mbox series

[0/8] soc: samsung: Add USIv2 driver

Message ID 20211127223253.19098-1-semen.protsenko@linaro.org (mailing list archive)
Headers show
Series soc: samsung: Add USIv2 driver | expand

Message

Sam Protsenko Nov. 27, 2021, 10:32 p.m. UTC
USIv2 IP-core provides selectable serial protocol (UART, SPI or
High-Speed I2C); only one can be chosen at a time. This series
implements USIv2 driver, which allows one to select particular USI
function in device tree, and also performs USI block initialization.

With that driver implemented, it's not needed to do USI initialization
in protocol drivers anymore, so that code is removed from the serial
driver.

Because USI driver is tristate (can be built as a module), serial driver
was reworked so it's possible to use its console part as a module too.
This way we can load serial driver module from user space and still have
serial console functional.

Make it impossible to build UART/SPI/I2C driver as a built-in when USIv2
driver built as a module: USIv2 configuration must be always done before
tinkering with particular protocol it implements.

Design features:
  - "reg" property contains USI registers start address (0xc0 offset);
    it's used in the driver to access USI_CON and USI_OPTION registers.
    This way all USI initialization (reset, HWACG, etc) can be done in
    USIv2 driver separately, rather than duplicating that code over
    UART/SPI/I2C drivers
  - System Register (system controller node) and its SW_CONF register
    offset are provided in "samsung,sysreg" property; it's used to
    select USI function (protocol to be used)
  - USI function is specified in "samsung,mode" property; integer value
    is used to simplify parsing
  - there is "samsung,clkreq-on" bool property, which makes driver
    disable HWACG control (needed for UART to work properly)
  - PCLK and IPCLK clocks are both provided to USI node; apparently both
    need to be enabled to access USI registers
  - protocol nodes are embedded (as a child nodes) in USI node; it
    allows correct init order, and reflects HW properly
  - USIv2 driver is a tristate: can be also useful from Android GKI
    requirements point of view
  - driver functions are implemented with further development in mind:
    we might want to add some SysFS interface later for example, or
    provide some functions to serial drivers with EXPORT_SYMBOL(), etc

Sam Protsenko (8):
  dt-bindings: soc: samsung: Add Exynos USIv2 bindings
  dt-bindings: soc: samsung: Add Exynos USIv2 bindings doc
  soc: samsung: Add USIv2 driver
  tty: serial: samsung: Remove USI initialization
  tty: serial: samsung: Enable console as module
  tty: serial: Make SERIAL_SAMSUNG=y impossible when EXYNOS_USI_V2=m
  i2c: Make I2C_EXYNOS5=y impossible when EXYNOS_USI_V2=m
  spi: Make SPI_S3C64XX=y impossible when EXYNOS_USI_V2=m

 .../bindings/soc/samsung/exynos-usi-v2.yaml   | 124 +++++++++
 drivers/i2c/busses/Kconfig                    |   1 +
 drivers/soc/samsung/Kconfig                   |  14 +
 drivers/soc/samsung/Makefile                  |   2 +
 drivers/soc/samsung/exynos-usi-v2.c           | 242 ++++++++++++++++++
 drivers/spi/Kconfig                           |   1 +
 drivers/tty/serial/Kconfig                    |   3 +-
 drivers/tty/serial/samsung_tty.c              |  57 ++---
 .../dt-bindings/soc/samsung,exynos-usi-v2.h   |  16 ++
 include/linux/serial_s3c.h                    |   9 -
 10 files changed, 425 insertions(+), 44 deletions(-)
 create mode 100644 Documentation/devicetree/bindings/soc/samsung/exynos-usi-v2.yaml
 create mode 100644 drivers/soc/samsung/exynos-usi-v2.c
 create mode 100644 include/dt-bindings/soc/samsung,exynos-usi-v2.h

Comments

David Virag Nov. 28, 2021, 3:15 a.m. UTC | #1
On Sun, 2021-11-28 at 00:32 +0200, Sam Protsenko wrote:
> USIv2 IP-core provides selectable serial protocol (UART, SPI or
> High-Speed I2C); only one can be chosen at a time. This series
> implements USIv2 driver, which allows one to select particular USI
> function in device tree, and also performs USI block initialization.
> 
> With that driver implemented, it's not needed to do USI
> initialization
> in protocol drivers anymore, so that code is removed from the serial
> driver.
> 

I think the downstream way of doing this (USI node reg being on the
SW_CONF register itself rather than an offset from uart/i2c/spi, the
USI driver only controlling the SW_CONF, and the uart/i2c/spi drivers
controlling their USI_CON and USI_OPTION regs) is cleaner, better, and
easier to adapt to USIv1 too.

For example: I'm sure this is the case on USIv2 devices too, but on
Exynos7885, different devices have USI modes configured differently.
For example a Samsung Galaxy A8 (2018) has all the USI blocks
configured as SPI while a Samsung Galaxy M20 has the first USI
configured as dual HSI2C, the second as HSI2C on the first 2 pins and
the third as HSI2C on the last 2 pins. With this way of doing
everything on USIv2 we'd need 3 disabled USIv2 nodes in the SoC DTSI
for one USI block, each for every protocol the USI block can do, all
having a single child for their protocol and each referencing the same
sysreg (not even sure if that's even supported). Then the board DTS
could enable the USI node it needs.

With the downstream way we could have just one USI node and we could
add the 3 protocols it can do disabled as seperate or child nodes. This
way the board DTS only needs to set the appropriate mode setting and
enable the protocol it needs. I'd say much better than having 3 USI
nodes for the same USI block.

Also this way is pretty USIv2 centric. Adding USIv1 support to this
driver is difficult this way because of the the lack of USI_CON and
USI_OPTION registers as a whole (so having nowhere to actually set the
reg of the USI node to, as the only thing USIv1 has is the SW_CONF
register). In my opinion being able to use the same driver and same
device tree layout for USIv1 and USIv2 is a definite plus

The only real drawback of that way is having to add code for USIv2
inside the UART, HSI2C, and SPI drivers but in my opinion the benefits
overweigh the drawbacks greatly. We could even make the uart/spi/hsi2c
drivers call a helper function in the USI driver to set their USI_CON
and USI_OPTION registers up so that code would be shared and not
duplicated. Wether this patch gets applied like this is not my choice
though, I'll let the people responsible decide
:-)

Anyways, soon enough I can write an USIv1 driver after I submit all the
7885 stuff I'm working on currently. If you want to, you can add USIv2
support to that driver, or if an USIv2 driver is already in upstream at
that point, if it is written in the downstream way I can add v1 support
to that, or if it's like this I'll have to make a whole seperate driver
with a whole seperate DT structure.

