* Re: [PATCH v8 1/3] riscv: Introduce CONFIG_RELOCATABLE
[not found] ` <20230215143626.453491-2-alexghiti@rivosinc.com>
@ 2023-02-22 12:29 ` Alexandre Ghiti
2023-02-24 15:58 ` Björn Töpel
2023-03-22 13:38 ` Alexandre Ghiti
0 siblings, 2 replies; 6+ messages in thread
From: Alexandre Ghiti @ 2023-02-22 12:29 UTC (permalink / raw)
To: Alexandre Ghiti, Michael Ellerman, Nicholas Piggin,
Christophe Leroy, Paul Walmsley, Palmer Dabbelt, Albert Ou,
linuxppc-dev, linux-kernel, linux-riscv, nathan, linux-kbuild,
llvm, ndesaulniers, Björn Töpel
+cc linux-kbuild, llvm, Nathan, Nick
On 2/15/23 15:36, Alexandre Ghiti wrote:
> From: Alexandre Ghiti <alex@ghiti.fr>
>
> This config allows to compile 64b kernel as PIE and to relocate it at
> any virtual address at runtime: this paves the way to KASLR.
> Runtime relocation is possible since relocation metadata are embedded into
> the kernel.
>
> Note that relocating at runtime introduces an overhead even if the
> kernel is loaded at the same address it was linked at and that the compiler
> options are those used in arm64 which uses the same RELA relocation
> format.
>
> Signed-off-by: Alexandre Ghiti <alex@ghiti.fr>
> ---
> arch/riscv/Kconfig | 14 +++++++++
> arch/riscv/Makefile | 7 +++--
> arch/riscv/kernel/efi-header.S | 6 ++--
> arch/riscv/kernel/vmlinux.lds.S | 10 ++++--
> arch/riscv/mm/Makefile | 4 +++
> arch/riscv/mm/init.c | 54 ++++++++++++++++++++++++++++++++-
> 6 files changed, 87 insertions(+), 8 deletions(-)
>
> diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
> index e2b656043abf..e0ee7ce4b2e3 100644
> --- a/arch/riscv/Kconfig
> +++ b/arch/riscv/Kconfig
> @@ -544,6 +544,20 @@ config COMPAT
>
> If you want to execute 32-bit userspace applications, say Y.
>
> +config RELOCATABLE
> + bool "Build a relocatable kernel"
> + depends on MMU && 64BIT && !XIP_KERNEL
> + help
> + This builds a kernel as a Position Independent Executable (PIE),
> + which retains all relocation metadata required to relocate the
> + kernel binary at runtime to a different virtual address than the
> + address it was linked at.
> + Since RISCV uses the RELA relocation format, this requires a
> + relocation pass at runtime even if the kernel is loaded at the
> + same address it was linked at.
> +
> + If unsure, say N.
> +
> endmenu # "Kernel features"
>
> menu "Boot options"
> diff --git a/arch/riscv/Makefile b/arch/riscv/Makefile
> index 82153960ac00..97c34136b027 100644
> --- a/arch/riscv/Makefile
> +++ b/arch/riscv/Makefile
> @@ -7,9 +7,12 @@
> #
>
> OBJCOPYFLAGS := -O binary
> -LDFLAGS_vmlinux :=
> +ifeq ($(CONFIG_RELOCATABLE),y)
> + LDFLAGS_vmlinux += -shared -Bsymbolic -z notext -z norelro
> + KBUILD_CFLAGS += -fPIE
> +endif
> ifeq ($(CONFIG_DYNAMIC_FTRACE),y)
> - LDFLAGS_vmlinux := --no-relax
> + LDFLAGS_vmlinux += --no-relax
> KBUILD_CPPFLAGS += -DCC_USING_PATCHABLE_FUNCTION_ENTRY
> CC_FLAGS_FTRACE := -fpatchable-function-entry=8
> endif
> diff --git a/arch/riscv/kernel/efi-header.S b/arch/riscv/kernel/efi-header.S
> index 8e733aa48ba6..f7ee09c4f12d 100644
> --- a/arch/riscv/kernel/efi-header.S
> +++ b/arch/riscv/kernel/efi-header.S
> @@ -33,7 +33,7 @@ optional_header:
> .byte 0x02 // MajorLinkerVersion
> .byte 0x14 // MinorLinkerVersion
> .long __pecoff_text_end - efi_header_end // SizeOfCode
> - .long __pecoff_data_virt_size // SizeOfInitializedData
> + .long __pecoff_data_virt_end - __pecoff_text_end // SizeOfInitializedData
> .long 0 // SizeOfUninitializedData
> .long __efistub_efi_pe_entry - _start // AddressOfEntryPoint
> .long efi_header_end - _start // BaseOfCode
> @@ -91,9 +91,9 @@ section_table:
> IMAGE_SCN_MEM_EXECUTE // Characteristics
>
> .ascii ".data\0\0\0"
> - .long __pecoff_data_virt_size // VirtualSize
> + .long __pecoff_data_virt_end - __pecoff_text_end // VirtualSize
> .long __pecoff_text_end - _start // VirtualAddress
> - .long __pecoff_data_raw_size // SizeOfRawData
> + .long __pecoff_data_raw_end - __pecoff_text_end // SizeOfRawData
> .long __pecoff_text_end - _start // PointerToRawData
>
> .long 0 // PointerToRelocations
