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_helpers.h"
22 #include "drv_priv.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, DRM_FORMAT_ABGR2101010,
42 DRM_FORMAT_ABGR16161616F };
43
44 static const uint32_t texture_source_formats[] = { DRM_FORMAT_NV12, DRM_FORMAT_R8,
45 DRM_FORMAT_YVU420, DRM_FORMAT_YVU420_ANDROID,
46 DRM_FORMAT_P010 };
47
48 /*
49 * Each macrotile consists of m x n (mostly 4 x 4) tiles.
50 * Pixel data pitch/stride is aligned with macrotile width.
51 * Pixel data height is aligned with macrotile height.
52 * Entire pixel data buffer is aligned with 4k(bytes).
53 */
get_ubwc_meta_size(uint32_t width,uint32_t height,uint32_t tile_width,uint32_t tile_height)54 static uint32_t get_ubwc_meta_size(uint32_t width, uint32_t height, uint32_t tile_width,
55 uint32_t tile_height)
56 {
57 uint32_t macrotile_width, macrotile_height;
58
59 macrotile_width = DIV_ROUND_UP(width, tile_width);
60 macrotile_height = DIV_ROUND_UP(height, tile_height);
61
62 // Align meta buffer width to 64 blocks
63 macrotile_width = ALIGN(macrotile_width, MACROTILE_WIDTH_ALIGN);
64
65 // Align meta buffer height to 16 blocks
66 macrotile_height = ALIGN(macrotile_height, MACROTILE_HEIGHT_ALIGN);
67
68 return ALIGN(macrotile_width * macrotile_height, PLANE_SIZE_ALIGN);
69 }
70
get_pitch_alignment(struct bo * bo)71 static unsigned get_pitch_alignment(struct bo *bo)
72 {
73 switch (bo->meta.format) {
74 case DRM_FORMAT_NV12:
75 return VENUS_STRIDE_ALIGN;
76 case DRM_FORMAT_P010:
77 return VENUS_STRIDE_ALIGN * 2;
78 case DRM_FORMAT_YVU420:
79 case DRM_FORMAT_YVU420_ANDROID:
80 /* TODO other YUV formats? */
81 /* Something (in the video stack?) assumes the U/V planes can use
82 * half the pitch as the Y plane.. to componsate, double the
83 * alignment:
84 */
85 return 2 * DEFAULT_ALIGNMENT;
86 default:
87 return DEFAULT_ALIGNMENT;
88 }
89 }
90
msm_calculate_layout(struct bo * bo)91 static void msm_calculate_layout(struct bo *bo)
92 {
93 uint32_t width, height;
94
95 width = bo->meta.width;
96 height = bo->meta.height;
97
98 /* NV12 format requires extra padding with platform
99 * specific alignments for venus driver
100 */
101 if (bo->meta.format == DRM_FORMAT_NV12 || bo->meta.format == DRM_FORMAT_P010) {
102 uint32_t y_stride, uv_stride, y_scanline, uv_scanline, y_plane, uv_plane, size,
103 extra_padding;
104
105 // P010 has the same layout as NV12. The difference is that each
106 // pixel in P010 takes 2 bytes, while in NV12 each pixel takes 1 byte.
107 if (bo->meta.format == DRM_FORMAT_P010)
108 width *= 2;
109
110 y_stride = ALIGN(width, get_pitch_alignment(bo));
111 uv_stride = ALIGN(width, get_pitch_alignment(bo));
112 y_scanline = ALIGN(height, VENUS_SCANLINE_ALIGN * 2);
113 uv_scanline = ALIGN(DIV_ROUND_UP(height, 2),
114 VENUS_SCANLINE_ALIGN * (bo->meta.tiling ? 2 : 1));
115 y_plane = y_stride * y_scanline;
116 uv_plane = uv_stride * uv_scanline;
117
118 if (bo->meta.tiling == MSM_UBWC_TILING) {
119 y_plane = ALIGN(y_plane, PLANE_SIZE_ALIGN);
120 uv_plane = ALIGN(uv_plane, PLANE_SIZE_ALIGN);
121 y_plane += get_ubwc_meta_size(width, height, 32, 8);
122 uv_plane += get_ubwc_meta_size(width >> 1, height >> 1, 16, 8);
123 extra_padding = NV12_UBWC_PADDING(y_stride);
124 } else {
125 extra_padding = NV12_LINEAR_PADDING;
126 }
127
128 bo->meta.strides[0] = y_stride;
129 bo->meta.sizes[0] = y_plane;
130 bo->meta.offsets[1] = y_plane;
131 bo->meta.strides[1] = uv_stride;
132 size = y_plane + uv_plane + extra_padding;
133 bo->meta.total_size = ALIGN(size, BUFFER_SIZE_ALIGN);
134 bo->meta.sizes[1] = bo->meta.total_size - bo->meta.sizes[0];
135 } else {
136 uint32_t stride, alignw, alignh;
137
138 alignw = ALIGN(width, get_pitch_alignment(bo));
139 /* HAL_PIXEL_FORMAT_YV12 requires that the buffer's height not be aligned.