Best regards,
David
Krzysztof Kozlowski Nov. 29, 2021, 9:02 a.m. UTC | #2
On 28/11/2021 04:15, David Virag wrote:
> On Sun, 2021-11-28 at 00:32 +0200, Sam Protsenko wrote:
>> USIv2 IP-core provides selectable serial protocol (UART, SPI or
>> High-Speed I2C); only one can be chosen at a time. This series
>> implements USIv2 driver, which allows one to select particular USI
>> function in device tree, and also performs USI block initialization.
>>
>> With that driver implemented, it's not needed to do USI
>> initialization
>> in protocol drivers anymore, so that code is removed from the serial
>> driver.
>>
> 
> I think the downstream way of doing this (USI node reg being on the
> SW_CONF register itself rather than an offset from uart/i2c/spi, the
> USI driver only controlling the SW_CONF, and the uart/i2c/spi drivers
> controlling their USI_CON and USI_OPTION regs) is cleaner, better, and
> easier to adapt to USIv1 too.
> 
> For example: I'm sure this is the case on USIv2 devices too, but on
> Exynos7885, different devices have USI modes configured differently.
> For example a Samsung Galaxy A8 (2018) has all the USI blocks
> configured as SPI while a Samsung Galaxy M20 has the first USI
> configured as dual HSI2C, the second as HSI2C on the first 2 pins and
> the third as HSI2C on the last 2 pins. With this way of doing
> everything on USIv2 we'd need 3 disabled USIv2 nodes in the SoC DTSI
> for one USI block, each for every protocol the USI block can do, all
> having a single child for their protocol and each referencing the same
> sysreg (not even sure if that's even supported). Then the board DTS
> could enable the USI node it needs.

It's not supported (one cannot have three same nodes with same unit
addresses), so this would be solved by dropping out unused interfaces,
commenting them out or storing everything under one USI:

usi@0x1abcdef0 {
  serial@.... {
    status = "okay";
  }

  i2c@.... {
    status = "disabled";
  }

  spi@.... {
    status = "disabled";
  }
}

> 
> With the downstream way we could have just one USI node and we could
> add the 3 protocols it can do disabled as seperate or child nodes. This
> way the board DTS only needs to set the appropriate mode setting and
> enable the protocol it needs. I'd say much better than having 3 USI
> nodes for the same USI block.

Then however you need to handle probe ordering and possible probe deferrals.

> 
> Also this way is pretty USIv2 centric. Adding USIv1 support to this
> driver is difficult this way because of the the lack of USI_CON and
> USI_OPTION registers as a whole (so having nowhere to actually set the
> reg of the USI node to, as the only thing USIv1 has is the SW_CONF
> register). 

How is it difficult? Not having a register is easy - noop on given platform.

> In my opinion being able to use the same driver and same
> device tree layout for USIv1 and USIv2 is a definite plus
> 
> The only real drawback of that way is having to add code for USIv2
> inside the UART, HSI2C, and SPI drivers but in my opinion the benefits
> overweigh the drawbacks greatly. We could even make the uart/spi/hsi2c
> drivers call a helper function in the USI driver to set their USI_CON
> and USI_OPTION registers up so that code would be shared and not
> duplicated. Wether this patch gets applied like this is not my choice
> though, I'll let the people responsible decide
> :-)
> 
> Anyways, soon enough I can write an USIv1 driver after I submit all the
> 7885 stuff I'm working on currently. If you want to, you can add USIv2
> support to that driver, or if an USIv2 driver is already in upstream at
> that point, if it is written in the downstream way I can add v1 support
> to that, or if it's like this I'll have to make a whole seperate driver
> with a whole seperate DT structure.
> 
> Best regards,
> David
> 


Best regards,
Krzysztof
Sam Protsenko Nov. 29, 2021, 1:56 p.m. UTC | #3
On Sun, 28 Nov 2021 at 05:15, David Virag <virag.david003@gmail.com> wrote:
>
> On Sun, 2021-11-28 at 00:32 +0200, Sam Protsenko wrote:
> > USIv2 IP-core provides selectable serial protocol (UART, SPI or
> > High-Speed I2C); only one can be chosen at a time. This series
> > implements USIv2 driver, which allows one to select particular USI
> > function in device tree, and also performs USI block initialization.
> >
> > With that driver implemented, it's not needed to do USI
> > initialization
> > in protocol drivers anymore, so that code is removed from the serial
> > driver.
> >
>
> I think the downstream way of doing this (USI node reg being on the
> SW_CONF register itself rather than an offset from uart/i2c/spi, the
> USI driver only controlling the SW_CONF, and the uart/i2c/spi drivers
> controlling their USI_CON and USI_OPTION regs) is cleaner, better, and
> easier to adapt to USIv1 too.
>

One reason why I think it's better to provide SW_CONF register via
syscon node, is that it helps us to avoid possible register access
conflicts in future, and also conflicts when requesting corresponding
resources. In other words, the System Register block can be used by
many consumers (drivers) in future; those consumers might try to
modify the same registers simultaneously, which might lead to race
conditions (as RMW operation is not atomic), so some kind of
serialization should be done (like locking in regmap), which is
provided by syscon. Also, that wouldn't even come to that: you just
can't request the same I/O area twice in Linux. So if SW_CONF is
passed via "reg" property to USI driver, and then we try to map the
whole System Register (or its portion that includes SW_CONF), that
request would fail.

Although passing one SW_CONF register via "reg" might look easier to
implement, it might also bring us all sort of problems later on. And I
think a good design should account for such pitfalls.

As for the USI registers: I really don't think that duplicating the
code for USI block reset across uart/i2c/spi drivers would help us to
accomplish anything. Why those drivers should be even aware of USI
reset? At least in USIv2 block, the USI registers and uart/i2c/spi
registers are not mixed: they are located at different and always
fixed addresses. We can benefit from that fact, and provide Device
Tree structure which reflects the hardware one, separating USI control
from actual protocol nodes.

