xref: /aosp_15_r20/external/minigbm/virtgpu_virgl.c (revision d95af8df99a05bcb8679a54dc3ab8e5cd312b38e)
1 /*
2  * Copyright 2017 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 #include <assert.h>
8 #include <errno.h>
9 #include <fcntl.h>
10 #include <pthread.h>
11 #include <stdatomic.h>
12 #include <stdint.h>
13 #include <stdio.h>
14 #include <string.h>
15 #include <sys/mman.h>
16 #include <unistd.h>
17 #include <xf86drm.h>
18 
19 #include "drv_helpers.h"
20 #include "drv_priv.h"
21 #include "external/virgl_hw.h"
22 #include "external/virgl_protocol.h"
23 #include "external/virtgpu_drm.h"
24 #include "external/virtgpu_gfxstream_protocol.h"
25 #include "util.h"
26 #include "virtgpu.h"
27 
28 #define PIPE_TEXTURE_2D 2
29 
30 // This comes from a combination of SwiftShader's VkPhysicalDeviceLimits::maxFramebufferWidth and
31 // VkPhysicalDeviceLimits::maxImageDimension2D (see https://crrev.com/c/1917130).
32 #define ANGLE_ON_SWIFTSHADER_MAX_TEXTURE_2D_SIZE 8192
33 
34 #ifndef MIN
35 #define MIN(a, b) ((a) < (b) ? (a) : (b))
36 #endif
37 #define VIRGL_2D_MAX_TEXTURE_2D_SIZE                                                               \
38 	MIN(ANGLE_ON_SWIFTSHADER_MAX_TEXTURE_2D_SIZE, MESA_LLVMPIPE_MAX_TEXTURE_2D_SIZE)
39 
40 static const uint32_t render_target_formats[] = { DRM_FORMAT_ABGR8888, DRM_FORMAT_ARGB8888,
41 						  DRM_FORMAT_RGB565, DRM_FORMAT_XBGR8888,
42 						  DRM_FORMAT_XRGB8888 };
43 
44 static const uint32_t dumb_texture_source_formats[] = {
45 	DRM_FORMAT_R8,		DRM_FORMAT_R16,		 DRM_FORMAT_YVU420,
46 	DRM_FORMAT_NV12,	DRM_FORMAT_NV21,	 DRM_FORMAT_YVU420_ANDROID,
47 	DRM_FORMAT_ABGR2101010, DRM_FORMAT_ABGR16161616F
48 };
49 
50 static const uint32_t texture_source_formats[] = {
51 	DRM_FORMAT_NV12,	DRM_FORMAT_NV21,	 DRM_FORMAT_R8,
52 	DRM_FORMAT_R16,	        DRM_FORMAT_RG88,	 DRM_FORMAT_YVU420_ANDROID,
53 	DRM_FORMAT_ABGR2101010, DRM_FORMAT_ABGR16161616F
54 };
55 
56 static const uint32_t depth_stencil_formats[] = {
57 	DRM_FORMAT_DEPTH16, DRM_FORMAT_DEPTH24, DRM_FORMAT_DEPTH24_STENCIL8,
58 	DRM_FORMAT_DEPTH32, DRM_FORMAT_DEPTH32_STENCIL8
59 };
60 
61 extern struct virtgpu_param params[];
62 
63 struct virgl_blob_metadata_cache {
64 	struct lru_entry entry;
65 	struct bo_metadata meta;
66 };
67 
68 #define lru_entry_to_metadata(entry) ((struct virgl_blob_metadata_cache *)(void *)(entry))
69 
70 #define MAX_CACHED_FORMATS 128
71 
72 struct virgl_priv {
73 	int caps_is_v2;
74 	union virgl_caps caps;
75 	int caps_is_gfxstream;
76 	struct vulkanCapset gfxstream_vulkan_caps;
77 	int host_gbm_enabled;
78 	atomic_int next_blob_id;
79 
80 	pthread_mutex_t host_blob_format_lock;
81 	struct lru virgl_blob_metadata_cache;
82 };
83 
translate_format(uint32_t drm_fourcc)84 static uint32_t translate_format(uint32_t drm_fourcc)
85 {
86 	switch (drm_fourcc) {
87 	case DRM_FORMAT_BGR888:
88 	case DRM_FORMAT_RGB888:
89 		return VIRGL_FORMAT_R8G8B8_UNORM;
90 	case DRM_FORMAT_XRGB8888:
91 		return VIRGL_FORMAT_B8G8R8X8_UNORM;
92 	case DRM_FORMAT_ARGB8888:
93 		return VIRGL_FORMAT_B8G8R8A8_UNORM;
94 	case DRM_FORMAT_XBGR8888:
95 		return VIRGL_FORMAT_R8G8B8X8_UNORM;
96 	case DRM_FORMAT_ABGR8888:
97 		return VIRGL_FORMAT_R8G8B8A8_UNORM;
98 	case DRM_FORMAT_ABGR16161616F:
99 		return VIRGL_FORMAT_R16G16B16A16_FLOAT;
100 	case DRM_FORMAT_ABGR2101010:
101 		return VIRGL_FORMAT_R10G10B10A2_UNORM;
102 	case DRM_FORMAT_RGB565:
103 		return VIRGL_FORMAT_B5G6R5_UNORM;
104 	case DRM_FORMAT_R8:
105 		return VIRGL_FORMAT_R8_UNORM;
106 	case DRM_FORMAT_R16:
107 		return VIRGL_FORMAT_R16_UNORM;
108 	case DRM_FORMAT_RG88:
109 		return VIRGL_FORMAT_R8G8_UNORM;
110 	case DRM_FORMAT_NV12:
111 		return VIRGL_FORMAT_NV12;
112 	case DRM_FORMAT_NV21:
113 		return VIRGL_FORMAT_NV21;
114 	case DRM_FORMAT_P010:
115 		return VIRGL_FORMAT_P010;
116 	case DRM_FORMAT_YVU420:
117 	case DRM_FORMAT_YVU420_ANDROID:
118 		return VIRGL_FORMAT_YV12;
119 	case DRM_FORMAT_DEPTH16:
120 		return VIRGL_FORMAT_Z16_UNORM;
121 	case DRM_FORMAT_DEPTH24:
122 		return VIRGL_FORMAT_Z24X8_UNORM;
123 	case DRM_FORMAT_DEPTH24_STENCIL8:
124 		return VIRGL_FORMAT_Z24_UNORM_S8_UINT;
125 	case DRM_FORMAT_DEPTH32:
126 		return VIRGL_FORMAT_Z32_FLOAT;
127 	case DRM_FORMAT_DEPTH32_STENCIL8:
128 		return VIRGL_FORMAT_Z32_FLOAT_S8X24_UINT;
129 	default:
130 		drv_loge("Unhandled format:%d\n", drm_fourcc);
131 		return 0;
132 	}
133 }
134 
virgl_bitmask_supports_format(struct virgl_supported_format_mask * supported,uint32_t drm_format)135 static bool virgl_bitmask_supports_format(struct virgl_supported_format_mask *supported,
136 					  uint32_t drm_format)
137 {
138 	uint32_t virgl_format = translate_format(drm_format);
139 	if (!virgl_format)
140 		return false;
141 
142 	uint32_t bitmask_index = virgl_format / 32;
143 	uint32_t bit_index = virgl_format % 32;
144 	return supported->bitmask[bitmask_index] & (1 << bit_index);
145 }
146 
147 // The metadata generated here for emulated buffers is slightly different than the metadata
148 // generated by drv_bo_from_format. In order to simplify transfers in the flush and invalidate
149 // functions below, the emulated buffers are oversized. For example, ignoring stride alignment
150 // requirements to demonstrate, a 6x6 YUV420 image buffer might have the following layout from
151 // drv_bo_from_format:
152 //
153 // | Y | Y | Y | Y | Y | Y |
154 // | Y | Y | Y | Y | Y | Y |
155 // | Y | Y | Y | Y | Y | Y |
156 // | Y | Y | Y | Y | Y | Y |
157 // | Y | Y | Y | Y | Y | Y |
158 // | Y | Y | Y | Y | Y | Y |
159 // | U | U | U | U | U | U |
160 // | U | U | U | V | V | V |
161 // | V | V | V | V | V | V |
162 //
163 // where each plane immediately follows the previous plane in memory. This layout makes it
164 // difficult to compute the transfers needed for example when the middle 2x2 region of the
165 // image is locked and needs to be flushed/invalidated.
166 //
167 // Emulated multi-plane buffers instead have a layout of:
168 //
169 // | Y | Y | Y | Y | Y | Y |
170 // | Y | Y | Y | Y | Y | Y |
171 // | Y | Y | Y | Y | Y | Y |
172 // | Y | Y | Y | Y | Y | Y |
173 // | Y | Y | Y | Y | Y | Y |
174 // | Y | Y | Y | Y | Y | Y |
175 // | U | U | U |   |   |   |
176 // | U | U | U |   |   |   |
177 // | U | U | U |   |   |   |
178 // | V | V | V |   |   |   |
179 // | V | V | V |   |   |   |
180 // | V | V | V |   |   |   |
181 //
182 // where each plane is placed as a sub-image (albeit with a very large stride) in order to
183 // simplify transfers into 3 sub-image transfers for the above example.
