[llvm-dev] LLD support for mach-o aliases (weak or otherwise)

303 views
Skip to first unread message

Michael Clark via llvm-dev

unread,
Jun 6, 2017, 7:09:29 PM6/6/17
to llvm-dev
Hi Folks,

I’m working on a port of musl libc to macos (arch triple is “x86_64-xnu-musl”) to solve some irreconcilable issues I’m having with libSystem.dylib. I don’t want to use glibc for various reasons, mainly because I want to static link. I have static PIE + ASLR working which is not actually supported by the Apple toolchain (*1), but I managed to get it to work. I’m sure Apple might say “Don’t do that”, but from looking at the history of the xnu kernel ABI, it seems to be very stable between versions.

In any case the musl libc source makes extensive use of weak aliases, perhaps to allow easier interposition of C library routines, however aliases, weak or otherwise are not currently supported by ld64.

It appears that the mach-o format supports aliases, but the functionality has not been exposed via the linker (ld64/LLD).


The musl code does the following which currently errors out saying aliases are not currently supported:

#undef weak_alias
#define weak_alias(old, new) \
        extern __typeof(old) new __attribute__((weak, alias(#old)))

and the macro is used internally like this:

int __pthread_join(pthread_t t, void **res)
{
        // implementation here
}

weak_alias(__pthread_join, pthread_join);

The problem is the actual export used by clients is an alias and I want to maintain source compatibility.

I seem to have found a way to semi-emulate aliases (at least within one module). My goal is to at least turn them into strong aliases somehow, so I can at a minimum make the musl source compatible with clang on macos. The following compiles but foo is not exported:

$ cat a.c
#include <stdio.h>

void foo() __attribute__((weak_import)) __asm("_bar");

void bar()
{
        printf("bar\n");
}

int main()
{
        foo();
}

$ cc -c a.c -o a.o
$ nm a.o
0000000000000000 T _bar
0000000000000020 T _main
                 U _printf

Any ideas how I can get foo as an exported symbol? 

Is weak alias or plan alias support planned for mach-o in LLD?

The goal at a minimum is to make the weak_alias macro emit a strong alias with clang/ld64 or clang/LLD? so I don’t need to diverge too much from the upstream musl source (as the lack of alias support currently requires me to rename function declarations in the source). Of course pthreads which I’m working on now are going to be completely different… but musl has support for architecture specific overrides in its build system.

BTW I now have some quite non-trivial programs compiling against musl-xnu + libcxx + libcxxabi on macos.

There are a lot of libcxx changes like this:

-#ifdef __APPLE__
+#if defined(__APPLE__) && !defined(_LIBCPP_HAS_MUSL_LIBC)

Michael.

Michael Clark via llvm-dev

unread,
Jun 7, 2017, 9:22:49 PM6/7/17
to llvm-dev
I believe I have solved the problem at least such that I can maintain source compatibility with musl’s use of a weak_alias macro and calls through the aliases. Thanks to Jameson who gave me some off-list advice. I would like to post this to the mailing list so that others who are trying to get weak aliases working on Mach-O might find this in Google versus what I found which was error messages and no apparent working solutions.

From the nm output, it seems the approach below is generating a strong direct alias, however in my case that is satisfactory as I don’t need interposition support, rather my goal is source compatibility. It seems weak_import has no effect when the symbol is defined in the same module. i.e. it does not create weak linkage, and if I don’t define the symbol with .set I get unresolved symbol on link.

It seems we need N_INDR support as per the blog post and in this commit, and some syntax to coax the assembler into making a weak indirect alias:


Here is my strong direct alias, versus ideally what should be a weak indirect alias. It works for my use case (source compat).


$ cat alias.c 
#include <stdio.h>

#if HAVE_ALIASES

#define weak_alias(old, new) \
        extern __typeof(old) new __attribute__((weak, alias(#old)))

#else

#define weak_alias(old, new) \
__asm__(".globl _" #new); \
__asm__(".set _" #new ", _" #old); \
extern __typeof(old) new __attribute__((weak_import))

#endif

void __foo()
{
printf("foo\n");
}

weak_alias(__foo, foo);

int main()
{
foo();
}
$ cc -DHAVE_ALIASES=1 -c alias.c -o alias.o
alias.c:22:1: error: only weak aliases are supported on darwin
weak_alias(__foo, foo);
^
alias.c:6:55: note: expanded from macro 'weak_alias'
        extern __typeof(old) new __attribute__((weak, alias(#old)))
                                                      ^
1 error generated.
$ cc -c alias.c -o alias.o
$ nm -x alias.o 
0000000000000000 0f 01 0000 00000001 ___foo
0000000000000000 0f 01 0000 00000003 _foo
0000000000000020 0f 01 0000 00000008 _main
0000000000000000 01 00 0000 0000000e _printf
$ cc alias.o
$ ./a.out 
foo

Michael Clark via llvm-dev

unread,
Jun 7, 2017, 11:31:10 PM6/7/17
to llvm-dev
I think I have weak indirect aliases working on Mach-O. Can anyone verify this?

