OSDN Git Service

Merge Android R
[android-x86/external-minigbm.git] / virtio_gpu.c
index 205f53a..ed67693 100644 (file)
@@ -4,18 +4,19 @@
  * found in the LICENSE file.
  */
 
+#include <assert.h>
 #include <errno.h>
 #include <stdint.h>
 #include <stdio.h>
 #include <string.h>
 #include <sys/mman.h>
-#include <virtgpu_drm.h>
 #include <xf86drm.h>
 
 #include "drv_priv.h"
 #include "helpers.h"
 #include "util.h"
 #include "virgl_hw.h"
+#include "virtgpu_drm.h"
 
 #ifndef PAGE_SIZE
 #define PAGE_SIZE 0x1000
 #define MESA_LLVMPIPE_TILE_ORDER 6
 #define MESA_LLVMPIPE_TILE_SIZE (1 << MESA_LLVMPIPE_TILE_ORDER)
 
+struct feature {
+       uint64_t feature;
+       const char *name;
+       uint32_t enabled;
+};
+
+enum feature_id {
+       feat_3d,
+       feat_capset_fix,
+       feat_max,
+};
+
+#define FEATURE(x)                                                                                 \
+       (struct feature)                                                                           \
+       {                                                                                          \
+               x, #x, 0                                                                           \
+       }
+
+static struct feature features[] = { FEATURE(VIRTGPU_PARAM_3D_FEATURES),
+                                    FEATURE(VIRTGPU_PARAM_CAPSET_QUERY_FIX) };
+
 static const uint32_t render_target_formats[] = { DRM_FORMAT_ABGR8888, DRM_FORMAT_ARGB8888,
                                                  DRM_FORMAT_RGB565, DRM_FORMAT_XBGR8888,
                                                  DRM_FORMAT_XRGB8888 };
 
-static const uint32_t dumb_texture_source_formats[] = { DRM_FORMAT_R8, DRM_FORMAT_YVU420,
-                                                       DRM_FORMAT_YVU420_ANDROID };
+static const uint32_t dumb_texture_source_formats[] = {
+       DRM_FORMAT_R8,   DRM_FORMAT_R16,  DRM_FORMAT_YVU420,
+       DRM_FORMAT_NV12, DRM_FORMAT_NV21, DRM_FORMAT_YVU420_ANDROID
+};
 
-static const uint32_t texture_source_formats[] = { DRM_FORMAT_NV12, DRM_FORMAT_R8, DRM_FORMAT_RG88,
-                                                  DRM_FORMAT_YVU420_ANDROID };
+static const uint32_t texture_source_formats[] = { DRM_FORMAT_NV12, DRM_FORMAT_NV21,
+                                                  DRM_FORMAT_R8,   DRM_FORMAT_R16,
+                                                  DRM_FORMAT_RG88, DRM_FORMAT_YVU420_ANDROID };
 
 struct virtio_gpu_priv {
-       int has_3d;
+       int caps_is_v2;
+       union virgl_caps caps;
+       int host_gbm_enabled;
 };
 
 static uint32_t translate_format(uint32_t drm_fourcc)
 {
        switch (drm_fourcc) {
+       case DRM_FORMAT_BGR888:
+       case DRM_FORMAT_RGB888:
+               return VIRGL_FORMAT_R8G8B8_UNORM;
        case DRM_FORMAT_XRGB8888:
                return VIRGL_FORMAT_B8G8R8X8_UNORM;
        case DRM_FORMAT_ARGB8888:
@@ -50,6 +80,8 @@ static uint32_t translate_format(uint32_t drm_fourcc)
                return VIRGL_FORMAT_R8G8B8X8_UNORM;
        case DRM_FORMAT_ABGR8888:
                return VIRGL_FORMAT_R8G8B8A8_UNORM;
+       case DRM_FORMAT_ABGR16161616F:
+               return VIRGL_FORMAT_R16G16B16A16_UNORM;
        case DRM_FORMAT_RGB565:
                return VIRGL_FORMAT_B5G6R5_UNORM;
        case DRM_FORMAT_R8:
@@ -58,6 +90,8 @@ static uint32_t translate_format(uint32_t drm_fourcc)
                return VIRGL_FORMAT_R8G8_UNORM;
        case DRM_FORMAT_NV12:
                return VIRGL_FORMAT_NV12;
+       case DRM_FORMAT_NV21:
+               return VIRGL_FORMAT_NV21;
        case DRM_FORMAT_YVU420:
        case DRM_FORMAT_YVU420_ANDROID:
                return VIRGL_FORMAT_YV12;
@@ -66,15 +100,303 @@ static uint32_t translate_format(uint32_t drm_fourcc)
        }
 }
 
