From: "Jonathan (Zhixiong) Zhang" zjzhang@codeaurora.org
On a platform with APEI (ACPI Platform Error Interface) enabled, firmware updates a memory region with hardware error record using nocache attribute. When OS reads the region, since it maps the region with cacahed attribute even though EFI memory map defines this region as uncached, OS gets stale data and errorneously reports there is no new HW error.
When ghes driver maps the memory region, it uses the cache attribute according to EFI memory map, if EFI memory map feature is enabled at runtime.
Since both arch/x86 and arch/ia64 implemented architecture agnostic EFI memory map attribue lookup function efi_memattributes(), the code is moved from arch/x86 into EFI subsystem and is declared as __weak; archs other than ia64 should not override the default implementation.
V10: 1. Fixed compilation issue on x86. V9: 1. Rebased to arm64-upstream-14543 of arm64/master. 2. Match strict MM type in arch_apei_get_mem_attribute(). V8: 1. For x86, always return PAGE_KERNEL for arch_apei_get_mem_attribute(). The rational is explained in comment. 2. Rebased to arm64-upstream-14201 of arm64/master, next-20150724 of linux-next/master. V7: 1. Added PROT_DEVICE_nGnRnE and PROT_NORMAL_WT to support all possible UEFI memory types for arm64. V6: 1. Implemented arch_apei_get_mem_attributes() for arm64 as inline function. 2. Rebased to efi-next-14364 of efi/next, pm+acpi-4.2-rc3 of linux-pm/master, arm64-upstream-13521 of arm64/master, next-20150720 of linux-next/master. V5: 1. Rebased to next-20150713 of linux-next/master, efi-next-14359 of efi/next, pm+acpi-4.2-rc2 of linux-pm/master, arm64-fixes-1215 of arm64/master. 2. Added comment for efi_mem_attributes(), explained why it is marked as __weak at the function definition site. V4: 1. Introduced arch_apei_get_mem_attributes() to allow arch specific implementation of getting pgprot_t appropriate for a physical address. 2. Implemented arch_apei_get_mem_attributes() for x86 and for arm64. V3: 1. Rebased to v4.1-rc7. 2. Moved efi_mem_attributes() from arch/x86 to drivers/firmware/efi and declared it as __weak. 3. Introduced ARCH_APEI_PAGE_KERNEL_UC to allow arch specific page protection type for UC. 4. Removed efi_ioremap(). It can not be used for GHES memory region mapping purpose since ioremap can not be used in atomic context. V2: 1. Rebased to v4.1-rc5. 2. Split removal of efi_mem_attributes() and creation of efi_ioremap() into two patches.
Jonathan (Zhixiong) Zhang (5): efi: x86: rearrange efi_mem_attributes() x86: acpi: implement arch_apei_get_mem_attributes() arm64: mm: add PROT_DEVICE_nGnRnE and PROT_NORMAL_WT arm64: apei: implement arch_apei_get_mem_attributes() acpi, apei: use appropriate pgprot_t to map GHES memory
arch/arm64/include/asm/acpi.h | 27 +++++++++++++++++++++++++++ arch/arm64/include/asm/memory.h | 1 + arch/arm64/include/asm/pgtable.h | 2 ++ arch/arm64/mm/proc.S | 4 +++- arch/x86/include/asm/acpi.h | 24 ++++++++++++++++++++++++ arch/x86/platform/efi/efi.c | 18 ------------------ drivers/acpi/apei/ghes.c | 6 ++++-- drivers/firmware/efi/efi.c | 31 +++++++++++++++++++++++++++++++ 8 files changed, 92 insertions(+), 21 deletions(-)
From: "Jonathan (Zhixiong) Zhang" zjzhang@codeaurora.org
x86 and ia64 implement efi_mem_attributes() differently. This function needs to be available for other arch (such as arm64) as well, such as for the purpose of ACPI/APEI.
ia64 efi does not setup memmap variable and does not set EFI_MEMMAP flag, so it needs to have its unique implementation of efi_mem_attributes().
Move efi_mem_attributes() implementation from x86 to efi, and declare it with __weak. It is recommended that other archs should not override the default implementation.
