Force loading ZLUDA through LD_PRELOAD (#447)

Certain applications (pytorch) decide that it's a great idea to distribute whole CUDA driver and link to it with DT_RPATH. This igores LD_LIBRARY_PATH.
This code defeats that evil mechanism through any means necessary

Cargo.lock

@@ -300,6 +300,22 @@ dependencies = [
  "cfg-if",
 ]
 
+[[package]]
+name = "ctor"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ec09e802f5081de6157da9a75701d6c713d8dc3ba52571fd4bd25f412644e8a6"
+dependencies = [
+ "ctor-proc-macro",
+ "dtor 0.0.6",
+]
+
+[[package]]
+name = "ctor-proc-macro"
+version = "0.0.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e2931af7e13dc045d8e9d26afccc6fa115d64e115c9c84b1166288b46f6782c2"
+
 [[package]]
 name = "cuda_macros"
 version = "0.0.0"
@@ -406,7 +422,16 @@ version = "0.0.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "97cbdf2ad6846025e8e25df05171abfb30e3ababa12ee0a0e44b9bbe570633a8"
 dependencies = [
- "dtor-proc-macro",
+ "dtor-proc-macro 0.0.5",
+]
+
+[[package]]
+name = "dtor"
+version = "0.0.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bbc66182e62c4e716e2d70f97beceea0de798923f8ca48fb82aa3134dc3cae12"
+dependencies = [
+ "dtor-proc-macro 0.0.6",
 ]
 
 [[package]]
@@ -415,6 +440,12 @@ version = "0.0.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7454e41ff9012c00d53cf7f475c5e3afa3b91b7c90568495495e8d9bf47a1055"
 
+[[package]]
+name = "dtor-proc-macro"
+version = "0.0.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f678cf4a922c215c63e0de95eb1ff08a958a81d47e485cf9da1e27bf6305cfa5"
+
 [[package]]
 name = "dynasm"
 version = "1.2.3"
@@ -1781,7 +1812,7 @@ dependencies = [
  "cuda_macros",
  "cuda_types",
  "dark_api",
- "dtor",
+ "dtor 0.0.7",
  "hip_runtime-sys",
  "lazy_static",
  "libc",
@@ -1874,6 +1905,14 @@ dependencies = [
  "cuda_types",
 ]
 
+[[package]]
+name = "zluda_preload"
+version = "0.0.0"
+dependencies = [
+ "ctor",
+ "unwrap_or",
+]
+
 [[package]]
 name = "zluda_redirect"
 version = "0.0.0"
@@ -1905,6 +1944,7 @@ dependencies = [
  "format",
  "goblin",
  "libc",
+ "libloading",
  "parking_lot",
  "paste",
  "ptx",
@@ -1956,6 +1996,7 @@ dependencies = [
  "cuda_types",
  "dark_api",
  "format",
+ "libc",
  "libloading",
 ]
 

Cargo.toml

@@ -33,6 +33,7 @@ members = [
     "zluda_fft",
     "zluda_inject",
     "zluda_ml",
+    "zluda_preload",
     "zluda_redirect",
     "zluda_sparse",
 ]

xtask/src/main.rs

@@ -155,6 +155,8 @@ struct Metadata {
 #[derive(Deserialize)]
 #[serde(deny_unknown_fields)]
 struct ZludaMetadata {
+    #[serde(default)]
+    linux_only: bool,
     #[serde(default)]
     windows_only: bool,
     #[serde(default)]
@@ -192,6 +194,9 @@ fn compile(b: Build) -> (PathBuf, String, Vec<Project>) {
         .into_iter()
         .filter_map(Project::try_new)
         .filter(|project| {
+            if project.meta.linux_only && cfg!(windows) {
+                return false;
+            }
             if project.meta.windows_only && cfg!(not(windows)) {
                 return false;
             }

zluda/Cargo.toml

@@ -22,7 +22,6 @@ lz4-sys = "1.9"
 tempfile = "3"
 paste = "1.0"
 rustc-hash = "1.1"
-dtor = "0.0.6"
 zluda_common = { path = "../zluda_common" }
 
 [target.'cfg(windows)'.dependencies]
@@ -30,7 +29,7 @@ winapi = { version = "0.3", features = ["heapapi", "std"] }
 