> diff --git a/arch/riscv/kernel/vmlinux.lds.S b/arch/riscv/kernel/vmlinux.lds.S
> index 4e6c88aa4d87..8be2de3be08c 100644
> --- a/arch/riscv/kernel/vmlinux.lds.S
> +++ b/arch/riscv/kernel/vmlinux.lds.S
> @@ -122,9 +122,15 @@ SECTIONS
> *(.sdata*)
> }
>
> + .rela.dyn : ALIGN(8) {
> + __rela_dyn_start = .;
> + *(.rela .rela*)
> + __rela_dyn_end = .;
> + }
> +
So I realized those relocations would be better in the init section so
we can get rid of them at some point. So I tried the following:
diff --git a/arch/riscv/kernel/vmlinux.lds.S
b/arch/riscv/kernel/vmlinux.lds.S
index 7ac215467fd5..6111023a89ef 100644
--- a/arch/riscv/kernel/vmlinux.lds.S
+++ b/arch/riscv/kernel/vmlinux.lds.S
@@ -93,6 +93,12 @@ SECTIONS
*(.rel.dyn*)
}
+ .rela.dyn : ALIGN(8) {
+ __rela_dyn_start = .;
+ *(.rela .rela*)
+ __rela_dyn_end = .;
+ }
+
__init_data_end = .;
. = ALIGN(8);
@@ -119,12 +125,6 @@ SECTIONS
*(.sdata*)
}
- .rela.dyn : ALIGN(8) {
- __rela_dyn_start = .;
- *(.rela .rela*)
- __rela_dyn_end = .;
- }
-
#ifdef CONFIG_EFI
.pecoff_edata_padding : { BYTE(0); . =
ALIGN(PECOFF_FILE_ALIGNMENT); }
__pecoff_data_raw_end = ABSOLUTE(.);
But then all the relocations in vmlinux end up being null:
vmlinux: file format elf64-littleriscv
$ riscv64-linux-gnu-objdump -R vmlinux
DYNAMIC RELOCATION RECORDS
OFFSET TYPE VALUE
0000000000000000 R_RISCV_NONE *ABS*
0000000000000000 R_RISCV_NONE *ABS*
....
I also noticed that re-linking vmlinux with the same command right
after works (ie, the relocations are now valid):
$ riscv64-linux-gnu-objdump -R vmlinux
vmlinux: file format elf64-littleriscv
DYNAMIC RELOCATION RECORDS
OFFSET TYPE VALUE
ffffffff82600718 R_RISCV_RELATIVE *ABS*-0x000000007d9ff8e8
ffffffff82600720 R_RISCV_RELATIVE *ABS*-0x000000007d9ff8e8
...
Below is the command used to generate this working vmlinux:
riscv64-unknown-linux-gnu-ld -melf64lriscv -z noexecstack
--no-warn-rwx-segments -shared -Bsymbolic -z notext -z norelro
--no-relax --build-id=sha1 --script=./arch/riscv/kernel/vmlinux.lds
-Map=vmlinux.map -o vmlinux --whole-archive vmlinux.a .vmlinux.export.o
init/version-timestamp.o --no-whole-archive --start-group
./drivers/firmware/efi/libstub/lib.a --end-group .tmp_vmlinux.kallsyms3.o
I tried a lot of things, but I struggle to understand, does anyone have
any idea? FYI, the same problem happens with LLVM.
Thanks,
Alex
> #ifdef CONFIG_EFI
> .pecoff_edata_padding : { BYTE(0); . = ALIGN(PECOFF_FILE_ALIGNMENT); }
> - __pecoff_data_raw_size = ABSOLUTE(. - __pecoff_text_end);
> + __pecoff_data_raw_end = ABSOLUTE(.);
> #endif
>
> /* End of data section */
> @@ -134,7 +140,7 @@ SECTIONS
>
> #ifdef CONFIG_EFI
> . = ALIGN(PECOFF_SECTION_ALIGNMENT);
> - __pecoff_data_virt_size = ABSOLUTE(. - __pecoff_text_end);
> + __pecoff_data_virt_end = ABSOLUTE(.);
> #endif
> _end = .;
>
> diff --git a/arch/riscv/mm/Makefile b/arch/riscv/mm/Makefile
> index 2ac177c05352..b85e9e82f082 100644
> --- a/arch/riscv/mm/Makefile
> +++ b/arch/riscv/mm/Makefile
> @@ -1,6 +1,10 @@
> # SPDX-License-Identifier: GPL-2.0-only
>
> CFLAGS_init.o := -mcmodel=medany
> +ifdef CONFIG_RELOCATABLE
> +CFLAGS_init.o += -fno-pie
> +endif
> +
> ifdef CONFIG_FTRACE
> CFLAGS_REMOVE_init.o = $(CC_FLAGS_FTRACE)
> CFLAGS_REMOVE_cacheflush.o = $(CC_FLAGS_FTRACE)
> diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
> index 7f01c2e56efe..3862696c2ac9 100644
> --- a/arch/riscv/mm/init.c
> +++ b/arch/riscv/mm/init.c
> @@ -20,6 +20,9 @@
> #include <linux/dma-map-ops.h>
> #include <linux/crash_dump.h>
> #include <linux/hugetlb.h>
> +#ifdef CONFIG_RELOCATABLE
> +#include <linux/elf.h>
> +#endif
>
> #include <asm/fixmap.h>
> #include <asm/tlbflush.h>
> @@ -146,7 +149,7 @@ static void __init print_vm_layout(void)
> print_ml("kasan", KASAN_SHADOW_START, KASAN_SHADOW_END);
> #endif
>
> - print_ml("kernel", (unsigned long)KERNEL_LINK_ADDR,
> + print_ml("kernel", (unsigned long)kernel_map.virt_addr,
> (unsigned long)ADDRESS_SPACE_END);
> }
> }
> @@ -854,6 +857,44 @@ static __init void set_satp_mode(uintptr_t dtb_pa)
> #error "setup_vm() is called from head.S before relocate so it should not use absolute addressing."
