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 <dlfcn.h>
11 #include <drm_fourcc.h>
12 #include <errno.h>
13 #include <inttypes.h>
14 #include <msm_drm.h>
15 #include <stdbool.h>
16 #include <stdio.h>
17 #include <string.h>
18 #include <sys/mman.h>
19 #include <xf86drm.h>
20
21 #include "drv_priv.h"
22 #include "helpers.h"
23 #include "util.h"
24
25 /* Alignment values are based on SDM845 Gfx IP */
26 #define DEFAULT_ALIGNMENT 64
27 #define BUFFER_SIZE_ALIGN 4096
28
29 #define VENUS_STRIDE_ALIGN 128
30 #define VENUS_SCANLINE_ALIGN 16
31 #define NV12_LINEAR_PADDING (12 * 1024)
32 #define NV12_UBWC_PADDING(y_stride) (MAX(16 * 1024, y_stride * 48))
33 #define MACROTILE_WIDTH_ALIGN 64
34 #define MACROTILE_HEIGHT_ALIGN 16
35 #define PLANE_SIZE_ALIGN 4096
36
37 #define MSM_UBWC_TILING 1
38
39 static const uint32_t render_target_formats[] = { DRM_FORMAT_ABGR8888, DRM_FORMAT_ARGB8888,
40 DRM_FORMAT_RGB565, DRM_FORMAT_XBGR8888,
41 DRM_FORMAT_XRGB8888 };
42
43 static const uint32_t texture_source_formats[] = { DRM_FORMAT_NV12, DRM_FORMAT_R8,
44 DRM_FORMAT_YVU420, DRM_FORMAT_YVU420_ANDROID };
45
46 /*
47 * Each macrotile consists of m x n (mostly 4 x 4) tiles.
48 * Pixel data pitch/stride is aligned with macrotile width.
49 * Pixel data height is aligned with macrotile height.
50 * Entire pixel data buffer is aligned with 4k(bytes).
51 */
get_ubwc_meta_size(uint32_t width,uint32_t height,uint32_t tile_width,uint32_t tile_height)52 static uint32_t get_ubwc_meta_size(uint32_t width, uint32_t height, uint32_t tile_width,
53 uint32_t tile_height)
54 {
55 uint32_t macrotile_width, macrotile_height;
56
57 macrotile_width = DIV_ROUND_UP(width, tile_width);
58 macrotile_height = DIV_ROUND_UP(height, tile_height);
59
60 // Align meta buffer width to 64 blocks
61 macrotile_width = ALIGN(macrotile_width, MACROTILE_WIDTH_ALIGN);
62
63 // Align meta buffer height to 16 blocks
64 macrotile_height = ALIGN(macrotile_height, MACROTILE_HEIGHT_ALIGN);
65
66 return ALIGN(macrotile_width * macrotile_height, PLANE_SIZE_ALIGN);
67 }
68
get_pitch_alignment(struct bo * bo)69 static unsigned get_pitch_alignment(struct bo *bo)
70 {
71 switch (bo->meta.format) {
72 case DRM_FORMAT_NV12:
73 return VENUS_STRIDE_ALIGN;
74 case DRM_FORMAT_YVU420:
75 case DRM_FORMAT_YVU420_ANDROID:
76 /* TODO other YUV formats? */
77 /* Something (in the video stack?) assumes the U/V planes can use
78 * half the pitch as the Y plane.. to componsate, double the
79 * alignment:
80 */
81 return 2 * DEFAULT_ALIGNMENT;
82 default:
83 return DEFAULT_ALIGNMENT;
84 }
85 }
86
msm_calculate_layout(struct bo * bo)87 static void msm_calculate_layout(struct bo *bo)
88 {
89 uint32_t width, height;
90
91 width = bo->meta.width;
92 height = bo->meta.height;
93
94 /* NV12 format requires extra padding with platform
95 * specific alignments for venus driver
96 */
97 if (bo->meta.format == DRM_FORMAT_NV12) {
98 uint32_t y_stride, uv_stride, y_scanline, uv_scanline, y_plane, uv_plane, size,
99 extra_padding;
100
101 y_stride = ALIGN(width, VENUS_STRIDE_ALIGN);
102 uv_stride = ALIGN(width, VENUS_STRIDE_ALIGN);
103 y_scanline = ALIGN(height, VENUS_SCANLINE_ALIGN * 2);
