Re: [RFC][PATCH 1/2] WIP: Devicetree bindings for Ion

From: Andrew
Date: Wed Oct 07 2015 - 15:24:21 EST


On 2015-10-07 21:36, Rob Herring wrote:
On Wed, Oct 7, 2015 at 5:36 AM, Andrew <andrew@xxxxxxxxxx> wrote:
On 2015-10-07 02:01, Laura Abbott wrote:

On 10/6/15 3:35 PM, Rob Herring wrote:

On Tue, Oct 6, 2015 at 3:47 PM, Laura Abbott <labbott@xxxxxxxxxxxxxxxxx>
wrote:

From: Laura Abbott <laura@xxxxxxxxxxxx>


This adds a base set of devicetree bindings for the Ion memory
manager. This supports setting up the generic set of heaps and
their properties.

Signed-off-by: Laura Abbott <laura@xxxxxxxxxxxx>
Signed-off-by: Andrew Andrianov <andrew@xxxxxxxxxx>
---
drivers/staging/android/ion/devicetree.txt | 53
++++++++++++++++++++++++++++++


I have no issue with this going in here, but I do have lots of issues
with this binding.

1 file changed, 53 insertions(+)
create mode 100644 drivers/staging/android/ion/devicetree.txt

diff --git a/drivers/staging/android/ion/devicetree.txt
b/drivers/staging/android/ion/devicetree.txt
new file mode 100644
index 0000000..4a0c941
--- /dev/null
+++ b/drivers/staging/android/ion/devicetree.txt
@@ -0,0 +1,53 @@
+Ion Memory Manager
+
+Ion is a memory manager that allows for sharing of buffers via dma-buf.
+Ion allows for different types of allocation via an abstraction called
+a 'heap'. A heap represents a specific type of memory. Each heap has
+a different type. There can be multiple instances of the same heap
+type.
+
+Required properties for Ion
+
+- compatible: "linux,ion"
+
+All child nodes of a linux,ion node are interpreted as heaps
+
+required properties for heaps
+
+- linux,ion-heap-id: The Ion heap id used for allocation selection
+- linux,ion-heap-type: Ion heap type defined in ion.h
+- linux,ion-heap-name: Human readble name of the heap
+
+
+Optional properties
+- memory-region: A phandle to a memory region. Required for DMA heap
type
+(see reserved-memory.txt for details on the reservation)
+- linux,ion-heap-align: Alignment for the heap.
+
+Example:
+
+ ion {
+ compatbile = "linux,ion";
+ #address-cells = <1>;
+ #size-cells = <0>;
+
+ ion-system-heap {
+ linux,ion-heap-id = <0>;
+ linux,ion-heap-type = <ION_SYSTEM_HEAP_TYPE>;
+ linux,ion-heap-name = "system";


How does this vary across platforms? Is all of this being pushed down
to DT, because there is no coordination of this at the kernel ABI
level across platforms. In other words, why can't heap 0 be hardcoded
as system heap in the driver. It seems to me any 1 of these 3
properties could be used to derive the other 2.


Right now there is no guarantee heap IDs will be the same across
platforms. The heap IDs are currently part of the userspace ABI
as well since userspace clients must pass in a mask of the heap
IDs to allocate from. If we assume all existing clients could change,
heaps such as the system heap could be mandated to have the same
heap ID but we'd still run into problems if you have multiple
heaps of the same type (e.g. multiple carveouts)

Vendors largely ignore the kernel-userspace ABI and anything in
staging is not a ABI. So arguments about what the ABI is currently is
pointless IMO.

Pushing an inconsistent kernel ABI to DT is not the answer.

Totally agree here.



I don't really like the idea of enforcing any IDs here. As of now
heap ids are generally something VERY platform-specific
(or even product-specific). Personally I'd prefer something like this
for userspace apps:

int id1 = ion_get_heap_id("camera");
if (id1 < 0) {
fprintf(stderr, "Invalid heap id");
exit(1);
}

int id2 = ion_get_heap_id("backup-heap");
if (id2 < 0) {
fprintf(stderr, "Invalid heap id");
exit(1);
}

We've learned that creating number spaces like this are bad (irqs,
gpios, /dev nodes). We should move away from that. Why should
userspace care about IDs or what the IDs are? An ID is just encoding
certain implicit requirements. So are the strings here. Users should
express what capabilities, restrictions, etc. they have, and then the
kernel can find the best heap.

I'd argue about that point, since sometimes kernel might NOT know
all the hardware details behind some of the heaps or how they are going to
be used. Hence it can't decide a proper heap. And that's where things get ugly.

Real-world example: There are several on-chip SRAM banks that make up several
heaps. Say, IM0, IM1, IM2.

Technically - they are all DMA, and all work. But the hardware guys hand you
a handful of weird recommendations, like:
* Decoder will work faster if you use bank IM2 for internal buffers,
and prefer DDR bank A for decoded frames.
* DSP should stick with IM1, and please prefer DDR bank B for buffers

When several such devices are involved in one chain - things may get
even weirder. Having manual control over where allocations are made allows
us to keep all these voodoo magicks away from the kernel and (hopefully) keep
vendors from dirty hacks into ion itself.


...

int ret = ion_alloc(fd, 0x100, 0x4,
(id1 | id2),
0, &handle);


What concerns kernel stuff, things are simpler - we may just pass the heap
to use
by a reference in devicetree node for that driver. Something like that:

...
ion-decoder-region : region_ddr {
linux,ion-heap-id = <1>;
linux,ion-heap-type = <ION_DMA_HEAP_TYPE>;
linux,ion-heap-name = "decoder_mem"
memory-region = <&camera_region>;
};
...
video_decoder@80180000 {
compatible = "acme,h266dec";
reg = <0x80180000 0x20000>,
reg-names = "registers";
interrupts = <12>;
interrupt-parent = <&vic1>;
ion-heaps-for-buffers = <&ion-decoder-region>

This is how memory-region is supposed to work. I don't see why we need
an additional level of indirection.

I don't see right now a clean way of getting ion heap id from driver when
having a memory-region referenced there. Correct me if if I'm all wrong about it.


Rob

--
Regards,
Andrew
RC Module
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/