tables and not
>to resources directly:

Sure. dpipe table -> resource mapping is the only one that makes sense.

>
>kernel_fdb--> dpipe_fdb -->/kvd/hash_single.
>
>> Arkadi Sharshevsky (10):
>>   devlink: Add per devlink instance lock
>>   devlink: Add support for resource abstraction
>>   devlink: Add support for reload
>>   devlink: Add relation between dpipe and resource
>>   mlxsw: pci: Add support for performing bus reset
>>   mlxsw: spectrum: Register KVD resources with devlink
>>   mlxsw: spectrum_dpipe: Connect dpipe tables to resources
>>   mlxsw: spectrum: Add support for getting kvdl occupancy
>>   mlxsw: pci: Add support for getting resource through devlink
>>   mlxsw: core: Add support for reload
>> 
>>  drivers/net/ethernet/mellanox/mlxsw/core.c         |  85 ++-
>>  drivers/net/ethernet/mellanox/mlxsw/core.h         |  16 +-
>>  drivers/net/ethernet/mellanox/mlxsw/i2c.c          |   5 +-
>>  drivers/net/ethernet/mellanox/mlxsw/pci.c          |  98 ++--
>>  drivers/net/ethernet/mellanox/mlxsw/spectrum.c     | 205 ++++++++
>>  drivers/net/ethernet/mellanox/mlxsw/spectrum.h     |  13 +
>>  .../net/ethernet/mellanox/mlxsw/spectrum_dpipe.c   |  72 ++-
>>  .../net/ethernet/mellanox/mlxsw/spectrum_kvdl.c    |  26 +
>>  include/net/devlink.h                              |  97 ++++
>>  include/uapi/linux/devlink.h                       |  21 +
>>  net/core/devlink.c                                 | 573 ++++++++++++++++++---
>>  11 files changed, 1079 insertions(+), 132 deletions(-)
>> 

Powered by blists - more mailing lists