> From: Michael Kelley <mikelley@xxxxxxxxxxxxx> > Sent: Wednesday, April 21, 2021 2:06 PM > ... > > Yes I think put_hvpcibus() and get_hvpcibus() can be removed, as we have > > changed to use > > a dedicated workqueue for hbus since they were introduced. > > > > But we still need to call tasklet_disable/enable() the same way > > hv_pci_suspend() does, the > > reason is that we need to protect hbus->state. This value needs to be > consistent for the > > driver. For example, a CPU may decide to schedule a work on a work queue > that we just > > flushed or destroyed, by reading the wrong hbus->state. > > > > Yes, I would agree the tasklet disable/enable are needed, especially since > tasklet_disable() > is what ensures that the tasklet is not currently running. > > If the hbus ref counting isn't needed any longer, I would strongly recommend > adding > a patch to the series that removes it. This synchronization stuff is hard > enough to > understand and reason about; having a leftover mechanism that doesn't really > do > anything useful makes it nearly impossible. :-) > > Dexuan -- I'm hoping you can take a look as well and see if you agree. > > Michael I also think we can remove the reference counting. But it looks like there is still race in hv_pci_remove() even with Long's patch: in hv_pci_remove(), we disable the tasklet, change hbus->state to hv_pcibus_removing, re-enable the tasklet and flush hbus->wq, and set hbus->state to hv_pcibus_removed -- what if the channel callback runs again? -- now hbus->state is no longer hv_pcibus_removing, so hv_pci_devices_present() -> hv_pci_start_relations_work() and hv_pci_eject_device() can still add new work items to hbus->wq, and the new work items may race with the vmbus_close(). It looks like we should remove the state hv_pcibus_removed? Thanks, -- Dexuan
