[PATCH] hv_netvsc: remove ndo_poll_controller

[PATCH] hv_netvsc: remove ndo_poll_controller

[Stephen Hemminger]

Similar to other patches from ERic.

As diagnosed by Song Liu, ndo_poll_controller() can
be very dangerous on loaded hosts, since the cpu
calling ndo_poll_controller() might steal all NAPI
contexts (for all RX/TX queues of the NIC). This capture
can last for unlimited amount of time, since one
cpu is generally not able to drain all the queues under load.

In netvsc driver it uses NAPI for TX completions. The default
poll_napi will do this for us now and avoid the capture.

Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: Eric Dumazet <edumazet@google.com>
---
 drivers/net/hyperv/netvsc_drv.c | 23 -----------------------
 1 file changed, 23 deletions(-)

diff --git a/drivers/net/hyperv/netvsc_drv.c b/drivers/net/hyperv/netvsc_drv.c
index ec699741170b..d7f843463bf2 100644
--- a/drivers/net/hyperv/netvsc_drv.c
+++ b/drivers/net/hyperv/netvsc_drv.c
@@ -1564,26 +1564,6 @@ netvsc_set_rxnfc(struct net_device *ndev, struct ethtool_rxnfc *info)
 	return -EOPNOTSUPP;
 }
 
-#ifdef CONFIG_NET_POLL_CONTROLLER
-static void netvsc_poll_controller(struct net_device *dev)
-{
-	struct net_device_context *ndc = netdev_priv(dev);
-	struct netvsc_device *ndev;
-	int i;
-
-	rcu_read_lock();
-	ndev = rcu_dereference(ndc->nvdev);
-	if (ndev) {
-		for (i = 0; i < ndev->num_chn; i++) {
-			struct netvsc_channel *nvchan = &ndev->chan_table[i];
-
-			napi_schedule(&nvchan->napi);
-		}
-	}
-	rcu_read_unlock();
-}
-#endif
-
 static u32 netvsc_get_rxfh_key_size(struct net_device *dev)
 {
 	return NETVSC_HASH_KEYLEN;
@@ -1811,9 +1791,6 @@ static const struct net_device_ops device_ops = {
 	.ndo_set_mac_address =		netvsc_set_mac_addr,
 	.ndo_select_queue =		netvsc_select_queue,
 	.ndo_get_stats64 =		netvsc_get_stats64,
-#ifdef CONFIG_NET_POLL_CONTROLLER
-	.ndo_poll_controller =		netvsc_poll_controller,
-#endif
 };
 
 /*
-- 
2.17.1

Re: [PATCH] hv_netvsc: remove ndo_poll_controller

[Stephen Hemminger]

On Sat, 29 Sep 2018 14:52:56 +0200
Stephen Hemminger <stephen@networkplumber.org> wrote:

> Similar to other patches from ERic.
> 
> As diagnosed by Song Liu, ndo_poll_controller() can
> be very dangerous on loaded hosts, since the cpu
> calling ndo_poll_controller() might steal all NAPI
> contexts (for all RX/TX queues of the NIC). This capture
> can last for unlimited amount of time, since one
> cpu is generally not able to drain all the queues under load.
> 
> In netvsc driver it uses NAPI for TX completions. The default
> poll_napi will do this for us now and avoid the capture.
> 
> Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
> Cc: Haiyang Zhang <haiyangz@microsoft.com>
> Cc: Eric Dumazet <edumazet@google.com>

Forgot to label this, it is for net-next.

RE: [PATCH] hv_netvsc: remove ndo_poll_controller

[Haiyang Zhang]

> -----Original Message-----
> From: Stephen Hemminger <stephen@networkplumber.org>
> Sent: Saturday, September 29, 2018 8:53 AM
> To: netdev@vger.kernel.org
> Cc: Stephen Hemminger <sthemmin@microsoft.com>; Haiyang Zhang
> <haiyangz@microsoft.com>; Eric Dumazet <edumazet@google.com>
> Subject: [PATCH] hv_netvsc: remove ndo_poll_controller
> 
> Similar to other patches from ERic.
> 
> As diagnosed by Song Liu, ndo_poll_controller() can be very dangerous on
> loaded hosts, since the cpu calling ndo_poll_controller() might steal all NAPI
> contexts (for all RX/TX queues of the NIC). This capture can last for unlimited
> amount of time, since one cpu is generally not able to drain all the queues
> under load.
> 
> In netvsc driver it uses NAPI for TX completions. The default poll_napi will do
> this for us now and avoid the capture.
> 
> Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
> Cc: Haiyang Zhang <haiyangz@microsoft.com>
> Cc: Eric Dumazet <edumazet@google.com>
> ---

Thank you.
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>

Re: [PATCH] hv_netvsc: remove ndo_poll_controller

[David Miller]

From: Stephen Hemminger <stephen@networkplumber.org>
Date: Sat, 29 Sep 2018 14:52:56 +0200

> Similar to other patches from ERic.
> 
> As diagnosed by Song Liu, ndo_poll_controller() can
> be very dangerous on loaded hosts, since the cpu
> calling ndo_poll_controller() might steal all NAPI
> contexts (for all RX/TX queues of the NIC). This capture
> can last for unlimited amount of time, since one
> cpu is generally not able to drain all the queues under load.
> 
> In netvsc driver it uses NAPI for TX completions. The default
> poll_napi will do this for us now and avoid the capture.
> 
> Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>

Applied to net-next.