From: Mike Rapoport rppt@linux.ibm.com
The semantics of pfn_valid() is to check presence of the memory map for a PFN and not whether a PFN is covered by the linear map. The memory map may be present for NOMAP memory regions, but they won't be mapped in the linear mapping. Accessing such regions via __va() when they are memremap()'ed will cause a crash.
On v5.4.y the crash happens on qemu-arm with UEFI [1]:
<1>[ 0.084476] 8<--- cut here --- <1>[ 0.084595] Unable to handle kernel paging request at virtual address dfb76000 <1>[ 0.084938] pgd = (ptrval) <1>[ 0.085038] [dfb76000] *pgd=5f7fe801, *pte=00000000, *ppte=00000000
...
<4>[ 0.093923] [<c0ed6ce8>] (memcpy) from [<c16a06f8>] (dmi_setup+0x60/0x418) <4>[ 0.094204] [<c16a06f8>] (dmi_setup) from [<c16a38d4>] (arm_dmi_init+0x8/0x10) <4>[ 0.094408] [<c16a38d4>] (arm_dmi_init) from [<c0302e9c>] (do_one_initcall+0x50/0x228) <4>[ 0.094619] [<c0302e9c>] (do_one_initcall) from [<c16011e4>] (kernel_init_freeable+0x15c/0x1f8) <4>[ 0.094841] [<c16011e4>] (kernel_init_freeable) from [<c0f028cc>] (kernel_init+0x8/0x10c) <4>[ 0.095057] [<c0f028cc>] (kernel_init) from [<c03010e8>] (ret_from_fork+0x14/0x2c)
On kernels v5.10.y and newer the same crash won't reproduce on ARM because commit b10d6bca8720 ("arch, drivers: replace for_each_membock() with for_each_mem_range()") changed the way memory regions are registered in the resource tree, but that merely covers up the problem.
On ARM64 memory resources registered in yet another way and there the issue of wrong usage of pfn_valid() to ensure availability of the linear map is also covered.
Implement arch_memremap_can_ram_remap() on ARM and ARM64 to prevent access to NOMAP regions via the linear mapping in memremap().
Link: https://lore.kernel.org/all/Yl65zxGgFzF1Okac@sirena.org.uk Reported-by: "kernelci.org bot" bot@kernelci.org Tested-by: Mark Brown broonie@kernel.org Cc: stable@vger.kernel.org # 5.4+ Signed-off-by: Mike Rapoport rppt@linux.ibm.com --- v2: don't remove pfn_valid() from try_ram_remap(), per Ard
arch/arm/include/asm/io.h | 3 +++ arch/arm/mm/ioremap.c | 8 ++++++++ arch/arm64/include/asm/io.h | 4 ++++ arch/arm64/mm/ioremap.c | 8 ++++++++ 4 files changed, 23 insertions(+)
diff --git a/arch/arm/include/asm/io.h b/arch/arm/include/asm/io.h index 0c70eb688a00..2a0739a2350b 100644 --- a/arch/arm/include/asm/io.h +++ b/arch/arm/include/asm/io.h @@ -440,6 +440,9 @@ extern void pci_iounmap(struct pci_dev *dev, void __iomem *addr); #define ARCH_HAS_VALID_PHYS_ADDR_RANGE extern int valid_phys_addr_range(phys_addr_t addr, size_t size); extern int valid_mmap_phys_addr_range(unsigned long pfn, size_t size); +extern bool arch_memremap_can_ram_remap(resource_size_t offset, size_t size, + unsigned long flags); +#define arch_memremap_can_ram_remap arch_memremap_can_ram_remap #endif
/* diff --git a/arch/arm/mm/ioremap.c b/arch/arm/mm/ioremap.c index aa08bcb72db9..290702328a33 100644 --- a/arch/arm/mm/ioremap.c +++ b/arch/arm/mm/ioremap.c @@ -493,3 +493,11 @@ void __init early_ioremap_init(void) { early_ioremap_setup(); } + +bool arch_memremap_can_ram_remap(resource_size_t offset, size_t size, + unsigned long flags) +{ + unsigned long pfn = PHYS_PFN(offset); + + return memblock_is_map_memory(pfn); +} diff --git a/arch/arm64/include/asm/io.h b/arch/arm64/include/asm/io.h index 7fd836bea7eb..3995652daf81 100644 --- a/arch/arm64/include/asm/io.h +++ b/arch/arm64/include/asm/io.h @@ -192,4 +192,8 @@ extern void __iomem *ioremap_cache(phys_addr_t phys_addr, size_t size); extern int valid_phys_addr_range(phys_addr_t addr, size_t size); extern int valid_mmap_phys_addr_range(unsigned long pfn, size_t size);
+extern bool arch_memremap_can_ram_remap(resource_size_t offset, size_t size, + unsigned long flags); +#define arch_memremap_can_ram_remap arch_memremap_can_ram_remap + #endif /* __ASM_IO_H */ diff --git a/arch/arm64/mm/ioremap.c b/arch/arm64/mm/ioremap.c index b7c81dacabf0..b21f91cd830d 100644 --- a/arch/arm64/mm/ioremap.c +++ b/arch/arm64/mm/ioremap.c @@ -99,3 +99,11 @@ void __init early_ioremap_init(void) { early_ioremap_setup(); } + +bool arch_memremap_can_ram_remap(resource_size_t offset, size_t size, + unsigned long flags) +{ + unsigned long pfn = PHYS_PFN(offset); + + return pfn_is_map_memory(pfn); +}
base-commit: b2d229d4ddb17db541098b83524d901257e93845
On Tue, 26 Apr 2022 at 08:01, Mike Rapoport rppt@kernel.org wrote:
From: Mike Rapoport rppt@linux.ibm.com
The semantics of pfn_valid() is to check presence of the memory map for a PFN and not whether a PFN is covered by the linear map. The memory map may be present for NOMAP memory regions, but they won't be mapped in the linear mapping. Accessing such regions via __va() when they are memremap()'ed will cause a crash.