Reviewed-by: Matt Fleming matt.fleming@intel.com Signed-off-by: Jonathan (Zhixiong) Zhang zjzhang@codeaurora.org --- arch/x86/platform/efi/efi.c | 18 ------------------ drivers/firmware/efi/efi.c | 31 +++++++++++++++++++++++++++++++ 2 files changed, 31 insertions(+), 18 deletions(-)
diff --git a/arch/x86/platform/efi/efi.c b/arch/x86/platform/efi/efi.c index cfba30f27392..1ea9cb05163b 100644 --- a/arch/x86/platform/efi/efi.c +++ b/arch/x86/platform/efi/efi.c @@ -952,24 +952,6 @@ u32 efi_mem_type(unsigned long phys_addr) return 0; }
-u64 efi_mem_attributes(unsigned long phys_addr) -{ - efi_memory_desc_t *md; - void *p; - - if (!efi_enabled(EFI_MEMMAP)) - return 0; - - for (p = memmap.map; p < memmap.map_end; p += memmap.desc_size) { - md = p; - if ((md->phys_addr <= phys_addr) && - (phys_addr < (md->phys_addr + - (md->num_pages << EFI_PAGE_SHIFT)))) - return md->attribute; - } - return 0; -} - static int __init arch_parse_efi_cmdline(char *str) { if (parse_option_str(str, "old_map")) diff --git a/drivers/firmware/efi/efi.c b/drivers/firmware/efi/efi.c index 46eb8a6e9393..713f92f6b254 100644 --- a/drivers/firmware/efi/efi.c +++ b/drivers/firmware/efi/efi.c @@ -602,3 +602,34 @@ char * __init efi_md_typeattr_format(char *buf, size_t size, attr & EFI_MEMORY_UC ? "UC" : ""); return buf; } + +/* + * efi_mem_attributes - lookup memmap attributes for physical address + * @phys_addr: the physical address to lookup + * + * Search in the EFI memory map for the region covering + * @phys_addr. Returns the EFI memory attributes if the region + * was found in the memory map, 0 otherwise. + * + * Despite being marked __weak, most architectures should *not* + * override this function. It is __weak solely for the benefit + * of ia64 which has a funky EFI memory map that doesn't work + * the same way as other architectures. + */ +u64 __weak efi_mem_attributes(unsigned long phys_addr) +{ + efi_memory_desc_t *md; + void *p; + + if (!efi_enabled(EFI_MEMMAP)) + return 0; + + for (p = memmap.map; p < memmap.map_end; p += memmap.desc_size) { + md = p; + if ((md->phys_addr <= phys_addr) && + (phys_addr < (md->phys_addr + + (md->num_pages << EFI_PAGE_SHIFT)))) + return md->attribute; + } + return 0; +}
From: "Jonathan (Zhixiong) Zhang" zjzhang@codeaurora.org
... to allow arch specific implementation of getting page protection type associated with a physical address.
On x86, we currently have no way to lookup the EFI memory map attributes for a region in a consistent way because the memmap is discarded after efi_free_boot_services(). So if you call efi_mem_attributes() during boot and at runtime, you could theoretically see different attributes.
Since we are yet to see any x86 platforms that require anything other than PAGE_KERNEL (some arm64 platforms require the equivalent of PAGE_KERNEL_NOCACHE), return that until we know differently.
Signed-off-by: Jonathan (Zhixiong) Zhang zjzhang@codeaurora.org --- arch/x86/include/asm/acpi.h | 24 ++++++++++++++++++++++++ 1 file changed, 24 insertions(+)
diff --git a/arch/x86/include/asm/acpi.h b/arch/x86/include/asm/acpi.h index 3a45668f6dc3..ffe22c6b17bd 100644 --- a/arch/x86/include/asm/acpi.h +++ b/arch/x86/include/asm/acpi.h @@ -32,6 +32,10 @@ #include <asm/mpspec.h> #include <asm/realmode.h>
+#ifdef CONFIG_ACPI_APEI +#include <asm/pgtable_types.h> +#endif + #ifdef CONFIG_ACPI extern int acpi_lapic; extern int acpi_ioapic; @@ -147,4 +151,24 @@ extern int x86_acpi_numa_init(void);
#define acpi_unlazy_tlb(x) leave_mm(x)
+#ifdef CONFIG_ACPI_APEI +static inline pgprot_t arch_apei_get_mem_attribute(phys_addr_t addr) +{ + /* + * We currently have no way to lookup the EFI memory map + * attributes for a region in a consistent way because the + * memmap is discarded after efi_free_boot_services(). So if + * you call efi_mem_attributes() during boot and at runtime, + * you could theoretically see different attributes. + * + * Since we are yet to see any x86 platforms that require + * anything other than PAGE_KERNEL (some arm64 platforms + * require the equivalent of PAGE_KERNEL_NOCACHE), return that + * until we know differently. + */ + + return PAGE_KERNEL; +} +#endif + #endif /* _ASM_X86_ACPI_H */
From: "Jonathan (Zhixiong) Zhang" zjzhang@codeaurora.org
UEFI spec 2.5 section 2.3.6.1 defines that EFI_MEMORY_[UC|WC|WT|WB] are possible EFI memory types for AArch64. Each of those EFI memory types is mapped to a corresponding AArch64 memory type. So we need to define PROT_DEVICE_nGnRnE and PROT_NORMWL_WT additionaly.
