xref: /openbmc/linux/drivers/gpu/drm/vc4/vc4_validate.c (revision e5f586c763a079349398e2b0c7c271386193ac34)
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 		DRM_ERROR("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 		DRM_ERROR("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_cma_object *
106 vc4_use_bo(struct vc4_exec_info *exec, uint32_t hindex)
107 {
108 	struct drm_gem_cma_object *obj;
109 	struct vc4_bo *bo;
110 
111 	if (hindex >= exec->bo_count) {
112 		DRM_ERROR("BO index %d greater than BO count %d\n",
113 			  hindex, exec->bo_count);
114 		return NULL;
115 	}
116 	obj = exec->bo[hindex];
117 	bo = to_vc4_bo(&obj->base);
118 
119 	if (bo->validated_shader) {
120 		DRM_ERROR("Trying to use shader BO as something other than "
121 			  "a shader\n");
122 		return NULL;
123 	}
124 
125 	return obj;
126 }
127 
128 static struct drm_gem_cma_object *
129 vc4_use_handle(struct vc4_exec_info *exec, uint32_t gem_handles_packet_index)
130 {
131 	return vc4_use_bo(exec, exec->bo_index[gem_handles_packet_index]);
132 }
133 
134 static bool
135 validate_bin_pos(struct vc4_exec_info *exec, void *untrusted, uint32_t pos)
136 {
137 	/* Note that the untrusted pointer passed to these functions is
138 	 * incremented past the packet byte.
139 	 */
140 	return (untrusted - 1 == exec->bin_u + pos);
141 }
142 
143 static uint32_t
144 gl_shader_rec_size(uint32_t pointer_bits)
145 {
146 	uint32_t attribute_count = pointer_bits & 7;
147 	bool extended = pointer_bits & 8;
148 
149 	if (attribute_count == 0)
150 		attribute_count = 8;
151 
152 	if (extended)
153 		return 100 + attribute_count * 4;
154 	else
155 		return 36 + attribute_count * 8;
156 }
157 
158 bool
159 vc4_check_tex_size(struct vc4_exec_info *exec, struct drm_gem_cma_object *fbo,
160 		   uint32_t offset, uint8_t tiling_format,
161 		   uint32_t width, uint32_t height, uint8_t cpp)
162 {
163 	uint32_t aligned_width, aligned_height, stride, size;
164 	uint32_t utile_w = utile_width(cpp);
165 	uint32_t utile_h = utile_height(cpp);
166 
167 	/* The shaded vertex format stores signed 12.4 fixed point
168 	 * (-2048,2047) offsets from the viewport center, so we should
169 	 * never have a render target larger than 4096.  The texture
170 	 * unit can only sample from 2048x2048, so it's even more
171 	 * restricted.  This lets us avoid worrying about overflow in
172 	 * our math.
173 	 */
174 	if (width > 4096 || height > 4096) {
175 		DRM_ERROR("Surface dimesions (%d,%d) too large", width, height);
176 		return false;
177 	}
178 
179 	switch (tiling_format) {
180 	case VC4_TILING_FORMAT_LINEAR:
181 		aligned_width = round_up(width, utile_w);
182 		aligned_height = height;
183 		break;
184 	case VC4_TILING_FORMAT_T:
185 		aligned_width = round_up(width, utile_w * 8);
186 		aligned_height = round_up(height, utile_h * 8);
187 		break;
188 	case VC4_TILING_FORMAT_LT:
189 		aligned_width = round_up(width, utile_w);
190 		aligned_height = round_up(height, utile_h);
191 		break;
192 	default:
193 		DRM_ERROR("buffer tiling %d unsupported\n", tiling_format);
194 		return false;
195 	}
196 
197 	stride = aligned_width * cpp;
198 	size = stride * aligned_height;
199 
200 	if (size + offset < size ||
201 	    size + offset > fbo->base.size) {
202 		DRM_ERROR("Overflow in %dx%d (%dx%d) fbo size (%d + %d > %zd)\n",
203 			  width, height,
204 			  aligned_width, aligned_height,
205 			  size, offset, fbo->base.size);
206 		return false;
207 	}
208 
209 	return true;
210 }
211 
212 static int
213 validate_flush(VALIDATE_ARGS)
214 {
215 	if (!validate_bin_pos(exec, untrusted, exec->args->bin_cl_size - 1)) {
216 		DRM_ERROR("Bin CL must end with VC4_PACKET_FLUSH\n");