On v5.4.y the crash happens on qemu-arm with UEFI [1]:
<1>[ 0.084476] 8<--- cut here --- <1>[ 0.084595] Unable to handle kernel paging request at virtual address dfb76000 <1>[ 0.084938] pgd = (ptrval) <1>[ 0.085038] [dfb76000] *pgd=5f7fe801, *pte=00000000, *ppte=00000000
...
<4>[ 0.093923] [<c0ed6ce8>] (memcpy) from [<c16a06f8>] (dmi_setup+0x60/0x418) <4>[ 0.094204] [<c16a06f8>] (dmi_setup) from [<c16a38d4>] (arm_dmi_init+0x8/0x10) <4>[ 0.094408] [<c16a38d4>] (arm_dmi_init) from [<c0302e9c>] (do_one_initcall+0x50/0x228) <4>[ 0.094619] [<c0302e9c>] (do_one_initcall) from [<c16011e4>] (kernel_init_freeable+0x15c/0x1f8) <4>[ 0.094841] [<c16011e4>] (kernel_init_freeable) from [<c0f028cc>] (kernel_init+0x8/0x10c) <4>[ 0.095057] [<c0f028cc>] (kernel_init) from [<c03010e8>] (ret_from_fork+0x14/0x2c)
On kernels v5.10.y and newer the same crash won't reproduce on ARM because commit b10d6bca8720 ("arch, drivers: replace for_each_membock() with for_each_mem_range()") changed the way memory regions are registered in the resource tree, but that merely covers up the problem.
On ARM64 memory resources registered in yet another way and there the issue of wrong usage of pfn_valid() to ensure availability of the linear map is also covered.
Implement arch_memremap_can_ram_remap() on ARM and ARM64 to prevent access to NOMAP regions via the linear mapping in memremap().
Link: https://lore.kernel.org/all/Yl65zxGgFzF1Okac@sirena.org.uk Reported-by: "kernelci.org bot" bot@kernelci.org Tested-by: Mark Brown broonie@kernel.org Cc: stable@vger.kernel.org # 5.4+ Signed-off-by: Mike Rapoport rppt@linux.ibm.com
v2: don't remove pfn_valid() from try_ram_remap(), per Ard
arch/arm/include/asm/io.h | 3 +++ arch/arm/mm/ioremap.c | 8 ++++++++ arch/arm64/include/asm/io.h | 4 ++++ arch/arm64/mm/ioremap.c | 8 ++++++++ 4 files changed, 23 insertions(+)
I think this looks reasonable, but you'll need to split it in two if it is going through the respective arch trees.
