Skip to contentSkip to conversation

arm64: mm: Ensure phys_to_ttbr on pgdir for idmap_cpu_replace_ttbr1

PATCH In review Weikang Guo 2 participants 5 messages Jul 22, 2025 – Jul 23, 2025 arm64

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 arm64: mm: Ensure phys_to_ttbr on pgdir for idmap_cpu_replace_ttbr1

Commit 5ffdfaedfa0a ("arm64: mm: Support Common Not Private translations") changed the contract of idmap_cpu_replace_ttbr1, requiring that the TTBR argument passed in should already be processed by phys_to_ttbr (i.e., in TTBR format, not just a raw physical address).

However, the current map_kernel implementation does not always convert the pgdir/ttbr argument via phys_to_ttbr before calling idmap_cpu_replace_ttbr1. This can lead to issues on systems with CONFIG_ARM64_PA_BITS_52 enabled, as the TTBR would not be properly folded per the ARMv8.2+ requirements.

---
Note: I do not currently have access to ARM64 hardware or an emulation environment that supports 52-bit physical address (PA52). I would greatly appreciate if anyone with such a platform could help test this patch. Thank you!

---
 arch/arm64/kernel/pi/map_kernel.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
arch/arm64/kernel/pi/map_kernel.c +3 -3
index 0f4bd7771859..05a04eb91e81 100644
@@ -18,7 +18,7 @@
 
 extern const u8 __eh_frame_start[], __eh_frame_end[];
 
-extern void idmap_cpu_replace_ttbr1(void *pgdir);
+extern void idmap_cpu_replace_ttbr1(phys_addr_t);
 
 static void __init map_segment(pgd_t *pg_dir, u64 *pgd, u64 va_offset,
 			       void *start, void *end, pgprot_t prot,
@@ -90,7 +90,7 @@ static void __init map_kernel(u64 kaslr_offset, u64 va_offset, int root_level)
 		    true, root_level);
 	dsb(ishst);
 
-	idmap_cpu_replace_ttbr1(init_pg_dir);
+	idmap_cpu_replace_ttbr1(phys_to_ttbr((u64)init_pg_dir));
 
 	if (twopass) {
 		if (IS_ENABLED(CONFIG_RELOCATABLE))
@@ -129,7 +129,7 @@ static void __init map_kernel(u64 kaslr_offset, u64 va_offset, int root_level)
 	/* Copy the root page table to its final location */
 	memcpy((void *)swapper_pg_dir + va_offset, init_pg_dir, PAGE_SIZE);
 	dsb(ishst);
-	idmap_cpu_replace_ttbr1(swapper_pg_dir);
+	idmap_cpu_replace_ttbr1(phys_to_ttbr((u64)swapper_pg_dir));
 }
 
 static void noinline __section(".idmap.text") set_ttbr0_for_lpa2(u64 ttbr)
Signature · 1 line
2.25.1
reviewer
#1 lore raw
Quoted 10 lines · Weikang Guo

Commit 5ffdfaedfa0a ("arm64: mm: Support Common Not Private translations") changed the contract of idmap_cpu_replace_ttbr1, requiring that the TTBR argument passed in should already be processed by phys_to_ttbr (i.e., in TTBR format, not just a raw physical address).

However, the current map_kernel implementation does not always convert the pgdir/ttbr argument via phys_to_ttbr before calling idmap_cpu_replace_ttbr1. This can lead to issues on systems with CONFIG_ARM64_PA_BITS_52 enabled, as the TTBR would not be properly folded per the ARMv8.2+ requirements.

For the cases below I don't believe that this is actually a problem. Since commit:

453dfcee70c5c344 ("arm64: booting: Require placement within 48-bit addressable memory")

... we require that the kernel Image (including any trailing unallocated bytes accounted for in image_size) are below the 48-bit address limit, and so there should be no difference between the PA and TTBR format.

We could probably test and enforce that in the early boot code somehow, if we're not doing that already.

If we were going to change things to avoid accidents in future, I think it would be better to enforce this with the type system. e.g. we could have a ttbr_val type that's distinct from phys_addr_t. Even then, for the idmap code I think it's better to avoid the phys_to_ttbr() dance, since that has runtime patching.

Mark.

Quoted 47 lines

---
Note: I do not currently have access to ARM64 hardware or an emulation environment that supports 52-bit physical address (PA52). I would greatly appreciate if anyone with such a platform could help test this patch. Thank you!

