[PATCH 1/6] drm/ttm: Add unampping of the entire device address space

[Daniel Vetter]

On Wed, Jun 10, 2020 at 04:05:04PM +0200, Christian König wrote:
> Am 10.06.20 um 15:54 schrieb Andrey Grodzovsky:
> > 
> > 
> > On 6/10/20 6:15 AM, Thomas Hellström (Intel) wrote:
> > > 
> > > 
> > > On 6/9/20 7:21 PM, Koenig, Christian wrote:
> > > > 
> > > > 
> > > > Am 09.06.2020 18:37 schrieb "Grodzovsky, Andrey"
> > > > <Andrey.Grodzovsky at amd.com>:
> > > > 
> > > > 
> > > >     On 6/5/20 2:40 PM, Christian König wrote:
> > > >     > Am 05.06.20 um 16:29 schrieb Andrey Grodzovsky:
> > > >     >>
> > > >     >> On 5/11/20 2:45 AM, Christian König wrote:
> > > >     >>> Am 09.05.20 um 20:51 schrieb Andrey Grodzovsky:
> > > >     >>>> Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky at amd.com>
> > > >     >>>> ---
> > > >     >>>> drivers/gpu/drm/ttm/ttm_bo.c    | 22 +++++++++++++++++++++-
> > > >     >>>> include/drm/ttm/ttm_bo_driver.h |  2 ++
> > > >     >>>>   2 files changed, 23 insertions(+), 1 deletion(-)
> > > >     >>>>
> > > >     >>>> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c
> > > >     >>>> b/drivers/gpu/drm/ttm/ttm_bo.c
> > > >     >>>> index c5b516f..eae61cc 100644
> > > >     >>>> --- a/drivers/gpu/drm/ttm/ttm_bo.c
> > > >     >>>> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
> > > >     >>>> @@ -1750,9 +1750,29 @@ void ttm_bo_unmap_virtual(struct
> > > >     >>>> ttm_buffer_object *bo)
> > > >     >>>> ttm_bo_unmap_virtual_locked(bo);
> > > >     >>>> ttm_mem_io_unlock(man);
> > > >     >>>>   }
> > > >     >>>> +EXPORT_SYMBOL(ttm_bo_unmap_virtual);
> > > >     >>>>   +void ttm_bo_unmap_virtual_address_space(struct
> > > >     ttm_bo_device *bdev)
> > > >     >>>> +{
> > > >     >>>> +    struct ttm_mem_type_manager *man;
> > > >     >>>> +    int i;
> > > >     >>>> -EXPORT_SYMBOL(ttm_bo_unmap_virtual);
> > > >     >>>
> > > >     >>>> +    for (i = 0; i < TTM_NUM_MEM_TYPES; i++) {
> > > >     >>>> +        man = &bdev->man[i];
> > > >     >>>> +        if (man->has_type && man->use_type)
> > > >     >>>> + ttm_mem_io_lock(man, false);
> > > >     >>>> +    }
> > > >     >>>
> > > >     >>> You should drop that it will just result in a deadlock
> > > >     warning for
> > > >     >>> Nouveau and has no effect at all.
> > > >     >>>
> > > >     >>> Apart from that looks good to me,
> > > >     >>> Christian.
> > > >     >>
> > > >     >>
> > > >     >> As I am considering to re-include this in V2 of the
> > > >     patchsets, can
> > > >     >> you clarify please why this will have no effect at all ?
> > > >     >
> > > >     > The locks are exclusive for Nouveau to allocate/free the io
> > > >     address
> > > >     > space.
> > > >     >
> > > >     > Since we don't do this here we don't need the locks.
> > > >     >
> > > >     > Christian.
> > > > 
> > > > 
> > > >     So basically calling unmap_mapping_range doesn't require any extra
> > > >     locking around it and whatever locks are taken within the function
> > > >     should be enough ?
> > > > 
> > > > 
> > > > 
> > > > I think so, yes.
> > > > 
> > > > Christian.
> > > 
> > > Yes, that's true. However, without the bo reservation, nothing stops
> > > a PTE from being immediately re-faulted back again. Even while
> > > unmap_mapping_range() is running.
> > > 
> > 
> > Can you explain more on this - specifically, which function to reserve
> > the BO, why BO reservation would prevent re-fault of the PTE ?
> > 
> 
> Thomas is talking about ttm_bo_reserver()/ttm_bo_unreserve(), but we don't
> need this because we unmap everything because the whole device is gone and
> not just manipulate a single BO.
> 
> > 
> > > So the device removed flag needs to be advertized before this
> > > function is run,
> > > 
> > 
> > I indeed intend to call this  right after calling drm_dev_unplug from
> > amdgpu_pci_remove while adding drm_dev_enter/exit in ttm_bo_vm_fault (or
> > in amdgpu specific wrapper since I don't see how can I access struct
> > drm_device from ttm_bo_vm_fault) and this in my understanding should
> > stop a PTE from being re-faulted back as you pointed out - so again I
> > don't see how  bo reservation would prevent it so it looks like I am
> > missing something...
> > 
> > 
> > > (perhaps with a memory barrier pair).
> > > 
> > 
> > drm_dev_unplug and drm_dev_enter/exit are RCU synchronized and so I
> > don't think require any extra memory barriers for visibility of the
> > removed flag being set
> > 
> 
> As far as I can see that should be perfectly sufficient.

Only if you have a drm_dev_enter/exit pair in your fault handler.
Otherwise you're still open to the races Thomas described. But aside from
that the drm_dev_unplug stuff has all the barriers and stuff to make sure
nothing escapes.

Failure to drm_dev_enter could then also trigger the special case where we
put a dummy page in place.
-Daniel

> 
> Christian.
> 
> > 
> > Andrey
> > 
> > 
> > > That should probably be added to the function documentation.
> > > 
> > > (Other than that, please add a commit message if respinning).
> > > 
> > > /Thomas
> > > 
> > > 
> > > 
> 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch