xref: /openbmc/linux/drivers/gpu/drm/vc4/vc4_validate.c (revision 1c2dd16a)
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  * Command list validator for VC4.
26  *
27  * The VC4 has no IOMMU between it and system memory.  So, a user with
28  * 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 texture, or
31  * uniform data, or vertex data).
32  *
33  * This validates command lists to ensure that all accesses are within
34  * the bounds of the GEM objects referenced.  It explicitly whitelists
35  * packets, and looks at the offsets in any address fields to make
36  * sure they're constrained within the BOs they reference.
37  *
38  * Note that because of the validation that's happening anyway, this
39  * is where GEM relocation processing happens.
40  */
41 
42 #include "uapi/drm/vc4_drm.h"
43 #include "vc4_drv.h"
44 #include "vc4_packet.h"
45 
46 #define VALIDATE_ARGS \
47 	struct vc4_exec_info *exec,			\
48 	void *validated,				\
49 	void *untrusted
50 
51 /** Return the width in pixels of a 64-byte microtile. */
52 static uint32_t
53 utile_width(int cpp)
54 {
55 	switch (cpp) {
56 	case 1:
57 	case 2:
58 		return 8;
59 	case 4:
60 		return 4;
61 	case 8:
62 		return 2;
63 	default:
64 		DRM_ERROR("unknown cpp: %d\n", cpp);
65 		return 1;
66 	}
67 }
68 
69 /** Return the height in pixels of a 64-byte microtile. */
70 static uint32_t
71 utile_height(int cpp)
72 {
73 	switch (cpp) {
74 	case 1:
75 		return 8;
76 	case 2:
77 	case 4:
78 	case 8:
79 		return 4;
80 	default:
81 		DRM_ERROR("unknown cpp: %d\n", cpp);
82 		return 1;
83 	}
84 }
85 
86 /**
87  * The texture unit decides what tiling format a particular miplevel is using
88  * this function, so we lay out our miptrees accordingly.
89  */
90 static bool
91 size_is_lt(uint32_t width, uint32_t height, int cpp)
92 {
93 	return (width <= 4 * utile_width(cpp) ||
94 		height <= 4 * utile_height(cpp));
95 }
96 
97 struct drm_gem_cma_object *
98 vc4_use_bo(struct vc4_exec_info *exec, uint32_t hindex)
99 {
100 	struct drm_gem_cma_object *obj;
101 	struct vc4_bo *bo;
102 
103 	if (hindex >= exec->bo_count) {
104 		DRM_ERROR("BO index %d greater than BO count %d\n",
105 			  hindex, exec->bo_count);
106 		return NULL;
107 	}
108 	obj = exec->bo[hindex];
109 	bo = to_vc4_bo(&obj->base);
110 
111 	if (bo->validated_shader) {
112 		DRM_ERROR("Trying to use shader BO as something other than "
113 			  "a shader\n");
114 		return NULL;
115 	}
116 
117 	return obj;
118 }
119 
120 static struct drm_gem_cma_object *
121 vc4_use_handle(struct vc4_exec_info *exec, uint32_t gem_handles_packet_index)
122 {
123 	return vc4_use_bo(exec, exec->bo_index[gem_handles_packet_index]);
124 }
125 
126 static bool
127 validate_bin_pos(struct vc4_exec_info *exec, void *untrusted, uint32_t pos)
128 {
129 	/* Note that the untrusted pointer passed to these functions is
130 	 * incremented past the packet byte.
131 	 */
132 	return (untrusted - 1 == exec->bin_u + pos);
133 }
134 
135 static uint32_t
136 gl_shader_rec_size(uint32_t pointer_bits)
137 {
138 	uint32_t attribute_count = pointer_bits & 7;
139 	bool extended = pointer_bits & 8;
140 
141 	if (attribute_count == 0)
142 		attribute_count = 8;
143 
144 	if (extended)
145 		return 100 + attribute_count * 4;
146 	else
147 		return 36 + attribute_count * 8;
148 }
149 
150 bool
151 vc4_check_tex_size(struct vc4_exec_info *exec, struct drm_gem_cma_object *fbo,
152 		   uint32_t offset, uint8_t tiling_format,
153 		   uint32_t width, uint32_t height, uint8_t cpp)
154 {
155 	uint32_t aligned_width, aligned_height, stride, size;
156 	uint32_t utile_w = utile_width(cpp);
157 	uint32_t utile_h = utile_height(cpp);
158 
159 	/* The shaded vertex format stores signed 12.4 fixed point
160 	 * (-2048,2047) offsets from the viewport center, so we should
161 	 * never have a render target larger than 4096.  The texture
162 	 * unit can only sample from 2048x2048, so it's even more
163 	 * restricted.  This lets us avoid worrying about overflow in
164 	 * our math.
