various native-lib-trace-related fixups

Co-Authored-By: Ralf Jung <post@ralfj.de>

Author: nia-e

src/alloc/isolated_alloc.rs

@@ -302,31 +302,28 @@ impl IsolatedAlloc {
         }
     }
 
-    /// Returns a vector of page addresses managed by the allocator.
-    pub fn pages(&self) -> Vec<usize> {
-        let mut pages: Vec<usize> =
-            self.page_ptrs.iter().map(|p| p.expose_provenance().get()).collect();
-        for (ptr, size) in self.huge_ptrs.iter() {
-            for i in 0..size / self.page_size {
-                pages.push(ptr.expose_provenance().get().strict_add(i * self.page_size));
-            }
-        }
-        pages
+    /// Returns a list of page addresses managed by the allocator.
+    pub fn pages(&self) -> impl Iterator<Item = usize> {
+        let pages = self.page_ptrs.iter().map(|p| p.expose_provenance().get());
+        pages.chain(self.huge_ptrs.iter().flat_map(|(ptr, size)| {
+            (0..size / self.page_size)
+                .map(|i| ptr.expose_provenance().get().strict_add(i * self.page_size))
+        }))
     }
 
     /// Protects all owned memory as `PROT_NONE`, preventing accesses.
     ///
     /// SAFETY: Accessing memory after this point will result in a segfault
     /// unless it is first unprotected.
-    pub unsafe fn prepare_ffi(&mut self) -> Result<(), nix::errno::Errno> {
+    pub unsafe fn start_ffi(&mut self) -> Result<(), nix::errno::Errno> {
         let prot = mman::ProtFlags::PROT_NONE;
         unsafe { self.mprotect(prot) }
     }
 
     /// Deprotects all owned memory by setting it to RW. Erroring here is very
     /// likely unrecoverable, so it may panic if applying those permissions
     /// fails.
-    pub fn unprep_ffi(&mut self) {
+    pub fn end_ffi(&mut self) {
         let prot = mman::ProtFlags::PROT_READ | mman::ProtFlags::PROT_WRITE;
         unsafe {
             self.mprotect(prot).unwrap();

src/bin/miri.rs

@@ -233,8 +233,6 @@ impl rustc_driver::Callbacks for MiriCompilerCalls {
         } else {
             let return_code = miri::eval_entry(tcx, entry_def_id, entry_type, &config, None)
                 .unwrap_or_else(|| {
-                    //#[cfg(target_os = "linux")]
-                    //miri::native_lib::register_retcode_sv(rustc_driver::EXIT_FAILURE);
                     tcx.dcx().abort_if_errors();
                     rustc_driver::EXIT_FAILURE
                 });
@@ -337,6 +335,9 @@ impl rustc_driver::Callbacks for MiriBeRustCompilerCalls {
 fn exit(exit_code: i32) -> ! {
     // Drop the tracing guard before exiting, so tracing calls are flushed correctly.
     deinit_loggers();
+    // Make sure the supervisor knows about the code code.
+    #[cfg(target_os = "linux")]
+    miri::native_lib::register_retcode_sv(exit_code);
     std::process::exit(exit_code);
 }
 
@@ -355,6 +356,11 @@ fn run_compiler_and_exit(
     args: &[String],
     callbacks: &mut (dyn rustc_driver::Callbacks + Send),
 ) -> ! {
+    // Install the ctrlc handler that sets `rustc_const_eval::CTRL_C_RECEIVED`, even if
+    // MIRI_BE_RUSTC is set. We do this late so that when `native_lib::init_sv` is called,
+    // there are no other threads.
+    rustc_driver::install_ctrlc_handler();
+
     // Invoke compiler, catch any unwinding panics and handle return code.
     let exit_code =
         rustc_driver::catch_with_exit_code(move || rustc_driver::run_compiler(args, callbacks));
@@ -439,10 +445,6 @@ fn main() {
     let args = rustc_driver::catch_fatal_errors(|| rustc_driver::args::raw_args(&early_dcx))
         .unwrap_or_else(|_| std::process::exit(rustc_driver::EXIT_FAILURE));
 