> #endif
>
> +#ifdef CONFIG_RELOCATABLE
> +extern unsigned long __rela_dyn_start, __rela_dyn_end;
> +
> +static void __init relocate_kernel(void)
> +{
> + Elf64_Rela *rela = (Elf64_Rela *)&__rela_dyn_start;
> + /*
> + * This holds the offset between the linked virtual address and the
> + * relocated virtual address.
> + */
> + uintptr_t reloc_offset = kernel_map.virt_addr - KERNEL_LINK_ADDR;
> + /*
> + * This holds the offset between kernel linked virtual address and
> + * physical address.
> + */
> + uintptr_t va_kernel_link_pa_offset = KERNEL_LINK_ADDR - kernel_map.phys_addr;
> +
> + for ( ; rela < (Elf64_Rela *)&__rela_dyn_end; rela++) {
> + Elf64_Addr addr = (rela->r_offset - va_kernel_link_pa_offset);
> + Elf64_Addr relocated_addr = rela->r_addend;
> +
> + if (rela->r_info != R_RISCV_RELATIVE)
> + continue;
> +
> + /*
> + * Make sure to not relocate vdso symbols like rt_sigreturn
> + * which are linked from the address 0 in vmlinux since
> + * vdso symbol addresses are actually used as an offset from
> + * mm->context.vdso in VDSO_OFFSET macro.
> + */
> + if (relocated_addr >= KERNEL_LINK_ADDR)
> + relocated_addr += reloc_offset;
> +
> + *(Elf64_Addr *)addr = relocated_addr;
> + }
> +}
> +#endif /* CONFIG_RELOCATABLE */
> +
> #ifdef CONFIG_XIP_KERNEL
> static void __init create_kernel_page_table(pgd_t *pgdir,
> __always_unused bool early)
> @@ -1039,6 +1080,17 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
> BUG_ON((kernel_map.virt_addr + kernel_map.size) > ADDRESS_SPACE_END - SZ_4K);
> #endif
>
> +#ifdef CONFIG_RELOCATABLE
> + /*
> + * Early page table uses only one PUD, which makes it possible
> + * to map PUD_SIZE aligned on PUD_SIZE: if the relocation offset
> + * makes the kernel cross over a PUD_SIZE boundary, raise a bug
> + * since a part of the kernel would not get mapped.
> + */
> + BUG_ON(PUD_SIZE - (kernel_map.virt_addr & (PUD_SIZE - 1)) < kernel_map.size);
> + relocate_kernel();
> +#endif
> +
> apply_early_boot_alternatives();
> pt_ops_set_early();
>
^ permalink raw reply related [flat|nested] 6+ messages in thread
* Re: [PATCH v8 1/3] riscv: Introduce CONFIG_RELOCATABLE
2023-02-22 12:29 ` [PATCH v8 1/3] riscv: Introduce CONFIG_RELOCATABLE Alexandre Ghiti
@ 2023-02-24 15:58 ` Björn Töpel
2023-03-22 18:25 ` Nick Desaulniers
2023-03-22 13:38 ` Alexandre Ghiti
1 sibling, 1 reply; 6+ messages in thread
From: Björn Töpel @ 2023-02-24 15:58 UTC (permalink / raw)
To: Alexandre Ghiti, Alexandre Ghiti, Michael Ellerman,
Nicholas Piggin, Christophe Leroy, Paul Walmsley, Palmer Dabbelt,
Albert Ou, linuxppc-dev, linux-kernel, linux-riscv, nathan,
linux-kbuild, llvm, ndesaulniers
Alexandre Ghiti <alex@ghiti.fr> writes:
> +cc linux-kbuild, llvm, Nathan, Nick
>
> On 2/15/23 15:36, Alexandre Ghiti wrote:
>> From: Alexandre Ghiti <alex@ghiti.fr>
>>
>> This config allows to compile 64b kernel as PIE and to relocate it at
>> any virtual address at runtime: this paves the way to KASLR.
>> Runtime relocation is possible since relocation metadata are embedded into
>> the kernel.
>>
>> Note that relocating at runtime introduces an overhead even if the
>> kernel is loaded at the same address it was linked at and that the compiler
>> options are those used in arm64 which uses the same RELA relocation
>> format.
>>
>> Signed-off-by: Alexandre Ghiti <alex@ghiti.fr>
>> ---
>> arch/riscv/Kconfig | 14 +++++++++
>> arch/riscv/Makefile | 7 +++--
>> arch/riscv/kernel/efi-header.S | 6 ++--
>> arch/riscv/kernel/vmlinux.lds.S | 10 ++++--
>> arch/riscv/mm/Makefile | 4 +++
>> arch/riscv/mm/init.c | 54 ++++++++++++++++++++++++++++++++-
>> 6 files changed, 87 insertions(+), 8 deletions(-)
>>
>> diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
>> index e2b656043abf..e0ee7ce4b2e3 100644
>> --- a/arch/riscv/Kconfig
>> +++ b/arch/riscv/Kconfig
>> @@ -544,6 +544,20 @@ config COMPAT
>>
>> If you want to execute 32-bit userspace applications, say Y.
>>
>> +config RELOCATABLE
>> + bool "Build a relocatable kernel"
>> + depends on MMU && 64BIT && !XIP_KERNEL
>> + help
>> + This builds a kernel as a Position Independent Executable (PIE),
>> + which retains all relocation metadata required to relocate the
>> + kernel binary at runtime to a different virtual address than the
>> + address it was linked at.
>> + Since RISCV uses the RELA relocation format, this requires a
>> + relocation pass at runtime even if the kernel is loaded at the
>> + same address it was linked at.
>> +
>> + If unsure, say N.
