int amdgpu_bo_export(amdgpu_bo_handle bo,
enum amdgpu_bo_handle_type type,
uint32_t *shared_handle)
{
int r;
switch (type) {
case amdgpu_bo_handle_type_gem_flink_name:
r = amdgpu_bo_export_flink(bo);
if (r)
return r;
*shared_handle = bo->flink_name;
return 0;
case amdgpu_bo_handle_type_kms:
amdgpu_add_handle_to_table(bo);
*shared_handle = bo->handle;
return 0;
case amdgpu_bo_handle_type_dma_buf_fd:
amdgpu_add_handle_to_table(bo);
return drmPrimeHandleToFD(bo->dev->fd, bo->handle, DRM_CLOEXEC,
(int*)shared_handle);
}
return -EINVAL;
}
int radeon_gem_prime_share_bo(struct radeon_bo *bo, int *handle)
{
struct radeon_bo_gem *bo_gem = (struct radeon_bo_gem*)bo;
int ret;
ret = drmPrimeHandleToFD(bo_gem->base.bom->fd, bo->handle, DRM_CLOEXEC, handle);
return ret;
}
static int amdgpu_bo_export_flink(amdgpu_bo_handle bo)
{
struct drm_gem_flink flink;
int fd, dma_fd;
uint32_t handle;
int r;
fd = bo->dev->fd;
handle = bo->handle;
if (bo->flink_name)
return 0;
if (bo->dev->flink_fd != bo->dev->fd) {
r = drmPrimeHandleToFD(bo->dev->fd, bo->handle, DRM_CLOEXEC,
&dma_fd);
if (!r) {
r = drmPrimeFDToHandle(bo->dev->flink_fd, dma_fd, &handle);
close(dma_fd);
}
if (r)
return r;
fd = bo->dev->flink_fd;
}
memset(&flink, 0, sizeof(flink));
flink.handle = handle;
r = drmIoctl(fd, DRM_IOCTL_GEM_FLINK, &flink);
if (r)
return r;
bo->flink_name = flink.name;
if (bo->dev->flink_fd != bo->dev->fd) {
struct drm_gem_close args = {};
args.handle = handle;
drmIoctl(bo->dev->flink_fd, DRM_IOCTL_GEM_CLOSE, &args);
}
pthread_mutex_lock(&bo->dev->bo_table_mutex);
util_hash_table_set(bo->dev->bo_flink_names,
(void*)(uintptr_t)bo->flink_name,
bo);
pthread_mutex_unlock(&bo->dev->bo_table_mutex);
return 0;
}
static boolean
vmw_drm_surface_get_handle(struct svga_winsys_screen *sws,
struct svga_winsys_surface *surface,
unsigned stride,
struct winsys_handle *whandle)
{
struct vmw_winsys_screen *vws = vmw_winsys_screen(sws);
struct vmw_svga_winsys_surface *vsrf;
int ret;
if (!surface)
return FALSE;
vsrf = vmw_svga_winsys_surface(surface);
whandle->handle = vsrf->sid;
whandle->stride = stride;
whandle->offset = 0;
switch (whandle->type) {
case DRM_API_HANDLE_TYPE_SHARED:
case DRM_API_HANDLE_TYPE_KMS:
whandle->handle = vsrf->sid;
break;
case DRM_API_HANDLE_TYPE_FD:
ret = drmPrimeHandleToFD(vws->ioctl.drm_fd, vsrf->sid, DRM_CLOEXEC,
(int *)&whandle->handle);
if (ret) {
vmw_error("Failed to get file descriptor from prime.\n");
return FALSE;
}
break;
default:
vmw_error("Attempt to export unsupported handle type %d.\n",
whandle->type);
return FALSE;
}
return TRUE;
}
static boolean virgl_drm_winsys_resource_get_handle(struct virgl_winsys *qws,
struct virgl_hw_res *res,
uint32_t stride,
struct winsys_handle *whandle)
{
struct virgl_drm_winsys *qdws = virgl_drm_winsys(qws);
struct drm_gem_flink flink;
if (!res)
return FALSE;
if (whandle->type == DRM_API_HANDLE_TYPE_SHARED) {
if (!res->flinked) {
memset(&flink, 0, sizeof(flink));
flink.handle = res->bo_handle;
if (drmIoctl(qdws->fd, DRM_IOCTL_GEM_FLINK, &flink)) {
return FALSE;
}
res->flinked = TRUE;
res->flink = flink.name;
mtx_lock(&qdws->bo_handles_mutex);
util_hash_table_set(qdws->bo_names, (void *)(uintptr_t)res->flink, res);
mtx_unlock(&qdws->bo_handles_mutex);
}
whandle->handle = res->flink;
} else if (whandle->type == DRM_API_HANDLE_TYPE_KMS) {
whandle->handle = res->bo_handle;
} else if (whandle->type == DRM_API_HANDLE_TYPE_FD) {
if (drmPrimeHandleToFD(qdws->fd, res->bo_handle, DRM_CLOEXEC, (int*)&whandle->handle))
return FALSE;
mtx_lock(&qdws->bo_handles_mutex);
util_hash_table_set(qdws->bo_handles, (void *)(uintptr_t)res->bo_handle, res);
mtx_unlock(&qdws->bo_handles_mutex);
}
whandle->stride = stride;
return TRUE;
}
drm_public
int drm_tegra_bo_to_dmabuf(struct drm_tegra_bo *bo, uint32_t *handle)
{
int prime_fd;
int err;
if (!bo || !handle)
return -EINVAL;
err = drmPrimeHandleToFD(bo->drm->fd, bo->handle, DRM_CLOEXEC,
&prime_fd);
if (err) {
VDBG_BO(bo, "faile err %d strerror(%s)\n",
err, strerror(-err));
return err;
}
*handle = prime_fd;
VDBG_BO(bo, "success prime_fd %u\n", prime_fd);
return 0;
}
static struct gbm_bo *gbm_kms_bo_create(struct gbm_device *gbm,
uint32_t width, uint32_t height,
uint32_t format, uint32_t usage)
{
struct gbm_kms_device *dev = (struct gbm_kms_device*)gbm;
struct gbm_kms_bo *bo;
unsigned attr[] = {
KMS_BO_TYPE, KMS_BO_TYPE_SCANOUT_X8R8G8B8,
KMS_WIDTH, 0,
KMS_HEIGHT, 0,
KMS_TERMINATE_PROP_LIST
};
GBM_DEBUG("%s: %s: %d\n", __FILE__, __func__, __LINE__);
if (!(bo = calloc(1, sizeof(struct gbm_kms_bo))))
return NULL;
switch (format) {
// 32bpp
case GBM_FORMAT_ARGB8888:
case GBM_BO_FORMAT_ARGB8888:
case GBM_FORMAT_XRGB8888:
case GBM_BO_FORMAT_XRGB8888:
break;
default:
// unsupported...