Note, the .set nor .weak_ref create an indirect alias, but after reading Tim’s patch, I changed the macro to use = and now the symbol has a lowercase t in nm output and .globl was no longer required to get it to appear in the symbol table.

/*

        $ cc -c alias.c -o alias.o
        $ cc alias.o
        $ ./a.out 
        foo
        $ nm alias.o
        0000000000000000 T ___foo
        0000000000000000 t _foo

        0000000000000020 T _main
                         U _printf
*/


#include <stdio.h>

#if HAVE_ALIASES

#define weak_alias(old, new) \
        extern __typeof(old) new __attribute__((weak, alias(#old)))

#else

#define weak_alias(old, new) \
        __asm__("_" #new " = _" #old); \

        extern __typeof(old) new __attribute__((weak_import))

#endif

void __foo()
{
        printf("foo\n");
}

weak_alias(__foo, foo);

int main()
{
        foo();
}

Michael Clark via llvm-dev

unread,
Jun 7, 2017, 11:50:41 PM6/7/17
to llvm-dev
Sorry for the noise…

The previous version generated a non-external symbol and didn’t work between modules. It needs .globl to be exported.

I don’t think = is generating N_INDR. In any case I have source compat, just not weak indirect aliases.

This version works between modules.

#define weak_alias(old, new) \
        __asm__(".globl _" #new); \
        __asm__("_" #new " = _" #old); \
        extern __typeof(old) new __attribute__((weak_import))

Michael Clark via llvm-dev

unread,
Jun 14, 2017, 5:47:37 PM6/14/17
to Louis Gerbarg, llvm-dev

On 15 Jun 2017, at 6:50 AM, Louis Gerbarg <lger...@apple.com> wrote:


On Jun 6, 2017, at 4:08 PM, Michael Clark via llvm-dev <llvm...@lists.llvm.org> wrote:

Hi Folks,

I’m working on a port of musl libc to macos (arch triple is “x86_64-xnu-musl”) to solve some irreconcilable issues I’m having with libSystem.dylib. I don’t want to use glibc for various reasons, mainly because I want to static link. I have static PIE + ASLR working which is not actually supported by the Apple toolchain (*1), but I managed to get it to work. I’m sure Apple might say “Don’t do that”, but from looking at the history of the xnu kernel ABI, it seems to be very stable between versions.

I am from Apple, and I will say “Don’t do that.” The kernel ABI for our platforms is not stable, we only guarantee stability at the dynamic link boundary (in this case public symbols exported from libSystem). While the kernel syscall numbers have not changed (though the kernel team reserves the right to do that), the parameter lists and argument marshaling for them certainly has changed. We also do not support static executables on our system.

We even had bincompat issues related to this i rolled their during the last major release (macOS 10.12 Sierra): Go implemented its own syscall support, which caused all of their binaries that used gettimeofday since the internal interface changed <https://github.com/golang/go/issues/16606>. More broadly you can look at a discussion of their issues here: <https://github.com/golang/go/issues/16606>. In their case they want to avoid invoking an external linker like ld64 or lld as opposed to avoiding libSystem, but the effect is the same, they shipped a tool that caused unsuspecting developers to have surprise bincompat issues that were entirely avoidable.

I’m aware that Go has had some issues with the XNU ABI boundary.

I’m working on a CPU simulator / binary translator and I need control of the process address space layout. It seems I may ultimately need to use Hypervisor.framework however that is a lot more work in the short term.

The issue I am having with libSystem.dylib is the lack of weak linkage (versus weak_import) i.e. weak aliases. I don’t want to use a wrapper binary with DYLD_INSERT_LIBRARIES. I want to interpose Libc symbols with some of the symbols present in my binary (memory allocator, mmap). Interposition support is somewhat lacking in the Mach-O toolchain and runtime linker despite the Mach-O format technically supporting what I need (N_INDR and N_WEAK_DEF).


If I could use N_INDR and N_WEAK_DEF to have early bound (runtime link time) interposition with symbols in my binary replacing the C library allocator and mmap, and have libSystem use my implementations then I would be happy. libSystem itself would need to use weak aliases. This is possible with C libraries on other platforms.

I’ve tried relentlessly to intercept the malloc_zone implementation. malloc_zone_register is not sufficient as some of the internal zones are tied to the internals of Libc and I am getting heap collisions with Libc allocated objects and my guest address space. On Linux I have enough control to do what I need and can interpose my symbols to implement versions of libc functions that I wish to override. The problem on darwin is that I am not able to interpose the malloc implementation until main starts, and at that point it is tool late as the C library already has created its internal zones. I’m also unable to interpose mmap. I have already looked at the interpose symbol tricks but they don’t meet my purposes  (not wanting to re-exec with DYLD_INSERT_LIBRARIES). Weak aliases from libSystem to the allocator implementation and various public symbols along with N_INDR and N_WEAK_DEF would be required for me to achieve what I need to achieve (somewhat similarly to the elegant internal implementation of musl libc).

