Re: [PATCH v11 5/5] drm/amdgpu: add drm buddy support to amdgpu

[Matthew Auld <matthew.auld@xxxxxxxxx>]

On 08/02/2022 11:20, Arunpravin wrote:
> On 04/02/22 6:53 pm, Christian König wrote:
> > Am 04.02.22 um 12:22 schrieb Arunpravin:
> > > On 28/01/22 7:48 pm, Matthew Auld wrote:
> > > > On Thu, 27 Jan 2022 at 14:11, Arunpravin
> > > > <Arunpravin.PaneerSelvam@xxxxxxx> wrote:
> > > > > - Remove drm_mm references and replace with drm buddy functionalities
> > > > > - Add res cursor support for drm buddy
> > > > >
> > > > > v2(Matthew Auld):
> > > > >     - replace spinlock with mutex as we call kmem_cache_zalloc
> > > > >       (..., GFP_KERNEL) in drm_buddy_alloc() function
> > > > >
> > > > >     - lock drm_buddy_block_trim() function as it calls
> > > > >       mark_free/mark_split are all globally visible
> > > > >
> > > > > v3(Matthew Auld):
> > > > >     - remove trim method error handling as we address the failure case
> > > > >       at drm_buddy_block_trim() function
> > > > >
> > > > > v4:
> > > > >     - fix warnings reported by kernel test robot <lkp@xxxxxxxxx>
> > > > >
> > > > > v5:
> > > > >     - fix merge conflict issue
> > > > >
> > > > > v6:
> > > > >     - fix warnings reported by kernel test robot <lkp@xxxxxxxxx>
> > > > >
> > > > > Signed-off-by: Arunpravin <Arunpravin.PaneerSelvam@xxxxxxx>
> > > > > ---
> > > > >    drivers/gpu/drm/Kconfig                       |   1 +
> > > > >    .../gpu/drm/amd/amdgpu/amdgpu_res_cursor.h    |  97 +++++--
> > > > >    drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h       |   7 +-
> > > > >    drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c  | 259 ++++++++++--------
> > > > >    4 files changed, 231 insertions(+), 133 deletions(-)
> > > > <snip>
> > > >
> > > > > -/**
> > > > > - * amdgpu_vram_mgr_virt_start - update virtual start address
> > > > > - *
> > > > > - * @mem: ttm_resource to update
> > > > > - * @node: just allocated node
> > > > > - *
> > > > > - * Calculate a virtual BO start address to easily check if everything is CPU
> > > > > - * accessible.
> > > > > - */
> > > > > -static void amdgpu_vram_mgr_virt_start(struct ttm_resource *mem,
> > > > > -                                      struct drm_mm_node *node)
> > > > > -{
> > > > > -       unsigned long start;
> > > > > -
> > > > > -       start = node->start + node->size;
> > > > > -       if (start > mem->num_pages)
> > > > > -               start -= mem->num_pages;
> > > > > -       else
> > > > > -               start = 0;
> > > > > -       mem->start = max(mem->start, start);
> > > > > -}
> > > > > -
> > > > >    /**
> > > > >     * amdgpu_vram_mgr_new - allocate new ranges
> > > > >     *
> > > > > @@ -366,13 +357,13 @@ static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
> > > > >                                  const struct ttm_place *place,
> > > > >                                  struct ttm_resource **res)
> > > > >    {
> > > > > -       unsigned long lpfn, num_nodes, pages_per_node, pages_left, pages;
> > > > > +       unsigned long lpfn, pages_per_node, pages_left, pages, n_pages;
> > > > > +       u64 vis_usage = 0, mem_bytes, max_bytes, min_page_size;
> > > > >           struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
> > > > >           struct amdgpu_device *adev = to_amdgpu_device(mgr);
> > > > > -       uint64_t vis_usage = 0, mem_bytes, max_bytes;
> > > > > -       struct ttm_range_mgr_node *node;
> > > > > -       struct drm_mm *mm = &mgr->mm;
> > > > > -       enum drm_mm_insert_mode mode;
> > > > > +       struct amdgpu_vram_mgr_node *node;
> > > > > +       struct drm_buddy *mm = &mgr->mm;
> > > > > +       struct drm_buddy_block *block;
> > > > >           unsigned i;
> > > > >           int r;
> > > > >
> > > > > @@ -391,10 +382,9 @@ static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
> > > > >                   goto error_sub;
> > > > >           }
> > > > >
> > > > > -       if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
> > > > > +       if (place->flags & TTM_PL_FLAG_CONTIGUOUS)
> > > > >                   pages_per_node = ~0ul;
> > > > > -               num_nodes = 1;
> > > > > -       } else {
> > > > > +       else {
> > > > >    #ifdef CONFIG_TRANSPARENT_HUGEPAGE
> > > > >                   pages_per_node = HPAGE_PMD_NR;
> > > > >    #else
> > > > > @@ -403,11 +393,9 @@ static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
> > > > >    #endif
