This article describes ELF interposition, the linker option -Bsymbolic, and its friends. In the end, it will discuss an ambitious plan which I dubbed "the Last Alliance of ELF and Men".

Motivated by a great post by Daniel Colascione ("Python is 1.3x faster when compiled in a way that re-examines shitty technical decisions from the 1990s.") and a recent rant from Linus Torvalds on shared objects' performance issues, I have summarized the current unfortunate ELF state and filed some GCC/binutils feature requests. I believe the performance of our shared object oriented world will be no slower than one with mostly statically linked executables.

(I wrote -fno-semantic-interposition first but then realized reorganization would improve readability, so moved some parts and added some stuff to this new article.)

The ELF pioneers had "dynamic linking should be similar to static linking" in mind when designing dynamic linking. I think it means two things: (a) no source-level annotation, (b) emulating archive member extraction. Let's discuss both concepts in detail.

No source-level annotation

Say, we have two default visibility functions f and g. g calls f. g is compiled and linked into a shared object. There are 3 cases for f.

First case: f is defined in the same translation unit of g. I will discuss this in depth in my next article -fno-semantic-interposition. One notable point: GCC -fpic suppresses interprocedural optimizations including inlining for non-vague-linkage external linkage functions.

1
2
3

void f() { ... }
void g() { f(); }

Second case: f is defined in a different object file which will be linked into the same shared object.

1
2
3
4
5
6
7
8

void f() { ... }


void f();
void g() { f(); }

We will see when linking a shared object, that the first two cases need a PLT entry for f even if there is a local definition.

Third case: f is defined in a different shared object or the executable. The symbol search on f cannot be prevented.

1
2
3
4
5
6
7
8

void f() { ... }


void f();
void g() { f(); }

You can see that in all three cases no annotation is required on f and g.

ELF interposition

A component is an executable or shared object, sometimes called a module. A dynamic symbol is STB_GLOBAL or STB_WEAK.

ELF interposition means the following properties. The next chapter will give a rigid definition.

• If a dynamic symbol is defined by multiple components, they don't conflict.
• For a symbol lookup (due to a relocation like R_*_JUMP_SLOT/R_*_GLOB_DAT/absolute relocation/etc), the definition from the first component wins.
• Definitions from subsequent components are overridden.

Let's see an example about duplicate definitions in two components.

1
2
3
4
5
6
7
8

void f() { ... }



void f() { ... }
void g() { f(); }

In the static linking case, we have ld ... a.a b.a. In the dynamic linking case, assuming a.so and b.so will be linked into the executable, we have ld ... a.so b.so.

So, is ld ... a.so b.so similar to ld ... a.a b.a? I.e. Can we think of ELF interposition as an emulation of archive member extraction?

If b.a(b.o) is not extracted (i.e. the archive member does not define a symbol which is referenced previously and not provided by a preceding object file), ld ... a.so b.so is somewhat similar to ld ... a.a b.a. The g definition in b.so is shadowed, just like the g definition in b.a(b.o) is not used. (There is still a difference, though: g is defined in the dynamic linking case while undefined (because the archive member is not extracted) in the static linking case.)

However, if b.a(b.o) is extracted, we will get a duplicate definition error from ld ... a.a b.a. Well, ld ... a.so b.so does not cause an error.

Therefore, I think "dynamic linking should be similar to static linking" as justification for interposition is lame. Interposition is a simple scheme, and convenient in some scenarios, but inefficient, error-prone, and less secure.

Specification

Since 2000-07-17, the ELF specification says the following for the STV_DEFAULT visibility (this is the default visibility. You get this unless you do thing like -fvisibility= or __attribute__((visibility(...)))):

Global and weak symbols are also preemptable, that is, they may by preempted by (typo: be) definitions of the same name in another component."

In Chapter 5 Dynamic Linking, the specification says:

When the dynamic linker creates the memory segments for an object file, the dependencies (recorded in DT_NEEDED entries of the dynamic structure) tell what shared objects are needed to supply the program's services. By repeatedly connecting referenced shared objects and their dependencies, the dynamic linker builds a complete process image. When resolving symbolic references, the dynamic linker examines the symbol tables with a breadth-first search. That is, it first looks at the symbol table of the executable program itself, then at the symbol tables of the DT_NEEDED entries (in order), and then at the second level DT_NEEDED entries, and so on. Shared object files must be readable by the process; other permissions are not required.

