Fix use-libc-auxv to use a weak dependency on getauxval. (#535)

* Fix `use-libc-auxv` to use a weak dependency on `getauxval`.

glibc <= 2.15 lacks `getauxval`, so avoid a strong dependency on it in
the `use-libc-auxv` implementation.

src/backend/linux_raw/mod.rs

@@ -14,6 +14,11 @@
 //! such as which pointers are array slices, out parameters, or in-out
 //! parameters, which integers are owned or borrowed file descriptors, etc.
 
+// Weak symbols used by the use-libc-auxv feature for glibc 2.15 support.
+#[cfg(feature = "use-libc-auxv")]
+#[macro_use]
+mod weak;
+
 #[macro_use]
 mod arch;
 mod conv;

src/backend/linux_raw/param/libc_auxv.rs

@@ -5,14 +5,15 @@
 //! This uses raw pointers to locate and read the kernel-provided auxv array.
 #![allow(unsafe_code)]
 
-#[cfg(any(feature = "param", feature = "runtime"))]
-use super::super::c;
 use super::super::elf::*;
 #[cfg(feature = "param")]
 use crate::ffi::CStr;
 #[cfg(feature = "runtime")]
 use core::slice;
 
+// `getauxval` wasn't supported in glibc until 2.16.
+weak!(fn getauxval(libc::c_ulong) -> *mut libc::c_void);
+
 #[cfg(feature = "param")]
 #[inline]
 pub(crate) fn page_size() -> usize {
@@ -22,35 +23,39 @@ pub(crate) fn page_size() -> usize {
 #[cfg(feature = "param")]
 #[inline]
 pub(crate) fn clock_ticks_per_second() -> u64 {
-    unsafe { libc::getauxval(libc::AT_CLKTCK) as u64 }
+    unsafe { libc::sysconf(libc::_SC_CLK_TCK) as u64 }
 }
 
 #[cfg(feature = "param")]
 #[inline]
 pub(crate) fn linux_hwcap() -> (usize, usize) {
-    unsafe {
-        (
-            libc::getauxval(libc::AT_HWCAP) as usize,
-            libc::getauxval(libc::AT_HWCAP2) as usize,
-        )
+    if let Some(libc_getauxval) = getauxval.get() {
+        unsafe {
+            let hwcap = libc_getauxval(libc::AT_HWCAP) as usize;
+            let hwcap2 = libc_getauxval(libc::AT_HWCAP2) as usize;
+            (hwcap, hwcap2)
+        }
+    } else {
+        (0, 0)
     }
 }
 
 #[cfg(feature = "param")]
 #[inline]
 pub(crate) fn linux_execfn() -> &'static CStr {
-    unsafe {
-        let execfn = libc::getauxval(libc::AT_EXECFN) as *const c::c_char;
-        CStr::from_ptr(execfn.cast())
+    if let Some(libc_getauxval) = getauxval.get() {
+        unsafe { CStr::from_ptr(libc_getauxval(libc::AT_EXECFN).cast()) }
+    } else {
+        cstr!("")
     }
 }
 
 #[cfg(feature = "runtime")]
 #[inline]
-pub(crate) fn exe_phdrs() -> (*const c::c_void, usize) {
+pub(crate) fn exe_phdrs() -> (*const libc::c_void, usize) {
     unsafe {
         (
-            libc::getauxval(libc::AT_PHDR) as *const c::c_void,
+            libc::getauxval(libc::AT_PHDR) as *const libc::c_void,
             libc::getauxval(libc::AT_PHNUM) as usize,
         )
     }
@@ -70,5 +75,9 @@ pub(in super::super) fn exe_phdrs_slice() -> &'static [Elf_Phdr] {
 /// so if we don't see it, this function returns a null pointer.
 #[inline]
 pub(in super::super) fn sysinfo_ehdr() -> *const Elf_Ehdr {
-    unsafe { libc::getauxval(linux_raw_sys::general::AT_SYSINFO_EHDR.into()) as *const Elf_Ehdr }
+    if let Some(libc_getauxval) = getauxval.get() {
+        unsafe { libc_getauxval(linux_raw_sys::general::AT_SYSINFO_EHDR.into()) as *const Elf_Ehdr }
+    } else {
+        core::ptr::null()
+    }
 }

