drm/i915: Move system memory to TTM for discrete
Subsystems (2)
cover drm/i915: Move system memory to TTM for discrete
Early implementation of moving system memory for discrete cards over to TTM. We first add the notion of objects being migratable under the object lock to i915 gem, and add some asserts to verify that objects are either locked or pinned when the placement is checked by the gem code.
Patch 2 and 3 deals with updating the i915 gem bookkeeping after a TTM move, Patch 4 moves system over from shmem to TTM for discrete and finally the last patch is more to be considered an RFC for migration implementation.
Much of this code is not testing covered, so we might have to add some selftests as well. Meanwhile this should not be merged but may well be reviewed.
Note that the mock device doesn't consider itself discrete so the TTM system path is not checked by the mock selftests.
Also testing if CI can handle base-commit: and prerequisite-patch-id:
Thomas Hellström (5):
drm/i915: Update object placement flags to be mutable drm/i915/ttm: Adjust gem flags and caching settings after a move drm/i915/ttm: Calculate the object placement at get_pages time drm/i915/ttm: Use TTM for system memory
drm/i915/ttm: Implement object migration
drivers/gpu/drm/i915/gem/i915_gem_internal.c | 4 +- drivers/gpu/drm/i915/gem/i915_gem_mman.c | 7 +- drivers/gpu/drm/i915/gem/i915_gem_object.c | 138 +++++++++ drivers/gpu/drm/i915/gem/i915_gem_object.h | 24 +- .../gpu/drm/i915/gem/i915_gem_object_types.h | 27 +- drivers/gpu/drm/i915/gem/i915_gem_pages.c | 2 +- drivers/gpu/drm/i915/gem/i915_gem_phys.c | 2 +- drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 10 +- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 267 +++++++++++++++--- drivers/gpu/drm/i915/gem/i915_gem_userptr.c | 4 +- .../drm/i915/gem/selftests/huge_gem_object.c | 4 +- .../gpu/drm/i915/gem/selftests/huge_pages.c | 5 +- .../drm/i915/gem/selftests/i915_gem_mman.c | 4 +- .../drm/i915/gem/selftests/i915_gem_phys.c | 3 +- drivers/gpu/drm/i915/i915_drv.h | 3 - drivers/gpu/drm/i915/intel_memory_region.c | 7 +- drivers/gpu/drm/i915/intel_memory_region.h | 8 + drivers/gpu/drm/i915/intel_region_ttm.c | 8 +- drivers/gpu/drm/i915/intel_region_ttm.h | 2 + 19 files changed, 437 insertions(+), 92 deletions(-)
base-commit: cd6eb5f605478f2fff85ec7ac39b7cf445d3deb9
prerequisite-patch-id: b3e7766bc492d04a5ace71fd0d4259689f3bd749
prerequisite-patch-id: 7978430e513e204addd7bc9bef63fd3a9ea25d7c
prerequisite-patch-id: f6aa8d141b22b8e532f8152e53f9c913f98458df
prerequisite-patch-id: 6486c9db619d88a3e9a4ffc01a89bdc722941a67
prerequisite-patch-id: b71878b47e3be71748aacb5d1b3955853bf80de5
prerequisite-patch-id: 1b67c82c32c02dfdb1217f224652276e044fb549
prerequisite-patch-id: 5229db39ec27eb77a8417ed7c9d54af40e0e9f33
prerequisite-patch-id: b66b1893401cc526b4466ea3c427512261d33dfd
prerequisite-patch-id: 706521b6f4858cf4d9ecd83102b347e695b96bb7
prerequisite-patch-id: 686dc7cfea6361c4adabc963996dc3d0d41b28b1
prerequisite-patch-id: 7be87140130a9c2d62a190ceafdacdebd51c5196
Signature · 1 line
2.31.1
PATCH 1/5 drm/i915: Update object placement flags to be mutable
The object ops i915_GEM_OBJECT_HAS_IOMEM and the object I915_BO_ALLOC_STRUCT_PAGE flags are considered immutable by much of our code. Introduce a new mem_flags member to hold these and make sure checks for these flags being set are either done under the object lock or with pages properly pinned. The flags will change during migration under the object lock.
--- drivers/gpu/drm/i915/gem/i915_gem_internal.c | 4 +- drivers/gpu/drm/i915/gem/i915_gem_mman.c | 7 +++- drivers/gpu/drm/i915/gem/i915_gem_object.c | 38 +++++++++++++++++++ drivers/gpu/drm/i915/gem/i915_gem_object.h | 14 ++----- .../gpu/drm/i915/gem/i915_gem_object_types.h | 20 +++++----- drivers/gpu/drm/i915/gem/i915_gem_pages.c | 2 +- drivers/gpu/drm/i915/gem/i915_gem_phys.c | 2 +- drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 10 +++-- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 2 +- drivers/gpu/drm/i915/gem/i915_gem_userptr.c | 4 +- .../drm/i915/gem/selftests/huge_gem_object.c | 4 +- .../gpu/drm/i915/gem/selftests/huge_pages.c | 5 +-- .../drm/i915/gem/selftests/i915_gem_mman.c | 4 +- .../drm/i915/gem/selftests/i915_gem_phys.c | 3 +- 14 files changed, 78 insertions(+), 41 deletions(-)
drivers/gpu/drm/i915/gem/i915_gem_internal.c +2 -2
@@ -177,8 +177,8 @@ i915_gem_object_create_internal(struct drm_i915_private *i915,return ERR_PTR (-ENOMEM );drm_gem_private_object_init (&i915->drm, &obj->base, size); -i915_gem_object_init (obj, &i915_gem_object_internal_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); +i915_gem_object_init (obj, &i915_gem_object_internal_ops, &lock_class,0 ); + obj->mem_flags |=I915_BO_FLAG_STRUCT_PAGE ;/* * Mark the object as volatile, such that the pages are marked as
drivers/gpu/drm/i915/gem/i915_gem_mman.c +6 -1
@@ -683,7 +683,7 @@ __assign_mmap_offset(struct drm_i915_gem_object *obj,if (mmap_type !=I915_MMAP_TYPE_GTT && !i915_gem_object_has_struct_page (obj) && - !i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM )) + !i915_gem_object_has_iomem (obj))return -ENODEV ; mmo =mmap_offset_attach (obj, mmap_type, file);
@@ -707,7 +707,12 @@ __assign_mmap_offset_handle(struct drm_file *file,if (!obj)return -ENOENT ; + err =i915_gem_object_lock_interruptible (obj,NULL ); +if (err) + goto out_put; err =__assign_mmap_offset (obj, mmap_type, offset, file); +i915_gem_object_unlock (obj); +out_put:i915_gem_object_put (obj);return err; }
drivers/gpu/drm/i915/gem/i915_gem_object.c +38 -0
@@ -475,6 +475,44 @@ bool i915_gem_object_migratable(struct drm_i915_gem_object *obj)return obj->mm.n_placements >1 ; } +/** +* i915_gem_object_has_struct_page - Whether the object is page-backed +* @obj: The object to query. +* +* This function should only be called while the object is locked or pinned, +* otherwise the page backing may change under the caller. +* +* Return: True if page-backed, false otherwise. +*/ +bool i915_gem_object_has_struct_page (const struct drm_i915_gem_object *obj) +{ +#ifdef CONFIG_LOCKDEP +if (IS_DGFX (to_i915 (obj->base.dev)) && +i915_gem_object_evictable ((void __force *)obj)) +assert_object_held_shared (obj); +#endif +return obj->mem_flags &I915_BO_FLAG_STRUCT_PAGE ; +} + +/** +* i915_gem_object_has_iomem - Whether the object is iomem-backed +* @obj: The object to query. +* +* This function should only be called while the object is locked or pinned, +* otherwise the iomem backing may change under the caller. +* +* Return: True if iomem-backed, false otherwise. +*/ +bool i915_gem_object_has_iomem (const struct drm_i915_gem_object *obj) +{ +#ifdef CONFIG_LOCKDEP +if (IS_DGFX (to_i915 (obj->base.dev)) && +i915_gem_object_evictable ((void __force *)obj)) +assert_object_held_shared (obj); +#endif +return obj->mem_flags &I915_BO_FLAG_IOMEM ; +} +void i915_gem_init__objects (struct drm_i915_private *i915) {INIT_WORK (&i915->mm.free_work, __i915_gem_free_work);
drivers/gpu/drm/i915/gem/i915_gem_object.h +3 -11
@@ -147,7 +147,7 @@ i915_gem_object_put(struct drm_i915_gem_object *obj)/* * If more than one potential simultaneous locker, assert held. */ -static inline void assert_object_held_shared (struct drm_i915_gem_object *obj) +static inline void assert_object_held_shared (const struct drm_i915_gem_object *obj) {/* * Note mm list lookup is protected by
@@ -261,17 +261,9 @@ i915_gem_object_type_has(const struct drm_i915_gem_object *obj,return obj->ops->flags & flags; } -static inline bool -i915_gem_object_has_struct_page (const struct drm_i915_gem_object *obj) -{ -return obj->flags &I915_BO_ALLOC_STRUCT_PAGE ; -} +bool i915_gem_object_has_struct_page (const struct drm_i915_gem_object *obj); -static inline bool -i915_gem_object_has_iomem (const struct drm_i915_gem_object *obj) -{ -return i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM ); -} +bool i915_gem_object_has_iomem (const struct drm_i915_gem_object *obj);static inline bool i915_gem_object_is_shrinkable (const struct drm_i915_gem_object *obj)
drivers/gpu/drm/i915/gem/i915_gem_object_types.h +10 -10
@@ -33,10 +33,9 @@ struct i915_lut_handle {struct drm_i915_gem_object_ops {unsigned int flags; -#define I915_GEM_OBJECT_HAS_IOMEM BIT(1) -#define I915_GEM_OBJECT_IS_SHRINKABLE BIT(2) -#define I915_GEM_OBJECT_IS_PROXY BIT(3) -#define I915_GEM_OBJECT_NO_MMAP BIT(4) +#define I915_GEM_OBJECT_IS_SHRINKABLE BIT(1) +#define I915_GEM_OBJECT_IS_PROXY BIT(2) +#define I915_GEM_OBJECT_NO_MMAP BIT(3)/* Interface between the GEM object and its backing storage. * get_pages() is called once prior to the use of the associated set
@@ -201,17 +200,18 @@ struct drm_i915_gem_object {unsigned long flags;#define I915_BO_ALLOC_CONTIGUOUS BIT(0)#define I915_BO_ALLOC_VOLATILE BIT(1) -#define I915_BO_ALLOC_STRUCT_PAGE BIT(2) -#define I915_BO_ALLOC_CPU_CLEAR BIT(3) -#define I915_BO_ALLOC_USER BIT(4) +#define I915_BO_ALLOC_CPU_CLEAR BIT(2) +#define I915_BO_ALLOC_USER BIT(3)#define I915_BO_ALLOC_FLAGS (I915_BO_ALLOC_CONTIGUOUS | \ I915_BO_ALLOC_VOLATILE | \ - I915_BO_ALLOC_STRUCT_PAGE | \ I915_BO_ALLOC_CPU_CLEAR | \ I915_BO_ALLOC_USER) -#define I915_BO_READONLY BIT(5) -#define I915_TILING_QUIRK_BIT 6/* unknown swizzling; do not release! */ +#define I915_BO_READONLY BIT(4) +#define I915_TILING_QUIRK_BIT 5/* unknown swizzling; do not release! */ + unsigned int mem_flags:2; +#define I915_BO_FLAG_STRUCT_PAGE BIT(0) +#define I915_BO_FLAG_IOMEM BIT(1)/* * Is the object to be mapped as read-only to the GPU * Only honoured if hardware has relevant pte bit
drivers/gpu/drm/i915/gem/i915_gem_pages.c +1 -1
@@ -351,7 +351,7 @@ void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj,int err;if (!i915_gem_object_has_struct_page (obj) && - !i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM )) + !i915_gem_object_has_iomem (obj))return ERR_PTR (-ENXIO );assert_object_held (obj);
drivers/gpu/drm/i915/gem/i915_gem_phys.c +1 -1
@@ -76,7 +76,7 @@ static int i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj)intel_gt_chipset_flush (&to_i915 (obj->base.dev)->gt);/* We're no longer struct page backed */ - obj->flags &= ~I915_BO_ALLOC_STRUCT_PAGE ; + obj->mem_flags &= ~I915_BO_FLAG_STRUCT_PAGE ;__i915_gem_object_set_pages (obj, st, sg->length);return 0 ;
drivers/gpu/drm/i915/gem/i915_gem_shmem.c +6 -4
@@ -284,6 +284,7 @@ __i915_gem_object_release_shmem(struct drm_i915_gem_object *obj, bool needs_clflush) {GEM_BUG_ON (obj->mm.madv == __I915_MADV_PURGED); +GEM_WARN_ON (IS_DGFX (to_i915 (obj->base.dev)));if (obj->mm.madv ==I915_MADV_DONTNEED ) obj->mm.dirty = false;
@@ -302,6 +303,7 @@ void i915_gem_object_put_pages_shmem(struct drm_i915_gem_object *obj, struct sg_struct pagevec pvec;struct page *page; +GEM_WARN_ON (IS_DGFX (to_i915 (obj->base.dev)));__i915_gem_object_release_shmem (obj, pages, true);i915_gem_gtt_finish_pages (obj, pages);
@@ -444,7 +446,7 @@ shmem_pread(struct drm_i915_gem_object *obj,static void shmem_release (struct drm_i915_gem_object *obj) { -if (obj->flags &I915_BO_ALLOC_STRUCT_PAGE ) +if (i915_gem_object_has_struct_page (obj))i915_gem_object_release_memory_region (obj);fput (obj->base.filp);
@@ -513,9 +515,8 @@ static int shmem_object_init(struct intel_memory_region *mem,mapping_set_gfp_mask (mapping, mask);GEM_BUG_ON (!(mapping_gfp_mask (mapping) & __GFP_RECLAIM)); -i915_gem_object_init (obj, &i915_gem_shmem_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); - +i915_gem_object_init (obj, &i915_gem_shmem_ops, &lock_class,0 ); + obj->mem_flags |=I915_BO_FLAG_STRUCT_PAGE ; obj->write_domain =I915_GEM_DOMAIN_CPU ; obj->read_domains =I915_GEM_DOMAIN_CPU ;
@@ -561,6 +562,7 @@ i915_gem_object_create_shmem_from_data(struct drm_i915_private *dev_priv,resource_size_t offset;int err; +GEM_WARN_ON (IS_DGFX (dev_priv)); obj =i915_gem_object_create_shmem (dev_priv,round_up (size,PAGE_SIZE ));if (IS_ERR (obj))return obj;
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +1 -1
@@ -563,7 +563,6 @@ static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj)const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = { .name ="i915_gem_object_ttm" , - .flags =I915_GEM_OBJECT_HAS_IOMEM , .get_pages = i915_ttm_get_pages, .put_pages = i915_ttm_put_pages,
