Re: [PATCH 2/4] x86/kprobes: Fix frame pointer annotations

[Andy Lutomirski]

> On May 9, 2019, at 1:14 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
>> On Thu, May 09, 2019 at 10:20:30AM +0900, Masami Hiramatsu wrote:
>> Hi Josh,
>> 
>> On Wed, 8 May 2019 13:48:48 -0500
>> Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
>> 
>>>> On Wed, May 08, 2019 at 05:39:07PM +0200, Peter Zijlstra wrote:
>>>>> On Wed, May 08, 2019 at 07:42:48AM -0500, Josh Poimboeuf wrote:
>>>>> On Wed, May 08, 2019 at 02:04:16PM +0200, Peter Zijlstra wrote:
>>>> 
>>>>>> Do the x86_64 variants also want some ORC annotation?
>>>>> 
>>>>> Maybe so.  Though it looks like regs->ip isn't saved.  The saved
>>>>> registers might need to be tweaked.  I'll need to look into it.
>>>> 
>>>> What all these sites do (and maybe we should look at unifying them
>>>> somehow) is turn a CALL frame (aka RET-IP) into an exception frame (aka
>>>> pt_regs).
>>>> 
>>>> So regs->ip will be the return address (which is fixed up to be the CALL
>>>> address in the handler).
>>> 
>>> But from what I can tell, trampoline_handler() hard-codes regs->ip to
>>> point to kretprobe_trampoline(), and the original return address is
>>> placed in regs->sp.
>>> 
>>> Masami, is there a reason why regs->ip doesn't have the original return
>>> address and regs->sp doesn't have the original SP?  I think that would
>>> help the unwinder understand things.
>> 
>> Yes, for regs->ip, there is a histrical reason. Since previously, we had
>> an int3 at trampoline, so the user (kretprobe) handler expects that
>> regs->ip is trampoline address and ri->ret_addr is original return address.
>> It is better to check the other archs, but I think it is possible to
>> change the regs->ip to original return address, since no one cares such
>> "fixed address". :)
>> 
>> For the regs->sp, there are 2 reasons.
>> 
>> For x86-64, it's just for over-optimizing (reduce stack usage).
>> I think we can make a gap for putting return address, something like
>> 
>>    "kretprobe_trampoline:\n"
>> #ifdef CONFIG_X86_64
>>    "    pushq %rsp\n"    /* Make a gap for return address */
>>    "    pushq 0(%rsp)\n"    /* Copy original stack pointer */
>>    "    pushfq\n"
>>    SAVE_REGS_STRING
>>    "    movq %rsp, %rdi\n"
>>    "    call trampoline_handler\n"
>>    /* Push the true return address to the bottom */
>>    "    movq %rax, 20*8(%rsp)\n"
>>    RESTORE_REGS_STRING
>>    "    popfq\n"
>>    "    addq $8, %rsp\n"    /* Skip original stack pointer */
>> 
>> For i386 (x86-32), there is no other way to keep &regs->sp as
>> the original stack pointer. It has to be changed with this series,
>> maybe as same as x86-64.
> 
> Right; I already fixed that in my patch changing i386's pt_regs.
> 
> But what I'd love to do is something like the belwo patch, and make all
> the trampolines (very much including ftrace) use that. Such that we then
> only have 1 copy of this magic (well, 2 because x86_64 also needs an
> implementation of this of course).
> 
> Changing ftrace over to this would be a little more work but it can
> easily chain things a little to get its original context back:
> 
> ENTRY(ftrace_regs_caller)
> GLOBAL(ftrace_regs_func)
>    push ftrace_stub
>    push ftrace_regs_handler
>    jmp call_to_exception_trampoline
> END(ftrace_regs_caller)
> 
> typedef void (*ftrace_func_t)(unsigned long, unsigned long, struct ftrace_op *, struct pt_regs *);
> 
> struct ftrace_regs_stack {
>    ftrace_func_t func;
>    unsigned long parent_ip;
> };
> 
> void ftrace_regs_handler(struct pr_regs *regs)
> {
>    struct ftrace_regs_stack *st = (void *)regs->sp;
>    ftrace_func_t func = st->func;
> 
>    regs->sp += sizeof(long); /* pop func */
> 
>    func(regs->ip, st->parent_ip, function_trace_op, regs);
> }
> 
> Hmm? I didn't look into the function_graph thing, but I imagine it can
> be added without too much pain.
> 
> ---
