linux/arch/arm/include/asm/memory.h

390 lines
10 KiB

ARM: mm: Make virt_to_pfn() a static inline Making virt_to_pfn() a static inline taking a strongly typed (const void *) makes the contract of a passing a pointer of that type to the function explicit and exposes any misuse of the macro virt_to_pfn() acting polymorphic and accepting many types such as (void *), (unitptr_t) or (unsigned long) as arguments without warnings. Doing this is a bit intrusive: virt_to_pfn() requires PHYS_PFN_OFFSET and PAGE_SHIFT to be defined, and this is defined in <asm/page.h>, so this must be included *before* <asm/memory.h>. The use of macros were obscuring the unclear inclusion order here, as the macros would eventually be resolved, but a static inline like this cannot be compiled with unresolved macros. The naive solution to include <asm/page.h> at the top of <asm/memory.h> does not work, because <asm/memory.h> sometimes includes <asm/page.h> at the end of itself, which would create a confusing inclusion loop. So instead, take the approach to always unconditionally include <asm/page.h> at the end of <asm/memory.h> arch/arm uses <asm/memory.h> explicitly in a lot of places, however it turns out that if we just unconditionally include <asm/memory.h> into <asm/page.h> and switch all inclusions of <asm/memory.h> to <asm/page.h> instead, we enforce the right order and <asm/memory.h> will always have access to the definitions. Put an inclusion guard in place making it impossible to include <asm/memory.h> explicitly. Link: https://lore.kernel.org/linux-mm/20220701160004.2ffff4e5ab59a55499f4c736@linux-foundation.org/ Signed-off-by: Linus Walleij <linus.walleij@linaro.org>
2022-06-02 10:18:32 +02:00
ARM: mm: Make virt_to_pfn() a static inline Making virt_to_pfn() a static inline taking a strongly typed (const void *) makes the contract of a passing a pointer of that type to the function explicit and exposes any misuse of the macro virt_to_pfn() acting polymorphic and accepting many types such as (void *), (unitptr_t) or (unsigned long) as arguments without warnings. Doing this is a bit intrusive: virt_to_pfn() requires PHYS_PFN_OFFSET and PAGE_SHIFT to be defined, and this is defined in <asm/page.h>, so this must be included *before* <asm/memory.h>. The use of macros were obscuring the unclear inclusion order here, as the macros would eventually be resolved, but a static inline like this cannot be compiled with unresolved macros. The naive solution to include <asm/page.h> at the top of <asm/memory.h> does not work, because <asm/memory.h> sometimes includes <asm/page.h> at the end of itself, which would create a confusing inclusion loop. So instead, take the approach to always unconditionally include <asm/page.h> at the end of <asm/memory.h> arch/arm uses <asm/memory.h> explicitly in a lot of places, however it turns out that if we just unconditionally include <asm/memory.h> into <asm/page.h> and switch all inclusions of <asm/memory.h> to <asm/page.h> instead, we enforce the right order and <asm/memory.h> will always have access to the definitions. Put an inclusion guard in place making it impossible to include <asm/memory.h> explicitly. Link: https://lore.kernel.org/linux-mm/20220701160004.2ffff4e5ab59a55499f4c736@linux-foundation.org/ Signed-off-by: Linus Walleij <linus.walleij@linaro.org>
2022-06-02 10:18:32 +02:00
ARM: 9015/2: Define the virtual space of KASan's shadow region Define KASAN_SHADOW_OFFSET,KASAN_SHADOW_START and KASAN_SHADOW_END for the Arm kernel address sanitizer. We are "stealing" lowmem (the 4GB addressable by a 32bit architecture) out of the virtual address space to use as shadow memory for KASan as follows: +----+ 0xffffffff | | | | |-> Static kernel image (vmlinux) BSS and page table | |/ +----+ PAGE_OFFSET | | | | |-> Loadable kernel modules virtual address space area | |/ +----+ MODULES_VADDR = KASAN_SHADOW_END | | | | |-> The shadow area of kernel virtual address. | |/ +----+-> TASK_SIZE (start of kernel space) = KASAN_SHADOW_START the | | shadow address of MODULES_VADDR | | | | | | | | |-> The user space area in lowmem. The kernel address | | | sanitizer do not use this space, nor does it map it. | | | | | | | | | | | | | |/ ------ 0 0 .. TASK_SIZE is the memory that can be used by shared userspace/kernelspace. It us used for userspace processes and for passing parameters and memory buffers in system calls etc. We do not need to shadow this area. KASAN_SHADOW_START: This value begins with the MODULE_VADDR's shadow address. It is the start of kernel virtual space. Since we have modules to load, we need to cover also that area with shadow memory so we can find memory bugs in modules. KASAN_SHADOW_END This value is the 0x100000000's shadow address: the mapping that would be after the end of the kernel memory at 0xffffffff. It is the end of kernel address sanitizer shadow area. It is also the start of the module area. KASAN_SHADOW_OFFSET: This value is used to map an address to the corresponding shadow address by the following formula: shadow_addr = (address >> 3) + KASAN_SHADOW_OFFSET; As you would expect, >> 3 is equal to dividing by 8, meaning each byte in the shadow memory covers 8 bytes of kernel memory, so one bit shadow memory per byte of kernel memory is used. The KASAN_SHADOW_OFFSET is provided in a Kconfig option depending on the VMSPLIT layout of the system: the kernel and userspace can split up lowmem in different ways according to needs, so we calculate the shadow offset depending on this. When kasan is enabled, the definition of TASK_SIZE is not an 8-bit rotated constant, so we need to modify the TASK_SIZE access code in the *.s file. The kernel and modules may use different amounts of memory, according to the VMSPLIT configuration, which in turn determines the PAGE_OFFSET. We use the following KASAN_SHADOW_OFFSETs depending on how the virtual memory is split up: - 0x1f000000 if we have 1G userspace / 3G kernelspace split: - The kernel address space is 3G (0xc0000000) - PAGE_OFFSET is then set to 0x40000000 so the kernel static image (vmlinux) uses addresses 0x40000000 .. 0xffffffff - On top of that we have the MODULES_VADDR which under the worst case (using ARM instructions) is PAGE_OFFSET - 16M (0x01000000) = 0x3f000000 so the modules use addresses 0x3f000000 .. 0x3fffffff - So the addresses 0x3f000000 .. 0xffffffff need to be covered with shadow memory. That is 0xc1000000 bytes of memory. - 1/8 of that is needed for its shadow memory, so 0x18200000 bytes of shadow memory is needed. We "steal" that from the remaining lowmem. - The KASAN_SHADOW_START becomes 0x26e00000, to KASAN_SHADOW_END at 0x3effffff. - Now we can calculate the KASAN_SHADOW_OFFSET for any kernel address as 0x3f000000 needs to map to the first byte of shadow memory and 0xffffffff needs to map to the last byte of shadow memory. Since: SHADOW_ADDR = (address >> 3) + KASAN_SHADOW_OFFSET 0x26e00000 = (0x3f000000 >> 3) + KASAN_SHADOW_OFFSET KASAN_SHADOW_OFFSET = 0x26e00000 - (0x3f000000 >> 3) KASAN_SHADOW_OFFSET = 0x26e00000 - 0x07e00000 KASAN_SHADOW_OFFSET = 0x1f000000 - 0x5f000000 if we have 2G userspace / 2G kernelspace split: - The kernel space is 2G (0x80000000) - PAGE_OFFSET is set to 0x80000000 so the kernel static image uses 0x80000000 .. 0xffffffff. - On top of that we have the MODULES_VADDR which under the worst case (using ARM instructions) is PAGE_OFFSET - 16M (0x01000000) = 0x7f000000 so the modules use addresses 0x7f000000 .. 0x7fffffff - So the addresses 0x7f000000 .. 0xffffffff need to be covered with shadow memory. That is 0x81000000 bytes of memory. - 1/8 of that is needed for its shadow memory, so 0x10200000 bytes of shadow memory is needed. We "steal" that from the remaining lowmem. - The KASAN_SHADOW_START becomes 0x6ee00000, to KASAN_SHADOW_END at 0x7effffff. - Now we can calculate the KASAN_SHADOW_OFFSET for any kernel address as 0x7f000000 needs to map to the first byte of shadow memory and 0xffffffff needs to map to the last byte of shadow memory. Since: SHADOW_ADDR = (address >> 3) + KASAN_SHADOW_OFFSET 0x6ee00000 = (0x7f000000 >> 3) + KASAN_SHADOW_OFFSET KASAN_SHADOW_OFFSET = 0x6ee00000 - (0x7f000000 >> 3) KASAN_SHADOW_OFFSET = 0x6ee00000 - 0x0fe00000 KASAN_SHADOW_OFFSET = 0x5f000000 - 0x9f000000 if we have 3G userspace / 1G kernelspace split, and this is the default split for ARM: - The kernel address space is 1GB (0x40000000) - PAGE_OFFSET is set to 0xc0000000 so the kernel static image uses 0xc0000000 .. 0xffffffff. - On top of that we have the MODULES_VADDR which under the worst case (using ARM instructions) is PAGE_OFFSET - 16M (0x01000000) = 0xbf000000 so the modules use addresses 0xbf000000 .. 0xbfffffff - So the addresses 0xbf000000 .. 0xffffffff need to be covered with shadow memory. That is 0x41000000 bytes of memory. - 1/8 of that is needed for its shadow memory, so 0x08200000 bytes of shadow memory is needed. We "steal" that from the remaining lowmem. - The KASAN_SHADOW_START becomes 0xb6e00000, to KASAN_SHADOW_END at 0xbfffffff. - Now we can calculate the KASAN_SHADOW_OFFSET for any kernel address as 0xbf000000 needs to map to the first byte of shadow memory and 0xffffffff needs to map to the last byte of shadow memory. Since: SHADOW_ADDR = (address >> 3) + KASAN_SHADOW_OFFSET 0xb6e00000 = (0xbf000000 >> 3) + KASAN_SHADOW_OFFSET KASAN_SHADOW_OFFSET = 0xb6e00000 - (0xbf000000 >> 3) KASAN_SHADOW_OFFSET = 0xb6e00000 - 0x17e00000 KASAN_SHADOW_OFFSET = 0x9f000000 - 0x8f000000 if we have 3G userspace / 1G kernelspace with full 1 GB low memory (VMSPLIT_3G_OPT): - The kernel address space is 1GB (0x40000000) - PAGE_OFFSET is set to 0xb0000000 so the kernel static image uses 0xb0000000 .. 0xffffffff. - On top of that we have the MODULES_VADDR which under the worst case (using ARM instructions) is PAGE_OFFSET - 16M (0x01000000) = 0xaf000000 so the modules use addresses 0xaf000000 .. 0xaffffff - So the addresses 0xaf000000 .. 0xffffffff need to be covered with shadow memory. That is 0x51000000 bytes of memory. - 1/8 of that is needed for its shadow memory, so 0x0a200000 bytes of shadow memory is needed. We "steal" that from the remaining lowmem. - The KASAN_SHADOW_START becomes 0xa4e00000, to KASAN_SHADOW_END at 0xaeffffff. - Now we can calculate the KASAN_SHADOW_OFFSET for any kernel address as 0xaf000000 needs to map to the first byte of shadow memory and 0xffffffff needs to map to the last byte of shadow memory. Since: SHADOW_ADDR = (address >> 3) + KASAN_SHADOW_OFFSET 0xa4e00000 = (0xaf000000 >> 3) + KASAN_SHADOW_OFFSET KASAN_SHADOW_OFFSET = 0xa4e00000 - (0xaf000000 >> 3) KASAN_SHADOW_OFFSET = 0xa4e00000 - 0x15e00000 KASAN_SHADOW_OFFSET = 0x8f000000 - The default value of 0xffffffff for KASAN_SHADOW_OFFSET is an error value. We should always match one of the above shadow offsets. When we do this, TASK_SIZE will sometimes get a bit odd values that will not fit into immediate mov assembly instructions. To account for this, we need to rewrite some assembly using TASK_SIZE like this: - mov r1, #TASK_SIZE + ldr r1, =TASK_SIZE or - cmp r4, #TASK_SIZE + ldr r0, =TASK_SIZE + cmp r4, r0 this is done to avoid the immediate #TASK_SIZE that need to fit into a limited number of bits. Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: kasan-dev@googlegroups.com Cc: Mike Rapoport <rppt@linux.ibm.com> Reviewed-by: Ard Biesheuvel <ardb@kernel.org> Tested-by: Ard Biesheuvel <ardb@kernel.org> # QEMU/KVM/mach-virt/LPAE/8G Tested-by: Florian Fainelli <f.fainelli@gmail.com> # Brahma SoCs Tested-by: Ahmad Fatoum <a.fatoum@pengutronix.de> # i.MX6Q Reported-by: Ard Biesheuvel <ardb@kernel.org> Signed-off-by: Abbott Liu <liuwenliang@huawei.com> Signed-off-by: Florian Fainelli <f.fainelli@gmail.com> Signed-off-by: Linus Walleij <linus.walleij@linaro.org> Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