184 //
185 // Additional note: the V-plane is not placed to the right of the U-plane due to some
186 // observed failures in media framework code which assumes the V-plane is not
187 // "row-interlaced" with the U-plane.
virgl_get_emulated_metadata(const struct bo * bo,struct bo_metadata * metadata)188 static void virgl_get_emulated_metadata(const struct bo *bo, struct bo_metadata *metadata)
189 {
190 	uint32_t y_plane_height;
191 	uint32_t c_plane_height;
192 	uint32_t original_width = bo->meta.width;
193 	uint32_t original_height = bo->meta.height;
194 
195 	metadata->format = DRM_FORMAT_R8;
196 	switch (bo->meta.format) {
197 	case DRM_FORMAT_NV12:
198 	case DRM_FORMAT_NV21:
199 		// Bi-planar
200 		metadata->num_planes = 2;
201 
202 		y_plane_height = original_height;
203 		c_plane_height = DIV_ROUND_UP(original_height, 2);
204 
205 		metadata->width = original_width;
206 		metadata->height = y_plane_height + c_plane_height;
207 
208 		// Y-plane (full resolution)
209 		metadata->strides[0] = metadata->width;
210 		metadata->offsets[0] = 0;
211 		metadata->sizes[0] = metadata->width * y_plane_height;
212 
213 		// CbCr-plane  (half resolution, interleaved, placed below Y-plane)
214 		metadata->strides[1] = metadata->width;
215 		metadata->offsets[1] = metadata->offsets[0] + metadata->sizes[0];
216 		metadata->sizes[1] = metadata->width * c_plane_height;
217 
218 		metadata->total_size = metadata->width * metadata->height;
219 		break;
220 	case DRM_FORMAT_YVU420:
221 	case DRM_FORMAT_YVU420_ANDROID:
222 		// Tri-planar
223 		metadata->num_planes = 3;
224 
225 		y_plane_height = original_height;
226 		c_plane_height = DIV_ROUND_UP(original_height, 2);
227 
228 		metadata->width = ALIGN(original_width, 32);
229 		metadata->height = y_plane_height + (2 * c_plane_height);
230 
231 		// Y-plane (full resolution)
232 		metadata->strides[0] = metadata->width;
233 		metadata->offsets[0] = 0;
234 		metadata->sizes[0] = metadata->width * original_height;
235 
236 		// Cb-plane (half resolution, placed below Y-plane)
237 		metadata->strides[1] = metadata->width;
238 		metadata->offsets[1] = metadata->offsets[0] + metadata->sizes[0];
239 		metadata->sizes[1] = metadata->width * c_plane_height;
240 
241 		// Cr-plane (half resolution, placed below Cb-plane)
242 		metadata->strides[2] = metadata->width;
243 		metadata->offsets[2] = metadata->offsets[1] + metadata->sizes[1];
244 		metadata->sizes[2] = metadata->width * c_plane_height;
245 
246 		metadata->total_size = metadata->width * metadata->height;
247 		break;
248 	default:
249 		break;
250 	}
251 }
252 
253 struct virtio_transfers_params {
254 	size_t xfers_needed;
255 	struct rectangle xfer_boxes[DRV_MAX_PLANES];
256 };
257 
virgl_get_emulated_transfers_params(const struct bo * bo,const struct rectangle * transfer_box,struct virtio_transfers_params * xfer_params)258 static void virgl_get_emulated_transfers_params(const struct bo *bo,
259 						const struct rectangle *transfer_box,
260 						struct virtio_transfers_params *xfer_params)
261 {
262 	uint32_t y_plane_height;
263 	uint32_t c_plane_height;
264 	struct bo_metadata emulated_metadata = { 0 };
265 
266 	if (transfer_box->x == 0 && transfer_box->y == 0 && transfer_box->width == bo->meta.width &&
267 	    transfer_box->height == bo->meta.height) {
268 		virgl_get_emulated_metadata(bo, &emulated_metadata);
269 
270 		xfer_params->xfers_needed = 1;
271 		xfer_params->xfer_boxes[0].x = 0;
272 		xfer_params->xfer_boxes[0].y = 0;
273 		xfer_params->xfer_boxes[0].width = emulated_metadata.width;
274 		xfer_params->xfer_boxes[0].height = emulated_metadata.height;
275 
276 		return;
277 	}
278 
279 	switch (bo->meta.format) {
280 	case DRM_FORMAT_NV12:
281 	case DRM_FORMAT_NV21:
282 		// Bi-planar
283 		xfer_params->xfers_needed = 2;
284 
285 		y_plane_height = bo->meta.height;
286 		c_plane_height = DIV_ROUND_UP(bo->meta.height, 2);
287 
288 		// Y-plane (full resolution)
289 		xfer_params->xfer_boxes[0].x = transfer_box->x;
290 		xfer_params->xfer_boxes[0].y = transfer_box->y;
291 		xfer_params->xfer_boxes[0].width = transfer_box->width;
292 		xfer_params->xfer_boxes[0].height = transfer_box->height;
293 
294 		// CbCr-plane (half resolution, interleaved, placed below Y-plane)
295 		xfer_params->xfer_boxes[1].x = transfer_box->x;
296 		xfer_params->xfer_boxes[1].y = transfer_box->y + y_plane_height;
297 		xfer_params->xfer_boxes[1].width = transfer_box->width;
298 		xfer_params->xfer_boxes[1].height = DIV_ROUND_UP(transfer_box->height, 2);
299 
300 		break;
301 	case DRM_FORMAT_YVU420:
302 	case DRM_FORMAT_YVU420_ANDROID:
303 		// Tri-planar
304 		xfer_params->xfers_needed = 3;
305 
306 		y_plane_height = bo->meta.height;
307 		c_plane_height = DIV_ROUND_UP(bo->meta.height, 2);
308 
309 		// Y-plane (full resolution)
310 		xfer_params->xfer_boxes[0].x = transfer_box->x;
311 		xfer_params->xfer_boxes[0].y = transfer_box->y;
312 		xfer_params->xfer_boxes[0].width = transfer_box->width;
313 		xfer_params->xfer_boxes[0].height = transfer_box->height;
314 
315 		// Cb-plane (half resolution, placed below Y-plane)
316 		xfer_params->xfer_boxes[1].x = transfer_box->x;
317 		xfer_params->xfer_boxes[1].y = transfer_box->y + y_plane_height;
318 		xfer_params->xfer_boxes[1].width = DIV_ROUND_UP(transfer_box->width, 2);
319 		xfer_params->xfer_boxes[1].height = DIV_ROUND_UP(transfer_box->height, 2);
320 
321 		// Cr-plane (half resolution, placed below Cb-plane)
322 		xfer_params->xfer_boxes[2].x = transfer_box->x;
323 		xfer_params->xfer_boxes[2].y = transfer_box->y + y_plane_height + c_plane_height;
324 		xfer_params->xfer_boxes[2].width = DIV_ROUND_UP(transfer_box->width, 2);
325 		xfer_params->xfer_boxes[2].height = DIV_ROUND_UP(transfer_box->height, 2);
326 
327 		break;
328 	}
329 }
330 
virgl_supports_combination_natively(struct driver * drv,uint32_t drm_format,uint64_t use_flags)331 static bool virgl_supports_combination_natively(struct driver *drv, uint32_t drm_format,
332 						uint64_t use_flags)
333 {
334 	struct virgl_priv *priv = (struct virgl_priv *)drv->priv;
335 
336 	if (priv->caps_is_gfxstream) {
337 		// If the data is invalid or an older version just accept all formats as previously
338 		if (priv->gfxstream_vulkan_caps.protocolVersion == 0 ||
339 		    priv->gfxstream_vulkan_caps.virglSupportedFormats[0] == 0)
340 			return true;
341 		bool supported_format = virgl_bitmask_supports_format(
342 		    (struct virgl_supported_format_mask *)&priv->gfxstream_vulkan_caps
343 			.virglSupportedFormats[0],
344 		    drm_format);
345 		return supported_format;
346 	}
347 	if (priv->caps.max_version == 0)
348 		return true;
349 
350 	if ((use_flags & BO_USE_RENDERING) &&
351 	    !virgl_bitmask_supports_format(&priv->caps.v1.render, drm_format))
352 		return false;
353 
354 	if ((use_flags & BO_USE_TEXTURE) &&
355 	    !virgl_bitmask_supports_format(&priv->caps.v1.sampler, drm_format))
356 		return false;
357 
358 	if ((use_flags & BO_USE_SCANOUT) && priv->caps_is_v2 &&
359 	    !virgl_bitmask_supports_format(&priv->caps.v2.scanout, drm_format))
360 		return false;
361 
362 	return true;
363 }
364 
365 // For virtio backends that do not support formats natively (e.g. multi-planar formats are not
366 // supported in virglrenderer when gbm is unavailable on the host machine), whether or not the
367 // format and usage combination can be handled as a blob (byte buffer).
virgl_supports_combination_through_emulation(struct driver * drv,uint32_t drm_format,uint64_t use_flags)368 static bool virgl_supports_combination_through_emulation(struct driver *drv, uint32_t drm_format,
369 							 uint64_t use_flags)
370 {
371 	struct virgl_priv *priv = (struct virgl_priv *)drv->priv;
372 
373 	// Only enable emulation on non-gbm virtio backends.