The wording remains unchanged since then, i.e. the evolution of dynamic linking has not contributed back to the specification.

This paragraph is probably difficult to follow. Let me rephrase it with some additions of dynamic loader behaviors. The dynamic loader does one critical job: resolving dynamic relocations and binding symbol references from one component to another. There is a flat namespace for symbol search. The dynamic loader computes a breadth-first search list (executable, needed0, needed1, needed2, needed0_of_needed0, needed1_of_needed0, ...). For each symbol reference, the dynamic loader iterates over the list and finds the first component which provides a definition. (For dlsym with an explicit handle, the symbol search uses the dependency order, a breadth-first search rooted at the handle.)

The implication is that STB_GLOBAL and STB_WEAK definitions are equivalent in terms of symbol search. A STB_WEAK definition can preempt a STB_GLOBAL definition.

While not mentioned in the ELF specification, many dynamic loader implementations allow the environment variable LD_PRELOAD to inject shared objects. The effect is like the LD_PRELOAD list is inserted at the beginning of the executable's DT_NEEDED list. The search list may look like executable, preload0, preload1, needed0, needed1, needed2, needed0_of_preload0, ..., needed0_of_needed0, needed1_of_needed0, ... (If the program calls dlopen with RTLD_GLOBAL, the newly loaded component and its dependencies (if not loaded) will be appended to the list.) Here is the algorithm:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17

fn load(c) {
  ...
  if tail != null { tail->next = &c }
  tail = &c
}

head = &exe
tail = null
load(exe)
for c in LD_PRELOAD {
  load(c)
}
for (c = head; c; c = c->next) {
  if c.loaded { continue; }
  for c1 in c.needed
    load(c1)
}

Preemptibility in the linker

The executable is always the first element of the search list, so a defined symbol of any binding in the executable cannot be preempted (interposed).

For a shared object, a default visibility STB_GLOBAL or STB_WEAK symbol can be preempted (interposed) because a preceding component may define a symbol of the same name.

In the following example, f is preemptible even if it is locally defined.

1
2
3

void f() { ... }
void g() { f(); }

To make it possible for the branch instruction to jump to a different definition, the linker resolves the branch target to a PLT entry. The PLT entry has a code sequence loading the resolved address from an associated GOT entry. The GOT entry is relocated by R_*_JUMP_SLOT.

Two places have costs:

• The dynamic loader performs a symbol search for R_*_JUMP_SLOT, and possibly for R_*_GLOB_DAT/absolute relocations.
• Every call site goes through a PLT indirection.

Alternative symbol search models

Solaris named the above the default search model and introduced an alternative model: direct bindings. With -z defs, one can ensure the dependencies are provided as part of the link and all symbol references are satisfied. The linker can record the bound component for each symbol reference.

Here is an example from Solaris's Linkers and Libraries Guide:

1
2
3

$ elfdump -y W.so.2
[6] [ DEPEND DIRECT ]          <self>  a
[7] [ DEPEND LAZY DIRECT ] [1] w.so.1  b

With the information about the component name, the dynamic loader can speed up its symbol search by just looking at one component. In particular, frequently the bound component is the component itself.

In Mac OS X, the two-level namespace introduced in 10.1 (default unless you use ld -flat_namespace) is a similar model.

Prelink can be conceived as a direct binding model without great ergonomics.

The ELF specification defines DF_SYMBOLIC which can be conceived as a special case of direct bindings. When a shared object is marked as DF_SYMBOLIC (set by ld -Bsymbolic), the symbol search checks the shared object itself before starting the linear search from the executable. It is quite common for a shared object to call STV_DEFAULT definitions in itself. DF_SYMBOLIC can improve the performance greatly.

-Bsymbolic

The linker option -Bsymbolic can be used together with -shared. ld -shared -Bsymbolic is very similar to -pie.

