Re: [musl] [RFC v2 0/2] xtensa FDPIC port

[Rich Felker <dalias@libc.org>]

On Thu, Apr 04, 2024 at 01:44:13AM -0700, Max Filippov wrote:
> On Wed, Apr 3, 2024 at 2:45 PM Rich Felker <dalias@libc.org> wrote:
> >
> > On Wed, Apr 03, 2024 at 04:55:56PM -0400, Rich Felker wrote:
> > > On Tue, Apr 02, 2024 at 07:30:57PM -0700, Max Filippov wrote:
> > > > On Thu, Mar 28, 2024 at 6:48 PM Rich Felker <dalias@libc.org> wrote:
> > > > > On Thu, Mar 28, 2024 at 05:48:50PM -0700, Max Filippov wrote:
> > > > > > On Thu, Mar 28, 2024 at 4:00 PM Rich Felker <dalias@libc.org> wrote:
> > > > > > > On Thu, Mar 28, 2024 at 01:03:17PM -0700, Max Filippov wrote:
> > > > > > > > functional/dlopen fails with the
> > > > > > > >   src/functional/dlopen.c:39: dlsym main failed: (null)
> > > > > > > > There's no failure in the dlsym call, but the pointers don't match.
> > > > > > >
> > > > > > > Is this something related to canonical function descriptors? Is it
> > > > > > > musl's fault or a bug in the tooling? I suspect the latter.
> > > > > >
> > > > > > Yes, dlsym() returns the pointer into def.dso->funcdescs,
> > > > > > but (void *)main returns the pointer to the canonical function
> > > > > > descriptor. I understand that the linker must use the
> > > > > > R_XTENSA_FUNCDESC relocation for the locally defined
> > > > > > global symbol instead of the .rofixup entries.
> > > > >
> > > > > If the xtensa FDPIC ABI is going to be that the linker makes canonical
> > > > > function descriptors, I think that's workable, but the dynamic linker
> > > > > would need a way to find and usee them. I'm not sure how that would
> > > > > work.
> > > > >
> > > > > The simple (but probably less efficient) way is to copy what SH did
> > > > > and have the dynamic linker always be responsible for them (load
> > > > > descriptor address from GOT).
> > > >
> > > > I've built and tested SH FDPIC toolchain, it fails this test in exactly
> > > > the same way: pointer loaded directly does not match the pointer
> > > > returned by dlsym().
> > >
> > > Yes, I've been able to reproduce this and it's a clear bug. There does
> > > not seem to be any way the dynamic linker could find these GOTFUNCDESC
> > > slots to use them as a canonical address for the function, and
> > > moreover, they're not even unique; there would be one per library.
> > >
> > > The code path for legitimize_pic_address in sh.c that emits
> > > GOTFUNCDESC has the wrong logic. A simple fix would be just making
> > > that path never be taken, but I'm not sure if that would break use of
> > > GOTFUNCDESC for pure-call purposes.
> > >
> > > The condition should probably be something like: if it's just used for
> > > a call (if this is even needed; pure call is probably handled
> > > elsewhere) or if the function is static or hidden, use GOTFUNCDESC;
> > > otherwise, use GOT.
> > >
> > > I might try patching it and see what happens.
> >
> > Attached patch seems to fix it. I'm not sure if this is the most
> > idiomatic way to write the predicate in gcc sources, but it should be
> > correct.
> 
> It's not what I observe. On my side it doesn't change the result of the
> dlopen test, and it also breaks building of all statically-linked tests.
> 
> There are no relocations against the symbol 'main' neither in the test
> built with the original gcc nor in the test built with the patched one.
> dlopen.o built with the original gcc had R_SH_GOTOFFFUNCDESC
> relocation against the symbol 'main', dlopen.o built with the patched
> gcc has R_SH_GOTOFF instead. The code generated with the patched
> gcc:
> 
>        if (dlsym(g, "main") != (void*)main) {
> 28c:   50 d1           mov.l   3d0 <main+0x3d0>,r1     ! 0 <main>
> 28e:   8c 31           add     r8,r1
> 290:   12 61           mov.l   @r1,r1
> 292:   13 62           mov     r1,r2
> 294:   8f 91           mov.w   3b6 <main+0x3b6>,r1     ! 1e0
> 296:   ec 31           add     r14,r1
> 298:   23 65           mov     r2,r5
> 29a:   1c 54           mov.l   @(48,r1),r4
> 29c:   83 6c           mov     r8,r12
> 29e:   4d d6           mov.l   3d4 <main+0x3d4>,r6     ! 130
> 2a0:   03 06           bsrf    r6
> 2a2:   09 00           nop
> 2a4:   03 61           mov     r0,r1
> 2a6:   4c d2           mov.l   3d8 <main+0x3d8>,r2     ! 0 <main>
> 2a8:   8c 32           add     r8,r2
> 2aa:   20 31           cmp/eq  r2,r1
> 2ac:   27 89           bt      2fe <main+0x2fe>
> ....
> 3d8:   00 00           .word 0x0000
>                        3d8: R_SH_GOTOFF        main
> 
> doesn't look right to me at all. Using R_SH_GOTOFF
> for the symbol in text doesn't make sense. Using R_SH_GOT
> (AFAIU that's what you meant it to be) doesn't make sense
> to me as well, as the value stored in the GOT would be the
> address of the main() entry point, not of its descriptor.
> 
> I believe that gcc need to generate R_SH_GOTFUNCDESC
> instead of R_SH_GOTOFFFUNCDESC for this test to work
> correctly, and that the linker need to put R_SH_FUNCDESC
> relocation against that GOT entry, so that the dynamic linker
> could put there the address of the function descriptor associated
> with the symbol.

Indeed I messed this up. Ignore that patch. You're right that use of
R_SH_GOTOFFFUNCDESC is wrong -- it's "Used for the FDPIC-relative
offset to the function descriptor itself", which is never a constant
except for static/hidden functions or when static linking.

However, you can test switching it to R_SH_GOTFUNCDESC just by using
-fPIC, and when you link, you still get the same problem, because the
linker is breaking it by binding at ld-time. So there are at least two
bugs in play here.

Rich