On 12/03/2019 20:46, David Hildenbrand wrote:
On 12.03.19 19:23, David Hildenbrand wrote:
I guess something like this could do the trick if I understood it correctly:
diff --git a/drivers/xen/balloon.c b/drivers/xen/balloon.c
index 39b229f9e256..d37dd5bb7a8f 100644
--- a/drivers/xen/balloon.c
+++ b/drivers/xen/balloon.c
@@ -604,6 +604,7 @@ int alloc_xenballooned_pages(int nr_pages, struct
page **pages)
while (pgno < nr_pages) {
page = balloon_retrieve(true);
if (page) {
+ __ClearPageOffline(page);
pages[pgno++] = page;
#ifdef CONFIG_XEN_HAVE_PVMMU
/*
@@ -645,8 +646,10 @@ void free_xenballooned_pages(int nr_pages, struct
page **pages)
mutex_lock(&balloon_mutex);
for (i = 0; i < nr_pages; i++) {
- if (pages[i])
+ if (pages[i]) {
+ __SetPageOffline(pages[i]);
balloon_append(pages[i]);
+ }
}
balloon_stats.target_unpopulated -= nr_pages;
At least this way, the pages allocated (and thus eventually mapped to
user space) would not be marked, but the other ones would remain marked
and could be excluded by makedumptool.
I think this patch should do the trick. Julien, could you give it a
try? On x86 I can't reproduce your problem easily as dom0 is PV with
plenty of unpopulated pages for grant memory not suffering from
missing "offline" bit.