[U-Boot] efi_loader: arch/arm/include/asm/setjmp.h
Hello Alex,
our current logic for removing the handler of the loaded image is wrong:
If a driver image calls Exit() with EFI_SUCCESS the image has to remain.
I tried to rework the logic and needed a second setjmp in cmd/bootefi.c to handle 'bootefi selftest'.
On arm this resulted in:
Building current source for 1 boards (1 thread, 4 jobs per thread) arm: + vexpress_ca15_tc2 +{standard input}: Assembler messages: +{standard input}:638: Error: symbol `jmp_target' is already defined +{standard input}:633: Error: symbol .text.do_bootefi_exec is in a different section +make[2]: *** [cmd/bootefi.o] Error 1 +make[1]: *** [cmd] Error 2 +make: *** [sub-make] Error 2
All registers that have to be restored should be stored in jmp_buf and not at any static address (jmp_target).
Otherwise nesting of setjmp will not work: U-Boot calling an EFI application, which loads another EFI application which exits to the first application which in turn exits to the U-Boot.
Are you able to rework this? Unfortunately I lack the assembler experience. The uClibc implementation might be a good starting point.
Best regards
Heinrich
On 07.10.17 22:23, Heinrich Schuchardt wrote:
Hello Alex,
our current logic for removing the handler of the loaded image is wrong:
If a driver image calls Exit() with EFI_SUCCESS the image has to remain.
I tried to rework the logic and needed a second setjmp in cmd/bootefi.c to handle 'bootefi selftest'.
On arm this resulted in:
Building current source for 1 boards (1 thread, 4 jobs per thread) arm: + vexpress_ca15_tc2 +{standard input}: Assembler messages: +{standard input}:638: Error: symbol `jmp_target' is already defined +{standard input}:633: Error: symbol .text.do_bootefi_exec is in a different section +make[2]: *** [cmd/bootefi.o] Error 1 +make[1]: *** [cmd] Error 2 +make: *** [sub-make] Error 2
All registers that have to be restored should be stored in jmp_buf and not at any static address (jmp_target).
jmp_target is not an address to store things at, it's an address to jump to :). We just happen to define the same label twice because of inlining.
Does the patch below fix it for you?
Either way, Philipp rewrote all of the setjmp/longjmp code anyway in his rockchip patch set.
Philipp, what path do you expect that to take? Should we fix the in-tree setjmp or just apply yours?
Alex
diff --git a/arch/arm/include/asm/setjmp.h b/arch/arm/include/asm/setjmp.h index c3399a7e15..2c58e2819e 100644 --- a/arch/arm/include/asm/setjmp.h +++ b/arch/arm/include/asm/setjmp.h @@ -22,13 +22,13 @@ static inline int setjmp(jmp_buf jmp)
#ifdef CONFIG_ARM64 asm volatile( - "adr x1, jmp_target\n" + "adr x1, 1f\n" "str x1, %0\n" "stp x26, x27, %1\n" "stp x28, x29, %2\n" "mov x1, sp\n" "str x1, %3\n" - "jmp_target: " + "1: " : "=m" (jmp->target), "=m" (jmp->regs[0]), "=m" (jmp->regs[2]), "=m" (jmp->regs[4]) : @@ -41,16 +41,16 @@ static inline int setjmp(jmp_buf jmp) asm volatile( #if CONFIG_IS_ENABLED(SYS_THUMB_BUILD) ".align 2\n" - "adr r0, jmp_target\n" + "adr r0, 1f\n" "add r0, r0, $1\n" #else - "adr r0, jmp_target\n" + "adr r0, 1f\n" #endif "mov r1, %0\n" "mov r2, sp\n" "stm r1!, {r0, r2, r4, r5, r6, r7}\n" ".align 2\n" - "jmp_target: \n" + "1: \n" : : "l" (&jmp->target) : "r0", "r1", "r2", "r3", /* "r4", "r5", "r6", "r7", */
participants (2)
-
Alexander Graf -
Heinrich Schuchardt