140 DRM_FORMAT_R8 of height one is used for JPEG camera output, so don't
141 height align that. */
142 if (bo->meta.format == DRM_FORMAT_YVU420_ANDROID ||
143 bo->meta.format == DRM_FORMAT_YVU420 ||
144 (bo->meta.format == DRM_FORMAT_R8 && height == 1)) {
145 assert(bo->meta.tiling != MSM_UBWC_TILING);
146 alignh = height;
147 } else {
148 alignh = ALIGN(height, DEFAULT_ALIGNMENT);
149 }
150
151 stride = drv_stride_from_format(bo->meta.format, alignw, 0);
152
153 /* Calculate size and assign stride, size, offset to each plane based on format */
154 drv_bo_from_format(bo, stride, 1, alignh, bo->meta.format);
155
156 /* For all RGB UBWC formats */
157 if (bo->meta.tiling == MSM_UBWC_TILING) {
158 bo->meta.sizes[0] += get_ubwc_meta_size(width, height, 16, 4);
159 bo->meta.total_size = bo->meta.sizes[0];
160 assert(IS_ALIGNED(bo->meta.total_size, BUFFER_SIZE_ALIGN));
161 }
162 }
163 }
164
is_ubwc_fmt(uint32_t format)165 static bool is_ubwc_fmt(uint32_t format)
166 {
167 switch (format) {
168 case DRM_FORMAT_XBGR8888:
169 case DRM_FORMAT_ABGR8888:
170 case DRM_FORMAT_XRGB8888:
171 case DRM_FORMAT_ARGB8888:
172 #ifndef QCOM_DISABLE_COMPRESSED_NV12
173 case DRM_FORMAT_NV12:
174 #endif
175 return 1;
176 default:
177 return 0;
178 }
179 }
180
msm_add_ubwc_combinations(struct driver * drv,const uint32_t * formats,uint32_t num_formats,struct format_metadata * metadata,uint64_t use_flags)181 static void msm_add_ubwc_combinations(struct driver *drv, const uint32_t *formats,
182 uint32_t num_formats, struct format_metadata *metadata,
183 uint64_t use_flags)
184 {
185 for (uint32_t i = 0; i < num_formats; i++) {
186 if (is_ubwc_fmt(formats[i])) {
187 struct combination combo = { .format = formats[i],
188 .metadata = *metadata,
189 .use_flags = use_flags };
190 drv_array_append(drv->combos, &combo);
191 }
192 }
193 }
194
195 /**
196 * Check for buggy apps that are known to not support modifiers, to avoid surprising them
197 * with a UBWC buffer.
198 */
should_avoid_ubwc(void)199 static bool should_avoid_ubwc(void)
200 {
201 #ifndef __ANDROID__
202 /* waffle is buggy and, requests a renderable buffer (which on qcom platforms, we
203 * want to use UBWC), and then passes it to the kernel discarding the modifier.
204 * So mesa ends up correctly rendering to as tiled+compressed, but kernel tries
205 * to display as linear. Other platforms do not see this issue, simply because
206 * they only use compressed (ex, AFBC) with the BO_USE_SCANOUT flag.
