Add Libdl.LazyLibrary

This provides an in-base mechanism to handle chained library
dependencies.  In essence, the `LazyLibrary` object can be used anywhere
a pointer to a library can be used (`dlopen`, `dlsym`, `ccall`, etc...)
but it delays loading the library (and its recursive dependencies) until
it is actually needed.

This is the foundational piece needed to upgrade JLLs to lazily-load
their libraries.  In this new scheme, JLLs would generally lose all
executable code and consist of nothing more than `LazyLibrary`
definitions.

Author: staticfloat

Makefile

@@ -84,6 +84,12 @@ julia-base: julia-deps $(build_sysconfdir)/julia/startup.jl $(build_man1dir)/jul
 julia-libccalltest: julia-deps
 	@$(MAKE) $(QUIET_MAKE) -C $(BUILDROOT)/src libccalltest
 
+julia-libccalllazyfoo: julia-deps
+	@$(MAKE) $(QUIET_MAKE) -C $(BUILDROOT)/src libccalllazyfoo
+
+julia-libccalllazybar: julia-deps julia-libccalllazyfoo
+	@$(MAKE) $(QUIET_MAKE) -C $(BUILDROOT)/src libccalllazybar
+
 julia-libllvmcalltest: julia-deps
 	@$(MAKE) $(QUIET_MAKE) -C $(BUILDROOT)/src libllvmcalltest
 
@@ -102,7 +108,8 @@ julia-sysimg-bc : julia-stdlib julia-base julia-cli-$(JULIA_BUILD_MODE) julia-sr
 julia-sysimg-release julia-sysimg-debug : julia-sysimg-% : julia-sysimg-ji julia-src-%
 	@$(MAKE) $(QUIET_MAKE) -C $(BUILDROOT) -f sysimage.mk sysimg-$*
 
-julia-debug julia-release : julia-% : julia-sysimg-% julia-src-% julia-symlink julia-libccalltest julia-libllvmcalltest julia-base-cache
+julia-debug julia-release : julia-% : julia-sysimg-% julia-src-% julia-symlink julia-libccalltest \
+                                      julia-libccalllazyfoo julia-libccalllazybar julia-libllvmcalltest julia-base-cache
 
 stdlibs-cache-release stdlibs-cache-debug : stdlibs-cache-% : julia-%
 	@$(MAKE) $(QUIET_MAKE) -C $(BUILDROOT) -f pkgimage.mk all-$*
@@ -189,7 +196,7 @@ JL_TARGETS := julia-debug
 endif
 
 # private libraries, that are installed in $(prefix)/lib/julia
-JL_PRIVATE_LIBS-0 := libccalltest libllvmcalltest
+JL_PRIVATE_LIBS-0 := libccalltest libccalllazyfoo libccalllazybar libllvmcalltest
 ifeq ($(JULIA_BUILD_MODE),release)
 JL_PRIVATE_LIBS-0 += libjulia-internal libjulia-codegen
 else ifeq ($(JULIA_BUILD_MODE),debug)

NEWS.md

@@ -27,6 +27,8 @@ Compiler/Runtime improvements
 
 * The `@pure` macro is now deprecated. Use `Base.@assume_effects :foldable` instead ([#48682]).
 * The mark phase of the Garbage Collector is now multi-threaded ([#48600]).
+* A new `LazyLibrary` type is exported from `Libdl` for use in building chained lazy library
+  loads, primarily to be used within JLLs ([#50074]).
 
 Command-line option changes
 ---------------------------

base/Base.jl

@@ -314,6 +314,12 @@ include("missing.jl")
 # version
 include("version.jl")
 
+# Concurrency (part 1)
+include("linked_list.jl")
+include("condition.jl")
+include("threads.jl")
+include("lock.jl")
+
 # system & environment
 include("sysinfo.jl")
 include("libc.jl")
@@ -328,11 +334,9 @@ const liblapack_name = libblas_name
 include("logging.jl")
 using .CoreLogging
 
-# Concurrency
-include("linked_list.jl")
-include("condition.jl")
-include("threads.jl")
-include("lock.jl")
+# Concurrency (part 2)
+# Note that `atomics.jl` here should be deprecated
+Core.eval(Threads, :(include("atomics.jl")))
 include("channels.jl")
 include("partr.jl")
 include("task.jl")

base/libdl.jl

@@ -9,7 +9,7 @@ import Base.DL_LOAD_PATH
 
 export DL_LOAD_PATH, RTLD_DEEPBIND, RTLD_FIRST, RTLD_GLOBAL, RTLD_LAZY, RTLD_LOCAL,
     RTLD_NODELETE, RTLD_NOLOAD, RTLD_NOW, dlclose, dlopen, dlopen_e, dlsym, dlsym_e,
-    dlpath, find_library, dlext, dllist
+    dlpath, find_library, dlext, dllist, LazyLibrary, LazyStringFunc
 
