On 20.11.2023 17.33, Mathias Nyman wrote:
On 18.11.2023 12.19, Kuen-Han Tsai wrote:
Hi Mathias
if (usb_endpoint_xfer_isoc(&urb->ep->desc))
@@ -1552,8 +1561,10 @@ static int xhci_urb_enqueue(struct usb_hcd *hcd, struct urb *urb, gfp_t mem_flag
num_tds = 1;
urb_priv = kzalloc(struct_size(urb_priv, td, num_tds), mem_flags);
kzalloc with spinlock held, should preferably be moved outside lock, otherwise should use GFP_ATOMIC
Thanks for pointing this out. I realize this patch is incorrect and it
is non-ideal to include many codes unrelated to xhci->devs[slot_id]
within the lock.
xhci_check_maxpacket() called here can't be called with spinlock held
It appears that xhci_check_maxpacket() might potentially lead to a
deadlock later if a spinlock is held. Is this the concern you were
referring to? If not, please let me know if there are any other
potential issues that I may have missed, thanks!
xhci_check_maxpacket() will allocate memory, wait for completion, and use the same lock,
so there are several issues here.
I actually think we shouldn't call xhci_check_maxpacket() at all while queuing urbs.
usb core knows when there was max packet size mismatch during enumeration.
I think we should add a hook to the hcd that usb core can call in these cases
I moved the max packet checks away from xhci_urb_enqueue() and fixed up the locking.
I can't trigger the original issue, but I tested it by setting incorrect initial max packet
size values.
If you have the chance to test this with your setup I'd appreciate it.
patches found here:
git://git.kernel.org/pub/scm/linux/kernel/git/mnyman/xhci.git fix_urb_enqueue_locking
https://git.kernel.org/pub/scm/linux/kernel/git/mnyman/xhci.git/log/?h=fix_urb_enqueue_locking
I'll add them to this thread as well
thanks
Mathias
