netdev.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Vitaly Kuznetsov <vkuznets@redhat.com>
To: KY Srinivasan <kys@microsoft.com>
Cc: "davem\@davemloft.net" <davem@davemloft.net>,
	"netdev\@vger.kernel.org" <netdev@vger.kernel.org>,
	"linux-kernel\@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	"devel\@linuxdriverproject.org" <devel@linuxdriverproject.org>,
	"olaf\@aepfle.de" <olaf@aepfle.de>,
	"apw\@canonical.com" <apw@canonical.com>,
	"jasowang\@redhat.com" <jasowang@redhat.com>,
	Dexuan Cui <decui@microsoft.com>
Subject: Re: [PATCH net-next 1/1] hv_netvsc: Wait for sub-channels to be processed during probe
Date: Fri, 17 Jul 2015 18:09:44 +0200	[thread overview]
Message-ID: <877fpyn4tz.fsf@vitty.brq.redhat.com> (raw)
In-Reply-To: <BY2PR0301MB16542B6DF558055287B1B7F1A0980@BY2PR0301MB1654.namprd03.prod.outlook.com> (KY Srinivasan's message of "Fri, 17 Jul 2015 15:53:54 +0000")

KY Srinivasan <kys@microsoft.com> writes:

>> -----Original Message-----
>> From: Vitaly Kuznetsov [mailto:vkuznets@redhat.com]
>> Sent: Friday, July 17, 2015 7:13 AM
>> To: KY Srinivasan
>> Cc: davem@davemloft.net; netdev@vger.kernel.org; linux-
>> kernel@vger.kernel.org; devel@linuxdriverproject.org; olaf@aepfle.de;
>> apw@canonical.com; jasowang@redhat.com; Dexuan Cui
>> Subject: Re: [PATCH net-next 1/1] hv_netvsc: Wait for sub-channels to be
>> processed during probe
>> 
>> "K. Y. Srinivasan" <kys@microsoft.com> writes:
>> 
>> > The current code returns from probe without waiting for the proper
>> handling
>> > of subchannels that may be requested. If the netvsc driver were to be
>> rapidly
>> > loaded/unloaded, we can  trigger a panic as the unload will be tearing
>> > down state that may not have been fully setup yet. We fix this issue by
>> making
>> > sure that we return from the probe call only after ensuring that the
>> > sub-channel offers in flight are properly handled.
>> >
>> > Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
>> > Reviewed-and-tested-by: Haiyang Zhang <haiyangz@microsoft.com
>> > ---
>> >  drivers/net/hyperv/hyperv_net.h   |    2 ++
>> >  drivers/net/hyperv/rndis_filter.c |   25 +++++++++++++++++++++++++
>> >  2 files changed, 27 insertions(+), 0 deletions(-)
>> >
>> > diff --git a/drivers/net/hyperv/hyperv_net.h
>> b/drivers/net/hyperv/hyperv_net.h
>> > index 26cd14c..925b75d 100644
>> > --- a/drivers/net/hyperv/hyperv_net.h
>> > +++ b/drivers/net/hyperv/hyperv_net.h
>> > @@ -671,6 +671,8 @@ struct netvsc_device {
>> >  	u32 send_table[VRSS_SEND_TAB_SIZE];
>> >  	u32 max_chn;
>> >  	u32 num_chn;
>> > +	spinlock_t sc_lock; /* Protects num_sc_offered variable */
>> > +	u32 num_sc_offered;
>> >  	atomic_t queue_sends[NR_CPUS];
>> >
>> >  	/* Holds rndis device info */
>> > diff --git a/drivers/net/hyperv/rndis_filter.c
>> b/drivers/net/hyperv/rndis_filter.c
>> > index 2e40417..2e09f3f 100644
>> > --- a/drivers/net/hyperv/rndis_filter.c
>> > +++ b/drivers/net/hyperv/rndis_filter.c
>> > @@ -984,9 +984,16 @@ static void netvsc_sc_open(struct vmbus_channel
>> *new_sc)
>> >  	struct netvsc_device *nvscdev;
>> >  	u16 chn_index = new_sc->offermsg.offer.sub_channel_index;
>> >  	int ret;
>> > +	unsigned long flags;
>> >
>> >  	nvscdev = hv_get_drvdata(new_sc->primary_channel->device_obj);
>> >
>> > +	spin_lock_irqsave(&nvscdev->sc_lock, flags);
>> > +	nvscdev->num_sc_offered--;
>> > +	spin_unlock_irqrestore(&nvscdev->sc_lock, flags);
>> > +	if (nvscdev->num_sc_offered == 0)
>> > +		complete(&nvscdev->channel_init_wait);
>> > +
>> >  	if (chn_index >= nvscdev->num_chn)
>> >  		return;
>> >
>> > @@ -1015,8 +1022,10 @@ int rndis_filter_device_add(struct hv_device
>> *dev,
>> >  	u32 rsscap_size = sizeof(struct ndis_recv_scale_cap);
>> >  	u32 mtu, size;
>> >  	u32 num_rss_qs;
>> > +	u32 sc_delta;
>> >  	const struct cpumask *node_cpu_mask;
>> >  	u32 num_possible_rss_qs;
>> > +	unsigned long flags;
>> >
>> >  	rndis_device = get_rndis_device();
>> >  	if (!rndis_device)
>> > @@ -1039,6 +1048,8 @@ int rndis_filter_device_add(struct hv_device
>> *dev,
>> >  	net_device->max_chn = 1;
>> >  	net_device->num_chn = 1;
>> >
>> > +	spin_lock_init(&net_device->sc_lock);
>> > +
>> >  	net_device->extension = rndis_device;
>> >  	rndis_device->net_dev = net_device;
>> >
>> > @@ -1116,6 +1127,9 @@ int rndis_filter_device_add(struct hv_device
>> *dev,
>> >  	num_possible_rss_qs = cpumask_weight(node_cpu_mask);
>> >  	net_device->num_chn = min(num_possible_rss_qs, num_rss_qs);
>> >
>> > +	num_rss_qs = net_device->num_chn - 1;
>> > +	net_device->num_sc_offered = num_rss_qs;
>> > +
>> >  	if (net_device->num_chn == 1)
>> >  		goto out;
>> >
>> > @@ -1157,11 +1171,22 @@ int rndis_filter_device_add(struct hv_device
>> *dev,
>> >
>> >  	ret = rndis_filter_set_rss_param(rndis_device, net_device-
>> >num_chn);
>> >
>> > +	/*
>> > +	 * Wait for the host to send us the sub-channel offers.
>> > +	 */
>> > +	spin_lock_irqsave(&net_device->sc_lock, flags);
>> > +	sc_delta = net_device->num_chn - 1 - num_rss_qs;
>> > +	net_device->num_sc_offered -= sc_delta;
>> > +	spin_unlock_irqrestore(&net_device->sc_lock, flags);
>> > +
>> > +	if (net_device->num_sc_offered != 0)
>> > +		wait_for_completion(&net_device->channel_init_wait);
>> 
>> I'd suggest we add an essentian timeout (big, let's say 30 sec.)
>> here. In case something goes wrong we don't really want to hang the
>> whole kernel for forever. Such bugs are hard to debug as if a 'kernel
>> hangs' is reported we can't be sure which wait caused it. We can even
>> have something like:
>> 
>>  t = wait_for_completion_timeout(&net_device->channel_init_wait, 30*HZ);
>>  BUG_ON(t == 0);
>> 
>> This is much better as we'll be sure what went wrong. (I know other
>> pieces of hyper-v code use wait_for_completion() without a timeout, this
>> is rather a general suggestion for all of them).
>
> There is some history here. Initially, I had timeout for calls where we could reasonably
> rollback state if we timed out. Some calls were subsequently changed to unconditional
> wait because under some load conditions, these timeouts would trigger (granted I did not
> have 30 second timeout; it was a 5 second timeout).
>
> Greg was opposed to calls to BUG_ON() in general for drivers.

