mac80211: do not call driver wake_tx_queue op during reconfig
Subsystems (2)
PATCH 1/1 mac80211: do not call driver wake_tx_queue op during reconfig
There are several scenarios in which mac80211 can call drv_wake_tx_queue after ieee80211_restart_hw has been called and has not yet completed. Driver private structs are considered uninitialized until mac80211 has uploaded the vifs, stations and keys again, so using private tx queue data during that time is not safe.
The driver can also not rely on drv_reconfig_complete to figure out when it is safe to accept drv_wake_tx_queue calls again, because it is only called after all tx queues are woken again.
To fix this, bail out early in drv_wake_tx_queue if local->in_reconfig is set.
--- net/mac80211/driver-ops.h | 3 +++ 1 file changed, 3 insertions(+)
net/mac80211/driver-ops.h +3 -0
@@ -1195,6 +1195,9 @@ static inline void drv_wake_tx_queue(struct ieee80211_local *local, {struct ieee80211_sub_if_data *sdata =vif_to_sdata (txq->txq.vif); +if (local->in_reconfig) +return ; +if (!check_sdata_in_driver (sdata))return ;
Signature · 1 line
2.17.0
Quoted 6 lines · Felix Fietkau
There are several scenarios in which mac80211 can call
drv_wake_tx_queueafterieee80211_restart_hwhas been called and has not yet completed. Driver private structs are considered uninitialized until mac80211 has uploaded the vifs, stations and keys again, so using private tx queue data during that time is not safe.
Quoted 6 lines
The driver can also not rely on
drv_reconfig_completeto figure out when it is safe to acceptdrv_wake_tx_queuecalls again, because it is only called after all tx queues are woken again.To fix this, bail out early in
drv_wake_tx_queueiflocal->in_reconfigis set.
This reminded me of a patch I posted back in 2016. The discussion just sort of ended on it, but curious if you have a new opinion on it after debugging the issue in this patch:
https://patchwork.kernel.org/patch/9457709/
For what its worth, I've been using the patch above since I posted it, and it seems to work well for ath9k and ath10k.
Thanks,
Ben
Quoted 22 lines
Signed-off-by Felix Fietkau <[email protected]>--- net/mac80211/driver-ops.h | 3 +++ 1 file changed, 3 insertions(+)net/mac80211/driver-ops.h +3
-0@@ -1195,6 +1195,9 @@ static inline void drv_wake_tx_queue(struct ieee80211_local *local, {struct ieee80211_sub_if_data *sdata =vif_to_sdata (txq->txq.vif); +if (local->in_reconfig) +return ; +if (!check_sdata_in_driver (sdata))return ;
Signature · 2 lines
Ben Greear <[email protected]> Candela Technologies Inc http://www.candelatech.com
Quoted 14 lines · Ben Greear
Quoted 6 lines · Felix Fietkau
There are several scenarios in which mac80211 can call
drv_wake_tx_queueafterieee80211_restart_hwhas been called and has not yet completed. Driver private structs are considered uninitialized until mac80211 has uploaded the vifs, stations and keys again, so using private tx queue data during that time is not safe.Quoted 6 lines
The driver can also not rely on
drv_reconfig_completeto figure out when it is safe to acceptdrv_wake_tx_queuecalls again, because it is only called after all tx queues are woken again.To fix this, bail out early in
drv_wake_tx_queueiflocal->in_reconfigis set.
Quoted 8 lines
This reminded me of a patch I posted back in 2016. The discussion just sort of ended on it, but curious if you have a new opinion on it after debugging the issue in this patch:
https://patchwork.kernel.org/patch/9457709/
For what its worth, I've been using the patch above since I posted it, and it seems to work well for ath9k and ath10k.
I agree with what Johannes wrote about that patch. Fixing this could likely be as simple as clearing IEEE80211_SDATA_IN_DRIVER on all interfaces before bringing any of them back up. That way the normal interface add logic applies without nasty special cases.
The reconfig code checks for ieee80211_sdata_running (which is unaffected), so I don't think we need to save the previous value of that flag.
- Felix
Quoted 23 lines · Felix Fietkau
Quoted 14 lines · Ben Greear
Quoted 6 lines · Felix Fietkau
There are several scenarios in which mac80211 can call
drv_wake_tx_queueafterieee80211_restart_hwhas been called and has not yet completed. Driver private structs are considered uninitialized until mac80211 has uploaded the vifs, stations and keys again, so using private tx queue data during that time is not safe.Quoted 6 lines
The driver can also not rely on
drv_reconfig_completeto figure out when it is safe to acceptdrv_wake_tx_queuecalls again, because it is only called after all tx queues are woken again.To fix this, bail out early in
drv_wake_tx_queueiflocal->in_reconfigis set.Quoted 8 lines
This reminded me of a patch I posted back in 2016. The discussion just sort of ended on it, but curious if you have a new opinion on it after debugging the issue in this patch:
https://patchwork.kernel.org/patch/9457709/
For what its worth, I've been using the patch above since I posted it, and it seems to work well for ath9k and ath10k.
