Most logical place to introduce TTM buffer objects is as an i915 gem object backend. We need to add some ops to account for added functionality like delayed delete and LRU list manipulation. Initially we support only LMEM and SYSTEM memory, but SYSTEM (which in this case means evicted LMEM objects) is not visible to i915 GEM yet. The plan is to move the i915 gem system region over to the TTM system memory type in upcoming patches. We set up GPU bindings directly both from LMEM and from the system region, as there is no need to use the legacy TTM_TT memory type. We reserve that for future porting of GGTT bindings to TTM. Remove the old lmem backend. Signed-off-by: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx> Reviewed-by: Matthew Auld <matthew.auld@xxxxxxxxx> --- v2: - Break out needed TTM functionality to a separate patch (Reported by Christian König). - Fix an unhandled error (Reported by Matthew Auld and Maarten Lankhorst) - Remove a stray leftover sg_table allocation (Reported by Matthew Auld) - Use ttm_tt_unpopulate() rather than ttm_tt_destroy() in the purge path as some TTM functionality relies on having a ttm_tt present for !is_iomem. v3: - Use ttm_bo_type_device for userspace visible objects so that TTM can allocate an address space offset for mmap'ing. - Fix up the destruction path (Reported by Matthew Auld) - Use ttm_bo_validate() for purging (Reported by Christian König) - Create ttm_tts write-combined as they are currently for eviction only and we want to maintain consistent write-combined caching for bos that are not in system only. (Suggested by Daniel Vetter) - Make struct ttm_placements static. - Add the ttm device funcs/ops to i915_gem_ttm.h for the region code. - Rename new->dst and old->src. Check for swapin in the move callback. v4: - Adapt to small interface change in ttm_move_memcpy. - Use a function to pull out the ttm driver from the backend. --- drivers/gpu/drm/i915/Makefile | 1 + drivers/gpu/drm/i915/gem/i915_gem_create.c | 9 +- drivers/gpu/drm/i915/gem/i915_gem_lmem.c | 84 --- drivers/gpu/drm/i915/gem/i915_gem_lmem.h | 5 - drivers/gpu/drm/i915/gem/i915_gem_object.c | 125 ++-- drivers/gpu/drm/i915/gem/i915_gem_object.h | 9 + .../gpu/drm/i915/gem/i915_gem_object_types.h | 27 +- drivers/gpu/drm/i915/gem/i915_gem_region.c | 6 +- drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 541 ++++++++++++++++++ drivers/gpu/drm/i915/gem/i915_gem_ttm.h | 48 ++ drivers/gpu/drm/i915/gt/intel_region_lmem.c | 3 +- drivers/gpu/drm/i915/i915_gem.c | 5 +- drivers/gpu/drm/i915/intel_memory_region.c | 1 - drivers/gpu/drm/i915/intel_memory_region.h | 1 - drivers/gpu/drm/i915/intel_region_ttm.c | 8 +- drivers/gpu/drm/i915/intel_region_ttm.h | 11 +- 16 files changed, 731 insertions(+), 153 deletions(-) create mode 100644 drivers/gpu/drm/i915/gem/i915_gem_ttm.c create mode 100644 drivers/gpu/drm/i915/gem/i915_gem_ttm.h diff --git a/drivers/gpu/drm/i915/Makefile b/drivers/gpu/drm/i915/Makefile index 4f22cac1c49b..f57dfc74d6ce 100644 --- a/drivers/gpu/drm/i915/Makefile +++ b/drivers/gpu/drm/i915/Makefile @@ -155,6 +155,7 @@ gem-y += \ gem/i915_gem_stolen.o \ gem/i915_gem_throttle.o \ gem/i915_gem_tiling.o \ + gem/i915_gem_ttm.o \ gem/i915_gem_userptr.o \ gem/i915_gem_wait.o \ gem/i915_gemfs.o diff --git a/drivers/gpu/drm/i915/gem/i915_gem_create.c b/drivers/gpu/drm/i915/gem/i915_gem_create.c index 548ddf39d853..93bf63bbaff1 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_create.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_create.c @@ -85,13 +85,10 @@ i915_gem_setup(struct drm_i915_gem_object *obj, u64 size) return -E2BIG; /* - * For now resort to CPU based clearing for device local-memory, in the - * near future this will use the blitter engine for accelerated, GPU - * based clearing. + * I915_BO_ALLOC_USER will make sure the object is cleared before + * any user access. */ - flags = 0; - if (mr->type == INTEL_MEMORY_LOCAL) - flags = I915_BO_ALLOC_CPU_CLEAR; + flags = I915_BO_ALLOC_USER; ret = mr->ops->init_object(mr, obj, size, flags); if (ret) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_lmem.c b/drivers/gpu/drm/i915/gem/i915_gem_lmem.c index 3b4aa28a076d..2b8cd15de1d9 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_lmem.