With my current solution (musl on xnu) I have successfully reserved 0x1000 – 0x7fff_0000_0000. Essentially the low 128TiB minus 4GiB at the top of the address space where I place my translator and translator stack. This is satisfactory for my user mode simulator to emulate Linux processes on macOS.

I think Hypervisor.framework is probably the correct interface to be using if I want to avoid the kernel ABI, however that is a lot more work that making syscall wrappers and I would need to implement communication from VM process to the host process.

I’m actually implementing Linux syscall emulation in a user simulator so the kernel ABI is probably the technically correct layer. The full system emulator ultimately needs to use Hypervisor.framework if I am to use hardware paging instead of soft MMU. I have two simulators, a user-mode sim that emulates the Linux ABI and a full system emulator: https://rv8.io/ and I really want to support RISC-V Linux on macOS in the user mode simulator.

Proper Linux ABI emulation on macOS would ultimately require kernel support, at minimum something like binfmt misc, but ideally a kext that implements another ABI personality (much like Linux ABI emulation on Windows) in addition to the BSD personality. In fact the FreeBSD linux compat could be used if the FreeBSD portion of XNU is synced up with current, and we’d get bug fixes for long standing issues like the macOS TCP_NOPUSH bug that has long since been fixed in FreeBSD.

Ultimately if you are doing this on your own for your own for fun thats great, but if this something you intend to ship to other people please reconsider. It is more than a theoretical concern that it will break.

Louis

_______________________________________________
LLVM Developers mailing list
llvm...@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

Michael Clark via llvm-dev

unread,
Jun 14, 2017, 8:51:19 PM6/14/17
to Louis Gerbarg, llvm-dev

On 15 Jun 2017, at 11:35 AM, Louis Gerbarg <lger...@apple.com> wrote:


On Jun 14, 2017, at 2:47 PM, Michael Clark via llvm-dev <llvm...@lists.llvm.org> wrote:


On 15 Jun 2017, at 6:50 AM, Louis Gerbarg <lger...@apple.com> wrote:


On Jun 6, 2017, at 4:08 PM, Michael Clark via llvm-dev <llvm...@lists.llvm.org> wrote:

Hi Folks,

I’m working on a port of musl libc to macos (arch triple is “x86_64-xnu-musl”) to solve some irreconcilable issues I’m having with libSystem.dylib. I don’t want to use glibc for various reasons, mainly because I want to static link. I have static PIE + ASLR working which is not actually supported by the Apple toolchain (*1), but I managed to get it to work. I’m sure Apple might say “Don’t do that”, but from looking at the history of the xnu kernel ABI, it seems to be very stable between versions.

I am from Apple, and I will say “Don’t do that.” The kernel ABI for our platforms is not stable, we only guarantee stability at the dynamic link boundary (in this case public symbols exported from libSystem). While the kernel syscall numbers have not changed (though the kernel team reserves the right to do that), the parameter lists and argument marshaling for them certainly has changed. We also do not support static executables on our system.

We even had bincompat issues related to this i rolled their during the last major release (macOS 10.12 Sierra): Go implemented its own syscall support, which caused all of their binaries that used gettimeofday since the internal interface changed <https://github.com/golang/go/issues/16606>. More broadly you can look at a discussion of their issues here: <https://github.com/golang/go/issues/16606>. In their case they want to avoid invoking an external linker like ld64 or lld as opposed to avoiding libSystem, but the effect is the same, they shipped a tool that caused unsuspecting developers to have surprise bincompat issues that were entirely avoidable.

I’m aware that Go has had some issues with the XNU ABI boundary.

I’m working on a CPU simulator / binary translator and I need control of the process address space layout. It seems I may ultimately need to use Hypervisor.framework however that is a lot more work in the short term.

I actually thought about mentioning Hypervisor.framework, but I was not sure about your use case. If you need full control of the address space (PAGE_ZERO control, overriding the shared cache mappings, etc) that is really the only supported mechanism.


The issue I am having with libSystem.dylib is the lack of weak linkage (versus weak_import) i.e. weak aliases. I don’t want to use a wrapper binary with DYLD_INSERT_LIBRARIES. I want to interpose Libc symbols with some of the symbols present in my binary (memory allocator, mmap). Interposition support is somewhat lacking in the Mach-O toolchain and runtime linker despite the Mach-O format technically supporting what I need (N_INDR and N_WEAK_DEF).


Dyld does not generally use nlists at runtime except for things like dladdr(), and has not for the last 10 years or so. Instead dyld uses a trie to publish exports, and and a small byte code language to describe binding imports. We still support using nlists for old binaries, but anything built with recent tools also contains the newer trie and bind op codes which will be used if they are available. I do not think our tries can express the sort of import semantics you want.

I see two ways of potentially doing it (short of hypervisor.framework). Both of them are a bit gross and have some bincompat risks, but given you are an open source project and can rev if need be that may not be an enormous issue:

You could specify a custom segment in your executable with zero file size and a vm size the blocks out the address range you need and then unmap it. There may be practical limits on it that prevent you from achieving what you want.

