Merge #1217

1217: Polymorphic host functions based on dynamic trampoline generation. r=losfair a=losfair

This PR implements polymorphic host functions by dynamically generating the "glue" code that translates platform arguments to an array in `runtime-core`.

TODO:

- ~~Multiple return values.~~ Deferring to a future multivalue PR.
- [x] Dynamic signatures for polymorphic functions.
- [x] Use a proper executable memory allocator.

Co-authored-by: losfair <zhy20000919@hotmail.com>
Co-authored-by: Ivan Enderlin <ivan.enderlin@hoa-project.net>
Co-authored-by: Heyang Zhou <zhy20000919@hotmail.com>

CHANGELOG.md

@@ -2,6 +2,7 @@
 
 ## **[Unreleased]**
 
+- [#1217](https://github.com/wasmerio/wasmer/pull/1217) Polymorphic host functions based on dynamic trampoline generation.
 - [#1252](https://github.com/wasmerio/wasmer/pull/1252) Allow `/` in wasi `--mapdir` wasm path.
 - [#1212](https://github.com/wasmerio/wasmer/pull/1212) Add support for GDB JIT debugging:
   - Add `--generate-debug-info` and `-g` flags to `wasmer run` to generate debug information during compilation. The debug info is passed via the GDB JIT interface to a debugger to allow source-level debugging of Wasm files. Currently only available on clif-backend.

lib/runtime-core-tests/tests/imports.rs

@@ -1,6 +1,13 @@
+use std::sync::Arc;
 use wasmer_runtime_core::{
-    compile_with, error::RuntimeError, imports, memory::Memory, typed_func::Func,
-    types::MemoryDescriptor, units::Pages, vm, Instance,
+    compile_with,
+    error::RuntimeError,
+    imports,
+    memory::Memory,
+    typed_func::{DynamicFunc, Func},
+    types::{FuncSig, MemoryDescriptor, Type, Value},
+    units::Pages,
+    vm, Instance,
 };
 use wasmer_runtime_core_tests::{get_compiler, wat2wasm};
 
@@ -68,6 +75,7 @@ fn imported_functions_forms(test: &dyn Fn(&Instance)) {
   (import "env" "memory" (memory 1 1))
   (import "env" "callback_fn" (func $callback_fn (type $type)))
   (import "env" "callback_closure" (func $callback_closure (type $type)))
+  (import "env" "callback_closure_dynamic" (func $callback_closure_dynamic (type $type)))
   (import "env" "callback_closure_with_env" (func $callback_closure_with_env (type $type)))
   (import "env" "callback_fn_with_vmctx" (func $callback_fn_with_vmctx (type $type)))
   (import "env" "callback_closure_with_vmctx" (func $callback_closure_with_vmctx (type $type)))
@@ -86,6 +94,10 @@ fn imported_functions_forms(test: &dyn Fn(&Instance)) {
     get_local 0
     call $callback_closure)
 
+  (func (export "function_closure_dynamic") (type $type)
+    get_local 0
+    call $callback_closure_dynamic)
+
   (func (export "function_closure_with_env") (type $type)
     get_local 0
     call $callback_closure_with_env)
@@ -142,6 +154,16 @@ fn imported_functions_forms(test: &dyn Fn(&Instance)) {
                 Ok(n + 1)
             }),
 
+            "callback_closure_dynamic" => DynamicFunc::new(
+                Arc::new(FuncSig::new(vec![Type::I32], vec![Type::I32])),
+                |_, params| -> Vec<Value> {
+                    match params[0] {
+                        Value::I32(x) => vec![Value::I32(x + 1)],
+                        _ => unreachable!()
+                    }
+                }
+            ),
+
             // Closure with a captured environment (a single variable + an instance of `Memory`).
             "callback_closure_with_env" => Func::new(move |n: i32| -> Result<i32, ()> {
                 let shift_ = shift + memory.view::<i32>()[0].get();
@@ -236,6 +258,7 @@ macro_rules! test {
 
 test!(test_fn, function_fn, Ok(2));
 test!(test_closure, function_closure, Ok(2));
+test!(test_closure_dynamic, function_closure_dynamic, Ok(2));
 test!(
     test_closure_with_env,
     function_closure_with_env,

lib/runtime-core/src/loader.rs

@@ -1,7 +1,9 @@
 //! The loader module functions are used to load an instance.
 use crate::{backend::RunnableModule, module::ModuleInfo, types::Type, types::Value, vm::Ctx};
 #[cfg(unix)]
-use libc::{mmap, mprotect, munmap, MAP_ANON, MAP_PRIVATE, PROT_EXEC, PROT_READ, PROT_WRITE};
+use libc::{
+    mmap, mprotect, munmap, MAP_ANON, MAP_NORESERVE, MAP_PRIVATE, PROT_EXEC, PROT_READ, PROT_WRITE,
+};
 use std::{
     fmt::Debug,
     ops::{Deref, DerefMut},
@@ -138,12 +140,12 @@ impl CodeMemory {
         unimplemented!("CodeMemory::new");
     }
 
