[PATCH 3/3] drm/amdgpu: Replace drm_mm with drm buddy manager

Add drm buddy allocator support for vram memory management

Signed-off-by: Arunpravin Arunpravin.PaneerSelvam@amd.com --- .../gpu/drm/amd/amdgpu/amdgpu_res_cursor.h | 97 +++++-- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h | 4 +- drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c | 251 ++++++++++-------- 3 files changed, 217 insertions(+), 135 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_res_cursor.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_res_cursor.h index acfa207cf970..2c17e948355e 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_res_cursor.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_res_cursor.h @@ -30,12 +30,15 @@ #include <drm/ttm/ttm_resource.h> #include <drm/ttm/ttm_range_manager.h>

+#include "amdgpu_vram_mgr.h" + /* state back for walking over vram_mgr and gtt_mgr allocations */ struct amdgpu_res_cursor { uint64_t start; uint64_t size; uint64_t remaining; - struct drm_mm_node *node; + void *node; + uint32_t mem_type; };

/** @@ -52,27 +55,63 @@ static inline void amdgpu_res_first(struct ttm_resource *res, uint64_t start, uint64_t size, struct amdgpu_res_cursor *cur) { + struct drm_buddy_block *block; + struct list_head *head, *next; struct drm_mm_node *node;

- if (!res || res->mem_type == TTM_PL_SYSTEM) { - cur->start = start; - cur->size = size; - cur->remaining = size; - cur->node = NULL; - WARN_ON(res && start + size > res->num_pages << PAGE_SHIFT); - return; - } + if (!res) + goto err_out;

BUG_ON(start + size > res->num_pages << PAGE_SHIFT);

- node = to_ttm_range_mgr_node(res)->mm_nodes; - while (start >= node->size << PAGE_SHIFT) - start -= node++->size << PAGE_SHIFT; + cur->mem_type = res->mem_type; + + switch (cur->mem_type) { + case TTM_PL_VRAM: + head = &to_amdgpu_vram_mgr_node(res)->blocks; + + block = list_first_entry_or_null(head, + struct drm_buddy_block, + link); + if (!block) + goto err_out; + + while (start >= node_size(block)) { + start -= node_size(block); + + next = block->link.next; + if (next != head) + block = list_entry(next, struct drm_buddy_block, link); + } + + cur->start = node_start(block) + start; + cur->size = min(node_size(block) - start, size); + cur->remaining = size; + cur->node = block; + break; + case TTM_PL_TT: + node = to_ttm_range_mgr_node(res)->mm_nodes; + while (start >= node->size << PAGE_SHIFT) + start -= node++->size << PAGE_SHIFT; + + cur->start = (node->start << PAGE_SHIFT) + start; + cur->size = min((node->size << PAGE_SHIFT) - start, size); + cur->remaining = size; + cur->node = node; + break; + default: + goto err_out; + }

- cur->start = (node->start << PAGE_SHIFT) + start; - cur->size = min((node->size << PAGE_SHIFT) - start, size); + return; + +err_out: + cur->start = start; + cur->size = size; cur->remaining = size; - cur->node = node; + cur->node = NULL; + WARN_ON(res && start + size > res->num_pages << PAGE_SHIFT); + return; }

/** @@ -85,7 +124,9 @@ static inline void amdgpu_res_first(struct ttm_resource *res, */ static inline void amdgpu_res_next(struct amdgpu_res_cursor *cur, uint64_t size) { - struct drm_mm_node *node = cur->node; + struct drm_buddy_block *block; + struct drm_mm_node *node; + struct list_head *next;

BUG_ON(size > cur->remaining);

@@ -99,9 +140,27 @@ static inline void amdgpu_res_next(struct amdgpu_res_cursor *cur, uint64_t size) return; }

- cur->node = ++node; - cur->start = node->start << PAGE_SHIFT; - cur->size = min(node->size << PAGE_SHIFT, cur->remaining); + switch (cur->mem_type) { + case TTM_PL_VRAM: + block = cur->node; + + next = block->link.next; + block = list_entry(next, struct drm_buddy_block, link); + + cur->node = block; + cur->start = node_start(block); + cur->size = min(node_size(block), cur->remaining); + break; + case TTM_PL_TT: + node = cur->node; + + cur->node = ++node; + cur->start = node->start << PAGE_SHIFT; + cur->size = min(node->size << PAGE_SHIFT, cur->remaining); + break; + default: + return; + } }

#endif diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h index ba5c864b8de1..a47a1aa5868f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h @@ -26,6 +26,7 @@

#include <linux/dma-direction.h> #include <drm/gpu_scheduler.h> +#include <drm/drm_buddy.h> #include "amdgpu.h"

#define AMDGPU_PL_GDS (TTM_PL_PRIV + 0) @@ -40,12 +41,13 @@

struct amdgpu_vram_mgr { struct ttm_resource_manager manager; - struct drm_mm mm; + struct drm_buddy_mm mm; spinlock_t lock; struct list_head reservations_pending; struct list_head reserved_pages; atomic64_t usage; atomic64_t vis_usage; + uint64_t default_page_size; };

struct amdgpu_gtt_mgr { diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c index 7b2b0980ec41..0c55a5ea1ed1 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c @@ -29,25 +29,9 @@ #include "amdgpu_vm.h" #include "amdgpu_res_cursor.h" #include "amdgpu_atomfirmware.h" +#include "amdgpu_vram_mgr.h" #include "atom.h"

-struct amdgpu_vram_reservation { - struct list_head node; - struct drm_mm_node mm_node; -}; - -static inline struct amdgpu_vram_mgr * -to_vram_mgr(struct ttm_resource_manager *man) -{ - return container_of(man, struct amdgpu_vram_mgr, manager); -} - -static inline struct amdgpu_device * -to_amdgpu_device(struct amdgpu_vram_mgr *mgr) -{ - return container_of(mgr, struct amdgpu_device, mman.vram_mgr); -} - /** * DOC: mem_info_vram_total * @@ -196,10 +180,10 @@ const struct attribute_group amdgpu_vram_mgr_attr_group = { * Calculate how many bytes of the MM node are inside visible VRAM */ static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev, - struct drm_mm_node *node) + struct drm_buddy_block *block) { - uint64_t start = node->start << PAGE_SHIFT; - uint64_t end = (node->size + node->start) << PAGE_SHIFT; + uint64_t start = node_start(block); + uint64_t end = start + node_size(block);

if (start >= adev->gmc.visible_vram_size) return 0; @@ -220,9 +204,9 @@ u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo) { struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); struct ttm_resource *res = bo->tbo.resource; - unsigned pages = res->num_pages; - struct drm_mm_node *mm; - u64 usage; + struct amdgpu_vram_mgr_node *node = to_amdgpu_vram_mgr_node(res); + struct drm_buddy_block *block; + u64 usage = 0;

if (amdgpu_gmc_vram_full_visible(&adev->gmc)) return amdgpu_bo_size(bo); @@ -230,9 +214,8 @@ u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo) if (res->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT) return 0;

- mm = &container_of(res, struct ttm_range_mgr_node, base)->mm_nodes[0]; - for (usage = 0; pages; pages -= mm->size, mm++) - usage += amdgpu_vram_mgr_vis_size(adev, mm); + list_for_each_entry(block, &node->blocks, link) + usage += amdgpu_vram_mgr_vis_size(adev, block);

return usage; } @@ -242,21 +225,30 @@ static void amdgpu_vram_mgr_do_reserve(struct ttm_resource_manager *man) { struct amdgpu_vram_mgr *mgr = to_vram_mgr(man); struct amdgpu_device *adev = to_amdgpu_device(mgr); - struct drm_mm *mm = &mgr->mm; + struct drm_buddy_mm *mm = &mgr->mm; struct amdgpu_vram_reservation *rsv, *temp; + struct drm_buddy_block *block; uint64_t vis_usage;

list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, node) { - if (drm_mm_reserve_node(mm, &rsv->mm_node)) + if (drm_buddy_alloc(mm, rsv->start, rsv->start + rsv->size, + rsv->size, rsv->min_size, &rsv->block, + rsv->flags)) continue;