goto error;
}
if (usage & GBM_BO_USE_CURSOR_64X64)
attr[1] = KMS_BO_TYPE_CURSOR_64X64_A8R8G8B8;
attr[3] = width;
attr[5] = height;
// Create BO
if (kms_bo_create(dev->kms, attr, &bo->bo))
goto error;
bo->base.gbm = gbm;
bo->base.width = width;
bo->base.height = height;
bo->base.format = format;
kms_bo_get_prop(bo->bo, KMS_HANDLE, &bo->base.handle.u32);
kms_bo_get_prop(bo->bo, KMS_PITCH, &bo->base.stride);
bo->size = bo->base.stride * bo->base.height;
if (drmPrimeHandleToFD(dev->base.base.fd, bo->base.handle.u32, DRM_CLOEXEC, &bo->fd)) {
GBM_DEBUG("%s: %s: drmPrimeHandleToFD() failed. %s\n", __FILE__, __func__, strerror(errno));
goto error;
}
// Map to the user space for bo_write
if (usage & GBM_BO_USE_WRITE) {
if (kms_bo_map(bo->bo, &bo->addr))
goto error;
}
return (struct gbm_bo*)bo;
error:
GBM_DEBUG("%s: %s: %d: ERROR!!!!\n", __FILE__, __func__, __LINE__);
gbm_kms_bo_destroy((struct gbm_bo*)bo);
return NULL;
}
static int kmsgrab_read_packet(AVFormatContext *avctx, AVPacket *pkt)
{
KMSGrabContext *ctx = avctx->priv_data;
drmModePlane *plane;
drmModeFB *fb;
AVDRMFrameDescriptor *desc;
AVFrame *frame;
int64_t now;
int err, fd;
now = av_gettime();
if (ctx->frame_last) {
int64_t delay;
while (1) {
delay = ctx->frame_last + ctx->frame_delay - now;
if (delay <= 0)
break;
av_usleep(delay);
now = av_gettime();
}
}
ctx->frame_last = now;
plane = drmModeGetPlane(ctx->hwctx->fd, ctx->plane_id);
if (!plane) {
av_log(avctx, AV_LOG_ERROR, "Failed to get plane "
"%"PRIu32".\n", ctx->plane_id);
return AVERROR(EIO);
}
if (!plane->fb_id) {
av_log(avctx, AV_LOG_ERROR, "Plane %"PRIu32" no longer has "
"an associated framebuffer.\n", ctx->plane_id);
return AVERROR(EIO);
}
fb = drmModeGetFB(ctx->hwctx->fd, plane->fb_id);
if (!fb) {
av_log(avctx, AV_LOG_ERROR, "Failed to get framebuffer "
"%"PRIu32".\n", plane->fb_id);
return AVERROR(EIO);
}
if (fb->width != ctx->width || fb->height != ctx->height) {
av_log(avctx, AV_LOG_ERROR, "Plane %"PRIu32" framebuffer "
"dimensions changed: now %"PRIu32"x%"PRIu32".\n",
ctx->plane_id, fb->width, fb->height);
return AVERROR(EIO);
}
if (!fb->handle) {
av_log(avctx, AV_LOG_ERROR, "No handle set on framebuffer.\n");
return AVERROR(EIO);
}
err = drmPrimeHandleToFD(ctx->hwctx->fd, fb->handle, O_RDONLY, &fd);
if (err < 0) {
err = errno;
av_log(avctx, AV_LOG_ERROR, "Failed to get PRIME fd from "
"framebuffer handle: %s.\n", strerror(errno));
return AVERROR(err);
}
desc = av_mallocz(sizeof(*desc));
if (!desc)
return AVERROR(ENOMEM);
*desc = (AVDRMFrameDescriptor) {
.nb_objects = 1,
.objects[0] = {
.fd = fd,
.size = fb->height * fb->pitch,
.format_modifier = ctx->drm_format_modifier,
},
.nb_layers = 1,
.layers[0] = {
.format = ctx->drm_format,
.nb_planes = 1,
.planes[0] = {
.object_index = 0,
.offset = 0,
.pitch = fb->pitch,
},
},
};
int amdgpu_bo_import(amdgpu_device_handle dev,
enum amdgpu_bo_handle_type type,
uint32_t shared_handle,
struct amdgpu_bo_import_result *output)
{
struct drm_gem_open open_arg = {};
struct amdgpu_bo *bo = NULL;
int r;
int dma_fd;
uint64_t dma_buf_size = 0;
/* Convert a DMA buf handle to a KMS handle now. */
if (type == amdgpu_bo_handle_type_dma_buf_fd) {
uint32_t handle;
off_t size;
/* Get a KMS handle. */
r = drmPrimeFDToHandle(dev->fd, shared_handle, &handle);
if (r) {
return r;
}
/* Query the buffer size. */
size = lseek(shared_handle, 0, SEEK_END);
if (size == (off_t)-1) {
amdgpu_close_kms_handle(dev, handle);
return -errno;
}
lseek(shared_handle, 0, SEEK_SET);
dma_buf_size = size;
shared_handle = handle;
}
/* We must maintain a list of pairs <handle, bo>, so that we always
* return the same amdgpu_bo instance for the same handle. */
pthread_mutex_lock(&dev->bo_table_mutex);
/* If we have already created a buffer with this handle, find it. */
switch (type) {
case amdgpu_bo_handle_type_gem_flink_name:
bo = util_hash_table_get(dev->bo_flink_names,
(void*)(uintptr_t)shared_handle);
break;
case amdgpu_bo_handle_type_dma_buf_fd:
bo = util_hash_table_get(dev->bo_handles,
(void*)(uintptr_t)shared_handle);
break;
case amdgpu_bo_handle_type_kms:
/* Importing a KMS handle in not allowed. */
pthread_mutex_unlock(&dev->bo_table_mutex);
return -EPERM;
default:
pthread_mutex_unlock(&dev->bo_table_mutex);
return -EINVAL;
}
if (bo) {
pthread_mutex_unlock(&dev->bo_table_mutex);
/* The buffer already exists, just bump the refcount. */
atomic_inc(&bo->refcount);
output->buf_handle = bo;
output->alloc_size = bo->alloc_size;
return 0;
}
bo = calloc(1, sizeof(struct amdgpu_bo));
if (!bo) {
pthread_mutex_unlock(&dev->bo_table_mutex);
if (type == amdgpu_bo_handle_type_dma_buf_fd) {
amdgpu_close_kms_handle(dev, shared_handle);
}
return -ENOMEM;
}
/* Open the handle. */
switch (type) {
case amdgpu_bo_handle_type_gem_flink_name:
open_arg.name = shared_handle;
r = drmIoctl(dev->flink_fd, DRM_IOCTL_GEM_OPEN, &open_arg);
if (r) {
free(bo);
pthread_mutex_unlock(&dev->bo_table_mutex);
return r;
}
bo->handle = open_arg.handle;
if (dev->flink_fd != dev->fd) {
r = drmPrimeHandleToFD(dev->flink_fd, bo->handle, DRM_CLOEXEC, &dma_fd);
if (r) {
free(bo);
pthread_mutex_unlock(&dev->bo_table_mutex);
return r;
}
r = drmPrimeFDToHandle(dev->fd, dma_fd, &bo->handle );
close(dma_fd);
if (r) {
free(bo);
pthread_mutex_unlock(&dev->bo_table_mutex);
return r;
}
}
bo->flink_name = shared_handle;
bo->alloc_size = open_arg.size;
util_hash_table_set(dev->bo_flink_names,
(void*)(uintptr_t)bo->flink_name, bo);
break;
case amdgpu_bo_handle_type_dma_buf_fd:
bo->handle = shared_handle;
bo->alloc_size = dma_buf_size;
break;
case amdgpu_bo_handle_type_kms:
assert(0); /* unreachable */
}
/* Initialize it. */
atomic_set(&bo->refcount, 1);
bo->dev = dev;
pthread_mutex_init(&bo->cpu_access_mutex, NULL);
util_hash_table_set(dev->bo_handles, (void*)(uintptr_t)bo->handle, bo);
pthread_mutex_unlock(&dev->bo_table_mutex);
output->buf_handle = bo;
output->alloc_size = bo->alloc_size;
return 0;
}
int
iris_bo_busy(struct iris_bo *bo)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
struct drm_i915_gem_busy busy = { .handle = bo->gem_handle };
int ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_BUSY, &busy);
if (ret == 0) {
bo->idle = !busy.busy;
return busy.busy;
}
return false;
}
int
iris_bo_madvise(struct iris_bo *bo, int state)
{
struct drm_i915_gem_madvise madv = {
.handle = bo->gem_handle,
.madv = state,
.retained = 1,
};
drm_ioctl(bo->bufmgr->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv);
return madv.retained;
}
/* drop the oldest entries that have been purged by the kernel */
static void
iris_bo_cache_purge_bucket(struct iris_bufmgr *bufmgr,
struct bo_cache_bucket *bucket)
{
list_for_each_entry_safe(struct iris_bo, bo, &bucket->head, head) {
if (iris_bo_madvise(bo, I915_MADV_DONTNEED))
break;
list_del(&bo->head);
bo_free(bo);
}
}
static struct iris_bo *
bo_calloc(void)
{
struct iris_bo *bo = calloc(1, sizeof(*bo));
if (bo) {
bo->hash = _mesa_hash_pointer(bo);
}
return bo;
}
static struct iris_bo *
bo_alloc_internal(struct iris_bufmgr *bufmgr,
const char *name,
uint64_t size,
enum iris_memory_zone memzone,
unsigned flags,
uint32_t tiling_mode,
uint32_t stride)
{
struct iris_bo *bo;
unsigned int page_size = getpagesize();
int ret;
struct bo_cache_bucket *bucket;
bool alloc_from_cache;
uint64_t bo_size;
bool zeroed = false;
if (flags & BO_ALLOC_ZEROED)
zeroed = true;
if ((flags & BO_ALLOC_COHERENT) && !bufmgr->has_llc) {
bo_size = MAX2(ALIGN(size, page_size), page_size);
bucket = NULL;
goto skip_cache;
}
/* Round the allocated size up to a power of two number of pages. */
bucket = bucket_for_size(bufmgr, size);
/* If we don't have caching at this size, don't actually round the
* allocation up.
