Re: [PATCH] netconsole: release the spinlock before __netpoll_cleanup()

[Veaceslav Falico <vfalico@redhat.com>]

On Wed, Mar 06, 2013 at 07:08:24PM -0500, Neil Horman wrote:
>On Wed, Mar 06, 2013 at 03:46:43PM +0100, Veaceslav Falico wrote:
>> Commit 3335f0ca130c201f8680e97f63612053fbc16e22 removed spinlock unlocking
>> before __netpoll_cleanup() in netconsole_netdev_event(), however we still
>> might sleep in __netpoll_cleanup() - via synchronize_srcu(). Revert it and
>> add a comment.
>>
>> Signed-off-by: Veaceslav Falico <vfalico@redhat.com>
>> ---
>>  drivers/net/netconsole.c |   10 ++++++++++
>>  1 files changed, 10 insertions(+), 0 deletions(-)
>>
>> diff --git a/drivers/net/netconsole.c b/drivers/net/netconsole.c
>> index 37add21..dd62b4c 100644
>> --- a/drivers/net/netconsole.c
>> +++ b/drivers/net/netconsole.c
>> @@ -680,7 +680,17 @@ static int netconsole_netdev_event(struct notifier_block *this,
>>  				 * rtnl_lock already held
>>  				 */
>>  				if (nt->np.dev) {
>> +					/*
>> +					 * we still might sleep in
>> +					 * __netpoll_cleanup(), so release
>> +					 * the lock
>> +					 */
>> +					spin_unlock_irqrestore(
>> +							      &target_list_lock,
>> +							      flags);
>>  					__netpoll_cleanup(&nt->np);
>> +					spin_lock_irqsave(&target_list_lock,
>> +							  flags);
>>  					dev_put(nt->np.dev);
>>  					nt->np.dev = NULL;
>>  				}
>> --
>> 1.7.1
>>
>Thanks for noticing this Vaeceslav, but you can't just drop and re-acquire the
>lock like this, as it protect the list_for_each_entry loop that you're in.  You
>can drop the lock in the above if clause, but then, after the nt->np.dev = NULL,
>go back an re-aquire the lock, and start the for loop.  I thought we had already
>done this for some other purpose in this code using a label and a goto, but I
>suppose I was mistaken

You're right, I somehow missed that restart goto, which was removed
earlier. Does that feel right (I've also added back the
netconsole_target_put()):

Subject: [PATCH] netconsole: release the spinlock before __netpoll_cleanup()

__netpoll_cleanup() might sleep in synchronize_srcu(), which was added to
avoid race in another situation, so we can't call it with the spinlock
target_list_lock held.

Add spin_unlock/lock before/after it and restart the loop.

Signed-off-by: Veaceslav Falico <vfalico@redhat.com>
---
  drivers/net/netconsole.c |   13 +++++++++++++
  1 files changed, 13 insertions(+), 0 deletions(-)

diff --git a/drivers/net/netconsole.c b/drivers/net/netconsole.c
index 37add21..267c26b 100644
--- a/drivers/net/netconsole.c
+++ b/drivers/net/netconsole.c
@@ -666,6 +666,7 @@ static int netconsole_netdev_event(struct notifier_block *this,
  		goto done;
  
  	spin_lock_irqsave(&target_list_lock, flags);
+restart:
  	list_for_each_entry(nt, &target_list, list) {
  		netconsole_target_get(nt);
  		if (nt->np.dev == dev) {
@@ -680,9 +681,21 @@ static int netconsole_netdev_event(struct notifier_block *this,
  				 * rtnl_lock already held
  				 */
  				if (nt->np.dev) {
+					/*
+					 * we still might sleep in
+					 * __netpoll_cleanup(), so release
+					 * the lock and restart
+					 */
+					spin_unlock_irqrestore(
+							      &target_list_lock,
+							      flags);
  					__netpoll_cleanup(&nt->np);
+					spin_lock_irqsave(&target_list_lock,
+							  flags);
  					dev_put(nt->np.dev);
  					nt->np.dev = NULL;
+					netconsole_target_put(nt);
+					goto restart;
  				}
  				nt->enabled = 0;
  				stopped = true;
-- 
1.7.1