On Wed, Aug 14, 2013 at 3:49 AM, Joerg Roedel <joro@xxxxxxxxxx> wrote:
> On Thu, Aug 08, 2013 at 11:28:44AM -0700, Grant Grundler wrote:
>> I can't speak to the previous BUG_ON(). I believe the EADDRESSINUSE
>> failures could be either WARN_ON or BUG_ON. This condition is
>> clearly a bug in the generic IOMMU allocator and I think that's why
>> KyongHo Cho used BUG_ON.
>>
>> Handing out duplicate addresses will generally lead to some sort of
>> data corruption or other fault depending on how robust the underlying
>> device drivers are written. So my preference is a BUG_ON to
>> immediately flag this condition instead of hoping a device driver will
>> correctly handling the dma mapping failure (Some do, most currently
>> don't).
>>
>> WARN_ON() + return -EADDRESSINUSE would be a good alternative.
>
> Even if it is a real BUG condition, I don't think it is worth to stop
> execution at this point. It makes debugging harder and the system less
> reliable. I prefer to go with the WARN_ON and an error return value.
I'm ok with WARN_ON and an error return value. This is "valid"
behavior. I expect this bug to never happen but if and when it does,
I want a clear symptom (e.g. WARN_ON) that it happened.
My concern is that historically, drivers did not get an error return
value on failure:
ftp://193.166.3.4/pub/linux/kernel/v2.3/patch-html/patch-2.3.47/linux_Documentation_DMA-mapping.txt.html
or later:
https://www.kernel.org/pub/linux/kernel/people/marcelo/linux-2.4/Documentation/DMA-mapping.txt
And thus, some drivers don't check or attempt to handle mapping
failures based on this existing code. Here is a recent example:
http://comments.gmane.org/gmane.linux.network/272969
I hope very few or none of those exist since Neil Horman demonstrated
"dma debugging" can flag this behavior.
Just for fun, I'll include this link : (apperently 2003 was a good
year for DMA talks :)
http://ols.fedoraproject.org/OLS/Reprints-2003/LinuxSymposium2003-2side.pdf
(three talks on DMA issues)
thanks
grant
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