> For example: I'm sure this is the case on USIv2 devices too, but on
> Exynos7885, different devices have USI modes configured differently.
> For example a Samsung Galaxy A8 (2018) has all the USI blocks
> configured as SPI while a Samsung Galaxy M20 has the first USI
> configured as dual HSI2C, the second as HSI2C on the first 2 pins and
> the third as HSI2C on the last 2 pins. With this way of doing
> everything on USIv2 we'd need 3 disabled USIv2 nodes in the SoC DTSI
> for one USI block, each for every protocol the USI block can do, all
> having a single child for their protocol and each referencing the same
> sysreg (not even sure if that's even supported). Then the board DTS
> could enable the USI node it needs.
>

If I'm following you correctly, then it's not like that. I guess
Krzysztof already replied to that, so I'll probably just repeat his
words. In that case you'll have something like this in your SoC dtsi
(for your USIv1 case of course, because dual HSI2C is not present in
USIv2):

<<<<<<<<<<<<<<<<<<<<<<<<< cut here >>>>>>>>>>>>>>>>>>>>>>>
usi1 {
    spi1 {
    };

    hsi2c1_1 {
    };

    hsi2c1_2 {
    };
};

usi2 {
    spi2 {
    };

    hsi2c2_1 {
    };
};


usi3 {
    spi3 {
    };

    hsi2c2_2 {
    };
};
<<<<<<<<<<<<<<<<<<<<<<<<< cut here >>>>>>>>>>>>>>>>>>>>>>>

and then in your board dts you just have to enable corresponding usi's
with proper modes, and enable chosen protocol nodes, like this:

<<<<<<<<<<<<<<<<<<<<<<<<< cut here >>>>>>>>>>>>>>>>>>>>>>>
&usi1 {
    status = "okay"
    samsung,mode = <USI_V1_DUAL_I2C>;
};

&hsi2c1_1 {
    status = "okay"
};

&hsi2c1_2 {
    status = "okay"
};
<<<<<<<<<<<<<<<<<<<<<<<<< cut here >>>>>>>>>>>>>>>>>>>>>>>

> With the downstream way we could have just one USI node and we could
> add the 3 protocols it can do disabled as seperate or child nodes. This
> way the board DTS only needs to set the appropriate mode setting and
> enable the protocol it needs. I'd say much better than having 3 USI
> nodes for the same USI block.
>

Not sure if with downstream USI driver you can actually have protocols
as sub-nodes in USI node though. It doesn't do anything like
of_platform_populate().

Also, with this USIv2 driver you can do the same thing you described:
you can have just one USI node with 3 protocols as sub-nodes (or you
can even have protocol nodes outside of USI node, but I'd not
recommend that).

Actually I can see that it's my fault for not describing that case in
bindings example. I'll make sure to do that in v2. You also got me
thinking about default mode: sometimes SW_CONF reset value chooses
some protocol. In that case maybe it'd useful to have something like
USI_V2_DEFAULT, to tell driver to not touch SW_CONF at all. And also I
can add USI_V2_NONE while at it, so that driver can write 0x0 to
SW_CONF: that way no protocol will be selected. Maybe that can be
beneficial for PM reasons, if some board doesn't use some USI blocks
at all. Do you think it's feasible to add those two values to
dt-bindings header? And is it possible to do so in USIv1?

> Also this way is pretty USIv2 centric. Adding USIv1 support to this
> driver is difficult this way because of the the lack of USI_CON and
> USI_OPTION registers as a whole (so having nowhere to actually set the
> reg of the USI node to, as the only thing USIv1 has is the SW_CONF
> register). In my opinion being able to use the same driver and same
> device tree layout for USIv1 and USIv2 is a definite plus
>

Well, it's USIv2 driver after all. I never expected it can be extended
for USIv1 support. If you think it can be reused for USIv1, it's fine
by me. But we need to consider next things:
  - rename the driver to just "usi.c" (and also its configuration symbol)
  - provide different compatible for USIv1 (and maybe corresponding driver data)
  - rework bindings (header and doc); make sure existing bindings are
intact (we shouldn't change already introduced interfaces)
  - in case of USIv1 compatible; don't try to tinker with USIv2 registers
  - samsung,clkreq-on won't be available in case of USIv1 compatible

Because I don't have USIv1 SoC TRM (and neither do I possess some
USIv1 board which I can use for test), I don't think it's my place to
add USIv1 support. But I think it's possible to do so, using my input
above.

I can see how it might be frustrating having to do some extra work
(comparing to just using the code existing in downstream). But I guess
that's the difference: vendor is mostly concerned about competitive
advantage and getting to market fast, while upstream is more concerned
about quality, considering all use cases, and having proper design.
Anyway, we can work together to make it right, and to have both
IP-cores support. In the worst case, if those are too different, we
can have two separate drivers for those.

> The only real drawback of that way is having to add code for USIv2
> inside the UART, HSI2C, and SPI drivers but in my opinion the benefits
> overweigh the drawbacks greatly. We could even make the uart/spi/hsi2c
> drivers call a helper function in the USI driver to set their USI_CON
> and USI_OPTION registers up so that code would be shared and not
> duplicated. Wether this patch gets applied like this is not my choice
> though, I'll let the people responsible decide
> :-)
>

I'd argue that there are a lot of real drawbacks of using downstream
driver as is. That's why I completely re-designed and re-implemented
it. Downstream driver can't be built and function as a module, it
doesn't respect System Register sharing between consumers, it leads to
USI reset code duplication scattered across protocol drivers (that
arguably shouldn't even be aware of that), it doesn't reflect HW
structure clearly, it's not holding clocks needed for registers access
(btw, sysreg clock can be provided in syscon node, exactly for that
reason). As Krzysztof said, it also can't handle correct probe order
and deferred probes. Downstream driver might work fine for some
particular use-cases the vendor has, but in upstream it's better to
cover more cases we can expect, as upstream kernel is used on more
platforms, with more user space variants, etc.

I don't really think protocol drivers should be aware of USI registers
at all, but if we they do -- we can provide some API from USIv2 driver
later, with EXPORT_SYMBOL(), referencing corresponding USI instance by
phandle or using some other mechanism for inter-driver communication.

Of course, it's not my place to decide on patch acceptance too. But I
was under the impression that maintainers would be ok with this course
of actions. Also, upstream kernel seems to already follow the same
design for some similar drivers. See for example
drivers/soc/qcom/qcom_gsbi.c.

> Anyways, soon enough I can write an USIv1 driver after I submit all the
> 7885 stuff I'm working on currently. If you want to, you can add USIv2
> support to that driver, or if an USIv2 driver is already in upstream at
> that point, if it is written in the downstream way I can add v1 support
> to that, or if it's like this I'll have to make a whole seperate driver
> with a whole seperate DT structure.
>

If it's like you said (USIv1 only touches the SW_CONF register), I
guess USIv2 driver can be extended for USIv1 case. I already provided
my thoughts on such rework above. It's probably better to consult with
Krzysztof first. I guess the only way to figure out if it's feasible
or it's better to have separate exynos-usi-v1.c for USIv1, is to try
and add USIv1 support into USIv2 driver and see how pretty or ugly it
is :) Whatever the way you decide to go with, please add me to Cc list
when sending USIv1 patches.

