From: Jack Lange <jacklange@cs.pitt.edu>
Date: Tue, 20 Sep 2011 19:40:31 +0000 (-0400)
Subject: fix barrier hang w/ unitialized secondary cores
X-Git-Url: http://v3vee.org/palacios/gitweb/gitweb.cgi?p=palacios.git;a=commitdiff_plain;h=94429f4b9268cf4d5f86dee5c3d929110811ac80

fix barrier hang w/ unitialized secondary cores
---

diff --git a/palacios/src/palacios/svm.c b/palacios/src/palacios/svm.c
index a52b4d0..115e19c 100644
--- a/palacios/src/palacios/svm.c
+++ b/palacios/src/palacios/svm.c
@@ -700,6 +700,9 @@ int v3_start_svm_guest(struct guest_info * info) {
 	}
 
 	PrintDebug("SVM core %u(on %u) initialized\n", info->vcpu_id, info->pcpu_id);
+
+	// We'll be paranoid about race conditions here
+	v3_wait_at_barrier(info);
     } 
 
     PrintDebug("SVM core %u(on %u): I am starting at CS=0x%x (base=0x%p, limit=0x%x),  RIP=0x%p\n", 
diff --git a/palacios/src/palacios/vmm_barrier.c b/palacios/src/palacios/vmm_barrier.c
index ab0aae4..7718fc8 100644
--- a/palacios/src/palacios/vmm_barrier.c
+++ b/palacios/src/palacios/vmm_barrier.c
@@ -104,6 +104,14 @@ int v3_raise_barrier(struct v3_vm_info * vm_info, struct guest_info * local_core
 	all_blocked = 1;
 
 	for (i = 0; i < vm_info->num_cores; i++) {
+	    
+	    // Tricky: If a core is not running then it is safe to ignore it. 
+	    // Whenever we transition a core to the RUNNING state we MUST immediately wait on the barrier. 
+	    // TODO: Wrap the state transitions in functions that do this automatically
+	    if (vm_info->cores[i].core_run_state != CORE_RUNNING) {
+		continue;
+	    }
+
 	    if (v3_bitmap_check(&(barrier->cpu_map), i) == 0) {
 		// There is still a core that is not waiting at the barrier
 		all_blocked = 0;
@@ -157,6 +165,8 @@ int v3_wait_at_barrier(struct guest_info * core) {
 	return 0;
     }
 
+    V3_Print("Core %d waiting at barrier\n", core->vcpu_id);
+
     /*  Barrier has been activated. 
      *  Wait here until it's lowered
      */
@@ -164,6 +174,7 @@ int v3_wait_at_barrier(struct guest_info * core) {
     
     // set cpu bit in barrier bitmap
     v3_bitmap_set(&(barrier->cpu_map), core->vcpu_id);
+    V3_Print("Core %d bit set as waiting\n", core->vcpu_id);
 
     // wait for cpu bit to clear
     while (v3_bitmap_check(&(barrier->cpu_map), core->vcpu_id) == 1) {
diff --git a/palacios/src/palacios/vmx.c b/palacios/src/palacios/vmx.c
index c7c66c5..e733eb5 100644
--- a/palacios/src/palacios/vmx.c
+++ b/palacios/src/palacios/vmx.c
@@ -955,6 +955,9 @@ int v3_start_vmx_guest(struct guest_info * info) {
         }
 	
 	PrintDebug("VMX core %u initialized\n", info->vcpu_id);
+
+	// We'll be paranoid about race conditions here
+	v3_wait_at_barrier(info);
     }