165 	 */
166 	if (width > 4096 || height > 4096) {
167 		DRM_ERROR("Surface dimesions (%d,%d) too large", width, height);
168 		return false;
169 	}
170 
171 	switch (tiling_format) {
172 	case VC4_TILING_FORMAT_LINEAR:
173 		aligned_width = round_up(width, utile_w);
174 		aligned_height = height;
175 		break;
176 	case VC4_TILING_FORMAT_T:
177 		aligned_width = round_up(width, utile_w * 8);
178 		aligned_height = round_up(height, utile_h * 8);
179 		break;
180 	case VC4_TILING_FORMAT_LT:
181 		aligned_width = round_up(width, utile_w);
182 		aligned_height = round_up(height, utile_h);
183 		break;
184 	default:
185 		DRM_ERROR("buffer tiling %d unsupported\n", tiling_format);
186 		return false;
187 	}
188 
189 	stride = aligned_width * cpp;
190 	size = stride * aligned_height;
191 
192 	if (size + offset < size ||
193 	    size + offset > fbo->base.size) {
194 		DRM_ERROR("Overflow in %dx%d (%dx%d) fbo size (%d + %d > %zd)\n",
195 			  width, height,
196 			  aligned_width, aligned_height,
197 			  size, offset, fbo->base.size);
198 		return false;
199 	}
200 
201 	return true;
202 }
203 
204 static int
205 validate_flush(VALIDATE_ARGS)
206 {
207 	if (!validate_bin_pos(exec, untrusted, exec->args->bin_cl_size - 1)) {
208 		DRM_ERROR("Bin CL must end with VC4_PACKET_FLUSH\n");
209 		return -EINVAL;
210 	}
211 	exec->found_flush = true;
212 
213 	return 0;
214 }
215 
216 static int
217 validate_start_tile_binning(VALIDATE_ARGS)
218 {
219 	if (exec->found_start_tile_binning_packet) {
220 		DRM_ERROR("Duplicate VC4_PACKET_START_TILE_BINNING\n");
221 		return -EINVAL;
222 	}
223 	exec->found_start_tile_binning_packet = true;
224 
225 	if (!exec->found_tile_binning_mode_config_packet) {
226 		DRM_ERROR("missing VC4_PACKET_TILE_BINNING_MODE_CONFIG\n");
227 		return -EINVAL;
228 	}
229 
230 	return 0;
231 }
232 
233 static int
234 validate_increment_semaphore(VALIDATE_ARGS)
235 {
236 	if (!validate_bin_pos(exec, untrusted, exec->args->bin_cl_size - 2)) {
237 		DRM_ERROR("Bin CL must end with "
238 			  "VC4_PACKET_INCREMENT_SEMAPHORE\n");
239 		return -EINVAL;
240 	}
241 	exec->found_increment_semaphore_packet = true;
242 
243 	return 0;
244 }
245 
246 static int
247 validate_indexed_prim_list(VALIDATE_ARGS)
248 {
249 	struct drm_gem_cma_object *ib;
250 	uint32_t length = *(uint32_t *)(untrusted + 1);
251 	uint32_t offset = *(uint32_t *)(untrusted + 5);
252 	uint32_t max_index = *(uint32_t *)(untrusted + 9);
253 	uint32_t index_size = (*(uint8_t *)(untrusted + 0) >> 4) ? 2 : 1;
254 	struct vc4_shader_state *shader_state;
255 
256 	/* Check overflow condition */
257 	if (exec->shader_state_count == 0) {
258 		DRM_ERROR("shader state must precede primitives\n");
259 		return -EINVAL;
260 	}
261 	shader_state = &exec->shader_state[exec->shader_state_count - 1];
262 
263 	if (max_index > shader_state->max_index)
264 		shader_state->max_index = max_index;
265 
266 	ib = vc4_use_handle(exec, 0);
267 	if (!ib)
268 		return -EINVAL;
269 
270 	exec->bin_dep_seqno = max(exec->bin_dep_seqno,
271 				  to_vc4_bo(&ib->base)->write_seqno);
272 
273 	if (offset > ib->base.size ||
274 	    (ib->base.size - offset) / index_size < length) {
275 		DRM_ERROR("IB access overflow (%d + %d*%d > %zd)\n",
276 			  offset, length, index_size, ib->base.size);
277 		return -EINVAL;
278 	}
279 
280 	*(uint32_t *)(validated + 5) = ib->paddr + offset;
281 
282 	return 0;
283 }
284 
285 static int
286 validate_gl_array_primitive(VALIDATE_ARGS)
287 {
288 	uint32_t length = *(uint32_t *)(untrusted + 1);
289 	uint32_t base_index = *(uint32_t *)(untrusted + 5);
290 	uint32_t max_index;
291 	struct vc4_shader_state *shader_state;
292 
293 	/* Check overflow condition */
