ARM Realview PCIX preparation

This patch adds the memory map PCI support to the include files for the Realview Northbridge based boards.

Signed-off-by Colin Tuckley <[email protected]>

Acked-by Catalin Marinas <[email protected]>

arch/arm/mach-realview/include/mach/hardware.h +2 -0

index 8a638d1..dcc64f2 100644
+/*
+ * PCI space physical addresses and sizes

Quoted 8 lines

arch/arm/mach-realview/include/mach/io.h +0 -0

index f05bcdf..aa97d18 100644
@@ -20,9 +20,25 @@
 #ifndef __ASM_ARM_ARCH_IO_H
 #define __ASM_ARM_ARCH_IO_H
 

arch/arm/mach-realview/include/mach/io.h line 26

Quoted 4 lines

arch/arm/mach-realview/include/mach/io.h +3 -0

+#include <asm/mach-types.h>
+#include <mach/hardware.h>
+
 #define IO_SPACE_LIMIT 0xffffffff

The IO_SPACE_LIMIT should match the REALVIEW_PB_PCI_IO_BASE above.

Quoted 5 lines · Colin Tuckley

This patch adds the memory map PCI support to the include files for the Realview Northbridge based boards.

Signed-off-by Colin Tuckley <[email protected]>

Acked-by Catalin Marinas <[email protected]>

Hi Colin,

I think we've discussed this before. I think it's good to have PCI support enabled in realview, but please do the PIO window properly.

Quoted 6 lines

arch/arm/mach-realview/include/mach/hardware.h +2 -0

index 8a638d1..dcc64f2 100644
+/*
+ * PCI space physical addresses and sizes

arch/arm/mach-realview/include/mach/hardware.h line 46

Quoted 10 lines

arch/arm/mach-realview/include/mach/hardware.h +10 -0

+ */
+#define REALVIEW_PB_PCI_BASE		0x90040000	/* PCI-X Unit base */
+#define REALVIEW_PB_PCI_BASE_SIZE	0x00010000	/* 4 Kb + 60Kb reserved */
+#define REALVIEW_PB_PCI_IO_BASE		0x90050000	/* IO Region on AHB */
+#define REALVIEW_PB_PCI_IO_SIZE		0x00010000	/* 64 Kb */
+#define REALVIEW_PB_PCI_IO_LIMIT       (REALVIEW_PB_PCI_IO_BASE + REALVIEW_PB_PCI_IO_SIZE - 1)
+#define REALVIEW_PB_PCI_MEM_BASE	0xA0000000	/* MEM Region on AHB */
+#define REALVIEW_PB_PCI_MEM_SIZE	0x20000000	/* 512 MB */
+
+#define REALVIEW_ISSP_REG_BASE		0x100E3000

As you write here, the I/O window size is 64KB, which is totally normal for PCI buses.

arch/arm/mach-realview/include/mach/hardware.h line 57

Quoted 10 lines

arch/arm/mach-realview/include/mach/hardware.h +10 -0

+#ifdef CONFIG_PCI
+#if !defined(__ASSEMBLY__)
+static inline unsigned int pcibios_min_io(void)
+{
+	if (machine_is_realview_pb11mp() || machine_is_realview_pba8() ||
+	    machine_is_realview_pbx())
+		return REALVIEW_PB_PCI_IO_BASE;
+	else
+		return 0;
+}

pcibios_min_io should be 0x1000 as a constant, in order to get the ISA addresses out of the way, but there is no need to make it board dependent.

arch/arm/mach-realview/include/mach/hardware.h line 67

Quoted 9 lines

arch/arm/mach-realview/include/mach/hardware.h +9 -0

+static inline unsigned int pcibios_min_mem(void)
+{
+	if (machine_is_realview_pb11mp() || machine_is_realview_pba8() ||
+	    machine_is_realview_pbx())
+		return REALVIEW_PB_PCI_MEM_BASE;
+	else
+		return 0;
+}
+#endif