src/backend/linux_raw/weak.rs

@@ -0,0 +1,228 @@
+// Implementation derived from `weak` in Rust's
+// library/std/src/sys/unix/weak.rs at revision
+// fd0cb0cdc21dd9c06025277d772108f8d42cb25f.
+
+#![allow(unsafe_code)]
+
+//! Support for "weak linkage" to symbols on Unix
+//!
+//! Some I/O operations we do in libstd require newer versions of OSes but we
+//! need to maintain binary compatibility with older releases for now. In order
+//! to use the new functionality when available we use this module for
+//! detection.
+//!
+//! One option to use here is weak linkage, but that is unfortunately only
+//! really workable on Linux. Hence, use dlsym to get the symbol value at
+//! runtime. This is also done for compatibility with older versions of glibc,
+//! and to avoid creating dependencies on `GLIBC_PRIVATE` symbols. It assumes
+//! that we've been dynamically linked to the library the symbol comes from,
+//! but that is currently always the case for things like libpthread/libc.
+//!
+//! A long time ago this used weak linkage for the `__pthread_get_minstack`
+//! symbol, but that caused Debian to detect an unnecessarily strict versioned
+//! dependency on libc6 (#23628).
+
+// There are a variety of `#[cfg]`s controlling which targets are involved in
+// each instance of `weak!` and `syscall!`. Rather than trying to unify all of
+// that, we'll just allow that some unix targets don't use this module at all.
+#![allow(dead_code, unused_macros)]
+#![allow(clippy::doc_markdown)]
+
+use crate::ffi::CStr;
+use core::ffi::c_void;
+use core::ptr::null_mut;
+use core::sync::atomic::{self, AtomicPtr, Ordering};
+use core::{marker, mem};
+
+const NULL: *mut c_void = null_mut();
+const INVALID: *mut c_void = 1 as *mut c_void;
+
+macro_rules! weak {
+    ($vis:vis fn $name:ident($($t:ty),*) -> $ret:ty) => (
+        #[allow(non_upper_case_globals)]
+        $vis static $name: $crate::backend::weak::Weak<unsafe extern fn($($t),*) -> $ret> =
+            $crate::backend::weak::Weak::new(concat!(stringify!($name), '\0'));
+    )
+}
+
+pub(crate) struct Weak<F> {
+    name: &'static str,
+    addr: AtomicPtr<c_void>,
+    _marker: marker::PhantomData<F>,
+}
+
+impl<F> Weak<F> {
+    pub(crate) const fn new(name: &'static str) -> Self {
+        Self {
+            name,
+            addr: AtomicPtr::new(INVALID),
+            _marker: marker::PhantomData,
+        }
+    }
+
+    pub(crate) fn get(&self) -> Option<F> {
+        assert_eq!(mem::size_of::<F>(), mem::size_of::<usize>());
+        unsafe {
+            // Relaxed is fine here because we fence before reading through the
+            // pointer (see the comment below).
+            match self.addr.load(Ordering::Relaxed) {
+                INVALID => self.initialize(),
+                NULL => None,
+                addr => {
+                    let func = mem::transmute_copy::<*mut c_void, F>(&addr);
+                    // The caller is presumably going to read through this value
+                    // (by calling the function we've dlsymed). This means we'd
+                    // need to have loaded it with at least C11's consume
+                    // ordering in order to be guaranteed that the data we read
+                    // from the pointer isn't from before the pointer was
+                    // stored. Rust has no equivalent to memory_order_consume,
+                    // so we use an acquire fence (sorry, ARM).
+                    //
+                    // Now, in practice this likely isn't needed even on CPUs
+                    // where relaxed and consume mean different things. The
+                    // symbols we're loading are probably present (or not) at
+                    // init, and even if they aren't the runtime dynamic loader
+                    // is extremely likely have sufficient barriers internally
+                    // (possibly implicitly, for example the ones provided by
+                    // invoking `mprotect`).
+                    //
+                    // That said, none of that's *guaranteed*, and so we fence.
+                    atomic::fence(Ordering::Acquire);
+                    Some(func)
+                }
+            }
+        }
+    }
+
+    // Cold because it should only happen during first-time initialization.
+    #[cold]
+    unsafe fn initialize(&self) -> Option<F> {
+        let val = fetch(self.name);
+        // This synchronizes with the acquire fence in `get`.
+        self.addr.store(val, Ordering::Release);
+
+        match val {
+            NULL => None,