+static bool virtio_gpu_bitmask_supports_format(struct virgl_supported_format_mask *supported,
+                                              uint32_t drm_format)
+{
+       uint32_t virgl_format = translate_format(drm_format);
+       if (!virgl_format) {
+               return false;
+       }
+
+       uint32_t bitmask_index = virgl_format / 32;
+       uint32_t bit_index = virgl_format % 32;
+       return supported->bitmask[bitmask_index] & (1 << bit_index);
+}
+
+// The metadata generated here for emulated buffers is slightly different than the metadata
+// generated by drv_bo_from_format. In order to simplify transfers in the flush and invalidate
+// functions below, the emulated buffers are oversized. For example, ignoring stride alignment
+// requirements to demonstrate, a 6x6 YUV420 image buffer might have the following layout from
+// drv_bo_from_format:
+//
+// | Y | Y | Y | Y | Y | Y |
+// | Y | Y | Y | Y | Y | Y |
+// | Y | Y | Y | Y | Y | Y |
+// | Y | Y | Y | Y | Y | Y |
+// | Y | Y | Y | Y | Y | Y |
+// | Y | Y | Y | Y | Y | Y |
+// | U | U | U | U | U | U |
+// | U | U | U | V | V | V |
+// | V | V | V | V | V | V |
+//
+// where each plane immediately follows the previous plane in memory. This layout makes it
+// difficult to compute the transfers needed for example when the middle 2x2 region of the
+// image is locked and needs to be flushed/invalidated.
+//
+// Emulated multi-plane buffers instead have a layout of:
+//
+// | Y | Y | Y | Y | Y | Y |
+// | Y | Y | Y | Y | Y | Y |
+// | Y | Y | Y | Y | Y | Y |
+// | Y | Y | Y | Y | Y | Y |
+// | Y | Y | Y | Y | Y | Y |
+// | Y | Y | Y | Y | Y | Y |
+// | U | U | U |   |   |   |
+// | U | U | U |   |   |   |
+// | U | U | U |   |   |   |
+// | V | V | V |   |   |   |
+// | V | V | V |   |   |   |
+// | V | V | V |   |   |   |
+//
+// where each plane is placed as a sub-image (albeit with a very large stride) in order to
+// simplify transfers into 3 sub-image transfers for the above example.
+//
+// Additional note: the V-plane is not placed to the right of the U-plane due to some
+// observed failures in media framework code which assumes the V-plane is not
+// "row-interlaced" with the U-plane.
+static void virtio_gpu_get_emulated_metadata(const struct bo *bo, struct bo_metadata *metadata)
+{
+       uint32_t y_plane_height;
+       uint32_t c_plane_height;
+       uint32_t original_width = bo->meta.width;
+       uint32_t original_height = bo->meta.height;
+
+       metadata->format = DRM_FORMAT_R8;
+       switch (bo->meta.format) {
+       case DRM_FORMAT_NV12:
+       case DRM_FORMAT_NV21:
+               // Bi-planar
+               metadata->num_planes = 2;
+
+               y_plane_height = original_height;
+               c_plane_height = DIV_ROUND_UP(original_height, 2);
+
+               metadata->width = original_width;
+               metadata->height = y_plane_height + c_plane_height;
+
+               // Y-plane (full resolution)
+               metadata->strides[0] = metadata->width;
+               metadata->offsets[0] = 0;
+               metadata->sizes[0] = metadata->width * y_plane_height;
+
+               // CbCr-plane  (half resolution, interleaved, placed below Y-plane)
+               metadata->strides[1] = metadata->width;
+               metadata->offsets[1] = metadata->offsets[0] + metadata->sizes[0];
+               metadata->sizes[1] = metadata->width * c_plane_height;
+
+               metadata->total_size = metadata->width * metadata->height;
+               break;
+       case DRM_FORMAT_YVU420:
+       case DRM_FORMAT_YVU420_ANDROID:
+               // Tri-planar
+               metadata->num_planes = 3;
+
+               y_plane_height = original_height;
+               c_plane_height = DIV_ROUND_UP(original_height, 2);
+
+               metadata->width = ALIGN(original_width, 32);
+               metadata->height = y_plane_height + (2 * c_plane_height);
+
+               // Y-plane (full resolution)
+               metadata->strides[0] = metadata->width;
+               metadata->offsets[0] = 0;
+               metadata->sizes[0] = metadata->width * original_height;
+
+               // Cb-plane (half resolution, placed below Y-plane)
+               metadata->strides[1] = metadata->width;
+               metadata->offsets[1] = metadata->offsets[0] + metadata->sizes[0];
+               metadata->sizes[1] = metadata->width * c_plane_height;
+
+               // Cr-plane (half resolution, placed below Cb-plane)
+               metadata->strides[2] = metadata->width;
+               metadata->offsets[2] = metadata->offsets[1] + metadata->sizes[1];
+               metadata->sizes[2] = metadata->width * c_plane_height;
+
+               metadata->total_size = metadata->width * metadata->height;
+               break;
+       default:
+               break;
+       }
+}
+
+struct virtio_transfers_params {
+       size_t xfers_needed;
+       struct rectangle xfer_boxes[DRV_MAX_PLANES];
+};
+
+static void virtio_gpu_get_emulated_transfers_params(const struct bo *bo,
+                                                    const struct rectangle *transfer_box,
+                                                    struct virtio_transfers_params *xfer_params)
+{
+       uint32_t y_plane_height;
+       uint32_t c_plane_height;
+       struct bo_metadata emulated_metadata;
+
+       if (transfer_box->x == 0 && transfer_box->y == 0 && transfer_box->width == bo->meta.width &&
+           transfer_box->height == bo->meta.height) {
+               virtio_gpu_get_emulated_metadata(bo, &emulated_metadata);
+
+               xfer_params->xfers_needed = 1;
+               xfer_params->xfer_boxes[0].x = 0;
+               xfer_params->xfer_boxes[0].y = 0;
+               xfer_params->xfer_boxes[0].width = emulated_metadata.width;
+               xfer_params->xfer_boxes[0].height = emulated_metadata.height;
+
+               return;
+       }
+
+       switch (bo->meta.format) {
+       case DRM_FORMAT_NV12:
+       case DRM_FORMAT_NV21:
+               // Bi-planar
+               xfer_params->xfers_needed = 2;
+
+               y_plane_height = bo->meta.height;
+               c_plane_height = DIV_ROUND_UP(bo->meta.height, 2);
+
+               // Y-plane (full resolution)
+               xfer_params->xfer_boxes[0].x = transfer_box->x;
