gpu: nvgpu: CDE support

This patch adds support for executing a precompiled GPU program to
allow exporting GPU buffers to other graphics units that have color
decompression engine (CDE) support.

Bug 1409151

Change-Id: Id0c930923f2449b85a6555de71d7ec93eed238ae
Signed-off-by: Arto Merilainen <amerilainen@nvidia.com>
Reviewed-on: http://git-master/r/360418
Reviewed-by: Lauri Peltonen <lpeltonen@nvidia.com>
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>

5 files changed, 1186 insertions, 0 deletions

diff --git a/drivers/gpu/nvgpu/gk20a/Makefile b/drivers/gpu/nvgpu/gk20a/Makefile
index 246f9447..aa9237b4 100644
--- a/drivers/gpu/nvgpu/gk20a/Makefile
+++ b/drivers/gpu/nvgpu/gk20a/Makefile
@@ -34,6 +34,7 @@ nvgpu-y := \
         hal.o \
         hal_gk20a.o \
         gk20a_allocator.o \
+        cde_gk20a.o \
         platform_gk20a_generic.o \
         tsg_gk20a.o
 nvgpu-$(CONFIG_TEGRA_GK20A) += platform_gk20a_tegra.o
diff --git a/drivers/gpu/nvgpu/gk20a/cde_gk20a.c b/drivers/gpu/nvgpu/gk20a/cde_gk20a.c
new file mode 100644
index 00000000..d01426be
--- /dev/null
+++ b/drivers/gpu/nvgpu/gk20a/cde_gk20a.c
@@ -0,0 +1,924 @@
+/*
+ * Color decompression engine support
+ *
+ * Copyright (c) 2014, NVIDIA Corporation.  All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include <linux/nvhost.h>
+#include <linux/dma-mapping.h>
+#include <linux/firmware.h>
+#include <linux/debugfs.h>
+#include <linux/dma-buf.h>
+
+#include "gk20a.h"
+#include "channel_gk20a.h"
+#include "mm_gk20a.h"
+#include "cde_gk20a.h"
+#include "fence_gk20a.h"
+#include "gr_gk20a.h"
+#include "debug_gk20a.h"
+
+#include "hw_ccsr_gk20a.h"
+#include "hw_pbdma_gk20a.h"
+
+void gk20a_cde_dump(struct gk20a_cde_ctx *cde_ctx)
+{
+        int i;
+        for (i = 0; i < cde_ctx->num_bufs; i++) {
+                struct gk20a_cde_mem_desc *target_mem = cde_ctx->mem + i;
+                u32 *target_mem_ptr = target_mem->cpuva;
+                int j = 0;
+
+                gk20a_dbg(gpu_dbg_cde, "cde: buffer=%d, size=%zu, gpuva=%llx\n",
+                          i, target_mem->num_bytes, target_mem->gpu_va);
+
+                for (j = 0; j < target_mem->num_bytes / sizeof(u32); j++)
+                        gk20a_dbg(gpu_dbg_cde, "0x%08x ", target_mem_ptr[j]);
+                gk20a_dbg(gpu_dbg_cde, "\n\n");
+        }
+}
+
+static void gk20a_deinit_cde_img(struct gk20a_cde_ctx *cde_ctx)
+{
+        struct device *dev = &cde_ctx->pdev->dev;
+        int i;
+
+        for (i = 0; i < cde_ctx->num_bufs; i++) {
+                struct gk20a_cde_mem_desc *mem = cde_ctx->mem + i;
+                gk20a_gmmu_unmap(cde_ctx->vm, mem->gpu_va, mem->num_bytes, 1);
+                gk20a_free_sgtable(&mem->sgt);
+                dma_free_coherent(dev, mem->num_bytes, mem->cpuva, mem->iova);
+        }
+
+        for (i = 0; i < cde_ctx->num_obj_ids; i++)
+                gk20a_free_obj_ctx(cde_ctx->ch,
+                        &(struct nvhost_free_obj_ctx_args)
+                        { cde_ctx->obj_ids[i] });
+
+        kfree(cde_ctx->init_cmd);
+        kfree(cde_ctx->convert_cmd);
+
+        cde_ctx->convert_cmd = NULL;
+        cde_ctx->init_cmd = NULL;
+        cde_ctx->num_bufs = 0;
+        cde_ctx->num_obj_ids = 0;
+        cde_ctx->num_params = 0;
+        cde_ctx->init_cmd_num_entries = 0;
+        cde_ctx->convert_cmd_num_entries = 0;
+}
+
+static int gk20a_cde_remove(struct gk20a_cde_ctx *cde_ctx)
+{
+        struct gk20a *g = cde_ctx->g;
+        struct channel_gk20a *ch = cde_ctx->ch;
+        struct vm_gk20a *vm = ch->vm;
+
+        /* free the channel */
+        gk20a_free_channel(cde_ctx->ch, true);
+
+        /* ..then release mapped memory */
+        gk20a_deinit_cde_img(cde_ctx);
+        gk20a_gmmu_unmap(vm, cde_ctx->backing_store_vaddr,
+                         g->gr.compbit_store.size, 1);
+
+        return 0;
+}
+
+int gk20a_cde_destroy(struct gk20a *g)
+{
+        struct gk20a_cde_app *cde_app = &g->cde_app;
+        struct gk20a_cde_ctx *cde_ctx = cde_app->cde_ctx;
+        int ret, i;
+
+        if (!cde_app->initialised)
+                return 0;
+
+        for (i = 0; i < ARRAY_SIZE(cde_app->cde_ctx); i++, cde_ctx++)
+                ret = gk20a_cde_remove(cde_ctx);
+
+        cde_app->initialised = false;
+        return ret;
+}
+
+static int gk20a_init_cde_buf(struct gk20a_cde_ctx *cde_ctx,
+                              const struct firmware *img,
+                              struct gk20a_cde_hdr_buf *buf)
+{
+        struct device *dev = &cde_ctx->pdev->dev;
+        struct gk20a_cde_mem_desc *mem;
+        int err;
+
+        /* check that the file can hold the buf */
+        if (buf->data_byte_offset != 0 &&
+            buf->data_byte_offset + buf->num_bytes > img->size) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: invalid data section. buffer idx = %d",
+                           cde_ctx->num_bufs);
+                return -EINVAL;
+        }
+
+        /* check that we have enough buf elems available */
+        if (cde_ctx->num_bufs > MAX_CDE_BUFS) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: invalid data section. buffer idx = %d",
+                           cde_ctx->num_bufs);
+                return -ENOMEM;
+        }
+
+        /* allocate buf */
+        mem = cde_ctx->mem + cde_ctx->num_bufs;
+        mem->num_bytes = buf->num_bytes;
+        mem->cpuva = dma_alloc_coherent(dev, mem->num_bytes, &mem->iova,
+                                        GFP_KERNEL);
