Skip to contentSkip to conversation

ARM: MAX_DMA_ADDRESS fixes

PATCH In review Florian Fainelli 5 participants 11 messages Mar 24, 2022 – Apr 29, 2022 Arm (32-bit)

lore
Share & tools
lore
#0 lore raw

cover ARM: MAX_DMA_ADDRESS fixes

This patch series clamps the MAX_DMA_ADDRESS to 32-bit in order to fit within a virtual address, and also fixes the off by one which was suggested by Geert.

Florian Fainelli (2):
ARM: Fix off by one in checking DMA zone size
ARM: Clamp MAX_DMA_ADDRESS to 32-bit

 arch/arm/include/asm/dma.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
Signature · 1 line
2.25.1
#1 lore raw

PATCH 1/2 ARM: Fix off by one in checking DMA zone size

The maximum DMA address is PAGE_OFFSET + arm_dma_zone_size - 1, fix this by doing the appropriate subtraction.

Fixes 650320181a08 ("ARM: change ARM_DMA_ZONE_SIZE into a variable")
---
 arch/arm/include/asm/dma.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm/include/asm/dma.h +1 -1
index a81dda65c576..f244ee68e814 100644
@@ -11,7 +11,7 @@
 #define MAX_DMA_ADDRESS	({ \
 	extern phys_addr_t arm_dma_zone_size; \
 	arm_dma_zone_size && arm_dma_zone_size < (0x10000000 - PAGE_OFFSET) ? \
-		(PAGE_OFFSET + arm_dma_zone_size) : 0xffffffffUL; })
+		(PAGE_OFFSET + arm_dma_zone_size - 1) : 0xffffffffUL; })
 #endif
 
 #ifdef CONFIG_ISA_DMA_API
Signature · 1 line
2.25.1
Reviewed-by on v1 1/2
#2 lore raw

Re: [PATCH 1/2] ARM: Fix off by one in checking DMA zone size

Quoted 6 lines · Florian Fainelli

The maximum DMA address is PAGE_OFFSET + arm_dma_zone_size - 1, fix this by doing the appropriate subtraction.

Fixes 650320181a08 ("ARM: change ARM_DMA_ZONE_SIZE into a variable")

Yours,
Linus Walleij

#3 lore raw

Re: [PATCH 1/2] ARM: Fix off by one in checking DMA zone size

In reply to Florian Fainelli
Quoted 2 lines · Florian Fainelli

The maximum DMA address is PAGE_OFFSET + arm_dma_zone_size - 1, fix this by doing the appropriate subtraction.

The question is... is MAX_DMA_ADDRESS inclusive or exclusive. The answer is rather indeterminant, unfortunately.

drivers/net/wan/cosa.c: if (b + len >= MAX_DMA_ADDRESS)

So, if the buffer address + buffer length is equal to MAX_DMA_ADDRESS, then the buffer is not DMA-able. To me this looks completely wrong. It's also completely wrong because 'b' is a "unsigned long" which means on 32-bit, this can wrap.

drivers/parport/parport_pc.c:

	unsigned long start = (unsigned long) buf;
	unsigned long end = (unsigned long) buf + length - 1;

	if (end < MAX_DMA_ADDRESS) {

So, "end" is the last byte of the buffer. If MAX_DMA_ADDRESS is the last byte of the virtual address space that supports DMA, then if the two are equal, we do not use DMA. This seems wrong to me, and points to it being an _exclusive_ value - it's the last byte plus one.

there's a bunch of memblock allocation functions that use __pa(MAX_DMA_ADDRESS) as the "min_addr" and if this is a minimum address, then surely this means that it is also an exclusive value.

So, I came to the conclusion that MAX_DMA_ADDRESS is supposed to be exclusive. In other words, where PAGE_OFFSET + sizeof(ram) if sizeof(ram) is the size in bytes of the RAM, limited to the maximum addressable virtual address that RAM can be mapped to.

I don't see an inclusive value being correct, at least not for the cases I've looekd at.

Signature · 2 lines
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTP is here! 40Mbps down 10Mbps up. Decent connectivity at last!
#4 lore raw

Re: [PATCH 1/2] ARM: Fix off by one in checking DMA zone size

Quoted 8 lines · Russell King (Oracle)
Quoted 2 lines · Florian Fainelli

The maximum DMA address is PAGE_OFFSET + arm_dma_zone_size - 1, fix this by doing the appropriate subtraction.

The question is... is MAX_DMA_ADDRESS inclusive or exclusive. The answer is rather indeterminant, unfortunately.

Quoted 6 lines

drivers/net/wan/cosa.c: if (b + len >= MAX_DMA_ADDRESS)

So, if the buffer address + buffer length is equal to MAX_DMA_ADDRESS, then the buffer is not DMA-able. To me this looks completely wrong. It's also completely wrong because 'b' is a "unsigned long" which means on 32-bit, this can wrap.

