usbnet.c, buf.overrun crash-bugfix, Kernel 2.6.12-rc1
PATCH 1/1 usbnet.c, buf.overrun crash-bugfix, Kernel 2.6.12-rc1
Atleast versions 2.6.5 - 2.6.12-rc1 crash if an USB device using usbnet sends oversized packet. Such packets occur most likely with broken device. Here's a patch that throws away such packet, to keep the machine
from crashing. Hopefully this doesn't leave memory unreleased. If it does, it's still better than crashing as such oversized packets are really rare.
drivers/usb/net/usbnet.c 2005-03-24 16:46:08.000000000 +0200 renamed from drivers/usb/net/usbnet.c 2005-03-18 03:34:13.000000000 +0200 +14 -3
@@ -2795,9 +2795,20 @@ struct usbnet *dev = entry->dev; int urb_status = urb->status; - skb_put (skb, urb->actual_length); - entry->state = rx_done; - entry->urb = NULL; + if (unlikely((skb->tail + urb->actual_length) > skb->end)) { + entry->state = rx_cleanup; + dev->stats.rx_errors++; + dev->stats.rx_length_errors++; + entry->urb = NULL; + printk(KERN_ERR + "USB RX packet too long, discarded. " + "Your slave device most likely is broken\n"); + /* lets hope upper level protocols will recover */ + } else { + skb_put(skb, urb->actual_length); + entry->state = rx_done; + entry->urb = NULL; + } switch (urb_status) { // success
Quoted 3 lines · Jakemuksen spammiosote
Atleast versions 2.6.5 - 2.6.12-rc1 crash if an USB device using usbnet sends oversized packet. Such packets occur most likely with broken device.
Care to mention what device(s) you saw this with? And what HCD?
drivers/usb/net/usbnet.c 2005-03-24 16:46:08.000000000 +0200 line 2798
Quoted 4 lines
drivers/usb/net/usbnet.c 2005-03-24 16:46:08.000000000 +0200 renamed from drivers/usb/net/usbnet.c 2005-03-18 03:34:13.000000000 +0200 +1
-3- skb_put (skb, urb->actual_length); - entry->state = rx_done; - entry->urb = NULL; + if (unlikely((skb->tail + urb->actual_length) > skb->end)) {
This logic looks wrong. If that ever happens, surely the problem is that the rx_submit() code submitted an urb with transfer_size that mismatched the SKB. The host controller isn't allowed to overrun the end of the buffer it's passed. And if it's tempted to do so, it's supposed to fill up to the end (skb->end in this case...) and then report urb->status of -EOVERFLOW.
If you insist on changing this bit of logic, then the best way to ignore the packet is just to force urb->status to -EOVERFLOW
Quoted 18 lines
drivers/usb/net/usbnet.c 2005-03-24 16:46:08.000000000 +0200 renamed from drivers/usb/net/usbnet.c 2005-03-18 03:34:13.000000000 +0200 +13
-0+ entry->state = rx_cleanup; + dev->stats.rx_errors++; + dev->stats.rx_length_errors++; + entry->urb = NULL; + printk(KERN_ERR + "USB RX packet too long, discarded. " + "Your slave device most likely is broken\n"); + /* lets hope upper level protocols will recover */ + } else { + skb_put(skb, urb->actual_length); + entry->state = rx_done; + entry->urb = NULL; + } switch (urb_status) { // success
Quoted 4 lines · David Brownell
Quoted 2 lines · Jakemuksen spammiosote
Atleast versions 2.6.5 - 2.6.12-rc1 crash if an USB device using usbnet sends oversized packet. Such packets occur most likely with broken
Care to mention what device(s) you saw this with? And what HCD?
I can't tell about the device(NDA), and don't remember the HCD and I can check it only after holidays.
Quoted 4 lines
Quoted 1 line
+ if (
unlikely((skb->tail + urb->actual_length) > skb->end)) {This logic looks wrong. If that ever happens, surely the problem is that the
rx_submit()code submitted an urb withtransfer_sizethat mismatched the SKB. The host controller isn't allowed to overrun the
Sounds reasonable. So, I'll go thru the HCD code instead if the responsibility is there. Am i the first one to run into such crash situation? If so, perhaps it's not ever worthy to fix in mainstream kernel, as the device causes the crash under very specific - 'abusing' one might say, situation only.
Quoted 10 lines · Jakemuksen spammiosote
Quoted 7 lines · David Brownell
Quoted 1 line · Jakemuksen spammiosote
+ if (
unlikely((skb->tail + urb->actual_length) > skb->end)) {This logic looks wrong. If that ever happens, surely the problem is that the
rx_submit()code submitted an urb withtransfer_sizethat mismatched the SKB. The host controller isn't allowed to overrun theSounds reasonable. So, I'll go thru the HCD code
Better yet, start with the code supporting that device you're under NDA for.
Quoted 5 lines
instead if the
responsibility is there. Am i the first one to run into such crash situation? If so, perhaps it's not ever worthy to fix in mainstream kernel, as the device causes the crash under very specific - 'abusing' one might say, situation only.
You're the first one to report such a problem.
- Dave