MT_NORMAL_WT is defined, and its encoding is added to MAIR_EL1 when initializing cpu.
Reviewed-by: Catalin Marinas catalin.marinas@arm.com Signed-off-by: Jonathan (Zhixiong) Zhang zjzhang@codeaurora.org --- arch/arm64/include/asm/memory.h | 1 + arch/arm64/include/asm/pgtable.h | 2 ++ arch/arm64/mm/proc.S | 4 +++- 3 files changed, 6 insertions(+), 1 deletion(-)
diff --git a/arch/arm64/include/asm/memory.h b/arch/arm64/include/asm/memory.h index f800d45ea226..4112b3d7468e 100644 --- a/arch/arm64/include/asm/memory.h +++ b/arch/arm64/include/asm/memory.h @@ -100,6 +100,7 @@ #define MT_DEVICE_GRE 2 #define MT_NORMAL_NC 3 #define MT_NORMAL 4 +#define MT_NORMAL_WT 5
/* * Memory types for Stage-2 translation diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h index 56283f8a675c..0a105e3254a1 100644 --- a/arch/arm64/include/asm/pgtable.h +++ b/arch/arm64/include/asm/pgtable.h @@ -61,8 +61,10 @@ extern void __pgd_error(const char *file, int line, unsigned long val); #define PROT_SECT_DEFAULT (PMD_TYPE_SECT | PMD_SECT_AF) #endif
+#define PROT_DEVICE_nGnRnE (PROT_DEFAULT | PTE_PXN | PTE_UXN | PTE_ATTRINDX(MT_DEVICE_nGnRnE)) #define PROT_DEVICE_nGnRE (PROT_DEFAULT | PTE_PXN | PTE_UXN | PTE_ATTRINDX(MT_DEVICE_nGnRE)) #define PROT_NORMAL_NC (PROT_DEFAULT | PTE_PXN | PTE_UXN | PTE_ATTRINDX(MT_NORMAL_NC)) +#define PROT_NORMAL_WT (PROT_DEFAULT | PTE_PXN | PTE_UXN | PTE_ATTRINDX(MT_NORMAL_WT)) #define PROT_NORMAL (PROT_DEFAULT | PTE_PXN | PTE_UXN | PTE_ATTRINDX(MT_NORMAL))
#define PROT_SECT_DEVICE_nGnRE (PROT_SECT_DEFAULT | PMD_SECT_PXN | PMD_SECT_UXN | PMD_ATTRINDX(MT_DEVICE_nGnRE)) diff --git a/arch/arm64/mm/proc.S b/arch/arm64/mm/proc.S index 39139a3aa16d..160a1b5ab9c6 100644 --- a/arch/arm64/mm/proc.S +++ b/arch/arm64/mm/proc.S @@ -167,12 +167,14 @@ ENTRY(__cpu_setup) * DEVICE_GRE 010 00001100 * NORMAL_NC 011 01000100 * NORMAL 100 11111111 + * NORMAL_WT 101 10111011 */ ldr x5, =MAIR(0x00, MT_DEVICE_nGnRnE) | \ MAIR(0x04, MT_DEVICE_nGnRE) | \ MAIR(0x0c, MT_DEVICE_GRE) | \ MAIR(0x44, MT_NORMAL_NC) | \ - MAIR(0xff, MT_NORMAL) + MAIR(0xff, MT_NORMAL) | \ + MAIR(0xbb, MT_NORMAL_WT) msr mair_el1, x5 /* * Prepare SCTLR
From: "Jonathan (Zhixiong) Zhang" zjzhang@codeaurora.org
Table 8 of UEFI 2.5 section 2.3.6.1 defines mappings from EFI memory types to MAIR attribute encodings for arm64.