2020-10-25 23:53:46 +01:00
ARM: 9015/2: Define the virtual space of KASan's shadow region Define KASAN_SHADOW_OFFSET,KASAN_SHADOW_START and KASAN_SHADOW_END for the Arm kernel address sanitizer. We are "stealing" lowmem (the 4GB addressable by a 32bit architecture) out of the virtual address space to use as shadow memory for KASan as follows: +----+ 0xffffffff | | | | |-> Static kernel image (vmlinux) BSS and page table | |/ +----+ PAGE_OFFSET | | | | |-> Loadable kernel modules virtual address space area | |/ +----+ MODULES_VADDR = KASAN_SHADOW_END | | | | |-> The shadow area of kernel virtual address. | |/ +----+-> TASK_SIZE (start of kernel space) = KASAN_SHADOW_START the | | shadow address of MODULES_VADDR | | | | | | | | |-> The user space area in lowmem. The kernel address | | | sanitizer do not use this space, nor does it map it. | | | | | | | | | | | | | |/ ------ 0 0 .. TASK_SIZE is the memory that can be used by shared userspace/kernelspace. It us used for userspace processes and for passing parameters and memory buffers in system calls etc. We do not need to shadow this area. KASAN_SHADOW_START: This value begins with the MODULE_VADDR's shadow address. It is the start of kernel virtual space. Since we have modules to load, we need to cover also that area with shadow memory so we can find memory bugs in modules. KASAN_SHADOW_END This value is the 0x100000000's shadow address: the mapping that would be after the end of the kernel memory at 0xffffffff. It is the end of kernel address sanitizer shadow area. It is also the start of the module area. KASAN_SHADOW_OFFSET: This value is used to map an address to the corresponding shadow address by the following formula: shadow_addr = (address >> 3) + KASAN_SHADOW_OFFSET; As you would expect, >> 3 is equal to dividing by 8, meaning each byte in the shadow memory covers 8 bytes of kernel memory, so one bit shadow memory per byte of kernel memory is used. The KASAN_SHADOW_OFFSET is provided in a Kconfig option depending on the VMSPLIT layout of the system: the kernel and userspace can split up lowmem in different ways according to needs, so we calculate the shadow offset depending on this. When kasan is enabled, the definition of TASK_SIZE is not an 8-bit rotated constant, so we need to modify the TASK_SIZE access code in the *.s file. The kernel and modules may use different amounts of memory, according to the VMSPLIT configuration, which in turn determines the PAGE_OFFSET. We use the following KASAN_SHADOW_OFFSETs depending on how the virtual memory is split up: - 0x1f000000 if we have 1G userspace / 3G kernelspace split: - The kernel address space is 3G (0xc0000000) - PAGE_OFFSET is then set to 0x40000000 so the kernel static image (vmlinux) uses addresses 0x40000000 .. 0xffffffff - On top of that we have the MODULES_VADDR which under the worst case (using ARM instructions) is PAGE_OFFSET - 16M (0x01000000) = 0x3f000000 so the modules use addresses 0x3f000000 .. 0x3fffffff - So the addresses 0x3f000000 .. 0xffffffff need to be covered with shadow memory. That is 0xc1000000 bytes of memory. - 1/8 of that is needed for its shadow memory, so 0x18200000 bytes of shadow memory is needed. We "steal" that from the remaining lowmem. - The KASAN_SHADOW_START becomes 0x26e00000, to KASAN_SHADOW_END at 0x3effffff. - Now we can calculate the KASAN_SHADOW_OFFSET for any kernel address as 0x3f000000 needs to map to the first byte of shadow memory and 0xffffffff needs to map to the last byte of shadow memory. Since: SHADOW_ADDR = (address >> 3) + KASAN_SHADOW_OFFSET 0x26e00000 = (0x3f000000 >> 3) + KASAN_SHADOW_OFFSET KASAN_SHADOW_OFFSET = 0x26e00000 - (0x3f000000 >> 3) KASAN_SHADOW_OFFSET = 0x26e00000 - 0x07e00000 KASAN_SHADOW_OFFSET = 0x1f000000 - 0x5f000000 if we have 2G userspace / 2G kernelspace split: - The kernel space is 2G (0x80000000) - PAGE_OFFSET is set to 0x80000000 so the kernel static image uses 0x80000000 .. 0xffffffff. - On top of that we have the MODULES_VADDR which under the worst case (using ARM instructions) is PAGE_OFFSET - 16M (0x01000000) = 0x7f000000 so the modules use addresses 0x7f000000 .. 0x7fffffff - So the addresses 0x7f000000 .. 0xffffffff need to be covered with shadow memory. That is 0x81000000 bytes of memory. - 1/8 of that is needed for its shadow memory, so 0x10200000 bytes of shadow memory is needed. We "steal" that from the remaining lowmem. - The KASAN_SHADOW_START becomes 0x6ee00000, to KASAN_SHADOW_END at 0x7effffff. - Now we can calculate the KASAN_SHADOW_OFFSET for any kernel address as 0x7f000000 needs to map to the first byte of shadow memory and 0xffffffff needs to map to the last byte of shadow memory. Since: SHADOW_ADDR = (address >> 3) + KASAN_SHADOW_OFFSET 0x6ee00000 = (0x7f000000 >> 3) + KASAN_SHADOW_OFFSET KASAN_SHADOW_OFFSET = 0x6ee00000 - (0x7f000000 >> 3) KASAN_SHADOW_OFFSET = 0x6ee00000 - 0x0fe00000 KASAN_SHADOW_OFFSET = 0x5f000000 - 0x9f000000 if we have 3G userspace / 1G kernelspace split, and this is the default split for ARM: - The kernel address space is 1GB (0x40000000) - PAGE_OFFSET is set to 0xc0000000 so the kernel static image uses 0xc0000000 .. 0xffffffff. - On top of that we have the MODULES_VADDR which under the worst case (using ARM instructions) is PAGE_OFFSET - 16M (0x01000000) = 0xbf000000 so the modules use addresses 0xbf000000 .. 0xbfffffff - So the addresses 0xbf000000 .. 0xffffffff need to be covered with shadow memory. That is 0x41000000 bytes of memory. - 1/8 of that is needed for its shadow memory, so 0x08200000 bytes of shadow memory is needed. We "steal" that from the remaining lowmem. - The KASAN_SHADOW_START becomes 0xb6e00000, to KASAN_SHADOW_END at 0xbfffffff. - Now we can calculate the KASAN_SHADOW_OFFSET for any kernel address as 0xbf000000 needs to map to the first byte of shadow memory and 0xffffffff needs to map to the last byte of shadow memory. Since: SHADOW_ADDR = (address >> 3) + KASAN_SHADOW_OFFSET 0xb6e00000 = (0xbf000000 >> 3) + KASAN_SHADOW_OFFSET KASAN_SHADOW_OFFSET = 0xb6e00000 - (0xbf000000 >> 3) KASAN_SHADOW_OFFSET = 0xb6e00000 - 0x17e00000 KASAN_SHADOW_OFFSET = 0x9f000000 - 0x8f000000 if we have 3G userspace / 1G kernelspace with full 1 GB low memory (VMSPLIT_3G_OPT): - The kernel address space is 1GB (0x40000000) - PAGE_OFFSET is set to 0xb0000000 so the kernel static image uses 0xb0000000 .. 0xffffffff. - On top of that we have the MODULES_VADDR which under the worst case (using ARM instructions) is PAGE_OFFSET - 16M (0x01000000) = 0xaf000000 so the modules use addresses 0xaf000000 .. 0xaffffff - So the addresses 0xaf000000 .. 0xffffffff need to be covered with shadow memory. That is 0x51000000 bytes of memory. - 1/8 of that is needed for its shadow memory, so 0x0a200000 bytes of shadow memory is needed. We "steal" that from the remaining lowmem. - The KASAN_SHADOW_START becomes 0xa4e00000, to KASAN_SHADOW_END at 0xaeffffff. - Now we can calculate the KASAN_SHADOW_OFFSET for any kernel address as 0xaf000000 needs to map to the first byte of shadow memory and 0xffffffff needs to map to the last byte of shadow memory. Since: SHADOW_ADDR = (address >> 3) + KASAN_SHADOW_OFFSET 0xa4e00000 = (0xaf000000 >> 3) + KASAN_SHADOW_OFFSET KASAN_SHADOW_OFFSET = 0xa4e00000 - (0xaf000000 >> 3) KASAN_SHADOW_OFFSET = 0xa4e00000 - 0x15e00000 KASAN_SHADOW_OFFSET = 0x8f000000 - The default value of 0xffffffff for KASAN_SHADOW_OFFSET is an error value. We should always match one of the above shadow offsets. When we do this, TASK_SIZE will sometimes get a bit odd values that will not fit into immediate mov assembly instructions. To account for this, we need to rewrite some assembly using TASK_SIZE like this: - mov r1, #TASK_SIZE + ldr r1, =TASK_SIZE or - cmp r4, #TASK_SIZE + ldr r0, =TASK_SIZE + cmp r4, r0 this is done to avoid the immediate #TASK_SIZE that need to fit into a limited number of bits. Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: kasan-dev@googlegroups.com Cc: Mike Rapoport <rppt@linux.ibm.com> Reviewed-by: Ard Biesheuvel <ardb@kernel.org> Tested-by: Ard Biesheuvel <ardb@kernel.org> # QEMU/KVM/mach-virt/LPAE/8G Tested-by: Florian Fainelli <f.fainelli@gmail.com> # Brahma SoCs Tested-by: Ahmad Fatoum <a.fatoum@pengutronix.de> # i.MX6Q Reported-by: Ard Biesheuvel <ardb@kernel.org> Signed-off-by: Abbott Liu <liuwenliang@huawei.com> Signed-off-by: Florian Fainelli <f.fainelli@gmail.com> Signed-off-by: Linus Walleij <linus.walleij@linaro.org> Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
2020-10-25 23:53:46 +01:00
ARM: 9015/2: Define the virtual space of KASan's shadow region Define KASAN_SHADOW_OFFSET,KASAN_SHADOW_START and KASAN_SHADOW_END for the Arm kernel address sanitizer. We are "stealing" lowmem (the 4GB addressable by a 32bit architecture) out of the virtual address space to use as shadow memory for KASan as follows: +----+ 0xffffffff | | | | |-> Static kernel image (vmlinux) BSS and page table | |/ +----+ PAGE_OFFSET | | | | |-> Loadable kernel modules virtual address space area | |/ +----+ MODULES_VADDR = KASAN_SHADOW_END | | | | |-> The shadow area of kernel virtual address. | |/ +----+-> TASK_SIZE (start of kernel space) = KASAN_SHADOW_START the | | shadow address of MODULES_VADDR | | | | | | | | |-> The user space area in lowmem. The kernel address | | | sanitizer do not use this space, nor does it map it. | | | | | | | | | | | | | |/ ------ 0 0 .. TASK_SIZE is the memory that can be used by shared userspace/kernelspace. It us used for userspace processes and for passing parameters and memory buffers in system calls etc. We do not need to shadow this area. KASAN_SHADOW_START: This value begins with the MODULE_VADDR's shadow address. It is the start of kernel virtual space. Since we have modules to load, we need to cover also that area with shadow memory so we can find memory bugs in modules. KASAN_SHADOW_END This value is the 0x100000000's shadow address: the mapping that would be after the end of the kernel memory at 0xffffffff. It is the end of kernel address sanitizer shadow area. It is also the start of the module area. KASAN_SHADOW_OFFSET: This value is used to map an address to the corresponding shadow address by the following formula: shadow_addr = (address >> 3) + KASAN_SHADOW_OFFSET; As you would expect, >> 3 is equal to dividing by 8, meaning each byte in the shadow memory covers 8 bytes of kernel memory, so one bit shadow memory per byte of kernel memory is used. The KASAN_SHADOW_OFFSET is provided in a Kconfig option depending on the VMSPLIT layout of the system: the kernel and userspace can split up lowmem in different ways according to needs, so we calculate the shadow offset depending on this. When kasan is enabled, the definition of TASK_SIZE is not an 8-bit rotated constant, so we need to modify the TASK_SIZE access code in the *.s file. The kernel and modules may use different amounts of memory, according to the VMSPLIT configuration, which in turn determines the PAGE_OFFSET. We use the following KASAN_SHADOW_OFFSETs depending on how the virtual memory is split up: - 0x1f000000 if we have 1G userspace / 3G kernelspace split: - The kernel address space is 3G (0xc0000000) - PAGE_OFFSET is then set to 0x40000000 so the kernel static image (vmlinux) uses addresses 0x40000000 .. 0xffffffff - On top of that we have the MODULES_VADDR which under the worst case (using ARM instructions) is PAGE_OFFSET - 16M (0x01000000) = 0x3f000000 so the modules use addresses 0x3f000000 .. 0x3fffffff - So the addresses 0x3f000000 .. 