je_malloc_usable_size not found

[aarlt]

I'm running:

• Which Parity version?: @ 3c27587
• Which operating system?: MacOS 10.13.6
• How installed?: from source

I having an issue building parity-clib. I get the following linker error, because of undefined symbols:

...
          Undefined symbols for architecture x86_64:
            "_je_malloc_usable_size", referenced from:
                heapsize::heap_size_of_impl::hb1048887d365f818 in libheapsize-3daab403b63f3097.rlib(heapsize-3daab403b63f3097.heapsize3.rcgu.o)
          ld: symbol(s) not found for architecture x86_64
          clang: error: linker command failed with exit code 1 (use -v to see invocation)

error: aborting due to previous error

error: Could not compile `parity-clib`.

To learn more, run the command again with --verbose.

I used cargo build -p parity-clib to build parity-clib.

[niklasad1]

Hey,

I think this is an issue with the heapsize crate crate because it assumes je_malloc exists and in debug builds clang disables je_malloc!

For further description see: servo/heapsize#80

I don't have any Mac machine to confirm this but may just try building it in release-mode:

$ cargo build -p parity-clib --release

@tomaka any thoughts on this? I guess this will also impact iOS?

[tomaka]

Shouldn't this also happen when compiling the parity binary?

[aarlt]

@niklasad1 unfortunately cargo build -p parity-clib --release is also not working.

@tomaka a normal parity build with cargo build --release --features final works - no linking problems.

[niklasad1]

I think this only applies dynamically linked libraries i.e, with rust terms dylib or cdylib. If I have understood this correctly heapsize assumes that symbol je_malloc_usable_size exist in the binary when it is statically linked it does but when it dynamically linked it might not!

[aarlt]

@niklasad1 yes, thats true. the static library is buildable, but im not very sure that this is usable - the symbol should be missing there too, am I right?

[aarlt]

Hmm.. after looking into https://github.com/servo/heapsize/blob/master/src/lib.rs it seems that there is an dependency to the jemalloc library. After just running brew install jemalloc to install that library the problem still persists. After checking the exported symbols of that lib, it looks like that the symbol _je_malloc_usable_size doesn't exist there, but there is a symbol named _malloc_usable_size.

[aarlt]

I found a workaround for this, but I'm not sure whether this problem will also exist on ios / android. Those are the first rust lines I wrote :D

However, I added the following function to parity-ethereum/parity-clib/src/lib.rs - this will just be a proxy function that will call the correct function exported by jemalloc on macos. If the code is correct I would be also able to create a pull request.

 #[no_mangle]
 #[cfg(target_os = "macos")]
 pub extern fn je_malloc_usable_size(_ptr: *const c_void) -> usize {
     unsafe {
       extern "C" {
         fn malloc_usable_size(ptr: *const c_void) -> usize;
       }
       return malloc_usable_size(_ptr);
     }
 }

I was able to compile & run the example with g++ main.cpp libparity.a -framework Security -framework IOKit -framework CoreFoundation -ljemalloc. I'm not sure whether the linking against jemalloc need to be manually added into your buildsystem.

[tomaka]

This is an extremely hazardous work-around. The code can break every time we upgrade Rust. Also it would need to be enabled only in debug if I understand correctly.

[aarlt]

@tomaka yeah that may be true. I guess thats really an issue within https://github.com/servo/heapsize. My understanding - I'm not a rust developer - https://github.com/servo/heapsize/blob/master/src/lib.rs#L44 seem to change the symbol name for macos, ios and android to be what is not correct - at least for macos. The symbol name need to be malloc_usable_size - as it would normally be from my understanding - but it will be changed to be je_malloc_usable_size.

What I don't understand: why is the stuff working for the normal parity-ethereum build. We should somehow see the same issues.

The code can break every time we upgrade Rust. Also it would need to be enabled only in debug if I understand correctly.

I don't understand why this may break every time if Rust is upgraded - its just an issue with the exported symbols of jemalloc. From my point of view the redefinition of the symbol name to je_malloc_usable_size is the real issue. I also don't get why this may only be needed in debug.

[aarlt]

I just created an issue here: servo/heapsize#97

[tomaka]

Unfortunately I think they have been having a pretty long discussion about this already, and the conclusion is that heapsize is wrong by design and that something else should be used.

[aarlt]

Ah ok.. sounds not good..

However, I think it might be a good idea to add cargo build -p parity-clib --release to your CI. I guess youre also doing there ios and android builds - so at least there seem to be an easy way to find out whether this is not only an macos issue.

[aarlt]

@tomaka please see servo/heapsize#97 (comment)

[aarlt]

@tomaka @niklasad1 I still don't really understand why there is no linking issue with the normal parity build - any ideas?

[aarlt]

@tomaka @niklasad1 I was able to understand what is happening here, but due to lack of rust & parity-ethereum codebase knowledge I will probably not yet be able to create fix for this issue.

@tomaka you where right, my work-around is extremely hazardous.

I just read https://doc.rust-lang.org/1.9.0/book/custom-allocators.html. rust is shipped with two default allocators - alloc_system and alloc_jemalloc. alloc_jemalloc seem not to work nicely on all systems, thats why it is sometimes just not used. But basically, if alloc_jemalloc is available on a specific platform, it will be used as a default allocator for binaries. In contrast to that, if a static- or dynamic library is build - it will use alloc_system. As I understood alloc_system will just use the standard allocation API like malloc and free.

That means at least for macos it looks like that alloc_jemalloc is available, thats why the default parity binary can be build. As described, for libraries it will use the alloc_system, where of course jemalloc will not be linked against. So my work-arround will basically mix different allocators - that is for sure quite dangerous.

I think the best solution is probably really to create a custom crate for heapsize, where somehow https://github.com/servo/heapsize/blob/master/src/lib.rs#L44 will need to be changed in a way that it will not redefine the symbol name - if compiled as a library.

[tomaka]

With Rust 1.28, adding je_malloc_usable_size ourselves could actually be a solution if we use the system allocator that is now stable.

[gguoss]

parity version: 4fec2f2
rustc 1.33.0-nightly
MacOS
Build parity error.

[gguoss]

Build Ok , this issue should close.