>> +
>> endmenu # "Kernel features"
>>
>> menu "Boot options"
>> diff --git a/arch/riscv/Makefile b/arch/riscv/Makefile
>> index 82153960ac00..97c34136b027 100644
>> --- a/arch/riscv/Makefile
>> +++ b/arch/riscv/Makefile
>> @@ -7,9 +7,12 @@
>> #
>>
>> OBJCOPYFLAGS := -O binary
>> -LDFLAGS_vmlinux :=
>> +ifeq ($(CONFIG_RELOCATABLE),y)
>> + LDFLAGS_vmlinux += -shared -Bsymbolic -z notext -z norelro
>> + KBUILD_CFLAGS += -fPIE
>> +endif
>> ifeq ($(CONFIG_DYNAMIC_FTRACE),y)
>> - LDFLAGS_vmlinux := --no-relax
>> + LDFLAGS_vmlinux += --no-relax
>> KBUILD_CPPFLAGS += -DCC_USING_PATCHABLE_FUNCTION_ENTRY
>> CC_FLAGS_FTRACE := -fpatchable-function-entry=8
>> endif
>> diff --git a/arch/riscv/kernel/efi-header.S b/arch/riscv/kernel/efi-header.S
>> index 8e733aa48ba6..f7ee09c4f12d 100644
>> --- a/arch/riscv/kernel/efi-header.S
>> +++ b/arch/riscv/kernel/efi-header.S
>> @@ -33,7 +33,7 @@ optional_header:
>> .byte 0x02 // MajorLinkerVersion
>> .byte 0x14 // MinorLinkerVersion
>> .long __pecoff_text_end - efi_header_end // SizeOfCode
>> - .long __pecoff_data_virt_size // SizeOfInitializedData
>> + .long __pecoff_data_virt_end - __pecoff_text_end // SizeOfInitializedData
>> .long 0 // SizeOfUninitializedData
>> .long __efistub_efi_pe_entry - _start // AddressOfEntryPoint
>> .long efi_header_end - _start // BaseOfCode
>> @@ -91,9 +91,9 @@ section_table:
>> IMAGE_SCN_MEM_EXECUTE // Characteristics
>>
>> .ascii ".data\0\0\0"
>> - .long __pecoff_data_virt_size // VirtualSize
>> + .long __pecoff_data_virt_end - __pecoff_text_end // VirtualSize
>> .long __pecoff_text_end - _start // VirtualAddress
>> - .long __pecoff_data_raw_size // SizeOfRawData
>> + .long __pecoff_data_raw_end - __pecoff_text_end // SizeOfRawData
>> .long __pecoff_text_end - _start // PointerToRawData
>>
>> .long 0 // PointerToRelocations
>> diff --git a/arch/riscv/kernel/vmlinux.lds.S b/arch/riscv/kernel/vmlinux.lds.S
>> index 4e6c88aa4d87..8be2de3be08c 100644
>> --- a/arch/riscv/kernel/vmlinux.lds.S
>> +++ b/arch/riscv/kernel/vmlinux.lds.S
>> @@ -122,9 +122,15 @@ SECTIONS
>> *(.sdata*)
>> }
>>
>> + .rela.dyn : ALIGN(8) {
>> + __rela_dyn_start = .;
>> + *(.rela .rela*)
>> + __rela_dyn_end = .;
>> + }
>> +
>
>
> So I realized those relocations would be better in the init section so
> we can get rid of them at some point. So I tried the following:
>
> diff --git a/arch/riscv/kernel/vmlinux.lds.S
> b/arch/riscv/kernel/vmlinux.lds.S
> index 7ac215467fd5..6111023a89ef 100644
> --- a/arch/riscv/kernel/vmlinux.lds.S
> +++ b/arch/riscv/kernel/vmlinux.lds.S
> @@ -93,6 +93,12 @@ SECTIONS
> *(.rel.dyn*)
> }
>
> + .rela.dyn : ALIGN(8) {
> + __rela_dyn_start = .;
> + *(.rela .rela*)
> + __rela_dyn_end = .;
> + }
> +
> __init_data_end = .;
>
> . = ALIGN(8);
> @@ -119,12 +125,6 @@ SECTIONS
> *(.sdata*)
> }
>
> - .rela.dyn : ALIGN(8) {
> - __rela_dyn_start = .;
> - *(.rela .rela*)
> - __rela_dyn_end = .;
> - }
> -
> #ifdef CONFIG_EFI
> .pecoff_edata_padding : { BYTE(0); . =
> ALIGN(PECOFF_FILE_ALIGNMENT); }
> __pecoff_data_raw_end = ABSOLUTE(.);
>
>
> But then all the relocations in vmlinux end up being null:
>
> vmlinux: file format elf64-littleriscv
>
> $ riscv64-linux-gnu-objdump -R vmlinux
>
> DYNAMIC RELOCATION RECORDS
> OFFSET TYPE VALUE
> 0000000000000000 R_RISCV_NONE *ABS*
> 0000000000000000 R_RISCV_NONE *ABS*
> ....
>
> I also noticed that re-linking vmlinux with the same command right
> after works (ie, the relocations are now valid):
>
> $ riscv64-linux-gnu-objdump -R vmlinux
>
> vmlinux: file format elf64-littleriscv
>
> DYNAMIC RELOCATION RECORDS
> OFFSET TYPE VALUE
> ffffffff82600718 R_RISCV_RELATIVE *ABS*-0x000000007d9ff8e8
> ffffffff82600720 R_RISCV_RELATIVE *ABS*-0x000000007d9ff8e8
> ...
>
> Below is the command used to generate this working vmlinux:
>
> riscv64-unknown-linux-gnu-ld -melf64lriscv -z noexecstack
> --no-warn-rwx-segments -shared -Bsymbolic -z notext -z norelro
> --no-relax --build-id=sha1 --script=./arch/riscv/kernel/vmlinux.lds
> -Map=vmlinux.map -o vmlinux --whole-archive vmlinux.a .vmlinux.export.o
> init/version-timestamp.o --no-whole-archive --start-group
> ./drivers/firmware/efi/libstub/lib.a --end-group .tmp_vmlinux.kallsyms3.o
>
> I tried a lot of things, but I struggle to understand, does anyone have
> any idea? FYI, the same problem happens with LLVM.
Don't ask me *why*, but adding --emit-relocs to your linker flags solves
"the NULL .rela.dyn" both for GCC and LLVM.
The downside is that you end up with a bunch of .rela cruft in your
vmlinux.
Björn
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v8 1/3] riscv: Introduce CONFIG_RELOCATABLE
2023-02-22 12:29 ` [PATCH v8 1/3] riscv: Introduce CONFIG_RELOCATABLE Alexandre Ghiti
2023-02-24 15:58 ` Björn Töpel
@ 2023-03-22 13:38 ` Alexandre Ghiti
1 sibling, 0 replies; 6+ messages in thread
From: Alexandre Ghiti @ 2023-03-22 13:38 UTC (permalink / raw)
To: Alexandre Ghiti, Michael Ellerman, Nicholas Piggin,
Christophe Leroy, Paul Walmsley, Palmer Dabbelt, Albert Ou,
linuxppc-dev, linux-kernel, linux-riscv, nathan, linux-kbuild,
llvm, ndesaulniers, Björn Töpel
@linux-kbuild: Does anyone has an idea to solve this?
Thanks!
On 2/22/23 13:29, Alexandre Ghiti wrote:
> +cc linux-kbuild, llvm, Nathan, Nick
>
> On 2/15/23 15:36, Alexandre Ghiti wrote:
>> From: Alexandre Ghiti <alex@ghiti.fr>
>>
>> This config allows to compile 64b kernel as PIE and to relocate it at
>> any virtual address at runtime: this paves the way to KASLR.
>> Runtime relocation is possible since relocation metadata are embedded
>> into
>> the kernel.
>>
>> Note that relocating at runtime introduces an overhead even if the
>> kernel is loaded at the same address it was linked at and that the
>> compiler
>> options are those used in arm64 which uses the same RELA relocation
>> format.
>>
>> Signed-off-by: Alexandre Ghiti <alex@ghiti.fr>
>> ---
>> arch/riscv/Kconfig | 14 +++++++++
>> arch/riscv/Makefile | 7 +++--
>> arch/riscv/kernel/efi-header.S | 6 ++--
>> arch/riscv/kernel/vmlinux.lds.S | 10 ++++--
>> arch/riscv/mm/Makefile | 4 +++
>> arch/riscv/mm/init.c | 54 ++++++++++++++++++++++++++++++++-
>> 6 files changed, 87 insertions(+), 8 deletions(-)
>>
>> diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
>> index e2b656043abf..e0ee7ce4b2e3 100644
>> --- a/arch/riscv/Kconfig
>> +++ b/arch/riscv/Kconfig
>> @@ -544,6 +544,20 @@ config COMPAT
>> If you want to execute 32-bit userspace applications, say Y.
>> +config RELOCATABLE
>> + bool "Build a relocatable kernel"
>> + depends on MMU && 64BIT && !XIP_KERNEL
>> + help
>> + This builds a kernel as a Position Independent Executable
>> (PIE),
>> + which retains all relocation metadata required to relocate
>> the
>> + kernel binary at runtime to a different virtual address
>> than the
>> + address it was linked at.
>> + Since RISCV uses the RELA relocation format, this requires a
>> + relocation pass at runtime even if the kernel is loaded at
>> the
>> + same address it was linked at.
>> +
>> + If unsure, say N.
>> +
>> endmenu # "Kernel features"
>> menu "Boot options"
>> diff --git a/arch/riscv/Makefile b/arch/riscv/Makefile
>> index 82153960ac00..97c34136b027 100644
>> --- a/arch/riscv/Makefile
>> +++ b/arch/riscv/Makefile
>> @@ -7,9 +7,12 @@
>> #
>> OBJCOPYFLAGS := -O binary
>> -LDFLAGS_vmlinux :=
>> +ifeq ($(CONFIG_RELOCATABLE),y)
>> + LDFLAGS_vmlinux += -shared -Bsymbolic -z notext -z norelro
>> + KBUILD_CFLAGS += -fPIE
>> +endif
>> ifeq ($(CONFIG_DYNAMIC_FTRACE),y)
>> - LDFLAGS_vmlinux := --no-relax
>> + LDFLAGS_vmlinux += --no-relax
>> KBUILD_CPPFLAGS += -DCC_USING_PATCHABLE_FUNCTION_ENTRY
>> CC_FLAGS_FTRACE := -fpatchable-function-entry=8
>> endif
>> diff --git a/arch/riscv/kernel/efi-header.S
>> b/arch/riscv/kernel/efi-header.S
>> index 8e733aa48ba6..f7ee09c4f12d 100644
>> --- a/arch/riscv/kernel/efi-header.S
>> +++ b/arch/riscv/kernel/efi-header.S
>> @@ -33,7 +33,7 @@ optional_header:
>> .byte 0x02 // MajorLinkerVersion
>> .byte 0x14 // MinorLinkerVersion
>> .long __pecoff_text_end - efi_header_end // SizeOfCode
>> - .long __pecoff_data_virt_size //
>> SizeOfInitializedData
>> + .long __pecoff_data_virt_end - __pecoff_text_end //
>> SizeOfInitializedData
>> .long 0 // SizeOfUninitializedData
>> .long __efistub_efi_pe_entry - _start //
>> AddressOfEntryPoint
>> .long efi_header_end - _start // BaseOfCode
>> @@ -91,9 +91,9 @@ section_table:
>> IMAGE_SCN_MEM_EXECUTE // Characteristics
>> .ascii ".data\0\0\0"
>> - .long __pecoff_data_virt_size // VirtualSize
>> + .long __pecoff_data_virt_end - __pecoff_text_end //
>> VirtualSize
>> .long __pecoff_text_end - _start // VirtualAddress
>> - .long __pecoff_data_raw_size // SizeOfRawData
>> + .long __pecoff_data_raw_end - __pecoff_text_end //
>> SizeOfRawData
