Message-ID: <ae506e45-c4fe-7c36-6c9c-67b47818e7ba@gmail.com>
Date:   Mon, 21 Dec 2020 14:55:58 -0800
From:   Florian Fainelli <f.fainelli@...il.com>
To:     Vladimir Oltean <olteanv@...il.com>
Cc:     netdev@...r.kernel.org, andrew@...n.ch,
        "David S. Miller" <davem@...emloft.net>,
        Jakub Kicinski <kuba@...nel.org>,
        Murali Krishna Policharla <murali.policharla@...adcom.com>,
        Vladimir Oltean <vladimir.oltean@....com>,
        "open list:BROADCOM SYSTEMPORT ETHERNET DRIVER" 
        <bcm-kernel-feedback-list@...adcom.com>,
        open list <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH net] net: systemport: set dev->max_mtu to
 UMAC_MAX_MTU_SIZE



On 12/21/2020 2:25 PM, Vladimir Oltean wrote:
> On Mon, Dec 21, 2020 at 01:49:03PM -0800, Florian Fainelli wrote:
>> On 12/18/2020 1:17 PM, Florian Fainelli wrote:
>>>>>>>>> SYSTEMPORT Lite does not actually validate the frame length, so setting
>>>>>>>>> a maximum number to the buffer size we allocate could work, but I don't
>>>>>>>>> see a reason to differentiate the two types of MACs here.
>>>>>>>>
>>>>>>>> And if the Lite doesn't validate the frame length, then shouldn't it
>>>>>>>> report a max_mtu equal to the max_mtu of the attached DSA switch, plus
>>>>>>>> the Broadcom tag length? Doesn't the b53 driver support jumbo frames?
>>>>>>>
>>>>>>> And how would I do that without create a horrible layering violation in
>>>>>>> either the systemport driver or DSA? Yes the b53 driver supports jumbo
>>>>>>> frames.
>>>>>>
>>>>>> Sorry, I don't understand where is the layering violation (maybe it doesn't
>>>>>> help me either that I'm not familiar with Broadcom architectures).
>>>>>>
>>>>>> Is the SYSTEMPORT Lite always used as a DSA master, or could it also be
>>>>>> used standalone? What would be the issue with hardcoding a max_mtu value
>>>>>> which is large enough for b53 to use jumbo frames?
>>>>>
>>>>> SYSTEMPORT Lite is always used as a DSA master AFAICT given its GMII
>>>>> Integration Block (GIB) was specifically designed with another MAC and
>>>>> particularly that of a switch on the other side.
>>>>>
>>>>> The layering violation I am concerned with is that we do not know ahead
>>>>> of time which b53 switch we are going to be interfaced with, and they
>>>>> have various limitations on the sizes they support. Right now b53 only
>>>>> concerns itself with returning JMS_MAX_SIZE, but I am fairly positive
>>>>> this needs fixing given the existing switches supported by the driver.
>>>>
>>>> Maybe we don't need to over-engineer this. As long as you report a large
>>>> enough max_mtu in the SYSTEMPORT Lite driver to accomodate for all
>>>> possible revisions of embedded switches, and the max_mtu of the switch
>>>> itself is still accurate and representative of the switch revision limits,
>>>> I think that's good enough.
>>>
>>> I suppose that is fair, v2 coming, thanks!
>>
>> I was going to issue a v2 for this patch, but given that we don't
>> allocate buffers larger than 2KiB and there is really no need to
>> implement ndo_change_mtu(), is there really a point not to use
>> UMAC_MAX_MTU_SIZE for both variants of the SYSTEMPORT MAC?
> 
> After your first reply that "the Lite doesn't validate the frame length", I was
> under the impression that it is sufficient to declare a larger max_mtu such as
> JMS_MAX_SIZE and 9K jumbo frames would just work. But with the current buffer
> allocation in bcm_sysport_rx_refill it clearly wouldn't. A stupid confusion
> really. So yeah, sorry for having you resend a v2 with no change.
> If it helps you could add to the patch:
> 
> Reviewed-by: Vladimir Oltean <olteanv@...il.com>
> 
> Thanks again for explaining.

No worries, Jakub, David, do you need me to resend or can you pick it up
from patchwork?
-- 
Florian

Powered by blists - more mailing lists