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_lmem.c @@ -4,74 +4,10 @@ */ #include "intel_memory_region.h" -#include "intel_region_ttm.h" #include "gem/i915_gem_region.h" #include "gem/i915_gem_lmem.h" #include "i915_drv.h" -static void lmem_put_pages(struct drm_i915_gem_object *obj, - struct sg_table *pages) -{ - intel_region_ttm_node_free(obj->mm.region, obj->mm.st_mm_node); - obj->mm.dirty = false; - sg_free_table(pages); - kfree(pages); -} - -static int lmem_get_pages(struct drm_i915_gem_object *obj) -{ - unsigned int flags; - struct sg_table *pages; - - flags = I915_ALLOC_MIN_PAGE_SIZE; - if (obj->flags & I915_BO_ALLOC_CONTIGUOUS) - flags |= I915_ALLOC_CONTIGUOUS; - - obj->mm.st_mm_node = intel_region_ttm_node_alloc(obj->mm.region, - obj->base.size, - flags); - if (IS_ERR(obj->mm.st_mm_node)) - return PTR_ERR(obj->mm.st_mm_node); - - /* Range manager is always contigous */ - if (obj->mm.region->is_range_manager) - obj->flags |= I915_BO_ALLOC_CONTIGUOUS; - pages = intel_region_ttm_node_to_st(obj->mm.region, obj->mm.st_mm_node); - if (IS_ERR(pages)) { - intel_region_ttm_node_free(obj->mm.region, obj->mm.st_mm_node); - return PTR_ERR(pages); - } - - __i915_gem_object_set_pages(obj, pages, i915_sg_dma_sizes(pages->sgl)); - - if (obj->flags & I915_BO_ALLOC_CPU_CLEAR) { - void __iomem *vaddr = - i915_gem_object_lmem_io_map(obj, 0, obj->base.size); - - if (!vaddr) { - struct sg_table *pages = - __i915_gem_object_unset_pages(obj); - - if (!IS_ERR_OR_NULL(pages)) - lmem_put_pages(obj, pages); - } - - memset_io(vaddr, 0, obj->base.size); - io_mapping_unmap(vaddr); - } - - return 0; -} - -const struct drm_i915_gem_object_ops i915_gem_lmem_obj_ops = { - .name = "i915_gem_object_lmem", - .flags = I915_GEM_OBJECT_HAS_IOMEM, - - .get_pages = lmem_get_pages, - .put_pages = lmem_put_pages, - .release = i915_gem_object_release_memory_region, -}; - void __iomem * i915_gem_object_lmem_io_map(struct drm_i915_gem_object *obj, unsigned long n, @@ -103,23 +39,3 @@ i915_gem_object_create_lmem(struct drm_i915_private *i915, return i915_gem_object_create_region(i915->mm.regions[INTEL_REGION_LMEM], size, flags); } - -int __i915_gem_lmem_object_init(struct intel_memory_region *mem, - struct drm_i915_gem_object *obj, - resource_size_t size, - unsigned int flags) -{ - static struct lock_class_key lock_class; - struct drm_i915_private *i915 = mem->i915; - - drm_gem_private_object_init(&i915->drm, &obj->base, size); - i915_gem_object_init(obj, &i915_gem_lmem_obj_ops, &lock_class, flags); - - obj->read_domains = I915_GEM_DOMAIN_WC | I915_GEM_DOMAIN_GTT; - - i915_gem_object_set_cache_coherency(obj, I915_CACHE_NONE); - - i915_gem_object_init_memory_region(obj, mem); - - return 0; -} diff --git a/drivers/gpu/drm/i915/gem/i915_gem_lmem.h b/drivers/gpu/drm/i915/gem/i915_gem_lmem.h index fac6bc5a5ebb..ea76fd11ccb0 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_lmem.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_lmem.h @@ -26,9 +26,4 @@ i915_gem_object_create_lmem(struct drm_i915_private *i915, resource_size_t size, unsigned int flags); -int __i915_gem_lmem_object_init(struct intel_memory_region *mem, - struct drm_i915_gem_object *obj, - resource_size_t size, - unsigned int flags); - #endif /* !__I915_GEM_LMEM_H */ diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c index 5706d471692d..16eac5ea9238 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c @@ -172,7 +172,7 @@ static void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *f } } -static void __i915_gem_free_object_rcu(struct rcu_head *head) +void __i915_gem_free_object_rcu(struct rcu_head *head) { struct drm_i915_gem_object *obj = container_of(head, typeof(*obj), rcu); @@ -208,59 +208,69 @@ static void __i915_gem_object_free_mmaps(struct drm_i915_gem_object *obj) } } -static void __i915_gem_free_objects(struct drm_i915_private *i915, - struct llist_node *freed) +void __i915_gem_free_object(struct drm_i915_gem_object *obj) { - struct drm_i915_gem_object *obj, *on; + trace_i915_gem_object_destroy(obj); - llist_for_each_entry_safe(obj, on, freed, freed) { - trace_i915_gem_object_destroy(obj); + if (!list_empty(&obj->vma.list)) { + struct i915_vma *vma; + + /* + * Note that the vma keeps an object reference while + * it is active, so it *should* not sleep while we + * destroy it. Our debug code errs insits it *might*. + * For the moment, play along. + */ + spin_lock(&obj->vma.lock); + while ((vma = list_first_entry_or_null(&obj->vma.list, + struct i915_vma, + obj_link))) { + GEM_BUG_ON(vma->obj != obj); + spin_unlock(&obj->vma.lock); - if (!list_empty(&obj->vma.list)) { - struct i915_vma *vma; + __i915_vma_put(vma); - /* - * Note that the vma keeps an object reference while - * it is active, so it *should* not sleep while we - * destroy it. Our debug code errs insits it *might*. - * For the moment, play along. - */ spin_lock(&obj->vma.lock); - while ((vma = list_first_entry_or_null(&obj->vma.list, - struct i915_vma, - obj_link))) { - GEM_BUG_ON(vma->obj != obj); - spin_unlock(&obj->vma.lock); + } + spin_unlock(&obj->vma.lock); + } - __i915_vma_put(vma); + __i915_gem_object_free_mmaps(obj); - spin_lock(&obj->vma.lock); - } - spin_unlock(&obj->vma.lock); - } + GEM_BUG_ON(!list_empty(&obj->lut_list)); - __i915_gem_object_free_mmaps(obj); + atomic_set(&obj->mm.pages_pin_count, 0); + __i915_gem_object_put_pages(obj); + GEM_BUG_ON(i915_gem_object_has_pages(obj)); + bitmap_free(obj->bit_17); - GEM_BUG_ON(!list_empty(&obj->lut_list)); + if (obj->base.import_attach) + drm_prime_gem_destroy(&obj->base, NULL); - atomic_set(&obj->mm.pages_pin_count, 0); - __i915_gem_object_put_pages(obj); - GEM_BUG_ON(i915_gem_object_has_pages(obj)); - bitmap_free(obj->bit_17); + drm_gem_free_mmap_offset(&obj->base); - if (obj->base.import_attach) - drm_prime_gem_destroy(&obj->base, NULL); + if (obj->ops->release) + obj->ops->release(obj); - drm_gem_free_mmap_offset(&obj->base); + if (obj->mm.n_placements > 1) + kfree(obj->mm.placements); - if (obj->ops->release) - obj->ops->release(obj); + if (obj->shares_resv_from) + i915_vm_resv_put(obj->shares_resv_from); +} - if (obj->mm.n_placements > 1) - kfree(obj->mm.placements); +static void __i915_gem_free_objects(struct drm_i915_private *i915, + struct llist_node *freed) +{ + struct drm_i915_gem_object *obj, *on; - if (obj->shares_resv_from) - i915_vm_resv_put(obj->shares_resv_from); + llist_for_each_entry_safe(obj, on, freed, freed) { + might_sleep(); + if (obj->ops->delayed_free) { + obj->ops->delayed_free(obj); + continue; + } + __i915_gem_free_object(obj); /* But keep the pointer alive for RCU-protected lookups */ call_rcu(&obj->rcu, __i915_gem_free_object_rcu); @@ -318,6 +328,7 @@ static void i915_gem_free_object(struct drm_gem_object *gem_obj) * worker and performing frees directly from subsequent allocations for * crude but effective memory throttling. */ + if (llist_add(&obj->freed, &i915->mm.free_list)) queue_work(i915->wq, &i915->mm.free_work); } @@ -410,6 +421,42 @@ int i915_gem_object_read_from_page(struct drm_i915_gem_object *obj, u64 offset, return 0; } +/** + * i915_gem_object_evictable - Whether object is likely evictable after unbind. + * @obj: The object to check + * + * This function checks whether the object is likely unvictable after unbind. + * If the object is not locked when checking, the result is only advisory. + * If the object is locked when checking, and the function returns true, + * then an eviction should indeed be possible. But since unlocked vma + * unpinning and unbinding is currently possible, the object can actually + * become evictable even if this function returns false. + * + * Return: true if the object may be evictable. False otherwise. + */ +bool i915_gem_object_evictable(struct drm_i915_gem_object *obj) +{ + struct i915_vma *vma; + int pin_count = atomic_read(&obj->mm.pages_pin_count); + + if (!pin_count) + return true; + + spin_lock(&obj->vma.lock); + list_for_each_entry(vma, &obj->vma.list, obj_link) { + if (i915_vma_is_pinned(vma)) { + spin_unlock(&obj->vma.lock); + return false; + } + if (atomic_read(&vma->pages_count)) + pin_count--; + } + spin_unlock(&obj->vma.lock); + GEM_WARN_ON(pin_count < 0); + + return pin_count == 0; +} + void i915_gem_init__objects(struct drm_i915_private *i915) { INIT_WORK(&i915->mm.free_work, __i915_gem_free_work); diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h b/drivers/gpu/drm/i915/gem/i915_gem_object.h index 2ebd79537aea..ae5930e307d5 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h @@ -200,6 +200,9 @@ static inline bool i915_gem_object_trylock(struct drm_i915_gem_object *obj) static inline void i915_gem_object_unlock(struct drm_i915_gem_object *obj) { + if (obj->ops->adjust_lru) + obj->ops->adjust_lru(obj); + dma_resv_unlock(obj->base.resv); } @@ -587,6 +590,12 @@ int i915_gem_object_read_from_page(struct drm_i915_gem_object *obj, u64 offset, bool i915_gem_object_is_shmem(const struct drm_i915_gem_object *obj); +void __i915_gem_free_object_rcu(struct rcu_head *head); + +void __i915_gem_free_object(struct drm_i915_gem_object *obj); + +bool i915_gem_object_evictable(struct drm_i915_gem_object *obj); + #ifdef CONFIG_MMU_NOTIFIER static inline bool i915_gem_object_is_userptr(struct drm_i915_gem_object *obj) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object_types.h b/drivers/gpu/drm/i915/gem/i915_gem_object_types.h index d047ea126029..68313474e6a6 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object_types.