svn commit: r186730 - in head: lib/libusb20 sys/dev/usb2/controller sys/dev/usb2/core sys/dev/usb2/ethernet sys/dev/usb2/image sys/dev/usb2/include sys/dev/usb2/serial sys/dev/usb2/sound sys/dev/us...

[Mark Tinguely]

>  On Wed, 7 Jan 2009 17:51:48 +0100
>  Hans Petter Selasky <hselasky at c2i.net> mentioned:
>
>  >  
>  >         cpu_idcache_wbinv_all(); /* and this line */
>  > 
>  >         bus_dmamap_sync(pc->tag, pc->map,
>  >             BUS_DMASYNC_POSTWRITE | BUS_DMASYNC_POSTREAD);
>  > }
>  > 
>  > Then everything worked.
>  > 
>  > I think the ones that have been working on busdma & cache sync issues recently 
>  > are to blame. I will try to investigate more.
>  > 
>
>  Hmm, this might be the same problem with duplicate dcache entries raj reported
>  recently on arm at . I tried replacing cpu_idcache_wbinv_all with busdma PREREAD
>  and PREWRITE sync, but this doesn't work.
>
>  BTW, even with cpu_idcache_wbinv_all change I'm still observing some CAM
>  request retires under high load, so we really seem to have a generic cache
>  issue somewhere.
>
>  - -- 
>  Stanislav Sedov
>  ST4096-RIPE

It could be the same problem if there is a duplicate kernel mapping. If
there is no duplicate kernel mapping, then there is another cache leak
somewhere.

Because a duplicate kernel mapping currently does not turn off the
caches in either mapping, even the wb/inv trick is a cache russian rollette
situation.

If someone wants to look at very rough code - I don't have a current box
to even compile it, I don't know anyone besides myself and Olivier Houchard
that is up on the ARM pmap code, but I am looking for someone who will
give this an eyeball. There can be consequences to adding pv_entrys where none
existed before. I don't want people to blindly download and attempt to run it.

My fix idea is to manage all mapped pages, with the expection of ficticous
pages. I don't think fictiicous pages can be mapped twice. We need to manage
them to turn caching off on all the mappings when there is multiple writers
or a writer/reader case.

--Mark Tinguely.