> > > > >                   pages_per_node = max_t(uint32_t, pages_per_node,
> > > > >                                          tbo->page_alignment);
> > > > > -               num_nodes = DIV_ROUND_UP_ULL(PFN_UP(mem_bytes), pages_per_node);
> > > > >           }
> > > > >
> > > > > -       node = kvmalloc(struct_size(node, mm_nodes, num_nodes),
> > > > > -                       GFP_KERNEL | __GFP_ZERO);
> > > > > +       node = kzalloc(sizeof(*node), GFP_KERNEL);
> > > > >           if (!node) {
> > > > >                   r = -ENOMEM;
> > > > >                   goto error_sub;
> > > > > @@ -415,9 +403,17 @@ static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
> > > > >
> > > > >           ttm_resource_init(tbo, place, &node->base);
> > > > >
> > > > > -       mode = DRM_MM_INSERT_BEST;
> > > > > +       INIT_LIST_HEAD(&node->blocks);
> > > > > +
> > > > >           if (place->flags & TTM_PL_FLAG_TOPDOWN)
> > > > > -               mode = DRM_MM_INSERT_HIGH;
> > > > > +               node->flags |= DRM_BUDDY_TOPDOWN_ALLOCATION;
> > > > > +
> > > > > +       if (place->fpfn || lpfn != man->size)
> > > > > +               /* Allocate blocks in desired range */
> > > > > +               node->flags |= DRM_BUDDY_RANGE_ALLOCATION;
> > > > > +
> > > > > +       min_page_size = mgr->default_page_size;
> > > > > +       BUG_ON(min_page_size < mm->chunk_size);
> > > > >
> > > > >           pages_left = node->base.num_pages;
> > > > >
> > > > > @@ -425,36 +421,61 @@ static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
> > > > >           pages = min(pages_left, 2UL << (30 - PAGE_SHIFT));
> > > > >
> > > > >           i = 0;
> > > > > -       spin_lock(&mgr->lock);
> > > > >           while (pages_left) {
> > > > > -               uint32_t alignment = tbo->page_alignment;
> > > > > -
> > > > >                   if (pages >= pages_per_node)
> > > > > -                       alignment = pages_per_node;
> > > > > -
> > > > > -               r = drm_mm_insert_node_in_range(mm, &node->mm_nodes[i], pages,
> > > > > -                                               alignment, 0, place->fpfn,
> > > > > -                                               lpfn, mode);
> > > > > -               if (unlikely(r)) {
> > > > > -                       if (pages > pages_per_node) {
> > > > > -                               if (is_power_of_2(pages))
> > > > > -                                       pages = pages / 2;
> > > > > -                               else
> > > > > -                                       pages = rounddown_pow_of_two(pages);
> > > > > -                               continue;
> > > > > -                       }
> > > > > -                       goto error_free;
> > > > > +                       pages = pages_per_node;
> > > > > +
> > > > > +               n_pages = pages;
> > > > > +
> > > > > +               if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
> > > > > +                       n_pages = roundup_pow_of_two(n_pages);
> > > > > +                       min_page_size = (u64)n_pages << PAGE_SHIFT;
> > > > > +
> > > > > +                       if (n_pages > lpfn)
> > > > > +                               lpfn = n_pages;
> > > > >                   }
> > > > >
> > > > > -               vis_usage += amdgpu_vram_mgr_vis_size(adev, &node->mm_nodes[i]);
> > > > > -               amdgpu_vram_mgr_virt_start(&node->base, &node->mm_nodes[i]);
> > > > > +               mutex_lock(&mgr->lock);
> > > > > +               r = drm_buddy_alloc_blocks(mm, (u64)place->fpfn << PAGE_SHIFT,
> > > > > +                                          (u64)lpfn << PAGE_SHIFT,
> > > > > +                                          (u64)n_pages << PAGE_SHIFT,
> > > > > +                                          min_page_size,
> > > > > +                                          &node->blocks,
> > > > > +                                          node->flags);
> > > > > +               mutex_unlock(&mgr->lock);
> > > > > +               if (unlikely(r))
> > > > > +                       goto error_free_blocks;
> > > > > +
> > > > >                   pages_left -= pages;
> > > > >                   ++i;
> > > > >
> > > > >                   if (pages > pages_left)
> > > > >                           pages = pages_left;
> > > > >           }
> > > > > -       spin_unlock(&mgr->lock);
> > > > > +
> > > > > +       /* Free unused pages for contiguous allocation */