0xffffffff need to be covered with shadow memory. That is 0xc1000000 bytes of memory. - 1/8 of that is needed for its shadow memory, so 0x18200000 bytes of shadow memory is needed. We "steal" that from the remaining lowmem. - The KASAN_SHADOW_START becomes 0x26e00000, to KASAN_SHADOW_END at 0x3effffff. - Now we can calculate the KASAN_SHADOW_OFFSET for any kernel address as 0x3f000000 needs to map to the first byte of shadow memory and 0xffffffff needs to map to the last byte of shadow memory. Since: SHADOW_ADDR = (address >> 3) + KASAN_SHADOW_OFFSET 0x26e00000 = (0x3f000000 >> 3) + KASAN_SHADOW_OFFSET KASAN_SHADOW_OFFSET = 0x26e00000 - (0x3f000000 >> 3) KASAN_SHADOW_OFFSET = 0x26e00000 - 0x07e00000 KASAN_SHADOW_OFFSET = 0x1f000000 - 0x5f000000 if we have 2G userspace / 2G kernelspace split: - The kernel space is 2G (0x80000000) - PAGE_OFFSET is set to 0x80000000 so the kernel static image uses 0x80000000 .. 0xffffffff. - On top of that we have the MODULES_VADDR which under the worst case (using ARM instructions) is PAGE_OFFSET - 16M (0x01000000) = 0x7f000000 so the modules use addresses 0x7f000000 .. 0x7fffffff - So the addresses 0x7f000000 .. 0xffffffff need to be covered with shadow memory. That is 0x81000000 bytes of memory. - 1/8 of that is needed for its shadow memory, so 0x10200000 bytes of shadow memory is needed. We "steal" that from the remaining lowmem. - The KASAN_SHADOW_START becomes 0x6ee00000, to KASAN_SHADOW_END at 0x7effffff. - Now we can calculate the KASAN_SHADOW_OFFSET for any kernel address as 0x7f000000 needs to map to the first byte of shadow memory and 0xffffffff needs to map to the last byte of shadow memory. Since: SHADOW_ADDR = (address >> 3) + KASAN_SHADOW_OFFSET 0x6ee00000 = (0x7f000000 >> 3) + KASAN_SHADOW_OFFSET KASAN_SHADOW_OFFSET = 0x6ee00000 - (0x7f000000 >> 3) KASAN_SHADOW_OFFSET = 0x6ee00000 - 0x0fe00000 KASAN_SHADOW_OFFSET = 0x5f000000 - 0x9f000000 if we have 3G userspace / 1G kernelspace split, and this is the default split for ARM: - The kernel address space is 1GB (0x40000000) - PAGE_OFFSET is set to 0xc0000000 so the kernel static image uses 0xc0000000 .. 0xffffffff. - On top of that we have the MODULES_VADDR which under the worst case (using ARM instructions) is PAGE_OFFSET - 16M (0x01000000) = 0xbf000000 so the modules use addresses 0xbf000000 .. 0xbfffffff - So the addresses 0xbf000000 .. 0xffffffff need to be covered with shadow memory. That is 0x41000000 bytes of memory. - 1/8 of that is needed for its shadow memory, so 0x08200000 bytes of shadow memory is needed. We "steal" that from the remaining lowmem. - The KASAN_SHADOW_START becomes 0xb6e00000, to KASAN_SHADOW_END at 0xbfffffff. - Now we can calculate the KASAN_SHADOW_OFFSET for any kernel address as 0xbf000000 needs to map to the first byte of shadow memory and 0xffffffff needs to map to the last byte of shadow memory. Since: SHADOW_ADDR = (address >> 3) + KASAN_SHADOW_OFFSET 0xb6e00000 = (0xbf000000 >> 3) + KASAN_SHADOW_OFFSET KASAN_SHADOW_OFFSET = 0xb6e00000 - (0xbf000000 >> 3) KASAN_SHADOW_OFFSET = 0xb6e00000 - 0x17e00000 KASAN_SHADOW_OFFSET = 0x9f000000 - 0x8f000000 if we have 3G userspace / 1G kernelspace with full 1 GB low memory (VMSPLIT_3G_OPT): - The kernel address space is 1GB (0x40000000) - PAGE_OFFSET is set to 0xb0000000 so the kernel static image uses 0xb0000000 .. 0xffffffff. - On top of that we have the MODULES_VADDR which under the worst case (using ARM instructions) is PAGE_OFFSET - 16M (0x01000000) = 0xaf000000 so the modules use addresses 0xaf000000 .. 0xaffffff - So the addresses 0xaf000000 .. 0xffffffff need to be covered with shadow memory. That is 0x51000000 bytes of memory. - 1/8 of that is needed for its shadow memory, so 0x0a200000 bytes of shadow memory is needed. We "steal" that from the remaining lowmem. - The KASAN_SHADOW_START becomes 0xa4e00000, to KASAN_SHADOW_END at 0xaeffffff. - Now we can calculate the KASAN_SHADOW_OFFSET for any kernel address as 0xaf000000 needs to map to the first byte of shadow memory and 0xffffffff needs to map to the last byte of shadow memory. Since: SHADOW_ADDR = (address >> 3) + KASAN_SHADOW_OFFSET 0xa4e00000 = (0xaf000000 >> 3) + KASAN_SHADOW_OFFSET KASAN_SHADOW_OFFSET = 0xa4e00000 - (0xaf000000 >> 3) KASAN_SHADOW_OFFSET = 0xa4e00000 - 0x15e00000 KASAN_SHADOW_OFFSET = 0x8f000000 - The default value of 0xffffffff for KASAN_SHADOW_OFFSET is an error value. We should always match one of the above shadow offsets. When we do this, TASK_SIZE will sometimes get a bit odd values that will not fit into immediate mov assembly instructions. To account for this, we need to rewrite some assembly using TASK_SIZE like this: - mov r1, #TASK_SIZE + ldr r1, =TASK_SIZE or - cmp r4, #TASK_SIZE + ldr r0, =TASK_SIZE + cmp r4, r0 this is done to avoid the immediate #TASK_SIZE that need to fit into a limited number of bits. Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: kasan-dev@googlegroups.com Cc: Mike Rapoport <rppt@linux.ibm.com> Reviewed-by: Ard Biesheuvel <ardb@kernel.org> Tested-by: Ard Biesheuvel <ardb@kernel.org> # QEMU/KVM/mach-virt/LPAE/8G Tested-by: Florian Fainelli <f.fainelli@gmail.com> # Brahma SoCs Tested-by: Ahmad Fatoum <a.fatoum@pengutronix.de> # i.MX6Q Reported-by: Ard Biesheuvel <ardb@kernel.org> Signed-off-by: Abbott Liu <liuwenliang@huawei.com> Signed-off-by: Florian Fainelli <f.fainelli@gmail.com> Signed-off-by: Linus Walleij <linus.walleij@linaro.org> Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
2020-10-25 23:53:46 +01:00
ARM: P2V: introduce phys_to_virt/virt_to_phys runtime patching This idea came from Nicolas, Eric Miao produced an initial version, which was then rewritten into this. Patch the physical to virtual translations at runtime. As we modify the code, this makes it incompatible with XIP kernels, but allows us to achieve this with minimal loss of performance. As many translations are of the form: physical = virtual + (PHYS_OFFSET - PAGE_OFFSET) virtual = physical - (PHYS_OFFSET - PAGE_OFFSET) we generate an 'add' instruction for __virt_to_phys(), and a 'sub' instruction for __phys_to_virt(). We calculate at run time (PHYS_OFFSET - PAGE_OFFSET) by comparing the address prior to MMU initialization with where it should be once the MMU has been initialized, and place this constant into the above add/sub instructions. Once we have (PHYS_OFFSET - PAGE_OFFSET), we can calculate the real PHYS_OFFSET as PAGE_OFFSET is a build-time constant, and save this for the C-mode PHYS_OFFSET variable definition to use. At present, we are unable to support Realview with Sparsemem enabled as this uses a complex mapping function, and MSM as this requires a constant which will not fit in our math instruction. Add a module version magic string for this feature to prevent incompatible modules being loaded. Tested-by: Tony Lindgren <tony@atomide.com> Reviewed-by: Nicolas Pitre <nicolas.pitre@linaro.org> Tested-by: Nicolas Pitre <nicolas.pitre@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2011-01-04 19:09:43 +00:00
ARM: Better virt_to_page() handling virt_to_page() is incredibly inefficient when virt-to-phys patching is enabled. This is because we end up with this calculation: page = &mem_map[asm virt_to_phys(addr) >> 12 - __pv_phys_offset >> 12] in assembly. The asm virt_to_phys() is equivalent this this operation: addr - PAGE_OFFSET + __pv_phys_offset and we can see that because this is assembly, the compiler has no chance to optimise some of that away. This should reduce down to: page = &mem_map[(addr - PAGE_OFFSET) >> 12] for the common cases. Permit the compiler to make this optimisation by giving it more of the information it needs - do this by providing a virt_to_pfn() macro. Another issue which makes this more complex is that __pv_phys_offset is a 64-bit type on all platforms. This is needlessly wasteful - if we store the physical offset as a PFN, we can save a lot of work having to deal with 64-bit values, which sometimes ends up producing incredibly horrid code: a4c: e3009000 movw r9, #0 a4c: R_ARM_MOVW_ABS_NC __pv_phys_offset a50: e3409000 movt r9, #0 ; r9 = &__pv_phys_offset a50: R_ARM_MOVT_ABS __pv_phys_offset a54: e3002000 movw r2, #0 a54: R_ARM_MOVW_ABS_NC __pv_phys_offset a58: e3402000 movt r2, #0 ; r2 = &__pv_phys_offset a58: R_ARM_MOVT_ABS __pv_phys_offset a5c: e5999004 ldr r9, [r9, #4] ; r9 = high word of __pv_phys_offset a60: e3001000 movw r1, #0 a60: R_ARM_MOVW_ABS_NC mem_map a64: e592c000 ldr ip, [r2] ; ip = low word of __pv_phys_offset Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2014-03-25 19:45:31 +00:00
ARM: P2V: introduce phys_to_virt/virt_to_phys runtime patching This idea came from Nicolas, Eric Miao produced an initial version, which was then rewritten into this. Patch the physical to virtual translations at runtime. As we modify the code, this makes it incompatible with XIP kernels, but allows us to achieve this with minimal loss of performance. As many translations are of the form: physical = virtual + (PHYS_OFFSET - PAGE_OFFSET) virtual = physical - (PHYS_OFFSET - PAGE_OFFSET) we generate an 'add' instruction for __virt_to_phys(), and a 'sub' instruction for __phys_to_virt(). We calculate at run time (PHYS_OFFSET - PAGE_OFFSET) by comparing the address prior to MMU initialization with where it should be once the MMU has been initialized, and place this constant into the above add/sub instructions. Once we have (PHYS_OFFSET - PAGE_OFFSET), we can calculate the real PHYS_OFFSET as PAGE_OFFSET is a build-time constant, and save this for the C-mode PHYS_OFFSET variable definition to use. At present, we are unable to support Realview with Sparsemem enabled as this uses a complex mapping function, and MSM as this requires a constant which will not fit in our math instruction. Add a module version magic string for this feature to prevent incompatible modules being loaded. Tested-by: Tony Lindgren <tony@atomide.com> Reviewed-by: Nicolas Pitre <nicolas.pitre@linaro.org> Tested-by: Nicolas Pitre <nicolas.pitre@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2011-01-04 19:09:43 +00:00
ARM: Better virt_to_page() handling virt_to_page() is incredibly inefficient when virt-to-phys patching is enabled. This is because we end up with this calculation: page = &mem_map[asm virt_to_phys(addr) >> 12 - __pv_phys_offset >> 12] in assembly. The asm virt_to_phys() is equivalent this this operation: addr - PAGE_OFFSET + __pv_phys_offset and we can see that because this is assembly, the compiler has no chance to optimise some of that away. This should reduce down to: page = &mem_map[(addr - PAGE_OFFSET) >> 12] for the common cases. Permit the compiler to make this optimisation by giving it more of the information it needs - do this by providing a virt_to_pfn() macro. Another issue which makes this more complex is that __pv_phys_offset is a 64-bit type on all platforms. This is needlessly wasteful - if we store the physical offset as a PFN, we can save a lot of work having to deal with 64-bit values, which sometimes ends up producing incredibly horrid code: a4c: e3009000 movw r9, #0 a4c: R_ARM_MOVW_ABS_NC __pv_phys_offset a50: e3409000 movt r9, #0 ; r9 = &__pv_phys_offset a50: R_ARM_MOVT_ABS __pv_phys_offset a54: e3002000 movw r2, #0 a54: R_ARM_MOVW_ABS_NC __pv_phys_offset a58: e3402000 movt r2, #0 ; r2 = &__pv_phys_offset a58: R_ARM_MOVT_ABS __pv_phys_offset a5c: e5999004 ldr r9, [r9, #4] ; r9 = high word of __pv_phys_offset a60: e3001000 movw r1, #0 a60: R_ARM_MOVW_ABS_NC mem_map a64: e592c000 ldr ip, [r2] ; ip = low word of __pv_phys_offset Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2014-03-25 19:45:31 +00:00