*/
if (bucket == NULL) {
bo_size = MAX2(ALIGN(size, page_size), page_size);
} else {
bo_size = bucket->size;
}
mtx_lock(&bufmgr->lock);
/* Get a buffer out of the cache if available */
retry:
alloc_from_cache = false;
if (bucket != NULL && !list_empty(&bucket->head)) {
/* If the last BO in the cache is idle, then reuse it. Otherwise,
* allocate a fresh buffer to avoid stalling.
*/
bo = LIST_ENTRY(struct iris_bo, bucket->head.next, head);
if (!iris_bo_busy(bo)) {
alloc_from_cache = true;
list_del(&bo->head);
}
if (alloc_from_cache) {
if (!iris_bo_madvise(bo, I915_MADV_WILLNEED)) {
bo_free(bo);
iris_bo_cache_purge_bucket(bufmgr, bucket);
goto retry;
}
if (bo_set_tiling_internal(bo, tiling_mode, stride)) {
bo_free(bo);
goto retry;
}
if (zeroed) {
void *map = iris_bo_map(NULL, bo, MAP_WRITE | MAP_RAW);
if (!map) {
bo_free(bo);
goto retry;
}
memset(map, 0, bo_size);
}
}
}
if (alloc_from_cache) {
/* If the cached BO isn't in the right memory zone, free the old
* memory and assign it a new address.
*/
if (memzone != iris_memzone_for_address(bo->gtt_offset)) {
vma_free(bufmgr, bo->gtt_offset, bo->size);
bo->gtt_offset = 0ull;
}
} else {
skip_cache:
bo = bo_calloc();
if (!bo)
goto err;
bo->size = bo_size;
bo->idle = true;
struct drm_i915_gem_create create = { .size = bo_size };
/* All new BOs we get from the kernel are zeroed, so we don't need to
* worry about that here.
*/
ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_CREATE, &create);
if (ret != 0) {
free(bo);
goto err;
}
bo->gem_handle = create.handle;
bo->bufmgr = bufmgr;
bo->tiling_mode = I915_TILING_NONE;
bo->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
bo->stride = 0;
if (bo_set_tiling_internal(bo, tiling_mode, stride))
goto err_free;
/* Calling set_domain() will allocate pages for the BO outside of the
* struct mutex lock in the kernel, which is more efficient than waiting
* to create them during the first execbuf that uses the BO.
*/
struct drm_i915_gem_set_domain sd = {
.handle = bo->gem_handle,
.read_domains = I915_GEM_DOMAIN_CPU,
.write_domain = 0,
};
if (drm_ioctl(bo->bufmgr->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &sd) != 0)
goto err_free;
}
bo->name = name;
p_atomic_set(&bo->refcount, 1);
bo->reusable = bucket && bufmgr->bo_reuse;
bo->cache_coherent = bufmgr->has_llc;
bo->index = -1;
bo->kflags = EXEC_OBJECT_SUPPORTS_48B_ADDRESS | EXEC_OBJECT_PINNED;
/* By default, capture all driver-internal buffers like shader kernels,
* surface states, dynamic states, border colors, and so on.
*/
if (memzone < IRIS_MEMZONE_OTHER)
bo->kflags |= EXEC_OBJECT_CAPTURE;
if (bo->gtt_offset == 0ull) {
bo->gtt_offset = vma_alloc(bufmgr, memzone, bo->size, 1);
if (bo->gtt_offset == 0ull)
goto err_free;
}
mtx_unlock(&bufmgr->lock);
if ((flags & BO_ALLOC_COHERENT) && !bo->cache_coherent) {
struct drm_i915_gem_caching arg = {
.handle = bo->gem_handle,
.caching = 1,
};
if (drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_SET_CACHING, &arg) == 0) {
bo->cache_coherent = true;
bo->reusable = false;
}
}
DBG("bo_create: buf %d (%s) (%s memzone) %llub\n", bo->gem_handle,
bo->name, memzone_name(memzone), (unsigned long long) size);
return bo;
err_free:
bo_free(bo);
err:
mtx_unlock(&bufmgr->lock);
return NULL;
}
struct iris_bo *
iris_bo_alloc(struct iris_bufmgr *bufmgr,
const char *name,
uint64_t size,
enum iris_memory_zone memzone)
{
return bo_alloc_internal(bufmgr, name, size, memzone,
0, I915_TILING_NONE, 0);
}
struct iris_bo *
iris_bo_alloc_tiled(struct iris_bufmgr *bufmgr, const char *name,
uint64_t size, enum iris_memory_zone memzone,
uint32_t tiling_mode, uint32_t pitch, unsigned flags)
{
return bo_alloc_internal(bufmgr, name, size, memzone,
flags, tiling_mode, pitch);
}
struct iris_bo *
iris_bo_create_userptr(struct iris_bufmgr *bufmgr, const char *name,
void *ptr, size_t size,
enum iris_memory_zone memzone)
{
struct iris_bo *bo;
bo = bo_calloc();
if (!bo)
return NULL;
struct drm_i915_gem_userptr arg = {
.user_ptr = (uintptr_t)ptr,
.user_size = size,
};
if (drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_USERPTR, &arg))
goto err_free;
bo->gem_handle = arg.handle;
/* Check the buffer for validity before we try and use it in a batch */
struct drm_i915_gem_set_domain sd = {
.handle = bo->gem_handle,
.read_domains = I915_GEM_DOMAIN_CPU,
};
if (drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &sd))
goto err_close;
bo->name = name;
bo->size = size;
bo->map_cpu = ptr;
bo->bufmgr = bufmgr;
bo->kflags = EXEC_OBJECT_SUPPORTS_48B_ADDRESS | EXEC_OBJECT_PINNED;
bo->gtt_offset = vma_alloc(bufmgr, memzone, size, 1);
if (bo->gtt_offset == 0ull)
goto err_close;
p_atomic_set(&bo->refcount, 1);
bo->userptr = true;
bo->cache_coherent = true;
bo->index = -1;
bo->idle = true;
return bo;
err_close:
drm_ioctl(bufmgr->fd, DRM_IOCTL_GEM_CLOSE, &bo->gem_handle);
err_free:
free(bo);
return NULL;
}
/**
* Returns a iris_bo wrapping the given buffer object handle.
*
* This can be used when one application needs to pass a buffer object
* to another.
*/
struct iris_bo *
iris_bo_gem_create_from_name(struct iris_bufmgr *bufmgr,
const char *name, unsigned int handle)
{
struct iris_bo *bo;
/* At the moment most applications only have a few named bo.
* For instance, in a DRI client only the render buffers passed
* between X and the client are named. And since X returns the
* alternating names for the front/back buffer a linear search
* provides a sufficiently fast match.
*/
mtx_lock(&bufmgr->lock);
bo = hash_find_bo(bufmgr->name_table, handle);
if (bo) {
iris_bo_reference(bo);
goto out;
}
struct drm_gem_open open_arg = { .name = handle };
int ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_GEM_OPEN, &open_arg);
if (ret != 0) {
DBG("Couldn't reference %s handle 0x%08x: %s\n",
name, handle, strerror(errno));
bo = NULL;
goto out;
}
/* Now see if someone has used a prime handle to get this
* object from the kernel before by looking through the list
* again for a matching gem_handle
*/
bo = hash_find_bo(bufmgr->handle_table, open_arg.handle);
if (bo) {
iris_bo_reference(bo);
goto out;
}
bo = bo_calloc();
if (!bo)
goto out;
p_atomic_set(&bo->refcount, 1);
bo->size = open_arg.size;
bo->gtt_offset = 0;
bo->bufmgr = bufmgr;
bo->gem_handle = open_arg.handle;
bo->name = name;
bo->global_name = handle;
bo->reusable = false;
bo->external = true;
bo->kflags = EXEC_OBJECT_SUPPORTS_48B_ADDRESS | EXEC_OBJECT_PINNED;
bo->gtt_offset = vma_alloc(bufmgr, IRIS_MEMZONE_OTHER, bo->size, 1);
_mesa_hash_table_insert(bufmgr->handle_table, &bo->gem_handle, bo);
_mesa_hash_table_insert(bufmgr->name_table, &bo->global_name, bo);
struct drm_i915_gem_get_tiling get_tiling = { .handle = bo->gem_handle };
ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_GET_TILING, &get_tiling);
if (ret != 0)
goto err_unref;
bo->tiling_mode = get_tiling.tiling_mode;
bo->swizzle_mode = get_tiling.swizzle_mode;
/* XXX stride is unknown */
DBG("bo_create_from_handle: %d (%s)\n", handle, bo->name);
out:
mtx_unlock(&bufmgr->lock);
return bo;
err_unref:
bo_free(bo);
mtx_unlock(&bufmgr->lock);
return NULL;
}
static void
bo_free(struct iris_bo *bo)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
if (bo->map_cpu && !bo->userptr) {
VG_NOACCESS(bo->map_cpu, bo->size);
munmap(bo->map_cpu, bo->size);
}
if (bo->map_wc) {
VG_NOACCESS(bo->map_wc, bo->size);
munmap(bo->map_wc, bo->size);
}
if (bo->map_gtt) {
VG_NOACCESS(bo->map_gtt, bo->size);
munmap(bo->map_gtt, bo->size);
}
if (bo->external) {
struct hash_entry *entry;
if (bo->global_name) {
entry = _mesa_hash_table_search(bufmgr->name_table, &bo->global_name);
_mesa_hash_table_remove(bufmgr->name_table, entry);
}
entry = _mesa_hash_table_search(bufmgr->handle_table, &bo->gem_handle);
_mesa_hash_table_remove(bufmgr->handle_table, entry);
}
/* Close this object */
struct drm_gem_close close = { .handle = bo->gem_handle };
int ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_GEM_CLOSE, &close);
if (ret != 0) {
DBG("DRM_IOCTL_GEM_CLOSE %d failed (%s): %s\n",
bo->gem_handle, bo->name, strerror(errno));
}
vma_free(bo->bufmgr, bo->gtt_offset, bo->size);
free(bo);
}
/** Frees all cached buffers significantly older than @time. */
static void
cleanup_bo_cache(struct iris_bufmgr *bufmgr, time_t time)
{
int i;
if (bufmgr->time == time)
return;
for (i = 0; i < bufmgr->num_buckets; i++) {
struct bo_cache_bucket *bucket = &bufmgr->cache_bucket[i];
list_for_each_entry_safe(struct iris_bo, bo, &bucket->head, head) {
if (time - bo->free_time <= 1)
break;
list_del(&bo->head);
bo_free(bo);
}
}
bufmgr->time = time;
}
static void
bo_unreference_final(struct iris_bo *bo, time_t time)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
struct bo_cache_bucket *bucket;
DBG("bo_unreference final: %d (%s)\n", bo->gem_handle, bo->name);
bucket = NULL;
if (bo->reusable)
bucket = bucket_for_size(bufmgr, bo->size);
/* Put the buffer into our internal cache for reuse if we can. */
if (bucket && iris_bo_madvise(bo, I915_MADV_DONTNEED)) {
bo->free_time = time;
bo->name = NULL;
list_addtail(&bo->head, &bucket->head);
} else {
bo_free(bo);
}
}
void
iris_bo_unreference(struct iris_bo *bo)
{
if (bo == NULL)
return;
assert(p_atomic_read(&bo->refcount) > 0);
if (atomic_add_unless(&bo->refcount, -1, 1)) {
struct iris_bufmgr *bufmgr = bo->bufmgr;
struct timespec time;
clock_gettime(CLOCK_MONOTONIC, &time);
mtx_lock(&bufmgr->lock);
if (p_atomic_dec_zero(&bo->refcount)) {
bo_unreference_final(bo, time.tv_sec);
cleanup_bo_cache(bufmgr, time.tv_sec);
}
mtx_unlock(&bufmgr->lock);
}
}
static void
bo_wait_with_stall_warning(struct pipe_debug_callback *dbg,
struct iris_bo *bo,
const char *action)
{
bool busy = dbg && !bo->idle;
double elapsed = unlikely(busy) ? -get_time() : 0.0;
iris_bo_wait_rendering(bo);
if (unlikely(busy)) {
elapsed += get_time();
if (elapsed > 1e-5) /* 0.01ms */ {
perf_debug(dbg, "%s a busy \"%s\" BO stalled and took %.03f ms.\n",
action, bo->name, elapsed * 1000);
}
}
}
static void
print_flags(unsigned flags)
{
if (flags & MAP_READ)
DBG("READ ");
if (flags & MAP_WRITE)
DBG("WRITE ");
if (flags & MAP_ASYNC)
DBG("ASYNC ");
if (flags & MAP_PERSISTENT)
DBG("PERSISTENT ");
if (flags & MAP_COHERENT)
DBG("COHERENT ");
if (flags & MAP_RAW)
DBG("RAW ");
DBG("\n");
}
static void *
iris_bo_map_cpu(struct pipe_debug_callback *dbg,
struct iris_bo *bo, unsigned flags)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
/* We disallow CPU maps for writing to non-coherent buffers, as the
* CPU map can become invalidated when a batch is flushed out, which
* can happen at unpredictable times. You should use WC maps instead.
*/
assert(bo->cache_coherent || !(flags & MAP_WRITE));
if (!bo->map_cpu) {
DBG("iris_bo_map_cpu: %d (%s)\n", bo->gem_handle, bo->name);
struct drm_i915_gem_mmap mmap_arg = {
.handle = bo->gem_handle,
.size = bo->size,
};
int ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_MMAP, &mmap_arg);
if (ret != 0) {
DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
__FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
return NULL;
}
void *map = (void *) (uintptr_t) mmap_arg.addr_ptr;
VG_DEFINED(map, bo->size);
if (p_atomic_cmpxchg(&bo->map_cpu, NULL, map)) {
VG_NOACCESS(map, bo->size);
munmap(map, bo->size);
}
}
assert(bo->map_cpu);
DBG("iris_bo_map_cpu: %d (%s) -> %p, ", bo->gem_handle, bo->name,
bo->map_cpu);
print_flags(flags);
if (!(flags & MAP_ASYNC)) {
bo_wait_with_stall_warning(dbg, bo, "CPU mapping");
}
if (!bo->cache_coherent && !bo->bufmgr->has_llc) {
/* If we're reusing an existing CPU mapping, the CPU caches may
* contain stale data from the last time we read from that mapping.
* (With the BO cache, it might even be data from a previous buffer!)
* Even if it's a brand new mapping, the kernel may have zeroed the
* buffer via CPU writes.
*
* We need to invalidate those cachelines so that we see the latest
* contents, and so long as we only read from the CPU mmap we do not
* need to write those cachelines back afterwards.
*
* On LLC, the emprical evidence suggests that writes from the GPU
* that bypass the LLC (i.e. for scanout) do *invalidate* the CPU
* cachelines. (Other reads, such as the display engine, bypass the
* LLC entirely requiring us to keep dirty pixels for the scanout
* out of any cache.)
*/
gen_invalidate_range(bo->map_cpu, bo->size);
}
return bo->map_cpu;
}
static void *
iris_bo_map_wc(struct pipe_debug_callback *dbg,
struct iris_bo *bo, unsigned flags)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
if (!bo->map_wc) {
DBG("iris_bo_map_wc: %d (%s)\n", bo->gem_handle, bo->name);
struct drm_i915_gem_mmap mmap_arg = {
.handle = bo->gem_handle,
.size = bo->size,
.flags = I915_MMAP_WC,
};
int ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_MMAP, &mmap_arg);
if (ret != 0) {
DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
__FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
return NULL;
}
void *map = (void *) (uintptr_t) mmap_arg.addr_ptr;
VG_DEFINED(map, bo->size);
if (p_atomic_cmpxchg(&bo->map_wc, NULL, map)) {
VG_NOACCESS(map, bo->size);
munmap(map, bo->size);
}
}
assert(bo->map_wc);
DBG("iris_bo_map_wc: %d (%s) -> %p\n", bo->gem_handle, bo->name, bo->map_wc);
print_flags(flags);
if (!(flags & MAP_ASYNC)) {
bo_wait_with_stall_warning(dbg, bo, "WC mapping");
}
return bo->map_wc;
}
/**
* Perform an uncached mapping via the GTT.
