race in usbnet.c in full RT
Hi,
seems there is a race in drivers/net/usbnet.c in full RT mode. To be honest I haven't hardly checked this on the latest kernel and latest RT patch but just took a look at usbnet.c and latest RT patch and haven't observed any related changes.
The usbnet_start_xmit() contains the following code:
... spin_lock_irqsave (&dev->txq.lock, flags);
#ifdef CONFIG_USB_NET1080
if (info->flags & FLAG_FRAMING_NC) {
header->packet_id = cpu_to_le16 ((u16)dev->dev_packet_id++);
put_unaligned (header->packet_id, &trailer->packet_id);
#if 0
devdbg (dev, "frame >tx h %d p %d id %d",
header->hdr_len, header->packet_len,
header->packet_id);
#endif
}#endif /* CONFIG_USB_NET1080 */
switch ((retval = usb_submit_urb (urb, GFP_ATOMIC))) {
case -EPIPE:
netif_stop_queue (net);
defer_kevent (dev, EVENT_TX_HALT);
break;
default:
devdbg (dev, "tx: submit urb err %d", retval);
break;
case 0:
net->trans_start = jiffies;
__skb_queue_tail (&dev->txq, skb);
if (dev->txq.qlen >= TX_QLEN (dev))
netif_stop_queue (net);
}
spin_unlock_irqrestore (&dev->txq.lock, flags);
...THe race in full RT is between tx_complete() routine and __skb_queue_tail (&dev->txq, skb): skb->list gets initialized in __skb_queue_tail() but may be dereferenced before initialization at defer_bh() called from tx_complete() (since tx_complete() and usbnet_start_xmit() are completely asynchronous routines).
in non-RT case spin_lock_irqsave (&dev->txq.lock, flags) disables interrupts and thus code from usb_submit_urb() call upto __skb_queue_tail (&dev->txq, skb) executes atomically. But in RT case interrupts are not disabled and usb_submit_urb() triggers an interrupt which may cause tx_complete() execution before __skb_queue_tail () call. And since skb->list gets initialized just at __skb_queue_tail(), call to tx_complete() (via defer_bh() which thus executes before __skb_queue_tail) dereferences NULL (skb->list) pointer.
Thus looks tx_complete() and usbnet_start_xmit() require a serialization. Please find proposed fix attached though not sure the patch will apply cleanly to the latest kernel.
Eugeny
drivers/usb/net/usbnet.c +2 -0
@@ -2820,6 +2820,8 @@ urb->dev =NULL ; entry->state = tx_done; +spin_lock_rt (&dev->txq.lock); +spin_unlock_rt (&dev->txq.lock);defer_bh (dev, skb); }
Quoted 9 lines · Eugeny S. Mints
in non-RT case
spin_lock_irqsave(&dev->txq.lock, flags) disables interrupts and thus code fromusb_submit_urb()call upto__skb_queue_tail(&dev->txq, skb) executes atomically. But in RT case interrupts are not disabled andusb_submit_urb()triggers an interrupt which may causetx_complete()execution before__skb_queue_tail() call. And sinceskb->listgets initialized just at__skb_queue_tail(), call totx_complete()(viadefer_bh()which thus executes before__skb_queue_tail) dereferences NULL (skb->list) pointer.
Quoted 3 lines
Thus looks
tx_complete()andusbnet_start_xmit()require a serialization. Please find proposed fix attached though not sure the patch will apply cleanly to the latest kernel.
Please fix whatever patch you use for "full RT mode" to not break valid assupmtions in drivers.
Quoted 4 lines · Eugeny S. Mints
seems there is a race in
drivers/net/usbnet.cin full RT mode. To be honest I haven't hardly checked this on the latest kernel and latest RT patch but just took a look atusbnet.cand latest RT patch and haven't observed any related changes.
thanks, i've applied your patch to my tree. Note that your patch is specific to the -RT kernel (both in terms of semantics and in term of API dependence), so it does not make any sense to apply it upstream. David, please ignore it.
Ingo
Quoted 7 lines · Ingo Molnar
Quoted 1 line · Eugeny S. Mints
seems there is a race in
drivers/net/usbnet.cin full RT mode. To be honest I haven't hardly checked this on the latest kernel and latest RT patch but just took a look atusbnet.cand latest RT patch and haven't observed any related changes.thanks, i've applied your patch to my tree. Note that your patch is specific to the -RT kernel (both in terms of semantics and in term of API dependence), so it does not make any sense to apply it upstream.
exactly. I put David in CC just to be sure usbnet.c has not been rewritten in the recent kerenls in sense of locking scheme.
Eugeny
Quoted 6 lines
David, please ignore it.
Ingo
Quoted 7 lines · Ingo Molnar
Quoted 4 lines · Eugeny S. Mints
seems there is a race in
drivers/net/usbnet.cin full RT mode. To be honest I haven't hardly checked this on the latest kernel and latest RT patch but just took a look atusbnet.cand latest RT patch and haven't observed any related changes.
Quoted 5 lines
thanks, i've applied your patch to my tree. Note that your patch is specific to the -RT kernel (both in terms of semantics and in term of API dependence), so it does not make any sense to apply it upstream. David, please ignore it.
Is this action only take place on the same CPU, or is this also an SMP problem? I would think if this is a race with full RT, that this may also be a race with SMP, unless the race is guaranteed to always happen on the same CPU. Then this is only a RT problem.
-- Steve
Quoted 7 lines · Steven Rostedt
Quoted 5 lines · Ingo Molnar
Quoted 1 line · Eugeny S. Mints
seems there is a race in
drivers/net/usbnet.cin full RT mode. To be honest I haven't hardly checked this on the latest kernel and latest RT patch but just took a look atusbnet.cand latest RT patch and haven't observed any related changes.
Quoted 8 lines
Quoted 2 lines
thanks, i've applied your patch to my tree. Note that your patch is specific to the -RT kernel (both in terms of semantics and in term of API dependence), so it does not make any sense to apply it upstream. David, please ignore it.
Is this action only take place on the same CPU, or is this also an SMP problem? I would think if this is a race with full RT, that this may also be a race with SMP, unless the race is guaranteed to always happen on the same CPU. Then this is only a RT problem.
thanks, good point. looks like it could be SMP problem but probably David is able to say it for sure as usb host code expert. David?
Eugeny
Quoted 6 lines
-- Steve