- dev_dbg(adev->dev, "Reservation 0x%llx - %lld, Succeeded\n", - rsv->mm_node.start, rsv->mm_node.size); + block = list_first_entry_or_null(&rsv->block, + struct drm_buddy_block, + link); + + if (block) { + dev_dbg(adev->dev, "Reservation 0x%llx - %lld, Succeeded\n", + rsv->start, rsv->size);

- vis_usage = amdgpu_vram_mgr_vis_size(adev, &rsv->mm_node); - atomic64_add(vis_usage, &mgr->vis_usage); - atomic64_add(rsv->mm_node.size << PAGE_SHIFT, &mgr->usage); - list_move(&rsv->node, &mgr->reserved_pages); + vis_usage = amdgpu_vram_mgr_vis_size(adev, block); + atomic64_add(vis_usage, &mgr->vis_usage); + atomic64_add(rsv->size, &mgr->usage); + list_move(&rsv->node, &mgr->reserved_pages); + } } }

@@ -280,11 +272,15 @@ int amdgpu_vram_mgr_reserve_range(struct ttm_resource_manager *man, return -ENOMEM;

INIT_LIST_HEAD(&rsv->node); - rsv->mm_node.start = start >> PAGE_SHIFT; - rsv->mm_node.size = size >> PAGE_SHIFT; + INIT_LIST_HEAD(&rsv->block); + + rsv->start = start; + rsv->size = size; + rsv->min_size = size; + rsv->flags |= DRM_BUDDY_RANGE_ALLOCATION;

spin_lock(&mgr->lock); - list_add_tail(&mgr->reservations_pending, &rsv->node); + list_add_tail(&rsv->node, &mgr->reservations_pending); amdgpu_vram_mgr_do_reserve(man); spin_unlock(&mgr->lock);

@@ -312,16 +308,16 @@ int amdgpu_vram_mgr_query_page_status(struct ttm_resource_manager *man, spin_lock(&mgr->lock);

list_for_each_entry(rsv, &mgr->reservations_pending, node) { - if ((rsv->mm_node.start <= start) && - (start < (rsv->mm_node.start + rsv->mm_node.size))) { + if ((rsv->start <= start) && + (start < (rsv->start + rsv->size))) { ret = -EBUSY; goto out; } }

list_for_each_entry(rsv, &mgr->reserved_pages, node) { - if ((rsv->mm_node.start <= start) && - (start < (rsv->mm_node.start + rsv->mm_node.size))) { + if ((rsv->start <= start) && + (start < (rsv->start + rsv->size))) { ret = 0; goto out; } @@ -333,28 +329,6 @@ int amdgpu_vram_mgr_query_page_status(struct ttm_resource_manager *man, return ret; }

-/** - * amdgpu_vram_mgr_virt_start - update virtual start address - * - * @mem: ttm_resource to update - * @node: just allocated node - * - * Calculate a virtual BO start address to easily check if everything is CPU - * accessible. - */ -static void amdgpu_vram_mgr_virt_start(struct ttm_resource *mem, - struct drm_mm_node *node) -{ - unsigned long start; - - start = node->start + node->size; - if (start > mem->num_pages) - start -= mem->num_pages; - else - start = 0; - mem->start = max(mem->start, start); -} - /** * amdgpu_vram_mgr_new - allocate new ranges * @@ -370,13 +344,13 @@ static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man, const struct ttm_place *place, struct ttm_resource **res) { - unsigned long lpfn, num_nodes, pages_per_node, pages_left, pages; + unsigned long lpfn, pages_per_node, pages_left, pages, n_pages; + uint64_t vis_usage = 0, mem_bytes, max_bytes, min_page_size; struct amdgpu_vram_mgr *mgr = to_vram_mgr(man); struct amdgpu_device *adev = to_amdgpu_device(mgr); - uint64_t vis_usage = 0, mem_bytes, max_bytes; - struct ttm_range_mgr_node *node; - struct drm_mm *mm = &mgr->mm; - enum drm_mm_insert_mode mode; + struct amdgpu_vram_mgr_node *node; + struct drm_buddy_mm *mm = &mgr->mm; + struct drm_buddy_block *block; unsigned i; int r;

@@ -395,10 +369,9 @@ static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man, goto error_sub; }

