Hi Thierry,
thanks for the reply and having this discussion!
On 24/11/17 13:31, Thierry Reding wrote:
> On Fri, Nov 24, 2017 at 12:04:12PM +0000, Andre Przywara wrote:
>> Hi,
>>
>> (adding linux-sunxi, which I forgot at the initial post).
>>
>> On 24/11/17 10:52, Thierry Reding wrote:
>>> On Fri, Nov 24, 2017 at 11:19:52AM +0100, Linus Walleij wrote:
>>>> On Mon, Nov 13, 2017 at 2:25 AM, Andre Przywara <andre.przywara@xxxxxxx> wrote:
>>>>
>>>>> So far all the Allwinner pinctrl drivers provided a table in the
>>>>> kernel to describe all the pins and the link between the pinctrl functions
>>>>> names (strings) and their respective mux values (register values).
>>>>>
>>>>> Extend the binding to put those mappings in the DT, so that any SoC can
>>>>> describe its pinctrl and GPIO data fully there instead of relying on
>>>>> tables.
>>>>> This uses a generic compatible name, to be prepended with an SoC
>>>>> specific name in the node.
>>>
>>> This seems backwards to me. I'm not sure if Rob has any hard rules on
>>> this, but in the past I've seen a lot of drivers stick this kind of data
>>> into drivers. I personally also prefer that approach because the data is
>>> completely static and there's no way for any specific board to customize
>>> it. So the tables are in fact implied completely by the SoC compatible
>>> string.
>>
>> But this is just *data*, and I believe it is actually package specific.
>> We need the DT to describe the relation between devices and pins anyway,
>> it's just that we use arbitrary strings today instead of the actual
>> register value. This is what the generic pinmux property fixes.
>
> Register values don't belong in a device tree.
I don't think that's true in a general way. This "pinmux" is already an
accepted property and used for exactly that purpose in other pinctrl
drivers:
$ git grep -l '[^,]pinmux = ' arch/arm{,64}/boot/dts
Plus the fsl,pinmux-ids property, which seems to serve the same purpose.
> And it's totally fine to
> have data in the kernel, too. We do it all the time.
I know, but that does not mean that I need to like it or should avoid
finding better solutions.
>>> Moving all of this data into device tree has a number of disadvantages:
>>>
>>> * Existing boards already use the static tables in the driver, and the
>>> device trees don't contain any data, so you can't get rid of any of
>>> the existing tables because it would break ABI.
>>
>> Yes, my DeLorean is in the garage, so I can't really change this anymore
>> ;-) But that doesn't mean we have to go on with this forever, I think.
>
> ABI stability means that, yes, you have to keep maintaining the existing
> tables forever, else old DTBs will stop working if you rip out the
> static tables that drivers depend on.
I think you got me wrong, I am totally not arguing ABI stability and its
consequences. That's why this binding is mostly for newer SoCs, but also
designed to stay compatible with the old binding - that's for other DT
users. Of course we need to keep the tables in the kernel.
>>> * Moving the table into the DT doesn't actually solve anything because
>>> the driver would have to validate the DT description to make sure it
>>> contains valid data. And in order to validate DT content, the driver
>>> would need a copy of the table anyway.
>>
>> I don't get what the driver would need to validate? We rely on DT
>> information to be correct anyway, otherwise your board just won't work.
>> If the DT is wrong, you have much bigger problems.
>
> Given that DT is an ABI you should treat it the same way as other ABIs.
> You can't rely on the DT being correct.
But it's not a *user* ABI, where unprivileged users can supply input data.
What else can you rely on then if not on the DT?
By definition you can't be able to validate every bit of it. How would
you know that the reg address of that SATA chip is correct and not
actually pointing to the "set-board-on-fire" device?
And what about those regulators, where it gives you the allowed voltage
range?
> Rather you have to make sure to
> validate it before you hand the content to hardware. If you allow direct
> register access to your hardware via DT and don't validate, it becomes
> really easy for people to exploit it.
If you can change the DT, you already have full control over the system.
