native-lib/trace: Intercept mmap and munmap

src/alloc/alloc_bytes.rs

@@ -12,8 +12,12 @@ use crate::helpers::ToU64 as _;
 
 #[derive(Clone, Debug)]
 pub enum MiriAllocParams {
+    /// Allocation was created by calling `alloc::alloc()`.
     Global,
+    /// Allocation came from the isolated allocator.
     Isolated(Rc<RefCell<IsolatedAlloc>>),
+    /// Page mapped elsewhere that we don't try to deallocate.
+    Forged(usize),
 }
 
 /// Allocation bytes that explicitly handle the layout of the data they're storing.
@@ -27,8 +31,7 @@ pub struct MiriAllocBytes {
     /// * If `self.layout.size() == 0`, then `self.ptr` was allocated with the equivalent layout with size 1.
     /// * Otherwise, `self.ptr` points to memory allocated with `self.layout`.
     ptr: *mut u8,
-    /// Whether this instance of `MiriAllocBytes` had its allocation created by calling `alloc::alloc()`
-    /// (`Global`) or the discrete allocator (`Isolated`)
+    /// Metadata on where this allocation came from and therefore how to deallocate it.
     params: MiriAllocParams,
 }
 
@@ -56,6 +59,13 @@ impl Drop for MiriAllocBytes {
                 MiriAllocParams::Global => alloc::dealloc(self.ptr, alloc_layout),
                 MiriAllocParams::Isolated(alloc) =>
                     alloc.borrow_mut().dealloc(self.ptr, alloc_layout),
+                // We can't nicely support mapping a page on one side of the FFI
+                // bound and freeing on the other, so just do nothing on an attempt
+                // to free.
+                //
+                // FIXME: Should emit an unsupported diagnostic when `libc::munmap()`
+                // is manually called on memory backed by forged bytes.
+                MiriAllocParams::Forged(_) => (),
             }
         }
     }
@@ -121,6 +131,7 @@ impl AllocBytes for MiriAllocBytes {
             match params {
                 MiriAllocParams::Global => alloc::alloc(layout),
                 MiriAllocParams::Isolated(alloc) => alloc.borrow_mut().alloc(layout),
+                MiriAllocParams::Forged(addr) => std::ptr::with_exposed_provenance_mut(*addr),
             }
         };
         let alloc_bytes = MiriAllocBytes::alloc_with(size.to_u64(), align, params, alloc_fn)
@@ -141,6 +152,7 @@ impl AllocBytes for MiriAllocBytes {
             match params {
                 MiriAllocParams::Global => alloc::alloc_zeroed(layout),
                 MiriAllocParams::Isolated(alloc) => alloc.borrow_mut().alloc_zeroed(layout),
+                MiriAllocParams::Forged(addr) => std::ptr::with_exposed_provenance_mut(*addr),
             }
         };
         MiriAllocBytes::alloc_with(size, align, params, alloc_fn).ok()

