On Thu, 2009-04-23 at 17:11 +0200, Thomas Hellstrom wrote: > Jerome Glisse wrote: > > On Thu, 2009-04-23 at 14:29 +0200, Thomas Hellstrom wrote: > > > >> Jerome Glisse wrote: > >> > >>> On Thu, 2009-04-23 at 13:51 +0200, Thomas Hellstrom wrote: > >>> > >>> > >>>> Jerome Glisse wrote: > >>>> > >>>> > >>>>> Hi Thomas, > >>>>> > >>>>> It seems my path for bo move from system to vram is completely > >>>>> wrong, i really have hard time to understand this bo move it > >>>>> looks way more complicated than it should be. Here is what i > >>>>> do. > >>>>> > >>>>> tmp_mem = *old_mem; > >>>>> tmp_mem.mm_node = NULL; > >>>>> tmp_mem.proposed_flags = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING; > >>>>> r = ttm_bo_mem_space(bo, &tmp_mem, interruptible, no_wait); > >>>>> r = ttm_bo_move_ttm(bo, true, no_wait, &tmp_mem); > >>>>> r = radeon_move_blit(bo, true, no_wait, new_mem, old_mem); > >>>>> accel_cleanup > >>>>> > >>>>> So now on resume i get a bind of 0 pages because > >>>>> bo->ttm.state != unpopulated. Before resume this path seems > >>>>> to work fine. > >>>>> > >>>>> Am i doing somethings wrong ? > >>>>> > >>>>> > >>>>> > >>>>> > >>>> The question is then how you get an unpopulated ttm without any pages in > >>>> it? > >>>> If it's swapped, then the state should indeed be unpopulated. > >>>> > >>>> Also, why do you pass old_mem to radeon_move_blit, as that should point > >>>> to a system region that > >>>> is no longer valid? > >>>> > >>>> /Thomas > >>>> > >>>> > >>> bo->ttm exist and seems to have pages in it. in ttm_bo_move_ttm > >>> *old_mem = tmp_mem so when i do move_blit with old_mem it's in > >>> fact using the tmp mem allocated. > >>> > >>> > >> Hmm. That's a bit confusing to me. Can't radeon_blit just use bo->mem? > >> Anyway, you say you get a bind operation of 0 pages? Why is the ttm_tt > >> code not doing the bind? > >> > >> /Thomas > >> > >> > > > > It's confusing for me too, radeon_blit is just a blit function don't > > do anythings beside emitting blit gpu command and a fence. It's the > > ttm_tt code which calls bind. > > > > So ttm_tt_bind is called by ttm_bo_move_ttm, ttm_tt_bind call populate > > but state != unpopulated thus populate return right away, then > > ttm_tt_bind call driver callback bind but what i have in the ttm_backend > > object is 0 pages so i am asked to bind 0 pages. > > > > I can't make sense of the call chain, i will dump the whole object > > before and after suspend. > > > > > This sounds like the ttm::state and the backend is out-of-sync. > So the only place I can find, where the ttm state can transition from > ttm_unpopulated is in > ttm_tt_populate, with one exception: > > If ttm_tt_unbind is called while the TT is not in the bound state, when > ttm->state == ttm_unpopulated, > it may transition from ttm_unpopulated to ttm_unbound. > > So in ttm_tt_unbind, if you put the > ttm->state = ttm_unbound > > within the if statement, that help finding out what's going wrong. > > /Thomas > >
This fix the problem, i think unbind is call in cleanup evict path. Cheers, Jerome ------------------------------------------------------------------------------ Stay on top of everything new and different, both inside and around Java (TM) technology - register by April 22, and save $200 on the JavaOne (SM) conference, June 2-5, 2009, San Francisco. 300 plus technical and hands-on sessions. Register today. Use priority code J9JMT32. http://p.sf.net/sfu/p -- _______________________________________________ Dri-devel mailing list Dri-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/dri-devel