>> .long __pecoff_text_end - _start // PointerToRawData
>> .long 0 // PointerToRelocations
>> diff --git a/arch/riscv/kernel/vmlinux.lds.S
>> b/arch/riscv/kernel/vmlinux.lds.S
>> index 4e6c88aa4d87..8be2de3be08c 100644
>> --- a/arch/riscv/kernel/vmlinux.lds.S
>> +++ b/arch/riscv/kernel/vmlinux.lds.S
>> @@ -122,9 +122,15 @@ SECTIONS
>> *(.sdata*)
>> }
>> + .rela.dyn : ALIGN(8) {
>> + __rela_dyn_start = .;
>> + *(.rela .rela*)
>> + __rela_dyn_end = .;
>> + }
>> +
>
>
> So I realized those relocations would be better in the init section so
> we can get rid of them at some point. So I tried the following:
>
> diff --git a/arch/riscv/kernel/vmlinux.lds.S
> b/arch/riscv/kernel/vmlinux.lds.S
> index 7ac215467fd5..6111023a89ef 100644
> --- a/arch/riscv/kernel/vmlinux.lds.S
> +++ b/arch/riscv/kernel/vmlinux.lds.S
> @@ -93,6 +93,12 @@ SECTIONS
> *(.rel.dyn*)
> }
>
> + .rela.dyn : ALIGN(8) {
> + __rela_dyn_start = .;
> + *(.rela .rela*)
> + __rela_dyn_end = .;
> + }
> +
> __init_data_end = .;
>
> . = ALIGN(8);
> @@ -119,12 +125,6 @@ SECTIONS
> *(.sdata*)
> }
>
> - .rela.dyn : ALIGN(8) {
> - __rela_dyn_start = .;
> - *(.rela .rela*)
> - __rela_dyn_end = .;
> - }
> -
> #ifdef CONFIG_EFI
> .pecoff_edata_padding : { BYTE(0); . =
> ALIGN(PECOFF_FILE_ALIGNMENT); }
> __pecoff_data_raw_end = ABSOLUTE(.);
>
>
> But then all the relocations in vmlinux end up being null:
>
> vmlinux: file format elf64-littleriscv
>
> $ riscv64-linux-gnu-objdump -R vmlinux
>
> DYNAMIC RELOCATION RECORDS
> OFFSET TYPE VALUE
> 0000000000000000 R_RISCV_NONE *ABS*
> 0000000000000000 R_RISCV_NONE *ABS*
> ....
>
> I also noticed that re-linking vmlinux with the same command right
> after works (ie, the relocations are now valid):
>
> $ riscv64-linux-gnu-objdump -R vmlinux
>
> vmlinux: file format elf64-littleriscv
>
> DYNAMIC RELOCATION RECORDS
> OFFSET TYPE VALUE
> ffffffff82600718 R_RISCV_RELATIVE *ABS*-0x000000007d9ff8e8
> ffffffff82600720 R_RISCV_RELATIVE *ABS*-0x000000007d9ff8e8
> ...
>
> Below is the command used to generate this working vmlinux:
>
> riscv64-unknown-linux-gnu-ld -melf64lriscv -z noexecstack
> --no-warn-rwx-segments -shared -Bsymbolic -z notext -z norelro
> --no-relax --build-id=sha1 --script=./arch/riscv/kernel/vmlinux.lds
> -Map=vmlinux.map -o vmlinux --whole-archive vmlinux.a
> .vmlinux.export.o init/version-timestamp.o --no-whole-archive
> --start-group ./drivers/firmware/efi/libstub/lib.a --end-group
> .tmp_vmlinux.kallsyms3.o
>
> I tried a lot of things, but I struggle to understand, does anyone
> have any idea? FYI, the same problem happens with LLVM.
>
> Thanks,
>
> Alex
>
>
>> #ifdef CONFIG_EFI
>> .pecoff_edata_padding : { BYTE(0); . =
>> ALIGN(PECOFF_FILE_ALIGNMENT); }
>> - __pecoff_data_raw_size = ABSOLUTE(. - __pecoff_text_end);
>> + __pecoff_data_raw_end = ABSOLUTE(.);
>> #endif
>> /* End of data section */
>> @@ -134,7 +140,7 @@ SECTIONS
>> #ifdef CONFIG_EFI
>> . = ALIGN(PECOFF_SECTION_ALIGNMENT);
>> - __pecoff_data_virt_size = ABSOLUTE(. - __pecoff_text_end);
>> + __pecoff_data_virt_end = ABSOLUTE(.);
>> #endif
>> _end = .;
>> diff --git a/arch/riscv/mm/Makefile b/arch/riscv/mm/Makefile
>> index 2ac177c05352..b85e9e82f082 100644
>> --- a/arch/riscv/mm/Makefile
>> +++ b/arch/riscv/mm/Makefile
>> @@ -1,6 +1,10 @@
>> # SPDX-License-Identifier: GPL-2.0-only
>> CFLAGS_init.o := -mcmodel=medany
>> +ifdef CONFIG_RELOCATABLE
>> +CFLAGS_init.o += -fno-pie
>> +endif
>> +
>> ifdef CONFIG_FTRACE
>> CFLAGS_REMOVE_init.o = $(CC_FLAGS_FTRACE)
>> CFLAGS_REMOVE_cacheflush.o = $(CC_FLAGS_FTRACE)
>> diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
>> index 7f01c2e56efe..3862696c2ac9 100644
>> --- a/arch/riscv/mm/init.c
>> +++ b/arch/riscv/mm/init.c
>> @@ -20,6 +20,9 @@
>> #include <linux/dma-map-ops.h>
>> #include <linux/crash_dump.h>
>> #include <linux/hugetlb.h>
>> +#ifdef CONFIG_RELOCATABLE
>> +#include <linux/elf.h>
>> +#endif
>> #include <asm/fixmap.h>
>> #include <asm/tlbflush.h>
>> @@ -146,7 +149,7 @@ static void __init print_vm_layout(void)
>> print_ml("kasan", KASAN_SHADOW_START, KASAN_SHADOW_END);
>> #endif
>> - print_ml("kernel", (unsigned long)KERNEL_LINK_ADDR,
>> + print_ml("kernel", (unsigned long)kernel_map.virt_addr,
>> (unsigned long)ADDRESS_SPACE_END);
>> }
>> }
>> @@ -854,6 +857,44 @@ static __init void set_satp_mode(uintptr_t dtb_pa)
>> #error "setup_vm() is called from head.S before relocate so it
>> should not use absolute addressing."