src/alloc/isolated_alloc.rs

@@ -26,6 +26,9 @@ pub struct IsolatedAlloc {
     /// Pointers to multiple-page-sized allocations. These must also be page-aligned,
     /// with their size stored as the second element of the vector.
     huge_ptrs: Vec<(NonNull<u8>, usize)>,
+    /// Addresses of pages that we don't actually manage, but which were allocated
+    /// by foreign code and where we need to track accesses.
+    forged_pages: Vec<NonNull<u8>>,
     /// The host (not emulated) page size.
     page_size: usize,
 }
@@ -37,6 +40,7 @@ impl IsolatedAlloc {
             page_ptrs: Vec::new(),
             huge_ptrs: Vec::new(),
             page_infos: Vec::new(),
+            forged_pages: Vec::new(),
             // SAFETY: `sysconf(_SC_PAGESIZE)` is always safe to call at runtime
             // See https://www.man7.org/linux/man-pages/man3/sysconf.3.html
             page_size: unsafe { libc::sysconf(libc::_SC_PAGESIZE).try_into().unwrap() },
@@ -308,8 +312,31 @@ impl IsolatedAlloc {
     /// Returns a list of page ranges managed by the allocator, given in terms of pointers
     /// and size (in bytes).
     pub fn pages(&self) -> impl Iterator<Item = (NonNull<u8>, usize)> {
-        let pages = self.page_ptrs.iter().map(|&p| (p, self.page_size));
-        pages.chain(self.huge_ptrs.iter().copied())
+        let with_pg_sz = |&p| (p, self.page_size);
+        let pages = self.page_ptrs.iter().map(with_pg_sz);
+        pages.chain(self.huge_ptrs.iter().copied()).chain(self.forged_pages.iter().map(with_pg_sz))
+    }
+
+    /// Makes the allocator also return this page address when `pages` is called.
+    pub fn forge_page(&mut self, addr: usize) {
+        assert!(addr.is_multiple_of(self.page_size), "Address is not page-aligned");
+        assert!(
+            !self.forged_pages.iter().any(|a| a.addr().get() == addr),
+            "Page already contained"
+        );
+        self.forged_pages.push(NonNull::new(std::ptr::with_exposed_provenance_mut(addr)).unwrap());
+    }
+
+    /// Deletes an entry from the list of forged pages.
+    pub fn remove_forged(&mut self, addr: usize) {
+        assert!(addr.is_multiple_of(self.page_size), "Address is not page-aligned");
+        let (index, _) = self
+            .forged_pages
+            .iter()
+            .enumerate()
+            .find(|(_, p_addr)| addr == p_addr.addr().get())
+            .expect("Page not contained");
+        self.forged_pages.remove(index);
     }
 }
 

src/shims/native_lib/mod.rs

@@ -26,11 +26,27 @@ use self::ffi::OwnedArg;
 use crate::*;
 
 /// The final results of an FFI trace, containing every relevant event detected
-/// by the tracer.
+/// by the tracer. Events are ordered sequentially by the real time they occurred.
+pub type MemEvents = Vec<SingleEvent>;
+
+/// Singular event occurring in an FFI call.
 #[derive(Serialize, Deserialize, Debug)]
-pub struct MemEvents {
-    /// An list of memory accesses that occurred, in the order they occurred in.
-    pub acc_events: Vec<AccessEvent>,
+pub enum SingleEvent {
+    Acc(AccessEvent),
+    Map(MapEvent),
+}
+
+/// A single page in the address space being modified. Addresses must always be a
+/// multiple of the system page size, and the event is assumed to span from the
+/// address to `addr + page_size`.
+///
+/// TODO: Support pages that are not (just) RW.
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub enum MapEvent {
+    /// A page was mapped with this base address.
+    Mmap(usize),
+    /// The page at this address was unmapped.
+    Munmap(usize),
 }
 
 /// A single memory access.
@@ -209,58 +225,68 @@ trait EvalContextExtPriv<'tcx>: crate::MiriInterpCxExt<'tcx> {
     }
 
     /// 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> {
+    /// ordered sequentially by when they occurred.
+    fn tracing_apply(&mut self, events: MemEvents) -> InterpResult<'tcx> {
+        let this = self.eval_context_mut();
+        for evt in events {
+            match evt {
+                SingleEvent::Acc(acc) => this.tracing_apply_access(acc)?,
+                SingleEvent::Map(map) => this.tracing_apply_mapping(map)?,
+            }
+        }
+        interp_ok(())
+    }
+
+    /// Applies the possible effects of a single memory access. Sizes are assumed
+    /// to be exact.
+    fn tracing_apply_access(&mut self, acc: AccessEvent) -> 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())
-                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(_) => {
-                        // If a provenance was read by the foreign code, expose it.
-                        for prov in alloc.provenance().get_range(this, overlap.into()) {
-                            this.expose_provenance(prov)?;
-                        }
+        let acc_rg = acc.get_range();
+        // LLVM at least permits vectorising accesses to adjacent allocations,
+        // so we cannot assume 1 access = 1 allocation. :(
+        let mut rg = acc_rg.addr..acc_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())
+            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 acc {
+                AccessEvent::Read(_) => {
+                    // If a provenance was read by the foreign code, expose it.
+                    for prov in alloc.provenance().get_range(this, overlap.into()) {
+                        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()),
-                                }),
-                            )
-                        }
+                }
+                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()),
+                            }),
+                        )
                     }
                 }
             }
@@ -269,6 +295,40 @@ trait EvalContextExtPriv<'tcx>: crate::MiriInterpCxExt<'tcx> {
         interp_ok(())
     }
 