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object_types.h @@ -63,6 +63,20 @@ struct drm_i915_gem_object_ops { const struct drm_i915_gem_pwrite *arg); int (*dmabuf_export)(struct drm_i915_gem_object *obj); + + /** + * adjust_lru - notify that the madvise value was updated + * @obj: The gem object + * + * The madvise value may have been updated, or object was recently + * referenced so act accordingly (Perhaps changing an LRU list etc). + */ + void (*adjust_lru)(struct drm_i915_gem_object *obj); + + /** + * delayed_free - Override the default delayed free implementation + */ + void (*delayed_free)(struct drm_i915_gem_object *obj); void (*release)(struct drm_i915_gem_object *obj); const char *name; /* friendly name for debug, e.g. lockdep classes */ @@ -187,12 +201,14 @@ struct drm_i915_gem_object { #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_FLAGS (I915_BO_ALLOC_CONTIGUOUS | \ I915_BO_ALLOC_VOLATILE | \ I915_BO_ALLOC_STRUCT_PAGE | \ - I915_BO_ALLOC_CPU_CLEAR) -#define I915_BO_READONLY BIT(4) -#define I915_TILING_QUIRK_BIT 5 /* unknown swizzling; do not release! */ + 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! */ /* * Is the object to be mapped as read-only to the GPU @@ -310,6 +326,11 @@ struct drm_i915_gem_object { bool dirty:1; } mm; + struct { + struct sg_table *cached_io_st; + bool created:1; + } ttm; + /** Record of address bit 17 of each page at last unbind. */ unsigned long *bit_17; diff --git a/drivers/gpu/drm/i915/gem/i915_gem_region.c b/drivers/gpu/drm/i915/gem/i915_gem_region.c index f25e6646c5b7..d1f1840540dd 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_region.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_region.c @@ -18,11 +18,7 @@ void i915_gem_object_init_memory_region(struct drm_i915_gem_object *obj, mutex_lock(&mem->objects.lock); - if (obj->flags & I915_BO_ALLOC_VOLATILE) - list_add(&obj->mm.region_link, &mem->objects.purgeable); - else - list_add(&obj->mm.region_link, &mem->objects.list); - + list_add(&obj->mm.region_link, &mem->objects.list); mutex_unlock(&mem->objects.lock); } diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c new file mode 100644 index 000000000000..17598930a99e --- /dev/null +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c @@ -0,0 +1,541 @@ +// SPDX-License-Identifier: MIT +/* + * Copyright © 2021 Intel Corporation + */ + +#include <drm/ttm/ttm_bo_driver.h> +#include <drm/ttm/ttm_placement.h> + +#include "i915_drv.h" +#include "intel_memory_region.h" +#include "intel_region_ttm.h" + +#include "gem/i915_gem_object.h" +#include "gem/i915_gem_region.h" +#include "gem/i915_gem_ttm.h" + +#define I915_PL_LMEM0 TTM_PL_PRIV +#define I915_PL_SYSTEM TTM_PL_SYSTEM +#define I915_PL_STOLEN TTM_PL_VRAM +#define I915_PL_GGTT TTM_PL_TT + +#define I915_TTM_PRIO_PURGE 0 +#define I915_TTM_PRIO_NO_PAGES 1 +#define I915_TTM_PRIO_HAS_PAGES 2 + +/** + * struct i915_ttm_tt - TTM page vector with additional private information + * @ttm: The base TTM page vector. + * @dev: The struct device used for dma mapping and unmapping. + * @cached_st: The cached scatter-gather table. + * + * Note that DMA may be going on right up to the point where the page- + * vector is unpopulated in delayed destroy. Hence keep the + * scatter-gather table mapped and cached up to that point. This is + * different from the cached gem object io scatter-gather table which + * doesn't have an associated dma mapping. + */ +struct i915_ttm_tt { + struct ttm_tt ttm; + struct device *dev; + 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 struct ttm_placement i915_sys_placement = { + .num_placement = 1, + .placement = &lmem0_sys_placement_flags[1], + .num_busy_placement = 1, + .busy_placement = &lmem0_sys_placement_flags[1], +}; + +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, + uint32_t page_flags) +{ + struct ttm_resource_manager *man = + ttm_manager_type(bo->bdev, bo->mem.mem_type); + struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); + struct i915_ttm_tt *i915_tt; + int ret; + + i915_tt = kzalloc(sizeof(*i915_tt), GFP_KERNEL); + if (!i915_tt) + return NULL; + + if (obj->flags & I915_BO_ALLOC_CPU_CLEAR && + man->use_tt) + page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC; + + ret = ttm_tt_init(&i915_tt->ttm, bo, page_flags, ttm_write_combined); + if (ret) { + kfree(i915_tt); + return NULL; + } + + i915_tt->dev = obj->base.dev->dev; + + return &i915_tt->ttm; +} + +static void i915_ttm_tt_unpopulate(struct ttm_device *bdev, struct ttm_tt *ttm) +{ + struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); + + if (i915_tt->cached_st) { + dma_unmap_sgtable(i915_tt->dev, i915_tt->cached_st, + DMA_BIDIRECTIONAL, 0); + sg_free_table(i915_tt->cached_st); + kfree(i915_tt->cached_st); + i915_tt->cached_st = NULL; + } + ttm_pool_free(&bdev->pool, ttm); +} + +static void i915_ttm_tt_destroy(struct ttm_device *bdev, struct ttm_tt *ttm) +{ + struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); + + ttm_tt_destroy_common(bdev, ttm); + kfree(i915_tt); +} + +static bool i915_ttm_eviction_valuable(struct ttm_buffer_object *bo, + const struct ttm_place *place) +{ + struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); + + /* Will do for now. Our pinned objects are still on TTM's LRU lists */ + if (!i915_gem_object_evictable(obj)) + return false; + + /* This isn't valid with a buddy allocator */ + return ttm_bo_eviction_valuable(bo, place); +} + +static void i915_ttm_evict_flags(struct ttm_buffer_object *bo, + struct ttm_placement *placement) +{ + *placement = i915_sys_placement; +} + +static int i915_ttm_move_notify(struct ttm_buffer_object *bo) +{ + struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); + int ret; + + 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; + + return 0; +} + +static void i915_ttm_free_cached_io_st(struct drm_i915_gem_object *obj) +{ + if (obj->ttm.cached_io_st) { + sg_free_table(obj->ttm.cached_io_st); + kfree(obj->ttm.cached_io_st); + obj->ttm.cached_io_st = NULL; + } +} + +static void i915_ttm_purge(struct drm_i915_gem_object *obj) +{ + struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); + struct ttm_operation_ctx ctx = { + .interruptible = true, + .no_wait_gpu = false, + }; + struct ttm_placement place = {}; + int ret; + + if (obj->mm.madv == __I915_MADV_PURGED) + return; + + /* TTM's purge interface. Note that we might be reentering. */ + ret = ttm_bo_validate(bo, &place, &ctx); + + if (!ret) { + i915_ttm_free_cached_io_st(obj); + obj->mm.madv = __I915_MADV_PURGED; + } +} + +static void i915_ttm_swap_notify(struct ttm_buffer_object *bo) +{ + struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); + int ret = i915_ttm_move_notify(bo); + + GEM_WARN_ON(ret); + GEM_WARN_ON(obj->ttm.cached_io_st); + if (!ret && obj->mm.madv != I915_MADV_WILLNEED) + i915_ttm_purge(obj); +} + +static void i915_ttm_delete_mem_notify(struct ttm_buffer_object *bo) +{ + struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); + + if (likely(obj)) { + /* This releases all gem object bindings to the backend. */ + __i915_gem_free_object(obj); + } +} + +static struct intel_memory_region * +i915_ttm_region(struct ttm_device *bdev, int ttm_mem_type) +{ + struct drm_i915_private *i915 = container_of(bdev, typeof(*i915), bdev); + + /* There's some room for optimization here... */ + GEM_BUG_ON(ttm_mem_type != I915_PL_SYSTEM && + ttm_mem_type < I915_PL_LMEM0); + if (ttm_mem_type == I915_PL_SYSTEM) + return intel_memory_region_lookup(i915, INTEL_MEMORY_SYSTEM, + 0); + + return intel_memory_region_lookup(i915, INTEL_MEMORY_LOCAL, + ttm_mem_type - I915_PL_LMEM0); +} + +static struct sg_table *i915_ttm_tt_get_st(struct ttm_tt *ttm) +{ + struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); + struct scatterlist *sg; + struct sg_table *st; + int ret; + + if (i915_tt->cached_st) + return i915_tt->cached_st; + + st = kzalloc(sizeof(*st), GFP_KERNEL); + if (!st) + return ERR_PTR(-ENOMEM); + + sg = __sg_alloc_table_from_pages + (st, ttm->pages, ttm->num_pages, 0, + (unsigned long)ttm->num_pages << PAGE_SHIFT, + i915_sg_segment_size(), NULL, 0, GFP_KERNEL); + if (IS_ERR(sg)) { + kfree(st); + return ERR_CAST(sg); + } + + ret = dma_map_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 0); + if (ret) { + sg_free_table(st); + kfree(st); + return ERR_PTR(ret); + } + + i915_tt->cached_st = st; + return st; +} + +static struct sg_table * +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) + return i915_ttm_tt_get_st(bo->ttm); + + return