Skip to contentSkip to conversation

Several races in "usbnet" module (kernel 4.1.x)

PATCH In review Eugene Shatokhin 6 participants 53 messages Jul 20, 2015 – Sep 8, 2015

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

Hi,

I have recently found several data races in "usbnet" module, checked on vanilla kernel 4.1.0 on x86_64. The races do actually happen, I have confirmed it by adding delays and using hardware breakpoints to detect the conflicting memory accesses (with RaceHound tool, https://github.com/winnukem/racehound).

I have not analyzed yet how harmful these races are (if they are), but it is better to report them anyway, I think.

Everything was checked using YOTA 4G LTE Modem that works via "usbnet" and "cdc_ether" kernel modules.
--------------------------

[Race #1]

Race on skb_queue ('next' pointer) between usbnet_stop() and rx_complete().

Reproduced that by unplugging the device while the system was downloading a large file from the Net.

Here is part of the call stack with the code where the changes to the queue happen:

#0 __skb_unlink (skbuff.h:1517)

	prev->next = next;
#1 defer_bh (usbnet.c:430)
	spin_lock_irqsave(&list->lock, flags);
	old_state = entry->state;
	entry->state = state;
	__skb_unlink(skb, list);
	spin_unlock(&list->lock);
	spin_lock(&dev->done.lock);
	__skb_queue_tail(&dev->done, skb);
	if (dev->done.qlen == 1)
		tasklet_schedule(&dev->bh);
	spin_unlock_irqrestore(&dev->done.lock, flags);
#2 rx_complete (usbnet.c:640)
	state = defer_bh(dev, skb, &dev->rxq, state);

At the same time, the following code repeatedly checks if the queue is empty and reads the same values concurrently with the above changes:

#0 usbnet_terminate_urbs (usbnet.c:765)

	/* maybe wait for deletions to finish. */
	while (!skb_queue_empty(&dev->rxq)
		&& !skb_queue_empty(&dev->txq)
		&& !skb_queue_empty(&dev->done)) {
			schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
			set_current_state(TASK_UNINTERRUPTIBLE);
			netif_dbg(dev, ifdown, dev->net,
				  "waited for %d urb completions\n", temp);
	}
#1  usbnet_stop (usbnet.c:806)
	if (!(info->flags & FLAG_AVOID_UNLINK_URBS))
		usbnet_terminate_urbs(dev);

For example, it is possible that the skb is removed from dev->rxq by __skb_unlink() before the check "!skb_queue_empty(&dev->rxq)" in usbnet_terminate_urbs() is made. It is also possible in this case that the skb is added to dev->done queue after "!skb_queue_empty(&dev->done)" is checked. So usbnet_terminate_urbs() may stop waiting and return while dev->done queue still has an item.
--------------------------

Unrelated the that race, if the goal of that while loop in usbnet_terminate_urbs() is to wait till all three queues (dev->rxq, dev->txq, dev->done) become empty, perhaps the code should be changed as follows:

	while (!skb_queue_empty(&dev->rxq)
-		&& !skb_queue_empty(&dev->txq)
-		&& !skb_queue_empty(&dev->done)) {
+		|| !skb_queue_empty(&dev->txq)
+		|| !skb_queue_empty(&dev->done)) {
			schedule_timeout(...));

--------------------------

[Race #2]

Races on dev->rx_qlen. Reproduced these by repeatedly changing MTU (1500 <-> 1400) while downloading large files.

dev->rx_qlen is written to here:
#0 usbnet_update_max_qlen (usbnet.c:351)

	case USB_SPEED_HIGH:
		dev->rx_qlen = MAX_QUEUE_MEMORY / dev->rx_urb_size;
#1  __handle_link_change (usbnet.c:1049)
	/* hard_mtu or rx_urb_size may change during link change */
	usbnet_update_max_qlen(dev);
#2  usbnet_deferred_kevent (usbnet.c:1172)
	if (test_bit (EVENT_LINK_CHANGE, &dev->flags))
		__handle_link_change(dev);

Here are the conflicting reads from dev->rx_qlen (via RX_QLEN(dev)), 3 code locations:

* usbnet_bh (usbnet.c:1492)

	if (temp < RX_QLEN(dev)) { ...
* usbnet_bh (usbnet.c:1499)
	if (dev->rxq.qlen < RX_QLEN(dev)) ...
* rx_alloc_submit (usbnet.c:1431)
	for (i = 0; i < 10 && dev->rxq.qlen < RX_QLEN(dev); i++) { ...

--------------------------

[Race #3]

Similar to race #2 but on dev->tx_qlen. I reproduced it the same way.

dev->tx_qlen is written to here:
#0 usbnet_update_max_qlen (usbnet.c:352)

	case USB_SPEED_HIGH:
		dev->rx_qlen = MAX_QUEUE_MEMORY / dev->rx_urb_size;
		dev->tx_qlen = MAX_QUEUE_MEMORY / dev->hard_mtu;
#1  __handle_link_change (usbnet.c:1049)
	/* hard_mtu or rx_urb_size may change during link change */
	usbnet_update_max_qlen(dev);
#2  usbnet_deferred_kevent (usbnet.c:1172)
	if (test_bit (EVENT_LINK_CHANGE, &dev->flags))
		__handle_link_change(dev);

Here are the conflicting reads from dev->tx_qlen (via TX_QLEN(dev)), 2 code locations:

* usbnet_bh (usbnet.c:1502)

		if (dev->txq.qlen < TX_QLEN (dev))
			netif_wake_queue (dev->net);
* usbnet_start_xmit (usbnet.c:1398)
		if (dev->txq.qlen >= TX_QLEN (dev))
			netif_stop_queue (net);

--------------------------

[Race #4]

Race on dev->flags. It happened when I unplugged the device while a large file was being downloaded.

dev->flags is set to 0 here:

#0 usbnet_stop (usbnet.c:816)

	/* deferred work (task, timer, softirq) must also stop.
	 * can't flush_scheduled_work() until we drop rtnl (later),
	 * else workers could deadlock; so make workers a NOP.
	 */
	dev->flags = 0;
	del_timer_sync (&dev->delay);
	tasklet_kill (&dev->bh);

And here, the code clears EVENT_RX_KILL bit in dev->flags, which may execute concurrently with the above operation:
#0 clear_bit (bitops.h:113, inlined)
#1 usbnet_bh (usbnet.c:1475)

	/* restart RX again after disabling due to high error rate */
	clear_bit(EVENT_RX_KILL, &dev->flags);

If clear_bit() is atomic w.r.t. setting dev->flags to 0, this race is not a problem, I guess. Otherwise, it may be.
--------------------------

[Race #5]

Race on dev->rx_urb_size. I reproduced it a similar way as the races #2 and #3 (changing MTU while downloading files).

dev->rx_urb_size is written to here:
#0 usbnet_change_mtu (usbnet.c:392)

	dev->rx_urb_size = dev->hard_mtu;

Here is the conflicting read from dev->rx_urb_size:
* rx_submit (usbnet.c:467)

	size_t			size = dev->rx_urb_size;

--------------------------

Regards,
Eugene

Signature · 2 lines
Eugene Shatokhin, ROSA
www.rosalab.com
maintainer
#1 lore raw
Quoted 58 lines · Eugene Shatokhin

Hi,

I have recently found several data races in "usbnet" module, checked on vanilla kernel 4.1.0 on x86_64. The races do actually happen, I have confirmed it by adding delays and using hardware breakpoints to detect the conflicting memory accesses (with RaceHound tool,
https://github.com/winnukem/racehound).

I have not analyzed yet how harmful these races are (if they are), but it is better to report them anyway, I think.

Everything was checked using YOTA 4G LTE Modem that works via "usbnet" and "cdc_ether" kernel modules.
--------------------------

[Race #1]

Race on skb_queue ('next' pointer) between usbnet_stop() and rx_complete().

Reproduced that by unplugging the device while the system was downloading a large file from the Net.

Here is part of the call stack with the code where the changes to the queue happen:

#0 __skb_unlink (skbuff.h:1517)

	prev->next = next;
#1 defer_bh (usbnet.c:430)
	spin_lock_irqsave(&list->lock, flags);
	old_state = entry->state;
	entry->state = state;
	__skb_unlink(skb, list);
	spin_unlock(&list->lock);
	spin_lock(&dev->done.lock);
	__skb_queue_tail(&dev->done, skb);
	if (dev->done.qlen == 1)
		tasklet_schedule(&dev->bh);
	spin_unlock_irqrestore(&dev->done.lock, flags);
#2 rx_complete (usbnet.c:640)
	state = defer_bh(dev, skb, &dev->rxq, state);

At the same time, the following code repeatedly checks if the queue is empty and reads the same values concurrently with the above changes:

#0 usbnet_terminate_urbs (usbnet.c:765)

	/* maybe wait for deletions to finish. */
	while (!skb_queue_empty(&dev->rxq)
		&& !skb_queue_empty(&dev->txq)
		&& !skb_queue_empty(&dev->done)) {
			schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
			set_current_state(TASK_UNINTERRUPTIBLE);
			netif_dbg(dev, ifdown, dev->net,
				  "waited for %d urb completions\n", temp);
	}
#1  usbnet_stop (usbnet.c:806)
	if (!(info->flags & FLAG_AVOID_UNLINK_URBS))
		usbnet_terminate_urbs(dev);
Quoted 6 lines

For example, it is possible that the skb is removed from dev->rxq by __skb_unlink() before the check "!skb_queue_empty(&dev->rxq)" in usbnet_terminate_urbs() is made. It is also possible in this case that the skb is added to dev->done queue after "!skb_queue_empty(&dev->done)" is checked. So usbnet_terminate_urbs() may stop waiting and return while dev->done queue still has an item.

Hi,

your analysis is correct and it looks like in addition to your proposed fix locking needs to be simplified and a common lock to be taken. Suggestions?

Regards
Oliver

#2 lore raw

21.07.2015 15:04, Oliver Neukum пишет:

Quoted 66 lines · Eugene Shatokhin
Quoted 58 lines

Hi,

I have recently found several data races in "usbnet" module, checked on vanilla kernel 4.1.0 on x86_64. The races do actually happen, I have confirmed it by adding delays and using hardware breakpoints to detect the conflicting memory accesses (with RaceHound tool,
https://github.com/winnukem/racehound).

I have not analyzed yet how harmful these races are (if they are), but it is better to report them anyway, I think.

Everything was checked using YOTA 4G LTE Modem that works via "usbnet" and "cdc_ether" kernel modules.
--------------------------

[Race #1]

Race on skb_queue ('next' pointer) between usbnet_stop() and rx_complete().

Reproduced that by unplugging the device while the system was downloading a large file from the Net.

Here is part of the call stack with the code where the changes to the queue happen:

#0 __skb_unlink (skbuff.h:1517)

	prev->next = next;
#1 defer_bh (usbnet.c:430)
	spin_lock_irqsave(&list->lock, flags);
	old_state = entry->state;
	entry->state = state;
	__skb_unlink(skb, list);
	spin_unlock(&list->lock);
	spin_lock(&dev->done.lock);
	__skb_queue_tail(&dev->done, skb);
	if (dev->done.qlen == 1)
		tasklet_schedule(&dev->bh);
	spin_unlock_irqrestore(&dev->done.lock, flags);
#2 rx_complete (usbnet.c:640)
	state = defer_bh(dev, skb, &dev->rxq, state);

At the same time, the following code repeatedly checks if the queue is empty and reads the same values concurrently with the above changes:

#0 usbnet_terminate_urbs (usbnet.c:765)

	/* maybe wait for deletions to finish. */
	while (!skb_queue_empty(&dev->rxq)
		&& !skb_queue_empty(&dev->txq)
		&& !skb_queue_empty(&dev->done)) {
			schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
			set_current_state(TASK_UNINTERRUPTIBLE);
			netif_dbg(dev, ifdown, dev->net,
				  "waited for %d urb completions\n", temp);
	}
#1  usbnet_stop (usbnet.c:806)
	if (!(info->flags & FLAG_AVOID_UNLINK_URBS))
		usbnet_terminate_urbs(dev);
Quoted 6 lines

For example, it is possible that the skb is removed from dev->rxq by __skb_unlink() before the check "!skb_queue_empty(&dev->rxq)" in usbnet_terminate_urbs() is made. It is also possible in this case that the skb is added to dev->done queue after "!skb_queue_empty(&dev->done)" is checked. So usbnet_terminate_urbs() may stop waiting and return while dev->done queue still has an item.

Quoted 5 lines

Hi,

your analysis is correct and it looks like in addition to your proposed fix locking needs to be simplified and a common lock to be taken. Suggestions?

Just an idea, I haven't tested it.

How about moving the operations with dev->done under &list->lock in defer_bh, while keeping dev->done.lock too and changing usbnet_terminate_urbs() as described below?

Like this:
@@ -428,12 +428,12 @@ static enum skb_state defer_bh(struct usbnet *dev, struct sk_buff *skb,
old_state = entry->state;
entry->state = state;
__skb_unlink(skb, list);
- spin_unlock(&list->lock);
spin_lock(&dev->done.lock);
__skb_queue_tail(&dev->done, skb);
if (dev->done.qlen == 1)
tasklet_schedule(&dev->bh);
- spin_unlock_irqrestore(&dev->done.lock, flags);
+ spin_unlock(&dev->done.lock);
+ spin_unlock_irqrestore(&list->lock, flags);
return old_state;
}
-------------------

usbnet_terminate_urbs() can then be changed as follows:

@@ -749,6 +749,20 @@ EXPORT_SYMBOL_GPL(usbnet_unlink_rx_urbs);

/*-------------------------------------------------------------------------*/

+static void wait_skb_queue_empty(struct sk_buff_head *q)
+{
+ unsigned long flags;
+
+ spin_lock_irqsave(&q->lock, flags);
+ while (!skb_queue_empty(q)) {
+ spin_unlock_irqrestore(&q->lock, flags);
+ schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
+ set_current_state(TASK_UNINTERRUPTIBLE);
+ spin_lock_irqsave(&q->lock, flags);
+ }
+ spin_unlock_irqrestore(&q->lock, flags);
+}
+
// precondition: never called in_interrupt
static void usbnet_terminate_urbs(struct usbnet *dev)
{
@@ -762,14 +776,11 @@ static void usbnet_terminate_urbs(struct usbnet *dev)
unlink_urbs(dev, &dev->rxq);

/* maybe wait for deletions to finish. */
- while (!skb_queue_empty(&dev->rxq)
- && !skb_queue_empty(&dev->txq)
- && !skb_queue_empty(&dev->done)) {
- schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
- set_current_state(TASK_UNINTERRUPTIBLE);
- netif_dbg(dev, ifdown, dev->net,
- "waited for %d urb completions\n", temp);
- }
+ wait_skb_queue_empty(&dev->rxq);
+ wait_skb_queue_empty(&dev->txq);
+ wait_skb_queue_empty(&dev->done);
+ netif_dbg(dev, ifdown, dev->net,
+ "waited for %d urb completions\n", temp);
set_current_state(TASK_RUNNING);
remove_wait_queue(&dev->wait, &wait);
}
-------------------

This way, when usbnet_terminate_urbs() finds dev->rxq or dev->txq empty, the skbs from these queues, if there were any, have already been queued to dev->done.

At the first glance, moving the code under list->lock in defer_bh() should not produce deadlocks. Still, I suppose, it is better to use lockdep to be sure.

Regards,
Eugene

--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

maintainer
#3 lore raw
Quoted 10 lines · Eugene Shatokhin

21.07.2015 15:04, Oliver Neukum пишет:

Quoted 3 lines

your analysis is correct and it looks like in addition to your proposed fix locking needs to be simplified and a common lock to be taken. Suggestions?