207 *
208 * See b/163137550
209 */
210 if (dlsym(RTLD_DEFAULT, "waffle_display_connect")) {
211 drv_logi("WARNING: waffle detected, disabling UBWC\n");
212 return true;
213 }
214 #endif
215 return false;
216 }
217
msm_init(struct driver * drv)218 static int msm_init(struct driver *drv)
219 {
220 struct format_metadata metadata;
221 uint64_t render_use_flags = BO_USE_RENDER_MASK | BO_USE_SCANOUT;
222 uint64_t texture_use_flags = BO_USE_TEXTURE_MASK | BO_USE_HW_VIDEO_DECODER;
223 /*
224 * NOTE: we actually could use tiled in the BO_USE_FRONT_RENDERING case,
225 * if we had a modifier for tiled-but-not-compressed. But we *cannot* use
226 * compressed in this case because the UBWC flags/meta data can be out of
227 * sync with pixel data while the GPU is writing a frame out to memory.
228 */
229 uint64_t sw_flags =
230 (BO_USE_RENDERSCRIPT | BO_USE_SW_MASK | BO_USE_LINEAR | BO_USE_FRONT_RENDERING);
231
232 drv_add_combinations(drv, render_target_formats, ARRAY_SIZE(render_target_formats),
233 &LINEAR_METADATA, render_use_flags);
234
235 drv_add_combinations(drv, texture_source_formats, ARRAY_SIZE(texture_source_formats),
236 &LINEAR_METADATA, texture_use_flags);
237
238 /* The camera stack standardizes on NV12 for YUV buffers. */
239 /* YVU420 and NV12 formats for camera, display and encoding. */
240 drv_modify_combination(drv, DRM_FORMAT_NV12, &LINEAR_METADATA,
241 BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE | BO_USE_SCANOUT |
242 BO_USE_HW_VIDEO_ENCODER);
243
244 /*
245 * R8 format is used for Android's HAL_PIXEL_FORMAT_BLOB and is used for JPEG snapshots
246 * from camera and input/output from hardware decoder/encoder.
247 */
248 drv_modify_combination(drv, DRM_FORMAT_R8, &LINEAR_METADATA,
249 BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE | BO_USE_HW_VIDEO_DECODER |
250 BO_USE_HW_VIDEO_ENCODER | BO_USE_GPU_DATA_BUFFER |
251 BO_USE_SENSOR_DIRECT_DATA);
252
253 /*
254 * Android also frequently requests YV12 formats for some camera implementations
255 * (including the external provider implmenetation).
256 */
257 drv_modify_combination(drv, DRM_FORMAT_YVU420_ANDROID, &LINEAR_METADATA,
258 BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE);
259 drv_modify_combination(drv, DRM_FORMAT_YVU420, &LINEAR_METADATA,
260 BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE);
261
262 /* Android CTS tests require this. */
263 drv_add_combination(drv, DRM_FORMAT_BGR888, &LINEAR_METADATA, BO_USE_SW_MASK);
264
265 #ifdef SC_7280
266 drv_modify_combination(drv, DRM_FORMAT_P010, &LINEAR_METADATA,
267 BO_USE_SCANOUT | BO_USE_HW_VIDEO_ENCODER);
268 #endif
269
270 drv_modify_linear_combinations(drv);
271
272 if (should_avoid_ubwc() || !drv->compression)
273 return 0;
274
275 metadata.tiling = MSM_UBWC_TILING;
276 metadata.priority = 2;
277 metadata.modifier = DRM_FORMAT_MOD_QCOM_COMPRESSED;
278
279 render_use_flags &= ~sw_flags;
280 texture_use_flags &= ~sw_flags;
281
282 msm_add_ubwc_combinations(drv, render_target_formats, ARRAY_SIZE(render_target_formats),
283 &metadata, render_use_flags);
284
285 msm_add_ubwc_combinations(drv, texture_source_formats, ARRAY_SIZE(texture_source_formats),
286 &metadata, texture_use_flags);
287
288 drv_modify_combination(drv, DRM_FORMAT_NV12, &metadata,
289 BO_USE_SCANOUT | BO_USE_HW_VIDEO_ENCODER);
290
