minor changes to yields to avoid guest lock ups due to irq flooding scenarios

[palacios.git] / palacios / src / palacios / svm_halt.c

diff --git a/palacios/src/palacios/svm_halt.c b/palacios/src/palacios/svm_halt.c
index 16239e8..2e6f9e2 100644 (file)
--- a/palacios/src/palacios/svm_halt.c
+++ b/palacios/src/palacios/svm_halt.c
@@ -30,7 +30,7 @@
 
 
 //
-// This should trigger a #GP if cpl!=0, otherwise, yield to host
+// This should trigger a #GP if cpl != 0, otherwise, yield to host
 //
 
 int v3_handle_svm_halt(struct guest_info * info) {
@@ -39,8 +39,8 @@ int v3_handle_svm_halt(struct guest_info * info) {
        v3_raise_exception(info, GPF_EXCEPTION);
     } else {
     
-       ullong_t yield_start = 0;
-       ullong_t yield_stop = 0;
+       uint64_t yield_start = 0;
+       uint64_t yield_stop = 0;
        uint32_t gap = 0;
        
        PrintDebug("CPU Yield\n");
@@ -53,7 +53,18 @@ int v3_handle_svm_halt(struct guest_info * info) {
        //v3_update_time(info, yield_stop - yield_start);
        gap = yield_stop - yield_start;
 
-       v3_raise_irq(info, 0);
+       /*  WARNING!!! WARNING!!!
+        *  
+        * DO NOT REMOVE THIS CONDITIONAL!!!
+        *
+        * It is common for an OS to issue an IO op, and then sit in a halt loop
+        * waiting for the device to complete and raise an irq.
+        * If you remove this then the timer interrupt will ALWAYS subvert the completion 
+        * interrupt and stall the guest.
+        */
+       if (!v3_intr_pending(info)) {
+           v3_raise_irq(info, 0);
+       }
 
        
        PrintDebug("CPU Yield Done (%d cycles)\n", gap);