That’s an interesting approach worth exploring. I wasn’t sure I could unmap the zero page at runtime.

I’m currently using -Wl,-pagezero_size,0x1000 which frees up the lower address space but of course the Libc allocator starts using this address space; the default on x86_64 with ld64 is a 4GiB zero page and that is where Libc normally allocates. I believe Libc is passing NULL for the address hint to mmap or vm_allocate and is getting the default address returned by the kernel, so it allocates at the lowest address possible. I’m likely going to have a similar problem if I unmap a region after start up and then call a Libc function that allocates using one of the internal zones. Even when I replaced the default malloc zone, it seems there were already other zones created and appeared to be used internally by Libc.

In fact, during this process, I’ve been working on minimising my Libc footprint, which is obviously required if I want to run in Hypervisor.framework. The early CRT initialisation hooks for C++ and image relocation stuff will be required, however I may well end up with a tiny stub that loads an ELF image if I do in fact use Hypervisor.framework. I’m using C++ so I can use vector, map, string, shared_ptr, etc. I find that using vector, map, string, shared_ptr are much safer than traditional C and raw pointers.

You could use implicit interposing. This is a feature add so that ASAN binaries can avoid the the whole re-exec with DYLD_INSERT_LIBRARIES issue. It is not guaranteed to be stable, but in practice it is probably the most stable option short of using a hypervisor. The way it works:

I saw the ASAN interposition patch that avoids DYLD_INSERT_LIBRARIES however I was not sure how it worked.

Define all the symbols in a dylib along with an interpose section (as though you were going to load it with DYLD_INSERT_LIBRARIES). Directly link your executable to binary. Dyld will discover the interpose during dependency analysis (before libSystem initializers are run) and apply the interpose. This has only been tested in the case of our sanitizer runtimes, but it *SHOULD* work.

This might be an approach worth exploring, as I can then interpose vm_allocate and/or mmap to add an address hint to coax Libc into using a reserved area of memory. I actually tried to get this to work but my interposed functions were not called as they were in the main executable. e.g.


So I guess I need a dependency on an additional dylib which has my interposed functions. It’s a pity dyld only searches dylibs and not the main executable for interpose sections (as it didn’t appear to work with an interpose section in the main executable).

If I could use N_INDR and N_WEAK_DEF to have early bound (runtime link time) interposition with symbols in my binary replacing the C library allocator and mmap, and have libSystem use my implementations then I would be happy. libSystem itself would need to use weak aliases. This is possible with C libraries on other platforms.

I’ve tried relentlessly to intercept the malloc_zone implementation. malloc_zone_register is not sufficient as some of the internal zones are tied to the internals of Libc and I am getting heap collisions with Libc allocated objects and my guest address space. On Linux I have enough control to do what I need and can interpose my symbols to implement versions of libc functions that I wish to override. The problem on darwin is that I am not able to interpose the malloc implementation until main starts, and at that point it is tool late as the C library already has created its internal zones. I’m also unable to interpose mmap. I have already looked at the interpose symbol tricks but they don’t meet my purposes  (not wanting to re-exec with DYLD_INSERT_LIBRARIES). Weak aliases from libSystem to the allocator implementation and various public symbols along with N_INDR and N_WEAK_DEF would be required for me to achieve what I need to achieve (somewhat similarly to the elegant internal implementation of musl libc).

With my current solution (musl on xnu) I have successfully reserved 0x1000 – 0x7fff_0000_0000. Essentially the low 128TiB minus 4GiB at the top of the address space where I place my translator and translator stack. This is satisfactory for my user mode simulator to emulate Linux processes on macOS.

I think Hypervisor.framework is probably the correct interface to be using if I want to avoid the kernel ABI, however that is a lot more work that making syscall wrappers and I would need to implement communication from VM process to the host process.

I think that this is probably your best choice from a binary compatibility standpoint in the long run. It is a lot of work, though I am not sure if it is really that much more work than trying to port a new libc or maintain a custom toolchain.

Yes, both of them are quite a bit of work. I need to get early boot code to switch the CPU into long mode and implement a virtual device to communicate with the host process, i.e. for console IO. Of course I need a thread implementation and a bunch of other things. It’s also quite a lot of work.

Thanks,
Michael.

Louis

Louis Gerbarg via llvm-dev

unread,
Jun 15, 2017, 1:54:10 AM6/15/17
to Michael Clark, llvm-dev
On Jun 6, 2017, at 4:08 PM, Michael Clark via llvm-dev <llvm...@lists.llvm.org> wrote:

Hi Folks,

I’m working on a port of musl libc to macos (arch triple is “x86_64-xnu-musl”) to solve some irreconcilable issues I’m having with libSystem.dylib. I don’t want to use glibc for various reasons, mainly because I want to static link. I have static PIE + ASLR working which is not actually supported by the Apple toolchain (*1), but I managed to get it to work. I’m sure Apple might say “Don’t do that”, but from looking at the history of the xnu kernel ABI, it seems to be very stable between versions.
I am from Apple, and I will say “Don’t do that.” The kernel ABI for our platforms is not stable, we only guarantee stability at the dynamic link boundary (in this case public symbols exported from libSystem). While the kernel syscall numbers have not changed (though the kernel team reserves the right to do that), the parameter lists and argument marshaling for them certainly has changed. We also do not support static executables on our system.