 [target.'cfg(not(windows))'.dependencies]
 libc = "0.2"
-dtor = "0.0.6"
+dtor = "0.0.7"
 
 [package.metadata.zluda]
 linux_symlinks = [

zluda/src/impl/memory.rs

@@ -1,5 +1,4 @@
 use hip_runtime_sys::*;
-use std::mem;
 
 pub(crate) fn alloc_v2(dptr: *mut hipDeviceptr_t, bytesize: usize) -> hipError_t {
     unsafe { hipMalloc(dptr.cast(), bytesize) }?;
@@ -36,7 +35,7 @@ pub(crate) fn get_address_range_v2(
 }
 
 pub(crate) fn set_d32_v2(dst: hipDeviceptr_t, ui: ::core::ffi::c_uint, n: usize) -> hipError_t {
-    unsafe { hipMemsetD32(dst, mem::transmute(ui), n) }
+    unsafe { hipMemsetD32(dst, ui as std::ffi::c_int, n) }
 }
 
 pub(crate) fn set_d8_v2(dst: hipDeviceptr_t, value: ::core::ffi::c_uchar, n: usize) -> hipError_t {

zluda_preload/Cargo.toml

@@ -0,0 +1,20 @@
+[package]
+name = "zluda_preload"
+version = "0.0.0"
+authors = ["Andrzej Janik <vosen@vosen.pl>"]
+edition = "2021"
+
+[lib]
+crate-type = ["cdylib"]
+
+[dependencies]
+ctor = "0.4.3"
+unwrap_or = "1.0.1"
+
+[package.metadata.zluda]
+linux_only = true
+linux_symlinks = [
+    "zluda_preload",
+    "trace/zluda_preload",
+    "trace_nvidia/zluda_preload",
+]

zluda_preload/README.md

@@ -0,0 +1,21 @@
+This crate is a last resort Linux-specific solution.
+Most of the time we can inject ourselves into a process by having users
+set `LD_LIBRARY_PATH`.
+Unfortunately, there is software out there which dynamically links to CUDA and
+CUDA performance libraries using RPATH. On Linux, dynamic linker operates
+using approximately this algorithm:
+* If path contains `/` treat the name as a (possibly relative) path and just use it
+* Otherwise return the first that succeeds:
+  * Library with this name already loaded into the process
+  * Try paths in `DT_RPATH` (if `DT_RUNPATH` is not present)
+  * Try paths in `LD_LIBRARY_PATH`
+  * Try paths in `DT_RUNPATH`
+  * Try system paths
+
+In order to defeat `DT_RPATH` this library needs to be preloaded with `LD_PRELOAD`.
+On initialization we also preload all the performance libraries. We also hijack
+`dlopen` and on every call to `dlopen` that tries to open a CUDA library we
+redirect it to our libraries
+
+We also expose `zluda_dlopen_noredirect` for the purpose of tracing libraries
+so they can load real underlying library and not just get redirected to themselves