Reviewed-by: Ard Biesheuvel ardb@kernel.org
diff --git a/arch/arm/include/asm/io.h b/arch/arm/include/asm/io.h index 0c70eb688a00..2a0739a2350b 100644 --- a/arch/arm/include/asm/io.h +++ b/arch/arm/include/asm/io.h @@ -440,6 +440,9 @@ extern void pci_iounmap(struct pci_dev *dev, void __iomem *addr); #define ARCH_HAS_VALID_PHYS_ADDR_RANGE extern int valid_phys_addr_range(phys_addr_t addr, size_t size); extern int valid_mmap_phys_addr_range(unsigned long pfn, size_t size); +extern bool arch_memremap_can_ram_remap(resource_size_t offset, size_t size,
unsigned long flags);+#define arch_memremap_can_ram_remap arch_memremap_can_ram_remap #endif
/* diff --git a/arch/arm/mm/ioremap.c b/arch/arm/mm/ioremap.c index aa08bcb72db9..290702328a33 100644 --- a/arch/arm/mm/ioremap.c +++ b/arch/arm/mm/ioremap.c @@ -493,3 +493,11 @@ void __init early_ioremap_init(void) { early_ioremap_setup(); }
+bool arch_memremap_can_ram_remap(resource_size_t offset, size_t size,
unsigned long flags)+{
unsigned long pfn = PHYS_PFN(offset);return memblock_is_map_memory(pfn);+} diff --git a/arch/arm64/include/asm/io.h b/arch/arm64/include/asm/io.h index 7fd836bea7eb..3995652daf81 100644 --- a/arch/arm64/include/asm/io.h +++ b/arch/arm64/include/asm/io.h @@ -192,4 +192,8 @@ extern void __iomem *ioremap_cache(phys_addr_t phys_addr, size_t size); extern int valid_phys_addr_range(phys_addr_t addr, size_t size); extern int valid_mmap_phys_addr_range(unsigned long pfn, size_t size);
+extern bool arch_memremap_can_ram_remap(resource_size_t offset, size_t size,
unsigned long flags);+#define arch_memremap_can_ram_remap arch_memremap_can_ram_remap
#endif /* __ASM_IO_H */ diff --git a/arch/arm64/mm/ioremap.c b/arch/arm64/mm/ioremap.c index b7c81dacabf0..b21f91cd830d 100644 --- a/arch/arm64/mm/ioremap.c +++ b/arch/arm64/mm/ioremap.c @@ -99,3 +99,11 @@ void __init early_ioremap_init(void) { early_ioremap_setup(); }
+bool arch_memremap_can_ram_remap(resource_size_t offset, size_t size,
unsigned long flags)+{
unsigned long pfn = PHYS_PFN(offset);return pfn_is_map_memory(pfn);+}
base-commit: b2d229d4ddb17db541098b83524d901257e93845
2.28.0
On Fri, Apr 29, 2022 at 03:38:59PM +0200, Ard Biesheuvel wrote:
On Tue, 26 Apr 2022 at 08:01, Mike Rapoport rppt@kernel.org wrote:
From: Mike Rapoport rppt@linux.ibm.com
The semantics of pfn_valid() is to check presence of the memory map for a PFN and not whether a PFN is covered by the linear map. The memory map may be present for NOMAP memory regions, but they won't be mapped in the linear mapping. Accessing such regions via __va() when they are memremap()'ed will cause a crash.
On v5.4.y the crash happens on qemu-arm with UEFI [1]:
<1>[ 0.084476] 8<--- cut here --- <1>[ 0.084595] Unable to handle kernel paging request at virtual address dfb76000 <1>[ 0.084938] pgd = (ptrval) <1>[ 0.085038] [dfb76000] *pgd=5f7fe801, *pte=00000000, *ppte=00000000
...
<4>[ 0.093923] [<c0ed6ce8>] (memcpy) from [<c16a06f8>] (dmi_setup+0x60/0x418) <4>[ 0.094204] [<c16a06f8>] (dmi_setup) from [<c16a38d4>] (arm_dmi_init+0x8/0x10) <4>[ 0.094408] [<c16a38d4>] (arm_dmi_init) from [<c0302e9c>] (do_one_initcall+0x50/0x228) <4>[ 0.094619] [<c0302e9c>] (do_one_initcall) from [<c16011e4>] (kernel_init_freeable+0x15c/0x1f8) <4>[ 0.094841] [<c16011e4>] (kernel_init_freeable) from [<c0f028cc>] (kernel_init+0x8/0x10c) <4>[ 0.095057] [<c0f028cc>] (kernel_init) from [<c03010e8>] (ret_from_fork+0x14/0x2c)
On kernels v5.10.y and newer the same crash won't reproduce on ARM because commit b10d6bca8720 ("arch, drivers: replace for_each_membock() with for_each_mem_range()") changed the way memory regions are registered in the resource tree, but that merely covers up the problem.
On ARM64 memory resources registered in yet another way and there the issue of wrong usage of pfn_valid() to ensure availability of the linear map is also covered.
Implement arch_memremap_can_ram_remap() on ARM and ARM64 to prevent access to NOMAP regions via the linear mapping in memremap().
Link: https://lore.kernel.org/all/Yl65zxGgFzF1Okac@sirena.org.uk Reported-by: "kernelci.org bot" bot@kernelci.org Tested-by: Mark Brown broonie@kernel.org Cc: stable@vger.kernel.org # 5.4+ Signed-off-by: Mike Rapoport rppt@linux.ibm.com
v2: don't remove pfn_valid() from try_ram_remap(), per Ard
arch/arm/include/asm/io.h | 3 +++ arch/arm/mm/ioremap.c | 8 ++++++++ arch/arm64/include/asm/io.h | 4 ++++ arch/arm64/mm/ioremap.c | 8 ++++++++ 4 files changed, 23 insertions(+)
I think this looks reasonable, but you'll need to split it in two if it is going through the respective arch trees.
I guess that would be best. Otherwise, if Andrew picks it up:
Acked-by: Catalin Marinas catalin.marinas@arm.com
linux-stable-mirror@lists.linaro.org
-
Ard Biesheuvel -
Catalin Marinas -
Mike Rapoport