This fixes an issue with gradual decommit in scenarios where we have … (

#84975)

This fixes an issue with gradual decommit in scenarios where we have almost exclusively background GCs.

The problem is that we turn off the gradual_decommit_in_progress_p flag and rely on the distribute_free_regions flag to turn it back on. If we only trigger a BGC however, distribute_free_regions won't get called and the flag will stay off, despite the global_regions_to_decommit list being non-empty.

This will cause regions to accumulate in the global_regions_to_decommit list, and eventual will cause the process to run out of memory.

The fix is to just test for the no gc region case in decommit_step. This will also turn off the gradual_decommit_in_progress_p flag for the no gc region case, but not for background GCs.

src/coreclr/gc/gc.cpp

@@ -6848,7 +6848,6 @@ void gc_heap::gc_thread_function ()
             END_TIMING(suspend_ee_during_log);
 
             proceed_with_gc_p = TRUE;
-            gradual_decommit_in_progress_p = FALSE;
 
             if (!should_proceed_with_gc())
             {
@@ -41526,6 +41525,12 @@ void gc_heap::decommit_ephemeral_segment_pages()
 // return true if we actually decommitted anything
 bool gc_heap::decommit_step (uint64_t step_milliseconds)
 {
+    if (settings.pause_mode == pause_no_gc)
+    {
+        // don't decommit at all if we have entered a no gc region
+        return false;
+    }
+
     size_t decommit_size = 0;
 
 #ifdef USE_REGIONS