Make protect page aligned

thread/thread.cpp

@@ -305,7 +305,8 @@ namespace photon
             assert(state == states::DONE);
             // `buf` and `stack_size` will always store on register
             // when calling deallocating.
-            mprotect(buf, PAGE_SIZE, PROT_READ | PROT_WRITE);
+            char* protect_head = (char*)align_up((uint64_t)buf, PAGE_SIZE);
+            mprotect(protect_head, PAGE_SIZE, PROT_READ | PROT_WRITE);
             photon_thread_dealloc(buf, stack_size);
         }
     };
@@ -935,14 +936,16 @@ R"(
         if (unlikely(!rq.current))
             LOG_ERROR_RETURN(ENOSYS, nullptr, "Photon not initialized in this vCPU (OS thread)");
         size_t randomizer = (rand() % 32) * (1024 + 8);
-        stack_size = align_up(randomizer + stack_size + sizeof(thread) + PAGE_SIZE, PAGE_SIZE);
+        stack_size = align_up(randomizer + stack_size + sizeof(thread), PAGE_SIZE);
+        stack_size += PAGE_SIZE * 2;  // extra 2 pages for alignment and set guard page
         char* ptr = (char*)photon_thread_alloc(stack_size);
-        mprotect(ptr, PAGE_SIZE, PROT_NONE);
-        uint64_t p = (uint64_t) ptr + stack_size - sizeof(thread) - randomizer;
+        char* protect_head = (char*)align_up((uint64_t)ptr, PAGE_SIZE);
+        mprotect(protect_head, PAGE_SIZE, PROT_NONE);
+        uint64_t p = (uint64_t)ptr + stack_size - sizeof(thread) - randomizer;
         p = align_down(p, 64);
-        auto th = new((char*) p) thread;
+        auto th = new ((char*)p) thread;
         th->buf = ptr;
-        th->stackful_alloc_top = ptr + PAGE_SIZE;
+        th->stackful_alloc_top = protect_head + PAGE_SIZE;
         th->start = start;
         th->stack_size = stack_size;
         th->arg = arg;

thread/timer.h

@@ -35,7 +35,7 @@ namespace photon
         // it has a `stack_size`, and the `on_timer` is invoked within the thread's context.
         // The timer object is deleted automatically after it is finished.
         Timer(uint64_t default_timeout, Entry on_timer, bool repeating = true,
-              uint64_t stack_size = 8UL * 1024 * 1024)
+              uint64_t stack_size = DEFAULT_STACK_SIZE)
         {
             _on_timer = on_timer;
             _default_timeout = default_timeout;