Just an idea, I haven't tested it.

How about moving the operations with dev->done under &list->lock in defer_bh, while keeping dev->done.lock too and changing

Why keep dev->done.lock?
Does it make sense at all?

Quoted 26 lines

usbnet_terminate_urbs() as described below?

Like this:

(quoted diff) +2 -3
@@ -428,12 +428,12 @@ static enum skb_state defer_bh(struct usbnet *dev, 
 struct sk_buff *skb,
  	old_state = entry->state;
  	entry->state = state;
  	__skb_unlink(skb, list);
-	spin_unlock(&list->lock);
  	spin_lock(&dev->done.lock);
  	__skb_queue_tail(&dev->done, skb);
  	if (dev->done.qlen == 1)
  		tasklet_schedule(&dev->bh);
-	spin_unlock_irqrestore(&dev->done.lock, flags);
+	spin_unlock(&dev->done.lock);
+	spin_unlock_irqrestore(&list->lock, flags);
  	return old_state;
  }
-------------------

usbnet_terminate_urbs() can then be changed as follows:

(quoted diff) +0 -0
@@ -749,6 +749,20 @@ EXPORT_SYMBOL_GPL(usbnet_unlink_rx_urbs);
 
 
 /*-------------------------------------------------------------------------*/

Review EXPORT_SYMBOL_GPL(usbnet_unlink_rx_urbs); · line 761

Quoted 10 lines
(quoted diff) +9 -0
 
+static void wait_skb_queue_empty(struct sk_buff_head *q)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&q->lock, flags);
+	while (!skb_queue_empty(q)) {
+		spin_unlock_irqrestore(&q->lock, flags);
+		schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
+		set_current_state(TASK_UNINTERRUPTIBLE);

I suppose you want to invert those lines

Review EXPORT_SYMBOL_GPL(usbnet_unlink_rx_urbs); · line 766

Quoted 5 lines
(quoted diff) +5 -0
+		spin_lock_irqsave(&q->lock, flags);
+	}
+	spin_unlock_irqrestore(&q->lock, flags);
+}
+

Your changes make sense, but it locks to me as if a lock would become totally redundant.

Regards
Oliver

#4 lore raw

27.07.2015 15:29, Oliver Neukum пишет:

Quoted 7 lines · Eugene Shatokhin
Quoted 6 lines

21.07.2015 15:04, Oliver Neukum пишет:

Quoted 3 lines

your analysis is correct and it looks like in addition to your proposed fix locking needs to be simplified and a common lock to be taken. Suggestions?

Quoted 7 lines
Quoted 4 lines

Just an idea, I haven't tested it.

How about moving the operations with dev->done under &list->lock in defer_bh, while keeping dev->done.lock too and changing

Why keep dev->done.lock?
Does it make sense at all?

I think it does.

Both skb_queue_tail(&dev->done, skb) called from rx_process() and skb_dequeue (&dev->done) called from usbnet_bh() take dev->done.lock internally. So, to synchronize accesses to dev->done, one needs that lock in defer_bh() too.

Quoted 28 lines
Quoted 28 lines

usbnet_terminate_urbs() as described below?

Like this:

(quoted diff) +2 -3
@@ -428,12 +428,12 @@ static enum skb_state defer_bh(struct usbnet *dev,
 struct sk_buff *skb,
   	old_state = entry->state;
   	entry->state = state;
   	__skb_unlink(skb, list);
-	spin_unlock(&list->lock);
   	spin_lock(&dev->done.lock);
   	__skb_queue_tail(&dev->done, skb);
   	if (dev->done.qlen == 1)
   		tasklet_schedule(&dev->bh);
-	spin_unlock_irqrestore(&dev->done.lock, flags);
+	spin_unlock(&dev->done.lock);
+	spin_unlock_irqrestore(&list->lock, flags);
   	return old_state;
   }
-------------------

usbnet_terminate_urbs() can then be changed as follows:

(quoted diff) +0 -0
@@ -749,6 +749,20 @@ EXPORT_SYMBOL_GPL(usbnet_unlink_rx_urbs);
 
 
 /*-------------------------------------------------------------------------*/
 
Quoted 11 lines

Review EXPORT_SYMBOL_GPL(usbnet_unlink_rx_urbs); · line 761

Quoted 9 lines
(quoted diff) +9 -0
+static void wait_skb_queue_empty(struct sk_buff_head *q)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&q->lock, flags);
+	while (!skb_queue_empty(q)) {
+		spin_unlock_irqrestore(&q->lock, flags);
+		schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
+		set_current_state(TASK_UNINTERRUPTIBLE);

I suppose you want to invert those lines

Do you mean +set_current_state(TASK_UNINTERRUPTIBLE); +schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS)); ?

Quoted 10 lines

Review EXPORT_SYMBOL_GPL(usbnet_unlink_rx_urbs); · line 766

Quoted 6 lines
(quoted diff) +5 -0
+		spin_lock_irqsave(&q->lock, flags);
+	}
+	spin_unlock_irqrestore(&q->lock, flags);
+}
+

Your changes make sense, but it locks to me as if a lock would become totally redundant.

Regards,

Eugene

maintainer
#5 lore raw
In reply to Eugene Shatokhin
Quoted 3 lines · Eugene Shatokhin

Races on dev->rx_qlen. Reproduced these by repeatedly changing MTU (1500
<-> 1400) while downloading large files.

Hi,

I don't see how it matters much. The number of buffers is just an optimization. As long as it eventually is corrected I don't see harm.

Regards
Oliver

maintainer
#6 lore raw
In reply to Eugene Shatokhin
Quoted 9 lines · Eugene Shatokhin

And here, the code clears EVENT_RX_KILL bit in dev->flags, which may execute concurrently with the above operation:
#0 clear_bit (bitops.h:113, inlined)
#1 usbnet_bh (usbnet.c:1475)

        /* restart RX again after disabling due to high error rate */
        clear_bit(EVENT_RX_KILL, &dev->flags);

If clear_bit() is atomic w.r.t. setting dev->flags to 0, this race is not a problem, I guess. Otherwise, it may be.

clear_bit is atomic with respect to other atomic operations. So how about this:

Regards
Oliver

Quoted 1 line

From 1c4e685b3a9c183e04c46b661830e5c7ed35b513 Mon Sep 17 00:00:00 2001

From: Oliver Neukum <oneukum-IBi9RG/[email protected]>
Date: Tue, 21 Jul 2015 16:19:40 +0200
Subject: [PATCH] usbnet: fix race between usbnet_stop() and the BH

Does this do the job?

---
 drivers/net/usb/usbnet.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)
drivers/net/usb/usbnet.c +6 -3
index 3c86b10..77a9a86 100644
@@ -778,7 +778,7 @@ int usbnet_stop (struct net_device *net)
 {
 	struct usbnet		*dev = netdev_priv(net);
 	struct driver_info	*info = dev->driver_info;
-	int			retval, pm;
+	int			retval, pm, mpn;
 
 	clear_bit(EVENT_DEV_OPEN, &dev->flags);
 	netif_stop_queue (net);
@@ -813,14 +813,17 @@ int usbnet_stop (struct net_device *net)
 	 * can't flush_scheduled_work() until we drop rtnl (later),
 	 * else workers could deadlock; so make workers a NOP.
 	 */
+	mpn = !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags);
 	dev->flags = 0;
 	del_timer_sync (&dev->delay);
 	tasklet_kill (&dev->bh);
+	mpn |= !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags);
+	/* in case the bh reset a flag */
+	dev->flags = 0;
 	if (!pm)
 		usb_autopm_put_interface(dev->intf);
 
-	if (info->manage_power &&
-	    !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags))
+	if (info->manage_power && mpn)
 		info->manage_power(dev, 0);
 	else
 		usb_autopm_put_interface(dev->intf);
Signature · 8 lines
2.1.4



--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#7 lore raw

21.07.2015 17:22, Oliver Neukum пишет:

Quoted 8 lines · Eugene Shatokhin
Quoted 7 lines

And here, the code clears EVENT_RX_KILL bit in dev->flags, which may execute concurrently with the above operation:
#0 clear_bit (bitops.h:113, inlined)
#1 usbnet_bh (usbnet.c:1475)

         /* restart RX again after disabling due to high error rate */
         clear_bit(EVENT_RX_KILL, &dev->flags);
Quoted 9 lines
Quoted 2 lines

If clear_bit() is atomic w.r.t. setting dev->flags to 0, this race is not a problem, I guess. Otherwise, it may be.

clear_bit is atomic with respect to other atomic operations. So how about this:

Regards
Oliver

Thanks for the quick replies!
My comments are below.

Quoted 26 lines
Quoted 1 line

From 1c4e685b3a9c183e04c46b661830e5c7ed35b513 Mon Sep 17 00:00:00 2001

From: Oliver Neukum <[email protected]>
Date: Tue, 21 Jul 2015 16:19:40 +0200
Subject: [PATCH] usbnet: fix race between usbnet_stop() and the BH

Does this do the job?

---
  drivers/net/usb/usbnet.c | 9 ++++++---
  1 file changed, 6 insertions(+), 3 deletions(-)
drivers/net/usb/usbnet.c +1 -1
index 3c86b10..77a9a86 100644
@@ -778,7 +778,7 @@ int usbnet_stop (struct net_device *net)
  {
  	struct usbnet		*dev = netdev_priv(net);
  	struct driver_info	*info = dev->driver_info;
-	int			retval, pm;
+	int			retval, pm, mpn;
 
  	clear_bit(EVENT_DEV_OPEN, &dev->flags);
  	netif_stop_queue (net);

drivers/net/usb/usbnet.c int usbnet_stop (struct net_device *net) · line 817

Quoted 5 lines
drivers/net/usb/usbnet.c +1 -0
@@ -813,14 +813,17 @@ int usbnet_stop (struct net_device *net)
  	 * can't flush_scheduled_work() until we drop rtnl (later),
  	 * else workers could deadlock; so make workers a NOP.
  	 */
 
+	mpn = !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags);

Right, I missed that. Indeed, if one needs EVENT_NO_RUNTIME_PM bit, one should get it before dev->flags is set to 0.

drivers/net/usb/usbnet.c int usbnet_stop (struct net_device *net) · line 817

Quoted 1 line
drivers/net/usb/usbnet.c +0 -0
  	dev->flags = 0;

I suppose usbnet_bh() cannot be re-scheduled at this point. And if it is running now, tasklet_kill will wait till it finishes. So, I guess, it would be enough to zero dev->flags after "tasklet_kill (&dev->bh);" rather than before it, like it is now.

Anyway, if it is needed to clear any particular flags to prevent re-scheduling of usbnet_bh(), this can be done here with clear_bit(). Not sure if there are such flags, I am by no means an expert in usbnet.

drivers/net/usb/usbnet.c int usbnet_stop (struct net_device *net) · line 819

Quoted 2 lines
drivers/net/usb/usbnet.c +0 -0
  	del_timer_sync (&dev->delay);
  	tasklet_kill (&dev->bh);

The following part is not necessary, I think. usbnet_bh() does not touch EVENT_NO_RUNTIME_PM bit explicitly and these bit operations are atomic w.r.t. each other.

drivers/net/usb/usbnet.c int usbnet_stop (struct net_device *net) · line 821

Quoted 2 lines
drivers/net/usb/usbnet.c +2 -0
+	mpn |= !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags);
+	/* in case the bh reset a flag */

But zeroing dev->flags here is necessary, I agree.

Quoted 9 lines
drivers/net/usb/usbnet.c +2 -2
+	dev->flags = 0;
 
  	if (!pm)
  		usb_autopm_put_interface(dev->intf);
 
-	if (info->manage_power &&
-	    !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags))
+	if (info->manage_power && mpn)
  		info->manage_power(dev, 0);
Quoted 3 lines
  	else
  		usb_autopm_put_interface(dev->intf);

Regards,
Eugene

maintainer
#8 lore raw
Quoted 8 lines · Eugene Shatokhin

The following part is not necessary, I think. usbnet_bh() does not touch
EVENT_NO_RUNTIME_PM bit explicitly and these bit operations are atomic
w.r.t. each other.

Quoted 2 lines
drivers/net/usb/usbnet.c +2 -0
+     mpn |= !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags);
+     /* in case the bh reset a flag */

Yes, they are atomic w.r.t. each other. And that limitation worries me.

I am considering architectures which do atomic operations with spinlocks. And this code mixes another operation into it. Can this happen?

CPU A CPU B

take lock
read old value
set value to 0
clear bit
write back changed value
release lock

Regards
Oliver

#9 lore raw

23.07.2015 12:15, Oliver Neukum пишет:

Quoted 16 lines · Eugene Shatokhin
Quoted 8 lines

The following part is not necessary, I think. usbnet_bh() does not touch
EVENT_NO_RUNTIME_PM bit explicitly and these bit operations are atomic
w.r.t. each other.

Quoted 2 lines
(quoted diff) +2 -0
+     mpn |= !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags);
+     /* in case the bh reset a flag */

Yes, they are atomic w.r.t. each other. And that limitation worries me.

I am considering architectures which do atomic operations with spinlocks. And this code mixes another operation into it. Can this happen?

Quoted 8 lines

CPU A CPU B

take lock
read old value
set value to 0
clear bit
write back changed value
release lock

From what I see now in Documentation/atomic_ops.txt, stores to the properly aligned memory locations are in fact atomic.

So, I think, the situation you described above cannot happen for dev->flags, which is good. No need to address that in the patch. The race might be harmless after all.

If I understand the code correctly now, dev->flags is set to 0 in usbnet_stop() so that the worker function (usbnet_deferred_kevent) would do nothing, should it start later. If so, how about adding memory barriers for all CPUs to see dev->flags is 0 before other things?

The patch could look like this then:

--------------------

drivers/net/usb/usbnet.c +9 -3
index 3c86b10..d87b9c7 100644
@@ -778,7 +778,7 @@ int usbnet_stop (struct net_device *net)
 {
 	struct usbnet		*dev = netdev_priv(net);
 	struct driver_info	*info = dev->driver_info;
-	int			retval, pm;
+	int			retval, pm, mpn;
 
 	clear_bit(EVENT_DEV_OPEN, &dev->flags);
 	netif_stop_queue (net);
@@ -813,14 +813,17 @@ int usbnet_stop (struct net_device *net)
 	 * can't flush_scheduled_work() until we drop rtnl (later),
 	 * else workers could deadlock; so make workers a NOP.
 	 */
+	mpn = !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags);
 	dev->flags = 0;
+	smp_mb(); /* make sure the workers see that dev->flags == 0 */
+
 	del_timer_sync (&dev->delay);
 	tasklet_kill (&dev->bh);
+
 	if (!pm)
 		usb_autopm_put_interface(dev->intf);
 
-	if (info->manage_power &&
-	    !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags))
+	if (info->manage_power && mpn)
 		info->manage_power(dev, 0);
 	else
 		usb_autopm_put_interface(dev->intf);
@@ -1078,6 +1081,9 @@ usbnet_deferred_kevent (struct work_struct *work)
 		container_of(work, struct usbnet, kevent);
 	int			status;
 
+	/* See the changes in dev->flags from other CPUs. */
+	smp_mb();
+
 	/* usb_clear_halt() needs a thread context */
 	if (test_bit (EVENT_TX_HALT, &dev->flags)) {
 		unlink_urbs (dev, &dev->txq);

--------------------

What do you think?

Regards,
Eugene

maintainer
#10 lore raw
Quoted 4 lines · Eugene Shatokhin

23.07.2015 12:15, Oliver Neukum пишет:

From what I see now in Documentation/atomic_ops.txt, stores to the
properly aligned memory locations are in fact atomic.

