[PATCH Xenial 4.4] UBUNTU: [Config] CONFIG_ZONE_DEVICE=y for amd64

Tim Gardner tcanonical at mail.tpi.com
Fri Jan 29 16:50:16 UTC 2016



On 01/27/2016 12:23 PM, Mark wrote:
> On Sun, January 24, 2016 21:42, Tim Gardner wrote:
>> On 01/23/2016 10:47 AM, Mark wrote:
>>> On Sat, January 23, 2016 14:13, Tim Gardner wrote:
>>>> In the event that upstream does not come up with a solution, how about
>>>> if we leave ZONE_DMA enabled in the low-latency flavour ? After all, it
>>>> seems most of the devices that this affects are audio related (except
>>>> for the parallel port).
>>>
>>> Another point of information. According to a post in fa.linux.kernel
>>> from
>>> 2007, some Broadcom NICs may need ZONE_DMA.
>>>
>>> https://groups.google.com/forum/#!original/fa.linux.kernel/bXnCP9QIlMI/V6pLHumoe0oJ
>>>
>>> "Just because you don't need ISA, doesn't mean you don't need
>>> ZONE_DMA. Various PCI cards have limitations on what addresses they
>>> can DMA to (24 bits on some old PCI sound cards iirc, and more recently
>>> a 30 bit DMA limitation on some Broadcom networking equipment.)"
>>>
>>> Another thread mentions the floppy controller, so systems with native
>>> floppy drives could be another casualty:
>>> https://groups.google.com/forum/#!original/fa.linux.kernel/rM_PYJc9ExE/brVe--Xr0FgJ
>>>
>>> Is there an easy way to grep the kernel source tree to get a list of all
>>> drivers which depend on ZONE_DMA? I tried that in the drivers
>>> subdirectory
>>> and came up with
>>> net/wireless/b43/dma.c
>>> net/wireless/b43legacy/dma.c
>>>
>>> Grepping for GFP_DMA returned many more results; someone more
>>> knowledgeable than me might be able to say whether any are relevant.
>>>
>>>
>>> Mark
>>>
>>
>> This seems to be a bit of an intractable problem, and I don't see an
>> upstream solution any time soon (if ever). Therefore, I've implemented a
>> compromise as detailed in
>> https://bugs.launchpad.net/ubuntu/xenial/+source/linux/+bug/1534647/comments/3
>>
>> Regardless of the impact on these legacy devices, I feel strongly that
>> NVDIMMs must be supported on amd64 servers for 16.04. Also, I do not
>> think this regression in support for legacy devices warrants a new amd64
>> flavour given the lowlatency compromise.
> 
> It seems there is some upstream discussion and an RFC patch (which did
> seem to fix the problem for me). That patch probably won't be committed
> as-is, but hopefully a solution will be found so this issue ceases to be a
> problem.
> 
> http://www.spinics.net/lists/kernel/msg2170767.html
> 
> I don't know whether any solution will get backported to kernel 4.4, but
> if not, perhaps backporting it could be one of the Ubuntu-specific kernel
> changes.
> 
> 
> Mark
> 
> 

Well, that thread does hold some promise. Perhaps you could drop a note
on this list when something gets merged into a maintainer tree or even
linux-next.

rtg
-- 
Tim Gardner tim.gardner at canonical.com




More information about the kernel-team mailing list