-Bsymbolic follows ELF DF_SYMBOLIC semantics: all defined symbols are non-preemptible. This can optimize relocation processing:

• function calls: a branch instruction (e.g. call [email protected]) will not create a PLT entry. The associated R_*_JUMP_SLOT dynamic relocation will be suppressed.
• variable access and function addresses: the GOT entry will not cause a R_*_GLOB_DAT dynamic relocation. On x86-64, with R_X86_64_GOTPCRELX/R_X86_64_REX_GOTPCRELX, the GOT indirection code sequence can be rewritten. However, the code sequence is still longer than that without GOT. On PowerPC64, there is a similar TOC optimization. On other architectures, there is no difference in code sequences.

-fno-semantic-interposition can save a PLT if all call sites are in the defining translation unit. -Bsymbolic can address cross-translation-unit pessimization which cannot be optimized with -fno-semantic-interposition. Personally I think -Bsymbolic claims most of the direct binding benefits.

As a data point, when building the Linux kernel's x86_64 defconfig with a clang -fPIC built clang, my build is 15% faster if I add -Bsymbolic-functions to libLLVM.so and libclang-cpp.so. I cannot tell the performance difference with a mostly statically linked PIE clang. From llvm-project 13.0.0 onwards, the build system uses -Bsymbolic-functions by default.

However, in practice, deployment of -Bsymbolic may run into pointer equality problems. We will discuss variables and functions separately. In practice the most serious problem is copy relocations.

Pointer equality for variables

An inline variable with external linkage and a local static variable defined in an inline function with external linkage are required to be unique. The address of such a variable seen by a -Bsymbolic linked shared object may be different from the address seen from outside the shared object.

Fortunately it is uncommon to export such a vague linkage variable to more than one component. Some projects roll their own typeid mechanism and need caution, e.g. llvm::Any::TypeID, mlir::TypeID.

1
2
3
4
5
6
7
8
9
10
11
12
13

inline int *addr() {
  static int data;
  return &data;
}


#include "a.h"
int *addr0 = addr();


#include "a.h"
int *addr1 = addr();

(ELF specific) In addition, a regular non-inline variable with external linkage can cause incompatibility problems due to copy relocations. GCC/Clang -fno-pic emits direct access relocations referencing a global variable. If the global variable turns out to be defined in a shared object, there will be a copy relocation in the executable. The object the shared object sees and the executable sees will be different.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

extern int var;


int var;


int *addr0() { return &var; }


int main() {
  
  
  int *addr1 = &var;
}

For Clang -fno-pic, the direct access relocation can be avoided with -fno-direct-access-external-data. GCC feature request: PR98112.

Since GCC 5, on x86-64, -fpie can cause copy relocations as well due to HAVE_LD_PIE_COPYRELOC. We should fix it. Pending GCC patch.

(In C++, typeid() on an incomplete class can define a typeinfo name object. A -Bsymbolic linked shared object may see a different copy, but the address can hardly cause a problem.).

Pointer equality for functions

Many objects in C++ are not clearly part of a single object file, but are required by the ODR to have a single definition. For example, C++ [dcl.inline]: "An inline function or variable with external or module linkage can be defined in multiple translation units ([basic.def.odr]), but is one entity with one address. A type or static variable defined in the body of such a function is therefore a single entity."

The address of an inline function seen by a -Bsymbolic linked shared object may be different from the address seen from outside the shared object. Fortunately such cases are rare. ELF/Mach-O programs may use -fvisibility-inlines-hidden to break such pointer equality. On Windows, correct dllexport and dllimport are needed to make pointer equality work. On the other hand, Windows link.exe enables identical COMDAT folding (/OPT:ICF) by default so different inline functions may have the same address.

On Mach-O, such symbols are placed into __LINKEDIT,__weak_binding so that dyld can coalesce the definitions across dylibs.

On Windows, you need to compile the DLL and the executable differently: the defining DLL needs __declspec(dllexport) on the inline function and the executable needs __declspec(dllimport). This is tricky.

1
2
3
4
5
6
7

__declspec(dllexport) inline void f() {}
__declspec(dllexport) void *g() { return (void *)&f; }



__declspec(dllimport) inline void f() {}

(ELF specific) In addition, a regular non-inline function with external linkage can cause incompatibility problems due to canonical PLT entries. GCC/Clang -fno-pic emits direct access relocations when taking the address of an external function. If the global variable turns out to be defined in a shared object, there will be a canonical PLT entry in the executable. The function address the shared object sees and the executable sees will be different.

