do not allow parallel flushing of snapshot md clusters since both md types are implicitly in linked list format

Author: ashwin-simplyblock

lib/blob/blobstore.c

@@ -14102,7 +14102,8 @@ blob_search_for_new_flush_job(struct spdk_blob *blob, struct t_flush_job *job) {
 				++blob->current_array_ordinal;
 				blob->next_idx_in_array = blob->active.extent_pages_array_size - 1; // next array is extent pages array
 			}
-		} else if (blob->current_array_ordinal <= 2) {
+		} else if (blob->current_array_ordinal <= 2 && blob->nflush_jobs_current == 0) // no parallelism allowed on md
+		{
 			job->is_md_job = true;
 
 			uint32_t *page_idxs_arr = blob->current_array_ordinal == 1 ? blob->active.extent_pages : blob->active.pages;
@@ -14127,6 +14128,10 @@ blob_search_for_new_flush_job(struct spdk_blob *blob, struct t_flush_job *job) {
 				++blob->current_array_ordinal;
 				blob->next_idx_in_array = blob->active.num_pages - 1; // next array after extent pages array is pages array
 			}
+		} else if (job->buf) // job will never be done again, free its buffer to save memory
+		{
+			spdk_free(job->buf);
+			job->buf = NULL;
 		}
 	}
 	return 0;

lib/blob/blobstore.h

@@ -204,8 +204,8 @@ struct spdk_blob {
 	are on the same array to safely flush data before metadata and less 
 	sensitive metadata before other metadata. 
 
-	However, the very first non-extent page must be the last md page to be flushed. 
-	For convenience sake, we'll just flush both types of md pages in reverse order.
+	However, extent pages and md pages are implicitly ordered like a linked list, so 
+	we must flush them one md cluster at a time in reverse order.
 	*/
 	uint64_t next_idx_in_array;