Solaris dldump (was: Pure space)

[ali_gnu2@emvision.com]

On 8/17/24 6:10 PM, Po Lu wrote:
> Hello Ali!
> 
> I think you underestimate the number of programs using dldump.  I've
> seen both Perl 5 and GNU Make hacked to save state with dldump, on
> Oracle Solaris, producing binaries that don't depend on the presence of
> a state file and probably start faster as well.  Meanwhile
> pdumper-dumped binaries appear to crash in an x86 Solaris 10 zone,
> though I don't really use this configuration and I'm not interested in
> trying the portable dumper on sparc:
> 
> core 'core' of 7021:        ../../src/bootstrap-emacs -batch --no-site-file --no-site-lisp -f batc
>   00007fffaf433dc2 ???????? ()
>   00007fffaf5eb3d7 ???????? ()
>   00007fffaf5ec590 ???????? ()
>   00007fffae3f351a _lwp_kill () + a
>   00007fffae3981b9 raise () + 19
>   00000000008baf90 terminate_due_to_signal () + c0
>   000000000090236e ???????? ()
>   0000000000902334 deliver_thread_signal () + 74
>   00000000009023b0 deliver_fatal_thread_signal () + 10
>   00000000009024ef handle_sigsegv () + 4f
>   00007fffae3edd16 __sighndlr () + 6
>   00007fffae3e25e2 call_user_handler () + 252
>   00007fffae3e280e sigacthandler () + ee
>   00007fffaf5ea82d ???????? ()
>   ffffffffffffffff ???????? ()
>   00000000009c77e7 lisp_align_malloc () + 4d7
>   00000000009c9dd2 make_float () + 42
>   00000000009d2e9d init_alloc () + d
>   00000000008bd373 main () + bb3
>   00000000006d15ab ???????? ()
> 

Hello!

    Is that stack from the s10 zone?

You're probably right that I don't know who is using
dldump(), outside of emacs, but not to worry, it's not
going away. It's a committed interface, so hard to remove,
and at the same time, isn't causing any problems. Nonetheless,
it's not our favored way to deploy emacs, and I wouldn't want
anyone to think we prefer its use, or require it.

We use pdumper on newer Solaris 11.4, both x86 and sparc,
with no reported issues. I wasn't aware of the Solaris 10
zone problems (haven't seen any reports). If you end up
looking at it, and think that the s10 zone is somehow at
fault, please feel free to contact me offline. However,
given that s10 is 20 years old, it wouldn't be unreasonable
to drop it off the support tail. From discussions with Rainer
Orth, who maintains gcc for Solaris, I believe that s10 support
for gcc has ended, or is very close to ending. My personal
opinion is that anyone happy to use a 20 year old OS should
have no problem using an older gcc, or emacs, so it's not
really the end of the road for those folks.

>> There are open source variants of Solaris for whom I don't
>> speak, but from what I know about our common code, they should
>> not be any more stuck on unexec() than we are. pdumper really
>> doesn't use any unix features that didn't exist decades ago.
> 
> I don't believe we try to support Illumos.  If Emacs should work, more
> power to them, but they have bigger fish to fry when GCC exception
> handling fails if an exception is raised the instant an object is
> unmapped, prompting dl_iterate_phdr to return -1.

I expect that we're both benefiting from your work anyway.
Isn't emacs still largely C (not C++)? I wouldn't expect
exception handling to be needed, so maybe it's OK.

I do know that dl_iterate_phdr() is a relatively recent
addition for us, and was done after the split, so that
is a case where the code is not common. No doubt the fix
for Illumos would not be difficult, if/when they get to it.

Thanks!

- Ali