[RFC,v2] ARC: allow to use IOC and non-IOC DMA devices simultaneously

The ARC HS processor provides an IOC port (I/O coherency bus
interface) that allows external devices such as DMA devices
to access memory through the cache hierarchy, providing
coherency between I/O transactions and the complete memory
hierarchy.

Some recent SoC with ARC HS (like HSDK) allow to select bus
port (IOC or non-IOC port) for connecting DMA devices in runtime.

With this patch we can use both HW-coherent and regular DMA
peripherals simultaneously.

For example we can connect USB and SDIO controllers through IOC port
(so we don't need to need to maintain cache coherency for these
devices manualy. All cache sync ops will be nop)
And we can connect Ethernet directly to RAM port (so we had to
maintain cache coherency manualy. Cache sync ops will be real
flush/invalidate operations)

Cache ops are set per-device and depends on "dma-coherent" device
tree property:
"dma_noncoherent_ops" are used if no "dma-coherent" property is
present (or IOC is disabled)
"dma_direct_ops" are used if "dma-coherent" property is present.

NOTE 1:
It works perfectly fine only if we don't have ZONE_HIGHMEM used
as IOC doesn't cover all physical memory. As for today It configured
to cover 1GiB starting from 0x8z (which is ZONE_NORMAL memory for
us). Transactions outside this region are sent on the non-coherent
I/O bus interface.
We can't configure IOC to cover all physical memory as it has several
limitations relating to aperture size and start address.

And if we get DMA buffer from ZONE_HIGHMEM memory we need to
do real flush/invalidate operations on that buffer, which is obviously
not done by "dma_direct_ops".

So I am not sure about "dma_direct_ops" using - probably we need to
create our special cache ops like "arc_ioc_ops" which will handle
ZONE_HIGHMEM case.

(BTW: current ARC dma_noncoherent_ops implementation also has same
problem if IOC and HIGHMEM are enabled.)

NOTE 2:
In this RFC only hsdk.dts changes are shown to reduce patch size.
AXS103 device tree changes are not shown.


Changes v1->v2 (Thanks to Vineet and Christoph):
 * Panic if both IOC and HIGHMEM_ZONE are enabled.
 * Move arch_setup_dma_ops to arch/arc/mm/dma.c
 * Tweak the boot printing about IOC
 * Print info about cache ops used for each device.
 * Refactor arch_setup_dma_ops

Signed-off-by: Eugeniy Paltsev <Eugeniy.Paltsev@synopsys.com>
---
 arch/arc/Kconfig                   |  1 +
 arch/arc/boot/dts/hsdk.dts         |  4 ++++
 arch/arc/include/asm/dma-mapping.h | 13 +++++++++++++
 arch/arc/mm/cache.c                | 20 +++++---------------
 arch/arc/mm/dma.c                  | 32 +++++++++++++++++++-------------
 5 files changed, 42 insertions(+), 28 deletions(-)
 create mode 100644 arch/arc/include/asm/dma-mapping.h

On Thu, Jun 28, 2018 at 05:14:52PM +0300, Eugeniy Paltsev wrote:
> And if we get DMA buffer from ZONE_HIGHMEM memory we need to
> do real flush/invalidate operations on that buffer, which is obviously
> not done by "dma_direct_ops".
> 
> So I am not sure about "dma_direct_ops" using - probably we need to
> create our special cache ops like "arc_ioc_ops" which will handle
> ZONE_HIGHMEM case.

FYI, I have a plan to merge dma_direct_ops and dma_noncoherent_ops
and make the decision to use cache flushing run time controllable
with a flag in struct device based on existing dynamic selection
in e.g. arm, arm64 and mips.  It will probably take a few more merge
windows to get there, but once it is done you should be able to take
easy advantage of it.

> +++ b/arch/arc/mm/dma.c
> @@ -19,6 +19,7 @@
>  #include <linux/dma-noncoherent.h>
>  #include <asm/cache.h>
>  #include <asm/cacheflush.h>
> +#include <linux/dma-mapping.h>

dma-noncoherent.h already includes dma-mapping.h, so this is not
required.

