[PATCH 0/6] omap: iommu: hwmod support and code reorganization

[Cousson, Benoit]

On 11/7/2010 10:43 AM, Ramirez Luna, Omar wrote:
> On Sat, Nov 6, 2010 at 1:31 PM, Cousson, Benoit<b-cousson at ti.com>  wrote:
>> s/ducati/ipu/
>> s/tesla/dsp/
>>
>> Please do not use internal codename for the changelog or even for the code.
>
> I picked this terminology from the driver, I didn't want to cause
> confusion, agree to the change.
>
>> On 11/5/2010 9:19 PM, Ramirez Luna, Omar wrote:
>>>
>>> Boot tested on omap 3430 and 3630, functionality on iva2 only.
>>>
>>> Introduced hwmod data and support for omap3 (iva2 mmu&    isp mmu) and
>>> omap4 (ducati mmu&    tesla mmu).
>>>
>>> Minor cleanup due to this changes.
>>>
>>> There is an issue (present without patches too):
>>>
>>> Doing a cycle of insmod-rmmod to iommu module and then full
>>> insmod of bridgedriver (with all dependencies) causes the next
>>> read of iommu registers to dump an unhandled fault log; this,
>>> because when rmmod of iommu is called it tries to clean the
>>> iommu tables and flush any old entry prior to removing the driver,
>>> however these reads/writes are attempted while the MMU is under
>>> reset and will timeout on the L3 IVA target agent, which will leave
>>> MMU unusable until the proper restore sequence to clear this L3
>>> error is executed.
>>>
>>> Fix would be to move page table allocation to iommu get and its
>>> correspondent free to iommu put, however it will fall entirely
>>> under iommu user responsibility, as it needs to clear the mmu
>>> reset bit, before calling iommu functions that read/write into
>>> mmu registers (which should apply only for first boot or on
>>> baseimage reload); trading this restriction in order to keep
>>> the same generic iommu code for all omap iommu devices.
>>>
>>> This issue has been seen on omap3 iva2 mmu, same restriction should
>>
>> I'm still not really understanding that issue.
>> In theory, the reset is deasserted when the omap_device is enabled and will
>> be re-asserted when the omap_device will be disable.
>
> Probably I'm missing the bits telling hwmod to handle its reset bit, I
> will recheck.

After reading the other patches of that series, this is what I realized 
as well. You need the entries with hw reset information to make that works.

>
>> Why is the mmu already under reset when the iommu is trying to clean the
>> tables?
>
> However, the driver disables the device on a call named iommu_put
> (where, from your comment, reset should be re-asserted ), and cleans
> the TLBs and page tables on removal of the driver (after iommu_put).
>
>> Could please describe the complete sequence?
>
> I'm assuming the dependencies are installed.
>
> on boot up:
> 1. insmod iommu.ko
> - device_enable is not called, unless a call to iommu_get is given,
> this will mean that the driver has one user and thus device_enable is
> called (however right now it doesn't matter if you call iommu_get or
> not)
>
> 2. rmmod iommu.ko
> - remove function will try to cleanup the TLBs writing into MMU
> registers, since MMU is on reset it will trigger a timeout on its TA.
>
> 3. insmod bridgedriver.ko
> - it is dependent on iommu.ko so it should have been installed before.
> Whenever bridge calls iommu_get, the external abort will be caused due
> to a read to MMU.
>
> As I commented, cleaning the page tables could be moved from driver's
> remove function to iommu_put, if omap device enable/disable is
> handling the reset bit probably this issue shouldn't be hit.

Yep, I think so as well.

Regards,
Benoit