-    /// Makes this code memory executable.
+    /// Makes this code memory executable and not writable.
     pub fn make_executable(&self) {
         unimplemented!("CodeMemory::make_executable");
     }
 
-    /// Makes this code memory writable.
+    /// Makes this code memory writable and not executable.
     pub fn make_writable(&self) {
         unimplemented!("CodeMemory::make_writable");
     }
@@ -169,7 +171,7 @@ impl CodeMemory {
                 std::ptr::null_mut(),
                 size,
                 PROT_READ | PROT_WRITE,
-                MAP_PRIVATE | MAP_ANON,
+                MAP_PRIVATE | MAP_ANON | MAP_NORESERVE,
                 -1,
                 0,
             )
@@ -183,19 +185,33 @@ impl CodeMemory {
         }
     }
 
-    /// Makes this code memory executable.
+    /// Makes this code memory executable and not writable.
     pub fn make_executable(&self) {
         if unsafe { mprotect(self.ptr as _, self.size, PROT_READ | PROT_EXEC) } != 0 {
             panic!("cannot set code memory to executable");
         }
     }
 
-    /// Makes this code memory writable.
+    /// Makes this code memory writable and not executable.
     pub fn make_writable(&self) {
         if unsafe { mprotect(self.ptr as _, self.size, PROT_READ | PROT_WRITE) } != 0 {
             panic!("cannot set code memory to writable");
         }
     }
+
+    /// Makes this code memory both writable and executable.
+    ///
+    /// Avoid using this if a combination `make_executable` and `make_writable` can be used.
+    pub fn make_writable_executable(&self) {
+        if unsafe { mprotect(self.ptr as _, self.size, PROT_READ | PROT_WRITE | PROT_EXEC) } != 0 {
+            panic!("cannot set code memory to writable and executable");
+        }
+    }
+
+    /// Returns the backing pointer of this code memory.
+    pub fn get_backing_ptr(&self) -> *mut u8 {
+        self.ptr
+    }
 }
 
 #[cfg(unix)]

lib/runtime-core/src/trampoline_x64.rs

@@ -8,7 +8,10 @@
 
 use crate::loader::CodeMemory;
 use crate::vm::Ctx;
+use std::collections::BTreeMap;
 use std::fmt;
+use std::ptr::NonNull;
+use std::sync::Mutex;
 use std::{mem, slice};
 
 lazy_static! {
@@ -29,6 +32,96 @@ lazy_static! {
             mem::transmute(ptr)
         }
     };
+
+    static ref TRAMPOLINES: TrampBuffer = TrampBuffer::new(64 * 1048576);
+}
+
+/// The global trampoline buffer.
+struct TrampBuffer {
+    /// A fixed-(virtual)-size executable+writable buffer for storing trampolines.
+    buffer: CodeMemory,
+
+    /// Allocation state.
+    alloc: Mutex<AllocState>,
+}
+
+/// The allocation state of a `TrampBuffer`.
+struct AllocState {
+    /// Records all allocated blocks in `buffer`.
+    ///
+    /// Maps the start address of each block to its end address.
+    blocks: BTreeMap<usize, usize>,
+}
+
+impl TrampBuffer {
+    /// Creates a trampoline buffer with a given (virtual) size.
+    fn new(size: usize) -> TrampBuffer {
+        let mem = CodeMemory::new(size);
+        mem.make_writable_executable();
+        TrampBuffer {
+            buffer: mem,
+            alloc: Mutex::new(AllocState {
+                blocks: BTreeMap::new(),
+            }),
+        }
+    }
+
+    /// Removes a previously-`insert`ed trampoline.
+    ///
+    /// For safety, refer to the public interface `TrampolineBufferBuilder::remove_global`.
+    unsafe fn remove(&self, start: NonNull<u8>) {
+        let start = start.as_ptr() as usize - self.buffer.get_backing_ptr() as usize;
+        let mut alloc = self.alloc.lock().unwrap();
+        alloc
+            .blocks
+            .remove(&start)
+            .expect("TrampBuffer::remove(): Attempting to remove a non-existent allocation.");
+    }
+
+    /// Allocates a region of executable memory and copies `buf` to the end of this region.
+    ///
+    /// Returns `None` if no memory is available.
+    fn insert(&self, buf: &[u8]) -> Option<NonNull<u8>> {
+        // First, assume an available start position...
+        let mut assumed_start: usize = 0;
+
+        let mut alloc = self.alloc.lock().unwrap();
+        let mut found = false;
+
+        // Then, try invalidating that assumption...
+        for (&start, &end) in &alloc.blocks {
+            if start - assumed_start < buf.len() {
+                // Unavailable. Move to next free block.
+                assumed_start = end;
+            } else {
+                // This free block can be used.
+                found = true;
+                break;
+            }
+        }
+
+        if !found {
+            // No previous free blocks were found. Try allocating at the end.
+            if self.buffer.len() - assumed_start < buf.len() {
+                // No more free space. Cannot allocate.
+                return None;
+            }
+        }
+
+        // Now we know `assumed_start` is valid.
+        let start = assumed_start;
+        alloc.blocks.insert(start, start + buf.len());
+
+        // We have unique ownership to `self.buffer[start..start + buf.len()]`.
+        let slice = unsafe {
+            std::slice::from_raw_parts_mut(
+                self.buffer.get_backing_ptr().offset(start as _),
+                buf.len(),
+            )
+        };
+        slice.copy_from_slice(buf);
+        Some(NonNull::new(slice.as_mut_ptr()).unwrap())
+    }
 }
 