We even had bincompat issues related to this i rolled their during the last major release (macOS 10.12 Sierra): Go implemented its own syscall support, which caused all of their binaries that used gettimeofday since the internal interface changed <https://github.com/golang/go/issues/16606>. More broadly you can look at a discussion of their issues here: <https://github.com/golang/go/issues/16606>. In their case they want to avoid invoking an external linker like ld64 or lld as opposed to avoiding libSystem, but the effect is the same, they shipped a tool that caused unsuspecting developers to have surprise bincompat issues that were entirely avoidable.

Ultimately if you are doing this on your own for your own for fun thats great, but if this something you intend to ship to other people please reconsider. It is more than a theoretical concern that it will break.

Louis

Louis Gerbarg via llvm-dev

unread,
Jun 15, 2017, 1:55:16 AM6/15/17
to Michael Clark, llvm-dev
On Jun 14, 2017, at 2:47 PM, Michael Clark via llvm-dev <llvm...@lists.llvm.org> wrote:


On 15 Jun 2017, at 6:50 AM, Louis Gerbarg <lger...@apple.com> wrote:


On Jun 6, 2017, at 4:08 PM, Michael Clark via llvm-dev <llvm...@lists.llvm.org> wrote:

Hi Folks,

I’m working on a port of musl libc to macos (arch triple is “x86_64-xnu-musl”) to solve some irreconcilable issues I’m having with libSystem.dylib. I don’t want to use glibc for various reasons, mainly because I want to static link. I have static PIE + ASLR working which is not actually supported by the Apple toolchain (*1), but I managed to get it to work. I’m sure Apple might say “Don’t do that”, but from looking at the history of the xnu kernel ABI, it seems to be very stable between versions.

I am from Apple, and I will say “Don’t do that.” The kernel ABI for our platforms is not stable, we only guarantee stability at the dynamic link boundary (in this case public symbols exported from libSystem). While the kernel syscall numbers have not changed (though the kernel team reserves the right to do that), the parameter lists and argument marshaling for them certainly has changed. We also do not support static executables on our system.

We even had bincompat issues related to this i rolled their during the last major release (macOS 10.12 Sierra): Go implemented its own syscall support, which caused all of their binaries that used gettimeofday since the internal interface changed <https://github.com/golang/go/issues/16606>. More broadly you can look at a discussion of their issues here: <https://github.com/golang/go/issues/16606>. In their case they want to avoid invoking an external linker like ld64 or lld as opposed to avoiding libSystem, but the effect is the same, they shipped a tool that caused unsuspecting developers to have surprise bincompat issues that were entirely avoidable.

I’m aware that Go has had some issues with the XNU ABI boundary.

I’m working on a CPU simulator / binary translator and I need control of the process address space layout. It seems I may ultimately need to use Hypervisor.framework however that is a lot more work in the short term.

I actually thought about mentioning Hypervisor.framework, but I was not sure about your use case. If you need full control of the address space (PAGE_ZERO control, overriding the shared cache mappings, etc) that is really the only supported mechanism.

The issue I am having with libSystem.dylib is the lack of weak linkage (versus weak_import) i.e. weak aliases. I don’t want to use a wrapper binary with DYLD_INSERT_LIBRARIES. I want to interpose Libc symbols with some of the symbols present in my binary (memory allocator, mmap). Interposition support is somewhat lacking in the Mach-O toolchain and runtime linker despite the Mach-O format technically supporting what I need (N_INDR and N_WEAK_DEF).


Dyld does not generally use nlists at runtime except for things like dladdr(), and has not for the last 10 years or so. Instead dyld uses a trie to publish exports, and and a small byte code language to describe binding imports. We still support using nlists for old binaries, but anything built with recent tools also contains the newer trie and bind op codes which will be used if they are available. I do not think our tries can express the sort of import semantics you want.

I see two ways of potentially doing it (short of hypervisor.framework). Both of them are a bit gross and have some bincompat risks, but given you are an open source project and can rev if need be that may not be an enormous issue:

You could specify a custom segment in your executable with zero file size and a vm size the blocks out the address range you need and then unmap it. There may be practical limits on it that prevent you from achieving what you want.

You could use implicit interposing. This is a feature add so that ASAN binaries can avoid the the whole re-exec with DYLD_INSERT_LIBRARIES issue. It is not guaranteed to be stable, but in practice it is probably the most stable option short of using a hypervisor. The way it works:

Define all the symbols in a dylib along with an interpose section (as though you were going to load it with DYLD_INSERT_LIBRARIES). Directly link your executable to binary. Dyld will discover the interpose during dependency analysis (before libSystem initializers are run) and apply the interpose. This has only been tested in the case of our sanitizer runtimes, but it *SHOULD* work.

