xref: /linux/drivers/gpu/drm/vc4/vc4_validate.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
1 /*
2  * Copyright © 2014 Broadcom
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  */
23 
24 /**
25  * DOC: Command list validator for VC4.
26  *
27  * Since the VC4 has no IOMMU between it and system memory, a user
28  * with access to execute command lists could escalate privilege by
29  * overwriting system memory (drawing to it as a framebuffer) or
30  * reading system memory it shouldn't (reading it as a vertex buffer
31  * or index buffer)
32  *
33  * We validate binner command lists to ensure that all accesses are
34  * within the bounds of the GEM objects referenced by the submitted
35  * job.  It explicitly whitelists packets, and looks at the offsets in
36  * any address fields to make sure they're contained within the BOs
37  * they reference.
38  *
39  * Note that because CL validation is already reading the
40  * user-submitted CL and writing the validated copy out to the memory
41  * that the GPU will actually read, this is also where GEM relocation
42  * processing (turning BO references into actual addresses for the GPU
43  * to use) happens.
44  */
45 
46 #include "uapi/drm/vc4_drm.h"
47 #include "vc4_drv.h"
48 #include "vc4_packet.h"
49 
50 #define VALIDATE_ARGS \
51 	struct vc4_exec_info *exec,			\
52 	void *validated,				\
53 	void *untrusted
54 
55 /** Return the width in pixels of a 64-byte microtile. */
56 static uint32_t
57 utile_width(int cpp)
58 {
59 	switch (cpp) {
60 	case 1:
61 	case 2:
62 		return 8;
63 	case 4:
64 		return 4;
65 	case 8:
66 		return 2;
67 	default:
68 		pr_err("unknown cpp: %d\n", cpp);
69 		return 1;
70 	}
71 }
72 
73 /** Return the height in pixels of a 64-byte microtile. */
74 static uint32_t
75 utile_height(int cpp)
76 {
77 	switch (cpp) {
78 	case 1:
79 		return 8;
80 	case 2:
81 	case 4:
82 	case 8:
83 		return 4;
84 	default:
85 		pr_err("unknown cpp: %d\n", cpp);
86 		return 1;
87 	}
88 }
89 
90 /**
91  * size_is_lt() - Returns whether a miplevel of the given size will
92  * use the lineartile (LT) tiling layout rather than the normal T
93  * tiling layout.
94  * @width: Width in pixels of the miplevel
95  * @height: Height in pixels of the miplevel
96  * @cpp: Bytes per pixel of the pixel format
97  */
98 static bool
99 size_is_lt(uint32_t width, uint32_t height, int cpp)
100 {
101 	return (width <= 4 * utile_width(cpp) ||
102 		height <= 4 * utile_height(cpp));
103 }
104 
105 struct drm_gem_dma_object *
106 vc4_use_bo(struct vc4_exec_info *exec, uint32_t hindex)
107 {
108 	struct vc4_dev *vc4 = exec->dev;
109 	struct drm_gem_dma_object *obj;
110 	struct vc4_bo *bo;
111 
112 	if (WARN_ON_ONCE(vc4->gen > VC4_GEN_4))
113 		return NULL;
114 
115 	if (hindex >= exec->bo_count) {
116 		DRM_DEBUG("BO index %d greater than BO count %d\n",
117 			  hindex, exec->bo_count);
118 		return NULL;
119 	}
120 	obj = to_drm_gem_dma_obj(exec->bo[hindex]);
121 	bo = to_vc4_bo(&obj->base);
122 
123 	if (bo->validated_shader) {
124 		DRM_DEBUG("Trying to use shader BO as something other than "
125 			  "a shader\n");
126 		return NULL;
127 	}
128 
129 	return obj;
130 }
131 
132 static struct drm_gem_dma_object *
133 vc4_use_handle(struct vc4_exec_info *exec, uint32_t gem_handles_packet_index)
134 {
135 	return vc4_use_bo(exec, exec->bo_index[gem_handles_packet_index]);
136 }
137 
138 static bool
139 validate_bin_pos(struct vc4_exec_info *exec, void *untrusted, uint32_t pos)
140 {
141 	/* Note that the untrusted pointer passed to these functions is
142 	 * incremented past the packet byte.
143 	 */
144 	return (untrusted - 1 == exec->bin_u + pos);
145 }
146 
147 static uint32_t
148 gl_shader_rec_size(uint32_t pointer_bits)
149 {
150 	uint32_t attribute_count = pointer_bits & 7;
151 	bool extended = pointer_bits & 8;
152 
153 	if (attribute_count == 0)
154 		attribute_count = 8;
155 
156 	if (extended)
157 		return 100 + attribute_count * 4;
158 	else
159 		return 36 + attribute_count * 8;
160 }
161 
162 bool
163 vc4_check_tex_size(struct vc4_exec_info *exec, struct drm_gem_dma_object *fbo,
164 		   uint32_t offset, uint8_t tiling_format,
165 		   uint32_t width, uint32_t height, uint8_t cpp)
166 {
167 	struct vc4_dev *vc4 = exec->dev;
168 	uint32_t aligned_width, aligned_height, stride, size;
169 	uint32_t utile_w = utile_width(cpp);
170 	uint32_t utile_h = utile_height(cpp);
171 
172 	if (WARN_ON_ONCE(vc4->gen > VC4_GEN_4))
173 		return false;
174 
175 	/* The shaded vertex format stores signed 12.4 fixed point
176 	 * (-2048,2047) offsets from the viewport center, so we should
177 	 * never have a render target larger than 4096.  The texture
178 	 * unit can only sample from 2048x2048, so it's even more
179 	 * restricted.  This lets us avoid worrying about overflow in
180 	 * our math.
181 	 */
182 	if (width > 4096 || height > 4096) {
183 		DRM_DEBUG("Surface dimensions (%d,%d) too large",
184 			  width, height);
185 		return false;
186 	}
187 
188 	switch (tiling_format) {
189 	case VC4_TILING_FORMAT_LINEAR:
190 		aligned_width = round_up(width, utile_w);
191 		aligned_height = height;
192 		break;
193 	case VC4_TILING_FORMAT_T:
194 		aligned_width = round_up(width, utile_w * 8);
195 		aligned_height = round_up(height, utile_h * 8);
196 		break;
197 	case VC4_TILING_FORMAT_LT:
198 		aligned_width = round_up(width, utile_w);
199 		aligned_height = round_up(height, utile_h);
200 		break;
201 	default:
202 		DRM_DEBUG("buffer tiling %d unsupported\n", tiling_format);
203 		return false;
204 	}
205 
206 	stride = aligned_width * cpp;
207 	size = stride * aligned_height;
208 
209 	if (size + offset < size ||
210 	    size + offset > fbo->base.size) {
211 		DRM_DEBUG("Overflow in %dx%d (%dx%d) fbo size (%d + %d > %zd)\n",
212 			  width, height,
213 			  aligned_width, aligned_height,
214 			  size, offset, fbo->base.size);
215 		return false;
216 	}
217 
218 	return true;
219 }
220 
221 static int
222 validate_flush(VALIDATE_ARGS)
223 {
224 	if (!validate_bin_pos(exec, untrusted, exec->args->bin_cl_size - 1)) {
225 		DRM_DEBUG("Bin CL must end with VC4_PACKET_FLUSH\n");
226 		return -EINVAL;
227 	}
228 	exec->found_flush = true;
229 
230 	return 0;
231 }
232 
233 static int
234 validate_start_tile_binning(VALIDATE_ARGS)
235 {
236 	if (exec->found_start_tile_binning_packet) {
237 		DRM_DEBUG("Duplicate VC4_PACKET_START_TILE_BINNING\n");
238 		return -EINVAL;
239 	}
240 	exec->found_start_tile_binning_packet = true;
241 
242 	if (!exec->found_tile_binning_mode_config_packet) {
243 		DRM_DEBUG("missing VC4_PACKET_TILE_BINNING_MODE_CONFIG\n");
244 		return -EINVAL;
245 	}
246 
247 	return 0;
248 }
249 
250 static int
251 validate_increment_semaphore(VALIDATE_ARGS)
252 {
253 	if (!validate_bin_pos(exec, untrusted, exec->args->bin_cl_size - 2)) {
254 		DRM_DEBUG("Bin CL must end with "
255 			  "VC4_PACKET_INCREMENT_SEMAPHORE\n");
256 		return -EINVAL;
257 	}
258 	exec->found_increment_semaphore_packet = true;
259 
260 	return 0;
261 }
262 
263 static int
264 validate_indexed_prim_list(VALIDATE_ARGS)
265 {
266 	struct drm_gem_dma_object *ib;
267 	uint32_t length = *(uint32_t *)(untrusted + 1);
268 	uint32_t offset = *(uint32_t *)(untrusted + 5);
269 	uint32_t max_index = *(uint32_t *)(untrusted + 9);
270 	uint32_t index_size = (*(uint8_t *)(untrusted + 0) >> 4) ? 2 : 1;
271 	struct vc4_shader_state *shader_state;
272 
273 	/* Check overflow condition */
274 	if (exec->shader_state_count == 0) {
275 		DRM_DEBUG("shader state must precede primitives\n");
276 		return -EINVAL;
277 	}
278 	shader_state = &exec->shader_state[exec->shader_state_count - 1];
279 
280 	if (max_index > shader_state->max_index)
281 		shader_state->max_index = max_index;
282 
283 	ib = vc4_use_handle(exec, 0);
284 	if (!ib)
285 		return -EINVAL;
286 
287 	exec->bin_dep_seqno = max(exec->bin_dep_seqno,
288 				  to_vc4_bo(&ib->base)->write_seqno);
289 
290 	if (offset > ib->base.size ||
291 	    (ib->base.size - offset) / index_size < length) {
292 		DRM_DEBUG("IB access overflow (%d + %d*%d > %zd)\n",
293 			  offset, length, index_size, ib->base.size);
294 		return -EINVAL;
295 	}
296 
297 	*(uint32_t *)(validated + 5) = ib->dma_addr + offset;
298 
299 	return 0;
300 }
301 
302 static int
303 validate_gl_array_primitive(VALIDATE_ARGS)
304 {
305 	uint32_t length = *(uint32_t *)(untrusted + 1);
306 	uint32_t base_index = *(uint32_t *)(untrusted + 5);
307 	uint32_t max_index;
308 	struct vc4_shader_state *shader_state;
309 
310 	/* Check overflow condition */
311 	if (exec->shader_state_count == 0) {
312 		DRM_DEBUG("shader state must precede primitives\n");
313 		return -EINVAL;
314 	}
315 	shader_state = &exec->shader_state[exec->shader_state_count - 1];
316 
317 	if (length + base_index < length) {
318 		DRM_DEBUG("primitive vertex count overflow\n");
319 		return -EINVAL;
320 	}
321 	max_index = length + base_index - 1;
322 
323 	if (max_index > shader_state->max_index)
324 		shader_state->max_index = max_index;
325 
326 	return 0;
327 }
328 
329 static int
330 validate_gl_shader_state(VALIDATE_ARGS)
331 {
332 	uint32_t i = exec->shader_state_count++;
333 
334 	if (i >= exec->shader_state_size) {
335 		DRM_DEBUG("More requests for shader states than declared\n");
336 		return -EINVAL;
337 	}
338 
339 	exec->shader_state[i].addr = *(uint32_t *)untrusted;
340 	exec->shader_state[i].max_index = 0;
341 
342 	if (exec->shader_state[i].addr & ~0xf) {
343 		DRM_DEBUG("high bits set in GL shader rec reference\n");
344 		return -EINVAL;
345 	}
346 
347 	*(uint32_t *)validated = (exec->shader_rec_p +
348 				  exec->shader_state[i].addr);
349 
350 	exec->shader_rec_p +=
351 		roundup(gl_shader_rec_size(exec->shader_state[i].addr), 16);
352 
353 	return 0;
354 }
355 
356 static int
357 validate_tile_binning_config(VALIDATE_ARGS)
358 {
359 	struct drm_device *dev = exec->exec_bo->base.dev;
360 	struct vc4_dev *vc4 = to_vc4_dev(dev);
361 	uint8_t flags;
362 	uint32_t tile_state_size;
363 	uint32_t tile_count, bin_addr;
364 	int bin_slot;
365 
366 	if (exec->found_tile_binning_mode_config_packet) {
367 		DRM_DEBUG("Duplicate VC4_PACKET_TILE_BINNING_MODE_CONFIG\n");
368 		return -EINVAL;
369 	}
370 	exec->found_tile_binning_mode_config_packet = true;
371 
372 	exec->bin_tiles_x = *(uint8_t *)(untrusted + 12);
373 	exec->bin_tiles_y = *(uint8_t *)(untrusted + 13);
374 	tile_count = exec->bin_tiles_x * exec->bin_tiles_y;
375 	flags = *(uint8_t *)(untrusted + 14);
376 
377 	if (exec->bin_tiles_x == 0 ||
378 	    exec->bin_tiles_y == 0) {
379 		DRM_DEBUG("Tile binning config of %dx%d too small\n",
380 			  exec->bin_tiles_x, exec->bin_tiles_y);
381 		return -EINVAL;
382 	}
383 
384 	if (flags & (VC4_BIN_CONFIG_DB_NON_MS |
385 		     VC4_BIN_CONFIG_TILE_BUFFER_64BIT)) {
386 		DRM_DEBUG("unsupported binning config flags 0x%02x\n", flags);
387 		return -EINVAL;
388 	}
389 
390 	bin_slot = vc4_v3d_get_bin_slot(vc4);
391 	if (bin_slot < 0) {
392 		if (bin_slot != -EINTR && bin_slot != -ERESTARTSYS) {
393 			drm_err(dev, "Failed to allocate binner memory: %d\n",
394 				bin_slot);
395 		}
396 		return bin_slot;
397 	}
398 
399 	/* The slot we allocated will only be used by this job, and is
400 	 * free when the job completes rendering.
401 	 */
402 	exec->bin_slots |= BIT(bin_slot);
403 	bin_addr = vc4->bin_bo->base.dma_addr + bin_slot * vc4->bin_alloc_size;
404 
405 	/* The tile state data array is 48 bytes per tile, and we put it at
406 	 * the start of a BO containing both it and the tile alloc.
407 	 */
408 	tile_state_size = 48 * tile_count;
409 
410 	/* Since the tile alloc array will follow us, align. */
411 	exec->tile_alloc_offset = bin_addr + roundup(tile_state_size, 4096);
412 
413 	*(uint8_t *)(validated + 14) =
414 		((flags & ~(VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_MASK |
415 			    VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_MASK)) |
416 		 VC4_BIN_CONFIG_AUTO_INIT_TSDA |
417 		 VC4_SET_FIELD(VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_32,
418 			       VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE) |
419 		 VC4_SET_FIELD(VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_128,
420 			       VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE));
421 
422 	/* tile alloc address. */
423 	*(uint32_t *)(validated + 0) = exec->tile_alloc_offset;
424 	/* tile alloc size. */
425 	*(uint32_t *)(validated + 4) = (bin_addr + vc4->bin_alloc_size -
426 					exec->tile_alloc_offset);
427 	/* tile state address. */
428 	*(uint32_t *)(validated + 8) = bin_addr;
429 
430 	return 0;
431 }
432 
433 static int
434 validate_gem_handles(VALIDATE_ARGS)
435 {
436 	memcpy(exec->bo_index, untrusted, sizeof(exec->bo_index));
437 	return 0;
438 }
439 
440 #define VC4_DEFINE_PACKET(packet, func) \
441 	[packet] = { packet ## _SIZE, #packet, func }
442 
443 static const struct cmd_info {
444 	uint16_t len;
445 	const char *name;
446 	int (*func)(struct vc4_exec_info *exec, void *validated,
447 		    void *untrusted);
448 } cmd_info[] = {
449 	VC4_DEFINE_PACKET(VC4_PACKET_HALT, NULL),
450 	VC4_DEFINE_PACKET(VC4_PACKET_NOP, NULL),
451 	VC4_DEFINE_PACKET(VC4_PACKET_FLUSH, validate_flush),
452 	VC4_DEFINE_PACKET(VC4_PACKET_FLUSH_ALL, NULL),
453 	VC4_DEFINE_PACKET(VC4_PACKET_START_TILE_BINNING,
454 			  validate_start_tile_binning),
455 	VC4_DEFINE_PACKET(VC4_PACKET_INCREMENT_SEMAPHORE,
456 			  validate_increment_semaphore),
457 
458 	VC4_DEFINE_PACKET(VC4_PACKET_GL_INDEXED_PRIMITIVE,
459 			  validate_indexed_prim_list),
460 	VC4_DEFINE_PACKET(VC4_PACKET_GL_ARRAY_PRIMITIVE,
461 			  validate_gl_array_primitive),
462 
463 	VC4_DEFINE_PACKET(VC4_PACKET_PRIMITIVE_LIST_FORMAT, NULL),
464 
465 	VC4_DEFINE_PACKET(VC4_PACKET_GL_SHADER_STATE, validate_gl_shader_state),
466 
467 	VC4_DEFINE_PACKET(VC4_PACKET_CONFIGURATION_BITS, NULL),
468 	VC4_DEFINE_PACKET(VC4_PACKET_FLAT_SHADE_FLAGS, NULL),
469 	VC4_DEFINE_PACKET(VC4_PACKET_POINT_SIZE, NULL),
470 	VC4_DEFINE_PACKET(VC4_PACKET_LINE_WIDTH, NULL),
471 	VC4_DEFINE_PACKET(VC4_PACKET_RHT_X_BOUNDARY, NULL),
472 	VC4_DEFINE_PACKET(VC4_PACKET_DEPTH_OFFSET, NULL),
473 	VC4_DEFINE_PACKET(VC4_PACKET_CLIP_WINDOW, NULL),
474 	VC4_DEFINE_PACKET(VC4_PACKET_VIEWPORT_OFFSET, NULL),
475 	VC4_DEFINE_PACKET(VC4_PACKET_CLIPPER_XY_SCALING, NULL),
476 	/* Note: The docs say this was also 105, but it was 106 in the
477 	 * initial userland code drop.
478 	 */
479 	VC4_DEFINE_PACKET(VC4_PACKET_CLIPPER_Z_SCALING, NULL),
480 
481 	VC4_DEFINE_PACKET(VC4_PACKET_TILE_BINNING_MODE_CONFIG,
482 			  validate_tile_binning_config),
483 
484 	VC4_DEFINE_PACKET(VC4_PACKET_GEM_HANDLES, validate_gem_handles),
485 };
486 
487 int
488 vc4_validate_bin_cl(struct drm_device *dev,
489 		    void *validated,
490 		    void *unvalidated,
491 		    struct vc4_exec_info *exec)
492 {
493 	struct vc4_dev *vc4 = to_vc4_dev(dev);
494 	uint32_t len = exec->args->bin_cl_size;
495 	uint32_t dst_offset = 0;
496 	uint32_t src_offset = 0;
497 
498 	if (WARN_ON_ONCE(vc4->gen > VC4_GEN_4))
499 		return -ENODEV;
500 
501 	while (src_offset < len) {
502 		void *dst_pkt = validated + dst_offset;
503 		void *src_pkt = unvalidated + src_offset;
504 		u8 cmd = *(uint8_t *)src_pkt;
505 		const struct cmd_info *info;
506 
507 		if (cmd >= ARRAY_SIZE(cmd_info)) {
508 			DRM_DEBUG("0x%08x: packet %d out of bounds\n",
509 				  src_offset, cmd);
510 			return -EINVAL;
511 		}
512 
513 		info = &cmd_info[cmd];
514 		if (!info->name) {
515 			DRM_DEBUG("0x%08x: packet %d invalid\n",
516 				  src_offset, cmd);
517 			return -EINVAL;
518 		}
519 
520 		if (src_offset + info->len > len) {
521 			DRM_DEBUG("0x%08x: packet %d (%s) length 0x%08x "
522 				  "exceeds bounds (0x%08x)\n",
523 				  src_offset, cmd, info->name, info->len,
524 				  src_offset + len);
525 			return -EINVAL;
526 		}
527 
528 		if (cmd != VC4_PACKET_GEM_HANDLES)
529 			memcpy(dst_pkt, src_pkt, info->len);
530 
531 		if (info->func && info->func(exec,
532 					     dst_pkt + 1,
533 					     src_pkt + 1)) {
534 			DRM_DEBUG("0x%08x: packet %d (%s) failed to validate\n",
535 				  src_offset, cmd, info->name);
536 			return -EINVAL;
537 		}
538 
539 		src_offset += info->len;
540 		/* GEM handle loading doesn't produce HW packets. */
541 		if (cmd != VC4_PACKET_GEM_HANDLES)
542 			dst_offset += info->len;
543 
544 		/* When the CL hits halt, it'll stop reading anything else. */
545 		if (cmd == VC4_PACKET_HALT)
546 			break;
547 	}
548 
549 	exec->ct0ea = exec->ct0ca + dst_offset;
550 
551 	if (!exec->found_start_tile_binning_packet) {
552 		DRM_DEBUG("Bin CL missing VC4_PACKET_START_TILE_BINNING\n");
553 		return -EINVAL;
554 	}
555 
556 	/* The bin CL must be ended with INCREMENT_SEMAPHORE and FLUSH.  The
557 	 * semaphore is used to trigger the render CL to start up, and the
558 	 * FLUSH is what caps the bin lists with
559 	 * VC4_PACKET_RETURN_FROM_SUB_LIST (so they jump back to the main
560 	 * render CL when they get called to) and actually triggers the queued
561 	 * semaphore increment.
562 	 */
563 	if (!exec->found_increment_semaphore_packet || !exec->found_flush) {
564 		DRM_DEBUG("Bin CL missing VC4_PACKET_INCREMENT_SEMAPHORE + "
565 			  "VC4_PACKET_FLUSH\n");
566 		return -EINVAL;
567 	}
568 
569 	return 0;
570 }
571 
572 static bool
573 reloc_tex(struct vc4_exec_info *exec,
574 	  void *uniform_data_u,
575 	  struct vc4_texture_sample_info *sample,
576 	  uint32_t texture_handle_index, bool is_cs)
577 {
578 	struct drm_gem_dma_object *tex;
579 	uint32_t p0 = *(uint32_t *)(uniform_data_u + sample->p_offset[0]);
580 	uint32_t p1 = *(uint32_t *)(uniform_data_u + sample->p_offset[1]);
581 	uint32_t p2 = (sample->p_offset[2] != ~0 ?
582 		       *(uint32_t *)(uniform_data_u + sample->p_offset[2]) : 0);
583 	uint32_t p3 = (sample->p_offset[3] != ~0 ?
584 		       *(uint32_t *)(uniform_data_u + sample->p_offset[3]) : 0);
585 	uint32_t *validated_p0 = exec->uniforms_v + sample->p_offset[0];
586 	uint32_t offset = p0 & VC4_TEX_P0_OFFSET_MASK;
587 	uint32_t miplevels = VC4_GET_FIELD(p0, VC4_TEX_P0_MIPLVLS);
588 	uint32_t width = VC4_GET_FIELD(p1, VC4_TEX_P1_WIDTH);
589 	uint32_t height = VC4_GET_FIELD(p1, VC4_TEX_P1_HEIGHT);
590 	uint32_t cpp, tiling_format, utile_w, utile_h;
591 	uint32_t i;
592 	uint32_t cube_map_stride = 0;
593 	enum vc4_texture_data_type type;
594 
595 	tex = vc4_use_bo(exec, texture_handle_index);
596 	if (!tex)
597 		return false;
598 
599 	if (sample->is_direct) {
600 		uint32_t remaining_size = tex->base.size - p0;
601 
602 		if (p0 > tex->base.size - 4) {
603 			DRM_DEBUG("UBO offset greater than UBO size\n");
604 			goto fail;
605 		}
606 		if (p1 > remaining_size - 4) {
607 			DRM_DEBUG("UBO clamp would allow reads "
608 				  "outside of UBO\n");
609 			goto fail;
610 		}
611 		*validated_p0 = tex->dma_addr + p0;
612 		return true;
613 	}
614 
615 	if (width == 0)
616 		width = 2048;
617 	if (height == 0)
618 		height = 2048;
619 
620 	if (p0 & VC4_TEX_P0_CMMODE_MASK) {
621 		if (VC4_GET_FIELD(p2, VC4_TEX_P2_PTYPE) ==
622 		    VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE)
623 			cube_map_stride = p2 & VC4_TEX_P2_CMST_MASK;
624 		if (VC4_GET_FIELD(p3, VC4_TEX_P2_PTYPE) ==
625 		    VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE) {
626 			if (cube_map_stride) {
627 				DRM_DEBUG("Cube map stride set twice\n");
628 				goto fail;
629 			}
630 
631 			cube_map_stride = p3 & VC4_TEX_P2_CMST_MASK;
632 		}
633 		if (!cube_map_stride) {
634 			DRM_DEBUG("Cube map stride not set\n");
635 			goto fail;
636 		}
637 	}
638 
639 	type = (VC4_GET_FIELD(p0, VC4_TEX_P0_TYPE) |
640 		(VC4_GET_FIELD(p1, VC4_TEX_P1_TYPE4) << 4));
641 
642 	switch (type) {
643 	case VC4_TEXTURE_TYPE_RGBA8888:
644 	case VC4_TEXTURE_TYPE_RGBX8888:
645 	case VC4_TEXTURE_TYPE_RGBA32R:
646 		cpp = 4;
647 		break;
648 	case VC4_TEXTURE_TYPE_RGBA4444:
649 	case VC4_TEXTURE_TYPE_RGBA5551:
650 	case VC4_TEXTURE_TYPE_RGB565:
651 	case VC4_TEXTURE_TYPE_LUMALPHA:
652 	case VC4_TEXTURE_TYPE_S16F:
653 	case VC4_TEXTURE_TYPE_S16:
654 		cpp = 2;
655 		break;
656 	case VC4_TEXTURE_TYPE_LUMINANCE:
657 	case VC4_TEXTURE_TYPE_ALPHA:
658 	case VC4_TEXTURE_TYPE_S8:
659 		cpp = 1;
660 		break;
661 	case VC4_TEXTURE_TYPE_ETC1:
662 		/* ETC1 is arranged as 64-bit blocks, where each block is 4x4
663 		 * pixels.
664 		 */
665 		cpp = 8;
666 		width = (width + 3) >> 2;
667 		height = (height + 3) >> 2;
668 		break;
669 	case VC4_TEXTURE_TYPE_BW1:
670 	case VC4_TEXTURE_TYPE_A4:
671 	case VC4_TEXTURE_TYPE_A1:
672 	case VC4_TEXTURE_TYPE_RGBA64:
673 	case VC4_TEXTURE_TYPE_YUV422R:
674 	default:
675 		DRM_DEBUG("Texture format %d unsupported\n", type);
676 		goto fail;
677 	}
678 	utile_w = utile_width(cpp);
679 	utile_h = utile_height(cpp);
680 
681 	if (type == VC4_TEXTURE_TYPE_RGBA32R) {
682 		tiling_format = VC4_TILING_FORMAT_LINEAR;
683 	} else {
684 		if (size_is_lt(width, height, cpp))
685 			tiling_format = VC4_TILING_FORMAT_LT;
686 		else
687 			tiling_format = VC4_TILING_FORMAT_T;
688 	}
689 
690 	if (!vc4_check_tex_size(exec, tex, offset + cube_map_stride * 5,
691 				tiling_format, width, height, cpp)) {
692 		goto fail;
693 	}
694 
695 	/* The mipmap levels are stored before the base of the texture.  Make
696 	 * sure there is actually space in the BO.
697 	 */
698 	for (i = 1; i <= miplevels; i++) {
699 		uint32_t level_width = max(width >> i, 1u);
700 		uint32_t level_height = max(height >> i, 1u);
701 		uint32_t aligned_width, aligned_height;
702 		uint32_t level_size;
703 
704 		/* Once the levels get small enough, they drop from T to LT. */
705 		if (tiling_format == VC4_TILING_FORMAT_T &&
706 		    size_is_lt(level_width, level_height, cpp)) {
707 			tiling_format = VC4_TILING_FORMAT_LT;
708 		}
709 
710 		switch (tiling_format) {
711 		case VC4_TILING_FORMAT_T:
712 			aligned_width = round_up(level_width, utile_w * 8);
713 			aligned_height = round_up(level_height, utile_h * 8);
714 			break;
715 		case VC4_TILING_FORMAT_LT:
716 			aligned_width = round_up(level_width, utile_w);
717 			aligned_height = round_up(level_height, utile_h);
718 			break;
719 		default:
720 			aligned_width = round_up(level_width, utile_w);
721 			aligned_height = level_height;
722 			break;
723 		}
724 
725 		level_size = aligned_width * cpp * aligned_height;
726 
727 		if (offset < level_size) {
728 			DRM_DEBUG("Level %d (%dx%d -> %dx%d) size %db "
729 				  "overflowed buffer bounds (offset %d)\n",
730 				  i, level_width, level_height,
731 				  aligned_width, aligned_height,
732 				  level_size, offset);
733 			goto fail;
734 		}
735 
736 		offset -= level_size;
737 	}
738 
739 	*validated_p0 = tex->dma_addr + p0;
740 
741 	if (is_cs) {
742 		exec->bin_dep_seqno = max(exec->bin_dep_seqno,
743 					  to_vc4_bo(&tex->base)->write_seqno);
744 	}
745 
746 	return true;
747  fail:
748 	DRM_INFO("Texture p0 at %d: 0x%08x\n", sample->p_offset[0], p0);
749 	DRM_INFO("Texture p1 at %d: 0x%08x\n", sample->p_offset[1], p1);
750 	DRM_INFO("Texture p2 at %d: 0x%08x\n", sample->p_offset[2], p2);
751 	DRM_INFO("Texture p3 at %d: 0x%08x\n", sample->p_offset[3], p3);
752 	return false;
753 }
754 
755 static int
756 validate_gl_shader_rec(struct drm_device *dev,
757 		       struct vc4_exec_info *exec,
758 		       struct vc4_shader_state *state)
759 {
760 	uint32_t *src_handles;
761 	void *pkt_u, *pkt_v;
762 	static const uint32_t shader_reloc_offsets[] = {
763 		4, /* fs */
764 		16, /* vs */
765 		28, /* cs */
766 	};
767 	uint32_t shader_reloc_count = ARRAY_SIZE(shader_reloc_offsets);
768 	struct drm_gem_dma_object *bo[ARRAY_SIZE(shader_reloc_offsets) + 8];
769 	uint32_t nr_attributes, nr_relocs, packet_size;
770 	int i;
771 
772 	nr_attributes = state->addr & 0x7;
773 	if (nr_attributes == 0)
774 		nr_attributes = 8;
775 	packet_size = gl_shader_rec_size(state->addr);
776 
777 	nr_relocs = ARRAY_SIZE(shader_reloc_offsets) + nr_attributes;
778 	if (nr_relocs * 4 > exec->shader_rec_size) {
779 		DRM_DEBUG("overflowed shader recs reading %d handles "
780 			  "from %d bytes left\n",
781 			  nr_relocs, exec->shader_rec_size);
782 		return -EINVAL;
783 	}
784 	src_handles = exec->shader_rec_u;
785 	exec->shader_rec_u += nr_relocs * 4;
786 	exec->shader_rec_size -= nr_relocs * 4;
787 
788 	if (packet_size > exec->shader_rec_size) {
789 		DRM_DEBUG("overflowed shader recs copying %db packet "
790 			  "from %d bytes left\n",
791 			  packet_size, exec->shader_rec_size);
792 		return -EINVAL;
793 	}
794 	pkt_u = exec->shader_rec_u;
795 	pkt_v = exec->shader_rec_v;
796 	memcpy(pkt_v, pkt_u, packet_size);
797 	exec->shader_rec_u += packet_size;
798 	/* Shader recs have to be aligned to 16 bytes (due to the attribute
799 	 * flags being in the low bytes), so round the next validated shader
800 	 * rec address up.  This should be safe, since we've got so many
801 	 * relocations in a shader rec packet.
802 	 */
803 	BUG_ON(roundup(packet_size, 16) - packet_size > nr_relocs * 4);
804 	exec->shader_rec_v += roundup(packet_size, 16);
805 	exec->shader_rec_size -= packet_size;
806 
807 	for (i = 0; i < shader_reloc_count; i++) {
808 		if (src_handles[i] > exec->bo_count) {
809 			DRM_DEBUG("Shader handle %d too big\n", src_handles[i]);
810 			return -EINVAL;
811 		}
812 
813 		bo[i] = to_drm_gem_dma_obj(exec->bo[src_handles[i]]);
814 		if (!bo[i])
815 			return -EINVAL;
816 	}
817 	for (i = shader_reloc_count; i < nr_relocs; i++) {
818 		bo[i] = vc4_use_bo(exec, src_handles[i]);
819 		if (!bo[i])
820 			return -EINVAL;
821 	}
822 
823 	if (((*(uint16_t *)pkt_u & VC4_SHADER_FLAG_FS_SINGLE_THREAD) == 0) !=
824 	    to_vc4_bo(&bo[0]->base)->validated_shader->is_threaded) {
825 		DRM_DEBUG("Thread mode of CL and FS do not match\n");
826 		return -EINVAL;
827 	}
828 
829 	if (to_vc4_bo(&bo[1]->base)->validated_shader->is_threaded ||
830 	    to_vc4_bo(&bo[2]->base)->validated_shader->is_threaded) {
831 		DRM_DEBUG("cs and vs cannot be threaded\n");
832 		return -EINVAL;
833 	}
834 
835 	for (i = 0; i < shader_reloc_count; i++) {
836 		struct vc4_validated_shader_info *validated_shader;
837 		uint32_t o = shader_reloc_offsets[i];
838 		uint32_t src_offset = *(uint32_t *)(pkt_u + o);
839 		uint32_t *texture_handles_u;
840 		void *uniform_data_u;
841 		uint32_t tex, uni;
842 
843 		*(uint32_t *)(pkt_v + o) = bo[i]->dma_addr + src_offset;
844 
845 		if (src_offset != 0) {
846 			DRM_DEBUG("Shaders must be at offset 0 of "
847 				  "the BO.\n");
848 			return -EINVAL;
849 		}
850 
851 		validated_shader = to_vc4_bo(&bo[i]->base)->validated_shader;
852 		if (!validated_shader)
853 			return -EINVAL;
854 
855 		if (validated_shader->uniforms_src_size >
856 		    exec->uniforms_size) {
857 			DRM_DEBUG("Uniforms src buffer overflow\n");
858 			return -EINVAL;
859 		}
860 
861 		texture_handles_u = exec->uniforms_u;
862 		uniform_data_u = (texture_handles_u +
863 				  validated_shader->num_texture_samples);
864 
865 		memcpy(exec->uniforms_v, uniform_data_u,
866 		       validated_shader->uniforms_size);
867 
868 		for (tex = 0;
869 		     tex < validated_shader->num_texture_samples;
870 		     tex++) {
871 			if (!reloc_tex(exec,
872 				       uniform_data_u,
873 				       &validated_shader->texture_samples[tex],
874 				       texture_handles_u[tex],
875 				       i == 2)) {
876 				return -EINVAL;
877 			}
878 		}
879 
880 		/* Fill in the uniform slots that need this shader's
881 		 * start-of-uniforms address (used for resetting the uniform
882 		 * stream in the presence of control flow).
883 		 */
884 		for (uni = 0;
885 		     uni < validated_shader->num_uniform_addr_offsets;
886 		     uni++) {
887 			uint32_t o = validated_shader->uniform_addr_offsets[uni];
888 			((uint32_t *)exec->uniforms_v)[o] = exec->uniforms_p;
889 		}
890 
891 		*(uint32_t *)(pkt_v + o + 4) = exec->uniforms_p;
892 
893 		exec->uniforms_u += validated_shader->uniforms_src_size;
894 		exec->uniforms_v += validated_shader->uniforms_size;
895 		exec->uniforms_p += validated_shader->uniforms_size;
896 	}
897 
898 	for (i = 0; i < nr_attributes; i++) {
899 		struct drm_gem_dma_object *vbo =
900 			bo[ARRAY_SIZE(shader_reloc_offsets) + i];
901 		uint32_t o = 36 + i * 8;
902 		uint32_t offset = *(uint32_t *)(pkt_u + o + 0);
903 		uint32_t attr_size = *(uint8_t *)(pkt_u + o + 4) + 1;
904 		uint32_t stride = *(uint8_t *)(pkt_u + o + 5);
905 		uint32_t max_index;
906 
907 		exec->bin_dep_seqno = max(exec->bin_dep_seqno,
908 					  to_vc4_bo(&vbo->base)->write_seqno);
909 
910 		if (state->addr & 0x8)
911 			stride |= (*(uint32_t *)(pkt_u + 100 + i * 4)) & ~0xff;
912 
913 		if (vbo->base.size < offset ||
914 		    vbo->base.size - offset < attr_size) {
915 			DRM_DEBUG("BO offset overflow (%d + %d > %zu)\n",
916 				  offset, attr_size, vbo->base.size);
917 			return -EINVAL;
918 		}
919 
920 		if (stride != 0) {
921 			max_index = ((vbo->base.size - offset - attr_size) /
922 				     stride);
923 			if (state->max_index > max_index) {
924 				DRM_DEBUG("primitives use index %d out of "
925 					  "supplied %d\n",
926 					  state->max_index, max_index);
927 				return -EINVAL;
928 			}
929 		}
930 
931 		*(uint32_t *)(pkt_v + o) = vbo->dma_addr + offset;
932 	}
933 
934 	return 0;
935 }
936 
937 int
938 vc4_validate_shader_recs(struct drm_device *dev,
939 			 struct vc4_exec_info *exec)
940 {
941 	struct vc4_dev *vc4 = to_vc4_dev(dev);
942 	uint32_t i;
943 	int ret = 0;
944 
945 	if (WARN_ON_ONCE(vc4->gen > VC4_GEN_4))
946 		return -ENODEV;
947 
948 	for (i = 0; i < exec->shader_state_count; i++) {
949 		ret = validate_gl_shader_rec(dev, exec, &exec->shader_state[i]);
950 		if (ret)
951 			return ret;
952 	}
953 
954 	return ret;
955 }
956