@@ -620,6 +619,7 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,i915_gem_object_init_memory_region (obj, mem);i915_gem_object_make_unshrinkable (obj); obj->read_domains =I915_GEM_DOMAIN_WC |I915_GEM_DOMAIN_GTT ; + obj->mem_flags |=I915_BO_FLAG_IOMEM ;i915_gem_object_set_cache_coherency (obj,I915_CACHE_NONE );INIT_RADIX_TREE (&obj->ttm.get_io_page.radix,GFP_KERNEL | __GFP_NOWARN);mutex_init (&obj->ttm.get_io_page.lock);
drivers/gpu/drm/i915/gem/i915_gem_userptr.c +2 -2
@@ -538,8 +538,8 @@ i915_gem_userptr_ioctl(struct drm_device *dev,return -ENOMEM ;drm_gem_private_object_init (dev, &obj->base, args->user_size); -i915_gem_object_init (obj, &i915_gem_userptr_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); +i915_gem_object_init (obj, &i915_gem_userptr_ops, &lock_class,0 ); + obj->mem_flags =I915_BO_FLAG_STRUCT_PAGE ; obj->read_domains =I915_GEM_DOMAIN_CPU ; obj->write_domain =I915_GEM_DOMAIN_CPU ;i915_gem_object_set_cache_coherency (obj,I915_CACHE_LLC );
drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c +2 -2
@@ -114,8 +114,8 @@ huge_gem_object(struct drm_i915_private *i915,return ERR_PTR (-ENOMEM );drm_gem_private_object_init (&i915->drm, &obj->base, dma_size); -i915_gem_object_init (obj, &huge_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); +i915_gem_object_init (obj, &huge_ops, &lock_class,0 ); + obj->mem_flags |=I915_BO_FLAG_STRUCT_PAGE ; obj->read_domains =I915_GEM_DOMAIN_CPU ; obj->write_domain =I915_GEM_DOMAIN_CPU ;
drivers/gpu/drm/i915/gem/selftests/huge_pages.c +2 -3
@@ -167,9 +167,8 @@ huge_pages_object(struct drm_i915_private *i915,return ERR_PTR (-ENOMEM );drm_gem_private_object_init (&i915->drm, &obj->base, size); -i915_gem_object_init (obj, &huge_page_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); - +i915_gem_object_init (obj, &huge_page_ops, &lock_class,0 ); + obj->mem_flags |=I915_BO_FLAG_STRUCT_PAGE ;i915_gem_object_set_volatile (obj); obj->write_domain =I915_GEM_DOMAIN_CPU ;
drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c +2 -2
@@ -837,7 +837,7 @@ static bool can_mmap(struct drm_i915_gem_object *obj, enum i915_mmap_type type)if (type !=I915_MMAP_TYPE_GTT && !i915_gem_object_has_struct_page (obj) && - !i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM )) + !i915_gem_object_has_iomem (obj))return false;return true;
@@ -991,7 +991,7 @@ static const char *repr_mmap_type(enum i915_mmap_type type)static bool can_access (const struct drm_i915_gem_object *obj) {return i915_gem_object_has_struct_page (obj) || -i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM ); +i915_gem_object_has_iomem (obj); } static int __igt_mmap_access(struct drm_i915_private *i915,
drivers/gpu/drm/i915/gem/selftests/i915_gem_phys.c +2 -1
@@ -25,13 +25,14 @@ static int mock_phys_object(void *arg) goto out; } +i915_gem_object_lock (obj,NULL );if (!i915_gem_object_has_struct_page (obj)) { +i915_gem_object_unlock (obj); err = -EINVAL ;pr_err ("shmem has no struct page\n" ); goto out_obj; } -i915_gem_object_lock (obj,NULL ); err =i915_gem_object_attach_phys (obj,PAGE_SIZE );i915_gem_object_unlock (obj);if (err) {
Signature · 6 lines
2.31.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH 3/5 drm/i915/ttm: Calculate the object placement at get_pages time
In reply to Thomas HellströmInstead of relying on a static placement, calculate at get_pages() time. This should work for LMEM regions and system for now. For stolen we need to take preallocated range into account. That well be added later.
--- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 92 ++++++++++++++++++------- drivers/gpu/drm/i915/intel_region_ttm.c | 8 ++- drivers/gpu/drm/i915/intel_region_ttm.h | 2 + 3 files changed, 75 insertions(+), 27 deletions(-)
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +67 -25
@@ -24,6 +24,11 @@#define I915_TTM_PRIO_NO_PAGES 1#define I915_TTM_PRIO_HAS_PAGES 2 +/* +* Size of struct ttm_place vector in on-stack struct ttm_placement allocs +*/ +#define I915_TTM_MAX_PLACEMENTS 10 +/** * struct i915_ttm_tt - TTM page vector with additional private information * @ttm: The base TTM page vector.
@@ -42,32 +47,18 @@ struct i915_ttm_tt {struct sg_table *cached_st; }; -static const struct ttm_place lmem0_sys_placement_flags[] = { - { - .fpfn =0 , - .lpfn =0 , - .mem_type =I915_PL_LMEM0 , - .flags =0 , - }, { - .fpfn =0 , - .lpfn =0 , - .mem_type =I915_PL_SYSTEM , - .flags =0 , - } -}; - -static struct ttm_placement i915_lmem0_placement = { - .num_placement =1 , - .placement = &lmem0_sys_placement_flags[0 ], - .num_busy_placement =1 , - .busy_placement = &lmem0_sys_placement_flags[0 ], +static const struct ttm_place sys_placement_flags = { + .fpfn =0 , + .lpfn =0 , + .mem_type =I915_PL_SYSTEM , + .flags =0 , };static struct ttm_placement i915_sys_placement = { .num_placement =1 , - .placement = &lmem0_sys_placement_flags[1 ], + .placement = &sys_placement_flags, .num_busy_placement =1 , - .busy_placement = &lmem0_sys_placement_flags[1 ], + .busy_placement = &sys_placement_flags, };static bool gpu_binds_iomem (struct ttm_resource *mem)
@@ -83,6 +74,55 @@ static bool cpu_maps_iomem(struct ttm_resource *mem)static void i915_ttm_adjust_lru (struct drm_i915_gem_object *obj); +static enum ttm_caching +i915_ttm_select_tt_caching (const struct drm_i915_gem_object *obj) +{ +/* +* Objects only allowed in system get cached cpu-mappings. +* Other objects get WC mapping for now. Even if in system. +*/ +if (obj->mm.region->type ==INTEL_MEMORY_SYSTEM && + obj->mm.n_placements <=1 ) +return ttm_cached; + +return ttm_write_combined; +} + +static void +i915_ttm_place_from_region (const struct intel_memory_region *mr, +struct ttm_place *place) +{ +memset (place,0 ,sizeof (*place)); + place->mem_type =intel_region_to_ttm_type (mr); +} + +static void +i915_ttm_placement_from_obj (const struct drm_i915_gem_object *obj, +struct ttm_place *requested, +struct ttm_place *busy, +struct ttm_placement *placement) +{ +unsigned int i; +unsigned int num_allowed = obj->mm.n_placements; + + placement->num_placement =1 ; +i915_ttm_place_from_region (num_allowed ? obj->mm.placements[0 ] : + obj->mm.region, requested); + +/* Cache this on object? */ + placement->num_busy_placement = num_allowed; +for (i =0 ; i < placement->num_busy_placement; ++i) +i915_ttm_place_from_region (obj->mm.placements[i], busy + i); + +if (num_allowed ==0 ) { + *busy = *requested; + placement->num_busy_placement =1 ; + } + + placement->placement = requested; + placement->busy_placement = busy; +} +static struct ttm_tt *i915_ttm_tt_create (struct ttm_buffer_object *bo,uint32_t page_flags) {
@@ -100,7 +140,8 @@ static struct ttm_tt *i915_ttm_tt_create(struct ttm_buffer_object *bo, man->use_tt) page_flags |=TTM_PAGE_FLAG_ZERO_ALLOC ; - ret =ttm_tt_init (&i915_tt->ttm, bo, page_flags, ttm_write_combined); + ret =ttm_tt_init (&i915_tt->ttm, bo, page_flags, +i915_ttm_select_tt_caching (obj));if (ret) {kfree (i915_tt);return NULL ;
@@ -465,10 +506,13 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) .no_wait_gpu = false, };struct sg_table *st; +struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS ]; +struct ttm_placement placement;int ret;/* Move to the requested placement. */ - ret =ttm_bo_validate (bo, &i915_lmem0_placement, &ctx); +i915_ttm_placement_from_obj (obj, &requested, busy, &placement); + ret =ttm_bo_validate (bo, &placement, &ctx);if (ret)return ret == -ENOSPC ? -ENXIO : ret;
@@ -684,7 +728,6 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,i915_gem_object_make_unshrinkable (obj);INIT_RADIX_TREE (&obj->ttm.get_io_page.radix,GFP_KERNEL | __GFP_NOWARN);mutex_init (&obj->ttm.get_io_page.lock); - bo_type = (obj->flags &I915_BO_ALLOC_USER ) ? ttm_bo_type_device : ttm_bo_type_kernel;
@@ -708,7 +751,6 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,i915_ttm_adjust_domains_after_cpu_move (obj);i915_ttm_adjust_gem_after_move (obj);i915_gem_object_unlock (obj); - out:/* i915 wants -ENXIO when out of memory region space. */ return (ret == -ENOSPC ) ? -ENXIO : ret;
drivers/gpu/drm/i915/intel_region_ttm.c +6 -2
@@ -49,12 +49,16 @@ void intel_region_ttm_device_fini(struct drm_i915_private *dev_priv)* driver-private types for now, reserving TTM_PL_VRAM for stolen * memory and TTM_PL_TT for GGTT use if decided to implement this. */ -static int intel_region_to_ttm_type (struct intel_memory_region *mem) +int intel_region_to_ttm_type (const struct intel_memory_region *mem) {int type;GEM_BUG_ON (mem->type !=INTEL_MEMORY_LOCAL && - mem->type !=INTEL_MEMORY_MOCK ); + mem->type !=INTEL_MEMORY_MOCK && + mem->type !=INTEL_MEMORY_SYSTEM ); + +if (mem->type ==INTEL_MEMORY_SYSTEM ) +return TTM_PL_SYSTEM ; type = mem->instance +TTM_PL_PRIV ;GEM_BUG_ON (type >=TTM_NUM_MEM_TYPES );
drivers/gpu/drm/i915/intel_region_ttm.h +2 -0
@@ -27,6 +27,8 @@ struct sg_table *intel_region_ttm_node_to_st(struct intel_memory_region *mem,void intel_region_ttm_node_free (struct intel_memory_region *mem,void *node); +int intel_region_to_ttm_type (const struct intel_memory_region *mem); +struct ttm_device_funcs *i915_ttm_driver (void );#ifdef CONFIG_DRM_I915_SELFTEST
Signature · 6 lines
2.31.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH 4/5 drm/i915/ttm: Use TTM for system memory
In reply to Thomas HellströmFor discrete, use TTM for both cached and WC system memory. That means we currently rely on the TTM memory accounting / shrinker. For cached system memory we should consider remaining shmem-backed, which can be implemented from our ttm_tt_populate calback. We can then also reuse our own very elaborate shrinker for that memory.
--- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 22 ++++++++++++++++++++++ drivers/gpu/drm/i915/i915_drv.h | 3 --- drivers/gpu/drm/i915/intel_memory_region.c | 7 ++++++- drivers/gpu/drm/i915/intel_memory_region.h | 8 ++++++++ 4 files changed, 36 insertions(+), 4 deletions(-)
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +22 -0
@@ -755,3 +755,25 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,/* i915 wants -ENXIO when out of memory region space. */ return (ret == -ENOSPC ) ? -ENXIO : ret; } + +static const struct intel_memory_region_ops ttm_system_region_ops = { + .init_object = __i915_gem_ttm_object_init, +}; + +struct intel_memory_region * +i915_gem_ttm_system_setup (struct drm_i915_private *i915, +u16 type,u16 instance) +{ +struct intel_memory_region *mr; + + mr =intel_memory_region_create (i915,0 , +totalram_pages () <<PAGE_SHIFT , +PAGE_SIZE ,0 , + type, instance, + &ttm_system_region_ops); +if (IS_ERR_OR_NULL (mr)) +return mr; + +intel_memory_region_set_name (mr,"system-ttm" ); +return mr; +}
drivers/gpu/drm/i915/i915_drv.h +0 -3
@@ -1768,9 +1768,6 @@ void i915_gem_cleanup_userptr(struct drm_i915_private *dev_priv);void i915_gem_init_early (struct drm_i915_private *dev_priv);void i915_gem_cleanup_early (struct drm_i915_private *dev_priv); -struct intel_memory_region *i915_gem_shmem_setup (struct drm_i915_private *i915, -u16 type,u16 instance); -static inline void i915_gem_drain_freed_objects (struct drm_i915_private *i915) {/*
drivers/gpu/drm/i915/intel_memory_region.c +6 -1
@@ -220,7 +220,12 @@ int intel_memory_regions_hw_probe(struct drm_i915_private *i915) instance = intel_region_map[i].instance;switch (type) {case INTEL_MEMORY_SYSTEM : - mem =i915_gem_shmem_setup (i915, type, instance); +if (IS_DGFX (i915)) + mem =i915_gem_ttm_system_setup (i915, type, + instance); +else + mem =i915_gem_shmem_setup (i915, type, + instance);break ;case INTEL_MEMORY_STOLEN_LOCAL : mem =i915_gem_stolen_lmem_setup (i915, type, instance);
drivers/gpu/drm/i915/intel_memory_region.h +8 -0
@@ -142,4 +142,12 @@ void intel_memory_region_unreserve(struct intel_memory_region *mem);int intel_memory_region_reserve (struct intel_memory_region *mem,resource_size_t offset,resource_size_t size); + +struct intel_memory_region * +i915_gem_ttm_system_setup (struct drm_i915_private *i915, +u16 type,u16 instance); +struct intel_memory_region * +i915_gem_shmem_setup (struct drm_i915_private *i915, +u16 type,u16 instance); +#endif
Signature · 6 lines
2.31.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH 2/5 drm/i915/ttm: Adjust gem flags and caching settings after a move
In reply to Thomas HellströmAfter a TTM move we need to update the i915 gem flags and caching settings to reflect the new placement.