They are, but again only with respect to each other.

Quoted 6 lines

So, I think, the situation you described above cannot happen for dev->flags, which is good. No need to address that in the patch. The race might be harmless after all.

If I understand the code correctly now, dev->flags is set to 0 in usbnet_stop() so that the worker function (usbnet_deferred_kevent) would

Yes, particularly not reschedule itself.

Quoted 2 lines

do nothing, should it start later. If so, how about adding memory barriers for all CPUs to see dev->flags is 0 before other things?

Taking a lock, as del_timer_sync() does, implies a memory barrier, as does a work.

Regards
Oliver

#11 lore raw

27.07.2015 13:00, Oliver Neukum пишет:

Quoted 7 lines · Eugene Shatokhin
Quoted 4 lines

23.07.2015 12:15, Oliver Neukum пишет:

From what I see now in Documentation/atomic_ops.txt, stores to the
properly aligned memory locations are in fact atomic.

They are, but again only with respect to each other.

You are right. The architectures like "sparc" and may be others, indeed, use spinlocks to implement atomic operations, including bit manupulation.

Well then, I can only think about clearing each flag individually (with clear_bit()) instead of using "dev->flags = 0".

Something like this:

-----------------

drivers/net/usb/usbnet.c +3 -1
index 3c86b10..826eefe 100644
@@ -779,6 +790,7 @@ int usbnet_stop (struct net_device *net)
        struct usbnet           *dev = netdev_priv(net);
        struct driver_info      *info = dev->driver_info;
        int                     retval, pm;
+       int                     e;
 
        clear_bit(EVENT_DEV_OPEN, &dev->flags);
        netif_stop_queue (net);
@@ -813,7 +825,8 @@ int usbnet_stop (struct net_device *net)
         * can't flush_scheduled_work() until we drop rtnl (later),
         * else workers could deadlock; so make workers a NOP.
         */
-       dev->flags = 0;
+       for (e = 0; e < EVENT_NUM_EVENTS; ++e)
+               clear_bit(e, &dev->flags)
        del_timer_sync (&dev->delay);
        tasklet_kill (&dev->bh);
        if (!pm)
include/linux/usb/usbnet.h +0 -0
index 6e0ce8c..7ad62da 100644
@@ -79,6 +79,7 @@ struct usbnet {
# define EVENT_RX_KILL 10
# define EVENT_LINK_CHANGE 11
# define EVENT_SET_RX_MODE 12
+# define EVENT_NUM_EVENTS 13 /* Or may be keep all these in an enum? */
};

static inline struct usb_driver *driver_of(struct usb_interface *intf)
-------------------

clear_bit() is atomic w.r.t. itself and other bit ops.

Quoted 8 lines
Quoted 7 lines

So, I think, the situation you described above cannot happen for dev->flags, which is good. No need to address that in the patch. The race might be harmless after all.

If I understand the code correctly now, dev->flags is set to 0 in usbnet_stop() so that the worker function (usbnet_deferred_kevent) would

Quoted 7 lines

Yes, particularly not reschedule itself.

Quoted 2 lines

do nothing, should it start later. If so, how about adding memory barriers for all CPUs to see dev->flags is 0 before other things?

Taking a lock, as del_timer_sync() does, implies a memory barrier, as does a work.

If so, then, yes, additional barriers are not needed.

Regards,
Eugene

#12 lore raw
In reply to Oliver Neukum

Hi,

21.07.2015 17:22, Oliver Neukum пишет:

Quoted 53 lines · Eugene Shatokhin
Quoted 9 lines

And here, the code clears EVENT_RX_KILL bit in dev->flags, which may execute concurrently with the above operation:
#0 clear_bit (bitops.h:113, inlined)
#1 usbnet_bh (usbnet.c:1475)

         /* restart RX again after disabling due to high error rate */
         clear_bit(EVENT_RX_KILL, &dev->flags);

If clear_bit() is atomic w.r.t. setting dev->flags to 0, this race is not a problem, I guess. Otherwise, it may be.

clear_bit is atomic with respect to other atomic operations. So how about this:

Regards
Oliver

Quoted 1 line

From 1c4e685b3a9c183e04c46b661830e5c7ed35b513 Mon Sep 17 00:00:00 2001

From: Oliver Neukum <[email protected]>
Date: Tue, 21 Jul 2015 16:19:40 +0200
Subject: [PATCH] usbnet: fix race between usbnet_stop() and the BH

Does this do the job?

---
  drivers/net/usb/usbnet.c | 9 ++++++---
  1 file changed, 6 insertions(+), 3 deletions(-)
drivers/net/usb/usbnet.c +4 -1
index 3c86b10..77a9a86 100644
@@ -778,7 +778,7 @@ int usbnet_stop (struct net_device *net)
  {
  	struct usbnet		*dev = netdev_priv(net);
  	struct driver_info	*info = dev->driver_info;
-	int			retval, pm;
+	int			retval, pm, mpn;
 
  	clear_bit(EVENT_DEV_OPEN, &dev->flags);
  	netif_stop_queue (net);
@@ -813,14 +813,17 @@ int usbnet_stop (struct net_device *net)
  	 * can't flush_scheduled_work() until we drop rtnl (later),
  	 * else workers could deadlock; so make workers a NOP.
  	 */
+	mpn = !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags);
  	dev->flags = 0;
  	del_timer_sync (&dev->delay);
  	tasklet_kill (&dev->bh);
+	mpn |= !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags);
+	/* in case the bh reset a flag */

drivers/net/usb/usbnet.c int usbnet_stop (struct net_device *net) · line 829

Quoted 10 lines
drivers/net/usb/usbnet.c +2 -2
+	dev->flags = 0;
  	if (!pm)
  		usb_autopm_put_interface(dev->intf);
 
-	if (info->manage_power &&
-	    !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags))
+	if (info->manage_power && mpn)
  		info->manage_power(dev, 0);
  	else
  		usb_autopm_put_interface(dev->intf);

From what we have discussed here, I have combined a patch that fixes the race #1 in usbnet_stop() and makes #4 harmless by using atomics. I will send it shortly.

I had to make some adjustments (e.g. using spin_lock_nested in one place for lockdep to see it is OK to take dev->done.lock there).

I have tested the patch on the mainline kernel 4.2-rc6 built for x86-64, with the same USB modem. So far, lockdep, Kmemleak (just in case) and my tools have not detected problems in the relevant parts of the code. The device and the driver seem to work well.

So, what is your opinion?

Regards,
Eugene

#13 lore raw

PATCH 1/1 usbnet: Fix two races between usbnet_stop() and the BH

Both races may happen when a device (e.g. YOTA 4G LTE Modem) is unplugged while the system is downloading a large file from the Net.

Hardware breakpoints and Kprobes with delays were used to confirm that the races do actually happen.

1. The first race is on skb_queue ('next' pointer) between usbnet_stop() and rx_complete(), which, in turn, calls usbnet_bh().

Here is a part of the call stack with the code where the changes to the queue happen. The line numbers are for the kernel 4.1.0:

*0 __skb_unlink (skbuff.h:1517)

    prev->next = next;
*1 defer_bh (usbnet.c:430)
    spin_lock_irqsave(&list->lock, flags);
    old_state = entry->state;
    entry->state = state;
    __skb_unlink(skb, list);
    spin_unlock(&list->lock);
    spin_lock(&dev->done.lock);
    __skb_queue_tail(&dev->done, skb);
    if (dev->done.qlen == 1)
        tasklet_schedule(&dev->bh);
    spin_unlock_irqrestore(&dev->done.lock, flags);
*2 rx_complete (usbnet.c:640)
    state = defer_bh(dev, skb, &dev->rxq, state);

At the same time, the following code repeatedly checks if the queue is empty and reads these values concurrently with the above changes:

*0 usbnet_terminate_urbs (usbnet.c:765)

    /* maybe wait for deletions to finish. */
    while (!skb_queue_empty(&dev->rxq)
        && !skb_queue_empty(&dev->txq)
        && !skb_queue_empty(&dev->done)) {
            schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
            set_current_state(TASK_UNINTERRUPTIBLE);
            netif_dbg(dev, ifdown, dev->net,
                  "waited for %d urb completions\n", temp);
    }
*1  usbnet_stop (usbnet.c:806)
    if (!(info->flags & FLAG_AVOID_UNLINK_URBS))
        usbnet_terminate_urbs(dev);

As a result, it is possible, for example, that the skb is removed from dev->rxq by __skb_unlink() before the check "!skb_queue_empty(&dev->rxq)" in usbnet_terminate_urbs() is made. It is also possible in this case that the skb is added to dev->done queue after "!skb_queue_empty(&dev->done)" is checked. So usbnet_terminate_urbs() may stop waiting and return while dev->done queue still has an item.

Locking in defer_bh() and usbnet_terminate_urbs() was revisited to avoid this race.

2. The second race is on dev->flags.

dev->flags is set to 0 here:
*0 usbnet_stop (usbnet.c:816)

    /* deferred work (task, timer, softirq) must also stop.
     * can't flush_scheduled_work() until we drop rtnl (later),
     * else workers could deadlock; so make workers a NOP.
     */
    dev->flags = 0;
    del_timer_sync (&dev->delay);
    tasklet_kill (&dev->bh);

And here, the code clears EVENT_RX_KILL bit in dev->flags, which may execute concurrently with the above operation:
*0 clear_bit (bitops.h:113, inlined)
*1 usbnet_bh (usbnet.c:1475)

    /* restart RX again after disabling due to high error rate */
    clear_bit(EVENT_RX_KILL, &dev->flags);

It seems, setting dev->flags to 0 is not necessarily atomic w.r.t. clear_bit() and other bit operations with dev->flags. It is safer to make it atomic and this way, make the race harmless.

While at it, the checking of EVENT_NO_RUNTIME_PM bit of dev->flags in usbnet_stop() was fixed too: the bit should be checked before dev->flags is cleared.

---
 drivers/net/usb/usbnet.c   | 49 ++++++++++++++++++++++++++++++++--------------
 include/linux/usb/usbnet.h | 33 +++++++++++++++++++------------
 2 files changed, 54 insertions(+), 28 deletions(-)
drivers/net/usb/usbnet.c +34 -15
index 3c86b10..a53124c 100644
@@ -428,12 +428,18 @@ static enum skb_state defer_bh(struct usbnet *dev, struct sk_buff *skb,
 	old_state = entry->state;
 	entry->state = state;
 	__skb_unlink(skb, list);
-	spin_unlock(&list->lock);
-	spin_lock(&dev->done.lock);
+
+	/* defer_bh() is never called with list == &dev->done.
+	 * spin_lock_nested() tells lockdep that it is OK to take
+	 * dev->done.lock here with list->lock held. *
+	 */
+	spin_lock_nested(&dev->done.lock, SINGLE_DEPTH_NESTING);
+
 	__skb_queue_tail(&dev->done, skb);
 	if (dev->done.qlen == 1)
 		tasklet_schedule(&dev->bh);
-	spin_unlock_irqrestore(&dev->done.lock, flags);
+	spin_unlock(&dev->done.lock);
+	spin_unlock_irqrestore(&list->lock, flags);
 	return old_state;
 }
 
@@ -749,6 +755,20 @@ EXPORT_SYMBOL_GPL(usbnet_unlink_rx_urbs);
 
 /*-------------------------------------------------------------------------*/
 
+static void wait_skb_queue_empty(struct sk_buff_head *q)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&q->lock, flags);
+	while (!skb_queue_empty(q)) {
+		spin_unlock_irqrestore(&q->lock, flags);
+		schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		spin_lock_irqsave(&q->lock, flags);
+	}
+	spin_unlock_irqrestore(&q->lock, flags);
+}
+
 // precondition: never called in_interrupt
 static void usbnet_terminate_urbs(struct usbnet *dev)
 {
@@ -762,14 +782,11 @@ static void usbnet_terminate_urbs(struct usbnet *dev)
 		unlink_urbs(dev, &dev->rxq);
 
 	/* maybe wait for deletions to finish. */
-	while (!skb_queue_empty(&dev->rxq)
-		&& !skb_queue_empty(&dev->txq)
-		&& !skb_queue_empty(&dev->done)) {
-			schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
-			set_current_state(TASK_UNINTERRUPTIBLE);
-			netif_dbg(dev, ifdown, dev->net,
-				  "waited for %d urb completions\n", temp);
-	}
+	wait_skb_queue_empty(&dev->rxq);
+	wait_skb_queue_empty(&dev->txq);
+	wait_skb_queue_empty(&dev->done);
+	netif_dbg(dev, ifdown, dev->net,
+		  "waited for %d urb completions\n", temp);
 	set_current_state(TASK_RUNNING);
 	remove_wait_queue(&dev->wait, &wait);
 }
@@ -778,7 +795,8 @@ int usbnet_stop (struct net_device *net)
 {
 	struct usbnet		*dev = netdev_priv(net);
 	struct driver_info	*info = dev->driver_info;
-	int			retval, pm;
+	int			retval, pm, mpn;
+	int			e;
 
 	clear_bit(EVENT_DEV_OPEN, &dev->flags);
 	netif_stop_queue (net);
@@ -813,14 +831,15 @@ int usbnet_stop (struct net_device *net)
 	 * can't flush_scheduled_work() until we drop rtnl (later),
 	 * else workers could deadlock; so make workers a NOP.
 	 */
-	dev->flags = 0;
+	mpn = !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags);
+	for (e = 0; e < EVENT_NUM_EVENTS; ++e)
+		clear_bit(e, &dev->flags);
 	del_timer_sync (&dev->delay);
 	tasklet_kill (&dev->bh);
 	if (!pm)
 		usb_autopm_put_interface(dev->intf);
 
-	if (info->manage_power &&
-	    !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags))
+	if (info->manage_power && mpn)
 		info->manage_power(dev, 0);
 	else
 		usb_autopm_put_interface(dev->intf);
include/linux/usb/usbnet.h +20 -13
index 6e0ce8c..39a57f5 100644
@@ -22,6 +22,23 @@
 #ifndef	__LINUX_USB_USBNET_H
 #define	__LINUX_USB_USBNET_H
 
+enum usbnet_event_bits {
+	EVENT_TX_HALT =		0,
+	EVENT_RX_HALT =		1,
+	EVENT_RX_MEMORY =	2,
+	EVENT_STS_SPLIT =	3,
+	EVENT_LINK_RESET =	4,
+	EVENT_RX_PAUSED =	5,
+	EVENT_DEV_ASLEEP =	6,
+	EVENT_DEV_OPEN =	7,
+	EVENT_DEVICE_REPORT_IDLE = 8,
+	EVENT_NO_RUNTIME_PM =	9,
+	EVENT_RX_KILL =		10,
+	EVENT_LINK_CHANGE =	11,
+	EVENT_SET_RX_MODE =	12,
+	EVENT_NUM_EVENTS	/* keep it last */
+};
+
 /* interface from usbnet core to each USB networking link we handle */
 struct usbnet {
 	/* housekeeping */
@@ -65,20 +82,10 @@ struct usbnet {
 	struct tasklet_struct	bh;
 
 	struct work_struct	kevent;
+
+	/* See enum usbnet_event_bits for the meaning of the individial
+	 * bits in 'flags'. */
 	unsigned long		flags;
-#		define EVENT_TX_HALT	0
-#		define EVENT_RX_HALT	1
-#		define EVENT_RX_MEMORY	2
-#		define EVENT_STS_SPLIT	3
-#		define EVENT_LINK_RESET	4
-#		define EVENT_RX_PAUSED	5
-#		define EVENT_DEV_ASLEEP 6
-#		define EVENT_DEV_OPEN	7
-#		define EVENT_DEVICE_REPORT_IDLE	8
-#		define EVENT_NO_RUNTIME_PM	9
-#		define EVENT_RX_KILL	10
-#		define EVENT_LINK_CHANGE	11
-#		define EVENT_SET_RX_MODE	12
 };
 
 static inline struct usb_driver *driver_of(struct usb_interface *intf)
Signature · 1 line
2.3.2
maintainer
#14 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

From: Eugene Shatokhin <[email protected]>
Date: Fri, 14 Aug 2015 19:58:36 +0300

Quoted 19 lines

2. The second race is on dev->flags.

dev->flags is set to 0 here:
*0 usbnet_stop (usbnet.c:816)

    /* deferred work (task, timer, softirq) must also stop.
     * can't flush_scheduled_work() until we drop rtnl (later),
     * else workers could deadlock; so make workers a NOP.
     */
    dev->flags = 0;
    del_timer_sync (&dev->delay);
    tasklet_kill (&dev->bh);

And here, the code clears EVENT_RX_KILL bit in dev->flags, which may execute concurrently with the above operation:
*0 clear_bit (bitops.h:113, inlined)
*1 usbnet_bh (usbnet.c:1475)

    /* restart RX again after disabling due to high error rate */
    clear_bit(EVENT_RX_KILL, &dev->flags);
Quoted 7 lines

It seems, setting dev->flags to 0 is not necessarily atomic w.r.t. clear_bit() and other bit operations with dev->flags. It is safer to make it atomic and this way, make the race harmless.

