Re: [PATCH 2/2] mfd: Add Baikal-T1 Boot Controller driver

From: Sergey Semin
Date: Fri Mar 27 2020 - 12:36:35 EST


Hello Miquel,

Thanks for clarifications. My comments are blow.

On Fri, Mar 27, 2020 at 11:25:30AM +0100, Miquel Raynal wrote:
> Hi Lee, Sergey,
>
> Lee Jones <lee.jones@xxxxxxxxxx> wrote on Fri, 27 Mar 2020 09:01:39
> +0000:
>
> > On Thu, 26 Mar 2020, Sergey Semin wrote:
> >
> > > Michael, Richard, Vignesh and MTD mailing list are Cc'ed to have their
> > > comments on the issue.
> > >
> > > My answers on the previous email are below.
> > >
> > > On Thu, Mar 26, 2020 at 09:13:13AM +0000, Lee Jones wrote:
> > > > On Wed, 25 Mar 2020, Sergey Semin wrote:
> > > >
> > > > > Hello Lee,
> > > > >
> > > > > On Wed, Mar 25, 2020 at 10:09:40AM +0000, Lee Jones wrote:
> > > > > > On Fri, 06 Mar 2020, Sergey.Semin@xxxxxxxxxxxxxxxxxxxx wrote:
> > > > > >
> > > > > > > From: Serge Semin <Sergey.Semin@xxxxxxxxxxxxxxxxxxxx>
> > > > > > >
> > > > > > > Baikal-T1 Boot Controller is an IP block embedded into the SoC and
> > > > > > > responsible for the chip proper pre-initialization and further
> > > > > > > booting up from some memory device. From the Linux kernel point of view
> > > > > > > it's just a multi-functional device, which exports three physically mapped
> > > > > > > ROMs and a single SPI controller.
> > > > > > >
> > > > > > > Primarily the ROMs are intended to be used for transparent access of
> > > > > > > the memory devices with system bootup code. First ROM device is an
> > > > > > > embedded into the SoC firmware, which is supposed to be used just for
> > > > > > > the chip debug and tests. Second ROM device provides a MMIO-based
> > > > > > > access to an external SPI flash. Third ROM mirrors either the Internal
> > > > > > > or SPI ROM region, depending on the state of the external BOOTCFG_{0,1}
> > > > > > > chip pins selecting the system boot device.
> > > > > > >
> > > > > > > External SPI flash can be also accessed by the DW APB SSI SPI controller
> > > > > > > embedded into the Baikal-T1 Boot Controller. In this case the memory mapped
> > > > > > > SPI flash region shouldn't be accessed.
> > > > > > >
> > > > > > > Taking into account all the peculiarities described above, we created
> > > > > > > an MFD-based driver for the Baikal-T1 controller. Aside from ordinary
> > > > > > > OF-based sub-device registration it also provides a simple API to
> > > > > > > serialize an access to the external SPI flash from either the MMIO-based
> > > > > > > SPI interface or embedded SPI controller.
> > > > > >
> > > > > > Not sure why this is being classified as an MFD.
> > > > > >
> > > > > > This is clearly 'just' a memory device.
> > > > > >
> > > > >
> > > > > Hm, I see this as a normal MFD device. The Boot controller provides a
> > > > > set of physically mapped ROMs and a DW APB SSI-based embedded SPI
> > > > > controller. Yes, the SPI controller is normally utilized to access an external
> > > > > flash device, and at boot stage it is used for it. But still it's a SPI
> > > > > controller which driver belongs to the kernel SPI subsystem. Moreover
> > > > > nothing prevents a platform designer from using it together with custom
> > > > > GPIO-based chip-selects to have additional devices on the SPI bus.
> > > > >
> > > > > As I said the problem is that an SPI flash might be accessed either with
> > > > > use of a physically mapped ROM or via the normal DW APB SPI controller.
> > > > > These two interfaces can't be used simultaneously, because the ROM is
> > > > > just an rtl state-machine, which is built to translate MMIO operations
> > > > > through the SPI controller registers to an external SPI-nor at CS0 of
> > > > > the interface. That's why first I need to make sure the interface is locked,
> > > > > then being enabled, then the corresponding driver can use it, then get
> > > > > to unlock. That's the point of having the __bt1_bc_spi_lock() and
> > > > > bt1_bc_spi_unlock() methods exported from the driver.
> > > > >
> > > > > I've got two drivers for MTD ROM and SPI controller based on that
> > > > > methods. But I haven't submitted them yet, because they belong to two
> > > > > different subsystems and I need to have this one being accepted first.
> > > >