Even WARN() in my opinion should do the job: we'll at least have
something in the log.

> Consequently, I have chosen to
> wait unconditionally in cases where there is no sensible rollback if we were to timeout (rather
> thank bug checking).

(irrelevant to this patch but) in that case I'd suggest we have
something like:

while (1) {
     t = wait_for_completion_timeout(&..., 10*HZ);
     if (t)
         break;
     WARN(1, "Timeout while waiting for hypervisor to reply! Keep waiting...");
}

The sole purpose of that is to have something in dmesg.

>
> In this specific case though, we can have a timeout and we don't need to bug check if we timeout.
>
> I will make this change and resubmit.

Thanks!

>
> Thanks,
>
> K. Y 
>> 
>> >  out:
>> >  	if (ret) {
>> >  		net_device->max_chn = 1;
>> >  		net_device->num_chn = 1;
>> >  	}
>> > +
>> >  	return 0; /* return 0 because primary channel can be used alone */
>> >
>> >  err_dev_remv:
>> 
>> --
>>   Vitaly

-- 
  Vitaly

  reply	other threads:[~2015-07-17 16:09 UTC|newest]

Thread overview: 11+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2015-07-16 19:17 [PATCH net-next 1/1] hv_netvsc: Wait for sub-channels to be processed during probe K. Y. Srinivasan
2015-07-17 10:00 ` Dexuan Cui
2015-07-17 15:33   ` KY Srinivasan
2015-07-20  9:39     ` Dexuan Cui
2015-07-20 19:00       ` KY Srinivasan
2015-07-17 10:07 ` Dexuan Cui
2015-07-17 15:34   ` KY Srinivasan
2015-07-17 14:12 ` Vitaly Kuznetsov
2015-07-17 15:53   ` KY Srinivasan
2015-07-17 16:09     ` Vitaly Kuznetsov [this message]
2015-07-17 16:13       ` KY Srinivasan

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=877fpyn4tz.fsf@vitty.brq.redhat.com \
    --to=vkuznets@redhat.com \
    --cc=apw@canonical.com \
    --cc=davem@davemloft.net \
    --cc=decui@microsoft.com \
    --cc=devel@linuxdriverproject.org \
    --cc=jasowang@redhat.com \
    --cc=kys@microsoft.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=netdev@vger.kernel.org \
    --cc=olaf@aepfle.de \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).