Enable per-thread tracking

Cargo.toml

@@ -6,5 +6,5 @@ description = "The Quick And Dirty Allocation Profiling Tool"
 license = "Apache-2.0"
 
 [dependencies]
-spin = "0.4"
+libc = { version = "0.2", default-features = false }
-libc = { version = "0.2", default-features = false }
+spin = "0.4"

src/lib.rs

@@ -1,8 +1,11 @@
 #![no_std]
+#![feature(alloc)]
 
+extern crate alloc;
 extern crate libc;
 extern crate spin;
 
+use alloc::collections::btree_map::BTreeMap;
 use libc::c_void;
 use libc::free;
 use libc::malloc;
@@ -10,18 +13,23 @@ use core::alloc::Layout;
 use core::alloc::GlobalAlloc;
 use core::sync::atomic::AtomicBool;
 use core::sync::atomic::Ordering;
+use spin::RwLock;
 
 mod const_init;
 use const_init::ConstInit;
 
+mod thread_id;
+
+// TODO: Doesn't check for race conditions
 static INTERNAL_ALLOCATION: AtomicBool = AtomicBool::new(false);
 
 pub struct QADAPTInternal {
-    pub has_allocated: AtomicBool
+    pub thread_has_allocated: BTreeMap<usize, AtomicBool>,
+    pub recording_enabled: BTreeMap<usize, AtomicBool>
 }
 
 pub struct QADAPT {
-    internal: spin::Once<QADAPTInternal>
+    internal: spin::Once<RwLock<QADAPTInternal>>
 }
 
 impl ConstInit for QADAPT {
@@ -30,10 +38,23 @@ impl ConstInit for QADAPT {
     };
 }
 
+
+
 unsafe impl GlobalAlloc for QADAPT {
     unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
         if !INTERNAL_ALLOCATION.load(Ordering::SeqCst) {
-            self.internal().has_allocated.store(true, Ordering::SeqCst);
+            let tid = thread_id::get();
+
+            // Need to use RAII guard because record_allocation() needs write access
+            let should_record = {
+                let internal = self.internal().read();
+                internal.recording_enabled.contains_key(&tid)
+                    && internal.recording_enabled.get(&tid).unwrap().load(Ordering::SeqCst)
+            };
+
+            if should_record {
+                self.record_allocation(thread_id::get())
+            }
         }
 
         malloc(layout.size()) as *mut u8
@@ -47,24 +68,70 @@ unsafe impl GlobalAlloc for QADAPT {
 impl QADAPT {
     pub const INIT: Self = <Self as ConstInit>::INIT;
 
-    fn internal(&self) -> &QADAPTInternal {
+    fn internal(&self) -> &RwLock<QADAPTInternal> {
 
         self.internal.call_once(|| {
             INTERNAL_ALLOCATION.store(true, Ordering::SeqCst);
             let q = QADAPTInternal {
-                has_allocated: AtomicBool::new(false)
+                thread_has_allocated: BTreeMap::new(),
+                recording_enabled: BTreeMap::new()
             };
             INTERNAL_ALLOCATION.store(false, Ordering::SeqCst);
 
-            q
+            RwLock::new(q)
         })
     }
 
-    pub fn clear_allocations(&self) {
+    pub fn reset_allocation_state(&self) {
-        self.internal().has_allocated.store(false, Ordering::SeqCst);
+        let internal = self.internal().write();
+        for (_tid, has_allocated) in &internal.thread_has_allocated {
+
+            has_allocated.store(false, Ordering::SeqCst);
+        }
+        for (_tid, recording_enabled) in &internal.recording_enabled {
+
+            recording_enabled.store(false, Ordering::SeqCst);
+        }
     }
 