104 uv_scanline = ALIGN(DIV_ROUND_UP(height, 2),
105 VENUS_SCANLINE_ALIGN * (bo->meta.tiling ? 2 : 1));
106 y_plane = y_stride * y_scanline;
107 uv_plane = uv_stride * uv_scanline;
108
109 if (bo->meta.tiling == MSM_UBWC_TILING) {
110 y_plane = ALIGN(y_plane, PLANE_SIZE_ALIGN);
111 uv_plane = ALIGN(uv_plane, PLANE_SIZE_ALIGN);
112 y_plane += get_ubwc_meta_size(width, height, 32, 8);
113 uv_plane += get_ubwc_meta_size(width >> 1, height >> 1, 16, 8);
114 extra_padding = NV12_UBWC_PADDING(y_stride);
115 } else {
116 extra_padding = NV12_LINEAR_PADDING;
117 }
118
119 bo->meta.strides[0] = y_stride;
120 bo->meta.sizes[0] = y_plane;
121 bo->meta.offsets[1] = y_plane;
122 bo->meta.strides[1] = uv_stride;
123 size = y_plane + uv_plane + extra_padding;
124 bo->meta.total_size = ALIGN(size, BUFFER_SIZE_ALIGN);
125 bo->meta.sizes[1] = bo->meta.total_size - bo->meta.sizes[0];
126 } else {
127 uint32_t stride, alignw, alignh;
128
129 alignw = ALIGN(width, get_pitch_alignment(bo));
130 /* HAL_PIXEL_FORMAT_YV12 requires that the buffer's height not be aligned.
131 DRM_FORMAT_R8 of height one is used for JPEG camera output, so don't
132 height align that. */
133 if (bo->meta.format == DRM_FORMAT_YVU420_ANDROID ||
134 (bo->meta.format == DRM_FORMAT_R8 && height == 1)) {
135 alignh = height;
136 } else {
137 alignh = ALIGN(height, DEFAULT_ALIGNMENT);
138 }
139
140 stride = drv_stride_from_format(bo->meta.format, alignw, 0);
141
142 /* Calculate size and assign stride, size, offset to each plane based on format */
143 drv_bo_from_format(bo, stride, alignh, bo->meta.format);
144
145 /* For all RGB UBWC formats */
146 if (bo->meta.tiling == MSM_UBWC_TILING) {
147 bo->meta.sizes[0] += get_ubwc_meta_size(width, height, 16, 4);
148 bo->meta.total_size = bo->meta.sizes[0];
149 assert(IS_ALIGNED(bo->meta.total_size, BUFFER_SIZE_ALIGN));
150 }
151 }
152 }
153
is_ubwc_fmt(uint32_t format)154 static bool is_ubwc_fmt(uint32_t format)
155 {
156 switch (format) {
157 case DRM_FORMAT_XBGR8888:
158 case DRM_FORMAT_ABGR8888:
159 case DRM_FORMAT_XRGB8888:
160 case DRM_FORMAT_ARGB8888:
161 case DRM_FORMAT_NV12:
162 return 1;
163 default:
164 return 0;
165 }
166 }
167
msm_add_ubwc_combinations(struct driver * drv,const uint32_t * formats,uint32_t num_formats,struct format_metadata * metadata,uint64_t use_flags)168 static void msm_add_ubwc_combinations(struct driver *drv, const uint32_t *formats,
169 uint32_t num_formats, struct format_metadata *metadata,
170 uint64_t use_flags)
171 {
172 for (uint32_t i = 0; i < num_formats; i++) {
173 if (is_ubwc_fmt(formats[i])) {
174 struct combination combo = { .format = formats[i],
175 .metadata = *metadata,
176 .use_flags = use_flags };
177 drv_array_append(drv->combos, &combo);
178 }
179 }
180 }
181
182 /**
183 * Check for buggy apps that are known to not support modifiers, to avoid surprising them
184 * with a UBWC buffer.
185 */
should_avoid_ubwc(void)186 static bool should_avoid_ubwc(void)
187 {
188 #ifndef __ANDROID__
189 /* waffle is buggy and, requests a renderable buffer (which on qcom platforms, we
190 * want to use UBWC), and then passes it to the kernel discarding the modifier.
191 * So mesa ends up correctly rendering to as tiled+compressed, but kernel tries
192 * to display as linear. Other platforms do not see this issue, simply because
193 * they only use compressed (ex, AFBC) with the BO_USE_SCANOUT flag.
194 *
195 * See b/163137550
196 */
197 if (dlsym(RTLD_DEFAULT, "waffle_display_connect")) {
198 drv_log("WARNING: waffle detected, disabling UBWC\n");
199 return true;
200 }
201
202 /* The video_decode_accelerator_tests needs to read back the frames
203 * to verify they are correct. The frame verification relies on
204 * computing the MD5 of the video frame. UBWC results in a different
205 * MD5. This turns off UBWC for gtest until a proper frame
206 * comparison can be made
207 * Rely on the same mechanism that waffle is using, but this time check
208 * for a dynamic library function that is present in chrome, but missing
209 * in gtest. Cups is not loaded for video tests.
210 *
211 * See b/171260705
212 */
213 if (!dlsym(RTLD_DEFAULT, "cupsFilePrintf")) {
214 drv_log("WARNING: gtest detected, disabling UBWC\n");
215 return true;
216 }
217 #endif
218 return false;
219 }
220
msm_init(struct driver * drv)221 static int msm_init(struct driver *drv)
222 {
223 struct format_metadata metadata;
224 uint64_t render_use_flags = BO_USE_RENDER_MASK | BO_USE_SCANOUT;
225 uint64_t texture_use_flags = BO_USE_TEXTURE_MASK | BO_USE_HW_VIDEO_DECODER;
226 /*
227 * NOTE: we actually could use tiled in the BO_USE_FRONT_RENDERING case,
228 * if we had a modifier for tiled-but-not-compressed. But we *cannot* use
229 * compressed in this case because the UBWC flags/meta data can be out of
230 * sync with pixel data while the GPU is writing a frame out to memory.
231 */
232 uint64_t sw_flags =
233 (BO_USE_RENDERSCRIPT | BO_USE_SW_MASK | BO_USE_LINEAR | BO_USE_FRONT_RENDERING);
234
235 drv_add_combinations(drv, render_target_formats, ARRAY_SIZE(render_target_formats),
236 &LINEAR_METADATA, render_use_flags);
237
238 drv_add_combinations(drv, texture_source_formats, ARRAY_SIZE(texture_source_formats),
239 &LINEAR_METADATA, texture_use_flags);
240
241 /* The camera stack standardizes on NV12 for YUV buffers. */
242 /* YVU420 and NV12 formats for camera, display and encoding. */
243 drv_modify_combination(drv, DRM_FORMAT_NV12, &LINEAR_METADATA,
244 BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE | BO_USE_SCANOUT |
245 BO_USE_HW_VIDEO_ENCODER);
246
247 /*
248 * R8 format is used for Android's HAL_PIXEL_FORMAT_BLOB and is used for JPEG snapshots
249 * from camera and input/output from hardware decoder/encoder.
250 */
251 drv_modify_combination(drv, DRM_FORMAT_R8, &LINEAR_METADATA,
252 BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE | BO_USE_HW_VIDEO_DECODER |
253 BO_USE_HW_VIDEO_ENCODER);
254
255 /* Android CTS tests require this. */
256 drv_add_combination(drv, DRM_FORMAT_BGR888, &LINEAR_METADATA, BO_USE_SW_MASK);
257
258 drv_modify_linear_combinations(drv);
259
260 if (should_avoid_ubwc() || !drv->compression)
261 return 0;
262
263 metadata.tiling = MSM_UBWC_TILING;
264 metadata.priority = 2;
265 metadata.modifier = DRM_FORMAT_MOD_QCOM_COMPRESSED;
266
267 render_use_flags &= ~sw_flags;
268 texture_use_flags &= ~sw_flags;
269
270 msm_add_ubwc_combinations(drv, render_target_formats, ARRAY_SIZE(render_target_formats),
271 &metadata, render_use_flags);
272
273 msm_add_ubwc_combinations(drv, texture_source_formats, ARRAY_SIZE(texture_source_formats),