374 	if (priv->host_gbm_enabled)
375 		return false;
376 
377 	if (use_flags & (BO_USE_RENDERING | BO_USE_SCANOUT))
378 		return false;
379 
380 	if (!virgl_supports_combination_natively(drv, DRM_FORMAT_R8, use_flags))
381 		return false;
382 
383 	return drm_format == DRM_FORMAT_NV12 || drm_format == DRM_FORMAT_NV21 ||
384 	       drm_format == DRM_FORMAT_YVU420 || drm_format == DRM_FORMAT_YVU420_ANDROID;
385 }
386 
387 // Adds the given buffer combination to the list of supported buffer combinations if the
388 // combination is supported by the virtio backend.
virgl_add_combination(struct driver * drv,uint32_t drm_format,struct format_metadata * metadata,uint64_t use_flags)389 static void virgl_add_combination(struct driver *drv, uint32_t drm_format,
390 				  struct format_metadata *metadata, uint64_t use_flags)
391 {
392 	if (params[param_3d].value) {
393 		if ((use_flags & BO_USE_SCANOUT) &&
394 		    !virgl_supports_combination_natively(drv, drm_format, BO_USE_SCANOUT)) {
395 			drv_logi("Strip scanout on format: %d\n", drm_format);
396 			use_flags &= ~BO_USE_SCANOUT;
397 		}
398 
399 		if (!virgl_supports_combination_natively(drv, drm_format, use_flags) &&
400 		    !virgl_supports_combination_through_emulation(drv, drm_format, use_flags)) {
401 			drv_logi("Skipping unsupported combination format:%d\n", drm_format);
402 			return;
403 		}
404 	}
405 
406 	drv_add_combination(drv, drm_format, metadata, use_flags);
407 }
408 
409 // Adds each given buffer combination to the list of supported buffer combinations if the
410 // combination supported by the virtio backend.
virgl_add_combinations(struct driver * drv,const uint32_t * drm_formats,uint32_t num_formats,struct format_metadata * metadata,uint64_t use_flags)411 static void virgl_add_combinations(struct driver *drv, const uint32_t *drm_formats,
412 				   uint32_t num_formats, struct format_metadata *metadata,
413 				   uint64_t use_flags)
414 {
415 	uint32_t i;
416 
417 	for (i = 0; i < num_formats; i++)
418 		virgl_add_combination(drv, drm_formats[i], metadata, use_flags);
419 }
420 
virgl_2d_dumb_bo_create(struct bo * bo,uint32_t width,uint32_t height,uint32_t format,uint64_t use_flags)421 static int virgl_2d_dumb_bo_create(struct bo *bo, uint32_t width, uint32_t height, uint32_t format,
422 				   uint64_t use_flags)
423 {
424 	if (bo->meta.format != DRM_FORMAT_R8) {
425 		width = ALIGN(width, MESA_LLVMPIPE_TILE_SIZE);
426 		height = ALIGN(height, MESA_LLVMPIPE_TILE_SIZE);
427 	}
428 
429 	return drv_dumb_bo_create_ex(bo, width, height, format, use_flags, BO_QUIRK_DUMB32BPP);
430 }
431 
handle_flag(uint64_t * flag,uint64_t check_flag,uint32_t * bind,uint32_t virgl_bind)432 static inline void handle_flag(uint64_t *flag, uint64_t check_flag, uint32_t *bind,
433 			       uint32_t virgl_bind)
434 {
435 	if ((*flag) & check_flag) {
436 		(*flag) &= ~check_flag;
437 		(*bind) |= virgl_bind;
438 	}
439 }
440 
compute_virgl_bind_flags(uint64_t use_flags)441 static uint32_t compute_virgl_bind_flags(uint64_t use_flags)
442 {
443 	/* In crosvm, VIRGL_BIND_SHARED means minigbm will allocate, not virglrenderer. */
444 	uint32_t bind = VIRGL_BIND_SHARED;
445 
446 	handle_flag(&use_flags, BO_USE_TEXTURE, &bind, VIRGL_BIND_SAMPLER_VIEW);
447 	handle_flag(&use_flags, BO_USE_RENDERING, &bind, VIRGL_BIND_RENDER_TARGET);
448 	handle_flag(&use_flags, BO_USE_SCANOUT, &bind, VIRGL_BIND_SCANOUT);
449 	handle_flag(&use_flags, BO_USE_CURSOR, &bind, VIRGL_BIND_CURSOR);
450 	handle_flag(&use_flags, BO_USE_LINEAR, &bind, VIRGL_BIND_LINEAR);
451 	handle_flag(&use_flags, BO_USE_SENSOR_DIRECT_DATA, &bind, VIRGL_BIND_LINEAR);
452 	handle_flag(&use_flags, BO_USE_GPU_DATA_BUFFER, &bind, VIRGL_BIND_LINEAR);
453 	handle_flag(&use_flags, BO_USE_FRONT_RENDERING, &bind, VIRGL_BIND_LINEAR);
454 
455 	if (use_flags & BO_USE_PROTECTED) {
456 		handle_flag(&use_flags, BO_USE_PROTECTED, &bind, VIRGL_BIND_MINIGBM_PROTECTED);
457 	} else {
458 		// Make sure we don't set both flags, since that could be mistaken for
459 		// protected. Give OFTEN priority over RARELY.
460 		if (use_flags & BO_USE_SW_READ_OFTEN) {
461 			handle_flag(&use_flags, BO_USE_SW_READ_OFTEN, &bind,
462 				    VIRGL_BIND_MINIGBM_SW_READ_OFTEN);
463 		} else {
464 			handle_flag(&use_flags, BO_USE_SW_READ_RARELY, &bind,
465 				    VIRGL_BIND_MINIGBM_SW_READ_RARELY);
466 		}
467 		if (use_flags & BO_USE_SW_WRITE_OFTEN) {
468 			handle_flag(&use_flags, BO_USE_SW_WRITE_OFTEN, &bind,
469 				    VIRGL_BIND_MINIGBM_SW_WRITE_OFTEN);
470 		} else {
471 			handle_flag(&use_flags, BO_USE_SW_WRITE_RARELY, &bind,
472 				    VIRGL_BIND_MINIGBM_SW_WRITE_RARELY);
473 		}
474 	}
475 
476 	handle_flag(&use_flags, BO_USE_CAMERA_WRITE, &bind, VIRGL_BIND_MINIGBM_CAMERA_WRITE);
477 	handle_flag(&use_flags, BO_USE_CAMERA_READ, &bind, VIRGL_BIND_MINIGBM_CAMERA_READ);
478 	handle_flag(&use_flags, BO_USE_HW_VIDEO_DECODER, &bind,
479 		    VIRGL_BIND_MINIGBM_HW_VIDEO_DECODER);
480 	handle_flag(&use_flags, BO_USE_HW_VIDEO_ENCODER, &bind,
481 		    VIRGL_BIND_MINIGBM_HW_VIDEO_ENCODER);
482 
483 	if (use_flags)
484 		drv_loge("Unhandled bo use flag: %llx\n", (unsigned long long)use_flags);
485 
486 	return bind;
487 }
488 
virgl_3d_bo_create(struct bo * bo,uint32_t width,uint32_t height,uint32_t format,uint64_t use_flags)489 static int virgl_3d_bo_create(struct bo *bo, uint32_t width, uint32_t height, uint32_t format,
490 			      uint64_t use_flags)
491 {
492 	int ret;
493 	size_t i;
494 	uint32_t stride;
495 	struct drm_virtgpu_resource_create res_create = { 0 };
496 	struct bo_metadata emulated_metadata = { 0 };
497 
498 	if (virgl_supports_combination_natively(bo->drv, format, use_flags)) {
499 		stride = drv_stride_from_format(format, width, 0);
500 		drv_bo_from_format(bo, stride, 1, height, format);
501 	} else {
502 		assert(virgl_supports_combination_through_emulation(bo->drv, format, use_flags));
503 
504 		virgl_get_emulated_metadata(bo, &emulated_metadata);
505 
506 		format = emulated_metadata.format;
507 		width = emulated_metadata.width;
508 		height = emulated_metadata.height;
509 		for (i = 0; i < emulated_metadata.num_planes; i++) {
510 			bo->meta.strides[i] = emulated_metadata.strides[i];
511 			bo->meta.offsets[i] = emulated_metadata.offsets[i];
512 			bo->meta.sizes[i] = emulated_metadata.sizes[i];
513 		}
514 		bo->meta.total_size = emulated_metadata.total_size;
515 	}
516 
517 	/*
518 	 * Setting the target is intended to ensure this resource gets bound as a 2D
519 	 * texture in the host renderer's GL state. All of these resource properties are
520 	 * sent unchanged by the kernel to the host, which in turn sends them unchanged to
521 	 * virglrenderer. When virglrenderer makes a resource, it will convert the target
522 	 * enum to the equivalent one in GL and then bind the resource to that target.