294 	if (exec->shader_state_count == 0) {
295 		DRM_ERROR("shader state must precede primitives\n");
296 		return -EINVAL;
297 	}
298 	shader_state = &exec->shader_state[exec->shader_state_count - 1];
299 
300 	if (length + base_index < length) {
301 		DRM_ERROR("primitive vertex count overflow\n");
302 		return -EINVAL;
303 	}
304 	max_index = length + base_index - 1;
305 
306 	if (max_index > shader_state->max_index)
307 		shader_state->max_index = max_index;
308 
309 	return 0;
310 }
311 
312 static int
313 validate_gl_shader_state(VALIDATE_ARGS)
314 {
315 	uint32_t i = exec->shader_state_count++;
316 
317 	if (i >= exec->shader_state_size) {
318 		DRM_ERROR("More requests for shader states than declared\n");
319 		return -EINVAL;
320 	}
321 
322 	exec->shader_state[i].addr = *(uint32_t *)untrusted;
323 	exec->shader_state[i].max_index = 0;
324 
325 	if (exec->shader_state[i].addr & ~0xf) {
326 		DRM_ERROR("high bits set in GL shader rec reference\n");
327 		return -EINVAL;
328 	}
329 
330 	*(uint32_t *)validated = (exec->shader_rec_p +
331 				  exec->shader_state[i].addr);
332 
333 	exec->shader_rec_p +=
334 		roundup(gl_shader_rec_size(exec->shader_state[i].addr), 16);
335 
336 	return 0;
337 }
338 
339 static int
340 validate_tile_binning_config(VALIDATE_ARGS)
341 {
342 	struct drm_device *dev = exec->exec_bo->base.dev;
343 	struct vc4_bo *tile_bo;
344 	uint8_t flags;
345 	uint32_t tile_state_size, tile_alloc_size;
346 	uint32_t tile_count;
347 
348 	if (exec->found_tile_binning_mode_config_packet) {
349 		DRM_ERROR("Duplicate VC4_PACKET_TILE_BINNING_MODE_CONFIG\n");
350 		return -EINVAL;
351 	}
352 	exec->found_tile_binning_mode_config_packet = true;
353 
354 	exec->bin_tiles_x = *(uint8_t *)(untrusted + 12);
355 	exec->bin_tiles_y = *(uint8_t *)(untrusted + 13);
356 	tile_count = exec->bin_tiles_x * exec->bin_tiles_y;
357 	flags = *(uint8_t *)(untrusted + 14);
358 
359 	if (exec->bin_tiles_x == 0 ||
360 	    exec->bin_tiles_y == 0) {
361 		DRM_ERROR("Tile binning config of %dx%d too small\n",
362 			  exec->bin_tiles_x, exec->bin_tiles_y);
363 		return -EINVAL;
364 	}
365 
366 	if (flags & (VC4_BIN_CONFIG_DB_NON_MS |
367 		     VC4_BIN_CONFIG_TILE_BUFFER_64BIT)) {
368 		DRM_ERROR("unsupported binning config flags 0x%02x\n", flags);
369 		return -EINVAL;
370 	}
371 
372 	/* The tile state data array is 48 bytes per tile, and we put it at
373 	 * the start of a BO containing both it and the tile alloc.
374 	 */
375 	tile_state_size = 48 * tile_count;
376 
377 	/* Since the tile alloc array will follow us, align. */
378 	exec->tile_alloc_offset = roundup(tile_state_size, 4096);
379 
380 	*(uint8_t *)(validated + 14) =
381 		((flags & ~(VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_MASK |
382 			    VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_MASK)) |
383 		 VC4_BIN_CONFIG_AUTO_INIT_TSDA |
384 		 VC4_SET_FIELD(VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_32,
385 			       VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE) |
386 		 VC4_SET_FIELD(VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_128,
387 			       VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE));
388 
389 	/* Initial block size. */
390 	tile_alloc_size = 32 * tile_count;
391 
392 	/*
393 	 * The initial allocation gets rounded to the next 256 bytes before
394 	 * the hardware starts fulfilling further allocations.
395 	 */
396 	tile_alloc_size = roundup(tile_alloc_size, 256);
397 
398 	/* Add space for the extra allocations.  This is what gets used first,
399 	 * before overflow memory.  It must have at least 4096 bytes, but we
400 	 * want to avoid overflow memory usage if possible.