ARM: P2V: introduce phys_to_virt/virt_to_phys runtime patching This idea came from Nicolas, Eric Miao produced an initial version, which was then rewritten into this. Patch the physical to virtual translations at runtime. As we modify the code, this makes it incompatible with XIP kernels, but allows us to achieve this with minimal loss of performance. As many translations are of the form: physical = virtual + (PHYS_OFFSET - PAGE_OFFSET) virtual = physical - (PHYS_OFFSET - PAGE_OFFSET) we generate an 'add' instruction for __virt_to_phys(), and a 'sub' instruction for __phys_to_virt(). We calculate at run time (PHYS_OFFSET - PAGE_OFFSET) by comparing the address prior to MMU initialization with where it should be once the MMU has been initialized, and place this constant into the above add/sub instructions. Once we have (PHYS_OFFSET - PAGE_OFFSET), we can calculate the real PHYS_OFFSET as PAGE_OFFSET is a build-time constant, and save this for the C-mode PHYS_OFFSET variable definition to use. At present, we are unable to support Realview with Sparsemem enabled as this uses a complex mapping function, and MSM as this requires a constant which will not fit in our math instruction. Add a module version magic string for this feature to prevent incompatible modules being loaded. Tested-by: Tony Lindgren <tony@atomide.com> Reviewed-by: Nicolas Pitre <nicolas.pitre@linaro.org> Tested-by: Nicolas Pitre <nicolas.pitre@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2011-01-04 19:09:43 +00:00
ARM: p2v: reduce p2v alignment requirement to 2 MiB The ARM kernel's linear map starts at PAGE_OFFSET, which maps to a physical address (PHYS_OFFSET) that is platform specific, and is discovered at boot. Since we don't want to slow down translations between physical and virtual addresses by keeping the offset in a variable in memory, we implement this by patching the code performing the translation, and putting the offset between PAGE_OFFSET and the start of physical RAM directly into the instruction opcodes. As we only patch up to 8 bits of offset, yielding 4 GiB >> 8 == 16 MiB of granularity, we have to round up PHYS_OFFSET to the next multiple if the start of physical RAM is not a multiple of 16 MiB. This wastes some physical RAM, since the memory that was skipped will now live below PAGE_OFFSET, making it inaccessible to the kernel. We can improve this by changing the patchable sequences and the patching logic to carry more bits of offset: 11 bits gives us 4 GiB >> 11 == 2 MiB of granularity, and so we will never waste more than that amount by rounding up the physical start of DRAM to the next multiple of 2 MiB. (Note that 2 MiB granularity guarantees that the linear mapping can be created efficiently, whereas less than 2 MiB may result in the linear mapping needing another level of page tables) This helps Zhen Lei's scenario, where the start of DRAM is known to be occupied. It also helps EFI boot, which relies on the firmware's page allocator to allocate space for the decompressed kernel as low as possible. And if the KASLR patches ever land for 32-bit, it will give us 3 more bits of randomization of the placement of the kernel inside the linear region. For the ARM code path, it simply comes down to using two add/sub instructions instead of one for the carryless version, and patching each of them with the correct immediate depending on the rotation field. For the LPAE calculation, which has to deal with a carry, it patches the MOVW instruction with up to 12 bits of offset (but we only need 11 bits anyway) For the Thumb2 code path, patching more than 11 bits of displacement would be somewhat cumbersome, but the 11 bits we need fit nicely into the second word of the u16[2] opcode, so we simply update the immediate assignment and the left shift to create an addend of the right magnitude. Suggested-by: Zhen Lei <thunder.leizhen@huawei.com> Acked-by: Nicolas Pitre <nico@fluxnic.net> Acked-by: Linus Walleij <linus.walleij@linaro.org> Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
2020-09-18 11:55:42 +03:00
ARM: mm: Correct virt_to_phys patching for 64 bit physical addresses The current phys_to_virt patching mechanism works only for 32 bit physical addresses and this patch extends the idea for 64bit physical addresses. The 64bit v2p patching mechanism patches the higher 8 bits of physical address with a constant using 'mov' instruction and lower 32bits are patched using 'add'. While this is correct, in those platforms where the lowmem addressable physical memory spawns across 4GB boundary, a carry bit can be produced as a result of addition of lower 32bits. This has to be taken in to account and added in to the upper. The patched __pv_offset and va are added in lower 32bits, where __pv_offset can be in two's complement form when PA_START < VA_START and that can result in a false carry bit. e.g 1) PA = 0x80000000; VA = 0xC0000000 __pv_offset = PA - VA = 0xC0000000 (2's complement) 2) PA = 0x2 80000000; VA = 0xC000000 __pv_offset = PA - VA = 0x1 C0000000 So adding __pv_offset + VA should never result in a true overflow for (1). So in order to differentiate between a true carry, a __pv_offset is extended to 64bit and the upper 32bits will have 0xffffffff if __pv_offset is 2's complement. So 'mvn #0' is inserted instead of 'mov' while patching for the same reason. Since mov, add, sub instruction are to patched with different constants inside the same stub, the rotation field of the opcode is using to differentiate between them. So the above examples for v2p translation becomes for VA=0xC0000000, 1) PA[63:32] = 0xffffffff PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x0 80000000 2) PA[63:32] = 0x1 PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x2 80000000 The above ideas were suggested by Nicolas Pitre <nico@linaro.org> as part of the review of first and second versions of the subject patch. There is no corresponding change on the phys_to_virt() side, because computations on the upper 32-bits would be discarded anyway. Cc: Russell King <linux@arm.linux.org.uk> Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Sricharan R <r.sricharan@ti.com> Signed-off-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
2013-07-29 20:26:22 +05:30
ARM: Better virt_to_page() handling virt_to_page() is incredibly inefficient when virt-to-phys patching is enabled. This is because we end up with this calculation: page = &mem_map[asm virt_to_phys(addr) >> 12 - __pv_phys_offset >> 12] in assembly. The asm virt_to_phys() is equivalent this this operation: addr - PAGE_OFFSET + __pv_phys_offset and we can see that because this is assembly, the compiler has no chance to optimise some of that away. This should reduce down to: page = &mem_map[(addr - PAGE_OFFSET) >> 12] for the common cases. Permit the compiler to make this optimisation by giving it more of the information it needs - do this by providing a virt_to_pfn() macro. Another issue which makes this more complex is that __pv_phys_offset is a 64-bit type on all platforms. This is needlessly wasteful - if we store the physical offset as a PFN, we can save a lot of work having to deal with 64-bit values, which sometimes ends up producing incredibly horrid code: a4c: e3009000 movw r9, #0 a4c: R_ARM_MOVW_ABS_NC __pv_phys_offset a50: e3409000 movt r9, #0 ; r9 = &__pv_phys_offset a50: R_ARM_MOVT_ABS __pv_phys_offset a54: e3002000 movw r2, #0 a54: R_ARM_MOVW_ABS_NC __pv_phys_offset a58: e3402000 movt r2, #0 ; r2 = &__pv_phys_offset a58: R_ARM_MOVT_ABS __pv_phys_offset a5c: e5999004 ldr r9, [r9, #4] ; r9 = high word of __pv_phys_offset a60: e3001000 movw r1, #0 a60: R_ARM_MOVW_ABS_NC mem_map a64: e592c000 ldr ip, [r2] ; ip = low word of __pv_phys_offset Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2014-03-25 19:45:31 +00:00
ARM: mm: Correct virt_to_phys patching for 64 bit physical addresses The current phys_to_virt patching mechanism works only for 32 bit physical addresses and this patch extends the idea for 64bit physical addresses. The 64bit v2p patching mechanism patches the higher 8 bits of physical address with a constant using 'mov' instruction and lower 32bits are patched using 'add'. While this is correct, in those platforms where the lowmem addressable physical memory spawns across 4GB boundary, a carry bit can be produced as a result of addition of lower 32bits. This has to be taken in to account and added in to the upper. The patched __pv_offset and va are added in lower 32bits, where __pv_offset can be in two's complement form when PA_START < VA_START and that can result in a false carry bit. e.g 1) PA = 0x80000000; VA = 0xC0000000 __pv_offset = PA - VA = 0xC0000000 (2's complement) 2) PA = 0x2 80000000; VA = 0xC000000 __pv_offset = PA - VA = 0x1 C0000000 So adding __pv_offset + VA should never result in a true overflow for (1). So in order to differentiate between a true carry, a __pv_offset is extended to 64bit and the upper 32bits will have 0xffffffff if __pv_offset is 2's complement. So 'mvn #0' is inserted instead of 'mov' while patching for the same reason. Since mov, add, sub instruction are to patched with different constants inside the same stub, the rotation field of the opcode is using to differentiate between them. So the above examples for v2p translation becomes for VA=0xC0000000, 1) PA[63:32] = 0xffffffff PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x0 80000000 2) PA[63:32] = 0x1 PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x2 80000000 The above ideas were suggested by Nicolas Pitre <nico@linaro.org> as part of the review of first and second versions of the subject patch. There is no corresponding change on the phys_to_virt() side, because computations on the upper 32-bits would be discarded anyway. Cc: Russell King <linux@arm.linux.org.uk> Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Sricharan R <r.sricharan@ti.com> Signed-off-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
2013-07-29 20:26:22 +05:30
ARM: Better virt_to_page() handling virt_to_page() is incredibly inefficient when virt-to-phys patching is enabled. This is because we end up with this calculation: page = &mem_map[asm virt_to_phys(addr) >> 12 - __pv_phys_offset >> 12] in assembly. The asm virt_to_phys() is equivalent this this operation: addr - PAGE_OFFSET + __pv_phys_offset and we can see that because this is assembly, the compiler has no chance to optimise some of that away. This should reduce down to: page = &mem_map[(addr - PAGE_OFFSET) >> 12] for the common cases. Permit the compiler to make this optimisation by giving it more of the information it needs - do this by providing a virt_to_pfn() macro. Another issue which makes this more complex is that __pv_phys_offset is a 64-bit type on all platforms. This is needlessly wasteful - if we store the physical offset as a PFN, we can save a lot of work having to deal with 64-bit values, which sometimes ends up producing incredibly horrid code: a4c: e3009000 movw r9, #0 a4c: R_ARM_MOVW_ABS_NC __pv_phys_offset a50: e3409000 movt r9, #0 ; r9 = &__pv_phys_offset a50: R_ARM_MOVT_ABS __pv_phys_offset a54: e3002000 movw r2, #0 a54: R_ARM_MOVW_ABS_NC __pv_phys_offset a58: e3402000 movt r2, #0 ; r2 = &__pv_phys_offset a58: R_ARM_MOVT_ABS __pv_phys_offset a5c: e5999004 ldr r9, [r9, #4] ; r9 = high word of __pv_phys_offset a60: e3001000 movw r1, #0 a60: R_ARM_MOVW_ABS_NC mem_map a64: e592c000 ldr ip, [r2] ; ip = low word of __pv_phys_offset Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2014-03-25 19:45:31 +00:00
