The work can submit URBs and the URBs can schedule the work. This cycle needs to be broken, when a device is to be stopped. Use a flag to do so. This is a design issue as old as the driver.
Signed-off-by: Oliver Neukum oneukum@suse.com Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2") CC: stable@vger.kernel.org ---
v2: fix PM reference issue
drivers/net/usb/usbnet.c | 37 ++++++++++++++++++++++++++++--------- include/linux/usb/usbnet.h | 17 +++++++++++++++++ 2 files changed, 45 insertions(+), 9 deletions(-)
diff --git a/drivers/net/usb/usbnet.c b/drivers/net/usb/usbnet.c index 18eb5ba436df..2506aa8c603e 100644 --- a/drivers/net/usb/usbnet.c +++ b/drivers/net/usb/usbnet.c @@ -464,10 +464,15 @@ static enum skb_state defer_bh(struct usbnet *dev, struct sk_buff *skb, void usbnet_defer_kevent (struct usbnet *dev, int work) { set_bit (work, &dev->flags); - if (!schedule_work (&dev->kevent)) - netdev_dbg(dev->net, "kevent %s may have been dropped\n", usbnet_event_names[work]); - else - netdev_dbg(dev->net, "kevent %s scheduled\n", usbnet_event_names[work]); + if (!usbnet_going_away(dev)) { + if (!schedule_work(&dev->kevent)) + netdev_dbg(dev->net, + "kevent %s may have been dropped\n", + usbnet_event_names[work]); + else + netdev_dbg(dev->net, + "kevent %s scheduled\n", usbnet_event_names[work]); + } } EXPORT_SYMBOL_GPL(usbnet_defer_kevent);
@@ -535,7 +540,8 @@ static int rx_submit (struct usbnet *dev, struct urb *urb, gfp_t flags) tasklet_schedule (&dev->bh); break; case 0: - __usbnet_queue_skb(&dev->rxq, skb, rx_start); + if (!usbnet_going_away(dev)) + __usbnet_queue_skb(&dev->rxq, skb, rx_start); } } else { netif_dbg(dev, ifdown, dev->net, "rx: stopped\n"); @@ -843,9 +849,18 @@ int usbnet_stop (struct net_device *net)
/* deferred work (timer, softirq, task) must also stop */ dev->flags = 0; - del_timer_sync (&dev->delay); - tasklet_kill (&dev->bh); + del_timer_sync(&dev->delay); + tasklet_kill(&dev->bh); cancel_work_sync(&dev->kevent); + + /* We have cyclic dependencies. Those calls are needed + * to break a cycle. We cannot fall into the gaps because + * we have a flag + */ + tasklet_kill(&dev->bh); + del_timer_sync(&dev->delay); + cancel_work_sync(&dev->kevent); + if (!pm) usb_autopm_put_interface(dev->intf);
@@ -1171,7 +1186,8 @@ usbnet_deferred_kevent (struct work_struct *work) status); } else { clear_bit (EVENT_RX_HALT, &dev->flags); - tasklet_schedule (&dev->bh); + if (!usbnet_going_away(dev)) + tasklet_schedule(&dev->bh); } }
@@ -1196,7 +1212,8 @@ usbnet_deferred_kevent (struct work_struct *work) usb_autopm_put_interface(dev->intf); fail_lowmem: if (resched) - tasklet_schedule (&dev->bh); + if (!usbnet_going_away(dev)) + tasklet_schedule(&dev->bh); } }
@@ -1559,6 +1576,7 @@ static void usbnet_bh (struct timer_list *t) } else if (netif_running (dev->net) && netif_device_present (dev->net) && netif_carrier_ok(dev->net) && + !usbnet_going_away(dev) && !timer_pending(&dev->delay) && !test_bit(EVENT_RX_PAUSED, &dev->flags) && !test_bit(EVENT_RX_HALT, &dev->flags)) { @@ -1606,6 +1624,7 @@ void usbnet_disconnect (struct usb_interface *intf) usb_set_intfdata(intf, NULL); if (!dev) return; + usbnet_mark_going_away(dev);
xdev = interface_to_usbdev (intf);
