Re: [PATCH] usb: gadget: rndis: don't use dev_get_stats()

[Eric Dumazet <eric.dumazet@gmail.com>]

Le lundi 20 septembre 2010 à 11:27 -0700, David Miller a écrit :
> From: Michal Nazarewicz <m.nazarewicz@samsung.com>
> Date: Mon, 20 Sep 2010 20:23:04 +0200
> 
> > This commit removes the call to dev_get_stats() from the
> > gen_ndis_query_resp() function.  Since spin_lock_bh() was
> > added to dev_txq_stats_fold() the call started causing
> > warnings.  This is because gen_ndis_query_resp() can be
> > (indirectly) called from rndis_command_complete() which is
> > called with interrupts disabled.
> > 
> > While at it, this commit also changes the
> > gen_ndis_query_resp() function so that "net" local variable is
> > used instead of "rndis_per_dev_params[configNr].dev" in all
> > places this is referenced.
> > 
> > Signed-off-by: Michal Nazarewicz <m.nazarewicz@samsung.com>
> 
> The way this works is dev_get_stats() takes that "temp" object the
> caller provides, it writes the correct statistics into it (with any
> necessary translations), and then passes back a pointer to it.
> 
> See net/core/dev.c:dev_get_stats()
> 
> You must use dev_get_stats() or else the statistics won't be properly
> converted.
> 
> Like I originally suggested, you need to rearrange the code in this
> driver such that the gen_ndis_query_resp() work happens in a tasklet,
> workqueue, or some other non-hardware-irq context.

Hmm, maybe its a bit difficult to fix this problem for stable kernel.

What we could do is assume rndis wont use a device driver that actually
needs txq tx stats folding, and just use following interim patch ?

diff --git a/net/core/dev.c b/net/core/dev.c
index b9b22a3..31d5424 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -5288,6 +5288,14 @@ void dev_txq_stats_fold(const struct net_device *dev,
 	unsigned int i;
 	struct netdev_queue *txq;
 
+	/* temporary hack : rndis calls us under hard irq */
+	if (WARN_ON_ONCE(in_irq())) {
+		stats->tx_bytes   = dev->stats.tx_bytes;
+		stats->tx_packets = dev->stats.tx_packets;
+		stats->tx_dropped = dev->stats.tx_dropped;
+		return;
+	}
+
 	for (i = 0; i < dev->num_tx_queues; i++) {
 		txq = netdev_get_tx_queue(dev, i);
 		spin_lock_bh(&txq->_xmit_lock);