On Tue, Jul 23, 2019 at 09:31:35PM +0800, Jason Wang wrote: > > On 2019/7/23 下午5:26, Michael S. Tsirkin wrote: > > On Tue, Jul 23, 2019 at 04:49:01PM +0800, Jason Wang wrote: > > > On 2019/7/23 下午4:10, Michael S. Tsirkin wrote: > > > > On Tue, Jul 23, 2019 at 03:53:06PM +0800, Jason Wang wrote: > > > > > On 2019/7/23 下午3:23, Michael S. Tsirkin wrote: > > > > > > > > Really let's just use kfree_rcu. It's way cleaner: fire and forget. > > > > > > > Looks not, you need rate limit the fire as you've figured out? > > > > > > See the discussion that followed. Basically no, it's good enough > > > > > > already and is only going to be better. > > > > > > > > > > > > > And in fact, > > > > > > > the synchronization is not even needed, does it help if I leave a comment to > > > > > > > explain? > > > > > > Let's try to figure it out in the mail first. I'm pretty sure the > > > > > > current logic is wrong. > > > > > Here is what the code what to achieve: > > > > > > > > > > - The map was protected by RCU > > > > > > > > > > - Writers are: MMU notifier invalidation callbacks, file operations (ioctls > > > > > etc), meta_prefetch (datapath) > > > > > > > > > > - Readers are: memory accessor > > > > > > > > > > Writer are synchronized through mmu_lock. RCU is used to synchronized > > > > > between writers and readers. > > > > > > > > > > The synchronize_rcu() in vhost_reset_vq_maps() was used to synchronized it > > > > > with readers (memory accessors) in the path of file operations. But in this > > > > > case, vq->mutex was already held, this means it has been serialized with > > > > > memory accessor. That's why I think it could be removed safely. > > > > > > > > > > Anything I miss here? > > > > > > > > > So invalidate callbacks need to reset the map, and they do > > > > not have vq mutex. How can they do this and free > > > > the map safely? They need synchronize_rcu or kfree_rcu right? > > > Invalidation callbacks need but file operations (e.g ioctl) not. > > > > > > > > > > And I worry somewhat that synchronize_rcu in an MMU notifier > > > > is a problem, MMU notifiers are supposed to be quick: > > > Looks not, since it can allow to be blocked and lots of driver depends on > > > this. (E.g mmu_notifier_range_blockable()). > > Right, they can block. So why don't we take a VQ mutex and be > > done with it then? No RCU tricks. > > > This is how I want to go with RFC and V1. But I end up with deadlock between > vq locks and some MM internal locks. So I decide to use RCU which is 100% > under the control of vhost. > > Thanks And I guess the deadlock is because GUP is taking mmu locks which are taken on mmu notifier path, right? How about we add a seqlock and take that in invalidate callbacks? We can then drop the VQ lock before GUP, and take it again immediately after. something like if (!vq_meta_mapped(vq)) { vq_meta_setup(&uaddrs); mutex_unlock(vq->mutex) vq_meta_map(&uaddrs); mutex_lock(vq->mutex) /* recheck both sock->private_data and seqlock count. */ if changed - bail out } And also requires that VQ uaddrs is defined like this: - writers must have both vq mutex and dev mutex - readers must have either vq mutex or dev mutex That's a big change though. For now, how about switching to a per-vq SRCU? That is only a little bit more expensive than RCU, and we can use synchronize_srcu_expedited. -- MST