2021-04-20 14:34:52

by Christophe JAILLET

Subject: [PATCH resend] xhci: Do not use GFP_KERNEL in (potentially) atomic context

'xhci_urb_enqueue()' is passed a 'mem_flags' argument, because "URBs may be
submitted in interrupt context" (see comment related to 'usb_submit_urb()'
in 'drivers/usb/core/urb.c')

So this flag should be used in all the calling chain.
Up to now, 'xhci_check_maxpacket()' which is only called from
'xhci_urb_enqueue()', uses GFP_KERNEL.

Be safe and pass the mem_flags to this function as well.

Fixes: ddba5cd0aeff ("xhci: Use command structures when queuing commands on the command ring")
Signed-off-by: Christophe JAILLET <[email protected]>
---
I'm not 100% sure of the Fixes tag. The commit is the only that introduced
this GFP_KERNEL, but I've not checked what was the behavior before that.

If the patch is correct, I guess that a cc stable should be welcome.

This patch was proposed on 14/08/20. It has been rebased on latest -next tree.
---
drivers/usb/host/xhci.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/drivers/usb/host/xhci.c b/drivers/usb/host/xhci.c
index ca9385d22f68..27283654ca08 100644
--- a/drivers/usb/host/xhci.c
+++ b/drivers/usb/host/xhci.c
@@ -1514,7 +1514,7 @@ static int xhci_configure_endpoint(struct xhci_hcd *xhci,
* we need to issue an evaluate context command and wait on it.
*/
static int xhci_check_maxpacket(struct xhci_hcd *xhci, unsigned int slot_id,
- unsigned int ep_index, struct urb *urb)
+ unsigned int ep_index, struct urb *urb, gfp_t mem_flags)
{
struct xhci_container_ctx *out_ctx;
struct xhci_input_control_ctx *ctrl_ctx;
@@ -1545,7 +1545,7 @@ static int xhci_check_maxpacket(struct xhci_hcd *xhci, unsigned int slot_id,
* changes max packet sizes.
*/

- command = xhci_alloc_command(xhci, true, GFP_KERNEL);
+ command = xhci_alloc_command(xhci, true, mem_flags);
if (!command)
return -ENOMEM;

@@ -1639,7 +1639,7 @@ static int xhci_urb_enqueue(struct usb_hcd *hcd, struct urb *urb, gfp_t mem_flag
*/
if (urb->dev->speed == USB_SPEED_FULL) {
ret = xhci_check_maxpacket(xhci, slot_id,
- ep_index, urb);
+ ep_index, urb, mem_flags);
if (ret < 0) {
xhci_urb_free_priv(urb_priv);
urb->hcpriv = NULL;
--
2.27.0

2021-04-21 08:01:54

by Mathias Nyman

Subject: Re: [PATCH resend] xhci: Do not use GFP_KERNEL in (potentially) atomic context

On 20.4.2021 15.32, Christophe JAILLET wrote:
> 'xhci_urb_enqueue()' is passed a 'mem_flags' argument, because "URBs may be
> submitted in interrupt context" (see comment related to 'usb_submit_urb()'
> in 'drivers/usb/core/urb.c')
>
> So this flag should be used in all the calling chain.
> Up to now, 'xhci_check_maxpacket()' which is only called from
> 'xhci_urb_enqueue()', uses GFP_KERNEL.
>
> Be safe and pass the mem_flags to this function as well.
>
> Fixes: ddba5cd0aeff ("xhci: Use command structures when queuing commands on the command ring")
> Signed-off-by: Christophe JAILLET <[email protected]>
> ---
> I'm not 100% sure of the Fixes tag. The commit is the only that introduced
> this GFP_KERNEL, but I've not checked what was the behavior before that.
>
> If the patch is correct, I guess that a cc stable should be welcome.
>
> This patch was proposed on 14/08/20. It has been rebased on latest -next tree.

Added to queue, and added stable tag

Thanks
-Mathias