> Best regards,
> David
Krzysztof Kozlowski Nov. 29, 2021, 5:35 p.m. UTC | #4
On 29/11/2021 14:56, Sam Protsenko wrote:
> On Sun, 28 Nov 2021 at 05:15, David Virag <virag.david003@gmail.com> wrote:
>>
>> Also this way is pretty USIv2 centric. Adding USIv1 support to this
>> driver is difficult this way because of the the lack of USI_CON and
>> USI_OPTION registers as a whole (so having nowhere to actually set the
>> reg of the USI node to, as the only thing USIv1 has is the SW_CONF
>> register). In my opinion being able to use the same driver and same
>> device tree layout for USIv1 and USIv2 is a definite plus
>>
> 
> Well, it's USIv2 driver after all. I never expected it can be extended
> for USIv1 support. If you think it can be reused for USIv1, it's fine
> by me. But we need to consider next things:
>   - rename the driver to just "usi.c" (and also its configuration symbol)
>   - provide different compatible for USIv1 (and maybe corresponding driver data)
>   - rework bindings (header and doc); make sure existing bindings are
> intact (we shouldn't change already introduced interfaces)
>   - in case of USIv1 compatible; don't try to tinker with USIv2 registers
>   - samsung,clkreq-on won't be available in case of USIv1 compatible

I expect this driver to be in future extended for USIv1 and I do not see
any problems in doing that for current Sam's approach. Most of our
drivers support several devices, sometimes with differences, and we
already have patterns solving it, e.g. ops structure or quirks bitmap.
Driver for new USIv1 compatible would skip setting USI_CON (or any other
unrelated register). Modification of SW_CONF could be shared or could be
also split, depending on complexity.

> 
> Because I don't have USIv1 SoC TRM (and neither do I possess some
> USIv1 board which I can use for test), I don't think it's my place to
> add USIv1 support. But I think it's possible to do so, using my input
> above.
> 
> I can see how it might be frustrating having to do some extra work
> (comparing to just using the code existing in downstream). But I guess
> that's the difference: vendor is mostly concerned about competitive
> advantage and getting to market fast, while upstream is more concerned
> about quality, considering all use cases, and having proper design.
> Anyway, we can work together to make it right, and to have both
> IP-cores support. In the worst case, if those are too different, we
> can have two separate drivers for those.
> 
>> The only real drawback of that way is having to add code for USIv2
>> inside the UART, HSI2C, and SPI drivers but in my opinion the benefits
>> overweigh the drawbacks greatly. We could even make the uart/spi/hsi2c
>> drivers call a helper function in the USI driver to set their USI_CON
>> and USI_OPTION registers up so that code would be shared and not
>> duplicated. Wether this patch gets applied like this is not my choice
>> though, I'll let the people responsible decide
>> :-)
>>
> 
> I'd argue that there are a lot of real drawbacks of using downstream
> driver as is. That's why I completely re-designed and re-implemented
> it. Downstream driver can't be built and function as a module, it
> doesn't respect System Register sharing between consumers, it leads to
> USI reset code duplication scattered across protocol drivers (that
> arguably shouldn't even be aware of that), it doesn't reflect HW
> structure clearly, it's not holding clocks needed for registers access
> (btw, sysreg clock can be provided in syscon node, exactly for that
> reason). As Krzysztof said, it also can't handle correct probe order
> and deferred probes. Downstream driver might work fine for some
> particular use-cases the vendor has, but in upstream it's better to
> cover more cases we can expect, as upstream kernel is used on more
> platforms, with more user space variants, etc.

Implementing USI in each of I2C/SPI/UART drivers is a big minus. Current
approach nicely encapsulates USI in dedicated driver without polluting
the other drivers with unrelated bus/protocol stuff.

Best regards,
Krzysztof
David Virag Nov. 29, 2021, 7:19 p.m. UTC | #5
On Mon, 2021-11-29 at 15:56 +0200, Sam Protsenko wrote:
> On Sun, 28 Nov 2021 at 05:15, David Virag <virag.david003@gmail.com>
> wrote:
> > 
> > On Sun, 2021-11-28 at 00:32 +0200, Sam Protsenko wrote:
> > > USIv2 IP-core provides selectable serial protocol (UART, SPI or
> > > High-Speed I2C); only one can be chosen at a time. This series
> > > implements USIv2 driver, which allows one to select particular USI
> > > function in device tree, and also performs USI block
> > > initialization.
> > > 
> > > With that driver implemented, it's not needed to do USI
> > > initialization
> > > in protocol drivers anymore, so that code is removed from the
> > > serial
> > > driver.
> > > 
> > 
> > I think the downstream way of doing this (USI node reg being on the
> > SW_CONF register itself rather than an offset from uart/i2c/spi, the
> > USI driver only controlling the SW_CONF, and the uart/i2c/spi drivers
> > controlling their USI_CON and USI_OPTION regs) is cleaner, better,
> > and
> > easier to adapt to USIv1 too.
> > 
> 
> One reason why I think it's better to provide SW_CONF register via
> syscon node, is that it helps us to avoid possible register access
> conflicts in future, and also conflicts when requesting corresponding
> resources. In other words, the System Register block can be used by
> many consumers (drivers) in future; those consumers might try to
> modify the same registers simultaneously, which might lead to race
> conditions (as RMW operation is not atomic), so some kind of
> serialization should be done (like locking in regmap), which is
> provided by syscon. Also, that wouldn't even come to that: you just
> can't request the same I/O area twice in Linux. So if SW_CONF is
> passed via "reg" property to USI driver, and then we try to map the
> whole System Register (or its portion that includes SW_CONF), that
> request would fail.

I've got to admit, that's something I didn't think about much, partly
because the lack of TRM on my hand, as I'm working with just vendor
kernel sources and consumer phones. What other things are in the sysreg
in your case? Looking at my vendor device tree, the USI SW_CONF
registers are at 0x10032000-0x10032008 in my case and the DT lacks
anything else close by (in the 0x1003xxxx region).

