Fix interactions of initial-exec TLS model and dlopen

(1) If an initial-exec relocation was used for a non-local symbol
(i.e. the definition of the symbol is in a different DSO), the
computation of the static TLS offset used the wrong DSO.
This would effectively mean the wrong address was computed
(PR toolchain/50277, PR pkg/57445).

Fix this by forcing the computation of the correct DSO (the one defining
the symbol).

This code uses __UNCONST to avoid the vast interface changes for this
special case.

(2) If symbols from a DSO loaded via dlopen are used with both
global-dynamic/local-dynamic and initial-exec relocations AND
a initial-exec relocation was resolved first in a thread, a split brain
situation could exist where the dynamic relocations would use one memory
block (separate allocation) and the initial-exec relocations the static
per-thread TLS space.

(3) If the initial-exec relocation in (2) is seen after any thread has
already used a GD/LD allocation, bail out. Since IE relocations are used
only in the GOT, this will prevent the dlopen. This is a bit more
aggressive than necessary, but a full blown reference counting doesn't
seem to be justified.

Add IRELATIVE support for ARM, X86 and PowerPC.

Replace COMBREL with just-in-time check in _rtld_relocate_nonplt_objects.

The COMBREL logic predates thread-safety of the dynamic linker and
breaks the use of shared locks for the common symbol lookup case. It is
unlikely to provide any benefit for lazy binding or PLT lookups, so
provide equivalent functionality in the non-PLT relocation handling loop
by checking if the symbol used by the current relocation is the same as
the one used during the last lookup. No inter-object cachine is done as
it is also unlikely to be benefical.

Testing with Firefox startup on AMD64 shows a small performance gain by
the new method.

branches: 1.41.8;
Add basic support for indirect functions. It allows providing a public
function symbol with an implementation choosen at run time.
Refactor calls to functions by address in ld.elf_so to create temporary
function descriptors on the stack, if the address is not leaked outside.

Limitations:
- no support for initialising static storage with function pointers
- no support for unnamed resolver functions

Inspired by FreeBSD's r228435 by kib@freebsd.org.

branches: 1.40.22;
Enable TLS for alpha - a 2nd class^Wtier system in pookaBSD^WNetBSD

alpha_reloc.c diff from joerg on port-alpha.

Add basic locking to ld.elf_so.

Introduce a new type Elf_Symindx for use in decoding the symbol hash table
section and allow this type to be overridden.

The ELF specification says it should always be uint32_t (Elf_Word), but
alpha decided to be different (not sure why). Define Elf_Symindx to be
uint64_t on alpha.

Alpha no longer uses non-standard definitions of Elf64_Sword and
Elf64_Word. Remove the ability to override these types.

Fixes ld.elf_so after Herculean effort from me and martin.

Reduce header pollution for mdreloc.c. Make Obj_Entry argument of
_rtld_relocate_nonplt_objects non-const in preparation for TLS support.

Wrap a few long lines.

Shut gcc up.

PR/39240: Satoshi Suetake: Don't fail when attempting to resolve weak symbols
when we are doing immediate binding, leave them alone and they will be dealt
with later during lazy binding. From skrll@

Remove trailing \n in calls to _rtld_error: a newline is automatically
added by a call to the function.

Copy build fix from i386: Fix const issues (cast const pointers
to "const uint8_t *" instead of "caddr_t").

branches: 1.31.4;
Refactor common code to _rtld_relocate_plt_object to i386 and arm so they
act like the other versions.
In _rtld_bind, if the result is 0, call _rtld_die.
Initialize _rtld_sym_zero.st_value to -_rtld_objself.maprelocbase. Now when
the symbol is resolved, st_value + maprelocbase will equal 0 and the above
check in _rtld_bind will fire and a call to NULL will be avoided.

branches: 1.30.12;
Remove MD COMBRELOC support since it's now MI.
Enable COMBRELOC for all arches

Remove leading __ from __(const|inline|signed|volatile) -- it is obsolete.

Add __RCSID.

KNF

appease gcc -Wuninitialized

Support LD_BIND_NOW for alpha.

ANSIfy and de-__P().

Skip the lazy PLT relocation if relocbase==0 (useful if libraries are loaded
at their VMA address).

Remove the `self' args to _rtld_relocate_objects() and
_rtld_relocate_nonplt_objects().

Push the _rtld_bind() interface into MD code -- it's just a trivial wrapper
anyway.

Add combreloc code here too.

It is possible for reloc entries to point to unaligned locations; handle
this.

Do not use the implicit addend on REFQUAD relocs (same as GLOB_DAT).

Nuke -DRTLD_RELOCATE_SELF and `dodebug' from orbit.

_rtld_bind_start() is not used in MI code, so declare it in the MD code.

Implement _rtld_relocate_nonplt_self() on Alpha.

Fix a serious performance problem for large programs on the Alpha.