ARM: P2V: introduce phys_to_virt/virt_to_phys runtime patching This idea came from Nicolas, Eric Miao produced an initial version, which was then rewritten into this. Patch the physical to virtual translations at runtime. As we modify the code, this makes it incompatible with XIP kernels, but allows us to achieve this with minimal loss of performance. As many translations are of the form: physical = virtual + (PHYS_OFFSET - PAGE_OFFSET) virtual = physical - (PHYS_OFFSET - PAGE_OFFSET) we generate an 'add' instruction for __virt_to_phys(), and a 'sub' instruction for __phys_to_virt(). We calculate at run time (PHYS_OFFSET - PAGE_OFFSET) by comparing the address prior to MMU initialization with where it should be once the MMU has been initialized, and place this constant into the above add/sub instructions. Once we have (PHYS_OFFSET - PAGE_OFFSET), we can calculate the real PHYS_OFFSET as PAGE_OFFSET is a build-time constant, and save this for the C-mode PHYS_OFFSET variable definition to use. At present, we are unable to support Realview with Sparsemem enabled as this uses a complex mapping function, and MSM as this requires a constant which will not fit in our math instruction. Add a module version magic string for this feature to prevent incompatible modules being loaded. Tested-by: Tony Lindgren <tony@atomide.com> Reviewed-by: Nicolas Pitre <nicolas.pitre@linaro.org> Tested-by: Nicolas Pitre <nicolas.pitre@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2011-01-04 19:09:43 +00:00
ARM: p2v: reduce p2v alignment requirement to 2 MiB The ARM kernel's linear map starts at PAGE_OFFSET, which maps to a physical address (PHYS_OFFSET) that is platform specific, and is discovered at boot. Since we don't want to slow down translations between physical and virtual addresses by keeping the offset in a variable in memory, we implement this by patching the code performing the translation, and putting the offset between PAGE_OFFSET and the start of physical RAM directly into the instruction opcodes. As we only patch up to 8 bits of offset, yielding 4 GiB >> 8 == 16 MiB of granularity, we have to round up PHYS_OFFSET to the next multiple if the start of physical RAM is not a multiple of 16 MiB. This wastes some physical RAM, since the memory that was skipped will now live below PAGE_OFFSET, making it inaccessible to the kernel. We can improve this by changing the patchable sequences and the patching logic to carry more bits of offset: 11 bits gives us 4 GiB >> 11 == 2 MiB of granularity, and so we will never waste more than that amount by rounding up the physical start of DRAM to the next multiple of 2 MiB. (Note that 2 MiB granularity guarantees that the linear mapping can be created efficiently, whereas less than 2 MiB may result in the linear mapping needing another level of page tables) This helps Zhen Lei's scenario, where the start of DRAM is known to be occupied. It also helps EFI boot, which relies on the firmware's page allocator to allocate space for the decompressed kernel as low as possible. And if the KASLR patches ever land for 32-bit, it will give us 3 more bits of randomization of the placement of the kernel inside the linear region. For the ARM code path, it simply comes down to using two add/sub instructions instead of one for the carryless version, and patching each of them with the correct immediate depending on the rotation field. For the LPAE calculation, which has to deal with a carry, it patches the MOVW instruction with up to 12 bits of offset (but we only need 11 bits anyway) For the Thumb2 code path, patching more than 11 bits of displacement would be somewhat cumbersome, but the 11 bits we need fit nicely into the second word of the u16[2] opcode, so we simply update the immediate assignment and the left shift to create an addend of the right magnitude. Suggested-by: Zhen Lei <thunder.leizhen@huawei.com> Acked-by: Nicolas Pitre <nico@fluxnic.net> Acked-by: Linus Walleij <linus.walleij@linaro.org> Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
2020-09-18 11:55:42 +03:00
ARM: P2V: introduce phys_to_virt/virt_to_phys runtime patching This idea came from Nicolas, Eric Miao produced an initial version, which was then rewritten into this. Patch the physical to virtual translations at runtime. As we modify the code, this makes it incompatible with XIP kernels, but allows us to achieve this with minimal loss of performance. As many translations are of the form: physical = virtual + (PHYS_OFFSET - PAGE_OFFSET) virtual = physical - (PHYS_OFFSET - PAGE_OFFSET) we generate an 'add' instruction for __virt_to_phys(), and a 'sub' instruction for __phys_to_virt(). We calculate at run time (PHYS_OFFSET - PAGE_OFFSET) by comparing the address prior to MMU initialization with where it should be once the MMU has been initialized, and place this constant into the above add/sub instructions. Once we have (PHYS_OFFSET - PAGE_OFFSET), we can calculate the real PHYS_OFFSET as PAGE_OFFSET is a build-time constant, and save this for the C-mode PHYS_OFFSET variable definition to use. At present, we are unable to support Realview with Sparsemem enabled as this uses a complex mapping function, and MSM as this requires a constant which will not fit in our math instruction. Add a module version magic string for this feature to prevent incompatible modules being loaded. Tested-by: Tony Lindgren <tony@atomide.com> Reviewed-by: Nicolas Pitre <nicolas.pitre@linaro.org> Tested-by: Nicolas Pitre <nicolas.pitre@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2011-01-04 19:09:43 +00:00
ARM: p2v: reduce p2v alignment requirement to 2 MiB The ARM kernel's linear map starts at PAGE_OFFSET, which maps to a physical address (PHYS_OFFSET) that is platform specific, and is discovered at boot. Since we don't want to slow down translations between physical and virtual addresses by keeping the offset in a variable in memory, we implement this by patching the code performing the translation, and putting the offset between PAGE_OFFSET and the start of physical RAM directly into the instruction opcodes. As we only patch up to 8 bits of offset, yielding 4 GiB >> 8 == 16 MiB of granularity, we have to round up PHYS_OFFSET to the next multiple if the start of physical RAM is not a multiple of 16 MiB. This wastes some physical RAM, since the memory that was skipped will now live below PAGE_OFFSET, making it inaccessible to the kernel. We can improve this by changing the patchable sequences and the patching logic to carry more bits of offset: 11 bits gives us 4 GiB >> 11 == 2 MiB of granularity, and so we will never waste more than that amount by rounding up the physical start of DRAM to the next multiple of 2 MiB. (Note that 2 MiB granularity guarantees that the linear mapping can be created efficiently, whereas less than 2 MiB may result in the linear mapping needing another level of page tables) This helps Zhen Lei's scenario, where the start of DRAM is known to be occupied. It also helps EFI boot, which relies on the firmware's page allocator to allocate space for the decompressed kernel as low as possible. And if the KASLR patches ever land for 32-bit, it will give us 3 more bits of randomization of the placement of the kernel inside the linear region. For the ARM code path, it simply comes down to using two add/sub instructions instead of one for the carryless version, and patching each of them with the correct immediate depending on the rotation field. For the LPAE calculation, which has to deal with a carry, it patches the MOVW instruction with up to 12 bits of offset (but we only need 11 bits anyway) For the Thumb2 code path, patching more than 11 bits of displacement would be somewhat cumbersome, but the 11 bits we need fit nicely into the second word of the u16[2] opcode, so we simply update the immediate assignment and the left shift to create an addend of the right magnitude. Suggested-by: Zhen Lei <thunder.leizhen@huawei.com> Acked-by: Nicolas Pitre <nico@fluxnic.net> Acked-by: Linus Walleij <linus.walleij@linaro.org> Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
2020-09-18 11:55:42 +03:00
ARM: P2V: introduce phys_to_virt/virt_to_phys runtime patching This idea came from Nicolas, Eric Miao produced an initial version, which was then rewritten into this. Patch the physical to virtual translations at runtime. As we modify the code, this makes it incompatible with XIP kernels, but allows us to achieve this with minimal loss of performance. As many translations are of the form: physical = virtual + (PHYS_OFFSET - PAGE_OFFSET) virtual = physical - (PHYS_OFFSET - PAGE_OFFSET) we generate an 'add' instruction for __virt_to_phys(), and a 'sub' instruction for __phys_to_virt(). We calculate at run time (PHYS_OFFSET - PAGE_OFFSET) by comparing the address prior to MMU initialization with where it should be once the MMU has been initialized, and place this constant into the above add/sub instructions. Once we have (PHYS_OFFSET - PAGE_OFFSET), we can calculate the real PHYS_OFFSET as PAGE_OFFSET is a build-time constant, and save this for the C-mode PHYS_OFFSET variable definition to use. At present, we are unable to support Realview with Sparsemem enabled as this uses a complex mapping function, and MSM as this requires a constant which will not fit in our math instruction. Add a module version magic string for this feature to prevent incompatible modules being loaded. Tested-by: Tony Lindgren <tony@atomide.com> Reviewed-by: Nicolas Pitre <nicolas.pitre@linaro.org> Tested-by: Nicolas Pitre <nicolas.pitre@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2011-01-04 19:09:43 +00:00
ARM: p2v: reduce p2v alignment requirement to 2 MiB The ARM kernel's linear map starts at PAGE_OFFSET, which maps to a physical address (PHYS_OFFSET) that is platform specific, and is discovered at boot. Since we don't want to slow down translations between physical and virtual addresses by keeping the offset in a variable in memory, we implement this by patching the code performing the translation, and putting the offset between PAGE_OFFSET and the start of physical RAM directly into the instruction opcodes. As we only patch up to 8 bits of offset, yielding 4 GiB >> 8 == 16 MiB of granularity, we have to round up PHYS_OFFSET to the next multiple if the start of physical RAM is not a multiple of 16 MiB. This wastes some physical RAM, since the memory that was skipped will now live below PAGE_OFFSET, making it inaccessible to the kernel. We can improve this by changing the patchable sequences and the patching logic to carry more bits of offset: 11 bits gives us 4 GiB >> 11 == 2 MiB of granularity, and so we will never waste more than that amount by rounding up the physical start of DRAM to the next multiple of 2 MiB. (Note that 2 MiB granularity guarantees that the linear mapping can be created efficiently, whereas less than 2 MiB may result in the linear mapping needing another level of page tables) This helps Zhen Lei's scenario, where the start of DRAM is known to be occupied. It also helps EFI boot, which relies on the firmware's page allocator to allocate space for the decompressed kernel as low as possible. And if the KASLR patches ever land for 32-bit, it will give us 3 more bits of randomization of the placement of the kernel inside the linear region. For the ARM code path, it simply comes down to using two add/sub instructions instead of one for the carryless version, and patching each of them with the correct immediate depending on the rotation field. For the LPAE calculation, which has to deal with a carry, it patches the MOVW instruction with up to 12 bits of offset (but we only need 11 bits anyway) For the Thumb2 code path, patching more than 11 bits of displacement would be somewhat cumbersome, but the 11 bits we need fit nicely into the second word of the u16[2] opcode, so we simply update the immediate assignment and the left shift to create an addend of the right magnitude. Suggested-by: Zhen Lei <thunder.leizhen@huawei.com> Acked-by: Nicolas Pitre <nico@fluxnic.net> Acked-by: Linus Walleij <linus.walleij@linaro.org> Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