Also introduce gpu_binds_iomem() and cpu_maps_iomem() to clean up the various ways we previously used to detect this.
Finally, initialize the TTM object reserved to be able to update flags and caching before anyone else gets hold of the object.
--- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 112 +++++++++++++++++++----- 1 file changed, 90 insertions(+), 22 deletions(-)
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +90 -22
@@ -70,6 +70,17 @@ static struct ttm_placement i915_sys_placement = { .busy_placement = &lmem0_sys_placement_flags[1], }; +static bool gpu_binds_iomem (struct ttm_resource *mem) +{ +return (mem->mem_type !=TTM_PL_SYSTEM ); +} + +static bool cpu_maps_iomem (struct ttm_resource *mem) +{ +/* Once / if we support GGTT, this is also false for cached ttm_tts */ +return (mem->mem_type !=TTM_PL_SYSTEM ); +} +static void i915_ttm_adjust_lru (struct drm_i915_gem_object *obj); static struct ttm_tt *i915_ttm_tt_create(struct ttm_buffer_object *bo,
@@ -175,6 +186,41 @@ static void i915_ttm_free_cached_io_st(struct drm_i915_gem_object *obj) obj->ttm.cached_io_st =NULL ; } +static void +i915_ttm_adjust_domains_after_cpu_move (struct drm_i915_gem_object *obj) +{ +struct ttm_buffer_object *bo =i915_gem_to_ttm (obj); + +if (cpu_maps_iomem (&bo->mem) || bo->ttm->caching != ttm_cached) { + obj->write_domain =I915_GEM_DOMAIN_WC ; + obj->read_domains =I915_GEM_DOMAIN_WC ; + }else { + obj->write_domain =I915_GEM_DOMAIN_CPU ; + obj->read_domains =I915_GEM_DOMAIN_CPU ; + } +} + +static void i915_ttm_adjust_gem_after_move (struct drm_i915_gem_object *obj) +{ +struct drm_i915_private *i915 =to_i915 (obj->base.dev); +struct ttm_buffer_object *bo =i915_gem_to_ttm (obj); +unsigned int cache_level; + + obj->mem_flags &= ~(I915_BO_FLAG_STRUCT_PAGE |I915_BO_FLAG_IOMEM ); + + obj->mem_flags |=cpu_maps_iomem (&bo->mem) ?I915_BO_FLAG_IOMEM : +I915_BO_FLAG_STRUCT_PAGE ; + +if ((HAS_LLC (i915) ||HAS_SNOOP (i915)) && !gpu_binds_iomem (&bo->mem) && + bo->ttm->caching == ttm_cached) { + cache_level =I915_CACHE_LLC ; + }else { + cache_level =I915_CACHE_NONE ; + } + +i915_gem_object_set_cache_coherency (obj, cache_level); +} +static void i915_ttm_purge (struct drm_i915_gem_object *obj) {struct ttm_buffer_object *bo =i915_gem_to_ttm (obj);
@@ -190,8 +236,10 @@ static void i915_ttm_purge(struct drm_i915_gem_object *obj)/* TTM's purge interface. Note that we might be reentering. */ ret =ttm_bo_validate (bo, &place, &ctx); -if (!ret) { + obj->write_domain =0 ; + obj->read_domains =0 ; +i915_ttm_adjust_gem_after_move (obj);i915_ttm_free_cached_io_st (obj); obj->mm.madv = __I915_MADV_PURGED; }
@@ -273,12 +321,15 @@ i915_ttm_resource_get_st(struct drm_i915_gem_object *obj, struct ttm_resource *res) {struct ttm_buffer_object *bo =i915_gem_to_ttm (obj); -struct ttm_resource_manager *man = -ttm_manager_type (bo->bdev, res->mem_type); -if (man->use_tt) +if (!gpu_binds_iomem (res))return i915_ttm_tt_get_st (bo->ttm); +/* +* If CPU mapping differs, we need to add the ttm_tt pages to +* the resulting st. Might make sense for GGTT. +*/ +GEM_WARN_ON (!cpu_maps_iomem (res));return intel_region_ttm_node_to_st (obj->mm.region, res->mm_node); }
@@ -290,8 +341,6 @@ static int i915_ttm_move(struct ttm_buffer_object *bo, bool evict,struct drm_i915_gem_object *obj =i915_ttm_to_gem (bo);struct ttm_resource_manager *dst_man =ttm_manager_type (bo->bdev, dst_mem->mem_type); -struct ttm_resource_manager *src_man = -ttm_manager_type (bo->bdev, bo->mem.mem_type);struct intel_memory_region *dst_reg, *src_reg;union {struct ttm_kmap_iter_tt tt;
@@ -332,34 +381,36 @@ static int i915_ttm_move(struct ttm_buffer_object *bo, bool evict,if (IS_ERR (dst_st))return PTR_ERR (dst_st); -/* If we start mapping GGTT, we can no longer use man::use_tt here. */ - dst_iter = dst_man->use_tt ? + dst_iter = !cpu_maps_iomem (dst_mem) ?ttm_kmap_iter_tt_init (&_dst_iter.tt, bo->ttm) :ttm_kmap_iter_iomap_init (&_dst_iter.io, &dst_reg->iomap, dst_st, dst_reg->region.start); - src_iter = src_man->use_tt ? + src_iter = !cpu_maps_iomem (&bo->mem) ?ttm_kmap_iter_tt_init (&_src_iter.tt, bo->ttm) :ttm_kmap_iter_iomap_init (&_src_iter.io, &src_reg->iomap, obj->ttm.cached_io_st, src_reg->region.start);ttm_move_memcpy (bo, dst_mem->num_pages, dst_iter, src_iter); +/* Below dst_mem becomes bo->mem. */ ttm_bo_move_sync_cleanup (bo, dst_mem); +i915_ttm_adjust_domains_after_cpu_move (obj);i915_ttm_free_cached_io_st (obj); -if (!dst_man->use_tt) { +if (gpu_binds_iomem (dst_mem) ||cpu_maps_iomem (dst_mem)) { obj->ttm.cached_io_st = dst_st; obj->ttm.get_io_page.sg_pos = dst_st->sgl; obj->ttm.get_io_page.sg_idx =0 ; } +i915_ttm_adjust_gem_after_move (obj);return 0 ; }static int i915_ttm_io_mem_reserve (struct ttm_device *bdev,struct ttm_resource *mem) { -if (mem->mem_type <I915_PL_LMEM0 ) +if (!cpu_maps_iomem (mem))return 0 ; mem->bus.caching = ttm_write_combined;
@@ -421,6 +472,16 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj)if (ret)return ret == -ENOSPC ? -ENXIO : ret; +i915_ttm_adjust_lru (obj); +if (bo->ttm && !ttm_tt_is_populated (bo->ttm)) { + ret =ttm_tt_populate (bo->bdev, bo->ttm, &ctx); +if (ret) +return ret; + +i915_ttm_adjust_domains_after_cpu_move (obj); +i915_ttm_adjust_gem_after_move (obj); + } +/* Object either has a page vector or is an iomem object */ st = bo->ttm ?i915_ttm_tt_get_st (bo->ttm) : obj->ttm.cached_io_st;if (IS_ERR (st))
@@ -428,8 +489,6 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj)__i915_gem_object_set_pages (obj, st,i915_sg_dma_sizes (st->sgl)); -i915_ttm_adjust_lru (obj); -return ret; }
@@ -563,6 +622,7 @@ static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj)const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = { .name ="i915_gem_object_ttm" , + .flags =I915_GEM_OBJECT_IS_SHRINKABLE , .get_pages = i915_ttm_get_pages, .put_pages = i915_ttm_put_pages,
@@ -599,6 +659,10 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem, {static struct lock_class_key lock_class;struct drm_i915_private *i915 = mem->i915; +struct ttm_operation_ctx ctx = { + .interruptible = true, + .no_wait_gpu = false, + };enum ttm_bo_type bo_type;size_t alignment =0 ;int ret;
@@ -618,15 +682,14 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,i915_gem_object_init (obj, &i915_gem_ttm_obj_ops, &lock_class, flags);i915_gem_object_init_memory_region (obj, mem);i915_gem_object_make_unshrinkable (obj); - obj->read_domains =I915_GEM_DOMAIN_WC |I915_GEM_DOMAIN_GTT ; - obj->mem_flags |=I915_BO_FLAG_IOMEM ; -i915_gem_object_set_cache_coherency (obj,I915_CACHE_NONE );INIT_RADIX_TREE (&obj->ttm.get_io_page.radix,GFP_KERNEL | __GFP_NOWARN);mutex_init (&obj->ttm.get_io_page.lock); bo_type = (obj->flags &I915_BO_ALLOC_USER ) ? ttm_bo_type_device : ttm_bo_type_kernel; + obj->base.vma_node.driver_private =i915_gem_to_ttm (obj); +/* * If this function fails, it will call the destructor, but * our caller still owns the object. So no freeing in the
@@ -634,14 +697,19 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,* Similarly, in delayed_destroy, we can't call ttm_bo_put() * until successful initialization. */ - obj->base.vma_node.driver_private =i915_gem_to_ttm (obj); - ret =ttm_bo_init (&i915->bdev,i915_gem_to_ttm (obj), size, - bo_type, &i915_sys_placement, alignment, - true,NULL ,NULL , i915_ttm_bo_destroy); + ret =ttm_bo_init_reserved (&i915->bdev,i915_gem_to_ttm (obj), size, + bo_type, &i915_sys_placement, alignment, + &ctx,NULL ,NULL , i915_ttm_bo_destroy); + +if (ret) + goto out; -if (!ret) - obj->ttm.created = true; + obj->ttm.created = true; +i915_ttm_adjust_domains_after_cpu_move (obj); +i915_ttm_adjust_gem_after_move (obj); +i915_gem_object_unlock (obj); +out:/* i915 wants -ENXIO when out of memory region space. */ return (ret == -ENOSPC ) ? -ENXIO : ret; }
Signature · 1 line
2.31.1
PATCH 5/5 drm/i915/ttm: Implement object migration
In reply to Thomas HellströmImplement object migration, needed primarily for dma-buf exports of objects currently residing in LMEM, until we land p2pdma. There are no users yet of this code.
--- drivers/gpu/drm/i915/gem/i915_gem_object.c | 100 ++++++++++++++++++ drivers/gpu/drm/i915/gem/i915_gem_object.h | 10 ++ .../gpu/drm/i915/gem/i915_gem_object_types.h | 7 ++ drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 51 ++++++++- 4 files changed, 164 insertions(+), 4 deletions(-)
drivers/gpu/drm/i915/gem/i915_gem_object.c +100 -0
@@ -475,6 +475,106 @@ bool i915_gem_object_migratable(struct drm_i915_gem_object *obj)return obj->mm.n_placements >1 ; } +bool i915_gem_object_can_migrate_to_region (struct drm_i915_gem_object *obj, +struct intel_memory_region *mr, +unsigned int *placement_index) +{ +unsigned int i; +unsigned int num_allowed = obj->mm.n_placements; + +if (!i915_gem_object_evictable (obj)) +return false; + +if (num_allowed ==0 && mr != obj->mm.region) +return false; + +if (num_allowed ==1 && mr != obj->mm.placements[0 ]) +return false; + +for (i =0 ; i < num_allowed; ++i) { +if (mr == obj->mm.placements[i]) { +if (placement_index) + *placement_index = i; +return true; + } + } + +return false; +} + +/** +* i915_gem_object_migrate_to_region_lazy - Lazily migrate an object +* @obj: The object to migrate. +* @mr: The region to migrate to. +* +* Check that @obj can migrate to @mr, and update all data necessary to +* make that happen on the next get_pages(). We sync and unbind gpu bindings +* and put pages. The word "lazy" means that the actual migration blit +* is not triggered by this function. +* +* Return: Zero on success, negative error code on failure. +*/ +int i915_gem_object_migrate_to_region_lazy (struct drm_i915_gem_object *obj, +struct intel_memory_region *mr) +{ +unsigned int index; +int ret; + +if (obj->mm.region == mr) +return 0 ; + +if (!i915_gem_object_can_migrate_to_region (obj, mr, &index)) +return -EINVAL ; + + ret =i915_gem_object_unbind (obj,I915_GEM_OBJECT_UNBIND_ACTIVE ); +if (ret) +return ret; + + ret =__i915_gem_object_put_pages (obj); +if (ret) +return ret; + +/* +* The next get_pages() will pick up the new desired placement +* and migrate. +*/ +if (obj->mm.override_region) { +intel_memory_region_put (obj->mm.override_region); + obj->mm.override_region =NULL ; + } + +if (index !=0 ) + obj->mm.override_region = +intel_memory_region_get (obj->mm.placements[index]); + +return 0 ; +} + +/** +* i915_gem_object_migrate_to_region - Migrate an object +* @obj: The object to migrate. +* @mr: The region to migrate to. +* +* Check that @obj can migrate to @mr, and migrate the object. +* The caller needs to check that the final region was the +* desired one since the object may have ended up elsewhere on +* lack of space in the desired region, and if there are other +* allowed placements. +* +* Return: Zero on success, negative error code on failure. +*/ +int i915_gem_object_migrate_to_region (struct drm_i915_gem_object *obj, +struct intel_memory_region *mr) +{ +int ret; + + ret =i915_gem_object_migrate_to_region_lazy (obj, mr); +if (ret) +return ret; + +return ____i915_gem_object_get_pages (obj); +} +/** * i915_gem_object_has_struct_page - Whether the object is page-backed * @obj: The object to query.
drivers/gpu/drm/i915/gem/i915_gem_object.h +10 -0
@@ -592,6 +592,16 @@ bool i915_gem_object_migratable(struct drm_i915_gem_object *obj);bool i915_gem_object_validates_to_lmem (struct drm_i915_gem_object *obj); +bool i915_gem_object_can_migrate_to_region (struct drm_i915_gem_object *obj, +struct intel_memory_region *mr, +unsigned int *placement_index); + +int i915_gem_object_migrate_to_region_lazy (struct drm_i915_gem_object *obj, +struct intel_memory_region *mr); + +int i915_gem_object_migrate_to_region (struct drm_i915_gem_object *obj, +struct intel_memory_region *mr); +#ifdef CONFIG_MMU_NOTIFIER static inline bool i915_gem_object_is_userptr (struct drm_i915_gem_object *obj)
drivers/gpu/drm/i915/gem/i915_gem_object_types.h +7 -0
@@ -264,6 +264,13 @@ struct drm_i915_gem_object {*/ struct intel_memory_region *region; +/** +* Override memory region for this object. Use to +* override the order of the placement list to migrate +* an object to the desired region. +*/ +struct intel_memory_region *override_region; +/** * Memory manager node allocated for this object. */
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +47 -4
72 lines not on this page · raw message
Signature · 1 line
2.31.1
Patchwork
CI[Intel-gfx] ✗ Fi.CI.BUILD: failure for drm/i915: Move system memory to TTM for discrete
In reply to Thomas Hellström== Series Details ==
== Series Details ==
Series: drm/i915: Move system memory to TTM for discrete
URL : https://patchwork.freedesktop.org/series/90898/
State : failure
== Summary ==
Applying: drm/i915: Update object placement flags to be mutable
Using index info to reconstruct a base tree...