 /// An opaque type for pointers to a callable memory location.
@@ -219,6 +312,27 @@ impl TrampolineBufferBuilder {
         idx
     }
 
+    /// Inserts this trampoline to the global trampoline buffer.
+    pub fn insert_global(self) -> Option<NonNull<u8>> {
+        TRAMPOLINES.insert(&self.code)
+    }
+
+    /// Removes the trampoline pointed to by `ptr` from the global trampoline buffer. Panics if `ptr`
+    /// does not point to any trampoline.
+    ///
+    /// # Safety
+    ///
+    /// Calling this function invalidates the trampoline `ptr` points to and recycles its memory. You
+    /// should ensure that `ptr` isn't used after calling `remove_global`.
+    pub unsafe fn remove_global(ptr: NonNull<u8>) {
+        TRAMPOLINES.remove(ptr);
+    }
+
+    /// Gets the current (non-executable) code in this builder.
+    pub fn code(&self) -> &[u8] {
+        &self.code
+    }
+
     /// Consumes the builder and builds the trampoline buffer.
     pub fn build(self) -> TrampolineBuffer {
         get_context(); // ensure lazy initialization is completed
@@ -292,4 +406,85 @@ mod tests {
         };
         assert_eq!(ret, 136);
     }
+
+    #[test]
+    fn test_many_global_trampolines() {
+        unsafe extern "C" fn inner(n: *const CallContext, args: *const u64) -> u64 {
+            let n = n as usize;
+            let mut result: u64 = 0;
+            for i in 0..n {
+                result += *args.offset(i as _);
+            }
+            result
+        }
+
+        // Use the smallest possible buffer size (page size) to check memory releasing logic.
+        let buffer = TrampBuffer::new(4096);
+
+        // Validate the previous trampoline instead of the current one to ensure that no overwrite happened.
+        let mut prev: Option<(NonNull<u8>, u64)> = None;
+
+        for i in 0..5000usize {
+            let mut builder = TrampolineBufferBuilder::new();
+            let n = i % 8;
+            builder.add_callinfo_trampoline(inner, n as _, n as _);
+            let ptr = buffer
+                .insert(builder.code())
+                .expect("cannot insert new code into global buffer");
+
+            if let Some((ptr, expected)) = prev.take() {
+                use std::mem::transmute;
+
+                // Test different argument counts.
+                unsafe {
+                    match expected {
+                        0 => {
+                            let f = transmute::<_, extern "C" fn() -> u64>(ptr);
+                            assert_eq!(f(), 0);
+                        }
+                        1 => {
+                            let f = transmute::<_, extern "C" fn(u64) -> u64>(ptr);
+                            assert_eq!(f(1), 1);
+                        }
+                        3 => {
+                            let f = transmute::<_, extern "C" fn(u64, u64) -> u64>(ptr);
+                            assert_eq!(f(1, 2), 3);
+                        }
+                        6 => {
+                            let f = transmute::<_, extern "C" fn(u64, u64, u64) -> u64>(ptr);
+                            assert_eq!(f(1, 2, 3), 6);
+                        }
+                        10 => {
+                            let f = transmute::<_, extern "C" fn(u64, u64, u64, u64) -> u64>(ptr);
+                            assert_eq!(f(1, 2, 3, 4), 10);
+                        }
+                        15 => {
+                            let f =
+                                transmute::<_, extern "C" fn(u64, u64, u64, u64, u64) -> u64>(ptr);
+                            assert_eq!(f(1, 2, 3, 4, 5), 15);
+                        }
+                        21 => {
+                            let f = transmute::<
+                                _,
+                                extern "C" fn(u64, u64, u64, u64, u64, u64) -> u64,
+                            >(ptr);
+                            assert_eq!(f(1, 2, 3, 4, 5, 6), 21);
+                        }
+                        28 => {
+                            let f = transmute::<
+                                _,
+                                extern "C" fn(u64, u64, u64, u64, u64, u64, u64) -> u64,
+                            >(ptr);
+                            assert_eq!(f(1, 2, 3, 4, 5, 6, 7), 28);
+                        }
+                        _ => unreachable!(),
+                    }
+                    buffer.remove(ptr);
+                }
+            }
+
+            let expected = (0..=n as u64).sum();
+            prev = Some((ptr, expected))
+        }
+    }
 }