>> #endif
>> +#ifdef CONFIG_RELOCATABLE
>> +extern unsigned long __rela_dyn_start, __rela_dyn_end;
>> +
>> +static void __init relocate_kernel(void)
>> +{
>> + Elf64_Rela *rela = (Elf64_Rela *)&__rela_dyn_start;
>> + /*
>> + * This holds the offset between the linked virtual address and the
>> + * relocated virtual address.
>> + */
>> + uintptr_t reloc_offset = kernel_map.virt_addr - KERNEL_LINK_ADDR;
>> + /*
>> + * This holds the offset between kernel linked virtual address and
>> + * physical address.
>> + */
>> + uintptr_t va_kernel_link_pa_offset = KERNEL_LINK_ADDR -
>> kernel_map.phys_addr;
>> +
>> + for ( ; rela < (Elf64_Rela *)&__rela_dyn_end; rela++) {
>> + Elf64_Addr addr = (rela->r_offset - va_kernel_link_pa_offset);
>> + Elf64_Addr relocated_addr = rela->r_addend;
>> +
>> + if (rela->r_info != R_RISCV_RELATIVE)
>> + continue;
>> +
>> + /*
>> + * Make sure to not relocate vdso symbols like rt_sigreturn
>> + * which are linked from the address 0 in vmlinux since
>> + * vdso symbol addresses are actually used as an offset from
>> + * mm->context.vdso in VDSO_OFFSET macro.
>> + */
>> + if (relocated_addr >= KERNEL_LINK_ADDR)
>> + relocated_addr += reloc_offset;
>> +
>> + *(Elf64_Addr *)addr = relocated_addr;
>> + }
>> +}
>> +#endif /* CONFIG_RELOCATABLE */
>> +
>> #ifdef CONFIG_XIP_KERNEL
>> static void __init create_kernel_page_table(pgd_t *pgdir,
>> __always_unused bool early)
>> @@ -1039,6 +1080,17 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
>> BUG_ON((kernel_map.virt_addr + kernel_map.size) >
>> ADDRESS_SPACE_END - SZ_4K);
>> #endif
>> +#ifdef CONFIG_RELOCATABLE
>> + /*
>> + * Early page table uses only one PUD, which makes it possible
>> + * to map PUD_SIZE aligned on PUD_SIZE: if the relocation offset
>> + * makes the kernel cross over a PUD_SIZE boundary, raise a bug
>> + * since a part of the kernel would not get mapped.
>> + */
>> + BUG_ON(PUD_SIZE - (kernel_map.virt_addr & (PUD_SIZE - 1)) <
>> kernel_map.size);
>> + relocate_kernel();
>> +#endif
>> +
>> apply_early_boot_alternatives();
>> pt_ops_set_early();
>
> _______________________________________________
> linux-riscv mailing list
> linux-riscv@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-riscv
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v8 1/3] riscv: Introduce CONFIG_RELOCATABLE
2023-02-24 15:58 ` Björn Töpel
@ 2023-03-22 18:25 ` Nick Desaulniers
2023-03-23 20:01 ` Fangrui Song
2023-03-24 10:34 ` Alexandre Ghiti
0 siblings, 2 replies; 6+ messages in thread
From: Nick Desaulniers @ 2023-03-22 18:25 UTC (permalink / raw)
To: Björn Töpel
Cc: Alexandre Ghiti, Alexandre Ghiti, Michael Ellerman,
Nicholas Piggin, Christophe Leroy, Paul Walmsley, Palmer Dabbelt,
Albert Ou, linuxppc-dev, linux-kernel, linux-riscv, nathan,
linux-kbuild, llvm
On Fri, Feb 24, 2023 at 7:58 AM Björn Töpel <bjorn@kernel.org> wrote:
>
> Alexandre Ghiti <alex@ghiti.fr> writes:
>
> > +cc linux-kbuild, llvm, Nathan, Nick
> >
> > On 2/15/23 15:36, Alexandre Ghiti wrote:
> >> From: Alexandre Ghiti <alex@ghiti.fr>
> >>
> > I tried a lot of things, but I struggle to understand, does anyone have
> > any idea? FYI, the same problem happens with LLVM.
Off the top of my head, no idea.
(Maybe as a follow up to this series, I wonder if pursuing
ARCH_HAS_RELR for ARCH=riscv is worthwhile?)
>
> Don't ask me *why*, but adding --emit-relocs to your linker flags solves
> "the NULL .rela.dyn" both for GCC and LLVM.
>
> The downside is that you end up with a bunch of .rela cruft in your
> vmlinux.
