On 02.05.21 08:33, Mike Rapoport wrote:

On Thu, Apr 29, 2021 at 02:25:18PM +0200, David Hildenbrand wrote:

quoted

Let's properly use page_offline_(start|end) to synchronize setting
PageOffline(), so we won't have valid page access to unplugged memory
regions from /proc/kcore.

Signed-off-by: David Hildenbrand <redacted>
---
  drivers/virtio/virtio_mem.c | 2 ++
  mm/util.c                   | 2 ++
  2 files changed, 4 insertions(+)

diff --git a/drivers/virtio/virtio_mem.c b/drivers/virtio/virtio_mem.c
index 10ec60d81e84..dc2a2e2b2ff8 100644
--- a/drivers/virtio/virtio_mem.c
+++ b/drivers/virtio/virtio_mem.c

@@ -1065,6 +1065,7 @@ static int virtio_mem_memory_notifier_cb(struct notifier_block *nb,
  static void virtio_mem_set_fake_offline(unsigned long pfn,
  					unsigned long nr_pages, bool onlined)
  {
+	page_offline_begin();
  	for (; nr_pages--; pfn++) {
  		struct page *page = pfn_to_page(pfn);
  

@@ -1075,6 +1076,7 @@ static void virtio_mem_set_fake_offline(unsigned long pfn,
  			ClearPageReserved(page);
  		}
  	}
+	page_offline_end();

I'm not really familiar with ballooning and memory hotplug, but is it the
only place that needs page_offline_{begin,end} ?

Existing balloon implementations that I am aware of (Hyper-V, XEN, 
virtio-balloon, vmware-balloon) usually allow reading inflated memory; 
doing so might result in unnecessary overhead in the hypervisor, so we 
really want to avoid it -- but it's strictly not forbidden and has been 
working forever. So we barely care about races: if there would be a rare 
race, we'd still be able to read that memory.

For virtio-mem, it'll be different in the future when using shmem, huge 
pages, !anonymous private mappings, ... as backing storage for a VM; 
there will be a virtio spec extension to document that virtio-mem 
changes that indicate the new behavior won't allow reading unplugged 
memory and doing so will result in undefined behavior.

-- 
Thanks,

David / dhildenb