274 &metadata, texture_use_flags);
275
276 drv_modify_combination(drv, DRM_FORMAT_NV12, &metadata,
277 BO_USE_SCANOUT | BO_USE_HW_VIDEO_ENCODER);
278
279 return 0;
280 }
281
msm_bo_create_for_modifier(struct bo * bo,uint32_t width,uint32_t height,uint32_t format,const uint64_t modifier)282 static int msm_bo_create_for_modifier(struct bo *bo, uint32_t width, uint32_t height,
283 uint32_t format, const uint64_t modifier)
284 {
285 struct drm_msm_gem_new req = { 0 };
286 int ret;
287 size_t i;
288
289 bo->meta.tiling = (modifier == DRM_FORMAT_MOD_QCOM_COMPRESSED) ? MSM_UBWC_TILING : 0;
290 msm_calculate_layout(bo);
291
292 req.flags = MSM_BO_WC | MSM_BO_SCANOUT;
293 req.size = bo->meta.total_size;
294
295 ret = drmIoctl(bo->drv->fd, DRM_IOCTL_MSM_GEM_NEW, &req);
296 if (ret) {
297 drv_log("DRM_IOCTL_MSM_GEM_NEW failed with %s\n", strerror(errno));
298 return -errno;
299 }
300
301 /*
302 * Though we use only one plane, we need to set handle for
303 * all planes to pass kernel checks
304 */
305 for (i = 0; i < bo->meta.num_planes; i++)
306 bo->handles[i].u32 = req.handle;
307
308 bo->meta.format_modifier = modifier;
309 return 0;
310 }
311
msm_bo_create_with_modifiers(struct bo * bo,uint32_t width,uint32_t height,uint32_t format,const uint64_t * modifiers,uint32_t count)312 static int msm_bo_create_with_modifiers(struct bo *bo, uint32_t width, uint32_t height,
313 uint32_t format, const uint64_t *modifiers, uint32_t count)
314 {
315 static const uint64_t modifier_order[] = {
316 DRM_FORMAT_MOD_QCOM_COMPRESSED,
317 DRM_FORMAT_MOD_LINEAR,
318 };
319
320 uint64_t modifier =
321 drv_pick_modifier(modifiers, count, modifier_order, ARRAY_SIZE(modifier_order));
322
323 if (!bo->drv->compression && modifier == DRM_FORMAT_MOD_QCOM_COMPRESSED)
324 modifier = DRM_FORMAT_MOD_LINEAR;
325
326 return msm_bo_create_for_modifier(bo, width, height, format, modifier);
327 }
328
329 /* 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)330 static int msm_bo_create(struct bo *bo, uint32_t width, uint32_t height, uint32_t format,
331 uint64_t flags)
332 {
333 struct combination *combo = drv_get_combination(bo->drv, format, flags);
334
335 if (!combo) {
336 drv_log("invalid format = %d, flags = %" PRIx64 " combination\n", format, flags);
337 return -EINVAL;
338 }
339
340 return msm_bo_create_for_modifier(bo, width, height, format, combo->metadata.modifier);
341 }
342
msm_bo_map(struct bo * bo,struct vma * vma,size_t plane,uint32_t map_flags)343 static void *msm_bo_map(struct bo *bo, struct vma *vma, size_t plane, uint32_t map_flags)
344 {
345 int ret;
346 struct drm_msm_gem_info req = { 0 };
347
348 req.handle = bo->handles[0].u32;
349 ret = drmIoctl(bo->drv->fd, DRM_IOCTL_MSM_GEM_INFO, &req);
350 if (ret) {
351 drv_log("DRM_IOCLT_MSM_GEM_INFO failed with %s\n", strerror(errno));
352 return MAP_FAILED;
353 }
354 vma->length = bo->meta.total_size;
355
356 return mmap(0, bo->meta.total_size, drv_get_prot(map_flags), MAP_SHARED, bo->drv->fd,
357 req.offset);
358 }
359
360 const struct backend backend_msm = {
361 .name = "msm",
362 .init = msm_init,
363 .bo_create = msm_bo_create,
364 .bo_create_with_modifiers = msm_bo_create_with_modifiers,
365 .bo_destroy = drv_gem_bo_destroy,
366 .bo_import = drv_prime_bo_import,
367 .bo_map = msm_bo_map,
368 .bo_unmap = drv_bo_munmap,
369 .resolve_format = drv_resolve_format_helper,
370 };
371 #endif /* DRV_MSM */
372