Large programs need multiple GOTs. The lazy binding stub in the PLT
can be reached from any of these GOTs, but the dynamic linker only
has enough information to fix up the first GOT entry. Thus, calls
through the other GOTs went through the time-consuming lazy binding
process on every call.

This fix rewrites the PLT entries themselves to bypass the lazy binding
for those GOT entries that the dynamic linker can't fixup.

Fix from FreeBSD.

Note that now that we patch up the PLT, we need to put back the "imb"
that was removed from the binder exit path.

Pass down an additional flag to _rtld_relocate_nonplt_objects() which
indicates whether we're relocating ld.elf_so itself. Use this in some places
rather than hackish tests on `dodebug'. (The Alpha and HPPA `dodebug' tests
were actually noops, because RTLD_RELOCATE_SELF is not set, and therefore
dodebug is always true.)

Introduce a new flag, `isdynamic', which is used to remember whether the
executable was of type ET_DYN. Use this instead of `mainprog' to determine
whether we need to do base-relative fixups of the PLT. (This allows loading
non-relocatable objects, should we desire to do that at some point...)

Re-poison a lot of consts, now that the mark shite is gone.

Split _rtld_relocate_plt_object() into two MD functions:
* _rtld_relocate_plt_lazy() fixes up all the relocs pointing to the PLT. (On
most platforms it just does a simple base-relative fixup; on SPARC it does
nothing.)
* _rtld_relocate_plt_object() does immediate binding for a PLT entry.
The basic gist is that this saves a bit of time on SPARC (where the iteration
through the pltrela table was gratuitous), and a little less time on all other
platforms. A whole lot of #ifdef'ed crap is moved out of reloc.c, too.

NOT tested on: hppa sh x86_64

Partially fix up some debug printf()s that don't need to use defobj.

A few things:
* Pass a symbol number to _rtld_find_symdef(), not a r_info.
* Don't try to do a symbol lookup when we find an unsupported relocation;
instead get the symbol name from the referencing object's strtab.
* Add preliminary support for `-z combreloc'-style startup optimization on
i386, `#ifdef COMBRELOC'.

Fix one indentation error.

Rename _rtld_relocate_nonplt_object() to _rtld_relocate_nonplt_objects(),
and push the outer loop into it. This actually shaves a couple % off startup
time at least on PCs.

Split _rtld_relocate_nonplt_object() into separate MD files.

Make _rtld_setup_pltgot() a consistent interface on all platforms.

There is no need to do an imb after each call to the binder, because we only
patch the GOT, not the PLT. However, do an imb when setting up the PLT thunk,
just in case.

branches: 1.2.2;
Add code to peek into the PLT and determine which PLT entry format
is being used.

Supply two lazy binding routines for Alpha: one that works with the
old PLT format, and one that works with the new.

XXX We currently always use _rtld_bind_start_old() in
_rtld_setup_alpha_pltgot(). We need to add code to peek
into one of the PLT entries to see which format it's in
and pick the correct binding routine.

Add IRELATIVE support for ARM, X86 and PowerPC.

Replace COMBREL with just-in-time check in _rtld_relocate_nonplt_objects.

The COMBREL logic predates thread-safety of the dynamic linker and
breaks the use of shared locks for the common symbol lookup case. It is
unlikely to provide any benefit for lazy binding or PLT lookups, so
provide equivalent functionality in the non-PLT relocation handling loop
by checking if the symbol used by the current relocation is the same as
the one used during the last lookup. No inter-object cachine is done as
it is also unlikely to be benefical.

Testing with Firefox startup on AMD64 shows a small performance gain by
the new method.

branches: 1.41.8;
Add basic support for indirect functions. It allows providing a public
function symbol with an implementation choosen at run time.
Refactor calls to functions by address in ld.elf_so to create temporary
function descriptors on the stack, if the address is not leaked outside.

Limitations:
- no support for initialising static storage with function pointers
- no support for unnamed resolver functions

Inspired by FreeBSD's r228435 by kib@freebsd.org.

branches: 1.40.22;
Enable TLS for alpha - a 2nd class^Wtier system in pookaBSD^WNetBSD

alpha_reloc.c diff from joerg on port-alpha.

Add basic locking to ld.elf_so.

Introduce a new type Elf_Symindx for use in decoding the symbol hash table
section and allow this type to be overridden.

The ELF specification says it should always be uint32_t (Elf_Word), but
alpha decided to be different (not sure why). Define Elf_Symindx to be
uint64_t on alpha.

Alpha no longer uses non-standard definitions of Elf64_Sword and
Elf64_Word. Remove the ability to override these types.

Fixes ld.elf_so after Herculean effort from me and martin.

Reduce header pollution for mdreloc.c. Make Obj_Entry argument of
_rtld_relocate_nonplt_objects non-const in preparation for TLS support.

Wrap a few long lines.

Shut gcc up.