-    pub fn has_allocated(&self) -> bool {
+    pub fn has_allocated_current(&self) -> bool {
-        self.internal().has_allocated.load(Ordering::SeqCst)
+        let tid = thread_id::get();
+        let internal = self.internal().read();
+
+        // UNWRAP: Already checked for existence
+        internal.thread_has_allocated.contains_key(&tid)
+            && internal.thread_has_allocated.get(&tid).unwrap().load(Ordering::SeqCst)
+    }
+
+    pub fn record_allocation(&self, thread_id: usize) {
+        let mut internal = self.internal().write();
+        if internal.thread_has_allocated.contains_key(&thread_id) {
+            // UNWRAP: Already checked for existence
+            internal.thread_has_allocated.get(&thread_id)
+                .unwrap().store(true, Ordering::SeqCst)
+        }
+        else {
+            INTERNAL_ALLOCATION.store(true, Ordering::SeqCst);
+            internal.thread_has_allocated.insert(thread_id, AtomicBool::new(true));
+            INTERNAL_ALLOCATION.store(false, Ordering::SeqCst);
+        }
+    }
+
+    pub fn enable_recording_current(&self) {
+        self.enable_recording(thread_id::get());
+    }
+
+    pub fn enable_recording(&self, tid: usize) {
+        let mut internal = self.internal().write();
+
+        if internal.recording_enabled.contains_key(&tid) {
+            // UNWRAP: Already checked for existence
+            internal.recording_enabled.get(&tid).unwrap().store(true, Ordering::SeqCst);
+        }
+        else {
+            INTERNAL_ALLOCATION.store(true, Ordering::SeqCst);
+            internal.recording_enabled.insert(tid, AtomicBool::new(true));
+            INTERNAL_ALLOCATION.store(false, Ordering::SeqCst);
+        }
     }
 }

src/thread_id.rs

@@ -0,0 +1,6 @@
+/// Taken from https://crates.io/crates/thread-id and re-purposed to be no-std safe
+use libc;
+
+pub fn get() -> usize {
+    unsafe { libc::pthread_self() as usize }
+}

tests/basic.rs

@@ -3,7 +3,6 @@ extern crate qadapt;
 use qadapt::QADAPT;
 use std::alloc::alloc;
 use std::alloc::Layout;
-use std::sync::atomic::Ordering;
 
 #[global_allocator]
 static A: QADAPT = QADAPT::INIT;
@@ -24,23 +23,24 @@ struct NonEmpty {
 
 #[test]
 fn allocation_flag() {
-    A.clear_allocations();
+    A.reset_allocation_state();
-    assert!(!A.has_allocated());
+    A.enable_recording_current();
+    assert!(!A.has_allocated_current());
 
     let _x = 24;
-    assert!(!A.has_allocated());
+    assert!(!A.has_allocated_current());
 
     let _x = Empty {};
-    assert!(!A.has_allocated());
+    assert!(!A.has_allocated_current());
 
     let _x = NonEmpty {
         _x: 42,
         _y: 84
     };
-    assert!(!A.has_allocated());
+    assert!(!A.has_allocated_current());
 
     let _x = Box::new(42);
-    assert!(A.has_allocated());
+    assert!(A.has_allocated_current());
 }
 
 #[inline(never)]
@@ -48,11 +48,12 @@ fn no_op() {}
 
 #[test]
 fn no_alloc_during_noop() {
-    A.clear_allocations();
+    A.reset_allocation_state();
-    assert!(!A.has_allocated());
+    A.enable_recording_current();
+    assert!(!A.has_allocated_current());
 
     no_op();
-    assert!(!A.has_allocated());
+    assert!(!A.has_allocated_current());
 }
 
 #[inline(never)]
@@ -62,9 +63,19 @@ fn allocates() {
 
 #[test]
 fn alloc_during_func_call() {
-    A.clear_allocations();
+    A.reset_allocation_state();
-    assert!(!A.has_allocated());
+    A.enable_recording_current();
+    assert!(!A.has_allocated_current());
 
     allocates();
-    assert!(A.has_allocated());
+    assert!(A.has_allocated_current());
+}
+
+#[test]
+fn allocates_unrecorded() {
+    A.reset_allocation_state();
+    assert!(!A.has_allocated_current());
+
+    allocates();
+    assert!(!A.has_allocated_current());
 }

tests/initial.rs

@@ -1,20 +1,18 @@
 extern crate qadapt;
 
 use qadapt::QADAPT;
-use std::sync::atomic::Ordering;
 
 #[global_allocator]
 static A: QADAPT = QADAPT::INIT;
 
 #[test]
 fn init() {
-    // Because the Allocator and its internals isn't the only "pre-main" allocation
+    assert!(!A.has_allocated_current());
-    // that happens, when starting up we expect to see that A has in fact allocated
+    A.reset_allocation_state();
-    assert!(A.has_allocated());
+    A.enable_recording_current();
 
-    A.clear_allocations();
+    assert!(!A.has_allocated_current());
-    assert!(!A.has_allocated());
 
     let _x = Box::new(42);
-    assert!(A.has_allocated());
+    assert!(A.has_allocated_current());
 }