trace: incorporate events

src/lib.rs

@@ -15,6 +15,7 @@
 #![feature(unqualified_local_imports)]
 #![feature(derive_coerce_pointee)]
 #![feature(arbitrary_self_types)]
+#![feature(iter_advance_by)]
 // Configure clippy and other lints
 #![allow(
     clippy::collapsible_else_if,

src/shims/native_lib/mod.rs

@@ -34,18 +34,25 @@ pub struct MemEvents {
 /// A single memory access.
 #[allow(dead_code)]
 #[cfg_attr(target_os = "linux", derive(serde::Serialize, serde::Deserialize))]
-#[derive(Debug)]
+#[derive(Clone, 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.
+    /// A read occurred on this memory range.
     Read(AccessRange),
-    /// A read may have occurred on this memory range.
+    /// A write 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),
+    /// If the second field is true, the access definitely happened.
+    Write(AccessRange, bool),
+}
+
+impl AccessEvent {
+    fn get_range(&self) -> AccessRange {
+        match self {
+            AccessEvent::Read(access_range) => access_range.clone(),
+            AccessEvent::Write(access_range, _) => access_range.clone(),
+        }
+    }
 }
 
 /// The memory touched by a given access.
@@ -59,6 +66,12 @@ pub struct AccessRange {
     pub size: usize,
 }
 
+impl AccessRange {
+    fn end(&self) -> usize {
+        self.addr.strict_add(self.size)
+    }
+}
+
 impl<'tcx> EvalContextExtPriv<'tcx> for crate::MiriInterpCx<'tcx> {}
 trait EvalContextExtPriv<'tcx>: crate::MiriInterpCxExt<'tcx> {
     /// Call native host function and return the output as an immediate.
@@ -196,6 +209,73 @@ trait EvalContextExtPriv<'tcx>: crate::MiriInterpCxExt<'tcx> {
         }
         None
     }
+
+    /// Applies the `events` to Miri's internal state. The event vector must be
+    /// ordered sequentially by when the accesses happened, and the sizes are
+    /// assumed to be exact.
+    fn tracing_apply_accesses(&mut self, events: MemEvents) -> InterpResult<'tcx> {
+        let this = self.eval_context_mut();
+
+        for evt in events.acc_events {
+            let evt_rg = evt.get_range();
+            // LLVM at least permits vectorising accesses to adjacent allocations,
+            // so we cannot assume 1 access = 1 allocation. :(
+            let mut rg = evt_rg.addr..evt_rg.end();
+            while let Some(curr) = rg.next() {
+                let Some(alloc_id) = this.alloc_id_from_addr(
+                    curr.to_u64(),
+                    rg.len().try_into().unwrap(),
+                    /* only_exposed_allocations */ true,
+                ) else {
+                    throw_ub_format!("Foreign code did an out-of-bounds access!")
+                };
+                let alloc = this.get_alloc_raw(alloc_id)?;
+                // The logical and physical address of the allocation coincide, so we can use
+                // this instead of `addr_from_alloc_id`.
+                let alloc_addr = alloc.get_bytes_unchecked_raw().addr();
+
+                // Determine the range inside the allocation that this access covers. This range is
+                // in terms of offsets from the start of `alloc`. The start of the overlap range
+                // will be `curr`; the end will be the minimum of the end of the allocation and the
+                // end of the access' range.
+                let overlap = curr.strict_sub(alloc_addr)
+                    ..std::cmp::min(alloc.len(), rg.end.strict_sub(alloc_addr));
+                // Skip forward however many bytes of the access are contained in the current
+                // allocation, subtracting 1 since the overlap range includes the current addr
+                // that was already popped off of the range.
+                rg.advance_by(overlap.len().strict_sub(1)).unwrap();
+
+                match evt {
+                    AccessEvent::Read(_) => {
+                        // FIXME: ProvenanceMap should have something like get_range().
+                        let p_map = alloc.provenance();
+                        for idx in overlap {
+                            // If a provenance was read by the foreign code, expose it.
+                            if let Some(prov) = p_map.get(Size::from_bytes(idx), this) {
+                                this.expose_provenance(prov)?;
+                            }
+                        }
+                    }
+                    AccessEvent::Write(_, certain) => {
+                        // Sometimes we aren't certain if a write happened, in which case we
+                        // only initialise that data if the allocation is mutable.
+                        if certain || alloc.mutability.is_mut() {
+                            let (alloc, cx) = this.get_alloc_raw_mut(alloc_id)?;
+                            alloc.process_native_write(
+                                &cx.tcx,
+                                Some(AllocRange {
+                                    start: Size::from_bytes(overlap.start),
+                                    size: Size::from_bytes(overlap.len()),
+                                }),
+                            )
+                        }
+                    }
+                }
+            }
+        }
+
+        interp_ok(())
+    }
 }
 
 impl<'tcx> EvalContextExt<'tcx> for crate::MiriInterpCx<'tcx> {}
@@ -221,6 +301,9 @@ pub trait EvalContextExt<'tcx>: crate::MiriInterpCxExt<'tcx> {
             }
         };
 
