Re: [RFC PATCH v5 6/8] drivers: vmware: balloon - report inflated memory

[Nadav Amit <namit@vmware.com>]

On Oct 19, 2022, at 12:56 PM, Alexander Atanasov <alexander.atanasov@virtuozzo.com> wrote:

> Update the inflated memory in the mm core on change.
> 
> Signed-off-by: Alexander Atanasov <alexander.atanasov@virtuozzo.com>
> ---
> drivers/misc/vmw_balloon.c | 1 +
> 1 file changed, 1 insertion(+)
> 
> diff --git a/drivers/misc/vmw_balloon.c b/drivers/misc/vmw_balloon.c
> index 91d4d2a285c5..3bfd845898f5 100644
> --- a/drivers/misc/vmw_balloon.c
> +++ b/drivers/misc/vmw_balloon.c
> @@ -1507,6 +1507,7 @@ static void vmballoon_work(struct work_struct *work)
> 	queue_delayed_work(system_freezable_wq,
> 			   dwork, round_jiffies_relative(HZ));
> 
> +	balloon_set_inflated_free(atomic64_read(&b->size) << 2);
> }

I don’t like it in general (I think that something like that should go into
some common infra).

But more concretely there are at least 2 problems here. First, queueing the
work should come last. Second, there are other places that change the
balloon size (e.g., vmballoon_reset()), which are not handled by this patch.

If you added calls to balloon_set_inflated_free() from these places, you can
get races while calling balloon_set_inflated_free() and the last value that
would be latched would be the wrong one. I don’t see anything in the logic
or comments that clarify how something like that should be resolved.