This driver is on its way out and will be removed in 5.19.

Quoted 11 lines

there's a bunch of memblock allocation functions that use __pa(MAX_DMA_ADDRESS) as the "min_addr" and if this is a minimum address, then surely this means that it is also an exclusive value.

So, I came to the conclusion that MAX_DMA_ADDRESS is supposed to be exclusive. In other words, where PAGE_OFFSET + sizeof(ram) if sizeof(ram) is the size in bytes of the RAM, limited to the maximum addressable virtual address that RAM can be mapped to.

I don't see an inclusive value being correct, at least not for the cases I've looekd at.

Looking at the other definitions, I see it's usually either ULONG_MAX (which is inclusive) or an exclusive limit, same as the existing version on arm:

arch/alpha/include/asm/io.h:#define IDENT_ADDR 0xffff800000000000UL
arch/alpha/include/asm/dma.h:#define MAX_DMA_ADDRESS (alpha_mv.mv_pci_tbi ? \
arch/alpha/include/asm/dma.h- ~0UL :
IDENT_ADDR + 0x01000000)
arch/arc/include/asm/dma.h:#define MAX_DMA_ADDRESS 0xC0000000 arch/hexagon/include/asm/dma.h:#define MAX_DMA_ADDRESS (PAGE_OFFSET) arch/ia64/mm/init.c:unsigned long MAX_DMA_ADDRESS = PAGE_OFFSET + 0x100000000UL; arch/m68k/include/asm/dma.h:#define MAX_DMA_ADDRESS PAGE_OFFSET
arch/mips/include/asm/dma.h:#define MAX_DMA_ADDRESS PAGE_OFFSET
arch/mips/include/asm/dma.h:#define MAX_DMA_ADDRESS
(PAGE_OFFSET + 0x01000000)
arch/parisc/include/asm/dma.h:#define MAX_DMA_ADDRESS (~0UL)
arch/powerpc/include/asm/dma.h:#define MAX_DMA_ADDRESS (~0UL)
arch/s390/include/asm/dma.h:#define MAX_DMA_ADDRESS 0x80000000
arch/sparc/include/asm/dma.h:#define MAX_DMA_ADDRESS (~0UL) arch/um/include/asm/dma.h:#define MAX_DMA_ADDRESS (uml_physmem)
arch/x86/include/asm/dma.h:#define MAX_DMA_ADDRESS (PAGE_OFFSET +
0x1000000)
arch/x86/include/asm/dma.h:#define MAX_DMA_PFN ((16UL * 1024 * 1024)

Quoted 1 line
Quoted 1 line

PAGE_SHIFT)

arch/x86/include/asm/dma.h:#define MAX_DMA_ADDRESS ((unsigned long)__va(MAX_DMA_PFN << PAGE_SHIFT))

The exceptions are:
arch/xtensa/include/asm/kmem_layout.h:#define XCHAL_KIO_SIZE

         0x10000000

arch/xtensa/include/asm/dma.h:#define MAX_DMA_ADDRESS (PAGE_OFFSET + XCHAL_KIO_SIZE - 1) arch/microblaze/include/asm/dma.h:#define MAX_DMA_ADDRESS (CONFIG_KERNEL_START + memory_size - 1)

Arnd

#5 lore raw

PATCH 2/2 ARM: Clamp MAX_DMA_ADDRESS to 32-bit

In reply to Florian Fainelli

MAX_DMA_ADDRESS is a virtual address, therefore it needs to fit within a 32-bit unsigned quantity. Platforms defining a DMA zone size in their machine descriptor can easily overflow this quantity depending on the DMA zone size and/or the PAGE_OFFSET setting.

In most cases this is harmless, however in the case of a CONFIG_DEBUG_VIRTUAL enabled, __virt_addr_valid() will be unable to return that MAX_DMA_ADDRESS is valid because the value passed to that function is an unsigned long which has already overflowed.

