Use `mmap` to map in task stacks and guard page

Also implement caching of stacks.

3 files changed, 103 insertions, 34 deletions

diff --git a/src/libgreen/context.rs b/src/libgreen/context.rs
index 8530e3e837e..3f2f74bbb8d 100644
--- a/src/libgreen/context.rs
+++ b/src/libgreen/context.rs
@@ -12,10 +12,9 @@ use std::libc::c_void;
 use std::uint;
 use std::cast::{transmute, transmute_mut_unsafe,
                 transmute_region, transmute_mut_region};
+use stack::Stack;
 use std::unstable::stack;
 
-use stack::StackSegment;
-
 // FIXME #7761: Registers is boxed so that it is 16-byte aligned, for storing
 // SSE regs.  It would be marginally better not to do this. In C++ we
 // use an attribute on a struct.
@@ -41,7 +40,7 @@ impl Context {
     }
 
     /// Create a new context that will resume execution by running proc()
-    pub fn new(start: proc(), stack: &mut StackSegment) -> Context {
+    pub fn new(start: proc(), stack: &mut Stack) -> Context {
         // The C-ABI function that is the task entry point
         //
         // Note that this function is a little sketchy. We're taking a
@@ -79,6 +78,7 @@ impl Context {
         // be passed to the spawn function.  Another unfortunate
         // allocation
         let start = ~start;
+
         initialize_call_frame(&mut *regs,
                               task_start_wrapper as *c_void,
                               unsafe { transmute(&*start) },
diff --git a/src/libgreen/coroutine.rs b/src/libgreen/coroutine.rs
index 7bc5d0accfe..3d7dc58a1b2 100644
--- a/src/libgreen/coroutine.rs
+++ b/src/libgreen/coroutine.rs
@@ -14,7 +14,7 @@
 use std::rt::env;
 
 use context::Context;
-use stack::{StackPool, StackSegment};
+use stack::{StackPool, Stack};
 
 /// A coroutine is nothing more than a (register context, stack) pair.
 pub struct Coroutine {
@@ -24,7 +24,7 @@ pub struct Coroutine {
     ///
     /// Servo needs this to be public in order to tell SpiderMonkey
     /// about the stack bounds.
-    current_stack_segment: StackSegment,
+    current_stack_segment: Stack,
 
     /// Always valid if the task is alive and not running.
     saved_context: Context
@@ -39,7 +39,7 @@ impl Coroutine {
             Some(size) => size,
             None => env::min_stack()
         };
-        let mut stack = stack_pool.take_segment(stack_size);
+        let mut stack = stack_pool.take_stack(stack_size);
         let initial_context = Context::new(start, &mut stack);
         Coroutine {
             current_stack_segment: stack,
@@ -49,7 +49,7 @@ impl Coroutine {
 
     pub fn empty() -> Coroutine {
         Coroutine {
-            current_stack_segment: StackSegment::new(0),
+            current_stack_segment: Stack::new(0),
             saved_context: Context::empty()
         }
     }
@@ -57,6 +57,6 @@ impl Coroutine {
     /// Destroy coroutine and try to reuse std::stack segment.
     pub fn recycle(self, stack_pool: &mut StackPool) {
         let Coroutine { current_stack_segment, .. } = self;
-        stack_pool.give_segment(current_stack_segment);
+        stack_pool.give_stack(current_stack_segment);
     }
 }
diff --git a/src/libgreen/stack.rs b/src/libgreen/stack.rs
index 7e6dd02dd67..a5d5174b91b 100644
--- a/src/libgreen/stack.rs
+++ b/src/libgreen/stack.rs
@@ -8,46 +8,101 @@
 // option. This file may not be copied, modified, or distributed
 // except according to those terms.
 
-use std::vec;
-use std::libc::{c_uint, uintptr_t};
+use std::rt::env::max_cached_stacks;
+use std::os::{errno, page_size, MemoryMap, MapReadable, MapWritable, MapNonStandardFlags};
+#[cfg(not(windows))]
+use std::libc::{MAP_STACK, MAP_PRIVATE, MAP_ANON};
+use std::libc::{c_uint, c_int, c_void, uintptr_t};
 
-pub struct StackSegment {
-    priv buf: ~[u8],
-    priv valgrind_id: c_uint
+/// A task's stack. The name "Stack" is a vestige of segmented stacks.
+pub struct Stack {
+    priv buf: MemoryMap,
+    priv min_size: uint,
+    priv valgrind_id: c_uint,
 }
 
-impl StackSegment {
-    pub fn new(size: uint) -> StackSegment {
-        unsafe {
-            // Crate a block of uninitialized values
-            let mut stack = vec::with_capacity(size);
-            stack.set_len(size);
+// Try to use MAP_STACK on platforms that support it (it's what we're doing
+// anyway), but some platforms don't support it at all. For example, it appears
+// that there's a bug in freebsd that MAP_STACK implies MAP_FIXED (so it always
+// fails): http://lists.freebsd.org/pipermail/freebsd-bugs/2011-July/044840.html
+#[cfg(not(windows), not(target_os = "freebsd"))]
+static STACK_FLAGS: c_int = MAP_STACK | MAP_PRIVATE | MAP_ANON;
+#[cfg(target_os = "freebsd")]
+static STACK_FLAGS: c_int = MAP_PRIVATE | MAP_ANON;
+#[cfg(windows)]
+static STACK_FLAGS: c_int = 0;
 