M drivers/gpu/drm/i915/gem/i915_gem_mman.c
M drivers/gpu/drm/i915/gem/i915_gem_object.c
M drivers/gpu/drm/i915/gem/i915_gem_object.h
M drivers/gpu/drm/i915/gem/i915_gem_object_types.h
M drivers/gpu/drm/i915/gem/i915_gem_pages.c
A drivers/gpu/drm/i915/gem/i915_gem_ttm.c
M drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c
Falling back to patching base and 3-way merge...
Auto-merging drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c
CONFLICT (modify/delete): drivers/gpu/drm/i915/gem/i915_gem_ttm.c deleted in HEAD and modified in drm/i915: Update object placement flags to be mutable. Version drm/i915: Update object placement flags to be mutable of drivers/gpu/drm/i915/gem/i915_gem_ttm.c left in tree. Auto-merging drivers/gpu/drm/i915/gem/i915_gem_pages.c
Auto-merging drivers/gpu/drm/i915/gem/i915_gem_object_types.h
CONFLICT (content): Merge conflict in drivers/gpu/drm/i915/gem/i915_gem_object_types.h
Auto-merging drivers/gpu/drm/i915/gem/i915_gem_object.h
Auto-merging drivers/gpu/drm/i915/gem/i915_gem_object.c
CONFLICT (content): Merge conflict in drivers/gpu/drm/i915/gem/i915_gem_object.c
Auto-merging drivers/gpu/drm/i915/gem/i915_gem_mman.c
CONFLICT (content): Merge conflict in drivers/gpu/drm/i915/gem/i915_gem_mman.c
error: Failed to merge in the changes.
hint: Use 'git am --show-current-patch=diff' to see the failed patch
Patch failed at 0001 drm/i915: Update object placement flags to be mutable
When you have resolved this problem, run "git am --continue".
If you prefer to skip this patch, run "git am --skip" instead.
To restore the original branch and stop patching, run "git am --abort".
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH 4/5] drm/i915/ttm: Use TTM for system memory
In reply to Thomas HellströmQuoted 30 lines · Thomas Hellström
For discrete, use TTM for both cached and WC system memory. That means we currently rely on the TTM memory accounting / shrinker. For cached system memory we should consider remaining shmem-backed, which can be implemented from our
ttm_tt_populatecalback. We can then also reuse our own very elaborate shrinker for that memory.Signed-off-by Thomas Hellström <[email protected]>--- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 22 ++++++++++++++++++++++ drivers/gpu/drm/i915/i915_drv.h | 3 --- drivers/gpu/drm/i915/intel_memory_region.c | 7 ++++++- drivers/gpu/drm/i915/intel_memory_region.h | 8 ++++++++ 4 files changed, 36 insertions(+), 4 deletions(-)drivers/gpu/drm/i915/gem/i915_gem_ttm.c +7
-0@@ -755,3 +755,25 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,/* i915 wants -ENXIO when out of memory region space. */ return (ret == -ENOSPC ) ? -ENXIO : ret; } + +static const struct intel_memory_region_ops ttm_system_region_ops = { + .init_object = __i915_gem_ttm_object_init, +}; + +struct intel_memory_region * +i915_gem_ttm_system_setup(struct drm_i915_private *i915,
drivers/gpu/drm/i915/gem/i915_gem_ttm.c int __i915_gem_ttm_object_init(struct intel_memory_region *mem, · line 774
Quoted 10 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +10
-0+ u16 type, u16 instance) +{ +struct intel_memory_region *mr; + + mr =intel_memory_region_create (i915,0 , +totalram_pages () <<PAGE_SHIFT , +PAGE_SIZE ,0 , + type, instance, + &ttm_system_region_ops); +if (IS_ERR_OR_NULL (mr))
region_create can't return NULL.
Quoted 57 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +5
-0+return mr; + +intel_memory_region_set_name (mr,"system-ttm" ); +return mr; +}drivers/gpu/drm/i915/i915_drv.h +0
-3@@ -1768,9 +1768,6 @@ void i915_gem_cleanup_userptr(struct drm_i915_private *dev_priv);void i915_gem_init_early (struct drm_i915_private *dev_priv);void i915_gem_cleanup_early (struct drm_i915_private *dev_priv); -struct intel_memory_region *i915_gem_shmem_setup (struct drm_i915_private *i915, -u16 type,u16 instance); -static inline void i915_gem_drain_freed_objects (struct drm_i915_private *i915) {/* drivers/gpu/drm/i915/intel_memory_region.c +6
-1@@ -220,7 +220,12 @@ int intel_memory_regions_hw_probe(struct drm_i915_private *i915) instance = intel_region_map[i].instance;switch (type) {case INTEL_MEMORY_SYSTEM : - mem =i915_gem_shmem_setup (i915, type, instance); +if (IS_DGFX (i915)) + mem =i915_gem_ttm_system_setup (i915, type, + instance); +else + mem =i915_gem_shmem_setup (i915, type, + instance);break ;case INTEL_MEMORY_STOLEN_LOCAL : mem =i915_gem_stolen_lmem_setup (i915, type, instance);drivers/gpu/drm/i915/intel_memory_region.h +8
-1@@ -142,4 +142,12 @@ void intel_memory_region_unreserve(struct intel_memory_region *mem);int intel_memory_region_reserve (struct intel_memory_region *mem,resource_size_t offset,resource_size_t size); + +struct intel_memory_region * +i915_gem_ttm_system_setup (struct drm_i915_private *i915, +u16 type,u16 instance); +struct intel_memory_region * +i915_gem_shmem_setup (struct drm_i915_private *i915, +u16 type,u16 instance); +#endif --2.31.1
Quoted 4 lines
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH 2/5] drm/i915/ttm: Adjust gem flags and caching settings after a move
In reply to Thomas HellströmQuoted 18 lines · Thomas Hellström
After a TTM move we need to update the i915 gem flags and caching settings to reflect the new placement.
Also introducegpu_binds_iomem()andcpu_maps_iomem()to clean up the various ways we previously used to detect this.
Finally, initialize the TTM object reserved to be able to update flags and caching before anyone else gets hold of the object.Signed-off-by Thomas Hellström <[email protected]>--- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 112 +++++++++++++++++++----- 1 file changed, 90 insertions(+), 22 deletions(-)drivers/gpu/drm/i915/gem/i915_gem_ttm.c +0
-0
drivers/gpu/drm/i915/gem/i915_gem_ttm.c static struct ttm_placement i915_sys_placement = { · line 73
Quoted 4 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +1
-0@@ -70,6 +70,17 @@ static struct ttm_placement i915_sys_placement = { .busy_placement = &lmem0_sys_placement_flags[1], }; +static bool gpu_binds_iomem (struct ttm_resource *mem)
What does gpu_binds mean here? It's mapped though the GTT?
drivers/gpu/drm/i915/gem/i915_gem_ttm.c static struct ttm_placement i915_sys_placement = { · line 82
Quoted 9 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +9
-0+{ +return (mem->mem_type !=TTM_PL_SYSTEM ); +} + +static bool cpu_maps_iomem (struct ttm_resource *mem) +{ +/* Once / if we support GGTT, this is also false for cached ttm_tts */ +return (mem->mem_type !=TTM_PL_SYSTEM ); +}
Can drop the extra brackets.
Quoted 24 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +17
-0+static void i915_ttm_adjust_lru (struct drm_i915_gem_object *obj); static struct ttm_tt *i915_ttm_tt_create(struct ttm_buffer_object *bo,@@ -175,6 +186,41 @@ static void i915_ttm_free_cached_io_st(struct drm_i915_gem_object *obj) obj->ttm.cached_io_st =NULL ; } +static void +i915_ttm_adjust_domains_after_cpu_move (struct drm_i915_gem_object *obj) +{ +struct ttm_buffer_object *bo =i915_gem_to_ttm (obj); + +if (cpu_maps_iomem (&bo->mem) || bo->ttm->caching != ttm_cached) { + obj->write_domain =I915_GEM_DOMAIN_WC ; + obj->read_domains =I915_GEM_DOMAIN_WC ; + }else { + obj->write_domain =I915_GEM_DOMAIN_CPU ; + obj->read_domains =I915_GEM_DOMAIN_CPU ; + } +} + +static void i915_ttm_adjust_gem_after_move (struct drm_i915_gem_object *obj) +{
drivers/gpu/drm/i915/gem/i915_gem_ttm.c static void i915_ttm_free_cached_io_st(struct drm_i915_gem_object *obj) · line 214
Quoted 10 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +10
-0+struct drm_i915_private *i915 =to_i915 (obj->base.dev); +struct ttm_buffer_object *bo =i915_gem_to_ttm (obj); +unsigned int cache_level; + + obj->mem_flags &= ~(I915_BO_FLAG_STRUCT_PAGE |I915_BO_FLAG_IOMEM ); + + obj->mem_flags |=cpu_maps_iomem (&bo->mem) ?I915_BO_FLAG_IOMEM : +I915_BO_FLAG_STRUCT_PAGE ; + + if ((HAS_LLC(i915) || HAS_SNOOP(i915)) && !gpu_binds_iomem(&bo->mem) &&
I think all modern hw has support for snooping or the fast shared LLC. Are we expecting to hit this path for non-dgpu? Also by default we will choose ttm_cached at least for platforms like dg1? Also pin_map() is still a separate interface at this point?