523 	 */
524 
525 	res_create.target = PIPE_TEXTURE_2D;
526 	res_create.format = translate_format(format);
527 	res_create.bind = compute_virgl_bind_flags(use_flags);
528 	res_create.width = width;
529 	res_create.height = height;
530 
531 	/* For virgl 3D */
532 	res_create.depth = 1;
533 	res_create.array_size = 1;
534 	res_create.last_level = 0;
535 	res_create.nr_samples = 0;
536 
537 	res_create.size = ALIGN(bo->meta.total_size, PAGE_SIZE); // PAGE_SIZE = 0x1000
538 	ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_RESOURCE_CREATE, &res_create);
539 	if (ret) {
540 		drv_loge("DRM_IOCTL_VIRTGPU_RESOURCE_CREATE failed with %s\n", strerror(errno));
541 		return ret;
542 	}
543 
544 	bo->handle.u32 = res_create.bo_handle;
545 
546 	return 0;
547 }
548 
virgl_3d_bo_map(struct bo * bo,struct vma * vma,uint32_t map_flags)549 static void *virgl_3d_bo_map(struct bo *bo, struct vma *vma, uint32_t map_flags)
550 {
551 	int ret;
552 	struct drm_virtgpu_map gem_map = { 0 };
553 
554 	gem_map.handle = bo->handle.u32;
555 	ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_MAP, &gem_map);
556 	if (ret) {
557 		drv_loge("DRM_IOCTL_VIRTGPU_MAP failed with %s\n", strerror(errno));
558 		return MAP_FAILED;
559 	}
560 
561 	vma->length = bo->meta.total_size;
562 	return mmap(0, bo->meta.total_size, drv_get_prot(map_flags), MAP_SHARED, bo->drv->fd,
563 		    gem_map.offset);
564 }
565 
virgl_3d_get_max_texture_2d_size(struct driver * drv)566 static uint32_t virgl_3d_get_max_texture_2d_size(struct driver *drv)
567 {
568 	struct virgl_priv *priv = (struct virgl_priv *)drv->priv;
569 
570 	if (priv->caps.v2.max_texture_2d_size)
571 		return priv->caps.v2.max_texture_2d_size;
572 
573 	return UINT32_MAX;
574 }
575 
virgl_get_caps(struct driver * drv,struct virgl_priv * priv)576 static int virgl_get_caps(struct driver *drv, struct virgl_priv *priv)
577 {
578 	int ret;
579 	struct drm_virtgpu_get_caps cap_args = { 0 };
580 
581 	memset(&priv->caps, 0, sizeof(union virgl_caps));
582 	priv->caps_is_v2 = 0;
583 	memset(&priv->gfxstream_vulkan_caps, 0, sizeof(struct vulkanCapset));
584 
585 	if (params[param_supported_capset_ids].value) {
586 		drv_logi("Supported CAPSET IDs: %u.", params[param_supported_capset_ids].value);
587 		if (params[param_supported_capset_ids].value & (1 << VIRTIO_GPU_CAPSET_VIRGL2)) {
588 			priv->caps_is_v2 = 1;
589 		} else if (params[param_supported_capset_ids].value &
590 			   (1 << VIRTIO_GPU_CAPSET_VIRGL)) {
591 			priv->caps_is_v2 = 0;
592 		} else if (params[param_supported_capset_ids].value &
593 			   (1 << VIRTIO_GPU_CAPSET_GFXSTREAM_VULKAN)) {
594 			priv->caps_is_gfxstream = 1;
595 		} else {
596 			drv_logi("Unrecognized CAPSET IDs: %u. Assuming all zero caps.",
597 				 params[param_supported_capset_ids].value);
598 			return 0;
599 		}
600 	} else if (params[param_capset_fix].value) {
601 		priv->caps_is_v2 = 1;
602 	}
603 
604 	cap_args.addr = (unsigned long long)&priv->caps;
605 	if (priv->caps_is_v2) {
606 		cap_args.cap_set_id = VIRTIO_GPU_CAPSET_VIRGL2;
607 		cap_args.size = sizeof(union virgl_caps);
608 	} else if (priv->caps_is_gfxstream) {
609 		cap_args.addr = (unsigned long long)&priv->gfxstream_vulkan_caps;
610 		cap_args.cap_set_id = VIRTIO_GPU_CAPSET_GFXSTREAM_VULKAN;
611 		cap_args.size = sizeof(struct vulkanCapset);
612 	} else {
613 		cap_args.cap_set_id = VIRTIO_GPU_CAPSET_VIRGL;
614 		cap_args.size = sizeof(struct virgl_caps_v1);
615 	}
616 
617 	ret = drmIoctl(drv->fd, DRM_IOCTL_VIRTGPU_GET_CAPS, &cap_args);
618 	if (ret) {
619 		drv_loge("DRM_IOCTL_VIRTGPU_GET_CAPS failed with %s\n", strerror(errno));
620 		priv->caps_is_v2 = 0;
621 		priv->caps_is_gfxstream = 0;
622 		cap_args.addr = (unsigned long long)&priv->caps;
623 
624 		// Fallback to v1
625 		cap_args.cap_set_id = VIRTIO_GPU_CAPSET_VIRGL;
626 		cap_args.size = sizeof(struct virgl_caps_v1);
627 
628 		ret = drmIoctl(drv->fd, DRM_IOCTL_VIRTGPU_GET_CAPS, &cap_args);
629 		if (ret)
630 			drv_loge("DRM_IOCTL_VIRTGPU_GET_CAPS failed with %s\n", strerror(errno));
631 	}
632 
633 	return ret;
634 }
635 
virgl_init_params_and_caps(struct driver * drv)636 static void virgl_init_params_and_caps(struct driver *drv)
637 {
638 	struct virgl_priv *priv = (struct virgl_priv *)drv->priv;
639 	if (params[param_3d].value) {
640 		virgl_get_caps(drv, priv);
641 
642 		// We use two criteria to determine whether host minigbm is used on the host for
643 		// swapchain allocations.
644 		//
645 		// (1) Host minigbm is only available via virglrenderer, and only virglrenderer
646 		//     advertises capabilities.
647 		// (2) Only host minigbm doesn't emulate YUV formats.  Checking this is a bit of a
648 		//     proxy, but it works.
