drm/ttm: restructure to allow driver to plug in alternate memory manager
[pandora-kernel.git] / drivers / gpu / drm / nouveau / nouveau_bo.c
index f6f4477..80353e2 100644 (file)
 #include <linux/log2.h>
 #include <linux/slab.h>
 
-int
-nouveau_bo_sync_gpu(struct nouveau_bo *nvbo, struct nouveau_channel *chan)
-{
-       struct nouveau_fence *prev_fence = nvbo->bo.sync_obj;
-       int ret;
-
-       if (!prev_fence || nouveau_fence_channel(prev_fence) == chan)
-               return 0;
-
-       spin_lock(&nvbo->bo.lock);
-       ret = ttm_bo_wait(&nvbo->bo, false, false, false);
-       spin_unlock(&nvbo->bo.lock);
-       return ret;
-}
-
 static void
 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
 {
@@ -58,8 +43,6 @@ nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
        struct drm_device *dev = dev_priv->dev;
        struct nouveau_bo *nvbo = nouveau_bo(bo);
 
-       ttm_bo_kunmap(&nvbo->kmap);
-
        if (unlikely(nvbo->gem))
                DRM_ERROR("bo %p still attached to GEM object\n", bo);
 
@@ -164,8 +147,6 @@ nouveau_bo_new(struct drm_device *dev, struct nouveau_channel *chan,
        nouveau_bo_fixup_align(dev, tile_mode, tile_flags, &align, &size);
        align >>= PAGE_SHIFT;
 
-       nvbo->placement.fpfn = 0;
-       nvbo->placement.lpfn = mappable ? dev_priv->fb_mappable_pages : 0;
        nouveau_bo_placement_set(nvbo, flags, 0);
 
        nvbo->channel = chan;
@@ -305,7 +286,8 @@ nouveau_bo_map(struct nouveau_bo *nvbo)
 void
 nouveau_bo_unmap(struct nouveau_bo *nvbo)
 {
-       ttm_bo_kunmap(&nvbo->kmap);
+       if (nvbo)
+               ttm_bo_kunmap(&nvbo->kmap);
 }
 
 u16
@@ -399,14 +381,19 @@ nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
                man->default_caching = TTM_PL_FLAG_CACHED;
                break;
        case TTM_PL_VRAM:
+               man->func = &ttm_bo_manager_func;
                man->flags = TTM_MEMTYPE_FLAG_FIXED |
                             TTM_MEMTYPE_FLAG_MAPPABLE;
                man->available_caching = TTM_PL_FLAG_UNCACHED |
                                         TTM_PL_FLAG_WC;
                man->default_caching = TTM_PL_FLAG_WC;
-               man->gpu_offset = dev_priv->vm_vram_base;
+               if (dev_priv->card_type == NV_50)
+                       man->gpu_offset = 0x40000000;
+               else
+                       man->gpu_offset = 0;
                break;
        case TTM_PL_TT:
+               man->func = &ttm_bo_manager_func;
                switch (dev_priv->gart_info.type) {
                case NOUVEAU_GART_AGP:
                        man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
@@ -469,19 +456,26 @@ nouveau_bo_move_accel_cleanup(struct nouveau_channel *chan,
        if (ret)
                return ret;
 
-       ret = ttm_bo_move_accel_cleanup(&nvbo->bo, fence, NULL,
-                                       evict || (nvbo->channel &&
-                                                 nvbo->channel != chan),
+       if (nvbo->channel) {
+               ret = nouveau_fence_sync(fence, nvbo->channel);
+               if (ret)
+                       goto out;
+       }
+
+       ret = ttm_bo_move_accel_cleanup(&nvbo->bo, fence, NULL, evict,
                                        no_wait_reserve, no_wait_gpu, new_mem);
+out:
        nouveau_fence_unref((void *)&fence);
        return ret;
 }
 
 static inline uint32_t
-nouveau_bo_mem_ctxdma(struct nouveau_bo *nvbo, struct nouveau_channel *chan,
-                     struct ttm_mem_reg *mem)
+nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
+                     struct nouveau_channel *chan, struct ttm_mem_reg *mem)
 {
-       if (chan == nouveau_bdev(nvbo->bo.bdev)->channel) {
+       struct nouveau_bo *nvbo = nouveau_bo(bo);
+
+       if (nvbo->no_vm) {
                if (mem->mem_type == TTM_PL_TT)
                        return NvDmaGART;
                return NvDmaVRAM;
@@ -493,86 +487,181 @@ nouveau_bo_mem_ctxdma(struct nouveau_bo *nvbo, struct nouveau_channel *chan,
 }
 
 static int
-nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
-                    bool no_wait_reserve, bool no_wait_gpu,
-                    struct ttm_mem_reg *new_mem)
+nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
+                 struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
 {
-       struct nouveau_bo *nvbo = nouveau_bo(bo);
        struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
-       struct ttm_mem_reg *old_mem = &bo->mem;
-       struct nouveau_channel *chan;
-       uint64_t src_offset, dst_offset;
-       uint32_t page_count;
+       struct nouveau_bo *nvbo = nouveau_bo(bo);
+       u64 length = (new_mem->num_pages << PAGE_SHIFT);
+       u64 src_offset, dst_offset;
        int ret;
 
-       chan = nvbo->channel;
-       if (!chan || nvbo->tile_flags || nvbo->no_vm)
-               chan = dev_priv->channel;
-
-       src_offset = old_mem->mm_node->start << PAGE_SHIFT;
-       dst_offset = new_mem->mm_node->start << PAGE_SHIFT;
-       if (chan != dev_priv->channel) {
-               if (old_mem->mem_type == TTM_PL_TT)
-                       src_offset += dev_priv->vm_gart_base;
-               else
+       src_offset = old_mem->start << PAGE_SHIFT;
+       dst_offset = new_mem->start << PAGE_SHIFT;
+       if (!nvbo->no_vm) {
+               if (old_mem->mem_type == TTM_PL_VRAM)
                        src_offset += dev_priv->vm_vram_base;
-
-               if (new_mem->mem_type == TTM_PL_TT)
-                       dst_offset += dev_priv->vm_gart_base;
                else
+                       src_offset += dev_priv->vm_gart_base;
+
+               if (new_mem->mem_type == TTM_PL_VRAM)
                        dst_offset += dev_priv->vm_vram_base;
+               else
+                       dst_offset += dev_priv->vm_gart_base;
        }
 
        ret = RING_SPACE(chan, 3);
        if (ret)
                return ret;
-       BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
-       OUT_RING(chan, nouveau_bo_mem_ctxdma(nvbo, chan, old_mem));
-       OUT_RING(chan, nouveau_bo_mem_ctxdma(nvbo, chan, new_mem));
 
-       if (dev_priv->card_type >= NV_50) {
-               ret = RING_SPACE(chan, 4);
+       BEGIN_RING(chan, NvSubM2MF, 0x0184, 2);
+       OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
+       OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
+
+       while (length) {
+               u32 amount, stride, height;
+
+               amount  = min(length, (u64)(4 * 1024 * 1024));
+               stride  = 16 * 4;
+               height  = amount / stride;
+
+               if (new_mem->mem_type == TTM_PL_VRAM && nvbo->tile_flags) {
+                       ret = RING_SPACE(chan, 8);
+                       if (ret)
+                               return ret;
+
+                       BEGIN_RING(chan, NvSubM2MF, 0x0200, 7);
+                       OUT_RING  (chan, 0);
+                       OUT_RING  (chan, 0);
+                       OUT_RING  (chan, stride);
+                       OUT_RING  (chan, height);
+                       OUT_RING  (chan, 1);
+                       OUT_RING  (chan, 0);
+                       OUT_RING  (chan, 0);
+               } else {
+                       ret = RING_SPACE(chan, 2);
+                       if (ret)
+                               return ret;
+
+                       BEGIN_RING(chan, NvSubM2MF, 0x0200, 1);
+                       OUT_RING  (chan, 1);
+               }
+               if (old_mem->mem_type == TTM_PL_VRAM && nvbo->tile_flags) {
+                       ret = RING_SPACE(chan, 8);
+                       if (ret)
+                               return ret;
+
+                       BEGIN_RING(chan, NvSubM2MF, 0x021c, 7);
+                       OUT_RING  (chan, 0);
+                       OUT_RING  (chan, 0);
+                       OUT_RING  (chan, stride);
+                       OUT_RING  (chan, height);
+                       OUT_RING  (chan, 1);
+                       OUT_RING  (chan, 0);
+                       OUT_RING  (chan, 0);
+               } else {
+                       ret = RING_SPACE(chan, 2);
+                       if (ret)
+                               return ret;
+
+                       BEGIN_RING(chan, NvSubM2MF, 0x021c, 1);
+                       OUT_RING  (chan, 1);
+               }
+
+               ret = RING_SPACE(chan, 14);
                if (ret)
                        return ret;
-               BEGIN_RING(chan, NvSubM2MF, 0x0200, 1);
-               OUT_RING(chan, 1);
-               BEGIN_RING(chan, NvSubM2MF, 0x021c, 1);
-               OUT_RING(chan, 1);
+
+               BEGIN_RING(chan, NvSubM2MF, 0x0238, 2);
+               OUT_RING  (chan, upper_32_bits(src_offset));
+               OUT_RING  (chan, upper_32_bits(dst_offset));
+               BEGIN_RING(chan, NvSubM2MF, 0x030c, 8);
+               OUT_RING  (chan, lower_32_bits(src_offset));
+               OUT_RING  (chan, lower_32_bits(dst_offset));
+               OUT_RING  (chan, stride);
+               OUT_RING  (chan, stride);
+               OUT_RING  (chan, stride);
+               OUT_RING  (chan, height);
+               OUT_RING  (chan, 0x00000101);
+               OUT_RING  (chan, 0x00000000);
+               BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
+               OUT_RING  (chan, 0);
+
+               length -= amount;
+               src_offset += amount;
+               dst_offset += amount;
        }
 
+       return 0;
+}
+
+static int
+nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
+                 struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
+{
+       u32 src_offset = old_mem->start << PAGE_SHIFT;
+       u32 dst_offset = new_mem->start << PAGE_SHIFT;
+       u32 page_count = new_mem->num_pages;
+       int ret;
+
+       ret = RING_SPACE(chan, 3);
+       if (ret)
+               return ret;
+
+       BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
+       OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
+       OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
+
        page_count = new_mem->num_pages;
        while (page_count) {
                int line_count = (page_count > 2047) ? 2047 : page_count;
 
-               if (dev_priv->card_type >= NV_50) {
-                       ret = RING_SPACE(chan, 3);
-                       if (ret)
-                               return ret;
-                       BEGIN_RING(chan, NvSubM2MF, 0x0238, 2);
-                       OUT_RING(chan, upper_32_bits(src_offset));
-                       OUT_RING(chan, upper_32_bits(dst_offset));
-               }
                ret = RING_SPACE(chan, 11);
                if (ret)
                        return ret;
+
                BEGIN_RING(chan, NvSubM2MF,
                                 NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
-               OUT_RING(chan, lower_32_bits(src_offset));
-               OUT_RING(chan, lower_32_bits(dst_offset));
-               OUT_RING(chan, PAGE_SIZE); /* src_pitch */
-               OUT_RING(chan, PAGE_SIZE); /* dst_pitch */
-               OUT_RING(chan, PAGE_SIZE); /* line_length */
-               OUT_RING(chan, line_count);
-               OUT_RING(chan, (1<<8)|(1<<0));
-               OUT_RING(chan, 0);
+               OUT_RING  (chan, src_offset);
+               OUT_RING  (chan, dst_offset);
+               OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
+               OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
+               OUT_RING  (chan, PAGE_SIZE); /* line_length */
+               OUT_RING  (chan, line_count);
+               OUT_RING  (chan, 0x00000101);
+               OUT_RING  (chan, 0x00000000);
                BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
-               OUT_RING(chan, 0);
+               OUT_RING  (chan, 0);
 
                page_count -= line_count;
                src_offset += (PAGE_SIZE * line_count);
                dst_offset += (PAGE_SIZE * line_count);
        }
 
+       return 0;
+}
+
+static int
+nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
+                    bool no_wait_reserve, bool no_wait_gpu,
+                    struct ttm_mem_reg *new_mem)
+{
+       struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
+       struct nouveau_bo *nvbo = nouveau_bo(bo);
+       struct nouveau_channel *chan;
+       int ret;
+
+       chan = nvbo->channel;
+       if (!chan || nvbo->no_vm)
+               chan = dev_priv->channel;
+
+       if (dev_priv->card_type < NV_50)
+               ret = nv04_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
+       else
+               ret = nv50_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
+       if (ret)
+               return ret;
+
        return nouveau_bo_move_accel_cleanup(chan, nvbo, evict, no_wait_reserve, no_wait_gpu, new_mem);
 }
 
@@ -606,12 +695,7 @@ nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
 
        ret = ttm_bo_move_ttm(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
 out:
-       if (tmp_mem.mm_node) {
-               spin_lock(&bo->bdev->glob->lru_lock);
-               drm_mm_put_block(tmp_mem.mm_node);
-               spin_unlock(&bo->bdev->glob->lru_lock);
-       }
-
+       ttm_bo_mem_put(bo, &tmp_mem);
        return ret;
 }
 
@@ -644,12 +728,7 @@ nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
                goto out;
 
 out:
-       if (tmp_mem.mm_node) {
-               spin_lock(&bo->bdev->glob->lru_lock);
-               drm_mm_put_block(tmp_mem.mm_node);
-               spin_unlock(&bo->bdev->glob->lru_lock);
-       }
-
+       ttm_bo_mem_put(bo, &tmp_mem);
        return ret;
 }
 
@@ -669,7 +748,7 @@ nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem,
                return 0;
        }
 
-       offset = new_mem->mm_node->start << PAGE_SHIFT;
+       offset = new_mem->start << PAGE_SHIFT;
 
        if (dev_priv->card_type == NV_50) {
                ret = nv50_mem_vm_bind_linear(dev,
@@ -719,12 +798,6 @@ nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
        if (ret)
                return ret;
 
-       /* Software copy if the card isn't up and running yet. */
-       if (!dev_priv->channel) {
-               ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
-               goto out;
-       }
-
        /* Fake bo copy. */
        if (old_mem->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
                BUG_ON(bo->mem.mm_node != NULL);
@@ -733,6 +806,12 @@ nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
                goto out;
        }
 
+       /* Software copy if the card isn't up and running yet. */
+       if (!dev_priv->channel) {
+               ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
+               goto out;
+       }
+
        /* Hardware assisted copy. */
        if (new_mem->mem_type == TTM_PL_SYSTEM)
                ret = nouveau_bo_move_flipd(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
@@ -783,14 +862,14 @@ nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
        case TTM_PL_TT:
 #if __OS_HAS_AGP
                if (dev_priv->gart_info.type == NOUVEAU_GART_AGP) {
-                       mem->bus.offset = mem->mm_node->start << PAGE_SHIFT;
+                       mem->bus.offset = mem->start << PAGE_SHIFT;
                        mem->bus.base = dev_priv->gart_info.aper_base;
                        mem->bus.is_iomem = true;
                }
 #endif
                break;
        case TTM_PL_VRAM:
-               mem->bus.offset = mem->mm_node->start << PAGE_SHIFT;
+               mem->bus.offset = mem->start << PAGE_SHIFT;
                mem->bus.base = pci_resource_start(dev->pdev, 1);
                mem->bus.is_iomem = true;
                break;
@@ -808,7 +887,26 @@ nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 static int
 nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
 {
-       return 0;
+       struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
+       struct nouveau_bo *nvbo = nouveau_bo(bo);
+
+       /* as long as the bo isn't in vram, and isn't tiled, we've got
+        * nothing to do here.
+        */
+       if (bo->mem.mem_type != TTM_PL_VRAM) {
+               if (dev_priv->card_type < NV_50 || !nvbo->tile_flags)
+                       return 0;
+       }
+
+       /* make sure bo is in mappable vram */
+       if (bo->mem.start + bo->mem.num_pages < dev_priv->fb_mappable_pages)
+               return 0;
+
+
+       nvbo->placement.fpfn = 0;
+       nvbo->placement.lpfn = dev_priv->fb_mappable_pages;
+       nouveau_bo_placement_set(nvbo, TTM_PL_VRAM, 0);
+       return ttm_bo_validate(bo, &nvbo->placement, false, true, false);
 }
 
 struct ttm_bo_driver nouveau_bo_driver = {