2014-11-26 09:07:27

by Jukka Rissanen

[permalink] [raw]
Subject: [PATCH v7] nl80211: Stop scheduled scan if netlink client disappears

An attribute NL80211_ATTR_SOCKET_OWNER can be set by the scan initiator.
If present, the attribute will cause the scan to be stopped if the client
dies.

Signed-off-by: Jukka Rissanen <[email protected]>
---
Hi,

v7:
- convert the cfg80211_sched_scan_request to __rcu pointer in order
to avoid races when accessing it
- reverting the patch v6, the port id is back in request struct

v6:
- moved owner netlink port id from cfg80211_sched_scan_request to
rdev in order to avoid possible races

v5:
- discarded the locking changes in v4
- instead of trying to schedule sched_scan_stop worker from
struct cfg80211_sched_scan_request, move the worker to wiphy
as that makes it easier to manage the sched_scan_stop worker.
There are also one scheduled scan / wiphy so it is also logical
to do it like this.

v4:
- rtnl locking issues fixed in patch 2

v3:
- backward compatibility define tweaked in patch 1
- added missing signed-off-by:

v2:
- split the patch
- In patch 1, use a generic NL80211_ATTR_SOCKET_OWNER attribute and
convert the old code that uses NL80211_ATTR_IFACE_SOCKET_OWNER to
use the new value. A define is provided for backward compatibility.
- Any pending schedule scan stop worker is cancelled when interface is
taken down in patch 2

Cheers,
Jukka


include/net/cfg80211.h | 2 ++
include/uapi/linux/nl80211.h | 3 +++
net/wireless/core.c | 28 +++++++++++++++++++++++++---
net/wireless/core.h | 4 +++-
net/wireless/nl80211.c | 40 +++++++++++++++++++++++++++++++++-------
net/wireless/scan.c | 27 +++++++++++++++++++++------
6 files changed, 87 insertions(+), 17 deletions(-)

diff --git a/include/net/cfg80211.h b/include/net/cfg80211.h
index bb748c4..0e540fc 100644
--- a/include/net/cfg80211.h
+++ b/include/net/cfg80211.h
@@ -1537,6 +1537,8 @@ struct cfg80211_sched_scan_request {
struct wiphy *wiphy;
struct net_device *dev;
unsigned long scan_start;
+ u32 owner_nlportid;
+ struct rcu_head rcu_head;

/* keep last */
struct ieee80211_channel *channels[0];
diff --git a/include/uapi/linux/nl80211.h b/include/uapi/linux/nl80211.h
index d775245..a0e3b32 100644
--- a/include/uapi/linux/nl80211.h
+++ b/include/uapi/linux/nl80211.h
@@ -1655,6 +1655,9 @@ enum nl80211_commands {
* @NL80211_ATTR_SOCKET_OWNER: Flag attribute, if set during interface
* creation then the new interface will be owned by the netlink socket
* that created it and will be destroyed when the socket is closed.
+ * If set during scheduled scan start then the new scan req will be
+ * owned by the netlink socket that created it and the scheduled scan will
+ * be stopped when the socket is closed.
*
* @NL80211_ATTR_TDLS_INITIATOR: flag attribute indicating the current end is
* the TDLS link initiator.
diff --git a/net/wireless/core.c b/net/wireless/core.c
index 4c2e501..368e276 100644
--- a/net/wireless/core.c
+++ b/net/wireless/core.c
@@ -320,6 +320,20 @@ static void cfg80211_destroy_iface_wk(struct work_struct *work)
rtnl_unlock();
}

+static void cfg80211_sched_scan_stop_wk(struct work_struct *work)
+{
+ struct cfg80211_registered_device *rdev;
+
+ rdev = container_of(work, struct cfg80211_registered_device,
+ sched_scan_stop_wk);
+
+ rtnl_lock();
+
+ __cfg80211_stop_sched_scan(rdev, false);
+
+ rtnl_unlock();
+}
+
/* exported functions */

struct wiphy *wiphy_new_nm(const struct cfg80211_ops *ops, int sizeof_priv,
@@ -406,6 +420,7 @@ use_default_name:
INIT_LIST_HEAD(&rdev->destroy_list);
spin_lock_init(&rdev->destroy_list_lock);
INIT_WORK(&rdev->destroy_work, cfg80211_destroy_iface_wk);
+ INIT_WORK(&rdev->sched_scan_stop_wk, cfg80211_sched_scan_stop_wk);

