On Tue, 2013-08-13 at 14:30 -0600, Bjorn Helgaas wrote:
> [+cc James in case he has opinions on the DMA mask question]
>
> On Tue, Aug 13, 2013 at 10:12 AM, Chris Metcalf <cmetcalf@xxxxxxxxxx> wrote:
> > (Trimming the quoted material a little to try to keep this email under control.)
> >
> > On 8/12/2013 4:42 PM, Bjorn Helgaas wrote:
> >> On Mon, Aug 12, 2013 at 1:42 PM, Chris Metcalf <cmetcalf@xxxxxxxxxx> wrote:
> >>> On 8/9/2013 6:42 PM, Bjorn Helgaas wrote:
> >>>> OK, so physical memory in the [3GB,4GB] range is unreachable via DMA
> >>>> as you describe. And even if DMA *could* reach it, the CPU couldn't
> >>>> see it because CPU accesses to that range would go to PCI for the
> >>>> memory-mapped BAR space, not to memory.
> >>> Right. Unreachability is only a problem if the DMA window overlaps [3G, 4G], and since the 64-bit DMA window is [1TB,2TB], the whole PA space can be reached by 64-bit capable devices.
> >> So the [0-1TB] memory range (including [3GB-4GB]) is reachable by
> >> 64-bit DMA to bus addresses [1TB-2TB]. But if the CPU can't see
> >> physical memory from [3GB-4GB], how is it useful to DMA there?
> >
> > Sorry, looking back I can see that the thread is a little confusing.
> > The CPU can see the whole PA space. The confusion comes from the BAR space
> > in [3GB, 4GB].
> >
> > On Tile, we define the CPU memory space as follows:
> >
> > [0, 1TB]: PA
> > [1TB + 3GB, 1TB + 4GB]: BAR space for RC port 0, in [3GB, 4GB]
> > [1TB + 3GB + N*4GB, 1TB + (1 + N)*4GB]: BAR space for RC port N, in [3GB, 4GB]
> >
> > The mapping from [1TB + 3GB + N*4GB, 1TB + (1 + N)*4GB] to [3GB, 4GB] is done by a
> > hardware PIO region, which generates PCI bus addresses in [3GB, 4GB] for MMIOs to
> > the BAR space.
>
> OK, I think I get it now. CPU address space:
> [0, 1TB]: physical memory
> [1TB + 3GB, 1TB + 4GB]: translated to bus address [3GB, 4GB] under RC port 0
> [1TB + 3GB + N*4GB, 1TB + (1 + N)*4GB]: translated to bus address
> [3GB, 4GB] under RC port N
>
> Bus address space:
> [0, 3GB]: 32-bit DMA reaches physical memory [0, 3GB]
> [3GB, 4GB]: 32-bit DMA (peer-to-peer DMA under local RC port, I guess?)
> [1TB, 2TB]: 64-bit DMA mapped via IOMMU to physical memory [0, 1TB]
>
> I guess the problem is that 32-bit DMA can't reach physical memory
> [3GB, 4GB], so you're using bounce buffers so the bus address is in
> [0, 3GB]. That makes sense, and I don't see another possibility other
> than just throwing away the [3GB, 4GB] range by leaving it out of the
> kernel allocator altogether, or using hardware (which tilegx probably
> doesn't have) to remap it somewhere else.
This is remarkably familiar. I think almost every system on earth has a
configuration similar to this. On PARISC, the top 256MB of memory on a
32 bit system is reserved for I/O access and is designated as "F Space".
What is unusual is that you seem to have responding memory behind the F
Space which is accessible to some bus entities. On PARISC 32 bit, the
memory is just lost (inaccessible) on 64 bit, it's remapped above 32GB
(and the low F Space window expanded to 1GB).
> So it seems like just a question of how you wrap this all up in
> dma_ops, and *that* is all arch stuff that I don't have an opinion on.
>
> >>> Unfortunately, the Megaraid driver doesn’t even call pci_set_consistent_dma_mask(dev, DMA_BIT_MASK(32)).
> >> If the Megaraid driver needs that call, but it's missing, why wouldn't
> >> we just add it?
> >
> > The Megaraid driver doesn’t strictly need that call on other platforms, because
> > by default the device coherent_dma_mask is DMA_BIT_MASK(32) and the consistent
> > memory pool doesn’t come from the bounce buffers on most other platforms.
> >
> > Of course, for the sake of correctness, this call should be added across all platforms.
> > ...
>
> > What is unique about Tile is that the PCI drivers must explicitly declare
> > its DMA capability by calling pci_set_dma_mask() and pci_set_consistent_dma_mask().
>
> It looks like the reason you need drivers to explicitly call
> pci_set_dma_mask() and pci_set_consistent_dma_mask() is because you
> have hooks in those functions to tweak the dma_ops, even though the
> mask itself might not be changed.
>
> That doesn't sound like a robust solution: we have well-known defaults
> for the mask sizes, and I don't think it's reasonable to expect
> drivers to explicitly set the mask even if they are happy with the
> defaults (though Documentation/DMA-API-HOWTO.txt does say that being
> explicit is good style). I'm afraid you'll just keep tripping over
> drivers that don't work on tilegx because they don't set the mask.
Right, it's not a robust solution at all. A DMA mask is just that: an
accessibility mask. The founding assumption is that an address line is
either connected or not, which is why the mask works. What you have is
two different classes of memory: 0-3GB which is usable for I/O and 3-4GB
which isn't. Surely what you need to do is implement ZONE_DMA32? which
stretches from 0-3GB, which means all kmallocs in the driver will be in
the right range. Then you need a bounce pfn at 3GB which means that
user space which gets the 3-4GB is bounced. There's a magic pfn in
BLK_BOUNCE_HIGH that was designed for this, but I'm not sure the design
contemplated BLK_BOUNCE_HIGH being different for 64 and 32 bit.
James
��.n��������+%������w��{.n�����{���"�)�� jg��� �����ݢj����G���� ���j:+v���w�m������w�������h�����٥
