2015-08-10 18:11:45

by Julien Grall

Subject: [PATCH] xen/xenbus: Don't leak memory when unmapping the ring on HVM backend

The commit ccc9d90a9a8b5c4ad7e9708ec41f75ff9e98d61d "xenbus_client:
Extend interface to support multi-page ring" removes the call to
free_xenballooned_pages in xenbus_unmap_ring_vfree_hvm.

This will result to not give back the pages to Linux and loose them
forever. It only happens when the backends are running in HVM domains.

Signed-off-by: Julien Grall <[email protected]>

---
Cc: Konrad Rzeszutek Wilk <[email protected]>
Cc: Boris Ostrovsky <[email protected]>
Cc: David Vrabel <[email protected]>
Cc: Wei Liu <[email protected]>

Appeared in Linux 4.1. HVM backend, which is always the case on ARM, will
leak every mapped ring (i.e ~12KB per domain with 1 disk and 1 vif).
---
drivers/xen/xenbus/xenbus_client.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/drivers/xen/xenbus/xenbus_client.c b/drivers/xen/xenbus/xenbus_client.c
index 9ad3272..e303535 100644
--- a/drivers/xen/xenbus/xenbus_client.c
+++ b/drivers/xen/xenbus/xenbus_client.c
@@ -814,8 +814,10 @@ static int xenbus_unmap_ring_vfree_hvm(struct xenbus_device *dev, void *vaddr)

rv = xenbus_unmap_ring(dev, node->handles, node->nr_handles,
addrs);
- if (!rv)
+ if (!rv) {
vunmap(vaddr);
+ free_xenballooned_pages(node->nr_handles, node->hvm.pages);
+ }
else
WARN(1, "Leaking %p, size %u page(s)\n", vaddr,
node->nr_handles);
--
2.1.4

2015-08-10 18:38:54

by Boris Ostrovsky

Subject: Re: [PATCH] xen/xenbus: Don't leak memory when unmapping the ring on HVM backend

On 08/10/2015 02:10 PM, Julien Grall wrote:
> The commit ccc9d90a9a8b5c4ad7e9708ec41f75ff9e98d61d "xenbus_client:
> Extend interface to support multi-page ring" removes the call to
> free_xenballooned_pages in xenbus_unmap_ring_vfree_hvm.
>
> This will result to not give back the pages to Linux and loose them
> forever. It only happens when the backends are running in HVM domains.
>
> Signed-off-by: Julien Grall <[email protected]>

Reviewed-by: Boris Ostrovsky <[email protected]>

>
> ---
> Cc: Konrad Rzeszutek Wilk <[email protected]>
> Cc: Boris Ostrovsky <[email protected]>
> Cc: David Vrabel <[email protected]>
> Cc: Wei Liu <[email protected]>
>
> Appeared in Linux 4.1. HVM backend, which is always the case on ARM, will
> leak every mapped ring (i.e ~12KB per domain with 1 disk and 1 vif).
> ---
> drivers/xen/xenbus/xenbus_client.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/xen/xenbus/xenbus_client.c b/drivers/xen/xenbus/xenbus_client.c
> index 9ad3272..e303535 100644
> --- a/drivers/xen/xenbus/xenbus_client.c
> +++ b/drivers/xen/xenbus/xenbus_client.c
> @@ -814,8 +814,10 @@ static int xenbus_unmap_ring_vfree_hvm(struct xenbus_device *dev, void *vaddr)
>
> rv = xenbus_unmap_ring(dev, node->handles, node->nr_handles,
> addrs);
> - if (!rv)
> + if (!rv) {
> vunmap(vaddr);
> + free_xenballooned_pages(node->nr_handles, node->hvm.pages);
> + }
> else
> WARN(1, "Leaking %p, size %u page(s)\n", vaddr,
> node->nr_handles);

2015-08-10 18:46:43

by Wei Liu

Subject: Re: [PATCH] xen/xenbus: Don't leak memory when unmapping the ring on HVM backend

On Mon, Aug 10, 2015 at 07:10:38PM +0100, Julien Grall wrote:
> The commit ccc9d90a9a8b5c4ad7e9708ec41f75ff9e98d61d "xenbus_client:
> Extend interface to support multi-page ring" removes the call to
> free_xenballooned_pages in xenbus_unmap_ring_vfree_hvm.
>
> This will result to not give back the pages to Linux and loose them
> forever. It only happens when the backends are running in HVM domains.
>
> Signed-off-by: Julien Grall <[email protected]>
>
> ---
> Cc: Konrad Rzeszutek Wilk <[email protected]>
> Cc: Boris Ostrovsky <[email protected]>
> Cc: David Vrabel <[email protected]>
> Cc: Wei Liu <[email protected]>
>

Reviewed-by: Wei Liu <[email protected]>

> Appeared in Linux 4.1. HVM backend, which is always the case on ARM, will
> leak every mapped ring (i.e ~12KB per domain with 1 disk and 1 vif).
> ---
> drivers/xen/xenbus/xenbus_client.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/xen/xenbus/xenbus_client.c b/drivers/xen/xenbus/xenbus_client.c
> index 9ad3272..e303535 100644
> --- a/drivers/xen/xenbus/xenbus_client.c
> +++ b/drivers/xen/xenbus/xenbus_client.c
> @@ -814,8 +814,10 @@ static int xenbus_unmap_ring_vfree_hvm(struct xenbus_device *dev, void *vaddr)
>
> rv = xenbus_unmap_ring(dev, node->handles, node->nr_handles,
> addrs);
> - if (!rv)
> + if (!rv) {
> vunmap(vaddr);
> + free_xenballooned_pages(node->nr_handles, node->hvm.pages);
> + }
> else
> WARN(1, "Leaking %p, size %u page(s)\n", vaddr,
> node->nr_handles);
> --
> 2.1.4

2015-08-11 10:05:05

by David Vrabel

Subject: Re: [Xen-devel] [PATCH] xen/xenbus: Don't leak memory when unmapping the ring on HVM backend

On 10/08/15 19:10, Julien Grall wrote:
> The commit ccc9d90a9a8b5c4ad7e9708ec41f75ff9e98d61d "xenbus_client:
> Extend interface to support multi-page ring" removes the call to
> free_xenballooned_pages in xenbus_unmap_ring_vfree_hvm.
>
> This will result to not give back the pages to Linux and loose them
> forever. It only happens when the backends are running in HVM domains.

Applied to for-linus-4.2 and tagged for stable, thanks.

David