 """
     DL_LOAD_PATH
@@ -45,6 +45,9 @@ applicable.
 """
 (RTLD_DEEPBIND, RTLD_FIRST, RTLD_GLOBAL, RTLD_LAZY, RTLD_LOCAL, RTLD_NODELETE, RTLD_NOLOAD, RTLD_NOW)
 
+# The default flags for `dlopen()`
+const default_rtld_flags = RTLD_LAZY | RTLD_DEEPBIND
+
 """
     dlsym(handle, sym; throw_error::Bool = true)
 
@@ -72,8 +75,8 @@ end
 Look up a symbol from a shared library handle, silently return `C_NULL` on lookup failure.
 This method is now deprecated in favor of `dlsym(handle, sym; throw_error=false)`.
 """
-function dlsym_e(hnd::Ptr, s::Union{Symbol,AbstractString})
-    return something(dlsym(hnd, s; throw_error=false), C_NULL)
+function dlsym_e(args...)
+    return something(dlsym(args...; throw_error=false), C_NULL)
 end
 
 """
@@ -110,10 +113,10 @@ If the library cannot be found, this method throws an error, unless the keyword
 """
 function dlopen end
 
-dlopen(s::Symbol, flags::Integer = RTLD_LAZY | RTLD_DEEPBIND; kwargs...) =
+dlopen(s::Symbol, flags::Integer = default_rtld_flags; kwargs...) =
     dlopen(string(s), flags; kwargs...)
 
-function dlopen(s::AbstractString, flags::Integer = RTLD_LAZY | RTLD_DEEPBIND; throw_error::Bool = true)
+function dlopen(s::AbstractString, flags::Integer = default_rtld_flags; throw_error::Bool = true)
     ret = ccall(:jl_load_dynamic_library, Ptr{Cvoid}, (Cstring,UInt32,Cint), s, flags, Cint(throw_error))
     if ret == C_NULL
         return nothing
@@ -314,4 +317,98 @@ function dllist()
     return dynamic_libraries
 end
 
+
+const LazyLibraryLock = Base.ReentrantLock()
+"""
+    LazyLibrary(name, flags = <default dlopen flags>,
+                dependencies = LazyLibrary[], on_load_callback = nothing)
+
+Represents a lazily-loaded library that opens itself and its dependencies on first usage
+in a `dlopen()`, `dlsym()`, or `ccall()` usage.  While this structure contains the
+ability to run arbitrary code on first load via `on_load_callback`, we caution that this
+should be used sparingly, as it is not expected that `ccall()` should result in large
+amounts of Julia code being run.  You may call `ccall()` from within the
+`on_load_callback` but only for the current library and its dependencies.
+"""
+mutable struct LazyLibrary
+    # Name and flags to open with
+    const name::AbstractString
+    const flags::UInt32
+
+    # Dependencies that must be loaded before we can load
+    dependencies::Vector{LazyLibrary}
+
+    # Function that gets called once upon initial load with the pointer as an argument
+    on_load_callback::Union{Nothing,Function}
+
+    # Pointer that we eventually fill out upon first `dlopen()`
+    @atomic handle::Ptr{Cvoid}
+    function LazyLibrary(name; flags = default_rtld_flags, dependencies = LazyLibrary[],
+                         on_load_callback = nothing)
+        return new(
+            name,
+            UInt32(flags),
+            collect(dependencies),
+            on_load_callback,
+            C_NULL,
+        )
+    end
+end
+
+# Register this LazyLibrary type with the C code during bootstrap,
+# so that `ccall()` lowering knows what a `LazyLibrary` is, and how
+# to `dlopen()` it.  This hack brought to you by Valentin. \[0_o]/
+Base.unsafe_store!(cglobal(:jl_lazy_library_type, Any), LazyLibrary)
+Base.unsafe_store!(cglobal(:jl_libdl_dlopen_func, Any), dlopen)
+
+"""
+    LazyStringFunc
+
+Helper type for `LazyString` to be used with `LazyLibrary` so that runtime-generated
+library paths can be used with lazy libraries.  Note that the value will only be computed
+once; it is cached after initial computation.  Typical usage:
+
+```
+libfoo = LazyLibrary(LazyStringFunc(() -> joinpath(artifact"foo/lib/libfoo.so.1.2.3")))
+```
+"""
+LazyStringFunc(f) = LazyString(_LazyStringFunc(f))
+
+struct _LazyStringFunc <: AbstractString
+    f::Function
+end
+Base.print(io::IO, lsf::_LazyStringFunc) = print(io, lsf.f())
+
+function dlopen(ll::LazyLibrary, flags::Integer = ll.flags; kwargs...)
+    handle = @atomic :acquire ll.handle
+    if handle != C_NULL
+        return handle
+    end
+
+    # Ensure that all dependencies are loaded
+    for dep in ll.dependencies
+        dlopen(dep; kwargs...)
+    end
+
+    # Only let a single thread run callbacks at a time
+    @lock LazyLibraryLock begin
+        # Check to see if another thread has already run this
+        handle = @atomic :acquire ll.handle
+        if handle == C_NULL
+            # Load our library
+            handle = dlopen(ll.name, flags; kwargs...)
+
+            # Invoke our on load callback, if it exists
+            if ll.on_load_callback !== nothing
+                ll.on_load_callback(handle)
+            end
+
+            @atomic :release ll.handle = handle
+        end
+    end
+
+    return handle
+end
+dlsym(ll::LazyLibrary, args...; kwargs...) = dlsym(dlopen(ll), args...; kwargs...)
+dlpath(ll::LazyLibrary) = dlpath(dlopen(ll))
 end # module Libdl

