Enable TSAN with FULL4G and T2C support

ThreadSanitizer (TSAN) can now detect race conditions across the entire
multi-threaded JIT pipeline with full 4GB address space emulation. This
enables testing of the tier-2 LLVM compilation thread while maintaining
production memory layout.

Memory Layout (TSAN-compatible):
- Main memory: MAP_FIXED at 0x7d0000000000 (4GB)
- JIT buffer: MAP_FIXED at 0x7d1000000000
- Both allocations within TSAN app range (0x7cf-0x7ff trillion)
- Prevents conflicts with TSAN shadow memory (0x02a-0x7ce trillion)

ASLR Mitigation:
- Added setarch -R wrapper for TSAN test execution
- Disables ASLR to prevent random allocations in shadow memory
- Only affects test runs, not production builds

SDL Conflict Resolution:
- SDL (uninstrumented system library) creates threads TSAN cannot track
- Disabled SDL when TSAN enabled to focus on built-in race detection
- Production builds still fully support SDL

Author: jserv

Makefile

@@ -80,6 +80,28 @@ endif
 ENABLE_ARCH_TEST ?= 0
 $(call set-feature, ARCH_TEST)
 
+# ThreadSanitizer support
+# TSAN on x86-64 memory layout:
+#   Shadow: 0x02a000000000 - 0x7cefffffffff (reserved by TSAN)
+#   App:    0x7cf000000000 - 0x7ffffffff000 (usable by application)
+#
+# We use MAP_FIXED to allocate FULL4G's 4GB memory at a fixed address
+# (0x7d0000000000) within TSAN's app range, ensuring compatibility.
+#
+# IMPORTANT: TSAN requires ASLR (Address Space Layout Randomization) to be
+# disabled to prevent system allocations from landing in TSAN's shadow memory.
+# Tests are run with 'setarch $(uname -m) -R' to disable ASLR.
+ENABLE_TSAN ?= 0
+ifeq ("$(ENABLE_TSAN)", "1")
+override ENABLE_SDL := 0       # SDL (uninstrumented system lib) creates threads TSAN cannot track
+override ENABLE_LTO := 0       # LTO interferes with TSAN instrumentation
+CFLAGS += -DTSAN_ENABLED       # Signal code to use TSAN-compatible allocations
+# Disable ASLR for TSAN tests to prevent allocations in TSAN shadow memory
+BIN_WRAPPER = setarch $(shell uname -m) -R
+else
+BIN_WRAPPER =
+endif
+
 # Enable link-time optimization (LTO)
 ENABLE_LTO ?= 1
 ifeq ($(call has, LTO), 1)
@@ -332,6 +354,12 @@ CFLAGS += -fsanitize=undefined -fno-sanitize=alignment -fno-sanitize-recover=all
 LDFLAGS += -fsanitize=undefined -fno-sanitize=alignment -fno-sanitize-recover=all
 endif
 
+# ThreadSanitizer flags (ENABLE_TSAN is set earlier to override SDL/FULL4G)
+ifeq ("$(ENABLE_TSAN)", "1")
+CFLAGS += -fsanitize=thread -g
+LDFLAGS += -fsanitize=thread
+endif
+
 $(OUT)/emulate.o: CFLAGS += -foptimize-sibling-calls -fomit-frame-pointer -fno-stack-check -fno-stack-protector
 
 # .DEFAULT_GOAL should be set to all since the very first target is not all
@@ -445,7 +473,7 @@ define check-test
 $(Q)true; \
 $(PRINTF) "Running $(3) ... "; \
 OUTPUT_FILE="$$(mktemp)"; \
-if (LC_ALL=C $(BIN) $(1) $(2) > "$$OUTPUT_FILE") && \
+if (LC_ALL=C $(BIN_WRAPPER) $(BIN) $(1) $(2) > "$$OUTPUT_FILE") && \
    [ "$$(cat "$$OUTPUT_FILE" | $(LOG_FILTER) | $(4))" = "$(5)" ]; then \
     $(call notice, [OK]); \
 else \

