Hi Dave, On 03/10/2015 07:59 PM, Dave Gerlach wrote: > On 03/10/2015 12:36 PM, Grygorii Strashko wrote: >> On 03/06/2015 07:45 PM, Tony Lindgren wrote: >>> * Dave Gerlach <d-gerlach@xxxxxx> [150306 09:28]: >>>> On 03/05/2015 06:41 PM, Tony Lindgren wrote: >>>>> * Tony Lindgren <tony@xxxxxxxxxxx> [150305 12:24]: >>>>>> * Dave Gerlach <d-gerlach@xxxxxx> [150305 11:53]: >>>>>>> On 03/05/2015 12:49 PM, Tony Lindgren wrote: >>>>>>>> * Paul Walmsley <paul@xxxxxxxxx> [150305 10:16]: >>>>>>>>> On Thu, 5 Mar 2015, Dave Gerlach wrote: >>>>>>>>> >>>>>>>>>> Introduce a dt property, ti,no-init, that prevents hwmod initialization. >>>>>>>>>> Even if a dt node is marked as disabled, hwmod still at least enables >>>>>>>>>> the hwmod and programs the sysconfig before attempting to idle it at >>>>>>>>>> boot. If an IP has been disabled by the hardware configuration on a >>>>>>>>>> platform, this will cause a hang due to writing to inactive registers. >>>>>>>>>> This property prevents that from happening by marking the hwmod as >>>>>>>>>> _HWMOD_STATE_DISABLED during init. >>>>>>>>> >>>>>>>>> I'm kind of wondering if hwmod should even touch a device if it's marked >>>>>>>>> as disabled in the DT. Tony, what do you think? >>>>>>>> >>>>>>>> Well nothing happens if a device is status = "disabled". No dev entry >>>>>>>> gets created for it at all and hwmod won't have any data for the device >>>>>>>> populated. The only way hwmod code could see that device if the device >>>>>>>> gets it's data from the legacy omap_hwmod_*_data.c instead of DT. >>>>>>>> >>>>>>> >>>>>>> We still need this for the sysconfig programming, correct? hwmod programs that >>>>>>> regardless of dt status and then idles the IP, >>>>>> >>>>>> Well hwmod does not even know about the IP IO addresses if it's marked >>>>>> with status = "disabled".. Which IP are you having problems with? >>>>>> >>>>>>> which is why I needed the ti,no-init for the epos evm. It isn't just a >>>>>>> matter of we shouldnt write to it because we don't want to use it; we >>>>>>> can't write to it because the module is held off so it causes an >>>>>>> external abort if we do. >>>>>> >>>>>> Well hard to say not knowing which module this is.. Pretty much all >>>>>> the modules have drivers and the driver just does pm_runtime_get() >>>>>> on it? >>>>> >>>>> Heh OK this thread is about the RTC driver, so I assume that's the >>>>> problem :) So if you set the rtc to status = "disabled" how can the >>>>> hwmod code do anything as AFAIK it won't even get the rtc IO address? >>>>> >>>>> Or am I missing something here? >>>> >>>> Perhaps I am mistaken, but from what I understand, all hwmods have _init and >>>> _setup called on them, and all hwmods read the IO address regardless of DT >>>> status at this point with _init_mpu_rt_base. In _setup, _setup_reset gets called >>>> which calls _enable for the hwmod, and this calls both _enable_sysc and >>>> _update_sysc_cache which touch the sysconfig register of the IP. And that's the problem :) What ePar says: “disabled” - Indicates that the device is not presently operational, but it might become operational in the future (for example, something is not plugged in, or switched off). and current OF implementation will not register corresponding device in DD core and, as result, drivers will be never ever bound with these devices. As i said before, such devices are invisible/absent/not present, so there are no reasons to touch them at all in Kernel, because result is unpredictable in general. And there are no guarantee that there will be no other issues, even if you'll fix this particular case. >>> >>> Oh OK, I think you're right. I was thinking of omap_device_build_from_dt(), >>> sorry. Looks like the hwmod IO address data does get populated even >>> for status = "disabled" although the dev entry won't get created and >>> omap_device_build_from_dt() never gets called. >>> >>>> Normally this is fine regardless of whether or not we are using an IP because >>>> the module will wake up for a moment, have it's sysc programmed, and then be put >>>> back to sleep later, potentially never to be woken again if we bind no driver >>>> for it, which is fine for 99% of cases. In the case of am43x epos evm, you can ^^ if status=“disabled”/"failed" there will be no device and no driver will have been bound :) >>>> take the same piece of silicon that will boot happily on the gp evm with the rtc >>>> hwmod in place and it will hang during boot on the epos evm because the board >>>> uses a pin to hold the RTC IP in reset. There is no way to detect this in >>>> software, the module can be turned on as normal using the clk_ctrl, but if you >>>> touch any of the