Are xfrm state_add/delete() calls serialized?

Are xfrm state_add/delete() calls serialized?

[Bharat Bhushan]

Hi All,

Have a very basic query related to .xdo_dev_state_add()/delete() ops supported by netdev driver. Can .xdo_dev_state_add()/delete() execute from other core while already in process of handling .xdo_dev_state_add()/delete() on one core? Or these calls are always serialized by stack? 
Wanted to know if we need proper locking while handling these ops in driver.

Thanks
-Bharat 

Re: Are xfrm state_add/delete() calls serialized?

[Leon Romanovsky]

On Mon, Sep 12, 2022 at 03:10:12PM +0000, Bharat Bhushan wrote:
> Hi All,
> 
> Have a very basic query related to .xdo_dev_state_add()/delete() ops supported by netdev driver. Can .xdo_dev_state_add()/delete() execute from other core while already in process of handling .xdo_dev_state_add()/delete() on one core? Or these calls are always serialized by stack?

It is protected from userspace callers with xfrm_cfg_mutex in xfrm_netlink_rcv().
However, stack triggered deletion can be in parallel. There is a lock
for that specific SA that is going to be deleted, and it is not global.

> Wanted to know if we need proper locking while handling these ops in driver.
> 
> Thanks
> -Bharat 

RE: [EXT] Re: Are xfrm state_add/delete() calls serialized?

[Bharat Bhushan]

Please see inline 

> -----Original Message-----
> From: Leon Romanovsky <leon@kernel.org>
> Sent: Monday, September 19, 2022 3:00 PM
> To: Bharat Bhushan <bbhushan2@marvell.com>
> Cc: netdev@vger.kernel.org
> Subject: [EXT] Re: Are xfrm state_add/delete() calls serialized?
> 
> External Email
> 
> ----------------------------------------------------------------------
> On Mon, Sep 12, 2022 at 03:10:12PM +0000, Bharat Bhushan wrote:
> > Hi All,
> >
> > Have a very basic query related to .xdo_dev_state_add()/delete() ops
> supported by netdev driver. Can .xdo_dev_state_add()/delete() execute from
> other core while already in process of handling .xdo_dev_state_add()/delete()
> on one core? Or these calls are always serialized by stack?
> 
> It is protected from userspace callers with xfrm_cfg_mutex in xfrm_netlink_rcv().

So all *_state_add() and _state_delete() are serialized from user.

> However, stack triggered deletion can be in parallel. There is a lock for that
> specific SA that is going to be deleted, and it is not global.

Just want to confirm m understanding, xfrm_state->lock is used by stack (example xfrm_timer_handler()) for deletion, but this lock is per SA (not global).
So _state_delete() of different SA can happen in parallel and also _state_delete() by stack can run in parallel to state addition from user.

Thanks
-Bharat

> 
> > Wanted to know if we need proper locking while handling these ops in driver.
> >
> > Thanks
> > -Bharat

Re: [EXT] Re: Are xfrm state_add/delete() calls serialized?

[Leon Romanovsky]

On Wed, Sep 21, 2022 at 05:42:22AM +0000, Bharat Bhushan wrote:
> Please see inline 
> 
> > -----Original Message-----
> > From: Leon Romanovsky <leon@kernel.org>
> > Sent: Monday, September 19, 2022 3:00 PM
> > To: Bharat Bhushan <bbhushan2@marvell.com>
> > Cc: netdev@vger.kernel.org
> > Subject: [EXT] Re: Are xfrm state_add/delete() calls serialized?
> > 
> > External Email
> > 
> > ----------------------------------------------------------------------
> > On Mon, Sep 12, 2022 at 03:10:12PM +0000, Bharat Bhushan wrote:
> > > Hi All,
> > >
> > > Have a very basic query related to .xdo_dev_state_add()/delete() ops
> > supported by netdev driver. Can .xdo_dev_state_add()/delete() execute from
> > other core while already in process of handling .xdo_dev_state_add()/delete()
> > on one core? Or these calls are always serialized by stack?
> > 
> > It is protected from userspace callers with xfrm_cfg_mutex in xfrm_netlink_rcv().
> 
> So all *_state_add() and _state_delete() are serialized from user.
> 
> > However, stack triggered deletion can be in parallel. There is a lock for that
> > specific SA that is going to be deleted, and it is not global.
> 
> Just want to confirm m understanding, xfrm_state->lock is used by stack (example xfrm_timer_handler()) for deletion, but this lock is per SA (not global).
> So _state_delete() of different SA can happen in parallel and also _state_delete() by stack can run in parallel to state addition from user.

Right

> 
> Thanks
> -Bharat
> 
> > 
> > > Wanted to know if we need proper locking while handling these ops in driver.
> > >
> > > Thanks
> > > -Bharat