On Thu, 2025-03-06 at 10:02 +0000, Hangbin Liu wrote:
For bond_ipsec_add_sa_all(), I will move the xso.real_dev = real_dev after .xdo_dev_state_add() in case the following situation.
xso.real_dev needs to be initialized before the call to xdo_dev_state_add, since many of the implementations look in xso.real_dev to determine on which device to operate on. So the ordering should be: - get the lock - set xso.real_dev to real_dev - release the lock - call xdo_dev_state_add - if it fails, reacquire the lock and set the device to NULL.
Unfortunately, this doesn't seem to protect against the scenario below, as after dropping the spinlock from bond_ipsec_add_sa_all, bond_ipsec_del_sa can freely call xdo_dev_state_delete() on real_dev before xdo_dev_state_add happens.
I don't know what to do in this case...
bond_ipsec_add_sa_all() spin_unlock(&ipsec->x->lock); ipsec->xs->xso.real_dev = real_dev; __xfrm_state_delete x->state = DEAD - bond_ipsec_del_sa() - .xdo_dev_state_delete() .xdo_dev_state_add()
Cosmin.