Re: [PATCH v2] arm64: dts: rockchip: rk356x: Fix PCIe register map and ranges

[Peter Geis]

On Fri, Oct 21, 2022 at 4:52 PM Mark Kettenis <mark.kettenis@xxxxxxxxx> wrote:
>
> > Date: Fri, 21 Oct 2022 21:32:48 +0200
> > From: Ondřej Jirman <megi@xxxxxx>
> >
> > On Fri, Oct 21, 2022 at 12:48:15PM -0400, Peter Geis wrote:
> > > On Fri, Oct 21, 2022 at 11:39 AM Ondřej Jirman <megi@xxxxxx> wrote:
> > > >
> > > > On Fri, Oct 21, 2022 at 09:07:50AM -0400, Peter Geis wrote:
> > > > > Good Morning Heiko,
> > > > >
> > > > > Apologies for just getting to this, I'm still in the middle of moving
> > > > > and just got my lab set back up.
> > > > >
> > > > > I've tested this patch series and it leads to the same regression with
> > > > > NVMe drives. A loop of md5sum on two identical 4GB random files
> > > > > produces the following results:
> > > > > d11cf0caa541b72551ca22dc5bef2de0  test-rand.img
> > > > > fad97e91da8d4fd554c895cafa89809b  test-rand2.img
> > > > > 2d56a7baa05c38535f4c19a2b371f90a  test-rand.img
> > > > > 74e8e6f93d7c3dc3ad250e91176f5901  test-rand2.img
> > > > > 25cfcfecf4dd529e4e9fbbe2be482053  test-rand.img
> > > > > 74e8e6f93d7c3dc3ad250e91176f5901  test-rand2.img
> > > > > b9637505bf88ed725f6d03deb7065dab  test-rand.img
> > > > > f7437e88d524ea92e097db51dce1c60d  test-rand2.img
> > > > >
> > > > > Before this patch series:
> > > > > d11cf0caa541b72551ca22dc5bef2de0  test-rand.img
> > > > > d11cf0caa541b72551ca22dc5bef2de0  test-rand2.img
> > > > > d11cf0caa541b72551ca22dc5bef2de0  test-rand.img
> > > > > d11cf0caa541b72551ca22dc5bef2de0  test-rand2.img
> > > > > d11cf0caa541b72551ca22dc5bef2de0  test-rand.img
> > > > > d11cf0caa541b72551ca22dc5bef2de0  test-rand2.img
> > > > > d11cf0caa541b72551ca22dc5bef2de0  test-rand.img
> > > > > d11cf0caa541b72551ca22dc5bef2de0  test-rand2.img
> > > > >
> > > > > Though I do love where this patch is going and would like to see if it
> > > > > can be made to work, in its current form it does not.
> > > >
> > > > Thanks for the test. Can you please also test v1? Also please share lspci -vvv
> > > > of your nvme drive, so that we can see allocated address ranges, etc.
> > >
> > > Good catch, with your patch as is, the following issue crops up:
> > > Region 0: Memory at 300000000 (64-bit, non-prefetchable) [size=16K]
> > > Region 2: I/O ports at 1000 [disabled] [size=256]
> > >
> > > However, with a simple fix, we can get this:
> > > Region 0: Memory at 300000000 (64-bit, non-prefetchable) [virtual] [size=16K]
> > > Region 2: I/O ports at 1000 [virtual] [size=256]
> > >
> > > and with it a working NVMe drive.
> > >
> > > Change the following range:
> > > 0x02000000 0x0 0x40000000 0x3 0x00000000 0x0 0x40000000>;
> > > to
> > > 0x02000000 0x0 0x00000000 0x3 0x00000000 0x0 0x40000000>;
> >
> > I've already tried this, but this unfrotunately breaks the wifi cards.
> > (those only use the I/O space) Maybe because I/O and memory address spaces
> > now overlap, I don't know. That's why I used the 1GiB offset for memory
> > space.
>
> Meanwhile, I have an NVMe drive that only works if mmio is completely
> untranslated.  This is an ADATA SX8000NP drive, which uses a Silicon
> Motion SM2260 controller.
>
> So for me, a working configuration has the following "ranges":
>
> ranges = <0x01000000 0x0 0x00000000 0x3 0x3fff0000 0x0 0x00010000>,
>          <0x02000000 0x0 0xf4000000 0x0 0xf4000000 0x0 0x02000000>,
>          <0x03000000 0x3 0x10000000 0x3 0x10000000 0x0 0x2fff0000>;
>
> This also needs changes to the "reg" propery:
>
> reg = <0x3 0xc0000000 0x0 0x00400000>,
>       <0x0 0xfe260000 0x0 0x00010000>,
>       <0x3 0x00000000 0x0 0x10000000>;

Now this is interesting. I've been reading up on PCIe ranges and what
is necessary for things to work properly, and I found this interesting
article from ARM:
https://developer.arm.com/documentation/102337/0000/Programmers-model/Memory-maps/AP-system-memory-map/PCIe-MMIO-and-ECAM-memory-regions

TLDR: We need a low region (below 4g) and a high region.

>From other articles I've gleaned that the config / io should probably
also be in the low range. As such I believe the other patch that was
sent to me may be the correct way to go. If both of you would try the
following reg / ranges:

reg = <0x3 0xc0000000 0x0 0x00400000>,
      <0x0 0xfe260000 0x0 0x00010000>,
      <0x0 0xf4000000 0x0 0x00100000>;

ranges = <0x01000000 0x0 0xf4100000 0x0 0xf4100000 0x0 0x00100000>,
<0x02000000 0x0 0xf4200000 0x0 0xf4200000 0x0 0x01e00000>,
<0x03000000 0x0 0x40000000 0x3 0x00000000 0x0 0x40000000>;

Very Respectfully,
Peter Geis

>
> Now admittedly, this is with OpenBSD running on EDK2 UEFI firmware
> from
>
>   https://github.com/jaredmcneill/quartz64_uefi
>
> that I modified to pass through the device tree and modify the ranges
> as above.  But the way my OpenBSD driver sets up the address
> translation windows matches what the mainline Linux driver does.
>
> I picked the ranges above to match the EDK2 configuration.  But it is
> a setup that maximizes the 32-bit mmio window.
>
> Cheers,
>
> Mark
>
> > > I still haven't tested this with other cards yet, and another patch
> > > that does similar work I've tested successfully as well with NVMe
> > > drives. I'll have to get back to you on the results of greater
> > > testing.
> > >
> > > Very Respectfully,
> > > Peter Geis
> > >
> > > >
> > > > kind regards,
> > > >         o.
> > > >
> > > > > Very Respectfully,
> > > > > Peter Geis
> >
> > _______________________________________________
> > linux-arm-kernel mailing list
> > linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
> > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel