Re: [RFC 01/32] Kconfig: introduce and depend on LEGACY_PCI

Em Tue, 28 Dec 2021 11:58:55 +0100
Niklas Schnelle [off-list ref] escreveu:

On Tue, 2021-12-28 at 10:15 +0100, Mauro Carvalho Chehab wrote:

quoted

Em Tue, 28 Dec 2021 09:21:23 +0100
Greg Kroah-Hartman [off-list ref] escreveu:
  

quoted

On Mon, Dec 27, 2021 at 05:42:46PM +0100, Niklas Schnelle wrote:  

quoted

--- a/drivers/pci/Kconfig
+++ b/drivers/pci/Kconfig

@@ -23,6 +23,17 @@ menuconfig PCI
 
 if PCI
 
+config LEGACY_PCI
+	bool "Enable support for legacy PCI devices"
+	depends on HAVE_PCI
+	help
+	   This option enables support for legacy PCI devices. This includes
+	   PCI devices attached directly or via a bridge on a PCI Express bus.
+	   It also includes compatibility features on PCI Express devices which
+	   make use of legacy I/O spaces.    

This Kconfig doesn't seem what it is needed there, as this should be an 
arch-dependent feature, and not something that the poor user should be
aware if a given architecture supports it or not. Also, the above will keep
causing warnings or errors with randconfigs.

Also, the "depends on HAVE_CPI" is bogus, as PCI already depends on 
HAVE_PCI:  

Ah yes you're right.

quoted

	menuconfig PCI
	bool "PCI support"
	depends on HAVE_PCI
	help
	  This option enables support for the PCI local bus, including
	  support for PCI-X and the foundations for PCI Express support.
	  Say 'Y' here unless you know what you are doing.

So, instead, I would expect that a new HAVE_xxx option would be
added at arch/*/Kconfig, like:

	config X86
		...
		select HAVE_PCI_DIRECT_IO

It would also make sense to document it at Documentation/features/.  

I'll look into that, thanks.

quoted

quoted

All you really care about is the "legacy" I/O spaces here, this isn't
tied to PCI specifically at all, right?

So why not just have a OLD_STYLE_IO config option or something like
that, to show that it's the i/o functions we care about here, not PCI at
all?

And maybe not call it "old" or "legacy" as time constantly goes forward,
just describe it as it is, "DIRECT_IO"?  

Agreed. HAVE_PCI_DIRECT_IO (or something similar) seems a more appropriate
name for it.

Thanks,
Mauro  

Hmm, I might be missing something here but that sounds a lot like the
HAS_IOPORT option added in patch 02.

We add both LEGACY_PCI and HAS_IOPORT to differentiate between two
cases. HAS_IOPORT is for PC-style devices that are not on a PCI card
while LEGACY_PCI is for PCI drivers that require port I/O. 

I didn't look at the other patches on this series, but why it is needed
to deal with them on a separate way? Won't "PCI" and "HAS_IOPORT" be enough? 

I mean, are there any architecture where HAVE_PCI=y and HAS_IOPORT=y
where LEGACY_PCI shall be "n"?

This
includes pre-PCIe devices as well as PCIe devices which require
features like I/O spaces. The "legacy" naming is comes from the PCIe
spec which in section 2.1.1.2 says "PCI Express supports I/O Space for
compatibility with legacy devices which require their use. Future
revisions of this specification may deprecate the use of I/O Space."

I would still avoid calling it LEGACY_PCI, as this sounds too generic.

I didn't read the PCI/PCIe specs, but I suspect that are a lot more
features that were/will be deprecated on PCI specs as time goes by.

So, I would, instead, use something like PCI_LEGACY_IO_SPACE or 
HAVE_PCI_LEGACY_IO_SPACE, in order to let it clear what "legacy"
means.

These two separate config options allow us to compile without support
for these legacy PCI devices even on a system where inb()/outb() and
friends are required for some PC style devices and for example ACPI.

Hmm... why this patch make SND_BT87X dependent on LEGACY_PCI?

I couldn't find any usage of inb/outb & friends on it:

	$ grep -E '(inb|outb|inw|outw|inl|outl)\b' ./sound/pci/bt87x.c

It uses, instead, readl/writel:

	static inline u32 snd_bt87x_readl(struct snd_bt87x *chip, u32 reg)
	{
	        return readl(chip->mmio + reg);
	}

	static inline void snd_bt87x_writel(struct snd_bt87x *chip, u32 reg, u32 value)
	{
	        writel(value, chip->mmio + reg);
	}

I failed to see what makes it different from VIDEO_BT848 and
DVB_BT8XX drivers. They all support exactly the same chipset:
Brooktree/Conexant BT8xx. On those devices, depending on the exact
model, up to three separate interfaces are provided, each one with
its own Kconfig var:

	- audio I/O (SND_BT87X);
	- video I/O (VIDEO_BT848);
	- MPEG-TS I/O (DVB_BT8XX).

Thanks,
Mauro