*
* Write access through the GTT is not quite fully coherent. On low power
* systems especially, like modern Atoms, we can observe reads from RAM before
* the write via GTT has landed. A write memory barrier that flushes the Write
* Combining Buffer (i.e. sfence/mfence) is not sufficient to order the later
* read after the write as the GTT write suffers a small delay through the GTT
* indirection. The kernel uses an uncached mmio read to ensure the GTT write
* is ordered with reads (either by the GPU, WB or WC) and unconditionally
* flushes prior to execbuf submission. However, if we are not informing the
* kernel about our GTT writes, it will not flush before earlier access, such
* as when using the cmdparser. Similarly, we need to be careful if we should
* ever issue a CPU read immediately following a GTT write.
*
* Telling the kernel about write access also has one more important
* side-effect. Upon receiving notification about the write, it cancels any
* scanout buffering for FBC/PSR and friends. Later FBC/PSR is then flushed by
* either SW_FINISH or DIRTYFB. The presumption is that we never write to the
* actual scanout via a mmaping, only to a backbuffer and so all the FBC/PSR
* tracking is handled on the buffer exchange instead.
*/
static void *
iris_bo_map_gtt(struct pipe_debug_callback *dbg,
struct iris_bo *bo, unsigned flags)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
/* Get a mapping of the buffer if we haven't before. */
if (bo->map_gtt == NULL) {
DBG("bo_map_gtt: mmap %d (%s)\n", bo->gem_handle, bo->name);
struct drm_i915_gem_mmap_gtt mmap_arg = { .handle = bo->gem_handle };
/* Get the fake offset back... */
int ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_MMAP_GTT, &mmap_arg);
if (ret != 0) {
DBG("%s:%d: Error preparing buffer map %d (%s): %s .\n",
__FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
return NULL;
}
/* and mmap it. */
void *map = mmap(0, bo->size, PROT_READ | PROT_WRITE,
MAP_SHARED, bufmgr->fd, mmap_arg.offset);
if (map == MAP_FAILED) {
DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
__FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
return NULL;
}
/* We don't need to use VALGRIND_MALLOCLIKE_BLOCK because Valgrind will
* already intercept this mmap call. However, for consistency between
* all the mmap paths, we mark the pointer as defined now and mark it
* as inaccessible afterwards.
*/
VG_DEFINED(map, bo->size);
if (p_atomic_cmpxchg(&bo->map_gtt, NULL, map)) {
VG_NOACCESS(map, bo->size);
munmap(map, bo->size);
}
}
assert(bo->map_gtt);
DBG("bo_map_gtt: %d (%s) -> %p, ", bo->gem_handle, bo->name, bo->map_gtt);
print_flags(flags);
if (!(flags & MAP_ASYNC)) {
bo_wait_with_stall_warning(dbg, bo, "GTT mapping");
}
return bo->map_gtt;
}
static bool
can_map_cpu(struct iris_bo *bo, unsigned flags)
{
if (bo->cache_coherent)
return true;
/* Even if the buffer itself is not cache-coherent (such as a scanout), on
* an LLC platform reads always are coherent (as they are performed via the
* central system agent). It is just the writes that we need to take special
* care to ensure that land in main memory and not stick in the CPU cache.
*/
if (!(flags & MAP_WRITE) && bo->bufmgr->has_llc)
return true;
/* If PERSISTENT or COHERENT are set, the mmapping needs to remain valid
* across batch flushes where the kernel will change cache domains of the
* bo, invalidating continued access to the CPU mmap on non-LLC device.
*
* Similarly, ASYNC typically means that the buffer will be accessed via
* both the CPU and the GPU simultaneously. Batches may be executed that
* use the BO even while it is mapped. While OpenGL technically disallows
* most drawing while non-persistent mappings are active, we may still use
* the GPU for blits or other operations, causing batches to happen at
* inconvenient times.
*
* If RAW is set, we expect the caller to be able to handle a WC buffer
* more efficiently than the involuntary clflushes.
*/
if (flags & (MAP_PERSISTENT | MAP_COHERENT | MAP_ASYNC | MAP_RAW))
return false;
return !(flags & MAP_WRITE);
}
void *
iris_bo_map(struct pipe_debug_callback *dbg,
struct iris_bo *bo, unsigned flags)
{
if (bo->tiling_mode != I915_TILING_NONE && !(flags & MAP_RAW))
return iris_bo_map_gtt(dbg, bo, flags);
void *map;
if (can_map_cpu(bo, flags))
map = iris_bo_map_cpu(dbg, bo, flags);
else
map = iris_bo_map_wc(dbg, bo, flags);
/* Allow the attempt to fail by falling back to the GTT where necessary.
*
* Not every buffer can be mmaped directly using the CPU (or WC), for
* example buffers that wrap stolen memory or are imported from other
* devices. For those, we have little choice but to use a GTT mmapping.
* However, if we use a slow GTT mmapping for reads where we expected fast
* access, that order of magnitude difference in throughput will be clearly
* expressed by angry users.
*
* We skip MAP_RAW because we want to avoid map_gtt's fence detiling.
*/
if (!map && !(flags & MAP_RAW)) {
perf_debug(dbg, "Fallback GTT mapping for %s with access flags %x\n",
bo->name, flags);
map = iris_bo_map_gtt(dbg, bo, flags);
}
return map;
}
/** Waits for all GPU rendering with the object to have completed. */
void
iris_bo_wait_rendering(struct iris_bo *bo)
{
/* We require a kernel recent enough for WAIT_IOCTL support.
* See intel_init_bufmgr()
*/
iris_bo_wait(bo, -1);
}
/**
* Waits on a BO for the given amount of time.
*
* @bo: buffer object to wait for
* @timeout_ns: amount of time to wait in nanoseconds.
* If value is less than 0, an infinite wait will occur.
*
* Returns 0 if the wait was successful ie. the last batch referencing the
* object has completed within the allotted time. Otherwise some negative return
* value describes the error. Of particular interest is -ETIME when the wait has
* failed to yield the desired result.
*
* Similar to iris_bo_wait_rendering except a timeout parameter allows
* the operation to give up after a certain amount of time. Another subtle
* difference is the internal locking semantics are different (this variant does
* not hold the lock for the duration of the wait). This makes the wait subject
* to a larger userspace race window.
*
* The implementation shall wait until the object is no longer actively
* referenced within a batch buffer at the time of the call. The wait will
* not guarantee that the buffer is re-issued via another thread, or an flinked
* handle. Userspace must make sure this race does not occur if such precision
* is important.
*
* Note that some kernels have broken the inifite wait for negative values
* promise, upgrade to latest stable kernels if this is the case.
*/
int
iris_bo_wait(struct iris_bo *bo, int64_t timeout_ns)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
/* If we know it's idle, don't bother with the kernel round trip */
if (bo->idle && !bo->external)
return 0;
struct drm_i915_gem_wait wait = {
.bo_handle = bo->gem_handle,
.timeout_ns = timeout_ns,
};
int ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_WAIT, &wait);
if (ret != 0)
return -errno;
bo->idle = true;
return ret;
}
void
iris_bufmgr_destroy(struct iris_bufmgr *bufmgr)
{
mtx_destroy(&bufmgr->lock);
/* Free any cached buffer objects we were going to reuse */
for (int i = 0; i < bufmgr->num_buckets; i++) {
struct bo_cache_bucket *bucket = &bufmgr->cache_bucket[i];
list_for_each_entry_safe(struct iris_bo, bo, &bucket->head, head) {
list_del(&bo->head);
bo_free(bo);
}
}
_mesa_hash_table_destroy(bufmgr->name_table, NULL);
_mesa_hash_table_destroy(bufmgr->handle_table, NULL);
for (int z = 0; z < IRIS_MEMZONE_COUNT; z++) {
if (z != IRIS_MEMZONE_BINDER)
util_vma_heap_finish(&bufmgr->vma_allocator[z]);
}
free(bufmgr);
}
static int
bo_set_tiling_internal(struct iris_bo *bo, uint32_t tiling_mode,
uint32_t stride)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
struct drm_i915_gem_set_tiling set_tiling;
int ret;
if (bo->global_name == 0 &&
tiling_mode == bo->tiling_mode && stride == bo->stride)
return 0;
memset(&set_tiling, 0, sizeof(set_tiling));
do {
/* set_tiling is slightly broken and overwrites the
* input on the error path, so we have to open code
* drm_ioctl.