-            let mut stk = StackSegment {
-                buf: stack,
-                valgrind_id: 0
-            };
+impl Stack {
+    pub fn new(size: uint) -> Stack {
+        // Map in a stack. Eventually we might be able to handle stack allocation failure, which
+        // would fail to spawn the task. But there's not many sensible things to do on OOM.
+        // Failure seems fine (and is what the old stack allocation did).
+        let stack = match MemoryMap::new(size, [MapReadable, MapWritable,
+                                         MapNonStandardFlags(STACK_FLAGS)]) {
+            Ok(map) => map,
+            Err(e) => fail!("Creating memory map for stack of size {} failed: {}", size, e)
+        };
 
-            // XXX: Using the FFI to call a C macro. Slow
-            stk.valgrind_id = rust_valgrind_stack_register(stk.start(), stk.end());
-            return stk;
+        // Change the last page to be inaccessible. This is to provide safety; when an FFI
+        // function overflows it will (hopefully) hit this guard page. It isn't guaranteed, but
+        // that's why FFI is unsafe. buf.data is guaranteed to be aligned properly.
+        if !protect_last_page(&stack) {
+            fail!("Could not memory-protect guard page. stack={:?}, errno={}",
+                  stack, errno());
         }
+
+        let mut stk = Stack {
+            buf: stack,
+            min_size: size,
+            valgrind_id: 0
+        };
+
+        // XXX: Using the FFI to call a C macro. Slow
+        stk.valgrind_id = unsafe { rust_valgrind_stack_register(stk.start(), stk.end()) };
+        return stk;
     }
 
     /// Point to the low end of the allocated stack
     pub fn start(&self) -> *uint {
-        self.buf.as_ptr() as *uint
+        self.buf.data as *uint
     }
 
     /// Point one word beyond the high end of the allocated stack
     pub fn end(&self) -> *uint {
         unsafe {
-            self.buf.as_ptr().offset(self.buf.len() as int) as *uint
+            self.buf.data.offset(self.buf.len as int) as *uint
         }
     }
 }
 
-impl Drop for StackSegment {
+// These use ToPrimitive so that we never need to worry about the sizes of whatever types these
+// (which we would with scalar casts). It's either a wrapper for a scalar cast or failure: fast, or
+// will fail during compilation.
+#[cfg(unix)]
+fn protect_last_page(stack: &MemoryMap) -> bool {
+    use std::libc::{mprotect, PROT_NONE, size_t};
+    unsafe {
+        // This may seem backwards: the start of the segment is the last page? Yes! The stack grows
+        // from higher addresses (the end of the allocated block) to lower addresses (the start of
+        // the allocated block).
+        let last_page = stack.data as *c_void;
+        mprotect(last_page, page_size() as size_t, PROT_NONE) != -1
+    }
+}
+
+#[cfg(windows)]
+fn protect_last_page(stack: &MemoryMap) -> bool {
+    use std::libc::{VirtualProtect, PAGE_NOACCESS, SIZE_T, LPDWORD, DWORD};
+    unsafe {
+        // see above
+        let last_page = stack.data as *mut c_void;
+        let mut old_prot: DWORD = 0;
+        VirtualProtect(last_page, page_size() as SIZE_T, PAGE_NOACCESS,
+                       &mut old_prot as LPDWORD) != 0
+    }
+}
+
+impl Drop for Stack {
     fn drop(&mut self) {
         unsafe {
             // XXX: Using the FFI to call a C macro. Slow
@@ -56,16 +111,30 @@ impl Drop for StackSegment {
     }
 }
 
-pub struct StackPool(());
+pub struct StackPool {
+    // Ideally this would be some datastructure that preserved ordering on Stack.min_size.
+    priv stacks: ~[Stack],
+}
 
 impl StackPool {
-    pub fn new() -> StackPool { StackPool(()) }
+    pub fn new() -> StackPool {
+        StackPool {
+            stacks: ~[],
+        }
+    }
 
-    pub fn take_segment(&self, min_size: uint) -> StackSegment {
-        StackSegment::new(min_size)
+    pub fn take_stack(&mut self, min_size: uint) -> Stack {
+        // Ideally this would be a binary search
+        match self.stacks.iter().position(|s| s.min_size < min_size) {
+            Some(idx) => self.stacks.swap_remove(idx),
+            None      => Stack::new(min_size)
+        }
     }
 
-    pub fn give_segment(&self, _stack: StackSegment) {
+    pub fn give_stack(&mut self, stack: Stack) {
+        if self.stacks.len() <= max_cached_stacks() {
+            self.stacks.push(stack)
+        }
     }
 }