217 		return -EINVAL;
218 	}
219 	exec->found_flush = true;
220 
221 	return 0;
222 }
223 
224 static int
225 validate_start_tile_binning(VALIDATE_ARGS)
226 {
227 	if (exec->found_start_tile_binning_packet) {
228 		DRM_ERROR("Duplicate VC4_PACKET_START_TILE_BINNING\n");
229 		return -EINVAL;
230 	}
231 	exec->found_start_tile_binning_packet = true;
232 
233 	if (!exec->found_tile_binning_mode_config_packet) {
234 		DRM_ERROR("missing VC4_PACKET_TILE_BINNING_MODE_CONFIG\n");
235 		return -EINVAL;
236 	}
237 
238 	return 0;
239 }
240 
241 static int
242 validate_increment_semaphore(VALIDATE_ARGS)
243 {
244 	if (!validate_bin_pos(exec, untrusted, exec->args->bin_cl_size - 2)) {
245 		DRM_ERROR("Bin CL must end with "
246 			  "VC4_PACKET_INCREMENT_SEMAPHORE\n");
247 		return -EINVAL;
248 	}
249 	exec->found_increment_semaphore_packet = true;
250 
251 	return 0;
252 }
253 
254 static int
255 validate_indexed_prim_list(VALIDATE_ARGS)
256 {
257 	struct drm_gem_cma_object *ib;
258 	uint32_t length = *(uint32_t *)(untrusted + 1);
259 	uint32_t offset = *(uint32_t *)(untrusted + 5);
260 	uint32_t max_index = *(uint32_t *)(untrusted + 9);
261 	uint32_t index_size = (*(uint8_t *)(untrusted + 0) >> 4) ? 2 : 1;
262 	struct vc4_shader_state *shader_state;
263 
264 	/* Check overflow condition */
265 	if (exec->shader_state_count == 0) {
266 		DRM_ERROR("shader state must precede primitives\n");
267 		return -EINVAL;
268 	}
269 	shader_state = &exec->shader_state[exec->shader_state_count - 1];
270 
271 	if (max_index > shader_state->max_index)
272 		shader_state->max_index = max_index;
273 
274 	ib = vc4_use_handle(exec, 0);
275 	if (!ib)
276 		return -EINVAL;
277 
278 	exec->bin_dep_seqno = max(exec->bin_dep_seqno,
279 				  to_vc4_bo(&ib->base)->write_seqno);
280 
281 	if (offset > ib->base.size ||
282 	    (ib->base.size - offset) / index_size < length) {
283 		DRM_ERROR("IB access overflow (%d + %d*%d > %zd)\n",
284 			  offset, length, index_size, ib->base.size);
285 		return -EINVAL;
286 	}
287 
288 	*(uint32_t *)(validated + 5) = ib->paddr + offset;
289 
290 	return 0;
291 }
292 
293 static int
294 validate_gl_array_primitive(VALIDATE_ARGS)
295 {
296 	uint32_t length = *(uint32_t *)(untrusted + 1);
297 	uint32_t base_index = *(uint32_t *)(untrusted + 5);
298 	uint32_t max_index;
299 	struct vc4_shader_state *shader_state;
300 
301 	/* Check overflow condition */
302 	if (exec->shader_state_count == 0) {
303 		DRM_ERROR("shader state must precede primitives\n");
304 		return -EINVAL;
305 	}
306 	shader_state = &exec->shader_state[exec->shader_state_count - 1];
307 
308 	if (length + base_index < length) {
309 		DRM_ERROR("primitive vertex count overflow\n");
310 		return -EINVAL;
311 	}
312 	max_index = length + base_index - 1;
313 
314 	if (max_index > shader_state->max_index)
315 		shader_state->max_index = max_index;
316 
317 	return 0;
318 }
319 
320 static int
321 validate_gl_shader_state(VALIDATE_ARGS)
322 {
323 	uint32_t i = exec->shader_state_count++;
324 
325 	if (i >= exec->shader_state_size) {
326 		DRM_ERROR("More requests for shader states than declared\n");
327 		return -EINVAL;
328 	}
329 
330 	exec->shader_state[i].addr = *(uint32_t *)untrusted;
331 	exec->shader_state[i].max_index = 0;
332 
333 	if (exec->shader_state[i].addr & ~0xf) {
334 		DRM_ERROR("high bits set in GL shader rec reference\n");
335 		return -EINVAL;
336 	}
337 
338 	*(uint32_t *)validated = (exec->shader_rec_p +
339 				  exec->shader_state[i].addr);
340 
341 	exec->shader_rec_p +=
342 		roundup(gl_shader_rec_size(exec->shader_state[i].addr), 16);
343 
344 	return 0;
345 }
346 
347 static int
348 validate_tile_binning_config(VALIDATE_ARGS)