While at it, the checking of EVENT_NO_RUNTIME_PM bit of dev->flags in usbnet_stop() was fixed too: the bit should be checked before dev->flags is cleared.

The fix for this is excessive.

Instead of all of this madness, looping over expensive clear_bit() atomics, just do whatever it takes to make sure that usbnet_bh() is quiesced and cannot execute any more. Then you can safely clear dev->flags normally.

#15 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

19.08.2015 04:54, David Miller пишет:

Quoted 30 lines

From: Eugene Shatokhin <[email protected]>
Date: Fri, 14 Aug 2015 19:58:36 +0300

Quoted 19 lines

2. The second race is on dev->flags.

dev->flags is set to 0 here:
*0 usbnet_stop (usbnet.c:816)

     /* deferred work (task, timer, softirq) must also stop.
      * can't flush_scheduled_work() until we drop rtnl (later),
      * else workers could deadlock; so make workers a NOP.
      */
     dev->flags = 0;
     del_timer_sync (&dev->delay);
     tasklet_kill (&dev->bh);

And here, the code clears EVENT_RX_KILL bit in dev->flags, which may execute concurrently with the above operation:
*0 clear_bit (bitops.h:113, inlined)
*1 usbnet_bh (usbnet.c:1475)

     /* restart RX again after disabling due to high error rate */
     clear_bit(EVENT_RX_KILL, &dev->flags);
Quoted 7 lines

It seems, setting dev->flags to 0 is not necessarily atomic w.r.t. clear_bit() and other bit operations with dev->flags. It is safer to make it atomic and this way, make the race harmless.

While at it, the checking of EVENT_NO_RUNTIME_PM bit of dev->flags in usbnet_stop() was fixed too: the bit should be checked before dev->flags is cleared.

Quoted 7 lines

The fix for this is excessive.

Instead of all of this madness, looping over expensive clear_bit() atomics, just do whatever it takes to make sure that usbnet_bh() is quiesced and cannot execute any more. Then you can safely clear dev->flags normally.

If I understand it correctly, it is to make sure usbnet_bh() is not scheduled again that dev->flags should be set to 0 first, one way or another. That is what this madness is for.

tasklet_kill() will wait then for the already running instance of usbnet_bh() (if one is running). After that, it is guaranteed BH is not running and will not be re-scheduled.

As for the performance concerns, I doubt that usbnet_stop() is anywhere on the critical path. I have been testing this patch for some time and haven't seen any new performance issues with it yet.

If needed, it is possible to measure and compare the time needed for usbnet_stop() before and after this patch and try to estimate the impact of this on the overall performance.

Regards,
Eugene

--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#16 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

Quoted 39 lines · Eugene Shatokhin

19.08.2015 04:54, David Miller пишет:

Quoted 30 lines

From: Eugene Shatokhin <[email protected]>
Date: Fri, 14 Aug 2015 19:58:36 +0300

Quoted 19 lines

2. The second race is on dev->flags.

dev->flags is set to 0 here:
*0 usbnet_stop (usbnet.c:816)

     /* deferred work (task, timer, softirq) must also stop.
      * can't flush_scheduled_work() until we drop rtnl (later),
      * else workers could deadlock; so make workers a NOP.
      */
     dev->flags = 0;
     del_timer_sync (&dev->delay);
     tasklet_kill (&dev->bh);

And here, the code clears EVENT_RX_KILL bit in dev->flags, which may execute concurrently with the above operation:
*0 clear_bit (bitops.h:113, inlined)
*1 usbnet_bh (usbnet.c:1475)

     /* restart RX again after disabling due to high error rate */
     clear_bit(EVENT_RX_KILL, &dev->flags);
Quoted 7 lines

It seems, setting dev->flags to 0 is not necessarily atomic w.r.t. clear_bit() and other bit operations with dev->flags. It is safer to make it atomic and this way, make the race harmless.

While at it, the checking of EVENT_NO_RUNTIME_PM bit of dev->flags in usbnet_stop() was fixed too: the bit should be checked before dev->flags is cleared.

Quoted 7 lines

The fix for this is excessive.

Instead of all of this madness, looping over expensive clear_bit() atomics, just do whatever it takes to make sure that usbnet_bh() is quiesced and cannot execute any more. Then you can safely clear dev->flags normally.

Quoted 3 lines

If I understand it correctly, it is to make sure usbnet_bh() is not scheduled again that dev->flags should be set to 0 first, one way or another. That is what this madness is for.

Assuming there is a race which may reorder these, exactly what difference does it make wrt EVENT_RX_KILL if you do

a)  clear_bit(EVENT_RX_KILL, &dev->flags);
    dev->flags = 0;

or

b)  dev->flags = 0;
    clear_bit(EVENT_RX_KILL, &dev->flags);

AFAICS, the result will be a cleared EVENT_RX_KILL bit in either case.

The EVENT_NO_RUNTIME_PM bug should definitely be fixed. Please split that out as a separate fix. It's a separate issue, and should be backported to all maintained stable releases it applies to (anything from v3.8 and newer)

Bjørn

#17 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

19.08.2015 13:54, Bjørn Mork пишет:

Quoted 51 lines
Quoted 39 lines · Eugene Shatokhin

19.08.2015 04:54, David Miller пишет:

Quoted 30 lines

From: Eugene Shatokhin <[email protected]>
Date: Fri, 14 Aug 2015 19:58:36 +0300

Quoted 19 lines

2. The second race is on dev->flags.

dev->flags is set to 0 here:
*0 usbnet_stop (usbnet.c:816)

      /* deferred work (task, timer, softirq) must also stop.
       * can't flush_scheduled_work() until we drop rtnl (later),
       * else workers could deadlock; so make workers a NOP.
       */
      dev->flags = 0;
      del_timer_sync (&dev->delay);
      tasklet_kill (&dev->bh);

And here, the code clears EVENT_RX_KILL bit in dev->flags, which may execute concurrently with the above operation:
*0 clear_bit (bitops.h:113, inlined)
*1 usbnet_bh (usbnet.c:1475)

      /* restart RX again after disabling due to high error rate */
      clear_bit(EVENT_RX_KILL, &dev->flags);
Quoted 7 lines

It seems, setting dev->flags to 0 is not necessarily atomic w.r.t. clear_bit() and other bit operations with dev->flags. It is safer to make it atomic and this way, make the race harmless.

While at it, the checking of EVENT_NO_RUNTIME_PM bit of dev->flags in usbnet_stop() was fixed too: the bit should be checked before dev->flags is cleared.

Quoted 7 lines

The fix for this is excessive.

Instead of all of this madness, looping over expensive clear_bit() atomics, just do whatever it takes to make sure that usbnet_bh() is quiesced and cannot execute any more. Then you can safely clear dev->flags normally.

Quoted 3 lines

If I understand it correctly, it is to make sure usbnet_bh() is not scheduled again that dev->flags should be set to 0 first, one way or another. That is what this madness is for.

Assuming there is a race which may reorder these, exactly what difference does it make wrt EVENT_RX_KILL if you do

a)  clear_bit(EVENT_RX_KILL, &dev->flags);
     dev->flags = 0;
Quoted 8 lines

or

b)  dev->flags = 0;
     clear_bit(EVENT_RX_KILL, &dev->flags);

AFAICS, the result will be a cleared EVENT_RX_KILL bit in either case.

Thanks for the review!

The problem is not in the reordering but rather in the fact that "dev->flags = 0" is not necessarily atomic w.r.t. "clear_bit(EVENT_RX_KILL, &dev->flags)", and vice versa.

So the following might be possible, although unlikely:

CPU0 CPU1
clear_bit: read dev->flags
clear_bit: clear EVENT_RX_KILL in the read value

dev->flags=0;

                 clear_bit: write updated dev->flags

As a result, dev->flags may become non-zero again.

I cannot prove yet that this is an impossible situation. If anyone can, please explain. If so, this part of the patch will not be needed.

Quoted 5 lines

The EVENT_NO_RUNTIME_PM bug should definitely be fixed. Please split that out as a separate fix. It's a separate issue, and should be backported to all maintained stable releases it applies to (anything from v3.8 and newer)

Yes, that makes sense. However, this fix was originally provided by Oliver Neukum rather than me, so I would like to hear his opinion as well first.

Quoted 4 lines

Bjørn

Regards,
Eugene

#18 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

Quoted 10 lines · Eugene Shatokhin

The problem is not in the reordering but rather in the fact that "dev->flags = 0" is not necessarily atomic
w.r.t. "clear_bit(EVENT_RX_KILL, &dev->flags)", and vice versa.

So the following might be possible, although unlikely:

CPU0 CPU1
clear_bit: read dev->flags
clear_bit: clear EVENT_RX_KILL in the read value

Quoted 5 lines

dev->flags=0;

                 clear_bit: write updated dev->flags

As a result, dev->flags may become non-zero again.

Ah, right. Thanks for explaining.

Quoted 2 lines

I cannot prove yet that this is an impossible situation. If anyone can, please explain. If so, this part of the patch will not be needed.

I wonder if we could simply move the dev->flags = 0 down a few lines to fix both issues? It doesn't seem to do anything useful except for resetting the flags to a sane initial state after the device is down.

Stopping the tasklet rescheduling etc depends only on netif_running(), which will be false when usbnet_stop is called. There is no need to touch dev->flags for this to happen.

Quoted 8 lines
Quoted 4 lines

The EVENT_NO_RUNTIME_PM bug should definitely be fixed. Please split that out as a separate fix. It's a separate issue, and should be backported to all maintained stable releases it applies to (anything from v3.8 and newer)

Yes, that makes sense. However, this fix was originally provided by Oliver Neukum rather than me, so I would like to hear his opinion as well first.

If what I write above is correct (please help me verify...), then maybe it does make sense to do these together anyway.

Bjørn

#19 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

19.08.2015 15:31, Bjørn Mork пишет:

Quoted 23 lines
Quoted 10 lines · Eugene Shatokhin

The problem is not in the reordering but rather in the fact that "dev->flags = 0" is not necessarily atomic
w.r.t. "clear_bit(EVENT_RX_KILL, &dev->flags)", and vice versa.

So the following might be possible, although unlikely:

CPU0 CPU1
clear_bit: read dev->flags
clear_bit: clear EVENT_RX_KILL in the read value

Quoted 5 lines

dev->flags=0;

                  clear_bit: write updated dev->flags

As a result, dev->flags may become non-zero again.

Ah, right. Thanks for explaining.

Quoted 2 lines

I cannot prove yet that this is an impossible situation. If anyone can, please explain. If so, this part of the patch will not be needed.

Quoted 7 lines

I wonder if we could simply move the dev->flags = 0 down a few lines to fix both issues? It doesn't seem to do anything useful except for resetting the flags to a sane initial state after the device is down.

Stopping the tasklet rescheduling etc depends only on netif_running(), which will be false when usbnet_stop is called. There is no need to touch dev->flags for this to happen.

That was one of the first ideas we discussed here. Unfortunately, it is probably not so simple.

Setting dev->flags to 0 makes some delayed operations do nothing and, among other things, not to reschedule usbnet_bh().

As you can see in drivers/net/usb/usbnet.c, usbnet_bh() can be called as a tasklet function and as a timer function in a number of situations (look for the usage of dev->bh and dev->delay there).

netif_running() is indeed false when usbnet_stop() runs, usbnet_stop() also disables Tx. This seems to be enough for many cases where usbnet_bh() is scheduled, but I am not so sure about the remaining ones, namely:

1. A work function, usbnet_deferred_kevent(), may reschedule usbnet_bh(). Looks like the workqueue is only stopped in usbnet_disconnect(), so a work item might be processed while usbnet_stop() works. Setting dev->flags to 0 makes the work function do nothing, by the way. See also the comment in usbnet_stop() about this.

A work item may be placed to this workqueue in a number of ways, by both usbnet module and the mini-drivers. It is not too easy to track all these situations.

2. rx_complete() and tx_complete() may schedule execution of usbnet_bh() as a tasklet or a timer function. These two are URB completion callbacks.

It seems, new Rx and Tx URBs cannot be submitted when usbnet_stop() clears dev->flags, indeed. But it does not prevent the completion handlers for the previously submitted URBs from running concurrently with usbnet_stop(). The latter waits for them to complete (via usbnet_terminate_urbs(dev)) but only if FLAG_AVOID_UNLINK_URBS is not set in info->flags. rndis_wlan, however, sets this flag for a few hardware models. So - no guarantees here as well.

If someone could list the particular bits of dev->flags that should be cleared to make sure no deferred call could reschedule usbnet_bh(), etc... Well, it would be enough to clear these first and use dev->flags = 0 later, after tasklet_kill() and del_timer_sync(). I cannot point out these particular bits now.

Besides, it is possible, although unlikely, that new event bits will be added to dev->flags in the future. And one will need to keep track of these to see if they should be cleared as well. I'd prever to play safer for now and clear them all.

Quoted 6 lines
Quoted 6 lines
Quoted 5 lines

The EVENT_NO_RUNTIME_PM bug should definitely be fixed. Please split that out as a separate fix. It's a separate issue, and should be backported to all maintained stable releases it applies to (anything from v3.8 and newer)

Quoted 6 lines
Quoted 3 lines

Yes, that makes sense. However, this fix was originally provided by Oliver Neukum rather than me, so I would like to hear his opinion as well first.

If what I write above is correct (please help me verify...), then maybe it does make sense to do these together anyway.

I think, your suggestion to make it a separate patch is right. Will do it in the next version of the patchset, hopefully soon.

Regards,
Eugene

#20 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

Quoted 32 lines · Eugene Shatokhin

19.08.2015 15:31, Bjørn Mork пишет:

Quoted 23 lines
Quoted 10 lines · Eugene Shatokhin

The problem is not in the reordering but rather in the fact that "dev->flags = 0" is not necessarily atomic
w.r.t. "clear_bit(EVENT_RX_KILL, &dev->flags)", and vice versa.

So the following might be possible, although unlikely:

CPU0 CPU1
clear_bit: read dev->flags
clear_bit: clear EVENT_RX_KILL in the read value

Quoted 5 lines

dev->flags=0;

                  clear_bit: write updated dev->flags

As a result, dev->flags may become non-zero again.

