cpu_install_idmap replaces ttbr0 to install idmap. replace_phys uses
 idmap_cpu_replace_ttbr1 which is located in idmap,
 thus it requires ttbr0 to point idmap. 
 
 However, if a kernel task has switched out right after
 cpu_install_idmap, when kernel task is loaded, its ttbr0 is
 mapped to reserved_pg_dir which doesn't contain any idmap
 functions.
 Therefore, the kernel task cannot find any functions to run,
 causing memory fault.
 
 As far as I can verify from the current arm64 tree,
 no in-tree caller invokes this function from a preemptible context.
 
 Nevertheless, future work might allow this function to be called in
 preemptive state. Encapsulating the idmap sequence with
 preempt_disable() and preempt_enable() prevents this possible problem.

Signed-off-by: Jeongin Yeo <leo.yeo@luaberry.com>
---
 arch/arm64/mm/mmu.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index 2ba01dc8ef82..c3da836ebe8d 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c
@@ -29,6 +29,7 @@
 #include <linux/mm_inline.h>
 #include <linux/pagewalk.h>
 #include <linux/stop_machine.h>
+#include <linux/preempt.h>
 
 #include <asm/barrier.h>
 #include <asm/cputype.h>
@@ -2136,7 +2137,7 @@ void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp)
 		ttbr1 |= TTBR_CNP_BIT;
 
 	replace_phys = (void *)__pa_symbol(idmap_cpu_replace_ttbr1);
-
+	preempt_disable();
 	cpu_install_idmap();
 
 	/*
@@ -2148,6 +2149,7 @@ void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp)
 	local_daif_restore(daif);
 
 	cpu_uninstall_idmap();
+	preempt_enable();
 }
 
 #ifdef CONFIG_ARCH_HAS_PKEYS
-- 
2.43.0

On Tue, Dec 02, 2025 at 12:42:23AM +0000, Jeongin Yeo wrote:
>  cpu_install_idmap replaces ttbr0 to install idmap. replace_phys uses
>  idmap_cpu_replace_ttbr1 which is located in idmap,
>  thus it requires ttbr0 to point idmap. 
>  
>  However, if a kernel task has switched out right after
>  cpu_install_idmap, when kernel task is loaded, its ttbr0 is
>  mapped to reserved_pg_dir which doesn't contain any idmap
>  functions.
>  Therefore, the kernel task cannot find any functions to run,
>  causing memory fault.
>  
>  As far as I can verify from the current arm64 tree,
>  no in-tree caller invokes this function from a preemptible context.
>  
>  Nevertheless, future work might allow this function to be called in
>  preemptive state. Encapsulating the idmap sequence with
>  preempt_disable() and preempt_enable() prevents this possible problem.
> 
> Signed-off-by: Jeongin Yeo <leo.yeo@luaberry.com>
> ---
>  arch/arm64/mm/mmu.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> index 2ba01dc8ef82..c3da836ebe8d 100644
> --- a/arch/arm64/mm/mmu.c
> +++ b/arch/arm64/mm/mmu.c
> @@ -29,6 +29,7 @@
>  #include <linux/mm_inline.h>
>  #include <linux/pagewalk.h>
>  #include <linux/stop_machine.h>
> +#include <linux/preempt.h>
>  
>  #include <asm/barrier.h>
>  #include <asm/cputype.h>
> @@ -2136,7 +2137,7 @@ void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp)
>  		ttbr1 |= TTBR_CNP_BIT;
>  
>  	replace_phys = (void *)__pa_symbol(idmap_cpu_replace_ttbr1);
> -
> +	preempt_disable();
>  	cpu_install_idmap();
>  
>  	/*
> @@ -2148,6 +2149,7 @@ void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp)
>  	local_daif_restore(daif);
>  
>  	cpu_uninstall_idmap();
> +	preempt_enable();
>  }

__cpu_replace_ttbr1() is a _very_ low-level helper with only a handful
of arch-specific callers and so we shouldn't be worrying about random
functions calling it with preemption enabled.

Will

On Tue, Dec 2, 2025 at 9:43 AM Jeongin Yeo <leo.yeo@luaberry.com> wrote:
>
>  cpu_install_idmap replaces ttbr0 to install idmap. replace_phys uses
>  idmap_cpu_replace_ttbr1 which is located in idmap,
>  thus it requires ttbr0 to point idmap.
>
>  However, if a kernel task has switched out right after
>  cpu_install_idmap, when kernel task is loaded, its ttbr0 is
>  mapped to reserved_pg_dir which doesn't contain any idmap
>  functions.
>  Therefore, the kernel task cannot find any functions to run,
>  causing memory fault.
>
>  As far as I can verify from the current arm64 tree,
>  no in-tree caller invokes this function from a preemptible context.
>
>  Nevertheless, future work might allow this function to be called in
>  preemptive state. Encapsulating the idmap sequence with
>  preempt_disable() and preempt_enable() prevents this possible problem.

Hi,

Gentle ping on this patch (sent on Dec 1).
Is there any feedback or concerns I should address?

Thanks.

>
> Signed-off-by: Jeongin Yeo <leo.yeo@luaberry.com>
> ---
>  arch/arm64/mm/mmu.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> index 2ba01dc8ef82..c3da836ebe8d 100644
> --- a/arch/arm64/mm/mmu.c
> +++ b/arch/arm64/mm/mmu.c
> @@ -29,6 +29,7 @@
>  #include <linux/mm_inline.h>
>  #include <linux/pagewalk.h>
>  #include <linux/stop_machine.h>
> +#include <linux/preempt.h>
>
>  #include <asm/barrier.h>
>  #include <asm/cputype.h>
> @@ -2136,7 +2137,7 @@ void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp)
>                 ttbr1 |= TTBR_CNP_BIT;
>
>         replace_phys = (void *)__pa_symbol(idmap_cpu_replace_ttbr1);
> -
> +       preempt_disable();
>         cpu_install_idmap();
>
>         /*
> @@ -2148,6 +2149,7 @@ void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp)
>         local_daif_restore(daif);
>
>         cpu_uninstall_idmap();
> +       preempt_enable();
>  }
>
>  #ifdef CONFIG_ARCH_HAS_PKEYS
> --
> 2.43.0
>