/**
* This is called when we need to set up GL rendering to a new X window.
*/
boolean
dri_create_buffer(__DRIscreen * sPriv,
__DRIdrawable * dPriv,
const struct gl_config * visual, boolean isPixmap)
{
struct dri_screen *screen = sPriv->driverPrivate;
struct dri_drawable *drawable = NULL;
if (isPixmap)
goto fail; /* not implemented */
drawable = CALLOC_STRUCT(dri_drawable);
if (drawable == NULL)
goto fail;
dri_fill_st_visual(&drawable->stvis, screen, visual);
/* setup the st_framebuffer_iface */
drawable->base.visual = &drawable->stvis;
drawable->base.flush_front = dri_st_framebuffer_flush_front;
drawable->base.validate = dri_st_framebuffer_validate;
drawable->base.st_manager_private = (void *) drawable;
drawable->screen = screen;
drawable->sPriv = sPriv;
drawable->dPriv = dPriv;
drawable->desired_fences = screen->default_throttle_frames;
if (drawable->desired_fences > DRI_SWAP_FENCES_MAX)
drawable->desired_fences = DRI_SWAP_FENCES_MAX;
dPriv->driverPrivate = (void *)drawable;
p_atomic_set(&drawable->base.stamp, 1);
return GL_TRUE;
fail:
FREE(drawable);
return GL_FALSE;
}
static void
st_context_notify_invalid_framebuffer(struct st_context_iface *stctxi,
struct st_framebuffer_iface *stfbi)
{
struct st_context *st = (struct st_context *) stctxi;
struct st_framebuffer *stfb;
/* either draw or read winsys fb */
stfb = st_ws_framebuffer(st->ctx->WinSysDrawBuffer);
if (!stfb || stfb->iface != stfbi)
stfb = st_ws_framebuffer(st->ctx->WinSysReadBuffer);
if (stfb && stfb->iface == stfbi) {
p_atomic_set(&stfb->revalidate, TRUE);
}
else {
/* This function is probably getting called when we've detected a
* change in a window's size but the currently bound context is
* not bound to that window.
* If the st_framebuffer_iface structure had a pointer to the
* corresponding st_framebuffer we'd be able to handle this.
*/
}
}
static struct svga_winsys_surface *
vmw_svga_winsys_surface_create(struct svga_winsys_screen *sws,
SVGA3dSurfaceFlags flags,
SVGA3dSurfaceFormat format,
unsigned usage,
SVGA3dSize size,
uint32 numLayers,
uint32 numMipLevels,
unsigned sampleCount)
{
struct vmw_winsys_screen *vws = vmw_winsys_screen(sws);
struct vmw_svga_winsys_surface *surface;
struct vmw_buffer_desc desc;
struct pb_manager *provider;
uint32_t buffer_size;
memset(&desc, 0, sizeof(desc));
surface = CALLOC_STRUCT(vmw_svga_winsys_surface);
if(!surface)
goto no_surface;
pipe_reference_init(&surface->refcnt, 1);
p_atomic_set(&surface->validated, 0);
surface->screen = vws;
pipe_mutex_init(surface->mutex);
surface->shared = !!(usage & SVGA_SURFACE_USAGE_SHARED);
provider = (surface->shared) ? vws->pools.gmr : vws->pools.mob_fenced;
/*
* Used for the backing buffer GB surfaces, and to approximate
* when to flush on non-GB hosts.
*/
buffer_size = svga3dsurface_get_serialized_size(format, size, numMipLevels,
numLayers);
if (flags & SVGA3D_SURFACE_BIND_STREAM_OUTPUT)
buffer_size += sizeof(SVGA3dDXSOState);
if (buffer_size > vws->ioctl.max_texture_size) {
goto no_sid;
}
if (sws->have_gb_objects) {
SVGAGuestPtr ptr = {0,0};
/*
* If the backing buffer size is small enough, try to allocate a
* buffer out of the buffer cache. Otherwise, let the kernel allocate
* a suitable buffer for us.