zluda_preload/src/lib.rs

@@ -0,0 +1,150 @@
+use std::{
+    ffi::{c_char, c_int, c_void, CStr},
+    mem,
+    path::PathBuf,
+    ptr::{self, NonNull},
+    sync::LazyLock,
+};
+use unwrap_or::unwrap_some_or;
+
+// Definition takes from `libc` crate:
+// https://github.com/rust-lang/libc/blob/cf82fdf3f22ccfa98ba120efc50d5f39ab2d52ff/src/unix/linux_like/linux/mod.rs#L2682
+const RTLD_NEXT: *mut c_void = -1isize as _;
+
+unsafe extern "C" {
+    fn dlsym(handle: *mut c_void, symbol: *const c_char) -> *mut c_void;
+    fn dladdr(addr: *const c_void, info: *mut DLInfo) -> c_int;
+}
+
+#[repr(C)]
+struct DLInfo {
+    dli_fname: *const c_char,
+    dli_fbase: *mut c_void,
+    dli_sname: *const c_char,
+    dli_saddr: *mut c_void,
+}
+
+static FILES_FOR_REDIRECT: [&'static str; 14] = [
+    "libcublas.so",
+    "libcublas.so.12",
+    "libcublasLt.so",
+    "libcublasLt.so.12",
+    "libcuda.so",
+    "libcuda.so.1",
+    "libcudnn.so",
+    "libcudnn.so.9",
+    "libcufft.so",
+    "libcufft.so.11",
+    "libcusparse.so",
+    "libcusparse.so.12",
+    "libnvidia-ml.so",
+    "libnvidia-ml.so.1",
+];
+
+// Global state, caching some computations that would be otherwise repeated on every `dlopen`
+struct GlobalState {
+    /// The original `dlopen` implementation from libdl.  
+    dlopen_next: Option<unsafe extern "C" fn(*const c_char, c_int) -> DlopenResult>,
+    /// The full paths of the file names from `FILES_FOR_REDIRECT` that will be used for redirection
+    replacement_paths: Option<[Vec<u8>; FILES_FOR_REDIRECT.len()]>,
+}
+
+static GLOBAL_STATE: LazyLock<GlobalState> = LazyLock::new(|| {
+    let dlopen_next = unsafe { mem::transmute(dlsym(RTLD_NEXT, c"dlopen".as_ptr())) };
+    let mut self_dlinfo = unsafe { mem::zeroed::<DLInfo>() };
+    let replacement_paths = if unsafe { dladdr(dlopen as _, &mut self_dlinfo) } != 0 {
+        unsafe { CStr::from_ptr(self_dlinfo.dli_fname) }
+            .to_str()
+            .ok()
+            .and_then(|path| {
+                let mut pathbuf = PathBuf::from(path);
+                if !pathbuf.pop() {
+                    return None;
+                }
+                Some(FILES_FOR_REDIRECT.map(|file| {
+                    let mut buffer = pathbuf.join(file).into_os_string().into_encoded_bytes();
+                    buffer.push(0);
+                    buffer
+                }))
+            })
+    } else {
+        None
+    };
+    GlobalState {
+        dlopen_next,
+        replacement_paths,
+    }
+});
+
+pub const RTLD_GLOBAL: c_int = 0x100;
+pub const RTLD_LAZY: c_int = 1;
+
+#[ctor::ctor]
+unsafe fn ctor() {
+    let GlobalState {
+        dlopen_next,
+        replacement_paths,
+    } = &*GLOBAL_STATE;
+    let dlopen_next = unwrap_some_or!(dlopen_next, return);
+    let replacement_paths = unwrap_some_or!(replacement_paths, return);
+    // We preload the paths to the files we want to redirect, because
+    // * We don't control dynamic linking when loading dependencies. We hijack
+    //   dlopen, but that only works if the dependency has been explicitly
+    //   loaded with dlopen. It does not intercept the loading of the dependencies
+    // * The first step that dynamic linker does is check if the file is already
+    //   loaded
+    for replacement in replacement_paths.into_iter() {
+        dlopen_next(replacement.as_ptr().cast(), RTLD_GLOBAL | RTLD_LAZY).ok();
+    }
+}
+
+type DlopenResult = Result<NonNull<c_void>, ()>;
+
+const _: fn() = || {
+    let _ = std::mem::transmute::<*mut c_void, DlopenResult>;
+};
+
+#[no_mangle]
+unsafe extern "C" fn dlopen(filename: *const c_char, flags: c_int) -> DlopenResult {
+    let GlobalState {
+        dlopen_next,
+        replacement_paths,
+    } = &*GLOBAL_STATE;
+    let dlopen_next = dlopen_next.ok_or(())?;
+    dlopen_redirect(dlopen_next, replacement_paths, filename, flags)
+        .or_else(|| dlopen_next(filename, flags).ok())
+        .ok_or(())
+}
+
+#[no_mangle]
+unsafe extern "C" fn zluda_dlopen_noredirect(
+    filename: *const c_char,
+    flags: c_int,
+) -> DlopenResult {
+    let dlopen_next = GLOBAL_STATE.dlopen_next.ok_or(())?;
+    dlopen_next(filename, flags)
+}
+
+unsafe fn dlopen_redirect<'a>(
+    dlopen_next: unsafe extern "C" fn(*const c_char, c_int) -> DlopenResult,
+    replacement_paths: &'a Option<[Vec<u8>; FILES_FOR_REDIRECT.len()]>,
+    input_path: *const c_char,
+    flags: c_int,
+) -> Option<NonNull<c_void>> {
+    if input_path == ptr::null() {
+        return None;
+    }
+    let input_path = CStr::from_ptr(input_path).to_str().ok()?;
+    let replacement_paths = replacement_paths.as_ref()?;
+    let replacement_path = FILES_FOR_REDIRECT
+        .into_iter()
+        .zip(replacement_paths.into_iter())
+        .find_map(|(file, path)| {
+            if input_path.ends_with(file) {
+                Some(path)
+            } else {
+                None
+            }
+        })?;
+    unsafe { dlopen_next(replacement_path.as_ptr() as _, flags) }.ok()
+}