291 return 0;
292 }
293
msm_bo_create_for_modifier(struct bo * bo,uint32_t width,uint32_t height,uint32_t format,const uint64_t modifier)294 static int msm_bo_create_for_modifier(struct bo *bo, uint32_t width, uint32_t height,
295 uint32_t format, const uint64_t modifier)
296 {
297 struct drm_msm_gem_new req = { 0 };
298 int ret;
299 size_t i;
300
301 bo->meta.tiling = (modifier == DRM_FORMAT_MOD_QCOM_COMPRESSED) ? MSM_UBWC_TILING : 0;
302 msm_calculate_layout(bo);
303
304 req.flags = MSM_BO_WC | MSM_BO_SCANOUT;
305 req.size = bo->meta.total_size;
306
307 ret = drmIoctl(bo->drv->fd, DRM_IOCTL_MSM_GEM_NEW, &req);
308 if (ret) {
309 drv_loge("DRM_IOCTL_MSM_GEM_NEW failed with %s\n", strerror(errno));
310 return -errno;
311 }
312
313 /*
314 * Though we use only one plane, we need to set handle for
315 * all planes to pass kernel checks
316 */
317 for (i = 0; i < bo->meta.num_planes; i++)
318 bo->handles[i].u32 = req.handle;
319
320 bo->meta.format_modifier = modifier;
321 return 0;
322 }
323
msm_bo_create_with_modifiers(struct bo * bo,uint32_t width,uint32_t height,uint32_t format,const uint64_t * modifiers,uint32_t count)324 static int msm_bo_create_with_modifiers(struct bo *bo, uint32_t width, uint32_t height,
325 uint32_t format, const uint64_t *modifiers, uint32_t count)
326 {
327 static const uint64_t modifier_order[] = {
328 DRM_FORMAT_MOD_QCOM_COMPRESSED,
329 DRM_FORMAT_MOD_LINEAR,
330 };
331
332 uint64_t modifier =
333 drv_pick_modifier(modifiers, count, modifier_order, ARRAY_SIZE(modifier_order));
334
335 if (!bo->drv->compression && modifier == DRM_FORMAT_MOD_QCOM_COMPRESSED)
336 modifier = DRM_FORMAT_MOD_LINEAR;
337
338 return msm_bo_create_for_modifier(bo, width, height, format, modifier);
339 }
340
341 /* 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)342 static int msm_bo_create(struct bo *bo, uint32_t width, uint32_t height, uint32_t format,
343 uint64_t flags)
344 {
345 struct combination *combo = drv_get_combination(bo->drv, format, flags);
346
347 if (!combo) {
348 drv_loge("invalid format = %d, flags = %" PRIx64 " combination\n", format, flags);
349 return -EINVAL;
350 }
351
352 return msm_bo_create_for_modifier(bo, width, height, format, combo->metadata.modifier);
353 }
354
msm_bo_map(struct bo * bo,struct vma * vma,uint32_t map_flags)355 static void *msm_bo_map(struct bo *bo, struct vma *vma, uint32_t map_flags)
356 {
357 int ret;
358 struct drm_msm_gem_info req = { 0 };
359
360 if (bo->meta.format_modifier)
361 return MAP_FAILED;
362
363 req.handle = bo->handles[0].u32;
364 ret = drmIoctl(bo->drv->fd, DRM_IOCTL_MSM_GEM_INFO, &req);
365 if (ret) {
366 drv_loge("DRM_IOCLT_MSM_GEM_INFO failed with %s\n", strerror(errno));
367 return MAP_FAILED;
368 }
369 vma->length = bo->meta.total_size;
370
371 return mmap(0, bo->meta.total_size, drv_get_prot(map_flags), MAP_SHARED, bo->drv->fd,
372 req.offset);
373 }
374
375 const struct backend backend_msm = {
376 .name = "msm",
377 .init = msm_init,
378 .bo_create = msm_bo_create,
379 .bo_create_with_modifiers = msm_bo_create_with_modifiers,
380 .bo_destroy = drv_gem_bo_destroy,
381 .bo_import = drv_prime_bo_import,
382 .bo_map = msm_bo_map,
383 .bo_unmap = drv_bo_munmap,
384 .resolve_format_and_use_flags = drv_resolve_format_and_use_flags_helper,
385 };
386 #endif /* DRV_MSM */
387