+            addr => Some(mem::transmute_copy::<*mut c_void, F>(&addr)),
+        }
+    }
+}
+
+unsafe fn fetch(name: &str) -> *mut c_void {
+    let name = match CStr::from_bytes_with_nul(name.as_bytes()) {
+        Ok(c_str) => c_str,
+        Err(..) => return null_mut(),
+    };
+    libc::dlsym(libc::RTLD_DEFAULT, name.as_ptr().cast())
+}
+
+#[cfg(not(any(target_os = "android", target_os = "linux")))]
+macro_rules! syscall {
+    (fn $name:ident($($arg_name:ident: $t:ty),*) via $_sys_name:ident -> $ret:ty) => (
+        unsafe fn $name($($arg_name: $t),*) -> $ret {
+            weak! { fn $name($($t),*) -> $ret }
+
+            if let Some(fun) = $name.get() {
+                fun($($arg_name),*)
+            } else {
+                libc_errno::set_errno(libc_errno::Errno(libc::ENOSYS));
+                -1
+            }
+        }
+    )
+}
+
+#[cfg(any(target_os = "android", target_os = "linux"))]
+macro_rules! syscall {
+    (fn $name:ident($($arg_name:ident: $t:ty),*) via $sys_name:ident -> $ret:ty) => (
+        unsafe fn $name($($arg_name:$t),*) -> $ret {
+            // This looks like a hack, but concat_idents only accepts idents
+            // (not paths).
+            use libc::*;
+
+            trait AsSyscallArg {
+                type SyscallArgType;
+                fn into_syscall_arg(self) -> Self::SyscallArgType;
+            }
+
+            // Pass pointer types as pointers, to preserve provenance.
+            impl<T> AsSyscallArg for *mut T {
+                type SyscallArgType = *mut T;
+                fn into_syscall_arg(self) -> Self::SyscallArgType { self }
+            }
+            impl<T> AsSyscallArg for *const T {
+                type SyscallArgType = *const T;
+                fn into_syscall_arg(self) -> Self::SyscallArgType { self }
+            }
+
+            // Pass `BorrowedFd` values as the integer value.
+            impl AsSyscallArg for $crate::fd::BorrowedFd<'_> {
+                type SyscallArgType = c::c_long;
+                fn into_syscall_arg(self) -> Self::SyscallArgType {
+                    $crate::fd::AsRawFd::as_raw_fd(&self) as _
+                }
+            }
+
+            // Coerce integer values into `c_long`.
+            impl AsSyscallArg for i32 {
+                type SyscallArgType = c::c_long;
+                fn into_syscall_arg(self) -> Self::SyscallArgType { self as _ }
+            }
+            impl AsSyscallArg for u32 {
+                type SyscallArgType = c::c_long;
+                fn into_syscall_arg(self) -> Self::SyscallArgType { self as _ }
+            }
+            impl AsSyscallArg for usize {
+                type SyscallArgType = c::c_long;
+                fn into_syscall_arg(self) -> Self::SyscallArgType { self as _ }
+            }
+
+            // `concat_idents is unstable, so we take an extra `sys_name`
+            // parameter and have our users do the concat for us for now.
+            /*
+            syscall(
+                concat_idents!(SYS_, $name),
+                $($arg_name.into_syscall_arg()),*
+            ) as $ret
+            */
+
+            syscall($sys_name, $($arg_name.into_syscall_arg()),*) as $ret
+        }
+    )
+}
+
+macro_rules! weakcall {
+    ($vis:vis fn $name:ident($($arg_name:ident: $t:ty),*) -> $ret:ty) => (
+        $vis unsafe fn $name($($arg_name: $t),*) -> $ret {
+            weak! { fn $name($($t),*) -> $ret }
+
+            // Use a weak symbol from libc when possible, allowing `LD_PRELOAD`
+            // interposition, but if it's not found just fail.
+            if let Some(fun) = $name.get() {
+                fun($($arg_name),*)
+            } else {
+                libc_errno::set_errno(libc_errno::Errno(libc::ENOSYS));
+                -1
+            }
+        }
+    )
+}
+
+/// A combination of `weakcall` and `syscall`. Use the libc function if it's
+/// available, and fall back to `libc::syscall` otherwise.
+macro_rules! weak_or_syscall {
+    ($vis:vis fn $name:ident($($arg_name:ident: $t:ty),*) via $sys_name:ident -> $ret:ty) => (
+        $vis unsafe fn $name($($arg_name: $t),*) -> $ret {
+            weak! { fn $name($($t),*) -> $ret }
+
+            // Use a weak symbol from libc when possible, allowing `LD_PRELOAD`
+            // interposition, but if it's not found just fail.
+            if let Some(fun) = $name.get() {
+                fun($($arg_name),*)
+            } else {
+                syscall! { fn $name($($arg_name: $t),*) via $sys_name -> $ret }
+                $name($($arg_name),*)
+            }
+        }
+    )
+}