2020-09-18 11:55:42 +03:00
ARM: mm: Correct virt_to_phys patching for 64 bit physical addresses The current phys_to_virt patching mechanism works only for 32 bit physical addresses and this patch extends the idea for 64bit physical addresses. The 64bit v2p patching mechanism patches the higher 8 bits of physical address with a constant using 'mov' instruction and lower 32bits are patched using 'add'. While this is correct, in those platforms where the lowmem addressable physical memory spawns across 4GB boundary, a carry bit can be produced as a result of addition of lower 32bits. This has to be taken in to account and added in to the upper. The patched __pv_offset and va are added in lower 32bits, where __pv_offset can be in two's complement form when PA_START < VA_START and that can result in a false carry bit. e.g 1) PA = 0x80000000; VA = 0xC0000000 __pv_offset = PA - VA = 0xC0000000 (2's complement) 2) PA = 0x2 80000000; VA = 0xC000000 __pv_offset = PA - VA = 0x1 C0000000 So adding __pv_offset + VA should never result in a true overflow for (1). So in order to differentiate between a true carry, a __pv_offset is extended to 64bit and the upper 32bits will have 0xffffffff if __pv_offset is 2's complement. So 'mvn #0' is inserted instead of 'mov' while patching for the same reason. Since mov, add, sub instruction are to patched with different constants inside the same stub, the rotation field of the opcode is using to differentiate between them. So the above examples for v2p translation becomes for VA=0xC0000000, 1) PA[63:32] = 0xffffffff PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x0 80000000 2) PA[63:32] = 0x1 PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x2 80000000 The above ideas were suggested by Nicolas Pitre <nico@linaro.org> as part of the review of first and second versions of the subject patch. There is no corresponding change on the phys_to_virt() side, because computations on the upper 32-bits would be discarded anyway. Cc: Russell King <linux@arm.linux.org.uk> Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Sricharan R <r.sricharan@ti.com> Signed-off-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
2013-07-29 20:26:22 +05:30
ARM: p2v: reduce p2v alignment requirement to 2 MiB The ARM kernel's linear map starts at PAGE_OFFSET, which maps to a physical address (PHYS_OFFSET) that is platform specific, and is discovered at boot. Since we don't want to slow down translations between physical and virtual addresses by keeping the offset in a variable in memory, we implement this by patching the code performing the translation, and putting the offset between PAGE_OFFSET and the start of physical RAM directly into the instruction opcodes. As we only patch up to 8 bits of offset, yielding 4 GiB >> 8 == 16 MiB of granularity, we have to round up PHYS_OFFSET to the next multiple if the start of physical RAM is not a multiple of 16 MiB. This wastes some physical RAM, since the memory that was skipped will now live below PAGE_OFFSET, making it inaccessible to the kernel. We can improve this by changing the patchable sequences and the patching logic to carry more bits of offset: 11 bits gives us 4 GiB >> 11 == 2 MiB of granularity, and so we will never waste more than that amount by rounding up the physical start of DRAM to the next multiple of 2 MiB. (Note that 2 MiB granularity guarantees that the linear mapping can be created efficiently, whereas less than 2 MiB may result in the linear mapping needing another level of page tables) This helps Zhen Lei's scenario, where the start of DRAM is known to be occupied. It also helps EFI boot, which relies on the firmware's page allocator to allocate space for the decompressed kernel as low as possible. And if the KASLR patches ever land for 32-bit, it will give us 3 more bits of randomization of the placement of the kernel inside the linear region. For the ARM code path, it simply comes down to using two add/sub instructions instead of one for the carryless version, and patching each of them with the correct immediate depending on the rotation field. For the LPAE calculation, which has to deal with a carry, it patches the MOVW instruction with up to 12 bits of offset (but we only need 11 bits anyway) For the Thumb2 code path, patching more than 11 bits of displacement would be somewhat cumbersome, but the 11 bits we need fit nicely into the second word of the u16[2] opcode, so we simply update the immediate assignment and the left shift to create an addend of the right magnitude. Suggested-by: Zhen Lei <thunder.leizhen@huawei.com> Acked-by: Nicolas Pitre <nico@fluxnic.net> Acked-by: Linus Walleij <linus.walleij@linaro.org> Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
2020-09-18 11:55:42 +03:00
ARM: mm: Correct virt_to_phys patching for 64 bit physical addresses The current phys_to_virt patching mechanism works only for 32 bit physical addresses and this patch extends the idea for 64bit physical addresses. The 64bit v2p patching mechanism patches the higher 8 bits of physical address with a constant using 'mov' instruction and lower 32bits are patched using 'add'. While this is correct, in those platforms where the lowmem addressable physical memory spawns across 4GB boundary, a carry bit can be produced as a result of addition of lower 32bits. This has to be taken in to account and added in to the upper. The patched __pv_offset and va are added in lower 32bits, where __pv_offset can be in two's complement form when PA_START < VA_START and that can result in a false carry bit. e.g 1) PA = 0x80000000; VA = 0xC0000000 __pv_offset = PA - VA = 0xC0000000 (2's complement) 2) PA = 0x2 80000000; VA = 0xC000000 __pv_offset = PA - VA = 0x1 C0000000 So adding __pv_offset + VA should never result in a true overflow for (1). So in order to differentiate between a true carry, a __pv_offset is extended to 64bit and the upper 32bits will have 0xffffffff if __pv_offset is 2's complement. So 'mvn #0' is inserted instead of 'mov' while patching for the same reason. Since mov, add, sub instruction are to patched with different constants inside the same stub, the rotation field of the opcode is using to differentiate between them. So the above examples for v2p translation becomes for VA=0xC0000000, 1) PA[63:32] = 0xffffffff PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x0 80000000 2) PA[63:32] = 0x1 PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x2 80000000 The above ideas were suggested by Nicolas Pitre <nico@linaro.org> as part of the review of first and second versions of the subject patch. There is no corresponding change on the phys_to_virt() side, because computations on the upper 32-bits would be discarded anyway. Cc: Russell King <linux@arm.linux.org.uk> Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Sricharan R <r.sricharan@ti.com> Signed-off-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
2013-07-29 20:26:22 +05:30
ARM: mm: Correct virt_to_phys patching for 64 bit physical addresses The current phys_to_virt patching mechanism works only for 32 bit physical addresses and this patch extends the idea for 64bit physical addresses. The 64bit v2p patching mechanism patches the higher 8 bits of physical address with a constant using 'mov' instruction and lower 32bits are patched using 'add'. While this is correct, in those platforms where the lowmem addressable physical memory spawns across 4GB boundary, a carry bit can be produced as a result of addition of lower 32bits. This has to be taken in to account and added in to the upper. The patched __pv_offset and va are added in lower 32bits, where __pv_offset can be in two's complement form when PA_START < VA_START and that can result in a false carry bit. e.g 1) PA = 0x80000000; VA = 0xC0000000 __pv_offset = PA - VA = 0xC0000000 (2's complement) 2) PA = 0x2 80000000; VA = 0xC000000 __pv_offset = PA - VA = 0x1 C0000000 So adding __pv_offset + VA should never result in a true overflow for (1). So in order to differentiate between a true carry, a __pv_offset is extended to 64bit and the upper 32bits will have 0xffffffff if __pv_offset is 2's complement. So 'mvn #0' is inserted instead of 'mov' while patching for the same reason. Since mov, add, sub instruction are to patched with different constants inside the same stub, the rotation field of the opcode is using to differentiate between them. So the above examples for v2p translation becomes for VA=0xC0000000, 1) PA[63:32] = 0xffffffff PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x0 80000000 2) PA[63:32] = 0x1 PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x2 80000000 The above ideas were suggested by Nicolas Pitre <nico@linaro.org> as part of the review of first and second versions of the subject patch. There is no corresponding change on the phys_to_virt() side, because computations on the upper 32-bits would be discarded anyway. Cc: Russell King <linux@arm.linux.org.uk> Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Sricharan R <r.sricharan@ti.com> Signed-off-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
2013-07-29 20:26:22 +05:30
ARM: mm: Correct virt_to_phys patching for 64 bit physical addresses The current phys_to_virt patching mechanism works only for 32 bit physical addresses and this patch extends the idea for 64bit physical addresses. The 64bit v2p patching mechanism patches the higher 8 bits of physical address with a constant using 'mov' instruction and lower 32bits are patched using 'add'. While this is correct, in those platforms where the lowmem addressable physical memory spawns across 4GB boundary, a carry bit can be produced as a result of addition of lower 32bits. This has to be taken in to account and added in to the upper. The patched __pv_offset and va are added in lower 32bits, where __pv_offset can be in two's complement form when PA_START < VA_START and that can result in a false carry bit. e.g 1) PA = 0x80000000; VA = 0xC0000000 __pv_offset = PA - VA = 0xC0000000 (2's complement) 2) PA = 0x2 80000000; VA = 0xC000000 __pv_offset = PA - VA = 0x1 C0000000 So adding __pv_offset + VA should never result in a true overflow for (1). So in order to differentiate between a true carry, a __pv_offset is extended to 64bit and the upper 32bits will have 0xffffffff if __pv_offset is 2's complement. So 'mvn #0' is inserted instead of 'mov' while patching for the same reason. Since mov, add, sub instruction are to patched with different constants inside the same stub, the rotation field of the opcode is using to differentiate between them. So the above examples for v2p translation becomes for VA=0xC0000000, 1) PA[63:32] = 0xffffffff PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x0 80000000 2) PA[63:32] = 0x1 PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x2 80000000 The above ideas were suggested by Nicolas Pitre <nico@linaro.org> as part of the review of first and second versions of the subject patch. There is no corresponding change on the phys_to_virt() side, because computations on the upper 32-bits would be discarded anyway. Cc: Russell King <linux@arm.linux.org.uk> Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Sricharan R <r.sricharan@ti.com> Signed-off-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
2013-07-29 20:26:22 +05:30