+               xfer_params->xfer_boxes[0].y = transfer_box->y;
+               xfer_params->xfer_boxes[0].width = transfer_box->width;
+               xfer_params->xfer_boxes[0].height = transfer_box->height;
+
+               // CbCr-plane (half resolution, interleaved, placed below Y-plane)
+               xfer_params->xfer_boxes[1].x = transfer_box->x;
+               xfer_params->xfer_boxes[1].y = transfer_box->y + y_plane_height;
+               xfer_params->xfer_boxes[1].width = transfer_box->width;
+               xfer_params->xfer_boxes[1].height = DIV_ROUND_UP(transfer_box->height, 2);
+
+               break;
+       case DRM_FORMAT_YVU420:
+       case DRM_FORMAT_YVU420_ANDROID:
+               // Tri-planar
+               xfer_params->xfers_needed = 3;
+
+               y_plane_height = bo->meta.height;
+               c_plane_height = DIV_ROUND_UP(bo->meta.height, 2);
+
+               // Y-plane (full resolution)
+               xfer_params->xfer_boxes[0].x = transfer_box->x;
+               xfer_params->xfer_boxes[0].y = transfer_box->y;
+               xfer_params->xfer_boxes[0].width = transfer_box->width;
+               xfer_params->xfer_boxes[0].height = transfer_box->height;
+
+               // Cb-plane (half resolution, placed below Y-plane)
+               xfer_params->xfer_boxes[1].x = transfer_box->x;
+               xfer_params->xfer_boxes[1].y = transfer_box->y + y_plane_height;
+               xfer_params->xfer_boxes[1].width = DIV_ROUND_UP(transfer_box->width, 2);
+               xfer_params->xfer_boxes[1].height = DIV_ROUND_UP(transfer_box->height, 2);
+
+               // Cr-plane (half resolution, placed below Cb-plane)
+               xfer_params->xfer_boxes[2].x = transfer_box->x;
+               xfer_params->xfer_boxes[2].y = transfer_box->y + y_plane_height + c_plane_height;
+               xfer_params->xfer_boxes[2].width = DIV_ROUND_UP(transfer_box->width, 2);
+               xfer_params->xfer_boxes[2].height = DIV_ROUND_UP(transfer_box->height, 2);
+
+               break;
+       }
+}
+
+static bool virtio_gpu_supports_combination_natively(struct driver *drv, uint32_t drm_format,
+                                                    uint64_t use_flags)
+{
+       struct virtio_gpu_priv *priv = (struct virtio_gpu_priv *)drv->priv;
+
+       if (priv->caps.max_version == 0) {
+               return true;
+       }
+
+       if ((use_flags & BO_USE_RENDERING) &&
+           !virtio_gpu_bitmask_supports_format(&priv->caps.v1.render, drm_format)) {
+               return false;
+       }
+
+       if ((use_flags & BO_USE_TEXTURE) &&
+           !virtio_gpu_bitmask_supports_format(&priv->caps.v1.sampler, drm_format)) {
+               return false;
+       }
+
+       if ((use_flags & BO_USE_SCANOUT) && priv->caps_is_v2 &&
+           !virtio_gpu_bitmask_supports_format(&priv->caps.v2.scanout, drm_format)) {
+               return false;
+       }
+
+       return true;
+}
+
+// For virtio backends that do not support formats natively (e.g. multi-planar formats are not
+// supported in virglrenderer when gbm is unavailable on the host machine), whether or not the
+// format and usage combination can be handled as a blob (byte buffer).
+static bool virtio_gpu_supports_combination_through_emulation(struct driver *drv,
+                                                             uint32_t drm_format,
+                                                             uint64_t use_flags)
+{
+       struct virtio_gpu_priv *priv = (struct virtio_gpu_priv *)drv->priv;
+
+       // Only enable emulation on non-gbm virtio backends.
+       if (priv->host_gbm_enabled) {
+               return false;
+       }
+
+       if (use_flags & (BO_USE_RENDERING | BO_USE_SCANOUT)) {
+               return false;
+       }
+
+       if (!virtio_gpu_supports_combination_natively(drv, DRM_FORMAT_R8, use_flags)) {
+               return false;
+       }
+
+       return drm_format == DRM_FORMAT_NV12 || drm_format == DRM_FORMAT_NV21 ||
+              drm_format == DRM_FORMAT_YVU420 || drm_format == DRM_FORMAT_YVU420_ANDROID;
+}
+
+// Adds the given buffer combination to the list of supported buffer combinations if the
+// combination is supported by the virtio backend.
+static void virtio_gpu_add_combination(struct driver *drv, uint32_t drm_format,
+                                      struct format_metadata *metadata, uint64_t use_flags)
+{
+       struct virtio_gpu_priv *priv = (struct virtio_gpu_priv *)drv->priv;
+
+       if (features[feat_3d].enabled && priv->caps.max_version >= 1) {
+               if ((use_flags & BO_USE_SCANOUT) && priv->caps_is_v2 &&
+                   !virtio_gpu_supports_combination_natively(drv, drm_format, use_flags)) {
+                       drv_log("Scanout format: %d\n", drm_format);
+                       use_flags &= ~BO_USE_SCANOUT;
+               }
+
+               if (!virtio_gpu_supports_combination_natively(drv, drm_format, use_flags) &&
+                   !virtio_gpu_supports_combination_through_emulation(drv, drm_format,
+                                                                      use_flags)) {
+                       drv_log("Skipping unsupported combination format:%d\n", drm_format);
+                       return;
+               }
+       }
+
+       drv_add_combination(drv, drm_format, metadata, use_flags);
+}
+
+// Adds each given buffer combination to the list of supported buffer combinations if the
+// combination supported by the virtio backend.
+static void virtio_gpu_add_combinations(struct driver *drv, const uint32_t *drm_formats,
+                                       uint32_t num_formats, struct format_metadata *metadata,
+                                       uint64_t use_flags)
+{
+       uint32_t i;
+
+       for (i = 0; i < num_formats; i++) {
+               virtio_gpu_add_combination(drv, drm_formats[i], metadata, use_flags);
+       }
+}
+
 static int virtio_dumb_bo_create(struct bo *bo, uint32_t width, uint32_t height, uint32_t format,
                                 uint64_t use_flags)
 {
-       if (bo->format != DRM_FORMAT_R8) {
+       if (bo->meta.format != DRM_FORMAT_R8) {
                width = ALIGN(width, MESA_LLVMPIPE_TILE_SIZE);
                height = ALIGN(height, MESA_LLVMPIPE_TILE_SIZE);
        }
 