+        if (!mem->cpuva) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: could not allocate device memory. buffer idx = %d",
+                           cde_ctx->num_bufs);
+                return -ENOMEM;
+        }
+
+        err = gk20a_get_sgtable(dev, &mem->sgt, mem->cpuva, mem->iova,
+                                mem->num_bytes);
+        if (err) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: could not get sg table. buffer idx = %d",
+                           cde_ctx->num_bufs);
+                err = -ENOMEM;
+                goto err_get_sgtable;
+        }
+
+        mem->gpu_va = gk20a_gmmu_map(cde_ctx->vm, &mem->sgt, mem->num_bytes, 0,
+                                     gk20a_mem_flag_none);
+        if (!mem->gpu_va) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: could not map buffer to gpuva. buffer idx = %d",
+                           cde_ctx->num_bufs);
+                err = -ENOMEM;
+                goto err_map_buffer;
+        }
+
+        /* copy the content */
+        if (buf->data_byte_offset != 0)
+                memcpy(mem->cpuva, img->data + buf->data_byte_offset,
+                       buf->num_bytes);
+
+        cde_ctx->num_bufs++;
+
+        return 0;
+
+err_map_buffer:
+        gk20a_free_sgtable(&mem->sgt);
+        kfree(mem->sgt);
+err_get_sgtable:
+        dma_free_coherent(dev, mem->num_bytes, &mem->cpuva, mem->iova);
+        return err;
+}
+
+static int gk20a_replace_data(struct gk20a_cde_ctx *cde_ctx, void *target,
+                              int type, s32 shift, u64 mask, u64 value)
+{
+        u32 *target_mem_ptr = target;
+        u64 *target_mem_ptr_u64 = target;
+        u64 current_value, new_value;
+
+        value = (shift >= 0) ? value << shift : value >> -shift;
+        value &= mask;
+
+        /* read current data from the location */
+        if (type == TYPE_PARAM_TYPE_U32)
+                current_value = *target_mem_ptr;
+        else if (type == TYPE_PARAM_TYPE_U64_LITTLE)
+                current_value = *target_mem_ptr_u64;
+        else if (type == TYPE_PARAM_TYPE_U64_BIG) {
+                current_value = *target_mem_ptr_u64;
+                current_value = (u64)(current_value >> 32) |
+                        (u64)(current_value << 32);
+        } else {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: unknown type. type=%d",
+                           type);
+                return -EINVAL;
+        }
+
+        current_value &= ~mask;
+        new_value = current_value | value;
+
+        /* store the element data back */
+        if (type == TYPE_PARAM_TYPE_U32)
+                *target_mem_ptr = (u32)new_value;
+        else if (type == TYPE_PARAM_TYPE_U64_LITTLE)
+                *target_mem_ptr_u64 = new_value;
+        else  {
+                new_value = (u64)(new_value >> 32) |
+                        (u64)(new_value << 32);
+                *target_mem_ptr_u64 = new_value;
+        }
+
+        return 0;
+}
+
+static int gk20a_init_cde_replace(struct gk20a_cde_ctx *cde_ctx,
+                                      const struct firmware *img,
+                                      struct gk20a_cde_hdr_replace *replace)
+{
+        struct gk20a_cde_mem_desc *source_mem;
+        struct gk20a_cde_mem_desc *target_mem;
+        u32 *target_mem_ptr;
+        u64 vaddr;
+        int err;
+
+        if (replace->target_buf >= cde_ctx->num_bufs ||
+            replace->source_buf >= cde_ctx->num_bufs) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: invalid buffer. target_buf=%u, source_buf=%u, num_bufs=%d",
+                           replace->target_buf, replace->source_buf,
+                           cde_ctx->num_bufs);
+                return -EINVAL;
+        }
+
+        source_mem = cde_ctx->mem + replace->source_buf;
+        target_mem = cde_ctx->mem + replace->target_buf;
+        target_mem_ptr = target_mem->cpuva;
+
+        if (source_mem->num_bytes < (replace->source_byte_offset + 3) ||
+            target_mem->num_bytes < (replace->target_byte_offset + 3)) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: invalid buffer offsets. target_buf_offs=%lld, source_buf_offs=%lld, source_buf_size=%zu, dest_buf_size=%zu",
+                           replace->target_byte_offset,
+                           replace->source_byte_offset,
+                         source_mem->num_bytes,
+                         target_mem->num_bytes);
+                return -EINVAL;
+        }
+
+        /* calculate the target pointer */
+        target_mem_ptr += (replace->target_byte_offset / sizeof(u32));
+
+        /* determine patch value */
+        vaddr = source_mem->gpu_va + replace->source_byte_offset;
+        err = gk20a_replace_data(cde_ctx, target_mem_ptr, replace->type,
+                                 replace->shift, replace->mask,
+                                 vaddr);
+        if (err) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: replace failed. err=%d, target_buf=%u, target_buf_offs=%lld, source_buf=%u, source_buf_offs=%lld",
+                           err, replace->target_buf,
+                           replace->target_byte_offset,
+                           replace->source_buf,
+                           replace->source_byte_offset);
+        }
+
+        return err;
+}
+
+static int gk20a_cde_patch_params(struct gk20a_cde_ctx *cde_ctx)
+{
+        struct gk20a *g = cde_ctx->g;
+        struct gk20a_cde_mem_desc *target_mem;
+        u32 *target_mem_ptr;
+        u64 new_data;
+        int user_id = 0, i, err;
+
+        for (i = 0; i < cde_ctx->num_params; i++) {