401 	 */
402 	tile_alloc_size += 1024 * 1024;
403 
404 	tile_bo = vc4_bo_create(dev, exec->tile_alloc_offset + tile_alloc_size,
405 				true);
406 	exec->tile_bo = &tile_bo->base;
407 	if (IS_ERR(exec->tile_bo))
408 		return PTR_ERR(exec->tile_bo);
409 	list_add_tail(&tile_bo->unref_head, &exec->unref_list);
410 
411 	/* tile alloc address. */
412 	*(uint32_t *)(validated + 0) = (exec->tile_bo->paddr +
413 					exec->tile_alloc_offset);
414 	/* tile alloc size. */
415 	*(uint32_t *)(validated + 4) = tile_alloc_size;
416 	/* tile state address. */
417 	*(uint32_t *)(validated + 8) = exec->tile_bo->paddr;
418 
419 	return 0;
420 }
421 
422 static int
423 validate_gem_handles(VALIDATE_ARGS)
424 {
425 	memcpy(exec->bo_index, untrusted, sizeof(exec->bo_index));
426 	return 0;
427 }
428 
429 #define VC4_DEFINE_PACKET(packet, func) \
430 	[packet] = { packet ## _SIZE, #packet, func }
431 
432 static const struct cmd_info {
433 	uint16_t len;
434 	const char *name;
435 	int (*func)(struct vc4_exec_info *exec, void *validated,
436 		    void *untrusted);
437 } cmd_info[] = {
438 	VC4_DEFINE_PACKET(VC4_PACKET_HALT, NULL),
439 	VC4_DEFINE_PACKET(VC4_PACKET_NOP, NULL),
440 	VC4_DEFINE_PACKET(VC4_PACKET_FLUSH, validate_flush),
441 	VC4_DEFINE_PACKET(VC4_PACKET_FLUSH_ALL, NULL),
442 	VC4_DEFINE_PACKET(VC4_PACKET_START_TILE_BINNING,
443 			  validate_start_tile_binning),
444 	VC4_DEFINE_PACKET(VC4_PACKET_INCREMENT_SEMAPHORE,
445 			  validate_increment_semaphore),
446 
447 	VC4_DEFINE_PACKET(VC4_PACKET_GL_INDEXED_PRIMITIVE,
448 			  validate_indexed_prim_list),
449 	VC4_DEFINE_PACKET(VC4_PACKET_GL_ARRAY_PRIMITIVE,
450 			  validate_gl_array_primitive),
451 
452 	VC4_DEFINE_PACKET(VC4_PACKET_PRIMITIVE_LIST_FORMAT, NULL),
453 
454 	VC4_DEFINE_PACKET(VC4_PACKET_GL_SHADER_STATE, validate_gl_shader_state),
455 
456 	VC4_DEFINE_PACKET(VC4_PACKET_CONFIGURATION_BITS, NULL),
457 	VC4_DEFINE_PACKET(VC4_PACKET_FLAT_SHADE_FLAGS, NULL),
458 	VC4_DEFINE_PACKET(VC4_PACKET_POINT_SIZE, NULL),
459 	VC4_DEFINE_PACKET(VC4_PACKET_LINE_WIDTH, NULL),
460 	VC4_DEFINE_PACKET(VC4_PACKET_RHT_X_BOUNDARY, NULL),
461 	VC4_DEFINE_PACKET(VC4_PACKET_DEPTH_OFFSET, NULL),
462 	VC4_DEFINE_PACKET(VC4_PACKET_CLIP_WINDOW, NULL),
463 	VC4_DEFINE_PACKET(VC4_PACKET_VIEWPORT_OFFSET, NULL),
464 	VC4_DEFINE_PACKET(VC4_PACKET_CLIPPER_XY_SCALING, NULL),
465 	/* Note: The docs say this was also 105, but it was 106 in the
466 	 * initial userland code drop.