-       return drv_dumb_bo_create(bo, width, height, format, use_flags);
+       return drv_dumb_bo_create_ex(bo, width, height, format, use_flags, BO_QUIRK_DUMB32BPP);
 }
 
 static inline void handle_flag(uint64_t *flag, uint64_t check_flag, uint32_t *bind,
@@ -94,7 +416,19 @@ static uint32_t use_flags_to_bind(uint64_t use_flags)
        handle_flag(&use_flags, BO_USE_TEXTURE, &bind, VIRGL_BIND_SAMPLER_VIEW);
        handle_flag(&use_flags, BO_USE_RENDERING, &bind, VIRGL_BIND_RENDER_TARGET);
        handle_flag(&use_flags, BO_USE_SCANOUT, &bind, VIRGL_BIND_SCANOUT);
-       // TODO (b/12983436): handle other use flags.
+       handle_flag(&use_flags, BO_USE_CURSOR, &bind, VIRGL_BIND_CURSOR);
+       handle_flag(&use_flags, BO_USE_LINEAR, &bind, VIRGL_BIND_LINEAR);
+
+       handle_flag(&use_flags, BO_USE_SW_READ_OFTEN, &bind, VIRGL_BIND_LINEAR);
+       handle_flag(&use_flags, BO_USE_SW_READ_RARELY, &bind, VIRGL_BIND_LINEAR);
+       handle_flag(&use_flags, BO_USE_SW_WRITE_OFTEN, &bind, VIRGL_BIND_LINEAR);
+       handle_flag(&use_flags, BO_USE_SW_WRITE_RARELY, &bind, VIRGL_BIND_LINEAR);
+
+       // All host drivers only support linear camera buffer formats. If
+       // that changes, this will need to be modified.
+       handle_flag(&use_flags, BO_USE_CAMERA_READ, &bind, VIRGL_BIND_LINEAR);
+       handle_flag(&use_flags, BO_USE_CAMERA_WRITE, &bind, VIRGL_BIND_LINEAR);
+
        if (use_flags) {
                drv_log("Unhandled bo use flag: %llx\n", (unsigned long long)use_flags);
        }
@@ -106,11 +440,30 @@ static int virtio_virgl_bo_create(struct bo *bo, uint32_t width, uint32_t height
                                  uint64_t use_flags)
 {
        int ret;
+       size_t i;
        uint32_t stride;
        struct drm_virtgpu_resource_create res_create;
+       struct bo_metadata emulated_metadata;
 
-       stride = drv_stride_from_format(format, width, 0);
-       drv_bo_from_format(bo, stride, height, format);
+       if (virtio_gpu_supports_combination_natively(bo->drv, format, use_flags)) {
+               stride = drv_stride_from_format(format, width, 0);
+               drv_bo_from_format(bo, stride, height, format);
+       } else {
+               assert(
+                   virtio_gpu_supports_combination_through_emulation(bo->drv, format, use_flags));
+
+               virtio_gpu_get_emulated_metadata(bo, &emulated_metadata);
+
+               format = emulated_metadata.format;
+               width = emulated_metadata.width;
+               height = emulated_metadata.height;
+               for (i = 0; i < emulated_metadata.num_planes; i++) {
+                       bo->meta.strides[i] = emulated_metadata.strides[i];
+                       bo->meta.offsets[i] = emulated_metadata.offsets[i];
+                       bo->meta.sizes[i] = emulated_metadata.sizes[i];
+               }
+               bo->meta.total_size = emulated_metadata.total_size;
+       }
 
        /*
         * Setting the target is intended to ensure this resource gets bound as a 2D
@@ -133,14 +486,14 @@ static int virtio_virgl_bo_create(struct bo *bo, uint32_t width, uint32_t height
        res_create.last_level = 0;
        res_create.nr_samples = 0;
 
