Re: [PATCH v2 2/2] reset: Reset controller driver for Intel LGM SoC

From: Martin Blumenstingl
Date: Mon Sep 02 2019 - 18:05:09 EST


Hi,

On Mon, Sep 2, 2019 at 11:45 AM Chuan Hua, Lei
<chuanhua.lei@xxxxxxxxxxxxxxx> wrote:
>
> Hi Martin,
>
>
> On 9/2/2019 5:38 AM, Martin Blumenstingl wrote:
> > Hi,
> >
> > On Fri, Aug 30, 2019 at 5:02 AM Chuan Hua, Lei
> > <chuanhua.lei@xxxxxxxxxxxxxxx> wrote:
> >> Hi Martin,
> >>
> >> On 8/30/2019 5:40 AM, Martin Blumenstingl wrote:
> >>> Hi,
> >>>
> >>> On Thu, Aug 29, 2019 at 4:51 AM Chuan Hua, Lei
> >>> <chuanhua.lei@xxxxxxxxxxxxxxx> wrote:
> >>>
> >>>>> I'm not surprised that we got some of the IP block layout for the
> >>>>> VRX200 RCU "wrong" - all "documentation" we have is the old Lantiq UGW
> >>>>> (BSP).
> >>>>> with proper documentation (as in a "public datasheet for the SoC") it
> >>>>> would be easy to spot these mistakes (at least I assume that the
> >>>>> quality of the Infineon / Lantiq datasheets is excellent).
> >>>>>
> >>>>> back to reset-intel-syscon:
> >>>>> assigning only one job to the RCU hardware is a good idea (in my opinion).
> >>>>> that brings up a question: why do we need the "syscon" compatible for
> >>>>> the RCU node?
> >>>>> this is typically used when registers are accessed by another IP block
> >>>>> and the other driver has to access these registers as well. does this
> >>>>> mean that there's more hidden in the RCU registers?
> >>>> As I mentioned, some other misc registers are put into RCU even they
> >>>> don't belong to reset functions.
> >>> OK, just be aware that there are also rules for syscon compatible
> >>> drivers, see for example: [0]
> >>> if Rob (dt-bindings maintainer) is happy with the documentation in
> >>> patch 1 then I'm fine with it as well.
> >>> for my own education I would appreciate if you could describe these
> >>> "other misc registers" with a few sentences (I assume that this can
> >>> also help Rob)
> >> For LGM, RCU is clean. There would be no MISC register after software's
> >> feedback. These misc registers will be moved to chiptop/misc
> >> groups(implemented by syscon). For legacy SoC, we do have a lot MISC
> >> registers for different SoCs.
> > OK, I think I understand now: chiptop != RCU
> > so RCU really only has one purpose: handling resets
> > while chiptop manages all the random bits
> >
> > does this means we don't need RCU to match "syscon"?
>
> If we don't support legacy SoC with the same driver, we don't need
> syscon, just regmap. Regmap is a must for us since we will use regmap
> proxy to implement secure rest via secure processor.
I think we should drop the syscon compatible for LGM then
even for the legacy SoCs the reset controller should not have a syscon
compatible: instead it should have a syscon parent (as the current
"lantiq,xrx200-reset" binding requires and as suggested by Rob for
another IP block: [0])

keeping regmap is great in my opinion because it's a nice API and gets
rid of some boilerplate
even better if it makes things easier for accessing the secure processor