*/
if (buffer_size < VMW_TRY_CACHED_SIZE && !surface->shared) {
struct pb_buffer *pb_buf;
surface->size = buffer_size;
desc.pb_desc.alignment = 4096;
desc.pb_desc.usage = 0;
pb_buf = provider->create_buffer(provider, buffer_size, &desc.pb_desc);
surface->buf = vmw_svga_winsys_buffer_wrap(pb_buf);
if (surface->buf && !vmw_gmr_bufmgr_region_ptr(pb_buf, &ptr))
assert(0);
}
surface->sid = vmw_ioctl_gb_surface_create(vws, flags, format, usage,
size, numLayers,
numMipLevels, sampleCount,
ptr.gmrId,
surface->buf ? NULL :
&desc.region);
if (surface->sid == SVGA3D_INVALID_ID && surface->buf) {
/*
* Kernel refused to allocate a surface for us.
* Perhaps something was wrong with our buffer?
* This is really a guard against future new size requirements
* on the backing buffers.
*/
vmw_svga_winsys_buffer_destroy(sws, surface->buf);
surface->buf = NULL;
surface->sid = vmw_ioctl_gb_surface_create(vws, flags, format, usage,
size, numLayers,
numMipLevels, sampleCount,
0, &desc.region);
if (surface->sid == SVGA3D_INVALID_ID)
goto no_sid;
}
/*
* If the kernel created the buffer for us, wrap it into a
* vmw_svga_winsys_buffer.
*/
if (surface->buf == NULL) {
struct pb_buffer *pb_buf;
surface->size = vmw_region_size(desc.region);
desc.pb_desc.alignment = 4096;
desc.pb_desc.usage = VMW_BUFFER_USAGE_SHARED;
pb_buf = provider->create_buffer(provider, surface->size,
&desc.pb_desc);
surface->buf = vmw_svga_winsys_buffer_wrap(pb_buf);
if (surface->buf == NULL) {
vmw_ioctl_region_destroy(desc.region);
vmw_ioctl_surface_destroy(vws, surface->sid);
goto no_sid;
}
}
} else {
surface->sid = vmw_ioctl_surface_create(vws, flags, format, usage,
size, numLayers, numMipLevels,
sampleCount);
if(surface->sid == SVGA3D_INVALID_ID)
goto no_sid;
/* Best estimate for surface size, used for early flushing. */
surface->size = buffer_size;
surface->buf = NULL;
}
return svga_winsys_surface(surface);
no_sid:
if (surface->buf)
vmw_svga_winsys_buffer_destroy(sws, surface->buf);
FREE(surface);
no_surface:
return NULL;
}
static struct svga_winsys_surface *
vmw_drm_surface_from_handle(struct svga_winsys_screen *sws,
struct winsys_handle *whandle,
SVGA3dSurfaceFormat *format)
{
struct vmw_svga_winsys_surface *vsrf;
struct svga_winsys_surface *ssrf;
struct vmw_winsys_screen *vws = vmw_winsys_screen(sws);
union drm_vmw_surface_reference_arg arg;
struct drm_vmw_surface_arg *req = &arg.req;
struct drm_vmw_surface_create_req *rep = &arg.rep;
uint32_t handle = 0;
struct drm_vmw_size size;
SVGA3dSize base_size;
int ret;
int i;
if (whandle->offset != 0) {
fprintf(stderr, "Attempt to import unsupported winsys offset %u\n",
whandle->offset);
return NULL;
}
switch (whandle->type) {
case DRM_API_HANDLE_TYPE_SHARED:
case DRM_API_HANDLE_TYPE_KMS:
handle = whandle->handle;
break;
case DRM_API_HANDLE_TYPE_FD:
ret = drmPrimeFDToHandle(vws->ioctl.drm_fd, whandle->handle,
&handle);
if (ret) {
vmw_error("Failed to get handle from prime fd %d.\n",
(int) whandle->handle);
return NULL;
}
break;
default:
vmw_error("Attempt to import unsupported handle type %d.\n",
whandle->type);
return NULL;
}
memset(&arg, 0, sizeof(arg));
req->sid = handle;
rep->size_addr = (unsigned long)&size;
ret = drmCommandWriteRead(vws->ioctl.drm_fd, DRM_VMW_REF_SURFACE,
&arg, sizeof(arg));
/*
* Need to close the handle we got from prime.