ARM: p2v: reduce p2v alignment requirement to 2 MiB The ARM kernel's linear map starts at PAGE_OFFSET, which maps to a physical address (PHYS_OFFSET) that is platform specific, and is discovered at boot. Since we don't want to slow down translations between physical and virtual addresses by keeping the offset in a variable in memory, we implement this by patching the code performing the translation, and putting the offset between PAGE_OFFSET and the start of physical RAM directly into the instruction opcodes. As we only patch up to 8 bits of offset, yielding 4 GiB >> 8 == 16 MiB of granularity, we have to round up PHYS_OFFSET to the next multiple if the start of physical RAM is not a multiple of 16 MiB. This wastes some physical RAM, since the memory that was skipped will now live below PAGE_OFFSET, making it inaccessible to the kernel. We can improve this by changing the patchable sequences and the patching logic to carry more bits of offset: 11 bits gives us 4 GiB >> 11 == 2 MiB of granularity, and so we will never waste more than that amount by rounding up the physical start of DRAM to the next multiple of 2 MiB. (Note that 2 MiB granularity guarantees that the linear mapping can be created efficiently, whereas less than 2 MiB may result in the linear mapping needing another level of page tables) This helps Zhen Lei's scenario, where the start of DRAM is known to be occupied. It also helps EFI boot, which relies on the firmware's page allocator to allocate space for the decompressed kernel as low as possible. And if the KASLR patches ever land for 32-bit, it will give us 3 more bits of randomization of the placement of the kernel inside the linear region. For the ARM code path, it simply comes down to using two add/sub instructions instead of one for the carryless version, and patching each of them with the correct immediate depending on the rotation field. For the LPAE calculation, which has to deal with a carry, it patches the MOVW instruction with up to 12 bits of offset (but we only need 11 bits anyway) For the Thumb2 code path, patching more than 11 bits of displacement would be somewhat cumbersome, but the 11 bits we need fit nicely into the second word of the u16[2] opcode, so we simply update the immediate assignment and the left shift to create an addend of the right magnitude. Suggested-by: Zhen Lei <thunder.leizhen@huawei.com> Acked-by: Nicolas Pitre <nico@fluxnic.net> Acked-by: Linus Walleij <linus.walleij@linaro.org> Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
2020-09-18 11:55:42 +03:00
ARM: p2v: reduce p2v alignment requirement to 2 MiB The ARM kernel's linear map starts at PAGE_OFFSET, which maps to a physical address (PHYS_OFFSET) that is platform specific, and is discovered at boot. Since we don't want to slow down translations between physical and virtual addresses by keeping the offset in a variable in memory, we implement this by patching the code performing the translation, and putting the offset between PAGE_OFFSET and the start of physical RAM directly into the instruction opcodes. As we only patch up to 8 bits of offset, yielding 4 GiB >> 8 == 16 MiB of granularity, we have to round up PHYS_OFFSET to the next multiple if the start of physical RAM is not a multiple of 16 MiB. This wastes some physical RAM, since the memory that was skipped will now live below PAGE_OFFSET, making it inaccessible to the kernel. We can improve this by changing the patchable sequences and the patching logic to carry more bits of offset: 11 bits gives us 4 GiB >> 11 == 2 MiB of granularity, and so we will never waste more than that amount by rounding up the physical start of DRAM to the next multiple of 2 MiB. (Note that 2 MiB granularity guarantees that the linear mapping can be created efficiently, whereas less than 2 MiB may result in the linear mapping needing another level of page tables) This helps Zhen Lei's scenario, where the start of DRAM is known to be occupied. It also helps EFI boot, which relies on the firmware's page allocator to allocate space for the decompressed kernel as low as possible. And if the KASLR patches ever land for 32-bit, it will give us 3 more bits of randomization of the placement of the kernel inside the linear region. For the ARM code path, it simply comes down to using two add/sub instructions instead of one for the carryless version, and patching each of them with the correct immediate depending on the rotation field. For the LPAE calculation, which has to deal with a carry, it patches the MOVW instruction with up to 12 bits of offset (but we only need 11 bits anyway) For the Thumb2 code path, patching more than 11 bits of displacement would be somewhat cumbersome, but the 11 bits we need fit nicely into the second word of the u16[2] opcode, so we simply update the immediate assignment and the left shift to create an addend of the right magnitude. Suggested-by: Zhen Lei <thunder.leizhen@huawei.com> Acked-by: Nicolas Pitre <nico@fluxnic.net> Acked-by: Linus Walleij <linus.walleij@linaro.org> Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
2020-09-18 11:55:42 +03:00
ARM: P2V: introduce phys_to_virt/virt_to_phys runtime patching This idea came from Nicolas, Eric Miao produced an initial version, which was then rewritten into this. Patch the physical to virtual translations at runtime. As we modify the code, this makes it incompatible with XIP kernels, but allows us to achieve this with minimal loss of performance. As many translations are of the form: physical = virtual + (PHYS_OFFSET - PAGE_OFFSET) virtual = physical - (PHYS_OFFSET - PAGE_OFFSET) we generate an 'add' instruction for __virt_to_phys(), and a 'sub' instruction for __phys_to_virt(). We calculate at run time (PHYS_OFFSET - PAGE_OFFSET) by comparing the address prior to MMU initialization with where it should be once the MMU has been initialized, and place this constant into the above add/sub instructions. Once we have (PHYS_OFFSET - PAGE_OFFSET), we can calculate the real PHYS_OFFSET as PAGE_OFFSET is a build-time constant, and save this for the C-mode PHYS_OFFSET variable definition to use. At present, we are unable to support Realview with Sparsemem enabled as this uses a complex mapping function, and MSM as this requires a constant which will not fit in our math instruction. Add a module version magic string for this feature to prevent incompatible modules being loaded. Tested-by: Tony Lindgren <tony@atomide.com> Reviewed-by: Nicolas Pitre <nicolas.pitre@linaro.org> Tested-by: Nicolas Pitre <nicolas.pitre@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2011-01-04 19:09:43 +00:00
ARM: mm: Correct virt_to_phys patching for 64 bit physical addresses The current phys_to_virt patching mechanism works only for 32 bit physical addresses and this patch extends the idea for 64bit physical addresses. The 64bit v2p patching mechanism patches the higher 8 bits of physical address with a constant using 'mov' instruction and lower 32bits are patched using 'add'. While this is correct, in those platforms where the lowmem addressable physical memory spawns across 4GB boundary, a carry bit can be produced as a result of addition of lower 32bits. This has to be taken in to account and added in to the upper. The patched __pv_offset and va are added in lower 32bits, where __pv_offset can be in two's complement form when PA_START < VA_START and that can result in a false carry bit. e.g 1) PA = 0x80000000; VA = 0xC0000000 __pv_offset = PA - VA = 0xC0000000 (2's complement) 2) PA = 0x2 80000000; VA = 0xC000000 __pv_offset = PA - VA = 0x1 C0000000 So adding __pv_offset + VA should never result in a true overflow for (1). So in order to differentiate between a true carry, a __pv_offset is extended to 64bit and the upper 32bits will have 0xffffffff if __pv_offset is 2's complement. So 'mvn #0' is inserted instead of 'mov' while patching for the same reason. Since mov, add, sub instruction are to patched with different constants inside the same stub, the rotation field of the opcode is using to differentiate between them. So the above examples for v2p translation becomes for VA=0xC0000000, 1) PA[63:32] = 0xffffffff PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x0 80000000 2) PA[63:32] = 0x1 PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x2 80000000 The above ideas were suggested by Nicolas Pitre <nico@linaro.org> as part of the review of first and second versions of the subject patch. There is no corresponding change on the phys_to_virt() side, because computations on the upper 32-bits would be discarded anyway. Cc: Russell King <linux@arm.linux.org.uk> Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Sricharan R <r.sricharan@ti.com> Signed-off-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
2013-07-29 20:26:22 +05:30
ARM: mm: Correct virt_to_phys patching for 64 bit physical addresses The current phys_to_virt patching mechanism works only for 32 bit physical addresses and this patch extends the idea for 64bit physical addresses. The 64bit v2p patching mechanism patches the higher 8 bits of physical address with a constant using 'mov' instruction and lower 32bits are patched using 'add'. While this is correct, in those platforms where the lowmem addressable physical memory spawns across 4GB boundary, a carry bit can be produced as a result of addition of lower 32bits. This has to be taken in to account and added in to the upper. The patched __pv_offset and va are added in lower 32bits, where __pv_offset can be in two's complement form when PA_START < VA_START and that can result in a false carry bit. e.g 1) PA = 0x80000000; VA = 0xC0000000 __pv_offset = PA - VA = 0xC0000000 (2's complement) 2) PA = 0x2 80000000; VA = 0xC000000 __pv_offset = PA - VA = 0x1 C0000000 So adding __pv_offset + VA should never result in a true overflow for (1). So in order to differentiate between a true carry, a __pv_offset is extended to 64bit and the upper 32bits will have 0xffffffff if __pv_offset is 2's complement. So 'mvn #0' is inserted instead of 'mov' while patching for the same reason. Since mov, add, sub instruction are to patched with different constants inside the same stub, the rotation field of the opcode is using to differentiate between them. So the above examples for v2p translation becomes for VA=0xC0000000, 1) PA[63:32] = 0xffffffff PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x0 80000000 2) PA[63:32] = 0x1 PA[31:0] = VA + 0xC0000000 --> results in a carry PA[63:32] = PA[63:32] + carry PA[63:0] = 0x2 80000000 The above ideas were suggested by Nicolas Pitre <nico@linaro.org> as part of the review of first and second versions of the subject patch. There is no corresponding change on the phys_to_virt() side, because computations on the upper 32-bits would be discarded anyway. Cc: Russell King <linux@arm.linux.org.uk> Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Sricharan R <r.sricharan@ti.com> Signed-off-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
2013-07-29 20:26:22 +05:30
