On Mon 13-06-16 22:52:43, Tetsuo Handa wrote: > Michal Hocko wrote: > > I have checked the vnet code and it doesn't seem to rely on > > copy_from_user/get_user AFAICS. Other users of use_mm() need to copy to > > the userspace only as well. So we should be perfectly safe to OOM reap > > address space even when it is shared by the kthread [1] so this is > > not really needed for the OOM correctness purpose. It would be much > > nicer if the kthread didn't pin the mm for two long outside of the OOM > > handling as well of course but that lowers the priority of the change. > > > > [1] http://lkml.kernel.org/r/[email protected] > > It seems to me that vhost code relies on copy from the userspace. > > use_mm(dev->mm) and unuse_mm(dev->mm) are used inside vhost_worker(). > work->fn(work) is initialized by vhost_work_init(). > vhost_scsi_open() passes vhost_scsi_complete_cmd_work() and > vhost_scsi_evt_work() as ->fn, and both functions call __get_user(). > > vhost_scsi_complete_cmd_work() { > vhost_signal() { > vhost_notify() { > __get_user() > } > } > } > > vhost_scsi_evt_work() { > vhost_scsi_do_evt_work() { > vhost_get_vq_desc() { > __get_user() / __copy_from_user() > get_indirect() { > copy_from_iter() > } > } > } > }
Ahh, I've missed those. Thanks for pointing this out! Let me try to find out whether the code is robust to see unexpected 0 when reading from the userspace. -- Michal Hocko SUSE Labs