> 
> Although passing one SW_CONF register via "reg" might look easier to
> implement, it might also bring us all sort of problems later on. And I
> think a good design should account for such pitfalls.
> 
> As for the USI registers: I really don't think that duplicating the
> code for USI block reset across uart/i2c/spi drivers would help us to
> accomplish anything. Why those drivers should be even aware of USI
> reset? At least in USIv2 block, the USI registers and uart/i2c/spi
> registers are not mixed: they are located at different and always
> fixed addresses. We can benefit from that fact, and provide Device
> Tree structure which reflects the hardware one, separating USI control
> from actual protocol nodes.
> 
> > For example: I'm sure this is the case on USIv2 devices too, but on
> > Exynos7885, different devices have USI modes configured differently.
> > For example a Samsung Galaxy A8 (2018) has all the USI blocks
> > configured as SPI while a Samsung Galaxy M20 has the first USI
> > configured as dual HSI2C, the second as HSI2C on the first 2 pins and
> > the third as HSI2C on the last 2 pins. With this way of doing
> > everything on USIv2 we'd need 3 disabled USIv2 nodes in the SoC DTSI
> > for one USI block, each for every protocol the USI block can do, all
> > having a single child for their protocol and each referencing the
> > same
> > sysreg (not even sure if that's even supported). Then the board DTS
> > could enable the USI node it needs.
> > 
> 
> If I'm following you correctly, then it's not like that. I guess
> Krzysztof already replied to that, so I'll probably just repeat his
> words. In that case you'll have something like this in your SoC dtsi
> (for your USIv1 case of course, because dual HSI2C is not present in
> USIv2):
> 
> <<<<<<<<<<<<<<<<<<<<<<<<< cut here >>>>>>>>>>>>>>>>>>>>>>>
> usi1 {
>     spi1 {
>     };
> 
>     hsi2c1_1 {
>     };
> 
>     hsi2c1_2 {
>     };
> };
> 
> usi2 {
>     spi2 {
>     };
> 
>     hsi2c2_1 {
>     };
> };
> 
> 
> usi3 {
>     spi3 {
>     };
> 
>     hsi2c2_2 {
>     };
> };
> <<<<<<<<<<<<<<<<<<<<<<<<< cut here >>>>>>>>>>>>>>>>>>>>>>>
> 
> and then in your board dts you just have to enable corresponding usi's
> with proper modes, and enable chosen protocol nodes, like this:
> 
> <<<<<<<<<<<<<<<<<<<<<<<<< cut here >>>>>>>>>>>>>>>>>>>>>>>
> &usi1 {
>     status = "okay"
>     samsung,mode = <USI_V1_DUAL_I2C>;
> };
> 
> &hsi2c1_1 {
>     status = "okay"
> };
> 
> &hsi2c1_2 {
>     status = "okay"
> };
> <<<<<<<<<<<<<<<<<<<<<<<<< cut here >>>>>>>>>>>>>>>>>>>>>>>

What got me confused is the following: Upon checking vendor drivers I
was under the impression that we have all 3 protocols at seperate
addresses, and the USI SW_CONF register kind of works like a
multiplexer for the USI pins to switch between protocols. Now I see
that I was wrong, and the addresses are in fact the same. Now on a
hardware level it might still work just as a multiplexer but it
swithches the entire address space for a whole different protocol
block. Dumb little misunderstanding on my part, never mind! They are on
the same address even on USIv1. Not sure how I haven't noticed that
before, I guess since I never started experimenting with USI before,
just looked at the code as a reference I assumed a lot of things.

> 
> > With the downstream way we could have just one USI node and we could
> > add the 3 protocols it can do disabled as seperate or child nodes.
> > This
> > way the board DTS only needs to set the appropriate mode setting and
> > enable the protocol it needs. I'd say much better than having 3 USI
> > nodes for the same USI block.
> > 
> 
> Not sure if with downstream USI driver you can actually have protocols
> as sub-nodes in USI node though. It doesn't do anything like
> of_platform_populate().

It can't as far as I'm aware, I was just thinking that did seem like a
good idea to keep.

> 
> Also, with this USIv2 driver you can do the same thing you described:
> you can have just one USI node with 3 protocols as sub-nodes (or you
> can even have protocol nodes outside of USI node, but I'd not
> recommend that).
> 
> Actually I can see that it's my fault for not describing that case in
> bindings example. I'll make sure to do that in v2. You also got me
> thinking about default mode: sometimes SW_CONF reset value chooses
> some protocol. In that case maybe it'd useful to have something like
> USI_V2_DEFAULT, to tell driver to not touch SW_CONF at all.

Not sure if that's useful, I'm thinking we specify some protocol for
the USIs in board dts anyways, and if we don't, then we probably don't
use that USI block anyways, so at a minimum all protocols should be
probably disabled in that case, and probably the USI block as a whole
too. (SoC dtsi has them disabled, board dts doesn't touch them, so they
remain disabled). May I know how do you think a defult mode would be
useful?

> And also I
> can add USI_V2_NONE while at it, so that driver can write 0x0 to
> SW_CONF: that way no protocol will be selected. Maybe that can be
> beneficial for PM reasons, if some board doesn't use some USI blocks
> at all. Do you think it's feasible to add those two values to
> dt-bindings header? And is it possible to do so in USIv1?

I think I saw some downstream driver do something similiar, that sounds
like a good idea. In USIv1 I can see the HSI2C driver writing 0 to the
SW_CONF register at pm suspend. Not sure why that's in the HSI2C driver
rather than the USI but I'm guessing it should do the same thing as for
you. I have no TRM though, so not sure. We'll probably just have to
assume that's how it works here, maybe someone that has access to an
USIv1 SoC TRM could confirm? Probably won't get any response from
anyone who has it though.
 
> 
> > Also this way is pretty USIv2 centric. Adding USIv1 support to this
> > driver is difficult this way because of the the lack of USI_CON and
> > USI_OPTION registers as a whole (so having nowhere to actually set
> > the
> > reg of the USI node to, as the only thing USIv1 has is the SW_CONF
> > register). In my opinion being able to use the same driver and same
> > device tree layout for USIv1 and USIv2 is a definite plus
> > 
> 
> Well, it's USIv2 driver after all. I never expected it can be extended
> for USIv1 support. If you think it can be reused for USIv1, it's fine
> by me. But we need to consider next things:
>   - rename the driver to just "usi.c" (and also its configuration
> symbol)
>   - provide different compatible for USIv1 (and maybe corresponding
> driver data)
>   - rework bindings (header and doc); make sure existing bindings are
> intact (we shouldn't change already introduced interfaces)
>   - in case of USIv1 compatible; don't try to tinker with USIv2
> registers
>   - samsung,clkreq-on won't be available in case of USIv1 compatible
> 
> Because I don't have USIv1 SoC TRM (and neither do I possess some
> USIv1 board which I can use for test), I don't think it's my place to
> add USIv1 support. But I think it's possible to do so, using my input
> above.
> 
> I can see how it might be frustrating having to do some extra work
> (comparing to just using the code existing in downstream). But I guess
> that's the difference: vendor is mostly concerned about competitive
> advantage and getting to market fast, while upstream is more concerned
> about quality, considering all use cases, and having proper design.