There was a patch just this week to use $(OBJCOPY) to strip these from
vmlinux (for x86). Looks like x86 uses --emit-relocs for KASLR:
https://lore.kernel.org/lkml/20230320121006.4863-1-petr.pavlu@suse.com/
--
Thanks,
~Nick Desaulniers
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v8 1/3] riscv: Introduce CONFIG_RELOCATABLE
2023-03-22 18:25 ` Nick Desaulniers
@ 2023-03-23 20:01 ` Fangrui Song
2023-03-24 10:34 ` Alexandre Ghiti
1 sibling, 0 replies; 6+ messages in thread
From: Fangrui Song @ 2023-03-23 20:01 UTC (permalink / raw)
To: Nick Desaulniers
Cc: Björn Töpel, Alexandre Ghiti, Alexandre Ghiti,
Michael Ellerman, Nicholas Piggin, Christophe Leroy,
Paul Walmsley, Palmer Dabbelt, Albert Ou, linuxppc-dev,
linux-kernel, linux-riscv, nathan, linux-kbuild, llvm
On Wed, Mar 22, 2023 at 11:26 AM Nick Desaulniers
<ndesaulniers@google.com> wrote:
>
> On Fri, Feb 24, 2023 at 7:58 AM Björn Töpel <bjorn@kernel.org> wrote:
> >
> > Alexandre Ghiti <alex@ghiti.fr> writes:
> >
> > > +cc linux-kbuild, llvm, Nathan, Nick
> > >
> > > On 2/15/23 15:36, Alexandre Ghiti wrote:
> > >> From: Alexandre Ghiti <alex@ghiti.fr>
> > >>
> > > I tried a lot of things, but I struggle to understand, does anyone have
> > > any idea? FYI, the same problem happens with LLVM.
>
> Off the top of my head, no idea.
>
> (Maybe as a follow up to this series, I wonder if pursuing
> ARCH_HAS_RELR for ARCH=riscv is worthwhile?)
(I had thought about this for my own fun, but the currently only
implementation arch/arm64/kernel/head.S uses assembly.
Every port needs to write some assembly for the same task, which is a pity.
In FreeBSD rtld, glibc, and musl, DT_RELR code is target-independent.)
> >
> > Don't ask me *why*, but adding --emit-relocs to your linker flags solves
> > "the NULL .rela.dyn" both for GCC and LLVM.
> >
> > The downside is that you end up with a bunch of .rela cruft in your
> > vmlinux.
>
> There was a patch just this week to use $(OBJCOPY) to strip these from
> vmlinux (for x86). Looks like x86 uses --emit-relocs for KASLR:
> https://lore.kernel.org/lkml/20230320121006.4863-1-petr.pavlu@suse.com/
> --
> Thanks,
> ~Nick Desaulniers
>
--
宋方睿
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v8 1/3] riscv: Introduce CONFIG_RELOCATABLE
2023-03-22 18:25 ` Nick Desaulniers
2023-03-23 20:01 ` Fangrui Song
@ 2023-03-24 10:34 ` Alexandre Ghiti
1 sibling, 0 replies; 6+ messages in thread
From: Alexandre Ghiti @ 2023-03-24 10:34 UTC (permalink / raw)
To: Nick Desaulniers, Björn Töpel
Cc: Alexandre Ghiti, Michael Ellerman, Nicholas Piggin,
Christophe Leroy, Paul Walmsley, Palmer Dabbelt, Albert Ou,
linuxppc-dev, linux-kernel, linux-riscv, nathan, linux-kbuild,
llvm
Hi Nick,
On 3/22/23 19:25, Nick Desaulniers wrote:
> On Fri, Feb 24, 2023 at 7:58 AM Björn Töpel <bjorn@kernel.org> wrote:
>> Alexandre Ghiti <alex@ghiti.fr> writes:
>>
>>> +cc linux-kbuild, llvm, Nathan, Nick
>>>
>>> On 2/15/23 15:36, Alexandre Ghiti wrote:
>>>> From: Alexandre Ghiti <alex@ghiti.fr>
>>>>
>>> I tried a lot of things, but I struggle to understand, does anyone have
>>> any idea? FYI, the same problem happens with LLVM.
> Off the top of my head, no idea.
>
> (Maybe as a follow up to this series, I wonder if pursuing
> ARCH_HAS_RELR for ARCH=riscv is worthwhile?)
IIUC, the goal for using RELR is to reduce the size of a kernel image:
right now, this is not my priority, but I'll add that to my todo list
because that may be useful to distros.
>
>> Don't ask me *why*, but adding --emit-relocs to your linker flags solves
>> "the NULL .rela.dyn" both for GCC and LLVM.
>>
>> The downside is that you end up with a bunch of .rela cruft in your
>> vmlinux.
> There was a patch just this week to use $(OBJCOPY) to strip these from
> vmlinux (for x86). Looks like x86 uses --emit-relocs for KASLR:
> https://lore.kernel.org/lkml/20230320121006.4863-1-petr.pavlu@suse.com/
That's nice, that would be an interesting intermediate step until we
find the issue here as I believe it is important to have the relocations
in the init section to save memory.
Thanks for your answer Nick, really appreciated,
Alex
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2023-03-24 10:34 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
[not found] <20230215143626.453491-1-alexghiti@rivosinc.com>
[not found] ` <20230215143626.453491-2-alexghiti@rivosinc.com>
2023-02-22 12:29 ` [PATCH v8 1/3] riscv: Introduce CONFIG_RELOCATABLE Alexandre Ghiti
2023-02-24 15:58 ` Björn Töpel
2023-03-22 18:25 ` Nick Desaulniers
2023-03-23 20:01 ` Fangrui Song
2023-03-24 10:34 ` Alexandre Ghiti
2023-03-22 13:38 ` Alexandre Ghiti
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).