-       res_create.size = ALIGN(bo->total_size, PAGE_SIZE); // PAGE_SIZE = 0x1000
+       res_create.size = ALIGN(bo->meta.total_size, PAGE_SIZE); // PAGE_SIZE = 0x1000
        ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_RESOURCE_CREATE, &res_create);
        if (ret) {
                drv_log("DRM_IOCTL_VIRTGPU_RESOURCE_CREATE failed with %s\n", strerror(errno));
                return ret;
        }
 
-       for (uint32_t plane = 0; plane < bo->num_planes; plane++)
+       for (uint32_t plane = 0; plane < bo->meta.num_planes; plane++)
                bo->handles[plane].u32 = res_create.bo_handle;
 
        return 0;
@@ -160,54 +513,131 @@ static void *virtio_virgl_bo_map(struct bo *bo, struct vma *vma, size_t plane, u
                return MAP_FAILED;
        }
 
-       vma->length = bo->total_size;
-       return mmap(0, bo->total_size, drv_get_prot(map_flags), MAP_SHARED, bo->drv->fd,
+       vma->length = bo->meta.total_size;
+       return mmap(0, bo->meta.total_size, drv_get_prot(map_flags), MAP_SHARED, bo->drv->fd,
                    gem_map.offset);
 }
 
-static int virtio_gpu_init(struct driver *drv)
+static int virtio_gpu_get_caps(struct driver *drv, union virgl_caps *caps, int *caps_is_v2)
 {
        int ret;
-       struct virtio_gpu_priv *priv;
-       struct drm_virtgpu_getparam args;
-
-       priv = calloc(1, sizeof(*priv));
-       drv->priv = priv;
+       struct drm_virtgpu_get_caps cap_args;
+
+       *caps_is_v2 = 0;
+       memset(&cap_args, 0, sizeof(cap_args));
+       cap_args.addr = (unsigned long long)caps;
+       if (features[feat_capset_fix].enabled) {
+               *caps_is_v2 = 1;
+               cap_args.cap_set_id = 2;
+               cap_args.size = sizeof(union virgl_caps);
+       } else {
+               cap_args.cap_set_id = 1;
+               cap_args.size = sizeof(struct virgl_caps_v1);
+       }
 
-       memset(&args, 0, sizeof(args));
-       args.param = VIRTGPU_PARAM_3D_FEATURES;
-       args.value = (uint64_t)(uintptr_t)&priv->has_3d;
-       ret = drmIoctl(drv->fd, DRM_IOCTL_VIRTGPU_GETPARAM, &args);
+       ret = drmIoctl(drv->fd, DRM_IOCTL_VIRTGPU_GET_CAPS, &cap_args);
        if (ret) {
-               drv_log("virtio 3D acceleration is not available\n");
-               /* Be paranoid */
-               priv->has_3d = 0;
+               drv_log("DRM_IOCTL_VIRTGPU_GET_CAPS failed with %s\n", strerror(errno));
+               *caps_is_v2 = 0;
+
+               // Fallback to v1
+               cap_args.cap_set_id = 1;
+               cap_args.size = sizeof(struct virgl_caps_v1);
+
+               ret = drmIoctl(drv->fd, DRM_IOCTL_VIRTGPU_GET_CAPS, &cap_args);
+               if (ret) {
+                       drv_log("DRM_IOCTL_VIRTGPU_GET_CAPS failed with %s\n", strerror(errno));
+               }
        }
 
-       /* This doesn't mean host can scanout everything, it just means host
-        * hypervisor can show it. */
-       drv_add_combinations(drv, render_target_formats, ARRAY_SIZE(render_target_formats),
-                            &LINEAR_METADATA, BO_USE_RENDER_MASK | BO_USE_SCANOUT);
+       return ret;
+}
 