If I could use N_INDR and N_WEAK_DEF to have early bound (runtime link time) interposition with symbols in my binary replacing the C library allocator and mmap, and have libSystem use my implementations then I would be happy. libSystem itself would need to use weak aliases. This is possible with C libraries on other platforms.

I’ve tried relentlessly to intercept the malloc_zone implementation. malloc_zone_register is not sufficient as some of the internal zones are tied to the internals of Libc and I am getting heap collisions with Libc allocated objects and my guest address space. On Linux I have enough control to do what I need and can interpose my symbols to implement versions of libc functions that I wish to override. The problem on darwin is that I am not able to interpose the malloc implementation until main starts, and at that point it is tool late as the C library already has created its internal zones. I’m also unable to interpose mmap. I have already looked at the interpose symbol tricks but they don’t meet my purposes  (not wanting to re-exec with DYLD_INSERT_LIBRARIES). Weak aliases from libSystem to the allocator implementation and various public symbols along with N_INDR and N_WEAK_DEF would be required for me to achieve what I need to achieve (somewhat similarly to the elegant internal implementation of musl libc).

With my current solution (musl on xnu) I have successfully reserved 0x1000 – 0x7fff_0000_0000. Essentially the low 128TiB minus 4GiB at the top of the address space where I place my translator and translator stack. This is satisfactory for my user mode simulator to emulate Linux processes on macOS.

I think Hypervisor.framework is probably the correct interface to be using if I want to avoid the kernel ABI, however that is a lot more work that making syscall wrappers and I would need to implement communication from VM process to the host process.
I think that this is probably your best choice from a binary compatibility standpoint in the long run. It is a lot of work, though I am not sure if it is really that much more work than trying to port a new libc or maintain a custom toolchain.

Louis

Michael Clark via llvm-dev

unread,
Jun 16, 2017, 6:13:22 AM6/16/17
to Louis Gerbarg, llvm-dev
Hi Louis,

I tried unmapping the 4GiB zero page at runtime however the vm_map is owned by the kernel and it appears it can’t be unmapped after process creation. This left the option of linking with -page_zero 0x1000 and somehow interposing the allocator.

I then tried the second approach of using implicit interposing and have come up with a workable solution where I set the zero page size to 0x1000 and then interpose mmap, vm_map, vm_allocate, mach_vm_map and mach_vm_allocate. Interestingly the dylib needs to be linked with an explicit image_base otherwise dyld will load the interposing dylib into the low address space, using the default map address. After quite a bit of experimenting I came up with what is a relatively clean solution, and it doesn’t require any knowledge of the internals of libsystem_malloc.dylib other than how it uses the default address hints (for entropy). I’ve tested with and without PIE however I need to link with -no-pie due to bugs in ld64 when altering both the zero page size and the image base.


In the test results, you can see mmap-test-vanilla running with vanilla libSystem.dylib. The test shows heap allocation addresses (from libsystem_malloc.dylib) along with addresses from explicit calls to mmap and vm_allocate with and without address hints (mapping 0x1000 fails of course). In the first case the lower 4GiB is not available.

In mmap-test-collision I have linked the program with -page-zero 0x1000 which frees up the address space, but you can see ibsystem_malloc.dylib heap allocations colliding/intermingling with the lower 4GiB of address space which I wish to reserve for the CPU simulator; my original problem. I have already tried with runtime registration of malloc zones however this didn’t work due to internal zones and early allocations in the CRT. That is the point I was at previously (heap collisions).

In the third test mmap-test-himem you can see the same program run with the interposed memory allocation functions. I’m able to reserve 0x1000 - 0x7ffe00000000 which should be satisfactory for emulating a Linux user process address space on macOS (as I’m loading ELFs).

Thanks for the hints; otherwise I might have been using musl-xnu, although I’m still keen to get that working for a variety of other reasons, but not for this project. This approach solves my immediate problem and I don’t need to build a sysroot of musl-xnu, libcxx, libcxxabi and libunwind to get my project to build on macOS. I actually already had my project working with musl-xnu, including stdio, signals, setjmp/longjmp, mach_time, static PIE, ASLR and a lot of other functionality, but its kind of a large and complicated build dependency. I still need to implement threads and semaphores however I was making rapid progress…

Cheers,
Michael.