+                struct gk20a_cde_hdr_param *param = cde_ctx->params + i;
+                target_mem = cde_ctx->mem + param->target_buf;
+                target_mem_ptr = target_mem->cpuva;
+                target_mem_ptr += (param->target_byte_offset / sizeof(u32));
+
+                switch (param->id) {
+                case TYPE_PARAM_COMPTAGS_PER_CACHELINE:
+                        new_data = g->gr.comptags_per_cacheline;
+                        break;
+                case TYPE_PARAM_GPU_CONFIGURATION:
+                        new_data = g->ltc_count * g->gr.slices_per_ltc *
+                                g->gr.cacheline_size;
+                        break;
+                case TYPE_PARAM_FIRSTPAGEOFFSET:
+                        new_data = cde_ctx->src_param_offset;
+                        break;
+                case TYPE_PARAM_NUMPAGES:
+                        new_data = cde_ctx->src_param_lines;
+                        break;
+                case TYPE_PARAM_BACKINGSTORE:
+                        new_data = cde_ctx->backing_store_vaddr;
+                        break;
+                case TYPE_PARAM_DESTINATION:
+                        new_data = cde_ctx->dest_vaddr;
+                        break;
+                case TYPE_PARAM_DESTINATION_SIZE:
+                        new_data = cde_ctx->dest_size;
+                        break;
+                case TYPE_PARAM_BACKINGSTORE_SIZE:
+                        new_data = g->gr.compbit_store.size;
+                        break;
+                case TYPE_PARAM_SOURCE_SMMU_ADDR:
+                        new_data = gk20a_mm_gpuva_to_iova(cde_ctx->vm,
+                                                          cde_ctx->src_vaddr);
+                        if (new_data == 0)
+                                err = -EINVAL;
+                        break;
+                default:
+                        user_id = param->id - NUM_RESERVED_PARAMS;
+                        if (user_id < 0 || user_id >= MAX_CDE_USER_PARAMS)
+                                continue;
+                        new_data = cde_ctx->user_param_values[user_id];
+                }
+
+                gk20a_dbg(gpu_dbg_cde, "cde: patch: idx_in_file=%d      param_id=%d     target_buf=%u   target_byte_offset=%lld data_value=0x%llx       data_offset/data_diff=%lld      data_type=%d    data_shift=%d   data_mask=0x%llx",
+                          i, param->id, param->target_buf,
+                          param->target_byte_offset, new_data,
+                          param->data_offset, param->type, param->shift,
+                          param->mask);
+
+                new_data += param->data_offset;
+
+                err = gk20a_replace_data(cde_ctx, target_mem_ptr, param->type,
+                                         param->shift, param->mask, new_data);
+
+                if (err) {
+                        gk20a_warn(&cde_ctx->pdev->dev, "cde: patch failed. err=%d, idx=%d, id=%d, target_buf=%u, target_buf_offs=%lld, patch_value=%llu",
+                                   err, i, param->id, param->target_buf,
+                                   param->target_byte_offset, new_data);
+                        return err;
+                }
+        }
+
+        return 0;
+}
+
+static int gk20a_init_cde_param(struct gk20a_cde_ctx *cde_ctx,
+                                const struct firmware *img,
+                                struct gk20a_cde_hdr_param *param)
+{
+        struct gk20a_cde_mem_desc *target_mem;
+
+        if (param->target_buf >= cde_ctx->num_bufs) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: invalid buffer parameter. param idx = %d, target_buf=%u, num_bufs=%u",
+                           cde_ctx->num_params, param->target_buf,
+                           cde_ctx->num_bufs);
+                return -EINVAL;
+        }
+
+        target_mem = cde_ctx->mem + param->target_buf;
+        if (target_mem->num_bytes < (param->target_byte_offset + 3)) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: invalid buffer parameter. param idx = %d, target_buf_offs=%lld, target_buf_size=%zu",
+                           cde_ctx->num_params, param->target_byte_offset,
+                           target_mem->num_bytes);
+                return -EINVAL;
+        }
+
+        /* does this parameter fit into our parameter structure */
+        if (cde_ctx->num_params >= MAX_CDE_PARAMS) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: no room for new parameters param idx = %d",
+                           cde_ctx->num_params);
+                return -ENOMEM;
+        }
+
+        /* is the given id valid? */
+        if (param->id >= NUM_RESERVED_PARAMS + MAX_CDE_USER_PARAMS) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: parameter id is not valid. param idx = %d, id=%u, max=%u",
+                           param->id, cde_ctx->num_params,
+                           NUM_RESERVED_PARAMS + MAX_CDE_USER_PARAMS);
+                return -EINVAL;
+        }
+
+        cde_ctx->params[cde_ctx->num_params] = *param;
+        cde_ctx->num_params++;
+
+        return 0;
+}
+
+static int gk20a_init_cde_required_class(struct gk20a_cde_ctx *cde_ctx,
+                                         const struct firmware *img,
+                                         u32 required_class)
+{
+        struct nvhost_alloc_obj_ctx_args alloc_obj_ctx;
+        int err;
+
+        if (cde_ctx->num_obj_ids >= MAX_CDE_OBJ_IDS) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: running out of class ids");
+                return -ENOMEM;
+        }
+
+        alloc_obj_ctx.class_num = required_class;
+        alloc_obj_ctx.padding = 0;
+
+        err = gk20a_alloc_obj_ctx(cde_ctx->ch, &alloc_obj_ctx);
+        if (err) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: failed to allocate ctx. err=%d",
+                           err);
+                return err;
+        }
+
+        cde_ctx->obj_ids[cde_ctx->num_obj_ids] = alloc_obj_ctx.obj_id;
+        cde_ctx->num_obj_ids++;
+
+        return 0;
+}
+
+static int gk20a_init_cde_command(struct gk20a_cde_ctx *cde_ctx,
+                                  const struct firmware *img,
+                                  u32 op,
+                                  struct gk20a_cde_cmd_elem *cmd_elem,
+                                  u32 num_elems)
+{
+        struct nvhost_gpfifo **gpfifo, *gpfifo_elem;
+        u32 *num_entries;
+        int i;
+
+        /* check command type */
+        if (op == TYPE_BUF_COMMAND_INIT) {
+                gpfifo = &cde_ctx->init_cmd;
+                num_entries = &cde_ctx->init_cmd_num_entries;
+        } else if (op == TYPE_BUF_COMMAND_CONVERT) {
+                gpfifo = &cde_ctx->convert_cmd;
+                num_entries = &cde_ctx->convert_cmd_num_entries;
+        } else {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: unknown command. op=%u",
+                           op);
+                return -EINVAL;
+        }
+
+        /* allocate gpfifo entries to be pushed */
+        *gpfifo = kzalloc(sizeof(struct nvhost_gpfifo) * num_elems,
+                          GFP_KERNEL);
+        if (!*gpfifo) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: could not allocate memory for gpfifo entries");
+                return -ENOMEM;
+        }
+
+        gpfifo_elem = *gpfifo;
+        for (i = 0; i < num_elems; i++, cmd_elem++, gpfifo_elem++) {
+                struct gk20a_cde_mem_desc *target_mem;
+
+                /* validate the current entry */
+                if (cmd_elem->target_buf >= cde_ctx->num_bufs) {
+                        gk20a_warn(&cde_ctx->pdev->dev, "cde: target buffer is not available (target=%u, num_bufs=%u)",
+                                   cmd_elem->target_buf, cde_ctx->num_bufs);
+                        return -EINVAL;
+                }
+
+                target_mem = cde_ctx->mem + cmd_elem->target_buf;
+                if (target_mem->num_bytes <
+                    cmd_elem->target_byte_offset + cmd_elem->num_bytes) {
+                        gk20a_warn(&cde_ctx->pdev->dev, "cde: target buffer cannot hold all entries (target_size=%zu, target_byte_offset=%lld, num_bytes=%llu)",
+                                   target_mem->num_bytes,
+                                   cmd_elem->target_byte_offset,
+                                   cmd_elem->num_bytes);
+                        return -EINVAL;
+                }
+
+                /* store the element into gpfifo */
+                gpfifo_elem->entry0 =
+                        u64_lo32(target_mem->gpu_va +
+                        cmd_elem->target_byte_offset);
+                gpfifo_elem->entry1 =
+                        u64_hi32(target_mem->gpu_va +
+                        cmd_elem->target_byte_offset) |
+                        pbdma_gp_entry1_length_f(cmd_elem->num_bytes /
+                                                 sizeof(u32));
+        }
+
+        *num_entries = num_elems;
+        return 0;
+}
+
+static int gk20a_init_cde_img(struct gk20a_cde_ctx *cde_ctx,
+                              const struct firmware *img)
+{
+        u32 *data = (u32 *)img->data;
+        u32 version, num_of_elems;
+        struct gk20a_cde_hdr_elem *elem;
+        u32 min_size = 0;
+        int err = 0;
+        int i;
+
+        min_size += 2 * sizeof(u32);
+        if (img->size < min_size) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: invalid image header");
+                return -EINVAL;
+        }
+
+        version = data[0];
+        num_of_elems = data[1];
+
+        min_size += num_of_elems * sizeof(*elem);
+        if (img->size < min_size) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: bad image");
+                return -EINVAL;
+        }
+
+        elem = (struct gk20a_cde_hdr_elem *)&data[2];
+        for (i = 0; i < num_of_elems; i++) {
+                int err = 0;
+                switch (elem->type) {
+                case TYPE_BUF:
+                        err = gk20a_init_cde_buf(cde_ctx, img, &elem->buf);
+                        break;
+                case TYPE_REPLACE:
+                        err = gk20a_init_cde_replace(cde_ctx, img,
+                                                     &elem->replace);
+                        break;
+                case TYPE_PARAM:
+                        err = gk20a_init_cde_param(cde_ctx, img, &elem->param);
+                        break;
+                case TYPE_REQUIRED_CLASS:
+                        err = gk20a_init_cde_required_class(cde_ctx, img,
+                                elem->required_class);
+                        break;
+                case TYPE_COMMAND:
+                {
+                        struct gk20a_cde_cmd_elem *cmd = (void *)
+                                &img->data[elem->command.data_byte_offset];
+                        err = gk20a_init_cde_command(cde_ctx, img,
+                                elem->command.op, cmd,
+                                elem->command.num_entries);
+                        break;
+                }
+                default:
