On Wed, 15 Apr 2020 at 12:18, Maxime Ripard <maxime@xxxxxxxxxx> wrote: > > On Wed, Apr 08, 2020 at 06:30:51PM +0300, Sakari Ailus wrote: > > Cc'ing Laurent as well. > > > > On Wed, Apr 08, 2020 at 06:28:57PM +0300, Sakari Ailus wrote: > > > Hi Maxime, > > > > > > On Wed, Apr 08, 2020 at 03:43:15PM +0200, Maxime Ripard wrote: > > > > On Wed, Apr 08, 2020 at 02:35:28PM +0200, Tomasz Figa wrote: > > > > > On Wed, Apr 8, 2020 at 2:21 PM Maxime Ripard <maxime@xxxxxxxxxx> wrote: > > > > > > On Tue, Apr 07, 2020 at 08:20:35PM +0300, Sakari Ailus wrote: > > > > > > > On Tue, Apr 07, 2020 at 06:46:06PM +0200, Tomasz Figa wrote: > > > > > > > > On Tue, Apr 7, 2020 at 6:40 PM Sakari Ailus <sakari.ailus@xxxxxx> wrote: > > > > > > > > > > > > > > > > > > On Tue, Apr 07, 2020 at 05:47:41PM +0200, Robert Foss wrote: > > > > > > > > > > On Tue, 7 Apr 2020 at 14:32, Maxime Ripard <maxime@xxxxxxxxxx> wrote: > > > > > > > > > > > > > > > > > > > > > > Hi Robert, > > > > > > > > > > > > > > > > > > > > > > On Tue, Apr 07, 2020 at 01:29:05PM +0200, Robert Foss wrote: > > > > > > > > > > > > On Tue, 7 Apr 2020 at 10:36, Maxime Ripard <maxime@xxxxxxxxxx> wrote: > > > > > > > > > > > > > On Mon, Apr 06, 2020 at 11:35:07AM +0300, Sakari Ailus wrote: > > > > > > > > > > > > > > > But that 19.2MHz is not a limitation of the device itself, it's a > > > > > > > > > > > > > > > limitation of our implementation, so we can instead implement > > > > > > > > > > > > > > > something equivalent in Linux using a clk_set_rate to 19.2MHz (to make > > > > > > > > > > > > > > > sure that our parent clock is configured at the right rate) and the > > > > > > > > > > > > > > > clk_get_rate and compare that to 19.2MHz (to make sure that it's not > > > > > > > > > > > > > > > been rounded too far apart from the frequency we expect). > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > This is doing exactly the same thing, except that we don't encode our > > > > > > > > > > > > > > > implementation limitations in the DT, but in the driver instead. > > > > > > > > > > > > > > > > > > > > > > > > > > > > What I really wanted to say that a driver that doesn't get the clock > > > > > > > > > > > > > > frequency from DT but still sets that frequency is broken. > > > > > > > > > > > > > > > > > > > > > > > > > > > > This frequency is highly system specific, and in many cases only a certain > > > > > > > > > > > > > > frequency is usable, for a few reasons: On many SoCs, not all common > > > > > > > > > > > > > > frequencies can be used (e.g. 9,6 MHz, 19,2 MHz and 24 MHz; while others > > > > > > > > > > > > > > are being used as well), and then that frequency affects the usable CSI-2 > > > > > > > > > > > > > > bus frequencies directly --- and of those, only safe, known-good ones > > > > > > > > > > > > > > should be used. IOW, getting the external clock frequency wrong typically > > > > > > > > > > > > > > has an effect that that none of the known-good CSI-2 bus clock frequencies > > > > > > > > > > > > > > are available. > > > > > > > > > > > > > > > > > > > > > > > > > > So clock-frequency is not about the "Frequency of the xvclk clock in > > > > > > > > > > > > > Hertz", but the frequency at which that clock must run on this > > > > > > > > > > > > > particular SoC / board to be functional? > > > > > > > > > > > > > > > > > > > > > > > > > > If so, then yeah, we should definitely keep it, but the documentation > > > > > > > > > > > > > of the binding should be made clearer as well. > > > > > > > > > > > > > > > > > > > > > > > > Alright so, let me summarise the desired approach then. > > > > > > > > > > > > > > > > > > > > > > There's a separate discussion on the same topic here: > > > > > > > > > > > https://lore.kernel.org/linux-media/20200407122106.GD4751@xxxxxxxxxxxxxxxxxxxxxxxxxx/ > > > > > > > > > > > > > > > > > > > > Thanks for the link. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > ACPI: > > > > > > > > > > > > - Fetch the "clock-frequency" property > > > > > > > > > > > > - Verify it to be 19.2Mhz > > > > > > > > > > > > > > > > > > > > > > > > DT: > > > > > > > > > > > > - Fetch the "clock-frequency" property > > > > > > > > > > > > - Verify it to be 19.2Mhz > > > > > > > > > > > > - Get xvclk clock > > > > > > > > > > > > - Get xvclk clock rate > > > > > > > > > > > > - Verify xvclk clock rate to be 19.2Mhz > > > > > > > > > > > > > > > > > > > > > > The current status is that you should > > > > > > > > > > > 's/clock-frequency/link-frequencies/', and in order to replace > > > > > > > > > > > assigned-clock-rates, you'll want to have a clk_set_rate to 19.2MHz > > > > > > > > > > > between steps 3 and 4 > > > > > > > > > > > > > > > > > > > > Would we want to 's/clock-frequency/link-frequencies/' for ACPI too? > > > > > > > > > > I imagine that would cause some breakage. > > > > > > > > > > > > > > > > > > It would, yes, and it would be no more correct on DT either. > > > > > > > > > > > > > > > > > > There are basically two possibilities here; either use the clock-frequency > > > > > > > > > property and set the frequency, or rely on assigned-clock-rates, and get > > > > > > > > > the frequency instead. > > > > > > > > > > > > > > > > > > The latter, while I understand it is generally preferred, comes with having > > > > > > > > > to figure out the register list set that closest matches the frequency > > > > > > > > > obtained. The former generally gets around this silently by the clock > > > > > > > > > driver setting the closest frequency it can support. > > > > > > > > > > > > > > > > Wouldn't the former actually cause problems, because the closest > > > > > > > > frequency the clock driver can support could be pretty far from the > > > > > > > > one requested? (E.g. 19.2 MHz vs 24 MHz) The driver needs to check the > > > > > > > > resulting frequency anyway. > > > > > > > > > > > > > > That's possible, yes; in this case there wouldn't be a guarantee the > > > > > > > frequency wouldn't be far off. > > > > > > > > > > > > assigned-clock-rates is really fragile... There's zero guarantee on > > > > > > how far the actual rate is going to be from the asked one, but more > > > > > > importantly you have zero guarantee on the time frame that rate is > > > > > > going to be enforced for. > > > > > > > > > > Is there such a guarantee if clk_set_rate() is called? > > > > > > > > with clk_set_rate itself, no, but... > > > > > > > > > > It's simply going to change the rate as a one-off thing, and if > > > > > > there's the next millisecond someone else is going to change its rate > > > > > > one way or another, it's going to do so and you won't have any > > > > > > notification. > > > > > > > > You can get notified, and you can use clk_set_rate_exclusive if you > > > > *really* want to enforce it. > > > > > > Is the conclusion then we should go back to relying on the clock-frequency > > > property? > > clock-frequency or link-frequencies. link-frequencies seems to be a > better fit here, but we don't really have the choice for older > bindings. Alright, so since this is a new binding, let's aim for link-frequencies then. I don't think I understand what they look like on the driver side. Do you know an example of what a link-frequencies based driver would look like? > > > > This has been discussed multiple times over the years, and I don't really > > > disagree with the above. The frequency is typically indeed hand-picked for > > > the hardware, and no other frequency should be used in any circumstances. > > > > > > No sensor driver I've seen has used clk_set_rate_exclusive() but I guess > > > they should. The absence of practical problems has been probably because of > > > two factors; firstly, these are typically clocks dedicated to the sensors > > > and secondly, good luck. > > My point was that at least with handling the clock rate within the > driver (as opposed to assigned-clock-rates) you have multiple options > in dealing with changing colck rates / parents (Modelling the sensor > clock as a clock itself, using clk_set_rate_exclusive, using a > notifier, etc).. Some are more intrusive to the rest of the system > than others (especially clk_set_rate_exclusive), so I'm not really > advocating for any here, but we should make sure we have them in the > first place. > > Maxime