src/emulate.c

@@ -304,6 +304,7 @@ static block_t *block_alloc(riscv_t *rv)
     block->hot2 = false;
     block->has_loops = false;
     block->n_invoke = 0;
+    block->func = NULL;
     INIT_LIST_HEAD(&block->list);
 #if RV32_HAS(T2C)
     block->compiled = false;
@@ -1176,22 +1177,32 @@ void rv_step(void *arg)
 #if RV32_HAS(JIT)
 #if RV32_HAS(T2C)
         /* executed through the tier-2 JIT compiler */
-        if (block->hot2) {
+        /* Use acquire semantics to ensure we see func write before using it */
+        if (__atomic_load_n(&block->hot2, __ATOMIC_ACQUIRE)) {
             ((exec_t2c_func_t) block->func)(rv);
             prev = NULL;
             continue;
         } /* check if invoking times of t1 generated code exceed threshold */
-        else if (!block->compiled && block->n_invoke >= THRESHOLD) {
-            block->compiled = true;
+        else if (!__atomic_load_n(&block->compiled, __ATOMIC_RELAXED) &&
+                 __atomic_load_n(&block->n_invoke, __ATOMIC_RELAXED) >=
+                     THRESHOLD) {
+            __atomic_store_n(&block->compiled, true, __ATOMIC_RELAXED);
             queue_entry_t *entry = malloc(sizeof(queue_entry_t));
             if (unlikely(!entry)) {
                 /* Malloc failed - reset compiled flag to allow retry later */
-                block->compiled = false;
+                __atomic_store_n(&block->compiled, false, __ATOMIC_RELAXED);
                 continue;
             }
-            entry->block = block;
+            /* Store cache key instead of pointer to prevent use-after-free */
+#if RV32_HAS(SYSTEM)
+            entry->key =
+                (uint64_t) block->pc_start | ((uint64_t) block->satp << 32);
+#else
+            entry->key = (uint64_t) block->pc_start;
+#endif
             pthread_mutex_lock(&rv->wait_queue_lock);
             list_add(&entry->list, &rv->wait_queue);
+            pthread_cond_signal(&rv->wait_queue_cond);
             pthread_mutex_unlock(&rv->wait_queue_lock);
         }
 #endif
@@ -1203,7 +1214,11 @@ void rv_step(void *arg)
          *       entry in compiled binary buffer.
          */
         if (block->hot) {
+#if RV32_HAS(T2C)
+            __atomic_fetch_add(&block->n_invoke, 1, __ATOMIC_RELAXED);
+#else
             block->n_invoke++;
+#endif
             ((exec_block_func_t) state->buf)(
                 rv, (uintptr_t) (state->buf + block->offset));
             prev = NULL;

src/io.c

@@ -27,12 +27,33 @@ memory_t *memory_new(uint32_t size)
         return NULL;
     assert(mem);
 #if HAVE_MMAP
+#if defined(TSAN_ENABLED) && defined(__x86_64__)
+    /* ThreadSanitizer compatibility: Use MAP_FIXED to allocate at a specific
+     * address within TSAN's app range (0x7cf000000000 - 0x7ffffffff000).
+     *
+     * Fixed address: 0x7d0000000000
+     * Size: up to 4GB (0x100000000)
+     * End: 0x7d0100000000 (well within app range)
+     *
+     * This guarantees the allocation won't land in TSAN's shadow memory,
+     * preventing "unexpected memory mapping" errors.
+     */
+    void *fixed_addr = (void *) 0x7d0000000000UL;
+    data_memory_base = mmap(fixed_addr, size, PROT_READ | PROT_WRITE,
+                            MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1, 0);
+    if (data_memory_base == MAP_FAILED) {
+        free(mem);
+        return NULL;
+    }
+#else
+    /* Standard allocation without TSAN */
     data_memory_base = mmap(NULL, size, PROT_READ | PROT_WRITE,
                             MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
     if (data_memory_base == MAP_FAILED) {
         free(mem);
         return NULL;
     }
+#endif
 #else
     data_memory_base = malloc(size);
     if (!data_memory_base) {