Quoted 186 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +53
-23+ bo->ttm->caching == ttm_cached) { + cache_level =I915_CACHE_LLC ; + }else { + cache_level =I915_CACHE_NONE ; + } + +i915_gem_object_set_cache_coherency (obj, cache_level); +} +static void i915_ttm_purge (struct drm_i915_gem_object *obj) {struct ttm_buffer_object *bo =i915_gem_to_ttm (obj);@@ -190,8 +236,10 @@ static void i915_ttm_purge(struct drm_i915_gem_object *obj)/* TTM's purge interface. Note that we might be reentering. */ ret =ttm_bo_validate (bo, &place, &ctx); -if (!ret) { + obj->write_domain =0 ; + obj->read_domains =0 ; +i915_ttm_adjust_gem_after_move (obj);i915_ttm_free_cached_io_st (obj); obj->mm.madv = __I915_MADV_PURGED; }@@ -273,12 +321,15 @@ i915_ttm_resource_get_st(struct drm_i915_gem_object *obj, struct ttm_resource *res) {struct ttm_buffer_object *bo =i915_gem_to_ttm (obj); -struct ttm_resource_manager *man = -ttm_manager_type (bo->bdev, res->mem_type); -if (man->use_tt) +if (!gpu_binds_iomem (res))return i915_ttm_tt_get_st (bo->ttm); +/* +* If CPU mapping differs, we need to add the ttm_tt pages to +* the resulting st. Might make sense for GGTT. +*/ +GEM_WARN_ON (!cpu_maps_iomem (res));return intel_region_ttm_node_to_st (obj->mm.region, res->mm_node); }@@ -290,8 +341,6 @@ static int i915_ttm_move(struct ttm_buffer_object *bo, bool evict,struct drm_i915_gem_object *obj =i915_ttm_to_gem (bo);struct ttm_resource_manager *dst_man =ttm_manager_type (bo->bdev, dst_mem->mem_type); -struct ttm_resource_manager *src_man = -ttm_manager_type (bo->bdev, bo->mem.mem_type);struct intel_memory_region *dst_reg, *src_reg;union {struct ttm_kmap_iter_tt tt;@@ -332,34 +381,36 @@ static int i915_ttm_move(struct ttm_buffer_object *bo, bool evict,if (IS_ERR (dst_st))return PTR_ERR (dst_st); -/* If we start mapping GGTT, we can no longer use man::use_tt here. */ - dst_iter = dst_man->use_tt ? + dst_iter = !cpu_maps_iomem (dst_mem) ?ttm_kmap_iter_tt_init (&_dst_iter.tt, bo->ttm) :ttm_kmap_iter_iomap_init (&_dst_iter.io, &dst_reg->iomap, dst_st, dst_reg->region.start); - src_iter = src_man->use_tt ? + src_iter = !cpu_maps_iomem (&bo->mem) ?ttm_kmap_iter_tt_init (&_src_iter.tt, bo->ttm) :ttm_kmap_iter_iomap_init (&_src_iter.io, &src_reg->iomap, obj->ttm.cached_io_st, src_reg->region.start);ttm_move_memcpy (bo, dst_mem->num_pages, dst_iter, src_iter); +/* Below dst_mem becomes bo->mem. */ ttm_bo_move_sync_cleanup (bo, dst_mem); +i915_ttm_adjust_domains_after_cpu_move (obj);i915_ttm_free_cached_io_st (obj); -if (!dst_man->use_tt) { +if (gpu_binds_iomem (dst_mem) ||cpu_maps_iomem (dst_mem)) { obj->ttm.cached_io_st = dst_st; obj->ttm.get_io_page.sg_pos = dst_st->sgl; obj->ttm.get_io_page.sg_idx =0 ; } +i915_ttm_adjust_gem_after_move (obj);return 0 ; }static int i915_ttm_io_mem_reserve (struct ttm_device *bdev,struct ttm_resource *mem) { -if (mem->mem_type <I915_PL_LMEM0 ) +if (!cpu_maps_iomem (mem))return 0 ; mem->bus.caching = ttm_write_combined;@@ -421,6 +472,16 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj)if (ret)return ret == -ENOSPC ? -ENXIO : ret; +i915_ttm_adjust_lru (obj); +if (bo->ttm && !ttm_tt_is_populated (bo->ttm)) { + ret =ttm_tt_populate (bo->bdev, bo->ttm, &ctx); +if (ret) +return ret; + +i915_ttm_adjust_domains_after_cpu_move (obj); +i915_ttm_adjust_gem_after_move (obj); + } +/* Object either has a page vector or is an iomem object */ st = bo->ttm ?i915_ttm_tt_get_st (bo->ttm) : obj->ttm.cached_io_st;if (IS_ERR (st))@@ -428,8 +489,6 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj)__i915_gem_object_set_pages (obj, st,i915_sg_dma_sizes (st->sgl)); -i915_ttm_adjust_lru (obj); -return ret; }@@ -563,6 +622,7 @@ static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj)const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = { .name ="i915_gem_object_ttm" , + .flags =I915_GEM_OBJECT_IS_SHRINKABLE , .get_pages = i915_ttm_get_pages, .put_pages = i915_ttm_put_pages,@@ -599,6 +659,10 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem, {static struct lock_class_key lock_class;struct drm_i915_private *i915 = mem->i915; +struct ttm_operation_ctx ctx = { + .interruptible = true, + .no_wait_gpu = false, + };enum ttm_bo_type bo_type;size_t alignment =0 ;int ret;@@ -618,15 +682,14 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,i915_gem_object_init (obj, &i915_gem_ttm_obj_ops, &lock_class, flags);i915_gem_object_init_memory_region (obj, mem);i915_gem_object_make_unshrinkable (obj); - obj->read_domains =I915_GEM_DOMAIN_WC |I915_GEM_DOMAIN_GTT ; - obj->mem_flags |=I915_BO_FLAG_IOMEM ; -i915_gem_object_set_cache_coherency (obj,I915_CACHE_NONE );INIT_RADIX_TREE (&obj->ttm.get_io_page.radix,GFP_KERNEL | __GFP_NOWARN);mutex_init (&obj->ttm.get_io_page.lock); bo_type = (obj->flags &I915_BO_ALLOC_USER ) ? ttm_bo_type_device : ttm_bo_type_kernel; + obj->base.vma_node.driver_private =i915_gem_to_ttm (obj); +/* * If this function fails, it will call the destructor, but * our caller still owns the object. So no freeing in the @@ -634,14 +697,19 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,* Similarly, in delayed_destroy, we can't call ttm_bo_put() * until successful initialization. */ - obj->base.vma_node.driver_private =i915_gem_to_ttm (obj); - ret =ttm_bo_init (&i915->bdev,i915_gem_to_ttm (obj), size, - bo_type, &i915_sys_placement, alignment, - true,NULL ,NULL , i915_ttm_bo_destroy); + ret =ttm_bo_init_reserved (&i915->bdev,i915_gem_to_ttm (obj), size, + bo_type, &i915_sys_placement, alignment, + &ctx,NULL ,NULL , i915_ttm_bo_destroy); + +if (ret) + goto out; -if (!ret) - obj->ttm.created = true; + obj->ttm.created = true; +i915_ttm_adjust_domains_after_cpu_move (obj); +i915_ttm_adjust_gem_after_move (obj); +i915_gem_object_unlock (obj); +out:/* i915 wants -ENXIO when out of memory region space. */ return (ret == -ENOSPC ) ? -ENXIO : ret; } --2.31.1
Quoted 4 lines
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH 3/5] drm/i915/ttm: Calculate the object placement at get_pages time
In reply to Thomas HellströmQuoted 6 lines · Thomas Hellström
Instead of relying on a static placement, calculate at
get_pages()time. This should work for LMEM regions and system for now. For stolen we need to take preallocated range into account. That well be added later.Signed-off-by Thomas Hellström <[email protected]>
Can we split this patch out and merge it? We can use this for feeding in the per BO flags.
Quoted 86 lines
--- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 92 ++++++++++++++++++------- drivers/gpu/drm/i915/intel_region_ttm.c | 8 ++- drivers/gpu/drm/i915/intel_region_ttm.h | 2 + 3 files changed, 75 insertions(+), 27 deletions(-)drivers/gpu/drm/i915/gem/i915_gem_ttm.c +32
-21@@ -24,6 +24,11 @@#define I915_TTM_PRIO_NO_PAGES 1#define I915_TTM_PRIO_HAS_PAGES 2 +/* +* Size of struct ttm_place vector in on-stack struct ttm_placement allocs +*/ +#define I915_TTM_MAX_PLACEMENTS 10 +/** * struct i915_ttm_tt - TTM page vector with additional private information * @ttm: The base TTM page vector. @@ -42,32 +47,18 @@ struct i915_ttm_tt {struct sg_table *cached_st; }; -static const struct ttm_place lmem0_sys_placement_flags[] = { - { - .fpfn =0 , - .lpfn =0 , - .mem_type =I915_PL_LMEM0 , - .flags =0 , - }, { - .fpfn =0 , - .lpfn =0 , - .mem_type =I915_PL_SYSTEM , - .flags =0 , - } -}; - -static struct ttm_placement i915_lmem0_placement = { - .num_placement =1 , - .placement = &lmem0_sys_placement_flags[0 ], - .num_busy_placement =1 , - .busy_placement = &lmem0_sys_placement_flags[0 ], +static const struct ttm_place sys_placement_flags = { + .fpfn =0 , + .lpfn =0 , + .mem_type =I915_PL_SYSTEM , + .flags =0 , };static struct ttm_placement i915_sys_placement = { .num_placement =1 , - .placement = &lmem0_sys_placement_flags[1 ], + .placement = &sys_placement_flags, .num_busy_placement =1 , - .busy_placement = &lmem0_sys_placement_flags[1 ], + .busy_placement = &sys_placement_flags, };static bool gpu_binds_iomem (struct ttm_resource *mem)@@ -83,6 +74,55 @@ static bool cpu_maps_iomem(struct ttm_resource *mem)static void i915_ttm_adjust_lru (struct drm_i915_gem_object *obj); +static enum ttm_caching +i915_ttm_select_tt_caching (const struct drm_i915_gem_object *obj) +{ +/* +* Objects only allowed in system get cached cpu-mappings. +* Other objects get WC mapping for now. Even if in system. +*/ +if (obj->mm.region->type ==INTEL_MEMORY_SYSTEM && + obj->mm.n_placements <=1 ) +return ttm_cached; + +return ttm_write_combined; +} + +static void +i915_ttm_place_from_region (const struct intel_memory_region *mr, +struct ttm_place *place) +{ +memset (place,0 ,sizeof (*place)); + place->mem_type =intel_region_to_ttm_type (mr);
drivers/gpu/drm/i915/gem/i915_gem_ttm.c static bool cpu_maps_iomem(struct ttm_resource *mem) · line 106
Quoted 10 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +10
-0+} + +static void +i915_ttm_placement_from_obj (const struct drm_i915_gem_object *obj, +struct ttm_place *requested, +struct ttm_place *busy, +struct ttm_placement *placement) +{ +unsigned int i; +unsigned int num_allowed = obj->mm.n_placements;
Style nit: Christmas tree
Quoted 106 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +25
-4+ + placement->num_placement =1 ; +i915_ttm_place_from_region (num_allowed ? obj->mm.placements[0 ] : + obj->mm.region, requested); + +/* Cache this on object? */ + placement->num_busy_placement = num_allowed; +for (i =0 ; i < placement->num_busy_placement; ++i) +i915_ttm_place_from_region (obj->mm.placements[i], busy + i); + +if (num_allowed ==0 ) { + *busy = *requested; + placement->num_busy_placement =1 ; + } + + placement->placement = requested; + placement->busy_placement = busy; +} +static struct ttm_tt *i915_ttm_tt_create (struct ttm_buffer_object *bo,uint32_t page_flags) {@@ -100,7 +140,8 @@ static struct ttm_tt *i915_ttm_tt_create(struct ttm_buffer_object *bo, man->use_tt) page_flags |=TTM_PAGE_FLAG_ZERO_ALLOC ; - ret =ttm_tt_init (&i915_tt->ttm, bo, page_flags, ttm_write_combined); + ret =ttm_tt_init (&i915_tt->ttm, bo, page_flags, +i915_ttm_select_tt_caching (obj));if (ret) {kfree (i915_tt);return NULL ;@@ -465,10 +506,13 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) .no_wait_gpu = false, };struct sg_table *st; +struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS ]; +struct ttm_placement placement;int ret;/* Move to the requested placement. */ - ret =ttm_bo_validate (bo, &i915_lmem0_placement, &ctx); +i915_ttm_placement_from_obj (obj, &requested, busy, &placement); + ret =ttm_bo_validate (bo, &placement, &ctx);if (ret)return ret == -ENOSPC ? -ENXIO : ret;@@ -684,7 +728,6 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,i915_gem_object_make_unshrinkable (obj);INIT_RADIX_TREE (&obj->ttm.get_io_page.radix,GFP_KERNEL | __GFP_NOWARN);mutex_init (&obj->ttm.get_io_page.lock); - bo_type = (obj->flags &I915_BO_ALLOC_USER ) ? ttm_bo_type_device : ttm_bo_type_kernel;@@ -708,7 +751,6 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,i915_ttm_adjust_domains_after_cpu_move (obj);i915_ttm_adjust_gem_after_move (obj);i915_gem_object_unlock (obj); - out:/* i915 wants -ENXIO when out of memory region space. */ return (ret == -ENOSPC ) ? -ENXIO : ret;drivers/gpu/drm/i915/intel_region_ttm.c +6
-2@@ -49,12 +49,16 @@ void intel_region_ttm_device_fini(struct drm_i915_private *dev_priv)* driver-private types for now, reserving TTM_PL_VRAM for stolen * memory and TTM_PL_TT for GGTT use if decided to implement this. */ -static int intel_region_to_ttm_type (struct intel_memory_region *mem) +int intel_region_to_ttm_type (const struct intel_memory_region *mem) {int type;GEM_BUG_ON (mem->type !=INTEL_MEMORY_LOCAL && - mem->type !=INTEL_MEMORY_MOCK ); + mem->type !=INTEL_MEMORY_MOCK && + mem->type !=INTEL_MEMORY_SYSTEM ); + +if (mem->type ==INTEL_MEMORY_SYSTEM ) +return TTM_PL_SYSTEM ; type = mem->instance +TTM_PL_PRIV ;GEM_BUG_ON (type >=TTM_NUM_MEM_TYPES );drivers/gpu/drm/i915/intel_region_ttm.h +2
-1@@ -27,6 +27,8 @@ struct sg_table *intel_region_ttm_node_to_st(struct intel_memory_region *mem,void intel_region_ttm_node_free (struct intel_memory_region *mem,void *node); +int intel_region_to_ttm_type (const struct intel_memory_region *mem); +struct ttm_device_funcs *i915_ttm_driver (void );#ifdef CONFIG_DRM_I915_SELFTEST --2.31.1
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH 1/5] drm/i915: Update object placement flags to be mutable
In reply to Thomas HellströmQuoted 7 lines · Thomas Hellström
The object ops
i915_GEM_OBJECT_HAS_IOMEMand the objectI915_BO_ALLOC_STRUCT_PAGEflags are considered immutable by much of our code. Introduce a newmem_flagsmember to hold these and make sure checks for these flags being set are either done under the object lock or with pages properly pinned. The flags will change during migration under the object lock.
What are the rules for the gem_object_ops flags? Like is_shrinkable? Can't we just move these there(IOMEM vs PAGE)?
