Hi Jason, On Mon, 23 Sept 2024 at 20:52, Jason Gunthorpe <jgg@xxxxxxxx> wrote: > > On Fri, Sep 20, 2024 at 02:14:02PM +0800, Yunsheng Lin wrote: > > > I am not sure what dose the API that allows netdev to "give" struct device > > to page_pool look like or how to implement the API yet, but the obvious way > > to stall the calling of device_del() is to wait for the inflight > > page to > > It is not device_del() you need to stall, but the remove() function of > the device driver. > > Once all drivers have been unbound the DMA API can be reconfigured and > all existing DMA mappings must be concluded before this happens, > otherwise there will be problems. > > So, stalling something like unregister_netdevice() would be a better > target - though stalling forever on driver unbind would not be > acceptable. TBH, I have doubts that even stalling it for small amounts of time is going to disrupt userspace and people are going to yell at us. I am gonna repeat myself here, but I think keeping a list of the inflight SKBs that we need to unmap when the interface goes down, is the most complex, but less disruptive solution Thanks /Ilias > > Jason
