Re: [PATCH] recordmcount: only record relocation of type R_AARCH64_CALL26 on arm64.

[Gregory Herrero <gregory.herrero@oracle.com>]

Hi Mark,

On Thu, Jul 23, 2020 at 12:52:16PM +0100, Mark Rutland wrote:
> Hi Gregory,
> 
> As a general thing, for patches affecting arm64 could you please Cc the
> linx-arm-kernel mailing list (linux-arm-kernel@lists.infradead.org).
> Some folk working on arm/arm64 aren't subscribed to LKML, and it means
> patches like this may get missed.
> 
Got it, I will do that next time.

> On Fri, Jul 17, 2020 at 04:33:38PM +0200, gregory.herrero@oracle.com wrote:
> > From: Gregory Herrero <gregory.herrero@oracle.com>
> > 
> > Currently, if a section has a relocation to '_mcount' symbol, a new
> > __mcount_loc entry will be added whatever the relocation type is.
> > This is problematic when a relocation to '_mcount' is in the middle of a
> > section and is not a call for ftrace use.
> > 
> > Such relocation could be generated with below code for example:
> >     bool is_mcount(unsigned long addr)
> >     {
> >         return (target == (unsigned long) &_mcount);
> >     }
> > 
> > With this snippet of code, ftrace will try to patch the mcount location
> > generated by this code on module load and fail with:
> > 
> >     Call trace:
> >      ftrace_bug+0xa0/0x28c
> >      ftrace_process_locs+0x2f4/0x430
> >      ftrace_module_init+0x30/0x38
> >      load_module+0x14f0/0x1e78
> >      __do_sys_finit_module+0x100/0x11c
> >      __arm64_sys_finit_module+0x28/0x34
> >      el0_svc_common+0x88/0x194
> >      el0_svc_handler+0x38/0x8c
> >      el0_svc+0x8/0xc
> >     ---[ end trace d828d06b36ad9d59 ]---
> >     ftrace failed to modify
> >     [<ffffa2dbf3a3a41c>] 0xffffa2dbf3a3a41c
> >      actual:   66:a9:3c:90
> >     Initializing ftrace call sites
> >     ftrace record flags: 2000000
> >      (0)
> >     expected tramp: ffffa2dc6cf66724
> 
> Which code specifically is this triggering for? Is this something in an
> upstream kernel, or out-of-tree patches?
> 
We faced this issue while porting Ksplice on ARM64 architecture.  So
that's an out-of-tree module. And we got this issue because we have
multiple references to '_mcount' like the one described in the commit
description of this patch.

> Can you say which toolchain you're using, too?
> 
We are using native gcc version: gcc (GCC) 7.3.0 20180125 (Red Hat 7.3.0-5)
And native binutils 2.31.1.

> > So Limit the relocation type to R_AARCH64_CALL26 as in perl version of
> > recordmcount.
> 
> Given our patching code expects each callsite to be:
> 
> 	bl	_mcount
> 
> ... this looks sane to me, and I *think* that's sound for modules too.
> 
Ok great.

> > Fixes: ed60453fa8f8 ("ARM: 6511/1: ftrace: add ARM support for C version of recordmcount")
> 
> That's a 32-bit arm commit. I suspect that was meant to be:
> 
> Fixes: af64d2aa872a1747 ("ftrace: Add arm64 support to recordmcount")
> 
Right.

> > Signed-off-by: Gregory Herrero <gregory.herrero@oracle.com>
> > ---
> >  scripts/recordmcount.c | 6 ++++++
> >  1 file changed, 6 insertions(+)
> > 
> > diff --git a/scripts/recordmcount.c b/scripts/recordmcount.c
> > index 7225107a9aaf..e59022b3f125 100644
> > --- a/scripts/recordmcount.c
> > +++ b/scripts/recordmcount.c
> > @@ -434,6 +434,11 @@ static int arm_is_fake_mcount(Elf32_Rel const *rp)
> >  	return 1;
> >  }
> >  
> > +static int arm64_is_fake_mcount(Elf64_Rel const *rp)
> > +{
> > +	return ELF64_R_TYPE(w(rp->r_info)) != R_AARCH64_CALL26;
> > +}
> > +
> >  /* 64-bit EM_MIPS has weird ELF64_Rela.r_info.
> >   * http://techpubs.sgi.com/library/manuals/4000/007-4658-001/pdf/007-4658-001.pdf
> >   * We interpret Table 29 Relocation Operation (Elf64_Rel, Elf64_Rela) [p.40]
> > @@ -547,6 +552,7 @@ static int do_file(char const *const fname)
> >  		make_nop = make_nop_arm64;
> >  		rel_type_nop = R_AARCH64_NONE;
> >  		ideal_nop = ideal_nop4_arm64;
> > +		is_fake_mcount64 = arm64_is_fake_mcount;
> >  		break;
> 
> As above, I think this is sound, but if you could answer my questions
> that'd be helpful.
> 
Thanks for the review,
Greg