It's not really the extra work, I just didn't see the benefits of this
way, and my misunderstanding caused me to not see how this would work.
I never really wanted to use the downstream driver as is, but in my
head I was thinking that "layout" should work.

> Anyway, we can work together to make it right, and to have both
> IP-cores support. In the worst case, if those are too different, we
> can have two separate drivers for those.
> 
> > The only real drawback of that way is having to add code for USIv2
> > inside the UART, HSI2C, and SPI drivers but in my opinion the
> > benefits
> > overweigh the drawbacks greatly. We could even make the
> > uart/spi/hsi2c
> > drivers call a helper function in the USI driver to set their
> > USI_CON
> > and USI_OPTION registers up so that code would be shared and not
> > duplicated. Wether this patch gets applied like this is not my
> > choice
> > though, I'll let the people responsible decide
> > :-)
> > 
> 
> I'd argue that there are a lot of real drawbacks of using downstream
> driver as is. That's why I completely re-designed and re-implemented
> it. Downstream driver can't be built and function as a module, it
> doesn't respect System Register sharing between consumers, it leads
> to
> USI reset code duplication scattered across protocol drivers (that
> arguably shouldn't even be aware of that), it doesn't reflect HW
> structure clearly, it's not holding clocks needed for registers
> access
> (btw, sysreg clock can be provided in syscon node, exactly for that
> reason). As Krzysztof said, it also can't handle correct probe order
> and deferred probes. Downstream driver might work fine for some
> particular use-cases the vendor has, but in upstream it's better to
> cover more cases we can expect, as upstream kernel is used on more
> platforms, with more user space variants, etc.

I do agree now, as I said a bit of a misunderstanding made me believe
this was wrong. (as if the addresses were different and the downstream
drivers worked the same way that would mean each USIv2 would have 3
sets of USI_CON and USI_OPTION registers for each protocol which would
definitely have to be handled somewhat differently.

> 
> I don't really think protocol drivers should be aware of USI
> registers
> at all, but if we they do -- we can provide some API from USIv2
> driver
> later, with EXPORT_SYMBOL(), referencing corresponding USI instance
> by
> phandle or using some other mechanism for inter-driver communication.
> 
> Of course, it's not my place to decide on patch acceptance too. But I
> was under the impression that maintainers would be ok with this
> course
> of actions. Also, upstream kernel seems to already follow the same
> design for some similar drivers. See for example
> drivers/soc/qcom/qcom_gsbi.c.
> 
> > Anyways, soon enough I can write an USIv1 driver after I submit all
> > the
> > 7885 stuff I'm working on currently. If you want to, you can add
> > USIv2
> > support to that driver, or if an USIv2 driver is already in
> > upstream at
> > that point, if it is written in the downstream way I can add v1
> > support
> > to that, or if it's like this I'll have to make a whole seperate
> > driver
> > with a whole seperate DT structure.
> > 
> 
> If it's like you said (USIv1 only touches the SW_CONF register), I
> guess USIv2 driver can be extended for USIv1 case. I already provided
> my thoughts on such rework above. It's probably better to consult
> with
> Krzysztof first. I guess the only way to figure out if it's feasible
> or it's better to have separate exynos-usi-v1.c for USIv1, is to try
> and add USIv1 support into USIv2 driver and see how pretty or ugly it
> is :) Whatever the way you decide to go with, please add me to Cc
> list
> when sending USIv1 patches.

Sure, I'll try doing it on top of the final version of your driver
then! Sorry for the misunderstanding there!

> 
> > Best regards,
> > David
Sam Protsenko Nov. 30, 2021, 12:01 a.m. UTC | #6
On Mon, 29 Nov 2021 at 21:19, David Virag <virag.david003@gmail.com> wrote:
>
> On Mon, 2021-11-29 at 15:56 +0200, Sam Protsenko wrote:
> > On Sun, 28 Nov 2021 at 05:15, David Virag <virag.david003@gmail.com>
> > wrote:
> > >
> > > On Sun, 2021-11-28 at 00:32 +0200, Sam Protsenko wrote:
> > > > USIv2 IP-core provides selectable serial protocol (UART, SPI or
> > > > High-Speed I2C); only one can be chosen at a time. This series
> > > > implements USIv2 driver, which allows one to select particular USI
> > > > function in device tree, and also performs USI block
> > > > initialization.
> > > >
> > > > With that driver implemented, it's not needed to do USI
> > > > initialization
> > > > in protocol drivers anymore, so that code is removed from the
> > > > serial
> > > > driver.
> > > >
> > >
> > > I think the downstream way of doing this (USI node reg being on the
> > > SW_CONF register itself rather than an offset from uart/i2c/spi, the
> > > USI driver only controlling the SW_CONF, and the uart/i2c/spi drivers
> > > controlling their USI_CON and USI_OPTION regs) is cleaner, better,
> > > and
> > > easier to adapt to USIv1 too.
> > >
> >
> > One reason why I think it's better to provide SW_CONF register via
> > syscon node, is that it helps us to avoid possible register access
> > conflicts in future, and also conflicts when requesting corresponding
> > resources. In other words, the System Register block can be used by
> > many consumers (drivers) in future; those consumers might try to
> > modify the same registers simultaneously, which might lead to race
> > conditions (as RMW operation is not atomic), so some kind of
> > serialization should be done (like locking in regmap), which is
> > provided by syscon. Also, that wouldn't even come to that: you just
> > can't request the same I/O area twice in Linux. So if SW_CONF is
> > passed via "reg" property to USI driver, and then we try to map the
> > whole System Register (or its portion that includes SW_CONF), that
> > request would fail.
>
> I've got to admit, that's something I didn't think about much, partly
> because the lack of TRM on my hand, as I'm working with just vendor
> kernel sources and consumer phones. What other things are in the sysreg
> in your case? Looking at my vendor device tree, the USI SW_CONF
> registers are at 0x10032000-0x10032008 in my case and the DT lacks
> anything else close by (in the 0x1003xxxx region).
>

Just in case, System Register is not a single register, but a register
block. In case of Exynos850 I have all sorts of registers in SYSREG.
Basically I have one SYSREG per domain, e.g. for PERI domain I have
SYSREG_PERI. Registers inside of each SYSREG may vary. SYREG_PERI has
IPCLK control register, SW_CONF registers, APB register, some USER
registers, etc, etc...

