[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20230807221754.51667-1-blarson@amd.com>
Date: Mon, 7 Aug 2023 15:17:54 -0700
From: Brad Larson <blarson@....com>
To: <arnd@...db.de>
CC: <adrian.hunter@...el.com>, <alcooperx@...il.com>,
<andy.shevchenko@...il.com>, <blarson@....com>,
<brendan.higgins@...ux.dev>, <briannorris@...omium.org>,
<broonie@...nel.org>, <catalin.marinas@....com>,
<conor+dt@...nel.org>, <davidgow@...gle.com>,
<devicetree@...r.kernel.org>, <fancer.lancer@...il.com>,
<gerg@...ux-m68k.org>, <gsomlo@...il.com>,
<hal.feng@...rfivetech.com>, <hasegawa-hitomi@...itsu.com>,
<j.neuschaefer@....net>, <joel@....id.au>, <kernel@...il.dk>,
<krzk@...nel.org>, <krzysztof.kozlowski+dt@...aro.org>,
<lee.jones@...aro.org>, <lee@...nel.org>,
<linux-arm-kernel@...ts.infradead.org>,
<linux-kernel@...r.kernel.org>, <linux-mmc@...r.kernel.org>,
<linux-spi@...r.kernel.org>, <p.zabel@...gutronix.de>,
<rdunlap@...radead.org>, <robh+dt@...nel.org>,
<samuel@...lland.org>, <skhan@...uxfoundation.org>,
<suravee.suthikulpanit@....com>, <thomas.lendacky@....com>,
<tonyhuang.sunplus@...il.com>, <ulf.hansson@...aro.org>,
<vaishnav.a@...com>, <walker.chen@...rfivetech.com>,
<will@...nel.org>, <zhuyinbo@...ngson.cn>
Subject: Re: [PATCH v14 8/8] soc: amd: Add support for AMD Pensando SoC Controller
Hi Arnd,
> On Wed, May 24, 2023, at 00:11, Brad Larson wrote:
>>> On Mon, May 15, 2023, at 20:16, Brad Larson wrote:
...
>>> Also, can you explain why this needs a low-lever user interface
>>> in the first place, rather than something that can be expressed
>>> using high-level abstractions as you already do with the reset
>>> control?
>>>
>>> All of the above should be part of the changelog text to get a
>>> driver like this merged. I don't think we can get a quick
>>> solution here though, so maybe you can start by removing the
>>> ioctl side and having the rest of the driver in drivers/reset?
Might be best to pull the whole thing for now until an acceptable
solution is reached. The reset function is a recovery mechanisim rarely
used where the byte access to the different IP at the 4 chip-selects
is needed for a system to boot.
>> In the original patchset I added a pensando compatible to spidev and that
>> was nacked in review and reusing some random compatible that made it into
>> spidev was just wrong. Further it was recommended this should be a system
>> specific driver and don't rely on a debug driver like spidev. I changed
>> over to a platform specific driver and at that time I also needed to include
>> a reset controller (emmc reset only). I put these in drivers/mfd and
>> drivers/reset. Review of the device tree for this approach went back and
>> forth to _not_ have four child nodes on the spi device each with the same
>> compatible. Decision was to squash the child nodes into the parent and put
>> the reset-controller there also. One driver and since its pensando
>> specific its currently in drivers/soc/amd.
>>
>> There are five different user processes and some utilities that access the
>> functionality in the cpld/fpga. You're correct, its passing messages that
>> are specific to the IP accessed via chip-select. No Elba system will boot
>> without this driver providing ioctl access.
> Thank you for the detailed summary. Moving away from spidev and
> from mfd seems all reasonable here. I'm still a bit confused by
> why you have multiple chipselects here that are for different
> subdevices but ended with a single user interface for all of them,
> but that's not a big deal.
The goal is to isolate the the kernel from device and platform specific
changes. All the IO to the spi connected CPLD/FPGA (design/cost dependent)
is a byte at a time or up to 16 bytes for internal flash mgmt. Performance
is not an issue and spidev was sufficient.
Maybe this paints the right picture to zero in on a correct approach.
Internal and external IP can be present at CS1/CS2 depending on the design
where the CS0 board controller registers get additions over time in a
backward compatible manner.
Design 1: FPGA
CS0: Board controller registers
CS1: Designware SPI to I2C to board peripherals
CS2: Lattice dual I2C master
CS3: Internal storage
Design 2: CPLD
CS0: Board controller registers
CS1: Not used or some other board specific registers
CS2: Lattice dual I2C master
CS3: Internal storage
> The main bit that sticks out about this high-level design is how
> it relies on user space utilities at all to understand the message
> format. From what I understand about the actual functionality of
> this device, it most closely resembles an embedded controller that
> you might find in a laptop or server machine, and those usually
> have kernel drivers in drivers/platform/ to interact with the
> device.
The dozens of registers at CS0 for board management are defined in
userspace programs or script. Only the regsiter offset/bit for
emmc reset is needed for the reset function in the patches.
> Has anyone tried to do it like that? Maybe it would help
> to see what the full protocol and the user space side looks
> like, in order to move some or all of it into the kernel.
Looking at drivers/platform its pretty sparse. What do you
recommend based on the design 1/2 variations?
Regards,
Brad
Powered by blists - more mailing lists