-       if (priv->has_3d) {
-               drv_add_combinations(drv, texture_source_formats,
-                                    ARRAY_SIZE(texture_source_formats), &LINEAR_METADATA,
-                                    BO_USE_TEXTURE_MASK);
+static void virtio_gpu_init_features_and_caps(struct driver *drv)
+{
+       struct virtio_gpu_priv *priv = (struct virtio_gpu_priv *)drv->priv;
+
+       for (uint32_t i = 0; i < ARRAY_SIZE(features); i++) {
+               struct drm_virtgpu_getparam params = { 0 };
+
+               params.param = features[i].feature;
+               params.value = (uint64_t)(uintptr_t)&features[i].enabled;
+               int ret = drmIoctl(drv->fd, DRM_IOCTL_VIRTGPU_GETPARAM, &params);
+               if (ret)
+                       drv_log("DRM_IOCTL_VIRTGPU_GET_PARAM failed with %s\n", strerror(errno));
+       }
+
+       if (features[feat_3d].enabled) {
+               virtio_gpu_get_caps(drv, &priv->caps, &priv->caps_is_v2);
+       }
+
+       // Multi-planar formats are currently only supported in virglrenderer through gbm.
+       priv->host_gbm_enabled =
+           virtio_gpu_supports_combination_natively(drv, DRM_FORMAT_NV12, BO_USE_TEXTURE);
+}
+
+static int virtio_gpu_init(struct driver *drv)
+{
+       struct virtio_gpu_priv *priv;
+
+       priv = calloc(1, sizeof(*priv));
+       drv->priv = priv;
+
+       virtio_gpu_init_features_and_caps(drv);
+
+       if (features[feat_3d].enabled) {
+               /* This doesn't mean host can scanout everything, it just means host
+                * hypervisor can show it. */
+               virtio_gpu_add_combinations(drv, render_target_formats,
+                                           ARRAY_SIZE(render_target_formats), &LINEAR_METADATA,
+                                           BO_USE_RENDER_MASK | BO_USE_SCANOUT);
+               virtio_gpu_add_combinations(drv, texture_source_formats,
+                                           ARRAY_SIZE(texture_source_formats), &LINEAR_METADATA,
+                                           BO_USE_TEXTURE_MASK);
        } else {
-               drv_add_combinations(drv, dumb_texture_source_formats,
-                                    ARRAY_SIZE(dumb_texture_source_formats), &LINEAR_METADATA,
-                                    BO_USE_TEXTURE_MASK);
-               drv_add_combination(drv, DRM_FORMAT_NV12, &LINEAR_METADATA,
-                                   BO_USE_SW_MASK | BO_USE_LINEAR);
+               /* Virtio primary plane only allows this format. */
+               virtio_gpu_add_combination(drv, DRM_FORMAT_XRGB8888, &LINEAR_METADATA,
+                                          BO_USE_RENDER_MASK | BO_USE_SCANOUT);
+               /* Virtio cursor plane only allows this format and Chrome cannot live without
+                * ARGB888 renderable format. */
+               virtio_gpu_add_combination(drv, DRM_FORMAT_ARGB8888, &LINEAR_METADATA,
+                                          BO_USE_RENDER_MASK | BO_USE_CURSOR);
+               /* Android needs more, but they cannot be bound as scanouts anymore after
+                * "drm/virtio: fix DRM_FORMAT_* handling" */
+               virtio_gpu_add_combinations(drv, render_target_formats,
+                                           ARRAY_SIZE(render_target_formats), &LINEAR_METADATA,
+                                           BO_USE_RENDER_MASK);
+               virtio_gpu_add_combinations(drv, dumb_texture_source_formats,
+                                           ARRAY_SIZE(dumb_texture_source_formats),
+                                           &LINEAR_METADATA, BO_USE_TEXTURE_MASK);
+               virtio_gpu_add_combination(drv, DRM_FORMAT_NV12, &LINEAR_METADATA,
+                                          BO_USE_SW_MASK | BO_USE_LINEAR);
+               virtio_gpu_add_combination(drv, DRM_FORMAT_NV21, &LINEAR_METADATA,
+                                          BO_USE_SW_MASK | BO_USE_LINEAR);
        }
 
        /* Android CTS tests require this. */
-       drv_add_combination(drv, DRM_FORMAT_BGR888, &LINEAR_METADATA, BO_USE_SW_MASK);
+       virtio_gpu_add_combination(drv, DRM_FORMAT_RGB888, &LINEAR_METADATA, BO_USE_SW_MASK);
+       virtio_gpu_add_combination(drv, DRM_FORMAT_BGR888, &LINEAR_METADATA, BO_USE_SW_MASK);
+       virtio_gpu_add_combination(drv, DRM_FORMAT_ABGR16161616F, &LINEAR_METADATA,
+                                  BO_USE_SW_MASK | BO_USE_TEXTURE_MASK);
 
        drv_modify_combination(drv, DRM_FORMAT_NV12, &LINEAR_METADATA,
-                              BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE);
+                              BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE | BO_USE_HW_VIDEO_DECODER |
+                                  BO_USE_HW_VIDEO_ENCODER);
+       drv_modify_combination(drv, DRM_FORMAT_NV21, &LINEAR_METADATA,
+                              BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE | BO_USE_HW_VIDEO_DECODER |
+                                  BO_USE_HW_VIDEO_ENCODER);
+       drv_modify_combination(drv, DRM_FORMAT_YVU420, &LINEAR_METADATA,
+                              BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE | BO_USE_HW_VIDEO_DECODER |
+                                  BO_USE_HW_VIDEO_ENCODER | BO_USE_RENDERSCRIPT);
+       drv_modify_combination(drv, DRM_FORMAT_YVU420_ANDROID, &LINEAR_METADATA,
+                              BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE | BO_USE_HW_VIDEO_DECODER |
+                                  BO_USE_HW_VIDEO_ENCODER | BO_USE_RENDERSCRIPT);
        drv_modify_combination(drv, DRM_FORMAT_R8, &LINEAR_METADATA,
-                              BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE);
+                              BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE | BO_USE_HW_VIDEO_DECODER);
+       drv_modify_combination(drv, DRM_FORMAT_R16, &LINEAR_METADATA,
+                              BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE | BO_USE_HW_VIDEO_DECODER);
 
        return drv_modify_linear_combinations(drv);
 }
@@ -221,8 +651,7 @@ static void virtio_gpu_close(struct driver *drv)
 static int virtio_gpu_bo_create(struct bo *bo, uint32_t width, uint32_t height, uint32_t format,
                                uint64_t use_flags)
 {
-       struct virtio_gpu_priv *priv = (struct virtio_gpu_priv *)bo->drv->priv;
-       if (priv->has_3d)
+       if (features[feat_3d].enabled)
                return virtio_virgl_bo_create(bo, width, height, format, use_flags);
        else
                return virtio_dumb_bo_create(bo, width, height, format, use_flags);
@@ -230,8 +659,7 @@ static int virtio_gpu_bo_create(struct bo *bo, uint32_t width, uint32_t height,
 
 static int virtio_gpu_bo_destroy(struct bo *bo)
 {
-       struct virtio_gpu_priv *priv = (struct virtio_gpu_priv *)bo->drv->priv;
-       if (priv->has_3d)
+       if (features[feat_3d].enabled)
                return drv_gem_bo_destroy(bo);
        else
                return drv_dumb_bo_destroy(bo);
@@ -239,8 +667,7 @@ static int virtio_gpu_bo_destroy(struct bo *bo)
 
 static void *virtio_gpu_bo_map(struct bo *bo, struct vma *vma, size_t plane, uint32_t map_flags)
 {
-       struct virtio_gpu_priv *priv = (struct virtio_gpu_priv *)bo->drv->priv;
-       if (priv->has_3d)
+       if (features[feat_3d].enabled)
                return virtio_virgl_bo_map(bo, vma, plane, map_flags);
        else
                return drv_dumb_bo_map(bo, vma, plane, map_flags);
@@ -249,23 +676,70 @@ static void *virtio_gpu_bo_map(struct bo *bo, struct vma *vma, size_t plane, uin
 static int virtio_gpu_bo_invalidate(struct bo *bo, struct mapping *mapping)
 {
        int ret;
+       size_t i;
        struct drm_virtgpu_3d_transfer_from_host xfer;
+       struct drm_virtgpu_3d_wait waitcmd;
+       struct virtio_transfers_params xfer_params;
        struct virtio_gpu_priv *priv = (struct virtio_gpu_priv *)bo->drv->priv;
 
-       if (!priv->has_3d)
+       if (!features[feat_3d].enabled)
+               return 0;
+
+       // Invalidate is only necessary if the host writes to the buffer.
+       if ((bo->meta.use_flags & (BO_USE_RENDERING | BO_USE_CAMERA_WRITE |
+                                  BO_USE_HW_VIDEO_ENCODER | BO_USE_HW_VIDEO_DECODER)) == 0)
                return 0;
 
        memset(&xfer, 0, sizeof(xfer));
        xfer.bo_handle = mapping->vma->handle;
-       xfer.box.x = mapping->rect.x;
-       xfer.box.y = mapping->rect.y;
-       xfer.box.w = mapping->rect.width;
-       xfer.box.h = mapping->rect.height;
-       xfer.box.d = 1;
 
-       ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_TRANSFER_FROM_HOST, &xfer);
+       if ((bo->meta.use_flags & BO_USE_RENDERING) == 0) {
+               // Unfortunately, the kernel doesn't actually pass the guest layer_stride
+               // and guest stride to the host (compare virtio_gpu.h and virtgpu_drm.h).
+               // For gbm based resources, we can work around this by using the level field
+               // to pass the stride to virglrenderer's gbm transfer code. However, we need
+               // to avoid doing this for resources which don't rely on that transfer code,
+               // which is resources with the BO_USE_RENDERING flag set.
+               // TODO(b/145993887): Send also stride when the patches are landed
+               if (priv->host_gbm_enabled) {
+                       xfer.level = bo->meta.strides[0];
+               }
+       }
+
+       if (virtio_gpu_supports_combination_natively(bo->drv, bo->meta.format,
+                                                    bo->meta.use_flags)) {
+               xfer_params.xfers_needed = 1;
+               xfer_params.xfer_boxes[0] = mapping->rect;
+       } else {
+               assert(virtio_gpu_supports_combination_through_emulation(bo->drv, bo->meta.format,
+                                                                        bo->meta.use_flags));
+
+               virtio_gpu_get_emulated_transfers_params(bo, &mapping->rect, &xfer_params);
+       }
+
+       for (i = 0; i < xfer_params.xfers_needed; i++) {
+               xfer.box.x = xfer_params.xfer_boxes[i].x;
+               xfer.box.y = xfer_params.xfer_boxes[i].y;
+               xfer.box.w = xfer_params.xfer_boxes[i].width;
+               xfer.box.h = xfer_params.xfer_boxes[i].height;
+               xfer.box.d = 1;
+
+               ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_TRANSFER_FROM_HOST, &xfer);
+               if (ret) {
+                       drv_log("DRM_IOCTL_VIRTGPU_TRANSFER_FROM_HOST failed with %s\n",
+                               strerror(errno));
+                       return -errno;
+               }
+       }
+
+       // The transfer needs to complete before invalidate returns so that any host changes
+       // are visible and to ensure the host doesn't overwrite subsequent guest changes.
+       // TODO(b/136733358): Support returning fences from transfers
+       memset(&waitcmd, 0, sizeof(waitcmd));
+       waitcmd.handle = mapping->vma->handle;
+       ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_WAIT, &waitcmd);
        if (ret) {
-               drv_log("DRM_IOCTL_VIRTGPU_TRANSFER_FROM_HOST failed with %s\n", strerror(errno));
+               drv_log("DRM_IOCTL_VIRTGPU_WAIT failed with %s\n", strerror(errno));
                return -errno;
        }
 
@@ -275,10 +749,13 @@ static int virtio_gpu_bo_invalidate(struct bo *bo, struct mapping *mapping)
 static int virtio_gpu_bo_flush(struct bo *bo, struct mapping *mapping)
 {
        int ret;
+       size_t i;
        struct drm_virtgpu_3d_transfer_to_host xfer;
+       struct drm_virtgpu_3d_wait waitcmd;
+       struct virtio_transfers_params xfer_params;
        struct virtio_gpu_priv *priv = (struct virtio_gpu_priv *)bo->drv->priv;
 
