• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2018 The Chromium OS Authors. All rights reserved.
3  * Use of this source code is governed by a BSD-style license that can be
4  * found in the LICENSE file.
5  */
6 
7 #ifdef DRV_MSM
8 
9 #include <assert.h>
10 #include <drm_fourcc.h>
11 #include <errno.h>
12 #include <inttypes.h>
13 #include <msm_drm.h>
14 #include <stdbool.h>
15 #include <stdio.h>
16 #include <string.h>
17 #include <sys/mman.h>
18 #include <xf86drm.h>
19 
20 #include "drv_priv.h"
21 #include "helpers.h"
22 #include "util.h"
23 
24 /* Alignment values are based on SDM845 Gfx IP */
25 #define DEFAULT_ALIGNMENT 64
26 #define BUFFER_SIZE_ALIGN 4096
27 
28 #define VENUS_STRIDE_ALIGN 128
29 #define VENUS_SCANLINE_ALIGN 16
30 #define NV12_LINEAR_PADDING (12 * 1024)
31 #define NV12_UBWC_PADDING(y_stride) (MAX(16 * 1024, y_stride * 48))
32 #define MACROTILE_WIDTH_ALIGN 64
33 #define MACROTILE_HEIGHT_ALIGN 16
34 #define PLANE_SIZE_ALIGN 4096
35 
36 #define MSM_UBWC_TILING 1
37 
38 static const uint32_t render_target_formats[] = { DRM_FORMAT_ABGR8888, DRM_FORMAT_ARGB8888,
39 						  DRM_FORMAT_RGB565, DRM_FORMAT_XBGR8888,
40 						  DRM_FORMAT_XRGB8888 };
41 
42 static const uint32_t texture_source_formats[] = { DRM_FORMAT_NV12, DRM_FORMAT_R8,
43 						   DRM_FORMAT_YVU420, DRM_FORMAT_YVU420_ANDROID };
44 
45 /*
46  * Each macrotile consists of m x n (mostly 4 x 4) tiles.
47  * Pixel data pitch/stride is aligned with macrotile width.
48  * Pixel data height is aligned with macrotile height.
49  * Entire pixel data buffer is aligned with 4k(bytes).
50  */
get_ubwc_meta_size(uint32_t width,uint32_t height,uint32_t tile_width,uint32_t tile_height)51 static uint32_t get_ubwc_meta_size(uint32_t width, uint32_t height, uint32_t tile_width,
52 				   uint32_t tile_height)
53 {
54 	uint32_t macrotile_width, macrotile_height;
55 
56 	macrotile_width = DIV_ROUND_UP(width, tile_width);
57 	macrotile_height = DIV_ROUND_UP(height, tile_height);
58 
59 	// Align meta buffer width to 64 blocks
60 	macrotile_width = ALIGN(macrotile_width, MACROTILE_WIDTH_ALIGN);
61 
62 	// Align meta buffer height to 16 blocks
63 	macrotile_height = ALIGN(macrotile_height, MACROTILE_HEIGHT_ALIGN);
64 
65 	return ALIGN(macrotile_width * macrotile_height, PLANE_SIZE_ALIGN);
66 }
67 
msm_calculate_layout(struct bo * bo)68 static void msm_calculate_layout(struct bo *bo)
69 {
70 	uint32_t width, height;
71 
72 	width = bo->meta.width;
73 	height = bo->meta.height;
74 
75 	/* NV12 format requires extra padding with platform
76 	 * specific alignments for venus driver
77 	 */
78 	if (bo->meta.format == DRM_FORMAT_NV12) {
79 		uint32_t y_stride, uv_stride, y_scanline, uv_scanline, y_plane, uv_plane, size,
80 		    extra_padding;
81 
82 		y_stride = ALIGN(width, VENUS_STRIDE_ALIGN);
83 		uv_stride = ALIGN(width, VENUS_STRIDE_ALIGN);
84 		y_scanline = ALIGN(height, VENUS_SCANLINE_ALIGN * 2);
85 		uv_scanline = ALIGN(DIV_ROUND_UP(height, 2), VENUS_SCANLINE_ALIGN);
86 		y_plane = y_stride * y_scanline;
87 		uv_plane = uv_stride * uv_scanline;
88 
89 		if (bo->meta.tiling == MSM_UBWC_TILING) {
90 			y_plane += get_ubwc_meta_size(width, height, 32, 8);
91 			uv_plane += get_ubwc_meta_size(width >> 1, height >> 1, 16, 8);
92 			extra_padding = NV12_UBWC_PADDING(y_stride);
