Re: [PATCH v2 3/9] tty: serial: fsl_lpuart: handle EPROBE_DEFER for DMA

[Michael Walle <michael@walle.cc>]

Hi Rob, Hi Leo,

Am 2020-02-28 00:03, schrieb Rob Herring:
> On Thu, Feb 27, 2020 at 4:49 PM Li Yang <leoyang.li@nxp.com> wrote:
>> 
>> On Thu, Feb 27, 2020 at 4:35 PM Rob Herring <robh+dt@kernel.org> 
>> wrote:
>> >
>> > On Fri, Feb 21, 2020 at 11:48 AM Michael Walle <michael@walle.cc> wrote:
>> > >
>> > > The DMA channel might not be available at the first probe time. This is
>> > > esp. the case if the DMA controller has an IOMMU mapping.
>> > >
>> > > Use the new dma_request_chan() API and handle EPROBE_DEFER errors. Also
>> > > reorder the code a bit, so that we don't prepare the whole UART just to
>> > > determine that the DMA channel is not ready yet and we have to undo all
>> > > the stuff. Try to map the DMA channels earlier.
>> >
>> > Changing this means you never probe successfully if you boot a kernel
>> > with the DMA driver disabled (or it's IOMMU disabled). Some other
>> > drivers request DMA in open() and can work either way.

Oh, I see.

>> We got this exact issue previously with another driver.  When the

What driver is it? I've been working on the i2c-mxs.c driver which has
the same problem. Ie. its not working with DMA when the IOMMU is 
enabled.
Now that I've learned that dma_request_chan() will return EPROBE_DEFER
if the actual DMA driver is not available, I don't think there is any
trick like this there. There is no function which would be called late
except you'd do something like on the first master_xfer() try to request
the DMA channels. But I don't think that would be the way to go.

-michael

>> required DMA driver is disabled, the DMA framework cannot figure out
>> this situation and keeps returning EPROBE_DEFER.  I'm wondering if we
>> should update the DMA framework to use your deferred probe timeout
>> mechanism.  Is it still only used for debug purpose?
> 
> It's undergoing some rework ATM to not just be for debug. However,
> it's not really going to help you if you care about the console
> because waiting for the timeout will be too late to register the
> console.