Ideally the DT comes as part of the firmware, so if you control this,
there are far less complicated ways to burn down the system. And even if
you supply the DT with your kernel, you must have control and could
trigger the set-fire device via /dev/mem or the like.
> This is not the same as saying we need to be able to fully validate all
> aspects of device tree. We can't, because some information simply does
> not exist outside of DT. However, I think it's a big mistake to trust a
> user to fully know about all intricacies of a pinmux and not make any
> mistake when writing the device tree.
When does a *user* write a device tree? What would a clueless Joe
Average expect from doing so? The device tree is written by either the
board/SoC vendor or by some kernel developers. It is not meant to be
changed by the user, apart from carefully crafted overlays, maybe. You
can have the same control by just changing the kernel (binary patching
the image file or re-compiling with
writel(MATCHES, base_addr + SET_FIRE)).
> What if one of the settings causes the board to go up in flames?
Then you better not play with it. But I don't think this is a valid
argument, really. What if gravity reverses tomorrow?
Are there any known issues with writing mux values to pinctrl registers?
I don't think the consequences would be different from putting low or
high to a GPIO, which you can do already from userland.
>> Actually we gain something, because we only commit information that can
>> actually be tested. Right now we have a lot of information which is
>> copied from the manual, and nobody knows if pin H24 on the A10 is really
>> PATA-CS1 or not. Plus we have bugs when creating the table, plus
>> copy&paste bugs. I found some while grep-ing for patterns - will send
>> fixes ASAP.
>
> That's a different matter. If you've got bugs in the tables, then go fix
> the tables. However the assumption here is that you've done at least a
> minimum of testing and your driver didn't cause your board to go up in
> flames.
Yes, but at the moment we can't test all of the settings we put in the
kernel table. Some pins are for devices for which we don't have drivers
in the kernel (yet), for some we probably never will (JTAG).
>From writing this table for the A64 before I know it's tedious and even
with review there are bugs left. If we can decrease this risk, I am all
for it.
> When patches were posted, people had the opportunity to review
> the tables for correctness. However, if you put all of the flexibility
> into DT, you also put all of the risk there. People may just make some
> stupid mistake and cause physical damage to their hardware.
I don't see how a DT is more risky. Conceptually I consider the DT being
part of the firmware, so one trust level above the kernel. I understand
that some people have different experiences with "firmware", but frankly
I believe this discussion is pointless. If you get the DT wrong, your
board won't work. If the system easily allows you to destroy it by means
of software, you might want to fix it in general, for instance by
telling firmware to not allow access to it from the kernel. Or by
designing hardware in a way that you don't need to expose critical
settings to software. That how x86 solves most of it and what makes it
hard to physically kill a rented root server.
Hiding stuff from the DT and putting molly guards in the kernel does not
really solve this problem.
>> In the moment all the table gives us is a mapping between a *string* and
>> the respective mux register value (per pin), plus the number of pins in
>> each bank. This can *easily* be put in the DT and should belong there.
>
> Why? This is data that is implied by the compatible string and static
> per SoC. There is no way you can change the mapping in DT. What does
> need to go into DT is the configuration of the pinmux, that is, what
> function is used for each pin on a given board.
Exactly, and that does not change on the *board* level, because there we
just refer to the pinctrl subnode phandles, defined in <soc>.dtsi.
And it's just there that I want to add the right pinmux value.
>> Actually I believe that the current binding is not correct, because it
>> makes those mux strings a part of the binding, though this is not
>> documented anywhere. A developer cannot take the binding and write a
>> working driver or even a DT without looking at the code.
>> Plus we already changed those names in the past (for instance commit
>> bc0f566a98c4), basically breaking compatibility.
>
> If you haven't documented the strings your binding is not complete.
> That's a bug and should be fixed. Also, it is occasionally acceptable to
> break compatibility (it's technically only breaking if somebody notices)
> and fixing bugs in bindings has in the past been one of the exceptions
> where breaking ABI was specifically allowed.
>
> However, the kind of breakage we're talking about here is total. If you
> rip out the static tables from your driver, you don't have any data to
> replace the missing information and none of the driver will work. This
> is different from the driver erroring out trying to configure a pin for
> the NAND function because it couldn't match the name.