*/
if (whandle->type == DRM_API_HANDLE_TYPE_FD)
vmw_ioctl_surface_destroy(vws, handle);
if (ret) {
/*
* Any attempt to share something other than a surface, like a dumb
* kms buffer, should fail here.
*/
vmw_error("Failed referencing shared surface. SID %d.\n"
"Error %d (%s).\n",
handle, ret, strerror(-ret));
return NULL;
}
if (rep->mip_levels[0] != 1) {
vmw_error("Incorrect number of mipmap levels on shared surface."
" SID %d, levels %d\n",
handle, rep->mip_levels[0]);
goto out_mip;
}
for (i=1; i < DRM_VMW_MAX_SURFACE_FACES; ++i) {
if (rep->mip_levels[i] != 0) {
vmw_error("Incorrect number of faces levels on shared surface."
" SID %d, face %d present.\n",
handle, i);
goto out_mip;
}
}
vsrf = CALLOC_STRUCT(vmw_svga_winsys_surface);
if (!vsrf)
goto out_mip;
pipe_reference_init(&vsrf->refcnt, 1);
p_atomic_set(&vsrf->validated, 0);
vsrf->screen = vws;
vsrf->sid = handle;
ssrf = svga_winsys_surface(vsrf);
*format = rep->format;
/* Estimate usage, for early flushing. */
base_size.width = size.width;
base_size.height = size.height;
base_size.depth = size.depth;
vsrf->size = svga3dsurface_get_serialized_size(rep->format, base_size,
rep->mip_levels[0],
FALSE);
return ssrf;
out_mip:
vmw_ioctl_surface_destroy(vws, handle);
return NULL;
}
/**
* vmw_drm_gb_surface_from_handle - Create a shared surface
*
* @sws: Screen to register the surface with.
* @whandle: struct winsys_handle identifying the kernel surface object
* @format: On successful return points to a value describing the
* surface format.
*
* Returns a refcounted pointer to a struct svga_winsys_surface
* embedded in a struct vmw_svga_winsys_surface on success or NULL
* on failure.
*/
static struct svga_winsys_surface *
vmw_drm_gb_surface_from_handle(struct svga_winsys_screen *sws,
struct winsys_handle *whandle,
SVGA3dSurfaceFormat *format)
{
struct vmw_svga_winsys_surface *vsrf;
struct svga_winsys_surface *ssrf;
struct vmw_winsys_screen *vws = vmw_winsys_screen(sws);
SVGA3dSurfaceFlags flags;
uint32_t mip_levels;
struct vmw_buffer_desc desc;
struct pb_manager *provider = vws->pools.gmr;
struct pb_buffer *pb_buf;
uint32_t handle;
int ret;
if (whandle->offset != 0) {
fprintf(stderr, "Attempt to import unsupported winsys offset %u\n",
whandle->offset);
return NULL;
}
ret = vmw_ioctl_gb_surface_ref(vws, whandle, &flags, format,
&mip_levels, &handle, &desc.region);
if (ret) {
fprintf(stderr, "Failed referencing shared surface. SID %d.\n"
"Error %d (%s).\n",
whandle->handle, ret, strerror(-ret));
return NULL;
}
if (mip_levels != 1) {
fprintf(stderr, "Incorrect number of mipmap levels on shared surface."