---
 arch/arm64/kernel/pi/map_kernel.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
arch/arm64/kernel/pi/map_kernel.c +3 -3
index 0f4bd7771859..05a04eb91e81 100644
@@ -18,7 +18,7 @@
 
 extern const u8 __eh_frame_start[], __eh_frame_end[];
 
-extern void idmap_cpu_replace_ttbr1(void *pgdir);
+extern void idmap_cpu_replace_ttbr1(phys_addr_t);
 
 static void __init map_segment(pgd_t *pg_dir, u64 *pgd, u64 va_offset,
 			       void *start, void *end, pgprot_t prot,
@@ -90,7 +90,7 @@ static void __init map_kernel(u64 kaslr_offset, u64 va_offset, int root_level)
 		    true, root_level);
 	dsb(ishst);
 
-	idmap_cpu_replace_ttbr1(init_pg_dir);
+	idmap_cpu_replace_ttbr1(phys_to_ttbr((u64)init_pg_dir));
 
 	if (twopass) {
 		if (IS_ENABLED(CONFIG_RELOCATABLE))
@@ -129,7 +129,7 @@ static void __init map_kernel(u64 kaslr_offset, u64 va_offset, int root_level)
 	/* Copy the root page table to its final location */
 	memcpy((void *)swapper_pg_dir + va_offset, init_pg_dir, PAGE_SIZE);
 	dsb(ishst);
-	idmap_cpu_replace_ttbr1(swapper_pg_dir);
+	idmap_cpu_replace_ttbr1(phys_to_ttbr((u64)swapper_pg_dir));
 }
 
 static void noinline __section(".idmap.text") set_ttbr0_for_lpa2(u64 ttbr)
Signature · 1 line
2.25.1
#2 lore raw
Quoted 24 lines · Mark Rutland
Quoted 10 lines · Weikang Guo

Commit 5ffdfaedfa0a ("arm64: mm: Support Common Not Private translations") changed the contract of idmap_cpu_replace_ttbr1, requiring that the TTBR argument passed in should already be processed by phys_to_ttbr (i.e., in TTBR format, not just a raw physical address).

However, the current map_kernel implementation does not always convert the pgdir/ttbr argument via phys_to_ttbr before calling idmap_cpu_replace_ttbr1. This can lead to issues on systems with CONFIG_ARM64_PA_BITS_52 enabled, as the TTBR would not be properly folded per the ARMv8.2+ requirements.

For the cases below I don't believe that this is actually a problem. Since commit:

453dfcee70c5c344 ("arm64: booting: Require placement within 48-bit addressable memory")

... we require that the kernel Image (including any trailing unallocated bytes accounted for in image_size) are below the 48-bit address limit, and so there should be no difference between the PA and TTBR format.

We could probably test and enforce that in the early boot code somehow, if we're not doing that already.

Quoted 8 lines

If we were going to change things to avoid accidents in future, I think it would be better to enforce this with the type system. e.g. we could have a ttbr_val type that's distinct from phys_addr_t. Even then, for the idmap code I think it's better to avoid the phys_to_ttbr() dance, since that has runtime patching.

Mark.

Thank you for your detailed explanation.

As you mentioned, if we can guarantee that the kernel image is always within the 48-bit PA range,then there is indeed no real difference between the PA and TTBR formats in this context.
In that case, does it mean that the conversion of reserved_pg_dirhere is also redundant? (There may be other similar cases.)

If we already ensure the kernel is always mapped below 48 bits, it does seem safe to remove phys_to_ttbrhere as well.

.macro __idmap_cpu_set_reserved_ttbr1, tmp1, tmp2
adrp \tmp1, reserved_pg_dir
phys_to_ttbr \tmp2, \tmp1 // This might not be needed either?
offset_ttbr1 \tmp2, \tmp1
msr ttbr1_el1, \tmp2
isb
tlbi vmalle1
dsb nsh
isb
.endm

Thanks again for the clarification!

WeiKang

reviewer
#3 lore raw
Quoted 31 lines · Weikang Guo
Quoted 30 lines · Mark Rutland
Quoted 10 lines · Weikang Guo

Commit 5ffdfaedfa0a ("arm64: mm: Support Common Not Private translations") changed the contract of idmap_cpu_replace_ttbr1, requiring that the TTBR argument passed in should already be processed by phys_to_ttbr (i.e., in TTBR format, not just a raw physical address).

However, the current map_kernel implementation does not always convert the pgdir/ttbr argument via phys_to_ttbr before calling idmap_cpu_replace_ttbr1. This can lead to issues on systems with CONFIG_ARM64_PA_BITS_52 enabled, as the TTBR would not be properly folded per the ARMv8.2+ requirements.

For the cases below I don't believe that this is actually a problem. Since commit:

453dfcee70c5c344 ("arm64: booting: Require placement within 48-bit addressable memory")

... we require that the kernel Image (including any trailing unallocated bytes accounted for in image_size) are below the 48-bit address limit, and so there should be no difference between the PA and TTBR format.

We could probably test and enforce that in the early boot code somehow, if we're not doing that already.

If we were going to change things to avoid accidents in future, I think it would be better to enforce this with the type system. e.g. we could have a ttbr_val type that's distinct from phys_addr_t. Even then, for the idmap code I think it's better to avoid the phys_to_ttbr() dance, since that has runtime patching.

Quoted 8 lines
Quoted 2 lines

Mark.

