Re: [PREVIEW 10/10] linkage: add .cfi_{start/end}proc to ENTRY/ENDPROC

From: Josh Poimboeuf <jpoimboe@redhat.com>
To: Jiri Slaby <jslaby@suse.cz>
Cc: mingo@redhat.com, tglx@linutronix.de, hpa@zytor.com,
	x86@kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PREVIEW 10/10] linkage: add .cfi_{start/end}proc to ENTRY/ENDPROC
Date: Fri, 17 Feb 2017 15:18:04 -0600	[thread overview]
Message-ID: <20170217211804.j6l2d7t5mfzqzmbt@treble> (raw)
In-Reply-To: <7f613a09-39f8-5477-ed5e-17024baf5876@suse.cz>

On Fri, Feb 17, 2017 at 03:26:36PM +0100, Jiri Slaby wrote:
> On 02/17/2017, 03:07 PM, Josh Poimboeuf wrote:
> > On Fri, Feb 17, 2017 at 02:36:15PM +0100, Jiri Slaby wrote:
> >> On 02/17/2017, 02:16 PM, Josh Poimboeuf wrote:
> >>> On Fri, Feb 17, 2017 at 11:47:57AM +0100, Jiri Slaby wrote:
> >>>> This is just a preview, not to merged now, only later with DWARF
> >>>> unwinder series. This is what the series will serve for (aside from
> >>>> cleanup and unification).
> >>>>
> >>>> I am aware of CFI_STARTPROC and CFI_ENDPROC left defined in other archs
> >>>> in spite of cfi annotations removal ages ago. For simplicity. I am using
> >>>> DW_ prefix here.
> >>>
> >>> If objtool is going to be generating CFI instructions, why not have it
> >>> generate .cfi_startproc and .cfi_endproc too?
> >>
> >> I tried that, but in many places it is very hard to recognize start
> >> and/or end of a function.
> > 
> > How so?  objtool already knows where the start and end of every function
> > is due to ENDPROC's use of the .type and .size macros.
> 
> Right, I did not realized that we are writing about different things. So
> yes, when the series is applied, we have ENTRY, ENDPROC et al. at all
> appropriate places. We can indeed use that info.
> 
> >> Having .cfi_startproc and .cfi_endproc in place makes it rather easy,
> >> actually reduced to "emit dwarf instructions for this code between
> >> here and there". Plus pre-prepared .eh_frame only to be extended.
> >> (.eh_frame_hdr has to be rehashed of course.)
> > 
> > Hm, but now objtool has to read *and* write CFI instead of just writing.
> 
> That is needed due to inline assembly anyway :/. So the way I wanted to
> go was: here you have code with possibly incomplete (but at least some)
> CFIs, fix the broken ones. Otherwise we would have to differentiate 2
> kind of files:
> * .c files with inline assembly (read-write = update CFIs)
> * .S files with "native" assembly (write eh_frame header and also CFIs)

Have you seen any real inline asm issues?  I had been thinking they
weren't a problem, because CFI instructions are only emitted for the
function prologue and epilogue.  Running objtool with CFI analysis
seemed to confirm that -- I don't remember seeing any inline asm issues
with CFI like we did with frame pointers.

So my thinking for objtool CFI was:

.c files: read-only
.S files: write-only

> > It would help to see the generation code.  Have you written it?
> 
> I started writing it and it is complete crap so far :P. Please see
> "objtool: generate dwarf for asm" at:
> https://git.kernel.org/cgit/linux/kernel/git/jirislaby/linux.git/log/?h=devel
> 
> And yes, the current code is for .S files without any .eh_frame.

Nice!  Does it work?

> Then I decided to clean up ENTRY/ENDPROC etc. first. It's up to
> discussion what route we want to go, but I would prefer the
> "read-write" since we have to implement it anyway.

Yes, that would make sense if we have to do read-write for .c files.

-- 
Josh