intel_region_ttm_node_to_st(obj->mm.region, res->mm_node); +} + +static int i915_ttm_move(struct ttm_buffer_object *bo, bool evict, + struct ttm_operation_ctx *ctx, + struct ttm_resource *dst_mem, + struct ttm_place *hop) +{ + 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; + struct ttm_kmap_iter_iomap io; + } _dst_iter, _src_iter; + struct ttm_kmap_iter *dst_iter, *src_iter; + struct sg_table *dst_st; + int ret; + + dst_reg = i915_ttm_region(bo->bdev, dst_mem->mem_type); + src_reg = i915_ttm_region(bo->bdev, bo->mem.mem_type); + GEM_BUG_ON(!dst_reg || !src_reg); + + /* Sync for now. We could do the actual copy async. */ + ret = ttm_bo_wait_ctx(bo, ctx); + if (ret) + return ret; + + ret = i915_ttm_move_notify(bo); + if (ret) + return ret; + + if (obj->mm.madv != I915_MADV_WILLNEED) { + i915_ttm_purge(obj); + ttm_resource_free(bo, dst_mem); + return 0; + } + + /* Populate ttm with pages if needed. Typically system memory. */ + if (bo->ttm && (dst_man->use_tt || + (bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED))) { + ret = ttm_tt_populate(bo->bdev, bo->ttm, ctx); + if (ret) + return ret; + } + + dst_st = i915_ttm_resource_get_st(obj, dst_mem); + 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 ? + 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 ? + 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); + ttm_bo_move_sync_cleanup(bo, dst_mem); + i915_ttm_free_cached_io_st(obj); + + if (!dst_man->use_tt) + obj->ttm.cached_io_st = dst_st; + + return 0; +} + +static struct ttm_device_funcs i915_ttm_bo_driver = { + .ttm_tt_create = i915_ttm_tt_create, + .ttm_tt_unpopulate = i915_ttm_tt_unpopulate, + .ttm_tt_destroy = i915_ttm_tt_destroy, + .eviction_valuable = i915_ttm_eviction_valuable, + .evict_flags = i915_ttm_evict_flags, + .move = i915_ttm_move, + .verify_access = NULL, + .swap_notify = i915_ttm_swap_notify, + .delete_mem_notify = i915_ttm_delete_mem_notify, +}; + +/** + * i915_ttm_driver - Return a pointer to the TTM device funcs + * + * Return: Pointer to statically allocated TTM device funcs. + */ +struct ttm_device_funcs *i915_ttm_driver(void) +{ + return &i915_ttm_bo_driver; +} + +static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) +{ + struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); + struct ttm_operation_ctx ctx = { + .interruptible = true, + .no_wait_gpu = false, + }; + struct sg_table *st; + int ret; + + /* Move to the requested placement. */ + ret = ttm_bo_validate(bo, &i915_lmem0_placement, &ctx); + if (ret) + return ret == -ENOSPC ? -ENXIO : ret; + + /* 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)) + return PTR_ERR(st); + + __i915_gem_object_set_pages(obj, st, i915_sg_dma_sizes(st->sgl)); + + i915_ttm_adjust_lru(obj); + + return ret; +} + +static void i915_ttm_put_pages(struct drm_i915_gem_object *obj, + struct sg_table *st) +{ + /* + * We're currently not called from a shrinker, so put_pages() + * typically means the object is about to destroyed, or called + * from move_notify(). So just avoid doing much for now. + * If the object is not destroyed next, The TTM eviction logic + * and shrinkers will move it out if needed. + */ + + i915_ttm_adjust_lru(obj); +} + +static void i915_ttm_adjust_lru(struct drm_i915_gem_object *obj) +{ + struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); + + /* + * Don't manipulate the TTM LRUs while in TTM bo destruction. + * We're called through i915_ttm_delete_mem_notify(). + */ + if (!kref_read(&bo->kref)) + return; + + /* + * Put on the correct LRU list depending on the MADV status + */ + spin_lock(&bo->bdev->lru_lock); + if (obj->mm.madv != I915_MADV_WILLNEED) { + bo->priority = I915_TTM_PRIO_PURGE; + } else if (!i915_gem_object_has_pages(obj)) { + if (bo->priority < I915_TTM_PRIO_HAS_PAGES) + bo->priority = I915_TTM_PRIO_HAS_PAGES; + } else { + if (bo->priority > I915_TTM_PRIO_NO_PAGES) + bo->priority = I915_TTM_PRIO_NO_PAGES; + } + + ttm_bo_move_to_lru_tail(bo, &bo->mem, NULL); + spin_unlock(&bo->bdev->lru_lock); +} + +/* + * TTM-backed gem object destruction requires some clarification. + * Basically we have two possibilities here. We can either rely on the + * i915 delayed destruction and put the TTM object when the object + * is idle. This would be detected by TTM which would bypass the + * TTM delayed destroy handling. The other approach is to put the TTM + * object early and rely on the TTM destroyed handling, and then free + * the