Re: [PATCH 00/10] ftrace: Add register_ftrace_direct()

[Josh Poimboeuf]

On Fri, Nov 08, 2019 at 04:28:34PM -0500, Steven Rostedt wrote:
> 
> Alexei mentioned that he would like a way to access the ftrace fentry
> code to jump directly to a custom eBPF trampoline instead of using
> ftrace regs caller, as he said it would be faster.
> 
> I proposed a new register_ftrace_direct() function that would allow
> this to happen and still work with the ftrace infrastructure. I posted
> a proof of concept patch here:
> 
>  https://lore.kernel.org/r/20191023122307.756b4978@xxxxxxxxxxxxxxxxxx
> 
> This patch series is a more complete version, and the start of the
> actual implementation. I haven't run it through my full test suite but
> it passes my smoke tests and some other custom tests I built.
> 
> I also realized that I need to make the sample modules depend on X86_64
> as it has inlined assembly in it that requires that dependency.
> 
> This is based on 5.4-rc6 plus the permanent patches that prevent
> a ftrace_ops from being disabled by /proc/sys/kernel/ftrace_enabled
> 
> Below is the tree that contains this code.
> 
>   git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace.git
> ftrace/direct

Here's the fix for the objtool warning:

From: Josh Poimboeuf <jpoimboe@xxxxxxxxxx>
Subject: [PATCH] ftrace/x86: Tell objtool to ignore nondeterministic ftrace stack layout

Objtool complains about the new ftrace direct trampoline code:

  arch/x86/kernel/ftrace_64.o: warning: objtool: ftrace_regs_caller()+0x190: stack state mismatch: cfa1=7+16 cfa2=7+24

Typically, code has a deterministic stack layout, such that at a given
instruction address, the stack frame size is always the same.

That's not the case for the new ftrace_regs_caller() code after it
adjusts the stack for the direct case.  Just plead ignorance and assume
it's always the non-direct path.  Note this creates a tiny window for
ORC to get confused.

Reported-by: Steven Rostedt <rostedt@xxxxxxxxxxx>
Signed-off-by: Josh Poimboeuf <jpoimboe@xxxxxxxxxx>
---
 arch/x86/include/asm/unwind_hints.h |  8 ++++++++
 arch/x86/kernel/ftrace_64.S         | 12 +++++++++++-
 2 files changed, 19 insertions(+), 1 deletion(-)

diff --git a/arch/x86/include/asm/unwind_hints.h b/arch/x86/include/asm/unwind_hints.h
index 0bcdb1279361..f5e2eb12cb71 100644
--- a/arch/x86/include/asm/unwind_hints.h
+++ b/arch/x86/include/asm/unwind_hints.h
@@ -86,6 +86,14 @@
 	UNWIND_HINT sp_offset=\sp_offset
 .endm
 
+.macro UNWIND_HINT_SAVE
+	UNWIND_HINT type=UNWIND_HINT_TYPE_SAVE
+.endm
+
+.macro UNWIND_HINT_RESTORE
+	UNWIND_HINT type=UNWIND_HINT_TYPE_RESTORE
+.endm
+
 #else /* !__ASSEMBLY__ */
 
 #define UNWIND_HINT(sp_reg, sp_offset, type, end)		\
diff --git a/arch/x86/kernel/ftrace_64.S b/arch/x86/kernel/ftrace_64.S
index 5d946ab40b52..1c79624a36b2 100644
--- a/arch/x86/kernel/ftrace_64.S
+++ b/arch/x86/kernel/ftrace_64.S
@@ -175,6 +175,8 @@ ENTRY(ftrace_regs_caller)
 	/* Save the current flags before any operations that can change them */
 	pushfq
 
+	UNWIND_HINT_SAVE
+
 	/* added 8 bytes to save flags */
 	save_mcount_regs 8
 	/* save_mcount_regs fills in first two parameters */
@@ -249,8 +251,16 @@ GLOBAL(ftrace_regs_call)
 1:	restore_mcount_regs
 
 
+2:
+	/*
+	 * The stack layout is nondetermistic here, depending on which path was
+	 * taken.  This confuses objtool and ORC, rightfully so.  For now,
+	 * pretend the stack always looks like the non-direct case.
+	 */
+	UNWIND_HINT_RESTORE
+
 	/* Restore flags */
-2:	popfq
+	popfq
 
 	/*
 	 * As this jmp to ftrace_epilogue can be a short jump
-- 
2.20.1