Re: [RFC v2 4/5] dt-bindings: of: Add plumbing for restricted DMA pool

From: Tomasz Figa
Date: Tue Aug 11 2020 - 05:15:58 EST


On Mon, Aug 3, 2020 at 5:15 PM Tomasz Figa <tfiga@xxxxxxxxxxxx> wrote:
>
> Hi Claire and Rob,
>
> On Mon, Aug 3, 2020 at 4:26 PM Claire Chang <tientzu@xxxxxxxxxxxx> wrote:
> >
> > On Sat, Aug 1, 2020 at 4:58 AM Rob Herring <robh@xxxxxxxxxx> wrote:
> > >
> > > On Tue, Jul 28, 2020 at 01:01:39PM +0800, Claire Chang wrote:
> > > > Introduce the new compatible string, device-swiotlb-pool, for restricted
> > > > DMA. One can specify the address and length of the device swiotlb memory
> > > > region by device-swiotlb-pool in the device tree.
> > > >
> > > > Signed-off-by: Claire Chang <tientzu@xxxxxxxxxxxx>
> > > > ---
> > > > .../reserved-memory/reserved-memory.txt | 35 +++++++++++++++++++
> > > > 1 file changed, 35 insertions(+)
> > > >
> > > > diff --git a/Documentation/devicetree/bindings/reserved-memory/reserved-memory.txt b/Documentation/devicetree/bindings/reserved-memory/reserved-memory.txt
> > > > index 4dd20de6977f..78850896e1d0 100644
> > > > --- a/Documentation/devicetree/bindings/reserved-memory/reserved-memory.txt
> > > > +++ b/Documentation/devicetree/bindings/reserved-memory/reserved-memory.txt
> > > > @@ -51,6 +51,24 @@ compatible (optional) - standard definition
> > > > used as a shared pool of DMA buffers for a set of devices. It can
> > > > be used by an operating system to instantiate the necessary pool
> > > > management subsystem if necessary.
> > > > + - device-swiotlb-pool: This indicates a region of memory meant to be
> > >
> > > swiotlb is a Linux thing. The binding should be independent.
> > Got it. Thanks for pointing this out.
> >
> > >
> > > > + used as a pool of device swiotlb buffers for a given device. When
> > > > + using this, the no-map and reusable properties must not be set, so the
> > > > + operating system can create a virtual mapping that will be used for
> > > > + synchronization. Also, there must be a restricted-dma property in the
> > > > + device node to specify the indexes of reserved-memory nodes. One can
> > > > + specify two reserved-memory nodes in the device tree. One with
> > > > + shared-dma-pool to handle the coherent DMA buffer allocation, and
> > > > + another one with device-swiotlb-pool for regular DMA to/from system
> > > > + memory, which would be subject to bouncing. The main purpose for
> > > > + restricted DMA is to mitigate the lack of DMA access control on
> > > > + systems without an IOMMU, which could result in the DMA accessing the
> > > > + system memory at unexpected times and/or unexpected addresses,
> > > > + possibly leading to data leakage or corruption. The feature on its own
> > > > + provides a basic level of protection against the DMA overwriting buffer
> > > > + contents at unexpected times. However, to protect against general data
> > > > + leakage and system memory corruption, the system needs to provide a
> > > > + way to restrict the DMA to a predefined memory region.
> > >
> > > I'm pretty sure we already support per device carveouts and I don't
> > > understand how this is different.
> > We use this to bounce streaming DMA in and out of a specially allocated region.
> > I'll try to merge this with the existing one (i.e., shared-dma-pool)
> > to see if that
> > makes things clearer.
> >
>
> Indeed, from the firmware point of view, this is just a carveout, for
> which we have the "shared-dma-pool" compatible string defined already.
>
> However, depending on the device and firmware setup, the way the
> carevout is used may change. I can see the following scenarios:
>
> 1) coherent DMA (dma_alloc_*) within a reserved pool and no
> non-coherent DMA (dma_map_*).
>
> This is how the "memory-region" property is handled today in Linux for
> devices which can only DMA from/to the given memory region. However,
> I'm not sure if no non-coherent DMA is actually enforced in any way by
> the DMA subsystem.
>
> 2) coherent DMA from a reserved pool and non-coherent DMA from system memory
>
> This is the case for the systems which have some dedicated part of
> memory which is guaranteed to be coherent with the DMA, but still can
> do non-coherent DMA to any part of the system memory. Linux handles it
> the same way as 1), which is what made me believe that 1) might not
> actually be handled correctly.
>
> 3) coherent DMA and bounced non-coherent DMA within a reserved pool
> 4) coherent DMA within one pool and bounced non-coherent within another pool
>
> These are the two cases we're interested in. Basically they make it
> possible for non-coherent DMA from arbitrary system memory to be
> bounced through a reserved pool, which the device has access to. The
> current series implements 4), but I'd argue that it:
>
> - is problematic from the firmware point of view, because on most of
> the systems, both pools would be just some carveouts and the fact that
> Linux would use one for coherent and the other for non-coherent DMA
> would be an OS implementation detail,
> - suffers from the static memory split between coherent and
> non-coherent DMA, which could either result in some wasted memory or
> the DMA stopped working after a kernel update if the driver changes
> its allocation pattern,
>
> and so we should rather go with 3).
>
> Now, from the firmware point of view, it doesn't matter how the OS
> uses the carveout, but I think it's still necessary to tell the OS
> about the device DMA capability. Right now we use "memory-region" for
> any kind of reserved memory, but looking at the above scenarios, there
> are 2 cases:
>
> a) the memory region is preferred for the device, e.g. it enables
> coherency, but the device can still DMA across the rest of the system
> memory. This is the case in scenario 2) and is kind of assumed in the
> Linux DMA subsystem, although it's certainly not the case for a lot of
> hardware, even if they use the "memory-region" binding.
>
> b) the memory region is the only region that the device can access.
> This is the case in scenarios 1), 3) and 4).
>
> For this, I'd like to propose a "restricted-dma-region" (feel free to
> suggest a better name) binding, which is explicitly specified to be
> the only DMA-able memory for this device and make Linux use the given
> pool for coherent DMA allocations and bouncing non-coherent DMA.
>
> What do you think?