- if (place->flags & TTM_PL_FLAG_CONTIGUOUS) { + if (place->flags & TTM_PL_FLAG_CONTIGUOUS) pages_per_node = ~0ul; - num_nodes = 1; - } else { + else { #ifdef CONFIG_TRANSPARENT_HUGEPAGE pages_per_node = HPAGE_PMD_NR; #else @@ -407,11 +380,10 @@ static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man, #endif pages_per_node = max_t(uint32_t, pages_per_node, tbo->page_alignment); - num_nodes = DIV_ROUND_UP_ULL(PFN_UP(mem_bytes), pages_per_node); }

- node = kvmalloc(struct_size(node, mm_nodes, num_nodes), - GFP_KERNEL | __GFP_ZERO); + node = kzalloc(sizeof(*node), GFP_KERNEL); + if (!node) { r = -ENOMEM; goto error_sub; @@ -419,9 +391,17 @@ static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,

ttm_resource_init(tbo, place, &node->base);

- mode = DRM_MM_INSERT_BEST; + INIT_LIST_HEAD(&node->blocks); + if (place->flags & TTM_PL_FLAG_TOPDOWN) - mode = DRM_MM_INSERT_HIGH; + node->flags |= DRM_BUDDY_TOPDOWN_ALLOCATION; + + if (place->fpfn || lpfn != man->size) + /* Allocate blocks in desired range */ + node->flags |= DRM_BUDDY_RANGE_ALLOCATION; + + min_page_size = mgr->default_page_size; + BUG_ON(min_page_size < mm->chunk_size);

pages_left = node->base.num_pages;

@@ -429,36 +409,63 @@ static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man, pages = min(pages_left, 2UL << (30 - PAGE_SHIFT));

i = 0; - spin_lock(&mgr->lock); while (pages_left) { - uint32_t alignment = tbo->page_alignment; - if (pages >= pages_per_node) - alignment = pages_per_node; - - r = drm_mm_insert_node_in_range(mm, &node->mm_nodes[i], pages, - alignment, 0, place->fpfn, - lpfn, mode); - if (unlikely(r)) { - if (pages > pages_per_node) { - if (is_power_of_2(pages)) - pages = pages / 2; - else - pages = rounddown_pow_of_two(pages); - continue; - } - goto error_free; + pages = pages_per_node; + + n_pages = pages; + + if (place->flags & TTM_PL_FLAG_CONTIGUOUS) { + n_pages = roundup_pow_of_two(n_pages); + min_page_size = (uint64_t)n_pages << PAGE_SHIFT; + + if (n_pages > lpfn) + lpfn = n_pages; }

- vis_usage += amdgpu_vram_mgr_vis_size(adev, &node->mm_nodes[i]); - amdgpu_vram_mgr_virt_start(&node->base, &node->mm_nodes[i]); + spin_lock(&mgr->lock); + r = drm_buddy_alloc(mm, (uint64_t)place->fpfn << PAGE_SHIFT, + (uint64_t)lpfn << PAGE_SHIFT, + (uint64_t)n_pages << PAGE_SHIFT, + min_page_size, &node->blocks, + node->flags); + spin_unlock(&mgr->lock); + + if (unlikely(r)) + goto error_free_blocks; + pages_left -= pages; ++i;

if (pages > pages_left) pages = pages_left; } - spin_unlock(&mgr->lock); + + /* Free unused pages for contiguous allocation */ + if (place->flags & TTM_PL_FLAG_CONTIGUOUS) { + uint64_t actual_size = (uint64_t)node->base.num_pages << PAGE_SHIFT; + + r = drm_buddy_free_unused_pages(mm, + actual_size, + &node->blocks); + + if (unlikely(r)) + goto error_free_blocks; + } + + list_for_each_entry(block, &node->blocks, link) + vis_usage += amdgpu_vram_mgr_vis_size(adev, block); + + block = list_first_entry_or_null(&node->blocks, + struct drm_buddy_block, + link); + + if (!block) { + r = -ENOSPC; + goto error_free_res; + } + + node->base.start = node_start(block) >> PAGE_SHIFT;

if (i == 1) node->base.placement |= TTM_PL_FLAG_CONTIGUOUS; @@ -472,12 +479,12 @@ static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man, *res = &node->base; return 0;