Ah, right. Thanks for explaining.

Quoted 2 lines

I cannot prove yet that this is an impossible situation. If anyone can, please explain. If so, this part of the patch will not be needed.

Quoted 7 lines

I wonder if we could simply move the dev->flags = 0 down a few lines to fix both issues? It doesn't seem to do anything useful except for resetting the flags to a sane initial state after the device is down.

Stopping the tasklet rescheduling etc depends only on netif_running(), which will be false when usbnet_stop is called. There is no need to touch dev->flags for this to happen.

Quoted 5 lines

That was one of the first ideas we discussed here. Unfortunately, it is probably not so simple.

Setting dev->flags to 0 makes some delayed operations do nothing and, among other things, not to reschedule usbnet_bh().

Yes, but I believe that is merely a side effect. You should never need to clear multiple flags to get the desired behaviour.

Quoted 16 lines

As you can see in drivers/net/usb/usbnet.c, usbnet_bh() can be called as a tasklet function and as a timer function in a number of situations (look for the usage of dev->bh and dev->delay there).

netif_running() is indeed false when usbnet_stop() runs, usbnet_stop() also disables Tx. This seems to be enough for many cases where usbnet_bh() is scheduled, but I am not so sure about the remaining ones, namely:

1. A work function, usbnet_deferred_kevent(), may reschedule usbnet_bh(). Looks like the workqueue is only stopped in usbnet_disconnect(), so a work item might be processed while usbnet_stop() works. Setting dev->flags to 0 makes the work function do nothing, by the way. See also the comment in usbnet_stop() about this.

Quoted 3 lines

A work item may be placed to this workqueue in a number of ways, by both usbnet module and the mini-drivers. It is not too easy to track all these situations.

That's an understatement :)

Quoted 11 lines

2. rx_complete() and tx_complete() may schedule execution of usbnet_bh() as a tasklet or a timer function. These two are URB completion callbacks.

It seems, new Rx and Tx URBs cannot be submitted when usbnet_stop() clears dev->flags, indeed. But it does not prevent the completion handlers for the previously submitted URBs from running concurrently with usbnet_stop(). The latter waits for them to complete (via usbnet_terminate_urbs(dev)) but only if FLAG_AVOID_UNLINK_URBS is not set in info->flags. rndis_wlan, however, sets this flag for a few hardware models. So - no guarantees here as well.

FLAG_AVOID_UNLINK_URBS looks like it should be replaced by the newer ability to keep the status urb active. I believe that must have been the real reason for adding it, based on the commit message and the effect the flag will have:

commit 1487cd5e76337555737cbc55d7d83f41460d198f
Author: Jussi Kivilinna <[email protected]>
Date: Thu Jul 30 19:41:20 2009 +0300

usbnet: allow "minidriver" to prevent urb unlinking on usbnet_stop

rndis_wlan devices freeze after running usbnet_stop several times. It appears that firmware freezes in state where it does not respond to any RNDIS commands and device have to be physically unplugged/replugged. This patch lets minidrivers to disable unlink_urbs on usbnet_stop through new info flag.

The rx urbs will not be resubmitted in any case, and there are of course no tx urbs being submitted. So the only effect of this flag is on the status/interrupt urb, which I can imagine some RNDIS devices wants active all the time.

So FLAG_AVOID_UNLINK_URBS should probably be removed and replaced calls to usbnet_status_start() and usbnet_status_stop(). This will require testing on some of the devices with the original firmware problem however.

In any case: I do not think this flag should be considered when trying to make usbnet_stop behaviour saner. It's only purpose is to deliberately break usbnet_stop by not actually stopping.

Quoted 5 lines

If someone could list the particular bits of dev->flags that should be cleared to make sure no deferred call could reschedule usbnet_bh(), etc... Well, it would be enough to clear these first and use dev->flags = 0 later, after tasklet_kill() and del_timer_sync(). I cannot point out these particular bits now.

I don't think any of the flags must be cleared. The sequence

        dev_close(dev->net);
	usbnet_terminate_urbs(dev);
	usbnet_purge_paused_rxq(dev);
	del_timer_sync (&dev->delay);
	tasklet_kill (&dev->bh);

should prevent any rescheduling of usbnet_bh

Quoted 4 lines

Besides, it is possible, although unlikely, that new event bits will be added to dev->flags in the future. And one will need to keep track of these to see if they should be cleared as well. I'd prever to play safer for now and clear them all.

I don't think we should ever make a flag which will _have_ to be reset for usbnet_stop. The only reason for clearing all flags is to reset the state before the next open.

Yes, I see that we currently need to clear EVENT_DEV_OPEN in usbnet_stop, but I really don't see what this flag gives us which isn't already provided by netif_running(). It looks like a duplicate.

Bjørn

#21 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

24.08.2015 16:29, Bjørn Mork пишет:

Quoted 125 lines
Quoted 32 lines · Eugene Shatokhin

19.08.2015 15:31, Bjørn Mork пишет:

Quoted 23 lines
Quoted 10 lines · Eugene Shatokhin

The problem is not in the reordering but rather in the fact that "dev->flags = 0" is not necessarily atomic
w.r.t. "clear_bit(EVENT_RX_KILL, &dev->flags)", and vice versa.

So the following might be possible, although unlikely:

CPU0 CPU1
clear_bit: read dev->flags
clear_bit: clear EVENT_RX_KILL in the read value

Quoted 5 lines

dev->flags=0;

                   clear_bit: write updated dev->flags

As a result, dev->flags may become non-zero again.

Ah, right. Thanks for explaining.

Quoted 2 lines

I cannot prove yet that this is an impossible situation. If anyone can, please explain. If so, this part of the patch will not be needed.

Quoted 7 lines

I wonder if we could simply move the dev->flags = 0 down a few lines to fix both issues? It doesn't seem to do anything useful except for resetting the flags to a sane initial state after the device is down.

Stopping the tasklet rescheduling etc depends only on netif_running(), which will be false when usbnet_stop is called. There is no need to touch dev->flags for this to happen.

Quoted 5 lines

That was one of the first ideas we discussed here. Unfortunately, it is probably not so simple.

Setting dev->flags to 0 makes some delayed operations do nothing and, among other things, not to reschedule usbnet_bh().

Yes, but I believe that is merely a side effect. You should never need to clear multiple flags to get the desired behaviour.

Quoted 16 lines

As you can see in drivers/net/usb/usbnet.c, usbnet_bh() can be called as a tasklet function and as a timer function in a number of situations (look for the usage of dev->bh and dev->delay there).

netif_running() is indeed false when usbnet_stop() runs, usbnet_stop() also disables Tx. This seems to be enough for many cases where usbnet_bh() is scheduled, but I am not so sure about the remaining ones, namely:

1. A work function, usbnet_deferred_kevent(), may reschedule usbnet_bh(). Looks like the workqueue is only stopped in usbnet_disconnect(), so a work item might be processed while usbnet_stop() works. Setting dev->flags to 0 makes the work function do nothing, by the way. See also the comment in usbnet_stop() about this.

Quoted 3 lines

A work item may be placed to this workqueue in a number of ways, by both usbnet module and the mini-drivers. It is not too easy to track all these situations.

That's an understatement :)

Quoted 11 lines

2. rx_complete() and tx_complete() may schedule execution of usbnet_bh() as a tasklet or a timer function. These two are URB completion callbacks.

It seems, new Rx and Tx URBs cannot be submitted when usbnet_stop() clears dev->flags, indeed. But it does not prevent the completion handlers for the previously submitted URBs from running concurrently with usbnet_stop(). The latter waits for them to complete (via usbnet_terminate_urbs(dev)) but only if FLAG_AVOID_UNLINK_URBS is not set in info->flags. rndis_wlan, however, sets this flag for a few hardware models. So - no guarantees here as well.

FLAG_AVOID_UNLINK_URBS looks like it should be replaced by the newer ability to keep the status urb active. I believe that must have been the real reason for adding it, based on the commit message and the effect the flag will have:

commit 1487cd5e76337555737cbc55d7d83f41460d198f
Author: Jussi Kivilinna <[email protected]>
Date: Thu Jul 30 19:41:20 2009 +0300

usbnet: allow "minidriver" to prevent urb unlinking on usbnet_stop

rndis_wlan devices freeze after running usbnet_stop several times. It appears that firmware freezes in state where it does not respond to any RNDIS commands and device have to be physically unplugged/replugged. This patch lets minidrivers to disable unlink_urbs on usbnet_stop through new info flag.

The rx urbs will not be resubmitted in any case, and there are of course no tx urbs being submitted. So the only effect of this flag is on the status/interrupt urb, which I can imagine some RNDIS devices wants active all the time.

So FLAG_AVOID_UNLINK_URBS should probably be removed and replaced calls to usbnet_status_start() and usbnet_status_stop(). This will require testing on some of the devices with the original firmware problem however.

In any case: I do not think this flag should be considered when trying to make usbnet_stop behaviour saner. It's only purpose is to deliberately break usbnet_stop by not actually stopping.

Quoted 5 lines

If someone could list the particular bits of dev->flags that should be cleared to make sure no deferred call could reschedule usbnet_bh(), etc... Well, it would be enough to clear these first and use dev->flags = 0 later, after tasklet_kill() and del_timer_sync(). I cannot point out these particular bits now.

I don't think any of the flags must be cleared. The sequence

Quoted 7 lines
         dev_close(dev->net);
	usbnet_terminate_urbs(dev);
	usbnet_purge_paused_rxq(dev);
	del_timer_sync (&dev->delay);
	tasklet_kill (&dev->bh);

should prevent any rescheduling of usbnet_bh

If so, then, I suppose, one could ignore that FLAG_AVOID_UNLINK_URBS for now and just move dev->flags = 0 down as you suggested and as we thought before.

The patch will become simpler, indeed.

Quoted 6 lines
Quoted 5 lines

Besides, it is possible, although unlikely, that new event bits will be added to dev->flags in the future. And one will need to keep track of these to see if they should be cleared as well. I'd prever to play safer for now and clear them all.

Quoted 8 lines

I don't think we should ever make a flag which will _have_ to be reset for usbnet_stop. The only reason for clearing all flags is to reset the state before the next open.

Yes, I see that we currently need to clear EVENT_DEV_OPEN in usbnet_stop, but I really don't see what this flag gives us which isn't already provided by netif_running(). It looks like a duplicate.

Regards,
Eugene

maintainer
#22 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

In reply to Bjørn Mork
Quoted 9 lines · Bjørn Mork
Quoted 7 lines · Eugene Shatokhin

19.08.2015 15:31, Bjørn Mork пишет:

Quoted 5 lines

Eugene Shatokhin <[email protected]> writes:

Stopping the tasklet rescheduling etc depends only on netif_running(), which will be false when usbnet_stop is called. There is no need to touch dev->flags for this to happen.

Quoted 8 lines
Quoted 5 lines

That was one of the first ideas we discussed here. Unfortunately, it is probably not so simple.

Setting dev->flags to 0 makes some delayed operations do nothing and, among other things, not to reschedule usbnet_bh().

Yes, but I believe that is merely a side effect. You should never need to clear multiple flags to get the desired behaviour.

Why? Is there any reason you cannot have a TX and an RX halt at the same time?

Quoted 9 lines
Quoted 9 lines

As you can see in drivers/net/usb/usbnet.c, usbnet_bh() can be called as a tasklet function and as a timer function in a number of situations (look for the usage of dev->bh and dev->delay there).

netif_running() is indeed false when usbnet_stop() runs, usbnet_stop() also disables Tx. This seems to be enough for many cases where usbnet_bh() is scheduled, but I am not so sure about the remaining ones, namely:

Quoted 6 lines
Quoted 6 lines

1. A work function, usbnet_deferred_kevent(), may reschedule usbnet_bh(). Looks like the workqueue is only stopped in usbnet_disconnect(), so a work item might be processed while usbnet_stop() works. Setting dev->flags to 0 makes the work function do nothing, by the way. See also the comment in usbnet_stop() about this.

Yes, this is the main reason the flags are collectively cleared. We could do them all with clear_bit(). Ugly though.

Quoted 5 lines
Quoted 3 lines

A work item may be placed to this workqueue in a number of ways, by both usbnet module and the mini-drivers. It is not too easy to track all these situations.

That's an understatement :)

Yes.

Quoted 4 lines

So FLAG_AVOID_UNLINK_URBS should probably be removed and replaced calls to usbnet_status_start() and usbnet_status_stop(). This will require testing on some of the devices with the original firmware problem however.

And there you point out the main problem.

Quoted 3 lines

In any case: I do not think this flag should be considered when trying to make usbnet_stop behaviour saner. It's only purpose is to deliberately break usbnet_stop by not actually stopping.

Yes.

Regards
Oliver

maintainer
#23 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

In reply to Eugene Shatokhin

From: Eugene Shatokhin <[email protected]>
Date: Wed, 19 Aug 2015 14:59:01 +0300

Quoted 11 lines

So the following might be possible, although unlikely:

CPU0 CPU1
clear_bit: read dev->flags
clear_bit: clear EVENT_RX_KILL in the read value

dev->flags=0;

                 clear_bit: write updated dev->flags

As a result, dev->flags may become non-zero again.

Is this really possible?

Stores really are "atomic" in the sense that the do their update in one indivisible operation.

Atomic operations like clear_bit also will behave that way.

If a clear_bit is in progress, the "dev->flags=0" store will not be able to grab the cache line exclusively until the clear_bit is done.

So I think the above sequent of events is completely impossible. Once a clear_bit starts, a write by another foreign agent on the bus is absolutely impossible to legally occur until the clear_bit completes.

I think this is a non-issue.

#24 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

Quoted 11 lines · David Miller

From: Eugene Shatokhin <[email protected]>
Date: Wed, 19 Aug 2015 14:59:01 +0300

Quoted 8 lines

So the following might be possible, although unlikely:

CPU0 CPU1
clear_bit: read dev->flags
clear_bit: clear EVENT_RX_KILL in the read value

dev->flags=0;

Quoted 8 lines
Quoted 3 lines
                 clear_bit: write updated dev->flags

As a result, dev->flags may become non-zero again.

Is this really possible?

Stores really are "atomic" in the sense that the do their update in one indivisible operation.

Provided you use ACCESS_ONCE or WRITE_ONCE or whatever people like to call it now.

Quoted 1 line

Atomic operations like clear_bit also will behave that way.

Are you certain about that? I couldn't find any mention of it in Documentation/atomic_ops.txt.

In theory, an architecture could implement atomic bit operations using a spinlock to insure atomicity. I don't know if any architectures do this, but if they do then the scenario above could arise.

Alan Stern

--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#25 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

Quoted 25 lines · Alan Stern
Quoted 11 lines · David Miller

From: Eugene Shatokhin <[email protected]>
Date: Wed, 19 Aug 2015 14:59:01 +0300

Quoted 8 lines

So the following might be possible, although unlikely:

CPU0 CPU1
clear_bit: read dev->flags
clear_bit: clear EVENT_RX_KILL in the read value

dev->flags=0;

Quoted 8 lines
Quoted 3 lines
                 clear_bit: write updated dev->flags

As a result, dev->flags may become non-zero again.

Is this really possible?

Stores really are "atomic" in the sense that the do their update in one indivisible operation.

Provided you use ACCESS_ONCE or WRITE_ONCE or whatever people like to call it now.

Quoted 8 lines
Quoted 1 line

Atomic operations like clear_bit also will behave that way.

Are you certain about that? I couldn't find any mention of it in Documentation/atomic_ops.txt.

In theory, an architecture could implement atomic bit operations using a spinlock to insure atomicity. I don't know if any architectures do this, but if they do then the scenario above could arise.

Now that I see this in writing, I realize it's not possible after all. clear_bit() et al. will work with a single unsigned long, which doesn't leave any place for spinlocks or other mechanisms. I was thinking of atomic_t.

So never mind...

Alan Stern

maintainer
#26 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

Quoted 4 lines · Alan Stern
Quoted 1 line

In theory, an architecture could implement atomic bit operations

using

Quoted 1 line

a spinlock to insure atomicity. I don't know if any architectures

do

Quoted 8 lines
Quoted 1 line

this, but if they do then the scenario above could arise.

Now that I see this in writing, I realize it's not possible after all.
clear_bit() et al. will work with a single unsigned long, which doesn't
leave any place for spinlocks or other mechanisms. I was thinking of atomic_t.

Refuting yourself you are making the assumption that the lock has to be inside the data structure. That is not true.

Regards
Oliver

maintainer
#27 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

In reply to Alan Stern

From: Alan Stern <[email protected]>
Date: Mon, 24 Aug 2015 14:06:15 -0400 (EDT)

Quoted 9 lines · David Miller
Quoted 1 line

Atomic operations like clear_bit also will behave that way.

Are you certain about that? I couldn't find any mention of it in Documentation/atomic_ops.txt.

In theory, an architecture could implement atomic bit operations using a spinlock to insure atomicity. I don't know if any architectures do this, but if they do then the scenario above could arise.

Indeed, we do have platforms like 32-bit sparc and parisc that do this.

So, taking that into consideration, this is a bit unfortunate and on such platforms we do have this problem.

#28 lore raw

Re: [PATCH] usbnet: Fix two races between usbnet_stop() and the BH

In reply to David Miller

24.08.2015 20:43, David Miller пишет:

Quoted 9 lines

From: Eugene Shatokhin <[email protected]>
Date: Wed, 19 Aug 2015 14:59:01 +0300

Quoted 6 lines

So the following might be possible, although unlikely:

CPU0 CPU1
clear_bit: read dev->flags
clear_bit: clear EVENT_RX_KILL in the read value

Quoted 7 lines
Quoted 5 lines

dev->flags=0;

                  clear_bit: write updated dev->flags

As a result, dev->flags may become non-zero again.

Is this really possible?

On x86, it is not possible, so this is not a problem. Perhaps, for ARM too. As for the other architectures supported by the kernel - not sure, no common guarantees, it seems. Anyway, this is not a critical issue, I agree.

OK, let us leave things as they are for this one and fix the rest.

Quoted 6 lines

Stores really are "atomic" in the sense that the do their update in one indivisible operation.

Atomic operations like clear_bit also will behave that way.

Quoted 9 lines

If a clear_bit is in progress, the "dev->flags=0" store will not be able to grab the cache line exclusively until the clear_bit is done.

So I think the above sequent of events is completely impossible. Once a clear_bit starts, a write by another foreign agent on the bus is absolutely impossible to legally occur until the clear_bit completes.

I think this is a non-issue.

Regards,
Eugene

maintainer
#29 lore raw
In reply to Eugene Shatokhin
Quoted 6 lines · Eugene Shatokhin

[Race #5]

Race on dev->rx_urb_size. I reproduced it a similar way as the races #2
and #3 (changing MTU while downloading files).

Quoted 7 lines

dev->rx_urb_size is written to here:
#0 usbnet_change_mtu (usbnet.c:392)

        dev->rx_urb_size = dev->hard_mtu;

Here is the conflicting read from dev->rx_urb_size:
* rx_submit (usbnet.c:467)

        size_t                  size = dev->rx_urb_size;

Yes, but what is the actual bad race? I mean, if you change the MTU in operation, there will be a race. That is just unavoidable.
Do we generate errors?

Regards
Oliver

--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#30 lore raw

23.07.2015 12:43, Oliver Neukum пишет:

Quoted 11 lines · Eugene Shatokhin
Quoted 10 lines

[Race #5]

Race on dev->rx_urb_size. I reproduced it a similar way as the races #2
and #3 (changing MTU while downloading files).

dev->rx_urb_size is written to here:
#0 usbnet_change_mtu (usbnet.c:392)

         dev->rx_urb_size = dev->hard_mtu;
Quoted 8 lines
Quoted 3 lines

Here is the conflicting read from dev->rx_urb_size:
* rx_submit (usbnet.c:467)

         size_t                  size = dev->rx_urb_size;

Yes, but what is the actual bad race? I mean, if you change the MTU in operation, there will be a race. That is just unavoidable.
Do we generate errors?

No, I have observed no problems due to this race so far.

Regards,
Eugene

#31 lore raw

cover usbnet: Fix 2 problems in usbnet_stop()

In reply to Eugene Shatokhin

The following problems found when investigating races in usbnet module are fixed here:

1. EVENT_NO_RUNTIME_PM bit of dev->flags should be read before it is cleared by "dev->flags = 0". Thanks to Oliver Neukum for spotting this problem and providing a fix.

2. A race on on skb_queue between usbnet_stop() and usbnet_bh().

Compared to the combined patch I sent earlier
("[PATCH] usbnet: Fix two races between usbnet_stop() and the BH"), this patch set has the following changes:

* The fix for handling of EVENT_NO_RUNTIME_PM is now in a separate patch.
* The fix for the race on dev->flags has been removed because the race is not considered harmful.

Regards,
Eugene

#32 lore raw

PATCH 1/2 usbnet: Get EVENT_NO_RUNTIME_PM bit before it is cleared

It is needed to check EVENT_NO_RUNTIME_PM bit of dev->flags in usbnet_stop(), but its value should be read before it is cleared when dev->flags is set to 0.

The problem was spotted and the fix was provided by Oliver Neukum <[email protected]>.

---
 drivers/net/usb/usbnet.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/net/usb/usbnet.c +4 -3
index 3c86b10..e049857 100644
@@ -778,7 +778,7 @@ int usbnet_stop (struct net_device *net)
 {
 	struct usbnet		*dev = netdev_priv(net);
 	struct driver_info	*info = dev->driver_info;
-	int			retval, pm;
+	int			retval, pm, mpn;
 
 	clear_bit(EVENT_DEV_OPEN, &dev->flags);
 	netif_stop_queue (net);
@@ -809,6 +809,8 @@ int usbnet_stop (struct net_device *net)
 
 	usbnet_purge_paused_rxq(dev);
 
+	mpn = !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags);
+
 	/* deferred work (task, timer, softirq) must also stop.
 	 * can't flush_scheduled_work() until we drop rtnl (later),
 	 * else workers could deadlock; so make workers a NOP.
@@ -819,8 +821,7 @@ int usbnet_stop (struct net_device *net)
 	if (!pm)
 		usb_autopm_put_interface(dev->intf);
 
-	if (info->manage_power &&
-	    !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags))
+	if (info->manage_power && mpn)
 		info->manage_power(dev, 0);
 	else
 		usb_autopm_put_interface(dev->intf);
Signature · 1 line
2.3.2
maintainer
#33 lore raw

Re: [PATCH 1/2] usbnet: Get EVENT_NO_RUNTIME_PM bit before it is cleared

Quoted 3 lines · Eugene Shatokhin

It is needed to check EVENT_NO_RUNTIME_PM bit of dev->flags in usbnet_stop(), but its value should be read before it is cleared when dev->flags is set to 0.

Can we agree that this at least is good and should go upstream and into stable?

Regards
Oliver

#34 lore raw

Re: [PATCH 1/2] usbnet: Get EVENT_NO_RUNTIME_PM bit before it is cleared

Quoted 7 lines · Oliver Neukum
Quoted 3 lines · Eugene Shatokhin

It is needed to check EVENT_NO_RUNTIME_PM bit of dev->flags in usbnet_stop(), but its value should be read before it is cleared when dev->flags is set to 0.

Can we agree that this at least is good and should go upstream and into stable?

I definitely agree.

Bjørn
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

Acked-by on v1 1/2
#35 lore raw

Re: [PATCH 1/2] usbnet: Get EVENT_NO_RUNTIME_PM bit before it is cleared

In reply to Eugene Shatokhin
Quoted 8 lines · Eugene Shatokhin

It is needed to check EVENT_NO_RUNTIME_PM bit of dev->flags in usbnet_stop(), but its value should be read before it is cleared when dev->flags is set to 0.

The problem was spotted and the fix was provided by Oliver Neukum <[email protected]>.

maintainer
#36 lore raw

Re: [PATCH 1/2] usbnet: Get EVENT_NO_RUNTIME_PM bit before it is cleared

In reply to Eugene Shatokhin

From: Eugene Shatokhin <[email protected]>
Date: Mon, 24 Aug 2015 23:13:42 +0300

Quoted 8 lines

It is needed to check EVENT_NO_RUNTIME_PM bit of dev->flags in usbnet_stop(), but its value should be read before it is cleared when dev->flags is set to 0.

The problem was spotted and the fix was provided by Oliver Neukum <[email protected]>.

Applied and queued up for -stable.

#37 lore raw

PATCH 2/2 usbnet: Fix a race between usbnet_stop() and the BH

In reply to Eugene Shatokhin

The race may happen when a device (e.g. YOTA 4G LTE Modem) is unplugged while the system is downloading a large file from the Net.

Hardware breakpoints and Kprobes with delays were used to confirm that the race does actually happen.

The race is on skb_queue ('next' pointer) between usbnet_stop() and rx_complete(), which, in turn, calls usbnet_bh().

Here is a part of the call stack with the code where the changes to the queue happen. The line numbers are for the kernel 4.1.0:

*0 __skb_unlink (skbuff.h:1517)

    prev->next = next;
*1 defer_bh (usbnet.c:430)
    spin_lock_irqsave(&list->lock, flags);
    old_state = entry->state;
    entry->state = state;
    __skb_unlink(skb, list);
    spin_unlock(&list->lock);
    spin_lock(&dev->done.lock);
    __skb_queue_tail(&dev->done, skb);
    if (dev->done.qlen == 1)
        tasklet_schedule(&dev->bh);
    spin_unlock_irqrestore(&dev->done.lock, flags);
*2 rx_complete (usbnet.c:640)
    state = defer_bh(dev, skb, &dev->rxq, state);

At the same time, the following code repeatedly checks if the queue is empty and reads these values concurrently with the above changes:

*0 usbnet_terminate_urbs (usbnet.c:765)

    /* maybe wait for deletions to finish. */
    while (!skb_queue_empty(&dev->rxq)
        && !skb_queue_empty(&dev->txq)
        && !skb_queue_empty(&dev->done)) {
            schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
            set_current_state(TASK_UNINTERRUPTIBLE);
            netif_dbg(dev, ifdown, dev->net,
                  "waited for %d urb completions\n", temp);
    }
*1  usbnet_stop (usbnet.c:806)
    if (!(info->flags & FLAG_AVOID_UNLINK_URBS))
        usbnet_terminate_urbs(dev);

As a result, it is possible, for example, that the skb is removed from dev->rxq by __skb_unlink() before the check "!skb_queue_empty(&dev->rxq)" in usbnet_terminate_urbs() is made. It is also possible in this case that the skb is added to dev->done queue after "!skb_queue_empty(&dev->done)" is checked. So usbnet_terminate_urbs() may stop waiting and return while dev->done queue still has an item.

Locking in defer_bh() and usbnet_terminate_urbs() was revisited to avoid this race.

---
 drivers/net/usb/usbnet.c | 39 ++++++++++++++++++++++++++++-----------
 1 file changed, 28 insertions(+), 11 deletions(-)
drivers/net/usb/usbnet.c +28 -11
index e049857..b4cf107 100644
@@ -428,12 +428,18 @@ static enum skb_state defer_bh(struct usbnet *dev, struct sk_buff *skb,
 	old_state = entry->state;
 	entry->state = state;
 	__skb_unlink(skb, list);
-	spin_unlock(&list->lock);
-	spin_lock(&dev->done.lock);
+
+	/* defer_bh() is never called with list == &dev->done.
+	 * spin_lock_nested() tells lockdep that it is OK to take
+	 * dev->done.lock here with list->lock held.
+	 */
+	spin_lock_nested(&dev->done.lock, SINGLE_DEPTH_NESTING);
+
 	__skb_queue_tail(&dev->done, skb);
 	if (dev->done.qlen == 1)
 		tasklet_schedule(&dev->bh);
-	spin_unlock_irqrestore(&dev->done.lock, flags);
+	spin_unlock(&dev->done.lock);
+	spin_unlock_irqrestore(&list->lock, flags);
 	return old_state;
 }
 
@@ -749,6 +755,20 @@ EXPORT_SYMBOL_GPL(usbnet_unlink_rx_urbs);
 
 /*-------------------------------------------------------------------------*/
 
+static void wait_skb_queue_empty(struct sk_buff_head *q)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&q->lock, flags);
+	while (!skb_queue_empty(q)) {
+		spin_unlock_irqrestore(&q->lock, flags);
+		schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		spin_lock_irqsave(&q->lock, flags);
+	}
+	spin_unlock_irqrestore(&q->lock, flags);
+}
+
 // precondition: never called in_interrupt
 static void usbnet_terminate_urbs(struct usbnet *dev)
 {
@@ -762,14 +782,11 @@ static void usbnet_terminate_urbs(struct usbnet *dev)
 		unlink_urbs(dev, &dev->rxq);
 
 	/* maybe wait for deletions to finish. */
-	while (!skb_queue_empty(&dev->rxq)
-		&& !skb_queue_empty(&dev->txq)
-		&& !skb_queue_empty(&dev->done)) {
-			schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
-			set_current_state(TASK_UNINTERRUPTIBLE);
-			netif_dbg(dev, ifdown, dev->net,
-				  "waited for %d urb completions\n", temp);
-	}
+	wait_skb_queue_empty(&dev->rxq);
+	wait_skb_queue_empty(&dev->txq);
+	wait_skb_queue_empty(&dev->done);
+	netif_dbg(dev, ifdown, dev->net,
+		  "waited for %d urb completions\n", temp);
 	set_current_state(TASK_RUNNING);
 	remove_wait_queue(&dev->wait, &wait);
 }
Signature · 1 line
2.3.2
#38 lore raw

Re: [PATCH 2/2] usbnet: Fix a race between usbnet_stop() and the BH

Quoted 45 lines · Eugene Shatokhin

The race may happen when a device (e.g. YOTA 4G LTE Modem) is unplugged while the system is downloading a large file from the Net.

Hardware breakpoints and Kprobes with delays were used to confirm that the race does actually happen.

The race is on skb_queue ('next' pointer) between usbnet_stop() and rx_complete(), which, in turn, calls usbnet_bh().

Here is a part of the call stack with the code where the changes to the queue happen. The line numbers are for the kernel 4.1.0:

*0 __skb_unlink (skbuff.h:1517)

    prev->next = next;
*1 defer_bh (usbnet.c:430)
    spin_lock_irqsave(&list->lock, flags);
    old_state = entry->state;
    entry->state = state;
    __skb_unlink(skb, list);
    spin_unlock(&list->lock);
    spin_lock(&dev->done.lock);
    __skb_queue_tail(&dev->done, skb);
    if (dev->done.qlen == 1)
        tasklet_schedule(&dev->bh);
    spin_unlock_irqrestore(&dev->done.lock, flags);
*2 rx_complete (usbnet.c:640)
    state = defer_bh(dev, skb, &dev->rxq, state);

At the same time, the following code repeatedly checks if the queue is empty and reads these values concurrently with the above changes:

*0 usbnet_terminate_urbs (usbnet.c:765)

    /* maybe wait for deletions to finish. */
    while (!skb_queue_empty(&dev->rxq)
        && !skb_queue_empty(&dev->txq)
        && !skb_queue_empty(&dev->done)) {
            schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
            set_current_state(TASK_UNINTERRUPTIBLE);
            netif_dbg(dev, ifdown, dev->net,
                  "waited for %d urb completions\n", temp);
    }
*1  usbnet_stop (usbnet.c:806)
    if (!(info->flags & FLAG_AVOID_UNLINK_URBS))
        usbnet_terminate_urbs(dev);
Quoted 7 lines

As a result, it is possible, for example, that the skb is removed from dev->rxq by __skb_unlink() before the check "!skb_queue_empty(&dev->rxq)" in usbnet_terminate_urbs() is made. It is also possible in this case that the skb is added to dev->done queue after "!skb_queue_empty(&dev->done)" is checked. So usbnet_terminate_urbs() may stop waiting and return while dev->done queue still has an item.

Exactly what problem will that result in? The tasklet_kill() will wait for the processing of the single element done queue, and everything will be fine. Or?

Bjørn

#39 lore raw

Re: [PATCH 2/2] usbnet: Fix a race between usbnet_stop() and the BH

25.08.2015 00:01, Bjørn Mork пишет:

Quoted 55 lines
Quoted 45 lines · Eugene Shatokhin

The race may happen when a device (e.g. YOTA 4G LTE Modem) is unplugged while the system is downloading a large file from the Net.

Hardware breakpoints and Kprobes with delays were used to confirm that the race does actually happen.

The race is on skb_queue ('next' pointer) between usbnet_stop() and rx_complete(), which, in turn, calls usbnet_bh().

Here is a part of the call stack with the code where the changes to the queue happen. The line numbers are for the kernel 4.1.0:

*0 __skb_unlink (skbuff.h:1517)

     prev->next = next;
*1 defer_bh (usbnet.c:430)
     spin_lock_irqsave(&list->lock, flags);
     old_state = entry->state;
     entry->state = state;
     __skb_unlink(skb, list);
     spin_unlock(&list->lock);
     spin_lock(&dev->done.lock);
     __skb_queue_tail(&dev->done, skb);
     if (dev->done.qlen == 1)
         tasklet_schedule(&dev->bh);
     spin_unlock_irqrestore(&dev->done.lock, flags);
*2 rx_complete (usbnet.c:640)
     state = defer_bh(dev, skb, &dev->rxq, state);

At the same time, the following code repeatedly checks if the queue is empty and reads these values concurrently with the above changes:

*0 usbnet_terminate_urbs (usbnet.c:765)

     /* maybe wait for deletions to finish. */
     while (!skb_queue_empty(&dev->rxq)
         && !skb_queue_empty(&dev->txq)
         && !skb_queue_empty(&dev->done)) {
             schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
             set_current_state(TASK_UNINTERRUPTIBLE);
             netif_dbg(dev, ifdown, dev->net,
                   "waited for %d urb completions\n", temp);
     }
*1  usbnet_stop (usbnet.c:806)
     if (!(info->flags & FLAG_AVOID_UNLINK_URBS))
         usbnet_terminate_urbs(dev);
Quoted 7 lines

As a result, it is possible, for example, that the skb is removed from dev->rxq by __skb_unlink() before the check "!skb_queue_empty(&dev->rxq)" in usbnet_terminate_urbs() is made. It is also possible in this case that the skb is added to dev->done queue after "!skb_queue_empty(&dev->done)" is checked. So usbnet_terminate_urbs() may stop waiting and return while dev->done queue still has an item.

Quoted 3 lines

Exactly what problem will that result in? The tasklet_kill() will wait for the processing of the single element done queue, and everything will be fine. Or?

Given enough time, what prevents defer_bh() from calling tasklet_schedule(&dev->bh) *after* usbnet_stop() calls tasklet_kill()?

Consider the following situation (assuming '&&' are changed to '||' in that while loop in usbnet_terminate_urbs() as they should be):

CPU0 CPU1
usbnet_stop() defer_bh() with list == dev->rxq
usbnet_terminate_urbs()
__skb_unlink() removes the last skb from dev->rxq. dev->rxq, dev->txq and dev->done are now empty.
while (!skb_queue_empty()...)
The loop ends because all 3
queues are now empty.

usbnet_terminate_urbs() ends.

usbnet_stop() continues:
usbnet_status_stop(dev);
...
del_timer_sync (&dev->delay);

  tasklet_kill (&dev->bh);
                                __skb_queue_tail(&dev->done, skb);
                                if (dev->done.qlen == 1)
                                  tasklet_schedule(&dev->bh);

The BH is scheduled at this point, which is not what was intended. The race window is small, but still.

Regards,
Eugene

#40 lore raw

Re: [PATCH 2/2] usbnet: Fix a race between usbnet_stop() and the BH

Quoted 83 lines · Eugene Shatokhin

25.08.2015 00:01, Bjørn Mork пишет:

Quoted 55 lines
Quoted 45 lines · Eugene Shatokhin

The race may happen when a device (e.g. YOTA 4G LTE Modem) is unplugged while the system is downloading a large file from the Net.

Hardware breakpoints and Kprobes with delays were used to confirm that the race does actually happen.

The race is on skb_queue ('next' pointer) between usbnet_stop() and rx_complete(), which, in turn, calls usbnet_bh().

Here is a part of the call stack with the code where the changes to the queue happen. The line numbers are for the kernel 4.1.0:

*0 __skb_unlink (skbuff.h:1517)

     prev->next = next;
*1 defer_bh (usbnet.c:430)
     spin_lock_irqsave(&list->lock, flags);
     old_state = entry->state;
     entry->state = state;
     __skb_unlink(skb, list);
     spin_unlock(&list->lock);
     spin_lock(&dev->done.lock);
     __skb_queue_tail(&dev->done, skb);
     if (dev->done.qlen == 1)
         tasklet_schedule(&dev->bh);
     spin_unlock_irqrestore(&dev->done.lock, flags);
*2 rx_complete (usbnet.c:640)
     state = defer_bh(dev, skb, &dev->rxq, state);

At the same time, the following code repeatedly checks if the queue is empty and reads these values concurrently with the above changes:

*0 usbnet_terminate_urbs (usbnet.c:765)

     /* maybe wait for deletions to finish. */
     while (!skb_queue_empty(&dev->rxq)
         && !skb_queue_empty(&dev->txq)
         && !skb_queue_empty(&dev->done)) {
             schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
             set_current_state(TASK_UNINTERRUPTIBLE);
             netif_dbg(dev, ifdown, dev->net,
                   "waited for %d urb completions\n", temp);
     }
*1  usbnet_stop (usbnet.c:806)
     if (!(info->flags & FLAG_AVOID_UNLINK_URBS))
         usbnet_terminate_urbs(dev);
Quoted 7 lines

As a result, it is possible, for example, that the skb is removed from dev->rxq by __skb_unlink() before the check "!skb_queue_empty(&dev->rxq)" in usbnet_terminate_urbs() is made. It is also possible in this case that the skb is added to dev->done queue after "!skb_queue_empty(&dev->done)" is checked. So usbnet_terminate_urbs() may stop waiting and return while dev->done queue still has an item.

Quoted 3 lines

Exactly what problem will that result in? The tasklet_kill() will wait for the processing of the single element done queue, and everything will be fine. Or?

Given enough time, what prevents defer_bh() from calling tasklet_schedule(&dev->bh) *after* usbnet_stop() calls tasklet_kill()?

Consider the following situation (assuming '&&' are changed to '||' in that while loop in usbnet_terminate_urbs() as they should be):

CPU0 CPU1
usbnet_stop() defer_bh() with list == dev->rxq
usbnet_terminate_urbs()
__skb_unlink() removes the last skb from dev->rxq. dev->rxq, dev->txq and dev->done are now empty.
while (!skb_queue_empty()...)
The loop ends because all 3
queues are now empty.

usbnet_terminate_urbs() ends.

usbnet_stop() continues:
usbnet_status_stop(dev);
...
del_timer_sync (&dev->delay);

Quoted 7 lines
  tasklet_kill (&dev->bh);
                                __skb_queue_tail(&dev->done, skb);
                                if (dev->done.qlen == 1)
                                  tasklet_schedule(&dev->bh);

The BH is scheduled at this point, which is not what was intended. The race window is small, but still.

I guess you are right. At least I cannot prove that you are not :)

There is a bit too much complexity involved here for me...

Bjørn

#41 lore raw

Re: [PATCH 2/2] usbnet: Fix a race between usbnet_stop() and the BH

28.08.2015 11:55, Bjørn Mork пишет:

Quoted 90 lines
Quoted 88 lines · Eugene Shatokhin

25.08.2015 00:01, Bjørn Mork пишет:

Quoted 55 lines
Quoted 45 lines · Eugene Shatokhin

The race may happen when a device (e.g. YOTA 4G LTE Modem) is unplugged while the system is downloading a large file from the Net.

Hardware breakpoints and Kprobes with delays were used to confirm that the race does actually happen.

The race is on skb_queue ('next' pointer) between usbnet_stop() and rx_complete(), which, in turn, calls usbnet_bh().

Here is a part of the call stack with the code where the changes to the queue happen. The line numbers are for the kernel 4.1.0:

*0 __skb_unlink (skbuff.h:1517)

      prev->next = next;
*1 defer_bh (usbnet.c:430)
      spin_lock_irqsave(&list->lock, flags);
      old_state = entry->state;
      entry->state = state;
      __skb_unlink(skb, list);
      spin_unlock(&list->lock);
      spin_lock(&dev->done.lock);
      __skb_queue_tail(&dev->done, skb);
      if (dev->done.qlen == 1)
          tasklet_schedule(&dev->bh);
      spin_unlock_irqrestore(&dev->done.lock, flags);
*2 rx_complete (usbnet.c:640)
      state = defer_bh(dev, skb, &dev->rxq, state);

At the same time, the following code repeatedly checks if the queue is empty and reads these values concurrently with the above changes:

*0 usbnet_terminate_urbs (usbnet.c:765)

      /* maybe wait for deletions to finish. */
      while (!skb_queue_empty(&dev->rxq)
          && !skb_queue_empty(&dev->txq)
          && !skb_queue_empty(&dev->done)) {
              schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
              set_current_state(TASK_UNINTERRUPTIBLE);
              netif_dbg(dev, ifdown, dev->net,
                    "waited for %d urb completions\n", temp);
      }
*1  usbnet_stop (usbnet.c:806)
      if (!(info->flags & FLAG_AVOID_UNLINK_URBS))
          usbnet_terminate_urbs(dev);
Quoted 7 lines

As a result, it is possible, for example, that the skb is removed from dev->rxq by __skb_unlink() before the check "!skb_queue_empty(&dev->rxq)" in usbnet_terminate_urbs() is made. It is also possible in this case that the skb is added to dev->done queue after "!skb_queue_empty(&dev->done)" is checked. So usbnet_terminate_urbs() may stop waiting and return while dev->done queue still has an item.

Quoted 3 lines

Exactly what problem will that result in? The tasklet_kill() will wait for the processing of the single element done queue, and everything will be fine. Or?

Given enough time, what prevents defer_bh() from calling tasklet_schedule(&dev->bh) *after* usbnet_stop() calls tasklet_kill()?

Consider the following situation (assuming '&&' are changed to '||' in that while loop in usbnet_terminate_urbs() as they should be):

CPU0 CPU1
usbnet_stop() defer_bh() with list == dev->rxq
usbnet_terminate_urbs()
__skb_unlink() removes the last skb from dev->rxq. dev->rxq, dev->txq and dev->done are now empty.
while (!skb_queue_empty()...)
The loop ends because all 3
queues are now empty.

usbnet_terminate_urbs() ends.

usbnet_stop() continues:

   usbnet_status_stop(dev);
   ...
   del_timer_sync (&dev->delay);
   tasklet_kill (&dev->bh);
                                 __skb_queue_tail(&dev->done, skb);
                                 if (dev->done.qlen == 1)
                                   tasklet_schedule(&dev->bh);
Quoted 6 lines
Quoted 2 lines

The BH is scheduled at this point, which is not what was intended. The race window is small, but still.

I guess you are right. At least I cannot prove that you are not :)

There is a bit too much complexity involved here for me...

:-)

Yes, it is quite complex.

I admit, it was easier for me to find the races in usbnet (the tools like KernelStrider and RaceHound do the dirty work) than to analyze their consequences. The latter often requires some time and effort, and so it did this time.

Well, any objections to this patch?

Regards,

Eugene

#42 lore raw

Re: [PATCH 2/2] usbnet: Fix a race between usbnet_stop() and the BH

Quoted 8 lines · Eugene Shatokhin

28.08.2015 11:55, Bjørn Mork пишет:

Quoted 3 lines

I guess you are right. At least I cannot prove that you are not :)

There is a bit too much complexity involved here for me...

:-)

Quoted 8 lines

Yes, it is quite complex.

I admit, it was easier for me to find the races in usbnet (the tools like KernelStrider and RaceHound do the dirty work) than to analyze their consequences. The latter often requires some time and effort, and so it did this time.

Well, any objections to this patch?

No objections from me.

But I would have liked it much better if the code became simpler instead of more complex.

Bjørn

#43 lore raw

Re: [PATCH 2/2] usbnet: Fix a race between usbnet_stop() and the BH

31.08.2015 10:32, Bjørn Mork пишет:

Quoted 16 lines
Quoted 15 lines · Eugene Shatokhin

28.08.2015 11:55, Bjørn Mork пишет:

Quoted 3 lines

I guess you are right. At least I cannot prove that you are not :)

There is a bit too much complexity involved here for me...

:-)

Yes, it is quite complex.

I admit, it was easier for me to find the races in usbnet (the tools like KernelStrider and RaceHound do the dirty work) than to analyze their consequences. The latter often requires some time and effort, and so it did this time.

Quoted 6 lines
Quoted 1 line

Well, any objections to this patch?

No objections from me.

But I would have liked it much better if the code became simpler instead of more complex.

Me too, but I can see no other way here. The code is simpler without locking, indeed, but locking is needed to prevent the problems described earlier.

One needs to make sure that checking if txq or rxq is empty in usbnet_terminate_urbs() cannot get inbetween of processing of these queues and dev->done in defer_bh(). So 'list' and 'dev->done' must be updated under a common lock in defer_bh(). list->lock is an obvious candidate for this.

For the same reason, skb_queue_empty(q) must be called under q->lock. So the code takes it, calls skb_queue_empty(q) and then releases it to wait a little. Rinse and repeat.

The last complex piece is that spin_lock_nested() in defer_bh. It is safe to take both list->lock and dev->done.lock there (defer_bh can only be called for list = dev->rxq or dev->txq but not for dev->done). For lockdep, however, this is suspicious because '*list' and 'dev->done' are of the same type so the lock class is the same. So it complained.

To tell lockdep it is OK to use such locking scheme in this particular case, the recommended pattern was used: spin_lock_nested with SINGLE_DEPTH_NESTING.

Regards,
Eugene

maintainer
#44 lore raw

Re: [PATCH 2/2] usbnet: Fix a race between usbnet_stop() and the BH

Quoted 8 lines · Eugene Shatokhin
Quoted 1 line

But I would have liked it much better if the code became simpler

instead

Quoted 1 line

of more complex.

Me too, but I can see no other way here. The code is simpler without locking, indeed, but locking is needed to prevent the problems described
earlier.

On a practical note, will you resend the patch?

Regards
Oliver

#45 lore raw

Re: [PATCH 2/2] usbnet: Fix a race between usbnet_stop() and the BH

01.09.2015 10:58, Oliver Neukum пишет:

Quoted 11 lines · Eugene Shatokhin
Quoted 8 lines
Quoted 1 line

But I would have liked it much better if the code became simpler

instead

Quoted 1 line

of more complex.