src/jit.c

@@ -2336,6 +2336,25 @@ struct jit_state *jit_state_init(size_t size)
 
     state->offset = 0;
     state->size = size;
+#if defined(TSAN_ENABLED) && defined(__x86_64__)
+    /* ThreadSanitizer compatibility: Allocate JIT code buffer at a fixed
+     * address above the main memory region to avoid conflicts.
+     *
+     * Main memory: 0x7d0000000000 - 0x7d0100000000 (4GB for FULL4G)
+     * JIT buffer:  0x7d1000000000 + size
+     *
+     * This keeps both allocations in TSAN's app range (0x7cf000000000 -
+     * 0x7ffffffff000) and prevents overlap with main memory or TSAN shadow.
+     */
+    void *jit_addr = (void *) 0x7d1000000000UL;
+    state->buf = mmap(jit_addr, size, PROT_READ | PROT_WRITE | PROT_EXEC,
+                      MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1, 0);
+    if (state->buf == MAP_FAILED) {
+        free(state);
+        return NULL;
+    }
+#else
+    /* Standard allocation without TSAN */
     state->buf = mmap(0, size, PROT_READ | PROT_WRITE | PROT_EXEC,
                       MAP_PRIVATE | MAP_ANONYMOUS
 #if defined(__APPLE__)
@@ -2347,8 +2366,7 @@ struct jit_state *jit_state_init(size_t size)
         free(state);
         return NULL;
     }
-    assert(state->buf != MAP_FAILED);
-
+#endif
     state->n_blocks = 0;
     set_reset(&state->set);
     reset_reg();

src/main.c

@@ -19,6 +19,28 @@
 #include "riscv.h"
 #include "utils.h"
 
+/* ThreadSanitizer configuration for FULL4G compatibility
+ *
+ * We use MAP_FIXED to allocate emulated memory at 0x7d0000000000, which is
+ * within TSAN's application memory range (0x7cf000000000 - 0x7ffffffff000).
+ * This avoids conflicts with TSAN's shadow memory and allows race detection
+ * to work with FULL4G's 4GB address space.
+ *
+ * Configuration optimizes for race detection with minimal overhead.
+ */
+#if defined(__SANITIZE_THREAD__)
+const char *__tsan_default_options()
+{
+    return "halt_on_error=0"          /* Continue after errors */
+           ":report_bugs=1"           /* Report data races */
+           ":second_deadlock_stack=1" /* Full deadlock info */
+           ":verbosity=0"             /* Reduce noise */
+           ":memory_limit_mb=0"       /* No memory limit */
+           ":history_size=7"          /* Larger race detection window */
+           ":io_sync=0";              /* Don't sync on I/O */
+}
+#endif
+
 /* enable program trace mode */
 #if !RV32_HAS(SYSTEM) || (RV32_HAS(SYSTEM) && RV32_HAS(ELF_LOADER))
 static bool opt_trace = false;