If the physical address has memory attributes defined by EFI memmap as EFI_MEMORY_[UC|WC|WT], return approprate page protection type according to the UEFI spec. Otherwise, return PAGE_KERNEL.
Reviewed-by: Catalin Marinas catalin.marinas@arm.com Acked-by: Hanjun Guo hanjun.guo@linaro.org Signed-off-by: Jonathan (Zhixiong) Zhang zjzhang@codeaurora.org --- This patch as is does not build on current tip of next branch of git.kernel.org/pub/scm/linux/kernel/git/mfleming/efi.git, a small tweak is needed. It builds on tip of linux-next/master of linus' repo and origin/master of arm64 repo. --- arch/arm64/include/asm/acpi.h | 26 ++++++++++++++++++++++++++ 1 file changed, 26 insertions(+)
diff --git a/arch/arm64/include/asm/acpi.h b/arch/arm64/include/asm/acpi.h index 1ff9e6eb5e02..1025d0401016 100644 --- a/arch/arm64/include/asm/acpi.h +++ b/arch/arm64/include/asm/acpi.h @@ -19,6 +19,11 @@ #include <asm/psci.h> #include <asm/smp_plat.h>
+#ifdef CONFIG_ACPI_APEI +#include <linux/efi.h> +#include <asm/pgtable.h> +#endif + /* Macros for consistency checks of the GICC subtable of MADT */ #define ACPI_MADT_GICC_LENGTH \ (acpi_gbl_FADT.header.revision < 6 ? 76 : 80) @@ -91,4 +95,26 @@ static inline const char *acpi_get_enable_method(int cpu) { return acpi_psci_present() ? "psci" : NULL; } + +#ifdef CONFIG_ACPI_APEI +/* + * According to "Table 8 Map: EFI memory types to AArch64 memory types" + * of UEFI 2.5 section 2.3.6.1, each EFI memory type is mapped to + * corresponding MAIR attribute encoding. + */ +static inline pgprot_t arch_apei_get_mem_attribute(phys_addr_t addr) +{ + u64 attr; + + attr = efi_mem_attributes(addr); + if (attr & EFI_MEMORY_UC) + return __pgprot(PROT_DEVICE_nGnRnE); + if (attr & EFI_MEMORY_WC) + return __pgprot(PROT_NORMAL_NC); + if (attr & EFI_MEMORY_WT) + return __pgprot(PROT_NORMAL_WT); + return __pgprot(PAGE_KERNEL); +} +#endif + #endif /*_ASM_ACPI_H*/
On 6 August 2015 at 15:40, Jonathan (Zhixiong) Zhang zjzhang@codeaurora.org wrote:
From: "Jonathan (Zhixiong) Zhang" zjzhang@codeaurora.org
Table 8 of UEFI 2.5 section 2.3.6.1 defines mappings from EFI memory types to MAIR attribute encodings for arm64.
If the physical address has memory attributes defined by EFI memmap as EFI_MEMORY_[UC|WC|WT], return approprate page protection type according to the UEFI spec. Otherwise, return PAGE_KERNEL.
Reviewed-by: Catalin Marinas catalin.marinas@arm.com Acked-by: Hanjun Guo hanjun.guo@linaro.org Signed-off-by: Jonathan (Zhixiong) Zhang zjzhang@codeaurora.org
This patch as is does not build on current tip of next branch of git.kernel.org/pub/scm/linux/kernel/git/mfleming/efi.git, a small tweak is needed. It builds on tip of linux-next/master of linus' repo and origin/master of arm64 repo.
arch/arm64/include/asm/acpi.h | 26 ++++++++++++++++++++++++++ 1 file changed, 26 insertions(+)
diff --git a/arch/arm64/include/asm/acpi.h b/arch/arm64/include/asm/acpi.h index 1ff9e6eb5e02..1025d0401016 100644 --- a/arch/arm64/include/asm/acpi.h +++ b/arch/arm64/include/asm/acpi.h @@ -19,6 +19,11 @@ #include <asm/psci.h> #include <asm/smp_plat.h>
+#ifdef CONFIG_ACPI_APEI +#include <linux/efi.h> +#include <asm/pgtable.h> +#endif
/* Macros for consistency checks of the GICC subtable of MADT */ #define ACPI_MADT_GICC_LENGTH \ (acpi_gbl_FADT.header.revision < 6 ? 76 : 80) @@ -91,4 +95,26 @@ static inline const char *acpi_get_enable_method(int cpu) { return acpi_psci_present() ? "psci" : NULL; }
+#ifdef CONFIG_ACPI_APEI +/*
- According to "Table 8 Map: EFI memory types to AArch64 memory types"
- of UEFI 2.5 section 2.3.6.1, each EFI memory type is mapped to
- corresponding MAIR attribute encoding.