#ifdef CONFIG_CFG80211_DEFAULT_PS
rdev->wiphy.flags |= WIPHY_FLAG_PS_ON_BY_DEFAULT;
@@ -764,6 +779,7 @@ void wiphy_unregister(struct wiphy *wiphy)
flush_work(&rdev->event_work);
cancel_delayed_work_sync(&rdev->dfs_update_channels_wk);
flush_work(&rdev->destroy_work);
+ flush_work(&rdev->sched_scan_stop_wk);

#ifdef CONFIG_PM
if (rdev->wiphy.wowlan_config && rdev->ops->set_wakeup)
@@ -854,8 +870,11 @@ void __cfg80211_leave(struct cfg80211_registered_device *rdev,
break;
case NL80211_IFTYPE_P2P_CLIENT:
case NL80211_IFTYPE_STATION:
- if (rdev->sched_scan_req && dev == rdev->sched_scan_req->dev)
+ rcu_read_lock();
+ if (rcu_access_pointer(rdev->sched_scan_req) &&
+ dev == rtnl_dereference(rdev->sched_scan_req)->dev)
__cfg80211_stop_sched_scan(rdev, false);
+ rcu_read_unlock();

#ifdef CONFIG_CFG80211_WEXT
kfree(wdev->wext.ie);
@@ -993,10 +1012,13 @@ static int cfg80211_netdev_notifier_call(struct notifier_block *nb,
___cfg80211_scan_done(rdev, false);
}

- if (WARN_ON(rdev->sched_scan_req &&
- rdev->sched_scan_req->dev == wdev->netdev)) {
+ rcu_read_lock();
+ if (WARN_ON(rcu_access_pointer(rdev->sched_scan_req) &&
+ rtnl_dereference(rdev->sched_scan_req)->dev ==
+ wdev->netdev)) {
__cfg80211_stop_sched_scan(rdev, false);
}
+ rcu_read_unlock();

rdev->opencount--;
wake_up(&rdev->dev_wait);
diff --git a/net/wireless/core.h b/net/wireless/core.h
index faa5b16..4e3630b 100644
--- a/net/wireless/core.h
+++ b/net/wireless/core.h
@@ -63,7 +63,7 @@ struct cfg80211_registered_device {
u32 bss_generation;
struct cfg80211_scan_request *scan_req; /* protected by RTNL */
struct sk_buff *scan_msg;
- struct cfg80211_sched_scan_request *sched_scan_req;
+ struct cfg80211_sched_scan_request __rcu *sched_scan_req;
unsigned long suspend_at;
struct work_struct scan_done_wk;
struct work_struct sched_scan_results_wk;
@@ -84,6 +84,8 @@ struct cfg80211_registered_device {
struct list_head destroy_list;
struct work_struct destroy_work;

+ struct work_struct sched_scan_stop_wk;
+
/* must be last because of the way we do wiphy_priv(),
* and it should at least be aligned to NETDEV_ALIGN */
struct wiphy wiphy __aligned(NETDEV_ALIGN);
diff --git a/net/wireless/nl80211.c b/net/wireless/nl80211.c
index 6e41777..11ed9df 100644
--- a/net/wireless/nl80211.c
+++ b/net/wireless/nl80211.c
@@ -6077,27 +6077,40 @@ static int nl80211_start_sched_scan(struct sk_buff *skb,
if (rdev->sched_scan_req)
return -EINPROGRESS;

- rdev->sched_scan_req = nl80211_parse_sched_scan(&rdev->wiphy, wdev,
- info->attrs);
+ rcu_assign_pointer(rdev->sched_scan_req,
+ nl80211_parse_sched_scan(&rdev->wiphy, wdev,
+ info->attrs));
+ synchronize_rcu();
+
err = PTR_ERR_OR_ZERO(rdev->sched_scan_req);
if (err)
goto out_err;