*/
set_tiling.handle = bo->gem_handle;
set_tiling.tiling_mode = tiling_mode;
set_tiling.stride = stride;
ret = ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_SET_TILING, &set_tiling);
} while (ret == -1 && (errno == EINTR || errno == EAGAIN));
if (ret == -1)
return -errno;
bo->tiling_mode = set_tiling.tiling_mode;
bo->swizzle_mode = set_tiling.swizzle_mode;
bo->stride = set_tiling.stride;
return 0;
}
int
iris_bo_get_tiling(struct iris_bo *bo, uint32_t *tiling_mode,
uint32_t *swizzle_mode)
{
*tiling_mode = bo->tiling_mode;
*swizzle_mode = bo->swizzle_mode;
return 0;
}
struct iris_bo *
iris_bo_import_dmabuf(struct iris_bufmgr *bufmgr, int prime_fd)
{
uint32_t handle;
struct iris_bo *bo;
mtx_lock(&bufmgr->lock);
int ret = drmPrimeFDToHandle(bufmgr->fd, prime_fd, &handle);
if (ret) {
DBG("import_dmabuf: failed to obtain handle from fd: %s\n",
strerror(errno));
mtx_unlock(&bufmgr->lock);
return NULL;
}
/*
* See if the kernel has already returned this buffer to us. Just as
* for named buffers, we must not create two bo's pointing at the same
* kernel object
*/
bo = hash_find_bo(bufmgr->handle_table, handle);
if (bo) {
iris_bo_reference(bo);
goto out;
}
bo = bo_calloc();
if (!bo)
goto out;
p_atomic_set(&bo->refcount, 1);
/* Determine size of bo. The fd-to-handle ioctl really should
* return the size, but it doesn't. If we have kernel 3.12 or
* later, we can lseek on the prime fd to get the size. Older
* kernels will just fail, in which case we fall back to the
* provided (estimated or guess size). */
ret = lseek(prime_fd, 0, SEEK_END);
if (ret != -1)
bo->size = ret;
bo->bufmgr = bufmgr;
bo->gem_handle = handle;
_mesa_hash_table_insert(bufmgr->handle_table, &bo->gem_handle, bo);
bo->name = "prime";
bo->reusable = false;
bo->external = true;
bo->kflags = EXEC_OBJECT_SUPPORTS_48B_ADDRESS | EXEC_OBJECT_PINNED;
bo->gtt_offset = vma_alloc(bufmgr, IRIS_MEMZONE_OTHER, bo->size, 1);
struct drm_i915_gem_get_tiling get_tiling = { .handle = bo->gem_handle };
if (drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_GET_TILING, &get_tiling))
goto err;
bo->tiling_mode = get_tiling.tiling_mode;
bo->swizzle_mode = get_tiling.swizzle_mode;
/* XXX stride is unknown */
out:
mtx_unlock(&bufmgr->lock);
return bo;
err:
bo_free(bo);
mtx_unlock(&bufmgr->lock);
return NULL;
}
static void
iris_bo_make_external_locked(struct iris_bo *bo)
{
if (!bo->external) {
_mesa_hash_table_insert(bo->bufmgr->handle_table, &bo->gem_handle, bo);
bo->external = true;
}
}
static void
iris_bo_make_external(struct iris_bo *bo)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
if (bo->external)
return;
mtx_lock(&bufmgr->lock);
iris_bo_make_external_locked(bo);
mtx_unlock(&bufmgr->lock);
}
int
iris_bo_export_dmabuf(struct iris_bo *bo, int *prime_fd)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
iris_bo_make_external(bo);
if (drmPrimeHandleToFD(bufmgr->fd, bo->gem_handle,
DRM_CLOEXEC, prime_fd) != 0)
return -errno;
bo->reusable = false;
return 0;
}
uint32_t
iris_bo_export_gem_handle(struct iris_bo *bo)
{
iris_bo_make_external(bo);
return bo->gem_handle;
}
int
iris_bo_flink(struct iris_bo *bo, uint32_t *name)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
if (!bo->global_name) {
struct drm_gem_flink flink = { .handle = bo->gem_handle };
if (drm_ioctl(bufmgr->fd, DRM_IOCTL_GEM_FLINK, &flink))
return -errno;
mtx_lock(&bufmgr->lock);
if (!bo->global_name) {
iris_bo_make_external_locked(bo);
bo->global_name = flink.name;
_mesa_hash_table_insert(bufmgr->name_table, &bo->global_name, bo);
}
mtx_unlock(&bufmgr->lock);
bo->reusable = false;
}
*name = bo->global_name;
return 0;
}
static void
add_bucket(struct iris_bufmgr *bufmgr, int size)
{
unsigned int i = bufmgr->num_buckets;
assert(i < ARRAY_SIZE(bufmgr->cache_bucket));
list_inithead(&bufmgr->cache_bucket[i].head);
bufmgr->cache_bucket[i].size = size;
bufmgr->num_buckets++;
assert(bucket_for_size(bufmgr, size) == &bufmgr->cache_bucket[i]);
assert(bucket_for_size(bufmgr, size - 2048) == &bufmgr->cache_bucket[i]);
assert(bucket_for_size(bufmgr, size + 1) != &bufmgr->cache_bucket[i]);
}
static void
init_cache_buckets(struct iris_bufmgr *bufmgr)
{
uint64_t size, cache_max_size = 64 * 1024 * 1024;
/* OK, so power of two buckets was too wasteful of memory.
* Give 3 other sizes between each power of two, to hopefully
* cover things accurately enough. (The alternative is
* probably to just go for exact matching of sizes, and assume
* that for things like composited window resize the tiled
* width/height alignment and rounding of sizes to pages will
* get us useful cache hit rates anyway)
*/
add_bucket(bufmgr, PAGE_SIZE);
add_bucket(bufmgr, PAGE_SIZE * 2);
add_bucket(bufmgr, PAGE_SIZE * 3);
/* Initialize the linked lists for BO reuse cache. */
for (size = 4 * PAGE_SIZE; size <= cache_max_size; size *= 2) {
add_bucket(bufmgr, size);
add_bucket(bufmgr, size + size * 1 / 4);
add_bucket(bufmgr, size + size * 2 / 4);
add_bucket(bufmgr, size + size * 3 / 4);
}
}
/*
* Destroy a exynos buffer object.
*
* @bo: a exynos buffer object to be destroyed.
*/
drm_public void exynos_bo_destroy(struct exynos_bo *bo)
{
if (!bo)
return;
if (bo->vaddr)
munmap(bo->vaddr, bo->size);
if (bo->handle) {
struct drm_gem_close req = {
.handle = bo->handle,
};
drmIoctl(bo->dev->fd, DRM_IOCTL_GEM_CLOSE, &req);
}
free(bo);
}
/*
* Get a exynos buffer object from a gem global object name.
*
* @dev: a exynos device object.
* @name: a gem global object name exported by another process.
*
* this interface is used to get a exynos buffer object from a gem
* global object name sent by another process for buffer sharing.
*
* if true, return a exynos buffer object else NULL.
*
*/
drm_public struct exynos_bo *
exynos_bo_from_name(struct exynos_device *dev, uint32_t name)
{
struct exynos_bo *bo;
struct drm_gem_open req = {
.name = name,
};
bo = calloc(sizeof(*bo), 1);
if (!bo) {
fprintf(stderr, "failed to allocate bo[%s].\n",
strerror(errno));
return NULL;
}
if (drmIoctl(dev->fd, DRM_IOCTL_GEM_OPEN, &req)) {
fprintf(stderr, "failed to open gem object[%s].\n",
strerror(errno));
goto err_free_bo;
}
bo->dev = dev;
bo->name = name;
bo->handle = req.handle;
return bo;
err_free_bo:
free(bo);
return NULL;
}
/*
* Get a gem global object name from a gem object handle.