Fixes e377cd8221eb ("ARM: 8640/1: Add support for CONFIG_DEBUG_VIRTUAL")
Fixes 2fb3ec5c9503 ("ARM: Replace platform definition of ISA_DMA_THRESHOLD/MAX_DMA_ADDRESS")
---
 arch/arm/include/asm/dma.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
arch/arm/include/asm/dma.h +3 -1
index f244ee68e814..ea47420babd4 100644
@@ -8,10 +8,12 @@
 #ifndef CONFIG_ZONE_DMA
 #define MAX_DMA_ADDRESS	0xffffffffUL
 #else
+#include <linux/minmax.h>
 #define MAX_DMA_ADDRESS	({ \
 	extern phys_addr_t arm_dma_zone_size; \
 	arm_dma_zone_size && arm_dma_zone_size < (0x10000000 - PAGE_OFFSET) ? \
-		(PAGE_OFFSET + arm_dma_zone_size - 1) : 0xffffffffUL; })
+		min_t(phys_addr_t, (PAGE_OFFSET + arm_dma_zone_size - 1), 0xffffffffUL) : \
+		0xffffffffUL; })
 #endif
 
 #ifdef CONFIG_ISA_DMA_API
Signature · 1 line
2.25.1
Reviewed-by on v1 2/2
#6 lore raw

Re: [PATCH 2/2] ARM: Clamp MAX_DMA_ADDRESS to 32-bit

Quoted 5 lines · Florian Fainelli

MAX_DMA_ADDRESS is a virtual address, therefore it needs to fit within a 32-bit unsigned quantity. Platforms defining a DMA zone size in their machine descriptor can easily overflow this quantity depending on the DMA zone size and/or the PAGE_OFFSET setting.

Quoted 8 lines

In most cases this is harmless, however in the case of a CONFIG_DEBUG_VIRTUAL enabled, __virt_addr_valid() will be unable to return that MAX_DMA_ADDRESS is valid because the value passed to that function is an unsigned long which has already overflowed.

Fixes e377cd8221eb ("ARM: 8640/1: Add support for CONFIG_DEBUG_VIRTUAL")
Fixes 2fb3ec5c9503 ("ARM: Replace platform definition of ISA_DMA_THRESHOLD/MAX_DMA_ADDRESS")

Yours,
Linus Walleij

#7 lore raw

Re: [PATCH 2/2] ARM: Clamp MAX_DMA_ADDRESS to 32-bit

In reply to Florian Fainelli

Hi Florian,

Quoted 17 lines · Florian Fainelli

MAX_DMA_ADDRESS is a virtual address, therefore it needs to fit within a 32-bit unsigned quantity. Platforms defining a DMA zone size in their machine descriptor can easily overflow this quantity depending on the DMA zone size and/or the PAGE_OFFSET setting.

In most cases this is harmless, however in the case of a CONFIG_DEBUG_VIRTUAL enabled, __virt_addr_valid() will be unable to return that MAX_DMA_ADDRESS is valid because the value passed to that function is an unsigned long which has already overflowed.

Fixes e377cd8221eb ("ARM: 8640/1: Add support for CONFIG_DEBUG_VIRTUAL")
Fixes 2fb3ec5c9503 ("ARM: Replace platform definition of ISA_DMA_THRESHOLD/MAX_DMA_ADDRESS")

arch/arm/include/asm/dma.h line 14

Quoted 7 lines
arch/arm/include/asm/dma.h +1 -0
@@ -8,10 +8,12 @@
 #ifndef CONFIG_ZONE_DMA
 #define MAX_DMA_ADDRESS        0xffffffffUL
 #else
+#include <linux/minmax.h>
 #define MAX_DMA_ADDRESS        ({ \
        extern phys_addr_t arm_dma_zone_size; \
        arm_dma_zone_size && arm_dma_zone_size < (0x10000000 - PAGE_OFFSET) ? \

"arm_dma_zone_size < (0x10000000 - PAGE_OFFSET)" looks like an overflow-avoiding version of "PAGE_OFFSET + arm_dma_zone_size < 0x10000000". However, PAGE_OFFSET is always larger than 0x10000000, so "0x10000000 - PAGE_OFFSET" is a rather large number?

arch/arm/include/asm/dma.h line 19

Quoted 6 lines
arch/arm/include/asm/dma.h +2 -1
-               (PAGE_OFFSET + arm_dma_zone_size - 1) : 0xffffffffUL; })
+               min_t(phys_addr_t, (PAGE_OFFSET + arm_dma_zone_size - 1), 0xffffffffUL) : \
+               0xffffffffUL; })
 #endif
 
 #ifdef CONFIG_ISA_DMA_API

Gr{oetje,eeting}s,

Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]

In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds

#8 lore raw

Re: [PATCH 2/2] ARM: Clamp MAX_DMA_ADDRESS to 32-bit

Quoted 17 lines · Geert Uytterhoeven

Hi Florian,

Quoted 13 lines · Florian Fainelli

MAX_DMA_ADDRESS is a virtual address, therefore it needs to fit within a 32-bit unsigned quantity. Platforms defining a DMA zone size in their machine descriptor can easily overflow this quantity depending on the DMA zone size and/or the PAGE_OFFSET setting.

In most cases this is harmless, however in the case of a CONFIG_DEBUG_VIRTUAL enabled, __virt_addr_valid() will be unable to return that MAX_DMA_ADDRESS is valid because the value passed to that function is an unsigned long which has already overflowed.