- err = rdev_sched_scan_start(rdev, dev, rdev->sched_scan_req);
+ err = rdev_sched_scan_start(rdev, dev,
+ rcu_access_pointer(rdev->sched_scan_req));
if (err)
goto out_free;

- rdev->sched_scan_req->dev = dev;
- rdev->sched_scan_req->wiphy = &rdev->wiphy;
+ rcu_read_lock();
+
+ if (info->attrs[NL80211_ATTR_SOCKET_OWNER])
+ rtnl_dereference(rdev->sched_scan_req)->owner_nlportid =
+ info->snd_portid;
+
+ rtnl_dereference(rdev->sched_scan_req)->dev = dev;
+ rtnl_dereference(rdev->sched_scan_req)->wiphy = &rdev->wiphy;
+
+ rcu_read_unlock();

nl80211_send_sched_scan(rdev, dev,
NL80211_CMD_START_SCHED_SCAN);
return 0;

out_free:
- kfree(rdev->sched_scan_req);
+ kfree_rcu(rcu_access_pointer(rdev->sched_scan_req), rcu_head);
out_err:
- rdev->sched_scan_req = NULL;
+ rcu_assign_pointer(rdev->sched_scan_req, NULL);
+ synchronize_rcu();
return err;
}

@@ -12475,6 +12488,13 @@ static int nl80211_netlink_notify(struct notifier_block * nb,

list_for_each_entry_rcu(rdev, &cfg80211_rdev_list, list) {
bool schedule_destroy_work = false;
+ bool schedule_scan_stop = false;
+ struct cfg80211_sched_scan_request __rcu *req =
+ rdev->sched_scan_req;
+
+ if (rcu_access_pointer(req) && notify->portid &&
+ rcu_dereference(req)->owner_nlportid == notify->portid)
+ schedule_scan_stop = true;

list_for_each_entry_rcu(wdev, &rdev->wdev_list, list) {
cfg80211_mlme_unregister_socket(wdev, notify->portid);
@@ -12505,6 +12525,12 @@ static int nl80211_netlink_notify(struct notifier_block * nb,
spin_unlock(&rdev->destroy_list_lock);
schedule_work(&rdev->destroy_work);
}
+ } else if (schedule_scan_stop) {
+ rcu_dereference(req)->owner_nlportid = 0;
+
+ if (rdev->ops->sched_scan_stop &&
+ rdev->wiphy.flags & WIPHY_FLAG_SUPPORTS_SCHED_SCAN)
+ schedule_work(&rdev->sched_scan_stop_wk);
}
}

diff --git a/net/wireless/scan.c b/net/wireless/scan.c
index bda39f1..782d04f 100644
--- a/net/wireless/scan.c
+++ b/net/wireless/scan.c
@@ -257,7 +257,9 @@ void __cfg80211_sched_scan_results(struct work_struct *wk)

rtnl_lock();

- request = rdev->sched_scan_req;
+ rcu_read_lock();
+ request = rtnl_dereference(rdev->sched_scan_req);
+ rcu_read_unlock();

/* we don't have sched_scan_req anymore if the scan is stopping */
if (request) {
@@ -277,9 +279,16 @@ void __cfg80211_sched_scan_results(struct work_struct *wk)

void cfg80211_sched_scan_results(struct wiphy *wiphy)
{
+ struct cfg80211_sched_scan_request *request;
+
trace_cfg80211_sched_scan_results(wiphy);
/* ignore if we're not scanning */
- if (wiphy_to_rdev(wiphy)->sched_scan_req)
+
+ rcu_read_lock();
+ request = rcu_dereference(wiphy_to_rdev(wiphy)->sched_scan_req);
+ rcu_read_unlock();
+
+ if (request)
queue_work(cfg80211_wq,
&wiphy_to_rdev(wiphy)->sched_scan_results_wk);
}
@@ -309,13 +318,16 @@ int __cfg80211_stop_sched_scan(struct cfg80211_registered_device *rdev,
bool driver_initiated)
{
struct net_device *dev;
+ struct cfg80211_sched_scan_request *req;

ASSERT_RTNL();