> > > > This is not a totally unique device/situation. I've seen (and NACKed)
> > > > this type of device before. You need to explain this to the MTD
> > > > (SPI-NOR?) maintainers. They should be in a good position to provide
> > > > guidance.
> > > >
> > >
> > > Sorry, I don't really understand your justification. The boot controller
> > > exports two types of devices: physically mapped ROMs and an DW APB
> > > SSI-based SPI controller. Aside from being able to access an externally
> > > attached SPI flash the SPI controller has normal SPI interface which in
> > > general can be used to access any other SPI-slave. The complexity of
> > > the case is that one of physically mapped ROM RTL uses the DW APB SSI
> > > controller to access an external SPI flash, which when done makes the
> > > DW APB SSI registers unusable on direct way. That's why I implemented the
> > > boot controller as an MFD. An alternation caused by this peculiarity
> > > will be submitted to drivers/mtd/maps/physmap-{core.c,baikal-t1-rom.c}
> > > later after this change is reviewed and accepted. Similar situation
> > > concerns a driver of DW APB SSI-based SPI controller. So according to
> > > the current design the boot controller with it' sub-devices will be
> > > declared in dts as follows:
> > >
> > > boot: boot@1f040000 {
> > > compatible = "be,bt1-boot-ctl";
> > > reg = <0x1f040000 0x100>;
> > > #address-cells = <1>;
> > > #size-cells = <1>;
> > > ranges;
> >
> > What control does this device offer in those 0x100 registers?
> >
> > > clocks = <&ccu_sys CCU_SYS_APB_CLK>;
> > > clock-names = "pclk";
> > >
> > > int_rom: rom@1bfc0000 {
> > > compatible = "be,bt1-int-rom", "mtd-rom";
> > > reg = <0x1bfc0000 0x10000>;
> > > ...
> > > };
> > >
> > > spi_rom: rom@1c000000 {
> > > compatible = "be,bt1-ssi-rom", "mtd-rom";
> > > reg = <0x1c000000 0x1000000>;
> > > ...
> > > };
> > >
> > > spi0: spi@1f040100 {
> > > compatible = "be,bt1-boot-ssi";
> > > reg = <0x1f040100 0xf00>;
> > > #address-cells = <1>;
> > > #size-cells = <0>;
> > >
> > > clocks = <&ccu_sys CCU_SYS_SPI_CLK>;
> > > clock-names = "ref";
> > >
> > > ...
> > > };
> > >
> > > boot_rom: rom@1fc00000 {
> > > compatible = "be,bt1-boot-rom", "mtd-rom";
> > > reg = <0x1fc00000 0x400000>;
> > > ...
> > > };
> > > };
> > >
> > > I dare to assume, that by saying: "Despite including the MFD API, I don't
> > > see it being used at all." you meant that since the driver doesn't
> > > redistribute any resource by declaring mfd_cell'es, doesn't
> > > register mfd-based sub-devices, and primary use-case of the boot
> > > controller is to access flash-devices, it should be just moved to the MTD
> > > subsystem. I don't think it would be better than to have a common part
> > > defined here in MFD while ROM-specific part - in MTD, and SPI-specific - in
> > > the SPI subsystems. I would consider Baikal-T1 Boot Controller being similar
> > > to drivers/mfd/qcom-spmi-pmic.c, drivers/mfd/atmel-flexcom.c, etc, but
> > > instead of having GPIO, RTC, UART, i2c, etc sub-devices (which are also
> > > fully defied in dts), it exports MMIO-based ROMs and SPI-controller.
> >
> > Are the ROMs all controlled via SPI?
> >
> > > Lee, explain me please what is the difference between these MFDs and
> > > Baikal-T1 Boot Controller, that makes one simple MFDs suitable for the
> > > MFD subsystem, while another isn't?
> >
> > Complexity.
> >
> > [NB: Please Don't assume that just because I accepted a driver into
> > MFD 6 years ago, that it would be accepted again today.]
> >
> > To me this looks like an MTD device which is controlled via SPI.
> >
> > The way I'm reading this currently, it might serve well to make the
> > MTD portion(s) children of the SPI controller. I still do not see a
> > compelling reason to warrant adding a superfluous MFD driver if at all
> > avoidable.
> >
> > > Miquel, Richard, Vignesh and everyone from MTD, who can help could you
> > > join this discussion and clarify whether the Baikal-T1 Boot Controller
> > > driver is supposed to be moved to the MTD subsystem? If so, then what is
> > > the better place to put it within the drivers/mtd/ directory tree?
> > >
>
> Thank you for the detailed explanation. I'll try to bring useful
> information to sort this out. IIUC, this bloc exports:
>