You can use something like this to find SYSREG info in your kernel:

    $ find drivers/ -type f -name '*7885*' -exec grep -Hni 'SYSREG' {} \;
    $ git grep -n --all-match -e SYSREG -e 1003 -- drivers/*7885*

Looking at Exynos7885 downstream kernel, you have next SYSREGs:

    SYSREG_PERI    0x10030000
    SYSREG_MIF0    0x10470000
    SYSREG_MIF1    0x10570000
    SYSREG_CPUCL0    0x10910000
    SYSREG_CPUCL1    0x10810000
    SYSREG_CPUCL2    0x10A10000
    SYSREG_APM    0x11C20000
    SYSREG_CORE    0x12010000
    SYSREG_FSYS    0x13420000

Those are base addresses for each sysreg. My wild guess, each SYSREG
size would be at least 0x10000.

SYSREG which contains SW_CONF registers for USI blocks is apparently
SYSREG_PERI. And SW_CONF offsets for each USI (inside of SYSREG_PERIO)
are:

    USI0: 0x2000
    USI1: 0x2004
    USI2: 0x2008

> >
> > Although passing one SW_CONF register via "reg" might look easier to
> > implement, it might also bring us all sort of problems later on. And I
> > think a good design should account for such pitfalls.
> >
> > As for the USI registers: I really don't think that duplicating the
> > code for USI block reset across uart/i2c/spi drivers would help us to
> > accomplish anything. Why those drivers should be even aware of USI
> > reset? At least in USIv2 block, the USI registers and uart/i2c/spi
> > registers are not mixed: they are located at different and always
> > fixed addresses. We can benefit from that fact, and provide Device
> > Tree structure which reflects the hardware one, separating USI control
> > from actual protocol nodes.
> >
> > > For example: I'm sure this is the case on USIv2 devices too, but on
> > > Exynos7885, different devices have USI modes configured differently.
> > > For example a Samsung Galaxy A8 (2018) has all the USI blocks
> > > configured as SPI while a Samsung Galaxy M20 has the first USI
> > > configured as dual HSI2C, the second as HSI2C on the first 2 pins and
> > > the third as HSI2C on the last 2 pins. With this way of doing
> > > everything on USIv2 we'd need 3 disabled USIv2 nodes in the SoC DTSI
> > > for one USI block, each for every protocol the USI block can do, all
> > > having a single child for their protocol and each referencing the
> > > same
> > > sysreg (not even sure if that's even supported). Then the board DTS
> > > could enable the USI node it needs.
> > >
> >
> > If I'm following you correctly, then it's not like that. I guess
> > Krzysztof already replied to that, so I'll probably just repeat his
> > words. In that case you'll have something like this in your SoC dtsi
> > (for your USIv1 case of course, because dual HSI2C is not present in
> > USIv2):
> >
> > <<<<<<<<<<<<<<<<<<<<<<<<< cut here >>>>>>>>>>>>>>>>>>>>>>>
> > usi1 {
> >     spi1 {
> >     };
> >
> >     hsi2c1_1 {
> >     };
> >
> >     hsi2c1_2 {
> >     };
> > };
> >
> > usi2 {
> >     spi2 {
> >     };
> >
> >     hsi2c2_1 {
> >     };
> > };
> >
> >
> > usi3 {
> >     spi3 {
> >     };
> >
> >     hsi2c2_2 {
> >     };
> > };
> > <<<<<<<<<<<<<<<<<<<<<<<<< cut here >>>>>>>>>>>>>>>>>>>>>>>
> >
> > and then in your board dts you just have to enable corresponding usi's
> > with proper modes, and enable chosen protocol nodes, like this:
> >
> > <<<<<<<<<<<<<<<<<<<<<<<<< cut here >>>>>>>>>>>>>>>>>>>>>>>
> > &usi1 {
> >     status = "okay"
> >     samsung,mode = <USI_V1_DUAL_I2C>;
> > };
> >
> > &hsi2c1_1 {
> >     status = "okay"
> > };
> >
> > &hsi2c1_2 {
> >     status = "okay"
> > };
> > <<<<<<<<<<<<<<<<<<<<<<<<< cut here >>>>>>>>>>>>>>>>>>>>>>>
>
> What got me confused is the following: Upon checking vendor drivers I
> was under the impression that we have all 3 protocols at seperate
> addresses, and the USI SW_CONF register kind of works like a
> multiplexer for the USI pins to switch between protocols. Now I see
> that I was wrong, and the addresses are in fact the same. Now on a
> hardware level it might still work just as a multiplexer but it
> swithches the entire address space for a whole different protocol
> block. Dumb little misunderstanding on my part, never mind! They are on
> the same address even on USIv1. Not sure how I haven't noticed that
> before, I guess since I never started experimenting with USI before,
> just looked at the code as a reference I assumed a lot of things.
>

Ah, yeah, USI block actually shares most of its internal circuits
within each protocol. So you can only choose one protocol per USI. I
should probably add that info to the bindings doc.

> >
> > > With the downstream way we could have just one USI node and we could
> > > add the 3 protocols it can do disabled as seperate or child nodes.
> > > This
> > > way the board DTS only needs to set the appropriate mode setting and
> > > enable the protocol it needs. I'd say much better than having 3 USI
> > > nodes for the same USI block.
> > >
> >
> > Not sure if with downstream USI driver you can actually have protocols
> > as sub-nodes in USI node though. It doesn't do anything like
> > of_platform_populate().
>
> It can't as far as I'm aware, I was just thinking that did seem like a
> good idea to keep.
>
> >
> > Also, with this USIv2 driver you can do the same thing you described:
> > you can have just one USI node with 3 protocols as sub-nodes (or you
> > can even have protocol nodes outside of USI node, but I'd not
> > recommend that).
> >
> > Actually I can see that it's my fault for not describing that case in
> > bindings example. I'll make sure to do that in v2. You also got me
> > thinking about default mode: sometimes SW_CONF reset value chooses
> > some protocol. In that case maybe it'd useful to have something like
> > USI_V2_DEFAULT, to tell driver to not touch SW_CONF at all.
>
> Not sure if that's useful, I'm thinking we specify some protocol for
> the USIs in board dts anyways, and if we don't, then we probably don't
> use that USI block anyways, so at a minimum all protocols should be
> probably disabled in that case, and probably the USI block as a whole
> too. (SoC dtsi has them disabled, board dts doesn't touch them, so they
> remain disabled). May I know how do you think a defult mode would be
> useful?
>

Yeah, you are right. I'll probably add USI_NONE configuration for 0x0.
Default one is really of no use.