93 		} else {
94 			extra_padding = NV12_LINEAR_PADDING;
95 		}
96 
97 		bo->meta.strides[0] = y_stride;
98 		bo->meta.sizes[0] = y_plane;
99 		bo->meta.offsets[1] = y_plane;
100 		bo->meta.strides[1] = uv_stride;
101 		size = y_plane + uv_plane + extra_padding;
102 		bo->meta.total_size = ALIGN(size, BUFFER_SIZE_ALIGN);
103 		bo->meta.sizes[1] = bo->meta.total_size - bo->meta.sizes[0];
104 	} else {
105 		uint32_t stride, alignw, alignh;
106 
107 		alignw = ALIGN(width, DEFAULT_ALIGNMENT);
108 		/* HAL_PIXEL_FORMAT_YV12 requires that the buffer's height not be aligned.
109 			DRM_FORMAT_R8 of height one is used for JPEG camera output, so don't
110 			height align that. */
111 		if (bo->meta.format == DRM_FORMAT_YVU420_ANDROID ||
112 		    (bo->meta.format == DRM_FORMAT_R8 && height == 1)) {
113 			alignh = height;
114 		} else {
115 			alignh = ALIGN(height, DEFAULT_ALIGNMENT);
116 		}
117 
118 		stride = drv_stride_from_format(bo->meta.format, alignw, 0);
119 
120 		/* Calculate size and assign stride, size, offset to each plane based on format */
121 		drv_bo_from_format(bo, stride, alignh, bo->meta.format);
122 
123 		/* For all RGB UBWC formats */
124 		if (bo->meta.tiling == MSM_UBWC_TILING) {
125 			bo->meta.sizes[0] += get_ubwc_meta_size(width, height, 16, 4);
126 			bo->meta.total_size = bo->meta.sizes[0];
127 			assert(IS_ALIGNED(bo->meta.total_size, BUFFER_SIZE_ALIGN));
128 		}
129 	}
130 }
131 
is_ubwc_fmt(uint32_t format)132 static bool is_ubwc_fmt(uint32_t format)
133 {
134 	switch (format) {
135 	case DRM_FORMAT_XBGR8888:
136 	case DRM_FORMAT_ABGR8888:
137 	case DRM_FORMAT_XRGB8888:
138 	case DRM_FORMAT_ARGB8888:
139 	case DRM_FORMAT_NV12:
140 		return 1;
141 	default:
142 		return 0;
143 	}
144 }
145 
msm_add_ubwc_combinations(struct driver * drv,const uint32_t * formats,uint32_t num_formats,struct format_metadata * metadata,uint64_t use_flags)146 static void msm_add_ubwc_combinations(struct driver *drv, const uint32_t *formats,
147 				      uint32_t num_formats, struct format_metadata *metadata,
148 				      uint64_t use_flags)
149 {
150 	for (uint32_t i = 0; i < num_formats; i++) {
151 		if (is_ubwc_fmt(formats[i])) {
152 			struct combination combo = { .format = formats[i],
153 						     .metadata = *metadata,
154 						     .use_flags = use_flags };
155 			drv_array_append(drv->combos, &combo);
156 		}
157 	}
158 }
159 
msm_init(struct driver * drv)160 static int msm_init(struct driver *drv)
161 {
162 	struct format_metadata metadata;
163 	uint64_t render_use_flags = BO_USE_RENDER_MASK | BO_USE_SCANOUT;
164 	uint64_t texture_use_flags = BO_USE_TEXTURE_MASK | BO_USE_HW_VIDEO_DECODER;
165 	uint64_t sw_flags = (BO_USE_RENDERSCRIPT | BO_USE_SW_WRITE_OFTEN | BO_USE_SW_READ_OFTEN |
166 			     BO_USE_LINEAR | BO_USE_PROTECTED);
167 
168 	drv_add_combinations(drv, render_target_formats, ARRAY_SIZE(render_target_formats),
169 			     &LINEAR_METADATA, render_use_flags);
170 
171 	drv_add_combinations(drv, texture_source_formats, ARRAY_SIZE(texture_source_formats),
172 			     &LINEAR_METADATA, texture_use_flags);
173 
174 	/*
175 	 * Chrome uses DMA-buf mmap to write to YV12 buffers, which are then accessed by the
176 	 * Video Encoder Accelerator (VEA). It could also support NV12 potentially in the future.
