On Fri, Apr 24, 2020 at 11:36 AM John Stultz <john.stultz@xxxxxxxxxx> wrote: > On Fri, Apr 24, 2020 at 10:12 AM Jack Pham <jackp@xxxxxxxxxxxxxx> wrote: > > On Tue, Apr 21, 2020 at 10:09:27PM -0700, John Stultz wrote: > > > Does something like this make sense? It's not causing trouble on > > > db845c either so far in my testing. > > > > Ok I'll bite... > > > > I'm now curious why it hasn't been a problem with the Qualcomm HW. Do > > you mind please capturing a similar trace log on the db845c? Would be > > good to see a side-by-side comparison and see if, first of all, whether > > the same S/G path is getting exercised (i.e. 16KiB OUT requests from ADB > > userspace using AIO which then get broken up into 4K chunks by f_fs), > > and what the behaviors of the reclaim_trb and giveback are when the > > transfer is completed. > > > > Preferably if you could get a trace without your patch applied that > > would be great. And maybe also one after your patch just to see if the > > traces are truly identical or not. > > Sure. I've captured logs in the same manner with and without on db845c > (against 5.7-rc2). See attached. > > I suspect the difference is the db845c is using an iommu (I don't > think it will boot without it) where hikey960 isn't, but I'll let you > take a look. And I've added my own printk debugging that I used to track the issue down on HiKey960 and on the db845c we are never hitting the case where the HWO flag is set when we call dwc3_gadget_ep_reclaim_trb_sg() thanks -john
