On 2/17/21 11:57 AM, Ard Biesheuvel wrote:
On Wed, 17 Feb 2021 at 18:16, Jeremy Linton <firstname.lastname@example.org> wrote:
Considering that the reason for the 1 GB device limit is the -3 GB DMA
On 2/17/21 1:55 AM, Ard Biesheuvel via groups.io wrote:
On Wed, 17 Feb 2021 at 08:30, Jeremy Linton <email@example.com> wrote:Some of these firmware/board revisions appear to need the 3G
Which case is that?
On 2/17/21 12:56 AM, Ard Biesheuvel wrote:
On Wed, 17 Feb 2021 at 07:18, jlinton <firstname.lastname@example.org> wrote:Ha, I gave up! There are more important things to fix, especially when I
What happened to the IORT change? Don't we need that to ensure that
From: Jeremy Linton <email@example.com>
The existing RPi3 ACPI entries for the Arasan
and SDHCI controllers need updating to work
with the RPi4. This is done by adding a caps
override for the legacy Arasan controller and
then adding an entirely new entry for the newer
Then we flip the default routing to make the eMMC2
the default for the SD card, so that the WiFi can
start working on the Arasan.
Additional we add a menu item to enable the SDMA/ADMA2
modes on the controller.
v2->v3: Various small review tweaks, whitespace, wording
Linux sizes ZONE_DMA appropriately?
found another case that couldn't just be fixed by the IORT tweaking
without more kernel patches.
translation. The EMMC seems to be one of the devices who's DT
descriptions are being modified by the lower level firmware (like the PCI).
translation, I'd assume that the former and the latter apply to the
same set of peripherals.
But are you saying the dma-ranges properties are manipulated by the VC
firmware? Or other DT properties related to DMA translation?
Yes, Its changing dma-range property associated with the emmc in the DT its being handed which is then shared with atf/etc.
By 'this error', I mean the splat resulting from mismatching DMA
Well with stock linux, the device will configure, startup and corrupt data.
The default in this set is PIO mode, no DMA, same as the Arasan. If IBut having a 1 GB limit for only the eMMC2 in the IORT and a matching
get motivated (or someone else does) they can pick up the pieces to
finish turning the DMA on in linux. It also simplifies that IORT disable
patch I posted separately since I don't have to worry about enabling it
for a limit <2G.
_DMA method in its container should not trigger this error, I'd
limits for XHCI between IORT and _DMA.
No, I don't see that. The PCI/XHCI is fine with the IORT changes.
Is the reason for the data corruption understood?
It runs but appears to the address translation portion doesn't get applied (the command rings appear to be ok/etc) to FS buffers reliably so garbage gets written to the disk as the wrong bus locations get used. Its somewhat odd because at a first glance the directory structure/etc come back so if one just mounts the FS and ls's it, then unmounts it all appears to be ok. The first indications something is wrong are usually FS corruption messages. I have an instrumented sdhci/etc driver splatting on addrs > 1G so that all looks ok.
I could also argue that PIO at low speeds is worse then PIO at high
? I'm not sure I understand. Right now in linux we don't have SD or
The sdhci_caps_mask choice is what flags the device as not supportingAre you saying switching to the Arasan disabled UHS mode, and this is
DMA modes unless the user enables it. Yes this hurts perf, but not
nearly as badly as disabling UHS mode because we can't lower the card
voltage with the standard sdhci registers (rather having to depend on a
nonstandard rpi mailbox call which isn't available without a _DSM() or
something equally undesirable).
why this is an improvement nonetheless?
wifi. With just the caps _DSD entry the arasan will configure but it
runs PIO mode all the time (including with DT). The cap is needed
because the arasan returns 0 in the standard SDHCI caps registers.
The emmc2 supports faster modes, but we can't easily switch the voltage
to support them because the standard voltage control registers aren't
wired correctly (AFAIK). Given the change detection doesn't work right
either (AFAIK), we could hack up the linux sdhci subsystem to not reset
the card at startup and leave faster cards at 1.8V, but that is uglier
than adding a _DSM() to forward the voltage change request to the rpi
mailbox. But again we are hacking up the iproc (or sdhci_acpi) driver.
IMHO, Given its just a perf thing, and this whole board is compromised
in so many ways, it just isn't worth trying to support low voltage/UHS.
Since the card is already running at the slower speeds, using PIO
instead of DMA isn't a huge hit.
speeds, given that the CPU will be tied up for longer to transfer the
same amount of data. >
So then we don't have to have a 1GYes, that is obviously a win.
IORT, or dynamic _DMA translation.
But this set is about enabling both the SD and WiFi. The latter requiresFair enough. I'm just slightly disappointed that we cannot use the
the SD to be bound to the emmc2. At the moment there isn't much in the
way of a perf advantage to switching the SD from the Arasan to the
emmc2, the benefit comes from being able to use the wifi..
eMMC2 in DMA mode even for the lower speed, but I guess it is not the
end of the world.
Well its never done, at some point it can be revisited to make it faster. Maybe someone will come up with a clever way to do the voltage switching too. The platform has an easy way to trap to el3, but I can't see how to utilize that without sdhci driver changes at the moment.
If everyone else is on board with this approach, I'll pick these up tomorrow.