leftover parts of the GEM object once TTM's destroyed list handling is + * complete. For now, we rely on the latter for two reasons: + * a) TTM can evict an object even when it's on the delayed destroy list, + * which in theory allows for complete eviction. + * b) There is work going on in TTM to allow freeing an object even when + * it's not idle, and using the TTM destroyed list handling could help us + * benefit from that. + */ +static void i915_ttm_delayed_free(struct drm_i915_gem_object *obj) +{ + if (obj->ttm.created) { + ttm_bo_put(i915_gem_to_ttm(obj)); + } else { + __i915_gem_free_object(obj); + call_rcu(&obj->rcu, __i915_gem_free_object_rcu); + } +} + +static 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, + .truncate = i915_ttm_purge, + .adjust_lru = i915_ttm_adjust_lru, + .delayed_free = i915_ttm_delayed_free, +}; + +void i915_ttm_bo_destroy(struct ttm_buffer_object *bo) +{ + struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); + + i915_gem_object_release_memory_region(obj); + if (obj->ttm.created) + call_rcu(&obj->rcu, __i915_gem_free_object_rcu); +} + +/** + * __i915_gem_ttm_object_init - Initialize a ttm-backed i915 gem object + * @mem: The initial memory region for the object. + * @obj: The gem object. + * @size: Object size in bytes. + * @flags: gem object flags. + * + * Return: 0 on success, negative error code on failure. + */ +int __i915_gem_ttm_object_init(struct intel_memory_region *mem, + struct drm_i915_gem_object *obj, + resource_size_t size, + unsigned int flags) +{ + static struct lock_class_key lock_class; + struct drm_i915_private *i915 = mem->i915; + enum ttm_bo_type bo_type; + size_t alignment = 0; + int ret; + + /* Adjust alignment to GPU- and CPU huge page sizes. */ + + if (mem->is_range_manager) { + if (size >= SZ_1G) + alignment = SZ_1G >> PAGE_SHIFT; + else if (size >= SZ_2M) + alignment = SZ_2M >> PAGE_SHIFT; + else if (size >= SZ_64K) + alignment = SZ_64K >> PAGE_SHIFT; + } + + drm_gem_private_object_init(&i915->drm, &obj->base, size); + 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; + i915_gem_object_set_cache_coherency(obj, I915_CACHE_NONE); + + bo_type = (obj->flags & I915_BO_ALLOC_USER) ? ttm_bo_type_device : + ttm_bo_type_kernel; + + /* + * If this function fails, it will call the destructor, but + * our caller still owns the object. So no freeing in the + * destructor until obj->ttm.created is true. + * Similarly, in delayed_destroy, we can't call ttm_bo_put() + * until successful initialization. + */ + 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); + + if (!ret) + obj->ttm.created = true; + + /* i915 wants -ENXIO when out of memory region space. */ + return (ret == -ENOSPC) ? -ENXIO : ret; +} diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.h b/drivers/gpu/drm/i915/gem/i915_gem_ttm.h new file mode 100644 index 000000000000..b8d3dcbb50df --- /dev/null +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.h @@ -0,0 +1,48 @@ +/* SPDX-License-Identifier: MIT */ +/* + * Copyright © 2021 Intel Corporation + */ +#ifndef _I915_GEM_TTM_H_ +#define _I915_GEM_TTM_H_ + +#include "gem/i915_gem_object_types.h" + +/** + * i915_gem_to_ttm - Convert a struct drm_i915_gem_object to a + * struct ttm_buffer_object. + * @obj: Pointer to the gem object. + * + * Return: Pointer to the embedded struct ttm_buffer_object. + */ +static inline struct ttm_buffer_object * +i915_gem_to_ttm(struct drm_i915_gem_object *obj) +{ + return &obj->__do_not_access; +} + +/* + * i915 ttm gem object destructor. Internal use only. + */ +void i915_ttm_bo_destroy(struct ttm_buffer_object *bo); + +/** + * i915_ttm_to_gem - Convert a struct ttm_buffer_object to an embedding + * struct drm_i915_gem_object. + * + * Return: Pointer to the embedding struct ttm_buffer_object, or NULL + * if the object was not an i915 ttm object. + */ +static inline struct drm_i915_gem_object * +i915_ttm_to_gem(struct ttm_buffer_object *bo) +{ + if (GEM_WARN_ON(bo->destroy != i915_ttm_bo_destroy)) + return NULL; + + return container_of(bo, struct drm_i915_gem_object, __do_not_access); +} + +int __i915_gem_ttm_object_init(struct intel_memory_region *mem, + struct drm_i915_gem_object *obj, + resource_size_t size, + unsigned int flags); +#endif diff --git a/drivers/gpu/drm/i915/gt/intel_region_lmem.c b/drivers/gpu/drm/i915/gt/intel_region_lmem.c index f7366b054f8e..4ae1f717a94c 100644 --- a/drivers/gpu/drm/i915/gt/intel_region_lmem.c +++ b/drivers/gpu/drm/i915/gt/intel_region_lmem.c @@ -9,6 +9,7 @@ #include "intel_region_ttm.h" #include "gem/i915_gem_lmem.h" #include "gem/i915_gem_region.h" +#include "gem/i915_gem_ttm.h" #include "intel_region_lmem.h" static int init_fake_lmem_bar(struct intel_memory_region *mem) @@ -107,7 +108,7 @@ region_lmem_init(struct intel_memory_region *mem) static const struct intel_memory_region_ops intel_region_lmem_ops = { .init = region_lmem_init, .release = region_lmem_release, - .init_object = __i915_gem_lmem_object_init, + .init_object = __i915_gem_ttm_object_init, }; struct intel_memory_region * diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index 0993d706f067..40fe9a147318 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -1005,8 +1005,11 @@ i915_gem_madvise_ioctl(struct drm_device *dev, void *data, } } - if (obj->mm.madv != __I915_MADV_PURGED) + if (obj->mm.madv != __I915_MADV_PURGED) { obj->mm.madv = args->madv; + if (obj->ops->adjust_lru) + obj->ops->adjust_lru(obj); + } if (i915_gem_object_has_pages(obj)) { unsigned long flags; diff --git a/drivers/gpu/drm/i915/intel_memory_region.c b/drivers/gpu/drm/i915/intel_memory_region.c index 4092cc987679..bd27e897d4d0 100644 --- a/drivers/gpu/drm/i915/intel_memory_region.c +++ b/drivers/gpu/drm/i915/intel_memory_region.c @@ -149,7 +149,6 @@ intel_memory_region_create(struct drm_i915_private *i915, mutex_init(&mem->objects.lock); INIT_LIST_HEAD(&mem->objects.list); - INIT_LIST_HEAD(&mem->objects.purgeable); INIT_LIST_HEAD(&mem->reserved); mutex_init(&mem->mm_lock); diff --git a/drivers/gpu/drm/i915/intel_memory_region.h b/drivers/gpu/drm/i915/intel_memory_region.h index e69cde13daf2..7b5fa97c0b59 100644 --- a/drivers/gpu/drm/i915/intel_memory_region.h +++ b/drivers/gpu/drm/i915/intel_memory_region.h @@ -100,7 +100,6 @@ struct intel_memory_region { struct { struct mutex lock; /* Protects access to objects */ struct list_head list; - struct list_head purgeable; } objects; size_t chunk_size; diff --git a/drivers/gpu/drm/i915/intel_region_ttm.c b/drivers/gpu/drm/i915/intel_region_ttm.c index c8ac118c21f6..0b41a1545570 100644 --- a/drivers/gpu/drm/i915/intel_region_ttm.c +++ b/drivers/gpu/drm/i915/intel_region_ttm.c @@ -10,6 +10,7 @@ #include "intel_region_ttm.h" +#include "gem/i915_gem_ttm.h" /* For the funcs/ops export only */ /** * DOC: TTM support structure * @@ -19,9 +20,6 @@ * i915 GEM regions to TTM memory types and resource managers. */ -/* A Zero-initialized driver for now. We don't have a TTM backend yet. */ -static struct ttm_device_funcs i915_ttm_bo_driver; - /** * intel_region_ttm_device_init - Initialize a TTM device * @dev_priv: Pointer to an i915 device private structure. @@ -32,7 +30,7 @@ int intel_region_ttm_device_init(struct drm_i915_private *dev_priv) { struct drm_device *drm = &dev_priv->drm; - return ttm_device_init(&dev_priv->bdev, &i915_ttm_bo_driver, + return ttm_device_init(&dev_priv->bdev, i915_ttm_driver(), drm->dev, drm->anon_inode->i_mapping, drm->vma_offset_manager, false, false); } @@ -172,6 +170,7 @@ struct sg_table *intel_region_ttm_node_to_st(struct intel_memory_region *mem, return i915_sg_from_mm_node(node, mem->region.start); } +#ifdef CONFIG_DRM_I915_SELFTEST /** * intel_region_ttm_node_alloc - Allocate memory resources from a region * @mem: The memory region, @@ -218,3 +217,4 @@ void *intel_region_ttm_node_alloc(struct intel_memory_region *mem, ret = -ENXIO; return ret ? ERR_PTR(ret) : res.mm_node; } +#endif diff --git a/drivers/gpu/drm/i915/intel_region_ttm.h b/drivers/gpu/drm/i915/intel_region_ttm.h index 1c82c6c3429d..eaa3eccfa252 100644 --- a/drivers/gpu/drm/i915/intel_region_ttm.h +++ b/drivers/gpu/drm/i915/intel_region_ttm.h @@ -11,6 +11,7 @@ struct drm_i915_private; struct intel_memory_region; +struct ttm_device_funcs; int intel_region_ttm_device_init(struct drm_i915_private *dev_priv); @@ -23,10 +24,14 @@ void intel_region_ttm_fini(struct intel_memory_region *mem); struct sg_table *intel_region_ttm_node_to_st(struct intel_memory_region *mem, void *node); +void intel_region_ttm_node_free(struct intel_memory_region *mem, + void *node); + +struct ttm_device_funcs *i915_ttm_driver(void); + +#ifdef CONFIG_DRM_I915_SELFTEST void *intel_region_ttm_node_alloc(struct intel_memory_region *mem, resource_size_t size, unsigned int flags); - -void intel_region_ttm_node_free(struct intel_memory_region *mem, - void *node); +#endif #endif /* _INTEL_REGION_TTM_H_ */ -- 2.31.1