- */
+static inline pgprot_t arch_apei_get_mem_attribute(phys_addr_t addr) +{
u64 attr;
attr = efi_mem_attributes(addr);
if (attr & EFI_MEMORY_UC)
return __pgprot(PROT_DEVICE_nGnRnE);
if (attr & EFI_MEMORY_WC)
return __pgprot(PROT_NORMAL_NC);
if (attr & EFI_MEMORY_WT)
return __pgprot(PROT_NORMAL_WT);
return __pgprot(PAGE_KERNEL);
The EFI memory types are not exclusive, and so many regions will have all of the above set. The UEFI spec does not define how to interpret these superimposed attributes, it is up to the OS to decide on a consistent approach.
For instance, this region (captured from a arm64 boot log with uefi_debug set on the command line)
[Runtime Data |RUN| | | | |WB|WT|WC|UC]
would be mapped uncached when following the above logic, while it makes more sense to map using PAGE_KERNEL in this case.
From the spec:
""" EFI_MEMORY_UC: The memory region supports being configured as not cacheable. EFI_MEMORY_WC: The memory region supports being configured as write combining. EFI_MEMORY_WT: The memory region supports being configured as cacheable with a “write through” policy. Writes that hit in the cache will also be written to main memory. EFI_MEMORY_WB: The memory region supports being configured as cacheable with a “write back” policy. Reads and writes that hit in the cache do not propagate to main memory. Dirty data is written back to main memory when a new cache line is allocated. """
Also, the final __pgprot() is redundant here, it is already present in the definition of PAGE_KERNEL.
On Fri, 07 Aug, at 11:00:17AM, Ard Biesheuvel wrote:
The EFI memory types are not exclusive, and so many regions will have all of the above set. The UEFI spec does not define how to interpret these superimposed attributes, it is up to the OS to decide on a consistent approach.
For instance, this region (captured from a arm64 boot log with uefi_debug set on the command line)
[Runtime Data |RUN| | | | |WB|WT|WC|UC]
would be mapped uncached when following the above logic, while it makes more sense to map using PAGE_KERNEL in this case.
Urgh... good point Ard. Right now this is limited to the GHES driver, so it's unclear whether this patch is buggy in practice or not.
Does it *ever* make sense to map a region as cacheable (WB/WT/WC) on arm64 for the APEI/GHES case? Does the firmware handle the necessary cache flushing?
From the spec:
""" EFI_MEMORY_UC: The memory region supports being configured as not cacheable. EFI_MEMORY_WC: The memory region supports being configured as write combining. EFI_MEMORY_WT: The memory region supports being configured as cacheable with a “write through” policy. Writes that hit in the cache will also be written to main memory. EFI_MEMORY_WB: The memory region supports being configured as cacheable with a “write back” policy. Reads and writes that hit in the cache do not propagate to main memory. Dirty data is written back to main memory when a new cache line is allocated. """
Jonathan, can you please provide the EFI memory map region attributes for the GHES region that requires this series?
On 7 August 2015 at 11:37, Matt Fleming matt@codeblueprint.co.uk wrote:
On Fri, 07 Aug, at 11:00:17AM, Ard Biesheuvel wrote:
The EFI memory types are not exclusive, and so many regions will have all of the above set. The UEFI spec does not define how to interpret these superimposed attributes, it is up to the OS to decide on a consistent approach.
For instance, this region (captured from a arm64 boot log with uefi_debug set on the command line)
[Runtime Data |RUN| | | | |WB|WT|WC|UC]
would be mapped uncached when following the above logic, while it makes more sense to map using PAGE_KERNEL in this case.
Urgh... good point Ard. Right now this is limited to the GHES driver, so it's unclear whether this patch is buggy in practice or not.
Does it *ever* make sense to map a region as cacheable (WB/WT/WC) on arm64 for the APEI/GHES case? Does the firmware handle the necessary cache flushing?