$ make test
cc -Wall -Wpedantic -c -o obj/mmap-himem.o mmap-himem.c
cc -dynamiclib \
-install_name @rpath/mmap-himem.dylib \
-image_base 0x7ffe80000000 \
-o lib/mmap-himem.dylib obj/mmap-himem.o
cc -Wall -Wpedantic -c -o obj/mmap-test.o mmap-test.c
cc -o bin/mmap-test-vanilla obj/mmap-test.o
cc -Wl,-no_pie -Wl,-pagezero_size,0x1000 \
-image_base 0x7ffe00000000 \
-o bin/mmap-test-collision obj/mmap-test.o
cc -Wl,-no_pie -Wl,-pagezero_size,0x1000 \
-image_base 0x7ffe00000000 \
-rpath lib lib/mmap-himem.dylib \
-o bin/mmap-test-himem obj/mmap-test.o

== mmap-test-vanilla ==
stak=0x7fff5053c7c8
text=0x10f6c4040
heap=0x10f6f9000
map1=0xffffffffffffffff
map2=0x7ff000000000
map3=0x10f71a000
map4=0x10f71b000

== mmap-test-collision ==
stak=0x7fff5fbff7c8
text=0x7ffe00001040
heap=0x34000
map1=0x1000
map2=0x7ff000000000
map3=0x55000
map4=0x56000

== mmap-test-himem ==
stak=0x7fff5fbff7d0
text=0x7ffe00001040
heap=0x7fff01002000
map1=0x1000
map2=0x7ff000000000
map3=0x7fff01026000
map4=0x7fff01028000

Michael.

Jean-Daniel via llvm-dev

unread,
Jun 17, 2017, 12:46:55 PM6/17/17
to Louis Gerbarg, llvm-dev
Le 15 juin 2017 à 01:35, Louis Gerbarg via llvm-dev <llvm...@lists.llvm.org> a écrit :


On Jun 14, 2017, at 2:47 PM, Michael Clark via llvm-dev <llvm...@lists.llvm.org> wrote:


On 15 Jun 2017, at 6:50 AM, Louis Gerbarg <lger...@apple.com> wrote:


On Jun 6, 2017, at 4:08 PM, Michael Clark via llvm-dev <llvm...@lists.llvm.org> wrote:

Hi Folks,

I’m working on a port of musl libc to macos (arch triple is “x86_64-xnu-musl”) to solve some irreconcilable issues I’m having with libSystem.dylib. I don’t want to use glibc for various reasons, mainly because I want to static link. I have static PIE + ASLR working which is not actually supported by the Apple toolchain (*1), but I managed to get it to work. I’m sure Apple might say “Don’t do that”, but from looking at the history of the xnu kernel ABI, it seems to be very stable between versions.

I am from Apple, and I will say “Don’t do that.” The kernel ABI for our platforms is not stable, we only guarantee stability at the dynamic link boundary (in this case public symbols exported from libSystem). While the kernel syscall numbers have not changed (though the kernel team reserves the right to do that), the parameter lists and argument marshaling for them certainly has changed. We also do not support static executables on our system.

We even had bincompat issues related to this i rolled their during the last major release (macOS 10.12 Sierra): Go implemented its own syscall support, which caused all of their binaries that used gettimeofday since the internal interface changed <https://github.com/golang/go/issues/16606>. More broadly you can look at a discussion of their issues here: <https://github.com/golang/go/issues/16606>. In their case they want to avoid invoking an external linker like ld64 or lld as opposed to avoiding libSystem, but the effect is the same, they shipped a tool that caused unsuspecting developers to have surprise bincompat issues that were entirely avoidable.

I’m aware that Go has had some issues with the XNU ABI boundary.

I’m working on a CPU simulator / binary translator and I need control of the process address space layout. It seems I may ultimately need to use Hypervisor.framework however that is a lot more work in the short term.

I actually thought about mentioning Hypervisor.framework, but I was not sure about your use case. If you need full control of the address space (PAGE_ZERO control, overriding the shared cache mappings, etc) that is really the only supported mechanism.

The issue I am having with libSystem.dylib is the lack of weak linkage (versus weak_import) i.e. weak aliases. I don’t want to use a wrapper binary with DYLD_INSERT_LIBRARIES. I want to interpose Libc symbols with some of the symbols present in my binary (memory allocator, mmap). Interposition support is somewhat lacking in the Mach-O toolchain and runtime linker despite the Mach-O format technically supporting what I need (N_INDR and N_WEAK_DEF).


Dyld does not generally use nlists at runtime except for things like dladdr(), and has not for the last 10 years or so. Instead dyld uses a trie to publish exports, and and a small byte code language to describe binding imports. We still support using nlists for old binaries, but anything built with recent tools also contains the newer trie and bind op codes which will be used if they are available. I do not think our tries can express the sort of import semantics you want.

I see two ways of potentially doing it (short of hypervisor.framework). Both of them are a bit gross and have some bincompat risks, but given you are an open source project and can rev if need be that may not be an enormous issue:

You could specify a custom segment in your executable with zero file size and a vm size the blocks out the address range you need and then unmap it. There may be practical limits on it that prevent you from achieving what you want.

You could use implicit interposing. This is a feature add so that ASAN binaries can avoid the the whole re-exec with DYLD_INSERT_LIBRARIES issue. It is not guaranteed to be stable, but in practice it is probably the most stable option short of using a hypervisor. The way it works:

Was is really added for ASAN ? I remember using dyld_interpose for a long time. It was already described in the Amit Singh edition of Mac OS X Internal book (which is older than clang). Or we are not talking about the same interpose ?

