Add workaround for musl lack of SONAME support

Project.toml

@@ -4,6 +4,7 @@ authors = ["Mosè Giordano", "Elliot Saba"]
 version = "1.3.0"
 
 [deps]
+Libdl = "8f399da3-3557-5675-b5ff-fb832c97cbdb"
 Preferences = "21216c6a-2e73-6563-6e65-726566657250"
 
 [compat]

src/JLLWrappers.jl

@@ -6,6 +6,7 @@ end
 
 if VERSION >= v"1.6.0-DEV"
     using Preferences
+    using Base.BinaryPlatforms
 end
 
 # We need to glue expressions together a lot

src/products/library_generators.jl

@@ -51,7 +51,10 @@ macro init_library_product(product_name, product_path, dlopen_flags)
             # of `ccall` with its path/SONAME will find this path immediately.
             # dlopen_flags === nothing means to not dlopen the library.
             if $(dlopen_flags) !== nothing
-                global $(handle_name) = dlopen($(path_name), $(dlopen_flags))
+                global $(handle_name) = Base.invokelatest(
+                    JLLWrappers.musl_soname_workaround,
+                    dlopen($(path_name), $(dlopen_flags))
+                )
                 push!(LIBPATH_list, dirname($(path_name)))
             end
         end,

src/runtime.jl

@@ -92,3 +92,24 @@ function get_julia_libpaths()
     end
     return JULIA_LIBDIRS
 end
+
+@static if VERSION >= v"1.6.0" && libc(HostPlatform()) == "musl"
+    include("runtime_musl_workaround.jl")
+
+    """
+        musl_soname_workaround(lib_handle::Ptr{Cvoid})
+
+    Applies a workaround for musl's lack of SONAME loading by twiddling
+    some internal datastructures.  See `src/runtime_musl_workaround.jl`
+    for more details.  Returns the altered library handle.
+    """
+    musl_soname_workaround(lib_handle::Ptr{Cvoid}) = replace_musl_shortname(lib_handle)
+else
+    """
+        musl_soname_workaround(lib_handle::Ptr{Cvoid})
+
+    Not applicable on this platform.  Returns the altered library handle.
+    """
+    musl_soname_workaround(lib_handle::Ptr{Cvoid}) = lib_handle
+end
+

