Re: [PATCH v2 8/8] habanalabs: enable 64-bit DMA mask in POWER9
From: Oded Gabbay
Date: Wed Jun 12 2019 - 02:30:47 EST
On Wed, Jun 12, 2019 at 1:53 AM Benjamin Herrenschmidt
> On Tue, 2019-06-11 at 20:22 +0300, Oded Gabbay wrote:
> > > So, to summarize:
> > > If I call pci_set_dma_mask with 48, then it fails on POWER9. However,
> > > in runtime, I don't know if its POWER9 or not, so upon failure I will
> > > call it again with 32, which makes our device pretty much unusable.
> > > If I call pci_set_dma_mask with 64, and do the dedicated configuration
> > > in Goya's PCIe controller, then it won't work on x86-64, because bit
> > > 59 will be set and the host won't like it (I checked it). In addition,
> > > I might get addresses above 50 bits, which my device can't generate.
> > >
> > > I hope this makes things more clear. Now, please explain to me how I
> > > can call pci_set_dma_mask without any regard to whether I run on
> > > x86-64 or POWER9, considering what I wrote above ?
> > >
> > > Thanks,
> > > Oded
> > Adding ppc mailing list.
> You can't. Your device is broken. Devices that don't support DMAing to
> the full 64-bit deserve to be added to the trash pile.
Hmm... right know they are added to customers data-centers but what do I know ;)
> As a result, getting it to work will require hacks. Some GPUs have
> similar issues and require similar hacks, it's unfortunate.
> Added a couple of guys on CC who might be able to help get those hacks
> It's still very fishy .. the idea is to detect the case where setting a
> 64-bit mask will give your system memory mapped at a fixed high address
> (1 << 59 in our case) and program that in your chip in the "Fixed high
> bits" register that you seem to have (also make sure it doesn't affect
> MSIs or it will break them).
MSI-X are working. The set of bit 59 doesn't apply to MSI-X
transactions (AFAICS from the PCIe controller spec we have).
> This will only work as long as all of the system memory can be
> addressed at an offset from that fixed address that itself fits your
> device addressing capabilities (50 bits in this case). It may or may
> not be the case but there's no way to check since the DMA mask logic
> won't really apply.
Understood. In the specific system we are integrated to, that is the
case - we have less then 48 bits. But, as you pointed out, it is not a
generic solution but with my H/W I can't give a generic fit-all
solution for POWER9. I'll settle for the best that I can do.
> You might want to consider fixing your HW in the next iteration... This
> is going to bite you when x86 increases the max physical memory for
> example, or on other architectures.
Understood and taken care of.