Message-ID: <87wmw43aa5.fsf@intel.com>
Date: Mon, 02 Oct 2023 15:54:42 -0700
From: Vinicius Costa Gomes <vinicius.gomes@...el.com>
To: Xabier Marquiegui <reibax@...il.com>
Cc: alex.maftei@....com, chrony-dev@...ony.tuxfamily.org,
 davem@...emloft.net, horms@...nel.org, mlichvar@...hat.com,
 netdev@...r.kernel.org, ntp-lists@...tcorallo.com, reibax@...il.com,
 richardcochran@...il.com, rrameshbabu@...dia.com, shuah@...nel.org
Subject: Re: [PATCH net-next v3 3/3] ptp: support event queue reader channel
 masks

Xabier Marquiegui <reibax@...il.com> writes:

> Vinicius Costa Gomes <vinicius.gomes@...el.com> writes:
>
>> Looking below, at the usability of the API, it feels too complicated, I
>> was trying to think, "how an application would change the mask for
>> itself": first it would need to know the PID of the process that created
>> the fd, then it would have to find the OID associated with that PID, and
>> then build the request.
>> 
>> And it has the problem of being error prone, for example, it's easy for
>> an application to override the mask of another, either by mistake or
>> else.
>> 
>> My suggestion is to keep things simple, the "SET" only receives the
>> 'mask', and it only changes the mask for that particular fd (which you
>> already did the hard work of allowing that). Seems to be less error prone.
>> 
>> At least in my mental model, I don't think much else is needed (we
>> expose only a "SET" operation), at least from the UAPI side of things.
>> 
>> For "debugging", i.e. discovering which applications have what masks,
>> then perhaps we could do it "on the side", for example, a debugfs entry
>> that lists all open file descriptors and their masks. Just an idea.
>> 
>> What do you think?
>
> Thank you very much for your input Vinicius. I really appreciate it.
>
> I totally agree with your observations. I had already thought about that angle
> myself, but I decided to go this route anyway because it was the only way I
> could think of meeting all of Richard's requirements at that time.
>
> Even if being error prone, being able to externally manipulate the channel
> masks is the only way I can think of to make this feature backwards compatible
> with existing software. One example of a piece of software that would need to
> be updated to support multiple channels is linuxptp. If you try to start ts2phc
> with multiple channels enabled and no masks, it refuses to work stating that
> unwanted channels are present. This would be easy to fix, incorporating the
> SET operation you mention, but it is still something that needs to be changed.
>

I never looked at this a lot, so, as always, I could be missing stuff.

But from the way I see things, the solution that seems better has two
parts:

1. Fix ts2phc to ignore events from channels that it cannot/doesn't want
   to handle. (Is this possible?)
2. Add the "set mask ioctl/alternative ideas, is then more like a
   optimization, to avoid waking up applications that don't want some
   events;

So we have 'ts2phc' working on "old" kernels and on "new" kernels it is
"just" more efficient.

> Now that I think of it, it is true that nothing prevents us from having both
> methods available: the simple and safe, and the complicated and unsafe.
>
> Even with that option, I also think that going exclusively with the safe
> and simple route is better.
>
> So, I wonder: Can we just do it and require changes in software that relies
> on this driver, or should we maintain compatibility at all cost?
>
> Thank you very much for sharing your knowledge and experience.


Cheers,
-- 
Vinicius

Powered by blists - more mailing lists