[PATCH v4.2] powerpc/modules: Don't try to restore r2 after a sibling call

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Thu, Nov 16, 2017 at 06:39:03PM +0530, Naveen N. Rao wrote:
> Josh Poimboeuf wrote:
> > On Wed, Nov 15, 2017 at 02:58:33PM +0530, Naveen N. Rao wrote:
> > > > +int instr_is_link_branch(unsigned int instr)
> > > > +{
> > > > +	return (instr_is_branch_iform(instr) || instr_is_branch_bform(instr)) &&
> > > > +	       (instr & BRANCH_SET_LINK);
> > > > +}
> > > > +
> > > 
> > > Nitpicking here, but since we're not considering the other branch forms,
> > > perhaps this can be renamed to instr_is_link_relative_branch() (or maybe
> > > instr_is_relative_branch_link()), just so we're clear :)
> > 
> > My understanding is that the absolute/relative bit isn't a "form", but
> > rather a bit that can be set for either the b-form (conditional) or the
> > i-form (unconditional).  And the above function isn't checking the
> > absolute bit, so it isn't necessarily a relative branch.  Or did I miss
> > something?
> 
> Ah, good point. I was coming from the fact that we are only considering the
> i-form and b-form branches and not the lr/ctr/tar based branches, which are
> always absolute branches, but can also set the link register.

Hm, RISC is more complicated than I realized ;-)

> Thinking about this more, aren't we only interested in relative branches
> here (for relocations), so can we actually filter out the absolute branches?
> Something like this?
> 
> int instr_is_relative_branch_link(unsigned int instr)
> {
> 	return ((instr_is_branch_iform(instr) || instr_is_branch_bform(instr)) &&
> 	       !(instr & BRANCH_ABSOLUTE) && (instr & BRANCH_SET_LINK));

Yeah, makes sense to me.  Here's another try (also untested).  If this
looks ok, Kamalesh would you mind testing again?

----8<----

From: Josh Poimboeuf <jpoimboe@redhat.com>
Subject: [PATCH v4.2] powerpc/modules: Don't try to restore r2 after a sibling call

When attempting to load a livepatch module, I got the following error:

  module_64: patch_module: Expect noop after relocate, got 3c820000

The error was triggered by the following code in
unregister_netdevice_queue():

  14c:   00 00 00 48     b       14c <unregister_netdevice_queue+0x14c>
                         14c: R_PPC64_REL24      net_set_todo
  150:   00 00 82 3c     addis   r4,r2,0

GCC didn't insert a nop after the branch to net_set_todo() because it's
a sibling call, so it never returns.  The nop isn't needed after the
branch in that case.

Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com>
---
 arch/powerpc/include/asm/code-patching.h |  1 +
 arch/powerpc/kernel/module_64.c          | 12 +++++++++++-
 arch/powerpc/lib/code-patching.c         |  5 +++++
 3 files changed, 17 insertions(+), 1 deletion(-)

diff --git a/arch/powerpc/include/asm/code-patching.h b/arch/powerpc/include/asm/code-patching.h
index abef812de7f8..2c895e8d07f7 100644
--- a/arch/powerpc/include/asm/code-patching.h
+++ b/arch/powerpc/include/asm/code-patching.h
@@ -33,6 +33,7 @@ int patch_branch(unsigned int *addr, unsigned long target, int flags);
 int patch_instruction(unsigned int *addr, unsigned int instr);
 
 int instr_is_relative_branch(unsigned int instr);
+int instr_is_relative_link_branch(unsigned int instr);
 int instr_is_branch_to_addr(const unsigned int *instr, unsigned long addr);
 unsigned long branch_target(const unsigned int *instr);
 unsigned int translate_branch(const unsigned int *dest,
diff --git a/arch/powerpc/kernel/module_64.c b/arch/powerpc/kernel/module_64.c
index 759104b99f9f..180c16f04063 100644
--- a/arch/powerpc/kernel/module_64.c
+++ b/arch/powerpc/kernel/module_64.c
@@ -487,7 +487,17 @@ static bool is_early_mcount_callsite(u32 *instruction)
    restore r2. */
 static int restore_r2(u32 *instruction, struct module *me)
 {
-	if (is_early_mcount_callsite(instruction - 1))
+	u32 *prev_insn = instruction - 1;
+
+	if (is_early_mcount_callsite(prev_insn))
+		return 1;
+
+	/*
+	 * Make sure the branch isn't a sibling call.  Sibling calls aren't
+	 * "link" branches and they don't return, so they don't need the r2
+	 * restore afterwards.
+	 */
+	if (!instr_is_relative_link_branch(*prev_insn))
 		return 1;
 
 	if (*instruction != PPC_INST_NOP) {
diff --git a/arch/powerpc/lib/code-patching.c b/arch/powerpc/lib/code-patching.c
index c9de03e0c1f1..d81aab7441f7 100644
--- a/arch/powerpc/lib/code-patching.c
+++ b/arch/powerpc/lib/code-patching.c
@@ -304,6 +304,11 @@ int instr_is_relative_branch(unsigned int instr)
 	return instr_is_branch_iform(instr) || instr_is_branch_bform(instr);
 }
 
+int instr_is_relative_link_branch(unsigned int instr)
+{
+	return instr_is_relative_branch(instr) && (instr & BRANCH_SET_LINK);
+}
+
 static unsigned long branch_iform_target(const unsigned int *instr)
 {
 	signed long imm;
-- 
2.13.6