by Peter De Schrijver

[permalink] [raw]

Subject: Re: [PATCH v7 3/5] misc: fuse: Add efuse driver for Tegra

On Wed, Jun 11, 2014 at 05:58:28PM +0200, Stephen Warren wrote:
> On 06/11/2014 09:25 AM, Peter De Schrijver wrote:
> > On Wed, Jun 11, 2014 at 02:47:31PM +0200, Mikko Perttunen wrote:
> >> On 05/06/14 16:09, Peter De Schrijver wrote:
> >> ...
> >>> +int tegra_fuse_readl(u32 offset, u32 *val)
> >>> +{
> >>> + if (!fuse_readl)
> >>> + return -ENXIO;
> >>> +
> >>> + *val = fuse_readl(offset);
> >>> +
> >>> + return 0;
> >>> +}
> >>> +
> >>
> >> -EPROBE_DEFER would be a better error value, so that drivers can work
> >
> > Ok.
> >
> >> even if they are initially probed before the fuse driver. Of course, if
> >> the fuse initialization is moved into machine init then this is a non-issue.
> >
> > The exported function will always be initialized later because on Tegra20 it
> > requires APB DMA to be available. If you read the fuses directly, the system
> > sometimes hangs.
>
> That's not true in the current code. IIRC, the bug was that *if* an APB
> DMA access to anything and a CPU access to the fuses happen at the same
> time, then there can be a hang. As such, the current fuse code accesses
> the fuses directly (without potential for a hang) if the APB DMA driver
> is not available, but once the driver becomes available, it reads the
> fuses through DMA instead. Does the new code not do that?
>

I'm not so sure about that. I have seen the hang when dumping all fuses using
sysfs in an otherwise idle system booted from initrd. I don't think there
should be any APB DMA activity going on then?

Cheers,

Peter.

2014-06-11 16:29:27

by Stephen Warren

[permalink] [raw]

Subject: Re: [PATCH v7 3/5] misc: fuse: Add efuse driver for Tegra

On 06/11/2014 10:19 AM, Peter De Schrijver wrote:
> On Wed, Jun 11, 2014 at 05:58:28PM +0200, Stephen Warren wrote:
>> On 06/11/2014 09:25 AM, Peter De Schrijver wrote:
>>> On Wed, Jun 11, 2014 at 02:47:31PM +0200, Mikko Perttunen wrote:
>>>> On 05/06/14 16:09, Peter De Schrijver wrote:
>>>> ...
>>>>> +int tegra_fuse_readl(u32 offset, u32 *val)
>>>>> +{
>>>>> + if (!fuse_readl)
>>>>> + return -ENXIO;
>>>>> +
>>>>> + *val = fuse_readl(offset);
>>>>> +
>>>>> + return 0;
>>>>> +}
>>>>> +
>>>>
>>>> -EPROBE_DEFER would be a better error value, so that drivers can work
>>>
>>> Ok.
>>>
>>>> even if they are initially probed before the fuse driver. Of course, if
>>>> the fuse initialization is moved into machine init then this is a non-issue.
>>>
>>> The exported function will always be initialized later because on Tegra20 it
>>> requires APB DMA to be available. If you read the fuses directly, the system
>>> sometimes hangs.
>>
>> That's not true in the current code. IIRC, the bug was that *if* an APB
>> DMA access to anything and a CPU access to the fuses happen at the same
>> time, then there can be a hang. As such, the current fuse code accesses
>> the fuses directly (without potential for a hang) if the APB DMA driver
>> is not available, but once the driver becomes available, it reads the
>> fuses through DMA instead. Does the new code not do that?
>>
>
> I'm not so sure about that. I have seen the hang when dumping all fuses using
> sysfs in an otherwise idle system booted from initrd. I don't think there
> should be any APB DMA activity going on then?

Hmm. Perhaps I'm misremembering the trigger for the bug then. Still, the
existing code works as I described. Perhaps that's dangerous and it
shouldn't though. Either way, I think we should have a standalone commit
that removes tegra_apb_readl_using_dma()'s fallback to
tegra_apb_readl_direct(), so any behaviour change that causes a problem
can be bisected easily.