177 	 */
178 	drv_modify_combination(drv, DRM_FORMAT_YVU420, &LINEAR_METADATA, BO_USE_HW_VIDEO_ENCODER);
179 	drv_modify_combination(drv, DRM_FORMAT_NV12, &LINEAR_METADATA, BO_USE_HW_VIDEO_ENCODER);
180 
181 	/* The camera stack standardizes on NV12 for YUV buffers. */
182 	drv_modify_combination(drv, DRM_FORMAT_NV12, &LINEAR_METADATA,
183 			       BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE | BO_USE_SCANOUT);
184 	/*
185 	 * R8 format is used for Android's HAL_PIXEL_FORMAT_BLOB and is used for JPEG snapshots
186 	 * from camera.
187 	 */
188 	drv_modify_combination(drv, DRM_FORMAT_R8, &LINEAR_METADATA,
189 			       BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE);
190 
191 	/* Android CTS tests require this. */
192 	drv_add_combination(drv, DRM_FORMAT_BGR888, &LINEAR_METADATA, BO_USE_SW_MASK);
193 
194 	drv_modify_linear_combinations(drv);
195 
196 	metadata.tiling = MSM_UBWC_TILING;
197 	metadata.priority = 2;
198 	metadata.modifier = DRM_FORMAT_MOD_QCOM_COMPRESSED;
199 
200 	render_use_flags &= ~sw_flags;
201 	texture_use_flags &= ~sw_flags;
202 
203 	msm_add_ubwc_combinations(drv, render_target_formats, ARRAY_SIZE(render_target_formats),
204 				  &metadata, render_use_flags);
205 
206 	msm_add_ubwc_combinations(drv, texture_source_formats, ARRAY_SIZE(texture_source_formats),
207 				  &metadata, texture_use_flags);
208 
209 	return 0;
210 }
211 
msm_bo_create_for_modifier(struct bo * bo,uint32_t width,uint32_t height,uint32_t format,const uint64_t modifier)212 static int msm_bo_create_for_modifier(struct bo *bo, uint32_t width, uint32_t height,
213 				      uint32_t format, const uint64_t modifier)
214 {
215 	struct drm_msm_gem_new req;
216 	int ret;
217 	size_t i;
218 
219 	bo->meta.tiling = (modifier == DRM_FORMAT_MOD_QCOM_COMPRESSED) ? MSM_UBWC_TILING : 0;
220 
221 	msm_calculate_layout(bo);
222 
223 	memset(&req, 0, sizeof(req));
224 	req.flags = MSM_BO_WC | MSM_BO_SCANOUT;
225 	req.size = bo->meta.total_size;
226 
227 	ret = drmIoctl(bo->drv->fd, DRM_IOCTL_MSM_GEM_NEW, &req);
228 	if (ret) {
229 		drv_log("DRM_IOCTL_MSM_GEM_NEW failed with %s\n", strerror(errno));
230 		return -errno;
231 	}
232 
233 	/*
234 	 * Though we use only one plane, we need to set handle for
235 	 * all planes to pass kernel checks
236 	 */
237 	for (i = 0; i < bo->meta.num_planes; i++) {
238 		bo->handles[i].u32 = req.handle;
239 		bo->meta.format_modifiers[i] = modifier;
240 	}
241 
242 	return 0;
243 }
244 