467 	 */
468 	VC4_DEFINE_PACKET(VC4_PACKET_CLIPPER_Z_SCALING, NULL),
469 
470 	VC4_DEFINE_PACKET(VC4_PACKET_TILE_BINNING_MODE_CONFIG,
471 			  validate_tile_binning_config),
472 
473 	VC4_DEFINE_PACKET(VC4_PACKET_GEM_HANDLES, validate_gem_handles),
474 };
475 
476 int
477 vc4_validate_bin_cl(struct drm_device *dev,
478 		    void *validated,
479 		    void *unvalidated,
480 		    struct vc4_exec_info *exec)
481 {
482 	uint32_t len = exec->args->bin_cl_size;
483 	uint32_t dst_offset = 0;
484 	uint32_t src_offset = 0;
485 
486 	while (src_offset < len) {
487 		void *dst_pkt = validated + dst_offset;
488 		void *src_pkt = unvalidated + src_offset;
489 		u8 cmd = *(uint8_t *)src_pkt;
490 		const struct cmd_info *info;
491 
492 		if (cmd >= ARRAY_SIZE(cmd_info)) {
493 			DRM_ERROR("0x%08x: packet %d out of bounds\n",
494 				  src_offset, cmd);
495 			return -EINVAL;
496 		}
497 
498 		info = &cmd_info[cmd];
499 		if (!info->name) {
500 			DRM_ERROR("0x%08x: packet %d invalid\n",
501 				  src_offset, cmd);
502 			return -EINVAL;
503 		}
504 
505 		if (src_offset + info->len > len) {
506 			DRM_ERROR("0x%08x: packet %d (%s) length 0x%08x "
507 				  "exceeds bounds (0x%08x)\n",
508 				  src_offset, cmd, info->name, info->len,
509 				  src_offset + len);
510 			return -EINVAL;
511 		}
512 
513 		if (cmd != VC4_PACKET_GEM_HANDLES)
514 			memcpy(dst_pkt, src_pkt, info->len);
515 
516 		if (info->func && info->func(exec,
517 					     dst_pkt + 1,
518 					     src_pkt + 1)) {
519 			DRM_ERROR("0x%08x: packet %d (%s) failed to validate\n",
520 				  src_offset, cmd, info->name);
521 			return -EINVAL;
522 		}
523 
524 		src_offset += info->len;
525 		/* GEM handle loading doesn't produce HW packets. */
526 		if (cmd != VC4_PACKET_GEM_HANDLES)
527 			dst_offset += info->len;
528 
529 		/* When the CL hits halt, it'll stop reading anything else. */
530 		if (cmd == VC4_PACKET_HALT)
531 			break;
532 	}
533 
534 	exec->ct0ea = exec->ct0ca + dst_offset;
535 
536 	if (!exec->found_start_tile_binning_packet) {
537 		DRM_ERROR("Bin CL missing VC4_PACKET_START_TILE_BINNING\n");
538 		return -EINVAL;
539 	}
540 
541 	/* The bin CL must be ended with INCREMENT_SEMAPHORE and FLUSH.  The
542 	 * semaphore is used to trigger the render CL to start up, and the
543 	 * FLUSH is what caps the bin lists with
544 	 * VC4_PACKET_RETURN_FROM_SUB_LIST (so they jump back to the main
545 	 * render CL when they get called to) and actually triggers the queued
546 	 * semaphore increment.
547 	 */
548 	if (!exec->found_increment_semaphore_packet || !exec->found_flush) {
549 		DRM_ERROR("Bin CL missing VC4_PACKET_INCREMENT_SEMAPHORE + "
550 			  "VC4_PACKET_FLUSH\n");
551 		return -EINVAL;
552 	}
553 
554 	return 0;
555 }
556 
557 static bool
558 reloc_tex(struct vc4_exec_info *exec,
559 	  void *uniform_data_u,
560 	  struct vc4_texture_sample_info *sample,
561 	  uint32_t texture_handle_index, bool is_cs)
562 {
563 	struct drm_gem_cma_object *tex;
564 	uint32_t p0 = *(uint32_t *)(uniform_data_u + sample->p_offset[0]);
565 	uint32_t p1 = *(uint32_t *)(uniform_data_u + sample->p_offset[1]);
566 	uint32_t p2 = (sample->p_offset[2] != ~0 ?
567 		       *(uint32_t *)(uniform_data_u + sample->p_offset[2]) : 0);
568 	uint32_t p3 = (sample->p_offset[3] != ~0 ?
569 		       *(uint32_t *)(uniform_data_u + sample->p_offset[3]) : 0);