> -	if ((is_isa_arcv2() && ioc_enable) ||
> -	    (attrs & DMA_ATTR_NON_CONSISTENT))
> +	if (attrs & DMA_ATTR_NON_CONSISTENT)
>  		need_coh = 0;

need_coh is only used twice, it might be cleaner to remove the
variable and just open code the DMA_ATTR_NON_CONSISTENT check.
And for extra points remove the need_kvaddr variable as well.

Also you probably want to do the equivalent change in arch_dma_free
as well.

> +void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
> +			const struct iommu_ops *iommu, bool coherent)
> +{
> +	/*
> +	 * IOC hardware snoops all DMA traffic keeping the caches consistent
> +	 * with memory - eliding need for any explicit cache maintenance of
> +	 * DMA buffers - so we can use dma_direct cache ops.
> +	 */
> +	if (is_isa_arcv2() && ioc_enable && coherent) {
> +		set_dma_ops(dev, &dma_direct_ops);
> +		dev_info(dev, "use dma_direct_ops cache ops\n");
> +	} else {
> +		set_dma_ops(dev, &dma_noncoherent_ops);
> +		dev_info(dev, "use dma_noncoherent_ops cache ops\n");
> +	}
> +}

Note that due to your use of asm-generic/dma-mapping.h we already
default to dma_noncoherent_ops if no per-device ops is set.  So we
could skip the else branch here I think.

On 06/28/2018 07:14 AM, Eugeniy Paltsev wrote:
> The ARC HS processor provides an IOC port (I/O coherency bus
> interface) that allows external devices such as DMA devices
> to access memory through the cache hierarchy, providing
> coherency between I/O transactions and the complete memory
> hierarchy.

You mention IOC port in first line, but the change has nothing to do with it -
this is confusing.
Make this less verbose and just expand on the subject you have - global vs. per
device dma settings, without going into all the internal details.

And again going back to my original point, you are changing the semantics (ioc and
non ioc to co-exist) but everything still goes thru IOC port. That is different
that what we had before - everything goes through IOC port and ioc dma ops ONLY if
IOC existed.

I understand the need to isolate the 2 things and that is fine. But then what you
should do is
1a. introduce the per ioc dma settings in ARC - and not touch the DTBs.
1b. introduce the IO port settings in platform code
2. Now switch the DT binding to use the per device ioc

And any panic etc for IOC + highmem etc should be a separate patch - so we can
easily revert it when time is right !

>
> Some recent SoC with ARC HS (like HSDK) allow to select bus
> port (IOC or non-IOC port) for connecting DMA devices in runtime.
>
> With this patch we can use both HW-coherent and regular DMA
> peripherals simultaneously.
>
> For example we can connect USB and SDIO controllers through IOC port
> (so we don't need to need to maintain cache coherency for these
> devices manualy. All cache sync ops will be nop)
> And we can connect Ethernet directly to RAM port (so we had to
> maintain cache coherency manualy. Cache sync ops will be real
> flush/invalidate operations)

I presume this has been tested with some stress testing bonie++ etc.

> +void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
> +			const struct iommu_ops *iommu, bool coherent)
> +{
> +	/*
> +	 * IOC hardware snoops all DMA traffic keeping the caches consistent
> +	 * with memory - eliding need for any explicit cache maintenance of
> +	 * DMA buffers - so we can use dma_direct cache ops.
> +	 */
> +	if (is_isa_arcv2() && ioc_enable && coherent) {
> +		set_dma_ops(dev, &dma_direct_ops);
> +		dev_info(dev, "use dma_direct_ops cache ops\n");
> +	} else {
> +		set_dma_ops(dev, &dma_noncoherent_ops);
> +		dev_info(dev, "use dma_noncoherent_ops cache ops\n");
> +	}
> +}

And further to Christoph's comment, if we decide to skip the branch, please add a
comment that it is already set in generic code etc.