+                        gk20a_warn(&cde_ctx->pdev->dev, "cde: unknown header element");
+                        err = -EINVAL;
+                }
+
+                if (err)
+                        goto deinit_image;
+
+                elem++;
+        }
+
+        if (!cde_ctx->init_cmd || !cde_ctx->init_cmd_num_entries) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: convert command not defined");
+                err = -EINVAL;
+                goto deinit_image;
+        }
+
+        if (!cde_ctx->convert_cmd || !cde_ctx->convert_cmd_num_entries) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: convert command not defined");
+                err = -EINVAL;
+                goto deinit_image;
+        }
+
+        return 0;
+
+deinit_image:
+        gk20a_deinit_cde_img(cde_ctx);
+        return err;
+}
+
+static int gk20a_cde_execute_buffer(struct gk20a_cde_ctx *cde_ctx,
+                                    u32 op, struct nvhost_fence *fence,
+                                    u32 flags, struct gk20a_fence **fence_out)
+{
+        struct nvhost_gpfifo *gpfifo = NULL;
+        int num_entries = 0;
+
+        /* check command type */
+        if (op == TYPE_BUF_COMMAND_INIT) {
+                gpfifo = cde_ctx->init_cmd;
+                num_entries = cde_ctx->init_cmd_num_entries;
+        } else if (op == TYPE_BUF_COMMAND_CONVERT) {
+                gpfifo = cde_ctx->convert_cmd;
+                num_entries = cde_ctx->convert_cmd_num_entries;
+        } else {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: unknown buffer");
+                return -EINVAL;
+        }
+
+        if (gpfifo == NULL || num_entries == 0) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: buffer not available");
+                return -ENOSYS;
+        }
+
+        return gk20a_submit_channel_gpfifo(cde_ctx->ch, gpfifo,
+                                           num_entries, flags, fence, fence_out);
+}
+
+int gk20a_cde_convert(struct gk20a *g, u32 src_fd, u32 dst_fd,
+                      s32 dst_kind, u64 dst_byte_offset,
+                      u32 dst_size, struct nvhost_fence *fence,
+                      u32 __flags, struct gk20a_cde_param *params,
+                      int num_params, struct gk20a_fence **fence_out)
+{
+        struct gk20a_cde_app *cde_app = &g->cde_app;
+        struct gk20a_comptags comptags;
+        struct gk20a_cde_ctx *cde_ctx;
+        struct dma_buf *src = NULL, *dst = NULL;
+        u64 dst_vaddr = 0, src_vaddr = 0;
+        u32 flags;
+        int err, i;
+
+        if (!cde_app->initialised) {
+                gk20a_warn(&g->dev->dev, "cde: conversion requrest but no image has been provided");
+                return -ENOSYS;
+        }
+
+        mutex_lock(&cde_app->mutex);
+
+        /* pick next free cde context */
+        cde_ctx = cde_app->cde_ctx + cde_app->cde_ctx_ptr;
+        cde_app->cde_ctx_ptr = (cde_app->cde_ctx_ptr + 1) %
+                ARRAY_SIZE(cde_app->cde_ctx);
+
+        /* First, get buffer references and map the buffers to local va */
+
+        dst = dma_buf_get(dst_fd);
+        if (IS_ERR(src)) {
+                dst = NULL;
+                err = -EINVAL;
+                goto exit_unlock;
+        }
+
+        /* ensure that the dst buffer has drvdata */
+        err = gk20a_dmabuf_alloc_drvdata(dst, &g->dev->dev);
+        if (err)
+                goto exit_unlock;
+
+        /* map the destination buffer */
+        dst_vaddr = gk20a_vm_map(g->cde_app.vm, dst, 0,
+                                 0, dst_kind, NULL, true,
+                                 gk20a_mem_flag_none,
+                                 0, 0);
+        if (!dst_vaddr) {
+                err = -EINVAL;
+                goto exit_unlock;
+        }
+
+        src = dma_buf_get(src_fd);
+        if (IS_ERR(src)) {
+                src = NULL;
+                err = -EINVAL;
+                goto exit_unlock;
+        }
+
+        /* ensure that the src buffer has drvdata */
+        err = gk20a_dmabuf_alloc_drvdata(src, &g->dev->dev);
+        if (err)
+                goto exit_unlock;
+
+        /* map the source buffer to prevent premature release */
+        src_vaddr = gk20a_vm_map(g->cde_app.vm, src, 0,
+                                 0, dst_kind, NULL, true,
+                                 gk20a_mem_flag_none,
+                                 0, 0);
+        if (!src_vaddr) {
+                err = -EINVAL;
+                goto exit_unlock;
+        }
+
+        if (!dst_size)
+                dst_size = dst->size - dst_byte_offset;
+
+        /* reload buffer converter if it has failed */
+        if (cde_ctx->ch->has_timedout) {
+                mutex_unlock(&cde_app->mutex);
+                err = gk20a_cde_reload(g);
+                if (err)
+                        return err;
+                mutex_lock(&cde_app->mutex);
+        }
+
+        /* wait for channel idle */
+        err = gk20a_channel_finish(cde_ctx->ch, 2000);
+        if (err) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: old work could not be finished");
+                goto exit_unlock;
+        }
+
+        /* disable the channel */
+        gk20a_writel(g, ccsr_channel_r(cde_ctx->ch->hw_chid),
+                gk20a_readl(g, ccsr_channel_r(cde_ctx->ch->hw_chid)) |
+                ccsr_channel_enable_clr_true_f());
+        gk20a_fifo_preempt_channel(g, cde_ctx->ch->hw_chid);
+        channel_gk20a_unbind(&g->fifo.channel[cde_ctx->ch->hw_chid]);
+
+        /* reinitialise the graphics context of the channel */