> >
> >>> [...]
> >>>>>>>>>> 4. Code not optimized and intel internal review not assessed.
> >>>>>>>>> insights from you (like the issue with the reset callback) are very
> >>>>>>>>> valuable - this shows that we should focus on having one driver.
> >>>>>>>>>
> >>>>>>>>>> Based on the above findings, I would suggest reset-lantiq.c to move to
> >>>>>>>>>> reset-intel-syscon.c
> >>>>>>>>> my concern with having two separate drivers is that it will be hard to
> >>>>>>>>> migrate from reset-lantiq to the "optimized" reset-intel-syscon
> >>>>>>>>> driver.
> >>>>>>>>> I don't have access to the datasheets for the any Lantiq/Intel SoC
> >>>>>>>>> (VRX200 and even older).
> >>>>>>>>> so debugging issues after switching from one driver to another is
> >>>>>>>>> tedious because I cannot tell which part of the driver is causing a
> >>>>>>>>> problem (it's either "all code from driver A" vs "all code from driver
> >>>>>>>>> B", meaning it's hard to narrow it down).
> >>>>>>>>> with separate commits/patches that are improving the reset-lantiq
> >>>>>>>>> driver I can do git bisect to find the cause of a problem on the older
> >>>>>>>>> SoCs (VRX200 for example)
> >>>>>>>> Our internal version supports XRX350/XRX500/PRX300(MIPS based) and
> >>>>>>>> latest Lighting Mountain(X86 based). Migration to reset-intel-syscon.c
> >>>>>>>> should be straight forward.
> >>>>>>> what about the _reset callback on the XRX350/XRX500/PRX300 SoCs - do
> >>>>>>> they only use level resets (_assert and _deassert) or are some reset
> >>>>>>> lines using reset pulses (_reset)?
> >>>>>>>
> >>>>>>> when we wanted to switch from reset-lantiq.c to reset-intel-syscon.c
> >>>>>>> we still had to add support for the _reset callback as this is missing
> >>>>>>> in reset-intel-syscon.c currently
> >>>>>> Yes. We have reset pulse(assert, then check the reset status).
> >>>>> only now I realized that the reset-intel-syscon driver does not seem
> >>>>> to use the status registers (instead it's looking at the reset
> >>>>> registers when checking the status).
> >>>>> what happened to the status registers - do they still exist in newer
> >>>>> SoCs (like LGM)? why are they not used?
> >>>> Reset status check is there. regmap_read_poll_timeout to check status
> >>>> big. Status register offset <4) from request register. For legacy, there
> >>>> is one exception, we can add soc specific data to handle it.
> >>> I see, thank you for the explanation
> >>> this won't work on VRX200 for example because the status register is
> >>> not always at (reset register - 0x4)
> >> As I mentioned, VRX200 and all legacy SoCs (MIPS based) can be solved
> >> with one soc data in the compatible array.
> >>
> >> For example(not same as upstream, but idea is similar)
> >>
> >> static u32 intel_stat_reg_off(struct intel_reset_data *data, u32 req_off)
> >> {
> >> if (data->soc_data->legacy && req_off == RCU_RST_REQ)
> >> return RCU_RST_STAT;
> >> else
> >> return req_off + 0x4;
> >> }
> >>
> >>>>> on VRX200 for example there seem to be some cases where the bits in
> >>>>> the reset and status registers are different (for example: the first
> >>>>> GPHY seems to use reset bit 31 but status bit 30)
> >>>>> this is currently not supported in reset-intel-syscon
> >>>> This is most tricky and ugly part for VRX200/Danube. Do you have any
> >>>> idea to handle this nicely?
> >>> with reset-lantiq we have the following register information:
> >>> a) reset offset: first reg property
> >>> b) status offset: second reg property
> >>> c) reset bit: first #reset-cell
> >>> d) status bit: second #reset-cell
> >>>
> >>> reset-intel-syscon derives half of this information from the two #reset-cells:
> >>> a) reset offset: first #reset-cell
> >>> b) status offset: reset offset - 0x4
> >>> c) reset bit: second #reset-cell
> >>> d) status bit: same as reset bit
> >>>
> >>> I cannot make any suggestion (yet) how to handle VRX200 and LGM in one
> >>> driver because I don't know enough about LGM (yet).
> >>> on VRX200 my understanding is that we have 64 reset bits (2x 32bit
> >>> registers) and 64 status bits (also 2x 32bit registers). each reset
> >>> bit has a corresponding status bit but the numbering may be different
> >>> it's not clear to me how many resets LGM supports and how they are
> >>> organized. for example: I think it makes a difference if "there are 64
> >>> registers with each one reset bit" versus "there are two registers
> >>> with 32 bits each"
> >>> please share some details how it's organized internally, then I can
> >>> try to come up with a suggestion.
> >> LGM reset organization is more clean compared with legacy SoCs. We have
> >> 8 x 32bit reset and status registers(more modules need to be reset,
> >> overall ideas are similar without big change). Their request and status
> >> bit is at the same register bit position. Hope this will help you.
> > have you already discussed using only one reset cell?
> > if there's only one big reset controller in RCU then why not let the
> > reset controller driver do it's job of translating a reset line? also
> > this represents the hardware best (dt-bindings should describe the
> > hardware, drivers then translate that into the various subsystems
> > offered by the kernel).
> >
> > we have to translate it into:
> > - status register and bit
> > - reset register and bit
> >
> > for LGM the implementation seems to be the easiest because the reset
> > line can be mapped easily to the registers and bit offsets (for
> > example like reset-meson.c does it, which also supports 256 reset
> > lines together with for example
> > include/dt-bindings/reset/amlogic,meson-g12a-reset.h. the latter is
> > nice to have but optional)
> When we implement this driver, we checked other drivers(hisilicon/*,
> reset-berlin.c and etc). After evaluation, we think register offset and
> register bit are easier for users to understand and use if they follow
> the hardware spec.
just so I know how the documentation looks like:
does the hardware spec document 8 registers, each with (up to) the 32
reset lines in it?