diff --git a/include/linux/usb/usbnet.h b/include/linux/usb/usbnet.h index 9f08a584d707..d02d6f16da46 100644 --- a/include/linux/usb/usbnet.h +++ b/include/linux/usb/usbnet.h @@ -76,8 +76,25 @@ struct usbnet { # define EVENT_LINK_CHANGE 11 # define EVENT_SET_RX_MODE 12 # define EVENT_NO_IP_ALIGN 13 +/* This one is special, as it indicates that the device is going away + * there are cyclic dependencies between tasklet, timer and bh + * that must be broken + */ +# define EVENT_UNPLUG 31 };
+static inline bool usbnet_going_away(struct usbnet *ubn) +{ + smp_mb__before_atomic(); /* against usbnet_mark_going_away() */ + return test_bit(EVENT_UNPLUG, &ubn->flags); +} + +static inline void usbnet_mark_going_away(struct usbnet *ubn) +{ + set_bit(EVENT_UNPLUG, &ubn->flags); + smp_mb__after_atomic(); /* against usbnet_going_away() */ +} + static inline struct usb_driver *driver_of(struct usb_interface *intf) { return to_usb_driver(intf->dev.driver);
On 9/5/24 15:46, Oliver Neukum wrote:
The work can submit URBs and the URBs can schedule the work. This cycle needs to be broken, when a device is to be stopped. Use a flag to do so. This is a design issue as old as the driver.
Signed-off-by: Oliver Neukum oneukum@suse.com Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2") CC: stable@vger.kernel.org
v2: fix PM reference issue
drivers/net/usb/usbnet.c | 37 ++++++++++++++++++++++++++++--------- include/linux/usb/usbnet.h | 17 +++++++++++++++++ 2 files changed, 45 insertions(+), 9 deletions(-)
diff --git a/drivers/net/usb/usbnet.c b/drivers/net/usb/usbnet.c index 18eb5ba436df..2506aa8c603e 100644 --- a/drivers/net/usb/usbnet.c +++ b/drivers/net/usb/usbnet.c @@ -464,10 +464,15 @@ static enum skb_state defer_bh(struct usbnet *dev, struct sk_buff *skb, void usbnet_defer_kevent (struct usbnet *dev, int work) { set_bit (work, &dev->flags);
- if (!schedule_work (&dev->kevent))
netdev_dbg(dev->net, "kevent %s may have been dropped\n", usbnet_event_names[work]);- else
netdev_dbg(dev->net, "kevent %s scheduled\n", usbnet_event_names[work]);
- if (!usbnet_going_away(dev)) {
if (!schedule_work(&dev->kevent))netdev_dbg(dev->net,"kevent %s may have been dropped\n",usbnet_event_names[work]);elsenetdev_dbg(dev->net,"kevent %s scheduled\n", usbnet_event_names[work]);- } } EXPORT_SYMBOL_GPL(usbnet_defer_kevent);
@@ -535,7 +540,8 @@ static int rx_submit (struct usbnet *dev, struct urb *urb, gfp_t flags) tasklet_schedule (&dev->bh); break; case 0:
__usbnet_queue_skb(&dev->rxq, skb, rx_start);
if (!usbnet_going_away(dev))__usbnet_queue_skb(&dev->rxq, skb, rx_start);@@ -843,9 +849,18 @@ int usbnet_stop (struct net_device *net) /* deferred work (timer, softirq, task) must also stop */ dev->flags = 0;
- del_timer_sync (&dev->delay);
- tasklet_kill (&dev->bh);
- del_timer_sync(&dev->delay);
- tasklet_kill(&dev->bh); cancel_work_sync(&dev->kevent);
- /* We have cyclic dependencies. Those calls are needed
* to break a cycle. We cannot fall into the gaps because* we have a flag*/- tasklet_kill(&dev->bh);
- del_timer_sync(&dev->delay);
- cancel_work_sync(&dev->kevent);
I guess you do the shutdown twice because a running tasklet or timer could re-schedule the others? If so, what prevent the rescheduling to happen in the 2nd iteration? why can't you add usbnet_going_away() checks on tasklet and timer reschedule point?