Me too, but I can see no other way here. The code is simpler without locking, indeed, but locking is needed to prevent the problems described
earlier.

On a practical note, will you resend the patch?

Yes, I will do it shortly. Thanks for your help!

Regards,

Eugene

#46 lore raw

PATCH 1/1 usbnet: Fix a race between usbnet_stop() and the BH

In reply to Oliver Neukum

The race may happen when a device (e.g. YOTA 4G LTE Modem) is unplugged while the system is downloading a large file from the Net.

Hardware breakpoints and Kprobes with delays were used to confirm that the race does actually happen.

The race is on skb_queue ('next' pointer) between usbnet_stop() and rx_complete(), which, in turn, calls usbnet_bh().

Here is a part of the call stack with the code where the changes to the queue happen. The line numbers are for the kernel 4.1.0:

*0 __skb_unlink (skbuff.h:1517)

    prev->next = next;
*1 defer_bh (usbnet.c:430)
    spin_lock_irqsave(&list->lock, flags);
    old_state = entry->state;
    entry->state = state;
    __skb_unlink(skb, list);
    spin_unlock(&list->lock);
    spin_lock(&dev->done.lock);
    __skb_queue_tail(&dev->done, skb);
    if (dev->done.qlen == 1)
        tasklet_schedule(&dev->bh);
    spin_unlock_irqrestore(&dev->done.lock, flags);
*2 rx_complete (usbnet.c:640)
    state = defer_bh(dev, skb, &dev->rxq, state);

At the same time, the following code repeatedly checks if the queue is empty and reads these values concurrently with the above changes:

*0 usbnet_terminate_urbs (usbnet.c:765)

    /* maybe wait for deletions to finish. */
    while (!skb_queue_empty(&dev->rxq)
        && !skb_queue_empty(&dev->txq)
        && !skb_queue_empty(&dev->done)) {
            schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
            set_current_state(TASK_UNINTERRUPTIBLE);
            netif_dbg(dev, ifdown, dev->net,
                  "waited for %d urb completions\n", temp);
    }
*1  usbnet_stop (usbnet.c:806)
    if (!(info->flags & FLAG_AVOID_UNLINK_URBS))
        usbnet_terminate_urbs(dev);

As a result, it is possible, for example, that the skb is removed from dev->rxq by __skb_unlink() before the check "!skb_queue_empty(&dev->rxq)" in usbnet_terminate_urbs() is made. It is also possible in this case that the skb is added to dev->done queue after "!skb_queue_empty(&dev->done)" is checked. So usbnet_terminate_urbs() may stop waiting and return while dev->done queue still has an item.

Locking in defer_bh() and usbnet_terminate_urbs() was revisited to avoid this race.

---
 drivers/net/usb/usbnet.c | 39 ++++++++++++++++++++++++++++-----------
 1 file changed, 28 insertions(+), 11 deletions(-)
drivers/net/usb/usbnet.c +28 -11
index e049857..b4cf107 100644
@@ -428,12 +428,18 @@ static enum skb_state defer_bh(struct usbnet *dev, struct sk_buff *skb,
 	old_state = entry->state;
 	entry->state = state;
 	__skb_unlink(skb, list);
-	spin_unlock(&list->lock);
-	spin_lock(&dev->done.lock);
+
+	/* defer_bh() is never called with list == &dev->done.
+	 * spin_lock_nested() tells lockdep that it is OK to take
+	 * dev->done.lock here with list->lock held.
+	 */
+	spin_lock_nested(&dev->done.lock, SINGLE_DEPTH_NESTING);
+
 	__skb_queue_tail(&dev->done, skb);
 	if (dev->done.qlen == 1)
 		tasklet_schedule(&dev->bh);
-	spin_unlock_irqrestore(&dev->done.lock, flags);
+	spin_unlock(&dev->done.lock);
+	spin_unlock_irqrestore(&list->lock, flags);
 	return old_state;
 }
 
@@ -749,6 +755,20 @@ EXPORT_SYMBOL_GPL(usbnet_unlink_rx_urbs);
 
 /*-------------------------------------------------------------------------*/
 
+static void wait_skb_queue_empty(struct sk_buff_head *q)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&q->lock, flags);
+	while (!skb_queue_empty(q)) {
+		spin_unlock_irqrestore(&q->lock, flags);
+		schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		spin_lock_irqsave(&q->lock, flags);
+	}
+	spin_unlock_irqrestore(&q->lock, flags);
+}
+
 // precondition: never called in_interrupt
 static void usbnet_terminate_urbs(struct usbnet *dev)
 {
@@ -762,14 +782,11 @@ static void usbnet_terminate_urbs(struct usbnet *dev)
 		unlink_urbs(dev, &dev->rxq);
 
 	/* maybe wait for deletions to finish. */
-	while (!skb_queue_empty(&dev->rxq)
-		&& !skb_queue_empty(&dev->txq)
-		&& !skb_queue_empty(&dev->done)) {
-			schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
-			set_current_state(TASK_UNINTERRUPTIBLE);
-			netif_dbg(dev, ifdown, dev->net,
-				  "waited for %d urb completions\n", temp);
-	}
+	wait_skb_queue_empty(&dev->rxq);
+	wait_skb_queue_empty(&dev->txq);
+	wait_skb_queue_empty(&dev->done);
+	netif_dbg(dev, ifdown, dev->net,
+		  "waited for %d urb completions\n", temp);
 	set_current_state(TASK_RUNNING);
 	remove_wait_queue(&dev->wait, &wait);
 }
Signature · 1 line
2.3.2
#47 lore raw

Re: [PATCH] usbnet: Fix a race between usbnet_stop() and the BH

01.09.2015 17:05, Eugene Shatokhin пишет:

Quoted 119 lines

The race may happen when a device (e.g. YOTA 4G LTE Modem) is unplugged while the system is downloading a large file from the Net.

Hardware breakpoints and Kprobes with delays were used to confirm that the race does actually happen.

The race is on skb_queue ('next' pointer) between usbnet_stop() and rx_complete(), which, in turn, calls usbnet_bh().

Here is a part of the call stack with the code where the changes to the queue happen. The line numbers are for the kernel 4.1.0:

*0 __skb_unlink (skbuff.h:1517)

     prev->next = next;
*1 defer_bh (usbnet.c:430)
     spin_lock_irqsave(&list->lock, flags);
     old_state = entry->state;
     entry->state = state;
     __skb_unlink(skb, list);
     spin_unlock(&list->lock);
     spin_lock(&dev->done.lock);
     __skb_queue_tail(&dev->done, skb);
     if (dev->done.qlen == 1)
         tasklet_schedule(&dev->bh);
     spin_unlock_irqrestore(&dev->done.lock, flags);
*2 rx_complete (usbnet.c:640)
     state = defer_bh(dev, skb, &dev->rxq, state);

At the same time, the following code repeatedly checks if the queue is empty and reads these values concurrently with the above changes:

*0 usbnet_terminate_urbs (usbnet.c:765)

     /* maybe wait for deletions to finish. */
     while (!skb_queue_empty(&dev->rxq)
         && !skb_queue_empty(&dev->txq)
         && !skb_queue_empty(&dev->done)) {
             schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
             set_current_state(TASK_UNINTERRUPTIBLE);
             netif_dbg(dev, ifdown, dev->net,
                   "waited for %d urb completions\n", temp);
     }
*1  usbnet_stop (usbnet.c:806)
     if (!(info->flags & FLAG_AVOID_UNLINK_URBS))
         usbnet_terminate_urbs(dev);

As a result, it is possible, for example, that the skb is removed from dev->rxq by __skb_unlink() before the check "!skb_queue_empty(&dev->rxq)" in usbnet_terminate_urbs() is made. It is also possible in this case that the skb is added to dev->done queue after "!skb_queue_empty(&dev->done)" is checked. So usbnet_terminate_urbs() may stop waiting and return while dev->done queue still has an item.

Locking in defer_bh() and usbnet_terminate_urbs() was revisited to avoid this race.

---
  drivers/net/usb/usbnet.c | 39 ++++++++++++++++++++++++++++-----------
  1 file changed, 28 insertions(+), 11 deletions(-)
drivers/net/usb/usbnet.c +23 -9
index e049857..b4cf107 100644
@@ -428,12 +428,18 @@ static enum skb_state defer_bh(struct usbnet *dev, struct sk_buff *skb,
  	old_state = entry->state;
  	entry->state = state;
  	__skb_unlink(skb, list);
-	spin_unlock(&list->lock);
-	spin_lock(&dev->done.lock);
+
+	/* defer_bh() is never called with list == &dev->done.
+	 * spin_lock_nested() tells lockdep that it is OK to take
+	 * dev->done.lock here with list->lock held.
+	 */
+	spin_lock_nested(&dev->done.lock, SINGLE_DEPTH_NESTING);
+
  	__skb_queue_tail(&dev->done, skb);
  	if (dev->done.qlen == 1)
  		tasklet_schedule(&dev->bh);
-	spin_unlock_irqrestore(&dev->done.lock, flags);
+	spin_unlock(&dev->done.lock);
+	spin_unlock_irqrestore(&list->lock, flags);
  	return old_state;
  }
 
@@ -749,6 +755,20 @@ EXPORT_SYMBOL_GPL(usbnet_unlink_rx_urbs);
 
  /*-------------------------------------------------------------------------*/
 
+static void wait_skb_queue_empty(struct sk_buff_head *q)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&q->lock, flags);
+	while (!skb_queue_empty(q)) {
+		spin_unlock_irqrestore(&q->lock, flags);
+		schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		spin_lock_irqsave(&q->lock, flags);
+	}
+	spin_unlock_irqrestore(&q->lock, flags);
+}
+
  // precondition: never called in_interrupt
  static void usbnet_terminate_urbs(struct usbnet *dev)
  {
@@ -762,14 +782,11 @@ static void usbnet_terminate_urbs(struct usbnet *dev)
  		unlink_urbs(dev, &dev->rxq);
 
  	/* maybe wait for deletions to finish. */
-	while (!skb_queue_empty(&dev->rxq)
-		&& !skb_queue_empty(&dev->txq)
-		&& !skb_queue_empty(&dev->done)) {
-			schedule_timeout(msecs_to_jiffies(UNLINK_TIMEOUT_MS));
-			set_current_state(TASK_UNINTERRUPTIBLE);
-			netif_dbg(dev, ifdown, dev->net,

drivers/net/usb/usbnet.c static void usbnet_terminate_urbs(struct usbnet *dev) · line 792

Quoted 10 lines
drivers/net/usb/usbnet.c +5 -2
-				  "waited for %d urb completions\n", temp);
-	}
+	wait_skb_queue_empty(&dev->rxq);
+	wait_skb_queue_empty(&dev->txq);
+	wait_skb_queue_empty(&dev->done);
+	netif_dbg(dev, ifdown, dev->net,
+		  "waited for %d urb completions\n", temp);
  	set_current_state(TASK_RUNNING);
  	remove_wait_queue(&dev->wait, &wait);
  }

I resent the patch to make it separate. What is the status of this now?

Regards,
Eugene

Reviewed-by
#48 lore raw

Re: [PATCH] usbnet: Fix a race between usbnet_stop() and the BH

Quoted 1 line · Eugene Shatokhin

I resent the patch to make it separate. What is the status of this now?

One of the many nice features of patchwork is that you don't need to ask those questions :)

See http://patchwork.ozlabs.org/patch/512856/

I really don't think it's appropriate for me to ack this, but I can add my

if that helps.

Bjørn

maintainer Acked-by
#49 lore raw

Re: [PATCH] usbnet: Fix a race between usbnet_stop() and the BH

Quoted 11 lines · Bjørn Mork
Quoted 1 line · Eugene Shatokhin

I resent the patch to make it separate. What is the status of this now?

One of the many nice features of patchwork is that you don't need to ask those questions :)

See http://patchwork.ozlabs.org/patch/512856/

I really don't think it's appropriate for me to ack this, but I can add my

Well, in case this helps:

Quoted 1 line

Regards
Oliver

maintainer Applied
#50 lore raw

Re: [PATCH] usbnet: Fix a race between usbnet_stop() and the BH

In reply to Eugene Shatokhin

From: Eugene Shatokhin <[email protected]>
Date: Tue, 1 Sep 2015 17:05:33 +0300

Quoted 11 lines

The race may happen when a device (e.g. YOTA 4G LTE Modem) is unplugged while the system is downloading a large file from the Net.

Hardware breakpoints and Kprobes with delays were used to confirm that the race does actually happen.

The race is on skb_queue ('next' pointer) between usbnet_stop() and rx_complete(), which, in turn, calls usbnet_bh().

Here is a part of the call stack with the code where the changes to the queue happen. The line numbers are for the kernel 4.1.0:

...

Quoted 8 lines

As a result, it is possible, for example, that the skb is removed from dev->rxq by __skb_unlink() before the check "!skb_queue_empty(&dev->rxq)" in usbnet_terminate_urbs() is made. It is also possible in this case that the skb is added to dev->done queue after "!skb_queue_empty(&dev->done)" is checked. So usbnet_terminate_urbs() may stop waiting and return while dev->done queue still has an item.

Quoted 4 lines

Locking in defer_bh() and usbnet_terminate_urbs() was revisited to avoid this race.

Applied, thanks.

#51 lore raw

Re: [PATCH 2/2] usbnet: Fix a race between usbnet_stop() and the BH

In reply to Eugene Shatokhin
Quoted 35 lines · Eugene Shatokhin
Quoted 1 line

Exactly what problem will that result in? The tasklet_kill() will

wait

Quoted 1 line

for the processing of the single element done queue, and everything

will

Quoted 1 line

be fine. Or?

Given enough time, what prevents defer_bh() from calling tasklet_schedule(&dev->bh) *after* usbnet_stop() calls tasklet_kill()?

Consider the following situation (assuming '&&' are changed to '||' in
that while loop in usbnet_terminate_urbs() as they should be):

CPU0 CPU1
usbnet_stop() defer_bh() with list == dev->rxq
usbnet_terminate_urbs()
__skb_unlink() removes the last skb from dev->rxq. dev->rxq, dev->txq and dev->done are now empty.
while (!skb_queue_empty()...)
The loop ends because all 3
queues are now empty.

usbnet_terminate_urbs() ends.

usbnet_stop() continues:

   usbnet_status_stop(dev);
   ...
   del_timer_sync (&dev->delay);
   tasklet_kill (&dev->bh);
                                 __skb_queue_tail(&dev->done, skb);
                                 if (dev->done.qlen == 1)
                                   tasklet_schedule(&dev->bh);
Quoted 3 lines

The BH is scheduled at this point, which is not what was intended. The
race window is small, but still.

This looks possible. I cannot come up with a better fix, although it isn't nice. Thanks for finding this.

Regards
Oliver

maintainer
#52 lore raw

Re: [PATCH 2/2] usbnet: Fix a race between usbnet_stop() and the BH

In reply to Eugene Shatokhin

From: Eugene Shatokhin <[email protected]>
Date: Mon, 24 Aug 2015 23:13:43 +0300

Quoted 11 lines

The race may happen when a device (e.g. YOTA 4G LTE Modem) is unplugged while the system is downloading a large file from the Net.

Hardware breakpoints and Kprobes with delays were used to confirm that the race does actually happen.

The race is on skb_queue ('next' pointer) between usbnet_stop() and rx_complete(), which, in turn, calls usbnet_bh().

Here is a part of the call stack with the code where the changes to the queue happen. The line numbers are for the kernel 4.1.0:

...

It looks like this patch needs more discussion/work.