No it does not. Currently, we only consider EFI_MEMORY_WB when discovering system RAM from the UEFI memory map, so the direct linear mapping should have a hole where the APEI/GHES regions lives if it doesn't have the WB attribute set. This means we can map it WT/WC/UC without violating architectural rules regarding mismatches attributes, but it requires an explicit ioremap()
From the spec:
""" EFI_MEMORY_UC: The memory region supports being configured as not cacheable. EFI_MEMORY_WC: The memory region supports being configured as write combining. EFI_MEMORY_WT: The memory region supports being configured as cacheable with a “write through” policy. Writes that hit in the cache will also be written to main memory. EFI_MEMORY_WB: The memory region supports being configured as cacheable with a “write back” policy. Reads and writes that hit in the cache do not propagate to main memory. Dirty data is written back to main memory when a new cache line is allocated. """
Jonathan, can you please provide the EFI memory map region attributes for the GHES region that requires this series?
-- Matt Fleming, Intel Open Source Technology Center
On 8/7/2015 2:50 AM, Ard Biesheuvel wrote:
On 7 August 2015 at 11:37, Matt Fleming matt@codeblueprint.co.uk wrote:
On Fri, 07 Aug, at 11:00:17AM, Ard Biesheuvel wrote:
The EFI memory types are not exclusive, and so many regions will have all of the above set. The UEFI spec does not define how to interpret these superimposed attributes, it is up to the OS to decide on a consistent approach.
For instance, this region (captured from a arm64 boot log with uefi_debug set on the command line)
[Runtime Data |RUN| | | | |WB|WT|WC|UC]
would be mapped uncached when following the above logic, while it makes more sense to map using PAGE_KERNEL in this case.
Urgh... good point Ard. Right now this is limited to the GHES driver, so it's unclear whether this patch is buggy in practice or not.
Does it *ever* make sense to map a region as cacheable (WB/WT/WC) on arm64 for the APEI/GHES case? Does the firmware handle the necessary cache flushing?
No it does not. Currently, we only consider EFI_MEMORY_WB when discovering system RAM from the UEFI memory map, so the direct linear mapping should have a hole where the APEI/GHES regions lives if it doesn't have the WB attribute set. This means we can map it WT/WC/UC without violating architectural rules regarding mismatches attributes, but it requires an explicit ioremap()
On some (future) arm64 platforms, APEI/GHES region may have full coherent access by platform. In such case, the APEI/GHES region have the same memory attributes as the rest of system RAM, such region do not need to be advised by UEFI as separate entry, but as part of system RAM memory region. That being said, for arm64 platforms that do not have WB capability for APEI/GHES region, such region should be mapped accordingly.
From the spec:
""" EFI_MEMORY_UC: The memory region supports being configured as not cacheable. EFI_MEMORY_WC: The memory region supports being configured as write combining. EFI_MEMORY_WT: The memory region supports being configured as cacheable with a “write through” policy. Writes that hit in the cache will also be written to main memory. EFI_MEMORY_WB: The memory region supports being configured as cacheable with a “write back” policy. Reads and writes that hit in the cache do not propagate to main memory. Dirty data is written back to main memory when a new cache line is allocated. """
Jonathan, can you please provide the EFI memory map region attributes for the GHES region that requires this series?
[Reserved | | | | | | | | |UC]
-- Matt Fleming, Intel Open Source Technology Center
On Fri, 07 Aug, at 11:16:03AM, Zhang, Jonathan Zhixiong wrote:
On some (future) arm64 platforms, APEI/GHES region may have full coherent access by platform. In such case, the APEI/GHES region have the same memory attributes as the rest of system RAM, such region do not need to be advised by UEFI as separate entry, but as part of system RAM memory region. That being said, for arm64 platforms that do not have WB capability for APEI/GHES region, such region should be mapped accordingly.
OK, so what I need to know right now is whether I need to drop this entire series from my pull request or whether you can send a patch on top of the existing ones in the EFI 'next' branch to address the mapping heuristic in arch_apei_get_mem_attributes().
Jonathan, can you please provide the EFI memory map region attributes for the GHES region that requires this series?
[Reserved | | | | | | | | |UC]
Assuming this memmap entry is indicative of most GHES region on arm64 right now, I think it's worth taking this patch as-is and addressing the issue Ard raised as a separate patch.
Does that work?