" SID %d, levels %d\n",
whandle->handle, mip_levels);
goto out_mip;
}
vsrf = CALLOC_STRUCT(vmw_svga_winsys_surface);
if (!vsrf)
goto out_mip;
pipe_reference_init(&vsrf->refcnt, 1);
p_atomic_set(&vsrf->validated, 0);
vsrf->screen = vws;
vsrf->sid = handle;
vsrf->size = vmw_region_size(desc.region);
/*
* Synchronize backing buffers of shared surfaces using the
* kernel, since we don't pass fence objects around between
* processes.
*/
desc.pb_desc.alignment = 4096;
desc.pb_desc.usage = VMW_BUFFER_USAGE_SHARED | VMW_BUFFER_USAGE_SYNC;
pb_buf = provider->create_buffer(provider, vsrf->size, &desc.pb_desc);
vsrf->buf = vmw_svga_winsys_buffer_wrap(pb_buf);
if (!vsrf->buf)
goto out_no_buf;
ssrf = svga_winsys_surface(vsrf);
return ssrf;
out_no_buf:
FREE(vsrf);
out_mip:
vmw_ioctl_region_destroy(desc.region);
vmw_ioctl_surface_destroy(vws, whandle->handle);
return NULL;
}
static struct pipe_buffer *
vmw_drm_buffer_from_handle(struct drm_api *drm_api,
struct pipe_screen *screen,
const char *name,
unsigned handle)
{
struct vmw_svga_winsys_surface *vsrf;
struct svga_winsys_surface *ssrf;
struct vmw_winsys_screen *vws =
vmw_winsys_screen(svga_winsys_screen(screen));
struct pipe_buffer *buf;
union drm_vmw_surface_reference_arg arg;
struct drm_vmw_surface_arg *req = &arg.req;
struct drm_vmw_surface_create_req *rep = &arg.rep;
int ret;
int i;
/**
* The vmware device specific handle is the hardware SID.
* FIXME: We probably want to move this to the ioctl implementations.
*/
memset(&arg, 0, sizeof(arg));
req->sid = handle;
ret = drmCommandWriteRead(vws->ioctl.drm_fd, DRM_VMW_REF_SURFACE,
&arg, sizeof(arg));
if (ret) {
fprintf(stderr, "Failed referencing shared surface. SID %d.\n"
"Error %d (%s).\n",
handle, ret, strerror(-ret));
return NULL;
}
if (rep->mip_levels[0] != 1) {
fprintf(stderr, "Incorrect number of mipmap levels on shared surface."
" SID %d, levels %d\n",
handle, rep->mip_levels[0]);
goto out_mip;
}
for (i=1; i < DRM_VMW_MAX_SURFACE_FACES; ++i) {
if (rep->mip_levels[i] != 0) {
fprintf(stderr, "Incorrect number of faces levels on shared surface."
" SID %d, face %d present.\n",
handle, i);
goto out_mip;
}
}
vsrf = CALLOC_STRUCT(vmw_svga_winsys_surface);
if (!vsrf)
goto out_mip;
pipe_reference_init(&vsrf->refcnt, 1);
p_atomic_set(&vsrf->validated, 0);
vsrf->sid = handle;
ssrf = svga_winsys_surface(vsrf);
buf = svga_screen_buffer_wrap_surface(screen, rep->format, ssrf);
if (!buf)
vmw_svga_winsys_surface_reference(&vsrf, NULL);
return buf;
out_mip:
vmw_ioctl_surface_destroy(vws, handle);
return NULL;
}
static boolean
st_api_make_current(struct st_api *stapi, struct st_context_iface *stctxi,
struct st_framebuffer_iface *stdrawi,
struct st_framebuffer_iface *streadi)
{
struct st_context *st = (struct st_context *) stctxi;
struct st_framebuffer *stdraw, *stread;
boolean ret;
_glapi_check_multithread();
if (st) {
/* reuse or create the draw fb */
stdraw = st_framebuffer_reuse_or_create(st->ctx->WinSysDrawBuffer,
stdrawi);