ARM: P2V: introduce phys_to_virt/virt_to_phys runtime patching This idea came from Nicolas, Eric Miao produced an initial version, which was then rewritten into this. Patch the physical to virtual translations at runtime. As we modify the code, this makes it incompatible with XIP kernels, but allows us to achieve this with minimal loss of performance. As many translations are of the form: physical = virtual + (PHYS_OFFSET - PAGE_OFFSET) virtual = physical - (PHYS_OFFSET - PAGE_OFFSET) we generate an 'add' instruction for __virt_to_phys(), and a 'sub' instruction for __phys_to_virt(). We calculate at run time (PHYS_OFFSET - PAGE_OFFSET) by comparing the address prior to MMU initialization with where it should be once the MMU has been initialized, and place this constant into the above add/sub instructions. Once we have (PHYS_OFFSET - PAGE_OFFSET), we can calculate the real PHYS_OFFSET as PAGE_OFFSET is a build-time constant, and save this for the C-mode PHYS_OFFSET variable definition to use. At present, we are unable to support Realview with Sparsemem enabled as this uses a complex mapping function, and MSM as this requires a constant which will not fit in our math instruction. Add a module version magic string for this feature to prevent incompatible modules being loaded. Tested-by: Tony Lindgren <tony@atomide.com> Reviewed-by: Nicolas Pitre <nicolas.pitre@linaro.org> Tested-by: Nicolas Pitre <nicolas.pitre@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2011-01-04 19:09:43 +00:00
ARM: P2V: introduce phys_to_virt/virt_to_phys runtime patching This idea came from Nicolas, Eric Miao produced an initial version, which was then rewritten into this. Patch the physical to virtual translations at runtime. As we modify the code, this makes it incompatible with XIP kernels, but allows us to achieve this with minimal loss of performance. As many translations are of the form: physical = virtual + (PHYS_OFFSET - PAGE_OFFSET) virtual = physical - (PHYS_OFFSET - PAGE_OFFSET) we generate an 'add' instruction for __virt_to_phys(), and a 'sub' instruction for __phys_to_virt(). We calculate at run time (PHYS_OFFSET - PAGE_OFFSET) by comparing the address prior to MMU initialization with where it should be once the MMU has been initialized, and place this constant into the above add/sub instructions. Once we have (PHYS_OFFSET - PAGE_OFFSET), we can calculate the real PHYS_OFFSET as PAGE_OFFSET is a build-time constant, and save this for the C-mode PHYS_OFFSET variable definition to use. At present, we are unable to support Realview with Sparsemem enabled as this uses a complex mapping function, and MSM as this requires a constant which will not fit in our math instruction. Add a module version magic string for this feature to prevent incompatible modules being loaded. Tested-by: Tony Lindgren <tony@atomide.com> Reviewed-by: Nicolas Pitre <nicolas.pitre@linaro.org> Tested-by: Nicolas Pitre <nicolas.pitre@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2011-01-04 19:09:43 +00:00
ARM: P2V: introduce phys_to_virt/virt_to_phys runtime patching This idea came from Nicolas, Eric Miao produced an initial version, which was then rewritten into this. Patch the physical to virtual translations at runtime. As we modify the code, this makes it incompatible with XIP kernels, but allows us to achieve this with minimal loss of performance. As many translations are of the form: physical = virtual + (PHYS_OFFSET - PAGE_OFFSET) virtual = physical - (PHYS_OFFSET - PAGE_OFFSET) we generate an 'add' instruction for __virt_to_phys(), and a 'sub' instruction for __phys_to_virt(). We calculate at run time (PHYS_OFFSET - PAGE_OFFSET) by comparing the address prior to MMU initialization with where it should be once the MMU has been initialized, and place this constant into the above add/sub instructions. Once we have (PHYS_OFFSET - PAGE_OFFSET), we can calculate the real PHYS_OFFSET as PAGE_OFFSET is a build-time constant, and save this for the C-mode PHYS_OFFSET variable definition to use. At present, we are unable to support Realview with Sparsemem enabled as this uses a complex mapping function, and MSM as this requires a constant which will not fit in our math instruction. Add a module version magic string for this feature to prevent incompatible modules being loaded. Tested-by: Tony Lindgren <tony@atomide.com> Reviewed-by: Nicolas Pitre <nicolas.pitre@linaro.org> Tested-by: Nicolas Pitre <nicolas.pitre@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2011-01-04 19:09:43 +00:00
ARM: P2V: introduce phys_to_virt/virt_to_phys runtime patching This idea came from Nicolas, Eric Miao produced an initial version, which was then rewritten into this. Patch the physical to virtual translations at runtime. As we modify the code, this makes it incompatible with XIP kernels, but allows us to achieve this with minimal loss of performance. As many translations are of the form: physical = virtual + (PHYS_OFFSET - PAGE_OFFSET) virtual = physical - (PHYS_OFFSET - PAGE_OFFSET) we generate an 'add' instruction for __virt_to_phys(), and a 'sub' instruction for __phys_to_virt(). We calculate at run time (PHYS_OFFSET - PAGE_OFFSET) by comparing the address prior to MMU initialization with where it should be once the MMU has been initialized, and place this constant into the above add/sub instructions. Once we have (PHYS_OFFSET - PAGE_OFFSET), we can calculate the real PHYS_OFFSET as PAGE_OFFSET is a build-time constant, and save this for the C-mode PHYS_OFFSET variable definition to use. At present, we are unable to support Realview with Sparsemem enabled as this uses a complex mapping function, and MSM as this requires a constant which will not fit in our math instruction. Add a module version magic string for this feature to prevent incompatible modules being loaded. Tested-by: Tony Lindgren <tony@atomide.com> Reviewed-by: Nicolas Pitre <nicolas.pitre@linaro.org> Tested-by: Nicolas Pitre <nicolas.pitre@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2011-01-04 19:09:43 +00:00
ARM: Better virt_to_page() handling virt_to_page() is incredibly inefficient when virt-to-phys patching is enabled. This is because we end up with this calculation: page = &mem_map[asm virt_to_phys(addr) >> 12 - __pv_phys_offset >> 12] in assembly. The asm virt_to_phys() is equivalent this this operation: addr - PAGE_OFFSET + __pv_phys_offset and we can see that because this is assembly, the compiler has no chance to optimise some of that away. This should reduce down to: page = &mem_map[(addr - PAGE_OFFSET) >> 12] for the common cases. Permit the compiler to make this optimisation by giving it more of the information it needs - do this by providing a virt_to_pfn() macro. Another issue which makes this more complex is that __pv_phys_offset is a 64-bit type on all platforms. This is needlessly wasteful - if we store the physical offset as a PFN, we can save a lot of work having to deal with 64-bit values, which sometimes ends up producing incredibly horrid code: a4c: e3009000 movw r9, #0 a4c: R_ARM_MOVW_ABS_NC __pv_phys_offset a50: e3409000 movt r9, #0 ; r9 = &__pv_phys_offset a50: R_ARM_MOVT_ABS __pv_phys_offset a54: e3002000 movw r2, #0 a54: R_ARM_MOVW_ABS_NC __pv_phys_offset a58: e3402000 movt r2, #0 ; r2 = &__pv_phys_offset a58: R_ARM_MOVT_ABS __pv_phys_offset a5c: e5999004 ldr r9, [r9, #4] ; r9 = high word of __pv_phys_offset a60: e3001000 movw r1, #0 a60: R_ARM_MOVW_ABS_NC mem_map a64: e592c000 ldr ip, [r2] ; ip = low word of __pv_phys_offset Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2014-03-25 19:45:31 +00:00
ARM: mm: Make virt_to_pfn() a static inline Making virt_to_pfn() a static inline taking a strongly typed (const void *) makes the contract of a passing a pointer of that type to the function explicit and exposes any misuse of the macro virt_to_pfn() acting polymorphic and accepting many types such as (void *), (unitptr_t) or (unsigned long) as arguments without warnings. Doing this is a bit intrusive: virt_to_pfn() requires PHYS_PFN_OFFSET and PAGE_SHIFT to be defined, and this is defined in <asm/page.h>, so this must be included *before* <asm/memory.h>. The use of macros were obscuring the unclear inclusion order here, as the macros would eventually be resolved, but a static inline like this cannot be compiled with unresolved macros. The naive solution to include <asm/page.h> at the top of <asm/memory.h> does not work, because <asm/memory.h> sometimes includes <asm/page.h> at the end of itself, which would create a confusing inclusion loop. So instead, take the approach to always unconditionally include <asm/page.h> at the end of <asm/memory.h> arch/arm uses <asm/memory.h> explicitly in a lot of places, however it turns out that if we just unconditionally include <asm/memory.h> into <asm/page.h> and switch all inclusions of <asm/memory.h> to <asm/page.h> instead, we enforce the right order and <asm/memory.h> will always have access to the definitions. Put an inclusion guard in place making it impossible to include <asm/memory.h> explicitly. Link: https://lore.kernel.org/linux-mm/20220701160004.2ffff4e5ab59a55499f4c736@linux-foundation.org/ Signed-off-by: Linus Walleij <linus.walleij@linaro.org>
2022-06-02 10:18:32 +02:00
ARM: Better virt_to_page() handling virt_to_page() is incredibly inefficient when virt-to-phys patching is enabled. This is because we end up with this calculation: page = &mem_map[asm virt_to_phys(addr) >> 12 - __pv_phys_offset >> 12] in assembly. The asm virt_to_phys() is equivalent this this operation: addr - PAGE_OFFSET + __pv_phys_offset and we can see that because this is assembly, the compiler has no chance to optimise some of that away. This should reduce down to: page = &mem_map[(addr - PAGE_OFFSET) >> 12] for the common cases. Permit the compiler to make this optimisation by giving it more of the information it needs - do this by providing a virt_to_pfn() macro. Another issue which makes this more complex is that __pv_phys_offset is a 64-bit type on all platforms. This is needlessly wasteful - if we store the physical offset as a PFN, we can save a lot of work having to deal with 64-bit values, which sometimes ends up producing incredibly horrid code: a4c: e3009000 movw r9, #0 a4c: R_ARM_MOVW_ABS_NC __pv_phys_offset a50: e3409000 movt r9, #0 ; r9 = &__pv_phys_offset a50: R_ARM_MOVT_ABS __pv_phys_offset a54: e3002000 movw r2, #0 a54: R_ARM_MOVW_ABS_NC __pv_phys_offset a58: e3402000 movt r2, #0 ; r2 = &__pv_phys_offset a58: R_ARM_MOVT_ABS __pv_phys_offset a5c: e5999004 ldr r9, [r9, #4] ; r9 = high word of __pv_phys_offset a60: e3001000 movw r1, #0 a60: R_ARM_MOVW_ABS_NC mem_map a64: e592c000 ldr ip, [r2] ; ip = low word of __pv_phys_offset Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2014-03-25 19:45:31 +00:00
ARM: Better virt_to_page() handling virt_to_page() is incredibly inefficient when virt-to-phys patching is enabled. This is because we end up with this calculation: page = &mem_map[asm virt_to_phys(addr) >> 12 - __pv_phys_offset >> 12] in assembly. The asm virt_to_phys() is equivalent this this operation: addr - PAGE_OFFSET + __pv_phys_offset and we can see that because this is assembly, the compiler has no chance to optimise some of that away. This should reduce down to: page = &mem_map[(addr - PAGE_OFFSET) >> 12] for the common cases. Permit the compiler to make this optimisation by giving it more of the information it needs - do this by providing a virt_to_pfn() macro. Another issue which makes this more complex is that __pv_phys_offset is a 64-bit type on all platforms. This is needlessly wasteful - if we store the physical offset as a PFN, we can save a lot of work having to deal with 64-bit values, which sometimes ends up producing incredibly horrid code: a4c: e3009000 movw r9, #0 a4c: R_ARM_MOVW_ABS_NC __pv_phys_offset a50: e3409000 movt r9, #0 ; r9 = &__pv_phys_offset a50: R_ARM_MOVT_ABS __pv_phys_offset a54: e3002000 movw r2, #0 a54: R_ARM_MOVW_ABS_NC __pv_phys_offset a58: e3402000 movt r2, #0 ; r2 = &__pv_phys_offset a58: R_ARM_MOVT_ABS __pv_phys_offset a5c: e5999004 ldr r9, [r9, #4] ; r9 = high word of __pv_phys_offset a60: e3001000 movw r1, #0 a60: R_ARM_MOVW_ABS_NC mem_map a64: e592c000 ldr ip, [r2] ; ip = low word of __pv_phys_offset Reviewed-by: Nicolas Pitre <nico@linaro.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2014-03-25 19:45:31 +00:00