RISC-V: Make modules loadable.
cover RISC-V: Make modules loadable.
This was enough to load enough modules to make virtio_scsi devices appear on the qemu virt machine on a highly modular distro kernel.
PATCH 1/9 RISC-V: Build non position-independent modules.
The relocations we have to handle for no-PIC are much simpler, and we can't take advantage of GOT/PLT sharing anyway.
--- arch/riscv/Makefile | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
arch/riscv/Makefile +2 -2
@@ -11,8 +11,8 @@ LDFLAGS := OBJCOPYFLAGS := -O binary LDFLAGS_vmlinux := -KBUILD_AFLAGS_MODULE += -fPIC -KBUILD_CFLAGS_MODULE += -fPIC +KBUILD_AFLAGS_MODULE += -fno-PIC -mcmodel=medany +KBUILD_CFLAGS_MODULE += -fno-PIC -mcmodel=medany KBUILD_DEFCONFIG = defconfig
Signature · 1 line
2.16.1
PATCH 2/9 RISC-V: Load modules within relative jump range of the kernel text.
In reply to Shea Levy--- arch/riscv/include/asm/pgtable.h | 9 +++++++++ arch/riscv/kernel/module.c | 11 +++++++++++ 2 files changed, 20 insertions(+)
arch/riscv/include/asm/pgtable.h +9 -0
@@ -25,6 +25,7 @@#include <asm/page.h> #include <asm/tlbflush.h> #include <linux/mm_types.h> +#include <linux/sizes.h> #ifdef CONFIG_64BIT #include <asm/pgtable-64.h>
@@ -425,6 +426,14 @@ static inline void pgtable_cache_init(void)#define TASK_SIZE VMALLOC_START#endif +/* +* The module space lives between the addresses given by TASK_SIZE +* and PAGE_OFFSET - it must be within 2G of the kernel text. +*/ +#define MODULES_SIZE (SZ_128M) +#define MODULES_VADDR (PAGE_OFFSET - MODULES_SIZE) +#define MODULES_END (VMALLOC_END) +#include <asm-generic/pgtable.h> #endif /* !__ASSEMBLY__ */
arch/riscv/kernel/module.c +11 -0
@@ -16,6 +16,8 @@#include <linux/err.h> #include <linux/errno.h> #include <linux/moduleloader.h> +#include <linux/vmalloc.h> +#include <asm/pgtable.h> static int apply_r_riscv_64_rela (struct module *me,u32 *location,Elf_Addr v) {
@@ -215,3 +217,12 @@ int apply_relocate_add(Elf_Shdr *sechdrs, const char *strtab,return 0 ; } + +void *module_alloc (unsigned long size) +{ +return __vmalloc_node_range (size,1 ,MODULES_VADDR , +MODULES_END ,GFP_KERNEL , +PAGE_KERNEL_EXEC ,0 , +NUMA_NO_NODE , +__builtin_return_address (0 )); +}
Signature · 1 line
2.16.1
PATCH 2/9 RISC-V: Load modules within relative jump range of the kernel text.
Quoted 49 lines · shea at shealevy.com
Signed-off-by Shea Levy <[email protected]>--- arch/riscv/include/asm/pgtable.h | 9 +++++++++ arch/riscv/kernel/module.c | 11 +++++++++++ 2 files changed, 20 insertions(+)arch/riscv/include/asm/pgtable.h +9
-0@@ -25,6 +25,7 @@#include <asm/page.h> #include <asm/tlbflush.h> #include <linux/mm_types.h> +#include <linux/sizes.h> #ifdef CONFIG_64BIT #include <asm/pgtable-64.h> @@ -425,6 +426,14 @@ static inline void pgtable_cache_init(void)#define TASK_SIZE VMALLOC_START#endif +/* +* The module space lives between the addresses given by TASK_SIZE +* and PAGE_OFFSET - it must be within 2G of the kernel text. +*/ +#define MODULES_SIZE (SZ_128M) +#define MODULES_VADDR (PAGE_OFFSET - MODULES_SIZE) +#define MODULES_END (VMALLOC_END) +#include <asm-generic/pgtable.h> #endif /* !__ASSEMBLY__ */ arch/riscv/kernel/module.c +2
-0@@ -16,6 +16,8 @@#include <linux/err.h> #include <linux/errno.h> #include <linux/moduleloader.h> +#include <linux/vmalloc.h> +#include <asm/pgtable.h> static int apply_r_riscv_64_rela (struct module *me,u32 *location,Elf_Addr v) {@@ -215,3 +217,12 @@ int apply_relocate_add(Elf_Shdr *sechdrs, const char *strtab,return 0 ;
arch/riscv/kernel/module.c int apply_relocate_add(Elf_Shdr *sechdrs, const char *strtab, · line 228
Quoted 10 lines
arch/riscv/kernel/module.c +9
-0} + +void *module_alloc (unsigned long size) +{ +return __vmalloc_node_range (size,1 ,MODULES_VADDR , +MODULES_END ,GFP_KERNEL , +PAGE_KERNEL_EXEC ,0 , +NUMA_NO_NODE , +__builtin_return_address (0 )); +}
arm64 aligns their modules to a page boundary. I think we need to align them to at least whatever the linker is expecting, which should be recorded in each segment in the ELF that's provided.
This code otherwise looks exactly teh same as MIPS, maybe we should make it generic somewhere? It's not that much, though, but it looks like it'll get bigger as things support KASLR and such.
PATCH 2/9 RISC-V: Load modules within relative jump range of the kernel text.
Hi Palmer,
Quoted 65 lines · Palmer Dabbelt
Quoted 49 lines · shea at shealevy.com
Signed-off-by Shea Levy <[email protected]>--- arch/riscv/include/asm/pgtable.h | 9 +++++++++ arch/riscv/kernel/module.c | 11 +++++++++++ 2 files changed, 20 insertions(+)arch/riscv/include/asm/pgtable.h +9
-0@@ -25,6 +25,7 @@#include <asm/page.h> #include <asm/tlbflush.h> #include <linux/mm_types.h> +#include <linux/sizes.h> #ifdef CONFIG_64BIT #include <asm/pgtable-64.h> @@ -425,6 +426,14 @@ static inline void pgtable_cache_init(void)#define TASK_SIZE VMALLOC_START#endif +/* +* The module space lives between the addresses given by TASK_SIZE +* and PAGE_OFFSET - it must be within 2G of the kernel text. +*/ +#define MODULES_SIZE (SZ_128M) +#define MODULES_VADDR (PAGE_OFFSET - MODULES_SIZE) +#define MODULES_END (VMALLOC_END) +#include <asm-generic/pgtable.h> #endif /* !__ASSEMBLY__ */ arch/riscv/kernel/module.c +2
-0@@ -16,6 +16,8 @@#include <linux/err.h> #include <linux/errno.h> #include <linux/moduleloader.h> +#include <linux/vmalloc.h> +#include <asm/pgtable.h> static int apply_r_riscv_64_rela (struct module *me,u32 *location,Elf_Addr v) {@@ -215,3 +217,12 @@ int apply_relocate_add(Elf_Shdr *sechdrs, const char *strtab,return 0 ;arch/riscv/kernel/module.c int apply_relocate_add(Elf_Shdr *sechdrs, const char *strtab, · line 228
Quoted 10 lines
arch/riscv/kernel/module.c +9
-0} + +void *module_alloc (unsigned long size) +{ +return __vmalloc_node_range (size,1 ,MODULES_VADDR , +MODULES_END ,GFP_KERNEL , +PAGE_KERNEL_EXEC ,0 , +NUMA_NO_NODE , +__builtin_return_address (0 )); +}arm64 aligns their modules to a page boundary. I think we need to align them to at least whatever the linker is expecting, which should be recorded in each segment in the ELF that's provided.
Quoted 3 lines
This code otherwise looks exactly teh same as MIPS, maybe we should make it generic somewhere? It's not that much, though, but it looks like it'll get bigger as things support KASLR and such.
Thanks, I'll include this fix when I resolve the RISCV_ALIGN issue, and
see if it makes sense to make generic.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 832 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-riscv/attachments/20180301/56fddeab/attachment.sig>
PATCH 3/9 RISC-V: Support RISCV_CALL relocations for non-PIC modules.
In reply to Shea LevyOnly PIC objects have RISCV_CALL_PLT relocations.
--- arch/riscv/kernel/module.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
arch/riscv/kernel/module.c +2 -2
@@ -94,7 +94,7 @@ static int apply_r_riscv_pcrel_lo12_s_rela(struct module *me, u32 *location,return 0 ; } -static int apply_r_riscv_call_plt_rela (struct module *me,u32 *location, +static int apply_r_riscv_call_rela (struct module *me,u32 *location,Elf_Addr v) {s64 offset = (void *)v - (void *)location;
@@ -129,7 +129,7 @@ static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, [R_RISCV_PCREL_HI20] = apply_r_riscv_pcrel_hi20_rela, [R_RISCV_PCREL_LO12_I] = apply_r_riscv_pcrel_lo12_i_rela, [R_RISCV_PCREL_LO12_S] = apply_r_riscv_pcrel_lo12_s_rela, - [R_RISCV_CALL_PLT] = apply_r_riscv_call_plt_rela, + [R_RISCV_CALL] = apply_r_riscv_call_rela, [R_RISCV_RELAX] = apply_r_riscv_relax_rela, };
Signature · 1 line
2.16.1
PATCH 3/9 RISC-V: Support RISCV_CALL relocations for non-PIC modules.
Quoted 12 lines · shea at shealevy.com
Only PIC objects have
RISCV_CALL_PLTrelocations.Signed-off-by Shea Levy <[email protected]>--- arch/riscv/kernel/module.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)arch/riscv/kernel/module.c +0
-0
arch/riscv/kernel/module.c static int apply_r_riscv_pcrel_lo12_s_rela(struct module *me, u32 *location, · line 98
Quoted 6 lines
arch/riscv/kernel/module.c +1
-1@@ -94,7 +94,7 @@ static int apply_r_riscv_pcrel_lo12_s_rela(struct module *me, u32 *location,return 0 ; } -static int apply_r_riscv_call_plt_rela (struct module *me,u32 *location, +static int apply_r_riscv_call_rela (struct module *me,u32 *location,Elf_Addr v)
This is misaligned, but don't bother submitting another patch set just for this. I'll fix it up.
Quoted 10 lines
arch/riscv/kernel/module.c +1
-1{s64 offset = (void *)v - (void *)location;@@ -129,7 +129,7 @@ static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, [R_RISCV_PCREL_HI20] = apply_r_riscv_pcrel_hi20_rela, [R_RISCV_PCREL_LO12_I] = apply_r_riscv_pcrel_lo12_i_rela, [R_RISCV_PCREL_LO12_S] = apply_r_riscv_pcrel_lo12_s_rela, - [R_RISCV_CALL_PLT] = apply_r_riscv_call_plt_rela, + [R_RISCV_CALL] = apply_r_riscv_call_rela, [R_RISCV_RELAX] = apply_r_riscv_relax_rela, };
PATCH 4/9 RISC-V: Support RISCV_RVC_BRANCH relocations in modules.
In reply to Shea Levy--- arch/riscv/kernel/module.c | 17 +++++++++++++++++ 1 file changed, 17 insertions(+)
arch/riscv/kernel/module.c +17 -0
@@ -38,6 +38,22 @@ static int apply_r_riscv_branch_rela(struct module *me, u32 *location,return 0 ; } +static int apply_r_riscv_rvc_branch_rela (struct module *me,u32 *location, +Elf_Addr v) +{ +s64 offset = (void *)v - (void *)location; +u32 imm8 = (offset &0x100 ) << (12 -8 ); +u32 imm7_6 = (offset &0xc0 ) >> (7 -6 ); +u32 imm5 = (offset &0x20 ) >> (5 -2 ); +u32 imm4_3 = (offset &0x18 ) << (11 -4 ); +u32 imm2_1 = (offset &0x6 ) << (4 -2 ); + +u16 *short_location = (u16 *) location; + + *short_location = (*short_location &0xe383 ) | imm8 | imm7_6 | imm5 | imm4_3 | imm2_1; +return 0 ; +} +static int apply_r_riscv_jal_rela (struct module *me,u32 *location,Elf_Addr v) {
@@ -125,6 +141,7 @@ static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, Elf_Addr v) = { [R_RISCV_64 ] = apply_r_riscv_64_rela, [R_RISCV_BRANCH ] = apply_r_riscv_branch_rela, + [R_RISCV_RVC_BRANCH ] = apply_r_riscv_rvc_branch_rela, [R_RISCV_JAL ] = apply_r_riscv_jal_rela, [R_RISCV_PCREL_HI20 ] = apply_r_riscv_pcrel_hi20_rela, [R_RISCV_PCREL_LO12_I ] = apply_r_riscv_pcrel_lo12_i_rela,
Signature · 1 line
2.16.1
PATCH 5/9 RISC-V: Support RISCV_RVC_JUMP relocations in modules.
In reply to Shea Levy--- arch/riscv/kernel/module.c | 20 ++++++++++++++++++++ 1 file changed, 20 insertions(+)
arch/riscv/kernel/module.c +20 -0
@@ -67,6 +67,25 @@ static int apply_r_riscv_jal_rela(struct module *me, u32 *location,return 0 ; } +static int apply_r_riscv_rvc_jump_rela (struct module *me,u32 *location, +Elf_Addr v) +{ +s64 offset = (void *)v - (void *)location; +u32 imm11 = (offset &0x800 ) << (12 -11 ); +u32 imm10 = (offset &0x400 ) >> (10 -8 ); +u32 imm9_8 = (offset &0x300 ) << (10 -9 ); +u32 imm7 = (offset &0x80 ) >> (7 -6 ); +u32 imm6 = (offset &0x40 ) << (7 -6 ); +u32 imm5 = (offset &0x20 ) >> (5 -2 ); +u32 imm4 = (offset &0x10 ) << (11 -4 ); +u32 imm3_1 = (offset &0xe ) << (5 -3 ); + +u16 *short_location = (u16 *) location; + + *short_location = (*short_location &0xe003 ) | imm11 | imm10 | imm9_8 | imm7 | imm6 | imm5 | imm4 | imm3_1; +return 0 ; +} +static int apply_r_riscv_pcrel_hi20_rela (struct module *me,u32 *location,Elf_Addr v) {
@@ -143,6 +162,7 @@ static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, [R_RISCV_BRANCH] = apply_r_riscv_branch_rela, [R_RISCV_RVC_BRANCH] = apply_r_riscv_rvc_branch_rela, [R_RISCV_JAL] = apply_r_riscv_jal_rela, + [R_RISCV_RVC_JUMP] = apply_r_riscv_rvc_jump_rela, [R_RISCV_PCREL_HI20] = apply_r_riscv_pcrel_hi20_rela, [R_RISCV_PCREL_LO12_I] = apply_r_riscv_pcrel_lo12_i_rela, [R_RISCV_PCREL_LO12_S] = apply_r_riscv_pcrel_lo12_s_rela,
Signature · 1 line
2.16.1
PATCH 6/9 RISC-V: Support RISCV_ADD32 relocations in modules.
In reply to Shea Levy--- arch/riscv/kernel/module.c | 8 ++++++++ 1 file changed, 8 insertions(+)
arch/riscv/kernel/module.c +8 -0
@@ -156,6 +156,13 @@ static int apply_r_riscv_relax_rela(struct module *me, u32 *location,return 0 ; } +static int apply_r_riscv_add32_rela (struct module *me,u32 *location, +Elf_Addr v) +{ + *location += v; +return 0 ; +} +static int (*reloc_handlers_rela[]) (struct module *me,u32 *location,Elf_Addr v) = { [R_RISCV_64 ] = apply_r_riscv_64_rela,
@@ -168,6 +175,7 @@ static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, [R_RISCV_PCREL_LO12_S] = apply_r_riscv_pcrel_lo12_s_rela, [R_RISCV_CALL] = apply_r_riscv_call_rela, [R_RISCV_RELAX] = apply_r_riscv_relax_rela, + [R_RISCV_ADD32] = apply_r_riscv_add32_rela, }; int apply_relocate_add(Elf_Shdr *sechdrs, const char *strtab,
Signature · 1 line
2.16.1
PATCH 7/9 RISC-V: Support RISCV_SUB32 relocations in modules.
In reply to Shea Levy--- arch/riscv/kernel/module.c | 8 ++++++++ 1 file changed, 8 insertions(+)
arch/riscv/kernel/module.c +8 -0
@@ -163,6 +163,13 @@ static int apply_r_riscv_add32_rela(struct module *me, u32 *location,return 0 ; } +static int apply_r_riscv_sub32_rela (struct module *me,u32 *location, +Elf_Addr v) +{ + *location -= v; +return 0 ; +} +static int (*reloc_handlers_rela[]) (struct module *me,u32 *location,Elf_Addr v) = { [R_RISCV_64 ] = apply_r_riscv_64_rela,
@@ -176,6 +183,7 @@ static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, [R_RISCV_CALL] = apply_r_riscv_call_rela, [R_RISCV_RELAX] = apply_r_riscv_relax_rela, [R_RISCV_ADD32] = apply_r_riscv_add32_rela, + [R_RISCV_SUB32] = apply_r_riscv_sub32_rela, }; int apply_relocate_add(Elf_Shdr *sechdrs, const char *strtab,
Signature · 1 line
2.16.1
PATCH 8/9 RISC-V: Support RISCV_ALIGN relocations in modules.
In reply to Shea LevyWhile legal, this noop implementation may be inefficient.
--- arch/riscv/kernel/module.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)
arch/riscv/kernel/module.c +4 -3
@@ -150,8 +150,8 @@ static int apply_r_riscv_call_rela(struct module *me, u32 *location,return 0 ; } -static int apply_r_riscv_relax_rela (struct module *me,u32 *location, -Elf_Addr v) +static int apply_r_riscv_noop_rela (struct module *me,u32 *location, +Elf_Addr v) {return 0 ; }
@@ -181,7 +181,8 @@ static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, [R_RISCV_PCREL_LO12_I] = apply_r_riscv_pcrel_lo12_i_rela, [R_RISCV_PCREL_LO12_S] = apply_r_riscv_pcrel_lo12_s_rela, [R_RISCV_CALL] = apply_r_riscv_call_rela, - [R_RISCV_RELAX] = apply_r_riscv_relax_rela, + [R_RISCV_RELAX] = apply_r_riscv_noop_rela, + [R_RISCV_ALIGN] = apply_r_riscv_noop_rela, [R_RISCV_ADD32] = apply_r_riscv_add32_rela, [R_RISCV_SUB32] = apply_r_riscv_sub32_rela, };
Signature · 1 line
2.16.1
PATCH 8/9 RISC-V: Support RISCV_ALIGN relocations in modules.
Quoted 23 lines · shea at shealevy.com
While legal, this noop implementation may be inefficient.
Signed-off-by Shea Levy <[email protected]>--- arch/riscv/kernel/module.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)arch/riscv/kernel/module.c +2
-2@@ -150,8 +150,8 @@ static int apply_r_riscv_call_rela(struct module *me, u32 *location,return 0 ; } -static int apply_r_riscv_relax_rela (struct module *me,u32 *location, -Elf_Addr v) +static int apply_r_riscv_noop_rela (struct module *me,u32 *location, +Elf_Addr v) {return 0 ; }
arch/riscv/kernel/module.c static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, · line 188
Quoted 9 lines
arch/riscv/kernel/module.c +2
-1@@ -181,7 +181,8 @@ static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, [R_RISCV_PCREL_LO12_I] = apply_r_riscv_pcrel_lo12_i_rela, [R_RISCV_PCREL_LO12_S] = apply_r_riscv_pcrel_lo12_s_rela, [R_RISCV_CALL] = apply_r_riscv_call_rela, - [R_RISCV_RELAX] = apply_r_riscv_relax_rela, + [R_RISCV_RELAX] = apply_r_riscv_noop_rela, + [R_RISCV_ALIGN] = apply_r_riscv_noop_rela, [R_RISCV_ADD32] = apply_r_riscv_add32_rela, [R_RISCV_SUB32] = apply_r_riscv_sub32_rela, };
The output of most RISC-V toolchains requires a relaxation pass for alignment for correctness. Even if you're enforcing this somewhere in your module build flow you should at least check for alignment when loading modules so there aren't silent breakages, if R_RISCV_ALIGN is escaping to the kernel's loader then it probably needs to be handled.
For example, I did this (not the best example, but there should be some way to generate this from C I was lazy):
diff --git a/crypto/authenc.c b/crypto/authenc.c
index d3d6d72fe649..4715d7f63f3e 100644
--- a/crypto/authenc.c
+++ b/crypto/authenc.c
@@ -42,6 +42,17 @@ struct authenc_request_ctx {
char tail[];
};
+__asm__ (
+".align 2"
+"\n\t.global align2"
+"\n\talign2:"
+"\n\tret"
+"\n\t.align 3"
+"\n\t.global align3"
+"\n\talign3:"
+"\n\tret"
+);
+
static void authenc_request_complete(struct aead_request *req, int err)
{
if (err != -EINPROGRESS)and ended up with some misaligned stuff:
$ riscv64-unknown-elf-objdump -d -r crypto/authenc.ko | head -n25
crypto/authenc.ko: file format elf64-littleriscv
Disassembly of section .text:
0000000000000000 <align2-0x2>:
0: 0001 nop
0: R_RISCV_ALIGN *ABS*+0x2
0000000000000002 <align2>:
2: 8082 ret
4: 0001 nop
4: R_RISCV_ALIGN *ABS*+0x6
6: 00000013 nop
000000000000000a <align3>:
a: 8082 ret
000000000000000c <crypto_authenc_extractkeys>:
c: 1141 addi sp,sp,-16
e: e422 sd s0,8(sp)
10: 0800 addi s0,sp,16
12: 470d li a4,3
14: 08c77063 bleu a2,a4,94 <.L4>PATCH 8/9 RISC-V: Support RISCV_ALIGN relocations in modules.
Hi Palmer,
Quoted 31 lines · Palmer Dabbelt
Quoted 30 lines · shea at shealevy.com
While legal, this noop implementation may be inefficient.
Signed-off-by Shea Levy <[email protected]>--- arch/riscv/kernel/module.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)arch/riscv/kernel/module.c +4
-3@@ -150,8 +150,8 @@ static int apply_r_riscv_call_rela(struct module *me, u32 *location,return 0 ; } -static int apply_r_riscv_relax_rela (struct module *me,u32 *location, -Elf_Addr v) +static int apply_r_riscv_noop_rela (struct module *me,u32 *location, +Elf_Addr v) {return 0 ; }@@ -181,7 +181,8 @@ static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, [R_RISCV_PCREL_LO12_I] = apply_r_riscv_pcrel_lo12_i_rela, [R_RISCV_PCREL_LO12_S] = apply_r_riscv_pcrel_lo12_s_rela, [R_RISCV_CALL] = apply_r_riscv_call_rela, - [R_RISCV_RELAX] = apply_r_riscv_relax_rela, + [R_RISCV_RELAX] = apply_r_riscv_noop_rela, + [R_RISCV_ALIGN] = apply_r_riscv_noop_rela, [R_RISCV_ADD32] = apply_r_riscv_add32_rela,
Quoted 9 lines
arch/riscv/kernel/module.c static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, · line 188
Quoted 2 lines
arch/riscv/kernel/module.c +0
-0[R_RISCV_SUB32] = apply_r_riscv_sub32_rela, };The output of most RISC-V toolchains requires a relaxation pass for alignment for correctness. Even if you're enforcing this somewhere in your module build flow you should at least check for alignment when loading modules so there aren't silent breakages, if
R_RISCV_ALIGNis escaping to the kernel's loader then it probably needs to be handled.
Well, I guess the question is: Do we want to do relaxations at all at load time? If so, I'll work on that and include RISCV_ALIGN. If not, is there any way to configure gcc and/or ld not to emit any relaxations? In the latter case we can just explicitly not list RISCV_ALIGN here and the module loader will fail with an unsupported relocation.
Quoted 47 lines
For example, I did this (not the best example, but there should be some way to generate this from C I was lazy):
diff --git a/crypto/authenc.c b/crypto/authenc.c index d3d6d72fe649..4715d7f63f3e 100644 --- a/crypto/authenc.c +++ b/crypto/authenc.c @@ -42,6 +42,17 @@ struct authenc_request_ctx { char tail[]; }; +__asm__ ( +".align 2" +"\n\t.global align2" +"\n\talign2:" +"\n\tret" +"\n\t.align 3" +"\n\t.global align3" +"\n\talign3:" +"\n\tret" +); + static void authenc_request_complete(struct aead_request *req, int err) { if (err != -EINPROGRESS)and ended up with some misaligned stuff:
$ riscv64-unknown-elf-objdump -d -r crypto/authenc.ko | head -n25crypto/authenc.ko: file format elf64-littleriscv Disassembly of section .text: 0000000000000000 <align2-0x2>: 0: 0001 nop 0: R_RISCV_ALIGN *ABS*+0x2 0000000000000002 <align2>: 2: 8082 ret 4: 0001 nop 4: R_RISCV_ALIGN *ABS*+0x6 6: 00000013 nop 000000000000000a <align3>:
Quoted 8 lines
a: 8082 ret
000000000000000c <crypto_authenc_extractkeys>:
c: 1141 addi sp,sp,-16
e: e422 sd s0,8(sp)
10: 0800 addi s0,sp,16
12: 470d li a4,3
14: 08c77063 bleu a2,a4,94 <.L4>-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 832 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-riscv/attachments/20180301/8416b6ce/attachment.sig>
PATCH 8/9 RISC-V: Support RISCV_ALIGN relocations in modules.
Quoted 45 lines · shea at shealevy.com
Hi Palmer,
Quoted 31 lines · Palmer Dabbelt
Quoted 30 lines · shea at shealevy.com
While legal, this noop implementation may be inefficient.
Signed-off-by Shea Levy <[email protected]>--- arch/riscv/kernel/module.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)arch/riscv/kernel/module.c +4
-3@@ -150,8 +150,8 @@ static int apply_r_riscv_call_rela(struct module *me, u32 *location,return 0 ; } -static int apply_r_riscv_relax_rela (struct module *me,u32 *location, -Elf_Addr v) +static int apply_r_riscv_noop_rela (struct module *me,u32 *location, +Elf_Addr v) {return 0 ; }@@ -181,7 +181,8 @@ static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, [R_RISCV_PCREL_LO12_I] = apply_r_riscv_pcrel_lo12_i_rela, [R_RISCV_PCREL_LO12_S] = apply_r_riscv_pcrel_lo12_s_rela, [R_RISCV_CALL] = apply_r_riscv_call_rela, - [R_RISCV_RELAX] = apply_r_riscv_relax_rela, + [R_RISCV_RELAX] = apply_r_riscv_noop_rela, + [R_RISCV_ALIGN] = apply_r_riscv_noop_rela, [R_RISCV_ADD32] = apply_r_riscv_add32_rela,Quoted 9 lines
arch/riscv/kernel/module.c static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, · line 188
Quoted 2 lines
arch/riscv/kernel/module.c +0
-0[R_RISCV_SUB32] = apply_r_riscv_sub32_rela, };The output of most RISC-V toolchains requires a relaxation pass for alignment for correctness. Even if you're enforcing this somewhere in your module build flow you should at least check for alignment when loading modules so there aren't silent breakages, if
R_RISCV_ALIGNis escaping to the kernel's loader then it probably needs to be handled.
Quoted 5 lines
Well, I guess the question is: Do we want to do relaxations at all at load time? If so, I'll work on that and include
RISCV_ALIGN. If not, is there any way to configure gcc and/or ld not to emit any relaxations? In the latter case we can just explicitly not listRISCV_ALIGNhere and the module loader will fail with an unsupported relocation.
Well, this is awkward: I remembered doing this (and telling people about "-mno-relax"), but apparently that was all just in my head. Here's a patch to add "-mno-relax" to GCC in the real world :)
https://gcc.gnu.org/ml/gcc-patches/2018-03/msg00062.html
I think the best thing to do is:
* Check if the complier supports "-mno-relax" and add it to the module build command-line.
* Change R_RISCV_ALIGN support in the kernel to check if the address happens to already be aligned (which happens sometimes), and if it's not then provide a helpful error message along the lines of "upgrade to GCC 8".
I guess we eventually want to support relaxation in the kernel, but only if there's a real performance problem that someone can point out.
PATCH 8/9 RISC-V: Support RISCV_ALIGN relocations in modules.
Quoted 59 lines · Palmer Dabbelt
Quoted 45 lines · shea at shealevy.com
Hi Palmer,
Quoted 31 lines · Palmer Dabbelt
Quoted 30 lines · shea at shealevy.com
While legal, this noop implementation may be inefficient.
Signed-off-by Shea Levy <[email protected]>--- arch/riscv/kernel/module.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)arch/riscv/kernel/module.c +4
-3@@ -150,8 +150,8 @@ static int apply_r_riscv_call_rela(struct module *me, u32 *location,return 0 ; } -static int apply_r_riscv_relax_rela (struct module *me,u32 *location, -Elf_Addr v) +static int apply_r_riscv_noop_rela (struct module *me,u32 *location, +Elf_Addr v) {return 0 ; }@@ -181,7 +181,8 @@ static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, [R_RISCV_PCREL_LO12_I] = apply_r_riscv_pcrel_lo12_i_rela, [R_RISCV_PCREL_LO12_S] = apply_r_riscv_pcrel_lo12_s_rela, [R_RISCV_CALL] = apply_r_riscv_call_rela, - [R_RISCV_RELAX] = apply_r_riscv_relax_rela, + [R_RISCV_RELAX] = apply_r_riscv_noop_rela, + [R_RISCV_ALIGN] = apply_r_riscv_noop_rela, [R_RISCV_ADD32] = apply_r_riscv_add32_rela,Quoted 9 lines
arch/riscv/kernel/module.c static int (*reloc_handlers_rela[]) (struct module *me, u32 *location, · line 188
Quoted 2 lines
arch/riscv/kernel/module.c +0
-0[R_RISCV_SUB32] = apply_r_riscv_sub32_rela, };The output of most RISC-V toolchains requires a relaxation pass for alignment for correctness. Even if you're enforcing this somewhere in your module build flow you should at least check for alignment when loading modules so there aren't silent breakages, if
R_RISCV_ALIGNis escaping to the kernel's loader then it probably needs to be handled.Quoted 5 lines
Well, I guess the question is: Do we want to do relaxations at all at load time? If so, I'll work on that and include
RISCV_ALIGN. If not, is there any way to configure gcc and/or ld not to emit any relaxations? In the latter case we can just explicitly not listRISCV_ALIGNhere and the module loader will fail with an unsupported relocation.Well, this is awkward: I remembered doing this (and telling people about "-mno-relax"), but apparently that was all just in my head. Here's a patch to add "-mno-relax" to GCC in the real world :)
https://gcc.gnu.org/ml/gcc-patches/2018-03/msg00062.htmlI think the best thing to do is:
Quoted 8 lines
* Check if the complier supports "-mno-relax" and add it to the module build command-line.
* ChangeR_RISCV_ALIGNsupport in the kernel to check if the address happens to already be aligned (which happens sometimes), and if it's not then provide a helpful error message along the lines of "upgrade to GCC 8".I guess we eventually want to support relaxation in the kernel, but only if there's a real performance problem that someone can point out.
OK, thanks! I'll look into that.
Actually, other than the initial implementation cost, is there any
potential *problem* with implementing relaxations? If I have time I
might want to look into that just for my own education.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 832 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-riscv/attachments/20180301/9c5d9289/attachment.sig>
PATCH 8/9 RISC-V: Support RISCV_ALIGN relocations in modules.
Quoted 14 lines · Shea Levy
Quoted 13 lines · Palmer Dabbelt
Well, this is awkward: I remembered doing this (and telling people about "-mno-relax"), but apparently that was all just in my head. Here's a patch to add "-mno-relax" to GCC in the real world :)
https://gcc.gnu.org/ml/gcc-patches/2018-03/msg00062.htmlI think the best thing to do is:
* Check if the complier supports "-mno-relax" and add it to the module build command-line.
* ChangeR_RISCV_ALIGNsupport in the kernel to check if the address happens to already be aligned (which happens sometimes), and if it's not then provide a helpful error message along the lines of "upgrade to GCC 8".
Quoted 8 lines
Quoted 2 lines
I guess we eventually want to support relaxation in the kernel, but only if there's a real performance problem that someone can point out.
OK, thanks! I'll look into that.
Actually, other than the initial implementation cost, is there any potential *problem* with implementing relaxations? If I have time I might want to look into that just for my own education.
It might be a reasonable exercise but I'd rather not upstream it.
-mno-relax .o files ought to be order of 2x smaller, so assuming you care about initramfs size we want to use that instead of complicating the loader.
(I'd be interested in real numbers for a larger module, e.g. btrfs.)
-s
PATCH 8/9 RISC-V: Support RISCV_ALIGN relocations in modules.
Quoted 24 lines · On Fri, 02 Mar 2018 01:12:26 PST (-0800), sorear2 at gmail.com wrote:
Quoted 14 lines · Shea Levy
Quoted 13 lines · Palmer Dabbelt
Well, this is awkward: I remembered doing this (and telling people about "-mno-relax"), but apparently that was all just in my head. Here's a patch to add "-mno-relax" to GCC in the real world :)
https://gcc.gnu.org/ml/gcc-patches/2018-03/msg00062.htmlI think the best thing to do is:
* Check if the complier supports "-mno-relax" and add it to the module build command-line.
* ChangeR_RISCV_ALIGNsupport in the kernel to check if the address happens to already be aligned (which happens sometimes), and if it's not then provide a helpful error message along the lines of "upgrade to GCC 8".Quoted 8 lines
Quoted 2 lines
I guess we eventually want to support relaxation in the kernel, but only if there's a real performance problem that someone can point out.
OK, thanks! I'll look into that.
Actually, other than the initial implementation cost, is there any potential *problem* with implementing relaxations? If I have time I might want to look into that just for my own education.
Quoted 7 lines
It might be a reasonable exercise but I'd rather not upstream it.
-mno-relax .o files ought to be order of 2x smaller, so assuming you care about initramfs size we want to use that instead of complicating the loader.
(I'd be interested in real numbers for a larger module, e.g. btrfs.)
It's actually kind of an odd one: using "-mrelax" instead of "-mno-relax" would result in:
* Larger initramfs sizes (the kernel modules would still have all the pre-relaxed instructions, plus the extra relocation info required to relax and the most pessimistic alignment possible).
* Larger, slower, more complicated module loading code (to handle all the relaxations).
* Smaller, faster modules once they've been loaded into the kernel.
I guess we'd need to see some numbers as to how much can actually be relaxed, as I suspect there won't be a whole lot of opportunities for relaxation in module code. I think this is probably not that high priority, but it would be fun :).
PATCH 9/9 RISC-V: Build a module-enabled kernel by default.
In reply to Shea Levy--- arch/riscv/configs/defconfig | 1 + 1 file changed, 1 insertion(+)
arch/riscv/configs/defconfig +1 -0
@@ -73,3 +73,4 @@ CONFIG_NFS_V4_2=y CONFIG_ROOT_NFS=y # CONFIG_RCU_TRACE is not set CONFIG_CRYPTO_USER_API_HASH=y +CONFIG_MODULES=y
Signature · 1 line
2.16.1
cover RISC-V: Make modules loadable.
In reply to Shea LevyQuoted 2 lines · shea at shealevy.com
This was enough to load enough modules to make
virtio_scsidevices appear on the qemu virt machine on a highly modular distro kernel.
Thanks for this, but there's a bit of an issue with R_RISCV_ALIGN. I'd be OK merging the change set if you emit errors on R_RISCV_ALIGN instances that actually require alignment, as IIRC it's possible to coerce GCC to emit code that doesn't require relaxation of alignment sequences. This should probably go in the module CFLAGS as well.
There are also some minor comments, see the per-patch messages.
cover RISC-V: Make modules loadable.
Quoted 6 lines · Palmer Dabbelt
Quoted 3 lines · shea at shealevy.com
This was enough to load enough modules to make
virtio_scsidevices appear on the qemu virt machine on a highly modular distro kernel.
Quoted 7 lines
Thanks for this, but there's a bit of an issue with
R_RISCV_ALIGN. I'd be OK merging the change set if you emit errors onR_RISCV_ALIGNinstances that actually require alignment, as IIRC it's possible to coerce GCC to emit code that doesn't require relaxation of alignment sequences. This should probably go in the module CFLAGS as well.There are also some minor comments, see the per-patch messages.
The question of whether processing R_RISCV_ALIGN is mandatory came up during lld patch review,
https://lists.llvm.org/pipermail/llvm-dev/2017-July/115407.html . My
position is that if gas is generating instances that require alignment, that is a bug, although I'm quite fine with hardening this.
-s
cover RISC-V: Make modules loadable.
Quoted 13 lines · On Mon, 26 Feb 2018 21:49:40 PST (-0800), sorear2 at gmail.com wrote:
Quoted 12 lines · Palmer Dabbelt
Quoted 3 lines · shea at shealevy.com
This was enough to load enough modules to make
virtio_scsidevices appear on the qemu virt machine on a highly modular distro kernel.Thanks for this, but there's a bit of an issue with
R_RISCV_ALIGN. I'd be OK merging the change set if you emit errors onR_RISCV_ALIGNinstances that actually require alignment, as IIRC it's possible to coerce GCC to emit code that doesn't require relaxation of alignment sequences. This should probably go in the module CFLAGS as well.
Quoted 7 lines
Quoted 1 line
There are also some minor comments, see the per-patch messages.
The question of whether processing
R_RISCV_ALIGNis mandatory came up during lld patch review,
https://lists.llvm.org/pipermail/llvm-dev/2017-July/115407.html . My
position is that if gas is generating instances that require alignment, that is a bug, although I'm quite fine with hardening this.
We'd need to add a new relocation in order to make that happen, as R_RISCV_ALIGN assumes. I wouldn't be opposed to it, but I thought the conclusion we came to was that there were two reasons to avoid processing R_RISCV_ALIGN:
* The linker doesn't support relaxation. We consider this a bug in the linker, and I don't want to add a relocation just for incomplete linkers.
* Relaxation is too time consuming. In this case you want to disable all relaxation, and GAS will already emit code without R_RISCV_ALIGN when run with relaxation disabled.
IIRC there was an email thread or a github issue about this.
cover RISC-V: Make modules loadable.
Quoted 3 lines · Palmer Dabbelt
* Relaxation is too time consuming. In this case you want to disable all relaxation, and GAS will already emit code without
R_RISCV_ALIGNwhen run with relaxation disabled.
Tangentially, I described a linear time algorithm for relaxation a while ago, but binutils is still using a quadratic time one. Would patches be accepted?
-s
cover RISC-V: Make modules loadable.
Quoted 8 lines · On Tue, 27 Feb 2018 10:26:48 PST (-0800), sorear2 at gmail.com wrote:
Quoted 3 lines · Palmer Dabbelt
* Relaxation is too time consuming. In this case you want to disable all relaxation, and GAS will already emit code without
R_RISCV_ALIGNwhen run with relaxation disabled.Tangentially, I described a linear time algorithm for relaxation a while ago, but binutils is still using a quadratic time one. Would patches be accepted?
Yes. There's a tiny bit of support in there for it via R_RISCV_DELETE, I think we just need to convert the rest of the relaxations over to that and then flip around the pass that handles R_RISCV_DELETE and we should be good.
Of course, R_RISCV_DELETE isn't a public relocation so if you have a better way to handle it then I'm all ears :).