On May 9, 2019, at 1:14 AM, Peter Zijlstra peterz@infradead.org 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@redhat.com 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 ®s->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: you’re doing popfl, but you’re 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.