Quoted 7 lines
I agree with what Johannes wrote about that patch. Fixing this could likely be as simple as clearing
IEEE80211_SDATA_IN_DRIVERon all interfaces before bringing any of them back up. That way the normal interface add logic applies without nasty special cases.
The reconfig code checks forieee80211_sdata_running(which is unaffected), so I don't think we need to save the previous value of that flag.
How about this? (untested)
---
net/mac80211/util.c +3 -4
@@ -1956,7 +1956,6 @@ static void ieee80211_flush_completed_scan(struct ieee80211_local *local,static void ieee80211_handle_reconfig_failure (struct ieee80211_local *local) { -struct ieee80211_sub_if_data *sdata;struct ieee80211_chanctx *ctx;/*
@@ -1980,9 +1979,6 @@ static void ieee80211_handle_reconfig_failure(struct ieee80211_local *local)*/ ieee80211_sched_scan_end (local); -list_for_each_entry (sdata, &local->interfaces, list) - sdata->flags &= ~IEEE80211_SDATA_IN_DRIVER ; -/* Mark channel contexts as not being in the driver any more to avoid * removing them from the driver during the shutdown process... */
@@ -2135,6 +2131,9 @@ int ieee80211_reconfig(struct ieee80211_local *local) local->started = false; +list_for_each_entry (sdata, &local->interfaces, list) + sdata->flags &= ~IEEE80211_SDATA_IN_DRIVER ; +/* * Upon resume hardware can sometimes be goofy due to * various platform / driver / bus issues, so restarting
Quoted 55 lines · Felix Fietkau
Quoted 23 lines · Felix Fietkau
Quoted 14 lines · Ben Greear
Quoted 6 lines · Felix Fietkau
There are several scenarios in which mac80211 can call
drv_wake_tx_queueafterieee80211_restart_hwhas been called and has not yet completed. Driver private structs are considered uninitialized until mac80211 has uploaded the vifs, stations and keys again, so using private tx queue data during that time is not safe.Quoted 6 lines
The driver can also not rely on
drv_reconfig_completeto figure out when it is safe to acceptdrv_wake_tx_queuecalls again, because it is only called after all tx queues are woken again.To fix this, bail out early in
drv_wake_tx_queueiflocal->in_reconfigis set.Quoted 8 lines
This reminded me of a patch I posted back in 2016. The discussion just sort of ended on it, but curious if you have a new opinion on it after debugging the issue in this patch:
https://patchwork.kernel.org/patch/9457709/
For what its worth, I've been using the patch above since I posted it, and it seems to work well for ath9k and ath10k.
Quoted 7 lines
I agree with what Johannes wrote about that patch. Fixing this could likely be as simple as clearing
IEEE80211_SDATA_IN_DRIVERon all interfaces before bringing any of them back up. That way the normal interface add logic applies without nasty special cases.
The reconfig code checks forieee80211_sdata_running(which is unaffected), so I don't think we need to save the previous value of that flag.How about this? (untested)
---net/mac80211/util.c +0
-4@@ -1956,7 +1956,6 @@ static void ieee80211_flush_completed_scan(struct ieee80211_local *local,static void ieee80211_handle_reconfig_failure (struct ieee80211_local *local) { -struct ieee80211_sub_if_data *sdata;struct ieee80211_chanctx *ctx;/* @@ -1980,9 +1979,6 @@ static void ieee80211_handle_reconfig_failure(struct ieee80211_local *local)*/ ieee80211_sched_scan_end (local); -list_for_each_entry (sdata, &local->interfaces, list) - sdata->flags &= ~IEEE80211_SDATA_IN_DRIVER ; -/* Mark channel contexts as not being in the driver any more to avoid * removing them from the driver during the shutdown process... */
net/mac80211/util.c int ieee80211_reconfig(struct ieee80211_local *local) · line 2139
Quoted 9 lines
net/mac80211/util.c +3
-0@@ -2135,6 +2131,9 @@ int ieee80211_reconfig(struct ieee80211_local *local) local->started = false; +list_for_each_entry (sdata, &local->interfaces, list) + sdata->flags &= ~IEEE80211_SDATA_IN_DRIVER ; +/* * Upon resume hardware can sometimes be goofy due to * various platform / driver / bus issues, so restarting
Let me see if anyone (Johannes?) has comments on this... I can test it out a bit later if no one suggests changes.
Thanks,
Ben
Signature · 2 lines
Ben Greear <[email protected]> Candela Technologies Inc http://www.candelatech.com