src/runtime_musl_workaround.jl

@@ -0,0 +1,148 @@
+## JLLWrappers musl SONAME workaround
+#
+# The problem is detailed in this thread [0], but in short:
+#
+# JLLs rely on a specific behavior of most `dlopen()` implementations; that if
+# a library with the same SONAME will not be loaded twice; e.g. if you first
+# load `/a/libfoo.so`, loading `/b/libbar.so` which declares a dependency on
+# `libfoo.so` will find the previously-loaded `libfoo.so` without needing to
+# search because the SONAME `libbar.so` looks for matches the SONAME of the
+# previously-loaded `libfoo.so`.  This allows JLLs to store libraries all over
+# the place, and directly `dlopen()` all dependencies before any dependents
+# would trigger a system-wide search.
+#
+# Musl does not do this.  They do have a mechanism for skipping the directory
+# search, but it is only invoked when loading a library without specifying
+# the full path [1].  This means that when checking for dependencies, musl
+# skips all libraries that were loaded by full path [2].  All that needs to
+# happen is that musl needs to record the `shortname` (e.g. SONAME) of all
+# libraries, but sadly there's no way to do that if we also want to specify
+# the library unambiguously [3,2].  Manipulating the environment to allow for
+# non-fully-specified searches to work (e.g. changing `LD_LIBRARY_PATH` then
+# invoking `dlopen("libfoo.so")`) won't work, as the environment is only read
+# at process initialization.  We are therefore backed into a corner and must
+# resort to heroic measures: manually inserting an appropriate `shortname`.
+#
+# [0] https://github.com/JuliaLang/julia/issues/40556
+# [1] https://github.com/ifduyue/musl/blob/aad50fcd791e009961621ddfbe3d4c245fd689a3/ldso/dynlink.c#L1163-L1164
+# [2] https://github.com/ifduyue/musl/blob/aad50fcd791e009961621ddfbe3d4c245fd689a3/ldso/dynlink.c#L1047-L1052
+# [3] https://github.com/ifduyue/musl/blob/aad50fcd791e009961621ddfbe3d4c245fd689a3/ldso/dynlink.c#L1043-L1044
+
+using Libdl
+
+# Use this to ensure the GC doesn't clean up values we insert into musl.
+manual_gc_roots = String[]
+
+## We define these structures so that Julia's internal struct padding logic
+## can do some arithmetic for us, instead of us needing to do manual offset
+## calculation ourselves, which is more error-prone.
+
+
+# This structure taken from `libc.h`
+# https://github.com/ifduyue/musl/blob/aad50fcd791e009961621ddfbe3d4c245fd689a3/src/internal/libc.h#L14-L18
+struct musl_tls_module
+    next::Ptr{musl_tls_module}
+    image::Ptr{musl_tls_module}
+    len::Csize_t
+    size::Csize_t
+    align::Csize_t
+    offset::Csize_t
+end
+
+# This structure taken from `ldso/dynlink.c`
+# https://github.com/ifduyue/musl/blob/aad50fcd791e009961621ddfbe3d4c245fd689a3/ldso/dynlink.c#L53-L107
+struct musl_dso
+    # Things we find mildly interesting
+    base::Ptr{Cvoid}
+    name::Ptr{UInt8}
+
+    # The wasteland of things we don't care about
+    dynv::Ptr{Csize_t}
+    next::Ptr{musl_dso}
+    prev::Ptr{musl_dso}
+
+    phdr::Ptr{Cvoid}
+    phnum::Cint
+    phentsize::Csize_t
+
+    syms::Ptr{Cvoid}
+    hashtab::Ptr{Cvoid}
+    ghashtab::Ptr{Cvoid}
+    versym::Ptr{Int16}
+    strings::Ptr{UInt8}
+    syms_next::Ptr{musl_dso}
+    lazy_next::Ptr{musl_dso}
+    lazy::Ptr{Csize_t}
+    lazy_cnt::Csize_t
+
+    map::Ptr{Cuchar}
+    map_len::Csize_t
+
+    # We assume that dev_t and ino_t are always `uint64_t`, even on 32-bit systems.
+    dev::UInt64
+    ino::UInt64
+    relocated::Cchar
+    constructed::Cchar
+    kernel_mapped::Cchar
+    mark::Cchar
+    bfs_built::Cchar
+    runtime_loaded::Cchar
+    # NOTE: struct layout rules should insert two bytes of space here
+    deps::Ptr{Ptr{musl_dso}}
+    needed_by::Ptr{musl_dso}
+    ndeps_direct::Csize_t
+    next_dep::Csize_t
+    ctor_visitor::Cint
+    rpath_orig::Ptr{UInt8}
+    rpath::Ptr{UInt8}
+
+    tls::musl_tls_module
+    tls_id::Csize_t
+    relro_start::Csize_t
+    relro_end::Csize_t
+    new_dtv::Ptr{Ptr{Cuint}}
+    new_tls::Ptr{UInt8}
+    td_index::Ptr{Cvoid}
+    fini_next::Ptr{musl_dso}
+
+    # Finally!  The field we're interested in!
+    shortname::Ptr{UInt8}
+
+    # We'll put this stuff at the end because it might be interesting to someone somewhere
+    loadmap::Ptr{Cvoid}
+    funcdesc::Ptr{Cvoid}
+    got::Ptr{Csize_t}
+end
+
+function replace_musl_shortname(lib_handle::Ptr{Cvoid})
+    # First, find the absolute path of the library we're talking about
+    lib_path = abspath(dlpath(lib_handle))
+
+    # Load the DSO object, which conveniently is the handle that `dlopen()`
+    # itself passes back to us.  Check to make sure it's what we expect, by
+    # inspecting the `name` field.  If it's not, something has gone wrong,
+    # and we should stop before touching anything else.
+    dso = unsafe_load(Ptr{musl_dso}(lib_handle))
+    dso_name = abspath(unsafe_string(dso.name))
+    if dso_name != lib_path
+        @debug("Unable to synchronize to DSO structure", name=dso_name, path=lib_path)
+        return lib_handle
+    end
+
+    # If the shortname is not NULL, break out.
+    if dso.shortname != C_NULL
+        @debug("shortname != NULL!", ptr=shortname_ptr, value=unsafe_string(shortname_ptr))
+        return lib_handle
+    end
+
+    # Calculate the offset of `shortname` from the base pointer of the DSO object
+    shortname_offset = fieldoffset(musl_dso, findfirst(==(:shortname), fieldnames(musl_dso)))
+
+    # Replace the shortname with the basename of lib_path.  Note that, in general, this
+    # should be the SONAME, but not always.  If we wanted to be pedantic, we should
+    # actually parse out the SONAME of this object.  But we don't want to be.
+    new_shortname = basename(lib_path)
+    push!(manual_gc_roots, new_shortname)
+    unsafe_store!(Ptr{Ptr{UInt8}}(lib_handle + shortname_offset), pointer(new_shortname))
+    return lib_handle
+end