570 	uint32_t *validated_p0 = exec->uniforms_v + sample->p_offset[0];
571 	uint32_t offset = p0 & VC4_TEX_P0_OFFSET_MASK;
572 	uint32_t miplevels = VC4_GET_FIELD(p0, VC4_TEX_P0_MIPLVLS);
573 	uint32_t width = VC4_GET_FIELD(p1, VC4_TEX_P1_WIDTH);
574 	uint32_t height = VC4_GET_FIELD(p1, VC4_TEX_P1_HEIGHT);
575 	uint32_t cpp, tiling_format, utile_w, utile_h;
576 	uint32_t i;
577 	uint32_t cube_map_stride = 0;
578 	enum vc4_texture_data_type type;
579 
580 	tex = vc4_use_bo(exec, texture_handle_index);
581 	if (!tex)
582 		return false;
583 
584 	if (sample->is_direct) {
585 		uint32_t remaining_size = tex->base.size - p0;
586 
587 		if (p0 > tex->base.size - 4) {
588 			DRM_ERROR("UBO offset greater than UBO size\n");
589 			goto fail;
590 		}
591 		if (p1 > remaining_size - 4) {
592 			DRM_ERROR("UBO clamp would allow reads "
593 				  "outside of UBO\n");
594 			goto fail;
595 		}
596 		*validated_p0 = tex->paddr + p0;
597 		return true;
598 	}
599 
600 	if (width == 0)
601 		width = 2048;
602 	if (height == 0)
603 		height = 2048;
604 
605 	if (p0 & VC4_TEX_P0_CMMODE_MASK) {
606 		if (VC4_GET_FIELD(p2, VC4_TEX_P2_PTYPE) ==
607 		    VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE)
608 			cube_map_stride = p2 & VC4_TEX_P2_CMST_MASK;
609 		if (VC4_GET_FIELD(p3, VC4_TEX_P2_PTYPE) ==
610 		    VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE) {
611 			if (cube_map_stride) {
612 				DRM_ERROR("Cube map stride set twice\n");
613 				goto fail;
614 			}
615 
616 			cube_map_stride = p3 & VC4_TEX_P2_CMST_MASK;
617 		}
618 		if (!cube_map_stride) {
619 			DRM_ERROR("Cube map stride not set\n");
620 			goto fail;
621 		}
622 	}
623 
624 	type = (VC4_GET_FIELD(p0, VC4_TEX_P0_TYPE) |
625 		(VC4_GET_FIELD(p1, VC4_TEX_P1_TYPE4) << 4));
626 
627 	switch (type) {
628 	case VC4_TEXTURE_TYPE_RGBA8888:
629 	case VC4_TEXTURE_TYPE_RGBX8888:
630 	case VC4_TEXTURE_TYPE_RGBA32R:
631 		cpp = 4;
632 		break;
633 	case VC4_TEXTURE_TYPE_RGBA4444:
634 	case VC4_TEXTURE_TYPE_RGBA5551:
635 	case VC4_TEXTURE_TYPE_RGB565:
636 	case VC4_TEXTURE_TYPE_LUMALPHA:
637 	case VC4_TEXTURE_TYPE_S16F:
638 	case VC4_TEXTURE_TYPE_S16:
639 		cpp = 2;
640 		break;
641 	case VC4_TEXTURE_TYPE_LUMINANCE:
642 	case VC4_TEXTURE_TYPE_ALPHA:
643 	case VC4_TEXTURE_TYPE_S8:
644 		cpp = 1;
645 		break;
646 	case VC4_TEXTURE_TYPE_ETC1:
647 		/* ETC1 is arranged as 64-bit blocks, where each block is 4x4
648 		 * pixels.
649 		 */
650 		cpp = 8;
651 		width = (width + 3) >> 2;
652 		height = (height + 3) >> 2;
653 		break;
654 	case VC4_TEXTURE_TYPE_BW1:
655 	case VC4_TEXTURE_TYPE_A4:
656 	case VC4_TEXTURE_TYPE_A1:
657 	case VC4_TEXTURE_TYPE_RGBA64:
658 	case VC4_TEXTURE_TYPE_YUV422R:
659 	default:
660 		DRM_ERROR("Texture format %d unsupported\n", type);
661 		goto fail;
662 	}
663 	utile_w = utile_width(cpp);
664 	utile_h = utile_height(cpp);
665 
666 	if (type == VC4_TEXTURE_TYPE_RGBA32R) {
667 		tiling_format = VC4_TILING_FORMAT_LINEAR;
668 	} else {
669 		if (size_is_lt(width, height, cpp))
670 			tiling_format = VC4_TILING_FORMAT_LT;
671 		else
672 			tiling_format = VC4_TILING_FORMAT_T;
673 	}
674 
675 	if (!vc4_check_tex_size(exec, tex, offset + cube_map_stride * 5,
676 				tiling_format, width, height, cpp)) {
677 		goto fail;
678 	}
679 
680 	/* The mipmap levels are stored before the base of the texture.  Make
681 	 * sure there is actually space in the BO.