src/riscv.c

@@ -206,19 +206,41 @@ static pthread_t t2c_thread;
 static void *t2c_runloop(void *arg)
 {
     riscv_t *rv = (riscv_t *) arg;
+    pthread_mutex_lock(&rv->wait_queue_lock);
     while (!rv->quit) {
-        if (!list_empty(&rv->wait_queue)) {
-            queue_entry_t *entry =
-                list_last_entry(&rv->wait_queue, queue_entry_t, list);
-            pthread_mutex_lock(&rv->wait_queue_lock);
-            list_del_init(&entry->list);
-            pthread_mutex_unlock(&rv->wait_queue_lock);
-            pthread_mutex_lock(&rv->cache_lock);
-            t2c_compile(rv, entry->block);
-            pthread_mutex_unlock(&rv->cache_lock);
-            free(entry);
-        }
+        /* Wait for work or quit signal */
+        while (list_empty(&rv->wait_queue) && !rv->quit)
+            pthread_cond_wait(&rv->wait_queue_cond, &rv->wait_queue_lock);
+
+        if (rv->quit)
+            break;
+
+        /* Extract work item while holding the lock */
+        queue_entry_t *entry =
+            list_last_entry(&rv->wait_queue, queue_entry_t, list);
+        list_del_init(&entry->list);
+        pthread_mutex_unlock(&rv->wait_queue_lock);
+
+        /* Perform compilation with cache lock */
+        pthread_mutex_lock(&rv->cache_lock);
+        /* Look up block from cache using the key (might have been evicted) */
+        uint32_t pc = (uint32_t) entry->key;
+        block_t *block = (block_t *) cache_get(rv->block_cache, pc, false);
+#if RV32_HAS(SYSTEM)
+        /* Verify SATP matches (for system mode) */
+        uint32_t satp = (uint32_t) (entry->key >> 32);
+        if (block && block->satp != satp)
+            block = NULL;
+#endif
+        /* Compile only if block still exists in cache */
+        if (block)
+            t2c_compile(rv, block);
+        pthread_mutex_unlock(&rv->cache_lock);
+        free(entry);
+
+        pthread_mutex_lock(&rv->wait_queue_lock);
     }
+    pthread_mutex_unlock(&rv->wait_queue_lock);
     return NULL;
 }
 #endif
@@ -777,6 +799,7 @@ riscv_t *rv_create(riscv_user_t rv_attr)
     /* prepare wait queue. */
     pthread_mutex_init(&rv->wait_queue_lock, NULL);
     pthread_mutex_init(&rv->cache_lock, NULL);
+    pthread_cond_init(&rv->wait_queue_cond, NULL);
     INIT_LIST_HEAD(&rv->wait_queue);
     /* activate the background compilation thread. */
     pthread_create(&t2c_thread, NULL, t2c_runloop, rv);
@@ -910,10 +933,24 @@ void rv_delete(riscv_t *rv)
     block_map_destroy(rv);
 #else
 #if RV32_HAS(T2C)
+    /* Signal the thread to quit */
+    pthread_mutex_lock(&rv->wait_queue_lock);
     rv->quit = true;
+    pthread_cond_signal(&rv->wait_queue_cond);
+    pthread_mutex_unlock(&rv->wait_queue_lock);
+
     pthread_join(t2c_thread, NULL);
+
+    /* Clean up any remaining entries in wait queue */
+    queue_entry_t *entry, *safe;
+    list_for_each_entry_safe (entry, safe, &rv->wait_queue, list) {
+        list_del(&entry->list);
+        free(entry);
+    }
+
     pthread_mutex_destroy(&rv->wait_queue_lock);
     pthread_mutex_destroy(&rv->cache_lock);
+    pthread_cond_destroy(&rv->wait_queue_cond);
     jit_cache_exit(rv->jit_cache);
 #endif
     jit_state_exit(rv->jit_state);

src/riscv_private.h

@@ -105,7 +105,7 @@ typedef struct block {
 
 #if RV32_HAS(JIT) && RV32_HAS(T2C)
 typedef struct {
-    block_t *block;
+    uint64_t key; /**< cache key (PC or PC|SATP) to look up block */
     struct list_head list;
 } queue_entry_t;
 #endif
@@ -197,6 +197,7 @@ struct riscv_internal {
 #if RV32_HAS(T2C)
     struct list_head wait_queue;
     pthread_mutex_t wait_queue_lock, cache_lock;
+    pthread_cond_t wait_queue_cond;
     volatile bool quit; /**< Determine the main thread is terminated or not */
 #endif
     void *jit_state;

src/t2c.c

@@ -346,7 +346,9 @@ void t2c_compile(riscv_t *rv, block_t *block)
 
     jit_cache_update(rv->jit_cache, key, block->func);
 
-    block->hot2 = true;
+    /* Use release semantics to ensure func write is visible before hot2 is set
+     */
+    __atomic_store_n(&block->hot2, true, __ATOMIC_RELEASE);
 }
 
 struct jit_cache *jit_cache_init()