On 01/25/2012 07:12 PM, Dave Airlie wrote: > On Wed, Jan 25, 2012 at 5:19 PM, Thomas Hellstrom<thomas at shipmail.org> > wrote: >> On 01/25/2012 04:37 PM, Jerome Glisse wrote: >>> On Wed, Jan 25, 2012 at 10:21 AM, Ben Skeggs<skeggsb at gmail.com> wrote: >>>> On Wed, 2012-01-25 at 15:33 +0100, Thomas Hellstrom wrote: >>>>> On 01/25/2012 10:41 AM, Ben Skeggs wrote: >>>>>> My main concern is that we blindly and unnecessarily set up GPU >>>>>> bindings and >>>>>> end up with unnecessary code in TTM, and furthermore that we >>>>>> communicate >>>>>> that bad practice to future driver writers. >>>>>> This "unnecessary code" is like 5 lines of cleanup if something fails, >>>>>> hardly anything to be jumping up and down about :) >>>>> It's just not TTM's business, unless the GPU maps are mappable by the >>>>> CPU as well. >>>>> Also, What if the mapping setup in move_notify() fails? >>>> It can't fail, and well, in nouveau's implementation it never will. >>>> It's simply a "fill the ptes for all the vmas currently associated with >>>> a bo". >>>> >>>> And well, it's about as much TTM's business as VRAM aperture allocation >>>> is.. I don't see the big deal, if you wan't to do it a different way in >>>> your driver, there's nothing stopping you. It's a lot of bother for >>>> essentially zero effort in TTM.. >>>> >>>>>>> Thomas, what do you suggest to move forward with this? Both of these >>>>>>> bugs are serious regressions that make nouveau unusable with the >>>>>>> current >>>>>>> 3.3-rc series. Ben. >>>>>>> >>>>>>> My number one choice would of course be to have the drivers set up >>>>>>> their >>>>>>> private GPU mappings after a >>>>>>> successful validate call, as I originally suggested and you agreed to. >>>>>>> >>>>>>> If that's not possible (I realize it's late in the release series), >>>>>>> I'll >>>>>>> ack this patch if you and Jerome agree not to block >>>>>>> attempts to move in that direction for future kernel releases. >>>>>> I can't say I'm entirely happy with the plan honestly. To me, it still >>>>>> seems more efficient to handle this when a move happens (comparatively >>>>>> rare) and "map new backing storage into every vm that has a reference" >>>>>> than to (on every single buffer of every single "exec" call) go "is >>>>>> this >>>>>> buffer mapped into this channel's vm? yes, ok; no, lets go map it". >>>>>> >>>>>> I'm not even sure how exactly I plan on storing this mapping >>>>>> efficiently >>>>>> yet.. Scanning the BO's linked list of VMs it's mapped into for "if >>>>>> (this_vma == chan->vma)" doesn't exactly sound performant. >>>>> As previously suggested, in the simplest case a bo could have a 'needs >>>>> remap' flag >>>>> that is set on gpu map teardown on move_notify(), and when this flag is >>>>> detected in validate, >>>>> go ahead and set up all needed maps and clear that flag. >>>>> >>>>> This is the simplest case and more or less equivalent to the current >>>>> solution, except >>>>> maps aren't set up unless needed by at least one channel and there is a >>>>> clear way >>>>> to handle errors when GPU maps are set up. >>>> Yes, right. That can be done, and gives exactly the same functionality >>>> as I *already* achieve with move_notify() but apparently have to change >>>> just because you've decided nouveau/radeon are both doing the >>>> WrongThing(tm). >>>> >>>> Anyhow, I care more that 3.3 works than how it works. So, whatever. If >>>> I must agree to this in order to get a regression fix in, then I guess I >>>> really have no choice in the matter. >>>> >>>> Ben. >>>> >>>>> A simple and straightforward fix that leaves the path open (if so >>>>> desired) to >>>>> handle finer channel granularity. >>>>> >>>>> Or am I missing something? >>>>> >>> I went over the code and Ben fix is ok with me, i need to test it a >>> bit on radeon side. >>> >>> For long term solution why not just move most of the >>> ttm_bo_handle_move_mem to the driver. It would obsolete the >>> move_notify callback. move notify callback was introduced because in >>> some case the driver never knew directly that a bo moved. It's obvious >>> that driver need to know every time. So instead of having an ha-doc >>> function for that. Let just move the handle move stuff into the >>> driver. Yes there will be some code duplication but it will avoid >>> anykind of weird error path and driver will be able to perform what >>> ever make sense. >> >> Yes, this is a solution that eliminates the need for TTM to support private >> GPU map setup. Code duplication can largely be avoided if we >> collect common code in a small utility function. > Please this, its reduces one more place TTM suffers from midlayer effects. > > I'd much prefer the drivers be in control and call into help common > code instead of the driver calling in and then lots of callbacks out. > > (and yes the drm suffers from this a lot as well). > > Dave. OK, let's work in this direction.
/Thomas