Thanks,
Paolo
On 10.09.24 11:58, Paolo Abeni wrote:
I guess you do the shutdown twice because a running tasklet or timer could re-schedule the others? If so, what prevent the rescheduling to happen in the 2nd iteration? why can't you add usbnet_going_away() checks on tasklet and timer reschedule point?
Hi,
I am not sure I fully understand the question. Technically the flag prevents it in cooperation with del_timer_sync(), which will wait for the timer handler to run to completion.
Hence if the timer handler has passed the the check the first time, it will see the flag the second time. I am not sure that answers the question, because AFAICT I have added the checks, but there is an inevitable window between the check and acting upon it.
Regards Oliver
On Thu, 5 Sep 2024 15:46:50 +0200 Oliver Neukum wrote:
+static inline bool usbnet_going_away(struct usbnet *ubn) +{
- smp_mb__before_atomic(); /* against usbnet_mark_going_away() */
- return test_bit(EVENT_UNPLUG, &ubn->flags);
+}
+static inline void usbnet_mark_going_away(struct usbnet *ubn) +{
- set_bit(EVENT_UNPLUG, &ubn->flags);
- smp_mb__after_atomic(); /* against usbnet_going_away() */
+}
I have sort of an inverse question to what Paolo asked :) AFAIU we need the double-cancel because checking the flag and scheduling are not atomic. But if we do that why the memory barriers? They make it seem like we're doing something clever with memory ordering, while really we're just depending on normal properties of the tasklet/timer/work APIs.
FTR disable_work_sync() would work nicely here but it'd be a PITA for backports.
Also - is this based on some report or syzbot? I'm a bit tempted to put this in net-next given how unlikely the race is vs how commonly used the driver is.
On 11.09.24 00:44, Jakub Kicinski wrote:
Hi,
I have sort of an inverse question to what Paolo asked :)
This is getting interesting.
AFAIU we need the double-cancel because checking the flag and scheduling are not atomic. But if we do that why the memory
Right.
barriers? They make it seem like we're doing something clever with memory ordering, while really we're just depending on normal properties of the tasklet/timer/work APIs.
Good question. I added this because they are used in usbnet_defer_kevent() which can be used in hard irq context. Are you saying I should check whether this is actually needed?
FTR disable_work_sync() would work nicely here but it'd be a PITA for backports.
So should I use it?
Also - is this based on some report or syzbot? I'm a bit tempted to put this in net-next given how unlikely the race is vs how commonly used the driver is.
Having found the thing with the random MAC I decided to look closely at the driver for overlooked stuff.
Regards Oliver
On Thu, 12 Sep 2024 11:37:14 +0200 Oliver Neukum wrote:
barriers? They make it seem like we're doing something clever with memory ordering, while really we're just depending on normal properties of the tasklet/timer/work APIs.
Good question. I added this because they are used in usbnet_defer_kevent() which can be used in hard irq context. Are you saying I should check whether this is actually needed?
I am slightly bolder, I'm saying that my reading of the code is that it is in fact not needed :) We build our "proof of correctness" on tasklet/timer/work APIs which already provide all necessary barriers.
FTR disable_work_sync() would work nicely here but it'd be a PITA for backports.
So should I use it?
Up to you. It'd avoid work rescheduling but the backport would be a pain, and off top of my head timer doesn't have a disable so we'd still need the flag.
linux-stable-mirror@lists.linaro.org
-
Jakub Kicinski -
Oliver Neukum -
Paolo Abeni