Merge pull request #4405 from nia-e/standalone-ptrace

trace: implement supervisor components for tracing

Author: oli-obk

src/tools/miri/Cargo.lock

@@ -121,6 +121,26 @@ dependencies = [
  "serde",
 ]
 
+[[package]]
+name = "capstone"
+version = "0.13.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "015ef5d5ca1743e3f94af9509ba6bd2886523cfee46e48d15c2ef5216fd4ac9a"
+dependencies = [
+ "capstone-sys",
+ "libc",
+]
+
+[[package]]
+name = "capstone-sys"
+version = "0.17.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2267cb8d16a1e4197863ec4284ffd1aec26fe7e57c58af46b02590a0235809a0"
+dependencies = [
+ "cc",
+ "libc",
+]
+
 [[package]]
 name = "cargo-platform"
 version = "0.1.9"
@@ -591,6 +611,7 @@ version = "0.1.0"
 dependencies = [
  "aes",
  "bitflags",
+ "capstone",
  "chrono",
  "chrono-tz",
  "colored",

src/tools/miri/Cargo.toml

@@ -44,6 +44,7 @@ libloading = "0.8"
 nix = { version = "0.30.1", features = ["mman", "ptrace", "signal"] }
 ipc-channel = "0.19.0"
 serde = { version = "1.0.219", features = ["derive"] }
+capstone = "0.13"
 
 [dev-dependencies]
 ui_test = "0.29.1"

src/tools/miri/src/alloc/isolated_alloc.rs

@@ -1,5 +1,7 @@
 use std::alloc::Layout;
+use std::ptr::NonNull;
 
+use nix::sys::mman;
 use rustc_index::bit_set::DenseBitSet;
 
 /// How many bytes of memory each bit in the bitset represents.
@@ -12,7 +14,7 @@ pub struct IsolatedAlloc {
     /// Pointers to page-aligned memory that has been claimed by the allocator.
     /// Every pointer here must point to a page-sized allocation claimed via
     /// mmap. These pointers are used for "small" allocations.
-    page_ptrs: Vec<*mut u8>,
+    page_ptrs: Vec<NonNull<u8>>,
     /// Metadata about which bytes have been allocated on each page. The length
     /// of this vector must be the same as that of `page_ptrs`, and the domain
     /// size of the bitset must be exactly `page_size / COMPRESSION_FACTOR`.
@@ -24,7 +26,7 @@ pub struct IsolatedAlloc {
     page_infos: Vec<DenseBitSet<usize>>,
     /// 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<(*mut u8, usize)>,
+    huge_ptrs: Vec<(NonNull<u8>, usize)>,
     /// The host (not emulated) page size.
     page_size: usize,
 }
@@ -137,7 +139,7 @@ impl IsolatedAlloc {
     unsafe fn alloc_small(
         page_size: usize,
         layout: Layout,
-        page: *mut u8,
+        page: NonNull<u8>,
         pinfo: &mut DenseBitSet<usize>,
         zeroed: bool,
     ) -> Option<*mut u8> {
@@ -164,15 +166,15 @@ impl IsolatedAlloc {
                         // zero out, even if we allocated more
                         ptr.write_bytes(0, layout.size());
                     }
-                    return Some(ptr);
+                    return Some(ptr.as_ptr());
                 }
             }
         }
         None
     }
 
     /// Expands the available memory pool by adding one page.
-    fn add_page(&mut self) -> (*mut u8, &mut DenseBitSet<usize>) {
+    fn add_page(&mut self) -> (NonNull<u8>, &mut DenseBitSet<usize>) {
         // SAFETY: mmap is always safe to call when requesting anonymous memory
         let page_ptr = unsafe {
             libc::mmap(
@@ -189,8 +191,8 @@ impl IsolatedAlloc {
         // `page_infos` has to have one bit for each `COMPRESSION_FACTOR`-sized chunk of bytes in the page.
         assert!(self.page_size % COMPRESSION_FACTOR == 0);
         self.page_infos.push(DenseBitSet::new_empty(self.page_size / COMPRESSION_FACTOR));
-        self.page_ptrs.push(page_ptr);
-        (page_ptr, self.page_infos.last_mut().unwrap())
+        self.page_ptrs.push(NonNull::new(page_ptr).unwrap());
+        (NonNull::new(page_ptr).unwrap(), self.page_infos.last_mut().unwrap())
     }
 