> > > > > +       if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
> > > > > +               u64 actual_size = (u64)node->base.num_pages << PAGE_SHIFT;
> > > > > +
> > > > > +               mutex_lock(&mgr->lock);
> > > > > +               drm_buddy_block_trim(mm,
> > > > > +                                    actual_size,
> > > > > +                                    &node->blocks);
> > > > > +               mutex_unlock(&mgr->lock);
> > > > > +       }
> > > > > +
> > > > > +       list_for_each_entry(block, &node->blocks, link)
> > > > > +               vis_usage += amdgpu_vram_mgr_vis_size(adev, block);
> > > > > +
> > > > > +       block = list_first_entry_or_null(&node->blocks,
> > > > > +                                        struct drm_buddy_block,
> > > > > +                                        link);
> > > > > +       if (!block) {
> > > > > +               r = -ENOENT;
> > > > > +               goto error_free_res;
> > > > > +       }
> > > > > +
> > > > > +       node->base.start = amdgpu_node_start(block) >> PAGE_SHIFT;
> > > > Hmm, does this work? It looks like there are various places checking
> > > > that res->start + res->num_pages <= visible_size, which IIUC should
> > > > only return true when the entire object is placed in the mappable
> > > > portion. i915 is doing something similar. Also it looks like
> > > > ttm_resource_compat() is potentially relying on this, like when moving
> > > > something from non-mappable -> mappable in
> > > > amdgpu_bo_fault_reserve_notify()?
> > > >
> > > > Perhaps something like:
> > > >
> > > > if (vis_usage == num_pages)
> > > >       base.start = 0;
> > > > else
> > > >       base.start = visible_size;
> > > >
> > > > Otherwise I guess just keep amdgpu_vram_mgr_virt_start()?
> > > hmm, I wonder how it works, may be we didn't go through the corner case
> > > where res->start + res->num_pages > visible_size
> > >
> > > in amdgpu if AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED flag is enabled, we
> > > set the ttm_place.lpfn = visible_pfn at
> > > amdgpu_bo_placement_from_domain(). Hence, in amdgpu_vram_mgr_new()
> > > function DRM_BUDDY_RANGE_ALLOCATION flag is enabled, which calls the
> > > alloc_range_bias() in drm_buddy.c.
> > >
> > > Here we get blocks chained together in random order complying
> > > visible_pfn range. say for instance num_pages = 13
> > > we may get,
> > > Block 1 addr - 500 (order-3)
> > > Block 2 addr - 400 (order-2)
> > > Block 3 addr - 600 (order-0)
> > >
> > > I think currently base.start = Block 1 start address fetched from the
> > > list and the address 500 assigned to it, which is good for the resource
> > > access since we access the blocks using the list link
> > >
> > > But for the check res->start + res->num_pages <= visible_size in few
> > > places, this doesn't work. AFAIK, keeping amdgpu_vram_mgr_virt_start()
> > > doesn't work since the function looks for nodes in continuous address to
> > > calculate the start address. AFAIK, assigning the start address (400 +
> > > num_pages <= visible_size) mislead in our case since we use linked list
> > >
> > > how about replacing the check with a bool type return function which
> > > checks the each block start address + block size <= visible_size?
> >
> > Yeah, we already have that in the TTM code. It's just not used
> > everywhere IIRC.
>
> Hi Christian,
> here we have a problem, many places in ttm and amdgpu, we are using the
> tbo->resource->start + bo->resource->num_pages operation, this doesn't
> work in case of drm buddy since it allocates blocks in different
> locations which are chained together using linked list.

AFAICT that was already the case with the existing code, since it looks 
like you can get an array of discontinuous drm_mm blocks. 
amdgpu_vram_mgr_virt_start() looks to be handling that by creating 
fake/virtual res->start offset, which is effectively the maximum end pfn 
over the range of drm_mm blocks(whilst also accounting for num_pages), 
and yes if there are multiple blocks then the res->start might be more 
of a "virtual" offset. AFAIK that scheme should also work as-is with the 
buddy.

> > The node->start can just be set to the invalid offset for now and should
> > be removed as soon as we don't need it any more.
> Assigning the start block offset to resource->start doesn't work,
> If we set node->start to invalid offset, we get an incorrect value?
> > Regards,
> > Christian.