if (streadi != stdrawi) {
/* do the same for the read fb */
stread = st_framebuffer_reuse_or_create(st->ctx->WinSysReadBuffer,
streadi);
}
else {
stread = NULL;
/* reuse the draw fb for the read fb */
if (stdraw)
st_framebuffer_reference(&stread, stdraw);
}
if (stdraw && stread) {
if (stctxi != st_api_get_current(stapi)) {
p_atomic_set(&stdraw->revalidate, TRUE);
p_atomic_set(&stread->revalidate, TRUE);
}
st_framebuffer_validate(stdraw, st);
if (stread != stdraw)
st_framebuffer_validate(stread, st);
/* modify the draw/read buffers of the context */
if (stdraw->iface) {
st_visual_to_default_buffer(stdraw->iface->visual,
&st->ctx->Color.DrawBuffer[0], NULL);
}
if (stread->iface) {
st_visual_to_default_buffer(stread->iface->visual,
&st->ctx->Pixel.ReadBuffer, NULL);
}
ret = _mesa_make_current(st->ctx, &stdraw->Base, &stread->Base);
}
else {
struct gl_framebuffer *incomplete = _mesa_get_incomplete_framebuffer();
ret = _mesa_make_current(st->ctx, incomplete, incomplete);
}
st_framebuffer_reference(&stdraw, NULL);
st_framebuffer_reference(&stread, NULL);
}
else {
ret = _mesa_make_current(NULL, NULL, NULL);
}
return ret;
}
/**
* Validate a framebuffer to make sure up-to-date pipe_textures are used.
*/
static void
st_framebuffer_validate(struct st_framebuffer *stfb, struct st_context *st)
{
struct pipe_context *pipe = st->pipe;
struct pipe_resource *textures[ST_ATTACHMENT_COUNT];
uint width, height;
unsigned i;
boolean changed = FALSE;
if (!p_atomic_read(&stfb->revalidate))
return;
/* validate the fb */
if (!stfb->iface->validate(stfb->iface, stfb->statts, stfb->num_statts, textures))
return;
width = stfb->Base.Width;
height = stfb->Base.Height;
for (i = 0; i < stfb->num_statts; i++) {
struct st_renderbuffer *strb;
struct pipe_surface *ps, surf_tmpl;
gl_buffer_index idx;
if (!textures[i])
continue;
idx = attachment_to_buffer_index(stfb->statts[i]);
if (idx >= BUFFER_COUNT) {
pipe_resource_reference(&textures[i], NULL);
continue;
}
strb = st_renderbuffer(stfb->Base.Attachment[idx].Renderbuffer);
assert(strb);
if (strb->texture == textures[i]) {
pipe_resource_reference(&textures[i], NULL);
continue;
}
memset(&surf_tmpl, 0, sizeof(surf_tmpl));
u_surface_default_template(&surf_tmpl, textures[i],
PIPE_BIND_RENDER_TARGET);
ps = pipe->create_surface(pipe, textures[i], &surf_tmpl);
if (ps) {
pipe_surface_reference(&strb->surface, ps);
pipe_resource_reference(&strb->texture, ps->texture);
/* ownership transfered */
pipe_surface_reference(&ps, NULL);
changed = TRUE;
strb->Base.Width = strb->surface->width;
strb->Base.Height = strb->surface->height;
width = strb->Base.Width;
height = strb->Base.Height;
}
pipe_resource_reference(&textures[i], NULL);
}
if (changed) {
st->dirty.st |= ST_NEW_FRAMEBUFFER;
_mesa_resize_framebuffer(st->ctx, &stfb->Base, width, height);
assert(stfb->Base.Width == width);
assert(stfb->Base.Height == height);
}
p_atomic_set(&stfb->revalidate, FALSE);
}
/**
* Bind the context to the given framebuffers.