     /// Allocates in multiples of one page on the host system.
@@ -212,7 +214,7 @@ impl IsolatedAlloc {
             .cast::<u8>()
         };
         assert_ne!(ret.addr(), usize::MAX, "mmap failed");
-        self.huge_ptrs.push((ret, size));
+        self.huge_ptrs.push((NonNull::new(ret).unwrap(), size));
         // huge_normalized_layout ensures that we've overallocated enough space
         // for this to be valid.
         ret.map_addr(|a| a.next_multiple_of(layout.align()))
@@ -246,7 +248,7 @@ impl IsolatedAlloc {
                 // from us pointing to this page, and we know it was allocated
                 // in add_page as exactly a single page.
                 unsafe {
-                    assert_eq!(libc::munmap(page_ptr.cast(), self.page_size), 0);
+                    assert_eq!(libc::munmap(page_ptr.as_ptr().cast(), self.page_size), 0);
                 }
             }
         }
@@ -265,7 +267,7 @@ impl IsolatedAlloc {
         // This could be made faster if the list was sorted -- the allocator isn't fully optimized at the moment.
         let pinfo = std::iter::zip(&mut self.page_ptrs, &mut self.page_infos)
             .enumerate()
-            .find(|(_, (page, _))| page.addr() == page_addr);
+            .find(|(_, (page, _))| page.addr().get() == page_addr);
         let Some((idx_of_pinfo, (_, pinfo))) = pinfo else {
             panic!("Freeing in an unallocated page: {ptr:?}\nHolding pages {:?}", self.page_ptrs)
         };
@@ -287,30 +289,67 @@ impl IsolatedAlloc {
             .huge_ptrs
             .iter()
             .position(|&(pg, size)| {
-                pg.addr() <= ptr.addr() && ptr.addr() < pg.addr().strict_add(size)
+                pg.addr().get() <= ptr.addr() && ptr.addr() < pg.addr().get().strict_add(size)
             })
             .expect("Freeing unallocated pages");
         // And kick it from the list
         let (un_offset_ptr, size2) = self.huge_ptrs.remove(idx);
         assert_eq!(size, size2, "got wrong layout in dealloc");
         // SAFETY: huge_ptrs contains allocations made with mmap with the size recorded there.
         unsafe {
-            let ret = libc::munmap(un_offset_ptr.cast(), size);
+            let ret = libc::munmap(un_offset_ptr.as_ptr().cast(), size);
             assert_eq!(ret, 0);
         }
     }
 
     /// Returns a vector of page addresses managed by the allocator.
     pub fn pages(&self) -> Vec<usize> {
-        let mut pages: Vec<_> =
-            self.page_ptrs.clone().into_iter().map(|p| p.expose_provenance()).collect();
-        for (ptr, size) in &self.huge_ptrs {
+        let mut pages: Vec<usize> =
+            self.page_ptrs.clone().into_iter().map(|p| p.expose_provenance().get()).collect();
+        self.huge_ptrs.iter().for_each(|(ptr, size)| {
             for i in 0..size / self.page_size {
-                pages.push(ptr.expose_provenance().strict_add(i * self.page_size));
+                pages.push(ptr.expose_provenance().get().strict_add(i * self.page_size));
             }
-        }
+        });
         pages
     }
+
+    /// 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> {
+        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) {
+        let prot = mman::ProtFlags::PROT_READ | mman::ProtFlags::PROT_WRITE;
+        unsafe {
+            self.mprotect(prot).unwrap();
+        }
+    }
+
+    /// Applies `prot` to every page managed by the allocator.
+    ///
+    /// SAFETY: Accessing memory in violation of the protection flags will
+    /// trigger a segfault.
+    unsafe fn mprotect(&mut self, prot: mman::ProtFlags) -> Result<(), nix::errno::Errno> {
+        for &pg in &self.page_ptrs {
+            unsafe {
+                mman::mprotect(pg.cast(), self.page_size, prot)?;
+            }
+        }
+        for &(hpg, size) in &self.huge_ptrs {
+            unsafe {
+                mman::mprotect(hpg.cast(), size.next_multiple_of(self.page_size), prot)?;
+            }
+        }
+        Ok(())
+    }
 }
 
 #[cfg(test)]

src/tools/miri/src/shims/native_lib.rs

@@ -229,7 +229,14 @@ pub trait EvalContextExt<'tcx>: crate::MiriInterpCxExt<'tcx> {
             .collect::<Vec<libffi::high::Arg<'_>>>();
 
         // Call the function and store output, depending on return type in the function signature.
-        let (ret, _) = this.call_native_with_args(link_name, dest, code_ptr, libffi_args)?;
+        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?}");
+        }
 
         this.write_immediate(*ret, dest)?;
         interp_ok(true)
@@ -250,15 +257,15 @@ unsafe fn do_native_call<T: libffi::high::CType>(
 
     unsafe {
         if let Some(alloc) = alloc {
-            // SAFETY: We don't touch the machine memory past this point
+            // SAFETY: We don't touch the machine memory past this point.
             let (guard, stack_ptr) = Supervisor::start_ffi(alloc.clone());
-            // SAFETY: Upheld by caller
+            // SAFETY: Upheld by caller.
             let ret = ffi::call(ptr, args);
             // SAFETY: We got the guard and stack pointer from start_ffi, and
-            // the allocator is the same
+            // the allocator is the same.
             (ret, Supervisor::end_ffi(guard, alloc, stack_ptr))
         } else {
-            // SAFETY: Upheld by caller
+            // SAFETY: Upheld by caller.
             (ffi::call(ptr, args), None)
         }
     }