Skip to contentSkip to conversation

riscv: Introduce CONFIG_RELOCATABLE

PATCH Superseded Alexandre Ghiti 3 participants 8 messages Jan 23, 2020 – Feb 4, 2020 RISC-V

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Alexandre Ghiti

author reviewer
#0 lore raw

PATCH 1/1 riscv: Introduce CONFIG_RELOCATABLE

This config allows to compile the kernel as PIE and to relocate it at any virtual address at runtime: this paves the way to KASLR and to 4-level page table folding at runtime. 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.

---
 arch/riscv/Kconfig              | 11 ++++
 arch/riscv/Makefile             |  5 +-
 arch/riscv/boot/loader.lds.S    |  2 +-
 arch/riscv/include/asm/page.h   |  5 +-
 arch/riscv/kernel/head.S        |  3 +-
 arch/riscv/kernel/vmlinux.lds.S | 10 ++--
 arch/riscv/mm/Makefile          |  4 ++
 arch/riscv/mm/init.c            | 92 ++++++++++++++++++++++++++++-----
 8 files changed, 110 insertions(+), 22 deletions(-)
arch/riscv/Kconfig +11 -0
index fa7dc03459e7..c652b4b850ce 100644
@@ -163,6 +163,17 @@ config PGTABLE_LEVELS
 	default 3 if 64BIT
 	default 2
 
+config RELOCATABLE
+	bool
+	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.
+
 source "arch/riscv/Kconfig.socs"
 
 menu "Platform type"
arch/riscv/Makefile +4 -1
index b9009a2fbaf5..5a115cf6a9c1 100644
@@ -9,7 +9,10 @@
 #
 
 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
 endif
arch/riscv/boot/loader.lds.S +1 -1
index 47a5003c2e28..a9ed218171aa 100644
@@ -7,7 +7,7 @@ ENTRY(_start)
 
 SECTIONS
 {
-	. = PAGE_OFFSET;
+	. = CONFIG_PAGE_OFFSET;
 
 	.payload : {
 		*(.payload)
arch/riscv/include/asm/page.h +3 -2
index ac699246ae7e..27c95da68ecb 100644
@@ -31,9 +31,9 @@
  * When not using MMU this corresponds to the first free page in
  * physical memory (aligned on a page boundary).
  */
-#define PAGE_OFFSET		_AC(CONFIG_PAGE_OFFSET, UL)
+#define PAGE_OFFSET		kernel_load_addr
 
-#define KERN_VIRT_SIZE (-PAGE_OFFSET)
+#define KERN_VIRT_SIZE		(-_AC(CONFIG_PAGE_OFFSET, UL))
 
 #ifndef __ASSEMBLY__
 
@@ -97,6 +97,7 @@ extern unsigned long pfn_base;
 #define ARCH_PFN_OFFSET		(PAGE_OFFSET >> PAGE_SHIFT)
 #endif /* CONFIG_MMU */
 
+extern unsigned long kernel_load_addr;
 extern unsigned long max_low_pfn;
 extern unsigned long min_low_pfn;
 
arch/riscv/kernel/head.S +2 -1
index 2227db63f895..5042b2b48a06 100644
@@ -126,7 +126,8 @@ clear_bss_done:
 #ifdef CONFIG_MMU
 relocate:
 	/* Relocate return address */
-	li a1, PAGE_OFFSET
+	la a1, kernel_load_addr
+	REG_L a1, 0(a1)
 	la a2, _start
 	sub a1, a1, a2
 	add ra, ra, a1
arch/riscv/kernel/vmlinux.lds.S +6 -4
index 12f42f96d46e..5095aee7c37e 100644
@@ -4,7 +4,7 @@
  * Copyright (C) 2017 SiFive
  */
 
-#define LOAD_OFFSET PAGE_OFFSET
+#define LOAD_OFFSET CONFIG_PAGE_OFFSET
 #include <asm/vmlinux.lds.h>
 #include <asm/page.h>
 #include <asm/cache.h>
@@ -70,9 +70,11 @@ SECTIONS
 
 	EXCEPTION_TABLE(0x10)
 
-	.rel.dyn : {
-		*(.rel.dyn*)
-	}
+        .rela.dyn : ALIGN(8) {
+		__rela_dyn_start = .;
+                *(.rela .rela*)
+		__rela_dyn_end = .;
+        }
 
 	_end = .;
 
arch/riscv/mm/Makefile +4 -0
index a1bd95c8047a..dcd3d806243f 100644
@@ -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 = -pg
 endif
arch/riscv/mm/init.c +79 -13

159 lines not on this page · raw message

Signature · 1 line
2.20.1
#1 lore raw
Quoted 31 lines · Alexandre Ghiti

This config allows to compile the kernel as PIE and to relocate it at any virtual address at runtime: this paves the way to KASLR and to 4-level page table folding at runtime. 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.

---
 arch/riscv/Kconfig              | 11 ++++
 arch/riscv/Makefile             |  5 +-
 arch/riscv/boot/loader.lds.S    |  2 +-
 arch/riscv/include/asm/page.h   |  5 +-
 arch/riscv/kernel/head.S        |  3 +-
 arch/riscv/kernel/vmlinux.lds.S | 10 ++--
 arch/riscv/mm/Makefile          |  4 ++
 arch/riscv/mm/init.c            | 92 ++++++++++++++++++++++++++++-----
 8 files changed, 110 insertions(+), 22 deletions(-)
arch/riscv/Kconfig +1 -0
index fa7dc03459e7..c652b4b850ce 100644
@@ -163,6 +163,17 @@ config PGTABLE_LEVELS
        default 3 if 64BIT
        default 2
 
+config RELOCATABLE

arch/riscv/Kconfig config PGTABLE_LEVELS · line 176

Quoted 10 lines
arch/riscv/Kconfig +10 -0
+       bool
+       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.
+

I am sure this cannot be used for NOMMU support because of the page table mappings.

I guess this should depend on MMU.

Quoted 24 lines
arch/riscv/Kconfig +0 -0
 source "arch/riscv/Kconfig.socs"
 
 menu "Platform type"
arch/riscv/Makefile +4 -1
index b9009a2fbaf5..5a115cf6a9c1 100644
@@ -9,7 +9,10 @@
 #
 
 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
 endif
arch/riscv/boot/loader.lds.S +0 -0
index 47a5003c2e28..a9ed218171aa 100644

arch/riscv/boot/loader.lds.S ENTRY(_start) · line 13

Quoted 8 lines
arch/riscv/boot/loader.lds.S +1 -1
@@ -7,7 +7,7 @@ ENTRY(_start)
 
 SECTIONS
 {
-       . = PAGE_OFFSET;
+       . = CONFIG_PAGE_OFFSET;
 
        .payload : {
                *(.payload)

The loader.lds.S is used by NOMMU support so this should be separate patch.

Quoted 17 lines
arch/riscv/include/asm/page.h +2 -2
index ac699246ae7e..27c95da68ecb 100644
@@ -31,9 +31,9 @@
  * When not using MMU this corresponds to the first free page in
  * physical memory (aligned on a page boundary).
  */
-#define PAGE_OFFSET            _AC(CONFIG_PAGE_OFFSET, UL)
+#define PAGE_OFFSET            kernel_load_addr
 
-#define KERN_VIRT_SIZE (-PAGE_OFFSET)
+#define KERN_VIRT_SIZE         (-_AC(CONFIG_PAGE_OFFSET, UL))
 
 #ifndef __ASSEMBLY__
 

arch/riscv/include/asm/page.h extern unsigned long pfn_base; · line 100

Quoted 4 lines
arch/riscv/include/asm/page.h +1 -0
@@ -97,6 +97,7 @@ extern unsigned long pfn_base;
 #define ARCH_PFN_OFFSET                (PAGE_OFFSET >> PAGE_SHIFT)
 #endif /* CONFIG_MMU */
 
+extern unsigned long kernel_load_addr;

Having this variable named kernel_virt_addr seems more logical to me.

Is it mandatory to use kernel_load_addr name ?

Quoted 234 lines
arch/riscv/include/asm/page.h +0 -0
 extern unsigned long max_low_pfn;
 extern unsigned long min_low_pfn;
 
arch/riscv/kernel/head.S +2 -1
index 2227db63f895..5042b2b48a06 100644
@@ -126,7 +126,8 @@ clear_bss_done:
 #ifdef CONFIG_MMU
 relocate:
        /* Relocate return address */
-       li a1, PAGE_OFFSET
+       la a1, kernel_load_addr
+       REG_L a1, 0(a1)
        la a2, _start
        sub a1, a1, a2
        add ra, ra, a1
arch/riscv/kernel/vmlinux.lds.S +6 -4
index 12f42f96d46e..5095aee7c37e 100644
@@ -4,7 +4,7 @@
  * Copyright (C) 2017 SiFive
  */
 
-#define LOAD_OFFSET PAGE_OFFSET
+#define LOAD_OFFSET CONFIG_PAGE_OFFSET
 #include <asm/vmlinux.lds.h>
 #include <asm/page.h>
 #include <asm/cache.h>
@@ -70,9 +70,11 @@ SECTIONS
 
        EXCEPTION_TABLE(0x10)
 
-       .rel.dyn : {
-               *(.rel.dyn*)
-       }
+        .rela.dyn : ALIGN(8) {
+               __rela_dyn_start = .;
+                *(.rela .rela*)
+               __rela_dyn_end = .;
+        }
 
        _end = .;
 
arch/riscv/mm/Makefile +4 -0
index a1bd95c8047a..dcd3d806243f 100644
@@ -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 = -pg
 endif
arch/riscv/mm/init.c +79 -14
index 965a8cf4829c..ac9a9f69abc0 100644
@@ -12,6 +12,9 @@
 #include <linux/sizes.h>
 #include <linux/of_fdt.h>
 #include <linux/libfdt.h>
+#ifdef CONFIG_RELOCATABLE
+#include <linux/elf.h>
+#endif
 
 #include <asm/fixmap.h>
 #include <asm/tlbflush.h>
@@ -28,6 +31,9 @@ EXPORT_SYMBOL(empty_zero_page);
 extern char _start[];
 void *dtb_early_va;
 
+unsigned long kernel_load_addr = _AC(CONFIG_PAGE_OFFSET, UL);
+EXPORT_SYMBOL(kernel_load_addr);
+
 static void __init zone_sizes_init(void)
 {
        unsigned long max_zone_pfns[MAX_NR_ZONES] = { 0, };
@@ -132,7 +138,8 @@ void __init setup_bootmem(void)
                phys_addr_t end = reg->base + reg->size;
 
                if (reg->base <= vmlinux_end && vmlinux_end <= end) {
-                       mem_size = min(reg->size, (phys_addr_t)-PAGE_OFFSET);
+                       mem_size = min(reg->size,
+                                      (phys_addr_t)-kernel_load_addr);
 
                        /*
                         * Remove memblock from the end of usable area to the
@@ -269,7 +276,7 @@ static phys_addr_t __init alloc_pmd(uintptr_t va)
        if (mmu_enabled)
                return memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
 
-       pmd_num = (va - PAGE_OFFSET) >> PGDIR_SHIFT;
+       pmd_num = (va - kernel_load_addr) >> PGDIR_SHIFT;
        BUG_ON(pmd_num >= NUM_EARLY_PMDS);
        return (uintptr_t)&early_pmd[pmd_num * PTRS_PER_PMD];
 }
@@ -370,6 +377,54 @@ static uintptr_t __init best_map_size(phys_addr_t base, phys_addr_t size)
 #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;
+
+#ifdef CONFIG_64BIT
+#define Elf_Rela Elf64_Rela
+#define Elf_Addr Elf64_Addr
+#else
+#define Elf_Rela Elf32_Rela
+#define Elf_Addr Elf32_Addr
+#endif
+
+void __init relocate_kernel(uintptr_t load_pa)
+{
+       Elf_Rela *rela = (Elf_Rela *)&__rela_dyn_start;
+       uintptr_t link_addr = _AC(CONFIG_PAGE_OFFSET, UL);
+       /*
+        * This holds the offset between the linked virtual address and the
+        * relocated virtual address.
+        */
+       uintptr_t reloc_offset = kernel_load_addr - link_addr;
+       /*
+        * This holds the offset between linked virtual address and physical
+        * address whereas va_pa_offset holds the offset between relocated
+        * virtual address and physical address.
+        */
+       uintptr_t va_link_pa_offset = link_addr - load_pa;
+
+       for ( ; rela < (Elf_Rela *)&__rela_dyn_end; rela++) {
+               Elf_Addr addr = (rela->r_offset - va_link_pa_offset);
+               Elf_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 >= link_addr)
+                       relocated_addr += reloc_offset;
+
+               *(Elf_Addr *)addr = relocated_addr;
+       }
+}
+#endif
+
 asmlinkage void __init setup_vm(uintptr_t dtb_pa)
 {
        uintptr_t va, end_va;
@@ -377,9 +432,20 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
        uintptr_t load_sz = (uintptr_t)(&_end) - load_pa;
        uintptr_t map_size = best_map_size(load_pa, MAX_EARLY_MAPPING_SIZE);
 
-       va_pa_offset = PAGE_OFFSET - load_pa;
+       va_pa_offset = kernel_load_addr - load_pa;
        pfn_base = PFN_DOWN(load_pa);
 
+#ifdef CONFIG_RELOCATABLE
+       /*
+        * Early page table uses only one PGDIR, which makes it possible
+        * to map 1GB aligned on 1GB: if the relocation offset makes the kernel
+        * cross over a 1G boundary, raise a bug since a part of the kernel
+        * would not get mapped.
+        */
+       BUG_ON(SZ_1G - (kernel_load_addr & (SZ_1G - 1)) < load_sz);
+       relocate_kernel(load_pa);
+#endif
+
        /*
         * Enforce boot alignment requirements of RV32 and
         * RV64 by only allowing PMD or PGD mappings.
@@ -387,7 +453,7 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
        BUG_ON(map_size == PAGE_SIZE);
 
        /* Sanity check alignment and size */
-       BUG_ON((PAGE_OFFSET % PGDIR_SIZE) != 0);
+       BUILD_BUG_ON((_AC(CONFIG_PAGE_OFFSET, UL) % PGDIR_SIZE) != 0);
        BUG_ON((load_pa % map_size) != 0);
        BUG_ON(load_sz > MAX_EARLY_MAPPING_SIZE);
 
@@ -400,13 +466,13 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
        create_pmd_mapping(fixmap_pmd, FIXADDR_START,
                           (uintptr_t)fixmap_pte, PMD_SIZE, PAGE_TABLE);
        /* Setup trampoline PGD and PMD */
-       create_pgd_mapping(trampoline_pg_dir, PAGE_OFFSET,
+       create_pgd_mapping(trampoline_pg_dir, kernel_load_addr,
                           (uintptr_t)trampoline_pmd, PGDIR_SIZE, PAGE_TABLE);
-       create_pmd_mapping(trampoline_pmd, PAGE_OFFSET,
+       create_pmd_mapping(trampoline_pmd, kernel_load_addr,
                           load_pa, PMD_SIZE, PAGE_KERNEL_EXEC);
 #else
        /* Setup trampoline PGD */
-       create_pgd_mapping(trampoline_pg_dir, PAGE_OFFSET,
+       create_pgd_mapping(trampoline_pg_dir, kernel_load_addr,
                           load_pa, PGDIR_SIZE, PAGE_KERNEL_EXEC);
 #endif
 
@@ -415,10 +481,10 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         * us to reach paging_init(). We map all memory banks later
         * in setup_vm_final() below.
         */
-       end_va = PAGE_OFFSET + load_sz;
-       for (va = PAGE_OFFSET; va < end_va; va += map_size)
+       end_va = kernel_load_addr + load_sz;
+       for (va = kernel_load_addr; va < end_va; va += map_size)
                create_pgd_mapping(early_pg_dir, va,
-                                  load_pa + (va - PAGE_OFFSET),
+                                  load_pa + (va - kernel_load_addr),
                                   map_size, PAGE_KERNEL_EXEC);
 
        /* Create fixed mapping for early FDT parsing */
@@ -457,9 +523,9 @@ static void __init setup_vm_final(void)
                        break;
                if (memblock_is_nomap(reg))
                        continue;
-               if (start <= __pa(PAGE_OFFSET) &&
-                   __pa(PAGE_OFFSET) < end)
-                       start = __pa(PAGE_OFFSET);
+               if (start <= __pa(kernel_load_addr) &&
+                   __pa(kernel_load_addr) < end)
+                       start = __pa(kernel_load_addr);
 
                map_size = best_map_size(start, end - start);
                for (pa = start; pa < end; pa += map_size) {
--
Quoted 2 lines

2.20.1

Apart from minor comments above, I will certainly try this patch my end.

Regards,
Anup

author reviewer
#2 lore raw

Hi Anup,

Quoted 44 lines · Anup Patel
Quoted 30 lines · Alexandre Ghiti

This config allows to compile the kernel as PIE and to relocate it at any virtual address at runtime: this paves the way to KASLR and to 4-level page table folding at runtime. 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.

---
  arch/riscv/Kconfig              | 11 ++++
  arch/riscv/Makefile             |  5 +-
  arch/riscv/boot/loader.lds.S    |  2 +-
  arch/riscv/include/asm/page.h   |  5 +-
  arch/riscv/kernel/head.S        |  3 +-
  arch/riscv/kernel/vmlinux.lds.S | 10 ++--
  arch/riscv/mm/Makefile          |  4 ++
  arch/riscv/mm/init.c            | 92 ++++++++++++++++++++++++++++-----
  8 files changed, 110 insertions(+), 22 deletions(-)
arch/riscv/Kconfig +1 -0
index fa7dc03459e7..c652b4b850ce 100644
@@ -163,6 +163,17 @@ config PGTABLE_LEVELS
         default 3 if 64BIT
         default 2
 
+config RELOCATABLE

arch/riscv/Kconfig config PGTABLE_LEVELS · line 176

Quoted 10 lines
arch/riscv/Kconfig +10 -0
+       bool
+       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.
+

I am sure this cannot be used for NOMMU support because of the page table mappings.

Quoted 1 line

I guess this should depend on MMU.

Yes you're right, I'll add a dependency on MMU in v2, thanks.

Quoted 26 lines
Quoted 26 lines
arch/riscv/Kconfig +0 -0
  source "arch/riscv/Kconfig.socs"
 
  menu "Platform type"
arch/riscv/Makefile +4 -1
index b9009a2fbaf5..5a115cf6a9c1 100644
@@ -9,7 +9,10 @@
  #
 
  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
  endif
arch/riscv/boot/loader.lds.S +0 -0
index 47a5003c2e28..a9ed218171aa 100644
@@ -7,7 +7,7 @@ ENTRY(_start)
 
Quoted 8 lines

arch/riscv/boot/loader.lds.S ENTRY(_start) · line 13

Quoted 7 lines
arch/riscv/boot/loader.lds.S +1 -1
  SECTIONS
  {
-       . = PAGE_OFFSET;
+       . = CONFIG_PAGE_OFFSET;
 
         .payload : {
                 *(.payload)

The loader.lds.S is used by NOMMU support so this should be separate patch.

This patch defines PAGE_OFFSET as a variable which will make NOMMU support not compilable, I'm not sure I should move that out of this patch, are you ?

Quoted 21 lines
Quoted 21 lines
arch/riscv/include/asm/page.h +2 -2
index ac699246ae7e..27c95da68ecb 100644
@@ -31,9 +31,9 @@
   * When not using MMU this corresponds to the first free page in
   * physical memory (aligned on a page boundary).
   */
-#define PAGE_OFFSET            _AC(CONFIG_PAGE_OFFSET, UL)
+#define PAGE_OFFSET            kernel_load_addr
 
-#define KERN_VIRT_SIZE (-PAGE_OFFSET)
+#define KERN_VIRT_SIZE         (-_AC(CONFIG_PAGE_OFFSET, UL))
 
  #ifndef __ASSEMBLY__
 
@@ -97,6 +97,7 @@ extern unsigned long pfn_base;
  #define ARCH_PFN_OFFSET                (PAGE_OFFSET >> PAGE_SHIFT)
  #endif /* CONFIG_MMU */
 
Quoted 5 lines

arch/riscv/include/asm/page.h extern unsigned long pfn_base; · line 100

Quoted 1 line
arch/riscv/include/asm/page.h +1 -0
+extern unsigned long kernel_load_addr;

Having this variable named kernel_virt_addr seems more logical to me.

Is it mandatory to use kernel_load_addr name ?

Not mandatory at all, I decided the naming. I agree with your suggestion, "load" refers better to physical addressing. I'll change to kernel_virt_addr in v2, thanks.

Quoted 232 lines
Quoted 232 lines
arch/riscv/include/asm/page.h +0 -0
  extern unsigned long max_low_pfn;
  extern unsigned long min_low_pfn;
 
arch/riscv/kernel/head.S +2 -1
index 2227db63f895..5042b2b48a06 100644
@@ -126,7 +126,8 @@ clear_bss_done:
  #ifdef CONFIG_MMU
  relocate:
         /* Relocate return address */
-       li a1, PAGE_OFFSET
+       la a1, kernel_load_addr
+       REG_L a1, 0(a1)
         la a2, _start
         sub a1, a1, a2
         add ra, ra, a1
arch/riscv/kernel/vmlinux.lds.S +6 -4
index 12f42f96d46e..5095aee7c37e 100644
@@ -4,7 +4,7 @@
   * Copyright (C) 2017 SiFive
   */
 
-#define LOAD_OFFSET PAGE_OFFSET
+#define LOAD_OFFSET CONFIG_PAGE_OFFSET
  #include <asm/vmlinux.lds.h>
  #include <asm/page.h>
  #include <asm/cache.h>
@@ -70,9 +70,11 @@ SECTIONS
 
         EXCEPTION_TABLE(0x10)
 
-       .rel.dyn : {
-               *(.rel.dyn*)
-       }
+        .rela.dyn : ALIGN(8) {
+               __rela_dyn_start = .;
+                *(.rela .rela*)
+               __rela_dyn_end = .;
+        }
 
         _end = .;
 
arch/riscv/mm/Makefile +4 -0
index a1bd95c8047a..dcd3d806243f 100644
@@ -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 = -pg
  endif
arch/riscv/mm/init.c +79 -13
index 965a8cf4829c..ac9a9f69abc0 100644
@@ -12,6 +12,9 @@
  #include <linux/sizes.h>
  #include <linux/of_fdt.h>
  #include <linux/libfdt.h>
+#ifdef CONFIG_RELOCATABLE
+#include <linux/elf.h>
+#endif
 
  #include <asm/fixmap.h>
  #include <asm/tlbflush.h>
@@ -28,6 +31,9 @@ EXPORT_SYMBOL(empty_zero_page);
  extern char _start[];
  void *dtb_early_va;
 
+unsigned long kernel_load_addr = _AC(CONFIG_PAGE_OFFSET, UL);
+EXPORT_SYMBOL(kernel_load_addr);
+
  static void __init zone_sizes_init(void)
  {
         unsigned long max_zone_pfns[MAX_NR_ZONES] = { 0, };
@@ -132,7 +138,8 @@ void __init setup_bootmem(void)
                 phys_addr_t end = reg->base + reg->size;
 
                 if (reg->base <= vmlinux_end && vmlinux_end <= end) {
-                       mem_size = min(reg->size, (phys_addr_t)-PAGE_OFFSET);
+                       mem_size = min(reg->size,
+                                      (phys_addr_t)-kernel_load_addr);
 
                         /*
                          * Remove memblock from the end of usable area to the
@@ -269,7 +276,7 @@ static phys_addr_t __init alloc_pmd(uintptr_t va)
         if (mmu_enabled)
                 return memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
 
-       pmd_num = (va - PAGE_OFFSET) >> PGDIR_SHIFT;
+       pmd_num = (va - kernel_load_addr) >> PGDIR_SHIFT;
         BUG_ON(pmd_num >= NUM_EARLY_PMDS);
         return (uintptr_t)&early_pmd[pmd_num * PTRS_PER_PMD];
  }
@@ -370,6 +377,54 @@ static uintptr_t __init best_map_size(phys_addr_t base, phys_addr_t size)
  #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;
+
+#ifdef CONFIG_64BIT
+#define Elf_Rela Elf64_Rela
+#define Elf_Addr Elf64_Addr
+#else
+#define Elf_Rela Elf32_Rela
+#define Elf_Addr Elf32_Addr
+#endif
+
+void __init relocate_kernel(uintptr_t load_pa)
+{
+       Elf_Rela *rela = (Elf_Rela *)&__rela_dyn_start;
+       uintptr_t link_addr = _AC(CONFIG_PAGE_OFFSET, UL);
+       /*
+        * This holds the offset between the linked virtual address and the
+        * relocated virtual address.
+        */
+       uintptr_t reloc_offset = kernel_load_addr - link_addr;
+       /*
+        * This holds the offset between linked virtual address and physical
+        * address whereas va_pa_offset holds the offset between relocated
+        * virtual address and physical address.
+        */
+       uintptr_t va_link_pa_offset = link_addr - load_pa;
+
+       for ( ; rela < (Elf_Rela *)&__rela_dyn_end; rela++) {
+               Elf_Addr addr = (rela->r_offset - va_link_pa_offset);
+               Elf_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 >= link_addr)
+                       relocated_addr += reloc_offset;
+
+               *(Elf_Addr *)addr = relocated_addr;
+       }
+}
+#endif
+
  asmlinkage void __init setup_vm(uintptr_t dtb_pa)
  {
         uintptr_t va, end_va;
@@ -377,9 +432,20 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         uintptr_t load_sz = (uintptr_t)(&_end) - load_pa;
         uintptr_t map_size = best_map_size(load_pa, MAX_EARLY_MAPPING_SIZE);
 
-       va_pa_offset = PAGE_OFFSET - load_pa;
+       va_pa_offset = kernel_load_addr - load_pa;
         pfn_base = PFN_DOWN(load_pa);
 
+#ifdef CONFIG_RELOCATABLE
+       /*
+        * Early page table uses only one PGDIR, which makes it possible
+        * to map 1GB aligned on 1GB: if the relocation offset makes the kernel
+        * cross over a 1G boundary, raise a bug since a part of the kernel
+        * would not get mapped.
+        */
+       BUG_ON(SZ_1G - (kernel_load_addr & (SZ_1G - 1)) < load_sz);
+       relocate_kernel(load_pa);
+#endif
+
         /*
          * Enforce boot alignment requirements of RV32 and
          * RV64 by only allowing PMD or PGD mappings.
@@ -387,7 +453,7 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         BUG_ON(map_size == PAGE_SIZE);
 
         /* Sanity check alignment and size */
-       BUG_ON((PAGE_OFFSET % PGDIR_SIZE) != 0);
+       BUILD_BUG_ON((_AC(CONFIG_PAGE_OFFSET, UL) % PGDIR_SIZE) != 0);
         BUG_ON((load_pa % map_size) != 0);
         BUG_ON(load_sz > MAX_EARLY_MAPPING_SIZE);
 
@@ -400,13 +466,13 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         create_pmd_mapping(fixmap_pmd, FIXADDR_START,
                            (uintptr_t)fixmap_pte, PMD_SIZE, PAGE_TABLE);
         /* Setup trampoline PGD and PMD */
-       create_pgd_mapping(trampoline_pg_dir, PAGE_OFFSET,
+       create_pgd_mapping(trampoline_pg_dir, kernel_load_addr,
                            (uintptr_t)trampoline_pmd, PGDIR_SIZE, PAGE_TABLE);
-       create_pmd_mapping(trampoline_pmd, PAGE_OFFSET,
+       create_pmd_mapping(trampoline_pmd, kernel_load_addr,
                            load_pa, PMD_SIZE, PAGE_KERNEL_EXEC);
  #else
         /* Setup trampoline PGD */
-       create_pgd_mapping(trampoline_pg_dir, PAGE_OFFSET,
+       create_pgd_mapping(trampoline_pg_dir, kernel_load_addr,
                            load_pa, PGDIR_SIZE, PAGE_KERNEL_EXEC);
  #endif
 
@@ -415,10 +481,10 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
          * us to reach paging_init(). We map all memory banks later
          * in setup_vm_final() below.
          */
-       end_va = PAGE_OFFSET + load_sz;
-       for (va = PAGE_OFFSET; va < end_va; va += map_size)
+       end_va = kernel_load_addr + load_sz;
+       for (va = kernel_load_addr; va < end_va; va += map_size)
                 create_pgd_mapping(early_pg_dir, va,
-                                  load_pa + (va - PAGE_OFFSET),
+                                  load_pa + (va - kernel_load_addr),
                                    map_size, PAGE_KERNEL_EXEC);
 
         /* Create fixed mapping for early FDT parsing */
@@ -457,9 +523,9 @@ static void __init setup_vm_final(void)
                         break;
                 if (memblock_is_nomap(reg))
                         continue;
-               if (start <= __pa(PAGE_OFFSET) &&
-                   __pa(PAGE_OFFSET) < end)
-                       start = __pa(PAGE_OFFSET);
+               if (start <= __pa(kernel_load_addr) &&
+                   __pa(kernel_load_addr) < end)
+                       start = __pa(kernel_load_addr);
 
Quoted 7 lines
Quoted 5 lines
arch/riscv/mm/init.c +0 -1
                 map_size = best_map_size(start, end - start);
                 for (pa = start; pa < end; pa += map_size) {
--

2.20.1

Apart from minor comments above, I will certainly try this patch my end.

Great, more testing is always welcome, thanks.

Quoted 3 lines

Regards,
Anup

Alex

Reviewed-by Tested-by on v1 1/1
#3 lore raw
In reply to Alexandre Ghiti
Quoted 315 lines · Alexandre Ghiti

This config allows to compile the kernel as PIE and to relocate it at any virtual address at runtime: this paves the way to KASLR and to 4-level page table folding at runtime. 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.

---
 arch/riscv/Kconfig              | 11 ++++
 arch/riscv/Makefile             |  5 +-
 arch/riscv/boot/loader.lds.S    |  2 +-
 arch/riscv/include/asm/page.h   |  5 +-
 arch/riscv/kernel/head.S        |  3 +-
 arch/riscv/kernel/vmlinux.lds.S | 10 ++--
 arch/riscv/mm/Makefile          |  4 ++
 arch/riscv/mm/init.c            | 92 ++++++++++++++++++++++++++++-----
 8 files changed, 110 insertions(+), 22 deletions(-)
arch/riscv/Kconfig +11 -0
index fa7dc03459e7..c652b4b850ce 100644
@@ -163,6 +163,17 @@ config PGTABLE_LEVELS
        default 3 if 64BIT
        default 2
 
+config RELOCATABLE
+       bool
+       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.
+
 source "arch/riscv/Kconfig.socs"
 
 menu "Platform type"
arch/riscv/Makefile +4 -1
index b9009a2fbaf5..5a115cf6a9c1 100644
@@ -9,7 +9,10 @@
 #
 
 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
 endif
arch/riscv/boot/loader.lds.S +1 -1
index 47a5003c2e28..a9ed218171aa 100644
@@ -7,7 +7,7 @@ ENTRY(_start)
 
 SECTIONS
 {
-       . = PAGE_OFFSET;
+       . = CONFIG_PAGE_OFFSET;
 
        .payload : {
                *(.payload)
arch/riscv/include/asm/page.h +3 -2
index ac699246ae7e..27c95da68ecb 100644
@@ -31,9 +31,9 @@
  * When not using MMU this corresponds to the first free page in
  * physical memory (aligned on a page boundary).
  */
-#define PAGE_OFFSET            _AC(CONFIG_PAGE_OFFSET, UL)
+#define PAGE_OFFSET            kernel_load_addr
 
-#define KERN_VIRT_SIZE (-PAGE_OFFSET)
+#define KERN_VIRT_SIZE         (-_AC(CONFIG_PAGE_OFFSET, UL))
 
 #ifndef __ASSEMBLY__
 
@@ -97,6 +97,7 @@ extern unsigned long pfn_base;
 #define ARCH_PFN_OFFSET                (PAGE_OFFSET >> PAGE_SHIFT)
 #endif /* CONFIG_MMU */
 
+extern unsigned long kernel_load_addr;
 extern unsigned long max_low_pfn;
 extern unsigned long min_low_pfn;
 
arch/riscv/kernel/head.S +2 -1
index 2227db63f895..5042b2b48a06 100644
@@ -126,7 +126,8 @@ clear_bss_done:
 #ifdef CONFIG_MMU
 relocate:
        /* Relocate return address */
-       li a1, PAGE_OFFSET
+       la a1, kernel_load_addr
+       REG_L a1, 0(a1)
        la a2, _start
        sub a1, a1, a2
        add ra, ra, a1
arch/riscv/kernel/vmlinux.lds.S +6 -4
index 12f42f96d46e..5095aee7c37e 100644
@@ -4,7 +4,7 @@
  * Copyright (C) 2017 SiFive
  */
 
-#define LOAD_OFFSET PAGE_OFFSET
+#define LOAD_OFFSET CONFIG_PAGE_OFFSET
 #include <asm/vmlinux.lds.h>
 #include <asm/page.h>
 #include <asm/cache.h>
@@ -70,9 +70,11 @@ SECTIONS
 
        EXCEPTION_TABLE(0x10)
 
-       .rel.dyn : {
-               *(.rel.dyn*)
-       }
+        .rela.dyn : ALIGN(8) {
+               __rela_dyn_start = .;
+                *(.rela .rela*)
+               __rela_dyn_end = .;
+        }
 
        _end = .;
 
arch/riscv/mm/Makefile +4 -0
index a1bd95c8047a..dcd3d806243f 100644
@@ -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 = -pg
 endif
arch/riscv/mm/init.c +76 -10
index 965a8cf4829c..ac9a9f69abc0 100644
@@ -12,6 +12,9 @@
 #include <linux/sizes.h>
 #include <linux/of_fdt.h>
 #include <linux/libfdt.h>
+#ifdef CONFIG_RELOCATABLE
+#include <linux/elf.h>
+#endif
 
 #include <asm/fixmap.h>
 #include <asm/tlbflush.h>
@@ -28,6 +31,9 @@ EXPORT_SYMBOL(empty_zero_page);
 extern char _start[];
 void *dtb_early_va;
 
+unsigned long kernel_load_addr = _AC(CONFIG_PAGE_OFFSET, UL);
+EXPORT_SYMBOL(kernel_load_addr);
+
 static void __init zone_sizes_init(void)
 {
        unsigned long max_zone_pfns[MAX_NR_ZONES] = { 0, };
@@ -132,7 +138,8 @@ void __init setup_bootmem(void)
                phys_addr_t end = reg->base + reg->size;
 
                if (reg->base <= vmlinux_end && vmlinux_end <= end) {
-                       mem_size = min(reg->size, (phys_addr_t)-PAGE_OFFSET);
+                       mem_size = min(reg->size,
+                                      (phys_addr_t)-kernel_load_addr);
 
                        /*
                         * Remove memblock from the end of usable area to the
@@ -269,7 +276,7 @@ static phys_addr_t __init alloc_pmd(uintptr_t va)
        if (mmu_enabled)
                return memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
 
-       pmd_num = (va - PAGE_OFFSET) >> PGDIR_SHIFT;
+       pmd_num = (va - kernel_load_addr) >> PGDIR_SHIFT;
        BUG_ON(pmd_num >= NUM_EARLY_PMDS);
        return (uintptr_t)&early_pmd[pmd_num * PTRS_PER_PMD];
 }
@@ -370,6 +377,54 @@ static uintptr_t __init best_map_size(phys_addr_t base, phys_addr_t size)
 #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;
+
+#ifdef CONFIG_64BIT
+#define Elf_Rela Elf64_Rela
+#define Elf_Addr Elf64_Addr
+#else
+#define Elf_Rela Elf32_Rela
+#define Elf_Addr Elf32_Addr
+#endif
+
+void __init relocate_kernel(uintptr_t load_pa)
+{
+       Elf_Rela *rela = (Elf_Rela *)&__rela_dyn_start;
+       uintptr_t link_addr = _AC(CONFIG_PAGE_OFFSET, UL);
+       /*
+        * This holds the offset between the linked virtual address and the
+        * relocated virtual address.
+        */
+       uintptr_t reloc_offset = kernel_load_addr - link_addr;
+       /*
+        * This holds the offset between linked virtual address and physical
+        * address whereas va_pa_offset holds the offset between relocated
+        * virtual address and physical address.
+        */
+       uintptr_t va_link_pa_offset = link_addr - load_pa;
+
+       for ( ; rela < (Elf_Rela *)&__rela_dyn_end; rela++) {
+               Elf_Addr addr = (rela->r_offset - va_link_pa_offset);
+               Elf_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 >= link_addr)
+                       relocated_addr += reloc_offset;
+
+               *(Elf_Addr *)addr = relocated_addr;
+       }
+}
+#endif
+
 asmlinkage void __init setup_vm(uintptr_t dtb_pa)
 {
        uintptr_t va, end_va;
@@ -377,9 +432,20 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
        uintptr_t load_sz = (uintptr_t)(&_end) - load_pa;
        uintptr_t map_size = best_map_size(load_pa, MAX_EARLY_MAPPING_SIZE);
 
-       va_pa_offset = PAGE_OFFSET - load_pa;
+       va_pa_offset = kernel_load_addr - load_pa;
        pfn_base = PFN_DOWN(load_pa);
 
+#ifdef CONFIG_RELOCATABLE
+       /*
+        * Early page table uses only one PGDIR, which makes it possible
+        * to map 1GB aligned on 1GB: if the relocation offset makes the kernel
+        * cross over a 1G boundary, raise a bug since a part of the kernel
+        * would not get mapped.
+        */
+       BUG_ON(SZ_1G - (kernel_load_addr & (SZ_1G - 1)) < load_sz);
+       relocate_kernel(load_pa);
+#endif
+
        /*
         * Enforce boot alignment requirements of RV32 and
         * RV64 by only allowing PMD or PGD mappings.
@@ -387,7 +453,7 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
        BUG_ON(map_size == PAGE_SIZE);
 
        /* Sanity check alignment and size */
-       BUG_ON((PAGE_OFFSET % PGDIR_SIZE) != 0);
+       BUILD_BUG_ON((_AC(CONFIG_PAGE_OFFSET, UL) % PGDIR_SIZE) != 0);
        BUG_ON((load_pa % map_size) != 0);
        BUG_ON(load_sz > MAX_EARLY_MAPPING_SIZE);
 
@@ -400,13 +466,13 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
        create_pmd_mapping(fixmap_pmd, FIXADDR_START,
                           (uintptr_t)fixmap_pte, PMD_SIZE, PAGE_TABLE);
        /* Setup trampoline PGD and PMD */
-       create_pgd_mapping(trampoline_pg_dir, PAGE_OFFSET,
+       create_pgd_mapping(trampoline_pg_dir, kernel_load_addr,
                           (uintptr_t)trampoline_pmd, PGDIR_SIZE, PAGE_TABLE);
-       create_pmd_mapping(trampoline_pmd, PAGE_OFFSET,
+       create_pmd_mapping(trampoline_pmd, kernel_load_addr,
                           load_pa, PMD_SIZE, PAGE_KERNEL_EXEC);
 #else
        /* Setup trampoline PGD */
-       create_pgd_mapping(trampoline_pg_dir, PAGE_OFFSET,
+       create_pgd_mapping(trampoline_pg_dir, kernel_load_addr,
                           load_pa, PGDIR_SIZE, PAGE_KERNEL_EXEC);
 #endif
 
@@ -415,10 +481,10 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         * us to reach paging_init(). We map all memory banks later
         * in setup_vm_final() below.
         */
-       end_va = PAGE_OFFSET + load_sz;
-       for (va = PAGE_OFFSET; va < end_va; va += map_size)
+       end_va = kernel_load_addr + load_sz;
+       for (va = kernel_load_addr; va < end_va; va += map_size)
                create_pgd_mapping(early_pg_dir, va,
-                                  load_pa + (va - PAGE_OFFSET),
+                                  load_pa + (va - kernel_load_addr),
                                   map_size, PAGE_KERNEL_EXEC);
 
        /* Create fixed mapping for early FDT parsing */

arch/riscv/mm/init.c static void __init setup_vm_final(void) · line 528

Quoted 9 lines
arch/riscv/mm/init.c +3 -3
@@ -457,9 +523,9 @@ static void __init setup_vm_final(void)
                        break;
                if (memblock_is_nomap(reg))
                        continue;
-               if (start <= __pa(PAGE_OFFSET) &&
-                   __pa(PAGE_OFFSET) < end)
-                       start = __pa(PAGE_OFFSET);
+               if (start <= __pa(kernel_load_addr) &&
+                   __pa(kernel_load_addr) < end)
+                       start = __pa(kernel_load_addr);

Here should use __pa_symbol() instead of __pa() for kernel symbols.

I'm working on KASLR on top of this patch, it's work to me.

Quoted 6 lines
arch/riscv/mm/init.c +0 -1
                map_size = best_map_size(start, end - start);
                for (pa = start; pa < end; pa += map_size) {
--

2.20.1

#4 lore raw
Quoted 329 lines · Zong Li
Quoted 315 lines · Alexandre Ghiti

This config allows to compile the kernel as PIE and to relocate it at any virtual address at runtime: this paves the way to KASLR and to 4-level page table folding at runtime. 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.

---
 arch/riscv/Kconfig              | 11 ++++
 arch/riscv/Makefile             |  5 +-
 arch/riscv/boot/loader.lds.S    |  2 +-
 arch/riscv/include/asm/page.h   |  5 +-
 arch/riscv/kernel/head.S        |  3 +-
 arch/riscv/kernel/vmlinux.lds.S | 10 ++--
 arch/riscv/mm/Makefile          |  4 ++
 arch/riscv/mm/init.c            | 92 ++++++++++++++++++++++++++++-----
 8 files changed, 110 insertions(+), 22 deletions(-)
arch/riscv/Kconfig +11 -0
index fa7dc03459e7..c652b4b850ce 100644
@@ -163,6 +163,17 @@ config PGTABLE_LEVELS
        default 3 if 64BIT
        default 2
 
+config RELOCATABLE
+       bool
+       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.
+
 source "arch/riscv/Kconfig.socs"
 
 menu "Platform type"
arch/riscv/Makefile +4 -1
index b9009a2fbaf5..5a115cf6a9c1 100644
@@ -9,7 +9,10 @@
 #
 
 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
 endif
arch/riscv/boot/loader.lds.S +1 -1
index 47a5003c2e28..a9ed218171aa 100644
@@ -7,7 +7,7 @@ ENTRY(_start)
 
 SECTIONS
 {
-       . = PAGE_OFFSET;
+       . = CONFIG_PAGE_OFFSET;
 
        .payload : {
                *(.payload)
arch/riscv/include/asm/page.h +3 -2
index ac699246ae7e..27c95da68ecb 100644
@@ -31,9 +31,9 @@
  * When not using MMU this corresponds to the first free page in
  * physical memory (aligned on a page boundary).
  */
-#define PAGE_OFFSET            _AC(CONFIG_PAGE_OFFSET, UL)
+#define PAGE_OFFSET            kernel_load_addr
 
-#define KERN_VIRT_SIZE (-PAGE_OFFSET)
+#define KERN_VIRT_SIZE         (-_AC(CONFIG_PAGE_OFFSET, UL))
 
 #ifndef __ASSEMBLY__
 
@@ -97,6 +97,7 @@ extern unsigned long pfn_base;
 #define ARCH_PFN_OFFSET                (PAGE_OFFSET >> PAGE_SHIFT)
 #endif /* CONFIG_MMU */
 
+extern unsigned long kernel_load_addr;
 extern unsigned long max_low_pfn;
 extern unsigned long min_low_pfn;
 
arch/riscv/kernel/head.S +2 -1
index 2227db63f895..5042b2b48a06 100644
@@ -126,7 +126,8 @@ clear_bss_done:
 #ifdef CONFIG_MMU
 relocate:
        /* Relocate return address */
-       li a1, PAGE_OFFSET
+       la a1, kernel_load_addr
+       REG_L a1, 0(a1)
        la a2, _start
        sub a1, a1, a2
        add ra, ra, a1
arch/riscv/kernel/vmlinux.lds.S +6 -4
index 12f42f96d46e..5095aee7c37e 100644
@@ -4,7 +4,7 @@
  * Copyright (C) 2017 SiFive
  */
 
-#define LOAD_OFFSET PAGE_OFFSET
+#define LOAD_OFFSET CONFIG_PAGE_OFFSET
 #include <asm/vmlinux.lds.h>
 #include <asm/page.h>
 #include <asm/cache.h>
@@ -70,9 +70,11 @@ SECTIONS
 
        EXCEPTION_TABLE(0x10)
 
-       .rel.dyn : {
-               *(.rel.dyn*)
-       }
+        .rela.dyn : ALIGN(8) {
+               __rela_dyn_start = .;
+                *(.rela .rela*)
+               __rela_dyn_end = .;
+        }
 
        _end = .;
 
arch/riscv/mm/Makefile +4 -0
index a1bd95c8047a..dcd3d806243f 100644
@@ -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 = -pg
 endif
arch/riscv/mm/init.c +76 -10
index 965a8cf4829c..ac9a9f69abc0 100644
@@ -12,6 +12,9 @@
 #include <linux/sizes.h>
 #include <linux/of_fdt.h>
 #include <linux/libfdt.h>
+#ifdef CONFIG_RELOCATABLE
+#include <linux/elf.h>
+#endif
 
 #include <asm/fixmap.h>
 #include <asm/tlbflush.h>
@@ -28,6 +31,9 @@ EXPORT_SYMBOL(empty_zero_page);
 extern char _start[];
 void *dtb_early_va;
 
+unsigned long kernel_load_addr = _AC(CONFIG_PAGE_OFFSET, UL);
+EXPORT_SYMBOL(kernel_load_addr);
+
 static void __init zone_sizes_init(void)
 {
        unsigned long max_zone_pfns[MAX_NR_ZONES] = { 0, };
@@ -132,7 +138,8 @@ void __init setup_bootmem(void)
                phys_addr_t end = reg->base + reg->size;
 
                if (reg->base <= vmlinux_end && vmlinux_end <= end) {
-                       mem_size = min(reg->size, (phys_addr_t)-PAGE_OFFSET);
+                       mem_size = min(reg->size,
+                                      (phys_addr_t)-kernel_load_addr);
 
                        /*
                         * Remove memblock from the end of usable area to the
@@ -269,7 +276,7 @@ static phys_addr_t __init alloc_pmd(uintptr_t va)
        if (mmu_enabled)
                return memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
 
-       pmd_num = (va - PAGE_OFFSET) >> PGDIR_SHIFT;
+       pmd_num = (va - kernel_load_addr) >> PGDIR_SHIFT;
        BUG_ON(pmd_num >= NUM_EARLY_PMDS);
        return (uintptr_t)&early_pmd[pmd_num * PTRS_PER_PMD];
 }
@@ -370,6 +377,54 @@ static uintptr_t __init best_map_size(phys_addr_t base, phys_addr_t size)
 #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;
+
+#ifdef CONFIG_64BIT
+#define Elf_Rela Elf64_Rela
+#define Elf_Addr Elf64_Addr
+#else
+#define Elf_Rela Elf32_Rela
+#define Elf_Addr Elf32_Addr
+#endif
+
+void __init relocate_kernel(uintptr_t load_pa)
+{
+       Elf_Rela *rela = (Elf_Rela *)&__rela_dyn_start;
+       uintptr_t link_addr = _AC(CONFIG_PAGE_OFFSET, UL);
+       /*
+        * This holds the offset between the linked virtual address and the
+        * relocated virtual address.
+        */
+       uintptr_t reloc_offset = kernel_load_addr - link_addr;
+       /*
+        * This holds the offset between linked virtual address and physical
+        * address whereas va_pa_offset holds the offset between relocated
+        * virtual address and physical address.
+        */
+       uintptr_t va_link_pa_offset = link_addr - load_pa;
+
+       for ( ; rela < (Elf_Rela *)&__rela_dyn_end; rela++) {
+               Elf_Addr addr = (rela->r_offset - va_link_pa_offset);
+               Elf_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 >= link_addr)
+                       relocated_addr += reloc_offset;
+
+               *(Elf_Addr *)addr = relocated_addr;
+       }
+}
+#endif
+
 asmlinkage void __init setup_vm(uintptr_t dtb_pa)
 {
        uintptr_t va, end_va;
@@ -377,9 +432,20 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
        uintptr_t load_sz = (uintptr_t)(&_end) - load_pa;
        uintptr_t map_size = best_map_size(load_pa, MAX_EARLY_MAPPING_SIZE);
 
-       va_pa_offset = PAGE_OFFSET - load_pa;
+       va_pa_offset = kernel_load_addr - load_pa;
        pfn_base = PFN_DOWN(load_pa);
 
+#ifdef CONFIG_RELOCATABLE
+       /*
+        * Early page table uses only one PGDIR, which makes it possible
+        * to map 1GB aligned on 1GB: if the relocation offset makes the kernel
+        * cross over a 1G boundary, raise a bug since a part of the kernel
+        * would not get mapped.
+        */
+       BUG_ON(SZ_1G - (kernel_load_addr & (SZ_1G - 1)) < load_sz);
+       relocate_kernel(load_pa);
+#endif
+
        /*
         * Enforce boot alignment requirements of RV32 and
         * RV64 by only allowing PMD or PGD mappings.
@@ -387,7 +453,7 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
        BUG_ON(map_size == PAGE_SIZE);
 
        /* Sanity check alignment and size */
-       BUG_ON((PAGE_OFFSET % PGDIR_SIZE) != 0);
+       BUILD_BUG_ON((_AC(CONFIG_PAGE_OFFSET, UL) % PGDIR_SIZE) != 0);
        BUG_ON((load_pa % map_size) != 0);
        BUG_ON(load_sz > MAX_EARLY_MAPPING_SIZE);
 
@@ -400,13 +466,13 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
        create_pmd_mapping(fixmap_pmd, FIXADDR_START,
                           (uintptr_t)fixmap_pte, PMD_SIZE, PAGE_TABLE);
        /* Setup trampoline PGD and PMD */
-       create_pgd_mapping(trampoline_pg_dir, PAGE_OFFSET,
+       create_pgd_mapping(trampoline_pg_dir, kernel_load_addr,
                           (uintptr_t)trampoline_pmd, PGDIR_SIZE, PAGE_TABLE);
-       create_pmd_mapping(trampoline_pmd, PAGE_OFFSET,
+       create_pmd_mapping(trampoline_pmd, kernel_load_addr,
                           load_pa, PMD_SIZE, PAGE_KERNEL_EXEC);
 #else
        /* Setup trampoline PGD */
-       create_pgd_mapping(trampoline_pg_dir, PAGE_OFFSET,
+       create_pgd_mapping(trampoline_pg_dir, kernel_load_addr,
                           load_pa, PGDIR_SIZE, PAGE_KERNEL_EXEC);
 #endif
 
@@ -415,10 +481,10 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         * us to reach paging_init(). We map all memory banks later
         * in setup_vm_final() below.
         */
-       end_va = PAGE_OFFSET + load_sz;
-       for (va = PAGE_OFFSET; va < end_va; va += map_size)
+       end_va = kernel_load_addr + load_sz;
+       for (va = kernel_load_addr; va < end_va; va += map_size)
                create_pgd_mapping(early_pg_dir, va,
-                                  load_pa + (va - PAGE_OFFSET),
+                                  load_pa + (va - kernel_load_addr),
                                   map_size, PAGE_KERNEL_EXEC);
 
        /* Create fixed mapping for early FDT parsing */

arch/riscv/mm/init.c static void __init setup_vm_final(void) · line 528

Quoted 9 lines
arch/riscv/mm/init.c +3 -3
@@ -457,9 +523,9 @@ static void __init setup_vm_final(void)
                        break;
                if (memblock_is_nomap(reg))
                        continue;
-               if (start <= __pa(PAGE_OFFSET) &&
-                   __pa(PAGE_OFFSET) < end)
-                       start = __pa(PAGE_OFFSET);
+               if (start <= __pa(kernel_load_addr) &&
+                   __pa(kernel_load_addr) < end)
+                       start = __pa(kernel_load_addr);

Here should use __pa_symbol() instead of __pa() for kernel symbols.

Quoted 4 lines

I'm working on KASLR on top of this patch, it's work to me.

Thanks for testing. I will not test since you already tested it.

Regards,
Anup

Quoted 7 lines
Quoted 7 lines
arch/riscv/mm/init.c +0 -1
                map_size = best_map_size(start, end - start);
                for (pa = start; pa < end; pa += map_size) {
--

2.20.1

author reviewer
#5 lore raw
In reply to Zong Li

Hi Zong,

Quoted 313 lines · Zong Li
Quoted 312 lines · Alexandre Ghiti

This config allows to compile the kernel as PIE and to relocate it at any virtual address at runtime: this paves the way to KASLR and to 4-level page table folding at runtime. 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.

---
  arch/riscv/Kconfig              | 11 ++++
  arch/riscv/Makefile             |  5 +-
  arch/riscv/boot/loader.lds.S    |  2 +-
  arch/riscv/include/asm/page.h   |  5 +-
  arch/riscv/kernel/head.S        |  3 +-
  arch/riscv/kernel/vmlinux.lds.S | 10 ++--
  arch/riscv/mm/Makefile          |  4 ++
  arch/riscv/mm/init.c            | 92 ++++++++++++++++++++++++++++-----
  8 files changed, 110 insertions(+), 22 deletions(-)
arch/riscv/Kconfig +11 -0
index fa7dc03459e7..c652b4b850ce 100644
@@ -163,6 +163,17 @@ config PGTABLE_LEVELS
         default 3 if 64BIT
         default 2
 
+config RELOCATABLE
+       bool
+       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.
+
  source "arch/riscv/Kconfig.socs"
 
  menu "Platform type"
arch/riscv/Makefile +4 -1
index b9009a2fbaf5..5a115cf6a9c1 100644
@@ -9,7 +9,10 @@
  #
 
  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
  endif
arch/riscv/boot/loader.lds.S +1 -1
index 47a5003c2e28..a9ed218171aa 100644
@@ -7,7 +7,7 @@ ENTRY(_start)
 
  SECTIONS
  {
-       . = PAGE_OFFSET;
+       . = CONFIG_PAGE_OFFSET;
 
         .payload : {
                 *(.payload)
arch/riscv/include/asm/page.h +3 -2
index ac699246ae7e..27c95da68ecb 100644
@@ -31,9 +31,9 @@
   * When not using MMU this corresponds to the first free page in
   * physical memory (aligned on a page boundary).
   */
-#define PAGE_OFFSET            _AC(CONFIG_PAGE_OFFSET, UL)
+#define PAGE_OFFSET            kernel_load_addr
 
-#define KERN_VIRT_SIZE (-PAGE_OFFSET)
+#define KERN_VIRT_SIZE         (-_AC(CONFIG_PAGE_OFFSET, UL))
 
  #ifndef __ASSEMBLY__
 
@@ -97,6 +97,7 @@ extern unsigned long pfn_base;
  #define ARCH_PFN_OFFSET                (PAGE_OFFSET >> PAGE_SHIFT)
  #endif /* CONFIG_MMU */
 
+extern unsigned long kernel_load_addr;
  extern unsigned long max_low_pfn;
  extern unsigned long min_low_pfn;
 
arch/riscv/kernel/head.S +2 -1
index 2227db63f895..5042b2b48a06 100644
@@ -126,7 +126,8 @@ clear_bss_done:
  #ifdef CONFIG_MMU
  relocate:
         /* Relocate return address */
-       li a1, PAGE_OFFSET
+       la a1, kernel_load_addr
+       REG_L a1, 0(a1)
         la a2, _start
         sub a1, a1, a2
         add ra, ra, a1
arch/riscv/kernel/vmlinux.lds.S +6 -4
index 12f42f96d46e..5095aee7c37e 100644
@@ -4,7 +4,7 @@
   * Copyright (C) 2017 SiFive
   */
 
-#define LOAD_OFFSET PAGE_OFFSET
+#define LOAD_OFFSET CONFIG_PAGE_OFFSET
  #include <asm/vmlinux.lds.h>
  #include <asm/page.h>
  #include <asm/cache.h>
@@ -70,9 +70,11 @@ SECTIONS
 
         EXCEPTION_TABLE(0x10)
 
-       .rel.dyn : {
-               *(.rel.dyn*)
-       }
+        .rela.dyn : ALIGN(8) {
+               __rela_dyn_start = .;
+                *(.rela .rela*)
+               __rela_dyn_end = .;
+        }
 
         _end = .;
 
arch/riscv/mm/Makefile +4 -0
index a1bd95c8047a..dcd3d806243f 100644
@@ -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 = -pg
  endif
arch/riscv/mm/init.c +76 -10
index 965a8cf4829c..ac9a9f69abc0 100644
@@ -12,6 +12,9 @@
  #include <linux/sizes.h>
  #include <linux/of_fdt.h>
  #include <linux/libfdt.h>
+#ifdef CONFIG_RELOCATABLE
+#include <linux/elf.h>
+#endif
 
  #include <asm/fixmap.h>
  #include <asm/tlbflush.h>
@@ -28,6 +31,9 @@ EXPORT_SYMBOL(empty_zero_page);
  extern char _start[];
  void *dtb_early_va;
 
+unsigned long kernel_load_addr = _AC(CONFIG_PAGE_OFFSET, UL);
+EXPORT_SYMBOL(kernel_load_addr);
+
  static void __init zone_sizes_init(void)
  {
         unsigned long max_zone_pfns[MAX_NR_ZONES] = { 0, };
@@ -132,7 +138,8 @@ void __init setup_bootmem(void)
                 phys_addr_t end = reg->base + reg->size;
 
                 if (reg->base <= vmlinux_end && vmlinux_end <= end) {
-                       mem_size = min(reg->size, (phys_addr_t)-PAGE_OFFSET);
+                       mem_size = min(reg->size,
+                                      (phys_addr_t)-kernel_load_addr);
 
                         /*
                          * Remove memblock from the end of usable area to the
@@ -269,7 +276,7 @@ static phys_addr_t __init alloc_pmd(uintptr_t va)
         if (mmu_enabled)
                 return memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
 
-       pmd_num = (va - PAGE_OFFSET) >> PGDIR_SHIFT;
+       pmd_num = (va - kernel_load_addr) >> PGDIR_SHIFT;
         BUG_ON(pmd_num >= NUM_EARLY_PMDS);
         return (uintptr_t)&early_pmd[pmd_num * PTRS_PER_PMD];
  }
@@ -370,6 +377,54 @@ static uintptr_t __init best_map_size(phys_addr_t base, phys_addr_t size)
  #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;
+
+#ifdef CONFIG_64BIT
+#define Elf_Rela Elf64_Rela
+#define Elf_Addr Elf64_Addr
+#else
+#define Elf_Rela Elf32_Rela
+#define Elf_Addr Elf32_Addr
+#endif
+
+void __init relocate_kernel(uintptr_t load_pa)
+{
+       Elf_Rela *rela = (Elf_Rela *)&__rela_dyn_start;
+       uintptr_t link_addr = _AC(CONFIG_PAGE_OFFSET, UL);
+       /*
+        * This holds the offset between the linked virtual address and the
+        * relocated virtual address.
+        */
+       uintptr_t reloc_offset = kernel_load_addr - link_addr;
+       /*
+        * This holds the offset between linked virtual address and physical
+        * address whereas va_pa_offset holds the offset between relocated
+        * virtual address and physical address.
+        */
+       uintptr_t va_link_pa_offset = link_addr - load_pa;
+
+       for ( ; rela < (Elf_Rela *)&__rela_dyn_end; rela++) {
+               Elf_Addr addr = (rela->r_offset - va_link_pa_offset);
+               Elf_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 >= link_addr)
+                       relocated_addr += reloc_offset;
+
+               *(Elf_Addr *)addr = relocated_addr;
+       }
+}
+#endif
+
  asmlinkage void __init setup_vm(uintptr_t dtb_pa)
  {
         uintptr_t va, end_va;
@@ -377,9 +432,20 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         uintptr_t load_sz = (uintptr_t)(&_end) - load_pa;
         uintptr_t map_size = best_map_size(load_pa, MAX_EARLY_MAPPING_SIZE);
 
-       va_pa_offset = PAGE_OFFSET - load_pa;
+       va_pa_offset = kernel_load_addr - load_pa;
         pfn_base = PFN_DOWN(load_pa);
 
+#ifdef CONFIG_RELOCATABLE
+       /*
+        * Early page table uses only one PGDIR, which makes it possible
+        * to map 1GB aligned on 1GB: if the relocation offset makes the kernel
+        * cross over a 1G boundary, raise a bug since a part of the kernel
+        * would not get mapped.
+        */
+       BUG_ON(SZ_1G - (kernel_load_addr & (SZ_1G - 1)) < load_sz);
+       relocate_kernel(load_pa);
+#endif
+
         /*
          * Enforce boot alignment requirements of RV32 and
          * RV64 by only allowing PMD or PGD mappings.
@@ -387,7 +453,7 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         BUG_ON(map_size == PAGE_SIZE);
 
         /* Sanity check alignment and size */
-       BUG_ON((PAGE_OFFSET % PGDIR_SIZE) != 0);
+       BUILD_BUG_ON((_AC(CONFIG_PAGE_OFFSET, UL) % PGDIR_SIZE) != 0);
         BUG_ON((load_pa % map_size) != 0);
         BUG_ON(load_sz > MAX_EARLY_MAPPING_SIZE);
 
@@ -400,13 +466,13 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         create_pmd_mapping(fixmap_pmd, FIXADDR_START,
                            (uintptr_t)fixmap_pte, PMD_SIZE, PAGE_TABLE);
         /* Setup trampoline PGD and PMD */
-       create_pgd_mapping(trampoline_pg_dir, PAGE_OFFSET,
+       create_pgd_mapping(trampoline_pg_dir, kernel_load_addr,
                            (uintptr_t)trampoline_pmd, PGDIR_SIZE, PAGE_TABLE);
-       create_pmd_mapping(trampoline_pmd, PAGE_OFFSET,
+       create_pmd_mapping(trampoline_pmd, kernel_load_addr,
                            load_pa, PMD_SIZE, PAGE_KERNEL_EXEC);
  #else
         /* Setup trampoline PGD */
-       create_pgd_mapping(trampoline_pg_dir, PAGE_OFFSET,
+       create_pgd_mapping(trampoline_pg_dir, kernel_load_addr,
                            load_pa, PGDIR_SIZE, PAGE_KERNEL_EXEC);
  #endif
 
@@ -415,10 +481,10 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
          * us to reach paging_init(). We map all memory banks later
          * in setup_vm_final() below.
          */
-       end_va = PAGE_OFFSET + load_sz;
-       for (va = PAGE_OFFSET; va < end_va; va += map_size)
+       end_va = kernel_load_addr + load_sz;
+       for (va = kernel_load_addr; va < end_va; va += map_size)
                 create_pgd_mapping(early_pg_dir, va,
-                                  load_pa + (va - PAGE_OFFSET),
+                                  load_pa + (va - kernel_load_addr),
                                    map_size, PAGE_KERNEL_EXEC);
 
Quoted 12 lines

arch/riscv/mm/init.c static void __init setup_vm_final(void) · line 528

Quoted 11 lines
arch/riscv/mm/init.c +3 -3
         /* Create fixed mapping for early FDT parsing */
@@ -457,9 +523,9 @@ static void __init setup_vm_final(void)
                         break;
                 if (memblock_is_nomap(reg))
                         continue;
-               if (start <= __pa(PAGE_OFFSET) &&
-                   __pa(PAGE_OFFSET) < end)
-                       start = __pa(PAGE_OFFSET);
+               if (start <= __pa(kernel_load_addr) &&
+                   __pa(kernel_load_addr) < end)
+                       start = __pa(kernel_load_addr);

Here should use __pa_symbol() instead of __pa() for kernel symbols.

I'll add it to the v2 too, thanks.

Quoted 1 line

I'm working on KASLR on top of this patch, it's work to me.

If I can help, don't hesitate, even if it is simply for testing :)

Quoted 3 lines

Thanks for that,

Alex

Quoted 6 lines
Quoted 6 lines
arch/riscv/mm/init.c +0 -1
                 map_size = best_map_size(start, end - start);
                 for (pa = start; pa < end; pa += map_size) {
--

2.20.1

#6 lore raw
Quoted 341 lines · Alex Ghiti

Hi Zong,

Quoted 313 lines · Zong Li
Quoted 312 lines · Alexandre Ghiti

This config allows to compile the kernel as PIE and to relocate it at any virtual address at runtime: this paves the way to KASLR and to 4-level page table folding at runtime. 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.

---
  arch/riscv/Kconfig              | 11 ++++
  arch/riscv/Makefile             |  5 +-
  arch/riscv/boot/loader.lds.S    |  2 +-
  arch/riscv/include/asm/page.h   |  5 +-
  arch/riscv/kernel/head.S        |  3 +-
  arch/riscv/kernel/vmlinux.lds.S | 10 ++--
  arch/riscv/mm/Makefile          |  4 ++
  arch/riscv/mm/init.c            | 92 ++++++++++++++++++++++++++++-----
  8 files changed, 110 insertions(+), 22 deletions(-)
arch/riscv/Kconfig +11 -0
index fa7dc03459e7..c652b4b850ce 100644
@@ -163,6 +163,17 @@ config PGTABLE_LEVELS
         default 3 if 64BIT
         default 2
 
+config RELOCATABLE
+       bool
+       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.
+
  source "arch/riscv/Kconfig.socs"
 
  menu "Platform type"
arch/riscv/Makefile +4 -1
index b9009a2fbaf5..5a115cf6a9c1 100644
@@ -9,7 +9,10 @@
  #
 
  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
  endif
arch/riscv/boot/loader.lds.S +1 -1
index 47a5003c2e28..a9ed218171aa 100644
@@ -7,7 +7,7 @@ ENTRY(_start)
 
  SECTIONS
  {
-       . = PAGE_OFFSET;
+       . = CONFIG_PAGE_OFFSET;
 
         .payload : {
                 *(.payload)
arch/riscv/include/asm/page.h +3 -2
index ac699246ae7e..27c95da68ecb 100644
@@ -31,9 +31,9 @@
   * When not using MMU this corresponds to the first free page in
   * physical memory (aligned on a page boundary).
   */
-#define PAGE_OFFSET            _AC(CONFIG_PAGE_OFFSET, UL)
+#define PAGE_OFFSET            kernel_load_addr
 
-#define KERN_VIRT_SIZE (-PAGE_OFFSET)
+#define KERN_VIRT_SIZE         (-_AC(CONFIG_PAGE_OFFSET, UL))
 
  #ifndef __ASSEMBLY__
 
@@ -97,6 +97,7 @@ extern unsigned long pfn_base;
  #define ARCH_PFN_OFFSET                (PAGE_OFFSET >> PAGE_SHIFT)
  #endif /* CONFIG_MMU */
 
+extern unsigned long kernel_load_addr;
  extern unsigned long max_low_pfn;
  extern unsigned long min_low_pfn;
 
arch/riscv/kernel/head.S +2 -1
index 2227db63f895..5042b2b48a06 100644
@@ -126,7 +126,8 @@ clear_bss_done:
  #ifdef CONFIG_MMU
  relocate:
         /* Relocate return address */
-       li a1, PAGE_OFFSET
+       la a1, kernel_load_addr
+       REG_L a1, 0(a1)
         la a2, _start
         sub a1, a1, a2
         add ra, ra, a1
arch/riscv/kernel/vmlinux.lds.S +6 -4
index 12f42f96d46e..5095aee7c37e 100644
@@ -4,7 +4,7 @@
   * Copyright (C) 2017 SiFive
   */
 
-#define LOAD_OFFSET PAGE_OFFSET
+#define LOAD_OFFSET CONFIG_PAGE_OFFSET
  #include <asm/vmlinux.lds.h>
  #include <asm/page.h>
  #include <asm/cache.h>
@@ -70,9 +70,11 @@ SECTIONS
 
         EXCEPTION_TABLE(0x10)
 
-       .rel.dyn : {
-               *(.rel.dyn*)
-       }
+        .rela.dyn : ALIGN(8) {
+               __rela_dyn_start = .;
+                *(.rela .rela*)
+               __rela_dyn_end = .;
+        }
 
         _end = .;
 
arch/riscv/mm/Makefile +4 -0
index a1bd95c8047a..dcd3d806243f 100644
@@ -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 = -pg
  endif
arch/riscv/mm/init.c +76 -10
index 965a8cf4829c..ac9a9f69abc0 100644
@@ -12,6 +12,9 @@
  #include <linux/sizes.h>
  #include <linux/of_fdt.h>
  #include <linux/libfdt.h>
+#ifdef CONFIG_RELOCATABLE
+#include <linux/elf.h>
+#endif
 
  #include <asm/fixmap.h>
  #include <asm/tlbflush.h>
@@ -28,6 +31,9 @@ EXPORT_SYMBOL(empty_zero_page);
  extern char _start[];
  void *dtb_early_va;
 
+unsigned long kernel_load_addr = _AC(CONFIG_PAGE_OFFSET, UL);
+EXPORT_SYMBOL(kernel_load_addr);
+
  static void __init zone_sizes_init(void)
  {
         unsigned long max_zone_pfns[MAX_NR_ZONES] = { 0, };
@@ -132,7 +138,8 @@ void __init setup_bootmem(void)
                 phys_addr_t end = reg->base + reg->size;
 
                 if (reg->base <= vmlinux_end && vmlinux_end <= end) {
-                       mem_size = min(reg->size, (phys_addr_t)-PAGE_OFFSET);
+                       mem_size = min(reg->size,
+                                      (phys_addr_t)-kernel_load_addr);
 
                         /*
                          * Remove memblock from the end of usable area to the
@@ -269,7 +276,7 @@ static phys_addr_t __init alloc_pmd(uintptr_t va)
         if (mmu_enabled)
                 return memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
 
-       pmd_num = (va - PAGE_OFFSET) >> PGDIR_SHIFT;
+       pmd_num = (va - kernel_load_addr) >> PGDIR_SHIFT;
         BUG_ON(pmd_num >= NUM_EARLY_PMDS);
         return (uintptr_t)&early_pmd[pmd_num * PTRS_PER_PMD];
  }
@@ -370,6 +377,54 @@ static uintptr_t __init best_map_size(phys_addr_t base, phys_addr_t size)
  #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;
+
+#ifdef CONFIG_64BIT
+#define Elf_Rela Elf64_Rela
+#define Elf_Addr Elf64_Addr
+#else
+#define Elf_Rela Elf32_Rela
+#define Elf_Addr Elf32_Addr
+#endif
+
+void __init relocate_kernel(uintptr_t load_pa)
+{
+       Elf_Rela *rela = (Elf_Rela *)&__rela_dyn_start;
+       uintptr_t link_addr = _AC(CONFIG_PAGE_OFFSET, UL);
+       /*
+        * This holds the offset between the linked virtual address and the
+        * relocated virtual address.
+        */
+       uintptr_t reloc_offset = kernel_load_addr - link_addr;
+       /*
+        * This holds the offset between linked virtual address and physical
+        * address whereas va_pa_offset holds the offset between relocated
+        * virtual address and physical address.
+        */
+       uintptr_t va_link_pa_offset = link_addr - load_pa;
+
+       for ( ; rela < (Elf_Rela *)&__rela_dyn_end; rela++) {
+               Elf_Addr addr = (rela->r_offset - va_link_pa_offset);
+               Elf_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 >= link_addr)
+                       relocated_addr += reloc_offset;
+
+               *(Elf_Addr *)addr = relocated_addr;
+       }
+}
+#endif
+
  asmlinkage void __init setup_vm(uintptr_t dtb_pa)
  {
         uintptr_t va, end_va;
@@ -377,9 +432,20 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         uintptr_t load_sz = (uintptr_t)(&_end) - load_pa;
         uintptr_t map_size = best_map_size(load_pa, MAX_EARLY_MAPPING_SIZE);
 
-       va_pa_offset = PAGE_OFFSET - load_pa;
+       va_pa_offset = kernel_load_addr - load_pa;
         pfn_base = PFN_DOWN(load_pa);
 
+#ifdef CONFIG_RELOCATABLE
+       /*
+        * Early page table uses only one PGDIR, which makes it possible
+        * to map 1GB aligned on 1GB: if the relocation offset makes the kernel
+        * cross over a 1G boundary, raise a bug since a part of the kernel
+        * would not get mapped.
+        */
+       BUG_ON(SZ_1G - (kernel_load_addr & (SZ_1G - 1)) < load_sz);
+       relocate_kernel(load_pa);
+#endif
+
         /*
          * Enforce boot alignment requirements of RV32 and
          * RV64 by only allowing PMD or PGD mappings.
@@ -387,7 +453,7 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         BUG_ON(map_size == PAGE_SIZE);
 
         /* Sanity check alignment and size */
-       BUG_ON((PAGE_OFFSET % PGDIR_SIZE) != 0);
+       BUILD_BUG_ON((_AC(CONFIG_PAGE_OFFSET, UL) % PGDIR_SIZE) != 0);
         BUG_ON((load_pa % map_size) != 0);
         BUG_ON(load_sz > MAX_EARLY_MAPPING_SIZE);
 
@@ -400,13 +466,13 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         create_pmd_mapping(fixmap_pmd, FIXADDR_START,
                            (uintptr_t)fixmap_pte, PMD_SIZE, PAGE_TABLE);
         /* Setup trampoline PGD and PMD */
-       create_pgd_mapping(trampoline_pg_dir, PAGE_OFFSET,
+       create_pgd_mapping(trampoline_pg_dir, kernel_load_addr,
                            (uintptr_t)trampoline_pmd, PGDIR_SIZE, PAGE_TABLE);
-       create_pmd_mapping(trampoline_pmd, PAGE_OFFSET,
+       create_pmd_mapping(trampoline_pmd, kernel_load_addr,
                            load_pa, PMD_SIZE, PAGE_KERNEL_EXEC);
  #else
         /* Setup trampoline PGD */
-       create_pgd_mapping(trampoline_pg_dir, PAGE_OFFSET,
+       create_pgd_mapping(trampoline_pg_dir, kernel_load_addr,
                            load_pa, PGDIR_SIZE, PAGE_KERNEL_EXEC);
  #endif
 
@@ -415,10 +481,10 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
          * us to reach paging_init(). We map all memory banks later
          * in setup_vm_final() below.
          */
-       end_va = PAGE_OFFSET + load_sz;
-       for (va = PAGE_OFFSET; va < end_va; va += map_size)
+       end_va = kernel_load_addr + load_sz;
+       for (va = kernel_load_addr; va < end_va; va += map_size)
                 create_pgd_mapping(early_pg_dir, va,
-                                  load_pa + (va - PAGE_OFFSET),
+                                  load_pa + (va - kernel_load_addr),
                                    map_size, PAGE_KERNEL_EXEC);
 
Quoted 12 lines

arch/riscv/mm/init.c static void __init setup_vm_final(void) · line 528

Quoted 11 lines
arch/riscv/mm/init.c +3 -3
         /* Create fixed mapping for early FDT parsing */
@@ -457,9 +523,9 @@ static void __init setup_vm_final(void)
                         break;
                 if (memblock_is_nomap(reg))
                         continue;
-               if (start <= __pa(PAGE_OFFSET) &&
-                   __pa(PAGE_OFFSET) < end)
-                       start = __pa(PAGE_OFFSET);
+               if (start <= __pa(kernel_load_addr) &&
+                   __pa(kernel_load_addr) < end)
+                       start = __pa(kernel_load_addr);

Here should use __pa_symbol() instead of __pa() for kernel symbols.

I'll add it to the v2 too, thanks.

Quoted 1 line

I'm working on KASLR on top of this patch, it's work to me.

If I can help, don't hesitate, even if it is simply for testing :)

Quoted 1 line
Quoted 15 lines
Quoted 2 lines

Thanks for that,

Alex

Quoted 6 lines
Quoted 6 lines
arch/riscv/mm/init.c +0 -1
                 map_size = best_map_size(start, end - start);
                 for (pa = start; pa < end; pa += map_size) {
--

2.20.1

Hi all,

I find that there is a risk when building kernel as PIE. The problem is that hart_lottery will be put in .bss section rather than .sdata section, so the value of hart_lottery will be cleared by main hart. I post a patch to mark hart_lottery as sdata section by using gcc attribution.

Please refer to https://patchwork.kernel.org/patch/11364457/.

Feel free to give any feedbacks and comments. Thanks.

#7 lore raw
Quoted 361 lines · Zong Li
Quoted 341 lines · Alex Ghiti

Hi Zong,

Quoted 313 lines · Zong Li
Quoted 312 lines · Alexandre Ghiti

This config allows to compile the kernel as PIE and to relocate it at any virtual address at runtime: this paves the way to KASLR and to 4-level page table folding at runtime. 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.

---
  arch/riscv/Kconfig              | 11 ++++
  arch/riscv/Makefile             |  5 +-
  arch/riscv/boot/loader.lds.S    |  2 +-
  arch/riscv/include/asm/page.h   |  5 +-
  arch/riscv/kernel/head.S        |  3 +-
  arch/riscv/kernel/vmlinux.lds.S | 10 ++--
  arch/riscv/mm/Makefile          |  4 ++
  arch/riscv/mm/init.c            | 92 ++++++++++++++++++++++++++++-----
  8 files changed, 110 insertions(+), 22 deletions(-)
arch/riscv/Kconfig +11 -0
index fa7dc03459e7..c652b4b850ce 100644
@@ -163,6 +163,17 @@ config PGTABLE_LEVELS
         default 3 if 64BIT
         default 2
 
+config RELOCATABLE
+       bool
+       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.
+
  source "arch/riscv/Kconfig.socs"
 
  menu "Platform type"
arch/riscv/Makefile +4 -1
index b9009a2fbaf5..5a115cf6a9c1 100644
@@ -9,7 +9,10 @@
  #
 
  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
  endif
arch/riscv/boot/loader.lds.S +1 -1
index 47a5003c2e28..a9ed218171aa 100644
@@ -7,7 +7,7 @@ ENTRY(_start)
 
  SECTIONS
  {
-       . = PAGE_OFFSET;
+       . = CONFIG_PAGE_OFFSET;
 
         .payload : {
                 *(.payload)
arch/riscv/include/asm/page.h +3 -2
index ac699246ae7e..27c95da68ecb 100644
@@ -31,9 +31,9 @@
   * When not using MMU this corresponds to the first free page in
   * physical memory (aligned on a page boundary).
   */
-#define PAGE_OFFSET            _AC(CONFIG_PAGE_OFFSET, UL)
+#define PAGE_OFFSET            kernel_load_addr
 
-#define KERN_VIRT_SIZE (-PAGE_OFFSET)
+#define KERN_VIRT_SIZE         (-_AC(CONFIG_PAGE_OFFSET, UL))
 
  #ifndef __ASSEMBLY__
 
@@ -97,6 +97,7 @@ extern unsigned long pfn_base;
  #define ARCH_PFN_OFFSET                (PAGE_OFFSET >> PAGE_SHIFT)
  #endif /* CONFIG_MMU */
 
+extern unsigned long kernel_load_addr;
  extern unsigned long max_low_pfn;
  extern unsigned long min_low_pfn;
 
arch/riscv/kernel/head.S +2 -1
index 2227db63f895..5042b2b48a06 100644
@@ -126,7 +126,8 @@ clear_bss_done:
  #ifdef CONFIG_MMU
  relocate:
         /* Relocate return address */
-       li a1, PAGE_OFFSET
+       la a1, kernel_load_addr
+       REG_L a1, 0(a1)
         la a2, _start
         sub a1, a1, a2
         add ra, ra, a1
arch/riscv/kernel/vmlinux.lds.S +6 -4
index 12f42f96d46e..5095aee7c37e 100644
@@ -4,7 +4,7 @@
   * Copyright (C) 2017 SiFive
   */
 
-#define LOAD_OFFSET PAGE_OFFSET
+#define LOAD_OFFSET CONFIG_PAGE_OFFSET
  #include <asm/vmlinux.lds.h>
  #include <asm/page.h>
  #include <asm/cache.h>
@@ -70,9 +70,11 @@ SECTIONS
 
         EXCEPTION_TABLE(0x10)
 
-       .rel.dyn : {
-               *(.rel.dyn*)
-       }
+        .rela.dyn : ALIGN(8) {
+               __rela_dyn_start = .;
+                *(.rela .rela*)
+               __rela_dyn_end = .;
+        }
 
         _end = .;
 
arch/riscv/mm/Makefile +4 -0
index a1bd95c8047a..dcd3d806243f 100644
@@ -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 = -pg
  endif
arch/riscv/mm/init.c +76 -10
index 965a8cf4829c..ac9a9f69abc0 100644
@@ -12,6 +12,9 @@
  #include <linux/sizes.h>
  #include <linux/of_fdt.h>
  #include <linux/libfdt.h>
+#ifdef CONFIG_RELOCATABLE
+#include <linux/elf.h>
+#endif
 
  #include <asm/fixmap.h>
  #include <asm/tlbflush.h>
@@ -28,6 +31,9 @@ EXPORT_SYMBOL(empty_zero_page);
  extern char _start[];
  void *dtb_early_va;
 
+unsigned long kernel_load_addr = _AC(CONFIG_PAGE_OFFSET, UL);
+EXPORT_SYMBOL(kernel_load_addr);
+
  static void __init zone_sizes_init(void)
  {
         unsigned long max_zone_pfns[MAX_NR_ZONES] = { 0, };
@@ -132,7 +138,8 @@ void __init setup_bootmem(void)
                 phys_addr_t end = reg->base + reg->size;
 
                 if (reg->base <= vmlinux_end && vmlinux_end <= end) {
-                       mem_size = min(reg->size, (phys_addr_t)-PAGE_OFFSET);
+                       mem_size = min(reg->size,
+                                      (phys_addr_t)-kernel_load_addr);
 
                         /*
                          * Remove memblock from the end of usable area to the
@@ -269,7 +276,7 @@ static phys_addr_t __init alloc_pmd(uintptr_t va)
         if (mmu_enabled)
                 return memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
 
-       pmd_num = (va - PAGE_OFFSET) >> PGDIR_SHIFT;
+       pmd_num = (va - kernel_load_addr) >> PGDIR_SHIFT;
         BUG_ON(pmd_num >= NUM_EARLY_PMDS);
         return (uintptr_t)&early_pmd[pmd_num * PTRS_PER_PMD];
  }
@@ -370,6 +377,54 @@ static uintptr_t __init best_map_size(phys_addr_t base, phys_addr_t size)
  #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;
+
+#ifdef CONFIG_64BIT
+#define Elf_Rela Elf64_Rela
+#define Elf_Addr Elf64_Addr
+#else
+#define Elf_Rela Elf32_Rela
+#define Elf_Addr Elf32_Addr
+#endif
+
+void __init relocate_kernel(uintptr_t load_pa)
+{
+       Elf_Rela *rela = (Elf_Rela *)&__rela_dyn_start;
+       uintptr_t link_addr = _AC(CONFIG_PAGE_OFFSET, UL);
+       /*
+        * This holds the offset between the linked virtual address and the
+        * relocated virtual address.
+        */
+       uintptr_t reloc_offset = kernel_load_addr - link_addr;
+       /*
+        * This holds the offset between linked virtual address and physical
+        * address whereas va_pa_offset holds the offset between relocated
+        * virtual address and physical address.
+        */
+       uintptr_t va_link_pa_offset = link_addr - load_pa;
+
+       for ( ; rela < (Elf_Rela *)&__rela_dyn_end; rela++) {
+               Elf_Addr addr = (rela->r_offset - va_link_pa_offset);
+               Elf_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 >= link_addr)
+                       relocated_addr += reloc_offset;
+
+               *(Elf_Addr *)addr = relocated_addr;
+       }
+}
+#endif
+
  asmlinkage void __init setup_vm(uintptr_t dtb_pa)
  {
         uintptr_t va, end_va;
@@ -377,9 +432,20 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         uintptr_t load_sz = (uintptr_t)(&_end) - load_pa;
         uintptr_t map_size = best_map_size(load_pa, MAX_EARLY_MAPPING_SIZE);
 
-       va_pa_offset = PAGE_OFFSET - load_pa;
+       va_pa_offset = kernel_load_addr - load_pa;
         pfn_base = PFN_DOWN(load_pa);
 
+#ifdef CONFIG_RELOCATABLE
+       /*
+        * Early page table uses only one PGDIR, which makes it possible
+        * to map 1GB aligned on 1GB: if the relocation offset makes the kernel
+        * cross over a 1G boundary, raise a bug since a part of the kernel
+        * would not get mapped.
+        */
+       BUG_ON(SZ_1G - (kernel_load_addr & (SZ_1G - 1)) < load_sz);
+       relocate_kernel(load_pa);
+#endif
+
         /*
          * Enforce boot alignment requirements of RV32 and
          * RV64 by only allowing PMD or PGD mappings.
@@ -387,7 +453,7 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         BUG_ON(map_size == PAGE_SIZE);
 
         /* Sanity check alignment and size */
-       BUG_ON((PAGE_OFFSET % PGDIR_SIZE) != 0);
+       BUILD_BUG_ON((_AC(CONFIG_PAGE_OFFSET, UL) % PGDIR_SIZE) != 0);
         BUG_ON((load_pa % map_size) != 0);
         BUG_ON(load_sz > MAX_EARLY_MAPPING_SIZE);
 
@@ -400,13 +466,13 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
         create_pmd_mapping(fixmap_pmd, FIXADDR_START,
                            (uintptr_t)fixmap_pte, PMD_SIZE, PAGE_TABLE);
         /* Setup trampoline PGD and PMD */
-       create_pgd_mapping(trampoline_pg_dir, PAGE_OFFSET,
+       create_pgd_mapping(trampoline_pg_dir, kernel_load_addr,
                            (uintptr_t)trampoline_pmd, PGDIR_SIZE, PAGE_TABLE);
-       create_pmd_mapping(trampoline_pmd, PAGE_OFFSET,
+       create_pmd_mapping(trampoline_pmd, kernel_load_addr,
                            load_pa, PMD_SIZE, PAGE_KERNEL_EXEC);
  #else
         /* Setup trampoline PGD */
-       create_pgd_mapping(trampoline_pg_dir, PAGE_OFFSET,
+       create_pgd_mapping(trampoline_pg_dir, kernel_load_addr,
                            load_pa, PGDIR_SIZE, PAGE_KERNEL_EXEC);
  #endif
 
@@ -415,10 +481,10 @@ asmlinkage void __init setup_vm(uintptr_t dtb_pa)
          * us to reach paging_init(). We map all memory banks later
          * in setup_vm_final() below.
          */
-       end_va = PAGE_OFFSET + load_sz;
-       for (va = PAGE_OFFSET; va < end_va; va += map_size)
+       end_va = kernel_load_addr + load_sz;
+       for (va = kernel_load_addr; va < end_va; va += map_size)
                 create_pgd_mapping(early_pg_dir, va,
-                                  load_pa + (va - PAGE_OFFSET),
+                                  load_pa + (va - kernel_load_addr),
                                    map_size, PAGE_KERNEL_EXEC);
 
Quoted 12 lines

arch/riscv/mm/init.c static void __init setup_vm_final(void) · line 528

Quoted 11 lines
arch/riscv/mm/init.c +3 -3
         /* Create fixed mapping for early FDT parsing */
@@ -457,9 +523,9 @@ static void __init setup_vm_final(void)
                         break;
                 if (memblock_is_nomap(reg))
                         continue;
-               if (start <= __pa(PAGE_OFFSET) &&
-                   __pa(PAGE_OFFSET) < end)
-                       start = __pa(PAGE_OFFSET);
+               if (start <= __pa(kernel_load_addr) &&
+                   __pa(kernel_load_addr) < end)
+                       start = __pa(kernel_load_addr);

Here should use __pa_symbol() instead of __pa() for kernel symbols.

I'll add it to the v2 too, thanks.

Quoted 1 line

I'm working on KASLR on top of this patch, it's work to me.

If I can help, don't hesitate, even if it is simply for testing :)

Quoted 1 line
Quoted 15 lines
Quoted 2 lines

Thanks for that,

Alex

Quoted 6 lines
Quoted 6 lines
arch/riscv/mm/init.c +0 -1
                 map_size = best_map_size(start, end - start);
                 for (pa = start; pa < end; pa += map_size) {
--

2.20.1

Hi all,

Quoted 7 lines

I find that there is a risk when building kernel as PIE. The problem is that hart_lottery will be put in .bss section rather than .sdata section, so the value of hart_lottery will be cleared by main hart. I post a patch to mark hart_lottery as sdata section by using gcc attribution.

Please refer to https://patchwork.kernel.org/patch/11364457/.

The CONFIG_RELOCATION has performance impact hence it is will not be enabled by default.

Also, we are moving away from fragile lottery based HART bring up to SBI HSM based HART bring up. I am sure by next year the lottery based HART bring-up will be obsolete and eventually we might even remove it altogether from Linux RISC-V kernel.

Regards,
Anup