drm/amdgpu: add a custom GTT memory manager v2
authorChristian König <christian.koenig@amd.com>
Fri, 9 Sep 2016 14:32:33 +0000 (16:32 +0200)
committerAlex Deucher <alexander.deucher@amd.com>
Wed, 28 Sep 2016 20:16:19 +0000 (16:16 -0400)
Only allocate address space when we really need it.

v2: fix a typo, add correct function description,
    stop leaking the node in the error case.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/Makefile
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c [new file with mode: 0644]
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h

index d15e9b080ce1b362227d5ceea4206a7554c277da..9ec262d4b8a2de493fcc97840775a7bca6bf6e4b 100644 (file)
@@ -23,7 +23,8 @@ amdgpu-y += amdgpu_device.o amdgpu_kms.o \
        amdgpu_pm.o atombios_dp.o amdgpu_afmt.o amdgpu_trace_points.o \
        atombios_encoders.o amdgpu_sa.o atombios_i2c.o \
        amdgpu_prime.o amdgpu_vm.o amdgpu_ib.o amdgpu_pll.o \
-       amdgpu_ucode.o amdgpu_bo_list.o amdgpu_ctx.o amdgpu_sync.o
+       amdgpu_ucode.o amdgpu_bo_list.o amdgpu_ctx.o amdgpu_sync.o \
+       amdgpu_gtt_mgr.o
 
 # add asic specific block
 amdgpu-$(CONFIG_DRM_AMDGPU_CIK)+= cik.o cik_ih.o kv_smc.o kv_dpm.o \
index b8412bcbad2a2355ed13d78ebb0a0078ea75d48d..b0f6e6957536a7827acc979856f3ef5280b3ce79 100644 (file)
@@ -648,7 +648,7 @@ static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
        if (!r && p->uf_entry.robj) {
                struct amdgpu_bo *uf = p->uf_entry.robj;
 
-               r = amdgpu_ttm_bind(uf->tbo.ttm, &uf->tbo.mem);
+               r = amdgpu_ttm_bind(&uf->tbo, &uf->tbo.mem);
                p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
        }
 
@@ -1192,7 +1192,7 @@ int amdgpu_cs_sysvm_access_required(struct amdgpu_cs_parser *parser)
        for (i = 0; i < parser->bo_list->num_entries; i++) {
                struct amdgpu_bo *bo = parser->bo_list->array[i].robj;
 
-               r = amdgpu_ttm_bind(bo->tbo.ttm, &bo->tbo.mem);
+               r = amdgpu_ttm_bind(&bo->tbo, &bo->tbo.mem);
                if (unlikely(r))
                        return r;
        }
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c
new file mode 100644 (file)
index 0000000..262e872
--- /dev/null
@@ -0,0 +1,238 @@
+/*
+ * Copyright 2016 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * Authors: Christian König
+ */
+
+#include <drm/drmP.h>
+#include "amdgpu.h"
+
+struct amdgpu_gtt_mgr {
+       struct drm_mm mm;
+       spinlock_t lock;
+       uint64_t available;
+};
+
+/**
+ * amdgpu_gtt_mgr_init - init GTT manager and DRM MM
+ *
+ * @man: TTM memory type manager
+ * @p_size: maximum size of GTT
+ *
+ * Allocate and initialize the GTT manager.
+ */
+static int amdgpu_gtt_mgr_init(struct ttm_mem_type_manager *man,
+                              unsigned long p_size)
+{
+       struct amdgpu_gtt_mgr *mgr;
+
+       mgr = kzalloc(sizeof(*mgr), GFP_KERNEL);
+       if (!mgr)
+               return -ENOMEM;
+
+       drm_mm_init(&mgr->mm, 0, p_size);
+       spin_lock_init(&mgr->lock);
+       mgr->available = p_size;
+       man->priv = mgr;
+       return 0;
+}
+
+/**
+ * amdgpu_gtt_mgr_fini - free and destroy GTT manager
+ *
+ * @man: TTM memory type manager
+ *
+ * Destroy and free the GTT manager, returns -EBUSY if ranges are still
+ * allocated inside it.
+ */
+static int amdgpu_gtt_mgr_fini(struct ttm_mem_type_manager *man)
+{
+       struct amdgpu_gtt_mgr *mgr = man->priv;
+
+       spin_lock(&mgr->lock);
+       if (!drm_mm_clean(&mgr->mm)) {
+               spin_unlock(&mgr->lock);
+               return -EBUSY;
+       }
+
+       drm_mm_takedown(&mgr->mm);
+       spin_unlock(&mgr->lock);
+       kfree(mgr);
+       man->priv = NULL;
+       return 0;
+}
+
+/**
+ * amdgpu_gtt_mgr_alloc - allocate new ranges
+ *
+ * @man: TTM memory type manager
+ * @tbo: TTM BO we need this range for
+ * @place: placement flags and restrictions
+ * @mem: the resulting mem object
+ *
+ * Allocate the address space for a node.
+ */
+int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
+                        struct ttm_buffer_object *tbo,
+                        const struct ttm_place *place,
+                        struct ttm_mem_reg *mem)
+{
+       struct amdgpu_gtt_mgr *mgr = man->priv;
+       struct drm_mm_node *node = mem->mm_node;
+       enum drm_mm_search_flags sflags = DRM_MM_SEARCH_BEST;
+       enum drm_mm_allocator_flags aflags = DRM_MM_CREATE_DEFAULT;
+       unsigned long fpfn, lpfn;
+       int r;
+
+       if (node->start != AMDGPU_BO_INVALID_OFFSET)
+               return 0;
+
+       if (place)
+               fpfn = place->fpfn;
+       else
+               fpfn = 0;
+
+       if (place && place->lpfn)
+               lpfn = place->lpfn;
+       else
+               lpfn = man->size;
+
+       if (place && place->flags & TTM_PL_FLAG_TOPDOWN) {
+               sflags = DRM_MM_SEARCH_BELOW;
+               aflags = DRM_MM_CREATE_TOP;
+       }
+
+       spin_lock(&mgr->lock);
+       r = drm_mm_insert_node_in_range_generic(&mgr->mm, node, mem->num_pages,
+                                               mem->page_alignment, 0,
+                                               fpfn, lpfn, sflags, aflags);
+       spin_unlock(&mgr->lock);
+
+       if (!r) {
+               mem->start = node->start;
+               tbo->offset = (tbo->mem.start << PAGE_SHIFT) +
+                   tbo->bdev->man[tbo->mem.mem_type].gpu_offset;
+       }
+
+       return r;
+}
+
+/**
+ * amdgpu_gtt_mgr_new - allocate a new node
+ *
+ * @man: TTM memory type manager
+ * @tbo: TTM BO we need this range for
+ * @place: placement flags and restrictions
+ * @mem: the resulting mem object
+ *
+ * Dummy, allocate the node but no space for it yet.
+ */
+static int amdgpu_gtt_mgr_new(struct ttm_mem_type_manager *man,
+                             struct ttm_buffer_object *tbo,
+                             const struct ttm_place *place,
+                             struct ttm_mem_reg *mem)
+{
+       struct amdgpu_gtt_mgr *mgr = man->priv;
+       struct drm_mm_node *node;
+       int r;
+
+       spin_lock(&mgr->lock);
+       if (mgr->available < mem->num_pages) {
+               spin_unlock(&mgr->lock);
+               return 0;
+       }
+       mgr->available -= mem->num_pages;
+       spin_unlock(&mgr->lock);
+
+       node = kzalloc(sizeof(*node), GFP_KERNEL);
+       if (!node)
+               return -ENOMEM;
+
+       node->start = AMDGPU_BO_INVALID_OFFSET;
+       mem->mm_node = node;
+
+       if (place->fpfn || place->lpfn || place->flags & TTM_PL_FLAG_TOPDOWN) {
+               r = amdgpu_gtt_mgr_alloc(man, tbo, place, mem);
+               if (unlikely(r)) {
+                       kfree(node);
+                       mem->mm_node = NULL;
+               }
+       } else {
+               mem->start = node->start;
+       }
+
+       return 0;
+}
+
+/**
+ * amdgpu_gtt_mgr_del - free ranges
+ *
+ * @man: TTM memory type manager
+ * @tbo: TTM BO we need this range for
+ * @place: placement flags and restrictions
+ * @mem: TTM memory object
+ *
+ * Free the allocated GTT again.
+ */
+static void amdgpu_gtt_mgr_del(struct ttm_mem_type_manager *man,
+                              struct ttm_mem_reg *mem)
+{
+       struct amdgpu_gtt_mgr *mgr = man->priv;
+       struct drm_mm_node *node = mem->mm_node;
+
+       if (!node)
+               return;
+
+       spin_lock(&mgr->lock);
+       if (node->start != AMDGPU_BO_INVALID_OFFSET)
+               drm_mm_remove_node(node);
+       mgr->available += mem->num_pages;
+       spin_unlock(&mgr->lock);
+
+       kfree(node);
+       mem->mm_node = NULL;
+}
+
+/**
+ * amdgpu_gtt_mgr_debug - dump VRAM table
+ *
+ * @man: TTM memory type manager
+ * @prefix: text prefix
+ *
+ * Dump the table content using printk.
+ */
+static void amdgpu_gtt_mgr_debug(struct ttm_mem_type_manager *man,
+                                 const char *prefix)
+{
+       struct amdgpu_gtt_mgr *mgr = man->priv;
+
+       spin_lock(&mgr->lock);
+       drm_mm_debug_table(&mgr->mm, prefix);
+       spin_unlock(&mgr->lock);
+}
+
+const struct ttm_mem_type_manager_func amdgpu_gtt_mgr_func = {
+       amdgpu_gtt_mgr_init,
+       amdgpu_gtt_mgr_fini,
+       amdgpu_gtt_mgr_new,
+       amdgpu_gtt_mgr_del,
+       amdgpu_gtt_mgr_debug
+};
index 428aa00025e4875637afaa51b1b1e6ebba2799b8..9b80dfedcb4a5227b6d42fbc16fb91270417fff9 100644 (file)
@@ -673,7 +673,7 @@ int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
                dev_err(bo->adev->dev, "%p pin failed\n", bo);
                goto error;
        }
-       r = amdgpu_ttm_bind(bo->tbo.ttm, &bo->tbo.mem);
+       r = amdgpu_ttm_bind(&bo->tbo, &bo->tbo.mem);
        if (unlikely(r)) {
                dev_err(bo->adev->dev, "%p bind failed\n", bo);
                goto error;
index dfb12237a6b08b8e061fb991106891948fe449b6..e21e823f67a56c8be5ac8a8d57f57540e0716c85 100644 (file)
@@ -160,7 +160,7 @@ static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
                man->default_caching = TTM_PL_FLAG_CACHED;
                break;
        case TTM_PL_TT:
-               man->func = &ttm_bo_manager_func;
+               man->func = &amdgpu_gtt_mgr_func;
                man->gpu_offset = adev->mc.gtt_start;
                man->available_caching = TTM_PL_MASK_CACHING;
                man->default_caching = TTM_PL_FLAG_CACHED;
@@ -277,7 +277,7 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 
        switch (old_mem->mem_type) {
        case TTM_PL_TT:
-               r = amdgpu_ttm_bind(bo->ttm, old_mem);
+               r = amdgpu_ttm_bind(bo, old_mem);
                if (r)
                        return r;
 
@@ -290,7 +290,7 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
        }
        switch (new_mem->mem_type) {
        case TTM_PL_TT:
-               r = amdgpu_ttm_bind(bo->ttm, new_mem);
+               r = amdgpu_ttm_bind(bo, new_mem);
                if (r)
                        return r;
 
@@ -675,7 +675,6 @@ static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
                        return r;
                }
        }
-       gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
        if (!ttm->num_pages) {
                WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
                     ttm->num_pages, bo_mem, ttm);
@@ -696,16 +695,25 @@ bool amdgpu_ttm_is_bound(struct ttm_tt *ttm)
        return gtt && !list_empty(&gtt->list);
 }
 
-int amdgpu_ttm_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem)
+int amdgpu_ttm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *bo_mem)
 {
-       struct amdgpu_ttm_tt *gtt = (void *)ttm;
+       struct ttm_tt *ttm = bo->ttm;
+       struct amdgpu_ttm_tt *gtt = (void *)bo->ttm;
        uint32_t flags;
        int r;
 
        if (!ttm || amdgpu_ttm_is_bound(ttm))
                return 0;
 
+       r = amdgpu_gtt_mgr_alloc(&bo->bdev->man[TTM_PL_TT], bo,
+                                NULL, bo_mem);
+       if (r) {
+               DRM_ERROR("Failed to allocate GTT address space (%d)\n", r);
+               return r;
+       }
+
        flags = amdgpu_ttm_tt_pte_flags(gtt->adev, ttm, bo_mem);
+       gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
        r = amdgpu_gart_bind(gtt->adev, gtt->offset, ttm->num_pages,
                ttm->pages, gtt->ttm.dma_address, flags);
 
index 3ee825f4de28f8e983add6ce62e6b5f11b7fbeff..9812c805326c383cc02bdbc537f056e2d98bd40c 100644 (file)
@@ -65,6 +65,13 @@ struct amdgpu_mman {
        struct amdgpu_mman_lru                  guard;
 };
 
+extern const struct ttm_mem_type_manager_func amdgpu_gtt_mgr_func;
+
+int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
+                        struct ttm_buffer_object *tbo,
+                        const struct ttm_place *place,
+                        struct ttm_mem_reg *mem);
+
 int amdgpu_copy_buffer(struct amdgpu_ring *ring,
                       uint64_t src_offset,
                       uint64_t dst_offset,
@@ -78,6 +85,6 @@ int amdgpu_fill_buffer(struct amdgpu_bo *bo,
 
 int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma);
 bool amdgpu_ttm_is_bound(struct ttm_tt *ttm);
-int amdgpu_ttm_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem);
+int amdgpu_ttm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *bo_mem);
 
 #endif