-error_free: - while (i--) - drm_mm_remove_node(&node->mm_nodes[i]); +error_free_blocks: + spin_lock(&mgr->lock); + drm_buddy_free_list(mm, &node->blocks); spin_unlock(&mgr->lock); - kvfree(node); - +error_free_res: + kfree(node); error_sub: atomic64_sub(mem_bytes, &mgr->usage); return r; @@ -494,28 +501,28 @@ static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man, static void amdgpu_vram_mgr_del(struct ttm_resource_manager *man, struct ttm_resource *res) { - struct ttm_range_mgr_node *node = to_ttm_range_mgr_node(res); + struct amdgpu_vram_mgr_node *node = to_amdgpu_vram_mgr_node(res); struct amdgpu_vram_mgr *mgr = to_vram_mgr(man); struct amdgpu_device *adev = to_amdgpu_device(mgr); + struct drm_buddy_mm *mm = &mgr->mm; + struct drm_buddy_block *block; uint64_t usage = 0, vis_usage = 0; - unsigned i, pages;

spin_lock(&mgr->lock); - for (i = 0, pages = res->num_pages; pages; - pages -= node->mm_nodes[i].size, ++i) { - struct drm_mm_node *mm = &node->mm_nodes[i]; - - drm_mm_remove_node(mm); - usage += mm->size << PAGE_SHIFT; - vis_usage += amdgpu_vram_mgr_vis_size(adev, mm); + list_for_each_entry(block, &node->blocks, link) { + usage += node_size(block); + vis_usage += amdgpu_vram_mgr_vis_size(adev, block); } + amdgpu_vram_mgr_do_reserve(man); + + drm_buddy_free_list(mm, &node->blocks); spin_unlock(&mgr->lock);

atomic64_sub(usage, &mgr->usage); atomic64_sub(vis_usage, &mgr->vis_usage);

- kvfree(node); + kfree(node); }

/** @@ -669,9 +676,18 @@ static void amdgpu_vram_mgr_debug(struct ttm_resource_manager *man, struct drm_printer *printer) { struct amdgpu_vram_mgr *mgr = to_vram_mgr(man); + struct drm_buddy_mm *mm = &mgr->mm; + struct drm_buddy_block *block;

spin_lock(&mgr->lock); - drm_mm_print(&mgr->mm, printer); + drm_printf(printer, "default_page_size: %lluKiB\n", + mgr->default_page_size >> 10); + + drm_buddy_print(mm, printer); + + drm_printf(printer, "reserved:\n"); + list_for_each_entry(block, &mgr->reserved_pages, link) + drm_buddy_block_print(mm, block, printer); spin_unlock(&mgr->lock);

drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n", @@ -696,15 +712,20 @@ int amdgpu_vram_mgr_init(struct amdgpu_device *adev) { struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr; struct ttm_resource_manager *man = &mgr->manager; + int err;

ttm_resource_manager_init(man, adev->gmc.real_vram_size >> PAGE_SHIFT);

man->func = &amdgpu_vram_mgr_func;

- drm_mm_init(&mgr->mm, 0, man->size); + err = drm_buddy_init(&mgr->mm, man->size << PAGE_SHIFT, PAGE_SIZE); + if (err) + return err; + spin_lock_init(&mgr->lock); INIT_LIST_HEAD(&mgr->reservations_pending); INIT_LIST_HEAD(&mgr->reserved_pages); + mgr->default_page_size = PAGE_SIZE;

ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, &mgr->manager); ttm_resource_manager_set_used(man, true); @@ -737,10 +758,10 @@ void amdgpu_vram_mgr_fini(struct amdgpu_device *adev) kfree(rsv);

list_for_each_entry_safe(rsv, temp, &mgr->reserved_pages, node) { - drm_mm_remove_node(&rsv->mm_node); + drm_buddy_free_list(&mgr->mm, &rsv->block); kfree(rsv); } - drm_mm_takedown(&mgr->mm); + drm_buddy_fini(&mgr->mm); spin_unlock(&mgr->lock);

ttm_resource_manager_cleanup(man);