lkml.org 
[lkml]   [2021]   [Mar]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH 0/3] Apple M1 DART IOMMU driver
On Fri, Mar 26, 2021 at 6:51 PM Sven Peter <sven@svenpeter.dev> wrote:
> On Fri, Mar 26, 2021, at 18:34, Robin Murphy wrote:
> > On 2021-03-26 17:26, Mark Kettenis wrote:
> > >
> > > Anyway, from my viewpoint having the information about the IOVA
> > > address space sit on the devices makes little sense. This information
> > > is needed by the DART driver, and there is no direct cnnection from
> > > the DART to the individual devices in the devicetree. The "iommus"
> > > property makes a connection in the opposite direction.
> >
> > What still seems unclear is whether these addressing limitations are a
> > property of the DART input interface, the device output interface, or
> > the interconnect between them. Although the observable end result
> > appears more or less the same either way, they are conceptually
> > different things which we have different abstractions to deal with.
>
> I'm not really sure if there is any way for us to figure out where these
> limitation comes from though.

My first guess was that this is done to partition the available address
address space in a way that allows one physical IOTLB to handle
multiple devices that each have their own page table for a subset
of the address space, as was done on old PowerPC IOMMUs.
However, the ranges you list don't really support that model.

> I've done some more experiments and looked at all DART nodes in Apple's Device
> Tree though. It seems that most (if not all) masters only connect 32 address
> lines even though the iommu can handle a much larger address space. I'll therefore
> remove the code to handle the full space for v2 since it's essentially dead
> code that can't be tested anyway.
>
>
> There are some exceptions though:
>
> There are the PCIe DARTs which have a different limitation which could be
> encoded as 'dma-ranges' in the pci bus node:
>
> name base size
> dart-apcie1: 00100000 3fe00000
> dart-apcie2: 00100000 3fe00000
> dart-apcie0: 00100000 3fe00000
> dart-apciec0: 00004000 7fffc000
> dart-apciec1: 80000000 7fffc000

This looks like they are reserving some address space in the beginning
and/or the end, and for apciec0, the address space is partitioned into
two equal-sized regions.

> Then there are also these display controller DARTs. If we wanted to use dma-ranges
> we could just put them in a single sub bus:
>
> name base size
> dart-disp0: 00000000 fc000000
> dart-dcp: 00000000 fc000000
> dart-dispext0: 00000000 fc000000
> dart-dcpext: 00000000 fc000000
>
>
> And finally we have these strange ones which might eventually each require
> another awkward sub-bus if we want to stick to the dma-ranges property.
>
> name base size
> dart-aop: 00030000 ffffffff ("always-on processor")
> dart-pmp: 00000000 bff00000 (no idea yet)

Here I also see a "pio-vm-size" property:

dart-pmp {
pio-vm-base = <0xc0000000>;
pio-vm-size = <0x40000000>;
vm-size = <0xbff00000>;
...
};

Which seems to give 3GB of address space to the normal iotlb,
plus the last 1GB to something else. The vm-base property is also
missing rather than zero, but that could just be part of their syntax
instead of a real difference.

Could it be that there are

> dart-sio: 0021c000 fbde4000 (at least their Secure Enclave/TPM co-processor)

Same here:
dart-sio {
vm-base = <0x0>;
vm-size = <0xfc000000>;
pio-vm-base = <0xfd000000>;
pio-vm-size = <0x2000000>;
pio-granularity = <0x1000000>;
}

There are clearly two distinct ranges that split up the 4GB space again,
with a small hole of 16MB (==pio-granularity) at the end of each range.

The "pio" name might indicate that this is a range of addresses that
can be programmed to point at I/O registers in another device, rather
than pointing to RAM.

Arnd

\
 
 \ /
  Last update: 2021-03-26 21:01    [W:0.218 / U:0.124 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site