649 		priv->host_gbm_enabled =
650 		    priv->caps.max_version > 0 &&
651 		    virgl_supports_combination_natively(drv, DRM_FORMAT_NV12, BO_USE_TEXTURE);
652 	}
653 }
654 
virgl_init(struct driver * drv)655 static int virgl_init(struct driver *drv)
656 {
657 	struct virgl_priv *priv;
658 
659 	priv = calloc(1, sizeof(*priv));
660 	if (!priv)
661 		return -ENOMEM;
662 
663 	int ret = pthread_mutex_init(&priv->host_blob_format_lock, NULL);
664 	if (ret)
665 		return ret;
666 
667 	drv->priv = priv;
668 	lru_init(&priv->virgl_blob_metadata_cache, MAX_CACHED_FORMATS);
669 
670 	virgl_init_params_and_caps(drv);
671 
672 	if (params[param_3d].value) {
673 		/* This doesn't mean host can scanout everything, it just means host
674 		 * hypervisor can show it. */
675 		virgl_add_combinations(drv, render_target_formats,
676 				       ARRAY_SIZE(render_target_formats), &LINEAR_METADATA,
677 				       BO_USE_RENDER_MASK | BO_USE_SCANOUT);
678 		virgl_add_combinations(drv, texture_source_formats,
679 				       ARRAY_SIZE(texture_source_formats), &LINEAR_METADATA,
680 				       BO_USE_TEXTURE_MASK);
681 		virgl_add_combinations(drv, depth_stencil_formats,
682 				       ARRAY_SIZE(depth_stencil_formats), &LINEAR_METADATA,
683 				       BO_USE_GPU_HW);
684 		/* NV12 with scanout must flow through virgl_add_combination, so that the native
685 		 * support is checked and scanout use_flag can be conditionally stripped. */
686 		virgl_add_combination(drv, DRM_FORMAT_NV12, &LINEAR_METADATA,
687 				      BO_USE_TEXTURE_MASK | BO_USE_CAMERA_READ |
688 					  BO_USE_CAMERA_WRITE | BO_USE_HW_VIDEO_DECODER |
689 					  BO_USE_HW_VIDEO_ENCODER | BO_USE_SCANOUT);
690 	} else {
691 		/* Virtio primary plane only allows this format. */
692 		virgl_add_combination(drv, DRM_FORMAT_XRGB8888, &LINEAR_METADATA,
693 				      BO_USE_RENDER_MASK | BO_USE_SCANOUT);
694 		/* Virtio cursor plane only allows this format and Chrome cannot live without
695 		 * ARGB888 renderable format. */
696 		virgl_add_combination(drv, DRM_FORMAT_ARGB8888, &LINEAR_METADATA,
697 				      BO_USE_RENDER_MASK | BO_USE_CURSOR);
698 		/* Android needs more, but they cannot be bound as scanouts anymore after
699 		 * "drm/virtio: fix DRM_FORMAT_* handling" */
700 		virgl_add_combinations(drv, render_target_formats,
701 				       ARRAY_SIZE(render_target_formats), &LINEAR_METADATA,
702 				       BO_USE_RENDER_MASK);
703 		virgl_add_combinations(drv, dumb_texture_source_formats,
704 				       ARRAY_SIZE(dumb_texture_source_formats), &LINEAR_METADATA,
705 				       BO_USE_TEXTURE_MASK);
706 		drv_modify_combination(drv, DRM_FORMAT_NV12, &LINEAR_METADATA,
707 				       BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE |
708 					   BO_USE_HW_VIDEO_DECODER | BO_USE_HW_VIDEO_ENCODER);
709 	}
710 
711 	/* Android CTS tests require this. */
712 	virgl_add_combination(drv, DRM_FORMAT_RGB888, &LINEAR_METADATA, BO_USE_SW_MASK);
713 	virgl_add_combination(drv, DRM_FORMAT_BGR888, &LINEAR_METADATA, BO_USE_SW_MASK);
714 	/* Android Camera CTS tests requires this. Additionally, the scanout usage is needed for
715 	 * Camera preview and is expected to be conditionally stripped by virgl_add_combination
716 	 * when not natively supported and instead handled by HWComposer. */
717 	virgl_add_combination(drv, DRM_FORMAT_P010, &LINEAR_METADATA,
718 			      BO_USE_SCANOUT | BO_USE_TEXTURE | BO_USE_SW_MASK |
719 				  BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE);
720 	/* Android VTS sensors hal tests require BO_USE_SENSOR_DIRECT_DATA. */
721 	drv_modify_combination(drv, DRM_FORMAT_R8, &LINEAR_METADATA,
722 			       BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE | BO_USE_HW_VIDEO_DECODER |
723 				   BO_USE_HW_VIDEO_ENCODER | BO_USE_SENSOR_DIRECT_DATA |
724 				   BO_USE_GPU_DATA_BUFFER);
725 
726 	if (!priv->host_gbm_enabled) {
727 		drv_modify_combination(drv, DRM_FORMAT_ABGR8888, &LINEAR_METADATA,
728 				       BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE |
729 					   BO_USE_HW_VIDEO_DECODER | BO_USE_HW_VIDEO_ENCODER);
730 		drv_modify_combination(drv, DRM_FORMAT_XBGR8888, &LINEAR_METADATA,
731 				       BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE |
732 					   BO_USE_HW_VIDEO_DECODER | BO_USE_HW_VIDEO_ENCODER);
733 		drv_modify_combination(drv, DRM_FORMAT_NV21, &LINEAR_METADATA,
734 				       BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE |
735 					   BO_USE_HW_VIDEO_DECODER | BO_USE_HW_VIDEO_ENCODER);
736 		drv_modify_combination(drv, DRM_FORMAT_R16, &LINEAR_METADATA,
737 				       BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE |
738 					   BO_USE_HW_VIDEO_DECODER);
739 		drv_modify_combination(drv, DRM_FORMAT_YVU420, &LINEAR_METADATA,
740 				       BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE |
741 					   BO_USE_HW_VIDEO_DECODER | BO_USE_HW_VIDEO_ENCODER);
742 		drv_modify_combination(drv, DRM_FORMAT_YVU420_ANDROID, &LINEAR_METADATA,
743 				       BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE |
744 					   BO_USE_HW_VIDEO_DECODER | BO_USE_HW_VIDEO_ENCODER);
745 	}
746 
747 	return drv_modify_linear_combinations(drv);
748 }
749 
virgl_close(struct driver * drv)750 static void virgl_close(struct driver *drv)
751 {
752 	free(drv->priv);
753 	drv->priv = NULL;
754 }
755 
blob_flags_from_use_flags(uint32_t use_flags)756 static uint32_t blob_flags_from_use_flags(uint32_t use_flags)
757 {
758 	uint32_t blob_flags = VIRTGPU_BLOB_FLAG_USE_SHAREABLE;
759 	if (use_flags & (BO_USE_SW_MASK | BO_USE_GPU_DATA_BUFFER))
760 		blob_flags |= VIRTGPU_BLOB_FLAG_USE_MAPPABLE;
761 
762 	// For now, all blob use cases are cross device. When we add wider
763 	// support for blobs, we can revisit making this unconditional.
764 	blob_flags |= VIRTGPU_BLOB_FLAG_USE_CROSS_DEVICE;
765 
766 	return blob_flags;
767 }
768 
virgl_blob_metadata_eq(struct lru_entry * entry,void * data)769 static bool virgl_blob_metadata_eq(struct lru_entry *entry, void *data)
770 {
771 	struct virgl_blob_metadata_cache *e = lru_entry_to_metadata(entry);
772 	struct bo_metadata *meta = data;
773 	uint32_t virgl_format1 = translate_format(e->meta.format);
774 	uint32_t virgl_format2 = translate_format(meta->format);
775 
776 	return e->meta.height == meta->height && e->meta.width == meta->width &&
777 	       e->meta.use_flags == meta->use_flags && virgl_format1 == virgl_format2;
778 }
779 
virgl_blob_do_create(struct driver * drv,uint32_t width,uint32_t height,uint32_t use_flags,uint32_t virgl_format,uint32_t total_size,uint32_t * bo_handle)780 static int virgl_blob_do_create(struct driver *drv, uint32_t width, uint32_t height,
781 				uint32_t use_flags, uint32_t virgl_format, uint32_t total_size,
782 				uint32_t *bo_handle)
783 {
784 	int ret;
785 	uint32_t cur_blob_id;
786 	uint32_t cmd[VIRGL_PIPE_RES_CREATE_SIZE + 1] = { 0 };
787 	struct drm_virtgpu_resource_create_blob drm_rc_blob = { 0 };
788 	struct virgl_priv *priv = (struct virgl_priv *)drv->priv;
789 	uint32_t virgl_bind_flags = compute_virgl_bind_flags(use_flags);
790 	uint32_t blob_flags = blob_flags_from_use_flags(use_flags);
791 
792 	cur_blob_id = atomic_fetch_add(&priv->next_blob_id, 1);
793 
794 	cmd[0] = VIRGL_CMD0(VIRGL_CCMD_PIPE_RESOURCE_CREATE, 0, VIRGL_PIPE_RES_CREATE_SIZE);
795 	cmd[VIRGL_PIPE_RES_CREATE_TARGET] = PIPE_TEXTURE_2D;
796 	cmd[VIRGL_PIPE_RES_CREATE_WIDTH] = width;
797 	cmd[VIRGL_PIPE_RES_CREATE_HEIGHT] = height;
798 	cmd[VIRGL_PIPE_RES_CREATE_FORMAT] = virgl_format;
799 	cmd[VIRGL_PIPE_RES_CREATE_BIND] = virgl_bind_flags;
800 	cmd[VIRGL_PIPE_RES_CREATE_DEPTH] = 1;
801 	cmd[VIRGL_PIPE_RES_CREATE_BLOB_ID] = cur_blob_id;
802 
803 	drm_rc_blob.cmd = (uint64_t)&cmd;
804 	drm_rc_blob.cmd_size = 4 * (VIRGL_PIPE_RES_CREATE_SIZE + 1);
805 	drm_rc_blob.size = total_size;
806 	drm_rc_blob.blob_mem = VIRTGPU_BLOB_MEM_HOST3D;
807 	drm_rc_blob.blob_flags = blob_flags;
808 	drm_rc_blob.blob_id = cur_blob_id;
809 
810 	ret = drmIoctl(drv->fd, DRM_IOCTL_VIRTGPU_RESOURCE_CREATE_BLOB, &drm_rc_blob);
811 	if (ret < 0) {
812 		drv_loge("DRM_VIRTGPU_RESOURCE_CREATE_BLOB failed with %s\n", strerror(errno));
813 		return -errno;
814 	}
815 
816 	*bo_handle = drm_rc_blob.bo_handle;
817 	return 0;
818 }
819 
820 // Queries the host layout for the requested buffer metadata.
821 //
822 // Of particular interest is total_size. This value is passed to the kernel when creating
823 // a buffer via drm_virtgpu_resource_create_blob.size, to specify how much "vram" to
824 // allocate for use when exposing the host buffer to the guest. As such, we need to know
825 // this value before allocating a buffer to ensure that the full host buffer is actually
826 // visible to the guest.
