Subject: Re: [PATCH] pci/sriov: Add an option to probe VFs or not before
 enabling SR-IOV
To: Alex Williamson <alex.williamson@redhat.com>,
        Gavin Shan <gwshan@linux.vnet.ibm.com>
References: <1490022874-54718-1-git-send-email-bodong@mellanox.com>
 <20170320230706.GA12252@gwshan>
 <7bfcfdcd-e0a8-f1e9-f112-fa35fdb845d7@mellanox.com>
 <20170320225708.6868676a@t450s.home> <20170321054305.GA12230@gwshan>
 <20170321000158.444991a3@t450s.home>
CC: <bhelgaas@google.com>, <linux-pci@vger.kernel.org>,
        <linux-kernel@vger.kernel.org>, <saeedm@mellanox.com>,
        Eli Cohen <eli@mellanox.com>
From: Bodong Wang <bodong@mellanox.com>
Message-ID: <7d9d8756-d737-66dd-4c0c-f407985f3db6@mellanox.com>
Date: Tue, 21 Mar 2017 08:43:32 -0500
User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:45.0) Gecko/20100101
 Thunderbird/45.4.0
MIME-Version: 1.0
In-Reply-To: <20170321000158.444991a3@t450s.home>
Content-Type: text/plain; charset="windows-1252"; format=flowed
Content-Transfer-Encoding: 7bit
SpamDiagnosticOutput: 1:99
SpamDiagnosticMetadata: NSPM
X-MS-Exchange-CrossTenant-OriginalArrivalTime: 21 Mar 2017 13:43:50.3781 (UTC)
X-MS-Exchange-CrossTenant-FromEntityHeader: Hosted
X-MS-Exchange-Transport-CrossTenantHeadersStamped: DB6PR0502MB3061
Sender: linux-kernel-owner@vger.kernel.org
List-ID: <linux-kernel.vger.kernel.org>
X-Mailing-List: linux-kernel@vger.kernel.org
Content-Length: 3036
Lines: 54

On 3/21/2017 1:01 AM, Alex Williamson wrote:
> On Tue, 21 Mar 2017 16:43:05 +1100
> Gavin Shan <gwshan@linux.vnet.ibm.com> wrote:
>
>> On Mon, Mar 20, 2017 at 10:57:08PM -0600, Alex Williamson wrote:
>>> On Mon, 20 Mar 2017 18:34:23 -0500
>>> Bodong Wang <bodong@mellanox.com> wrote:
>> .../...
>>
>>>>> Bodong, I'm not sure if there is a requirement to load driver for the
>>>>> specified number of VFs? That indicates no driver will be loaded for
>>>>> other VFs. If so, this interface might serve the purpose as well.
>>>> Gavin, thanks for the review. That is indeed an interesting suggestion.
>>>> Theoretically,  we can change that probe_vfs from boolean to integer.
>>>> And use it as a counter to probe the first N VFs(if N < total_vfs).
>>>> Let's see if there are any objections.
>>> Is it just me or does this seem like a confusing user interface, ie. to
>>> get binary on/off behavior a user now needs to 'cat total_vfs >
>>> sriov_probe_vfs'.  It's not very intuitive, what's the use case for it?
>>>   
>> After it's changed to integer, it accepts number. If users want to load
>> driver for all VFs and don't want to check the maximal number of VFs,
>> they can simply write 0xffffffff. So "on" and "off" are replaced with 0xffffffff
>> and 0, but users has to press the keyboard more times though.
>>
>> drivers/net/ethernet/mellanox/mlx4/main.c::probe_vfs_argc allows to specify
>> the number of VFs with which we're going to bind drivers. Less time is needed
>> to enable SRIOV capability. As I had in some development environment: assume
>> PF supports 256 VFs and I'm going to enable all of them, but I only want to
>> load driver for two of them, then test the data path on those two VFs. Besides,
>> I can image the VF needn't a driver in host if it's going to be passed to guest.
>> Not sure how much sense it makes.
> Yes, I understand what you're trying to do, but I still think it's
> confusing for a user interface.  This also doesn't answer what's the
> practical, typical user case you see where it's useful to probe some
> VFs but not others.  The case listed is a development case where you
> could just as easily disable all probing, then manually bind the first
> two VFs to the host driver.  Which is the better design, impose a
> confusing interface on all users to simplify an obscure development
> environment or simplify the user interface and assume developers know
> how to bind devices otherwise?  Thanks,
>
> Alex

I agree with Alex on this concern. Sometimes, I need to probe 1 or 2 VFs 
to host side just for development purpose. Bind/unbind satisfy this case 
perfectly, and that's how current implementation is designed. But, from 
Gavin's use case, it will be a pain to bind/unbind hundreds of VFs. So, 
I want to understand how common this use case is. If it's not common, I 
prefer to keep current design because 1) the interface is much easier to 
understand and use 2) less error prone because no need to check current 
total_vfs and maintain a static counter to enable N vfs.

Bodong