Re: [PATCH v3] strip: keep .ctf section in stripped file

[Nix <nix@esperi.org.uk>]

On 7 Mar 2023, Mark Wielaard said:

> Hi Nick, Hi Guillermo,
>
> On Sat, 2023-03-04 at 08:00 -0600, Guillermo E. Martinez wrote:
>> Hello,
>> 
>> On Fri, Mar 03, 2023 at 12:24:19PM +0000, Nick Clifton wrote:
>> > Hi Mark,
>> > 
>> > > > > I am mainly wondering why binutils strip already seems to keep ".ctf"
>> > > > > sections (even without -g).
>> > 
>> > > Right. But I am wondering how that mechanism works with binutils strip.
>> > > Apparently we do something different in eu-strip which makes it
>> > > necessary to add a --keep-ctf option based on the section name. It
>> > > would be good if we figured out how/what we can do to keep the
>> > > different strip utilities in sync.
>> > 
>> > I think that the function that does this is "is_strip_section_1" in
>> 
>> That's right, this function decided whether the section will be striped
>> out by default, if it has set `SEC_DEBUGGING' in BFD section flags then
>> section is removed by the default.  For `.ctf' this flag is not set
>> because in `_bfd_elf_make_section_from_shdr' when BFD sections are
>> building it uses the section's name: ".debug", ".gnu.debuglto_.debug_",
>> ".gnu.linkonce.wi.", ".zdebug", "lines", ".stab", etc, to set
>> `SEC_DEBUGGING' flag.
>> 
>> > binutils/objcopy.c.  If an input section has the BSF_DEBUGGING flag
>> > set (an internal flag to he BFD library, but basically it should be
>> > set for all debug sections, including .ctf sections I think), then
>> > the basic decision is to keep the section unless -g is used.
>
> OK, that is interesting. So given the .ctf section is NOT marked as
> SEC_DEBUGGING binutils strip basically never strips it.

Can I chime in?

This behaviour is in line with the intended purpose of CTF, which is to
give C programs introspection into into their type systems at all times,
via libctf, so we can use it for ABI checking, shared library interface
construction at runtime (for dlopen()) and things like that. This is why
keeping CTF small is basically our sole format priority: if it's big,
people will be tempted to strip it out, and the format becomes useless.
(This is also why it's not loadable -- libctf reads it itself so that
it doesn't cost time paging it in if it's not used.)

(This is also why libctf doesn't follow gnu_debuglink pointers: it can
only read CTF from the ELF object itself.)

> While eu-strip does strip it by default (since it is a non-loadable
> PROGBITS section), but keeps it with -g (which only strips the
> explicitly named .debug sections).

It should probably keep it regardless, unless you do some sort of
"really brutal strip" that strips out literally everything not needed
for the program to start at all. It might well be necessary at runtime
(in future, anyway).

> So binutils strip and eu-strip aren't totally identical with the
> default flags, but with --keep-section=.ctf and --remove-section=.ctf
> they seem to do the same thing.

Yeah, this is a bug -- I simply forgot that eu-strip existed :/ sorry!

> So I am not sure we really need a --keep-ctf flag for eu-strip (unless
> we also get it for binutils strip). My preference would be to just
> recommend a user use --keep-section=.ctf or --remove-section=.ctf to
> indicate what they want.

My strong preference would be to make eu-strip do what binutils strip
does. If CTF is not in the binary, the whole format is useless. If
you've got room for the DWARF, the CTF is a strict subset of it anyway
(it's generated from it, after all).

-- 
NULL && (void)