Re: [PATCH 2/2] iommu/omap: Add support for configuring dsp iommus on DRA7xx

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



* Suman Anna <s-anna@xxxxxx> [150724 09:27]:
> Hi Tony,
> 
> On 07/23/2015 11:30 PM, Tony Lindgren wrote:
> > * Suman Anna <s-anna@xxxxxx> [150723 09:25]:
> >> Hi Tony,
> >>
> >> On 07/23/2015 02:24 AM, Tony Lindgren wrote:
> >>> * Suman Anna <s-anna@xxxxxx> [150722 09:25]:
> >>>> On 07/22/2015 12:26 AM, Tony Lindgren wrote:
> >>>>>
> >>>>> I don't like using syscon for tinkering directly with SoC registers.
> >>>>
> >>>> This is not a SoC-level register, but a register within a sub-module of
> >>>> the DSP processor sub-system. The DSP_SYSTEM sub-module in general is
> >>>> described in Section 5.3.3 of the TRM [1], and it implements different
> >>>> functionalities like the PRCM handshaking, wakeup logic and DSP
> >>>> subsystem top-level configuration. It is a module present within the DSP
> >>>> processor sub-system, so can only be accessed when the sub-system is
> >>>> clocked and the appropriate reset is released.
> >>>
> >>> OK so if it's specific to the DSP driver along the lines of sysc and
> >>> syss registers.
> >>
> >> There will be those registers too within the MMU config register space,
> >> even for DRA7xx MMUs. This is different, think of it like a register in
> >> the Control module except that it is present within the DSP sub-system
> >> instead of at the SoC level.
> > 
> > And what is taking care of pm_runtime_get here to ensure the module
> > is powered and clocked?
> 
> pm_runtime_get_sync is indeed getting invoked, its just the current
> patch doesn't include the code block that invokes it. The function that
> invokes omap2_iommu_enable does so after a call to the
> pm_runtime_get_sync() call.

OK 

> > I think you are missing a layer here and it's the Linux kernel side
> > device driver for DSP that initializes things.
> 
> We already have separate drivers for MMUs (omap-iommu) and the processor
> management (omap-rproc). The former manages all the low-level
> programming sequences for the MMUs, while the latter manages the
> low-level reset etc, and is a client user of the respective IOMMU
> device. Both integrate into the respective frameworks. The IOMMU API
> invocations are handled in the remoteproc core, with the OMAP remoteproc
> driver publishing itself as having an MMU with the remoteproc core. The
> IOMMU API invoke the appropriate iommu_ops.
> 
> You can lookup the functions rproc_enable_iommu()/rproc_disable_iommu()
> in the remoteproc core. The IOMMU enabling sequences happen within the
> iommu_attach_device() API. The call flow is
> iommu_attach_device()->omap_iommu_attach_dev()->omap_iommu_attach()->
> iommu_enable()->
>    omap_device_deassert_hardreset, pm_runtime_get_sync, omap2_iommu_enable.

OK. The thing to check here is that you have a separate device driver
for each sysc/syss device registers. This is because each hardware module
can be independently clocked and idled. Otherwise things will break at
some point. And no "things are configured for autoidle" or "we're not
doing PM is not a good excuse here" :)

Can you please check that? If the remoteproc driver and iommu driver
are tinkering with registers in separate hardware modules, we have
a layering violation. My guess is that we have at least two hardware
modules involved here, one for the iommu and one for the DSP device.
 
> >>> Typically we handle these registers by mapping them to the PM runtime
> >>> functions for the interconnect so we can reset and idle the hardware
> >>> modules even if there is no driver, see for example
> >>> omap54xx_mmu_dsp_hwmod.
> >>
> >> I haven't yet submitted the DRA7xx hwmods, but they will look almost
> >> exactly like the OMAP5 ones. The reset and idle on these are in general
> >> not effective at boot time since these are also controlled by a
> >> hard-reset line, so that's left to the drivers to deal with it through
> >> the omap_device_deassert_hardreset API.
> > 
> > If the MMU configuration is one time init, it may make sense to add
> > it to the hwmod reset function. However, if the Linux kernel side
> > driver needs to configure things depending on the DSP firmware, it
> > should be done in the kernel side device driver.
> 
> The MMU configuration comes into picture whenever the remoteproc driver
> is being booted and shut down, and also during suspend (no support for
> this yet in mainline on OMAP rproc). Today, the hwmod
> _enable/_idle/reset functions alone are not enough to power on the OMAP
> remoteproc/iommus. We need sequencing calls to both
> omap_device_assert/_deassert_hardreset (done through pdata-quirks) and
> pm_runtime API to achieve this.

Right and that's why I'm worried that we have multiple hardware modules
involved :)  
 
> >>>>> We should use some Linux generic framework for configuring these
> >>>>> bits to avoid nasty dependencies between various hardware modules
> >>>>> on the SoC.
> >>>>>
> >>>>> What does DSP_SYS_MMU_CONFIG register do? It seems it's probably
> >>>>> a regulator or a gate clock? If so, it should be set up as a
> >>>>> regulator or a clock and then the omap-iommu driver can just
> >>>>> use regulator or clcok framework to request the resource.
> >>>>
> >>>> No, its neither. It is a control bit that dictates whether the
> >>>> processor/EDMA addresses go through the respective MMU or not. The
> >>>> register currently has 4 bits (bit 0 in each nibble), one each for
> >>>> enabling each MMU and requesting an MMU abort on each. The MMU
> >>>> integration and enablement notes are detailed in Section 5.3.6 of the
> >>>> TRM [1], and the DSP_SYS_MMU_CONFIG register layout is in Table 5-28
> >>>> (Page 1641).
> >>>
> >>> OK yeah seems like it should be handled by the DSP driver during
> >>> probe after doing pm_runtime_get. Then the driver can configure
> >>> things like IOMMU and load firmware. Or am I missing something here?
> >>
> >> The DSP (remoteproc) driver uses the generic IOMMU API, and all the
> >> IOMMU enabling sequence is transparent to the DSP driver. So, any
> >> configuration/enabling sequence specific to the IOMMU has to be handled
> >> within the respective IOMMU platform driver that gets integrated into
> >> the IOMMU API.
> > 
> > To me it seems you still need a DSP kernel driver to manage the
> > DSP hardware for PM runtime. I don't see how iommu and remoteproc
> > alone would be sufficient here. You end up sprinkling DSP driver
> > specific code to the iommu and remoteproc layers.
> 
> I am not sure what you envision to be in this DSP kernel driver, because
> all the needed functionality is already handled by existing drivers. The
> OMAP remoteproc driver is the kernel driver to deal with the device
> management for both IPUs and DSPs on OMAP4, OMAP5 and DRA7 (DT support
> to be added).

OK maybe check the syss/sysc registers involved here for each hardware
module here and which driver tinkers with which registers? This will
make things a lot easier in the long run for sure.

Regrads,

Tony

--
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



[Index of Archives]     [Linux Arm (vger)]     [ARM Kernel]     [ARM MSM]     [Linux Tegra]     [Linux WPAN Networking]     [Linux Wireless Networking]     [Maemo Users]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite Trails]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux