diff mbox series

[v8,1/3] riscv: Introduce CONFIG_RELOCATABLE

Message ID 20230215143626.453491-2-alexghiti@rivosinc.com (mailing list archive)
State Superseded
Delegated to: Palmer Dabbelt
Headers show
Series Introduce 64b relocatable kernel | expand

Checks

Context Check Description
conchuod/cover_letter success Series has a cover letter
conchuod/tree_selection success Guessed tree name to be for-next
conchuod/fixes_present success Fixes tag not required for -next series
conchuod/maintainers_pattern success MAINTAINERS pattern errors before the patch: 13 and now 13
conchuod/verify_signedoff success Signed-off-by tag matches author and committer
conchuod/kdoc success Errors and warnings before: 0 this patch: 0
conchuod/build_rv64_clang_allmodconfig success Errors and warnings before: 2475 this patch: 0
conchuod/module_param success Was 0 now: 0
conchuod/build_rv64_gcc_allmodconfig success Errors and warnings before: 17342 this patch: 1
conchuod/alphanumeric_selects warning Out of order selects before the patch: 727 and now 741
conchuod/build_rv32_defconfig success Build OK
conchuod/dtb_warn_rv64 success Errors and warnings before: 2 this patch: 2
conchuod/header_inline success No static functions without inline keyword in header files
conchuod/checkpatch warning WARNING: Do not crash the kernel unless it is absolutely unavoidable--use WARN_ON_ONCE() plus recovery code (if feasible) instead of BUG() or variants
conchuod/source_inline success Was 0 now: 0
conchuod/build_rv64_nommu_k210_defconfig success Build OK
conchuod/verify_fixes success No Fixes tag
conchuod/build_rv64_nommu_virt_defconfig success Build OK

Commit Message

Alexandre Ghiti Feb. 15, 2023, 2:36 p.m. UTC
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(-)

Comments

Alexandre Ghiti Feb. 22, 2023, 12:29 p.m. UTC | #1
+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();
>
Björn Töpel Feb. 24, 2023, 3:58 p.m. UTC | #2
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
Alexandre Ghiti March 22, 2023, 1:38 p.m. UTC | #3
@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
Nick Desaulniers March 22, 2023, 6:25 p.m. UTC | #4
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/
Fangrui Song March 23, 2023, 8:01 p.m. UTC | #5
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
>
Alexandre Ghiti March 24, 2023, 10:34 a.m. UTC | #6
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
Andreas Schwab May 9, 2023, 7:07 p.m. UTC | #7
That does not work with UEFI booting:

Loading Linux 6.4.0-rc1-1.g668187d-default ...
Loading initial ramdisk ...
Unhandled exception: Instruction access fault
EPC: ffffffff80016d56 RA: 000000008020334e TVAL: 0000007f80016d56
EPC: ffffffff002d1d56 RA: 00000000004be34e reloc adjusted
Unhandled exception: Load access fault
EPC: 00000000fff462d4 RA: 00000000fff462d0 TVAL: ffffffff80016d56
EPC: 00000000802012d4 RA: 00000000802012d0 reloc adjusted

Code: c825 8e0d 05b3 40b4 d0ef 0636 7493 ffe4 (d783 0004)
UEFI image [0x00000000fe65e000:0x00000000fe6e3fff] '/efi\boot\bootriscv64.efi'
UEFI image [0x00000000daa82000:0x00000000dcc2afff]
Alexandre Ghiti May 9, 2023, 7:55 p.m. UTC | #8
On 5/9/23 21:07, Andreas Schwab wrote:
> That does not work with UEFI booting:
>
> Loading Linux 6.4.0-rc1-1.g668187d-default ...
> Loading initial ramdisk ...
> Unhandled exception: Instruction access fault
> EPC: ffffffff80016d56 RA: 000000008020334e TVAL: 0000007f80016d56
> EPC: ffffffff002d1d56 RA: 00000000004be34e reloc adjusted
> Unhandled exception: Load access fault
> EPC: 00000000fff462d4 RA: 00000000fff462d0 TVAL: ffffffff80016d56
> EPC: 00000000802012d4 RA: 00000000802012d0 reloc adjusted
>
> Code: c825 8e0d 05b3 40b4 d0ef 0636 7493 ffe4 (d783 0004)
> UEFI image [0x00000000fe65e000:0x00000000fe6e3fff] '/efi\boot\bootriscv64.efi'
> UEFI image [0x00000000daa82000:0x00000000dcc2afff]
>

I need more details please, as I have a UEFI bootflow and it works great 
(KASLR is based on a relocatable kernel and works fine in UEFI too).

Thanks,

Alex
Andreas Schwab May 11, 2023, 6:18 p.m. UTC | #9
On Mai 09 2023, Alexandre Ghiti wrote:

> On 5/9/23 21:07, Andreas Schwab wrote:
>> That does not work with UEFI booting:
>>
>> Loading Linux 6.4.0-rc1-1.g668187d-default ...
>> Loading initial ramdisk ...
>> Unhandled exception: Instruction access fault
>> EPC: ffffffff80016d56 RA: 000000008020334e TVAL: 0000007f80016d56
>> EPC: ffffffff002d1d56 RA: 00000000004be34e reloc adjusted
>> Unhandled exception: Load access fault
>> EPC: 00000000fff462d4 RA: 00000000fff462d0 TVAL: ffffffff80016d56
>> EPC: 00000000802012d4 RA: 00000000802012d0 reloc adjusted
>>
>> Code: c825 8e0d 05b3 40b4 d0ef 0636 7493 ffe4 (d783 0004)
>> UEFI image [0x00000000fe65e000:0x00000000fe6e3fff] '/efi\boot\bootriscv64.efi'
>> UEFI image [0x00000000daa82000:0x00000000dcc2afff]
>>
>
> I need more details please, as I have a UEFI bootflow and it works great
> (KASLR is based on a relocatable kernel and works fine in UEFI too).

It also crashes without UEFI.  Disabling CONFIG_RELOCATABLE fixes that.
This was tested on the HiFive Unmatched board.
The kernel image I tested is available from
<https://download.opensuse.org/repositories/Kernel:/HEAD/RISCV/>.  The
same kernel with CONFIG_RELOCATABLE disabled is available from
<https://download.opensuse.org/repositories/home:/Andreas_Schwab:/riscv:/kernel/standard/>.
Palmer Dabbelt May 12, 2023, 6:41 p.m. UTC | #10
On Thu, 11 May 2023 11:18:23 PDT (-0700), schwab@linux-m68k.org wrote:
> On Mai 09 2023, Alexandre Ghiti wrote:
>
>> On 5/9/23 21:07, Andreas Schwab wrote:
>>> That does not work with UEFI booting:
>>>
>>> Loading Linux 6.4.0-rc1-1.g668187d-default ...
>>> Loading initial ramdisk ...
>>> Unhandled exception: Instruction access fault
>>> EPC: ffffffff80016d56 RA: 000000008020334e TVAL: 0000007f80016d56
>>> EPC: ffffffff002d1d56 RA: 00000000004be34e reloc adjusted
>>> Unhandled exception: Load access fault
>>> EPC: 00000000fff462d4 RA: 00000000fff462d0 TVAL: ffffffff80016d56
>>> EPC: 00000000802012d4 RA: 00000000802012d0 reloc adjusted
>>>
>>> Code: c825 8e0d 05b3 40b4 d0ef 0636 7493 ffe4 (d783 0004)
>>> UEFI image [0x00000000fe65e000:0x00000000fe6e3fff] '/efi\boot\bootriscv64.efi'
>>> UEFI image [0x00000000daa82000:0x00000000dcc2afff]
>>>
>>
>> I need more details please, as I have a UEFI bootflow and it works great
>> (KASLR is based on a relocatable kernel and works fine in UEFI too).
>
> It also crashes without UEFI.  Disabling CONFIG_RELOCATABLE fixes that.
> This was tested on the HiFive Unmatched board.
> The kernel image I tested is available from
> <https://download.opensuse.org/repositories/Kernel:/HEAD/RISCV/>.  The
> same kernel with CONFIG_RELOCATABLE disabled is available from
> <https://download.opensuse.org/repositories/home:/Andreas_Schwab:/riscv:/kernel/standard/>.

Sorry I missed this earlier, there's been some other reports of boot 
failures on rc1 showing up but those were all a lot more vague.  Just 
setting CONFIG_RELOCATABLE=y doesn't manifest a boot failure on QEMU on 
my end and I don't have an UNmatched floating around.

Alex says he's going to look into it (and IIRC he has my Unmatched...).
Alexandre Ghiti May 19, 2023, 11:08 a.m. UTC | #11
On 5/11/23 20:18, Andreas Schwab wrote:
> On Mai 09 2023, Alexandre Ghiti wrote:
>
>> On 5/9/23 21:07, Andreas Schwab wrote:
>>> That does not work with UEFI booting:
>>>
>>> Loading Linux 6.4.0-rc1-1.g668187d-default ...
>>> Loading initial ramdisk ...
>>> Unhandled exception: Instruction access fault
>>> EPC: ffffffff80016d56 RA: 000000008020334e TVAL: 0000007f80016d56
>>> EPC: ffffffff002d1d56 RA: 00000000004be34e reloc adjusted
>>> Unhandled exception: Load access fault
>>> EPC: 00000000fff462d4 RA: 00000000fff462d0 TVAL: ffffffff80016d56
>>> EPC: 00000000802012d4 RA: 00000000802012d0 reloc adjusted
>>>
>>> Code: c825 8e0d 05b3 40b4 d0ef 0636 7493 ffe4 (d783 0004)
>>> UEFI image [0x00000000fe65e000:0x00000000fe6e3fff] '/efi\boot\bootriscv64.efi'
>>> UEFI image [0x00000000daa82000:0x00000000dcc2afff]
>>>
>> I need more details please, as I have a UEFI bootflow and it works great
>> (KASLR is based on a relocatable kernel and works fine in UEFI too).
> It also crashes without UEFI.  Disabling CONFIG_RELOCATABLE fixes that.
> This was tested on the HiFive Unmatched board.
> The kernel image I tested is available from
> <https://download.opensuse.org/repositories/Kernel:/HEAD/RISCV/>.  The
> same kernel with CONFIG_RELOCATABLE disabled is available from
> <https://download.opensuse.org/repositories/home:/Andreas_Schwab:/riscv:/kernel/standard/>.
>

I have tested the following patch successfully, can you give it a try 
while I make sure this is the only place I forgot to add the -fno-pie flag?

diff --git a/arch/riscv/kernel/Makefile b/arch/riscv/kernel/Makefile
index fbdccc21418a..153864e4f399 100644
--- a/arch/riscv/kernel/Makefile
+++ b/arch/riscv/kernel/Makefile
@@ -23,6 +23,10 @@ ifdef CONFIG_FTRACE
  CFLAGS_REMOVE_alternative.o = $(CC_FLAGS_FTRACE)
  CFLAGS_REMOVE_cpufeature.o = $(CC_FLAGS_FTRACE)
  endif
+ifdef CONFIG_RELOCATABLE
+CFLAGS_alternative.o += -fno-pie
+CFLAGS_cpufeature.o += -fno-pie
+endif
  ifdef CONFIG_KASAN
  KASAN_SANITIZE_alternative.o := n
  KASAN_SANITIZE_cpufeature.o := n

Thanks

Alex
Andreas Schwab May 19, 2023, 9:48 p.m. UTC | #12
On Mai 19 2023, Alexandre Ghiti wrote:

> I have tested the following patch successfully, can you give it a try
> while I make sure this is the only place I forgot to add the -fno-pie
> flag?
>
> diff --git a/arch/riscv/kernel/Makefile b/arch/riscv/kernel/Makefile
> index fbdccc21418a..153864e4f399 100644
> --- a/arch/riscv/kernel/Makefile
> +++ b/arch/riscv/kernel/Makefile
> @@ -23,6 +23,10 @@ ifdef CONFIG_FTRACE
>  CFLAGS_REMOVE_alternative.o = $(CC_FLAGS_FTRACE)
>  CFLAGS_REMOVE_cpufeature.o = $(CC_FLAGS_FTRACE)
>  endif
> +ifdef CONFIG_RELOCATABLE
> +CFLAGS_alternative.o += -fno-pie
> +CFLAGS_cpufeature.o += -fno-pie
> +endif
>  ifdef CONFIG_KASAN
>  KASAN_SANITIZE_alternative.o := n
>  KASAN_SANITIZE_cpufeature.o := n

I can confirm that this fixes the crash.
Palmer Dabbelt May 19, 2023, 9:55 p.m. UTC | #13
On Fri, 19 May 2023 14:48:59 PDT (-0700), schwab@linux-m68k.org wrote:
> On Mai 19 2023, Alexandre Ghiti wrote:
>
>> I have tested the following patch successfully, can you give it a try
>> while I make sure this is the only place I forgot to add the -fno-pie
>> flag?
>>
>> diff --git a/arch/riscv/kernel/Makefile b/arch/riscv/kernel/Makefile
>> index fbdccc21418a..153864e4f399 100644
>> --- a/arch/riscv/kernel/Makefile
>> +++ b/arch/riscv/kernel/Makefile
>> @@ -23,6 +23,10 @@ ifdef CONFIG_FTRACE
>>  CFLAGS_REMOVE_alternative.o = $(CC_FLAGS_FTRACE)
>>  CFLAGS_REMOVE_cpufeature.o = $(CC_FLAGS_FTRACE)
>>  endif
>> +ifdef CONFIG_RELOCATABLE
>> +CFLAGS_alternative.o += -fno-pie
>> +CFLAGS_cpufeature.o += -fno-pie
>> +endif
>>  ifdef CONFIG_KASAN
>>  KASAN_SANITIZE_alternative.o := n
>>  KASAN_SANITIZE_cpufeature.o := n
>
> I can confirm that this fixes the crash.

Thanks.  Alex: can you send a patch?
Alexandre Ghiti May 22, 2023, 10:48 a.m. UTC | #14
On 19/05/2023 23:55, Palmer Dabbelt wrote:
> On Fri, 19 May 2023 14:48:59 PDT (-0700), schwab@linux-m68k.org wrote:
>> On Mai 19 2023, Alexandre Ghiti wrote:
>>
>>> I have tested the following patch successfully, can you give it a try
>>> while I make sure this is the only place I forgot to add the -fno-pie
>>> flag?
>>>
>>> diff --git a/arch/riscv/kernel/Makefile b/arch/riscv/kernel/Makefile
>>> index fbdccc21418a..153864e4f399 100644
>>> --- a/arch/riscv/kernel/Makefile
>>> +++ b/arch/riscv/kernel/Makefile
>>> @@ -23,6 +23,10 @@ ifdef CONFIG_FTRACE
>>>  CFLAGS_REMOVE_alternative.o = $(CC_FLAGS_FTRACE)
>>>  CFLAGS_REMOVE_cpufeature.o = $(CC_FLAGS_FTRACE)
>>>  endif
>>> +ifdef CONFIG_RELOCATABLE
>>> +CFLAGS_alternative.o += -fno-pie
>>> +CFLAGS_cpufeature.o += -fno-pie
>>> +endif
>>>  ifdef CONFIG_KASAN
>>>  KASAN_SANITIZE_alternative.o := n
>>>  KASAN_SANITIZE_cpufeature.o := n
>>
>> I can confirm that this fixes the crash.
>
> Thanks.  Alex: can you send a patch?


I don't think this patch alone will work, all the code in early 
alternatives must be compiled with -fno-pie, but I'm a bit scared that's 
a "big" constraint. For now, I see 2 solutions:

- Document somewhere the fact that anything called from early 
alternatives must be compiled with -fno-pie
- Or relocate once with physical address, call early alternatives, and 
then do the final virtual relocation

Both options can be cumbersome in their own way, if anyone has an 
opinion, I'd be happy to discuss that :)
diff mbox series

Patch

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 = .;
+	}
+
 #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();