827 //
828 // Note that we can't reuse these test buffers as actual allocations because our guess for
829 // total_size is insufficient if width!=stride or padding!=0.
virgl_blob_get_host_format(struct driver * drv,struct bo_metadata * meta)830 static int virgl_blob_get_host_format(struct driver *drv, struct bo_metadata *meta)
831 {
832 	struct virgl_priv *priv = (struct virgl_priv *)drv->priv;
833 	int num_planes = drv_num_planes_from_format(meta->format);
834 
835 	pthread_mutex_lock(&priv->host_blob_format_lock);
836 	if (meta->format == DRM_FORMAT_R8) {
837 		meta->offsets[0] = 0;
838 		meta->sizes[0] = meta->width;
839 		meta->strides[0] = meta->width;
840 		meta->total_size = meta->width;
841 	} else {
842 		uint32_t virgl_format = translate_format(meta->format);
843 		struct virgl_blob_metadata_cache *entry;
844 
845 		entry = lru_entry_to_metadata(
846 		    lru_find(&priv->virgl_blob_metadata_cache, virgl_blob_metadata_eq, meta));
847 
848 		if (!entry) {
849 			uint32_t total_size = 0;
850 			for (int i = 0; i < num_planes; i++) {
851 				uint32_t stride =
852 				    drv_stride_from_format(meta->format, meta->width, i);
853 				total_size +=
854 				    drv_size_from_format(meta->format, stride, meta->height, i);
855 			}
856 
857 			uint32_t handle;
858 			int ret =
859 			    virgl_blob_do_create(drv, meta->width, meta->height, meta->use_flags,
860 						 virgl_format, total_size, &handle);
861 			if (ret) {
862 				pthread_mutex_unlock(&priv->host_blob_format_lock);
863 				return ret;
864 			}
865 
866 			struct drm_virtgpu_resource_info_cros info = { 0 };
867 			info.bo_handle = handle;
868 			info.type = VIRTGPU_RESOURCE_INFO_TYPE_EXTENDED;
869 			int info_ret =
870 			    drmIoctl(drv->fd, DRM_IOCTL_VIRTGPU_RESOURCE_INFO_CROS, &info);
871 
872 			struct drm_gem_close gem_close = { 0 };
873 			gem_close.handle = handle;
874 			int close_ret = drmIoctl(drv->fd, DRM_IOCTL_GEM_CLOSE, &gem_close);
875 			if (close_ret)
876 				drv_loge("DRM_IOCTL_GEM_CLOSE failed (handle=%x) error %d\n",
877 					 handle, close_ret);
878 
879 			if (info_ret) {
880 				pthread_mutex_unlock(&priv->host_blob_format_lock);
881 				drv_loge("Getting resource info failed with %s\n", strerror(errno));
882 				return info_ret;
883 			}
884 
885 			entry = calloc(1, sizeof(*entry));
886 			entry->meta = *meta;
887 
888 			for (int i = 0; i < num_planes; i++) {
889 				entry->meta.strides[i] = info.strides[i];
890 				entry->meta.sizes[i] =
891 				    info.strides[i] *
892 				    drv_height_from_format(meta->format, meta->height, i);
893 				entry->meta.offsets[i] = info.offsets[i];
894 			}
895 			entry->meta.total_size =
896 			    entry->meta.offsets[num_planes - 1] + entry->meta.sizes[num_planes - 1];
897 			entry->meta.format_modifier = info.format_modifier;
898 
899 			lru_insert(&priv->virgl_blob_metadata_cache, &entry->entry);
900 		}
901 
902 		memcpy(meta->offsets, entry->meta.offsets, sizeof(meta->offsets));
903 		memcpy(meta->sizes, entry->meta.sizes, sizeof(meta->sizes));
904 		memcpy(meta->strides, entry->meta.strides, sizeof(meta->strides));
905 		meta->total_size = entry->meta.total_size;
906 		meta->format_modifier = entry->meta.format_modifier;
907 	}
908 	pthread_mutex_unlock(&priv->host_blob_format_lock);
909 
910 	meta->total_size = ALIGN(meta->total_size, PAGE_SIZE);
911 	meta->tiling = blob_flags_from_use_flags(meta->use_flags);
912 
913 	return 0;
914 }
915 
virgl_bo_create_blob(struct driver * drv,struct bo * bo)916 static int virgl_bo_create_blob(struct driver *drv, struct bo *bo)
917 {
918 	int ret;
919 	uint32_t virgl_format = translate_format(bo->meta.format);
920 	uint32_t bo_handle;
921 
922 	virgl_blob_get_host_format(drv, &bo->meta);
923 	ret = virgl_blob_do_create(drv, bo->meta.width, bo->meta.height, bo->meta.use_flags,
924 				   virgl_format, bo->meta.total_size, &bo_handle);
925 	if (ret)
926 		return ret;
927 
928 	bo->handle.u32 = bo_handle;
929 
930 	return 0;
931 }
932 
should_use_blob(struct driver * drv,uint32_t format,uint64_t use_flags)933 static bool should_use_blob(struct driver *drv, uint32_t format, uint64_t use_flags)
934 {
935 	struct virgl_priv *priv = (struct virgl_priv *)drv->priv;
936 
937 	// TODO(gurchetansingh): remove once all minigbm users are blob-safe
938 #ifndef VIRTIO_GPU_NEXT
939 	return false;
940 #endif
941 
942 	// Only use blob when host gbm is available
943 	if (!priv->host_gbm_enabled)
944 		return false;
945 
946 	// Use regular resources if only the GPU needs efficient access. Blob resource is a better
947 	// fit for BO_USE_GPU_DATA_BUFFER which is mapped to VIRGL_BIND_LINEAR.
948 	if (!(use_flags & (BO_USE_SW_READ_OFTEN | BO_USE_SW_WRITE_OFTEN | BO_USE_LINEAR |
949 			   BO_USE_NON_GPU_HW | BO_USE_GPU_DATA_BUFFER)))
950 		return false;
951 
952 	switch (format) {
953 	case DRM_FORMAT_R8:
954 		// Formats with strictly defined strides are supported
955 		return true;
956 	case DRM_FORMAT_ABGR8888:
957 		// Formats used with frequent software reads are supported as long as they
958 		// will not be used by non-GPU hardware.
959 		return (use_flags & BO_USE_SW_READ_OFTEN) && !(use_flags & BO_USE_NON_GPU_HW);
960 	case DRM_FORMAT_YVU420_ANDROID:
961 	case DRM_FORMAT_NV12:
962 		// Zero copy buffers are exposed for guest software access via a persistent
963 		// mapping, with no flush/invalidate messages. However, the virtio-video
964 		// device relies transfers to/from the host waiting on implicit fences in
965 		// the host kernel to synchronize with hardware output. As such, we can only
966 		// use zero copy if the guest doesn't need software access.
967 		return (use_flags & BO_USE_SW_MASK) == 0;
968 	default:
969 		return false;
970 	}
971 }
972 
virgl_bo_create(struct bo * bo,uint32_t width,uint32_t height,uint32_t format,uint64_t use_flags)973 static int virgl_bo_create(struct bo *bo, uint32_t width, uint32_t height, uint32_t format,
974 			   uint64_t use_flags)
975 {
976 	if (params[param_resource_blob].value && params[param_host_visible].value &&
977 	    should_use_blob(bo->drv, format, use_flags))
978 		return virgl_bo_create_blob(bo->drv, bo);
979 
980 	if (params[param_3d].value)
981 		return virgl_3d_bo_create(bo, width, height, format, use_flags);
982 	else
983 		return virgl_2d_dumb_bo_create(bo, width, height, format, use_flags);
984 }
985 
virgl_bo_create_with_modifiers(struct bo * bo,uint32_t width,uint32_t height,uint32_t format,const uint64_t * modifiers,uint32_t count)986 static int virgl_bo_create_with_modifiers(struct bo *bo, uint32_t width, uint32_t height,
987 					  uint32_t format, const uint64_t *modifiers,
988 					  uint32_t count)
989 {
990 	uint64_t use_flags = 0;
991 
992 	for (uint32_t i = 0; i < count; i++) {
993 		if (modifiers[i] == DRM_FORMAT_MOD_LINEAR) {
994 			return virgl_bo_create(bo, width, height, format, use_flags);
995 		}
996 	}
997 
998 	return -EINVAL;
999 }
1000 
virgl_bo_destroy(struct bo * bo)1001 static int virgl_bo_destroy(struct bo *bo)
1002 {
1003 	if (params[param_3d].value)
1004 		return drv_gem_bo_destroy(bo);
1005 	else
1006 		return drv_dumb_bo_destroy(bo);
1007 }
1008 
virgl_bo_map(struct bo * bo,struct vma * vma,uint32_t map_flags)1009 static void *virgl_bo_map(struct bo *bo, struct vma *vma, uint32_t map_flags)
1010 {
1011 	if (params[param_3d].value)
1012 		return virgl_3d_bo_map(bo, vma, map_flags);
1013 	else
1014 		return drv_dumb_bo_map(bo, vma, map_flags);
1015 }
1016 
is_arc_screen_capture_bo(struct bo * bo)1017 static bool is_arc_screen_capture_bo(struct bo *bo)
1018 {
1019 	struct drm_prime_handle prime_handle = {};
1020 	int ret, fd;
1021 	char tmp[256];
1022 
1023 	if (bo->meta.num_planes != 1 ||
1024 	    (bo->meta.format != DRM_FORMAT_ABGR8888 && bo->meta.format != DRM_FORMAT_ARGB8888 &&
1025 	     bo->meta.format != DRM_FORMAT_XRGB8888 && bo->meta.format != DRM_FORMAT_XBGR8888))
1026 		return false;
1027 	prime_handle.handle = bo->handle.u32;
1028 	ret = drmIoctl(bo->drv->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &prime_handle);
1029 	if (ret < 0)
1030 		return false;
1031 	snprintf(tmp, sizeof(tmp), "/proc/self/fdinfo/%d", prime_handle.fd);
1032 	fd = open(tmp, O_RDONLY);
1033 	if (fd < 0) {
1034 		close(prime_handle.fd);
1035 		return false;
1036 	}
1037 	ret = read(fd, tmp, sizeof(tmp) - 1);
1038 	close(prime_handle.fd);
1039 	close(fd);
1040 	if (ret < 0)
1041 		return false;
1042 	tmp[ret] = 0;
1043 
1044 	return strstr(tmp, "ARC-SCREEN-CAP");
1045 }
1046 
virgl_bo_invalidate(struct bo * bo,struct mapping * mapping)1047 static int virgl_bo_invalidate(struct bo *bo, struct mapping *mapping)
1048 {
1049 	int ret;
1050 	size_t i;
1051 	struct drm_virtgpu_3d_transfer_from_host xfer = { 0 };
1052 	struct drm_virtgpu_3d_wait waitcmd = { 0 };
1053 	struct virtio_transfers_params xfer_params;
1054 	struct virgl_priv *priv = (struct virgl_priv *)bo->drv->priv;
1055 	uint64_t host_write_flags;
1056 
1057 	if (!params[param_3d].value)
1058 		return 0;
1059 
1060 	// Invalidate is only necessary if the host writes to the buffer. The encoder and
1061 	// decoder flags don't differentiate between input and output buffers, but we can
1062 	// use the format to determine whether this buffer could be encoder/decoder output.