-       if (!priv->has_3d)
+       if (!features[feat_3d].enabled)
                return 0;
 
        if (!(mapping->vma->map_flags & BO_MAP_WRITE))
@@ -286,16 +763,53 @@ static int virtio_gpu_bo_flush(struct bo *bo, struct mapping *mapping)
 
        memset(&xfer, 0, sizeof(xfer));
        xfer.bo_handle = mapping->vma->handle;
-       xfer.box.x = mapping->rect.x;
-       xfer.box.y = mapping->rect.y;
-       xfer.box.w = mapping->rect.width;
-       xfer.box.h = mapping->rect.height;
-       xfer.box.d = 1;
 
-       ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_TRANSFER_TO_HOST, &xfer);
-       if (ret) {
-               drv_log("DRM_IOCTL_VIRTGPU_TRANSFER_TO_HOST failed with %s\n", strerror(errno));
-               return -errno;
+       // Unfortunately, the kernel doesn't actually pass the guest layer_stride and
+       // guest stride to the host (compare virtio_gpu.h and virtgpu_drm.h). We can use
+       // the level to work around this.
+       if (priv->host_gbm_enabled) {
+               xfer.level = bo->meta.strides[0];
+       }
+
+       if (virtio_gpu_supports_combination_natively(bo->drv, bo->meta.format,
+                                                    bo->meta.use_flags)) {
+               xfer_params.xfers_needed = 1;
+               xfer_params.xfer_boxes[0] = mapping->rect;
+       } else {
+               assert(virtio_gpu_supports_combination_through_emulation(bo->drv, bo->meta.format,
+                                                                        bo->meta.use_flags));
+
+               virtio_gpu_get_emulated_transfers_params(bo, &mapping->rect, &xfer_params);
+       }
+
+       for (i = 0; i < xfer_params.xfers_needed; i++) {
+               xfer.box.x = xfer_params.xfer_boxes[i].x;
+               xfer.box.y = xfer_params.xfer_boxes[i].y;
+               xfer.box.w = xfer_params.xfer_boxes[i].width;
+               xfer.box.h = xfer_params.xfer_boxes[i].height;
+               xfer.box.d = 1;
+
+               ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_TRANSFER_TO_HOST, &xfer);
+               if (ret) {
+                       drv_log("DRM_IOCTL_VIRTGPU_TRANSFER_TO_HOST failed with %s\n",
+                               strerror(errno));
+                       return -errno;
+               }
+       }
+
+       // If the buffer is only accessed by the host GPU, then the flush is ordered
+       // with subsequent commands. However, if other host hardware can access the
+       // buffer, we need to wait for the transfer to complete for consistency.
+       // TODO(b/136733358): Support returning fences from transfers
+       if (bo->meta.use_flags & BO_USE_NON_GPU_HW) {
+               memset(&waitcmd, 0, sizeof(waitcmd));
+               waitcmd.handle = mapping->vma->handle;
+
+               ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_WAIT, &waitcmd);
+               if (ret) {
+                       drv_log("DRM_IOCTL_VIRTGPU_WAIT failed with %s\n", strerror(errno));
+                       return -errno;
+               }
        }
 
        return 0;
@@ -303,8 +817,6 @@ static int virtio_gpu_bo_flush(struct bo *bo, struct mapping *mapping)
 
 static uint32_t virtio_gpu_resolve_format(struct driver *drv, uint32_t format, uint64_t use_flags)
 {
-       struct virtio_gpu_priv *priv = (struct virtio_gpu_priv *)drv->priv;
-
        switch (format) {
        case DRM_FORMAT_FLEX_IMPLEMENTATION_DEFINED:
                /* Camera subsystem requires NV12. */
@@ -317,15 +829,46 @@ static uint32_t virtio_gpu_resolve_format(struct driver *drv, uint32_t format, u
                 * All of our host drivers prefer NV12 as their flexible media format.
                 * If that changes, this will need to be modified.
                 */
-               if (priv->has_3d)
+               if (features[feat_3d].enabled)
                        return DRM_FORMAT_NV12;
                else
-                       return DRM_FORMAT_YVU420;
+                       return DRM_FORMAT_YVU420_ANDROID;
        default:
                return format;
        }
 }
 
+static int virtio_gpu_resource_info(struct bo *bo, uint32_t strides[DRV_MAX_PLANES],
+                                   uint32_t offsets[DRV_MAX_PLANES])
+{
+       int ret;
+       struct drm_virtgpu_resource_info res_info;
+
+       if (!features[feat_3d].enabled)
+               return 0;
+
+       memset(&res_info, 0, sizeof(res_info));
+       res_info.bo_handle = bo->handles[0].u32;
+       ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_RESOURCE_INFO, &res_info);
+       if (ret) {
+               drv_log("DRM_IOCTL_VIRTGPU_RESOURCE_INFO failed with %s\n", strerror(errno));
+               return ret;
+       }
+
+       for (uint32_t plane = 0; plane < bo->meta.num_planes; plane++) {
+               /*
+                * Currently, kernel v4.14 (Betty) doesn't have the extended resource info
+                * ioctl.
+                */
+               if (res_info.strides[plane]) {
+                       strides[plane] = res_info.strides[plane];
+                       offsets[plane] = res_info.offsets[plane];
+               }
+       }
+
+       return 0;
+}
+
 const struct backend backend_virtio_gpu = {
        .name = "virtio_gpu",
        .init = virtio_gpu_init,
@@ -338,4 +881,5 @@ const struct backend backend_virtio_gpu = {
        .bo_invalidate = virtio_gpu_bo_invalidate,
        .bo_flush = virtio_gpu_bo_flush,
        .resolve_format = virtio_gpu_resolve_format,
+       .resource_info = virtio_gpu_resource_info,
 };