zluda_trace/Cargo.toml

@@ -27,6 +27,7 @@ rustc-hash = "1.1.0"
 cglue = "0.3.5"
 zstd-safe = { version = "7.2.4", features = ["std"] }
 unwrap_or = "1.0.1"
+libloading = "0.8"
 
 [target.'cfg(windows)'.dependencies]
 winapi = { version = "0.3", features = ["libloaderapi", "debugapi", "std"] }

zluda_trace/src/lib.rs

@@ -56,8 +56,8 @@ pub(crate) struct CudaDynamicFns {
 
 impl CudaDynamicFns {
     pub(crate) unsafe fn load_library(path: &str) -> Option<Self> {
-        let lib_handle = NonNull::new(os::load_library(path));
-        lib_handle.map(|lib_handle| CudaDynamicFns {
+        let lib_handle = os::dlopen_local_noredirect(path).ok()?;
+        Some(CudaDynamicFns {
             lib_handle,
             fn_table: CudaFnTable::default(),
         })

zluda_trace/src/os_unix.rs

@@ -1,15 +1,17 @@
 use cuda_types::cuda::CUuuid;
-use std::ffi::{c_void, CStr, CString};
+use std::borrow::Cow;
+use std::ffi::{c_void, CStr};
 use std::mem;
+use std::ptr::NonNull;
 
 pub(crate) const LIBCUDA_DEFAULT_PATH: &str = "/usr/lib/x86_64-linux-gnu/libcuda.so.1";
 
-pub unsafe fn load_library(libcuda_path: &str) -> *mut c_void {
-    let libcuda_path = CString::new(libcuda_path).unwrap();
-    libc::dlopen(
-        libcuda_path.as_ptr() as *const _,
-        libc::RTLD_LOCAL | libc::RTLD_NOW,
-    )
+pub fn dlopen_local_noredirect<'a>(
+    path: impl Into<Cow<'a, str>>,
+) -> Result<NonNull<c_void>, libloading::Error> {
+    let lib: libloading::os::unix::Library =
+        zluda_trace_common::dlopen_local_noredirect(path)?.into();
+    NonNull::new(lib.into_raw()).ok_or(libloading::Error::DlOpenUnknown)
 }
 
 pub unsafe fn get_proc_address(handle: *mut c_void, func: &CStr) -> *mut c_void {

zluda_trace/src/os_win.rs

@@ -1,22 +1,30 @@
+use cuda_types::cuda::CUuuid;
+use std::borrow::Cow;
+use std::os::windows::io::AsRawHandle;
+use std::ptr::NonNull;
 use std::{
     ffi::{c_void, CStr},
     mem, ptr,
     sync::LazyLock,
 };
-
-use std::os::windows::io::AsRawHandle;
 use winapi::{
     shared::minwindef::{FARPROC, HMODULE},
     um::debugapi::OutputDebugStringA,
     um::libloaderapi::{GetProcAddress, LoadLibraryW},
 };
 
-use cuda_types::cuda::CUuuid;
-
 pub(crate) const LIBCUDA_DEFAULT_PATH: &'static str = "C:\\Windows\\System32\\nvcuda.dll";
 const LOAD_LIBRARY_NO_REDIRECT: &'static [u8] = b"ZludaLoadLibraryW_NoRedirect\0";
 const GET_PROC_ADDRESS_NO_REDIRECT: &'static [u8] = b"ZludaGetProcAddress_NoRedirect\0";
 
+pub fn dlopen_local_noredirect<'a>(
+    path: impl Into<Cow<'a, str>>,
+) -> Result<NonNull<c_void>, libloading::Error> {
+    let lib: libloading::os::windows::Library =
+        zluda_trace_common::dlopen_local_noredirect(path)?.into();
+    NonNull::new(lib.into_raw() as *mut _).ok_or(libloading::Error::DlOpenUnknown)
+}
+
 static PLATFORM_LIBRARY: LazyLock<PlatformLibrary> =
     LazyLock::new(|| unsafe { PlatformLibrary::new() });
 