I think this is a misunderstanding. As mentioned above I never proposed
to remove the tables from the kernel. I think I said this explicitly in
several places:
PATCH 0/3:
> Of course we can't remove the existing tables, to keep compatiblity
> with older DTs, but at least we don't need to add more tables in the
> future.
PATCH 1/3:
> +The binding described above can be extended in this manner to be
supported
> +by *both* an existing driver and some generic driver. Existing
drivers will
> +ignore the new properties and revert to their internal table instead.
I totally understand the concept of compatibility, ask Maxime, he hates
me for that ;-)
> Also, device tree bindings are not documentation for how to write a
> driver. They are not a replacement for hardware documentation. Nobody
> should be expected to be able to write an OS driver solely based on a
> device tree binding. Device tree bindings are more of a configuration
> interface specification for OS drivers.
Yes, but together with the hardware docs you should be able to write a
driver. And here you can't, because you are missing the strings. So a
BSD developer has to look at Linux code.
Sure you can go ahead and add those strings to the binding, but I wonder
how complicated it has to get until we just go with the obviously
simplest way and just add this innocent mux value to the DT.
>>> I don't think you're going to do yourself any favours by pushing this. I
>>> also don't see the commit description give any reason why you want to
>>> move the table into device tree. Do you see any advantages in doing so?
>>
>> We stop adding tables with SoC specific *data* in the kernel *code*
>> base. With being boolean Kconfig options, this gets added to every
>> single-image kernel.
>
> The kernel is full of data:
>
> $ objdump -h build/arm64/vmlinux
> [...]
> 1 .text 009c99c0 ffff000008081000 ffff000008081000 00011000 2**11
> CONTENTS, ALLOC, LOAD, READONLY, CODE
> 2 .rodata 00403bf8 ffff000008a50000 ffff000008a50000 009e0000 2**12
> CONTENTS, ALLOC, LOAD, DATA
> [...]
>
> So that's about 40% of the kernel image. Code really is no good without
> data to process.
But how much of this is SoC specific configuration data? How much is it
in x86? Yes, historically we had and have a lot of configuration data in
ARM kernels. But that doesn't mean that we have to continue with this or
even increase the share.
>> More important: those tables help Linux, but other DT consumers (*BSD,
>> U-Boot) have to replicate them, which is just wrong, IMHO.
>
> Yeah, I've heard this before. To be honest, I think these tables are the
> kind of data that you should generate, and once you do that it becomes
> extremely cheap to add the data to other DT consumers.
I don't get what's the point of duplicating this information everywhere.
We already have places in the DT to put this information:
function = "uart1";
So we add "pinmux = <2>" *once* to *the* DT, and everyone gets this for
free. If U-Boot wants to enable MMC0, it just follows the phandle from
pinctrl-0 and writes the value(s) from "pinmux" to the registers derived
from the member of the "pins" string list. No SoC knowledge required.
> And let's face it: the really difficult part of adding pinmux support is
> to write the driver (or subsystem if you don't have one yet). Adding the
> data is really the easy part.
I know, I did this before. But currently you have to write both the DT
part *and* the kernel table. And check patch 3/3, that's what the table
gets reduced to. So you avoid writing 601 lines, instead add 23 lines to
the DT.
>> I believe the kernel is a nice collection of really good code for
>> complicated file systems, high performance network protocols and
>> sophisticated memory management, among others. It shouldn't be a dumping
>> ground for arbitrary, very SoC specific information. Cf. LinusT 2011.
>> DT is out there to fix this, so we should do so.
>
> Every driver is very SoC specific information. There's never been an
> objection to having SoC specific drivers in the kernel. And back at the
> time the discussion was as much about the development process and code
> structure than it was about board files.
>
> The majority of the improvements over the years have been achieved by
> moving drivers out of arch/arm and moving board files to DT. The goal
> was never to get rid of all data.
Sure, not all data. But if we have the relatively easy opportunity to
avoid further addition of data, we should do it, I believe.
This significantly reduces the amount of kernel code we need to add to
support new SoCs.
Cheers,
Andre.
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