*/
static boolean
vg_context_bind_framebuffers(struct st_context_iface *stctxi,
struct st_framebuffer_iface *stdrawi,
struct st_framebuffer_iface *streadi)
{
struct vg_context *ctx = (struct vg_context *) stctxi;
struct st_framebuffer *stfb;
enum st_attachment_type strb_att;
/* the draw and read framebuffers must be the same */
if (stdrawi != streadi)
return FALSE;
p_atomic_set(&ctx->draw_buffer_invalid, TRUE);
strb_att = (stdrawi) ? choose_attachment(stdrawi) : ST_ATTACHMENT_INVALID;
if (ctx->draw_buffer) {
stfb = ctx->draw_buffer;
/* free the existing fb */
if (!stdrawi ||
stfb->strb_att != strb_att ||
stfb->strb->format != stdrawi->visual->color_format) {
destroy_renderbuffer(stfb->strb);
destroy_renderbuffer(stfb->dsrb);
FREE(stfb);
ctx->draw_buffer = NULL;
}
}
if (!stdrawi)
return TRUE;
if (strb_att == ST_ATTACHMENT_INVALID)
return FALSE;
/* create a new fb */
if (!ctx->draw_buffer) {
stfb = CALLOC_STRUCT(st_framebuffer);
if (!stfb)
return FALSE;
stfb->strb = create_renderbuffer(stdrawi->visual->color_format);
if (!stfb->strb) {
FREE(stfb);
return FALSE;
}
stfb->dsrb = create_renderbuffer(ctx->ds_format);
if (!stfb->dsrb) {
FREE(stfb->strb);
FREE(stfb);
return FALSE;
}
stfb->width = 0;
stfb->height = 0;
stfb->strb_att = strb_att;
ctx->draw_buffer = stfb;
}
ctx->draw_buffer->iface = stdrawi;
return TRUE;
}
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);
}
}
struct radeon_winsys *radeon_drm_winsys_create(int fd)
{
struct radeon_drm_winsys *ws;
if (!fd_tab) {
fd_tab = util_hash_table_create(hash_fd, compare_fd);
}
ws = util_hash_table_get(fd_tab, intptr_to_pointer(fd));
if (ws) {
pipe_reference(NULL, &ws->base.reference);
return &ws->base;
}
ws = CALLOC_STRUCT(radeon_drm_winsys);
if (!ws) {
return NULL;
}
ws->fd = fd;
util_hash_table_set(fd_tab, intptr_to_pointer(fd), ws);
if (!do_winsys_init(ws))
goto fail;
/* Create managers. */
ws->kman = radeon_bomgr_create(ws);
if (!ws->kman)
goto fail;
ws->cman = pb_cache_manager_create(ws->kman, 1000000);
if (!ws->cman)
goto fail;
if (ws->gen >= DRV_R600) {
ws->surf_man = radeon_surface_manager_new(fd);
if (!ws->surf_man)
goto fail;
}
/* init reference */
pipe_reference_init(&ws->base.reference, 1);
/* Set functions. */
ws->base.destroy = radeon_winsys_destroy;
ws->base.query_info = radeon_query_info;
ws->base.cs_request_feature = radeon_cs_request_feature;
ws->base.surface_init = radeon_drm_winsys_surface_init;
ws->base.surface_best = radeon_drm_winsys_surface_best;
ws->base.query_value = radeon_query_value;
radeon_bomgr_init_functions(ws);
radeon_drm_cs_init_functions(ws);
pipe_mutex_init(ws->hyperz_owner_mutex);
pipe_mutex_init(ws->cmask_owner_mutex);
pipe_mutex_init(ws->cs_stack_lock);
p_atomic_set(&ws->ncs, 0);
pipe_semaphore_init(&ws->cs_queued, 0);
pipe_condvar_init(ws->cs_queue_empty);
if (ws->num_cpus > 1 && debug_get_option_thread())
ws->thread = pipe_thread_create(radeon_drm_cs_emit_ioctl, ws);
return &ws->base;
fail:
if (ws->cman)
ws->cman->destroy(ws->cman);
if (ws->kman)
ws->kman->destroy(ws->kman);
if (ws->surf_man)
radeon_surface_manager_free(ws->surf_man);
FREE(ws);
return NULL;
}