@@ -73,14 +81,6 @@ impl PlatformLibrary {
     }
 }
 
-pub unsafe fn load_library(libcuda_path: &str) -> *mut c_void {
-    let libcuda_path_uf16 = libcuda_path
-        .encode_utf16()
-        .chain(std::iter::once(0))
-        .collect::<Vec<_>>();
-    (PLATFORM_LIBRARY.LoadLibraryW)(libcuda_path_uf16.as_ptr()) as _
-}
-
 pub unsafe fn get_proc_address(handle: *mut c_void, func: &CStr) -> *mut c_void {
     (PLATFORM_LIBRARY.GetProcAddress)(handle as _, func.as_ptr() as _) as _
 }

zluda_trace_blas/src/lib.rs

@@ -7,7 +7,7 @@ fn get_library() -> Option<Library> {
     let cuda_lib = std::env::var("ZLUDA_BLAS_LIB")
         .ok()
         .unwrap_or_else(|| "/usr/local/cuda/lib64/libcublas.so".to_string());
-    unsafe { Library::new(cuda_lib) }.ok()
+    zluda_trace_common::dlopen_local_noredirect(cuda_lib).ok()
 }
 
 macro_rules! unimplemented {

zluda_trace_blaslt/src/lib.rs

@@ -7,7 +7,7 @@ fn get_library() -> Option<Library> {
     let cuda_lib = std::env::var("ZLUDA_BLASLT_LIB")
         .ok()
         .unwrap_or_else(|| "/usr/local/cuda/lib64/libcublasLt.so".to_string());
-    unsafe { Library::new(cuda_lib) }.ok()
+    zluda_trace_common::dlopen_local_noredirect(cuda_lib).ok()
 }
 
 macro_rules! unimplemented {

zluda_trace_common/Cargo.toml

@@ -11,3 +11,6 @@ cuda_types = { path = "../cuda_types" }
 dark_api = { path = "../dark_api" }
 format = { path = "../format" }
 cglue = "0.3.5"
+
+[target.'cfg(not(windows))'.dependencies]
+libc = "0.2"

zluda_trace_common/src/lib.rs

@@ -5,7 +5,7 @@ use cuda_types::{
     cusparse::cusparseStatus_tConsts,
 };
 use dark_api::ByteVecFfi;
-use std::{ffi::c_void, num::NonZero, ptr, sync::LazyLock};
+use std::{borrow::Cow, ffi::c_void, num::NonZero, ptr, sync::LazyLock};
 
 pub fn get_export_table() -> Option<::dark_api::zluda_trace::ZludaTraceInternal> {
     static CU_GET_EXPORT_TABLE: LazyLock<
@@ -38,30 +38,95 @@ fn open_driver() -> Result<libloading::Library, libloading::Error> {
     os::open_driver()
 }
 
+pub fn dlopen_local_noredirect<'a>(
+    path: impl Into<Cow<'a, str>>,
+) -> Result<libloading::Library, libloading::Error> {
+    unsafe { os::dlopen_local_noredirect(path.into()) }
+}
+
 #[cfg(unix)]
 pub(crate) mod os {
+    use libc::{c_char, c_int};
     use libloading::os;
-
-    const RTLD_NOLOAD: i32 = 0x4;
+    use std::{borrow::Cow, ffi::c_void, mem};
 