*
* @bo: a exynos buffer object including gem handle.
* @name: a gem global object name to be got by kernel driver.
*
* this interface is used to get a gem global object name from a gem object
* handle to a buffer that wants to share it with another process.
*
* if true, return 0 else negative.
*/
drm_public int exynos_bo_get_name(struct exynos_bo *bo, uint32_t *name)
{
if (!bo->name) {
struct drm_gem_flink req = {
.handle = bo->handle,
};
int ret;
ret = drmIoctl(bo->dev->fd, DRM_IOCTL_GEM_FLINK, &req);
if (ret) {
fprintf(stderr, "failed to get gem global name[%s].\n",
strerror(errno));
return ret;
}
bo->name = req.name;
}
*name = bo->name;
return 0;
}
drm_public uint32_t exynos_bo_handle(struct exynos_bo *bo)
{
return bo->handle;
}
/*
* Mmap a buffer to user space.
*
* @bo: a exynos buffer object including a gem object handle to be mmapped
* to user space.
*
* if true, user pointer mmaped else NULL.
*/
drm_public void *exynos_bo_map(struct exynos_bo *bo)
{
if (!bo->vaddr) {
struct exynos_device *dev = bo->dev;
struct drm_exynos_gem_mmap req = {
.handle = bo->handle,
.size = bo->size,
};
int ret;
ret = drmIoctl(dev->fd, DRM_IOCTL_EXYNOS_GEM_MMAP, &req);
if (ret) {
fprintf(stderr, "failed to mmap[%s].\n",
strerror(errno));
return NULL;
}
bo->vaddr = (void *)(uintptr_t)req.mapped;
}
return bo->vaddr;
}
/*
* Export gem object to dmabuf as file descriptor.
*
* @dev: exynos device object
* @handle: gem handle to export as file descriptor of dmabuf
* @fd: file descriptor returned from kernel
*
* @return: 0 on success, -1 on error, and errno will be set
*/
drm_public int
exynos_prime_handle_to_fd(struct exynos_device *dev, uint32_t handle, int *fd)
{
return drmPrimeHandleToFD(dev->fd, handle, 0, fd);
}
/*
* Import file descriptor into gem handle.
*
* @dev: exynos device object
* @fd: file descriptor of dmabuf to import
* @handle: gem handle returned from kernel
*
* @return: 0 on success, -1 on error, and errno will be set
*/
drm_public int
exynos_prime_fd_to_handle(struct exynos_device *dev, int fd, uint32_t *handle)
{
return drmPrimeFDToHandle(dev->fd, fd, handle);
}
/*
* Request Wireless Display connection or disconnection.
*
* @dev: a exynos device object.
* @connect: indicate whether connectoin or disconnection request.
* @ext: indicate whether edid data includes extentions data or not.
* @edid: a pointer to edid data from Wireless Display device.
*
* this interface is used to request Virtual Display driver connection or
* disconnection. for this, user should get a edid data from the Wireless
* Display device and then send that data to kernel driver with connection
* request
*
* if true, return 0 else negative.
*/
drm_public int
exynos_vidi_connection(struct exynos_device *dev, uint32_t connect,
uint32_t ext, void *edid)
{
struct drm_exynos_vidi_connection req = {
.connection = connect,
.extensions = ext,
.edid = (uint64_t)(uintptr_t)edid,
};
int ret;
ret = drmIoctl(dev->fd, DRM_IOCTL_EXYNOS_VIDI_CONNECTION, &req);
if (ret) {
fprintf(stderr, "failed to request vidi connection[%s].\n",
strerror(errno));
return ret;
}
return 0;
}
/* Called under table_lock */
drm_private void bo_del(struct fd_bo *bo)
{
VG_BO_FREE(bo);
if (bo->map)
drm_munmap(bo->map, bo->size);
/* TODO probably bo's in bucket list get removed from
* handle table??
*/
if (bo->handle) {
struct drm_gem_close req = {
.handle = bo->handle,
};
drmHashDelete(bo->dev->handle_table, bo->handle);
if (bo->name)
drmHashDelete(bo->dev->name_table, bo->name);
drmIoctl(bo->dev->fd, DRM_IOCTL_GEM_CLOSE, &req);
}
bo->funcs->destroy(bo);
}
int fd_bo_get_name(struct fd_bo *bo, uint32_t *name)
{
if (!bo->name) {
struct drm_gem_flink req = {
.handle = bo->handle,
};
int ret;
ret = drmIoctl(bo->dev->fd, DRM_IOCTL_GEM_FLINK, &req);
if (ret) {
return ret;
}
pthread_mutex_lock(&table_lock);
set_name(bo, req.name);
pthread_mutex_unlock(&table_lock);
bo->bo_reuse = FALSE;
}
*name = bo->name;
return 0;
}
uint32_t fd_bo_handle(struct fd_bo *bo)
{
return bo->handle;
}
int fd_bo_dmabuf(struct fd_bo *bo)
{
int ret, prime_fd;
ret = drmPrimeHandleToFD(bo->dev->fd, bo->handle, DRM_CLOEXEC,
&prime_fd);
if (ret) {
ERROR_MSG("failed to get dmabuf fd: %d", ret);
return ret;
}
bo->bo_reuse = FALSE;
return prime_fd;
}
uint32_t fd_bo_size(struct fd_bo *bo)
{
return bo->size;
}
void * fd_bo_map(struct fd_bo *bo)
{
if (!bo->map) {
uint64_t offset;
int ret;
ret = bo->funcs->offset(bo, &offset);
if (ret) {
return NULL;
}
bo->map = drm_mmap(0, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
bo->dev->fd, offset);
if (bo->map == MAP_FAILED) {
ERROR_MSG("mmap failed: %s", strerror(errno));
bo->map = NULL;
}
}
return bo->map;
}
/* a bit odd to take the pipe as an arg, but it's a, umm, quirk of kgsl.. */
int fd_bo_cpu_prep(struct fd_bo *bo, struct fd_pipe *pipe, uint32_t op)
{
return bo->funcs->cpu_prep(bo, pipe, op);
}
void fd_bo_cpu_fini(struct fd_bo *bo)
{
bo->funcs->cpu_fini(bo);
}
struct renderonly_scanout *
renderonly_create_kms_dumb_buffer_for_resource(struct pipe_resource *rsc,
struct renderonly *ro)
{
struct pipe_screen *screen = rsc->screen;
struct renderonly_scanout *scanout;
struct winsys_handle handle;
int prime_fd, err;
struct drm_mode_create_dumb create_dumb = {
.width = rsc->width0,
.height = rsc->height0,
.bpp = 32,
};
struct drm_mode_destroy_dumb destroy_dumb = { };
scanout = CALLOC_STRUCT(renderonly_scanout);
if (!scanout)
return NULL;
/* create dumb buffer at scanout GPU */
err = ioctl(ro->kms_fd, DRM_IOCTL_MODE_CREATE_DUMB, &create_dumb);
if (err < 0) {
fprintf(stderr, "DRM_IOCTL_MODE_CREATE_DUMB failed: %s\n",
strerror(errno));
goto free_scanout;
}
scanout->handle = create_dumb.handle;
scanout->stride = create_dumb.pitch;
/* export dumb buffer */
err = drmPrimeHandleToFD(ro->kms_fd, create_dumb.handle, O_CLOEXEC,
&prime_fd);
if (err < 0) {
fprintf(stderr, "failed to export dumb buffer: %s\n", strerror(errno));
goto free_dumb;
}
/* import dumb buffer */
handle.type = DRM_API_HANDLE_TYPE_FD;
handle.handle = prime_fd;
handle.stride = create_dumb.pitch;
scanout->prime = screen->resource_from_handle(screen, rsc,
&handle, PIPE_HANDLE_USAGE_READ_WRITE);
if (!scanout->prime) {
fprintf(stderr, "failed to create resource_from_handle: %s\n", strerror(errno));
goto free_dumb;
}
return scanout;
free_dumb:
destroy_dumb.handle = scanout->handle;
ioctl(ro->kms_fd, DRM_IOCTL_MODE_DESTROY_DUMB, &destroy_dumb);
free_scanout:
FREE(scanout);
return NULL;
}
struct renderonly_scanout *
renderonly_create_gpu_import_for_resource(struct pipe_resource *rsc,
struct renderonly *ro)
{
struct pipe_screen *screen = rsc->screen;
struct renderonly_scanout *scanout;
boolean status;
int fd, err;
struct winsys_handle handle = {
.type = DRM_API_HANDLE_TYPE_FD
};
scanout = CALLOC_STRUCT(renderonly_scanout);
if (!scanout)
return NULL;
status = screen->resource_get_handle(screen, NULL, rsc, &handle,
PIPE_HANDLE_USAGE_READ_WRITE);
if (!status)
goto free_scanout;
scanout->stride = handle.stride;
fd = handle.handle;
err = drmPrimeFDToHandle(ro->kms_fd, fd, &scanout->handle);
close(fd);
if (err < 0) {
fprintf(stderr, "drmPrimeFDToHandle() failed: %s\n", strerror(errno));
goto free_scanout;
}
return scanout;
free_scanout:
FREE(scanout);
return NULL;
}
/*
* Destroy a exynos buffer object.