reset-meson.c does it like that, but the difference there is that the
reset registers are continuous because there's no status register in
between
so your existing way of describing the reset line seems fine if Rob is
happy with it as well

> > we can then implement special translation logic (in other words: a
> > separate of_xlate callback) for VRX200 which then has to do more
> > "magic" (like you have shown in your example code above: "if the reset
> > line belongs to the second set of 32 reset lines then use reset offset
> > X and status offset Y" - or even use a translation table as
> > reset-imx7.c does)
> >
> > the current binding is a mix of specifying reset register and bit in
> > .dts but calculating the status register.
> > I missed the calculation of the status register until you pointed it out earlier
> But we still don't have a good solution for VRX200 status bit issues.
> Before we solve this issue, it is very difficult to use one driver for
OK, let me summarize what we have so far.

all SoC have the following "shared" logic so far:
- all reset_control_ops callbacks are the same on VRX200 and LGM
(assuming we fix the issues you found in the reset-lantiq.c
implementation)
- internally we should use regmap (LGM for accessing the secure
processor, earlier SoCs because the parent is a syscon)
- each reset line consists of a reset register offset and bit as well
as a status register offset and bit

however, we have differences in:
- how to map the registers (LGM maps the RCU registers directly while
earlier SoCs fetch the parent syscon)
- calculation of the status register
- calculation of the status bit

I see two ways to use one common driver for LGM and the earlier SoCs:

1) use a reset line mapping table as for example reset-imx7.c does.
this would include reset register, reset bit, status register and status bit.
LGM can use a macro to get rid of the duplication between status bit
and reset bit (and the status register offset if you prefer)
this case would use #reset-cells = <1> and we wouldn't need to
implement the of_xlate callback

2) on VRX200 (and probably the older SoCs as well) we can encode the
following information in one 32-bit value:
- reset register (max value: 0x48)
- status register (max value: 0x24)
- reset bit (max value: 32)
- status bit (max value: 32)

if this also works for LGM we can determine all required information
for a reset line in the of_xlate callback and translate it to one
32-bit value.
LGM and earlier SoCs would each use it's own of_xlate implementation.
the reset_control_ops callback would then unpack the 32-bit value
("unsigned long id") into the reset register and bit as well as status
register and bit (as needed)

both ways can work, but it depends on what the dt-bindings maintainers
(like Rob) think of the binding itself.
(dt-bindings follow what the hardware implements, the driver only does
the translation between a vendor specific binding and a given
subsystem)
so we first need Rob's ACK on the binding, then we can figure out the
best driver implementation for that binding


Martin


[0] https://lkml.org/lkml/2019/8/27/849