On Thu, Aug 20, 2026 at 04:00:30PM +0100, Steven Price wrote:
> diff --git a/drivers/irqchip/irq-gic-v3-its.c
> b/drivers/irqchip/irq-gic-v3-its.c
> index 6f5811aae59c..a055837832bc 100644
> --- a/drivers/irqchip/irq-gic-v3-its.c
> +++ b/drivers/irqchip/irq-gic-v3-its.c
> @@ -213,16 +213,18 @@ static gfp_t gfp_flags_quirk;
> static struct page *its_alloc_pages_node(int node, gfp_t gfp,
> unsigned int order)
> {
> + bool want_zero = gfp & __GFP_ZERO;
> struct page *page;
> int ret = 0;
>
> - page = alloc_pages_node(node, gfp | gfp_flags_quirk, order);
> + page = alloc_pages_node(node, (gfp & ~__GFP_ZERO) | gfp_flags_quirk,
> + order);
I don't think pKVM does any scrubbing on set_memory_decrypted(), so it
potentially exposes confidential guest data before it reaches
clear_pages() below.
>
> if (!page)
> return NULL;
>
> ret = set_memory_decrypted((unsigned long)page_address(page),
> - 1 << order);
> + BIT(order));
> /*
> * If set_memory_decrypted() fails then we don't know what state the
> * page is in, so we can't free it. Instead we leak it.
> @@ -231,6 +233,9 @@ static struct page *its_alloc_pages_node(int node, gfp_t
> gfp,
> if (ret)
> return NULL;
>
> + if (want_zero)
> + clear_pages(page_address(page), BIT(order));
> +
> return page;
> }
>
> --
> 2.43.0
--
Catalin