+        // Do we have ptrace?
+        let tracing = trace::Supervisor::is_enabled();
+
         // Get the function arguments, and convert them to `libffi`-compatible form.
         let mut libffi_args = Vec::<CArg>::with_capacity(args.len());
         for arg in args.iter() {
@@ -240,9 +323,7 @@ pub trait EvalContextExt<'tcx>: crate::MiriInterpCxExt<'tcx> {
                 // The first time this happens, print a warning.
                 if !this.machine.native_call_mem_warned.replace(true) {
                     // Newly set, so first time we get here.
-                    this.emit_diagnostic(NonHaltingDiagnostic::NativeCallSharedMem {
-                        tracing: self::trace::Supervisor::is_enabled(),
-                    });
+                    this.emit_diagnostic(NonHaltingDiagnostic::NativeCallSharedMem { tracing });
                 }
 
                 this.expose_provenance(prov)?;
@@ -269,15 +350,23 @@ pub trait EvalContextExt<'tcx>: crate::MiriInterpCxExt<'tcx> {
             // be read by FFI. The `black_box` is defensive programming as LLVM likes
             // to (incorrectly) optimize away ptr2int casts whose result is unused.
             std::hint::black_box(alloc.get_bytes_unchecked_raw().expose_provenance());
-            // Expose all provenances in this allocation, since the native code can do $whatever.
-            for prov in alloc.provenance().provenances() {
-                this.expose_provenance(prov)?;
+
+            if !tracing {
+                // Expose all provenances in this allocation, since the native code can do $whatever.
+                // Can be skipped when tracing; in that case we'll expose just the actually-read parts later.
+                for prov in alloc.provenance().provenances() {
+                    this.expose_provenance(prov)?;
+                }
             }
 
             // Prepare for possible write from native code if mutable.
             if info.mutbl.is_mut() {
                 let (alloc, cx) = this.get_alloc_raw_mut(alloc_id)?;
-                alloc.process_native_write(&cx.tcx, None);
+                // These writes could initialize everything and wreck havoc with the pointers.
+                // We can skip that when tracing; in that case we'll later do that only for the memory that got actually written.
+                if !tracing {
+                    alloc.process_native_write(&cx.tcx, None);
+                }
                 // Also expose *mutable* provenance for the interpreter-level allocation.
                 std::hint::black_box(alloc.get_bytes_unchecked_raw_mut().expose_provenance());
             }
@@ -289,10 +378,8 @@ pub trait EvalContextExt<'tcx>: crate::MiriInterpCxExt<'tcx> {
         let (ret, maybe_memevents) =
             this.call_native_with_args(link_name, dest, code_ptr, libffi_args)?;
 
-        if cfg!(target_os = "linux")
-            && let Some(events) = maybe_memevents
-        {
-            trace!("Registered FFI events:\n{events:#0x?}");
+        if tracing {
+            this.tracing_apply_accesses(maybe_memevents.unwrap())?;
         }
 
         this.write_immediate(*ret, dest)?;

src/shims/native_lib/trace/parent.rs

@@ -58,11 +58,11 @@ impl ArchIndependentRegs for libc::user_regs_struct {
 #[rustfmt::skip]
 impl ArchIndependentRegs for libc::user_regs_struct {
     #[inline]
-    fn ip(&self) -> usize { self.eip.try_into().unwrap() }
+    fn ip(&self) -> usize { self.eip.cast_unsigned().try_into().unwrap() }
     #[inline]
-    fn set_ip(&mut self, ip: usize) { self.eip = ip.try_into().unwrap() }
+    fn set_ip(&mut self, ip: usize) { self.eip = ip.cast_signed().try_into().unwrap() }
     #[inline]
-    fn set_sp(&mut self, sp: usize) { self.esp = sp.try_into().unwrap() }
+    fn set_sp(&mut self, sp: usize) { self.esp = sp.cast_signed().try_into().unwrap() }
 }
 