Quoted 371 lines
Signed-off-by Thomas Hellström <[email protected]>--- drivers/gpu/drm/i915/gem/i915_gem_internal.c | 4 +- drivers/gpu/drm/i915/gem/i915_gem_mman.c | 7 +++- drivers/gpu/drm/i915/gem/i915_gem_object.c | 38 +++++++++++++++++++ drivers/gpu/drm/i915/gem/i915_gem_object.h | 14 ++----- .../gpu/drm/i915/gem/i915_gem_object_types.h | 20 +++++----- drivers/gpu/drm/i915/gem/i915_gem_pages.c | 2 +- drivers/gpu/drm/i915/gem/i915_gem_phys.c | 2 +- drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 10 +++-- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 2 +- drivers/gpu/drm/i915/gem/i915_gem_userptr.c | 4 +- .../drm/i915/gem/selftests/huge_gem_object.c | 4 +- .../gpu/drm/i915/gem/selftests/huge_pages.c | 5 +-- .../drm/i915/gem/selftests/i915_gem_mman.c | 4 +- .../drm/i915/gem/selftests/i915_gem_phys.c | 3 +- 14 files changed, 78 insertions(+), 41 deletions(-)drivers/gpu/drm/i915/gem/i915_gem_internal.c +2
-2@@ -177,8 +177,8 @@ i915_gem_object_create_internal(struct drm_i915_private *i915,return ERR_PTR (-ENOMEM );drm_gem_private_object_init (&i915->drm, &obj->base, size); -i915_gem_object_init (obj, &i915_gem_object_internal_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); +i915_gem_object_init (obj, &i915_gem_object_internal_ops, &lock_class,0 ); + obj->mem_flags |=I915_BO_FLAG_STRUCT_PAGE ;/* * Mark the object as volatile, such that the pages are marked as drivers/gpu/drm/i915/gem/i915_gem_mman.c +6
-1@@ -683,7 +683,7 @@ __assign_mmap_offset(struct drm_i915_gem_object *obj,if (mmap_type !=I915_MMAP_TYPE_GTT && !i915_gem_object_has_struct_page (obj) && - !i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM )) + !i915_gem_object_has_iomem (obj))return -ENODEV ; mmo =mmap_offset_attach (obj, mmap_type, file);@@ -707,7 +707,12 @@ __assign_mmap_offset_handle(struct drm_file *file,if (!obj)return -ENOENT ; + err =i915_gem_object_lock_interruptible (obj,NULL ); +if (err) + goto out_put; err =__assign_mmap_offset (obj, mmap_type, offset, file); +i915_gem_object_unlock (obj); +out_put:i915_gem_object_put (obj);return err; }drivers/gpu/drm/i915/gem/i915_gem_object.c +38
-0@@ -475,6 +475,44 @@ bool i915_gem_object_migratable(struct drm_i915_gem_object *obj)return obj->mm.n_placements >1 ; } +/** +* i915_gem_object_has_struct_page - Whether the object is page-backed +* @obj: The object to query. +* +* This function should only be called while the object is locked or pinned, +* otherwise the page backing may change under the caller. +* +* Return: True if page-backed, false otherwise. +*/ +bool i915_gem_object_has_struct_page (const struct drm_i915_gem_object *obj) +{ +#ifdef CONFIG_LOCKDEP +if (IS_DGFX (to_i915 (obj->base.dev)) && +i915_gem_object_evictable ((void __force *)obj)) +assert_object_held_shared (obj); +#endif +return obj->mem_flags &I915_BO_FLAG_STRUCT_PAGE ; +} + +/** +* i915_gem_object_has_iomem - Whether the object is iomem-backed +* @obj: The object to query. +* +* This function should only be called while the object is locked or pinned, +* otherwise the iomem backing may change under the caller. +* +* Return: True if iomem-backed, false otherwise. +*/ +bool i915_gem_object_has_iomem (const struct drm_i915_gem_object *obj) +{ +#ifdef CONFIG_LOCKDEP +if (IS_DGFX (to_i915 (obj->base.dev)) && +i915_gem_object_evictable ((void __force *)obj)) +assert_object_held_shared (obj); +#endif +return obj->mem_flags &I915_BO_FLAG_IOMEM ; +} +void i915_gem_init__objects (struct drm_i915_private *i915) {INIT_WORK (&i915->mm.free_work, __i915_gem_free_work);drivers/gpu/drm/i915/gem/i915_gem_object.h +3
-11@@ -147,7 +147,7 @@ i915_gem_object_put(struct drm_i915_gem_object *obj)/* * If more than one potential simultaneous locker, assert held. */ -static inline void assert_object_held_shared (struct drm_i915_gem_object *obj) +static inline void assert_object_held_shared (const struct drm_i915_gem_object *obj) {/* * Note mm list lookup is protected by @@ -261,17 +261,9 @@ i915_gem_object_type_has(const struct drm_i915_gem_object *obj,return obj->ops->flags & flags; } -static inline bool -i915_gem_object_has_struct_page (const struct drm_i915_gem_object *obj) -{ -return obj->flags &I915_BO_ALLOC_STRUCT_PAGE ; -} +bool i915_gem_object_has_struct_page (const struct drm_i915_gem_object *obj); -static inline bool -i915_gem_object_has_iomem (const struct drm_i915_gem_object *obj) -{ -return i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM ); -} +bool i915_gem_object_has_iomem (const struct drm_i915_gem_object *obj);static inline bool i915_gem_object_is_shrinkable (const struct drm_i915_gem_object *obj)drivers/gpu/drm/i915/gem/i915_gem_object_types.h +10
-10@@ -33,10 +33,9 @@ struct i915_lut_handle {struct drm_i915_gem_object_ops {unsigned int flags; -#define I915_GEM_OBJECT_HAS_IOMEM BIT(1) -#define I915_GEM_OBJECT_IS_SHRINKABLE BIT(2) -#define I915_GEM_OBJECT_IS_PROXY BIT(3) -#define I915_GEM_OBJECT_NO_MMAP BIT(4) +#define I915_GEM_OBJECT_IS_SHRINKABLE BIT(1) +#define I915_GEM_OBJECT_IS_PROXY BIT(2) +#define I915_GEM_OBJECT_NO_MMAP BIT(3)/* Interface between the GEM object and its backing storage. * get_pages() is called once prior to the use of the associated set @@ -201,17 +200,18 @@ struct drm_i915_gem_object {unsigned long flags;#define I915_BO_ALLOC_CONTIGUOUS BIT(0)#define I915_BO_ALLOC_VOLATILE BIT(1) -#define I915_BO_ALLOC_STRUCT_PAGE BIT(2) -#define I915_BO_ALLOC_CPU_CLEAR BIT(3) -#define I915_BO_ALLOC_USER BIT(4) +#define I915_BO_ALLOC_CPU_CLEAR BIT(2) +#define I915_BO_ALLOC_USER BIT(3)#define I915_BO_ALLOC_FLAGS (I915_BO_ALLOC_CONTIGUOUS | \ I915_BO_ALLOC_VOLATILE | \ - I915_BO_ALLOC_STRUCT_PAGE | \ I915_BO_ALLOC_CPU_CLEAR | \ I915_BO_ALLOC_USER) -#define I915_BO_READONLY BIT(5) -#define I915_TILING_QUIRK_BIT 6/* unknown swizzling; do not release! */ +#define I915_BO_READONLY BIT(4) +#define I915_TILING_QUIRK_BIT 5/* unknown swizzling; do not release! */ + unsigned int mem_flags:2; +#define I915_BO_FLAG_STRUCT_PAGE BIT(0) +#define I915_BO_FLAG_IOMEM BIT(1)/* * Is the object to be mapped as read-only to the GPU * Only honoured if hardware has relevant pte bit drivers/gpu/drm/i915/gem/i915_gem_pages.c +1
-1@@ -351,7 +351,7 @@ void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj,int err;if (!i915_gem_object_has_struct_page (obj) && - !i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM )) + !i915_gem_object_has_iomem (obj))return ERR_PTR (-ENXIO );assert_object_held (obj);drivers/gpu/drm/i915/gem/i915_gem_phys.c +1
-1@@ -76,7 +76,7 @@ static int i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj)intel_gt_chipset_flush (&to_i915 (obj->base.dev)->gt);/* We're no longer struct page backed */ - obj->flags &= ~I915_BO_ALLOC_STRUCT_PAGE ; + obj->mem_flags &= ~I915_BO_FLAG_STRUCT_PAGE ;__i915_gem_object_set_pages (obj, st, sg->length);return 0 ;drivers/gpu/drm/i915/gem/i915_gem_shmem.c +6
-4@@ -284,6 +284,7 @@ __i915_gem_object_release_shmem(struct drm_i915_gem_object *obj, bool needs_clflush) {GEM_BUG_ON (obj->mm.madv == __I915_MADV_PURGED); +GEM_WARN_ON (IS_DGFX (to_i915 (obj->base.dev)));if (obj->mm.madv ==I915_MADV_DONTNEED ) obj->mm.dirty = false;@@ -302,6 +303,7 @@ void i915_gem_object_put_pages_shmem(struct drm_i915_gem_object *obj, struct sg_struct pagevec pvec;struct page *page; +GEM_WARN_ON (IS_DGFX (to_i915 (obj->base.dev)));__i915_gem_object_release_shmem (obj, pages, true);i915_gem_gtt_finish_pages (obj, pages);@@ -444,7 +446,7 @@ shmem_pread(struct drm_i915_gem_object *obj,static void shmem_release (struct drm_i915_gem_object *obj) { -if (obj->flags &I915_BO_ALLOC_STRUCT_PAGE ) +if (i915_gem_object_has_struct_page (obj))i915_gem_object_release_memory_region (obj);fput (obj->base.filp);@@ -513,9 +515,8 @@ static int shmem_object_init(struct intel_memory_region *mem,mapping_set_gfp_mask (mapping, mask);GEM_BUG_ON (!(mapping_gfp_mask (mapping) & __GFP_RECLAIM)); -i915_gem_object_init (obj, &i915_gem_shmem_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); - +i915_gem_object_init (obj, &i915_gem_shmem_ops, &lock_class,0 ); + obj->mem_flags |=I915_BO_FLAG_STRUCT_PAGE ; obj->write_domain =I915_GEM_DOMAIN_CPU ; obj->read_domains =I915_GEM_DOMAIN_CPU ;@@ -561,6 +562,7 @@ i915_gem_object_create_shmem_from_data(struct drm_i915_private *dev_priv,resource_size_t offset;int err; +GEM_WARN_ON (IS_DGFX (dev_priv)); obj =i915_gem_object_create_shmem (dev_priv,round_up (size,PAGE_SIZE ));if (IS_ERR (obj))return obj;drivers/gpu/drm/i915/gem/i915_gem_ttm.c +1
-1@@ -563,7 +563,6 @@ static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj)const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = { .name ="i915_gem_object_ttm" , - .flags =I915_GEM_OBJECT_HAS_IOMEM , .get_pages = i915_ttm_get_pages, .put_pages = i915_ttm_put_pages,@@ -620,6 +619,7 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,i915_gem_object_init_memory_region (obj, mem);i915_gem_object_make_unshrinkable (obj); obj->read_domains =I915_GEM_DOMAIN_WC |I915_GEM_DOMAIN_GTT ; + obj->mem_flags |=I915_BO_FLAG_IOMEM ;i915_gem_object_set_cache_coherency (obj,I915_CACHE_NONE );INIT_RADIX_TREE (&obj->ttm.get_io_page.radix,GFP_KERNEL | __GFP_NOWARN);mutex_init (&obj->ttm.get_io_page.lock);drivers/gpu/drm/i915/gem/i915_gem_userptr.c +2
-2@@ -538,8 +538,8 @@ i915_gem_userptr_ioctl(struct drm_device *dev,return -ENOMEM ;drm_gem_private_object_init (dev, &obj->base, args->user_size); -i915_gem_object_init (obj, &i915_gem_userptr_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); +i915_gem_object_init (obj, &i915_gem_userptr_ops, &lock_class,0 ); + obj->mem_flags =I915_BO_FLAG_STRUCT_PAGE ; obj->read_domains =I915_GEM_DOMAIN_CPU ; obj->write_domain =I915_GEM_DOMAIN_CPU ;i915_gem_object_set_cache_coherency (obj,I915_CACHE_LLC );drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c +2
-2@@ -114,8 +114,8 @@ huge_gem_object(struct drm_i915_private *i915,return ERR_PTR (-ENOMEM );drm_gem_private_object_init (&i915->drm, &obj->base, dma_size); -i915_gem_object_init (obj, &huge_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); +i915_gem_object_init (obj, &huge_ops, &lock_class,0 ); + obj->mem_flags |=I915_BO_FLAG_STRUCT_PAGE ; obj->read_domains =I915_GEM_DOMAIN_CPU ; obj->write_domain =I915_GEM_DOMAIN_CPU ;drivers/gpu/drm/i915/gem/selftests/huge_pages.c +2
-3@@ -167,9 +167,8 @@ huge_pages_object(struct drm_i915_private *i915,return ERR_PTR (-ENOMEM );drm_gem_private_object_init (&i915->drm, &obj->base, size); -i915_gem_object_init (obj, &huge_page_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); - +i915_gem_object_init (obj, &huge_page_ops, &lock_class,0 ); + obj->mem_flags |=I915_BO_FLAG_STRUCT_PAGE ;i915_gem_object_set_volatile (obj); obj->write_domain =I915_GEM_DOMAIN_CPU ;drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c +2
-2@@ -837,7 +837,7 @@ static bool can_mmap(struct drm_i915_gem_object *obj, enum i915_mmap_type type)if (type !=I915_MMAP_TYPE_GTT && !i915_gem_object_has_struct_page (obj) && - !i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM )) + !i915_gem_object_has_iomem (obj))return false;return true;@@ -991,7 +991,7 @@ static const char *repr_mmap_type(enum i915_mmap_type type)static bool can_access (const struct drm_i915_gem_object *obj) {return i915_gem_object_has_struct_page (obj) || -i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM ); +i915_gem_object_has_iomem (obj); } static int __igt_mmap_access(struct drm_i915_private *i915,drivers/gpu/drm/i915/gem/selftests/i915_gem_phys.c +2
-2@@ -25,13 +25,14 @@ static int mock_phys_object(void *arg) goto out; } +i915_gem_object_lock (obj,NULL );if (!i915_gem_object_has_struct_page (obj)) { +i915_gem_object_unlock (obj); err = -EINVAL ;pr_err ("shmem has no struct page\n" ); goto out_obj; } -i915_gem_object_lock (obj,NULL ); err =i915_gem_object_attach_phys (obj,PAGE_SIZE );i915_gem_object_unlock (obj);if (err) { --2.31.1
Quoted 4 lines
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH 2/5] drm/i915/ttm: Adjust gem flags and caching settings after a move
In reply to Matthew AuldThanks for reviewing, Matthew.
Quoted 22 lines · Matthew Auld
Quoted 20 lines · Thomas Hellström
After a TTM move we need to update the i915 gem flags and caching settings to reflect the new placement.
Also introducegpu_binds_iomem()andcpu_maps_iomem()to clean up the various ways we previously used to detect this.
Finally, initialize the TTM object reserved to be able to update flags and caching before anyone else gets hold of the object.Signed-off-by Thomas Hellström <[email protected]>--- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 112 +++++++++++++++++++----- 1 file changed, 90 insertions(+), 22 deletions(-)drivers/gpu/drm/i915/gem/i915_gem_ttm.c +0
-0@@ -70,6 +70,17 @@ static struct ttm_placement i915_sys_placement = { .busy_placement = &lmem0_sys_placement_flags[1], };
Quoted 2 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c static struct ttm_placement i915_sys_placement = { · line 73
Quoted 1 line
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +1
-0+static bool gpu_binds_iomem (struct ttm_resource *mem)What does
gpu_bindsmean here? It's mapped though the GTT?
gpu_bind_XX would here translate to "make xx visible to the gpu", We could of course use the term "gpu_maps" if you think that fits better.