PR/39240: Satoshi Suetake: Don't fail when attempting to resolve weak symbols
when we are doing immediate binding, leave them alone and they will be dealt
with later during lazy binding. From skrll@

Remove trailing \n in calls to _rtld_error: a newline is automatically
added by a call to the function.

Copy build fix from i386: Fix const issues (cast const pointers
to "const uint8_t *" instead of "caddr_t").

branches: 1.31.4;
Refactor common code to _rtld_relocate_plt_object to i386 and arm so they
act like the other versions.
In _rtld_bind, if the result is 0, call _rtld_die.
Initialize _rtld_sym_zero.st_value to -_rtld_objself.maprelocbase. Now when
the symbol is resolved, st_value + maprelocbase will equal 0 and the above
check in _rtld_bind will fire and a call to NULL will be avoided.

branches: 1.30.12;
Remove MD COMBRELOC support since it's now MI.
Enable COMBRELOC for all arches

Remove leading __ from __(const|inline|signed|volatile) -- it is obsolete.

Add __RCSID.

KNF

appease gcc -Wuninitialized

Support LD_BIND_NOW for alpha.

ANSIfy and de-__P().

Skip the lazy PLT relocation if relocbase==0 (useful if libraries are loaded
at their VMA address).

Remove the `self' args to _rtld_relocate_objects() and
_rtld_relocate_nonplt_objects().

Push the _rtld_bind() interface into MD code -- it's just a trivial wrapper
anyway.

Add combreloc code here too.

It is possible for reloc entries to point to unaligned locations; handle
this.

Do not use the implicit addend on REFQUAD relocs (same as GLOB_DAT).

Nuke -DRTLD_RELOCATE_SELF and `dodebug' from orbit.

_rtld_bind_start() is not used in MI code, so declare it in the MD code.

Implement _rtld_relocate_nonplt_self() on Alpha.

Fix a serious performance problem for large programs on the Alpha.

Large programs need multiple GOTs. The lazy binding stub in the PLT
can be reached from any of these GOTs, but the dynamic linker only
has enough information to fix up the first GOT entry. Thus, calls
through the other GOTs went through the time-consuming lazy binding
process on every call.

This fix rewrites the PLT entries themselves to bypass the lazy binding
for those GOT entries that the dynamic linker can't fixup.

Fix from FreeBSD.

Note that now that we patch up the PLT, we need to put back the "imb"
that was removed from the binder exit path.

Pass down an additional flag to _rtld_relocate_nonplt_objects() which
indicates whether we're relocating ld.elf_so itself. Use this in some places
rather than hackish tests on `dodebug'. (The Alpha and HPPA `dodebug' tests
were actually noops, because RTLD_RELOCATE_SELF is not set, and therefore
dodebug is always true.)

Introduce a new flag, `isdynamic', which is used to remember whether the
executable was of type ET_DYN. Use this instead of `mainprog' to determine
whether we need to do base-relative fixups of the PLT. (This allows loading
non-relocatable objects, should we desire to do that at some point...)

Re-poison a lot of consts, now that the mark shite is gone.

Split _rtld_relocate_plt_object() into two MD functions:
* _rtld_relocate_plt_lazy() fixes up all the relocs pointing to the PLT. (On
most platforms it just does a simple base-relative fixup; on SPARC it does
nothing.)
* _rtld_relocate_plt_object() does immediate binding for a PLT entry.
The basic gist is that this saves a bit of time on SPARC (where the iteration
through the pltrela table was gratuitous), and a little less time on all other
platforms. A whole lot of #ifdef'ed crap is moved out of reloc.c, too.

NOT tested on: hppa sh x86_64

Partially fix up some debug printf()s that don't need to use defobj.

A few things:
* Pass a symbol number to _rtld_find_symdef(), not a r_info.
* Don't try to do a symbol lookup when we find an unsupported relocation;
instead get the symbol name from the referencing object's strtab.
* Add preliminary support for `-z combreloc'-style startup optimization on
i386, `#ifdef COMBRELOC'.

Fix one indentation error.

Rename _rtld_relocate_nonplt_object() to _rtld_relocate_nonplt_objects(),
and push the outer loop into it. This actually shaves a couple % off startup
time at least on PCs.

Split _rtld_relocate_nonplt_object() into separate MD files.

Make _rtld_setup_pltgot() a consistent interface on all platforms.

There is no need to do an imb after each call to the binder, because we only
patch the GOT, not the PLT. However, do an imb when setting up the PLT thunk,
just in case.

branches: 1.2.2;
Add code to peek into the PLT and determine which PLT entry format
is being used.

Supply two lazy binding routines for Alpha: one that works with the
old PLT format, and one that works with the new.

XXX We currently always use _rtld_bind_start_old() in
_rtld_setup_alpha_pltgot(). We need to add code to peek
into one of the PLT entries to see which format it's in
and pick the correct binding routine.