> --- a/arch/x86/entry/entry_32.S
> +++ b/arch/x86/entry/entry_32.S
> @@ -1576,3 +1576,100 @@ ENTRY(rewind_stack_do_exit)
>    call    do_exit
> 1:    jmp 1b
> END(rewind_stack_do_exit)
> +
> +/*
> + * Transforms a CALL frame into an exception frame; IOW it pretends the CALL we
> + * just did was in fact scribbled with an INT3.
> + *
> + * Use this trampoline like:
> + *
> + *   PUSH $func
> + *   JMP call_to_exception_trampoline
> + *
> + * $func will see regs->ip point at the CALL instruction and must therefore
> + * modify regs->ip in order to make progress (just like a normal INT3 scribbled
> + * CALL).
> + *
> + * NOTE: we do not restore any of the segment registers.
> + */
> +ENTRY(call_to_exception_trampoline)
> +    /*
> +     * On entry the stack looks like:
> +     *
> +     *   2*4(%esp) <previous context>
> +     *   1*4(%esp) RET-IP
> +     *   0*4(%esp) func
> +     *
> +     * transform this into:
> +     *
> +     *  19*4(%esp) <previous context>
> +     *  18*4(%esp) gap / RET-IP
> +     *  17*4(%esp) gap / func
> +     *  16*4(%esp) ss
> +     *  15*4(%esp) sp / <previous context>
> +     *  14*4(%esp) flags
> +     *  13*4(%esp) cs
> +     *  12*4(%esp) ip / RET-IP
> +     *  11*4(%esp) orig_eax
> +     *  10*4(%esp) gs
> +     *   9*4(%esp) fs
> +     *   8*4(%esp) es
> +     *   7*4(%esp) ds
> +     *   6*4(%esp) eax
> +     *   5*4(%esp) ebp
> +     *   4*4(%esp) edi
> +     *   3*4(%esp) esi
> +     *   2*4(%esp) edx
> +     *   1*4(%esp) ecx
> +     *   0*4(%esp) ebx
> +     */
> +    pushl    %ss
> +    pushl    %esp        # points at ss
> +    addl    $3*4, (%esp)    #   point it at <previous context>
> +    pushfl
> +    pushl    %cs
> +    pushl    5*4(%esp)    # RET-IP
> +    subl    5, (%esp)    #   point at CALL instruction
> +    pushl    $-1
> +    pushl    %gs
> +    pushl    %fs
> +    pushl    %es
> +    pushl    %ds
> +    pushl    %eax
> +    pushl    %ebp
> +    pushl    %edi
> +    pushl    %esi
> +    pushl    %edx
> +    pushl    %ecx
> +    pushl    %ebx
> +
> +    ENCODE_FRAME_POINTER
> +
> +    movl    %esp, %eax    # 1st argument: pt_regs
> +
> +    movl    17*4(%esp), %ebx    # func
> +    CALL_NOSPEC %ebx
> +
> +    movl    PT_OLDESP(%esp), %eax
> +
> +    movl    PT_EIP(%esp), %ecx
> +    movl    %ecx, -1*4(%eax)
> +
> +    movl    PT_EFLAGS(%esp), %ecx
> +    movl    %ecx, -2*4(%eax)
> +
> +    movl    PT_EAX(%esp), %ecx
> +    movl    %ecx, -3*4(%eax)
> +
> +    popl    %ebx
> +    popl    %ecx
> +    popl    %edx
> +    popl    %esi
> +    popl    %edi
> +    popl    %ebp
> +
> +    lea    -3*4(%eax), %esp
> +    popl    %eax
> +    popfl
> +    ret
> +END(call_to_exception_trampoline)
> --- a/arch/x86/kernel/kprobes/core.c
> +++ b/arch/x86/kernel/kprobes/core.c
> @@ -731,29 +731,8 @@ asm(
>    ".global kretprobe_trampoline\n"
>    ".type kretprobe_trampoline, @function\n"
>    "kretprobe_trampoline:\n"
> -    /* We don't bother saving the ss register */
> -#ifdef CONFIG_X86_64
> -    "    pushq %rsp\n"
> -    "    pushfq\n"
> -    SAVE_REGS_STRING
> -    "    movq %rsp, %rdi\n"
> -    "    call trampoline_handler\n"
> -    /* Replace saved sp with true return address. */
> -    "    movq %rax, 19*8(%rsp)\n"
> -    RESTORE_REGS_STRING
> -    "    popfq\n"
> -#else
> -    "    pushl %esp\n"
> -    "    pushfl\n"
> -    SAVE_REGS_STRING
> -    "    movl %esp, %eax\n"
> -    "    call trampoline_handler\n"
> -    /* Replace saved sp with true return address. */
> -    "    movl %eax, 15*4(%esp)\n"
> -    RESTORE_REGS_STRING
> -    "    popfl\n"
> -#endif
> -    "    ret\n"
> +    "push trampoline_handler\n"
> +    "jmp call_to_exception_trampoline\n"
>    ".size kretprobe_trampoline, .-kretprobe_trampoline\n"
> );


Potentially minor nit: youâre doing popfl, but youâre not doing TRACE_IRQ_whatever.  This makes me think that you should either add the tracing (ugh!) or you should maybe just skip the popfl.