Thank you for your detailed explanation.

As you mentioned, if we can guarantee that the kernel image is always within the 48-bit PA range,then there is indeed no real difference between the PA and TTBR formats in this context.

Yep.

To be clear, I'm saying that there's no functional problem in practice, and hence the description in the commit message is more alarming than necessary.

Since the conversion is trivial I'm not against applying the conversion consistently, but if we do that I think we should enforce that through the type system so that missing conversions will be identified by the compiler.

Quoted 2 lines

In that case, does it mean that the conversion of reserved_pg_dirhere is also redundant? (There may be other similar cases.)

Yes, 'reserved_pg_dir' should be part of the kernel Image and hence below the 48-bit limit.

Quoted 2 lines

If we already ensure the kernel is always mapped below 48 bits, it does seem safe to remove phys_to_ttbrhere as well.

I assume for both instances of 'here' above you're referring to the macro below.

Quoted 10 lines

.macro __idmap_cpu_set_reserved_ttbr1, tmp1, tmp2
adrp \tmp1, reserved_pg_dir
phys_to_ttbr \tmp2, \tmp1 // This might not be needed either?
offset_ttbr1 \tmp2, \tmp1
msr ttbr1_el1, \tmp2
isb
tlbi vmalle1
dsb nsh
isb
.endm

Yes, the phys_to_ttbr conversion above isn't strictly necessary today.

Mark.

#4 lore raw
Quoted 41 lines · Mark Rutland
Quoted 31 lines · Weikang Guo
Quoted 30 lines · Mark Rutland
Quoted 10 lines · Weikang Guo

Commit 5ffdfaedfa0a ("arm64: mm: Support Common Not Private translations") changed the contract of idmap_cpu_replace_ttbr1, requiring that the TTBR argument passed in should already be processed by phys_to_ttbr (i.e., in TTBR format, not just a raw physical address).

However, the current map_kernel implementation does not always convert the pgdir/ttbr argument via phys_to_ttbr before calling idmap_cpu_replace_ttbr1. This can lead to issues on systems with CONFIG_ARM64_PA_BITS_52 enabled, as the TTBR would not be properly folded per the ARMv8.2+ requirements.

For the cases below I don't believe that this is actually a problem. Since commit:

453dfcee70c5c344 ("arm64: booting: Require placement within 48-bit addressable memory")

... we require that the kernel Image (including any trailing unallocated bytes accounted for in image_size) are below the 48-bit address limit, and so there should be no difference between the PA and TTBR format.

We could probably test and enforce that in the early boot code somehow, if we're not doing that already.

If we were going to change things to avoid accidents in future, I think it would be better to enforce this with the type system. e.g. we could have a ttbr_val type that's distinct from phys_addr_t. Even then, for the idmap code I think it's better to avoid the phys_to_ttbr() dance, since that has runtime patching.

Quoted 8 lines
Quoted 2 lines

Mark.

Thank you for your detailed explanation.

As you mentioned, if we can guarantee that the kernel image is always within the 48-bit PA range,then there is indeed no real difference between the PA and TTBR formats in this context.

Quoted 5 lines

Yep.

To be clear, I'm saying that there's no functional problem in practice, and hence the description in the commit message is more alarming than necessary.

Okay, I understand what you mean! and this is clear if the image is always guaranteed to be within the 48-bit address range, phys_to_ttbr will not change anything.

Quoted 5 lines

Since the conversion is trivial I'm not against applying the conversion consistently, but if we do that I think we should enforce that through the type system so that missing conversions will be identified by the compiler.

For a function call it should work.

Quoted 4 lines
Quoted 3 lines

In that case, does it mean that the conversion of reserved_pg_dirhere is also redundant? (There may be other similar cases.)

Quoted 8 lines

Yes, 'reserved_pg_dir' should be part of the kernel Image and hence below the 48-bit limit.

Quoted 2 lines

If we already ensure the kernel is always mapped below 48 bits, it does seem safe to remove phys_to_ttbrhere as well.

I assume for both instances of 'here' above you're referring to the macro below.

Yes, I meant the __idmap_cpu_set_reserved_ttbr1 macro.

Quoted 12 lines
Quoted 11 lines

.macro __idmap_cpu_set_reserved_ttbr1, tmp1, tmp2
adrp \tmp1, reserved_pg_dir
phys_to_ttbr \tmp2, \tmp1 // This might not be needed either?
offset_ttbr1 \tmp2, \tmp1
msr ttbr1_el1, \tmp2
isb
tlbi vmalle1
dsb nsh
isb
.endm

Quoted 1 line

Yes, the phys_to_ttbr conversion above isn't strictly necessary today.

Thanks for confirming this!

Quoted 2 lines

Mark.

Mark, finally, would you prefer that we introduce a new type to ensure we always get a properly converted TTBR, or keep things as they are, or perhaps explicitly state that we do not support kernel images above the 48-bit range and remove those unnecessary conversions?

Weikang