Skip to contentSkip to conversation

at76_usb: fix Oops on disconnect

PATCH In review Jason Andryuk 2 participants 6 messages Feb 5, 2009 – Feb 19, 2009

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 at76_usb: fix Oops on disconnect

flush_workqueue needs to be called instead of the generic one and the associated functions need to be modified to prevent re-adding themselves to the workqueue.

--

(quoted diff) +0 -0

b/drivers/staging/at76_usb/at76_usb.c
index 185533e..211a6dd 100644

drivers/staging/at76_usb/at76_usb.c +3 -0
@@ -1967,6 +1967,9 @@ static void at76_dwork_hw_scan(struct work_struct *work)
 					      dwork_hw_scan.work);
 	int ret;
 
+	if (priv->device_unplugged)
+		return;
+
 	ret = at76_get_cmd_status(priv->udev, CMD_SCAN);
 	at76_dbg(DBG_MAC80211, "%s: CMD_SCAN status 0x%02x", __func__, ret);
 
@@ -2000,6 +2003,9 @@ static int at76_hw_scan(struct ieee80211_hw *hw,

u8 *ssid, size_t len)
at76_dbg(DBG_MAC80211, "%s():", __func__); at76_dbg_dump(DBG_MAC80211, ssid, len, "ssid %zd bytes:", len);

+ if (priv->device_unplugged)
+ return 0;
+
mutex_lock(&priv->mtx);

ieee80211_stop_queues(hw);
@@ -2102,6 +2108,12 @@ static void at76_configure_filter(struct ieee80211_hw *hw,
flags = changed_flags & AT76_SUPPORTED_FILTERS; *total_flags = AT76_SUPPORTED_FILTERS;

+ /* We bail out of the function to prevent work from being queued
+ * in the case when the device is no longer present. We do so
+ * after updating total_flags to avoid a WARN_ON in mac80211. */
+ if (priv->device_unplugged)
+ return;
+
/* FIXME: access to priv->promisc should be protected with
* priv->mtx, but it's impossible because this function needs to be
* atomic */
@@ -2511,6 +2523,8 @@ static int at76_init_new_device(struct at76_priv *priv,
/* mac80211 initialisation */ priv->hw->wiphy->bands[IEEE80211_BAND_2GHZ] = &at76_supported_band;

+ priv->device_unplugged = 0;
+
if (FIRMWARE_IS_WPA(priv->fw_version) &&
(at76_is_503rfmd(priv->board_type) ||
at76_is_505(priv->board_type)))
@@ -2561,13 +2575,11 @@ static void at76_delete_device(struct at76_priv *priv)
/* The device is gone, don't bother turning it off */ priv->device_unplugged = 1;

- if (priv->mac80211_registered)
- ieee80211_unregister_hw(priv->hw);
-
- /* assuming we used keventd, it must quiesce too */
- flush_scheduled_work();
+ if (priv->mac80211_registered) {
+ flush_workqueue(priv->hw->workqueue);

- kfree(priv->bulk_out_buffer);
+ ieee80211_unregister_hw(priv->hw);
+ }

if (priv->tx_urb) {
usb_kill_urb(priv->tx_urb);
@@ -2580,8 +2592,12 @@ static void at76_delete_device(struct at76_priv *priv)

at76_dbg(DBG_PROC_ENTRY, "%s: unlinked urbs", __func__);

- if (priv->rx_skb)
+ /* move me after urb unlink? */
+ kfree(priv->bulk_out_buffer);
+
+ if (priv->rx_skb) {
kfree_skb(priv->rx_skb);
+ }

usb_put_dev(priv->udev);

#1 lore raw
Quoted 3 lines · Jason Andryuk

flush_workqueue needs to be called instead of the generic one and the associated functions need to be modified to prevent re-adding themselves to the workqueue.

Good fix. Can you please port this to the at76c50x-usb driver I posted this morning?

Kalle

#2 lore raw

I'll take a look at at76c50x-usb in general and bringing over this patch specifically. Unfortunately, I still had one disconnect oops with the patch installed. Netconsole was not configured, so I did not catch it. There were several other successful disconnects without Oops though.

It may need locking as none is currently present, but does a one way transition device_unplugged 0 -> 1 need a lock?

Jason

Quoted 10 lines · Kalle Valo
Quoted 3 lines · Jason Andryuk

flush_workqueue needs to be called instead of the generic one and the associated functions need to be modified to prevent re-adding themselves to the workqueue.

Good fix. Can you please port this to the at76c50x-usb driver I posted this morning?

Kalle

#3 lore raw
Quoted 5 lines · Jason Andryuk

I'll take a look at at76c50x-usb in general and bringing over this patch specifically. Unfortunately, I still had one disconnect oops with the patch installed. Netconsole was not configured, so I did not catch it. There were several other successful disconnects without Oops though.

Progress is always good :)

Quoted 2 lines

It may need locking as none is currently present, but does a one way transition device_unplugged 0 -> 1 need a lock?

Using int might be enough but I think it's not guarenteed anywhere. Better to use atomic operations directly or use a proper lock. But I think the issues are more than that. For example, the disconnect handler is not taking the mutex.

Signature · 1 line
Kalle Valo
#4 lore raw

PATCH 1/1 at76c50x-usb: additional disconnect fixes

Additional attempts to fix Oops on disconnect, that appear to be successful. However, some may be extraneous.

The cancel_delayed_work call is probably the most necessary. The device_unplugged check may not be necessary. del_timer_sync may not be necessary either, but the Oops I was receiving was related to timers. Hence the addition.

--

drivers/net/wireless/at76c50x-usb.c +6 -0
index 7a03251..52b0be5 100644
@@ -1497,6 +1497,9 @@ static void at76_work_set_promisc(struct work_struct *work)
 					      work_set_promisc);
 	int ret = 0;
 
+	if (priv->device_unplugged)
+		return;
+
 	mutex_lock(&priv->mtx);
 
 	priv->mib_buf.type = MIB_LOCAL;
@@ -2297,6 +2300,7 @@ static void at76_delete_device(struct at76_priv *priv)
 	tasklet_kill(&priv->rx_tasklet);
 
 	if (priv->mac80211_registered) {
+		cancel_delayed_work(&priv->dwork_hw_scan);
 		flush_workqueue(priv->hw->workqueue);
 		ieee80211_unregister_hw(priv->hw);
 	}
@@ -2314,6 +2318,8 @@ static void at76_delete_device(struct at76_priv *priv)
 
 	kfree(priv->bulk_out_buffer);
 
+	del_timer_sync(&ledtrig_tx_timer);
+
 	if (priv->rx_skb)
 		kfree_skb(priv->rx_skb);
 
#5 lore raw

Re: [PATCH] at76c50x-usb: additional disconnect fixes

Quoted 7 lines · Jason Andryuk

Additional attempts to fix Oops on disconnect, that appear to be successful. However, some may be extraneous.

The cancel_delayed_work call is probably the most necessary. The device_unplugged check may not be necessary. del_timer_sync may not be necessary either, but the Oops I was receiving was related to timers. Hence the addition.

Thanks, looks good. I'll take this and do similarly as with the previous patches.

Signature · 1 line
Kalle Valo