682 	 */
683 	for (i = 1; i <= miplevels; i++) {
684 		uint32_t level_width = max(width >> i, 1u);
685 		uint32_t level_height = max(height >> i, 1u);
686 		uint32_t aligned_width, aligned_height;
687 		uint32_t level_size;
688 
689 		/* Once the levels get small enough, they drop from T to LT. */
690 		if (tiling_format == VC4_TILING_FORMAT_T &&
691 		    size_is_lt(level_width, level_height, cpp)) {
692 			tiling_format = VC4_TILING_FORMAT_LT;
693 		}
694 
695 		switch (tiling_format) {
696 		case VC4_TILING_FORMAT_T:
697 			aligned_width = round_up(level_width, utile_w * 8);
698 			aligned_height = round_up(level_height, utile_h * 8);
699 			break;
700 		case VC4_TILING_FORMAT_LT:
701 			aligned_width = round_up(level_width, utile_w);
702 			aligned_height = round_up(level_height, utile_h);
703 			break;
704 		default:
705 			aligned_width = round_up(level_width, utile_w);
706 			aligned_height = level_height;
707 			break;
708 		}
709 
710 		level_size = aligned_width * cpp * aligned_height;
711 
712 		if (offset < level_size) {
713 			DRM_ERROR("Level %d (%dx%d -> %dx%d) size %db "
714 				  "overflowed buffer bounds (offset %d)\n",
715 				  i, level_width, level_height,
716 				  aligned_width, aligned_height,
717 				  level_size, offset);
718 			goto fail;
719 		}
720 
721 		offset -= level_size;
722 	}
723 
724 	*validated_p0 = tex->paddr + p0;
725 
726 	if (is_cs) {
727 		exec->bin_dep_seqno = max(exec->bin_dep_seqno,
728 					  to_vc4_bo(&tex->base)->write_seqno);
729 	}
730 
731 	return true;
732  fail:
733 	DRM_INFO("Texture p0 at %d: 0x%08x\n", sample->p_offset[0], p0);
734 	DRM_INFO("Texture p1 at %d: 0x%08x\n", sample->p_offset[1], p1);
735 	DRM_INFO("Texture p2 at %d: 0x%08x\n", sample->p_offset[2], p2);
736 	DRM_INFO("Texture p3 at %d: 0x%08x\n", sample->p_offset[3], p3);
737 	return false;
738 }
739 
740 static int
741 validate_gl_shader_rec(struct drm_device *dev,
742 		       struct vc4_exec_info *exec,
743 		       struct vc4_shader_state *state)
744 {
745 	uint32_t *src_handles;
746 	void *pkt_u, *pkt_v;
747 	static const uint32_t shader_reloc_offsets[] = {
748 		4, /* fs */
749 		16, /* vs */
750 		28, /* cs */
751 	};
752 	uint32_t shader_reloc_count = ARRAY_SIZE(shader_reloc_offsets);
753 	struct drm_gem_cma_object *bo[shader_reloc_count + 8];
754 	uint32_t nr_attributes, nr_relocs, packet_size;
755 	int i;
756 
757 	nr_attributes = state->addr & 0x7;
758 	if (nr_attributes == 0)
759 		nr_attributes = 8;
760 	packet_size = gl_shader_rec_size(state->addr);
761 
762 	nr_relocs = ARRAY_SIZE(shader_reloc_offsets) + nr_attributes;
763 	if (nr_relocs * 4 > exec->shader_rec_size) {
764 		DRM_ERROR("overflowed shader recs reading %d handles "
765 			  "from %d bytes left\n",
766 			  nr_relocs, exec->shader_rec_size);
767 		return -EINVAL;
768 	}
769 	src_handles = exec->shader_rec_u;
770 	exec->shader_rec_u += nr_relocs * 4;
771 	exec->shader_rec_size -= nr_relocs * 4;
772 
773 	if (packet_size > exec->shader_rec_size) {
774 		DRM_ERROR("overflowed shader recs copying %db packet "
775 			  "from %d bytes left\n",
776 			  packet_size, exec->shader_rec_size);
777 		return -EINVAL;
778 	}
779 	pkt_u = exec->shader_rec_u;
780 	pkt_v = exec->shader_rec_v;
781 	memcpy(pkt_v, pkt_u, packet_size);
782 	exec->shader_rec_u += packet_size;
783 	/* Shader recs have to be aligned to 16 bytes (due to the attribute
784 	 * flags being in the low bytes), so round the next validated shader
785 	 * rec address up.  This should be safe, since we've got so many
786 	 * relocations in a shader rec packet.
