Fixes for winkle state.

    The memory profiling tools sometimes encountered a condition where
    the kernel stack was becoming corrupted.  I tracked it down to the
    winkle code storing the winkle-save state at the wrong end of the
    stack.  Moving the winkle-save area to the bottom of the stack,
    which is where I originally intended it to go.

    Also noticed that the task issuing the winkle was in "running"
    state while waiting for the cores to come out of winkle.  Ensure
    that the kernel updates the task state with a non-running status
    while we are waiting for winkle to complete.

Change-Id: I07a56ea6f24cbc09362f9227d81915da5bc9f148
Reviewed-on: http://gfw160.austin.ibm.com:8080/gerrit/1737
Tested-by: Jenkins Server
Reviewed-by: Douglas R. Gilbert <dgilbert@us.ibm.com>
Reviewed-by: Daniel M. Crowell <dcrowell@us.ibm.com>

1 files changed, 9 insertions, 2 deletions

diff --git a/src/kernel/cpumgr.C b/src/kernel/cpumgr.C
index 791d14703..db79a15bf 100644
--- a/src/kernel/cpumgr.C
+++ b/src/kernel/cpumgr.C
@@ -177,8 +177,15 @@ void CpuManager::startCPU(ssize_t i)
         }
         cpu->scheduler = &Singleton<Scheduler>::instance();
         cpu->scheduler_extra = NULL;
-        cpu->kernel_stack =
-            (void*) (((uint64_t)PageManager::allocatePage(4)) + 16320);
+
+        const size_t kernel_page_count = 4;
+        const size_t kernel_page_offset = kernel_page_count * PAGESIZE -
+                                          8 * sizeof(uint64_t);
+        cpu->kernel_stack_bottom = PageManager::allocatePage(kernel_page_count);
+        cpu->kernel_stack = reinterpret_cast<void*>(
+            reinterpret_cast<uintptr_t>(cpu->kernel_stack_bottom) +
+            kernel_page_offset);
+
         cpu->xscom_mutex = (mutex_t)MUTEX_INITIALIZER;
 
         // Create idle task.