1063 	host_write_flags = BO_USE_RENDERING | BO_USE_CAMERA_WRITE | BO_USE_GPU_DATA_BUFFER;
1064 	if (bo->meta.format == DRM_FORMAT_R8)
1065 		host_write_flags |= BO_USE_HW_VIDEO_ENCODER;
1066 	else
1067 		host_write_flags |= BO_USE_HW_VIDEO_DECODER;
1068 
1069 	// TODO(b/267892346): Revert this workaround after migrating to virtgpu_cross_domain
1070 	// backend since it's a special arc only behavior.
1071 	if (!(bo->meta.use_flags & (BO_USE_ARC_SCREEN_CAP_PROBED | BO_USE_RENDERING))) {
1072 		bo->meta.use_flags |= BO_USE_ARC_SCREEN_CAP_PROBED;
1073 		if (is_arc_screen_capture_bo(bo)) {
1074 			bo->meta.use_flags |= BO_USE_RENDERING;
1075 		}
1076 	}
1077 
1078 	if ((bo->meta.use_flags & host_write_flags) == 0)
1079 		return 0;
1080 
1081 	if (params[param_resource_blob].value && (bo->meta.tiling & VIRTGPU_BLOB_FLAG_USE_MAPPABLE))
1082 		return 0;
1083 
1084 	xfer.bo_handle = mapping->vma->handle;
1085 
1086 	if (mapping->rect.x || mapping->rect.y) {
1087 		/*
1088 		 * virglrenderer uses the box parameters and assumes that offset == 0 for planar
1089 		 * images
1090 		 */
1091 		if (bo->meta.num_planes == 1) {
1092 			xfer.offset =
1093 			    (bo->meta.strides[0] * mapping->rect.y) +
1094 			    drv_bytes_per_pixel_from_format(bo->meta.format, 0) * mapping->rect.x;
1095 		}
1096 	}
1097 
1098 	if ((bo->meta.use_flags & BO_USE_RENDERING) == 0) {
1099 		// Unfortunately, the kernel doesn't actually pass the guest layer_stride
1100 		// and guest stride to the host (compare virgl.h and virtgpu_drm.h).
1101 		// For gbm based resources, we can work around this by using the level field
1102 		// to pass the stride to virglrenderer's gbm transfer code. However, we need
1103 		// to avoid doing this for resources which don't rely on that transfer code,
1104 		// which is resources with the BO_USE_RENDERING flag set.
1105 		// TODO(b/145993887): Send also stride when the patches are landed
1106 		if (priv->host_gbm_enabled)
1107 			xfer.level = bo->meta.strides[0];
1108 	}
1109 
1110 	if (virgl_supports_combination_natively(bo->drv, bo->meta.format, bo->meta.use_flags)) {
1111 		xfer_params.xfers_needed = 1;
1112 		xfer_params.xfer_boxes[0] = mapping->rect;
1113 	} else {
1114 		assert(virgl_supports_combination_through_emulation(bo->drv, bo->meta.format,
1115 								    bo->meta.use_flags));
1116 
1117 		virgl_get_emulated_transfers_params(bo, &mapping->rect, &xfer_params);
1118 	}
1119 
1120 	for (i = 0; i < xfer_params.xfers_needed; i++) {
1121 		xfer.box.x = xfer_params.xfer_boxes[i].x;
1122 		xfer.box.y = xfer_params.xfer_boxes[i].y;
1123 		xfer.box.w = xfer_params.xfer_boxes[i].width;
1124 		xfer.box.h = xfer_params.xfer_boxes[i].height;
1125 		xfer.box.d = 1;
1126 
1127 		ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_TRANSFER_FROM_HOST, &xfer);
1128 		if (ret) {
1129 			drv_loge("DRM_IOCTL_VIRTGPU_TRANSFER_FROM_HOST failed with %s\n",
1130 				 strerror(errno));
1131 			return -errno;
1132 		}
1133 	}
1134 
1135 	// The transfer needs to complete before invalidate returns so that any host changes
1136 	// are visible and to ensure the host doesn't overwrite subsequent guest changes.
1137 	// TODO(b/136733358): Support returning fences from transfers
1138 	waitcmd.handle = mapping->vma->handle;
1139 	ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_WAIT, &waitcmd);
1140 	if (ret) {
1141 		drv_loge("DRM_IOCTL_VIRTGPU_WAIT failed with %s\n", strerror(errno));
1142 		return -errno;
1143 	}
1144 
1145 	return 0;
1146 }
1147 
virgl_bo_flush(struct bo * bo,struct mapping * mapping)1148 static int virgl_bo_flush(struct bo *bo, struct mapping *mapping)
1149 {
1150 	int ret;
1151 	size_t i;
1152 	struct drm_virtgpu_3d_transfer_to_host xfer = { 0 };
1153 	struct drm_virtgpu_3d_wait waitcmd = { 0 };
1154 	struct virtio_transfers_params xfer_params;
1155 	struct virgl_priv *priv = (struct virgl_priv *)bo->drv->priv;
1156 
1157 	if (!params[param_3d].value)
1158 		return 0;
1159 
1160 	if (!(mapping->vma->map_flags & BO_MAP_WRITE))
1161 		return 0;
1162 
1163 	if (params[param_resource_blob].value && (bo->meta.tiling & VIRTGPU_BLOB_FLAG_USE_MAPPABLE))
1164 		return 0;
1165 
1166 	xfer.bo_handle = mapping->vma->handle;
1167 
1168 	if (mapping->rect.x || mapping->rect.y) {
1169 		/*
1170 		 * virglrenderer uses the box parameters and assumes that offset == 0 for planar
1171 		 * images
1172 		 */
1173 		if (bo->meta.num_planes == 1) {
1174 			xfer.offset =
1175 			    (bo->meta.strides[0] * mapping->rect.y) +
1176 			    drv_bytes_per_pixel_from_format(bo->meta.format, 0) * mapping->rect.x;
1177 		}
1178 	}
1179 
1180 	// Unfortunately, the kernel doesn't actually pass the guest layer_stride and
1181 	// guest stride to the host (compare virgl.h and virtgpu_drm.h). We can use
1182 	// the level to work around this.
1183 	if (priv->host_gbm_enabled)
1184 		xfer.level = bo->meta.strides[0];
1185 
1186 	if (virgl_supports_combination_natively(bo->drv, bo->meta.format, bo->meta.use_flags)) {
1187 		xfer_params.xfers_needed = 1;
1188 		xfer_params.xfer_boxes[0] = mapping->rect;
1189 	} else {
1190 		assert(virgl_supports_combination_through_emulation(bo->drv, bo->meta.format,
1191 								    bo->meta.use_flags));
1192 
1193 		virgl_get_emulated_transfers_params(bo, &mapping->rect, &xfer_params);
1194 	}
1195 
1196 	for (i = 0; i < xfer_params.xfers_needed; i++) {
1197 		xfer.box.x = xfer_params.xfer_boxes[i].x;
1198 		xfer.box.y = xfer_params.xfer_boxes[i].y;
1199 		xfer.box.w = xfer_params.xfer_boxes[i].width;
1200 		xfer.box.h = xfer_params.xfer_boxes[i].height;
1201 		xfer.box.d = 1;
1202 
1203 		ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_TRANSFER_TO_HOST, &xfer);
1204 		if (ret) {
1205 			drv_loge("DRM_IOCTL_VIRTGPU_TRANSFER_TO_HOST failed with %s\n",
1206 				 strerror(errno));
1207 			return -errno;
1208 		}
1209 	}
1210 
1211 	// If the buffer is only accessed by the host GPU, then the flush is ordered
1212 	// with subsequent commands. However, if other host hardware can access the
1213 	// buffer, we need to wait for the transfer to complete for consistency.