787 	 */
788 	BUG_ON(roundup(packet_size, 16) - packet_size > nr_relocs * 4);
789 	exec->shader_rec_v += roundup(packet_size, 16);
790 	exec->shader_rec_size -= packet_size;
791 
792 	for (i = 0; i < shader_reloc_count; i++) {
793 		if (src_handles[i] > exec->bo_count) {
794 			DRM_ERROR("Shader handle %d too big\n", src_handles[i]);
795 			return -EINVAL;
796 		}
797 
798 		bo[i] = exec->bo[src_handles[i]];
799 		if (!bo[i])
800 			return -EINVAL;
801 	}
802 	for (i = shader_reloc_count; i < nr_relocs; i++) {
803 		bo[i] = vc4_use_bo(exec, src_handles[i]);
804 		if (!bo[i])
805 			return -EINVAL;
806 	}
807 
808 	if (((*(uint16_t *)pkt_u & VC4_SHADER_FLAG_FS_SINGLE_THREAD) == 0) !=
809 	    to_vc4_bo(&bo[0]->base)->validated_shader->is_threaded) {
810 		DRM_ERROR("Thread mode of CL and FS do not match\n");
811 		return -EINVAL;
812 	}
813 
814 	if (to_vc4_bo(&bo[1]->base)->validated_shader->is_threaded ||
815 	    to_vc4_bo(&bo[2]->base)->validated_shader->is_threaded) {
816 		DRM_ERROR("cs and vs cannot be threaded\n");
817 		return -EINVAL;
818 	}
819 
820 	for (i = 0; i < shader_reloc_count; i++) {
821 		struct vc4_validated_shader_info *validated_shader;
822 		uint32_t o = shader_reloc_offsets[i];
823 		uint32_t src_offset = *(uint32_t *)(pkt_u + o);
824 		uint32_t *texture_handles_u;
825 		void *uniform_data_u;
826 		uint32_t tex, uni;
827 
828 		*(uint32_t *)(pkt_v + o) = bo[i]->paddr + src_offset;
829 
830 		if (src_offset != 0) {
831 			DRM_ERROR("Shaders must be at offset 0 of "
832 				  "the BO.\n");
833 			return -EINVAL;
834 		}
835 
836 		validated_shader = to_vc4_bo(&bo[i]->base)->validated_shader;
837 		if (!validated_shader)
838 			return -EINVAL;
839 
840 		if (validated_shader->uniforms_src_size >
841 		    exec->uniforms_size) {
842 			DRM_ERROR("Uniforms src buffer overflow\n");
843 			return -EINVAL;
844 		}
845 
846 		texture_handles_u = exec->uniforms_u;
847 		uniform_data_u = (texture_handles_u +
848 				  validated_shader->num_texture_samples);
849 
850 		memcpy(exec->uniforms_v, uniform_data_u,
851 		       validated_shader->uniforms_size);
852 
853 		for (tex = 0;
854 		     tex < validated_shader->num_texture_samples;
855 		     tex++) {
856 			if (!reloc_tex(exec,
857 				       uniform_data_u,
858 				       &validated_shader->texture_samples[tex],
859 				       texture_handles_u[tex],
860 				       i == 2)) {
861 				return -EINVAL;
862 			}
863 		}
864 
865 		/* Fill in the uniform slots that need this shader's
866 		 * start-of-uniforms address (used for resetting the uniform
867 		 * stream in the presence of control flow).
868 		 */
869 		for (uni = 0;
870 		     uni < validated_shader->num_uniform_addr_offsets;
871 		     uni++) {
872 			uint32_t o = validated_shader->uniform_addr_offsets[uni];
873 			((uint32_t *)exec->uniforms_v)[o] = exec->uniforms_p;
874 		}
875 
876 		*(uint32_t *)(pkt_v + o + 4) = exec->uniforms_p;
877 
878 		exec->uniforms_u += validated_shader->uniforms_src_size;
879 		exec->uniforms_v += validated_shader->uniforms_size;
880 		exec->uniforms_p += validated_shader->uniforms_size;
881 	}
882 
883 	for (i = 0; i < nr_attributes; i++) {
884 		struct drm_gem_cma_object *vbo =
885 			bo[ARRAY_SIZE(shader_reloc_offsets) + i];
886 		uint32_t o = 36 + i * 8;
887 		uint32_t offset = *(uint32_t *)(pkt_u + o + 0);
888 		uint32_t attr_size = *(uint8_t *)(pkt_u + o + 4) + 1;
889 		uint32_t stride = *(uint8_t *)(pkt_u + o + 5);
890 		uint32_t max_index;
891 
892 		exec->bin_dep_seqno = max(exec->bin_dep_seqno,
893 					  to_vc4_bo(&vbo->base)->write_seqno);
894 
895 		if (state->addr & 0x8)
896 			stride |= (*(uint32_t *)(pkt_u + 100 + i * 4)) & ~0xff;
897 
898 		if (vbo->base.size < offset ||
899 		    vbo->base.size - offset < attr_size) {
900 			DRM_ERROR("BO offset overflow (%d + %d > %zu)\n",
901 				  offset, attr_size, vbo->base.size);
902 			return -EINVAL;
903 		}
904 
905 		if (stride != 0) {
906 			max_index = ((vbo->base.size - offset - attr_size) /
907 				     stride);
908 			if (state->max_index > max_index) {
909 				DRM_ERROR("primitives use index %d out of "
910 					  "supplied %d\n",
911 					  state->max_index, max_index);
912 				return -EINVAL;
913 			}
914 		}
915 
916 		*(uint32_t *)(pkt_v + o) = vbo->paddr + offset;
917 	}
918 
919 	return 0;
920 }
921 
922 int
923 vc4_validate_shader_recs(struct drm_device *dev,
924 			 struct vc4_exec_info *exec)
925 {
926 	uint32_t i;
927 	int ret = 0;
928 
929 	for (i = 0; i < exec->shader_state_count; i++) {
930 		ret = validate_gl_shader_rec(dev, exec, &exec->shader_state[i]);
931 		if (ret)
932 			return ret;
933 	}
934 
935 	return ret;
936 }
937