 /// A unified event representing something happening on the child process. Wraps
@@ -386,7 +386,17 @@ fn capstone_find_events(
                         acc_events.push(AccessEvent::Read(push.clone()));
                     }
                     if acc_ty.is_writable() {
-                        acc_events.push(AccessEvent::Write(push));
+                        // FIXME: This could be made certain; either determine all cases where
+                        // only reads happen, or have an intermediate mempr_* function to first
+                        // map the page(s) as readonly and check if a segfault occurred.
+
+                        // Per https://docs.rs/iced-x86/latest/iced_x86/enum.OpAccess.html,
+                        // we know that the possible access types are Read, CondRead, Write,
+                        // CondWrite, ReadWrite, and ReadCondWrite. Since we got a segfault
+                        // we know some kind of access happened so Cond{Read, Write}s are
+                        // certain reads and writes; the only uncertainty is with an RW op
+                        // as it might be a ReadCondWrite with the write condition unmet.
+                        acc_events.push(AccessEvent::Write(push, !acc_ty.is_readable()));
                     }
 
                     return true;
@@ -442,8 +452,7 @@ fn handle_segfault(
     // Get information on what caused the segfault. This contains the address
     // that triggered it.
     let siginfo = ptrace::getsiginfo(pid).unwrap();
-    // All x86, ARM, etc. instructions only have at most one memory operand
-    // (thankfully!)
+    // All x86 instructions only have at most one memory operand (thankfully!)
     // SAFETY: si_addr is safe to call.
     let addr = unsafe { siginfo.si_addr().addr() };
     let page_addr = addr.strict_sub(addr.strict_rem(page_size));
@@ -490,7 +499,7 @@ fn handle_segfault(
     ptrace::write(
         pid,
         (&raw const PAGE_ADDR).cast_mut().cast(),
-        libc::c_long::try_from(page_addr).unwrap(),
+        libc::c_long::try_from(page_addr.cast_signed()).unwrap(),
     )
     .unwrap();
 

tests/native-lib/fail/tracing/partial_init.rs

@@ -0,0 +1,25 @@
+//@only-target: x86_64-unknown-linux-gnu i686-unknown-linux-gnu
+//@compile-flags: -Zmiri-native-lib-enable-tracing
+
+extern "C" {
+    fn init_n(n: i32, ptr: *mut u8);
+}
+
+fn main() {
+    partial_init();
+}
+
+// Initialise the first 2 elements of the slice from native code, and check
+// that the 3rd is correctly deemed uninit.
+fn partial_init() {
+    let mut slice = std::mem::MaybeUninit::<[u8; 3]>::uninit();
+    let slice_ptr = slice.as_mut_ptr().cast::<u8>();
+    unsafe {
+        // Initialize the first two elements.
+        init_n(2, slice_ptr);
+        assert!(*slice_ptr == 0);
+        assert!(*slice_ptr.offset(1) == 0);
+        // Reading the third is UB!
+        let _val = *slice_ptr.offset(2); //~ ERROR: Undefined Behavior: using uninitialized data
+    }
+}

tests/native-lib/fail/tracing/partial_init.stderr

@@ -0,0 +1,39 @@
+warning: sharing memory with a native function called via FFI
+  --> tests/native-lib/fail/tracing/partial_init.rs:LL:CC
+   |
+LL |         init_n(2, slice_ptr);
+   |         ^^^^^^^^^^^^^^^^^^^^ sharing memory with a native function
+   |
+   = help: when memory is shared with a native function call, Miri can only track initialisation and provenance on a best-effort basis
+   = help: in particular, Miri assumes that the native call initializes all memory it has written to
+   = help: Miri also assumes that any part of this memory may be a pointer that is permitted to point to arbitrary exposed memory
+   = help: what this means is that Miri will easily miss Undefined Behavior related to incorrect usage of this shared memory, so you should not take a clean Miri run as a signal that your FFI code is UB-free
+   = help: tracing memory accesses in native code is not yet fully implemented, so there can be further imprecisions beyond what is documented here
+   = note: BACKTRACE:
+   = note: inside `partial_init` at tests/native-lib/fail/tracing/partial_init.rs:LL:CC
+note: inside `main`
+  --> tests/native-lib/fail/tracing/partial_init.rs:LL:CC
+   |
+LL |     partial_init();
+   |     ^^^^^^^^^^^^^^
+
+error: Undefined Behavior: using uninitialized data, but this operation requires initialized memory
+  --> tests/native-lib/fail/tracing/partial_init.rs:LL:CC
+   |
+LL |         let _val = *slice_ptr.offset(2);
+   |                    ^^^^^^^^^^^^^^^^^^^^ Undefined Behavior occurred here
+   |
+   = help: this indicates a bug in the program: it performed an invalid operation, and caused Undefined Behavior
+   = help: see https://doc.rust-lang.org/nightly/reference/behavior-considered-undefined.html for further information
+   = note: BACKTRACE:
+   = note: inside `partial_init` at tests/native-lib/fail/tracing/partial_init.rs:LL:CC
+note: inside `main`
+  --> tests/native-lib/fail/tracing/partial_init.rs:LL:CC
+   |
+LL |     partial_init();
+   |     ^^^^^^^^^^^^^^
+
+note: some details are omitted, run with `MIRIFLAGS=-Zmiri-backtrace=full` for a verbose backtrace
+
+error: aborting due to 1 previous error; 1 warning emitted
+

tests/native-lib/fail/tracing/unexposed_reachable_alloc.rs

@@ -0,0 +1,29 @@
+//@only-target: x86_64-unknown-linux-gnu i686-unknown-linux-gnu
+//@compile-flags: -Zmiri-permissive-provenance -Zmiri-native-lib-enable-tracing
+
+extern "C" {
+    fn do_one_deref(ptr: *const *const *const i32) -> usize;
+}
+
+fn main() {
+    unexposed_reachable_alloc();
+}
+
+// Expose 2 pointers by virtue of doing a native read and assert that the 3rd in
+// the chain remains properly unexposed.
+fn unexposed_reachable_alloc() {
+    let inner = 42;
+    let intermediate_a = &raw const inner;
+    let intermediate_b = &raw const intermediate_a;
+    let exposed = &raw const intermediate_b;
+    // Discard the return value; it's just there so the access in C doesn't get optimised away.
+    unsafe { do_one_deref(exposed) };
+    // Native read should have exposed the address of intermediate_b...
+    let valid: *const i32 = std::ptr::with_exposed_provenance(intermediate_b.addr());
+    // but not of intermediate_a.
+    let invalid: *const i32 = std::ptr::with_exposed_provenance(intermediate_a.addr());
+    unsafe {
+        let _ok = *valid;
+        let _not_ok = *invalid; //~ ERROR: Undefined Behavior: memory access failed: attempting to access
+    }
+}

tests/native-lib/fail/tracing/unexposed_reachable_alloc.stderr

@@ -0,0 +1,39 @@
+warning: sharing memory with a native function called via FFI
+  --> tests/native-lib/fail/tracing/unexposed_reachable_alloc.rs:LL:CC
+   |
+LL |     unsafe { do_one_deref(exposed) };
+   |              ^^^^^^^^^^^^^^^^^^^^^ sharing memory with a native function
+   |
+   = help: when memory is shared with a native function call, Miri can only track initialisation and provenance on a best-effort basis
+   = help: in particular, Miri assumes that the native call initializes all memory it has written to
+   = help: Miri also assumes that any part of this memory may be a pointer that is permitted to point to arbitrary exposed memory
+   = help: what this means is that Miri will easily miss Undefined Behavior related to incorrect usage of this shared memory, so you should not take a clean Miri run as a signal that your FFI code is UB-free
+   = help: tracing memory accesses in native code is not yet fully implemented, so there can be further imprecisions beyond what is documented here
+   = note: BACKTRACE:
+   = note: inside `unexposed_reachable_alloc` at tests/native-lib/fail/tracing/unexposed_reachable_alloc.rs:LL:CC
+note: inside `main`
+  --> tests/native-lib/fail/tracing/unexposed_reachable_alloc.rs:LL:CC
+   |
+LL |     unexposed_reachable_alloc();
+   |     ^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+error: Undefined Behavior: memory access failed: attempting to access 4 bytes, but got $HEX[noalloc] which is a dangling pointer (it has no provenance)
+  --> tests/native-lib/fail/tracing/unexposed_reachable_alloc.rs:LL:CC
+   |
+LL |         let _not_ok = *invalid;
+   |                       ^^^^^^^^ Undefined Behavior occurred here
+   |
+   = help: this indicates a bug in the program: it performed an invalid operation, and caused Undefined Behavior
+   = help: see https://doc.rust-lang.org/nightly/reference/behavior-considered-undefined.html for further information
+   = note: BACKTRACE:
+   = note: inside `unexposed_reachable_alloc` at tests/native-lib/fail/tracing/unexposed_reachable_alloc.rs:LL:CC
+note: inside `main`
+  --> tests/native-lib/fail/tracing/unexposed_reachable_alloc.rs:LL:CC
+   |
+LL |     unexposed_reachable_alloc();
+   |     ^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+note: some details are omitted, run with `MIRIFLAGS=-Zmiri-backtrace=full` for a verbose backtrace
+
+error: aborting due to 1 previous error; 1 warning emitted
+

tests/native-lib/pass/ptr_read_access.rs

@@ -1,3 +1,7 @@
+//@revisions: trace notrace
+//@[trace] only-target: x86_64-unknown-linux-gnu i686-unknown-linux-gnu
+//@[trace] compile-flags: -Zmiri-native-lib-enable-tracing
+
 fn main() {
     test_access_pointer();
     test_access_simple();