349 {
350 	struct drm_device *dev = exec->exec_bo->base.dev;
351 	struct vc4_bo *tile_bo;
352 	uint8_t flags;
353 	uint32_t tile_state_size, tile_alloc_size;
354 	uint32_t tile_count;
355 
356 	if (exec->found_tile_binning_mode_config_packet) {
357 		DRM_ERROR("Duplicate VC4_PACKET_TILE_BINNING_MODE_CONFIG\n");
358 		return -EINVAL;
359 	}
360 	exec->found_tile_binning_mode_config_packet = true;
361 
362 	exec->bin_tiles_x = *(uint8_t *)(untrusted + 12);
363 	exec->bin_tiles_y = *(uint8_t *)(untrusted + 13);
364 	tile_count = exec->bin_tiles_x * exec->bin_tiles_y;
365 	flags = *(uint8_t *)(untrusted + 14);
366 
367 	if (exec->bin_tiles_x == 0 ||
368 	    exec->bin_tiles_y == 0) {
369 		DRM_ERROR("Tile binning config of %dx%d too small\n",
370 			  exec->bin_tiles_x, exec->bin_tiles_y);
371 		return -EINVAL;
372 	}
373 
374 	if (flags & (VC4_BIN_CONFIG_DB_NON_MS |
375 		     VC4_BIN_CONFIG_TILE_BUFFER_64BIT)) {
376 		DRM_ERROR("unsupported binning config flags 0x%02x\n", flags);
377 		return -EINVAL;
378 	}
379 
380 	/* The tile state data array is 48 bytes per tile, and we put it at
381 	 * the start of a BO containing both it and the tile alloc.
382 	 */
383 	tile_state_size = 48 * tile_count;
384 
385 	/* Since the tile alloc array will follow us, align. */
386 	exec->tile_alloc_offset = roundup(tile_state_size, 4096);
387 
388 	*(uint8_t *)(validated + 14) =
389 		((flags & ~(VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_MASK |
390 			    VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_MASK)) |
391 		 VC4_BIN_CONFIG_AUTO_INIT_TSDA |
392 		 VC4_SET_FIELD(VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_32,
393 			       VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE) |
394 		 VC4_SET_FIELD(VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_128,
395 			       VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE));
396 
397 	/* Initial block size. */
398 	tile_alloc_size = 32 * tile_count;
399 
400 	/*
401 	 * The initial allocation gets rounded to the next 256 bytes before
402 	 * the hardware starts fulfilling further allocations.
403 	 */
404 	tile_alloc_size = roundup(tile_alloc_size, 256);
405 
406 	/* Add space for the extra allocations.  This is what gets used first,
407 	 * before overflow memory.  It must have at least 4096 bytes, but we
408 	 * want to avoid overflow memory usage if possible.
409 	 */
410 	tile_alloc_size += 1024 * 1024;
411 
412 	tile_bo = vc4_bo_create(dev, exec->tile_alloc_offset + tile_alloc_size,
413 				true);
414 	exec->tile_bo = &tile_bo->base;
415 	if (IS_ERR(exec->tile_bo))
416 		return PTR_ERR(exec->tile_bo);
417 	list_add_tail(&tile_bo->unref_head, &exec->unref_list);
418 
419 	/* tile alloc address. */
420 	*(uint32_t *)(validated + 0) = (exec->tile_bo->paddr +
421 					exec->tile_alloc_offset);
422 	/* tile alloc size. */
423 	*(uint32_t *)(validated + 4) = tile_alloc_size;
424 	/* tile state address. */
425 	*(uint32_t *)(validated + 8) = exec->tile_bo->paddr;
426 
427 	return 0;
428 }
429 
430 static int
431 validate_gem_handles(VALIDATE_ARGS)
432 {
433 	memcpy(exec->bo_index, untrusted, sizeof(exec->bo_index));
434 	return 0;
435 }
436 
437 #define VC4_DEFINE_PACKET(packet, func) \
438 	[packet] = { packet ## _SIZE, #packet, func }
439 
440 static const struct cmd_info {
441 	uint16_t len;
442 	const char *name;
443 	int (*func)(struct vc4_exec_info *exec, void *validated,
444 		    void *untrusted);
445 } cmd_info[] = {
446 	VC4_DEFINE_PACKET(VC4_PACKET_HALT, NULL),
447 	VC4_DEFINE_PACKET(VC4_PACKET_NOP, NULL),
448 	VC4_DEFINE_PACKET(VC4_PACKET_FLUSH, validate_flush),
449 	VC4_DEFINE_PACKET(VC4_PACKET_FLUSH_ALL, NULL),
450 	VC4_DEFINE_PACKET(VC4_PACKET_START_TILE_BINNING,
451 			  validate_start_tile_binning),
452 	VC4_DEFINE_PACKET(VC4_PACKET_INCREMENT_SEMAPHORE,
453 			  validate_increment_semaphore),
454 
455 	VC4_DEFINE_PACKET(VC4_PACKET_GL_INDEXED_PRIMITIVE,
456 			  validate_indexed_prim_list),
457 	VC4_DEFINE_PACKET(VC4_PACKET_GL_ARRAY_PRIMITIVE,
458 			  validate_gl_array_primitive),
459 
460 	VC4_DEFINE_PACKET(VC4_PACKET_PRIMITIVE_LIST_FORMAT, NULL),
461 
462 	VC4_DEFINE_PACKET(VC4_PACKET_GL_SHADER_STATE, validate_gl_shader_state),
463 
464 	VC4_DEFINE_PACKET(VC4_PACKET_CONFIGURATION_BITS, NULL),
465 	VC4_DEFINE_PACKET(VC4_PACKET_FLAT_SHADE_FLAGS, NULL),
466 	VC4_DEFINE_PACKET(VC4_PACKET_POINT_SIZE, NULL),
467 	VC4_DEFINE_PACKET(VC4_PACKET_LINE_WIDTH, NULL),
468 	VC4_DEFINE_PACKET(VC4_PACKET_RHT_X_BOUNDARY, NULL),
469 	VC4_DEFINE_PACKET(VC4_PACKET_DEPTH_OFFSET, NULL),
470 	VC4_DEFINE_PACKET(VC4_PACKET_CLIP_WINDOW, NULL),
471 	VC4_DEFINE_PACKET(VC4_PACKET_VIEWPORT_OFFSET, NULL),
472 	VC4_DEFINE_PACKET(VC4_PACKET_CLIPPER_XY_SCALING, NULL),
473 	/* Note: The docs say this was also 105, but it was 106 in the
474 	 * initial userland code drop.
475 	 */
476 	VC4_DEFINE_PACKET(VC4_PACKET_CLIPPER_Z_SCALING, NULL),
477 
478 	VC4_DEFINE_PACKET(VC4_PACKET_TILE_BINNING_MODE_CONFIG,
479 			  validate_tile_binning_config),
480 
481 	VC4_DEFINE_PACKET(VC4_PACKET_GEM_HANDLES, validate_gem_handles),
482 };
483 
484 int
485 vc4_validate_bin_cl(struct drm_device *dev,
486 		    void *validated,
487 		    void *unvalidated,
488 		    struct vc4_exec_info *exec)
489 {
490 	uint32_t len = exec->args->bin_cl_size;
491 	uint32_t dst_offset = 0;
492 	uint32_t src_offset = 0;
493 
494 	while (src_offset < len) {
495 		void *dst_pkt = validated + dst_offset;
496 		void *src_pkt = unvalidated + src_offset;
497 		u8 cmd = *(uint8_t *)src_pkt;
498 		const struct cmd_info *info;
499 
500 		if (cmd >= ARRAY_SIZE(cmd_info)) {
501 			DRM_ERROR("0x%08x: packet %d out of bounds\n",
502 				  src_offset, cmd);
503 			return -EINVAL;
504 		}
505 
506 		info = &cmd_info[cmd];
507 		if (!info->name) {
508 			DRM_ERROR("0x%08x: packet %d invalid\n",
509 				  src_offset, cmd);
510 			return -EINVAL;
511 		}
512 
513 		if (src_offset + info->len > len) {
514 			DRM_ERROR("0x%08x: packet %d (%s) length 0x%08x "
515 				  "exceeds bounds (0x%08x)\n",
516 				  src_offset, cmd, info->name, info->len,
517 				  src_offset + len);
518 			return -EINVAL;
519 		}
520 
521 		if (cmd != VC4_PACKET_GEM_HANDLES)
522 			memcpy(dst_pkt, src_pkt, info->len);
523 
524 		if (info->func && info->func(exec,
525 					     dst_pkt + 1,
526 					     src_pkt + 1)) {
527 			DRM_ERROR("0x%08x: packet %d (%s) failed to validate\n",
528 				  src_offset, cmd, info->name);
529 			return -EINVAL;
530 		}
531 
532 		src_offset += info->len;
533 		/* GEM handle loading doesn't produce HW packets. */
534 		if (cmd != VC4_PACKET_GEM_HANDLES)
535 			dst_offset += info->len;
536 
537 		/* When the CL hits halt, it'll stop reading anything else. */
538 		if (cmd == VC4_PACKET_HALT)
539 			break;
540 	}
541 
542 	exec->ct0ea = exec->ct0ca + dst_offset;
543 
544 	if (!exec->found_start_tile_binning_packet) {
545 		DRM_ERROR("Bin CL missing VC4_PACKET_START_TILE_BINNING\n");
546 		return -EINVAL;
547 	}
548 
549 	/* The bin CL must be ended with INCREMENT_SEMAPHORE and FLUSH.  The
550 	 * semaphore is used to trigger the render CL to start up, and the
551 	 * FLUSH is what caps the bin lists with
552 	 * VC4_PACKET_RETURN_FROM_SUB_LIST (so they jump back to the main
553 	 * render CL when they get called to) and actually triggers the queued
554 	 * semaphore increment.
555 	 */
556 	if (!exec->found_increment_semaphore_packet || !exec->found_flush) {
557 		DRM_ERROR("Bin CL missing VC4_PACKET_INCREMENT_SEMAPHORE + "
558 			  "VC4_PACKET_FLUSH\n");
559 		return -EINVAL;
560 	}
561 
562 	return 0;
563 }
564 
565 static bool
566 reloc_tex(struct vc4_exec_info *exec,
567 	  void *uniform_data_u,
568 	  struct vc4_texture_sample_info *sample,
569 	  uint32_t texture_handle_index, bool is_cs)
570 {
571 	struct drm_gem_cma_object *tex;
572 	uint32_t p0 = *(uint32_t *)(uniform_data_u + sample->p_offset[0]);
573 	uint32_t p1 = *(uint32_t *)(uniform_data_u + sample->p_offset[1]);
574 	uint32_t p2 = (sample->p_offset[2] != ~0 ?
575 		       *(uint32_t *)(uniform_data_u + sample->p_offset[2]) : 0);
576 	uint32_t p3 = (sample->p_offset[3] != ~0 ?
577 		       *(uint32_t *)(uniform_data_u + sample->p_offset[3]) : 0);
578 	uint32_t *validated_p0 = exec->uniforms_v + sample->p_offset[0];
579 	uint32_t offset = p0 & VC4_TEX_P0_OFFSET_MASK;
580 	uint32_t miplevels = VC4_GET_FIELD(p0, VC4_TEX_P0_MIPLVLS);
581 	uint32_t width = VC4_GET_FIELD(p1, VC4_TEX_P1_WIDTH);
582 	uint32_t height = VC4_GET_FIELD(p1, VC4_TEX_P1_HEIGHT);
583 	uint32_t cpp, tiling_format, utile_w, utile_h;
584 	uint32_t i;
585 	uint32_t cube_map_stride = 0;
586 	enum vc4_texture_data_type type;
587 
588 	tex = vc4_use_bo(exec, texture_handle_index);
589 	if (!tex)
590 		return false;
591 
592 	if (sample->is_direct) {
593 		uint32_t remaining_size = tex->base.size - p0;
594 
595 		if (p0 > tex->base.size - 4) {
596 			DRM_ERROR("UBO offset greater than UBO size\n");
597 			goto fail;
598 		}
599 		if (p1 > remaining_size - 4) {
600 			DRM_ERROR("UBO clamp would allow reads "
601 				  "outside of UBO\n");
602 			goto fail;
603 		}
604 		*validated_p0 = tex->paddr + p0;
605 		return true;
606 	}
607 
608 	if (width == 0)
609 		width = 2048;
610 	if (height == 0)
611 		height = 2048;
612 
613 	if (p0 & VC4_TEX_P0_CMMODE_MASK) {
614 		if (VC4_GET_FIELD(p2, VC4_TEX_P2_PTYPE) ==
615 		    VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE)
616 			cube_map_stride = p2 & VC4_TEX_P2_CMST_MASK;
617 		if (VC4_GET_FIELD(p3, VC4_TEX_P2_PTYPE) ==
618 		    VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE) {
619 			if (cube_map_stride) {
620 				DRM_ERROR("Cube map stride set twice\n");
621 				goto fail;
622 			}
623 
624 			cube_map_stride = p3 & VC4_TEX_P2_CMST_MASK;
625 		}
626 		if (!cube_map_stride) {
627 			DRM_ERROR("Cube map stride not set\n");
628 			goto fail;
629 		}
630 	}
631 
632 	type = (VC4_GET_FIELD(p0, VC4_TEX_P0_TYPE) |
633 		(VC4_GET_FIELD(p1, VC4_TEX_P1_TYPE4) << 4));
634 
635 	switch (type) {
636 	case VC4_TEXTURE_TYPE_RGBA8888:
637 	case VC4_TEXTURE_TYPE_RGBX8888:
638 	case VC4_TEXTURE_TYPE_RGBA32R:
639 		cpp = 4;
640 		break;
641 	case VC4_TEXTURE_TYPE_RGBA4444:
642 	case VC4_TEXTURE_TYPE_RGBA5551:
643 	case VC4_TEXTURE_TYPE_RGB565:
644 	case VC4_TEXTURE_TYPE_LUMALPHA:
645 	case VC4_TEXTURE_TYPE_S16F:
646 	case VC4_TEXTURE_TYPE_S16:
647 		cpp = 2;
648 		break;
649 	case VC4_TEXTURE_TYPE_LUMINANCE:
650 	case VC4_TEXTURE_TYPE_ALPHA:
651 	case VC4_TEXTURE_TYPE_S8:
652 		cpp = 1;
653 		break;
654 	case VC4_TEXTURE_TYPE_ETC1:
655 		/* ETC1 is arranged as 64-bit blocks, where each block is 4x4
656 		 * pixels.
657 		 */
658 		cpp = 8;
659 		width = (width + 3) >> 2;
660 		height = (height + 3) >> 2;
661 		break;
662 	case VC4_TEXTURE_TYPE_BW1:
663 	case VC4_TEXTURE_TYPE_A4:
664 	case VC4_TEXTURE_TYPE_A1:
665 	case VC4_TEXTURE_TYPE_RGBA64:
666 	case VC4_TEXTURE_TYPE_YUV422R:
667 	default:
668 		DRM_ERROR("Texture format %d unsupported\n", type);
669 		goto fail;
670 	}
671 	utile_w = utile_width(cpp);
672 	utile_h = utile_height(cpp);
673 
674 	if (type == VC4_TEXTURE_TYPE_RGBA32R) {
675 		tiling_format = VC4_TILING_FORMAT_LINEAR;
676 	} else {
677 		if (size_is_lt(width, height, cpp))
678 			tiling_format = VC4_TILING_FORMAT_LT;
679 		else
680 			tiling_format = VC4_TILING_FORMAT_T;
681 	}
682 
683 	if (!vc4_check_tex_size(exec, tex, offset + cube_map_stride * 5,
684 				tiling_format, width, height, cpp)) {
685 		goto fail;
686 	}
687 
688 	/* The mipmap levels are stored before the base of the texture.  Make
689 	 * sure there is actually space in the BO.
690 	 */
691 	for (i = 1; i <= miplevels; i++) {
692 		uint32_t level_width = max(width >> i, 1u);
693 		uint32_t level_height = max(height >> i, 1u);
694 		uint32_t aligned_width, aligned_height;
695 		uint32_t level_size;
696 
697 		/* Once the levels get small enough, they drop from T to LT. */
698 		if (tiling_format == VC4_TILING_FORMAT_T &&
699 		    size_is_lt(level_width, level_height, cpp)) {
700 			tiling_format = VC4_TILING_FORMAT_LT;
701 		}
702 
703 		switch (tiling_format) {
704 		case VC4_TILING_FORMAT_T:
705 			aligned_width = round_up(level_width, utile_w * 8);
706 			aligned_height = round_up(level_height, utile_h * 8);
707 			break;
708 		case VC4_TILING_FORMAT_LT:
709 			aligned_width = round_up(level_width, utile_w);
710 			aligned_height = round_up(level_height, utile_h);
711 			break;
712 		default:
713 			aligned_width = round_up(level_width, utile_w);
714 			aligned_height = level_height;
715 			break;
716 		}
717 
718 		level_size = aligned_width * cpp * aligned_height;
719 
720 		if (offset < level_size) {
721 			DRM_ERROR("Level %d (%dx%d -> %dx%d) size %db "
722 				  "overflowed buffer bounds (offset %d)\n",
723 				  i, level_width, level_height,
724 				  aligned_width, aligned_height,
725 				  level_size, offset);
726 			goto fail;
727 		}
728 
729 		offset -= level_size;
730 	}
731 
732 	*validated_p0 = tex->paddr + p0;
733 
734 	if (is_cs) {
735 		exec->bin_dep_seqno = max(exec->bin_dep_seqno,
736 					  to_vc4_bo(&tex->base)->write_seqno);
737 	}
738 
739 	return true;
740  fail:
741 	DRM_INFO("Texture p0 at %d: 0x%08x\n", sample->p_offset[0], p0);
742 	DRM_INFO("Texture p1 at %d: 0x%08x\n", sample->p_offset[1], p1);
743 	DRM_INFO("Texture p2 at %d: 0x%08x\n", sample->p_offset[2], p2);
744 	DRM_INFO("Texture p3 at %d: 0x%08x\n", sample->p_offset[3], p3);
745 	return false;
746 }
747 
748 static int
749 validate_gl_shader_rec(struct drm_device *dev,
750 		       struct vc4_exec_info *exec,
751 		       struct vc4_shader_state *state)
752 {
753 	uint32_t *src_handles;
754 	void *pkt_u, *pkt_v;
755 	static const uint32_t shader_reloc_offsets[] = {
756 		4, /* fs */
757 		16, /* vs */
758 		28, /* cs */
759 	};
760 	uint32_t shader_reloc_count = ARRAY_SIZE(shader_reloc_offsets);
761 	struct drm_gem_cma_object *bo[shader_reloc_count + 8];
762 	uint32_t nr_attributes, nr_relocs, packet_size;
763 	int i;
764 
765 	nr_attributes = state->addr & 0x7;
766 	if (nr_attributes == 0)
767 		nr_attributes = 8;
768 	packet_size = gl_shader_rec_size(state->addr);
769 
770 	nr_relocs = ARRAY_SIZE(shader_reloc_offsets) + nr_attributes;
771 	if (nr_relocs * 4 > exec->shader_rec_size) {
772 		DRM_ERROR("overflowed shader recs reading %d handles "
773 			  "from %d bytes left\n",
774 			  nr_relocs, exec->shader_rec_size);
775 		return -EINVAL;
776 	}
777 	src_handles = exec->shader_rec_u;
778 	exec->shader_rec_u += nr_relocs * 4;
779 	exec->shader_rec_size -= nr_relocs * 4;
780 
781 	if (packet_size > exec->shader_rec_size) {
782 		DRM_ERROR("overflowed shader recs copying %db packet "
783 			  "from %d bytes left\n",
784 			  packet_size, exec->shader_rec_size);
785 		return -EINVAL;
786 	}
787 	pkt_u = exec->shader_rec_u;
788 	pkt_v = exec->shader_rec_v;
789 	memcpy(pkt_v, pkt_u, packet_size);
790 	exec->shader_rec_u += packet_size;
791 	/* Shader recs have to be aligned to 16 bytes (due to the attribute
792 	 * flags being in the low bytes), so round the next validated shader
793 	 * rec address up.  This should be safe, since we've got so many
794 	 * relocations in a shader rec packet.
795 	 */
796 	BUG_ON(roundup(packet_size, 16) - packet_size > nr_relocs * 4);
797 	exec->shader_rec_v += roundup(packet_size, 16);
798 	exec->shader_rec_size -= packet_size;
799 
800 	for (i = 0; i < shader_reloc_count; i++) {
801 		if (src_handles[i] > exec->bo_count) {
802 			DRM_ERROR("Shader handle %d too big\n", src_handles[i]);
803 			return -EINVAL;
804 		}
805 
806 		bo[i] = exec->bo[src_handles[i]];
807 		if (!bo[i])
808 			return -EINVAL;
809 	}
810 	for (i = shader_reloc_count; i < nr_relocs; i++) {
811 		bo[i] = vc4_use_bo(exec, src_handles[i]);
812 		if (!bo[i])
813 			return -EINVAL;
814 	}
815 
816 	if (((*(uint16_t *)pkt_u & VC4_SHADER_FLAG_FS_SINGLE_THREAD) == 0) !=
817 	    to_vc4_bo(&bo[0]->base)->validated_shader->is_threaded) {
818 		DRM_ERROR("Thread mode of CL and FS do not match\n");
819 		return -EINVAL;
820 	}
821 
822 	if (to_vc4_bo(&bo[1]->base)->validated_shader->is_threaded ||
823 	    to_vc4_bo(&bo[2]->base)->validated_shader->is_threaded) {
824 		DRM_ERROR("cs and vs cannot be threaded\n");
825 		return -EINVAL;
826 	}
827 
828 	for (i = 0; i < shader_reloc_count; i++) {
829 		struct vc4_validated_shader_info *validated_shader;
830 		uint32_t o = shader_reloc_offsets[i];
831 		uint32_t src_offset = *(uint32_t *)(pkt_u + o);
832 		uint32_t *texture_handles_u;
833 		void *uniform_data_u;
834 		uint32_t tex, uni;
835 
836 		*(uint32_t *)(pkt_v + o) = bo[i]->paddr + src_offset;
837 
838 		if (src_offset != 0) {
839 			DRM_ERROR("Shaders must be at offset 0 of "
840 				  "the BO.\n");
841 			return -EINVAL;
842 		}
843 
844 		validated_shader = to_vc4_bo(&bo[i]->base)->validated_shader;
845 		if (!validated_shader)
846 			return -EINVAL;
847 
848 		if (validated_shader->uniforms_src_size >
849 		    exec->uniforms_size) {
850 			DRM_ERROR("Uniforms src buffer overflow\n");
851 			return -EINVAL;
852 		}
853 
854 		texture_handles_u = exec->uniforms_u;
855 		uniform_data_u = (texture_handles_u +
856 				  validated_shader->num_texture_samples);
857 
858 		memcpy(exec->uniforms_v, uniform_data_u,
859 		       validated_shader->uniforms_size);
860 
861 		for (tex = 0;
862 		     tex < validated_shader->num_texture_samples;
863 		     tex++) {
864 			if (!reloc_tex(exec,
865 				       uniform_data_u,
866 				       &validated_shader->texture_samples[tex],
867 				       texture_handles_u[tex],
868 				       i == 2)) {
869 				return -EINVAL;
870 			}
871 		}
872 
873 		/* Fill in the uniform slots that need this shader's
874 		 * start-of-uniforms address (used for resetting the uniform
875 		 * stream in the presence of control flow).
876 		 */
877 		for (uni = 0;
878 		     uni < validated_shader->num_uniform_addr_offsets;
879 		     uni++) {
880 			uint32_t o = validated_shader->uniform_addr_offsets[uni];
881 			((uint32_t *)exec->uniforms_v)[o] = exec->uniforms_p;
882 		}
883 
884 		*(uint32_t *)(pkt_v + o + 4) = exec->uniforms_p;
885 
886 		exec->uniforms_u += validated_shader->uniforms_src_size;
887 		exec->uniforms_v += validated_shader->uniforms_size;
888 		exec->uniforms_p += validated_shader->uniforms_size;
889 	}
890 
891 	for (i = 0; i < nr_attributes; i++) {
892 		struct drm_gem_cma_object *vbo =
893 			bo[ARRAY_SIZE(shader_reloc_offsets) + i];
894 		uint32_t o = 36 + i * 8;
895 		uint32_t offset = *(uint32_t *)(pkt_u + o + 0);
896 		uint32_t attr_size = *(uint8_t *)(pkt_u + o + 4) + 1;
897 		uint32_t stride = *(uint8_t *)(pkt_u + o + 5);
898 		uint32_t max_index;
899 
900 		exec->bin_dep_seqno = max(exec->bin_dep_seqno,
901 					  to_vc4_bo(&vbo->base)->write_seqno);
902 
903 		if (state->addr & 0x8)
904 			stride |= (*(uint32_t *)(pkt_u + 100 + i * 4)) & ~0xff;
905 
906 		if (vbo->base.size < offset ||
907 		    vbo->base.size - offset < attr_size) {
908 			DRM_ERROR("BO offset overflow (%d + %d > %zu)\n",
909 				  offset, attr_size, vbo->base.size);
910 			return -EINVAL;
911 		}
912 
913 		if (stride != 0) {
914 			max_index = ((vbo->base.size - offset - attr_size) /
915 				     stride);
916 			if (state->max_index > max_index) {
917 				DRM_ERROR("primitives use index %d out of "
918 					  "supplied %d\n",
919 					  state->max_index, max_index);
920 				return -EINVAL;
921 			}
922 		}
923 
924 		*(uint32_t *)(pkt_v + o) = vbo->paddr + offset;
925 	}
926 
927 	return 0;
928 }
929 
930 int
931 vc4_validate_shader_recs(struct drm_device *dev,
932 			 struct vc4_exec_info *exec)
933 {
934 	uint32_t i;
935 	int ret = 0;
936 
937 	for (i = 0; i < exec->shader_state_count; i++) {
938 		ret = validate_gl_shader_rec(dev, exec, &exec->shader_state[i]);
939 		if (ret)
940 			return ret;
941 	}
942 
943 	return ret;
944 }
945