Just hardcode this to REALVIEW_PB_PCI_MEM_BASE

arch/arm/mach-realview/include/mach/io.h line 26

Quoted 12 lines

arch/arm/mach-realview/include/mach/io.h +3 -0

index f05bcdf..aa97d18 100644
@@ -20,9 +20,25 @@
 #ifndef __ASM_ARM_ARCH_IO_H
 #define __ASM_ARM_ARCH_IO_H
 
+#include <asm/mach-types.h>
+#include <mach/hardware.h>
+
 #define IO_SPACE_LIMIT 0xffffffff

The IO_SPACE_LIMIT should match the REALVIEW_PB_PCI_IO_BASE above.

arch/arm/mach-realview/include/mach/io.h line 39

Quoted 13 lines

arch/arm/mach-realview/include/mach/io.h +12 -1

-#define __io(a)		__typesafe_io(a)
+static inline void __iomem *__io(unsigned long addr)
+{
+#ifdef CONFIG_PCI
+	/* check for PCI I/O space */
+	if (addr >= REALVIEW_PB_PCI_IO_BASE && addr <= REALVIEW_PB_PCI_IO_LIMIT)
+		return (void __iomem *)((addr - REALVIEW_PB_PCI_IO_BASE) + REALVIEW_PCI_IO_VBASE);
+	else
+		return (void __iomem *)addr;
+#else
+	return (void __iomem *)addr;
+#endif
+}

And there is no need to do these tricks when you only have a single PCI bus. Just define the PIO window to be based on zero and 64K in size, and hardcode that, to end up with

#define __io(a) (REALVIEW_PCI_IO_VBASE + (a & IO_SPACE_LIMIT))

Arnd

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel at lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

-----Original Message-----
from Russell King - ARM Linux
Subject: Re: [PATCH 1/3 v3] ARM Realview PCIX map include file changes

Please can you answer Arnd's concerns.

-----Original Message-----
From: Arnd Bergmann [mailto:arnd at arndb.de]
Subject: Re: [PATCH 1/3 v3] ARM Realview PCIX map include file changes

I think we've discussed this before. I think it's good to have PCI support enabled in realview, but please do the PIO window properly.

Review

Quoted 1 line

(quoted diff) +0 -0

b/arch/arm/mach-realview/include/mach/hardware.h

As you write here, the I/O window size is 64KB, which is totally normal for PCI buses.

arch/arm/mach-realview/include/mach/hardware.h line 57

Quoted 10 lines

arch/arm/mach-realview/include/mach/hardware.h +10 -0

+#ifdef CONFIG_PCI
+#if !defined(__ASSEMBLY__)
+static inline unsigned int pcibios_min_io(void)
+{
+   if (machine_is_realview_pb11mp() || machine_is_realview_pba8() ||
+       machine_is_realview_pbx())
+           return REALVIEW_PB_PCI_IO_BASE;
+   else
+           return 0;
+}

pcibios_min_io should be 0x1000 as a constant, in order to get the ISA addresses out of the way, but there is no need to make it board dependent.

arch/arm/mach-realview/include/mach/hardware.h line 67

Quoted 10 lines

arch/arm/mach-realview/include/mach/hardware.h +9 -0

+static inline unsigned int pcibios_min_mem(void)
+{
+   if (machine_is_realview_pb11mp() || machine_is_realview_pba8() ||
+       machine_is_realview_pbx())
+           return REALVIEW_PB_PCI_MEM_BASE;
+   else
+           return 0;
+}
+#endif

Just hardcode this to REALVIEW_PB_PCI_MEM_BASE

Quoted 1 line

#define IO_SPACE_LIMIT 0xffffffff

The IO_SPACE_LIMIT should match the REALVIEW_PB_PCI_IO_SIZE above.

arch/arm/mach-realview/include/mach/io.h

Quoted 6 lines

arch/arm/mach-realview/include/mach/io.h +5 -1

-#define __io(a)            __typesafe_io(a)
+static inline void __iomem *__io(unsigned long addr)
+{
+#ifdef CONFIG_PCI
+   /* check for PCI I/O space */
+   if (addr >= REALVIEW_PB_PCI_IO_BASE && addr <=

REALVIEW_PB_PCI_IO_LIMIT)

