On Mon, Mar 22, 2021 at 09:09:50PM -0700, Dipen Patel wrote:

On 3/22/21 7:59 PM, Kent Gibson wrote:

quoted

On Mon, Mar 22, 2021 at 06:53:10PM -0700, Dipen Patel wrote:

quoted

On 3/22/21 5:32 PM, Kent Gibson wrote:

quoted

On Mon, Mar 22, 2021 at 01:21:46PM -0700, Dipen Patel wrote:

quoted

Hi Linus and Kent,

[snip]

quoted

quoted

In response to all your comments above...

Firstly, I'm not suggesting that other kernel modules would use the
cdev lineevents, only that they would use the same mechanism that
gpiolib-cdev would use to timestamp the lineevents for userspace.

Sure, I just wanted to mention the different scenarios and wanted to know
how can we fit all those together. Having said that, shouldn't this serve
an opportunity to extend the linevent framework to accommodate kernel
drivers as a clients?

If we can't, then there is a risk of duplicating lineevent mechanism in all
of those kernel drivers or at least in GTE framework/infrastructure as far
as GPIO related GTE part is concerned.
 

In-kernel the lineevents are just IRQs so anything needing a "lineevent"
can request the IRQ directly.  Or am I missing something?

In the GPIO context, I meant we can extend line_event_timestamp to kernel
drivers as well in that way, both userspace and kernel drivers requesting
particular GPIO for the hardware timestamp would be managed by same
lineevent_* infrastructure from the gpiolib. Something like lineevent_create
version of the kernel drivers, so if they need hardware timestamp on the
GPIO line, they can request with some flags. In that way, GTE can leverage
linevent* codes from gpiolib to cover its both the GPIO related use cases i.e.
userspace app and kernel drivers.

I still don't see what that gives you that is better than an IRQ and a
function to provide the timestamp for that IRQ.  What specific features
of a lineevent are you after?

The gpiolib-cdev code is there to provide a palettable API for userspace,
and the bulk of that code is specific to the userspace API.
Reusing that code for clients within the kernel is just introducing
pointless overhead when they can get what they need more directly.

There may be a case for some additional gpiolib/irq helper functions, but
I don't see gpiolib-cdev as a good fit for that role.

quoted

quoted

quoted

As to that mechanism, my current thinking is that the approach of
associating GTE event FIFO entries with particular physical IRQ events is
problematic, as keeping the two in sync would be difficult, if not
impossible.

A more robust approach is to ignore the physical IRQs and instead service
the GTE event FIFO, generating IRQs from those events in software -
essentially a pre-timestamped IRQ.  The IRQ framework could provide the
timestamping functionality, equivalent to line_event_timestamp(), for
the IRQ handler/thread and in this case provide the timestamp from the GTE
event.

I have not fully understood above two paragraphs (except about
lineevent_event_timestamp related part).

I have no idea what it means to "ignore the physical IRQs and service the
GTE event FIFO". Just like GPIO clients, there could be IRQ clients which
want to monitor certain IRQ line, like ethernet driver wanted to retrieve
timestamp for its IRQ line and so on.

I mean that in the IRQ framework, rather than enabling the physical IRQ
line it would leave that masked and would instead enable the FIFO line to
service the FIFO, configure the GTE to generate the events for that
line, and then generate IRQs in response to the FIFO events.
That way the client requesting the IRQ is guaranteed to only receive an
IRQ that corresponds to a GTE FIFO event and the timestamp stored in the
IRQ framework would match.

I do not think we need to do such things, for example, below is
the rough sequence how GTE can notify its clients be it GPIO or IRQ
lines. I believe this will also help understand better ongoing GPIO
discussions.

1. Configure GTE FIFO watermark or threshold, lets assume 1, i.e
   generate GTE interrupt when FIFO depth is 1.
2. In the GTE ISR or ISR thread, drain internal FIFO entries
3. Through GTE driver's internal mapping, identify which IRQ or
   GPIO number this entry belongs to. (This is possible as GTE
   has predefined bits for each supported signals, for example GTE
   supports 40 GPIOs and 352 IRQ lines, and it has multliple GTE instances
   which can take care all of them)
4. GTE driver pushes the event data (in this case it will be timestamp and
   direction of the event ie.rising or falling) to the GTE generic framework
5. GTE framework will store per event data to its per client/event sw FIFO
6. wake up any sleeping client thread
7. Points 3 to 6 are happening in GTE ISR context. 
8. gte_retrieve_event (which can block if no event) at later convenient
   time do whatever it wants with it. We can extend it to non blocking
   version where some sort of client callbacks can be implemented.

Don't see where that conflicts with my suggestion except that I moved
everything, other than the hardware specific configuration, under the IRQ
umbrella as a lot of the functionality, in the abstract, aligns with IRQ.

If you want to reduce duplication shouldn't GTE be integrated into IRQ
rather than providing a new service?

As I see it, GTE is just another event source for IRQ, and I'd be
exploring that path. But that is just my view.

Cheers,
Kent.