base/threads.jl

@@ -8,7 +8,6 @@ module Threads
 global Condition # we'll define this later, make sure we don't import Base.Condition
 
 include("threadingconstructs.jl")
-include("atomics.jl")
 include("locks-mt.jl")
 
 end

doc/src/devdocs/locks.md

@@ -91,6 +91,8 @@ may result in pernicious and hard-to-find deadlocks. BE VERY CAREFUL!
 >
 >     > this may continue to be held after releasing the iolock, or acquired without it,
 >     > but be very careful to never attempt to acquire the iolock while holding it
+>
+>   * Libdl.LazyLibrary lock
 
 
 The following is the root lock, meaning no other lock shall be held when trying to acquire it:

src/Makefile

@@ -249,6 +249,8 @@ $(build_includedir)/julia/uv/*.h: $(LIBUV_INC)/uv/*.h | $(build_includedir)/juli
 	$(INSTALL_F) $^ $(build_includedir)/julia/uv
 
 libccalltest: $(build_shlibdir)/libccalltest.$(SHLIB_EXT)
+libccalllazyfoo: $(build_shlibdir)/libccalllazyfoo.$(SHLIB_EXT)
+libccalllazybar: $(build_shlibdir)/libccalllazybar.$(SHLIB_EXT)
 libllvmcalltest: $(build_shlibdir)/libllvmcalltest.$(SHLIB_EXT)
 
 ifeq ($(OS), Linux)
@@ -273,6 +275,12 @@ endif
 	mv [email protected] $@
 	$(INSTALL_NAME_CMD)libccalltest.$(SHLIB_EXT) $@
 
+$(build_shlibdir)/libccalllazyfoo.$(SHLIB_EXT): $(SRCDIR)/ccalllazyfoo.c
+	@$(call PRINT_CC, $(CC) $(JCFLAGS) $(JL_CFLAGS) $(JCPPFLAGS) $(FLAGS) -O3 $< $(fPIC) -shared -o $@ $(LDFLAGS) $(COMMON_LIBPATHS) $(call SONAME_FLAGS,ccalllazyfoo.$(SHLIB_EXT)))
+
+$(build_shlibdir)/libccalllazybar.$(SHLIB_EXT): $(SRCDIR)/ccalllazybar.c $(build_shlibdir)/libccalllazyfoo.$(SHLIB_EXT)
+	@$(call PRINT_CC, $(CC) $(JCFLAGS) $(JL_CFLAGS) $(JCPPFLAGS) $(FLAGS) -O3 $< $(fPIC) -shared -o $@ $(LDFLAGS) $(COMMON_LIBPATHS) $(call SONAME_FLAGS,ccalllazybar.$(SHLIB_EXT)) -lccalllazyfoo)
+
 $(build_shlibdir)/libllvmcalltest.$(SHLIB_EXT): $(SRCDIR)/llvmcalltest.cpp $(LLVM_CONFIG_ABSOLUTE)
 	@$(call PRINT_CC, $(CXX) $(LLVM_CXXFLAGS) $(FLAGS) $(CPPFLAGS) $(CXXFLAGS) -O3 $< $(fPIC) -shared -o $@ $(LDFLAGS) $(COMMON_LIBPATHS) $(NO_WHOLE_ARCHIVE) $(CG_LLVMLINK)) -lpthread
 

src/ccall.cpp

@@ -651,6 +651,9 @@ static void interpret_symbol_arg(jl_codectx_t &ctx, native_sym_arg_t &out, jl_va
                 f_lib = jl_symbol_name((jl_sym_t*)t1);
             else if (jl_is_string(t1))
                 f_lib = jl_string_data(t1);
+            else if (jl_lazy_library_type != NULL && jl_isa(t1, (jl_value_t *)jl_lazy_library_type)) {
+                out.lib_expr = t1;
+            }
             else
                 f_name = NULL;
         }

src/ccalllazybar.c

@@ -0,0 +1,10 @@
+// This file is a part of Julia. License is MIT: https://julialang.org/license
+
+#include "ccalltest_common.h"
+
+// We expect this to come from `libccalllazyfoo`
+extern int foo(int);
+
+DLLEXPORT int bar(int a) {
+    return foo(a + 1);
+}

src/ccalllazyfoo.c

@@ -0,0 +1,7 @@
+// This file is a part of Julia. License is MIT: https://julialang.org/license
+
+#include "ccalltest_common.h"
+
+DLLEXPORT int foo(int a) {
+    return a*2;
+}