Quoted 40 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c static struct ttm_placement i915_sys_placement = { · line 82
Quoted 10 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +9
-0+{ +return (mem->mem_type !=TTM_PL_SYSTEM ); +} + +static bool cpu_maps_iomem (struct ttm_resource *mem) +{ +/* Once / if we support GGTT, this is also false for cached ttm_tts */ +return (mem->mem_type !=TTM_PL_SYSTEM ); +}Can drop the extra brackets.
Quoted 28 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +21
-0+static void i915_ttm_adjust_lru (struct drm_i915_gem_object *obj); static struct ttm_tt *i915_ttm_tt_create(struct ttm_buffer_object *bo,@@ -175,6 +186,41 @@ static void i915_ttm_free_cached_io_st(struct drm_i915_gem_object *obj) obj->ttm.cached_io_st =NULL ; } +static void +i915_ttm_adjust_domains_after_cpu_move (struct drm_i915_gem_object *obj) +{ +struct ttm_buffer_object *bo =i915_gem_to_ttm (obj); + +if (cpu_maps_iomem (&bo->mem) || bo->ttm->caching != ttm_cached) { + obj->write_domain =I915_GEM_DOMAIN_WC ; + obj->read_domains =I915_GEM_DOMAIN_WC ; + }else { + obj->write_domain =I915_GEM_DOMAIN_CPU ; + obj->read_domains =I915_GEM_DOMAIN_CPU ; + } +} + +static void i915_ttm_adjust_gem_after_move (struct drm_i915_gem_object *obj) +{ +struct drm_i915_private *i915 =to_i915 (obj->base.dev); +struct ttm_buffer_object *bo =i915_gem_to_ttm (obj); +unsigned int cache_level; +
Quoted 8 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c static void i915_ttm_free_cached_io_st(struct drm_i915_gem_object *obj) · line 214
Quoted 6 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +6
-0+ obj->mem_flags &= ~(I915_BO_FLAG_STRUCT_PAGE |I915_BO_FLAG_IOMEM ); + + obj->mem_flags |=cpu_maps_iomem (&bo->mem) ?I915_BO_FLAG_IOMEM : +I915_BO_FLAG_STRUCT_PAGE ; + + if ((HAS_LLC(i915) || HAS_SNOOP(i915)) && !gpu_binds_iomem(&bo->mem) &&I think all modern hw has support for snooping or the fast shared LLC. Are we expecting to hit this path for non-dgpu?
Not initially, no. Possibly later.
Quoted 2 lines
Also by default we
will choosettm_cachedat least for platforms like dg1?
Not for evicted LMEM. They still have ttm_write_combine. User-space can still access those through mmap() while evicted.
Quoted 2 lines
Also
pin_map()
is still a separate interface at this point?
Yes, I had a comment on that in your caching patch that I reviewed, we need to make sure even for pin_map that we only map using the single supported caching attributes for the BO.
Quoted 186 lines
Quoted 186 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +53
-23+ bo->ttm->caching == ttm_cached) { + cache_level =I915_CACHE_LLC ; + }else { + cache_level =I915_CACHE_NONE ; + } + +i915_gem_object_set_cache_coherency (obj, cache_level); +} +static void i915_ttm_purge (struct drm_i915_gem_object *obj) {struct ttm_buffer_object *bo =i915_gem_to_ttm (obj);@@ -190,8 +236,10 @@ static void i915_ttm_purge(struct drm_i915_gem_object *obj)/* TTM's purge interface. Note that we might be reentering. */ ret =ttm_bo_validate (bo, &place, &ctx); -if (!ret) { + obj->write_domain =0 ; + obj->read_domains =0 ; +i915_ttm_adjust_gem_after_move (obj);i915_ttm_free_cached_io_st (obj); obj->mm.madv = __I915_MADV_PURGED; }@@ -273,12 +321,15 @@ i915_ttm_resource_get_st(struct drm_i915_gem_object *obj, struct ttm_resource *res) {struct ttm_buffer_object *bo =i915_gem_to_ttm (obj); -struct ttm_resource_manager *man = -ttm_manager_type (bo->bdev, res->mem_type); -if (man->use_tt) +if (!gpu_binds_iomem (res))return i915_ttm_tt_get_st (bo->ttm); +/* +* If CPU mapping differs, we need to add the ttm_tt pages to +* the resulting st. Might make sense for GGTT. +*/ +GEM_WARN_ON (!cpu_maps_iomem (res));return intel_region_ttm_node_to_st (obj->mm.region, res->mm_node); }@@ -290,8 +341,6 @@ static int i915_ttm_move(struct ttm_buffer_object *bo, bool evict,struct drm_i915_gem_object *obj =i915_ttm_to_gem (bo);struct ttm_resource_manager *dst_man =ttm_manager_type (bo->bdev, dst_mem->mem_type); -struct ttm_resource_manager *src_man = -ttm_manager_type (bo->bdev, bo->mem.mem_type);struct intel_memory_region *dst_reg, *src_reg;union {struct ttm_kmap_iter_tt tt;@@ -332,34 +381,36 @@ static int i915_ttm_move(struct ttm_buffer_object *bo, bool evict,if (IS_ERR (dst_st))return PTR_ERR (dst_st); -/* If we start mapping GGTT, we can no longer use man::use_tt here. */ - dst_iter = dst_man->use_tt ? + dst_iter = !cpu_maps_iomem (dst_mem) ?ttm_kmap_iter_tt_init (&_dst_iter.tt, bo->ttm) :ttm_kmap_iter_iomap_init (&_dst_iter.io, &dst_reg->iomap, dst_st, dst_reg->region.start); - src_iter = src_man->use_tt ? + src_iter = !cpu_maps_iomem (&bo->mem) ?ttm_kmap_iter_tt_init (&_src_iter.tt, bo->ttm) :ttm_kmap_iter_iomap_init (&_src_iter.io, &src_reg->iomap, obj->ttm.cached_io_st, src_reg->region.start);ttm_move_memcpy (bo, dst_mem->num_pages, dst_iter, src_iter); +/* Below dst_mem becomes bo->mem. */ ttm_bo_move_sync_cleanup (bo, dst_mem); +i915_ttm_adjust_domains_after_cpu_move (obj);i915_ttm_free_cached_io_st (obj); -if (!dst_man->use_tt) { +if (gpu_binds_iomem (dst_mem) ||cpu_maps_iomem (dst_mem)) { obj->ttm.cached_io_st = dst_st; obj->ttm.get_io_page.sg_pos = dst_st->sgl; obj->ttm.get_io_page.sg_idx =0 ; } +i915_ttm_adjust_gem_after_move (obj);return 0 ; }static int i915_ttm_io_mem_reserve (struct ttm_device *bdev,struct ttm_resource *mem) { -if (mem->mem_type <I915_PL_LMEM0 ) +if (!cpu_maps_iomem (mem))return 0 ; mem->bus.caching = ttm_write_combined;@@ -421,6 +472,16 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj)if (ret)return ret == -ENOSPC ? -ENXIO : ret; +i915_ttm_adjust_lru (obj); +if (bo->ttm && !ttm_tt_is_populated (bo->ttm)) { + ret =ttm_tt_populate (bo->bdev, bo->ttm, &ctx); +if (ret) +return ret; + +i915_ttm_adjust_domains_after_cpu_move (obj); +i915_ttm_adjust_gem_after_move (obj); + } +/* Object either has a page vector or is an iomem object */ st = bo->ttm ?i915_ttm_tt_get_st (bo->ttm) : obj->ttm.cached_io_st;if (IS_ERR (st))@@ -428,8 +489,6 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj)__i915_gem_object_set_pages (obj, st,i915_sg_dma_sizes (st->sgl)); -i915_ttm_adjust_lru (obj); -return ret; }@@ -563,6 +622,7 @@ static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj)const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = { .name ="i915_gem_object_ttm" , + .flags =I915_GEM_OBJECT_IS_SHRINKABLE , .get_pages = i915_ttm_get_pages, .put_pages = i915_ttm_put_pages,@@ -599,6 +659,10 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem, {static struct lock_class_key lock_class;struct drm_i915_private *i915 = mem->i915; +struct ttm_operation_ctx ctx = { + .interruptible = true, + .no_wait_gpu = false, + };enum ttm_bo_type bo_type;size_t alignment =0 ;int ret;@@ -618,15 +682,14 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,i915_gem_object_init (obj, &i915_gem_ttm_obj_ops, &lock_class, flags);i915_gem_object_init_memory_region (obj, mem);i915_gem_object_make_unshrinkable (obj); - obj->read_domains =I915_GEM_DOMAIN_WC |I915_GEM_DOMAIN_GTT ; - obj->mem_flags |=I915_BO_FLAG_IOMEM ; -i915_gem_object_set_cache_coherency (obj,I915_CACHE_NONE );INIT_RADIX_TREE (&obj->ttm.get_io_page.radix,GFP_KERNEL | __GFP_NOWARN);mutex_init (&obj->ttm.get_io_page.lock); bo_type = (obj->flags &I915_BO_ALLOC_USER ) ? ttm_bo_type_device : ttm_bo_type_kernel; + obj->base.vma_node.driver_private =i915_gem_to_ttm (obj); +/* * If this function fails, it will call the destructor, but * our caller still owns the object. So no freeing in the @@ -634,14 +697,19 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,* Similarly, in delayed_destroy, we can't call ttm_bo_put() * until successful initialization. */ - obj->base.vma_node.driver_private =i915_gem_to_ttm (obj); - ret =ttm_bo_init (&i915->bdev,i915_gem_to_ttm (obj), size, - bo_type, &i915_sys_placement, alignment, - true,NULL ,NULL , i915_ttm_bo_destroy); + ret =ttm_bo_init_reserved (&i915->bdev,i915_gem_to_ttm (obj), size, + bo_type, &i915_sys_placement, alignment, + &ctx,NULL ,NULL , i915_ttm_bo_destroy); + +if (ret) + goto out; -if (!ret) - obj->ttm.created = true; + obj->ttm.created = true; +i915_ttm_adjust_domains_after_cpu_move (obj); +i915_ttm_adjust_gem_after_move (obj); +i915_gem_object_unlock (obj); +out:/* i915 wants -ENXIO when out of memory region space. */ return (ret == -ENOSPC ) ? -ENXIO : ret; } --2.31.1
Quoted 4 lines
Quoted 4 lines
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH 1/5] drm/i915: Update object placement flags to be mutable
In reply to Matthew AuldQuoted 10 lines · Matthew Auld
Quoted 6 lines · Thomas Hellström
The object ops
i915_GEM_OBJECT_HAS_IOMEMand the objectI915_BO_ALLOC_STRUCT_PAGEflags are considered immutable by much of our code. Introduce a newmem_flagsmember to hold these and make sure checks for these flags being set are either done under the object lock or with pages properly pinned. The flags will change during migration under the object lock.What are the rules for the
gem_object_opsflags? Likeis_shrinkable? Can't we just move these there(IOMEM vs PAGE)?
But the ops flags are immutable, right? We expect the IOMEM and PAGE flags to change when an object is migrated. We could of course flip the ops under the lock, but I'd consider that a bit risky?