Fixes e377cd8221eb ("ARM: 8640/1: Add support for CONFIG_DEBUG_VIRTUAL")
Fixes 2fb3ec5c9503 ("ARM: Replace platform definition of ISA_DMA_THRESHOLD/MAX_DMA_ADDRESS")
Quoted 16 lines

arch/arm/include/asm/dma.h line 14

Quoted 10 lines
arch/arm/include/asm/dma.h +1 -0
@@ -8,10 +8,12 @@
  #ifndef CONFIG_ZONE_DMA
  #define MAX_DMA_ADDRESS        0xffffffffUL
  #else
+#include <linux/minmax.h>
  #define MAX_DMA_ADDRESS        ({ \
         extern phys_addr_t arm_dma_zone_size; \
         arm_dma_zone_size && arm_dma_zone_size < (0x10000000 - PAGE_OFFSET) ? \

"arm_dma_zone_size < (0x10000000 - PAGE_OFFSET)" looks like an overflow-avoiding version of "PAGE_OFFSET + arm_dma_zone_size < 0x10000000". However, PAGE_OFFSET is always larger than 0x10000000, so "0x10000000 - PAGE_OFFSET" is a rather large number?

Yes it is a large number, though I am not too sure what the intention of this check was in the first place, whether it denoted the largest known DMA size of any machine at the time, or if it has any relationship to lowmem (in which case it does not account for its exact value since it can be changed indirectly via vmalloc= on the command line).

We ought to be just fine with keeping only:

min_t(phys_addr_t, (PAGE_OFFSET + arm_dma_zone_size - 1), 0xffffffffUL);

Santosh, what did 0x10000000 mean when you added it?

Signature · 1 line
Florian
#9 lore raw

Re: [PATCH 2/2] ARM: Clamp MAX_DMA_ADDRESS to 32-bit

In reply to Geert Uytterhoeven
Quoted 17 lines · Geert Uytterhoeven

Hi Florian,

Quoted 13 lines · Florian Fainelli

MAX_DMA_ADDRESS is a virtual address, therefore it needs to fit within a 32-bit unsigned quantity. Platforms defining a DMA zone size in their machine descriptor can easily overflow this quantity depending on the DMA zone size and/or the PAGE_OFFSET setting.

In most cases this is harmless, however in the case of a CONFIG_DEBUG_VIRTUAL enabled, __virt_addr_valid() will be unable to return that MAX_DMA_ADDRESS is valid because the value passed to that function is an unsigned long which has already overflowed.

Fixes e377cd8221eb ("ARM: 8640/1: Add support for CONFIG_DEBUG_VIRTUAL")
Fixes 2fb3ec5c9503 ("ARM: Replace platform definition of ISA_DMA_THRESHOLD/MAX_DMA_ADDRESS")
Quoted 16 lines

arch/arm/include/asm/dma.h line 14

Quoted 10 lines
arch/arm/include/asm/dma.h +1 -0
@@ -8,10 +8,12 @@
 #ifndef CONFIG_ZONE_DMA
 #define MAX_DMA_ADDRESS        0xffffffffUL
 #else
+#include <linux/minmax.h>
 #define MAX_DMA_ADDRESS        ({ \
        extern phys_addr_t arm_dma_zone_size; \
        arm_dma_zone_size && arm_dma_zone_size < (0x10000000 - PAGE_OFFSET) ? \

"arm_dma_zone_size < (0x10000000 - PAGE_OFFSET)" looks like an overflow-avoiding version of "PAGE_OFFSET + arm_dma_zone_size < 0x10000000". However, PAGE_OFFSET is always larger than 0x10000000, so "0x10000000 - PAGE_OFFSET" is a rather large number?

This, to me, looks like it should have been:

	arm_dma_zone_size && arm_dma_zone_size < (0x100000000 - PAGE_OFFSET) ? \

since that is the only thing that would make sense - it's the virtual space remaining between PAGE_OFFSET and the top of addressable space.

However, 0x100000000 isn't correct - since the vectors live at 0xffff0000, and we have the vmalloc space etc.

Signature · 2 lines
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTP is here! 40Mbps down 10Mbps up. Decent connectivity at last!
#10 lore raw
In reply to Florian Fainelli
Quoted 7 lines · Florian Fainelli

This patch series clamps the MAX_DMA_ADDRESS to 32-bit in order to fit within a virtual address, and also fixes the off by one which was suggested by Geert.

Florian Fainelli (2):
ARM: Fix off by one in checking DMA zone size
ARM: Clamp MAX_DMA_ADDRESS to 32-bit

Russell, do you have any feedback on these two patches?

Signature · 1 line
Florian