+        gr_gk20a_load_golden_ctx_image(g, cde_ctx->ch);
+
+        /* re-enable the channel */
+        g->ops.fifo.bind_channel(&g->fifo.channel[cde_ctx->ch->hw_chid]);
+        gk20a_writel(g, ccsr_channel_r(cde_ctx->ch->hw_chid),
+                gk20a_readl(g, ccsr_channel_r(cde_ctx->ch->hw_chid)) |
+                ccsr_channel_enable_set_true_f());
+
+        /* store source buffer compression tags */
+        gk20a_get_comptags(&g->dev->dev, src, &comptags);
+        cde_ctx->src_vaddr = src_vaddr;
+        cde_ctx->src_param_offset = comptags.offset;
+        cde_ctx->src_param_lines = comptags.lines;
+
+        /* store information about destination */
+        cde_ctx->dest_vaddr = dst_vaddr + dst_byte_offset;
+        cde_ctx->dest_size = dst_size;
+
+        /* remove existing argument data */
+        memset(cde_ctx->user_param_values, 0,
+               sizeof(cde_ctx->user_param_values));
+
+        /* read user space arguments for the conversion */
+        for (i = 0; i < num_params; i++) {
+                struct gk20a_cde_param *param = params + i;
+                int id = param->id - NUM_RESERVED_PARAMS;
+
+                if (id < 0 || id >= MAX_CDE_USER_PARAMS) {
+                        gk20a_warn(&cde_ctx->pdev->dev, "cde: unknown user parameter");
+                        err = -EINVAL;
+                        goto exit_unlock;
+                }
+                cde_ctx->user_param_values[id] = param->value;
+        }
+
+        /* patch data */
+        err = gk20a_cde_patch_params(cde_ctx);
+        if (err) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: failed to patch parameters");
+                goto exit_unlock;
+        }
+
+        gk20a_dbg(gpu_dbg_cde, "cde: buffer=cbc, size=%zu, gpuva=%llx\n",
+                 g->gr.compbit_store.size, cde_ctx->backing_store_vaddr);
+        gk20a_dbg(gpu_dbg_cde, "cde: buffer=dst, size=%llu, gpuva=%llx\n",
+                 cde_ctx->dest_size, cde_ctx->dest_vaddr);
+        gk20a_cde_dump(cde_ctx);
+
+        /* execute the init push buffer */
+        err = gk20a_cde_execute_buffer(cde_ctx, TYPE_BUF_COMMAND_INIT,
+                                       NULL, 0, NULL);
+        if (err)
+                goto exit_unlock;
+
+        /* take always the postfence as it is needed for protecting the
+         * cde context */
+        flags = __flags | NVHOST_SUBMIT_GPFIFO_FLAGS_FENCE_GET;
+
+        /* execute the conversion buffer */
+        err = gk20a_cde_execute_buffer(cde_ctx, TYPE_BUF_COMMAND_CONVERT,
+                                       fence, flags, fence_out);
+
+exit_unlock:
+
+        /* unmap the buffers - channel holds references to them now */
+        if (dst_vaddr)
+                gk20a_vm_unmap(g->cde_app.vm, dst_vaddr);
+        if (src_vaddr)
+                gk20a_vm_unmap(g->cde_app.vm, src_vaddr);
+
+        /* drop dmabuf refs if work was aborted */
+        if (err && src)
+                dma_buf_put(src);
+        if (err && dst)
+                dma_buf_put(dst);
+
+        mutex_unlock(&cde_app->mutex);
+
+        return err;
+}
+
+int gk20a_cde_load(struct gk20a_cde_ctx *cde_ctx)
+{
+        struct gk20a *g = cde_ctx->g;
+        const struct firmware *img;
+        struct channel_gk20a *ch;
+        struct gr_gk20a *gr = &g->gr;
+        int err = 0;
+        u64 vaddr;
+
+        img = gk20a_request_firmware(g, "gpu2cde.bin");
+        if (!img) {
+                dev_err(&cde_ctx->pdev->dev, "cde: could not fetch the firmware");
+                return -ENOSYS;
+        }
+
+        ch = gk20a_open_new_channel(g);
+        if (!ch) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: gk20a channel not available");
+                err = -ENOMEM;
+                goto err_get_gk20a_channel;
+        }
+
+        /* bind the channel to the vm */
+        gk20a_vm_get(&g->mm.pmu.vm);
+        ch->vm = &g->mm.pmu.vm;
+        err = channel_gk20a_commit_va(ch);
+        if (err) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: could not bind vm");
+                goto err_commit_va;
+        }
+
+        /* allocate gpfifo (1024 should be more than enough) */
+        err = gk20a_alloc_channel_gpfifo(ch,
+                &(struct nvhost_alloc_gpfifo_args){1024, 0});
+        if (err) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: unable to allocate gpfifo");
+                goto err_alloc_gpfifo;
+        }
+
+        /* map backing store to gpu virtual space */
+        vaddr = gk20a_gmmu_map(ch->vm, &gr->compbit_store.sgt,
+                               g->gr.compbit_store.size, 0,
+                               gk20a_mem_flag_none);
+
+        if (!vaddr) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: cannot map compression bit backing store");
+                goto err_map_backingstore;
+        }
+
+        /* store initialisation data */
+        cde_ctx->ch = ch;
+        cde_ctx->vm = ch->vm;
+        cde_ctx->backing_store_vaddr = vaddr;
+
+        /* initialise the firmware */
+        err = gk20a_init_cde_img(cde_ctx, img);
+        if (err) {
+                gk20a_warn(&cde_ctx->pdev->dev, "cde: image initialisation failed");
+                goto err_init_cde_img;
+        }
+
+        /* initialisation done */
+        release_firmware(img);
+
+        return 0;
+
+err_init_cde_img:
+        gk20a_gmmu_unmap(ch->vm, vaddr, g->gr.compbit_store.size, 1);
+err_map_backingstore:
+err_alloc_gpfifo:
+        gk20a_vm_put(ch->vm);
+err_commit_va:
+err_get_gk20a_channel:
+        release_firmware(img);
+        dev_err(&cde_ctx->pdev->dev, "cde: couldn't initialise buffer converter: %d",
+                err);
+        return err;
+}
+
+int gk20a_cde_reload(struct gk20a *g)
+{
+        struct gk20a_cde_app *cde_app = &g->cde_app;
+        struct gk20a_cde_ctx *cde_ctx = cde_app->cde_ctx;
+        int err, i;
+
+        if (!cde_app->initialised) {
+                gk20a_busy(g->dev);
+                gk20a_init_cde_support(g);
+                gk20a_idle(g->dev);
+                if (!cde_app->initialised)
+                        return -ENOSYS;
+                return 0;
+        }
+
+        gk20a_busy(g->dev);
+        mutex_lock(&cde_app->mutex);
+        for (i = 0; i < ARRAY_SIZE(cde_app->cde_ctx); i++, cde_ctx++) {
+                gk20a_cde_remove(cde_ctx);
+                err = gk20a_cde_load(cde_ctx);
+        }
+
+        cde_app->cde_ctx_ptr = 0;
+        mutex_unlock(&cde_app->mutex);
+
+        gk20a_idle(g->dev);
+        return err;
+}
+
+int gk20a_init_cde_support(struct gk20a *g)
+{
+        struct gk20a_cde_app *cde_app = &g->cde_app;
+        struct gk20a_cde_ctx *cde_ctx = cde_app->cde_ctx;
+        int ret, i;
+
+        if (cde_app->initialised)
+                return 0;
+
+        mutex_init(&cde_app->mutex);
+        mutex_lock(&cde_app->mutex);
+
+        for (i = 0; i < ARRAY_SIZE(cde_app->cde_ctx); i++, cde_ctx++) {
+                cde_ctx->g = g;
+                cde_ctx->pdev = g->dev;
+                ret = gk20a_cde_load(cde_ctx);
+                if (ret)
+                        goto err_init_instance;
+        }
+
+        /* take shadow to the vm for general usage */
+        cde_app->vm = cde_app->cde_ctx->vm;
+
+        cde_app->cde_ctx_ptr = 0;
+        cde_app->initialised = true;
+        mutex_unlock(&cde_app->mutex);
+
+        return 0;
+
+err_init_instance:
+
+        /* deinitialise initialised channels */
+        while (i--) {
+                gk20a_cde_remove(cde_ctx);
+                cde_ctx--;
+        }
+        return ret;
+}
diff --git a/drivers/gpu/nvgpu/gk20a/cde_gk20a.h b/drivers/gpu/nvgpu/gk20a/cde_gk20a.h
new file mode 100644
index 00000000..784ae8b4
--- /dev/null
+++ b/drivers/gpu/nvgpu/gk20a/cde_gk20a.h
@@ -0,0 +1,254 @@
+/*
+ * GK20A color decompression engine support
+ *
+ * Copyright (c) 2014, NVIDIA Corporation.  All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _CDE_GK20A_H_
+#define _CDE_GK20A_H_
+
+#include "mm_gk20a.h"
+
+#define MAX_CDE_BUFS            10
+#define MAX_CDE_PARAMS          64
+#define MAX_CDE_USER_PARAMS     32
+#define MAX_CDE_OBJ_IDS         4
+
+struct dma_buf;
+struct gk20a;
+
+/*
+ * this element defines a buffer that is allocated and mapped into gpu address
+ * space. data_byte_offset defines the beginning of the buffer inside the
+ * firmare. num_bytes defines how many bytes the firmware contains.
+ *
+ * If data_byte_offset is zero, we allocate an empty buffer.
+ */
+
+struct gk20a_cde_hdr_buf {
+        u64 data_byte_offset;
+        u64 num_bytes;
+};
+
+/*
+ * this element defines a constant patching in buffers. It basically
+ * computes physical address to <source_buf>+source_byte_offset. The
+ * address is then modified into patch value as per:
+ *    value = (current_value & ~mask) | (address << shift) & mask .
+ *
+ * The type field defines the register size as:
+ *  0=u32,
+ *  1=u64 (little endian),
+ *  2=u64 (big endian)
+ */
+
+struct gk20a_cde_hdr_replace {
+        u32 target_buf;
+        u32 source_buf;
+        s32 shift;
+        u32 type;
+        s64 target_byte_offset;
+        s64 source_byte_offset;
+        u64 mask;
+};
+
+enum {
+        TYPE_PARAM_TYPE_U32 = 0,
+        TYPE_PARAM_TYPE_U64_LITTLE,
+        TYPE_PARAM_TYPE_U64_BIG
+};
+
+/*
+ * this element defines a runtime patching in buffers. Parameters with id from
+ * 0 to 1024 are reserved for special usage as follows:
+ *   0 = comptags_per_cacheline,
+ *   1 = slices_per_fbp,
+ *   2 = num_fbps
+ *   3 = source buffer first page offset
+ *   4 = source buffer block height log2
+ *   5 = backing store memory address
+ *   6 = destination memory address
+ *   7 = destination size (bytes)
+ *   8 = backing store size (bytes)
+ *   9 = cache line size
+ *
+ * Parameters above id 1024 are user-specified. I.e. they determine where a
+ * parameters from user space should be placed in buffers, what is their
+ * type, etc.
+ *
+ * Once the value is available, we add data_offset to the value.
+ *
+ * The value address is then modified into patch value as per:
+ *    value = (current_value & ~mask) | (address << shift) & mask .
+ *
+ * The type field defines the register size as:
+ *  0=u32,
+ *  1=u64 (little endian),
+ *  2=u64 (big endian)
+ */
+
+struct gk20a_cde_hdr_param {
+        u32 id;
+        u32 target_buf;
+        s32 shift;
+        u32 type;
+        s64 data_offset;
+        s64 target_byte_offset;
+        u64 mask;
+};
+
+enum {
+        TYPE_PARAM_COMPTAGS_PER_CACHELINE = 0,
+        TYPE_PARAM_GPU_CONFIGURATION,
+        TYPE_PARAM_FIRSTPAGEOFFSET,
+        TYPE_PARAM_NUMPAGES,
+        TYPE_PARAM_BACKINGSTORE,
+        TYPE_PARAM_DESTINATION,
+        TYPE_PARAM_DESTINATION_SIZE,
+        TYPE_PARAM_BACKINGSTORE_SIZE,
+        TYPE_PARAM_SOURCE_SMMU_ADDR,
+        NUM_RESERVED_PARAMS = 1024,
+};
+
+/*
+ * This header element defines a command. The op field determines whether the
+ * element is defining an init (0) or convert command (1). data_byte_offset
+ * denotes the beginning address of command elements in the file.