1214 	// TODO(b/136733358): Support returning fences from transfers
1215 	if (bo->meta.use_flags & BO_USE_NON_GPU_HW) {
1216 		waitcmd.handle = mapping->vma->handle;
1217 
1218 		ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_WAIT, &waitcmd);
1219 		if (ret) {
1220 			drv_loge("DRM_IOCTL_VIRTGPU_WAIT failed with %s\n", strerror(errno));
1221 			return -errno;
1222 		}
1223 	}
1224 
1225 	return 0;
1226 }
1227 
virgl_3d_resolve_format_and_use_flags(struct driver * drv,uint32_t format,uint64_t use_flags,uint32_t * out_format,uint64_t * out_use_flags)1228 static void virgl_3d_resolve_format_and_use_flags(struct driver *drv, uint32_t format,
1229 						  uint64_t use_flags, uint32_t *out_format,
1230 						  uint64_t *out_use_flags)
1231 {
1232 	*out_format = format;
1233 	*out_use_flags = use_flags;
1234 
1235 	/* resolve flexible format into explicit format */
1236 	switch (format) {
1237 	case DRM_FORMAT_FLEX_IMPLEMENTATION_DEFINED:
1238 		/* Camera subsystem requires NV12. */
1239 		if (use_flags & (BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE)) {
1240 			*out_format = DRM_FORMAT_NV12;
1241 		} else {
1242 			/* HACK: See b/28671744 and b/264408280 */
1243 			*out_format = DRM_FORMAT_XBGR8888;
1244 			*out_use_flags &= ~BO_USE_HW_VIDEO_ENCODER;
1245 			*out_use_flags |= BO_USE_LINEAR;
1246 		}
1247 		break;
1248 	case DRM_FORMAT_FLEX_YCbCr_420_888:
1249 		/* All of our host drivers prefer NV12 as their flexible media format.
1250 		 * If that changes, this will need to be modified. */
1251 		*out_format = DRM_FORMAT_NV12;
1252 		break;
1253 	default:
1254 		break;
1255 	}
1256 
1257 	/* resolve explicit format */
1258 	switch (*out_format) {
1259 	case DRM_FORMAT_NV12:
1260 	case DRM_FORMAT_ABGR8888:
1261 	case DRM_FORMAT_ARGB8888:
1262 	case DRM_FORMAT_RGB565:
1263 	case DRM_FORMAT_XBGR8888:
1264 	case DRM_FORMAT_XRGB8888:
1265 		/* These are the scanout capable formats to the guest. Strip scanout use_flag if the
1266 		 * host does not natively support scanout on the requested format. */
1267 		if ((*out_use_flags & BO_USE_SCANOUT) &&
1268 		    !virgl_supports_combination_natively(drv, *out_format, BO_USE_SCANOUT))
1269 			*out_use_flags &= ~BO_USE_SCANOUT;
1270 		break;
1271 	case DRM_FORMAT_YVU420_ANDROID:
1272 		*out_use_flags &= ~BO_USE_SCANOUT;
1273 		/* HACK: See b/172389166. Also see gbm_bo_create. */
1274 		*out_use_flags |= BO_USE_LINEAR;
1275 		break;
1276 	default:
1277 		break;
1278 	}
1279 }
1280 
virgl_2d_resolve_format_and_use_flags(uint32_t format,uint64_t use_flags,uint32_t * out_format,uint64_t * out_use_flags)1281 static void virgl_2d_resolve_format_and_use_flags(uint32_t format, uint64_t use_flags,
1282 						  uint32_t *out_format, uint64_t *out_use_flags)
1283 {
1284 	*out_format = format;
1285 	*out_use_flags = use_flags;
1286 
1287 	/* HACK: See crrev/c/1849773 */
1288 	if (format != DRM_FORMAT_XRGB8888)
1289 		*out_use_flags &= ~BO_USE_SCANOUT;
1290 
1291 	switch (format) {
1292 	case DRM_FORMAT_FLEX_IMPLEMENTATION_DEFINED:
1293 		/* Camera subsystem requires NV12. */
1294 		if (use_flags & (BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE)) {
1295 			*out_format = DRM_FORMAT_NV12;
1296 		} else {
1297 			/* HACK: See b/28671744 */
1298 			*out_format = DRM_FORMAT_XBGR8888;
1299 			*out_use_flags &= ~BO_USE_HW_VIDEO_ENCODER;
1300 		}
1301 		break;
1302 	case DRM_FORMAT_FLEX_YCbCr_420_888:
1303 		*out_format = DRM_FORMAT_YVU420_ANDROID;
1304 		/* fallthrough */
1305 	case DRM_FORMAT_YVU420_ANDROID:
1306 		*out_use_flags &= ~BO_USE_SCANOUT;
1307 		/* HACK: See b/172389166. Also see gbm_bo_create. */
1308 		*out_use_flags |= BO_USE_LINEAR;
1309 		break;
1310 	default:
1311 		break;
1312 	}
1313 }
1314 
virgl_resolve_format_and_use_flags(struct driver * drv,uint32_t format,uint64_t use_flags,uint32_t * out_format,uint64_t * out_use_flags)1315 static void virgl_resolve_format_and_use_flags(struct driver *drv, uint32_t format,
1316 					       uint64_t use_flags, uint32_t *out_format,
1317 					       uint64_t *out_use_flags)
1318 {
1319 	if (params[param_3d].value) {
1320 		return virgl_3d_resolve_format_and_use_flags(drv, format, use_flags, out_format,
1321 							     out_use_flags);
1322 	} else {
1323 		return virgl_2d_resolve_format_and_use_flags(format, use_flags, out_format,
1324 							     out_use_flags);
1325 	}
1326 }
1327 
virgl_resource_info(struct bo * bo,uint32_t strides[DRV_MAX_PLANES],uint32_t offsets[DRV_MAX_PLANES],uint64_t * format_modifier)1328 static int virgl_resource_info(struct bo *bo, uint32_t strides[DRV_MAX_PLANES],
1329 			       uint32_t offsets[DRV_MAX_PLANES], uint64_t *format_modifier)
1330 {
1331 	int ret;
1332 	struct drm_virtgpu_resource_info_cros res_info = { 0 };
1333 
1334 	if (!params[param_3d].value)
1335 		return 0;
1336 
1337 	res_info.bo_handle = bo->handle.u32;
1338 	res_info.type = VIRTGPU_RESOURCE_INFO_TYPE_EXTENDED;
1339 	ret = drmIoctl(bo->drv->fd, DRM_IOCTL_VIRTGPU_RESOURCE_INFO_CROS, &res_info);
1340 	if (ret) {
1341 		drv_loge("DRM_IOCTL_VIRTGPU_RESOURCE_INFO failed with %s\n", strerror(errno));
1342 		return ret;
1343 	}
1344 
1345 	for (uint32_t plane = 0; plane < DRV_MAX_PLANES; plane++) {
1346 		/*
1347 		 * Currently, kernel v4.14 (Betty) doesn't have the extended resource info
1348 		 * ioctl.
1349 		 */
1350 		if (!res_info.strides[plane])
1351 			break;
1352 
1353 		strides[plane] = res_info.strides[plane];
1354 		offsets[plane] = res_info.offsets[plane];
1355 	}
1356 	*format_modifier = res_info.format_modifier;
1357 
1358 	return 0;
1359 }
1360 
virgl_get_max_texture_2d_size(struct driver * drv)1361 static uint32_t virgl_get_max_texture_2d_size(struct driver *drv)
1362 {
1363 	if (params[param_3d].value)
1364 		return virgl_3d_get_max_texture_2d_size(drv);
1365 	else
1366 		return VIRGL_2D_MAX_TEXTURE_2D_SIZE;
1367 }
1368 
1369 const struct backend virtgpu_virgl = { .name = "virtgpu_virgl",
1370 				       .init = virgl_init,
1371 				       .close = virgl_close,
1372 				       .bo_create = virgl_bo_create,
1373 				       .bo_create_with_modifiers = virgl_bo_create_with_modifiers,
1374 				       .bo_destroy = virgl_bo_destroy,
1375 				       .bo_import = drv_prime_bo_import,
1376 				       .bo_map = virgl_bo_map,
1377 				       .bo_unmap = drv_bo_munmap,
1378 				       .bo_invalidate = virgl_bo_invalidate,
1379 				       .bo_flush = virgl_bo_flush,
1380 				       .resolve_format_and_use_flags =
1381 					   virgl_resolve_format_and_use_flags,
1382 				       .resource_info = virgl_resource_info,
1383 				       .get_max_texture_2d_size = virgl_get_max_texture_2d_size };
1384