On 7 August 2015 at 20:57, Matt Fleming matt@codeblueprint.co.uk wrote:
On Fri, 07 Aug, at 11:16:03AM, Zhang, Jonathan Zhixiong wrote:
On some (future) arm64 platforms, APEI/GHES region may have full coherent access by platform. In such case, the APEI/GHES region have the same memory attributes as the rest of system RAM, such region do not need to be advised by UEFI as separate entry, but as part of system RAM memory region. That being said, for arm64 platforms that do not have WB capability for APEI/GHES region, such region should be mapped accordingly.
OK, so what I need to know right now is whether I need to drop this entire series from my pull request or whether you can send a patch on top of the existing ones in the EFI 'next' branch to address the mapping heuristic in arch_apei_get_mem_attributes().
Jonathan, can you please provide the EFI memory map region attributes for the GHES region that requires this series?
[Reserved | | | | | | | | |UC]
Assuming this memmap entry is indicative of most GHES region on arm64 right now, I think it's worth taking this patch as-is and addressing the issue Ard raised as a separate patch.
Does that work?
I think that is fine.
So we'll expect two patches on top of Matt's -next branch: - one that removes the redundant __pgprot - one that inverts the order in which the memory attributes are tested
It would be good to have these in the same release so that the behavior does not change between releases.
Thanks, Ard.
On 8/8/2015 1:11 AM, Ard Biesheuvel wrote:
On 7 August 2015 at 20:57, Matt Fleming matt@codeblueprint.co.uk wrote:
On Fri, 07 Aug, at 11:16:03AM, Zhang, Jonathan Zhixiong wrote:
On some (future) arm64 platforms, APEI/GHES region may have full coherent access by platform. In such case, the APEI/GHES region have the same memory attributes as the rest of system RAM, such region do not need to be advised by UEFI as separate entry, but as part of system RAM memory region. That being said, for arm64 platforms that do not have WB capability for APEI/GHES region, such region should be mapped accordingly.
OK, so what I need to know right now is whether I need to drop this entire series from my pull request or whether you can send a patch on top of the existing ones in the EFI 'next' branch to address the mapping heuristic in arch_apei_get_mem_attributes().
Jonathan, can you please provide the EFI memory map region attributes for the GHES region that requires this series?
[Reserved | | | | | | | | |UC]
Assuming this memmap entry is indicative of most GHES region on arm64 right now, I think it's worth taking this patch as-is and addressing the issue Ard raised as a separate patch.
Does that work?
I think that is fine.
So we'll expect two patches on top of Matt's -next branch:
- one that removes the redundant __pgprot
- one that inverts the order in which the memory attributes are tested
It would be good to have these in the same release so that the behavior does not change between releases.
I have above mentioned two changes ready to go. I will send V11 of the patch set out in the next hour.
Thanks, Ard.
Thanks for the review, Ard. Comments inline below.
On 8/7/2015 2:00 AM, Ard Biesheuvel wrote:
On 6 August 2015 at 15:40, Jonathan (Zhixiong) Zhang zjzhang@codeaurora.org wrote:
From: "Jonathan (Zhixiong) Zhang" zjzhang@codeaurora.org
Table 8 of UEFI 2.5 section 2.3.6.1 defines mappings from EFI memory types to MAIR attribute encodings for arm64.
If the physical address has memory attributes defined by EFI memmap as EFI_MEMORY_[UC|WC|WT], return approprate page protection type according to the UEFI spec. Otherwise, return PAGE_KERNEL.
Reviewed-by: Catalin Marinas catalin.marinas@arm.com Acked-by: Hanjun Guo hanjun.guo@linaro.org Signed-off-by: Jonathan (Zhixiong) Zhang zjzhang@codeaurora.org
This patch as is does not build on current tip of next branch of git.kernel.org/pub/scm/linux/kernel/git/mfleming/efi.git, a small tweak is needed. It builds on tip of linux-next/master of linus' repo and origin/master of arm64 repo.