Quoted 1 line

+ return (void __iomem *)((addr - REALVIEW_PB_PCI_IO_BASE) +

REALVIEW_PCI_IO_VBASE);

arch/arm/mach-realview/include/mach/io.h line 39

Quoted 6 lines

arch/arm/mach-realview/include/mach/io.h +6 -0

+   else
+           return (void __iomem *)addr;
+#else
+   return (void __iomem *)addr;
+#endif
+}

And there is no need to do these tricks when you only have a single PCI bus. Just define the PIO window to be based on zero and 64K in size, and hardcode that, to end up with

#define __io(a) (REALVIEW_PCI_IO_VBASE + (a & IO_SPACE_LIMIT))

Quoted 3 lines

pcibios_min_io should be 0x1000 as a constant, in order to get the ISA addresses out of the way, but there is no need to make it board dependent.

REALVIEW_PB_PCI_IO_BASE is currently 0x90050000, which works, so why would I change pcibios_min_io to 0x1000 ?

-----Original Message-----
From: Russell King - ARM Linux [mailto:linux at arm.linux.org.uk]
Subject: Re: [PATCH 1/3 v3] ARM Realview PCIX map include file changes

All the points in this email all go together.

This is how it works:

1. The PCI IO window is _supposed_ to be 64K in size.

2. "pcibios_min_io" sets the minimum offset into the PCI IO window which
PCI IO BARs should be assigned. It is assumed that a PCI IO BAR
value
of 0 corresponds with the virtual base address of this window.
3. inb() et.al. take the PCI IO offset and _not_ the physical address nor the virtual address of the desired access.

Ah, that wasn't clear from the original email - I was making the changes incrementally and testing as I went.

Quoted 3 lines

This is how it works:

1. The PCI IO window is supposed to be 64K in size.

Yes, that bit was obvious after a bit of thought. I suspect the original value was just a cut 'n paste error.

Quoted 7 lines

2. "pcibios_min_io" sets the minimum offset into the PCI IO window which
PCI IO BARs should be assigned. It is assumed that a PCI IO BAR
value
of 0 corresponds with the virtual base address of this window.
3. inb() et.al. take the PCI IO offset and not the physical address nor the virtual address of the desired access.

That seems sensible.

However, after changing both __io() and pcibios_min_io as Arnd suggested the boot still hangs after " Uncompressing Linux... done, booting the kernel."

I did notice that there seems to have been some code changes and refactoring in the pci sub-system between 3.1 and 2.6.38 where I was testing before. Are there any significant changes I should be aware of?

Quoted 5 lines

-----Original Message-----
From: Russell King - ARM Linux [mailto:linux at arm.linux.org.uk]
Subject: Re: [PATCH 1/3 v3] ARM Realview PCIX map include file changes

All the points in this email all go together.

Ah, that wasn't clear from the original email - I was making the changes incrementally and testing as I went.

Quoted 3 lines

This is how it works:

1. The PCI IO window is _supposed_ to be 64K in size.

Yes, that bit was obvious after a bit of thought. I suspect the original value was just a cut 'n paste error.

Quoted 7 lines

2. "pcibios_min_io" sets the minimum offset into the PCI IO window which
? ?PCI IO BARs should be assigned. ?It is assumed that a PCI IO BAR value
? ?of 0 corresponds with the virtual base address of this window.
3. inb() et.al. take the PCI IO offset and _not_ the physical address nor the virtual address of the desired access.

That seems sensible.

However, after changing both __io() and pcibios_min_io as Arnd suggested the boot still hangs after " Uncompressing Linux... done, booting the kernel."

Quoted 1 line

However, after changing both __io() and pcibios_min_io as Arnd suggested the boot still hangs after " Uncompressing Linux... done, booting the kernel."

I haven't followed the patch series in detail but the __io() is also used by __io_address() in the RealView code and the former is just expected to do some type casting. By doing some arithmetic on the pointer, the static mappings for various peripherals (GIC etc.) are wrong.

So if you convert __io() as per Arnd's suggestion, make sure that __io_address() is redefined, something like below:

#define __io_address(n) __typesafe_io(IO_ADDRESS(n))

Alternatively we could leave __io() as in the original kernel and define RealView specific in*/out* macros.