xen: let alloc_xenballooned_pages() fail if not enough memory free

commit a1078e8 upstream.

Instead of trying to allocate pages with GFP_USER in
add_ballooned_pages() check the available free memory via
si_mem_available(). GFP_USER is far less limiting memory exhaustion
than the test via si_mem_available().

This will avoid dom0 running out of memory due to excessive foreign
page mappings especially on ARM and on x86 in PVH mode, as those don't
have a pre-ballooned area which can be used for foreign mappings.

As the normal ballooning suffers from the same problem don't balloon
down more than si_mem_available() pages in one iteration. At the same
time limit the default maximum number of retries.

This is part of XSA-300.

Signed-off-by: Juergen Gross <jgross@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

drivers/xen/balloon.c

@@ -538,8 +538,15 @@ static void balloon_process(struct work_struct *work)
 				state = reserve_additional_memory();
 		}
 
-		if (credit < 0)
-			state = decrease_reservation(-credit, GFP_BALLOON);
+		if (credit < 0) {
+			long n_pages;
+
+			n_pages = min(-credit, si_mem_available());
+			state = decrease_reservation(n_pages, GFP_BALLOON);
+			if (state == BP_DONE && n_pages != -credit &&
+			    n_pages < totalreserve_pages)
+				state = BP_EAGAIN;
+		}
 
 		state = update_schedule(state);
 
@@ -578,6 +585,9 @@ static int add_ballooned_pages(int nr_pages)
 		}
 	}
 
+	if (si_mem_available() < nr_pages)
+		return -ENOMEM;
+
 	st = decrease_reservation(nr_pages, GFP_USER);
 	if (st != BP_DONE)
 		return -ENOMEM;
@@ -710,7 +720,7 @@ static int __init balloon_init(void)
 	balloon_stats.schedule_delay = 1;
 	balloon_stats.max_schedule_delay = 32;
 	balloon_stats.retry_count = 1;
-	balloon_stats.max_retry_count = RETRY_UNLIMITED;
+	balloon_stats.max_retry_count = 4;
 
 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 	set_online_page_callback(&xen_online_page);