arch/arm64/include/asm/acpi.h | 26 ++++++++++++++++++++++++++ 1 file changed, 26 insertions(+)
diff --git a/arch/arm64/include/asm/acpi.h b/arch/arm64/include/asm/acpi.h index 1ff9e6eb5e02..1025d0401016 100644 --- a/arch/arm64/include/asm/acpi.h +++ b/arch/arm64/include/asm/acpi.h @@ -19,6 +19,11 @@ #include <asm/psci.h> #include <asm/smp_plat.h>
+#ifdef CONFIG_ACPI_APEI +#include <linux/efi.h> +#include <asm/pgtable.h> +#endif
- /* Macros for consistency checks of the GICC subtable of MADT */ #define ACPI_MADT_GICC_LENGTH \ (acpi_gbl_FADT.header.revision < 6 ? 76 : 80)
@@ -91,4 +95,26 @@ static inline const char *acpi_get_enable_method(int cpu) { return acpi_psci_present() ? "psci" : NULL; }
+#ifdef CONFIG_ACPI_APEI +/*
- According to "Table 8 Map: EFI memory types to AArch64 memory types"
- of UEFI 2.5 section 2.3.6.1, each EFI memory type is mapped to
- corresponding MAIR attribute encoding.
- */
+static inline pgprot_t arch_apei_get_mem_attribute(phys_addr_t addr) +{
u64 attr;
attr = efi_mem_attributes(addr);
if (attr & EFI_MEMORY_UC)
return __pgprot(PROT_DEVICE_nGnRnE);
if (attr & EFI_MEMORY_WC)
return __pgprot(PROT_NORMAL_NC);
if (attr & EFI_MEMORY_WT)
return __pgprot(PROT_NORMAL_WT);
return __pgprot(PAGE_KERNEL);
The EFI memory types are not exclusive, and so many regions will have all of the above set. The UEFI spec does not define how to interpret these superimposed attributes, it is up to the OS to decide on a consistent approach.
For instance, this region (captured from a arm64 boot log with uefi_debug set on the command line)
[Runtime Data |RUN| | | | |WB|WT|WC|UC]
would be mapped uncached when following the above logic, while it makes more sense to map using PAGE_KERNEL in this case.
From the spec:
""" EFI_MEMORY_UC: The memory region supports being configured as not cacheable. EFI_MEMORY_WC: The memory region supports being configured as write combining. EFI_MEMORY_WT: The memory region supports being configured as cacheable with a “write through” policy. Writes that hit in the cache will also be written to main memory. EFI_MEMORY_WB: The memory region supports being configured as cacheable with a “write back” policy. Reads and writes that hit in the cache do not propagate to main memory. Dirty data is written back to main memory when a new cache line is allocated. """
EFI memory map advises all possible (memory type) capabilities of a memory region. As you said, Linux should decide on a consistent approach. What about us using the most efficient capability when the attributes are superimposed? In such case, we will simply revert the sequence in the above code.
Also, the final __pgprot() is redundant here, it is already present in the definition of PAGE_KERNEL.
Yes, will fix it.
On Fri, 07 Aug, at 10:40:20AM, Zhang, Jonathan Zhixiong wrote:
Also, the final __pgprot() is redundant here, it is already present in the definition of PAGE_KERNEL.
Yes, will fix it.
I've already picked up this series and sent it as a pull request. Please write the patch that cleans up __pgprot() ontop of the 'next' branch at,
git://git.kernel.org/pub/scm/linux/kernel/git/mfleming/efi.git
From: "Jonathan (Zhixiong) Zhang" zjzhang@codeaurora.org
With ACPI APEI firmware first handling, generic hardware error record is updated by firmware in GHES memory region. On an arm64 platform, firmware updates GHES memory region with uncached access attribute, and then Linux reads stale data from cache.
GHES memory region should be mapped with page protection type according to what is returned from arch_apei_get_mem_attribute(), instead of always with PAGE_KERNEL (eg. cached attribute).
Acked-by: Borislav Petkov bp@suse.de Signed-off-by: Jonathan (Zhixiong) Zhang zjzhang@codeaurora.org --- drivers/acpi/apei/ghes.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c index 2bfd53cbfe80..0aa37c57acec 100644 --- a/drivers/acpi/apei/ghes.c +++ b/drivers/acpi/apei/ghes.c @@ -164,8 +164,10 @@ static void __iomem *ghes_ioremap_pfn_irq(u64 pfn) unsigned long vaddr;
vaddr = (unsigned long)GHES_IOREMAP_IRQ_PAGE(ghes_ioremap_area->addr); - ioremap_page_range(vaddr, vaddr + PAGE_SIZE, - pfn << PAGE_SHIFT, PAGE_KERNEL); + ioremap_page_range(vaddr, + vaddr + PAGE_SIZE, + pfn << PAGE_SHIFT, + arch_apei_get_mem_attribute(pfn << PAGE_SHIFT));
return (void __iomem *)vaddr; }