> 1/ One ROM located in the SoC
> 2/ The access to a possible second ROM over SPI
>

Absolutely right. In this case these are the physically mapped ROMs in
the mtd subsystem notation. So they are the subjects of
drivers/mtd/maps/physmap-* driver. Though these ROMs require a
word-aligned access, so I had to implement a dedicated
driver/mtd/maps/physmap-bt1-rom.c code. I'll submit it after we
finish with this patchset, settle the issues raised in its framework.

> And also:
>
> 3/ Access to the SPI controller itself

Absolutely right. It's a set of DW APB SSI registers, so normally it
would be a subject of drivers/spi/{spi-dw-mmio.c,spi-dw.c} driver, but
due to very limited resources (no DMA, no IRQ, just 8 bytes FIFO, a
single native chip-select) and a racy access from 2/, that code won't
work, but would only after a serious refactoring, so I had to create
a dedicated driver for it.

> 4/ Access to 1/ or 2/ depending on an external switch.
>

Yes, it's another physically mapped ROM, which is a mirror of either 1/
or 2/ depending on the SoC boot mode strapped at the chip startup.

So to speak 1/ is always mapped (it's Internal ROM), 2/ is a physically
mapped SPI ROM, 3/ is an SPI controller itself primarily dedicated to
access the SPI ROM, and 4/ is a mirror of 1/ or 2/. The problem is in
simultaneous access to 2/ and 3/ and to 4/ when the SPI-boot mode is enabled.
Here is the Boot Controller structure together with comments of how my
current drivers design is using it:

Boot Controller (0x1f040000 0x100) - root MFD device.
|
+---+-> 1/ Internal ROM (0x1bfc0000 0x10000) - seen as /dev/mtdX (64KB of internal firmware).
| \
+---+-> 4/ Boot ROM (0x1fc00000 0x400000) - seen as /dev/mtdY (4MB mirror of either Internal ROM or SPI ROM depending on the SoC boot mode strapped at the chip startup).
| /
+---+-> 3/ SPI ROM (0x1c000000 0x1000000) - seen as /dev/mtdZ.
| |
| \ - This is a switch, which is handled by a flag in the Boot Controller register 0x1f040000.
+-- +-> 2/ DW APB SSI registers (0x1f040100 0xf00) - seen as SPI-dev +------> Connected to CS0 external SPI-nor flash - seen as mtdN
|
+------> Might be connected to any other SPI-slave with using GPIO-based chip-select.

So as you can see, a system software can access either SPI/Boot ROM or DW APB SSI
registers at a time, not simultaneously. That's why I needed a dedicated
API, which would serialize an access to the DW APB SSI registers. This
is what the MFD driver submitted by this patchset is intended for. It
provides a lock, which makes sure that either SPI/Boot ROM or DW APB SSI
is enabled while lock is held. In addition the MFD driver is responsible
for the sub-devices population.

See the https://lkml.org/lkml/2020/3/27/273 message for some more
details.

> IMHO:
>
> 1/ Might be seen as a random MTD device, its driver should be in
> /drivers/mtd/devices I guess.

Agreed. I've got a driver for it, but it will be
drivers/mtd/maps/physmap-bt1-rom.c (will use the common code
drivers/mtd/maps/physmap-core.c), since it's just a ROM. Alas the ROM
will have no direct mapping due to the word-aligned access restriction
(no-unaligned-direct-access property). Due to this restriction I also had
to create a dedicated code in drivers/mtd/maps/physmap-bt1-rom.c instead
of using generic "mtd-rom" compatible binding.

> 3/ Is a SPI controller, its driver should be memory agnostic and
> located in /driver/spi/. However, it should probably implement
> the spi-mem infrastructure *and* the direct-mapping infrastructure
> which would automatically cover 2/ as well. The reg property of 2/
> should probably be part of 3/.
>

My new SPI-controller driver does implement the spi-mem infrastructure,
but it also implements a generic SPI callbacks so to allow any SPI-slave
(not only the Boot SPI-flash) being connected to the SPI bus. The
generic SPI callbacks are also required if GPIO-based chip-select is
available for the SPI-flash (don't really know why you permitted this
restriction in spi_mem_exec_op() instead of enabling the slave device at
the time of mem_op-based communications). Since 2/ and 4/ are handled by
drivers/mtd/maps/physmap-bt1-rom.c, there is no direct-mapping
implemented for this SPI-controller.

> For 4/ I don't know what is the right thing to do yet.
>

That's one of the reason why I implemented it in a way it's implemented.
1/, 2/ and 4/ are the physically mapped ROMs handled by the
drivers/mtd/maps/physmap-bt1-rom.c driver, while 3/ is just an SPI-controller.
And all of them are parts of a single MFD - Baikal-T1 Boot Controller.
Moreover current Boot Controller DT-nodes structure is as much as possible
close to the structure of the device and its subnodes from hardware
point of view. Such design looked neat and elegant for me. It also
covered all parts of the RTL block.

> What do you think?
>

Yes, your design was my first thought. I also had in mind to implement
2/ as a direct mapping of 3/ . But then I didn't find spi-nor using the
direct mapping interface, I also didn't come up with a solution for 4/
and seeing 1/, 2/ and 4/ having the same properties I decided that it
would be better to make them being exported as a dedicated mtd'es.

So, Miquèl, do you still think that my decision was wrong and it would be
better to create your version of the design? If so, folks, what to do with 4/
then, just drop?

Regards,
-Sergey

> > > > > Recently I've sent an RFC regarding a different question, but it
> > > > > concerns the Baikal-T1 system controller and the boot controller as being part
> > > > > of it:
> > > > >
> > > > > https://lkml.org/lkml/2020/3/22/393
>
>
> Thanks,
> Miquèl