Re: [PATCH v2 0/5] Fix ELF / FDPIC ELF core dumping, and use mmap_sem properly in there

[Rob Landley]

On 5/1/20 1:00 AM, Greg Ungerer wrote:
>> This sounds correct. My understanding of FLAT shared library support
>> is that it's really bad and based on having preassigned slot indices
>> for each library on the system, and a global array per-process to give
>> to data base address for each library. Libraries are compiled to know
>> their own slot numbers so that they just load from fixed_reg[slot_id]
>> to get what's effectively their GOT pointer.

fdpic is to elf what binflt is to a.out, and a.out shared libraries were never
pretty. Or easy.

>> I'm not sure if anybody has actually used this in over a decade. Last
>> time I looked the tooling appeared broken, but in this domain lots of
>> users have forked private tooling that's not publicly available or at
>> least not publicly indexed, so it's hard to say for sure.
> 
> Be at least 12 or 13 years since I last had a working shared library
> build for m68knommu. I have not bothered with it since then, not that I
> even used it much when it worked. Seemed more pain than it was worth.

Shared libraries worked fine with fdpic on sh2 last I checked, it's basically
just ELF PIC with the ability to move the 4 segments (text/rodata/bss/data)
independently of each other. (4 base pointers, no waiting.)

I don't think I've _ever_ used shared binflt libraries. I left myself
breadcrumbs back when I was wrestling with that stuff:

  https://landley.net/notes-2014.html#07-12-2014

But it looks like that last time I touched anything using elf2flt was:

  https://landley.net/notes-2018.html#08-05-2018

And that was just because arm's fdpic support stayed out of tree for years so I
dug up binflt and gave it another go. (It sucked so much I wound up building
static pie for cortex-m, taking the efficiency hit, and moving on. Running pie
binaries on nommu _works_, it's just incredibly inefficient. Since the writeable
and readable segments of the ELF are all relative to the same single base
pointer, you can't share the read-only parts of the binaries without address
remapping, so if you launch 4 instances of PIE bash on nommu you've loaded 4
instances of the bash text and rodata, and of course none of it can even be
demand faulted. In theory shared libraries _do_ help there but I hit some ld.so
bug and didn't want to debug a half-assed solution, so big hammer and moved on
until arm fdpic got merged and fixed it _properly_...)

Rob

P.S. The reason for binflt is bare metal hardware engineers who are conceptually
uncomfortable with software love them, because it's as close to "objcopy -O
binary" as they can get. Meanwhile on j-core we've had an 8k ROM boot loader
that loads vmlinux images and does the ELF relocations for 5 years now, and ever
since the switch to device tree that's our _only_ way to feed a dtb to the
kernel without statically linking it in, so it's ELF all the way down for us.