-    // Install the ctrlc handler that sets `rustc_const_eval::CTRL_C_RECEIVED`, even if
-    // MIRI_BE_RUSTC is set.
-    rustc_driver::install_ctrlc_handler();
-
     // If the environment asks us to actually be rustc, then do that.
     if let Some(crate_kind) = env::var_os("MIRI_BE_RUSTC") {
         // Earliest rustc setup.
@@ -750,15 +752,15 @@ fn main() {
 
     debug!("rustc arguments: {:?}", rustc_args);
     debug!("crate arguments: {:?}", miri_config.args);
-    #[cfg(target_os = "linux")]
     if !miri_config.native_lib.is_empty() && miri_config.native_lib_enable_tracing {
-        // FIXME: This should display a diagnostic / warning on error
-        // SAFETY: If any other threads exist at this point (namely for the ctrlc
-        // handler), they will not interact with anything on the main rustc/Miri
-        // thread in an async-signal-unsafe way such as by accessing shared
-        // semaphores, etc.; the handler only calls `sleep()` and `exit()`, which
-        // are async-signal-safe, as is accessing atomics
-        //let _ = unsafe { miri::native_lib::init_sv() };
+        // SAFETY: No other threads are running
+        #[cfg(target_os = "linux")]
+        if unsafe { miri::native_lib::init_sv() }.is_err() {
+            eprintln!(
+                "warning: The native-lib tracer could not be started. Is this an x86 Linux system, and does Miri have permissions to ptrace?\n\
+                Falling back to non-tracing native-lib mode."
+            );
+        }
     }
     run_compiler_and_exit(
         &rustc_args,

src/lib.rs

@@ -96,10 +96,10 @@ pub use rustc_const_eval::interpret::{self, AllocMap, Provenance as _};
 use rustc_middle::{bug, span_bug};
 use tracing::{info, trace};
 
-//#[cfg(target_os = "linux")]
-//pub mod native_lib {
-//    pub use crate::shims::{init_sv, register_retcode_sv};
-//}
+#[cfg(target_os = "linux")]
+pub mod native_lib {
+    pub use crate::shims::{init_sv, register_retcode_sv};
+}
 
 // Type aliases that set the provenance parameter.
 pub type Pointer = interpret::Pointer<Option<machine::Provenance>>;

src/shims/mod.rs

@@ -22,8 +22,8 @@ pub mod tls;
 pub mod unwind;
 
 pub use self::files::FdTable;
-//#[cfg(target_os = "linux")]
-//pub use self::native_lib::trace::{init_sv, register_retcode_sv};
+#[cfg(target_os = "linux")]
+pub use self::native_lib::trace::{init_sv, register_retcode_sv};
 pub use self::unix::{DirTable, EpollInterestTable};
 
 /// What needs to be done after emulating an item (a shim or an intrinsic) is done.

src/shims/native_lib/mod.rs

@@ -1,9 +1,5 @@
 //! Implements calling functions from a native library.
 
-// FIXME: disabled since it fails to build on many targets.
-//#[cfg(target_os = "linux")]
-//pub mod trace;
-
 use std::ops::Deref;
 
 use libffi::high::call as ffi;
@@ -13,14 +9,55 @@ use rustc_middle::mir::interpret::Pointer;
 use rustc_middle::ty::{self as ty, IntTy, UintTy};
 use rustc_span::Symbol;
 
-//#[cfg(target_os = "linux")]
-//use self::trace::Supervisor;
+#[cfg_attr(
+    not(all(
+        target_os = "linux",
+        target_env = "gnu",
+        any(target_arch = "x86", target_arch = "x86_64")
+    )),
+    path = "trace/stub.rs"
+)]
+pub mod trace;
+
 use crate::*;
 
-//#[cfg(target_os = "linux")]
-//type CallResult<'tcx> = InterpResult<'tcx, (ImmTy<'tcx>, Option<self::trace::messages::MemEvents>)>;
-//#[cfg(not(target_os = "linux"))]
-type CallResult<'tcx> = InterpResult<'tcx, (ImmTy<'tcx>, Option<!>)>;
+/// The final results of an FFI trace, containing every relevant event detected
+/// by the tracer.
+#[allow(dead_code)]
+#[cfg_attr(target_os = "linux", derive(serde::Serialize, serde::Deserialize))]
+#[derive(Debug)]
+pub struct MemEvents {
+    /// An list of memory accesses that occurred, in the order they occurred in.
+    pub acc_events: Vec<AccessEvent>,
+}
+
+/// A single memory access.
+#[allow(dead_code)]
+#[cfg_attr(target_os = "linux", derive(serde::Serialize, serde::Deserialize))]
+#[derive(Debug)]
+pub enum AccessEvent {
+    /// A read may have occurred on this memory range.
+    /// Some instructions *may* read memory without *always* doing that,
+    /// so this can be an over-approximation.
+    /// The range info, however, is reliable if the access did happen.
+    Read(AccessRange),
+    /// A read may have occurred on this memory range.
+    /// Some instructions *may* write memory without *always* doing that,
+    /// so this can be an over-approximation.
+    /// The range info, however, is reliable if the access did happen.
+    Write(AccessRange),
+}
+
+/// The memory touched by a given access.
+#[allow(dead_code)]
+#[cfg_attr(target_os = "linux", derive(serde::Serialize, serde::Deserialize))]
+#[derive(Clone, Debug)]
+pub struct AccessRange {
+    /// The base address in memory where an access occurred.
+    pub addr: usize,
+    /// The number of bytes affected from the base.
+    pub size: usize,
+}
 
 impl<'tcx> EvalContextExtPriv<'tcx> for crate::MiriInterpCx<'tcx> {}
 trait EvalContextExtPriv<'tcx>: crate::MiriInterpCxExt<'tcx> {
@@ -31,18 +68,17 @@ trait EvalContextExtPriv<'tcx>: crate::MiriInterpCxExt<'tcx> {
         dest: &MPlaceTy<'tcx>,
         ptr: CodePtr,
         libffi_args: Vec<libffi::high::Arg<'a>>,
-    ) -> CallResult<'tcx> {
+    ) -> InterpResult<'tcx, (crate::ImmTy<'tcx>, Option<MemEvents>)> {
         let this = self.eval_context_mut();
-        //#[cfg(target_os = "linux")]
-        //let alloc = this.machine.allocator.as_ref().unwrap();
-
-        // SAFETY: We don't touch the machine memory past this point.
-        //#[cfg(target_os = "linux")]
-        //let (guard, stack_ptr) = unsafe { Supervisor::start_ffi(alloc) };
+        #[cfg(target_os = "linux")]
+        let alloc = this.machine.allocator.as_ref().unwrap();
+        #[cfg(not(target_os = "linux"))]
+        // Placeholder value.
+        let alloc = ();
 
-        // Call the function (`ptr`) with arguments `libffi_args`, and obtain the return value
-        // as the specified primitive integer type
-        let res = 'res: {
+        trace::Supervisor::do_ffi(alloc, || {
+            // Call the function (`ptr`) with arguments `libffi_args`, and obtain the return value
+            // as the specified primitive integer type
             let scalar = match dest.layout.ty.kind() {
                 // ints
                 ty::Int(IntTy::I8) => {
@@ -93,31 +129,22 @@ trait EvalContextExtPriv<'tcx>: crate::MiriInterpCxExt<'tcx> {
                 // have the output_type `Tuple([])`.
                 ty::Tuple(t_list) if (*t_list).deref().is_empty() => {
                     unsafe { ffi::call::<()>(ptr, libffi_args.as_slice()) };
-                    break 'res interp_ok(ImmTy::uninit(dest.layout));
+                    return interp_ok(ImmTy::uninit(dest.layout));
                 }
                 ty::RawPtr(..) => {
                     let x = unsafe { ffi::call::<*const ()>(ptr, libffi_args.as_slice()) };
                     let ptr = Pointer::new(Provenance::Wildcard, Size::from_bytes(x.addr()));
                     Scalar::from_pointer(ptr, this)
                 }
                 _ =>
-                    break 'res Err(err_unsup_format!(
+                    return Err(err_unsup_format!(
                         "unsupported return type for native call: {:?}",
                         link_name
                     ))
                     .into(),
             };
             interp_ok(ImmTy::from_scalar(scalar, dest.layout))
-        };
-
-        // SAFETY: We got the guard and stack pointer from start_ffi, and
-        // the allocator is the same
-        //#[cfg(target_os = "linux")]
-        //let events = unsafe { Supervisor::end_ffi(alloc, guard, stack_ptr) };
-        //#[cfg(not(target_os = "linux"))]
-        let events = None;
-
-        interp_ok((res?, events))
+        })
     }
 
     /// Get the pointer to the function of the specified name in the shared object file,
@@ -214,10 +241,7 @@ pub trait EvalContextExt<'tcx>: crate::MiriInterpCxExt<'tcx> {
                 if !this.machine.native_call_mem_warned.replace(true) {
                     // Newly set, so first time we get here.
                     this.emit_diagnostic(NonHaltingDiagnostic::NativeCallSharedMem {
-                        //#[cfg(target_os = "linux")]
-                        //tracing: self::trace::Supervisor::is_enabled(),
-                        //#[cfg(not(target_os = "linux"))]
-                        tracing: false,
+                        tracing: self::trace::Supervisor::is_enabled(),
                     });
                 }