> > And also I
> > can add USI_V2_NONE while at it, so that driver can write 0x0 to
> > SW_CONF: that way no protocol will be selected. Maybe that can be
> > beneficial for PM reasons, if some board doesn't use some USI blocks
> > at all. Do you think it's feasible to add those two values to
> > dt-bindings header? And is it possible to do so in USIv1?
>
> I think I saw some downstream driver do something similiar, that sounds
> like a good idea. In USIv1 I can see the HSI2C driver writing 0 to the
> SW_CONF register at pm suspend. Not sure why that's in the HSI2C driver
> rather than the USI but I'm guessing it should do the same thing as for
> you. I have no TRM though, so not sure. We'll probably just have to
> assume that's how it works here, maybe someone that has access to an
> USIv1 SoC TRM could confirm? Probably won't get any response from
> anyone who has it though.
>

I guess it's enough to have that kernel source code to figure out
essentials. When you set 0x0, no protocol is chosen, so we can imagine
roughly what happens inside of USI IP-core (internal circuits are not
connected, muxes are opened, etc). As I understand, 0x0 might be the
reset value for some SW_CONF registers, so it'll appear on PM resume,
so one should set SW_CONF on resume again (which is done in my driver
already).

> >
> > > Also this way is pretty USIv2 centric. Adding USIv1 support to this
> > > driver is difficult this way because of the the lack of USI_CON and
> > > USI_OPTION registers as a whole (so having nowhere to actually set
> > > the
> > > reg of the USI node to, as the only thing USIv1 has is the SW_CONF
> > > register). In my opinion being able to use the same driver and same
> > > device tree layout for USIv1 and USIv2 is a definite plus
> > >
> >
> > Well, it's USIv2 driver after all. I never expected it can be extended
> > for USIv1 support. If you think it can be reused for USIv1, it's fine
> > by me. But we need to consider next things:
> >   - rename the driver to just "usi.c" (and also its configuration
> > symbol)
> >   - provide different compatible for USIv1 (and maybe corresponding
> > driver data)
> >   - rework bindings (header and doc); make sure existing bindings are
> > intact (we shouldn't change already introduced interfaces)
> >   - in case of USIv1 compatible; don't try to tinker with USIv2
> > registers
> >   - samsung,clkreq-on won't be available in case of USIv1 compatible
> >
> > Because I don't have USIv1 SoC TRM (and neither do I possess some
> > USIv1 board which I can use for test), I don't think it's my place to
> > add USIv1 support. But I think it's possible to do so, using my input
> > above.
> >
> > I can see how it might be frustrating having to do some extra work
> > (comparing to just using the code existing in downstream). But I guess
> > that's the difference: vendor is mostly concerned about competitive
> > advantage and getting to market fast, while upstream is more concerned
> > about quality, considering all use cases, and having proper design.
>
> It's not really the extra work, I just didn't see the benefits of this
> way, and my misunderstanding caused me to not see how this would work.
> I never really wanted to use the downstream driver as is, but in my
> head I was thinking that "layout" should work.
>
> > Anyway, we can work together to make it right, and to have both
> > IP-cores support. In the worst case, if those are too different, we
> > can have two separate drivers for those.
> >
> > > The only real drawback of that way is having to add code for USIv2
> > > inside the UART, HSI2C, and SPI drivers but in my opinion the
> > > benefits
> > > overweigh the drawbacks greatly. We could even make the
> > > uart/spi/hsi2c
> > > drivers call a helper function in the USI driver to set their
> > > USI_CON
> > > and USI_OPTION registers up so that code would be shared and not
> > > duplicated. Wether this patch gets applied like this is not my
> > > choice
> > > though, I'll let the people responsible decide
> > > :-)
> > >
> >
> > I'd argue that there are a lot of real drawbacks of using downstream
> > driver as is. That's why I completely re-designed and re-implemented
> > it. Downstream driver can't be built and function as a module, it
> > doesn't respect System Register sharing between consumers, it leads
> > to
> > USI reset code duplication scattered across protocol drivers (that
> > arguably shouldn't even be aware of that), it doesn't reflect HW
> > structure clearly, it's not holding clocks needed for registers
> > access
> > (btw, sysreg clock can be provided in syscon node, exactly for that
> > reason). As Krzysztof said, it also can't handle correct probe order
> > and deferred probes. Downstream driver might work fine for some
> > particular use-cases the vendor has, but in upstream it's better to
> > cover more cases we can expect, as upstream kernel is used on more
> > platforms, with more user space variants, etc.
>
> I do agree now, as I said a bit of a misunderstanding made me believe
> this was wrong. (as if the addresses were different and the downstream
> drivers worked the same way that would mean each USIv2 would have 3
> sets of USI_CON and USI_OPTION registers for each protocol which would
> definitely have to be handled somewhat differently.
>

I've checked USIv2 driver code in Exynos7885 kernel (publicly
available), and it looks like it would be relatively easy to add that
to the driver I submitted. Please wait for my series to be Acked or
applied, then you can go ahead and send your additions on top of that.
I don't want to do that, as I don't have any HW I can validate that,
so it doesn't make much sense.

> >
> > I don't really think protocol drivers should be aware of USI
> > registers
> > at all, but if we they do -- we can provide some API from USIv2
> > driver
> > later, with EXPORT_SYMBOL(), referencing corresponding USI instance
> > by
> > phandle or using some other mechanism for inter-driver communication.
> >
> > Of course, it's not my place to decide on patch acceptance too. But I
> > was under the impression that maintainers would be ok with this
> > course
> > of actions. Also, upstream kernel seems to already follow the same
> > design for some similar drivers. See for example
> > drivers/soc/qcom/qcom_gsbi.c.
> >
> > > Anyways, soon enough I can write an USIv1 driver after I submit all
> > > the
> > > 7885 stuff I'm working on currently. If you want to, you can add
> > > USIv2
> > > support to that driver, or if an USIv2 driver is already in
> > > upstream at
> > > that point, if it is written in the downstream way I can add v1
> > > support
> > > to that, or if it's like this I'll have to make a whole seperate
> > > driver
> > > with a whole seperate DT structure.
> > >
> >
> > If it's like you said (USIv1 only touches the SW_CONF register), I
> > guess USIv2 driver can be extended for USIv1 case. I already provided
> > my thoughts on such rework above. It's probably better to consult
> > with
> > Krzysztof first. I guess the only way to figure out if it's feasible
> > or it's better to have separate exynos-usi-v1.c for USIv1, is to try
> > and add USIv1 support into USIv2 driver and see how pretty or ugly it
> > is :) Whatever the way you decide to go with, please add me to Cc
> > list
> > when sending USIv1 patches.
>
> Sure, I'll try doing it on top of the final version of your driver
> then! Sorry for the misunderstanding there!
>
> >
> > > Best regards,
> > > David
>