1
2
3
4
5
6
7
8
9
10

void fun();


void fun() {}
void *addr0() { return (void *)&fun; }



void *addr1() { return (void *)&fun; }

-Bsymbolic-functions

The function incompatibility problems are uncommon. It is often benign when the function address seen by a shared object is different from outside the shared object. However, the variable case is usually severe: the executable and a shared object may act on different copies of a variable supposed to be the same entity.

In practice, we can usually use the linker option -Bsymbolic-functions. The option applies to STT_FUNC symbols in ld.lld and non-STT_OBJECT symbols in GNU ld and gold, avoiding variable incompatibility problems.

-Bsymbolic-non-weak-functions

The previous paragraph "Pointer equality for functions" has mentioned two problems. To address the pointer equality problem for vague linkage functions, we can introduce a linker option -Bsymbolic-non-weak-functions which applies to STT_FUNC STB_GLOBAL symbols to bypass vague linkage STB_WEAK symbols. GNU ld feature request: PR27871.

-Bsymbolic-non-weak-functions provides an escape hatch for rare cases where definition interposition is needed: such declarations can be annotated as __attribute__((weak)).

To address the canonical PLT entry problem, I think the best is to change the compiler's -fno-pic behavior: use GOT to take the address of an external default visibility function. GCC feature request: PR100593. A stage 2 clang is byte identical when I apply the change.

Most pieces of software will just work, because using GOT is how -fpie/-fpic work. The Linux kernel is an unfortunate outlier. Its x86 port has something like asm("call %P[new2]" :: [new2]"i"(clear_page_rep)); and such "i" usage only works if the non-definition clear_page_rep is directly accessed. In addition, more ports start to assume .got is empty. I think the Linux kernel should use an explicit option to enable absolute relocations for function declaration addresses.

Technically we could add an alternative linker option -Bsymbolic-plt which resolves PLT-generating relocations at link time if the referenced symbol is defined and keeps the previous behavior for other relocation types (GOT-generating, absolute, PC-relative, etc). However, I don't like making preemptibility different for different relocation types. In addition, once we fix the canonical PLT entry problem (we should, as it is evil like copy relocations), such an option will provide no additional usage (its effect on weak definitions can be solved with -fvisibility-inlines=hidden).

Visibility

-fvisibility=protected

A non-default visibility symbol cannot be preempted, even if the binding is STB_WEAK. -fvisibility=protected can make all definitions protected and thus non-preemptible, nullifying the performance benefit of -fno-semantic-interposition and -Bsymbolic. Note: if you want a definition to be preemptible, you will need a default visibility attribute, even if it is weak (e.g. __attribute__((weak,visibility("default")))).

However, -fvisibility=protected shares the same problem with -Bsymbolic: too coarse-grained. It can cause the same sets of problems as discussed above in Pointer equality for variables and functions. Notably, a shared object built with -fvisibility=protected is incompatible with -fno-pic global variable access.

In GCC/binutils's x86 port, there were some changes in the wrong direction during the GCC 5 era. As a result, there are some STT_OBJECT issues resulting in poor Clang interoperability (and also gold).

1
2
3
4
5
6
7
8
9
10
11

% cat a.s
leaq foo(%rip), %rax

.data
.global foo
.protected foo
foo:
% gcc -fuse-ld=bfd -shared a.s
/usr/bin/ld.bfd: /tmp/ccchu3Xo.o: relocation R_X86_64_PC32 against protected symbol `foo' can not be used when making a shared object
/usr/bin/ld.bfd: final link failed: bad value
collect2: error: ld returned 1 exit status

See Copy relocations, canonical PLT entries and protected visibility for details. There is no problem when you only use Clang and LLD.

-fvisibility=hidden

-fvisibility=hidden can make all definitions hidden and thus non-preemptible, nullifying the performance benefit of -fno-semantic-interposition.

-fvisibility=hidden requires annotation of exported symbols (__attribute__((visibility("default")))). The explicit annotation sometimes makes it inconvenient to split and join libraries.

However, projects with Windows portability in mind will define macros to dispatch to either the visibility attribute or __declspec(dllexport).

Personally I think that explicit annotation is the correct way of exposing API. However, the adoption may not be that high.