IP registers you get an abort. >>> >>> OK sounds like some dts property is needed to signal this. Seems It might be much more simple to just remove RTC device from DT, as it's disabled and unaccessible == not present: - add am4372-rtc.dtsi and put RTC device definition there - include am4372-rtc.dtsi in board files where allowed >> >> As I understand, there is device "RTC" and this device has status 'disabled', >> so there is reasonable question why do we need to touch it at all? >> The HWMOD is some kind of SoC description, which was needed before DT. >> Now with DT in place it seems unreasonable to touch any IP block which: >> - are not defined in DT >> - has status 'disabled'. >> in above cases the u-boot has to dial with IP block if it's invisible for Kernel. >> >> The HWMOD framework intended to reset and put in some predefined states IPs which are >> visible for the Kernel, so then proper driver can start working with IP or IP will be >> kept in some low-power state if there is no driver. > > Currently we still need this hwmod behavior as not all IPs are in an ideal state > for PM by default. Not sure if you saw my last email in this thread but I gave > an example for USB hwmod: > > If no USB driver is bound even just letting omap_device idle it using the USB_ > CLKCTRL in the PRCM will not work properly, because the USB on am33xx expects > it's SYSCONFIG STANDBY_MODE to be software supervised, the default state > (SMART_STANDBY) does not let the IP go into standby so the IP gets stuck in a > partially off state but does not actually shut down as expected. > > So we need to be able to touch the dt nodes even if they are marked as disabled > because we need to get the IO address to have access to the SYSCONFIG register. > I am not trying to introduce any major new changes to omap_hwmod layer, I am > just trying to allow it to work in more situations. Oh yes! And this happen just due to historical reasons ;) The HWMOD is ancient thing :P which was created at time when: - all OMAP's code was built-in - there were no DT in place. And it actually was created to describe configuration and integration of HW. But now there is DT which is intended for the same purposes and HWMOD data is still here (as i know) just because process of migration to DT is not fully completed yet. More over, now we have multiplatform; restriction to put SoC's erratas and hacks in kernel; and all drivers going to be converted to loadable modules. In my opinion, this is not good when we have device marked as 'disabled', but Kernel still tries to access them. More over, we need to completely remove device's definitions from DT to prevent Kernel of touching such devices. So, let's not make things worse than they are :( and just remove RTC device where it's not needed. > > This flag would only exist to PREVENT any hwmod action from happening in certain > cases where we cannot even touch the IP for whatever reason (in the first case > that this patch intends to be used for, the RTC IP on am43x-epos-evm which is > disabled by the board configuration). > >> >> Such kind of changes really confused :(, because when I'm looking on DT and see - status='disabled' >> I suppose that such IP will not be touched by Kernel. But it's NOT really 'disabled'. >> It's kinda 'partially disabled' or 'not present' or 'unused' .. >> (if someone will decide to continue with such approach - it could be more >> useful to define new value for 'status' parameter :P). >> Also, just imaging that OMAP kernel is used with Xen as DomU and how hard will >> it be to find and cut off all such tricks :( >> >> Did I miss something? >> >>> >>>> So we need to prevent this from happening but of course we can't selectively >>>> choose when the rtc hwmod gets added based on which board we are using so it >>>> seemed a DT flag was appropriate to indicate that we do not want to init the rtc >>>> IP at all only on this board. >>>> >>>> Without this flag in place but with the rtc hwmod added, the am43x-epos-evm >>>> fails booting with an imprecise abort. >>> >>> OK thanks for explaining it. I'm fine with this patch, Paul may have >>> other issues. The other option would be to use status = "disabled" to >>> not touch the device at all like Paul suggested. I wonder if that's >>> going to break PM on some devices though.. That how it should be as for me, but unfortunately, it could break PM :(, because It's common assumption now that Kernel will reset & disable ALL OMAP's devices regardless of DT. It could be possible if it will be handled properly outside of the kernel - in u-boot for example. regards, -grygorii -- To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
