On Sat, Mar 04, 2017 at 12:51:12AM +0000, Abel Vesa wrote:
> diff --git a/arch/arm/kernel/entry-ftrace.S b/arch/arm/kernel/entry-ftrace.S
> index c73c403..93f9abb 100644
> --- a/arch/arm/kernel/entry-ftrace.S
> +++ b/arch/arm/kernel/entry-ftrace.S
> @@ -92,12 +92,78 @@
> 2: mcount_exit
> .endm
>
> +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
> +
> +.macro __ftrace_regs_caller
> +
> + sub sp, sp, #8 @ space for CPSR and OLD_R0 (not used)
> +
> + add ip, sp, #12 @ move in IP the value of SP as it was
> + @ before the push {lr} of the mcount mechanism
> + stmdb sp!, {ip,lr,pc}
This doesn't build with CONFIG_THUMB2_KERNEL:
entry-ftrace.S:285: Error: PC not allowed in register list -- `stmdb
sp!,{ip,lr,pc}'
Saving PC in STMDB is prohibited in Thumb and deprecated in ARM.
> + stmdb sp!, {r0-r11,lr}
> +
> + ldr r0, [sp, #S_LR] @ replace PC with LR
> + str r0, [sp, #S_PC] @ into pt_regs
> +
> + ldr r1, [sp, #PT_REGS_SIZE] @ replace new LR with
> + str r1, [sp, #S_LR] @ previous LR into pt_regs
> +
> + @ stack content at this point:
> + @ 0 4 48 52 56 60 64 68 72
> + @ R0 | R1 | ... | LR | SP + 4 | previous LR | LR | PSR | OLD_R0 |
> previous LR |
So this code only adjust the SP by 72 bytes. Since the stack is not
aligned to 8 bytes at entry to mcount, this code calls into C functions
with a misaligned stack. This is a violation of the procedure call
standard.
Also, I believe this should have unwind annotations to indicate where it
saved the registers. See mcount_enter for an example.
> +
> + mov r3, sp @ struct pt_regs*
> + ldr r2, =function_trace_op
> + ldr r2, [r2] @ pointer to the current
> + @ function tracing op
> + ldr r1, [sp, #PT_REGS_SIZE] @ lr of instrumented func
> + mcount_adjust_addr r0, lr @ instrumented function
> +
> + .globl ftrace_regs_call
> +ftrace_regs_call:
> + bl ftrace_stub
> +
> +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> + .globl ftrace_graph_regs_call
> +ftrace_graph_regs_call:
> + mov r0, r0
> +#endif
> + @ pop saved regs
> + ldmia sp, {r0-r15}
This doesn't build either:
entry-ftrace.S:285: Error: LR and PC should not both be in register list --
`ldmia sp,{r0-r15}'
Restoring LR and PC together is prohibited in Thumb and deprecated in
ARM. It's the same case with SP.
> +.endm
> +
> +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> +.macro __ftrace_graph_regs_caller
> +
> + sub r0, fp, #4 @ lr of instrumented routine (parent)
> +
> + @ called from __ftrace_regs_caller
> + ldr r1, [sp, #S_PC] @ instrumented routine (func)
> + mcount_adjust_addr r1, r1
> +
> + mov r2, fp @ frame pointer
> + bl prepare_ftrace_return
> +
> + @ pop registers saved in ftrace_regs_caller
> + ldmia sp, {r0-r15}
This doesn't get built on CONFIG_THUMB2_KERNEL since there's no suppport
for CONFIG_FUNCTION_GRAPH_TRACER there, but this is also using
deprecated operands.
> +.endm
> +#endif
> +#endif
...
> diff --git a/arch/arm/kernel/ftrace.c b/arch/arm/kernel/ftrace.c
> index 3f17594..f165265 100644
> --- a/arch/arm/kernel/ftrace.c
> +++ b/arch/arm/kernel/ftrace.c
> @@ -29,11 +29,6 @@
> #endif
>
> #ifdef CONFIG_DYNAMIC_FTRACE
> -#ifdef CONFIG_OLD_MCOUNT
> -#define OLD_MCOUNT_ADDR ((unsigned long) mcount)
> -#define OLD_FTRACE_ADDR ((unsigned long) ftrace_caller_old)
> -
> -#define OLD_NOP 0xe1a00000 /* mov r0, r0 */
>
> static int __ftrace_modify_code(void *data)
> {
> @@ -51,6 +46,12 @@ void arch_ftrace_update_code(int command)
> stop_machine(__ftrace_modify_code, &command, NULL);
> }
>
> +#ifdef CONFIG_OLD_MCOUNT
> +#define OLD_MCOUNT_ADDR ((unsigned long) mcount)
> +#define OLD_FTRACE_ADDR ((unsigned long) ftrace_caller_old)
> +
> +#define OLD_NOP 0xe1a00000 /* mov r0, r0 */
> +
> static unsigned long ftrace_nop_replace(struct dyn_ftrace *rec)
This chunk above fixes the problem which Nicolai pointed out earlier. Since
it's a bug fix for the current dynamic ftrace code, it should really be
submitted as a separate patch. Feel free to use some variant of this
commit message which I wrote when I ran into this yesterday:
ARM: ftrace: fix dynamic ftrace with DEBUG_RODATA and !FRAME_POINTER
The support for dynamic ftrace with CONFIG_DEBUG_RODATA involves
overriding the weak arch_ftrace_update_code() with a variant which makes
the kernel text writable around the patching.
This override was however added under the CONFIG_OLD_MCOUNT ifdef, and
CONFIG_OLD_MCOUNT is only enabled if frame pointers are enabled.
This leads to non-functional dynamic ftrace (ftrace triggers a
WARN_ON()) when CONFIG_DEBUG_RODATA is enabled and CONFIG_FRAME_POINTER
is not.
Move the override out of that ifdef and into the CONFIG_DYNAMIC_FTRACE
ifdef where it belongs.
Fixes: 80d6b0c2eed2a ("ARM: mm: allow text and rodata sections to be
read-only")
> {
> return rec->arch.old_mcount ? OLD_NOP : NOP;
> @@ -139,6 +140,15 @@ int ftrace_update_ftrace_func(ftrace_func_t func)
>
> ret = ftrace_modify_code(pc, 0, new, false);
>
> +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
This code (and the rest of your additions to this file below), are already
under CONFIG_DYNAMIC_FTRACE and afaics CONFIG_DYNAMIC_FTRACE_WITH_REGS is
always enabled when the former is enabled, so could you please drop
these ifdefs?
The code in entry-ftrace.S under CONFIG_DYNAMIC_FTRACE_WITH_REGS would
also benefit from just using CONFIG_DYNAMIC_FTRACE everywhere instead.