*
* @bo: a exynos buffer object to be destroyed.
*/
drm_public void exynos_bo_destroy(struct exynos_bo *bo)
{
if (!bo)
return;
if (bo->vaddr)
munmap(bo->vaddr, bo->size);
if (bo->handle) {
struct drm_gem_close req = {
.handle = bo->handle,
};
drmIoctl(bo->dev->fd, DRM_IOCTL_GEM_CLOSE, &req);
}
free(bo);
}
/*
* Get a exynos buffer object from a gem global object name.
*
* @dev: a exynos device object.
* @name: a gem global object name exported by another process.
*
* this interface is used to get a exynos buffer object from a gem
* global object name sent by another process for buffer sharing.
*
* if true, return a exynos buffer object else NULL.
*
*/
drm_public struct exynos_bo *
exynos_bo_from_name(struct exynos_device *dev, uint32_t name)
{
struct exynos_bo *bo;
struct drm_gem_open req = {
.name = name,
};
bo = calloc(sizeof(*bo), 1);
if (!bo) {
fprintf(stderr, "failed to allocate bo[%s].\n",
strerror(errno));
return NULL;
}
if (drmIoctl(dev->fd, DRM_IOCTL_GEM_OPEN, &req)) {
fprintf(stderr, "failed to open gem object[%s].\n",
strerror(errno));
goto err_free_bo;
}
bo->dev = dev;
bo->name = name;
bo->handle = req.handle;
return bo;
err_free_bo:
free(bo);
return NULL;
}
/*
* Get a gem global object name from a gem object handle.
*
* @bo: a exynos buffer object including gem handle.
* @name: a gem global object name to be got by kernel driver.
*
* this interface is used to get a gem global object name from a gem object
* handle to a buffer that wants to share it with another process.
*
* if true, return 0 else negative.
*/
drm_public int exynos_bo_get_name(struct exynos_bo *bo, uint32_t *name)
{
if (!bo->name) {
struct drm_gem_flink req = {
.handle = bo->handle,
};
int ret;
ret = drmIoctl(bo->dev->fd, DRM_IOCTL_GEM_FLINK, &req);
if (ret) {
fprintf(stderr, "failed to get gem global name[%s].\n",
strerror(errno));
return ret;
}
bo->name = req.name;
}
*name = bo->name;
return 0;
}
drm_public uint32_t exynos_bo_handle(struct exynos_bo *bo)
{
return bo->handle;
}
/*
* Mmap a buffer to user space.
*
* @bo: a exynos buffer object including a gem object handle to be mmapped
* to user space.
*
* if true, user pointer mmapped else NULL.
*/
drm_public void *exynos_bo_map(struct exynos_bo *bo)
{
if (!bo->vaddr) {
struct exynos_device *dev = bo->dev;
struct drm_mode_map_dumb arg;
void *map = NULL;
int ret;
memset(&arg, 0, sizeof(arg));
arg.handle = bo->handle;
ret = drmIoctl(dev->fd, DRM_IOCTL_MODE_MAP_DUMB, &arg);
if (ret) {
fprintf(stderr, "failed to map dumb buffer[%s].\n",
strerror(errno));
return NULL;
}
map = drm_mmap(0, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
dev->fd, arg.offset);
if (map != MAP_FAILED)
bo->vaddr = map;
}
return bo->vaddr;
}
/*
* Export gem object to dmabuf as file descriptor.
*
* @dev: exynos device object
* @handle: gem handle to export as file descriptor of dmabuf
* @fd: file descriptor returned from kernel
*
* @return: 0 on success, -1 on error, and errno will be set
*/
drm_public int
exynos_prime_handle_to_fd(struct exynos_device *dev, uint32_t handle, int *fd)
{
return drmPrimeHandleToFD(dev->fd, handle, 0, fd);
}
/*
* Import file descriptor into gem handle.
*
* @dev: exynos device object
* @fd: file descriptor of dmabuf to import
* @handle: gem handle returned from kernel
*
* @return: 0 on success, -1 on error, and errno will be set
*/
drm_public int
exynos_prime_fd_to_handle(struct exynos_device *dev, int fd, uint32_t *handle)
{
return drmPrimeFDToHandle(dev->fd, fd, handle);
}
/*
* Request Wireless Display connection or disconnection.
*
* @dev: a exynos device object.
* @connect: indicate whether connectoin or disconnection request.
* @ext: indicate whether edid data includes extensions data or not.
* @edid: a pointer to edid data from Wireless Display device.
*
* this interface is used to request Virtual Display driver connection or
* disconnection. for this, user should get a edid data from the Wireless
* Display device and then send that data to kernel driver with connection
* request
*
* if true, return 0 else negative.
*/
drm_public int
exynos_vidi_connection(struct exynos_device *dev, uint32_t connect,
uint32_t ext, void *edid)
{
struct drm_exynos_vidi_connection req = {
.connection = connect,
.extensions = ext,
.edid = (uint64_t)(uintptr_t)edid,
};
int ret;
ret = drmIoctl(dev->fd, DRM_IOCTL_EXYNOS_VIDI_CONNECTION, &req);
if (ret) {
fprintf(stderr, "failed to request vidi connection[%s].\n",
strerror(errno));
return ret;
}
return 0;
}
static void
exynos_handle_vendor(int fd, struct drm_event *e, void *ctx)
{
struct drm_exynos_g2d_event *g2d;
struct exynos_event_context *ectx = ctx;
switch (e->type) {
case DRM_EXYNOS_G2D_EVENT:
if (ectx->version < 1 || ectx->g2d_event_handler == NULL)
break;
g2d = (struct drm_exynos_g2d_event *)e;
ectx->g2d_event_handler(fd, g2d->cmdlist_no, g2d->tv_sec,
g2d->tv_usec, U642VOID(g2d->user_data));
break;
default:
break;
}
}
drm_public int
exynos_handle_event(struct exynos_device *dev, struct exynos_event_context *ctx)
{
char buffer[1024];
int len, i;
struct drm_event *e;
struct drm_event_vblank *vblank;
drmEventContextPtr evctx = &ctx->base;
/* The DRM read semantics guarantees that we always get only
* complete events. */
len = read(dev->fd, buffer, sizeof buffer);
if (len == 0)
return 0;
if (len < (int)sizeof *e)
return -1;
i = 0;
while (i < len) {
e = (struct drm_event *)(buffer + i);
switch (e->type) {
case DRM_EVENT_VBLANK:
if (evctx->version < 1 ||
evctx->vblank_handler == NULL)
break;
vblank = (struct drm_event_vblank *) e;
evctx->vblank_handler(dev->fd,
vblank->sequence,
vblank->tv_sec,
vblank->tv_usec,
U642VOID (vblank->user_data));
break;
case DRM_EVENT_FLIP_COMPLETE:
if (evctx->version < 2 ||
evctx->page_flip_handler == NULL)
break;
vblank = (struct drm_event_vblank *) e;
evctx->page_flip_handler(dev->fd,
vblank->sequence,
vblank->tv_sec,
vblank->tv_usec,
U642VOID (vblank->user_data));
break;
default:
exynos_handle_vendor(dev->fd, e, evctx);
break;
}
i += e->length;
}
return 0;
}