Rob, Robin, we'd appreciate your feedback on this when you have a
chance to take a look again. Thanks!

Best regards,
Tomasz

>
> Best regards,
> Tomasz
>
> > >
> > > What is the last sentence supposed to imply? You need an IOMMU?
> > The main purpose is to mitigate the lack of DMA access control on
> > systems without an IOMMU.
> > For example, we plan to use this plus a MPU for our PCIe WiFi which is
> > not behind an IOMMU.
> >
> > >
> > > > - vendor specific string in the form <vendor>,[<device>-]<usage>
> > > > no-map (optional) - empty property
> > > > - Indicates the operating system must not create a virtual mapping
> > > > @@ -117,6 +135,16 @@ one for multimedia processing (named multimedia-memory@77000000, 64MiB).
> > > > compatible = "acme,multimedia-memory";
> > > > reg = <0x77000000 0x4000000>;
> > > > };
> > > > +
> > > > + wifi_coherent_mem_region: wifi_coherent_mem_region {
> > > > + compatible = "shared-dma-pool";
> > > > + reg = <0x50000000 0x400000>;
> > > > + };
> > > > +
> > > > + wifi_device_swiotlb_region: wifi_device_swiotlb_region {
> > > > + compatible = "device-swiotlb-pool";
> > > > + reg = <0x50400000 0x4000000>;
> > > > + };
> > > > };
> > > >
> > > > /* ... */
> > > > @@ -135,4 +163,11 @@ one for multimedia processing (named multimedia-memory@77000000, 64MiB).
> > > > memory-region = <&multimedia_reserved>;
> > > > /* ... */
> > > > };
> > > > +
> > > > + pcie_wifi: pcie_wifi@0,0 {
> > > > + memory-region = <&wifi_coherent_mem_region>,
> > > > + <&wifi_device_swiotlb_region>;
> > > > + restricted-dma = <0>, <1>;
> > > > + /* ... */
> > > > + };
> > > > };
> > > > --
> > > > 2.28.0.rc0.142.g3c755180ce-goog
> > > >