msm_bo_create_with_modifiers(struct bo * bo,uint32_t width,uint32_t height,uint32_t format,const uint64_t * modifiers,uint32_t count)245 static int msm_bo_create_with_modifiers(struct bo *bo, uint32_t width, uint32_t height,
246 					uint32_t format, const uint64_t *modifiers, uint32_t count)
247 {
248 	static const uint64_t modifier_order[] = {
249 		DRM_FORMAT_MOD_QCOM_COMPRESSED,
250 		DRM_FORMAT_MOD_LINEAR,
251 	};
252 
253 	uint64_t modifier =
254 	    drv_pick_modifier(modifiers, count, modifier_order, ARRAY_SIZE(modifier_order));
255 
256 	return msm_bo_create_for_modifier(bo, width, height, format, modifier);
257 }
258 
259 /* msm_bo_create will create linear buffers for now */
msm_bo_create(struct bo * bo,uint32_t width,uint32_t height,uint32_t format,uint64_t flags)260 static int msm_bo_create(struct bo *bo, uint32_t width, uint32_t height, uint32_t format,
261 			 uint64_t flags)
262 {
263 	struct combination *combo = drv_get_combination(bo->drv, format, flags);
264 
265 	if (!combo) {
266 		drv_log("invalid format = %d, flags = %" PRIx64 " combination\n", format, flags);
267 		return -EINVAL;
268 	}
269 
270 	return msm_bo_create_for_modifier(bo, width, height, format, combo->metadata.modifier);
271 }
272 
msm_bo_map(struct bo * bo,struct vma * vma,size_t plane,uint32_t map_flags)273 static void *msm_bo_map(struct bo *bo, struct vma *vma, size_t plane, uint32_t map_flags)
274 {
275 	int ret;
276 	struct drm_msm_gem_info req;
277 
278 	memset(&req, 0, sizeof(req));
279 	req.handle = bo->handles[0].u32;
280 
281 	ret = drmIoctl(bo->drv->fd, DRM_IOCTL_MSM_GEM_INFO, &req);
282 	if (ret) {
283 		drv_log("DRM_IOCLT_MSM_GEM_INFO failed with %s\n", strerror(errno));
284 		return MAP_FAILED;
285 	}
286 	vma->length = bo->meta.total_size;
287 
288 	return mmap(0, bo->meta.total_size, drv_get_prot(map_flags), MAP_SHARED, bo->drv->fd,
289 		    req.offset);
290 }
291 
msm_resolve_format(struct driver * drv,uint32_t format,uint64_t use_flags)292 static uint32_t msm_resolve_format(struct driver *drv, uint32_t format, uint64_t use_flags)
293 {
294 	switch (format) {
295 	case DRM_FORMAT_FLEX_YCbCr_420_888:
296 		return DRM_FORMAT_NV12;
297 	default:
298 		return format;
299 	}
300 }
301 
302 const struct backend backend_msm = {
303 	.name = "msm",
304 	.init = msm_init,
305 	.bo_create = msm_bo_create,
306 	.bo_create_with_modifiers = msm_bo_create_with_modifiers,
307 	.bo_destroy = drv_gem_bo_destroy,
308 	.bo_import = drv_prime_bo_import,
309 	.bo_map = msm_bo_map,
310 	.bo_unmap = drv_bo_munmap,
311 	.resolve_format = msm_resolve_format,
312 };
313 #endif /* DRV_MSM */
314