Re: [PATCH] intel-iommu: Quiesce devices before disabling IOMMU

[Takao Indoh <indou.takao@jp.fujitsu.com>]

(2013/09/09 18:07), David Woodhouse wrote:
> On Wed, 2013-08-21 at 16:15 +0900, Takao Indoh wrote:
>>
>> This causes problem on kdump. Devices are working in first kernel, and
>> after switching to second kernel and initializing IOMMU, many DMAR faults
>> occur and it causes problems like driver error or PCI SERR, at last
>> kdump fails. This patch fixes this problem.
> 
> I'm not sure I'd call this a fix.
> 
> If the driver is so broken that it cannot get the device working again
> after a fault, surely the driver needs to be fixed?

Yes,this problem may be solved by fixing driver. Actually megaraid sas
driver is recently fixed for this problem. (See commit 6431f5d7)

But I think root cause of this problem is initializing IOMMU while DMA
is still working, and I want to solve the root cause rather than
handling it in each driver, otherwise we have to fix driver each time we
find this kind of problem.

> 
> If the system is suffering an IRQ storm because device doesn't give up
> after the first few faults, then we should switch off the fault
> *reporting* for that device so that its faults get ignored (until it
> next actually sets up a DMA mapping, or something).

In such a case, yeah limiting messages is enough.

> 
> For the IOMMU code to reset individual devices, just because they still
> have an active DMA mapping even if they're not *doing* DMA, seems wrong.
> You'll even end up resetting devices just because they have an RMRR,
> won't you? (Although I wouldn't lose any sleep over that, I suppose. In
> fact it might be a *feature*... :)
 
Right, current code is resetting devices which *may* be doing DMA. The
ideal way is finding devices which are actually doing DMA and reset only
them but I don't know how we can do this, though I think current code
is sufficient.

Thanks,
Takao Indoh