-fvisibility-inlines-hidden

The C++ specific -fvisibility-inlines-hidden is a safer subset of -fvisibility=hidden. The option just violates pointer equality for inline function definitions. As discussed above, this is usually safe.

Interaction with LD_PRELOAD

There are several types of LD_PRELOAD usage.

First, use LD_PRELOAD=same_soname.so to replace a DT_NEEDED entry with the same SONAME. Both -fno-semantic-interposition and -Bsymbolic are compatible with such usage.

Second, use LD_PRELOAD=malloc.so to intercept some functions not defined in the application or any of its shared object dependencies. Both -fno-semantic-interposition and -Bsymbolic are compatible. Common examples include malloc replacement and fakeroot, both interposing some libc.so functions.

1

void *f() { return malloc(0xb612); }

Third, use LD_PRELOAD=different_soname.so to replace a non-vague-linkage function defined in a shared object dependency and the SONAME is different. (This usage is unlikely compatible with C++'s one definition rule.) Such usage is incompatible with -Bsymbolic, -Bsymbolic-functions, and (if non-weak) -Bsymbolic-non-weak-functions. If -fno-semantic-interposition causes an inlining and the call site is not intercepted, there is an incompatibility issue.

Note: it is unfair and usually incorrect to just state "it breaks LD_PRELOAD". Please categorize your LD_PRELOAD use cases.

The Last Alliance of ELF and Men

You may want to read my -fno-semantic-interposition first. This section formulates an ambitious plan "the Last Alliance of ELF and Men" and also serves as a summary. For easy navigation, You can click the {}-style links to jump back to the anchors defined in previous paragraphs.

Choice (a) Compiler option oriented

We need a variant of -fvisibility=protected which applies to STB_GLOBAL STT_FUNC symbols.

The GNU ld protected symbol bugs described in Copy relocations, canonical PLT entries and protected visibility must be fixed.

Choice (b) Linker option oriented

We need -Bsymbolic-non-weak-functions (a variant of -Bsymbolic-functions) which only applies to STB_GLOBAL symbols (i.e. STB_WEAK symbols are excluded). {-Bsymbolic-non-weak-functions}

I wish that distributions default to -fno-semantic-interposition and (in the long term) -Wl,-Bsymbolic-non-weak-functions, bringing back the lost performance for decades. macOS (Mach-O), Windows (PE-COFF), and Solaris (ELF) direct bindings have set up precedent so there is a good chance that most pieces of portable software are already in a good state.

We need a linker option to cancel default -Bsymbolic-non-weak-functions. I have added -Bno-symbolic to GNU ld and gold (binutils 2.37; PR27834) and ld.lld 13.

(From Peter Smith) The linker can introduce a debugging option for executables to catch accidental interposition, say, --warn-interposition: "Warning symbol S of type STT_FUNC is defined in executable A and shared objects B and C, using definition in A."

Next steps

There is some amount of work to annotate software which cannot be built with the non-weak function -fvisibility=protected or -Wl,-Bsymbolic-non-weak-functions.

In return, I estimate that many pieces of software may be 5% to 20% faster (CPython is 1.3x faster) and a few percent smaller in size.

It turns out that some GCC folks don't like a configure-time option making -fno-semantic-interposition the default: PR100937.

There is a trade-off and the downside is that LD_PRELOAD replacing a function definition in a shared object will be a non-default choice. The users can build the software by themselves.

GCC -fno-pic should use GOT to take the address of an external default visibility function. PR100593. {-fno-pic_got} Clang's behavior just emulates GCC.

The first stage of the plan is heavily throttled at GCC folks' mercy. The second stage of the plan is throttled at distributions' interest.

Copy relocations

With our function oriented non-interposition-by-default plan, we will not be blocked by copy relocation elimination. Given the long historical issues, I don't expect that copy relocations can be addressed any time soon (say: 10 years). Nevertheless, here is the plan.

We should fix the GCC 5 x86-64 -fpie HAVE_LD_PIE_COPYRELOC mistake. PATCH. {HAVE_LD_PIE_COPYRELOC}

GCC can introduce -fno-direct-access-external-data to avoid -fno-pic copy relocations. PR98112. In the x86-64 world, there are actually more problems with copy relocations and protected data symbols in GCC and GNU ld. I will recommend that interested readers read the summary of Copy relocations, canonical PLT entries and protected visibility.