2013-04-05 20:37:00

by Alan Ott

[permalink] [raw]
Subject: [PATCH] mac802154: Keep track of the channel when changed

Prevent set_channel() from getting called every time a packet is sent. This
looks like it was an oversight.

Signed-off-by: Alan Ott <[email protected]>
---
net/mac802154/tx.c | 3 +++
1 file changed, 3 insertions(+)

diff --git a/net/mac802154/tx.c b/net/mac802154/tx.c
index 3fd3e07..6d16473 100644
--- a/net/mac802154/tx.c
+++ b/net/mac802154/tx.c
@@ -58,6 +58,9 @@ static void mac802154_xmit_worker(struct work_struct *work)
pr_debug("set_channel failed\n");
goto out;
}
+
+ xw->priv->phy->current_channel = xw->chan;
+ xw->priv->phy->current_page = xw->page;
}

res = xw->priv->ops->xmit(&xw->priv->hw, xw->skb);
--
1.7.11.2


2013-04-05 21:06:27

by Werner Almesberger

[permalink] [raw]
Subject: Re: [Linux-zigbee-devel] [PATCH] mac802154: Keep track of the channel when changed

Alan Ott wrote:
> Prevent set_channel() from getting called every time a packet is sent. This
> looks like it was an oversight.

at86rf230.c and derivatives avoid this problem by setting
phy->current_* in the *_channel function.

But I'd agree that it's nicer to do this in one place, not in
every driver.

In case a driver had a weird failure mode in which it leaves the
original channel but only makes it halfway to the new channel, it
could still set phy->current_* and return an error. So there's no
loss of functionality with your change.

- Werner

2013-04-05 21:20:31

by Alan Ott

[permalink] [raw]
Subject: Re: [Linux-zigbee-devel] [PATCH] mac802154: Keep track of the channel when changed

On 04/05/2013 05:05 PM, Werner Almesberger wrote:
> Alan Ott wrote:
>> Prevent set_channel() from getting called every time a packet is sent. This
>> looks like it was an oversight.
> at86rf230.c and derivatives avoid this problem by setting
> phy->current_* in the *_channel function.
>
> But I'd agree that it's nicer to do this in one place, not in
> every driver.
>
> In case a driver had a weird failure mode in which it leaves the
> original channel but only makes it halfway to the new channel, it
> could still set phy->current_* and return an error. So there's no
> loss of functionality with your change.

Hmm... I just noticed that mib.c does the same thing (and doesn't set
phy->current_*). I'll need to fix that one too (and resubmit). :(

Alan.


2013-04-05 23:03:28

by Alan Ott

[permalink] [raw]
Subject: [PATCH v2] mac802154: Keep track of the channel when changed

Two sections checked whether the current channel != the new channel
without ever setting the current channel variables.

1. net/mac802154/tx.c: Prevent set_channel() from getting called every
time a packet is sent.

2. net/mac802154/mib.c: Lock (pib_lock) accesses to current_channel and
current_page and make sure they are updated when the channel has been
changed.

Signed-off-by: Alan Ott <[email protected]>
---
net/mac802154/mib.c | 12 +++++++++++-
net/mac802154/tx.c | 3 +++
2 files changed, 14 insertions(+), 1 deletion(-)

diff --git a/net/mac802154/mib.c b/net/mac802154/mib.c
index f03e55f..8ded97c 100644
--- a/net/mac802154/mib.c
+++ b/net/mac802154/mib.c
@@ -176,9 +176,15 @@ static void phy_chan_notify(struct work_struct *work)
struct mac802154_sub_if_data *priv = netdev_priv(nw->dev);
int res;

+ mutex_lock(&priv->hw->phy->pib_lock);
res = hw->ops->set_channel(&hw->hw, priv->page, priv->chan);
if (res)
pr_debug("set_channel failed\n");
+ else {
+ priv->hw->phy->current_channel = priv->chan;
+ priv->hw->phy->current_page = priv->page;
+ }
+ mutex_unlock(&priv->hw->phy->pib_lock);

kfree(nw);
}
@@ -195,8 +201,11 @@ void mac802154_dev_set_page_channel(struct net_device *dev, u8 page, u8 chan)
priv->chan = chan;
spin_unlock_bh(&priv->mib_lock);

+ mutex_lock(&priv->hw->phy->pib_lock);
if (priv->hw->phy->current_channel != priv->chan ||
priv->hw->phy->current_page != priv->page) {
+ mutex_unlock(&priv->hw->phy->pib_lock);
+
work = kzalloc(sizeof(*work), GFP_ATOMIC);
if (!work)
return;
@@ -204,5 +213,6 @@ void mac802154_dev_set_page_channel(struct net_device *dev, u8 page, u8 chan)
INIT_WORK(&work->work, phy_chan_notify);
work->dev = dev;
queue_work(priv->hw->dev_workqueue, &work->work);
- }
+ } else
+ mutex_unlock(&priv->hw->phy->pib_lock);
}
diff --git a/net/mac802154/tx.c b/net/mac802154/tx.c
index 3fd3e07..6d16473 100644
--- a/net/mac802154/tx.c
+++ b/net/mac802154/tx.c
@@ -58,6 +58,9 @@ static void mac802154_xmit_worker(struct work_struct *work)
pr_debug("set_channel failed\n");
goto out;
}
+
+ xw->priv->phy->current_channel = xw->chan;
+ xw->priv->phy->current_page = xw->page;
}

res = xw->priv->ops->xmit(&xw->priv->hw, xw->skb);
--
1.7.11.2

2013-04-08 16:09:57

by David Miller

[permalink] [raw]
Subject: Re: [PATCH v2] mac802154: Keep track of the channel when changed

From: Alan Ott <[email protected]>
Date: Fri, 5 Apr 2013 19:03:10 -0400

> Two sections checked whether the current channel != the new channel
> without ever setting the current channel variables.
>
> 1. net/mac802154/tx.c: Prevent set_channel() from getting called every
> time a packet is sent.
>
> 2. net/mac802154/mib.c: Lock (pib_lock) accesses to current_channel and
> current_page and make sure they are updated when the channel has been
> changed.
>
> Signed-off-by: Alan Ott <[email protected]>

Applied.