Quoted 374 lines
Quoted 374 lines
Signed-off-by Thomas Hellström <[email protected]>--- drivers/gpu/drm/i915/gem/i915_gem_internal.c | 4 +- drivers/gpu/drm/i915/gem/i915_gem_mman.c | 7 +++- drivers/gpu/drm/i915/gem/i915_gem_object.c | 38 +++++++++++++++++++ drivers/gpu/drm/i915/gem/i915_gem_object.h | 14 ++----- .../gpu/drm/i915/gem/i915_gem_object_types.h | 20 +++++----- drivers/gpu/drm/i915/gem/i915_gem_pages.c | 2 +- drivers/gpu/drm/i915/gem/i915_gem_phys.c | 2 +- drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 10 +++-- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 2 +- drivers/gpu/drm/i915/gem/i915_gem_userptr.c | 4 +- .../drm/i915/gem/selftests/huge_gem_object.c | 4 +- .../gpu/drm/i915/gem/selftests/huge_pages.c | 5 +-- .../drm/i915/gem/selftests/i915_gem_mman.c | 4 +- .../drm/i915/gem/selftests/i915_gem_phys.c | 3 +- 14 files changed, 78 insertions(+), 41 deletions(-)drivers/gpu/drm/i915/gem/i915_gem_internal.c +2
-2@@ -177,8 +177,8 @@ i915_gem_object_create_internal(struct drm_i915_private *i915,return ERR_PTR (-ENOMEM );drm_gem_private_object_init (&i915->drm, &obj->base, size); -i915_gem_object_init (obj, &i915_gem_object_internal_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); +i915_gem_object_init (obj, &i915_gem_object_internal_ops, &lock_class,0 ); + obj->mem_flags |=I915_BO_FLAG_STRUCT_PAGE ;/* * Mark the object as volatile, such that the pages are marked as drivers/gpu/drm/i915/gem/i915_gem_mman.c +6
-1@@ -683,7 +683,7 @@ __assign_mmap_offset(struct drm_i915_gem_object *obj,if (mmap_type !=I915_MMAP_TYPE_GTT && !i915_gem_object_has_struct_page (obj) && - !i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM )) + !i915_gem_object_has_iomem (obj))return -ENODEV ; mmo =mmap_offset_attach (obj, mmap_type, file);@@ -707,7 +707,12 @@ __assign_mmap_offset_handle(struct drm_file *file,if (!obj)return -ENOENT ; + err =i915_gem_object_lock_interruptible (obj,NULL ); +if (err) + goto out_put; err =__assign_mmap_offset (obj, mmap_type, offset, file); +i915_gem_object_unlock (obj); +out_put:i915_gem_object_put (obj);return err; }drivers/gpu/drm/i915/gem/i915_gem_object.c +38
-0@@ -475,6 +475,44 @@ bool i915_gem_object_migratable(struct drm_i915_gem_object *obj)return obj->mm.n_placements >1 ; } +/** +* i915_gem_object_has_struct_page - Whether the object is page-backed +* @obj: The object to query. +* +* This function should only be called while the object is locked or pinned, +* otherwise the page backing may change under the caller. +* +* Return: True if page-backed, false otherwise. +*/ +bool i915_gem_object_has_struct_page (const struct drm_i915_gem_object *obj) +{ +#ifdef CONFIG_LOCKDEP +if (IS_DGFX (to_i915 (obj->base.dev)) && +i915_gem_object_evictable ((void __force *)obj)) +assert_object_held_shared (obj); +#endif +return obj->mem_flags &I915_BO_FLAG_STRUCT_PAGE ; +} + +/** +* i915_gem_object_has_iomem - Whether the object is iomem-backed +* @obj: The object to query. +* +* This function should only be called while the object is locked or pinned, +* otherwise the iomem backing may change under the caller. +* +* Return: True if iomem-backed, false otherwise. +*/ +bool i915_gem_object_has_iomem (const struct drm_i915_gem_object *obj) +{ +#ifdef CONFIG_LOCKDEP +if (IS_DGFX (to_i915 (obj->base.dev)) && +i915_gem_object_evictable ((void __force *)obj)) +assert_object_held_shared (obj); +#endif +return obj->mem_flags &I915_BO_FLAG_IOMEM ; +} +void i915_gem_init__objects (struct drm_i915_private *i915) {INIT_WORK (&i915->mm.free_work, __i915_gem_free_work);drivers/gpu/drm/i915/gem/i915_gem_object.h +3
-11@@ -147,7 +147,7 @@ i915_gem_object_put(struct drm_i915_gem_object *obj)/* * If more than one potential simultaneous locker, assert held. */ -static inline void assert_object_held_shared (struct drm_i915_gem_object *obj) +static inline void assert_object_held_shared (const struct drm_i915_gem_object *obj) {/* * Note mm list lookup is protected by @@ -261,17 +261,9 @@ i915_gem_object_type_has(const struct drm_i915_gem_object *obj,return obj->ops->flags & flags; } -static inline bool -i915_gem_object_has_struct_page (const struct drm_i915_gem_object *obj) -{ -return obj->flags &I915_BO_ALLOC_STRUCT_PAGE ; -} +bool i915_gem_object_has_struct_page (const struct drm_i915_gem_object *obj); -static inline bool -i915_gem_object_has_iomem (const struct drm_i915_gem_object *obj) -{ -return i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM ); -} +bool i915_gem_object_has_iomem (const struct drm_i915_gem_object *obj);static inline bool i915_gem_object_is_shrinkable (const struct drm_i915_gem_object *obj)drivers/gpu/drm/i915/gem/i915_gem_object_types.h +10
-10@@ -33,10 +33,9 @@ struct i915_lut_handle {struct drm_i915_gem_object_ops {unsigned int flags; -#define I915_GEM_OBJECT_HAS_IOMEM BIT(1) -#define I915_GEM_OBJECT_IS_SHRINKABLE BIT(2) -#define I915_GEM_OBJECT_IS_PROXY BIT(3) -#define I915_GEM_OBJECT_NO_MMAP BIT(4) +#define I915_GEM_OBJECT_IS_SHRINKABLE BIT(1) +#define I915_GEM_OBJECT_IS_PROXY BIT(2) +#define I915_GEM_OBJECT_NO_MMAP BIT(3)/* Interface between the GEM object and its backing storage. * get_pages() is called once prior to the use of the associated set @@ -201,17 +200,18 @@ struct drm_i915_gem_object {unsigned long flags;#define I915_BO_ALLOC_CONTIGUOUS BIT(0)#define I915_BO_ALLOC_VOLATILE BIT(1) -#define I915_BO_ALLOC_STRUCT_PAGE BIT(2) -#define I915_BO_ALLOC_CPU_CLEAR BIT(3) -#define I915_BO_ALLOC_USER BIT(4) +#define I915_BO_ALLOC_CPU_CLEAR BIT(2) +#define I915_BO_ALLOC_USER BIT(3)#define I915_BO_ALLOC_FLAGS (I915_BO_ALLOC_CONTIGUOUS | \ I915_BO_ALLOC_VOLATILE | \ - I915_BO_ALLOC_STRUCT_PAGE | \ I915_BO_ALLOC_CPU_CLEAR | \ I915_BO_ALLOC_USER) -#define I915_BO_READONLY BIT(5) -#define I915_TILING_QUIRK_BIT 6/* unknown swizzling; do not release! */ +#define I915_BO_READONLY BIT(4) +#define I915_TILING_QUIRK_BIT 5/* unknown swizzling; do not release! */ + unsigned int mem_flags:2; +#define I915_BO_FLAG_STRUCT_PAGE BIT(0) +#define I915_BO_FLAG_IOMEM BIT(1)/* * Is the object to be mapped as read-only to the GPU * Only honoured if hardware has relevant pte bit drivers/gpu/drm/i915/gem/i915_gem_pages.c +1
-1@@ -351,7 +351,7 @@ void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj,int err;if (!i915_gem_object_has_struct_page (obj) && - !i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM )) + !i915_gem_object_has_iomem (obj))return ERR_PTR (-ENXIO );assert_object_held (obj);drivers/gpu/drm/i915/gem/i915_gem_phys.c +1
-1@@ -76,7 +76,7 @@ static int i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj)intel_gt_chipset_flush (&to_i915 (obj->base.dev)->gt);/* We're no longer struct page backed */ - obj->flags &= ~I915_BO_ALLOC_STRUCT_PAGE ; + obj->mem_flags &= ~I915_BO_FLAG_STRUCT_PAGE ;__i915_gem_object_set_pages (obj, st, sg->length);return 0 ;drivers/gpu/drm/i915/gem/i915_gem_shmem.c +6
-4@@ -284,6 +284,7 @@ __i915_gem_object_release_shmem(struct drm_i915_gem_object *obj, bool needs_clflush) {GEM_BUG_ON (obj->mm.madv == __I915_MADV_PURGED); +GEM_WARN_ON (IS_DGFX (to_i915 (obj->base.dev)));if (obj->mm.madv ==I915_MADV_DONTNEED ) obj->mm.dirty = false;@@ -302,6 +303,7 @@ void i915_gem_object_put_pages_shmem(struct drm_i915_gem_object *obj, struct sg_struct pagevec pvec;struct page *page; +GEM_WARN_ON (IS_DGFX (to_i915 (obj->base.dev)));__i915_gem_object_release_shmem (obj, pages, true);i915_gem_gtt_finish_pages (obj, pages);@@ -444,7 +446,7 @@ shmem_pread(struct drm_i915_gem_object *obj,static void shmem_release (struct drm_i915_gem_object *obj) { -if (obj->flags &I915_BO_ALLOC_STRUCT_PAGE ) +if (i915_gem_object_has_struct_page (obj))i915_gem_object_release_memory_region (obj);fput (obj->base.filp);@@ -513,9 +515,8 @@ static int shmem_object_init(struct intel_memory_region *mem,mapping_set_gfp_mask (mapping, mask);GEM_BUG_ON (!(mapping_gfp_mask (mapping) & __GFP_RECLAIM)); -i915_gem_object_init (obj, &i915_gem_shmem_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); - +i915_gem_object_init (obj, &i915_gem_shmem_ops, &lock_class,0 ); + obj->mem_flags |=I915_BO_FLAG_STRUCT_PAGE ; obj->write_domain =I915_GEM_DOMAIN_CPU ; obj->read_domains =I915_GEM_DOMAIN_CPU ;@@ -561,6 +562,7 @@ i915_gem_object_create_shmem_from_data(struct drm_i915_private *dev_priv,resource_size_t offset;int err; +GEM_WARN_ON (IS_DGFX (dev_priv)); obj =i915_gem_object_create_shmem (dev_priv,round_up (size,PAGE_SIZE ));if (IS_ERR (obj))return obj;drivers/gpu/drm/i915/gem/i915_gem_ttm.c +1
-1@@ -563,7 +563,6 @@ static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj)const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = { .name ="i915_gem_object_ttm" , - .flags =I915_GEM_OBJECT_HAS_IOMEM , .get_pages = i915_ttm_get_pages, .put_pages = i915_ttm_put_pages,@@ -620,6 +619,7 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,i915_gem_object_init_memory_region (obj, mem);i915_gem_object_make_unshrinkable (obj); obj->read_domains =I915_GEM_DOMAIN_WC |I915_GEM_DOMAIN_GTT ; + obj->mem_flags |=I915_BO_FLAG_IOMEM ;i915_gem_object_set_cache_coherency (obj,I915_CACHE_NONE );INIT_RADIX_TREE (&obj->ttm.get_io_page.radix,GFP_KERNEL | __GFP_NOWARN);mutex_init (&obj->ttm.get_io_page.lock);drivers/gpu/drm/i915/gem/i915_gem_userptr.c +2
-2@@ -538,8 +538,8 @@ i915_gem_userptr_ioctl(struct drm_device *dev,return -ENOMEM ;drm_gem_private_object_init (dev, &obj->base, args->user_size); -i915_gem_object_init (obj, &i915_gem_userptr_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); +i915_gem_object_init (obj, &i915_gem_userptr_ops, &lock_class,0 ); + obj->mem_flags =I915_BO_FLAG_STRUCT_PAGE ; obj->read_domains =I915_GEM_DOMAIN_CPU ; obj->write_domain =I915_GEM_DOMAIN_CPU ;i915_gem_object_set_cache_coherency (obj,I915_CACHE_LLC );drivers/gpu/drm/i915/gem/selftests/huge_gem_object.c +2
-2@@ -114,8 +114,8 @@ huge_gem_object(struct drm_i915_private *i915,return ERR_PTR (-ENOMEM );drm_gem_private_object_init (&i915->drm, &obj->base, dma_size); -i915_gem_object_init (obj, &huge_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); +i915_gem_object_init (obj, &huge_ops, &lock_class,0 ); + obj->mem_flags |=I915_BO_FLAG_STRUCT_PAGE ; obj->read_domains =I915_GEM_DOMAIN_CPU ; obj->write_domain =I915_GEM_DOMAIN_CPU ;drivers/gpu/drm/i915/gem/selftests/huge_pages.c +2
-3@@ -167,9 +167,8 @@ huge_pages_object(struct drm_i915_private *i915,return ERR_PTR (-ENOMEM );drm_gem_private_object_init (&i915->drm, &obj->base, size); -i915_gem_object_init (obj, &huge_page_ops, &lock_class, -I915_BO_ALLOC_STRUCT_PAGE ); - +i915_gem_object_init (obj, &huge_page_ops, &lock_class,0 ); + obj->mem_flags |=I915_BO_FLAG_STRUCT_PAGE ;i915_gem_object_set_volatile (obj); obj->write_domain =I915_GEM_DOMAIN_CPU ;drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c +2
-2@@ -837,7 +837,7 @@ static bool can_mmap(struct drm_i915_gem_object *obj, enum i915_mmap_type type)if (type !=I915_MMAP_TYPE_GTT && !i915_gem_object_has_struct_page (obj) && - !i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM )) + !i915_gem_object_has_iomem (obj))return false;return true;@@ -991,7 +991,7 @@ static const char *repr_mmap_type(enum i915_mmap_type type)static bool can_access (const struct drm_i915_gem_object *obj) {return i915_gem_object_has_struct_page (obj) || -i915_gem_object_type_has (obj,I915_GEM_OBJECT_HAS_IOMEM ); +i915_gem_object_has_iomem (obj); } static int __igt_mmap_access(struct drm_i915_private *i915,drivers/gpu/drm/i915/gem/selftests/i915_gem_phys.c +2
-2@@ -25,13 +25,14 @@ static int mock_phys_object(void *arg) goto out; } +i915_gem_object_lock (obj,NULL );if (!i915_gem_object_has_struct_page (obj)) { +i915_gem_object_unlock (obj); err = -EINVAL ;pr_err ("shmem has no struct page\n" ); goto out_obj; } -i915_gem_object_lock (obj,NULL ); err =i915_gem_object_attach_phys (obj,PAGE_SIZE );i915_gem_object_unlock (obj);if (err) { --2.31.1
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH 4/5] drm/i915/ttm: Use TTM for system memory
In reply to Matthew AuldQuoted 34 lines · Matthew Auld
Quoted 32 lines · Thomas Hellström
For discrete, use TTM for both cached and WC system memory. That means we currently rely on the TTM memory accounting / shrinker. For cached system memory we should consider remaining shmem-backed, which can be implemented from our
ttm_tt_populatecalback. We can then also reuse our own very elaborate shrinker for that memory.Signed-off-by Thomas Hellström <[email protected]>--- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 22 ++++++++++++++++++++++ drivers/gpu/drm/i915/i915_drv.h | 3 --- drivers/gpu/drm/i915/intel_memory_region.c | 7 ++++++- drivers/gpu/drm/i915/intel_memory_region.h | 8 ++++++++ 4 files changed, 36 insertions(+), 4 deletions(-)drivers/gpu/drm/i915/gem/i915_gem_ttm.c +10
-0@@ -755,3 +755,25 @@ int __i915_gem_ttm_object_init(struct intel_memory_region *mem,/* i915 wants -ENXIO when out of memory region space. */ return (ret == -ENOSPC ) ? -ENXIO : ret; } + +static const struct intel_memory_region_ops ttm_system_region_ops = { + .init_object = __i915_gem_ttm_object_init, +}; + +struct intel_memory_region * +i915_gem_ttm_system_setup (struct drm_i915_private *i915, +u16 type,u16 instance) +{ +struct intel_memory_region *mr;
Quoted 8 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c int __i915_gem_ttm_object_init(struct intel_memory_region *mem, · line 774
Quoted 7 lines
drivers/gpu/drm/i915/gem/i915_gem_ttm.c +7
-0+ + mr =intel_memory_region_create (i915,0 , +totalram_pages () <<PAGE_SHIFT , +PAGE_SIZE ,0 , + type, instance, + &ttm_system_region_ops); +if (IS_ERR_OR_NULL (mr))
region_createcan't return NULL.
OK, will fix.