+    /// Forges an allocation corresponding to a page mapping.
+    fn tracing_apply_mapping(&mut self, map: MapEvent) -> InterpResult<'tcx> {
+        let this = self.eval_context_mut();
+        let kind = MemoryKind::Machine(MiriMemoryKind::Mmap);
+
+        match map {
+            MapEvent::Mmap(addr) => {
+                let page_size = this.machine.page_size;
+                // Pretend an allocation was created at this address, and register
+                // it with the machine's allocator so it can track it.
+                let forged = Allocation::new(
+                    Size::from_bytes(page_size),
+                    rustc_abi::Align::from_bytes(page_size).unwrap(),
+                    AllocInit::Zero,
+                    crate::alloc::MiriAllocParams::Forged(addr),
+                );
+                let ptr = this.insert_allocation(forged, kind)?;
+                this.expose_provenance(ptr.provenance)?;
+                // Also make sure accesses on this page are intercepted.
+                this.machine.allocator.as_mut().unwrap().borrow_mut().forge_page(addr);
+            }
+            MapEvent::Munmap(addr) => {
+                let ptr = this.ptr_from_addr_cast(addr.to_u64())?;
+                // This will call `munmap` on already-unmapped memory; that's fine,
+                // since we intentionally ignore the returned error from `munmap`
+                // to allow this without more invasive changes.
+                this.deallocate_ptr(ptr, None, kind)?;
+                this.machine.allocator.as_mut().unwrap().borrow_mut().remove_forged(addr);
+            }
+        }
+
+        interp_ok(())
+    }
+
     /// Extract the value from the result of reading an operand from the machine
     /// and convert it to a `OwnedArg`.
     fn op_to_ffi_arg(&self, v: &OpTy<'tcx>, tracing: bool) -> InterpResult<'tcx, OwnedArg> {
@@ -492,7 +552,7 @@ pub trait EvalContextExt<'tcx>: crate::MiriInterpCxExt<'tcx> {
             this.call_native_with_args(link_name, dest, code_ptr, &mut libffi_args)?;
 
         if tracing {
-            this.tracing_apply_accesses(maybe_memevents.unwrap())?;
+            this.tracing_apply(maybe_memevents.unwrap())?;
         }
 
         this.write_immediate(*ret, dest)?;

src/shims/native_lib/trace/child.rs

@@ -55,6 +55,17 @@ impl Supervisor {
         Ok(())
     }
 
+    unsafe fn protect_pages_ignore_errs(
+        pages: impl Iterator<Item = (NonNull<u8>, usize)>,
+        prot: mman::ProtFlags,
+    ) {
+        for (pg, sz) in pages {
+            unsafe {
+                let _ = mman::mprotect(pg.cast(), sz, prot);
+            };
+        }
+    }
+
     /// Performs an arbitrary FFI call, enabling tracing from the supervisor.
     /// As this locks the supervisor via a mutex, no other threads may enter FFI
     /// until this function returns.
@@ -113,11 +124,12 @@ impl Supervisor {
 
             // SAFETY: We set memory back to normal, so this is safe.
             unsafe {
-                Self::protect_pages(
+                // Use the error-ignoring variant here, since it's possible that
+                // foreign code may have unmapped a page of ours.
+                Self::protect_pages_ignore_errs(
                     alloc.pages(),
                     mman::ProtFlags::PROT_READ | mman::ProtFlags::PROT_WRITE,
-                )
-                .unwrap();
+                );
             }
 
             // Signal the supervisor that we are done. Will block until the supervisor continues us.