- if (!rdev->sched_scan_req)
+ if (!rcu_access_pointer(rdev->sched_scan_req))
return -ENOENT;

- dev = rdev->sched_scan_req->dev;
+ rcu_read_lock();
+ dev = rtnl_dereference(rdev->sched_scan_req)->dev;
+ rcu_read_unlock();

if (!driver_initiated) {
int err = rdev_sched_scan_stop(rdev, dev);
@@ -325,8 +337,11 @@ int __cfg80211_stop_sched_scan(struct cfg80211_registered_device *rdev,

nl80211_send_sched_scan(rdev, dev, NL80211_CMD_SCHED_SCAN_STOPPED);

- kfree(rdev->sched_scan_req);
- rdev->sched_scan_req = NULL;
+ req = rcu_access_pointer(rdev->sched_scan_req);
+
+ rcu_assign_pointer(rdev->sched_scan_req, NULL);
+
+ kfree_rcu(req, rcu_head);

return 0;
}
--
1.8.3.1



2014-11-26 09:32:23

by Johannes Berg

[permalink] [raw]
Subject: Re: [PATCH v7] nl80211: Stop scheduled scan if netlink client disappears

On Wed, 2014-11-26 at 11:07 +0200, Jukka Rissanen wrote:
> An attribute NL80211_ATTR_SOCKET_OWNER can be set by the scan initiator.
> If present, the attribute will cause the scan to be stopped if the client
> dies.
>
> Signed-off-by: Jukka Rissanen <[email protected]>
> ---
> Hi,
>
> v7:
> - convert the cfg80211_sched_scan_request to __rcu pointer in order
> to avoid races when accessing it

You totally need to revisit how to use RCU.

Just some examples:
* don't synchronize_rcu() after assignment - what did you do that for?
* don't use rcu_read_lock() when you already have the write lock (rtnl)
* don't use rcu_access_pointer(), use
- local variables
- rcu_dereference_protected() (or rather rtnl_dereference())
- rcu_dereference()
* actually fix it to assign the values before rcu_assign_pointer() so
they're
all published together ...

Also, please make that a separate patch so it isn't mixed with the
functional changes.

johannes


2014-11-26 10:31:01

by Jukka Rissanen

[permalink] [raw]
Subject: Re: [PATCH v7] nl80211: Stop scheduled scan if netlink client disappears

On ke, 2014-11-26 at 10:32 +0100, Johannes Berg wrote:
> On Wed, 2014-11-26 at 11:07 +0200, Jukka Rissanen wrote:
> > An attribute NL80211_ATTR_SOCKET_OWNER can be set by the scan initiator.
> > If present, the attribute will cause the scan to be stopped if the client
> > dies.
> >
> > Signed-off-by: Jukka Rissanen <[email protected]>
> > ---
> > Hi,
> >
> > v7:
> > - convert the cfg80211_sched_scan_request to __rcu pointer in order
> > to avoid races when accessing it
>
> You totally need to revisit how to use RCU.

Yes, probably. This is very new stuff to me :)

>
> Just some examples:
> * don't synchronize_rcu() after assignment - what did you do that for?

Following the examples described in Documentation/RCU/whatisRCU.txt

> * don't use rcu_read_lock() when you already have the write lock (rtnl)
> * don't use rcu_access_pointer(), use
> - local variables
> - rcu_dereference_protected() (or rather rtnl_dereference())
> - rcu_dereference()
> * actually fix it to assign the values before rcu_assign_pointer() so
> they're
> all published together ...
>
> Also, please make that a separate patch so it isn't mixed with the
> functional changes.
>
> johannes
>

Jukka



2014-11-26 10:45:13

by Johannes Berg

[permalink] [raw]
Subject: Re: [PATCH v7] nl80211: Stop scheduled scan if netlink client disappears

On Wed, 2014-11-26 at 12:30 +0200, Jukka Rissanen wrote:

> > Just some examples:
> > * don't synchronize_rcu() after assignment - what did you do that for?
>
> Following the examples described in Documentation/RCU/whatisRCU.txt

Hmm, I don't think that documentation says that - was where you took it
from the example of replacing? But it's only needed for freeing, not for
assigning.

johannes