+ */
+
+struct gk20a_cde_hdr_command {
+        u32 op;
+        u32 num_entries;
+        u64 data_byte_offset;
+};
+
+enum {
+        TYPE_BUF_COMMAND_INIT = 0,
+        TYPE_BUF_COMMAND_CONVERT
+};
+
+/*
+ * This is a command element defines one entry inside push buffer. target_buf
+ * defines the buffer including the pushbuffer entries, target_byte_offset the
+ * offset inside the buffer and num_bytes the number of words in the buffer.
+ */
+
+struct gk20a_cde_cmd_elem {
+        u32 target_buf;
+        u32 padding;
+        u64 target_byte_offset;
+        u64 num_bytes;
+};
+
+/*
+ * Following defines a single header element. Each element has a type and
+ * some of the data structures.
+ */
+
+struct gk20a_cde_hdr_elem {
+        u32 type;
+        u32 padding;
+        union {
+                struct gk20a_cde_hdr_buf buf;
+                struct gk20a_cde_hdr_replace replace;
+                struct gk20a_cde_hdr_param param;
+                u32 required_class;
+                struct gk20a_cde_hdr_command command;
+        };
+};
+
+enum {
+        TYPE_BUF = 0,
+        TYPE_REPLACE,
+        TYPE_PARAM,
+        TYPE_REQUIRED_CLASS,
+        TYPE_COMMAND
+};
+
+struct gk20a_cde_mem_desc {
+        struct sg_table *sgt;
+        dma_addr_t iova;
+        void *cpuva;
+        size_t num_bytes;
+        u64 gpu_va;
+};
+
+struct gk20a_cde_param {
+        u32 id;
+        u32 padding;
+        u64 value;
+};
+
+struct gk20a_cde_ctx {
+        struct gk20a *g;
+        struct platform_device *pdev;
+
+        /* channel related data */
+        struct channel_gk20a *ch;
+        struct vm_gk20a *vm;
+
+        /* buf converter configuration */
+        struct gk20a_cde_mem_desc mem[MAX_CDE_BUFS];
+        int num_bufs;
+
+        /* buffer patching params (where should patching be done) */
+        struct gk20a_cde_hdr_param params[MAX_CDE_PARAMS];
+        int num_params;
+
+        /* storage for user space parameter values */
+        u32 user_param_values[MAX_CDE_USER_PARAMS];
+
+        u64 src_smmu_addr;
+        u32 src_param_offset;
+        u32 src_param_lines;
+
+        u64 src_vaddr;
+
+        u64 dest_vaddr;
+        u64 dest_size;
+
+        u32 obj_ids[MAX_CDE_OBJ_IDS];
+        int num_obj_ids;
+
+        u64 backing_store_vaddr;
+
+        struct nvhost_gpfifo *init_cmd;
+        int init_cmd_num_entries;
+
+        struct nvhost_gpfifo *convert_cmd;
+        int convert_cmd_num_entries;
+
+        struct kobj_attribute attr;
+};
+
+struct gk20a_cde_app {
+        bool initialised;
+        struct mutex mutex;
+        struct vm_gk20a *vm;
+
+        struct gk20a_cde_ctx cde_ctx[1];
+        int cde_ctx_ptr;
+};
+
+int gk20a_cde_destroy(struct gk20a *g);
+int gk20a_init_cde_support(struct gk20a *g);
+int gk20a_cde_reload(struct gk20a *g);
+int gk20a_cde_convert(struct gk20a *g, u32 src_fd, u32 dst_fd,
+                      s32 dst_kind, u64 dst_word_offset,
+                      u32 dst_size, struct nvhost_fence *fence,
+                      u32 __flags, struct gk20a_cde_param *params,
+                      int num_params, struct gk20a_fence **fence_out);
+
+#endif
diff --git a/drivers/gpu/nvgpu/gk20a/gk20a.c b/drivers/gpu/nvgpu/gk20a/gk20a.c
index fa6e0cce..2975798f 100644
--- a/drivers/gpu/nvgpu/gk20a/gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/gk20a.c
@@ -976,6 +976,7 @@ static int gk20a_pm_finalize_poweron(struct device *dev)
                 goto done;
         }
 
+
         gk20a_channel_resume(g);
         set_user_nice(current, nice_value);
 
@@ -983,6 +984,8 @@ static int gk20a_pm_finalize_poweron(struct device *dev)
 
         trace_gk20a_finalize_poweron_done(dev_name(dev));
 
+        if (IS_ENABLED(CONFIG_GK20A_CDE))
+                gk20a_init_cde_support(g);
 done:
         return err;
 }
diff --git a/drivers/gpu/nvgpu/gk20a/gk20a.h b/drivers/gpu/nvgpu/gk20a/gk20a.h
index 05ed9270..b2ecade5 100644
--- a/drivers/gpu/nvgpu/gk20a/gk20a.h
+++ b/drivers/gpu/nvgpu/gk20a/gk20a.h
@@ -48,6 +48,7 @@ struct acr_gm20b;
 #include "therm_gk20a.h"
 #include "platform_gk20a.h"
 #include "gm20b/acr_gm20b.h"
+#include "cde_gk20a.h"
 
 extern struct platform_device tegra_gk20a_device;
 
@@ -356,6 +357,8 @@ struct gk20a {
         struct gk20a_scale_profile *scale_profile;
 
         struct device_dma_parameters dma_parms;
+
+        struct gk20a_cde_app cde_app;
 };
 
 static inline unsigned long gk20a_get_gr_idle_timeout(struct gk20a *g)
@@ -422,6 +425,7 @@ enum gk20a_dbg_categories {
         gpu_dbg_clk     = BIT(7),  /* gk20a clk */
         gpu_dbg_map     = BIT(8),  /* mem mappings */
         gpu_dbg_gpu_dbg = BIT(9),  /* gpu debugger/profiler */
+        gpu_dbg_cde     = BIT(10), /* cde info messages */
         gpu_dbg_mem     = BIT(31), /* memory accesses, very verbose */
 };