     pub fn open_driver() -> Result<libloading::Library, libloading::Error> {
         unsafe {
-            os::unix::Library::open(Some("libcuda.so.1"), RTLD_NOLOAD | os::unix::RTLD_LAZY)
+            os::unix::Library::open(
+                Some("libcuda.so.1"),
+                libc::RTLD_NOLOAD | os::unix::RTLD_LAZY,
+            )
             .or_else(|_| {
-                    os::unix::Library::open(Some("libcuda.so"), RTLD_NOLOAD | os::unix::RTLD_LAZY)
+                os::unix::Library::open(Some("libcuda.so"), libc::RTLD_NOLOAD | os::unix::RTLD_LAZY)
             })
             .map(Into::into)
         }
     }
+
+    pub unsafe fn dlopen_local_noredirect<'a>(
+        path: Cow<'a, str>,
+    ) -> Result<libloading::Library, libloading::Error> {
+        fn terminate_with_nul<'a>(path: Cow<'a, str>) -> Cow<'a, str> {
+            let path = if !path.ends_with('\0') {
+                let mut path = path.into_owned();
+                path.push('\0');
+                Cow::Owned(path)
+            } else {
+                path
+            };
+            path
+        }
+        let zluda_dlopen_noredirect =
+            unsafe { libc::dlsym(libc::RTLD_DEFAULT, c"zluda_dlopen_noredirect".as_ptr()) };
+        let zluda_dlopen_noredirect = mem::transmute::<
+            _,
+            Option<unsafe extern "C" fn(*const c_char, c_int) -> *mut c_void>,
+        >(zluda_dlopen_noredirect);
+        let dlopen = zluda_dlopen_noredirect.unwrap_or(libc::dlopen);
+        let path = terminate_with_nul(path);
+        Ok(libloading::os::unix::Library::from_raw(dlopen(
+            path.as_ptr().cast(),
+            os::unix::RTLD_LOCAL | os::unix::RTLD_LAZY,
+        ))
+        .into())
+    }
 }
 
 #[cfg(windows)]
 pub(crate) mod os {
     use libloading::os;
+    use std::borrow::Cow;
 
     pub fn open_driver() -> Result<libloading::Library, libloading::Error> {
         os::windows::Library::open_already_loaded("nvcuda").map(Into::into)
     }
+
+    pub unsafe fn dlopen_local_noredirect<'a>(
+        path: Cow<'a, str>,
+    ) -> Result<libloading::Library, libloading::Error> {
+        fn terminate_with_nul(mut path: Vec<u16>) -> Vec<u16> {
+            if path.last().copied() == Some(0) {
+                path.push(0);
+            }
+            path
+        }
+        let driver = open_driver()?;
+        match driver.get::<unsafe extern "C" fn(*const u16) -> isize>(
+            c"ZludaLoadLibraryW_NoRedirect".to_bytes_with_nul(),
+        ) {
+            Ok(load_library) => {
+                let symbol = load_library(
+                    terminate_with_nul(path.encode_utf16().collect::<Vec<u16>>()).as_ptr(),
+                );
+                if symbol == 0 {
+                    Err(libloading::Error::LoadLibraryExWUnknown)
+                } else {
+                    Ok(libloading::os::windows::Library::from_raw(symbol).into())
+                }
+            }
+            Err(_) => libloading::Library::new(&*path),
+        }
+    }
 }
 
 pub trait ReprUsize {

zluda_trace_dnn/src/lib.rs

@@ -7,7 +7,7 @@ fn get_library() -> Option<Library> {
     let cuda_lib = std::env::var("ZLUDA_DNN_LIB")
         .ok()
         .unwrap_or_else(|| "/usr/lib/x86_64-linux-gnu/libcudnn.so.9".to_string());
-    unsafe { Library::new(cuda_lib) }.ok()
+    zluda_trace_common::dlopen_local_noredirect(cuda_lib).ok()
 }
 
 macro_rules! unimplemented {

zluda_trace_fft/src/lib.rs

@@ -7,7 +7,7 @@ fn get_library() -> Option<Library> {
     let cuda_lib = std::env::var("ZLUDA_FFT_LIB")
         .ok()
         .unwrap_or_else(|| "/usr/local/cuda/lib64/libcufft.so".to_string());
-    unsafe { Library::new(cuda_lib) }.ok()
+    zluda_trace_common::dlopen_local_noredirect(cuda_lib).ok()
 }
 
 macro_rules! unimplemented {

zluda_trace_sparse/src/lib.rs

@@ -7,7 +7,7 @@ fn get_library() -> Option<Library> {
     let cuda_lib = std::env::var("ZLUDA_SPARSE_LIB")
         .ok()
         .unwrap_or_else(|| "/usr/local/cuda/lib64/libcusparse.so".to_string());
-    unsafe { Library::new(cuda_lib) }.ok()
+    zluda_trace_common::dlopen_local_noredirect(cuda_lib).ok()
 }
 
 macro_rules! unimplemented {