Louis Gerbarg via llvm-dev

unread,
Jun 22, 2017, 3:44:21 PM6/22/17
to Jean-Daniel, llvm-dev
On Jun 17, 2017, at 9:46 AM, Jean-Daniel <mai...@xenonium.com> wrote:


Le 15 juin 2017 à 01:35, Louis Gerbarg via llvm-dev <llvm...@lists.llvm.org> a écrit :


On Jun 14, 2017, at 2:47 PM, Michael Clark via llvm-dev <llvm...@lists.llvm.org> wrote:


On 15 Jun 2017, at 6:50 AM, Louis Gerbarg <lger...@apple.com> wrote:


On Jun 6, 2017, at 4:08 PM, Michael Clark via llvm-dev <llvm...@lists.llvm.org> wrote:

Hi Folks,

I’m working on a port of musl libc to macos (arch triple is “x86_64-xnu-musl”) to solve some irreconcilable issues I’m having with libSystem.dylib. I don’t want to use glibc for various reasons, mainly because I want to static link. I have static PIE + ASLR working which is not actually supported by the Apple toolchain (*1), but I managed to get it to work. I’m sure Apple might say “Don’t do that”, but from looking at the history of the xnu kernel ABI, it seems to be very stable between versions.

I am from Apple, and I will say “Don’t do that.” The kernel ABI for our platforms is not stable, we only guarantee stability at the dynamic link boundary (in this case public symbols exported from libSystem). While the kernel syscall numbers have not changed (though the kernel team reserves the right to do that), the parameter lists and argument marshaling for them certainly has changed. We also do not support static executables on our system.

We even had bincompat issues related to this i rolled their during the last major release (macOS 10.12 Sierra): Go implemented its own syscall support, which caused all of their binaries that used gettimeofday since the internal interface changed <https://github.com/golang/go/issues/16606>. More broadly you can look at a discussion of their issues here: <https://github.com/golang/go/issues/16606>. In their case they want to avoid invoking an external linker like ld64 or lld as opposed to avoiding libSystem, but the effect is the same, they shipped a tool that caused unsuspecting developers to have surprise bincompat issues that were entirely avoidable.

I’m aware that Go has had some issues with the XNU ABI boundary.

I’m working on a CPU simulator / binary translator and I need control of the process address space layout. It seems I may ultimately need to use Hypervisor.framework however that is a lot more work in the short term.

I actually thought about mentioning Hypervisor.framework, but I was not sure about your use case. If you need full control of the address space (PAGE_ZERO control, overriding the shared cache mappings, etc) that is really the only supported mechanism.

The issue I am having with libSystem.dylib is the lack of weak linkage (versus weak_import) i.e. weak aliases. I don’t want to use a wrapper binary with DYLD_INSERT_LIBRARIES. I want to interpose Libc symbols with some of the symbols present in my binary (memory allocator, mmap). Interposition support is somewhat lacking in the Mach-O toolchain and runtime linker despite the Mach-O format technically supporting what I need (N_INDR and N_WEAK_DEF).


Dyld does not generally use nlists at runtime except for things like dladdr(), and has not for the last 10 years or so. Instead dyld uses a trie to publish exports, and and a small byte code language to describe binding imports. We still support using nlists for old binaries, but anything built with recent tools also contains the newer trie and bind op codes which will be used if they are available. I do not think our tries can express the sort of import semantics you want.

I see two ways of potentially doing it (short of hypervisor.framework). Both of them are a bit gross and have some bincompat risks, but given you are an open source project and can rev if need be that may not be an enormous issue:

You could specify a custom segment in your executable with zero file size and a vm size the blocks out the address range you need and then unmap it. There may be practical limits on it that prevent you from achieving what you want.

You could use implicit interposing. This is a feature add so that ASAN binaries can avoid the the whole re-exec with DYLD_INSERT_LIBRARIES issue. It is not guaranteed to be stable, but in practice it is probably the most stable option short of using a hypervisor. The way it works:

Was is really added for ASAN ? I remember using dyld_interpose for a long time. It was already described in the Amit Singh edition of Mac OS X Internal book (which is older than clang). Or we are not talking about the same interpose ?

Interposing has been around for a long time, but interposes would only be enabled by explicitly interposing the library at launch via DYLD_INSERT_LIBRARIES. In the case of ASAN where the compiled binaries are ABI dependent on the ASAN runtime support (and thus cannot really run without the interposed memory allocators) it doesn’t make sense to use the environment to control interposing. Instead the interposes are included in the ASAN runtime dylib (which the applications built for ASAN link to), dyld discovers them during dependency analysis and uses them. This implicit discovery mechanism was added for ASAN, though the underlying mechanism for specifying interposed functions is the same.

Louis
Reply all
Reply to author
Forward
0 new messages